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 "i915_drv.h"
  30 #include "intel_drv.h"
  31 #include "../../../platform/x86/intel_ips.h"
  32 #include <linux/module.h>
  33
  34 /* FBC, or Frame Buffer Compression, is a technique employed to compress the
  35  * framebuffer contents in-memory, aiming at reducing the required bandwidth
  36  * during in-memory transfers and, therefore, reduce the power packet.
  37  *
  38  * The benefits of FBC are mostly visible with solid backgrounds and
  39  * variation-less patterns.
  40  *
  41  * FBC-related functionality can be enabled by the means of the
  42  * i915.i915_enable_fbc parameter
  43  */
  44
  45 static void i8xx_disable_fbc(struct drm_device *dev)
  46 {
  47         struct drm_i915_private *dev_priv = dev->dev_private;
  48         u32 fbc_ctl;
  49
  50         /* Disable compression */
  51         fbc_ctl = I915_READ(FBC_CONTROL);
  52         if ((fbc_ctl & FBC_CTL_EN) == 0)
  53                 return;
  54
  55         fbc_ctl &= ~FBC_CTL_EN;
  56         I915_WRITE(FBC_CONTROL, fbc_ctl);
  57
  58         /* Wait for compressing bit to clear */
  59         if (wait_for((I915_READ(FBC_STATUS) & FBC_STAT_COMPRESSING) == 0, 10)) {
  60                 DRM_DEBUG_KMS("FBC idle timed out\n");
  61                 return;
  62         }
  63
  64         DRM_DEBUG_KMS("disabled FBC\n");
  65 }
  66
  67 static void i8xx_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
  68 {
  69         struct drm_device *dev = crtc->dev;
  70         struct drm_i915_private *dev_priv = dev->dev_private;
  71         struct drm_framebuffer *fb = crtc->fb;
  72         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
  73         struct drm_i915_gem_object *obj = intel_fb->obj;
  74         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
  75         int cfb_pitch;
  76         int plane, i;
  77         u32 fbc_ctl, fbc_ctl2;
  78
  79         cfb_pitch = dev_priv->cfb_size / FBC_LL_SIZE;
  80         if (fb->pitches[0] < cfb_pitch)
  81                 cfb_pitch = fb->pitches[0];
  82
  83         /* FBC_CTL wants 64B units */
  84         cfb_pitch = (cfb_pitch / 64) - 1;
  85         plane = intel_crtc->plane == 0 ? FBC_CTL_PLANEA : FBC_CTL_PLANEB;
  86
  87         /* Clear old tags */
  88         for (i = 0; i < (FBC_LL_SIZE / 32) + 1; i++)
  89                 I915_WRITE(FBC_TAG + (i * 4), 0);
  90
  91         /* Set it up... */
  92         fbc_ctl2 = FBC_CTL_FENCE_DBL | FBC_CTL_IDLE_IMM | FBC_CTL_CPU_FENCE;
  93         fbc_ctl2 |= plane;
  94         I915_WRITE(FBC_CONTROL2, fbc_ctl2);
  95         I915_WRITE(FBC_FENCE_OFF, crtc->y);
  96
  97         /* enable it... */
  98         fbc_ctl = FBC_CTL_EN | FBC_CTL_PERIODIC;
  99         if (IS_I945GM(dev))
 100                 fbc_ctl |= FBC_CTL_C3_IDLE; /* 945 needs special SR handling */
 101         fbc_ctl |= (cfb_pitch & 0xff) << FBC_CTL_STRIDE_SHIFT;
 102         fbc_ctl |= (interval & 0x2fff) << FBC_CTL_INTERVAL_SHIFT;
 103         fbc_ctl |= obj->fence_reg;
 104         I915_WRITE(FBC_CONTROL, fbc_ctl);
 105
 106         DRM_DEBUG_KMS("enabled FBC, pitch %d, yoff %d, plane %d, ",
 107                       cfb_pitch, crtc->y, intel_crtc->plane);
 108 }
 109
 110 static bool i8xx_fbc_enabled(struct drm_device *dev)
 111 {
 112         struct drm_i915_private *dev_priv = dev->dev_private;
 113
 114         return I915_READ(FBC_CONTROL) & FBC_CTL_EN;
 115 }
 116
 117 static void g4x_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 118 {
 119         struct drm_device *dev = crtc->dev;
 120         struct drm_i915_private *dev_priv = dev->dev_private;
 121         struct drm_framebuffer *fb = crtc->fb;
 122         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 123         struct drm_i915_gem_object *obj = intel_fb->obj;
 124         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 125         int plane = intel_crtc->plane == 0 ? DPFC_CTL_PLANEA : DPFC_CTL_PLANEB;
 126         unsigned long stall_watermark = 200;
 127         u32 dpfc_ctl;
 128
 129         dpfc_ctl = plane | DPFC_SR_EN | DPFC_CTL_LIMIT_1X;
 130         dpfc_ctl |= DPFC_CTL_FENCE_EN | obj->fence_reg;
 131         I915_WRITE(DPFC_CHICKEN, DPFC_HT_MODIFY);
 132
 133         I915_WRITE(DPFC_RECOMP_CTL, DPFC_RECOMP_STALL_EN |
 134                    (stall_watermark << DPFC_RECOMP_STALL_WM_SHIFT) |
 135                    (interval << DPFC_RECOMP_TIMER_COUNT_SHIFT));
 136         I915_WRITE(DPFC_FENCE_YOFF, crtc->y);
 137
 138         /* enable it... */
 139         I915_WRITE(DPFC_CONTROL, I915_READ(DPFC_CONTROL) | DPFC_CTL_EN);
 140
 141         DRM_DEBUG_KMS("enabled fbc on plane %d\n", intel_crtc->plane);
 142 }
 143
 144 static void g4x_disable_fbc(struct drm_device *dev)
 145 {
 146         struct drm_i915_private *dev_priv = dev->dev_private;
 147         u32 dpfc_ctl;
 148
 149         /* Disable compression */
 150         dpfc_ctl = I915_READ(DPFC_CONTROL);
 151         if (dpfc_ctl & DPFC_CTL_EN) {
 152                 dpfc_ctl &= ~DPFC_CTL_EN;
 153                 I915_WRITE(DPFC_CONTROL, dpfc_ctl);
 154
 155                 DRM_DEBUG_KMS("disabled FBC\n");
 156         }
 157 }
 158
 159 static bool g4x_fbc_enabled(struct drm_device *dev)
 160 {
 161         struct drm_i915_private *dev_priv = dev->dev_private;
 162
 163         return I915_READ(DPFC_CONTROL) & DPFC_CTL_EN;
 164 }
 165
 166 static void sandybridge_blit_fbc_update(struct drm_device *dev)
 167 {
 168         struct drm_i915_private *dev_priv = dev->dev_private;
 169         u32 blt_ecoskpd;
 170
 171         /* Make sure blitter notifies FBC of writes */
 172         gen6_gt_force_wake_get(dev_priv);
 173         blt_ecoskpd = I915_READ(GEN6_BLITTER_ECOSKPD);
 174         blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY <<
 175                 GEN6_BLITTER_LOCK_SHIFT;
 176         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 177         blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY;
 178         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 179         blt_ecoskpd &= ~(GEN6_BLITTER_FBC_NOTIFY <<
 180                          GEN6_BLITTER_LOCK_SHIFT);
 181         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 182         POSTING_READ(GEN6_BLITTER_ECOSKPD);
 183         gen6_gt_force_wake_put(dev_priv);
 184 }
 185
 186 static void ironlake_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 187 {
 188         struct drm_device *dev = crtc->dev;
 189         struct drm_i915_private *dev_priv = dev->dev_private;
 190         struct drm_framebuffer *fb = crtc->fb;
 191         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 192         struct drm_i915_gem_object *obj = intel_fb->obj;
 193         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 194         int plane = intel_crtc->plane == 0 ? DPFC_CTL_PLANEA : DPFC_CTL_PLANEB;
 195         unsigned long stall_watermark = 200;
 196         u32 dpfc_ctl;
 197
 198         dpfc_ctl = I915_READ(ILK_DPFC_CONTROL);
 199         dpfc_ctl &= DPFC_RESERVED;
 200         dpfc_ctl |= (plane | DPFC_CTL_LIMIT_1X);
 201         /* Set persistent mode for front-buffer rendering, ala X. */
 202         dpfc_ctl |= DPFC_CTL_PERSISTENT_MODE;
 203         dpfc_ctl |= (DPFC_CTL_FENCE_EN | obj->fence_reg);
 204         I915_WRITE(ILK_DPFC_CHICKEN, DPFC_HT_MODIFY);
 205
 206         I915_WRITE(ILK_DPFC_RECOMP_CTL, DPFC_RECOMP_STALL_EN |
 207                    (stall_watermark << DPFC_RECOMP_STALL_WM_SHIFT) |
 208                    (interval << DPFC_RECOMP_TIMER_COUNT_SHIFT));
 209         I915_WRITE(ILK_DPFC_FENCE_YOFF, crtc->y);
 210         I915_WRITE(ILK_FBC_RT_BASE, obj->gtt_offset | ILK_FBC_RT_VALID);
 211         /* enable it... */
 212         I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
 213
 214         if (IS_GEN6(dev)) {
 215                 I915_WRITE(SNB_DPFC_CTL_SA,
 216                            SNB_CPU_FENCE_ENABLE | obj->fence_reg);
 217                 I915_WRITE(DPFC_CPU_FENCE_OFFSET, crtc->y);
 218                 sandybridge_blit_fbc_update(dev);
 219         }
 220
 221         DRM_DEBUG_KMS("enabled fbc on plane %d\n", intel_crtc->plane);
 222 }
 223
 224 static void ironlake_disable_fbc(struct drm_device *dev)
 225 {
 226         struct drm_i915_private *dev_priv = dev->dev_private;
 227         u32 dpfc_ctl;
 228
 229         /* Disable compression */
 230         dpfc_ctl = I915_READ(ILK_DPFC_CONTROL);
 231         if (dpfc_ctl & DPFC_CTL_EN) {
 232                 dpfc_ctl &= ~DPFC_CTL_EN;
 233                 I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl);
 234
 235                 DRM_DEBUG_KMS("disabled FBC\n");
 236         }
 237 }
 238
 239 static bool ironlake_fbc_enabled(struct drm_device *dev)
 240 {
 241         struct drm_i915_private *dev_priv = dev->dev_private;
 242
 243         return I915_READ(ILK_DPFC_CONTROL) & DPFC_CTL_EN;
 244 }
 245
 246 bool intel_fbc_enabled(struct drm_device *dev)
 247 {
 248         struct drm_i915_private *dev_priv = dev->dev_private;
 249
 250         if (!dev_priv->display.fbc_enabled)
 251                 return false;
 252
 253         return dev_priv->display.fbc_enabled(dev);
 254 }
 255
 256 static void intel_fbc_work_fn(struct work_struct *__work)
 257 {
 258         struct intel_fbc_work *work =
 259                 container_of(to_delayed_work(__work),
 260                              struct intel_fbc_work, work);
 261         struct drm_device *dev = work->crtc->dev;
 262         struct drm_i915_private *dev_priv = dev->dev_private;
 263
 264         mutex_lock(&dev->struct_mutex);
 265         if (work == dev_priv->fbc_work) {
 266                 /* Double check that we haven't switched fb without cancelling
 267                  * the prior work.
 268                  */
 269                 if (work->crtc->fb == work->fb) {
 270                         dev_priv->display.enable_fbc(work->crtc,
 271                                                      work->interval);
 272
 273                         dev_priv->cfb_plane = to_intel_crtc(work->crtc)->plane;
 274                         dev_priv->cfb_fb = work->crtc->fb->base.id;
 275                         dev_priv->cfb_y = work->crtc->y;
 276                 }
 277
 278                 dev_priv->fbc_work = NULL;
 279         }
 280         mutex_unlock(&dev->struct_mutex);
 281
 282         kfree(work);
 283 }
 284
 285 static void intel_cancel_fbc_work(struct drm_i915_private *dev_priv)
 286 {
 287         if (dev_priv->fbc_work == NULL)
 288                 return;
 289
 290         DRM_DEBUG_KMS("cancelling pending FBC enable\n");
 291
 292         /* Synchronisation is provided by struct_mutex and checking of
 293          * dev_priv->fbc_work, so we can perform the cancellation
 294          * entirely asynchronously.
 295          */
 296         if (cancel_delayed_work(&dev_priv->fbc_work->work))
 297                 /* tasklet was killed before being run, clean up */
 298                 kfree(dev_priv->fbc_work);
 299
 300         /* Mark the work as no longer wanted so that if it does
 301          * wake-up (because the work was already running and waiting
 302          * for our mutex), it will discover that is no longer
 303          * necessary to run.
 304          */
 305         dev_priv->fbc_work = NULL;
 306 }
 307
 308 void intel_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 309 {
 310         struct intel_fbc_work *work;
 311         struct drm_device *dev = crtc->dev;
 312         struct drm_i915_private *dev_priv = dev->dev_private;
 313
 314         if (!dev_priv->display.enable_fbc)
 315                 return;
 316
 317         intel_cancel_fbc_work(dev_priv);
 318
 319         work = kzalloc(sizeof *work, GFP_KERNEL);
 320         if (work == NULL) {
 321                 dev_priv->display.enable_fbc(crtc, interval);
 322                 return;
 323         }
 324
 325         work->crtc = crtc;
 326         work->fb = crtc->fb;
 327         work->interval = interval;
 328         INIT_DELAYED_WORK(&work->work, intel_fbc_work_fn);
 329
 330         dev_priv->fbc_work = work;
 331
 332         DRM_DEBUG_KMS("scheduling delayed FBC enable\n");
 333
 334         /* Delay the actual enabling to let pageflipping cease and the
 335          * display to settle before starting the compression. Note that
 336          * this delay also serves a second purpose: it allows for a
 337          * vblank to pass after disabling the FBC before we attempt
 338          * to modify the control registers.
 339          *
 340          * A more complicated solution would involve tracking vblanks
 341          * following the termination of the page-flipping sequence
 342          * and indeed performing the enable as a co-routine and not
 343          * waiting synchronously upon the vblank.
 344          */
 345         schedule_delayed_work(&work->work, msecs_to_jiffies(50));
 346 }
 347
 348 void intel_disable_fbc(struct drm_device *dev)
 349 {
 350         struct drm_i915_private *dev_priv = dev->dev_private;
 351
 352         intel_cancel_fbc_work(dev_priv);
 353
 354         if (!dev_priv->display.disable_fbc)
 355                 return;
 356
 357         dev_priv->display.disable_fbc(dev);
 358         dev_priv->cfb_plane = -1;
 359 }
 360
 361 /**
 362  * intel_update_fbc - enable/disable FBC as needed
 363  * @dev: the drm_device
 364  *
 365  * Set up the framebuffer compression hardware at mode set time.  We
 366  * enable it if possible:
 367  *   - plane A only (on pre-965)
 368  *   - no pixel mulitply/line duplication
 369  *   - no alpha buffer discard
 370  *   - no dual wide
 371  *   - framebuffer <= 2048 in width, 1536 in height
 372  *
 373  * We can't assume that any compression will take place (worst case),
 374  * so the compressed buffer has to be the same size as the uncompressed
 375  * one.  It also must reside (along with the line length buffer) in
 376  * stolen memory.
 377  *
 378  * We need to enable/disable FBC on a global basis.
 379  */
 380 void intel_update_fbc(struct drm_device *dev)
 381 {
 382         struct drm_i915_private *dev_priv = dev->dev_private;
 383         struct drm_crtc *crtc = NULL, *tmp_crtc;
 384         struct intel_crtc *intel_crtc;
 385         struct drm_framebuffer *fb;
 386         struct intel_framebuffer *intel_fb;
 387         struct drm_i915_gem_object *obj;
 388         int enable_fbc;
 389
 390         if (!i915_powersave)
 391                 return;
 392
 393         if (!I915_HAS_FBC(dev))
 394                 return;
 395
 396         /*
 397          * If FBC is already on, we just have to verify that we can
 398          * keep it that way...
 399          * Need to disable if:
 400          *   - more than one pipe is active
 401          *   - changing FBC params (stride, fence, mode)
 402          *   - new fb is too large to fit in compressed buffer
 403          *   - going to an unsupported config (interlace, pixel multiply, etc.)
 404          */
 405         list_for_each_entry(tmp_crtc, &dev->mode_config.crtc_list, head) {
 406                 if (tmp_crtc->enabled &&
 407                     !to_intel_crtc(tmp_crtc)->primary_disabled &&
 408                     tmp_crtc->fb) {
 409                         if (crtc) {
 410                                 DRM_DEBUG_KMS("more than one pipe active, disabling compression\n");
 411                                 dev_priv->no_fbc_reason = FBC_MULTIPLE_PIPES;
 412                                 goto out_disable;
 413                         }
 414                         crtc = tmp_crtc;
 415                 }
 416         }
 417
 418         if (!crtc || crtc->fb == NULL) {
 419                 DRM_DEBUG_KMS("no output, disabling\n");
 420                 dev_priv->no_fbc_reason = FBC_NO_OUTPUT;
 421                 goto out_disable;
 422         }
 423
 424         intel_crtc = to_intel_crtc(crtc);
 425         fb = crtc->fb;
 426         intel_fb = to_intel_framebuffer(fb);
 427         obj = intel_fb->obj;
 428
 429         enable_fbc = i915_enable_fbc;
 430         if (enable_fbc < 0) {
 431                 DRM_DEBUG_KMS("fbc set to per-chip default\n");
 432                 enable_fbc = 1;
 433                 if (INTEL_INFO(dev)->gen <= 6)
 434                         enable_fbc = 0;
 435         }
 436         if (!enable_fbc) {
 437                 DRM_DEBUG_KMS("fbc disabled per module param\n");
 438                 dev_priv->no_fbc_reason = FBC_MODULE_PARAM;
 439                 goto out_disable;
 440         }
 441         if (intel_fb->obj->base.size > dev_priv->cfb_size) {
 442                 DRM_DEBUG_KMS("framebuffer too large, disabling "
 443                               "compression\n");
 444                 dev_priv->no_fbc_reason = FBC_STOLEN_TOO_SMALL;
 445                 goto out_disable;
 446         }
 447         if ((crtc->mode.flags & DRM_MODE_FLAG_INTERLACE) ||
 448             (crtc->mode.flags & DRM_MODE_FLAG_DBLSCAN)) {
 449                 DRM_DEBUG_KMS("mode incompatible with compression, "
 450                               "disabling\n");
 451                 dev_priv->no_fbc_reason = FBC_UNSUPPORTED_MODE;
 452                 goto out_disable;
 453         }
 454         if ((crtc->mode.hdisplay > 2048) ||
 455             (crtc->mode.vdisplay > 1536)) {
 456                 DRM_DEBUG_KMS("mode too large for compression, disabling\n");
 457                 dev_priv->no_fbc_reason = FBC_MODE_TOO_LARGE;
 458                 goto out_disable;
 459         }
 460         if ((IS_I915GM(dev) || IS_I945GM(dev)) && intel_crtc->plane != 0) {
 461                 DRM_DEBUG_KMS("plane not 0, disabling compression\n");
 462                 dev_priv->no_fbc_reason = FBC_BAD_PLANE;
 463                 goto out_disable;
 464         }
 465
 466         /* The use of a CPU fence is mandatory in order to detect writes
 467          * by the CPU to the scanout and trigger updates to the FBC.
 468          */
 469         if (obj->tiling_mode != I915_TILING_X ||
 470             obj->fence_reg == I915_FENCE_REG_NONE) {
 471                 DRM_DEBUG_KMS("framebuffer not tiled or fenced, disabling compression\n");
 472                 dev_priv->no_fbc_reason = FBC_NOT_TILED;
 473                 goto out_disable;
 474         }
 475
 476         /* If the kernel debugger is active, always disable compression */
 477         if (in_dbg_master())
 478                 goto out_disable;
 479
 480         /* If the scanout has not changed, don't modify the FBC settings.
 481          * Note that we make the fundamental assumption that the fb->obj
 482          * cannot be unpinned (and have its GTT offset and fence revoked)
 483          * without first being decoupled from the scanout and FBC disabled.
 484          */
 485         if (dev_priv->cfb_plane == intel_crtc->plane &&
 486             dev_priv->cfb_fb == fb->base.id &&
 487             dev_priv->cfb_y == crtc->y)
 488                 return;
 489
 490         if (intel_fbc_enabled(dev)) {
 491                 /* We update FBC along two paths, after changing fb/crtc
 492                  * configuration (modeswitching) and after page-flipping
 493                  * finishes. For the latter, we know that not only did
 494                  * we disable the FBC at the start of the page-flip
 495                  * sequence, but also more than one vblank has passed.
