2 * Copyright © 2012 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eugeni Dodonov <eugeni.dodonov@intel.com>
28 #include <linux/cpufreq.h>
29 #include <drm/drm_plane_helper.h>
31 #include "intel_drv.h"
32 #include "../../../platform/x86/intel_ips.h"
33 #include <linux/module.h>
38 * RC6 is a special power stage which allows the GPU to enter an very
39 * low-voltage mode when idle, using down to 0V while at this stage. This
40 * stage is entered automatically when the GPU is idle when RC6 support is
41 * enabled, and as soon as new workload arises GPU wakes up automatically as well.
43 * There are different RC6 modes available in Intel GPU, which differentiate
44 * among each other with the latency required to enter and leave RC6 and
45 * voltage consumed by the GPU in different states.
47 * The combination of the following flags define which states GPU is allowed
48 * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
49 * RC6pp is deepest RC6. Their support by hardware varies according to the
50 * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
51 * which brings the most power savings; deeper states save more power, but
52 * require higher latency to switch to and wake up.
54 #define INTEL_RC6_ENABLE (1<<0)
55 #define INTEL_RC6p_ENABLE (1<<1)
56 #define INTEL_RC6pp_ENABLE (1<<2)
58 static void gen9_init_clock_gating(struct drm_device *dev)
60 struct drm_i915_private *dev_priv = dev->dev_private;
62 /* See Bspec note for PSR2_CTL bit 31, Wa#828:skl,bxt,kbl */
63 I915_WRITE(CHICKEN_PAR1_1,
64 I915_READ(CHICKEN_PAR1_1) | SKL_EDP_PSR_FIX_RDWRAP);
66 I915_WRITE(GEN8_CONFIG0,
67 I915_READ(GEN8_CONFIG0) | GEN9_DEFAULT_FIXES);
69 /* WaEnableChickenDCPR:skl,bxt,kbl */
70 I915_WRITE(GEN8_CHICKEN_DCPR_1,
71 I915_READ(GEN8_CHICKEN_DCPR_1) | MASK_WAKEMEM);
73 /* WaFbcTurnOffFbcWatermark:skl,bxt,kbl */
74 /* WaFbcWakeMemOn:skl,bxt,kbl */
75 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
77 DISP_FBC_MEMORY_WAKE);
79 /* WaFbcHighMemBwCorruptionAvoidance:skl,bxt,kbl */
80 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
81 ILK_DPFC_DISABLE_DUMMY0);
84 static void bxt_init_clock_gating(struct drm_device *dev)
86 struct drm_i915_private *dev_priv = to_i915(dev);
88 gen9_init_clock_gating(dev);
90 /* WaDisableSDEUnitClockGating:bxt */
91 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
92 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
96 * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
98 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
99 GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ);
102 * Wa: Backlight PWM may stop in the asserted state, causing backlight
105 if (IS_BXT_REVID(dev_priv, BXT_REVID_B0, REVID_FOREVER))
106 I915_WRITE(GEN9_CLKGATE_DIS_0, I915_READ(GEN9_CLKGATE_DIS_0) |
107 PWM1_GATING_DIS | PWM2_GATING_DIS);
110 static void i915_pineview_get_mem_freq(struct drm_device *dev)
112 struct drm_i915_private *dev_priv = to_i915(dev);
115 tmp = I915_READ(CLKCFG);
117 switch (tmp & CLKCFG_FSB_MASK) {
119 dev_priv->fsb_freq = 533; /* 133*4 */
122 dev_priv->fsb_freq = 800; /* 200*4 */
125 dev_priv->fsb_freq = 667; /* 167*4 */
128 dev_priv->fsb_freq = 400; /* 100*4 */
132 switch (tmp & CLKCFG_MEM_MASK) {
134 dev_priv->mem_freq = 533;
137 dev_priv->mem_freq = 667;
140 dev_priv->mem_freq = 800;
144 /* detect pineview DDR3 setting */
145 tmp = I915_READ(CSHRDDR3CTL);
146 dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
149 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
151 struct drm_i915_private *dev_priv = to_i915(dev);
154 ddrpll = I915_READ16(DDRMPLL1);
155 csipll = I915_READ16(CSIPLL0);
157 switch (ddrpll & 0xff) {
159 dev_priv->mem_freq = 800;
162 dev_priv->mem_freq = 1066;
165 dev_priv->mem_freq = 1333;
168 dev_priv->mem_freq = 1600;
171 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
173 dev_priv->mem_freq = 0;
177 dev_priv->ips.r_t = dev_priv->mem_freq;
179 switch (csipll & 0x3ff) {
181 dev_priv->fsb_freq = 3200;
184 dev_priv->fsb_freq = 3733;
187 dev_priv->fsb_freq = 4266;
190 dev_priv->fsb_freq = 4800;
193 dev_priv->fsb_freq = 5333;
196 dev_priv->fsb_freq = 5866;
199 dev_priv->fsb_freq = 6400;
202 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
204 dev_priv->fsb_freq = 0;
208 if (dev_priv->fsb_freq == 3200) {
209 dev_priv->ips.c_m = 0;
210 } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
211 dev_priv->ips.c_m = 1;
213 dev_priv->ips.c_m = 2;
217 static const struct cxsr_latency cxsr_latency_table[] = {
218 {1, 0, 800, 400, 3382, 33382, 3983, 33983}, /* DDR2-400 SC */
219 {1, 0, 800, 667, 3354, 33354, 3807, 33807}, /* DDR2-667 SC */
220 {1, 0, 800, 800, 3347, 33347, 3763, 33763}, /* DDR2-800 SC */
221 {1, 1, 800, 667, 6420, 36420, 6873, 36873}, /* DDR3-667 SC */
222 {1, 1, 800, 800, 5902, 35902, 6318, 36318}, /* DDR3-800 SC */
224 {1, 0, 667, 400, 3400, 33400, 4021, 34021}, /* DDR2-400 SC */
225 {1, 0, 667, 667, 3372, 33372, 3845, 33845}, /* DDR2-667 SC */
226 {1, 0, 667, 800, 3386, 33386, 3822, 33822}, /* DDR2-800 SC */
227 {1, 1, 667, 667, 6438, 36438, 6911, 36911}, /* DDR3-667 SC */
228 {1, 1, 667, 800, 5941, 35941, 6377, 36377}, /* DDR3-800 SC */
230 {1, 0, 400, 400, 3472, 33472, 4173, 34173}, /* DDR2-400 SC */
231 {1, 0, 400, 667, 3443, 33443, 3996, 33996}, /* DDR2-667 SC */
232 {1, 0, 400, 800, 3430, 33430, 3946, 33946}, /* DDR2-800 SC */
233 {1, 1, 400, 667, 6509, 36509, 7062, 37062}, /* DDR3-667 SC */
234 {1, 1, 400, 800, 5985, 35985, 6501, 36501}, /* DDR3-800 SC */
236 {0, 0, 800, 400, 3438, 33438, 4065, 34065}, /* DDR2-400 SC */
237 {0, 0, 800, 667, 3410, 33410, 3889, 33889}, /* DDR2-667 SC */
238 {0, 0, 800, 800, 3403, 33403, 3845, 33845}, /* DDR2-800 SC */
239 {0, 1, 800, 667, 6476, 36476, 6955, 36955}, /* DDR3-667 SC */
240 {0, 1, 800, 800, 5958, 35958, 6400, 36400}, /* DDR3-800 SC */
242 {0, 0, 667, 400, 3456, 33456, 4103, 34106}, /* DDR2-400 SC */
243 {0, 0, 667, 667, 3428, 33428, 3927, 33927}, /* DDR2-667 SC */
244 {0, 0, 667, 800, 3443, 33443, 3905, 33905}, /* DDR2-800 SC */
245 {0, 1, 667, 667, 6494, 36494, 6993, 36993}, /* DDR3-667 SC */
246 {0, 1, 667, 800, 5998, 35998, 6460, 36460}, /* DDR3-800 SC */
248 {0, 0, 400, 400, 3528, 33528, 4255, 34255}, /* DDR2-400 SC */
249 {0, 0, 400, 667, 3500, 33500, 4079, 34079}, /* DDR2-667 SC */
250 {0, 0, 400, 800, 3487, 33487, 4029, 34029}, /* DDR2-800 SC */
251 {0, 1, 400, 667, 6566, 36566, 7145, 37145}, /* DDR3-667 SC */
252 {0, 1, 400, 800, 6042, 36042, 6584, 36584}, /* DDR3-800 SC */
255 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
260 const struct cxsr_latency *latency;
263 if (fsb == 0 || mem == 0)
266 for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
267 latency = &cxsr_latency_table[i];
268 if (is_desktop == latency->is_desktop &&
269 is_ddr3 == latency->is_ddr3 &&
270 fsb == latency->fsb_freq && mem == latency->mem_freq)
274 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
279 static void chv_set_memory_dvfs(struct drm_i915_private *dev_priv, bool enable)
283 mutex_lock(&dev_priv->rps.hw_lock);
285 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
287 val &= ~FORCE_DDR_HIGH_FREQ;
289 val |= FORCE_DDR_HIGH_FREQ;
290 val &= ~FORCE_DDR_LOW_FREQ;
291 val |= FORCE_DDR_FREQ_REQ_ACK;
292 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
294 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
295 FORCE_DDR_FREQ_REQ_ACK) == 0, 3))
296 DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
298 mutex_unlock(&dev_priv->rps.hw_lock);
301 static void chv_set_memory_pm5(struct drm_i915_private *dev_priv, bool enable)
305 mutex_lock(&dev_priv->rps.hw_lock);
307 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
309 val |= DSP_MAXFIFO_PM5_ENABLE;
311 val &= ~DSP_MAXFIFO_PM5_ENABLE;
312 vlv_punit_write(dev_priv, PUNIT_REG_DSPFREQ, val);
314 mutex_unlock(&dev_priv->rps.hw_lock);
317 #define FW_WM(value, plane) \
318 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
320 void intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
322 struct drm_device *dev = &dev_priv->drm;
325 if (IS_VALLEYVIEW(dev) || IS_CHERRYVIEW(dev)) {
326 I915_WRITE(FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
327 POSTING_READ(FW_BLC_SELF_VLV);
328 dev_priv->wm.vlv.cxsr = enable;
329 } else if (IS_G4X(dev) || IS_CRESTLINE(dev)) {
330 I915_WRITE(FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
331 POSTING_READ(FW_BLC_SELF);
332 } else if (IS_PINEVIEW(dev)) {
333 val = I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN;
334 val |= enable ? PINEVIEW_SELF_REFRESH_EN : 0;
335 I915_WRITE(DSPFW3, val);
336 POSTING_READ(DSPFW3);
337 } else if (IS_I945G(dev) || IS_I945GM(dev)) {
338 val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
339 _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
340 I915_WRITE(FW_BLC_SELF, val);
341 POSTING_READ(FW_BLC_SELF);
342 } else if (IS_I915GM(dev)) {
344 * FIXME can't find a bit like this for 915G, and
345 * and yet it does have the related watermark in
346 * FW_BLC_SELF. What's going on?
348 val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
349 _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
350 I915_WRITE(INSTPM, val);
351 POSTING_READ(INSTPM);
356 DRM_DEBUG_KMS("memory self-refresh is %s\n",
357 enable ? "enabled" : "disabled");
362 * Latency for FIFO fetches is dependent on several factors:
363 * - memory configuration (speed, channels)
365 * - current MCH state
366 * It can be fairly high in some situations, so here we assume a fairly
367 * pessimal value. It's a tradeoff between extra memory fetches (if we
368 * set this value too high, the FIFO will fetch frequently to stay full)
369 * and power consumption (set it too low to save power and we might see
370 * FIFO underruns and display "flicker").
372 * A value of 5us seems to be a good balance; safe for very low end
373 * platforms but not overly aggressive on lower latency configs.
375 static const int pessimal_latency_ns = 5000;
377 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
378 ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
380 static int vlv_get_fifo_size(struct drm_device *dev,
381 enum pipe pipe, int plane)
383 struct drm_i915_private *dev_priv = to_i915(dev);
384 int sprite0_start, sprite1_start, size;
387 uint32_t dsparb, dsparb2, dsparb3;
389 dsparb = I915_READ(DSPARB);
390 dsparb2 = I915_READ(DSPARB2);
391 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 0, 0);
392 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 8, 4);
395 dsparb = I915_READ(DSPARB);
396 dsparb2 = I915_READ(DSPARB2);
397 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 16, 8);
398 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 24, 12);
401 dsparb2 = I915_READ(DSPARB2);
402 dsparb3 = I915_READ(DSPARB3);
403 sprite0_start = VLV_FIFO_START(dsparb3, dsparb2, 0, 16);
404 sprite1_start = VLV_FIFO_START(dsparb3, dsparb2, 8, 20);
412 size = sprite0_start;
415 size = sprite1_start - sprite0_start;
418 size = 512 - 1 - sprite1_start;
424 DRM_DEBUG_KMS("Pipe %c %s %c FIFO size: %d\n",
425 pipe_name(pipe), plane == 0 ? "primary" : "sprite",
426 plane == 0 ? plane_name(pipe) : sprite_name(pipe, plane - 1),
432 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
434 struct drm_i915_private *dev_priv = to_i915(dev);
435 uint32_t dsparb = I915_READ(DSPARB);
438 size = dsparb & 0x7f;
440 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
442 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
443 plane ? "B" : "A", size);
448 static int i830_get_fifo_size(struct drm_device *dev, int plane)
450 struct drm_i915_private *dev_priv = to_i915(dev);
451 uint32_t dsparb = I915_READ(DSPARB);
454 size = dsparb & 0x1ff;
456 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
457 size >>= 1; /* Convert to cachelines */
459 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
460 plane ? "B" : "A", size);
465 static int i845_get_fifo_size(struct drm_device *dev, int plane)
467 struct drm_i915_private *dev_priv = to_i915(dev);
468 uint32_t dsparb = I915_READ(DSPARB);
471 size = dsparb & 0x7f;
472 size >>= 2; /* Convert to cachelines */
474 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
481 /* Pineview has different values for various configs */
482 static const struct intel_watermark_params pineview_display_wm = {
483 .fifo_size = PINEVIEW_DISPLAY_FIFO,
484 .max_wm = PINEVIEW_MAX_WM,
485 .default_wm = PINEVIEW_DFT_WM,
486 .guard_size = PINEVIEW_GUARD_WM,
487 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
489 static const struct intel_watermark_params pineview_display_hplloff_wm = {
490 .fifo_size = PINEVIEW_DISPLAY_FIFO,
491 .max_wm = PINEVIEW_MAX_WM,
492 .default_wm = PINEVIEW_DFT_HPLLOFF_WM,
493 .guard_size = PINEVIEW_GUARD_WM,
494 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
496 static const struct intel_watermark_params pineview_cursor_wm = {
497 .fifo_size = PINEVIEW_CURSOR_FIFO,
498 .max_wm = PINEVIEW_CURSOR_MAX_WM,
499 .default_wm = PINEVIEW_CURSOR_DFT_WM,
500 .guard_size = PINEVIEW_CURSOR_GUARD_WM,
501 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
503 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
504 .fifo_size = PINEVIEW_CURSOR_FIFO,
505 .max_wm = PINEVIEW_CURSOR_MAX_WM,
506 .default_wm = PINEVIEW_CURSOR_DFT_WM,
507 .guard_size = PINEVIEW_CURSOR_GUARD_WM,
508 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
510 static const struct intel_watermark_params g4x_wm_info = {
511 .fifo_size = G4X_FIFO_SIZE,
512 .max_wm = G4X_MAX_WM,
513 .default_wm = G4X_MAX_WM,
515 .cacheline_size = G4X_FIFO_LINE_SIZE,
517 static const struct intel_watermark_params g4x_cursor_wm_info = {
518 .fifo_size = I965_CURSOR_FIFO,
519 .max_wm = I965_CURSOR_MAX_WM,
520 .default_wm = I965_CURSOR_DFT_WM,
522 .cacheline_size = G4X_FIFO_LINE_SIZE,
524 static const struct intel_watermark_params i965_cursor_wm_info = {
525 .fifo_size = I965_CURSOR_FIFO,
526 .max_wm = I965_CURSOR_MAX_WM,
527 .default_wm = I965_CURSOR_DFT_WM,
529 .cacheline_size = I915_FIFO_LINE_SIZE,
531 static const struct intel_watermark_params i945_wm_info = {
532 .fifo_size = I945_FIFO_SIZE,
533 .max_wm = I915_MAX_WM,
536 .cacheline_size = I915_FIFO_LINE_SIZE,
538 static const struct intel_watermark_params i915_wm_info = {
539 .fifo_size = I915_FIFO_SIZE,
540 .max_wm = I915_MAX_WM,
543 .cacheline_size = I915_FIFO_LINE_SIZE,
545 static const struct intel_watermark_params i830_a_wm_info = {
546 .fifo_size = I855GM_FIFO_SIZE,
547 .max_wm = I915_MAX_WM,
550 .cacheline_size = I830_FIFO_LINE_SIZE,
552 static const struct intel_watermark_params i830_bc_wm_info = {
553 .fifo_size = I855GM_FIFO_SIZE,
554 .max_wm = I915_MAX_WM/2,
557 .cacheline_size = I830_FIFO_LINE_SIZE,
559 static const struct intel_watermark_params i845_wm_info = {
560 .fifo_size = I830_FIFO_SIZE,
561 .max_wm = I915_MAX_WM,
564 .cacheline_size = I830_FIFO_LINE_SIZE,
568 * intel_calculate_wm - calculate watermark level
569 * @clock_in_khz: pixel clock
570 * @wm: chip FIFO params
571 * @cpp: bytes per pixel
572 * @latency_ns: memory latency for the platform
574 * Calculate the watermark level (the level at which the display plane will
575 * start fetching from memory again). Each chip has a different display
576 * FIFO size and allocation, so the caller needs to figure that out and pass
577 * in the correct intel_watermark_params structure.
579 * As the pixel clock runs, the FIFO will be drained at a rate that depends
580 * on the pixel size. When it reaches the watermark level, it'll start
581 * fetching FIFO line sized based chunks from memory until the FIFO fills
582 * past the watermark point. If the FIFO drains completely, a FIFO underrun
583 * will occur, and a display engine hang could result.
585 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
586 const struct intel_watermark_params *wm,
587 int fifo_size, int cpp,
588 unsigned long latency_ns)
590 long entries_required, wm_size;
593 * Note: we need to make sure we don't overflow for various clock &
595 * clocks go from a few thousand to several hundred thousand.
596 * latency is usually a few thousand
598 entries_required = ((clock_in_khz / 1000) * cpp * latency_ns) /
600 entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
602 DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
604 wm_size = fifo_size - (entries_required + wm->guard_size);
606 DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
608 /* Don't promote wm_size to unsigned... */
609 if (wm_size > (long)wm->max_wm)
610 wm_size = wm->max_wm;
612 wm_size = wm->default_wm;
615 * Bspec seems to indicate that the value shouldn't be lower than
616 * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
617 * Lets go for 8 which is the burst size since certain platforms
618 * already use a hardcoded 8 (which is what the spec says should be
627 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
629 struct drm_crtc *crtc, *enabled = NULL;
631 for_each_crtc(dev, crtc) {
632 if (intel_crtc_active(crtc)) {
642 static void pineview_update_wm(struct drm_crtc *unused_crtc)
644 struct drm_device *dev = unused_crtc->dev;
645 struct drm_i915_private *dev_priv = to_i915(dev);
646 struct drm_crtc *crtc;
647 const struct cxsr_latency *latency;
651 latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
652 dev_priv->fsb_freq, dev_priv->mem_freq);
654 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
655 intel_set_memory_cxsr(dev_priv, false);
659 crtc = single_enabled_crtc(dev);
661 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
662 int cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
663 int clock = adjusted_mode->crtc_clock;
666 wm = intel_calculate_wm(clock, &pineview_display_wm,
667 pineview_display_wm.fifo_size,
668 cpp, latency->display_sr);
669 reg = I915_READ(DSPFW1);
670 reg &= ~DSPFW_SR_MASK;
671 reg |= FW_WM(wm, SR);
672 I915_WRITE(DSPFW1, reg);
673 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
676 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
677 pineview_display_wm.fifo_size,
678 cpp, latency->cursor_sr);
679 reg = I915_READ(DSPFW3);
680 reg &= ~DSPFW_CURSOR_SR_MASK;
681 reg |= FW_WM(wm, CURSOR_SR);
682 I915_WRITE(DSPFW3, reg);
684 /* Display HPLL off SR */
685 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
686 pineview_display_hplloff_wm.fifo_size,
687 cpp, latency->display_hpll_disable);
688 reg = I915_READ(DSPFW3);
689 reg &= ~DSPFW_HPLL_SR_MASK;
690 reg |= FW_WM(wm, HPLL_SR);
691 I915_WRITE(DSPFW3, reg);
693 /* cursor HPLL off SR */
694 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
695 pineview_display_hplloff_wm.fifo_size,
696 cpp, latency->cursor_hpll_disable);
697 reg = I915_READ(DSPFW3);
698 reg &= ~DSPFW_HPLL_CURSOR_MASK;
699 reg |= FW_WM(wm, HPLL_CURSOR);
700 I915_WRITE(DSPFW3, reg);
701 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
703 intel_set_memory_cxsr(dev_priv, true);
705 intel_set_memory_cxsr(dev_priv, false);
709 static bool g4x_compute_wm0(struct drm_device *dev,
711 const struct intel_watermark_params *display,
712 int display_latency_ns,
713 const struct intel_watermark_params *cursor,
714 int cursor_latency_ns,
718 struct drm_crtc *crtc;
719 const struct drm_display_mode *adjusted_mode;
720 int htotal, hdisplay, clock, cpp;
721 int line_time_us, line_count;
722 int entries, tlb_miss;
724 crtc = intel_get_crtc_for_plane(dev, plane);
725 if (!intel_crtc_active(crtc)) {
726 *cursor_wm = cursor->guard_size;
727 *plane_wm = display->guard_size;
731 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
732 clock = adjusted_mode->crtc_clock;
733 htotal = adjusted_mode->crtc_htotal;
734 hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
735 cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
737 /* Use the small buffer method to calculate plane watermark */
738 entries = ((clock * cpp / 1000) * display_latency_ns) / 1000;
739 tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
742 entries = DIV_ROUND_UP(entries, display->cacheline_size);
743 *plane_wm = entries + display->guard_size;
744 if (*plane_wm > (int)display->max_wm)
745 *plane_wm = display->max_wm;
747 /* Use the large buffer method to calculate cursor watermark */
748 line_time_us = max(htotal * 1000 / clock, 1);
749 line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
750 entries = line_count * crtc->cursor->state->crtc_w * cpp;
751 tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
754 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
755 *cursor_wm = entries + cursor->guard_size;
756 if (*cursor_wm > (int)cursor->max_wm)
757 *cursor_wm = (int)cursor->max_wm;
763 * Check the wm result.
765 * If any calculated watermark values is larger than the maximum value that
766 * can be programmed into the associated watermark register, that watermark
769 static bool g4x_check_srwm(struct drm_device *dev,
770 int display_wm, int cursor_wm,
771 const struct intel_watermark_params *display,
772 const struct intel_watermark_params *cursor)
774 DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
775 display_wm, cursor_wm);
777 if (display_wm > display->max_wm) {
778 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
779 display_wm, display->max_wm);
783 if (cursor_wm > cursor->max_wm) {
784 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
785 cursor_wm, cursor->max_wm);
789 if (!(display_wm || cursor_wm)) {
790 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
797 static bool g4x_compute_srwm(struct drm_device *dev,
800 const struct intel_watermark_params *display,
801 const struct intel_watermark_params *cursor,
802 int *display_wm, int *cursor_wm)
804 struct drm_crtc *crtc;
805 const struct drm_display_mode *adjusted_mode;
806 int hdisplay, htotal, cpp, clock;
807 unsigned long line_time_us;
808 int line_count, line_size;
813 *display_wm = *cursor_wm = 0;
817 crtc = intel_get_crtc_for_plane(dev, plane);
818 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
819 clock = adjusted_mode->crtc_clock;
820 htotal = adjusted_mode->crtc_htotal;
821 hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
822 cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
824 line_time_us = max(htotal * 1000 / clock, 1);
825 line_count = (latency_ns / line_time_us + 1000) / 1000;
826 line_size = hdisplay * cpp;
828 /* Use the minimum of the small and large buffer method for primary */
829 small = ((clock * cpp / 1000) * latency_ns) / 1000;
830 large = line_count * line_size;
832 entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
833 *display_wm = entries + display->guard_size;
835 /* calculate the self-refresh watermark for display cursor */
836 entries = line_count * cpp * crtc->cursor->state->crtc_w;
837 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
838 *cursor_wm = entries + cursor->guard_size;
840 return g4x_check_srwm(dev,
841 *display_wm, *cursor_wm,
845 #define FW_WM_VLV(value, plane) \
846 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
848 static void vlv_write_wm_values(struct intel_crtc *crtc,
849 const struct vlv_wm_values *wm)
851 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
852 enum pipe pipe = crtc->pipe;
854 I915_WRITE(VLV_DDL(pipe),
855 (wm->ddl[pipe].cursor << DDL_CURSOR_SHIFT) |
856 (wm->ddl[pipe].sprite[1] << DDL_SPRITE_SHIFT(1)) |
857 (wm->ddl[pipe].sprite[0] << DDL_SPRITE_SHIFT(0)) |
858 (wm->ddl[pipe].primary << DDL_PLANE_SHIFT));
861 FW_WM(wm->sr.plane, SR) |
862 FW_WM(wm->pipe[PIPE_B].cursor, CURSORB) |
863 FW_WM_VLV(wm->pipe[PIPE_B].primary, PLANEB) |
864 FW_WM_VLV(wm->pipe[PIPE_A].primary, PLANEA));
866 FW_WM_VLV(wm->pipe[PIPE_A].sprite[1], SPRITEB) |
867 FW_WM(wm->pipe[PIPE_A].cursor, CURSORA) |
868 FW_WM_VLV(wm->pipe[PIPE_A].sprite[0], SPRITEA));
870 FW_WM(wm->sr.cursor, CURSOR_SR));
872 if (IS_CHERRYVIEW(dev_priv)) {
873 I915_WRITE(DSPFW7_CHV,
874 FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
875 FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
876 I915_WRITE(DSPFW8_CHV,
877 FW_WM_VLV(wm->pipe[PIPE_C].sprite[1], SPRITEF) |
878 FW_WM_VLV(wm->pipe[PIPE_C].sprite[0], SPRITEE));
879 I915_WRITE(DSPFW9_CHV,
880 FW_WM_VLV(wm->pipe[PIPE_C].primary, PLANEC) |
881 FW_WM(wm->pipe[PIPE_C].cursor, CURSORC));
883 FW_WM(wm->sr.plane >> 9, SR_HI) |
884 FW_WM(wm->pipe[PIPE_C].sprite[1] >> 8, SPRITEF_HI) |
885 FW_WM(wm->pipe[PIPE_C].sprite[0] >> 8, SPRITEE_HI) |
886 FW_WM(wm->pipe[PIPE_C].primary >> 8, PLANEC_HI) |
887 FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
888 FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
889 FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
890 FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
891 FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
892 FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
895 FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
896 FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
898 FW_WM(wm->sr.plane >> 9, SR_HI) |
899 FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
900 FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
901 FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
902 FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
903 FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
904 FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
907 /* zero (unused) WM1 watermarks */
908 I915_WRITE(DSPFW4, 0);
909 I915_WRITE(DSPFW5, 0);
910 I915_WRITE(DSPFW6, 0);
911 I915_WRITE(DSPHOWM1, 0);
913 POSTING_READ(DSPFW1);
921 VLV_WM_LEVEL_DDR_DVFS,
924 /* latency must be in 0.1us units. */
925 static unsigned int vlv_wm_method2(unsigned int pixel_rate,
926 unsigned int pipe_htotal,
927 unsigned int horiz_pixels,
929 unsigned int latency)
933 ret = (latency * pixel_rate) / (pipe_htotal * 10000);
934 ret = (ret + 1) * horiz_pixels * cpp;
935 ret = DIV_ROUND_UP(ret, 64);
940 static void vlv_setup_wm_latency(struct drm_device *dev)
942 struct drm_i915_private *dev_priv = to_i915(dev);
944 /* all latencies in usec */
945 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM2] = 3;
947 dev_priv->wm.max_level = VLV_WM_LEVEL_PM2;
949 if (IS_CHERRYVIEW(dev_priv)) {
950 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM5] = 12;
951 dev_priv->wm.pri_latency[VLV_WM_LEVEL_DDR_DVFS] = 33;
953 dev_priv->wm.max_level = VLV_WM_LEVEL_DDR_DVFS;
957 static uint16_t vlv_compute_wm_level(struct intel_plane *plane,
958 struct intel_crtc *crtc,
959 const struct intel_plane_state *state,
962 struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
963 int clock, htotal, cpp, width, wm;
965 if (dev_priv->wm.pri_latency[level] == 0)
968 if (!state->base.visible)
971 cpp = drm_format_plane_cpp(state->base.fb->pixel_format, 0);
972 clock = crtc->config->base.adjusted_mode.crtc_clock;
973 htotal = crtc->config->base.adjusted_mode.crtc_htotal;
974 width = crtc->config->pipe_src_w;
975 if (WARN_ON(htotal == 0))
978 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
980 * FIXME the formula gives values that are
981 * too big for the cursor FIFO, and hence we
982 * would never be able to use cursors. For
983 * now just hardcode the watermark.
