2 * Copyright © 2012 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eugeni Dodonov <eugeni.dodonov@intel.com>
28 #include <linux/cpufreq.h>
30 #include "intel_drv.h"
31 #include "../../../platform/x86/intel_ips.h"
32 #include <linux/module.h>
35 * RC6 is a special power stage which allows the GPU to enter an very
36 * low-voltage mode when idle, using down to 0V while at this stage. This
37 * stage is entered automatically when the GPU is idle when RC6 support is
38 * enabled, and as soon as new workload arises GPU wakes up automatically as well.
40 * There are different RC6 modes available in Intel GPU, which differentiate
41 * among each other with the latency required to enter and leave RC6 and
42 * voltage consumed by the GPU in different states.
44 * The combination of the following flags define which states GPU is allowed
45 * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
46 * RC6pp is deepest RC6. Their support by hardware varies according to the
47 * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
48 * which brings the most power savings; deeper states save more power, but
49 * require higher latency to switch to and wake up.
51 #define INTEL_RC6_ENABLE (1<<0)
52 #define INTEL_RC6p_ENABLE (1<<1)
53 #define INTEL_RC6pp_ENABLE (1<<2)
55 static void bxt_init_clock_gating(struct drm_device
*dev
)
57 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
59 /* WaDisableSDEUnitClockGating:bxt */
60 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
61 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
65 * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
67 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
68 GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ
);
71 * Wa: Backlight PWM may stop in the asserted state, causing backlight
74 if (IS_BXT_REVID(dev_priv
, BXT_REVID_B0
, REVID_FOREVER
))
75 I915_WRITE(GEN9_CLKGATE_DIS_0
, I915_READ(GEN9_CLKGATE_DIS_0
) |
76 PWM1_GATING_DIS
| PWM2_GATING_DIS
);
79 static void i915_pineview_get_mem_freq(struct drm_device
*dev
)
81 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
84 tmp
= I915_READ(CLKCFG
);
86 switch (tmp
& CLKCFG_FSB_MASK
) {
88 dev_priv
->fsb_freq
= 533; /* 133*4 */
91 dev_priv
->fsb_freq
= 800; /* 200*4 */
94 dev_priv
->fsb_freq
= 667; /* 167*4 */
97 dev_priv
->fsb_freq
= 400; /* 100*4 */
101 switch (tmp
& CLKCFG_MEM_MASK
) {
103 dev_priv
->mem_freq
= 533;
106 dev_priv
->mem_freq
= 667;
109 dev_priv
->mem_freq
= 800;
113 /* detect pineview DDR3 setting */
114 tmp
= I915_READ(CSHRDDR3CTL
);
115 dev_priv
->is_ddr3
= (tmp
& CSHRDDR3CTL_DDR3
) ? 1 : 0;
118 static void i915_ironlake_get_mem_freq(struct drm_device
*dev
)
120 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
123 ddrpll
= I915_READ16(DDRMPLL1
);
124 csipll
= I915_READ16(CSIPLL0
);
126 switch (ddrpll
& 0xff) {
128 dev_priv
->mem_freq
= 800;
131 dev_priv
->mem_freq
= 1066;
134 dev_priv
->mem_freq
= 1333;
137 dev_priv
->mem_freq
= 1600;
140 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
142 dev_priv
->mem_freq
= 0;
146 dev_priv
->ips
.r_t
= dev_priv
->mem_freq
;
148 switch (csipll
& 0x3ff) {
150 dev_priv
->fsb_freq
= 3200;
153 dev_priv
->fsb_freq
= 3733;
156 dev_priv
->fsb_freq
= 4266;
159 dev_priv
->fsb_freq
= 4800;
162 dev_priv
->fsb_freq
= 5333;
165 dev_priv
->fsb_freq
= 5866;
168 dev_priv
->fsb_freq
= 6400;
171 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
173 dev_priv
->fsb_freq
= 0;
177 if (dev_priv
->fsb_freq
== 3200) {
178 dev_priv
->ips
.c_m
= 0;
179 } else if (dev_priv
->fsb_freq
> 3200 && dev_priv
->fsb_freq
<= 4800) {
180 dev_priv
->ips
.c_m
= 1;
182 dev_priv
->ips
.c_m
= 2;
186 static const struct cxsr_latency cxsr_latency_table
[] = {
187 {1, 0, 800, 400, 3382, 33382, 3983, 33983}, /* DDR2-400 SC */
188 {1, 0, 800, 667, 3354, 33354, 3807, 33807}, /* DDR2-667 SC */
189 {1, 0, 800, 800, 3347, 33347, 3763, 33763}, /* DDR2-800 SC */
190 {1, 1, 800, 667, 6420, 36420, 6873, 36873}, /* DDR3-667 SC */
191 {1, 1, 800, 800, 5902, 35902, 6318, 36318}, /* DDR3-800 SC */
193 {1, 0, 667, 400, 3400, 33400, 4021, 34021}, /* DDR2-400 SC */
194 {1, 0, 667, 667, 3372, 33372, 3845, 33845}, /* DDR2-667 SC */
195 {1, 0, 667, 800, 3386, 33386, 3822, 33822}, /* DDR2-800 SC */
196 {1, 1, 667, 667, 6438, 36438, 6911, 36911}, /* DDR3-667 SC */
197 {1, 1, 667, 800, 5941, 35941, 6377, 36377}, /* DDR3-800 SC */
199 {1, 0, 400, 400, 3472, 33472, 4173, 34173}, /* DDR2-400 SC */
200 {1, 0, 400, 667, 3443, 33443, 3996, 33996}, /* DDR2-667 SC */
201 {1, 0, 400, 800, 3430, 33430, 3946, 33946}, /* DDR2-800 SC */
202 {1, 1, 400, 667, 6509, 36509, 7062, 37062}, /* DDR3-667 SC */
203 {1, 1, 400, 800, 5985, 35985, 6501, 36501}, /* DDR3-800 SC */
205 {0, 0, 800, 400, 3438, 33438, 4065, 34065}, /* DDR2-400 SC */
206 {0, 0, 800, 667, 3410, 33410, 3889, 33889}, /* DDR2-667 SC */
207 {0, 0, 800, 800, 3403, 33403, 3845, 33845}, /* DDR2-800 SC */
208 {0, 1, 800, 667, 6476, 36476, 6955, 36955}, /* DDR3-667 SC */
209 {0, 1, 800, 800, 5958, 35958, 6400, 36400}, /* DDR3-800 SC */
211 {0, 0, 667, 400, 3456, 33456, 4103, 34106}, /* DDR2-400 SC */
212 {0, 0, 667, 667, 3428, 33428, 3927, 33927}, /* DDR2-667 SC */
213 {0, 0, 667, 800, 3443, 33443, 3905, 33905}, /* DDR2-800 SC */
214 {0, 1, 667, 667, 6494, 36494, 6993, 36993}, /* DDR3-667 SC */
215 {0, 1, 667, 800, 5998, 35998, 6460, 36460}, /* DDR3-800 SC */
217 {0, 0, 400, 400, 3528, 33528, 4255, 34255}, /* DDR2-400 SC */
218 {0, 0, 400, 667, 3500, 33500, 4079, 34079}, /* DDR2-667 SC */
219 {0, 0, 400, 800, 3487, 33487, 4029, 34029}, /* DDR2-800 SC */
220 {0, 1, 400, 667, 6566, 36566, 7145, 37145}, /* DDR3-667 SC */
221 {0, 1, 400, 800, 6042, 36042, 6584, 36584}, /* DDR3-800 SC */
224 static const struct cxsr_latency
*intel_get_cxsr_latency(int is_desktop
,
229 const struct cxsr_latency
*latency
;
232 if (fsb
== 0 || mem
== 0)
235 for (i
= 0; i
< ARRAY_SIZE(cxsr_latency_table
); i
++) {
236 latency
= &cxsr_latency_table
[i
];
237 if (is_desktop
== latency
->is_desktop
&&
238 is_ddr3
== latency
->is_ddr3
&&
239 fsb
== latency
->fsb_freq
&& mem
== latency
->mem_freq
)
243 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
248 static void chv_set_memory_dvfs(struct drm_i915_private
*dev_priv
, bool enable
)
252 mutex_lock(&dev_priv
->rps
.hw_lock
);
254 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
);
256 val
&= ~FORCE_DDR_HIGH_FREQ
;
258 val
|= FORCE_DDR_HIGH_FREQ
;
259 val
&= ~FORCE_DDR_LOW_FREQ
;
260 val
|= FORCE_DDR_FREQ_REQ_ACK
;
261 vlv_punit_write(dev_priv
, PUNIT_REG_DDR_SETUP2
, val
);
263 if (wait_for((vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
) &
264 FORCE_DDR_FREQ_REQ_ACK
) == 0, 3))
265 DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
267 mutex_unlock(&dev_priv
->rps
.hw_lock
);
270 static void chv_set_memory_pm5(struct drm_i915_private
*dev_priv
, bool enable
)
274 mutex_lock(&dev_priv
->rps
.hw_lock
);
276 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DSPFREQ
);
278 val
|= DSP_MAXFIFO_PM5_ENABLE
;
280 val
&= ~DSP_MAXFIFO_PM5_ENABLE
;
281 vlv_punit_write(dev_priv
, PUNIT_REG_DSPFREQ
, val
);
283 mutex_unlock(&dev_priv
->rps
.hw_lock
);
286 #define FW_WM(value, plane) \
287 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
289 void intel_set_memory_cxsr(struct drm_i915_private
*dev_priv
, bool enable
)
291 struct drm_device
*dev
= dev_priv
->dev
;
294 if (IS_VALLEYVIEW(dev
) || IS_CHERRYVIEW(dev
)) {
295 I915_WRITE(FW_BLC_SELF_VLV
, enable
? FW_CSPWRDWNEN
: 0);
296 POSTING_READ(FW_BLC_SELF_VLV
);
297 dev_priv
->wm
.vlv
.cxsr
= enable
;
298 } else if (IS_G4X(dev
) || IS_CRESTLINE(dev
)) {
299 I915_WRITE(FW_BLC_SELF
, enable
? FW_BLC_SELF_EN
: 0);
300 POSTING_READ(FW_BLC_SELF
);
301 } else if (IS_PINEVIEW(dev
)) {
302 val
= I915_READ(DSPFW3
) & ~PINEVIEW_SELF_REFRESH_EN
;
303 val
|= enable
? PINEVIEW_SELF_REFRESH_EN
: 0;
304 I915_WRITE(DSPFW3
, val
);
305 POSTING_READ(DSPFW3
);
306 } else if (IS_I945G(dev
) || IS_I945GM(dev
)) {
307 val
= enable
? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN
) :
308 _MASKED_BIT_DISABLE(FW_BLC_SELF_EN
);
309 I915_WRITE(FW_BLC_SELF
, val
);
310 POSTING_READ(FW_BLC_SELF
);
311 } else if (IS_I915GM(dev
)) {
312 val
= enable
? _MASKED_BIT_ENABLE(INSTPM_SELF_EN
) :
313 _MASKED_BIT_DISABLE(INSTPM_SELF_EN
);
314 I915_WRITE(INSTPM
, val
);
315 POSTING_READ(INSTPM
);
320 DRM_DEBUG_KMS("memory self-refresh is %s\n",
321 enable
? "enabled" : "disabled");
326 * Latency for FIFO fetches is dependent on several factors:
327 * - memory configuration (speed, channels)
329 * - current MCH state
330 * It can be fairly high in some situations, so here we assume a fairly
331 * pessimal value. It's a tradeoff between extra memory fetches (if we
332 * set this value too high, the FIFO will fetch frequently to stay full)
333 * and power consumption (set it too low to save power and we might see
334 * FIFO underruns and display "flicker").
336 * A value of 5us seems to be a good balance; safe for very low end
337 * platforms but not overly aggressive on lower latency configs.
339 static const int pessimal_latency_ns
= 5000;
341 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
342 ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
344 static int vlv_get_fifo_size(struct drm_device
*dev
,
345 enum pipe pipe
, int plane
)
347 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
348 int sprite0_start
, sprite1_start
, size
;
351 uint32_t dsparb
, dsparb2
, dsparb3
;
353 dsparb
= I915_READ(DSPARB
);
354 dsparb2
= I915_READ(DSPARB2
);
355 sprite0_start
= VLV_FIFO_START(dsparb
, dsparb2
, 0, 0);
356 sprite1_start
= VLV_FIFO_START(dsparb
, dsparb2
, 8, 4);
359 dsparb
= I915_READ(DSPARB
);
360 dsparb2
= I915_READ(DSPARB2
);
361 sprite0_start
= VLV_FIFO_START(dsparb
, dsparb2
, 16, 8);
362 sprite1_start
= VLV_FIFO_START(dsparb
, dsparb2
, 24, 12);
365 dsparb2
= I915_READ(DSPARB2
);
366 dsparb3
= I915_READ(DSPARB3
);
367 sprite0_start
= VLV_FIFO_START(dsparb3
, dsparb2
, 0, 16);
368 sprite1_start
= VLV_FIFO_START(dsparb3
, dsparb2
, 8, 20);
376 size
= sprite0_start
;
379 size
= sprite1_start
- sprite0_start
;
382 size
= 512 - 1 - sprite1_start
;
388 DRM_DEBUG_KMS("Pipe %c %s %c FIFO size: %d\n",
389 pipe_name(pipe
), plane
== 0 ? "primary" : "sprite",
390 plane
== 0 ? plane_name(pipe
) : sprite_name(pipe
, plane
- 1),
396 static int i9xx_get_fifo_size(struct drm_device
*dev
, int plane
)
398 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
399 uint32_t dsparb
= I915_READ(DSPARB
);
402 size
= dsparb
& 0x7f;
404 size
= ((dsparb
>> DSPARB_CSTART_SHIFT
) & 0x7f) - size
;
406 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb
,
407 plane
? "B" : "A", size
);
412 static int i830_get_fifo_size(struct drm_device
*dev
, int plane
)
414 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
415 uint32_t dsparb
= I915_READ(DSPARB
);
418 size
= dsparb
& 0x1ff;
420 size
= ((dsparb
>> DSPARB_BEND_SHIFT
) & 0x1ff) - size
;
421 size
>>= 1; /* Convert to cachelines */
423 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb
,
424 plane
? "B" : "A", size
);
429 static int i845_get_fifo_size(struct drm_device
*dev
, int plane
)
431 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
432 uint32_t dsparb
= I915_READ(DSPARB
);
435 size
= dsparb
& 0x7f;
436 size
>>= 2; /* Convert to cachelines */
438 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb
,
445 /* Pineview has different values for various configs */
446 static const struct intel_watermark_params pineview_display_wm
= {
447 .fifo_size
= PINEVIEW_DISPLAY_FIFO
,
448 .max_wm
= PINEVIEW_MAX_WM
,
449 .default_wm
= PINEVIEW_DFT_WM
,
450 .guard_size
= PINEVIEW_GUARD_WM
,
451 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
453 static const struct intel_watermark_params pineview_display_hplloff_wm
= {
454 .fifo_size
= PINEVIEW_DISPLAY_FIFO
,
455 .max_wm
= PINEVIEW_MAX_WM
,
456 .default_wm
= PINEVIEW_DFT_HPLLOFF_WM
,
457 .guard_size
= PINEVIEW_GUARD_WM
,
458 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
460 static const struct intel_watermark_params pineview_cursor_wm
= {
461 .fifo_size
= PINEVIEW_CURSOR_FIFO
,
462 .max_wm
= PINEVIEW_CURSOR_MAX_WM
,
463 .default_wm
= PINEVIEW_CURSOR_DFT_WM
,
464 .guard_size
= PINEVIEW_CURSOR_GUARD_WM
,
465 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
467 static const struct intel_watermark_params pineview_cursor_hplloff_wm
= {
468 .fifo_size
= PINEVIEW_CURSOR_FIFO
,
469 .max_wm
= PINEVIEW_CURSOR_MAX_WM
,
470 .default_wm
= PINEVIEW_CURSOR_DFT_WM
,
471 .guard_size
= PINEVIEW_CURSOR_GUARD_WM
,
472 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
474 static const struct intel_watermark_params g4x_wm_info
= {
475 .fifo_size
= G4X_FIFO_SIZE
,
476 .max_wm
= G4X_MAX_WM
,
477 .default_wm
= G4X_MAX_WM
,
479 .cacheline_size
= G4X_FIFO_LINE_SIZE
,
481 static const struct intel_watermark_params g4x_cursor_wm_info
= {
482 .fifo_size
= I965_CURSOR_FIFO
,
483 .max_wm
= I965_CURSOR_MAX_WM
,
484 .default_wm
= I965_CURSOR_DFT_WM
,
486 .cacheline_size
= G4X_FIFO_LINE_SIZE
,
488 static const struct intel_watermark_params valleyview_wm_info
= {
489 .fifo_size
= VALLEYVIEW_FIFO_SIZE
,
490 .max_wm
= VALLEYVIEW_MAX_WM
,
491 .default_wm
= VALLEYVIEW_MAX_WM
,
493 .cacheline_size
= G4X_FIFO_LINE_SIZE
,
495 static const struct intel_watermark_params valleyview_cursor_wm_info
= {
496 .fifo_size
= I965_CURSOR_FIFO
,
497 .max_wm
= VALLEYVIEW_CURSOR_MAX_WM
,
498 .default_wm
= I965_CURSOR_DFT_WM
,
500 .cacheline_size
= G4X_FIFO_LINE_SIZE
,
502 static const struct intel_watermark_params i965_cursor_wm_info
= {
503 .fifo_size
= I965_CURSOR_FIFO
,
504 .max_wm
= I965_CURSOR_MAX_WM
,
505 .default_wm
= I965_CURSOR_DFT_WM
,
507 .cacheline_size
= I915_FIFO_LINE_SIZE
,
509 static const struct intel_watermark_params i945_wm_info
= {
510 .fifo_size
= I945_FIFO_SIZE
,
511 .max_wm
= I915_MAX_WM
,
514 .cacheline_size
= I915_FIFO_LINE_SIZE
,
516 static const struct intel_watermark_params i915_wm_info
= {
517 .fifo_size
= I915_FIFO_SIZE
,
518 .max_wm
= I915_MAX_WM
,
521 .cacheline_size
= I915_FIFO_LINE_SIZE
,
523 static const struct intel_watermark_params i830_a_wm_info
= {
524 .fifo_size
= I855GM_FIFO_SIZE
,
525 .max_wm
= I915_MAX_WM
,
528 .cacheline_size
= I830_FIFO_LINE_SIZE
,
530 static const struct intel_watermark_params i830_bc_wm_info
= {
531 .fifo_size
= I855GM_FIFO_SIZE
,
532 .max_wm
= I915_MAX_WM
/2,
535 .cacheline_size
= I830_FIFO_LINE_SIZE
,
537 static const struct intel_watermark_params i845_wm_info
= {
538 .fifo_size
= I830_FIFO_SIZE
,
539 .max_wm
= I915_MAX_WM
,
542 .cacheline_size
= I830_FIFO_LINE_SIZE
,
546 * intel_calculate_wm - calculate watermark level
547 * @clock_in_khz: pixel clock
548 * @wm: chip FIFO params
549 * @pixel_size: display pixel size
550 * @latency_ns: memory latency for the platform
552 * Calculate the watermark level (the level at which the display plane will
553 * start fetching from memory again). Each chip has a different display
554 * FIFO size and allocation, so the caller needs to figure that out and pass
555 * in the correct intel_watermark_params structure.
557 * As the pixel clock runs, the FIFO will be drained at a rate that depends
558 * on the pixel size. When it reaches the watermark level, it'll start
559 * fetching FIFO line sized based chunks from memory until the FIFO fills
560 * past the watermark point. If the FIFO drains completely, a FIFO underrun
561 * will occur, and a display engine hang could result.
563 static unsigned long intel_calculate_wm(unsigned long clock_in_khz
,
564 const struct intel_watermark_params
*wm
,
567 unsigned long latency_ns
)
569 long entries_required
, wm_size
;
572 * Note: we need to make sure we don't overflow for various clock &
574 * clocks go from a few thousand to several hundred thousand.
575 * latency is usually a few thousand
577 entries_required
= ((clock_in_khz
/ 1000) * pixel_size
* latency_ns
) /
579 entries_required
= DIV_ROUND_UP(entries_required
, wm
->cacheline_size
);
581 DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required
);
583 wm_size
= fifo_size
- (entries_required
+ wm
->guard_size
);
585 DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size
);
587 /* Don't promote wm_size to unsigned... */
588 if (wm_size
> (long)wm
->max_wm
)
589 wm_size
= wm
->max_wm
;
591 wm_size
= wm
->default_wm
;
594 * Bspec seems to indicate that the value shouldn't be lower than
595 * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
596 * Lets go for 8 which is the burst size since certain platforms
597 * already use a hardcoded 8 (which is what the spec says should be
606 static struct drm_crtc
*single_enabled_crtc(struct drm_device
*dev
)
608 struct drm_crtc
*crtc
, *enabled
= NULL
;
610 for_each_crtc(dev
, crtc
) {
611 if (intel_crtc_active(crtc
)) {
621 static void pineview_update_wm(struct drm_crtc
*unused_crtc
)
623 struct drm_device
*dev
= unused_crtc
->dev
;
624 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
625 struct drm_crtc
*crtc
;
626 const struct cxsr_latency
*latency
;
630 latency
= intel_get_cxsr_latency(IS_PINEVIEW_G(dev
), dev_priv
->is_ddr3
,
631 dev_priv
->fsb_freq
, dev_priv
->mem_freq
);
633 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
634 intel_set_memory_cxsr(dev_priv
, false);
638 crtc
= single_enabled_crtc(dev
);
640 const struct drm_display_mode
*adjusted_mode
= &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
641 int pixel_size
= crtc
->primary
->state
->fb
->bits_per_pixel
/ 8;
642 int clock
= adjusted_mode
->crtc_clock
;
645 wm
= intel_calculate_wm(clock
, &pineview_display_wm
,
646 pineview_display_wm
.fifo_size
,
647 pixel_size
, latency
->display_sr
);
648 reg
= I915_READ(DSPFW1
);
649 reg
&= ~DSPFW_SR_MASK
;
650 reg
|= FW_WM(wm
, SR
);
651 I915_WRITE(DSPFW1
, reg
);
652 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg
);
655 wm
= intel_calculate_wm(clock
, &pineview_cursor_wm
,
656 pineview_display_wm
.fifo_size
,
657 pixel_size
, latency
->cursor_sr
);
658 reg
= I915_READ(DSPFW3
);
659 reg
&= ~DSPFW_CURSOR_SR_MASK
;
660 reg
|= FW_WM(wm
, CURSOR_SR
);
661 I915_WRITE(DSPFW3
, reg
);
663 /* Display HPLL off SR */
664 wm
= intel_calculate_wm(clock
, &pineview_display_hplloff_wm
,
665 pineview_display_hplloff_wm
.fifo_size
,
666 pixel_size
, latency
->display_hpll_disable
);
667 reg
= I915_READ(DSPFW3
);
668 reg
&= ~DSPFW_HPLL_SR_MASK
;
669 reg
|= FW_WM(wm
, HPLL_SR
);
670 I915_WRITE(DSPFW3
, reg
);
672 /* cursor HPLL off SR */
673 wm
= intel_calculate_wm(clock
, &pineview_cursor_hplloff_wm
,
674 pineview_display_hplloff_wm
.fifo_size
,
675 pixel_size
, latency
->cursor_hpll_disable
);
676 reg
= I915_READ(DSPFW3
);
677 reg
&= ~DSPFW_HPLL_CURSOR_MASK
;
678 reg
|= FW_WM(wm
, HPLL_CURSOR
);
679 I915_WRITE(DSPFW3
, reg
);
680 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg
);
682 intel_set_memory_cxsr(dev_priv
, true);
684 intel_set_memory_cxsr(dev_priv
, false);
688 static bool g4x_compute_wm0(struct drm_device
*dev
,
690 const struct intel_watermark_params
*display
,
691 int display_latency_ns
,
692 const struct intel_watermark_params
*cursor
,
693 int cursor_latency_ns
,
697 struct drm_crtc
*crtc
;
698 const struct drm_display_mode
*adjusted_mode
;
699 int htotal
, hdisplay
, clock
, pixel_size
;
700 int line_time_us
, line_count
;
701 int entries
, tlb_miss
;
703 crtc
= intel_get_crtc_for_plane(dev
, plane
);
704 if (!intel_crtc_active(crtc
)) {
705 *cursor_wm
= cursor
->guard_size
;
706 *plane_wm
= display
->guard_size
;
710 adjusted_mode
= &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
711 clock
= adjusted_mode
->crtc_clock
;
712 htotal
= adjusted_mode
->crtc_htotal
;
713 hdisplay
= to_intel_crtc(crtc
)->config
->pipe_src_w
;
714 pixel_size
= crtc
->primary
->state
->fb
->bits_per_pixel
/ 8;
716 /* Use the small buffer method to calculate plane watermark */
717 entries
= ((clock
* pixel_size
/ 1000) * display_latency_ns
) / 1000;
718 tlb_miss
= display
->fifo_size
*display
->cacheline_size
- hdisplay
* 8;
721 entries
= DIV_ROUND_UP(entries
, display
->cacheline_size
);
722 *plane_wm
= entries
+ display
->guard_size
;
723 if (*plane_wm
> (int)display
->max_wm
)
724 *plane_wm
= display
->max_wm
;
726 /* Use the large buffer method to calculate cursor watermark */
727 line_time_us
= max(htotal
* 1000 / clock
, 1);
728 line_count
= (cursor_latency_ns
/ line_time_us
+ 1000) / 1000;
729 entries
= line_count
* crtc
->cursor
->state
->crtc_w
* pixel_size
;
730 tlb_miss
= cursor
->fifo_size
*cursor
->cacheline_size
- hdisplay
* 8;
733 entries
= DIV_ROUND_UP(entries
, cursor
->cacheline_size
);
734 *cursor_wm
= entries
+ cursor
->guard_size
;
735 if (*cursor_wm
> (int)cursor
->max_wm
)
736 *cursor_wm
= (int)cursor
->max_wm
;
742 * Check the wm result.
744 * If any calculated watermark values is larger than the maximum value that
745 * can be programmed into the associated watermark register, that watermark
748 static bool g4x_check_srwm(struct drm_device
*dev
,
749 int display_wm
, int cursor_wm
,
750 const struct intel_watermark_params
*display
,
751 const struct intel_watermark_params
*cursor
)
753 DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
754 display_wm
, cursor_wm
);
756 if (display_wm
> display
->max_wm
) {
757 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
758 display_wm
, display
->max_wm
);
762 if (cursor_wm
> cursor
->max_wm
) {
763 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
764 cursor_wm
, cursor
->max_wm
);
768 if (!(display_wm
|| cursor_wm
)) {
769 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
776 static bool g4x_compute_srwm(struct drm_device
*dev
,
779 const struct intel_watermark_params
*display
,
780 const struct intel_watermark_params
*cursor
,
781 int *display_wm
, int *cursor_wm
)
783 struct drm_crtc
*crtc
;
784 const struct drm_display_mode
*adjusted_mode
;
785 int hdisplay
, htotal
, pixel_size
, clock
;
786 unsigned long line_time_us
;
787 int line_count
, line_size
;
792 *display_wm
= *cursor_wm
= 0;
796 crtc
= intel_get_crtc_for_plane(dev
, plane
);
797 adjusted_mode
= &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
798 clock
= adjusted_mode
->crtc_clock
;
799 htotal
= adjusted_mode
->crtc_htotal
;
800 hdisplay
= to_intel_crtc(crtc
)->config
->pipe_src_w
;
801 pixel_size
= crtc
->primary
->state
->fb
->bits_per_pixel
/ 8;
803 line_time_us
= max(htotal
* 1000 / clock
, 1);
804 line_count
= (latency_ns
/ line_time_us
+ 1000) / 1000;
805 line_size
= hdisplay
* pixel_size
;
807 /* Use the minimum of the small and large buffer method for primary */
808 small
= ((clock
* pixel_size
/ 1000) * latency_ns
) / 1000;
809 large
= line_count
* line_size
;
811 entries
= DIV_ROUND_UP(min(small
, large
), display
->cacheline_size
);
812 *display_wm
= entries
+ display
->guard_size
;
814 /* calculate the self-refresh watermark for display cursor */
815 entries
= line_count
* pixel_size
* crtc
->cursor
->state
->crtc_w
;
816 entries
= DIV_ROUND_UP(entries
, cursor
->cacheline_size
);
817 *cursor_wm
= entries
+ cursor
->guard_size
;
819 return g4x_check_srwm(dev
,
820 *display_wm
, *cursor_wm
,
824 #define FW_WM_VLV(value, plane) \
825 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
827 static void vlv_write_wm_values(struct intel_crtc
*crtc
,
828 const struct vlv_wm_values
*wm
)
830 struct drm_i915_private
*dev_priv
= to_i915(crtc
->base
.dev
);
831 enum pipe pipe
= crtc
->pipe
;
833 I915_WRITE(VLV_DDL(pipe
),
834 (wm
->ddl
[pipe
].cursor
<< DDL_CURSOR_SHIFT
) |
835 (wm
->ddl
[pipe
].sprite
[1] << DDL_SPRITE_SHIFT(1)) |
836 (wm
->ddl
[pipe
].sprite
[0] << DDL_SPRITE_SHIFT(0)) |
837 (wm
->ddl
[pipe
].primary
<< DDL_PLANE_SHIFT
));
840 FW_WM(wm
->sr
.plane
, SR
) |
841 FW_WM(wm
->pipe
[PIPE_B
].cursor
, CURSORB
) |
842 FW_WM_VLV(wm
->pipe
[PIPE_B
].primary
, PLANEB
) |
843 FW_WM_VLV(wm
->pipe
[PIPE_A
].primary
, PLANEA
));
845 FW_WM_VLV(wm
->pipe
[PIPE_A
].sprite
[1], SPRITEB
) |
846 FW_WM(wm
->pipe
[PIPE_A
].cursor
, CURSORA
) |
847 FW_WM_VLV(wm
->pipe
[PIPE_A
].sprite
[0], SPRITEA
));
849 FW_WM(wm
->sr
.cursor
, CURSOR_SR
));
851 if (IS_CHERRYVIEW(dev_priv
)) {
852 I915_WRITE(DSPFW7_CHV
,
853 FW_WM_VLV(wm
->pipe
[PIPE_B
].sprite
[1], SPRITED
) |
854 FW_WM_VLV(wm
->pipe
[PIPE_B
].sprite
[0], SPRITEC
));
855 I915_WRITE(DSPFW8_CHV
,
856 FW_WM_VLV(wm
->pipe
[PIPE_C
].sprite
[1], SPRITEF
) |
857 FW_WM_VLV(wm
->pipe
[PIPE_C
].sprite
[0], SPRITEE
));
858 I915_WRITE(DSPFW9_CHV
,
859 FW_WM_VLV(wm
->pipe
[PIPE_C
].primary
, PLANEC
) |
860 FW_WM(wm
->pipe
[PIPE_C
].cursor
, CURSORC
));
862 FW_WM(wm
->sr
.plane
>> 9, SR_HI
) |
863 FW_WM(wm
->pipe
[PIPE_C
].sprite
[1] >> 8, SPRITEF_HI
) |
864 FW_WM(wm
->pipe
[PIPE_C
].sprite
[0] >> 8, SPRITEE_HI
) |
865 FW_WM(wm
->pipe
[PIPE_C
].primary
>> 8, PLANEC_HI
) |
866 FW_WM(wm
->pipe
[PIPE_B
].sprite
[1] >> 8, SPRITED_HI
) |
867 FW_WM(wm
->pipe
[PIPE_B
].sprite
[0] >> 8, SPRITEC_HI
) |
868 FW_WM(wm
->pipe
[PIPE_B
].primary
>> 8, PLANEB_HI
) |
869 FW_WM(wm
->pipe
[PIPE_A
].sprite
[1] >> 8, SPRITEB_HI
) |
870 FW_WM(wm
->pipe
[PIPE_A
].sprite
[0] >> 8, SPRITEA_HI
) |
871 FW_WM(wm
->pipe
[PIPE_A
].primary
>> 8, PLANEA_HI
));
874 FW_WM_VLV(wm
->pipe
[PIPE_B
].sprite
[1], SPRITED
) |
875 FW_WM_VLV(wm
->pipe
[PIPE_B
].sprite
[0], SPRITEC
));
877 FW_WM(wm
->sr
.plane
>> 9, SR_HI
) |
878 FW_WM(wm
->pipe
[PIPE_B
].sprite
[1] >> 8, SPRITED_HI
) |
879 FW_WM(wm
->pipe
[PIPE_B
].sprite
[0] >> 8, SPRITEC_HI
) |
880 FW_WM(wm
->pipe
[PIPE_B
].primary
>> 8, PLANEB_HI
) |
881 FW_WM(wm
->pipe
[PIPE_A
].sprite
[1] >> 8, SPRITEB_HI
) |
882 FW_WM(wm
->pipe
[PIPE_A
].sprite
[0] >> 8, SPRITEA_HI
) |
883 FW_WM(wm
->pipe
[PIPE_A
].primary
>> 8, PLANEA_HI
));
886 /* zero (unused) WM1 watermarks */
887 I915_WRITE(DSPFW4
, 0);
888 I915_WRITE(DSPFW5
, 0);
889 I915_WRITE(DSPFW6
, 0);
890 I915_WRITE(DSPHOWM1
, 0);
892 POSTING_READ(DSPFW1
);
900 VLV_WM_LEVEL_DDR_DVFS
,
903 /* latency must be in 0.1us units. */
904 static unsigned int vlv_wm_method2(unsigned int pixel_rate
,
905 unsigned int pipe_htotal
,
906 unsigned int horiz_pixels
,
907 unsigned int bytes_per_pixel
,
908 unsigned int latency
)
912 ret
= (latency
* pixel_rate
) / (pipe_htotal
* 10000);
913 ret
= (ret
+ 1) * horiz_pixels
* bytes_per_pixel
;
914 ret
= DIV_ROUND_UP(ret
, 64);
919 static void vlv_setup_wm_latency(struct drm_device
*dev
)
921 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
923 /* all latencies in usec */
924 dev_priv
->wm
.pri_latency
[VLV_WM_LEVEL_PM2
] = 3;
926 dev_priv
->wm
.max_level
= VLV_WM_LEVEL_PM2
;
928 if (IS_CHERRYVIEW(dev_priv
)) {
929 dev_priv
->wm
.pri_latency
[VLV_WM_LEVEL_PM5
] = 12;
930 dev_priv
->wm
.pri_latency
[VLV_WM_LEVEL_DDR_DVFS
] = 33;
932 dev_priv
->wm
.max_level
= VLV_WM_LEVEL_DDR_DVFS
;
936 static uint16_t vlv_compute_wm_level(struct intel_plane
*plane
,
937 struct intel_crtc
*crtc
,
938 const struct intel_plane_state
*state
,
941 struct drm_i915_private
*dev_priv
= to_i915(plane
->base
.dev
);
942 int clock
, htotal
, pixel_size
, width
, wm
;
944 if (dev_priv
->wm
.pri_latency
[level
] == 0)
950 pixel_size
= drm_format_plane_cpp(state
->base
.fb
->pixel_format
, 0);
951 clock
= crtc
->config
->base
.adjusted_mode
.crtc_clock
;
952 htotal
= crtc
->config
->base
.adjusted_mode
.crtc_htotal
;
953 width
= crtc
->config
->pipe_src_w
;
954 if (WARN_ON(htotal
== 0))
957 if (plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
) {
959 * FIXME the formula gives values that are
960 * too big for the cursor FIFO, and hence we
961 * would never be able to use cursors. For
962 * now just hardcode the watermark.