 496                  *
 497                  * For the former case of modeswitching, it is possible
 498                  * to switch between two FBC valid configurations
 499                  * instantaneously so we do need to disable the FBC
 500                  * before we can modify its control registers. We also
 501                  * have to wait for the next vblank for that to take
 502                  * effect. However, since we delay enabling FBC we can
 503                  * assume that a vblank has passed since disabling and
 504                  * that we can safely alter the registers in the deferred
 505                  * callback.
 506                  *
 507                  * In the scenario that we go from a valid to invalid
 508                  * and then back to valid FBC configuration we have
 509                  * no strict enforcement that a vblank occurred since
 510                  * disabling the FBC. However, along all current pipe
 511                  * disabling paths we do need to wait for a vblank at
 512                  * some point. And we wait before enabling FBC anyway.
 513                  */
 514                 DRM_DEBUG_KMS("disabling active FBC for update\n");
 515                 intel_disable_fbc(dev);
 516         }
 517
 518         intel_enable_fbc(crtc, 500);
 519         return;
 520
 521 out_disable:
 522         /* Multiple disables should be harmless */
 523         if (intel_fbc_enabled(dev)) {
 524                 DRM_DEBUG_KMS("unsupported config, disabling FBC\n");
 525                 intel_disable_fbc(dev);
 526         }
 527 }
 528
 529 static void i915_pineview_get_mem_freq(struct drm_device *dev)
 530 {
 531         drm_i915_private_t *dev_priv = dev->dev_private;
 532         u32 tmp;
 533
 534         tmp = I915_READ(CLKCFG);
 535
 536         switch (tmp & CLKCFG_FSB_MASK) {
 537         case CLKCFG_FSB_533:
 538                 dev_priv->fsb_freq = 533; /* 133*4 */
 539                 break;
 540         case CLKCFG_FSB_800:
 541                 dev_priv->fsb_freq = 800; /* 200*4 */
 542                 break;
 543         case CLKCFG_FSB_667:
 544                 dev_priv->fsb_freq =  667; /* 167*4 */
 545                 break;
 546         case CLKCFG_FSB_400:
 547                 dev_priv->fsb_freq = 400; /* 100*4 */
 548                 break;
 549         }
 550
 551         switch (tmp & CLKCFG_MEM_MASK) {
 552         case CLKCFG_MEM_533:
 553                 dev_priv->mem_freq = 533;
 554                 break;
 555         case CLKCFG_MEM_667:
 556                 dev_priv->mem_freq = 667;
 557                 break;
 558         case CLKCFG_MEM_800:
 559                 dev_priv->mem_freq = 800;
 560                 break;
 561         }
 562
 563         /* detect pineview DDR3 setting */
 564         tmp = I915_READ(CSHRDDR3CTL);
 565         dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
 566 }
 567
 568 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
 569 {
 570         drm_i915_private_t *dev_priv = dev->dev_private;
 571         u16 ddrpll, csipll;
 572
 573         ddrpll = I915_READ16(DDRMPLL1);
 574         csipll = I915_READ16(CSIPLL0);
 575
 576         switch (ddrpll & 0xff) {
 577         case 0xc:
 578                 dev_priv->mem_freq = 800;
 579                 break;
 580         case 0x10:
 581                 dev_priv->mem_freq = 1066;
 582                 break;
 583         case 0x14:
 584                 dev_priv->mem_freq = 1333;
 585                 break;
 586         case 0x18:
 587                 dev_priv->mem_freq = 1600;
 588                 break;
 589         default:
 590                 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
 591                                  ddrpll & 0xff);
 592                 dev_priv->mem_freq = 0;
 593                 break;
 594         }
 595
 596         dev_priv->r_t = dev_priv->mem_freq;
 597
 598         switch (csipll & 0x3ff) {
 599         case 0x00c:
 600                 dev_priv->fsb_freq = 3200;
 601                 break;
 602         case 0x00e:
 603                 dev_priv->fsb_freq = 3733;
 604                 break;
 605         case 0x010:
 606                 dev_priv->fsb_freq = 4266;
 607                 break;
 608         case 0x012:
 609                 dev_priv->fsb_freq = 4800;
 610                 break;
 611         case 0x014:
 612                 dev_priv->fsb_freq = 5333;
 613                 break;
 614         case 0x016:
 615                 dev_priv->fsb_freq = 5866;
 616                 break;
 617         case 0x018:
 618                 dev_priv->fsb_freq = 6400;
 619                 break;
 620         default:
 621                 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
 622                                  csipll & 0x3ff);
 623                 dev_priv->fsb_freq = 0;
 624                 break;
 625         }
 626
 627         if (dev_priv->fsb_freq == 3200) {
 628                 dev_priv->c_m = 0;
 629         } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
 630                 dev_priv->c_m = 1;
 631         } else {
 632                 dev_priv->c_m = 2;
 633         }
 634 }
 635
 636 static const struct cxsr_latency cxsr_latency_table[] = {
 637         {1, 0, 800, 400, 3382, 33382, 3983, 33983},    /* DDR2-400 SC */
 638         {1, 0, 800, 667, 3354, 33354, 3807, 33807},    /* DDR2-667 SC */
 639         {1, 0, 800, 800, 3347, 33347, 3763, 33763},    /* DDR2-800 SC */
 640         {1, 1, 800, 667, 6420, 36420, 6873, 36873},    /* DDR3-667 SC */
 641         {1, 1, 800, 800, 5902, 35902, 6318, 36318},    /* DDR3-800 SC */
 642
 643         {1, 0, 667, 400, 3400, 33400, 4021, 34021},    /* DDR2-400 SC */
 644         {1, 0, 667, 667, 3372, 33372, 3845, 33845},    /* DDR2-667 SC */
 645         {1, 0, 667, 800, 3386, 33386, 3822, 33822},    /* DDR2-800 SC */
 646         {1, 1, 667, 667, 6438, 36438, 6911, 36911},    /* DDR3-667 SC */
 647         {1, 1, 667, 800, 5941, 35941, 6377, 36377},    /* DDR3-800 SC */
 648
 649         {1, 0, 400, 400, 3472, 33472, 4173, 34173},    /* DDR2-400 SC */
 650         {1, 0, 400, 667, 3443, 33443, 3996, 33996},    /* DDR2-667 SC */
 651         {1, 0, 400, 800, 3430, 33430, 3946, 33946},    /* DDR2-800 SC */
 652         {1, 1, 400, 667, 6509, 36509, 7062, 37062},    /* DDR3-667 SC */
 653         {1, 1, 400, 800, 5985, 35985, 6501, 36501},    /* DDR3-800 SC */
 654
 655         {0, 0, 800, 400, 3438, 33438, 4065, 34065},    /* DDR2-400 SC */
 656         {0, 0, 800, 667, 3410, 33410, 3889, 33889},    /* DDR2-667 SC */
 657         {0, 0, 800, 800, 3403, 33403, 3845, 33845},    /* DDR2-800 SC */
 658         {0, 1, 800, 667, 6476, 36476, 6955, 36955},    /* DDR3-667 SC */
 659         {0, 1, 800, 800, 5958, 35958, 6400, 36400},    /* DDR3-800 SC */
 660
 661         {0, 0, 667, 400, 3456, 33456, 4103, 34106},    /* DDR2-400 SC */
 662         {0, 0, 667, 667, 3428, 33428, 3927, 33927},    /* DDR2-667 SC */
 663         {0, 0, 667, 800, 3443, 33443, 3905, 33905},    /* DDR2-800 SC */
 664         {0, 1, 667, 667, 6494, 36494, 6993, 36993},    /* DDR3-667 SC */
 665         {0, 1, 667, 800, 5998, 35998, 6460, 36460},    /* DDR3-800 SC */
 666
 667         {0, 0, 400, 400, 3528, 33528, 4255, 34255},    /* DDR2-400 SC */
 668         {0, 0, 400, 667, 3500, 33500, 4079, 34079},    /* DDR2-667 SC */
 669         {0, 0, 400, 800, 3487, 33487, 4029, 34029},    /* DDR2-800 SC */
 670         {0, 1, 400, 667, 6566, 36566, 7145, 37145},    /* DDR3-667 SC */
 671         {0, 1, 400, 800, 6042, 36042, 6584, 36584},    /* DDR3-800 SC */
 672 };
 673
 674 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
 675                                                          int is_ddr3,
 676                                                          int fsb,
 677                                                          int mem)
 678 {
 679         const struct cxsr_latency *latency;
 680         int i;
 681
 682         if (fsb == 0 || mem == 0)
 683                 return NULL;
 684
 685         for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
 686                 latency = &cxsr_latency_table[i];
 687                 if (is_desktop == latency->is_desktop &&
 688                     is_ddr3 == latency->is_ddr3 &&
 689                     fsb == latency->fsb_freq && mem == latency->mem_freq)
 690                         return latency;
 691         }
 692
 693         DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 694
 695         return NULL;
 696 }
 697
 698 static void pineview_disable_cxsr(struct drm_device *dev)
 699 {
 700         struct drm_i915_private *dev_priv = dev->dev_private;
 701
 702         /* deactivate cxsr */
 703         I915_WRITE(DSPFW3, I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN);
 704 }
 705
 706 /*
 707  * Latency for FIFO fetches is dependent on several factors:
 708  *   - memory configuration (speed, channels)
 709  *   - chipset
 710  *   - current MCH state
 711  * It can be fairly high in some situations, so here we assume a fairly
 712  * pessimal value.  It's a tradeoff between extra memory fetches (if we
 713  * set this value too high, the FIFO will fetch frequently to stay full)
 714  * and power consumption (set it too low to save power and we might see
 715  * FIFO underruns and display "flicker").
 716  *
 717  * A value of 5us seems to be a good balance; safe for very low end
 718  * platforms but not overly aggressive on lower latency configs.
 719  */
 720 static const int latency_ns = 5000;
 721
 722 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
 723 {
 724         struct drm_i915_private *dev_priv = dev->dev_private;
 725         uint32_t dsparb = I915_READ(DSPARB);
 726         int size;
 727
 728         size = dsparb & 0x7f;
 729         if (plane)
 730                 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
 731
 732         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 733                       plane ? "B" : "A", size);
 734
 735         return size;
 736 }
 737
 738 static int i85x_get_fifo_size(struct drm_device *dev, int plane)
 739 {
 740         struct drm_i915_private *dev_priv = dev->dev_private;
 741         uint32_t dsparb = I915_READ(DSPARB);
 742         int size;
 743
 744         size = dsparb & 0x1ff;
 745         if (plane)
 746                 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
 747         size >>= 1; /* Convert to cachelines */
 748
 749         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 750                       plane ? "B" : "A", size);
 751
 752         return size;
 753 }
 754
 755 static int i845_get_fifo_size(struct drm_device *dev, int plane)
 756 {
 757         struct drm_i915_private *dev_priv = dev->dev_private;
 758         uint32_t dsparb = I915_READ(DSPARB);
 759         int size;
 760
 761         size = dsparb & 0x7f;
 762         size >>= 2; /* Convert to cachelines */
 763
 764         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 765                       plane ? "B" : "A",
 766                       size);
 767
 768         return size;
 769 }
 770
 771 static int i830_get_fifo_size(struct drm_device *dev, int plane)
 772 {
 773         struct drm_i915_private *dev_priv = dev->dev_private;
 774         uint32_t dsparb = I915_READ(DSPARB);
 775         int size;
 776
 777         size = dsparb & 0x7f;
 778         size >>= 1; /* Convert to cachelines */
 779
 780         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 781                       plane ? "B" : "A", size);
 782
 783         return size;
 784 }
 785
 786 /* Pineview has different values for various configs */
 787 static const struct intel_watermark_params pineview_display_wm = {
 788         PINEVIEW_DISPLAY_FIFO,
 789         PINEVIEW_MAX_WM,
 790         PINEVIEW_DFT_WM,
 791         PINEVIEW_GUARD_WM,
 792         PINEVIEW_FIFO_LINE_SIZE
 793 };
 794 static const struct intel_watermark_params pineview_display_hplloff_wm = {
 795         PINEVIEW_DISPLAY_FIFO,
 796         PINEVIEW_MAX_WM,
 797         PINEVIEW_DFT_HPLLOFF_WM,
 798         PINEVIEW_GUARD_WM,
 799         PINEVIEW_FIFO_LINE_SIZE
 800 };
 801 static const struct intel_watermark_params pineview_cursor_wm = {
 802         PINEVIEW_CURSOR_FIFO,
 803         PINEVIEW_CURSOR_MAX_WM,
 804         PINEVIEW_CURSOR_DFT_WM,
 805         PINEVIEW_CURSOR_GUARD_WM,
 806         PINEVIEW_FIFO_LINE_SIZE,
 807 };
 808 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
 809         PINEVIEW_CURSOR_FIFO,
 810         PINEVIEW_CURSOR_MAX_WM,
 811         PINEVIEW_CURSOR_DFT_WM,
 812         PINEVIEW_CURSOR_GUARD_WM,
 813         PINEVIEW_FIFO_LINE_SIZE
 814 };
 815 static const struct intel_watermark_params g4x_wm_info = {
 816         G4X_FIFO_SIZE,
 817         G4X_MAX_WM,
 818         G4X_MAX_WM,
 819         2,
 820         G4X_FIFO_LINE_SIZE,
 821 };
 822 static const struct intel_watermark_params g4x_cursor_wm_info = {
 823         I965_CURSOR_FIFO,
 824         I965_CURSOR_MAX_WM,
 825         I965_CURSOR_DFT_WM,
 826         2,
 827         G4X_FIFO_LINE_SIZE,
 828 };
 829 static const struct intel_watermark_params valleyview_wm_info = {
 830         VALLEYVIEW_FIFO_SIZE,
 831         VALLEYVIEW_MAX_WM,
 832         VALLEYVIEW_MAX_WM,
 833         2,
 834         G4X_FIFO_LINE_SIZE,
 835 };
 836 static const struct intel_watermark_params valleyview_cursor_wm_info = {
 837         I965_CURSOR_FIFO,
 838         VALLEYVIEW_CURSOR_MAX_WM,
 839         I965_CURSOR_DFT_WM,
 840         2,
 841         G4X_FIFO_LINE_SIZE,
 842 };
 843 static const struct intel_watermark_params i965_cursor_wm_info = {
 844         I965_CURSOR_FIFO,
 845         I965_CURSOR_MAX_WM,
 846         I965_CURSOR_DFT_WM,
 847         2,
 848         I915_FIFO_LINE_SIZE,
 849 };
 850 static const struct intel_watermark_params i945_wm_info = {
 851         I945_FIFO_SIZE,
 852         I915_MAX_WM,
 853         1,
 854         2,
 855         I915_FIFO_LINE_SIZE
 856 };
 857 static const struct intel_watermark_params i915_wm_info = {
 858         I915_FIFO_SIZE,
 859         I915_MAX_WM,
 860         1,
 861         2,
 862         I915_FIFO_LINE_SIZE
 863 };
 864 static const struct intel_watermark_params i855_wm_info = {
 865         I855GM_FIFO_SIZE,
 866         I915_MAX_WM,
 867         1,
 868         2,
 869         I830_FIFO_LINE_SIZE
 870 };
 871 static const struct intel_watermark_params i830_wm_info = {
 872         I830_FIFO_SIZE,
 873         I915_MAX_WM,
 874         1,
 875         2,
 876         I830_FIFO_LINE_SIZE
 877 };
 878
 879 static const struct intel_watermark_params ironlake_display_wm_info = {
 880         ILK_DISPLAY_FIFO,
 881         ILK_DISPLAY_MAXWM,
 882         ILK_DISPLAY_DFTWM,
 883         2,
 884         ILK_FIFO_LINE_SIZE
 885 };
 886 static const struct intel_watermark_params ironlake_cursor_wm_info = {
 887         ILK_CURSOR_FIFO,
 888         ILK_CURSOR_MAXWM,
 889         ILK_CURSOR_DFTWM,
 890         2,
 891         ILK_FIFO_LINE_SIZE
 892 };
 893 static const struct intel_watermark_params ironlake_display_srwm_info = {
 894         ILK_DISPLAY_SR_FIFO,
 895         ILK_DISPLAY_MAX_SRWM,
 896         ILK_DISPLAY_DFT_SRWM,
 897         2,
 898         ILK_FIFO_LINE_SIZE
 899 };
 900 static const struct intel_watermark_params ironlake_cursor_srwm_info = {
 901         ILK_CURSOR_SR_FIFO,
 902         ILK_CURSOR_MAX_SRWM,
 903         ILK_CURSOR_DFT_SRWM,
 904         2,
 905         ILK_FIFO_LINE_SIZE
 906 };
 907
 908 static const struct intel_watermark_params sandybridge_display_wm_info = {
 909         SNB_DISPLAY_FIFO,
 910         SNB_DISPLAY_MAXWM,
 911         SNB_DISPLAY_DFTWM,
 912         2,
 913         SNB_FIFO_LINE_SIZE
 914 };
 915 static const struct intel_watermark_params sandybridge_cursor_wm_info = {
 916         SNB_CURSOR_FIFO,
 917         SNB_CURSOR_MAXWM,
 918         SNB_CURSOR_DFTWM,
 919         2,
 920         SNB_FIFO_LINE_SIZE
 921 };
 922 static const struct intel_watermark_params sandybridge_display_srwm_info = {
 923         SNB_DISPLAY_SR_FIFO,
 924         SNB_DISPLAY_MAX_SRWM,
 925         SNB_DISPLAY_DFT_SRWM,
 926         2,
 927         SNB_FIFO_LINE_SIZE
 928 };
 929 static const struct intel_watermark_params sandybridge_cursor_srwm_info = {
 930         SNB_CURSOR_SR_FIFO,
 931         SNB_CURSOR_MAX_SRWM,
 932         SNB_CURSOR_DFT_SRWM,
 933         2,
 934         SNB_FIFO_LINE_SIZE
 935 };
 936
 937
 938 /**
 939  * intel_calculate_wm - calculate watermark level
 940  * @clock_in_khz: pixel clock
 941  * @wm: chip FIFO params
 942  * @pixel_size: display pixel size
 943  * @latency_ns: memory latency for the platform
 944  *
 945  * Calculate the watermark level (the level at which the display plane will
 946  * start fetching from memory again).  Each chip has a different display
 947  * FIFO size and allocation, so the caller needs to figure that out and pass
 948  * in the correct intel_watermark_params structure.
 949  *
 950  * As the pixel clock runs, the FIFO will be drained at a rate that depends
 951  * on the pixel size.  When it reaches the watermark level, it'll start
 952  * fetching FIFO line sized based chunks from memory until the FIFO fills
 953  * past the watermark point.  If the FIFO drains completely, a FIFO underrun
 954  * will occur, and a display engine hang could result.
 955  */
 956 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
 957                                         const struct intel_watermark_params *wm,
 958                                         int fifo_size,
 959                                         int pixel_size,
 960                                         unsigned long latency_ns)
 961 {
 962         long entries_required, wm_size;
 963
 964         /*
 965          * Note: we need to make sure we don't overflow for various clock &
 966          * latency values.
 967          * clocks go from a few thousand to several hundred thousand.
