drivers/gpu/drm/i915/intel_pm.c

   1 /*
   2  * Copyright © 2012 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Eugeni Dodonov <eugeni.dodonov@intel.com>
  25  *
  26  */
  27
  28 #include <linux/cpufreq.h>
  29 #include <drm/drm_plane_helper.h>
  30 #include "i915_drv.h"
  31 #include "intel_drv.h"
  32 #include "../../../platform/x86/intel_ips.h"
  33 #include <linux/module.h>
  34 #include <drm/drm_atomic_helper.h>
  35
  36 /**
  37  * DOC: RC6
  38  *
  39  * RC6 is a special power stage which allows the GPU to enter an very
  40  * low-voltage mode when idle, using down to 0V while at this stage.  This
  41  * stage is entered automatically when the GPU is idle when RC6 support is
  42  * enabled, and as soon as new workload arises GPU wakes up automatically as well.
  43  *
  44  * There are different RC6 modes available in Intel GPU, which differentiate
  45  * among each other with the latency required to enter and leave RC6 and
  46  * voltage consumed by the GPU in different states.
  47  *
  48  * The combination of the following flags define which states GPU is allowed
  49  * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
  50  * RC6pp is deepest RC6. Their support by hardware varies according to the
  51  * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
  52  * which brings the most power savings; deeper states save more power, but
  53  * require higher latency to switch to and wake up.
  54  */
  55 #define INTEL_RC6_ENABLE                        (1<<0)
  56 #define INTEL_RC6p_ENABLE                       (1<<1)
  57 #define INTEL_RC6pp_ENABLE                      (1<<2)
  58
  59 static void gen9_init_clock_gating(struct drm_i915_private *dev_priv)
  60 {
  61         /* See Bspec note for PSR2_CTL bit 31, Wa#828:skl,bxt,kbl */
  62         I915_WRITE(CHICKEN_PAR1_1,
  63                    I915_READ(CHICKEN_PAR1_1) | SKL_EDP_PSR_FIX_RDWRAP);
  64
  65         I915_WRITE(GEN8_CONFIG0,
  66                    I915_READ(GEN8_CONFIG0) | GEN9_DEFAULT_FIXES);
  67
  68         /* WaEnableChickenDCPR:skl,bxt,kbl,glk */
  69         I915_WRITE(GEN8_CHICKEN_DCPR_1,
  70                    I915_READ(GEN8_CHICKEN_DCPR_1) | MASK_WAKEMEM);
  71
  72         /* WaFbcTurnOffFbcWatermark:skl,bxt,kbl */
  73         /* WaFbcWakeMemOn:skl,bxt,kbl,glk */
  74         I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
  75                    DISP_FBC_WM_DIS |
  76                    DISP_FBC_MEMORY_WAKE);
  77
  78         /* WaFbcHighMemBwCorruptionAvoidance:skl,bxt,kbl */
  79         I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
  80                    ILK_DPFC_DISABLE_DUMMY0);
  81 }
  82
  83 static void bxt_init_clock_gating(struct drm_i915_private *dev_priv)
  84 {
  85         gen9_init_clock_gating(dev_priv);
  86
  87         /* WaDisableSDEUnitClockGating:bxt */
  88         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
  89                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
  90
  91         /*
  92          * FIXME:
  93          * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
  94          */
  95         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
  96                    GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ);
  97
  98         /*
  99          * Wa: Backlight PWM may stop in the asserted state, causing backlight
 100          * to stay fully on.
 101          */
 102         I915_WRITE(GEN9_CLKGATE_DIS_0, I915_READ(GEN9_CLKGATE_DIS_0) |
 103                    PWM1_GATING_DIS | PWM2_GATING_DIS);
 104 }
 105
 106 static void glk_init_clock_gating(struct drm_i915_private *dev_priv)
 107 {
 108         gen9_init_clock_gating(dev_priv);
 109
 110         /*
 111          * WaDisablePWMClockGating:glk
 112          * Backlight PWM may stop in the asserted state, causing backlight
 113          * to stay fully on.
 114          */
 115         I915_WRITE(GEN9_CLKGATE_DIS_0, I915_READ(GEN9_CLKGATE_DIS_0) |
 116                    PWM1_GATING_DIS | PWM2_GATING_DIS);
 117
 118         /* WaDDIIOTimeout:glk */
 119         if (IS_GLK_REVID(dev_priv, 0, GLK_REVID_A1)) {
 120                 u32 val = I915_READ(CHICKEN_MISC_2);
 121                 val &= ~(GLK_CL0_PWR_DOWN |
 122                          GLK_CL1_PWR_DOWN |
 123                          GLK_CL2_PWR_DOWN);
 124                 I915_WRITE(CHICKEN_MISC_2, val);
 125         }
 126
 127 }
 128
 129 static void i915_pineview_get_mem_freq(struct drm_i915_private *dev_priv)
 130 {
 131         u32 tmp;
 132
 133         tmp = I915_READ(CLKCFG);
 134
 135         switch (tmp & CLKCFG_FSB_MASK) {
 136         case CLKCFG_FSB_533:
 137                 dev_priv->fsb_freq = 533; /* 133*4 */
 138                 break;
 139         case CLKCFG_FSB_800:
 140                 dev_priv->fsb_freq = 800; /* 200*4 */
 141                 break;
 142         case CLKCFG_FSB_667:
 143                 dev_priv->fsb_freq =  667; /* 167*4 */
 144                 break;
 145         case CLKCFG_FSB_400:
 146                 dev_priv->fsb_freq = 400; /* 100*4 */
 147                 break;
 148         }
 149
 150         switch (tmp & CLKCFG_MEM_MASK) {
 151         case CLKCFG_MEM_533:
 152                 dev_priv->mem_freq = 533;
 153                 break;
 154         case CLKCFG_MEM_667:
 155                 dev_priv->mem_freq = 667;
 156                 break;
 157         case CLKCFG_MEM_800:
 158                 dev_priv->mem_freq = 800;
 159                 break;
 160         }
 161
 162         /* detect pineview DDR3 setting */
 163         tmp = I915_READ(CSHRDDR3CTL);
 164         dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
 165 }
 166
 167 static void i915_ironlake_get_mem_freq(struct drm_i915_private *dev_priv)
 168 {
 169         u16 ddrpll, csipll;
 170
 171         ddrpll = I915_READ16(DDRMPLL1);
 172         csipll = I915_READ16(CSIPLL0);
 173
 174         switch (ddrpll & 0xff) {
 175         case 0xc:
 176                 dev_priv->mem_freq = 800;
 177                 break;
 178         case 0x10:
 179                 dev_priv->mem_freq = 1066;
 180                 break;
 181         case 0x14:
 182                 dev_priv->mem_freq = 1333;
 183                 break;
 184         case 0x18:
 185                 dev_priv->mem_freq = 1600;
 186                 break;
 187         default:
 188                 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
 189                                  ddrpll & 0xff);
 190                 dev_priv->mem_freq = 0;
 191                 break;
 192         }
 193
 194         dev_priv->ips.r_t = dev_priv->mem_freq;
 195
 196         switch (csipll & 0x3ff) {
 197         case 0x00c:
 198                 dev_priv->fsb_freq = 3200;
 199                 break;
 200         case 0x00e:
 201                 dev_priv->fsb_freq = 3733;
 202                 break;
 203         case 0x010:
 204                 dev_priv->fsb_freq = 4266;
 205                 break;
 206         case 0x012:
 207                 dev_priv->fsb_freq = 4800;
 208                 break;
 209         case 0x014:
 210                 dev_priv->fsb_freq = 5333;
 211                 break;
 212         case 0x016:
 213                 dev_priv->fsb_freq = 5866;
 214                 break;
 215         case 0x018:
 216                 dev_priv->fsb_freq = 6400;
 217                 break;
 218         default:
 219                 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
 220                                  csipll & 0x3ff);
 221                 dev_priv->fsb_freq = 0;
 222                 break;
 223         }
 224
 225         if (dev_priv->fsb_freq == 3200) {
 226                 dev_priv->ips.c_m = 0;
 227         } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
 228                 dev_priv->ips.c_m = 1;
 229         } else {
 230                 dev_priv->ips.c_m = 2;
 231         }
 232 }
 233
 234 static const struct cxsr_latency cxsr_latency_table[] = {
 235         {1, 0, 800, 400, 3382, 33382, 3983, 33983},    /* DDR2-400 SC */
 236         {1, 0, 800, 667, 3354, 33354, 3807, 33807},    /* DDR2-667 SC */
 237         {1, 0, 800, 800, 3347, 33347, 3763, 33763},    /* DDR2-800 SC */
 238         {1, 1, 800, 667, 6420, 36420, 6873, 36873},    /* DDR3-667 SC */
 239         {1, 1, 800, 800, 5902, 35902, 6318, 36318},    /* DDR3-800 SC */
 240
 241         {1, 0, 667, 400, 3400, 33400, 4021, 34021},    /* DDR2-400 SC */
 242         {1, 0, 667, 667, 3372, 33372, 3845, 33845},    /* DDR2-667 SC */
 243         {1, 0, 667, 800, 3386, 33386, 3822, 33822},    /* DDR2-800 SC */
 244         {1, 1, 667, 667, 6438, 36438, 6911, 36911},    /* DDR3-667 SC */
 245         {1, 1, 667, 800, 5941, 35941, 6377, 36377},    /* DDR3-800 SC */
 246
 247         {1, 0, 400, 400, 3472, 33472, 4173, 34173},    /* DDR2-400 SC */
 248         {1, 0, 400, 667, 3443, 33443, 3996, 33996},    /* DDR2-667 SC */
 249         {1, 0, 400, 800, 3430, 33430, 3946, 33946},    /* DDR2-800 SC */
 250         {1, 1, 400, 667, 6509, 36509, 7062, 37062},    /* DDR3-667 SC */
 251         {1, 1, 400, 800, 5985, 35985, 6501, 36501},    /* DDR3-800 SC */
 252
 253         {0, 0, 800, 400, 3438, 33438, 4065, 34065},    /* DDR2-400 SC */
 254         {0, 0, 800, 667, 3410, 33410, 3889, 33889},    /* DDR2-667 SC */
 255         {0, 0, 800, 800, 3403, 33403, 3845, 33845},    /* DDR2-800 SC */
 256         {0, 1, 800, 667, 6476, 36476, 6955, 36955},    /* DDR3-667 SC */
 257         {0, 1, 800, 800, 5958, 35958, 6400, 36400},    /* DDR3-800 SC */
 258
 259         {0, 0, 667, 400, 3456, 33456, 4103, 34106},    /* DDR2-400 SC */
 260         {0, 0, 667, 667, 3428, 33428, 3927, 33927},    /* DDR2-667 SC */
 261         {0, 0, 667, 800, 3443, 33443, 3905, 33905},    /* DDR2-800 SC */
 262         {0, 1, 667, 667, 6494, 36494, 6993, 36993},    /* DDR3-667 SC */
 263         {0, 1, 667, 800, 5998, 35998, 6460, 36460},    /* DDR3-800 SC */
 264
 265         {0, 0, 400, 400, 3528, 33528, 4255, 34255},    /* DDR2-400 SC */
 266         {0, 0, 400, 667, 3500, 33500, 4079, 34079},    /* DDR2-667 SC */
 267         {0, 0, 400, 800, 3487, 33487, 4029, 34029},    /* DDR2-800 SC */
 268         {0, 1, 400, 667, 6566, 36566, 7145, 37145},    /* DDR3-667 SC */
 269         {0, 1, 400, 800, 6042, 36042, 6584, 36584},    /* DDR3-800 SC */
 270 };
 271
 272 static const struct cxsr_latency *intel_get_cxsr_latency(bool is_desktop,
 273                                                          bool is_ddr3,
 274                                                          int fsb,
 275                                                          int mem)
 276 {
 277         const struct cxsr_latency *latency;
 278         int i;
 279
 280         if (fsb == 0 || mem == 0)
 281                 return NULL;
 282
 283         for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
 284                 latency = &cxsr_latency_table[i];
 285                 if (is_desktop == latency->is_desktop &&
 286                     is_ddr3 == latency->is_ddr3 &&
 287                     fsb == latency->fsb_freq && mem == latency->mem_freq)
 288                         return latency;
 289         }
 290
 291         DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 292
 293         return NULL;
 294 }
 295
 296 static void chv_set_memory_dvfs(struct drm_i915_private *dev_priv, bool enable)
 297 {
 298         u32 val;
 299
 300         mutex_lock(&dev_priv->rps.hw_lock);
 301
 302         val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
 303         if (enable)
 304                 val &= ~FORCE_DDR_HIGH_FREQ;
 305         else
 306                 val |= FORCE_DDR_HIGH_FREQ;
 307         val &= ~FORCE_DDR_LOW_FREQ;
 308         val |= FORCE_DDR_FREQ_REQ_ACK;
 309         vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
 310
 311         if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
 312                       FORCE_DDR_FREQ_REQ_ACK) == 0, 3))
 313                 DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
 314
 315         mutex_unlock(&dev_priv->rps.hw_lock);
 316 }
 317
 318 static void chv_set_memory_pm5(struct drm_i915_private *dev_priv, bool enable)
 319 {
 320         u32 val;
 321
 322         mutex_lock(&dev_priv->rps.hw_lock);
 323
 324         val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
 325         if (enable)
 326                 val |= DSP_MAXFIFO_PM5_ENABLE;
 327         else
 328                 val &= ~DSP_MAXFIFO_PM5_ENABLE;
 329         vlv_punit_write(dev_priv, PUNIT_REG_DSPFREQ, val);
 330
 331         mutex_unlock(&dev_priv->rps.hw_lock);
 332 }
 333
 334 #define FW_WM(value, plane) \
 335         (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
 336
 337 static bool _intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
 338 {
 339         bool was_enabled;
 340         u32 val;
 341
 342         if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
 343                 was_enabled = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
 344                 I915_WRITE(FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
 345                 POSTING_READ(FW_BLC_SELF_VLV);
 346         } else if (IS_G4X(dev_priv) || IS_I965GM(dev_priv)) {
 347                 was_enabled = I915_READ(FW_BLC_SELF) & FW_BLC_SELF_EN;
 348                 I915_WRITE(FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
 349                 POSTING_READ(FW_BLC_SELF);
 350         } else if (IS_PINEVIEW(dev_priv)) {
 351                 val = I915_READ(DSPFW3);
 352                 was_enabled = val & PINEVIEW_SELF_REFRESH_EN;
 353                 if (enable)
 354                         val |= PINEVIEW_SELF_REFRESH_EN;
 355                 else
 356                         val &= ~PINEVIEW_SELF_REFRESH_EN;
 357                 I915_WRITE(DSPFW3, val);
 358                 POSTING_READ(DSPFW3);
 359         } else if (IS_I945G(dev_priv) || IS_I945GM(dev_priv)) {
 360                 was_enabled = I915_READ(FW_BLC_SELF) & FW_BLC_SELF_EN;
 361                 val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
 362                                _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
 363                 I915_WRITE(FW_BLC_SELF, val);
 364                 POSTING_READ(FW_BLC_SELF);
 365         } else if (IS_I915GM(dev_priv)) {
 366                 /*
 367                  * FIXME can't find a bit like this for 915G, and
 368                  * and yet it does have the related watermark in
 369                  * FW_BLC_SELF. What's going on?
 370                  */
 371                 was_enabled = I915_READ(INSTPM) & INSTPM_SELF_EN;
 372                 val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
 373                                _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
 374                 I915_WRITE(INSTPM, val);
 375                 POSTING_READ(INSTPM);
 376         } else {
 377                 return false;
 378         }
 379
 380         trace_intel_memory_cxsr(dev_priv, was_enabled, enable);
 381
 382         DRM_DEBUG_KMS("memory self-refresh is %s (was %s)\n",
 383                       enableddisabled(enable),
 384                       enableddisabled(was_enabled));
 385
 386         return was_enabled;
 387 }
 388
 389 bool intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
 390 {
 391         bool ret;
 392
 393         mutex_lock(&dev_priv->wm.wm_mutex);
 394         ret = _intel_set_memory_cxsr(dev_priv, enable);
 395         dev_priv->wm.vlv.cxsr = enable;
 396         mutex_unlock(&dev_priv->wm.wm_mutex);
 397
 398         return ret;
 399 }
 400
 401 /*
 402  * Latency for FIFO fetches is dependent on several factors:
 403  *   - memory configuration (speed, channels)
 404  *   - chipset
 405  *   - current MCH state
 406  * It can be fairly high in some situations, so here we assume a fairly
 407  * pessimal value.  It's a tradeoff between extra memory fetches (if we
 408  * set this value too high, the FIFO will fetch frequently to stay full)
 409  * and power consumption (set it too low to save power and we might see
 410  * FIFO underruns and display "flicker").
 411  *
 412  * A value of 5us seems to be a good balance; safe for very low end
 413  * platforms but not overly aggressive on lower latency configs.
 414  */
 415 static const int pessimal_latency_ns = 5000;
 416
 417 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
 418         ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
 419
 420 static void vlv_get_fifo_size(struct intel_crtc_state *crtc_state)
 421 {
 422         struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
 423         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
 424         struct vlv_fifo_state *fifo_state = &crtc_state->wm.vlv.fifo_state;
 425         enum pipe pipe = crtc->pipe;
 426         int sprite0_start, sprite1_start;
 427
 428         switch (pipe) {
 429                 uint32_t dsparb, dsparb2, dsparb3;
 430         case PIPE_A:
 431                 dsparb = I915_READ(DSPARB);
 432                 dsparb2 = I915_READ(DSPARB2);
 433                 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 0, 0);
 434                 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 8, 4);
 435                 break;
 436         case PIPE_B:
 437                 dsparb = I915_READ(DSPARB);
 438                 dsparb2 = I915_READ(DSPARB2);
 439                 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 16, 8);
 440                 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 24, 12);
 441                 break;
 442         case PIPE_C:
 443                 dsparb2 = I915_READ(DSPARB2);
 444                 dsparb3 = I915_READ(DSPARB3);
 445                 sprite0_start = VLV_FIFO_START(dsparb3, dsparb2, 0, 16);
 446                 sprite1_start = VLV_FIFO_START(dsparb3, dsparb2, 8, 20);
 447                 break;
 448         default:
 449                 MISSING_CASE(pipe);
 450                 return;
 451         }
 452
 453         fifo_state->plane[PLANE_PRIMARY] = sprite0_start;
 454         fifo_state->plane[PLANE_SPRITE0] = sprite1_start - sprite0_start;
 455         fifo_state->plane[PLANE_SPRITE1] = 511 - sprite1_start;
 456         fifo_state->plane[PLANE_CURSOR] = 63;
 457
 458         DRM_DEBUG_KMS("Pipe %c FIFO size: %d/%d/%d/%d\n",
 459                       pipe_name(pipe),
 460                       fifo_state->plane[PLANE_PRIMARY],
 461                       fifo_state->plane[PLANE_SPRITE0],
 462                       fifo_state->plane[PLANE_SPRITE1],
 463                       fifo_state->plane[PLANE_CURSOR]);
 464 }
 465
 466 static int i9xx_get_fifo_size(struct drm_i915_private *dev_priv, int plane)
 467 {
 468         uint32_t dsparb = I915_READ(DSPARB);
 469         int size;
 470
 471         size = dsparb & 0x7f;
 472         if (plane)
 473                 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
 474
 475         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 476                       plane ? "B" : "A", size);
 477
 478         return size;
 479 }
 480
 481 static int i830_get_fifo_size(struct drm_i915_private *dev_priv, int plane)
 482 {
 483         uint32_t dsparb = I915_READ(DSPARB);
 484         int size;
 485
 486         size = dsparb & 0x1ff;
 487         if (plane)
 488                 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
 489         size >>= 1; /* Convert to cachelines */
 490
 491         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 492                       plane ? "B" : "A", size);
 493
 494         return size;
 495 }
 496
 497 static int i845_get_fifo_size(struct drm_i915_private *dev_priv, int plane)
 498 {
 499         uint32_t dsparb = I915_READ(DSPARB);
 500         int size;
 501
 502         size = dsparb & 0x7f;
 503         size >>= 2; /* Convert to cachelines */
 504
 505         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 506                       plane ? "B" : "A",
 507                       size);
 508
 509         return size;
 510 }
 511
 512 /* Pineview has different values for various configs */
 513 static const struct intel_watermark_params pineview_display_wm = {
 514         .fifo_size = PINEVIEW_DISPLAY_FIFO,
 515         .max_wm = PINEVIEW_MAX_WM,
 516         .default_wm = PINEVIEW_DFT_WM,
 517         .guard_size = PINEVIEW_GUARD_WM,
 518         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 519 };
 520 static const struct intel_watermark_params pineview_display_hplloff_wm = {
 521         .fifo_size = PINEVIEW_DISPLAY_FIFO,
 522         .max_wm = PINEVIEW_MAX_WM,
 523         .default_wm = PINEVIEW_DFT_HPLLOFF_WM,
 524         .guard_size = PINEVIEW_GUARD_WM,
 525         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 526 };
 527 static const struct intel_watermark_params pineview_cursor_wm = {
 528         .fifo_size = PINEVIEW_CURSOR_FIFO,
 529         .max_wm = PINEVIEW_CURSOR_MAX_WM,
 530         .default_wm = PINEVIEW_CURSOR_DFT_WM,
 531         .guard_size = PINEVIEW_CURSOR_GUARD_WM,
 532         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 533 };
 534 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
 535         .fifo_size = PINEVIEW_CURSOR_FIFO,
 536         .max_wm = PINEVIEW_CURSOR_MAX_WM,
 537         .default_wm = PINEVIEW_CURSOR_DFT_WM,
 538         .guard_size = PINEVIEW_CURSOR_GUARD_WM,
 539         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 540 };
 541 static const struct intel_watermark_params g4x_wm_info = {
 542         .fifo_size = G4X_FIFO_SIZE,
 543         .max_wm = G4X_MAX_WM,
 544         .default_wm = G4X_MAX_WM,
 545         .guard_size = 2,
 546         .cacheline_size = G4X_FIFO_LINE_SIZE,
 547 };
 548 static const struct intel_watermark_params g4x_cursor_wm_info = {
 549         .fifo_size = I965_CURSOR_FIFO,
 550         .max_wm = I965_CURSOR_MAX_WM,
 551         .default_wm = I965_CURSOR_DFT_WM,
 552         .guard_size = 2,
 553         .cacheline_size = G4X_FIFO_LINE_SIZE,
 554 };
 555 static const struct intel_watermark_params i965_cursor_wm_info = {
 556         .fifo_size = I965_CURSOR_FIFO,
 557         .max_wm = I965_CURSOR_MAX_WM,
 558         .default_wm = I965_CURSOR_DFT_WM,
 559         .guard_size = 2,
 560         .cacheline_size = I915_FIFO_LINE_SIZE,
 561 };
 562 static const struct intel_watermark_params i945_wm_info = {
 563         .fifo_size = I945_FIFO_SIZE,
 564         .max_wm = I915_MAX_WM,
 565         .default_wm = 1,
 566         .guard_size = 2,
 567         .cacheline_size = I915_FIFO_LINE_SIZE,
 568 };
 569 static const struct intel_watermark_params i915_wm_info = {
 570         .fifo_size = I915_FIFO_SIZE,
 571         .max_wm = I915_MAX_WM,
 572         .default_wm = 1,
 573         .guard_size = 2,
 574         .cacheline_size = I915_FIFO_LINE_SIZE,
 575 };
 576 static const struct intel_watermark_params i830_a_wm_info = {
 577         .fifo_size = I855GM_FIFO_SIZE,
 578         .max_wm = I915_MAX_WM,
 579         .default_wm = 1,
 580         .guard_size = 2,
 581         .cacheline_size = I830_FIFO_LINE_SIZE,
 582 };
 583 static const struct intel_watermark_params i830_bc_wm_info = {
 584         .fifo_size = I855GM_FIFO_SIZE,
 585         .max_wm = I915_MAX_WM/2,
 586         .default_wm = 1,
 587         .guard_size = 2,
 588         .cacheline_size = I830_FIFO_LINE_SIZE,
 589 };
 590 static const struct intel_watermark_params i845_wm_info = {
 591         .fifo_size = I830_FIFO_SIZE,
 592         .max_wm = I915_MAX_WM,
 593         .default_wm = 1,
 594         .guard_size = 2,
 595         .cacheline_size = I830_FIFO_LINE_SIZE,
 596 };
 597
 598 /**
 599  * intel_calculate_wm - calculate watermark level
 600  * @clock_in_khz: pixel clock
 601  * @wm: chip FIFO params
 602  * @cpp: bytes per pixel
 603  * @latency_ns: memory latency for the platform
 604  *
 605  * Calculate the watermark level (the level at which the display plane will
 606  * start fetching from memory again).  Each chip has a different display
 607  * FIFO size and allocation, so the caller needs to figure that out and pass
 608  * in the correct intel_watermark_params structure.
 609  *
 610  * As the pixel clock runs, the FIFO will be drained at a rate that depends
 611  * on the pixel size.  When it reaches the watermark level, it'll start
 612  * fetching FIFO line sized based chunks from memory until the FIFO fills
 613  * past the watermark point.  If the FIFO drains completely, a FIFO underrun
 614  * will occur, and a display engine hang could result.
 615  */
 616 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
 617                                         const struct intel_watermark_params *wm,
 618                                         int fifo_size, int cpp,
 619                                         unsigned long latency_ns)
 620 {
 621         long entries_required, wm_size;
 622
 623         /*
 624          * Note: we need to make sure we don't overflow for various clock &
 625          * latency values.
 626          * clocks go from a few thousand to several hundred thousand.
 627          * latency is usually a few thousand
 628          */
 629         entries_required = ((clock_in_khz / 1000) * cpp * latency_ns) /
 630                 1000;
 631         entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
 632
 633         DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
 634
 635         wm_size = fifo_size - (entries_required + wm->guard_size);
 636
 637         DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
 638
 639         /* Don't promote wm_size to unsigned... */
 640         if (wm_size > (long)wm->max_wm)
 641                 wm_size = wm->max_wm;
 642         if (wm_size <= 0)
 643                 wm_size = wm->default_wm;
 644
 645         /*
 646          * Bspec seems to indicate that the value shouldn't be lower than
 647          * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
 648          * Lets go for 8 which is the burst size since certain platforms
 649          * already use a hardcoded 8 (which is what the spec says should be
 650          * done).
 651          */
 652         if (wm_size <= 8)
 653                 wm_size = 8;
 654
 655         return wm_size;
 656 }
 657
 658 static bool intel_wm_plane_visible(const struct intel_crtc_state *crtc_state,
 659                                    const struct intel_plane_state *plane_state)
 660 {
 661         struct intel_plane *plane = to_intel_plane(plane_state->base.plane);
 662
 663         /* FIXME check the 'enable' instead */
 664         if (!crtc_state->base.active)
 665                 return false;
 666
 667         /*
 668          * Treat cursor with fb as always visible since cursor updates
 669          * can happen faster than the vrefresh rate, and the current
 670          * watermark code doesn't handle that correctly. Cursor updates
 671          * which set/clear the fb or change the cursor size are going
 672          * to get throttled by intel_legacy_cursor_update() to work
 673          * around this problem with the watermark code.
 674          */
 675         if (plane->id == PLANE_CURSOR)
 676                 return plane_state->base.fb != NULL;
 677         else
 678                 return plane_state->base.visible;
 679 }
 680
 681 static struct intel_crtc *single_enabled_crtc(struct drm_i915_private *dev_priv)
 682 {
 683         struct intel_crtc *crtc, *enabled = NULL;
 684
 685         for_each_intel_crtc(&dev_priv->drm, crtc) {
 686                 if (intel_crtc_active(crtc)) {
 687                         if (enabled)
 688                                 return NULL;
 689                         enabled = crtc;
 690                 }
 691         }
 692
 693         return enabled;
 694 }
 695
 696 static void pineview_update_wm(struct intel_crtc *unused_crtc)
 697 {
 698         struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
 699         struct intel_crtc *crtc;
 700         const struct cxsr_latency *latency;
 701         u32 reg;
 702         unsigned long wm;
 703
 704         latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv),
 705                                          dev_priv->is_ddr3,
 706                                          dev_priv->fsb_freq,
 707                                          dev_priv->mem_freq);
 708         if (!latency) {
 709                 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 710                 intel_set_memory_cxsr(dev_priv, false);
 711                 return;
 712         }
 713
 714         crtc = single_enabled_crtc(dev_priv);
 715         if (crtc) {
 716                 const struct drm_display_mode *adjusted_mode =
 717                         &crtc->config->base.adjusted_mode;
 718                 const struct drm_framebuffer *fb =
 719                         crtc->base.primary->state->fb;
 720                 int cpp = fb->format->cpp[0];
 721                 int clock = adjusted_mode->crtc_clock;
 722
 723                 /* Display SR */
 724                 wm = intel_calculate_wm(clock, &pineview_display_wm,
 725                                         pineview_display_wm.fifo_size,
 726                                         cpp, latency->display_sr);
 727                 reg = I915_READ(DSPFW1);
 728                 reg &= ~DSPFW_SR_MASK;
 729                 reg |= FW_WM(wm, SR);
 730                 I915_WRITE(DSPFW1, reg);
 731                 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
 732
 733                 /* cursor SR */
 734                 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
 735                                         pineview_display_wm.fifo_size,
 736                                         cpp, latency->cursor_sr);
 737                 reg = I915_READ(DSPFW3);
 738                 reg &= ~DSPFW_CURSOR_SR_MASK;
 739                 reg |= FW_WM(wm, CURSOR_SR);
 740                 I915_WRITE(DSPFW3, reg);
 741
 742                 /* Display HPLL off SR */
 743                 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
 744                                         pineview_display_hplloff_wm.fifo_size,
 745                                         cpp, latency->display_hpll_disable);
 746                 reg = I915_READ(DSPFW3);
 747                 reg &= ~DSPFW_HPLL_SR_MASK;
 748                 reg |= FW_WM(wm, HPLL_SR);
 749                 I915_WRITE(DSPFW3, reg);
 750
 751                 /* cursor HPLL off SR */
 752                 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
 753                                         pineview_display_hplloff_wm.fifo_size,
 754                                         cpp, latency->cursor_hpll_disable);
 755                 reg = I915_READ(DSPFW3);
 756                 reg &= ~DSPFW_HPLL_CURSOR_MASK;
 757                 reg |= FW_WM(wm, HPLL_CURSOR);
 758                 I915_WRITE(DSPFW3, reg);
 759                 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
 760
 761                 intel_set_memory_cxsr(dev_priv, true);
 762         } else {
 763                 intel_set_memory_cxsr(dev_priv, false);
 764         }
 765 }
 766
 767 static bool g4x_compute_wm0(struct drm_i915_private *dev_priv,
 768                             int plane,
 769                             const struct intel_watermark_params *display,
 770                             int display_latency_ns,
 771                             const struct intel_watermark_params *cursor,
 772                             int cursor_latency_ns,
 773                             int *plane_wm,
 774                             int *cursor_wm)
 775 {
 776         struct intel_crtc *crtc;
 777         const struct drm_display_mode *adjusted_mode;
 778         const struct drm_framebuffer *fb;
 779         int htotal, hdisplay, clock, cpp;
 780         int line_time_us, line_count;
 781         int entries, tlb_miss;
 782
 783         crtc = intel_get_crtc_for_plane(dev_priv, plane);
 784         if (!intel_crtc_active(crtc)) {
 785                 *cursor_wm = cursor->guard_size;
 786                 *plane_wm = display->guard_size;
 787                 return false;
 788         }
 789
 790         adjusted_mode = &crtc->config->base.adjusted_mode;
 791         fb = crtc->base.primary->state->fb;
 792         clock = adjusted_mode->crtc_clock;
 793         htotal = adjusted_mode->crtc_htotal;
 794         hdisplay = crtc->config->pipe_src_w;
 795         cpp = fb->format->cpp[0];
 796
 797         /* Use the small buffer method to calculate plane watermark */
 798         entries = ((clock * cpp / 1000) * display_latency_ns) / 1000;
 799         tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
 800         if (tlb_miss > 0)
 801                 entries += tlb_miss;
 802         entries = DIV_ROUND_UP(entries, display->cacheline_size);
 803         *plane_wm = entries + display->guard_size;
 804         if (*plane_wm > (int)display->max_wm)
 805                 *plane_wm = display->max_wm;
 806
 807         /* Use the large buffer method to calculate cursor watermark */
 808         line_time_us = max(htotal * 1000 / clock, 1);
 809         line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
 810         entries = line_count * crtc->base.cursor->state->crtc_w * cpp;
 811         tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
 812         if (tlb_miss > 0)
 813                 entries += tlb_miss;
 814         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
 815         *cursor_wm = entries + cursor->guard_size;
 816         if (*cursor_wm > (int)cursor->max_wm)
 817                 *cursor_wm = (int)cursor->max_wm;
 818
 819         return true;
 820 }
 821
 822 /*
 823  * Check the wm result.
 824  *
 825  * If any calculated watermark values is larger than the maximum value that
 826  * can be programmed into the associated watermark register, that watermark
 827  * must be disabled.
 828  */
 829 static bool g4x_check_srwm(struct drm_i915_private *dev_priv,
 830                            int display_wm, int cursor_wm,
 831                            const struct intel_watermark_params *display,
 832                            const struct intel_watermark_params *cursor)
 833 {
 834         DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
 835                       display_wm, cursor_wm);
 836
 837         if (display_wm > display->max_wm) {
 838                 DRM_DEBUG_KMS("display watermark is too large(%d/%u), disabling\n",
 839                               display_wm, display->max_wm);
 840                 return false;
 841         }
 842
 843         if (cursor_wm > cursor->max_wm) {
 844                 DRM_DEBUG_KMS("cursor watermark is too large(%d/%u), disabling\n",
 845                               cursor_wm, cursor->max_wm);
 846                 return false;
 847         }
 848
 849         if (!(display_wm || cursor_wm)) {
 850                 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
 851                 return false;
 852         }
 853
 854         return true;
 855 }
 856
 857 static bool g4x_compute_srwm(struct drm_i915_private *dev_priv,
 858                              int plane,
 859                              int latency_ns,
 860                              const struct intel_watermark_params *display,
 861                              const struct intel_watermark_params *cursor,
 862                              int *display_wm, int *cursor_wm)
 863 {
 864         struct intel_crtc *crtc;
 865         const struct drm_display_mode *adjusted_mode;
 866         const struct drm_framebuffer *fb;
 867         int hdisplay, htotal, cpp, clock;
 868         unsigned long line_time_us;
 869         int line_count, line_size;
 870         int small, large;
 871         int entries;
 872
 873         if (!latency_ns) {
 874                 *display_wm = *cursor_wm = 0;
 875                 return false;
 876         }
 877
 878         crtc = intel_get_crtc_for_plane(dev_priv, plane);
 879         adjusted_mode = &crtc->config->base.adjusted_mode;
 880         fb = crtc->base.primary->state->fb;
 881         clock = adjusted_mode->crtc_clock;
 882         htotal = adjusted_mode->crtc_htotal;
 883         hdisplay = crtc->config->pipe_src_w;
 884         cpp = fb->format->cpp[0];
 885
 886         line_time_us = max(htotal * 1000 / clock, 1);
 887         line_count = (latency_ns / line_time_us + 1000) / 1000;
 888         line_size = hdisplay * cpp;
 889
 890         /* Use the minimum of the small and large buffer method for primary */
 891         small = ((clock * cpp / 1000) * latency_ns) / 1000;
 892         large = line_count * line_size;
 893
 894         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
 895         *display_wm = entries + display->guard_size;
 896
 897         /* calculate the self-refresh watermark for display cursor */
 898         entries = line_count * cpp * crtc->base.cursor->state->crtc_w;
 899         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
 900         *cursor_wm = entries + cursor->guard_size;
 901
 902         return g4x_check_srwm(dev_priv,
 903                               *display_wm, *cursor_wm,
 904                               display, cursor);
 905 }
 906
 907 #define FW_WM_VLV(value, plane) \
 908         (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
 909
 910 static void vlv_write_wm_values(struct drm_i915_private *dev_priv,
 911                                 const struct vlv_wm_values *wm)
 912 {
 913         enum pipe pipe;
 914
 915         for_each_pipe(dev_priv, pipe) {
 916                 trace_vlv_wm(intel_get_crtc_for_pipe(dev_priv, pipe), wm);
 917
 918                 I915_WRITE(VLV_DDL(pipe),
 919                            (wm->ddl[pipe].plane[PLANE_CURSOR] << DDL_CURSOR_SHIFT) |
 920                            (wm->ddl[pipe].plane[PLANE_SPRITE1] << DDL_SPRITE_SHIFT(1)) |
 921                            (wm->ddl[pipe].plane[PLANE_SPRITE0] << DDL_SPRITE_SHIFT(0)) |
 922                            (wm->ddl[pipe].plane[PLANE_PRIMARY] << DDL_PLANE_SHIFT));
 923         }
 924
 925         /*
 926          * Zero the (unused) WM1 watermarks, and also clear all the
 927          * high order bits so that there are no out of bounds values
 928          * present in the registers during the reprogramming.
 929          */
 930         I915_WRITE(DSPHOWM, 0);
 931         I915_WRITE(DSPHOWM1, 0);
 932         I915_WRITE(DSPFW4, 0);
 933         I915_WRITE(DSPFW5, 0);
 934         I915_WRITE(DSPFW6, 0);
 935
 936         I915_WRITE(DSPFW1,
 937                    FW_WM(wm->sr.plane, SR) |
 938                    FW_WM(wm->pipe[PIPE_B].plane[PLANE_CURSOR], CURSORB) |
 939                    FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_PRIMARY], PLANEB) |
 940                    FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_PRIMARY], PLANEA));
 941         I915_WRITE(DSPFW2,
 942                    FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_SPRITE1], SPRITEB) |
 943                    FW_WM(wm->pipe[PIPE_A].plane[PLANE_CURSOR], CURSORA) |
 944                    FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_SPRITE0], SPRITEA));
 945         I915_WRITE(DSPFW3,
 946                    FW_WM(wm->sr.cursor, CURSOR_SR));
 947
 948         if (IS_CHERRYVIEW(dev_priv)) {
 949                 I915_WRITE(DSPFW7_CHV,
 950                            FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE1], SPRITED) |
 951                            FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEC));
 952                 I915_WRITE(DSPFW8_CHV,
 953                            FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_SPRITE1], SPRITEF) |
 954                            FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_SPRITE0], SPRITEE));
 955                 I915_WRITE(DSPFW9_CHV,
 956                            FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_PRIMARY], PLANEC) |
 957                            FW_WM(wm->pipe[PIPE_C].plane[PLANE_CURSOR], CURSORC));
 958                 I915_WRITE(DSPHOWM,
 959                            FW_WM(wm->sr.plane >> 9, SR_HI) |
 960                            FW_WM(wm->pipe[PIPE_C].plane[PLANE_SPRITE1] >> 8, SPRITEF_HI) |
 961                            FW_WM(wm->pipe[PIPE_C].plane[PLANE_SPRITE0] >> 8, SPRITEE_HI) |
 962                            FW_WM(wm->pipe[PIPE_C].plane[PLANE_PRIMARY] >> 8, PLANEC_HI) |
 963                            FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE1] >> 8, SPRITED_HI) |
 964                            FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0] >> 8, SPRITEC_HI) |
 965                            FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY] >> 8, PLANEB_HI) |
 966                            FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE1] >> 8, SPRITEB_HI) |
 967                            FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0] >> 8, SPRITEA_HI) |
 968                            FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY] >> 8, PLANEA_HI));
 969         } else {
 970                 I915_WRITE(DSPFW7,
 971                            FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE1], SPRITED) |
 972                            FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEC));
 973                 I915_WRITE(DSPHOWM,
 974                            FW_WM(wm->sr.plane >> 9, SR_HI) |
 975                            FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE1] >> 8, SPRITED_HI) |
 976                            FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0] >> 8, SPRITEC_HI) |
 977                            FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY] >> 8, PLANEB_HI) |
 978                            FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE1] >> 8, SPRITEB_HI) |
 979                            FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0] >> 8, SPRITEA_HI) |
 980                            FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY] >> 8, PLANEA_HI));
 981         }
 982
 983         POSTING_READ(DSPFW1);
 984 }
 985
 986 #undef FW_WM_VLV
 987
 988 /* latency must be in 0.1us units. */
 989 static unsigned int vlv_wm_method2(unsigned int pixel_rate,
 990                                    unsigned int pipe_htotal,
 991                                    unsigned int horiz_pixels,
 992                                    unsigned int cpp,
 993                                    unsigned int latency)
 994 {
 995         unsigned int ret;
 996
 997         ret = (latency * pixel_rate) / (pipe_htotal * 10000);
 998         ret = (ret + 1) * horiz_pixels * cpp;
 999         ret = DIV_ROUND_UP(ret, 64);
1000
1001         return ret;
1002 }
1003
1004 static void vlv_setup_wm_latency(struct drm_i915_private *dev_priv)
1005 {
1006         /* all latencies in usec */
1007         dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM2] = 3;
1008
1009         dev_priv->wm.max_level = VLV_WM_LEVEL_PM2;
1010
1011         if (IS_CHERRYVIEW(dev_priv)) {
1012                 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM5] = 12;
1013                 dev_priv->wm.pri_latency[VLV_WM_LEVEL_DDR_DVFS] = 33;
1014
1015                 dev_priv->wm.max_level = VLV_WM_LEVEL_DDR_DVFS;
1016         }
1017 }
1018
1019 static uint16_t vlv_compute_wm_level(const struct intel_crtc_state *crtc_state,
1020                                      const struct intel_plane_state *plane_state,
1021                                      int level)
1022 {
1023         struct intel_plane *plane = to_intel_plane(plane_state->base.plane);
1024         struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
1025         const struct drm_display_mode *adjusted_mode =
1026                 &crtc_state->base.adjusted_mode;
1027         int clock, htotal, cpp, width, wm;
1028
1029         if (dev_priv->wm.pri_latency[level] == 0)
1030                 return USHRT_MAX;
1031
1032         if (!plane_state->base.visible)
1033                 return 0;
1034
1035         cpp = plane_state->base.fb->format->cpp[0];
1036         clock = adjusted_mode->crtc_clock;
1037         htotal = adjusted_mode->crtc_htotal;
1038         width = crtc_state->pipe_src_w;
1039         if (WARN_ON(htotal == 0))
1040                 htotal = 1;
1041
1042         if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
1043                 /*
1044                  * FIXME the formula gives values that are
1045                  * too big for the cursor FIFO, and hence we
1046                  * would never be able to use cursors. For
1047                  * now just hardcode the watermark.