966 wm
= vlv_wm_method2(clock
, htotal
, width
, pixel_size
,
967 dev_priv
->wm
.pri_latency
[level
] * 10);
970 return min_t(int, wm
, USHRT_MAX
);
973 static void vlv_compute_fifo(struct intel_crtc
*crtc
)
975 struct drm_device
*dev
= crtc
->base
.dev
;
976 struct vlv_wm_state
*wm_state
= &crtc
->wm_state
;
977 struct intel_plane
*plane
;
978 unsigned int total_rate
= 0;
979 const int fifo_size
= 512 - 1;
980 int fifo_extra
, fifo_left
= fifo_size
;
982 for_each_intel_plane_on_crtc(dev
, crtc
, plane
) {
983 struct intel_plane_state
*state
=
984 to_intel_plane_state(plane
->base
.state
);
986 if (plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
)
989 if (state
->visible
) {
990 wm_state
->num_active_planes
++;
991 total_rate
+= drm_format_plane_cpp(state
->base
.fb
->pixel_format
, 0);
995 for_each_intel_plane_on_crtc(dev
, crtc
, plane
) {
996 struct intel_plane_state
*state
=
997 to_intel_plane_state(plane
->base
.state
);
1000 if (plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
) {
1001 plane
->wm
.fifo_size
= 63;
1005 if (!state
->visible
) {
1006 plane
->wm
.fifo_size
= 0;
1010 rate
= drm_format_plane_cpp(state
->base
.fb
->pixel_format
, 0);
1011 plane
->wm
.fifo_size
= fifo_size
* rate
/ total_rate
;
1012 fifo_left
-= plane
->wm
.fifo_size
;
1015 fifo_extra
= DIV_ROUND_UP(fifo_left
, wm_state
->num_active_planes
?: 1);
1017 /* spread the remainder evenly */
1018 for_each_intel_plane_on_crtc(dev
, crtc
, plane
) {
1024 if (plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
)
1027 /* give it all to the first plane if none are active */
1028 if (plane
->wm
.fifo_size
== 0 &&
1029 wm_state
->num_active_planes
)
1032 plane_extra
= min(fifo_extra
, fifo_left
);
1033 plane
->wm
.fifo_size
+= plane_extra
;
1034 fifo_left
-= plane_extra
;
1037 WARN_ON(fifo_left
!= 0);
1040 static void vlv_invert_wms(struct intel_crtc
*crtc
)
1042 struct vlv_wm_state
*wm_state
= &crtc
->wm_state
;
1045 for (level
= 0; level
< wm_state
->num_levels
; level
++) {
1046 struct drm_device
*dev
= crtc
->base
.dev
;
1047 const int sr_fifo_size
= INTEL_INFO(dev
)->num_pipes
* 512 - 1;
1048 struct intel_plane
*plane
;
1050 wm_state
->sr
[level
].plane
= sr_fifo_size
- wm_state
->sr
[level
].plane
;
1051 wm_state
->sr
[level
].cursor
= 63 - wm_state
->sr
[level
].cursor
;
1053 for_each_intel_plane_on_crtc(dev
, crtc
, plane
) {
1054 switch (plane
->base
.type
) {
1056 case DRM_PLANE_TYPE_CURSOR
:
1057 wm_state
->wm
[level
].cursor
= plane
->wm
.fifo_size
-
1058 wm_state
->wm
[level
].cursor
;
1060 case DRM_PLANE_TYPE_PRIMARY
:
1061 wm_state
->wm
[level
].primary
= plane
->wm
.fifo_size
-
1062 wm_state
->wm
[level
].primary
;
1064 case DRM_PLANE_TYPE_OVERLAY
:
1065 sprite
= plane
->plane
;
1066 wm_state
->wm
[level
].sprite
[sprite
] = plane
->wm
.fifo_size
-
1067 wm_state
->wm
[level
].sprite
[sprite
];
1074 static void vlv_compute_wm(struct intel_crtc
*crtc
)
1076 struct drm_device
*dev
= crtc
->base
.dev
;
1077 struct vlv_wm_state
*wm_state
= &crtc
->wm_state
;
1078 struct intel_plane
*plane
;
1079 int sr_fifo_size
= INTEL_INFO(dev
)->num_pipes
* 512 - 1;
1082 memset(wm_state
, 0, sizeof(*wm_state
));
1084 wm_state
->cxsr
= crtc
->pipe
!= PIPE_C
&& crtc
->wm
.cxsr_allowed
;
1085 wm_state
->num_levels
= to_i915(dev
)->wm
.max_level
+ 1;
1087 wm_state
->num_active_planes
= 0;
1089 vlv_compute_fifo(crtc
);
1091 if (wm_state
->num_active_planes
!= 1)
1092 wm_state
->cxsr
= false;
1094 if (wm_state
->cxsr
) {
1095 for (level
= 0; level
< wm_state
->num_levels
; level
++) {
1096 wm_state
->sr
[level
].plane
= sr_fifo_size
;
1097 wm_state
->sr
[level
].cursor
= 63;
1101 for_each_intel_plane_on_crtc(dev
, crtc
, plane
) {
1102 struct intel_plane_state
*state
=
1103 to_intel_plane_state(plane
->base
.state
);
1105 if (!state
->visible
)
1108 /* normal watermarks */
1109 for (level
= 0; level
< wm_state
->num_levels
; level
++) {
1110 int wm
= vlv_compute_wm_level(plane
, crtc
, state
, level
);
1111 int max_wm
= plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
? 63 : 511;
1114 if (WARN_ON(level
== 0 && wm
> max_wm
))
1117 if (wm
> plane
->wm
.fifo_size
)
1120 switch (plane
->base
.type
) {
1122 case DRM_PLANE_TYPE_CURSOR
:
1123 wm_state
->wm
[level
].cursor
= wm
;
1125 case DRM_PLANE_TYPE_PRIMARY
:
1126 wm_state
->wm
[level
].primary
= wm
;
1128 case DRM_PLANE_TYPE_OVERLAY
:
1129 sprite
= plane
->plane
;
1130 wm_state
->wm
[level
].sprite
[sprite
] = wm
;
1135 wm_state
->num_levels
= level
;
1137 if (!wm_state
->cxsr
)
1140 /* maxfifo watermarks */
1141 switch (plane
->base
.type
) {
1143 case DRM_PLANE_TYPE_CURSOR
:
1144 for (level
= 0; level
< wm_state
->num_levels
; level
++)
1145 wm_state
->sr
[level
].cursor
=
1146 wm_state
->wm
[level
].cursor
;
1148 case DRM_PLANE_TYPE_PRIMARY
:
1149 for (level
= 0; level
< wm_state
->num_levels
; level
++)
1150 wm_state
->sr
[level
].plane
=
1151 min(wm_state
->sr
[level
].plane
,
1152 wm_state
->wm
[level
].primary
);
1154 case DRM_PLANE_TYPE_OVERLAY
:
1155 sprite
= plane
->plane
;
1156 for (level
= 0; level
< wm_state
->num_levels
; level
++)
1157 wm_state
->sr
[level
].plane
=
1158 min(wm_state
->sr
[level
].plane
,
1159 wm_state
->wm
[level
].sprite
[sprite
]);
1164 /* clear any (partially) filled invalid levels */
1165 for (level
= wm_state
->num_levels
; level
< to_i915(dev
)->wm
.max_level
+ 1; level
++) {
1166 memset(&wm_state
->wm
[level
], 0, sizeof(wm_state
->wm
[level
]));
1167 memset(&wm_state
->sr
[level
], 0, sizeof(wm_state
->sr
[level
]));
1170 vlv_invert_wms(crtc
);
1173 #define VLV_FIFO(plane, value) \
1174 (((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1176 static void vlv_pipe_set_fifo_size(struct intel_crtc
*crtc
)
1178 struct drm_device
*dev
= crtc
->base
.dev
;
1179 struct drm_i915_private
*dev_priv
= to_i915(dev
);
1180 struct intel_plane
*plane
;
1181 int sprite0_start
= 0, sprite1_start
= 0, fifo_size
= 0;
1183 for_each_intel_plane_on_crtc(dev
, crtc
, plane
) {
1184 if (plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
) {
1185 WARN_ON(plane
->wm
.fifo_size
!= 63);
1189 if (plane
->base
.type
== DRM_PLANE_TYPE_PRIMARY
)
1190 sprite0_start
= plane
->wm
.fifo_size
;
1191 else if (plane
->plane
== 0)
1192 sprite1_start
= sprite0_start
+ plane
->wm
.fifo_size
;
1194 fifo_size
= sprite1_start
+ plane
->wm
.fifo_size
;
1197 WARN_ON(fifo_size
!= 512 - 1);
1199 DRM_DEBUG_KMS("Pipe %c FIFO split %d / %d / %d\n",
1200 pipe_name(crtc
->pipe
), sprite0_start
,
1201 sprite1_start
, fifo_size
);
1203 switch (crtc
->pipe
) {
1204 uint32_t dsparb
, dsparb2
, dsparb3
;
1206 dsparb
= I915_READ(DSPARB
);
1207 dsparb2
= I915_READ(DSPARB2
);
1209 dsparb
&= ~(VLV_FIFO(SPRITEA
, 0xff) |
1210 VLV_FIFO(SPRITEB
, 0xff));
1211 dsparb
|= (VLV_FIFO(SPRITEA
, sprite0_start
) |
1212 VLV_FIFO(SPRITEB
, sprite1_start
));
1214 dsparb2
&= ~(VLV_FIFO(SPRITEA_HI
, 0x1) |
1215 VLV_FIFO(SPRITEB_HI
, 0x1));
1216 dsparb2
|= (VLV_FIFO(SPRITEA_HI
, sprite0_start
>> 8) |
1217 VLV_FIFO(SPRITEB_HI
, sprite1_start
>> 8));
1219 I915_WRITE(DSPARB
, dsparb
);
1220 I915_WRITE(DSPARB2
, dsparb2
);
1223 dsparb
= I915_READ(DSPARB
);
1224 dsparb2
= I915_READ(DSPARB2
);
1226 dsparb
&= ~(VLV_FIFO(SPRITEC
, 0xff) |
1227 VLV_FIFO(SPRITED
, 0xff));
1228 dsparb
|= (VLV_FIFO(SPRITEC
, sprite0_start
) |
1229 VLV_FIFO(SPRITED
, sprite1_start
));
1231 dsparb2
&= ~(VLV_FIFO(SPRITEC_HI
, 0xff) |
1232 VLV_FIFO(SPRITED_HI
, 0xff));
1233 dsparb2
|= (VLV_FIFO(SPRITEC_HI
, sprite0_start
>> 8) |
1234 VLV_FIFO(SPRITED_HI
, sprite1_start
>> 8));
1236 I915_WRITE(DSPARB
, dsparb
);
1237 I915_WRITE(DSPARB2
, dsparb2
);
1240 dsparb3
= I915_READ(DSPARB3
);
1241 dsparb2
= I915_READ(DSPARB2
);
1243 dsparb3
&= ~(VLV_FIFO(SPRITEE
, 0xff) |
1244 VLV_FIFO(SPRITEF
, 0xff));
1245 dsparb3
|= (VLV_FIFO(SPRITEE
, sprite0_start
) |
1246 VLV_FIFO(SPRITEF
, sprite1_start
));
1248 dsparb2
&= ~(VLV_FIFO(SPRITEE_HI
, 0xff) |
1249 VLV_FIFO(SPRITEF_HI
, 0xff));
1250 dsparb2
|= (VLV_FIFO(SPRITEE_HI
, sprite0_start
>> 8) |
1251 VLV_FIFO(SPRITEF_HI
, sprite1_start
>> 8));
1253 I915_WRITE(DSPARB3
, dsparb3
);
1254 I915_WRITE(DSPARB2
, dsparb2
);
1263 static void vlv_merge_wm(struct drm_device
*dev
,
1264 struct vlv_wm_values
*wm
)
1266 struct intel_crtc
*crtc
;
1267 int num_active_crtcs
= 0;
1269 wm
->level
= to_i915(dev
)->wm
.max_level
;
1272 for_each_intel_crtc(dev
, crtc
) {
1273 const struct vlv_wm_state
*wm_state
= &crtc
->wm_state
;
1278 if (!wm_state
->cxsr
)
1282 wm
->level
= min_t(int, wm
->level
, wm_state
->num_levels
- 1);
1285 if (num_active_crtcs
!= 1)
1288 if (num_active_crtcs
> 1)
1289 wm
->level
= VLV_WM_LEVEL_PM2
;
1291 for_each_intel_crtc(dev
, crtc
) {
1292 struct vlv_wm_state
*wm_state
= &crtc
->wm_state
;
1293 enum pipe pipe
= crtc
->pipe
;
1298 wm
->pipe
[pipe
] = wm_state
->wm
[wm
->level
];
1300 wm
->sr
= wm_state
->sr
[wm
->level
];
1302 wm
->ddl
[pipe
].primary
= DDL_PRECISION_HIGH
| 2;
1303 wm
->ddl
[pipe
].sprite
[0] = DDL_PRECISION_HIGH
| 2;
1304 wm
->ddl
[pipe
].sprite
[1] = DDL_PRECISION_HIGH
| 2;
1305 wm
->ddl
[pipe
].cursor
= DDL_PRECISION_HIGH
| 2;
1309 static void vlv_update_wm(struct drm_crtc
*crtc
)
1311 struct drm_device
*dev
= crtc
->dev
;
1312 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1313 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
1314 enum pipe pipe
= intel_crtc
->pipe
;
1315 struct vlv_wm_values wm
= {};
1317 vlv_compute_wm(intel_crtc
);
1318 vlv_merge_wm(dev
, &wm
);
1320 if (memcmp(&dev_priv
->wm
.vlv
, &wm
, sizeof(wm
)) == 0) {
1321 /* FIXME should be part of crtc atomic commit */
1322 vlv_pipe_set_fifo_size(intel_crtc
);
1326 if (wm
.level
< VLV_WM_LEVEL_DDR_DVFS
&&
1327 dev_priv
->wm
.vlv
.level
>= VLV_WM_LEVEL_DDR_DVFS
)
1328 chv_set_memory_dvfs(dev_priv
, false);
1330 if (wm
.level
< VLV_WM_LEVEL_PM5
&&
1331 dev_priv
->wm
.vlv
.level
>= VLV_WM_LEVEL_PM5
)
1332 chv_set_memory_pm5(dev_priv
, false);
1334 if (!wm
.cxsr
&& dev_priv
->wm
.vlv
.cxsr
)
1335 intel_set_memory_cxsr(dev_priv
, false);
1337 /* FIXME should be part of crtc atomic commit */
1338 vlv_pipe_set_fifo_size(intel_crtc
);
1340 vlv_write_wm_values(intel_crtc
, &wm
);
1342 DRM_DEBUG_KMS("Setting FIFO watermarks - %c: plane=%d, cursor=%d, "
1343 "sprite0=%d, sprite1=%d, SR: plane=%d, cursor=%d level=%d cxsr=%d\n",
1344 pipe_name(pipe
), wm
.pipe
[pipe
].primary
, wm
.pipe
[pipe
].cursor
,
1345 wm
.pipe
[pipe
].sprite
[0], wm
.pipe
[pipe
].sprite
[1],
1346 wm
.sr
.plane
, wm
.sr
.cursor
, wm
.level
, wm
.cxsr
);
1348 if (wm
.cxsr
&& !dev_priv
->wm
.vlv
.cxsr
)
1349 intel_set_memory_cxsr(dev_priv
, true);
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
, true);
1355 if (wm
.level
>= VLV_WM_LEVEL_DDR_DVFS
&&
1356 dev_priv
->wm
.vlv
.level
< VLV_WM_LEVEL_DDR_DVFS
)
1357 chv_set_memory_dvfs(dev_priv
, true);
1359 dev_priv
->wm
.vlv
= wm
;
1362 #define single_plane_enabled(mask) is_power_of_2(mask)
1364 static void g4x_update_wm(struct drm_crtc
*crtc
)
1366 struct drm_device
*dev
= crtc
->dev
;
1367 static const int sr_latency_ns
= 12000;
1368 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1369 int planea_wm
, planeb_wm
, cursora_wm
, cursorb_wm
;
1370 int plane_sr
, cursor_sr
;
1371 unsigned int enabled
= 0;
1374 if (g4x_compute_wm0(dev
, PIPE_A
,
1375 &g4x_wm_info
, pessimal_latency_ns
,
1376 &g4x_cursor_wm_info
, pessimal_latency_ns
,
1377 &planea_wm
, &cursora_wm
))
1378 enabled
|= 1 << PIPE_A
;
1380 if (g4x_compute_wm0(dev
, PIPE_B
,
1381 &g4x_wm_info
, pessimal_latency_ns
,
1382 &g4x_cursor_wm_info
, pessimal_latency_ns
,
1383 &planeb_wm
, &cursorb_wm
))
1384 enabled
|= 1 << PIPE_B
;
1386 if (single_plane_enabled(enabled
) &&
1387 g4x_compute_srwm(dev
, ffs(enabled
) - 1,
1390 &g4x_cursor_wm_info
,
1391 &plane_sr
, &cursor_sr
)) {
1392 cxsr_enabled
= true;
1394 cxsr_enabled
= false;
1395 intel_set_memory_cxsr(dev_priv
, false);
1396 plane_sr
= cursor_sr
= 0;
1399 DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, "
1400 "B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1401 planea_wm
, cursora_wm
,
1402 planeb_wm
, cursorb_wm
,
1403 plane_sr
, cursor_sr
);
1406 FW_WM(plane_sr
, SR
) |
1407 FW_WM(cursorb_wm
, CURSORB
) |
1408 FW_WM(planeb_wm
, PLANEB
) |
1409 FW_WM(planea_wm
, PLANEA
));
1411 (I915_READ(DSPFW2
) & ~DSPFW_CURSORA_MASK
) |
1412 FW_WM(cursora_wm
, CURSORA
));
1413 /* HPLL off in SR has some issues on G4x... disable it */
1415 (I915_READ(DSPFW3
) & ~(DSPFW_HPLL_SR_EN
| DSPFW_CURSOR_SR_MASK
)) |
1416 FW_WM(cursor_sr
, CURSOR_SR
));
1419 intel_set_memory_cxsr(dev_priv
, true);
1422 static void i965_update_wm(struct drm_crtc
*unused_crtc
)
1424 struct drm_device
*dev
= unused_crtc
->dev
;
1425 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1426 struct drm_crtc
*crtc
;
1431 /* Calc sr entries for one plane configs */
1432 crtc
= single_enabled_crtc(dev
);
1434 /* self-refresh has much higher latency */
1435 static const int sr_latency_ns
= 12000;
1436 const struct drm_display_mode
*adjusted_mode
= &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
1437 int clock
= adjusted_mode
->crtc_clock
;
1438 int htotal
= adjusted_mode
->crtc_htotal
;
1439 int hdisplay
= to_intel_crtc(crtc
)->config
->pipe_src_w
;
1440 int pixel_size
= crtc
->primary
->state
->fb
->bits_per_pixel
/ 8;
1441 unsigned long line_time_us
;
1444 line_time_us
= max(htotal
* 1000 / clock
, 1);
1446 /* Use ns/us then divide to preserve precision */
1447 entries
= (((sr_latency_ns
/ line_time_us
) + 1000) / 1000) *
1448 pixel_size
* hdisplay
;
1449 entries
= DIV_ROUND_UP(entries
, I915_FIFO_LINE_SIZE
);
1450 srwm
= I965_FIFO_SIZE
- entries
;
1454 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1457 entries
= (((sr_latency_ns
/ line_time_us
) + 1000) / 1000) *
1458 pixel_size
* crtc
->cursor
->state
->crtc_w
;
1459 entries
= DIV_ROUND_UP(entries
,
1460 i965_cursor_wm_info
.cacheline_size
);
1461 cursor_sr
= i965_cursor_wm_info
.fifo_size
-
1462 (entries
+ i965_cursor_wm_info
.guard_size
);
1464 if (cursor_sr
> i965_cursor_wm_info
.max_wm
)
1465 cursor_sr
= i965_cursor_wm_info
.max_wm
;
1467 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1468 "cursor %d\n", srwm
, cursor_sr
);
1470 cxsr_enabled
= true;
1472 cxsr_enabled
= false;
1473 /* Turn off self refresh if both pipes are enabled */
1474 intel_set_memory_cxsr(dev_priv
, false);
1477 DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1480 /* 965 has limitations... */
1481 I915_WRITE(DSPFW1
, FW_WM(srwm
, SR
) |
1485 I915_WRITE(DSPFW2
, FW_WM(8, CURSORA
) |
1486 FW_WM(8, PLANEC_OLD
));
1487 /* update cursor SR watermark */
1488 I915_WRITE(DSPFW3
, FW_WM(cursor_sr
, CURSOR_SR
));
1491 intel_set_memory_cxsr(dev_priv
, true);
1496 static void i9xx_update_wm(struct drm_crtc
*unused_crtc
)
1498 struct drm_device
*dev
= unused_crtc
->dev
;
1499 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1500 const struct intel_watermark_params
*wm_info
;
1505 int planea_wm
, planeb_wm
;
1506 struct drm_crtc
*crtc
, *enabled
= NULL
;
1509 wm_info
= &i945_wm_info
;
1510 else if (!IS_GEN2(dev
))
1511 wm_info
= &i915_wm_info
;
1513 wm_info
= &i830_a_wm_info
;
1515 fifo_size
= dev_priv
->display
.get_fifo_size(dev
, 0);
1516 crtc
= intel_get_crtc_for_plane(dev
, 0);
1517 if (intel_crtc_active(crtc
)) {
1518 const struct drm_display_mode
*adjusted_mode
;
1519 int cpp
= crtc
->primary
->state
->fb
->bits_per_pixel
/ 8;
1523 adjusted_mode
= &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
1524 planea_wm
= intel_calculate_wm(adjusted_mode
->crtc_clock
,
1525 wm_info
, fifo_size
, cpp
,
1526 pessimal_latency_ns
);
1529 planea_wm
= fifo_size
- wm_info
->guard_size
;
1530 if (planea_wm
> (long)wm_info
->max_wm
)
1531 planea_wm
= wm_info
->max_wm
;
1535 wm_info
= &i830_bc_wm_info
;
1537 fifo_size
= dev_priv
->display
.get_fifo_size(dev
, 1);
1538 crtc
= intel_get_crtc_for_plane(dev
, 1);
1539 if (intel_crtc_active(crtc
)) {
1540 const struct drm_display_mode
*adjusted_mode
;
1541 int cpp
= crtc
->primary
->state
->fb
->bits_per_pixel
/ 8;
1545 adjusted_mode
= &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
1546 planeb_wm
= intel_calculate_wm(adjusted_mode
->crtc_clock
,
1547 wm_info
, fifo_size
, cpp
,
1548 pessimal_latency_ns
);
1549 if (enabled
== NULL
)
1554 planeb_wm
= fifo_size
- wm_info
->guard_size
;
1555 if (planeb_wm
> (long)wm_info
->max_wm
)
1556 planeb_wm
= wm_info
->max_wm
;
1559 DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm
, planeb_wm
);
1561 if (IS_I915GM(dev
) && enabled
) {
1562 struct drm_i915_gem_object
*obj
;
1564 obj
= intel_fb_obj(enabled
->primary
->state
->fb
);
1566 /* self-refresh seems busted with untiled */
1567 if (obj
->tiling_mode
== I915_TILING_NONE
)
1572 * Overlay gets an aggressive default since video jitter is bad.
1576 /* Play safe and disable self-refresh before adjusting watermarks. */
1577 intel_set_memory_cxsr(dev_priv
, false);
1579 /* Calc sr entries for one plane configs */
1580 if (HAS_FW_BLC(dev
) && enabled
) {
1581 /* self-refresh has much higher latency */
1582 static const int sr_latency_ns
= 6000;
1583 const struct drm_display_mode
*adjusted_mode
= &to_intel_crtc(enabled
)->config
->base
.adjusted_mode
;
1584 int clock
= adjusted_mode
->crtc_clock
;
1585 int htotal
= adjusted_mode
->crtc_htotal
;
1586 int hdisplay
= to_intel_crtc(enabled
)->config
->pipe_src_w
;
1587 int pixel_size
= enabled
->primary
->state
->fb
->bits_per_pixel
/ 8;
1588 unsigned long line_time_us
;
1591 line_time_us
= max(htotal
* 1000 / clock
, 1);
1593 /* Use ns/us then divide to preserve precision */
1594 entries
= (((sr_latency_ns
/ line_time_us
) + 1000) / 1000) *
1595 pixel_size
* hdisplay
;
1596 entries
= DIV_ROUND_UP(entries
, wm_info
->cacheline_size
);
1597 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries
);
1598 srwm
= wm_info
->fifo_size
- entries
;
1602 if (IS_I945G(dev
) || IS_I945GM(dev
))
1603 I915_WRITE(FW_BLC_SELF
,
1604 FW_BLC_SELF_FIFO_MASK
| (srwm
& 0xff));
1605 else if (IS_I915GM(dev
))
1606 I915_WRITE(FW_BLC_SELF
, srwm
& 0x3f);
1609 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1610 planea_wm
, planeb_wm
, cwm
, srwm
);
1612 fwater_lo
= ((planeb_wm
& 0x3f) << 16) | (planea_wm
& 0x3f);
1613 fwater_hi
= (cwm
& 0x1f);
1615 /* Set request length to 8 cachelines per fetch */
1616 fwater_lo
= fwater_lo
| (1 << 24) | (1 << 8);
1617 fwater_hi
= fwater_hi
| (1 << 8);
1619 I915_WRITE(FW_BLC
, fwater_lo
);
1620 I915_WRITE(FW_BLC2
, fwater_hi
);
1623 intel_set_memory_cxsr(dev_priv
, true);
1626 static void i845_update_wm(struct drm_crtc
*unused_crtc
)
1628 struct drm_device
*dev
= unused_crtc
->dev
;
1629 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1630 struct drm_crtc
*crtc
;
1631 const struct drm_display_mode
*adjusted_mode
;
1635 crtc
= single_enabled_crtc(dev
);
1639 adjusted_mode
= &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
1640 planea_wm
= intel_calculate_wm(adjusted_mode
->crtc_clock
,
1642 dev_priv
->display
.get_fifo_size(dev
, 0),
1643 4, pessimal_latency_ns
);
1644 fwater_lo
= I915_READ(FW_BLC
) & ~0xfff;
1645 fwater_lo
|= (3<<8) | planea_wm
;
1647 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm
);
1649 I915_WRITE(FW_BLC
, fwater_lo
);
1652 uint32_t ilk_pipe_pixel_rate(const struct intel_crtc_state
*pipe_config
)
1654 uint32_t pixel_rate
;
1656 pixel_rate
= pipe_config
->base
.adjusted_mode
.crtc_clock
;
1658 /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
1659 * adjust the pixel_rate here. */
1661 if (pipe_config
->pch_pfit
.enabled
) {
1662 uint64_t pipe_w
, pipe_h
, pfit_w
, pfit_h
;
1663 uint32_t pfit_size
= pipe_config
->pch_pfit
.size
;
1665 pipe_w
= pipe_config
->pipe_src_w
;
1666 pipe_h
= pipe_config
->pipe_src_h
;
1668 pfit_w
= (pfit_size
>> 16) & 0xFFFF;
1669 pfit_h
= pfit_size
& 0xFFFF;
1670 if (pipe_w
< pfit_w
)
1672 if (pipe_h
< pfit_h
)
1675 if (WARN_ON(!pfit_w
|| !pfit_h
))
1678 pixel_rate
= div_u64((uint64_t) pixel_rate
* pipe_w
* pipe_h
,
1685 /* latency must be in 0.1us units. */
1686 static uint32_t ilk_wm_method1(uint32_t pixel_rate
, uint8_t bytes_per_pixel
,
1691 if (WARN(latency
== 0, "Latency value missing\n"))
1694 ret
= (uint64_t) pixel_rate
* bytes_per_pixel
* latency
;
1695 ret
= DIV_ROUND_UP_ULL(ret
, 64 * 10000) + 2;
1700 /* latency must be in 0.1us units. */
1701 static uint32_t ilk_wm_method2(uint32_t pixel_rate
, uint32_t pipe_htotal
,
1702 uint32_t horiz_pixels
, uint8_t bytes_per_pixel
,
1707 if (WARN(latency
== 0, "Latency value missing\n"))
1709 if (WARN_ON(!pipe_htotal
))
1712 ret
= (latency
* pixel_rate
) / (pipe_htotal
* 10000);
1713 ret
= (ret
+ 1) * horiz_pixels
* bytes_per_pixel
;
1714 ret
= DIV_ROUND_UP(ret
, 64) + 2;
1718 static uint32_t ilk_wm_fbc(uint32_t pri_val
, uint32_t horiz_pixels
,
1719 uint8_t bytes_per_pixel
)
1722 * Neither of these should be possible since this function shouldn't be
1723 * called if the CRTC is off or the plane is invisible. But let's be
1724 * extra paranoid to avoid a potential divide-by-zero if we screw up
1725 * elsewhere in the driver.
1727 if (WARN_ON(!bytes_per_pixel
))
1729 if (WARN_ON(!horiz_pixels
))
1732 return DIV_ROUND_UP(pri_val
* 64, horiz_pixels
* bytes_per_pixel
) + 2;
1735 struct ilk_wm_maximums
{
1743 * For both WM_PIPE and WM_LP.
1744 * mem_value must be in 0.1us units.
1746 static uint32_t ilk_compute_pri_wm(const struct intel_crtc_state
*cstate
,
1747 const struct intel_plane_state
*pstate
,
1751 int bpp
= pstate
->base
.fb
? pstate
->base
.fb
->bits_per_pixel
/ 8 : 0;
1752 uint32_t method1
, method2
;
1754 if (!cstate
->base
.active
|| !pstate
->visible
)
1757 method1
= ilk_wm_method1(ilk_pipe_pixel_rate(cstate
), bpp
, mem_value
);
1762 method2
= ilk_wm_method2(ilk_pipe_pixel_rate(cstate
),
1763 cstate
->base
.adjusted_mode
.crtc_htotal
,
1764 drm_rect_width(&pstate
->dst
),
1768 return min(method1
, method2
);
1772 * For both WM_PIPE and WM_LP.
1773 * mem_value must be in 0.1us units.