 968          * latency is usually a few thousand
 969          */
 970         entries_required = ((clock_in_khz / 1000) * pixel_size * latency_ns) /
 971                 1000;
 972         entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
 973
 974         DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
 975
 976         wm_size = fifo_size - (entries_required + wm->guard_size);
 977
 978         DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
 979
 980         /* Don't promote wm_size to unsigned... */
 981         if (wm_size > (long)wm->max_wm)
 982                 wm_size = wm->max_wm;
 983         if (wm_size <= 0)
 984                 wm_size = wm->default_wm;
 985         return wm_size;
 986 }
 987
 988 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
 989 {
 990         struct drm_crtc *crtc, *enabled = NULL;
 991
 992         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
 993                 if (crtc->enabled && crtc->fb) {
 994                         if (enabled)
 995                                 return NULL;
 996                         enabled = crtc;
 997                 }
 998         }
 999
1000         return enabled;
1001 }
1002
1003 static void pineview_update_wm(struct drm_device *dev)
1004 {
1005         struct drm_i915_private *dev_priv = dev->dev_private;
1006         struct drm_crtc *crtc;
1007         const struct cxsr_latency *latency;
1008         u32 reg;
1009         unsigned long wm;
1010
1011         latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
1012                                          dev_priv->fsb_freq, dev_priv->mem_freq);
1013         if (!latency) {
1014                 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
1015                 pineview_disable_cxsr(dev);
1016                 return;
1017         }
1018
1019         crtc = single_enabled_crtc(dev);
1020         if (crtc) {
1021                 int clock = crtc->mode.clock;
1022                 int pixel_size = crtc->fb->bits_per_pixel / 8;
1023
1024                 /* Display SR */
1025                 wm = intel_calculate_wm(clock, &pineview_display_wm,
1026                                         pineview_display_wm.fifo_size,
1027                                         pixel_size, latency->display_sr);
1028                 reg = I915_READ(DSPFW1);
1029                 reg &= ~DSPFW_SR_MASK;
1030                 reg |= wm << DSPFW_SR_SHIFT;
1031                 I915_WRITE(DSPFW1, reg);
1032                 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
1033
1034                 /* cursor SR */
1035                 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
1036                                         pineview_display_wm.fifo_size,
1037                                         pixel_size, latency->cursor_sr);
1038                 reg = I915_READ(DSPFW3);
1039                 reg &= ~DSPFW_CURSOR_SR_MASK;
1040                 reg |= (wm & 0x3f) << DSPFW_CURSOR_SR_SHIFT;
1041                 I915_WRITE(DSPFW3, reg);
1042
1043                 /* Display HPLL off SR */
1044                 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
1045                                         pineview_display_hplloff_wm.fifo_size,
1046                                         pixel_size, latency->display_hpll_disable);
1047                 reg = I915_READ(DSPFW3);
1048                 reg &= ~DSPFW_HPLL_SR_MASK;
1049                 reg |= wm & DSPFW_HPLL_SR_MASK;
1050                 I915_WRITE(DSPFW3, reg);
1051
1052                 /* cursor HPLL off SR */
1053                 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
1054                                         pineview_display_hplloff_wm.fifo_size,
1055                                         pixel_size, latency->cursor_hpll_disable);
1056                 reg = I915_READ(DSPFW3);
1057                 reg &= ~DSPFW_HPLL_CURSOR_MASK;
1058                 reg |= (wm & 0x3f) << DSPFW_HPLL_CURSOR_SHIFT;
1059                 I915_WRITE(DSPFW3, reg);
1060                 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
1061
1062                 /* activate cxsr */
1063                 I915_WRITE(DSPFW3,
1064                            I915_READ(DSPFW3) | PINEVIEW_SELF_REFRESH_EN);
1065                 DRM_DEBUG_KMS("Self-refresh is enabled\n");
1066         } else {
1067                 pineview_disable_cxsr(dev);
1068                 DRM_DEBUG_KMS("Self-refresh is disabled\n");
1069         }
1070 }
1071
1072 static bool g4x_compute_wm0(struct drm_device *dev,
1073                             int plane,
1074                             const struct intel_watermark_params *display,
1075                             int display_latency_ns,
1076                             const struct intel_watermark_params *cursor,
1077                             int cursor_latency_ns,
1078                             int *plane_wm,
1079                             int *cursor_wm)
1080 {
1081         struct drm_crtc *crtc;
1082         int htotal, hdisplay, clock, pixel_size;
1083         int line_time_us, line_count;
1084         int entries, tlb_miss;
1085
1086         crtc = intel_get_crtc_for_plane(dev, plane);
1087         if (crtc->fb == NULL || !crtc->enabled) {
1088                 *cursor_wm = cursor->guard_size;
1089                 *plane_wm = display->guard_size;
1090                 return false;
1091         }
1092
1093         htotal = crtc->mode.htotal;
1094         hdisplay = crtc->mode.hdisplay;
1095         clock = crtc->mode.clock;
1096         pixel_size = crtc->fb->bits_per_pixel / 8;
1097
1098         /* Use the small buffer method to calculate plane watermark */
1099         entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
1100         tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
1101         if (tlb_miss > 0)
1102                 entries += tlb_miss;
1103         entries = DIV_ROUND_UP(entries, display->cacheline_size);
1104         *plane_wm = entries + display->guard_size;
1105         if (*plane_wm > (int)display->max_wm)
1106                 *plane_wm = display->max_wm;
1107
1108         /* Use the large buffer method to calculate cursor watermark */
1109         line_time_us = ((htotal * 1000) / clock);
1110         line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
1111         entries = line_count * 64 * pixel_size;
1112         tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
1113         if (tlb_miss > 0)
1114                 entries += tlb_miss;
1115         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1116         *cursor_wm = entries + cursor->guard_size;
1117         if (*cursor_wm > (int)cursor->max_wm)
1118                 *cursor_wm = (int)cursor->max_wm;
1119
1120         return true;
1121 }
1122
1123 /*
1124  * Check the wm result.
1125  *
1126  * If any calculated watermark values is larger than the maximum value that
1127  * can be programmed into the associated watermark register, that watermark
1128  * must be disabled.
1129  */
1130 static bool g4x_check_srwm(struct drm_device *dev,
1131                            int display_wm, int cursor_wm,
1132                            const struct intel_watermark_params *display,
1133                            const struct intel_watermark_params *cursor)
1134 {
1135         DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
1136                       display_wm, cursor_wm);
1137
1138         if (display_wm > display->max_wm) {
1139                 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
1140                               display_wm, display->max_wm);
1141                 return false;
1142         }
1143
1144         if (cursor_wm > cursor->max_wm) {
1145                 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
1146                               cursor_wm, cursor->max_wm);
1147                 return false;
1148         }
1149
1150         if (!(display_wm || cursor_wm)) {
1151                 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
1152                 return false;
1153         }
1154
1155         return true;
1156 }
1157
1158 static bool g4x_compute_srwm(struct drm_device *dev,
1159                              int plane,
1160                              int latency_ns,
1161                              const struct intel_watermark_params *display,
1162                              const struct intel_watermark_params *cursor,
1163                              int *display_wm, int *cursor_wm)
1164 {
1165         struct drm_crtc *crtc;
1166         int hdisplay, htotal, pixel_size, clock;
1167         unsigned long line_time_us;
1168         int line_count, line_size;
1169         int small, large;
1170         int entries;
1171
1172         if (!latency_ns) {
1173                 *display_wm = *cursor_wm = 0;
1174                 return false;
1175         }
1176
1177         crtc = intel_get_crtc_for_plane(dev, plane);
1178         hdisplay = crtc->mode.hdisplay;
1179         htotal = crtc->mode.htotal;
1180         clock = crtc->mode.clock;
1181         pixel_size = crtc->fb->bits_per_pixel / 8;
1182
1183         line_time_us = (htotal * 1000) / clock;
1184         line_count = (latency_ns / line_time_us + 1000) / 1000;
1185         line_size = hdisplay * pixel_size;
1186
1187         /* Use the minimum of the small and large buffer method for primary */
1188         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1189         large = line_count * line_size;
1190
1191         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1192         *display_wm = entries + display->guard_size;
1193
1194         /* calculate the self-refresh watermark for display cursor */
1195         entries = line_count * pixel_size * 64;
1196         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1197         *cursor_wm = entries + cursor->guard_size;
1198
1199         return g4x_check_srwm(dev,
1200                               *display_wm, *cursor_wm,
1201                               display, cursor);
1202 }
1203
1204 static bool vlv_compute_drain_latency(struct drm_device *dev,
1205                                      int plane,
1206                                      int *plane_prec_mult,
1207                                      int *plane_dl,
1208                                      int *cursor_prec_mult,
1209                                      int *cursor_dl)
1210 {
1211         struct drm_crtc *crtc;
1212         int clock, pixel_size;
1213         int entries;
1214
1215         crtc = intel_get_crtc_for_plane(dev, plane);
1216         if (crtc->fb == NULL || !crtc->enabled)
1217                 return false;
1218
1219         clock = crtc->mode.clock;       /* VESA DOT Clock */
1220         pixel_size = crtc->fb->bits_per_pixel / 8;      /* BPP */
1221
1222         entries = (clock / 1000) * pixel_size;
1223         *plane_prec_mult = (entries > 256) ?
1224                 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1225         *plane_dl = (64 * (*plane_prec_mult) * 4) / ((clock / 1000) *
1226                                                      pixel_size);
1227
1228         entries = (clock / 1000) * 4;   /* BPP is always 4 for cursor */
1229         *cursor_prec_mult = (entries > 256) ?
1230                 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1231         *cursor_dl = (64 * (*cursor_prec_mult) * 4) / ((clock / 1000) * 4);
1232
1233         return true;
1234 }
1235
1236 /*
1237  * Update drain latency registers of memory arbiter
1238  *
1239  * Valleyview SoC has a new memory arbiter and needs drain latency registers
1240  * to be programmed. Each plane has a drain latency multiplier and a drain
1241  * latency value.
1242  */
1243
1244 static void vlv_update_drain_latency(struct drm_device *dev)
1245 {
1246         struct drm_i915_private *dev_priv = dev->dev_private;
1247         int planea_prec, planea_dl, planeb_prec, planeb_dl;
1248         int cursora_prec, cursora_dl, cursorb_prec, cursorb_dl;
1249         int plane_prec_mult, cursor_prec_mult; /* Precision multiplier is
1250                                                         either 16 or 32 */
1251
1252         /* For plane A, Cursor A */
1253         if (vlv_compute_drain_latency(dev, 0, &plane_prec_mult, &planea_dl,
1254                                       &cursor_prec_mult, &cursora_dl)) {
1255                 cursora_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1256                         DDL_CURSORA_PRECISION_32 : DDL_CURSORA_PRECISION_16;
1257                 planea_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1258                         DDL_PLANEA_PRECISION_32 : DDL_PLANEA_PRECISION_16;
1259
1260                 I915_WRITE(VLV_DDL1, cursora_prec |
1261                                 (cursora_dl << DDL_CURSORA_SHIFT) |
1262                                 planea_prec | planea_dl);
1263         }
1264
1265         /* For plane B, Cursor B */
1266         if (vlv_compute_drain_latency(dev, 1, &plane_prec_mult, &planeb_dl,
1267                                       &cursor_prec_mult, &cursorb_dl)) {
1268                 cursorb_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1269                         DDL_CURSORB_PRECISION_32 : DDL_CURSORB_PRECISION_16;
1270                 planeb_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1271                         DDL_PLANEB_PRECISION_32 : DDL_PLANEB_PRECISION_16;
1272
1273                 I915_WRITE(VLV_DDL2, cursorb_prec |
1274                                 (cursorb_dl << DDL_CURSORB_SHIFT) |
1275                                 planeb_prec | planeb_dl);
1276         }
1277 }
1278
1279 #define single_plane_enabled(mask) is_power_of_2(mask)
1280
1281 static void valleyview_update_wm(struct drm_device *dev)
1282 {
1283         static const int sr_latency_ns = 12000;
1284         struct drm_i915_private *dev_priv = dev->dev_private;
1285         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1286         int plane_sr, cursor_sr;
1287         unsigned int enabled = 0;
1288
1289         vlv_update_drain_latency(dev);
1290
1291         if (g4x_compute_wm0(dev, 0,
1292                             &valleyview_wm_info, latency_ns,
1293                             &valleyview_cursor_wm_info, latency_ns,
1294                             &planea_wm, &cursora_wm))
1295                 enabled |= 1;
1296
1297         if (g4x_compute_wm0(dev, 1,
1298                             &valleyview_wm_info, latency_ns,
1299                             &valleyview_cursor_wm_info, latency_ns,
1300                             &planeb_wm, &cursorb_wm))
1301                 enabled |= 2;
1302
1303         plane_sr = cursor_sr = 0;
1304         if (single_plane_enabled(enabled) &&
1305             g4x_compute_srwm(dev, ffs(enabled) - 1,
1306                              sr_latency_ns,
1307                              &valleyview_wm_info,
1308                              &valleyview_cursor_wm_info,
1309                              &plane_sr, &cursor_sr))
1310                 I915_WRITE(FW_BLC_SELF_VLV, FW_CSPWRDWNEN);
1311         else
1312                 I915_WRITE(FW_BLC_SELF_VLV,
1313                            I915_READ(FW_BLC_SELF_VLV) & ~FW_CSPWRDWNEN);
1314
1315         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1316                       planea_wm, cursora_wm,
1317                       planeb_wm, cursorb_wm,
1318                       plane_sr, cursor_sr);
1319
1320         I915_WRITE(DSPFW1,
1321                    (plane_sr << DSPFW_SR_SHIFT) |
1322                    (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1323                    (planeb_wm << DSPFW_PLANEB_SHIFT) |
1324                    planea_wm);
1325         I915_WRITE(DSPFW2,
1326                    (I915_READ(DSPFW2) & DSPFW_CURSORA_MASK) |
1327                    (cursora_wm << DSPFW_CURSORA_SHIFT));
1328         I915_WRITE(DSPFW3,
1329                    (I915_READ(DSPFW3) | (cursor_sr << DSPFW_CURSOR_SR_SHIFT)));
1330 }
1331
1332 static void g4x_update_wm(struct drm_device *dev)
1333 {
1334         static const int sr_latency_ns = 12000;
1335         struct drm_i915_private *dev_priv = dev->dev_private;
1336         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1337         int plane_sr, cursor_sr;
1338         unsigned int enabled = 0;
1339
1340         if (g4x_compute_wm0(dev, 0,
1341                             &g4x_wm_info, latency_ns,
1342                             &g4x_cursor_wm_info, latency_ns,
1343                             &planea_wm, &cursora_wm))
1344                 enabled |= 1;
1345
1346         if (g4x_compute_wm0(dev, 1,
1347                             &g4x_wm_info, latency_ns,
1348                             &g4x_cursor_wm_info, latency_ns,
1349                             &planeb_wm, &cursorb_wm))
1350                 enabled |= 2;
1351
1352         plane_sr = cursor_sr = 0;
1353         if (single_plane_enabled(enabled) &&
1354             g4x_compute_srwm(dev, ffs(enabled) - 1,
1355                              sr_latency_ns,
1356                              &g4x_wm_info,
1357                              &g4x_cursor_wm_info,
1358                              &plane_sr, &cursor_sr))
1359                 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1360         else
1361                 I915_WRITE(FW_BLC_SELF,
1362                            I915_READ(FW_BLC_SELF) & ~FW_BLC_SELF_EN);
1363
1364         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1365                       planea_wm, cursora_wm,
1366                       planeb_wm, cursorb_wm,
1367                       plane_sr, cursor_sr);
1368
1369         I915_WRITE(DSPFW1,
1370                    (plane_sr << DSPFW_SR_SHIFT) |
1371                    (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1372                    (planeb_wm << DSPFW_PLANEB_SHIFT) |
1373                    planea_wm);
1374         I915_WRITE(DSPFW2,
1375                    (I915_READ(DSPFW2) & DSPFW_CURSORA_MASK) |
1376                    (cursora_wm << DSPFW_CURSORA_SHIFT));
1377         /* HPLL off in SR has some issues on G4x... disable it */
1378         I915_WRITE(DSPFW3,
1379                    (I915_READ(DSPFW3) & ~DSPFW_HPLL_SR_EN) |
1380                    (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1381 }
1382
1383 static void i965_update_wm(struct drm_device *dev)
1384 {
1385         struct drm_i915_private *dev_priv = dev->dev_private;
1386         struct drm_crtc *crtc;
1387         int srwm = 1;
1388         int cursor_sr = 16;
1389
1390         /* Calc sr entries for one plane configs */
1391         crtc = single_enabled_crtc(dev);
1392         if (crtc) {
1393                 /* self-refresh has much higher latency */
1394                 static const int sr_latency_ns = 12000;
1395                 int clock = crtc->mode.clock;
1396                 int htotal = crtc->mode.htotal;
1397                 int hdisplay = crtc->mode.hdisplay;
1398                 int pixel_size = crtc->fb->bits_per_pixel / 8;
1399                 unsigned long line_time_us;
1400                 int entries;
1401
1402                 line_time_us = ((htotal * 1000) / clock);
1403
1404                 /* Use ns/us then divide to preserve precision */
1405                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1406                         pixel_size * hdisplay;
1407                 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1408                 srwm = I965_FIFO_SIZE - entries;
1409                 if (srwm < 0)
1410                         srwm = 1;
1411                 srwm &= 0x1ff;
1412                 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1413                               entries, srwm);
1414
1415                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1416                         pixel_size * 64;
1417                 entries = DIV_ROUND_UP(entries,
1418                                           i965_cursor_wm_info.cacheline_size);
1419                 cursor_sr = i965_cursor_wm_info.fifo_size -
1420                         (entries + i965_cursor_wm_info.guard_size);
1421
1422                 if (cursor_sr > i965_cursor_wm_info.max_wm)
1423                         cursor_sr = i965_cursor_wm_info.max_wm;
1424
1425                 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1426                               "cursor %d\n", srwm, cursor_sr);
1427
1428                 if (IS_CRESTLINE(dev))
1429                         I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1430         } else {
1431                 /* Turn off self refresh if both pipes are enabled */
1432                 if (IS_CRESTLINE(dev))
1433                         I915_WRITE(FW_BLC_SELF, I915_READ(FW_BLC_SELF)
1434                                    & ~FW_BLC_SELF_EN);
1435         }
1436
1437         DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1438                       srwm);
1439
1440         /* 965 has limitations... */
1441         I915_WRITE(DSPFW1, (srwm << DSPFW_SR_SHIFT) |
1442                    (8 << 16) | (8 << 8) | (8 << 0));
1443         I915_WRITE(DSPFW2, (8 << 8) | (8 << 0));
1444         /* update cursor SR watermark */
1445         I915_WRITE(DSPFW3, (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1446 }
1447
1448 static void i9xx_update_wm(struct drm_device *dev)
1449 {
1450         struct drm_i915_private *dev_priv = dev->dev_private;
1451         const struct intel_watermark_params *wm_info;
1452         uint32_t fwater_lo;
1453         uint32_t fwater_hi;
1454         int cwm, srwm = 1;
1455         int fifo_size;
1456         int planea_wm, planeb_wm;
1457         struct drm_crtc *crtc, *enabled = NULL;
1458
1459         if (IS_I945GM(dev))
1460                 wm_info = &i945_wm_info;
1461         else if (!IS_GEN2(dev))
1462                 wm_info = &i915_wm_info;
1463         else
1464                 wm_info = &i855_wm_info;
1465
1466         fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1467         crtc = intel_get_crtc_for_plane(dev, 0);
1468         if (crtc->enabled && crtc->fb) {
1469                 planea_wm = intel_calculate_wm(crtc->mode.clock,
1470                                                wm_info, fifo_size,
1471                                                crtc->fb->bits_per_pixel / 8,
1472                                                latency_ns);
1473                 enabled = crtc;
1474         } else
1475                 planea_wm = fifo_size - wm_info->guard_size;
1476
1477         fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1478         crtc = intel_get_crtc_for_plane(dev, 1);
1479         if (crtc->enabled && crtc->fb) {
1480                 planeb_wm = intel_calculate_wm(crtc->mode.clock,
1481                                                wm_info, fifo_size,
1482                                                crtc->fb->bits_per_pixel / 8,
1483                                                latency_ns);
1484                 if (enabled == NULL)
1485                         enabled = crtc;
1486                 else
1487                         enabled = NULL;
1488         } else
1489                 planeb_wm = fifo_size - wm_info->guard_size;
1490
1491         DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1492
1493         /*
1494          * Overlay gets an aggressive default since video jitter is bad.