987 wm = vlv_wm_method2(clock, htotal, width, cpp,
988 dev_priv->wm.pri_latency[level] * 10);
991 return min_t(int, wm, USHRT_MAX);
994 static void vlv_compute_fifo(struct intel_crtc *crtc)
996 struct drm_device *dev = crtc->base.dev;
997 struct vlv_wm_state *wm_state = &crtc->wm_state;
998 struct intel_plane *plane;
999 unsigned int total_rate = 0;
1000 const int fifo_size = 512 - 1;
1001 int fifo_extra, fifo_left = fifo_size;
1003 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1004 struct intel_plane_state *state =
1005 to_intel_plane_state(plane->base.state);
1007 if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
1010 if (state->base.visible) {
1011 wm_state->num_active_planes++;
1012 total_rate += drm_format_plane_cpp(state->base.fb->pixel_format, 0);
1016 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1017 struct intel_plane_state *state =
1018 to_intel_plane_state(plane->base.state);
1021 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
1022 plane->wm.fifo_size = 63;
1026 if (!state->base.visible) {
1027 plane->wm.fifo_size = 0;
1031 rate = drm_format_plane_cpp(state->base.fb->pixel_format, 0);
1032 plane->wm.fifo_size = fifo_size * rate / total_rate;
1033 fifo_left -= plane->wm.fifo_size;
1036 fifo_extra = DIV_ROUND_UP(fifo_left, wm_state->num_active_planes ?: 1);
1038 /* spread the remainder evenly */
1039 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1045 if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
1048 /* give it all to the first plane if none are active */
1049 if (plane->wm.fifo_size == 0 &&
1050 wm_state->num_active_planes)
1053 plane_extra = min(fifo_extra, fifo_left);
1054 plane->wm.fifo_size += plane_extra;
1055 fifo_left -= plane_extra;
1058 WARN_ON(fifo_left != 0);
1061 static void vlv_invert_wms(struct intel_crtc *crtc)
1063 struct vlv_wm_state *wm_state = &crtc->wm_state;
1066 for (level = 0; level < wm_state->num_levels; level++) {
1067 struct drm_device *dev = crtc->base.dev;
1068 const int sr_fifo_size = INTEL_INFO(dev)->num_pipes * 512 - 1;
1069 struct intel_plane *plane;
1071 wm_state->sr[level].plane = sr_fifo_size - wm_state->sr[level].plane;
1072 wm_state->sr[level].cursor = 63 - wm_state->sr[level].cursor;
1074 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1075 switch (plane->base.type) {
1077 case DRM_PLANE_TYPE_CURSOR:
1078 wm_state->wm[level].cursor = plane->wm.fifo_size -
1079 wm_state->wm[level].cursor;
1081 case DRM_PLANE_TYPE_PRIMARY:
1082 wm_state->wm[level].primary = plane->wm.fifo_size -
1083 wm_state->wm[level].primary;
1085 case DRM_PLANE_TYPE_OVERLAY:
1086 sprite = plane->plane;
1087 wm_state->wm[level].sprite[sprite] = plane->wm.fifo_size -
1088 wm_state->wm[level].sprite[sprite];
1095 static void vlv_compute_wm(struct intel_crtc *crtc)
1097 struct drm_device *dev = crtc->base.dev;
1098 struct vlv_wm_state *wm_state = &crtc->wm_state;
1099 struct intel_plane *plane;
1100 int sr_fifo_size = INTEL_INFO(dev)->num_pipes * 512 - 1;
1103 memset(wm_state, 0, sizeof(*wm_state));
1105 wm_state->cxsr = crtc->pipe != PIPE_C && crtc->wm.cxsr_allowed;
1106 wm_state->num_levels = to_i915(dev)->wm.max_level + 1;
1108 wm_state->num_active_planes = 0;
1110 vlv_compute_fifo(crtc);
1112 if (wm_state->num_active_planes != 1)
1113 wm_state->cxsr = false;
1115 if (wm_state->cxsr) {
1116 for (level = 0; level < wm_state->num_levels; level++) {
1117 wm_state->sr[level].plane = sr_fifo_size;
1118 wm_state->sr[level].cursor = 63;
1122 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1123 struct intel_plane_state *state =
1124 to_intel_plane_state(plane->base.state);
1126 if (!state->base.visible)
1129 /* normal watermarks */
1130 for (level = 0; level < wm_state->num_levels; level++) {
1131 int wm = vlv_compute_wm_level(plane, crtc, state, level);
1132 int max_wm = plane->base.type == DRM_PLANE_TYPE_CURSOR ? 63 : 511;
1135 if (WARN_ON(level == 0 && wm > max_wm))
1138 if (wm > plane->wm.fifo_size)
1141 switch (plane->base.type) {
1143 case DRM_PLANE_TYPE_CURSOR:
1144 wm_state->wm[level].cursor = wm;
1146 case DRM_PLANE_TYPE_PRIMARY:
1147 wm_state->wm[level].primary = wm;
1149 case DRM_PLANE_TYPE_OVERLAY:
1150 sprite = plane->plane;
1151 wm_state->wm[level].sprite[sprite] = wm;
1156 wm_state->num_levels = level;
1158 if (!wm_state->cxsr)
1161 /* maxfifo watermarks */
1162 switch (plane->base.type) {
1164 case DRM_PLANE_TYPE_CURSOR:
1165 for (level = 0; level < wm_state->num_levels; level++)
1166 wm_state->sr[level].cursor =
1167 wm_state->wm[level].cursor;
1169 case DRM_PLANE_TYPE_PRIMARY:
1170 for (level = 0; level < wm_state->num_levels; level++)
1171 wm_state->sr[level].plane =
1172 min(wm_state->sr[level].plane,
1173 wm_state->wm[level].primary);
1175 case DRM_PLANE_TYPE_OVERLAY:
1176 sprite = plane->plane;
1177 for (level = 0; level < wm_state->num_levels; level++)
1178 wm_state->sr[level].plane =
1179 min(wm_state->sr[level].plane,
1180 wm_state->wm[level].sprite[sprite]);
1185 /* clear any (partially) filled invalid levels */
1186 for (level = wm_state->num_levels; level < to_i915(dev)->wm.max_level + 1; level++) {
1187 memset(&wm_state->wm[level], 0, sizeof(wm_state->wm[level]));
1188 memset(&wm_state->sr[level], 0, sizeof(wm_state->sr[level]));
1191 vlv_invert_wms(crtc);
1194 #define VLV_FIFO(plane, value) \
1195 (((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1197 static void vlv_pipe_set_fifo_size(struct intel_crtc *crtc)
1199 struct drm_device *dev = crtc->base.dev;
1200 struct drm_i915_private *dev_priv = to_i915(dev);
1201 struct intel_plane *plane;
1202 int sprite0_start = 0, sprite1_start = 0, fifo_size = 0;
1204 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1205 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
1206 WARN_ON(plane->wm.fifo_size != 63);
1210 if (plane->base.type == DRM_PLANE_TYPE_PRIMARY)
1211 sprite0_start = plane->wm.fifo_size;
1212 else if (plane->plane == 0)
1213 sprite1_start = sprite0_start + plane->wm.fifo_size;
1215 fifo_size = sprite1_start + plane->wm.fifo_size;
1218 WARN_ON(fifo_size != 512 - 1);
1220 DRM_DEBUG_KMS("Pipe %c FIFO split %d / %d / %d\n",
1221 pipe_name(crtc->pipe), sprite0_start,
1222 sprite1_start, fifo_size);
1224 switch (crtc->pipe) {
1225 uint32_t dsparb, dsparb2, dsparb3;
1227 dsparb = I915_READ(DSPARB);
1228 dsparb2 = I915_READ(DSPARB2);
1230 dsparb &= ~(VLV_FIFO(SPRITEA, 0xff) |
1231 VLV_FIFO(SPRITEB, 0xff));
1232 dsparb |= (VLV_FIFO(SPRITEA, sprite0_start) |
1233 VLV_FIFO(SPRITEB, sprite1_start));
1235 dsparb2 &= ~(VLV_FIFO(SPRITEA_HI, 0x1) |
1236 VLV_FIFO(SPRITEB_HI, 0x1));
1237 dsparb2 |= (VLV_FIFO(SPRITEA_HI, sprite0_start >> 8) |
1238 VLV_FIFO(SPRITEB_HI, sprite1_start >> 8));
1240 I915_WRITE(DSPARB, dsparb);
1241 I915_WRITE(DSPARB2, dsparb2);
1244 dsparb = I915_READ(DSPARB);
1245 dsparb2 = I915_READ(DSPARB2);
1247 dsparb &= ~(VLV_FIFO(SPRITEC, 0xff) |
1248 VLV_FIFO(SPRITED, 0xff));
1249 dsparb |= (VLV_FIFO(SPRITEC, sprite0_start) |
1250 VLV_FIFO(SPRITED, sprite1_start));
1252 dsparb2 &= ~(VLV_FIFO(SPRITEC_HI, 0xff) |
1253 VLV_FIFO(SPRITED_HI, 0xff));
1254 dsparb2 |= (VLV_FIFO(SPRITEC_HI, sprite0_start >> 8) |
1255 VLV_FIFO(SPRITED_HI, sprite1_start >> 8));
1257 I915_WRITE(DSPARB, dsparb);
1258 I915_WRITE(DSPARB2, dsparb2);
1261 dsparb3 = I915_READ(DSPARB3);
1262 dsparb2 = I915_READ(DSPARB2);
1264 dsparb3 &= ~(VLV_FIFO(SPRITEE, 0xff) |
1265 VLV_FIFO(SPRITEF, 0xff));
1266 dsparb3 |= (VLV_FIFO(SPRITEE, sprite0_start) |
1267 VLV_FIFO(SPRITEF, sprite1_start));
1269 dsparb2 &= ~(VLV_FIFO(SPRITEE_HI, 0xff) |
1270 VLV_FIFO(SPRITEF_HI, 0xff));
1271 dsparb2 |= (VLV_FIFO(SPRITEE_HI, sprite0_start >> 8) |
1272 VLV_FIFO(SPRITEF_HI, sprite1_start >> 8));
1274 I915_WRITE(DSPARB3, dsparb3);
1275 I915_WRITE(DSPARB2, dsparb2);
1284 static void vlv_merge_wm(struct drm_device *dev,
1285 struct vlv_wm_values *wm)
1287 struct intel_crtc *crtc;
1288 int num_active_crtcs = 0;
1290 wm->level = to_i915(dev)->wm.max_level;
1293 for_each_intel_crtc(dev, crtc) {
1294 const struct vlv_wm_state *wm_state = &crtc->wm_state;
1299 if (!wm_state->cxsr)
1303 wm->level = min_t(int, wm->level, wm_state->num_levels - 1);
1306 if (num_active_crtcs != 1)
1309 if (num_active_crtcs > 1)
1310 wm->level = VLV_WM_LEVEL_PM2;
1312 for_each_intel_crtc(dev, crtc) {
1313 struct vlv_wm_state *wm_state = &crtc->wm_state;
1314 enum pipe pipe = crtc->pipe;
1319 wm->pipe[pipe] = wm_state->wm[wm->level];
1321 wm->sr = wm_state->sr[wm->level];
1323 wm->ddl[pipe].primary = DDL_PRECISION_HIGH | 2;
1324 wm->ddl[pipe].sprite[0] = DDL_PRECISION_HIGH | 2;
1325 wm->ddl[pipe].sprite[1] = DDL_PRECISION_HIGH | 2;
1326 wm->ddl[pipe].cursor = DDL_PRECISION_HIGH | 2;
1330 static void vlv_update_wm(struct drm_crtc *crtc)
1332 struct drm_device *dev = crtc->dev;
1333 struct drm_i915_private *dev_priv = to_i915(dev);
1334 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1335 enum pipe pipe = intel_crtc->pipe;
1336 struct vlv_wm_values wm = {};
1338 vlv_compute_wm(intel_crtc);
1339 vlv_merge_wm(dev, &wm);
1341 if (memcmp(&dev_priv->wm.vlv, &wm, sizeof(wm)) == 0) {
1342 /* FIXME should be part of crtc atomic commit */
1343 vlv_pipe_set_fifo_size(intel_crtc);
1347 if (wm.level < VLV_WM_LEVEL_DDR_DVFS &&
1348 dev_priv->wm.vlv.level >= VLV_WM_LEVEL_DDR_DVFS)
1349 chv_set_memory_dvfs(dev_priv, false);
1351 if (wm.level < VLV_WM_LEVEL_PM5 &&
1352 dev_priv->wm.vlv.level >= VLV_WM_LEVEL_PM5)
1353 chv_set_memory_pm5(dev_priv, false);
1355 if (!wm.cxsr && dev_priv->wm.vlv.cxsr)
1356 intel_set_memory_cxsr(dev_priv, false);
1358 /* FIXME should be part of crtc atomic commit */
1359 vlv_pipe_set_fifo_size(intel_crtc);
1361 vlv_write_wm_values(intel_crtc, &wm);
1363 DRM_DEBUG_KMS("Setting FIFO watermarks - %c: plane=%d, cursor=%d, "
1364 "sprite0=%d, sprite1=%d, SR: plane=%d, cursor=%d level=%d cxsr=%d\n",
1365 pipe_name(pipe), wm.pipe[pipe].primary, wm.pipe[pipe].cursor,
1366 wm.pipe[pipe].sprite[0], wm.pipe[pipe].sprite[1],
1367 wm.sr.plane, wm.sr.cursor, wm.level, wm.cxsr);
1369 if (wm.cxsr && !dev_priv->wm.vlv.cxsr)
1370 intel_set_memory_cxsr(dev_priv, true);
1372 if (wm.level >= VLV_WM_LEVEL_PM5 &&
1373 dev_priv->wm.vlv.level < VLV_WM_LEVEL_PM5)
1374 chv_set_memory_pm5(dev_priv, true);
1376 if (wm.level >= VLV_WM_LEVEL_DDR_DVFS &&
1377 dev_priv->wm.vlv.level < VLV_WM_LEVEL_DDR_DVFS)
1378 chv_set_memory_dvfs(dev_priv, true);
1380 dev_priv->wm.vlv = wm;
1383 #define single_plane_enabled(mask) is_power_of_2(mask)
1385 static void g4x_update_wm(struct drm_crtc *crtc)
1387 struct drm_device *dev = crtc->dev;
1388 static const int sr_latency_ns = 12000;
1389 struct drm_i915_private *dev_priv = to_i915(dev);
1390 int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1391 int plane_sr, cursor_sr;
1392 unsigned int enabled = 0;
1395 if (g4x_compute_wm0(dev, PIPE_A,
1396 &g4x_wm_info, pessimal_latency_ns,
1397 &g4x_cursor_wm_info, pessimal_latency_ns,
1398 &planea_wm, &cursora_wm))
1399 enabled |= 1 << PIPE_A;
1401 if (g4x_compute_wm0(dev, PIPE_B,
1402 &g4x_wm_info, pessimal_latency_ns,
1403 &g4x_cursor_wm_info, pessimal_latency_ns,
1404 &planeb_wm, &cursorb_wm))
1405 enabled |= 1 << PIPE_B;
1407 if (single_plane_enabled(enabled) &&
1408 g4x_compute_srwm(dev, ffs(enabled) - 1,
1411 &g4x_cursor_wm_info,
1412 &plane_sr, &cursor_sr)) {
1413 cxsr_enabled = true;
1415 cxsr_enabled = false;
1416 intel_set_memory_cxsr(dev_priv, false);
1417 plane_sr = cursor_sr = 0;
1420 DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, "
1421 "B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1422 planea_wm, cursora_wm,
1423 planeb_wm, cursorb_wm,
1424 plane_sr, cursor_sr);
1427 FW_WM(plane_sr, SR) |
1428 FW_WM(cursorb_wm, CURSORB) |
1429 FW_WM(planeb_wm, PLANEB) |
1430 FW_WM(planea_wm, PLANEA));
1432 (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1433 FW_WM(cursora_wm, CURSORA));
1434 /* HPLL off in SR has some issues on G4x... disable it */
1436 (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1437 FW_WM(cursor_sr, CURSOR_SR));
1440 intel_set_memory_cxsr(dev_priv, true);
1443 static void i965_update_wm(struct drm_crtc *unused_crtc)
1445 struct drm_device *dev = unused_crtc->dev;
1446 struct drm_i915_private *dev_priv = to_i915(dev);
1447 struct drm_crtc *crtc;
1452 /* Calc sr entries for one plane configs */
1453 crtc = single_enabled_crtc(dev);
1455 /* self-refresh has much higher latency */
1456 static const int sr_latency_ns = 12000;
1457 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1458 int clock = adjusted_mode->crtc_clock;
1459 int htotal = adjusted_mode->crtc_htotal;
1460 int hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
1461 int cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
1462 unsigned long line_time_us;
1465 line_time_us = max(htotal * 1000 / clock, 1);
1467 /* Use ns/us then divide to preserve precision */
1468 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1470 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1471 srwm = I965_FIFO_SIZE - entries;
1475 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1478 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1479 cpp * crtc->cursor->state->crtc_w;
1480 entries = DIV_ROUND_UP(entries,
1481 i965_cursor_wm_info.cacheline_size);
1482 cursor_sr = i965_cursor_wm_info.fifo_size -
1483 (entries + i965_cursor_wm_info.guard_size);
1485 if (cursor_sr > i965_cursor_wm_info.max_wm)
1486 cursor_sr = i965_cursor_wm_info.max_wm;
1488 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1489 "cursor %d\n", srwm, cursor_sr);
1491 cxsr_enabled = true;
1493 cxsr_enabled = false;
1494 /* Turn off self refresh if both pipes are enabled */
1495 intel_set_memory_cxsr(dev_priv, false);
1498 DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1501 /* 965 has limitations... */
1502 I915_WRITE(DSPFW1, FW_WM(srwm, SR) |
1506 I915_WRITE(DSPFW2, FW_WM(8, CURSORA) |
1507 FW_WM(8, PLANEC_OLD));
1508 /* update cursor SR watermark */
1509 I915_WRITE(DSPFW3, FW_WM(cursor_sr, CURSOR_SR));
1512 intel_set_memory_cxsr(dev_priv, true);
1517 static void i9xx_update_wm(struct drm_crtc *unused_crtc)
1519 struct drm_device *dev = unused_crtc->dev;
1520 struct drm_i915_private *dev_priv = to_i915(dev);
1521 const struct intel_watermark_params *wm_info;
1526 int planea_wm, planeb_wm;
1527 struct drm_crtc *crtc, *enabled = NULL;
1530 wm_info = &i945_wm_info;
1531 else if (!IS_GEN2(dev))
1532 wm_info = &i915_wm_info;
1534 wm_info = &i830_a_wm_info;
1536 fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1537 crtc = intel_get_crtc_for_plane(dev, 0);
1538 if (intel_crtc_active(crtc)) {
1539 const struct drm_display_mode *adjusted_mode;
1540 int cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
1544 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1545 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1546 wm_info, fifo_size, cpp,
1547 pessimal_latency_ns);
1550 planea_wm = fifo_size - wm_info->guard_size;
1551 if (planea_wm > (long)wm_info->max_wm)
1552 planea_wm = wm_info->max_wm;
1556 wm_info = &i830_bc_wm_info;
1558 fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1559 crtc = intel_get_crtc_for_plane(dev, 1);
1560 if (intel_crtc_active(crtc)) {
1561 const struct drm_display_mode *adjusted_mode;
1562 int cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
1566 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1567 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1568 wm_info, fifo_size, cpp,
1569 pessimal_latency_ns);
1570 if (enabled == NULL)
1575 planeb_wm = fifo_size - wm_info->guard_size;
1576 if (planeb_wm > (long)wm_info->max_wm)
1577 planeb_wm = wm_info->max_wm;
1580 DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1582 if (IS_I915GM(dev) && enabled) {
1583 struct drm_i915_gem_object *obj;
1585 obj = intel_fb_obj(enabled->primary->state->fb);
1587 /* self-refresh seems busted with untiled */
1588 if (!i915_gem_object_is_tiled(obj))
1593 * Overlay gets an aggressive default since video jitter is bad.
1597 /* Play safe and disable self-refresh before adjusting watermarks. */
1598 intel_set_memory_cxsr(dev_priv, false);
1600 /* Calc sr entries for one plane configs */
1601 if (HAS_FW_BLC(dev) && enabled) {
1602 /* self-refresh has much higher latency */
1603 static const int sr_latency_ns = 6000;
1604 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(enabled)->config->base.adjusted_mode;
1605 int clock = adjusted_mode->crtc_clock;
1606 int htotal = adjusted_mode->crtc_htotal;
1607 int hdisplay = to_intel_crtc(enabled)->config->pipe_src_w;
1608 int cpp = drm_format_plane_cpp(enabled->primary->state->fb->pixel_format, 0);
1609 unsigned long line_time_us;
1612 if (IS_I915GM(dev) || IS_I945GM(dev))
1615 line_time_us = max(htotal * 1000 / clock, 1);
1617 /* Use ns/us then divide to preserve precision */
1618 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1620 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1621 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1622 srwm = wm_info->fifo_size - entries;
1626 if (IS_I945G(dev) || IS_I945GM(dev))
1627 I915_WRITE(FW_BLC_SELF,
1628 FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1630 I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1633 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1634 planea_wm, planeb_wm, cwm, srwm);
1636 fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1637 fwater_hi = (cwm & 0x1f);
1639 /* Set request length to 8 cachelines per fetch */
1640 fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1641 fwater_hi = fwater_hi | (1 << 8);
1643 I915_WRITE(FW_BLC, fwater_lo);
1644 I915_WRITE(FW_BLC2, fwater_hi);
1647 intel_set_memory_cxsr(dev_priv, true);
1650 static void i845_update_wm(struct drm_crtc *unused_crtc)
1652 struct drm_device *dev = unused_crtc->dev;
1653 struct drm_i915_private *dev_priv = to_i915(dev);
1654 struct drm_crtc *crtc;
1655 const struct drm_display_mode *adjusted_mode;
1659 crtc = single_enabled_crtc(dev);
1663 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1664 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1666 dev_priv->display.get_fifo_size(dev, 0),
1667 4, pessimal_latency_ns);
1668 fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1669 fwater_lo |= (3<<8) | planea_wm;
1671 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1673 I915_WRITE(FW_BLC, fwater_lo);
1676 uint32_t ilk_pipe_pixel_rate(const struct intel_crtc_state *pipe_config)
1678 uint32_t pixel_rate;
1680 pixel_rate = pipe_config->base.adjusted_mode.crtc_clock;
1682 /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
1683 * adjust the pixel_rate here. */
1685 if (pipe_config->pch_pfit.enabled) {
1686 uint64_t pipe_w, pipe_h, pfit_w, pfit_h;
1687 uint32_t pfit_size = pipe_config->pch_pfit.size;
1689 pipe_w = pipe_config->pipe_src_w;
1690 pipe_h = pipe_config->pipe_src_h;
1692 pfit_w = (pfit_size >> 16) & 0xFFFF;
1693 pfit_h = pfit_size & 0xFFFF;
1694 if (pipe_w < pfit_w)
1696 if (pipe_h < pfit_h)
1699 if (WARN_ON(!pfit_w || !pfit_h))
1702 pixel_rate = div_u64((uint64_t) pixel_rate * pipe_w * pipe_h,
1709 /* latency must be in 0.1us units. */
1710 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t cpp, uint32_t latency)
1714 if (WARN(latency == 0, "Latency value missing\n"))
1717 ret = (uint64_t) pixel_rate * cpp * latency;
1718 ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
1723 /* latency must be in 0.1us units. */
1724 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
1725 uint32_t horiz_pixels, uint8_t cpp,
1730 if (WARN(latency == 0, "Latency value missing\n"))
1732 if (WARN_ON(!pipe_htotal))
1735 ret = (latency * pixel_rate) / (pipe_htotal * 10000);
1736 ret = (ret + 1) * horiz_pixels * cpp;
1737 ret = DIV_ROUND_UP(ret, 64) + 2;
1741 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
1745 * Neither of these should be possible since this function shouldn't be
1746 * called if the CRTC is off or the plane is invisible. But let's be
1747 * extra paranoid to avoid a potential divide-by-zero if we screw up
1748 * elsewhere in the driver.
1752 if (WARN_ON(!horiz_pixels))
1755 return DIV_ROUND_UP(pri_val * 64, horiz_pixels * cpp) + 2;
1758 struct ilk_wm_maximums {
1766 * For both WM_PIPE and WM_LP.
1767 * mem_value must be in 0.1us units.
1769 static uint32_t ilk_compute_pri_wm(const struct intel_crtc_state *cstate,
1770 const struct intel_plane_state *pstate,
1774 int cpp = pstate->base.fb ?
1775 drm_format_plane_cpp(pstate->base.fb->pixel_format, 0) : 0;
1776 uint32_t method1, method2;
1778 if (!cstate->base.active || !pstate->base.visible)
1781 method1 = ilk_wm_method1(ilk_pipe_pixel_rate(cstate), cpp, mem_value);
1786 method2 = ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1787 cstate->base.adjusted_mode.crtc_htotal,
1788 drm_rect_width(&pstate->base.dst),
1791 return min(method1, method2);
1795 * For both WM_PIPE and WM_LP.
1796 * mem_value must be in 0.1us units.
1798 static uint32_t ilk_compute_spr_wm(const struct intel_crtc_state *cstate,
1799 const struct intel_plane_state *pstate,
1802 int cpp = pstate->base.fb ?
1803 drm_format_plane_cpp(pstate->base.fb->pixel_format, 0) : 0;
1804 uint32_t method1, method2;
1806 if (!cstate->base.active || !pstate->base.visible)
1809 method1 = ilk_wm_method1(ilk_pipe_pixel_rate(cstate), cpp, mem_value);
1810 method2 = ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1811 cstate->base.adjusted_mode.crtc_htotal,
1812 drm_rect_width(&pstate->base.dst),
1814 return min(method1, method2);
1818 * For both WM_PIPE and WM_LP.
1819 * mem_value must be in 0.1us units.
1821 static uint32_t ilk_compute_cur_wm(const struct intel_crtc_state *cstate,
1822 const struct intel_plane_state *pstate,
1826 * We treat the cursor plane as always-on for the purposes of watermark
1827 * calculation. Until we have two-stage watermark programming merged,
1828 * this is necessary to avoid flickering.
1831 int width = pstate->base.visible ? pstate->base.crtc_w : 64;
1833 if (!cstate->base.active)
1836 return ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1837 cstate->base.adjusted_mode.crtc_htotal,
1838 width, cpp, mem_value);
1841 /* Only for WM_LP. */
1842 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state *cstate,
1843 const struct intel_plane_state *pstate,
1846 int cpp = pstate->base.fb ?
1847 drm_format_plane_cpp(pstate->base.fb->pixel_format, 0) : 0;
1849 if (!cstate->base.active || !pstate->base.visible)
1852 return ilk_wm_fbc(pri_val, drm_rect_width(&pstate->base.dst), cpp);
1855 static unsigned int ilk_display_fifo_size(const struct drm_device *dev)
1857 if (INTEL_INFO(dev)->gen >= 8)
1859 else if (INTEL_INFO(dev)->gen >= 7)
1865 static unsigned int ilk_plane_wm_reg_max(const struct drm_device *dev,
1866 int level, bool is_sprite)
1868 if (INTEL_INFO(dev)->gen >= 8)
1869 /* BDW primary/sprite plane watermarks */
1870 return level == 0 ? 255 : 2047;
1871 else if (INTEL_INFO(dev)->gen >= 7)
1872 /* IVB/HSW primary/sprite plane watermarks */
1873 return level == 0 ? 127 : 1023;
1874 else if (!is_sprite)
1875 /* ILK/SNB primary plane watermarks */
1876 return level == 0 ? 127 : 511;
1878 /* ILK/SNB sprite plane watermarks */
1879 return level == 0 ? 63 : 255;
1882 static unsigned int ilk_cursor_wm_reg_max(const struct drm_device *dev,
1885 if (INTEL_INFO(dev)->gen >= 7)
1886 return level == 0 ? 63 : 255;
1888 return level == 0 ? 31 : 63;
1891 static unsigned int ilk_fbc_wm_reg_max(const struct drm_device *dev)
1893 if (INTEL_INFO(dev)->gen >= 8)
1899 /* Calculate the maximum primary/sprite plane watermark */
1900 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
1902 const struct intel_wm_config *config,
1903 enum intel_ddb_partitioning ddb_partitioning,
1906 unsigned int fifo_size = ilk_display_fifo_size(dev);
1908 /* if sprites aren't enabled, sprites get nothing */
1909 if (is_sprite && !config->sprites_enabled)
1912 /* HSW allows LP1+ watermarks even with multiple pipes */
1913 if (level == 0 || config->num_pipes_active > 1) {
1914 fifo_size /= INTEL_INFO(dev)->num_pipes;
1917 * For some reason the non self refresh
1918 * FIFO size is only half of the self
1919 * refresh FIFO size on ILK/SNB.
1921 if (INTEL_INFO(dev)->gen <= 6)
1925 if (config->sprites_enabled) {
1926 /* level 0 is always calculated with 1:1 split */
1927 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
1936 /* clamp to max that the registers can hold */
1937 return min(fifo_size, ilk_plane_wm_reg_max(dev, level, is_sprite));
1940 /* Calculate the maximum cursor plane watermark */
1941 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
1943 const struct intel_wm_config *config)
1945 /* HSW LP1+ watermarks w/ multiple pipes */
1946 if (level > 0 && config->num_pipes_active > 1)
1949 /* otherwise just report max that registers can hold */
1950 return ilk_cursor_wm_reg_max(dev, level);
1953 static void ilk_compute_wm_maximums(const struct drm_device *dev,
1955 const struct intel_wm_config *config,
1956 enum intel_ddb_partitioning ddb_partitioning,
1957 struct ilk_wm_maximums *max)
1959 max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
1960 max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
1961 max->cur = ilk_cursor_wm_max(dev, level, config);
1962 max->fbc = ilk_fbc_wm_reg_max(dev);
1965 static void ilk_compute_wm_reg_maximums(struct drm_device *dev,
1967 struct ilk_wm_maximums *max)
1969 max->pri = ilk_plane_wm_reg_max(dev, level, false);
1970 max->spr = ilk_plane_wm_reg_max(dev, level, true);
1971 max->cur = ilk_cursor_wm_reg_max(dev, level);
1972 max->fbc = ilk_fbc_wm_reg_max(dev);
1975 static bool ilk_validate_wm_level(int level,
1976 const struct ilk_wm_maximums *max,
1977 struct intel_wm_level *result)
1981 /* already determined to be invalid? */
1982 if (!result->enable)
1985 result->enable = result->pri_val <= max->pri &&
1986 result->spr_val <= max->spr &&
1987 result->cur_val <= max->cur;
1989 ret = result->enable;
1992 * HACK until we can pre-compute everything,
1993 * and thus fail gracefully if LP0 watermarks
1996 if (level == 0 && !result->enable) {
1997 if (result->pri_val > max->pri)
1998 DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
1999 level, result->pri_val, max->pri);
2000 if (result->spr_val > max->spr)
2001 DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
2002 level, result->spr_val, max->spr);
2003 if (result->cur_val > max->cur)
2004 DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
2005 level, result->cur_val, max->cur);
2007 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
2008 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
2009 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
2010 result->enable = true;
2016 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
2017 const struct intel_crtc *intel_crtc,
2019 struct intel_crtc_state *cstate,
2020 struct intel_plane_state *pristate,
2021 struct intel_plane_state *sprstate,
2022 struct intel_plane_state *curstate,
2023 struct intel_wm_level *result)
2025 uint16_t pri_latency = dev_priv->wm.pri_latency[level];
2026 uint16_t spr_latency = dev_priv->wm.spr_latency[level];
2027 uint16_t cur_latency = dev_priv->wm.cur_latency[level];
2029 /* WM1+ latency values stored in 0.5us units */
2037 result->pri_val = ilk_compute_pri_wm(cstate, pristate,
2038 pri_latency, level);
2039 result->fbc_val = ilk_compute_fbc_wm(cstate, pristate, result->pri_val);
2043 result->spr_val = ilk_compute_spr_wm(cstate, sprstate, spr_latency);
2046 result->cur_val = ilk_compute_cur_wm(cstate, curstate, cur_latency);
2048 result->enable = true;
2052 hsw_compute_linetime_wm(const struct intel_crtc_state *cstate)
2054 const struct intel_atomic_state *intel_state =
2055 to_intel_atomic_state(cstate->base.state);
2056 const struct drm_display_mode *adjusted_mode =
2057 &cstate->base.adjusted_mode;
2058 u32 linetime, ips_linetime;
2060 if (!cstate->base.active)
2062 if (WARN_ON(adjusted_mode->crtc_clock == 0))
2064 if (WARN_ON(intel_state->cdclk == 0))
2067 /* The WM are computed with base on how long it takes to fill a single
2068 * row at the given clock rate, multiplied by 8.