1775 static uint32_t ilk_compute_spr_wm(const struct intel_crtc_state
*cstate
,
1776 const struct intel_plane_state
*pstate
,
1779 int bpp
= pstate
->base
.fb
? pstate
->base
.fb
->bits_per_pixel
/ 8 : 0;
1780 uint32_t method1
, method2
;
1782 if (!cstate
->base
.active
|| !pstate
->visible
)
1785 method1
= ilk_wm_method1(ilk_pipe_pixel_rate(cstate
), bpp
, mem_value
);
1786 method2
= ilk_wm_method2(ilk_pipe_pixel_rate(cstate
),
1787 cstate
->base
.adjusted_mode
.crtc_htotal
,
1788 drm_rect_width(&pstate
->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_cur_wm(const struct intel_crtc_state
*cstate
,
1799 const struct intel_plane_state
*pstate
,
1802 int bpp
= pstate
->base
.fb
? pstate
->base
.fb
->bits_per_pixel
/ 8 : 0;
1804 if (!cstate
->base
.active
|| !pstate
->visible
)
1807 return ilk_wm_method2(ilk_pipe_pixel_rate(cstate
),
1808 cstate
->base
.adjusted_mode
.crtc_htotal
,
1809 drm_rect_width(&pstate
->dst
),
1814 /* Only for WM_LP. */
1815 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state
*cstate
,
1816 const struct intel_plane_state
*pstate
,
1819 int bpp
= pstate
->base
.fb
? pstate
->base
.fb
->bits_per_pixel
/ 8 : 0;
1821 if (!cstate
->base
.active
|| !pstate
->visible
)
1824 return ilk_wm_fbc(pri_val
, drm_rect_width(&pstate
->dst
), bpp
);
1827 static unsigned int ilk_display_fifo_size(const struct drm_device
*dev
)
1829 if (INTEL_INFO(dev
)->gen
>= 8)
1831 else if (INTEL_INFO(dev
)->gen
>= 7)
1837 static unsigned int ilk_plane_wm_reg_max(const struct drm_device
*dev
,
1838 int level
, bool is_sprite
)
1840 if (INTEL_INFO(dev
)->gen
>= 8)
1841 /* BDW primary/sprite plane watermarks */
1842 return level
== 0 ? 255 : 2047;
1843 else if (INTEL_INFO(dev
)->gen
>= 7)
1844 /* IVB/HSW primary/sprite plane watermarks */
1845 return level
== 0 ? 127 : 1023;
1846 else if (!is_sprite
)
1847 /* ILK/SNB primary plane watermarks */
1848 return level
== 0 ? 127 : 511;
1850 /* ILK/SNB sprite plane watermarks */
1851 return level
== 0 ? 63 : 255;
1854 static unsigned int ilk_cursor_wm_reg_max(const struct drm_device
*dev
,
1857 if (INTEL_INFO(dev
)->gen
>= 7)
1858 return level
== 0 ? 63 : 255;
1860 return level
== 0 ? 31 : 63;
1863 static unsigned int ilk_fbc_wm_reg_max(const struct drm_device
*dev
)
1865 if (INTEL_INFO(dev
)->gen
>= 8)
1871 /* Calculate the maximum primary/sprite plane watermark */
1872 static unsigned int ilk_plane_wm_max(const struct drm_device
*dev
,
1874 const struct intel_wm_config
*config
,
1875 enum intel_ddb_partitioning ddb_partitioning
,
1878 unsigned int fifo_size
= ilk_display_fifo_size(dev
);
1880 /* if sprites aren't enabled, sprites get nothing */
1881 if (is_sprite
&& !config
->sprites_enabled
)
1884 /* HSW allows LP1+ watermarks even with multiple pipes */
1885 if (level
== 0 || config
->num_pipes_active
> 1) {
1886 fifo_size
/= INTEL_INFO(dev
)->num_pipes
;
1889 * For some reason the non self refresh
1890 * FIFO size is only half of the self
1891 * refresh FIFO size on ILK/SNB.
1893 if (INTEL_INFO(dev
)->gen
<= 6)
1897 if (config
->sprites_enabled
) {
1898 /* level 0 is always calculated with 1:1 split */
1899 if (level
> 0 && ddb_partitioning
== INTEL_DDB_PART_5_6
) {
1908 /* clamp to max that the registers can hold */
1909 return min(fifo_size
, ilk_plane_wm_reg_max(dev
, level
, is_sprite
));
1912 /* Calculate the maximum cursor plane watermark */
1913 static unsigned int ilk_cursor_wm_max(const struct drm_device
*dev
,
1915 const struct intel_wm_config
*config
)
1917 /* HSW LP1+ watermarks w/ multiple pipes */
1918 if (level
> 0 && config
->num_pipes_active
> 1)
1921 /* otherwise just report max that registers can hold */
1922 return ilk_cursor_wm_reg_max(dev
, level
);
1925 static void ilk_compute_wm_maximums(const struct drm_device
*dev
,
1927 const struct intel_wm_config
*config
,
1928 enum intel_ddb_partitioning ddb_partitioning
,
1929 struct ilk_wm_maximums
*max
)
1931 max
->pri
= ilk_plane_wm_max(dev
, level
, config
, ddb_partitioning
, false);
1932 max
->spr
= ilk_plane_wm_max(dev
, level
, config
, ddb_partitioning
, true);
1933 max
->cur
= ilk_cursor_wm_max(dev
, level
, config
);
1934 max
->fbc
= ilk_fbc_wm_reg_max(dev
);
1937 static void ilk_compute_wm_reg_maximums(struct drm_device
*dev
,
1939 struct ilk_wm_maximums
*max
)
1941 max
->pri
= ilk_plane_wm_reg_max(dev
, level
, false);
1942 max
->spr
= ilk_plane_wm_reg_max(dev
, level
, true);
1943 max
->cur
= ilk_cursor_wm_reg_max(dev
, level
);
1944 max
->fbc
= ilk_fbc_wm_reg_max(dev
);
1947 static bool ilk_validate_wm_level(int level
,
1948 const struct ilk_wm_maximums
*max
,
1949 struct intel_wm_level
*result
)
1953 /* already determined to be invalid? */
1954 if (!result
->enable
)
1957 result
->enable
= result
->pri_val
<= max
->pri
&&
1958 result
->spr_val
<= max
->spr
&&
1959 result
->cur_val
<= max
->cur
;
1961 ret
= result
->enable
;
1964 * HACK until we can pre-compute everything,
1965 * and thus fail gracefully if LP0 watermarks
1968 if (level
== 0 && !result
->enable
) {
1969 if (result
->pri_val
> max
->pri
)
1970 DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
1971 level
, result
->pri_val
, max
->pri
);
1972 if (result
->spr_val
> max
->spr
)
1973 DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
1974 level
, result
->spr_val
, max
->spr
);
1975 if (result
->cur_val
> max
->cur
)
1976 DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
1977 level
, result
->cur_val
, max
->cur
);
1979 result
->pri_val
= min_t(uint32_t, result
->pri_val
, max
->pri
);
1980 result
->spr_val
= min_t(uint32_t, result
->spr_val
, max
->spr
);
1981 result
->cur_val
= min_t(uint32_t, result
->cur_val
, max
->cur
);
1982 result
->enable
= true;
1988 static void ilk_compute_wm_level(const struct drm_i915_private
*dev_priv
,
1989 const struct intel_crtc
*intel_crtc
,
1991 struct intel_crtc_state
*cstate
,
1992 struct intel_plane_state
*pristate
,
1993 struct intel_plane_state
*sprstate
,
1994 struct intel_plane_state
*curstate
,
1995 struct intel_wm_level
*result
)
1997 uint16_t pri_latency
= dev_priv
->wm
.pri_latency
[level
];
1998 uint16_t spr_latency
= dev_priv
->wm
.spr_latency
[level
];
1999 uint16_t cur_latency
= dev_priv
->wm
.cur_latency
[level
];
2001 /* WM1+ latency values stored in 0.5us units */
2008 result
->pri_val
= ilk_compute_pri_wm(cstate
, pristate
,
2009 pri_latency
, level
);
2010 result
->spr_val
= ilk_compute_spr_wm(cstate
, sprstate
, spr_latency
);
2011 result
->cur_val
= ilk_compute_cur_wm(cstate
, curstate
, cur_latency
);
2012 result
->fbc_val
= ilk_compute_fbc_wm(cstate
, pristate
, result
->pri_val
);
2013 result
->enable
= true;
2017 hsw_compute_linetime_wm(struct drm_device
*dev
,
2018 struct intel_crtc_state
*cstate
)
2020 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2021 const struct drm_display_mode
*adjusted_mode
=
2022 &cstate
->base
.adjusted_mode
;
2023 u32 linetime
, ips_linetime
;
2025 if (!cstate
->base
.active
)
2027 if (WARN_ON(adjusted_mode
->crtc_clock
== 0))
2029 if (WARN_ON(dev_priv
->cdclk_freq
== 0))
2032 /* The WM are computed with base on how long it takes to fill a single
2033 * row at the given clock rate, multiplied by 8.
2035 linetime
= DIV_ROUND_CLOSEST(adjusted_mode
->crtc_htotal
* 1000 * 8,
2036 adjusted_mode
->crtc_clock
);
2037 ips_linetime
= DIV_ROUND_CLOSEST(adjusted_mode
->crtc_htotal
* 1000 * 8,
2038 dev_priv
->cdclk_freq
);
2040 return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime
) |
2041 PIPE_WM_LINETIME_TIME(linetime
);
2044 static void intel_read_wm_latency(struct drm_device
*dev
, uint16_t wm
[8])
2046 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2051 int level
, max_level
= ilk_wm_max_level(dev
);
2053 /* read the first set of memory latencies[0:3] */
2054 val
= 0; /* data0 to be programmed to 0 for first set */
2055 mutex_lock(&dev_priv
->rps
.hw_lock
);
2056 ret
= sandybridge_pcode_read(dev_priv
,
2057 GEN9_PCODE_READ_MEM_LATENCY
,
2059 mutex_unlock(&dev_priv
->rps
.hw_lock
);
2062 DRM_ERROR("SKL Mailbox read error = %d\n", ret
);
2066 wm
[0] = val
& GEN9_MEM_LATENCY_LEVEL_MASK
;
2067 wm
[1] = (val
>> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT
) &
2068 GEN9_MEM_LATENCY_LEVEL_MASK
;
2069 wm
[2] = (val
>> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT
) &
2070 GEN9_MEM_LATENCY_LEVEL_MASK
;
2071 wm
[3] = (val
>> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT
) &
2072 GEN9_MEM_LATENCY_LEVEL_MASK
;
2074 /* read the second set of memory latencies[4:7] */
2075 val
= 1; /* data0 to be programmed to 1 for second set */
2076 mutex_lock(&dev_priv
->rps
.hw_lock
);
2077 ret
= sandybridge_pcode_read(dev_priv
,
2078 GEN9_PCODE_READ_MEM_LATENCY
,
2080 mutex_unlock(&dev_priv
->rps
.hw_lock
);
2082 DRM_ERROR("SKL Mailbox read error = %d\n", ret
);
2086 wm
[4] = val
& GEN9_MEM_LATENCY_LEVEL_MASK
;
2087 wm
[5] = (val
>> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT
) &
2088 GEN9_MEM_LATENCY_LEVEL_MASK
;
2089 wm
[6] = (val
>> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT
) &
2090 GEN9_MEM_LATENCY_LEVEL_MASK
;
2091 wm
[7] = (val
>> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT
) &
2092 GEN9_MEM_LATENCY_LEVEL_MASK
;
2095 * WaWmMemoryReadLatency:skl
2097 * punit doesn't take into account the read latency so we need
2098 * to add 2us to the various latency levels we retrieve from
2100 * - W0 is a bit special in that it's the only level that
2101 * can't be disabled if we want to have display working, so
2102 * we always add 2us there.
2103 * - For levels >=1, punit returns 0us latency when they are
2104 * disabled, so we respect that and don't add 2us then
2106 * Additionally, if a level n (n > 1) has a 0us latency, all
2107 * levels m (m >= n) need to be disabled. We make sure to
2108 * sanitize the values out of the punit to satisfy this
2112 for (level
= 1; level
<= max_level
; level
++)
2116 for (i
= level
+ 1; i
<= max_level
; i
++)
2121 } else if (IS_HASWELL(dev
) || IS_BROADWELL(dev
)) {
2122 uint64_t sskpd
= I915_READ64(MCH_SSKPD
);
2124 wm
[0] = (sskpd
>> 56) & 0xFF;
2126 wm
[0] = sskpd
& 0xF;
2127 wm
[1] = (sskpd
>> 4) & 0xFF;
2128 wm
[2] = (sskpd
>> 12) & 0xFF;
2129 wm
[3] = (sskpd
>> 20) & 0x1FF;
2130 wm
[4] = (sskpd
>> 32) & 0x1FF;
2131 } else if (INTEL_INFO(dev
)->gen
>= 6) {
2132 uint32_t sskpd
= I915_READ(MCH_SSKPD
);
2134 wm
[0] = (sskpd
>> SSKPD_WM0_SHIFT
) & SSKPD_WM_MASK
;
2135 wm
[1] = (sskpd
>> SSKPD_WM1_SHIFT
) & SSKPD_WM_MASK
;
2136 wm
[2] = (sskpd
>> SSKPD_WM2_SHIFT
) & SSKPD_WM_MASK
;
2137 wm
[3] = (sskpd
>> SSKPD_WM3_SHIFT
) & SSKPD_WM_MASK
;
2138 } else if (INTEL_INFO(dev
)->gen
>= 5) {
2139 uint32_t mltr
= I915_READ(MLTR_ILK
);
2141 /* ILK primary LP0 latency is 700 ns */
2143 wm
[1] = (mltr
>> MLTR_WM1_SHIFT
) & ILK_SRLT_MASK
;
2144 wm
[2] = (mltr
>> MLTR_WM2_SHIFT
) & ILK_SRLT_MASK
;
2148 static void intel_fixup_spr_wm_latency(struct drm_device
*dev
, uint16_t wm
[5])
2150 /* ILK sprite LP0 latency is 1300 ns */
2151 if (INTEL_INFO(dev
)->gen
== 5)
2155 static void intel_fixup_cur_wm_latency(struct drm_device
*dev
, uint16_t wm
[5])
2157 /* ILK cursor LP0 latency is 1300 ns */
2158 if (INTEL_INFO(dev
)->gen
== 5)
2161 /* WaDoubleCursorLP3Latency:ivb */
2162 if (IS_IVYBRIDGE(dev
))
2166 int ilk_wm_max_level(const struct drm_device
*dev
)
2168 /* how many WM levels are we expecting */
2169 if (INTEL_INFO(dev
)->gen
>= 9)
2171 else if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
2173 else if (INTEL_INFO(dev
)->gen
>= 6)
2179 static void intel_print_wm_latency(struct drm_device
*dev
,
2181 const uint16_t wm
[8])
2183 int level
, max_level
= ilk_wm_max_level(dev
);
2185 for (level
= 0; level
<= max_level
; level
++) {
2186 unsigned int latency
= wm
[level
];
2189 DRM_ERROR("%s WM%d latency not provided\n",
2195 * - latencies are in us on gen9.
2196 * - before then, WM1+ latency values are in 0.5us units
2203 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2204 name
, level
, wm
[level
],
2205 latency
/ 10, latency
% 10);
2209 static bool ilk_increase_wm_latency(struct drm_i915_private
*dev_priv
,
2210 uint16_t wm
[5], uint16_t min
)
2212 int level
, max_level
= ilk_wm_max_level(dev_priv
->dev
);
2217 wm
[0] = max(wm
[0], min
);
2218 for (level
= 1; level
<= max_level
; level
++)
2219 wm
[level
] = max_t(uint16_t, wm
[level
], DIV_ROUND_UP(min
, 5));
2224 static void snb_wm_latency_quirk(struct drm_device
*dev
)
2226 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2230 * The BIOS provided WM memory latency values are often
2231 * inadequate for high resolution displays. Adjust them.
2233 changed
= ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.pri_latency
, 12) |
2234 ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.spr_latency
, 12) |
2235 ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.cur_latency
, 12);
2240 DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2241 intel_print_wm_latency(dev
, "Primary", dev_priv
->wm
.pri_latency
);
2242 intel_print_wm_latency(dev
, "Sprite", dev_priv
->wm
.spr_latency
);
2243 intel_print_wm_latency(dev
, "Cursor", dev_priv
->wm
.cur_latency
);
2246 static void ilk_setup_wm_latency(struct drm_device
*dev
)
2248 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2250 intel_read_wm_latency(dev
, dev_priv
->wm
.pri_latency
);
2252 memcpy(dev_priv
->wm
.spr_latency
, dev_priv
->wm
.pri_latency
,
2253 sizeof(dev_priv
->wm
.pri_latency
));
2254 memcpy(dev_priv
->wm
.cur_latency
, dev_priv
->wm
.pri_latency
,
2255 sizeof(dev_priv
->wm
.pri_latency
));
2257 intel_fixup_spr_wm_latency(dev
, dev_priv
->wm
.spr_latency
);
2258 intel_fixup_cur_wm_latency(dev
, dev_priv
->wm
.cur_latency
);
2260 intel_print_wm_latency(dev
, "Primary", dev_priv
->wm
.pri_latency
);
2261 intel_print_wm_latency(dev
, "Sprite", dev_priv
->wm
.spr_latency
);
2262 intel_print_wm_latency(dev
, "Cursor", dev_priv
->wm
.cur_latency
);
2265 snb_wm_latency_quirk(dev
);
2268 static void skl_setup_wm_latency(struct drm_device
*dev
)
2270 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2272 intel_read_wm_latency(dev
, dev_priv
->wm
.skl_latency
);
2273 intel_print_wm_latency(dev
, "Gen9 Plane", dev_priv
->wm
.skl_latency
);
2276 static bool ilk_validate_pipe_wm(struct drm_device
*dev
,
2277 struct intel_pipe_wm
*pipe_wm
)
2279 /* LP0 watermark maximums depend on this pipe alone */
2280 const struct intel_wm_config config
= {
2281 .num_pipes_active
= 1,
2282 .sprites_enabled
= pipe_wm
->sprites_enabled
,
2283 .sprites_scaled
= pipe_wm
->sprites_scaled
,
2285 struct ilk_wm_maximums max
;
2287 /* LP0 watermarks always use 1/2 DDB partitioning */
2288 ilk_compute_wm_maximums(dev
, 0, &config
, INTEL_DDB_PART_1_2
, &max
);
2290 /* At least LP0 must be valid */
2291 if (!ilk_validate_wm_level(0, &max
, &pipe_wm
->wm
[0])) {
2292 DRM_DEBUG_KMS("LP0 watermark invalid\n");
2299 /* Compute new watermarks for the pipe */
2300 static int ilk_compute_pipe_wm(struct intel_crtc
*intel_crtc
,
2301 struct drm_atomic_state
*state
)
2303 struct intel_pipe_wm
*pipe_wm
;
2304 struct drm_device
*dev
= intel_crtc
->base
.dev
;
2305 const struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2306 struct intel_crtc_state
*cstate
= NULL
;
2307 struct intel_plane
*intel_plane
;
2308 struct drm_plane_state
*ps
;
2309 struct intel_plane_state
*pristate
= NULL
;
2310 struct intel_plane_state
*sprstate
= NULL
;
2311 struct intel_plane_state
*curstate
= NULL
;
2312 int level
, max_level
= ilk_wm_max_level(dev
);
2313 struct ilk_wm_maximums max
;
2315 cstate
= intel_atomic_get_crtc_state(state
, intel_crtc
);
2317 return PTR_ERR(cstate
);
2319 pipe_wm
= &cstate
->wm
.optimal
.ilk
;
2321 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
2322 ps
= drm_atomic_get_plane_state(state
,
2323 &intel_plane
->base
);
2327 if (intel_plane
->base
.type
== DRM_PLANE_TYPE_PRIMARY
)
2328 pristate
= to_intel_plane_state(ps
);
2329 else if (intel_plane
->base
.type
== DRM_PLANE_TYPE_OVERLAY
)
2330 sprstate
= to_intel_plane_state(ps
);
2331 else if (intel_plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
)
2332 curstate
= to_intel_plane_state(ps
);
2335 pipe_wm
->pipe_enabled
= cstate
->base
.active
;
2336 pipe_wm
->sprites_enabled
= sprstate
->visible
;
2337 pipe_wm
->sprites_scaled
= sprstate
->visible
&&
2338 (drm_rect_width(&sprstate
->dst
) != drm_rect_width(&sprstate
->src
) >> 16 ||
2339 drm_rect_height(&sprstate
->dst
) != drm_rect_height(&sprstate
->src
) >> 16);
2341 /* ILK/SNB: LP2+ watermarks only w/o sprites */
2342 if (INTEL_INFO(dev
)->gen
<= 6 && sprstate
->visible
)
2345 /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2346 if (pipe_wm
->sprites_scaled
)
2349 ilk_compute_wm_level(dev_priv
, intel_crtc
, 0, cstate
,
2350 pristate
, sprstate
, curstate
, &pipe_wm
->wm
[0]);
2352 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
2353 pipe_wm
->linetime
= hsw_compute_linetime_wm(dev
, cstate
);
2355 if (!ilk_validate_pipe_wm(dev
, pipe_wm
))
2358 ilk_compute_wm_reg_maximums(dev
, 1, &max
);
2360 for (level
= 1; level
<= max_level
; level
++) {
2361 struct intel_wm_level wm
= {};
2363 ilk_compute_wm_level(dev_priv
, intel_crtc
, level
, cstate
,
2364 pristate
, sprstate
, curstate
, &wm
);
2367 * Disable any watermark level that exceeds the
2368 * register maximums since such watermarks are
2371 if (!ilk_validate_wm_level(level
, &max
, &wm
))
2374 pipe_wm
->wm
[level
] = wm
;
2381 * Build a set of 'intermediate' watermark values that satisfy both the old
2382 * state and the new state. These can be programmed to the hardware
2385 static int ilk_compute_intermediate_wm(struct drm_device
*dev
,
2386 struct intel_crtc
*intel_crtc
,
2387 struct intel_crtc_state
*newstate
)
2389 struct intel_pipe_wm
*a
= &newstate
->wm
.intermediate
;
2390 struct intel_pipe_wm
*b
= &intel_crtc
->wm
.active
.ilk
;
2391 int level
, max_level
= ilk_wm_max_level(dev
);
2394 * Start with the final, target watermarks, then combine with the
2395 * currently active watermarks to get values that are safe both before
2396 * and after the vblank.
2398 *a
= newstate
->wm
.optimal
.ilk
;
2399 a
->pipe_enabled
|= b
->pipe_enabled
;
2400 a
->sprites_enabled
|= b
->sprites_enabled
;
2401 a
->sprites_scaled
|= b
->sprites_scaled
;
2403 for (level
= 0; level
<= max_level
; level
++) {
2404 struct intel_wm_level
*a_wm
= &a
->wm
[level
];
2405 const struct intel_wm_level
*b_wm
= &b
->wm
[level
];
2407 a_wm
->enable
&= b_wm
->enable
;
2408 a_wm
->pri_val
= max(a_wm
->pri_val
, b_wm
->pri_val
);
2409 a_wm
->spr_val
= max(a_wm
->spr_val
, b_wm
->spr_val
);
2410 a_wm
->cur_val
= max(a_wm
->cur_val
, b_wm
->cur_val
);
2411 a_wm
->fbc_val
= max(a_wm
->fbc_val
, b_wm
->fbc_val
);
2415 * We need to make sure that these merged watermark values are
2416 * actually a valid configuration themselves. If they're not,
2417 * there's no safe way to transition from the old state to
2418 * the new state, so we need to fail the atomic transaction.
2420 if (!ilk_validate_pipe_wm(dev
, a
))
2424 * If our intermediate WM are identical to the final WM, then we can
2425 * omit the post-vblank programming; only update if it's different.
2427 if (memcmp(a
, &newstate
->wm
.optimal
.ilk
, sizeof(*a
)) != 0)
2428 newstate
->wm
.need_postvbl_update
= false;
2434 * Merge the watermarks from all active pipes for a specific level.
2436 static void ilk_merge_wm_level(struct drm_device
*dev
,
2438 struct intel_wm_level
*ret_wm
)
2440 const struct intel_crtc
*intel_crtc
;
2442 ret_wm
->enable
= true;
2444 for_each_intel_crtc(dev
, intel_crtc
) {
2445 const struct intel_pipe_wm
*active
= &intel_crtc
->wm
.active
.ilk
;
2446 const struct intel_wm_level
*wm
= &active
->wm
[level
];
2448 if (!active
->pipe_enabled
)
2452 * The watermark values may have been used in the past,
2453 * so we must maintain them in the registers for some
2454 * time even if the level is now disabled.
2457 ret_wm
->enable
= false;
2459 ret_wm
->pri_val
= max(ret_wm
->pri_val
, wm
->pri_val
);
2460 ret_wm
->spr_val
= max(ret_wm
->spr_val
, wm
->spr_val
);
2461 ret_wm
->cur_val
= max(ret_wm
->cur_val
, wm
->cur_val
);
2462 ret_wm
->fbc_val
= max(ret_wm
->fbc_val
, wm
->fbc_val
);
2467 * Merge all low power watermarks for all active pipes.
2469 static void ilk_wm_merge(struct drm_device
*dev
,
2470 const struct intel_wm_config
*config
,
2471 const struct ilk_wm_maximums
*max
,
2472 struct intel_pipe_wm
*merged
)
2474 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2475 int level
, max_level
= ilk_wm_max_level(dev
);
2476 int last_enabled_level
= max_level
;
2478 /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2479 if ((INTEL_INFO(dev
)->gen
<= 6 || IS_IVYBRIDGE(dev
)) &&
2480 config
->num_pipes_active
> 1)
2483 /* ILK: FBC WM must be disabled always */
2484 merged
->fbc_wm_enabled
= INTEL_INFO(dev
)->gen
>= 6;
2486 /* merge each WM1+ level */
2487 for (level
= 1; level
<= max_level
; level
++) {
2488 struct intel_wm_level
*wm
= &merged
->wm
[level
];
2490 ilk_merge_wm_level(dev
, level
, wm
);
2492 if (level
> last_enabled_level
)
2494 else if (!ilk_validate_wm_level(level
, max
, wm
))
2495 /* make sure all following levels get disabled */
2496 last_enabled_level
= level
- 1;
2499 * The spec says it is preferred to disable
2500 * FBC WMs instead of disabling a WM level.
2502 if (wm
->fbc_val
> max
->fbc
) {
2504 merged
->fbc_wm_enabled
= false;
2509 /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2511 * FIXME this is racy. FBC might get enabled later.
2512 * What we should check here is whether FBC can be
2513 * enabled sometime later.
2515 if (IS_GEN5(dev
) && !merged
->fbc_wm_enabled
&&
2516 intel_fbc_is_active(dev_priv
)) {
2517 for (level
= 2; level
<= max_level
; level
++) {
2518 struct intel_wm_level
*wm
= &merged
->wm
[level
];
2525 static int ilk_wm_lp_to_level(int wm_lp
, const struct intel_pipe_wm
*pipe_wm
)
2527 /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2528 return wm_lp
+ (wm_lp
>= 2 && pipe_wm
->wm
[4].enable
);
2531 /* The value we need to program into the WM_LPx latency field */
2532 static unsigned int ilk_wm_lp_latency(struct drm_device
*dev
, int level
)
2534 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2536 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
2539 return dev_priv
->wm
.pri_latency
[level
];
2542 static void ilk_compute_wm_results(struct drm_device
*dev
,
2543 const struct intel_pipe_wm
*merged
,
2544 enum intel_ddb_partitioning partitioning
,
2545 struct ilk_wm_values
*results
)
2547 struct intel_crtc
*intel_crtc
;
2550 results
->enable_fbc_wm
= merged
->fbc_wm_enabled
;
2551 results
->partitioning
= partitioning
;
2553 /* LP1+ register values */
2554 for (wm_lp
= 1; wm_lp
<= 3; wm_lp
++) {
2555 const struct intel_wm_level
*r
;
2557 level
= ilk_wm_lp_to_level(wm_lp
, merged
);
2559 r
= &merged
->wm
[level
];
2562 * Maintain the watermark values even if the level is
2563 * disabled. Doing otherwise could cause underruns.
2565 results
->wm_lp
[wm_lp
- 1] =
2566 (ilk_wm_lp_latency(dev
, level
) << WM1_LP_LATENCY_SHIFT
) |
2567 (r
->pri_val
<< WM1_LP_SR_SHIFT
) |
2571 results
->wm_lp
[wm_lp
- 1] |= WM1_LP_SR_EN
;
2573 if (INTEL_INFO(dev
)->gen
>= 8)
2574 results
->wm_lp
[wm_lp
- 1] |=
2575 r
->fbc_val
<< WM1_LP_FBC_SHIFT_BDW
;
2577 results
->wm_lp
[wm_lp
- 1] |=
2578 r
->fbc_val
<< WM1_LP_FBC_SHIFT
;
2581 * Always set WM1S_LP_EN when spr_val != 0, even if the
2582 * level is disabled. Doing otherwise could cause underruns.
2584 if (INTEL_INFO(dev
)->gen
<= 6 && r
->spr_val
) {
2585 WARN_ON(wm_lp
!= 1);
2586 results
->wm_lp_spr
[wm_lp
- 1] = WM1S_LP_EN
| r
->spr_val
;
2588 results
->wm_lp_spr
[wm_lp
- 1] = r
->spr_val
;
2591 /* LP0 register values */
2592 for_each_intel_crtc(dev
, intel_crtc
) {
2593 enum pipe pipe
= intel_crtc
->pipe
;
2594 const struct intel_wm_level
*r
=
2595 &intel_crtc
->wm
.active
.ilk
.wm
[0];
2597 if (WARN_ON(!r
->enable
))
2600 results
->wm_linetime
[pipe
] = intel_crtc
->wm
.active
.ilk
.linetime
;
2602 results
->wm_pipe
[pipe
] =
2603 (r
->pri_val
<< WM0_PIPE_PLANE_SHIFT
) |
2604 (r
->spr_val
<< WM0_PIPE_SPRITE_SHIFT
) |
2609 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2610 * case both are at the same level. Prefer r1 in case they're the same. */
2611 static struct intel_pipe_wm
*ilk_find_best_result(struct drm_device
*dev
,
2612 struct intel_pipe_wm
*r1
,
2613 struct intel_pipe_wm
*r2
)
2615 int level
, max_level
= ilk_wm_max_level(dev
);
2616 int level1
= 0, level2
= 0;
2618 for (level
= 1; level
<= max_level
; level
++) {
2619 if (r1
->wm
[level
].enable
)
2621 if (r2
->wm
[level
].enable
)
2625 if (level1
== level2
) {
2626 if (r2
->fbc_wm_enabled
&& !r1
->fbc_wm_enabled
)
2630 } else if (level1
> level2
) {
2637 /* dirty bits used to track which watermarks need changes */
2638 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2639 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2640 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2641 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2642 #define WM_DIRTY_FBC (1 << 24)
2643 #define WM_DIRTY_DDB (1 << 25)
2645 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private
*dev_priv
,
2646 const struct ilk_wm_values
*old
,
2647 const struct ilk_wm_values
*new)
2649 unsigned int dirty
= 0;
2653 for_each_pipe(dev_priv
, pipe
) {
2654 if (old
->wm_linetime
[pipe
] != new->wm_linetime
[pipe
]) {
2655 dirty
|= WM_DIRTY_LINETIME(pipe
);
2656 /* Must disable LP1+ watermarks too */
2657 dirty
|= WM_DIRTY_LP_ALL
;
2660 if (old
->wm_pipe
[pipe
] != new->wm_pipe
[pipe
]) {
2661 dirty
|= WM_DIRTY_PIPE(pipe
);
2662 /* Must disable LP1+ watermarks too */
2663 dirty
|= WM_DIRTY_LP_ALL
;
2667 if (old
->enable_fbc_wm
!= new->enable_fbc_wm
) {
2668 dirty
|= WM_DIRTY_FBC
;
2669 /* Must disable LP1+ watermarks too */
2670 dirty
|= WM_DIRTY_LP_ALL
;
2673 if (old
->partitioning
!= new->partitioning
) {
2674 dirty
|= WM_DIRTY_DDB
;
2675 /* Must disable LP1+ watermarks too */
2676 dirty
|= WM_DIRTY_LP_ALL
;
2679 /* LP1+ watermarks already deemed dirty, no need to continue */
2680 if (dirty
& WM_DIRTY_LP_ALL
)
2683 /* Find the lowest numbered LP1+ watermark in need of an update... */
2684 for (wm_lp
= 1; wm_lp
<= 3; wm_lp
++) {
2685 if (old
->wm_lp
[wm_lp
- 1] != new->wm_lp
[wm_lp
- 1] ||
2686 old
->wm_lp_spr
[wm_lp
- 1] != new->wm_lp_spr
[wm_lp
- 1])
2690 /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2691 for (; wm_lp
<= 3; wm_lp
++)
2692 dirty
|= WM_DIRTY_LP(wm_lp
);
2697 static bool _ilk_disable_lp_wm(struct drm_i915_private
*dev_priv
,
2700 struct ilk_wm_values
*previous
= &dev_priv
->wm
.hw
;
2701 bool changed
= false;
2703 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp
[2] & WM1_LP_SR_EN
) {
2704 previous
->wm_lp
[2] &= ~WM1_LP_SR_EN
;
2705 I915_WRITE(WM3_LP_ILK
, previous
->wm_lp
[2]);
2708 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp
[1] & WM1_LP_SR_EN
) {
2709 previous
->wm_lp
[1] &= ~WM1_LP_SR_EN
;
2710 I915_WRITE(WM2_LP_ILK
, previous
->wm_lp
[1]);
2713 if (dirty
& WM_DIRTY_LP(1) && previous
->wm_lp
[0] & WM1_LP_SR_EN
) {
2714 previous
->wm_lp
[0] &= ~WM1_LP_SR_EN
;
2715 I915_WRITE(WM1_LP_ILK
, previous
->wm_lp
[0]);
2720 * Don't touch WM1S_LP_EN here.
2721 * Doing so could cause underruns.
2728 * The spec says we shouldn't write when we don't need, because every write
2729 * causes WMs to be re-evaluated, expending some power.