1048                  */
1049                 wm = 63;
1050         } else {
1051                 wm = vlv_wm_method2(clock, htotal, width, cpp,
1052                                     dev_priv->wm.pri_latency[level] * 10);
1053         }
1054
1055         return min_t(int, wm, USHRT_MAX);
1056 }
1057
1058 static bool vlv_need_sprite0_fifo_workaround(unsigned int active_planes)
1059 {
1060         return (active_planes & (BIT(PLANE_SPRITE0) |
1061                                  BIT(PLANE_SPRITE1))) == BIT(PLANE_SPRITE1);
1062 }
1063
1064 static int vlv_compute_fifo(struct intel_crtc_state *crtc_state)
1065 {
1066         struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1067         const struct vlv_pipe_wm *raw =
1068                 &crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM2];
1069         struct vlv_fifo_state *fifo_state = &crtc_state->wm.vlv.fifo_state;
1070         unsigned int active_planes = crtc_state->active_planes & ~BIT(PLANE_CURSOR);
1071         int num_active_planes = hweight32(active_planes);
1072         const int fifo_size = 511;
1073         int fifo_extra, fifo_left = fifo_size;
1074         int sprite0_fifo_extra = 0;
1075         unsigned int total_rate;
1076         enum plane_id plane_id;
1077
1078         /*
1079          * When enabling sprite0 after sprite1 has already been enabled
1080          * we tend to get an underrun unless sprite0 already has some
1081          * FIFO space allcoated. Hence we always allocate at least one
1082          * cacheline for sprite0 whenever sprite1 is enabled.
1083          *
1084          * All other plane enable sequences appear immune to this problem.
1085          */
1086         if (vlv_need_sprite0_fifo_workaround(active_planes))
1087                 sprite0_fifo_extra = 1;
1088
1089         total_rate = raw->plane[PLANE_PRIMARY] +
1090                 raw->plane[PLANE_SPRITE0] +
1091                 raw->plane[PLANE_SPRITE1] +
1092                 sprite0_fifo_extra;
1093
1094         if (total_rate > fifo_size)
1095                 return -EINVAL;
1096
1097         if (total_rate == 0)
1098                 total_rate = 1;
1099
1100         for_each_plane_id_on_crtc(crtc, plane_id) {
1101                 unsigned int rate;
1102
1103                 if ((active_planes & BIT(plane_id)) == 0) {
1104                         fifo_state->plane[plane_id] = 0;
1105                         continue;
1106                 }
1107
1108                 rate = raw->plane[plane_id];
1109                 fifo_state->plane[plane_id] = fifo_size * rate / total_rate;
1110                 fifo_left -= fifo_state->plane[plane_id];
1111         }
1112
1113         fifo_state->plane[PLANE_SPRITE0] += sprite0_fifo_extra;
1114         fifo_left -= sprite0_fifo_extra;
1115
1116         fifo_state->plane[PLANE_CURSOR] = 63;
1117
1118         fifo_extra = DIV_ROUND_UP(fifo_left, num_active_planes ?: 1);
1119
1120         /* spread the remainder evenly */
1121         for_each_plane_id_on_crtc(crtc, plane_id) {
1122                 int plane_extra;
1123
1124                 if (fifo_left == 0)
1125                         break;
1126
1127                 if ((active_planes & BIT(plane_id)) == 0)
1128                         continue;
1129
1130                 plane_extra = min(fifo_extra, fifo_left);
1131                 fifo_state->plane[plane_id] += plane_extra;
1132                 fifo_left -= plane_extra;
1133         }
1134
1135         WARN_ON(active_planes != 0 && fifo_left != 0);
1136
1137         /* give it all to the first plane if none are active */
1138         if (active_planes == 0) {
1139                 WARN_ON(fifo_left != fifo_size);
1140                 fifo_state->plane[PLANE_PRIMARY] = fifo_left;
1141         }
1142
1143         return 0;
1144 }
1145
1146 static int vlv_num_wm_levels(struct drm_i915_private *dev_priv)
1147 {
1148         return dev_priv->wm.max_level + 1;
1149 }
1150
1151 /* mark all levels starting from 'level' as invalid */
1152 static void vlv_invalidate_wms(struct intel_crtc *crtc,
1153                                struct vlv_wm_state *wm_state, int level)
1154 {
1155         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1156
1157         for (; level < vlv_num_wm_levels(dev_priv); level++) {
1158                 enum plane_id plane_id;
1159
1160                 for_each_plane_id_on_crtc(crtc, plane_id)
1161                         wm_state->wm[level].plane[plane_id] = USHRT_MAX;
1162
1163                 wm_state->sr[level].cursor = USHRT_MAX;
1164                 wm_state->sr[level].plane = USHRT_MAX;
1165         }
1166 }
1167
1168 static u16 vlv_invert_wm_value(u16 wm, u16 fifo_size)
1169 {
1170         if (wm > fifo_size)
1171                 return USHRT_MAX;
1172         else
1173                 return fifo_size - wm;
1174 }
1175
1176 /*
1177  * Starting from 'level' set all higher
1178  * levels to 'value' in the "raw" watermarks.
1179  */
1180 static bool vlv_raw_plane_wm_set(struct intel_crtc_state *crtc_state,
1181                                  int level, enum plane_id plane_id, u16 value)
1182 {
1183         struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1184         int num_levels = vlv_num_wm_levels(dev_priv);
1185         bool dirty = false;
1186
1187         for (; level < num_levels; level++) {
1188                 struct vlv_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1189
1190                 dirty |= raw->plane[plane_id] != value;
1191                 raw->plane[plane_id] = value;
1192         }
1193
1194         return dirty;
1195 }
1196
1197 static bool vlv_plane_wm_compute(struct intel_crtc_state *crtc_state,
1198                                  const struct intel_plane_state *plane_state)
1199 {
1200         struct intel_plane *plane = to_intel_plane(plane_state->base.plane);
1201         enum plane_id plane_id = plane->id;
1202         int num_levels = vlv_num_wm_levels(to_i915(plane->base.dev));
1203         int level;
1204         bool dirty = false;
1205
1206         if (!plane_state->base.visible) {
1207                 dirty |= vlv_raw_plane_wm_set(crtc_state, 0, plane_id, 0);
1208                 goto out;
1209         }
1210
1211         for (level = 0; level < num_levels; level++) {
1212                 struct vlv_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1213                 int wm = vlv_compute_wm_level(crtc_state, plane_state, level);
1214                 int max_wm = plane_id == PLANE_CURSOR ? 63 : 511;
1215
1216                 if (wm > max_wm)
1217                         break;
1218
1219                 dirty |= raw->plane[plane_id] != wm;
1220                 raw->plane[plane_id] = wm;
1221         }
1222
1223         /* mark all higher levels as invalid */
1224         dirty |= vlv_raw_plane_wm_set(crtc_state, level, plane_id, USHRT_MAX);
1225
1226 out:
1227         if (dirty)
1228                 DRM_DEBUG_KMS("%s wms: [0]=%d,[1]=%d,[2]=%d\n",
1229                               plane->base.name,
1230                               crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM2].plane[plane_id],
1231                               crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM5].plane[plane_id],
1232                               crtc_state->wm.vlv.raw[VLV_WM_LEVEL_DDR_DVFS].plane[plane_id]);
1233
1234         return dirty;
1235 }
1236
1237 static bool vlv_plane_wm_is_valid(const struct intel_crtc_state *crtc_state,
1238                                   enum plane_id plane_id, int level)
1239 {
1240         const struct vlv_pipe_wm *raw =
1241                 &crtc_state->wm.vlv.raw[level];
1242         const struct vlv_fifo_state *fifo_state =
1243                 &crtc_state->wm.vlv.fifo_state;
1244
1245         return raw->plane[plane_id] <= fifo_state->plane[plane_id];
1246 }
1247
1248 static bool vlv_crtc_wm_is_valid(const struct intel_crtc_state *crtc_state, int level)
1249 {
1250         return vlv_plane_wm_is_valid(crtc_state, PLANE_PRIMARY, level) &&
1251                 vlv_plane_wm_is_valid(crtc_state, PLANE_SPRITE0, level) &&
1252                 vlv_plane_wm_is_valid(crtc_state, PLANE_SPRITE1, level) &&
1253                 vlv_plane_wm_is_valid(crtc_state, PLANE_CURSOR, level);
1254 }
1255
1256 static int vlv_compute_pipe_wm(struct intel_crtc_state *crtc_state)
1257 {
1258         struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1259         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1260         struct intel_atomic_state *state =
1261                 to_intel_atomic_state(crtc_state->base.state);
1262         struct vlv_wm_state *wm_state = &crtc_state->wm.vlv.optimal;
1263         const struct vlv_fifo_state *fifo_state =
1264                 &crtc_state->wm.vlv.fifo_state;
1265         int num_active_planes = hweight32(crtc_state->active_planes &
1266                                           ~BIT(PLANE_CURSOR));
1267         bool needs_modeset = drm_atomic_crtc_needs_modeset(&crtc_state->base);
1268         struct intel_plane_state *plane_state;
1269         struct intel_plane *plane;
1270         enum plane_id plane_id;
1271         int level, ret, i;
1272         unsigned int dirty = 0;
1273
1274         for_each_intel_plane_in_state(state, plane, plane_state, i) {
1275                 const struct intel_plane_state *old_plane_state =
1276                         to_intel_plane_state(plane->base.state);
1277
1278                 if (plane_state->base.crtc != &crtc->base &&
1279                     old_plane_state->base.crtc != &crtc->base)
1280                         continue;
1281
1282                 if (vlv_plane_wm_compute(crtc_state, plane_state))
1283                         dirty |= BIT(plane->id);
1284         }
1285
1286         /*
1287          * DSPARB registers may have been reset due to the
1288          * power well being turned off. Make sure we restore
1289          * them to a consistent state even if no primary/sprite
1290          * planes are initially active.
1291          */
1292         if (needs_modeset)
1293                 crtc_state->fifo_changed = true;
1294
1295         if (!dirty)
1296                 return 0;
1297
1298         /* cursor changes don't warrant a FIFO recompute */
1299         if (dirty & ~BIT(PLANE_CURSOR)) {
1300                 const struct intel_crtc_state *old_crtc_state =
1301                         to_intel_crtc_state(crtc->base.state);
1302                 const struct vlv_fifo_state *old_fifo_state =
1303                         &old_crtc_state->wm.vlv.fifo_state;
1304
1305                 ret = vlv_compute_fifo(crtc_state);
1306                 if (ret)
1307                         return ret;
1308
1309                 if (needs_modeset ||
1310                     memcmp(old_fifo_state, fifo_state,
1311                            sizeof(*fifo_state)) != 0)
1312                         crtc_state->fifo_changed = true;
1313         }
1314
1315         /* initially allow all levels */
1316         wm_state->num_levels = vlv_num_wm_levels(dev_priv);
1317         /*
1318          * Note that enabling cxsr with no primary/sprite planes
1319          * enabled can wedge the pipe. Hence we only allow cxsr
1320          * with exactly one enabled primary/sprite plane.
1321          */
1322         wm_state->cxsr = crtc->pipe != PIPE_C && num_active_planes == 1;
1323
1324         for (level = 0; level < wm_state->num_levels; level++) {
1325                 const struct vlv_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1326                 const int sr_fifo_size = INTEL_INFO(dev_priv)->num_pipes * 512 - 1;
1327
1328                 if (!vlv_crtc_wm_is_valid(crtc_state, level))
1329                         break;
1330
1331                 for_each_plane_id_on_crtc(crtc, plane_id) {
1332                         wm_state->wm[level].plane[plane_id] =
1333                                 vlv_invert_wm_value(raw->plane[plane_id],
1334                                                     fifo_state->plane[plane_id]);
1335                 }
1336
1337                 wm_state->sr[level].plane =
1338                         vlv_invert_wm_value(max3(raw->plane[PLANE_PRIMARY],
1339                                                  raw->plane[PLANE_SPRITE0],
1340                                                  raw->plane[PLANE_SPRITE1]),
1341                                             sr_fifo_size);
1342
1343                 wm_state->sr[level].cursor =
1344                         vlv_invert_wm_value(raw->plane[PLANE_CURSOR],
1345                                             63);
1346         }
1347
1348         if (level == 0)
1349                 return -EINVAL;
1350
1351         /* limit to only levels we can actually handle */
1352         wm_state->num_levels = level;
1353
1354         /* invalidate the higher levels */
1355         vlv_invalidate_wms(crtc, wm_state, level);
1356
1357         return 0;
1358 }
1359
1360 #define VLV_FIFO(plane, value) \
1361         (((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1362
1363 static void vlv_atomic_update_fifo(struct intel_atomic_state *state,
1364                                    struct intel_crtc_state *crtc_state)
1365 {
1366         struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1367         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1368         const struct vlv_fifo_state *fifo_state =
1369                 &crtc_state->wm.vlv.fifo_state;
1370         int sprite0_start, sprite1_start, fifo_size;
1371
1372         if (!crtc_state->fifo_changed)
1373                 return;
1374
1375         sprite0_start = fifo_state->plane[PLANE_PRIMARY];
1376         sprite1_start = fifo_state->plane[PLANE_SPRITE0] + sprite0_start;
1377         fifo_size = fifo_state->plane[PLANE_SPRITE1] + sprite1_start;
1378
1379         WARN_ON(fifo_state->plane[PLANE_CURSOR] != 63);
1380         WARN_ON(fifo_size != 511);
1381
1382         trace_vlv_fifo_size(crtc, sprite0_start, sprite1_start, fifo_size);
1383
1384         /*
1385          * uncore.lock serves a double purpose here. It allows us to
1386          * use the less expensive I915_{READ,WRITE}_FW() functions, and
1387          * it protects the DSPARB registers from getting clobbered by
1388          * parallel updates from multiple pipes.
1389          *
1390          * intel_pipe_update_start() has already disabled interrupts
1391          * for us, so a plain spin_lock() is sufficient here.
1392          */
1393         spin_lock(&dev_priv->uncore.lock);
1394
1395         switch (crtc->pipe) {
1396                 uint32_t dsparb, dsparb2, dsparb3;
1397         case PIPE_A:
1398                 dsparb = I915_READ_FW(DSPARB);
1399                 dsparb2 = I915_READ_FW(DSPARB2);
1400
1401                 dsparb &= ~(VLV_FIFO(SPRITEA, 0xff) |
1402                             VLV_FIFO(SPRITEB, 0xff));
1403                 dsparb |= (VLV_FIFO(SPRITEA, sprite0_start) |
1404                            VLV_FIFO(SPRITEB, sprite1_start));
1405
1406                 dsparb2 &= ~(VLV_FIFO(SPRITEA_HI, 0x1) |
1407                              VLV_FIFO(SPRITEB_HI, 0x1));
1408                 dsparb2 |= (VLV_FIFO(SPRITEA_HI, sprite0_start >> 8) |
1409                            VLV_FIFO(SPRITEB_HI, sprite1_start >> 8));
1410
1411                 I915_WRITE_FW(DSPARB, dsparb);
1412                 I915_WRITE_FW(DSPARB2, dsparb2);
1413                 break;
1414         case PIPE_B:
1415                 dsparb = I915_READ_FW(DSPARB);
1416                 dsparb2 = I915_READ_FW(DSPARB2);
1417
1418                 dsparb &= ~(VLV_FIFO(SPRITEC, 0xff) |
1419                             VLV_FIFO(SPRITED, 0xff));
1420                 dsparb |= (VLV_FIFO(SPRITEC, sprite0_start) |
1421                            VLV_FIFO(SPRITED, sprite1_start));
1422
1423                 dsparb2 &= ~(VLV_FIFO(SPRITEC_HI, 0xff) |
1424                              VLV_FIFO(SPRITED_HI, 0xff));
1425                 dsparb2 |= (VLV_FIFO(SPRITEC_HI, sprite0_start >> 8) |
1426                            VLV_FIFO(SPRITED_HI, sprite1_start >> 8));
1427
1428                 I915_WRITE_FW(DSPARB, dsparb);
1429                 I915_WRITE_FW(DSPARB2, dsparb2);
1430                 break;
1431         case PIPE_C:
1432                 dsparb3 = I915_READ_FW(DSPARB3);
1433                 dsparb2 = I915_READ_FW(DSPARB2);
1434
1435                 dsparb3 &= ~(VLV_FIFO(SPRITEE, 0xff) |
1436                              VLV_FIFO(SPRITEF, 0xff));
1437                 dsparb3 |= (VLV_FIFO(SPRITEE, sprite0_start) |
1438                             VLV_FIFO(SPRITEF, sprite1_start));
1439
1440                 dsparb2 &= ~(VLV_FIFO(SPRITEE_HI, 0xff) |
1441                              VLV_FIFO(SPRITEF_HI, 0xff));
1442                 dsparb2 |= (VLV_FIFO(SPRITEE_HI, sprite0_start >> 8) |
1443                            VLV_FIFO(SPRITEF_HI, sprite1_start >> 8));
1444
1445                 I915_WRITE_FW(DSPARB3, dsparb3);
1446                 I915_WRITE_FW(DSPARB2, dsparb2);
1447                 break;
1448         default:
1449                 break;
1450         }
1451
1452         POSTING_READ_FW(DSPARB);
1453
1454         spin_unlock(&dev_priv->uncore.lock);
1455 }
1456
1457 #undef VLV_FIFO
1458
1459 static int vlv_compute_intermediate_wm(struct drm_device *dev,
1460                                        struct intel_crtc *crtc,
1461                                        struct intel_crtc_state *crtc_state)
1462 {
1463         struct vlv_wm_state *intermediate = &crtc_state->wm.vlv.intermediate;
1464         const struct vlv_wm_state *optimal = &crtc_state->wm.vlv.optimal;
1465         const struct vlv_wm_state *active = &crtc->wm.active.vlv;
1466         int level;
1467
1468         intermediate->num_levels = min(optimal->num_levels, active->num_levels);
1469         intermediate->cxsr = optimal->cxsr && active->cxsr &&
1470                 !crtc_state->disable_cxsr;
1471
1472         for (level = 0; level < intermediate->num_levels; level++) {
1473                 enum plane_id plane_id;
1474
1475                 for_each_plane_id_on_crtc(crtc, plane_id) {
1476                         intermediate->wm[level].plane[plane_id] =
1477                                 min(optimal->wm[level].plane[plane_id],
1478                                     active->wm[level].plane[plane_id]);
1479                 }
1480
1481                 intermediate->sr[level].plane = min(optimal->sr[level].plane,
1482                                                     active->sr[level].plane);
1483                 intermediate->sr[level].cursor = min(optimal->sr[level].cursor,
1484                                                      active->sr[level].cursor);
1485         }
1486
1487         vlv_invalidate_wms(crtc, intermediate, level);
1488
1489         /*
1490          * If our intermediate WM are identical to the final WM, then we can
1491          * omit the post-vblank programming; only update if it's different.
1492          */
1493         if (memcmp(intermediate, optimal, sizeof(*intermediate)) != 0)
1494                 crtc_state->wm.need_postvbl_update = true;
1495
1496         return 0;
1497 }
1498
1499 static void vlv_merge_wm(struct drm_i915_private *dev_priv,
1500                          struct vlv_wm_values *wm)
1501 {
1502         struct intel_crtc *crtc;
1503         int num_active_crtcs = 0;
1504
1505         wm->level = dev_priv->wm.max_level;
1506         wm->cxsr = true;
1507
1508         for_each_intel_crtc(&dev_priv->drm, crtc) {
1509                 const struct vlv_wm_state *wm_state = &crtc->wm.active.vlv;
1510
1511                 if (!crtc->active)
1512                         continue;
1513
1514                 if (!wm_state->cxsr)
1515                         wm->cxsr = false;
1516
1517                 num_active_crtcs++;
1518                 wm->level = min_t(int, wm->level, wm_state->num_levels - 1);
1519         }
1520
1521         if (num_active_crtcs != 1)
1522                 wm->cxsr = false;
1523
1524         if (num_active_crtcs > 1)
1525                 wm->level = VLV_WM_LEVEL_PM2;
1526
1527         for_each_intel_crtc(&dev_priv->drm, crtc) {
1528                 const struct vlv_wm_state *wm_state = &crtc->wm.active.vlv;
1529                 enum pipe pipe = crtc->pipe;
1530
1531                 wm->pipe[pipe] = wm_state->wm[wm->level];
1532                 if (crtc->active && wm->cxsr)
1533                         wm->sr = wm_state->sr[wm->level];
1534
1535                 wm->ddl[pipe].plane[PLANE_PRIMARY] = DDL_PRECISION_HIGH | 2;
1536                 wm->ddl[pipe].plane[PLANE_SPRITE0] = DDL_PRECISION_HIGH | 2;
1537                 wm->ddl[pipe].plane[PLANE_SPRITE1] = DDL_PRECISION_HIGH | 2;
1538                 wm->ddl[pipe].plane[PLANE_CURSOR] = DDL_PRECISION_HIGH | 2;
1539         }
1540 }
1541
1542 static bool is_disabling(int old, int new, int threshold)
1543 {
1544         return old >= threshold && new < threshold;
1545 }
1546
1547 static bool is_enabling(int old, int new, int threshold)
1548 {
1549         return old < threshold && new >= threshold;
1550 }
1551
1552 static void vlv_program_watermarks(struct drm_i915_private *dev_priv)
1553 {
1554         struct vlv_wm_values *old_wm = &dev_priv->wm.vlv;
1555         struct vlv_wm_values new_wm = {};
1556
1557         vlv_merge_wm(dev_priv, &new_wm);
1558
1559         if (memcmp(old_wm, &new_wm, sizeof(new_wm)) == 0)
1560                 return;
1561
1562         if (is_disabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_DDR_DVFS))
1563                 chv_set_memory_dvfs(dev_priv, false);
1564
1565         if (is_disabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_PM5))
1566                 chv_set_memory_pm5(dev_priv, false);
1567
1568         if (is_disabling(old_wm->cxsr, new_wm.cxsr, true))
1569                 _intel_set_memory_cxsr(dev_priv, false);
1570
1571         vlv_write_wm_values(dev_priv, &new_wm);
1572
1573         if (is_enabling(old_wm->cxsr, new_wm.cxsr, true))
1574                 _intel_set_memory_cxsr(dev_priv, true);
1575
1576         if (is_enabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_PM5))
1577                 chv_set_memory_pm5(dev_priv, true);
1578
1579         if (is_enabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_DDR_DVFS))
1580                 chv_set_memory_dvfs(dev_priv, true);
1581
1582         *old_wm = new_wm;
1583 }
1584
1585 static void vlv_initial_watermarks(struct intel_atomic_state *state,
1586                                    struct intel_crtc_state *crtc_state)
1587 {
1588         struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1589         struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1590
1591         mutex_lock(&dev_priv->wm.wm_mutex);
1592         crtc->wm.active.vlv = crtc_state->wm.vlv.intermediate;
1593         vlv_program_watermarks(dev_priv);
1594         mutex_unlock(&dev_priv->wm.wm_mutex);
1595 }
1596
1597 static void vlv_optimize_watermarks(struct intel_atomic_state *state,
1598                                     struct intel_crtc_state *crtc_state)
1599 {
1600         struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1601         struct intel_crtc *intel_crtc = to_intel_crtc(crtc_state->base.crtc);
1602
1603         if (!crtc_state->wm.need_postvbl_update)
1604                 return;
1605
1606         mutex_lock(&dev_priv->wm.wm_mutex);
1607         intel_crtc->wm.active.vlv = crtc_state->wm.vlv.optimal;
1608         vlv_program_watermarks(dev_priv);
1609         mutex_unlock(&dev_priv->wm.wm_mutex);
1610 }
1611
1612 #define single_plane_enabled(mask) is_power_of_2(mask)
1613
1614 static void g4x_update_wm(struct intel_crtc *crtc)
1615 {
1616         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1617         static const int sr_latency_ns = 12000;
1618         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1619         int plane_sr, cursor_sr;
1620         unsigned int enabled = 0;
1621         bool cxsr_enabled;
1622
1623         if (g4x_compute_wm0(dev_priv, PIPE_A,
1624                             &g4x_wm_info, pessimal_latency_ns,
1625                             &g4x_cursor_wm_info, pessimal_latency_ns,
1626                             &planea_wm, &cursora_wm))
1627                 enabled |= 1 << PIPE_A;
1628
1629         if (g4x_compute_wm0(dev_priv, PIPE_B,
1630                             &g4x_wm_info, pessimal_latency_ns,
1631                             &g4x_cursor_wm_info, pessimal_latency_ns,
1632                             &planeb_wm, &cursorb_wm))
1633                 enabled |= 1 << PIPE_B;
1634
1635         if (single_plane_enabled(enabled) &&
1636             g4x_compute_srwm(dev_priv, ffs(enabled) - 1,
1637                              sr_latency_ns,
1638                              &g4x_wm_info,
1639                              &g4x_cursor_wm_info,
1640                              &plane_sr, &cursor_sr)) {
1641                 cxsr_enabled = true;
1642         } else {
1643                 cxsr_enabled = false;
1644                 intel_set_memory_cxsr(dev_priv, false);
1645                 plane_sr = cursor_sr = 0;
1646         }
1647
1648         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, "
1649                       "B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1650                       planea_wm, cursora_wm,
1651                       planeb_wm, cursorb_wm,
1652                       plane_sr, cursor_sr);
1653
1654         I915_WRITE(DSPFW1,
1655                    FW_WM(plane_sr, SR) |
1656                    FW_WM(cursorb_wm, CURSORB) |
1657                    FW_WM(planeb_wm, PLANEB) |
1658                    FW_WM(planea_wm, PLANEA));
1659         I915_WRITE(DSPFW2,
1660                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1661                    FW_WM(cursora_wm, CURSORA));
1662         /* HPLL off in SR has some issues on G4x... disable it */
1663         I915_WRITE(DSPFW3,
1664                    (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1665                    FW_WM(cursor_sr, CURSOR_SR));
1666
1667         if (cxsr_enabled)
1668                 intel_set_memory_cxsr(dev_priv, true);
1669 }
1670
1671 static void i965_update_wm(struct intel_crtc *unused_crtc)
1672 {
1673         struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
1674         struct intel_crtc *crtc;
1675         int srwm = 1;
1676         int cursor_sr = 16;
1677         bool cxsr_enabled;
1678
1679         /* Calc sr entries for one plane configs */
1680         crtc = single_enabled_crtc(dev_priv);
1681         if (crtc) {
1682                 /* self-refresh has much higher latency */
1683                 static const int sr_latency_ns = 12000;
1684                 const struct drm_display_mode *adjusted_mode =
1685                         &crtc->config->base.adjusted_mode;
1686                 const struct drm_framebuffer *fb =
1687                         crtc->base.primary->state->fb;
1688                 int clock = adjusted_mode->crtc_clock;
1689                 int htotal = adjusted_mode->crtc_htotal;
1690                 int hdisplay = crtc->config->pipe_src_w;
1691                 int cpp = fb->format->cpp[0];
1692                 unsigned long line_time_us;
1693                 int entries;
1694
1695                 line_time_us = max(htotal * 1000 / clock, 1);
1696
1697                 /* Use ns/us then divide to preserve precision */
1698                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1699                         cpp * hdisplay;
1700                 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1701                 srwm = I965_FIFO_SIZE - entries;
1702                 if (srwm < 0)
1703                         srwm = 1;
1704                 srwm &= 0x1ff;
1705                 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1706                               entries, srwm);
1707
1708                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1709                         cpp * crtc->base.cursor->state->crtc_w;
1710                 entries = DIV_ROUND_UP(entries,
1711                                           i965_cursor_wm_info.cacheline_size);
1712                 cursor_sr = i965_cursor_wm_info.fifo_size -
1713                         (entries + i965_cursor_wm_info.guard_size);
1714
1715                 if (cursor_sr > i965_cursor_wm_info.max_wm)
1716                         cursor_sr = i965_cursor_wm_info.max_wm;
1717
1718                 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1719                               "cursor %d\n", srwm, cursor_sr);
1720
1721                 cxsr_enabled = true;
1722         } else {
1723                 cxsr_enabled = false;
1724                 /* Turn off self refresh if both pipes are enabled */
1725                 intel_set_memory_cxsr(dev_priv, false);
1726         }
1727
1728         DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1729                       srwm);
1730
1731         /* 965 has limitations... */
1732         I915_WRITE(DSPFW1, FW_WM(srwm, SR) |
1733                    FW_WM(8, CURSORB) |
1734                    FW_WM(8, PLANEB) |
1735                    FW_WM(8, PLANEA));
1736         I915_WRITE(DSPFW2, FW_WM(8, CURSORA) |
1737                    FW_WM(8, PLANEC_OLD));
1738         /* update cursor SR watermark */
1739         I915_WRITE(DSPFW3, FW_WM(cursor_sr, CURSOR_SR));
1740
1741         if (cxsr_enabled)
1742                 intel_set_memory_cxsr(dev_priv, true);
1743 }
1744
1745 #undef FW_WM
1746
1747 static void i9xx_update_wm(struct intel_crtc *unused_crtc)
1748 {
1749         struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
1750         const struct intel_watermark_params *wm_info;
1751         uint32_t fwater_lo;
1752         uint32_t fwater_hi;
1753         int cwm, srwm = 1;
1754         int fifo_size;
1755         int planea_wm, planeb_wm;
1756         struct intel_crtc *crtc, *enabled = NULL;
1757
1758         if (IS_I945GM(dev_priv))
1759                 wm_info = &i945_wm_info;
1760         else if (!IS_GEN2(dev_priv))
1761                 wm_info = &i915_wm_info;
1762         else
1763                 wm_info = &i830_a_wm_info;
1764
1765         fifo_size = dev_priv->display.get_fifo_size(dev_priv, 0);
1766         crtc = intel_get_crtc_for_plane(dev_priv, 0);
1767         if (intel_crtc_active(crtc)) {
1768                 const struct drm_display_mode *adjusted_mode =
1769                         &crtc->config->base.adjusted_mode;
1770                 const struct drm_framebuffer *fb =
1771                         crtc->base.primary->state->fb;
1772                 int cpp;
1773
1774                 if (IS_GEN2(dev_priv))
1775                         cpp = 4;
1776                 else
1777                         cpp = fb->format->cpp[0];
1778
1779                 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1780                                                wm_info, fifo_size, cpp,
1781                                                pessimal_latency_ns);
1782                 enabled = crtc;
1783         } else {
1784                 planea_wm = fifo_size - wm_info->guard_size;
1785                 if (planea_wm > (long)wm_info->max_wm)
1786                         planea_wm = wm_info->max_wm;
1787         }
1788
1789         if (IS_GEN2(dev_priv))
1790                 wm_info = &i830_bc_wm_info;
1791
1792         fifo_size = dev_priv->display.get_fifo_size(dev_priv, 1);
1793         crtc = intel_get_crtc_for_plane(dev_priv, 1);
1794         if (intel_crtc_active(crtc)) {
1795                 const struct drm_display_mode *adjusted_mode =
1796                         &crtc->config->base.adjusted_mode;
1797                 const struct drm_framebuffer *fb =
1798                         crtc->base.primary->state->fb;
1799                 int cpp;
1800
1801                 if (IS_GEN2(dev_priv))
1802                         cpp = 4;
1803                 else
1804                         cpp = fb->format->cpp[0];
1805
1806                 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1807                                                wm_info, fifo_size, cpp,
1808                                                pessimal_latency_ns);
1809                 if (enabled == NULL)
1810                         enabled = crtc;
1811                 else
1812                         enabled = NULL;
1813         } else {
1814                 planeb_wm = fifo_size - wm_info->guard_size;
1815                 if (planeb_wm > (long)wm_info->max_wm)
1816                         planeb_wm = wm_info->max_wm;
1817         }
1818
1819         DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1820
1821         if (IS_I915GM(dev_priv) && enabled) {
1822                 struct drm_i915_gem_object *obj;
1823
1824                 obj = intel_fb_obj(enabled->base.primary->state->fb);
1825
1826                 /* self-refresh seems busted with untiled */
1827                 if (!i915_gem_object_is_tiled(obj))
1828                         enabled = NULL;
1829         }
1830
1831         /*
1832          * Overlay gets an aggressive default since video jitter is bad.
1833          */
1834         cwm = 2;
1835
1836         /* Play safe and disable self-refresh before adjusting watermarks. */
1837         intel_set_memory_cxsr(dev_priv, false);
1838
1839         /* Calc sr entries for one plane configs */
1840         if (HAS_FW_BLC(dev_priv) && enabled) {
1841                 /* self-refresh has much higher latency */
1842                 static const int sr_latency_ns = 6000;
1843                 const struct drm_display_mode *adjusted_mode =
1844                         &enabled->config->base.adjusted_mode;
1845                 const struct drm_framebuffer *fb =
1846                         enabled->base.primary->state->fb;
1847                 int clock = adjusted_mode->crtc_clock;
1848                 int htotal = adjusted_mode->crtc_htotal;
1849                 int hdisplay = enabled->config->pipe_src_w;
1850                 int cpp;
1851                 unsigned long line_time_us;
1852                 int entries;
1853
1854                 if (IS_I915GM(dev_priv) || IS_I945GM(dev_priv))
1855                         cpp = 4;
1856                 else
1857                         cpp = fb->format->cpp[0];
1858
1859                 line_time_us = max(htotal * 1000 / clock, 1);
1860
1861                 /* Use ns/us then divide to preserve precision */
1862                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1863                         cpp * hdisplay;
1864                 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1865                 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1866                 srwm = wm_info->fifo_size - entries;
1867                 if (srwm < 0)
1868                         srwm = 1;
1869
1870                 if (IS_I945G(dev_priv) || IS_I945GM(dev_priv))
1871                         I915_WRITE(FW_BLC_SELF,
1872                                    FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1873                 else
1874                         I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1875         }
1876
1877         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1878                       planea_wm, planeb_wm, cwm, srwm);
1879
1880         fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1881         fwater_hi = (cwm & 0x1f);
1882
1883         /* Set request length to 8 cachelines per fetch */
1884         fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1885         fwater_hi = fwater_hi | (1 << 8);
1886
1887         I915_WRITE(FW_BLC, fwater_lo);
1888         I915_WRITE(FW_BLC2, fwater_hi);
1889
1890         if (enabled)
1891                 intel_set_memory_cxsr(dev_priv, true);
1892 }
1893
1894 static void i845_update_wm(struct intel_crtc *unused_crtc)
1895 {
1896         struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
1897         struct intel_crtc *crtc;
1898         const struct drm_display_mode *adjusted_mode;
1899         uint32_t fwater_lo;
1900         int planea_wm;
1901
1902         crtc = single_enabled_crtc(dev_priv);
1903         if (crtc == NULL)
1904                 return;
1905
1906         adjusted_mode = &crtc->config->base.adjusted_mode;
1907         planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1908                                        &i845_wm_info,
1909                                        dev_priv->display.get_fifo_size(dev_priv, 0),
1910                                        4, pessimal_latency_ns);
1911         fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1912         fwater_lo |= (3<<8) | planea_wm;
1913
1914         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1915
1916         I915_WRITE(FW_BLC, fwater_lo);
1917 }
1918
1919 /* latency must be in 0.1us units. */
1920 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t cpp, uint32_t latency)
1921 {
1922         uint64_t ret;
1923
1924         if (WARN(latency == 0, "Latency value missing\n"))
1925                 return UINT_MAX;
1926
1927         ret = (uint64_t) pixel_rate * cpp * latency;
1928         ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
1929
1930         return ret;
1931 }
1932
1933 /* latency must be in 0.1us units. */
1934 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
1935                                uint32_t horiz_pixels, uint8_t cpp,
1936                                uint32_t latency)
1937 {
1938         uint32_t ret;
1939
1940         if (WARN(latency == 0, "Latency value missing\n"))
1941                 return UINT_MAX;
1942         if (WARN_ON(!pipe_htotal))
1943                 return UINT_MAX;
1944
1945         ret = (latency * pixel_rate) / (pipe_htotal * 10000);
1946         ret = (ret + 1) * horiz_pixels * cpp;
1947         ret = DIV_ROUND_UP(ret, 64) + 2;
1948         return ret;
1949 }
1950
1951 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
1952                            uint8_t cpp)
1953 {
1954         /*
1955          * Neither of these should be possible since this function shouldn't be
1956          * called if the CRTC is off or the plane is invisible.  But let's be
1957          * extra paranoid to avoid a potential divide-by-zero if we screw up
1958          * elsewhere in the driver.
1959          */
1960         if (WARN_ON(!cpp))
1961                 return 0;
1962         if (WARN_ON(!horiz_pixels))
1963                 return 0;
1964
1965         return DIV_ROUND_UP(pri_val * 64, horiz_pixels * cpp) + 2;
1966 }
1967
1968 struct ilk_wm_maximums {
1969         uint16_t pri;
1970         uint16_t spr;
1971         uint16_t cur;
1972         uint16_t fbc;
1973 };
1974
1975 /*
1976  * For both WM_PIPE and WM_LP.
1977  * mem_value must be in 0.1us units.
