2 * Copyright © 2012 Intel Corporation
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:
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
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
24 * Eugeni Dodonov <eugeni.dodonov@intel.com>
28 #include <linux/cpufreq.h>
29 #include <drm/drm_plane_helper.h>
31 #include "intel_drv.h"
32 #include "../../../platform/x86/intel_ips.h"
33 #include <linux/module.h>
38 * RC6 is a special power stage which allows the GPU to enter an very
39 * low-voltage mode when idle, using down to 0V while at this stage. This
40 * stage is entered automatically when the GPU is idle when RC6 support is
41 * enabled, and as soon as new workload arises GPU wakes up automatically as well.
43 * There are different RC6 modes available in Intel GPU, which differentiate
44 * among each other with the latency required to enter and leave RC6 and
45 * voltage consumed by the GPU in different states.
47 * The combination of the following flags define which states GPU is allowed
48 * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
49 * RC6pp is deepest RC6. Their support by hardware varies according to the
50 * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
51 * which brings the most power savings; deeper states save more power, but
52 * require higher latency to switch to and wake up.
54 #define INTEL_RC6_ENABLE (1<<0)
55 #define INTEL_RC6p_ENABLE (1<<1)
56 #define INTEL_RC6pp_ENABLE (1<<2)
58 static void gen9_init_clock_gating(struct drm_device *dev)
60 struct drm_i915_private *dev_priv = dev->dev_private;
62 /* See Bspec note for PSR2_CTL bit 31, Wa#828:skl,bxt,kbl */
63 I915_WRITE(CHICKEN_PAR1_1,
64 I915_READ(CHICKEN_PAR1_1) | SKL_EDP_PSR_FIX_RDWRAP);
66 I915_WRITE(GEN8_CONFIG0,
67 I915_READ(GEN8_CONFIG0) | GEN9_DEFAULT_FIXES);
69 /* WaEnableChickenDCPR:skl,bxt,kbl */
70 I915_WRITE(GEN8_CHICKEN_DCPR_1,
71 I915_READ(GEN8_CHICKEN_DCPR_1) | MASK_WAKEMEM);
73 /* WaFbcTurnOffFbcWatermark:skl,bxt,kbl */
74 /* WaFbcWakeMemOn:skl,bxt,kbl */
75 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
77 DISP_FBC_MEMORY_WAKE);
79 /* WaFbcHighMemBwCorruptionAvoidance:skl,bxt,kbl */
80 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
81 ILK_DPFC_DISABLE_DUMMY0);
84 static void bxt_init_clock_gating(struct drm_device *dev)
86 struct drm_i915_private *dev_priv = to_i915(dev);
88 gen9_init_clock_gating(dev);
90 /* WaDisableSDEUnitClockGating:bxt */
91 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
92 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
96 * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
98 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
99 GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ);
102 * Wa: Backlight PWM may stop in the asserted state, causing backlight
105 if (IS_BXT_REVID(dev_priv, BXT_REVID_B0, REVID_FOREVER))
106 I915_WRITE(GEN9_CLKGATE_DIS_0, I915_READ(GEN9_CLKGATE_DIS_0) |
107 PWM1_GATING_DIS | PWM2_GATING_DIS);
110 static void i915_pineview_get_mem_freq(struct drm_device *dev)
112 struct drm_i915_private *dev_priv = to_i915(dev);
115 tmp = I915_READ(CLKCFG);
117 switch (tmp & CLKCFG_FSB_MASK) {
119 dev_priv->fsb_freq = 533; /* 133*4 */
122 dev_priv->fsb_freq = 800; /* 200*4 */
125 dev_priv->fsb_freq = 667; /* 167*4 */
128 dev_priv->fsb_freq = 400; /* 100*4 */
132 switch (tmp & CLKCFG_MEM_MASK) {
134 dev_priv->mem_freq = 533;
137 dev_priv->mem_freq = 667;
140 dev_priv->mem_freq = 800;
144 /* detect pineview DDR3 setting */
145 tmp = I915_READ(CSHRDDR3CTL);
146 dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
149 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
151 struct drm_i915_private *dev_priv = to_i915(dev);
154 ddrpll = I915_READ16(DDRMPLL1);
155 csipll = I915_READ16(CSIPLL0);
157 switch (ddrpll & 0xff) {
159 dev_priv->mem_freq = 800;
162 dev_priv->mem_freq = 1066;
165 dev_priv->mem_freq = 1333;
168 dev_priv->mem_freq = 1600;
171 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
173 dev_priv->mem_freq = 0;
177 dev_priv->ips.r_t = dev_priv->mem_freq;
179 switch (csipll & 0x3ff) {
181 dev_priv->fsb_freq = 3200;
184 dev_priv->fsb_freq = 3733;
187 dev_priv->fsb_freq = 4266;
190 dev_priv->fsb_freq = 4800;
193 dev_priv->fsb_freq = 5333;
196 dev_priv->fsb_freq = 5866;
199 dev_priv->fsb_freq = 6400;
202 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
204 dev_priv->fsb_freq = 0;
208 if (dev_priv->fsb_freq == 3200) {
209 dev_priv->ips.c_m = 0;
210 } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
211 dev_priv->ips.c_m = 1;
213 dev_priv->ips.c_m = 2;
217 static const struct cxsr_latency cxsr_latency_table[] = {
218 {1, 0, 800, 400, 3382, 33382, 3983, 33983}, /* DDR2-400 SC */
219 {1, 0, 800, 667, 3354, 33354, 3807, 33807}, /* DDR2-667 SC */
220 {1, 0, 800, 800, 3347, 33347, 3763, 33763}, /* DDR2-800 SC */
221 {1, 1, 800, 667, 6420, 36420, 6873, 36873}, /* DDR3-667 SC */
222 {1, 1, 800, 800, 5902, 35902, 6318, 36318}, /* DDR3-800 SC */
224 {1, 0, 667, 400, 3400, 33400, 4021, 34021}, /* DDR2-400 SC */
225 {1, 0, 667, 667, 3372, 33372, 3845, 33845}, /* DDR2-667 SC */
226 {1, 0, 667, 800, 3386, 33386, 3822, 33822}, /* DDR2-800 SC */
227 {1, 1, 667, 667, 6438, 36438, 6911, 36911}, /* DDR3-667 SC */
228 {1, 1, 667, 800, 5941, 35941, 6377, 36377}, /* DDR3-800 SC */
230 {1, 0, 400, 400, 3472, 33472, 4173, 34173}, /* DDR2-400 SC */
231 {1, 0, 400, 667, 3443, 33443, 3996, 33996}, /* DDR2-667 SC */
232 {1, 0, 400, 800, 3430, 33430, 3946, 33946}, /* DDR2-800 SC */
233 {1, 1, 400, 667, 6509, 36509, 7062, 37062}, /* DDR3-667 SC */
234 {1, 1, 400, 800, 5985, 35985, 6501, 36501}, /* DDR3-800 SC */
236 {0, 0, 800, 400, 3438, 33438, 4065, 34065}, /* DDR2-400 SC */
237 {0, 0, 800, 667, 3410, 33410, 3889, 33889}, /* DDR2-667 SC */
238 {0, 0, 800, 800, 3403, 33403, 3845, 33845}, /* DDR2-800 SC */
239 {0, 1, 800, 667, 6476, 36476, 6955, 36955}, /* DDR3-667 SC */
240 {0, 1, 800, 800, 5958, 35958, 6400, 36400}, /* DDR3-800 SC */
242 {0, 0, 667, 400, 3456, 33456, 4103, 34106}, /* DDR2-400 SC */
243 {0, 0, 667, 667, 3428, 33428, 3927, 33927}, /* DDR2-667 SC */
244 {0, 0, 667, 800, 3443, 33443, 3905, 33905}, /* DDR2-800 SC */
245 {0, 1, 667, 667, 6494, 36494, 6993, 36993}, /* DDR3-667 SC */
246 {0, 1, 667, 800, 5998, 35998, 6460, 36460}, /* DDR3-800 SC */
248 {0, 0, 400, 400, 3528, 33528, 4255, 34255}, /* DDR2-400 SC */
249 {0, 0, 400, 667, 3500, 33500, 4079, 34079}, /* DDR2-667 SC */
250 {0, 0, 400, 800, 3487, 33487, 4029, 34029}, /* DDR2-800 SC */
251 {0, 1, 400, 667, 6566, 36566, 7145, 37145}, /* DDR3-667 SC */
252 {0, 1, 400, 800, 6042, 36042, 6584, 36584}, /* DDR3-800 SC */
255 static const struct cxsr_latency *intel_get_cxsr_latency(bool is_desktop,
260 const struct cxsr_latency *latency;
263 if (fsb == 0 || mem == 0)
266 for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
267 latency = &cxsr_latency_table[i];
268 if (is_desktop == latency->is_desktop &&
269 is_ddr3 == latency->is_ddr3 &&
270 fsb == latency->fsb_freq && mem == latency->mem_freq)
274 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
279 static void chv_set_memory_dvfs(struct drm_i915_private *dev_priv, bool enable)
283 mutex_lock(&dev_priv->rps.hw_lock);
285 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
287 val &= ~FORCE_DDR_HIGH_FREQ;
289 val |= FORCE_DDR_HIGH_FREQ;
290 val &= ~FORCE_DDR_LOW_FREQ;
291 val |= FORCE_DDR_FREQ_REQ_ACK;
292 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
294 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
295 FORCE_DDR_FREQ_REQ_ACK) == 0, 3))
296 DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
298 mutex_unlock(&dev_priv->rps.hw_lock);
301 static void chv_set_memory_pm5(struct drm_i915_private *dev_priv, bool enable)
305 mutex_lock(&dev_priv->rps.hw_lock);
307 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
309 val |= DSP_MAXFIFO_PM5_ENABLE;
311 val &= ~DSP_MAXFIFO_PM5_ENABLE;
312 vlv_punit_write(dev_priv, PUNIT_REG_DSPFREQ, val);
314 mutex_unlock(&dev_priv->rps.hw_lock);
317 #define FW_WM(value, plane) \
318 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
320 void intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
322 struct drm_device *dev = &dev_priv->drm;
325 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
326 I915_WRITE(FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
327 POSTING_READ(FW_BLC_SELF_VLV);
328 dev_priv->wm.vlv.cxsr = enable;
329 } else if (IS_G4X(dev_priv) || IS_CRESTLINE(dev_priv)) {
330 I915_WRITE(FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
331 POSTING_READ(FW_BLC_SELF);
332 } else if (IS_PINEVIEW(dev)) {
333 val = I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN;
334 val |= enable ? PINEVIEW_SELF_REFRESH_EN : 0;
335 I915_WRITE(DSPFW3, val);
336 POSTING_READ(DSPFW3);
337 } else if (IS_I945G(dev_priv) || IS_I945GM(dev_priv)) {
338 val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
339 _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
340 I915_WRITE(FW_BLC_SELF, val);
341 POSTING_READ(FW_BLC_SELF);
342 } else if (IS_I915GM(dev_priv)) {
344 * FIXME can't find a bit like this for 915G, and
345 * and yet it does have the related watermark in
346 * FW_BLC_SELF. What's going on?
348 val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
349 _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
350 I915_WRITE(INSTPM, val);
351 POSTING_READ(INSTPM);
356 DRM_DEBUG_KMS("memory self-refresh is %s\n",
357 enable ? "enabled" : "disabled");
362 * Latency for FIFO fetches is dependent on several factors:
363 * - memory configuration (speed, channels)
365 * - current MCH state
366 * It can be fairly high in some situations, so here we assume a fairly
367 * pessimal value. It's a tradeoff between extra memory fetches (if we
368 * set this value too high, the FIFO will fetch frequently to stay full)
369 * and power consumption (set it too low to save power and we might see
370 * FIFO underruns and display "flicker").
372 * A value of 5us seems to be a good balance; safe for very low end
373 * platforms but not overly aggressive on lower latency configs.
375 static const int pessimal_latency_ns = 5000;
377 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
378 ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
380 static int vlv_get_fifo_size(struct drm_device *dev,
381 enum pipe pipe, int plane)
383 struct drm_i915_private *dev_priv = to_i915(dev);
384 int sprite0_start, sprite1_start, size;
387 uint32_t dsparb, dsparb2, dsparb3;
389 dsparb = I915_READ(DSPARB);
390 dsparb2 = I915_READ(DSPARB2);
391 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 0, 0);
392 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 8, 4);
395 dsparb = I915_READ(DSPARB);
396 dsparb2 = I915_READ(DSPARB2);
397 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 16, 8);
398 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 24, 12);
401 dsparb2 = I915_READ(DSPARB2);
402 dsparb3 = I915_READ(DSPARB3);
403 sprite0_start = VLV_FIFO_START(dsparb3, dsparb2, 0, 16);
404 sprite1_start = VLV_FIFO_START(dsparb3, dsparb2, 8, 20);
412 size = sprite0_start;
415 size = sprite1_start - sprite0_start;
418 size = 512 - 1 - sprite1_start;
424 DRM_DEBUG_KMS("Pipe %c %s %c FIFO size: %d\n",
425 pipe_name(pipe), plane == 0 ? "primary" : "sprite",
426 plane == 0 ? plane_name(pipe) : sprite_name(pipe, plane - 1),
432 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
434 struct drm_i915_private *dev_priv = to_i915(dev);
435 uint32_t dsparb = I915_READ(DSPARB);
438 size = dsparb & 0x7f;
440 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
442 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
443 plane ? "B" : "A", size);
448 static int i830_get_fifo_size(struct drm_device *dev, int plane)
450 struct drm_i915_private *dev_priv = to_i915(dev);
451 uint32_t dsparb = I915_READ(DSPARB);
454 size = dsparb & 0x1ff;
456 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
457 size >>= 1; /* Convert to cachelines */
459 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
460 plane ? "B" : "A", size);
465 static int i845_get_fifo_size(struct drm_device *dev, int plane)
467 struct drm_i915_private *dev_priv = to_i915(dev);
468 uint32_t dsparb = I915_READ(DSPARB);
471 size = dsparb & 0x7f;
472 size >>= 2; /* Convert to cachelines */
474 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
481 /* Pineview has different values for various configs */
482 static const struct intel_watermark_params pineview_display_wm = {
483 .fifo_size = PINEVIEW_DISPLAY_FIFO,
484 .max_wm = PINEVIEW_MAX_WM,
485 .default_wm = PINEVIEW_DFT_WM,
486 .guard_size = PINEVIEW_GUARD_WM,
487 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
489 static const struct intel_watermark_params pineview_display_hplloff_wm = {
490 .fifo_size = PINEVIEW_DISPLAY_FIFO,
491 .max_wm = PINEVIEW_MAX_WM,
492 .default_wm = PINEVIEW_DFT_HPLLOFF_WM,
493 .guard_size = PINEVIEW_GUARD_WM,
494 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
496 static const struct intel_watermark_params pineview_cursor_wm = {
497 .fifo_size = PINEVIEW_CURSOR_FIFO,
498 .max_wm = PINEVIEW_CURSOR_MAX_WM,
499 .default_wm = PINEVIEW_CURSOR_DFT_WM,
500 .guard_size = PINEVIEW_CURSOR_GUARD_WM,
501 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
503 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
504 .fifo_size = PINEVIEW_CURSOR_FIFO,
505 .max_wm = PINEVIEW_CURSOR_MAX_WM,
506 .default_wm = PINEVIEW_CURSOR_DFT_WM,
507 .guard_size = PINEVIEW_CURSOR_GUARD_WM,
508 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
510 static const struct intel_watermark_params g4x_wm_info = {
511 .fifo_size = G4X_FIFO_SIZE,
512 .max_wm = G4X_MAX_WM,
513 .default_wm = G4X_MAX_WM,
515 .cacheline_size = G4X_FIFO_LINE_SIZE,
517 static const struct intel_watermark_params g4x_cursor_wm_info = {
518 .fifo_size = I965_CURSOR_FIFO,
519 .max_wm = I965_CURSOR_MAX_WM,
520 .default_wm = I965_CURSOR_DFT_WM,
522 .cacheline_size = G4X_FIFO_LINE_SIZE,
524 static const struct intel_watermark_params i965_cursor_wm_info = {
525 .fifo_size = I965_CURSOR_FIFO,
526 .max_wm = I965_CURSOR_MAX_WM,
527 .default_wm = I965_CURSOR_DFT_WM,
529 .cacheline_size = I915_FIFO_LINE_SIZE,
531 static const struct intel_watermark_params i945_wm_info = {
532 .fifo_size = I945_FIFO_SIZE,
533 .max_wm = I915_MAX_WM,
536 .cacheline_size = I915_FIFO_LINE_SIZE,
538 static const struct intel_watermark_params i915_wm_info = {
539 .fifo_size = I915_FIFO_SIZE,
540 .max_wm = I915_MAX_WM,
543 .cacheline_size = I915_FIFO_LINE_SIZE,
545 static const struct intel_watermark_params i830_a_wm_info = {
546 .fifo_size = I855GM_FIFO_SIZE,
547 .max_wm = I915_MAX_WM,
550 .cacheline_size = I830_FIFO_LINE_SIZE,
552 static const struct intel_watermark_params i830_bc_wm_info = {
553 .fifo_size = I855GM_FIFO_SIZE,
554 .max_wm = I915_MAX_WM/2,
557 .cacheline_size = I830_FIFO_LINE_SIZE,
559 static const struct intel_watermark_params i845_wm_info = {
560 .fifo_size = I830_FIFO_SIZE,
561 .max_wm = I915_MAX_WM,
564 .cacheline_size = I830_FIFO_LINE_SIZE,
568 * intel_calculate_wm - calculate watermark level
569 * @clock_in_khz: pixel clock
570 * @wm: chip FIFO params
571 * @cpp: bytes per pixel
572 * @latency_ns: memory latency for the platform
574 * Calculate the watermark level (the level at which the display plane will
575 * start fetching from memory again). Each chip has a different display
576 * FIFO size and allocation, so the caller needs to figure that out and pass
577 * in the correct intel_watermark_params structure.
579 * As the pixel clock runs, the FIFO will be drained at a rate that depends
580 * on the pixel size. When it reaches the watermark level, it'll start
581 * fetching FIFO line sized based chunks from memory until the FIFO fills
582 * past the watermark point. If the FIFO drains completely, a FIFO underrun
583 * will occur, and a display engine hang could result.
585 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
586 const struct intel_watermark_params *wm,
587 int fifo_size, int cpp,
588 unsigned long latency_ns)
590 long entries_required, wm_size;
593 * Note: we need to make sure we don't overflow for various clock &
595 * clocks go from a few thousand to several hundred thousand.
596 * latency is usually a few thousand
598 entries_required = ((clock_in_khz / 1000) * cpp * latency_ns) /
600 entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
602 DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
604 wm_size = fifo_size - (entries_required + wm->guard_size);
606 DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
608 /* Don't promote wm_size to unsigned... */
609 if (wm_size > (long)wm->max_wm)
610 wm_size = wm->max_wm;
612 wm_size = wm->default_wm;
615 * Bspec seems to indicate that the value shouldn't be lower than
616 * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
617 * Lets go for 8 which is the burst size since certain platforms
618 * already use a hardcoded 8 (which is what the spec says should be
627 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
629 struct drm_crtc *crtc, *enabled = NULL;
631 for_each_crtc(dev, crtc) {
632 if (intel_crtc_active(crtc)) {
642 static void pineview_update_wm(struct drm_crtc *unused_crtc)
644 struct drm_device *dev = unused_crtc->dev;
645 struct drm_i915_private *dev_priv = to_i915(dev);
646 struct drm_crtc *crtc;
647 const struct cxsr_latency *latency;
651 latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv),
656 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
657 intel_set_memory_cxsr(dev_priv, false);
661 crtc = single_enabled_crtc(dev);
663 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
664 int cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
665 int clock = adjusted_mode->crtc_clock;
668 wm = intel_calculate_wm(clock, &pineview_display_wm,
669 pineview_display_wm.fifo_size,
670 cpp, latency->display_sr);
671 reg = I915_READ(DSPFW1);
672 reg &= ~DSPFW_SR_MASK;
673 reg |= FW_WM(wm, SR);
674 I915_WRITE(DSPFW1, reg);
675 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
678 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
679 pineview_display_wm.fifo_size,
680 cpp, latency->cursor_sr);
681 reg = I915_READ(DSPFW3);
682 reg &= ~DSPFW_CURSOR_SR_MASK;
683 reg |= FW_WM(wm, CURSOR_SR);
684 I915_WRITE(DSPFW3, reg);
686 /* Display HPLL off SR */
687 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
688 pineview_display_hplloff_wm.fifo_size,
689 cpp, latency->display_hpll_disable);
690 reg = I915_READ(DSPFW3);
691 reg &= ~DSPFW_HPLL_SR_MASK;
692 reg |= FW_WM(wm, HPLL_SR);
693 I915_WRITE(DSPFW3, reg);
695 /* cursor HPLL off SR */
696 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
697 pineview_display_hplloff_wm.fifo_size,
698 cpp, latency->cursor_hpll_disable);
699 reg = I915_READ(DSPFW3);
700 reg &= ~DSPFW_HPLL_CURSOR_MASK;
701 reg |= FW_WM(wm, HPLL_CURSOR);
702 I915_WRITE(DSPFW3, reg);
703 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
705 intel_set_memory_cxsr(dev_priv, true);
707 intel_set_memory_cxsr(dev_priv, false);
711 static bool g4x_compute_wm0(struct drm_device *dev,
713 const struct intel_watermark_params *display,
714 int display_latency_ns,
715 const struct intel_watermark_params *cursor,
716 int cursor_latency_ns,
720 struct drm_crtc *crtc;
721 const struct drm_display_mode *adjusted_mode;
722 int htotal, hdisplay, clock, cpp;
723 int line_time_us, line_count;
724 int entries, tlb_miss;
726 crtc = intel_get_crtc_for_plane(dev, plane);
727 if (!intel_crtc_active(crtc)) {
728 *cursor_wm = cursor->guard_size;
729 *plane_wm = display->guard_size;
733 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
734 clock = adjusted_mode->crtc_clock;
735 htotal = adjusted_mode->crtc_htotal;
736 hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
737 cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
739 /* Use the small buffer method to calculate plane watermark */
740 entries = ((clock * cpp / 1000) * display_latency_ns) / 1000;
741 tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
744 entries = DIV_ROUND_UP(entries, display->cacheline_size);
745 *plane_wm = entries + display->guard_size;
746 if (*plane_wm > (int)display->max_wm)
747 *plane_wm = display->max_wm;
749 /* Use the large buffer method to calculate cursor watermark */
750 line_time_us = max(htotal * 1000 / clock, 1);
751 line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
752 entries = line_count * crtc->cursor->state->crtc_w * cpp;
753 tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
756 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
757 *cursor_wm = entries + cursor->guard_size;
758 if (*cursor_wm > (int)cursor->max_wm)
759 *cursor_wm = (int)cursor->max_wm;
765 * Check the wm result.
767 * If any calculated watermark values is larger than the maximum value that
768 * can be programmed into the associated watermark register, that watermark
771 static bool g4x_check_srwm(struct drm_device *dev,
772 int display_wm, int cursor_wm,
773 const struct intel_watermark_params *display,
774 const struct intel_watermark_params *cursor)
776 DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
777 display_wm, cursor_wm);
779 if (display_wm > display->max_wm) {
780 DRM_DEBUG_KMS("display watermark is too large(%d/%u), disabling\n",
781 display_wm, display->max_wm);
785 if (cursor_wm > cursor->max_wm) {
786 DRM_DEBUG_KMS("cursor watermark is too large(%d/%u), disabling\n",
787 cursor_wm, cursor->max_wm);
791 if (!(display_wm || cursor_wm)) {
792 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
799 static bool g4x_compute_srwm(struct drm_device *dev,
802 const struct intel_watermark_params *display,
803 const struct intel_watermark_params *cursor,
804 int *display_wm, int *cursor_wm)
806 struct drm_crtc *crtc;
807 const struct drm_display_mode *adjusted_mode;
808 int hdisplay, htotal, cpp, clock;
809 unsigned long line_time_us;
810 int line_count, line_size;
815 *display_wm = *cursor_wm = 0;
819 crtc = intel_get_crtc_for_plane(dev, plane);
820 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
821 clock = adjusted_mode->crtc_clock;
822 htotal = adjusted_mode->crtc_htotal;
823 hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
824 cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
826 line_time_us = max(htotal * 1000 / clock, 1);
827 line_count = (latency_ns / line_time_us + 1000) / 1000;
828 line_size = hdisplay * cpp;
830 /* Use the minimum of the small and large buffer method for primary */
831 small = ((clock * cpp / 1000) * latency_ns) / 1000;
832 large = line_count * line_size;
834 entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
835 *display_wm = entries + display->guard_size;
837 /* calculate the self-refresh watermark for display cursor */
838 entries = line_count * cpp * crtc->cursor->state->crtc_w;
839 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
840 *cursor_wm = entries + cursor->guard_size;
842 return g4x_check_srwm(dev,
843 *display_wm, *cursor_wm,
847 #define FW_WM_VLV(value, plane) \
848 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
850 static void vlv_write_wm_values(struct intel_crtc *crtc,
851 const struct vlv_wm_values *wm)
853 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
854 enum pipe pipe = crtc->pipe;
856 I915_WRITE(VLV_DDL(pipe),
857 (wm->ddl[pipe].cursor << DDL_CURSOR_SHIFT) |
858 (wm->ddl[pipe].sprite[1] << DDL_SPRITE_SHIFT(1)) |
859 (wm->ddl[pipe].sprite[0] << DDL_SPRITE_SHIFT(0)) |
860 (wm->ddl[pipe].primary << DDL_PLANE_SHIFT));
863 FW_WM(wm->sr.plane, SR) |
864 FW_WM(wm->pipe[PIPE_B].cursor, CURSORB) |
865 FW_WM_VLV(wm->pipe[PIPE_B].primary, PLANEB) |
866 FW_WM_VLV(wm->pipe[PIPE_A].primary, PLANEA));
868 FW_WM_VLV(wm->pipe[PIPE_A].sprite[1], SPRITEB) |
869 FW_WM(wm->pipe[PIPE_A].cursor, CURSORA) |
870 FW_WM_VLV(wm->pipe[PIPE_A].sprite[0], SPRITEA));
872 FW_WM(wm->sr.cursor, CURSOR_SR));
874 if (IS_CHERRYVIEW(dev_priv)) {
875 I915_WRITE(DSPFW7_CHV,
876 FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
877 FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
878 I915_WRITE(DSPFW8_CHV,
879 FW_WM_VLV(wm->pipe[PIPE_C].sprite[1], SPRITEF) |
880 FW_WM_VLV(wm->pipe[PIPE_C].sprite[0], SPRITEE));
881 I915_WRITE(DSPFW9_CHV,
882 FW_WM_VLV(wm->pipe[PIPE_C].primary, PLANEC) |
883 FW_WM(wm->pipe[PIPE_C].cursor, CURSORC));
885 FW_WM(wm->sr.plane >> 9, SR_HI) |
886 FW_WM(wm->pipe[PIPE_C].sprite[1] >> 8, SPRITEF_HI) |
887 FW_WM(wm->pipe[PIPE_C].sprite[0] >> 8, SPRITEE_HI) |
888 FW_WM(wm->pipe[PIPE_C].primary >> 8, PLANEC_HI) |
889 FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
890 FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
891 FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
892 FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
893 FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
894 FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
897 FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
898 FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
900 FW_WM(wm->sr.plane >> 9, SR_HI) |
901 FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
902 FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
903 FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
904 FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
905 FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
906 FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
909 /* zero (unused) WM1 watermarks */
910 I915_WRITE(DSPFW4, 0);
911 I915_WRITE(DSPFW5, 0);
912 I915_WRITE(DSPFW6, 0);
913 I915_WRITE(DSPHOWM1, 0);
915 POSTING_READ(DSPFW1);
923 VLV_WM_LEVEL_DDR_DVFS,
926 /* latency must be in 0.1us units. */
927 static unsigned int vlv_wm_method2(unsigned int pixel_rate,
928 unsigned int pipe_htotal,
929 unsigned int horiz_pixels,
931 unsigned int latency)
935 ret = (latency * pixel_rate) / (pipe_htotal * 10000);
936 ret = (ret + 1) * horiz_pixels * cpp;
937 ret = DIV_ROUND_UP(ret, 64);
942 static void vlv_setup_wm_latency(struct drm_device *dev)
944 struct drm_i915_private *dev_priv = to_i915(dev);
946 /* all latencies in usec */
947 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM2] = 3;
949 dev_priv->wm.max_level = VLV_WM_LEVEL_PM2;
951 if (IS_CHERRYVIEW(dev_priv)) {
952 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM5] = 12;
953 dev_priv->wm.pri_latency[VLV_WM_LEVEL_DDR_DVFS] = 33;
955 dev_priv->wm.max_level = VLV_WM_LEVEL_DDR_DVFS;
959 static uint16_t vlv_compute_wm_level(struct intel_plane *plane,
960 struct intel_crtc *crtc,
961 const struct intel_plane_state *state,
964 struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
965 int clock, htotal, cpp, width, wm;
967 if (dev_priv->wm.pri_latency[level] == 0)
970 if (!state->base.visible)
973 cpp = drm_format_plane_cpp(state->base.fb->pixel_format, 0);
974 clock = crtc->config->base.adjusted_mode.crtc_clock;
975 htotal = crtc->config->base.adjusted_mode.crtc_htotal;
976 width = crtc->config->pipe_src_w;
977 if (WARN_ON(htotal == 0))
980 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
982 * FIXME the formula gives values that are
983 * too big for the cursor FIFO, and hence we
984 * would never be able to use cursors. For
985 * now just hardcode the watermark.
