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 pixel_rate
= div_u64((uint64_t) pixel_rate
* pipe_w
* pipe_h
,
1682 /* latency must be in 0.1us units. */
1683 static uint32_t ilk_wm_method1(uint32_t pixel_rate
, uint8_t bytes_per_pixel
,
1688 if (WARN(latency
== 0, "Latency value missing\n"))
1691 ret
= (uint64_t) pixel_rate
* bytes_per_pixel
* latency
;
1692 ret
= DIV_ROUND_UP_ULL(ret
, 64 * 10000) + 2;
1697 /* latency must be in 0.1us units. */
1698 static uint32_t ilk_wm_method2(uint32_t pixel_rate
, uint32_t pipe_htotal
,
1699 uint32_t horiz_pixels
, uint8_t bytes_per_pixel
,
1704 if (WARN(latency
== 0, "Latency value missing\n"))
1707 ret
= (latency
* pixel_rate
) / (pipe_htotal
* 10000);
1708 ret
= (ret
+ 1) * horiz_pixels
* bytes_per_pixel
;
1709 ret
= DIV_ROUND_UP(ret
, 64) + 2;
1713 static uint32_t ilk_wm_fbc(uint32_t pri_val
, uint32_t horiz_pixels
,
1714 uint8_t bytes_per_pixel
)
1716 return DIV_ROUND_UP(pri_val
* 64, horiz_pixels
* bytes_per_pixel
) + 2;
1719 struct ilk_wm_maximums
{
1727 * For both WM_PIPE and WM_LP.
1728 * mem_value must be in 0.1us units.
1730 static uint32_t ilk_compute_pri_wm(const struct intel_crtc_state
*cstate
,
1731 const struct intel_plane_state
*pstate
,
1735 int bpp
= pstate
->base
.fb
? pstate
->base
.fb
->bits_per_pixel
/ 8 : 0;
1736 uint32_t method1
, method2
;
1738 if (!cstate
->base
.active
|| !pstate
->visible
)
1741 method1
= ilk_wm_method1(ilk_pipe_pixel_rate(cstate
), bpp
, mem_value
);
1746 method2
= ilk_wm_method2(ilk_pipe_pixel_rate(cstate
),
1747 cstate
->base
.adjusted_mode
.crtc_htotal
,
1748 drm_rect_width(&pstate
->dst
),
1752 return min(method1
, method2
);
1756 * For both WM_PIPE and WM_LP.
1757 * mem_value must be in 0.1us units.
1759 static uint32_t ilk_compute_spr_wm(const struct intel_crtc_state
*cstate
,
1760 const struct intel_plane_state
*pstate
,
1763 int bpp
= pstate
->base
.fb
? pstate
->base
.fb
->bits_per_pixel
/ 8 : 0;
1764 uint32_t method1
, method2
;
1766 if (!cstate
->base
.active
|| !pstate
->visible
)
1769 method1
= ilk_wm_method1(ilk_pipe_pixel_rate(cstate
), bpp
, mem_value
);
1770 method2
= ilk_wm_method2(ilk_pipe_pixel_rate(cstate
),
1771 cstate
->base
.adjusted_mode
.crtc_htotal
,
1772 drm_rect_width(&pstate
->dst
),
1775 return min(method1
, method2
);
1779 * For both WM_PIPE and WM_LP.
1780 * mem_value must be in 0.1us units.
1782 static uint32_t ilk_compute_cur_wm(const struct intel_crtc_state
*cstate
,
1783 const struct intel_plane_state
*pstate
,
1786 int bpp
= pstate
->base
.fb
? pstate
->base
.fb
->bits_per_pixel
/ 8 : 0;
1788 if (!cstate
->base
.active
|| !pstate
->visible
)
1791 return ilk_wm_method2(ilk_pipe_pixel_rate(cstate
),
1792 cstate
->base
.adjusted_mode
.crtc_htotal
,
1793 drm_rect_width(&pstate
->dst
),
1798 /* Only for WM_LP. */
1799 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state
*cstate
,
1800 const struct intel_plane_state
*pstate
,
1803 int bpp
= pstate
->base
.fb
? pstate
->base
.fb
->bits_per_pixel
/ 8 : 0;
1805 if (!cstate
->base
.active
|| !pstate
->visible
)
1808 return ilk_wm_fbc(pri_val
, drm_rect_width(&pstate
->dst
), bpp
);
1811 static unsigned int ilk_display_fifo_size(const struct drm_device
*dev
)
1813 if (INTEL_INFO(dev
)->gen
>= 8)
1815 else if (INTEL_INFO(dev
)->gen
>= 7)
1821 static unsigned int ilk_plane_wm_reg_max(const struct drm_device
*dev
,
1822 int level
, bool is_sprite
)
1824 if (INTEL_INFO(dev
)->gen
>= 8)
1825 /* BDW primary/sprite plane watermarks */
1826 return level
== 0 ? 255 : 2047;
1827 else if (INTEL_INFO(dev
)->gen
>= 7)
1828 /* IVB/HSW primary/sprite plane watermarks */
1829 return level
== 0 ? 127 : 1023;
1830 else if (!is_sprite
)
1831 /* ILK/SNB primary plane watermarks */
1832 return level
== 0 ? 127 : 511;
1834 /* ILK/SNB sprite plane watermarks */
1835 return level
== 0 ? 63 : 255;
1838 static unsigned int ilk_cursor_wm_reg_max(const struct drm_device
*dev
,
1841 if (INTEL_INFO(dev
)->gen
>= 7)
1842 return level
== 0 ? 63 : 255;
1844 return level
== 0 ? 31 : 63;
1847 static unsigned int ilk_fbc_wm_reg_max(const struct drm_device
*dev
)
1849 if (INTEL_INFO(dev
)->gen
>= 8)
1855 /* Calculate the maximum primary/sprite plane watermark */
1856 static unsigned int ilk_plane_wm_max(const struct drm_device
*dev
,
1858 const struct intel_wm_config
*config
,
1859 enum intel_ddb_partitioning ddb_partitioning
,
1862 unsigned int fifo_size
= ilk_display_fifo_size(dev
);
1864 /* if sprites aren't enabled, sprites get nothing */
1865 if (is_sprite
&& !config
->sprites_enabled
)
1868 /* HSW allows LP1+ watermarks even with multiple pipes */
1869 if (level
== 0 || config
->num_pipes_active
> 1) {
1870 fifo_size
/= INTEL_INFO(dev
)->num_pipes
;
1873 * For some reason the non self refresh
1874 * FIFO size is only half of the self
1875 * refresh FIFO size on ILK/SNB.