1978  */
1979 static uint32_t ilk_compute_pri_wm(const struct intel_crtc_state *cstate,
1980                                    const struct intel_plane_state *pstate,
1981                                    uint32_t mem_value,
1982                                    bool is_lp)
1983 {
1984         uint32_t method1, method2;
1985         int cpp;
1986
1987         if (!intel_wm_plane_visible(cstate, pstate))
1988                 return 0;
1989
1990         cpp = pstate->base.fb->format->cpp[0];
1991
1992         method1 = ilk_wm_method1(cstate->pixel_rate, cpp, mem_value);
1993
1994         if (!is_lp)
1995                 return method1;
1996
1997         method2 = ilk_wm_method2(cstate->pixel_rate,
1998                                  cstate->base.adjusted_mode.crtc_htotal,
1999                                  drm_rect_width(&pstate->base.dst),
2000                                  cpp, mem_value);
2001
2002         return min(method1, method2);
2003 }
2004
2005 /*
2006  * For both WM_PIPE and WM_LP.
2007  * mem_value must be in 0.1us units.
2008  */
2009 static uint32_t ilk_compute_spr_wm(const struct intel_crtc_state *cstate,
2010                                    const struct intel_plane_state *pstate,
2011                                    uint32_t mem_value)
2012 {
2013         uint32_t method1, method2;
2014         int cpp;
2015
2016         if (!intel_wm_plane_visible(cstate, pstate))
2017                 return 0;
2018
2019         cpp = pstate->base.fb->format->cpp[0];
2020
2021         method1 = ilk_wm_method1(cstate->pixel_rate, cpp, mem_value);
2022         method2 = ilk_wm_method2(cstate->pixel_rate,
2023                                  cstate->base.adjusted_mode.crtc_htotal,
2024                                  drm_rect_width(&pstate->base.dst),
2025                                  cpp, mem_value);
2026         return min(method1, method2);
2027 }
2028
2029 /*
2030  * For both WM_PIPE and WM_LP.
2031  * mem_value must be in 0.1us units.
2032  */
2033 static uint32_t ilk_compute_cur_wm(const struct intel_crtc_state *cstate,
2034                                    const struct intel_plane_state *pstate,
2035                                    uint32_t mem_value)
2036 {
2037         int cpp;
2038
2039         if (!intel_wm_plane_visible(cstate, pstate))
2040                 return 0;
2041
2042         cpp = pstate->base.fb->format->cpp[0];
2043
2044         return ilk_wm_method2(cstate->pixel_rate,
2045                               cstate->base.adjusted_mode.crtc_htotal,
2046                               pstate->base.crtc_w, cpp, mem_value);
2047 }
2048
2049 /* Only for WM_LP. */
2050 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state *cstate,
2051                                    const struct intel_plane_state *pstate,
2052                                    uint32_t pri_val)
2053 {
2054         int cpp;
2055
2056         if (!intel_wm_plane_visible(cstate, pstate))
2057                 return 0;
2058
2059         cpp = pstate->base.fb->format->cpp[0];
2060
2061         return ilk_wm_fbc(pri_val, drm_rect_width(&pstate->base.dst), cpp);
2062 }
2063
2064 static unsigned int
2065 ilk_display_fifo_size(const struct drm_i915_private *dev_priv)
2066 {
2067         if (INTEL_GEN(dev_priv) >= 8)
2068                 return 3072;
2069         else if (INTEL_GEN(dev_priv) >= 7)
2070                 return 768;
2071         else
2072                 return 512;
2073 }
2074
2075 static unsigned int
2076 ilk_plane_wm_reg_max(const struct drm_i915_private *dev_priv,
2077                      int level, bool is_sprite)
2078 {
2079         if (INTEL_GEN(dev_priv) >= 8)
2080                 /* BDW primary/sprite plane watermarks */
2081                 return level == 0 ? 255 : 2047;
2082         else if (INTEL_GEN(dev_priv) >= 7)
2083                 /* IVB/HSW primary/sprite plane watermarks */
2084                 return level == 0 ? 127 : 1023;
2085         else if (!is_sprite)
2086                 /* ILK/SNB primary plane watermarks */
2087                 return level == 0 ? 127 : 511;
2088         else
2089                 /* ILK/SNB sprite plane watermarks */
2090                 return level == 0 ? 63 : 255;
2091 }
2092
2093 static unsigned int
2094 ilk_cursor_wm_reg_max(const struct drm_i915_private *dev_priv, int level)
2095 {
2096         if (INTEL_GEN(dev_priv) >= 7)
2097                 return level == 0 ? 63 : 255;
2098         else
2099                 return level == 0 ? 31 : 63;
2100 }
2101
2102 static unsigned int ilk_fbc_wm_reg_max(const struct drm_i915_private *dev_priv)
2103 {
2104         if (INTEL_GEN(dev_priv) >= 8)
2105                 return 31;
2106         else
2107                 return 15;
2108 }
2109
2110 /* Calculate the maximum primary/sprite plane watermark */
2111 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
2112                                      int level,
2113                                      const struct intel_wm_config *config,
2114                                      enum intel_ddb_partitioning ddb_partitioning,
2115                                      bool is_sprite)
2116 {
2117         struct drm_i915_private *dev_priv = to_i915(dev);
2118         unsigned int fifo_size = ilk_display_fifo_size(dev_priv);
2119
2120         /* if sprites aren't enabled, sprites get nothing */
2121         if (is_sprite && !config->sprites_enabled)
2122                 return 0;
2123
2124         /* HSW allows LP1+ watermarks even with multiple pipes */
2125         if (level == 0 || config->num_pipes_active > 1) {
2126                 fifo_size /= INTEL_INFO(dev_priv)->num_pipes;
2127
2128                 /*
2129                  * For some reason the non self refresh
2130                  * FIFO size is only half of the self
2131                  * refresh FIFO size on ILK/SNB.
2132                  */
2133                 if (INTEL_GEN(dev_priv) <= 6)
2134                         fifo_size /= 2;
2135         }
2136
2137         if (config->sprites_enabled) {
2138                 /* level 0 is always calculated with 1:1 split */
2139                 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
2140                         if (is_sprite)
2141                                 fifo_size *= 5;
2142                         fifo_size /= 6;
2143                 } else {
2144                         fifo_size /= 2;
2145                 }
2146         }
2147
2148         /* clamp to max that the registers can hold */
2149         return min(fifo_size, ilk_plane_wm_reg_max(dev_priv, level, is_sprite));
2150 }
2151
2152 /* Calculate the maximum cursor plane watermark */
2153 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
2154                                       int level,
2155                                       const struct intel_wm_config *config)
2156 {
2157         /* HSW LP1+ watermarks w/ multiple pipes */
2158         if (level > 0 && config->num_pipes_active > 1)
2159                 return 64;
2160
2161         /* otherwise just report max that registers can hold */
2162         return ilk_cursor_wm_reg_max(to_i915(dev), level);
2163 }
2164
2165 static void ilk_compute_wm_maximums(const struct drm_device *dev,
2166                                     int level,
2167                                     const struct intel_wm_config *config,
2168                                     enum intel_ddb_partitioning ddb_partitioning,
2169                                     struct ilk_wm_maximums *max)
2170 {
2171         max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
2172         max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
2173         max->cur = ilk_cursor_wm_max(dev, level, config);
2174         max->fbc = ilk_fbc_wm_reg_max(to_i915(dev));
2175 }
2176
2177 static void ilk_compute_wm_reg_maximums(const struct drm_i915_private *dev_priv,
2178                                         int level,
2179                                         struct ilk_wm_maximums *max)
2180 {
2181         max->pri = ilk_plane_wm_reg_max(dev_priv, level, false);
2182         max->spr = ilk_plane_wm_reg_max(dev_priv, level, true);
2183         max->cur = ilk_cursor_wm_reg_max(dev_priv, level);
2184         max->fbc = ilk_fbc_wm_reg_max(dev_priv);
2185 }
2186
2187 static bool ilk_validate_wm_level(int level,
2188                                   const struct ilk_wm_maximums *max,
2189                                   struct intel_wm_level *result)
2190 {
2191         bool ret;
2192
2193         /* already determined to be invalid? */
2194         if (!result->enable)
2195                 return false;
2196
2197         result->enable = result->pri_val <= max->pri &&
2198                          result->spr_val <= max->spr &&
2199                          result->cur_val <= max->cur;
2200
2201         ret = result->enable;
2202
2203         /*
2204          * HACK until we can pre-compute everything,
2205          * and thus fail gracefully if LP0 watermarks
2206          * are exceeded...
2207          */
2208         if (level == 0 && !result->enable) {
2209                 if (result->pri_val > max->pri)
2210                         DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
2211                                       level, result->pri_val, max->pri);
2212                 if (result->spr_val > max->spr)
2213                         DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
2214                                       level, result->spr_val, max->spr);
2215                 if (result->cur_val > max->cur)
2216                         DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
2217                                       level, result->cur_val, max->cur);
2218
2219                 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
2220                 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
2221                 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
2222                 result->enable = true;
2223         }
2224
2225         return ret;
2226 }
2227
2228 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
2229                                  const struct intel_crtc *intel_crtc,
2230                                  int level,
2231                                  struct intel_crtc_state *cstate,
2232                                  struct intel_plane_state *pristate,
2233                                  struct intel_plane_state *sprstate,
2234                                  struct intel_plane_state *curstate,
2235                                  struct intel_wm_level *result)
2236 {
2237         uint16_t pri_latency = dev_priv->wm.pri_latency[level];
2238         uint16_t spr_latency = dev_priv->wm.spr_latency[level];
2239         uint16_t cur_latency = dev_priv->wm.cur_latency[level];
2240
2241         /* WM1+ latency values stored in 0.5us units */
2242         if (level > 0) {
2243                 pri_latency *= 5;
2244                 spr_latency *= 5;
2245                 cur_latency *= 5;
2246         }
2247
2248         if (pristate) {
2249                 result->pri_val = ilk_compute_pri_wm(cstate, pristate,
2250                                                      pri_latency, level);
2251                 result->fbc_val = ilk_compute_fbc_wm(cstate, pristate, result->pri_val);
2252         }
2253
2254         if (sprstate)
2255                 result->spr_val = ilk_compute_spr_wm(cstate, sprstate, spr_latency);
2256
2257         if (curstate)
2258                 result->cur_val = ilk_compute_cur_wm(cstate, curstate, cur_latency);
2259
2260         result->enable = true;
2261 }
2262
2263 static uint32_t
2264 hsw_compute_linetime_wm(const struct intel_crtc_state *cstate)
2265 {
2266         const struct intel_atomic_state *intel_state =
2267                 to_intel_atomic_state(cstate->base.state);
2268         const struct drm_display_mode *adjusted_mode =
2269                 &cstate->base.adjusted_mode;
2270         u32 linetime, ips_linetime;
2271
2272         if (!cstate->base.active)
2273                 return 0;
2274         if (WARN_ON(adjusted_mode->crtc_clock == 0))
2275                 return 0;
2276         if (WARN_ON(intel_state->cdclk.logical.cdclk == 0))
2277                 return 0;
2278
2279         /* The WM are computed with base on how long it takes to fill a single
2280          * row at the given clock rate, multiplied by 8.
2281          * */
2282         linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2283                                      adjusted_mode->crtc_clock);
2284         ips_linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2285                                          intel_state->cdclk.logical.cdclk);
2286
2287         return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2288                PIPE_WM_LINETIME_TIME(linetime);
2289 }
2290
2291 static void intel_read_wm_latency(struct drm_i915_private *dev_priv,
2292                                   uint16_t wm[8])
2293 {
2294         if (IS_GEN9(dev_priv)) {
2295                 uint32_t val;
2296                 int ret, i;
2297                 int level, max_level = ilk_wm_max_level(dev_priv);
2298
2299                 /* read the first set of memory latencies[0:3] */
2300                 val = 0; /* data0 to be programmed to 0 for first set */
2301                 mutex_lock(&dev_priv->rps.hw_lock);
2302                 ret = sandybridge_pcode_read(dev_priv,
2303                                              GEN9_PCODE_READ_MEM_LATENCY,
2304                                              &val);
2305                 mutex_unlock(&dev_priv->rps.hw_lock);
2306
2307                 if (ret) {
2308                         DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2309                         return;
2310                 }
2311
2312                 wm[0] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2313                 wm[1] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2314                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2315                 wm[2] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2316                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2317                 wm[3] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2318                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2319
2320                 /* read the second set of memory latencies[4:7] */
2321                 val = 1; /* data0 to be programmed to 1 for second set */
2322                 mutex_lock(&dev_priv->rps.hw_lock);
2323                 ret = sandybridge_pcode_read(dev_priv,
2324                                              GEN9_PCODE_READ_MEM_LATENCY,
2325                                              &val);
2326                 mutex_unlock(&dev_priv->rps.hw_lock);
2327                 if (ret) {
2328                         DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2329                         return;
2330                 }
2331
2332                 wm[4] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2333                 wm[5] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2334                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2335                 wm[6] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2336                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2337                 wm[7] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2338                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2339
2340                 /*
2341                  * If a level n (n > 1) has a 0us latency, all levels m (m >= n)
2342                  * need to be disabled. We make sure to sanitize the values out
2343                  * of the punit to satisfy this requirement.
2344                  */
2345                 for (level = 1; level <= max_level; level++) {
2346                         if (wm[level] == 0) {
2347                                 for (i = level + 1; i <= max_level; i++)
2348                                         wm[i] = 0;
2349                                 break;
2350                         }
2351                 }
2352
2353                 /*
2354                  * WaWmMemoryReadLatency:skl,glk
2355                  *
2356                  * punit doesn't take into account the read latency so we need
2357                  * to add 2us to the various latency levels we retrieve from the
2358                  * punit when level 0 response data us 0us.
2359                  */
2360                 if (wm[0] == 0) {
2361                         wm[0] += 2;
2362                         for (level = 1; level <= max_level; level++) {
2363                                 if (wm[level] == 0)
2364                                         break;
2365                                 wm[level] += 2;
2366                         }
2367                 }
2368
2369         } else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
2370                 uint64_t sskpd = I915_READ64(MCH_SSKPD);
2371
2372                 wm[0] = (sskpd >> 56) & 0xFF;
2373                 if (wm[0] == 0)
2374                         wm[0] = sskpd & 0xF;
2375                 wm[1] = (sskpd >> 4) & 0xFF;
2376                 wm[2] = (sskpd >> 12) & 0xFF;
2377                 wm[3] = (sskpd >> 20) & 0x1FF;
2378                 wm[4] = (sskpd >> 32) & 0x1FF;
2379         } else if (INTEL_GEN(dev_priv) >= 6) {
2380                 uint32_t sskpd = I915_READ(MCH_SSKPD);
2381
2382                 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2383                 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2384                 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2385                 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2386         } else if (INTEL_GEN(dev_priv) >= 5) {
2387                 uint32_t mltr = I915_READ(MLTR_ILK);
2388
2389                 /* ILK primary LP0 latency is 700 ns */
2390                 wm[0] = 7;
2391                 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2392                 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2393         }
2394 }
2395
2396 static void intel_fixup_spr_wm_latency(struct drm_i915_private *dev_priv,
2397                                        uint16_t wm[5])
2398 {
2399         /* ILK sprite LP0 latency is 1300 ns */
2400         if (IS_GEN5(dev_priv))
2401                 wm[0] = 13;
2402 }
2403
2404 static void intel_fixup_cur_wm_latency(struct drm_i915_private *dev_priv,
2405                                        uint16_t wm[5])
2406 {
2407         /* ILK cursor LP0 latency is 1300 ns */
2408         if (IS_GEN5(dev_priv))
2409                 wm[0] = 13;
2410
2411         /* WaDoubleCursorLP3Latency:ivb */
2412         if (IS_IVYBRIDGE(dev_priv))
2413                 wm[3] *= 2;
2414 }
2415
2416 int ilk_wm_max_level(const struct drm_i915_private *dev_priv)
2417 {
2418         /* how many WM levels are we expecting */
2419         if (INTEL_GEN(dev_priv) >= 9)
2420                 return 7;
2421         else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
2422                 return 4;
2423         else if (INTEL_GEN(dev_priv) >= 6)
2424                 return 3;
2425         else
2426                 return 2;
2427 }
2428
2429 static void intel_print_wm_latency(struct drm_i915_private *dev_priv,
2430                                    const char *name,
2431                                    const uint16_t wm[8])
2432 {
2433         int level, max_level = ilk_wm_max_level(dev_priv);
2434
2435         for (level = 0; level <= max_level; level++) {
2436                 unsigned int latency = wm[level];
2437
2438                 if (latency == 0) {
2439                         DRM_ERROR("%s WM%d latency not provided\n",
2440                                   name, level);
2441                         continue;
2442                 }
2443
2444                 /*
2445                  * - latencies are in us on gen9.
2446                  * - before then, WM1+ latency values are in 0.5us units
2447                  */
2448                 if (IS_GEN9(dev_priv))
2449                         latency *= 10;
2450                 else if (level > 0)
2451                         latency *= 5;
2452
2453                 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2454                               name, level, wm[level],
2455                               latency / 10, latency % 10);
2456         }
2457 }
2458
2459 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
2460                                     uint16_t wm[5], uint16_t min)
2461 {
2462         int level, max_level = ilk_wm_max_level(dev_priv);
2463
2464         if (wm[0] >= min)
2465                 return false;
2466
2467         wm[0] = max(wm[0], min);
2468         for (level = 1; level <= max_level; level++)
2469                 wm[level] = max_t(uint16_t, wm[level], DIV_ROUND_UP(min, 5));
2470
2471         return true;
2472 }
2473
2474 static void snb_wm_latency_quirk(struct drm_i915_private *dev_priv)
2475 {
2476         bool changed;
2477
2478         /*
2479          * The BIOS provided WM memory latency values are often
2480          * inadequate for high resolution displays. Adjust them.
2481          */
2482         changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
2483                 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
2484                 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
2485
2486         if (!changed)
2487                 return;
2488
2489         DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2490         intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
2491         intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
2492         intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
2493 }
2494
2495 static void ilk_setup_wm_latency(struct drm_i915_private *dev_priv)
2496 {
2497         intel_read_wm_latency(dev_priv, dev_priv->wm.pri_latency);
2498
2499         memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2500                sizeof(dev_priv->wm.pri_latency));
2501         memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2502                sizeof(dev_priv->wm.pri_latency));
2503
2504         intel_fixup_spr_wm_latency(dev_priv, dev_priv->wm.spr_latency);
2505         intel_fixup_cur_wm_latency(dev_priv, dev_priv->wm.cur_latency);
2506
2507         intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
2508         intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
2509         intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
2510
2511         if (IS_GEN6(dev_priv))
2512                 snb_wm_latency_quirk(dev_priv);
2513 }
2514
2515 static void skl_setup_wm_latency(struct drm_i915_private *dev_priv)
2516 {
2517         intel_read_wm_latency(dev_priv, dev_priv->wm.skl_latency);
2518         intel_print_wm_latency(dev_priv, "Gen9 Plane", dev_priv->wm.skl_latency);
2519 }
2520
2521 static bool ilk_validate_pipe_wm(struct drm_device *dev,
2522                                  struct intel_pipe_wm *pipe_wm)
2523 {
2524         /* LP0 watermark maximums depend on this pipe alone */
2525         const struct intel_wm_config config = {
2526                 .num_pipes_active = 1,
2527                 .sprites_enabled = pipe_wm->sprites_enabled,
2528                 .sprites_scaled = pipe_wm->sprites_scaled,
2529         };
2530         struct ilk_wm_maximums max;
2531
2532         /* LP0 watermarks always use 1/2 DDB partitioning */
2533         ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
2534
2535         /* At least LP0 must be valid */
2536         if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0])) {
2537                 DRM_DEBUG_KMS("LP0 watermark invalid\n");
2538                 return false;
2539         }
2540
2541         return true;
2542 }
2543
2544 /* Compute new watermarks for the pipe */
2545 static int ilk_compute_pipe_wm(struct intel_crtc_state *cstate)
2546 {
2547         struct drm_atomic_state *state = cstate->base.state;
2548         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
2549         struct intel_pipe_wm *pipe_wm;
2550         struct drm_device *dev = state->dev;
2551         const struct drm_i915_private *dev_priv = to_i915(dev);
2552         struct intel_plane *intel_plane;
2553         struct intel_plane_state *pristate = NULL;
2554         struct intel_plane_state *sprstate = NULL;
2555         struct intel_plane_state *curstate = NULL;
2556         int level, max_level = ilk_wm_max_level(dev_priv), usable_level;
2557         struct ilk_wm_maximums max;
2558
2559         pipe_wm = &cstate->wm.ilk.optimal;
2560
2561         for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
2562                 struct intel_plane_state *ps;
2563
2564                 ps = intel_atomic_get_existing_plane_state(state,
2565                                                            intel_plane);
2566                 if (!ps)
2567                         continue;
2568
2569                 if (intel_plane->base.type == DRM_PLANE_TYPE_PRIMARY)
2570                         pristate = ps;
2571                 else if (intel_plane->base.type == DRM_PLANE_TYPE_OVERLAY)
2572                         sprstate = ps;
2573                 else if (intel_plane->base.type == DRM_PLANE_TYPE_CURSOR)
2574                         curstate = ps;
2575         }
2576
2577         pipe_wm->pipe_enabled = cstate->base.active;
2578         if (sprstate) {
2579                 pipe_wm->sprites_enabled = sprstate->base.visible;
2580                 pipe_wm->sprites_scaled = sprstate->base.visible &&
2581                         (drm_rect_width(&sprstate->base.dst) != drm_rect_width(&sprstate->base.src) >> 16 ||
2582                          drm_rect_height(&sprstate->base.dst) != drm_rect_height(&sprstate->base.src) >> 16);
2583         }
2584
2585         usable_level = max_level;
2586
2587         /* ILK/SNB: LP2+ watermarks only w/o sprites */
2588         if (INTEL_GEN(dev_priv) <= 6 && pipe_wm->sprites_enabled)
2589                 usable_level = 1;
2590
2591         /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2592         if (pipe_wm->sprites_scaled)
2593                 usable_level = 0;
2594
2595         ilk_compute_wm_level(dev_priv, intel_crtc, 0, cstate,
2596                              pristate, sprstate, curstate, &pipe_wm->raw_wm[0]);
2597
2598         memset(&pipe_wm->wm, 0, sizeof(pipe_wm->wm));
2599         pipe_wm->wm[0] = pipe_wm->raw_wm[0];
2600
2601         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
2602                 pipe_wm->linetime = hsw_compute_linetime_wm(cstate);
2603
2604         if (!ilk_validate_pipe_wm(dev, pipe_wm))
2605                 return -EINVAL;
2606
2607         ilk_compute_wm_reg_maximums(dev_priv, 1, &max);
2608
2609         for (level = 1; level <= max_level; level++) {
2610                 struct intel_wm_level *wm = &pipe_wm->raw_wm[level];
2611
2612                 ilk_compute_wm_level(dev_priv, intel_crtc, level, cstate,
2613                                      pristate, sprstate, curstate, wm);
2614
2615                 /*
2616                  * Disable any watermark level that exceeds the
2617                  * register maximums since such watermarks are
2618                  * always invalid.
2619                  */
2620                 if (level > usable_level)
2621                         continue;
2622
2623                 if (ilk_validate_wm_level(level, &max, wm))
2624                         pipe_wm->wm[level] = *wm;
2625                 else
2626                         usable_level = level;
2627         }
2628
2629         return 0;
2630 }
2631
2632 /*
2633  * Build a set of 'intermediate' watermark values that satisfy both the old
2634  * state and the new state.  These can be programmed to the hardware
2635  * immediately.
2636  */
2637 static int ilk_compute_intermediate_wm(struct drm_device *dev,
2638                                        struct intel_crtc *intel_crtc,
2639                                        struct intel_crtc_state *newstate)
2640 {
2641         struct intel_pipe_wm *a = &newstate->wm.ilk.intermediate;
2642         struct intel_pipe_wm *b = &intel_crtc->wm.active.ilk;
2643         int level, max_level = ilk_wm_max_level(to_i915(dev));
2644
2645         /*
2646          * Start with the final, target watermarks, then combine with the
2647          * currently active watermarks to get values that are safe both before
2648          * and after the vblank.
2649          */
2650         *a = newstate->wm.ilk.optimal;
2651         a->pipe_enabled |= b->pipe_enabled;
2652         a->sprites_enabled |= b->sprites_enabled;
2653         a->sprites_scaled |= b->sprites_scaled;
2654
2655         for (level = 0; level <= max_level; level++) {
2656                 struct intel_wm_level *a_wm = &a->wm[level];
2657                 const struct intel_wm_level *b_wm = &b->wm[level];
2658
2659                 a_wm->enable &= b_wm->enable;
2660                 a_wm->pri_val = max(a_wm->pri_val, b_wm->pri_val);
2661                 a_wm->spr_val = max(a_wm->spr_val, b_wm->spr_val);
2662                 a_wm->cur_val = max(a_wm->cur_val, b_wm->cur_val);
2663                 a_wm->fbc_val = max(a_wm->fbc_val, b_wm->fbc_val);
2664         }
2665
2666         /*
2667          * We need to make sure that these merged watermark values are
2668          * actually a valid configuration themselves.  If they're not,
2669          * there's no safe way to transition from the old state to
2670          * the new state, so we need to fail the atomic transaction.
2671          */
2672         if (!ilk_validate_pipe_wm(dev, a))
2673                 return -EINVAL;
2674
2675         /*
2676          * If our intermediate WM are identical to the final WM, then we can
2677          * omit the post-vblank programming; only update if it's different.
2678          */
2679         if (memcmp(a, &newstate->wm.ilk.optimal, sizeof(*a)) != 0)
2680                 newstate->wm.need_postvbl_update = true;
2681
2682         return 0;
2683 }
2684
2685 /*
2686  * Merge the watermarks from all active pipes for a specific level.
2687  */
2688 static void ilk_merge_wm_level(struct drm_device *dev,
2689                                int level,
2690                                struct intel_wm_level *ret_wm)
2691 {
2692         const struct intel_crtc *intel_crtc;
2693
2694         ret_wm->enable = true;
2695
2696         for_each_intel_crtc(dev, intel_crtc) {
2697                 const struct intel_pipe_wm *active = &intel_crtc->wm.active.ilk;
2698                 const struct intel_wm_level *wm = &active->wm[level];
2699
2700                 if (!active->pipe_enabled)
2701                         continue;
2702
2703                 /*
2704                  * The watermark values may have been used in the past,
2705                  * so we must maintain them in the registers for some
2706                  * time even if the level is now disabled.
2707                  */
2708                 if (!wm->enable)
2709                         ret_wm->enable = false;
2710
2711                 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
2712                 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
2713                 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
2714                 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
2715         }
2716 }
2717
2718 /*
2719  * Merge all low power watermarks for all active pipes.
2720  */
2721 static void ilk_wm_merge(struct drm_device *dev,
2722                          const struct intel_wm_config *config,
2723                          const struct ilk_wm_maximums *max,
2724                          struct intel_pipe_wm *merged)
2725 {
2726         struct drm_i915_private *dev_priv = to_i915(dev);
2727         int level, max_level = ilk_wm_max_level(dev_priv);
2728         int last_enabled_level = max_level;
2729
2730         /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2731         if ((INTEL_GEN(dev_priv) <= 6 || IS_IVYBRIDGE(dev_priv)) &&
2732             config->num_pipes_active > 1)
2733                 last_enabled_level = 0;
2734
2735         /* ILK: FBC WM must be disabled always */
2736         merged->fbc_wm_enabled = INTEL_GEN(dev_priv) >= 6;
2737
2738         /* merge each WM1+ level */
2739         for (level = 1; level <= max_level; level++) {
2740                 struct intel_wm_level *wm = &merged->wm[level];
2741
2742                 ilk_merge_wm_level(dev, level, wm);
2743
2744                 if (level > last_enabled_level)
2745                         wm->enable = false;
2746                 else if (!ilk_validate_wm_level(level, max, wm))
2747                         /* make sure all following levels get disabled */
2748                         last_enabled_level = level - 1;
2749
2750                 /*
2751                  * The spec says it is preferred to disable
2752                  * FBC WMs instead of disabling a WM level.
2753                  */
2754                 if (wm->fbc_val > max->fbc) {
2755                         if (wm->enable)
2756                                 merged->fbc_wm_enabled = false;
2757                         wm->fbc_val = 0;
2758                 }
2759         }
2760
2761         /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2762         /*
2763          * FIXME this is racy. FBC might get enabled later.
2764          * What we should check here is whether FBC can be
2765          * enabled sometime later.
2766          */
2767         if (IS_GEN5(dev_priv) && !merged->fbc_wm_enabled &&
2768             intel_fbc_is_active(dev_priv)) {
2769                 for (level = 2; level <= max_level; level++) {
2770                         struct intel_wm_level *wm = &merged->wm[level];
2771
2772                         wm->enable = false;
2773                 }
2774         }
2775 }
2776
2777 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
2778 {
2779         /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2780         return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
2781 }
2782
2783 /* The value we need to program into the WM_LPx latency field */
2784 static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
2785 {
2786         struct drm_i915_private *dev_priv = to_i915(dev);
2787
2788         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
2789                 return 2 * level;
2790         else
2791                 return dev_priv->wm.pri_latency[level];
2792 }
2793
2794 static void ilk_compute_wm_results(struct drm_device *dev,
2795                                    const struct intel_pipe_wm *merged,
2796                                    enum intel_ddb_partitioning partitioning,
2797                                    struct ilk_wm_values *results)
2798 {
2799         struct drm_i915_private *dev_priv = to_i915(dev);
2800         struct intel_crtc *intel_crtc;
2801         int level, wm_lp;
2802
2803         results->enable_fbc_wm = merged->fbc_wm_enabled;
2804         results->partitioning = partitioning;
2805
2806         /* LP1+ register values */
2807         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2808                 const struct intel_wm_level *r;
2809
2810                 level = ilk_wm_lp_to_level(wm_lp, merged);
2811
2812                 r = &merged->wm[level];
2813
2814                 /*
2815                  * Maintain the watermark values even if the level is
2816                  * disabled. Doing otherwise could cause underruns.
2817                  */
2818                 results->wm_lp[wm_lp - 1] =
2819                         (ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
2820                         (r->pri_val << WM1_LP_SR_SHIFT) |
2821                         r->cur_val;
2822
2823                 if (r->enable)
2824                         results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
2825
2826                 if (INTEL_GEN(dev_priv) >= 8)
2827                         results->wm_lp[wm_lp - 1] |=
2828                                 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
2829                 else
2830                         results->wm_lp[wm_lp - 1] |=
2831                                 r->fbc_val << WM1_LP_FBC_SHIFT;
2832
2833                 /*
2834                  * Always set WM1S_LP_EN when spr_val != 0, even if the
2835                  * level is disabled. Doing otherwise could cause underruns.
2836                  */
2837                 if (INTEL_GEN(dev_priv) <= 6 && r->spr_val) {
2838                         WARN_ON(wm_lp != 1);
2839                         results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
2840                 } else
2841                         results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2842         }
2843
2844         /* LP0 register values */
2845         for_each_intel_crtc(dev, intel_crtc) {
2846                 enum pipe pipe = intel_crtc->pipe;
2847                 const struct intel_wm_level *r =
2848                         &intel_crtc->wm.active.ilk.wm[0];
2849
2850                 if (WARN_ON(!r->enable))
2851                         continue;
2852
2853                 results->wm_linetime[pipe] = intel_crtc->wm.active.ilk.linetime;
2854
2855                 results->wm_pipe[pipe] =
2856                         (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
2857                         (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
2858                         r->cur_val;
2859         }
2860 }
2861
2862 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2863  * case both are at the same level. Prefer r1 in case they're the same. */
2864 static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
2865                                                   struct intel_pipe_wm *r1,
2866                                                   struct intel_pipe_wm *r2)
2867 {
2868         int level, max_level = ilk_wm_max_level(to_i915(dev));
2869         int level1 = 0, level2 = 0;
2870
2871         for (level = 1; level <= max_level; level++) {
2872                 if (r1->wm[level].enable)
2873                         level1 = level;
2874                 if (r2->wm[level].enable)
2875                         level2 = level;
2876         }
2877
2878         if (level1 == level2) {
2879                 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
2880                         return r2;
2881                 else
2882                         return r1;
2883         } else if (level1 > level2) {
2884                 return r1;
2885         } else {
2886                 return r2;
2887         }
2888 }
2889
2890 /* dirty bits used to track which watermarks need changes */
2891 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2892 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2893 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2894 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2895 #define WM_DIRTY_FBC (1 << 24)
2896 #define WM_DIRTY_DDB (1 << 25)
2897
2898 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private *dev_priv,
2899                                          const struct ilk_wm_values *old,
2900                                          const struct ilk_wm_values *new)
2901 {
2902         unsigned int dirty = 0;
2903         enum pipe pipe;
2904         int wm_lp;
2905
2906         for_each_pipe(dev_priv, pipe) {
2907                 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
2908                         dirty |= WM_DIRTY_LINETIME(pipe);
2909                         /* Must disable LP1+ watermarks too */
2910                         dirty |= WM_DIRTY_LP_ALL;
2911                 }
2912
2913                 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
2914                         dirty |= WM_DIRTY_PIPE(pipe);
2915                         /* Must disable LP1+ watermarks too */
2916                         dirty |= WM_DIRTY_LP_ALL;
2917                 }
2918         }
2919
2920         if (old->enable_fbc_wm != new->enable_fbc_wm) {
2921                 dirty |= WM_DIRTY_FBC;
2922                 /* Must disable LP1+ watermarks too */
2923                 dirty |= WM_DIRTY_LP_ALL;
2924         }
2925
2926         if (old->partitioning != new->partitioning) {
2927                 dirty |= WM_DIRTY_DDB;
2928                 /* Must disable LP1+ watermarks too */
2929                 dirty |= WM_DIRTY_LP_ALL;
2930         }
2931
2932         /* LP1+ watermarks already deemed dirty, no need to continue */
2933         if (dirty & WM_DIRTY_LP_ALL)
2934                 return dirty;
2935
2936         /* Find the lowest numbered LP1+ watermark in need of an update... */
2937         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2938                 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
2939                     old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
2940                         break;
2941         }
2942
2943         /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2944         for (; wm_lp <= 3; wm_lp++)
2945                 dirty |= WM_DIRTY_LP(wm_lp);
2946
2947         return dirty;
2948 }
2949
2950 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
2951                                unsigned int dirty)
2952 {
2953         struct ilk_wm_values *previous = &dev_priv->wm.hw;
2954         bool changed = false;
2955
2956         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
2957                 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
2958                 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
2959                 changed = true;
2960         }
2961         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
2962                 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
2963                 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
2964                 changed = true;
2965         }
2966         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
2967                 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
2968                 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
2969                 changed = true;
2970         }
2971
2972         /*
2973          * Don't touch WM1S_LP_EN here.
2974          * Doing so could cause underruns.
2975          */
2976
2977         return changed;
2978 }
2979
2980 /*
2981  * The spec says we shouldn't write when we don't need, because every write
2982  * causes WMs to be re-evaluated, expending some power.
2983  */
2984 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
2985                                 struct ilk_wm_values *results)
2986 {
2987         struct ilk_wm_values *previous = &dev_priv->wm.hw;
2988         unsigned int dirty;
2989         uint32_t val;
2990
2991         dirty = ilk_compute_wm_dirty(dev_priv, previous, results);
2992         if (!dirty)
2993                 return;
2994
2995         _ilk_disable_lp_wm(dev_priv, dirty);
2996
2997         if (dirty & WM_DIRTY_PIPE(PIPE_A))
2998                 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2999         if (dirty & WM_DIRTY_PIPE(PIPE_B))
3000                 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
3001         if (dirty & WM_DIRTY_PIPE(PIPE_C))
3002                 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
3003
3004         if (dirty & WM_DIRTY_LINETIME(PIPE_A))
3005                 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
3006         if (dirty & WM_DIRTY_LINETIME(PIPE_B))
3007                 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
3008         if (dirty & WM_DIRTY_LINETIME(PIPE_C))
3009                 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
3010
3011         if (dirty & WM_DIRTY_DDB) {
3012                 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
3013                         val = I915_READ(WM_MISC);
3014                         if (results->partitioning == INTEL_DDB_PART_1_2)
3015                                 val &= ~WM_MISC_DATA_PARTITION_5_6;
3016                         else
3017                                 val |= WM_MISC_DATA_PARTITION_5_6;
3018                         I915_WRITE(WM_MISC, val);
3019                 } else {
3020                         val = I915_READ(DISP_ARB_CTL2);
3021                         if (results->partitioning == INTEL_DDB_PART_1_2)
3022                                 val &= ~DISP_DATA_PARTITION_5_6;
3023                         else
3024                                 val |= DISP_DATA_PARTITION_5_6;
3025                         I915_WRITE(DISP_ARB_CTL2, val);
3026                 }
3027         }
3028
3029         if (dirty & WM_DIRTY_FBC) {
3030                 val = I915_READ(DISP_ARB_CTL);
3031                 if (results->enable_fbc_wm)
3032                         val &= ~DISP_FBC_WM_DIS;
3033                 else
3034                         val |= DISP_FBC_WM_DIS;
3035                 I915_WRITE(DISP_ARB_CTL, val);
3036         }
3037
3038         if (dirty & WM_DIRTY_LP(1) &&
3039             previous->wm_lp_spr[0] != results->wm_lp_spr[0])
3040                 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
3041
3042         if (INTEL_GEN(dev_priv) >= 7) {
3043                 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
3044                         I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
3045                 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
3046                         I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
3047         }
3048
3049         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
3050                 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
3051         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
3052                 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
3053         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
3054                 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
3055
3056         dev_priv->wm.hw = *results;
3057 }
3058
3059 bool ilk_disable_lp_wm(struct drm_device *dev)
3060 {
3061         struct drm_i915_private *dev_priv = to_i915(dev);
3062
3063         return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
3064 }
3065
3066 #define SKL_SAGV_BLOCK_TIME     30 /* µs */
3067
3068 /*
3069  * FIXME: We still don't have the proper code detect if we need to apply the WA,
3070  * so assume we'll always need it in order to avoid underruns.
3071  */
3072 static bool skl_needs_memory_bw_wa(struct intel_atomic_state *state)
3073 {
3074         struct drm_i915_private *dev_priv = to_i915(state->base.dev);
3075
3076         if (IS_GEN9_BC(dev_priv) || IS_BROXTON(dev_priv))
3077                 return true;
3078
3079         return false;
3080 }
3081
3082 static bool
3083 intel_has_sagv(struct drm_i915_private *dev_priv)
3084 {
3085         if (IS_KABYLAKE(dev_priv))
3086                 return true;
3087
3088         if (IS_SKYLAKE(dev_priv) &&
3089             dev_priv->sagv_status != I915_SAGV_NOT_CONTROLLED)
3090                 return true;
3091
3092         return false;
3093 }
3094
3095 /*
3096  * SAGV dynamically adjusts the system agent voltage and clock frequencies
3097  * depending on power and performance requirements. The display engine access
3098  * to system memory is blocked during the adjustment time. Because of the
3099  * blocking time, having this enabled can cause full system hangs and/or pipe
3100  * underruns if we don't meet all of the following requirements:
3101  *
3102  *  - <= 1 pipe enabled
3103  *  - All planes can enable watermarks for latencies >= SAGV engine block time
3104  *  - We're not using an interlaced display configuration
3105  */
3106 int
3107 intel_enable_sagv(struct drm_i915_private *dev_priv)
3108 {
3109         int ret;
3110
3111         if (!intel_has_sagv(dev_priv))
3112                 return 0;
3113
3114         if (dev_priv->sagv_status == I915_SAGV_ENABLED)
3115                 return 0;
3116
3117         DRM_DEBUG_KMS("Enabling the SAGV\n");
3118         mutex_lock(&dev_priv->rps.hw_lock);
3119
3120         ret = sandybridge_pcode_write(dev_priv, GEN9_PCODE_SAGV_CONTROL,
3121                                       GEN9_SAGV_ENABLE);
3122
3123         /* We don't need to wait for the SAGV when enabling */
3124         mutex_unlock(&dev_priv->rps.hw_lock);
3125
3126         /*
3127          * Some skl systems, pre-release machines in particular,
3128          * don't actually have an SAGV.
