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>
30 #include "intel_drv.h"
31 #include "../../../platform/x86/intel_ips.h"
32 #include <linux/module.h>
35 * RC6 is a special power stage which allows the GPU to enter an very
36 * low-voltage mode when idle, using down to 0V while at this stage. This
37 * stage is entered automatically when the GPU is idle when RC6 support is
38 * enabled, and as soon as new workload arises GPU wakes up automatically as well.
40 * There are different RC6 modes available in Intel GPU, which differentiate
41 * among each other with the latency required to enter and leave RC6 and
42 * voltage consumed by the GPU in different states.
44 * The combination of the following flags define which states GPU is allowed
45 * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
46 * RC6pp is deepest RC6. Their support by hardware varies according to the
47 * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
48 * which brings the most power savings; deeper states save more power, but
49 * require higher latency to switch to and wake up.
51 #define INTEL_RC6_ENABLE (1<<0)
52 #define INTEL_RC6p_ENABLE (1<<1)
53 #define INTEL_RC6pp_ENABLE (1<<2)
55 static void gen9_init_clock_gating(struct drm_device *dev)
57 struct drm_i915_private *dev_priv = dev->dev_private;
59 /* WaEnableLbsSlaRetryTimerDecrement:skl */
60 I915_WRITE(BDW_SCRATCH1, I915_READ(BDW_SCRATCH1) |
61 GEN9_LBS_SLA_RETRY_TIMER_DECREMENT_ENABLE);
64 static void skl_init_clock_gating(struct drm_device *dev)
66 struct drm_i915_private *dev_priv = dev->dev_private;
68 gen9_init_clock_gating(dev);
70 if (INTEL_REVID(dev) == SKL_REVID_A0) {
72 * WaDisableSDEUnitClockGating:skl
73 * WaSetGAPSunitClckGateDisable:skl
75 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
76 GEN8_GAPSUNIT_CLOCK_GATE_DISABLE |
77 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
80 if (INTEL_REVID(dev) <= SKL_REVID_D0) {
81 /* WaDisableHDCInvalidation:skl */
82 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) |
83 BDW_DISABLE_HDC_INVALIDATION);
85 /* WaDisableChickenBitTSGBarrierAckForFFSliceCS:skl */
86 I915_WRITE(FF_SLICE_CS_CHICKEN2,
87 I915_READ(FF_SLICE_CS_CHICKEN2) |
88 GEN9_TSG_BARRIER_ACK_DISABLE);
91 if (INTEL_REVID(dev) <= SKL_REVID_E0)
92 /* WaDisableLSQCROPERFforOCL:skl */
93 I915_WRITE(GEN8_L3SQCREG4, I915_READ(GEN8_L3SQCREG4) |
94 GEN8_LQSC_RO_PERF_DIS);
97 static void i915_pineview_get_mem_freq(struct drm_device *dev)
99 struct drm_i915_private *dev_priv = dev->dev_private;
102 tmp = I915_READ(CLKCFG);
104 switch (tmp & CLKCFG_FSB_MASK) {
106 dev_priv->fsb_freq = 533; /* 133*4 */
109 dev_priv->fsb_freq = 800; /* 200*4 */
112 dev_priv->fsb_freq = 667; /* 167*4 */
115 dev_priv->fsb_freq = 400; /* 100*4 */
119 switch (tmp & CLKCFG_MEM_MASK) {
121 dev_priv->mem_freq = 533;
124 dev_priv->mem_freq = 667;
127 dev_priv->mem_freq = 800;
131 /* detect pineview DDR3 setting */
132 tmp = I915_READ(CSHRDDR3CTL);
133 dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
136 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
138 struct drm_i915_private *dev_priv = dev->dev_private;
141 ddrpll = I915_READ16(DDRMPLL1);
142 csipll = I915_READ16(CSIPLL0);
144 switch (ddrpll & 0xff) {
146 dev_priv->mem_freq = 800;
149 dev_priv->mem_freq = 1066;
152 dev_priv->mem_freq = 1333;
155 dev_priv->mem_freq = 1600;
158 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
160 dev_priv->mem_freq = 0;
164 dev_priv->ips.r_t = dev_priv->mem_freq;
166 switch (csipll & 0x3ff) {
168 dev_priv->fsb_freq = 3200;
171 dev_priv->fsb_freq = 3733;
174 dev_priv->fsb_freq = 4266;
177 dev_priv->fsb_freq = 4800;
180 dev_priv->fsb_freq = 5333;
183 dev_priv->fsb_freq = 5866;
186 dev_priv->fsb_freq = 6400;
189 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
191 dev_priv->fsb_freq = 0;
195 if (dev_priv->fsb_freq == 3200) {
196 dev_priv->ips.c_m = 0;
197 } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
198 dev_priv->ips.c_m = 1;
200 dev_priv->ips.c_m = 2;
204 static const struct cxsr_latency cxsr_latency_table[] = {
205 {1, 0, 800, 400, 3382, 33382, 3983, 33983}, /* DDR2-400 SC */
206 {1, 0, 800, 667, 3354, 33354, 3807, 33807}, /* DDR2-667 SC */
207 {1, 0, 800, 800, 3347, 33347, 3763, 33763}, /* DDR2-800 SC */
208 {1, 1, 800, 667, 6420, 36420, 6873, 36873}, /* DDR3-667 SC */
209 {1, 1, 800, 800, 5902, 35902, 6318, 36318}, /* DDR3-800 SC */
211 {1, 0, 667, 400, 3400, 33400, 4021, 34021}, /* DDR2-400 SC */
212 {1, 0, 667, 667, 3372, 33372, 3845, 33845}, /* DDR2-667 SC */
213 {1, 0, 667, 800, 3386, 33386, 3822, 33822}, /* DDR2-800 SC */
214 {1, 1, 667, 667, 6438, 36438, 6911, 36911}, /* DDR3-667 SC */
215 {1, 1, 667, 800, 5941, 35941, 6377, 36377}, /* DDR3-800 SC */
217 {1, 0, 400, 400, 3472, 33472, 4173, 34173}, /* DDR2-400 SC */
218 {1, 0, 400, 667, 3443, 33443, 3996, 33996}, /* DDR2-667 SC */
219 {1, 0, 400, 800, 3430, 33430, 3946, 33946}, /* DDR2-800 SC */
220 {1, 1, 400, 667, 6509, 36509, 7062, 37062}, /* DDR3-667 SC */
221 {1, 1, 400, 800, 5985, 35985, 6501, 36501}, /* DDR3-800 SC */
223 {0, 0, 800, 400, 3438, 33438, 4065, 34065}, /* DDR2-400 SC */
224 {0, 0, 800, 667, 3410, 33410, 3889, 33889}, /* DDR2-667 SC */
225 {0, 0, 800, 800, 3403, 33403, 3845, 33845}, /* DDR2-800 SC */
226 {0, 1, 800, 667, 6476, 36476, 6955, 36955}, /* DDR3-667 SC */
227 {0, 1, 800, 800, 5958, 35958, 6400, 36400}, /* DDR3-800 SC */
229 {0, 0, 667, 400, 3456, 33456, 4103, 34106}, /* DDR2-400 SC */
230 {0, 0, 667, 667, 3428, 33428, 3927, 33927}, /* DDR2-667 SC */
231 {0, 0, 667, 800, 3443, 33443, 3905, 33905}, /* DDR2-800 SC */
232 {0, 1, 667, 667, 6494, 36494, 6993, 36993}, /* DDR3-667 SC */
233 {0, 1, 667, 800, 5998, 35998, 6460, 36460}, /* DDR3-800 SC */
235 {0, 0, 400, 400, 3528, 33528, 4255, 34255}, /* DDR2-400 SC */
236 {0, 0, 400, 667, 3500, 33500, 4079, 34079}, /* DDR2-667 SC */
237 {0, 0, 400, 800, 3487, 33487, 4029, 34029}, /* DDR2-800 SC */
238 {0, 1, 400, 667, 6566, 36566, 7145, 37145}, /* DDR3-667 SC */
239 {0, 1, 400, 800, 6042, 36042, 6584, 36584}, /* DDR3-800 SC */
242 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
247 const struct cxsr_latency *latency;
250 if (fsb == 0 || mem == 0)
253 for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
254 latency = &cxsr_latency_table[i];
255 if (is_desktop == latency->is_desktop &&
256 is_ddr3 == latency->is_ddr3 &&
257 fsb == latency->fsb_freq && mem == latency->mem_freq)
261 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
266 static void chv_set_memory_dvfs(struct drm_i915_private *dev_priv, bool enable)
270 mutex_lock(&dev_priv->rps.hw_lock);
272 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
274 val &= ~FORCE_DDR_HIGH_FREQ;
276 val |= FORCE_DDR_HIGH_FREQ;
277 val &= ~FORCE_DDR_LOW_FREQ;
278 val |= FORCE_DDR_FREQ_REQ_ACK;
279 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
281 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
282 FORCE_DDR_FREQ_REQ_ACK) == 0, 3))
283 DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
285 mutex_unlock(&dev_priv->rps.hw_lock);
288 static void chv_set_memory_pm5(struct drm_i915_private *dev_priv, bool enable)
292 mutex_lock(&dev_priv->rps.hw_lock);
294 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
296 val |= DSP_MAXFIFO_PM5_ENABLE;
298 val &= ~DSP_MAXFIFO_PM5_ENABLE;
299 vlv_punit_write(dev_priv, PUNIT_REG_DSPFREQ, val);
301 mutex_unlock(&dev_priv->rps.hw_lock);
304 #define FW_WM(value, plane) \
305 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
307 void intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
309 struct drm_device *dev = dev_priv->dev;
312 if (IS_VALLEYVIEW(dev)) {
313 I915_WRITE(FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
314 if (IS_CHERRYVIEW(dev))
315 chv_set_memory_pm5(dev_priv, enable);
316 } else if (IS_G4X(dev) || IS_CRESTLINE(dev)) {
317 I915_WRITE(FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
318 } else if (IS_PINEVIEW(dev)) {
319 val = I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN;
320 val |= enable ? PINEVIEW_SELF_REFRESH_EN : 0;
321 I915_WRITE(DSPFW3, val);
322 } else if (IS_I945G(dev) || IS_I945GM(dev)) {
323 val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
324 _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
325 I915_WRITE(FW_BLC_SELF, val);
326 } else if (IS_I915GM(dev)) {
327 val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
328 _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
329 I915_WRITE(INSTPM, val);
334 DRM_DEBUG_KMS("memory self-refresh is %s\n",
335 enable ? "enabled" : "disabled");
340 * Latency for FIFO fetches is dependent on several factors:
341 * - memory configuration (speed, channels)
343 * - current MCH state
344 * It can be fairly high in some situations, so here we assume a fairly
345 * pessimal value. It's a tradeoff between extra memory fetches (if we
346 * set this value too high, the FIFO will fetch frequently to stay full)
347 * and power consumption (set it too low to save power and we might see
348 * FIFO underruns and display "flicker").
350 * A value of 5us seems to be a good balance; safe for very low end
351 * platforms but not overly aggressive on lower latency configs.
353 static const int pessimal_latency_ns = 5000;
355 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
356 ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
358 static int vlv_get_fifo_size(struct drm_device *dev,
359 enum pipe pipe, int plane)
361 struct drm_i915_private *dev_priv = dev->dev_private;
362 int sprite0_start, sprite1_start, size;
365 uint32_t dsparb, dsparb2, dsparb3;
367 dsparb = I915_READ(DSPARB);
368 dsparb2 = I915_READ(DSPARB2);
369 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 0, 0);
370 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 8, 4);
373 dsparb = I915_READ(DSPARB);
374 dsparb2 = I915_READ(DSPARB2);
375 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 16, 8);
376 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 24, 12);
379 dsparb2 = I915_READ(DSPARB2);
380 dsparb3 = I915_READ(DSPARB3);
381 sprite0_start = VLV_FIFO_START(dsparb3, dsparb2, 0, 16);
382 sprite1_start = VLV_FIFO_START(dsparb3, dsparb2, 8, 20);
390 size = sprite0_start;
393 size = sprite1_start - sprite0_start;
396 size = 512 - 1 - sprite1_start;
402 DRM_DEBUG_KMS("Pipe %c %s %c FIFO size: %d\n",
403 pipe_name(pipe), plane == 0 ? "primary" : "sprite",
404 plane == 0 ? plane_name(pipe) : sprite_name(pipe, plane - 1),
410 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
412 struct drm_i915_private *dev_priv = dev->dev_private;
413 uint32_t dsparb = I915_READ(DSPARB);
416 size = dsparb & 0x7f;
418 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
420 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
421 plane ? "B" : "A", size);
426 static int i830_get_fifo_size(struct drm_device *dev, int plane)
428 struct drm_i915_private *dev_priv = dev->dev_private;
429 uint32_t dsparb = I915_READ(DSPARB);
432 size = dsparb & 0x1ff;
434 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
435 size >>= 1; /* Convert to cachelines */
437 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
438 plane ? "B" : "A", size);
443 static int i845_get_fifo_size(struct drm_device *dev, int plane)
445 struct drm_i915_private *dev_priv = dev->dev_private;
446 uint32_t dsparb = I915_READ(DSPARB);
449 size = dsparb & 0x7f;
450 size >>= 2; /* Convert to cachelines */
452 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
459 /* Pineview has different values for various configs */
460 static const struct intel_watermark_params pineview_display_wm = {
461 .fifo_size = PINEVIEW_DISPLAY_FIFO,
462 .max_wm = PINEVIEW_MAX_WM,
463 .default_wm = PINEVIEW_DFT_WM,
464 .guard_size = PINEVIEW_GUARD_WM,
465 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
467 static const struct intel_watermark_params pineview_display_hplloff_wm = {
468 .fifo_size = PINEVIEW_DISPLAY_FIFO,
469 .max_wm = PINEVIEW_MAX_WM,
470 .default_wm = PINEVIEW_DFT_HPLLOFF_WM,
471 .guard_size = PINEVIEW_GUARD_WM,
472 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
474 static const struct intel_watermark_params pineview_cursor_wm = {
475 .fifo_size = PINEVIEW_CURSOR_FIFO,
476 .max_wm = PINEVIEW_CURSOR_MAX_WM,
477 .default_wm = PINEVIEW_CURSOR_DFT_WM,
478 .guard_size = PINEVIEW_CURSOR_GUARD_WM,
479 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
481 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
482 .fifo_size = PINEVIEW_CURSOR_FIFO,
483 .max_wm = PINEVIEW_CURSOR_MAX_WM,
484 .default_wm = PINEVIEW_CURSOR_DFT_WM,
485 .guard_size = PINEVIEW_CURSOR_GUARD_WM,
486 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
488 static const struct intel_watermark_params g4x_wm_info = {
489 .fifo_size = G4X_FIFO_SIZE,
490 .max_wm = G4X_MAX_WM,
491 .default_wm = G4X_MAX_WM,
493 .cacheline_size = G4X_FIFO_LINE_SIZE,
495 static const struct intel_watermark_params g4x_cursor_wm_info = {
496 .fifo_size = I965_CURSOR_FIFO,
497 .max_wm = I965_CURSOR_MAX_WM,
498 .default_wm = I965_CURSOR_DFT_WM,
500 .cacheline_size = G4X_FIFO_LINE_SIZE,
502 static const struct intel_watermark_params valleyview_wm_info = {
503 .fifo_size = VALLEYVIEW_FIFO_SIZE,
504 .max_wm = VALLEYVIEW_MAX_WM,
505 .default_wm = VALLEYVIEW_MAX_WM,
507 .cacheline_size = G4X_FIFO_LINE_SIZE,
509 static const struct intel_watermark_params valleyview_cursor_wm_info = {
510 .fifo_size = I965_CURSOR_FIFO,
511 .max_wm = VALLEYVIEW_CURSOR_MAX_WM,
512 .default_wm = I965_CURSOR_DFT_WM,
514 .cacheline_size = G4X_FIFO_LINE_SIZE,
516 static const struct intel_watermark_params i965_cursor_wm_info = {
517 .fifo_size = I965_CURSOR_FIFO,
518 .max_wm = I965_CURSOR_MAX_WM,
519 .default_wm = I965_CURSOR_DFT_WM,
521 .cacheline_size = I915_FIFO_LINE_SIZE,
523 static const struct intel_watermark_params i945_wm_info = {
524 .fifo_size = I945_FIFO_SIZE,
525 .max_wm = I915_MAX_WM,
528 .cacheline_size = I915_FIFO_LINE_SIZE,
530 static const struct intel_watermark_params i915_wm_info = {
531 .fifo_size = I915_FIFO_SIZE,
532 .max_wm = I915_MAX_WM,
535 .cacheline_size = I915_FIFO_LINE_SIZE,
537 static const struct intel_watermark_params i830_a_wm_info = {
538 .fifo_size = I855GM_FIFO_SIZE,
539 .max_wm = I915_MAX_WM,
542 .cacheline_size = I830_FIFO_LINE_SIZE,
544 static const struct intel_watermark_params i830_bc_wm_info = {
545 .fifo_size = I855GM_FIFO_SIZE,
546 .max_wm = I915_MAX_WM/2,
549 .cacheline_size = I830_FIFO_LINE_SIZE,
551 static const struct intel_watermark_params i845_wm_info = {
552 .fifo_size = I830_FIFO_SIZE,
553 .max_wm = I915_MAX_WM,
556 .cacheline_size = I830_FIFO_LINE_SIZE,
560 * intel_calculate_wm - calculate watermark level
561 * @clock_in_khz: pixel clock
562 * @wm: chip FIFO params
563 * @pixel_size: display pixel size
564 * @latency_ns: memory latency for the platform
566 * Calculate the watermark level (the level at which the display plane will
567 * start fetching from memory again). Each chip has a different display
568 * FIFO size and allocation, so the caller needs to figure that out and pass
569 * in the correct intel_watermark_params structure.
571 * As the pixel clock runs, the FIFO will be drained at a rate that depends
572 * on the pixel size. When it reaches the watermark level, it'll start
573 * fetching FIFO line sized based chunks from memory until the FIFO fills
574 * past the watermark point. If the FIFO drains completely, a FIFO underrun
575 * will occur, and a display engine hang could result.
577 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
578 const struct intel_watermark_params *wm,
581 unsigned long latency_ns)
583 long entries_required, wm_size;
586 * Note: we need to make sure we don't overflow for various clock &
588 * clocks go from a few thousand to several hundred thousand.
589 * latency is usually a few thousand
591 entries_required = ((clock_in_khz / 1000) * pixel_size * latency_ns) /
593 entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
595 DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
597 wm_size = fifo_size - (entries_required + wm->guard_size);
599 DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
601 /* Don't promote wm_size to unsigned... */
602 if (wm_size > (long)wm->max_wm)
603 wm_size = wm->max_wm;
605 wm_size = wm->default_wm;
608 * Bspec seems to indicate that the value shouldn't be lower than
609 * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
610 * Lets go for 8 which is the burst size since certain platforms
611 * already use a hardcoded 8 (which is what the spec says should be
620 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
622 struct drm_crtc *crtc, *enabled = NULL;
624 for_each_crtc(dev, crtc) {
625 if (intel_crtc_active(crtc)) {
635 static void pineview_update_wm(struct drm_crtc *unused_crtc)
637 struct drm_device *dev = unused_crtc->dev;
638 struct drm_i915_private *dev_priv = dev->dev_private;
639 struct drm_crtc *crtc;
640 const struct cxsr_latency *latency;
644 latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
645 dev_priv->fsb_freq, dev_priv->mem_freq);
647 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
648 intel_set_memory_cxsr(dev_priv, false);
652 crtc = single_enabled_crtc(dev);
654 const struct drm_display_mode *adjusted_mode;
655 int pixel_size = crtc->primary->state->fb->bits_per_pixel / 8;
658 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
659 clock = adjusted_mode->crtc_clock;
662 wm = intel_calculate_wm(clock, &pineview_display_wm,
663 pineview_display_wm.fifo_size,
664 pixel_size, latency->display_sr);
665 reg = I915_READ(DSPFW1);
666 reg &= ~DSPFW_SR_MASK;
667 reg |= FW_WM(wm, SR);
668 I915_WRITE(DSPFW1, reg);
669 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
672 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
673 pineview_display_wm.fifo_size,
674 pixel_size, latency->cursor_sr);
675 reg = I915_READ(DSPFW3);
676 reg &= ~DSPFW_CURSOR_SR_MASK;
677 reg |= FW_WM(wm, CURSOR_SR);
678 I915_WRITE(DSPFW3, reg);
680 /* Display HPLL off SR */
681 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
682 pineview_display_hplloff_wm.fifo_size,
683 pixel_size, latency->display_hpll_disable);
684 reg = I915_READ(DSPFW3);
685 reg &= ~DSPFW_HPLL_SR_MASK;
686 reg |= FW_WM(wm, HPLL_SR);
687 I915_WRITE(DSPFW3, reg);
689 /* cursor HPLL off SR */
690 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
691 pineview_display_hplloff_wm.fifo_size,
692 pixel_size, latency->cursor_hpll_disable);
693 reg = I915_READ(DSPFW3);
694 reg &= ~DSPFW_HPLL_CURSOR_MASK;
695 reg |= FW_WM(wm, HPLL_CURSOR);
696 I915_WRITE(DSPFW3, reg);
697 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
699 intel_set_memory_cxsr(dev_priv, true);
701 intel_set_memory_cxsr(dev_priv, false);
705 static bool g4x_compute_wm0(struct drm_device *dev,
707 const struct intel_watermark_params *display,
708 int display_latency_ns,
709 const struct intel_watermark_params *cursor,
710 int cursor_latency_ns,
714 struct drm_crtc *crtc;
715 const struct drm_display_mode *adjusted_mode;
716 int htotal, hdisplay, clock, pixel_size;
717 int line_time_us, line_count;
718 int entries, tlb_miss;
720 crtc = intel_get_crtc_for_plane(dev, plane);
721 if (!intel_crtc_active(crtc)) {
722 *cursor_wm = cursor->guard_size;
723 *plane_wm = display->guard_size;
727 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
728 clock = adjusted_mode->crtc_clock;
729 htotal = adjusted_mode->crtc_htotal;
730 hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
731 pixel_size = crtc->primary->state->fb->bits_per_pixel / 8;
733 /* Use the small buffer method to calculate plane watermark */
734 entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
735 tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
738 entries = DIV_ROUND_UP(entries, display->cacheline_size);
739 *plane_wm = entries + display->guard_size;
740 if (*plane_wm > (int)display->max_wm)
741 *plane_wm = display->max_wm;
743 /* Use the large buffer method to calculate cursor watermark */
744 line_time_us = max(htotal * 1000 / clock, 1);
745 line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
746 entries = line_count * crtc->cursor->state->crtc_w * pixel_size;
747 tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
750 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
751 *cursor_wm = entries + cursor->guard_size;
752 if (*cursor_wm > (int)cursor->max_wm)
753 *cursor_wm = (int)cursor->max_wm;
759 * Check the wm result.