2731 static void ilk_write_wm_values(struct drm_i915_private
*dev_priv
,
2732 struct ilk_wm_values
*results
)
2734 struct drm_device
*dev
= dev_priv
->dev
;
2735 struct ilk_wm_values
*previous
= &dev_priv
->wm
.hw
;
2739 dirty
= ilk_compute_wm_dirty(dev_priv
, previous
, results
);
2743 _ilk_disable_lp_wm(dev_priv
, dirty
);
2745 if (dirty
& WM_DIRTY_PIPE(PIPE_A
))
2746 I915_WRITE(WM0_PIPEA_ILK
, results
->wm_pipe
[0]);
2747 if (dirty
& WM_DIRTY_PIPE(PIPE_B
))
2748 I915_WRITE(WM0_PIPEB_ILK
, results
->wm_pipe
[1]);
2749 if (dirty
& WM_DIRTY_PIPE(PIPE_C
))
2750 I915_WRITE(WM0_PIPEC_IVB
, results
->wm_pipe
[2]);
2752 if (dirty
& WM_DIRTY_LINETIME(PIPE_A
))
2753 I915_WRITE(PIPE_WM_LINETIME(PIPE_A
), results
->wm_linetime
[0]);
2754 if (dirty
& WM_DIRTY_LINETIME(PIPE_B
))
2755 I915_WRITE(PIPE_WM_LINETIME(PIPE_B
), results
->wm_linetime
[1]);
2756 if (dirty
& WM_DIRTY_LINETIME(PIPE_C
))
2757 I915_WRITE(PIPE_WM_LINETIME(PIPE_C
), results
->wm_linetime
[2]);
2759 if (dirty
& WM_DIRTY_DDB
) {
2760 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
)) {
2761 val
= I915_READ(WM_MISC
);
2762 if (results
->partitioning
== INTEL_DDB_PART_1_2
)
2763 val
&= ~WM_MISC_DATA_PARTITION_5_6
;
2765 val
|= WM_MISC_DATA_PARTITION_5_6
;
2766 I915_WRITE(WM_MISC
, val
);
2768 val
= I915_READ(DISP_ARB_CTL2
);
2769 if (results
->partitioning
== INTEL_DDB_PART_1_2
)
2770 val
&= ~DISP_DATA_PARTITION_5_6
;
2772 val
|= DISP_DATA_PARTITION_5_6
;
2773 I915_WRITE(DISP_ARB_CTL2
, val
);
2777 if (dirty
& WM_DIRTY_FBC
) {
2778 val
= I915_READ(DISP_ARB_CTL
);
2779 if (results
->enable_fbc_wm
)
2780 val
&= ~DISP_FBC_WM_DIS
;
2782 val
|= DISP_FBC_WM_DIS
;
2783 I915_WRITE(DISP_ARB_CTL
, val
);
2786 if (dirty
& WM_DIRTY_LP(1) &&
2787 previous
->wm_lp_spr
[0] != results
->wm_lp_spr
[0])
2788 I915_WRITE(WM1S_LP_ILK
, results
->wm_lp_spr
[0]);
2790 if (INTEL_INFO(dev
)->gen
>= 7) {
2791 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp_spr
[1] != results
->wm_lp_spr
[1])
2792 I915_WRITE(WM2S_LP_IVB
, results
->wm_lp_spr
[1]);
2793 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp_spr
[2] != results
->wm_lp_spr
[2])
2794 I915_WRITE(WM3S_LP_IVB
, results
->wm_lp_spr
[2]);
2797 if (dirty
& WM_DIRTY_LP(1) && previous
->wm_lp
[0] != results
->wm_lp
[0])
2798 I915_WRITE(WM1_LP_ILK
, results
->wm_lp
[0]);
2799 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp
[1] != results
->wm_lp
[1])
2800 I915_WRITE(WM2_LP_ILK
, results
->wm_lp
[1]);
2801 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp
[2] != results
->wm_lp
[2])
2802 I915_WRITE(WM3_LP_ILK
, results
->wm_lp
[2]);
2804 dev_priv
->wm
.hw
= *results
;
2807 bool ilk_disable_lp_wm(struct drm_device
*dev
)
2809 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2811 return _ilk_disable_lp_wm(dev_priv
, WM_DIRTY_LP_ALL
);
2815 * On gen9, we need to allocate Display Data Buffer (DDB) portions to the
2816 * different active planes.
2819 #define SKL_DDB_SIZE 896 /* in blocks */
2820 #define BXT_DDB_SIZE 512
2823 * Return the index of a plane in the SKL DDB and wm result arrays. Primary
2824 * plane is always in slot 0, cursor is always in slot I915_MAX_PLANES-1, and
2825 * other universal planes are in indices 1..n. Note that this may leave unused
2826 * indices between the top "sprite" plane and the cursor.
2829 skl_wm_plane_id(const struct intel_plane
*plane
)
2831 switch (plane
->base
.type
) {
2832 case DRM_PLANE_TYPE_PRIMARY
:
2834 case DRM_PLANE_TYPE_CURSOR
:
2835 return PLANE_CURSOR
;
2836 case DRM_PLANE_TYPE_OVERLAY
:
2837 return plane
->plane
+ 1;
2839 MISSING_CASE(plane
->base
.type
);
2840 return plane
->plane
;
2845 skl_ddb_get_pipe_allocation_limits(struct drm_device
*dev
,
2846 const struct intel_crtc_state
*cstate
,
2847 const struct intel_wm_config
*config
,
2848 struct skl_ddb_entry
*alloc
/* out */)
2850 struct drm_crtc
*for_crtc
= cstate
->base
.crtc
;
2851 struct drm_crtc
*crtc
;
2852 unsigned int pipe_size
, ddb_size
;
2853 int nth_active_pipe
;
2855 if (!cstate
->base
.active
) {
2861 if (IS_BROXTON(dev
))
2862 ddb_size
= BXT_DDB_SIZE
;
2864 ddb_size
= SKL_DDB_SIZE
;
2866 ddb_size
-= 4; /* 4 blocks for bypass path allocation */
2868 nth_active_pipe
= 0;
2869 for_each_crtc(dev
, crtc
) {
2870 if (!to_intel_crtc(crtc
)->active
)
2873 if (crtc
== for_crtc
)
2879 pipe_size
= ddb_size
/ config
->num_pipes_active
;
2880 alloc
->start
= nth_active_pipe
* ddb_size
/ config
->num_pipes_active
;
2881 alloc
->end
= alloc
->start
+ pipe_size
;
2884 static unsigned int skl_cursor_allocation(const struct intel_wm_config
*config
)
2886 if (config
->num_pipes_active
== 1)
2892 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry
*entry
, u32 reg
)
2894 entry
->start
= reg
& 0x3ff;
2895 entry
->end
= (reg
>> 16) & 0x3ff;
2900 void skl_ddb_get_hw_state(struct drm_i915_private
*dev_priv
,
2901 struct skl_ddb_allocation
*ddb
/* out */)
2907 memset(ddb
, 0, sizeof(*ddb
));
2909 for_each_pipe(dev_priv
, pipe
) {
2910 if (!intel_display_power_is_enabled(dev_priv
, POWER_DOMAIN_PIPE(pipe
)))
2913 for_each_plane(dev_priv
, pipe
, plane
) {
2914 val
= I915_READ(PLANE_BUF_CFG(pipe
, plane
));
2915 skl_ddb_entry_init_from_hw(&ddb
->plane
[pipe
][plane
],
2919 val
= I915_READ(CUR_BUF_CFG(pipe
));
2920 skl_ddb_entry_init_from_hw(&ddb
->plane
[pipe
][PLANE_CURSOR
],
2926 skl_plane_relative_data_rate(const struct intel_crtc_state
*cstate
,
2927 const struct drm_plane_state
*pstate
,
2930 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
2931 struct drm_framebuffer
*fb
= pstate
->fb
;
2933 /* for planar format */
2934 if (fb
->pixel_format
== DRM_FORMAT_NV12
) {
2935 if (y
) /* y-plane data rate */
2936 return intel_crtc
->config
->pipe_src_w
*
2937 intel_crtc
->config
->pipe_src_h
*
2938 drm_format_plane_cpp(fb
->pixel_format
, 0);
2939 else /* uv-plane data rate */
2940 return (intel_crtc
->config
->pipe_src_w
/2) *
2941 (intel_crtc
->config
->pipe_src_h
/2) *
2942 drm_format_plane_cpp(fb
->pixel_format
, 1);
2945 /* for packed formats */
2946 return intel_crtc
->config
->pipe_src_w
*
2947 intel_crtc
->config
->pipe_src_h
*
2948 drm_format_plane_cpp(fb
->pixel_format
, 0);
2952 * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
2953 * a 8192x4096@32bpp framebuffer:
2954 * 3 * 4096 * 8192 * 4 < 2^32
2957 skl_get_total_relative_data_rate(const struct intel_crtc_state
*cstate
)
2959 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
2960 struct drm_device
*dev
= intel_crtc
->base
.dev
;
2961 const struct intel_plane
*intel_plane
;
2962 unsigned int total_data_rate
= 0;
2964 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
2965 const struct drm_plane_state
*pstate
= intel_plane
->base
.state
;
2967 if (pstate
->fb
== NULL
)
2970 if (intel_plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
)
2974 total_data_rate
+= skl_plane_relative_data_rate(cstate
,
2978 if (pstate
->fb
->pixel_format
== DRM_FORMAT_NV12
)
2980 total_data_rate
+= skl_plane_relative_data_rate(cstate
,
2985 return total_data_rate
;
2989 skl_allocate_pipe_ddb(struct intel_crtc_state
*cstate
,
2990 struct skl_ddb_allocation
*ddb
/* out */)
2992 struct drm_crtc
*crtc
= cstate
->base
.crtc
;
2993 struct drm_device
*dev
= crtc
->dev
;
2994 struct drm_i915_private
*dev_priv
= to_i915(dev
);
2995 struct intel_wm_config
*config
= &dev_priv
->wm
.config
;
2996 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
2997 struct intel_plane
*intel_plane
;
2998 enum pipe pipe
= intel_crtc
->pipe
;
2999 struct skl_ddb_entry
*alloc
= &ddb
->pipe
[pipe
];
3000 uint16_t alloc_size
, start
, cursor_blocks
;
3001 uint16_t minimum
[I915_MAX_PLANES
];
3002 uint16_t y_minimum
[I915_MAX_PLANES
];
3003 unsigned int total_data_rate
;
3005 skl_ddb_get_pipe_allocation_limits(dev
, cstate
, config
, alloc
);
3006 alloc_size
= skl_ddb_entry_size(alloc
);
3007 if (alloc_size
== 0) {
3008 memset(ddb
->plane
[pipe
], 0, sizeof(ddb
->plane
[pipe
]));
3009 memset(&ddb
->plane
[pipe
][PLANE_CURSOR
], 0,
3010 sizeof(ddb
->plane
[pipe
][PLANE_CURSOR
]));
3014 cursor_blocks
= skl_cursor_allocation(config
);
3015 ddb
->plane
[pipe
][PLANE_CURSOR
].start
= alloc
->end
- cursor_blocks
;
3016 ddb
->plane
[pipe
][PLANE_CURSOR
].end
= alloc
->end
;
3018 alloc_size
-= cursor_blocks
;
3019 alloc
->end
-= cursor_blocks
;
3021 /* 1. Allocate the mininum required blocks for each active plane */
3022 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
3023 struct drm_plane
*plane
= &intel_plane
->base
;
3024 struct drm_framebuffer
*fb
= plane
->state
->fb
;
3025 int id
= skl_wm_plane_id(intel_plane
);
3029 if (plane
->type
== DRM_PLANE_TYPE_CURSOR
)
3033 alloc_size
-= minimum
[id
];
3034 y_minimum
[id
] = (fb
->pixel_format
== DRM_FORMAT_NV12
) ? 8 : 0;
3035 alloc_size
-= y_minimum
[id
];
3039 * 2. Distribute the remaining space in proportion to the amount of
3040 * data each plane needs to fetch from memory.
3042 * FIXME: we may not allocate every single block here.
3044 total_data_rate
= skl_get_total_relative_data_rate(cstate
);
3046 start
= alloc
->start
;
3047 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
3048 struct drm_plane
*plane
= &intel_plane
->base
;
3049 struct drm_plane_state
*pstate
= intel_plane
->base
.state
;
3050 unsigned int data_rate
, y_data_rate
;
3051 uint16_t plane_blocks
, y_plane_blocks
= 0;
3052 int id
= skl_wm_plane_id(intel_plane
);
3054 if (pstate
->fb
== NULL
)
3056 if (plane
->type
== DRM_PLANE_TYPE_CURSOR
)
3059 data_rate
= skl_plane_relative_data_rate(cstate
, pstate
, 0);
3062 * allocation for (packed formats) or (uv-plane part of planar format):
3063 * promote the expression to 64 bits to avoid overflowing, the
3064 * result is < available as data_rate / total_data_rate < 1
3066 plane_blocks
= minimum
[id
];
3067 plane_blocks
+= div_u64((uint64_t)alloc_size
* data_rate
,
3070 ddb
->plane
[pipe
][id
].start
= start
;
3071 ddb
->plane
[pipe
][id
].end
= start
+ plane_blocks
;
3073 start
+= plane_blocks
;
3076 * allocation for y_plane part of planar format:
3078 if (pstate
->fb
->pixel_format
== DRM_FORMAT_NV12
) {
3079 y_data_rate
= skl_plane_relative_data_rate(cstate
,
3082 y_plane_blocks
= y_minimum
[id
];
3083 y_plane_blocks
+= div_u64((uint64_t)alloc_size
* y_data_rate
,
3086 ddb
->y_plane
[pipe
][id
].start
= start
;
3087 ddb
->y_plane
[pipe
][id
].end
= start
+ y_plane_blocks
;
3089 start
+= y_plane_blocks
;
3096 static uint32_t skl_pipe_pixel_rate(const struct intel_crtc_state
*config
)
3098 /* TODO: Take into account the scalers once we support them */
3099 return config
->base
.adjusted_mode
.crtc_clock
;
3103 * The max latency should be 257 (max the punit can code is 255 and we add 2us
3104 * for the read latency) and bytes_per_pixel should always be <= 8, so that
3105 * should allow pixel_rate up to ~2 GHz which seems sufficient since max
3106 * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
3108 static uint32_t skl_wm_method1(uint32_t pixel_rate
, uint8_t bytes_per_pixel
,
3111 uint32_t wm_intermediate_val
, ret
;
3116 wm_intermediate_val
= latency
* pixel_rate
* bytes_per_pixel
/ 512;
3117 ret
= DIV_ROUND_UP(wm_intermediate_val
, 1000);
3122 static uint32_t skl_wm_method2(uint32_t pixel_rate
, uint32_t pipe_htotal
,
3123 uint32_t horiz_pixels
, uint8_t bytes_per_pixel
,
3124 uint64_t tiling
, uint32_t latency
)
3127 uint32_t plane_bytes_per_line
, plane_blocks_per_line
;
3128 uint32_t wm_intermediate_val
;
3133 plane_bytes_per_line
= horiz_pixels
* bytes_per_pixel
;
3135 if (tiling
== I915_FORMAT_MOD_Y_TILED
||
3136 tiling
== I915_FORMAT_MOD_Yf_TILED
) {
3137 plane_bytes_per_line
*= 4;
3138 plane_blocks_per_line
= DIV_ROUND_UP(plane_bytes_per_line
, 512);
3139 plane_blocks_per_line
/= 4;
3141 plane_blocks_per_line
= DIV_ROUND_UP(plane_bytes_per_line
, 512);
3144 wm_intermediate_val
= latency
* pixel_rate
;
3145 ret
= DIV_ROUND_UP(wm_intermediate_val
, pipe_htotal
* 1000) *
3146 plane_blocks_per_line
;
3151 static bool skl_ddb_allocation_changed(const struct skl_ddb_allocation
*new_ddb
,
3152 const struct intel_crtc
*intel_crtc
)
3154 struct drm_device
*dev
= intel_crtc
->base
.dev
;
3155 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3156 const struct skl_ddb_allocation
*cur_ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
3159 * If ddb allocation of pipes changed, it may require recalculation of
3162 if (memcmp(new_ddb
->pipe
, cur_ddb
->pipe
, sizeof(new_ddb
->pipe
)))
3168 static bool skl_compute_plane_wm(const struct drm_i915_private
*dev_priv
,
3169 struct intel_crtc_state
*cstate
,
3170 struct intel_plane
*intel_plane
,
3171 uint16_t ddb_allocation
,
3173 uint16_t *out_blocks
, /* out */
3174 uint8_t *out_lines
/* out */)
3176 struct drm_plane
*plane
= &intel_plane
->base
;
3177 struct drm_framebuffer
*fb
= plane
->state
->fb
;
3178 uint32_t latency
= dev_priv
->wm
.skl_latency
[level
];
3179 uint32_t method1
, method2
;
3180 uint32_t plane_bytes_per_line
, plane_blocks_per_line
;
3181 uint32_t res_blocks
, res_lines
;
3182 uint32_t selected_result
;
3183 uint8_t bytes_per_pixel
;
3185 if (latency
== 0 || !cstate
->base
.active
|| !fb
)
3188 bytes_per_pixel
= drm_format_plane_cpp(fb
->pixel_format
, 0);
3189 method1
= skl_wm_method1(skl_pipe_pixel_rate(cstate
),
3192 method2
= skl_wm_method2(skl_pipe_pixel_rate(cstate
),
3193 cstate
->base
.adjusted_mode
.crtc_htotal
,
3199 plane_bytes_per_line
= cstate
->pipe_src_w
* bytes_per_pixel
;
3200 plane_blocks_per_line
= DIV_ROUND_UP(plane_bytes_per_line
, 512);
3202 if (fb
->modifier
[0] == I915_FORMAT_MOD_Y_TILED
||
3203 fb
->modifier
[0] == I915_FORMAT_MOD_Yf_TILED
) {
3204 uint32_t min_scanlines
= 4;
3205 uint32_t y_tile_minimum
;
3206 if (intel_rotation_90_or_270(plane
->state
->rotation
)) {
3207 int bpp
= (fb
->pixel_format
== DRM_FORMAT_NV12
) ?
3208 drm_format_plane_cpp(fb
->pixel_format
, 1) :
3209 drm_format_plane_cpp(fb
->pixel_format
, 0);
3219 WARN(1, "Unsupported pixel depth for rotation");
3222 y_tile_minimum
= plane_blocks_per_line
* min_scanlines
;
3223 selected_result
= max(method2
, y_tile_minimum
);
3225 if ((ddb_allocation
/ plane_blocks_per_line
) >= 1)
3226 selected_result
= min(method1
, method2
);
3228 selected_result
= method1
;
3231 res_blocks
= selected_result
+ 1;
3232 res_lines
= DIV_ROUND_UP(selected_result
, plane_blocks_per_line
);
3234 if (level
>= 1 && level
<= 7) {
3235 if (fb
->modifier
[0] == I915_FORMAT_MOD_Y_TILED
||
3236 fb
->modifier
[0] == I915_FORMAT_MOD_Yf_TILED
)
3242 if (res_blocks
>= ddb_allocation
|| res_lines
> 31)
3245 *out_blocks
= res_blocks
;
3246 *out_lines
= res_lines
;
3251 static void skl_compute_wm_level(const struct drm_i915_private
*dev_priv
,
3252 struct skl_ddb_allocation
*ddb
,
3253 struct intel_crtc_state
*cstate
,
3255 struct skl_wm_level
*result
)
3257 struct drm_device
*dev
= dev_priv
->dev
;
3258 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
3259 struct intel_plane
*intel_plane
;
3260 uint16_t ddb_blocks
;
3261 enum pipe pipe
= intel_crtc
->pipe
;
3263 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
3264 int i
= skl_wm_plane_id(intel_plane
);
3266 ddb_blocks
= skl_ddb_entry_size(&ddb
->plane
[pipe
][i
]);
3268 result
->plane_en
[i
] = skl_compute_plane_wm(dev_priv
,
3273 &result
->plane_res_b
[i
],
3274 &result
->plane_res_l
[i
]);
3279 skl_compute_linetime_wm(struct intel_crtc_state
*cstate
)
3281 if (!cstate
->base
.active
)
3284 if (WARN_ON(skl_pipe_pixel_rate(cstate
) == 0))
3287 return DIV_ROUND_UP(8 * cstate
->base
.adjusted_mode
.crtc_htotal
* 1000,
3288 skl_pipe_pixel_rate(cstate
));
3291 static void skl_compute_transition_wm(struct intel_crtc_state
*cstate
,
3292 struct skl_wm_level
*trans_wm
/* out */)
3294 struct drm_crtc
*crtc
= cstate
->base
.crtc
;
3295 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3296 struct intel_plane
*intel_plane
;
3298 if (!cstate
->base
.active
)
3301 /* Until we know more, just disable transition WMs */
3302 for_each_intel_plane_on_crtc(crtc
->dev
, intel_crtc
, intel_plane
) {
3303 int i
= skl_wm_plane_id(intel_plane
);
3305 trans_wm
->plane_en
[i
] = false;
3309 static void skl_compute_pipe_wm(struct intel_crtc_state
*cstate
,
3310 struct skl_ddb_allocation
*ddb
,
3311 struct skl_pipe_wm
*pipe_wm
)
3313 struct drm_device
*dev
= cstate
->base
.crtc
->dev
;
3314 const struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3315 int level
, max_level
= ilk_wm_max_level(dev
);
3317 for (level
= 0; level
<= max_level
; level
++) {
3318 skl_compute_wm_level(dev_priv
, ddb
, cstate
,
3319 level
, &pipe_wm
->wm
[level
]);
3321 pipe_wm
->linetime
= skl_compute_linetime_wm(cstate
);
3323 skl_compute_transition_wm(cstate
, &pipe_wm
->trans_wm
);
3326 static void skl_compute_wm_results(struct drm_device
*dev
,
3327 struct skl_pipe_wm
*p_wm
,
3328 struct skl_wm_values
*r
,
3329 struct intel_crtc
*intel_crtc
)
3331 int level
, max_level
= ilk_wm_max_level(dev
);
3332 enum pipe pipe
= intel_crtc
->pipe
;
3336 for (level
= 0; level
<= max_level
; level
++) {
3337 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++) {
3340 temp
|= p_wm
->wm
[level
].plane_res_l
[i
] <<
3341 PLANE_WM_LINES_SHIFT
;
3342 temp
|= p_wm
->wm
[level
].plane_res_b
[i
];
3343 if (p_wm
->wm
[level
].plane_en
[i
])
3344 temp
|= PLANE_WM_EN
;
3346 r
->plane
[pipe
][i
][level
] = temp
;
3351 temp
|= p_wm
->wm
[level
].plane_res_l
[PLANE_CURSOR
] << PLANE_WM_LINES_SHIFT
;
3352 temp
|= p_wm
->wm
[level
].plane_res_b
[PLANE_CURSOR
];
3354 if (p_wm
->wm
[level
].plane_en
[PLANE_CURSOR
])
3355 temp
|= PLANE_WM_EN
;
3357 r
->plane
[pipe
][PLANE_CURSOR
][level
] = temp
;
3361 /* transition WMs */
3362 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++) {
3364 temp
|= p_wm
->trans_wm
.plane_res_l
[i
] << PLANE_WM_LINES_SHIFT
;
3365 temp
|= p_wm
->trans_wm
.plane_res_b
[i
];
3366 if (p_wm
->trans_wm
.plane_en
[i
])
3367 temp
|= PLANE_WM_EN
;
3369 r
->plane_trans
[pipe
][i
] = temp
;
3373 temp
|= p_wm
->trans_wm
.plane_res_l
[PLANE_CURSOR
] << PLANE_WM_LINES_SHIFT
;
3374 temp
|= p_wm
->trans_wm
.plane_res_b
[PLANE_CURSOR
];
3375 if (p_wm
->trans_wm
.plane_en
[PLANE_CURSOR
])
3376 temp
|= PLANE_WM_EN
;
3378 r
->plane_trans
[pipe
][PLANE_CURSOR
] = temp
;
3380 r
->wm_linetime
[pipe
] = p_wm
->linetime
;
3383 static void skl_ddb_entry_write(struct drm_i915_private
*dev_priv
,
3385 const struct skl_ddb_entry
*entry
)
3388 I915_WRITE(reg
, (entry
->end
- 1) << 16 | entry
->start
);
3393 static void skl_write_wm_values(struct drm_i915_private
*dev_priv
,
3394 const struct skl_wm_values
*new)
3396 struct drm_device
*dev
= dev_priv
->dev
;
3397 struct intel_crtc
*crtc
;
3399 for_each_intel_crtc(dev
, crtc
) {
3400 int i
, level
, max_level
= ilk_wm_max_level(dev
);
3401 enum pipe pipe
= crtc
->pipe
;
3403 if (!new->dirty
[pipe
])
3406 I915_WRITE(PIPE_WM_LINETIME(pipe
), new->wm_linetime
[pipe
]);
3408 for (level
= 0; level
<= max_level
; level
++) {
3409 for (i
= 0; i
< intel_num_planes(crtc
); i
++)
3410 I915_WRITE(PLANE_WM(pipe
, i
, level
),
3411 new->plane
[pipe
][i
][level
]);
3412 I915_WRITE(CUR_WM(pipe
, level
),
3413 new->plane
[pipe
][PLANE_CURSOR
][level
]);
3415 for (i
= 0; i
< intel_num_planes(crtc
); i
++)
3416 I915_WRITE(PLANE_WM_TRANS(pipe
, i
),
3417 new->plane_trans
[pipe
][i
]);
3418 I915_WRITE(CUR_WM_TRANS(pipe
),
3419 new->plane_trans
[pipe
][PLANE_CURSOR
]);
3421 for (i
= 0; i
< intel_num_planes(crtc
); i
++) {
3422 skl_ddb_entry_write(dev_priv
,
3423 PLANE_BUF_CFG(pipe
, i
),
3424 &new->ddb
.plane
[pipe
][i
]);
3425 skl_ddb_entry_write(dev_priv
,
3426 PLANE_NV12_BUF_CFG(pipe
, i
),
3427 &new->ddb
.y_plane
[pipe
][i
]);
3430 skl_ddb_entry_write(dev_priv
, CUR_BUF_CFG(pipe
),
3431 &new->ddb
.plane
[pipe
][PLANE_CURSOR
]);
3436 * When setting up a new DDB allocation arrangement, we need to correctly
3437 * sequence the times at which the new allocations for the pipes are taken into
3438 * account or we'll have pipes fetching from space previously allocated to
3441 * Roughly the sequence looks like:
3442 * 1. re-allocate the pipe(s) with the allocation being reduced and not
3443 * overlapping with a previous light-up pipe (another way to put it is:
3444 * pipes with their new allocation strickly included into their old ones).
3445 * 2. re-allocate the other pipes that get their allocation reduced
3446 * 3. allocate the pipes having their allocation increased
3448 * Steps 1. and 2. are here to take care of the following case:
3449 * - Initially DDB looks like this:
3452 * - pipe B has a reduced DDB allocation that overlaps with the old pipe C
3456 * We need to sequence the re-allocation: C, B, A (and not B, C, A).
3460 skl_wm_flush_pipe(struct drm_i915_private
*dev_priv
, enum pipe pipe
, int pass
)
3464 DRM_DEBUG_KMS("flush pipe %c (pass %d)\n", pipe_name(pipe
), pass
);
3466 for_each_plane(dev_priv
, pipe
, plane
) {
3467 I915_WRITE(PLANE_SURF(pipe
, plane
),
3468 I915_READ(PLANE_SURF(pipe
, plane
)));
3470 I915_WRITE(CURBASE(pipe
), I915_READ(CURBASE(pipe
)));
3474 skl_ddb_allocation_included(const struct skl_ddb_allocation
*old
,
3475 const struct skl_ddb_allocation
*new,
3478 uint16_t old_size
, new_size
;
3480 old_size
= skl_ddb_entry_size(&old
->pipe
[pipe
]);
3481 new_size
= skl_ddb_entry_size(&new->pipe
[pipe
]);
3483 return old_size
!= new_size
&&
3484 new->pipe
[pipe
].start
>= old
->pipe
[pipe
].start
&&
3485 new->pipe
[pipe
].end
<= old
->pipe
[pipe
].end
;
3488 static void skl_flush_wm_values(struct drm_i915_private
*dev_priv
,
3489 struct skl_wm_values
*new_values
)
3491 struct drm_device
*dev
= dev_priv
->dev
;
3492 struct skl_ddb_allocation
*cur_ddb
, *new_ddb
;
3493 bool reallocated
[I915_MAX_PIPES
] = {};
3494 struct intel_crtc
*crtc
;
3497 new_ddb
= &new_values
->ddb
;
3498 cur_ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
3501 * First pass: flush the pipes with the new allocation contained into
3504 * We'll wait for the vblank on those pipes to ensure we can safely
3505 * re-allocate the freed space without this pipe fetching from it.
3507 for_each_intel_crtc(dev
, crtc
) {
3513 if (!skl_ddb_allocation_included(cur_ddb
, new_ddb
, pipe
))
3516 skl_wm_flush_pipe(dev_priv
, pipe
, 1);
3517 intel_wait_for_vblank(dev
, pipe
);
3519 reallocated
[pipe
] = true;
3524 * Second pass: flush the pipes that are having their allocation
3525 * reduced, but overlapping with a previous allocation.
3527 * Here as well we need to wait for the vblank to make sure the freed
3528 * space is not used anymore.
3530 for_each_intel_crtc(dev
, crtc
) {
3536 if (reallocated
[pipe
])
3539 if (skl_ddb_entry_size(&new_ddb
->pipe
[pipe
]) <
3540 skl_ddb_entry_size(&cur_ddb
->pipe
[pipe
])) {
3541 skl_wm_flush_pipe(dev_priv
, pipe
, 2);
3542 intel_wait_for_vblank(dev
, pipe
);
3543 reallocated
[pipe
] = true;
3548 * Third pass: flush the pipes that got more space allocated.
3550 * We don't need to actively wait for the update here, next vblank
3551 * will just get more DDB space with the correct WM values.
3553 for_each_intel_crtc(dev
, crtc
) {
3560 * At this point, only the pipes more space than before are
3561 * left to re-allocate.
3563 if (reallocated
[pipe
])
3566 skl_wm_flush_pipe(dev_priv
, pipe
, 3);
3570 static bool skl_update_pipe_wm(struct drm_crtc
*crtc
,
3571 struct skl_ddb_allocation
*ddb
, /* out */
3572 struct skl_pipe_wm
*pipe_wm
/* out */)
3574 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3575 struct intel_crtc_state
*cstate
= to_intel_crtc_state(crtc
->state
);
3577 skl_allocate_pipe_ddb(cstate
, ddb
);
3578 skl_compute_pipe_wm(cstate
, ddb
, pipe_wm
);
3580 if (!memcmp(&intel_crtc
->wm
.active
.skl
, pipe_wm
, sizeof(*pipe_wm
)))
3583 intel_crtc
->wm
.active
.skl
= *pipe_wm
;
3588 static void skl_update_other_pipe_wm(struct drm_device
*dev
,
3589 struct drm_crtc
*crtc
,
3590 struct skl_wm_values
*r
)
3592 struct intel_crtc
*intel_crtc
;
3593 struct intel_crtc
*this_crtc
= to_intel_crtc(crtc
);
3596 * If the WM update hasn't changed the allocation for this_crtc (the
3597 * crtc we are currently computing the new WM values for), other
3598 * enabled crtcs will keep the same allocation and we don't need to
3599 * recompute anything for them.
3601 if (!skl_ddb_allocation_changed(&r
->ddb
, this_crtc
))
3605 * Otherwise, because of this_crtc being freshly enabled/disabled, the
3606 * other active pipes need new DDB allocation and WM values.
3608 for_each_intel_crtc(dev
, intel_crtc
) {
3609 struct skl_pipe_wm pipe_wm
= {};
3612 if (this_crtc
->pipe
== intel_crtc
->pipe
)
3615 if (!intel_crtc
->active
)
3618 wm_changed
= skl_update_pipe_wm(&intel_crtc
->base
,
3622 * If we end up re-computing the other pipe WM values, it's
3623 * because it was really needed, so we expect the WM values to
3626 WARN_ON(!wm_changed
);
3628 skl_compute_wm_results(dev
, &pipe_wm
, r
, intel_crtc
);
3629 r
->dirty
[intel_crtc
->pipe
] = true;
3633 static void skl_clear_wm(struct skl_wm_values
*watermarks
, enum pipe pipe
)
3635 watermarks
->wm_linetime
[pipe
] = 0;
3636 memset(watermarks
->plane
[pipe
], 0,
3637 sizeof(uint32_t) * 8 * I915_MAX_PLANES
);
3638 memset(watermarks
->plane_trans
[pipe
],
3639 0, sizeof(uint32_t) * I915_MAX_PLANES
);
3640 watermarks
->plane_trans
[pipe
][PLANE_CURSOR
] = 0;
3642 /* Clear ddb entries for pipe */
3643 memset(&watermarks
->ddb
.pipe
[pipe
], 0, sizeof(struct skl_ddb_entry
));
3644 memset(&watermarks
->ddb
.plane
[pipe
], 0,
3645 sizeof(struct skl_ddb_entry
) * I915_MAX_PLANES
);
3646 memset(&watermarks
->ddb
.y_plane
[pipe
], 0,
3647 sizeof(struct skl_ddb_entry
) * I915_MAX_PLANES
);
3648 memset(&watermarks
->ddb
.plane
[pipe
][PLANE_CURSOR
], 0,
3649 sizeof(struct skl_ddb_entry
));
3653 static void skl_update_wm(struct drm_crtc
*crtc
)
3655 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3656 struct drm_device
*dev
= crtc
->dev
;
3657 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3658 struct skl_wm_values
*results
= &dev_priv
->wm
.skl_results
;
3659 struct intel_crtc_state
*cstate
= to_intel_crtc_state(crtc
->state
);
3660 struct skl_pipe_wm
*pipe_wm
= &cstate
->wm
.optimal
.skl
;
3663 /* Clear all dirty flags */
3664 memset(results
->dirty
, 0, sizeof(bool) * I915_MAX_PIPES
);
3666 skl_clear_wm(results
, intel_crtc
->pipe
);
3668 if (!skl_update_pipe_wm(crtc
, &results
->ddb
, pipe_wm
))
3671 skl_compute_wm_results(dev
, pipe_wm
, results
, intel_crtc
);
3672 results
->dirty
[intel_crtc
->pipe
] = true;
3674 skl_update_other_pipe_wm(dev
, crtc
, results
);
3675 skl_write_wm_values(dev_priv
, results
);
3676 skl_flush_wm_values(dev_priv
, results
);
3678 /* store the new configuration */
3679 dev_priv
->wm
.skl_hw
= *results
;
3682 static void ilk_program_watermarks(struct drm_i915_private
*dev_priv
)
3684 struct drm_device
*dev
= dev_priv
->dev
;
3685 struct intel_pipe_wm lp_wm_1_2
= {}, lp_wm_5_6
= {}, *best_lp_wm
;
3686 struct ilk_wm_maximums max
;
3687 struct intel_wm_config
*config
= &dev_priv
->wm
.config
;
3688 struct ilk_wm_values results
= {};
3689 enum intel_ddb_partitioning partitioning
;
3691 ilk_compute_wm_maximums(dev
, 1, config
, INTEL_DDB_PART_1_2
, &max
);
3692 ilk_wm_merge(dev
, config
, &max
, &lp_wm_1_2
);
3694 /* 5/6 split only in single pipe config on IVB+ */
3695 if (INTEL_INFO(dev
)->gen
>= 7 &&
3696 config
->num_pipes_active
== 1 && config
->sprites_enabled
) {
3697 ilk_compute_wm_maximums(dev
, 1, config
, INTEL_DDB_PART_5_6
, &max
);
3698 ilk_wm_merge(dev
, config
, &max
, &lp_wm_5_6
);
3700 best_lp_wm
= ilk_find_best_result(dev
, &lp_wm_1_2
, &lp_wm_5_6
);
3702 best_lp_wm
= &lp_wm_1_2
;
3705 partitioning
= (best_lp_wm
== &lp_wm_1_2
) ?