1877 if (INTEL_INFO(dev
)->gen
<= 6)
1881 if (config
->sprites_enabled
) {
1882 /* level 0 is always calculated with 1:1 split */
1883 if (level
> 0 && ddb_partitioning
== INTEL_DDB_PART_5_6
) {
1892 /* clamp to max that the registers can hold */
1893 return min(fifo_size
, ilk_plane_wm_reg_max(dev
, level
, is_sprite
));
1896 /* Calculate the maximum cursor plane watermark */
1897 static unsigned int ilk_cursor_wm_max(const struct drm_device
*dev
,
1899 const struct intel_wm_config
*config
)
1901 /* HSW LP1+ watermarks w/ multiple pipes */
1902 if (level
> 0 && config
->num_pipes_active
> 1)
1905 /* otherwise just report max that registers can hold */
1906 return ilk_cursor_wm_reg_max(dev
, level
);
1909 static void ilk_compute_wm_maximums(const struct drm_device
*dev
,
1911 const struct intel_wm_config
*config
,
1912 enum intel_ddb_partitioning ddb_partitioning
,
1913 struct ilk_wm_maximums
*max
)
1915 max
->pri
= ilk_plane_wm_max(dev
, level
, config
, ddb_partitioning
, false);
1916 max
->spr
= ilk_plane_wm_max(dev
, level
, config
, ddb_partitioning
, true);
1917 max
->cur
= ilk_cursor_wm_max(dev
, level
, config
);
1918 max
->fbc
= ilk_fbc_wm_reg_max(dev
);
1921 static void ilk_compute_wm_reg_maximums(struct drm_device
*dev
,
1923 struct ilk_wm_maximums
*max
)
1925 max
->pri
= ilk_plane_wm_reg_max(dev
, level
, false);
1926 max
->spr
= ilk_plane_wm_reg_max(dev
, level
, true);
1927 max
->cur
= ilk_cursor_wm_reg_max(dev
, level
);
1928 max
->fbc
= ilk_fbc_wm_reg_max(dev
);
1931 static bool ilk_validate_wm_level(int level
,
1932 const struct ilk_wm_maximums
*max
,
1933 struct intel_wm_level
*result
)
1937 /* already determined to be invalid? */
1938 if (!result
->enable
)
1941 result
->enable
= result
->pri_val
<= max
->pri
&&
1942 result
->spr_val
<= max
->spr
&&
1943 result
->cur_val
<= max
->cur
;
1945 ret
= result
->enable
;
1948 * HACK until we can pre-compute everything,
1949 * and thus fail gracefully if LP0 watermarks
1952 if (level
== 0 && !result
->enable
) {
1953 if (result
->pri_val
> max
->pri
)
1954 DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
1955 level
, result
->pri_val
, max
->pri
);
1956 if (result
->spr_val
> max
->spr
)
1957 DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
1958 level
, result
->spr_val
, max
->spr
);
1959 if (result
->cur_val
> max
->cur
)
1960 DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
1961 level
, result
->cur_val
, max
->cur
);
1963 result
->pri_val
= min_t(uint32_t, result
->pri_val
, max
->pri
);
1964 result
->spr_val
= min_t(uint32_t, result
->spr_val
, max
->spr
);
1965 result
->cur_val
= min_t(uint32_t, result
->cur_val
, max
->cur
);
1966 result
->enable
= true;
1972 static void ilk_compute_wm_level(const struct drm_i915_private
*dev_priv
,
1973 const struct intel_crtc
*intel_crtc
,
1975 struct intel_crtc_state
*cstate
,
1976 struct intel_plane_state
*pristate
,
1977 struct intel_plane_state
*sprstate
,
1978 struct intel_plane_state
*curstate
,
1979 struct intel_wm_level
*result
)
1981 uint16_t pri_latency
= dev_priv
->wm
.pri_latency
[level
];
1982 uint16_t spr_latency
= dev_priv
->wm
.spr_latency
[level
];
1983 uint16_t cur_latency
= dev_priv
->wm
.cur_latency
[level
];
1985 /* WM1+ latency values stored in 0.5us units */
1992 result
->pri_val
= ilk_compute_pri_wm(cstate
, pristate
,
1993 pri_latency
, level
);
1994 result
->spr_val
= ilk_compute_spr_wm(cstate
, sprstate
, spr_latency
);
1995 result
->cur_val
= ilk_compute_cur_wm(cstate
, curstate
, cur_latency
);
1996 result
->fbc_val
= ilk_compute_fbc_wm(cstate
, pristate
, result
->pri_val
);
1997 result
->enable
= true;
2001 hsw_compute_linetime_wm(struct drm_device
*dev
, struct drm_crtc
*crtc
)
2003 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2004 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
2005 const struct drm_display_mode
*adjusted_mode
= &intel_crtc
->config
->base
.adjusted_mode
;
2006 u32 linetime
, ips_linetime
;
2008 if (!intel_crtc
->active
)
2011 /* The WM are computed with base on how long it takes to fill a single
2012 * row at the given clock rate, multiplied by 8.