1495          */
1496         cwm = 2;
1497
1498         /* Play safe and disable self-refresh before adjusting watermarks. */
1499         if (IS_I945G(dev) || IS_I945GM(dev))
1500                 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN_MASK | 0);
1501         else if (IS_I915GM(dev))
1502                 I915_WRITE(INSTPM, I915_READ(INSTPM) & ~INSTPM_SELF_EN);
1503
1504         /* Calc sr entries for one plane configs */
1505         if (HAS_FW_BLC(dev) && enabled) {
1506                 /* self-refresh has much higher latency */
1507                 static const int sr_latency_ns = 6000;
1508                 int clock = enabled->mode.clock;
1509                 int htotal = enabled->mode.htotal;
1510                 int hdisplay = enabled->mode.hdisplay;
1511                 int pixel_size = enabled->fb->bits_per_pixel / 8;
1512                 unsigned long line_time_us;
1513                 int entries;
1514
1515                 line_time_us = (htotal * 1000) / clock;
1516
1517                 /* Use ns/us then divide to preserve precision */
1518                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1519                         pixel_size * hdisplay;
1520                 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1521                 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1522                 srwm = wm_info->fifo_size - entries;
1523                 if (srwm < 0)
1524                         srwm = 1;
1525
1526                 if (IS_I945G(dev) || IS_I945GM(dev))
1527                         I915_WRITE(FW_BLC_SELF,
1528                                    FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1529                 else if (IS_I915GM(dev))
1530                         I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1531         }
1532
1533         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1534                       planea_wm, planeb_wm, cwm, srwm);
1535
1536         fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1537         fwater_hi = (cwm & 0x1f);
1538
1539         /* Set request length to 8 cachelines per fetch */
1540         fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1541         fwater_hi = fwater_hi | (1 << 8);
1542
1543         I915_WRITE(FW_BLC, fwater_lo);
1544         I915_WRITE(FW_BLC2, fwater_hi);
1545
1546         if (HAS_FW_BLC(dev)) {
1547                 if (enabled) {
1548                         if (IS_I945G(dev) || IS_I945GM(dev))
1549                                 I915_WRITE(FW_BLC_SELF,
1550                                            FW_BLC_SELF_EN_MASK | FW_BLC_SELF_EN);
1551                         else if (IS_I915GM(dev))
1552                                 I915_WRITE(INSTPM, I915_READ(INSTPM) | INSTPM_SELF_EN);
1553                         DRM_DEBUG_KMS("memory self refresh enabled\n");
1554                 } else
1555                         DRM_DEBUG_KMS("memory self refresh disabled\n");
1556         }
1557 }
1558
1559 static void i830_update_wm(struct drm_device *dev)
1560 {
1561         struct drm_i915_private *dev_priv = dev->dev_private;
1562         struct drm_crtc *crtc;
1563         uint32_t fwater_lo;
1564         int planea_wm;
1565
1566         crtc = single_enabled_crtc(dev);
1567         if (crtc == NULL)
1568                 return;
1569
1570         planea_wm = intel_calculate_wm(crtc->mode.clock, &i830_wm_info,
1571                                        dev_priv->display.get_fifo_size(dev, 0),
1572                                        crtc->fb->bits_per_pixel / 8,
1573                                        latency_ns);
1574         fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1575         fwater_lo |= (3<<8) | planea_wm;
1576
1577         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1578
1579         I915_WRITE(FW_BLC, fwater_lo);
1580 }
1581
1582 #define ILK_LP0_PLANE_LATENCY           700
1583 #define ILK_LP0_CURSOR_LATENCY          1300
1584
1585 /*
1586  * Check the wm result.
1587  *
1588  * If any calculated watermark values is larger than the maximum value that
1589  * can be programmed into the associated watermark register, that watermark
1590  * must be disabled.
1591  */
1592 static bool ironlake_check_srwm(struct drm_device *dev, int level,
1593                                 int fbc_wm, int display_wm, int cursor_wm,
1594                                 const struct intel_watermark_params *display,
1595                                 const struct intel_watermark_params *cursor)
1596 {
1597         struct drm_i915_private *dev_priv = dev->dev_private;
1598
1599         DRM_DEBUG_KMS("watermark %d: display plane %d, fbc lines %d,"
1600                       " cursor %d\n", level, display_wm, fbc_wm, cursor_wm);
1601
1602         if (fbc_wm > SNB_FBC_MAX_SRWM) {
1603                 DRM_DEBUG_KMS("fbc watermark(%d) is too large(%d), disabling wm%d+\n",
1604                               fbc_wm, SNB_FBC_MAX_SRWM, level);
1605
1606                 /* fbc has it's own way to disable FBC WM */
1607                 I915_WRITE(DISP_ARB_CTL,
1608                            I915_READ(DISP_ARB_CTL) | DISP_FBC_WM_DIS);
1609                 return false;
1610         }
1611
1612         if (display_wm > display->max_wm) {
1613                 DRM_DEBUG_KMS("display watermark(%d) is too large(%d), disabling wm%d+\n",
1614                               display_wm, SNB_DISPLAY_MAX_SRWM, level);
1615                 return false;
1616         }
1617
1618         if (cursor_wm > cursor->max_wm) {
1619                 DRM_DEBUG_KMS("cursor watermark(%d) is too large(%d), disabling wm%d+\n",
1620                               cursor_wm, SNB_CURSOR_MAX_SRWM, level);
1621                 return false;
1622         }
1623
1624         if (!(fbc_wm || display_wm || cursor_wm)) {
1625                 DRM_DEBUG_KMS("latency %d is 0, disabling wm%d+\n", level, level);
1626                 return false;
1627         }
1628
1629         return true;
1630 }
1631
1632 /*
1633  * Compute watermark values of WM[1-3],
1634  */
1635 static bool ironlake_compute_srwm(struct drm_device *dev, int level, int plane,
1636                                   int latency_ns,
1637                                   const struct intel_watermark_params *display,
1638                                   const struct intel_watermark_params *cursor,
1639                                   int *fbc_wm, int *display_wm, int *cursor_wm)
1640 {
1641         struct drm_crtc *crtc;
1642         unsigned long line_time_us;
1643         int hdisplay, htotal, pixel_size, clock;
1644         int line_count, line_size;
1645         int small, large;
1646         int entries;
1647
1648         if (!latency_ns) {
1649                 *fbc_wm = *display_wm = *cursor_wm = 0;
1650                 return false;
1651         }
1652
1653         crtc = intel_get_crtc_for_plane(dev, plane);
1654         hdisplay = crtc->mode.hdisplay;
1655         htotal = crtc->mode.htotal;
1656         clock = crtc->mode.clock;
1657         pixel_size = crtc->fb->bits_per_pixel / 8;
1658
1659         line_time_us = (htotal * 1000) / clock;
1660         line_count = (latency_ns / line_time_us + 1000) / 1000;
1661         line_size = hdisplay * pixel_size;
1662
1663         /* Use the minimum of the small and large buffer method for primary */
1664         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1665         large = line_count * line_size;
1666
1667         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1668         *display_wm = entries + display->guard_size;
1669
1670         /*
1671          * Spec says:
1672          * FBC WM = ((Final Primary WM * 64) / number of bytes per line) + 2
1673          */
1674         *fbc_wm = DIV_ROUND_UP(*display_wm * 64, line_size) + 2;
1675
1676         /* calculate the self-refresh watermark for display cursor */
1677         entries = line_count * pixel_size * 64;
1678         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1679         *cursor_wm = entries + cursor->guard_size;
1680
1681         return ironlake_check_srwm(dev, level,
1682                                    *fbc_wm, *display_wm, *cursor_wm,
1683                                    display, cursor);
1684 }
1685
1686 static void ironlake_update_wm(struct drm_device *dev)
1687 {
1688         struct drm_i915_private *dev_priv = dev->dev_private;
1689         int fbc_wm, plane_wm, cursor_wm;
1690         unsigned int enabled;
1691
1692         enabled = 0;
1693         if (g4x_compute_wm0(dev, 0,
1694                             &ironlake_display_wm_info,
1695                             ILK_LP0_PLANE_LATENCY,
1696                             &ironlake_cursor_wm_info,
1697                             ILK_LP0_CURSOR_LATENCY,
1698                             &plane_wm, &cursor_wm)) {
1699                 I915_WRITE(WM0_PIPEA_ILK,
1700                            (plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm);
1701                 DRM_DEBUG_KMS("FIFO watermarks For pipe A -"
1702                               " plane %d, " "cursor: %d\n",
1703                               plane_wm, cursor_wm);
1704                 enabled |= 1;
1705         }
1706
1707         if (g4x_compute_wm0(dev, 1,
1708                             &ironlake_display_wm_info,
1709                             ILK_LP0_PLANE_LATENCY,
1710                             &ironlake_cursor_wm_info,
1711                             ILK_LP0_CURSOR_LATENCY,
1712                             &plane_wm, &cursor_wm)) {
1713                 I915_WRITE(WM0_PIPEB_ILK,
1714                            (plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm);
1715                 DRM_DEBUG_KMS("FIFO watermarks For pipe B -"
1716                               " plane %d, cursor: %d\n",
1717                               plane_wm, cursor_wm);
1718                 enabled |= 2;
1719         }
1720
1721         /*
1722          * Calculate and update the self-refresh watermark only when one
1723          * display plane is used.
1724          */
1725         I915_WRITE(WM3_LP_ILK, 0);
1726         I915_WRITE(WM2_LP_ILK, 0);
1727         I915_WRITE(WM1_LP_ILK, 0);
1728
1729         if (!single_plane_enabled(enabled))
1730                 return;
1731         enabled = ffs(enabled) - 1;
1732
1733         /* WM1 */
1734         if (!ironlake_compute_srwm(dev, 1, enabled,
1735                                    ILK_READ_WM1_LATENCY() * 500,
1736                                    &ironlake_display_srwm_info,
1737                                    &ironlake_cursor_srwm_info,
1738                                    &fbc_wm, &plane_wm, &cursor_wm))
1739                 return;
1740
1741         I915_WRITE(WM1_LP_ILK,
1742                    WM1_LP_SR_EN |
1743                    (ILK_READ_WM1_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1744                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1745                    (plane_wm << WM1_LP_SR_SHIFT) |
1746                    cursor_wm);
1747
1748         /* WM2 */
1749         if (!ironlake_compute_srwm(dev, 2, enabled,
1750                                    ILK_READ_WM2_LATENCY() * 500,
1751                                    &ironlake_display_srwm_info,
1752                                    &ironlake_cursor_srwm_info,
1753                                    &fbc_wm, &plane_wm, &cursor_wm))
1754                 return;
1755
1756         I915_WRITE(WM2_LP_ILK,
1757                    WM2_LP_EN |
1758                    (ILK_READ_WM2_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1759                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1760                    (plane_wm << WM1_LP_SR_SHIFT) |
1761                    cursor_wm);
1762
1763         /*
1764          * WM3 is unsupported on ILK, probably because we don't have latency
1765          * data for that power state
1766          */
1767 }
1768
1769 static void sandybridge_update_wm(struct drm_device *dev)
1770 {
1771         struct drm_i915_private *dev_priv = dev->dev_private;
1772         int latency = SNB_READ_WM0_LATENCY() * 100;     /* In unit 0.1us */
1773         u32 val;
1774         int fbc_wm, plane_wm, cursor_wm;
1775         unsigned int enabled;
1776
1777         enabled = 0;
1778         if (g4x_compute_wm0(dev, 0,
1779                             &sandybridge_display_wm_info, latency,
1780                             &sandybridge_cursor_wm_info, latency,
1781                             &plane_wm, &cursor_wm)) {
1782                 val = I915_READ(WM0_PIPEA_ILK);
1783                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
1784                 I915_WRITE(WM0_PIPEA_ILK, val |
1785                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
1786                 DRM_DEBUG_KMS("FIFO watermarks For pipe A -"
1787                               " plane %d, " "cursor: %d\n",
1788                               plane_wm, cursor_wm);
1789                 enabled |= 1;
1790         }
1791
1792         if (g4x_compute_wm0(dev, 1,
1793                             &sandybridge_display_wm_info, latency,
1794                             &sandybridge_cursor_wm_info, latency,
1795                             &plane_wm, &cursor_wm)) {
1796                 val = I915_READ(WM0_PIPEB_ILK);
1797                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
1798                 I915_WRITE(WM0_PIPEB_ILK, val |
1799                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
1800                 DRM_DEBUG_KMS("FIFO watermarks For pipe B -"
1801                               " plane %d, cursor: %d\n",
1802                               plane_wm, cursor_wm);
1803                 enabled |= 2;
1804         }
1805
1806         if ((dev_priv->num_pipe == 3) &&
1807             g4x_compute_wm0(dev, 2,
1808                             &sandybridge_display_wm_info, latency,
1809                             &sandybridge_cursor_wm_info, latency,
1810                             &plane_wm, &cursor_wm)) {
1811                 val = I915_READ(WM0_PIPEC_IVB);
1812                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
1813                 I915_WRITE(WM0_PIPEC_IVB, val |
1814                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
1815                 DRM_DEBUG_KMS("FIFO watermarks For pipe C -"
1816                               " plane %d, cursor: %d\n",
1817                               plane_wm, cursor_wm);
1818                 enabled |= 3;
1819         }
1820
1821         /*
1822          * Calculate and update the self-refresh watermark only when one
1823          * display plane is used.
1824          *
1825          * SNB support 3 levels of watermark.
1826          *
1827          * WM1/WM2/WM2 watermarks have to be enabled in the ascending order,
1828          * and disabled in the descending order
1829          *
1830          */
1831         I915_WRITE(WM3_LP_ILK, 0);
1832         I915_WRITE(WM2_LP_ILK, 0);
1833         I915_WRITE(WM1_LP_ILK, 0);
1834
1835         if (!single_plane_enabled(enabled) ||
1836             dev_priv->sprite_scaling_enabled)
1837                 return;
1838         enabled = ffs(enabled) - 1;
1839
1840         /* WM1 */
1841         if (!ironlake_compute_srwm(dev, 1, enabled,
1842                                    SNB_READ_WM1_LATENCY() * 500,
1843                                    &sandybridge_display_srwm_info,
1844                                    &sandybridge_cursor_srwm_info,
1845                                    &fbc_wm, &plane_wm, &cursor_wm))
1846                 return;
1847
1848         I915_WRITE(WM1_LP_ILK,
1849                    WM1_LP_SR_EN |
1850                    (SNB_READ_WM1_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1851                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1852                    (plane_wm << WM1_LP_SR_SHIFT) |
1853                    cursor_wm);
1854
1855         /* WM2 */
1856         if (!ironlake_compute_srwm(dev, 2, enabled,
1857                                    SNB_READ_WM2_LATENCY() * 500,
1858                                    &sandybridge_display_srwm_info,
1859                                    &sandybridge_cursor_srwm_info,
1860                                    &fbc_wm, &plane_wm, &cursor_wm))
1861                 return;
1862
1863         I915_WRITE(WM2_LP_ILK,
1864                    WM2_LP_EN |
1865                    (SNB_READ_WM2_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1866                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1867                    (plane_wm << WM1_LP_SR_SHIFT) |
1868                    cursor_wm);
1869
1870         /* WM3 */
1871         if (!ironlake_compute_srwm(dev, 3, enabled,
1872                                    SNB_READ_WM3_LATENCY() * 500,
1873                                    &sandybridge_display_srwm_info,
1874                                    &sandybridge_cursor_srwm_info,
1875                                    &fbc_wm, &plane_wm, &cursor_wm))
1876                 return;
1877
1878         I915_WRITE(WM3_LP_ILK,
1879                    WM3_LP_EN |
1880                    (SNB_READ_WM3_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1881                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1882                    (plane_wm << WM1_LP_SR_SHIFT) |
1883                    cursor_wm);
1884 }
1885
1886 static void
1887 haswell_update_linetime_wm(struct drm_device *dev, int pipe,
1888                                  struct drm_display_mode *mode)
1889 {
1890         struct drm_i915_private *dev_priv = dev->dev_private;
1891         u32 temp;
1892
1893         temp = I915_READ(PIPE_WM_LINETIME(pipe));
1894         temp &= ~PIPE_WM_LINETIME_MASK;
1895
1896         /* The WM are computed with base on how long it takes to fill a single
1897          * row at the given clock rate, multiplied by 8.
1898          * */
1899         temp |= PIPE_WM_LINETIME_TIME(
1900                 ((mode->crtc_hdisplay * 1000) / mode->clock) * 8);
1901
1902         /* IPS watermarks are only used by pipe A, and are ignored by
1903          * pipes B and C.  They are calculated similarly to the common
1904          * linetime values, except that we are using CD clock frequency
1905          * in MHz instead of pixel rate for the division.
1906          *
1907          * This is a placeholder for the IPS watermark calculation code.
1908          */
1909
1910         I915_WRITE(PIPE_WM_LINETIME(pipe), temp);
1911 }
1912
1913 static bool
1914 sandybridge_compute_sprite_wm(struct drm_device *dev, int plane,
1915                               uint32_t sprite_width, int pixel_size,
1916                               const struct intel_watermark_params *display,
1917                               int display_latency_ns, int *sprite_wm)
1918 {
1919         struct drm_crtc *crtc;
1920         int clock;
1921         int entries, tlb_miss;
1922
1923         crtc = intel_get_crtc_for_plane(dev, plane);
1924         if (crtc->fb == NULL || !crtc->enabled) {
1925                 *sprite_wm = display->guard_size;
1926                 return false;
1927         }
1928
1929         clock = crtc->mode.clock;
1930
1931         /* Use the small buffer method to calculate the sprite watermark */
1932         entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
1933         tlb_miss = display->fifo_size*display->cacheline_size -
1934                 sprite_width * 8;
1935         if (tlb_miss > 0)
1936                 entries += tlb_miss;
1937         entries = DIV_ROUND_UP(entries, display->cacheline_size);
1938         *sprite_wm = entries + display->guard_size;
1939         if (*sprite_wm > (int)display->max_wm)
1940                 *sprite_wm = display->max_wm;
1941
1942         return true;
1943 }
1944
1945 static bool
1946 sandybridge_compute_sprite_srwm(struct drm_device *dev, int plane,
1947                                 uint32_t sprite_width, int pixel_size,
1948                                 const struct intel_watermark_params *display,
1949                                 int latency_ns, int *sprite_wm)
1950 {
1951         struct drm_crtc *crtc;
1952         unsigned long line_time_us;
1953         int clock;
1954         int line_count, line_size;
1955         int small, large;
1956         int entries;
1957
1958         if (!latency_ns) {
1959                 *sprite_wm = 0;
1960                 return false;
1961         }
1962
1963         crtc = intel_get_crtc_for_plane(dev, plane);
1964         clock = crtc->mode.clock;
1965         if (!clock) {
1966                 *sprite_wm = 0;
1967                 return false;
1968         }
1969
1970         line_time_us = (sprite_width * 1000) / clock;
1971         if (!line_time_us) {
1972                 *sprite_wm = 0;
1973                 return false;
1974         }
1975
1976         line_count = (latency_ns / line_time_us + 1000) / 1000;
1977         line_size = sprite_width * pixel_size;
1978
1979         /* Use the minimum of the small and large buffer method for primary */
1980         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1981         large = line_count * line_size;
1982
1983         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1984         *sprite_wm = entries + display->guard_size;
1985
1986         return *sprite_wm > 0x3ff ? false : true;
1987 }
1988
1989 static void sandybridge_update_sprite_wm(struct drm_device *dev, int pipe,
1990                                          uint32_t sprite_width, int pixel_size)
1991 {
1992         struct drm_i915_private *dev_priv = dev->dev_private;
1993         int latency = SNB_READ_WM0_LATENCY() * 100;     /* In unit 0.1us */
1994         u32 val;
1995         int sprite_wm, reg;
1996         int ret;
1997
1998         switch (pipe) {
1999         case 0:
2000                 reg = WM0_PIPEA_ILK;
2001                 break;
2002         case 1:
2003                 reg = WM0_PIPEB_ILK;
2004                 break;
2005         case 2:
2006                 reg = WM0_PIPEC_IVB;
2007                 break;
2008         default:
2009                 return; /* bad pipe */
2010         }
2011
2012         ret = sandybridge_compute_sprite_wm(dev, pipe, sprite_width, pixel_size,
2013                                             &sandybridge_display_wm_info,
2014                                             latency, &sprite_wm);
2015         if (!ret) {
2016                 DRM_DEBUG_KMS("failed to compute sprite wm for pipe %d\n",
2017                               pipe);
2018                 return;
2019         }
2020
2021         val = I915_READ(reg);
2022         val &= ~WM0_PIPE_SPRITE_MASK;
2023         I915_WRITE(reg, val | (sprite_wm << WM0_PIPE_SPRITE_SHIFT));
2024         DRM_DEBUG_KMS("sprite watermarks For pipe %d - %d\n", pipe, sprite_wm);
2025
2026
2027         ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
2028                                               pixel_size,
2029                                               &sandybridge_display_srwm_info,
2030                                               SNB_READ_WM1_LATENCY() * 500,
2031                                               &sprite_wm);
2032         if (!ret) {
2033                 DRM_DEBUG_KMS("failed to compute sprite lp1 wm on pipe %d\n",
2034                               pipe);
2035                 return;
2036         }
2037         I915_WRITE(WM1S_LP_ILK, sprite_wm);
2038
2039         /* Only IVB has two more LP watermarks for sprite */
2040         if (!IS_IVYBRIDGE(dev))
2041                 return;
2042
2043         ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
2044                                               pixel_size,
2045                                               &sandybridge_display_srwm_info,
2046                                               SNB_READ_WM2_LATENCY() * 500,
2047                                               &sprite_wm);
2048         if (!ret) {
2049                 DRM_DEBUG_KMS("failed to compute sprite lp2 wm on pipe %d\n",
2050                               pipe);
2051                 return;
2052         }
2053         I915_WRITE(WM2S_LP_IVB, sprite_wm);
2054
2055         ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
2056                                               pixel_size,
2057                                               &sandybridge_display_srwm_info,
2058                                               SNB_READ_WM3_LATENCY() * 500,
2059                                               &sprite_wm);
2060         if (!ret) {
2061                 DRM_DEBUG_KMS("failed to compute sprite lp3 wm on pipe %d\n",
2062                               pipe);
2063                 return;
2064         }
2065         I915_WRITE(WM3S_LP_IVB, sprite_wm);
2066 }
2067
2068 /**
2069  * intel_update_watermarks - update FIFO watermark values based on current modes
2070  *
2071  * Calculate watermark values for the various WM regs based on current mode
2072  * and plane configuration.