989 wm = vlv_wm_method2(clock, htotal, width, cpp,
990 dev_priv->wm.pri_latency[level] * 10);
993 return min_t(int, wm, USHRT_MAX);
996 static void vlv_compute_fifo(struct intel_crtc *crtc)
998 struct drm_device *dev = crtc->base.dev;
999 struct vlv_wm_state *wm_state = &crtc->wm_state;
1000 struct intel_plane *plane;
1001 unsigned int total_rate = 0;
1002 const int fifo_size = 512 - 1;
1003 int fifo_extra, fifo_left = fifo_size;
1005 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1006 struct intel_plane_state *state =
1007 to_intel_plane_state(plane->base.state);
1009 if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
1012 if (state->base.visible) {
1013 wm_state->num_active_planes++;
1014 total_rate += drm_format_plane_cpp(state->base.fb->pixel_format, 0);
1018 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1019 struct intel_plane_state *state =
1020 to_intel_plane_state(plane->base.state);
1023 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
1024 plane->wm.fifo_size = 63;
1028 if (!state->base.visible) {
1029 plane->wm.fifo_size = 0;
1033 rate = drm_format_plane_cpp(state->base.fb->pixel_format, 0);
1034 plane->wm.fifo_size = fifo_size * rate / total_rate;
1035 fifo_left -= plane->wm.fifo_size;
1038 fifo_extra = DIV_ROUND_UP(fifo_left, wm_state->num_active_planes ?: 1);
1040 /* spread the remainder evenly */
1041 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1047 if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
1050 /* give it all to the first plane if none are active */
1051 if (plane->wm.fifo_size == 0 &&
1052 wm_state->num_active_planes)
1055 plane_extra = min(fifo_extra, fifo_left);
1056 plane->wm.fifo_size += plane_extra;
1057 fifo_left -= plane_extra;
1060 WARN_ON(fifo_left != 0);
1063 static void vlv_invert_wms(struct intel_crtc *crtc)
1065 struct vlv_wm_state *wm_state = &crtc->wm_state;
1068 for (level = 0; level < wm_state->num_levels; level++) {
1069 struct drm_device *dev = crtc->base.dev;
1070 const int sr_fifo_size = INTEL_INFO(dev)->num_pipes * 512 - 1;
1071 struct intel_plane *plane;
1073 wm_state->sr[level].plane = sr_fifo_size - wm_state->sr[level].plane;
1074 wm_state->sr[level].cursor = 63 - wm_state->sr[level].cursor;
1076 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1077 switch (plane->base.type) {
1079 case DRM_PLANE_TYPE_CURSOR:
1080 wm_state->wm[level].cursor = plane->wm.fifo_size -
1081 wm_state->wm[level].cursor;
1083 case DRM_PLANE_TYPE_PRIMARY:
1084 wm_state->wm[level].primary = plane->wm.fifo_size -
1085 wm_state->wm[level].primary;
1087 case DRM_PLANE_TYPE_OVERLAY:
1088 sprite = plane->plane;
1089 wm_state->wm[level].sprite[sprite] = plane->wm.fifo_size -
1090 wm_state->wm[level].sprite[sprite];
1097 static void vlv_compute_wm(struct intel_crtc *crtc)
1099 struct drm_device *dev = crtc->base.dev;
1100 struct vlv_wm_state *wm_state = &crtc->wm_state;
1101 struct intel_plane *plane;
1102 int sr_fifo_size = INTEL_INFO(dev)->num_pipes * 512 - 1;
1105 memset(wm_state, 0, sizeof(*wm_state));
1107 wm_state->cxsr = crtc->pipe != PIPE_C && crtc->wm.cxsr_allowed;
1108 wm_state->num_levels = to_i915(dev)->wm.max_level + 1;
1110 wm_state->num_active_planes = 0;
1112 vlv_compute_fifo(crtc);
1114 if (wm_state->num_active_planes != 1)
1115 wm_state->cxsr = false;
1117 if (wm_state->cxsr) {
1118 for (level = 0; level < wm_state->num_levels; level++) {
1119 wm_state->sr[level].plane = sr_fifo_size;
1120 wm_state->sr[level].cursor = 63;
1124 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1125 struct intel_plane_state *state =
1126 to_intel_plane_state(plane->base.state);
1128 if (!state->base.visible)
1131 /* normal watermarks */
1132 for (level = 0; level < wm_state->num_levels; level++) {
1133 int wm = vlv_compute_wm_level(plane, crtc, state, level);
1134 int max_wm = plane->base.type == DRM_PLANE_TYPE_CURSOR ? 63 : 511;
1137 if (WARN_ON(level == 0 && wm > max_wm))
1140 if (wm > plane->wm.fifo_size)
1143 switch (plane->base.type) {
1145 case DRM_PLANE_TYPE_CURSOR:
1146 wm_state->wm[level].cursor = wm;
1148 case DRM_PLANE_TYPE_PRIMARY:
1149 wm_state->wm[level].primary = wm;
1151 case DRM_PLANE_TYPE_OVERLAY:
1152 sprite = plane->plane;
1153 wm_state->wm[level].sprite[sprite] = wm;
1158 wm_state->num_levels = level;
1160 if (!wm_state->cxsr)
1163 /* maxfifo watermarks */
1164 switch (plane->base.type) {
1166 case DRM_PLANE_TYPE_CURSOR:
1167 for (level = 0; level < wm_state->num_levels; level++)
1168 wm_state->sr[level].cursor =
1169 wm_state->wm[level].cursor;
1171 case DRM_PLANE_TYPE_PRIMARY:
1172 for (level = 0; level < wm_state->num_levels; level++)
1173 wm_state->sr[level].plane =
1174 min(wm_state->sr[level].plane,
1175 wm_state->wm[level].primary);
1177 case DRM_PLANE_TYPE_OVERLAY:
1178 sprite = plane->plane;
1179 for (level = 0; level < wm_state->num_levels; level++)
1180 wm_state->sr[level].plane =
1181 min(wm_state->sr[level].plane,
1182 wm_state->wm[level].sprite[sprite]);
1187 /* clear any (partially) filled invalid levels */
1188 for (level = wm_state->num_levels; level < to_i915(dev)->wm.max_level + 1; level++) {
1189 memset(&wm_state->wm[level], 0, sizeof(wm_state->wm[level]));
1190 memset(&wm_state->sr[level], 0, sizeof(wm_state->sr[level]));
1193 vlv_invert_wms(crtc);
1196 #define VLV_FIFO(plane, value) \
1197 (((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1199 static void vlv_pipe_set_fifo_size(struct intel_crtc *crtc)
1201 struct drm_device *dev = crtc->base.dev;
1202 struct drm_i915_private *dev_priv = to_i915(dev);
1203 struct intel_plane *plane;
1204 int sprite0_start = 0, sprite1_start = 0, fifo_size = 0;
1206 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1207 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
1208 WARN_ON(plane->wm.fifo_size != 63);
1212 if (plane->base.type == DRM_PLANE_TYPE_PRIMARY)
1213 sprite0_start = plane->wm.fifo_size;
1214 else if (plane->plane == 0)
1215 sprite1_start = sprite0_start + plane->wm.fifo_size;
1217 fifo_size = sprite1_start + plane->wm.fifo_size;
1220 WARN_ON(fifo_size != 512 - 1);
1222 DRM_DEBUG_KMS("Pipe %c FIFO split %d / %d / %d\n",
1223 pipe_name(crtc->pipe), sprite0_start,
1224 sprite1_start, fifo_size);
1226 switch (crtc->pipe) {
1227 uint32_t dsparb, dsparb2, dsparb3;
1229 dsparb = I915_READ(DSPARB);
1230 dsparb2 = I915_READ(DSPARB2);
1232 dsparb &= ~(VLV_FIFO(SPRITEA, 0xff) |
1233 VLV_FIFO(SPRITEB, 0xff));
1234 dsparb |= (VLV_FIFO(SPRITEA, sprite0_start) |
1235 VLV_FIFO(SPRITEB, sprite1_start));
1237 dsparb2 &= ~(VLV_FIFO(SPRITEA_HI, 0x1) |
1238 VLV_FIFO(SPRITEB_HI, 0x1));
1239 dsparb2 |= (VLV_FIFO(SPRITEA_HI, sprite0_start >> 8) |
1240 VLV_FIFO(SPRITEB_HI, sprite1_start >> 8));
1242 I915_WRITE(DSPARB, dsparb);
1243 I915_WRITE(DSPARB2, dsparb2);
1246 dsparb = I915_READ(DSPARB);
1247 dsparb2 = I915_READ(DSPARB2);
1249 dsparb &= ~(VLV_FIFO(SPRITEC, 0xff) |
1250 VLV_FIFO(SPRITED, 0xff));
1251 dsparb |= (VLV_FIFO(SPRITEC, sprite0_start) |
1252 VLV_FIFO(SPRITED, sprite1_start));
1254 dsparb2 &= ~(VLV_FIFO(SPRITEC_HI, 0xff) |
1255 VLV_FIFO(SPRITED_HI, 0xff));
1256 dsparb2 |= (VLV_FIFO(SPRITEC_HI, sprite0_start >> 8) |
1257 VLV_FIFO(SPRITED_HI, sprite1_start >> 8));
1259 I915_WRITE(DSPARB, dsparb);
1260 I915_WRITE(DSPARB2, dsparb2);
1263 dsparb3 = I915_READ(DSPARB3);
1264 dsparb2 = I915_READ(DSPARB2);
1266 dsparb3 &= ~(VLV_FIFO(SPRITEE, 0xff) |
1267 VLV_FIFO(SPRITEF, 0xff));
1268 dsparb3 |= (VLV_FIFO(SPRITEE, sprite0_start) |
1269 VLV_FIFO(SPRITEF, sprite1_start));
1271 dsparb2 &= ~(VLV_FIFO(SPRITEE_HI, 0xff) |
1272 VLV_FIFO(SPRITEF_HI, 0xff));
1273 dsparb2 |= (VLV_FIFO(SPRITEE_HI, sprite0_start >> 8) |
1274 VLV_FIFO(SPRITEF_HI, sprite1_start >> 8));
1276 I915_WRITE(DSPARB3, dsparb3);
1277 I915_WRITE(DSPARB2, dsparb2);
1286 static void vlv_merge_wm(struct drm_device *dev,
1287 struct vlv_wm_values *wm)
1289 struct intel_crtc *crtc;
1290 int num_active_crtcs = 0;
1292 wm->level = to_i915(dev)->wm.max_level;
1295 for_each_intel_crtc(dev, crtc) {
1296 const struct vlv_wm_state *wm_state = &crtc->wm_state;
1301 if (!wm_state->cxsr)
1305 wm->level = min_t(int, wm->level, wm_state->num_levels - 1);
1308 if (num_active_crtcs != 1)
1311 if (num_active_crtcs > 1)
1312 wm->level = VLV_WM_LEVEL_PM2;
1314 for_each_intel_crtc(dev, crtc) {
1315 struct vlv_wm_state *wm_state = &crtc->wm_state;
1316 enum pipe pipe = crtc->pipe;
1321 wm->pipe[pipe] = wm_state->wm[wm->level];
1323 wm->sr = wm_state->sr[wm->level];
1325 wm->ddl[pipe].primary = DDL_PRECISION_HIGH | 2;
1326 wm->ddl[pipe].sprite[0] = DDL_PRECISION_HIGH | 2;
1327 wm->ddl[pipe].sprite[1] = DDL_PRECISION_HIGH | 2;
1328 wm->ddl[pipe].cursor = DDL_PRECISION_HIGH | 2;
1332 static void vlv_update_wm(struct drm_crtc *crtc)
1334 struct drm_device *dev = crtc->dev;
1335 struct drm_i915_private *dev_priv = to_i915(dev);
1336 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1337 enum pipe pipe = intel_crtc->pipe;
1338 struct vlv_wm_values wm = {};
1340 vlv_compute_wm(intel_crtc);
1341 vlv_merge_wm(dev, &wm);
1343 if (memcmp(&dev_priv->wm.vlv, &wm, sizeof(wm)) == 0) {
1344 /* FIXME should be part of crtc atomic commit */
1345 vlv_pipe_set_fifo_size(intel_crtc);
1349 if (wm.level < VLV_WM_LEVEL_DDR_DVFS &&
1350 dev_priv->wm.vlv.level >= VLV_WM_LEVEL_DDR_DVFS)
1351 chv_set_memory_dvfs(dev_priv, false);
1353 if (wm.level < VLV_WM_LEVEL_PM5 &&
1354 dev_priv->wm.vlv.level >= VLV_WM_LEVEL_PM5)
1355 chv_set_memory_pm5(dev_priv, false);
1357 if (!wm.cxsr && dev_priv->wm.vlv.cxsr)
1358 intel_set_memory_cxsr(dev_priv, false);
1360 /* FIXME should be part of crtc atomic commit */
1361 vlv_pipe_set_fifo_size(intel_crtc);
1363 vlv_write_wm_values(intel_crtc, &wm);
1365 DRM_DEBUG_KMS("Setting FIFO watermarks - %c: plane=%d, cursor=%d, "
1366 "sprite0=%d, sprite1=%d, SR: plane=%d, cursor=%d level=%d cxsr=%d\n",
1367 pipe_name(pipe), wm.pipe[pipe].primary, wm.pipe[pipe].cursor,
1368 wm.pipe[pipe].sprite[0], wm.pipe[pipe].sprite[1],
1369 wm.sr.plane, wm.sr.cursor, wm.level, wm.cxsr);
1371 if (wm.cxsr && !dev_priv->wm.vlv.cxsr)
1372 intel_set_memory_cxsr(dev_priv, true);
1374 if (wm.level >= VLV_WM_LEVEL_PM5 &&
1375 dev_priv->wm.vlv.level < VLV_WM_LEVEL_PM5)
1376 chv_set_memory_pm5(dev_priv, true);
1378 if (wm.level >= VLV_WM_LEVEL_DDR_DVFS &&
1379 dev_priv->wm.vlv.level < VLV_WM_LEVEL_DDR_DVFS)
1380 chv_set_memory_dvfs(dev_priv, true);
1382 dev_priv->wm.vlv = wm;
1385 #define single_plane_enabled(mask) is_power_of_2(mask)
1387 static void g4x_update_wm(struct drm_crtc *crtc)
1389 struct drm_device *dev = crtc->dev;
1390 static const int sr_latency_ns = 12000;
1391 struct drm_i915_private *dev_priv = to_i915(dev);
1392 int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1393 int plane_sr, cursor_sr;
1394 unsigned int enabled = 0;
1397 if (g4x_compute_wm0(dev, PIPE_A,
1398 &g4x_wm_info, pessimal_latency_ns,
1399 &g4x_cursor_wm_info, pessimal_latency_ns,
1400 &planea_wm, &cursora_wm))
1401 enabled |= 1 << PIPE_A;
1403 if (g4x_compute_wm0(dev, PIPE_B,
1404 &g4x_wm_info, pessimal_latency_ns,
1405 &g4x_cursor_wm_info, pessimal_latency_ns,
1406 &planeb_wm, &cursorb_wm))
1407 enabled |= 1 << PIPE_B;
1409 if (single_plane_enabled(enabled) &&
1410 g4x_compute_srwm(dev, ffs(enabled) - 1,
1413 &g4x_cursor_wm_info,
1414 &plane_sr, &cursor_sr)) {
1415 cxsr_enabled = true;
1417 cxsr_enabled = false;
1418 intel_set_memory_cxsr(dev_priv, false);
1419 plane_sr = cursor_sr = 0;
1422 DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, "
1423 "B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1424 planea_wm, cursora_wm,
1425 planeb_wm, cursorb_wm,
1426 plane_sr, cursor_sr);
1429 FW_WM(plane_sr, SR) |
1430 FW_WM(cursorb_wm, CURSORB) |
1431 FW_WM(planeb_wm, PLANEB) |
1432 FW_WM(planea_wm, PLANEA));
1434 (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1435 FW_WM(cursora_wm, CURSORA));
1436 /* HPLL off in SR has some issues on G4x... disable it */
1438 (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1439 FW_WM(cursor_sr, CURSOR_SR));
1442 intel_set_memory_cxsr(dev_priv, true);
1445 static void i965_update_wm(struct drm_crtc *unused_crtc)
1447 struct drm_device *dev = unused_crtc->dev;
1448 struct drm_i915_private *dev_priv = to_i915(dev);
1449 struct drm_crtc *crtc;
1454 /* Calc sr entries for one plane configs */
1455 crtc = single_enabled_crtc(dev);
1457 /* self-refresh has much higher latency */
1458 static const int sr_latency_ns = 12000;
1459 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1460 int clock = adjusted_mode->crtc_clock;
1461 int htotal = adjusted_mode->crtc_htotal;
1462 int hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
1463 int cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
1464 unsigned long line_time_us;
1467 line_time_us = max(htotal * 1000 / clock, 1);
1469 /* Use ns/us then divide to preserve precision */
1470 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1472 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1473 srwm = I965_FIFO_SIZE - entries;
1477 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1480 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1481 cpp * crtc->cursor->state->crtc_w;
1482 entries = DIV_ROUND_UP(entries,
1483 i965_cursor_wm_info.cacheline_size);
1484 cursor_sr = i965_cursor_wm_info.fifo_size -
1485 (entries + i965_cursor_wm_info.guard_size);
1487 if (cursor_sr > i965_cursor_wm_info.max_wm)
1488 cursor_sr = i965_cursor_wm_info.max_wm;
1490 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1491 "cursor %d\n", srwm, cursor_sr);
1493 cxsr_enabled = true;
1495 cxsr_enabled = false;
1496 /* Turn off self refresh if both pipes are enabled */
1497 intel_set_memory_cxsr(dev_priv, false);
1500 DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1503 /* 965 has limitations... */
1504 I915_WRITE(DSPFW1, FW_WM(srwm, SR) |
1508 I915_WRITE(DSPFW2, FW_WM(8, CURSORA) |
1509 FW_WM(8, PLANEC_OLD));
1510 /* update cursor SR watermark */
1511 I915_WRITE(DSPFW3, FW_WM(cursor_sr, CURSOR_SR));
1514 intel_set_memory_cxsr(dev_priv, true);
1519 static void i9xx_update_wm(struct drm_crtc *unused_crtc)
1521 struct drm_device *dev = unused_crtc->dev;
1522 struct drm_i915_private *dev_priv = to_i915(dev);
1523 const struct intel_watermark_params *wm_info;
1528 int planea_wm, planeb_wm;
1529 struct drm_crtc *crtc, *enabled = NULL;
1532 wm_info = &i945_wm_info;
1533 else if (!IS_GEN2(dev_priv))
1534 wm_info = &i915_wm_info;
1536 wm_info = &i830_a_wm_info;
1538 fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1539 crtc = intel_get_crtc_for_plane(dev, 0);
1540 if (intel_crtc_active(crtc)) {
1541 const struct drm_display_mode *adjusted_mode;
1542 int cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
1543 if (IS_GEN2(dev_priv))
1546 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1547 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1548 wm_info, fifo_size, cpp,
1549 pessimal_latency_ns);
1552 planea_wm = fifo_size - wm_info->guard_size;
1553 if (planea_wm > (long)wm_info->max_wm)
1554 planea_wm = wm_info->max_wm;
1557 if (IS_GEN2(dev_priv))
1558 wm_info = &i830_bc_wm_info;
1560 fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1561 crtc = intel_get_crtc_for_plane(dev, 1);
1562 if (intel_crtc_active(crtc)) {
1563 const struct drm_display_mode *adjusted_mode;
1564 int cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
1565 if (IS_GEN2(dev_priv))
1568 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1569 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1570 wm_info, fifo_size, cpp,
1571 pessimal_latency_ns);
1572 if (enabled == NULL)
1577 planeb_wm = fifo_size - wm_info->guard_size;
1578 if (planeb_wm > (long)wm_info->max_wm)
1579 planeb_wm = wm_info->max_wm;
1582 DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1584 if (IS_I915GM(dev_priv) && enabled) {
1585 struct drm_i915_gem_object *obj;
1587 obj = intel_fb_obj(enabled->primary->state->fb);
1589 /* self-refresh seems busted with untiled */
1590 if (!i915_gem_object_is_tiled(obj))
1595 * Overlay gets an aggressive default since video jitter is bad.
1599 /* Play safe and disable self-refresh before adjusting watermarks. */
1600 intel_set_memory_cxsr(dev_priv, false);
1602 /* Calc sr entries for one plane configs */
1603 if (HAS_FW_BLC(dev) && enabled) {
1604 /* self-refresh has much higher latency */
1605 static const int sr_latency_ns = 6000;
1606 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(enabled)->config->base.adjusted_mode;
1607 int clock = adjusted_mode->crtc_clock;
1608 int htotal = adjusted_mode->crtc_htotal;
1609 int hdisplay = to_intel_crtc(enabled)->config->pipe_src_w;
1610 int cpp = drm_format_plane_cpp(enabled->primary->state->fb->pixel_format, 0);
1611 unsigned long line_time_us;
1614 if (IS_I915GM(dev_priv) || IS_I945GM(dev_priv))
1617 line_time_us = max(htotal * 1000 / clock, 1);
1619 /* Use ns/us then divide to preserve precision */
1620 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1622 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1623 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1624 srwm = wm_info->fifo_size - entries;
1628 if (IS_I945G(dev_priv) || IS_I945GM(dev_priv))
1629 I915_WRITE(FW_BLC_SELF,
1630 FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1632 I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1635 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1636 planea_wm, planeb_wm, cwm, srwm);
1638 fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1639 fwater_hi = (cwm & 0x1f);
1641 /* Set request length to 8 cachelines per fetch */
1642 fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1643 fwater_hi = fwater_hi | (1 << 8);
1645 I915_WRITE(FW_BLC, fwater_lo);
1646 I915_WRITE(FW_BLC2, fwater_hi);
1649 intel_set_memory_cxsr(dev_priv, true);
1652 static void i845_update_wm(struct drm_crtc *unused_crtc)
1654 struct drm_device *dev = unused_crtc->dev;
1655 struct drm_i915_private *dev_priv = to_i915(dev);
1656 struct drm_crtc *crtc;
1657 const struct drm_display_mode *adjusted_mode;
1661 crtc = single_enabled_crtc(dev);
1665 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1666 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1668 dev_priv->display.get_fifo_size(dev, 0),
1669 4, pessimal_latency_ns);
1670 fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1671 fwater_lo |= (3<<8) | planea_wm;
1673 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1675 I915_WRITE(FW_BLC, fwater_lo);
1678 uint32_t ilk_pipe_pixel_rate(const struct intel_crtc_state *pipe_config)
1680 uint32_t pixel_rate;
1682 pixel_rate = pipe_config->base.adjusted_mode.crtc_clock;
1684 /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
1685 * adjust the pixel_rate here. */
1687 if (pipe_config->pch_pfit.enabled) {
1688 uint64_t pipe_w, pipe_h, pfit_w, pfit_h;
1689 uint32_t pfit_size = pipe_config->pch_pfit.size;
1691 pipe_w = pipe_config->pipe_src_w;
1692 pipe_h = pipe_config->pipe_src_h;
1694 pfit_w = (pfit_size >> 16) & 0xFFFF;
1695 pfit_h = pfit_size & 0xFFFF;
1696 if (pipe_w < pfit_w)
1698 if (pipe_h < pfit_h)
1701 if (WARN_ON(!pfit_w || !pfit_h))
1704 pixel_rate = div_u64((uint64_t) pixel_rate * pipe_w * pipe_h,
1711 /* latency must be in 0.1us units. */
1712 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t cpp, uint32_t latency)
1716 if (WARN(latency == 0, "Latency value missing\n"))
1719 ret = (uint64_t) pixel_rate * cpp * latency;
1720 ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
1725 /* latency must be in 0.1us units. */
1726 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
1727 uint32_t horiz_pixels, uint8_t cpp,
1732 if (WARN(latency == 0, "Latency value missing\n"))
1734 if (WARN_ON(!pipe_htotal))
1737 ret = (latency * pixel_rate) / (pipe_htotal * 10000);
1738 ret = (ret + 1) * horiz_pixels * cpp;
1739 ret = DIV_ROUND_UP(ret, 64) + 2;
1743 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
1747 * Neither of these should be possible since this function shouldn't be
1748 * called if the CRTC is off or the plane is invisible. But let's be
1749 * extra paranoid to avoid a potential divide-by-zero if we screw up
1750 * elsewhere in the driver.
1754 if (WARN_ON(!horiz_pixels))
1757 return DIV_ROUND_UP(pri_val * 64, horiz_pixels * cpp) + 2;
1760 struct ilk_wm_maximums {
1768 * For both WM_PIPE and WM_LP.
1769 * mem_value must be in 0.1us units.
1771 static uint32_t ilk_compute_pri_wm(const struct intel_crtc_state *cstate,
1772 const struct intel_plane_state *pstate,
1776 int cpp = pstate->base.fb ?
1777 drm_format_plane_cpp(pstate->base.fb->pixel_format, 0) : 0;
1778 uint32_t method1, method2;
1780 if (!cstate->base.active || !pstate->base.visible)
1783 method1 = ilk_wm_method1(ilk_pipe_pixel_rate(cstate), cpp, mem_value);
1788 method2 = ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1789 cstate->base.adjusted_mode.crtc_htotal,
1790 drm_rect_width(&pstate->base.dst),
1793 return min(method1, method2);
1797 * For both WM_PIPE and WM_LP.
1798 * mem_value must be in 0.1us units.
1800 static uint32_t ilk_compute_spr_wm(const struct intel_crtc_state *cstate,
1801 const struct intel_plane_state *pstate,
1804 int cpp = pstate->base.fb ?
1805 drm_format_plane_cpp(pstate->base.fb->pixel_format, 0) : 0;
1806 uint32_t method1, method2;
1808 if (!cstate->base.active || !pstate->base.visible)
1811 method1 = ilk_wm_method1(ilk_pipe_pixel_rate(cstate), cpp, mem_value);
1812 method2 = ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1813 cstate->base.adjusted_mode.crtc_htotal,
1814 drm_rect_width(&pstate->base.dst),
1816 return min(method1, method2);
1820 * For both WM_PIPE and WM_LP.
1821 * mem_value must be in 0.1us units.
1823 static uint32_t ilk_compute_cur_wm(const struct intel_crtc_state *cstate,
1824 const struct intel_plane_state *pstate,
1828 * We treat the cursor plane as always-on for the purposes of watermark
1829 * calculation. Until we have two-stage watermark programming merged,
1830 * this is necessary to avoid flickering.
1833 int width = pstate->base.visible ? pstate->base.crtc_w : 64;
1835 if (!cstate->base.active)
1838 return ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1839 cstate->base.adjusted_mode.crtc_htotal,
1840 width, cpp, mem_value);
1843 /* Only for WM_LP. */
1844 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state *cstate,
1845 const struct intel_plane_state *pstate,
1848 int cpp = pstate->base.fb ?
1849 drm_format_plane_cpp(pstate->base.fb->pixel_format, 0) : 0;
1851 if (!cstate->base.active || !pstate->base.visible)
1854 return ilk_wm_fbc(pri_val, drm_rect_width(&pstate->base.dst), cpp);
1857 static unsigned int ilk_display_fifo_size(const struct drm_device *dev)
1859 if (INTEL_INFO(dev)->gen >= 8)
1861 else if (INTEL_INFO(dev)->gen >= 7)
1867 static unsigned int ilk_plane_wm_reg_max(const struct drm_device *dev,
1868 int level, bool is_sprite)
1870 if (INTEL_INFO(dev)->gen >= 8)
1871 /* BDW primary/sprite plane watermarks */
1872 return level == 0 ? 255 : 2047;
1873 else if (INTEL_INFO(dev)->gen >= 7)
1874 /* IVB/HSW primary/sprite plane watermarks */
1875 return level == 0 ? 127 : 1023;
1876 else if (!is_sprite)
1877 /* ILK/SNB primary plane watermarks */
1878 return level == 0 ? 127 : 511;
1880 /* ILK/SNB sprite plane watermarks */
1881 return level == 0 ? 63 : 255;
1884 static unsigned int ilk_cursor_wm_reg_max(const struct drm_device *dev,
1887 if (INTEL_INFO(dev)->gen >= 7)
1888 return level == 0 ? 63 : 255;
1890 return level == 0 ? 31 : 63;
1893 static unsigned int ilk_fbc_wm_reg_max(const struct drm_device *dev)
1895 if (INTEL_INFO(dev)->gen >= 8)
1901 /* Calculate the maximum primary/sprite plane watermark */
1902 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
1904 const struct intel_wm_config *config,
1905 enum intel_ddb_partitioning ddb_partitioning,
1908 unsigned int fifo_size = ilk_display_fifo_size(dev);
1910 /* if sprites aren't enabled, sprites get nothing */
1911 if (is_sprite && !config->sprites_enabled)
1914 /* HSW allows LP1+ watermarks even with multiple pipes */
1915 if (level == 0 || config->num_pipes_active > 1) {
1916 fifo_size /= INTEL_INFO(dev)->num_pipes;
1919 * For some reason the non self refresh
1920 * FIFO size is only half of the self
1921 * refresh FIFO size on ILK/SNB.
1923 if (INTEL_INFO(dev)->gen <= 6)
1927 if (config->sprites_enabled) {
1928 /* level 0 is always calculated with 1:1 split */
1929 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
1938 /* clamp to max that the registers can hold */
1939 return min(fifo_size, ilk_plane_wm_reg_max(dev, level, is_sprite));
1942 /* Calculate the maximum cursor plane watermark */
1943 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
1945 const struct intel_wm_config *config)
1947 /* HSW LP1+ watermarks w/ multiple pipes */
1948 if (level > 0 && config->num_pipes_active > 1)
1951 /* otherwise just report max that registers can hold */
1952 return ilk_cursor_wm_reg_max(dev, level);
1955 static void ilk_compute_wm_maximums(const struct drm_device *dev,
1957 const struct intel_wm_config *config,
1958 enum intel_ddb_partitioning ddb_partitioning,
1959 struct ilk_wm_maximums *max)
1961 max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
1962 max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
1963 max->cur = ilk_cursor_wm_max(dev, level, config);
1964 max->fbc = ilk_fbc_wm_reg_max(dev);
1967 static void ilk_compute_wm_reg_maximums(struct drm_device *dev,
1969 struct ilk_wm_maximums *max)
1971 max->pri = ilk_plane_wm_reg_max(dev, level, false);
1972 max->spr = ilk_plane_wm_reg_max(dev, level, true);
1973 max->cur = ilk_cursor_wm_reg_max(dev, level);
1974 max->fbc = ilk_fbc_wm_reg_max(dev);
1977 static bool ilk_validate_wm_level(int level,
1978 const struct ilk_wm_maximums *max,
1979 struct intel_wm_level *result)
1983 /* already determined to be invalid? */
1984 if (!result->enable)
1987 result->enable = result->pri_val <= max->pri &&
1988 result->spr_val <= max->spr &&
1989 result->cur_val <= max->cur;
1991 ret = result->enable;
1994 * HACK until we can pre-compute everything,
1995 * and thus fail gracefully if LP0 watermarks
1998 if (level == 0 && !result->enable) {
1999 if (result->pri_val > max->pri)
2000 DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
2001 level, result->pri_val, max->pri);
2002 if (result->spr_val > max->spr)
2003 DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
2004 level, result->spr_val, max->spr);
2005 if (result->cur_val > max->cur)
2006 DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
2007 level, result->cur_val, max->cur);
2009 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
2010 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
2011 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
2012 result->enable = true;
2018 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
2019 const struct intel_crtc *intel_crtc,
2021 struct intel_crtc_state *cstate,
2022 struct intel_plane_state *pristate,
2023 struct intel_plane_state *sprstate,
2024 struct intel_plane_state *curstate,
2025 struct intel_wm_level *result)
2027 uint16_t pri_latency = dev_priv->wm.pri_latency[level];
2028 uint16_t spr_latency = dev_priv->wm.spr_latency[level];
2029 uint16_t cur_latency = dev_priv->wm.cur_latency[level];
2031 /* WM1+ latency values stored in 0.5us units */
2039 result->pri_val = ilk_compute_pri_wm(cstate, pristate,
2040 pri_latency, level);
2041 result->fbc_val = ilk_compute_fbc_wm(cstate, pristate, result->pri_val);
2045 result->spr_val = ilk_compute_spr_wm(cstate, sprstate, spr_latency);
2048 result->cur_val = ilk_compute_cur_wm(cstate, curstate, cur_latency);
2050 result->enable = true;
2054 hsw_compute_linetime_wm(const struct intel_crtc_state *cstate)
2056 const struct intel_atomic_state *intel_state =
2057 to_intel_atomic_state(cstate->base.state);
2058 const struct drm_display_mode *adjusted_mode =
2059 &cstate->base.adjusted_mode;
2060 u32 linetime, ips_linetime;
2062 if (!cstate->base.active)
2064 if (WARN_ON(adjusted_mode->crtc_clock == 0))
2066 if (WARN_ON(intel_state->cdclk == 0))
2069 /* The WM are computed with base on how long it takes to fill a single
2070 * row at the given clock rate, multiplied by 8.