2014 linetime
= DIV_ROUND_CLOSEST(adjusted_mode
->crtc_htotal
* 1000 * 8,
2015 adjusted_mode
->crtc_clock
);
2016 ips_linetime
= DIV_ROUND_CLOSEST(adjusted_mode
->crtc_htotal
* 1000 * 8,
2017 dev_priv
->cdclk_freq
);
2019 return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime
) |
2020 PIPE_WM_LINETIME_TIME(linetime
);
2023 static void intel_read_wm_latency(struct drm_device
*dev
, uint16_t wm
[8])
2025 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2030 int level
, max_level
= ilk_wm_max_level(dev
);
2032 /* read the first set of memory latencies[0:3] */
2033 val
= 0; /* data0 to be programmed to 0 for first set */
2034 mutex_lock(&dev_priv
->rps
.hw_lock
);
2035 ret
= sandybridge_pcode_read(dev_priv
,
2036 GEN9_PCODE_READ_MEM_LATENCY
,
2038 mutex_unlock(&dev_priv
->rps
.hw_lock
);
2041 DRM_ERROR("SKL Mailbox read error = %d\n", ret
);
2045 wm
[0] = val
& GEN9_MEM_LATENCY_LEVEL_MASK
;
2046 wm
[1] = (val
>> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT
) &
2047 GEN9_MEM_LATENCY_LEVEL_MASK
;
2048 wm
[2] = (val
>> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT
) &
2049 GEN9_MEM_LATENCY_LEVEL_MASK
;
2050 wm
[3] = (val
>> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT
) &
2051 GEN9_MEM_LATENCY_LEVEL_MASK
;
2053 /* read the second set of memory latencies[4:7] */
2054 val
= 1; /* data0 to be programmed to 1 for second 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
);
2061 DRM_ERROR("SKL Mailbox read error = %d\n", ret
);
2065 wm
[4] = val
& GEN9_MEM_LATENCY_LEVEL_MASK
;
2066 wm
[5] = (val
>> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT
) &
2067 GEN9_MEM_LATENCY_LEVEL_MASK
;
2068 wm
[6] = (val
>> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT
) &
2069 GEN9_MEM_LATENCY_LEVEL_MASK
;
2070 wm
[7] = (val
>> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT
) &
2071 GEN9_MEM_LATENCY_LEVEL_MASK
;
2074 * WaWmMemoryReadLatency:skl
2076 * punit doesn't take into account the read latency so we need
2077 * to add 2us to the various latency levels we retrieve from
2079 * - W0 is a bit special in that it's the only level that
2080 * can't be disabled if we want to have display working, so
2081 * we always add 2us there.
2082 * - For levels >=1, punit returns 0us latency when they are
2083 * disabled, so we respect that and don't add 2us then
2085 * Additionally, if a level n (n > 1) has a 0us latency, all
2086 * levels m (m >= n) need to be disabled. We make sure to
2087 * sanitize the values out of the punit to satisfy this
2091 for (level
= 1; level
<= max_level
; level
++)
2095 for (i
= level
+ 1; i
<= max_level
; i
++)
2100 } else if (IS_HASWELL(dev
) || IS_BROADWELL(dev
)) {
2101 uint64_t sskpd
= I915_READ64(MCH_SSKPD
);
2103 wm
[0] = (sskpd
>> 56) & 0xFF;
2105 wm
[0] = sskpd
& 0xF;
2106 wm
[1] = (sskpd
>> 4) & 0xFF;
2107 wm
[2] = (sskpd
>> 12) & 0xFF;
2108 wm
[3] = (sskpd
>> 20) & 0x1FF;
2109 wm
[4] = (sskpd
>> 32) & 0x1FF;
2110 } else if (INTEL_INFO(dev
)->gen
>= 6) {
2111 uint32_t sskpd
= I915_READ(MCH_SSKPD
);
2113 wm
[0] = (sskpd
>> SSKPD_WM0_SHIFT
) & SSKPD_WM_MASK
;
2114 wm
[1] = (sskpd
>> SSKPD_WM1_SHIFT
) & SSKPD_WM_MASK
;
2115 wm
[2] = (sskpd
>> SSKPD_WM2_SHIFT
) & SSKPD_WM_MASK
;
2116 wm
[3] = (sskpd
>> SSKPD_WM3_SHIFT
) & SSKPD_WM_MASK
;
2117 } else if (INTEL_INFO(dev
)->gen
>= 5) {
2118 uint32_t mltr
= I915_READ(MLTR_ILK
);
2120 /* ILK primary LP0 latency is 700 ns */
2122 wm
[1] = (mltr
>> MLTR_WM1_SHIFT
) & ILK_SRLT_MASK
;
2123 wm
[2] = (mltr
>> MLTR_WM2_SHIFT
) & ILK_SRLT_MASK
;
2127 static void intel_fixup_spr_wm_latency(struct drm_device
*dev
, uint16_t wm
[5])
2129 /* ILK sprite LP0 latency is 1300 ns */
2130 if (INTEL_INFO(dev
)->gen
== 5)
2134 static void intel_fixup_cur_wm_latency(struct drm_device
*dev
, uint16_t wm
[5])
2136 /* ILK cursor LP0 latency is 1300 ns */
2137 if (INTEL_INFO(dev
)->gen
== 5)
2140 /* WaDoubleCursorLP3Latency:ivb */
2141 if (IS_IVYBRIDGE(dev
))
2145 int ilk_wm_max_level(const struct drm_device
*dev
)
2147 /* how many WM levels are we expecting */
2148 if (INTEL_INFO(dev
)->gen
>= 9)
2150 else if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
2152 else if (INTEL_INFO(dev
)->gen
>= 6)
2158 static void intel_print_wm_latency(struct drm_device
*dev
,
2160 const uint16_t wm
[8])
2162 int level
, max_level
= ilk_wm_max_level(dev
);
2164 for (level
= 0; level
<= max_level
; level
++) {
2165 unsigned int latency
= wm
[level
];
2168 DRM_ERROR("%s WM%d latency not provided\n",
2174 * - latencies are in us on gen9.