3129          */
3130         if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
3131                 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
3132                 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
3133                 return 0;
3134         } else if (ret < 0) {
3135                 DRM_ERROR("Failed to enable the SAGV\n");
3136                 return ret;
3137         }
3138
3139         dev_priv->sagv_status = I915_SAGV_ENABLED;
3140         return 0;
3141 }
3142
3143 int
3144 intel_disable_sagv(struct drm_i915_private *dev_priv)
3145 {
3146         int ret;
3147
3148         if (!intel_has_sagv(dev_priv))
3149                 return 0;
3150
3151         if (dev_priv->sagv_status == I915_SAGV_DISABLED)
3152                 return 0;
3153
3154         DRM_DEBUG_KMS("Disabling the SAGV\n");
3155         mutex_lock(&dev_priv->rps.hw_lock);
3156
3157         /* bspec says to keep retrying for at least 1 ms */
3158         ret = skl_pcode_request(dev_priv, GEN9_PCODE_SAGV_CONTROL,
3159                                 GEN9_SAGV_DISABLE,
3160                                 GEN9_SAGV_IS_DISABLED, GEN9_SAGV_IS_DISABLED,
3161                                 1);
3162         mutex_unlock(&dev_priv->rps.hw_lock);
3163
3164         /*
3165          * Some skl systems, pre-release machines in particular,
3166          * don't actually have an SAGV.
3167          */
3168         if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
3169                 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
3170                 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
3171                 return 0;
3172         } else if (ret < 0) {
3173                 DRM_ERROR("Failed to disable the SAGV (%d)\n", ret);
3174                 return ret;
3175         }
3176
3177         dev_priv->sagv_status = I915_SAGV_DISABLED;
3178         return 0;
3179 }
3180
3181 bool intel_can_enable_sagv(struct drm_atomic_state *state)
3182 {
3183         struct drm_device *dev = state->dev;
3184         struct drm_i915_private *dev_priv = to_i915(dev);
3185         struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3186         struct intel_crtc *crtc;
3187         struct intel_plane *plane;
3188         struct intel_crtc_state *cstate;
3189         enum pipe pipe;
3190         int level, latency;
3191
3192         if (!intel_has_sagv(dev_priv))
3193                 return false;
3194
3195         /*
3196          * SKL workaround: bspec recommends we disable the SAGV when we have
3197          * more then one pipe enabled
3198          *
3199          * If there are no active CRTCs, no additional checks need be performed
3200          */
3201         if (hweight32(intel_state->active_crtcs) == 0)
3202                 return true;
3203         else if (hweight32(intel_state->active_crtcs) > 1)
3204                 return false;
3205
3206         /* Since we're now guaranteed to only have one active CRTC... */
3207         pipe = ffs(intel_state->active_crtcs) - 1;
3208         crtc = intel_get_crtc_for_pipe(dev_priv, pipe);
3209         cstate = to_intel_crtc_state(crtc->base.state);
3210
3211         if (crtc->base.state->adjusted_mode.flags & DRM_MODE_FLAG_INTERLACE)
3212                 return false;
3213
3214         for_each_intel_plane_on_crtc(dev, crtc, plane) {
3215                 struct skl_plane_wm *wm =
3216                         &cstate->wm.skl.optimal.planes[plane->id];
3217
3218                 /* Skip this plane if it's not enabled */
3219                 if (!wm->wm[0].plane_en)
3220                         continue;
3221
3222                 /* Find the highest enabled wm level for this plane */
3223                 for (level = ilk_wm_max_level(dev_priv);
3224                      !wm->wm[level].plane_en; --level)
3225                      { }
3226
3227                 latency = dev_priv->wm.skl_latency[level];
3228
3229                 if (skl_needs_memory_bw_wa(intel_state) &&
3230                     plane->base.state->fb->modifier ==
3231                     I915_FORMAT_MOD_X_TILED)
3232                         latency += 15;
3233
3234                 /*
3235                  * If any of the planes on this pipe don't enable wm levels
3236                  * that incur memory latencies higher then 30µs we can't enable
3237                  * the SAGV
3238                  */
3239                 if (latency < SKL_SAGV_BLOCK_TIME)
3240                         return false;
3241         }
3242
3243         return true;
3244 }
3245
3246 static void
3247 skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
3248                                    const struct intel_crtc_state *cstate,
3249                                    struct skl_ddb_entry *alloc, /* out */
3250                                    int *num_active /* out */)
3251 {
3252         struct drm_atomic_state *state = cstate->base.state;
3253         struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3254         struct drm_i915_private *dev_priv = to_i915(dev);
3255         struct drm_crtc *for_crtc = cstate->base.crtc;
3256         unsigned int pipe_size, ddb_size;
3257         int nth_active_pipe;
3258
3259         if (WARN_ON(!state) || !cstate->base.active) {
3260                 alloc->start = 0;
3261                 alloc->end = 0;
3262                 *num_active = hweight32(dev_priv->active_crtcs);
3263                 return;
3264         }
3265
3266         if (intel_state->active_pipe_changes)
3267                 *num_active = hweight32(intel_state->active_crtcs);
3268         else
3269                 *num_active = hweight32(dev_priv->active_crtcs);
3270
3271         ddb_size = INTEL_INFO(dev_priv)->ddb_size;
3272         WARN_ON(ddb_size == 0);
3273
3274         ddb_size -= 4; /* 4 blocks for bypass path allocation */
3275
3276         /*
3277          * If the state doesn't change the active CRTC's, then there's
3278          * no need to recalculate; the existing pipe allocation limits
3279          * should remain unchanged.  Note that we're safe from racing
3280          * commits since any racing commit that changes the active CRTC
3281          * list would need to grab _all_ crtc locks, including the one
3282          * we currently hold.
3283          */
3284         if (!intel_state->active_pipe_changes) {
3285                 /*
3286                  * alloc may be cleared by clear_intel_crtc_state,
3287                  * copy from old state to be sure
3288                  */
3289                 *alloc = to_intel_crtc_state(for_crtc->state)->wm.skl.ddb;
3290                 return;
3291         }
3292
3293         nth_active_pipe = hweight32(intel_state->active_crtcs &
3294                                     (drm_crtc_mask(for_crtc) - 1));
3295         pipe_size = ddb_size / hweight32(intel_state->active_crtcs);
3296         alloc->start = nth_active_pipe * ddb_size / *num_active;
3297         alloc->end = alloc->start + pipe_size;
3298 }
3299
3300 static unsigned int skl_cursor_allocation(int num_active)
3301 {
3302         if (num_active == 1)
3303                 return 32;
3304
3305         return 8;
3306 }
3307
3308 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry *entry, u32 reg)
3309 {
3310         entry->start = reg & 0x3ff;
3311         entry->end = (reg >> 16) & 0x3ff;
3312         if (entry->end)
3313                 entry->end += 1;
3314 }
3315
3316 void skl_ddb_get_hw_state(struct drm_i915_private *dev_priv,
3317                           struct skl_ddb_allocation *ddb /* out */)
3318 {
3319         struct intel_crtc *crtc;
3320
3321         memset(ddb, 0, sizeof(*ddb));
3322
3323         for_each_intel_crtc(&dev_priv->drm, crtc) {
3324                 enum intel_display_power_domain power_domain;
3325                 enum plane_id plane_id;
3326                 enum pipe pipe = crtc->pipe;
3327
3328                 power_domain = POWER_DOMAIN_PIPE(pipe);
3329                 if (!intel_display_power_get_if_enabled(dev_priv, power_domain))
3330                         continue;
3331
3332                 for_each_plane_id_on_crtc(crtc, plane_id) {
3333                         u32 val;
3334
3335                         if (plane_id != PLANE_CURSOR)
3336                                 val = I915_READ(PLANE_BUF_CFG(pipe, plane_id));
3337                         else
3338                                 val = I915_READ(CUR_BUF_CFG(pipe));
3339
3340                         skl_ddb_entry_init_from_hw(&ddb->plane[pipe][plane_id], val);
3341                 }
3342
3343                 intel_display_power_put(dev_priv, power_domain);
3344         }
3345 }
3346
3347 /*
3348  * Determines the downscale amount of a plane for the purposes of watermark calculations.
3349  * The bspec defines downscale amount as:
3350  *
3351  * """
3352  * Horizontal down scale amount = maximum[1, Horizontal source size /
3353  *                                           Horizontal destination size]
3354  * Vertical down scale amount = maximum[1, Vertical source size /
3355  *                                         Vertical destination size]
3356  * Total down scale amount = Horizontal down scale amount *
3357  *                           Vertical down scale amount
3358  * """
3359  *
3360  * Return value is provided in 16.16 fixed point form to retain fractional part.
3361  * Caller should take care of dividing & rounding off the value.
3362  */
3363 static uint32_t
3364 skl_plane_downscale_amount(const struct intel_crtc_state *cstate,
3365                            const struct intel_plane_state *pstate)
3366 {
3367         struct intel_plane *plane = to_intel_plane(pstate->base.plane);
3368         uint32_t downscale_h, downscale_w;
3369         uint32_t src_w, src_h, dst_w, dst_h;
3370
3371         if (WARN_ON(!intel_wm_plane_visible(cstate, pstate)))
3372                 return DRM_PLANE_HELPER_NO_SCALING;
3373
3374         /* n.b., src is 16.16 fixed point, dst is whole integer */
3375         if (plane->id == PLANE_CURSOR) {
3376                 /*
3377                  * Cursors only support 0/180 degree rotation,
3378                  * hence no need to account for rotation here.
3379                  */
3380                 src_w = pstate->base.src_w;
3381                 src_h = pstate->base.src_h;
3382                 dst_w = pstate->base.crtc_w;
3383                 dst_h = pstate->base.crtc_h;
3384         } else {
3385                 /*
3386                  * Src coordinates are already rotated by 270 degrees for
3387                  * the 90/270 degree plane rotation cases (to match the
3388                  * GTT mapping), hence no need to account for rotation here.
3389                  */
3390                 src_w = drm_rect_width(&pstate->base.src);
3391                 src_h = drm_rect_height(&pstate->base.src);
3392                 dst_w = drm_rect_width(&pstate->base.dst);
3393                 dst_h = drm_rect_height(&pstate->base.dst);
3394         }
3395
3396         downscale_h = max(src_h / dst_h, (uint32_t)DRM_PLANE_HELPER_NO_SCALING);
3397         downscale_w = max(src_w / dst_w, (uint32_t)DRM_PLANE_HELPER_NO_SCALING);
3398
3399         /* Provide result in 16.16 fixed point */
3400         return (uint64_t)downscale_w * downscale_h >> 16;
3401 }
3402
3403 static unsigned int
3404 skl_plane_relative_data_rate(const struct intel_crtc_state *cstate,
3405                              const struct drm_plane_state *pstate,
3406                              int y)
3407 {
3408         struct intel_plane *plane = to_intel_plane(pstate->plane);
3409         struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
3410         uint32_t down_scale_amount, data_rate;
3411         uint32_t width = 0, height = 0;
3412         struct drm_framebuffer *fb;
3413         u32 format;
3414
3415         if (!intel_pstate->base.visible)
3416                 return 0;
3417
3418         fb = pstate->fb;
3419         format = fb->format->format;
3420
3421         if (plane->id == PLANE_CURSOR)
3422                 return 0;
3423         if (y && format != DRM_FORMAT_NV12)
3424                 return 0;
3425
3426         /*
3427          * Src coordinates are already rotated by 270 degrees for
3428          * the 90/270 degree plane rotation cases (to match the
3429          * GTT mapping), hence no need to account for rotation here.
3430          */
3431         width = drm_rect_width(&intel_pstate->base.src) >> 16;
3432         height = drm_rect_height(&intel_pstate->base.src) >> 16;
3433
3434         /* for planar format */
3435         if (format == DRM_FORMAT_NV12) {
3436                 if (y)  /* y-plane data rate */
3437                         data_rate = width * height *
3438                                 fb->format->cpp[0];
3439                 else    /* uv-plane data rate */
3440                         data_rate = (width / 2) * (height / 2) *
3441                                 fb->format->cpp[1];
3442         } else {
3443                 /* for packed formats */
3444                 data_rate = width * height * fb->format->cpp[0];
3445         }
3446
3447         down_scale_amount = skl_plane_downscale_amount(cstate, intel_pstate);
3448
3449         return (uint64_t)data_rate * down_scale_amount >> 16;
3450 }
3451
3452 /*
3453  * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
3454  * a 8192x4096@32bpp framebuffer:
3455  *   3 * 4096 * 8192  * 4 < 2^32
3456  */
3457 static unsigned int
3458 skl_get_total_relative_data_rate(struct intel_crtc_state *intel_cstate,
3459                                  unsigned *plane_data_rate,
3460                                  unsigned *plane_y_data_rate)
3461 {
3462         struct drm_crtc_state *cstate = &intel_cstate->base;
3463         struct drm_atomic_state *state = cstate->state;
3464         struct drm_plane *plane;
3465         const struct drm_plane_state *pstate;
3466         unsigned int total_data_rate = 0;
3467
3468         if (WARN_ON(!state))
3469                 return 0;
3470
3471         /* Calculate and cache data rate for each plane */
3472         drm_atomic_crtc_state_for_each_plane_state(plane, pstate, cstate) {
3473                 enum plane_id plane_id = to_intel_plane(plane)->id;
3474                 unsigned int rate;
3475
3476                 /* packed/uv */
3477                 rate = skl_plane_relative_data_rate(intel_cstate,
3478                                                     pstate, 0);
3479                 plane_data_rate[plane_id] = rate;
3480
3481                 total_data_rate += rate;
3482
3483                 /* y-plane */
3484                 rate = skl_plane_relative_data_rate(intel_cstate,
3485                                                     pstate, 1);
3486                 plane_y_data_rate[plane_id] = rate;
3487
3488                 total_data_rate += rate;
3489         }
3490
3491         return total_data_rate;
3492 }
3493
3494 static uint16_t
3495 skl_ddb_min_alloc(const struct drm_plane_state *pstate,
3496                   const int y)
3497 {
3498         struct drm_framebuffer *fb = pstate->fb;
3499         struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
3500         uint32_t src_w, src_h;
3501         uint32_t min_scanlines = 8;
3502         uint8_t plane_bpp;
3503
3504         if (WARN_ON(!fb))
3505                 return 0;
3506
3507         /* For packed formats, no y-plane, return 0 */
3508         if (y && fb->format->format != DRM_FORMAT_NV12)
3509                 return 0;
3510
3511         /* For Non Y-tile return 8-blocks */
3512         if (fb->modifier != I915_FORMAT_MOD_Y_TILED &&
3513             fb->modifier != I915_FORMAT_MOD_Yf_TILED)
3514                 return 8;
3515
3516         /*
3517          * Src coordinates are already rotated by 270 degrees for
3518          * the 90/270 degree plane rotation cases (to match the
3519          * GTT mapping), hence no need to account for rotation here.
3520          */
3521         src_w = drm_rect_width(&intel_pstate->base.src) >> 16;
3522         src_h = drm_rect_height(&intel_pstate->base.src) >> 16;
3523
3524         /* Halve UV plane width and height for NV12 */
3525         if (fb->format->format == DRM_FORMAT_NV12 && !y) {
3526                 src_w /= 2;
3527                 src_h /= 2;
3528         }
3529
3530         if (fb->format->format == DRM_FORMAT_NV12 && !y)
3531                 plane_bpp = fb->format->cpp[1];
3532         else
3533                 plane_bpp = fb->format->cpp[0];
3534
3535         if (drm_rotation_90_or_270(pstate->rotation)) {
3536                 switch (plane_bpp) {
3537                 case 1:
3538                         min_scanlines = 32;
3539                         break;
3540                 case 2:
3541                         min_scanlines = 16;
3542                         break;
3543                 case 4:
3544                         min_scanlines = 8;
3545                         break;
3546                 case 8:
3547                         min_scanlines = 4;
3548                         break;
3549                 default:
3550                         WARN(1, "Unsupported pixel depth %u for rotation",
3551                              plane_bpp);
3552                         min_scanlines = 32;
3553                 }
3554         }
3555
3556         return DIV_ROUND_UP((4 * src_w * plane_bpp), 512) * min_scanlines/4 + 3;
3557 }
3558
3559 static void
3560 skl_ddb_calc_min(const struct intel_crtc_state *cstate, int num_active,
3561                  uint16_t *minimum, uint16_t *y_minimum)
3562 {
3563         const struct drm_plane_state *pstate;
3564         struct drm_plane *plane;
3565
3566         drm_atomic_crtc_state_for_each_plane_state(plane, pstate, &cstate->base) {
3567                 enum plane_id plane_id = to_intel_plane(plane)->id;
3568
3569                 if (plane_id == PLANE_CURSOR)
3570                         continue;
3571
3572                 if (!pstate->visible)
3573                         continue;
3574
3575                 minimum[plane_id] = skl_ddb_min_alloc(pstate, 0);
3576                 y_minimum[plane_id] = skl_ddb_min_alloc(pstate, 1);
3577         }
3578
3579         minimum[PLANE_CURSOR] = skl_cursor_allocation(num_active);
3580 }
3581
3582 static int
3583 skl_allocate_pipe_ddb(struct intel_crtc_state *cstate,
3584                       struct skl_ddb_allocation *ddb /* out */)
3585 {
3586         struct drm_atomic_state *state = cstate->base.state;
3587         struct drm_crtc *crtc = cstate->base.crtc;
3588         struct drm_device *dev = crtc->dev;
3589         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3590         enum pipe pipe = intel_crtc->pipe;
3591         struct skl_ddb_entry *alloc = &cstate->wm.skl.ddb;
3592         uint16_t alloc_size, start;
3593         uint16_t minimum[I915_MAX_PLANES] = {};
3594         uint16_t y_minimum[I915_MAX_PLANES] = {};
3595         unsigned int total_data_rate;
3596         enum plane_id plane_id;
3597         int num_active;
3598         unsigned plane_data_rate[I915_MAX_PLANES] = {};
3599         unsigned plane_y_data_rate[I915_MAX_PLANES] = {};
3600
3601         /* Clear the partitioning for disabled planes. */
3602         memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
3603         memset(ddb->y_plane[pipe], 0, sizeof(ddb->y_plane[pipe]));
3604
3605         if (WARN_ON(!state))
3606                 return 0;
3607
3608         if (!cstate->base.active) {
3609                 alloc->start = alloc->end = 0;
3610                 return 0;
3611         }
3612
3613         skl_ddb_get_pipe_allocation_limits(dev, cstate, alloc, &num_active);
3614         alloc_size = skl_ddb_entry_size(alloc);
3615         if (alloc_size == 0) {
3616                 memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
3617                 return 0;
3618         }
3619
3620         skl_ddb_calc_min(cstate, num_active, minimum, y_minimum);
3621
3622         /*
3623          * 1. Allocate the mininum required blocks for each active plane
3624          * and allocate the cursor, it doesn't require extra allocation
3625          * proportional to the data rate.
3626          */
3627
3628         for_each_plane_id_on_crtc(intel_crtc, plane_id) {
3629                 alloc_size -= minimum[plane_id];
3630                 alloc_size -= y_minimum[plane_id];
3631         }
3632
3633         ddb->plane[pipe][PLANE_CURSOR].start = alloc->end - minimum[PLANE_CURSOR];
3634         ddb->plane[pipe][PLANE_CURSOR].end = alloc->end;
3635
3636         /*
3637          * 2. Distribute the remaining space in proportion to the amount of
3638          * data each plane needs to fetch from memory.
3639          *
3640          * FIXME: we may not allocate every single block here.
3641          */
3642         total_data_rate = skl_get_total_relative_data_rate(cstate,
3643                                                            plane_data_rate,
3644                                                            plane_y_data_rate);
3645         if (total_data_rate == 0)
3646                 return 0;
3647
3648         start = alloc->start;
3649         for_each_plane_id_on_crtc(intel_crtc, plane_id) {
3650                 unsigned int data_rate, y_data_rate;
3651                 uint16_t plane_blocks, y_plane_blocks = 0;
3652
3653                 if (plane_id == PLANE_CURSOR)
3654                         continue;
3655
3656                 data_rate = plane_data_rate[plane_id];
3657
3658                 /*
3659                  * allocation for (packed formats) or (uv-plane part of planar format):
3660                  * promote the expression to 64 bits to avoid overflowing, the
3661                  * result is < available as data_rate / total_data_rate < 1
3662                  */
3663                 plane_blocks = minimum[plane_id];
3664                 plane_blocks += div_u64((uint64_t)alloc_size * data_rate,
3665                                         total_data_rate);
3666
3667                 /* Leave disabled planes at (0,0) */
3668                 if (data_rate) {
3669                         ddb->plane[pipe][plane_id].start = start;
3670                         ddb->plane[pipe][plane_id].end = start + plane_blocks;
3671                 }
3672
3673                 start += plane_blocks;
3674
3675                 /*
3676                  * allocation for y_plane part of planar format:
3677                  */
3678                 y_data_rate = plane_y_data_rate[plane_id];
3679
3680                 y_plane_blocks = y_minimum[plane_id];
3681                 y_plane_blocks += div_u64((uint64_t)alloc_size * y_data_rate,
3682                                         total_data_rate);
3683
3684                 if (y_data_rate) {
3685                         ddb->y_plane[pipe][plane_id].start = start;
3686                         ddb->y_plane[pipe][plane_id].end = start + y_plane_blocks;
3687                 }
3688
3689                 start += y_plane_blocks;
3690         }
3691
3692         return 0;
3693 }
3694
3695 /*
3696  * The max latency should be 257 (max the punit can code is 255 and we add 2us
3697  * for the read latency) and cpp should always be <= 8, so that
3698  * should allow pixel_rate up to ~2 GHz which seems sufficient since max
3699  * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
3700 */
3701 static uint_fixed_16_16_t skl_wm_method1(uint32_t pixel_rate, uint8_t cpp,
3702                                          uint32_t latency)
3703 {
3704         uint32_t wm_intermediate_val;
3705         uint_fixed_16_16_t ret;
3706
3707         if (latency == 0)
3708                 return FP_16_16_MAX;
3709
3710         wm_intermediate_val = latency * pixel_rate * cpp;
3711         ret = fixed_16_16_div_round_up_u64(wm_intermediate_val, 1000 * 512);
3712         return ret;
3713 }
3714
3715 static uint_fixed_16_16_t skl_wm_method2(uint32_t pixel_rate,
3716                         uint32_t pipe_htotal,
3717                         uint32_t latency,
3718                         uint_fixed_16_16_t plane_blocks_per_line)
3719 {
3720         uint32_t wm_intermediate_val;
3721         uint_fixed_16_16_t ret;
3722
3723         if (latency == 0)
3724                 return FP_16_16_MAX;
3725
3726         wm_intermediate_val = latency * pixel_rate;
3727         wm_intermediate_val = DIV_ROUND_UP(wm_intermediate_val,
3728                                            pipe_htotal * 1000);
3729         ret = mul_u32_fixed_16_16(wm_intermediate_val, plane_blocks_per_line);
3730         return ret;
3731 }
3732
3733 static uint32_t skl_adjusted_plane_pixel_rate(const struct intel_crtc_state *cstate,
3734                                               struct intel_plane_state *pstate)
3735 {
3736         uint64_t adjusted_pixel_rate;
3737         uint64_t downscale_amount;
3738         uint64_t pixel_rate;
3739
3740         /* Shouldn't reach here on disabled planes... */
3741         if (WARN_ON(!intel_wm_plane_visible(cstate, pstate)))
3742                 return 0;
3743
3744         /*
3745          * Adjusted plane pixel rate is just the pipe's adjusted pixel rate
3746          * with additional adjustments for plane-specific scaling.
3747          */
3748         adjusted_pixel_rate = cstate->pixel_rate;
3749         downscale_amount = skl_plane_downscale_amount(cstate, pstate);
3750
3751         pixel_rate = adjusted_pixel_rate * downscale_amount >> 16;
3752         WARN_ON(pixel_rate != clamp_t(uint32_t, pixel_rate, 0, ~0));
3753
3754         return pixel_rate;
3755 }
3756
3757 static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv,
3758                                 struct intel_crtc_state *cstate,
3759                                 struct intel_plane_state *intel_pstate,
3760                                 uint16_t ddb_allocation,
3761                                 int level,
3762                                 uint16_t *out_blocks, /* out */
3763                                 uint8_t *out_lines, /* out */
3764                                 bool *enabled /* out */)
3765 {
3766         struct intel_plane *plane = to_intel_plane(intel_pstate->base.plane);
3767         struct drm_plane_state *pstate = &intel_pstate->base;
3768         struct drm_framebuffer *fb = pstate->fb;
3769         uint32_t latency = dev_priv->wm.skl_latency[level];
3770         uint_fixed_16_16_t method1, method2;
3771         uint_fixed_16_16_t plane_blocks_per_line;
3772         uint_fixed_16_16_t selected_result;
3773         uint32_t interm_pbpl;
3774         uint32_t plane_bytes_per_line;
3775         uint32_t res_blocks, res_lines;
3776         uint8_t cpp;
3777         uint32_t width = 0, height = 0;
3778         uint32_t plane_pixel_rate;
3779         uint_fixed_16_16_t y_tile_minimum;
3780         uint32_t y_min_scanlines;
3781         struct intel_atomic_state *state =
3782                 to_intel_atomic_state(cstate->base.state);
3783         bool apply_memory_bw_wa = skl_needs_memory_bw_wa(state);
3784         bool y_tiled, x_tiled;
3785
3786         if (latency == 0 ||
3787             !intel_wm_plane_visible(cstate, intel_pstate)) {
3788                 *enabled = false;
3789                 return 0;
3790         }
3791
3792         y_tiled = fb->modifier == I915_FORMAT_MOD_Y_TILED ||
3793                   fb->modifier == I915_FORMAT_MOD_Yf_TILED;
3794         x_tiled = fb->modifier == I915_FORMAT_MOD_X_TILED;
3795
3796         /* Display WA #1141: kbl. */
3797         if (IS_KABYLAKE(dev_priv) && dev_priv->ipc_enabled)
3798                 latency += 4;
3799
3800         if (apply_memory_bw_wa && x_tiled)
3801                 latency += 15;
3802
3803         if (plane->id == PLANE_CURSOR) {
3804                 width = intel_pstate->base.crtc_w;
3805                 height = intel_pstate->base.crtc_h;
3806         } else {
3807                 /*
3808                  * Src coordinates are already rotated by 270 degrees for
3809                  * the 90/270 degree plane rotation cases (to match the
3810                  * GTT mapping), hence no need to account for rotation here.
3811                  */
3812                 width = drm_rect_width(&intel_pstate->base.src) >> 16;
3813                 height = drm_rect_height(&intel_pstate->base.src) >> 16;
3814         }
3815
3816         cpp = fb->format->cpp[0];
3817         plane_pixel_rate = skl_adjusted_plane_pixel_rate(cstate, intel_pstate);
3818
3819         if (drm_rotation_90_or_270(pstate->rotation)) {
3820                 int cpp = (fb->format->format == DRM_FORMAT_NV12) ?
3821                         fb->format->cpp[1] :
3822                         fb->format->cpp[0];
3823
3824                 switch (cpp) {
3825                 case 1:
3826                         y_min_scanlines = 16;
3827                         break;
3828                 case 2:
3829                         y_min_scanlines = 8;
3830                         break;
3831                 case 4:
3832                         y_min_scanlines = 4;
3833                         break;
3834                 default:
3835                         MISSING_CASE(cpp);
3836                         return -EINVAL;
3837                 }
3838         } else {
3839                 y_min_scanlines = 4;
3840         }
3841
3842         if (apply_memory_bw_wa)
3843                 y_min_scanlines *= 2;
3844
3845         plane_bytes_per_line = width * cpp;
3846         if (y_tiled) {
3847                 interm_pbpl = DIV_ROUND_UP(plane_bytes_per_line *
3848                                            y_min_scanlines, 512);
3849                 plane_blocks_per_line =
3850                       fixed_16_16_div_round_up(interm_pbpl, y_min_scanlines);
3851         } else if (x_tiled) {
3852                 interm_pbpl = DIV_ROUND_UP(plane_bytes_per_line, 512);
3853                 plane_blocks_per_line = u32_to_fixed_16_16(interm_pbpl);
3854         } else {
3855                 interm_pbpl = DIV_ROUND_UP(plane_bytes_per_line, 512) + 1;
3856                 plane_blocks_per_line = u32_to_fixed_16_16(interm_pbpl);
3857         }
3858
3859         method1 = skl_wm_method1(plane_pixel_rate, cpp, latency);
3860         method2 = skl_wm_method2(plane_pixel_rate,
3861                                  cstate->base.adjusted_mode.crtc_htotal,
3862                                  latency,
3863                                  plane_blocks_per_line);
3864
3865         y_tile_minimum = mul_u32_fixed_16_16(y_min_scanlines,
3866                                              plane_blocks_per_line);
3867
3868         if (y_tiled) {
3869                 selected_result = max_fixed_16_16(method2, y_tile_minimum);
3870         } else {
3871                 if ((cpp * cstate->base.adjusted_mode.crtc_htotal / 512 < 1) &&
3872                     (plane_bytes_per_line / 512 < 1))
3873                         selected_result = method2;
3874                 else if ((ddb_allocation /
3875                         fixed_16_16_to_u32_round_up(plane_blocks_per_line)) >= 1)
3876                         selected_result = min_fixed_16_16(method1, method2);
3877                 else
3878                         selected_result = method1;
3879         }
3880
3881         res_blocks = fixed_16_16_to_u32_round_up(selected_result) + 1;
3882         res_lines = DIV_ROUND_UP(selected_result.val,
3883                                  plane_blocks_per_line.val);
3884
3885         if (level >= 1 && level <= 7) {
3886                 if (y_tiled) {
3887                         res_blocks += fixed_16_16_to_u32_round_up(y_tile_minimum);
3888                         res_lines += y_min_scanlines;
3889                 } else {
3890                         res_blocks++;
3891                 }
3892         }
3893
3894         if (res_blocks >= ddb_allocation || res_lines > 31) {
3895                 *enabled = false;
3896
3897                 /*
3898                  * If there are no valid level 0 watermarks, then we can't
3899                  * support this display configuration.
3900                  */
3901                 if (level) {
3902                         return 0;
3903                 } else {
3904                         struct drm_plane *plane = pstate->plane;
3905
3906                         DRM_DEBUG_KMS("Requested display configuration exceeds system watermark limitations\n");
3907                         DRM_DEBUG_KMS("[PLANE:%d:%s] blocks required = %u/%u, lines required = %u/31\n",
3908                                       plane->base.id, plane->name,
3909                                       res_blocks, ddb_allocation, res_lines);
3910                         return -EINVAL;
3911                 }
3912         }
3913
3914         *out_blocks = res_blocks;
3915         *out_lines = res_lines;
3916         *enabled = true;
3917
3918         return 0;
3919 }
3920
3921 static int
3922 skl_compute_wm_level(const struct drm_i915_private *dev_priv,
3923                      struct skl_ddb_allocation *ddb,
3924                      struct intel_crtc_state *cstate,
3925                      struct intel_plane *intel_plane,
3926                      int level,
3927                      struct skl_wm_level *result)
3928 {
3929         struct drm_atomic_state *state = cstate->base.state;
3930         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
3931         struct drm_plane *plane = &intel_plane->base;
3932         struct intel_plane_state *intel_pstate = NULL;
3933         uint16_t ddb_blocks;
3934         enum pipe pipe = intel_crtc->pipe;
3935         int ret;
3936
3937         if (state)
3938                 intel_pstate =
3939                         intel_atomic_get_existing_plane_state(state,
3940                                                               intel_plane);
3941
3942         /*
3943          * Note: If we start supporting multiple pending atomic commits against
3944          * the same planes/CRTC's in the future, plane->state will no longer be
3945          * the correct pre-state to use for the calculations here and we'll
3946          * need to change where we get the 'unchanged' plane data from.
3947          *
3948          * For now this is fine because we only allow one queued commit against
3949          * a CRTC.  Even if the plane isn't modified by this transaction and we
3950          * don't have a plane lock, we still have the CRTC's lock, so we know
3951          * that no other transactions are racing with us to update it.
3952          */
3953         if (!intel_pstate)
3954                 intel_pstate = to_intel_plane_state(plane->state);
3955
3956         WARN_ON(!intel_pstate->base.fb);
3957
3958         ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][intel_plane->id]);
3959
3960         ret = skl_compute_plane_wm(dev_priv,
3961                                    cstate,
3962                                    intel_pstate,
3963                                    ddb_blocks,
3964                                    level,
3965                                    &result->plane_res_b,
3966                                    &result->plane_res_l,
3967                                    &result->plane_en);
3968         if (ret)
3969                 return ret;
3970
3971         return 0;
3972 }
3973
3974 static uint32_t
3975 skl_compute_linetime_wm(struct intel_crtc_state *cstate)
3976 {
3977         struct drm_atomic_state *state = cstate->base.state;
3978         struct drm_i915_private *dev_priv = to_i915(state->dev);
3979         uint32_t pixel_rate;
3980         uint32_t linetime_wm;
3981
3982         if (!cstate->base.active)
3983                 return 0;
3984
3985         pixel_rate = cstate->pixel_rate;
3986
3987         if (WARN_ON(pixel_rate == 0))
3988                 return 0;
3989
3990         linetime_wm = DIV_ROUND_UP(8 * cstate->base.adjusted_mode.crtc_htotal *
3991                                    1000, pixel_rate);
3992
3993         /* Display WA #1135: bxt. */
3994         if (IS_BROXTON(dev_priv) && dev_priv->ipc_enabled)
3995                 linetime_wm = DIV_ROUND_UP(linetime_wm, 2);
3996
3997         return linetime_wm;
3998 }
3999
4000 static void skl_compute_transition_wm(struct intel_crtc_state *cstate,
4001                                       struct skl_wm_level *trans_wm /* out */)
4002 {
4003         if (!cstate->base.active)
4004                 return;
4005
4006         /* Until we know more, just disable transition WMs */
4007         trans_wm->plane_en = false;
4008 }
4009
4010 static int skl_build_pipe_wm(struct intel_crtc_state *cstate,
4011                              struct skl_ddb_allocation *ddb,
4012                              struct skl_pipe_wm *pipe_wm)
4013 {
4014         struct drm_device *dev = cstate->base.crtc->dev;
4015         const struct drm_i915_private *dev_priv = to_i915(dev);
4016         struct intel_plane *intel_plane;
4017         struct skl_plane_wm *wm;
4018         int level, max_level = ilk_wm_max_level(dev_priv);
4019         int ret;
4020
4021         /*
4022          * We'll only calculate watermarks for planes that are actually
4023          * enabled, so make sure all other planes are set as disabled.
4024          */
4025         memset(pipe_wm->planes, 0, sizeof(pipe_wm->planes));
4026
4027         for_each_intel_plane_mask(&dev_priv->drm,
4028                                   intel_plane,
4029                                   cstate->base.plane_mask) {
4030                 wm = &pipe_wm->planes[intel_plane->id];
4031
4032                 for (level = 0; level <= max_level; level++) {
4033                         ret = skl_compute_wm_level(dev_priv, ddb, cstate,
4034                                                    intel_plane, level,
4035                                                    &wm->wm[level]);
4036                         if (ret)
4037                                 return ret;
4038                 }
4039                 skl_compute_transition_wm(cstate, &wm->trans_wm);
4040         }
4041         pipe_wm->linetime = skl_compute_linetime_wm(cstate);
4042
4043         return 0;
4044 }
4045
4046 static void skl_ddb_entry_write(struct drm_i915_private *dev_priv,
4047                                 i915_reg_t reg,
4048                                 const struct skl_ddb_entry *entry)
4049 {
4050         if (entry->end)
4051                 I915_WRITE(reg, (entry->end - 1) << 16 | entry->start);
4052         else
4053                 I915_WRITE(reg, 0);
4054 }
4055
4056 static void skl_write_wm_level(struct drm_i915_private *dev_priv,
4057                                i915_reg_t reg,
4058                                const struct skl_wm_level *level)
4059 {
4060         uint32_t val = 0;
4061
4062         if (level->plane_en) {
4063                 val |= PLANE_WM_EN;
4064                 val |= level->plane_res_b;
4065                 val |= level->plane_res_l << PLANE_WM_LINES_SHIFT;
4066         }
4067
4068         I915_WRITE(reg, val);
4069 }
4070
4071 static void skl_write_plane_wm(struct intel_crtc *intel_crtc,
4072                                const struct skl_plane_wm *wm,
4073                                const struct skl_ddb_allocation *ddb,
4074                                enum plane_id plane_id)
4075 {
4076         struct drm_crtc *crtc = &intel_crtc->base;
4077         struct drm_device *dev = crtc->dev;
4078         struct drm_i915_private *dev_priv = to_i915(dev);
4079         int level, max_level = ilk_wm_max_level(dev_priv);
4080         enum pipe pipe = intel_crtc->pipe;
4081
4082         for (level = 0; level <= max_level; level++) {
4083                 skl_write_wm_level(dev_priv, PLANE_WM(pipe, plane_id, level),
4084                                    &wm->wm[level]);
4085         }
4086         skl_write_wm_level(dev_priv, PLANE_WM_TRANS(pipe, plane_id),
4087                            &wm->trans_wm);
4088
4089         skl_ddb_entry_write(dev_priv, PLANE_BUF_CFG(pipe, plane_id),
4090                             &ddb->plane[pipe][plane_id]);
4091         skl_ddb_entry_write(dev_priv, PLANE_NV12_BUF_CFG(pipe, plane_id),
4092                             &ddb->y_plane[pipe][plane_id]);
4093 }
4094
4095 static void skl_write_cursor_wm(struct intel_crtc *intel_crtc,
4096                                 const struct skl_plane_wm *wm,
4097                                 const struct skl_ddb_allocation *ddb)
4098 {
4099         struct drm_crtc *crtc = &intel_crtc->base;
4100         struct drm_device *dev = crtc->dev;
4101         struct drm_i915_private *dev_priv = to_i915(dev);
4102         int level, max_level = ilk_wm_max_level(dev_priv);
4103         enum pipe pipe = intel_crtc->pipe;
4104
4105         for (level = 0; level <= max_level; level++) {
4106                 skl_write_wm_level(dev_priv, CUR_WM(pipe, level),
4107                                    &wm->wm[level]);
4108         }
4109         skl_write_wm_level(dev_priv, CUR_WM_TRANS(pipe), &wm->trans_wm);
4110
4111         skl_ddb_entry_write(dev_priv, CUR_BUF_CFG(pipe),
4112                             &ddb->plane[pipe][PLANE_CURSOR]);
4113 }
4114
4115 bool skl_wm_level_equals(const struct skl_wm_level *l1,
4116                          const struct skl_wm_level *l2)
4117 {
4118         if (l1->plane_en != l2->plane_en)
4119                 return false;
4120
4121         /* If both planes aren't enabled, the rest shouldn't matter */
4122         if (!l1->plane_en)
4123                 return true;
4124
4125         return (l1->plane_res_l == l2->plane_res_l &&
4126                 l1->plane_res_b == l2->plane_res_b);
4127 }
4128
4129 static inline bool skl_ddb_entries_overlap(const struct skl_ddb_entry *a,
4130                                            const struct skl_ddb_entry *b)
4131 {
4132         return a->start < b->end && b->start < a->end;
4133 }
4134
4135 bool skl_ddb_allocation_overlaps(const struct skl_ddb_entry **entries,
4136                                  const struct skl_ddb_entry *ddb,
4137                                  int ignore)
4138 {
4139         int i;
4140
4141         for (i = 0; i < I915_MAX_PIPES; i++)
4142                 if (i != ignore && entries[i] &&
4143                     skl_ddb_entries_overlap(ddb, entries[i]))
4144                         return true;
4145
4146         return false;
4147 }
4148
4149 static int skl_update_pipe_wm(struct drm_crtc_state *cstate,
4150                               const struct skl_pipe_wm *old_pipe_wm,
4151                               struct skl_pipe_wm *pipe_wm, /* out */
4152                               struct skl_ddb_allocation *ddb, /* out */
4153                               bool *changed /* out */)
4154 {
4155         struct intel_crtc_state *intel_cstate = to_intel_crtc_state(cstate);
4156         int ret;
4157
4158         ret = skl_build_pipe_wm(intel_cstate, ddb, pipe_wm);
4159         if (ret)
4160                 return ret;
4161
4162         if (!memcmp(old_pipe_wm, pipe_wm, sizeof(*pipe_wm)))
4163                 *changed = false;
4164         else
4165                 *changed = true;
4166
4167         return 0;
4168 }
4169
4170 static uint32_t
4171 pipes_modified(struct drm_atomic_state *state)
4172 {
4173         struct drm_crtc *crtc;
4174         struct drm_crtc_state *cstate;
4175         uint32_t i, ret = 0;
4176
4177         for_each_new_crtc_in_state(state, crtc, cstate, i)
4178                 ret |= drm_crtc_mask(crtc);
4179
4180         return ret;
4181 }
4182
4183 static int
4184 skl_ddb_add_affected_planes(struct intel_crtc_state *cstate)
4185 {
4186         struct drm_atomic_state *state = cstate->base.state;
4187         struct drm_device *dev = state->dev;
4188         struct drm_crtc *crtc = cstate->base.crtc;
4189         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4190         struct drm_i915_private *dev_priv = to_i915(dev);
4191         struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
4192         struct skl_ddb_allocation *new_ddb = &intel_state->wm_results.ddb;
4193         struct skl_ddb_allocation *cur_ddb = &dev_priv->wm.skl_hw.ddb;
4194         struct drm_plane_state *plane_state;
4195         struct drm_plane *plane;
4196         enum pipe pipe = intel_crtc->pipe;
4197
4198         WARN_ON(!drm_atomic_get_existing_crtc_state(state, crtc));
4199
4200         drm_for_each_plane_mask(plane, dev, cstate->base.plane_mask) {
4201                 enum plane_id plane_id = to_intel_plane(plane)->id;
4202
4203                 if (skl_ddb_entry_equal(&cur_ddb->plane[pipe][plane_id],
4204                                         &new_ddb->plane[pipe][plane_id]) &&
4205                     skl_ddb_entry_equal(&cur_ddb->y_plane[pipe][plane_id],
4206                                         &new_ddb->y_plane[pipe][plane_id]))
4207                         continue;
4208
4209                 plane_state = drm_atomic_get_plane_state(state, plane);
4210                 if (IS_ERR(plane_state))
4211                         return PTR_ERR(plane_state);
4212         }
4213
4214         return 0;
4215 }
4216
4217 static int
4218 skl_compute_ddb(struct drm_atomic_state *state)
4219 {
4220         struct drm_device *dev = state->dev;
4221         struct drm_i915_private *dev_priv = to_i915(dev);
4222         struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
4223         struct intel_crtc *intel_crtc;
4224         struct skl_ddb_allocation *ddb = &intel_state->wm_results.ddb;
4225         uint32_t realloc_pipes = pipes_modified(state);
4226         int ret;
4227
4228         /*
4229          * If this is our first atomic update following hardware readout,
4230          * we can't trust the DDB that the BIOS programmed for us.  Let's
4231          * pretend that all pipes switched active status so that we'll
4232          * ensure a full DDB recompute.