2070 linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2071 adjusted_mode->crtc_clock);
2072 ips_linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2073 intel_state->cdclk);
2075 return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2076 PIPE_WM_LINETIME_TIME(linetime);
2079 static void intel_read_wm_latency(struct drm_device *dev, uint16_t wm[8])
2081 struct drm_i915_private *dev_priv = to_i915(dev);
2086 int level, max_level = ilk_wm_max_level(dev);
2088 /* read the first set of memory latencies[0:3] */
2089 val = 0; /* data0 to be programmed to 0 for first set */
2090 mutex_lock(&dev_priv->rps.hw_lock);
2091 ret = sandybridge_pcode_read(dev_priv,
2092 GEN9_PCODE_READ_MEM_LATENCY,
2094 mutex_unlock(&dev_priv->rps.hw_lock);
2097 DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2101 wm[0] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2102 wm[1] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2103 GEN9_MEM_LATENCY_LEVEL_MASK;
2104 wm[2] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2105 GEN9_MEM_LATENCY_LEVEL_MASK;
2106 wm[3] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2107 GEN9_MEM_LATENCY_LEVEL_MASK;
2109 /* read the second set of memory latencies[4:7] */
2110 val = 1; /* data0 to be programmed to 1 for second set */
2111 mutex_lock(&dev_priv->rps.hw_lock);
2112 ret = sandybridge_pcode_read(dev_priv,
2113 GEN9_PCODE_READ_MEM_LATENCY,
2115 mutex_unlock(&dev_priv->rps.hw_lock);
2117 DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2121 wm[4] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2122 wm[5] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2123 GEN9_MEM_LATENCY_LEVEL_MASK;
2124 wm[6] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2125 GEN9_MEM_LATENCY_LEVEL_MASK;
2126 wm[7] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2127 GEN9_MEM_LATENCY_LEVEL_MASK;
2130 * If a level n (n > 1) has a 0us latency, all levels m (m >= n)
2131 * need to be disabled. We make sure to sanitize the values out
2132 * of the punit to satisfy this requirement.
2134 for (level = 1; level <= max_level; level++) {
2135 if (wm[level] == 0) {
2136 for (i = level + 1; i <= max_level; i++)
2143 * WaWmMemoryReadLatency:skl
2145 * punit doesn't take into account the read latency so we need
2146 * to add 2us to the various latency levels we retrieve from the
2147 * punit when level 0 response data us 0us.
2151 for (level = 1; level <= max_level; level++) {
2158 } else if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2159 uint64_t sskpd = I915_READ64(MCH_SSKPD);
2161 wm[0] = (sskpd >> 56) & 0xFF;
2163 wm[0] = sskpd & 0xF;
2164 wm[1] = (sskpd >> 4) & 0xFF;
2165 wm[2] = (sskpd >> 12) & 0xFF;
2166 wm[3] = (sskpd >> 20) & 0x1FF;
2167 wm[4] = (sskpd >> 32) & 0x1FF;
2168 } else if (INTEL_INFO(dev)->gen >= 6) {
2169 uint32_t sskpd = I915_READ(MCH_SSKPD);
2171 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2172 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2173 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2174 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2175 } else if (INTEL_INFO(dev)->gen >= 5) {
2176 uint32_t mltr = I915_READ(MLTR_ILK);
2178 /* ILK primary LP0 latency is 700 ns */
2180 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2181 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2185 static void intel_fixup_spr_wm_latency(struct drm_device *dev, uint16_t wm[5])
2187 /* ILK sprite LP0 latency is 1300 ns */
2192 static void intel_fixup_cur_wm_latency(struct drm_device *dev, uint16_t wm[5])
2194 /* ILK cursor LP0 latency is 1300 ns */
2198 /* WaDoubleCursorLP3Latency:ivb */
2199 if (IS_IVYBRIDGE(dev))
2203 int ilk_wm_max_level(const struct drm_device *dev)
2205 /* how many WM levels are we expecting */
2206 if (INTEL_INFO(dev)->gen >= 9)
2208 else if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2210 else if (INTEL_INFO(dev)->gen >= 6)
2216 static void intel_print_wm_latency(struct drm_device *dev,
2218 const uint16_t wm[8])
2220 int level, max_level = ilk_wm_max_level(dev);
2222 for (level = 0; level <= max_level; level++) {
2223 unsigned int latency = wm[level];
2226 DRM_ERROR("%s WM%d latency not provided\n",
2232 * - latencies are in us on gen9.
2233 * - before then, WM1+ latency values are in 0.5us units
2240 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2241 name, level, wm[level],
2242 latency / 10, latency % 10);
2246 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
2247 uint16_t wm[5], uint16_t min)
2249 int level, max_level = ilk_wm_max_level(&dev_priv->drm);
2254 wm[0] = max(wm[0], min);
2255 for (level = 1; level <= max_level; level++)
2256 wm[level] = max_t(uint16_t, wm[level], DIV_ROUND_UP(min, 5));
2261 static void snb_wm_latency_quirk(struct drm_device *dev)
2263 struct drm_i915_private *dev_priv = to_i915(dev);
2267 * The BIOS provided WM memory latency values are often
2268 * inadequate for high resolution displays. Adjust them.
2270 changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
2271 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
2272 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
2277 DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2278 intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2279 intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2280 intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2283 static void ilk_setup_wm_latency(struct drm_device *dev)
2285 struct drm_i915_private *dev_priv = to_i915(dev);
2287 intel_read_wm_latency(dev, dev_priv->wm.pri_latency);
2289 memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2290 sizeof(dev_priv->wm.pri_latency));
2291 memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2292 sizeof(dev_priv->wm.pri_latency));
2294 intel_fixup_spr_wm_latency(dev, dev_priv->wm.spr_latency);
2295 intel_fixup_cur_wm_latency(dev, dev_priv->wm.cur_latency);
2297 intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2298 intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2299 intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2302 snb_wm_latency_quirk(dev);
2305 static void skl_setup_wm_latency(struct drm_device *dev)
2307 struct drm_i915_private *dev_priv = to_i915(dev);
2309 intel_read_wm_latency(dev, dev_priv->wm.skl_latency);
2310 intel_print_wm_latency(dev, "Gen9 Plane", dev_priv->wm.skl_latency);
2313 static bool ilk_validate_pipe_wm(struct drm_device *dev,
2314 struct intel_pipe_wm *pipe_wm)
2316 /* LP0 watermark maximums depend on this pipe alone */
2317 const struct intel_wm_config config = {
2318 .num_pipes_active = 1,
2319 .sprites_enabled = pipe_wm->sprites_enabled,
2320 .sprites_scaled = pipe_wm->sprites_scaled,
2322 struct ilk_wm_maximums max;
2324 /* LP0 watermarks always use 1/2 DDB partitioning */
2325 ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
2327 /* At least LP0 must be valid */
2328 if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0])) {
2329 DRM_DEBUG_KMS("LP0 watermark invalid\n");
2336 /* Compute new watermarks for the pipe */
2337 static int ilk_compute_pipe_wm(struct intel_crtc_state *cstate)
2339 struct drm_atomic_state *state = cstate->base.state;
2340 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
2341 struct intel_pipe_wm *pipe_wm;
2342 struct drm_device *dev = state->dev;
2343 const struct drm_i915_private *dev_priv = to_i915(dev);
2344 struct intel_plane *intel_plane;
2345 struct intel_plane_state *pristate = NULL;
2346 struct intel_plane_state *sprstate = NULL;
2347 struct intel_plane_state *curstate = NULL;
2348 int level, max_level = ilk_wm_max_level(dev), usable_level;
2349 struct ilk_wm_maximums max;
2351 pipe_wm = &cstate->wm.ilk.optimal;
2353 for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
2354 struct intel_plane_state *ps;
2356 ps = intel_atomic_get_existing_plane_state(state,
2361 if (intel_plane->base.type == DRM_PLANE_TYPE_PRIMARY)
2363 else if (intel_plane->base.type == DRM_PLANE_TYPE_OVERLAY)
2365 else if (intel_plane->base.type == DRM_PLANE_TYPE_CURSOR)
2369 pipe_wm->pipe_enabled = cstate->base.active;
2371 pipe_wm->sprites_enabled = sprstate->base.visible;
2372 pipe_wm->sprites_scaled = sprstate->base.visible &&
2373 (drm_rect_width(&sprstate->base.dst) != drm_rect_width(&sprstate->base.src) >> 16 ||
2374 drm_rect_height(&sprstate->base.dst) != drm_rect_height(&sprstate->base.src) >> 16);
2377 usable_level = max_level;
2379 /* ILK/SNB: LP2+ watermarks only w/o sprites */
2380 if (INTEL_INFO(dev)->gen <= 6 && pipe_wm->sprites_enabled)
2383 /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2384 if (pipe_wm->sprites_scaled)
2387 ilk_compute_wm_level(dev_priv, intel_crtc, 0, cstate,
2388 pristate, sprstate, curstate, &pipe_wm->raw_wm[0]);
2390 memset(&pipe_wm->wm, 0, sizeof(pipe_wm->wm));
2391 pipe_wm->wm[0] = pipe_wm->raw_wm[0];
2393 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2394 pipe_wm->linetime = hsw_compute_linetime_wm(cstate);
2396 if (!ilk_validate_pipe_wm(dev, pipe_wm))
2399 ilk_compute_wm_reg_maximums(dev, 1, &max);
2401 for (level = 1; level <= max_level; level++) {
2402 struct intel_wm_level *wm = &pipe_wm->raw_wm[level];
2404 ilk_compute_wm_level(dev_priv, intel_crtc, level, cstate,
2405 pristate, sprstate, curstate, wm);
2408 * Disable any watermark level that exceeds the
2409 * register maximums since such watermarks are
2412 if (level > usable_level)
2415 if (ilk_validate_wm_level(level, &max, wm))
2416 pipe_wm->wm[level] = *wm;
2418 usable_level = level;
2425 * Build a set of 'intermediate' watermark values that satisfy both the old
2426 * state and the new state. These can be programmed to the hardware
2429 static int ilk_compute_intermediate_wm(struct drm_device *dev,
2430 struct intel_crtc *intel_crtc,
2431 struct intel_crtc_state *newstate)
2433 struct intel_pipe_wm *a = &newstate->wm.ilk.intermediate;
2434 struct intel_pipe_wm *b = &intel_crtc->wm.active.ilk;
2435 int level, max_level = ilk_wm_max_level(dev);
2438 * Start with the final, target watermarks, then combine with the
2439 * currently active watermarks to get values that are safe both before
2440 * and after the vblank.
2442 *a = newstate->wm.ilk.optimal;
2443 a->pipe_enabled |= b->pipe_enabled;
2444 a->sprites_enabled |= b->sprites_enabled;
2445 a->sprites_scaled |= b->sprites_scaled;
2447 for (level = 0; level <= max_level; level++) {
2448 struct intel_wm_level *a_wm = &a->wm[level];
2449 const struct intel_wm_level *b_wm = &b->wm[level];
2451 a_wm->enable &= b_wm->enable;
2452 a_wm->pri_val = max(a_wm->pri_val, b_wm->pri_val);
2453 a_wm->spr_val = max(a_wm->spr_val, b_wm->spr_val);
2454 a_wm->cur_val = max(a_wm->cur_val, b_wm->cur_val);
2455 a_wm->fbc_val = max(a_wm->fbc_val, b_wm->fbc_val);
2459 * We need to make sure that these merged watermark values are
2460 * actually a valid configuration themselves. If they're not,
2461 * there's no safe way to transition from the old state to
2462 * the new state, so we need to fail the atomic transaction.
2464 if (!ilk_validate_pipe_wm(dev, a))
2468 * If our intermediate WM are identical to the final WM, then we can
2469 * omit the post-vblank programming; only update if it's different.
2471 if (memcmp(a, &newstate->wm.ilk.optimal, sizeof(*a)) == 0)
2472 newstate->wm.need_postvbl_update = false;
2478 * Merge the watermarks from all active pipes for a specific level.
2480 static void ilk_merge_wm_level(struct drm_device *dev,
2482 struct intel_wm_level *ret_wm)
2484 const struct intel_crtc *intel_crtc;
2486 ret_wm->enable = true;
2488 for_each_intel_crtc(dev, intel_crtc) {
2489 const struct intel_pipe_wm *active = &intel_crtc->wm.active.ilk;
2490 const struct intel_wm_level *wm = &active->wm[level];
2492 if (!active->pipe_enabled)
2496 * The watermark values may have been used in the past,
2497 * so we must maintain them in the registers for some
2498 * time even if the level is now disabled.
2501 ret_wm->enable = false;
2503 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
2504 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
2505 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
2506 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
2511 * Merge all low power watermarks for all active pipes.
2513 static void ilk_wm_merge(struct drm_device *dev,
2514 const struct intel_wm_config *config,
2515 const struct ilk_wm_maximums *max,
2516 struct intel_pipe_wm *merged)
2518 struct drm_i915_private *dev_priv = to_i915(dev);
2519 int level, max_level = ilk_wm_max_level(dev);
2520 int last_enabled_level = max_level;
2522 /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2523 if ((INTEL_INFO(dev)->gen <= 6 || IS_IVYBRIDGE(dev)) &&
2524 config->num_pipes_active > 1)
2525 last_enabled_level = 0;
2527 /* ILK: FBC WM must be disabled always */
2528 merged->fbc_wm_enabled = INTEL_INFO(dev)->gen >= 6;
2530 /* merge each WM1+ level */
2531 for (level = 1; level <= max_level; level++) {
2532 struct intel_wm_level *wm = &merged->wm[level];
2534 ilk_merge_wm_level(dev, level, wm);
2536 if (level > last_enabled_level)
2538 else if (!ilk_validate_wm_level(level, max, wm))
2539 /* make sure all following levels get disabled */
2540 last_enabled_level = level - 1;
2543 * The spec says it is preferred to disable
2544 * FBC WMs instead of disabling a WM level.
2546 if (wm->fbc_val > max->fbc) {
2548 merged->fbc_wm_enabled = false;
2553 /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2555 * FIXME this is racy. FBC might get enabled later.
2556 * What we should check here is whether FBC can be
2557 * enabled sometime later.
2559 if (IS_GEN5(dev) && !merged->fbc_wm_enabled &&
2560 intel_fbc_is_active(dev_priv)) {
2561 for (level = 2; level <= max_level; level++) {
2562 struct intel_wm_level *wm = &merged->wm[level];
2569 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
2571 /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2572 return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
2575 /* The value we need to program into the WM_LPx latency field */
2576 static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
2578 struct drm_i915_private *dev_priv = to_i915(dev);
2580 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2583 return dev_priv->wm.pri_latency[level];
2586 static void ilk_compute_wm_results(struct drm_device *dev,
2587 const struct intel_pipe_wm *merged,
2588 enum intel_ddb_partitioning partitioning,
2589 struct ilk_wm_values *results)
2591 struct intel_crtc *intel_crtc;
2594 results->enable_fbc_wm = merged->fbc_wm_enabled;
2595 results->partitioning = partitioning;
2597 /* LP1+ register values */
2598 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2599 const struct intel_wm_level *r;
2601 level = ilk_wm_lp_to_level(wm_lp, merged);
2603 r = &merged->wm[level];
2606 * Maintain the watermark values even if the level is
2607 * disabled. Doing otherwise could cause underruns.
2609 results->wm_lp[wm_lp - 1] =
2610 (ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
2611 (r->pri_val << WM1_LP_SR_SHIFT) |
2615 results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
2617 if (INTEL_INFO(dev)->gen >= 8)
2618 results->wm_lp[wm_lp - 1] |=
2619 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
2621 results->wm_lp[wm_lp - 1] |=
2622 r->fbc_val << WM1_LP_FBC_SHIFT;
2625 * Always set WM1S_LP_EN when spr_val != 0, even if the
2626 * level is disabled. Doing otherwise could cause underruns.
2628 if (INTEL_INFO(dev)->gen <= 6 && r->spr_val) {
2629 WARN_ON(wm_lp != 1);
2630 results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
2632 results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2635 /* LP0 register values */
2636 for_each_intel_crtc(dev, intel_crtc) {
2637 enum pipe pipe = intel_crtc->pipe;
2638 const struct intel_wm_level *r =
2639 &intel_crtc->wm.active.ilk.wm[0];
2641 if (WARN_ON(!r->enable))
2644 results->wm_linetime[pipe] = intel_crtc->wm.active.ilk.linetime;
2646 results->wm_pipe[pipe] =
2647 (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
2648 (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
2653 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2654 * case both are at the same level. Prefer r1 in case they're the same. */
2655 static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
2656 struct intel_pipe_wm *r1,
2657 struct intel_pipe_wm *r2)
2659 int level, max_level = ilk_wm_max_level(dev);
2660 int level1 = 0, level2 = 0;
2662 for (level = 1; level <= max_level; level++) {
2663 if (r1->wm[level].enable)
2665 if (r2->wm[level].enable)
2669 if (level1 == level2) {
2670 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
2674 } else if (level1 > level2) {
2681 /* dirty bits used to track which watermarks need changes */
2682 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2683 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2684 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2685 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2686 #define WM_DIRTY_FBC (1 << 24)
2687 #define WM_DIRTY_DDB (1 << 25)
2689 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private *dev_priv,
2690 const struct ilk_wm_values *old,
2691 const struct ilk_wm_values *new)
2693 unsigned int dirty = 0;
2697 for_each_pipe(dev_priv, pipe) {
2698 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
2699 dirty |= WM_DIRTY_LINETIME(pipe);
2700 /* Must disable LP1+ watermarks too */
2701 dirty |= WM_DIRTY_LP_ALL;
2704 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
2705 dirty |= WM_DIRTY_PIPE(pipe);
2706 /* Must disable LP1+ watermarks too */
2707 dirty |= WM_DIRTY_LP_ALL;
2711 if (old->enable_fbc_wm != new->enable_fbc_wm) {
2712 dirty |= WM_DIRTY_FBC;
2713 /* Must disable LP1+ watermarks too */
2714 dirty |= WM_DIRTY_LP_ALL;
2717 if (old->partitioning != new->partitioning) {
2718 dirty |= WM_DIRTY_DDB;
2719 /* Must disable LP1+ watermarks too */
2720 dirty |= WM_DIRTY_LP_ALL;
2723 /* LP1+ watermarks already deemed dirty, no need to continue */
2724 if (dirty & WM_DIRTY_LP_ALL)
2727 /* Find the lowest numbered LP1+ watermark in need of an update... */
2728 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2729 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
2730 old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
2734 /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2735 for (; wm_lp <= 3; wm_lp++)
2736 dirty |= WM_DIRTY_LP(wm_lp);
2741 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
2744 struct ilk_wm_values *previous = &dev_priv->wm.hw;
2745 bool changed = false;
2747 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
2748 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
2749 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
2752 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
2753 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
2754 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
2757 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
2758 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
2759 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
2764 * Don't touch WM1S_LP_EN here.
2765 * Doing so could cause underruns.
2772 * The spec says we shouldn't write when we don't need, because every write
2773 * causes WMs to be re-evaluated, expending some power.
2775 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
2776 struct ilk_wm_values *results)
2778 struct drm_device *dev = &dev_priv->drm;
2779 struct ilk_wm_values *previous = &dev_priv->wm.hw;
2783 dirty = ilk_compute_wm_dirty(dev_priv, previous, results);
2787 _ilk_disable_lp_wm(dev_priv, dirty);
2789 if (dirty & WM_DIRTY_PIPE(PIPE_A))
2790 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2791 if (dirty & WM_DIRTY_PIPE(PIPE_B))
2792 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
2793 if (dirty & WM_DIRTY_PIPE(PIPE_C))
2794 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
2796 if (dirty & WM_DIRTY_LINETIME(PIPE_A))
2797 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
2798 if (dirty & WM_DIRTY_LINETIME(PIPE_B))
2799 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
2800 if (dirty & WM_DIRTY_LINETIME(PIPE_C))
2801 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
2803 if (dirty & WM_DIRTY_DDB) {
2804 if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2805 val = I915_READ(WM_MISC);
2806 if (results->partitioning == INTEL_DDB_PART_1_2)
2807 val &= ~WM_MISC_DATA_PARTITION_5_6;
2809 val |= WM_MISC_DATA_PARTITION_5_6;
2810 I915_WRITE(WM_MISC, val);
2812 val = I915_READ(DISP_ARB_CTL2);
2813 if (results->partitioning == INTEL_DDB_PART_1_2)
2814 val &= ~DISP_DATA_PARTITION_5_6;
2816 val |= DISP_DATA_PARTITION_5_6;
2817 I915_WRITE(DISP_ARB_CTL2, val);
2821 if (dirty & WM_DIRTY_FBC) {
2822 val = I915_READ(DISP_ARB_CTL);
2823 if (results->enable_fbc_wm)
2824 val &= ~DISP_FBC_WM_DIS;
2826 val |= DISP_FBC_WM_DIS;
2827 I915_WRITE(DISP_ARB_CTL, val);
2830 if (dirty & WM_DIRTY_LP(1) &&
2831 previous->wm_lp_spr[0] != results->wm_lp_spr[0])
2832 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
2834 if (INTEL_INFO(dev)->gen >= 7) {
2835 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
2836 I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
2837 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
2838 I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
2841 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
2842 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
2843 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
2844 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
2845 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
2846 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
2848 dev_priv->wm.hw = *results;
2851 bool ilk_disable_lp_wm(struct drm_device *dev)
2853 struct drm_i915_private *dev_priv = to_i915(dev);
2855 return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
2858 #define SKL_SAGV_BLOCK_TIME 30 /* µs */
2861 * Return the index of a plane in the SKL DDB and wm result arrays. Primary
2862 * plane is always in slot 0, cursor is always in slot I915_MAX_PLANES-1, and
2863 * other universal planes are in indices 1..n. Note that this may leave unused
2864 * indices between the top "sprite" plane and the cursor.
2867 skl_wm_plane_id(const struct intel_plane *plane)
2869 switch (plane->base.type) {
2870 case DRM_PLANE_TYPE_PRIMARY:
2872 case DRM_PLANE_TYPE_CURSOR:
2873 return PLANE_CURSOR;
2874 case DRM_PLANE_TYPE_OVERLAY:
2875 return plane->plane + 1;
2877 MISSING_CASE(plane->base.type);
2878 return plane->plane;
2883 * FIXME: We still don't have the proper code detect if we need to apply the WA,
2884 * so assume we'll always need it in order to avoid underruns.
2886 static bool skl_needs_memory_bw_wa(struct intel_atomic_state *state)
2888 struct drm_i915_private *dev_priv = to_i915(state->base.dev);
2890 if (IS_SKYLAKE(dev_priv) || IS_BROXTON(dev_priv) ||
2891 IS_KABYLAKE(dev_priv))
2898 intel_has_sagv(struct drm_i915_private *dev_priv)
2900 if (IS_KABYLAKE(dev_priv))
2903 if (IS_SKYLAKE(dev_priv) &&
2904 dev_priv->sagv_status != I915_SAGV_NOT_CONTROLLED)
2911 * SAGV dynamically adjusts the system agent voltage and clock frequencies
2912 * depending on power and performance requirements. The display engine access
2913 * to system memory is blocked during the adjustment time. Because of the
2914 * blocking time, having this enabled can cause full system hangs and/or pipe
2915 * underruns if we don't meet all of the following requirements:
2917 * - <= 1 pipe enabled
2918 * - All planes can enable watermarks for latencies >= SAGV engine block time
2919 * - We're not using an interlaced display configuration
2922 intel_enable_sagv(struct drm_i915_private *dev_priv)
2926 if (!intel_has_sagv(dev_priv))
2929 if (dev_priv->sagv_status == I915_SAGV_ENABLED)
2932 DRM_DEBUG_KMS("Enabling the SAGV\n");
2933 mutex_lock(&dev_priv->rps.hw_lock);
2935 ret = sandybridge_pcode_write(dev_priv, GEN9_PCODE_SAGV_CONTROL,
2938 /* We don't need to wait for the SAGV when enabling */
2939 mutex_unlock(&dev_priv->rps.hw_lock);
2942 * Some skl systems, pre-release machines in particular,
2943 * don't actually have an SAGV.
2945 if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
2946 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
2947 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
2949 } else if (ret < 0) {
2950 DRM_ERROR("Failed to enable the SAGV\n");
2954 dev_priv->sagv_status = I915_SAGV_ENABLED;
2959 intel_disable_sagv(struct drm_i915_private *dev_priv)
2963 if (!intel_has_sagv(dev_priv))
2966 if (dev_priv->sagv_status == I915_SAGV_DISABLED)
2969 DRM_DEBUG_KMS("Disabling the SAGV\n");
2970 mutex_lock(&dev_priv->rps.hw_lock);
2972 /* bspec says to keep retrying for at least 1 ms */
2973 ret = skl_pcode_request(dev_priv, GEN9_PCODE_SAGV_CONTROL,
2975 GEN9_SAGV_IS_DISABLED, GEN9_SAGV_IS_DISABLED,
2977 mutex_unlock(&dev_priv->rps.hw_lock);
2980 * Some skl systems, pre-release machines in particular,
2981 * don't actually have an SAGV.
2983 if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
2984 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
2985 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
2987 } else if (ret < 0) {
2988 DRM_ERROR("Failed to disable the SAGV (%d)\n", ret);
2992 dev_priv->sagv_status = I915_SAGV_DISABLED;
2996 bool intel_can_enable_sagv(struct drm_atomic_state *state)
2998 struct drm_device *dev = state->dev;
2999 struct drm_i915_private *dev_priv = to_i915(dev);
3000 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3001 struct intel_crtc *crtc;
3002 struct intel_plane *plane;
3004 int level, id, latency;
3006 if (!intel_has_sagv(dev_priv))
3010 * SKL workaround: bspec recommends we disable the SAGV when we have
3011 * more then one pipe enabled
3013 * If there are no active CRTCs, no additional checks need be performed
3015 if (hweight32(intel_state->active_crtcs) == 0)
3017 else if (hweight32(intel_state->active_crtcs) > 1)
3020 /* Since we're now guaranteed to only have one active CRTC... */
3021 pipe = ffs(intel_state->active_crtcs) - 1;
3022 crtc = to_intel_crtc(dev_priv->pipe_to_crtc_mapping[pipe]);
3024 if (crtc->base.state->mode.flags & DRM_MODE_FLAG_INTERLACE)
3027 for_each_intel_plane_on_crtc(dev, crtc, plane) {
3028 id = skl_wm_plane_id(plane);
3030 /* Skip this plane if it's not enabled */
3031 if (intel_state->wm_results.plane[pipe][id][0] == 0)
3034 /* Find the highest enabled wm level for this plane */
3035 for (level = ilk_wm_max_level(dev);
3036 intel_state->wm_results.plane[pipe][id][level] == 0; --level)
3039 latency = dev_priv->wm.skl_latency[level];
3041 if (skl_needs_memory_bw_wa(intel_state) &&
3042 plane->base.state->fb->modifier[0] ==
3043 I915_FORMAT_MOD_X_TILED)
3047 * If any of the planes on this pipe don't enable wm levels
3048 * that incur memory latencies higher then 30µs we can't enable
3051 if (latency < SKL_SAGV_BLOCK_TIME)
3059 skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
3060 const struct intel_crtc_state *cstate,
3061 struct skl_ddb_entry *alloc, /* out */
3062 int *num_active /* out */)
3064 struct drm_atomic_state *state = cstate->base.state;
3065 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3066 struct drm_i915_private *dev_priv = to_i915(dev);
3067 struct drm_crtc *for_crtc = cstate->base.crtc;
3068 unsigned int pipe_size, ddb_size;
3069 int nth_active_pipe;
3070 int pipe = to_intel_crtc(for_crtc)->pipe;
3072 if (WARN_ON(!state) || !cstate->base.active) {
3075 *num_active = hweight32(dev_priv->active_crtcs);
3079 if (intel_state->active_pipe_changes)
3080 *num_active = hweight32(intel_state->active_crtcs);
3082 *num_active = hweight32(dev_priv->active_crtcs);
3084 ddb_size = INTEL_INFO(dev_priv)->ddb_size;
3085 WARN_ON(ddb_size == 0);
3087 ddb_size -= 4; /* 4 blocks for bypass path allocation */
3090 * If the state doesn't change the active CRTC's, then there's
3091 * no need to recalculate; the existing pipe allocation limits
3092 * should remain unchanged. Note that we're safe from racing
3093 * commits since any racing commit that changes the active CRTC
3094 * list would need to grab _all_ crtc locks, including the one
3095 * we currently hold.
3097 if (!intel_state->active_pipe_changes) {
3098 *alloc = dev_priv->wm.skl_hw.ddb.pipe[pipe];
3102 nth_active_pipe = hweight32(intel_state->active_crtcs &
3103 (drm_crtc_mask(for_crtc) - 1));
3104 pipe_size = ddb_size / hweight32(intel_state->active_crtcs);
3105 alloc->start = nth_active_pipe * ddb_size / *num_active;
3106 alloc->end = alloc->start + pipe_size;
3109 static unsigned int skl_cursor_allocation(int num_active)
3111 if (num_active == 1)
3117 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry *entry, u32 reg)
3119 entry->start = reg & 0x3ff;
3120 entry->end = (reg >> 16) & 0x3ff;
3125 void skl_ddb_get_hw_state(struct drm_i915_private *dev_priv,
3126 struct skl_ddb_allocation *ddb /* out */)
3132 memset(ddb, 0, sizeof(*ddb));
3134 for_each_pipe(dev_priv, pipe) {
3135 enum intel_display_power_domain power_domain;
3137 power_domain = POWER_DOMAIN_PIPE(pipe);
3138 if (!intel_display_power_get_if_enabled(dev_priv, power_domain))
3141 for_each_plane(dev_priv, pipe, plane) {
3142 val = I915_READ(PLANE_BUF_CFG(pipe, plane));
3143 skl_ddb_entry_init_from_hw(&ddb->plane[pipe][plane],
3147 val = I915_READ(CUR_BUF_CFG(pipe));
3148 skl_ddb_entry_init_from_hw(&ddb->plane[pipe][PLANE_CURSOR],
3151 intel_display_power_put(dev_priv, power_domain);
3156 * Determines the downscale amount of a plane for the purposes of watermark calculations.