2175 * - before then, WM1+ latency values are in 0.5us units
2182 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2183 name
, level
, wm
[level
],
2184 latency
/ 10, latency
% 10);
2188 static bool ilk_increase_wm_latency(struct drm_i915_private
*dev_priv
,
2189 uint16_t wm
[5], uint16_t min
)
2191 int level
, max_level
= ilk_wm_max_level(dev_priv
->dev
);
2196 wm
[0] = max(wm
[0], min
);
2197 for (level
= 1; level
<= max_level
; level
++)
2198 wm
[level
] = max_t(uint16_t, wm
[level
], DIV_ROUND_UP(min
, 5));
2203 static void snb_wm_latency_quirk(struct drm_device
*dev
)
2205 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2209 * The BIOS provided WM memory latency values are often
2210 * inadequate for high resolution displays. Adjust them.
2212 changed
= ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.pri_latency
, 12) |
2213 ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.spr_latency
, 12) |
2214 ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.cur_latency
, 12);
2219 DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2220 intel_print_wm_latency(dev
, "Primary", dev_priv
->wm
.pri_latency
);
2221 intel_print_wm_latency(dev
, "Sprite", dev_priv
->wm
.spr_latency
);
2222 intel_print_wm_latency(dev
, "Cursor", dev_priv
->wm
.cur_latency
);
2225 static void ilk_setup_wm_latency(struct drm_device
*dev
)
2227 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2229 intel_read_wm_latency(dev
, dev_priv
->wm
.pri_latency
);
2231 memcpy(dev_priv
->wm
.spr_latency
, dev_priv
->wm
.pri_latency
,
2232 sizeof(dev_priv
->wm
.pri_latency
));
2233 memcpy(dev_priv
->wm
.cur_latency
, dev_priv
->wm
.pri_latency
,
2234 sizeof(dev_priv
->wm
.pri_latency
));
2236 intel_fixup_spr_wm_latency(dev
, dev_priv
->wm
.spr_latency
);
2237 intel_fixup_cur_wm_latency(dev
, dev_priv
->wm
.cur_latency
);
2239 intel_print_wm_latency(dev
, "Primary", dev_priv
->wm
.pri_latency
);
2240 intel_print_wm_latency(dev
, "Sprite", dev_priv
->wm
.spr_latency
);
2241 intel_print_wm_latency(dev
, "Cursor", dev_priv
->wm
.cur_latency
);
2244 snb_wm_latency_quirk(dev
);
2247 static void skl_setup_wm_latency(struct drm_device
*dev
)
2249 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2251 intel_read_wm_latency(dev
, dev_priv
->wm
.skl_latency
);
2252 intel_print_wm_latency(dev
, "Gen9 Plane", dev_priv
->wm
.skl_latency
);
2255 /* Compute new watermarks for the pipe */
2256 static int ilk_compute_pipe_wm(struct intel_crtc
*intel_crtc
,
2257 struct drm_atomic_state
*state
)
2259 struct intel_pipe_wm
*pipe_wm
;
2260 struct drm_device
*dev
= intel_crtc
->base
.dev
;
2261 const struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2262 struct intel_crtc_state
*cstate
= NULL
;
2263 struct intel_plane
*intel_plane
;
2264 struct drm_plane_state
*ps
;
2265 struct intel_plane_state
*pristate
= NULL
;
2266 struct intel_plane_state
*sprstate
= NULL
;
2267 struct intel_plane_state
*curstate
= NULL
;
2268 int level
, max_level
= ilk_wm_max_level(dev
);
2269 /* LP0 watermark maximums depend on this pipe alone */
2270 struct intel_wm_config config
= {
2271 .num_pipes_active
= 1,
2273 struct ilk_wm_maximums max
;
2275 cstate
= intel_atomic_get_crtc_state(state
, intel_crtc
);
2277 return PTR_ERR(cstate
);
2279 pipe_wm
= &cstate
->wm
.optimal
.ilk
;
2281 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
2282 ps
= drm_atomic_get_plane_state(state
,
2283 &intel_plane
->base
);
2287 if (intel_plane
->base
.type
== DRM_PLANE_TYPE_PRIMARY
)
2288 pristate
= to_intel_plane_state(ps
);
2289 else if (intel_plane
->base
.type
== DRM_PLANE_TYPE_OVERLAY
)
2290 sprstate
= to_intel_plane_state(ps
);
2291 else if (intel_plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
)
2292 curstate
= to_intel_plane_state(ps
);
2295 config
.sprites_enabled
= sprstate
->visible
;
2296 config
.sprites_scaled
= sprstate
->visible
&&
2297 (drm_rect_width(&sprstate
->dst
) != drm_rect_width(&sprstate
->src
) >> 16 ||
2298 drm_rect_height(&sprstate
->dst
) != drm_rect_height(&sprstate
->src
) >> 16);
2300 pipe_wm
->pipe_enabled
= cstate
->base
.active
;
2301 pipe_wm
->sprites_enabled
= config
.sprites_enabled
;
2302 pipe_wm
->sprites_scaled
= config
.sprites_scaled
;
2304 /* ILK/SNB: LP2+ watermarks only w/o sprites */
2305 if (INTEL_INFO(dev
)->gen
<= 6 && sprstate
->visible
)
2308 /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2309 if (config
.sprites_scaled
)
2312 ilk_compute_wm_level(dev_priv
, intel_crtc
, 0, cstate
,
2313 pristate
, sprstate
, curstate
, &pipe_wm
->wm
[0]);
2315 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
2316 pipe_wm
->linetime
= hsw_compute_linetime_wm(dev
,
2319 /* LP0 watermarks always use 1/2 DDB partitioning */
2320 ilk_compute_wm_maximums(dev
, 0, &config
, INTEL_DDB_PART_1_2
, &max
);
2322 /* At least LP0 must be valid */
2323 if (!ilk_validate_wm_level(0, &max
, &pipe_wm
->wm
[0]))
2326 ilk_compute_wm_reg_maximums(dev
, 1, &max
);
2328 for (level
= 1; level
<= max_level
; level
++) {
2329 struct intel_wm_level wm
= {};
2331 ilk_compute_wm_level(dev_priv
, intel_crtc
, level
, cstate
,
2332 pristate
, sprstate
, curstate
, &wm
);
2335 * Disable any watermark level that exceeds the
2336 * register maximums since such watermarks are
2339 if (!ilk_validate_wm_level(level
, &max
, &wm
))
2342 pipe_wm
->wm
[level
] = wm
;
2349 * Merge the watermarks from all active pipes for a specific level.