761 * If any calculated watermark values is larger than the maximum value that
762 * can be programmed into the associated watermark register, that watermark
765 static bool g4x_check_srwm(struct drm_device *dev,
766 int display_wm, int cursor_wm,
767 const struct intel_watermark_params *display,
768 const struct intel_watermark_params *cursor)
770 DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
771 display_wm, cursor_wm);
773 if (display_wm > display->max_wm) {
774 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
775 display_wm, display->max_wm);
779 if (cursor_wm > cursor->max_wm) {
780 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
781 cursor_wm, cursor->max_wm);
785 if (!(display_wm || cursor_wm)) {
786 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
793 static bool g4x_compute_srwm(struct drm_device *dev,
796 const struct intel_watermark_params *display,
797 const struct intel_watermark_params *cursor,
798 int *display_wm, int *cursor_wm)
800 struct drm_crtc *crtc;
801 const struct drm_display_mode *adjusted_mode;
802 int hdisplay, htotal, pixel_size, clock;
803 unsigned long line_time_us;
804 int line_count, line_size;
809 *display_wm = *cursor_wm = 0;
813 crtc = intel_get_crtc_for_plane(dev, plane);
814 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
815 clock = adjusted_mode->crtc_clock;
816 htotal = adjusted_mode->crtc_htotal;
817 hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
818 pixel_size = crtc->primary->state->fb->bits_per_pixel / 8;
820 line_time_us = max(htotal * 1000 / clock, 1);
821 line_count = (latency_ns / line_time_us + 1000) / 1000;
822 line_size = hdisplay * pixel_size;
824 /* Use the minimum of the small and large buffer method for primary */
825 small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
826 large = line_count * line_size;
828 entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
829 *display_wm = entries + display->guard_size;
831 /* calculate the self-refresh watermark for display cursor */
832 entries = line_count * pixel_size * crtc->cursor->state->crtc_w;
833 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
834 *cursor_wm = entries + cursor->guard_size;
836 return g4x_check_srwm(dev,
837 *display_wm, *cursor_wm,
841 #define FW_WM_VLV(value, plane) \
842 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
844 static void vlv_write_wm_values(struct intel_crtc *crtc,
845 const struct vlv_wm_values *wm)
847 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
848 enum pipe pipe = crtc->pipe;
850 I915_WRITE(VLV_DDL(pipe),
851 (wm->ddl[pipe].cursor << DDL_CURSOR_SHIFT) |
852 (wm->ddl[pipe].sprite[1] << DDL_SPRITE_SHIFT(1)) |
853 (wm->ddl[pipe].sprite[0] << DDL_SPRITE_SHIFT(0)) |
854 (wm->ddl[pipe].primary << DDL_PLANE_SHIFT));
857 FW_WM(wm->sr.plane, SR) |
858 FW_WM(wm->pipe[PIPE_B].cursor, CURSORB) |
859 FW_WM_VLV(wm->pipe[PIPE_B].primary, PLANEB) |
860 FW_WM_VLV(wm->pipe[PIPE_A].primary, PLANEA));
862 FW_WM_VLV(wm->pipe[PIPE_A].sprite[1], SPRITEB) |
863 FW_WM(wm->pipe[PIPE_A].cursor, CURSORA) |
864 FW_WM_VLV(wm->pipe[PIPE_A].sprite[0], SPRITEA));
866 FW_WM(wm->sr.cursor, CURSOR_SR));
868 if (IS_CHERRYVIEW(dev_priv)) {
869 I915_WRITE(DSPFW7_CHV,
870 FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
871 FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
872 I915_WRITE(DSPFW8_CHV,
873 FW_WM_VLV(wm->pipe[PIPE_C].sprite[1], SPRITEF) |
874 FW_WM_VLV(wm->pipe[PIPE_C].sprite[0], SPRITEE));
875 I915_WRITE(DSPFW9_CHV,
876 FW_WM_VLV(wm->pipe[PIPE_C].primary, PLANEC) |
877 FW_WM(wm->pipe[PIPE_C].cursor, CURSORC));
879 FW_WM(wm->sr.plane >> 9, SR_HI) |
880 FW_WM(wm->pipe[PIPE_C].sprite[1] >> 8, SPRITEF_HI) |
881 FW_WM(wm->pipe[PIPE_C].sprite[0] >> 8, SPRITEE_HI) |
882 FW_WM(wm->pipe[PIPE_C].primary >> 8, PLANEC_HI) |
883 FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
884 FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
885 FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
886 FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
887 FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
888 FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
891 FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
892 FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
894 FW_WM(wm->sr.plane >> 9, SR_HI) |
895 FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
896 FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
897 FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
898 FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
899 FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
900 FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
903 POSTING_READ(DSPFW1);
905 dev_priv->wm.vlv = *wm;
910 static uint8_t vlv_compute_drain_latency(struct drm_crtc *crtc,
911 struct drm_plane *plane)
913 struct drm_device *dev = crtc->dev;
914 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
915 int entries, prec_mult, drain_latency, pixel_size;
916 int clock = intel_crtc->config->base.adjusted_mode.crtc_clock;
917 const int high_precision = IS_CHERRYVIEW(dev) ? 16 : 64;
920 * FIXME the plane might have an fb
921 * but be invisible (eg. due to clipping)
923 if (!intel_crtc->active || !plane->state->fb)
926 if (WARN(clock == 0, "Pixel clock is zero!\n"))
929 pixel_size = drm_format_plane_cpp(plane->state->fb->pixel_format, 0);
931 if (WARN(pixel_size == 0, "Pixel size is zero!\n"))
934 entries = DIV_ROUND_UP(clock, 1000) * pixel_size;
936 prec_mult = high_precision;
937 drain_latency = 64 * prec_mult * 4 / entries;
939 if (drain_latency > DRAIN_LATENCY_MASK) {
941 drain_latency = 64 * prec_mult * 4 / entries;
944 if (drain_latency > DRAIN_LATENCY_MASK)
945 drain_latency = DRAIN_LATENCY_MASK;
947 return drain_latency | (prec_mult == high_precision ?
948 DDL_PRECISION_HIGH : DDL_PRECISION_LOW);
951 static int vlv_compute_wm(struct intel_crtc *crtc,
952 struct intel_plane *plane,
955 int clock, entries, pixel_size;
958 * FIXME the plane might have an fb
959 * but be invisible (eg. due to clipping)
961 if (!crtc->active || !plane->base.state->fb)
964 pixel_size = drm_format_plane_cpp(plane->base.state->fb->pixel_format, 0);
965 clock = crtc->config->base.adjusted_mode.crtc_clock;
967 entries = DIV_ROUND_UP(clock, 1000) * pixel_size;
970 * Set up the watermark such that we don't start issuing memory
971 * requests until we are within PND's max deadline value (256us).
972 * Idea being to be idle as long as possible while still taking
973 * advatange of PND's deadline scheduling. The limit of 8
974 * cachelines (used when the FIFO will anyway drain in less time
975 * than 256us) should match what we would be done if trickle
978 return fifo_size - clamp(DIV_ROUND_UP(256 * entries, 64), 0, fifo_size - 8);
981 static bool vlv_compute_sr_wm(struct drm_device *dev,
982 struct vlv_wm_values *wm)
984 struct drm_i915_private *dev_priv = to_i915(dev);
985 struct drm_crtc *crtc;
986 enum pipe pipe = INVALID_PIPE;
989 struct intel_plane *plane;
991 wm->sr.cursor = wm->sr.plane = 0;
993 crtc = single_enabled_crtc(dev);
994 /* maxfifo not supported on pipe C */
995 if (crtc && to_intel_crtc(crtc)->pipe != PIPE_C) {
996 pipe = to_intel_crtc(crtc)->pipe;
997 num_planes = !!wm->pipe[pipe].primary +
998 !!wm->pipe[pipe].sprite[0] +
999 !!wm->pipe[pipe].sprite[1];
1000 fifo_size = INTEL_INFO(dev_priv)->num_pipes * 512 - 1;
1003 if (fifo_size == 0 || num_planes > 1)
1006 wm->sr.cursor = vlv_compute_wm(to_intel_crtc(crtc),
1007 to_intel_plane(crtc->cursor), 0x3f);
1009 list_for_each_entry(plane, &dev->mode_config.plane_list, base.head) {
1010 if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
1013 if (plane->pipe != pipe)
1016 wm->sr.plane = vlv_compute_wm(to_intel_crtc(crtc),
1018 if (wm->sr.plane != 0)
1025 static void valleyview_update_wm(struct drm_crtc *crtc)
1027 struct drm_device *dev = crtc->dev;
1028 struct drm_i915_private *dev_priv = dev->dev_private;
1029 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1030 enum pipe pipe = intel_crtc->pipe;
1032 struct vlv_wm_values wm = dev_priv->wm.vlv;
1034 wm.ddl[pipe].primary = vlv_compute_drain_latency(crtc, crtc->primary);
1035 wm.pipe[pipe].primary = vlv_compute_wm(intel_crtc,
1036 to_intel_plane(crtc->primary),
1037 vlv_get_fifo_size(dev, pipe, 0));
1039 wm.ddl[pipe].cursor = vlv_compute_drain_latency(crtc, crtc->cursor);
1040 wm.pipe[pipe].cursor = vlv_compute_wm(intel_crtc,
1041 to_intel_plane(crtc->cursor),
1044 cxsr_enabled = vlv_compute_sr_wm(dev, &wm);
1046 if (memcmp(&wm, &dev_priv->wm.vlv, sizeof(wm)) == 0)
1049 DRM_DEBUG_KMS("Setting FIFO watermarks - %c: plane=%d, cursor=%d, "
1050 "SR: plane=%d, cursor=%d\n", pipe_name(pipe),
1051 wm.pipe[pipe].primary, wm.pipe[pipe].cursor,
1052 wm.sr.plane, wm.sr.cursor);
1055 * FIXME DDR DVFS introduces massive memory latencies which
1056 * are not known to system agent so any deadline specified
1057 * by the display may not be respected. To support DDR DVFS
1058 * the watermark code needs to be rewritten to essentially
1059 * bypass deadline mechanism and rely solely on the
1060 * watermarks. For now disable DDR DVFS.
1062 if (IS_CHERRYVIEW(dev_priv))
1063 chv_set_memory_dvfs(dev_priv, false);
1066 intel_set_memory_cxsr(dev_priv, false);
1068 vlv_write_wm_values(intel_crtc, &wm);
1071 intel_set_memory_cxsr(dev_priv, true);
1074 static void valleyview_update_sprite_wm(struct drm_plane *plane,
1075 struct drm_crtc *crtc,
1076 uint32_t sprite_width,
1077 uint32_t sprite_height,
1079 bool enabled, bool scaled)
1081 struct drm_device *dev = crtc->dev;
1082 struct drm_i915_private *dev_priv = dev->dev_private;
1083 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1084 enum pipe pipe = intel_crtc->pipe;
1085 int sprite = to_intel_plane(plane)->plane;
1087 struct vlv_wm_values wm = dev_priv->wm.vlv;
1090 wm.ddl[pipe].sprite[sprite] =
1091 vlv_compute_drain_latency(crtc, plane);
1093 wm.pipe[pipe].sprite[sprite] =
1094 vlv_compute_wm(intel_crtc,
1095 to_intel_plane(plane),
1096 vlv_get_fifo_size(dev, pipe, sprite+1));
1098 wm.ddl[pipe].sprite[sprite] = 0;
1099 wm.pipe[pipe].sprite[sprite] = 0;
1102 cxsr_enabled = vlv_compute_sr_wm(dev, &wm);
1104 if (memcmp(&wm, &dev_priv->wm.vlv, sizeof(wm)) == 0)
1107 DRM_DEBUG_KMS("Setting FIFO watermarks - %c: sprite %c=%d, "
1108 "SR: plane=%d, cursor=%d\n", pipe_name(pipe),
1109 sprite_name(pipe, sprite),
1110 wm.pipe[pipe].sprite[sprite],
1111 wm.sr.plane, wm.sr.cursor);
1114 intel_set_memory_cxsr(dev_priv, false);
1116 vlv_write_wm_values(intel_crtc, &wm);
1119 intel_set_memory_cxsr(dev_priv, true);
1122 #define single_plane_enabled(mask) is_power_of_2(mask)
1124 static void g4x_update_wm(struct drm_crtc *crtc)
1126 struct drm_device *dev = crtc->dev;
1127 static const int sr_latency_ns = 12000;
1128 struct drm_i915_private *dev_priv = dev->dev_private;
1129 int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1130 int plane_sr, cursor_sr;
1131 unsigned int enabled = 0;
1134 if (g4x_compute_wm0(dev, PIPE_A,
1135 &g4x_wm_info, pessimal_latency_ns,
1136 &g4x_cursor_wm_info, pessimal_latency_ns,
1137 &planea_wm, &cursora_wm))
1138 enabled |= 1 << PIPE_A;
1140 if (g4x_compute_wm0(dev, PIPE_B,
1141 &g4x_wm_info, pessimal_latency_ns,
1142 &g4x_cursor_wm_info, pessimal_latency_ns,
1143 &planeb_wm, &cursorb_wm))
1144 enabled |= 1 << PIPE_B;
1146 if (single_plane_enabled(enabled) &&
1147 g4x_compute_srwm(dev, ffs(enabled) - 1,
1150 &g4x_cursor_wm_info,
1151 &plane_sr, &cursor_sr)) {
1152 cxsr_enabled = true;
1154 cxsr_enabled = false;
1155 intel_set_memory_cxsr(dev_priv, false);
1156 plane_sr = cursor_sr = 0;
1159 DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, "
1160 "B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1161 planea_wm, cursora_wm,
1162 planeb_wm, cursorb_wm,
1163 plane_sr, cursor_sr);
1166 FW_WM(plane_sr, SR) |
1167 FW_WM(cursorb_wm, CURSORB) |
1168 FW_WM(planeb_wm, PLANEB) |
1169 FW_WM(planea_wm, PLANEA));
1171 (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1172 FW_WM(cursora_wm, CURSORA));
1173 /* HPLL off in SR has some issues on G4x... disable it */
1175 (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1176 FW_WM(cursor_sr, CURSOR_SR));
1179 intel_set_memory_cxsr(dev_priv, true);
1182 static void i965_update_wm(struct drm_crtc *unused_crtc)
1184 struct drm_device *dev = unused_crtc->dev;
1185 struct drm_i915_private *dev_priv = dev->dev_private;
1186 struct drm_crtc *crtc;
1191 /* Calc sr entries for one plane configs */
1192 crtc = single_enabled_crtc(dev);
1194 /* self-refresh has much higher latency */
1195 static const int sr_latency_ns = 12000;
1196 const struct drm_display_mode *adjusted_mode =
1197 &to_intel_crtc(crtc)->config->base.adjusted_mode;
1198 int clock = adjusted_mode->crtc_clock;
1199 int htotal = adjusted_mode->crtc_htotal;
1200 int hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
1201 int pixel_size = crtc->primary->state->fb->bits_per_pixel / 8;
1202 unsigned long line_time_us;
1205 line_time_us = max(htotal * 1000 / clock, 1);
1207 /* Use ns/us then divide to preserve precision */
1208 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1209 pixel_size * hdisplay;
1210 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1211 srwm = I965_FIFO_SIZE - entries;
1215 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1218 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1219 pixel_size * crtc->cursor->state->crtc_w;
1220 entries = DIV_ROUND_UP(entries,
1221 i965_cursor_wm_info.cacheline_size);
1222 cursor_sr = i965_cursor_wm_info.fifo_size -
1223 (entries + i965_cursor_wm_info.guard_size);
1225 if (cursor_sr > i965_cursor_wm_info.max_wm)
1226 cursor_sr = i965_cursor_wm_info.max_wm;
1228 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1229 "cursor %d\n", srwm, cursor_sr);
1231 cxsr_enabled = true;
1233 cxsr_enabled = false;
1234 /* Turn off self refresh if both pipes are enabled */
1235 intel_set_memory_cxsr(dev_priv, false);
1238 DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1241 /* 965 has limitations... */
1242 I915_WRITE(DSPFW1, FW_WM(srwm, SR) |
1246 I915_WRITE(DSPFW2, FW_WM(8, CURSORA) |
1247 FW_WM(8, PLANEC_OLD));
1248 /* update cursor SR watermark */
1249 I915_WRITE(DSPFW3, FW_WM(cursor_sr, CURSOR_SR));
1252 intel_set_memory_cxsr(dev_priv, true);
1257 static void i9xx_update_wm(struct drm_crtc *unused_crtc)
1259 struct drm_device *dev = unused_crtc->dev;
1260 struct drm_i915_private *dev_priv = dev->dev_private;
1261 const struct intel_watermark_params *wm_info;
1266 int planea_wm, planeb_wm;
1267 struct drm_crtc *crtc, *enabled = NULL;
1270 wm_info = &i945_wm_info;
1271 else if (!IS_GEN2(dev))
1272 wm_info = &i915_wm_info;
1274 wm_info = &i830_a_wm_info;
1276 fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1277 crtc = intel_get_crtc_for_plane(dev, 0);
1278 if (intel_crtc_active(crtc)) {
1279 const struct drm_display_mode *adjusted_mode;
1280 int cpp = crtc->primary->state->fb->bits_per_pixel / 8;
1284 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1285 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1286 wm_info, fifo_size, cpp,
1287 pessimal_latency_ns);
1290 planea_wm = fifo_size - wm_info->guard_size;
1291 if (planea_wm > (long)wm_info->max_wm)
1292 planea_wm = wm_info->max_wm;
1296 wm_info = &i830_bc_wm_info;
1298 fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1299 crtc = intel_get_crtc_for_plane(dev, 1);
1300 if (intel_crtc_active(crtc)) {
1301 const struct drm_display_mode *adjusted_mode;
1302 int cpp = crtc->primary->state->fb->bits_per_pixel / 8;
1306 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1307 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1308 wm_info, fifo_size, cpp,
1309 pessimal_latency_ns);
1310 if (enabled == NULL)
1315 planeb_wm = fifo_size - wm_info->guard_size;
1316 if (planeb_wm > (long)wm_info->max_wm)
1317 planeb_wm = wm_info->max_wm;
1320 DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1322 if (IS_I915GM(dev) && enabled) {
1323 struct drm_i915_gem_object *obj;
1325 obj = intel_fb_obj(enabled->primary->state->fb);
1327 /* self-refresh seems busted with untiled */
1328 if (obj->tiling_mode == I915_TILING_NONE)
1333 * Overlay gets an aggressive default since video jitter is bad.
1337 /* Play safe and disable self-refresh before adjusting watermarks. */
1338 intel_set_memory_cxsr(dev_priv, false);
1340 /* Calc sr entries for one plane configs */
1341 if (HAS_FW_BLC(dev) && enabled) {
1342 /* self-refresh has much higher latency */
1343 static const int sr_latency_ns = 6000;
1344 const struct drm_display_mode *adjusted_mode =
1345 &to_intel_crtc(enabled)->config->base.adjusted_mode;
1346 int clock = adjusted_mode->crtc_clock;
1347 int htotal = adjusted_mode->crtc_htotal;
1348 int hdisplay = to_intel_crtc(enabled)->config->pipe_src_w;
1349 int pixel_size = enabled->primary->state->fb->bits_per_pixel / 8;
1350 unsigned long line_time_us;
1353 line_time_us = max(htotal * 1000 / clock, 1);
1355 /* Use ns/us then divide to preserve precision */
1356 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1357 pixel_size * hdisplay;
1358 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1359 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1360 srwm = wm_info->fifo_size - entries;
1364 if (IS_I945G(dev) || IS_I945GM(dev))
1365 I915_WRITE(FW_BLC_SELF,
1366 FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1367 else if (IS_I915GM(dev))
1368 I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1371 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1372 planea_wm, planeb_wm, cwm, srwm);
1374 fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1375 fwater_hi = (cwm & 0x1f);
1377 /* Set request length to 8 cachelines per fetch */
1378 fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1379 fwater_hi = fwater_hi | (1 << 8);
1381 I915_WRITE(FW_BLC, fwater_lo);
1382 I915_WRITE(FW_BLC2, fwater_hi);
1385 intel_set_memory_cxsr(dev_priv, true);
1388 static void i845_update_wm(struct drm_crtc *unused_crtc)
1390 struct drm_device *dev = unused_crtc->dev;
1391 struct drm_i915_private *dev_priv = dev->dev_private;
1392 struct drm_crtc *crtc;
1393 const struct drm_display_mode *adjusted_mode;
1397 crtc = single_enabled_crtc(dev);
1401 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1402 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1404 dev_priv->display.get_fifo_size(dev, 0),
1405 4, pessimal_latency_ns);
1406 fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1407 fwater_lo |= (3<<8) | planea_wm;
1409 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1411 I915_WRITE(FW_BLC, fwater_lo);
1414 static uint32_t ilk_pipe_pixel_rate(struct drm_device *dev,
1415 struct drm_crtc *crtc)
1417 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1418 uint32_t pixel_rate;
1420 pixel_rate = intel_crtc->config->base.adjusted_mode.crtc_clock;
1422 /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
1423 * adjust the pixel_rate here. */
1425 if (intel_crtc->config->pch_pfit.enabled) {
1426 uint64_t pipe_w, pipe_h, pfit_w, pfit_h;
1427 uint32_t pfit_size = intel_crtc->config->pch_pfit.size;
1429 pipe_w = intel_crtc->config->pipe_src_w;
1430 pipe_h = intel_crtc->config->pipe_src_h;
1431 pfit_w = (pfit_size >> 16) & 0xFFFF;
1432 pfit_h = pfit_size & 0xFFFF;
1433 if (pipe_w < pfit_w)
1435 if (pipe_h < pfit_h)
1438 pixel_rate = div_u64((uint64_t) pixel_rate * pipe_w * pipe_h,
1445 /* latency must be in 0.1us units. */
1446 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
1451 if (WARN(latency == 0, "Latency value missing\n"))
1454 ret = (uint64_t) pixel_rate * bytes_per_pixel * latency;
1455 ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
1460 /* latency must be in 0.1us units. */
1461 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
1462 uint32_t horiz_pixels, uint8_t bytes_per_pixel,
1467 if (WARN(latency == 0, "Latency value missing\n"))
1470 ret = (latency * pixel_rate) / (pipe_htotal * 10000);
1471 ret = (ret + 1) * horiz_pixels * bytes_per_pixel;
1472 ret = DIV_ROUND_UP(ret, 64) + 2;
1476 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
1477 uint8_t bytes_per_pixel)
1479 return DIV_ROUND_UP(pri_val * 64, horiz_pixels * bytes_per_pixel) + 2;
1482 struct skl_pipe_wm_parameters {
1484 uint32_t pipe_htotal;
1485 uint32_t pixel_rate; /* in KHz */
1486 struct intel_plane_wm_parameters plane[I915_MAX_PLANES];
1487 struct intel_plane_wm_parameters cursor;
1490 struct ilk_pipe_wm_parameters {
1492 uint32_t pipe_htotal;
1493 uint32_t pixel_rate;
1494 struct intel_plane_wm_parameters pri;
1495 struct intel_plane_wm_parameters spr;
1496 struct intel_plane_wm_parameters cur;
1499 struct ilk_wm_maximums {
1506 /* used in computing the new watermarks state */
1507 struct intel_wm_config {
1508 unsigned int num_pipes_active;
1509 bool sprites_enabled;
1510 bool sprites_scaled;
1514 * For both WM_PIPE and WM_LP.
1515 * mem_value must be in 0.1us units.
1517 static uint32_t ilk_compute_pri_wm(const struct ilk_pipe_wm_parameters *params,
1521 uint32_t method1, method2;
1523 if (!params->active || !params->pri.enabled)
1526 method1 = ilk_wm_method1(params->pixel_rate,
1527 params->pri.bytes_per_pixel,
1533 method2 = ilk_wm_method2(params->pixel_rate,
1534 params->pipe_htotal,
1535 params->pri.horiz_pixels,
1536 params->pri.bytes_per_pixel,
1539 return min(method1, method2);
1543 * For both WM_PIPE and WM_LP.
1544 * mem_value must be in 0.1us units.
1546 static uint32_t ilk_compute_spr_wm(const struct ilk_pipe_wm_parameters *params,
1549 uint32_t method1, method2;
1551 if (!params->active || !params->spr.enabled)
1554 method1 = ilk_wm_method1(params->pixel_rate,
1555 params->spr.bytes_per_pixel,
1557 method2 = ilk_wm_method2(params->pixel_rate,
1558 params->pipe_htotal,
1559 params->spr.horiz_pixels,
1560 params->spr.bytes_per_pixel,
1562 return min(method1, method2);
1566 * For both WM_PIPE and WM_LP.
1567 * mem_value must be in 0.1us units.
1569 static uint32_t ilk_compute_cur_wm(const struct ilk_pipe_wm_parameters *params,
1572 if (!params->active || !params->cur.enabled)
1575 return ilk_wm_method2(params->pixel_rate,
1576 params->pipe_htotal,
1577 params->cur.horiz_pixels,
1578 params->cur.bytes_per_pixel,
1582 /* Only for WM_LP. */
1583 static uint32_t ilk_compute_fbc_wm(const struct ilk_pipe_wm_parameters *params,
1586 if (!params->active || !params->pri.enabled)
1589 return ilk_wm_fbc(pri_val,
1590 params->pri.horiz_pixels,
1591 params->pri.bytes_per_pixel);
1594 static unsigned int ilk_display_fifo_size(const struct drm_device *dev)
1596 if (INTEL_INFO(dev)->gen >= 8)
1598 else if (INTEL_INFO(dev)->gen >= 7)
1604 static unsigned int ilk_plane_wm_reg_max(const struct drm_device *dev,
1605 int level, bool is_sprite)
1607 if (INTEL_INFO(dev)->gen >= 8)
1608 /* BDW primary/sprite plane watermarks */
1609 return level == 0 ? 255 : 2047;
1610 else if (INTEL_INFO(dev)->gen >= 7)
1611 /* IVB/HSW primary/sprite plane watermarks */
1612 return level == 0 ? 127 : 1023;
1613 else if (!is_sprite)
1614 /* ILK/SNB primary plane watermarks */
1615 return level == 0 ? 127 : 511;
1617 /* ILK/SNB sprite plane watermarks */
1618 return level == 0 ? 63 : 255;
1621 static unsigned int ilk_cursor_wm_reg_max(const struct drm_device *dev,
1624 if (INTEL_INFO(dev)->gen >= 7)
1625 return level == 0 ? 63 : 255;
1627 return level == 0 ? 31 : 63;
1630 static unsigned int ilk_fbc_wm_reg_max(const struct drm_device *dev)
1632 if (INTEL_INFO(dev)->gen >= 8)
1638 /* Calculate the maximum primary/sprite plane watermark */
1639 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
1641 const struct intel_wm_config *config,
1642 enum intel_ddb_partitioning ddb_partitioning,
1645 unsigned int fifo_size = ilk_display_fifo_size(dev);
1647 /* if sprites aren't enabled, sprites get nothing */
1648 if (is_sprite && !config->sprites_enabled)
1651 /* HSW allows LP1+ watermarks even with multiple pipes */
1652 if (level == 0 || config->num_pipes_active > 1) {
1653 fifo_size /= INTEL_INFO(dev)->num_pipes;
1656 * For some reason the non self refresh
1657 * FIFO size is only half of the self
1658 * refresh FIFO size on ILK/SNB.