2072 linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2073 adjusted_mode->crtc_clock);
2074 ips_linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2075 intel_state->cdclk);
2077 return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2078 PIPE_WM_LINETIME_TIME(linetime);
2081 static void intel_read_wm_latency(struct drm_device *dev, uint16_t wm[8])
2083 struct drm_i915_private *dev_priv = to_i915(dev);
2085 if (IS_GEN9(dev_priv)) {
2088 int level, max_level = ilk_wm_max_level(dev_priv);
2090 /* read the first set of memory latencies[0:3] */
2091 val = 0; /* data0 to be programmed to 0 for first set */
2092 mutex_lock(&dev_priv->rps.hw_lock);
2093 ret = sandybridge_pcode_read(dev_priv,
2094 GEN9_PCODE_READ_MEM_LATENCY,
2096 mutex_unlock(&dev_priv->rps.hw_lock);
2099 DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2103 wm[0] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2104 wm[1] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2105 GEN9_MEM_LATENCY_LEVEL_MASK;
2106 wm[2] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2107 GEN9_MEM_LATENCY_LEVEL_MASK;
2108 wm[3] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2109 GEN9_MEM_LATENCY_LEVEL_MASK;
2111 /* read the second set of memory latencies[4:7] */
2112 val = 1; /* data0 to be programmed to 1 for second set */
2113 mutex_lock(&dev_priv->rps.hw_lock);
2114 ret = sandybridge_pcode_read(dev_priv,
2115 GEN9_PCODE_READ_MEM_LATENCY,
2117 mutex_unlock(&dev_priv->rps.hw_lock);
2119 DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2123 wm[4] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2124 wm[5] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2125 GEN9_MEM_LATENCY_LEVEL_MASK;
2126 wm[6] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2127 GEN9_MEM_LATENCY_LEVEL_MASK;
2128 wm[7] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2129 GEN9_MEM_LATENCY_LEVEL_MASK;
2132 * If a level n (n > 1) has a 0us latency, all levels m (m >= n)
2133 * need to be disabled. We make sure to sanitize the values out
2134 * of the punit to satisfy this requirement.
2136 for (level = 1; level <= max_level; level++) {
2137 if (wm[level] == 0) {
2138 for (i = level + 1; i <= max_level; i++)
2145 * WaWmMemoryReadLatency:skl
2147 * punit doesn't take into account the read latency so we need
2148 * to add 2us to the various latency levels we retrieve from the
2149 * punit when level 0 response data us 0us.
2153 for (level = 1; level <= max_level; level++) {
2160 } else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
2161 uint64_t sskpd = I915_READ64(MCH_SSKPD);
2163 wm[0] = (sskpd >> 56) & 0xFF;
2165 wm[0] = sskpd & 0xF;
2166 wm[1] = (sskpd >> 4) & 0xFF;
2167 wm[2] = (sskpd >> 12) & 0xFF;
2168 wm[3] = (sskpd >> 20) & 0x1FF;
2169 wm[4] = (sskpd >> 32) & 0x1FF;
2170 } else if (INTEL_INFO(dev)->gen >= 6) {
2171 uint32_t sskpd = I915_READ(MCH_SSKPD);
2173 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2174 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2175 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2176 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2177 } else if (INTEL_INFO(dev)->gen >= 5) {
2178 uint32_t mltr = I915_READ(MLTR_ILK);
2180 /* ILK primary LP0 latency is 700 ns */
2182 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2183 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2187 static void intel_fixup_spr_wm_latency(struct drm_i915_private *dev_priv,
2190 /* ILK sprite LP0 latency is 1300 ns */
2191 if (IS_GEN5(dev_priv))
2195 static void intel_fixup_cur_wm_latency(struct drm_i915_private *dev_priv,
2198 /* ILK cursor LP0 latency is 1300 ns */
2199 if (IS_GEN5(dev_priv))
2202 /* WaDoubleCursorLP3Latency:ivb */
2203 if (IS_IVYBRIDGE(dev_priv))
2207 int ilk_wm_max_level(const struct drm_i915_private *dev_priv)
2209 /* how many WM levels are we expecting */
2210 if (INTEL_GEN(dev_priv) >= 9)
2212 else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
2214 else if (INTEL_GEN(dev_priv) >= 6)
2220 static void intel_print_wm_latency(struct drm_i915_private *dev_priv,
2222 const uint16_t wm[8])
2224 int level, max_level = ilk_wm_max_level(dev_priv);
2226 for (level = 0; level <= max_level; level++) {
2227 unsigned int latency = wm[level];
2230 DRM_ERROR("%s WM%d latency not provided\n",
2236 * - latencies are in us on gen9.
2237 * - before then, WM1+ latency values are in 0.5us units
2239 if (IS_GEN9(dev_priv))
2244 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2245 name, level, wm[level],
2246 latency / 10, latency % 10);
2250 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
2251 uint16_t wm[5], uint16_t min)
2253 int level, max_level = ilk_wm_max_level(dev_priv);
2258 wm[0] = max(wm[0], min);
2259 for (level = 1; level <= max_level; level++)
2260 wm[level] = max_t(uint16_t, wm[level], DIV_ROUND_UP(min, 5));
2265 static void snb_wm_latency_quirk(struct drm_device *dev)
2267 struct drm_i915_private *dev_priv = to_i915(dev);
2271 * The BIOS provided WM memory latency values are often
2272 * inadequate for high resolution displays. Adjust them.
2274 changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
2275 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
2276 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
2281 DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2282 intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
2283 intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
2284 intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
2287 static void ilk_setup_wm_latency(struct drm_device *dev)
2289 struct drm_i915_private *dev_priv = to_i915(dev);
2291 intel_read_wm_latency(dev, dev_priv->wm.pri_latency);
2293 memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2294 sizeof(dev_priv->wm.pri_latency));
2295 memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2296 sizeof(dev_priv->wm.pri_latency));
2298 intel_fixup_spr_wm_latency(dev_priv, dev_priv->wm.spr_latency);
2299 intel_fixup_cur_wm_latency(dev_priv, dev_priv->wm.cur_latency);
2301 intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
2302 intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
2303 intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
2305 if (IS_GEN6(dev_priv))
2306 snb_wm_latency_quirk(dev);
2309 static void skl_setup_wm_latency(struct drm_device *dev)
2311 struct drm_i915_private *dev_priv = to_i915(dev);
2313 intel_read_wm_latency(dev, dev_priv->wm.skl_latency);
2314 intel_print_wm_latency(dev_priv, "Gen9 Plane", dev_priv->wm.skl_latency);
2317 static bool ilk_validate_pipe_wm(struct drm_device *dev,
2318 struct intel_pipe_wm *pipe_wm)
2320 /* LP0 watermark maximums depend on this pipe alone */
2321 const struct intel_wm_config config = {
2322 .num_pipes_active = 1,
2323 .sprites_enabled = pipe_wm->sprites_enabled,
2324 .sprites_scaled = pipe_wm->sprites_scaled,
2326 struct ilk_wm_maximums max;
2328 /* LP0 watermarks always use 1/2 DDB partitioning */
2329 ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
2331 /* At least LP0 must be valid */
2332 if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0])) {
2333 DRM_DEBUG_KMS("LP0 watermark invalid\n");
2340 /* Compute new watermarks for the pipe */
2341 static int ilk_compute_pipe_wm(struct intel_crtc_state *cstate)
2343 struct drm_atomic_state *state = cstate->base.state;
2344 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
2345 struct intel_pipe_wm *pipe_wm;
2346 struct drm_device *dev = state->dev;
2347 const struct drm_i915_private *dev_priv = to_i915(dev);
2348 struct intel_plane *intel_plane;
2349 struct intel_plane_state *pristate = NULL;
2350 struct intel_plane_state *sprstate = NULL;
2351 struct intel_plane_state *curstate = NULL;
2352 int level, max_level = ilk_wm_max_level(dev_priv), usable_level;
2353 struct ilk_wm_maximums max;
2355 pipe_wm = &cstate->wm.ilk.optimal;
2357 for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
2358 struct intel_plane_state *ps;
2360 ps = intel_atomic_get_existing_plane_state(state,
2365 if (intel_plane->base.type == DRM_PLANE_TYPE_PRIMARY)
2367 else if (intel_plane->base.type == DRM_PLANE_TYPE_OVERLAY)
2369 else if (intel_plane->base.type == DRM_PLANE_TYPE_CURSOR)
2373 pipe_wm->pipe_enabled = cstate->base.active;
2375 pipe_wm->sprites_enabled = sprstate->base.visible;
2376 pipe_wm->sprites_scaled = sprstate->base.visible &&
2377 (drm_rect_width(&sprstate->base.dst) != drm_rect_width(&sprstate->base.src) >> 16 ||
2378 drm_rect_height(&sprstate->base.dst) != drm_rect_height(&sprstate->base.src) >> 16);
2381 usable_level = max_level;
2383 /* ILK/SNB: LP2+ watermarks only w/o sprites */
2384 if (INTEL_INFO(dev)->gen <= 6 && pipe_wm->sprites_enabled)
2387 /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2388 if (pipe_wm->sprites_scaled)
2391 ilk_compute_wm_level(dev_priv, intel_crtc, 0, cstate,
2392 pristate, sprstate, curstate, &pipe_wm->raw_wm[0]);
2394 memset(&pipe_wm->wm, 0, sizeof(pipe_wm->wm));
2395 pipe_wm->wm[0] = pipe_wm->raw_wm[0];
2397 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
2398 pipe_wm->linetime = hsw_compute_linetime_wm(cstate);
2400 if (!ilk_validate_pipe_wm(dev, pipe_wm))
2403 ilk_compute_wm_reg_maximums(dev, 1, &max);
2405 for (level = 1; level <= max_level; level++) {
2406 struct intel_wm_level *wm = &pipe_wm->raw_wm[level];
2408 ilk_compute_wm_level(dev_priv, intel_crtc, level, cstate,
2409 pristate, sprstate, curstate, wm);
2412 * Disable any watermark level that exceeds the
2413 * register maximums since such watermarks are
2416 if (level > usable_level)
2419 if (ilk_validate_wm_level(level, &max, wm))
2420 pipe_wm->wm[level] = *wm;
2422 usable_level = level;
2429 * Build a set of 'intermediate' watermark values that satisfy both the old
2430 * state and the new state. These can be programmed to the hardware
2433 static int ilk_compute_intermediate_wm(struct drm_device *dev,
2434 struct intel_crtc *intel_crtc,
2435 struct intel_crtc_state *newstate)
2437 struct intel_pipe_wm *a = &newstate->wm.ilk.intermediate;
2438 struct intel_pipe_wm *b = &intel_crtc->wm.active.ilk;
2439 int level, max_level = ilk_wm_max_level(to_i915(dev));
2442 * Start with the final, target watermarks, then combine with the
2443 * currently active watermarks to get values that are safe both before
2444 * and after the vblank.
2446 *a = newstate->wm.ilk.optimal;
2447 a->pipe_enabled |= b->pipe_enabled;
2448 a->sprites_enabled |= b->sprites_enabled;
2449 a->sprites_scaled |= b->sprites_scaled;
2451 for (level = 0; level <= max_level; level++) {
2452 struct intel_wm_level *a_wm = &a->wm[level];
2453 const struct intel_wm_level *b_wm = &b->wm[level];
2455 a_wm->enable &= b_wm->enable;
2456 a_wm->pri_val = max(a_wm->pri_val, b_wm->pri_val);
2457 a_wm->spr_val = max(a_wm->spr_val, b_wm->spr_val);
2458 a_wm->cur_val = max(a_wm->cur_val, b_wm->cur_val);
2459 a_wm->fbc_val = max(a_wm->fbc_val, b_wm->fbc_val);
2463 * We need to make sure that these merged watermark values are
2464 * actually a valid configuration themselves. If they're not,
2465 * there's no safe way to transition from the old state to
2466 * the new state, so we need to fail the atomic transaction.
2468 if (!ilk_validate_pipe_wm(dev, a))
2472 * If our intermediate WM are identical to the final WM, then we can
2473 * omit the post-vblank programming; only update if it's different.
2475 if (memcmp(a, &newstate->wm.ilk.optimal, sizeof(*a)) == 0)
2476 newstate->wm.need_postvbl_update = false;
2482 * Merge the watermarks from all active pipes for a specific level.
2484 static void ilk_merge_wm_level(struct drm_device *dev,
2486 struct intel_wm_level *ret_wm)
2488 const struct intel_crtc *intel_crtc;
2490 ret_wm->enable = true;
2492 for_each_intel_crtc(dev, intel_crtc) {
2493 const struct intel_pipe_wm *active = &intel_crtc->wm.active.ilk;
2494 const struct intel_wm_level *wm = &active->wm[level];
2496 if (!active->pipe_enabled)
2500 * The watermark values may have been used in the past,
2501 * so we must maintain them in the registers for some
2502 * time even if the level is now disabled.
2505 ret_wm->enable = false;
2507 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
2508 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
2509 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
2510 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
2515 * Merge all low power watermarks for all active pipes.
2517 static void ilk_wm_merge(struct drm_device *dev,
2518 const struct intel_wm_config *config,
2519 const struct ilk_wm_maximums *max,
2520 struct intel_pipe_wm *merged)
2522 struct drm_i915_private *dev_priv = to_i915(dev);
2523 int level, max_level = ilk_wm_max_level(dev_priv);
2524 int last_enabled_level = max_level;
2526 /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2527 if ((INTEL_GEN(dev_priv) <= 6 || IS_IVYBRIDGE(dev_priv)) &&
2528 config->num_pipes_active > 1)
2529 last_enabled_level = 0;
2531 /* ILK: FBC WM must be disabled always */
2532 merged->fbc_wm_enabled = INTEL_INFO(dev)->gen >= 6;
2534 /* merge each WM1+ level */
2535 for (level = 1; level <= max_level; level++) {
2536 struct intel_wm_level *wm = &merged->wm[level];
2538 ilk_merge_wm_level(dev, level, wm);
2540 if (level > last_enabled_level)
2542 else if (!ilk_validate_wm_level(level, max, wm))
2543 /* make sure all following levels get disabled */
2544 last_enabled_level = level - 1;
2547 * The spec says it is preferred to disable
2548 * FBC WMs instead of disabling a WM level.
2550 if (wm->fbc_val > max->fbc) {
2552 merged->fbc_wm_enabled = false;
2557 /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2559 * FIXME this is racy. FBC might get enabled later.
2560 * What we should check here is whether FBC can be
2561 * enabled sometime later.
2563 if (IS_GEN5(dev_priv) && !merged->fbc_wm_enabled &&
2564 intel_fbc_is_active(dev_priv)) {
2565 for (level = 2; level <= max_level; level++) {
2566 struct intel_wm_level *wm = &merged->wm[level];
2573 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
2575 /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2576 return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
2579 /* The value we need to program into the WM_LPx latency field */
2580 static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
2582 struct drm_i915_private *dev_priv = to_i915(dev);
2584 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
2587 return dev_priv->wm.pri_latency[level];
2590 static void ilk_compute_wm_results(struct drm_device *dev,
2591 const struct intel_pipe_wm *merged,
2592 enum intel_ddb_partitioning partitioning,
2593 struct ilk_wm_values *results)
2595 struct intel_crtc *intel_crtc;
2598 results->enable_fbc_wm = merged->fbc_wm_enabled;
2599 results->partitioning = partitioning;
2601 /* LP1+ register values */
2602 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2603 const struct intel_wm_level *r;
2605 level = ilk_wm_lp_to_level(wm_lp, merged);
2607 r = &merged->wm[level];
2610 * Maintain the watermark values even if the level is
2611 * disabled. Doing otherwise could cause underruns.
2613 results->wm_lp[wm_lp - 1] =
2614 (ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
2615 (r->pri_val << WM1_LP_SR_SHIFT) |
2619 results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
2621 if (INTEL_INFO(dev)->gen >= 8)
2622 results->wm_lp[wm_lp - 1] |=
2623 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
2625 results->wm_lp[wm_lp - 1] |=
2626 r->fbc_val << WM1_LP_FBC_SHIFT;
2629 * Always set WM1S_LP_EN when spr_val != 0, even if the
2630 * level is disabled. Doing otherwise could cause underruns.
2632 if (INTEL_INFO(dev)->gen <= 6 && r->spr_val) {
2633 WARN_ON(wm_lp != 1);
2634 results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
2636 results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2639 /* LP0 register values */
2640 for_each_intel_crtc(dev, intel_crtc) {
2641 enum pipe pipe = intel_crtc->pipe;
2642 const struct intel_wm_level *r =
2643 &intel_crtc->wm.active.ilk.wm[0];
2645 if (WARN_ON(!r->enable))
2648 results->wm_linetime[pipe] = intel_crtc->wm.active.ilk.linetime;
2650 results->wm_pipe[pipe] =
2651 (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
2652 (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
2657 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2658 * case both are at the same level. Prefer r1 in case they're the same. */
2659 static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
2660 struct intel_pipe_wm *r1,
2661 struct intel_pipe_wm *r2)
2663 int level, max_level = ilk_wm_max_level(to_i915(dev));
2664 int level1 = 0, level2 = 0;
2666 for (level = 1; level <= max_level; level++) {
2667 if (r1->wm[level].enable)
2669 if (r2->wm[level].enable)
2673 if (level1 == level2) {
2674 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
2678 } else if (level1 > level2) {
2685 /* dirty bits used to track which watermarks need changes */
2686 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2687 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2688 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2689 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2690 #define WM_DIRTY_FBC (1 << 24)
2691 #define WM_DIRTY_DDB (1 << 25)
2693 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private *dev_priv,
2694 const struct ilk_wm_values *old,
2695 const struct ilk_wm_values *new)
2697 unsigned int dirty = 0;
2701 for_each_pipe(dev_priv, pipe) {
2702 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
2703 dirty |= WM_DIRTY_LINETIME(pipe);
2704 /* Must disable LP1+ watermarks too */
2705 dirty |= WM_DIRTY_LP_ALL;
2708 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
2709 dirty |= WM_DIRTY_PIPE(pipe);
2710 /* Must disable LP1+ watermarks too */
2711 dirty |= WM_DIRTY_LP_ALL;
2715 if (old->enable_fbc_wm != new->enable_fbc_wm) {
2716 dirty |= WM_DIRTY_FBC;
2717 /* Must disable LP1+ watermarks too */
2718 dirty |= WM_DIRTY_LP_ALL;
2721 if (old->partitioning != new->partitioning) {
2722 dirty |= WM_DIRTY_DDB;
2723 /* Must disable LP1+ watermarks too */
2724 dirty |= WM_DIRTY_LP_ALL;
2727 /* LP1+ watermarks already deemed dirty, no need to continue */
2728 if (dirty & WM_DIRTY_LP_ALL)
2731 /* Find the lowest numbered LP1+ watermark in need of an update... */
2732 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2733 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
2734 old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
2738 /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2739 for (; wm_lp <= 3; wm_lp++)
2740 dirty |= WM_DIRTY_LP(wm_lp);
2745 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
2748 struct ilk_wm_values *previous = &dev_priv->wm.hw;
2749 bool changed = false;
2751 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
2752 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
2753 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
2756 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
2757 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
2758 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
2761 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
2762 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
2763 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
2768 * Don't touch WM1S_LP_EN here.
2769 * Doing so could cause underruns.
2776 * The spec says we shouldn't write when we don't need, because every write
2777 * causes WMs to be re-evaluated, expending some power.
2779 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
2780 struct ilk_wm_values *results)
2782 struct drm_device *dev = &dev_priv->drm;
2783 struct ilk_wm_values *previous = &dev_priv->wm.hw;
2787 dirty = ilk_compute_wm_dirty(dev_priv, previous, results);
2791 _ilk_disable_lp_wm(dev_priv, dirty);
2793 if (dirty & WM_DIRTY_PIPE(PIPE_A))
2794 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2795 if (dirty & WM_DIRTY_PIPE(PIPE_B))
2796 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
2797 if (dirty & WM_DIRTY_PIPE(PIPE_C))
2798 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
2800 if (dirty & WM_DIRTY_LINETIME(PIPE_A))
2801 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
2802 if (dirty & WM_DIRTY_LINETIME(PIPE_B))
2803 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
2804 if (dirty & WM_DIRTY_LINETIME(PIPE_C))
2805 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
2807 if (dirty & WM_DIRTY_DDB) {
2808 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
2809 val = I915_READ(WM_MISC);
2810 if (results->partitioning == INTEL_DDB_PART_1_2)
2811 val &= ~WM_MISC_DATA_PARTITION_5_6;
2813 val |= WM_MISC_DATA_PARTITION_5_6;
2814 I915_WRITE(WM_MISC, val);
2816 val = I915_READ(DISP_ARB_CTL2);
2817 if (results->partitioning == INTEL_DDB_PART_1_2)
2818 val &= ~DISP_DATA_PARTITION_5_6;
2820 val |= DISP_DATA_PARTITION_5_6;
2821 I915_WRITE(DISP_ARB_CTL2, val);
2825 if (dirty & WM_DIRTY_FBC) {
2826 val = I915_READ(DISP_ARB_CTL);
2827 if (results->enable_fbc_wm)
2828 val &= ~DISP_FBC_WM_DIS;
2830 val |= DISP_FBC_WM_DIS;
2831 I915_WRITE(DISP_ARB_CTL, val);
2834 if (dirty & WM_DIRTY_LP(1) &&
2835 previous->wm_lp_spr[0] != results->wm_lp_spr[0])
2836 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
2838 if (INTEL_INFO(dev)->gen >= 7) {
2839 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
2840 I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
2841 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
2842 I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
2845 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
2846 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
2847 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
2848 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
2849 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
2850 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
2852 dev_priv->wm.hw = *results;
2855 bool ilk_disable_lp_wm(struct drm_device *dev)
2857 struct drm_i915_private *dev_priv = to_i915(dev);
2859 return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
2862 #define SKL_SAGV_BLOCK_TIME 30 /* µs */
2865 * Return the index of a plane in the SKL DDB and wm result arrays. Primary
2866 * plane is always in slot 0, cursor is always in slot I915_MAX_PLANES-1, and
2867 * other universal planes are in indices 1..n. Note that this may leave unused
2868 * indices between the top "sprite" plane and the cursor.
2871 skl_wm_plane_id(const struct intel_plane *plane)
2873 switch (plane->base.type) {
2874 case DRM_PLANE_TYPE_PRIMARY:
2876 case DRM_PLANE_TYPE_CURSOR:
2877 return PLANE_CURSOR;
2878 case DRM_PLANE_TYPE_OVERLAY:
2879 return plane->plane + 1;
2881 MISSING_CASE(plane->base.type);
2882 return plane->plane;
2887 intel_has_sagv(struct drm_i915_private *dev_priv)
2889 if (IS_KABYLAKE(dev_priv))
2892 if (IS_SKYLAKE(dev_priv) &&
2893 dev_priv->sagv_status != I915_SAGV_NOT_CONTROLLED)
2900 * SAGV dynamically adjusts the system agent voltage and clock frequencies
2901 * depending on power and performance requirements. The display engine access
2902 * to system memory is blocked during the adjustment time. Because of the
2903 * blocking time, having this enabled can cause full system hangs and/or pipe
2904 * underruns if we don't meet all of the following requirements:
2906 * - <= 1 pipe enabled
2907 * - All planes can enable watermarks for latencies >= SAGV engine block time
2908 * - We're not using an interlaced display configuration
2911 intel_enable_sagv(struct drm_i915_private *dev_priv)
2915 if (!intel_has_sagv(dev_priv))
2918 if (dev_priv->sagv_status == I915_SAGV_ENABLED)
2921 DRM_DEBUG_KMS("Enabling the SAGV\n");
2922 mutex_lock(&dev_priv->rps.hw_lock);
2924 ret = sandybridge_pcode_write(dev_priv, GEN9_PCODE_SAGV_CONTROL,
2927 /* We don't need to wait for the SAGV when enabling */
2928 mutex_unlock(&dev_priv->rps.hw_lock);
2931 * Some skl systems, pre-release machines in particular,
2932 * don't actually have an SAGV.
2934 if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
2935 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
2936 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
2938 } else if (ret < 0) {
2939 DRM_ERROR("Failed to enable the SAGV\n");
2943 dev_priv->sagv_status = I915_SAGV_ENABLED;
2948 intel_do_sagv_disable(struct drm_i915_private *dev_priv)
2951 uint32_t temp = GEN9_SAGV_DISABLE;
2953 ret = sandybridge_pcode_read(dev_priv, GEN9_PCODE_SAGV_CONTROL,
2958 return temp & GEN9_SAGV_IS_DISABLED;
2962 intel_disable_sagv(struct drm_i915_private *dev_priv)
2966 if (!intel_has_sagv(dev_priv))
2969 if (dev_priv->sagv_status == I915_SAGV_DISABLED)
2972 DRM_DEBUG_KMS("Disabling the SAGV\n");
2973 mutex_lock(&dev_priv->rps.hw_lock);
2975 /* bspec says to keep retrying for at least 1 ms */
2976 ret = wait_for(result = intel_do_sagv_disable(dev_priv), 1);
2977 mutex_unlock(&dev_priv->rps.hw_lock);
2979 if (ret == -ETIMEDOUT) {
2980 DRM_ERROR("Request to disable SAGV timed out\n");
2985 * Some skl systems, pre-release machines in particular,
2986 * don't actually have an SAGV.
2988 if (IS_SKYLAKE(dev_priv) && result == -ENXIO) {
2989 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
2990 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
2992 } else if (result < 0) {
2993 DRM_ERROR("Failed to disable the SAGV\n");
2997 dev_priv->sagv_status = I915_SAGV_DISABLED;
3001 bool intel_can_enable_sagv(struct drm_atomic_state *state)
3003 struct drm_device *dev = state->dev;
3004 struct drm_i915_private *dev_priv = to_i915(dev);
3005 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3006 struct drm_crtc *crtc;
3010 if (!intel_has_sagv(dev_priv))
3014 * SKL workaround: bspec recommends we disable the SAGV when we have
3015 * more then one pipe enabled
3017 * If there are no active CRTCs, no additional checks need be performed
3019 if (hweight32(intel_state->active_crtcs) == 0)
3021 else if (hweight32(intel_state->active_crtcs) > 1)
3024 /* Since we're now guaranteed to only have one active CRTC... */
3025 pipe = ffs(intel_state->active_crtcs) - 1;
3026 crtc = dev_priv->pipe_to_crtc_mapping[pipe];
3028 if (crtc->state->mode.flags & DRM_MODE_FLAG_INTERLACE)
3031 for_each_plane(dev_priv, pipe, plane) {
3032 /* Skip this plane if it's not enabled */
3033 if (intel_state->wm_results.plane[pipe][plane][0] == 0)
3036 /* Find the highest enabled wm level for this plane */
3037 for (level = ilk_wm_max_level(dev_priv);
3038 intel_state->wm_results.plane[pipe][plane][level] == 0; --level)
3042 * If any of the planes on this pipe don't enable wm levels
3043 * that incur memory latencies higher then 30µs we can't enable
3046 if (dev_priv->wm.skl_latency[level] < SKL_SAGV_BLOCK_TIME)
3054 skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
3055 const struct intel_crtc_state *cstate,
3056 struct skl_ddb_entry *alloc, /* out */
3057 int *num_active /* out */)
3059 struct drm_atomic_state *state = cstate->base.state;
3060 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3061 struct drm_i915_private *dev_priv = to_i915(dev);
3062 struct drm_crtc *for_crtc = cstate->base.crtc;
3063 unsigned int pipe_size, ddb_size;
3064 int nth_active_pipe;
3066 if (WARN_ON(!state) || !cstate->base.active) {
3069 *num_active = hweight32(dev_priv->active_crtcs);
3073 if (intel_state->active_pipe_changes)
3074 *num_active = hweight32(intel_state->active_crtcs);
3076 *num_active = hweight32(dev_priv->active_crtcs);
3078 ddb_size = INTEL_INFO(dev_priv)->ddb_size;
3079 WARN_ON(ddb_size == 0);
3081 ddb_size -= 4; /* 4 blocks for bypass path allocation */
3084 * If the state doesn't change the active CRTC's, then there's
3085 * no need to recalculate; the existing pipe allocation limits
3086 * should remain unchanged. Note that we're safe from racing
3087 * commits since any racing commit that changes the active CRTC
3088 * list would need to grab _all_ crtc locks, including the one
3089 * we currently hold.