2351 static void ilk_merge_wm_level(struct drm_device
*dev
,
2353 struct intel_wm_level
*ret_wm
)
2355 const struct intel_crtc
*intel_crtc
;
2357 ret_wm
->enable
= true;
2359 for_each_intel_crtc(dev
, intel_crtc
) {
2360 const struct intel_crtc_state
*cstate
=
2361 to_intel_crtc_state(intel_crtc
->base
.state
);
2362 const struct intel_pipe_wm
*active
= &cstate
->wm
.optimal
.ilk
;
2363 const struct intel_wm_level
*wm
= &active
->wm
[level
];
2365 if (!active
->pipe_enabled
)
2369 * The watermark values may have been used in the past,
2370 * so we must maintain them in the registers for some
2371 * time even if the level is now disabled.
2374 ret_wm
->enable
= false;
2376 ret_wm
->pri_val
= max(ret_wm
->pri_val
, wm
->pri_val
);
2377 ret_wm
->spr_val
= max(ret_wm
->spr_val
, wm
->spr_val
);
2378 ret_wm
->cur_val
= max(ret_wm
->cur_val
, wm
->cur_val
);
2379 ret_wm
->fbc_val
= max(ret_wm
->fbc_val
, wm
->fbc_val
);
2384 * Merge all low power watermarks for all active pipes.
2386 static void ilk_wm_merge(struct drm_device
*dev
,
2387 const struct intel_wm_config
*config
,
2388 const struct ilk_wm_maximums
*max
,
2389 struct intel_pipe_wm
*merged
)
2391 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2392 int level
, max_level
= ilk_wm_max_level(dev
);
2393 int last_enabled_level
= max_level
;
2395 /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2396 if ((INTEL_INFO(dev
)->gen
<= 6 || IS_IVYBRIDGE(dev
)) &&
2397 config
->num_pipes_active
> 1)
2400 /* ILK: FBC WM must be disabled always */
2401 merged
->fbc_wm_enabled
= INTEL_INFO(dev
)->gen
>= 6;
2403 /* merge each WM1+ level */
2404 for (level
= 1; level
<= max_level
; level
++) {
2405 struct intel_wm_level
*wm
= &merged
->wm
[level
];
2407 ilk_merge_wm_level(dev
, level
, wm
);
2409 if (level
> last_enabled_level
)
2411 else if (!ilk_validate_wm_level(level
, max
, wm
))
2412 /* make sure all following levels get disabled */
2413 last_enabled_level
= level
- 1;
2416 * The spec says it is preferred to disable
2417 * FBC WMs instead of disabling a WM level.
2419 if (wm
->fbc_val
> max
->fbc
) {
2421 merged
->fbc_wm_enabled
= false;
2426 /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2428 * FIXME this is racy. FBC might get enabled later.
2429 * What we should check here is whether FBC can be
2430 * enabled sometime later.
2432 if (IS_GEN5(dev
) && !merged
->fbc_wm_enabled
&&
2433 intel_fbc_is_active(dev_priv
)) {
2434 for (level
= 2; level
<= max_level
; level
++) {
2435 struct intel_wm_level
*wm
= &merged
->wm
[level
];
2442 static int ilk_wm_lp_to_level(int wm_lp
, const struct intel_pipe_wm
*pipe_wm
)
2444 /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2445 return wm_lp
+ (wm_lp
>= 2 && pipe_wm
->wm
[4].enable
);
2448 /* The value we need to program into the WM_LPx latency field */
2449 static unsigned int ilk_wm_lp_latency(struct drm_device
*dev
, int level
)
2451 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2453 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
2456 return dev_priv
->wm
.pri_latency
[level
];
2459 static void ilk_compute_wm_results(struct drm_device
*dev
,
2460 const struct intel_pipe_wm
*merged
,
2461 enum intel_ddb_partitioning partitioning
,
2462 struct ilk_wm_values
*results
)
2464 struct intel_crtc
*intel_crtc
;
2467 results
->enable_fbc_wm
= merged
->fbc_wm_enabled
;
2468 results
->partitioning
= partitioning
;
2470 /* LP1+ register values */
2471 for (wm_lp
= 1; wm_lp
<= 3; wm_lp
++) {
2472 const struct intel_wm_level
*r
;
2474 level
= ilk_wm_lp_to_level(wm_lp
, merged
);
2476 r
= &merged
->wm
[level
];
2479 * Maintain the watermark values even if the level is
2480 * disabled. Doing otherwise could cause underruns.
2482 results
->wm_lp
[wm_lp
- 1] =
2483 (ilk_wm_lp_latency(dev
, level
) << WM1_LP_LATENCY_SHIFT
) |
2484 (r
->pri_val
<< WM1_LP_SR_SHIFT
) |
2488 results
->wm_lp
[wm_lp
- 1] |= WM1_LP_SR_EN
;
2490 if (INTEL_INFO(dev
)->gen
>= 8)
2491 results
->wm_lp
[wm_lp
- 1] |=
2492 r
->fbc_val
<< WM1_LP_FBC_SHIFT_BDW
;
2494 results
->wm_lp
[wm_lp
- 1] |=
2495 r
->fbc_val
<< WM1_LP_FBC_SHIFT
;
2498 * Always set WM1S_LP_EN when spr_val != 0, even if the
2499 * level is disabled. Doing otherwise could cause underruns.