1660 if (INTEL_INFO(dev)->gen <= 6)
1664 if (config->sprites_enabled) {
1665 /* level 0 is always calculated with 1:1 split */
1666 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
1675 /* clamp to max that the registers can hold */
1676 return min(fifo_size, ilk_plane_wm_reg_max(dev, level, is_sprite));
1679 /* Calculate the maximum cursor plane watermark */
1680 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
1682 const struct intel_wm_config *config)
1684 /* HSW LP1+ watermarks w/ multiple pipes */
1685 if (level > 0 && config->num_pipes_active > 1)
1688 /* otherwise just report max that registers can hold */
1689 return ilk_cursor_wm_reg_max(dev, level);
1692 static void ilk_compute_wm_maximums(const struct drm_device *dev,
1694 const struct intel_wm_config *config,
1695 enum intel_ddb_partitioning ddb_partitioning,
1696 struct ilk_wm_maximums *max)
1698 max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
1699 max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
1700 max->cur = ilk_cursor_wm_max(dev, level, config);
1701 max->fbc = ilk_fbc_wm_reg_max(dev);
1704 static void ilk_compute_wm_reg_maximums(struct drm_device *dev,
1706 struct ilk_wm_maximums *max)
1708 max->pri = ilk_plane_wm_reg_max(dev, level, false);
1709 max->spr = ilk_plane_wm_reg_max(dev, level, true);
1710 max->cur = ilk_cursor_wm_reg_max(dev, level);
1711 max->fbc = ilk_fbc_wm_reg_max(dev);
1714 static bool ilk_validate_wm_level(int level,
1715 const struct ilk_wm_maximums *max,
1716 struct intel_wm_level *result)
1720 /* already determined to be invalid? */
1721 if (!result->enable)
1724 result->enable = result->pri_val <= max->pri &&
1725 result->spr_val <= max->spr &&
1726 result->cur_val <= max->cur;
1728 ret = result->enable;
1731 * HACK until we can pre-compute everything,
1732 * and thus fail gracefully if LP0 watermarks
1735 if (level == 0 && !result->enable) {
1736 if (result->pri_val > max->pri)
1737 DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
1738 level, result->pri_val, max->pri);
1739 if (result->spr_val > max->spr)
1740 DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
1741 level, result->spr_val, max->spr);
1742 if (result->cur_val > max->cur)
1743 DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
1744 level, result->cur_val, max->cur);
1746 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
1747 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
1748 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
1749 result->enable = true;
1755 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
1757 const struct ilk_pipe_wm_parameters *p,
1758 struct intel_wm_level *result)
1760 uint16_t pri_latency = dev_priv->wm.pri_latency[level];
1761 uint16_t spr_latency = dev_priv->wm.spr_latency[level];
1762 uint16_t cur_latency = dev_priv->wm.cur_latency[level];
1764 /* WM1+ latency values stored in 0.5us units */
1771 result->pri_val = ilk_compute_pri_wm(p, pri_latency, level);
1772 result->spr_val = ilk_compute_spr_wm(p, spr_latency);
1773 result->cur_val = ilk_compute_cur_wm(p, cur_latency);
1774 result->fbc_val = ilk_compute_fbc_wm(p, result->pri_val);
1775 result->enable = true;
1779 hsw_compute_linetime_wm(struct drm_device *dev, struct drm_crtc *crtc)
1781 struct drm_i915_private *dev_priv = dev->dev_private;
1782 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1783 struct drm_display_mode *mode = &intel_crtc->config->base.adjusted_mode;
1784 u32 linetime, ips_linetime;
1786 if (!intel_crtc->active)
1789 /* The WM are computed with base on how long it takes to fill a single
1790 * row at the given clock rate, multiplied by 8.
1792 linetime = DIV_ROUND_CLOSEST(mode->crtc_htotal * 1000 * 8,
1794 ips_linetime = DIV_ROUND_CLOSEST(mode->crtc_htotal * 1000 * 8,
1795 intel_ddi_get_cdclk_freq(dev_priv));
1797 return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
1798 PIPE_WM_LINETIME_TIME(linetime);
1801 static void intel_read_wm_latency(struct drm_device *dev, uint16_t wm[8])
1803 struct drm_i915_private *dev_priv = dev->dev_private;
1808 int level, max_level = ilk_wm_max_level(dev);
1810 /* read the first set of memory latencies[0:3] */
1811 val = 0; /* data0 to be programmed to 0 for first set */
1812 mutex_lock(&dev_priv->rps.hw_lock);
1813 ret = sandybridge_pcode_read(dev_priv,
1814 GEN9_PCODE_READ_MEM_LATENCY,
1816 mutex_unlock(&dev_priv->rps.hw_lock);
1819 DRM_ERROR("SKL Mailbox read error = %d\n", ret);
1823 wm[0] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
1824 wm[1] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
1825 GEN9_MEM_LATENCY_LEVEL_MASK;
1826 wm[2] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
1827 GEN9_MEM_LATENCY_LEVEL_MASK;
1828 wm[3] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
1829 GEN9_MEM_LATENCY_LEVEL_MASK;
1831 /* read the second set of memory latencies[4:7] */
1832 val = 1; /* data0 to be programmed to 1 for second set */
1833 mutex_lock(&dev_priv->rps.hw_lock);
1834 ret = sandybridge_pcode_read(dev_priv,
1835 GEN9_PCODE_READ_MEM_LATENCY,
1837 mutex_unlock(&dev_priv->rps.hw_lock);
1839 DRM_ERROR("SKL Mailbox read error = %d\n", ret);
1843 wm[4] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
1844 wm[5] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
1845 GEN9_MEM_LATENCY_LEVEL_MASK;
1846 wm[6] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
1847 GEN9_MEM_LATENCY_LEVEL_MASK;
1848 wm[7] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
1849 GEN9_MEM_LATENCY_LEVEL_MASK;
1852 * WaWmMemoryReadLatency:skl
1854 * punit doesn't take into account the read latency so we need
1855 * to add 2us to the various latency levels we retrieve from
1857 * - W0 is a bit special in that it's the only level that
1858 * can't be disabled if we want to have display working, so
1859 * we always add 2us there.
1860 * - For levels >=1, punit returns 0us latency when they are
1861 * disabled, so we respect that and don't add 2us then
1863 * Additionally, if a level n (n > 1) has a 0us latency, all
1864 * levels m (m >= n) need to be disabled. We make sure to
1865 * sanitize the values out of the punit to satisfy this
1869 for (level = 1; level <= max_level; level++)
1873 for (i = level + 1; i <= max_level; i++)
1878 } else if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
1879 uint64_t sskpd = I915_READ64(MCH_SSKPD);
1881 wm[0] = (sskpd >> 56) & 0xFF;
1883 wm[0] = sskpd & 0xF;
1884 wm[1] = (sskpd >> 4) & 0xFF;
1885 wm[2] = (sskpd >> 12) & 0xFF;
1886 wm[3] = (sskpd >> 20) & 0x1FF;
1887 wm[4] = (sskpd >> 32) & 0x1FF;
1888 } else if (INTEL_INFO(dev)->gen >= 6) {
1889 uint32_t sskpd = I915_READ(MCH_SSKPD);
1891 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
1892 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
1893 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
1894 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
1895 } else if (INTEL_INFO(dev)->gen >= 5) {
1896 uint32_t mltr = I915_READ(MLTR_ILK);
1898 /* ILK primary LP0 latency is 700 ns */
1900 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
1901 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
1905 static void intel_fixup_spr_wm_latency(struct drm_device *dev, uint16_t wm[5])
1907 /* ILK sprite LP0 latency is 1300 ns */
1908 if (INTEL_INFO(dev)->gen == 5)
1912 static void intel_fixup_cur_wm_latency(struct drm_device *dev, uint16_t wm[5])
1914 /* ILK cursor LP0 latency is 1300 ns */
1915 if (INTEL_INFO(dev)->gen == 5)
1918 /* WaDoubleCursorLP3Latency:ivb */
1919 if (IS_IVYBRIDGE(dev))
1923 int ilk_wm_max_level(const struct drm_device *dev)
1925 /* how many WM levels are we expecting */
1928 else if (IS_HASWELL(dev) || IS_BROADWELL(dev))
1930 else if (INTEL_INFO(dev)->gen >= 6)
1936 static void intel_print_wm_latency(struct drm_device *dev,
1938 const uint16_t wm[8])
1940 int level, max_level = ilk_wm_max_level(dev);
1942 for (level = 0; level <= max_level; level++) {
1943 unsigned int latency = wm[level];
1946 DRM_ERROR("%s WM%d latency not provided\n",
1952 * - latencies are in us on gen9.
1953 * - before then, WM1+ latency values are in 0.5us units
1960 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
1961 name, level, wm[level],
1962 latency / 10, latency % 10);
1966 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
1967 uint16_t wm[5], uint16_t min)
1969 int level, max_level = ilk_wm_max_level(dev_priv->dev);
1974 wm[0] = max(wm[0], min);
1975 for (level = 1; level <= max_level; level++)
1976 wm[level] = max_t(uint16_t, wm[level], DIV_ROUND_UP(min, 5));
1981 static void snb_wm_latency_quirk(struct drm_device *dev)
1983 struct drm_i915_private *dev_priv = dev->dev_private;
1987 * The BIOS provided WM memory latency values are often
1988 * inadequate for high resolution displays. Adjust them.
1990 changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
1991 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
1992 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
1997 DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
1998 intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
1999 intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2000 intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2003 static void ilk_setup_wm_latency(struct drm_device *dev)
2005 struct drm_i915_private *dev_priv = dev->dev_private;
2007 intel_read_wm_latency(dev, dev_priv->wm.pri_latency);
2009 memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2010 sizeof(dev_priv->wm.pri_latency));
2011 memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2012 sizeof(dev_priv->wm.pri_latency));
2014 intel_fixup_spr_wm_latency(dev, dev_priv->wm.spr_latency);
2015 intel_fixup_cur_wm_latency(dev, dev_priv->wm.cur_latency);
2017 intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2018 intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2019 intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2022 snb_wm_latency_quirk(dev);
2025 static void skl_setup_wm_latency(struct drm_device *dev)
2027 struct drm_i915_private *dev_priv = dev->dev_private;
2029 intel_read_wm_latency(dev, dev_priv->wm.skl_latency);
2030 intel_print_wm_latency(dev, "Gen9 Plane", dev_priv->wm.skl_latency);
2033 static void ilk_compute_wm_parameters(struct drm_crtc *crtc,
2034 struct ilk_pipe_wm_parameters *p)
2036 struct drm_device *dev = crtc->dev;
2037 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2038 enum pipe pipe = intel_crtc->pipe;
2039 struct drm_plane *plane;
2041 if (!intel_crtc->active)
2045 p->pipe_htotal = intel_crtc->config->base.adjusted_mode.crtc_htotal;
2046 p->pixel_rate = ilk_pipe_pixel_rate(dev, crtc);
2048 if (crtc->primary->state->fb) {
2049 p->pri.enabled = true;
2050 p->pri.bytes_per_pixel =
2051 crtc->primary->state->fb->bits_per_pixel / 8;
2053 p->pri.enabled = false;
2054 p->pri.bytes_per_pixel = 0;
2057 if (crtc->cursor->state->fb) {
2058 p->cur.enabled = true;
2059 p->cur.bytes_per_pixel = 4;
2061 p->cur.enabled = false;
2062 p->cur.bytes_per_pixel = 0;
2064 p->pri.horiz_pixels = intel_crtc->config->pipe_src_w;
2065 p->cur.horiz_pixels = intel_crtc->base.cursor->state->crtc_w;
2067 drm_for_each_legacy_plane(plane, &dev->mode_config.plane_list) {
2068 struct intel_plane *intel_plane = to_intel_plane(plane);
2070 if (intel_plane->pipe == pipe) {
2071 p->spr = intel_plane->wm;
2077 static void ilk_compute_wm_config(struct drm_device *dev,
2078 struct intel_wm_config *config)
2080 struct intel_crtc *intel_crtc;
2082 /* Compute the currently _active_ config */
2083 for_each_intel_crtc(dev, intel_crtc) {
2084 const struct intel_pipe_wm *wm = &intel_crtc->wm.active;
2086 if (!wm->pipe_enabled)
2089 config->sprites_enabled |= wm->sprites_enabled;
2090 config->sprites_scaled |= wm->sprites_scaled;
2091 config->num_pipes_active++;
2095 /* Compute new watermarks for the pipe */
2096 static bool intel_compute_pipe_wm(struct drm_crtc *crtc,
2097 const struct ilk_pipe_wm_parameters *params,
2098 struct intel_pipe_wm *pipe_wm)
2100 struct drm_device *dev = crtc->dev;
2101 const struct drm_i915_private *dev_priv = dev->dev_private;
2102 int level, max_level = ilk_wm_max_level(dev);
2103 /* LP0 watermark maximums depend on this pipe alone */
2104 struct intel_wm_config config = {
2105 .num_pipes_active = 1,
2106 .sprites_enabled = params->spr.enabled,
2107 .sprites_scaled = params->spr.scaled,
2109 struct ilk_wm_maximums max;
2111 pipe_wm->pipe_enabled = params->active;
2112 pipe_wm->sprites_enabled = params->spr.enabled;
2113 pipe_wm->sprites_scaled = params->spr.scaled;
2115 /* ILK/SNB: LP2+ watermarks only w/o sprites */
2116 if (INTEL_INFO(dev)->gen <= 6 && params->spr.enabled)
2119 /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2120 if (params->spr.scaled)
2123 ilk_compute_wm_level(dev_priv, 0, params, &pipe_wm->wm[0]);
2125 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2126 pipe_wm->linetime = hsw_compute_linetime_wm(dev, crtc);
2128 /* LP0 watermarks always use 1/2 DDB partitioning */
2129 ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
2131 /* At least LP0 must be valid */
2132 if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0]))
2135 ilk_compute_wm_reg_maximums(dev, 1, &max);
2137 for (level = 1; level <= max_level; level++) {
2138 struct intel_wm_level wm = {};
2140 ilk_compute_wm_level(dev_priv, level, params, &wm);
2143 * Disable any watermark level that exceeds the
2144 * register maximums since such watermarks are
2147 if (!ilk_validate_wm_level(level, &max, &wm))
2150 pipe_wm->wm[level] = wm;
2157 * Merge the watermarks from all active pipes for a specific level.
2159 static void ilk_merge_wm_level(struct drm_device *dev,
2161 struct intel_wm_level *ret_wm)
2163 const struct intel_crtc *intel_crtc;
2165 ret_wm->enable = true;
2167 for_each_intel_crtc(dev, intel_crtc) {
2168 const struct intel_pipe_wm *active = &intel_crtc->wm.active;
2169 const struct intel_wm_level *wm = &active->wm[level];
2171 if (!active->pipe_enabled)
2175 * The watermark values may have been used in the past,
2176 * so we must maintain them in the registers for some
2177 * time even if the level is now disabled.
2180 ret_wm->enable = false;
2182 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
2183 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
2184 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
2185 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
2190 * Merge all low power watermarks for all active pipes.
2192 static void ilk_wm_merge(struct drm_device *dev,
2193 const struct intel_wm_config *config,
2194 const struct ilk_wm_maximums *max,
2195 struct intel_pipe_wm *merged)
2197 int level, max_level = ilk_wm_max_level(dev);
2198 int last_enabled_level = max_level;
2200 /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2201 if ((INTEL_INFO(dev)->gen <= 6 || IS_IVYBRIDGE(dev)) &&
2202 config->num_pipes_active > 1)
2205 /* ILK: FBC WM must be disabled always */
2206 merged->fbc_wm_enabled = INTEL_INFO(dev)->gen >= 6;
2208 /* merge each WM1+ level */
2209 for (level = 1; level <= max_level; level++) {
2210 struct intel_wm_level *wm = &merged->wm[level];
2212 ilk_merge_wm_level(dev, level, wm);
2214 if (level > last_enabled_level)
2216 else if (!ilk_validate_wm_level(level, max, wm))
2217 /* make sure all following levels get disabled */
2218 last_enabled_level = level - 1;
2221 * The spec says it is preferred to disable
2222 * FBC WMs instead of disabling a WM level.
2224 if (wm->fbc_val > max->fbc) {
2226 merged->fbc_wm_enabled = false;
2231 /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2233 * FIXME this is racy. FBC might get enabled later.
2234 * What we should check here is whether FBC can be
2235 * enabled sometime later.
2237 if (IS_GEN5(dev) && !merged->fbc_wm_enabled && intel_fbc_enabled(dev)) {
2238 for (level = 2; level <= max_level; level++) {
2239 struct intel_wm_level *wm = &merged->wm[level];
2246 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
2248 /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2249 return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
2252 /* The value we need to program into the WM_LPx latency field */
2253 static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
2255 struct drm_i915_private *dev_priv = dev->dev_private;
2257 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2260 return dev_priv->wm.pri_latency[level];
2263 static void ilk_compute_wm_results(struct drm_device *dev,
2264 const struct intel_pipe_wm *merged,
2265 enum intel_ddb_partitioning partitioning,
2266 struct ilk_wm_values *results)
2268 struct intel_crtc *intel_crtc;
2271 results->enable_fbc_wm = merged->fbc_wm_enabled;
2272 results->partitioning = partitioning;
2274 /* LP1+ register values */
2275 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2276 const struct intel_wm_level *r;
2278 level = ilk_wm_lp_to_level(wm_lp, merged);
2280 r = &merged->wm[level];
2283 * Maintain the watermark values even if the level is
2284 * disabled. Doing otherwise could cause underruns.
2286 results->wm_lp[wm_lp - 1] =
2287 (ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
2288 (r->pri_val << WM1_LP_SR_SHIFT) |
2292 results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
2294 if (INTEL_INFO(dev)->gen >= 8)
2295 results->wm_lp[wm_lp - 1] |=
2296 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
2298 results->wm_lp[wm_lp - 1] |=
2299 r->fbc_val << WM1_LP_FBC_SHIFT;
2302 * Always set WM1S_LP_EN when spr_val != 0, even if the
2303 * level is disabled. Doing otherwise could cause underruns.
2305 if (INTEL_INFO(dev)->gen <= 6 && r->spr_val) {
2306 WARN_ON(wm_lp != 1);
2307 results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
2309 results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2312 /* LP0 register values */
2313 for_each_intel_crtc(dev, intel_crtc) {
2314 enum pipe pipe = intel_crtc->pipe;
2315 const struct intel_wm_level *r =
2316 &intel_crtc->wm.active.wm[0];
2318 if (WARN_ON(!r->enable))
2321 results->wm_linetime[pipe] = intel_crtc->wm.active.linetime;
2323 results->wm_pipe[pipe] =
2324 (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
2325 (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
2330 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2331 * case both are at the same level. Prefer r1 in case they're the same. */
2332 static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
2333 struct intel_pipe_wm *r1,
2334 struct intel_pipe_wm *r2)
2336 int level, max_level = ilk_wm_max_level(dev);
2337 int level1 = 0, level2 = 0;
2339 for (level = 1; level <= max_level; level++) {
2340 if (r1->wm[level].enable)
2342 if (r2->wm[level].enable)
2346 if (level1 == level2) {
2347 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
2351 } else if (level1 > level2) {
2358 /* dirty bits used to track which watermarks need changes */
2359 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2360 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2361 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2362 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2363 #define WM_DIRTY_FBC (1 << 24)
2364 #define WM_DIRTY_DDB (1 << 25)
2366 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private *dev_priv,
2367 const struct ilk_wm_values *old,
2368 const struct ilk_wm_values *new)
2370 unsigned int dirty = 0;
2374 for_each_pipe(dev_priv, pipe) {
2375 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
2376 dirty |= WM_DIRTY_LINETIME(pipe);
2377 /* Must disable LP1+ watermarks too */
2378 dirty |= WM_DIRTY_LP_ALL;
2381 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
2382 dirty |= WM_DIRTY_PIPE(pipe);
2383 /* Must disable LP1+ watermarks too */
2384 dirty |= WM_DIRTY_LP_ALL;
2388 if (old->enable_fbc_wm != new->enable_fbc_wm) {
2389 dirty |= WM_DIRTY_FBC;
2390 /* Must disable LP1+ watermarks too */
2391 dirty |= WM_DIRTY_LP_ALL;
2394 if (old->partitioning != new->partitioning) {
2395 dirty |= WM_DIRTY_DDB;
2396 /* Must disable LP1+ watermarks too */
2397 dirty |= WM_DIRTY_LP_ALL;
2400 /* LP1+ watermarks already deemed dirty, no need to continue */
2401 if (dirty & WM_DIRTY_LP_ALL)
2404 /* Find the lowest numbered LP1+ watermark in need of an update... */
2405 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2406 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
2407 old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
2411 /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2412 for (; wm_lp <= 3; wm_lp++)
2413 dirty |= WM_DIRTY_LP(wm_lp);
2418 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
2421 struct ilk_wm_values *previous = &dev_priv->wm.hw;
2422 bool changed = false;
2424 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
2425 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
2426 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
2429 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
2430 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
2431 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
2434 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
2435 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
2436 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
2441 * Don't touch WM1S_LP_EN here.
2442 * Doing so could cause underruns.
2449 * The spec says we shouldn't write when we don't need, because every write
2450 * causes WMs to be re-evaluated, expending some power.
2452 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
2453 struct ilk_wm_values *results)
2455 struct drm_device *dev = dev_priv->dev;
2456 struct ilk_wm_values *previous = &dev_priv->wm.hw;
2460 dirty = ilk_compute_wm_dirty(dev_priv, previous, results);
2464 _ilk_disable_lp_wm(dev_priv, dirty);
2466 if (dirty & WM_DIRTY_PIPE(PIPE_A))
2467 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2468 if (dirty & WM_DIRTY_PIPE(PIPE_B))
2469 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
2470 if (dirty & WM_DIRTY_PIPE(PIPE_C))
2471 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
2473 if (dirty & WM_DIRTY_LINETIME(PIPE_A))
2474 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
2475 if (dirty & WM_DIRTY_LINETIME(PIPE_B))
2476 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
2477 if (dirty & WM_DIRTY_LINETIME(PIPE_C))
2478 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
2480 if (dirty & WM_DIRTY_DDB) {
2481 if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2482 val = I915_READ(WM_MISC);
2483 if (results->partitioning == INTEL_DDB_PART_1_2)
2484 val &= ~WM_MISC_DATA_PARTITION_5_6;
2486 val |= WM_MISC_DATA_PARTITION_5_6;
2487 I915_WRITE(WM_MISC, val);
2489 val = I915_READ(DISP_ARB_CTL2);
2490 if (results->partitioning == INTEL_DDB_PART_1_2)
2491 val &= ~DISP_DATA_PARTITION_5_6;
2493 val |= DISP_DATA_PARTITION_5_6;
2494 I915_WRITE(DISP_ARB_CTL2, val);
2498 if (dirty & WM_DIRTY_FBC) {
2499 val = I915_READ(DISP_ARB_CTL);
2500 if (results->enable_fbc_wm)
2501 val &= ~DISP_FBC_WM_DIS;
2503 val |= DISP_FBC_WM_DIS;
2504 I915_WRITE(DISP_ARB_CTL, val);
2507 if (dirty & WM_DIRTY_LP(1) &&
2508 previous->wm_lp_spr[0] != results->wm_lp_spr[0])
2509 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
2511 if (INTEL_INFO(dev)->gen >= 7) {
2512 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
2513 I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
2514 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
2515 I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
2518 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
2519 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
2520 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
2521 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
2522 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
2523 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
2525 dev_priv->wm.hw = *results;
2528 static bool ilk_disable_lp_wm(struct drm_device *dev)
2530 struct drm_i915_private *dev_priv = dev->dev_private;
2532 return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
2536 * On gen9, we need to allocate Display Data Buffer (DDB) portions to the
2537 * different active planes.