3157 * The bspec defines downscale amount as:
3160 * Horizontal down scale amount = maximum[1, Horizontal source size /
3161 * Horizontal destination size]
3162 * Vertical down scale amount = maximum[1, Vertical source size /
3163 * Vertical destination size]
3164 * Total down scale amount = Horizontal down scale amount *
3165 * Vertical down scale amount
3168 * Return value is provided in 16.16 fixed point form to retain fractional part.
3169 * Caller should take care of dividing & rounding off the value.
3172 skl_plane_downscale_amount(const struct intel_plane_state *pstate)
3174 uint32_t downscale_h, downscale_w;
3175 uint32_t src_w, src_h, dst_w, dst_h;
3177 if (WARN_ON(!pstate->base.visible))
3178 return DRM_PLANE_HELPER_NO_SCALING;
3180 /* n.b., src is 16.16 fixed point, dst is whole integer */
3181 src_w = drm_rect_width(&pstate->base.src);
3182 src_h = drm_rect_height(&pstate->base.src);
3183 dst_w = drm_rect_width(&pstate->base.dst);
3184 dst_h = drm_rect_height(&pstate->base.dst);
3185 if (intel_rotation_90_or_270(pstate->base.rotation))
3188 downscale_h = max(src_h / dst_h, (uint32_t)DRM_PLANE_HELPER_NO_SCALING);
3189 downscale_w = max(src_w / dst_w, (uint32_t)DRM_PLANE_HELPER_NO_SCALING);
3191 /* Provide result in 16.16 fixed point */
3192 return (uint64_t)downscale_w * downscale_h >> 16;
3196 skl_plane_relative_data_rate(const struct intel_crtc_state *cstate,
3197 const struct drm_plane_state *pstate,
3200 struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
3201 struct drm_framebuffer *fb = pstate->fb;
3202 uint32_t down_scale_amount, data_rate;
3203 uint32_t width = 0, height = 0;
3204 unsigned format = fb ? fb->pixel_format : DRM_FORMAT_XRGB8888;
3206 if (!intel_pstate->base.visible)
3208 if (pstate->plane->type == DRM_PLANE_TYPE_CURSOR)
3210 if (y && format != DRM_FORMAT_NV12)
3213 width = drm_rect_width(&intel_pstate->base.src) >> 16;
3214 height = drm_rect_height(&intel_pstate->base.src) >> 16;
3216 if (intel_rotation_90_or_270(pstate->rotation))
3217 swap(width, height);
3219 /* for planar format */
3220 if (format == DRM_FORMAT_NV12) {
3221 if (y) /* y-plane data rate */
3222 data_rate = width * height *
3223 drm_format_plane_cpp(format, 0);
3224 else /* uv-plane data rate */
3225 data_rate = (width / 2) * (height / 2) *
3226 drm_format_plane_cpp(format, 1);
3228 /* for packed formats */
3229 data_rate = width * height * drm_format_plane_cpp(format, 0);
3232 down_scale_amount = skl_plane_downscale_amount(intel_pstate);
3234 return (uint64_t)data_rate * down_scale_amount >> 16;
3238 * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
3239 * a 8192x4096@32bpp framebuffer:
3240 * 3 * 4096 * 8192 * 4 < 2^32
3243 skl_get_total_relative_data_rate(struct intel_crtc_state *intel_cstate)
3245 struct drm_crtc_state *cstate = &intel_cstate->base;
3246 struct drm_atomic_state *state = cstate->state;
3247 struct drm_crtc *crtc = cstate->crtc;
3248 struct drm_device *dev = crtc->dev;
3249 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3250 const struct drm_plane *plane;
3251 const struct intel_plane *intel_plane;
3252 struct drm_plane_state *pstate;
3253 unsigned int rate, total_data_rate = 0;
3257 if (WARN_ON(!state))
3260 /* Calculate and cache data rate for each plane */
3261 for_each_plane_in_state(state, plane, pstate, i) {
3262 id = skl_wm_plane_id(to_intel_plane(plane));
3263 intel_plane = to_intel_plane(plane);
3265 if (intel_plane->pipe != intel_crtc->pipe)
3269 rate = skl_plane_relative_data_rate(intel_cstate,
3271 intel_cstate->wm.skl.plane_data_rate[id] = rate;
3274 rate = skl_plane_relative_data_rate(intel_cstate,
3276 intel_cstate->wm.skl.plane_y_data_rate[id] = rate;
3279 /* Calculate CRTC's total data rate from cached values */
3280 for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
3281 int id = skl_wm_plane_id(intel_plane);
3284 total_data_rate += intel_cstate->wm.skl.plane_data_rate[id];
3285 total_data_rate += intel_cstate->wm.skl.plane_y_data_rate[id];
3288 return total_data_rate;
3292 skl_ddb_min_alloc(const struct drm_plane_state *pstate,
3295 struct drm_framebuffer *fb = pstate->fb;
3296 struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
3297 uint32_t src_w, src_h;
3298 uint32_t min_scanlines = 8;
3304 /* For packed formats, no y-plane, return 0 */
3305 if (y && fb->pixel_format != DRM_FORMAT_NV12)
3308 /* For Non Y-tile return 8-blocks */
3309 if (fb->modifier[0] != I915_FORMAT_MOD_Y_TILED &&
3310 fb->modifier[0] != I915_FORMAT_MOD_Yf_TILED)
3313 src_w = drm_rect_width(&intel_pstate->base.src) >> 16;
3314 src_h = drm_rect_height(&intel_pstate->base.src) >> 16;
3316 if (intel_rotation_90_or_270(pstate->rotation))
3319 /* Halve UV plane width and height for NV12 */
3320 if (fb->pixel_format == DRM_FORMAT_NV12 && !y) {
3325 if (fb->pixel_format == DRM_FORMAT_NV12 && !y)
3326 plane_bpp = drm_format_plane_cpp(fb->pixel_format, 1);
3328 plane_bpp = drm_format_plane_cpp(fb->pixel_format, 0);
3330 if (intel_rotation_90_or_270(pstate->rotation)) {
3331 switch (plane_bpp) {
3345 WARN(1, "Unsupported pixel depth %u for rotation",
3351 return DIV_ROUND_UP((4 * src_w * plane_bpp), 512) * min_scanlines/4 + 3;
3355 skl_allocate_pipe_ddb(struct intel_crtc_state *cstate,
3356 struct skl_ddb_allocation *ddb /* out */)
3358 struct drm_atomic_state *state = cstate->base.state;
3359 struct drm_crtc *crtc = cstate->base.crtc;
3360 struct drm_device *dev = crtc->dev;
3361 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3362 struct intel_plane *intel_plane;
3363 struct drm_plane *plane;
3364 struct drm_plane_state *pstate;
3365 enum pipe pipe = intel_crtc->pipe;
3366 struct skl_ddb_entry *alloc = &ddb->pipe[pipe];
3367 uint16_t alloc_size, start, cursor_blocks;
3368 uint16_t *minimum = cstate->wm.skl.minimum_blocks;
3369 uint16_t *y_minimum = cstate->wm.skl.minimum_y_blocks;
3370 unsigned int total_data_rate;
3374 /* Clear the partitioning for disabled planes. */
3375 memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
3376 memset(ddb->y_plane[pipe], 0, sizeof(ddb->y_plane[pipe]));
3378 if (WARN_ON(!state))
3381 if (!cstate->base.active) {
3382 ddb->pipe[pipe].start = ddb->pipe[pipe].end = 0;
3386 skl_ddb_get_pipe_allocation_limits(dev, cstate, alloc, &num_active);
3387 alloc_size = skl_ddb_entry_size(alloc);
3388 if (alloc_size == 0) {
3389 memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
3393 cursor_blocks = skl_cursor_allocation(num_active);
3394 ddb->plane[pipe][PLANE_CURSOR].start = alloc->end - cursor_blocks;
3395 ddb->plane[pipe][PLANE_CURSOR].end = alloc->end;
3397 alloc_size -= cursor_blocks;
3399 /* 1. Allocate the mininum required blocks for each active plane */
3400 for_each_plane_in_state(state, plane, pstate, i) {
3401 intel_plane = to_intel_plane(plane);
3402 id = skl_wm_plane_id(intel_plane);
3404 if (intel_plane->pipe != pipe)
3407 if (!to_intel_plane_state(pstate)->base.visible) {
3412 if (plane->type == DRM_PLANE_TYPE_CURSOR) {
3418 minimum[id] = skl_ddb_min_alloc(pstate, 0);
3419 y_minimum[id] = skl_ddb_min_alloc(pstate, 1);
3422 for (i = 0; i < PLANE_CURSOR; i++) {
3423 alloc_size -= minimum[i];
3424 alloc_size -= y_minimum[i];
3428 * 2. Distribute the remaining space in proportion to the amount of
3429 * data each plane needs to fetch from memory.
3431 * FIXME: we may not allocate every single block here.
3433 total_data_rate = skl_get_total_relative_data_rate(cstate);
3434 if (total_data_rate == 0)
3437 start = alloc->start;
3438 for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
3439 unsigned int data_rate, y_data_rate;
3440 uint16_t plane_blocks, y_plane_blocks = 0;
3441 int id = skl_wm_plane_id(intel_plane);
3443 data_rate = cstate->wm.skl.plane_data_rate[id];
3446 * allocation for (packed formats) or (uv-plane part of planar format):
3447 * promote the expression to 64 bits to avoid overflowing, the
3448 * result is < available as data_rate / total_data_rate < 1
3450 plane_blocks = minimum[id];
3451 plane_blocks += div_u64((uint64_t)alloc_size * data_rate,
3454 /* Leave disabled planes at (0,0) */
3456 ddb->plane[pipe][id].start = start;
3457 ddb->plane[pipe][id].end = start + plane_blocks;
3460 start += plane_blocks;
3463 * allocation for y_plane part of planar format:
3465 y_data_rate = cstate->wm.skl.plane_y_data_rate[id];
3467 y_plane_blocks = y_minimum[id];
3468 y_plane_blocks += div_u64((uint64_t)alloc_size * y_data_rate,
3472 ddb->y_plane[pipe][id].start = start;
3473 ddb->y_plane[pipe][id].end = start + y_plane_blocks;
3476 start += y_plane_blocks;
3483 * The max latency should be 257 (max the punit can code is 255 and we add 2us
3484 * for the read latency) and cpp should always be <= 8, so that
3485 * should allow pixel_rate up to ~2 GHz which seems sufficient since max
3486 * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
3488 static uint32_t skl_wm_method1(uint32_t pixel_rate, uint8_t cpp, uint32_t latency)
3490 uint32_t wm_intermediate_val, ret;
3495 wm_intermediate_val = latency * pixel_rate * cpp / 512;
3496 ret = DIV_ROUND_UP(wm_intermediate_val, 1000);
3501 static uint32_t skl_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
3502 uint32_t latency, uint32_t plane_blocks_per_line)
3505 uint32_t wm_intermediate_val;
3510 wm_intermediate_val = latency * pixel_rate;
3511 ret = DIV_ROUND_UP(wm_intermediate_val, pipe_htotal * 1000) *
3512 plane_blocks_per_line;
3517 static uint32_t skl_adjusted_plane_pixel_rate(const struct intel_crtc_state *cstate,
3518 struct intel_plane_state *pstate)
3520 uint64_t adjusted_pixel_rate;
3521 uint64_t downscale_amount;
3522 uint64_t pixel_rate;
3524 /* Shouldn't reach here on disabled planes... */
3525 if (WARN_ON(!pstate->base.visible))
3529 * Adjusted plane pixel rate is just the pipe's adjusted pixel rate
3530 * with additional adjustments for plane-specific scaling.
3532 adjusted_pixel_rate = ilk_pipe_pixel_rate(cstate);
3533 downscale_amount = skl_plane_downscale_amount(pstate);
3535 pixel_rate = adjusted_pixel_rate * downscale_amount >> 16;
3536 WARN_ON(pixel_rate != clamp_t(uint32_t, pixel_rate, 0, ~0));
3541 static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv,
3542 struct intel_crtc_state *cstate,
3543 struct intel_plane_state *intel_pstate,
3544 uint16_t ddb_allocation,
3546 uint16_t *out_blocks, /* out */
3547 uint8_t *out_lines, /* out */
3548 bool *enabled /* out */)
3550 struct drm_plane_state *pstate = &intel_pstate->base;
3551 struct drm_framebuffer *fb = pstate->fb;
3552 uint32_t latency = dev_priv->wm.skl_latency[level];
3553 uint32_t method1, method2;
3554 uint32_t plane_bytes_per_line, plane_blocks_per_line;
3555 uint32_t res_blocks, res_lines;
3556 uint32_t selected_result;
3558 uint32_t width = 0, height = 0;
3559 uint32_t plane_pixel_rate;
3560 uint32_t y_tile_minimum, y_min_scanlines;
3561 struct intel_atomic_state *state =
3562 to_intel_atomic_state(cstate->base.state);
3563 bool apply_memory_bw_wa = skl_needs_memory_bw_wa(state);
3565 if (latency == 0 || !cstate->base.active || !intel_pstate->base.visible) {
3570 if (apply_memory_bw_wa && fb->modifier[0] == I915_FORMAT_MOD_X_TILED)
3573 width = drm_rect_width(&intel_pstate->base.src) >> 16;
3574 height = drm_rect_height(&intel_pstate->base.src) >> 16;
3576 if (intel_rotation_90_or_270(pstate->rotation))
3577 swap(width, height);
3579 cpp = drm_format_plane_cpp(fb->pixel_format, 0);
3580 plane_pixel_rate = skl_adjusted_plane_pixel_rate(cstate, intel_pstate);
3582 if (intel_rotation_90_or_270(pstate->rotation)) {
3583 int cpp = (fb->pixel_format == DRM_FORMAT_NV12) ?
3584 drm_format_plane_cpp(fb->pixel_format, 1) :
3585 drm_format_plane_cpp(fb->pixel_format, 0);
3589 y_min_scanlines = 16;
3592 y_min_scanlines = 8;
3595 WARN(1, "Unsupported pixel depth for rotation");
3597 y_min_scanlines = 4;
3601 y_min_scanlines = 4;
3604 if (apply_memory_bw_wa)
3605 y_min_scanlines *= 2;
3607 plane_bytes_per_line = width * cpp;
3608 if (fb->modifier[0] == I915_FORMAT_MOD_Y_TILED ||
3609 fb->modifier[0] == I915_FORMAT_MOD_Yf_TILED) {
3610 plane_blocks_per_line =
3611 DIV_ROUND_UP(plane_bytes_per_line * y_min_scanlines, 512);
3612 plane_blocks_per_line /= y_min_scanlines;
3613 } else if (fb->modifier[0] == DRM_FORMAT_MOD_NONE) {
3614 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512)
3617 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
3620 method1 = skl_wm_method1(plane_pixel_rate, cpp, latency);
3621 method2 = skl_wm_method2(plane_pixel_rate,
3622 cstate->base.adjusted_mode.crtc_htotal,
3624 plane_blocks_per_line);
3626 y_tile_minimum = plane_blocks_per_line * y_min_scanlines;
3628 if (fb->modifier[0] == I915_FORMAT_MOD_Y_TILED ||
3629 fb->modifier[0] == I915_FORMAT_MOD_Yf_TILED) {
3630 selected_result = max(method2, y_tile_minimum);
3632 if ((ddb_allocation / plane_blocks_per_line) >= 1)
3633 selected_result = min(method1, method2);
3635 selected_result = method1;
3638 res_blocks = selected_result + 1;
3639 res_lines = DIV_ROUND_UP(selected_result, plane_blocks_per_line);
3641 if (level >= 1 && level <= 7) {
3642 if (fb->modifier[0] == I915_FORMAT_MOD_Y_TILED ||
3643 fb->modifier[0] == I915_FORMAT_MOD_Yf_TILED) {
3644 res_blocks += y_tile_minimum;
3645 res_lines += y_min_scanlines;
3651 if (res_blocks >= ddb_allocation || res_lines > 31) {
3655 * If there are no valid level 0 watermarks, then we can't
3656 * support this display configuration.
3661 DRM_DEBUG_KMS("Requested display configuration exceeds system watermark limitations\n");
3662 DRM_DEBUG_KMS("Plane %d.%d: blocks required = %u/%u, lines required = %u/31\n",
3663 to_intel_crtc(cstate->base.crtc)->pipe,
3664 skl_wm_plane_id(to_intel_plane(pstate->plane)),
3665 res_blocks, ddb_allocation, res_lines);
3671 *out_blocks = res_blocks;
3672 *out_lines = res_lines;
3679 skl_compute_wm_level(const struct drm_i915_private *dev_priv,
3680 struct skl_ddb_allocation *ddb,
3681 struct intel_crtc_state *cstate,
3683 struct skl_wm_level *result)
3685 struct drm_atomic_state *state = cstate->base.state;
3686 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
3687 struct drm_plane *plane;
3688 struct intel_plane *intel_plane;
3689 struct intel_plane_state *intel_pstate;
3690 uint16_t ddb_blocks;
3691 enum pipe pipe = intel_crtc->pipe;
3695 * We'll only calculate watermarks for planes that are actually
3696 * enabled, so make sure all other planes are set as disabled.
3698 memset(result, 0, sizeof(*result));
3700 for_each_intel_plane_mask(&dev_priv->drm,
3702 cstate->base.plane_mask) {
3703 int i = skl_wm_plane_id(intel_plane);
3705 plane = &intel_plane->base;
3706 intel_pstate = NULL;
3709 intel_atomic_get_existing_plane_state(state,
3713 * Note: If we start supporting multiple pending atomic commits
3714 * against the same planes/CRTC's in the future, plane->state
3715 * will no longer be the correct pre-state to use for the
3716 * calculations here and we'll need to change where we get the
3717 * 'unchanged' plane data from.
3719 * For now this is fine because we only allow one queued commit
3720 * against a CRTC. Even if the plane isn't modified by this
3721 * transaction and we don't have a plane lock, we still have
3722 * the CRTC's lock, so we know that no other transactions are
3723 * racing with us to update it.
3726 intel_pstate = to_intel_plane_state(plane->state);
3728 WARN_ON(!intel_pstate->base.fb);
3730 ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][i]);
3732 ret = skl_compute_plane_wm(dev_priv,
3737 &result->plane_res_b[i],
3738 &result->plane_res_l[i],
3739 &result->plane_en[i]);
3748 skl_compute_linetime_wm(struct intel_crtc_state *cstate)
3750 if (!cstate->base.active)
3753 if (WARN_ON(ilk_pipe_pixel_rate(cstate) == 0))
3756 return DIV_ROUND_UP(8 * cstate->base.adjusted_mode.crtc_htotal * 1000,
3757 ilk_pipe_pixel_rate(cstate));
3760 static void skl_compute_transition_wm(struct intel_crtc_state *cstate,
3761 struct skl_wm_level *trans_wm /* out */)
3763 struct drm_crtc *crtc = cstate->base.crtc;
3764 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3765 struct intel_plane *intel_plane;
3767 if (!cstate->base.active)
3770 /* Until we know more, just disable transition WMs */
3771 for_each_intel_plane_on_crtc(crtc->dev, intel_crtc, intel_plane) {
3772 int i = skl_wm_plane_id(intel_plane);
3774 trans_wm->plane_en[i] = false;
3778 static int skl_build_pipe_wm(struct intel_crtc_state *cstate,
3779 struct skl_ddb_allocation *ddb,
3780 struct skl_pipe_wm *pipe_wm)
3782 struct drm_device *dev = cstate->base.crtc->dev;
3783 const struct drm_i915_private *dev_priv = to_i915(dev);
3784 int level, max_level = ilk_wm_max_level(dev);
3787 for (level = 0; level <= max_level; level++) {
3788 ret = skl_compute_wm_level(dev_priv, ddb, cstate,
3789 level, &pipe_wm->wm[level]);
3793 pipe_wm->linetime = skl_compute_linetime_wm(cstate);
3795 skl_compute_transition_wm(cstate, &pipe_wm->trans_wm);
3800 static void skl_compute_wm_results(struct drm_device *dev,
3801 struct skl_pipe_wm *p_wm,
3802 struct skl_wm_values *r,
3803 struct intel_crtc *intel_crtc)
3805 int level, max_level = ilk_wm_max_level(dev);
3806 enum pipe pipe = intel_crtc->pipe;
3810 for (level = 0; level <= max_level; level++) {
3811 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3814 temp |= p_wm->wm[level].plane_res_l[i] <<
3815 PLANE_WM_LINES_SHIFT;
3816 temp |= p_wm->wm[level].plane_res_b[i];
3817 if (p_wm->wm[level].plane_en[i])
3818 temp |= PLANE_WM_EN;
3820 r->plane[pipe][i][level] = temp;
3825 temp |= p_wm->wm[level].plane_res_l[PLANE_CURSOR] << PLANE_WM_LINES_SHIFT;
3826 temp |= p_wm->wm[level].plane_res_b[PLANE_CURSOR];
3828 if (p_wm->wm[level].plane_en[PLANE_CURSOR])
3829 temp |= PLANE_WM_EN;
3831 r->plane[pipe][PLANE_CURSOR][level] = temp;
3835 /* transition WMs */
3836 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3838 temp |= p_wm->trans_wm.plane_res_l[i] << PLANE_WM_LINES_SHIFT;
3839 temp |= p_wm->trans_wm.plane_res_b[i];
3840 if (p_wm->trans_wm.plane_en[i])
3841 temp |= PLANE_WM_EN;
3843 r->plane_trans[pipe][i] = temp;
3847 temp |= p_wm->trans_wm.plane_res_l[PLANE_CURSOR] << PLANE_WM_LINES_SHIFT;
3848 temp |= p_wm->trans_wm.plane_res_b[PLANE_CURSOR];
3849 if (p_wm->trans_wm.plane_en[PLANE_CURSOR])
3850 temp |= PLANE_WM_EN;
3852 r->plane_trans[pipe][PLANE_CURSOR] = temp;
3854 r->wm_linetime[pipe] = p_wm->linetime;
3857 static void skl_ddb_entry_write(struct drm_i915_private *dev_priv,
3859 const struct skl_ddb_entry *entry)
3862 I915_WRITE(reg, (entry->end - 1) << 16 | entry->start);
3867 void skl_write_plane_wm(struct intel_crtc *intel_crtc,
3868 const struct skl_wm_values *wm,
3871 struct drm_crtc *crtc = &intel_crtc->base;
3872 struct drm_device *dev = crtc->dev;
3873 struct drm_i915_private *dev_priv = to_i915(dev);
3874 int level, max_level = ilk_wm_max_level(dev);
3875 enum pipe pipe = intel_crtc->pipe;
3877 for (level = 0; level <= max_level; level++) {
3878 I915_WRITE(PLANE_WM(pipe, plane, level),
3879 wm->plane[pipe][plane][level]);
3881 I915_WRITE(PLANE_WM_TRANS(pipe, plane), wm->plane_trans[pipe][plane]);
3883 skl_ddb_entry_write(dev_priv, PLANE_BUF_CFG(pipe, plane),
3884 &wm->ddb.plane[pipe][plane]);
3885 skl_ddb_entry_write(dev_priv, PLANE_NV12_BUF_CFG(pipe, plane),
3886 &wm->ddb.y_plane[pipe][plane]);
3889 void skl_write_cursor_wm(struct intel_crtc *intel_crtc,
3890 const struct skl_wm_values *wm)
3892 struct drm_crtc *crtc = &intel_crtc->base;
3893 struct drm_device *dev = crtc->dev;
3894 struct drm_i915_private *dev_priv = to_i915(dev);
3895 int level, max_level = ilk_wm_max_level(dev);
3896 enum pipe pipe = intel_crtc->pipe;
3898 for (level = 0; level <= max_level; level++) {
3899 I915_WRITE(CUR_WM(pipe, level),
3900 wm->plane[pipe][PLANE_CURSOR][level]);
3902 I915_WRITE(CUR_WM_TRANS(pipe), wm->plane_trans[pipe][PLANE_CURSOR]);
3904 skl_ddb_entry_write(dev_priv, CUR_BUF_CFG(pipe),
3905 &wm->ddb.plane[pipe][PLANE_CURSOR]);
3908 bool skl_ddb_allocation_equals(const struct skl_ddb_allocation *old,
3909 const struct skl_ddb_allocation *new,
3912 return new->pipe[pipe].start == old->pipe[pipe].start &&
3913 new->pipe[pipe].end == old->pipe[pipe].end;
3916 static inline bool skl_ddb_entries_overlap(const struct skl_ddb_entry *a,
3917 const struct skl_ddb_entry *b)
3919 return a->start < b->end && b->start < a->end;
3922 bool skl_ddb_allocation_overlaps(struct drm_atomic_state *state,
3923 const struct skl_ddb_allocation *old,
3924 const struct skl_ddb_allocation *new,
3927 struct drm_device *dev = state->dev;
3928 struct intel_crtc *intel_crtc;
3931 for_each_intel_crtc(dev, intel_crtc) {
3932 otherp = intel_crtc->pipe;
3937 if (skl_ddb_entries_overlap(&new->pipe[pipe],
3938 &old->pipe[otherp]))
3945 static int skl_update_pipe_wm(struct drm_crtc_state *cstate,
3946 struct skl_ddb_allocation *ddb, /* out */
3947 struct skl_pipe_wm *pipe_wm, /* out */
3948 bool *changed /* out */)
3950 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->crtc);
3951 struct intel_crtc_state *intel_cstate = to_intel_crtc_state(cstate);
3954 ret = skl_build_pipe_wm(intel_cstate, ddb, pipe_wm);
3958 if (!memcmp(&intel_crtc->wm.active.skl, pipe_wm, sizeof(*pipe_wm)))
3967 pipes_modified(struct drm_atomic_state *state)
3969 struct drm_crtc *crtc;
3970 struct drm_crtc_state *cstate;
3971 uint32_t i, ret = 0;
3973 for_each_crtc_in_state(state, crtc, cstate, i)
3974 ret |= drm_crtc_mask(crtc);
3980 skl_ddb_add_affected_planes(struct intel_crtc_state *cstate)
3982 struct drm_atomic_state *state = cstate->base.state;
3983 struct drm_device *dev = state->dev;
3984 struct drm_crtc *crtc = cstate->base.crtc;
3985 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3986 struct drm_i915_private *dev_priv = to_i915(dev);
3987 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3988 struct skl_ddb_allocation *new_ddb = &intel_state->wm_results.ddb;
3989 struct skl_ddb_allocation *cur_ddb = &dev_priv->wm.skl_hw.ddb;
3990 struct drm_plane_state *plane_state;
3991 struct drm_plane *plane;
3992 enum pipe pipe = intel_crtc->pipe;
3995 WARN_ON(!drm_atomic_get_existing_crtc_state(state, crtc));
3997 drm_for_each_plane_mask(plane, dev, crtc->state->plane_mask) {
3998 id = skl_wm_plane_id(to_intel_plane(plane));
4000 if (skl_ddb_entry_equal(&cur_ddb->plane[pipe][id],
4001 &new_ddb->plane[pipe][id]) &&
4002 skl_ddb_entry_equal(&cur_ddb->y_plane[pipe][id],
4003 &new_ddb->y_plane[pipe][id]))
4006 plane_state = drm_atomic_get_plane_state(state, plane);
4007 if (IS_ERR(plane_state))
4008 return PTR_ERR(plane_state);
4015 skl_compute_ddb(struct drm_atomic_state *state)
4017 struct drm_device *dev = state->dev;
4018 struct drm_i915_private *dev_priv = to_i915(dev);
4019 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
4020 struct intel_crtc *intel_crtc;
4021 struct skl_ddb_allocation *ddb = &intel_state->wm_results.ddb;
4022 uint32_t realloc_pipes = pipes_modified(state);
4026 * If this is our first atomic update following hardware readout,
4027 * we can't trust the DDB that the BIOS programmed for us. Let's
4028 * pretend that all pipes switched active status so that we'll
4029 * ensure a full DDB recompute.
4031 if (dev_priv->wm.distrust_bios_wm) {
4032 ret = drm_modeset_lock(&dev->mode_config.connection_mutex,
4033 state->acquire_ctx);
4037 intel_state->active_pipe_changes = ~0;
4040 * We usually only initialize intel_state->active_crtcs if we
4041 * we're doing a modeset; make sure this field is always
4042 * initialized during the sanitization process that happens
4043 * on the first commit too.
4045 if (!intel_state->modeset)
4046 intel_state->active_crtcs = dev_priv->active_crtcs;
4050 * If the modeset changes which CRTC's are active, we need to
4051 * recompute the DDB allocation for *all* active pipes, even
4052 * those that weren't otherwise being modified in any way by this
4053 * atomic commit. Due to the shrinking of the per-pipe allocations
4054 * when new active CRTC's are added, it's possible for a pipe that
4055 * we were already using and aren't changing at all here to suddenly
4056 * become invalid if its DDB needs exceeds its new allocation.
4058 * Note that if we wind up doing a full DDB recompute, we can't let
4059 * any other display updates race with this transaction, so we need
4060 * to grab the lock on *all* CRTC's.
4062 if (intel_state->active_pipe_changes) {
4064 intel_state->wm_results.dirty_pipes = ~0;
4068 * We're not recomputing for the pipes not included in the commit, so
4069 * make sure we start with the current state.
4071 memcpy(ddb, &dev_priv->wm.skl_hw.ddb, sizeof(*ddb));
4073 for_each_intel_crtc_mask(dev, intel_crtc, realloc_pipes) {
4074 struct intel_crtc_state *cstate;
4076 cstate = intel_atomic_get_crtc_state(state, intel_crtc);
4078 return PTR_ERR(cstate);
4080 ret = skl_allocate_pipe_ddb(cstate, ddb);
4084 ret = skl_ddb_add_affected_planes(cstate);
4093 skl_copy_wm_for_pipe(struct skl_wm_values *dst,
4094 struct skl_wm_values *src,
4097 dst->wm_linetime[pipe] = src->wm_linetime[pipe];
4098 memcpy(dst->plane[pipe], src->plane[pipe],
4099 sizeof(dst->plane[pipe]));
4100 memcpy(dst->plane_trans[pipe], src->plane_trans[pipe],
4101 sizeof(dst->plane_trans[pipe]));
4103 dst->ddb.pipe[pipe] = src->ddb.pipe[pipe];
4104 memcpy(dst->ddb.y_plane[pipe], src->ddb.y_plane[pipe],
4105 sizeof(dst->ddb.y_plane[pipe]));
4106 memcpy(dst->ddb.plane[pipe], src->ddb.plane[pipe],
4107 sizeof(dst->ddb.plane[pipe]));
4111 skl_compute_wm(struct drm_atomic_state *state)
4113 struct drm_crtc *crtc;
4114 struct drm_crtc_state *cstate;
4115 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
4116 struct skl_wm_values *results = &intel_state->wm_results;
4117 struct drm_device *dev = state->dev;
4118 struct skl_pipe_wm *pipe_wm;
4119 bool changed = false;
4123 * When we distrust bios wm we always need to recompute to set the
4124 * expected DDB allocations for each CRTC.