4233          */
4234         if (dev_priv->wm.distrust_bios_wm) {
4235                 ret = drm_modeset_lock(&dev->mode_config.connection_mutex,
4236                                        state->acquire_ctx);
4237                 if (ret)
4238                         return ret;
4239
4240                 intel_state->active_pipe_changes = ~0;
4241
4242                 /*
4243                  * We usually only initialize intel_state->active_crtcs if we
4244                  * we're doing a modeset; make sure this field is always
4245                  * initialized during the sanitization process that happens
4246                  * on the first commit too.
4247                  */
4248                 if (!intel_state->modeset)
4249                         intel_state->active_crtcs = dev_priv->active_crtcs;
4250         }
4251
4252         /*
4253          * If the modeset changes which CRTC's are active, we need to
4254          * recompute the DDB allocation for *all* active pipes, even
4255          * those that weren't otherwise being modified in any way by this
4256          * atomic commit.  Due to the shrinking of the per-pipe allocations
4257          * when new active CRTC's are added, it's possible for a pipe that
4258          * we were already using and aren't changing at all here to suddenly
4259          * become invalid if its DDB needs exceeds its new allocation.
4260          *
4261          * Note that if we wind up doing a full DDB recompute, we can't let
4262          * any other display updates race with this transaction, so we need
4263          * to grab the lock on *all* CRTC's.
4264          */
4265         if (intel_state->active_pipe_changes) {
4266                 realloc_pipes = ~0;
4267                 intel_state->wm_results.dirty_pipes = ~0;
4268         }
4269
4270         /*
4271          * We're not recomputing for the pipes not included in the commit, so
4272          * make sure we start with the current state.
4273          */
4274         memcpy(ddb, &dev_priv->wm.skl_hw.ddb, sizeof(*ddb));
4275
4276         for_each_intel_crtc_mask(dev, intel_crtc, realloc_pipes) {
4277                 struct intel_crtc_state *cstate;
4278
4279                 cstate = intel_atomic_get_crtc_state(state, intel_crtc);
4280                 if (IS_ERR(cstate))
4281                         return PTR_ERR(cstate);
4282
4283                 ret = skl_allocate_pipe_ddb(cstate, ddb);
4284                 if (ret)
4285                         return ret;
4286
4287                 ret = skl_ddb_add_affected_planes(cstate);
4288                 if (ret)
4289                         return ret;
4290         }
4291
4292         return 0;
4293 }
4294
4295 static void
4296 skl_copy_wm_for_pipe(struct skl_wm_values *dst,
4297                      struct skl_wm_values *src,
4298                      enum pipe pipe)
4299 {
4300         memcpy(dst->ddb.y_plane[pipe], src->ddb.y_plane[pipe],
4301                sizeof(dst->ddb.y_plane[pipe]));
4302         memcpy(dst->ddb.plane[pipe], src->ddb.plane[pipe],
4303                sizeof(dst->ddb.plane[pipe]));
4304 }
4305
4306 static void
4307 skl_print_wm_changes(const struct drm_atomic_state *state)
4308 {
4309         const struct drm_device *dev = state->dev;
4310         const struct drm_i915_private *dev_priv = to_i915(dev);
4311         const struct intel_atomic_state *intel_state =
4312                 to_intel_atomic_state(state);
4313         const struct drm_crtc *crtc;
4314         const struct drm_crtc_state *cstate;
4315         const struct intel_plane *intel_plane;
4316         const struct skl_ddb_allocation *old_ddb = &dev_priv->wm.skl_hw.ddb;
4317         const struct skl_ddb_allocation *new_ddb = &intel_state->wm_results.ddb;
4318         int i;
4319
4320         for_each_new_crtc_in_state(state, crtc, cstate, i) {
4321                 const struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4322                 enum pipe pipe = intel_crtc->pipe;
4323
4324                 for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
4325                         enum plane_id plane_id = intel_plane->id;
4326                         const struct skl_ddb_entry *old, *new;
4327
4328                         old = &old_ddb->plane[pipe][plane_id];
4329                         new = &new_ddb->plane[pipe][plane_id];
4330
4331                         if (skl_ddb_entry_equal(old, new))
4332                                 continue;
4333
4334                         DRM_DEBUG_ATOMIC("[PLANE:%d:%s] ddb (%d - %d) -> (%d - %d)\n",
4335                                          intel_plane->base.base.id,
4336                                          intel_plane->base.name,
4337                                          old->start, old->end,
4338                                          new->start, new->end);
4339                 }
4340         }
4341 }
4342
4343 static int
4344 skl_compute_wm(struct drm_atomic_state *state)
4345 {
4346         struct drm_crtc *crtc;
4347         struct drm_crtc_state *cstate;
4348         struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
4349         struct skl_wm_values *results = &intel_state->wm_results;
4350         struct drm_device *dev = state->dev;
4351         struct skl_pipe_wm *pipe_wm;
4352         bool changed = false;
4353         int ret, i;
4354
4355         /*
4356          * When we distrust bios wm we always need to recompute to set the
4357          * expected DDB allocations for each CRTC.
4358          */
4359         if (to_i915(dev)->wm.distrust_bios_wm)
4360                 changed = true;
4361
4362         /*
4363          * If this transaction isn't actually touching any CRTC's, don't
4364          * bother with watermark calculation.  Note that if we pass this
4365          * test, we're guaranteed to hold at least one CRTC state mutex,
4366          * which means we can safely use values like dev_priv->active_crtcs
4367          * since any racing commits that want to update them would need to
4368          * hold _all_ CRTC state mutexes.
4369          */
4370         for_each_new_crtc_in_state(state, crtc, cstate, i)
4371                 changed = true;
4372
4373         if (!changed)
4374                 return 0;
4375
4376         /* Clear all dirty flags */
4377         results->dirty_pipes = 0;
4378
4379         ret = skl_compute_ddb(state);
4380         if (ret)
4381                 return ret;
4382
4383         /*
4384          * Calculate WM's for all pipes that are part of this transaction.
4385          * Note that the DDB allocation above may have added more CRTC's that
4386          * weren't otherwise being modified (and set bits in dirty_pipes) if
4387          * pipe allocations had to change.
4388          *
4389          * FIXME:  Now that we're doing this in the atomic check phase, we
4390          * should allow skl_update_pipe_wm() to return failure in cases where
4391          * no suitable watermark values can be found.
4392          */
4393         for_each_new_crtc_in_state(state, crtc, cstate, i) {
4394                 struct intel_crtc_state *intel_cstate =
4395                         to_intel_crtc_state(cstate);
4396                 const struct skl_pipe_wm *old_pipe_wm =
4397                         &to_intel_crtc_state(crtc->state)->wm.skl.optimal;
4398
4399                 pipe_wm = &intel_cstate->wm.skl.optimal;
4400                 ret = skl_update_pipe_wm(cstate, old_pipe_wm, pipe_wm,
4401                                          &results->ddb, &changed);
4402                 if (ret)
4403                         return ret;
4404
4405                 if (changed)
4406                         results->dirty_pipes |= drm_crtc_mask(crtc);
4407
4408                 if ((results->dirty_pipes & drm_crtc_mask(crtc)) == 0)
4409                         /* This pipe's WM's did not change */
4410                         continue;
4411
4412                 intel_cstate->update_wm_pre = true;
4413         }
4414
4415         skl_print_wm_changes(state);
4416
4417         return 0;
4418 }
4419
4420 static void skl_atomic_update_crtc_wm(struct intel_atomic_state *state,
4421                                       struct intel_crtc_state *cstate)
4422 {
4423         struct intel_crtc *crtc = to_intel_crtc(cstate->base.crtc);
4424         struct drm_i915_private *dev_priv = to_i915(state->base.dev);
4425         struct skl_pipe_wm *pipe_wm = &cstate->wm.skl.optimal;
4426         const struct skl_ddb_allocation *ddb = &state->wm_results.ddb;
4427         enum pipe pipe = crtc->pipe;
4428         enum plane_id plane_id;
4429
4430         if (!(state->wm_results.dirty_pipes & drm_crtc_mask(&crtc->base)))
4431                 return;
4432
4433         I915_WRITE(PIPE_WM_LINETIME(pipe), pipe_wm->linetime);
4434
4435         for_each_plane_id_on_crtc(crtc, plane_id) {
4436                 if (plane_id != PLANE_CURSOR)
4437                         skl_write_plane_wm(crtc, &pipe_wm->planes[plane_id],
4438                                            ddb, plane_id);
4439                 else
4440                         skl_write_cursor_wm(crtc, &pipe_wm->planes[plane_id],
4441                                             ddb);
4442         }
4443 }
4444
4445 static void skl_initial_wm(struct intel_atomic_state *state,
4446                            struct intel_crtc_state *cstate)
4447 {
4448         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4449         struct drm_device *dev = intel_crtc->base.dev;
4450         struct drm_i915_private *dev_priv = to_i915(dev);
4451         struct skl_wm_values *results = &state->wm_results;
4452         struct skl_wm_values *hw_vals = &dev_priv->wm.skl_hw;
4453         enum pipe pipe = intel_crtc->pipe;
4454
4455         if ((results->dirty_pipes & drm_crtc_mask(&intel_crtc->base)) == 0)
4456                 return;
4457
4458         mutex_lock(&dev_priv->wm.wm_mutex);
4459
4460         if (cstate->base.active_changed)
4461                 skl_atomic_update_crtc_wm(state, cstate);
4462
4463         skl_copy_wm_for_pipe(hw_vals, results, pipe);
4464
4465         mutex_unlock(&dev_priv->wm.wm_mutex);
4466 }
4467
4468 static void ilk_compute_wm_config(struct drm_device *dev,
4469                                   struct intel_wm_config *config)
4470 {
4471         struct intel_crtc *crtc;
4472
4473         /* Compute the currently _active_ config */
4474         for_each_intel_crtc(dev, crtc) {
4475                 const struct intel_pipe_wm *wm = &crtc->wm.active.ilk;
4476
4477                 if (!wm->pipe_enabled)
4478                         continue;
4479
4480                 config->sprites_enabled |= wm->sprites_enabled;
4481                 config->sprites_scaled |= wm->sprites_scaled;
4482                 config->num_pipes_active++;
4483         }
4484 }
4485
4486 static void ilk_program_watermarks(struct drm_i915_private *dev_priv)
4487 {
4488         struct drm_device *dev = &dev_priv->drm;
4489         struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
4490         struct ilk_wm_maximums max;
4491         struct intel_wm_config config = {};
4492         struct ilk_wm_values results = {};
4493         enum intel_ddb_partitioning partitioning;
4494
4495         ilk_compute_wm_config(dev, &config);
4496
4497         ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
4498         ilk_wm_merge(dev, &config, &max, &lp_wm_1_2);
4499
4500         /* 5/6 split only in single pipe config on IVB+ */
4501         if (INTEL_GEN(dev_priv) >= 7 &&
4502             config.num_pipes_active == 1 && config.sprites_enabled) {
4503                 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
4504                 ilk_wm_merge(dev, &config, &max, &lp_wm_5_6);
4505
4506                 best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
4507         } else {
4508                 best_lp_wm = &lp_wm_1_2;
4509         }
4510
4511         partitioning = (best_lp_wm == &lp_wm_1_2) ?
4512                        INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
4513
4514         ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
4515
4516         ilk_write_wm_values(dev_priv, &results);
4517 }
4518
4519 static void ilk_initial_watermarks(struct intel_atomic_state *state,
4520                                    struct intel_crtc_state *cstate)
4521 {
4522         struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
4523         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4524
4525         mutex_lock(&dev_priv->wm.wm_mutex);
4526         intel_crtc->wm.active.ilk = cstate->wm.ilk.intermediate;
4527         ilk_program_watermarks(dev_priv);
4528         mutex_unlock(&dev_priv->wm.wm_mutex);
4529 }
4530
4531 static void ilk_optimize_watermarks(struct intel_atomic_state *state,
4532                                     struct intel_crtc_state *cstate)
4533 {
4534         struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
4535         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4536
4537         mutex_lock(&dev_priv->wm.wm_mutex);
4538         if (cstate->wm.need_postvbl_update) {
4539                 intel_crtc->wm.active.ilk = cstate->wm.ilk.optimal;
4540                 ilk_program_watermarks(dev_priv);
4541         }
4542         mutex_unlock(&dev_priv->wm.wm_mutex);
4543 }
4544
4545 static inline void skl_wm_level_from_reg_val(uint32_t val,
4546                                              struct skl_wm_level *level)
4547 {
4548         level->plane_en = val & PLANE_WM_EN;
4549         level->plane_res_b = val & PLANE_WM_BLOCKS_MASK;
4550         level->plane_res_l = (val >> PLANE_WM_LINES_SHIFT) &
4551                 PLANE_WM_LINES_MASK;
4552 }
4553
4554 void skl_pipe_wm_get_hw_state(struct drm_crtc *crtc,
4555                               struct skl_pipe_wm *out)
4556 {
4557         struct drm_i915_private *dev_priv = to_i915(crtc->dev);
4558         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4559         enum pipe pipe = intel_crtc->pipe;
4560         int level, max_level;
4561         enum plane_id plane_id;
4562         uint32_t val;
4563
4564         max_level = ilk_wm_max_level(dev_priv);
4565
4566         for_each_plane_id_on_crtc(intel_crtc, plane_id) {
4567                 struct skl_plane_wm *wm = &out->planes[plane_id];
4568
4569                 for (level = 0; level <= max_level; level++) {
4570                         if (plane_id != PLANE_CURSOR)
4571                                 val = I915_READ(PLANE_WM(pipe, plane_id, level));
4572                         else
4573                                 val = I915_READ(CUR_WM(pipe, level));
4574
4575                         skl_wm_level_from_reg_val(val, &wm->wm[level]);
4576                 }
4577
4578                 if (plane_id != PLANE_CURSOR)
4579                         val = I915_READ(PLANE_WM_TRANS(pipe, plane_id));
4580                 else
4581                         val = I915_READ(CUR_WM_TRANS(pipe));
4582
4583                 skl_wm_level_from_reg_val(val, &wm->trans_wm);
4584         }
4585
4586         if (!intel_crtc->active)
4587                 return;
4588
4589         out->linetime = I915_READ(PIPE_WM_LINETIME(pipe));
4590 }
4591
4592 void skl_wm_get_hw_state(struct drm_device *dev)
4593 {
4594         struct drm_i915_private *dev_priv = to_i915(dev);
4595         struct skl_wm_values *hw = &dev_priv->wm.skl_hw;
4596         struct skl_ddb_allocation *ddb = &dev_priv->wm.skl_hw.ddb;
4597         struct drm_crtc *crtc;
4598         struct intel_crtc *intel_crtc;
4599         struct intel_crtc_state *cstate;
4600
4601         skl_ddb_get_hw_state(dev_priv, ddb);
4602         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
4603                 intel_crtc = to_intel_crtc(crtc);
4604                 cstate = to_intel_crtc_state(crtc->state);
4605
4606                 skl_pipe_wm_get_hw_state(crtc, &cstate->wm.skl.optimal);
4607
4608                 if (intel_crtc->active)
4609                         hw->dirty_pipes |= drm_crtc_mask(crtc);
4610         }
4611
4612         if (dev_priv->active_crtcs) {
4613                 /* Fully recompute DDB on first atomic commit */
4614                 dev_priv->wm.distrust_bios_wm = true;
4615         } else {
4616                 /* Easy/common case; just sanitize DDB now if everything off */
4617                 memset(ddb, 0, sizeof(*ddb));
4618         }
4619 }
4620
4621 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
4622 {
4623         struct drm_device *dev = crtc->dev;
4624         struct drm_i915_private *dev_priv = to_i915(dev);
4625         struct ilk_wm_values *hw = &dev_priv->wm.hw;
4626         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4627         struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
4628         struct intel_pipe_wm *active = &cstate->wm.ilk.optimal;
4629         enum pipe pipe = intel_crtc->pipe;
4630         static const i915_reg_t wm0_pipe_reg[] = {
4631                 [PIPE_A] = WM0_PIPEA_ILK,
4632                 [PIPE_B] = WM0_PIPEB_ILK,
4633                 [PIPE_C] = WM0_PIPEC_IVB,
4634         };
4635
4636         hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
4637         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
4638                 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
4639
4640         memset(active, 0, sizeof(*active));
4641
4642         active->pipe_enabled = intel_crtc->active;
4643
4644         if (active->pipe_enabled) {
4645                 u32 tmp = hw->wm_pipe[pipe];
4646
4647                 /*
4648                  * For active pipes LP0 watermark is marked as
4649                  * enabled, and LP1+ watermaks as disabled since
4650                  * we can't really reverse compute them in case
4651                  * multiple pipes are active.
4652                  */
4653                 active->wm[0].enable = true;
4654                 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
4655                 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
4656                 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
4657                 active->linetime = hw->wm_linetime[pipe];
4658         } else {
4659                 int level, max_level = ilk_wm_max_level(dev_priv);
4660
4661                 /*
4662                  * For inactive pipes, all watermark levels
4663                  * should be marked as enabled but zeroed,
4664                  * which is what we'd compute them to.
4665                  */
4666                 for (level = 0; level <= max_level; level++)
4667                         active->wm[level].enable = true;
4668         }
4669
4670         intel_crtc->wm.active.ilk = *active;
4671 }
4672
4673 #define _FW_WM(value, plane) \
4674         (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
4675 #define _FW_WM_VLV(value, plane) \
4676         (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
4677
4678 static void vlv_read_wm_values(struct drm_i915_private *dev_priv,
4679                                struct vlv_wm_values *wm)
4680 {
4681         enum pipe pipe;
4682         uint32_t tmp;
4683
4684         for_each_pipe(dev_priv, pipe) {
4685                 tmp = I915_READ(VLV_DDL(pipe));
4686
4687                 wm->ddl[pipe].plane[PLANE_PRIMARY] =
4688                         (tmp >> DDL_PLANE_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4689                 wm->ddl[pipe].plane[PLANE_CURSOR] =
4690                         (tmp >> DDL_CURSOR_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4691                 wm->ddl[pipe].plane[PLANE_SPRITE0] =
4692                         (tmp >> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4693                 wm->ddl[pipe].plane[PLANE_SPRITE1] =
4694                         (tmp >> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4695         }
4696
4697         tmp = I915_READ(DSPFW1);
4698         wm->sr.plane = _FW_WM(tmp, SR);
4699         wm->pipe[PIPE_B].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORB);
4700         wm->pipe[PIPE_B].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEB);
4701         wm->pipe[PIPE_A].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEA);
4702
4703         tmp = I915_READ(DSPFW2);
4704         wm->pipe[PIPE_A].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITEB);
4705         wm->pipe[PIPE_A].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORA);
4706         wm->pipe[PIPE_A].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEA);
4707
4708         tmp = I915_READ(DSPFW3);
4709         wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
4710
4711         if (IS_CHERRYVIEW(dev_priv)) {
4712                 tmp = I915_READ(DSPFW7_CHV);
4713                 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITED);
4714                 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEC);
4715
4716                 tmp = I915_READ(DSPFW8_CHV);
4717                 wm->pipe[PIPE_C].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITEF);
4718                 wm->pipe[PIPE_C].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEE);
4719
4720                 tmp = I915_READ(DSPFW9_CHV);
4721                 wm->pipe[PIPE_C].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEC);
4722                 wm->pipe[PIPE_C].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORC);
4723
4724                 tmp = I915_READ(DSPHOWM);
4725                 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
4726                 wm->pipe[PIPE_C].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEF_HI) << 8;
4727                 wm->pipe[PIPE_C].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEE_HI) << 8;
4728                 wm->pipe[PIPE_C].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEC_HI) << 8;
4729                 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITED_HI) << 8;
4730                 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
4731                 wm->pipe[PIPE_B].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEB_HI) << 8;
4732                 wm->pipe[PIPE_A].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
4733                 wm->pipe[PIPE_A].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
4734                 wm->pipe[PIPE_A].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEA_HI) << 8;
4735         } else {
4736                 tmp = I915_READ(DSPFW7);
4737                 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITED);
4738                 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEC);
4739
4740                 tmp = I915_READ(DSPHOWM);
4741                 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
4742                 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITED_HI) << 8;
4743                 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
4744                 wm->pipe[PIPE_B].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEB_HI) << 8;
4745                 wm->pipe[PIPE_A].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
4746                 wm->pipe[PIPE_A].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
4747                 wm->pipe[PIPE_A].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEA_HI) << 8;
4748         }
4749 }
4750
4751 #undef _FW_WM
4752 #undef _FW_WM_VLV
4753
4754 void vlv_wm_get_hw_state(struct drm_device *dev)
4755 {
4756         struct drm_i915_private *dev_priv = to_i915(dev);
4757         struct vlv_wm_values *wm = &dev_priv->wm.vlv;
4758         struct intel_crtc *crtc;
4759         u32 val;
4760
4761         vlv_read_wm_values(dev_priv, wm);
4762
4763         wm->cxsr = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
4764         wm->level = VLV_WM_LEVEL_PM2;
4765
4766         if (IS_CHERRYVIEW(dev_priv)) {
4767                 mutex_lock(&dev_priv->rps.hw_lock);
4768
4769                 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
4770                 if (val & DSP_MAXFIFO_PM5_ENABLE)
4771                         wm->level = VLV_WM_LEVEL_PM5;
4772
4773                 /*
4774                  * If DDR DVFS is disabled in the BIOS, Punit
4775                  * will never ack the request. So if that happens
4776                  * assume we don't have to enable/disable DDR DVFS
4777                  * dynamically. To test that just set the REQ_ACK
4778                  * bit to poke the Punit, but don't change the
4779                  * HIGH/LOW bits so that we don't actually change
4780                  * the current state.
4781                  */
4782                 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
4783                 val |= FORCE_DDR_FREQ_REQ_ACK;
4784                 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
4785
4786                 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
4787                               FORCE_DDR_FREQ_REQ_ACK) == 0, 3)) {
4788                         DRM_DEBUG_KMS("Punit not acking DDR DVFS request, "
4789                                       "assuming DDR DVFS is disabled\n");
4790                         dev_priv->wm.max_level = VLV_WM_LEVEL_PM5;
4791                 } else {
4792                         val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
4793                         if ((val & FORCE_DDR_HIGH_FREQ) == 0)
4794                                 wm->level = VLV_WM_LEVEL_DDR_DVFS;
4795                 }
4796
4797                 mutex_unlock(&dev_priv->rps.hw_lock);
4798         }
4799
4800         for_each_intel_crtc(dev, crtc) {
4801                 struct intel_crtc_state *crtc_state =
4802                         to_intel_crtc_state(crtc->base.state);
4803                 struct vlv_wm_state *active = &crtc->wm.active.vlv;
4804                 const struct vlv_fifo_state *fifo_state =
4805                         &crtc_state->wm.vlv.fifo_state;
4806                 enum pipe pipe = crtc->pipe;
4807                 enum plane_id plane_id;
4808                 int level;
4809
4810                 vlv_get_fifo_size(crtc_state);
4811
4812                 active->num_levels = wm->level + 1;
4813                 active->cxsr = wm->cxsr;
4814
4815                 for (level = 0; level < active->num_levels; level++) {
4816                         struct vlv_pipe_wm *raw =
4817                                 &crtc_state->wm.vlv.raw[level];
4818
4819                         active->sr[level].plane = wm->sr.plane;
4820                         active->sr[level].cursor = wm->sr.cursor;
4821
4822                         for_each_plane_id_on_crtc(crtc, plane_id) {
4823                                 active->wm[level].plane[plane_id] =
4824                                         wm->pipe[pipe].plane[plane_id];
4825
4826                                 raw->plane[plane_id] =
4827                                         vlv_invert_wm_value(active->wm[level].plane[plane_id],
4828                                                             fifo_state->plane[plane_id]);
4829                         }
4830                 }
4831
4832                 for_each_plane_id_on_crtc(crtc, plane_id)
4833                         vlv_raw_plane_wm_set(crtc_state, level,
4834                                              plane_id, USHRT_MAX);
4835                 vlv_invalidate_wms(crtc, active, level);
4836
4837                 crtc_state->wm.vlv.optimal = *active;
4838                 crtc_state->wm.vlv.intermediate = *active;
4839
4840                 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
4841                               pipe_name(pipe),
4842                               wm->pipe[pipe].plane[PLANE_PRIMARY],
4843                               wm->pipe[pipe].plane[PLANE_CURSOR],
4844                               wm->pipe[pipe].plane[PLANE_SPRITE0],
4845                               wm->pipe[pipe].plane[PLANE_SPRITE1]);
4846         }
4847
4848         DRM_DEBUG_KMS("Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
4849                       wm->sr.plane, wm->sr.cursor, wm->level, wm->cxsr);
4850 }
4851
4852 void vlv_wm_sanitize(struct drm_i915_private *dev_priv)
4853 {
4854         struct intel_plane *plane;
4855         struct intel_crtc *crtc;
4856
4857         mutex_lock(&dev_priv->wm.wm_mutex);
4858
4859         for_each_intel_plane(&dev_priv->drm, plane) {
4860                 struct intel_crtc *crtc =
4861                         intel_get_crtc_for_pipe(dev_priv, plane->pipe);
4862                 struct intel_crtc_state *crtc_state =
4863                         to_intel_crtc_state(crtc->base.state);
4864                 struct intel_plane_state *plane_state =
4865                         to_intel_plane_state(plane->base.state);
4866                 struct vlv_wm_state *wm_state = &crtc_state->wm.vlv.optimal;
4867                 const struct vlv_fifo_state *fifo_state =
4868                         &crtc_state->wm.vlv.fifo_state;
4869                 enum plane_id plane_id = plane->id;
4870                 int level;
4871
4872                 if (plane_state->base.visible)
4873                         continue;
4874
4875                 for (level = 0; level < wm_state->num_levels; level++) {
4876                         struct vlv_pipe_wm *raw =
4877                                 &crtc_state->wm.vlv.raw[level];
4878
4879                         raw->plane[plane_id] = 0;
4880
4881                         wm_state->wm[level].plane[plane_id] =
4882                                 vlv_invert_wm_value(raw->plane[plane_id],
4883                                                     fifo_state->plane[plane_id]);
4884                 }
4885         }
4886
4887         for_each_intel_crtc(&dev_priv->drm, crtc) {
4888                 struct intel_crtc_state *crtc_state =
4889                         to_intel_crtc_state(crtc->base.state);
4890
4891                 crtc_state->wm.vlv.intermediate =
4892                         crtc_state->wm.vlv.optimal;
4893                 crtc->wm.active.vlv = crtc_state->wm.vlv.optimal;
4894         }
4895
4896         vlv_program_watermarks(dev_priv);
4897
4898         mutex_unlock(&dev_priv->wm.wm_mutex);
4899 }
4900
4901 void ilk_wm_get_hw_state(struct drm_device *dev)
4902 {
4903         struct drm_i915_private *dev_priv = to_i915(dev);
4904         struct ilk_wm_values *hw = &dev_priv->wm.hw;
4905         struct drm_crtc *crtc;
4906
4907         for_each_crtc(dev, crtc)
4908                 ilk_pipe_wm_get_hw_state(crtc);
4909
4910         hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
4911         hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
4912         hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
4913
4914         hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
4915         if (INTEL_GEN(dev_priv) >= 7) {
4916                 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
4917                 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
4918         }
4919
4920         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
4921                 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
4922                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4923         else if (IS_IVYBRIDGE(dev_priv))
4924                 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
4925                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4926
4927         hw->enable_fbc_wm =
4928                 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
4929 }
4930
4931 /**
4932  * intel_update_watermarks - update FIFO watermark values based on current modes
4933  *
4934  * Calculate watermark values for the various WM regs based on current mode
4935  * and plane configuration.
4936  *
4937  * There are several cases to deal with here:
4938  *   - normal (i.e. non-self-refresh)
4939  *   - self-refresh (SR) mode
4940  *   - lines are large relative to FIFO size (buffer can hold up to 2)
4941  *   - lines are small relative to FIFO size (buffer can hold more than 2
4942  *     lines), so need to account for TLB latency
4943  *
4944  *   The normal calculation is:
4945  *     watermark = dotclock * bytes per pixel * latency
4946  *   where latency is platform & configuration dependent (we assume pessimal
4947  *   values here).
4948  *
4949  *   The SR calculation is:
4950  *     watermark = (trunc(latency/line time)+1) * surface width *
4951  *       bytes per pixel
4952  *   where
4953  *     line time = htotal / dotclock
4954  *     surface width = hdisplay for normal plane and 64 for cursor
4955  *   and latency is assumed to be high, as above.
4956  *
4957  * The final value programmed to the register should always be rounded up,
4958  * and include an extra 2 entries to account for clock crossings.
4959  *
4960  * We don't use the sprite, so we can ignore that.  And on Crestline we have
4961  * to set the non-SR watermarks to 8.
4962  */
4963 void intel_update_watermarks(struct intel_crtc *crtc)
4964 {
4965         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
4966
4967         if (dev_priv->display.update_wm)
4968                 dev_priv->display.update_wm(crtc);
4969 }
4970
4971 /*
4972  * Lock protecting IPS related data structures
4973  */
4974 DEFINE_SPINLOCK(mchdev_lock);
4975
4976 /* Global for IPS driver to get at the current i915 device. Protected by
4977  * mchdev_lock. */
4978 static struct drm_i915_private *i915_mch_dev;
4979
4980 bool ironlake_set_drps(struct drm_i915_private *dev_priv, u8 val)
4981 {
4982         u16 rgvswctl;
4983
4984         lockdep_assert_held(&mchdev_lock);
4985
4986         rgvswctl = I915_READ16(MEMSWCTL);
4987         if (rgvswctl & MEMCTL_CMD_STS) {
4988                 DRM_DEBUG("gpu busy, RCS change rejected\n");
4989                 return false; /* still busy with another command */
4990         }
4991
4992         rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
4993                 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
4994         I915_WRITE16(MEMSWCTL, rgvswctl);
4995         POSTING_READ16(MEMSWCTL);
4996
4997         rgvswctl |= MEMCTL_CMD_STS;
4998         I915_WRITE16(MEMSWCTL, rgvswctl);
4999
5000         return true;
5001 }
5002
5003 static void ironlake_enable_drps(struct drm_i915_private *dev_priv)
5004 {
5005         u32 rgvmodectl;
5006         u8 fmax, fmin, fstart, vstart;
5007
5008         spin_lock_irq(&mchdev_lock);
5009
5010         rgvmodectl = I915_READ(MEMMODECTL);
5011
5012         /* Enable temp reporting */
5013         I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
5014         I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
5015
5016         /* 100ms RC evaluation intervals */
5017         I915_WRITE(RCUPEI, 100000);
5018         I915_WRITE(RCDNEI, 100000);
5019
5020         /* Set max/min thresholds to 90ms and 80ms respectively */
5021         I915_WRITE(RCBMAXAVG, 90000);
5022         I915_WRITE(RCBMINAVG, 80000);
5023
5024         I915_WRITE(MEMIHYST, 1);
5025
5026         /* Set up min, max, and cur for interrupt handling */
5027         fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
5028         fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
5029         fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
5030                 MEMMODE_FSTART_SHIFT;
5031
5032         vstart = (I915_READ(PXVFREQ(fstart)) & PXVFREQ_PX_MASK) >>
5033                 PXVFREQ_PX_SHIFT;
5034
5035         dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
5036         dev_priv->ips.fstart = fstart;
5037
5038         dev_priv->ips.max_delay = fstart;
5039         dev_priv->ips.min_delay = fmin;
5040         dev_priv->ips.cur_delay = fstart;
5041
5042         DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
5043                          fmax, fmin, fstart);
5044
5045         I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
5046
5047         /*
5048          * Interrupts will be enabled in ironlake_irq_postinstall
5049          */
5050
5051         I915_WRITE(VIDSTART, vstart);
5052         POSTING_READ(VIDSTART);
5053
5054         rgvmodectl |= MEMMODE_SWMODE_EN;
5055         I915_WRITE(MEMMODECTL, rgvmodectl);
5056
5057         if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
5058                 DRM_ERROR("stuck trying to change perf mode\n");
5059         mdelay(1);
5060
5061         ironlake_set_drps(dev_priv, fstart);
5062
5063         dev_priv->ips.last_count1 = I915_READ(DMIEC) +
5064                 I915_READ(DDREC) + I915_READ(CSIEC);
5065         dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
5066         dev_priv->ips.last_count2 = I915_READ(GFXEC);
5067         dev_priv->ips.last_time2 = ktime_get_raw_ns();
5068
5069         spin_unlock_irq(&mchdev_lock);
5070 }
5071
5072 static void ironlake_disable_drps(struct drm_i915_private *dev_priv)
5073 {
5074         u16 rgvswctl;
5075
5076         spin_lock_irq(&mchdev_lock);
5077
5078         rgvswctl = I915_READ16(MEMSWCTL);
5079
5080         /* Ack interrupts, disable EFC interrupt */
5081         I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
5082         I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
5083         I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
5084         I915_WRITE(DEIIR, DE_PCU_EVENT);
5085         I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
5086
5087         /* Go back to the starting frequency */
5088         ironlake_set_drps(dev_priv, dev_priv->ips.fstart);
5089         mdelay(1);
5090         rgvswctl |= MEMCTL_CMD_STS;
5091         I915_WRITE(MEMSWCTL, rgvswctl);
5092         mdelay(1);
5093
5094         spin_unlock_irq(&mchdev_lock);
5095 }
5096
5097 /* There's a funny hw issue where the hw returns all 0 when reading from
5098  * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
5099  * ourselves, instead of doing a rmw cycle (which might result in us clearing
5100  * all limits and the gpu stuck at whatever frequency it is at atm).
5101  */
5102 static u32 intel_rps_limits(struct drm_i915_private *dev_priv, u8 val)
5103 {
5104         u32 limits;
5105
5106         /* Only set the down limit when we've reached the lowest level to avoid
5107          * getting more interrupts, otherwise leave this clear. This prevents a
5108          * race in the hw when coming out of rc6: There's a tiny window where
5109          * the hw runs at the minimal clock before selecting the desired
5110          * frequency, if the down threshold expires in that window we will not
5111          * receive a down interrupt. */
5112         if (IS_GEN9(dev_priv)) {
5113                 limits = (dev_priv->rps.max_freq_softlimit) << 23;
5114                 if (val <= dev_priv->rps.min_freq_softlimit)
5115                         limits |= (dev_priv->rps.min_freq_softlimit) << 14;
5116         } else {
5117                 limits = dev_priv->rps.max_freq_softlimit << 24;
5118                 if (val <= dev_priv->rps.min_freq_softlimit)
5119                         limits |= dev_priv->rps.min_freq_softlimit << 16;
5120         }
5121
5122         return limits;
5123 }
5124
5125 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
5126 {
5127         int new_power;
5128         u32 threshold_up = 0, threshold_down = 0; /* in % */
5129         u32 ei_up = 0, ei_down = 0;
5130
5131         new_power = dev_priv->rps.power;
5132         switch (dev_priv->rps.power) {
5133         case LOW_POWER:
5134                 if (val > dev_priv->rps.efficient_freq + 1 &&
5135                     val > dev_priv->rps.cur_freq)
5136                         new_power = BETWEEN;
5137                 break;
5138
5139         case BETWEEN:
5140                 if (val <= dev_priv->rps.efficient_freq &&
5141                     val < dev_priv->rps.cur_freq)
5142                         new_power = LOW_POWER;
5143                 else if (val >= dev_priv->rps.rp0_freq &&
5144                          val > dev_priv->rps.cur_freq)
5145                         new_power = HIGH_POWER;
5146                 break;
5147
5148         case HIGH_POWER:
5149                 if (val < (dev_priv->rps.rp1_freq + dev_priv->rps.rp0_freq) >> 1 &&
5150                     val < dev_priv->rps.cur_freq)
5151                         new_power = BETWEEN;
5152                 break;
5153         }
5154         /* Max/min bins are special */
5155         if (val <= dev_priv->rps.min_freq_softlimit)
5156                 new_power = LOW_POWER;
5157         if (val >= dev_priv->rps.max_freq_softlimit)
5158                 new_power = HIGH_POWER;
5159         if (new_power == dev_priv->rps.power)
5160                 return;
5161
5162         /* Note the units here are not exactly 1us, but 1280ns. */
5163         switch (new_power) {
5164         case LOW_POWER:
5165                 /* Upclock if more than 95% busy over 16ms */
5166                 ei_up = 16000;
5167                 threshold_up = 95;
5168
5169                 /* Downclock if less than 85% busy over 32ms */
5170                 ei_down = 32000;
5171                 threshold_down = 85;
5172                 break;
5173
5174         case BETWEEN:
5175                 /* Upclock if more than 90% busy over 13ms */
5176                 ei_up = 13000;
5177                 threshold_up = 90;
5178
5179                 /* Downclock if less than 75% busy over 32ms */
5180                 ei_down = 32000;
5181                 threshold_down = 75;
5182                 break;
5183
5184         case HIGH_POWER:
5185                 /* Upclock if more than 85% busy over 10ms */
5186                 ei_up = 10000;
5187                 threshold_up = 85;
5188
5189                 /* Downclock if less than 60% busy over 32ms */
5190                 ei_down = 32000;
5191                 threshold_down = 60;
5192                 break;
5193         }
5194
5195         /* When byt can survive without system hang with dynamic
5196          * sw freq adjustments, this restriction can be lifted.