2540 #define SKL_DDB_SIZE 896 /* in blocks */
2543 skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
2544 struct drm_crtc *for_crtc,
2545 const struct intel_wm_config *config,
2546 const struct skl_pipe_wm_parameters *params,
2547 struct skl_ddb_entry *alloc /* out */)
2549 struct drm_crtc *crtc;
2550 unsigned int pipe_size, ddb_size;
2551 int nth_active_pipe;
2553 if (!params->active) {
2559 ddb_size = SKL_DDB_SIZE;
2561 ddb_size -= 4; /* 4 blocks for bypass path allocation */
2563 nth_active_pipe = 0;
2564 for_each_crtc(dev, crtc) {
2565 if (!to_intel_crtc(crtc)->active)
2568 if (crtc == for_crtc)
2574 pipe_size = ddb_size / config->num_pipes_active;
2575 alloc->start = nth_active_pipe * ddb_size / config->num_pipes_active;
2576 alloc->end = alloc->start + pipe_size;
2579 static unsigned int skl_cursor_allocation(const struct intel_wm_config *config)
2581 if (config->num_pipes_active == 1)
2587 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry *entry, u32 reg)
2589 entry->start = reg & 0x3ff;
2590 entry->end = (reg >> 16) & 0x3ff;
2595 void skl_ddb_get_hw_state(struct drm_i915_private *dev_priv,
2596 struct skl_ddb_allocation *ddb /* out */)
2602 for_each_pipe(dev_priv, pipe) {
2603 for_each_plane(dev_priv, pipe, plane) {
2604 val = I915_READ(PLANE_BUF_CFG(pipe, plane));
2605 skl_ddb_entry_init_from_hw(&ddb->plane[pipe][plane],
2609 val = I915_READ(CUR_BUF_CFG(pipe));
2610 skl_ddb_entry_init_from_hw(&ddb->cursor[pipe], val);
2615 skl_plane_relative_data_rate(const struct intel_plane_wm_parameters *p)
2617 return p->horiz_pixels * p->vert_pixels * p->bytes_per_pixel;
2621 * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
2622 * a 8192x4096@32bpp framebuffer:
2623 * 3 * 4096 * 8192 * 4 < 2^32
2626 skl_get_total_relative_data_rate(struct intel_crtc *intel_crtc,
2627 const struct skl_pipe_wm_parameters *params)
2629 unsigned int total_data_rate = 0;
2632 for (plane = 0; plane < intel_num_planes(intel_crtc); plane++) {
2633 const struct intel_plane_wm_parameters *p;
2635 p = ¶ms->plane[plane];
2639 total_data_rate += skl_plane_relative_data_rate(p);
2642 return total_data_rate;
2646 skl_allocate_pipe_ddb(struct drm_crtc *crtc,
2647 const struct intel_wm_config *config,
2648 const struct skl_pipe_wm_parameters *params,
2649 struct skl_ddb_allocation *ddb /* out */)
2651 struct drm_device *dev = crtc->dev;
2652 struct drm_i915_private *dev_priv = dev->dev_private;
2653 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2654 enum pipe pipe = intel_crtc->pipe;
2655 struct skl_ddb_entry *alloc = &ddb->pipe[pipe];
2656 uint16_t alloc_size, start, cursor_blocks;
2657 uint16_t minimum[I915_MAX_PLANES];
2658 unsigned int total_data_rate;
2661 skl_ddb_get_pipe_allocation_limits(dev, crtc, config, params, alloc);
2662 alloc_size = skl_ddb_entry_size(alloc);
2663 if (alloc_size == 0) {
2664 memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
2665 memset(&ddb->cursor[pipe], 0, sizeof(ddb->cursor[pipe]));
2669 cursor_blocks = skl_cursor_allocation(config);
2670 ddb->cursor[pipe].start = alloc->end - cursor_blocks;
2671 ddb->cursor[pipe].end = alloc->end;
2673 alloc_size -= cursor_blocks;
2674 alloc->end -= cursor_blocks;
2676 /* 1. Allocate the mininum required blocks for each active plane */
2677 for_each_plane(dev_priv, pipe, plane) {
2678 const struct intel_plane_wm_parameters *p;
2680 p = ¶ms->plane[plane];
2685 alloc_size -= minimum[plane];
2689 * 2. Distribute the remaining space in proportion to the amount of
2690 * data each plane needs to fetch from memory.
2692 * FIXME: we may not allocate every single block here.
2694 total_data_rate = skl_get_total_relative_data_rate(intel_crtc, params);
2696 start = alloc->start;
2697 for (plane = 0; plane < intel_num_planes(intel_crtc); plane++) {
2698 const struct intel_plane_wm_parameters *p;
2699 unsigned int data_rate;
2700 uint16_t plane_blocks;
2702 p = ¶ms->plane[plane];
2706 data_rate = skl_plane_relative_data_rate(p);
2709 * promote the expression to 64 bits to avoid overflowing, the
2710 * result is < available as data_rate / total_data_rate < 1
2712 plane_blocks = minimum[plane];
2713 plane_blocks += div_u64((uint64_t)alloc_size * data_rate,
2716 ddb->plane[pipe][plane].start = start;
2717 ddb->plane[pipe][plane].end = start + plane_blocks;
2719 start += plane_blocks;
2724 static uint32_t skl_pipe_pixel_rate(const struct intel_crtc_state *config)
2726 /* TODO: Take into account the scalers once we support them */
2727 return config->base.adjusted_mode.crtc_clock;
2731 * The max latency should be 257 (max the punit can code is 255 and we add 2us
2732 * for the read latency) and bytes_per_pixel should always be <= 8, so that
2733 * should allow pixel_rate up to ~2 GHz which seems sufficient since max
2734 * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
2736 static uint32_t skl_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
2739 uint32_t wm_intermediate_val, ret;
2744 wm_intermediate_val = latency * pixel_rate * bytes_per_pixel / 512;
2745 ret = DIV_ROUND_UP(wm_intermediate_val, 1000);
2750 static uint32_t skl_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
2751 uint32_t horiz_pixels, uint8_t bytes_per_pixel,
2752 uint64_t tiling, uint32_t latency)
2755 uint32_t plane_bytes_per_line, plane_blocks_per_line;
2756 uint32_t wm_intermediate_val;
2761 plane_bytes_per_line = horiz_pixels * bytes_per_pixel;
2763 if (tiling == I915_FORMAT_MOD_Y_TILED ||
2764 tiling == I915_FORMAT_MOD_Yf_TILED) {
2765 plane_bytes_per_line *= 4;
2766 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
2767 plane_blocks_per_line /= 4;
2769 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
2772 wm_intermediate_val = latency * pixel_rate;
2773 ret = DIV_ROUND_UP(wm_intermediate_val, pipe_htotal * 1000) *
2774 plane_blocks_per_line;
2779 static bool skl_ddb_allocation_changed(const struct skl_ddb_allocation *new_ddb,
2780 const struct intel_crtc *intel_crtc)
2782 struct drm_device *dev = intel_crtc->base.dev;
2783 struct drm_i915_private *dev_priv = dev->dev_private;
2784 const struct skl_ddb_allocation *cur_ddb = &dev_priv->wm.skl_hw.ddb;
2785 enum pipe pipe = intel_crtc->pipe;
2787 if (memcmp(new_ddb->plane[pipe], cur_ddb->plane[pipe],
2788 sizeof(new_ddb->plane[pipe])))
2791 if (memcmp(&new_ddb->cursor[pipe], &cur_ddb->cursor[pipe],
2792 sizeof(new_ddb->cursor[pipe])))
2798 static void skl_compute_wm_global_parameters(struct drm_device *dev,
2799 struct intel_wm_config *config)
2801 struct drm_crtc *crtc;
2802 struct drm_plane *plane;
2804 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
2805 config->num_pipes_active += to_intel_crtc(crtc)->active;
2807 /* FIXME: I don't think we need those two global parameters on SKL */
2808 list_for_each_entry(plane, &dev->mode_config.plane_list, head) {
2809 struct intel_plane *intel_plane = to_intel_plane(plane);
2811 config->sprites_enabled |= intel_plane->wm.enabled;
2812 config->sprites_scaled |= intel_plane->wm.scaled;
2816 static void skl_compute_wm_pipe_parameters(struct drm_crtc *crtc,
2817 struct skl_pipe_wm_parameters *p)
2819 struct drm_device *dev = crtc->dev;
2820 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2821 enum pipe pipe = intel_crtc->pipe;
2822 struct drm_plane *plane;
2823 struct drm_framebuffer *fb;
2824 int i = 1; /* Index for sprite planes start */
2826 p->active = intel_crtc->active;
2828 p->pipe_htotal = intel_crtc->config->base.adjusted_mode.crtc_htotal;
2829 p->pixel_rate = skl_pipe_pixel_rate(intel_crtc->config);
2831 fb = crtc->primary->state->fb;
2833 p->plane[0].enabled = true;
2834 p->plane[0].bytes_per_pixel = fb->bits_per_pixel / 8;
2835 p->plane[0].tiling = fb->modifier[0];
2837 p->plane[0].enabled = false;
2838 p->plane[0].bytes_per_pixel = 0;
2839 p->plane[0].tiling = DRM_FORMAT_MOD_NONE;
2841 p->plane[0].horiz_pixels = intel_crtc->config->pipe_src_w;
2842 p->plane[0].vert_pixels = intel_crtc->config->pipe_src_h;
2843 p->plane[0].rotation = crtc->primary->state->rotation;
2845 fb = crtc->cursor->state->fb;
2847 p->cursor.enabled = true;
2848 p->cursor.bytes_per_pixel = fb->bits_per_pixel / 8;
2849 p->cursor.horiz_pixels = crtc->cursor->state->crtc_w;
2850 p->cursor.vert_pixels = crtc->cursor->state->crtc_h;
2852 p->cursor.enabled = false;
2853 p->cursor.bytes_per_pixel = 0;
2854 p->cursor.horiz_pixels = 64;
2855 p->cursor.vert_pixels = 64;
2859 list_for_each_entry(plane, &dev->mode_config.plane_list, head) {
2860 struct intel_plane *intel_plane = to_intel_plane(plane);
2862 if (intel_plane->pipe == pipe &&
2863 plane->type == DRM_PLANE_TYPE_OVERLAY)
2864 p->plane[i++] = intel_plane->wm;
2868 static bool skl_compute_plane_wm(const struct drm_i915_private *dev_priv,
2869 struct skl_pipe_wm_parameters *p,
2870 struct intel_plane_wm_parameters *p_params,
2871 uint16_t ddb_allocation,
2873 uint16_t *out_blocks, /* out */
2874 uint8_t *out_lines /* out */)
2876 uint32_t latency = dev_priv->wm.skl_latency[level];
2877 uint32_t method1, method2;
2878 uint32_t plane_bytes_per_line, plane_blocks_per_line;
2879 uint32_t res_blocks, res_lines;
2880 uint32_t selected_result;
2882 if (latency == 0 || !p->active || !p_params->enabled)
2885 method1 = skl_wm_method1(p->pixel_rate,
2886 p_params->bytes_per_pixel,
2888 method2 = skl_wm_method2(p->pixel_rate,
2890 p_params->horiz_pixels,
2891 p_params->bytes_per_pixel,
2895 plane_bytes_per_line = p_params->horiz_pixels *
2896 p_params->bytes_per_pixel;
2897 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
2899 if (p_params->tiling == I915_FORMAT_MOD_Y_TILED ||
2900 p_params->tiling == I915_FORMAT_MOD_Yf_TILED) {
2901 uint32_t min_scanlines = 4;
2902 uint32_t y_tile_minimum;
2903 if (intel_rotation_90_or_270(p_params->rotation)) {
2904 switch (p_params->bytes_per_pixel) {
2912 WARN(1, "Unsupported pixel depth for rotation");
2915 y_tile_minimum = plane_blocks_per_line * min_scanlines;
2916 selected_result = max(method2, y_tile_minimum);
2918 if ((ddb_allocation / plane_blocks_per_line) >= 1)
2919 selected_result = min(method1, method2);
2921 selected_result = method1;
2924 res_blocks = selected_result + 1;
2925 res_lines = DIV_ROUND_UP(selected_result, plane_blocks_per_line);
2927 if (level >= 1 && level <= 7) {
2928 if (p_params->tiling == I915_FORMAT_MOD_Y_TILED ||
2929 p_params->tiling == I915_FORMAT_MOD_Yf_TILED)
2935 if (res_blocks >= ddb_allocation || res_lines > 31)
2938 *out_blocks = res_blocks;
2939 *out_lines = res_lines;
2944 static void skl_compute_wm_level(const struct drm_i915_private *dev_priv,
2945 struct skl_ddb_allocation *ddb,
2946 struct skl_pipe_wm_parameters *p,
2950 struct skl_wm_level *result)
2952 uint16_t ddb_blocks;
2955 for (i = 0; i < num_planes; i++) {
2956 ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][i]);
2958 result->plane_en[i] = skl_compute_plane_wm(dev_priv,
2962 &result->plane_res_b[i],
2963 &result->plane_res_l[i]);
2966 ddb_blocks = skl_ddb_entry_size(&ddb->cursor[pipe]);
2967 result->cursor_en = skl_compute_plane_wm(dev_priv, p, &p->cursor,
2969 &result->cursor_res_b,
2970 &result->cursor_res_l);
2974 skl_compute_linetime_wm(struct drm_crtc *crtc, struct skl_pipe_wm_parameters *p)
2976 if (!to_intel_crtc(crtc)->active)
2979 return DIV_ROUND_UP(8 * p->pipe_htotal * 1000, p->pixel_rate);
2983 static void skl_compute_transition_wm(struct drm_crtc *crtc,
2984 struct skl_pipe_wm_parameters *params,
2985 struct skl_wm_level *trans_wm /* out */)
2987 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2990 if (!params->active)
2993 /* Until we know more, just disable transition WMs */
2994 for (i = 0; i < intel_num_planes(intel_crtc); i++)
2995 trans_wm->plane_en[i] = false;
2996 trans_wm->cursor_en = false;
2999 static void skl_compute_pipe_wm(struct drm_crtc *crtc,
3000 struct skl_ddb_allocation *ddb,
3001 struct skl_pipe_wm_parameters *params,
3002 struct skl_pipe_wm *pipe_wm)
3004 struct drm_device *dev = crtc->dev;
3005 const struct drm_i915_private *dev_priv = dev->dev_private;
3006 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3007 int level, max_level = ilk_wm_max_level(dev);
3009 for (level = 0; level <= max_level; level++) {
3010 skl_compute_wm_level(dev_priv, ddb, params, intel_crtc->pipe,
3011 level, intel_num_planes(intel_crtc),
3012 &pipe_wm->wm[level]);
3014 pipe_wm->linetime = skl_compute_linetime_wm(crtc, params);
3016 skl_compute_transition_wm(crtc, params, &pipe_wm->trans_wm);
3019 static void skl_compute_wm_results(struct drm_device *dev,
3020 struct skl_pipe_wm_parameters *p,
3021 struct skl_pipe_wm *p_wm,
3022 struct skl_wm_values *r,
3023 struct intel_crtc *intel_crtc)
3025 int level, max_level = ilk_wm_max_level(dev);
3026 enum pipe pipe = intel_crtc->pipe;
3030 for (level = 0; level <= max_level; level++) {
3031 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3034 temp |= p_wm->wm[level].plane_res_l[i] <<
3035 PLANE_WM_LINES_SHIFT;
3036 temp |= p_wm->wm[level].plane_res_b[i];
3037 if (p_wm->wm[level].plane_en[i])
3038 temp |= PLANE_WM_EN;
3040 r->plane[pipe][i][level] = temp;
3045 temp |= p_wm->wm[level].cursor_res_l << PLANE_WM_LINES_SHIFT;
3046 temp |= p_wm->wm[level].cursor_res_b;
3048 if (p_wm->wm[level].cursor_en)
3049 temp |= PLANE_WM_EN;
3051 r->cursor[pipe][level] = temp;
3055 /* transition WMs */
3056 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3058 temp |= p_wm->trans_wm.plane_res_l[i] << PLANE_WM_LINES_SHIFT;
3059 temp |= p_wm->trans_wm.plane_res_b[i];
3060 if (p_wm->trans_wm.plane_en[i])
3061 temp |= PLANE_WM_EN;
3063 r->plane_trans[pipe][i] = temp;
3067 temp |= p_wm->trans_wm.cursor_res_l << PLANE_WM_LINES_SHIFT;
3068 temp |= p_wm->trans_wm.cursor_res_b;
3069 if (p_wm->trans_wm.cursor_en)
3070 temp |= PLANE_WM_EN;
3072 r->cursor_trans[pipe] = temp;
3074 r->wm_linetime[pipe] = p_wm->linetime;
3077 static void skl_ddb_entry_write(struct drm_i915_private *dev_priv, uint32_t reg,
3078 const struct skl_ddb_entry *entry)
3081 I915_WRITE(reg, (entry->end - 1) << 16 | entry->start);
3086 static void skl_write_wm_values(struct drm_i915_private *dev_priv,
3087 const struct skl_wm_values *new)
3089 struct drm_device *dev = dev_priv->dev;
3090 struct intel_crtc *crtc;
3092 list_for_each_entry(crtc, &dev->mode_config.crtc_list, base.head) {
3093 int i, level, max_level = ilk_wm_max_level(dev);
3094 enum pipe pipe = crtc->pipe;
3096 if (!new->dirty[pipe])
3099 I915_WRITE(PIPE_WM_LINETIME(pipe), new->wm_linetime[pipe]);
3101 for (level = 0; level <= max_level; level++) {
3102 for (i = 0; i < intel_num_planes(crtc); i++)
3103 I915_WRITE(PLANE_WM(pipe, i, level),
3104 new->plane[pipe][i][level]);
3105 I915_WRITE(CUR_WM(pipe, level),
3106 new->cursor[pipe][level]);
3108 for (i = 0; i < intel_num_planes(crtc); i++)
3109 I915_WRITE(PLANE_WM_TRANS(pipe, i),
3110 new->plane_trans[pipe][i]);
3111 I915_WRITE(CUR_WM_TRANS(pipe), new->cursor_trans[pipe]);
3113 for (i = 0; i < intel_num_planes(crtc); i++)
3114 skl_ddb_entry_write(dev_priv,
3115 PLANE_BUF_CFG(pipe, i),
3116 &new->ddb.plane[pipe][i]);
3118 skl_ddb_entry_write(dev_priv, CUR_BUF_CFG(pipe),
3119 &new->ddb.cursor[pipe]);
3124 * When setting up a new DDB allocation arrangement, we need to correctly
3125 * sequence the times at which the new allocations for the pipes are taken into
3126 * account or we'll have pipes fetching from space previously allocated to
3129 * Roughly the sequence looks like:
3130 * 1. re-allocate the pipe(s) with the allocation being reduced and not
3131 * overlapping with a previous light-up pipe (another way to put it is:
3132 * pipes with their new allocation strickly included into their old ones).
3133 * 2. re-allocate the other pipes that get their allocation reduced
3134 * 3. allocate the pipes having their allocation increased
3136 * Steps 1. and 2. are here to take care of the following case:
3137 * - Initially DDB looks like this:
3140 * - pipe B has a reduced DDB allocation that overlaps with the old pipe C
3144 * We need to sequence the re-allocation: C, B, A (and not B, C, A).
3148 skl_wm_flush_pipe(struct drm_i915_private *dev_priv, enum pipe pipe, int pass)
3152 DRM_DEBUG_KMS("flush pipe %c (pass %d)\n", pipe_name(pipe), pass);
3154 for_each_plane(dev_priv, pipe, plane) {
3155 I915_WRITE(PLANE_SURF(pipe, plane),
3156 I915_READ(PLANE_SURF(pipe, plane)));
3158 I915_WRITE(CURBASE(pipe), I915_READ(CURBASE(pipe)));
3162 skl_ddb_allocation_included(const struct skl_ddb_allocation *old,
3163 const struct skl_ddb_allocation *new,
3166 uint16_t old_size, new_size;
3168 old_size = skl_ddb_entry_size(&old->pipe[pipe]);
3169 new_size = skl_ddb_entry_size(&new->pipe[pipe]);
3171 return old_size != new_size &&
3172 new->pipe[pipe].start >= old->pipe[pipe].start &&
3173 new->pipe[pipe].end <= old->pipe[pipe].end;
3176 static void skl_flush_wm_values(struct drm_i915_private *dev_priv,
3177 struct skl_wm_values *new_values)
3179 struct drm_device *dev = dev_priv->dev;
3180 struct skl_ddb_allocation *cur_ddb, *new_ddb;
3181 bool reallocated[I915_MAX_PIPES] = {false, false, false};
3182 struct intel_crtc *crtc;
3185 new_ddb = &new_values->ddb;
3186 cur_ddb = &dev_priv->wm.skl_hw.ddb;
3189 * First pass: flush the pipes with the new allocation contained into
3192 * We'll wait for the vblank on those pipes to ensure we can safely
3193 * re-allocate the freed space without this pipe fetching from it.
3195 for_each_intel_crtc(dev, crtc) {
3201 if (!skl_ddb_allocation_included(cur_ddb, new_ddb, pipe))
3204 skl_wm_flush_pipe(dev_priv, pipe, 1);
3205 intel_wait_for_vblank(dev, pipe);
3207 reallocated[pipe] = true;
3212 * Second pass: flush the pipes that are having their allocation
3213 * reduced, but overlapping with a previous allocation.
3215 * Here as well we need to wait for the vblank to make sure the freed
3216 * space is not used anymore.
3218 for_each_intel_crtc(dev, crtc) {
3224 if (reallocated[pipe])
3227 if (skl_ddb_entry_size(&new_ddb->pipe[pipe]) <
3228 skl_ddb_entry_size(&cur_ddb->pipe[pipe])) {
3229 skl_wm_flush_pipe(dev_priv, pipe, 2);
3230 intel_wait_for_vblank(dev, pipe);
3231 reallocated[pipe] = true;
3236 * Third pass: flush the pipes that got more space allocated.
3238 * We don't need to actively wait for the update here, next vblank
3239 * will just get more DDB space with the correct WM values.
3241 for_each_intel_crtc(dev, crtc) {
3248 * At this point, only the pipes more space than before are
3249 * left to re-allocate.
3251 if (reallocated[pipe])
3254 skl_wm_flush_pipe(dev_priv, pipe, 3);
3258 static bool skl_update_pipe_wm(struct drm_crtc *crtc,
3259 struct skl_pipe_wm_parameters *params,
3260 struct intel_wm_config *config,
3261 struct skl_ddb_allocation *ddb, /* out */
3262 struct skl_pipe_wm *pipe_wm /* out */)
3264 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3266 skl_compute_wm_pipe_parameters(crtc, params);
3267 skl_allocate_pipe_ddb(crtc, config, params, ddb);
3268 skl_compute_pipe_wm(crtc, ddb, params, pipe_wm);
3270 if (!memcmp(&intel_crtc->wm.skl_active, pipe_wm, sizeof(*pipe_wm)))
3273 intel_crtc->wm.skl_active = *pipe_wm;
3277 static void skl_update_other_pipe_wm(struct drm_device *dev,
3278 struct drm_crtc *crtc,
3279 struct intel_wm_config *config,
3280 struct skl_wm_values *r)
3282 struct intel_crtc *intel_crtc;
3283 struct intel_crtc *this_crtc = to_intel_crtc(crtc);
3286 * If the WM update hasn't changed the allocation for this_crtc (the
3287 * crtc we are currently computing the new WM values for), other
3288 * enabled crtcs will keep the same allocation and we don't need to
3289 * recompute anything for them.
3291 if (!skl_ddb_allocation_changed(&r->ddb, this_crtc))
3295 * Otherwise, because of this_crtc being freshly enabled/disabled, the
3296 * other active pipes need new DDB allocation and WM values.
3298 list_for_each_entry(intel_crtc, &dev->mode_config.crtc_list,
3300 struct skl_pipe_wm_parameters params = {};
3301 struct skl_pipe_wm pipe_wm = {};
3304 if (this_crtc->pipe == intel_crtc->pipe)
3307 if (!intel_crtc->active)
3310 wm_changed = skl_update_pipe_wm(&intel_crtc->base,
3315 * If we end up re-computing the other pipe WM values, it's
3316 * because it was really needed, so we expect the WM values to
3319 WARN_ON(!wm_changed);
3321 skl_compute_wm_results(dev, ¶ms, &pipe_wm, r, intel_crtc);
3322 r->dirty[intel_crtc->pipe] = true;
3326 static void skl_update_wm(struct drm_crtc *crtc)
3328 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3329 struct drm_device *dev = crtc->dev;
3330 struct drm_i915_private *dev_priv = dev->dev_private;
3331 struct skl_pipe_wm_parameters params = {};
3332 struct skl_wm_values *results = &dev_priv->wm.skl_results;
3333 struct skl_pipe_wm pipe_wm = {};
3334 struct intel_wm_config config = {};
3336 memset(results, 0, sizeof(*results));
3338 skl_compute_wm_global_parameters(dev, &config);
3340 if (!skl_update_pipe_wm(crtc, ¶ms, &config,
3341 &results->ddb, &pipe_wm))
3344 skl_compute_wm_results(dev, ¶ms, &pipe_wm, results, intel_crtc);
3345 results->dirty[intel_crtc->pipe] = true;
3347 skl_update_other_pipe_wm(dev, crtc, &config, results);
3348 skl_write_wm_values(dev_priv, results);
3349 skl_flush_wm_values(dev_priv, results);
3351 /* store the new configuration */
3352 dev_priv->wm.skl_hw = *results;
3356 skl_update_sprite_wm(struct drm_plane *plane, struct drm_crtc *crtc,
3357 uint32_t sprite_width, uint32_t sprite_height,
3358 int pixel_size, bool enabled, bool scaled)
3360 struct intel_plane *intel_plane = to_intel_plane(plane);
3361 struct drm_framebuffer *fb = plane->state->fb;
3363 intel_plane->wm.enabled = enabled;
3364 intel_plane->wm.scaled = scaled;
3365 intel_plane->wm.horiz_pixels = sprite_width;
3366 intel_plane->wm.vert_pixels = sprite_height;
3367 intel_plane->wm.bytes_per_pixel = pixel_size;
3368 intel_plane->wm.tiling = DRM_FORMAT_MOD_NONE;
3370 * Framebuffer can be NULL on plane disable, but it does not
3371 * matter for watermarks if we assume no tiling in that case.