4126 if (to_i915(dev)->wm.distrust_bios_wm)
4130 * If this transaction isn't actually touching any CRTC's, don't
4131 * bother with watermark calculation. Note that if we pass this
4132 * test, we're guaranteed to hold at least one CRTC state mutex,
4133 * which means we can safely use values like dev_priv->active_crtcs
4134 * since any racing commits that want to update them would need to
4135 * hold _all_ CRTC state mutexes.
4137 for_each_crtc_in_state(state, crtc, cstate, i)
4143 /* Clear all dirty flags */
4144 results->dirty_pipes = 0;
4146 ret = skl_compute_ddb(state);
4151 * Calculate WM's for all pipes that are part of this transaction.
4152 * Note that the DDB allocation above may have added more CRTC's that
4153 * weren't otherwise being modified (and set bits in dirty_pipes) if
4154 * pipe allocations had to change.
4156 * FIXME: Now that we're doing this in the atomic check phase, we
4157 * should allow skl_update_pipe_wm() to return failure in cases where
4158 * no suitable watermark values can be found.
4160 for_each_crtc_in_state(state, crtc, cstate, i) {
4161 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4162 struct intel_crtc_state *intel_cstate =
4163 to_intel_crtc_state(cstate);
4165 pipe_wm = &intel_cstate->wm.skl.optimal;
4166 ret = skl_update_pipe_wm(cstate, &results->ddb, pipe_wm,
4172 results->dirty_pipes |= drm_crtc_mask(crtc);
4174 if ((results->dirty_pipes & drm_crtc_mask(crtc)) == 0)
4175 /* This pipe's WM's did not change */
4178 intel_cstate->update_wm_pre = true;
4179 skl_compute_wm_results(crtc->dev, pipe_wm, results, intel_crtc);
4185 static void skl_update_wm(struct drm_crtc *crtc)
4187 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4188 struct drm_device *dev = crtc->dev;
4189 struct drm_i915_private *dev_priv = to_i915(dev);
4190 struct skl_wm_values *results = &dev_priv->wm.skl_results;
4191 struct skl_wm_values *hw_vals = &dev_priv->wm.skl_hw;
4192 struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
4193 struct skl_pipe_wm *pipe_wm = &cstate->wm.skl.optimal;
4194 enum pipe pipe = intel_crtc->pipe;
4196 if ((results->dirty_pipes & drm_crtc_mask(crtc)) == 0)
4199 intel_crtc->wm.active.skl = *pipe_wm;
4201 mutex_lock(&dev_priv->wm.wm_mutex);
4204 * If this pipe isn't active already, we're going to be enabling it
4205 * very soon. Since it's safe to update a pipe's ddb allocation while
4206 * the pipe's shut off, just do so here. Already active pipes will have
4207 * their watermarks updated once we update their planes.
4209 if (crtc->state->active_changed) {
4212 for (plane = 0; plane < intel_num_planes(intel_crtc); plane++)
4213 skl_write_plane_wm(intel_crtc, results, plane);
4215 skl_write_cursor_wm(intel_crtc, results);
4218 skl_copy_wm_for_pipe(hw_vals, results, pipe);
4220 mutex_unlock(&dev_priv->wm.wm_mutex);
4223 static void ilk_compute_wm_config(struct drm_device *dev,
4224 struct intel_wm_config *config)
4226 struct intel_crtc *crtc;
4228 /* Compute the currently _active_ config */
4229 for_each_intel_crtc(dev, crtc) {
4230 const struct intel_pipe_wm *wm = &crtc->wm.active.ilk;
4232 if (!wm->pipe_enabled)
4235 config->sprites_enabled |= wm->sprites_enabled;
4236 config->sprites_scaled |= wm->sprites_scaled;
4237 config->num_pipes_active++;
4241 static void ilk_program_watermarks(struct drm_i915_private *dev_priv)
4243 struct drm_device *dev = &dev_priv->drm;
4244 struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
4245 struct ilk_wm_maximums max;
4246 struct intel_wm_config config = {};
4247 struct ilk_wm_values results = {};
4248 enum intel_ddb_partitioning partitioning;
4250 ilk_compute_wm_config(dev, &config);
4252 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
4253 ilk_wm_merge(dev, &config, &max, &lp_wm_1_2);
4255 /* 5/6 split only in single pipe config on IVB+ */
4256 if (INTEL_INFO(dev)->gen >= 7 &&
4257 config.num_pipes_active == 1 && config.sprites_enabled) {
4258 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
4259 ilk_wm_merge(dev, &config, &max, &lp_wm_5_6);
4261 best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
4263 best_lp_wm = &lp_wm_1_2;
4266 partitioning = (best_lp_wm == &lp_wm_1_2) ?
4267 INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
4269 ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
4271 ilk_write_wm_values(dev_priv, &results);
4274 static void ilk_initial_watermarks(struct intel_crtc_state *cstate)
4276 struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
4277 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4279 mutex_lock(&dev_priv->wm.wm_mutex);
4280 intel_crtc->wm.active.ilk = cstate->wm.ilk.intermediate;
4281 ilk_program_watermarks(dev_priv);
4282 mutex_unlock(&dev_priv->wm.wm_mutex);
4285 static void ilk_optimize_watermarks(struct intel_crtc_state *cstate)
4287 struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
4288 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4290 mutex_lock(&dev_priv->wm.wm_mutex);
4291 if (cstate->wm.need_postvbl_update) {
4292 intel_crtc->wm.active.ilk = cstate->wm.ilk.optimal;
4293 ilk_program_watermarks(dev_priv);
4295 mutex_unlock(&dev_priv->wm.wm_mutex);
4298 static void skl_pipe_wm_active_state(uint32_t val,
4299 struct skl_pipe_wm *active,
4305 bool is_enabled = (val & PLANE_WM_EN) != 0;
4309 active->wm[level].plane_en[i] = is_enabled;
4310 active->wm[level].plane_res_b[i] =
4311 val & PLANE_WM_BLOCKS_MASK;
4312 active->wm[level].plane_res_l[i] =
4313 (val >> PLANE_WM_LINES_SHIFT) &
4314 PLANE_WM_LINES_MASK;
4316 active->wm[level].plane_en[PLANE_CURSOR] = is_enabled;
4317 active->wm[level].plane_res_b[PLANE_CURSOR] =
4318 val & PLANE_WM_BLOCKS_MASK;
4319 active->wm[level].plane_res_l[PLANE_CURSOR] =
4320 (val >> PLANE_WM_LINES_SHIFT) &
4321 PLANE_WM_LINES_MASK;
4325 active->trans_wm.plane_en[i] = is_enabled;
4326 active->trans_wm.plane_res_b[i] =
4327 val & PLANE_WM_BLOCKS_MASK;
4328 active->trans_wm.plane_res_l[i] =
4329 (val >> PLANE_WM_LINES_SHIFT) &
4330 PLANE_WM_LINES_MASK;
4332 active->trans_wm.plane_en[PLANE_CURSOR] = is_enabled;
4333 active->trans_wm.plane_res_b[PLANE_CURSOR] =
4334 val & PLANE_WM_BLOCKS_MASK;
4335 active->trans_wm.plane_res_l[PLANE_CURSOR] =
4336 (val >> PLANE_WM_LINES_SHIFT) &
4337 PLANE_WM_LINES_MASK;
4342 static void skl_pipe_wm_get_hw_state(struct drm_crtc *crtc)
4344 struct drm_device *dev = crtc->dev;
4345 struct drm_i915_private *dev_priv = to_i915(dev);
4346 struct skl_wm_values *hw = &dev_priv->wm.skl_hw;
4347 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4348 struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
4349 struct skl_pipe_wm *active = &cstate->wm.skl.optimal;
4350 enum pipe pipe = intel_crtc->pipe;
4351 int level, i, max_level;
4354 max_level = ilk_wm_max_level(dev);
4356 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
4358 for (level = 0; level <= max_level; level++) {
4359 for (i = 0; i < intel_num_planes(intel_crtc); i++)
4360 hw->plane[pipe][i][level] =
4361 I915_READ(PLANE_WM(pipe, i, level));
4362 hw->plane[pipe][PLANE_CURSOR][level] = I915_READ(CUR_WM(pipe, level));
4365 for (i = 0; i < intel_num_planes(intel_crtc); i++)
4366 hw->plane_trans[pipe][i] = I915_READ(PLANE_WM_TRANS(pipe, i));
4367 hw->plane_trans[pipe][PLANE_CURSOR] = I915_READ(CUR_WM_TRANS(pipe));
4369 if (!intel_crtc->active)
4372 hw->dirty_pipes |= drm_crtc_mask(crtc);
4374 active->linetime = hw->wm_linetime[pipe];
4376 for (level = 0; level <= max_level; level++) {
4377 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
4378 temp = hw->plane[pipe][i][level];
4379 skl_pipe_wm_active_state(temp, active, false,
4382 temp = hw->plane[pipe][PLANE_CURSOR][level];
4383 skl_pipe_wm_active_state(temp, active, false, true, i, level);
4386 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
4387 temp = hw->plane_trans[pipe][i];
4388 skl_pipe_wm_active_state(temp, active, true, false, i, 0);
4391 temp = hw->plane_trans[pipe][PLANE_CURSOR];
4392 skl_pipe_wm_active_state(temp, active, true, true, i, 0);
4394 intel_crtc->wm.active.skl = *active;
4397 void skl_wm_get_hw_state(struct drm_device *dev)
4399 struct drm_i915_private *dev_priv = to_i915(dev);
4400 struct skl_ddb_allocation *ddb = &dev_priv->wm.skl_hw.ddb;
4401 struct drm_crtc *crtc;
4403 skl_ddb_get_hw_state(dev_priv, ddb);
4404 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
4405 skl_pipe_wm_get_hw_state(crtc);
4407 if (dev_priv->active_crtcs) {
4408 /* Fully recompute DDB on first atomic commit */
4409 dev_priv->wm.distrust_bios_wm = true;
4411 /* Easy/common case; just sanitize DDB now if everything off */
4412 memset(ddb, 0, sizeof(*ddb));
4416 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
4418 struct drm_device *dev = crtc->dev;
4419 struct drm_i915_private *dev_priv = to_i915(dev);
4420 struct ilk_wm_values *hw = &dev_priv->wm.hw;
4421 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4422 struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
4423 struct intel_pipe_wm *active = &cstate->wm.ilk.optimal;
4424 enum pipe pipe = intel_crtc->pipe;
4425 static const i915_reg_t wm0_pipe_reg[] = {
4426 [PIPE_A] = WM0_PIPEA_ILK,
4427 [PIPE_B] = WM0_PIPEB_ILK,
4428 [PIPE_C] = WM0_PIPEC_IVB,
4431 hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
4432 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
4433 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
4435 memset(active, 0, sizeof(*active));
4437 active->pipe_enabled = intel_crtc->active;
4439 if (active->pipe_enabled) {
4440 u32 tmp = hw->wm_pipe[pipe];
4443 * For active pipes LP0 watermark is marked as
4444 * enabled, and LP1+ watermaks as disabled since
4445 * we can't really reverse compute them in case
4446 * multiple pipes are active.
4448 active->wm[0].enable = true;
4449 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
4450 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
4451 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
4452 active->linetime = hw->wm_linetime[pipe];
4454 int level, max_level = ilk_wm_max_level(dev);
4457 * For inactive pipes, all watermark levels
4458 * should be marked as enabled but zeroed,
4459 * which is what we'd compute them to.
4461 for (level = 0; level <= max_level; level++)
4462 active->wm[level].enable = true;
4465 intel_crtc->wm.active.ilk = *active;
4468 #define _FW_WM(value, plane) \
4469 (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
4470 #define _FW_WM_VLV(value, plane) \
4471 (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
4473 static void vlv_read_wm_values(struct drm_i915_private *dev_priv,
4474 struct vlv_wm_values *wm)
4479 for_each_pipe(dev_priv, pipe) {
4480 tmp = I915_READ(VLV_DDL(pipe));
4482 wm->ddl[pipe].primary =
4483 (tmp >> DDL_PLANE_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4484 wm->ddl[pipe].cursor =
4485 (tmp >> DDL_CURSOR_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4486 wm->ddl[pipe].sprite[0] =
4487 (tmp >> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4488 wm->ddl[pipe].sprite[1] =
4489 (tmp >> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4492 tmp = I915_READ(DSPFW1);
4493 wm->sr.plane = _FW_WM(tmp, SR);
4494 wm->pipe[PIPE_B].cursor = _FW_WM(tmp, CURSORB);
4495 wm->pipe[PIPE_B].primary = _FW_WM_VLV(tmp, PLANEB);
4496 wm->pipe[PIPE_A].primary = _FW_WM_VLV(tmp, PLANEA);
4498 tmp = I915_READ(DSPFW2);
4499 wm->pipe[PIPE_A].sprite[1] = _FW_WM_VLV(tmp, SPRITEB);
4500 wm->pipe[PIPE_A].cursor = _FW_WM(tmp, CURSORA);
4501 wm->pipe[PIPE_A].sprite[0] = _FW_WM_VLV(tmp, SPRITEA);
4503 tmp = I915_READ(DSPFW3);
4504 wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
4506 if (IS_CHERRYVIEW(dev_priv)) {
4507 tmp = I915_READ(DSPFW7_CHV);
4508 wm->pipe[PIPE_B].sprite[1] = _FW_WM_VLV(tmp, SPRITED);
4509 wm->pipe[PIPE_B].sprite[0] = _FW_WM_VLV(tmp, SPRITEC);
4511 tmp = I915_READ(DSPFW8_CHV);
4512 wm->pipe[PIPE_C].sprite[1] = _FW_WM_VLV(tmp, SPRITEF);
4513 wm->pipe[PIPE_C].sprite[0] = _FW_WM_VLV(tmp, SPRITEE);
4515 tmp = I915_READ(DSPFW9_CHV);
4516 wm->pipe[PIPE_C].primary = _FW_WM_VLV(tmp, PLANEC);
4517 wm->pipe[PIPE_C].cursor = _FW_WM(tmp, CURSORC);
4519 tmp = I915_READ(DSPHOWM);
4520 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
4521 wm->pipe[PIPE_C].sprite[1] |= _FW_WM(tmp, SPRITEF_HI) << 8;
4522 wm->pipe[PIPE_C].sprite[0] |= _FW_WM(tmp, SPRITEE_HI) << 8;
4523 wm->pipe[PIPE_C].primary |= _FW_WM(tmp, PLANEC_HI) << 8;
4524 wm->pipe[PIPE_B].sprite[1] |= _FW_WM(tmp, SPRITED_HI) << 8;
4525 wm->pipe[PIPE_B].sprite[0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
4526 wm->pipe[PIPE_B].primary |= _FW_WM(tmp, PLANEB_HI) << 8;
4527 wm->pipe[PIPE_A].sprite[1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
4528 wm->pipe[PIPE_A].sprite[0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
4529 wm->pipe[PIPE_A].primary |= _FW_WM(tmp, PLANEA_HI) << 8;
4531 tmp = I915_READ(DSPFW7);
4532 wm->pipe[PIPE_B].sprite[1] = _FW_WM_VLV(tmp, SPRITED);
4533 wm->pipe[PIPE_B].sprite[0] = _FW_WM_VLV(tmp, SPRITEC);
4535 tmp = I915_READ(DSPHOWM);
4536 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
4537 wm->pipe[PIPE_B].sprite[1] |= _FW_WM(tmp, SPRITED_HI) << 8;
4538 wm->pipe[PIPE_B].sprite[0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
4539 wm->pipe[PIPE_B].primary |= _FW_WM(tmp, PLANEB_HI) << 8;
4540 wm->pipe[PIPE_A].sprite[1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
4541 wm->pipe[PIPE_A].sprite[0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
4542 wm->pipe[PIPE_A].primary |= _FW_WM(tmp, PLANEA_HI) << 8;
4549 void vlv_wm_get_hw_state(struct drm_device *dev)
4551 struct drm_i915_private *dev_priv = to_i915(dev);
4552 struct vlv_wm_values *wm = &dev_priv->wm.vlv;
4553 struct intel_plane *plane;
4557 vlv_read_wm_values(dev_priv, wm);
4559 for_each_intel_plane(dev, plane) {
4560 switch (plane->base.type) {
4562 case DRM_PLANE_TYPE_CURSOR:
4563 plane->wm.fifo_size = 63;
4565 case DRM_PLANE_TYPE_PRIMARY:
4566 plane->wm.fifo_size = vlv_get_fifo_size(dev, plane->pipe, 0);
4568 case DRM_PLANE_TYPE_OVERLAY:
4569 sprite = plane->plane;
4570 plane->wm.fifo_size = vlv_get_fifo_size(dev, plane->pipe, sprite + 1);
4575 wm->cxsr = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
4576 wm->level = VLV_WM_LEVEL_PM2;
4578 if (IS_CHERRYVIEW(dev_priv)) {
4579 mutex_lock(&dev_priv->rps.hw_lock);
4581 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
4582 if (val & DSP_MAXFIFO_PM5_ENABLE)
4583 wm->level = VLV_WM_LEVEL_PM5;
4586 * If DDR DVFS is disabled in the BIOS, Punit
4587 * will never ack the request. So if that happens
4588 * assume we don't have to enable/disable DDR DVFS
4589 * dynamically. To test that just set the REQ_ACK
4590 * bit to poke the Punit, but don't change the
4591 * HIGH/LOW bits so that we don't actually change
4592 * the current state.
4594 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
4595 val |= FORCE_DDR_FREQ_REQ_ACK;
4596 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
4598 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
4599 FORCE_DDR_FREQ_REQ_ACK) == 0, 3)) {
4600 DRM_DEBUG_KMS("Punit not acking DDR DVFS request, "
4601 "assuming DDR DVFS is disabled\n");
4602 dev_priv->wm.max_level = VLV_WM_LEVEL_PM5;
4604 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
4605 if ((val & FORCE_DDR_HIGH_FREQ) == 0)
4606 wm->level = VLV_WM_LEVEL_DDR_DVFS;
4609 mutex_unlock(&dev_priv->rps.hw_lock);
4612 for_each_pipe(dev_priv, pipe)
4613 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
4614 pipe_name(pipe), wm->pipe[pipe].primary, wm->pipe[pipe].cursor,
4615 wm->pipe[pipe].sprite[0], wm->pipe[pipe].sprite[1]);
4617 DRM_DEBUG_KMS("Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
4618 wm->sr.plane, wm->sr.cursor, wm->level, wm->cxsr);
4621 void ilk_wm_get_hw_state(struct drm_device *dev)
4623 struct drm_i915_private *dev_priv = to_i915(dev);
4624 struct ilk_wm_values *hw = &dev_priv->wm.hw;
4625 struct drm_crtc *crtc;
4627 for_each_crtc(dev, crtc)
4628 ilk_pipe_wm_get_hw_state(crtc);
4630 hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
4631 hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
4632 hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
4634 hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
4635 if (INTEL_INFO(dev)->gen >= 7) {
4636 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
4637 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
4640 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
4641 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
4642 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4643 else if (IS_IVYBRIDGE(dev))
4644 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
4645 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4648 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
4652 * intel_update_watermarks - update FIFO watermark values based on current modes
4654 * Calculate watermark values for the various WM regs based on current mode
4655 * and plane configuration.
4657 * There are several cases to deal with here:
4658 * - normal (i.e. non-self-refresh)
4659 * - self-refresh (SR) mode
4660 * - lines are large relative to FIFO size (buffer can hold up to 2)
4661 * - lines are small relative to FIFO size (buffer can hold more than 2
4662 * lines), so need to account for TLB latency
4664 * The normal calculation is:
4665 * watermark = dotclock * bytes per pixel * latency
4666 * where latency is platform & configuration dependent (we assume pessimal
4669 * The SR calculation is:
4670 * watermark = (trunc(latency/line time)+1) * surface width *
4673 * line time = htotal / dotclock
4674 * surface width = hdisplay for normal plane and 64 for cursor
4675 * and latency is assumed to be high, as above.
4677 * The final value programmed to the register should always be rounded up,
4678 * and include an extra 2 entries to account for clock crossings.
4680 * We don't use the sprite, so we can ignore that. And on Crestline we have
4681 * to set the non-SR watermarks to 8.
4683 void intel_update_watermarks(struct drm_crtc *crtc)
4685 struct drm_i915_private *dev_priv = to_i915(crtc->dev);
4687 if (dev_priv->display.update_wm)
4688 dev_priv->display.update_wm(crtc);
4692 * Lock protecting IPS related data structures
4694 DEFINE_SPINLOCK(mchdev_lock);
4696 /* Global for IPS driver to get at the current i915 device. Protected by
4698 static struct drm_i915_private *i915_mch_dev;
4700 bool ironlake_set_drps(struct drm_i915_private *dev_priv, u8 val)
4704 assert_spin_locked(&mchdev_lock);
4706 rgvswctl = I915_READ16(MEMSWCTL);
4707 if (rgvswctl & MEMCTL_CMD_STS) {
4708 DRM_DEBUG("gpu busy, RCS change rejected\n");
4709 return false; /* still busy with another command */
4712 rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
4713 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
4714 I915_WRITE16(MEMSWCTL, rgvswctl);
4715 POSTING_READ16(MEMSWCTL);
4717 rgvswctl |= MEMCTL_CMD_STS;
4718 I915_WRITE16(MEMSWCTL, rgvswctl);
4723 static void ironlake_enable_drps(struct drm_i915_private *dev_priv)
4726 u8 fmax, fmin, fstart, vstart;
4728 spin_lock_irq(&mchdev_lock);
4730 rgvmodectl = I915_READ(MEMMODECTL);
4732 /* Enable temp reporting */
4733 I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
4734 I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
4736 /* 100ms RC evaluation intervals */
4737 I915_WRITE(RCUPEI, 100000);
4738 I915_WRITE(RCDNEI, 100000);
4740 /* Set max/min thresholds to 90ms and 80ms respectively */
4741 I915_WRITE(RCBMAXAVG, 90000);
4742 I915_WRITE(RCBMINAVG, 80000);
4744 I915_WRITE(MEMIHYST, 1);
4746 /* Set up min, max, and cur for interrupt handling */
4747 fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
4748 fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
4749 fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
4750 MEMMODE_FSTART_SHIFT;
4752 vstart = (I915_READ(PXVFREQ(fstart)) & PXVFREQ_PX_MASK) >>
4755 dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
4756 dev_priv->ips.fstart = fstart;
4758 dev_priv->ips.max_delay = fstart;
4759 dev_priv->ips.min_delay = fmin;
4760 dev_priv->ips.cur_delay = fstart;
4762 DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
4763 fmax, fmin, fstart);
4765 I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
4768 * Interrupts will be enabled in ironlake_irq_postinstall
4771 I915_WRITE(VIDSTART, vstart);
4772 POSTING_READ(VIDSTART);
4774 rgvmodectl |= MEMMODE_SWMODE_EN;
4775 I915_WRITE(MEMMODECTL, rgvmodectl);
4777 if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
4778 DRM_ERROR("stuck trying to change perf mode\n");
4781 ironlake_set_drps(dev_priv, fstart);
4783 dev_priv->ips.last_count1 = I915_READ(DMIEC) +
4784 I915_READ(DDREC) + I915_READ(CSIEC);
4785 dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
4786 dev_priv->ips.last_count2 = I915_READ(GFXEC);
4787 dev_priv->ips.last_time2 = ktime_get_raw_ns();
4789 spin_unlock_irq(&mchdev_lock);
4792 static void ironlake_disable_drps(struct drm_i915_private *dev_priv)
4796 spin_lock_irq(&mchdev_lock);
4798 rgvswctl = I915_READ16(MEMSWCTL);
4800 /* Ack interrupts, disable EFC interrupt */
4801 I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
4802 I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
4803 I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
4804 I915_WRITE(DEIIR, DE_PCU_EVENT);
4805 I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
4807 /* Go back to the starting frequency */
4808 ironlake_set_drps(dev_priv, dev_priv->ips.fstart);
4810 rgvswctl |= MEMCTL_CMD_STS;
4811 I915_WRITE(MEMSWCTL, rgvswctl);
4814 spin_unlock_irq(&mchdev_lock);
4817 /* There's a funny hw issue where the hw returns all 0 when reading from
4818 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
4819 * ourselves, instead of doing a rmw cycle (which might result in us clearing
4820 * all limits and the gpu stuck at whatever frequency it is at atm).
4822 static u32 intel_rps_limits(struct drm_i915_private *dev_priv, u8 val)
4826 /* Only set the down limit when we've reached the lowest level to avoid
4827 * getting more interrupts, otherwise leave this clear. This prevents a
4828 * race in the hw when coming out of rc6: There's a tiny window where
4829 * the hw runs at the minimal clock before selecting the desired
4830 * frequency, if the down threshold expires in that window we will not
4831 * receive a down interrupt. */
4832 if (IS_GEN9(dev_priv)) {
4833 limits = (dev_priv->rps.max_freq_softlimit) << 23;
4834 if (val <= dev_priv->rps.min_freq_softlimit)
4835 limits |= (dev_priv->rps.min_freq_softlimit) << 14;
4837 limits = dev_priv->rps.max_freq_softlimit << 24;
4838 if (val <= dev_priv->rps.min_freq_softlimit)
4839 limits |= dev_priv->rps.min_freq_softlimit << 16;
4845 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
4848 u32 threshold_up = 0, threshold_down = 0; /* in % */
4849 u32 ei_up = 0, ei_down = 0;
4851 new_power = dev_priv->rps.power;
4852 switch (dev_priv->rps.power) {
4854 if (val > dev_priv->rps.efficient_freq + 1 &&
4855 val > dev_priv->rps.cur_freq)
4856 new_power = BETWEEN;
4860 if (val <= dev_priv->rps.efficient_freq &&
4861 val < dev_priv->rps.cur_freq)
4862 new_power = LOW_POWER;
4863 else if (val >= dev_priv->rps.rp0_freq &&
4864 val > dev_priv->rps.cur_freq)
4865 new_power = HIGH_POWER;
4869 if (val < (dev_priv->rps.rp1_freq + dev_priv->rps.rp0_freq) >> 1 &&
4870 val < dev_priv->rps.cur_freq)
4871 new_power = BETWEEN;
4874 /* Max/min bins are special */
4875 if (val <= dev_priv->rps.min_freq_softlimit)
4876 new_power = LOW_POWER;
4877 if (val >= dev_priv->rps.max_freq_softlimit)
4878 new_power = HIGH_POWER;
4879 if (new_power == dev_priv->rps.power)
4882 /* Note the units here are not exactly 1us, but 1280ns. */
4883 switch (new_power) {
4885 /* Upclock if more than 95% busy over 16ms */
4889 /* Downclock if less than 85% busy over 32ms */
4891 threshold_down = 85;
4895 /* Upclock if more than 90% busy over 13ms */
4899 /* Downclock if less than 75% busy over 32ms */
4901 threshold_down = 75;
4905 /* Upclock if more than 85% busy over 10ms */
4909 /* Downclock if less than 60% busy over 32ms */
4911 threshold_down = 60;
4915 /* When byt can survive without system hang with dynamic
4916 * sw freq adjustments, this restriction can be lifted.
4918 if (IS_VALLEYVIEW(dev_priv))
4921 I915_WRITE(GEN6_RP_UP_EI,
4922 GT_INTERVAL_FROM_US(dev_priv, ei_up));
4923 I915_WRITE(GEN6_RP_UP_THRESHOLD,
4924 GT_INTERVAL_FROM_US(dev_priv,
4925 ei_up * threshold_up / 100));
4927 I915_WRITE(GEN6_RP_DOWN_EI,
4928 GT_INTERVAL_FROM_US(dev_priv, ei_down));
4929 I915_WRITE(GEN6_RP_DOWN_THRESHOLD,
4930 GT_INTERVAL_FROM_US(dev_priv,
4931 ei_down * threshold_down / 100));
4933 I915_WRITE(GEN6_RP_CONTROL,
4934 GEN6_RP_MEDIA_TURBO |
4935 GEN6_RP_MEDIA_HW_NORMAL_MODE |
4936 GEN6_RP_MEDIA_IS_GFX |
4938 GEN6_RP_UP_BUSY_AVG |
4939 GEN6_RP_DOWN_IDLE_AVG);
4942 dev_priv->rps.power = new_power;
4943 dev_priv->rps.up_threshold = threshold_up;
4944 dev_priv->rps.down_threshold = threshold_down;
4945 dev_priv->rps.last_adj = 0;
4948 static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
4952 /* We use UP_EI_EXPIRED interupts for both up/down in manual mode */
4953 if (val > dev_priv->rps.min_freq_softlimit)
4954 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
4955 if (val < dev_priv->rps.max_freq_softlimit)
4956 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD;
4958 mask &= dev_priv->pm_rps_events;
4960 return gen6_sanitize_rps_pm_mask(dev_priv, ~mask);
4963 /* gen6_set_rps is called to update the frequency request, but should also be
4964 * called when the range (min_delay and max_delay) is modified so that we can
4965 * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
4966 static void gen6_set_rps(struct drm_i915_private *dev_priv, u8 val)
4968 /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
4969 if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1))
4972 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4973 WARN_ON(val > dev_priv->rps.max_freq);
4974 WARN_ON(val < dev_priv->rps.min_freq);
4976 /* min/max delay may still have been modified so be sure to
4977 * write the limits value.
4979 if (val != dev_priv->rps.cur_freq) {
4980 gen6_set_rps_thresholds(dev_priv, val);
4982 if (IS_GEN9(dev_priv))
4983 I915_WRITE(GEN6_RPNSWREQ,
4984 GEN9_FREQUENCY(val));
4985 else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
4986 I915_WRITE(GEN6_RPNSWREQ,
4987 HSW_FREQUENCY(val));
4989 I915_WRITE(GEN6_RPNSWREQ,
4990 GEN6_FREQUENCY(val) |
4992 GEN6_AGGRESSIVE_TURBO);
4995 /* Make sure we continue to get interrupts
4996 * until we hit the minimum or maximum frequencies.
4998 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, intel_rps_limits(dev_priv, val));
4999 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
5001 POSTING_READ(GEN6_RPNSWREQ);
5003 dev_priv->rps.cur_freq = val;
5004 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
5007 static void valleyview_set_rps(struct drm_i915_private *dev_priv, u8 val)
5009 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5010 WARN_ON(val > dev_priv->rps.max_freq);
5011 WARN_ON(val < dev_priv->rps.min_freq);
5013 if (WARN_ONCE(IS_CHERRYVIEW(dev_priv) && (val & 1),
5014 "Odd GPU freq value\n"))
5017 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
5019 if (val != dev_priv->rps.cur_freq) {
5020 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
5021 if (!IS_CHERRYVIEW(dev_priv))
5022 gen6_set_rps_thresholds(dev_priv, val);
5025 dev_priv->rps.cur_freq = val;
5026 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
5029 /* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
5031 * * If Gfx is Idle, then
5032 * 1. Forcewake Media well.
5033 * 2. Request idle freq.
5034 * 3. Release Forcewake of Media well.