5197          */
5198         if (IS_VALLEYVIEW(dev_priv))
5199                 goto skip_hw_write;
5200
5201         I915_WRITE(GEN6_RP_UP_EI,
5202                    GT_INTERVAL_FROM_US(dev_priv, ei_up));
5203         I915_WRITE(GEN6_RP_UP_THRESHOLD,
5204                    GT_INTERVAL_FROM_US(dev_priv,
5205                                        ei_up * threshold_up / 100));
5206
5207         I915_WRITE(GEN6_RP_DOWN_EI,
5208                    GT_INTERVAL_FROM_US(dev_priv, ei_down));
5209         I915_WRITE(GEN6_RP_DOWN_THRESHOLD,
5210                    GT_INTERVAL_FROM_US(dev_priv,
5211                                        ei_down * threshold_down / 100));
5212
5213         I915_WRITE(GEN6_RP_CONTROL,
5214                    GEN6_RP_MEDIA_TURBO |
5215                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
5216                    GEN6_RP_MEDIA_IS_GFX |
5217                    GEN6_RP_ENABLE |
5218                    GEN6_RP_UP_BUSY_AVG |
5219                    GEN6_RP_DOWN_IDLE_AVG);
5220
5221 skip_hw_write:
5222         dev_priv->rps.power = new_power;
5223         dev_priv->rps.up_threshold = threshold_up;
5224         dev_priv->rps.down_threshold = threshold_down;
5225         dev_priv->rps.last_adj = 0;
5226 }
5227
5228 static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
5229 {
5230         u32 mask = 0;
5231
5232         /* We use UP_EI_EXPIRED interupts for both up/down in manual mode */
5233         if (val > dev_priv->rps.min_freq_softlimit)
5234                 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
5235         if (val < dev_priv->rps.max_freq_softlimit)
5236                 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD;
5237
5238         mask &= dev_priv->pm_rps_events;
5239
5240         return gen6_sanitize_rps_pm_mask(dev_priv, ~mask);
5241 }
5242
5243 /* gen6_set_rps is called to update the frequency request, but should also be
5244  * called when the range (min_delay and max_delay) is modified so that we can
5245  * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
5246 static int gen6_set_rps(struct drm_i915_private *dev_priv, u8 val)
5247 {
5248         /* min/max delay may still have been modified so be sure to
5249          * write the limits value.
5250          */
5251         if (val != dev_priv->rps.cur_freq) {
5252                 gen6_set_rps_thresholds(dev_priv, val);
5253
5254                 if (IS_GEN9(dev_priv))
5255                         I915_WRITE(GEN6_RPNSWREQ,
5256                                    GEN9_FREQUENCY(val));
5257                 else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
5258                         I915_WRITE(GEN6_RPNSWREQ,
5259                                    HSW_FREQUENCY(val));
5260                 else
5261                         I915_WRITE(GEN6_RPNSWREQ,
5262                                    GEN6_FREQUENCY(val) |
5263                                    GEN6_OFFSET(0) |
5264                                    GEN6_AGGRESSIVE_TURBO);
5265         }
5266
5267         /* Make sure we continue to get interrupts
5268          * until we hit the minimum or maximum frequencies.
5269          */
5270         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, intel_rps_limits(dev_priv, val));
5271         I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
5272
5273         dev_priv->rps.cur_freq = val;
5274         trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
5275
5276         return 0;
5277 }
5278
5279 static int valleyview_set_rps(struct drm_i915_private *dev_priv, u8 val)
5280 {
5281         int err;
5282
5283         if (WARN_ONCE(IS_CHERRYVIEW(dev_priv) && (val & 1),
5284                       "Odd GPU freq value\n"))
5285                 val &= ~1;
5286
5287         I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
5288
5289         if (val != dev_priv->rps.cur_freq) {
5290                 err = vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
5291                 if (err)
5292                         return err;
5293
5294                 gen6_set_rps_thresholds(dev_priv, val);
5295         }
5296
5297         dev_priv->rps.cur_freq = val;
5298         trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
5299
5300         return 0;
5301 }
5302
5303 /* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
5304  *
5305  * * If Gfx is Idle, then
5306  * 1. Forcewake Media well.
5307  * 2. Request idle freq.
5308  * 3. Release Forcewake of Media well.
5309 */
5310 static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
5311 {
5312         u32 val = dev_priv->rps.idle_freq;
5313         int err;
5314
5315         if (dev_priv->rps.cur_freq <= val)
5316                 return;
5317
5318         /* The punit delays the write of the frequency and voltage until it
5319          * determines the GPU is awake. During normal usage we don't want to
5320          * waste power changing the frequency if the GPU is sleeping (rc6).
5321          * However, the GPU and driver is now idle and we do not want to delay
5322          * switching to minimum voltage (reducing power whilst idle) as we do
5323          * not expect to be woken in the near future and so must flush the
5324          * change by waking the device.
5325          *
5326          * We choose to take the media powerwell (either would do to trick the
5327          * punit into committing the voltage change) as that takes a lot less
5328          * power than the render powerwell.
5329          */
5330         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_MEDIA);
5331         err = valleyview_set_rps(dev_priv, val);
5332         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_MEDIA);
5333
5334         if (err)
5335                 DRM_ERROR("Failed to set RPS for idle\n");
5336 }
5337
5338 void gen6_rps_busy(struct drm_i915_private *dev_priv)
5339 {
5340         mutex_lock(&dev_priv->rps.hw_lock);
5341         if (dev_priv->rps.enabled) {
5342                 u8 freq;
5343
5344                 if (dev_priv->pm_rps_events & GEN6_PM_RP_UP_EI_EXPIRED)
5345                         gen6_rps_reset_ei(dev_priv);
5346                 I915_WRITE(GEN6_PMINTRMSK,
5347                            gen6_rps_pm_mask(dev_priv, dev_priv->rps.cur_freq));
5348
5349                 gen6_enable_rps_interrupts(dev_priv);
5350
5351                 /* Use the user's desired frequency as a guide, but for better
5352                  * performance, jump directly to RPe as our starting frequency.
5353                  */
5354                 freq = max(dev_priv->rps.cur_freq,
5355                            dev_priv->rps.efficient_freq);
5356
5357                 if (intel_set_rps(dev_priv,
5358                                   clamp(freq,
5359                                         dev_priv->rps.min_freq_softlimit,
5360                                         dev_priv->rps.max_freq_softlimit)))
5361                         DRM_DEBUG_DRIVER("Failed to set idle frequency\n");
5362         }
5363         mutex_unlock(&dev_priv->rps.hw_lock);
5364 }
5365
5366 void gen6_rps_idle(struct drm_i915_private *dev_priv)
5367 {
5368         /* Flush our bottom-half so that it does not race with us
5369          * setting the idle frequency and so that it is bounded by
5370          * our rpm wakeref. And then disable the interrupts to stop any
5371          * futher RPS reclocking whilst we are asleep.
5372          */
5373         gen6_disable_rps_interrupts(dev_priv);
5374
5375         mutex_lock(&dev_priv->rps.hw_lock);
5376         if (dev_priv->rps.enabled) {
5377                 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
5378                         vlv_set_rps_idle(dev_priv);
5379                 else
5380                         gen6_set_rps(dev_priv, dev_priv->rps.idle_freq);
5381                 dev_priv->rps.last_adj = 0;
5382                 I915_WRITE(GEN6_PMINTRMSK,
5383                            gen6_sanitize_rps_pm_mask(dev_priv, ~0));
5384         }
5385         mutex_unlock(&dev_priv->rps.hw_lock);
5386
5387         spin_lock(&dev_priv->rps.client_lock);
5388         while (!list_empty(&dev_priv->rps.clients))
5389                 list_del_init(dev_priv->rps.clients.next);
5390         spin_unlock(&dev_priv->rps.client_lock);
5391 }
5392
5393 void gen6_rps_boost(struct drm_i915_private *dev_priv,
5394                     struct intel_rps_client *rps,
5395                     unsigned long submitted)
5396 {
5397         /* This is intentionally racy! We peek at the state here, then
5398          * validate inside the RPS worker.
5399          */
5400         if (!(dev_priv->gt.awake &&
5401               dev_priv->rps.enabled &&
5402               dev_priv->rps.cur_freq < dev_priv->rps.boost_freq))
5403                 return;
5404
5405         /* Force a RPS boost (and don't count it against the client) if
5406          * the GPU is severely congested.
5407          */
5408         if (rps && time_after(jiffies, submitted + DRM_I915_THROTTLE_JIFFIES))
5409                 rps = NULL;
5410
5411         spin_lock(&dev_priv->rps.client_lock);
5412         if (rps == NULL || list_empty(&rps->link)) {
5413                 spin_lock_irq(&dev_priv->irq_lock);
5414                 if (dev_priv->rps.interrupts_enabled) {
5415                         dev_priv->rps.client_boost = true;
5416                         schedule_work(&dev_priv->rps.work);
5417                 }
5418                 spin_unlock_irq(&dev_priv->irq_lock);
5419
5420                 if (rps != NULL) {
5421                         list_add(&rps->link, &dev_priv->rps.clients);
5422                         rps->boosts++;
5423                 } else
5424                         dev_priv->rps.boosts++;
5425         }
5426         spin_unlock(&dev_priv->rps.client_lock);
5427 }
5428
5429 int intel_set_rps(struct drm_i915_private *dev_priv, u8 val)
5430 {
5431         int err;
5432
5433         lockdep_assert_held(&dev_priv->rps.hw_lock);
5434         GEM_BUG_ON(val > dev_priv->rps.max_freq);
5435         GEM_BUG_ON(val < dev_priv->rps.min_freq);
5436
5437         if (!dev_priv->rps.enabled) {
5438                 dev_priv->rps.cur_freq = val;
5439                 return 0;
5440         }
5441
5442         if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
5443                 err = valleyview_set_rps(dev_priv, val);
5444         else
5445                 err = gen6_set_rps(dev_priv, val);
5446
5447         return err;
5448 }
5449
5450 static void gen9_disable_rc6(struct drm_i915_private *dev_priv)
5451 {
5452         I915_WRITE(GEN6_RC_CONTROL, 0);
5453         I915_WRITE(GEN9_PG_ENABLE, 0);
5454 }
5455
5456 static void gen9_disable_rps(struct drm_i915_private *dev_priv)
5457 {
5458         I915_WRITE(GEN6_RP_CONTROL, 0);
5459 }
5460
5461 static void gen6_disable_rps(struct drm_i915_private *dev_priv)
5462 {
5463         I915_WRITE(GEN6_RC_CONTROL, 0);
5464         I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
5465         I915_WRITE(GEN6_RP_CONTROL, 0);
5466 }
5467
5468 static void cherryview_disable_rps(struct drm_i915_private *dev_priv)
5469 {
5470         I915_WRITE(GEN6_RC_CONTROL, 0);
5471 }
5472
5473 static void valleyview_disable_rps(struct drm_i915_private *dev_priv)
5474 {
5475         /* we're doing forcewake before Disabling RC6,
5476          * This what the BIOS expects when going into suspend */
5477         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5478
5479         I915_WRITE(GEN6_RC_CONTROL, 0);
5480
5481         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5482 }
5483
5484 static void intel_print_rc6_info(struct drm_i915_private *dev_priv, u32 mode)
5485 {
5486         if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
5487                 if (mode & (GEN7_RC_CTL_TO_MODE | GEN6_RC_CTL_EI_MODE(1)))
5488                         mode = GEN6_RC_CTL_RC6_ENABLE;
5489                 else
5490                         mode = 0;
5491         }
5492         if (HAS_RC6p(dev_priv))
5493                 DRM_DEBUG_DRIVER("Enabling RC6 states: "
5494                                  "RC6 %s RC6p %s RC6pp %s\n",
5495                                  onoff(mode & GEN6_RC_CTL_RC6_ENABLE),
5496                                  onoff(mode & GEN6_RC_CTL_RC6p_ENABLE),
5497                                  onoff(mode & GEN6_RC_CTL_RC6pp_ENABLE));
5498
5499         else
5500                 DRM_DEBUG_DRIVER("Enabling RC6 states: RC6 %s\n",
5501                                  onoff(mode & GEN6_RC_CTL_RC6_ENABLE));
5502 }
5503
5504 static bool bxt_check_bios_rc6_setup(struct drm_i915_private *dev_priv)
5505 {
5506         struct i915_ggtt *ggtt = &dev_priv->ggtt;
5507         bool enable_rc6 = true;
5508         unsigned long rc6_ctx_base;
5509         u32 rc_ctl;
5510         int rc_sw_target;
5511
5512         rc_ctl = I915_READ(GEN6_RC_CONTROL);
5513         rc_sw_target = (I915_READ(GEN6_RC_STATE) & RC_SW_TARGET_STATE_MASK) >>
5514                        RC_SW_TARGET_STATE_SHIFT;
5515         DRM_DEBUG_DRIVER("BIOS enabled RC states: "
5516                          "HW_CTRL %s HW_RC6 %s SW_TARGET_STATE %x\n",
5517                          onoff(rc_ctl & GEN6_RC_CTL_HW_ENABLE),
5518                          onoff(rc_ctl & GEN6_RC_CTL_RC6_ENABLE),
5519                          rc_sw_target);
5520
5521         if (!(I915_READ(RC6_LOCATION) & RC6_CTX_IN_DRAM)) {
5522                 DRM_DEBUG_DRIVER("RC6 Base location not set properly.\n");
5523                 enable_rc6 = false;
5524         }
5525
5526         /*
5527          * The exact context size is not known for BXT, so assume a page size
5528          * for this check.
5529          */
5530         rc6_ctx_base = I915_READ(RC6_CTX_BASE) & RC6_CTX_BASE_MASK;
5531         if (!((rc6_ctx_base >= ggtt->stolen_reserved_base) &&
5532               (rc6_ctx_base + PAGE_SIZE <= ggtt->stolen_reserved_base +
5533                                         ggtt->stolen_reserved_size))) {
5534                 DRM_DEBUG_DRIVER("RC6 Base address not as expected.\n");
5535                 enable_rc6 = false;
5536         }
5537
5538         if (!(((I915_READ(PWRCTX_MAXCNT_RCSUNIT) & IDLE_TIME_MASK) > 1) &&
5539               ((I915_READ(PWRCTX_MAXCNT_VCSUNIT0) & IDLE_TIME_MASK) > 1) &&
5540               ((I915_READ(PWRCTX_MAXCNT_BCSUNIT) & IDLE_TIME_MASK) > 1) &&
5541               ((I915_READ(PWRCTX_MAXCNT_VECSUNIT) & IDLE_TIME_MASK) > 1))) {
5542                 DRM_DEBUG_DRIVER("Engine Idle wait time not set properly.\n");
5543                 enable_rc6 = false;
5544         }
5545
5546         if (!I915_READ(GEN8_PUSHBUS_CONTROL) ||
5547             !I915_READ(GEN8_PUSHBUS_ENABLE) ||
5548             !I915_READ(GEN8_PUSHBUS_SHIFT)) {
5549                 DRM_DEBUG_DRIVER("Pushbus not setup properly.\n");
5550                 enable_rc6 = false;
5551         }
5552
5553         if (!I915_READ(GEN6_GFXPAUSE)) {
5554                 DRM_DEBUG_DRIVER("GFX pause not setup properly.\n");
5555                 enable_rc6 = false;
5556         }
5557
5558         if (!I915_READ(GEN8_MISC_CTRL0)) {
5559                 DRM_DEBUG_DRIVER("GPM control not setup properly.\n");
5560                 enable_rc6 = false;
5561         }
5562
5563         return enable_rc6;
5564 }
5565
5566 int sanitize_rc6_option(struct drm_i915_private *dev_priv, int enable_rc6)
5567 {
5568         /* No RC6 before Ironlake and code is gone for ilk. */
5569         if (INTEL_INFO(dev_priv)->gen < 6)
5570                 return 0;
5571
5572         if (!enable_rc6)
5573                 return 0;
5574
5575         if (IS_GEN9_LP(dev_priv) && !bxt_check_bios_rc6_setup(dev_priv)) {
5576                 DRM_INFO("RC6 disabled by BIOS\n");
5577                 return 0;
5578         }
5579
5580         /* Respect the kernel parameter if it is set */
5581         if (enable_rc6 >= 0) {
5582                 int mask;
5583
5584                 if (HAS_RC6p(dev_priv))
5585                         mask = INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE |
5586                                INTEL_RC6pp_ENABLE;
5587                 else
5588                         mask = INTEL_RC6_ENABLE;
5589
5590                 if ((enable_rc6 & mask) != enable_rc6)
5591                         DRM_DEBUG_DRIVER("Adjusting RC6 mask to %d "
5592                                          "(requested %d, valid %d)\n",
5593                                          enable_rc6 & mask, enable_rc6, mask);
5594
5595                 return enable_rc6 & mask;
5596         }
5597
5598         if (IS_IVYBRIDGE(dev_priv))
5599                 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
5600
5601         return INTEL_RC6_ENABLE;
5602 }
5603
5604 static void gen6_init_rps_frequencies(struct drm_i915_private *dev_priv)
5605 {
5606         /* All of these values are in units of 50MHz */
5607
5608         /* static values from HW: RP0 > RP1 > RPn (min_freq) */
5609         if (IS_GEN9_LP(dev_priv)) {
5610                 u32 rp_state_cap = I915_READ(BXT_RP_STATE_CAP);
5611                 dev_priv->rps.rp0_freq = (rp_state_cap >> 16) & 0xff;
5612                 dev_priv->rps.rp1_freq = (rp_state_cap >>  8) & 0xff;
5613                 dev_priv->rps.min_freq = (rp_state_cap >>  0) & 0xff;
5614         } else {
5615                 u32 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
5616                 dev_priv->rps.rp0_freq = (rp_state_cap >>  0) & 0xff;
5617                 dev_priv->rps.rp1_freq = (rp_state_cap >>  8) & 0xff;
5618                 dev_priv->rps.min_freq = (rp_state_cap >> 16) & 0xff;
5619         }
5620         /* hw_max = RP0 until we check for overclocking */
5621         dev_priv->rps.max_freq = dev_priv->rps.rp0_freq;
5622
5623         dev_priv->rps.efficient_freq = dev_priv->rps.rp1_freq;
5624         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv) ||
5625             IS_GEN9_BC(dev_priv)) {
5626                 u32 ddcc_status = 0;
5627
5628                 if (sandybridge_pcode_read(dev_priv,
5629                                            HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL,
5630                                            &ddcc_status) == 0)
5631                         dev_priv->rps.efficient_freq =
5632                                 clamp_t(u8,
5633                                         ((ddcc_status >> 8) & 0xff),
5634                                         dev_priv->rps.min_freq,
5635                                         dev_priv->rps.max_freq);
5636         }
5637
5638         if (IS_GEN9_BC(dev_priv)) {
5639                 /* Store the frequency values in 16.66 MHZ units, which is
5640                  * the natural hardware unit for SKL
5641                  */
5642                 dev_priv->rps.rp0_freq *= GEN9_FREQ_SCALER;
5643                 dev_priv->rps.rp1_freq *= GEN9_FREQ_SCALER;
5644                 dev_priv->rps.min_freq *= GEN9_FREQ_SCALER;
5645                 dev_priv->rps.max_freq *= GEN9_FREQ_SCALER;
5646                 dev_priv->rps.efficient_freq *= GEN9_FREQ_SCALER;
5647         }
5648 }
5649
5650 static void reset_rps(struct drm_i915_private *dev_priv,
5651                       int (*set)(struct drm_i915_private *, u8))
5652 {
5653         u8 freq = dev_priv->rps.cur_freq;
5654
5655         /* force a reset */
5656         dev_priv->rps.power = -1;
5657         dev_priv->rps.cur_freq = -1;
5658
5659         if (set(dev_priv, freq))
5660                 DRM_ERROR("Failed to reset RPS to initial values\n");
5661 }
5662
5663 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
5664 static void gen9_enable_rps(struct drm_i915_private *dev_priv)
5665 {
5666         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5667
5668         /* Program defaults and thresholds for RPS*/
5669         I915_WRITE(GEN6_RC_VIDEO_FREQ,
5670                 GEN9_FREQUENCY(dev_priv->rps.rp1_freq));
5671
5672         /* 1 second timeout*/
5673         I915_WRITE(GEN6_RP_DOWN_TIMEOUT,
5674                 GT_INTERVAL_FROM_US(dev_priv, 1000000));
5675
5676         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 0xa);
5677
5678         /* Leaning on the below call to gen6_set_rps to program/setup the
5679          * Up/Down EI & threshold registers, as well as the RP_CONTROL,
5680          * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
5681         reset_rps(dev_priv, gen6_set_rps);
5682
5683         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5684 }
5685
5686 static void gen9_enable_rc6(struct drm_i915_private *dev_priv)
5687 {
5688         struct intel_engine_cs *engine;
5689         enum intel_engine_id id;
5690         uint32_t rc6_mask = 0;
5691
5692         /* 1a: Software RC state - RC0 */
5693         I915_WRITE(GEN6_RC_STATE, 0);
5694
5695         /* 1b: Get forcewake during program sequence. Although the driver
5696          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5697         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5698
5699         /* 2a: Disable RC states. */
5700         I915_WRITE(GEN6_RC_CONTROL, 0);
5701
5702         /* 2b: Program RC6 thresholds.*/
5703
5704         /* WaRsDoubleRc6WrlWithCoarsePowerGating: Doubling WRL only when CPG is enabled */
5705         if (IS_SKYLAKE(dev_priv))
5706                 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 108 << 16);
5707         else
5708                 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16);
5709         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5710         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
5711         for_each_engine(engine, dev_priv, id)
5712                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5713
5714         if (HAS_GUC(dev_priv))
5715                 I915_WRITE(GUC_MAX_IDLE_COUNT, 0xA);
5716
5717         I915_WRITE(GEN6_RC_SLEEP, 0);
5718
5719         /* 2c: Program Coarse Power Gating Policies. */
5720         I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS, 25);
5721         I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS, 25);
5722
5723         /* 3a: Enable RC6 */
5724         if (intel_enable_rc6() & INTEL_RC6_ENABLE)
5725                 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
5726         DRM_INFO("RC6 %s\n", onoff(rc6_mask & GEN6_RC_CTL_RC6_ENABLE));
5727         I915_WRITE(GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */
5728         I915_WRITE(GEN6_RC_CONTROL,
5729                    GEN6_RC_CTL_HW_ENABLE | GEN6_RC_CTL_EI_MODE(1) | rc6_mask);
5730
5731         /*
5732          * 3b: Enable Coarse Power Gating only when RC6 is enabled.
5733          * WaRsDisableCoarsePowerGating:skl,bxt - Render/Media PG need to be disabled with RC6.
5734          */
5735         if (NEEDS_WaRsDisableCoarsePowerGating(dev_priv))
5736                 I915_WRITE(GEN9_PG_ENABLE, 0);
5737         else
5738                 I915_WRITE(GEN9_PG_ENABLE, (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ?
5739                                 (GEN9_RENDER_PG_ENABLE | GEN9_MEDIA_PG_ENABLE) : 0);
5740
5741         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5742 }
5743
5744 static void gen8_enable_rps(struct drm_i915_private *dev_priv)
5745 {
5746         struct intel_engine_cs *engine;
5747         enum intel_engine_id id;
5748         uint32_t rc6_mask = 0;
5749
5750         /* 1a: Software RC state - RC0 */
5751         I915_WRITE(GEN6_RC_STATE, 0);
5752
5753         /* 1c & 1d: Get forcewake during program sequence. Although the driver
5754          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5755         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5756
5757         /* 2a: Disable RC states. */
5758         I915_WRITE(GEN6_RC_CONTROL, 0);
5759
5760         /* 2b: Program RC6 thresholds.*/
5761         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
5762         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5763         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
5764         for_each_engine(engine, dev_priv, id)
5765                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5766         I915_WRITE(GEN6_RC_SLEEP, 0);
5767         if (IS_BROADWELL(dev_priv))
5768                 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */
5769         else
5770                 I915_WRITE(GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
5771
5772         /* 3: Enable RC6 */
5773         if (intel_enable_rc6() & INTEL_RC6_ENABLE)
5774                 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
5775         intel_print_rc6_info(dev_priv, rc6_mask);
5776         if (IS_BROADWELL(dev_priv))
5777                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5778                                 GEN7_RC_CTL_TO_MODE |
5779                                 rc6_mask);
5780         else
5781                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5782                                 GEN6_RC_CTL_EI_MODE(1) |
5783                                 rc6_mask);
5784
5785         /* 4 Program defaults and thresholds for RPS*/
5786         I915_WRITE(GEN6_RPNSWREQ,
5787                    HSW_FREQUENCY(dev_priv->rps.rp1_freq));
5788         I915_WRITE(GEN6_RC_VIDEO_FREQ,
5789                    HSW_FREQUENCY(dev_priv->rps.rp1_freq));
5790         /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
5791         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
5792
5793         /* Docs recommend 900MHz, and 300 MHz respectively */
5794         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
5795                    dev_priv->rps.max_freq_softlimit << 24 |
5796                    dev_priv->rps.min_freq_softlimit << 16);
5797
5798         I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
5799         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
5800         I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
5801         I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
5802
5803         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5804
5805         /* 5: Enable RPS */
5806         I915_WRITE(GEN6_RP_CONTROL,
5807                    GEN6_RP_MEDIA_TURBO |
5808                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
5809                    GEN6_RP_MEDIA_IS_GFX |
5810                    GEN6_RP_ENABLE |
5811                    GEN6_RP_UP_BUSY_AVG |
5812                    GEN6_RP_DOWN_IDLE_AVG);
5813
5814         /* 6: Ring frequency + overclocking (our driver does this later */
5815
5816         reset_rps(dev_priv, gen6_set_rps);
5817
5818         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5819 }
5820
5821 static void gen6_enable_rps(struct drm_i915_private *dev_priv)
5822 {
5823         struct intel_engine_cs *engine;
5824         enum intel_engine_id id;
5825         u32 rc6vids, rc6_mask = 0;
5826         u32 gtfifodbg;
5827         int rc6_mode;
5828         int ret;
5829
5830         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5831
5832         /* Here begins a magic sequence of register writes to enable
5833          * auto-downclocking.
5834          *
5835          * Perhaps there might be some value in exposing these to
5836          * userspace...
5837          */
5838         I915_WRITE(GEN6_RC_STATE, 0);
5839
5840         /* Clear the DBG now so we don't confuse earlier errors */
5841         gtfifodbg = I915_READ(GTFIFODBG);
5842         if (gtfifodbg) {
5843                 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
5844                 I915_WRITE(GTFIFODBG, gtfifodbg);
5845         }
5846
5847         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5848
5849         /* disable the counters and set deterministic thresholds */
5850         I915_WRITE(GEN6_RC_CONTROL, 0);
5851
5852         I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
5853         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
5854         I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
5855         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
5856         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
5857
5858         for_each_engine(engine, dev_priv, id)
5859                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5860
5861         I915_WRITE(GEN6_RC_SLEEP, 0);
5862         I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
5863         if (IS_IVYBRIDGE(dev_priv))
5864                 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
5865         else
5866                 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
5867         I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
5868         I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
5869
5870         /* Check if we are enabling RC6 */
5871         rc6_mode = intel_enable_rc6();
5872         if (rc6_mode & INTEL_RC6_ENABLE)
5873                 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
5874
5875         /* We don't use those on Haswell */
5876         if (!IS_HASWELL(dev_priv)) {
5877                 if (rc6_mode & INTEL_RC6p_ENABLE)
5878                         rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
5879
5880                 if (rc6_mode & INTEL_RC6pp_ENABLE)
5881                         rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
5882         }
5883
5884         intel_print_rc6_info(dev_priv, rc6_mask);
5885
5886         I915_WRITE(GEN6_RC_CONTROL,
5887                    rc6_mask |
5888                    GEN6_RC_CTL_EI_MODE(1) |
5889                    GEN6_RC_CTL_HW_ENABLE);
5890
5891         /* Power down if completely idle for over 50ms */
5892         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
5893         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5894
5895         reset_rps(dev_priv, gen6_set_rps);
5896
5897         rc6vids = 0;
5898         ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
5899         if (IS_GEN6(dev_priv) && ret) {
5900                 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
5901         } else if (IS_GEN6(dev_priv) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
5902                 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
5903                           GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
5904                 rc6vids &= 0xffff00;
5905                 rc6vids |= GEN6_ENCODE_RC6_VID(450);
5906                 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
5907                 if (ret)
5908                         DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
5909         }
5910
5911         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5912 }
5913
5914 static void gen6_update_ring_freq(struct drm_i915_private *dev_priv)
5915 {
5916         int min_freq = 15;
5917         unsigned int gpu_freq;
5918         unsigned int max_ia_freq, min_ring_freq;
5919         unsigned int max_gpu_freq, min_gpu_freq;
5920         int scaling_factor = 180;
5921         struct cpufreq_policy *policy;
5922
5923         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5924
5925         policy = cpufreq_cpu_get(0);
5926         if (policy) {
5927                 max_ia_freq = policy->cpuinfo.max_freq;
5928                 cpufreq_cpu_put(policy);
5929         } else {
5930                 /*
5931                  * Default to measured freq if none found, PCU will ensure we
5932                  * don't go over
5933                  */
5934                 max_ia_freq = tsc_khz;
5935         }
5936
5937         /* Convert from kHz to MHz */
5938         max_ia_freq /= 1000;
5939
5940         min_ring_freq = I915_READ(DCLK) & 0xf;
5941         /* convert DDR frequency from units of 266.6MHz to bandwidth */
5942         min_ring_freq = mult_frac(min_ring_freq, 8, 3);
5943
5944         if (IS_GEN9_BC(dev_priv)) {
5945                 /* Convert GT frequency to 50 HZ units */
5946                 min_gpu_freq = dev_priv->rps.min_freq / GEN9_FREQ_SCALER;
5947                 max_gpu_freq = dev_priv->rps.max_freq / GEN9_FREQ_SCALER;
5948         } else {
5949                 min_gpu_freq = dev_priv->rps.min_freq;
5950                 max_gpu_freq = dev_priv->rps.max_freq;
5951         }
5952
5953         /*
5954          * For each potential GPU frequency, load a ring frequency we'd like
5955          * to use for memory access.  We do this by specifying the IA frequency
5956          * the PCU should use as a reference to determine the ring frequency.
5957          */
5958         for (gpu_freq = max_gpu_freq; gpu_freq >= min_gpu_freq; gpu_freq--) {
5959                 int diff = max_gpu_freq - gpu_freq;
5960                 unsigned int ia_freq = 0, ring_freq = 0;
5961
5962                 if (IS_GEN9_BC(dev_priv)) {
5963                         /*
5964                          * ring_freq = 2 * GT. ring_freq is in 100MHz units
5965                          * No floor required for ring frequency on SKL.
5966                          */
5967                         ring_freq = gpu_freq;
5968                 } else if (INTEL_INFO(dev_priv)->gen >= 8) {
5969                         /* max(2 * GT, DDR). NB: GT is 50MHz units */
5970                         ring_freq = max(min_ring_freq, gpu_freq);
5971                 } else if (IS_HASWELL(dev_priv)) {
5972                         ring_freq = mult_frac(gpu_freq, 5, 4);
5973                         ring_freq = max(min_ring_freq, ring_freq);
5974                         /* leave ia_freq as the default, chosen by cpufreq */
5975                 } else {
5976                         /* On older processors, there is no separate ring
5977                          * clock domain, so in order to boost the bandwidth
5978                          * of the ring, we need to upclock the CPU (ia_freq).
5979                          *
5980                          * For GPU frequencies less than 750MHz,
5981                          * just use the lowest ring freq.
5982                          */
5983                         if (gpu_freq < min_freq)
5984                                 ia_freq = 800;
5985                         else
5986                                 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
5987                         ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
5988                 }
5989
5990                 sandybridge_pcode_write(dev_priv,
5991                                         GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
5992                                         ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
5993                                         ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
5994                                         gpu_freq);
5995         }
5996 }
5997
5998 static int cherryview_rps_max_freq(struct drm_i915_private *dev_priv)
5999 {
6000         u32 val, rp0;
6001
6002         val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
6003
6004         switch (INTEL_INFO(dev_priv)->sseu.eu_total) {
6005         case 8:
6006                 /* (2 * 4) config */
6007                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT);
6008                 break;
6009         case 12:
6010                 /* (2 * 6) config */
6011                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT);
6012                 break;
6013         case 16:
6014                 /* (2 * 8) config */
6015         default:
6016                 /* Setting (2 * 8) Min RP0 for any other combination */
6017                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT);
6018                 break;
6019         }
6020
6021         rp0 = (rp0 & FB_GFX_FREQ_FUSE_MASK);
6022
6023         return rp0;
6024 }
6025
6026 static int cherryview_rps_rpe_freq(struct drm_i915_private *dev_priv)
6027 {
6028         u32 val, rpe;
6029
6030         val = vlv_punit_read(dev_priv, PUNIT_GPU_DUTYCYCLE_REG);
6031         rpe = (val >> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK;
6032
6033         return rpe;
6034 }
6035
6036 static int cherryview_rps_guar_freq(struct drm_i915_private *dev_priv)
6037 {
6038         u32 val, rp1;
6039
6040         val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
6041         rp1 = (val & FB_GFX_FREQ_FUSE_MASK);
6042
6043         return rp1;
6044 }
6045
6046 static u32 cherryview_rps_min_freq(struct drm_i915_private *dev_priv)
6047 {
6048         u32 val, rpn;
6049
6050         val = vlv_punit_read(dev_priv, FB_GFX_FMIN_AT_VMIN_FUSE);
6051         rpn = ((val >> FB_GFX_FMIN_AT_VMIN_FUSE_SHIFT) &
6052                        FB_GFX_FREQ_FUSE_MASK);
6053
6054         return rpn;
6055 }
6056
6057 static int valleyview_rps_guar_freq(struct drm_i915_private *dev_priv)
6058 {
6059         u32 val, rp1;
6060
6061         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
6062
6063         rp1 = (val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT;
6064
6065         return rp1;
6066 }
6067
6068 static int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
6069 {
6070         u32 val, rp0;
6071
6072         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
6073
6074         rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
6075         /* Clamp to max */
6076         rp0 = min_t(u32, rp0, 0xea);
6077
6078         return rp0;
6079 }
6080
6081 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
6082 {
6083         u32 val, rpe;
6084
6085         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
6086         rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
6087         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
6088         rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
6089
6090         return rpe;
6091 }
6092
6093 static int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
6094 {
6095         u32 val;
6096
6097         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
6098         /*
6099          * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
6100          * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
6101          * a BYT-M B0 the above register contains 0xbf. Moreover when setting
6102          * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
6103          * to make sure it matches what Punit accepts.
6104          */
6105         return max_t(u32, val, 0xc0);
6106 }
6107
6108 /* Check that the pctx buffer wasn't move under us. */
6109 static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
6110 {
6111         unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
6112
6113         WARN_ON(pctx_addr != dev_priv->mm.stolen_base +
6114                              dev_priv->vlv_pctx->stolen->start);
6115 }
6116
6117
6118 /* Check that the pcbr address is not empty. */
6119 static void cherryview_check_pctx(struct drm_i915_private *dev_priv)
6120 {
6121         unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
6122
6123         WARN_ON((pctx_addr >> VLV_PCBR_ADDR_SHIFT) == 0);
6124 }
6125
6126 static void cherryview_setup_pctx(struct drm_i915_private *dev_priv)
6127 {
6128         struct i915_ggtt *ggtt = &dev_priv->ggtt;
6129         unsigned long pctx_paddr, paddr;
6130         u32 pcbr;
6131         int pctx_size = 32*1024;
6132
6133         pcbr = I915_READ(VLV_PCBR);
6134         if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) {
6135                 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
6136                 paddr = (dev_priv->mm.stolen_base +
6137                          (ggtt->stolen_size - pctx_size));
6138
6139                 pctx_paddr = (paddr & (~4095));
6140                 I915_WRITE(VLV_PCBR, pctx_paddr);
6141         }
6142
6143         DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
6144 }
6145
6146 static void valleyview_setup_pctx(struct drm_i915_private *dev_priv)
6147 {
6148         struct drm_i915_gem_object *pctx;
6149         unsigned long pctx_paddr;
6150         u32 pcbr;
6151         int pctx_size = 24*1024;
6152
6153         pcbr = I915_READ(VLV_PCBR);
6154         if (pcbr) {
6155                 /* BIOS set it up already, grab the pre-alloc'd space */
6156                 int pcbr_offset;
6157
6158                 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
6159                 pctx = i915_gem_object_create_stolen_for_preallocated(dev_priv,
6160                                                                       pcbr_offset,
6161                                                                       I915_GTT_OFFSET_NONE,
6162                                                                       pctx_size);
6163                 goto out;
6164         }
6165
6166         DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
6167
6168         /*
6169          * From the Gunit register HAS:
6170          * The Gfx driver is expected to program this register and ensure
6171          * proper allocation within Gfx stolen memory.  For example, this
6172          * register should be programmed such than the PCBR range does not
6173          * overlap with other ranges, such as the frame buffer, protected
6174          * memory, or any other relevant ranges.