3374 intel_plane->wm.tiling = fb->modifier[0];
3375 intel_plane->wm.rotation = plane->state->rotation;
3377 skl_update_wm(crtc);
3380 static void ilk_update_wm(struct drm_crtc *crtc)
3382 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3383 struct drm_device *dev = crtc->dev;
3384 struct drm_i915_private *dev_priv = dev->dev_private;
3385 struct ilk_wm_maximums max;
3386 struct ilk_pipe_wm_parameters params = {};
3387 struct ilk_wm_values results = {};
3388 enum intel_ddb_partitioning partitioning;
3389 struct intel_pipe_wm pipe_wm = {};
3390 struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
3391 struct intel_wm_config config = {};
3393 ilk_compute_wm_parameters(crtc, ¶ms);
3395 intel_compute_pipe_wm(crtc, ¶ms, &pipe_wm);
3397 if (!memcmp(&intel_crtc->wm.active, &pipe_wm, sizeof(pipe_wm)))
3400 intel_crtc->wm.active = pipe_wm;
3402 ilk_compute_wm_config(dev, &config);
3404 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
3405 ilk_wm_merge(dev, &config, &max, &lp_wm_1_2);
3407 /* 5/6 split only in single pipe config on IVB+ */
3408 if (INTEL_INFO(dev)->gen >= 7 &&
3409 config.num_pipes_active == 1 && config.sprites_enabled) {
3410 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
3411 ilk_wm_merge(dev, &config, &max, &lp_wm_5_6);
3413 best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
3415 best_lp_wm = &lp_wm_1_2;
3418 partitioning = (best_lp_wm == &lp_wm_1_2) ?
3419 INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
3421 ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
3423 ilk_write_wm_values(dev_priv, &results);
3427 ilk_update_sprite_wm(struct drm_plane *plane,
3428 struct drm_crtc *crtc,
3429 uint32_t sprite_width, uint32_t sprite_height,
3430 int pixel_size, bool enabled, bool scaled)
3432 struct drm_device *dev = plane->dev;
3433 struct intel_plane *intel_plane = to_intel_plane(plane);
3435 intel_plane->wm.enabled = enabled;
3436 intel_plane->wm.scaled = scaled;
3437 intel_plane->wm.horiz_pixels = sprite_width;
3438 intel_plane->wm.vert_pixels = sprite_width;
3439 intel_plane->wm.bytes_per_pixel = pixel_size;
3442 * IVB workaround: must disable low power watermarks for at least
3443 * one frame before enabling scaling. LP watermarks can be re-enabled
3444 * when scaling is disabled.
3446 * WaCxSRDisabledForSpriteScaling:ivb
3448 if (IS_IVYBRIDGE(dev) && scaled && ilk_disable_lp_wm(dev))
3449 intel_wait_for_vblank(dev, intel_plane->pipe);
3451 ilk_update_wm(crtc);
3454 static void skl_pipe_wm_active_state(uint32_t val,
3455 struct skl_pipe_wm *active,
3461 bool is_enabled = (val & PLANE_WM_EN) != 0;
3465 active->wm[level].plane_en[i] = is_enabled;
3466 active->wm[level].plane_res_b[i] =
3467 val & PLANE_WM_BLOCKS_MASK;
3468 active->wm[level].plane_res_l[i] =
3469 (val >> PLANE_WM_LINES_SHIFT) &
3470 PLANE_WM_LINES_MASK;
3472 active->wm[level].cursor_en = is_enabled;
3473 active->wm[level].cursor_res_b =
3474 val & PLANE_WM_BLOCKS_MASK;
3475 active->wm[level].cursor_res_l =
3476 (val >> PLANE_WM_LINES_SHIFT) &
3477 PLANE_WM_LINES_MASK;
3481 active->trans_wm.plane_en[i] = is_enabled;
3482 active->trans_wm.plane_res_b[i] =
3483 val & PLANE_WM_BLOCKS_MASK;
3484 active->trans_wm.plane_res_l[i] =
3485 (val >> PLANE_WM_LINES_SHIFT) &
3486 PLANE_WM_LINES_MASK;
3488 active->trans_wm.cursor_en = is_enabled;
3489 active->trans_wm.cursor_res_b =
3490 val & PLANE_WM_BLOCKS_MASK;
3491 active->trans_wm.cursor_res_l =
3492 (val >> PLANE_WM_LINES_SHIFT) &
3493 PLANE_WM_LINES_MASK;
3498 static void skl_pipe_wm_get_hw_state(struct drm_crtc *crtc)
3500 struct drm_device *dev = crtc->dev;
3501 struct drm_i915_private *dev_priv = dev->dev_private;
3502 struct skl_wm_values *hw = &dev_priv->wm.skl_hw;
3503 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3504 struct skl_pipe_wm *active = &intel_crtc->wm.skl_active;
3505 enum pipe pipe = intel_crtc->pipe;
3506 int level, i, max_level;
3509 max_level = ilk_wm_max_level(dev);
3511 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
3513 for (level = 0; level <= max_level; level++) {
3514 for (i = 0; i < intel_num_planes(intel_crtc); i++)
3515 hw->plane[pipe][i][level] =
3516 I915_READ(PLANE_WM(pipe, i, level));
3517 hw->cursor[pipe][level] = I915_READ(CUR_WM(pipe, level));
3520 for (i = 0; i < intel_num_planes(intel_crtc); i++)
3521 hw->plane_trans[pipe][i] = I915_READ(PLANE_WM_TRANS(pipe, i));
3522 hw->cursor_trans[pipe] = I915_READ(CUR_WM_TRANS(pipe));
3524 if (!intel_crtc->active)
3527 hw->dirty[pipe] = true;
3529 active->linetime = hw->wm_linetime[pipe];
3531 for (level = 0; level <= max_level; level++) {
3532 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3533 temp = hw->plane[pipe][i][level];
3534 skl_pipe_wm_active_state(temp, active, false,
3537 temp = hw->cursor[pipe][level];
3538 skl_pipe_wm_active_state(temp, active, false, true, i, level);
3541 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3542 temp = hw->plane_trans[pipe][i];
3543 skl_pipe_wm_active_state(temp, active, true, false, i, 0);
3546 temp = hw->cursor_trans[pipe];
3547 skl_pipe_wm_active_state(temp, active, true, true, i, 0);
3550 void skl_wm_get_hw_state(struct drm_device *dev)
3552 struct drm_i915_private *dev_priv = dev->dev_private;
3553 struct skl_ddb_allocation *ddb = &dev_priv->wm.skl_hw.ddb;
3554 struct drm_crtc *crtc;
3556 skl_ddb_get_hw_state(dev_priv, ddb);
3557 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
3558 skl_pipe_wm_get_hw_state(crtc);
3561 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
3563 struct drm_device *dev = crtc->dev;
3564 struct drm_i915_private *dev_priv = dev->dev_private;
3565 struct ilk_wm_values *hw = &dev_priv->wm.hw;
3566 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3567 struct intel_pipe_wm *active = &intel_crtc->wm.active;
3568 enum pipe pipe = intel_crtc->pipe;
3569 static const unsigned int wm0_pipe_reg[] = {
3570 [PIPE_A] = WM0_PIPEA_ILK,
3571 [PIPE_B] = WM0_PIPEB_ILK,
3572 [PIPE_C] = WM0_PIPEC_IVB,
3575 hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
3576 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
3577 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
3579 active->pipe_enabled = intel_crtc->active;
3581 if (active->pipe_enabled) {
3582 u32 tmp = hw->wm_pipe[pipe];
3585 * For active pipes LP0 watermark is marked as
3586 * enabled, and LP1+ watermaks as disabled since
3587 * we can't really reverse compute them in case
3588 * multiple pipes are active.
3590 active->wm[0].enable = true;
3591 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
3592 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
3593 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
3594 active->linetime = hw->wm_linetime[pipe];
3596 int level, max_level = ilk_wm_max_level(dev);
3599 * For inactive pipes, all watermark levels
3600 * should be marked as enabled but zeroed,
3601 * which is what we'd compute them to.
3603 for (level = 0; level <= max_level; level++)
3604 active->wm[level].enable = true;
3608 void ilk_wm_get_hw_state(struct drm_device *dev)
3610 struct drm_i915_private *dev_priv = dev->dev_private;
3611 struct ilk_wm_values *hw = &dev_priv->wm.hw;
3612 struct drm_crtc *crtc;
3614 for_each_crtc(dev, crtc)
3615 ilk_pipe_wm_get_hw_state(crtc);
3617 hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
3618 hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
3619 hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
3621 hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
3622 if (INTEL_INFO(dev)->gen >= 7) {
3623 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
3624 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
3627 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
3628 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
3629 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
3630 else if (IS_IVYBRIDGE(dev))
3631 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
3632 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
3635 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
3639 * intel_update_watermarks - update FIFO watermark values based on current modes
3641 * Calculate watermark values for the various WM regs based on current mode
3642 * and plane configuration.
3644 * There are several cases to deal with here:
3645 * - normal (i.e. non-self-refresh)
3646 * - self-refresh (SR) mode
3647 * - lines are large relative to FIFO size (buffer can hold up to 2)
3648 * - lines are small relative to FIFO size (buffer can hold more than 2
3649 * lines), so need to account for TLB latency
3651 * The normal calculation is:
3652 * watermark = dotclock * bytes per pixel * latency
3653 * where latency is platform & configuration dependent (we assume pessimal
3656 * The SR calculation is:
3657 * watermark = (trunc(latency/line time)+1) * surface width *
3660 * line time = htotal / dotclock
3661 * surface width = hdisplay for normal plane and 64 for cursor
3662 * and latency is assumed to be high, as above.
3664 * The final value programmed to the register should always be rounded up,
3665 * and include an extra 2 entries to account for clock crossings.
3667 * We don't use the sprite, so we can ignore that. And on Crestline we have
3668 * to set the non-SR watermarks to 8.
3670 void intel_update_watermarks(struct drm_crtc *crtc)
3672 struct drm_i915_private *dev_priv = crtc->dev->dev_private;
3674 if (dev_priv->display.update_wm)
3675 dev_priv->display.update_wm(crtc);
3678 void intel_update_sprite_watermarks(struct drm_plane *plane,
3679 struct drm_crtc *crtc,
3680 uint32_t sprite_width,
3681 uint32_t sprite_height,
3683 bool enabled, bool scaled)
3685 struct drm_i915_private *dev_priv = plane->dev->dev_private;
3687 if (dev_priv->display.update_sprite_wm)
3688 dev_priv->display.update_sprite_wm(plane, crtc,
3689 sprite_width, sprite_height,
3690 pixel_size, enabled, scaled);
3694 * Lock protecting IPS related data structures
3696 DEFINE_SPINLOCK(mchdev_lock);
3698 /* Global for IPS driver to get at the current i915 device. Protected by
3700 static struct drm_i915_private *i915_mch_dev;
3702 bool ironlake_set_drps(struct drm_device *dev, u8 val)
3704 struct drm_i915_private *dev_priv = dev->dev_private;
3707 assert_spin_locked(&mchdev_lock);
3709 rgvswctl = I915_READ16(MEMSWCTL);
3710 if (rgvswctl & MEMCTL_CMD_STS) {
3711 DRM_DEBUG("gpu busy, RCS change rejected\n");
3712 return false; /* still busy with another command */
3715 rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
3716 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
3717 I915_WRITE16(MEMSWCTL, rgvswctl);
3718 POSTING_READ16(MEMSWCTL);
3720 rgvswctl |= MEMCTL_CMD_STS;
3721 I915_WRITE16(MEMSWCTL, rgvswctl);
3726 static void ironlake_enable_drps(struct drm_device *dev)
3728 struct drm_i915_private *dev_priv = dev->dev_private;
3729 u32 rgvmodectl = I915_READ(MEMMODECTL);
3730 u8 fmax, fmin, fstart, vstart;
3732 spin_lock_irq(&mchdev_lock);
3734 /* Enable temp reporting */
3735 I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
3736 I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
3738 /* 100ms RC evaluation intervals */
3739 I915_WRITE(RCUPEI, 100000);
3740 I915_WRITE(RCDNEI, 100000);
3742 /* Set max/min thresholds to 90ms and 80ms respectively */
3743 I915_WRITE(RCBMAXAVG, 90000);
3744 I915_WRITE(RCBMINAVG, 80000);
3746 I915_WRITE(MEMIHYST, 1);
3748 /* Set up min, max, and cur for interrupt handling */
3749 fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
3750 fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
3751 fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
3752 MEMMODE_FSTART_SHIFT;
3754 vstart = (I915_READ(PXVFREQ_BASE + (fstart * 4)) & PXVFREQ_PX_MASK) >>
3757 dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
3758 dev_priv->ips.fstart = fstart;
3760 dev_priv->ips.max_delay = fstart;
3761 dev_priv->ips.min_delay = fmin;
3762 dev_priv->ips.cur_delay = fstart;
3764 DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
3765 fmax, fmin, fstart);
3767 I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
3770 * Interrupts will be enabled in ironlake_irq_postinstall
3773 I915_WRITE(VIDSTART, vstart);
3774 POSTING_READ(VIDSTART);
3776 rgvmodectl |= MEMMODE_SWMODE_EN;
3777 I915_WRITE(MEMMODECTL, rgvmodectl);
3779 if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
3780 DRM_ERROR("stuck trying to change perf mode\n");
3783 ironlake_set_drps(dev, fstart);
3785 dev_priv->ips.last_count1 = I915_READ(0x112e4) + I915_READ(0x112e8) +
3787 dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
3788 dev_priv->ips.last_count2 = I915_READ(0x112f4);
3789 dev_priv->ips.last_time2 = ktime_get_raw_ns();
3791 spin_unlock_irq(&mchdev_lock);
3794 static void ironlake_disable_drps(struct drm_device *dev)
3796 struct drm_i915_private *dev_priv = dev->dev_private;
3799 spin_lock_irq(&mchdev_lock);
3801 rgvswctl = I915_READ16(MEMSWCTL);
3803 /* Ack interrupts, disable EFC interrupt */
3804 I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
3805 I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
3806 I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
3807 I915_WRITE(DEIIR, DE_PCU_EVENT);
3808 I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
3810 /* Go back to the starting frequency */
3811 ironlake_set_drps(dev, dev_priv->ips.fstart);
3813 rgvswctl |= MEMCTL_CMD_STS;
3814 I915_WRITE(MEMSWCTL, rgvswctl);
3817 spin_unlock_irq(&mchdev_lock);
3820 /* There's a funny hw issue where the hw returns all 0 when reading from
3821 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
3822 * ourselves, instead of doing a rmw cycle (which might result in us clearing
3823 * all limits and the gpu stuck at whatever frequency it is at atm).
3825 static u32 intel_rps_limits(struct drm_i915_private *dev_priv, u8 val)
3829 /* Only set the down limit when we've reached the lowest level to avoid
3830 * getting more interrupts, otherwise leave this clear. This prevents a
3831 * race in the hw when coming out of rc6: There's a tiny window where
3832 * the hw runs at the minimal clock before selecting the desired
3833 * frequency, if the down threshold expires in that window we will not
3834 * receive a down interrupt. */
3835 if (IS_GEN9(dev_priv->dev)) {
3836 limits = (dev_priv->rps.max_freq_softlimit) << 23;
3837 if (val <= dev_priv->rps.min_freq_softlimit)
3838 limits |= (dev_priv->rps.min_freq_softlimit) << 14;
3840 limits = dev_priv->rps.max_freq_softlimit << 24;
3841 if (val <= dev_priv->rps.min_freq_softlimit)
3842 limits |= dev_priv->rps.min_freq_softlimit << 16;
3848 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
3851 u32 threshold_up = 0, threshold_down = 0; /* in % */
3852 u32 ei_up = 0, ei_down = 0;
3854 new_power = dev_priv->rps.power;
3855 switch (dev_priv->rps.power) {
3857 if (val > dev_priv->rps.efficient_freq + 1 && val > dev_priv->rps.cur_freq)
3858 new_power = BETWEEN;
3862 if (val <= dev_priv->rps.efficient_freq && val < dev_priv->rps.cur_freq)
3863 new_power = LOW_POWER;
3864 else if (val >= dev_priv->rps.rp0_freq && val > dev_priv->rps.cur_freq)
3865 new_power = HIGH_POWER;
3869 if (val < (dev_priv->rps.rp1_freq + dev_priv->rps.rp0_freq) >> 1 && val < dev_priv->rps.cur_freq)
3870 new_power = BETWEEN;
3873 /* Max/min bins are special */
3874 if (val <= dev_priv->rps.min_freq_softlimit)
3875 new_power = LOW_POWER;
3876 if (val >= dev_priv->rps.max_freq_softlimit)
3877 new_power = HIGH_POWER;
3878 if (new_power == dev_priv->rps.power)
3881 /* Note the units here are not exactly 1us, but 1280ns. */
3882 switch (new_power) {
3884 /* Upclock if more than 95% busy over 16ms */
3888 /* Downclock if less than 85% busy over 32ms */
3890 threshold_down = 85;
3894 /* Upclock if more than 90% busy over 13ms */
3898 /* Downclock if less than 75% busy over 32ms */
3900 threshold_down = 75;
3904 /* Upclock if more than 85% busy over 10ms */
3908 /* Downclock if less than 60% busy over 32ms */
3910 threshold_down = 60;
3914 I915_WRITE(GEN6_RP_UP_EI,
3915 GT_INTERVAL_FROM_US(dev_priv, ei_up));
3916 I915_WRITE(GEN6_RP_UP_THRESHOLD,
3917 GT_INTERVAL_FROM_US(dev_priv, (ei_up * threshold_up / 100)));
3919 I915_WRITE(GEN6_RP_DOWN_EI,
3920 GT_INTERVAL_FROM_US(dev_priv, ei_down));
3921 I915_WRITE(GEN6_RP_DOWN_THRESHOLD,
3922 GT_INTERVAL_FROM_US(dev_priv, (ei_down * threshold_down / 100)));
3924 I915_WRITE(GEN6_RP_CONTROL,
3925 GEN6_RP_MEDIA_TURBO |
3926 GEN6_RP_MEDIA_HW_NORMAL_MODE |
3927 GEN6_RP_MEDIA_IS_GFX |
3929 GEN6_RP_UP_BUSY_AVG |
3930 GEN6_RP_DOWN_IDLE_AVG);
3932 dev_priv->rps.power = new_power;
3933 dev_priv->rps.last_adj = 0;
3936 static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
3940 if (val > dev_priv->rps.min_freq_softlimit)
3941 mask |= GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
3942 if (val < dev_priv->rps.max_freq_softlimit)
3943 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD;
3945 mask &= dev_priv->pm_rps_events;
3947 return gen6_sanitize_rps_pm_mask(dev_priv, ~mask);
3950 /* gen6_set_rps is called to update the frequency request, but should also be
3951 * called when the range (min_delay and max_delay) is modified so that we can
3952 * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
3953 static void gen6_set_rps(struct drm_device *dev, u8 val)
3955 struct drm_i915_private *dev_priv = dev->dev_private;
3957 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3958 WARN_ON(val > dev_priv->rps.max_freq);
3959 WARN_ON(val < dev_priv->rps.min_freq);
3961 /* min/max delay may still have been modified so be sure to
3962 * write the limits value.
3964 if (val != dev_priv->rps.cur_freq) {
3965 gen6_set_rps_thresholds(dev_priv, val);
3968 I915_WRITE(GEN6_RPNSWREQ,
3969 GEN9_FREQUENCY(val));
3970 else if (IS_HASWELL(dev) || IS_BROADWELL(dev))
3971 I915_WRITE(GEN6_RPNSWREQ,
3972 HSW_FREQUENCY(val));
3974 I915_WRITE(GEN6_RPNSWREQ,
3975 GEN6_FREQUENCY(val) |
3977 GEN6_AGGRESSIVE_TURBO);
3980 /* Make sure we continue to get interrupts
3981 * until we hit the minimum or maximum frequencies.
3983 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, intel_rps_limits(dev_priv, val));
3984 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
3986 POSTING_READ(GEN6_RPNSWREQ);
3988 dev_priv->rps.cur_freq = val;
3989 trace_intel_gpu_freq_change(val * 50);
3992 static void valleyview_set_rps(struct drm_device *dev, u8 val)
3994 struct drm_i915_private *dev_priv = dev->dev_private;
3996 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3997 WARN_ON(val > dev_priv->rps.max_freq);
3998 WARN_ON(val < dev_priv->rps.min_freq);
4000 if (WARN_ONCE(IS_CHERRYVIEW(dev) && (val & 1),
4001 "Odd GPU freq value\n"))
4004 if (val != dev_priv->rps.cur_freq)
4005 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
4007 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
4009 dev_priv->rps.cur_freq = val;
4010 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
4013 /* vlv_set_rps_idle: Set the frequency to Rpn if Gfx clocks are down
4015 * * If Gfx is Idle, then
4016 * 1. Mask Turbo interrupts
4017 * 2. Bring up Gfx clock
4018 * 3. Change the freq to Rpn and wait till P-Unit updates freq
4019 * 4. Clear the Force GFX CLK ON bit so that Gfx can down
4020 * 5. Unmask Turbo interrupts
4022 static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
4024 struct drm_device *dev = dev_priv->dev;
4025 u32 val = dev_priv->rps.idle_freq;
4027 /* CHV and latest VLV don't need to force the gfx clock */
4028 if (IS_CHERRYVIEW(dev) || dev->pdev->revision >= 0xd) {
4029 valleyview_set_rps(dev_priv->dev, val);
4034 * When we are idle. Drop to min voltage state.