2501 if (INTEL_INFO(dev
)->gen
<= 6 && r
->spr_val
) {
2502 WARN_ON(wm_lp
!= 1);
2503 results
->wm_lp_spr
[wm_lp
- 1] = WM1S_LP_EN
| r
->spr_val
;
2505 results
->wm_lp_spr
[wm_lp
- 1] = r
->spr_val
;
2508 /* LP0 register values */
2509 for_each_intel_crtc(dev
, intel_crtc
) {
2510 const struct intel_crtc_state
*cstate
=
2511 to_intel_crtc_state(intel_crtc
->base
.state
);
2512 enum pipe pipe
= intel_crtc
->pipe
;
2513 const struct intel_wm_level
*r
= &cstate
->wm
.optimal
.ilk
.wm
[0];
2515 if (WARN_ON(!r
->enable
))
2518 results
->wm_linetime
[pipe
] = cstate
->wm
.optimal
.ilk
.linetime
;
2520 results
->wm_pipe
[pipe
] =
2521 (r
->pri_val
<< WM0_PIPE_PLANE_SHIFT
) |
2522 (r
->spr_val
<< WM0_PIPE_SPRITE_SHIFT
) |
2527 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2528 * case both are at the same level. Prefer r1 in case they're the same. */
2529 static struct intel_pipe_wm
*ilk_find_best_result(struct drm_device
*dev
,
2530 struct intel_pipe_wm
*r1
,
2531 struct intel_pipe_wm
*r2
)
2533 int level
, max_level
= ilk_wm_max_level(dev
);
2534 int level1
= 0, level2
= 0;
2536 for (level
= 1; level
<= max_level
; level
++) {
2537 if (r1
->wm
[level
].enable
)
2539 if (r2
->wm
[level
].enable
)
2543 if (level1
== level2
) {
2544 if (r2
->fbc_wm_enabled
&& !r1
->fbc_wm_enabled
)
2548 } else if (level1
> level2
) {
2555 /* dirty bits used to track which watermarks need changes */
2556 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2557 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2558 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2559 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2560 #define WM_DIRTY_FBC (1 << 24)
2561 #define WM_DIRTY_DDB (1 << 25)
2563 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private
*dev_priv
,
2564 const struct ilk_wm_values
*old
,
2565 const struct ilk_wm_values
*new)
2567 unsigned int dirty
= 0;
2571 for_each_pipe(dev_priv
, pipe
) {
2572 if (old
->wm_linetime
[pipe
] != new->wm_linetime
[pipe
]) {
2573 dirty
|= WM_DIRTY_LINETIME(pipe
);
2574 /* Must disable LP1+ watermarks too */
2575 dirty
|= WM_DIRTY_LP_ALL
;
2578 if (old
->wm_pipe
[pipe
] != new->wm_pipe
[pipe
]) {
2579 dirty
|= WM_DIRTY_PIPE(pipe
);
2580 /* Must disable LP1+ watermarks too */
2581 dirty
|= WM_DIRTY_LP_ALL
;
2585 if (old
->enable_fbc_wm
!= new->enable_fbc_wm
) {
2586 dirty
|= WM_DIRTY_FBC
;
2587 /* Must disable LP1+ watermarks too */
2588 dirty
|= WM_DIRTY_LP_ALL
;
2591 if (old
->partitioning
!= new->partitioning
) {
2592 dirty
|= WM_DIRTY_DDB
;
2593 /* Must disable LP1+ watermarks too */
2594 dirty
|= WM_DIRTY_LP_ALL
;
2597 /* LP1+ watermarks already deemed dirty, no need to continue */
2598 if (dirty
& WM_DIRTY_LP_ALL
)
2601 /* Find the lowest numbered LP1+ watermark in need of an update... */
2602 for (wm_lp
= 1; wm_lp
<= 3; wm_lp
++) {
2603 if (old
->wm_lp
[wm_lp
- 1] != new->wm_lp
[wm_lp
- 1] ||
2604 old
->wm_lp_spr
[wm_lp
- 1] != new->wm_lp_spr
[wm_lp
- 1])
2608 /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2609 for (; wm_lp
<= 3; wm_lp
++)
2610 dirty
|= WM_DIRTY_LP(wm_lp
);
2615 static bool _ilk_disable_lp_wm(struct drm_i915_private
*dev_priv
,
2618 struct ilk_wm_values
*previous
= &dev_priv
->wm
.hw
;
2619 bool changed
= false;
2621 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp
[2] & WM1_LP_SR_EN
) {
2622 previous
->wm_lp
[2] &= ~WM1_LP_SR_EN
;
2623 I915_WRITE(WM3_LP_ILK
, previous
->wm_lp
[2]);
2626 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp
[1] & WM1_LP_SR_EN
) {
2627 previous
->wm_lp
[1] &= ~WM1_LP_SR_EN
;
2628 I915_WRITE(WM2_LP_ILK
, previous
->wm_lp
[1]);
2631 if (dirty
& WM_DIRTY_LP(1) && previous
->wm_lp
[0] & WM1_LP_SR_EN
) {
2632 previous
->wm_lp
[0] &= ~WM1_LP_SR_EN
;
2633 I915_WRITE(WM1_LP_ILK
, previous
->wm_lp
[0]);
2638 * Don't touch WM1S_LP_EN here.
2639 * Doing so could cause underruns.
2646 * The spec says we shouldn't write when we don't need, because every write
2647 * causes WMs to be re-evaluated, expending some power.