6175          */
6176         pctx = i915_gem_object_create_stolen(dev_priv, pctx_size);
6177         if (!pctx) {
6178                 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
6179                 goto out;
6180         }
6181
6182         pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
6183         I915_WRITE(VLV_PCBR, pctx_paddr);
6184
6185 out:
6186         DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
6187         dev_priv->vlv_pctx = pctx;
6188 }
6189
6190 static void valleyview_cleanup_pctx(struct drm_i915_private *dev_priv)
6191 {
6192         if (WARN_ON(!dev_priv->vlv_pctx))
6193                 return;
6194
6195         i915_gem_object_put(dev_priv->vlv_pctx);
6196         dev_priv->vlv_pctx = NULL;
6197 }
6198
6199 static void vlv_init_gpll_ref_freq(struct drm_i915_private *dev_priv)
6200 {
6201         dev_priv->rps.gpll_ref_freq =
6202                 vlv_get_cck_clock(dev_priv, "GPLL ref",
6203                                   CCK_GPLL_CLOCK_CONTROL,
6204                                   dev_priv->czclk_freq);
6205
6206         DRM_DEBUG_DRIVER("GPLL reference freq: %d kHz\n",
6207                          dev_priv->rps.gpll_ref_freq);
6208 }
6209
6210 static void valleyview_init_gt_powersave(struct drm_i915_private *dev_priv)
6211 {
6212         u32 val;
6213
6214         valleyview_setup_pctx(dev_priv);
6215
6216         vlv_init_gpll_ref_freq(dev_priv);
6217
6218         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
6219         switch ((val >> 6) & 3) {
6220         case 0:
6221         case 1:
6222                 dev_priv->mem_freq = 800;
6223                 break;
6224         case 2:
6225                 dev_priv->mem_freq = 1066;
6226                 break;
6227         case 3:
6228                 dev_priv->mem_freq = 1333;
6229                 break;
6230         }
6231         DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
6232
6233         dev_priv->rps.max_freq = valleyview_rps_max_freq(dev_priv);
6234         dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
6235         DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
6236                          intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
6237                          dev_priv->rps.max_freq);
6238
6239         dev_priv->rps.efficient_freq = valleyview_rps_rpe_freq(dev_priv);
6240         DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
6241                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
6242                          dev_priv->rps.efficient_freq);
6243
6244         dev_priv->rps.rp1_freq = valleyview_rps_guar_freq(dev_priv);
6245         DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
6246                          intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
6247                          dev_priv->rps.rp1_freq);
6248
6249         dev_priv->rps.min_freq = valleyview_rps_min_freq(dev_priv);
6250         DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
6251                          intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
6252                          dev_priv->rps.min_freq);
6253 }
6254
6255 static void cherryview_init_gt_powersave(struct drm_i915_private *dev_priv)
6256 {
6257         u32 val;
6258
6259         cherryview_setup_pctx(dev_priv);
6260
6261         vlv_init_gpll_ref_freq(dev_priv);
6262
6263         mutex_lock(&dev_priv->sb_lock);
6264         val = vlv_cck_read(dev_priv, CCK_FUSE_REG);
6265         mutex_unlock(&dev_priv->sb_lock);
6266
6267         switch ((val >> 2) & 0x7) {
6268         case 3:
6269                 dev_priv->mem_freq = 2000;
6270                 break;
6271         default:
6272                 dev_priv->mem_freq = 1600;
6273                 break;
6274         }
6275         DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
6276
6277         dev_priv->rps.max_freq = cherryview_rps_max_freq(dev_priv);
6278         dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
6279         DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
6280                          intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
6281                          dev_priv->rps.max_freq);
6282
6283         dev_priv->rps.efficient_freq = cherryview_rps_rpe_freq(dev_priv);
6284         DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
6285                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
6286                          dev_priv->rps.efficient_freq);
6287
6288         dev_priv->rps.rp1_freq = cherryview_rps_guar_freq(dev_priv);
6289         DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
6290                          intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
6291                          dev_priv->rps.rp1_freq);
6292
6293         dev_priv->rps.min_freq = cherryview_rps_min_freq(dev_priv);
6294         DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
6295                          intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
6296                          dev_priv->rps.min_freq);
6297
6298         WARN_ONCE((dev_priv->rps.max_freq |
6299                    dev_priv->rps.efficient_freq |
6300                    dev_priv->rps.rp1_freq |
6301                    dev_priv->rps.min_freq) & 1,
6302                   "Odd GPU freq values\n");
6303 }
6304
6305 static void valleyview_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
6306 {
6307         valleyview_cleanup_pctx(dev_priv);
6308 }
6309
6310 static void cherryview_enable_rps(struct drm_i915_private *dev_priv)
6311 {
6312         struct intel_engine_cs *engine;
6313         enum intel_engine_id id;
6314         u32 gtfifodbg, val, rc6_mode = 0, pcbr;
6315
6316         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6317
6318         gtfifodbg = I915_READ(GTFIFODBG) & ~(GT_FIFO_SBDEDICATE_FREE_ENTRY_CHV |
6319                                              GT_FIFO_FREE_ENTRIES_CHV);
6320         if (gtfifodbg) {
6321                 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
6322                                  gtfifodbg);
6323                 I915_WRITE(GTFIFODBG, gtfifodbg);
6324         }
6325
6326         cherryview_check_pctx(dev_priv);
6327
6328         /* 1a & 1b: Get forcewake during program sequence. Although the driver
6329          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
6330         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6331
6332         /*  Disable RC states. */
6333         I915_WRITE(GEN6_RC_CONTROL, 0);
6334
6335         /* 2a: Program RC6 thresholds.*/
6336         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
6337         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
6338         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
6339
6340         for_each_engine(engine, dev_priv, id)
6341                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6342         I915_WRITE(GEN6_RC_SLEEP, 0);
6343
6344         /* TO threshold set to 500 us ( 0x186 * 1.28 us) */
6345         I915_WRITE(GEN6_RC6_THRESHOLD, 0x186);
6346
6347         /* allows RC6 residency counter to work */
6348         I915_WRITE(VLV_COUNTER_CONTROL,
6349                    _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
6350                                       VLV_MEDIA_RC6_COUNT_EN |
6351                                       VLV_RENDER_RC6_COUNT_EN));
6352
6353         /* For now we assume BIOS is allocating and populating the PCBR  */
6354         pcbr = I915_READ(VLV_PCBR);
6355
6356         /* 3: Enable RC6 */
6357         if ((intel_enable_rc6() & INTEL_RC6_ENABLE) &&
6358             (pcbr >> VLV_PCBR_ADDR_SHIFT))
6359                 rc6_mode = GEN7_RC_CTL_TO_MODE;
6360
6361         I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
6362
6363         /* 4 Program defaults and thresholds for RPS*/
6364         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
6365         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
6366         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
6367         I915_WRITE(GEN6_RP_UP_EI, 66000);
6368         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
6369
6370         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
6371
6372         /* 5: Enable RPS */
6373         I915_WRITE(GEN6_RP_CONTROL,
6374                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
6375                    GEN6_RP_MEDIA_IS_GFX |
6376                    GEN6_RP_ENABLE |
6377                    GEN6_RP_UP_BUSY_AVG |
6378                    GEN6_RP_DOWN_IDLE_AVG);
6379
6380         /* Setting Fixed Bias */
6381         val = VLV_OVERRIDE_EN |
6382                   VLV_SOC_TDP_EN |
6383                   CHV_BIAS_CPU_50_SOC_50;
6384         vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
6385
6386         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
6387
6388         /* RPS code assumes GPLL is used */
6389         WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
6390
6391         DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
6392         DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
6393
6394         reset_rps(dev_priv, valleyview_set_rps);
6395
6396         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6397 }
6398
6399 static void valleyview_enable_rps(struct drm_i915_private *dev_priv)
6400 {
6401         struct intel_engine_cs *engine;
6402         enum intel_engine_id id;
6403         u32 gtfifodbg, val, rc6_mode = 0;
6404
6405         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6406
6407         valleyview_check_pctx(dev_priv);
6408
6409         gtfifodbg = I915_READ(GTFIFODBG);
6410         if (gtfifodbg) {
6411                 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
6412                                  gtfifodbg);
6413                 I915_WRITE(GTFIFODBG, gtfifodbg);
6414         }
6415
6416         /* If VLV, Forcewake all wells, else re-direct to regular path */
6417         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6418
6419         /*  Disable RC states. */
6420         I915_WRITE(GEN6_RC_CONTROL, 0);
6421
6422         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
6423         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
6424         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
6425         I915_WRITE(GEN6_RP_UP_EI, 66000);
6426         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
6427
6428         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
6429
6430         I915_WRITE(GEN6_RP_CONTROL,
6431                    GEN6_RP_MEDIA_TURBO |
6432                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
6433                    GEN6_RP_MEDIA_IS_GFX |
6434                    GEN6_RP_ENABLE |
6435                    GEN6_RP_UP_BUSY_AVG |
6436                    GEN6_RP_DOWN_IDLE_CONT);
6437
6438         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
6439         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
6440         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
6441
6442         for_each_engine(engine, dev_priv, id)
6443                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6444
6445         I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
6446
6447         /* allows RC6 residency counter to work */
6448         I915_WRITE(VLV_COUNTER_CONTROL,
6449                    _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
6450                                       VLV_MEDIA_RC0_COUNT_EN |
6451                                       VLV_RENDER_RC0_COUNT_EN |
6452                                       VLV_MEDIA_RC6_COUNT_EN |
6453                                       VLV_RENDER_RC6_COUNT_EN));
6454
6455         if (intel_enable_rc6() & INTEL_RC6_ENABLE)
6456                 rc6_mode = GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
6457
6458         intel_print_rc6_info(dev_priv, rc6_mode);
6459
6460         I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
6461
6462         /* Setting Fixed Bias */
6463         val = VLV_OVERRIDE_EN |
6464                   VLV_SOC_TDP_EN |
6465                   VLV_BIAS_CPU_125_SOC_875;
6466         vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
6467
6468         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
6469
6470         /* RPS code assumes GPLL is used */
6471         WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
6472
6473         DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
6474         DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
6475
6476         reset_rps(dev_priv, valleyview_set_rps);
6477
6478         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6479 }
6480
6481 static unsigned long intel_pxfreq(u32 vidfreq)
6482 {
6483         unsigned long freq;
6484         int div = (vidfreq & 0x3f0000) >> 16;
6485         int post = (vidfreq & 0x3000) >> 12;
6486         int pre = (vidfreq & 0x7);
6487
6488         if (!pre)
6489                 return 0;
6490
6491         freq = ((div * 133333) / ((1<<post) * pre));
6492
6493         return freq;
6494 }
6495
6496 static const struct cparams {
6497         u16 i;
6498         u16 t;
6499         u16 m;
6500         u16 c;
6501 } cparams[] = {
6502         { 1, 1333, 301, 28664 },
6503         { 1, 1066, 294, 24460 },
6504         { 1, 800, 294, 25192 },
6505         { 0, 1333, 276, 27605 },
6506         { 0, 1066, 276, 27605 },
6507         { 0, 800, 231, 23784 },
6508 };
6509
6510 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
6511 {
6512         u64 total_count, diff, ret;
6513         u32 count1, count2, count3, m = 0, c = 0;
6514         unsigned long now = jiffies_to_msecs(jiffies), diff1;
6515         int i;
6516
6517         lockdep_assert_held(&mchdev_lock);
6518
6519         diff1 = now - dev_priv->ips.last_time1;
6520
6521         /* Prevent division-by-zero if we are asking too fast.
6522          * Also, we don't get interesting results if we are polling
6523          * faster than once in 10ms, so just return the saved value
6524          * in such cases.
6525          */
6526         if (diff1 <= 10)
6527                 return dev_priv->ips.chipset_power;
6528
6529         count1 = I915_READ(DMIEC);
6530         count2 = I915_READ(DDREC);
6531         count3 = I915_READ(CSIEC);
6532
6533         total_count = count1 + count2 + count3;
6534
6535         /* FIXME: handle per-counter overflow */
6536         if (total_count < dev_priv->ips.last_count1) {
6537                 diff = ~0UL - dev_priv->ips.last_count1;
6538                 diff += total_count;
6539         } else {
6540                 diff = total_count - dev_priv->ips.last_count1;
6541         }
6542
6543         for (i = 0; i < ARRAY_SIZE(cparams); i++) {
6544                 if (cparams[i].i == dev_priv->ips.c_m &&
6545                     cparams[i].t == dev_priv->ips.r_t) {
6546                         m = cparams[i].m;
6547                         c = cparams[i].c;
6548                         break;
6549                 }
6550         }
6551
6552         diff = div_u64(diff, diff1);
6553         ret = ((m * diff) + c);
6554         ret = div_u64(ret, 10);
6555
6556         dev_priv->ips.last_count1 = total_count;
6557         dev_priv->ips.last_time1 = now;
6558
6559         dev_priv->ips.chipset_power = ret;
6560
6561         return ret;
6562 }
6563
6564 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
6565 {
6566         unsigned long val;
6567
6568         if (INTEL_INFO(dev_priv)->gen != 5)
6569                 return 0;
6570
6571         spin_lock_irq(&mchdev_lock);
6572
6573         val = __i915_chipset_val(dev_priv);
6574
6575         spin_unlock_irq(&mchdev_lock);
6576
6577         return val;
6578 }
6579
6580 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
6581 {
6582         unsigned long m, x, b;
6583         u32 tsfs;
6584
6585         tsfs = I915_READ(TSFS);
6586
6587         m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
6588         x = I915_READ8(TR1);
6589
6590         b = tsfs & TSFS_INTR_MASK;
6591
6592         return ((m * x) / 127) - b;
6593 }
6594
6595 static int _pxvid_to_vd(u8 pxvid)
6596 {
6597         if (pxvid == 0)
6598                 return 0;
6599
6600         if (pxvid >= 8 && pxvid < 31)
6601                 pxvid = 31;
6602
6603         return (pxvid + 2) * 125;
6604 }
6605
6606 static u32 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
6607 {
6608         const int vd = _pxvid_to_vd(pxvid);
6609         const int vm = vd - 1125;
6610
6611         if (INTEL_INFO(dev_priv)->is_mobile)
6612                 return vm > 0 ? vm : 0;
6613
6614         return vd;
6615 }
6616
6617 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
6618 {
6619         u64 now, diff, diffms;
6620         u32 count;
6621
6622         lockdep_assert_held(&mchdev_lock);
6623
6624         now = ktime_get_raw_ns();
6625         diffms = now - dev_priv->ips.last_time2;
6626         do_div(diffms, NSEC_PER_MSEC);
6627
6628         /* Don't divide by 0 */
6629         if (!diffms)
6630                 return;
6631
6632         count = I915_READ(GFXEC);
6633
6634         if (count < dev_priv->ips.last_count2) {
6635                 diff = ~0UL - dev_priv->ips.last_count2;
6636                 diff += count;
6637         } else {
6638                 diff = count - dev_priv->ips.last_count2;
6639         }
6640
6641         dev_priv->ips.last_count2 = count;
6642         dev_priv->ips.last_time2 = now;
6643
6644         /* More magic constants... */
6645         diff = diff * 1181;
6646         diff = div_u64(diff, diffms * 10);
6647         dev_priv->ips.gfx_power = diff;
6648 }
6649
6650 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
6651 {
6652         if (INTEL_INFO(dev_priv)->gen != 5)
6653                 return;
6654
6655         spin_lock_irq(&mchdev_lock);
6656
6657         __i915_update_gfx_val(dev_priv);
6658
6659         spin_unlock_irq(&mchdev_lock);
6660 }
6661
6662 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
6663 {
6664         unsigned long t, corr, state1, corr2, state2;
6665         u32 pxvid, ext_v;
6666
6667         lockdep_assert_held(&mchdev_lock);
6668
6669         pxvid = I915_READ(PXVFREQ(dev_priv->rps.cur_freq));
6670         pxvid = (pxvid >> 24) & 0x7f;
6671         ext_v = pvid_to_extvid(dev_priv, pxvid);
6672
6673         state1 = ext_v;
6674
6675         t = i915_mch_val(dev_priv);
6676
6677         /* Revel in the empirically derived constants */
6678
6679         /* Correction factor in 1/100000 units */
6680         if (t > 80)
6681                 corr = ((t * 2349) + 135940);
6682         else if (t >= 50)
6683                 corr = ((t * 964) + 29317);
6684         else /* < 50 */
6685                 corr = ((t * 301) + 1004);
6686
6687         corr = corr * ((150142 * state1) / 10000 - 78642);
6688         corr /= 100000;
6689         corr2 = (corr * dev_priv->ips.corr);
6690
6691         state2 = (corr2 * state1) / 10000;
6692         state2 /= 100; /* convert to mW */
6693
6694         __i915_update_gfx_val(dev_priv);
6695
6696         return dev_priv->ips.gfx_power + state2;
6697 }
6698
6699 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
6700 {
6701         unsigned long val;
6702
6703         if (INTEL_INFO(dev_priv)->gen != 5)
6704                 return 0;
6705
6706         spin_lock_irq(&mchdev_lock);
6707
6708         val = __i915_gfx_val(dev_priv);
6709
6710         spin_unlock_irq(&mchdev_lock);
6711
6712         return val;
6713 }
6714
6715 /**
6716  * i915_read_mch_val - return value for IPS use
6717  *
6718  * Calculate and return a value for the IPS driver to use when deciding whether
6719  * we have thermal and power headroom to increase CPU or GPU power budget.
6720  */
6721 unsigned long i915_read_mch_val(void)
6722 {
6723         struct drm_i915_private *dev_priv;
6724         unsigned long chipset_val, graphics_val, ret = 0;
6725
6726         spin_lock_irq(&mchdev_lock);
6727         if (!i915_mch_dev)
6728                 goto out_unlock;
6729         dev_priv = i915_mch_dev;
6730
6731         chipset_val = __i915_chipset_val(dev_priv);
6732         graphics_val = __i915_gfx_val(dev_priv);
6733
6734         ret = chipset_val + graphics_val;
6735
6736 out_unlock:
6737         spin_unlock_irq(&mchdev_lock);
6738
6739         return ret;
6740 }
6741 EXPORT_SYMBOL_GPL(i915_read_mch_val);
6742
6743 /**
6744  * i915_gpu_raise - raise GPU frequency limit
6745  *
6746  * Raise the limit; IPS indicates we have thermal headroom.
6747  */
6748 bool i915_gpu_raise(void)
6749 {
6750         struct drm_i915_private *dev_priv;
6751         bool ret = true;
6752
6753         spin_lock_irq(&mchdev_lock);
6754         if (!i915_mch_dev) {
6755                 ret = false;
6756                 goto out_unlock;
6757         }
6758         dev_priv = i915_mch_dev;
6759
6760         if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
6761                 dev_priv->ips.max_delay--;
6762
6763 out_unlock:
6764         spin_unlock_irq(&mchdev_lock);
6765
6766         return ret;
6767 }
6768 EXPORT_SYMBOL_GPL(i915_gpu_raise);
6769
6770 /**
6771  * i915_gpu_lower - lower GPU frequency limit
6772  *
6773  * IPS indicates we're close to a thermal limit, so throttle back the GPU
6774  * frequency maximum.
6775  */
6776 bool i915_gpu_lower(void)
6777 {
6778         struct drm_i915_private *dev_priv;
6779         bool ret = true;
6780
6781         spin_lock_irq(&mchdev_lock);
6782         if (!i915_mch_dev) {
6783                 ret = false;
6784                 goto out_unlock;
6785         }
6786         dev_priv = i915_mch_dev;
6787
6788         if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
6789                 dev_priv->ips.max_delay++;
6790
6791 out_unlock:
6792         spin_unlock_irq(&mchdev_lock);
6793
6794         return ret;
6795 }
6796 EXPORT_SYMBOL_GPL(i915_gpu_lower);
6797
6798 /**
6799  * i915_gpu_busy - indicate GPU business to IPS
6800  *
6801  * Tell the IPS driver whether or not the GPU is busy.
6802  */
6803 bool i915_gpu_busy(void)
6804 {
6805         bool ret = false;
6806
6807         spin_lock_irq(&mchdev_lock);
6808         if (i915_mch_dev)
6809                 ret = i915_mch_dev->gt.awake;
6810         spin_unlock_irq(&mchdev_lock);
6811
6812         return ret;
6813 }
6814 EXPORT_SYMBOL_GPL(i915_gpu_busy);
6815
6816 /**
6817  * i915_gpu_turbo_disable - disable graphics turbo
6818  *
6819  * Disable graphics turbo by resetting the max frequency and setting the
6820  * current frequency to the default.
6821  */
6822 bool i915_gpu_turbo_disable(void)
6823 {
6824         struct drm_i915_private *dev_priv;
6825         bool ret = true;
6826
6827         spin_lock_irq(&mchdev_lock);
6828         if (!i915_mch_dev) {
6829                 ret = false;
6830                 goto out_unlock;
6831         }
6832         dev_priv = i915_mch_dev;
6833
6834         dev_priv->ips.max_delay = dev_priv->ips.fstart;
6835
6836         if (!ironlake_set_drps(dev_priv, dev_priv->ips.fstart))
6837                 ret = false;
6838
6839 out_unlock:
6840         spin_unlock_irq(&mchdev_lock);
6841
6842         return ret;
6843 }
6844 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
6845
6846 /**
6847  * Tells the intel_ips driver that the i915 driver is now loaded, if
6848  * IPS got loaded first.
6849  *
6850  * This awkward dance is so that neither module has to depend on the
6851  * other in order for IPS to do the appropriate communication of
6852  * GPU turbo limits to i915.
6853  */
6854 static void
6855 ips_ping_for_i915_load(void)
6856 {
6857         void (*link)(void);
6858
6859         link = symbol_get(ips_link_to_i915_driver);
6860         if (link) {
6861                 link();
6862                 symbol_put(ips_link_to_i915_driver);
6863         }
6864 }
6865
6866 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
6867 {
6868         /* We only register the i915 ips part with intel-ips once everything is
6869          * set up, to avoid intel-ips sneaking in and reading bogus values. */
6870         spin_lock_irq(&mchdev_lock);
6871         i915_mch_dev = dev_priv;
6872         spin_unlock_irq(&mchdev_lock);
6873
6874         ips_ping_for_i915_load();
6875 }
6876
6877 void intel_gpu_ips_teardown(void)
6878 {
6879         spin_lock_irq(&mchdev_lock);
6880         i915_mch_dev = NULL;
6881         spin_unlock_irq(&mchdev_lock);
6882 }
6883
6884 static void intel_init_emon(struct drm_i915_private *dev_priv)
6885 {
6886         u32 lcfuse;
6887         u8 pxw[16];
6888         int i;
6889
6890         /* Disable to program */
6891         I915_WRITE(ECR, 0);
6892         POSTING_READ(ECR);
6893
6894         /* Program energy weights for various events */
6895         I915_WRITE(SDEW, 0x15040d00);
6896         I915_WRITE(CSIEW0, 0x007f0000);
6897         I915_WRITE(CSIEW1, 0x1e220004);
6898         I915_WRITE(CSIEW2, 0x04000004);
6899
6900         for (i = 0; i < 5; i++)
6901                 I915_WRITE(PEW(i), 0);
6902         for (i = 0; i < 3; i++)
6903                 I915_WRITE(DEW(i), 0);
6904
6905         /* Program P-state weights to account for frequency power adjustment */
6906         for (i = 0; i < 16; i++) {
6907                 u32 pxvidfreq = I915_READ(PXVFREQ(i));
6908                 unsigned long freq = intel_pxfreq(pxvidfreq);
6909                 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
6910                         PXVFREQ_PX_SHIFT;
6911                 unsigned long val;
6912
6913                 val = vid * vid;
6914                 val *= (freq / 1000);
6915                 val *= 255;
6916                 val /= (127*127*900);
6917                 if (val > 0xff)
6918                         DRM_ERROR("bad pxval: %ld\n", val);
6919                 pxw[i] = val;
6920         }
6921         /* Render standby states get 0 weight */
6922         pxw[14] = 0;
6923         pxw[15] = 0;
6924
6925         for (i = 0; i < 4; i++) {
6926                 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
6927                         (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
6928                 I915_WRITE(PXW(i), val);
6929         }
6930
6931         /* Adjust magic regs to magic values (more experimental results) */
6932         I915_WRITE(OGW0, 0);
6933         I915_WRITE(OGW1, 0);
6934         I915_WRITE(EG0, 0x00007f00);
6935         I915_WRITE(EG1, 0x0000000e);
6936         I915_WRITE(EG2, 0x000e0000);
6937         I915_WRITE(EG3, 0x68000300);
6938         I915_WRITE(EG4, 0x42000000);
6939         I915_WRITE(EG5, 0x00140031);
6940         I915_WRITE(EG6, 0);
6941         I915_WRITE(EG7, 0);
6942
6943         for (i = 0; i < 8; i++)
6944                 I915_WRITE(PXWL(i), 0);
6945
6946         /* Enable PMON + select events */
6947         I915_WRITE(ECR, 0x80000019);
6948
6949         lcfuse = I915_READ(LCFUSE02);
6950
6951         dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
6952 }
6953
6954 void intel_init_gt_powersave(struct drm_i915_private *dev_priv)
6955 {
6956         /*
6957          * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
6958          * requirement.
6959          */
6960         if (!i915.enable_rc6) {
6961                 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
6962                 intel_runtime_pm_get(dev_priv);
6963         }
6964
6965         mutex_lock(&dev_priv->drm.struct_mutex);
6966         mutex_lock(&dev_priv->rps.hw_lock);
6967
6968         /* Initialize RPS limits (for userspace) */
6969         if (IS_CHERRYVIEW(dev_priv))
6970                 cherryview_init_gt_powersave(dev_priv);
6971         else if (IS_VALLEYVIEW(dev_priv))
6972                 valleyview_init_gt_powersave(dev_priv);
6973         else if (INTEL_GEN(dev_priv) >= 6)
6974                 gen6_init_rps_frequencies(dev_priv);
6975
6976         /* Derive initial user preferences/limits from the hardware limits */
6977         dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
6978         dev_priv->rps.cur_freq = dev_priv->rps.idle_freq;
6979
6980         dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
6981         dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
6982
6983         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
6984                 dev_priv->rps.min_freq_softlimit =
6985                         max_t(int,
6986                               dev_priv->rps.efficient_freq,
6987                               intel_freq_opcode(dev_priv, 450));
6988
6989         /* After setting max-softlimit, find the overclock max freq */
6990         if (IS_GEN6(dev_priv) ||
6991             IS_IVYBRIDGE(dev_priv) || IS_HASWELL(dev_priv)) {
6992                 u32 params = 0;
6993
6994                 sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &params);
6995                 if (params & BIT(31)) { /* OC supported */
6996                         DRM_DEBUG_DRIVER("Overclocking supported, max: %dMHz, overclock: %dMHz\n",
6997                                          (dev_priv->rps.max_freq & 0xff) * 50,
6998                                          (params & 0xff) * 50);
6999                         dev_priv->rps.max_freq = params & 0xff;
7000                 }
7001         }
7002
7003         /* Finally allow us to boost to max by default */
7004         dev_priv->rps.boost_freq = dev_priv->rps.max_freq;
7005
7006         mutex_unlock(&dev_priv->rps.hw_lock);
7007         mutex_unlock(&dev_priv->drm.struct_mutex);
7008
7009         intel_autoenable_gt_powersave(dev_priv);
7010 }
7011
7012 void intel_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
7013 {
7014         if (IS_VALLEYVIEW(dev_priv))
7015                 valleyview_cleanup_gt_powersave(dev_priv);
7016
7017         if (!i915.enable_rc6)
7018                 intel_runtime_pm_put(dev_priv);
7019 }
7020
7021 /**
7022  * intel_suspend_gt_powersave - suspend PM work and helper threads
7023  * @dev_priv: i915 device
7024  *
7025  * We don't want to disable RC6 or other features here, we just want
7026  * to make sure any work we've queued has finished and won't bother
7027  * us while we're suspended.
7028  */
7029 void intel_suspend_gt_powersave(struct drm_i915_private *dev_priv)
7030 {
7031         if (INTEL_GEN(dev_priv) < 6)
7032                 return;
7033
7034         if (cancel_delayed_work_sync(&dev_priv->rps.autoenable_work))
7035                 intel_runtime_pm_put(dev_priv);
7036
7037         /* gen6_rps_idle() will be called later to disable interrupts */
7038 }
7039
7040 void intel_sanitize_gt_powersave(struct drm_i915_private *dev_priv)
7041 {
7042         dev_priv->rps.enabled = true; /* force disabling */
7043         intel_disable_gt_powersave(dev_priv);
7044
7045         gen6_reset_rps_interrupts(dev_priv);
7046 }
7047
7048 void intel_disable_gt_powersave(struct drm_i915_private *dev_priv)
7049 {
7050         if (!READ_ONCE(dev_priv->rps.enabled))
7051                 return;
7052
7053         mutex_lock(&dev_priv->rps.hw_lock);
7054
7055         if (INTEL_GEN(dev_priv) >= 9) {
7056                 gen9_disable_rc6(dev_priv);
7057                 gen9_disable_rps(dev_priv);
7058         } else if (IS_CHERRYVIEW(dev_priv)) {
7059                 cherryview_disable_rps(dev_priv);
7060         } else if (IS_VALLEYVIEW(dev_priv)) {
7061                 valleyview_disable_rps(dev_priv);
7062         } else if (INTEL_GEN(dev_priv) >= 6) {
7063                 gen6_disable_rps(dev_priv);
7064         }  else if (IS_IRONLAKE_M(dev_priv)) {
7065                 ironlake_disable_drps(dev_priv);
7066         }
7067
7068         dev_priv->rps.enabled = false;
7069         mutex_unlock(&dev_priv->rps.hw_lock);
7070 }
7071
7072 void intel_enable_gt_powersave(struct drm_i915_private *dev_priv)
7073 {
7074         /* We shouldn't be disabling as we submit, so this should be less
7075          * racy than it appears!
7076          */
7077         if (READ_ONCE(dev_priv->rps.enabled))
7078                 return;
7079
7080         /* Powersaving is controlled by the host when inside a VM */
7081         if (intel_vgpu_active(dev_priv))
7082                 return;
7083
7084         mutex_lock(&dev_priv->rps.hw_lock);
7085
7086         if (IS_CHERRYVIEW(dev_priv)) {
7087                 cherryview_enable_rps(dev_priv);
7088         } else if (IS_VALLEYVIEW(dev_priv)) {
7089                 valleyview_enable_rps(dev_priv);
7090         } else if (INTEL_GEN(dev_priv) >= 9) {
7091                 gen9_enable_rc6(dev_priv);
7092                 gen9_enable_rps(dev_priv);
7093                 if (IS_GEN9_BC(dev_priv))
7094                         gen6_update_ring_freq(dev_priv);
7095         } else if (IS_BROADWELL(dev_priv)) {
7096                 gen8_enable_rps(dev_priv);
7097                 gen6_update_ring_freq(dev_priv);
7098         } else if (INTEL_GEN(dev_priv) >= 6) {
7099                 gen6_enable_rps(dev_priv);
7100                 gen6_update_ring_freq(dev_priv);
7101         } else if (IS_IRONLAKE_M(dev_priv)) {
7102                 ironlake_enable_drps(dev_priv);
7103                 intel_init_emon(dev_priv);
7104         }
7105
7106         WARN_ON(dev_priv->rps.max_freq < dev_priv->rps.min_freq);
7107         WARN_ON(dev_priv->rps.idle_freq > dev_priv->rps.max_freq);
7108
7109         WARN_ON(dev_priv->rps.efficient_freq < dev_priv->rps.min_freq);
7110         WARN_ON(dev_priv->rps.efficient_freq > dev_priv->rps.max_freq);
7111
7112         dev_priv->rps.enabled = true;
7113         mutex_unlock(&dev_priv->rps.hw_lock);
7114 }
7115
7116 static void __intel_autoenable_gt_powersave(struct work_struct *work)
7117 {
7118         struct drm_i915_private *dev_priv =
7119                 container_of(work, typeof(*dev_priv), rps.autoenable_work.work);
7120         struct intel_engine_cs *rcs;
7121         struct drm_i915_gem_request *req;
7122
7123         if (READ_ONCE(dev_priv->rps.enabled))
7124                 goto out;
7125
7126         rcs = dev_priv->engine[RCS];
7127         if (rcs->last_retired_context)
7128                 goto out;
7129
7130         if (!rcs->init_context)
7131                 goto out;
7132
7133         mutex_lock(&dev_priv->drm.struct_mutex);
7134
7135         req = i915_gem_request_alloc(rcs, dev_priv->kernel_context);
7136         if (IS_ERR(req))
7137                 goto unlock;
7138
7139         if (!i915.enable_execlists && i915_switch_context(req) == 0)
7140                 rcs->init_context(req);
7141
7142         /* Mark the device busy, calling intel_enable_gt_powersave() */
7143         i915_add_request(req);
7144
7145 unlock:
7146         mutex_unlock(&dev_priv->drm.struct_mutex);
7147 out:
7148         intel_runtime_pm_put(dev_priv);
7149 }
7150
7151 void intel_autoenable_gt_powersave(struct drm_i915_private *dev_priv)
7152 {
7153         if (READ_ONCE(dev_priv->rps.enabled))
7154                 return;
7155
7156         if (IS_IRONLAKE_M(dev_priv)) {
7157                 ironlake_enable_drps(dev_priv);
7158                 intel_init_emon(dev_priv);
7159         } else if (INTEL_INFO(dev_priv)->gen >= 6) {
7160                 /*
7161                  * PCU communication is slow and this doesn't need to be
7162                  * done at any specific time, so do this out of our fast path
7163                  * to make resume and init faster.
7164                  *
7165                  * We depend on the HW RC6 power context save/restore
7166                  * mechanism when entering D3 through runtime PM suspend. So
7167                  * disable RPM until RPS/RC6 is properly setup. We can only
7168                  * get here via the driver load/system resume/runtime resume
7169                  * paths, so the _noresume version is enough (and in case of
7170                  * runtime resume it's necessary).
7171                  */
7172                 if (queue_delayed_work(dev_priv->wq,
7173                                        &dev_priv->rps.autoenable_work,
7174                                        round_jiffies_up_relative(HZ)))
7175                         intel_runtime_pm_get_noresume(dev_priv);
7176         }
7177 }
7178
7179 static void ibx_init_clock_gating(struct drm_i915_private *dev_priv)
7180 {
7181         /*
7182          * On Ibex Peak and Cougar Point, we need to disable clock
7183          * gating for the panel power sequencer or it will fail to
7184          * start up when no ports are active.
7185          */
7186         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
7187 }
7188
7189 static void g4x_disable_trickle_feed(struct drm_i915_private *dev_priv)
7190 {
7191         enum pipe pipe;
7192
7193         for_each_pipe(dev_priv, pipe) {
7194                 I915_WRITE(DSPCNTR(pipe),
7195                            I915_READ(DSPCNTR(pipe)) |
7196                            DISPPLANE_TRICKLE_FEED_DISABLE);
7197
7198                 I915_WRITE(DSPSURF(pipe), I915_READ(DSPSURF(pipe)));
7199                 POSTING_READ(DSPSURF(pipe));
7200         }
7201 }
7202
7203 static void ilk_init_lp_watermarks(struct drm_i915_private *dev_priv)
7204 {
7205         I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
7206         I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
7207         I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
7208
7209         /*
7210          * Don't touch WM1S_LP_EN here.
7211          * Doing so could cause underruns.
7212          */
7213 }
7214
7215 static void ironlake_init_clock_gating(struct drm_i915_private *dev_priv)
7216 {
7217         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
7218
7219         /*
7220          * Required for FBC
7221          * WaFbcDisableDpfcClockGating:ilk
7222          */
7223         dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
7224                    ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
7225                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
7226
7227         I915_WRITE(PCH_3DCGDIS0,
7228                    MARIUNIT_CLOCK_GATE_DISABLE |
7229                    SVSMUNIT_CLOCK_GATE_DISABLE);
7230         I915_WRITE(PCH_3DCGDIS1,
7231                    VFMUNIT_CLOCK_GATE_DISABLE);
7232
7233         /*
7234          * According to the spec the following bits should be set in
7235          * order to enable memory self-refresh
7236          * The bit 22/21 of 0x42004
7237          * The bit 5 of 0x42020
7238          * The bit 15 of 0x45000
7239          */
7240         I915_WRITE(ILK_DISPLAY_CHICKEN2,
7241                    (I915_READ(ILK_DISPLAY_CHICKEN2) |
7242                     ILK_DPARB_GATE | ILK_VSDPFD_FULL));
7243         dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
7244         I915_WRITE(DISP_ARB_CTL,
7245                    (I915_READ(DISP_ARB_CTL) |
7246                     DISP_FBC_WM_DIS));
7247
7248         ilk_init_lp_watermarks(dev_priv);
7249
7250         /*
7251          * Based on the document from hardware guys the following bits
7252          * should be set unconditionally in order to enable FBC.
7253          * The bit 22 of 0x42000
7254          * The bit 22 of 0x42004
7255          * The bit 7,8,9 of 0x42020.
7256          */
7257         if (IS_IRONLAKE_M(dev_priv)) {
7258                 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
7259                 I915_WRITE(ILK_DISPLAY_CHICKEN1,
7260                            I915_READ(ILK_DISPLAY_CHICKEN1) |
7261                            ILK_FBCQ_DIS);
7262                 I915_WRITE(ILK_DISPLAY_CHICKEN2,
7263                            I915_READ(ILK_DISPLAY_CHICKEN2) |
7264                            ILK_DPARB_GATE);
7265         }
7266
7267         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
7268
7269         I915_WRITE(ILK_DISPLAY_CHICKEN2,
7270                    I915_READ(ILK_DISPLAY_CHICKEN2) |
7271                    ILK_ELPIN_409_SELECT);
7272         I915_WRITE(_3D_CHICKEN2,
7273                    _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
7274                    _3D_CHICKEN2_WM_READ_PIPELINED);
7275
7276         /* WaDisableRenderCachePipelinedFlush:ilk */
7277         I915_WRITE(CACHE_MODE_0,
7278                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
7279
7280         /* WaDisable_RenderCache_OperationalFlush:ilk */
7281         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7282
7283         g4x_disable_trickle_feed(dev_priv);
7284
7285         ibx_init_clock_gating(dev_priv);
7286 }
7287
7288 static void cpt_init_clock_gating(struct drm_i915_private *dev_priv)
7289 {
7290         int pipe;
7291         uint32_t val;
7292
7293         /*
7294          * On Ibex Peak and Cougar Point, we need to disable clock
7295          * gating for the panel power sequencer or it will fail to
7296          * start up when no ports are active.
7297          */
7298         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
7299                    PCH_DPLUNIT_CLOCK_GATE_DISABLE |
7300                    PCH_CPUNIT_CLOCK_GATE_DISABLE);
7301         I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
7302                    DPLS_EDP_PPS_FIX_DIS);
7303         /* The below fixes the weird display corruption, a few pixels shifted
7304          * downward, on (only) LVDS of some HP laptops with IVY.
7305          */
7306         for_each_pipe(dev_priv, pipe) {
7307                 val = I915_READ(TRANS_CHICKEN2(pipe));
7308                 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
7309                 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
7310                 if (dev_priv->vbt.fdi_rx_polarity_inverted)
7311                         val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
7312                 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
7313                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
7314                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
7315                 I915_WRITE(TRANS_CHICKEN2(pipe), val);
7316         }
7317         /* WADP0ClockGatingDisable */
7318         for_each_pipe(dev_priv, pipe) {
7319                 I915_WRITE(TRANS_CHICKEN1(pipe),
7320                            TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
7321         }
7322 }
7323
7324 static void gen6_check_mch_setup(struct drm_i915_private *dev_priv)
7325 {
7326         uint32_t tmp;
7327
7328         tmp = I915_READ(MCH_SSKPD);
7329         if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL)
7330                 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
7331                               tmp);
7332 }
7333
7334 static void gen6_init_clock_gating(struct drm_i915_private *dev_priv)
7335 {
7336         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
7337
7338         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
7339
7340         I915_WRITE(ILK_DISPLAY_CHICKEN2,
7341                    I915_READ(ILK_DISPLAY_CHICKEN2) |
7342                    ILK_ELPIN_409_SELECT);
7343
7344         /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
7345         I915_WRITE(_3D_CHICKEN,
7346                    _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
7347
7348         /* WaDisable_RenderCache_OperationalFlush:snb */
7349         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7350
7351         /*
7352          * BSpec recoomends 8x4 when MSAA is used,
7353          * however in practice 16x4 seems fastest.
7354          *
7355          * Note that PS/WM thread counts depend on the WIZ hashing
7356          * disable bit, which we don't touch here, but it's good
7357          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7358          */
7359         I915_WRITE(GEN6_GT_MODE,
7360                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7361
7362         ilk_init_lp_watermarks(dev_priv);
7363
7364         I915_WRITE(CACHE_MODE_0,
7365                    _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
7366
7367         I915_WRITE(GEN6_UCGCTL1,
7368                    I915_READ(GEN6_UCGCTL1) |
7369                    GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
7370                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
7371
7372         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
7373          * gating disable must be set.  Failure to set it results in
7374          * flickering pixels due to Z write ordering failures after
7375          * some amount of runtime in the Mesa "fire" demo, and Unigine
7376          * Sanctuary and Tropics, and apparently anything else with
7377          * alpha test or pixel discard.
7378          *
7379          * According to the spec, bit 11 (RCCUNIT) must also be set,
7380          * but we didn't debug actual testcases to find it out.
7381          *
7382          * WaDisableRCCUnitClockGating:snb
7383          * WaDisableRCPBUnitClockGating:snb
7384          */
7385         I915_WRITE(GEN6_UCGCTL2,
7386                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
7387                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
7388
7389         /* WaStripsFansDisableFastClipPerformanceFix:snb */
7390         I915_WRITE(_3D_CHICKEN3,
7391                    _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
7392
7393         /*
7394          * Bspec says:
7395          * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
7396          * 3DSTATE_SF number of SF output attributes is more than 16."