2073  *
2074  * There are several cases to deal with here:
2075  *   - normal (i.e. non-self-refresh)
2076  *   - self-refresh (SR) mode
2077  *   - lines are large relative to FIFO size (buffer can hold up to 2)
2078  *   - lines are small relative to FIFO size (buffer can hold more than 2
2079  *     lines), so need to account for TLB latency
2080  *
2081  *   The normal calculation is:
2082  *     watermark = dotclock * bytes per pixel * latency
2083  *   where latency is platform & configuration dependent (we assume pessimal
2084  *   values here).
2085  *
2086  *   The SR calculation is:
2087  *     watermark = (trunc(latency/line time)+1) * surface width *
2088  *       bytes per pixel
2089  *   where
2090  *     line time = htotal / dotclock
2091  *     surface width = hdisplay for normal plane and 64 for cursor
2092  *   and latency is assumed to be high, as above.
2093  *
2094  * The final value programmed to the register should always be rounded up,
2095  * and include an extra 2 entries to account for clock crossings.
2096  *
2097  * We don't use the sprite, so we can ignore that.  And on Crestline we have
2098  * to set the non-SR watermarks to 8.
2099  */
2100 void intel_update_watermarks(struct drm_device *dev)
2101 {
2102         struct drm_i915_private *dev_priv = dev->dev_private;
2103
2104         if (dev_priv->display.update_wm)
2105                 dev_priv->display.update_wm(dev);
2106 }
2107
2108 void intel_update_linetime_watermarks(struct drm_device *dev,
2109                 int pipe, struct drm_display_mode *mode)
2110 {
2111         struct drm_i915_private *dev_priv = dev->dev_private;
2112
2113         if (dev_priv->display.update_linetime_wm)
2114                 dev_priv->display.update_linetime_wm(dev, pipe, mode);
2115 }
2116
2117 void intel_update_sprite_watermarks(struct drm_device *dev, int pipe,
2118                                     uint32_t sprite_width, int pixel_size)
2119 {
2120         struct drm_i915_private *dev_priv = dev->dev_private;
2121
2122         if (dev_priv->display.update_sprite_wm)
2123                 dev_priv->display.update_sprite_wm(dev, pipe, sprite_width,
2124                                                    pixel_size);
2125 }
2126
2127 static struct drm_i915_gem_object *
2128 intel_alloc_context_page(struct drm_device *dev)
2129 {
2130         struct drm_i915_gem_object *ctx;
2131         int ret;
2132
2133         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2134
2135         ctx = i915_gem_alloc_object(dev, 4096);
2136         if (!ctx) {
2137                 DRM_DEBUG("failed to alloc power context, RC6 disabled\n");
2138                 return NULL;
2139         }
2140
2141         ret = i915_gem_object_pin(ctx, 4096, true, false);
2142         if (ret) {
2143                 DRM_ERROR("failed to pin power context: %d\n", ret);
2144                 goto err_unref;
2145         }
2146
2147         ret = i915_gem_object_set_to_gtt_domain(ctx, 1);
2148         if (ret) {
2149                 DRM_ERROR("failed to set-domain on power context: %d\n", ret);
2150                 goto err_unpin;
2151         }
2152
2153         return ctx;
2154
2155 err_unpin:
2156         i915_gem_object_unpin(ctx);
2157 err_unref:
2158         drm_gem_object_unreference(&ctx->base);
2159         mutex_unlock(&dev->struct_mutex);
2160         return NULL;
2161 }
2162
2163 /**
2164  * Lock protecting IPS related data structures
2165  *   - i915_mch_dev
2166  *   - dev_priv->max_delay
2167  *   - dev_priv->min_delay
2168  *   - dev_priv->fmax
2169  *   - dev_priv->gpu_busy
2170  *   - dev_priv->gfx_power
2171  */
2172 DEFINE_SPINLOCK(mchdev_lock);
2173
2174 /* Global for IPS driver to get at the current i915 device. Protected by
2175  * mchdev_lock. */
2176 static struct drm_i915_private *i915_mch_dev;
2177
2178 bool ironlake_set_drps(struct drm_device *dev, u8 val)
2179 {
2180         struct drm_i915_private *dev_priv = dev->dev_private;
2181         u16 rgvswctl;
2182
2183         assert_spin_locked(&mchdev_lock);
2184
2185         rgvswctl = I915_READ16(MEMSWCTL);
2186         if (rgvswctl & MEMCTL_CMD_STS) {
2187                 DRM_DEBUG("gpu busy, RCS change rejected\n");
2188                 return false; /* still busy with another command */
2189         }
2190
2191         rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
2192                 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
2193         I915_WRITE16(MEMSWCTL, rgvswctl);
2194         POSTING_READ16(MEMSWCTL);
2195
2196         rgvswctl |= MEMCTL_CMD_STS;
2197         I915_WRITE16(MEMSWCTL, rgvswctl);
2198
2199         return true;
2200 }
2201
2202 static void ironlake_enable_drps(struct drm_device *dev)
2203 {
2204         struct drm_i915_private *dev_priv = dev->dev_private;
2205         u32 rgvmodectl = I915_READ(MEMMODECTL);
2206         u8 fmax, fmin, fstart, vstart;
2207
2208         spin_lock_irq(&mchdev_lock);
2209
2210         /* Enable temp reporting */
2211         I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
2212         I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
2213
2214         /* 100ms RC evaluation intervals */
2215         I915_WRITE(RCUPEI, 100000);
2216         I915_WRITE(RCDNEI, 100000);
2217
2218         /* Set max/min thresholds to 90ms and 80ms respectively */
2219         I915_WRITE(RCBMAXAVG, 90000);
2220         I915_WRITE(RCBMINAVG, 80000);
2221
2222         I915_WRITE(MEMIHYST, 1);
2223
2224         /* Set up min, max, and cur for interrupt handling */
2225         fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
2226         fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
2227         fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
2228                 MEMMODE_FSTART_SHIFT;
2229
2230         vstart = (I915_READ(PXVFREQ_BASE + (fstart * 4)) & PXVFREQ_PX_MASK) >>
2231                 PXVFREQ_PX_SHIFT;
2232
2233         dev_priv->fmax = fmax; /* IPS callback will increase this */
2234         dev_priv->fstart = fstart;
2235
2236         dev_priv->max_delay = fstart;
2237         dev_priv->min_delay = fmin;
2238         dev_priv->cur_delay = fstart;
2239
2240         DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
2241                          fmax, fmin, fstart);
2242
2243         I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
2244
2245         /*
2246          * Interrupts will be enabled in ironlake_irq_postinstall
2247          */
2248
2249         I915_WRITE(VIDSTART, vstart);
2250         POSTING_READ(VIDSTART);
2251
2252         rgvmodectl |= MEMMODE_SWMODE_EN;
2253         I915_WRITE(MEMMODECTL, rgvmodectl);
2254
2255         if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
2256                 DRM_ERROR("stuck trying to change perf mode\n");
2257         mdelay(1);
2258
2259         ironlake_set_drps(dev, fstart);
2260
2261         dev_priv->last_count1 = I915_READ(0x112e4) + I915_READ(0x112e8) +
2262                 I915_READ(0x112e0);
2263         dev_priv->last_time1 = jiffies_to_msecs(jiffies);
2264         dev_priv->last_count2 = I915_READ(0x112f4);
2265         getrawmonotonic(&dev_priv->last_time2);
2266
2267         spin_unlock_irq(&mchdev_lock);
2268 }
2269
2270 static void ironlake_disable_drps(struct drm_device *dev)
2271 {
2272         struct drm_i915_private *dev_priv = dev->dev_private;
2273         u16 rgvswctl;
2274
2275         spin_lock_irq(&mchdev_lock);
2276
2277         rgvswctl = I915_READ16(MEMSWCTL);
2278
2279         /* Ack interrupts, disable EFC interrupt */
2280         I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
2281         I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
2282         I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
2283         I915_WRITE(DEIIR, DE_PCU_EVENT);
2284         I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
2285
2286         /* Go back to the starting frequency */
2287         ironlake_set_drps(dev, dev_priv->fstart);
2288         mdelay(1);
2289         rgvswctl |= MEMCTL_CMD_STS;
2290         I915_WRITE(MEMSWCTL, rgvswctl);
2291         mdelay(1);
2292
2293         spin_unlock_irq(&mchdev_lock);
2294 }
2295
2296 /* There's a funny hw issue where the hw returns all 0 when reading from
2297  * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
2298  * ourselves, instead of doing a rmw cycle (which might result in us clearing
2299  * all limits and the gpu stuck at whatever frequency it is at atm).
2300  */
2301 static u32 gen6_rps_limits(struct drm_i915_private *dev_priv, u8 *val)
2302 {
2303         u32 limits;
2304
2305         limits = 0;
2306
2307         if (*val >= dev_priv->rps.max_delay)
2308                 *val = dev_priv->rps.max_delay;
2309         limits |= dev_priv->rps.max_delay << 24;
2310
2311         /* Only set the down limit when we've reached the lowest level to avoid
2312          * getting more interrupts, otherwise leave this clear. This prevents a
2313          * race in the hw when coming out of rc6: There's a tiny window where
2314          * the hw runs at the minimal clock before selecting the desired
2315          * frequency, if the down threshold expires in that window we will not
2316          * receive a down interrupt. */
2317         if (*val <= dev_priv->rps.min_delay) {
2318                 *val = dev_priv->rps.min_delay;
2319                 limits |= dev_priv->rps.min_delay << 16;
2320         }
2321
2322         return limits;
2323 }
2324
2325 void gen6_set_rps(struct drm_device *dev, u8 val)
2326 {
2327         struct drm_i915_private *dev_priv = dev->dev_private;
2328         u32 limits = gen6_rps_limits(dev_priv, &val);
2329
2330         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2331
2332         if (val == dev_priv->rps.cur_delay)
2333                 return;
2334
2335         I915_WRITE(GEN6_RPNSWREQ,
2336                    GEN6_FREQUENCY(val) |
2337                    GEN6_OFFSET(0) |
2338                    GEN6_AGGRESSIVE_TURBO);
2339
2340         /* Make sure we continue to get interrupts
2341          * until we hit the minimum or maximum frequencies.
2342          */
2343         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, limits);
2344
2345         dev_priv->rps.cur_delay = val;
2346 }
2347
2348 static void gen6_disable_rps(struct drm_device *dev)
2349 {
2350         struct drm_i915_private *dev_priv = dev->dev_private;
2351
2352         I915_WRITE(GEN6_RC_CONTROL, 0);
2353         I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
2354         I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
2355         I915_WRITE(GEN6_PMIER, 0);
2356         /* Complete PM interrupt masking here doesn't race with the rps work
2357          * item again unmasking PM interrupts because that is using a different
2358          * register (PMIMR) to mask PM interrupts. The only risk is in leaving
2359          * stale bits in PMIIR and PMIMR which gen6_enable_rps will clean up. */
2360
2361         spin_lock_irq(&dev_priv->rps.lock);
2362         dev_priv->rps.pm_iir = 0;
2363         spin_unlock_irq(&dev_priv->rps.lock);
2364
2365         I915_WRITE(GEN6_PMIIR, I915_READ(GEN6_PMIIR));
2366 }
2367
2368 int intel_enable_rc6(const struct drm_device *dev)
2369 {
2370         /* Respect the kernel parameter if it is set */
2371         if (i915_enable_rc6 >= 0)
2372                 return i915_enable_rc6;
2373
2374         if (INTEL_INFO(dev)->gen == 5) {
2375 #ifdef CONFIG_INTEL_IOMMU
2376                 /* Disable rc6 on ilk if VT-d is on. */
2377                 if (intel_iommu_gfx_mapped)
2378                         return false;
2379 #endif
2380                 DRM_DEBUG_DRIVER("Ironlake: only RC6 available\n");
2381                 return INTEL_RC6_ENABLE;
2382         }
2383
2384         if (IS_HASWELL(dev)) {
2385                 DRM_DEBUG_DRIVER("Haswell: only RC6 available\n");
2386                 return INTEL_RC6_ENABLE;
2387         }
2388
2389         /* snb/ivb have more than one rc6 state. */
2390         if (INTEL_INFO(dev)->gen == 6) {
2391                 DRM_DEBUG_DRIVER("Sandybridge: deep RC6 disabled\n");
2392                 return INTEL_RC6_ENABLE;
2393         }
2394
2395         DRM_DEBUG_DRIVER("RC6 and deep RC6 enabled\n");
2396         return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
2397 }
2398
2399 static void gen6_enable_rps(struct drm_device *dev)
2400 {
2401         struct drm_i915_private *dev_priv = dev->dev_private;
2402         struct intel_ring_buffer *ring;
2403         u32 rp_state_cap;
2404         u32 gt_perf_status;
2405         u32 pcu_mbox, rc6_mask = 0;
2406         u32 gtfifodbg;
2407         int rc6_mode;
2408         int i;
2409
2410         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2411
2412         /* Here begins a magic sequence of register writes to enable
2413          * auto-downclocking.
2414          *
2415          * Perhaps there might be some value in exposing these to
2416          * userspace...
2417          */
2418         I915_WRITE(GEN6_RC_STATE, 0);
2419
2420         /* Clear the DBG now so we don't confuse earlier errors */
2421         if ((gtfifodbg = I915_READ(GTFIFODBG))) {
2422                 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
2423                 I915_WRITE(GTFIFODBG, gtfifodbg);
2424         }
2425
2426         gen6_gt_force_wake_get(dev_priv);
2427
2428         rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
2429         gt_perf_status = I915_READ(GEN6_GT_PERF_STATUS);
2430
2431         /* In units of 100MHz */
2432         dev_priv->rps.max_delay = rp_state_cap & 0xff;
2433         dev_priv->rps.min_delay = (rp_state_cap & 0xff0000) >> 16;
2434         dev_priv->rps.cur_delay = 0;
2435
2436         /* disable the counters and set deterministic thresholds */
2437         I915_WRITE(GEN6_RC_CONTROL, 0);
2438
2439         I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
2440         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
2441         I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
2442         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
2443         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
2444
2445         for_each_ring(ring, dev_priv, i)
2446                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
2447
2448         I915_WRITE(GEN6_RC_SLEEP, 0);
2449         I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
2450         I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
2451         I915_WRITE(GEN6_RC6p_THRESHOLD, 100000);
2452         I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
2453
2454         /* Check if we are enabling RC6 */
2455         rc6_mode = intel_enable_rc6(dev_priv->dev);
2456         if (rc6_mode & INTEL_RC6_ENABLE)
2457                 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
2458
2459         /* We don't use those on Haswell */
2460         if (!IS_HASWELL(dev)) {
2461                 if (rc6_mode & INTEL_RC6p_ENABLE)
2462                         rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
2463
2464                 if (rc6_mode & INTEL_RC6pp_ENABLE)
2465                         rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
2466         }
2467
2468         DRM_INFO("Enabling RC6 states: RC6 %s, RC6p %s, RC6pp %s\n",
2469                         (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off",
2470                         (rc6_mask & GEN6_RC_CTL_RC6p_ENABLE) ? "on" : "off",
2471                         (rc6_mask & GEN6_RC_CTL_RC6pp_ENABLE) ? "on" : "off");
2472
2473         I915_WRITE(GEN6_RC_CONTROL,
2474                    rc6_mask |
2475                    GEN6_RC_CTL_EI_MODE(1) |
2476                    GEN6_RC_CTL_HW_ENABLE);
2477
2478         I915_WRITE(GEN6_RPNSWREQ,
2479                    GEN6_FREQUENCY(10) |
2480                    GEN6_OFFSET(0) |
2481                    GEN6_AGGRESSIVE_TURBO);
2482         I915_WRITE(GEN6_RC_VIDEO_FREQ,
2483                    GEN6_FREQUENCY(12));
2484
2485         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
2486         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
2487                    dev_priv->rps.max_delay << 24 |
2488                    dev_priv->rps.min_delay << 16);
2489
2490         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
2491         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
2492         I915_WRITE(GEN6_RP_UP_EI, 66000);
2493         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
2494
2495         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
2496         I915_WRITE(GEN6_RP_CONTROL,
2497                    GEN6_RP_MEDIA_TURBO |
2498                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
2499                    GEN6_RP_MEDIA_IS_GFX |
2500                    GEN6_RP_ENABLE |
2501                    GEN6_RP_UP_BUSY_AVG |
2502                    (IS_HASWELL(dev) ? GEN7_RP_DOWN_IDLE_AVG : GEN6_RP_DOWN_IDLE_CONT));
2503
2504         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
2505                      500))
2506                 DRM_ERROR("timeout waiting for pcode mailbox to become idle\n");
2507
2508         I915_WRITE(GEN6_PCODE_DATA, 0);
2509         I915_WRITE(GEN6_PCODE_MAILBOX,
2510                    GEN6_PCODE_READY |
2511                    GEN6_PCODE_WRITE_MIN_FREQ_TABLE);
2512         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
2513                      500))
2514                 DRM_ERROR("timeout waiting for pcode mailbox to finish\n");
2515
2516         /* Check for overclock support */
2517         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
2518                      500))
2519                 DRM_ERROR("timeout waiting for pcode mailbox to become idle\n");
2520         I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_READ_OC_PARAMS);
2521         pcu_mbox = I915_READ(GEN6_PCODE_DATA);
2522         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
2523                      500))
2524                 DRM_ERROR("timeout waiting for pcode mailbox to finish\n");
2525         if (pcu_mbox & (1<<31)) { /* OC supported */
2526                 dev_priv->rps.max_delay = pcu_mbox & 0xff;
2527                 DRM_DEBUG_DRIVER("overclocking supported, adjusting frequency max to %dMHz\n", pcu_mbox * 50);
2528         }
2529
2530         gen6_set_rps(dev_priv->dev, (gt_perf_status & 0xff00) >> 8);
2531
2532         /* requires MSI enabled */
2533         I915_WRITE(GEN6_PMIER, GEN6_PM_DEFERRED_EVENTS);
2534         spin_lock_irq(&dev_priv->rps.lock);
2535         WARN_ON(dev_priv->rps.pm_iir != 0);
2536         I915_WRITE(GEN6_PMIMR, 0);
2537         spin_unlock_irq(&dev_priv->rps.lock);
2538         /* enable all PM interrupts */
2539         I915_WRITE(GEN6_PMINTRMSK, 0);
2540
2541         gen6_gt_force_wake_put(dev_priv);
2542 }
2543
2544 static void gen6_update_ring_freq(struct drm_device *dev)
2545 {
2546         struct drm_i915_private *dev_priv = dev->dev_private;
2547         int min_freq = 15;
2548         int gpu_freq, ia_freq, max_ia_freq;
2549         int scaling_factor = 180;
2550
2551         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2552
2553         max_ia_freq = cpufreq_quick_get_max(0);
2554         /*
2555          * Default to measured freq if none found, PCU will ensure we don't go
2556          * over
2557          */
2558         if (!max_ia_freq)
2559                 max_ia_freq = tsc_khz;
2560
2561         /* Convert from kHz to MHz */
2562         max_ia_freq /= 1000;
2563
2564         /*
2565          * For each potential GPU frequency, load a ring frequency we'd like
2566          * to use for memory access.  We do this by specifying the IA frequency
2567          * the PCU should use as a reference to determine the ring frequency.
2568          */
2569         for (gpu_freq = dev_priv->rps.max_delay; gpu_freq >= dev_priv->rps.min_delay;
2570              gpu_freq--) {
2571                 int diff = dev_priv->rps.max_delay - gpu_freq;
2572
2573                 /*
2574                  * For GPU frequencies less than 750MHz, just use the lowest
2575                  * ring freq.