3706 INTEL_DDB_PART_1_2
: INTEL_DDB_PART_5_6
;
3708 ilk_compute_wm_results(dev
, best_lp_wm
, partitioning
, &results
);
3710 ilk_write_wm_values(dev_priv
, &results
);
3713 static void ilk_initial_watermarks(struct intel_crtc_state
*cstate
)
3715 struct drm_i915_private
*dev_priv
= to_i915(cstate
->base
.crtc
->dev
);
3716 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
3718 mutex_lock(&dev_priv
->wm
.wm_mutex
);
3719 intel_crtc
->wm
.active
.ilk
= cstate
->wm
.intermediate
;
3720 ilk_program_watermarks(dev_priv
);
3721 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
3724 static void ilk_optimize_watermarks(struct intel_crtc_state
*cstate
)
3726 struct drm_i915_private
*dev_priv
= to_i915(cstate
->base
.crtc
->dev
);
3727 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
3729 mutex_lock(&dev_priv
->wm
.wm_mutex
);
3730 if (cstate
->wm
.need_postvbl_update
) {
3731 intel_crtc
->wm
.active
.ilk
= cstate
->wm
.optimal
.ilk
;
3732 ilk_program_watermarks(dev_priv
);
3734 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
3737 static void skl_pipe_wm_active_state(uint32_t val
,
3738 struct skl_pipe_wm
*active
,
3744 bool is_enabled
= (val
& PLANE_WM_EN
) != 0;
3748 active
->wm
[level
].plane_en
[i
] = is_enabled
;
3749 active
->wm
[level
].plane_res_b
[i
] =
3750 val
& PLANE_WM_BLOCKS_MASK
;
3751 active
->wm
[level
].plane_res_l
[i
] =
3752 (val
>> PLANE_WM_LINES_SHIFT
) &
3753 PLANE_WM_LINES_MASK
;
3755 active
->wm
[level
].plane_en
[PLANE_CURSOR
] = is_enabled
;
3756 active
->wm
[level
].plane_res_b
[PLANE_CURSOR
] =
3757 val
& PLANE_WM_BLOCKS_MASK
;
3758 active
->wm
[level
].plane_res_l
[PLANE_CURSOR
] =
3759 (val
>> PLANE_WM_LINES_SHIFT
) &
3760 PLANE_WM_LINES_MASK
;
3764 active
->trans_wm
.plane_en
[i
] = is_enabled
;
3765 active
->trans_wm
.plane_res_b
[i
] =
3766 val
& PLANE_WM_BLOCKS_MASK
;
3767 active
->trans_wm
.plane_res_l
[i
] =
3768 (val
>> PLANE_WM_LINES_SHIFT
) &
3769 PLANE_WM_LINES_MASK
;
3771 active
->trans_wm
.plane_en
[PLANE_CURSOR
] = is_enabled
;
3772 active
->trans_wm
.plane_res_b
[PLANE_CURSOR
] =
3773 val
& PLANE_WM_BLOCKS_MASK
;
3774 active
->trans_wm
.plane_res_l
[PLANE_CURSOR
] =
3775 (val
>> PLANE_WM_LINES_SHIFT
) &
3776 PLANE_WM_LINES_MASK
;
3781 static void skl_pipe_wm_get_hw_state(struct drm_crtc
*crtc
)
3783 struct drm_device
*dev
= crtc
->dev
;
3784 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3785 struct skl_wm_values
*hw
= &dev_priv
->wm
.skl_hw
;
3786 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3787 struct intel_crtc_state
*cstate
= to_intel_crtc_state(crtc
->state
);
3788 struct skl_pipe_wm
*active
= &cstate
->wm
.optimal
.skl
;
3789 enum pipe pipe
= intel_crtc
->pipe
;
3790 int level
, i
, max_level
;
3793 max_level
= ilk_wm_max_level(dev
);
3795 hw
->wm_linetime
[pipe
] = I915_READ(PIPE_WM_LINETIME(pipe
));
3797 for (level
= 0; level
<= max_level
; level
++) {
3798 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++)
3799 hw
->plane
[pipe
][i
][level
] =
3800 I915_READ(PLANE_WM(pipe
, i
, level
));
3801 hw
->plane
[pipe
][PLANE_CURSOR
][level
] = I915_READ(CUR_WM(pipe
, level
));
3804 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++)
3805 hw
->plane_trans
[pipe
][i
] = I915_READ(PLANE_WM_TRANS(pipe
, i
));
3806 hw
->plane_trans
[pipe
][PLANE_CURSOR
] = I915_READ(CUR_WM_TRANS(pipe
));
3808 if (!intel_crtc
->active
)
3811 hw
->dirty
[pipe
] = true;
3813 active
->linetime
= hw
->wm_linetime
[pipe
];
3815 for (level
= 0; level
<= max_level
; level
++) {
3816 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++) {
3817 temp
= hw
->plane
[pipe
][i
][level
];
3818 skl_pipe_wm_active_state(temp
, active
, false,
3821 temp
= hw
->plane
[pipe
][PLANE_CURSOR
][level
];
3822 skl_pipe_wm_active_state(temp
, active
, false, true, i
, level
);
3825 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++) {
3826 temp
= hw
->plane_trans
[pipe
][i
];
3827 skl_pipe_wm_active_state(temp
, active
, true, false, i
, 0);
3830 temp
= hw
->plane_trans
[pipe
][PLANE_CURSOR
];
3831 skl_pipe_wm_active_state(temp
, active
, true, true, i
, 0);
3833 intel_crtc
->wm
.active
.skl
= *active
;
3836 void skl_wm_get_hw_state(struct drm_device
*dev
)
3838 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3839 struct skl_ddb_allocation
*ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
3840 struct drm_crtc
*crtc
;
3842 skl_ddb_get_hw_state(dev_priv
, ddb
);
3843 list_for_each_entry(crtc
, &dev
->mode_config
.crtc_list
, head
)
3844 skl_pipe_wm_get_hw_state(crtc
);
3847 static void ilk_pipe_wm_get_hw_state(struct drm_crtc
*crtc
)
3849 struct drm_device
*dev
= crtc
->dev
;
3850 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3851 struct ilk_wm_values
*hw
= &dev_priv
->wm
.hw
;
3852 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3853 struct intel_crtc_state
*cstate
= to_intel_crtc_state(crtc
->state
);
3854 struct intel_pipe_wm
*active
= &cstate
->wm
.optimal
.ilk
;
3855 enum pipe pipe
= intel_crtc
->pipe
;
3856 static const i915_reg_t wm0_pipe_reg
[] = {
3857 [PIPE_A
] = WM0_PIPEA_ILK
,
3858 [PIPE_B
] = WM0_PIPEB_ILK
,
3859 [PIPE_C
] = WM0_PIPEC_IVB
,
3862 hw
->wm_pipe
[pipe
] = I915_READ(wm0_pipe_reg
[pipe
]);
3863 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
3864 hw
->wm_linetime
[pipe
] = I915_READ(PIPE_WM_LINETIME(pipe
));
3866 active
->pipe_enabled
= intel_crtc
->active
;
3868 if (active
->pipe_enabled
) {
3869 u32 tmp
= hw
->wm_pipe
[pipe
];
3872 * For active pipes LP0 watermark is marked as
3873 * enabled, and LP1+ watermaks as disabled since
3874 * we can't really reverse compute them in case
3875 * multiple pipes are active.
3877 active
->wm
[0].enable
= true;
3878 active
->wm
[0].pri_val
= (tmp
& WM0_PIPE_PLANE_MASK
) >> WM0_PIPE_PLANE_SHIFT
;
3879 active
->wm
[0].spr_val
= (tmp
& WM0_PIPE_SPRITE_MASK
) >> WM0_PIPE_SPRITE_SHIFT
;
3880 active
->wm
[0].cur_val
= tmp
& WM0_PIPE_CURSOR_MASK
;
3881 active
->linetime
= hw
->wm_linetime
[pipe
];
3883 int level
, max_level
= ilk_wm_max_level(dev
);
3886 * For inactive pipes, all watermark levels
3887 * should be marked as enabled but zeroed,
3888 * which is what we'd compute them to.
3890 for (level
= 0; level
<= max_level
; level
++)
3891 active
->wm
[level
].enable
= true;
3894 intel_crtc
->wm
.active
.ilk
= *active
;
3897 #define _FW_WM(value, plane) \
3898 (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
3899 #define _FW_WM_VLV(value, plane) \
3900 (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
3902 static void vlv_read_wm_values(struct drm_i915_private
*dev_priv
,
3903 struct vlv_wm_values
*wm
)
3908 for_each_pipe(dev_priv
, pipe
) {
3909 tmp
= I915_READ(VLV_DDL(pipe
));
3911 wm
->ddl
[pipe
].primary
=
3912 (tmp
>> DDL_PLANE_SHIFT
) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
3913 wm
->ddl
[pipe
].cursor
=
3914 (tmp
>> DDL_CURSOR_SHIFT
) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
3915 wm
->ddl
[pipe
].sprite
[0] =
3916 (tmp
>> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
3917 wm
->ddl
[pipe
].sprite
[1] =
3918 (tmp
>> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
3921 tmp
= I915_READ(DSPFW1
);
3922 wm
->sr
.plane
= _FW_WM(tmp
, SR
);
3923 wm
->pipe
[PIPE_B
].cursor
= _FW_WM(tmp
, CURSORB
);
3924 wm
->pipe
[PIPE_B
].primary
= _FW_WM_VLV(tmp
, PLANEB
);
3925 wm
->pipe
[PIPE_A
].primary
= _FW_WM_VLV(tmp
, PLANEA
);
3927 tmp
= I915_READ(DSPFW2
);
3928 wm
->pipe
[PIPE_A
].sprite
[1] = _FW_WM_VLV(tmp
, SPRITEB
);
3929 wm
->pipe
[PIPE_A
].cursor
= _FW_WM(tmp
, CURSORA
);
3930 wm
->pipe
[PIPE_A
].sprite
[0] = _FW_WM_VLV(tmp
, SPRITEA
);
3932 tmp
= I915_READ(DSPFW3
);
3933 wm
->sr
.cursor
= _FW_WM(tmp
, CURSOR_SR
);
3935 if (IS_CHERRYVIEW(dev_priv
)) {
3936 tmp
= I915_READ(DSPFW7_CHV
);
3937 wm
->pipe
[PIPE_B
].sprite
[1] = _FW_WM_VLV(tmp
, SPRITED
);
3938 wm
->pipe
[PIPE_B
].sprite
[0] = _FW_WM_VLV(tmp
, SPRITEC
);
3940 tmp
= I915_READ(DSPFW8_CHV
);
3941 wm
->pipe
[PIPE_C
].sprite
[1] = _FW_WM_VLV(tmp
, SPRITEF
);
3942 wm
->pipe
[PIPE_C
].sprite
[0] = _FW_WM_VLV(tmp
, SPRITEE
);
3944 tmp
= I915_READ(DSPFW9_CHV
);
3945 wm
->pipe
[PIPE_C
].primary
= _FW_WM_VLV(tmp
, PLANEC
);
3946 wm
->pipe
[PIPE_C
].cursor
= _FW_WM(tmp
, CURSORC
);
3948 tmp
= I915_READ(DSPHOWM
);
3949 wm
->sr
.plane
|= _FW_WM(tmp
, SR_HI
) << 9;
3950 wm
->pipe
[PIPE_C
].sprite
[1] |= _FW_WM(tmp
, SPRITEF_HI
) << 8;
3951 wm
->pipe
[PIPE_C
].sprite
[0] |= _FW_WM(tmp
, SPRITEE_HI
) << 8;
3952 wm
->pipe
[PIPE_C
].primary
|= _FW_WM(tmp
, PLANEC_HI
) << 8;
3953 wm
->pipe
[PIPE_B
].sprite
[1] |= _FW_WM(tmp
, SPRITED_HI
) << 8;
3954 wm
->pipe
[PIPE_B
].sprite
[0] |= _FW_WM(tmp
, SPRITEC_HI
) << 8;
3955 wm
->pipe
[PIPE_B
].primary
|= _FW_WM(tmp
, PLANEB_HI
) << 8;
3956 wm
->pipe
[PIPE_A
].sprite
[1] |= _FW_WM(tmp
, SPRITEB_HI
) << 8;
3957 wm
->pipe
[PIPE_A
].sprite
[0] |= _FW_WM(tmp
, SPRITEA_HI
) << 8;
3958 wm
->pipe
[PIPE_A
].primary
|= _FW_WM(tmp
, PLANEA_HI
) << 8;
3960 tmp
= I915_READ(DSPFW7
);
3961 wm
->pipe
[PIPE_B
].sprite
[1] = _FW_WM_VLV(tmp
, SPRITED
);
3962 wm
->pipe
[PIPE_B
].sprite
[0] = _FW_WM_VLV(tmp
, SPRITEC
);
3964 tmp
= I915_READ(DSPHOWM
);
3965 wm
->sr
.plane
|= _FW_WM(tmp
, SR_HI
) << 9;
3966 wm
->pipe
[PIPE_B
].sprite
[1] |= _FW_WM(tmp
, SPRITED_HI
) << 8;
3967 wm
->pipe
[PIPE_B
].sprite
[0] |= _FW_WM(tmp
, SPRITEC_HI
) << 8;
3968 wm
->pipe
[PIPE_B
].primary
|= _FW_WM(tmp
, PLANEB_HI
) << 8;
3969 wm
->pipe
[PIPE_A
].sprite
[1] |= _FW_WM(tmp
, SPRITEB_HI
) << 8;
3970 wm
->pipe
[PIPE_A
].sprite
[0] |= _FW_WM(tmp
, SPRITEA_HI
) << 8;
3971 wm
->pipe
[PIPE_A
].primary
|= _FW_WM(tmp
, PLANEA_HI
) << 8;
3978 void vlv_wm_get_hw_state(struct drm_device
*dev
)
3980 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3981 struct vlv_wm_values
*wm
= &dev_priv
->wm
.vlv
;
3982 struct intel_plane
*plane
;
3986 vlv_read_wm_values(dev_priv
, wm
);
3988 for_each_intel_plane(dev
, plane
) {
3989 switch (plane
->base
.type
) {
3991 case DRM_PLANE_TYPE_CURSOR
:
3992 plane
->wm
.fifo_size
= 63;
3994 case DRM_PLANE_TYPE_PRIMARY
:
3995 plane
->wm
.fifo_size
= vlv_get_fifo_size(dev
, plane
->pipe
, 0);
3997 case DRM_PLANE_TYPE_OVERLAY
:
3998 sprite
= plane
->plane
;
3999 plane
->wm
.fifo_size
= vlv_get_fifo_size(dev
, plane
->pipe
, sprite
+ 1);
4004 wm
->cxsr
= I915_READ(FW_BLC_SELF_VLV
) & FW_CSPWRDWNEN
;
4005 wm
->level
= VLV_WM_LEVEL_PM2
;
4007 if (IS_CHERRYVIEW(dev_priv
)) {
4008 mutex_lock(&dev_priv
->rps
.hw_lock
);
4010 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DSPFREQ
);
4011 if (val
& DSP_MAXFIFO_PM5_ENABLE
)
4012 wm
->level
= VLV_WM_LEVEL_PM5
;
4015 * If DDR DVFS is disabled in the BIOS, Punit
4016 * will never ack the request. So if that happens
4017 * assume we don't have to enable/disable DDR DVFS
4018 * dynamically. To test that just set the REQ_ACK
4019 * bit to poke the Punit, but don't change the
4020 * HIGH/LOW bits so that we don't actually change
4021 * the current state.
4023 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
);
4024 val
|= FORCE_DDR_FREQ_REQ_ACK
;
4025 vlv_punit_write(dev_priv
, PUNIT_REG_DDR_SETUP2
, val
);
4027 if (wait_for((vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
) &
4028 FORCE_DDR_FREQ_REQ_ACK
) == 0, 3)) {
4029 DRM_DEBUG_KMS("Punit not acking DDR DVFS request, "
4030 "assuming DDR DVFS is disabled\n");
4031 dev_priv
->wm
.max_level
= VLV_WM_LEVEL_PM5
;
4033 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
);
4034 if ((val
& FORCE_DDR_HIGH_FREQ
) == 0)
4035 wm
->level
= VLV_WM_LEVEL_DDR_DVFS
;
4038 mutex_unlock(&dev_priv
->rps
.hw_lock
);
4041 for_each_pipe(dev_priv
, pipe
)
4042 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
4043 pipe_name(pipe
), wm
->pipe
[pipe
].primary
, wm
->pipe
[pipe
].cursor
,
4044 wm
->pipe
[pipe
].sprite
[0], wm
->pipe
[pipe
].sprite
[1]);
4046 DRM_DEBUG_KMS("Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
4047 wm
->sr
.plane
, wm
->sr
.cursor
, wm
->level
, wm
->cxsr
);
4050 void ilk_wm_get_hw_state(struct drm_device
*dev
)
4052 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4053 struct ilk_wm_values
*hw
= &dev_priv
->wm
.hw
;
4054 struct drm_crtc
*crtc
;
4056 for_each_crtc(dev
, crtc
)
4057 ilk_pipe_wm_get_hw_state(crtc
);
4059 hw
->wm_lp
[0] = I915_READ(WM1_LP_ILK
);
4060 hw
->wm_lp
[1] = I915_READ(WM2_LP_ILK
);
4061 hw
->wm_lp
[2] = I915_READ(WM3_LP_ILK
);
4063 hw
->wm_lp_spr
[0] = I915_READ(WM1S_LP_ILK
);
4064 if (INTEL_INFO(dev
)->gen
>= 7) {
4065 hw
->wm_lp_spr
[1] = I915_READ(WM2S_LP_IVB
);
4066 hw
->wm_lp_spr
[2] = I915_READ(WM3S_LP_IVB
);
4069 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
4070 hw
->partitioning
= (I915_READ(WM_MISC
) & WM_MISC_DATA_PARTITION_5_6
) ?
4071 INTEL_DDB_PART_5_6
: INTEL_DDB_PART_1_2
;
4072 else if (IS_IVYBRIDGE(dev
))
4073 hw
->partitioning
= (I915_READ(DISP_ARB_CTL2
) & DISP_DATA_PARTITION_5_6
) ?
4074 INTEL_DDB_PART_5_6
: INTEL_DDB_PART_1_2
;
4077 !(I915_READ(DISP_ARB_CTL
) & DISP_FBC_WM_DIS
);
4081 * intel_update_watermarks - update FIFO watermark values based on current modes
4083 * Calculate watermark values for the various WM regs based on current mode
4084 * and plane configuration.
4086 * There are several cases to deal with here:
4087 * - normal (i.e. non-self-refresh)
4088 * - self-refresh (SR) mode
4089 * - lines are large relative to FIFO size (buffer can hold up to 2)
4090 * - lines are small relative to FIFO size (buffer can hold more than 2
4091 * lines), so need to account for TLB latency
4093 * The normal calculation is:
4094 * watermark = dotclock * bytes per pixel * latency
4095 * where latency is platform & configuration dependent (we assume pessimal
4098 * The SR calculation is:
4099 * watermark = (trunc(latency/line time)+1) * surface width *
4102 * line time = htotal / dotclock
4103 * surface width = hdisplay for normal plane and 64 for cursor
4104 * and latency is assumed to be high, as above.
4106 * The final value programmed to the register should always be rounded up,
4107 * and include an extra 2 entries to account for clock crossings.
4109 * We don't use the sprite, so we can ignore that. And on Crestline we have
4110 * to set the non-SR watermarks to 8.
4112 void intel_update_watermarks(struct drm_crtc
*crtc
)
4114 struct drm_i915_private
*dev_priv
= crtc
->dev
->dev_private
;
4116 if (dev_priv
->display
.update_wm
)
4117 dev_priv
->display
.update_wm(crtc
);
4121 * Lock protecting IPS related data structures
4123 DEFINE_SPINLOCK(mchdev_lock
);
4125 /* Global for IPS driver to get at the current i915 device. Protected by
4127 static struct drm_i915_private
*i915_mch_dev
;
4129 bool ironlake_set_drps(struct drm_device
*dev
, u8 val
)
4131 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4134 assert_spin_locked(&mchdev_lock
);
4136 rgvswctl
= I915_READ16(MEMSWCTL
);
4137 if (rgvswctl
& MEMCTL_CMD_STS
) {
4138 DRM_DEBUG("gpu busy, RCS change rejected\n");
4139 return false; /* still busy with another command */
4142 rgvswctl
= (MEMCTL_CMD_CHFREQ
<< MEMCTL_CMD_SHIFT
) |
4143 (val
<< MEMCTL_FREQ_SHIFT
) | MEMCTL_SFCAVM
;
4144 I915_WRITE16(MEMSWCTL
, rgvswctl
);
4145 POSTING_READ16(MEMSWCTL
);
4147 rgvswctl
|= MEMCTL_CMD_STS
;
4148 I915_WRITE16(MEMSWCTL
, rgvswctl
);
4153 static void ironlake_enable_drps(struct drm_device
*dev
)
4155 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4156 u32 rgvmodectl
= I915_READ(MEMMODECTL
);
4157 u8 fmax
, fmin
, fstart
, vstart
;
4159 spin_lock_irq(&mchdev_lock
);
4161 /* Enable temp reporting */
4162 I915_WRITE16(PMMISC
, I915_READ(PMMISC
) | MCPPCE_EN
);
4163 I915_WRITE16(TSC1
, I915_READ(TSC1
) | TSE
);
4165 /* 100ms RC evaluation intervals */
4166 I915_WRITE(RCUPEI
, 100000);
4167 I915_WRITE(RCDNEI
, 100000);
4169 /* Set max/min thresholds to 90ms and 80ms respectively */
4170 I915_WRITE(RCBMAXAVG
, 90000);
4171 I915_WRITE(RCBMINAVG
, 80000);
4173 I915_WRITE(MEMIHYST
, 1);
4175 /* Set up min, max, and cur for interrupt handling */
4176 fmax
= (rgvmodectl
& MEMMODE_FMAX_MASK
) >> MEMMODE_FMAX_SHIFT
;
4177 fmin
= (rgvmodectl
& MEMMODE_FMIN_MASK
);
4178 fstart
= (rgvmodectl
& MEMMODE_FSTART_MASK
) >>
4179 MEMMODE_FSTART_SHIFT
;
4181 vstart
= (I915_READ(PXVFREQ(fstart
)) & PXVFREQ_PX_MASK
) >>
4184 dev_priv
->ips
.fmax
= fmax
; /* IPS callback will increase this */
4185 dev_priv
->ips
.fstart
= fstart
;
4187 dev_priv
->ips
.max_delay
= fstart
;
4188 dev_priv
->ips
.min_delay
= fmin
;
4189 dev_priv
->ips
.cur_delay
= fstart
;
4191 DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
4192 fmax
, fmin
, fstart
);
4194 I915_WRITE(MEMINTREN
, MEMINT_CX_SUPR_EN
| MEMINT_EVAL_CHG_EN
);
4197 * Interrupts will be enabled in ironlake_irq_postinstall
4200 I915_WRITE(VIDSTART
, vstart
);
4201 POSTING_READ(VIDSTART
);
4203 rgvmodectl
|= MEMMODE_SWMODE_EN
;
4204 I915_WRITE(MEMMODECTL
, rgvmodectl
);
4206 if (wait_for_atomic((I915_READ(MEMSWCTL
) & MEMCTL_CMD_STS
) == 0, 10))
4207 DRM_ERROR("stuck trying to change perf mode\n");
4210 ironlake_set_drps(dev
, fstart
);
4212 dev_priv
->ips
.last_count1
= I915_READ(DMIEC
) +
4213 I915_READ(DDREC
) + I915_READ(CSIEC
);
4214 dev_priv
->ips
.last_time1
= jiffies_to_msecs(jiffies
);
4215 dev_priv
->ips
.last_count2
= I915_READ(GFXEC
);
4216 dev_priv
->ips
.last_time2
= ktime_get_raw_ns();
4218 spin_unlock_irq(&mchdev_lock
);
4221 static void ironlake_disable_drps(struct drm_device
*dev
)
4223 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4226 spin_lock_irq(&mchdev_lock
);
4228 rgvswctl
= I915_READ16(MEMSWCTL
);
4230 /* Ack interrupts, disable EFC interrupt */
4231 I915_WRITE(MEMINTREN
, I915_READ(MEMINTREN
) & ~MEMINT_EVAL_CHG_EN
);
4232 I915_WRITE(MEMINTRSTS
, MEMINT_EVAL_CHG
);
4233 I915_WRITE(DEIER
, I915_READ(DEIER
) & ~DE_PCU_EVENT
);
4234 I915_WRITE(DEIIR
, DE_PCU_EVENT
);
4235 I915_WRITE(DEIMR
, I915_READ(DEIMR
) | DE_PCU_EVENT
);
4237 /* Go back to the starting frequency */
4238 ironlake_set_drps(dev
, dev_priv
->ips
.fstart
);
4240 rgvswctl
|= MEMCTL_CMD_STS
;
4241 I915_WRITE(MEMSWCTL
, rgvswctl
);
4244 spin_unlock_irq(&mchdev_lock
);
4247 /* There's a funny hw issue where the hw returns all 0 when reading from
4248 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
4249 * ourselves, instead of doing a rmw cycle (which might result in us clearing
4250 * all limits and the gpu stuck at whatever frequency it is at atm).
4252 static u32
intel_rps_limits(struct drm_i915_private
*dev_priv
, u8 val
)
4256 /* Only set the down limit when we've reached the lowest level to avoid
4257 * getting more interrupts, otherwise leave this clear. This prevents a
4258 * race in the hw when coming out of rc6: There's a tiny window where
4259 * the hw runs at the minimal clock before selecting the desired
4260 * frequency, if the down threshold expires in that window we will not
4261 * receive a down interrupt. */
4262 if (IS_GEN9(dev_priv
->dev
)) {
4263 limits
= (dev_priv
->rps
.max_freq_softlimit
) << 23;
4264 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
4265 limits
|= (dev_priv
->rps
.min_freq_softlimit
) << 14;
4267 limits
= dev_priv
->rps
.max_freq_softlimit
<< 24;
4268 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
4269 limits
|= dev_priv
->rps
.min_freq_softlimit
<< 16;
4275 static void gen6_set_rps_thresholds(struct drm_i915_private
*dev_priv
, u8 val
)
4278 u32 threshold_up
= 0, threshold_down
= 0; /* in % */
4279 u32 ei_up
= 0, ei_down
= 0;
4281 new_power
= dev_priv
->rps
.power
;
4282 switch (dev_priv
->rps
.power
) {
4284 if (val
> dev_priv
->rps
.efficient_freq
+ 1 && val
> dev_priv
->rps
.cur_freq
)
4285 new_power
= BETWEEN
;
4289 if (val
<= dev_priv
->rps
.efficient_freq
&& val
< dev_priv
->rps
.cur_freq
)
4290 new_power
= LOW_POWER
;
4291 else if (val
>= dev_priv
->rps
.rp0_freq
&& val
> dev_priv
->rps
.cur_freq
)
4292 new_power
= HIGH_POWER
;
4296 if (val
< (dev_priv
->rps
.rp1_freq
+ dev_priv
->rps
.rp0_freq
) >> 1 && val
< dev_priv
->rps
.cur_freq
)
4297 new_power
= BETWEEN
;
4300 /* Max/min bins are special */
4301 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
4302 new_power
= LOW_POWER
;
4303 if (val
>= dev_priv
->rps
.max_freq_softlimit
)
4304 new_power
= HIGH_POWER
;
4305 if (new_power
== dev_priv
->rps
.power
)
4308 /* Note the units here are not exactly 1us, but 1280ns. */
4309 switch (new_power
) {
4311 /* Upclock if more than 95% busy over 16ms */
4315 /* Downclock if less than 85% busy over 32ms */
4317 threshold_down
= 85;
4321 /* Upclock if more than 90% busy over 13ms */
4325 /* Downclock if less than 75% busy over 32ms */
4327 threshold_down
= 75;
4331 /* Upclock if more than 85% busy over 10ms */
4335 /* Downclock if less than 60% busy over 32ms */
4337 threshold_down
= 60;
4341 I915_WRITE(GEN6_RP_UP_EI
,
4342 GT_INTERVAL_FROM_US(dev_priv
, ei_up
));
4343 I915_WRITE(GEN6_RP_UP_THRESHOLD
,
4344 GT_INTERVAL_FROM_US(dev_priv
, (ei_up
* threshold_up
/ 100)));
4346 I915_WRITE(GEN6_RP_DOWN_EI
,
4347 GT_INTERVAL_FROM_US(dev_priv
, ei_down
));
4348 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
,
4349 GT_INTERVAL_FROM_US(dev_priv
, (ei_down
* threshold_down
/ 100)));
4351 I915_WRITE(GEN6_RP_CONTROL
,
4352 GEN6_RP_MEDIA_TURBO
|
4353 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
4354 GEN6_RP_MEDIA_IS_GFX
|
4356 GEN6_RP_UP_BUSY_AVG
|
4357 GEN6_RP_DOWN_IDLE_AVG
);
4359 dev_priv
->rps
.power
= new_power
;
4360 dev_priv
->rps
.up_threshold
= threshold_up
;
4361 dev_priv
->rps
.down_threshold
= threshold_down
;
4362 dev_priv
->rps
.last_adj
= 0;
4365 static u32
gen6_rps_pm_mask(struct drm_i915_private
*dev_priv
, u8 val
)
4369 if (val
> dev_priv
->rps
.min_freq_softlimit
)
4370 mask
|= GEN6_PM_RP_DOWN_EI_EXPIRED
| GEN6_PM_RP_DOWN_THRESHOLD
| GEN6_PM_RP_DOWN_TIMEOUT
;
4371 if (val
< dev_priv
->rps
.max_freq_softlimit
)
4372 mask
|= GEN6_PM_RP_UP_EI_EXPIRED
| GEN6_PM_RP_UP_THRESHOLD
;
4374 mask
&= dev_priv
->pm_rps_events
;
4376 return gen6_sanitize_rps_pm_mask(dev_priv
, ~mask
);
4379 /* gen6_set_rps is called to update the frequency request, but should also be
4380 * called when the range (min_delay and max_delay) is modified so that we can
4381 * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
4382 static void gen6_set_rps(struct drm_device
*dev
, u8 val
)
4384 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4386 /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
4387 if (IS_BXT_REVID(dev
, 0, BXT_REVID_A1
))
4390 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
4391 WARN_ON(val
> dev_priv
->rps
.max_freq
);
4392 WARN_ON(val
< dev_priv
->rps
.min_freq
);
4394 /* min/max delay may still have been modified so be sure to
4395 * write the limits value.
4397 if (val
!= dev_priv
->rps
.cur_freq
) {
4398 gen6_set_rps_thresholds(dev_priv
, val
);
4401 I915_WRITE(GEN6_RPNSWREQ
,
4402 GEN9_FREQUENCY(val
));
4403 else if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
4404 I915_WRITE(GEN6_RPNSWREQ
,
4405 HSW_FREQUENCY(val
));
4407 I915_WRITE(GEN6_RPNSWREQ
,
4408 GEN6_FREQUENCY(val
) |
4410 GEN6_AGGRESSIVE_TURBO
);
4413 /* Make sure we continue to get interrupts
4414 * until we hit the minimum or maximum frequencies.
4416 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS
, intel_rps_limits(dev_priv
, val
));
4417 I915_WRITE(GEN6_PMINTRMSK
, gen6_rps_pm_mask(dev_priv
, val
));
4419 POSTING_READ(GEN6_RPNSWREQ
);
4421 dev_priv
->rps
.cur_freq
= val
;
4422 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv
, val
));
4425 static void valleyview_set_rps(struct drm_device
*dev
, u8 val
)
4427 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4429 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
4430 WARN_ON(val
> dev_priv
->rps
.max_freq
);
4431 WARN_ON(val
< dev_priv
->rps
.min_freq
);
4433 if (WARN_ONCE(IS_CHERRYVIEW(dev
) && (val
& 1),
4434 "Odd GPU freq value\n"))
4437 I915_WRITE(GEN6_PMINTRMSK
, gen6_rps_pm_mask(dev_priv
, val
));
4439 if (val
!= dev_priv
->rps
.cur_freq
) {
4440 vlv_punit_write(dev_priv
, PUNIT_REG_GPU_FREQ_REQ
, val
);
4441 if (!IS_CHERRYVIEW(dev_priv
))
4442 gen6_set_rps_thresholds(dev_priv
, val
);
4445 dev_priv
->rps
.cur_freq
= val
;
4446 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv
, val
));
4449 /* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
4451 * * If Gfx is Idle, then
4452 * 1. Forcewake Media well.