7397          */
7398         I915_WRITE(_3D_CHICKEN3,
7399                    _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
7400
7401         /*
7402          * According to the spec the following bits should be
7403          * set in order to enable memory self-refresh and fbc:
7404          * The bit21 and bit22 of 0x42000
7405          * The bit21 and bit22 of 0x42004
7406          * The bit5 and bit7 of 0x42020
7407          * The bit14 of 0x70180
7408          * The bit14 of 0x71180
7409          *
7410          * WaFbcAsynchFlipDisableFbcQueue:snb
7411          */
7412         I915_WRITE(ILK_DISPLAY_CHICKEN1,
7413                    I915_READ(ILK_DISPLAY_CHICKEN1) |
7414                    ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
7415         I915_WRITE(ILK_DISPLAY_CHICKEN2,
7416                    I915_READ(ILK_DISPLAY_CHICKEN2) |
7417                    ILK_DPARB_GATE | ILK_VSDPFD_FULL);
7418         I915_WRITE(ILK_DSPCLK_GATE_D,
7419                    I915_READ(ILK_DSPCLK_GATE_D) |
7420                    ILK_DPARBUNIT_CLOCK_GATE_ENABLE  |
7421                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
7422
7423         g4x_disable_trickle_feed(dev_priv);
7424
7425         cpt_init_clock_gating(dev_priv);
7426
7427         gen6_check_mch_setup(dev_priv);
7428 }
7429
7430 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
7431 {
7432         uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
7433
7434         /*
7435          * WaVSThreadDispatchOverride:ivb,vlv
7436          *
7437          * This actually overrides the dispatch
7438          * mode for all thread types.
7439          */
7440         reg &= ~GEN7_FF_SCHED_MASK;
7441         reg |= GEN7_FF_TS_SCHED_HW;
7442         reg |= GEN7_FF_VS_SCHED_HW;
7443         reg |= GEN7_FF_DS_SCHED_HW;
7444
7445         I915_WRITE(GEN7_FF_THREAD_MODE, reg);
7446 }
7447
7448 static void lpt_init_clock_gating(struct drm_i915_private *dev_priv)
7449 {
7450         /*
7451          * TODO: this bit should only be enabled when really needed, then
7452          * disabled when not needed anymore in order to save power.
7453          */
7454         if (HAS_PCH_LPT_LP(dev_priv))
7455                 I915_WRITE(SOUTH_DSPCLK_GATE_D,
7456                            I915_READ(SOUTH_DSPCLK_GATE_D) |
7457                            PCH_LP_PARTITION_LEVEL_DISABLE);
7458
7459         /* WADPOClockGatingDisable:hsw */
7460         I915_WRITE(TRANS_CHICKEN1(PIPE_A),
7461                    I915_READ(TRANS_CHICKEN1(PIPE_A)) |
7462                    TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
7463 }
7464
7465 static void lpt_suspend_hw(struct drm_i915_private *dev_priv)
7466 {
7467         if (HAS_PCH_LPT_LP(dev_priv)) {
7468                 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
7469
7470                 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
7471                 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
7472         }
7473 }
7474
7475 static void gen8_set_l3sqc_credits(struct drm_i915_private *dev_priv,
7476                                    int general_prio_credits,
7477                                    int high_prio_credits)
7478 {
7479         u32 misccpctl;
7480
7481         /* WaTempDisableDOPClkGating:bdw */
7482         misccpctl = I915_READ(GEN7_MISCCPCTL);
7483         I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
7484
7485         I915_WRITE(GEN8_L3SQCREG1,
7486                    L3_GENERAL_PRIO_CREDITS(general_prio_credits) |
7487                    L3_HIGH_PRIO_CREDITS(high_prio_credits));
7488
7489         /*
7490          * Wait at least 100 clocks before re-enabling clock gating.
7491          * See the definition of L3SQCREG1 in BSpec.
7492          */
7493         POSTING_READ(GEN8_L3SQCREG1);
7494         udelay(1);
7495         I915_WRITE(GEN7_MISCCPCTL, misccpctl);
7496 }
7497
7498 static void kabylake_init_clock_gating(struct drm_i915_private *dev_priv)
7499 {
7500         gen9_init_clock_gating(dev_priv);
7501
7502         /* WaDisableSDEUnitClockGating:kbl */
7503         if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
7504                 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7505                            GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7506
7507         /* WaDisableGamClockGating:kbl */
7508         if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
7509                 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
7510                            GEN6_GAMUNIT_CLOCK_GATE_DISABLE);
7511
7512         /* WaFbcNukeOnHostModify:kbl */
7513         I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
7514                    ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
7515 }
7516
7517 static void skylake_init_clock_gating(struct drm_i915_private *dev_priv)
7518 {
7519         gen9_init_clock_gating(dev_priv);
7520
7521         /* WAC6entrylatency:skl */
7522         I915_WRITE(FBC_LLC_READ_CTRL, I915_READ(FBC_LLC_READ_CTRL) |
7523                    FBC_LLC_FULLY_OPEN);
7524
7525         /* WaFbcNukeOnHostModify:skl */
7526         I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
7527                    ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
7528 }
7529
7530 static void broadwell_init_clock_gating(struct drm_i915_private *dev_priv)
7531 {
7532         enum pipe pipe;
7533
7534         ilk_init_lp_watermarks(dev_priv);
7535
7536         /* WaSwitchSolVfFArbitrationPriority:bdw */
7537         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
7538
7539         /* WaPsrDPAMaskVBlankInSRD:bdw */
7540         I915_WRITE(CHICKEN_PAR1_1,
7541                    I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
7542
7543         /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
7544         for_each_pipe(dev_priv, pipe) {
7545                 I915_WRITE(CHICKEN_PIPESL_1(pipe),
7546                            I915_READ(CHICKEN_PIPESL_1(pipe)) |
7547                            BDW_DPRS_MASK_VBLANK_SRD);
7548         }
7549
7550         /* WaVSRefCountFullforceMissDisable:bdw */
7551         /* WaDSRefCountFullforceMissDisable:bdw */
7552         I915_WRITE(GEN7_FF_THREAD_MODE,
7553                    I915_READ(GEN7_FF_THREAD_MODE) &
7554                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
7555
7556         I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
7557                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
7558
7559         /* WaDisableSDEUnitClockGating:bdw */
7560         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7561                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7562
7563         /* WaProgramL3SqcReg1Default:bdw */
7564         gen8_set_l3sqc_credits(dev_priv, 30, 2);
7565
7566         /*
7567          * WaGttCachingOffByDefault:bdw
7568          * GTT cache may not work with big pages, so if those
7569          * are ever enabled GTT cache may need to be disabled.
7570          */
7571         I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
7572
7573         /* WaKVMNotificationOnConfigChange:bdw */
7574         I915_WRITE(CHICKEN_PAR2_1, I915_READ(CHICKEN_PAR2_1)
7575                    | KVM_CONFIG_CHANGE_NOTIFICATION_SELECT);
7576
7577         lpt_init_clock_gating(dev_priv);
7578
7579         /* WaDisableDopClockGating:bdw
7580          *
7581          * Also see the CHICKEN2 write in bdw_init_workarounds() to disable DOP
7582          * clock gating.
7583          */
7584         I915_WRITE(GEN6_UCGCTL1,
7585                    I915_READ(GEN6_UCGCTL1) | GEN6_EU_TCUNIT_CLOCK_GATE_DISABLE);
7586 }
7587
7588 static void haswell_init_clock_gating(struct drm_i915_private *dev_priv)
7589 {
7590         ilk_init_lp_watermarks(dev_priv);
7591
7592         /* L3 caching of data atomics doesn't work -- disable it. */
7593         I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
7594         I915_WRITE(HSW_ROW_CHICKEN3,
7595                    _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
7596
7597         /* This is required by WaCatErrorRejectionIssue:hsw */
7598         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7599                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7600                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7601
7602         /* WaVSRefCountFullforceMissDisable:hsw */
7603         I915_WRITE(GEN7_FF_THREAD_MODE,
7604                    I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
7605
7606         /* WaDisable_RenderCache_OperationalFlush:hsw */
7607         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7608
7609         /* enable HiZ Raw Stall Optimization */
7610         I915_WRITE(CACHE_MODE_0_GEN7,
7611                    _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
7612
7613         /* WaDisable4x2SubspanOptimization:hsw */
7614         I915_WRITE(CACHE_MODE_1,
7615                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7616
7617         /*
7618          * BSpec recommends 8x4 when MSAA is used,
7619          * however in practice 16x4 seems fastest.
7620          *
7621          * Note that PS/WM thread counts depend on the WIZ hashing
7622          * disable bit, which we don't touch here, but it's good
7623          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7624          */
7625         I915_WRITE(GEN7_GT_MODE,
7626                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7627
7628         /* WaSampleCChickenBitEnable:hsw */
7629         I915_WRITE(HALF_SLICE_CHICKEN3,
7630                    _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE));
7631
7632         /* WaSwitchSolVfFArbitrationPriority:hsw */
7633         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
7634
7635         /* WaRsPkgCStateDisplayPMReq:hsw */
7636         I915_WRITE(CHICKEN_PAR1_1,
7637                    I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
7638
7639         lpt_init_clock_gating(dev_priv);
7640 }
7641
7642 static void ivybridge_init_clock_gating(struct drm_i915_private *dev_priv)
7643 {
7644         uint32_t snpcr;
7645
7646         ilk_init_lp_watermarks(dev_priv);
7647
7648         I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
7649
7650         /* WaDisableEarlyCull:ivb */
7651         I915_WRITE(_3D_CHICKEN3,
7652                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
7653
7654         /* WaDisableBackToBackFlipFix:ivb */
7655         I915_WRITE(IVB_CHICKEN3,
7656                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7657                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
7658
7659         /* WaDisablePSDDualDispatchEnable:ivb */
7660         if (IS_IVB_GT1(dev_priv))
7661                 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
7662                            _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
7663
7664         /* WaDisable_RenderCache_OperationalFlush:ivb */
7665         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7666
7667         /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
7668         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
7669                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
7670
7671         /* WaApplyL3ControlAndL3ChickenMode:ivb */
7672         I915_WRITE(GEN7_L3CNTLREG1,
7673                         GEN7_WA_FOR_GEN7_L3_CONTROL);
7674         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
7675                    GEN7_WA_L3_CHICKEN_MODE);
7676         if (IS_IVB_GT1(dev_priv))
7677                 I915_WRITE(GEN7_ROW_CHICKEN2,
7678                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7679         else {
7680                 /* must write both registers */
7681                 I915_WRITE(GEN7_ROW_CHICKEN2,
7682                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7683                 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
7684                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7685         }
7686
7687         /* WaForceL3Serialization:ivb */
7688         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
7689                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
7690
7691         /*
7692          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7693          * This implements the WaDisableRCZUnitClockGating:ivb workaround.
7694          */
7695         I915_WRITE(GEN6_UCGCTL2,
7696                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7697
7698         /* This is required by WaCatErrorRejectionIssue:ivb */
7699         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7700                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7701                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7702
7703         g4x_disable_trickle_feed(dev_priv);
7704
7705         gen7_setup_fixed_func_scheduler(dev_priv);
7706
7707         if (0) { /* causes HiZ corruption on ivb:gt1 */
7708                 /* enable HiZ Raw Stall Optimization */
7709                 I915_WRITE(CACHE_MODE_0_GEN7,
7710                            _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
7711         }
7712
7713         /* WaDisable4x2SubspanOptimization:ivb */
7714         I915_WRITE(CACHE_MODE_1,
7715                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7716
7717         /*
7718          * BSpec recommends 8x4 when MSAA is used,
7719          * however in practice 16x4 seems fastest.
7720          *
7721          * Note that PS/WM thread counts depend on the WIZ hashing
7722          * disable bit, which we don't touch here, but it's good
7723          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7724          */
7725         I915_WRITE(GEN7_GT_MODE,
7726                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7727
7728         snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
7729         snpcr &= ~GEN6_MBC_SNPCR_MASK;
7730         snpcr |= GEN6_MBC_SNPCR_MED;
7731         I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
7732
7733         if (!HAS_PCH_NOP(dev_priv))
7734                 cpt_init_clock_gating(dev_priv);
7735
7736         gen6_check_mch_setup(dev_priv);
7737 }
7738
7739 static void valleyview_init_clock_gating(struct drm_i915_private *dev_priv)
7740 {
7741         /* WaDisableEarlyCull:vlv */
7742         I915_WRITE(_3D_CHICKEN3,
7743                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
7744
7745         /* WaDisableBackToBackFlipFix:vlv */
7746         I915_WRITE(IVB_CHICKEN3,
7747                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7748                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
7749
7750         /* WaPsdDispatchEnable:vlv */
7751         /* WaDisablePSDDualDispatchEnable:vlv */
7752         I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
7753                    _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
7754                                       GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
7755
7756         /* WaDisable_RenderCache_OperationalFlush:vlv */
7757         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7758
7759         /* WaForceL3Serialization:vlv */
7760         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
7761                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
7762
7763         /* WaDisableDopClockGating:vlv */
7764         I915_WRITE(GEN7_ROW_CHICKEN2,
7765                    _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7766
7767         /* This is required by WaCatErrorRejectionIssue:vlv */
7768         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7769                    I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7770                    GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7771
7772         gen7_setup_fixed_func_scheduler(dev_priv);
7773
7774         /*
7775          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7776          * This implements the WaDisableRCZUnitClockGating:vlv workaround.
7777          */
7778         I915_WRITE(GEN6_UCGCTL2,
7779                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7780
7781         /* WaDisableL3Bank2xClockGate:vlv
7782          * Disabling L3 clock gating- MMIO 940c[25] = 1
7783          * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
7784         I915_WRITE(GEN7_UCGCTL4,
7785                    I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
7786
7787         /*
7788          * BSpec says this must be set, even though
7789          * WaDisable4x2SubspanOptimization isn't listed for VLV.
7790          */
7791         I915_WRITE(CACHE_MODE_1,
7792                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7793
7794         /*
7795          * BSpec recommends 8x4 when MSAA is used,
7796          * however in practice 16x4 seems fastest.
7797          *
7798          * Note that PS/WM thread counts depend on the WIZ hashing
7799          * disable bit, which we don't touch here, but it's good
7800          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7801          */
7802         I915_WRITE(GEN7_GT_MODE,
7803                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7804
7805         /*
7806          * WaIncreaseL3CreditsForVLVB0:vlv
7807          * This is the hardware default actually.
7808          */
7809         I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
7810
7811         /*
7812          * WaDisableVLVClockGating_VBIIssue:vlv
7813          * Disable clock gating on th GCFG unit to prevent a delay
7814          * in the reporting of vblank events.
7815          */
7816         I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
7817 }
7818
7819 static void cherryview_init_clock_gating(struct drm_i915_private *dev_priv)
7820 {
7821         /* WaVSRefCountFullforceMissDisable:chv */
7822         /* WaDSRefCountFullforceMissDisable:chv */
7823         I915_WRITE(GEN7_FF_THREAD_MODE,
7824                    I915_READ(GEN7_FF_THREAD_MODE) &
7825                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
7826
7827         /* WaDisableSemaphoreAndSyncFlipWait:chv */
7828         I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
7829                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
7830
7831         /* WaDisableCSUnitClockGating:chv */
7832         I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
7833                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
7834
7835         /* WaDisableSDEUnitClockGating:chv */
7836         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7837                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7838
7839         /*
7840          * WaProgramL3SqcReg1Default:chv
7841          * See gfxspecs/Related Documents/Performance Guide/
7842          * LSQC Setting Recommendations.
7843          */
7844         gen8_set_l3sqc_credits(dev_priv, 38, 2);
7845
7846         /*
7847          * GTT cache may not work with big pages, so if those
7848          * are ever enabled GTT cache may need to be disabled.
7849          */
7850         I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
7851 }
7852
7853 static void g4x_init_clock_gating(struct drm_i915_private *dev_priv)
7854 {
7855         uint32_t dspclk_gate;
7856
7857         I915_WRITE(RENCLK_GATE_D1, 0);
7858         I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
7859                    GS_UNIT_CLOCK_GATE_DISABLE |
7860                    CL_UNIT_CLOCK_GATE_DISABLE);
7861         I915_WRITE(RAMCLK_GATE_D, 0);
7862         dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
7863                 OVRUNIT_CLOCK_GATE_DISABLE |
7864                 OVCUNIT_CLOCK_GATE_DISABLE;
7865         if (IS_GM45(dev_priv))
7866                 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
7867         I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
7868
7869         /* WaDisableRenderCachePipelinedFlush */
7870         I915_WRITE(CACHE_MODE_0,
7871                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
7872
7873         /* WaDisable_RenderCache_OperationalFlush:g4x */
7874         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7875
7876         g4x_disable_trickle_feed(dev_priv);
7877 }
7878
7879 static void crestline_init_clock_gating(struct drm_i915_private *dev_priv)
7880 {
7881         I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
7882         I915_WRITE(RENCLK_GATE_D2, 0);
7883         I915_WRITE(DSPCLK_GATE_D, 0);
7884         I915_WRITE(RAMCLK_GATE_D, 0);
7885         I915_WRITE16(DEUC, 0);
7886         I915_WRITE(MI_ARB_STATE,
7887                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7888
7889         /* WaDisable_RenderCache_OperationalFlush:gen4 */
7890         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7891 }
7892
7893 static void broadwater_init_clock_gating(struct drm_i915_private *dev_priv)
7894 {
7895         I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
7896                    I965_RCC_CLOCK_GATE_DISABLE |
7897                    I965_RCPB_CLOCK_GATE_DISABLE |
7898                    I965_ISC_CLOCK_GATE_DISABLE |
7899                    I965_FBC_CLOCK_GATE_DISABLE);
7900         I915_WRITE(RENCLK_GATE_D2, 0);
7901         I915_WRITE(MI_ARB_STATE,
7902                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7903
7904         /* WaDisable_RenderCache_OperationalFlush:gen4 */
7905         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7906 }
7907
7908 static void gen3_init_clock_gating(struct drm_i915_private *dev_priv)
7909 {
7910         u32 dstate = I915_READ(D_STATE);
7911
7912         dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
7913                 DSTATE_DOT_CLOCK_GATING;
7914         I915_WRITE(D_STATE, dstate);
7915
7916         if (IS_PINEVIEW(dev_priv))
7917                 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
7918
7919         /* IIR "flip pending" means done if this bit is set */
7920         I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
7921
7922         /* interrupts should cause a wake up from C3 */
7923         I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
7924
7925         /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
7926         I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
7927
7928         I915_WRITE(MI_ARB_STATE,
7929                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7930 }
7931
7932 static void i85x_init_clock_gating(struct drm_i915_private *dev_priv)
7933 {
7934         I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
7935
7936         /* interrupts should cause a wake up from C3 */
7937         I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
7938                    _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
7939
7940         I915_WRITE(MEM_MODE,
7941                    _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE));
7942 }
7943
7944 static void i830_init_clock_gating(struct drm_i915_private *dev_priv)
7945 {
7946         I915_WRITE(MEM_MODE,
7947                    _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE) |
7948                    _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE));
7949 }
7950
7951 void intel_init_clock_gating(struct drm_i915_private *dev_priv)
7952 {
7953         dev_priv->display.init_clock_gating(dev_priv);
7954 }
7955
7956 void intel_suspend_hw(struct drm_i915_private *dev_priv)
7957 {
7958         if (HAS_PCH_LPT(dev_priv))
7959                 lpt_suspend_hw(dev_priv);
7960 }
7961
7962 static void nop_init_clock_gating(struct drm_i915_private *dev_priv)
7963 {
7964         DRM_DEBUG_KMS("No clock gating settings or workarounds applied.\n");
7965 }
7966
7967 /**
7968  * intel_init_clock_gating_hooks - setup the clock gating hooks
7969  * @dev_priv: device private
7970  *
7971  * Setup the hooks that configure which clocks of a given platform can be
7972  * gated and also apply various GT and display specific workarounds for these
7973  * platforms. Note that some GT specific workarounds are applied separately
7974  * when GPU contexts or batchbuffers start their execution.
7975  */
7976 void intel_init_clock_gating_hooks(struct drm_i915_private *dev_priv)
7977 {
7978         if (IS_SKYLAKE(dev_priv))
7979                 dev_priv->display.init_clock_gating = skylake_init_clock_gating;
7980         else if (IS_KABYLAKE(dev_priv))
7981                 dev_priv->display.init_clock_gating = kabylake_init_clock_gating;
7982         else if (IS_BROXTON(dev_priv))
7983                 dev_priv->display.init_clock_gating = bxt_init_clock_gating;
7984         else if (IS_GEMINILAKE(dev_priv))
7985                 dev_priv->display.init_clock_gating = glk_init_clock_gating;
7986         else if (IS_BROADWELL(dev_priv))
7987                 dev_priv->display.init_clock_gating = broadwell_init_clock_gating;
7988         else if (IS_CHERRYVIEW(dev_priv))
7989                 dev_priv->display.init_clock_gating = cherryview_init_clock_gating;
7990         else if (IS_HASWELL(dev_priv))
7991                 dev_priv->display.init_clock_gating = haswell_init_clock_gating;
7992         else if (IS_IVYBRIDGE(dev_priv))
7993                 dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
7994         else if (IS_VALLEYVIEW(dev_priv))
7995                 dev_priv->display.init_clock_gating = valleyview_init_clock_gating;
7996         else if (IS_GEN6(dev_priv))
7997                 dev_priv->display.init_clock_gating = gen6_init_clock_gating;
7998         else if (IS_GEN5(dev_priv))
7999                 dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
8000         else if (IS_G4X(dev_priv))
8001                 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
8002         else if (IS_I965GM(dev_priv))
8003                 dev_priv->display.init_clock_gating = crestline_init_clock_gating;
8004         else if (IS_I965G(dev_priv))
8005                 dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
8006         else if (IS_GEN3(dev_priv))
8007                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
8008         else if (IS_I85X(dev_priv) || IS_I865G(dev_priv))
8009                 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
8010         else if (IS_GEN2(dev_priv))
8011                 dev_priv->display.init_clock_gating = i830_init_clock_gating;
8012         else {
8013                 MISSING_CASE(INTEL_DEVID(dev_priv));
8014                 dev_priv->display.init_clock_gating = nop_init_clock_gating;
8015         }
8016 }
8017
8018 /* Set up chip specific power management-related functions */
8019 void intel_init_pm(struct drm_i915_private *dev_priv)
8020 {
8021         intel_fbc_init(dev_priv);
8022
8023         /* For cxsr */
8024         if (IS_PINEVIEW(dev_priv))
8025                 i915_pineview_get_mem_freq(dev_priv);
8026         else if (IS_GEN5(dev_priv))
8027                 i915_ironlake_get_mem_freq(dev_priv);
8028
8029         /* For FIFO watermark updates */
8030         if (INTEL_GEN(dev_priv) >= 9) {
8031                 skl_setup_wm_latency(dev_priv);
8032                 dev_priv->display.initial_watermarks = skl_initial_wm;
8033                 dev_priv->display.atomic_update_watermarks = skl_atomic_update_crtc_wm;
8034                 dev_priv->display.compute_global_watermarks = skl_compute_wm;
8035         } else if (HAS_PCH_SPLIT(dev_priv)) {
8036                 ilk_setup_wm_latency(dev_priv);
8037
8038                 if ((IS_GEN5(dev_priv) && dev_priv->wm.pri_latency[1] &&
8039                      dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
8040                     (!IS_GEN5(dev_priv) && dev_priv->wm.pri_latency[0] &&
8041                      dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
8042                         dev_priv->display.compute_pipe_wm = ilk_compute_pipe_wm;
8043                         dev_priv->display.compute_intermediate_wm =
8044                                 ilk_compute_intermediate_wm;
8045                         dev_priv->display.initial_watermarks =
8046                                 ilk_initial_watermarks;
8047                         dev_priv->display.optimize_watermarks =
8048                                 ilk_optimize_watermarks;
8049                 } else {
8050                         DRM_DEBUG_KMS("Failed to read display plane latency. "
8051                                       "Disable CxSR\n");
8052                 }
8053         } else if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
8054                 vlv_setup_wm_latency(dev_priv);
8055                 dev_priv->display.compute_pipe_wm = vlv_compute_pipe_wm;
8056                 dev_priv->display.compute_intermediate_wm = vlv_compute_intermediate_wm;
8057                 dev_priv->display.initial_watermarks = vlv_initial_watermarks;
8058                 dev_priv->display.optimize_watermarks = vlv_optimize_watermarks;
8059                 dev_priv->display.atomic_update_watermarks = vlv_atomic_update_fifo;
8060         } else if (IS_PINEVIEW(dev_priv)) {
8061                 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv),
8062                                             dev_priv->is_ddr3,
8063                                             dev_priv->fsb_freq,
8064                                             dev_priv->mem_freq)) {
8065                         DRM_INFO("failed to find known CxSR latency "
8066                                  "(found ddr%s fsb freq %d, mem freq %d), "
8067                                  "disabling CxSR\n",
8068                                  (dev_priv->is_ddr3 == 1) ? "3" : "2",
8069                                  dev_priv->fsb_freq, dev_priv->mem_freq);
8070                         /* Disable CxSR and never update its watermark again */
8071                         intel_set_memory_cxsr(dev_priv, false);
8072                         dev_priv->display.update_wm = NULL;
8073                 } else
8074                         dev_priv->display.update_wm = pineview_update_wm;
8075         } else if (IS_G4X(dev_priv)) {
8076                 dev_priv->display.update_wm = g4x_update_wm;
8077         } else if (IS_GEN4(dev_priv)) {
8078                 dev_priv->display.update_wm = i965_update_wm;
8079         } else if (IS_GEN3(dev_priv)) {
8080                 dev_priv->display.update_wm = i9xx_update_wm;
8081                 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
8082         } else if (IS_GEN2(dev_priv)) {
8083                 if (INTEL_INFO(dev_priv)->num_pipes == 1) {
8084                         dev_priv->display.update_wm = i845_update_wm;
8085                         dev_priv->display.get_fifo_size = i845_get_fifo_size;
8086                 } else {
8087                         dev_priv->display.update_wm = i9xx_update_wm;
8088                         dev_priv->display.get_fifo_size = i830_get_fifo_size;
8089                 }
8090         } else {
8091                 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
8092         }
8093 }
8094
8095 static inline int gen6_check_mailbox_status(struct drm_i915_private *dev_priv)
8096 {
8097         uint32_t flags =
8098                 I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_ERROR_MASK;
8099
8100         switch (flags) {
8101         case GEN6_PCODE_SUCCESS:
8102                 return 0;
8103         case GEN6_PCODE_UNIMPLEMENTED_CMD:
8104         case GEN6_PCODE_ILLEGAL_CMD:
8105                 return -ENXIO;
8106         case GEN6_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
8107         case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
8108                 return -EOVERFLOW;
8109         case GEN6_PCODE_TIMEOUT:
8110                 return -ETIMEDOUT;
8111         default:
8112                 MISSING_CASE(flags);
8113                 return 0;
8114         }
8115 }
8116
8117 static inline int gen7_check_mailbox_status(struct drm_i915_private *dev_priv)
8118 {
8119         uint32_t flags =
8120                 I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_ERROR_MASK;
8121
8122         switch (flags) {
8123         case GEN6_PCODE_SUCCESS:
8124                 return 0;
8125         case GEN6_PCODE_ILLEGAL_CMD:
8126                 return -ENXIO;
8127         case GEN7_PCODE_TIMEOUT:
8128                 return -ETIMEDOUT;
8129         case GEN7_PCODE_ILLEGAL_DATA:
8130                 return -EINVAL;
8131         case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
8132                 return -EOVERFLOW;
8133         default:
8134                 MISSING_CASE(flags);
8135                 return 0;
8136         }
8137 }
8138
8139 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u32 mbox, u32 *val)
8140 {
8141         int status;
8142
8143         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
8144
8145         /* GEN6_PCODE_* are outside of the forcewake domain, we can
8146          * use te fw I915_READ variants to reduce the amount of work
8147          * required when reading/writing.
8148          */
8149
8150         if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
8151                 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
8152                 return -EAGAIN;
8153         }
8154
8155         I915_WRITE_FW(GEN6_PCODE_DATA, *val);
8156         I915_WRITE_FW(GEN6_PCODE_DATA1, 0);
8157         I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
8158
8159         if (intel_wait_for_register_fw(dev_priv,
8160                                        GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
8161                                        500)) {
8162                 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
8163                 return -ETIMEDOUT;
8164         }
8165
8166         *val = I915_READ_FW(GEN6_PCODE_DATA);
8167         I915_WRITE_FW(GEN6_PCODE_DATA, 0);
8168
8169         if (INTEL_GEN(dev_priv) > 6)
8170                 status = gen7_check_mailbox_status(dev_priv);
8171         else
8172                 status = gen6_check_mailbox_status(dev_priv);
8173
8174         if (status) {
8175                 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed: %d\n",
8176                                  status);
8177                 return status;
8178         }
8179
8180         return 0;
8181 }
8182
8183 int sandybridge_pcode_write(struct drm_i915_private *dev_priv,
8184                             u32 mbox, u32 val)
8185 {
8186         int status;
8187
8188         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
8189
8190         /* GEN6_PCODE_* are outside of the forcewake domain, we can
8191          * use te fw I915_READ variants to reduce the amount of work
8192          * required when reading/writing.
8193          */
8194
8195         if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
8196                 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
8197                 return -EAGAIN;
8198         }
8199
8200         I915_WRITE_FW(GEN6_PCODE_DATA, val);
8201         I915_WRITE_FW(GEN6_PCODE_DATA1, 0);
8202         I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
8203
8204         if (intel_wait_for_register_fw(dev_priv,
8205                                        GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
8206                                        500)) {
8207                 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
8208                 return -ETIMEDOUT;
8209         }
8210
8211         I915_WRITE_FW(GEN6_PCODE_DATA, 0);
8212
8213         if (INTEL_GEN(dev_priv) > 6)
8214                 status = gen7_check_mailbox_status(dev_priv);
8215         else
8216                 status = gen6_check_mailbox_status(dev_priv);
8217
8218         if (status) {
8219                 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed: %d\n",
8220                                  status);
8221                 return status;
8222         }
8223
8224         return 0;
8225 }
8226
8227 static bool skl_pcode_try_request(struct drm_i915_private *dev_priv, u32 mbox,
8228                                   u32 request, u32 reply_mask, u32 reply,
8229                                   u32 *status)
8230 {
8231         u32 val = request;
8232
8233         *status = sandybridge_pcode_read(dev_priv, mbox, &val);
8234
8235         return *status || ((val & reply_mask) == reply);
8236 }
8237
8238 /**
8239  * skl_pcode_request - send PCODE request until acknowledgment
8240  * @dev_priv: device private
8241  * @mbox: PCODE mailbox ID the request is targeted for
8242  * @request: request ID
8243  * @reply_mask: mask used to check for request acknowledgment
8244  * @reply: value used to check for request acknowledgment
8245  * @timeout_base_ms: timeout for polling with preemption enabled
8246  *
8247  * Keep resending the @request to @mbox until PCODE acknowledges it, PCODE
8248  * reports an error or an overall timeout of @timeout_base_ms+50 ms expires.
8249  * The request is acknowledged once the PCODE reply dword equals @reply after
8250  * applying @reply_mask. Polling is first attempted with preemption enabled
8251  * for @timeout_base_ms and if this times out for another 50 ms with
8252  * preemption disabled.
8253  *
8254  * Returns 0 on success, %-ETIMEDOUT in case of a timeout, <0 in case of some
8255  * other error as reported by PCODE.
8256  */
8257 int skl_pcode_request(struct drm_i915_private *dev_priv, u32 mbox, u32 request,
8258                       u32 reply_mask, u32 reply, int timeout_base_ms)
8259 {
8260         u32 status;
8261         int ret;
8262
8263         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
8264
8265 #define COND skl_pcode_try_request(dev_priv, mbox, request, reply_mask, reply, \
8266                                    &status)
8267
8268         /*
8269          * Prime the PCODE by doing a request first. Normally it guarantees
8270          * that a subsequent request, at most @timeout_base_ms later, succeeds.
8271          * _wait_for() doesn't guarantee when its passed condition is evaluated
8272          * first, so send the first request explicitly.
8273          */
8274         if (COND) {
8275                 ret = 0;
8276                 goto out;
8277         }
8278         ret = _wait_for(COND, timeout_base_ms * 1000, 10);
8279         if (!ret)
8280                 goto out;
8281
8282         /*
8283          * The above can time out if the number of requests was low (2 in the
8284          * worst case) _and_ PCODE was busy for some reason even after a
8285          * (queued) request and @timeout_base_ms delay. As a workaround retry
8286          * the poll with preemption disabled to maximize the number of
8287          * requests. Increase the timeout from @timeout_base_ms to 50ms to
8288          * account for interrupts that could reduce the number of these
8289          * requests, and for any quirks of the PCODE firmware that delays
8290          * the request completion.
8291          */
8292         DRM_DEBUG_KMS("PCODE timeout, retrying with preemption disabled\n");
8293         WARN_ON_ONCE(timeout_base_ms > 3);
8294         preempt_disable();
8295         ret = wait_for_atomic(COND, 50);
8296         preempt_enable();
8297
8298 out:
8299         return ret ? ret : status;
8300 #undef COND
8301 }
8302
8303 static int byt_gpu_freq(struct drm_i915_private *dev_priv, int val)
8304 {
8305         /*
8306          * N = val - 0xb7
8307          * Slow = Fast = GPLL ref * N
8308          */
8309         return DIV_ROUND_CLOSEST(dev_priv->rps.gpll_ref_freq * (val - 0xb7), 1000);
8310 }
8311
8312 static int byt_freq_opcode(struct drm_i915_private *dev_priv, int val)
8313 {
8314         return DIV_ROUND_CLOSEST(1000 * val, dev_priv->rps.gpll_ref_freq) + 0xb7;
8315 }
8316
8317 static int chv_gpu_freq(struct drm_i915_private *dev_priv, int val)
8318 {
8319         /*
8320          * N = val / 2
8321          * CU (slow) = CU2x (fast) / 2 = GPLL ref * N / 2
8322          */
8323         return DIV_ROUND_CLOSEST(dev_priv->rps.gpll_ref_freq * val, 2 * 2 * 1000);
8324 }
8325
8326 static int chv_freq_opcode(struct drm_i915_private *dev_priv, int val)
8327 {
8328         /* CHV needs even values */
8329         return DIV_ROUND_CLOSEST(2 * 1000 * val, dev_priv->rps.gpll_ref_freq) * 2;
8330 }
8331
8332 int intel_gpu_freq(struct drm_i915_private *dev_priv, int val)
8333 {
8334         if (IS_GEN9(dev_priv))
8335                 return DIV_ROUND_CLOSEST(val * GT_FREQUENCY_MULTIPLIER,
8336                                          GEN9_FREQ_SCALER);
8337         else if (IS_CHERRYVIEW(dev_priv))
8338                 return chv_gpu_freq(dev_priv, val);
8339         else if (IS_VALLEYVIEW(dev_priv))
8340                 return byt_gpu_freq(dev_priv, val);
8341         else
8342                 return val * GT_FREQUENCY_MULTIPLIER;
8343 }
8344
8345 int intel_freq_opcode(struct drm_i915_private *dev_priv, int val)
8346 {
8347         if (IS_GEN9(dev_priv))
8348                 return DIV_ROUND_CLOSEST(val * GEN9_FREQ_SCALER,
8349                                          GT_FREQUENCY_MULTIPLIER);
8350         else if (IS_CHERRYVIEW(dev_priv))
8351                 return chv_freq_opcode(dev_priv, val);
8352         else if (IS_VALLEYVIEW(dev_priv))
8353                 return byt_freq_opcode(dev_priv, val);
8354         else
8355                 return DIV_ROUND_CLOSEST(val, GT_FREQUENCY_MULTIPLIER);
8356 }
8357
8358 struct request_boost {
8359         struct work_struct work;
8360         struct drm_i915_gem_request *req;
8361 };
8362
8363 static void __intel_rps_boost_work(struct work_struct *work)
8364 {
8365         struct request_boost *boost = container_of(work, struct request_boost, work);
8366         struct drm_i915_gem_request *req = boost->req;
8367
8368         if (!i915_gem_request_completed(req))
8369                 gen6_rps_boost(req->i915, NULL, req->emitted_jiffies);
8370
8371         i915_gem_request_put(req);
8372         kfree(boost);
8373 }
8374
8375 void intel_queue_rps_boost_for_request(struct drm_i915_gem_request *req)
8376 {
8377         struct request_boost *boost;
8378
8379         if (req == NULL || INTEL_GEN(req->i915) < 6)
8380                 return;
8381
8382         if (i915_gem_request_completed(req))
8383                 return;
8384
8385         boost = kmalloc(sizeof(*boost), GFP_ATOMIC);
8386         if (boost == NULL)
8387                 return;
8388
8389         boost->req = i915_gem_request_get(req);
8390
8391         INIT_WORK(&boost->work, __intel_rps_boost_work);
8392         queue_work(req->i915->wq, &boost->work);
8393 }
8394
8395 void intel_pm_setup(struct drm_i915_private *dev_priv)
8396 {
8397         mutex_init(&dev_priv->rps.hw_lock);
8398         spin_lock_init(&dev_priv->rps.client_lock);
8399
8400         INIT_DELAYED_WORK(&dev_priv->rps.autoenable_work,
8401                           __intel_autoenable_gt_powersave);
8402         INIT_LIST_HEAD(&dev_priv->rps.clients);
8403
8404         dev_priv->pm.suspended = false;
8405         atomic_set(&dev_priv->pm.wakeref_count, 0);
8406 }
8407
8408 static u64 vlv_residency_raw(struct drm_i915_private *dev_priv,
8409                              const i915_reg_t reg)
8410 {
8411         u32 lower, upper, tmp;
8412         int loop = 2;
8413
8414         /* The register accessed do not need forcewake. We borrow
8415          * uncore lock to prevent concurrent access to range reg.
8416          */
8417         spin_lock_irq(&dev_priv->uncore.lock);
8418
8419         /* vlv and chv residency counters are 40 bits in width.
8420          * With a control bit, we can choose between upper or lower
8421          * 32bit window into this counter.
8422          *
8423          * Although we always use the counter in high-range mode elsewhere,
8424          * userspace may attempt to read the value before rc6 is initialised,
8425          * before we have set the default VLV_COUNTER_CONTROL value. So always
8426          * set the high bit to be safe.
8427          */
8428         I915_WRITE_FW(VLV_COUNTER_CONTROL,
8429                       _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH));
8430         upper = I915_READ_FW(reg);
8431         do {
8432                 tmp = upper;
8433
8434                 I915_WRITE_FW(VLV_COUNTER_CONTROL,
8435                               _MASKED_BIT_DISABLE(VLV_COUNT_RANGE_HIGH));
8436                 lower = I915_READ_FW(reg);
8437
8438                 I915_WRITE_FW(VLV_COUNTER_CONTROL,
8439                               _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH));
8440                 upper = I915_READ_FW(reg);
8441         } while (upper != tmp && --loop);
8442
8443         /* Everywhere else we always use VLV_COUNTER_CONTROL with the
8444          * VLV_COUNT_RANGE_HIGH bit set - so it is safe to leave it set
8445          * now.
8446          */
8447
8448         spin_unlock_irq(&dev_priv->uncore.lock);
8449
8450         return lower | (u64)upper << 8;
8451 }
8452
8453 u64 intel_rc6_residency_us(struct drm_i915_private *dev_priv,
8454                            const i915_reg_t reg)
8455 {
8456         u64 time_hw, units, div;
8457
8458         if (!intel_enable_rc6())
8459                 return 0;
8460
8461         intel_runtime_pm_get(dev_priv);
8462
8463         /* On VLV and CHV, residency time is in CZ units rather than 1.28us */
8464         if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
8465                 units = 1000;
8466                 div = dev_priv->czclk_freq;
8467
8468                 time_hw = vlv_residency_raw(dev_priv, reg);
8469         } else if (IS_GEN9_LP(dev_priv)) {
8470                 units = 1000;
8471                 div = 1200;             /* 833.33ns */
8472
8473                 time_hw = I915_READ(reg);
8474         } else {
8475                 units = 128000; /* 1.28us */
8476                 div = 100000;
8477
8478                 time_hw = I915_READ(reg);
8479         }
8480
8481         intel_runtime_pm_put(dev_priv);
8482         return DIV_ROUND_UP_ULL(time_hw * units, div);
8483 }