5036 static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
5038 u32 val = dev_priv->rps.idle_freq;
5040 if (dev_priv->rps.cur_freq <= val)
5043 /* Wake up the media well, as that takes a lot less
5044 * power than the Render well. */
5045 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_MEDIA);
5046 valleyview_set_rps(dev_priv, val);
5047 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_MEDIA);
5050 void gen6_rps_busy(struct drm_i915_private *dev_priv)
5052 mutex_lock(&dev_priv->rps.hw_lock);
5053 if (dev_priv->rps.enabled) {
5054 if (dev_priv->pm_rps_events & GEN6_PM_RP_UP_EI_EXPIRED)
5055 gen6_rps_reset_ei(dev_priv);
5056 I915_WRITE(GEN6_PMINTRMSK,
5057 gen6_rps_pm_mask(dev_priv, dev_priv->rps.cur_freq));
5059 gen6_enable_rps_interrupts(dev_priv);
5061 /* Ensure we start at the user's desired frequency */
5062 intel_set_rps(dev_priv,
5063 clamp(dev_priv->rps.cur_freq,
5064 dev_priv->rps.min_freq_softlimit,
5065 dev_priv->rps.max_freq_softlimit));
5067 mutex_unlock(&dev_priv->rps.hw_lock);
5070 void gen6_rps_idle(struct drm_i915_private *dev_priv)
5072 /* Flush our bottom-half so that it does not race with us
5073 * setting the idle frequency and so that it is bounded by
5074 * our rpm wakeref. And then disable the interrupts to stop any
5075 * futher RPS reclocking whilst we are asleep.
5077 gen6_disable_rps_interrupts(dev_priv);
5079 mutex_lock(&dev_priv->rps.hw_lock);
5080 if (dev_priv->rps.enabled) {
5081 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
5082 vlv_set_rps_idle(dev_priv);
5084 gen6_set_rps(dev_priv, dev_priv->rps.idle_freq);
5085 dev_priv->rps.last_adj = 0;
5086 I915_WRITE(GEN6_PMINTRMSK,
5087 gen6_sanitize_rps_pm_mask(dev_priv, ~0));
5089 mutex_unlock(&dev_priv->rps.hw_lock);
5091 spin_lock(&dev_priv->rps.client_lock);
5092 while (!list_empty(&dev_priv->rps.clients))
5093 list_del_init(dev_priv->rps.clients.next);
5094 spin_unlock(&dev_priv->rps.client_lock);
5097 void gen6_rps_boost(struct drm_i915_private *dev_priv,
5098 struct intel_rps_client *rps,
5099 unsigned long submitted)
5101 /* This is intentionally racy! We peek at the state here, then
5102 * validate inside the RPS worker.
5104 if (!(dev_priv->gt.awake &&
5105 dev_priv->rps.enabled &&
5106 dev_priv->rps.cur_freq < dev_priv->rps.boost_freq))
5109 /* Force a RPS boost (and don't count it against the client) if
5110 * the GPU is severely congested.
5112 if (rps && time_after(jiffies, submitted + DRM_I915_THROTTLE_JIFFIES))
5115 spin_lock(&dev_priv->rps.client_lock);
5116 if (rps == NULL || list_empty(&rps->link)) {
5117 spin_lock_irq(&dev_priv->irq_lock);
5118 if (dev_priv->rps.interrupts_enabled) {
5119 dev_priv->rps.client_boost = true;
5120 schedule_work(&dev_priv->rps.work);
5122 spin_unlock_irq(&dev_priv->irq_lock);
5125 list_add(&rps->link, &dev_priv->rps.clients);
5128 dev_priv->rps.boosts++;
5130 spin_unlock(&dev_priv->rps.client_lock);
5133 void intel_set_rps(struct drm_i915_private *dev_priv, u8 val)
5135 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
5136 valleyview_set_rps(dev_priv, val);
5138 gen6_set_rps(dev_priv, val);
5141 static void gen9_disable_rc6(struct drm_i915_private *dev_priv)
5143 I915_WRITE(GEN6_RC_CONTROL, 0);
5144 I915_WRITE(GEN9_PG_ENABLE, 0);
5147 static void gen9_disable_rps(struct drm_i915_private *dev_priv)
5149 I915_WRITE(GEN6_RP_CONTROL, 0);
5152 static void gen6_disable_rps(struct drm_i915_private *dev_priv)
5154 I915_WRITE(GEN6_RC_CONTROL, 0);
5155 I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
5156 I915_WRITE(GEN6_RP_CONTROL, 0);
5159 static void cherryview_disable_rps(struct drm_i915_private *dev_priv)
5161 I915_WRITE(GEN6_RC_CONTROL, 0);
5164 static void valleyview_disable_rps(struct drm_i915_private *dev_priv)
5166 /* we're doing forcewake before Disabling RC6,
5167 * This what the BIOS expects when going into suspend */
5168 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5170 I915_WRITE(GEN6_RC_CONTROL, 0);
5172 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5175 static void intel_print_rc6_info(struct drm_i915_private *dev_priv, u32 mode)
5177 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
5178 if (mode & (GEN7_RC_CTL_TO_MODE | GEN6_RC_CTL_EI_MODE(1)))
5179 mode = GEN6_RC_CTL_RC6_ENABLE;
5183 if (HAS_RC6p(dev_priv))
5184 DRM_DEBUG_DRIVER("Enabling RC6 states: "
5185 "RC6 %s RC6p %s RC6pp %s\n",
5186 onoff(mode & GEN6_RC_CTL_RC6_ENABLE),
5187 onoff(mode & GEN6_RC_CTL_RC6p_ENABLE),
5188 onoff(mode & GEN6_RC_CTL_RC6pp_ENABLE));
5191 DRM_DEBUG_DRIVER("Enabling RC6 states: RC6 %s\n",
5192 onoff(mode & GEN6_RC_CTL_RC6_ENABLE));
5195 static bool bxt_check_bios_rc6_setup(struct drm_i915_private *dev_priv)
5197 struct i915_ggtt *ggtt = &dev_priv->ggtt;
5198 bool enable_rc6 = true;
5199 unsigned long rc6_ctx_base;
5203 rc_ctl = I915_READ(GEN6_RC_CONTROL);
5204 rc_sw_target = (I915_READ(GEN6_RC_STATE) & RC_SW_TARGET_STATE_MASK) >>
5205 RC_SW_TARGET_STATE_SHIFT;
5206 DRM_DEBUG_DRIVER("BIOS enabled RC states: "
5207 "HW_CTRL %s HW_RC6 %s SW_TARGET_STATE %x\n",
5208 onoff(rc_ctl & GEN6_RC_CTL_HW_ENABLE),
5209 onoff(rc_ctl & GEN6_RC_CTL_RC6_ENABLE),
5212 if (!(I915_READ(RC6_LOCATION) & RC6_CTX_IN_DRAM)) {
5213 DRM_DEBUG_DRIVER("RC6 Base location not set properly.\n");
5218 * The exact context size is not known for BXT, so assume a page size
5221 rc6_ctx_base = I915_READ(RC6_CTX_BASE) & RC6_CTX_BASE_MASK;
5222 if (!((rc6_ctx_base >= ggtt->stolen_reserved_base) &&
5223 (rc6_ctx_base + PAGE_SIZE <= ggtt->stolen_reserved_base +
5224 ggtt->stolen_reserved_size))) {
5225 DRM_DEBUG_DRIVER("RC6 Base address not as expected.\n");
5229 if (!(((I915_READ(PWRCTX_MAXCNT_RCSUNIT) & IDLE_TIME_MASK) > 1) &&
5230 ((I915_READ(PWRCTX_MAXCNT_VCSUNIT0) & IDLE_TIME_MASK) > 1) &&
5231 ((I915_READ(PWRCTX_MAXCNT_BCSUNIT) & IDLE_TIME_MASK) > 1) &&
5232 ((I915_READ(PWRCTX_MAXCNT_VECSUNIT) & IDLE_TIME_MASK) > 1))) {
5233 DRM_DEBUG_DRIVER("Engine Idle wait time not set properly.\n");
5237 if (!I915_READ(GEN8_PUSHBUS_CONTROL) ||
5238 !I915_READ(GEN8_PUSHBUS_ENABLE) ||
5239 !I915_READ(GEN8_PUSHBUS_SHIFT)) {
5240 DRM_DEBUG_DRIVER("Pushbus not setup properly.\n");
5244 if (!I915_READ(GEN6_GFXPAUSE)) {
5245 DRM_DEBUG_DRIVER("GFX pause not setup properly.\n");
5249 if (!I915_READ(GEN8_MISC_CTRL0)) {
5250 DRM_DEBUG_DRIVER("GPM control not setup properly.\n");
5257 int sanitize_rc6_option(struct drm_i915_private *dev_priv, int enable_rc6)
5259 /* No RC6 before Ironlake and code is gone for ilk. */
5260 if (INTEL_INFO(dev_priv)->gen < 6)
5266 if (IS_BROXTON(dev_priv) && !bxt_check_bios_rc6_setup(dev_priv)) {
5267 DRM_INFO("RC6 disabled by BIOS\n");
5271 /* Respect the kernel parameter if it is set */
5272 if (enable_rc6 >= 0) {
5275 if (HAS_RC6p(dev_priv))
5276 mask = INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE |
5279 mask = INTEL_RC6_ENABLE;
5281 if ((enable_rc6 & mask) != enable_rc6)
5282 DRM_DEBUG_DRIVER("Adjusting RC6 mask to %d "
5283 "(requested %d, valid %d)\n",
5284 enable_rc6 & mask, enable_rc6, mask);
5286 return enable_rc6 & mask;
5289 if (IS_IVYBRIDGE(dev_priv))
5290 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
5292 return INTEL_RC6_ENABLE;
5295 static void gen6_init_rps_frequencies(struct drm_i915_private *dev_priv)
5297 /* All of these values are in units of 50MHz */
5299 /* static values from HW: RP0 > RP1 > RPn (min_freq) */
5300 if (IS_BROXTON(dev_priv)) {
5301 u32 rp_state_cap = I915_READ(BXT_RP_STATE_CAP);
5302 dev_priv->rps.rp0_freq = (rp_state_cap >> 16) & 0xff;
5303 dev_priv->rps.rp1_freq = (rp_state_cap >> 8) & 0xff;
5304 dev_priv->rps.min_freq = (rp_state_cap >> 0) & 0xff;
5306 u32 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
5307 dev_priv->rps.rp0_freq = (rp_state_cap >> 0) & 0xff;
5308 dev_priv->rps.rp1_freq = (rp_state_cap >> 8) & 0xff;
5309 dev_priv->rps.min_freq = (rp_state_cap >> 16) & 0xff;
5311 /* hw_max = RP0 until we check for overclocking */
5312 dev_priv->rps.max_freq = dev_priv->rps.rp0_freq;
5314 dev_priv->rps.efficient_freq = dev_priv->rps.rp1_freq;
5315 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv) ||
5316 IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5317 u32 ddcc_status = 0;
5319 if (sandybridge_pcode_read(dev_priv,
5320 HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL,
5322 dev_priv->rps.efficient_freq =
5324 ((ddcc_status >> 8) & 0xff),
5325 dev_priv->rps.min_freq,
5326 dev_priv->rps.max_freq);
5329 if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5330 /* Store the frequency values in 16.66 MHZ units, which is
5331 * the natural hardware unit for SKL
5333 dev_priv->rps.rp0_freq *= GEN9_FREQ_SCALER;
5334 dev_priv->rps.rp1_freq *= GEN9_FREQ_SCALER;
5335 dev_priv->rps.min_freq *= GEN9_FREQ_SCALER;
5336 dev_priv->rps.max_freq *= GEN9_FREQ_SCALER;
5337 dev_priv->rps.efficient_freq *= GEN9_FREQ_SCALER;
5341 static void reset_rps(struct drm_i915_private *dev_priv,
5342 void (*set)(struct drm_i915_private *, u8))
5344 u8 freq = dev_priv->rps.cur_freq;
5347 dev_priv->rps.power = -1;
5348 dev_priv->rps.cur_freq = -1;
5350 set(dev_priv, freq);
5353 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
5354 static void gen9_enable_rps(struct drm_i915_private *dev_priv)
5356 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5358 /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
5359 if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) {
5361 * BIOS could leave the Hw Turbo enabled, so need to explicitly
5362 * clear out the Control register just to avoid inconsitency
5363 * with debugfs interface, which will show Turbo as enabled
5364 * only and that is not expected by the User after adding the
5365 * WaGsvDisableTurbo. Apart from this there is no problem even
5366 * if the Turbo is left enabled in the Control register, as the
5367 * Up/Down interrupts would remain masked.
5369 gen9_disable_rps(dev_priv);
5370 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5374 /* Program defaults and thresholds for RPS*/
5375 I915_WRITE(GEN6_RC_VIDEO_FREQ,
5376 GEN9_FREQUENCY(dev_priv->rps.rp1_freq));
5378 /* 1 second timeout*/
5379 I915_WRITE(GEN6_RP_DOWN_TIMEOUT,
5380 GT_INTERVAL_FROM_US(dev_priv, 1000000));
5382 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 0xa);
5384 /* Leaning on the below call to gen6_set_rps to program/setup the
5385 * Up/Down EI & threshold registers, as well as the RP_CONTROL,
5386 * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
5387 reset_rps(dev_priv, gen6_set_rps);
5389 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5392 static void gen9_enable_rc6(struct drm_i915_private *dev_priv)
5394 struct intel_engine_cs *engine;
5395 uint32_t rc6_mask = 0;
5397 /* 1a: Software RC state - RC0 */
5398 I915_WRITE(GEN6_RC_STATE, 0);
5400 /* 1b: Get forcewake during program sequence. Although the driver
5401 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5402 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5404 /* 2a: Disable RC states. */
5405 I915_WRITE(GEN6_RC_CONTROL, 0);
5407 /* 2b: Program RC6 thresholds.*/
5409 /* WaRsDoubleRc6WrlWithCoarsePowerGating: Doubling WRL only when CPG is enabled */
5410 if (IS_SKYLAKE(dev_priv))
5411 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 108 << 16);
5413 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16);
5414 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5415 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
5416 for_each_engine(engine, dev_priv)
5417 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5419 if (HAS_GUC(dev_priv))
5420 I915_WRITE(GUC_MAX_IDLE_COUNT, 0xA);
5422 I915_WRITE(GEN6_RC_SLEEP, 0);
5424 /* 2c: Program Coarse Power Gating Policies. */
5425 I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS, 25);
5426 I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS, 25);
5428 /* 3a: Enable RC6 */
5429 if (intel_enable_rc6() & INTEL_RC6_ENABLE)
5430 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
5431 DRM_INFO("RC6 %s\n", onoff(rc6_mask & GEN6_RC_CTL_RC6_ENABLE));
5432 /* WaRsUseTimeoutMode */
5433 if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_D0) ||
5434 IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) {
5435 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us */
5436 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5437 GEN7_RC_CTL_TO_MODE |
5440 I915_WRITE(GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */
5441 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5442 GEN6_RC_CTL_EI_MODE(1) |
5447 * 3b: Enable Coarse Power Gating only when RC6 is enabled.
5448 * WaRsDisableCoarsePowerGating:skl,bxt - Render/Media PG need to be disabled with RC6.
5450 if (NEEDS_WaRsDisableCoarsePowerGating(dev_priv))
5451 I915_WRITE(GEN9_PG_ENABLE, 0);
5453 I915_WRITE(GEN9_PG_ENABLE, (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ?
5454 (GEN9_RENDER_PG_ENABLE | GEN9_MEDIA_PG_ENABLE) : 0);
5456 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5459 static void gen8_enable_rps(struct drm_i915_private *dev_priv)
5461 struct intel_engine_cs *engine;
5462 uint32_t rc6_mask = 0;
5464 /* 1a: Software RC state - RC0 */
5465 I915_WRITE(GEN6_RC_STATE, 0);
5467 /* 1c & 1d: Get forcewake during program sequence. Although the driver
5468 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5469 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5471 /* 2a: Disable RC states. */
5472 I915_WRITE(GEN6_RC_CONTROL, 0);
5474 /* 2b: Program RC6 thresholds.*/
5475 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
5476 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5477 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
5478 for_each_engine(engine, dev_priv)
5479 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5480 I915_WRITE(GEN6_RC_SLEEP, 0);
5481 if (IS_BROADWELL(dev_priv))
5482 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */
5484 I915_WRITE(GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
5487 if (intel_enable_rc6() & INTEL_RC6_ENABLE)
5488 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
5489 intel_print_rc6_info(dev_priv, rc6_mask);
5490 if (IS_BROADWELL(dev_priv))
5491 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5492 GEN7_RC_CTL_TO_MODE |
5495 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5496 GEN6_RC_CTL_EI_MODE(1) |
5499 /* 4 Program defaults and thresholds for RPS*/
5500 I915_WRITE(GEN6_RPNSWREQ,
5501 HSW_FREQUENCY(dev_priv->rps.rp1_freq));
5502 I915_WRITE(GEN6_RC_VIDEO_FREQ,
5503 HSW_FREQUENCY(dev_priv->rps.rp1_freq));
5504 /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
5505 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
5507 /* Docs recommend 900MHz, and 300 MHz respectively */
5508 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
5509 dev_priv->rps.max_freq_softlimit << 24 |
5510 dev_priv->rps.min_freq_softlimit << 16);
5512 I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
5513 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
5514 I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
5515 I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
5517 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5520 I915_WRITE(GEN6_RP_CONTROL,
5521 GEN6_RP_MEDIA_TURBO |
5522 GEN6_RP_MEDIA_HW_NORMAL_MODE |
5523 GEN6_RP_MEDIA_IS_GFX |
5525 GEN6_RP_UP_BUSY_AVG |
5526 GEN6_RP_DOWN_IDLE_AVG);
5528 /* 6: Ring frequency + overclocking (our driver does this later */
5530 reset_rps(dev_priv, gen6_set_rps);
5532 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5535 static void gen6_enable_rps(struct drm_i915_private *dev_priv)
5537 struct intel_engine_cs *engine;
5538 u32 rc6vids, rc6_mask = 0;
5543 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5545 /* Here begins a magic sequence of register writes to enable
5546 * auto-downclocking.
5548 * Perhaps there might be some value in exposing these to
5551 I915_WRITE(GEN6_RC_STATE, 0);
5553 /* Clear the DBG now so we don't confuse earlier errors */
5554 gtfifodbg = I915_READ(GTFIFODBG);
5556 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
5557 I915_WRITE(GTFIFODBG, gtfifodbg);
5560 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5562 /* disable the counters and set deterministic thresholds */
5563 I915_WRITE(GEN6_RC_CONTROL, 0);
5565 I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
5566 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
5567 I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
5568 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
5569 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
5571 for_each_engine(engine, dev_priv)
5572 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5574 I915_WRITE(GEN6_RC_SLEEP, 0);
5575 I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
5576 if (IS_IVYBRIDGE(dev_priv))
5577 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
5579 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
5580 I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
5581 I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
5583 /* Check if we are enabling RC6 */
5584 rc6_mode = intel_enable_rc6();
5585 if (rc6_mode & INTEL_RC6_ENABLE)
5586 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
5588 /* We don't use those on Haswell */
5589 if (!IS_HASWELL(dev_priv)) {
5590 if (rc6_mode & INTEL_RC6p_ENABLE)
5591 rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
5593 if (rc6_mode & INTEL_RC6pp_ENABLE)
5594 rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
5597 intel_print_rc6_info(dev_priv, rc6_mask);
5599 I915_WRITE(GEN6_RC_CONTROL,
5601 GEN6_RC_CTL_EI_MODE(1) |
5602 GEN6_RC_CTL_HW_ENABLE);
5604 /* Power down if completely idle for over 50ms */
5605 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
5606 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5608 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_MIN_FREQ_TABLE, 0);
5610 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
5612 reset_rps(dev_priv, gen6_set_rps);
5615 ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
5616 if (IS_GEN6(dev_priv) && ret) {
5617 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
5618 } else if (IS_GEN6(dev_priv) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
5619 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
5620 GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
5621 rc6vids &= 0xffff00;
5622 rc6vids |= GEN6_ENCODE_RC6_VID(450);
5623 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
5625 DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
5628 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5631 static void gen6_update_ring_freq(struct drm_i915_private *dev_priv)
5634 unsigned int gpu_freq;
5635 unsigned int max_ia_freq, min_ring_freq;
5636 unsigned int max_gpu_freq, min_gpu_freq;
5637 int scaling_factor = 180;
5638 struct cpufreq_policy *policy;
5640 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5642 policy = cpufreq_cpu_get(0);
5644 max_ia_freq = policy->cpuinfo.max_freq;
5645 cpufreq_cpu_put(policy);
5648 * Default to measured freq if none found, PCU will ensure we
5651 max_ia_freq = tsc_khz;
5654 /* Convert from kHz to MHz */
5655 max_ia_freq /= 1000;
5657 min_ring_freq = I915_READ(DCLK) & 0xf;
5658 /* convert DDR frequency from units of 266.6MHz to bandwidth */
5659 min_ring_freq = mult_frac(min_ring_freq, 8, 3);
5661 if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5662 /* Convert GT frequency to 50 HZ units */
5663 min_gpu_freq = dev_priv->rps.min_freq / GEN9_FREQ_SCALER;
5664 max_gpu_freq = dev_priv->rps.max_freq / GEN9_FREQ_SCALER;
5666 min_gpu_freq = dev_priv->rps.min_freq;
5667 max_gpu_freq = dev_priv->rps.max_freq;
5671 * For each potential GPU frequency, load a ring frequency we'd like
5672 * to use for memory access. We do this by specifying the IA frequency
5673 * the PCU should use as a reference to determine the ring frequency.
5675 for (gpu_freq = max_gpu_freq; gpu_freq >= min_gpu_freq; gpu_freq--) {
5676 int diff = max_gpu_freq - gpu_freq;
5677 unsigned int ia_freq = 0, ring_freq = 0;
5679 if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5681 * ring_freq = 2 * GT. ring_freq is in 100MHz units
5682 * No floor required for ring frequency on SKL.
5684 ring_freq = gpu_freq;
5685 } else if (INTEL_INFO(dev_priv)->gen >= 8) {
5686 /* max(2 * GT, DDR). NB: GT is 50MHz units */
5687 ring_freq = max(min_ring_freq, gpu_freq);
5688 } else if (IS_HASWELL(dev_priv)) {
5689 ring_freq = mult_frac(gpu_freq, 5, 4);
5690 ring_freq = max(min_ring_freq, ring_freq);
5691 /* leave ia_freq as the default, chosen by cpufreq */
5693 /* On older processors, there is no separate ring
5694 * clock domain, so in order to boost the bandwidth
5695 * of the ring, we need to upclock the CPU (ia_freq).
5697 * For GPU frequencies less than 750MHz,
5698 * just use the lowest ring freq.
5700 if (gpu_freq < min_freq)
5703 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
5704 ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
5707 sandybridge_pcode_write(dev_priv,
5708 GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
5709 ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
5710 ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
5715 static int cherryview_rps_max_freq(struct drm_i915_private *dev_priv)
5719 val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
5721 switch (INTEL_INFO(dev_priv)->sseu.eu_total) {
5723 /* (2 * 4) config */
5724 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT);
5727 /* (2 * 6) config */
5728 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT);
5731 /* (2 * 8) config */
5733 /* Setting (2 * 8) Min RP0 for any other combination */
5734 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT);
5738 rp0 = (rp0 & FB_GFX_FREQ_FUSE_MASK);
5743 static int cherryview_rps_rpe_freq(struct drm_i915_private *dev_priv)
5747 val = vlv_punit_read(dev_priv, PUNIT_GPU_DUTYCYCLE_REG);
5748 rpe = (val >> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK;
5753 static int cherryview_rps_guar_freq(struct drm_i915_private *dev_priv)
5757 val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
5758 rp1 = (val & FB_GFX_FREQ_FUSE_MASK);
5763 static int valleyview_rps_guar_freq(struct drm_i915_private *dev_priv)
5767 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
5769 rp1 = (val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT;
5774 static int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
5778 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
5780 rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
5782 rp0 = min_t(u32, rp0, 0xea);
5787 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
5791 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
5792 rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
5793 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
5794 rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
5799 static int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
5803 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
5805 * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
5806 * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
5807 * a BYT-M B0 the above register contains 0xbf. Moreover when setting
5808 * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
5809 * to make sure it matches what Punit accepts.
5811 return max_t(u32, val, 0xc0);
5814 /* Check that the pctx buffer wasn't move under us. */
5815 static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
5817 unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
5819 WARN_ON(pctx_addr != dev_priv->mm.stolen_base +
5820 dev_priv->vlv_pctx->stolen->start);
5824 /* Check that the pcbr address is not empty. */
5825 static void cherryview_check_pctx(struct drm_i915_private *dev_priv)
5827 unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
5829 WARN_ON((pctx_addr >> VLV_PCBR_ADDR_SHIFT) == 0);
5832 static void cherryview_setup_pctx(struct drm_i915_private *dev_priv)
5834 struct i915_ggtt *ggtt = &dev_priv->ggtt;
5835 unsigned long pctx_paddr, paddr;
5837 int pctx_size = 32*1024;
5839 pcbr = I915_READ(VLV_PCBR);
5840 if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) {
5841 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5842 paddr = (dev_priv->mm.stolen_base +
5843 (ggtt->stolen_size - pctx_size));
5845 pctx_paddr = (paddr & (~4095));
5846 I915_WRITE(VLV_PCBR, pctx_paddr);
5849 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
5852 static void valleyview_setup_pctx(struct drm_i915_private *dev_priv)
5854 struct drm_i915_gem_object *pctx;
5855 unsigned long pctx_paddr;
5857 int pctx_size = 24*1024;
5859 pcbr = I915_READ(VLV_PCBR);
5861 /* BIOS set it up already, grab the pre-alloc'd space */
5864 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
5865 pctx = i915_gem_object_create_stolen_for_preallocated(&dev_priv->drm,
5867 I915_GTT_OFFSET_NONE,
5872 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5875 * From the Gunit register HAS:
5876 * The Gfx driver is expected to program this register and ensure
5877 * proper allocation within Gfx stolen memory. For example, this
5878 * register should be programmed such than the PCBR range does not
5879 * overlap with other ranges, such as the frame buffer, protected
5880 * memory, or any other relevant ranges.