2576                  */
2577                 if (gpu_freq < min_freq)
2578                         ia_freq = 800;
2579                 else
2580                         ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
2581                 ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
2582
2583                 I915_WRITE(GEN6_PCODE_DATA,
2584                            (ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT) |
2585                            gpu_freq);
2586                 I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY |
2587                            GEN6_PCODE_WRITE_MIN_FREQ_TABLE);
2588                 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) &
2589                               GEN6_PCODE_READY) == 0, 10)) {
2590                         DRM_ERROR("pcode write of freq table timed out\n");
2591                         continue;
2592                 }
2593         }
2594 }
2595
2596 void ironlake_teardown_rc6(struct drm_device *dev)
2597 {
2598         struct drm_i915_private *dev_priv = dev->dev_private;
2599
2600         if (dev_priv->renderctx) {
2601                 i915_gem_object_unpin(dev_priv->renderctx);
2602                 drm_gem_object_unreference(&dev_priv->renderctx->base);
2603                 dev_priv->renderctx = NULL;
2604         }
2605
2606         if (dev_priv->pwrctx) {
2607                 i915_gem_object_unpin(dev_priv->pwrctx);
2608                 drm_gem_object_unreference(&dev_priv->pwrctx->base);
2609                 dev_priv->pwrctx = NULL;
2610         }
2611 }
2612
2613 static void ironlake_disable_rc6(struct drm_device *dev)
2614 {
2615         struct drm_i915_private *dev_priv = dev->dev_private;
2616
2617         if (I915_READ(PWRCTXA)) {
2618                 /* Wake the GPU, prevent RC6, then restore RSTDBYCTL */
2619                 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) | RCX_SW_EXIT);
2620                 wait_for(((I915_READ(RSTDBYCTL) & RSX_STATUS_MASK) == RSX_STATUS_ON),
2621                          50);
2622
2623                 I915_WRITE(PWRCTXA, 0);
2624                 POSTING_READ(PWRCTXA);
2625
2626                 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
2627                 POSTING_READ(RSTDBYCTL);
2628         }
2629 }
2630
2631 static int ironlake_setup_rc6(struct drm_device *dev)
2632 {
2633         struct drm_i915_private *dev_priv = dev->dev_private;
2634
2635         if (dev_priv->renderctx == NULL)
2636                 dev_priv->renderctx = intel_alloc_context_page(dev);
2637         if (!dev_priv->renderctx)
2638                 return -ENOMEM;
2639
2640         if (dev_priv->pwrctx == NULL)
2641                 dev_priv->pwrctx = intel_alloc_context_page(dev);
2642         if (!dev_priv->pwrctx) {
2643                 ironlake_teardown_rc6(dev);
2644                 return -ENOMEM;
2645         }
2646
2647         return 0;
2648 }
2649
2650 static void ironlake_enable_rc6(struct drm_device *dev)
2651 {
2652         struct drm_i915_private *dev_priv = dev->dev_private;
2653         struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
2654         int ret;
2655
2656         /* rc6 disabled by default due to repeated reports of hanging during
2657          * boot and resume.
2658          */
2659         if (!intel_enable_rc6(dev))
2660                 return;
2661
2662         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2663
2664         ret = ironlake_setup_rc6(dev);
2665         if (ret)
2666                 return;
2667
2668         /*
2669          * GPU can automatically power down the render unit if given a page
2670          * to save state.
2671          */
2672         ret = intel_ring_begin(ring, 6);
2673         if (ret) {
2674                 ironlake_teardown_rc6(dev);
2675                 return;
2676         }
2677
2678         intel_ring_emit(ring, MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN);
2679         intel_ring_emit(ring, MI_SET_CONTEXT);
2680         intel_ring_emit(ring, dev_priv->renderctx->gtt_offset |
2681                         MI_MM_SPACE_GTT |
2682                         MI_SAVE_EXT_STATE_EN |
2683                         MI_RESTORE_EXT_STATE_EN |
2684                         MI_RESTORE_INHIBIT);
2685         intel_ring_emit(ring, MI_SUSPEND_FLUSH);
2686         intel_ring_emit(ring, MI_NOOP);
2687         intel_ring_emit(ring, MI_FLUSH);
2688         intel_ring_advance(ring);
2689
2690         /*
2691          * Wait for the command parser to advance past MI_SET_CONTEXT. The HW
2692          * does an implicit flush, combined with MI_FLUSH above, it should be
2693          * safe to assume that renderctx is valid
2694          */
2695         ret = intel_wait_ring_idle(ring);
2696         if (ret) {
2697                 DRM_ERROR("failed to enable ironlake power power savings\n");
2698                 ironlake_teardown_rc6(dev);
2699                 return;
2700         }
2701
2702         I915_WRITE(PWRCTXA, dev_priv->pwrctx->gtt_offset | PWRCTX_EN);
2703         I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
2704 }
2705
2706 static unsigned long intel_pxfreq(u32 vidfreq)
2707 {
2708         unsigned long freq;
2709         int div = (vidfreq & 0x3f0000) >> 16;
2710         int post = (vidfreq & 0x3000) >> 12;
2711         int pre = (vidfreq & 0x7);
2712
2713         if (!pre)
2714                 return 0;
2715
2716         freq = ((div * 133333) / ((1<<post) * pre));
2717
2718         return freq;
2719 }
2720
2721 static const struct cparams {
2722         u16 i;
2723         u16 t;
2724         u16 m;
2725         u16 c;
2726 } cparams[] = {
2727         { 1, 1333, 301, 28664 },
2728         { 1, 1066, 294, 24460 },
2729         { 1, 800, 294, 25192 },
2730         { 0, 1333, 276, 27605 },
2731         { 0, 1066, 276, 27605 },
2732         { 0, 800, 231, 23784 },
2733 };
2734
2735 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
2736 {
2737         u64 total_count, diff, ret;
2738         u32 count1, count2, count3, m = 0, c = 0;
2739         unsigned long now = jiffies_to_msecs(jiffies), diff1;
2740         int i;
2741
2742         assert_spin_locked(&mchdev_lock);
2743
2744         diff1 = now - dev_priv->last_time1;
2745
2746         /* Prevent division-by-zero if we are asking too fast.
2747          * Also, we don't get interesting results if we are polling
2748          * faster than once in 10ms, so just return the saved value
2749          * in such cases.
2750          */
2751         if (diff1 <= 10)
2752                 return dev_priv->chipset_power;
2753
2754         count1 = I915_READ(DMIEC);
2755         count2 = I915_READ(DDREC);
2756         count3 = I915_READ(CSIEC);
2757
2758         total_count = count1 + count2 + count3;
2759
2760         /* FIXME: handle per-counter overflow */
2761         if (total_count < dev_priv->last_count1) {
2762                 diff = ~0UL - dev_priv->last_count1;
2763                 diff += total_count;
2764         } else {
2765                 diff = total_count - dev_priv->last_count1;
2766         }
2767
2768         for (i = 0; i < ARRAY_SIZE(cparams); i++) {
2769                 if (cparams[i].i == dev_priv->c_m &&
2770                     cparams[i].t == dev_priv->r_t) {
2771                         m = cparams[i].m;
2772                         c = cparams[i].c;
2773                         break;
2774                 }
2775         }
2776
2777         diff = div_u64(diff, diff1);
2778         ret = ((m * diff) + c);
2779         ret = div_u64(ret, 10);
2780
2781         dev_priv->last_count1 = total_count;
2782         dev_priv->last_time1 = now;
2783
2784         dev_priv->chipset_power = ret;
2785
2786         return ret;
2787 }
2788
2789 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
2790 {
2791         unsigned long m, x, b;
2792         u32 tsfs;
2793
2794         tsfs = I915_READ(TSFS);
2795
2796         m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
2797         x = I915_READ8(TR1);
2798
2799         b = tsfs & TSFS_INTR_MASK;
2800
2801         return ((m * x) / 127) - b;
2802 }
2803
2804 static u16 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
2805 {
2806         static const struct v_table {
2807                 u16 vd; /* in .1 mil */
2808                 u16 vm; /* in .1 mil */
2809         } v_table[] = {
2810                 { 0, 0, },
2811                 { 375, 0, },
2812                 { 500, 0, },
2813                 { 625, 0, },
2814                 { 750, 0, },
2815                 { 875, 0, },
2816                 { 1000, 0, },
2817                 { 1125, 0, },
2818                 { 4125, 3000, },
2819                 { 4125, 3000, },
2820                 { 4125, 3000, },
2821                 { 4125, 3000, },
2822                 { 4125, 3000, },
2823                 { 4125, 3000, },
2824                 { 4125, 3000, },
2825                 { 4125, 3000, },
2826                 { 4125, 3000, },
2827                 { 4125, 3000, },
2828                 { 4125, 3000, },
2829                 { 4125, 3000, },
2830                 { 4125, 3000, },
2831                 { 4125, 3000, },
2832                 { 4125, 3000, },
2833                 { 4125, 3000, },
2834                 { 4125, 3000, },
2835                 { 4125, 3000, },
2836                 { 4125, 3000, },
2837                 { 4125, 3000, },
2838                 { 4125, 3000, },
2839                 { 4125, 3000, },
2840                 { 4125, 3000, },
2841                 { 4125, 3000, },
2842                 { 4250, 3125, },
2843                 { 4375, 3250, },
2844                 { 4500, 3375, },
2845                 { 4625, 3500, },
2846                 { 4750, 3625, },
2847                 { 4875, 3750, },
2848                 { 5000, 3875, },
2849                 { 5125, 4000, },
2850                 { 5250, 4125, },
2851                 { 5375, 4250, },
2852                 { 5500, 4375, },
2853                 { 5625, 4500, },
2854                 { 5750, 4625, },
2855                 { 5875, 4750, },
2856                 { 6000, 4875, },
2857                 { 6125, 5000, },
2858                 { 6250, 5125, },
2859                 { 6375, 5250, },
2860                 { 6500, 5375, },
2861                 { 6625, 5500, },
2862                 { 6750, 5625, },
2863                 { 6875, 5750, },
2864                 { 7000, 5875, },
2865                 { 7125, 6000, },
2866                 { 7250, 6125, },
2867                 { 7375, 6250, },
2868                 { 7500, 6375, },
2869                 { 7625, 6500, },
2870                 { 7750, 6625, },
2871                 { 7875, 6750, },
2872                 { 8000, 6875, },
2873                 { 8125, 7000, },
2874                 { 8250, 7125, },
2875                 { 8375, 7250, },
2876                 { 8500, 7375, },
2877                 { 8625, 7500, },
2878                 { 8750, 7625, },
2879                 { 8875, 7750, },
2880                 { 9000, 7875, },
2881                 { 9125, 8000, },
2882                 { 9250, 8125, },
2883                 { 9375, 8250, },
2884                 { 9500, 8375, },
2885                 { 9625, 8500, },
2886                 { 9750, 8625, },
2887                 { 9875, 8750, },
2888                 { 10000, 8875, },
2889                 { 10125, 9000, },
2890                 { 10250, 9125, },
2891                 { 10375, 9250, },
2892                 { 10500, 9375, },
2893                 { 10625, 9500, },
2894                 { 10750, 9625, },
2895                 { 10875, 9750, },
2896                 { 11000, 9875, },
2897                 { 11125, 10000, },
2898                 { 11250, 10125, },
2899                 { 11375, 10250, },
2900                 { 11500, 10375, },
2901                 { 11625, 10500, },
2902                 { 11750, 10625, },
2903                 { 11875, 10750, },
2904                 { 12000, 10875, },
2905                 { 12125, 11000, },
2906                 { 12250, 11125, },
2907                 { 12375, 11250, },
2908                 { 12500, 11375, },
2909                 { 12625, 11500, },
2910                 { 12750, 11625, },
2911                 { 12875, 11750, },
2912                 { 13000, 11875, },
2913                 { 13125, 12000, },
2914                 { 13250, 12125, },
2915                 { 13375, 12250, },
2916                 { 13500, 12375, },
2917                 { 13625, 12500, },
2918                 { 13750, 12625, },
2919                 { 13875, 12750, },
2920                 { 14000, 12875, },
2921                 { 14125, 13000, },
2922                 { 14250, 13125, },
2923                 { 14375, 13250, },
2924                 { 14500, 13375, },
2925                 { 14625, 13500, },
2926                 { 14750, 13625, },
2927                 { 14875, 13750, },
2928                 { 15000, 13875, },
2929                 { 15125, 14000, },
2930                 { 15250, 14125, },
2931                 { 15375, 14250, },
2932                 { 15500, 14375, },
2933                 { 15625, 14500, },
2934                 { 15750, 14625, },
2935                 { 15875, 14750, },
2936                 { 16000, 14875, },
2937                 { 16125, 15000, },
2938         };
2939         if (dev_priv->info->is_mobile)
2940                 return v_table[pxvid].vm;
2941         else
2942                 return v_table[pxvid].vd;
2943 }
2944
2945 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
2946 {
2947         struct timespec now, diff1;
2948         u64 diff;
2949         unsigned long diffms;
2950         u32 count;
2951
2952         assert_spin_locked(&mchdev_lock);
2953
2954         getrawmonotonic(&now);
2955         diff1 = timespec_sub(now, dev_priv->last_time2);
2956
2957         /* Don't divide by 0 */
2958         diffms = diff1.tv_sec * 1000 + diff1.tv_nsec / 1000000;
2959         if (!diffms)
2960                 return;
2961
2962         count = I915_READ(GFXEC);
2963
2964         if (count < dev_priv->last_count2) {
2965                 diff = ~0UL - dev_priv->last_count2;
2966                 diff += count;
2967         } else {
2968                 diff = count - dev_priv->last_count2;
2969         }
2970
2971         dev_priv->last_count2 = count;
2972         dev_priv->last_time2 = now;
2973
2974         /* More magic constants... */
2975         diff = diff * 1181;
2976         diff = div_u64(diff, diffms * 10);
2977         dev_priv->gfx_power = diff;
2978 }
2979
2980 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
2981 {
2982         if (dev_priv->info->gen != 5)
2983                 return;
2984
2985         spin_lock_irq(&mchdev_lock);
2986
2987         __i915_update_gfx_val(dev_priv);
2988
2989         spin_unlock_irq(&mchdev_lock);
2990 }
2991
2992 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
2993 {
2994         unsigned long t, corr, state1, corr2, state2;
2995         u32 pxvid, ext_v;
2996
2997         assert_spin_locked(&mchdev_lock);
2998
2999         pxvid = I915_READ(PXVFREQ_BASE + (dev_priv->rps.cur_delay * 4));
3000         pxvid = (pxvid >> 24) & 0x7f;
3001         ext_v = pvid_to_extvid(dev_priv, pxvid);
3002
3003         state1 = ext_v;
3004
3005         t = i915_mch_val(dev_priv);
3006
3007         /* Revel in the empirically derived constants */
3008
3009         /* Correction factor in 1/100000 units */
3010         if (t > 80)
3011                 corr = ((t * 2349) + 135940);
3012         else if (t >= 50)
3013                 corr = ((t * 964) + 29317);
3014         else /* < 50 */
3015                 corr = ((t * 301) + 1004);
3016
3017         corr = corr * ((150142 * state1) / 10000 - 78642);
3018         corr /= 100000;
3019         corr2 = (corr * dev_priv->corr);
3020
3021         state2 = (corr2 * state1) / 10000;
3022         state2 /= 100; /* convert to mW */
3023
3024         __i915_update_gfx_val(dev_priv);
3025
3026         return dev_priv->gfx_power + state2;
3027 }
3028
3029 /**
3030  * i915_read_mch_val - return value for IPS use
3031  *
3032  * Calculate and return a value for the IPS driver to use when deciding whether
3033  * we have thermal and power headroom to increase CPU or GPU power budget.
3034  */
3035 unsigned long i915_read_mch_val(void)
3036 {
3037         struct drm_i915_private *dev_priv;
3038         unsigned long chipset_val, graphics_val, ret = 0;
3039
3040         spin_lock_irq(&mchdev_lock);
3041         if (!i915_mch_dev)
3042                 goto out_unlock;
3043         dev_priv = i915_mch_dev;
3044
3045         chipset_val = i915_chipset_val(dev_priv);
3046         graphics_val = i915_gfx_val(dev_priv);
3047
3048         ret = chipset_val + graphics_val;
3049
3050 out_unlock:
3051         spin_unlock_irq(&mchdev_lock);
3052
3053         return ret;
3054 }
3055 EXPORT_SYMBOL_GPL(i915_read_mch_val);
3056
3057 /**
3058  * i915_gpu_raise - raise GPU frequency limit
3059  *
3060  * Raise the limit; IPS indicates we have thermal headroom.
3061  */
3062 bool i915_gpu_raise(void)
3063 {
3064         struct drm_i915_private *dev_priv;
3065         bool ret = true;
3066
3067         spin_lock_irq(&mchdev_lock);
3068         if (!i915_mch_dev) {
3069                 ret = false;
3070                 goto out_unlock;
3071         }
3072         dev_priv = i915_mch_dev;
3073
3074         if (dev_priv->max_delay > dev_priv->fmax)
3075                 dev_priv->max_delay--;
3076
3077 out_unlock:
3078         spin_unlock_irq(&mchdev_lock);
3079
3080         return ret;
3081 }
3082 EXPORT_SYMBOL_GPL(i915_gpu_raise);
3083
3084 /**
3085  * i915_gpu_lower - lower GPU frequency limit
3086  *
3087  * IPS indicates we're close to a thermal limit, so throttle back the GPU
3088  * frequency maximum.
3089  */
3090 bool i915_gpu_lower(void)
3091 {
3092         struct drm_i915_private *dev_priv;
3093         bool ret = true;
3094
3095         spin_lock_irq(&mchdev_lock);
3096         if (!i915_mch_dev) {
3097                 ret = false;
3098                 goto out_unlock;
3099         }
3100         dev_priv = i915_mch_dev;
3101
3102         if (dev_priv->max_delay < dev_priv->min_delay)
3103                 dev_priv->max_delay++;
3104
3105 out_unlock:
3106         spin_unlock_irq(&mchdev_lock);
3107
3108         return ret;
3109 }
3110 EXPORT_SYMBOL_GPL(i915_gpu_lower);
3111
3112 /**
3113  * i915_gpu_busy - indicate GPU business to IPS
3114  *
3115  * Tell the IPS driver whether or not the GPU is busy.
3116  */
3117 bool i915_gpu_busy(void)
3118 {
3119         struct drm_i915_private *dev_priv;
3120         struct intel_ring_buffer *ring;
3121         bool ret = false;
3122         int i;
3123
3124         spin_lock_irq(&mchdev_lock);
3125         if (!i915_mch_dev)
3126                 goto out_unlock;
3127         dev_priv = i915_mch_dev;
3128
3129         for_each_ring(ring, dev_priv, i)
3130                 ret |= !list_empty(&ring->request_list);
3131
3132 out_unlock:
3133         spin_unlock_irq(&mchdev_lock);
3134
3135         return ret;
3136 }
3137 EXPORT_SYMBOL_GPL(i915_gpu_busy);
3138
3139 /**
3140  * i915_gpu_turbo_disable - disable graphics turbo
3141  *
3142  * Disable graphics turbo by resetting the max frequency and setting the
3143  * current frequency to the default.
3144  */
3145 bool i915_gpu_turbo_disable(void)
3146 {
3147         struct drm_i915_private *dev_priv;
3148         bool ret = true;
3149
3150         spin_lock_irq(&mchdev_lock);
3151         if (!i915_mch_dev) {
3152                 ret = false;
3153                 goto out_unlock;
3154         }
3155         dev_priv = i915_mch_dev;
3156
3157         dev_priv->max_delay = dev_priv->fstart;
3158
3159         if (!ironlake_set_drps(dev_priv->dev, dev_priv->fstart))
3160                 ret = false;
3161
3162 out_unlock:
3163         spin_unlock_irq(&mchdev_lock);
3164
3165         return ret;
3166 }
3167 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
3168
3169 /**
3170  * Tells the intel_ips driver that the i915 driver is now loaded, if
3171  * IPS got loaded first.
3172  *
3173  * This awkward dance is so that neither module has to depend on the
3174  * other in order for IPS to do the appropriate communication of
3175  * GPU turbo limits to i915.