3091 if (!intel_state->active_pipe_changes) {
3092 *alloc = to_intel_crtc(for_crtc)->hw_ddb;
3096 nth_active_pipe = hweight32(intel_state->active_crtcs &
3097 (drm_crtc_mask(for_crtc) - 1));
3098 pipe_size = ddb_size / hweight32(intel_state->active_crtcs);
3099 alloc->start = nth_active_pipe * ddb_size / *num_active;
3100 alloc->end = alloc->start + pipe_size;
3103 static unsigned int skl_cursor_allocation(int num_active)
3105 if (num_active == 1)
3111 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry *entry, u32 reg)
3113 entry->start = reg & 0x3ff;
3114 entry->end = (reg >> 16) & 0x3ff;
3119 void skl_ddb_get_hw_state(struct drm_i915_private *dev_priv,
3120 struct skl_ddb_allocation *ddb /* out */)
3126 memset(ddb, 0, sizeof(*ddb));
3128 for_each_pipe(dev_priv, pipe) {
3129 enum intel_display_power_domain power_domain;
3131 power_domain = POWER_DOMAIN_PIPE(pipe);
3132 if (!intel_display_power_get_if_enabled(dev_priv, power_domain))
3135 for_each_plane(dev_priv, pipe, plane) {
3136 val = I915_READ(PLANE_BUF_CFG(pipe, plane));
3137 skl_ddb_entry_init_from_hw(&ddb->plane[pipe][plane],
3141 val = I915_READ(CUR_BUF_CFG(pipe));
3142 skl_ddb_entry_init_from_hw(&ddb->plane[pipe][PLANE_CURSOR],
3145 intel_display_power_put(dev_priv, power_domain);
3150 * Determines the downscale amount of a plane for the purposes of watermark calculations.
3151 * The bspec defines downscale amount as:
3154 * Horizontal down scale amount = maximum[1, Horizontal source size /
3155 * Horizontal destination size]
3156 * Vertical down scale amount = maximum[1, Vertical source size /
3157 * Vertical destination size]
3158 * Total down scale amount = Horizontal down scale amount *
3159 * Vertical down scale amount
3162 * Return value is provided in 16.16 fixed point form to retain fractional part.
3163 * Caller should take care of dividing & rounding off the value.
3166 skl_plane_downscale_amount(const struct intel_plane_state *pstate)
3168 uint32_t downscale_h, downscale_w;
3169 uint32_t src_w, src_h, dst_w, dst_h;
3171 if (WARN_ON(!pstate->base.visible))
3172 return DRM_PLANE_HELPER_NO_SCALING;
3174 /* n.b., src is 16.16 fixed point, dst is whole integer */
3175 src_w = drm_rect_width(&pstate->base.src);
3176 src_h = drm_rect_height(&pstate->base.src);
3177 dst_w = drm_rect_width(&pstate->base.dst);
3178 dst_h = drm_rect_height(&pstate->base.dst);
3179 if (intel_rotation_90_or_270(pstate->base.rotation))
3182 downscale_h = max(src_h / dst_h, (uint32_t)DRM_PLANE_HELPER_NO_SCALING);
3183 downscale_w = max(src_w / dst_w, (uint32_t)DRM_PLANE_HELPER_NO_SCALING);
3185 /* Provide result in 16.16 fixed point */
3186 return (uint64_t)downscale_w * downscale_h >> 16;
3190 skl_plane_relative_data_rate(const struct intel_crtc_state *cstate,
3191 const struct drm_plane_state *pstate,
3194 struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
3195 struct drm_framebuffer *fb = pstate->fb;
3196 uint32_t down_scale_amount, data_rate;
3197 uint32_t width = 0, height = 0;
3198 unsigned format = fb ? fb->pixel_format : DRM_FORMAT_XRGB8888;
3200 if (!intel_pstate->base.visible)
3202 if (pstate->plane->type == DRM_PLANE_TYPE_CURSOR)
3204 if (y && format != DRM_FORMAT_NV12)
3207 width = drm_rect_width(&intel_pstate->base.src) >> 16;
3208 height = drm_rect_height(&intel_pstate->base.src) >> 16;
3210 if (intel_rotation_90_or_270(pstate->rotation))
3211 swap(width, height);
3213 /* for planar format */
3214 if (format == DRM_FORMAT_NV12) {
3215 if (y) /* y-plane data rate */
3216 data_rate = width * height *
3217 drm_format_plane_cpp(format, 0);
3218 else /* uv-plane data rate */
3219 data_rate = (width / 2) * (height / 2) *
3220 drm_format_plane_cpp(format, 1);
3222 /* for packed formats */
3223 data_rate = width * height * drm_format_plane_cpp(format, 0);
3226 down_scale_amount = skl_plane_downscale_amount(intel_pstate);
3228 return (uint64_t)data_rate * down_scale_amount >> 16;
3232 * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
3233 * a 8192x4096@32bpp framebuffer:
3234 * 3 * 4096 * 8192 * 4 < 2^32
3237 skl_get_total_relative_data_rate(struct intel_crtc_state *intel_cstate)
3239 struct drm_crtc_state *cstate = &intel_cstate->base;
3240 struct drm_atomic_state *state = cstate->state;
3241 struct drm_crtc *crtc = cstate->crtc;
3242 struct drm_device *dev = crtc->dev;
3243 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3244 const struct drm_plane *plane;
3245 const struct intel_plane *intel_plane;
3246 struct drm_plane_state *pstate;
3247 unsigned int rate, total_data_rate = 0;
3251 if (WARN_ON(!state))
3254 /* Calculate and cache data rate for each plane */
3255 for_each_plane_in_state(state, plane, pstate, i) {
3256 id = skl_wm_plane_id(to_intel_plane(plane));
3257 intel_plane = to_intel_plane(plane);
3259 if (intel_plane->pipe != intel_crtc->pipe)
3263 rate = skl_plane_relative_data_rate(intel_cstate,
3265 intel_cstate->wm.skl.plane_data_rate[id] = rate;
3268 rate = skl_plane_relative_data_rate(intel_cstate,
3270 intel_cstate->wm.skl.plane_y_data_rate[id] = rate;
3273 /* Calculate CRTC's total data rate from cached values */
3274 for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
3275 int id = skl_wm_plane_id(intel_plane);
3278 total_data_rate += intel_cstate->wm.skl.plane_data_rate[id];
3279 total_data_rate += intel_cstate->wm.skl.plane_y_data_rate[id];
3282 return total_data_rate;
3286 skl_ddb_min_alloc(const struct drm_plane_state *pstate,
3289 struct drm_framebuffer *fb = pstate->fb;
3290 struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
3291 uint32_t src_w, src_h;
3292 uint32_t min_scanlines = 8;
3298 /* For packed formats, no y-plane, return 0 */
3299 if (y && fb->pixel_format != DRM_FORMAT_NV12)
3302 /* For Non Y-tile return 8-blocks */
3303 if (fb->modifier[0] != I915_FORMAT_MOD_Y_TILED &&
3304 fb->modifier[0] != I915_FORMAT_MOD_Yf_TILED)
3307 src_w = drm_rect_width(&intel_pstate->base.src) >> 16;
3308 src_h = drm_rect_height(&intel_pstate->base.src) >> 16;
3310 if (intel_rotation_90_or_270(pstate->rotation))
3313 /* Halve UV plane width and height for NV12 */
3314 if (fb->pixel_format == DRM_FORMAT_NV12 && !y) {
3319 if (fb->pixel_format == DRM_FORMAT_NV12 && !y)
3320 plane_bpp = drm_format_plane_cpp(fb->pixel_format, 1);
3322 plane_bpp = drm_format_plane_cpp(fb->pixel_format, 0);
3324 if (intel_rotation_90_or_270(pstate->rotation)) {
3325 switch (plane_bpp) {
3339 WARN(1, "Unsupported pixel depth %u for rotation",
3345 return DIV_ROUND_UP((4 * src_w * plane_bpp), 512) * min_scanlines/4 + 3;
3349 skl_allocate_pipe_ddb(struct intel_crtc_state *cstate,
3350 struct skl_ddb_allocation *ddb /* out */)
3352 struct drm_atomic_state *state = cstate->base.state;
3353 struct drm_crtc *crtc = cstate->base.crtc;
3354 struct drm_device *dev = crtc->dev;
3355 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3356 struct intel_plane *intel_plane;
3357 struct drm_plane *plane;
3358 struct drm_plane_state *pstate;
3359 enum pipe pipe = intel_crtc->pipe;
3360 struct skl_ddb_entry *alloc = &cstate->wm.skl.ddb;
3361 uint16_t alloc_size, start, cursor_blocks;
3362 uint16_t *minimum = cstate->wm.skl.minimum_blocks;
3363 uint16_t *y_minimum = cstate->wm.skl.minimum_y_blocks;
3364 unsigned int total_data_rate;
3368 /* Clear the partitioning for disabled planes. */
3369 memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
3370 memset(ddb->y_plane[pipe], 0, sizeof(ddb->y_plane[pipe]));
3372 if (WARN_ON(!state))
3375 if (!cstate->base.active) {
3376 alloc->start = alloc->end = 0;
3380 skl_ddb_get_pipe_allocation_limits(dev, cstate, alloc, &num_active);
3381 alloc_size = skl_ddb_entry_size(alloc);
3382 if (alloc_size == 0) {
3383 memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
3387 cursor_blocks = skl_cursor_allocation(num_active);
3388 ddb->plane[pipe][PLANE_CURSOR].start = alloc->end - cursor_blocks;
3389 ddb->plane[pipe][PLANE_CURSOR].end = alloc->end;
3391 alloc_size -= cursor_blocks;
3393 /* 1. Allocate the mininum required blocks for each active plane */
3394 for_each_plane_in_state(state, plane, pstate, i) {
3395 intel_plane = to_intel_plane(plane);
3396 id = skl_wm_plane_id(intel_plane);
3398 if (intel_plane->pipe != pipe)
3401 if (!to_intel_plane_state(pstate)->base.visible) {
3406 if (plane->type == DRM_PLANE_TYPE_CURSOR) {
3412 minimum[id] = skl_ddb_min_alloc(pstate, 0);
3413 y_minimum[id] = skl_ddb_min_alloc(pstate, 1);
3416 for (i = 0; i < PLANE_CURSOR; i++) {
3417 alloc_size -= minimum[i];
3418 alloc_size -= y_minimum[i];
3422 * 2. Distribute the remaining space in proportion to the amount of
3423 * data each plane needs to fetch from memory.
3425 * FIXME: we may not allocate every single block here.
3427 total_data_rate = skl_get_total_relative_data_rate(cstate);
3428 if (total_data_rate == 0)
3431 start = alloc->start;
3432 for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
3433 unsigned int data_rate, y_data_rate;
3434 uint16_t plane_blocks, y_plane_blocks = 0;
3435 int id = skl_wm_plane_id(intel_plane);
3437 data_rate = cstate->wm.skl.plane_data_rate[id];
3440 * allocation for (packed formats) or (uv-plane part of planar format):
3441 * promote the expression to 64 bits to avoid overflowing, the
3442 * result is < available as data_rate / total_data_rate < 1
3444 plane_blocks = minimum[id];
3445 plane_blocks += div_u64((uint64_t)alloc_size * data_rate,
3448 /* Leave disabled planes at (0,0) */
3450 ddb->plane[pipe][id].start = start;
3451 ddb->plane[pipe][id].end = start + plane_blocks;
3454 start += plane_blocks;
3457 * allocation for y_plane part of planar format:
3459 y_data_rate = cstate->wm.skl.plane_y_data_rate[id];
3461 y_plane_blocks = y_minimum[id];
3462 y_plane_blocks += div_u64((uint64_t)alloc_size * y_data_rate,
3466 ddb->y_plane[pipe][id].start = start;
3467 ddb->y_plane[pipe][id].end = start + y_plane_blocks;
3470 start += y_plane_blocks;
3476 static uint32_t skl_pipe_pixel_rate(const struct intel_crtc_state *config)
3478 /* TODO: Take into account the scalers once we support them */
3479 return config->base.adjusted_mode.crtc_clock;
3483 * The max latency should be 257 (max the punit can code is 255 and we add 2us
3484 * for the read latency) and cpp should always be <= 8, so that
3485 * should allow pixel_rate up to ~2 GHz which seems sufficient since max
3486 * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
3488 static uint32_t skl_wm_method1(uint32_t pixel_rate, uint8_t cpp, uint32_t latency)
3490 uint32_t wm_intermediate_val, ret;
3495 wm_intermediate_val = latency * pixel_rate * cpp / 512;
3496 ret = DIV_ROUND_UP(wm_intermediate_val, 1000);
3501 static uint32_t skl_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
3502 uint32_t latency, uint32_t plane_blocks_per_line)
3505 uint32_t wm_intermediate_val;
3510 wm_intermediate_val = latency * pixel_rate;
3511 ret = DIV_ROUND_UP(wm_intermediate_val, pipe_htotal * 1000) *
3512 plane_blocks_per_line;
3517 static uint32_t skl_adjusted_plane_pixel_rate(const struct intel_crtc_state *cstate,
3518 struct intel_plane_state *pstate)
3520 uint64_t adjusted_pixel_rate;
3521 uint64_t downscale_amount;
3522 uint64_t pixel_rate;
3524 /* Shouldn't reach here on disabled planes... */
3525 if (WARN_ON(!pstate->base.visible))
3529 * Adjusted plane pixel rate is just the pipe's adjusted pixel rate
3530 * with additional adjustments for plane-specific scaling.
3532 adjusted_pixel_rate = skl_pipe_pixel_rate(cstate);
3533 downscale_amount = skl_plane_downscale_amount(pstate);
3535 pixel_rate = adjusted_pixel_rate * downscale_amount >> 16;
3536 WARN_ON(pixel_rate != clamp_t(uint32_t, pixel_rate, 0, ~0));
3541 static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv,
3542 struct intel_crtc_state *cstate,
3543 struct intel_plane_state *intel_pstate,
3544 uint16_t ddb_allocation,
3546 uint16_t *out_blocks, /* out */
3547 uint8_t *out_lines, /* out */
3548 bool *enabled /* out */)
3550 struct drm_plane_state *pstate = &intel_pstate->base;
3551 struct drm_framebuffer *fb = pstate->fb;
3552 uint32_t latency = dev_priv->wm.skl_latency[level];
3553 uint32_t method1, method2;
3554 uint32_t plane_bytes_per_line, plane_blocks_per_line;
3555 uint32_t res_blocks, res_lines;
3556 uint32_t selected_result;
3558 uint32_t width = 0, height = 0;
3559 uint32_t plane_pixel_rate;
3560 uint32_t y_tile_minimum, y_min_scanlines;
3562 if (latency == 0 || !cstate->base.active || !intel_pstate->base.visible) {
3567 width = drm_rect_width(&intel_pstate->base.src) >> 16;
3568 height = drm_rect_height(&intel_pstate->base.src) >> 16;
3570 if (intel_rotation_90_or_270(pstate->rotation))
3571 swap(width, height);
3573 cpp = drm_format_plane_cpp(fb->pixel_format, 0);
3574 plane_pixel_rate = skl_adjusted_plane_pixel_rate(cstate, intel_pstate);
3576 if (intel_rotation_90_or_270(pstate->rotation)) {
3577 int cpp = (fb->pixel_format == DRM_FORMAT_NV12) ?
3578 drm_format_plane_cpp(fb->pixel_format, 1) :
3579 drm_format_plane_cpp(fb->pixel_format, 0);
3583 y_min_scanlines = 16;
3586 y_min_scanlines = 8;
3589 y_min_scanlines = 4;
3596 y_min_scanlines = 4;
3599 plane_bytes_per_line = width * cpp;
3600 if (fb->modifier[0] == I915_FORMAT_MOD_Y_TILED ||
3601 fb->modifier[0] == I915_FORMAT_MOD_Yf_TILED) {
3602 plane_blocks_per_line =
3603 DIV_ROUND_UP(plane_bytes_per_line * y_min_scanlines, 512);
3604 plane_blocks_per_line /= y_min_scanlines;
3605 } else if (fb->modifier[0] == DRM_FORMAT_MOD_NONE) {
3606 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512)
3609 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
3612 method1 = skl_wm_method1(plane_pixel_rate, cpp, latency);
3613 method2 = skl_wm_method2(plane_pixel_rate,
3614 cstate->base.adjusted_mode.crtc_htotal,
3616 plane_blocks_per_line);
3618 y_tile_minimum = plane_blocks_per_line * y_min_scanlines;
3620 if (fb->modifier[0] == I915_FORMAT_MOD_Y_TILED ||
3621 fb->modifier[0] == I915_FORMAT_MOD_Yf_TILED) {
3622 selected_result = max(method2, y_tile_minimum);
3624 if ((cpp * cstate->base.adjusted_mode.crtc_htotal / 512 < 1) &&
3625 (plane_bytes_per_line / 512 < 1))
3626 selected_result = method2;
3627 else if ((ddb_allocation / plane_blocks_per_line) >= 1)
3628 selected_result = min(method1, method2);
3630 selected_result = method1;
3633 res_blocks = selected_result + 1;
3634 res_lines = DIV_ROUND_UP(selected_result, plane_blocks_per_line);
3636 if (level >= 1 && level <= 7) {
3637 if (fb->modifier[0] == I915_FORMAT_MOD_Y_TILED ||
3638 fb->modifier[0] == I915_FORMAT_MOD_Yf_TILED) {
3639 res_blocks += y_tile_minimum;
3640 res_lines += y_min_scanlines;
3646 if (res_blocks >= ddb_allocation || res_lines > 31) {
3650 * If there are no valid level 0 watermarks, then we can't
3651 * support this display configuration.
3656 DRM_DEBUG_KMS("Requested display configuration exceeds system watermark limitations\n");
3657 DRM_DEBUG_KMS("Plane %d.%d: blocks required = %u/%u, lines required = %u/31\n",
3658 to_intel_crtc(cstate->base.crtc)->pipe,
3659 skl_wm_plane_id(to_intel_plane(pstate->plane)),
3660 res_blocks, ddb_allocation, res_lines);
3666 *out_blocks = res_blocks;
3667 *out_lines = res_lines;
3674 skl_compute_wm_level(const struct drm_i915_private *dev_priv,
3675 struct skl_ddb_allocation *ddb,
3676 struct intel_crtc_state *cstate,
3677 struct intel_plane *intel_plane,
3679 struct skl_wm_level *result)
3681 struct drm_atomic_state *state = cstate->base.state;
3682 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
3683 struct drm_plane *plane = &intel_plane->base;
3684 struct intel_plane_state *intel_pstate = NULL;
3685 uint16_t ddb_blocks;
3686 enum pipe pipe = intel_crtc->pipe;
3688 int i = skl_wm_plane_id(intel_plane);
3692 intel_atomic_get_existing_plane_state(state,
3696 * Note: If we start supporting multiple pending atomic commits against
3697 * the same planes/CRTC's in the future, plane->state will no longer be
3698 * the correct pre-state to use for the calculations here and we'll
3699 * need to change where we get the 'unchanged' plane data from.
3701 * For now this is fine because we only allow one queued commit against
3702 * a CRTC. Even if the plane isn't modified by this transaction and we
3703 * don't have a plane lock, we still have the CRTC's lock, so we know
3704 * that no other transactions are racing with us to update it.
3707 intel_pstate = to_intel_plane_state(plane->state);
3709 WARN_ON(!intel_pstate->base.fb);
3711 ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][i]);
3713 ret = skl_compute_plane_wm(dev_priv,
3718 &result->plane_res_b,
3719 &result->plane_res_l,
3728 skl_compute_linetime_wm(struct intel_crtc_state *cstate)
3730 if (!cstate->base.active)
3733 if (WARN_ON(skl_pipe_pixel_rate(cstate) == 0))
3736 return DIV_ROUND_UP(8 * cstate->base.adjusted_mode.crtc_htotal * 1000,
3737 skl_pipe_pixel_rate(cstate));
3740 static void skl_compute_transition_wm(struct intel_crtc_state *cstate,
3741 struct skl_wm_level *trans_wm /* out */)
3743 if (!cstate->base.active)
3746 /* Until we know more, just disable transition WMs */
3747 trans_wm->plane_en = false;
3750 static int skl_build_pipe_wm(struct intel_crtc_state *cstate,
3751 struct skl_ddb_allocation *ddb,
3752 struct skl_pipe_wm *pipe_wm)
3754 struct drm_device *dev = cstate->base.crtc->dev;
3755 const struct drm_i915_private *dev_priv = to_i915(dev);
3756 struct intel_plane *intel_plane;
3757 struct skl_plane_wm *wm;
3758 int level, max_level = ilk_wm_max_level(dev_priv);
3762 * We'll only calculate watermarks for planes that are actually
3763 * enabled, so make sure all other planes are set as disabled.
3765 memset(pipe_wm->planes, 0, sizeof(pipe_wm->planes));
3767 for_each_intel_plane_mask(&dev_priv->drm,
3769 cstate->base.plane_mask) {
3770 wm = &pipe_wm->planes[skl_wm_plane_id(intel_plane)];
3772 for (level = 0; level <= max_level; level++) {
3773 ret = skl_compute_wm_level(dev_priv, ddb, cstate,
3779 skl_compute_transition_wm(cstate, &wm->trans_wm);
3781 pipe_wm->linetime = skl_compute_linetime_wm(cstate);
3786 static void skl_compute_wm_results(struct drm_device *dev,
3787 struct skl_pipe_wm *p_wm,
3788 struct skl_wm_values *r,
3789 struct intel_crtc *intel_crtc)
3791 int level, max_level = ilk_wm_max_level(to_i915(dev));
3792 struct skl_plane_wm *plane_wm;
3793 enum pipe pipe = intel_crtc->pipe;
3797 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3798 plane_wm = &p_wm->planes[i];
3800 for (level = 0; level <= max_level; level++) {
3803 temp |= plane_wm->wm[level].plane_res_l <<
3804 PLANE_WM_LINES_SHIFT;
3805 temp |= plane_wm->wm[level].plane_res_b;
3806 if (plane_wm->wm[level].plane_en)
3807 temp |= PLANE_WM_EN;
3809 r->plane[pipe][i][level] = temp;
3813 for (level = 0; level <= max_level; level++) {
3814 plane_wm = &p_wm->planes[PLANE_CURSOR];
3816 temp |= plane_wm->wm[level].plane_res_l << PLANE_WM_LINES_SHIFT;
3817 temp |= plane_wm->wm[level].plane_res_b;
3818 if (plane_wm->wm[level].plane_en)
3819 temp |= PLANE_WM_EN;
3821 r->plane[pipe][PLANE_CURSOR][level] = temp;
3824 /* transition WMs */
3825 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3826 plane_wm = &p_wm->planes[i];
3828 temp |= plane_wm->trans_wm.plane_res_l << PLANE_WM_LINES_SHIFT;
3829 temp |= plane_wm->trans_wm.plane_res_b;
3830 if (plane_wm->trans_wm.plane_en)
3831 temp |= PLANE_WM_EN;
3833 r->plane_trans[pipe][i] = temp;
3836 plane_wm = &p_wm->planes[PLANE_CURSOR];
3838 temp |= plane_wm->trans_wm.plane_res_l << PLANE_WM_LINES_SHIFT;
3839 temp |= plane_wm->trans_wm.plane_res_b;
3840 if (plane_wm->trans_wm.plane_en)
3841 temp |= PLANE_WM_EN;
3843 r->plane_trans[pipe][PLANE_CURSOR] = temp;
3846 static void skl_ddb_entry_write(struct drm_i915_private *dev_priv,
3848 const struct skl_ddb_entry *entry)
3851 I915_WRITE(reg, (entry->end - 1) << 16 | entry->start);
3856 void skl_write_plane_wm(struct intel_crtc *intel_crtc,
3857 const struct skl_wm_values *wm,
3860 struct drm_crtc *crtc = &intel_crtc->base;
3861 struct drm_device *dev = crtc->dev;
3862 struct drm_i915_private *dev_priv = to_i915(dev);
3863 int level, max_level = ilk_wm_max_level(dev_priv);
3864 enum pipe pipe = intel_crtc->pipe;
3866 for (level = 0; level <= max_level; level++) {
3867 I915_WRITE(PLANE_WM(pipe, plane, level),
3868 wm->plane[pipe][plane][level]);
3870 I915_WRITE(PLANE_WM_TRANS(pipe, plane), wm->plane_trans[pipe][plane]);
3872 skl_ddb_entry_write(dev_priv, PLANE_BUF_CFG(pipe, plane),
3873 &wm->ddb.plane[pipe][plane]);
3874 skl_ddb_entry_write(dev_priv, PLANE_NV12_BUF_CFG(pipe, plane),
3875 &wm->ddb.y_plane[pipe][plane]);
3878 void skl_write_cursor_wm(struct intel_crtc *intel_crtc,
3879 const struct skl_wm_values *wm)
3881 struct drm_crtc *crtc = &intel_crtc->base;
3882 struct drm_device *dev = crtc->dev;
3883 struct drm_i915_private *dev_priv = to_i915(dev);
3884 int level, max_level = ilk_wm_max_level(dev_priv);
3885 enum pipe pipe = intel_crtc->pipe;
3887 for (level = 0; level <= max_level; level++) {
3888 I915_WRITE(CUR_WM(pipe, level),
3889 wm->plane[pipe][PLANE_CURSOR][level]);
3891 I915_WRITE(CUR_WM_TRANS(pipe), wm->plane_trans[pipe][PLANE_CURSOR]);
3893 skl_ddb_entry_write(dev_priv, CUR_BUF_CFG(pipe),
3894 &wm->ddb.plane[pipe][PLANE_CURSOR]);
3897 static inline bool skl_ddb_entries_overlap(const struct skl_ddb_entry *a,
3898 const struct skl_ddb_entry *b)
3900 return a->start < b->end && b->start < a->end;
3903 bool skl_ddb_allocation_overlaps(struct drm_atomic_state *state,
3904 struct intel_crtc *intel_crtc)
3906 struct drm_crtc *other_crtc;
3907 struct drm_crtc_state *other_cstate;
3908 struct intel_crtc *other_intel_crtc;
3909 const struct skl_ddb_entry *ddb =
3910 &to_intel_crtc_state(intel_crtc->base.state)->wm.skl.ddb;
3913 for_each_crtc_in_state(state, other_crtc, other_cstate, i) {
3914 other_intel_crtc = to_intel_crtc(other_crtc);
3916 if (other_intel_crtc == intel_crtc)
3919 if (skl_ddb_entries_overlap(ddb, &other_intel_crtc->hw_ddb))
3926 static int skl_update_pipe_wm(struct drm_crtc_state *cstate,
3927 struct skl_ddb_allocation *ddb, /* out */
3928 struct skl_pipe_wm *pipe_wm, /* out */
3929 bool *changed /* out */)
3931 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->crtc);
3932 struct intel_crtc_state *intel_cstate = to_intel_crtc_state(cstate);
3935 ret = skl_build_pipe_wm(intel_cstate, ddb, pipe_wm);
3939 if (!memcmp(&intel_crtc->wm.active.skl, pipe_wm, sizeof(*pipe_wm)))
3948 pipes_modified(struct drm_atomic_state *state)
3950 struct drm_crtc *crtc;
3951 struct drm_crtc_state *cstate;
3952 uint32_t i, ret = 0;
3954 for_each_crtc_in_state(state, crtc, cstate, i)
3955 ret |= drm_crtc_mask(crtc);
3961 skl_ddb_add_affected_planes(struct intel_crtc_state *cstate)
3963 struct drm_atomic_state *state = cstate->base.state;
3964 struct drm_device *dev = state->dev;
3965 struct drm_crtc *crtc = cstate->base.crtc;
3966 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3967 struct drm_i915_private *dev_priv = to_i915(dev);
3968 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3969 struct skl_ddb_allocation *new_ddb = &intel_state->wm_results.ddb;
3970 struct skl_ddb_allocation *cur_ddb = &dev_priv->wm.skl_hw.ddb;
3971 struct drm_plane_state *plane_state;
3972 struct drm_plane *plane;
3973 enum pipe pipe = intel_crtc->pipe;
3976 WARN_ON(!drm_atomic_get_existing_crtc_state(state, crtc));
3978 drm_for_each_plane_mask(plane, dev, crtc->state->plane_mask) {
3979 id = skl_wm_plane_id(to_intel_plane(plane));
3981 if (skl_ddb_entry_equal(&cur_ddb->plane[pipe][id],
3982 &new_ddb->plane[pipe][id]) &&
3983 skl_ddb_entry_equal(&cur_ddb->y_plane[pipe][id],
3984 &new_ddb->y_plane[pipe][id]))
3987 plane_state = drm_atomic_get_plane_state(state, plane);
3988 if (IS_ERR(plane_state))
3989 return PTR_ERR(plane_state);
3996 skl_compute_ddb(struct drm_atomic_state *state)
3998 struct drm_device *dev = state->dev;
3999 struct drm_i915_private *dev_priv = to_i915(dev);
4000 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
4001 struct intel_crtc *intel_crtc;
4002 struct skl_ddb_allocation *ddb = &intel_state->wm_results.ddb;
4003 uint32_t realloc_pipes = pipes_modified(state);
4007 * If this is our first atomic update following hardware readout,
4008 * we can't trust the DDB that the BIOS programmed for us. Let's
4009 * pretend that all pipes switched active status so that we'll
4010 * ensure a full DDB recompute.
4012 if (dev_priv->wm.distrust_bios_wm) {
4013 ret = drm_modeset_lock(&dev->mode_config.connection_mutex,
4014 state->acquire_ctx);
4018 intel_state->active_pipe_changes = ~0;
4021 * We usually only initialize intel_state->active_crtcs if we
4022 * we're doing a modeset; make sure this field is always
4023 * initialized during the sanitization process that happens
4024 * on the first commit too.
4026 if (!intel_state->modeset)
4027 intel_state->active_crtcs = dev_priv->active_crtcs;
4031 * If the modeset changes which CRTC's are active, we need to
4032 * recompute the DDB allocation for *all* active pipes, even
4033 * those that weren't otherwise being modified in any way by this
4034 * atomic commit. Due to the shrinking of the per-pipe allocations
4035 * when new active CRTC's are added, it's possible for a pipe that
4036 * we were already using and aren't changing at all here to suddenly
4037 * become invalid if its DDB needs exceeds its new allocation.
4039 * Note that if we wind up doing a full DDB recompute, we can't let
4040 * any other display updates race with this transaction, so we need
4041 * to grab the lock on *all* CRTC's.
4043 if (intel_state->active_pipe_changes) {
4045 intel_state->wm_results.dirty_pipes = ~0;
4049 * We're not recomputing for the pipes not included in the commit, so
4050 * make sure we start with the current state.