4037 if (dev_priv->rps.cur_freq <= val)
4040 /* Mask turbo interrupt so that they will not come in between */
4041 I915_WRITE(GEN6_PMINTRMSK,
4042 gen6_sanitize_rps_pm_mask(dev_priv, ~0));
4044 vlv_force_gfx_clock(dev_priv, true);
4046 dev_priv->rps.cur_freq = val;
4048 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
4050 if (wait_for(((vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS))
4051 & GENFREQSTATUS) == 0, 100))
4052 DRM_ERROR("timed out waiting for Punit\n");
4054 vlv_force_gfx_clock(dev_priv, false);
4056 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
4059 void gen6_rps_busy(struct drm_i915_private *dev_priv)
4061 mutex_lock(&dev_priv->rps.hw_lock);
4062 if (dev_priv->rps.enabled) {
4063 if (dev_priv->pm_rps_events & (GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_UP_EI_EXPIRED))
4064 gen6_rps_reset_ei(dev_priv);
4065 I915_WRITE(GEN6_PMINTRMSK,
4066 gen6_rps_pm_mask(dev_priv, dev_priv->rps.cur_freq));
4068 mutex_unlock(&dev_priv->rps.hw_lock);
4071 void gen6_rps_idle(struct drm_i915_private *dev_priv)
4073 struct drm_device *dev = dev_priv->dev;
4075 mutex_lock(&dev_priv->rps.hw_lock);
4076 if (dev_priv->rps.enabled) {
4077 if (IS_VALLEYVIEW(dev))
4078 vlv_set_rps_idle(dev_priv);
4080 gen6_set_rps(dev_priv->dev, dev_priv->rps.idle_freq);
4081 dev_priv->rps.last_adj = 0;
4082 I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
4084 mutex_unlock(&dev_priv->rps.hw_lock);
4087 void gen6_rps_boost(struct drm_i915_private *dev_priv)
4091 mutex_lock(&dev_priv->rps.hw_lock);
4092 val = dev_priv->rps.max_freq_softlimit;
4093 if (dev_priv->rps.enabled &&
4094 dev_priv->mm.busy &&
4095 dev_priv->rps.cur_freq < val) {
4096 intel_set_rps(dev_priv->dev, val);
4097 dev_priv->rps.last_adj = 0;
4099 mutex_unlock(&dev_priv->rps.hw_lock);
4102 void intel_set_rps(struct drm_device *dev, u8 val)
4104 if (IS_VALLEYVIEW(dev))
4105 valleyview_set_rps(dev, val);
4107 gen6_set_rps(dev, val);
4110 static void gen9_disable_rps(struct drm_device *dev)
4112 struct drm_i915_private *dev_priv = dev->dev_private;
4114 I915_WRITE(GEN6_RC_CONTROL, 0);
4115 I915_WRITE(GEN9_PG_ENABLE, 0);
4118 static void gen6_disable_rps(struct drm_device *dev)
4120 struct drm_i915_private *dev_priv = dev->dev_private;
4122 I915_WRITE(GEN6_RC_CONTROL, 0);
4123 I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
4126 static void cherryview_disable_rps(struct drm_device *dev)
4128 struct drm_i915_private *dev_priv = dev->dev_private;
4130 I915_WRITE(GEN6_RC_CONTROL, 0);
4133 static void valleyview_disable_rps(struct drm_device *dev)
4135 struct drm_i915_private *dev_priv = dev->dev_private;
4137 /* we're doing forcewake before Disabling RC6,
4138 * This what the BIOS expects when going into suspend */
4139 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4141 I915_WRITE(GEN6_RC_CONTROL, 0);
4143 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4146 static void intel_print_rc6_info(struct drm_device *dev, u32 mode)
4148 if (IS_VALLEYVIEW(dev)) {
4149 if (mode & (GEN7_RC_CTL_TO_MODE | GEN6_RC_CTL_EI_MODE(1)))
4150 mode = GEN6_RC_CTL_RC6_ENABLE;
4155 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s RC6p %s RC6pp %s\n",
4156 (mode & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off",
4157 (mode & GEN6_RC_CTL_RC6p_ENABLE) ? "on" : "off",
4158 (mode & GEN6_RC_CTL_RC6pp_ENABLE) ? "on" : "off");
4161 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s\n",
4162 (mode & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off");
4165 static int sanitize_rc6_option(const struct drm_device *dev, int enable_rc6)
4167 /* No RC6 before Ironlake */
4168 if (INTEL_INFO(dev)->gen < 5)
4171 /* RC6 is only on Ironlake mobile not on desktop */
4172 if (INTEL_INFO(dev)->gen == 5 && !IS_IRONLAKE_M(dev))
4175 /* Respect the kernel parameter if it is set */
4176 if (enable_rc6 >= 0) {
4180 mask = INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE |
4183 mask = INTEL_RC6_ENABLE;
4185 if ((enable_rc6 & mask) != enable_rc6)
4186 DRM_DEBUG_KMS("Adjusting RC6 mask to %d (requested %d, valid %d)\n",
4187 enable_rc6 & mask, enable_rc6, mask);
4189 return enable_rc6 & mask;
4192 /* Disable RC6 on Ironlake */
4193 if (INTEL_INFO(dev)->gen == 5)
4196 if (IS_IVYBRIDGE(dev))
4197 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
4199 return INTEL_RC6_ENABLE;
4202 int intel_enable_rc6(const struct drm_device *dev)
4204 return i915.enable_rc6;
4207 static void gen6_init_rps_frequencies(struct drm_device *dev)
4209 struct drm_i915_private *dev_priv = dev->dev_private;
4210 uint32_t rp_state_cap;
4211 u32 ddcc_status = 0;
4214 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
4215 /* All of these values are in units of 50MHz */
4216 dev_priv->rps.cur_freq = 0;
4217 /* static values from HW: RP0 > RP1 > RPn (min_freq) */
4218 dev_priv->rps.rp0_freq = (rp_state_cap >> 0) & 0xff;
4219 dev_priv->rps.rp1_freq = (rp_state_cap >> 8) & 0xff;
4220 dev_priv->rps.min_freq = (rp_state_cap >> 16) & 0xff;
4221 if (IS_SKYLAKE(dev)) {
4222 /* Store the frequency values in 16.66 MHZ units, which is
4223 the natural hardware unit for SKL */
4224 dev_priv->rps.rp0_freq *= GEN9_FREQ_SCALER;
4225 dev_priv->rps.rp1_freq *= GEN9_FREQ_SCALER;
4226 dev_priv->rps.min_freq *= GEN9_FREQ_SCALER;
4228 /* hw_max = RP0 until we check for overclocking */
4229 dev_priv->rps.max_freq = dev_priv->rps.rp0_freq;
4231 dev_priv->rps.efficient_freq = dev_priv->rps.rp1_freq;
4232 if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
4233 ret = sandybridge_pcode_read(dev_priv,
4234 HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL,
4237 dev_priv->rps.efficient_freq =
4239 ((ddcc_status >> 8) & 0xff),
4240 dev_priv->rps.min_freq,
4241 dev_priv->rps.max_freq);
4244 dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
4246 /* Preserve min/max settings in case of re-init */
4247 if (dev_priv->rps.max_freq_softlimit == 0)
4248 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
4250 if (dev_priv->rps.min_freq_softlimit == 0) {
4251 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
4252 dev_priv->rps.min_freq_softlimit =
4253 /* max(RPe, 450 MHz) */
4254 max(dev_priv->rps.efficient_freq, (u8) 9);
4256 dev_priv->rps.min_freq_softlimit =
4257 dev_priv->rps.min_freq;
4261 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
4262 static void gen9_enable_rps(struct drm_device *dev)
4264 struct drm_i915_private *dev_priv = dev->dev_private;
4266 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4268 gen6_init_rps_frequencies(dev);
4270 /* Program defaults and thresholds for RPS*/
4271 I915_WRITE(GEN6_RC_VIDEO_FREQ,
4272 GEN9_FREQUENCY(dev_priv->rps.rp1_freq));
4274 /* 1 second timeout*/
4275 I915_WRITE(GEN6_RP_DOWN_TIMEOUT,
4276 GT_INTERVAL_FROM_US(dev_priv, 1000000));
4278 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 0xa);
4280 /* Leaning on the below call to gen6_set_rps to program/setup the
4281 * Up/Down EI & threshold registers, as well as the RP_CONTROL,
4282 * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
4283 dev_priv->rps.power = HIGH_POWER; /* force a reset */
4284 gen6_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
4286 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4289 static void gen9_enable_rc6(struct drm_device *dev)
4291 struct drm_i915_private *dev_priv = dev->dev_private;
4292 struct intel_engine_cs *ring;
4293 uint32_t rc6_mask = 0;
4296 /* 1a: Software RC state - RC0 */
4297 I915_WRITE(GEN6_RC_STATE, 0);
4299 /* 1b: Get forcewake during program sequence. Although the driver
4300 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4301 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4303 /* 2a: Disable RC states. */
4304 I915_WRITE(GEN6_RC_CONTROL, 0);
4306 /* 2b: Program RC6 thresholds.*/
4307 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16);
4308 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
4309 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
4310 for_each_ring(ring, dev_priv, unused)
4311 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4312 I915_WRITE(GEN6_RC_SLEEP, 0);
4313 I915_WRITE(GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */
4315 /* 2c: Program Coarse Power Gating Policies. */
4316 I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS, 25);
4317 I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS, 25);
4319 /* 3a: Enable RC6 */
4320 if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
4321 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
4322 DRM_INFO("RC6 %s\n", (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ?
4324 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4325 GEN6_RC_CTL_EI_MODE(1) |
4328 /* 3b: Enable Coarse Power Gating only when RC6 is enabled */
4329 I915_WRITE(GEN9_PG_ENABLE, (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ? 3 : 0);
4331 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4335 static void gen8_enable_rps(struct drm_device *dev)
4337 struct drm_i915_private *dev_priv = dev->dev_private;
4338 struct intel_engine_cs *ring;
4339 uint32_t rc6_mask = 0;
4342 /* 1a: Software RC state - RC0 */
4343 I915_WRITE(GEN6_RC_STATE, 0);
4345 /* 1c & 1d: Get forcewake during program sequence. Although the driver
4346 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4347 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4349 /* 2a: Disable RC states. */
4350 I915_WRITE(GEN6_RC_CONTROL, 0);
4352 /* Initialize rps frequencies */
4353 gen6_init_rps_frequencies(dev);
4355 /* 2b: Program RC6 thresholds.*/
4356 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
4357 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
4358 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
4359 for_each_ring(ring, dev_priv, unused)
4360 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4361 I915_WRITE(GEN6_RC_SLEEP, 0);
4362 if (IS_BROADWELL(dev))
4363 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */
4365 I915_WRITE(GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
4368 if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
4369 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
4370 intel_print_rc6_info(dev, rc6_mask);
4371 if (IS_BROADWELL(dev))
4372 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4373 GEN7_RC_CTL_TO_MODE |
4376 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4377 GEN6_RC_CTL_EI_MODE(1) |
4380 /* 4 Program defaults and thresholds for RPS*/
4381 I915_WRITE(GEN6_RPNSWREQ,
4382 HSW_FREQUENCY(dev_priv->rps.rp1_freq));
4383 I915_WRITE(GEN6_RC_VIDEO_FREQ,
4384 HSW_FREQUENCY(dev_priv->rps.rp1_freq));
4385 /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
4386 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
4388 /* Docs recommend 900MHz, and 300 MHz respectively */
4389 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
4390 dev_priv->rps.max_freq_softlimit << 24 |
4391 dev_priv->rps.min_freq_softlimit << 16);
4393 I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
4394 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
4395 I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
4396 I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
4398 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
4401 I915_WRITE(GEN6_RP_CONTROL,
4402 GEN6_RP_MEDIA_TURBO |
4403 GEN6_RP_MEDIA_HW_NORMAL_MODE |
4404 GEN6_RP_MEDIA_IS_GFX |
4406 GEN6_RP_UP_BUSY_AVG |
4407 GEN6_RP_DOWN_IDLE_AVG);
4409 /* 6: Ring frequency + overclocking (our driver does this later */
4411 dev_priv->rps.power = HIGH_POWER; /* force a reset */
4412 gen6_set_rps(dev_priv->dev, dev_priv->rps.idle_freq);
4414 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4417 static void gen6_enable_rps(struct drm_device *dev)
4419 struct drm_i915_private *dev_priv = dev->dev_private;
4420 struct intel_engine_cs *ring;
4421 u32 rc6vids, pcu_mbox = 0, rc6_mask = 0;
4426 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4428 /* Here begins a magic sequence of register writes to enable
4429 * auto-downclocking.
4431 * Perhaps there might be some value in exposing these to
4434 I915_WRITE(GEN6_RC_STATE, 0);
4436 /* Clear the DBG now so we don't confuse earlier errors */
4437 if ((gtfifodbg = I915_READ(GTFIFODBG))) {
4438 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
4439 I915_WRITE(GTFIFODBG, gtfifodbg);
4442 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4444 /* Initialize rps frequencies */
4445 gen6_init_rps_frequencies(dev);
4447 /* disable the counters and set deterministic thresholds */
4448 I915_WRITE(GEN6_RC_CONTROL, 0);
4450 I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
4451 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
4452 I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
4453 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
4454 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
4456 for_each_ring(ring, dev_priv, i)
4457 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4459 I915_WRITE(GEN6_RC_SLEEP, 0);
4460 I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
4461 if (IS_IVYBRIDGE(dev))
4462 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
4464 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
4465 I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
4466 I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
4468 /* Check if we are enabling RC6 */
4469 rc6_mode = intel_enable_rc6(dev_priv->dev);
4470 if (rc6_mode & INTEL_RC6_ENABLE)
4471 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
4473 /* We don't use those on Haswell */
4474 if (!IS_HASWELL(dev)) {
4475 if (rc6_mode & INTEL_RC6p_ENABLE)
4476 rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
4478 if (rc6_mode & INTEL_RC6pp_ENABLE)
4479 rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
4482 intel_print_rc6_info(dev, rc6_mask);
4484 I915_WRITE(GEN6_RC_CONTROL,
4486 GEN6_RC_CTL_EI_MODE(1) |
4487 GEN6_RC_CTL_HW_ENABLE);
4489 /* Power down if completely idle for over 50ms */
4490 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
4491 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
4493 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_MIN_FREQ_TABLE, 0);
4495 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
4497 ret = sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &pcu_mbox);
4498 if (!ret && (pcu_mbox & (1<<31))) { /* OC supported */
4499 DRM_DEBUG_DRIVER("Overclocking supported. Max: %dMHz, Overclock max: %dMHz\n",
4500 (dev_priv->rps.max_freq_softlimit & 0xff) * 50,
4501 (pcu_mbox & 0xff) * 50);
4502 dev_priv->rps.max_freq = pcu_mbox & 0xff;
4505 dev_priv->rps.power = HIGH_POWER; /* force a reset */
4506 gen6_set_rps(dev_priv->dev, dev_priv->rps.idle_freq);
4509 ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
4510 if (IS_GEN6(dev) && ret) {
4511 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
4512 } else if (IS_GEN6(dev) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
4513 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
4514 GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
4515 rc6vids &= 0xffff00;
4516 rc6vids |= GEN6_ENCODE_RC6_VID(450);
4517 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
4519 DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
4522 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4525 static void __gen6_update_ring_freq(struct drm_device *dev)
4527 struct drm_i915_private *dev_priv = dev->dev_private;
4529 unsigned int gpu_freq;
4530 unsigned int max_ia_freq, min_ring_freq;
4531 int scaling_factor = 180;
4532 struct cpufreq_policy *policy;
4534 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4536 policy = cpufreq_cpu_get(0);
4538 max_ia_freq = policy->cpuinfo.max_freq;
4539 cpufreq_cpu_put(policy);
4542 * Default to measured freq if none found, PCU will ensure we
4545 max_ia_freq = tsc_khz;
4548 /* Convert from kHz to MHz */
4549 max_ia_freq /= 1000;
4551 min_ring_freq = I915_READ(DCLK) & 0xf;
4552 /* convert DDR frequency from units of 266.6MHz to bandwidth */
4553 min_ring_freq = mult_frac(min_ring_freq, 8, 3);
4556 * For each potential GPU frequency, load a ring frequency we'd like
4557 * to use for memory access. We do this by specifying the IA frequency
4558 * the PCU should use as a reference to determine the ring frequency.
4560 for (gpu_freq = dev_priv->rps.max_freq; gpu_freq >= dev_priv->rps.min_freq;
4562 int diff = dev_priv->rps.max_freq - gpu_freq;
4563 unsigned int ia_freq = 0, ring_freq = 0;
4565 if (INTEL_INFO(dev)->gen >= 8) {
4566 /* max(2 * GT, DDR). NB: GT is 50MHz units */
4567 ring_freq = max(min_ring_freq, gpu_freq);
4568 } else if (IS_HASWELL(dev)) {
4569 ring_freq = mult_frac(gpu_freq, 5, 4);
4570 ring_freq = max(min_ring_freq, ring_freq);
4571 /* leave ia_freq as the default, chosen by cpufreq */
4573 /* On older processors, there is no separate ring
4574 * clock domain, so in order to boost the bandwidth
4575 * of the ring, we need to upclock the CPU (ia_freq).
4577 * For GPU frequencies less than 750MHz,
4578 * just use the lowest ring freq.
4580 if (gpu_freq < min_freq)
4583 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
4584 ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
4587 sandybridge_pcode_write(dev_priv,
4588 GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
4589 ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
4590 ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
4595 void gen6_update_ring_freq(struct drm_device *dev)
4597 struct drm_i915_private *dev_priv = dev->dev_private;
4599 if (INTEL_INFO(dev)->gen < 6 || IS_VALLEYVIEW(dev))
4602 mutex_lock(&dev_priv->rps.hw_lock);
4603 __gen6_update_ring_freq(dev);
4604 mutex_unlock(&dev_priv->rps.hw_lock);
4607 static int cherryview_rps_max_freq(struct drm_i915_private *dev_priv)
4609 struct drm_device *dev = dev_priv->dev;
4612 if (dev->pdev->revision >= 0x20) {
4613 val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
4615 switch (INTEL_INFO(dev)->eu_total) {
4617 /* (2 * 4) config */
4618 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT);
4621 /* (2 * 6) config */
4622 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT);
4625 /* (2 * 8) config */
4627 /* Setting (2 * 8) Min RP0 for any other combination */
4628 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT);
4631 rp0 = (rp0 & FB_GFX_FREQ_FUSE_MASK);
4633 /* For pre-production hardware */
4634 val = vlv_punit_read(dev_priv, PUNIT_GPU_STATUS_REG);
4635 rp0 = (val >> PUNIT_GPU_STATUS_MAX_FREQ_SHIFT) &
4636 PUNIT_GPU_STATUS_MAX_FREQ_MASK;
4641 static int cherryview_rps_rpe_freq(struct drm_i915_private *dev_priv)
4645 val = vlv_punit_read(dev_priv, PUNIT_GPU_DUTYCYCLE_REG);
4646 rpe = (val >> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK;
4651 static int cherryview_rps_guar_freq(struct drm_i915_private *dev_priv)
4653 struct drm_device *dev = dev_priv->dev;
4656 if (dev->pdev->revision >= 0x20) {
4657 val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
4658 rp1 = (val & FB_GFX_FREQ_FUSE_MASK);
4660 /* For pre-production hardware */
4661 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
4662 rp1 = ((val >> PUNIT_GPU_STATUS_MAX_FREQ_SHIFT) &
4663 PUNIT_GPU_STATUS_MAX_FREQ_MASK);
4668 static int cherryview_rps_min_freq(struct drm_i915_private *dev_priv)
4670 struct drm_device *dev = dev_priv->dev;
4673 if (dev->pdev->revision >= 0x20) {
4674 val = vlv_punit_read(dev_priv, FB_GFX_FMIN_AT_VMIN_FUSE);
4675 rpn = ((val >> FB_GFX_FMIN_AT_VMIN_FUSE_SHIFT) &
4676 FB_GFX_FREQ_FUSE_MASK);
4677 } else { /* For pre-production hardware */
4678 val = vlv_punit_read(dev_priv, PUNIT_GPU_STATUS_REG);
4679 rpn = ((val >> PUNIT_GPU_STATIS_GFX_MIN_FREQ_SHIFT) &
4680 PUNIT_GPU_STATUS_GFX_MIN_FREQ_MASK);
4686 static int valleyview_rps_guar_freq(struct drm_i915_private *dev_priv)
4690 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
4692 rp1 = (val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT;
4697 static int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
4701 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
4703 rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
4705 rp0 = min_t(u32, rp0, 0xea);
4710 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
4714 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
4715 rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
4716 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
4717 rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
4722 static int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
4724 return vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
4727 /* Check that the pctx buffer wasn't move under us. */
4728 static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
4730 unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
4732 WARN_ON(pctx_addr != dev_priv->mm.stolen_base +
4733 dev_priv->vlv_pctx->stolen->start);
4737 /* Check that the pcbr address is not empty. */
4738 static void cherryview_check_pctx(struct drm_i915_private *dev_priv)
4740 unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
4742 WARN_ON((pctx_addr >> VLV_PCBR_ADDR_SHIFT) == 0);
4745 static void cherryview_setup_pctx(struct drm_device *dev)
4747 struct drm_i915_private *dev_priv = dev->dev_private;
4748 unsigned long pctx_paddr, paddr;
4749 struct i915_gtt *gtt = &dev_priv->gtt;
4751 int pctx_size = 32*1024;
4753 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
4755 pcbr = I915_READ(VLV_PCBR);
4756 if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) {
4757 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
4758 paddr = (dev_priv->mm.stolen_base +
4759 (gtt->stolen_size - pctx_size));
4761 pctx_paddr = (paddr & (~4095));
4762 I915_WRITE(VLV_PCBR, pctx_paddr);
4765 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
4768 static void valleyview_setup_pctx(struct drm_device *dev)
4770 struct drm_i915_private *dev_priv = dev->dev_private;
4771 struct drm_i915_gem_object *pctx;
4772 unsigned long pctx_paddr;
4774 int pctx_size = 24*1024;
4776 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
4778 pcbr = I915_READ(VLV_PCBR);
4780 /* BIOS set it up already, grab the pre-alloc'd space */
4783 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
4784 pctx = i915_gem_object_create_stolen_for_preallocated(dev_priv->dev,
4786 I915_GTT_OFFSET_NONE,
4791 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
4794 * From the Gunit register HAS:
4795 * The Gfx driver is expected to program this register and ensure
4796 * proper allocation within Gfx stolen memory. For example, this
4797 * register should be programmed such than the PCBR range does not
4798 * overlap with other ranges, such as the frame buffer, protected
4799 * memory, or any other relevant ranges.
4801 pctx = i915_gem_object_create_stolen(dev, pctx_size);
4803 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
4807 pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
4808 I915_WRITE(VLV_PCBR, pctx_paddr);
4811 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
4812 dev_priv->vlv_pctx = pctx;
4815 static void valleyview_cleanup_pctx(struct drm_device *dev)
4817 struct drm_i915_private *dev_priv = dev->dev_private;
4819 if (WARN_ON(!dev_priv->vlv_pctx))
4822 drm_gem_object_unreference(&dev_priv->vlv_pctx->base);
4823 dev_priv->vlv_pctx = NULL;
4826 static void valleyview_init_gt_powersave(struct drm_device *dev)
4828 struct drm_i915_private *dev_priv = dev->dev_private;
4831 valleyview_setup_pctx(dev);
4833 mutex_lock(&dev_priv->rps.hw_lock);
4835 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
4836 switch ((val >> 6) & 3) {
4839 dev_priv->mem_freq = 800;
4842 dev_priv->mem_freq = 1066;
4845 dev_priv->mem_freq = 1333;
4848 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
4850 dev_priv->rps.max_freq = valleyview_rps_max_freq(dev_priv);
4851 dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
4852 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
4853 intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
4854 dev_priv->rps.max_freq);
4856 dev_priv->rps.efficient_freq = valleyview_rps_rpe_freq(dev_priv);
4857 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
4858 intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
4859 dev_priv->rps.efficient_freq);
4861 dev_priv->rps.rp1_freq = valleyview_rps_guar_freq(dev_priv);
4862 DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
4863 intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
4864 dev_priv->rps.rp1_freq);
4866 dev_priv->rps.min_freq = valleyview_rps_min_freq(dev_priv);
4867 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
4868 intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
4869 dev_priv->rps.min_freq);
4871 dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
4873 /* Preserve min/max settings in case of re-init */
4874 if (dev_priv->rps.max_freq_softlimit == 0)
4875 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
4877 if (dev_priv->rps.min_freq_softlimit == 0)
4878 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
4880 mutex_unlock(&dev_priv->rps.hw_lock);
4883 static void cherryview_init_gt_powersave(struct drm_device *dev)
4885 struct drm_i915_private *dev_priv = dev->dev_private;
4888 cherryview_setup_pctx(dev);
4890 mutex_lock(&dev_priv->rps.hw_lock);
4892 mutex_lock(&dev_priv->dpio_lock);
4893 val = vlv_cck_read(dev_priv, CCK_FUSE_REG);
4894 mutex_unlock(&dev_priv->dpio_lock);
4896 switch ((val >> 2) & 0x7) {
4899 dev_priv->rps.cz_freq = 200;
4900 dev_priv->mem_freq = 1600;
4903 dev_priv->rps.cz_freq = 267;
4904 dev_priv->mem_freq = 1600;
4907 dev_priv->rps.cz_freq = 333;
4908 dev_priv->mem_freq = 2000;
4911 dev_priv->rps.cz_freq = 320;
4912 dev_priv->mem_freq = 1600;
4915 dev_priv->rps.cz_freq = 400;
4916 dev_priv->mem_freq = 1600;
4919 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
4921 dev_priv->rps.max_freq = cherryview_rps_max_freq(dev_priv);
4922 dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
4923 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
4924 intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
4925 dev_priv->rps.max_freq);
4927 dev_priv->rps.efficient_freq = cherryview_rps_rpe_freq(dev_priv);
4928 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
4929 intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
4930 dev_priv->rps.efficient_freq);
4932 dev_priv->rps.rp1_freq = cherryview_rps_guar_freq(dev_priv);
4933 DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
4934 intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
4935 dev_priv->rps.rp1_freq);
4937 dev_priv->rps.min_freq = cherryview_rps_min_freq(dev_priv);
4938 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
4939 intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
4940 dev_priv->rps.min_freq);
4942 WARN_ONCE((dev_priv->rps.max_freq |
4943 dev_priv->rps.efficient_freq |
4944 dev_priv->rps.rp1_freq |
4945 dev_priv->rps.min_freq) & 1,
4946 "Odd GPU freq values\n");
4948 dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
4950 /* Preserve min/max settings in case of re-init */
4951 if (dev_priv->rps.max_freq_softlimit == 0)
4952 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
4954 if (dev_priv->rps.min_freq_softlimit == 0)
4955 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
4957 mutex_unlock(&dev_priv->rps.hw_lock);
4960 static void valleyview_cleanup_gt_powersave(struct drm_device *dev)
4962 valleyview_cleanup_pctx(dev);
4965 static void cherryview_enable_rps(struct drm_device *dev)
4967 struct drm_i915_private *dev_priv = dev->dev_private;
4968 struct intel_engine_cs *ring;
4969 u32 gtfifodbg, val, rc6_mode = 0, pcbr;
4972 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4974 gtfifodbg = I915_READ(GTFIFODBG);
4976 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
4978 I915_WRITE(GTFIFODBG, gtfifodbg);
4981 cherryview_check_pctx(dev_priv);
4983 /* 1a & 1b: Get forcewake during program sequence. Although the driver
4984 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4985 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4987 /* Disable RC states. */
4988 I915_WRITE(GEN6_RC_CONTROL, 0);
4990 /* 2a: Program RC6 thresholds.*/
4991 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
4992 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
4993 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
4995 for_each_ring(ring, dev_priv, i)
4996 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4997 I915_WRITE(GEN6_RC_SLEEP, 0);
4999 /* TO threshold set to 1750 us ( 0x557 * 1.28 us) */
5000 I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
5002 /* allows RC6 residency counter to work */
5003 I915_WRITE(VLV_COUNTER_CONTROL,
5004 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
5005 VLV_MEDIA_RC6_COUNT_EN |
5006 VLV_RENDER_RC6_COUNT_EN));
5008 /* For now we assume BIOS is allocating and populating the PCBR */
5009 pcbr = I915_READ(VLV_PCBR);
5012 if ((intel_enable_rc6(dev) & INTEL_RC6_ENABLE) &&
5013 (pcbr >> VLV_PCBR_ADDR_SHIFT))
5014 rc6_mode = GEN7_RC_CTL_TO_MODE;
5016 I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
5018 /* 4 Program defaults and thresholds for RPS*/
5019 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
5020 I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
5021 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
5022 I915_WRITE(GEN6_RP_UP_EI, 66000);
5023 I915_WRITE(GEN6_RP_DOWN_EI, 350000);
5025 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5028 I915_WRITE(GEN6_RP_CONTROL,
5029 GEN6_RP_MEDIA_HW_NORMAL_MODE |
5030 GEN6_RP_MEDIA_IS_GFX |
5032 GEN6_RP_UP_BUSY_AVG |
5033 GEN6_RP_DOWN_IDLE_AVG);
5035 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5037 /* RPS code assumes GPLL is used */
5038 WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
5040 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", val & GPLLENABLE ? "yes" : "no");
5041 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
5043 dev_priv->rps.cur_freq = (val >> 8) & 0xff;
5044 DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
5045 intel_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
5046 dev_priv->rps.cur_freq);
5048 DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
5049 intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5050 dev_priv->rps.efficient_freq);
5052 valleyview_set_rps(dev_priv->dev, dev_priv->rps.efficient_freq);
5054 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5057 static void valleyview_enable_rps(struct drm_device *dev)
5059 struct drm_i915_private *dev_priv = dev->dev_private;
5060 struct intel_engine_cs *ring;
5061 u32 gtfifodbg, val, rc6_mode = 0;
5064 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5066 valleyview_check_pctx(dev_priv);
5068 if ((gtfifodbg = I915_READ(GTFIFODBG))) {
5069 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5071 I915_WRITE(GTFIFODBG, gtfifodbg);
5074 /* If VLV, Forcewake all wells, else re-direct to regular path */
5075 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5077 /* Disable RC states. */
5078 I915_WRITE(GEN6_RC_CONTROL, 0);
5080 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
5081 I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
5082 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
5083 I915_WRITE(GEN6_RP_UP_EI, 66000);
5084 I915_WRITE(GEN6_RP_DOWN_EI, 350000);
5086 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5088 I915_WRITE(GEN6_RP_CONTROL,
5089 GEN6_RP_MEDIA_TURBO |
5090 GEN6_RP_MEDIA_HW_NORMAL_MODE |
5091 GEN6_RP_MEDIA_IS_GFX |
5093 GEN6_RP_UP_BUSY_AVG |
5094 GEN6_RP_DOWN_IDLE_CONT);
5096 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
5097 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
5098 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
5100 for_each_ring(ring, dev_priv, i)
5101 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
5103 I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
5105 /* allows RC6 residency counter to work */
5106 I915_WRITE(VLV_COUNTER_CONTROL,
5107 _MASKED_BIT_ENABLE(VLV_MEDIA_RC0_COUNT_EN |
5108 VLV_RENDER_RC0_COUNT_EN |
5109 VLV_MEDIA_RC6_COUNT_EN |
5110 VLV_RENDER_RC6_COUNT_EN));
5112 if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
5113 rc6_mode = GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
5115 intel_print_rc6_info(dev, rc6_mode);
5117 I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
5119 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5121 /* RPS code assumes GPLL is used */
5122 WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
5124 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", val & GPLLENABLE ? "yes" : "no");
5125 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
5127 dev_priv->rps.cur_freq = (val >> 8) & 0xff;
5128 DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
5129 intel_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
5130 dev_priv->rps.cur_freq);
5132 DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
5133 intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5134 dev_priv->rps.efficient_freq);
5136 valleyview_set_rps(dev_priv->dev, dev_priv->rps.efficient_freq);
5138 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5141 static unsigned long intel_pxfreq(u32 vidfreq)
5144 int div = (vidfreq & 0x3f0000) >> 16;
5145 int post = (vidfreq & 0x3000) >> 12;
5146 int pre = (vidfreq & 0x7);
5151 freq = ((div * 133333) / ((1<<post) * pre));
5156 static const struct cparams {
5162 { 1, 1333, 301, 28664 },
5163 { 1, 1066, 294, 24460 },
5164 { 1, 800, 294, 25192 },
5165 { 0, 1333, 276, 27605 },
5166 { 0, 1066, 276, 27605 },
5167 { 0, 800, 231, 23784 },
5170 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
5172 u64 total_count, diff, ret;
5173 u32 count1, count2, count3, m = 0, c = 0;
5174 unsigned long now = jiffies_to_msecs(jiffies), diff1;
5177 assert_spin_locked(&mchdev_lock);
5179 diff1 = now - dev_priv->ips.last_time1;
5181 /* Prevent division-by-zero if we are asking too fast.