4453 * 2. Request idle freq.
4454 * 3. Release Forcewake of Media well.
4456 static void vlv_set_rps_idle(struct drm_i915_private
*dev_priv
)
4458 u32 val
= dev_priv
->rps
.idle_freq
;
4460 if (dev_priv
->rps
.cur_freq
<= val
)
4463 /* Wake up the media well, as that takes a lot less
4464 * power than the Render well. */
4465 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_MEDIA
);
4466 valleyview_set_rps(dev_priv
->dev
, val
);
4467 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_MEDIA
);
4470 void gen6_rps_busy(struct drm_i915_private
*dev_priv
)
4472 mutex_lock(&dev_priv
->rps
.hw_lock
);
4473 if (dev_priv
->rps
.enabled
) {
4474 if (dev_priv
->pm_rps_events
& (GEN6_PM_RP_DOWN_EI_EXPIRED
| GEN6_PM_RP_UP_EI_EXPIRED
))
4475 gen6_rps_reset_ei(dev_priv
);
4476 I915_WRITE(GEN6_PMINTRMSK
,
4477 gen6_rps_pm_mask(dev_priv
, dev_priv
->rps
.cur_freq
));
4479 mutex_unlock(&dev_priv
->rps
.hw_lock
);
4482 void gen6_rps_idle(struct drm_i915_private
*dev_priv
)
4484 struct drm_device
*dev
= dev_priv
->dev
;
4486 mutex_lock(&dev_priv
->rps
.hw_lock
);
4487 if (dev_priv
->rps
.enabled
) {
4488 if (IS_VALLEYVIEW(dev
) || IS_CHERRYVIEW(dev
))
4489 vlv_set_rps_idle(dev_priv
);
4491 gen6_set_rps(dev_priv
->dev
, dev_priv
->rps
.idle_freq
);
4492 dev_priv
->rps
.last_adj
= 0;
4493 I915_WRITE(GEN6_PMINTRMSK
, 0xffffffff);
4495 mutex_unlock(&dev_priv
->rps
.hw_lock
);
4497 spin_lock(&dev_priv
->rps
.client_lock
);
4498 while (!list_empty(&dev_priv
->rps
.clients
))
4499 list_del_init(dev_priv
->rps
.clients
.next
);
4500 spin_unlock(&dev_priv
->rps
.client_lock
);
4503 void gen6_rps_boost(struct drm_i915_private
*dev_priv
,
4504 struct intel_rps_client
*rps
,
4505 unsigned long submitted
)
4507 /* This is intentionally racy! We peek at the state here, then
4508 * validate inside the RPS worker.
4510 if (!(dev_priv
->mm
.busy
&&
4511 dev_priv
->rps
.enabled
&&
4512 dev_priv
->rps
.cur_freq
< dev_priv
->rps
.max_freq_softlimit
))
4515 /* Force a RPS boost (and don't count it against the client) if
4516 * the GPU is severely congested.
4518 if (rps
&& time_after(jiffies
, submitted
+ DRM_I915_THROTTLE_JIFFIES
))
4521 spin_lock(&dev_priv
->rps
.client_lock
);
4522 if (rps
== NULL
|| list_empty(&rps
->link
)) {
4523 spin_lock_irq(&dev_priv
->irq_lock
);
4524 if (dev_priv
->rps
.interrupts_enabled
) {
4525 dev_priv
->rps
.client_boost
= true;
4526 queue_work(dev_priv
->wq
, &dev_priv
->rps
.work
);
4528 spin_unlock_irq(&dev_priv
->irq_lock
);
4531 list_add(&rps
->link
, &dev_priv
->rps
.clients
);
4534 dev_priv
->rps
.boosts
++;
4536 spin_unlock(&dev_priv
->rps
.client_lock
);
4539 void intel_set_rps(struct drm_device
*dev
, u8 val
)
4541 if (IS_VALLEYVIEW(dev
) || IS_CHERRYVIEW(dev
))
4542 valleyview_set_rps(dev
, val
);
4544 gen6_set_rps(dev
, val
);
4547 static void gen9_disable_rps(struct drm_device
*dev
)
4549 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4551 I915_WRITE(GEN6_RC_CONTROL
, 0);
4552 I915_WRITE(GEN9_PG_ENABLE
, 0);
4555 static void gen6_disable_rps(struct drm_device
*dev
)
4557 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4559 I915_WRITE(GEN6_RC_CONTROL
, 0);
4560 I915_WRITE(GEN6_RPNSWREQ
, 1 << 31);
4563 static void cherryview_disable_rps(struct drm_device
*dev
)
4565 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4567 I915_WRITE(GEN6_RC_CONTROL
, 0);
4570 static void valleyview_disable_rps(struct drm_device
*dev
)
4572 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4574 /* we're doing forcewake before Disabling RC6,
4575 * This what the BIOS expects when going into suspend */
4576 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4578 I915_WRITE(GEN6_RC_CONTROL
, 0);
4580 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4583 static void intel_print_rc6_info(struct drm_device
*dev
, u32 mode
)
4585 if (IS_VALLEYVIEW(dev
) || IS_CHERRYVIEW(dev
)) {
4586 if (mode
& (GEN7_RC_CTL_TO_MODE
| GEN6_RC_CTL_EI_MODE(1)))
4587 mode
= GEN6_RC_CTL_RC6_ENABLE
;
4592 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s RC6p %s RC6pp %s\n",
4593 (mode
& GEN6_RC_CTL_RC6_ENABLE
) ? "on" : "off",
4594 (mode
& GEN6_RC_CTL_RC6p_ENABLE
) ? "on" : "off",
4595 (mode
& GEN6_RC_CTL_RC6pp_ENABLE
) ? "on" : "off");
4598 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s\n",
4599 (mode
& GEN6_RC_CTL_RC6_ENABLE
) ? "on" : "off");
4602 static int sanitize_rc6_option(const struct drm_device
*dev
, int enable_rc6
)
4604 /* No RC6 before Ironlake and code is gone for ilk. */
4605 if (INTEL_INFO(dev
)->gen
< 6)
4608 /* Respect the kernel parameter if it is set */
4609 if (enable_rc6
>= 0) {
4613 mask
= INTEL_RC6_ENABLE
| INTEL_RC6p_ENABLE
|
4616 mask
= INTEL_RC6_ENABLE
;
4618 if ((enable_rc6
& mask
) != enable_rc6
)
4619 DRM_DEBUG_KMS("Adjusting RC6 mask to %d (requested %d, valid %d)\n",
4620 enable_rc6
& mask
, enable_rc6
, mask
);
4622 return enable_rc6
& mask
;
4625 if (IS_IVYBRIDGE(dev
))
4626 return (INTEL_RC6_ENABLE
| INTEL_RC6p_ENABLE
);
4628 return INTEL_RC6_ENABLE
;
4631 int intel_enable_rc6(const struct drm_device
*dev
)
4633 return i915
.enable_rc6
;
4636 static void gen6_init_rps_frequencies(struct drm_device
*dev
)
4638 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4639 uint32_t rp_state_cap
;
4640 u32 ddcc_status
= 0;
4643 /* All of these values are in units of 50MHz */
4644 dev_priv
->rps
.cur_freq
= 0;
4645 /* static values from HW: RP0 > RP1 > RPn (min_freq) */
4646 if (IS_BROXTON(dev
)) {
4647 rp_state_cap
= I915_READ(BXT_RP_STATE_CAP
);
4648 dev_priv
->rps
.rp0_freq
= (rp_state_cap
>> 16) & 0xff;
4649 dev_priv
->rps
.rp1_freq
= (rp_state_cap
>> 8) & 0xff;
4650 dev_priv
->rps
.min_freq
= (rp_state_cap
>> 0) & 0xff;
4652 rp_state_cap
= I915_READ(GEN6_RP_STATE_CAP
);
4653 dev_priv
->rps
.rp0_freq
= (rp_state_cap
>> 0) & 0xff;
4654 dev_priv
->rps
.rp1_freq
= (rp_state_cap
>> 8) & 0xff;
4655 dev_priv
->rps
.min_freq
= (rp_state_cap
>> 16) & 0xff;
4658 /* hw_max = RP0 until we check for overclocking */
4659 dev_priv
->rps
.max_freq
= dev_priv
->rps
.rp0_freq
;
4661 dev_priv
->rps
.efficient_freq
= dev_priv
->rps
.rp1_freq
;
4662 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
) ||
4663 IS_SKYLAKE(dev
) || IS_KABYLAKE(dev
)) {
4664 ret
= sandybridge_pcode_read(dev_priv
,
4665 HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL
,
4668 dev_priv
->rps
.efficient_freq
=
4670 ((ddcc_status
>> 8) & 0xff),
4671 dev_priv
->rps
.min_freq
,
4672 dev_priv
->rps
.max_freq
);
4675 if (IS_SKYLAKE(dev
) || IS_KABYLAKE(dev
)) {
4676 /* Store the frequency values in 16.66 MHZ units, which is
4677 the natural hardware unit for SKL */
4678 dev_priv
->rps
.rp0_freq
*= GEN9_FREQ_SCALER
;
4679 dev_priv
->rps
.rp1_freq
*= GEN9_FREQ_SCALER
;
4680 dev_priv
->rps
.min_freq
*= GEN9_FREQ_SCALER
;
4681 dev_priv
->rps
.max_freq
*= GEN9_FREQ_SCALER
;
4682 dev_priv
->rps
.efficient_freq
*= GEN9_FREQ_SCALER
;
4685 dev_priv
->rps
.idle_freq
= dev_priv
->rps
.min_freq
;
4687 /* Preserve min/max settings in case of re-init */
4688 if (dev_priv
->rps
.max_freq_softlimit
== 0)
4689 dev_priv
->rps
.max_freq_softlimit
= dev_priv
->rps
.max_freq
;
4691 if (dev_priv
->rps
.min_freq_softlimit
== 0) {
4692 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
4693 dev_priv
->rps
.min_freq_softlimit
=
4694 max_t(int, dev_priv
->rps
.efficient_freq
,
4695 intel_freq_opcode(dev_priv
, 450));
4697 dev_priv
->rps
.min_freq_softlimit
=
4698 dev_priv
->rps
.min_freq
;
4702 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
4703 static void gen9_enable_rps(struct drm_device
*dev
)
4705 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4707 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4709 gen6_init_rps_frequencies(dev
);
4711 /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
4712 if (IS_BXT_REVID(dev
, 0, BXT_REVID_A1
)) {
4713 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4717 /* Program defaults and thresholds for RPS*/
4718 I915_WRITE(GEN6_RC_VIDEO_FREQ
,
4719 GEN9_FREQUENCY(dev_priv
->rps
.rp1_freq
));
4721 /* 1 second timeout*/
4722 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
,
4723 GT_INTERVAL_FROM_US(dev_priv
, 1000000));
4725 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 0xa);
4727 /* Leaning on the below call to gen6_set_rps to program/setup the
4728 * Up/Down EI & threshold registers, as well as the RP_CONTROL,
4729 * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
4730 dev_priv
->rps
.power
= HIGH_POWER
; /* force a reset */
4731 gen6_set_rps(dev_priv
->dev
, dev_priv
->rps
.min_freq_softlimit
);
4733 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4736 static void gen9_enable_rc6(struct drm_device
*dev
)
4738 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4739 struct intel_engine_cs
*ring
;
4740 uint32_t rc6_mask
= 0;
4743 /* 1a: Software RC state - RC0 */
4744 I915_WRITE(GEN6_RC_STATE
, 0);
4746 /* 1b: Get forcewake during program sequence. Although the driver
4747 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4748 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4750 /* 2a: Disable RC states. */
4751 I915_WRITE(GEN6_RC_CONTROL
, 0);
4753 /* 2b: Program RC6 thresholds.*/
4755 /* WaRsDoubleRc6WrlWithCoarsePowerGating: Doubling WRL only when CPG is enabled */
4756 if (IS_SKYLAKE(dev
))
4757 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 108 << 16);
4759 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 54 << 16);
4760 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
4761 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
4762 for_each_ring(ring
, dev_priv
, unused
)
4763 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
4765 if (HAS_GUC_UCODE(dev
))
4766 I915_WRITE(GUC_MAX_IDLE_COUNT
, 0xA);
4768 I915_WRITE(GEN6_RC_SLEEP
, 0);
4770 /* 2c: Program Coarse Power Gating Policies. */
4771 I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS
, 25);
4772 I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS
, 25);
4774 /* 3a: Enable RC6 */
4775 if (intel_enable_rc6(dev
) & INTEL_RC6_ENABLE
)
4776 rc6_mask
= GEN6_RC_CTL_RC6_ENABLE
;
4777 DRM_INFO("RC6 %s\n", (rc6_mask
& GEN6_RC_CTL_RC6_ENABLE
) ?
4779 /* WaRsUseTimeoutMode */
4780 if (IS_SKL_REVID(dev
, 0, SKL_REVID_D0
) ||
4781 IS_BXT_REVID(dev
, 0, BXT_REVID_A1
)) {
4782 I915_WRITE(GEN6_RC6_THRESHOLD
, 625); /* 800us */
4783 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
4784 GEN7_RC_CTL_TO_MODE
|
4787 I915_WRITE(GEN6_RC6_THRESHOLD
, 37500); /* 37.5/125ms per EI */
4788 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
4789 GEN6_RC_CTL_EI_MODE(1) |
4794 * 3b: Enable Coarse Power Gating only when RC6 is enabled.
4795 * WaRsDisableCoarsePowerGating:skl,bxt - Render/Media PG need to be disabled with RC6.
4797 if (NEEDS_WaRsDisableCoarsePowerGating(dev
))
4798 I915_WRITE(GEN9_PG_ENABLE
, 0);
4800 I915_WRITE(GEN9_PG_ENABLE
, (rc6_mask
& GEN6_RC_CTL_RC6_ENABLE
) ?
4801 (GEN9_RENDER_PG_ENABLE
| GEN9_MEDIA_PG_ENABLE
) : 0);
4803 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4807 static void gen8_enable_rps(struct drm_device
*dev
)
4809 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4810 struct intel_engine_cs
*ring
;
4811 uint32_t rc6_mask
= 0;
4814 /* 1a: Software RC state - RC0 */
4815 I915_WRITE(GEN6_RC_STATE
, 0);
4817 /* 1c & 1d: Get forcewake during program sequence. Although the driver
4818 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4819 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4821 /* 2a: Disable RC states. */
4822 I915_WRITE(GEN6_RC_CONTROL
, 0);
4824 /* Initialize rps frequencies */
4825 gen6_init_rps_frequencies(dev
);
4827 /* 2b: Program RC6 thresholds.*/
4828 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16);
4829 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
4830 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
4831 for_each_ring(ring
, dev_priv
, unused
)
4832 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
4833 I915_WRITE(GEN6_RC_SLEEP
, 0);
4834 if (IS_BROADWELL(dev
))
4835 I915_WRITE(GEN6_RC6_THRESHOLD
, 625); /* 800us/1.28 for TO */
4837 I915_WRITE(GEN6_RC6_THRESHOLD
, 50000); /* 50/125ms per EI */
4840 if (intel_enable_rc6(dev
) & INTEL_RC6_ENABLE
)
4841 rc6_mask
= GEN6_RC_CTL_RC6_ENABLE
;
4842 intel_print_rc6_info(dev
, rc6_mask
);
4843 if (IS_BROADWELL(dev
))
4844 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
4845 GEN7_RC_CTL_TO_MODE
|
4848 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
4849 GEN6_RC_CTL_EI_MODE(1) |
4852 /* 4 Program defaults and thresholds for RPS*/
4853 I915_WRITE(GEN6_RPNSWREQ
,
4854 HSW_FREQUENCY(dev_priv
->rps
.rp1_freq
));
4855 I915_WRITE(GEN6_RC_VIDEO_FREQ
,
4856 HSW_FREQUENCY(dev_priv
->rps
.rp1_freq
));
4857 /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
4858 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 100000000 / 128); /* 1 second timeout */
4860 /* Docs recommend 900MHz, and 300 MHz respectively */
4861 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS
,
4862 dev_priv
->rps
.max_freq_softlimit
<< 24 |
4863 dev_priv
->rps
.min_freq_softlimit
<< 16);
4865 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 7600000 / 128); /* 76ms busyness per EI, 90% */
4866 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 31300000 / 128); /* 313ms busyness per EI, 70%*/
4867 I915_WRITE(GEN6_RP_UP_EI
, 66000); /* 84.48ms, XXX: random? */
4868 I915_WRITE(GEN6_RP_DOWN_EI
, 350000); /* 448ms, XXX: random? */
4870 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
4873 I915_WRITE(GEN6_RP_CONTROL
,
4874 GEN6_RP_MEDIA_TURBO
|
4875 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
4876 GEN6_RP_MEDIA_IS_GFX
|
4878 GEN6_RP_UP_BUSY_AVG
|
4879 GEN6_RP_DOWN_IDLE_AVG
);
4881 /* 6: Ring frequency + overclocking (our driver does this later */
4883 dev_priv
->rps
.power
= HIGH_POWER
; /* force a reset */
4884 gen6_set_rps(dev_priv
->dev
, dev_priv
->rps
.idle_freq
);
4886 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4889 static void gen6_enable_rps(struct drm_device
*dev
)
4891 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4892 struct intel_engine_cs
*ring
;
4893 u32 rc6vids
, pcu_mbox
= 0, rc6_mask
= 0;
4898 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
4900 /* Here begins a magic sequence of register writes to enable
4901 * auto-downclocking.
4903 * Perhaps there might be some value in exposing these to
4906 I915_WRITE(GEN6_RC_STATE
, 0);
4908 /* Clear the DBG now so we don't confuse earlier errors */
4909 if ((gtfifodbg
= I915_READ(GTFIFODBG
))) {
4910 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg
);
4911 I915_WRITE(GTFIFODBG
, gtfifodbg
);
4914 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4916 /* Initialize rps frequencies */
4917 gen6_init_rps_frequencies(dev
);
4919 /* disable the counters and set deterministic thresholds */
4920 I915_WRITE(GEN6_RC_CONTROL
, 0);
4922 I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT
, 1000 << 16);
4923 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16 | 30);
4924 I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT
, 30);
4925 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000);
4926 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25);
4928 for_each_ring(ring
, dev_priv
, i
)
4929 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
4931 I915_WRITE(GEN6_RC_SLEEP
, 0);
4932 I915_WRITE(GEN6_RC1e_THRESHOLD
, 1000);
4933 if (IS_IVYBRIDGE(dev
))
4934 I915_WRITE(GEN6_RC6_THRESHOLD
, 125000);
4936 I915_WRITE(GEN6_RC6_THRESHOLD
, 50000);
4937 I915_WRITE(GEN6_RC6p_THRESHOLD
, 150000);
4938 I915_WRITE(GEN6_RC6pp_THRESHOLD
, 64000); /* unused */
4940 /* Check if we are enabling RC6 */
4941 rc6_mode
= intel_enable_rc6(dev_priv
->dev
);
4942 if (rc6_mode
& INTEL_RC6_ENABLE
)
4943 rc6_mask
|= GEN6_RC_CTL_RC6_ENABLE
;
4945 /* We don't use those on Haswell */
4946 if (!IS_HASWELL(dev
)) {
4947 if (rc6_mode
& INTEL_RC6p_ENABLE
)
4948 rc6_mask
|= GEN6_RC_CTL_RC6p_ENABLE
;
4950 if (rc6_mode
& INTEL_RC6pp_ENABLE
)
4951 rc6_mask
|= GEN6_RC_CTL_RC6pp_ENABLE
;
4954 intel_print_rc6_info(dev
, rc6_mask
);
4956 I915_WRITE(GEN6_RC_CONTROL
,
4958 GEN6_RC_CTL_EI_MODE(1) |
4959 GEN6_RC_CTL_HW_ENABLE
);
4961 /* Power down if completely idle for over 50ms */
4962 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 50000);
4963 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
4965 ret
= sandybridge_pcode_write(dev_priv
, GEN6_PCODE_WRITE_MIN_FREQ_TABLE
, 0);
4967 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
4969 ret
= sandybridge_pcode_read(dev_priv
, GEN6_READ_OC_PARAMS
, &pcu_mbox
);
4970 if (!ret
&& (pcu_mbox
& (1<<31))) { /* OC supported */
4971 DRM_DEBUG_DRIVER("Overclocking supported. Max: %dMHz, Overclock max: %dMHz\n",
4972 (dev_priv
->rps
.max_freq_softlimit
& 0xff) * 50,
4973 (pcu_mbox
& 0xff) * 50);
4974 dev_priv
->rps
.max_freq
= pcu_mbox
& 0xff;
4977 dev_priv
->rps
.power
= HIGH_POWER
; /* force a reset */
4978 gen6_set_rps(dev_priv
->dev
, dev_priv
->rps
.idle_freq
);
4981 ret
= sandybridge_pcode_read(dev_priv
, GEN6_PCODE_READ_RC6VIDS
, &rc6vids
);
4982 if (IS_GEN6(dev
) && ret
) {
4983 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
4984 } else if (IS_GEN6(dev
) && (GEN6_DECODE_RC6_VID(rc6vids
& 0xff) < 450)) {
4985 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
4986 GEN6_DECODE_RC6_VID(rc6vids
& 0xff), 450);
4987 rc6vids
&= 0xffff00;
4988 rc6vids
|= GEN6_ENCODE_RC6_VID(450);
4989 ret
= sandybridge_pcode_write(dev_priv
, GEN6_PCODE_WRITE_RC6VIDS
, rc6vids
);
4991 DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
4994 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4997 static void __gen6_update_ring_freq(struct drm_device
*dev
)
4999 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5001 unsigned int gpu_freq
;
5002 unsigned int max_ia_freq
, min_ring_freq
;
5003 unsigned int max_gpu_freq
, min_gpu_freq
;
5004 int scaling_factor
= 180;
5005 struct cpufreq_policy
*policy
;
5007 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
5009 policy
= cpufreq_cpu_get(0);
5011 max_ia_freq
= policy
->cpuinfo
.max_freq
;
5012 cpufreq_cpu_put(policy
);
5015 * Default to measured freq if none found, PCU will ensure we
5018 max_ia_freq
= tsc_khz
;
5021 /* Convert from kHz to MHz */
5022 max_ia_freq
/= 1000;
5024 min_ring_freq
= I915_READ(DCLK
) & 0xf;
5025 /* convert DDR frequency from units of 266.6MHz to bandwidth */
5026 min_ring_freq
= mult_frac(min_ring_freq
, 8, 3);
5028 if (IS_SKYLAKE(dev
) || IS_KABYLAKE(dev
)) {
5029 /* Convert GT frequency to 50 HZ units */
5030 min_gpu_freq
= dev_priv
->rps
.min_freq
/ GEN9_FREQ_SCALER
;
5031 max_gpu_freq
= dev_priv
->rps
.max_freq
/ GEN9_FREQ_SCALER
;
5033 min_gpu_freq
= dev_priv
->rps
.min_freq
;
5034 max_gpu_freq
= dev_priv
->rps
.max_freq
;
5038 * For each potential GPU frequency, load a ring frequency we'd like
5039 * to use for memory access. We do this by specifying the IA frequency
5040 * the PCU should use as a reference to determine the ring frequency.
5042 for (gpu_freq
= max_gpu_freq
; gpu_freq
>= min_gpu_freq
; gpu_freq
--) {
5043 int diff
= max_gpu_freq
- gpu_freq
;
5044 unsigned int ia_freq
= 0, ring_freq
= 0;
5046 if (IS_SKYLAKE(dev
) || IS_KABYLAKE(dev
)) {
5048 * ring_freq = 2 * GT. ring_freq is in 100MHz units
5049 * No floor required for ring frequency on SKL.
5051 ring_freq
= gpu_freq
;
5052 } else if (INTEL_INFO(dev
)->gen
>= 8) {
5053 /* max(2 * GT, DDR). NB: GT is 50MHz units */
5054 ring_freq
= max(min_ring_freq
, gpu_freq
);
5055 } else if (IS_HASWELL(dev
)) {
5056 ring_freq
= mult_frac(gpu_freq
, 5, 4);
5057 ring_freq
= max(min_ring_freq
, ring_freq
);
5058 /* leave ia_freq as the default, chosen by cpufreq */
5060 /* On older processors, there is no separate ring
5061 * clock domain, so in order to boost the bandwidth
5062 * of the ring, we need to upclock the CPU (ia_freq).
5064 * For GPU frequencies less than 750MHz,
5065 * just use the lowest ring freq.
5067 if (gpu_freq
< min_freq
)
5070 ia_freq
= max_ia_freq
- ((diff
* scaling_factor
) / 2);
5071 ia_freq
= DIV_ROUND_CLOSEST(ia_freq
, 100);
5074 sandybridge_pcode_write(dev_priv
,
5075 GEN6_PCODE_WRITE_MIN_FREQ_TABLE
,
5076 ia_freq
<< GEN6_PCODE_FREQ_IA_RATIO_SHIFT
|
5077 ring_freq
<< GEN6_PCODE_FREQ_RING_RATIO_SHIFT
|
5082 void gen6_update_ring_freq(struct drm_device
*dev
)
5084 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5086 if (!HAS_CORE_RING_FREQ(dev
))
5089 mutex_lock(&dev_priv
->rps
.hw_lock
);
5090 __gen6_update_ring_freq(dev
);
5091 mutex_unlock(&dev_priv
->rps
.hw_lock
);
5094 static int cherryview_rps_max_freq(struct drm_i915_private
*dev_priv
)
5096 struct drm_device
*dev
= dev_priv
->dev
;
5099 val
= vlv_punit_read(dev_priv
, FB_GFX_FMAX_AT_VMAX_FUSE
);
5101 switch (INTEL_INFO(dev
)->eu_total
) {
5103 /* (2 * 4) config */
5104 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT
);
5107 /* (2 * 6) config */
5108 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT
);
5111 /* (2 * 8) config */
5113 /* Setting (2 * 8) Min RP0 for any other combination */
5114 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT
);
5118 rp0
= (rp0
& FB_GFX_FREQ_FUSE_MASK
);
5123 static int cherryview_rps_rpe_freq(struct drm_i915_private
*dev_priv
)
5127 val
= vlv_punit_read(dev_priv
, PUNIT_GPU_DUTYCYCLE_REG
);
5128 rpe
= (val
>> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT
) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK
;
5133 static int cherryview_rps_guar_freq(struct drm_i915_private
*dev_priv
)
5137 val
= vlv_punit_read(dev_priv
, FB_GFX_FMAX_AT_VMAX_FUSE
);
5138 rp1
= (val
& FB_GFX_FREQ_FUSE_MASK
);
5143 static int valleyview_rps_guar_freq(struct drm_i915_private
*dev_priv
)
5147 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FREQ_FUSE
);
5149 rp1
= (val
& FB_GFX_FGUARANTEED_FREQ_FUSE_MASK
) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT
;
5154 static int valleyview_rps_max_freq(struct drm_i915_private
*dev_priv
)
5158 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FREQ_FUSE
);
5160 rp0
= (val
& FB_GFX_MAX_FREQ_FUSE_MASK
) >> FB_GFX_MAX_FREQ_FUSE_SHIFT
;
5162 rp0
= min_t(u32
, rp0
, 0xea);
5167 static int valleyview_rps_rpe_freq(struct drm_i915_private
*dev_priv
)
5171 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FMAX_FUSE_LO
);
5172 rpe
= (val
& FB_FMAX_VMIN_FREQ_LO_MASK
) >> FB_FMAX_VMIN_FREQ_LO_SHIFT
;
5173 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FMAX_FUSE_HI
);
5174 rpe
|= (val
& FB_FMAX_VMIN_FREQ_HI_MASK
) << 5;
5179 static int valleyview_rps_min_freq(struct drm_i915_private
*dev_priv
)
5183 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_LFM
) & 0xff;
5185 * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
5186 * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
5187 * a BYT-M B0 the above register contains 0xbf. Moreover when setting
5188 * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
5189 * to make sure it matches what Punit accepts.
5191 return max_t(u32
, val
, 0xc0);
5194 /* Check that the pctx buffer wasn't move under us. */
5195 static void valleyview_check_pctx(struct drm_i915_private
*dev_priv
)
5197 unsigned long pctx_addr
= I915_READ(VLV_PCBR
) & ~4095;
5199 WARN_ON(pctx_addr
!= dev_priv
->mm
.stolen_base
+
5200 dev_priv
->vlv_pctx
->stolen
->start
);
5204 /* Check that the pcbr address is not empty. */
5205 static void cherryview_check_pctx(struct drm_i915_private
*dev_priv
)
5207 unsigned long pctx_addr
= I915_READ(VLV_PCBR
) & ~4095;
5209 WARN_ON((pctx_addr
>> VLV_PCBR_ADDR_SHIFT
) == 0);
5212 static void cherryview_setup_pctx(struct drm_device
*dev
)
5214 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5215 unsigned long pctx_paddr
, paddr
;
5216 struct i915_gtt
*gtt
= &dev_priv
->gtt
;
5218 int pctx_size
= 32*1024;
5220 WARN_ON(!mutex_is_locked(&dev
->struct_mutex
));
5222 pcbr
= I915_READ(VLV_PCBR
);
5223 if ((pcbr
>> VLV_PCBR_ADDR_SHIFT
) == 0) {
5224 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5225 paddr
= (dev_priv
->mm
.stolen_base
+
5226 (gtt
->stolen_size
- pctx_size
));
5228 pctx_paddr
= (paddr
& (~4095));
5229 I915_WRITE(VLV_PCBR
, pctx_paddr
);
5232 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR
));
5235 static void valleyview_setup_pctx(struct drm_device
*dev
)
5237 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5238 struct drm_i915_gem_object
*pctx
;
5239 unsigned long pctx_paddr
;
5241 int pctx_size
= 24*1024;
5243 WARN_ON(!mutex_is_locked(&dev
->struct_mutex
));
5245 pcbr
= I915_READ(VLV_PCBR
);
5247 /* BIOS set it up already, grab the pre-alloc'd space */
5250 pcbr_offset
= (pcbr
& (~4095)) - dev_priv
->mm
.stolen_base
;
5251 pctx
= i915_gem_object_create_stolen_for_preallocated(dev_priv
->dev
,
5253 I915_GTT_OFFSET_NONE
,
5258 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5261 * From the Gunit register HAS:
5262 * The Gfx driver is expected to program this register and ensure
5263 * proper allocation within Gfx stolen memory. For example, this
5264 * register should be programmed such than the PCBR range does not
5265 * overlap with other ranges, such as the frame buffer, protected
5266 * memory, or any other relevant ranges.