4052 memcpy(ddb, &dev_priv->wm.skl_hw.ddb, sizeof(*ddb));
4054 for_each_intel_crtc_mask(dev, intel_crtc, realloc_pipes) {
4055 struct intel_crtc_state *cstate;
4057 cstate = intel_atomic_get_crtc_state(state, intel_crtc);
4059 return PTR_ERR(cstate);
4061 ret = skl_allocate_pipe_ddb(cstate, ddb);
4065 ret = skl_ddb_add_affected_planes(cstate);
4074 skl_copy_wm_for_pipe(struct skl_wm_values *dst,
4075 struct skl_wm_values *src,
4078 memcpy(dst->plane[pipe], src->plane[pipe],
4079 sizeof(dst->plane[pipe]));
4080 memcpy(dst->plane_trans[pipe], src->plane_trans[pipe],
4081 sizeof(dst->plane_trans[pipe]));
4083 memcpy(dst->ddb.y_plane[pipe], src->ddb.y_plane[pipe],
4084 sizeof(dst->ddb.y_plane[pipe]));
4085 memcpy(dst->ddb.plane[pipe], src->ddb.plane[pipe],
4086 sizeof(dst->ddb.plane[pipe]));
4090 skl_compute_wm(struct drm_atomic_state *state)
4092 struct drm_crtc *crtc;
4093 struct drm_crtc_state *cstate;
4094 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
4095 struct skl_wm_values *results = &intel_state->wm_results;
4096 struct skl_pipe_wm *pipe_wm;
4097 bool changed = false;
4101 * If this transaction isn't actually touching any CRTC's, don't
4102 * bother with watermark calculation. Note that if we pass this
4103 * test, we're guaranteed to hold at least one CRTC state mutex,
4104 * which means we can safely use values like dev_priv->active_crtcs
4105 * since any racing commits that want to update them would need to
4106 * hold _all_ CRTC state mutexes.
4108 for_each_crtc_in_state(state, crtc, cstate, i)
4113 /* Clear all dirty flags */
4114 results->dirty_pipes = 0;
4116 ret = skl_compute_ddb(state);
4121 * Calculate WM's for all pipes that are part of this transaction.
4122 * Note that the DDB allocation above may have added more CRTC's that
4123 * weren't otherwise being modified (and set bits in dirty_pipes) if
4124 * pipe allocations had to change.
4126 * FIXME: Now that we're doing this in the atomic check phase, we
4127 * should allow skl_update_pipe_wm() to return failure in cases where
4128 * no suitable watermark values can be found.
4130 for_each_crtc_in_state(state, crtc, cstate, i) {
4131 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4132 struct intel_crtc_state *intel_cstate =
4133 to_intel_crtc_state(cstate);
4135 pipe_wm = &intel_cstate->wm.skl.optimal;
4136 ret = skl_update_pipe_wm(cstate, &results->ddb, pipe_wm,
4142 results->dirty_pipes |= drm_crtc_mask(crtc);
4144 if ((results->dirty_pipes & drm_crtc_mask(crtc)) == 0)
4145 /* This pipe's WM's did not change */
4148 intel_cstate->update_wm_pre = true;
4149 skl_compute_wm_results(crtc->dev, pipe_wm, results, intel_crtc);
4155 static void skl_update_wm(struct drm_crtc *crtc)
4157 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4158 struct drm_device *dev = crtc->dev;
4159 struct drm_i915_private *dev_priv = to_i915(dev);
4160 struct skl_wm_values *results = &dev_priv->wm.skl_results;
4161 struct skl_wm_values *hw_vals = &dev_priv->wm.skl_hw;
4162 struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
4163 struct skl_pipe_wm *pipe_wm = &cstate->wm.skl.optimal;
4164 enum pipe pipe = intel_crtc->pipe;
4166 if ((results->dirty_pipes & drm_crtc_mask(crtc)) == 0)
4169 intel_crtc->wm.active.skl = *pipe_wm;
4171 mutex_lock(&dev_priv->wm.wm_mutex);
4174 * If this pipe isn't active already, we're going to be enabling it
4175 * very soon. Since it's safe to update a pipe's ddb allocation while
4176 * the pipe's shut off, just do so here. Already active pipes will have
4177 * their watermarks updated once we update their planes.
4179 if (crtc->state->active_changed) {
4182 for (plane = 0; plane < intel_num_planes(intel_crtc); plane++)
4183 skl_write_plane_wm(intel_crtc, results, plane);
4185 skl_write_cursor_wm(intel_crtc, results);
4188 skl_copy_wm_for_pipe(hw_vals, results, pipe);
4190 intel_crtc->hw_ddb = cstate->wm.skl.ddb;
4192 mutex_unlock(&dev_priv->wm.wm_mutex);
4195 static void ilk_compute_wm_config(struct drm_device *dev,
4196 struct intel_wm_config *config)
4198 struct intel_crtc *crtc;
4200 /* Compute the currently _active_ config */
4201 for_each_intel_crtc(dev, crtc) {
4202 const struct intel_pipe_wm *wm = &crtc->wm.active.ilk;
4204 if (!wm->pipe_enabled)
4207 config->sprites_enabled |= wm->sprites_enabled;
4208 config->sprites_scaled |= wm->sprites_scaled;
4209 config->num_pipes_active++;
4213 static void ilk_program_watermarks(struct drm_i915_private *dev_priv)
4215 struct drm_device *dev = &dev_priv->drm;
4216 struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
4217 struct ilk_wm_maximums max;
4218 struct intel_wm_config config = {};
4219 struct ilk_wm_values results = {};
4220 enum intel_ddb_partitioning partitioning;
4222 ilk_compute_wm_config(dev, &config);
4224 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
4225 ilk_wm_merge(dev, &config, &max, &lp_wm_1_2);
4227 /* 5/6 split only in single pipe config on IVB+ */
4228 if (INTEL_INFO(dev)->gen >= 7 &&
4229 config.num_pipes_active == 1 && config.sprites_enabled) {
4230 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
4231 ilk_wm_merge(dev, &config, &max, &lp_wm_5_6);
4233 best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
4235 best_lp_wm = &lp_wm_1_2;
4238 partitioning = (best_lp_wm == &lp_wm_1_2) ?
4239 INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
4241 ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
4243 ilk_write_wm_values(dev_priv, &results);
4246 static void ilk_initial_watermarks(struct intel_crtc_state *cstate)
4248 struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
4249 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4251 mutex_lock(&dev_priv->wm.wm_mutex);
4252 intel_crtc->wm.active.ilk = cstate->wm.ilk.intermediate;
4253 ilk_program_watermarks(dev_priv);
4254 mutex_unlock(&dev_priv->wm.wm_mutex);
4257 static void ilk_optimize_watermarks(struct intel_crtc_state *cstate)
4259 struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
4260 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4262 mutex_lock(&dev_priv->wm.wm_mutex);
4263 if (cstate->wm.need_postvbl_update) {
4264 intel_crtc->wm.active.ilk = cstate->wm.ilk.optimal;
4265 ilk_program_watermarks(dev_priv);
4267 mutex_unlock(&dev_priv->wm.wm_mutex);
4270 static void skl_pipe_wm_active_state(uint32_t val,
4271 struct skl_pipe_wm *active,
4277 bool is_enabled = (val & PLANE_WM_EN) != 0;
4281 active->planes[i].wm[level].plane_en = is_enabled;
4282 active->planes[i].wm[level].plane_res_b =
4283 val & PLANE_WM_BLOCKS_MASK;
4284 active->planes[i].wm[level].plane_res_l =
4285 (val >> PLANE_WM_LINES_SHIFT) &
4286 PLANE_WM_LINES_MASK;
4288 active->planes[PLANE_CURSOR].wm[level].plane_en =
4290 active->planes[PLANE_CURSOR].wm[level].plane_res_b =
4291 val & PLANE_WM_BLOCKS_MASK;
4292 active->planes[PLANE_CURSOR].wm[level].plane_res_l =
4293 (val >> PLANE_WM_LINES_SHIFT) &
4294 PLANE_WM_LINES_MASK;
4298 active->planes[i].trans_wm.plane_en = is_enabled;
4299 active->planes[i].trans_wm.plane_res_b =
4300 val & PLANE_WM_BLOCKS_MASK;
4301 active->planes[i].trans_wm.plane_res_l =
4302 (val >> PLANE_WM_LINES_SHIFT) &
4303 PLANE_WM_LINES_MASK;
4305 active->planes[PLANE_CURSOR].trans_wm.plane_en =
4307 active->planes[PLANE_CURSOR].trans_wm.plane_res_b =
4308 val & PLANE_WM_BLOCKS_MASK;
4309 active->planes[PLANE_CURSOR].trans_wm.plane_res_l =
4310 (val >> PLANE_WM_LINES_SHIFT) &
4311 PLANE_WM_LINES_MASK;
4316 static void skl_pipe_wm_get_hw_state(struct drm_crtc *crtc)
4318 struct drm_device *dev = crtc->dev;
4319 struct drm_i915_private *dev_priv = to_i915(dev);
4320 struct skl_wm_values *hw = &dev_priv->wm.skl_hw;
4321 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4322 struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
4323 struct skl_pipe_wm *active = &cstate->wm.skl.optimal;
4324 enum pipe pipe = intel_crtc->pipe;
4325 int level, i, max_level;
4328 max_level = ilk_wm_max_level(dev_priv);
4330 for (level = 0; level <= max_level; level++) {
4331 for (i = 0; i < intel_num_planes(intel_crtc); i++)
4332 hw->plane[pipe][i][level] =
4333 I915_READ(PLANE_WM(pipe, i, level));
4334 hw->plane[pipe][PLANE_CURSOR][level] = I915_READ(CUR_WM(pipe, level));
4337 for (i = 0; i < intel_num_planes(intel_crtc); i++)
4338 hw->plane_trans[pipe][i] = I915_READ(PLANE_WM_TRANS(pipe, i));
4339 hw->plane_trans[pipe][PLANE_CURSOR] = I915_READ(CUR_WM_TRANS(pipe));
4341 if (!intel_crtc->active)
4344 hw->dirty_pipes |= drm_crtc_mask(crtc);
4346 active->linetime = I915_READ(PIPE_WM_LINETIME(pipe));
4348 for (level = 0; level <= max_level; level++) {
4349 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
4350 temp = hw->plane[pipe][i][level];
4351 skl_pipe_wm_active_state(temp, active, false,
4354 temp = hw->plane[pipe][PLANE_CURSOR][level];
4355 skl_pipe_wm_active_state(temp, active, false, true, i, level);
4358 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
4359 temp = hw->plane_trans[pipe][i];
4360 skl_pipe_wm_active_state(temp, active, true, false, i, 0);
4363 temp = hw->plane_trans[pipe][PLANE_CURSOR];
4364 skl_pipe_wm_active_state(temp, active, true, true, i, 0);
4366 intel_crtc->wm.active.skl = *active;
4369 void skl_wm_get_hw_state(struct drm_device *dev)
4371 struct drm_i915_private *dev_priv = to_i915(dev);
4372 struct skl_ddb_allocation *ddb = &dev_priv->wm.skl_hw.ddb;
4373 struct drm_crtc *crtc;
4375 skl_ddb_get_hw_state(dev_priv, ddb);
4376 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
4377 skl_pipe_wm_get_hw_state(crtc);
4379 if (dev_priv->active_crtcs) {
4380 /* Fully recompute DDB on first atomic commit */
4381 dev_priv->wm.distrust_bios_wm = true;
4383 /* Easy/common case; just sanitize DDB now if everything off */
4384 memset(ddb, 0, sizeof(*ddb));
4388 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
4390 struct drm_device *dev = crtc->dev;
4391 struct drm_i915_private *dev_priv = to_i915(dev);
4392 struct ilk_wm_values *hw = &dev_priv->wm.hw;
4393 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4394 struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
4395 struct intel_pipe_wm *active = &cstate->wm.ilk.optimal;
4396 enum pipe pipe = intel_crtc->pipe;
4397 static const i915_reg_t wm0_pipe_reg[] = {
4398 [PIPE_A] = WM0_PIPEA_ILK,
4399 [PIPE_B] = WM0_PIPEB_ILK,
4400 [PIPE_C] = WM0_PIPEC_IVB,
4403 hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
4404 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
4405 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
4407 memset(active, 0, sizeof(*active));
4409 active->pipe_enabled = intel_crtc->active;
4411 if (active->pipe_enabled) {
4412 u32 tmp = hw->wm_pipe[pipe];
4415 * For active pipes LP0 watermark is marked as
4416 * enabled, and LP1+ watermaks as disabled since
4417 * we can't really reverse compute them in case
4418 * multiple pipes are active.
4420 active->wm[0].enable = true;
4421 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
4422 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
4423 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
4424 active->linetime = hw->wm_linetime[pipe];
4426 int level, max_level = ilk_wm_max_level(dev_priv);
4429 * For inactive pipes, all watermark levels
4430 * should be marked as enabled but zeroed,
4431 * which is what we'd compute them to.
4433 for (level = 0; level <= max_level; level++)
4434 active->wm[level].enable = true;
4437 intel_crtc->wm.active.ilk = *active;
4440 #define _FW_WM(value, plane) \
4441 (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
4442 #define _FW_WM_VLV(value, plane) \
4443 (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
4445 static void vlv_read_wm_values(struct drm_i915_private *dev_priv,
4446 struct vlv_wm_values *wm)
4451 for_each_pipe(dev_priv, pipe) {
4452 tmp = I915_READ(VLV_DDL(pipe));
4454 wm->ddl[pipe].primary =
4455 (tmp >> DDL_PLANE_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4456 wm->ddl[pipe].cursor =
4457 (tmp >> DDL_CURSOR_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4458 wm->ddl[pipe].sprite[0] =
4459 (tmp >> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4460 wm->ddl[pipe].sprite[1] =
4461 (tmp >> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4464 tmp = I915_READ(DSPFW1);
4465 wm->sr.plane = _FW_WM(tmp, SR);
4466 wm->pipe[PIPE_B].cursor = _FW_WM(tmp, CURSORB);
4467 wm->pipe[PIPE_B].primary = _FW_WM_VLV(tmp, PLANEB);
4468 wm->pipe[PIPE_A].primary = _FW_WM_VLV(tmp, PLANEA);
4470 tmp = I915_READ(DSPFW2);
4471 wm->pipe[PIPE_A].sprite[1] = _FW_WM_VLV(tmp, SPRITEB);
4472 wm->pipe[PIPE_A].cursor = _FW_WM(tmp, CURSORA);
4473 wm->pipe[PIPE_A].sprite[0] = _FW_WM_VLV(tmp, SPRITEA);
4475 tmp = I915_READ(DSPFW3);
4476 wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
4478 if (IS_CHERRYVIEW(dev_priv)) {
4479 tmp = I915_READ(DSPFW7_CHV);
4480 wm->pipe[PIPE_B].sprite[1] = _FW_WM_VLV(tmp, SPRITED);
4481 wm->pipe[PIPE_B].sprite[0] = _FW_WM_VLV(tmp, SPRITEC);
4483 tmp = I915_READ(DSPFW8_CHV);
4484 wm->pipe[PIPE_C].sprite[1] = _FW_WM_VLV(tmp, SPRITEF);
4485 wm->pipe[PIPE_C].sprite[0] = _FW_WM_VLV(tmp, SPRITEE);
4487 tmp = I915_READ(DSPFW9_CHV);
4488 wm->pipe[PIPE_C].primary = _FW_WM_VLV(tmp, PLANEC);
4489 wm->pipe[PIPE_C].cursor = _FW_WM(tmp, CURSORC);
4491 tmp = I915_READ(DSPHOWM);
4492 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
4493 wm->pipe[PIPE_C].sprite[1] |= _FW_WM(tmp, SPRITEF_HI) << 8;
4494 wm->pipe[PIPE_C].sprite[0] |= _FW_WM(tmp, SPRITEE_HI) << 8;
4495 wm->pipe[PIPE_C].primary |= _FW_WM(tmp, PLANEC_HI) << 8;
4496 wm->pipe[PIPE_B].sprite[1] |= _FW_WM(tmp, SPRITED_HI) << 8;
4497 wm->pipe[PIPE_B].sprite[0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
4498 wm->pipe[PIPE_B].primary |= _FW_WM(tmp, PLANEB_HI) << 8;
4499 wm->pipe[PIPE_A].sprite[1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
4500 wm->pipe[PIPE_A].sprite[0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
4501 wm->pipe[PIPE_A].primary |= _FW_WM(tmp, PLANEA_HI) << 8;
4503 tmp = I915_READ(DSPFW7);
4504 wm->pipe[PIPE_B].sprite[1] = _FW_WM_VLV(tmp, SPRITED);
4505 wm->pipe[PIPE_B].sprite[0] = _FW_WM_VLV(tmp, SPRITEC);
4507 tmp = I915_READ(DSPHOWM);
4508 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
4509 wm->pipe[PIPE_B].sprite[1] |= _FW_WM(tmp, SPRITED_HI) << 8;
4510 wm->pipe[PIPE_B].sprite[0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
4511 wm->pipe[PIPE_B].primary |= _FW_WM(tmp, PLANEB_HI) << 8;
4512 wm->pipe[PIPE_A].sprite[1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
4513 wm->pipe[PIPE_A].sprite[0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
4514 wm->pipe[PIPE_A].primary |= _FW_WM(tmp, PLANEA_HI) << 8;
4521 void vlv_wm_get_hw_state(struct drm_device *dev)
4523 struct drm_i915_private *dev_priv = to_i915(dev);
4524 struct vlv_wm_values *wm = &dev_priv->wm.vlv;
4525 struct intel_plane *plane;
4529 vlv_read_wm_values(dev_priv, wm);
4531 for_each_intel_plane(dev, plane) {
4532 switch (plane->base.type) {
4534 case DRM_PLANE_TYPE_CURSOR:
4535 plane->wm.fifo_size = 63;
4537 case DRM_PLANE_TYPE_PRIMARY:
4538 plane->wm.fifo_size = vlv_get_fifo_size(dev, plane->pipe, 0);
4540 case DRM_PLANE_TYPE_OVERLAY:
4541 sprite = plane->plane;
4542 plane->wm.fifo_size = vlv_get_fifo_size(dev, plane->pipe, sprite + 1);
4547 wm->cxsr = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
4548 wm->level = VLV_WM_LEVEL_PM2;
4550 if (IS_CHERRYVIEW(dev_priv)) {
4551 mutex_lock(&dev_priv->rps.hw_lock);
4553 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
4554 if (val & DSP_MAXFIFO_PM5_ENABLE)
4555 wm->level = VLV_WM_LEVEL_PM5;
4558 * If DDR DVFS is disabled in the BIOS, Punit
4559 * will never ack the request. So if that happens
4560 * assume we don't have to enable/disable DDR DVFS
4561 * dynamically. To test that just set the REQ_ACK
4562 * bit to poke the Punit, but don't change the
4563 * HIGH/LOW bits so that we don't actually change
4564 * the current state.
4566 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
4567 val |= FORCE_DDR_FREQ_REQ_ACK;
4568 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
4570 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
4571 FORCE_DDR_FREQ_REQ_ACK) == 0, 3)) {
4572 DRM_DEBUG_KMS("Punit not acking DDR DVFS request, "
4573 "assuming DDR DVFS is disabled\n");
4574 dev_priv->wm.max_level = VLV_WM_LEVEL_PM5;
4576 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
4577 if ((val & FORCE_DDR_HIGH_FREQ) == 0)
4578 wm->level = VLV_WM_LEVEL_DDR_DVFS;
4581 mutex_unlock(&dev_priv->rps.hw_lock);
4584 for_each_pipe(dev_priv, pipe)
4585 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
4586 pipe_name(pipe), wm->pipe[pipe].primary, wm->pipe[pipe].cursor,
4587 wm->pipe[pipe].sprite[0], wm->pipe[pipe].sprite[1]);
4589 DRM_DEBUG_KMS("Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
4590 wm->sr.plane, wm->sr.cursor, wm->level, wm->cxsr);
4593 void ilk_wm_get_hw_state(struct drm_device *dev)
4595 struct drm_i915_private *dev_priv = to_i915(dev);
4596 struct ilk_wm_values *hw = &dev_priv->wm.hw;
4597 struct drm_crtc *crtc;
4599 for_each_crtc(dev, crtc)
4600 ilk_pipe_wm_get_hw_state(crtc);
4602 hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
4603 hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
4604 hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
4606 hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
4607 if (INTEL_INFO(dev)->gen >= 7) {
4608 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
4609 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
4612 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
4613 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
4614 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4615 else if (IS_IVYBRIDGE(dev_priv))
4616 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
4617 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4620 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
4624 * intel_update_watermarks - update FIFO watermark values based on current modes
4626 * Calculate watermark values for the various WM regs based on current mode
4627 * and plane configuration.
4629 * There are several cases to deal with here:
4630 * - normal (i.e. non-self-refresh)
4631 * - self-refresh (SR) mode
4632 * - lines are large relative to FIFO size (buffer can hold up to 2)
4633 * - lines are small relative to FIFO size (buffer can hold more than 2
4634 * lines), so need to account for TLB latency
4636 * The normal calculation is:
4637 * watermark = dotclock * bytes per pixel * latency
4638 * where latency is platform & configuration dependent (we assume pessimal
4641 * The SR calculation is:
4642 * watermark = (trunc(latency/line time)+1) * surface width *
4645 * line time = htotal / dotclock
4646 * surface width = hdisplay for normal plane and 64 for cursor
4647 * and latency is assumed to be high, as above.
4649 * The final value programmed to the register should always be rounded up,
4650 * and include an extra 2 entries to account for clock crossings.
4652 * We don't use the sprite, so we can ignore that. And on Crestline we have
4653 * to set the non-SR watermarks to 8.
4655 void intel_update_watermarks(struct drm_crtc *crtc)
4657 struct drm_i915_private *dev_priv = to_i915(crtc->dev);
4659 if (dev_priv->display.update_wm)
4660 dev_priv->display.update_wm(crtc);
4664 * Lock protecting IPS related data structures
4666 DEFINE_SPINLOCK(mchdev_lock);
4668 /* Global for IPS driver to get at the current i915 device. Protected by
4670 static struct drm_i915_private *i915_mch_dev;
4672 bool ironlake_set_drps(struct drm_i915_private *dev_priv, u8 val)
4676 assert_spin_locked(&mchdev_lock);
4678 rgvswctl = I915_READ16(MEMSWCTL);
4679 if (rgvswctl & MEMCTL_CMD_STS) {
4680 DRM_DEBUG("gpu busy, RCS change rejected\n");
4681 return false; /* still busy with another command */
4684 rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
4685 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
4686 I915_WRITE16(MEMSWCTL, rgvswctl);
4687 POSTING_READ16(MEMSWCTL);
4689 rgvswctl |= MEMCTL_CMD_STS;
4690 I915_WRITE16(MEMSWCTL, rgvswctl);
4695 static void ironlake_enable_drps(struct drm_i915_private *dev_priv)
4698 u8 fmax, fmin, fstart, vstart;
4700 spin_lock_irq(&mchdev_lock);
4702 rgvmodectl = I915_READ(MEMMODECTL);
4704 /* Enable temp reporting */
4705 I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
4706 I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
4708 /* 100ms RC evaluation intervals */
4709 I915_WRITE(RCUPEI, 100000);
4710 I915_WRITE(RCDNEI, 100000);
4712 /* Set max/min thresholds to 90ms and 80ms respectively */
4713 I915_WRITE(RCBMAXAVG, 90000);
4714 I915_WRITE(RCBMINAVG, 80000);
4716 I915_WRITE(MEMIHYST, 1);
4718 /* Set up min, max, and cur for interrupt handling */
4719 fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
4720 fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
4721 fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
4722 MEMMODE_FSTART_SHIFT;
4724 vstart = (I915_READ(PXVFREQ(fstart)) & PXVFREQ_PX_MASK) >>
4727 dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
4728 dev_priv->ips.fstart = fstart;
4730 dev_priv->ips.max_delay = fstart;
4731 dev_priv->ips.min_delay = fmin;
4732 dev_priv->ips.cur_delay = fstart;
4734 DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
4735 fmax, fmin, fstart);
4737 I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
4740 * Interrupts will be enabled in ironlake_irq_postinstall
4743 I915_WRITE(VIDSTART, vstart);
4744 POSTING_READ(VIDSTART);
4746 rgvmodectl |= MEMMODE_SWMODE_EN;
4747 I915_WRITE(MEMMODECTL, rgvmodectl);
4749 if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
4750 DRM_ERROR("stuck trying to change perf mode\n");
4753 ironlake_set_drps(dev_priv, fstart);
4755 dev_priv->ips.last_count1 = I915_READ(DMIEC) +
4756 I915_READ(DDREC) + I915_READ(CSIEC);
4757 dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
4758 dev_priv->ips.last_count2 = I915_READ(GFXEC);
4759 dev_priv->ips.last_time2 = ktime_get_raw_ns();
4761 spin_unlock_irq(&mchdev_lock);
4764 static void ironlake_disable_drps(struct drm_i915_private *dev_priv)
4768 spin_lock_irq(&mchdev_lock);
4770 rgvswctl = I915_READ16(MEMSWCTL);
4772 /* Ack interrupts, disable EFC interrupt */
4773 I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
4774 I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
4775 I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
4776 I915_WRITE(DEIIR, DE_PCU_EVENT);
4777 I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
4779 /* Go back to the starting frequency */
4780 ironlake_set_drps(dev_priv, dev_priv->ips.fstart);
4782 rgvswctl |= MEMCTL_CMD_STS;
4783 I915_WRITE(MEMSWCTL, rgvswctl);
4786 spin_unlock_irq(&mchdev_lock);
4789 /* There's a funny hw issue where the hw returns all 0 when reading from
4790 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
4791 * ourselves, instead of doing a rmw cycle (which might result in us clearing
4792 * all limits and the gpu stuck at whatever frequency it is at atm).
4794 static u32 intel_rps_limits(struct drm_i915_private *dev_priv, u8 val)
4798 /* Only set the down limit when we've reached the lowest level to avoid
4799 * getting more interrupts, otherwise leave this clear. This prevents a
4800 * race in the hw when coming out of rc6: There's a tiny window where
4801 * the hw runs at the minimal clock before selecting the desired
4802 * frequency, if the down threshold expires in that window we will not
4803 * receive a down interrupt. */
4804 if (IS_GEN9(dev_priv)) {
4805 limits = (dev_priv->rps.max_freq_softlimit) << 23;
4806 if (val <= dev_priv->rps.min_freq_softlimit)
4807 limits |= (dev_priv->rps.min_freq_softlimit) << 14;
4809 limits = dev_priv->rps.max_freq_softlimit << 24;
4810 if (val <= dev_priv->rps.min_freq_softlimit)
4811 limits |= dev_priv->rps.min_freq_softlimit << 16;
4817 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
4820 u32 threshold_up = 0, threshold_down = 0; /* in % */
4821 u32 ei_up = 0, ei_down = 0;
4823 new_power = dev_priv->rps.power;
4824 switch (dev_priv->rps.power) {
4826 if (val > dev_priv->rps.efficient_freq + 1 &&
4827 val > dev_priv->rps.cur_freq)
4828 new_power = BETWEEN;
4832 if (val <= dev_priv->rps.efficient_freq &&
4833 val < dev_priv->rps.cur_freq)
4834 new_power = LOW_POWER;
4835 else if (val >= dev_priv->rps.rp0_freq &&
4836 val > dev_priv->rps.cur_freq)
4837 new_power = HIGH_POWER;
4841 if (val < (dev_priv->rps.rp1_freq + dev_priv->rps.rp0_freq) >> 1 &&
4842 val < dev_priv->rps.cur_freq)
4843 new_power = BETWEEN;
4846 /* Max/min bins are special */
4847 if (val <= dev_priv->rps.min_freq_softlimit)
4848 new_power = LOW_POWER;
4849 if (val >= dev_priv->rps.max_freq_softlimit)
4850 new_power = HIGH_POWER;
4851 if (new_power == dev_priv->rps.power)
4854 /* Note the units here are not exactly 1us, but 1280ns. */
4855 switch (new_power) {
4857 /* Upclock if more than 95% busy over 16ms */
4861 /* Downclock if less than 85% busy over 32ms */
4863 threshold_down = 85;
4867 /* Upclock if more than 90% busy over 13ms */
4871 /* Downclock if less than 75% busy over 32ms */
4873 threshold_down = 75;
4877 /* Upclock if more than 85% busy over 10ms */
4881 /* Downclock if less than 60% busy over 32ms */
4883 threshold_down = 60;
4887 I915_WRITE(GEN6_RP_UP_EI,
4888 GT_INTERVAL_FROM_US(dev_priv, ei_up));
4889 I915_WRITE(GEN6_RP_UP_THRESHOLD,
4890 GT_INTERVAL_FROM_US(dev_priv,
4891 ei_up * threshold_up / 100));
4893 I915_WRITE(GEN6_RP_DOWN_EI,
4894 GT_INTERVAL_FROM_US(dev_priv, ei_down));
4895 I915_WRITE(GEN6_RP_DOWN_THRESHOLD,
4896 GT_INTERVAL_FROM_US(dev_priv,
4897 ei_down * threshold_down / 100));
4899 I915_WRITE(GEN6_RP_CONTROL,
4900 GEN6_RP_MEDIA_TURBO |
4901 GEN6_RP_MEDIA_HW_NORMAL_MODE |
4902 GEN6_RP_MEDIA_IS_GFX |
4904 GEN6_RP_UP_BUSY_AVG |
4905 GEN6_RP_DOWN_IDLE_AVG);
4907 dev_priv->rps.power = new_power;
4908 dev_priv->rps.up_threshold = threshold_up;
4909 dev_priv->rps.down_threshold = threshold_down;
4910 dev_priv->rps.last_adj = 0;
4913 static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
4917 if (val > dev_priv->rps.min_freq_softlimit)
4918 mask |= GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
4919 if (val < dev_priv->rps.max_freq_softlimit)
4920 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD;
4922 mask &= dev_priv->pm_rps_events;
4924 return gen6_sanitize_rps_pm_mask(dev_priv, ~mask);
4927 /* gen6_set_rps is called to update the frequency request, but should also be
4928 * called when the range (min_delay and max_delay) is modified so that we can
4929 * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
4930 static void gen6_set_rps(struct drm_i915_private *dev_priv, u8 val)
4932 /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
4933 if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1))
4936 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4937 WARN_ON(val > dev_priv->rps.max_freq);
4938 WARN_ON(val < dev_priv->rps.min_freq);
4940 /* min/max delay may still have been modified so be sure to
4941 * write the limits value.
4943 if (val != dev_priv->rps.cur_freq) {
4944 gen6_set_rps_thresholds(dev_priv, val);
4946 if (IS_GEN9(dev_priv))
4947 I915_WRITE(GEN6_RPNSWREQ,
4948 GEN9_FREQUENCY(val));
4949 else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
4950 I915_WRITE(GEN6_RPNSWREQ,
4951 HSW_FREQUENCY(val));
4953 I915_WRITE(GEN6_RPNSWREQ,
4954 GEN6_FREQUENCY(val) |
4956 GEN6_AGGRESSIVE_TURBO);
4959 /* Make sure we continue to get interrupts
4960 * until we hit the minimum or maximum frequencies.