2649 static void ilk_write_wm_values(struct drm_i915_private
*dev_priv
,
2650 struct ilk_wm_values
*results
)
2652 struct drm_device
*dev
= dev_priv
->dev
;
2653 struct ilk_wm_values
*previous
= &dev_priv
->wm
.hw
;
2657 dirty
= ilk_compute_wm_dirty(dev_priv
, previous
, results
);
2661 _ilk_disable_lp_wm(dev_priv
, dirty
);
2663 if (dirty
& WM_DIRTY_PIPE(PIPE_A
))
2664 I915_WRITE(WM0_PIPEA_ILK
, results
->wm_pipe
[0]);
2665 if (dirty
& WM_DIRTY_PIPE(PIPE_B
))
2666 I915_WRITE(WM0_PIPEB_ILK
, results
->wm_pipe
[1]);
2667 if (dirty
& WM_DIRTY_PIPE(PIPE_C
))
2668 I915_WRITE(WM0_PIPEC_IVB
, results
->wm_pipe
[2]);
2670 if (dirty
& WM_DIRTY_LINETIME(PIPE_A
))
2671 I915_WRITE(PIPE_WM_LINETIME(PIPE_A
), results
->wm_linetime
[0]);
2672 if (dirty
& WM_DIRTY_LINETIME(PIPE_B
))
2673 I915_WRITE(PIPE_WM_LINETIME(PIPE_B
), results
->wm_linetime
[1]);
2674 if (dirty
& WM_DIRTY_LINETIME(PIPE_C
))
2675 I915_WRITE(PIPE_WM_LINETIME(PIPE_C
), results
->wm_linetime
[2]);
2677 if (dirty
& WM_DIRTY_DDB
) {
2678 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
)) {
2679 val
= I915_READ(WM_MISC
);
2680 if (results
->partitioning
== INTEL_DDB_PART_1_2
)
2681 val
&= ~WM_MISC_DATA_PARTITION_5_6
;
2683 val
|= WM_MISC_DATA_PARTITION_5_6
;
2684 I915_WRITE(WM_MISC
, val
);
2686 val
= I915_READ(DISP_ARB_CTL2
);
2687 if (results
->partitioning
== INTEL_DDB_PART_1_2
)
2688 val
&= ~DISP_DATA_PARTITION_5_6
;
2690 val
|= DISP_DATA_PARTITION_5_6
;
2691 I915_WRITE(DISP_ARB_CTL2
, val
);
2695 if (dirty
& WM_DIRTY_FBC
) {
2696 val
= I915_READ(DISP_ARB_CTL
);
2697 if (results
->enable_fbc_wm
)
2698 val
&= ~DISP_FBC_WM_DIS
;
2700 val
|= DISP_FBC_WM_DIS
;
2701 I915_WRITE(DISP_ARB_CTL
, val
);
2704 if (dirty
& WM_DIRTY_LP(1) &&
2705 previous
->wm_lp_spr
[0] != results
->wm_lp_spr
[0])
2706 I915_WRITE(WM1S_LP_ILK
, results
->wm_lp_spr
[0]);
2708 if (INTEL_INFO(dev
)->gen
>= 7) {
2709 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp_spr
[1] != results
->wm_lp_spr
[1])
2710 I915_WRITE(WM2S_LP_IVB
, results
->wm_lp_spr
[1]);
2711 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp_spr
[2] != results
->wm_lp_spr
[2])
2712 I915_WRITE(WM3S_LP_IVB
, results
->wm_lp_spr
[2]);
2715 if (dirty
& WM_DIRTY_LP(1) && previous
->wm_lp
[0] != results
->wm_lp
[0])
2716 I915_WRITE(WM1_LP_ILK
, results
->wm_lp
[0]);
2717 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp
[1] != results
->wm_lp
[1])
2718 I915_WRITE(WM2_LP_ILK
, results
->wm_lp
[1]);
2719 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp
[2] != results
->wm_lp
[2])
2720 I915_WRITE(WM3_LP_ILK
, results
->wm_lp
[2]);
2722 dev_priv
->wm
.hw
= *results
;
2725 static bool ilk_disable_lp_wm(struct drm_device
*dev
)
2727 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2729 return _ilk_disable_lp_wm(dev_priv
, WM_DIRTY_LP_ALL
);
2733 * On gen9, we need to allocate Display Data Buffer (DDB) portions to the
2734 * different active planes.
2737 #define SKL_DDB_SIZE 896 /* in blocks */
2738 #define BXT_DDB_SIZE 512
2741 * Return the index of a plane in the SKL DDB and wm result arrays. Primary
2742 * plane is always in slot 0, cursor is always in slot I915_MAX_PLANES-1, and
2743 * other universal planes are in indices 1..n. Note that this may leave unused
2744 * indices between the top "sprite" plane and the cursor.