5882 pctx = i915_gem_object_create_stolen(&dev_priv->drm, pctx_size);
5884 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
5888 pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
5889 I915_WRITE(VLV_PCBR, pctx_paddr);
5892 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
5893 dev_priv->vlv_pctx = pctx;
5896 static void valleyview_cleanup_pctx(struct drm_i915_private *dev_priv)
5898 if (WARN_ON(!dev_priv->vlv_pctx))
5901 i915_gem_object_put_unlocked(dev_priv->vlv_pctx);
5902 dev_priv->vlv_pctx = NULL;
5905 static void vlv_init_gpll_ref_freq(struct drm_i915_private *dev_priv)
5907 dev_priv->rps.gpll_ref_freq =
5908 vlv_get_cck_clock(dev_priv, "GPLL ref",
5909 CCK_GPLL_CLOCK_CONTROL,
5910 dev_priv->czclk_freq);
5912 DRM_DEBUG_DRIVER("GPLL reference freq: %d kHz\n",
5913 dev_priv->rps.gpll_ref_freq);
5916 static void valleyview_init_gt_powersave(struct drm_i915_private *dev_priv)
5920 valleyview_setup_pctx(dev_priv);
5922 vlv_init_gpll_ref_freq(dev_priv);
5924 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5925 switch ((val >> 6) & 3) {
5928 dev_priv->mem_freq = 800;
5931 dev_priv->mem_freq = 1066;
5934 dev_priv->mem_freq = 1333;
5937 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
5939 dev_priv->rps.max_freq = valleyview_rps_max_freq(dev_priv);
5940 dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
5941 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5942 intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
5943 dev_priv->rps.max_freq);
5945 dev_priv->rps.efficient_freq = valleyview_rps_rpe_freq(dev_priv);
5946 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5947 intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5948 dev_priv->rps.efficient_freq);
5950 dev_priv->rps.rp1_freq = valleyview_rps_guar_freq(dev_priv);
5951 DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
5952 intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
5953 dev_priv->rps.rp1_freq);
5955 dev_priv->rps.min_freq = valleyview_rps_min_freq(dev_priv);
5956 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5957 intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
5958 dev_priv->rps.min_freq);
5961 static void cherryview_init_gt_powersave(struct drm_i915_private *dev_priv)
5965 cherryview_setup_pctx(dev_priv);
5967 vlv_init_gpll_ref_freq(dev_priv);
5969 mutex_lock(&dev_priv->sb_lock);
5970 val = vlv_cck_read(dev_priv, CCK_FUSE_REG);
5971 mutex_unlock(&dev_priv->sb_lock);
5973 switch ((val >> 2) & 0x7) {
5975 dev_priv->mem_freq = 2000;
5978 dev_priv->mem_freq = 1600;
5981 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
5983 dev_priv->rps.max_freq = cherryview_rps_max_freq(dev_priv);
5984 dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
5985 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5986 intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
5987 dev_priv->rps.max_freq);
5989 dev_priv->rps.efficient_freq = cherryview_rps_rpe_freq(dev_priv);
5990 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5991 intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5992 dev_priv->rps.efficient_freq);
5994 dev_priv->rps.rp1_freq = cherryview_rps_guar_freq(dev_priv);
5995 DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
5996 intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
5997 dev_priv->rps.rp1_freq);
5999 /* PUnit validated range is only [RPe, RP0] */
6000 dev_priv->rps.min_freq = dev_priv->rps.efficient_freq;
6001 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
6002 intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
6003 dev_priv->rps.min_freq);
6005 WARN_ONCE((dev_priv->rps.max_freq |
6006 dev_priv->rps.efficient_freq |
6007 dev_priv->rps.rp1_freq |
6008 dev_priv->rps.min_freq) & 1,
6009 "Odd GPU freq values\n");
6012 static void valleyview_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
6014 valleyview_cleanup_pctx(dev_priv);
6017 static void cherryview_enable_rps(struct drm_i915_private *dev_priv)
6019 struct intel_engine_cs *engine;
6020 u32 gtfifodbg, val, rc6_mode = 0, pcbr;
6022 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6024 gtfifodbg = I915_READ(GTFIFODBG) & ~(GT_FIFO_SBDEDICATE_FREE_ENTRY_CHV |
6025 GT_FIFO_FREE_ENTRIES_CHV);
6027 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
6029 I915_WRITE(GTFIFODBG, gtfifodbg);
6032 cherryview_check_pctx(dev_priv);
6034 /* 1a & 1b: Get forcewake during program sequence. Although the driver
6035 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
6036 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6038 /* Disable RC states. */
6039 I915_WRITE(GEN6_RC_CONTROL, 0);
6041 /* 2a: Program RC6 thresholds.*/
6042 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
6043 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
6044 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
6046 for_each_engine(engine, dev_priv)
6047 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6048 I915_WRITE(GEN6_RC_SLEEP, 0);
6050 /* TO threshold set to 500 us ( 0x186 * 1.28 us) */
6051 I915_WRITE(GEN6_RC6_THRESHOLD, 0x186);
6053 /* allows RC6 residency counter to work */
6054 I915_WRITE(VLV_COUNTER_CONTROL,
6055 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
6056 VLV_MEDIA_RC6_COUNT_EN |
6057 VLV_RENDER_RC6_COUNT_EN));
6059 /* For now we assume BIOS is allocating and populating the PCBR */
6060 pcbr = I915_READ(VLV_PCBR);
6063 if ((intel_enable_rc6() & INTEL_RC6_ENABLE) &&
6064 (pcbr >> VLV_PCBR_ADDR_SHIFT))
6065 rc6_mode = GEN7_RC_CTL_TO_MODE;
6067 I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
6069 /* 4 Program defaults and thresholds for RPS*/
6070 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
6071 I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
6072 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
6073 I915_WRITE(GEN6_RP_UP_EI, 66000);
6074 I915_WRITE(GEN6_RP_DOWN_EI, 350000);
6076 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
6079 I915_WRITE(GEN6_RP_CONTROL,
6080 GEN6_RP_MEDIA_HW_NORMAL_MODE |
6081 GEN6_RP_MEDIA_IS_GFX |
6083 GEN6_RP_UP_BUSY_AVG |
6084 GEN6_RP_DOWN_IDLE_AVG);
6086 /* Setting Fixed Bias */
6087 val = VLV_OVERRIDE_EN |
6089 CHV_BIAS_CPU_50_SOC_50;
6090 vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
6092 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
6094 /* RPS code assumes GPLL is used */
6095 WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
6097 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
6098 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
6100 reset_rps(dev_priv, valleyview_set_rps);
6102 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6105 static void valleyview_enable_rps(struct drm_i915_private *dev_priv)
6107 struct intel_engine_cs *engine;
6108 u32 gtfifodbg, val, rc6_mode = 0;
6110 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6112 valleyview_check_pctx(dev_priv);
6114 gtfifodbg = I915_READ(GTFIFODBG);
6116 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
6118 I915_WRITE(GTFIFODBG, gtfifodbg);
6121 /* If VLV, Forcewake all wells, else re-direct to regular path */
6122 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6124 /* Disable RC states. */
6125 I915_WRITE(GEN6_RC_CONTROL, 0);
6127 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
6128 I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
6129 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
6130 I915_WRITE(GEN6_RP_UP_EI, 66000);
6131 I915_WRITE(GEN6_RP_DOWN_EI, 350000);
6133 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
6135 I915_WRITE(GEN6_RP_CONTROL,
6136 GEN6_RP_MEDIA_TURBO |
6137 GEN6_RP_MEDIA_HW_NORMAL_MODE |
6138 GEN6_RP_MEDIA_IS_GFX |
6140 GEN6_RP_UP_BUSY_AVG |
6141 GEN6_RP_DOWN_IDLE_CONT);
6143 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
6144 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
6145 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
6147 for_each_engine(engine, dev_priv)
6148 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6150 I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
6152 /* allows RC6 residency counter to work */
6153 I915_WRITE(VLV_COUNTER_CONTROL,
6154 _MASKED_BIT_ENABLE(VLV_MEDIA_RC0_COUNT_EN |
6155 VLV_RENDER_RC0_COUNT_EN |
6156 VLV_MEDIA_RC6_COUNT_EN |
6157 VLV_RENDER_RC6_COUNT_EN));
6159 if (intel_enable_rc6() & INTEL_RC6_ENABLE)
6160 rc6_mode = GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
6162 intel_print_rc6_info(dev_priv, rc6_mode);
6164 I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
6166 /* Setting Fixed Bias */
6167 val = VLV_OVERRIDE_EN |
6169 VLV_BIAS_CPU_125_SOC_875;
6170 vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
6172 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
6174 /* RPS code assumes GPLL is used */
6175 WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
6177 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
6178 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
6180 reset_rps(dev_priv, valleyview_set_rps);
6182 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6185 static unsigned long intel_pxfreq(u32 vidfreq)
6188 int div = (vidfreq & 0x3f0000) >> 16;
6189 int post = (vidfreq & 0x3000) >> 12;
6190 int pre = (vidfreq & 0x7);
6195 freq = ((div * 133333) / ((1<<post) * pre));
6200 static const struct cparams {
6206 { 1, 1333, 301, 28664 },
6207 { 1, 1066, 294, 24460 },
6208 { 1, 800, 294, 25192 },
6209 { 0, 1333, 276, 27605 },
6210 { 0, 1066, 276, 27605 },
6211 { 0, 800, 231, 23784 },
6214 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
6216 u64 total_count, diff, ret;
6217 u32 count1, count2, count3, m = 0, c = 0;
6218 unsigned long now = jiffies_to_msecs(jiffies), diff1;
6221 assert_spin_locked(&mchdev_lock);
6223 diff1 = now - dev_priv->ips.last_time1;
6225 /* Prevent division-by-zero if we are asking too fast.
6226 * Also, we don't get interesting results if we are polling
6227 * faster than once in 10ms, so just return the saved value
6231 return dev_priv->ips.chipset_power;
6233 count1 = I915_READ(DMIEC);
6234 count2 = I915_READ(DDREC);
6235 count3 = I915_READ(CSIEC);
6237 total_count = count1 + count2 + count3;
6239 /* FIXME: handle per-counter overflow */
6240 if (total_count < dev_priv->ips.last_count1) {
6241 diff = ~0UL - dev_priv->ips.last_count1;
6242 diff += total_count;
6244 diff = total_count - dev_priv->ips.last_count1;
6247 for (i = 0; i < ARRAY_SIZE(cparams); i++) {
6248 if (cparams[i].i == dev_priv->ips.c_m &&
6249 cparams[i].t == dev_priv->ips.r_t) {
6256 diff = div_u64(diff, diff1);
6257 ret = ((m * diff) + c);
6258 ret = div_u64(ret, 10);
6260 dev_priv->ips.last_count1 = total_count;
6261 dev_priv->ips.last_time1 = now;
6263 dev_priv->ips.chipset_power = ret;
6268 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
6272 if (INTEL_INFO(dev_priv)->gen != 5)
6275 spin_lock_irq(&mchdev_lock);
6277 val = __i915_chipset_val(dev_priv);
6279 spin_unlock_irq(&mchdev_lock);
6284 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
6286 unsigned long m, x, b;
6289 tsfs = I915_READ(TSFS);
6291 m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
6292 x = I915_READ8(TR1);
6294 b = tsfs & TSFS_INTR_MASK;
6296 return ((m * x) / 127) - b;
6299 static int _pxvid_to_vd(u8 pxvid)
6304 if (pxvid >= 8 && pxvid < 31)
6307 return (pxvid + 2) * 125;
6310 static u32 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
6312 const int vd = _pxvid_to_vd(pxvid);
6313 const int vm = vd - 1125;
6315 if (INTEL_INFO(dev_priv)->is_mobile)
6316 return vm > 0 ? vm : 0;
6321 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
6323 u64 now, diff, diffms;
6326 assert_spin_locked(&mchdev_lock);
6328 now = ktime_get_raw_ns();
6329 diffms = now - dev_priv->ips.last_time2;
6330 do_div(diffms, NSEC_PER_MSEC);
6332 /* Don't divide by 0 */
6336 count = I915_READ(GFXEC);
6338 if (count < dev_priv->ips.last_count2) {
6339 diff = ~0UL - dev_priv->ips.last_count2;
6342 diff = count - dev_priv->ips.last_count2;
6345 dev_priv->ips.last_count2 = count;
6346 dev_priv->ips.last_time2 = now;
6348 /* More magic constants... */
6350 diff = div_u64(diff, diffms * 10);
6351 dev_priv->ips.gfx_power = diff;
6354 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
6356 if (INTEL_INFO(dev_priv)->gen != 5)
6359 spin_lock_irq(&mchdev_lock);
6361 __i915_update_gfx_val(dev_priv);
6363 spin_unlock_irq(&mchdev_lock);
6366 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
6368 unsigned long t, corr, state1, corr2, state2;
6371 assert_spin_locked(&mchdev_lock);
6373 pxvid = I915_READ(PXVFREQ(dev_priv->rps.cur_freq));
6374 pxvid = (pxvid >> 24) & 0x7f;
6375 ext_v = pvid_to_extvid(dev_priv, pxvid);
6379 t = i915_mch_val(dev_priv);
6381 /* Revel in the empirically derived constants */
6383 /* Correction factor in 1/100000 units */
6385 corr = ((t * 2349) + 135940);
6387 corr = ((t * 964) + 29317);
6389 corr = ((t * 301) + 1004);
6391 corr = corr * ((150142 * state1) / 10000 - 78642);
6393 corr2 = (corr * dev_priv->ips.corr);
6395 state2 = (corr2 * state1) / 10000;
6396 state2 /= 100; /* convert to mW */
6398 __i915_update_gfx_val(dev_priv);
6400 return dev_priv->ips.gfx_power + state2;
6403 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
6407 if (INTEL_INFO(dev_priv)->gen != 5)
6410 spin_lock_irq(&mchdev_lock);
6412 val = __i915_gfx_val(dev_priv);
6414 spin_unlock_irq(&mchdev_lock);
6420 * i915_read_mch_val - return value for IPS use
6422 * Calculate and return a value for the IPS driver to use when deciding whether
6423 * we have thermal and power headroom to increase CPU or GPU power budget.
6425 unsigned long i915_read_mch_val(void)
6427 struct drm_i915_private *dev_priv;
6428 unsigned long chipset_val, graphics_val, ret = 0;
6430 spin_lock_irq(&mchdev_lock);
6433 dev_priv = i915_mch_dev;
6435 chipset_val = __i915_chipset_val(dev_priv);
6436 graphics_val = __i915_gfx_val(dev_priv);
6438 ret = chipset_val + graphics_val;
6441 spin_unlock_irq(&mchdev_lock);
6445 EXPORT_SYMBOL_GPL(i915_read_mch_val);
6448 * i915_gpu_raise - raise GPU frequency limit
6450 * Raise the limit; IPS indicates we have thermal headroom.
6452 bool i915_gpu_raise(void)
6454 struct drm_i915_private *dev_priv;
6457 spin_lock_irq(&mchdev_lock);
6458 if (!i915_mch_dev) {
6462 dev_priv = i915_mch_dev;
6464 if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
6465 dev_priv->ips.max_delay--;
6468 spin_unlock_irq(&mchdev_lock);
6472 EXPORT_SYMBOL_GPL(i915_gpu_raise);
6475 * i915_gpu_lower - lower GPU frequency limit
6477 * IPS indicates we're close to a thermal limit, so throttle back the GPU
6478 * frequency maximum.
6480 bool i915_gpu_lower(void)
6482 struct drm_i915_private *dev_priv;
6485 spin_lock_irq(&mchdev_lock);
6486 if (!i915_mch_dev) {
6490 dev_priv = i915_mch_dev;
6492 if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
6493 dev_priv->ips.max_delay++;
6496 spin_unlock_irq(&mchdev_lock);
6500 EXPORT_SYMBOL_GPL(i915_gpu_lower);
6503 * i915_gpu_busy - indicate GPU business to IPS
6505 * Tell the IPS driver whether or not the GPU is busy.
6507 bool i915_gpu_busy(void)
6511 spin_lock_irq(&mchdev_lock);
6513 ret = i915_mch_dev->gt.awake;
6514 spin_unlock_irq(&mchdev_lock);
6518 EXPORT_SYMBOL_GPL(i915_gpu_busy);
6521 * i915_gpu_turbo_disable - disable graphics turbo
6523 * Disable graphics turbo by resetting the max frequency and setting the
6524 * current frequency to the default.
6526 bool i915_gpu_turbo_disable(void)
6528 struct drm_i915_private *dev_priv;
6531 spin_lock_irq(&mchdev_lock);
6532 if (!i915_mch_dev) {
6536 dev_priv = i915_mch_dev;
6538 dev_priv->ips.max_delay = dev_priv->ips.fstart;
6540 if (!ironlake_set_drps(dev_priv, dev_priv->ips.fstart))
6544 spin_unlock_irq(&mchdev_lock);
6548 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
6551 * Tells the intel_ips driver that the i915 driver is now loaded, if
6552 * IPS got loaded first.
6554 * This awkward dance is so that neither module has to depend on the
6555 * other in order for IPS to do the appropriate communication of
6556 * GPU turbo limits to i915.
6559 ips_ping_for_i915_load(void)
6563 link = symbol_get(ips_link_to_i915_driver);
6566 symbol_put(ips_link_to_i915_driver);
6570 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
6572 /* We only register the i915 ips part with intel-ips once everything is
6573 * set up, to avoid intel-ips sneaking in and reading bogus values. */
6574 spin_lock_irq(&mchdev_lock);
6575 i915_mch_dev = dev_priv;
6576 spin_unlock_irq(&mchdev_lock);
6578 ips_ping_for_i915_load();
6581 void intel_gpu_ips_teardown(void)
6583 spin_lock_irq(&mchdev_lock);
6584 i915_mch_dev = NULL;
6585 spin_unlock_irq(&mchdev_lock);
6588 static void intel_init_emon(struct drm_i915_private *dev_priv)
6594 /* Disable to program */
6598 /* Program energy weights for various events */
6599 I915_WRITE(SDEW, 0x15040d00);
6600 I915_WRITE(CSIEW0, 0x007f0000);
6601 I915_WRITE(CSIEW1, 0x1e220004);
6602 I915_WRITE(CSIEW2, 0x04000004);
6604 for (i = 0; i < 5; i++)
6605 I915_WRITE(PEW(i), 0);
6606 for (i = 0; i < 3; i++)
6607 I915_WRITE(DEW(i), 0);
6609 /* Program P-state weights to account for frequency power adjustment */
6610 for (i = 0; i < 16; i++) {
6611 u32 pxvidfreq = I915_READ(PXVFREQ(i));
6612 unsigned long freq = intel_pxfreq(pxvidfreq);
6613 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
6618 val *= (freq / 1000);
6620 val /= (127*127*900);
6622 DRM_ERROR("bad pxval: %ld\n", val);
6625 /* Render standby states get 0 weight */
6629 for (i = 0; i < 4; i++) {
6630 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
6631 (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
6632 I915_WRITE(PXW(i), val);
6635 /* Adjust magic regs to magic values (more experimental results) */
6636 I915_WRITE(OGW0, 0);
6637 I915_WRITE(OGW1, 0);
6638 I915_WRITE(EG0, 0x00007f00);
6639 I915_WRITE(EG1, 0x0000000e);
6640 I915_WRITE(EG2, 0x000e0000);
6641 I915_WRITE(EG3, 0x68000300);
6642 I915_WRITE(EG4, 0x42000000);
6643 I915_WRITE(EG5, 0x00140031);
6647 for (i = 0; i < 8; i++)
6648 I915_WRITE(PXWL(i), 0);
6650 /* Enable PMON + select events */
6651 I915_WRITE(ECR, 0x80000019);
6653 lcfuse = I915_READ(LCFUSE02);
6655 dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
6658 void intel_init_gt_powersave(struct drm_i915_private *dev_priv)
6661 * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
6664 if (!i915.enable_rc6) {
6665 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
6666 intel_runtime_pm_get(dev_priv);
6669 mutex_lock(&dev_priv->drm.struct_mutex);
6670 mutex_lock(&dev_priv->rps.hw_lock);
6672 /* Initialize RPS limits (for userspace) */
6673 if (IS_CHERRYVIEW(dev_priv))
6674 cherryview_init_gt_powersave(dev_priv);
6675 else if (IS_VALLEYVIEW(dev_priv))
6676 valleyview_init_gt_powersave(dev_priv);
6677 else if (INTEL_GEN(dev_priv) >= 6)
6678 gen6_init_rps_frequencies(dev_priv);
6680 /* Derive initial user preferences/limits from the hardware limits */
6681 dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
6682 dev_priv->rps.cur_freq = dev_priv->rps.idle_freq;
6684 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
6685 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
6687 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
6688 dev_priv->rps.min_freq_softlimit =
6690 dev_priv->rps.efficient_freq,
6691 intel_freq_opcode(dev_priv, 450));
6693 /* After setting max-softlimit, find the overclock max freq */
6694 if (IS_GEN6(dev_priv) ||
6695 IS_IVYBRIDGE(dev_priv) || IS_HASWELL(dev_priv)) {
6698 sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, ¶ms);
6699 if (params & BIT(31)) { /* OC supported */
6700 DRM_DEBUG_DRIVER("Overclocking supported, max: %dMHz, overclock: %dMHz\n",
6701 (dev_priv->rps.max_freq & 0xff) * 50,
6702 (params & 0xff) * 50);
6703 dev_priv->rps.max_freq = params & 0xff;
6707 /* Finally allow us to boost to max by default */
6708 dev_priv->rps.boost_freq = dev_priv->rps.max_freq;
6710 mutex_unlock(&dev_priv->rps.hw_lock);
6711 mutex_unlock(&dev_priv->drm.struct_mutex);
6713 intel_autoenable_gt_powersave(dev_priv);
6716 void intel_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
6718 if (IS_VALLEYVIEW(dev_priv))
6719 valleyview_cleanup_gt_powersave(dev_priv);
6721 if (!i915.enable_rc6)
6722 intel_runtime_pm_put(dev_priv);
6726 * intel_suspend_gt_powersave - suspend PM work and helper threads
6727 * @dev_priv: i915 device
6729 * We don't want to disable RC6 or other features here, we just want
6730 * to make sure any work we've queued has finished and won't bother
6731 * us while we're suspended.
6733 void intel_suspend_gt_powersave(struct drm_i915_private *dev_priv)
6735 if (INTEL_GEN(dev_priv) < 6)
6738 if (cancel_delayed_work_sync(&dev_priv->rps.autoenable_work))
6739 intel_runtime_pm_put(dev_priv);
6741 /* gen6_rps_idle() will be called later to disable interrupts */
6744 void intel_sanitize_gt_powersave(struct drm_i915_private *dev_priv)
6746 dev_priv->rps.enabled = true; /* force disabling */
6747 intel_disable_gt_powersave(dev_priv);
6749 gen6_reset_rps_interrupts(dev_priv);
6752 void intel_disable_gt_powersave(struct drm_i915_private *dev_priv)
6754 if (!READ_ONCE(dev_priv->rps.enabled))
6757 mutex_lock(&dev_priv->rps.hw_lock);
6759 if (INTEL_GEN(dev_priv) >= 9) {
6760 gen9_disable_rc6(dev_priv);
6761 gen9_disable_rps(dev_priv);
6762 } else if (IS_CHERRYVIEW(dev_priv)) {
6763 cherryview_disable_rps(dev_priv);
6764 } else if (IS_VALLEYVIEW(dev_priv)) {
6765 valleyview_disable_rps(dev_priv);
6766 } else if (INTEL_GEN(dev_priv) >= 6) {
6767 gen6_disable_rps(dev_priv);
6768 } else if (IS_IRONLAKE_M(dev_priv)) {
6769 ironlake_disable_drps(dev_priv);
6772 dev_priv->rps.enabled = false;
6773 mutex_unlock(&dev_priv->rps.hw_lock);
6776 void intel_enable_gt_powersave(struct drm_i915_private *dev_priv)
6778 /* We shouldn't be disabling as we submit, so this should be less
6779 * racy than it appears!
6781 if (READ_ONCE(dev_priv->rps.enabled))
6784 /* Powersaving is controlled by the host when inside a VM */
6785 if (intel_vgpu_active(dev_priv))
6788 mutex_lock(&dev_priv->rps.hw_lock);
6790 if (IS_CHERRYVIEW(dev_priv)) {
6791 cherryview_enable_rps(dev_priv);
6792 } else if (IS_VALLEYVIEW(dev_priv)) {
6793 valleyview_enable_rps(dev_priv);
6794 } else if (INTEL_GEN(dev_priv) >= 9) {
6795 gen9_enable_rc6(dev_priv);
6796 gen9_enable_rps(dev_priv);
6797 if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv))
6798 gen6_update_ring_freq(dev_priv);
6799 } else if (IS_BROADWELL(dev_priv)) {
6800 gen8_enable_rps(dev_priv);
6801 gen6_update_ring_freq(dev_priv);
6802 } else if (INTEL_GEN(dev_priv) >= 6) {
6803 gen6_enable_rps(dev_priv);
6804 gen6_update_ring_freq(dev_priv);
6805 } else if (IS_IRONLAKE_M(dev_priv)) {
6806 ironlake_enable_drps(dev_priv);
6807 intel_init_emon(dev_priv);
6810 WARN_ON(dev_priv->rps.max_freq < dev_priv->rps.min_freq);
6811 WARN_ON(dev_priv->rps.idle_freq > dev_priv->rps.max_freq);
6813 WARN_ON(dev_priv->rps.efficient_freq < dev_priv->rps.min_freq);
6814 WARN_ON(dev_priv->rps.efficient_freq > dev_priv->rps.max_freq);
6816 dev_priv->rps.enabled = true;
6817 mutex_unlock(&dev_priv->rps.hw_lock);
6820 static void __intel_autoenable_gt_powersave(struct work_struct *work)
6822 struct drm_i915_private *dev_priv =
6823 container_of(work, typeof(*dev_priv), rps.autoenable_work.work);
6824 struct intel_engine_cs *rcs;
6825 struct drm_i915_gem_request *req;
6827 if (READ_ONCE(dev_priv->rps.enabled))
6830 rcs = &dev_priv->engine[RCS];
6831 if (rcs->last_context)
6834 if (!rcs->init_context)
6837 mutex_lock(&dev_priv->drm.struct_mutex);
6839 req = i915_gem_request_alloc(rcs, dev_priv->kernel_context);
6843 if (!i915.enable_execlists && i915_switch_context(req) == 0)
6844 rcs->init_context(req);
6846 /* Mark the device busy, calling intel_enable_gt_powersave() */
6847 i915_add_request_no_flush(req);
6850 mutex_unlock(&dev_priv->drm.struct_mutex);
6852 intel_runtime_pm_put(dev_priv);
6855 void intel_autoenable_gt_powersave(struct drm_i915_private *dev_priv)
6857 if (READ_ONCE(dev_priv->rps.enabled))
6860 if (IS_IRONLAKE_M(dev_priv)) {
6861 ironlake_enable_drps(dev_priv);
6862 intel_init_emon(dev_priv);
6863 } else if (INTEL_INFO(dev_priv)->gen >= 6) {
6865 * PCU communication is slow and this doesn't need to be
6866 * done at any specific time, so do this out of our fast path
6867 * to make resume and init faster.
6869 * We depend on the HW RC6 power context save/restore
6870 * mechanism when entering D3 through runtime PM suspend. So
6871 * disable RPM until RPS/RC6 is properly setup. We can only
6872 * get here via the driver load/system resume/runtime resume
6873 * paths, so the _noresume version is enough (and in case of
6874 * runtime resume it's necessary).
6876 if (queue_delayed_work(dev_priv->wq,
6877 &dev_priv->rps.autoenable_work,
6878 round_jiffies_up_relative(HZ)))
6879 intel_runtime_pm_get_noresume(dev_priv);
6883 static void ibx_init_clock_gating(struct drm_device *dev)
6885 struct drm_i915_private *dev_priv = to_i915(dev);
6888 * On Ibex Peak and Cougar Point, we need to disable clock
6889 * gating for the panel power sequencer or it will fail to
6890 * start up when no ports are active.
6892 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
6895 static void g4x_disable_trickle_feed(struct drm_device *dev)
6897 struct drm_i915_private *dev_priv = to_i915(dev);
6900 for_each_pipe(dev_priv, pipe) {
6901 I915_WRITE(DSPCNTR(pipe),
6902 I915_READ(DSPCNTR(pipe)) |
6903 DISPPLANE_TRICKLE_FEED_DISABLE);
6905 I915_WRITE(DSPSURF(pipe), I915_READ(DSPSURF(pipe)));
6906 POSTING_READ(DSPSURF(pipe));
6910 static void ilk_init_lp_watermarks(struct drm_device *dev)
6912 struct drm_i915_private *dev_priv = to_i915(dev);
6914 I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
6915 I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
6916 I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
6919 * Don't touch WM1S_LP_EN here.
6920 * Doing so could cause underruns.
6924 static void ironlake_init_clock_gating(struct drm_device *dev)
6926 struct drm_i915_private *dev_priv = to_i915(dev);
6927 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6931 * WaFbcDisableDpfcClockGating:ilk
6933 dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
6934 ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
6935 ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
6937 I915_WRITE(PCH_3DCGDIS0,
6938 MARIUNIT_CLOCK_GATE_DISABLE |
6939 SVSMUNIT_CLOCK_GATE_DISABLE);
6940 I915_WRITE(PCH_3DCGDIS1,
6941 VFMUNIT_CLOCK_GATE_DISABLE);
6944 * According to the spec the following bits should be set in
6945 * order to enable memory self-refresh
6946 * The bit 22/21 of 0x42004
6947 * The bit 5 of 0x42020
6948 * The bit 15 of 0x45000
6950 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6951 (I915_READ(ILK_DISPLAY_CHICKEN2) |
6952 ILK_DPARB_GATE | ILK_VSDPFD_FULL));
6953 dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
6954 I915_WRITE(DISP_ARB_CTL,
6955 (I915_READ(DISP_ARB_CTL) |
6958 ilk_init_lp_watermarks(dev);
6961 * Based on the document from hardware guys the following bits
6962 * should be set unconditionally in order to enable FBC.
6963 * The bit 22 of 0x42000
6964 * The bit 22 of 0x42004
6965 * The bit 7,8,9 of 0x42020.
6967 if (IS_IRONLAKE_M(dev)) {
6968 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
6969 I915_WRITE(ILK_DISPLAY_CHICKEN1,
6970 I915_READ(ILK_DISPLAY_CHICKEN1) |
6972 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6973 I915_READ(ILK_DISPLAY_CHICKEN2) |
6977 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
6979 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6980 I915_READ(ILK_DISPLAY_CHICKEN2) |
6981 ILK_ELPIN_409_SELECT);
6982 I915_WRITE(_3D_CHICKEN2,
6983 _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
6984 _3D_CHICKEN2_WM_READ_PIPELINED);
6986 /* WaDisableRenderCachePipelinedFlush:ilk */
6987 I915_WRITE(CACHE_MODE_0,
6988 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
6990 /* WaDisable_RenderCache_OperationalFlush:ilk */
6991 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6993 g4x_disable_trickle_feed(dev);
6995 ibx_init_clock_gating(dev);
6998 static void cpt_init_clock_gating(struct drm_device *dev)
7000 struct drm_i915_private *dev_priv = to_i915(dev);
7005 * On Ibex Peak and Cougar Point, we need to disable clock
7006 * gating for the panel power sequencer or it will fail to
7007 * start up when no ports are active.
7009 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
7010 PCH_DPLUNIT_CLOCK_GATE_DISABLE |
7011 PCH_CPUNIT_CLOCK_GATE_DISABLE);
7012 I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
7013 DPLS_EDP_PPS_FIX_DIS);
7014 /* The below fixes the weird display corruption, a few pixels shifted
7015 * downward, on (only) LVDS of some HP laptops with IVY.
7017 for_each_pipe(dev_priv, pipe) {
7018 val = I915_READ(TRANS_CHICKEN2(pipe));
7019 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
7020 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
7021 if (dev_priv->vbt.fdi_rx_polarity_inverted)
7022 val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
7023 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
7024 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
7025 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
7026 I915_WRITE(TRANS_CHICKEN2(pipe), val);
7028 /* WADP0ClockGatingDisable */
7029 for_each_pipe(dev_priv, pipe) {
7030 I915_WRITE(TRANS_CHICKEN1(pipe),
7031 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
7035 static void gen6_check_mch_setup(struct drm_device *dev)
7037 struct drm_i915_private *dev_priv = to_i915(dev);
7040 tmp = I915_READ(MCH_SSKPD);
7041 if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL)
7042 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
7046 static void gen6_init_clock_gating(struct drm_device *dev)
7048 struct drm_i915_private *dev_priv = to_i915(dev);
7049 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
7051 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
7053 I915_WRITE(ILK_DISPLAY_CHICKEN2,
7054 I915_READ(ILK_DISPLAY_CHICKEN2) |
7055 ILK_ELPIN_409_SELECT);
7057 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
7058 I915_WRITE(_3D_CHICKEN,
7059 _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
7061 /* WaDisable_RenderCache_OperationalFlush:snb */
7062 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7065 * BSpec recoomends 8x4 when MSAA is used,
7066 * however in practice 16x4 seems fastest.
7068 * Note that PS/WM thread counts depend on the WIZ hashing
7069 * disable bit, which we don't touch here, but it's good
7070 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7072 I915_WRITE(GEN6_GT_MODE,
7073 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7075 ilk_init_lp_watermarks(dev);
7077 I915_WRITE(CACHE_MODE_0,
7078 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
7080 I915_WRITE(GEN6_UCGCTL1,
7081 I915_READ(GEN6_UCGCTL1) |
7082 GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
7083 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
7085 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
7086 * gating disable must be set. Failure to set it results in
7087 * flickering pixels due to Z write ordering failures after
7088 * some amount of runtime in the Mesa "fire" demo, and Unigine
7089 * Sanctuary and Tropics, and apparently anything else with
7090 * alpha test or pixel discard.
7092 * According to the spec, bit 11 (RCCUNIT) must also be set,
7093 * but we didn't debug actual testcases to find it out.
7095 * WaDisableRCCUnitClockGating:snb
7096 * WaDisableRCPBUnitClockGating:snb
7098 I915_WRITE(GEN6_UCGCTL2,
7099 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
7100 GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
7102 /* WaStripsFansDisableFastClipPerformanceFix:snb */
7103 I915_WRITE(_3D_CHICKEN3,
7104 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
7108 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
7109 * 3DSTATE_SF number of SF output attributes is more than 16."
7111 I915_WRITE(_3D_CHICKEN3,
7112 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
7115 * According to the spec the following bits should be
7116 * set in order to enable memory self-refresh and fbc:
7117 * The bit21 and bit22 of 0x42000
7118 * The bit21 and bit22 of 0x42004
7119 * The bit5 and bit7 of 0x42020
7120 * The bit14 of 0x70180
7121 * The bit14 of 0x71180
7123 * WaFbcAsynchFlipDisableFbcQueue:snb
7125 I915_WRITE(ILK_DISPLAY_CHICKEN1,
7126 I915_READ(ILK_DISPLAY_CHICKEN1) |
7127 ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
7128 I915_WRITE(ILK_DISPLAY_CHICKEN2,
7129 I915_READ(ILK_DISPLAY_CHICKEN2) |
7130 ILK_DPARB_GATE | ILK_VSDPFD_FULL);
7131 I915_WRITE(ILK_DSPCLK_GATE_D,
7132 I915_READ(ILK_DSPCLK_GATE_D) |
7133 ILK_DPARBUNIT_CLOCK_GATE_ENABLE |
7134 ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
7136 g4x_disable_trickle_feed(dev);
7138 cpt_init_clock_gating(dev);
7140 gen6_check_mch_setup(dev);
7143 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
7145 uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
7148 * WaVSThreadDispatchOverride:ivb,vlv
7150 * This actually overrides the dispatch
7151 * mode for all thread types.