3176  */
3177 static void
3178 ips_ping_for_i915_load(void)
3179 {
3180         void (*link)(void);
3181
3182         link = symbol_get(ips_link_to_i915_driver);
3183         if (link) {
3184                 link();
3185                 symbol_put(ips_link_to_i915_driver);
3186         }
3187 }
3188
3189 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
3190 {
3191         /* We only register the i915 ips part with intel-ips once everything is
3192          * set up, to avoid intel-ips sneaking in and reading bogus values. */
3193         spin_lock_irq(&mchdev_lock);
3194         i915_mch_dev = dev_priv;
3195         spin_unlock_irq(&mchdev_lock);
3196
3197         ips_ping_for_i915_load();
3198 }
3199
3200 void intel_gpu_ips_teardown(void)
3201 {
3202         spin_lock_irq(&mchdev_lock);
3203         i915_mch_dev = NULL;
3204         spin_unlock_irq(&mchdev_lock);
3205 }
3206 static void intel_init_emon(struct drm_device *dev)
3207 {
3208         struct drm_i915_private *dev_priv = dev->dev_private;
3209         u32 lcfuse;
3210         u8 pxw[16];
3211         int i;
3212
3213         /* Disable to program */
3214         I915_WRITE(ECR, 0);
3215         POSTING_READ(ECR);
3216
3217         /* Program energy weights for various events */
3218         I915_WRITE(SDEW, 0x15040d00);
3219         I915_WRITE(CSIEW0, 0x007f0000);
3220         I915_WRITE(CSIEW1, 0x1e220004);
3221         I915_WRITE(CSIEW2, 0x04000004);
3222
3223         for (i = 0; i < 5; i++)
3224                 I915_WRITE(PEW + (i * 4), 0);
3225         for (i = 0; i < 3; i++)
3226                 I915_WRITE(DEW + (i * 4), 0);
3227
3228         /* Program P-state weights to account for frequency power adjustment */
3229         for (i = 0; i < 16; i++) {
3230                 u32 pxvidfreq = I915_READ(PXVFREQ_BASE + (i * 4));
3231                 unsigned long freq = intel_pxfreq(pxvidfreq);
3232                 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
3233                         PXVFREQ_PX_SHIFT;
3234                 unsigned long val;
3235
3236                 val = vid * vid;
3237                 val *= (freq / 1000);
3238                 val *= 255;
3239                 val /= (127*127*900);
3240                 if (val > 0xff)
3241                         DRM_ERROR("bad pxval: %ld\n", val);
3242                 pxw[i] = val;
3243         }
3244         /* Render standby states get 0 weight */
3245         pxw[14] = 0;
3246         pxw[15] = 0;
3247
3248         for (i = 0; i < 4; i++) {
3249                 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
3250                         (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
3251                 I915_WRITE(PXW + (i * 4), val);
3252         }
3253
3254         /* Adjust magic regs to magic values (more experimental results) */
3255         I915_WRITE(OGW0, 0);
3256         I915_WRITE(OGW1, 0);
3257         I915_WRITE(EG0, 0x00007f00);
3258         I915_WRITE(EG1, 0x0000000e);
3259         I915_WRITE(EG2, 0x000e0000);
3260         I915_WRITE(EG3, 0x68000300);
3261         I915_WRITE(EG4, 0x42000000);
3262         I915_WRITE(EG5, 0x00140031);
3263         I915_WRITE(EG6, 0);
3264         I915_WRITE(EG7, 0);
3265
3266         for (i = 0; i < 8; i++)
3267                 I915_WRITE(PXWL + (i * 4), 0);
3268
3269         /* Enable PMON + select events */
3270         I915_WRITE(ECR, 0x80000019);
3271
3272         lcfuse = I915_READ(LCFUSE02);
3273
3274         dev_priv->corr = (lcfuse & LCFUSE_HIV_MASK);
3275 }
3276
3277 void intel_disable_gt_powersave(struct drm_device *dev)
3278 {
3279         if (IS_IRONLAKE_M(dev)) {
3280                 ironlake_disable_drps(dev);
3281                 ironlake_disable_rc6(dev);
3282         } else if (INTEL_INFO(dev)->gen >= 6 && !IS_VALLEYVIEW(dev)) {
3283                 gen6_disable_rps(dev);
3284         }
3285 }
3286
3287 void intel_enable_gt_powersave(struct drm_device *dev)
3288 {
3289         if (IS_IRONLAKE_M(dev)) {
3290                 ironlake_enable_drps(dev);
3291                 ironlake_enable_rc6(dev);
3292                 intel_init_emon(dev);
3293         } else if ((IS_GEN6(dev) || IS_GEN7(dev)) && !IS_VALLEYVIEW(dev)) {
3294                 gen6_enable_rps(dev);
3295                 gen6_update_ring_freq(dev);
3296         }
3297 }
3298
3299 static void ironlake_init_clock_gating(struct drm_device *dev)
3300 {
3301         struct drm_i915_private *dev_priv = dev->dev_private;
3302         uint32_t dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE;
3303
3304         /* Required for FBC */
3305         dspclk_gate |= DPFCUNIT_CLOCK_GATE_DISABLE |
3306                 DPFCRUNIT_CLOCK_GATE_DISABLE |
3307                 DPFDUNIT_CLOCK_GATE_DISABLE;
3308         /* Required for CxSR */
3309         dspclk_gate |= DPARBUNIT_CLOCK_GATE_DISABLE;
3310
3311         I915_WRITE(PCH_3DCGDIS0,
3312                    MARIUNIT_CLOCK_GATE_DISABLE |
3313                    SVSMUNIT_CLOCK_GATE_DISABLE);
3314         I915_WRITE(PCH_3DCGDIS1,
3315                    VFMUNIT_CLOCK_GATE_DISABLE);
3316
3317         I915_WRITE(PCH_DSPCLK_GATE_D, dspclk_gate);
3318
3319         /*
3320          * According to the spec the following bits should be set in
3321          * order to enable memory self-refresh
3322          * The bit 22/21 of 0x42004
3323          * The bit 5 of 0x42020
3324          * The bit 15 of 0x45000
3325          */
3326         I915_WRITE(ILK_DISPLAY_CHICKEN2,
3327                    (I915_READ(ILK_DISPLAY_CHICKEN2) |
3328                     ILK_DPARB_GATE | ILK_VSDPFD_FULL));
3329         I915_WRITE(ILK_DSPCLK_GATE,
3330                    (I915_READ(ILK_DSPCLK_GATE) |
3331                     ILK_DPARB_CLK_GATE));
3332         I915_WRITE(DISP_ARB_CTL,
3333                    (I915_READ(DISP_ARB_CTL) |
3334                     DISP_FBC_WM_DIS));
3335         I915_WRITE(WM3_LP_ILK, 0);
3336         I915_WRITE(WM2_LP_ILK, 0);
3337         I915_WRITE(WM1_LP_ILK, 0);
3338
3339         /*
3340          * Based on the document from hardware guys the following bits
3341          * should be set unconditionally in order to enable FBC.
3342          * The bit 22 of 0x42000
3343          * The bit 22 of 0x42004
3344          * The bit 7,8,9 of 0x42020.
3345          */
3346         if (IS_IRONLAKE_M(dev)) {
3347                 I915_WRITE(ILK_DISPLAY_CHICKEN1,
3348                            I915_READ(ILK_DISPLAY_CHICKEN1) |
3349                            ILK_FBCQ_DIS);
3350                 I915_WRITE(ILK_DISPLAY_CHICKEN2,
3351                            I915_READ(ILK_DISPLAY_CHICKEN2) |
3352                            ILK_DPARB_GATE);
3353                 I915_WRITE(ILK_DSPCLK_GATE,
3354                            I915_READ(ILK_DSPCLK_GATE) |
3355                            ILK_DPFC_DIS1 |
3356                            ILK_DPFC_DIS2 |
3357                            ILK_CLK_FBC);
3358         }
3359
3360         I915_WRITE(ILK_DISPLAY_CHICKEN2,
3361                    I915_READ(ILK_DISPLAY_CHICKEN2) |
3362                    ILK_ELPIN_409_SELECT);
3363         I915_WRITE(_3D_CHICKEN2,
3364                    _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
3365                    _3D_CHICKEN2_WM_READ_PIPELINED);
3366 }
3367
3368 static void gen6_init_clock_gating(struct drm_device *dev)
3369 {
3370         struct drm_i915_private *dev_priv = dev->dev_private;
3371         int pipe;
3372         uint32_t dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE;
3373
3374         I915_WRITE(PCH_DSPCLK_GATE_D, dspclk_gate);
3375
3376         I915_WRITE(ILK_DISPLAY_CHICKEN2,
3377                    I915_READ(ILK_DISPLAY_CHICKEN2) |
3378                    ILK_ELPIN_409_SELECT);
3379
3380         I915_WRITE(WM3_LP_ILK, 0);
3381         I915_WRITE(WM2_LP_ILK, 0);
3382         I915_WRITE(WM1_LP_ILK, 0);
3383
3384         I915_WRITE(CACHE_MODE_0,
3385                    _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
3386
3387         I915_WRITE(GEN6_UCGCTL1,
3388                    I915_READ(GEN6_UCGCTL1) |
3389                    GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
3390                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
3391
3392         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
3393          * gating disable must be set.  Failure to set it results in
3394          * flickering pixels due to Z write ordering failures after
3395          * some amount of runtime in the Mesa "fire" demo, and Unigine
3396          * Sanctuary and Tropics, and apparently anything else with
3397          * alpha test or pixel discard.
3398          *
3399          * According to the spec, bit 11 (RCCUNIT) must also be set,
3400          * but we didn't debug actual testcases to find it out.
3401          *
3402          * Also apply WaDisableVDSUnitClockGating and
3403          * WaDisableRCPBUnitClockGating.
3404          */
3405         I915_WRITE(GEN6_UCGCTL2,
3406                    GEN7_VDSUNIT_CLOCK_GATE_DISABLE |
3407                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
3408                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
3409
3410         /* Bspec says we need to always set all mask bits. */
3411         I915_WRITE(_3D_CHICKEN, (0xFFFF << 16) |
3412                    _3D_CHICKEN_SF_DISABLE_FASTCLIP_CULL);
3413
3414         /*
3415          * According to the spec the following bits should be
3416          * set in order to enable memory self-refresh and fbc:
3417          * The bit21 and bit22 of 0x42000
3418          * The bit21 and bit22 of 0x42004
3419          * The bit5 and bit7 of 0x42020
3420          * The bit14 of 0x70180
3421          * The bit14 of 0x71180
3422          */
3423         I915_WRITE(ILK_DISPLAY_CHICKEN1,
3424                    I915_READ(ILK_DISPLAY_CHICKEN1) |
3425                    ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
3426         I915_WRITE(ILK_DISPLAY_CHICKEN2,
3427                    I915_READ(ILK_DISPLAY_CHICKEN2) |
3428                    ILK_DPARB_GATE | ILK_VSDPFD_FULL);
3429         I915_WRITE(ILK_DSPCLK_GATE,
3430                    I915_READ(ILK_DSPCLK_GATE) |
3431                    ILK_DPARB_CLK_GATE  |
3432                    ILK_DPFD_CLK_GATE);
3433
3434         I915_WRITE(GEN6_MBCTL, I915_READ(GEN6_MBCTL) |
3435                    GEN6_MBCTL_ENABLE_BOOT_FETCH);
3436
3437         for_each_pipe(pipe) {
3438                 I915_WRITE(DSPCNTR(pipe),
3439                            I915_READ(DSPCNTR(pipe)) |
3440                            DISPPLANE_TRICKLE_FEED_DISABLE);
3441                 intel_flush_display_plane(dev_priv, pipe);
3442         }
3443 }
3444
3445 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
3446 {
3447         uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
3448
3449         reg &= ~GEN7_FF_SCHED_MASK;
3450         reg |= GEN7_FF_TS_SCHED_HW;
3451         reg |= GEN7_FF_VS_SCHED_HW;
3452         reg |= GEN7_FF_DS_SCHED_HW;
3453
3454         I915_WRITE(GEN7_FF_THREAD_MODE, reg);
3455 }
3456
3457 static void haswell_init_clock_gating(struct drm_device *dev)
3458 {
3459         struct drm_i915_private *dev_priv = dev->dev_private;
3460         int pipe;
3461         uint32_t dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE;
3462
3463         I915_WRITE(PCH_DSPCLK_GATE_D, dspclk_gate);
3464
3465         I915_WRITE(WM3_LP_ILK, 0);
3466         I915_WRITE(WM2_LP_ILK, 0);
3467         I915_WRITE(WM1_LP_ILK, 0);
3468
3469         /* According to the spec, bit 13 (RCZUNIT) must be set on IVB.
3470          * This implements the WaDisableRCZUnitClockGating workaround.
3471          */
3472         I915_WRITE(GEN6_UCGCTL2, GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
3473
3474         I915_WRITE(ILK_DSPCLK_GATE, IVB_VRHUNIT_CLK_GATE);
3475
3476         I915_WRITE(IVB_CHICKEN3,
3477                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
3478                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
3479
3480         /* Apply the WaDisableRHWOOptimizationForRenderHang workaround. */
3481         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
3482                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
3483
3484         /* WaApplyL3ControlAndL3ChickenMode requires those two on Ivy Bridge */
3485         I915_WRITE(GEN7_L3CNTLREG1,
3486                         GEN7_WA_FOR_GEN7_L3_CONTROL);
3487         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
3488                         GEN7_WA_L3_CHICKEN_MODE);
3489
3490         /* This is required by WaCatErrorRejectionIssue */
3491         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
3492                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
3493                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
3494
3495         for_each_pipe(pipe) {
3496                 I915_WRITE(DSPCNTR(pipe),
3497                            I915_READ(DSPCNTR(pipe)) |
3498                            DISPPLANE_TRICKLE_FEED_DISABLE);
3499                 intel_flush_display_plane(dev_priv, pipe);
3500         }
3501
3502         gen7_setup_fixed_func_scheduler(dev_priv);
3503
3504         /* WaDisable4x2SubspanOptimization */
3505         I915_WRITE(CACHE_MODE_1,
3506                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
3507
3508         /* XXX: This is a workaround for early silicon revisions and should be
3509          * removed later.
3510          */
3511         I915_WRITE(WM_DBG,
3512                         I915_READ(WM_DBG) |
3513                         WM_DBG_DISALLOW_MULTIPLE_LP |
3514                         WM_DBG_DISALLOW_SPRITE |
3515                         WM_DBG_DISALLOW_MAXFIFO);
3516
3517 }
3518
3519 static void ivybridge_init_clock_gating(struct drm_device *dev)
3520 {
3521         struct drm_i915_private *dev_priv = dev->dev_private;
3522         int pipe;
3523         uint32_t dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE;
3524         uint32_t snpcr;
3525
3526         I915_WRITE(PCH_DSPCLK_GATE_D, dspclk_gate);
3527
3528         I915_WRITE(WM3_LP_ILK, 0);
3529         I915_WRITE(WM2_LP_ILK, 0);
3530         I915_WRITE(WM1_LP_ILK, 0);
3531
3532         I915_WRITE(ILK_DSPCLK_GATE, IVB_VRHUNIT_CLK_GATE);
3533
3534         I915_WRITE(IVB_CHICKEN3,
3535                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
3536                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
3537
3538         /* Apply the WaDisableRHWOOptimizationForRenderHang workaround. */
3539         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
3540                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
3541
3542         /* WaApplyL3ControlAndL3ChickenMode requires those two on Ivy Bridge */
3543         I915_WRITE(GEN7_L3CNTLREG1,
3544                         GEN7_WA_FOR_GEN7_L3_CONTROL);
3545         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
3546                         GEN7_WA_L3_CHICKEN_MODE);
3547
3548         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
3549          * gating disable must be set.  Failure to set it results in
3550          * flickering pixels due to Z write ordering failures after
3551          * some amount of runtime in the Mesa "fire" demo, and Unigine
3552          * Sanctuary and Tropics, and apparently anything else with
3553          * alpha test or pixel discard.
3554          *
3555          * According to the spec, bit 11 (RCCUNIT) must also be set,
3556          * but we didn't debug actual testcases to find it out.
3557          *
3558          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
3559          * This implements the WaDisableRCZUnitClockGating workaround.
3560          */
3561         I915_WRITE(GEN6_UCGCTL2,
3562                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE |
3563                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
3564
3565         /* This is required by WaCatErrorRejectionIssue */
3566         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
3567                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
3568                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
3569
3570         for_each_pipe(pipe) {
3571                 I915_WRITE(DSPCNTR(pipe),
3572                            I915_READ(DSPCNTR(pipe)) |
3573                            DISPPLANE_TRICKLE_FEED_DISABLE);
3574                 intel_flush_display_plane(dev_priv, pipe);
3575         }
3576
3577         I915_WRITE(GEN6_MBCTL, I915_READ(GEN6_MBCTL) |
3578                    GEN6_MBCTL_ENABLE_BOOT_FETCH);
3579
3580         gen7_setup_fixed_func_scheduler(dev_priv);
3581
3582         /* WaDisable4x2SubspanOptimization */
3583         I915_WRITE(CACHE_MODE_1,
3584                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
3585
3586         snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
3587         snpcr &= ~GEN6_MBC_SNPCR_MASK;
3588         snpcr |= GEN6_MBC_SNPCR_MED;
3589         I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
3590 }
3591
3592 static void valleyview_init_clock_gating(struct drm_device *dev)
3593 {
3594         struct drm_i915_private *dev_priv = dev->dev_private;
3595         int pipe;
3596         uint32_t dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE;
3597
3598         I915_WRITE(PCH_DSPCLK_GATE_D, dspclk_gate);
3599
3600         I915_WRITE(WM3_LP_ILK, 0);
3601         I915_WRITE(WM2_LP_ILK, 0);
3602         I915_WRITE(WM1_LP_ILK, 0);
3603
3604         I915_WRITE(ILK_DSPCLK_GATE, IVB_VRHUNIT_CLK_GATE);
3605
3606         I915_WRITE(IVB_CHICKEN3,
3607                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
3608                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
3609
3610         /* Apply the WaDisableRHWOOptimizationForRenderHang workaround. */
3611         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
3612                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
3613
3614         /* WaApplyL3ControlAndL3ChickenMode requires those two on Ivy Bridge */
3615         I915_WRITE(GEN7_L3CNTLREG1, GEN7_WA_FOR_GEN7_L3_CONTROL);
3616         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER, GEN7_WA_L3_CHICKEN_MODE);
3617
3618         /* This is required by WaCatErrorRejectionIssue */
3619         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
3620                    I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
3621                    GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
3622
3623         I915_WRITE(GEN6_MBCTL, I915_READ(GEN6_MBCTL) |
3624                    GEN6_MBCTL_ENABLE_BOOT_FETCH);
3625
3626
3627         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
3628          * gating disable must be set.  Failure to set it results in
3629          * flickering pixels due to Z write ordering failures after
3630          * some amount of runtime in the Mesa "fire" demo, and Unigine
3631          * Sanctuary and Tropics, and apparently anything else with
3632          * alpha test or pixel discard.
3633          *
3634          * According to the spec, bit 11 (RCCUNIT) must also be set,
3635          * but we didn't debug actual testcases to find it out.
3636          *
3637          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
3638          * This implements the WaDisableRCZUnitClockGating workaround.
3639          *
3640          * Also apply WaDisableVDSUnitClockGating and
3641          * WaDisableRCPBUnitClockGating.