5182 * Also, we don't get interesting results if we are polling
5183 * faster than once in 10ms, so just return the saved value
5187 return dev_priv->ips.chipset_power;
5189 count1 = I915_READ(DMIEC);
5190 count2 = I915_READ(DDREC);
5191 count3 = I915_READ(CSIEC);
5193 total_count = count1 + count2 + count3;
5195 /* FIXME: handle per-counter overflow */
5196 if (total_count < dev_priv->ips.last_count1) {
5197 diff = ~0UL - dev_priv->ips.last_count1;
5198 diff += total_count;
5200 diff = total_count - dev_priv->ips.last_count1;
5203 for (i = 0; i < ARRAY_SIZE(cparams); i++) {
5204 if (cparams[i].i == dev_priv->ips.c_m &&
5205 cparams[i].t == dev_priv->ips.r_t) {
5212 diff = div_u64(diff, diff1);
5213 ret = ((m * diff) + c);
5214 ret = div_u64(ret, 10);
5216 dev_priv->ips.last_count1 = total_count;
5217 dev_priv->ips.last_time1 = now;
5219 dev_priv->ips.chipset_power = ret;
5224 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
5226 struct drm_device *dev = dev_priv->dev;
5229 if (INTEL_INFO(dev)->gen != 5)
5232 spin_lock_irq(&mchdev_lock);
5234 val = __i915_chipset_val(dev_priv);
5236 spin_unlock_irq(&mchdev_lock);
5241 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
5243 unsigned long m, x, b;
5246 tsfs = I915_READ(TSFS);
5248 m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
5249 x = I915_READ8(TR1);
5251 b = tsfs & TSFS_INTR_MASK;
5253 return ((m * x) / 127) - b;
5256 static int _pxvid_to_vd(u8 pxvid)
5261 if (pxvid >= 8 && pxvid < 31)
5264 return (pxvid + 2) * 125;
5267 static u32 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
5269 struct drm_device *dev = dev_priv->dev;
5270 const int vd = _pxvid_to_vd(pxvid);
5271 const int vm = vd - 1125;
5273 if (INTEL_INFO(dev)->is_mobile)
5274 return vm > 0 ? vm : 0;
5279 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
5281 u64 now, diff, diffms;
5284 assert_spin_locked(&mchdev_lock);
5286 now = ktime_get_raw_ns();
5287 diffms = now - dev_priv->ips.last_time2;
5288 do_div(diffms, NSEC_PER_MSEC);
5290 /* Don't divide by 0 */
5294 count = I915_READ(GFXEC);
5296 if (count < dev_priv->ips.last_count2) {
5297 diff = ~0UL - dev_priv->ips.last_count2;
5300 diff = count - dev_priv->ips.last_count2;
5303 dev_priv->ips.last_count2 = count;
5304 dev_priv->ips.last_time2 = now;
5306 /* More magic constants... */
5308 diff = div_u64(diff, diffms * 10);
5309 dev_priv->ips.gfx_power = diff;
5312 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
5314 struct drm_device *dev = dev_priv->dev;
5316 if (INTEL_INFO(dev)->gen != 5)
5319 spin_lock_irq(&mchdev_lock);
5321 __i915_update_gfx_val(dev_priv);
5323 spin_unlock_irq(&mchdev_lock);
5326 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
5328 unsigned long t, corr, state1, corr2, state2;
5331 assert_spin_locked(&mchdev_lock);
5333 pxvid = I915_READ(PXVFREQ_BASE + (dev_priv->rps.cur_freq * 4));
5334 pxvid = (pxvid >> 24) & 0x7f;
5335 ext_v = pvid_to_extvid(dev_priv, pxvid);
5339 t = i915_mch_val(dev_priv);
5341 /* Revel in the empirically derived constants */
5343 /* Correction factor in 1/100000 units */
5345 corr = ((t * 2349) + 135940);
5347 corr = ((t * 964) + 29317);
5349 corr = ((t * 301) + 1004);
5351 corr = corr * ((150142 * state1) / 10000 - 78642);
5353 corr2 = (corr * dev_priv->ips.corr);
5355 state2 = (corr2 * state1) / 10000;
5356 state2 /= 100; /* convert to mW */
5358 __i915_update_gfx_val(dev_priv);
5360 return dev_priv->ips.gfx_power + state2;
5363 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
5365 struct drm_device *dev = dev_priv->dev;
5368 if (INTEL_INFO(dev)->gen != 5)
5371 spin_lock_irq(&mchdev_lock);
5373 val = __i915_gfx_val(dev_priv);
5375 spin_unlock_irq(&mchdev_lock);
5381 * i915_read_mch_val - return value for IPS use
5383 * Calculate and return a value for the IPS driver to use when deciding whether
5384 * we have thermal and power headroom to increase CPU or GPU power budget.
5386 unsigned long i915_read_mch_val(void)
5388 struct drm_i915_private *dev_priv;
5389 unsigned long chipset_val, graphics_val, ret = 0;
5391 spin_lock_irq(&mchdev_lock);
5394 dev_priv = i915_mch_dev;
5396 chipset_val = __i915_chipset_val(dev_priv);
5397 graphics_val = __i915_gfx_val(dev_priv);
5399 ret = chipset_val + graphics_val;
5402 spin_unlock_irq(&mchdev_lock);
5406 EXPORT_SYMBOL_GPL(i915_read_mch_val);
5409 * i915_gpu_raise - raise GPU frequency limit
5411 * Raise the limit; IPS indicates we have thermal headroom.
5413 bool i915_gpu_raise(void)
5415 struct drm_i915_private *dev_priv;
5418 spin_lock_irq(&mchdev_lock);
5419 if (!i915_mch_dev) {
5423 dev_priv = i915_mch_dev;
5425 if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
5426 dev_priv->ips.max_delay--;
5429 spin_unlock_irq(&mchdev_lock);
5433 EXPORT_SYMBOL_GPL(i915_gpu_raise);
5436 * i915_gpu_lower - lower GPU frequency limit
5438 * IPS indicates we're close to a thermal limit, so throttle back the GPU
5439 * frequency maximum.
5441 bool i915_gpu_lower(void)
5443 struct drm_i915_private *dev_priv;
5446 spin_lock_irq(&mchdev_lock);
5447 if (!i915_mch_dev) {
5451 dev_priv = i915_mch_dev;
5453 if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
5454 dev_priv->ips.max_delay++;
5457 spin_unlock_irq(&mchdev_lock);
5461 EXPORT_SYMBOL_GPL(i915_gpu_lower);
5464 * i915_gpu_busy - indicate GPU business to IPS
5466 * Tell the IPS driver whether or not the GPU is busy.
5468 bool i915_gpu_busy(void)
5470 struct drm_i915_private *dev_priv;
5471 struct intel_engine_cs *ring;
5475 spin_lock_irq(&mchdev_lock);
5478 dev_priv = i915_mch_dev;
5480 for_each_ring(ring, dev_priv, i)
5481 ret |= !list_empty(&ring->request_list);
5484 spin_unlock_irq(&mchdev_lock);
5488 EXPORT_SYMBOL_GPL(i915_gpu_busy);
5491 * i915_gpu_turbo_disable - disable graphics turbo
5493 * Disable graphics turbo by resetting the max frequency and setting the
5494 * current frequency to the default.
5496 bool i915_gpu_turbo_disable(void)
5498 struct drm_i915_private *dev_priv;
5501 spin_lock_irq(&mchdev_lock);
5502 if (!i915_mch_dev) {
5506 dev_priv = i915_mch_dev;
5508 dev_priv->ips.max_delay = dev_priv->ips.fstart;
5510 if (!ironlake_set_drps(dev_priv->dev, dev_priv->ips.fstart))
5514 spin_unlock_irq(&mchdev_lock);
5518 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
5521 * Tells the intel_ips driver that the i915 driver is now loaded, if
5522 * IPS got loaded first.
5524 * This awkward dance is so that neither module has to depend on the
5525 * other in order for IPS to do the appropriate communication of
5526 * GPU turbo limits to i915.
5529 ips_ping_for_i915_load(void)
5533 link = symbol_get(ips_link_to_i915_driver);
5536 symbol_put(ips_link_to_i915_driver);
5540 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
5542 /* We only register the i915 ips part with intel-ips once everything is
5543 * set up, to avoid intel-ips sneaking in and reading bogus values. */
5544 spin_lock_irq(&mchdev_lock);
5545 i915_mch_dev = dev_priv;
5546 spin_unlock_irq(&mchdev_lock);
5548 ips_ping_for_i915_load();
5551 void intel_gpu_ips_teardown(void)
5553 spin_lock_irq(&mchdev_lock);
5554 i915_mch_dev = NULL;
5555 spin_unlock_irq(&mchdev_lock);
5558 static void intel_init_emon(struct drm_device *dev)
5560 struct drm_i915_private *dev_priv = dev->dev_private;
5565 /* Disable to program */
5569 /* Program energy weights for various events */
5570 I915_WRITE(SDEW, 0x15040d00);
5571 I915_WRITE(CSIEW0, 0x007f0000);
5572 I915_WRITE(CSIEW1, 0x1e220004);
5573 I915_WRITE(CSIEW2, 0x04000004);
5575 for (i = 0; i < 5; i++)
5576 I915_WRITE(PEW + (i * 4), 0);
5577 for (i = 0; i < 3; i++)
5578 I915_WRITE(DEW + (i * 4), 0);
5580 /* Program P-state weights to account for frequency power adjustment */
5581 for (i = 0; i < 16; i++) {
5582 u32 pxvidfreq = I915_READ(PXVFREQ_BASE + (i * 4));
5583 unsigned long freq = intel_pxfreq(pxvidfreq);
5584 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
5589 val *= (freq / 1000);
5591 val /= (127*127*900);
5593 DRM_ERROR("bad pxval: %ld\n", val);
5596 /* Render standby states get 0 weight */
5600 for (i = 0; i < 4; i++) {
5601 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
5602 (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
5603 I915_WRITE(PXW + (i * 4), val);
5606 /* Adjust magic regs to magic values (more experimental results) */
5607 I915_WRITE(OGW0, 0);
5608 I915_WRITE(OGW1, 0);
5609 I915_WRITE(EG0, 0x00007f00);
5610 I915_WRITE(EG1, 0x0000000e);
5611 I915_WRITE(EG2, 0x000e0000);
5612 I915_WRITE(EG3, 0x68000300);
5613 I915_WRITE(EG4, 0x42000000);
5614 I915_WRITE(EG5, 0x00140031);
5618 for (i = 0; i < 8; i++)
5619 I915_WRITE(PXWL + (i * 4), 0);
5621 /* Enable PMON + select events */
5622 I915_WRITE(ECR, 0x80000019);
5624 lcfuse = I915_READ(LCFUSE02);
5626 dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
5629 void intel_init_gt_powersave(struct drm_device *dev)
5631 i915.enable_rc6 = sanitize_rc6_option(dev, i915.enable_rc6);
5633 if (IS_CHERRYVIEW(dev))
5634 cherryview_init_gt_powersave(dev);
5635 else if (IS_VALLEYVIEW(dev))
5636 valleyview_init_gt_powersave(dev);
5639 void intel_cleanup_gt_powersave(struct drm_device *dev)
5641 if (IS_CHERRYVIEW(dev))
5643 else if (IS_VALLEYVIEW(dev))
5644 valleyview_cleanup_gt_powersave(dev);
5647 static void gen6_suspend_rps(struct drm_device *dev)
5649 struct drm_i915_private *dev_priv = dev->dev_private;
5651 flush_delayed_work(&dev_priv->rps.delayed_resume_work);
5653 gen6_disable_rps_interrupts(dev);
5657 * intel_suspend_gt_powersave - suspend PM work and helper threads
5660 * We don't want to disable RC6 or other features here, we just want
5661 * to make sure any work we've queued has finished and won't bother
5662 * us while we're suspended.
5664 void intel_suspend_gt_powersave(struct drm_device *dev)
5666 struct drm_i915_private *dev_priv = dev->dev_private;
5668 if (INTEL_INFO(dev)->gen < 6)
5671 gen6_suspend_rps(dev);
5673 /* Force GPU to min freq during suspend */
5674 gen6_rps_idle(dev_priv);
5677 void intel_disable_gt_powersave(struct drm_device *dev)
5679 struct drm_i915_private *dev_priv = dev->dev_private;
5681 if (IS_IRONLAKE_M(dev)) {
5682 ironlake_disable_drps(dev);
5683 } else if (INTEL_INFO(dev)->gen >= 6) {
5684 intel_suspend_gt_powersave(dev);
5686 mutex_lock(&dev_priv->rps.hw_lock);
5687 if (INTEL_INFO(dev)->gen >= 9)
5688 gen9_disable_rps(dev);
5689 else if (IS_CHERRYVIEW(dev))
5690 cherryview_disable_rps(dev);
5691 else if (IS_VALLEYVIEW(dev))
5692 valleyview_disable_rps(dev);
5694 gen6_disable_rps(dev);
5696 dev_priv->rps.enabled = false;
5697 mutex_unlock(&dev_priv->rps.hw_lock);
5701 static void intel_gen6_powersave_work(struct work_struct *work)
5703 struct drm_i915_private *dev_priv =
5704 container_of(work, struct drm_i915_private,
5705 rps.delayed_resume_work.work);
5706 struct drm_device *dev = dev_priv->dev;
5708 mutex_lock(&dev_priv->rps.hw_lock);
5710 gen6_reset_rps_interrupts(dev);
5712 if (IS_CHERRYVIEW(dev)) {
5713 cherryview_enable_rps(dev);
5714 } else if (IS_VALLEYVIEW(dev)) {
5715 valleyview_enable_rps(dev);
5716 } else if (INTEL_INFO(dev)->gen >= 9) {
5717 gen9_enable_rc6(dev);
5718 gen9_enable_rps(dev);
5719 __gen6_update_ring_freq(dev);
5720 } else if (IS_BROADWELL(dev)) {
5721 gen8_enable_rps(dev);
5722 __gen6_update_ring_freq(dev);
5724 gen6_enable_rps(dev);
5725 __gen6_update_ring_freq(dev);
5728 WARN_ON(dev_priv->rps.max_freq < dev_priv->rps.min_freq);
5729 WARN_ON(dev_priv->rps.idle_freq > dev_priv->rps.max_freq);
5731 WARN_ON(dev_priv->rps.efficient_freq < dev_priv->rps.min_freq);
5732 WARN_ON(dev_priv->rps.efficient_freq > dev_priv->rps.max_freq);
5734 dev_priv->rps.enabled = true;
5736 gen6_enable_rps_interrupts(dev);
5738 mutex_unlock(&dev_priv->rps.hw_lock);
5740 intel_runtime_pm_put(dev_priv);
5743 void intel_enable_gt_powersave(struct drm_device *dev)
5745 struct drm_i915_private *dev_priv = dev->dev_private;
5747 /* Powersaving is controlled by the host when inside a VM */
5748 if (intel_vgpu_active(dev))
5751 if (IS_IRONLAKE_M(dev)) {
5752 mutex_lock(&dev->struct_mutex);
5753 ironlake_enable_drps(dev);
5754 intel_init_emon(dev);
5755 mutex_unlock(&dev->struct_mutex);
5756 } else if (INTEL_INFO(dev)->gen >= 6) {
5758 * PCU communication is slow and this doesn't need to be
5759 * done at any specific time, so do this out of our fast path
5760 * to make resume and init faster.
5762 * We depend on the HW RC6 power context save/restore
5763 * mechanism when entering D3 through runtime PM suspend. So
5764 * disable RPM until RPS/RC6 is properly setup. We can only
5765 * get here via the driver load/system resume/runtime resume
5766 * paths, so the _noresume version is enough (and in case of
5767 * runtime resume it's necessary).
5769 if (schedule_delayed_work(&dev_priv->rps.delayed_resume_work,
5770 round_jiffies_up_relative(HZ)))
5771 intel_runtime_pm_get_noresume(dev_priv);
5775 void intel_reset_gt_powersave(struct drm_device *dev)
5777 struct drm_i915_private *dev_priv = dev->dev_private;
5779 if (INTEL_INFO(dev)->gen < 6)
5782 gen6_suspend_rps(dev);
5783 dev_priv->rps.enabled = false;
5786 static void ibx_init_clock_gating(struct drm_device *dev)
5788 struct drm_i915_private *dev_priv = dev->dev_private;
5791 * On Ibex Peak and Cougar Point, we need to disable clock
5792 * gating for the panel power sequencer or it will fail to
5793 * start up when no ports are active.
5795 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
5798 static void g4x_disable_trickle_feed(struct drm_device *dev)
5800 struct drm_i915_private *dev_priv = dev->dev_private;
5803 for_each_pipe(dev_priv, pipe) {
5804 I915_WRITE(DSPCNTR(pipe),
5805 I915_READ(DSPCNTR(pipe)) |
5806 DISPPLANE_TRICKLE_FEED_DISABLE);
5807 intel_flush_primary_plane(dev_priv, pipe);
5811 static void ilk_init_lp_watermarks(struct drm_device *dev)
5813 struct drm_i915_private *dev_priv = dev->dev_private;
5815 I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
5816 I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
5817 I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
5820 * Don't touch WM1S_LP_EN here.
5821 * Doing so could cause underruns.
5825 static void ironlake_init_clock_gating(struct drm_device *dev)
5827 struct drm_i915_private *dev_priv = dev->dev_private;
5828 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
5832 * WaFbcDisableDpfcClockGating:ilk
5834 dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
5835 ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
5836 ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
5838 I915_WRITE(PCH_3DCGDIS0,
5839 MARIUNIT_CLOCK_GATE_DISABLE |
5840 SVSMUNIT_CLOCK_GATE_DISABLE);
5841 I915_WRITE(PCH_3DCGDIS1,
5842 VFMUNIT_CLOCK_GATE_DISABLE);
5845 * According to the spec the following bits should be set in
5846 * order to enable memory self-refresh
5847 * The bit 22/21 of 0x42004
5848 * The bit 5 of 0x42020
5849 * The bit 15 of 0x45000
5851 I915_WRITE(ILK_DISPLAY_CHICKEN2,
5852 (I915_READ(ILK_DISPLAY_CHICKEN2) |
5853 ILK_DPARB_GATE | ILK_VSDPFD_FULL));
5854 dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
5855 I915_WRITE(DISP_ARB_CTL,
5856 (I915_READ(DISP_ARB_CTL) |
5859 ilk_init_lp_watermarks(dev);
5862 * Based on the document from hardware guys the following bits
5863 * should be set unconditionally in order to enable FBC.
5864 * The bit 22 of 0x42000
5865 * The bit 22 of 0x42004
5866 * The bit 7,8,9 of 0x42020.
5868 if (IS_IRONLAKE_M(dev)) {
5869 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
5870 I915_WRITE(ILK_DISPLAY_CHICKEN1,
5871 I915_READ(ILK_DISPLAY_CHICKEN1) |
5873 I915_WRITE(ILK_DISPLAY_CHICKEN2,
5874 I915_READ(ILK_DISPLAY_CHICKEN2) |
5878 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
5880 I915_WRITE(ILK_DISPLAY_CHICKEN2,
5881 I915_READ(ILK_DISPLAY_CHICKEN2) |
5882 ILK_ELPIN_409_SELECT);
5883 I915_WRITE(_3D_CHICKEN2,
5884 _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
5885 _3D_CHICKEN2_WM_READ_PIPELINED);
5887 /* WaDisableRenderCachePipelinedFlush:ilk */
5888 I915_WRITE(CACHE_MODE_0,
5889 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
5891 /* WaDisable_RenderCache_OperationalFlush:ilk */
5892 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5894 g4x_disable_trickle_feed(dev);
5896 ibx_init_clock_gating(dev);
5899 static void cpt_init_clock_gating(struct drm_device *dev)
5901 struct drm_i915_private *dev_priv = dev->dev_private;
5906 * On Ibex Peak and Cougar Point, we need to disable clock
5907 * gating for the panel power sequencer or it will fail to
5908 * start up when no ports are active.
5910 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
5911 PCH_DPLUNIT_CLOCK_GATE_DISABLE |
5912 PCH_CPUNIT_CLOCK_GATE_DISABLE);
5913 I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
5914 DPLS_EDP_PPS_FIX_DIS);
5915 /* The below fixes the weird display corruption, a few pixels shifted
5916 * downward, on (only) LVDS of some HP laptops with IVY.
5918 for_each_pipe(dev_priv, pipe) {
5919 val = I915_READ(TRANS_CHICKEN2(pipe));
5920 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
5921 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
5922 if (dev_priv->vbt.fdi_rx_polarity_inverted)
5923 val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
5924 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
5925 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
5926 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
5927 I915_WRITE(TRANS_CHICKEN2(pipe), val);
5929 /* WADP0ClockGatingDisable */
5930 for_each_pipe(dev_priv, pipe) {
5931 I915_WRITE(TRANS_CHICKEN1(pipe),
5932 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
5936 static void gen6_check_mch_setup(struct drm_device *dev)
5938 struct drm_i915_private *dev_priv = dev->dev_private;
5941 tmp = I915_READ(MCH_SSKPD);
5942 if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL)
5943 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
5947 static void gen6_init_clock_gating(struct drm_device *dev)
5949 struct drm_i915_private *dev_priv = dev->dev_private;
5950 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
5952 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
5954 I915_WRITE(ILK_DISPLAY_CHICKEN2,
5955 I915_READ(ILK_DISPLAY_CHICKEN2) |
5956 ILK_ELPIN_409_SELECT);
5958 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
5959 I915_WRITE(_3D_CHICKEN,
5960 _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
5962 /* WaDisable_RenderCache_OperationalFlush:snb */
5963 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5966 * BSpec recoomends 8x4 when MSAA is used,
5967 * however in practice 16x4 seems fastest.