4962 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, intel_rps_limits(dev_priv, val));
4963 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
4965 POSTING_READ(GEN6_RPNSWREQ);
4967 dev_priv->rps.cur_freq = val;
4968 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
4971 static void valleyview_set_rps(struct drm_i915_private *dev_priv, u8 val)
4973 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4974 WARN_ON(val > dev_priv->rps.max_freq);
4975 WARN_ON(val < dev_priv->rps.min_freq);
4977 if (WARN_ONCE(IS_CHERRYVIEW(dev_priv) && (val & 1),
4978 "Odd GPU freq value\n"))
4981 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
4983 if (val != dev_priv->rps.cur_freq) {
4984 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
4985 if (!IS_CHERRYVIEW(dev_priv))
4986 gen6_set_rps_thresholds(dev_priv, val);
4989 dev_priv->rps.cur_freq = val;
4990 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
4993 /* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
4995 * * If Gfx is Idle, then
4996 * 1. Forcewake Media well.
4997 * 2. Request idle freq.
4998 * 3. Release Forcewake of Media well.
5000 static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
5002 u32 val = dev_priv->rps.idle_freq;
5004 if (dev_priv->rps.cur_freq <= val)
5007 /* Wake up the media well, as that takes a lot less
5008 * power than the Render well. */
5009 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_MEDIA);
5010 valleyview_set_rps(dev_priv, val);
5011 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_MEDIA);
5014 void gen6_rps_busy(struct drm_i915_private *dev_priv)
5016 mutex_lock(&dev_priv->rps.hw_lock);
5017 if (dev_priv->rps.enabled) {
5018 if (dev_priv->pm_rps_events & (GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_UP_EI_EXPIRED))
5019 gen6_rps_reset_ei(dev_priv);
5020 I915_WRITE(GEN6_PMINTRMSK,
5021 gen6_rps_pm_mask(dev_priv, dev_priv->rps.cur_freq));
5023 gen6_enable_rps_interrupts(dev_priv);
5025 /* Ensure we start at the user's desired frequency */
5026 intel_set_rps(dev_priv,
5027 clamp(dev_priv->rps.cur_freq,
5028 dev_priv->rps.min_freq_softlimit,
5029 dev_priv->rps.max_freq_softlimit));
5031 mutex_unlock(&dev_priv->rps.hw_lock);
5034 void gen6_rps_idle(struct drm_i915_private *dev_priv)
5036 /* Flush our bottom-half so that it does not race with us
5037 * setting the idle frequency and so that it is bounded by
5038 * our rpm wakeref. And then disable the interrupts to stop any
5039 * futher RPS reclocking whilst we are asleep.
5041 gen6_disable_rps_interrupts(dev_priv);
5043 mutex_lock(&dev_priv->rps.hw_lock);
5044 if (dev_priv->rps.enabled) {
5045 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
5046 vlv_set_rps_idle(dev_priv);
5048 gen6_set_rps(dev_priv, dev_priv->rps.idle_freq);
5049 dev_priv->rps.last_adj = 0;
5050 I915_WRITE(GEN6_PMINTRMSK,
5051 gen6_sanitize_rps_pm_mask(dev_priv, ~0));
5053 mutex_unlock(&dev_priv->rps.hw_lock);
5055 spin_lock(&dev_priv->rps.client_lock);
5056 while (!list_empty(&dev_priv->rps.clients))
5057 list_del_init(dev_priv->rps.clients.next);
5058 spin_unlock(&dev_priv->rps.client_lock);
5061 void gen6_rps_boost(struct drm_i915_private *dev_priv,
5062 struct intel_rps_client *rps,
5063 unsigned long submitted)
5065 /* This is intentionally racy! We peek at the state here, then
5066 * validate inside the RPS worker.
5068 if (!(dev_priv->gt.awake &&
5069 dev_priv->rps.enabled &&
5070 dev_priv->rps.cur_freq < dev_priv->rps.boost_freq))
5073 /* Force a RPS boost (and don't count it against the client) if
5074 * the GPU is severely congested.
5076 if (rps && time_after(jiffies, submitted + DRM_I915_THROTTLE_JIFFIES))
5079 spin_lock(&dev_priv->rps.client_lock);
5080 if (rps == NULL || list_empty(&rps->link)) {
5081 spin_lock_irq(&dev_priv->irq_lock);
5082 if (dev_priv->rps.interrupts_enabled) {
5083 dev_priv->rps.client_boost = true;
5084 schedule_work(&dev_priv->rps.work);
5086 spin_unlock_irq(&dev_priv->irq_lock);
5089 list_add(&rps->link, &dev_priv->rps.clients);
5092 dev_priv->rps.boosts++;
5094 spin_unlock(&dev_priv->rps.client_lock);
5097 void intel_set_rps(struct drm_i915_private *dev_priv, u8 val)
5099 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
5100 valleyview_set_rps(dev_priv, val);
5102 gen6_set_rps(dev_priv, val);
5105 static void gen9_disable_rc6(struct drm_i915_private *dev_priv)
5107 I915_WRITE(GEN6_RC_CONTROL, 0);
5108 I915_WRITE(GEN9_PG_ENABLE, 0);
5111 static void gen9_disable_rps(struct drm_i915_private *dev_priv)
5113 I915_WRITE(GEN6_RP_CONTROL, 0);
5116 static void gen6_disable_rps(struct drm_i915_private *dev_priv)
5118 I915_WRITE(GEN6_RC_CONTROL, 0);
5119 I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
5120 I915_WRITE(GEN6_RP_CONTROL, 0);
5123 static void cherryview_disable_rps(struct drm_i915_private *dev_priv)
5125 I915_WRITE(GEN6_RC_CONTROL, 0);
5128 static void valleyview_disable_rps(struct drm_i915_private *dev_priv)
5130 /* we're doing forcewake before Disabling RC6,
5131 * This what the BIOS expects when going into suspend */
5132 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5134 I915_WRITE(GEN6_RC_CONTROL, 0);
5136 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5139 static void intel_print_rc6_info(struct drm_i915_private *dev_priv, u32 mode)
5141 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
5142 if (mode & (GEN7_RC_CTL_TO_MODE | GEN6_RC_CTL_EI_MODE(1)))
5143 mode = GEN6_RC_CTL_RC6_ENABLE;
5147 if (HAS_RC6p(dev_priv))
5148 DRM_DEBUG_DRIVER("Enabling RC6 states: "
5149 "RC6 %s RC6p %s RC6pp %s\n",
5150 onoff(mode & GEN6_RC_CTL_RC6_ENABLE),
5151 onoff(mode & GEN6_RC_CTL_RC6p_ENABLE),
5152 onoff(mode & GEN6_RC_CTL_RC6pp_ENABLE));
5155 DRM_DEBUG_DRIVER("Enabling RC6 states: RC6 %s\n",
5156 onoff(mode & GEN6_RC_CTL_RC6_ENABLE));
5159 static bool bxt_check_bios_rc6_setup(struct drm_i915_private *dev_priv)
5161 struct i915_ggtt *ggtt = &dev_priv->ggtt;
5162 bool enable_rc6 = true;
5163 unsigned long rc6_ctx_base;
5167 rc_ctl = I915_READ(GEN6_RC_CONTROL);
5168 rc_sw_target = (I915_READ(GEN6_RC_STATE) & RC_SW_TARGET_STATE_MASK) >>
5169 RC_SW_TARGET_STATE_SHIFT;
5170 DRM_DEBUG_DRIVER("BIOS enabled RC states: "
5171 "HW_CTRL %s HW_RC6 %s SW_TARGET_STATE %x\n",
5172 onoff(rc_ctl & GEN6_RC_CTL_HW_ENABLE),
5173 onoff(rc_ctl & GEN6_RC_CTL_RC6_ENABLE),
5176 if (!(I915_READ(RC6_LOCATION) & RC6_CTX_IN_DRAM)) {
5177 DRM_DEBUG_DRIVER("RC6 Base location not set properly.\n");
5182 * The exact context size is not known for BXT, so assume a page size
5185 rc6_ctx_base = I915_READ(RC6_CTX_BASE) & RC6_CTX_BASE_MASK;
5186 if (!((rc6_ctx_base >= ggtt->stolen_reserved_base) &&
5187 (rc6_ctx_base + PAGE_SIZE <= ggtt->stolen_reserved_base +
5188 ggtt->stolen_reserved_size))) {
5189 DRM_DEBUG_DRIVER("RC6 Base address not as expected.\n");
5193 if (!(((I915_READ(PWRCTX_MAXCNT_RCSUNIT) & IDLE_TIME_MASK) > 1) &&
5194 ((I915_READ(PWRCTX_MAXCNT_VCSUNIT0) & IDLE_TIME_MASK) > 1) &&
5195 ((I915_READ(PWRCTX_MAXCNT_BCSUNIT) & IDLE_TIME_MASK) > 1) &&
5196 ((I915_READ(PWRCTX_MAXCNT_VECSUNIT) & IDLE_TIME_MASK) > 1))) {
5197 DRM_DEBUG_DRIVER("Engine Idle wait time not set properly.\n");
5201 if (!I915_READ(GEN8_PUSHBUS_CONTROL) ||
5202 !I915_READ(GEN8_PUSHBUS_ENABLE) ||
5203 !I915_READ(GEN8_PUSHBUS_SHIFT)) {
5204 DRM_DEBUG_DRIVER("Pushbus not setup properly.\n");
5208 if (!I915_READ(GEN6_GFXPAUSE)) {
5209 DRM_DEBUG_DRIVER("GFX pause not setup properly.\n");
5213 if (!I915_READ(GEN8_MISC_CTRL0)) {
5214 DRM_DEBUG_DRIVER("GPM control not setup properly.\n");
5221 int sanitize_rc6_option(struct drm_i915_private *dev_priv, int enable_rc6)
5223 /* No RC6 before Ironlake and code is gone for ilk. */
5224 if (INTEL_INFO(dev_priv)->gen < 6)
5230 if (IS_BROXTON(dev_priv) && !bxt_check_bios_rc6_setup(dev_priv)) {
5231 DRM_INFO("RC6 disabled by BIOS\n");
5235 /* Respect the kernel parameter if it is set */
5236 if (enable_rc6 >= 0) {
5239 if (HAS_RC6p(dev_priv))
5240 mask = INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE |
5243 mask = INTEL_RC6_ENABLE;
5245 if ((enable_rc6 & mask) != enable_rc6)
5246 DRM_DEBUG_DRIVER("Adjusting RC6 mask to %d "
5247 "(requested %d, valid %d)\n",
5248 enable_rc6 & mask, enable_rc6, mask);
5250 return enable_rc6 & mask;
5253 if (IS_IVYBRIDGE(dev_priv))
5254 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
5256 return INTEL_RC6_ENABLE;
5259 static void gen6_init_rps_frequencies(struct drm_i915_private *dev_priv)
5261 /* All of these values are in units of 50MHz */
5263 /* static values from HW: RP0 > RP1 > RPn (min_freq) */
5264 if (IS_BROXTON(dev_priv)) {
5265 u32 rp_state_cap = I915_READ(BXT_RP_STATE_CAP);
5266 dev_priv->rps.rp0_freq = (rp_state_cap >> 16) & 0xff;
5267 dev_priv->rps.rp1_freq = (rp_state_cap >> 8) & 0xff;
5268 dev_priv->rps.min_freq = (rp_state_cap >> 0) & 0xff;
5270 u32 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
5271 dev_priv->rps.rp0_freq = (rp_state_cap >> 0) & 0xff;
5272 dev_priv->rps.rp1_freq = (rp_state_cap >> 8) & 0xff;
5273 dev_priv->rps.min_freq = (rp_state_cap >> 16) & 0xff;
5275 /* hw_max = RP0 until we check for overclocking */
5276 dev_priv->rps.max_freq = dev_priv->rps.rp0_freq;
5278 dev_priv->rps.efficient_freq = dev_priv->rps.rp1_freq;
5279 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv) ||
5280 IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5281 u32 ddcc_status = 0;
5283 if (sandybridge_pcode_read(dev_priv,
5284 HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL,
5286 dev_priv->rps.efficient_freq =
5288 ((ddcc_status >> 8) & 0xff),
5289 dev_priv->rps.min_freq,
5290 dev_priv->rps.max_freq);
5293 if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5294 /* Store the frequency values in 16.66 MHZ units, which is
5295 * the natural hardware unit for SKL
5297 dev_priv->rps.rp0_freq *= GEN9_FREQ_SCALER;
5298 dev_priv->rps.rp1_freq *= GEN9_FREQ_SCALER;
5299 dev_priv->rps.min_freq *= GEN9_FREQ_SCALER;
5300 dev_priv->rps.max_freq *= GEN9_FREQ_SCALER;
5301 dev_priv->rps.efficient_freq *= GEN9_FREQ_SCALER;
5305 static void reset_rps(struct drm_i915_private *dev_priv,
5306 void (*set)(struct drm_i915_private *, u8))
5308 u8 freq = dev_priv->rps.cur_freq;
5311 dev_priv->rps.power = -1;
5312 dev_priv->rps.cur_freq = -1;
5314 set(dev_priv, freq);
5317 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
5318 static void gen9_enable_rps(struct drm_i915_private *dev_priv)
5320 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5322 /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
5323 if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) {
5325 * BIOS could leave the Hw Turbo enabled, so need to explicitly
5326 * clear out the Control register just to avoid inconsitency
5327 * with debugfs interface, which will show Turbo as enabled
5328 * only and that is not expected by the User after adding the
5329 * WaGsvDisableTurbo. Apart from this there is no problem even
5330 * if the Turbo is left enabled in the Control register, as the
5331 * Up/Down interrupts would remain masked.
5333 gen9_disable_rps(dev_priv);
5334 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5338 /* Program defaults and thresholds for RPS*/
5339 I915_WRITE(GEN6_RC_VIDEO_FREQ,
5340 GEN9_FREQUENCY(dev_priv->rps.rp1_freq));
5342 /* 1 second timeout*/
5343 I915_WRITE(GEN6_RP_DOWN_TIMEOUT,
5344 GT_INTERVAL_FROM_US(dev_priv, 1000000));
5346 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 0xa);
5348 /* Leaning on the below call to gen6_set_rps to program/setup the
5349 * Up/Down EI & threshold registers, as well as the RP_CONTROL,
5350 * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
5351 reset_rps(dev_priv, gen6_set_rps);
5353 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5356 static void gen9_enable_rc6(struct drm_i915_private *dev_priv)
5358 struct intel_engine_cs *engine;
5359 enum intel_engine_id id;
5360 uint32_t rc6_mask = 0;
5362 /* 1a: Software RC state - RC0 */
5363 I915_WRITE(GEN6_RC_STATE, 0);
5365 /* 1b: Get forcewake during program sequence. Although the driver
5366 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5367 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5369 /* 2a: Disable RC states. */
5370 I915_WRITE(GEN6_RC_CONTROL, 0);
5372 /* 2b: Program RC6 thresholds.*/
5374 /* WaRsDoubleRc6WrlWithCoarsePowerGating: Doubling WRL only when CPG is enabled */
5375 if (IS_SKYLAKE(dev_priv))
5376 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 108 << 16);
5378 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16);
5379 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5380 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
5381 for_each_engine(engine, dev_priv, id)
5382 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5384 if (HAS_GUC(dev_priv))
5385 I915_WRITE(GUC_MAX_IDLE_COUNT, 0xA);
5387 I915_WRITE(GEN6_RC_SLEEP, 0);
5389 /* 2c: Program Coarse Power Gating Policies. */
5390 I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS, 25);
5391 I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS, 25);
5393 /* 3a: Enable RC6 */
5394 if (intel_enable_rc6() & INTEL_RC6_ENABLE)
5395 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
5396 DRM_INFO("RC6 %s\n", onoff(rc6_mask & GEN6_RC_CTL_RC6_ENABLE));
5397 /* WaRsUseTimeoutMode:bxt */
5398 if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) {
5399 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us */
5400 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5401 GEN7_RC_CTL_TO_MODE |
5404 I915_WRITE(GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */
5405 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5406 GEN6_RC_CTL_EI_MODE(1) |
5411 * 3b: Enable Coarse Power Gating only when RC6 is enabled.
5412 * WaRsDisableCoarsePowerGating:skl,bxt - Render/Media PG need to be disabled with RC6.
5414 if (NEEDS_WaRsDisableCoarsePowerGating(dev_priv))
5415 I915_WRITE(GEN9_PG_ENABLE, 0);
5417 I915_WRITE(GEN9_PG_ENABLE, (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ?
5418 (GEN9_RENDER_PG_ENABLE | GEN9_MEDIA_PG_ENABLE) : 0);
5420 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5423 static void gen8_enable_rps(struct drm_i915_private *dev_priv)
5425 struct intel_engine_cs *engine;
5426 enum intel_engine_id id;
5427 uint32_t rc6_mask = 0;
5429 /* 1a: Software RC state - RC0 */
5430 I915_WRITE(GEN6_RC_STATE, 0);
5432 /* 1c & 1d: Get forcewake during program sequence. Although the driver
5433 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5434 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5436 /* 2a: Disable RC states. */
5437 I915_WRITE(GEN6_RC_CONTROL, 0);
5439 /* 2b: Program RC6 thresholds.*/
5440 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
5441 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5442 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
5443 for_each_engine(engine, dev_priv, id)
5444 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5445 I915_WRITE(GEN6_RC_SLEEP, 0);
5446 if (IS_BROADWELL(dev_priv))
5447 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */
5449 I915_WRITE(GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
5452 if (intel_enable_rc6() & INTEL_RC6_ENABLE)
5453 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
5454 intel_print_rc6_info(dev_priv, rc6_mask);
5455 if (IS_BROADWELL(dev_priv))
5456 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5457 GEN7_RC_CTL_TO_MODE |
5460 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5461 GEN6_RC_CTL_EI_MODE(1) |
5464 /* 4 Program defaults and thresholds for RPS*/
5465 I915_WRITE(GEN6_RPNSWREQ,
5466 HSW_FREQUENCY(dev_priv->rps.rp1_freq));
5467 I915_WRITE(GEN6_RC_VIDEO_FREQ,
5468 HSW_FREQUENCY(dev_priv->rps.rp1_freq));
5469 /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
5470 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
5472 /* Docs recommend 900MHz, and 300 MHz respectively */
5473 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
5474 dev_priv->rps.max_freq_softlimit << 24 |
5475 dev_priv->rps.min_freq_softlimit << 16);
5477 I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
5478 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
5479 I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
5480 I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
5482 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5485 I915_WRITE(GEN6_RP_CONTROL,
5486 GEN6_RP_MEDIA_TURBO |
5487 GEN6_RP_MEDIA_HW_NORMAL_MODE |
5488 GEN6_RP_MEDIA_IS_GFX |
5490 GEN6_RP_UP_BUSY_AVG |
5491 GEN6_RP_DOWN_IDLE_AVG);
5493 /* 6: Ring frequency + overclocking (our driver does this later */
5495 reset_rps(dev_priv, gen6_set_rps);
5497 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5500 static void gen6_enable_rps(struct drm_i915_private *dev_priv)
5502 struct intel_engine_cs *engine;
5503 enum intel_engine_id id;
5504 u32 rc6vids, rc6_mask = 0;
5509 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5511 /* Here begins a magic sequence of register writes to enable
5512 * auto-downclocking.
5514 * Perhaps there might be some value in exposing these to
5517 I915_WRITE(GEN6_RC_STATE, 0);
5519 /* Clear the DBG now so we don't confuse earlier errors */
5520 gtfifodbg = I915_READ(GTFIFODBG);
5522 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
5523 I915_WRITE(GTFIFODBG, gtfifodbg);
5526 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5528 /* disable the counters and set deterministic thresholds */
5529 I915_WRITE(GEN6_RC_CONTROL, 0);
5531 I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
5532 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
5533 I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
5534 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
5535 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
5537 for_each_engine(engine, dev_priv, id)
5538 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5540 I915_WRITE(GEN6_RC_SLEEP, 0);
5541 I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
5542 if (IS_IVYBRIDGE(dev_priv))
5543 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
5545 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
5546 I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
5547 I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
5549 /* Check if we are enabling RC6 */
5550 rc6_mode = intel_enable_rc6();
5551 if (rc6_mode & INTEL_RC6_ENABLE)
5552 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
5554 /* We don't use those on Haswell */
5555 if (!IS_HASWELL(dev_priv)) {
5556 if (rc6_mode & INTEL_RC6p_ENABLE)
5557 rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
5559 if (rc6_mode & INTEL_RC6pp_ENABLE)
5560 rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
5563 intel_print_rc6_info(dev_priv, rc6_mask);
5565 I915_WRITE(GEN6_RC_CONTROL,
5567 GEN6_RC_CTL_EI_MODE(1) |
5568 GEN6_RC_CTL_HW_ENABLE);
5570 /* Power down if completely idle for over 50ms */
5571 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
5572 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5574 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_MIN_FREQ_TABLE, 0);
5576 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
5578 reset_rps(dev_priv, gen6_set_rps);
5581 ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
5582 if (IS_GEN6(dev_priv) && ret) {
5583 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
5584 } else if (IS_GEN6(dev_priv) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
5585 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
5586 GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
5587 rc6vids &= 0xffff00;
5588 rc6vids |= GEN6_ENCODE_RC6_VID(450);
5589 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
5591 DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
5594 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5597 static void gen6_update_ring_freq(struct drm_i915_private *dev_priv)
5600 unsigned int gpu_freq;
5601 unsigned int max_ia_freq, min_ring_freq;
5602 unsigned int max_gpu_freq, min_gpu_freq;
5603 int scaling_factor = 180;
5604 struct cpufreq_policy *policy;
5606 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5608 policy = cpufreq_cpu_get(0);
5610 max_ia_freq = policy->cpuinfo.max_freq;
5611 cpufreq_cpu_put(policy);
5614 * Default to measured freq if none found, PCU will ensure we
5617 max_ia_freq = tsc_khz;
5620 /* Convert from kHz to MHz */
5621 max_ia_freq /= 1000;
5623 min_ring_freq = I915_READ(DCLK) & 0xf;
5624 /* convert DDR frequency from units of 266.6MHz to bandwidth */
5625 min_ring_freq = mult_frac(min_ring_freq, 8, 3);
5627 if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5628 /* Convert GT frequency to 50 HZ units */
5629 min_gpu_freq = dev_priv->rps.min_freq / GEN9_FREQ_SCALER;
5630 max_gpu_freq = dev_priv->rps.max_freq / GEN9_FREQ_SCALER;
5632 min_gpu_freq = dev_priv->rps.min_freq;
5633 max_gpu_freq = dev_priv->rps.max_freq;
5637 * For each potential GPU frequency, load a ring frequency we'd like
5638 * to use for memory access. We do this by specifying the IA frequency
5639 * the PCU should use as a reference to determine the ring frequency.
5641 for (gpu_freq = max_gpu_freq; gpu_freq >= min_gpu_freq; gpu_freq--) {
5642 int diff = max_gpu_freq - gpu_freq;
5643 unsigned int ia_freq = 0, ring_freq = 0;
5645 if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5647 * ring_freq = 2 * GT. ring_freq is in 100MHz units
5648 * No floor required for ring frequency on SKL.
5650 ring_freq = gpu_freq;
5651 } else if (INTEL_INFO(dev_priv)->gen >= 8) {
5652 /* max(2 * GT, DDR). NB: GT is 50MHz units */
5653 ring_freq = max(min_ring_freq, gpu_freq);
5654 } else if (IS_HASWELL(dev_priv)) {
5655 ring_freq = mult_frac(gpu_freq, 5, 4);
5656 ring_freq = max(min_ring_freq, ring_freq);
5657 /* leave ia_freq as the default, chosen by cpufreq */
5659 /* On older processors, there is no separate ring
5660 * clock domain, so in order to boost the bandwidth
5661 * of the ring, we need to upclock the CPU (ia_freq).
5663 * For GPU frequencies less than 750MHz,
5664 * just use the lowest ring freq.
5666 if (gpu_freq < min_freq)
5669 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
5670 ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
5673 sandybridge_pcode_write(dev_priv,
5674 GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
5675 ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
5676 ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
5681 static int cherryview_rps_max_freq(struct drm_i915_private *dev_priv)
5685 val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
5687 switch (INTEL_INFO(dev_priv)->sseu.eu_total) {
5689 /* (2 * 4) config */
5690 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT);
5693 /* (2 * 6) config */
5694 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT);
5697 /* (2 * 8) config */
5699 /* Setting (2 * 8) Min RP0 for any other combination */
5700 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT);
5704 rp0 = (rp0 & FB_GFX_FREQ_FUSE_MASK);
5709 static int cherryview_rps_rpe_freq(struct drm_i915_private *dev_priv)
5713 val = vlv_punit_read(dev_priv, PUNIT_GPU_DUTYCYCLE_REG);
5714 rpe = (val >> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK;
5719 static int cherryview_rps_guar_freq(struct drm_i915_private *dev_priv)
5723 val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
5724 rp1 = (val & FB_GFX_FREQ_FUSE_MASK);
5729 static int valleyview_rps_guar_freq(struct drm_i915_private *dev_priv)
5733 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
5735 rp1 = (val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT;
5740 static int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
5744 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
5746 rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
5748 rp0 = min_t(u32, rp0, 0xea);
5753 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
5757 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
5758 rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
5759 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
5760 rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
5765 static int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
5769 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
5771 * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
5772 * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
5773 * a BYT-M B0 the above register contains 0xbf. Moreover when setting
5774 * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
5775 * to make sure it matches what Punit accepts.
5777 return max_t(u32, val, 0xc0);
5780 /* Check that the pctx buffer wasn't move under us. */
5781 static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
5783 unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
5785 WARN_ON(pctx_addr != dev_priv->mm.stolen_base +
5786 dev_priv->vlv_pctx->stolen->start);
5790 /* Check that the pcbr address is not empty. */
5791 static void cherryview_check_pctx(struct drm_i915_private *dev_priv)
5793 unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
5795 WARN_ON((pctx_addr >> VLV_PCBR_ADDR_SHIFT) == 0);
5798 static void cherryview_setup_pctx(struct drm_i915_private *dev_priv)
5800 struct i915_ggtt *ggtt = &dev_priv->ggtt;
5801 unsigned long pctx_paddr, paddr;
5803 int pctx_size = 32*1024;
5805 pcbr = I915_READ(VLV_PCBR);
5806 if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) {
5807 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5808 paddr = (dev_priv->mm.stolen_base +
5809 (ggtt->stolen_size - pctx_size));
5811 pctx_paddr = (paddr & (~4095));
5812 I915_WRITE(VLV_PCBR, pctx_paddr);
5815 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
5818 static void valleyview_setup_pctx(struct drm_i915_private *dev_priv)
5820 struct drm_i915_gem_object *pctx;
5821 unsigned long pctx_paddr;
5823 int pctx_size = 24*1024;
5825 pcbr = I915_READ(VLV_PCBR);
5827 /* BIOS set it up already, grab the pre-alloc'd space */
5830 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
5831 pctx = i915_gem_object_create_stolen_for_preallocated(&dev_priv->drm,
5833 I915_GTT_OFFSET_NONE,
5838 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5841 * From the Gunit register HAS:
5842 * The Gfx driver is expected to program this register and ensure
5843 * proper allocation within Gfx stolen memory. For example, this
5844 * register should be programmed such than the PCBR range does not
5845 * overlap with other ranges, such as the frame buffer, protected
5846 * memory, or any other relevant ranges.