5268 pctx
= i915_gem_object_create_stolen(dev
, pctx_size
);
5270 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
5274 pctx_paddr
= dev_priv
->mm
.stolen_base
+ pctx
->stolen
->start
;
5275 I915_WRITE(VLV_PCBR
, pctx_paddr
);
5278 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR
));
5279 dev_priv
->vlv_pctx
= pctx
;
5282 static void valleyview_cleanup_pctx(struct drm_device
*dev
)
5284 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5286 if (WARN_ON(!dev_priv
->vlv_pctx
))
5289 drm_gem_object_unreference(&dev_priv
->vlv_pctx
->base
);
5290 dev_priv
->vlv_pctx
= NULL
;
5293 static void valleyview_init_gt_powersave(struct drm_device
*dev
)
5295 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5298 valleyview_setup_pctx(dev
);
5300 mutex_lock(&dev_priv
->rps
.hw_lock
);
5302 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
5303 switch ((val
>> 6) & 3) {
5306 dev_priv
->mem_freq
= 800;
5309 dev_priv
->mem_freq
= 1066;
5312 dev_priv
->mem_freq
= 1333;
5315 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv
->mem_freq
);
5317 dev_priv
->rps
.max_freq
= valleyview_rps_max_freq(dev_priv
);
5318 dev_priv
->rps
.rp0_freq
= dev_priv
->rps
.max_freq
;
5319 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5320 intel_gpu_freq(dev_priv
, dev_priv
->rps
.max_freq
),
5321 dev_priv
->rps
.max_freq
);
5323 dev_priv
->rps
.efficient_freq
= valleyview_rps_rpe_freq(dev_priv
);
5324 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5325 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
5326 dev_priv
->rps
.efficient_freq
);
5328 dev_priv
->rps
.rp1_freq
= valleyview_rps_guar_freq(dev_priv
);
5329 DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
5330 intel_gpu_freq(dev_priv
, dev_priv
->rps
.rp1_freq
),
5331 dev_priv
->rps
.rp1_freq
);
5333 dev_priv
->rps
.min_freq
= valleyview_rps_min_freq(dev_priv
);
5334 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5335 intel_gpu_freq(dev_priv
, dev_priv
->rps
.min_freq
),
5336 dev_priv
->rps
.min_freq
);
5338 dev_priv
->rps
.idle_freq
= dev_priv
->rps
.min_freq
;
5340 /* Preserve min/max settings in case of re-init */
5341 if (dev_priv
->rps
.max_freq_softlimit
== 0)
5342 dev_priv
->rps
.max_freq_softlimit
= dev_priv
->rps
.max_freq
;
5344 if (dev_priv
->rps
.min_freq_softlimit
== 0)
5345 dev_priv
->rps
.min_freq_softlimit
= dev_priv
->rps
.min_freq
;
5347 mutex_unlock(&dev_priv
->rps
.hw_lock
);
5350 static void cherryview_init_gt_powersave(struct drm_device
*dev
)
5352 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5355 cherryview_setup_pctx(dev
);
5357 mutex_lock(&dev_priv
->rps
.hw_lock
);
5359 mutex_lock(&dev_priv
->sb_lock
);
5360 val
= vlv_cck_read(dev_priv
, CCK_FUSE_REG
);
5361 mutex_unlock(&dev_priv
->sb_lock
);
5363 switch ((val
>> 2) & 0x7) {
5365 dev_priv
->mem_freq
= 2000;
5368 dev_priv
->mem_freq
= 1600;
5371 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv
->mem_freq
);
5373 dev_priv
->rps
.max_freq
= cherryview_rps_max_freq(dev_priv
);
5374 dev_priv
->rps
.rp0_freq
= dev_priv
->rps
.max_freq
;
5375 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5376 intel_gpu_freq(dev_priv
, dev_priv
->rps
.max_freq
),
5377 dev_priv
->rps
.max_freq
);
5379 dev_priv
->rps
.efficient_freq
= cherryview_rps_rpe_freq(dev_priv
);
5380 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5381 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
5382 dev_priv
->rps
.efficient_freq
);
5384 dev_priv
->rps
.rp1_freq
= cherryview_rps_guar_freq(dev_priv
);
5385 DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
5386 intel_gpu_freq(dev_priv
, dev_priv
->rps
.rp1_freq
),
5387 dev_priv
->rps
.rp1_freq
);
5389 /* PUnit validated range is only [RPe, RP0] */
5390 dev_priv
->rps
.min_freq
= dev_priv
->rps
.efficient_freq
;
5391 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5392 intel_gpu_freq(dev_priv
, dev_priv
->rps
.min_freq
),
5393 dev_priv
->rps
.min_freq
);
5395 WARN_ONCE((dev_priv
->rps
.max_freq
|
5396 dev_priv
->rps
.efficient_freq
|
5397 dev_priv
->rps
.rp1_freq
|
5398 dev_priv
->rps
.min_freq
) & 1,
5399 "Odd GPU freq values\n");
5401 dev_priv
->rps
.idle_freq
= dev_priv
->rps
.min_freq
;
5403 /* Preserve min/max settings in case of re-init */
5404 if (dev_priv
->rps
.max_freq_softlimit
== 0)
5405 dev_priv
->rps
.max_freq_softlimit
= dev_priv
->rps
.max_freq
;
5407 if (dev_priv
->rps
.min_freq_softlimit
== 0)
5408 dev_priv
->rps
.min_freq_softlimit
= dev_priv
->rps
.min_freq
;
5410 mutex_unlock(&dev_priv
->rps
.hw_lock
);
5413 static void valleyview_cleanup_gt_powersave(struct drm_device
*dev
)
5415 valleyview_cleanup_pctx(dev
);
5418 static void cherryview_enable_rps(struct drm_device
*dev
)
5420 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5421 struct intel_engine_cs
*ring
;
5422 u32 gtfifodbg
, val
, rc6_mode
= 0, pcbr
;
5425 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
5427 gtfifodbg
= I915_READ(GTFIFODBG
);
5429 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5431 I915_WRITE(GTFIFODBG
, gtfifodbg
);
5434 cherryview_check_pctx(dev_priv
);
5436 /* 1a & 1b: Get forcewake during program sequence. Although the driver
5437 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5438 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
5440 /* Disable RC states. */
5441 I915_WRITE(GEN6_RC_CONTROL
, 0);
5443 /* 2a: Program RC6 thresholds.*/
5444 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16);
5445 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
5446 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
5448 for_each_ring(ring
, dev_priv
, i
)
5449 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
5450 I915_WRITE(GEN6_RC_SLEEP
, 0);
5452 /* TO threshold set to 500 us ( 0x186 * 1.28 us) */
5453 I915_WRITE(GEN6_RC6_THRESHOLD
, 0x186);
5455 /* allows RC6 residency counter to work */
5456 I915_WRITE(VLV_COUNTER_CONTROL
,
5457 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH
|
5458 VLV_MEDIA_RC6_COUNT_EN
|
5459 VLV_RENDER_RC6_COUNT_EN
));
5461 /* For now we assume BIOS is allocating and populating the PCBR */
5462 pcbr
= I915_READ(VLV_PCBR
);
5465 if ((intel_enable_rc6(dev
) & INTEL_RC6_ENABLE
) &&
5466 (pcbr
>> VLV_PCBR_ADDR_SHIFT
))
5467 rc6_mode
= GEN7_RC_CTL_TO_MODE
;
5469 I915_WRITE(GEN6_RC_CONTROL
, rc6_mode
);
5471 /* 4 Program defaults and thresholds for RPS*/
5472 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 1000000);
5473 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 59400);
5474 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 245000);
5475 I915_WRITE(GEN6_RP_UP_EI
, 66000);
5476 I915_WRITE(GEN6_RP_DOWN_EI
, 350000);
5478 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
5481 I915_WRITE(GEN6_RP_CONTROL
,
5482 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
5483 GEN6_RP_MEDIA_IS_GFX
|
5485 GEN6_RP_UP_BUSY_AVG
|
5486 GEN6_RP_DOWN_IDLE_AVG
);
5488 /* Setting Fixed Bias */
5489 val
= VLV_OVERRIDE_EN
|
5491 CHV_BIAS_CPU_50_SOC_50
;
5492 vlv_punit_write(dev_priv
, VLV_TURBO_SOC_OVERRIDE
, val
);
5494 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
5496 /* RPS code assumes GPLL is used */
5497 WARN_ONCE((val
& GPLLENABLE
) == 0, "GPLL not enabled\n");
5499 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val
& GPLLENABLE
));
5500 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val
);
5502 dev_priv
->rps
.cur_freq
= (val
>> 8) & 0xff;
5503 DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
5504 intel_gpu_freq(dev_priv
, dev_priv
->rps
.cur_freq
),
5505 dev_priv
->rps
.cur_freq
);
5507 DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
5508 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
5509 dev_priv
->rps
.efficient_freq
);
5511 valleyview_set_rps(dev_priv
->dev
, dev_priv
->rps
.efficient_freq
);
5513 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
5516 static void valleyview_enable_rps(struct drm_device
*dev
)
5518 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5519 struct intel_engine_cs
*ring
;
5520 u32 gtfifodbg
, val
, rc6_mode
= 0;
5523 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
5525 valleyview_check_pctx(dev_priv
);
5527 if ((gtfifodbg
= I915_READ(GTFIFODBG
))) {
5528 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5530 I915_WRITE(GTFIFODBG
, gtfifodbg
);
5533 /* If VLV, Forcewake all wells, else re-direct to regular path */
5534 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
5536 /* Disable RC states. */
5537 I915_WRITE(GEN6_RC_CONTROL
, 0);
5539 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 1000000);
5540 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 59400);
5541 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 245000);
5542 I915_WRITE(GEN6_RP_UP_EI
, 66000);
5543 I915_WRITE(GEN6_RP_DOWN_EI
, 350000);
5545 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
5547 I915_WRITE(GEN6_RP_CONTROL
,
5548 GEN6_RP_MEDIA_TURBO
|
5549 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
5550 GEN6_RP_MEDIA_IS_GFX
|
5552 GEN6_RP_UP_BUSY_AVG
|
5553 GEN6_RP_DOWN_IDLE_CONT
);
5555 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 0x00280000);
5556 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000);
5557 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25);
5559 for_each_ring(ring
, dev_priv
, i
)
5560 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
5562 I915_WRITE(GEN6_RC6_THRESHOLD
, 0x557);
5564 /* allows RC6 residency counter to work */
5565 I915_WRITE(VLV_COUNTER_CONTROL
,
5566 _MASKED_BIT_ENABLE(VLV_MEDIA_RC0_COUNT_EN
|
5567 VLV_RENDER_RC0_COUNT_EN
|
5568 VLV_MEDIA_RC6_COUNT_EN
|
5569 VLV_RENDER_RC6_COUNT_EN
));
5571 if (intel_enable_rc6(dev
) & INTEL_RC6_ENABLE
)
5572 rc6_mode
= GEN7_RC_CTL_TO_MODE
| VLV_RC_CTL_CTX_RST_PARALLEL
;
5574 intel_print_rc6_info(dev
, rc6_mode
);
5576 I915_WRITE(GEN6_RC_CONTROL
, rc6_mode
);
5578 /* Setting Fixed Bias */
5579 val
= VLV_OVERRIDE_EN
|
5581 VLV_BIAS_CPU_125_SOC_875
;
5582 vlv_punit_write(dev_priv
, VLV_TURBO_SOC_OVERRIDE
, val
);
5584 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
5586 /* RPS code assumes GPLL is used */
5587 WARN_ONCE((val
& GPLLENABLE
) == 0, "GPLL not enabled\n");
5589 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val
& GPLLENABLE
));
5590 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val
);
5592 dev_priv
->rps
.cur_freq
= (val
>> 8) & 0xff;
5593 DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
5594 intel_gpu_freq(dev_priv
, dev_priv
->rps
.cur_freq
),
5595 dev_priv
->rps
.cur_freq
);
5597 DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
5598 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
5599 dev_priv
->rps
.efficient_freq
);
5601 valleyview_set_rps(dev_priv
->dev
, dev_priv
->rps
.efficient_freq
);
5603 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
5606 static unsigned long intel_pxfreq(u32 vidfreq
)
5609 int div
= (vidfreq
& 0x3f0000) >> 16;
5610 int post
= (vidfreq
& 0x3000) >> 12;
5611 int pre
= (vidfreq
& 0x7);
5616 freq
= ((div
* 133333) / ((1<<post
) * pre
));
5621 static const struct cparams
{
5627 { 1, 1333, 301, 28664 },
5628 { 1, 1066, 294, 24460 },
5629 { 1, 800, 294, 25192 },
5630 { 0, 1333, 276, 27605 },
5631 { 0, 1066, 276, 27605 },
5632 { 0, 800, 231, 23784 },
5635 static unsigned long __i915_chipset_val(struct drm_i915_private
*dev_priv
)
5637 u64 total_count
, diff
, ret
;
5638 u32 count1
, count2
, count3
, m
= 0, c
= 0;
5639 unsigned long now
= jiffies_to_msecs(jiffies
), diff1
;
5642 assert_spin_locked(&mchdev_lock
);
5644 diff1
= now
- dev_priv
->ips
.last_time1
;
5646 /* Prevent division-by-zero if we are asking too fast.
5647 * Also, we don't get interesting results if we are polling
5648 * faster than once in 10ms, so just return the saved value
5652 return dev_priv
->ips
.chipset_power
;
5654 count1
= I915_READ(DMIEC
);
5655 count2
= I915_READ(DDREC
);
5656 count3
= I915_READ(CSIEC
);
5658 total_count
= count1
+ count2
+ count3
;
5660 /* FIXME: handle per-counter overflow */
5661 if (total_count
< dev_priv
->ips
.last_count1
) {
5662 diff
= ~0UL - dev_priv
->ips
.last_count1
;
5663 diff
+= total_count
;
5665 diff
= total_count
- dev_priv
->ips
.last_count1
;
5668 for (i
= 0; i
< ARRAY_SIZE(cparams
); i
++) {
5669 if (cparams
[i
].i
== dev_priv
->ips
.c_m
&&
5670 cparams
[i
].t
== dev_priv
->ips
.r_t
) {
5677 diff
= div_u64(diff
, diff1
);
5678 ret
= ((m
* diff
) + c
);
5679 ret
= div_u64(ret
, 10);
5681 dev_priv
->ips
.last_count1
= total_count
;
5682 dev_priv
->ips
.last_time1
= now
;
5684 dev_priv
->ips
.chipset_power
= ret
;
5689 unsigned long i915_chipset_val(struct drm_i915_private
*dev_priv
)
5691 struct drm_device
*dev
= dev_priv
->dev
;
5694 if (INTEL_INFO(dev
)->gen
!= 5)
5697 spin_lock_irq(&mchdev_lock
);
5699 val
= __i915_chipset_val(dev_priv
);
5701 spin_unlock_irq(&mchdev_lock
);
5706 unsigned long i915_mch_val(struct drm_i915_private
*dev_priv
)
5708 unsigned long m
, x
, b
;
5711 tsfs
= I915_READ(TSFS
);
5713 m
= ((tsfs
& TSFS_SLOPE_MASK
) >> TSFS_SLOPE_SHIFT
);
5714 x
= I915_READ8(TR1
);
5716 b
= tsfs
& TSFS_INTR_MASK
;
5718 return ((m
* x
) / 127) - b
;
5721 static int _pxvid_to_vd(u8 pxvid
)
5726 if (pxvid
>= 8 && pxvid
< 31)
5729 return (pxvid
+ 2) * 125;
5732 static u32
pvid_to_extvid(struct drm_i915_private
*dev_priv
, u8 pxvid
)
5734 struct drm_device
*dev
= dev_priv
->dev
;
5735 const int vd
= _pxvid_to_vd(pxvid
);
5736 const int vm
= vd
- 1125;
5738 if (INTEL_INFO(dev
)->is_mobile
)
5739 return vm
> 0 ? vm
: 0;
5744 static void __i915_update_gfx_val(struct drm_i915_private
*dev_priv
)
5746 u64 now
, diff
, diffms
;
5749 assert_spin_locked(&mchdev_lock
);
5751 now
= ktime_get_raw_ns();
5752 diffms
= now
- dev_priv
->ips
.last_time2
;
5753 do_div(diffms
, NSEC_PER_MSEC
);
5755 /* Don't divide by 0 */
5759 count
= I915_READ(GFXEC
);
5761 if (count
< dev_priv
->ips
.last_count2
) {
5762 diff
= ~0UL - dev_priv
->ips
.last_count2
;
5765 diff
= count
- dev_priv
->ips
.last_count2
;
5768 dev_priv
->ips
.last_count2
= count
;
5769 dev_priv
->ips
.last_time2
= now
;
5771 /* More magic constants... */
5773 diff
= div_u64(diff
, diffms
* 10);
5774 dev_priv
->ips
.gfx_power
= diff
;
5777 void i915_update_gfx_val(struct drm_i915_private
*dev_priv
)
5779 struct drm_device
*dev
= dev_priv
->dev
;
5781 if (INTEL_INFO(dev
)->gen
!= 5)
5784 spin_lock_irq(&mchdev_lock
);
5786 __i915_update_gfx_val(dev_priv
);
5788 spin_unlock_irq(&mchdev_lock
);
5791 static unsigned long __i915_gfx_val(struct drm_i915_private
*dev_priv
)
5793 unsigned long t
, corr
, state1
, corr2
, state2
;
5796 assert_spin_locked(&mchdev_lock
);
5798 pxvid
= I915_READ(PXVFREQ(dev_priv
->rps
.cur_freq
));
5799 pxvid
= (pxvid
>> 24) & 0x7f;
5800 ext_v
= pvid_to_extvid(dev_priv
, pxvid
);
5804 t
= i915_mch_val(dev_priv
);
5806 /* Revel in the empirically derived constants */
5808 /* Correction factor in 1/100000 units */
5810 corr
= ((t
* 2349) + 135940);
5812 corr
= ((t
* 964) + 29317);
5814 corr
= ((t
* 301) + 1004);
5816 corr
= corr
* ((150142 * state1
) / 10000 - 78642);
5818 corr2
= (corr
* dev_priv
->ips
.corr
);
5820 state2
= (corr2
* state1
) / 10000;
5821 state2
/= 100; /* convert to mW */
5823 __i915_update_gfx_val(dev_priv
);
5825 return dev_priv
->ips
.gfx_power
+ state2
;
5828 unsigned long i915_gfx_val(struct drm_i915_private
*dev_priv
)
5830 struct drm_device
*dev
= dev_priv
->dev
;
5833 if (INTEL_INFO(dev
)->gen
!= 5)
5836 spin_lock_irq(&mchdev_lock
);
5838 val
= __i915_gfx_val(dev_priv
);
5840 spin_unlock_irq(&mchdev_lock
);
5846 * i915_read_mch_val - return value for IPS use
5848 * Calculate and return a value for the IPS driver to use when deciding whether
5849 * we have thermal and power headroom to increase CPU or GPU power budget.
5851 unsigned long i915_read_mch_val(void)
5853 struct drm_i915_private
*dev_priv
;
5854 unsigned long chipset_val
, graphics_val
, ret
= 0;
5856 spin_lock_irq(&mchdev_lock
);
5859 dev_priv
= i915_mch_dev
;
5861 chipset_val
= __i915_chipset_val(dev_priv
);
5862 graphics_val
= __i915_gfx_val(dev_priv
);
5864 ret
= chipset_val
+ graphics_val
;
5867 spin_unlock_irq(&mchdev_lock
);
5871 EXPORT_SYMBOL_GPL(i915_read_mch_val
);
5874 * i915_gpu_raise - raise GPU frequency limit
5876 * Raise the limit; IPS indicates we have thermal headroom.
5878 bool i915_gpu_raise(void)
5880 struct drm_i915_private
*dev_priv
;
5883 spin_lock_irq(&mchdev_lock
);
5884 if (!i915_mch_dev
) {
5888 dev_priv
= i915_mch_dev
;
5890 if (dev_priv
->ips
.max_delay
> dev_priv
->ips
.fmax
)
5891 dev_priv
->ips
.max_delay
--;
5894 spin_unlock_irq(&mchdev_lock
);
5898 EXPORT_SYMBOL_GPL(i915_gpu_raise
);
5901 * i915_gpu_lower - lower GPU frequency limit
5903 * IPS indicates we're close to a thermal limit, so throttle back the GPU
5904 * frequency maximum.
5906 bool i915_gpu_lower(void)
5908 struct drm_i915_private
*dev_priv
;
5911 spin_lock_irq(&mchdev_lock
);
5912 if (!i915_mch_dev
) {
5916 dev_priv
= i915_mch_dev
;
5918 if (dev_priv
->ips
.max_delay
< dev_priv
->ips
.min_delay
)
5919 dev_priv
->ips
.max_delay
++;
5922 spin_unlock_irq(&mchdev_lock
);
5926 EXPORT_SYMBOL_GPL(i915_gpu_lower
);
5929 * i915_gpu_busy - indicate GPU business to IPS
5931 * Tell the IPS driver whether or not the GPU is busy.
5933 bool i915_gpu_busy(void)
5935 struct drm_i915_private
*dev_priv
;
5936 struct intel_engine_cs
*ring
;
5940 spin_lock_irq(&mchdev_lock
);
5943 dev_priv
= i915_mch_dev
;
5945 for_each_ring(ring
, dev_priv
, i
)
5946 ret
|= !list_empty(&ring
->request_list
);
5949 spin_unlock_irq(&mchdev_lock
);
5953 EXPORT_SYMBOL_GPL(i915_gpu_busy
);
5956 * i915_gpu_turbo_disable - disable graphics turbo
5958 * Disable graphics turbo by resetting the max frequency and setting the
5959 * current frequency to the default.
5961 bool i915_gpu_turbo_disable(void)
5963 struct drm_i915_private
*dev_priv
;
5966 spin_lock_irq(&mchdev_lock
);
5967 if (!i915_mch_dev
) {
5971 dev_priv
= i915_mch_dev
;
5973 dev_priv
->ips
.max_delay
= dev_priv
->ips
.fstart
;
5975 if (!ironlake_set_drps(dev_priv
->dev
, dev_priv
->ips
.fstart
))
5979 spin_unlock_irq(&mchdev_lock
);
5983 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable
);
5986 * Tells the intel_ips driver that the i915 driver is now loaded, if
5987 * IPS got loaded first.
5989 * This awkward dance is so that neither module has to depend on the
5990 * other in order for IPS to do the appropriate communication of
5991 * GPU turbo limits to i915.
5994 ips_ping_for_i915_load(void)
5998 link
= symbol_get(ips_link_to_i915_driver
);
6001 symbol_put(ips_link_to_i915_driver
);
6005 void intel_gpu_ips_init(struct drm_i915_private
*dev_priv
)
6007 /* We only register the i915 ips part with intel-ips once everything is
6008 * set up, to avoid intel-ips sneaking in and reading bogus values. */
6009 spin_lock_irq(&mchdev_lock
);
6010 i915_mch_dev
= dev_priv
;
6011 spin_unlock_irq(&mchdev_lock
);
6013 ips_ping_for_i915_load();
6016 void intel_gpu_ips_teardown(void)
6018 spin_lock_irq(&mchdev_lock
);
6019 i915_mch_dev
= NULL
;
6020 spin_unlock_irq(&mchdev_lock
);
6023 static void intel_init_emon(struct drm_device
*dev
)
6025 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6030 /* Disable to program */
6034 /* Program energy weights for various events */
6035 I915_WRITE(SDEW
, 0x15040d00);
6036 I915_WRITE(CSIEW0
, 0x007f0000);
6037 I915_WRITE(CSIEW1
, 0x1e220004);
6038 I915_WRITE(CSIEW2
, 0x04000004);
6040 for (i
= 0; i
< 5; i
++)
6041 I915_WRITE(PEW(i
), 0);
6042 for (i
= 0; i
< 3; i
++)
6043 I915_WRITE(DEW(i
), 0);
6045 /* Program P-state weights to account for frequency power adjustment */
6046 for (i
= 0; i
< 16; i
++) {
6047 u32 pxvidfreq
= I915_READ(PXVFREQ(i
));
6048 unsigned long freq
= intel_pxfreq(pxvidfreq
);
6049 unsigned long vid
= (pxvidfreq
& PXVFREQ_PX_MASK
) >>
6054 val
*= (freq
/ 1000);
6056 val
/= (127*127*900);
6058 DRM_ERROR("bad pxval: %ld\n", val
);
6061 /* Render standby states get 0 weight */
6065 for (i
= 0; i
< 4; i
++) {
6066 u32 val
= (pxw
[i
*4] << 24) | (pxw
[(i
*4)+1] << 16) |
6067 (pxw
[(i
*4)+2] << 8) | (pxw
[(i
*4)+3]);
6068 I915_WRITE(PXW(i
), val
);
6071 /* Adjust magic regs to magic values (more experimental results) */
6072 I915_WRITE(OGW0
, 0);
6073 I915_WRITE(OGW1
, 0);
6074 I915_WRITE(EG0
, 0x00007f00);
6075 I915_WRITE(EG1
, 0x0000000e);
6076 I915_WRITE(EG2
, 0x000e0000);
6077 I915_WRITE(EG3
, 0x68000300);
6078 I915_WRITE(EG4
, 0x42000000);
6079 I915_WRITE(EG5
, 0x00140031);
6083 for (i
= 0; i
< 8; i
++)
6084 I915_WRITE(PXWL(i
), 0);
6086 /* Enable PMON + select events */
6087 I915_WRITE(ECR
, 0x80000019);
6089 lcfuse
= I915_READ(LCFUSE02
);
6091 dev_priv
->ips
.corr
= (lcfuse
& LCFUSE_HIV_MASK
);
6094 void intel_init_gt_powersave(struct drm_device
*dev
)
6096 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6098 i915
.enable_rc6
= sanitize_rc6_option(dev
, i915
.enable_rc6
);
6100 * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
6103 if (!i915
.enable_rc6
) {
6104 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
6105 intel_runtime_pm_get(dev_priv
);
6108 if (IS_CHERRYVIEW(dev
))
6109 cherryview_init_gt_powersave(dev
);
6110 else if (IS_VALLEYVIEW(dev
))
6111 valleyview_init_gt_powersave(dev
);
6114 void intel_cleanup_gt_powersave(struct drm_device
*dev
)
6116 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6118 if (IS_CHERRYVIEW(dev
))
6120 else if (IS_VALLEYVIEW(dev
))
6121 valleyview_cleanup_gt_powersave(dev
);
6123 if (!i915
.enable_rc6
)
6124 intel_runtime_pm_put(dev_priv
);
6127 static void gen6_suspend_rps(struct drm_device
*dev
)
6129 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6131 flush_delayed_work(&dev_priv
->rps
.delayed_resume_work
);
6133 gen6_disable_rps_interrupts(dev
);
6137 * intel_suspend_gt_powersave - suspend PM work and helper threads
6140 * We don't want to disable RC6 or other features here, we just want
6141 * to make sure any work we've queued has finished and won't bother
6142 * us while we're suspended.
6144 void intel_suspend_gt_powersave(struct drm_device
*dev
)
6146 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6148 if (INTEL_INFO(dev
)->gen
< 6)
6151 gen6_suspend_rps(dev
);
6153 /* Force GPU to min freq during suspend */
6154 gen6_rps_idle(dev_priv
);
6157 void intel_disable_gt_powersave(struct drm_device
*dev
)
6159 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6161 if (IS_IRONLAKE_M(dev
)) {
6162 ironlake_disable_drps(dev
);
6163 } else if (INTEL_INFO(dev
)->gen
>= 6) {
6164 intel_suspend_gt_powersave(dev
);
6166 mutex_lock(&dev_priv
->rps
.hw_lock
);
6167 if (INTEL_INFO(dev
)->gen
>= 9)
6168 gen9_disable_rps(dev
);
6169 else if (IS_CHERRYVIEW(dev
))
6170 cherryview_disable_rps(dev
);
6171 else if (IS_VALLEYVIEW(dev
))
6172 valleyview_disable_rps(dev
);
6174 gen6_disable_rps(dev
);
6176 dev_priv
->rps
.enabled
= false;
6177 mutex_unlock(&dev_priv
->rps
.hw_lock
);
6181 static void intel_gen6_powersave_work(struct work_struct
*work
)
6183 struct drm_i915_private
*dev_priv
=
6184 container_of(work
, struct drm_i915_private
,
6185 rps
.delayed_resume_work
.work
);
6186 struct drm_device
*dev
= dev_priv
->dev
;
6188 mutex_lock(&dev_priv
->rps
.hw_lock
);
6190 gen6_reset_rps_interrupts(dev
);
6192 if (IS_CHERRYVIEW(dev
)) {
6193 cherryview_enable_rps(dev
);
6194 } else if (IS_VALLEYVIEW(dev
)) {
6195 valleyview_enable_rps(dev
);
6196 } else if (INTEL_INFO(dev
)->gen
>= 9) {
6197 gen9_enable_rc6(dev
);
6198 gen9_enable_rps(dev
);
6199 if (IS_SKYLAKE(dev
) || IS_KABYLAKE(dev
))
6200 __gen6_update_ring_freq(dev
);
6201 } else if (IS_BROADWELL(dev
)) {
6202 gen8_enable_rps(dev
);
6203 __gen6_update_ring_freq(dev
);
6205 gen6_enable_rps(dev
);
6206 __gen6_update_ring_freq(dev
);
6209 WARN_ON(dev_priv
->rps
.max_freq
< dev_priv
->rps
.min_freq
);
6210 WARN_ON(dev_priv
->rps
.idle_freq
> dev_priv
->rps
.max_freq
);
6212 WARN_ON(dev_priv
->rps
.efficient_freq
< dev_priv
->rps
.min_freq
);
6213 WARN_ON(dev_priv
->rps
.efficient_freq
> dev_priv
->rps
.max_freq
);
6215 dev_priv
->rps
.enabled
= true;
6217 gen6_enable_rps_interrupts(dev
);
6219 mutex_unlock(&dev_priv
->rps
.hw_lock
);
6221 intel_runtime_pm_put(dev_priv
);
6224 void intel_enable_gt_powersave(struct drm_device
*dev
)
6226 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6228 /* Powersaving is controlled by the host when inside a VM */
6229 if (intel_vgpu_active(dev
))
6232 if (IS_IRONLAKE_M(dev
)) {
6233 mutex_lock(&dev
->struct_mutex
);
6234 ironlake_enable_drps(dev
);
6235 intel_init_emon(dev
);
6236 mutex_unlock(&dev
->struct_mutex
);
6237 } else if (INTEL_INFO(dev
)->gen
>= 6) {
6239 * PCU communication is slow and this doesn't need to be
6240 * done at any specific time, so do this out of our fast path
6241 * to make resume and init faster.
6243 * We depend on the HW RC6 power context save/restore
6244 * mechanism when entering D3 through runtime PM suspend. So
6245 * disable RPM until RPS/RC6 is properly setup. We can only
6246 * get here via the driver load/system resume/runtime resume
6247 * paths, so the _noresume version is enough (and in case of
6248 * runtime resume it's necessary).
6250 if (schedule_delayed_work(&dev_priv
->rps
.delayed_resume_work
,
6251 round_jiffies_up_relative(HZ
)))
6252 intel_runtime_pm_get_noresume(dev_priv
);
6256 void intel_reset_gt_powersave(struct drm_device
*dev
)
6258 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6260 if (INTEL_INFO(dev
)->gen
< 6)
6263 gen6_suspend_rps(dev
);
6264 dev_priv
->rps
.enabled
= false;
6267 static void ibx_init_clock_gating(struct drm_device
*dev
)
6269 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6272 * On Ibex Peak and Cougar Point, we need to disable clock
6273 * gating for the panel power sequencer or it will fail to
6274 * start up when no ports are active.
6276 I915_WRITE(SOUTH_DSPCLK_GATE_D
, PCH_DPLSUNIT_CLOCK_GATE_DISABLE
);
6279 static void g4x_disable_trickle_feed(struct drm_device
*dev
)
6281 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6284 for_each_pipe(dev_priv
, pipe
) {
6285 I915_WRITE(DSPCNTR(pipe
),
6286 I915_READ(DSPCNTR(pipe
)) |
6287 DISPPLANE_TRICKLE_FEED_DISABLE
);
6289 I915_WRITE(DSPSURF(pipe
), I915_READ(DSPSURF(pipe
)));
6290 POSTING_READ(DSPSURF(pipe
));
6294 static void ilk_init_lp_watermarks(struct drm_device
*dev
)
6296 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6298 I915_WRITE(WM3_LP_ILK
, I915_READ(WM3_LP_ILK
) & ~WM1_LP_SR_EN
);
6299 I915_WRITE(WM2_LP_ILK
, I915_READ(WM2_LP_ILK
) & ~WM1_LP_SR_EN
);
6300 I915_WRITE(WM1_LP_ILK
, I915_READ(WM1_LP_ILK
) & ~WM1_LP_SR_EN
);
6303 * Don't touch WM1S_LP_EN here.
6304 * Doing so could cause underruns.
6308 static void ironlake_init_clock_gating(struct drm_device
*dev
)
6310 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6311 uint32_t dspclk_gate
= ILK_VRHUNIT_CLOCK_GATE_DISABLE
;
6315 * WaFbcDisableDpfcClockGating:ilk
6317 dspclk_gate
|= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE
|
6318 ILK_DPFCUNIT_CLOCK_GATE_DISABLE
|
6319 ILK_DPFDUNIT_CLOCK_GATE_ENABLE
;
6321 I915_WRITE(PCH_3DCGDIS0
,
6322 MARIUNIT_CLOCK_GATE_DISABLE
|
6323 SVSMUNIT_CLOCK_GATE_DISABLE
);
6324 I915_WRITE(PCH_3DCGDIS1
,
6325 VFMUNIT_CLOCK_GATE_DISABLE
);
6328 * According to the spec the following bits should be set in
6329 * order to enable memory self-refresh
6330 * The bit 22/21 of 0x42004
6331 * The bit 5 of 0x42020
6332 * The bit 15 of 0x45000
6334 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6335 (I915_READ(ILK_DISPLAY_CHICKEN2
) |
6336 ILK_DPARB_GATE
| ILK_VSDPFD_FULL
));
6337 dspclk_gate
|= ILK_DPARBUNIT_CLOCK_GATE_ENABLE
;
6338 I915_WRITE(DISP_ARB_CTL
,
6339 (I915_READ(DISP_ARB_CTL
) |
6342 ilk_init_lp_watermarks(dev
);
6345 * Based on the document from hardware guys the following bits
6346 * should be set unconditionally in order to enable FBC.
6347 * The bit 22 of 0x42000
6348 * The bit 22 of 0x42004
6349 * The bit 7,8,9 of 0x42020.
6351 if (IS_IRONLAKE_M(dev
)) {
6352 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
6353 I915_WRITE(ILK_DISPLAY_CHICKEN1
,
6354 I915_READ(ILK_DISPLAY_CHICKEN1
) |
6356 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6357 I915_READ(ILK_DISPLAY_CHICKEN2
) |
6361 I915_WRITE(ILK_DSPCLK_GATE_D
, dspclk_gate
);
6363 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6364 I915_READ(ILK_DISPLAY_CHICKEN2
) |
6365 ILK_ELPIN_409_SELECT
);
6366 I915_WRITE(_3D_CHICKEN2
,
6367 _3D_CHICKEN2_WM_READ_PIPELINED
<< 16 |
6368 _3D_CHICKEN2_WM_READ_PIPELINED
);
6370 /* WaDisableRenderCachePipelinedFlush:ilk */
6371 I915_WRITE(CACHE_MODE_0
,
6372 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE
));
6374 /* WaDisable_RenderCache_OperationalFlush:ilk */
6375 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6377 g4x_disable_trickle_feed(dev
);
6379 ibx_init_clock_gating(dev
);
6382 static void cpt_init_clock_gating(struct drm_device
*dev
)
6384 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6389 * On Ibex Peak and Cougar Point, we need to disable clock
6390 * gating for the panel power sequencer or it will fail to
6391 * start up when no ports are active.
6393 I915_WRITE(SOUTH_DSPCLK_GATE_D
, PCH_DPLSUNIT_CLOCK_GATE_DISABLE
|
6394 PCH_DPLUNIT_CLOCK_GATE_DISABLE
|
6395 PCH_CPUNIT_CLOCK_GATE_DISABLE
);
6396 I915_WRITE(SOUTH_CHICKEN2
, I915_READ(SOUTH_CHICKEN2
) |
6397 DPLS_EDP_PPS_FIX_DIS
);
6398 /* The below fixes the weird display corruption, a few pixels shifted
6399 * downward, on (only) LVDS of some HP laptops with IVY.