5969 * Note that PS/WM thread counts depend on the WIZ hashing
5970 * disable bit, which we don't touch here, but it's good
5971 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
5973 I915_WRITE(GEN6_GT_MODE,
5974 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
5976 ilk_init_lp_watermarks(dev);
5978 I915_WRITE(CACHE_MODE_0,
5979 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
5981 I915_WRITE(GEN6_UCGCTL1,
5982 I915_READ(GEN6_UCGCTL1) |
5983 GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
5984 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
5986 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
5987 * gating disable must be set. Failure to set it results in
5988 * flickering pixels due to Z write ordering failures after
5989 * some amount of runtime in the Mesa "fire" demo, and Unigine
5990 * Sanctuary and Tropics, and apparently anything else with
5991 * alpha test or pixel discard.
5993 * According to the spec, bit 11 (RCCUNIT) must also be set,
5994 * but we didn't debug actual testcases to find it out.
5996 * WaDisableRCCUnitClockGating:snb
5997 * WaDisableRCPBUnitClockGating:snb
5999 I915_WRITE(GEN6_UCGCTL2,
6000 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
6001 GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
6003 /* WaStripsFansDisableFastClipPerformanceFix:snb */
6004 I915_WRITE(_3D_CHICKEN3,
6005 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
6009 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
6010 * 3DSTATE_SF number of SF output attributes is more than 16."
6012 I915_WRITE(_3D_CHICKEN3,
6013 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
6016 * According to the spec the following bits should be
6017 * set in order to enable memory self-refresh and fbc:
6018 * The bit21 and bit22 of 0x42000
6019 * The bit21 and bit22 of 0x42004
6020 * The bit5 and bit7 of 0x42020
6021 * The bit14 of 0x70180
6022 * The bit14 of 0x71180
6024 * WaFbcAsynchFlipDisableFbcQueue:snb
6026 I915_WRITE(ILK_DISPLAY_CHICKEN1,
6027 I915_READ(ILK_DISPLAY_CHICKEN1) |
6028 ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
6029 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6030 I915_READ(ILK_DISPLAY_CHICKEN2) |
6031 ILK_DPARB_GATE | ILK_VSDPFD_FULL);
6032 I915_WRITE(ILK_DSPCLK_GATE_D,
6033 I915_READ(ILK_DSPCLK_GATE_D) |
6034 ILK_DPARBUNIT_CLOCK_GATE_ENABLE |
6035 ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
6037 g4x_disable_trickle_feed(dev);
6039 cpt_init_clock_gating(dev);
6041 gen6_check_mch_setup(dev);
6044 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
6046 uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
6049 * WaVSThreadDispatchOverride:ivb,vlv
6051 * This actually overrides the dispatch
6052 * mode for all thread types.
6054 reg &= ~GEN7_FF_SCHED_MASK;
6055 reg |= GEN7_FF_TS_SCHED_HW;
6056 reg |= GEN7_FF_VS_SCHED_HW;
6057 reg |= GEN7_FF_DS_SCHED_HW;
6059 I915_WRITE(GEN7_FF_THREAD_MODE, reg);
6062 static void lpt_init_clock_gating(struct drm_device *dev)
6064 struct drm_i915_private *dev_priv = dev->dev_private;
6067 * TODO: this bit should only be enabled when really needed, then
6068 * disabled when not needed anymore in order to save power.
6070 if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE)
6071 I915_WRITE(SOUTH_DSPCLK_GATE_D,
6072 I915_READ(SOUTH_DSPCLK_GATE_D) |
6073 PCH_LP_PARTITION_LEVEL_DISABLE);
6075 /* WADPOClockGatingDisable:hsw */
6076 I915_WRITE(_TRANSA_CHICKEN1,
6077 I915_READ(_TRANSA_CHICKEN1) |
6078 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
6081 static void lpt_suspend_hw(struct drm_device *dev)
6083 struct drm_i915_private *dev_priv = dev->dev_private;
6085 if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE) {
6086 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
6088 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
6089 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
6093 static void broadwell_init_clock_gating(struct drm_device *dev)
6095 struct drm_i915_private *dev_priv = dev->dev_private;
6098 I915_WRITE(WM3_LP_ILK, 0);
6099 I915_WRITE(WM2_LP_ILK, 0);
6100 I915_WRITE(WM1_LP_ILK, 0);
6102 /* WaSwitchSolVfFArbitrationPriority:bdw */
6103 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
6105 /* WaPsrDPAMaskVBlankInSRD:bdw */
6106 I915_WRITE(CHICKEN_PAR1_1,
6107 I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
6109 /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
6110 for_each_pipe(dev_priv, pipe) {
6111 I915_WRITE(CHICKEN_PIPESL_1(pipe),
6112 I915_READ(CHICKEN_PIPESL_1(pipe)) |
6113 BDW_DPRS_MASK_VBLANK_SRD);
6116 /* WaVSRefCountFullforceMissDisable:bdw */
6117 /* WaDSRefCountFullforceMissDisable:bdw */
6118 I915_WRITE(GEN7_FF_THREAD_MODE,
6119 I915_READ(GEN7_FF_THREAD_MODE) &
6120 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
6122 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
6123 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
6125 /* WaDisableSDEUnitClockGating:bdw */
6126 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
6127 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
6129 lpt_init_clock_gating(dev);
6132 static void haswell_init_clock_gating(struct drm_device *dev)
6134 struct drm_i915_private *dev_priv = dev->dev_private;
6136 ilk_init_lp_watermarks(dev);
6138 /* L3 caching of data atomics doesn't work -- disable it. */
6139 I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
6140 I915_WRITE(HSW_ROW_CHICKEN3,
6141 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
6143 /* This is required by WaCatErrorRejectionIssue:hsw */
6144 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
6145 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
6146 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
6148 /* WaVSRefCountFullforceMissDisable:hsw */
6149 I915_WRITE(GEN7_FF_THREAD_MODE,
6150 I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
6152 /* WaDisable_RenderCache_OperationalFlush:hsw */
6153 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6155 /* enable HiZ Raw Stall Optimization */
6156 I915_WRITE(CACHE_MODE_0_GEN7,
6157 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
6159 /* WaDisable4x2SubspanOptimization:hsw */
6160 I915_WRITE(CACHE_MODE_1,
6161 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
6164 * BSpec recommends 8x4 when MSAA is used,
6165 * however in practice 16x4 seems fastest.
6167 * Note that PS/WM thread counts depend on the WIZ hashing
6168 * disable bit, which we don't touch here, but it's good
6169 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6171 I915_WRITE(GEN7_GT_MODE,
6172 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6174 /* WaSampleCChickenBitEnable:hsw */
6175 I915_WRITE(HALF_SLICE_CHICKEN3,
6176 _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE));
6178 /* WaSwitchSolVfFArbitrationPriority:hsw */
6179 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
6181 /* WaRsPkgCStateDisplayPMReq:hsw */
6182 I915_WRITE(CHICKEN_PAR1_1,
6183 I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
6185 lpt_init_clock_gating(dev);
6188 static void ivybridge_init_clock_gating(struct drm_device *dev)
6190 struct drm_i915_private *dev_priv = dev->dev_private;
6193 ilk_init_lp_watermarks(dev);
6195 I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
6197 /* WaDisableEarlyCull:ivb */
6198 I915_WRITE(_3D_CHICKEN3,
6199 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
6201 /* WaDisableBackToBackFlipFix:ivb */
6202 I915_WRITE(IVB_CHICKEN3,
6203 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
6204 CHICKEN3_DGMG_DONE_FIX_DISABLE);
6206 /* WaDisablePSDDualDispatchEnable:ivb */
6207 if (IS_IVB_GT1(dev))
6208 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
6209 _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
6211 /* WaDisable_RenderCache_OperationalFlush:ivb */
6212 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6214 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
6215 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
6216 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
6218 /* WaApplyL3ControlAndL3ChickenMode:ivb */
6219 I915_WRITE(GEN7_L3CNTLREG1,
6220 GEN7_WA_FOR_GEN7_L3_CONTROL);
6221 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
6222 GEN7_WA_L3_CHICKEN_MODE);
6223 if (IS_IVB_GT1(dev))
6224 I915_WRITE(GEN7_ROW_CHICKEN2,
6225 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6227 /* must write both registers */
6228 I915_WRITE(GEN7_ROW_CHICKEN2,
6229 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6230 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
6231 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6234 /* WaForceL3Serialization:ivb */
6235 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
6236 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
6239 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6240 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
6242 I915_WRITE(GEN6_UCGCTL2,
6243 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
6245 /* This is required by WaCatErrorRejectionIssue:ivb */
6246 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
6247 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
6248 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
6250 g4x_disable_trickle_feed(dev);
6252 gen7_setup_fixed_func_scheduler(dev_priv);
6254 if (0) { /* causes HiZ corruption on ivb:gt1 */
6255 /* enable HiZ Raw Stall Optimization */
6256 I915_WRITE(CACHE_MODE_0_GEN7,
6257 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
6260 /* WaDisable4x2SubspanOptimization:ivb */
6261 I915_WRITE(CACHE_MODE_1,
6262 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
6265 * BSpec recommends 8x4 when MSAA is used,
6266 * however in practice 16x4 seems fastest.
6268 * Note that PS/WM thread counts depend on the WIZ hashing
6269 * disable bit, which we don't touch here, but it's good
6270 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6272 I915_WRITE(GEN7_GT_MODE,
6273 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6275 snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
6276 snpcr &= ~GEN6_MBC_SNPCR_MASK;
6277 snpcr |= GEN6_MBC_SNPCR_MED;
6278 I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
6280 if (!HAS_PCH_NOP(dev))
6281 cpt_init_clock_gating(dev);
6283 gen6_check_mch_setup(dev);
6286 static void vlv_init_display_clock_gating(struct drm_i915_private *dev_priv)
6288 I915_WRITE(DSPCLK_GATE_D, VRHUNIT_CLOCK_GATE_DISABLE);
6291 * Disable trickle feed and enable pnd deadline calculation
6293 I915_WRITE(MI_ARB_VLV, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE);
6294 I915_WRITE(CBR1_VLV, 0);
6297 static void valleyview_init_clock_gating(struct drm_device *dev)
6299 struct drm_i915_private *dev_priv = dev->dev_private;
6301 vlv_init_display_clock_gating(dev_priv);
6303 /* WaDisableEarlyCull:vlv */
6304 I915_WRITE(_3D_CHICKEN3,
6305 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
6307 /* WaDisableBackToBackFlipFix:vlv */
6308 I915_WRITE(IVB_CHICKEN3,
6309 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
6310 CHICKEN3_DGMG_DONE_FIX_DISABLE);
6312 /* WaPsdDispatchEnable:vlv */
6313 /* WaDisablePSDDualDispatchEnable:vlv */
6314 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
6315 _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
6316 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
6318 /* WaDisable_RenderCache_OperationalFlush:vlv */
6319 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6321 /* WaForceL3Serialization:vlv */
6322 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
6323 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
6325 /* WaDisableDopClockGating:vlv */
6326 I915_WRITE(GEN7_ROW_CHICKEN2,
6327 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6329 /* This is required by WaCatErrorRejectionIssue:vlv */
6330 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
6331 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
6332 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
6334 gen7_setup_fixed_func_scheduler(dev_priv);
6337 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6338 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
6340 I915_WRITE(GEN6_UCGCTL2,
6341 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
6343 /* WaDisableL3Bank2xClockGate:vlv
6344 * Disabling L3 clock gating- MMIO 940c[25] = 1
6345 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
6346 I915_WRITE(GEN7_UCGCTL4,
6347 I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
6350 * BSpec says this must be set, even though
6351 * WaDisable4x2SubspanOptimization isn't listed for VLV.
6353 I915_WRITE(CACHE_MODE_1,
6354 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
6357 * BSpec recommends 8x4 when MSAA is used,
6358 * however in practice 16x4 seems fastest.
6360 * Note that PS/WM thread counts depend on the WIZ hashing
6361 * disable bit, which we don't touch here, but it's good
6362 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6364 I915_WRITE(GEN7_GT_MODE,
6365 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6368 * WaIncreaseL3CreditsForVLVB0:vlv
6369 * This is the hardware default actually.
6371 I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
6374 * WaDisableVLVClockGating_VBIIssue:vlv
6375 * Disable clock gating on th GCFG unit to prevent a delay
6376 * in the reporting of vblank events.
6378 I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
6381 static void cherryview_init_clock_gating(struct drm_device *dev)
6383 struct drm_i915_private *dev_priv = dev->dev_private;
6385 vlv_init_display_clock_gating(dev_priv);
6387 /* WaVSRefCountFullforceMissDisable:chv */
6388 /* WaDSRefCountFullforceMissDisable:chv */
6389 I915_WRITE(GEN7_FF_THREAD_MODE,
6390 I915_READ(GEN7_FF_THREAD_MODE) &
6391 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
6393 /* WaDisableSemaphoreAndSyncFlipWait:chv */
6394 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
6395 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
6397 /* WaDisableCSUnitClockGating:chv */
6398 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
6399 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
6401 /* WaDisableSDEUnitClockGating:chv */
6402 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
6403 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
6406 static void g4x_init_clock_gating(struct drm_device *dev)
6408 struct drm_i915_private *dev_priv = dev->dev_private;
6409 uint32_t dspclk_gate;
6411 I915_WRITE(RENCLK_GATE_D1, 0);
6412 I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
6413 GS_UNIT_CLOCK_GATE_DISABLE |
6414 CL_UNIT_CLOCK_GATE_DISABLE);
6415 I915_WRITE(RAMCLK_GATE_D, 0);
6416 dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
6417 OVRUNIT_CLOCK_GATE_DISABLE |
6418 OVCUNIT_CLOCK_GATE_DISABLE;
6420 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
6421 I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
6423 /* WaDisableRenderCachePipelinedFlush */
6424 I915_WRITE(CACHE_MODE_0,
6425 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
6427 /* WaDisable_RenderCache_OperationalFlush:g4x */
6428 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6430 g4x_disable_trickle_feed(dev);
6433 static void crestline_init_clock_gating(struct drm_device *dev)
6435 struct drm_i915_private *dev_priv = dev->dev_private;
6437 I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
6438 I915_WRITE(RENCLK_GATE_D2, 0);
6439 I915_WRITE(DSPCLK_GATE_D, 0);
6440 I915_WRITE(RAMCLK_GATE_D, 0);
6441 I915_WRITE16(DEUC, 0);
6442 I915_WRITE(MI_ARB_STATE,
6443 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
6445 /* WaDisable_RenderCache_OperationalFlush:gen4 */
6446 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6449 static void broadwater_init_clock_gating(struct drm_device *dev)
6451 struct drm_i915_private *dev_priv = dev->dev_private;
6453 I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
6454 I965_RCC_CLOCK_GATE_DISABLE |
6455 I965_RCPB_CLOCK_GATE_DISABLE |
6456 I965_ISC_CLOCK_GATE_DISABLE |
6457 I965_FBC_CLOCK_GATE_DISABLE);
6458 I915_WRITE(RENCLK_GATE_D2, 0);
6459 I915_WRITE(MI_ARB_STATE,
6460 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
6462 /* WaDisable_RenderCache_OperationalFlush:gen4 */
6463 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6466 static void gen3_init_clock_gating(struct drm_device *dev)
6468 struct drm_i915_private *dev_priv = dev->dev_private;
6469 u32 dstate = I915_READ(D_STATE);
6471 dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
6472 DSTATE_DOT_CLOCK_GATING;
6473 I915_WRITE(D_STATE, dstate);
6475 if (IS_PINEVIEW(dev))
6476 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
6478 /* IIR "flip pending" means done if this bit is set */
6479 I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
6481 /* interrupts should cause a wake up from C3 */
6482 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
6484 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
6485 I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
6487 I915_WRITE(MI_ARB_STATE,
6488 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
6491 static void i85x_init_clock_gating(struct drm_device *dev)
6493 struct drm_i915_private *dev_priv = dev->dev_private;
6495 I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
6497 /* interrupts should cause a wake up from C3 */
6498 I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
6499 _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
6501 I915_WRITE(MEM_MODE,
6502 _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE));
6505 static void i830_init_clock_gating(struct drm_device *dev)
6507 struct drm_i915_private *dev_priv = dev->dev_private;
6509 I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
6511 I915_WRITE(MEM_MODE,
6512 _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE) |
6513 _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE));
6516 void intel_init_clock_gating(struct drm_device *dev)
6518 struct drm_i915_private *dev_priv = dev->dev_private;
6520 if (dev_priv->display.init_clock_gating)
6521 dev_priv->display.init_clock_gating(dev);
6524 void intel_suspend_hw(struct drm_device *dev)
6526 if (HAS_PCH_LPT(dev))
6527 lpt_suspend_hw(dev);
6530 /* Set up chip specific power management-related functions */
6531 void intel_init_pm(struct drm_device *dev)
6533 struct drm_i915_private *dev_priv = dev->dev_private;
6535 intel_fbc_init(dev_priv);
6538 if (IS_PINEVIEW(dev))
6539 i915_pineview_get_mem_freq(dev);
6540 else if (IS_GEN5(dev))
6541 i915_ironlake_get_mem_freq(dev);
6543 /* For FIFO watermark updates */
6544 if (INTEL_INFO(dev)->gen >= 9) {
6545 skl_setup_wm_latency(dev);
6547 dev_priv->display.init_clock_gating = skl_init_clock_gating;
6548 dev_priv->display.update_wm = skl_update_wm;
6549 dev_priv->display.update_sprite_wm = skl_update_sprite_wm;
6550 } else if (HAS_PCH_SPLIT(dev)) {
6551 ilk_setup_wm_latency(dev);
6553 if ((IS_GEN5(dev) && dev_priv->wm.pri_latency[1] &&
6554 dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
6555 (!IS_GEN5(dev) && dev_priv->wm.pri_latency[0] &&
6556 dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
6557 dev_priv->display.update_wm = ilk_update_wm;
6558 dev_priv->display.update_sprite_wm = ilk_update_sprite_wm;
6560 DRM_DEBUG_KMS("Failed to read display plane latency. "
6565 dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
6566 else if (IS_GEN6(dev))
6567 dev_priv->display.init_clock_gating = gen6_init_clock_gating;
6568 else if (IS_IVYBRIDGE(dev))
6569 dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
6570 else if (IS_HASWELL(dev))
6571 dev_priv->display.init_clock_gating = haswell_init_clock_gating;
6572 else if (INTEL_INFO(dev)->gen == 8)
6573 dev_priv->display.init_clock_gating = broadwell_init_clock_gating;
6574 } else if (IS_CHERRYVIEW(dev)) {
6575 dev_priv->display.update_wm = valleyview_update_wm;
6576 dev_priv->display.update_sprite_wm = valleyview_update_sprite_wm;
6577 dev_priv->display.init_clock_gating =
6578 cherryview_init_clock_gating;
6579 } else if (IS_VALLEYVIEW(dev)) {
6580 dev_priv->display.update_wm = valleyview_update_wm;
6581 dev_priv->display.update_sprite_wm = valleyview_update_sprite_wm;
6582 dev_priv->display.init_clock_gating =
6583 valleyview_init_clock_gating;
6584 } else if (IS_PINEVIEW(dev)) {
6585 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
6588 dev_priv->mem_freq)) {
6589 DRM_INFO("failed to find known CxSR latency "
6590 "(found ddr%s fsb freq %d, mem freq %d), "
6592 (dev_priv->is_ddr3 == 1) ? "3" : "2",
6593 dev_priv->fsb_freq, dev_priv->mem_freq);
6594 /* Disable CxSR and never update its watermark again */
6595 intel_set_memory_cxsr(dev_priv, false);
6596 dev_priv->display.update_wm = NULL;
6598 dev_priv->display.update_wm = pineview_update_wm;
6599 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
6600 } else if (IS_G4X(dev)) {
6601 dev_priv->display.update_wm = g4x_update_wm;
6602 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
6603 } else if (IS_GEN4(dev)) {
6604 dev_priv->display.update_wm = i965_update_wm;
6605 if (IS_CRESTLINE(dev))
6606 dev_priv->display.init_clock_gating = crestline_init_clock_gating;
6607 else if (IS_BROADWATER(dev))
6608 dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
6609 } else if (IS_GEN3(dev)) {
6610 dev_priv->display.update_wm = i9xx_update_wm;
6611 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
6612 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
6613 } else if (IS_GEN2(dev)) {
6614 if (INTEL_INFO(dev)->num_pipes == 1) {
6615 dev_priv->display.update_wm = i845_update_wm;
6616 dev_priv->display.get_fifo_size = i845_get_fifo_size;
6618 dev_priv->display.update_wm = i9xx_update_wm;
6619 dev_priv->display.get_fifo_size = i830_get_fifo_size;
6622 if (IS_I85X(dev) || IS_I865G(dev))
6623 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
6625 dev_priv->display.init_clock_gating = i830_init_clock_gating;
6627 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
6631 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u32 mbox, u32 *val)
6633 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6635 if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
6636 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
6640 I915_WRITE(GEN6_PCODE_DATA, *val);
6641 I915_WRITE(GEN6_PCODE_DATA1, 0);
6642 I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
6644 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
6646 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
6650 *val = I915_READ(GEN6_PCODE_DATA);
6651 I915_WRITE(GEN6_PCODE_DATA, 0);
6656 int sandybridge_pcode_write(struct drm_i915_private *dev_priv, u32 mbox, u32 val)
6658 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6660 if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
6661 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
6665 I915_WRITE(GEN6_PCODE_DATA, val);
6666 I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
6668 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
6670 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
6674 I915_WRITE(GEN6_PCODE_DATA, 0);
6679 static int vlv_gpu_freq_div(unsigned int czclk_freq)
6681 switch (czclk_freq) {
6696 static int byt_gpu_freq(struct drm_i915_private *dev_priv, int val)
6698 int div, czclk_freq = DIV_ROUND_CLOSEST(dev_priv->mem_freq, 4);
6700 div = vlv_gpu_freq_div(czclk_freq);
6704 return DIV_ROUND_CLOSEST(czclk_freq * (val + 6 - 0xbd), div);
6707 static int byt_freq_opcode(struct drm_i915_private *dev_priv, int val)
6709 int mul, czclk_freq = DIV_ROUND_CLOSEST(dev_priv->mem_freq, 4);
6711 mul = vlv_gpu_freq_div(czclk_freq);
6715 return DIV_ROUND_CLOSEST(mul * val, czclk_freq) + 0xbd - 6;
6718 static int chv_gpu_freq(struct drm_i915_private *dev_priv, int val)
6720 int div, czclk_freq = dev_priv->rps.cz_freq;
6722 div = vlv_gpu_freq_div(czclk_freq) / 2;
6726 return DIV_ROUND_CLOSEST(czclk_freq * val, 2 * div) / 2;
6729 static int chv_freq_opcode(struct drm_i915_private *dev_priv, int val)
6731 int mul, czclk_freq = dev_priv->rps.cz_freq;
6733 mul = vlv_gpu_freq_div(czclk_freq) / 2;
6737 /* CHV needs even values */
6738 return DIV_ROUND_CLOSEST(val * 2 * mul, czclk_freq) * 2;
6741 int intel_gpu_freq(struct drm_i915_private *dev_priv, int val)
6743 if (IS_GEN9(dev_priv->dev))
6744 return (val * GT_FREQUENCY_MULTIPLIER) / GEN9_FREQ_SCALER;
6745 else if (IS_CHERRYVIEW(dev_priv->dev))
6746 return chv_gpu_freq(dev_priv, val);
6747 else if (IS_VALLEYVIEW(dev_priv->dev))
6748 return byt_gpu_freq(dev_priv, val);
6750 return val * GT_FREQUENCY_MULTIPLIER;
6753 int intel_freq_opcode(struct drm_i915_private *dev_priv, int val)
6755 if (IS_GEN9(dev_priv->dev))
6756 return (val * GEN9_FREQ_SCALER) / GT_FREQUENCY_MULTIPLIER;
6757 else if (IS_CHERRYVIEW(dev_priv->dev))
6758 return chv_freq_opcode(dev_priv, val);
6759 else if (IS_VALLEYVIEW(dev_priv->dev))
6760 return byt_freq_opcode(dev_priv, val);
6762 return val / GT_FREQUENCY_MULTIPLIER;
6765 void intel_pm_setup(struct drm_device *dev)
6767 struct drm_i915_private *dev_priv = dev->dev_private;
6769 mutex_init(&dev_priv->rps.hw_lock);
6771 INIT_DELAYED_WORK(&dev_priv->rps.delayed_resume_work,
6772 intel_gen6_powersave_work);
6774 dev_priv->pm.suspended = false;