3642          */
3643         I915_WRITE(GEN6_UCGCTL2,
3644                    GEN7_VDSUNIT_CLOCK_GATE_DISABLE |
3645                    GEN7_TDLUNIT_CLOCK_GATE_DISABLE |
3646                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE |
3647                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
3648                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
3649
3650         I915_WRITE(GEN7_UCGCTL4, GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
3651
3652         for_each_pipe(pipe) {
3653                 I915_WRITE(DSPCNTR(pipe),
3654                            I915_READ(DSPCNTR(pipe)) |
3655                            DISPPLANE_TRICKLE_FEED_DISABLE);
3656                 intel_flush_display_plane(dev_priv, pipe);
3657         }
3658
3659         I915_WRITE(CACHE_MODE_1,
3660                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
3661
3662         /*
3663          * On ValleyView, the GUnit needs to signal the GT
3664          * when flip and other events complete.  So enable
3665          * all the GUnit->GT interrupts here
3666          */
3667         I915_WRITE(VLV_DPFLIPSTAT, PIPEB_LINE_COMPARE_INT_EN |
3668                    PIPEB_HLINE_INT_EN | PIPEB_VBLANK_INT_EN |
3669                    SPRITED_FLIPDONE_INT_EN | SPRITEC_FLIPDONE_INT_EN |
3670                    PLANEB_FLIPDONE_INT_EN | PIPEA_LINE_COMPARE_INT_EN |
3671                    PIPEA_HLINE_INT_EN | PIPEA_VBLANK_INT_EN |
3672                    SPRITEB_FLIPDONE_INT_EN | SPRITEA_FLIPDONE_INT_EN |
3673                    PLANEA_FLIPDONE_INT_EN);
3674 }
3675
3676 static void g4x_init_clock_gating(struct drm_device *dev)
3677 {
3678         struct drm_i915_private *dev_priv = dev->dev_private;
3679         uint32_t dspclk_gate;
3680
3681         I915_WRITE(RENCLK_GATE_D1, 0);
3682         I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
3683                    GS_UNIT_CLOCK_GATE_DISABLE |
3684                    CL_UNIT_CLOCK_GATE_DISABLE);
3685         I915_WRITE(RAMCLK_GATE_D, 0);
3686         dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
3687                 OVRUNIT_CLOCK_GATE_DISABLE |
3688                 OVCUNIT_CLOCK_GATE_DISABLE;
3689         if (IS_GM45(dev))
3690                 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
3691         I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
3692 }
3693
3694 static void crestline_init_clock_gating(struct drm_device *dev)
3695 {
3696         struct drm_i915_private *dev_priv = dev->dev_private;
3697
3698         I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
3699         I915_WRITE(RENCLK_GATE_D2, 0);
3700         I915_WRITE(DSPCLK_GATE_D, 0);
3701         I915_WRITE(RAMCLK_GATE_D, 0);
3702         I915_WRITE16(DEUC, 0);
3703 }
3704
3705 static void broadwater_init_clock_gating(struct drm_device *dev)
3706 {
3707         struct drm_i915_private *dev_priv = dev->dev_private;
3708
3709         I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
3710                    I965_RCC_CLOCK_GATE_DISABLE |
3711                    I965_RCPB_CLOCK_GATE_DISABLE |
3712                    I965_ISC_CLOCK_GATE_DISABLE |
3713                    I965_FBC_CLOCK_GATE_DISABLE);
3714         I915_WRITE(RENCLK_GATE_D2, 0);
3715 }
3716
3717 static void gen3_init_clock_gating(struct drm_device *dev)
3718 {
3719         struct drm_i915_private *dev_priv = dev->dev_private;
3720         u32 dstate = I915_READ(D_STATE);
3721
3722         dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
3723                 DSTATE_DOT_CLOCK_GATING;
3724         I915_WRITE(D_STATE, dstate);
3725
3726         if (IS_PINEVIEW(dev))
3727                 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
3728 }
3729
3730 static void i85x_init_clock_gating(struct drm_device *dev)
3731 {
3732         struct drm_i915_private *dev_priv = dev->dev_private;
3733
3734         I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
3735 }
3736
3737 static void i830_init_clock_gating(struct drm_device *dev)
3738 {
3739         struct drm_i915_private *dev_priv = dev->dev_private;
3740
3741         I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
3742 }
3743
3744 static void ibx_init_clock_gating(struct drm_device *dev)
3745 {
3746         struct drm_i915_private *dev_priv = dev->dev_private;
3747
3748         /*
3749          * On Ibex Peak and Cougar Point, we need to disable clock
3750          * gating for the panel power sequencer or it will fail to
3751          * start up when no ports are active.
3752          */
3753         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
3754 }
3755
3756 static void cpt_init_clock_gating(struct drm_device *dev)
3757 {
3758         struct drm_i915_private *dev_priv = dev->dev_private;
3759         int pipe;
3760
3761         /*
3762          * On Ibex Peak and Cougar Point, we need to disable clock
3763          * gating for the panel power sequencer or it will fail to
3764          * start up when no ports are active.
3765          */
3766         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
3767         I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
3768                    DPLS_EDP_PPS_FIX_DIS);
3769         /* Without this, mode sets may fail silently on FDI */
3770         for_each_pipe(pipe)
3771                 I915_WRITE(TRANS_CHICKEN2(pipe), TRANS_AUTOTRAIN_GEN_STALL_DIS);
3772 }
3773
3774 void intel_init_clock_gating(struct drm_device *dev)
3775 {
3776         struct drm_i915_private *dev_priv = dev->dev_private;
3777
3778         dev_priv->display.init_clock_gating(dev);
3779
3780         if (dev_priv->display.init_pch_clock_gating)
3781                 dev_priv->display.init_pch_clock_gating(dev);
3782 }
3783
3784 /* Starting with Haswell, we have different power wells for
3785  * different parts of the GPU. This attempts to enable them all.
3786  */
3787 void intel_init_power_wells(struct drm_device *dev)
3788 {
3789         struct drm_i915_private *dev_priv = dev->dev_private;
3790         unsigned long power_wells[] = {
3791                 HSW_PWR_WELL_CTL1,
3792                 HSW_PWR_WELL_CTL2,
3793                 HSW_PWR_WELL_CTL4
3794         };
3795         int i;
3796
3797         if (!IS_HASWELL(dev))
3798                 return;
3799
3800         mutex_lock(&dev->struct_mutex);
3801
3802         for (i = 0; i < ARRAY_SIZE(power_wells); i++) {
3803                 int well = I915_READ(power_wells[i]);
3804
3805                 if ((well & HSW_PWR_WELL_STATE) == 0) {
3806                         I915_WRITE(power_wells[i], well & HSW_PWR_WELL_ENABLE);
3807                         if (wait_for(I915_READ(power_wells[i] & HSW_PWR_WELL_STATE), 20))
3808                                 DRM_ERROR("Error enabling power well %lx\n", power_wells[i]);
3809                 }
3810         }
3811
3812         mutex_unlock(&dev->struct_mutex);
3813 }
3814
3815 /* Set up chip specific power management-related functions */
3816 void intel_init_pm(struct drm_device *dev)
3817 {
3818         struct drm_i915_private *dev_priv = dev->dev_private;
3819
3820         if (I915_HAS_FBC(dev)) {
3821                 if (HAS_PCH_SPLIT(dev)) {
3822                         dev_priv->display.fbc_enabled = ironlake_fbc_enabled;
3823                         dev_priv->display.enable_fbc = ironlake_enable_fbc;
3824                         dev_priv->display.disable_fbc = ironlake_disable_fbc;
3825                 } else if (IS_GM45(dev)) {
3826                         dev_priv->display.fbc_enabled = g4x_fbc_enabled;
3827                         dev_priv->display.enable_fbc = g4x_enable_fbc;
3828                         dev_priv->display.disable_fbc = g4x_disable_fbc;
3829                 } else if (IS_CRESTLINE(dev)) {
3830                         dev_priv->display.fbc_enabled = i8xx_fbc_enabled;
3831                         dev_priv->display.enable_fbc = i8xx_enable_fbc;
3832                         dev_priv->display.disable_fbc = i8xx_disable_fbc;
3833                 }
3834                 /* 855GM needs testing */
3835         }
3836
3837         /* For cxsr */
3838         if (IS_PINEVIEW(dev))
3839                 i915_pineview_get_mem_freq(dev);
3840         else if (IS_GEN5(dev))
3841                 i915_ironlake_get_mem_freq(dev);
3842
3843         /* For FIFO watermark updates */
3844         if (HAS_PCH_SPLIT(dev)) {
3845                 if (HAS_PCH_IBX(dev))
3846                         dev_priv->display.init_pch_clock_gating = ibx_init_clock_gating;
3847                 else if (HAS_PCH_CPT(dev))
3848                         dev_priv->display.init_pch_clock_gating = cpt_init_clock_gating;
3849
3850                 if (IS_GEN5(dev)) {
3851                         if (I915_READ(MLTR_ILK) & ILK_SRLT_MASK)
3852                                 dev_priv->display.update_wm = ironlake_update_wm;
3853                         else {
3854                                 DRM_DEBUG_KMS("Failed to get proper latency. "
3855                                               "Disable CxSR\n");
3856                                 dev_priv->display.update_wm = NULL;
3857                         }
3858                         dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
3859                 } else if (IS_GEN6(dev)) {
3860                         if (SNB_READ_WM0_LATENCY()) {
3861                                 dev_priv->display.update_wm = sandybridge_update_wm;
3862                                 dev_priv->display.update_sprite_wm = sandybridge_update_sprite_wm;
3863                         } else {
3864                                 DRM_DEBUG_KMS("Failed to read display plane latency. "
3865                                               "Disable CxSR\n");
3866                                 dev_priv->display.update_wm = NULL;
3867                         }
3868                         dev_priv->display.init_clock_gating = gen6_init_clock_gating;
3869                 } else if (IS_IVYBRIDGE(dev)) {
3870                         /* FIXME: detect B0+ stepping and use auto training */
3871                         if (SNB_READ_WM0_LATENCY()) {
3872                                 dev_priv->display.update_wm = sandybridge_update_wm;
3873                                 dev_priv->display.update_sprite_wm = sandybridge_update_sprite_wm;
3874                         } else {
3875                                 DRM_DEBUG_KMS("Failed to read display plane latency. "
3876                                               "Disable CxSR\n");
3877                                 dev_priv->display.update_wm = NULL;
3878                         }
3879                         dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
3880                 } else if (IS_HASWELL(dev)) {
3881                         if (SNB_READ_WM0_LATENCY()) {
3882                                 dev_priv->display.update_wm = sandybridge_update_wm;
3883                                 dev_priv->display.update_sprite_wm = sandybridge_update_sprite_wm;
3884                                 dev_priv->display.update_linetime_wm = haswell_update_linetime_wm;
3885                         } else {
3886                                 DRM_DEBUG_KMS("Failed to read display plane latency. "
3887                                               "Disable CxSR\n");
3888                                 dev_priv->display.update_wm = NULL;
3889                         }
3890                         dev_priv->display.init_clock_gating = haswell_init_clock_gating;
3891                 } else
3892                         dev_priv->display.update_wm = NULL;
3893         } else if (IS_VALLEYVIEW(dev)) {
3894                 dev_priv->display.update_wm = valleyview_update_wm;
3895                 dev_priv->display.init_clock_gating =
3896                         valleyview_init_clock_gating;
3897         } else if (IS_PINEVIEW(dev)) {
3898                 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
3899                                             dev_priv->is_ddr3,
3900                                             dev_priv->fsb_freq,
3901                                             dev_priv->mem_freq)) {
3902                         DRM_INFO("failed to find known CxSR latency "
3903                                  "(found ddr%s fsb freq %d, mem freq %d), "
3904                                  "disabling CxSR\n",
3905                                  (dev_priv->is_ddr3 == 1) ? "3" : "2",
3906                                  dev_priv->fsb_freq, dev_priv->mem_freq);
3907                         /* Disable CxSR and never update its watermark again */
3908                         pineview_disable_cxsr(dev);
3909                         dev_priv->display.update_wm = NULL;
3910                 } else
3911                         dev_priv->display.update_wm = pineview_update_wm;
3912                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
3913         } else if (IS_G4X(dev)) {
3914                 dev_priv->display.update_wm = g4x_update_wm;
3915                 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
3916         } else if (IS_GEN4(dev)) {
3917                 dev_priv->display.update_wm = i965_update_wm;
3918                 if (IS_CRESTLINE(dev))
3919                         dev_priv->display.init_clock_gating = crestline_init_clock_gating;
3920                 else if (IS_BROADWATER(dev))
3921                         dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
3922         } else if (IS_GEN3(dev)) {
3923                 dev_priv->display.update_wm = i9xx_update_wm;
3924                 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
3925                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
3926         } else if (IS_I865G(dev)) {
3927                 dev_priv->display.update_wm = i830_update_wm;
3928                 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
3929                 dev_priv->display.get_fifo_size = i830_get_fifo_size;
3930         } else if (IS_I85X(dev)) {
3931                 dev_priv->display.update_wm = i9xx_update_wm;
3932                 dev_priv->display.get_fifo_size = i85x_get_fifo_size;
3933                 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
3934         } else {
3935                 dev_priv->display.update_wm = i830_update_wm;
3936                 dev_priv->display.init_clock_gating = i830_init_clock_gating;
3937                 if (IS_845G(dev))
3938                         dev_priv->display.get_fifo_size = i845_get_fifo_size;
3939                 else
3940                         dev_priv->display.get_fifo_size = i830_get_fifo_size;
3941         }
3942 }
3943
3944 static void __gen6_gt_wait_for_thread_c0(struct drm_i915_private *dev_priv)
3945 {
3946         u32 gt_thread_status_mask;
3947
3948         if (IS_HASWELL(dev_priv->dev))
3949                 gt_thread_status_mask = GEN6_GT_THREAD_STATUS_CORE_MASK_HSW;
3950         else
3951                 gt_thread_status_mask = GEN6_GT_THREAD_STATUS_CORE_MASK;
3952
3953         /* w/a for a sporadic read returning 0 by waiting for the GT
3954          * thread to wake up.
3955          */
3956         if (wait_for_atomic_us((I915_READ_NOTRACE(GEN6_GT_THREAD_STATUS_REG) & gt_thread_status_mask) == 0, 500))
3957                 DRM_ERROR("GT thread status wait timed out\n");
3958 }
3959
3960 static void __gen6_gt_force_wake_get(struct drm_i915_private *dev_priv)
3961 {
3962         u32 forcewake_ack;
3963
3964         if (IS_HASWELL(dev_priv->dev))
3965                 forcewake_ack = FORCEWAKE_ACK_HSW;
3966         else
3967                 forcewake_ack = FORCEWAKE_ACK;
3968
3969         if (wait_for_atomic_us((I915_READ_NOTRACE(forcewake_ack) & 1) == 0, 500))
3970                 DRM_ERROR("Force wake wait timed out\n");
3971
3972         I915_WRITE_NOTRACE(FORCEWAKE, 1);
3973         POSTING_READ(FORCEWAKE);
3974
3975         if (wait_for_atomic_us((I915_READ_NOTRACE(forcewake_ack) & 1), 500))
3976                 DRM_ERROR("Force wake wait timed out\n");
3977
3978         __gen6_gt_wait_for_thread_c0(dev_priv);
3979 }
3980
3981 static void __gen6_gt_force_wake_mt_get(struct drm_i915_private *dev_priv)
3982 {
3983         u32 forcewake_ack;
3984
3985         if (IS_HASWELL(dev_priv->dev))
3986                 forcewake_ack = FORCEWAKE_ACK_HSW;
3987         else
3988                 forcewake_ack = FORCEWAKE_MT_ACK;
3989
3990         if (wait_for_atomic_us((I915_READ_NOTRACE(forcewake_ack) & 1) == 0, 500))
3991                 DRM_ERROR("Force wake wait timed out\n");
3992
3993         I915_WRITE_NOTRACE(FORCEWAKE_MT, _MASKED_BIT_ENABLE(1));
3994         POSTING_READ(FORCEWAKE_MT);
3995
3996         if (wait_for_atomic_us((I915_READ_NOTRACE(forcewake_ack) & 1), 500))
3997                 DRM_ERROR("Force wake wait timed out\n");
3998
3999         __gen6_gt_wait_for_thread_c0(dev_priv);
4000 }
4001
4002 /*
4003  * Generally this is called implicitly by the register read function. However,
4004  * if some sequence requires the GT to not power down then this function should
4005  * be called at the beginning of the sequence followed by a call to
4006  * gen6_gt_force_wake_put() at the end of the sequence.
4007  */
4008 void gen6_gt_force_wake_get(struct drm_i915_private *dev_priv)
4009 {
4010         unsigned long irqflags;
4011
4012         spin_lock_irqsave(&dev_priv->gt_lock, irqflags);
4013         if (dev_priv->forcewake_count++ == 0)
4014                 dev_priv->gt.force_wake_get(dev_priv);
4015         spin_unlock_irqrestore(&dev_priv->gt_lock, irqflags);
4016 }
4017
4018 void gen6_gt_check_fifodbg(struct drm_i915_private *dev_priv)
4019 {
4020         u32 gtfifodbg;
4021         gtfifodbg = I915_READ_NOTRACE(GTFIFODBG);
4022         if (WARN(gtfifodbg & GT_FIFO_CPU_ERROR_MASK,
4023              "MMIO read or write has been dropped %x\n", gtfifodbg))
4024                 I915_WRITE_NOTRACE(GTFIFODBG, GT_FIFO_CPU_ERROR_MASK);
4025 }
4026
4027 static void __gen6_gt_force_wake_put(struct drm_i915_private *dev_priv)
4028 {
4029         I915_WRITE_NOTRACE(FORCEWAKE, 0);
4030         POSTING_READ(FORCEWAKE);
4031         gen6_gt_check_fifodbg(dev_priv);
4032 }
4033
4034 static void __gen6_gt_force_wake_mt_put(struct drm_i915_private *dev_priv)
4035 {
4036         I915_WRITE_NOTRACE(FORCEWAKE_MT, _MASKED_BIT_DISABLE(1));
4037         POSTING_READ(FORCEWAKE_MT);
4038         gen6_gt_check_fifodbg(dev_priv);
4039 }
4040
4041 /*
4042  * see gen6_gt_force_wake_get()
4043  */
4044 void gen6_gt_force_wake_put(struct drm_i915_private *dev_priv)
4045 {
4046         unsigned long irqflags;
4047
4048         spin_lock_irqsave(&dev_priv->gt_lock, irqflags);
4049         if (--dev_priv->forcewake_count == 0)
4050                 dev_priv->gt.force_wake_put(dev_priv);
4051         spin_unlock_irqrestore(&dev_priv->gt_lock, irqflags);
4052 }
4053
4054 int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
4055 {
4056         int ret = 0;
4057
4058         if (dev_priv->gt_fifo_count < GT_FIFO_NUM_RESERVED_ENTRIES) {
4059                 int loop = 500;
4060                 u32 fifo = I915_READ_NOTRACE(GT_FIFO_FREE_ENTRIES);
4061                 while (fifo <= GT_FIFO_NUM_RESERVED_ENTRIES && loop--) {
4062                         udelay(10);
4063                         fifo = I915_READ_NOTRACE(GT_FIFO_FREE_ENTRIES);
4064                 }
4065                 if (WARN_ON(loop < 0 && fifo <= GT_FIFO_NUM_RESERVED_ENTRIES))
4066                         ++ret;
4067                 dev_priv->gt_fifo_count = fifo;
4068         }
4069         dev_priv->gt_fifo_count--;
4070
4071         return ret;
4072 }
4073
4074 static void vlv_force_wake_get(struct drm_i915_private *dev_priv)
4075 {
4076         /* Already awake? */
4077         if ((I915_READ(0x130094) & 0xa1) == 0xa1)
4078                 return;
4079
4080         I915_WRITE_NOTRACE(FORCEWAKE_VLV, 0xffffffff);
4081         POSTING_READ(FORCEWAKE_VLV);
4082
4083         if (wait_for_atomic_us((I915_READ_NOTRACE(FORCEWAKE_ACK_VLV) & 1), 500))
4084                 DRM_ERROR("Force wake wait timed out\n");
4085
4086         __gen6_gt_wait_for_thread_c0(dev_priv);
4087 }
4088
4089 static void vlv_force_wake_put(struct drm_i915_private *dev_priv)
4090 {
4091         I915_WRITE_NOTRACE(FORCEWAKE_VLV, 0xffff0000);
4092         /* FIXME: confirm VLV behavior with Punit folks */
4093         POSTING_READ(FORCEWAKE_VLV);
4094 }
4095
4096 void intel_gt_init(struct drm_device *dev)
4097 {
4098         struct drm_i915_private *dev_priv = dev->dev_private;
4099
4100         spin_lock_init(&dev_priv->gt_lock);
4101
4102         if (IS_VALLEYVIEW(dev)) {
4103                 dev_priv->gt.force_wake_get = vlv_force_wake_get;
4104                 dev_priv->gt.force_wake_put = vlv_force_wake_put;
4105         } else if (INTEL_INFO(dev)->gen >= 6) {
4106                 dev_priv->gt.force_wake_get = __gen6_gt_force_wake_get;
4107                 dev_priv->gt.force_wake_put = __gen6_gt_force_wake_put;
4108
4109                 /* IVB configs may use multi-threaded forcewake */
4110                 if (IS_IVYBRIDGE(dev) || IS_HASWELL(dev)) {
4111                         u32 ecobus;
4112
4113                         /* A small trick here - if the bios hasn't configured
4114                          * MT forcewake, and if the device is in RC6, then
4115                          * force_wake_mt_get will not wake the device and the
4116                          * ECOBUS read will return zero. Which will be
4117                          * (correctly) interpreted by the test below as MT
4118                          * forcewake being disabled.
4119                          */
4120                         mutex_lock(&dev->struct_mutex);
4121                         __gen6_gt_force_wake_mt_get(dev_priv);
4122                         ecobus = I915_READ_NOTRACE(ECOBUS);
4123                         __gen6_gt_force_wake_mt_put(dev_priv);
4124                         mutex_unlock(&dev->struct_mutex);
4125
4126                         if (ecobus & FORCEWAKE_MT_ENABLE) {
4127                                 DRM_DEBUG_KMS("Using MT version of forcewake\n");
4128                                 dev_priv->gt.force_wake_get =
4129                                         __gen6_gt_force_wake_mt_get;
4130                                 dev_priv->gt.force_wake_put =
4131                                         __gen6_gt_force_wake_mt_put;
4132                         }
4133                 }
4134         }
4135 }
4136