7153 reg &= ~GEN7_FF_SCHED_MASK;
7154 reg |= GEN7_FF_TS_SCHED_HW;
7155 reg |= GEN7_FF_VS_SCHED_HW;
7156 reg |= GEN7_FF_DS_SCHED_HW;
7158 I915_WRITE(GEN7_FF_THREAD_MODE, reg);
7161 static void lpt_init_clock_gating(struct drm_device *dev)
7163 struct drm_i915_private *dev_priv = to_i915(dev);
7166 * TODO: this bit should only be enabled when really needed, then
7167 * disabled when not needed anymore in order to save power.
7169 if (HAS_PCH_LPT_LP(dev))
7170 I915_WRITE(SOUTH_DSPCLK_GATE_D,
7171 I915_READ(SOUTH_DSPCLK_GATE_D) |
7172 PCH_LP_PARTITION_LEVEL_DISABLE);
7174 /* WADPOClockGatingDisable:hsw */
7175 I915_WRITE(TRANS_CHICKEN1(PIPE_A),
7176 I915_READ(TRANS_CHICKEN1(PIPE_A)) |
7177 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
7180 static void lpt_suspend_hw(struct drm_device *dev)
7182 struct drm_i915_private *dev_priv = to_i915(dev);
7184 if (HAS_PCH_LPT_LP(dev)) {
7185 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
7187 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
7188 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
7192 static void gen8_set_l3sqc_credits(struct drm_i915_private *dev_priv,
7193 int general_prio_credits,
7194 int high_prio_credits)
7198 /* WaTempDisableDOPClkGating:bdw */
7199 misccpctl = I915_READ(GEN7_MISCCPCTL);
7200 I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
7202 I915_WRITE(GEN8_L3SQCREG1,
7203 L3_GENERAL_PRIO_CREDITS(general_prio_credits) |
7204 L3_HIGH_PRIO_CREDITS(high_prio_credits));
7207 * Wait at least 100 clocks before re-enabling clock gating.
7208 * See the definition of L3SQCREG1 in BSpec.
7210 POSTING_READ(GEN8_L3SQCREG1);
7212 I915_WRITE(GEN7_MISCCPCTL, misccpctl);
7215 static void kabylake_init_clock_gating(struct drm_device *dev)
7217 struct drm_i915_private *dev_priv = dev->dev_private;
7219 gen9_init_clock_gating(dev);
7221 /* WaDisableSDEUnitClockGating:kbl */
7222 if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
7223 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7224 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7226 /* WaDisableGamClockGating:kbl */
7227 if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
7228 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
7229 GEN6_GAMUNIT_CLOCK_GATE_DISABLE);
7231 /* WaFbcNukeOnHostModify:kbl */
7232 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
7233 ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
7236 static void skylake_init_clock_gating(struct drm_device *dev)
7238 struct drm_i915_private *dev_priv = dev->dev_private;
7240 gen9_init_clock_gating(dev);
7242 /* WAC6entrylatency:skl */
7243 I915_WRITE(FBC_LLC_READ_CTRL, I915_READ(FBC_LLC_READ_CTRL) |
7244 FBC_LLC_FULLY_OPEN);
7246 /* WaFbcNukeOnHostModify:skl */
7247 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
7248 ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
7251 static void broadwell_init_clock_gating(struct drm_device *dev)
7253 struct drm_i915_private *dev_priv = to_i915(dev);
7256 ilk_init_lp_watermarks(dev);
7258 /* WaSwitchSolVfFArbitrationPriority:bdw */
7259 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
7261 /* WaPsrDPAMaskVBlankInSRD:bdw */
7262 I915_WRITE(CHICKEN_PAR1_1,
7263 I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
7265 /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
7266 for_each_pipe(dev_priv, pipe) {
7267 I915_WRITE(CHICKEN_PIPESL_1(pipe),
7268 I915_READ(CHICKEN_PIPESL_1(pipe)) |
7269 BDW_DPRS_MASK_VBLANK_SRD);
7272 /* WaVSRefCountFullforceMissDisable:bdw */
7273 /* WaDSRefCountFullforceMissDisable:bdw */
7274 I915_WRITE(GEN7_FF_THREAD_MODE,
7275 I915_READ(GEN7_FF_THREAD_MODE) &
7276 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
7278 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
7279 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
7281 /* WaDisableSDEUnitClockGating:bdw */
7282 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7283 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7285 /* WaProgramL3SqcReg1Default:bdw */
7286 gen8_set_l3sqc_credits(dev_priv, 30, 2);
7289 * WaGttCachingOffByDefault:bdw
7290 * GTT cache may not work with big pages, so if those
7291 * are ever enabled GTT cache may need to be disabled.
7293 I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
7295 /* WaKVMNotificationOnConfigChange:bdw */
7296 I915_WRITE(CHICKEN_PAR2_1, I915_READ(CHICKEN_PAR2_1)
7297 | KVM_CONFIG_CHANGE_NOTIFICATION_SELECT);
7299 lpt_init_clock_gating(dev);
7302 static void haswell_init_clock_gating(struct drm_device *dev)
7304 struct drm_i915_private *dev_priv = to_i915(dev);
7306 ilk_init_lp_watermarks(dev);
7308 /* L3 caching of data atomics doesn't work -- disable it. */
7309 I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
7310 I915_WRITE(HSW_ROW_CHICKEN3,
7311 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
7313 /* This is required by WaCatErrorRejectionIssue:hsw */
7314 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7315 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7316 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7318 /* WaVSRefCountFullforceMissDisable:hsw */
7319 I915_WRITE(GEN7_FF_THREAD_MODE,
7320 I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
7322 /* WaDisable_RenderCache_OperationalFlush:hsw */
7323 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7325 /* enable HiZ Raw Stall Optimization */
7326 I915_WRITE(CACHE_MODE_0_GEN7,
7327 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
7329 /* WaDisable4x2SubspanOptimization:hsw */
7330 I915_WRITE(CACHE_MODE_1,
7331 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7334 * BSpec recommends 8x4 when MSAA is used,
7335 * however in practice 16x4 seems fastest.
7337 * Note that PS/WM thread counts depend on the WIZ hashing
7338 * disable bit, which we don't touch here, but it's good
7339 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7341 I915_WRITE(GEN7_GT_MODE,
7342 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7344 /* WaSampleCChickenBitEnable:hsw */
7345 I915_WRITE(HALF_SLICE_CHICKEN3,
7346 _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE));
7348 /* WaSwitchSolVfFArbitrationPriority:hsw */
7349 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
7351 /* WaRsPkgCStateDisplayPMReq:hsw */
7352 I915_WRITE(CHICKEN_PAR1_1,
7353 I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
7355 lpt_init_clock_gating(dev);
7358 static void ivybridge_init_clock_gating(struct drm_device *dev)
7360 struct drm_i915_private *dev_priv = to_i915(dev);
7363 ilk_init_lp_watermarks(dev);
7365 I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
7367 /* WaDisableEarlyCull:ivb */
7368 I915_WRITE(_3D_CHICKEN3,
7369 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
7371 /* WaDisableBackToBackFlipFix:ivb */
7372 I915_WRITE(IVB_CHICKEN3,
7373 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7374 CHICKEN3_DGMG_DONE_FIX_DISABLE);
7376 /* WaDisablePSDDualDispatchEnable:ivb */
7377 if (IS_IVB_GT1(dev))
7378 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
7379 _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
7381 /* WaDisable_RenderCache_OperationalFlush:ivb */
7382 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7384 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
7385 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
7386 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
7388 /* WaApplyL3ControlAndL3ChickenMode:ivb */
7389 I915_WRITE(GEN7_L3CNTLREG1,
7390 GEN7_WA_FOR_GEN7_L3_CONTROL);
7391 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
7392 GEN7_WA_L3_CHICKEN_MODE);
7393 if (IS_IVB_GT1(dev))
7394 I915_WRITE(GEN7_ROW_CHICKEN2,
7395 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7397 /* must write both registers */
7398 I915_WRITE(GEN7_ROW_CHICKEN2,
7399 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7400 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
7401 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7404 /* WaForceL3Serialization:ivb */
7405 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
7406 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
7409 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7410 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
7412 I915_WRITE(GEN6_UCGCTL2,
7413 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7415 /* This is required by WaCatErrorRejectionIssue:ivb */
7416 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7417 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7418 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7420 g4x_disable_trickle_feed(dev);
7422 gen7_setup_fixed_func_scheduler(dev_priv);
7424 if (0) { /* causes HiZ corruption on ivb:gt1 */
7425 /* enable HiZ Raw Stall Optimization */
7426 I915_WRITE(CACHE_MODE_0_GEN7,
7427 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
7430 /* WaDisable4x2SubspanOptimization:ivb */
7431 I915_WRITE(CACHE_MODE_1,
7432 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7435 * BSpec recommends 8x4 when MSAA is used,
7436 * however in practice 16x4 seems fastest.
7438 * Note that PS/WM thread counts depend on the WIZ hashing
7439 * disable bit, which we don't touch here, but it's good
7440 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7442 I915_WRITE(GEN7_GT_MODE,
7443 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7445 snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
7446 snpcr &= ~GEN6_MBC_SNPCR_MASK;
7447 snpcr |= GEN6_MBC_SNPCR_MED;
7448 I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
7450 if (!HAS_PCH_NOP(dev))
7451 cpt_init_clock_gating(dev);
7453 gen6_check_mch_setup(dev);
7456 static void valleyview_init_clock_gating(struct drm_device *dev)
7458 struct drm_i915_private *dev_priv = to_i915(dev);
7460 /* WaDisableEarlyCull:vlv */
7461 I915_WRITE(_3D_CHICKEN3,
7462 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
7464 /* WaDisableBackToBackFlipFix:vlv */
7465 I915_WRITE(IVB_CHICKEN3,
7466 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7467 CHICKEN3_DGMG_DONE_FIX_DISABLE);
7469 /* WaPsdDispatchEnable:vlv */
7470 /* WaDisablePSDDualDispatchEnable:vlv */
7471 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
7472 _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
7473 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
7475 /* WaDisable_RenderCache_OperationalFlush:vlv */
7476 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7478 /* WaForceL3Serialization:vlv */
7479 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
7480 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
7482 /* WaDisableDopClockGating:vlv */
7483 I915_WRITE(GEN7_ROW_CHICKEN2,
7484 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7486 /* This is required by WaCatErrorRejectionIssue:vlv */
7487 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7488 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7489 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7491 gen7_setup_fixed_func_scheduler(dev_priv);
7494 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7495 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
7497 I915_WRITE(GEN6_UCGCTL2,
7498 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7500 /* WaDisableL3Bank2xClockGate:vlv
7501 * Disabling L3 clock gating- MMIO 940c[25] = 1
7502 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
7503 I915_WRITE(GEN7_UCGCTL4,
7504 I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
7507 * BSpec says this must be set, even though
7508 * WaDisable4x2SubspanOptimization isn't listed for VLV.
7510 I915_WRITE(CACHE_MODE_1,
7511 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7514 * BSpec recommends 8x4 when MSAA is used,
7515 * however in practice 16x4 seems fastest.
7517 * Note that PS/WM thread counts depend on the WIZ hashing
7518 * disable bit, which we don't touch here, but it's good
7519 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7521 I915_WRITE(GEN7_GT_MODE,
7522 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7525 * WaIncreaseL3CreditsForVLVB0:vlv
7526 * This is the hardware default actually.
7528 I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
7531 * WaDisableVLVClockGating_VBIIssue:vlv
7532 * Disable clock gating on th GCFG unit to prevent a delay
7533 * in the reporting of vblank events.
7535 I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
7538 static void cherryview_init_clock_gating(struct drm_device *dev)
7540 struct drm_i915_private *dev_priv = to_i915(dev);
7542 /* WaVSRefCountFullforceMissDisable:chv */
7543 /* WaDSRefCountFullforceMissDisable:chv */
7544 I915_WRITE(GEN7_FF_THREAD_MODE,
7545 I915_READ(GEN7_FF_THREAD_MODE) &
7546 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
7548 /* WaDisableSemaphoreAndSyncFlipWait:chv */
7549 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
7550 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
7552 /* WaDisableCSUnitClockGating:chv */
7553 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
7554 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
7556 /* WaDisableSDEUnitClockGating:chv */
7557 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7558 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7561 * WaProgramL3SqcReg1Default:chv
7562 * See gfxspecs/Related Documents/Performance Guide/
7563 * LSQC Setting Recommendations.
7565 gen8_set_l3sqc_credits(dev_priv, 38, 2);
7568 * GTT cache may not work with big pages, so if those
7569 * are ever enabled GTT cache may need to be disabled.
7571 I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
7574 static void g4x_init_clock_gating(struct drm_device *dev)
7576 struct drm_i915_private *dev_priv = to_i915(dev);
7577 uint32_t dspclk_gate;
7579 I915_WRITE(RENCLK_GATE_D1, 0);
7580 I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
7581 GS_UNIT_CLOCK_GATE_DISABLE |
7582 CL_UNIT_CLOCK_GATE_DISABLE);
7583 I915_WRITE(RAMCLK_GATE_D, 0);
7584 dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
7585 OVRUNIT_CLOCK_GATE_DISABLE |
7586 OVCUNIT_CLOCK_GATE_DISABLE;
7588 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
7589 I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
7591 /* WaDisableRenderCachePipelinedFlush */
7592 I915_WRITE(CACHE_MODE_0,
7593 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
7595 /* WaDisable_RenderCache_OperationalFlush:g4x */
7596 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7598 g4x_disable_trickle_feed(dev);
7601 static void crestline_init_clock_gating(struct drm_device *dev)
7603 struct drm_i915_private *dev_priv = to_i915(dev);
7605 I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
7606 I915_WRITE(RENCLK_GATE_D2, 0);
7607 I915_WRITE(DSPCLK_GATE_D, 0);
7608 I915_WRITE(RAMCLK_GATE_D, 0);
7609 I915_WRITE16(DEUC, 0);
7610 I915_WRITE(MI_ARB_STATE,
7611 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7613 /* WaDisable_RenderCache_OperationalFlush:gen4 */
7614 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7617 static void broadwater_init_clock_gating(struct drm_device *dev)
7619 struct drm_i915_private *dev_priv = to_i915(dev);
7621 I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
7622 I965_RCC_CLOCK_GATE_DISABLE |
7623 I965_RCPB_CLOCK_GATE_DISABLE |
7624 I965_ISC_CLOCK_GATE_DISABLE |
7625 I965_FBC_CLOCK_GATE_DISABLE);
7626 I915_WRITE(RENCLK_GATE_D2, 0);
7627 I915_WRITE(MI_ARB_STATE,
7628 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7630 /* WaDisable_RenderCache_OperationalFlush:gen4 */
7631 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7634 static void gen3_init_clock_gating(struct drm_device *dev)
7636 struct drm_i915_private *dev_priv = to_i915(dev);
7637 u32 dstate = I915_READ(D_STATE);
7639 dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
7640 DSTATE_DOT_CLOCK_GATING;
7641 I915_WRITE(D_STATE, dstate);
7643 if (IS_PINEVIEW(dev))
7644 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
7646 /* IIR "flip pending" means done if this bit is set */
7647 I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
7649 /* interrupts should cause a wake up from C3 */
7650 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
7652 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
7653 I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
7655 I915_WRITE(MI_ARB_STATE,
7656 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7659 static void i85x_init_clock_gating(struct drm_device *dev)
7661 struct drm_i915_private *dev_priv = to_i915(dev);
7663 I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
7665 /* interrupts should cause a wake up from C3 */
7666 I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
7667 _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
7669 I915_WRITE(MEM_MODE,
7670 _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE));
7673 static void i830_init_clock_gating(struct drm_device *dev)
7675 struct drm_i915_private *dev_priv = to_i915(dev);
7677 I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
7679 I915_WRITE(MEM_MODE,
7680 _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE) |
7681 _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE));
7684 void intel_init_clock_gating(struct drm_device *dev)
7686 struct drm_i915_private *dev_priv = to_i915(dev);
7688 dev_priv->display.init_clock_gating(dev);
7691 void intel_suspend_hw(struct drm_device *dev)
7693 if (HAS_PCH_LPT(dev))
7694 lpt_suspend_hw(dev);
7697 static void nop_init_clock_gating(struct drm_device *dev)
7699 DRM_DEBUG_KMS("No clock gating settings or workarounds applied.\n");
7703 * intel_init_clock_gating_hooks - setup the clock gating hooks
7704 * @dev_priv: device private
7706 * Setup the hooks that configure which clocks of a given platform can be
7707 * gated and also apply various GT and display specific workarounds for these
7708 * platforms. Note that some GT specific workarounds are applied separately
7709 * when GPU contexts or batchbuffers start their execution.
7711 void intel_init_clock_gating_hooks(struct drm_i915_private *dev_priv)
7713 if (IS_SKYLAKE(dev_priv))
7714 dev_priv->display.init_clock_gating = skylake_init_clock_gating;
7715 else if (IS_KABYLAKE(dev_priv))
7716 dev_priv->display.init_clock_gating = kabylake_init_clock_gating;
7717 else if (IS_BROXTON(dev_priv))
7718 dev_priv->display.init_clock_gating = bxt_init_clock_gating;
7719 else if (IS_BROADWELL(dev_priv))
7720 dev_priv->display.init_clock_gating = broadwell_init_clock_gating;
7721 else if (IS_CHERRYVIEW(dev_priv))
7722 dev_priv->display.init_clock_gating = cherryview_init_clock_gating;
7723 else if (IS_HASWELL(dev_priv))
7724 dev_priv->display.init_clock_gating = haswell_init_clock_gating;
7725 else if (IS_IVYBRIDGE(dev_priv))
7726 dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
7727 else if (IS_VALLEYVIEW(dev_priv))
7728 dev_priv->display.init_clock_gating = valleyview_init_clock_gating;
7729 else if (IS_GEN6(dev_priv))
7730 dev_priv->display.init_clock_gating = gen6_init_clock_gating;
7731 else if (IS_GEN5(dev_priv))
7732 dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
7733 else if (IS_G4X(dev_priv))
7734 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
7735 else if (IS_CRESTLINE(dev_priv))
7736 dev_priv->display.init_clock_gating = crestline_init_clock_gating;
7737 else if (IS_BROADWATER(dev_priv))
7738 dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
7739 else if (IS_GEN3(dev_priv))
7740 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
7741 else if (IS_I85X(dev_priv) || IS_I865G(dev_priv))
7742 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
7743 else if (IS_GEN2(dev_priv))
7744 dev_priv->display.init_clock_gating = i830_init_clock_gating;
7746 MISSING_CASE(INTEL_DEVID(dev_priv));
7747 dev_priv->display.init_clock_gating = nop_init_clock_gating;
7751 /* Set up chip specific power management-related functions */
7752 void intel_init_pm(struct drm_device *dev)
7754 struct drm_i915_private *dev_priv = to_i915(dev);
7756 intel_fbc_init(dev_priv);
7759 if (IS_PINEVIEW(dev))
7760 i915_pineview_get_mem_freq(dev);
7761 else if (IS_GEN5(dev))
7762 i915_ironlake_get_mem_freq(dev);
7764 /* For FIFO watermark updates */
7765 if (INTEL_INFO(dev)->gen >= 9) {
7766 skl_setup_wm_latency(dev);
7767 dev_priv->display.update_wm = skl_update_wm;
7768 dev_priv->display.compute_global_watermarks = skl_compute_wm;
7769 } else if (HAS_PCH_SPLIT(dev)) {
7770 ilk_setup_wm_latency(dev);
7772 if ((IS_GEN5(dev) && dev_priv->wm.pri_latency[1] &&
7773 dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
7774 (!IS_GEN5(dev) && dev_priv->wm.pri_latency[0] &&
7775 dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
7776 dev_priv->display.compute_pipe_wm = ilk_compute_pipe_wm;
7777 dev_priv->display.compute_intermediate_wm =
7778 ilk_compute_intermediate_wm;
7779 dev_priv->display.initial_watermarks =
7780 ilk_initial_watermarks;
7781 dev_priv->display.optimize_watermarks =
7782 ilk_optimize_watermarks;
7784 DRM_DEBUG_KMS("Failed to read display plane latency. "
7787 } else if (IS_CHERRYVIEW(dev)) {
7788 vlv_setup_wm_latency(dev);
7789 dev_priv->display.update_wm = vlv_update_wm;
7790 } else if (IS_VALLEYVIEW(dev)) {
7791 vlv_setup_wm_latency(dev);
7792 dev_priv->display.update_wm = vlv_update_wm;
7793 } else if (IS_PINEVIEW(dev)) {
7794 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
7797 dev_priv->mem_freq)) {
7798 DRM_INFO("failed to find known CxSR latency "
7799 "(found ddr%s fsb freq %d, mem freq %d), "
7801 (dev_priv->is_ddr3 == 1) ? "3" : "2",
7802 dev_priv->fsb_freq, dev_priv->mem_freq);
7803 /* Disable CxSR and never update its watermark again */
7804 intel_set_memory_cxsr(dev_priv, false);
7805 dev_priv->display.update_wm = NULL;
7807 dev_priv->display.update_wm = pineview_update_wm;
7808 } else if (IS_G4X(dev)) {
7809 dev_priv->display.update_wm = g4x_update_wm;
7810 } else if (IS_GEN4(dev)) {
7811 dev_priv->display.update_wm = i965_update_wm;
7812 } else if (IS_GEN3(dev)) {
7813 dev_priv->display.update_wm = i9xx_update_wm;
7814 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
7815 } else if (IS_GEN2(dev)) {
7816 if (INTEL_INFO(dev)->num_pipes == 1) {
7817 dev_priv->display.update_wm = i845_update_wm;
7818 dev_priv->display.get_fifo_size = i845_get_fifo_size;
7820 dev_priv->display.update_wm = i9xx_update_wm;
7821 dev_priv->display.get_fifo_size = i830_get_fifo_size;
7824 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
7828 static inline int gen6_check_mailbox_status(struct drm_i915_private *dev_priv)
7831 I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_ERROR_MASK;
7834 case GEN6_PCODE_SUCCESS:
7836 case GEN6_PCODE_UNIMPLEMENTED_CMD:
7837 case GEN6_PCODE_ILLEGAL_CMD:
7839 case GEN6_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
7840 case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
7842 case GEN6_PCODE_TIMEOUT:
7850 static inline int gen7_check_mailbox_status(struct drm_i915_private *dev_priv)
7853 I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_ERROR_MASK;
7856 case GEN6_PCODE_SUCCESS:
7858 case GEN6_PCODE_ILLEGAL_CMD:
7860 case GEN7_PCODE_TIMEOUT:
7862 case GEN7_PCODE_ILLEGAL_DATA:
7864 case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
7867 MISSING_CASE(flags);
7872 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u32 mbox, u32 *val)
7876 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7878 /* GEN6_PCODE_* are outside of the forcewake domain, we can
7879 * use te fw I915_READ variants to reduce the amount of work
7880 * required when reading/writing.
7883 if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
7884 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
7888 I915_WRITE_FW(GEN6_PCODE_DATA, *val);
7889 I915_WRITE_FW(GEN6_PCODE_DATA1, 0);
7890 I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
7892 if (intel_wait_for_register_fw(dev_priv,
7893 GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
7895 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
7899 *val = I915_READ_FW(GEN6_PCODE_DATA);
7900 I915_WRITE_FW(GEN6_PCODE_DATA, 0);
7902 if (INTEL_GEN(dev_priv) > 6)
7903 status = gen7_check_mailbox_status(dev_priv);
7905 status = gen6_check_mailbox_status(dev_priv);
7908 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed: %d\n",
7916 int sandybridge_pcode_write(struct drm_i915_private *dev_priv,
7921 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7923 /* GEN6_PCODE_* are outside of the forcewake domain, we can
7924 * use te fw I915_READ variants to reduce the amount of work
7925 * required when reading/writing.
7928 if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
7929 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
7933 I915_WRITE_FW(GEN6_PCODE_DATA, val);
7934 I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
7936 if (intel_wait_for_register_fw(dev_priv,
7937 GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
7939 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
7943 I915_WRITE_FW(GEN6_PCODE_DATA, 0);
7945 if (INTEL_GEN(dev_priv) > 6)
7946 status = gen7_check_mailbox_status(dev_priv);
7948 status = gen6_check_mailbox_status(dev_priv);
7951 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed: %d\n",
7959 static bool skl_pcode_try_request(struct drm_i915_private *dev_priv, u32 mbox,
7960 u32 request, u32 reply_mask, u32 reply,
7965 *status = sandybridge_pcode_read(dev_priv, mbox, &val);
7967 return *status || ((val & reply_mask) == reply);
7971 * skl_pcode_request - send PCODE request until acknowledgment
7972 * @dev_priv: device private
7973 * @mbox: PCODE mailbox ID the request is targeted for
7974 * @request: request ID
7975 * @reply_mask: mask used to check for request acknowledgment
7976 * @reply: value used to check for request acknowledgment
7977 * @timeout_base_ms: timeout for polling with preemption enabled
7979 * Keep resending the @request to @mbox until PCODE acknowledges it, PCODE
7980 * reports an error or an overall timeout of @timeout_base_ms+50 ms expires.
7981 * The request is acknowledged once the PCODE reply dword equals @reply after
7982 * applying @reply_mask. Polling is first attempted with preemption enabled
7983 * for @timeout_base_ms and if this times out for another 50 ms with
7984 * preemption disabled.
7986 * Returns 0 on success, %-ETIMEDOUT in case of a timeout, <0 in case of some
7987 * other error as reported by PCODE.
7989 int skl_pcode_request(struct drm_i915_private *dev_priv, u32 mbox, u32 request,
7990 u32 reply_mask, u32 reply, int timeout_base_ms)
7995 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7997 #define COND skl_pcode_try_request(dev_priv, mbox, request, reply_mask, reply, \
8001 * Prime the PCODE by doing a request first. Normally it guarantees
8002 * that a subsequent request, at most @timeout_base_ms later, succeeds.
8003 * _wait_for() doesn't guarantee when its passed condition is evaluated
8004 * first, so send the first request explicitly.
8010 ret = _wait_for(COND, timeout_base_ms * 1000, 10);
8015 * The above can time out if the number of requests was low (2 in the
8016 * worst case) _and_ PCODE was busy for some reason even after a
8017 * (queued) request and @timeout_base_ms delay. As a workaround retry
8018 * the poll with preemption disabled to maximize the number of
8019 * requests. Increase the timeout from @timeout_base_ms to 50ms to
8020 * account for interrupts that could reduce the number of these
8021 * requests, and for any quirks of the PCODE firmware that delays
8022 * the request completion.
8024 DRM_DEBUG_KMS("PCODE timeout, retrying with preemption disabled\n");
8025 WARN_ON_ONCE(timeout_base_ms > 3);
8027 ret = wait_for_atomic(COND, 50);
8031 return ret ? ret : status;
8035 static int byt_gpu_freq(struct drm_i915_private *dev_priv, int val)
8039 * Slow = Fast = GPLL ref * N
8041 return DIV_ROUND_CLOSEST(dev_priv->rps.gpll_ref_freq * (val - 0xb7), 1000);
8044 static int byt_freq_opcode(struct drm_i915_private *dev_priv, int val)
8046 return DIV_ROUND_CLOSEST(1000 * val, dev_priv->rps.gpll_ref_freq) + 0xb7;
8049 static int chv_gpu_freq(struct drm_i915_private *dev_priv, int val)
8053 * CU (slow) = CU2x (fast) / 2 = GPLL ref * N / 2
8055 return DIV_ROUND_CLOSEST(dev_priv->rps.gpll_ref_freq * val, 2 * 2 * 1000);
8058 static int chv_freq_opcode(struct drm_i915_private *dev_priv, int val)
8060 /* CHV needs even values */
8061 return DIV_ROUND_CLOSEST(2 * 1000 * val, dev_priv->rps.gpll_ref_freq) * 2;
8064 int intel_gpu_freq(struct drm_i915_private *dev_priv, int val)
8066 if (IS_GEN9(dev_priv))
8067 return DIV_ROUND_CLOSEST(val * GT_FREQUENCY_MULTIPLIER,
8069 else if (IS_CHERRYVIEW(dev_priv))
8070 return chv_gpu_freq(dev_priv, val);
8071 else if (IS_VALLEYVIEW(dev_priv))
8072 return byt_gpu_freq(dev_priv, val);
8074 return val * GT_FREQUENCY_MULTIPLIER;
8077 int intel_freq_opcode(struct drm_i915_private *dev_priv, int val)
8079 if (IS_GEN9(dev_priv))
8080 return DIV_ROUND_CLOSEST(val * GEN9_FREQ_SCALER,
8081 GT_FREQUENCY_MULTIPLIER);
8082 else if (IS_CHERRYVIEW(dev_priv))
8083 return chv_freq_opcode(dev_priv, val);
8084 else if (IS_VALLEYVIEW(dev_priv))
8085 return byt_freq_opcode(dev_priv, val);
8087 return DIV_ROUND_CLOSEST(val, GT_FREQUENCY_MULTIPLIER);
8090 struct request_boost {
8091 struct work_struct work;
8092 struct drm_i915_gem_request *req;
8095 static void __intel_rps_boost_work(struct work_struct *work)
8097 struct request_boost *boost = container_of(work, struct request_boost, work);
8098 struct drm_i915_gem_request *req = boost->req;
8100 if (!i915_gem_request_completed(req))
8101 gen6_rps_boost(req->i915, NULL, req->emitted_jiffies);
8103 i915_gem_request_put(req);
8107 void intel_queue_rps_boost_for_request(struct drm_i915_gem_request *req)
8109 struct request_boost *boost;
8111 if (req == NULL || INTEL_GEN(req->i915) < 6)
8114 if (i915_gem_request_completed(req))
8117 boost = kmalloc(sizeof(*boost), GFP_ATOMIC);
8121 boost->req = i915_gem_request_get(req);
8123 INIT_WORK(&boost->work, __intel_rps_boost_work);
8124 queue_work(req->i915->wq, &boost->work);
8127 void intel_pm_setup(struct drm_device *dev)
8129 struct drm_i915_private *dev_priv = to_i915(dev);
8131 mutex_init(&dev_priv->rps.hw_lock);
8132 spin_lock_init(&dev_priv->rps.client_lock);
8134 INIT_DELAYED_WORK(&dev_priv->rps.autoenable_work,
8135 __intel_autoenable_gt_powersave);
8136 INIT_LIST_HEAD(&dev_priv->rps.clients);
8138 dev_priv->pm.suspended = false;
8139 atomic_set(&dev_priv->pm.wakeref_count, 0);
8140 atomic_set(&dev_priv->pm.atomic_seq, 0);