5848 pctx = i915_gem_object_create_stolen(&dev_priv->drm, pctx_size);
5850 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
5854 pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
5855 I915_WRITE(VLV_PCBR, pctx_paddr);
5858 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
5859 dev_priv->vlv_pctx = pctx;
5862 static void valleyview_cleanup_pctx(struct drm_i915_private *dev_priv)
5864 if (WARN_ON(!dev_priv->vlv_pctx))
5867 i915_gem_object_put_unlocked(dev_priv->vlv_pctx);
5868 dev_priv->vlv_pctx = NULL;
5871 static void vlv_init_gpll_ref_freq(struct drm_i915_private *dev_priv)
5873 dev_priv->rps.gpll_ref_freq =
5874 vlv_get_cck_clock(dev_priv, "GPLL ref",
5875 CCK_GPLL_CLOCK_CONTROL,
5876 dev_priv->czclk_freq);
5878 DRM_DEBUG_DRIVER("GPLL reference freq: %d kHz\n",
5879 dev_priv->rps.gpll_ref_freq);
5882 static void valleyview_init_gt_powersave(struct drm_i915_private *dev_priv)
5886 valleyview_setup_pctx(dev_priv);
5888 vlv_init_gpll_ref_freq(dev_priv);
5890 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5891 switch ((val >> 6) & 3) {
5894 dev_priv->mem_freq = 800;
5897 dev_priv->mem_freq = 1066;
5900 dev_priv->mem_freq = 1333;
5903 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
5905 dev_priv->rps.max_freq = valleyview_rps_max_freq(dev_priv);
5906 dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
5907 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5908 intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
5909 dev_priv->rps.max_freq);
5911 dev_priv->rps.efficient_freq = valleyview_rps_rpe_freq(dev_priv);
5912 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5913 intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5914 dev_priv->rps.efficient_freq);
5916 dev_priv->rps.rp1_freq = valleyview_rps_guar_freq(dev_priv);
5917 DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
5918 intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
5919 dev_priv->rps.rp1_freq);
5921 dev_priv->rps.min_freq = valleyview_rps_min_freq(dev_priv);
5922 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5923 intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
5924 dev_priv->rps.min_freq);
5927 static void cherryview_init_gt_powersave(struct drm_i915_private *dev_priv)
5931 cherryview_setup_pctx(dev_priv);
5933 vlv_init_gpll_ref_freq(dev_priv);
5935 mutex_lock(&dev_priv->sb_lock);
5936 val = vlv_cck_read(dev_priv, CCK_FUSE_REG);
5937 mutex_unlock(&dev_priv->sb_lock);
5939 switch ((val >> 2) & 0x7) {
5941 dev_priv->mem_freq = 2000;
5944 dev_priv->mem_freq = 1600;
5947 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
5949 dev_priv->rps.max_freq = cherryview_rps_max_freq(dev_priv);
5950 dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
5951 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5952 intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
5953 dev_priv->rps.max_freq);
5955 dev_priv->rps.efficient_freq = cherryview_rps_rpe_freq(dev_priv);
5956 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5957 intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5958 dev_priv->rps.efficient_freq);
5960 dev_priv->rps.rp1_freq = cherryview_rps_guar_freq(dev_priv);
5961 DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
5962 intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
5963 dev_priv->rps.rp1_freq);
5965 /* PUnit validated range is only [RPe, RP0] */
5966 dev_priv->rps.min_freq = dev_priv->rps.efficient_freq;
5967 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5968 intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
5969 dev_priv->rps.min_freq);
5971 WARN_ONCE((dev_priv->rps.max_freq |
5972 dev_priv->rps.efficient_freq |
5973 dev_priv->rps.rp1_freq |
5974 dev_priv->rps.min_freq) & 1,
5975 "Odd GPU freq values\n");
5978 static void valleyview_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
5980 valleyview_cleanup_pctx(dev_priv);
5983 static void cherryview_enable_rps(struct drm_i915_private *dev_priv)
5985 struct intel_engine_cs *engine;
5986 enum intel_engine_id id;
5987 u32 gtfifodbg, val, rc6_mode = 0, pcbr;
5989 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5991 gtfifodbg = I915_READ(GTFIFODBG) & ~(GT_FIFO_SBDEDICATE_FREE_ENTRY_CHV |
5992 GT_FIFO_FREE_ENTRIES_CHV);
5994 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5996 I915_WRITE(GTFIFODBG, gtfifodbg);
5999 cherryview_check_pctx(dev_priv);
6001 /* 1a & 1b: Get forcewake during program sequence. Although the driver
6002 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
6003 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6005 /* Disable RC states. */
6006 I915_WRITE(GEN6_RC_CONTROL, 0);
6008 /* 2a: Program RC6 thresholds.*/
6009 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
6010 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
6011 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
6013 for_each_engine(engine, dev_priv, id)
6014 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6015 I915_WRITE(GEN6_RC_SLEEP, 0);
6017 /* TO threshold set to 500 us ( 0x186 * 1.28 us) */
6018 I915_WRITE(GEN6_RC6_THRESHOLD, 0x186);
6020 /* allows RC6 residency counter to work */
6021 I915_WRITE(VLV_COUNTER_CONTROL,
6022 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
6023 VLV_MEDIA_RC6_COUNT_EN |
6024 VLV_RENDER_RC6_COUNT_EN));
6026 /* For now we assume BIOS is allocating and populating the PCBR */
6027 pcbr = I915_READ(VLV_PCBR);
6030 if ((intel_enable_rc6() & INTEL_RC6_ENABLE) &&
6031 (pcbr >> VLV_PCBR_ADDR_SHIFT))
6032 rc6_mode = GEN7_RC_CTL_TO_MODE;
6034 I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
6036 /* 4 Program defaults and thresholds for RPS*/
6037 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
6038 I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
6039 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
6040 I915_WRITE(GEN6_RP_UP_EI, 66000);
6041 I915_WRITE(GEN6_RP_DOWN_EI, 350000);
6043 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
6046 I915_WRITE(GEN6_RP_CONTROL,
6047 GEN6_RP_MEDIA_HW_NORMAL_MODE |
6048 GEN6_RP_MEDIA_IS_GFX |
6050 GEN6_RP_UP_BUSY_AVG |
6051 GEN6_RP_DOWN_IDLE_AVG);
6053 /* Setting Fixed Bias */
6054 val = VLV_OVERRIDE_EN |
6056 CHV_BIAS_CPU_50_SOC_50;
6057 vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
6059 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
6061 /* RPS code assumes GPLL is used */
6062 WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
6064 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
6065 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
6067 reset_rps(dev_priv, valleyview_set_rps);
6069 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6072 static void valleyview_enable_rps(struct drm_i915_private *dev_priv)
6074 struct intel_engine_cs *engine;
6075 enum intel_engine_id id;
6076 u32 gtfifodbg, val, rc6_mode = 0;
6078 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6080 valleyview_check_pctx(dev_priv);
6082 gtfifodbg = I915_READ(GTFIFODBG);
6084 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
6086 I915_WRITE(GTFIFODBG, gtfifodbg);
6089 /* If VLV, Forcewake all wells, else re-direct to regular path */
6090 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6092 /* Disable RC states. */
6093 I915_WRITE(GEN6_RC_CONTROL, 0);
6095 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
6096 I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
6097 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
6098 I915_WRITE(GEN6_RP_UP_EI, 66000);
6099 I915_WRITE(GEN6_RP_DOWN_EI, 350000);
6101 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
6103 I915_WRITE(GEN6_RP_CONTROL,
6104 GEN6_RP_MEDIA_TURBO |
6105 GEN6_RP_MEDIA_HW_NORMAL_MODE |
6106 GEN6_RP_MEDIA_IS_GFX |
6108 GEN6_RP_UP_BUSY_AVG |
6109 GEN6_RP_DOWN_IDLE_CONT);
6111 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
6112 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
6113 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
6115 for_each_engine(engine, dev_priv, id)
6116 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6118 I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
6120 /* allows RC6 residency counter to work */
6121 I915_WRITE(VLV_COUNTER_CONTROL,
6122 _MASKED_BIT_ENABLE(VLV_MEDIA_RC0_COUNT_EN |
6123 VLV_RENDER_RC0_COUNT_EN |
6124 VLV_MEDIA_RC6_COUNT_EN |
6125 VLV_RENDER_RC6_COUNT_EN));
6127 if (intel_enable_rc6() & INTEL_RC6_ENABLE)
6128 rc6_mode = GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
6130 intel_print_rc6_info(dev_priv, rc6_mode);
6132 I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
6134 /* Setting Fixed Bias */
6135 val = VLV_OVERRIDE_EN |
6137 VLV_BIAS_CPU_125_SOC_875;
6138 vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
6140 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
6142 /* RPS code assumes GPLL is used */
6143 WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
6145 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
6146 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
6148 reset_rps(dev_priv, valleyview_set_rps);
6150 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6153 static unsigned long intel_pxfreq(u32 vidfreq)
6156 int div = (vidfreq & 0x3f0000) >> 16;
6157 int post = (vidfreq & 0x3000) >> 12;
6158 int pre = (vidfreq & 0x7);
6163 freq = ((div * 133333) / ((1<<post) * pre));
6168 static const struct cparams {
6174 { 1, 1333, 301, 28664 },
6175 { 1, 1066, 294, 24460 },
6176 { 1, 800, 294, 25192 },
6177 { 0, 1333, 276, 27605 },
6178 { 0, 1066, 276, 27605 },
6179 { 0, 800, 231, 23784 },
6182 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
6184 u64 total_count, diff, ret;
6185 u32 count1, count2, count3, m = 0, c = 0;
6186 unsigned long now = jiffies_to_msecs(jiffies), diff1;
6189 assert_spin_locked(&mchdev_lock);
6191 diff1 = now - dev_priv->ips.last_time1;
6193 /* Prevent division-by-zero if we are asking too fast.
6194 * Also, we don't get interesting results if we are polling
6195 * faster than once in 10ms, so just return the saved value
6199 return dev_priv->ips.chipset_power;
6201 count1 = I915_READ(DMIEC);
6202 count2 = I915_READ(DDREC);
6203 count3 = I915_READ(CSIEC);
6205 total_count = count1 + count2 + count3;
6207 /* FIXME: handle per-counter overflow */
6208 if (total_count < dev_priv->ips.last_count1) {
6209 diff = ~0UL - dev_priv->ips.last_count1;
6210 diff += total_count;
6212 diff = total_count - dev_priv->ips.last_count1;
6215 for (i = 0; i < ARRAY_SIZE(cparams); i++) {
6216 if (cparams[i].i == dev_priv->ips.c_m &&
6217 cparams[i].t == dev_priv->ips.r_t) {
6224 diff = div_u64(diff, diff1);
6225 ret = ((m * diff) + c);
6226 ret = div_u64(ret, 10);
6228 dev_priv->ips.last_count1 = total_count;
6229 dev_priv->ips.last_time1 = now;
6231 dev_priv->ips.chipset_power = ret;
6236 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
6240 if (INTEL_INFO(dev_priv)->gen != 5)
6243 spin_lock_irq(&mchdev_lock);
6245 val = __i915_chipset_val(dev_priv);
6247 spin_unlock_irq(&mchdev_lock);
6252 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
6254 unsigned long m, x, b;
6257 tsfs = I915_READ(TSFS);
6259 m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
6260 x = I915_READ8(TR1);
6262 b = tsfs & TSFS_INTR_MASK;
6264 return ((m * x) / 127) - b;
6267 static int _pxvid_to_vd(u8 pxvid)
6272 if (pxvid >= 8 && pxvid < 31)
6275 return (pxvid + 2) * 125;
6278 static u32 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
6280 const int vd = _pxvid_to_vd(pxvid);
6281 const int vm = vd - 1125;
6283 if (INTEL_INFO(dev_priv)->is_mobile)
6284 return vm > 0 ? vm : 0;
6289 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
6291 u64 now, diff, diffms;
6294 assert_spin_locked(&mchdev_lock);
6296 now = ktime_get_raw_ns();
6297 diffms = now - dev_priv->ips.last_time2;
6298 do_div(diffms, NSEC_PER_MSEC);
6300 /* Don't divide by 0 */
6304 count = I915_READ(GFXEC);
6306 if (count < dev_priv->ips.last_count2) {
6307 diff = ~0UL - dev_priv->ips.last_count2;
6310 diff = count - dev_priv->ips.last_count2;
6313 dev_priv->ips.last_count2 = count;
6314 dev_priv->ips.last_time2 = now;
6316 /* More magic constants... */
6318 diff = div_u64(diff, diffms * 10);
6319 dev_priv->ips.gfx_power = diff;
6322 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
6324 if (INTEL_INFO(dev_priv)->gen != 5)
6327 spin_lock_irq(&mchdev_lock);
6329 __i915_update_gfx_val(dev_priv);
6331 spin_unlock_irq(&mchdev_lock);
6334 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
6336 unsigned long t, corr, state1, corr2, state2;
6339 assert_spin_locked(&mchdev_lock);
6341 pxvid = I915_READ(PXVFREQ(dev_priv->rps.cur_freq));
6342 pxvid = (pxvid >> 24) & 0x7f;
6343 ext_v = pvid_to_extvid(dev_priv, pxvid);
6347 t = i915_mch_val(dev_priv);
6349 /* Revel in the empirically derived constants */
6351 /* Correction factor in 1/100000 units */
6353 corr = ((t * 2349) + 135940);
6355 corr = ((t * 964) + 29317);
6357 corr = ((t * 301) + 1004);
6359 corr = corr * ((150142 * state1) / 10000 - 78642);
6361 corr2 = (corr * dev_priv->ips.corr);
6363 state2 = (corr2 * state1) / 10000;
6364 state2 /= 100; /* convert to mW */
6366 __i915_update_gfx_val(dev_priv);
6368 return dev_priv->ips.gfx_power + state2;
6371 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
6375 if (INTEL_INFO(dev_priv)->gen != 5)
6378 spin_lock_irq(&mchdev_lock);
6380 val = __i915_gfx_val(dev_priv);
6382 spin_unlock_irq(&mchdev_lock);
6388 * i915_read_mch_val - return value for IPS use
6390 * Calculate and return a value for the IPS driver to use when deciding whether
6391 * we have thermal and power headroom to increase CPU or GPU power budget.
6393 unsigned long i915_read_mch_val(void)
6395 struct drm_i915_private *dev_priv;
6396 unsigned long chipset_val, graphics_val, ret = 0;
6398 spin_lock_irq(&mchdev_lock);
6401 dev_priv = i915_mch_dev;
6403 chipset_val = __i915_chipset_val(dev_priv);
6404 graphics_val = __i915_gfx_val(dev_priv);
6406 ret = chipset_val + graphics_val;
6409 spin_unlock_irq(&mchdev_lock);
6413 EXPORT_SYMBOL_GPL(i915_read_mch_val);
6416 * i915_gpu_raise - raise GPU frequency limit
6418 * Raise the limit; IPS indicates we have thermal headroom.
6420 bool i915_gpu_raise(void)
6422 struct drm_i915_private *dev_priv;
6425 spin_lock_irq(&mchdev_lock);
6426 if (!i915_mch_dev) {
6430 dev_priv = i915_mch_dev;
6432 if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
6433 dev_priv->ips.max_delay--;
6436 spin_unlock_irq(&mchdev_lock);
6440 EXPORT_SYMBOL_GPL(i915_gpu_raise);
6443 * i915_gpu_lower - lower GPU frequency limit
6445 * IPS indicates we're close to a thermal limit, so throttle back the GPU
6446 * frequency maximum.
6448 bool i915_gpu_lower(void)
6450 struct drm_i915_private *dev_priv;
6453 spin_lock_irq(&mchdev_lock);
6454 if (!i915_mch_dev) {
6458 dev_priv = i915_mch_dev;
6460 if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
6461 dev_priv->ips.max_delay++;
6464 spin_unlock_irq(&mchdev_lock);
6468 EXPORT_SYMBOL_GPL(i915_gpu_lower);
6471 * i915_gpu_busy - indicate GPU business to IPS
6473 * Tell the IPS driver whether or not the GPU is busy.
6475 bool i915_gpu_busy(void)
6479 spin_lock_irq(&mchdev_lock);
6481 ret = i915_mch_dev->gt.awake;
6482 spin_unlock_irq(&mchdev_lock);
6486 EXPORT_SYMBOL_GPL(i915_gpu_busy);
6489 * i915_gpu_turbo_disable - disable graphics turbo
6491 * Disable graphics turbo by resetting the max frequency and setting the
6492 * current frequency to the default.
6494 bool i915_gpu_turbo_disable(void)
6496 struct drm_i915_private *dev_priv;
6499 spin_lock_irq(&mchdev_lock);
6500 if (!i915_mch_dev) {
6504 dev_priv = i915_mch_dev;
6506 dev_priv->ips.max_delay = dev_priv->ips.fstart;
6508 if (!ironlake_set_drps(dev_priv, dev_priv->ips.fstart))
6512 spin_unlock_irq(&mchdev_lock);
6516 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
6519 * Tells the intel_ips driver that the i915 driver is now loaded, if
6520 * IPS got loaded first.
6522 * This awkward dance is so that neither module has to depend on the
6523 * other in order for IPS to do the appropriate communication of
6524 * GPU turbo limits to i915.
6527 ips_ping_for_i915_load(void)
6531 link = symbol_get(ips_link_to_i915_driver);
6534 symbol_put(ips_link_to_i915_driver);
6538 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
6540 /* We only register the i915 ips part with intel-ips once everything is
6541 * set up, to avoid intel-ips sneaking in and reading bogus values. */
6542 spin_lock_irq(&mchdev_lock);
6543 i915_mch_dev = dev_priv;
6544 spin_unlock_irq(&mchdev_lock);
6546 ips_ping_for_i915_load();
6549 void intel_gpu_ips_teardown(void)
6551 spin_lock_irq(&mchdev_lock);
6552 i915_mch_dev = NULL;
6553 spin_unlock_irq(&mchdev_lock);
6556 static void intel_init_emon(struct drm_i915_private *dev_priv)
6562 /* Disable to program */
6566 /* Program energy weights for various events */
6567 I915_WRITE(SDEW, 0x15040d00);
6568 I915_WRITE(CSIEW0, 0x007f0000);
6569 I915_WRITE(CSIEW1, 0x1e220004);
6570 I915_WRITE(CSIEW2, 0x04000004);
6572 for (i = 0; i < 5; i++)
6573 I915_WRITE(PEW(i), 0);
6574 for (i = 0; i < 3; i++)
6575 I915_WRITE(DEW(i), 0);
6577 /* Program P-state weights to account for frequency power adjustment */
6578 for (i = 0; i < 16; i++) {
6579 u32 pxvidfreq = I915_READ(PXVFREQ(i));
6580 unsigned long freq = intel_pxfreq(pxvidfreq);
6581 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
6586 val *= (freq / 1000);
6588 val /= (127*127*900);
6590 DRM_ERROR("bad pxval: %ld\n", val);
6593 /* Render standby states get 0 weight */
6597 for (i = 0; i < 4; i++) {
6598 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
6599 (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
6600 I915_WRITE(PXW(i), val);
6603 /* Adjust magic regs to magic values (more experimental results) */
6604 I915_WRITE(OGW0, 0);
6605 I915_WRITE(OGW1, 0);
6606 I915_WRITE(EG0, 0x00007f00);
6607 I915_WRITE(EG1, 0x0000000e);
6608 I915_WRITE(EG2, 0x000e0000);
6609 I915_WRITE(EG3, 0x68000300);
6610 I915_WRITE(EG4, 0x42000000);
6611 I915_WRITE(EG5, 0x00140031);
6615 for (i = 0; i < 8; i++)
6616 I915_WRITE(PXWL(i), 0);
6618 /* Enable PMON + select events */
6619 I915_WRITE(ECR, 0x80000019);
6621 lcfuse = I915_READ(LCFUSE02);
6623 dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
6626 void intel_init_gt_powersave(struct drm_i915_private *dev_priv)
6629 * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
6632 if (!i915.enable_rc6) {
6633 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
6634 intel_runtime_pm_get(dev_priv);
6637 mutex_lock(&dev_priv->drm.struct_mutex);
6638 mutex_lock(&dev_priv->rps.hw_lock);
6640 /* Initialize RPS limits (for userspace) */
6641 if (IS_CHERRYVIEW(dev_priv))
6642 cherryview_init_gt_powersave(dev_priv);
6643 else if (IS_VALLEYVIEW(dev_priv))
6644 valleyview_init_gt_powersave(dev_priv);
6645 else if (INTEL_GEN(dev_priv) >= 6)
6646 gen6_init_rps_frequencies(dev_priv);
6648 /* Derive initial user preferences/limits from the hardware limits */
6649 dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
6650 dev_priv->rps.cur_freq = dev_priv->rps.idle_freq;
6652 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
6653 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
6655 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
6656 dev_priv->rps.min_freq_softlimit =
6658 dev_priv->rps.efficient_freq,
6659 intel_freq_opcode(dev_priv, 450));
6661 /* After setting max-softlimit, find the overclock max freq */
6662 if (IS_GEN6(dev_priv) ||
6663 IS_IVYBRIDGE(dev_priv) || IS_HASWELL(dev_priv)) {
6666 sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, ¶ms);
6667 if (params & BIT(31)) { /* OC supported */
6668 DRM_DEBUG_DRIVER("Overclocking supported, max: %dMHz, overclock: %dMHz\n",
6669 (dev_priv->rps.max_freq & 0xff) * 50,
6670 (params & 0xff) * 50);
6671 dev_priv->rps.max_freq = params & 0xff;
6675 /* Finally allow us to boost to max by default */
6676 dev_priv->rps.boost_freq = dev_priv->rps.max_freq;
6678 mutex_unlock(&dev_priv->rps.hw_lock);
6679 mutex_unlock(&dev_priv->drm.struct_mutex);
6681 intel_autoenable_gt_powersave(dev_priv);
6684 void intel_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
6686 if (IS_VALLEYVIEW(dev_priv))
6687 valleyview_cleanup_gt_powersave(dev_priv);
6689 if (!i915.enable_rc6)
6690 intel_runtime_pm_put(dev_priv);
6694 * intel_suspend_gt_powersave - suspend PM work and helper threads
6695 * @dev_priv: i915 device
6697 * We don't want to disable RC6 or other features here, we just want
6698 * to make sure any work we've queued has finished and won't bother
6699 * us while we're suspended.
6701 void intel_suspend_gt_powersave(struct drm_i915_private *dev_priv)
6703 if (INTEL_GEN(dev_priv) < 6)
6706 if (cancel_delayed_work_sync(&dev_priv->rps.autoenable_work))
6707 intel_runtime_pm_put(dev_priv);
6709 /* gen6_rps_idle() will be called later to disable interrupts */
6712 void intel_sanitize_gt_powersave(struct drm_i915_private *dev_priv)
6714 dev_priv->rps.enabled = true; /* force disabling */
6715 intel_disable_gt_powersave(dev_priv);
6717 gen6_reset_rps_interrupts(dev_priv);
6720 void intel_disable_gt_powersave(struct drm_i915_private *dev_priv)
6722 if (!READ_ONCE(dev_priv->rps.enabled))
6725 mutex_lock(&dev_priv->rps.hw_lock);
6727 if (INTEL_GEN(dev_priv) >= 9) {
6728 gen9_disable_rc6(dev_priv);
6729 gen9_disable_rps(dev_priv);
6730 } else if (IS_CHERRYVIEW(dev_priv)) {
6731 cherryview_disable_rps(dev_priv);
6732 } else if (IS_VALLEYVIEW(dev_priv)) {
6733 valleyview_disable_rps(dev_priv);
6734 } else if (INTEL_GEN(dev_priv) >= 6) {
6735 gen6_disable_rps(dev_priv);
6736 } else if (IS_IRONLAKE_M(dev_priv)) {
6737 ironlake_disable_drps(dev_priv);
6740 dev_priv->rps.enabled = false;
6741 mutex_unlock(&dev_priv->rps.hw_lock);
6744 void intel_enable_gt_powersave(struct drm_i915_private *dev_priv)
6746 /* We shouldn't be disabling as we submit, so this should be less
6747 * racy than it appears!
6749 if (READ_ONCE(dev_priv->rps.enabled))
6752 /* Powersaving is controlled by the host when inside a VM */
6753 if (intel_vgpu_active(dev_priv))
6756 mutex_lock(&dev_priv->rps.hw_lock);
6758 if (IS_CHERRYVIEW(dev_priv)) {
6759 cherryview_enable_rps(dev_priv);
6760 } else if (IS_VALLEYVIEW(dev_priv)) {
6761 valleyview_enable_rps(dev_priv);
6762 } else if (INTEL_GEN(dev_priv) >= 9) {
6763 gen9_enable_rc6(dev_priv);
6764 gen9_enable_rps(dev_priv);
6765 if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv))
6766 gen6_update_ring_freq(dev_priv);
6767 } else if (IS_BROADWELL(dev_priv)) {
6768 gen8_enable_rps(dev_priv);
6769 gen6_update_ring_freq(dev_priv);
6770 } else if (INTEL_GEN(dev_priv) >= 6) {
6771 gen6_enable_rps(dev_priv);
6772 gen6_update_ring_freq(dev_priv);
6773 } else if (IS_IRONLAKE_M(dev_priv)) {
6774 ironlake_enable_drps(dev_priv);
6775 intel_init_emon(dev_priv);
6778 WARN_ON(dev_priv->rps.max_freq < dev_priv->rps.min_freq);
6779 WARN_ON(dev_priv->rps.idle_freq > dev_priv->rps.max_freq);
6781 WARN_ON(dev_priv->rps.efficient_freq < dev_priv->rps.min_freq);
6782 WARN_ON(dev_priv->rps.efficient_freq > dev_priv->rps.max_freq);
6784 dev_priv->rps.enabled = true;
6785 mutex_unlock(&dev_priv->rps.hw_lock);
6788 static void __intel_autoenable_gt_powersave(struct work_struct *work)
6790 struct drm_i915_private *dev_priv =
6791 container_of(work, typeof(*dev_priv), rps.autoenable_work.work);
6792 struct intel_engine_cs *rcs;
6793 struct drm_i915_gem_request *req;
6795 if (READ_ONCE(dev_priv->rps.enabled))
6798 rcs = dev_priv->engine[RCS];
6799 if (rcs->last_context)
6802 if (!rcs->init_context)
6805 mutex_lock(&dev_priv->drm.struct_mutex);
6807 req = i915_gem_request_alloc(rcs, dev_priv->kernel_context);
6811 if (!i915.enable_execlists && i915_switch_context(req) == 0)
6812 rcs->init_context(req);
6814 /* Mark the device busy, calling intel_enable_gt_powersave() */
6815 i915_add_request_no_flush(req);
6818 mutex_unlock(&dev_priv->drm.struct_mutex);
6820 intel_runtime_pm_put(dev_priv);
6823 void intel_autoenable_gt_powersave(struct drm_i915_private *dev_priv)
6825 if (READ_ONCE(dev_priv->rps.enabled))
6828 if (IS_IRONLAKE_M(dev_priv)) {
6829 ironlake_enable_drps(dev_priv);
6830 intel_init_emon(dev_priv);
6831 } else if (INTEL_INFO(dev_priv)->gen >= 6) {
6833 * PCU communication is slow and this doesn't need to be
6834 * done at any specific time, so do this out of our fast path
6835 * to make resume and init faster.
6837 * We depend on the HW RC6 power context save/restore
6838 * mechanism when entering D3 through runtime PM suspend. So
6839 * disable RPM until RPS/RC6 is properly setup. We can only
6840 * get here via the driver load/system resume/runtime resume
6841 * paths, so the _noresume version is enough (and in case of
6842 * runtime resume it's necessary).
6844 if (queue_delayed_work(dev_priv->wq,
6845 &dev_priv->rps.autoenable_work,
6846 round_jiffies_up_relative(HZ)))
6847 intel_runtime_pm_get_noresume(dev_priv);
6851 static void ibx_init_clock_gating(struct drm_device *dev)
6853 struct drm_i915_private *dev_priv = to_i915(dev);
6856 * On Ibex Peak and Cougar Point, we need to disable clock
6857 * gating for the panel power sequencer or it will fail to
6858 * start up when no ports are active.
6860 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
6863 static void g4x_disable_trickle_feed(struct drm_device *dev)
6865 struct drm_i915_private *dev_priv = to_i915(dev);
6868 for_each_pipe(dev_priv, pipe) {
6869 I915_WRITE(DSPCNTR(pipe),
6870 I915_READ(DSPCNTR(pipe)) |
6871 DISPPLANE_TRICKLE_FEED_DISABLE);
6873 I915_WRITE(DSPSURF(pipe), I915_READ(DSPSURF(pipe)));
6874 POSTING_READ(DSPSURF(pipe));
6878 static void ilk_init_lp_watermarks(struct drm_device *dev)
6880 struct drm_i915_private *dev_priv = to_i915(dev);
6882 I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
6883 I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
6884 I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
6887 * Don't touch WM1S_LP_EN here.
6888 * Doing so could cause underruns.
6892 static void ironlake_init_clock_gating(struct drm_device *dev)
6894 struct drm_i915_private *dev_priv = to_i915(dev);
6895 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6899 * WaFbcDisableDpfcClockGating:ilk
6901 dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
6902 ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
6903 ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
6905 I915_WRITE(PCH_3DCGDIS0,
6906 MARIUNIT_CLOCK_GATE_DISABLE |
6907 SVSMUNIT_CLOCK_GATE_DISABLE);
6908 I915_WRITE(PCH_3DCGDIS1,
6909 VFMUNIT_CLOCK_GATE_DISABLE);
6912 * According to the spec the following bits should be set in
6913 * order to enable memory self-refresh
6914 * The bit 22/21 of 0x42004
6915 * The bit 5 of 0x42020
6916 * The bit 15 of 0x45000
6918 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6919 (I915_READ(ILK_DISPLAY_CHICKEN2) |
6920 ILK_DPARB_GATE | ILK_VSDPFD_FULL));
6921 dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
6922 I915_WRITE(DISP_ARB_CTL,
6923 (I915_READ(DISP_ARB_CTL) |
6926 ilk_init_lp_watermarks(dev);
6929 * Based on the document from hardware guys the following bits
6930 * should be set unconditionally in order to enable FBC.
6931 * The bit 22 of 0x42000
6932 * The bit 22 of 0x42004
6933 * The bit 7,8,9 of 0x42020.
6935 if (IS_IRONLAKE_M(dev_priv)) {
6936 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
6937 I915_WRITE(ILK_DISPLAY_CHICKEN1,
6938 I915_READ(ILK_DISPLAY_CHICKEN1) |
6940 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6941 I915_READ(ILK_DISPLAY_CHICKEN2) |
6945 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
6947 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6948 I915_READ(ILK_DISPLAY_CHICKEN2) |
6949 ILK_ELPIN_409_SELECT);
6950 I915_WRITE(_3D_CHICKEN2,
6951 _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
6952 _3D_CHICKEN2_WM_READ_PIPELINED);
6954 /* WaDisableRenderCachePipelinedFlush:ilk */
6955 I915_WRITE(CACHE_MODE_0,
6956 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
6958 /* WaDisable_RenderCache_OperationalFlush:ilk */
6959 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6961 g4x_disable_trickle_feed(dev);
6963 ibx_init_clock_gating(dev);
6966 static void cpt_init_clock_gating(struct drm_device *dev)
6968 struct drm_i915_private *dev_priv = to_i915(dev);
6973 * On Ibex Peak and Cougar Point, we need to disable clock
6974 * gating for the panel power sequencer or it will fail to
6975 * start up when no ports are active.
6977 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
6978 PCH_DPLUNIT_CLOCK_GATE_DISABLE |
6979 PCH_CPUNIT_CLOCK_GATE_DISABLE);
6980 I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
6981 DPLS_EDP_PPS_FIX_DIS);
6982 /* The below fixes the weird display corruption, a few pixels shifted
6983 * downward, on (only) LVDS of some HP laptops with IVY.
6985 for_each_pipe(dev_priv, pipe) {
6986 val = I915_READ(TRANS_CHICKEN2(pipe));
6987 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
6988 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6989 if (dev_priv->vbt.fdi_rx_polarity_inverted)
6990 val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6991 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
6992 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
6993 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
6994 I915_WRITE(TRANS_CHICKEN2(pipe), val);
6996 /* WADP0ClockGatingDisable */
6997 for_each_pipe(dev_priv, pipe) {
6998 I915_WRITE(TRANS_CHICKEN1(pipe),
6999 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
7003 static void gen6_check_mch_setup(struct drm_device *dev)
7005 struct drm_i915_private *dev_priv = to_i915(dev);
7008 tmp = I915_READ(MCH_SSKPD);
7009 if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL)
7010 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
7014 static void gen6_init_clock_gating(struct drm_device *dev)
7016 struct drm_i915_private *dev_priv = to_i915(dev);
7017 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
7019 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
7021 I915_WRITE(ILK_DISPLAY_CHICKEN2,
7022 I915_READ(ILK_DISPLAY_CHICKEN2) |
7023 ILK_ELPIN_409_SELECT);
7025 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
7026 I915_WRITE(_3D_CHICKEN,
7027 _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
7029 /* WaDisable_RenderCache_OperationalFlush:snb */
7030 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7033 * BSpec recoomends 8x4 when MSAA is used,
7034 * however in practice 16x4 seems fastest.