6401 for_each_pipe(dev_priv
, pipe
) {
6402 val
= I915_READ(TRANS_CHICKEN2(pipe
));
6403 val
|= TRANS_CHICKEN2_TIMING_OVERRIDE
;
6404 val
&= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED
;
6405 if (dev_priv
->vbt
.fdi_rx_polarity_inverted
)
6406 val
|= TRANS_CHICKEN2_FDI_POLARITY_REVERSED
;
6407 val
&= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK
;
6408 val
&= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER
;
6409 val
&= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH
;
6410 I915_WRITE(TRANS_CHICKEN2(pipe
), val
);
6412 /* WADP0ClockGatingDisable */
6413 for_each_pipe(dev_priv
, pipe
) {
6414 I915_WRITE(TRANS_CHICKEN1(pipe
),
6415 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE
);
6419 static void gen6_check_mch_setup(struct drm_device
*dev
)
6421 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6424 tmp
= I915_READ(MCH_SSKPD
);
6425 if ((tmp
& MCH_SSKPD_WM0_MASK
) != MCH_SSKPD_WM0_VAL
)
6426 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
6430 static void gen6_init_clock_gating(struct drm_device
*dev
)
6432 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6433 uint32_t dspclk_gate
= ILK_VRHUNIT_CLOCK_GATE_DISABLE
;
6435 I915_WRITE(ILK_DSPCLK_GATE_D
, dspclk_gate
);
6437 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6438 I915_READ(ILK_DISPLAY_CHICKEN2
) |
6439 ILK_ELPIN_409_SELECT
);
6441 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
6442 I915_WRITE(_3D_CHICKEN
,
6443 _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB
));
6445 /* WaDisable_RenderCache_OperationalFlush:snb */
6446 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6449 * BSpec recoomends 8x4 when MSAA is used,
6450 * however in practice 16x4 seems fastest.
6452 * Note that PS/WM thread counts depend on the WIZ hashing
6453 * disable bit, which we don't touch here, but it's good
6454 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6456 I915_WRITE(GEN6_GT_MODE
,
6457 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
6459 ilk_init_lp_watermarks(dev
);
6461 I915_WRITE(CACHE_MODE_0
,
6462 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB
));
6464 I915_WRITE(GEN6_UCGCTL1
,
6465 I915_READ(GEN6_UCGCTL1
) |
6466 GEN6_BLBUNIT_CLOCK_GATE_DISABLE
|
6467 GEN6_CSUNIT_CLOCK_GATE_DISABLE
);
6469 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
6470 * gating disable must be set. Failure to set it results in
6471 * flickering pixels due to Z write ordering failures after
6472 * some amount of runtime in the Mesa "fire" demo, and Unigine
6473 * Sanctuary and Tropics, and apparently anything else with
6474 * alpha test or pixel discard.
6476 * According to the spec, bit 11 (RCCUNIT) must also be set,
6477 * but we didn't debug actual testcases to find it out.
6479 * WaDisableRCCUnitClockGating:snb
6480 * WaDisableRCPBUnitClockGating:snb
6482 I915_WRITE(GEN6_UCGCTL2
,
6483 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE
|
6484 GEN6_RCCUNIT_CLOCK_GATE_DISABLE
);
6486 /* WaStripsFansDisableFastClipPerformanceFix:snb */
6487 I915_WRITE(_3D_CHICKEN3
,
6488 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL
));
6492 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
6493 * 3DSTATE_SF number of SF output attributes is more than 16."
6495 I915_WRITE(_3D_CHICKEN3
,
6496 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH
));
6499 * According to the spec the following bits should be
6500 * set in order to enable memory self-refresh and fbc:
6501 * The bit21 and bit22 of 0x42000
6502 * The bit21 and bit22 of 0x42004
6503 * The bit5 and bit7 of 0x42020
6504 * The bit14 of 0x70180
6505 * The bit14 of 0x71180
6507 * WaFbcAsynchFlipDisableFbcQueue:snb
6509 I915_WRITE(ILK_DISPLAY_CHICKEN1
,
6510 I915_READ(ILK_DISPLAY_CHICKEN1
) |
6511 ILK_FBCQ_DIS
| ILK_PABSTRETCH_DIS
);
6512 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6513 I915_READ(ILK_DISPLAY_CHICKEN2
) |
6514 ILK_DPARB_GATE
| ILK_VSDPFD_FULL
);
6515 I915_WRITE(ILK_DSPCLK_GATE_D
,
6516 I915_READ(ILK_DSPCLK_GATE_D
) |
6517 ILK_DPARBUNIT_CLOCK_GATE_ENABLE
|
6518 ILK_DPFDUNIT_CLOCK_GATE_ENABLE
);
6520 g4x_disable_trickle_feed(dev
);
6522 cpt_init_clock_gating(dev
);
6524 gen6_check_mch_setup(dev
);
6527 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private
*dev_priv
)
6529 uint32_t reg
= I915_READ(GEN7_FF_THREAD_MODE
);
6532 * WaVSThreadDispatchOverride:ivb,vlv
6534 * This actually overrides the dispatch
6535 * mode for all thread types.
6537 reg
&= ~GEN7_FF_SCHED_MASK
;
6538 reg
|= GEN7_FF_TS_SCHED_HW
;
6539 reg
|= GEN7_FF_VS_SCHED_HW
;
6540 reg
|= GEN7_FF_DS_SCHED_HW
;
6542 I915_WRITE(GEN7_FF_THREAD_MODE
, reg
);
6545 static void lpt_init_clock_gating(struct drm_device
*dev
)
6547 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6550 * TODO: this bit should only be enabled when really needed, then
6551 * disabled when not needed anymore in order to save power.
6553 if (HAS_PCH_LPT_LP(dev
))
6554 I915_WRITE(SOUTH_DSPCLK_GATE_D
,
6555 I915_READ(SOUTH_DSPCLK_GATE_D
) |
6556 PCH_LP_PARTITION_LEVEL_DISABLE
);
6558 /* WADPOClockGatingDisable:hsw */
6559 I915_WRITE(TRANS_CHICKEN1(PIPE_A
),
6560 I915_READ(TRANS_CHICKEN1(PIPE_A
)) |
6561 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE
);
6564 static void lpt_suspend_hw(struct drm_device
*dev
)
6566 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6568 if (HAS_PCH_LPT_LP(dev
)) {
6569 uint32_t val
= I915_READ(SOUTH_DSPCLK_GATE_D
);
6571 val
&= ~PCH_LP_PARTITION_LEVEL_DISABLE
;
6572 I915_WRITE(SOUTH_DSPCLK_GATE_D
, val
);
6576 static void broadwell_init_clock_gating(struct drm_device
*dev
)
6578 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6582 ilk_init_lp_watermarks(dev
);
6584 /* WaSwitchSolVfFArbitrationPriority:bdw */
6585 I915_WRITE(GAM_ECOCHK
, I915_READ(GAM_ECOCHK
) | HSW_ECOCHK_ARB_PRIO_SOL
);
6587 /* WaPsrDPAMaskVBlankInSRD:bdw */
6588 I915_WRITE(CHICKEN_PAR1_1
,
6589 I915_READ(CHICKEN_PAR1_1
) | DPA_MASK_VBLANK_SRD
);
6591 /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
6592 for_each_pipe(dev_priv
, pipe
) {
6593 I915_WRITE(CHICKEN_PIPESL_1(pipe
),
6594 I915_READ(CHICKEN_PIPESL_1(pipe
)) |
6595 BDW_DPRS_MASK_VBLANK_SRD
);
6598 /* WaVSRefCountFullforceMissDisable:bdw */
6599 /* WaDSRefCountFullforceMissDisable:bdw */
6600 I915_WRITE(GEN7_FF_THREAD_MODE
,
6601 I915_READ(GEN7_FF_THREAD_MODE
) &
6602 ~(GEN8_FF_DS_REF_CNT_FFME
| GEN7_FF_VS_REF_CNT_FFME
));
6604 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL
,
6605 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE
));
6607 /* WaDisableSDEUnitClockGating:bdw */
6608 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
6609 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
6612 * WaProgramL3SqcReg1Default:bdw
6613 * WaTempDisableDOPClkGating:bdw
6615 misccpctl
= I915_READ(GEN7_MISCCPCTL
);
6616 I915_WRITE(GEN7_MISCCPCTL
, misccpctl
& ~GEN7_DOP_CLOCK_GATE_ENABLE
);
6617 I915_WRITE(GEN8_L3SQCREG1
, BDW_WA_L3SQCREG1_DEFAULT
);
6618 I915_WRITE(GEN7_MISCCPCTL
, misccpctl
);
6621 * WaGttCachingOffByDefault:bdw
6622 * GTT cache may not work with big pages, so if those
6623 * are ever enabled GTT cache may need to be disabled.
6625 I915_WRITE(HSW_GTT_CACHE_EN
, GTT_CACHE_EN_ALL
);
6627 lpt_init_clock_gating(dev
);
6630 static void haswell_init_clock_gating(struct drm_device
*dev
)
6632 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6634 ilk_init_lp_watermarks(dev
);
6636 /* L3 caching of data atomics doesn't work -- disable it. */
6637 I915_WRITE(HSW_SCRATCH1
, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE
);
6638 I915_WRITE(HSW_ROW_CHICKEN3
,
6639 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE
));
6641 /* This is required by WaCatErrorRejectionIssue:hsw */
6642 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
6643 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
6644 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
6646 /* WaVSRefCountFullforceMissDisable:hsw */
6647 I915_WRITE(GEN7_FF_THREAD_MODE
,
6648 I915_READ(GEN7_FF_THREAD_MODE
) & ~GEN7_FF_VS_REF_CNT_FFME
);
6650 /* WaDisable_RenderCache_OperationalFlush:hsw */
6651 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6653 /* enable HiZ Raw Stall Optimization */
6654 I915_WRITE(CACHE_MODE_0_GEN7
,
6655 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE
));
6657 /* WaDisable4x2SubspanOptimization:hsw */
6658 I915_WRITE(CACHE_MODE_1
,
6659 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
6662 * BSpec recommends 8x4 when MSAA is used,
6663 * however in practice 16x4 seems fastest.
6665 * Note that PS/WM thread counts depend on the WIZ hashing
6666 * disable bit, which we don't touch here, but it's good
6667 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6669 I915_WRITE(GEN7_GT_MODE
,
6670 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
6672 /* WaSampleCChickenBitEnable:hsw */
6673 I915_WRITE(HALF_SLICE_CHICKEN3
,
6674 _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE
));
6676 /* WaSwitchSolVfFArbitrationPriority:hsw */
6677 I915_WRITE(GAM_ECOCHK
, I915_READ(GAM_ECOCHK
) | HSW_ECOCHK_ARB_PRIO_SOL
);
6679 /* WaRsPkgCStateDisplayPMReq:hsw */
6680 I915_WRITE(CHICKEN_PAR1_1
,
6681 I915_READ(CHICKEN_PAR1_1
) | FORCE_ARB_IDLE_PLANES
);
6683 lpt_init_clock_gating(dev
);
6686 static void ivybridge_init_clock_gating(struct drm_device
*dev
)
6688 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6691 ilk_init_lp_watermarks(dev
);
6693 I915_WRITE(ILK_DSPCLK_GATE_D
, ILK_VRHUNIT_CLOCK_GATE_DISABLE
);
6695 /* WaDisableEarlyCull:ivb */
6696 I915_WRITE(_3D_CHICKEN3
,
6697 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL
));
6699 /* WaDisableBackToBackFlipFix:ivb */
6700 I915_WRITE(IVB_CHICKEN3
,
6701 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE
|
6702 CHICKEN3_DGMG_DONE_FIX_DISABLE
);
6704 /* WaDisablePSDDualDispatchEnable:ivb */
6705 if (IS_IVB_GT1(dev
))
6706 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1
,
6707 _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE
));
6709 /* WaDisable_RenderCache_OperationalFlush:ivb */
6710 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6712 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
6713 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1
,
6714 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC
);
6716 /* WaApplyL3ControlAndL3ChickenMode:ivb */
6717 I915_WRITE(GEN7_L3CNTLREG1
,
6718 GEN7_WA_FOR_GEN7_L3_CONTROL
);
6719 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER
,
6720 GEN7_WA_L3_CHICKEN_MODE
);
6721 if (IS_IVB_GT1(dev
))
6722 I915_WRITE(GEN7_ROW_CHICKEN2
,
6723 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
6725 /* must write both registers */
6726 I915_WRITE(GEN7_ROW_CHICKEN2
,
6727 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
6728 I915_WRITE(GEN7_ROW_CHICKEN2_GT2
,
6729 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
6732 /* WaForceL3Serialization:ivb */
6733 I915_WRITE(GEN7_L3SQCREG4
, I915_READ(GEN7_L3SQCREG4
) &
6734 ~L3SQ_URB_READ_CAM_MATCH_DISABLE
);
6737 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6738 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
6740 I915_WRITE(GEN6_UCGCTL2
,
6741 GEN6_RCZUNIT_CLOCK_GATE_DISABLE
);
6743 /* This is required by WaCatErrorRejectionIssue:ivb */
6744 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
6745 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
6746 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
6748 g4x_disable_trickle_feed(dev
);
6750 gen7_setup_fixed_func_scheduler(dev_priv
);
6752 if (0) { /* causes HiZ corruption on ivb:gt1 */
6753 /* enable HiZ Raw Stall Optimization */
6754 I915_WRITE(CACHE_MODE_0_GEN7
,
6755 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE
));
6758 /* WaDisable4x2SubspanOptimization:ivb */
6759 I915_WRITE(CACHE_MODE_1
,
6760 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
6763 * BSpec recommends 8x4 when MSAA is used,
6764 * however in practice 16x4 seems fastest.
6766 * Note that PS/WM thread counts depend on the WIZ hashing
6767 * disable bit, which we don't touch here, but it's good
6768 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6770 I915_WRITE(GEN7_GT_MODE
,
6771 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
6773 snpcr
= I915_READ(GEN6_MBCUNIT_SNPCR
);
6774 snpcr
&= ~GEN6_MBC_SNPCR_MASK
;
6775 snpcr
|= GEN6_MBC_SNPCR_MED
;
6776 I915_WRITE(GEN6_MBCUNIT_SNPCR
, snpcr
);
6778 if (!HAS_PCH_NOP(dev
))
6779 cpt_init_clock_gating(dev
);
6781 gen6_check_mch_setup(dev
);
6784 static void vlv_init_display_clock_gating(struct drm_i915_private
*dev_priv
)
6786 I915_WRITE(DSPCLK_GATE_D
, VRHUNIT_CLOCK_GATE_DISABLE
);
6789 * Disable trickle feed and enable pnd deadline calculation
6791 I915_WRITE(MI_ARB_VLV
, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
);
6792 I915_WRITE(CBR1_VLV
, 0);
6795 static void valleyview_init_clock_gating(struct drm_device
*dev
)
6797 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6799 vlv_init_display_clock_gating(dev_priv
);
6801 /* WaDisableEarlyCull:vlv */
6802 I915_WRITE(_3D_CHICKEN3
,
6803 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL
));
6805 /* WaDisableBackToBackFlipFix:vlv */
6806 I915_WRITE(IVB_CHICKEN3
,
6807 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE
|
6808 CHICKEN3_DGMG_DONE_FIX_DISABLE
);
6810 /* WaPsdDispatchEnable:vlv */
6811 /* WaDisablePSDDualDispatchEnable:vlv */
6812 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1
,
6813 _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP
|
6814 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE
));
6816 /* WaDisable_RenderCache_OperationalFlush:vlv */
6817 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6819 /* WaForceL3Serialization:vlv */
6820 I915_WRITE(GEN7_L3SQCREG4
, I915_READ(GEN7_L3SQCREG4
) &
6821 ~L3SQ_URB_READ_CAM_MATCH_DISABLE
);
6823 /* WaDisableDopClockGating:vlv */
6824 I915_WRITE(GEN7_ROW_CHICKEN2
,
6825 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
6827 /* This is required by WaCatErrorRejectionIssue:vlv */
6828 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
6829 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
6830 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
6832 gen7_setup_fixed_func_scheduler(dev_priv
);
6835 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6836 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
6838 I915_WRITE(GEN6_UCGCTL2
,
6839 GEN6_RCZUNIT_CLOCK_GATE_DISABLE
);
6841 /* WaDisableL3Bank2xClockGate:vlv
6842 * Disabling L3 clock gating- MMIO 940c[25] = 1
6843 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
6844 I915_WRITE(GEN7_UCGCTL4
,
6845 I915_READ(GEN7_UCGCTL4
) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE
);
6848 * BSpec says this must be set, even though
6849 * WaDisable4x2SubspanOptimization isn't listed for VLV.
6851 I915_WRITE(CACHE_MODE_1
,
6852 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
6855 * BSpec recommends 8x4 when MSAA is used,
6856 * however in practice 16x4 seems fastest.
6858 * Note that PS/WM thread counts depend on the WIZ hashing
6859 * disable bit, which we don't touch here, but it's good
6860 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6862 I915_WRITE(GEN7_GT_MODE
,
6863 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
6866 * WaIncreaseL3CreditsForVLVB0:vlv
6867 * This is the hardware default actually.
6869 I915_WRITE(GEN7_L3SQCREG1
, VLV_B0_WA_L3SQCREG1_VALUE
);
6872 * WaDisableVLVClockGating_VBIIssue:vlv
6873 * Disable clock gating on th GCFG unit to prevent a delay
6874 * in the reporting of vblank events.
6876 I915_WRITE(VLV_GUNIT_CLOCK_GATE
, GCFG_DIS
);
6879 static void cherryview_init_clock_gating(struct drm_device
*dev
)
6881 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6883 vlv_init_display_clock_gating(dev_priv
);
6885 /* WaVSRefCountFullforceMissDisable:chv */
6886 /* WaDSRefCountFullforceMissDisable:chv */
6887 I915_WRITE(GEN7_FF_THREAD_MODE
,
6888 I915_READ(GEN7_FF_THREAD_MODE
) &
6889 ~(GEN8_FF_DS_REF_CNT_FFME
| GEN7_FF_VS_REF_CNT_FFME
));
6891 /* WaDisableSemaphoreAndSyncFlipWait:chv */
6892 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL
,
6893 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE
));
6895 /* WaDisableCSUnitClockGating:chv */
6896 I915_WRITE(GEN6_UCGCTL1
, I915_READ(GEN6_UCGCTL1
) |
6897 GEN6_CSUNIT_CLOCK_GATE_DISABLE
);
6899 /* WaDisableSDEUnitClockGating:chv */
6900 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
6901 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
6904 * GTT cache may not work with big pages, so if those
6905 * are ever enabled GTT cache may need to be disabled.
6907 I915_WRITE(HSW_GTT_CACHE_EN
, GTT_CACHE_EN_ALL
);
6910 static void g4x_init_clock_gating(struct drm_device
*dev
)
6912 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6913 uint32_t dspclk_gate
;
6915 I915_WRITE(RENCLK_GATE_D1
, 0);
6916 I915_WRITE(RENCLK_GATE_D2
, VF_UNIT_CLOCK_GATE_DISABLE
|
6917 GS_UNIT_CLOCK_GATE_DISABLE
|
6918 CL_UNIT_CLOCK_GATE_DISABLE
);
6919 I915_WRITE(RAMCLK_GATE_D
, 0);
6920 dspclk_gate
= VRHUNIT_CLOCK_GATE_DISABLE
|
6921 OVRUNIT_CLOCK_GATE_DISABLE
|
6922 OVCUNIT_CLOCK_GATE_DISABLE
;
6924 dspclk_gate
|= DSSUNIT_CLOCK_GATE_DISABLE
;
6925 I915_WRITE(DSPCLK_GATE_D
, dspclk_gate
);
6927 /* WaDisableRenderCachePipelinedFlush */
6928 I915_WRITE(CACHE_MODE_0
,
6929 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE
));
6931 /* WaDisable_RenderCache_OperationalFlush:g4x */
6932 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6934 g4x_disable_trickle_feed(dev
);
6937 static void crestline_init_clock_gating(struct drm_device
*dev
)
6939 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6941 I915_WRITE(RENCLK_GATE_D1
, I965_RCC_CLOCK_GATE_DISABLE
);
6942 I915_WRITE(RENCLK_GATE_D2
, 0);
6943 I915_WRITE(DSPCLK_GATE_D
, 0);
6944 I915_WRITE(RAMCLK_GATE_D
, 0);
6945 I915_WRITE16(DEUC
, 0);
6946 I915_WRITE(MI_ARB_STATE
,
6947 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
6949 /* WaDisable_RenderCache_OperationalFlush:gen4 */
6950 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6953 static void broadwater_init_clock_gating(struct drm_device
*dev
)
6955 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6957 I915_WRITE(RENCLK_GATE_D1
, I965_RCZ_CLOCK_GATE_DISABLE
|
6958 I965_RCC_CLOCK_GATE_DISABLE
|
6959 I965_RCPB_CLOCK_GATE_DISABLE
|
6960 I965_ISC_CLOCK_GATE_DISABLE
|
6961 I965_FBC_CLOCK_GATE_DISABLE
);
6962 I915_WRITE(RENCLK_GATE_D2
, 0);
6963 I915_WRITE(MI_ARB_STATE
,
6964 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
6966 /* WaDisable_RenderCache_OperationalFlush:gen4 */
6967 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6970 static void gen3_init_clock_gating(struct drm_device
*dev
)
6972 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6973 u32 dstate
= I915_READ(D_STATE
);
6975 dstate
|= DSTATE_PLL_D3_OFF
| DSTATE_GFX_CLOCK_GATING
|
6976 DSTATE_DOT_CLOCK_GATING
;
6977 I915_WRITE(D_STATE
, dstate
);
6979 if (IS_PINEVIEW(dev
))
6980 I915_WRITE(ECOSKPD
, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY
));
6982 /* IIR "flip pending" means done if this bit is set */
6983 I915_WRITE(ECOSKPD
, _MASKED_BIT_DISABLE(ECO_FLIP_DONE
));
6985 /* interrupts should cause a wake up from C3 */
6986 I915_WRITE(INSTPM
, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN
));
6988 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
6989 I915_WRITE(MI_ARB_STATE
, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE
));
6991 I915_WRITE(MI_ARB_STATE
,
6992 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
6995 static void i85x_init_clock_gating(struct drm_device
*dev
)
6997 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6999 I915_WRITE(RENCLK_GATE_D1
, SV_CLOCK_GATE_DISABLE
);
7001 /* interrupts should cause a wake up from C3 */
7002 I915_WRITE(MI_STATE
, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN
) |
7003 _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE
));
7005 I915_WRITE(MEM_MODE
,
7006 _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE
));
7009 static void i830_init_clock_gating(struct drm_device
*dev
)
7011 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
7013 I915_WRITE(DSPCLK_GATE_D
, OVRUNIT_CLOCK_GATE_DISABLE
);
7015 I915_WRITE(MEM_MODE
,
7016 _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE
) |
7017 _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE
));
7020 void intel_init_clock_gating(struct drm_device
*dev
)
7022 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
7024 if (dev_priv
->display
.init_clock_gating
)
7025 dev_priv
->display
.init_clock_gating(dev
);
7028 void intel_suspend_hw(struct drm_device
*dev
)
7030 if (HAS_PCH_LPT(dev
))
7031 lpt_suspend_hw(dev
);
7034 /* Set up chip specific power management-related functions */
7035 void intel_init_pm(struct drm_device
*dev
)
7037 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
7039 intel_fbc_init(dev_priv
);
7042 if (IS_PINEVIEW(dev
))
7043 i915_pineview_get_mem_freq(dev
);
7044 else if (IS_GEN5(dev
))
7045 i915_ironlake_get_mem_freq(dev
);
7047 /* For FIFO watermark updates */
7048 if (INTEL_INFO(dev
)->gen
>= 9) {
7049 skl_setup_wm_latency(dev
);
7051 if (IS_BROXTON(dev
))
7052 dev_priv
->display
.init_clock_gating
=
7053 bxt_init_clock_gating
;
7054 dev_priv
->display
.update_wm
= skl_update_wm
;
7055 } else if (HAS_PCH_SPLIT(dev
)) {
7056 ilk_setup_wm_latency(dev
);
7058 if ((IS_GEN5(dev
) && dev_priv
->wm
.pri_latency
[1] &&
7059 dev_priv
->wm
.spr_latency
[1] && dev_priv
->wm
.cur_latency
[1]) ||
7060 (!IS_GEN5(dev
) && dev_priv
->wm
.pri_latency
[0] &&
7061 dev_priv
->wm
.spr_latency
[0] && dev_priv
->wm
.cur_latency
[0])) {
7062 dev_priv
->display
.compute_pipe_wm
= ilk_compute_pipe_wm
;
7063 dev_priv
->display
.compute_intermediate_wm
=
7064 ilk_compute_intermediate_wm
;
7065 dev_priv
->display
.initial_watermarks
=
7066 ilk_initial_watermarks
;
7067 dev_priv
->display
.optimize_watermarks
=
7068 ilk_optimize_watermarks
;
7070 DRM_DEBUG_KMS("Failed to read display plane latency. "
7075 dev_priv
->display
.init_clock_gating
= ironlake_init_clock_gating
;
7076 else if (IS_GEN6(dev
))
7077 dev_priv
->display
.init_clock_gating
= gen6_init_clock_gating
;
7078 else if (IS_IVYBRIDGE(dev
))
7079 dev_priv
->display
.init_clock_gating
= ivybridge_init_clock_gating
;
7080 else if (IS_HASWELL(dev
))
7081 dev_priv
->display
.init_clock_gating
= haswell_init_clock_gating
;
7082 else if (INTEL_INFO(dev
)->gen
== 8)
7083 dev_priv
->display
.init_clock_gating
= broadwell_init_clock_gating
;
7084 } else if (IS_CHERRYVIEW(dev
)) {
7085 vlv_setup_wm_latency(dev
);
7087 dev_priv
->display
.update_wm
= vlv_update_wm
;
7088 dev_priv
->display
.init_clock_gating
=
7089 cherryview_init_clock_gating
;
7090 } else if (IS_VALLEYVIEW(dev
)) {
7091 vlv_setup_wm_latency(dev
);
7093 dev_priv
->display
.update_wm
= vlv_update_wm
;
7094 dev_priv
->display
.init_clock_gating
=
7095 valleyview_init_clock_gating
;
7096 } else if (IS_PINEVIEW(dev
)) {
7097 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev
),
7100 dev_priv
->mem_freq
)) {
7101 DRM_INFO("failed to find known CxSR latency "
7102 "(found ddr%s fsb freq %d, mem freq %d), "
7104 (dev_priv
->is_ddr3
== 1) ? "3" : "2",
7105 dev_priv
->fsb_freq
, dev_priv
->mem_freq
);
7106 /* Disable CxSR and never update its watermark again */
7107 intel_set_memory_cxsr(dev_priv
, false);
7108 dev_priv
->display
.update_wm
= NULL
;
7110 dev_priv
->display
.update_wm
= pineview_update_wm
;
7111 dev_priv
->display
.init_clock_gating
= gen3_init_clock_gating
;
7112 } else if (IS_G4X(dev
)) {
7113 dev_priv
->display
.update_wm
= g4x_update_wm
;
7114 dev_priv
->display
.init_clock_gating
= g4x_init_clock_gating
;
7115 } else if (IS_GEN4(dev
)) {
7116 dev_priv
->display
.update_wm
= i965_update_wm
;
7117 if (IS_CRESTLINE(dev
))
7118 dev_priv
->display
.init_clock_gating
= crestline_init_clock_gating
;
7119 else if (IS_BROADWATER(dev
))
7120 dev_priv
->display
.init_clock_gating
= broadwater_init_clock_gating
;
7121 } else if (IS_GEN3(dev
)) {
7122 dev_priv
->display
.update_wm
= i9xx_update_wm
;
7123 dev_priv
->display
.get_fifo_size
= i9xx_get_fifo_size
;
7124 dev_priv
->display
.init_clock_gating
= gen3_init_clock_gating
;
7125 } else if (IS_GEN2(dev
)) {
7126 if (INTEL_INFO(dev
)->num_pipes
== 1) {
7127 dev_priv
->display
.update_wm
= i845_update_wm
;
7128 dev_priv
->display
.get_fifo_size
= i845_get_fifo_size
;
7130 dev_priv
->display
.update_wm
= i9xx_update_wm
;
7131 dev_priv
->display
.get_fifo_size
= i830_get_fifo_size
;
7134 if (IS_I85X(dev
) || IS_I865G(dev
))
7135 dev_priv
->display
.init_clock_gating
= i85x_init_clock_gating
;
7137 dev_priv
->display
.init_clock_gating
= i830_init_clock_gating
;
7139 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
7143 int sandybridge_pcode_read(struct drm_i915_private
*dev_priv
, u32 mbox
, u32
*val
)
7145 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
7147 if (I915_READ(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) {
7148 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
7152 I915_WRITE(GEN6_PCODE_DATA
, *val
);
7153 I915_WRITE(GEN6_PCODE_DATA1
, 0);
7154 I915_WRITE(GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
| mbox
);
7156 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) == 0,
7158 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox
);
7162 *val
= I915_READ(GEN6_PCODE_DATA
);
7163 I915_WRITE(GEN6_PCODE_DATA
, 0);
7168 int sandybridge_pcode_write(struct drm_i915_private
*dev_priv
, u32 mbox
, u32 val
)
7170 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
7172 if (I915_READ(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) {
7173 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
7177 I915_WRITE(GEN6_PCODE_DATA
, val
);
7178 I915_WRITE(GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
| mbox
);
7180 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) == 0,
7182 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox
);
7186 I915_WRITE(GEN6_PCODE_DATA
, 0);
7191 static int vlv_gpu_freq_div(unsigned int czclk_freq
)
7193 switch (czclk_freq
) {
7208 static int byt_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
7210 int div
, czclk_freq
= DIV_ROUND_CLOSEST(dev_priv
->czclk_freq
, 1000);
7212 div
= vlv_gpu_freq_div(czclk_freq
);
7216 return DIV_ROUND_CLOSEST(czclk_freq
* (val
+ 6 - 0xbd), div
);
7219 static int byt_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
7221 int mul
, czclk_freq
= DIV_ROUND_CLOSEST(dev_priv
->czclk_freq
, 1000);
7223 mul
= vlv_gpu_freq_div(czclk_freq
);
7227 return DIV_ROUND_CLOSEST(mul
* val
, czclk_freq
) + 0xbd - 6;
7230 static int chv_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
7232 int div
, czclk_freq
= DIV_ROUND_CLOSEST(dev_priv
->czclk_freq
, 1000);
7234 div
= vlv_gpu_freq_div(czclk_freq
) / 2;
7238 return DIV_ROUND_CLOSEST(czclk_freq
* val
, 2 * div
) / 2;
7241 static int chv_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
7243 int mul
, czclk_freq
= DIV_ROUND_CLOSEST(dev_priv
->czclk_freq
, 1000);
7245 mul
= vlv_gpu_freq_div(czclk_freq
) / 2;
7249 /* CHV needs even values */
7250 return DIV_ROUND_CLOSEST(val
* 2 * mul
, czclk_freq
) * 2;
7253 int intel_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
7255 if (IS_GEN9(dev_priv
->dev
))
7256 return DIV_ROUND_CLOSEST(val
* GT_FREQUENCY_MULTIPLIER
,
7258 else if (IS_CHERRYVIEW(dev_priv
->dev
))
7259 return chv_gpu_freq(dev_priv
, val
);
7260 else if (IS_VALLEYVIEW(dev_priv
->dev
))
7261 return byt_gpu_freq(dev_priv
, val
);
7263 return val
* GT_FREQUENCY_MULTIPLIER
;
7266 int intel_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
7268 if (IS_GEN9(dev_priv
->dev
))
7269 return DIV_ROUND_CLOSEST(val
* GEN9_FREQ_SCALER
,
7270 GT_FREQUENCY_MULTIPLIER
);
7271 else if (IS_CHERRYVIEW(dev_priv
->dev
))
7272 return chv_freq_opcode(dev_priv
, val
);
7273 else if (IS_VALLEYVIEW(dev_priv
->dev
))
7274 return byt_freq_opcode(dev_priv
, val
);
7276 return DIV_ROUND_CLOSEST(val
, GT_FREQUENCY_MULTIPLIER
);
7279 struct request_boost
{
7280 struct work_struct work
;
7281 struct drm_i915_gem_request
*req
;
7284 static void __intel_rps_boost_work(struct work_struct
*work
)
7286 struct request_boost
*boost
= container_of(work
, struct request_boost
, work
);
7287 struct drm_i915_gem_request
*req
= boost
->req
;
7289 if (!i915_gem_request_completed(req
, true))
7290 gen6_rps_boost(to_i915(req
->ring
->dev
), NULL
,
7291 req
->emitted_jiffies
);
7293 i915_gem_request_unreference__unlocked(req
);
7297 void intel_queue_rps_boost_for_request(struct drm_device
*dev
,
7298 struct drm_i915_gem_request
*req
)
7300 struct request_boost
*boost
;
7302 if (req
== NULL
|| INTEL_INFO(dev
)->gen
< 6)
7305 if (i915_gem_request_completed(req
, true))
7308 boost
= kmalloc(sizeof(*boost
), GFP_ATOMIC
);
7312 i915_gem_request_reference(req
);
7315 INIT_WORK(&boost
->work
, __intel_rps_boost_work
);
7316 queue_work(to_i915(dev
)->wq
, &boost
->work
);
7319 void intel_pm_setup(struct drm_device
*dev
)
7321 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
7323 mutex_init(&dev_priv
->rps
.hw_lock
);
7324 spin_lock_init(&dev_priv
->rps
.client_lock
);
7326 INIT_DELAYED_WORK(&dev_priv
->rps
.delayed_resume_work
,
7327 intel_gen6_powersave_work
);
7328 INIT_LIST_HEAD(&dev_priv
->rps
.clients
);
7329 INIT_LIST_HEAD(&dev_priv
->rps
.semaphores
.link
);
7330 INIT_LIST_HEAD(&dev_priv
->rps
.mmioflips
.link
);
7332 dev_priv
->pm
.suspended
= false;
7333 atomic_set(&dev_priv
->pm
.wakeref_count
, 0);
7334 atomic_set(&dev_priv
->pm
.atomic_seq
, 0);