7036 * Note that PS/WM thread counts depend on the WIZ hashing
7037 * disable bit, which we don't touch here, but it's good
7038 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7040 I915_WRITE(GEN6_GT_MODE,
7041 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7043 ilk_init_lp_watermarks(dev);
7045 I915_WRITE(CACHE_MODE_0,
7046 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
7048 I915_WRITE(GEN6_UCGCTL1,
7049 I915_READ(GEN6_UCGCTL1) |
7050 GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
7051 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
7053 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
7054 * gating disable must be set. Failure to set it results in
7055 * flickering pixels due to Z write ordering failures after
7056 * some amount of runtime in the Mesa "fire" demo, and Unigine
7057 * Sanctuary and Tropics, and apparently anything else with
7058 * alpha test or pixel discard.
7060 * According to the spec, bit 11 (RCCUNIT) must also be set,
7061 * but we didn't debug actual testcases to find it out.
7063 * WaDisableRCCUnitClockGating:snb
7064 * WaDisableRCPBUnitClockGating:snb
7066 I915_WRITE(GEN6_UCGCTL2,
7067 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
7068 GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
7070 /* WaStripsFansDisableFastClipPerformanceFix:snb */
7071 I915_WRITE(_3D_CHICKEN3,
7072 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
7076 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
7077 * 3DSTATE_SF number of SF output attributes is more than 16."
7079 I915_WRITE(_3D_CHICKEN3,
7080 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
7083 * According to the spec the following bits should be
7084 * set in order to enable memory self-refresh and fbc:
7085 * The bit21 and bit22 of 0x42000
7086 * The bit21 and bit22 of 0x42004
7087 * The bit5 and bit7 of 0x42020
7088 * The bit14 of 0x70180
7089 * The bit14 of 0x71180
7091 * WaFbcAsynchFlipDisableFbcQueue:snb
7093 I915_WRITE(ILK_DISPLAY_CHICKEN1,
7094 I915_READ(ILK_DISPLAY_CHICKEN1) |
7095 ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
7096 I915_WRITE(ILK_DISPLAY_CHICKEN2,
7097 I915_READ(ILK_DISPLAY_CHICKEN2) |
7098 ILK_DPARB_GATE | ILK_VSDPFD_FULL);
7099 I915_WRITE(ILK_DSPCLK_GATE_D,
7100 I915_READ(ILK_DSPCLK_GATE_D) |
7101 ILK_DPARBUNIT_CLOCK_GATE_ENABLE |
7102 ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
7104 g4x_disable_trickle_feed(dev);
7106 cpt_init_clock_gating(dev);
7108 gen6_check_mch_setup(dev);
7111 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
7113 uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
7116 * WaVSThreadDispatchOverride:ivb,vlv
7118 * This actually overrides the dispatch
7119 * mode for all thread types.
7121 reg &= ~GEN7_FF_SCHED_MASK;
7122 reg |= GEN7_FF_TS_SCHED_HW;
7123 reg |= GEN7_FF_VS_SCHED_HW;
7124 reg |= GEN7_FF_DS_SCHED_HW;
7126 I915_WRITE(GEN7_FF_THREAD_MODE, reg);
7129 static void lpt_init_clock_gating(struct drm_device *dev)
7131 struct drm_i915_private *dev_priv = to_i915(dev);
7134 * TODO: this bit should only be enabled when really needed, then
7135 * disabled when not needed anymore in order to save power.
7137 if (HAS_PCH_LPT_LP(dev_priv))
7138 I915_WRITE(SOUTH_DSPCLK_GATE_D,
7139 I915_READ(SOUTH_DSPCLK_GATE_D) |
7140 PCH_LP_PARTITION_LEVEL_DISABLE);
7142 /* WADPOClockGatingDisable:hsw */
7143 I915_WRITE(TRANS_CHICKEN1(PIPE_A),
7144 I915_READ(TRANS_CHICKEN1(PIPE_A)) |
7145 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
7148 static void lpt_suspend_hw(struct drm_device *dev)
7150 struct drm_i915_private *dev_priv = to_i915(dev);
7152 if (HAS_PCH_LPT_LP(dev_priv)) {
7153 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
7155 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
7156 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
7160 static void gen8_set_l3sqc_credits(struct drm_i915_private *dev_priv,
7161 int general_prio_credits,
7162 int high_prio_credits)
7166 /* WaTempDisableDOPClkGating:bdw */
7167 misccpctl = I915_READ(GEN7_MISCCPCTL);
7168 I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
7170 I915_WRITE(GEN8_L3SQCREG1,
7171 L3_GENERAL_PRIO_CREDITS(general_prio_credits) |
7172 L3_HIGH_PRIO_CREDITS(high_prio_credits));
7175 * Wait at least 100 clocks before re-enabling clock gating.
7176 * See the definition of L3SQCREG1 in BSpec.
7178 POSTING_READ(GEN8_L3SQCREG1);
7180 I915_WRITE(GEN7_MISCCPCTL, misccpctl);
7183 static void kabylake_init_clock_gating(struct drm_device *dev)
7185 struct drm_i915_private *dev_priv = dev->dev_private;
7187 gen9_init_clock_gating(dev);
7189 /* WaDisableSDEUnitClockGating:kbl */
7190 if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
7191 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7192 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7194 /* WaDisableGamClockGating:kbl */
7195 if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
7196 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
7197 GEN6_GAMUNIT_CLOCK_GATE_DISABLE);
7199 /* WaFbcNukeOnHostModify:kbl */
7200 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
7201 ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
7204 static void skylake_init_clock_gating(struct drm_device *dev)
7206 struct drm_i915_private *dev_priv = dev->dev_private;
7208 gen9_init_clock_gating(dev);
7210 /* WAC6entrylatency:skl */
7211 I915_WRITE(FBC_LLC_READ_CTRL, I915_READ(FBC_LLC_READ_CTRL) |
7212 FBC_LLC_FULLY_OPEN);
7214 /* WaFbcNukeOnHostModify:skl */
7215 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
7216 ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
7219 static void broadwell_init_clock_gating(struct drm_device *dev)
7221 struct drm_i915_private *dev_priv = to_i915(dev);
7224 ilk_init_lp_watermarks(dev);
7226 /* WaSwitchSolVfFArbitrationPriority:bdw */
7227 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
7229 /* WaPsrDPAMaskVBlankInSRD:bdw */
7230 I915_WRITE(CHICKEN_PAR1_1,
7231 I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
7233 /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
7234 for_each_pipe(dev_priv, pipe) {
7235 I915_WRITE(CHICKEN_PIPESL_1(pipe),
7236 I915_READ(CHICKEN_PIPESL_1(pipe)) |
7237 BDW_DPRS_MASK_VBLANK_SRD);
7240 /* WaVSRefCountFullforceMissDisable:bdw */
7241 /* WaDSRefCountFullforceMissDisable:bdw */
7242 I915_WRITE(GEN7_FF_THREAD_MODE,
7243 I915_READ(GEN7_FF_THREAD_MODE) &
7244 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
7246 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
7247 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
7249 /* WaDisableSDEUnitClockGating:bdw */
7250 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7251 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7253 /* WaProgramL3SqcReg1Default:bdw */
7254 gen8_set_l3sqc_credits(dev_priv, 30, 2);
7257 * WaGttCachingOffByDefault:bdw
7258 * GTT cache may not work with big pages, so if those
7259 * are ever enabled GTT cache may need to be disabled.
7261 I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
7263 /* WaKVMNotificationOnConfigChange:bdw */
7264 I915_WRITE(CHICKEN_PAR2_1, I915_READ(CHICKEN_PAR2_1)
7265 | KVM_CONFIG_CHANGE_NOTIFICATION_SELECT);
7267 lpt_init_clock_gating(dev);
7270 static void haswell_init_clock_gating(struct drm_device *dev)
7272 struct drm_i915_private *dev_priv = to_i915(dev);
7274 ilk_init_lp_watermarks(dev);
7276 /* L3 caching of data atomics doesn't work -- disable it. */
7277 I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
7278 I915_WRITE(HSW_ROW_CHICKEN3,
7279 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
7281 /* This is required by WaCatErrorRejectionIssue:hsw */
7282 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7283 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7284 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7286 /* WaVSRefCountFullforceMissDisable:hsw */
7287 I915_WRITE(GEN7_FF_THREAD_MODE,
7288 I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
7290 /* WaDisable_RenderCache_OperationalFlush:hsw */
7291 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7293 /* enable HiZ Raw Stall Optimization */
7294 I915_WRITE(CACHE_MODE_0_GEN7,
7295 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
7297 /* WaDisable4x2SubspanOptimization:hsw */
7298 I915_WRITE(CACHE_MODE_1,
7299 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7302 * BSpec recommends 8x4 when MSAA is used,
7303 * however in practice 16x4 seems fastest.
7305 * Note that PS/WM thread counts depend on the WIZ hashing
7306 * disable bit, which we don't touch here, but it's good
7307 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7309 I915_WRITE(GEN7_GT_MODE,
7310 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7312 /* WaSampleCChickenBitEnable:hsw */
7313 I915_WRITE(HALF_SLICE_CHICKEN3,
7314 _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE));
7316 /* WaSwitchSolVfFArbitrationPriority:hsw */
7317 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
7319 /* WaRsPkgCStateDisplayPMReq:hsw */
7320 I915_WRITE(CHICKEN_PAR1_1,
7321 I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
7323 lpt_init_clock_gating(dev);
7326 static void ivybridge_init_clock_gating(struct drm_device *dev)
7328 struct drm_i915_private *dev_priv = to_i915(dev);
7331 ilk_init_lp_watermarks(dev);
7333 I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
7335 /* WaDisableEarlyCull:ivb */
7336 I915_WRITE(_3D_CHICKEN3,
7337 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
7339 /* WaDisableBackToBackFlipFix:ivb */
7340 I915_WRITE(IVB_CHICKEN3,
7341 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7342 CHICKEN3_DGMG_DONE_FIX_DISABLE);
7344 /* WaDisablePSDDualDispatchEnable:ivb */
7345 if (IS_IVB_GT1(dev_priv))
7346 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
7347 _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
7349 /* WaDisable_RenderCache_OperationalFlush:ivb */
7350 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7352 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
7353 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
7354 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
7356 /* WaApplyL3ControlAndL3ChickenMode:ivb */
7357 I915_WRITE(GEN7_L3CNTLREG1,
7358 GEN7_WA_FOR_GEN7_L3_CONTROL);
7359 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
7360 GEN7_WA_L3_CHICKEN_MODE);
7361 if (IS_IVB_GT1(dev_priv))
7362 I915_WRITE(GEN7_ROW_CHICKEN2,
7363 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7365 /* must write both registers */
7366 I915_WRITE(GEN7_ROW_CHICKEN2,
7367 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7368 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
7369 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7372 /* WaForceL3Serialization:ivb */
7373 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
7374 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
7377 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7378 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
7380 I915_WRITE(GEN6_UCGCTL2,
7381 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7383 /* This is required by WaCatErrorRejectionIssue:ivb */
7384 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7385 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7386 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7388 g4x_disable_trickle_feed(dev);
7390 gen7_setup_fixed_func_scheduler(dev_priv);
7392 if (0) { /* causes HiZ corruption on ivb:gt1 */
7393 /* enable HiZ Raw Stall Optimization */
7394 I915_WRITE(CACHE_MODE_0_GEN7,
7395 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
7398 /* WaDisable4x2SubspanOptimization:ivb */
7399 I915_WRITE(CACHE_MODE_1,
7400 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7403 * BSpec recommends 8x4 when MSAA is used,
7404 * however in practice 16x4 seems fastest.
7406 * Note that PS/WM thread counts depend on the WIZ hashing
7407 * disable bit, which we don't touch here, but it's good
7408 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7410 I915_WRITE(GEN7_GT_MODE,
7411 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7413 snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
7414 snpcr &= ~GEN6_MBC_SNPCR_MASK;
7415 snpcr |= GEN6_MBC_SNPCR_MED;
7416 I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
7418 if (!HAS_PCH_NOP(dev_priv))
7419 cpt_init_clock_gating(dev);
7421 gen6_check_mch_setup(dev);
7424 static void valleyview_init_clock_gating(struct drm_device *dev)
7426 struct drm_i915_private *dev_priv = to_i915(dev);
7428 /* WaDisableEarlyCull:vlv */
7429 I915_WRITE(_3D_CHICKEN3,
7430 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
7432 /* WaDisableBackToBackFlipFix:vlv */
7433 I915_WRITE(IVB_CHICKEN3,
7434 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7435 CHICKEN3_DGMG_DONE_FIX_DISABLE);
7437 /* WaPsdDispatchEnable:vlv */
7438 /* WaDisablePSDDualDispatchEnable:vlv */
7439 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
7440 _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
7441 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
7443 /* WaDisable_RenderCache_OperationalFlush:vlv */
7444 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7446 /* WaForceL3Serialization:vlv */
7447 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
7448 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
7450 /* WaDisableDopClockGating:vlv */
7451 I915_WRITE(GEN7_ROW_CHICKEN2,
7452 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7454 /* This is required by WaCatErrorRejectionIssue:vlv */
7455 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7456 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7457 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7459 gen7_setup_fixed_func_scheduler(dev_priv);
7462 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7463 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
7465 I915_WRITE(GEN6_UCGCTL2,
7466 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7468 /* WaDisableL3Bank2xClockGate:vlv
7469 * Disabling L3 clock gating- MMIO 940c[25] = 1
7470 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
7471 I915_WRITE(GEN7_UCGCTL4,
7472 I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
7475 * BSpec says this must be set, even though
7476 * WaDisable4x2SubspanOptimization isn't listed for VLV.
7478 I915_WRITE(CACHE_MODE_1,
7479 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7482 * BSpec recommends 8x4 when MSAA is used,
7483 * however in practice 16x4 seems fastest.
7485 * Note that PS/WM thread counts depend on the WIZ hashing
7486 * disable bit, which we don't touch here, but it's good
7487 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7489 I915_WRITE(GEN7_GT_MODE,
7490 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7493 * WaIncreaseL3CreditsForVLVB0:vlv
7494 * This is the hardware default actually.
7496 I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
7499 * WaDisableVLVClockGating_VBIIssue:vlv
7500 * Disable clock gating on th GCFG unit to prevent a delay
7501 * in the reporting of vblank events.
7503 I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
7506 static void cherryview_init_clock_gating(struct drm_device *dev)
7508 struct drm_i915_private *dev_priv = to_i915(dev);
7510 /* WaVSRefCountFullforceMissDisable:chv */
7511 /* WaDSRefCountFullforceMissDisable:chv */
7512 I915_WRITE(GEN7_FF_THREAD_MODE,
7513 I915_READ(GEN7_FF_THREAD_MODE) &
7514 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
7516 /* WaDisableSemaphoreAndSyncFlipWait:chv */
7517 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
7518 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
7520 /* WaDisableCSUnitClockGating:chv */
7521 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
7522 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
7524 /* WaDisableSDEUnitClockGating:chv */
7525 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7526 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7529 * WaProgramL3SqcReg1Default:chv
7530 * See gfxspecs/Related Documents/Performance Guide/
7531 * LSQC Setting Recommendations.
7533 gen8_set_l3sqc_credits(dev_priv, 38, 2);
7536 * GTT cache may not work with big pages, so if those
7537 * are ever enabled GTT cache may need to be disabled.
7539 I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
7542 static void g4x_init_clock_gating(struct drm_device *dev)
7544 struct drm_i915_private *dev_priv = to_i915(dev);
7545 uint32_t dspclk_gate;
7547 I915_WRITE(RENCLK_GATE_D1, 0);
7548 I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
7549 GS_UNIT_CLOCK_GATE_DISABLE |
7550 CL_UNIT_CLOCK_GATE_DISABLE);
7551 I915_WRITE(RAMCLK_GATE_D, 0);
7552 dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
7553 OVRUNIT_CLOCK_GATE_DISABLE |
7554 OVCUNIT_CLOCK_GATE_DISABLE;
7555 if (IS_GM45(dev_priv))
7556 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
7557 I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
7559 /* WaDisableRenderCachePipelinedFlush */
7560 I915_WRITE(CACHE_MODE_0,
7561 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
7563 /* WaDisable_RenderCache_OperationalFlush:g4x */
7564 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7566 g4x_disable_trickle_feed(dev);
7569 static void crestline_init_clock_gating(struct drm_device *dev)
7571 struct drm_i915_private *dev_priv = to_i915(dev);
7573 I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
7574 I915_WRITE(RENCLK_GATE_D2, 0);
7575 I915_WRITE(DSPCLK_GATE_D, 0);
7576 I915_WRITE(RAMCLK_GATE_D, 0);
7577 I915_WRITE16(DEUC, 0);
7578 I915_WRITE(MI_ARB_STATE,
7579 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7581 /* WaDisable_RenderCache_OperationalFlush:gen4 */
7582 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7585 static void broadwater_init_clock_gating(struct drm_device *dev)
7587 struct drm_i915_private *dev_priv = to_i915(dev);
7589 I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
7590 I965_RCC_CLOCK_GATE_DISABLE |
7591 I965_RCPB_CLOCK_GATE_DISABLE |
7592 I965_ISC_CLOCK_GATE_DISABLE |
7593 I965_FBC_CLOCK_GATE_DISABLE);
7594 I915_WRITE(RENCLK_GATE_D2, 0);
7595 I915_WRITE(MI_ARB_STATE,
7596 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7598 /* WaDisable_RenderCache_OperationalFlush:gen4 */
7599 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7602 static void gen3_init_clock_gating(struct drm_device *dev)
7604 struct drm_i915_private *dev_priv = to_i915(dev);
7605 u32 dstate = I915_READ(D_STATE);
7607 dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
7608 DSTATE_DOT_CLOCK_GATING;
7609 I915_WRITE(D_STATE, dstate);
7611 if (IS_PINEVIEW(dev))
7612 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
7614 /* IIR "flip pending" means done if this bit is set */
7615 I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
7617 /* interrupts should cause a wake up from C3 */
7618 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
7620 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
7621 I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
7623 I915_WRITE(MI_ARB_STATE,
7624 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7627 static void i85x_init_clock_gating(struct drm_device *dev)
7629 struct drm_i915_private *dev_priv = to_i915(dev);
7631 I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
7633 /* interrupts should cause a wake up from C3 */
7634 I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
7635 _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
7637 I915_WRITE(MEM_MODE,
7638 _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE));
7641 static void i830_init_clock_gating(struct drm_device *dev)
7643 struct drm_i915_private *dev_priv = to_i915(dev);
7645 I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
7647 I915_WRITE(MEM_MODE,
7648 _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE) |
7649 _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE));
7652 void intel_init_clock_gating(struct drm_device *dev)
7654 struct drm_i915_private *dev_priv = to_i915(dev);
7656 dev_priv->display.init_clock_gating(dev);
7659 void intel_suspend_hw(struct drm_device *dev)
7661 if (HAS_PCH_LPT(to_i915(dev)))
7662 lpt_suspend_hw(dev);
7665 static void nop_init_clock_gating(struct drm_device *dev)
7667 DRM_DEBUG_KMS("No clock gating settings or workarounds applied.\n");
7671 * intel_init_clock_gating_hooks - setup the clock gating hooks
7672 * @dev_priv: device private
7674 * Setup the hooks that configure which clocks of a given platform can be
7675 * gated and also apply various GT and display specific workarounds for these
7676 * platforms. Note that some GT specific workarounds are applied separately
7677 * when GPU contexts or batchbuffers start their execution.
7679 void intel_init_clock_gating_hooks(struct drm_i915_private *dev_priv)
7681 if (IS_SKYLAKE(dev_priv))
7682 dev_priv->display.init_clock_gating = skylake_init_clock_gating;
7683 else if (IS_KABYLAKE(dev_priv))
7684 dev_priv->display.init_clock_gating = kabylake_init_clock_gating;
7685 else if (IS_BROXTON(dev_priv))
7686 dev_priv->display.init_clock_gating = bxt_init_clock_gating;
7687 else if (IS_BROADWELL(dev_priv))
7688 dev_priv->display.init_clock_gating = broadwell_init_clock_gating;
7689 else if (IS_CHERRYVIEW(dev_priv))
7690 dev_priv->display.init_clock_gating = cherryview_init_clock_gating;
7691 else if (IS_HASWELL(dev_priv))
7692 dev_priv->display.init_clock_gating = haswell_init_clock_gating;
7693 else if (IS_IVYBRIDGE(dev_priv))
7694 dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
7695 else if (IS_VALLEYVIEW(dev_priv))
7696 dev_priv->display.init_clock_gating = valleyview_init_clock_gating;
7697 else if (IS_GEN6(dev_priv))
7698 dev_priv->display.init_clock_gating = gen6_init_clock_gating;
7699 else if (IS_GEN5(dev_priv))
7700 dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
7701 else if (IS_G4X(dev_priv))
7702 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
7703 else if (IS_CRESTLINE(dev_priv))
7704 dev_priv->display.init_clock_gating = crestline_init_clock_gating;
7705 else if (IS_BROADWATER(dev_priv))
7706 dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
7707 else if (IS_GEN3(dev_priv))
7708 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
7709 else if (IS_I85X(dev_priv) || IS_I865G(dev_priv))
7710 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
7711 else if (IS_GEN2(dev_priv))
7712 dev_priv->display.init_clock_gating = i830_init_clock_gating;
7714 MISSING_CASE(INTEL_DEVID(dev_priv));
7715 dev_priv->display.init_clock_gating = nop_init_clock_gating;
7719 /* Set up chip specific power management-related functions */
7720 void intel_init_pm(struct drm_device *dev)
7722 struct drm_i915_private *dev_priv = to_i915(dev);
7724 intel_fbc_init(dev_priv);
7727 if (IS_PINEVIEW(dev))
7728 i915_pineview_get_mem_freq(dev);
7729 else if (IS_GEN5(dev_priv))
7730 i915_ironlake_get_mem_freq(dev);
7732 /* For FIFO watermark updates */
7733 if (INTEL_INFO(dev)->gen >= 9) {
7734 skl_setup_wm_latency(dev);
7735 dev_priv->display.update_wm = skl_update_wm;
7736 dev_priv->display.compute_global_watermarks = skl_compute_wm;
7737 } else if (HAS_PCH_SPLIT(dev_priv)) {
7738 ilk_setup_wm_latency(dev);
7740 if ((IS_GEN5(dev_priv) && dev_priv->wm.pri_latency[1] &&
7741 dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
7742 (!IS_GEN5(dev_priv) && dev_priv->wm.pri_latency[0] &&
7743 dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
7744 dev_priv->display.compute_pipe_wm = ilk_compute_pipe_wm;
7745 dev_priv->display.compute_intermediate_wm =
7746 ilk_compute_intermediate_wm;
7747 dev_priv->display.initial_watermarks =
7748 ilk_initial_watermarks;
7749 dev_priv->display.optimize_watermarks =
7750 ilk_optimize_watermarks;
7752 DRM_DEBUG_KMS("Failed to read display plane latency. "
7755 } else if (IS_CHERRYVIEW(dev_priv)) {
7756 vlv_setup_wm_latency(dev);
7757 dev_priv->display.update_wm = vlv_update_wm;
7758 } else if (IS_VALLEYVIEW(dev_priv)) {
7759 vlv_setup_wm_latency(dev);
7760 dev_priv->display.update_wm = vlv_update_wm;
7761 } else if (IS_PINEVIEW(dev)) {
7762 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv),
7765 dev_priv->mem_freq)) {
7766 DRM_INFO("failed to find known CxSR latency "
7767 "(found ddr%s fsb freq %d, mem freq %d), "
7769 (dev_priv->is_ddr3 == 1) ? "3" : "2",
7770 dev_priv->fsb_freq, dev_priv->mem_freq);
7771 /* Disable CxSR and never update its watermark again */
7772 intel_set_memory_cxsr(dev_priv, false);
7773 dev_priv->display.update_wm = NULL;
7775 dev_priv->display.update_wm = pineview_update_wm;
7776 } else if (IS_G4X(dev_priv)) {
7777 dev_priv->display.update_wm = g4x_update_wm;
7778 } else if (IS_GEN4(dev_priv)) {
7779 dev_priv->display.update_wm = i965_update_wm;
7780 } else if (IS_GEN3(dev_priv)) {
7781 dev_priv->display.update_wm = i9xx_update_wm;
7782 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
7783 } else if (IS_GEN2(dev_priv)) {
7784 if (INTEL_INFO(dev)->num_pipes == 1) {
7785 dev_priv->display.update_wm = i845_update_wm;
7786 dev_priv->display.get_fifo_size = i845_get_fifo_size;
7788 dev_priv->display.update_wm = i9xx_update_wm;
7789 dev_priv->display.get_fifo_size = i830_get_fifo_size;
7792 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
7796 static inline int gen6_check_mailbox_status(struct drm_i915_private *dev_priv)
7799 I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_ERROR_MASK;
7802 case GEN6_PCODE_SUCCESS:
7804 case GEN6_PCODE_UNIMPLEMENTED_CMD:
7805 case GEN6_PCODE_ILLEGAL_CMD:
7807 case GEN6_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
7808 case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
7810 case GEN6_PCODE_TIMEOUT:
7818 static inline int gen7_check_mailbox_status(struct drm_i915_private *dev_priv)
7821 I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_ERROR_MASK;
7824 case GEN6_PCODE_SUCCESS:
7826 case GEN6_PCODE_ILLEGAL_CMD:
7828 case GEN7_PCODE_TIMEOUT:
7830 case GEN7_PCODE_ILLEGAL_DATA:
7832 case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
7835 MISSING_CASE(flags);
7840 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u32 mbox, u32 *val)
7844 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7846 /* GEN6_PCODE_* are outside of the forcewake domain, we can
7847 * use te fw I915_READ variants to reduce the amount of work
7848 * required when reading/writing.
7851 if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
7852 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
7856 I915_WRITE_FW(GEN6_PCODE_DATA, *val);
7857 I915_WRITE_FW(GEN6_PCODE_DATA1, 0);
7858 I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
7860 if (intel_wait_for_register_fw(dev_priv,
7861 GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
7863 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
7867 *val = I915_READ_FW(GEN6_PCODE_DATA);
7868 I915_WRITE_FW(GEN6_PCODE_DATA, 0);
7870 if (INTEL_GEN(dev_priv) > 6)
7871 status = gen7_check_mailbox_status(dev_priv);
7873 status = gen6_check_mailbox_status(dev_priv);
7876 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed: %d\n",
7884 int sandybridge_pcode_write(struct drm_i915_private *dev_priv,
7889 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7891 /* GEN6_PCODE_* are outside of the forcewake domain, we can
7892 * use te fw I915_READ variants to reduce the amount of work
7893 * required when reading/writing.
7896 if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
7897 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
7901 I915_WRITE_FW(GEN6_PCODE_DATA, val);
7902 I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
7904 if (intel_wait_for_register_fw(dev_priv,
7905 GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
7907 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
7911 I915_WRITE_FW(GEN6_PCODE_DATA, 0);
7913 if (INTEL_GEN(dev_priv) > 6)
7914 status = gen7_check_mailbox_status(dev_priv);
7916 status = gen6_check_mailbox_status(dev_priv);
7919 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed: %d\n",
7927 static int byt_gpu_freq(struct drm_i915_private *dev_priv, int val)
7931 * Slow = Fast = GPLL ref * N
7933 return DIV_ROUND_CLOSEST(dev_priv->rps.gpll_ref_freq * (val - 0xb7), 1000);
7936 static int byt_freq_opcode(struct drm_i915_private *dev_priv, int val)
7938 return DIV_ROUND_CLOSEST(1000 * val, dev_priv->rps.gpll_ref_freq) + 0xb7;
7941 static int chv_gpu_freq(struct drm_i915_private *dev_priv, int val)
7945 * CU (slow) = CU2x (fast) / 2 = GPLL ref * N / 2
7947 return DIV_ROUND_CLOSEST(dev_priv->rps.gpll_ref_freq * val, 2 * 2 * 1000);
7950 static int chv_freq_opcode(struct drm_i915_private *dev_priv, int val)
7952 /* CHV needs even values */
7953 return DIV_ROUND_CLOSEST(2 * 1000 * val, dev_priv->rps.gpll_ref_freq) * 2;
7956 int intel_gpu_freq(struct drm_i915_private *dev_priv, int val)
7958 if (IS_GEN9(dev_priv))
7959 return DIV_ROUND_CLOSEST(val * GT_FREQUENCY_MULTIPLIER,
7961 else if (IS_CHERRYVIEW(dev_priv))
7962 return chv_gpu_freq(dev_priv, val);
7963 else if (IS_VALLEYVIEW(dev_priv))
7964 return byt_gpu_freq(dev_priv, val);
7966 return val * GT_FREQUENCY_MULTIPLIER;
7969 int intel_freq_opcode(struct drm_i915_private *dev_priv, int val)
7971 if (IS_GEN9(dev_priv))
7972 return DIV_ROUND_CLOSEST(val * GEN9_FREQ_SCALER,
7973 GT_FREQUENCY_MULTIPLIER);
7974 else if (IS_CHERRYVIEW(dev_priv))
7975 return chv_freq_opcode(dev_priv, val);
7976 else if (IS_VALLEYVIEW(dev_priv))
7977 return byt_freq_opcode(dev_priv, val);
7979 return DIV_ROUND_CLOSEST(val, GT_FREQUENCY_MULTIPLIER);
7982 struct request_boost {
7983 struct work_struct work;
7984 struct drm_i915_gem_request *req;
7987 static void __intel_rps_boost_work(struct work_struct *work)
7989 struct request_boost *boost = container_of(work, struct request_boost, work);
7990 struct drm_i915_gem_request *req = boost->req;
7992 if (!i915_gem_request_completed(req))
7993 gen6_rps_boost(req->i915, NULL, req->emitted_jiffies);
7995 i915_gem_request_put(req);
7999 void intel_queue_rps_boost_for_request(struct drm_i915_gem_request *req)
8001 struct request_boost *boost;
8003 if (req == NULL || INTEL_GEN(req->i915) < 6)
8006 if (i915_gem_request_completed(req))
8009 boost = kmalloc(sizeof(*boost), GFP_ATOMIC);
8013 boost->req = i915_gem_request_get(req);
8015 INIT_WORK(&boost->work, __intel_rps_boost_work);
8016 queue_work(req->i915->wq, &boost->work);
8019 void intel_pm_setup(struct drm_device *dev)
8021 struct drm_i915_private *dev_priv = to_i915(dev);
8023 mutex_init(&dev_priv->rps.hw_lock);
8024 spin_lock_init(&dev_priv->rps.client_lock);
8026 INIT_DELAYED_WORK(&dev_priv->rps.autoenable_work,
8027 __intel_autoenable_gt_powersave);
8028 INIT_LIST_HEAD(&dev_priv->rps.clients);
8030 dev_priv->pm.suspended = false;
8031 atomic_set(&dev_priv->pm.wakeref_count, 0);
8032 atomic_set(&dev_priv->pm.atomic_seq, 0);