2747 skl_wm_plane_id(const struct intel_plane
*plane
)
2749 switch (plane
->base
.type
) {
2750 case DRM_PLANE_TYPE_PRIMARY
:
2752 case DRM_PLANE_TYPE_CURSOR
:
2753 return PLANE_CURSOR
;
2754 case DRM_PLANE_TYPE_OVERLAY
:
2755 return plane
->plane
+ 1;
2757 MISSING_CASE(plane
->base
.type
);
2758 return plane
->plane
;
2763 skl_ddb_get_pipe_allocation_limits(struct drm_device
*dev
,
2764 const struct intel_crtc_state
*cstate
,
2765 const struct intel_wm_config
*config
,
2766 struct skl_ddb_entry
*alloc
/* out */)
2768 struct drm_crtc
*for_crtc
= cstate
->base
.crtc
;
2769 struct drm_crtc
*crtc
;
2770 unsigned int pipe_size
, ddb_size
;
2771 int nth_active_pipe
;
2773 if (!cstate
->base
.active
) {
2779 if (IS_BROXTON(dev
))
2780 ddb_size
= BXT_DDB_SIZE
;
2782 ddb_size
= SKL_DDB_SIZE
;
2784 ddb_size
-= 4; /* 4 blocks for bypass path allocation */
2786 nth_active_pipe
= 0;
2787 for_each_crtc(dev
, crtc
) {
2788 if (!to_intel_crtc(crtc
)->active
)
2791 if (crtc
== for_crtc
)
2797 pipe_size
= ddb_size
/ config
->num_pipes_active
;
2798 alloc
->start
= nth_active_pipe
* ddb_size
/ config
->num_pipes_active
;
2799 alloc
->end
= alloc
->start
+ pipe_size
;
2802 static unsigned int skl_cursor_allocation(const struct intel_wm_config
*config
)
2804 if (config
->num_pipes_active
== 1)
2810 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry
*entry
, u32 reg
)
2812 entry
->start
= reg
& 0x3ff;
2813 entry
->end
= (reg
>> 16) & 0x3ff;
2818 void skl_ddb_get_hw_state(struct drm_i915_private
*dev_priv
,
2819 struct skl_ddb_allocation
*ddb
/* out */)
2825 memset(ddb
, 0, sizeof(*ddb
));
2827 for_each_pipe(dev_priv
, pipe
) {
2828 if (!intel_display_power_is_enabled(dev_priv
, POWER_DOMAIN_PIPE(pipe
)))
2831 for_each_plane(dev_priv
, pipe
, plane
) {
2832 val
= I915_READ(PLANE_BUF_CFG(pipe
, plane
));
2833 skl_ddb_entry_init_from_hw(&ddb
->plane
[pipe
][plane
],
2837 val
= I915_READ(CUR_BUF_CFG(pipe
));
2838 skl_ddb_entry_init_from_hw(&ddb
->plane
[pipe
][PLANE_CURSOR
],
2844 skl_plane_relative_data_rate(const struct intel_crtc_state
*cstate
,
2845 const struct drm_plane_state
*pstate
,
2848 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
2849 struct drm_framebuffer
*fb
= pstate
->fb
;
2851 /* for planar format */
2852 if (fb
->pixel_format
== DRM_FORMAT_NV12
) {
2853 if (y
) /* y-plane data rate */
2854 return intel_crtc
->config
->pipe_src_w
*
2855 intel_crtc
->config
->pipe_src_h
*
2856 drm_format_plane_cpp(fb
->pixel_format
, 0);
2857 else /* uv-plane data rate */
2858 return (intel_crtc
->config
->pipe_src_w
/2) *
2859 (intel_crtc
->config
->pipe_src_h
/2) *
2860 drm_format_plane_cpp(fb
->pixel_format
, 1);
2863 /* for packed formats */
2864 return intel_crtc
->config
->pipe_src_w
*
2865 intel_crtc
->config
->pipe_src_h
*
2866 drm_format_plane_cpp(fb
->pixel_format
, 0);
2870 * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
2871 * a 8192x4096@32bpp framebuffer:
2872 * 3 * 4096 * 8192 * 4 < 2^32
2875 skl_get_total_relative_data_rate(const struct intel_crtc_state
*cstate
)
2877 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
2878 struct drm_device
*dev
= intel_crtc
->base
.dev
;
2879 const struct intel_plane
*intel_plane
;
2880 unsigned int total_data_rate
= 0;
2882 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
2883 const struct drm_plane_state
*pstate
= intel_plane
->base
.state
;
2885 if (pstate
->fb
== NULL
)
2888 if (intel_plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
)
2892 total_data_rate
+= skl_plane_relative_data_rate(cstate
,
2896 if (pstate
->fb
->pixel_format
== DRM_FORMAT_NV12
)
2898 total_data_rate
+= skl_plane_relative_data_rate(cstate
,
2903 return total_data_rate
;
2907 skl_allocate_pipe_ddb(struct intel_crtc_state
*cstate
,
2908 struct skl_ddb_allocation
*ddb
/* out */)
2910 struct drm_crtc
*crtc
= cstate
->base
.crtc
;
2911 struct drm_device
*dev
= crtc
->dev
;
2912 struct drm_i915_private
*dev_priv
= to_i915(dev
);
2913 struct intel_wm_config
*config
= &dev_priv
->wm
.config
;
2914 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
2915 struct intel_plane
*intel_plane
;
2916 enum pipe pipe
= intel_crtc
->pipe
;
2917 struct skl_ddb_entry
*alloc
= &ddb
->pipe
[pipe
];
2918 uint16_t alloc_size
, start
, cursor_blocks
;
2919 uint16_t minimum
[I915_MAX_PLANES
];
2920 uint16_t y_minimum
[I915_MAX_PLANES
];
2921 unsigned int total_data_rate
;
2923 skl_ddb_get_pipe_allocation_limits(dev
, cstate
, config
, alloc
);
2924 alloc_size
= skl_ddb_entry_size(alloc
);
2925 if (alloc_size
== 0) {
2926 memset(ddb
->plane
[pipe
], 0, sizeof(ddb
->plane
[pipe
]));
2927 memset(&ddb
->plane
[pipe
][PLANE_CURSOR
], 0,
2928 sizeof(ddb
->plane
[pipe
][PLANE_CURSOR
]));
2932 cursor_blocks
= skl_cursor_allocation(config
);
2933 ddb
->plane
[pipe
][PLANE_CURSOR
].start
= alloc
->end
- cursor_blocks
;
2934 ddb
->plane
[pipe
][PLANE_CURSOR
].end
= alloc
->end
;
2936 alloc_size
-= cursor_blocks
;
2937 alloc
->end
-= cursor_blocks
;
2939 /* 1. Allocate the mininum required blocks for each active plane */
2940 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
2941 struct drm_plane
*plane
= &intel_plane
->base
;
2942 struct drm_framebuffer
*fb
= plane
->state
->fb
;
2943 int id
= skl_wm_plane_id(intel_plane
);
2947 if (plane
->type
== DRM_PLANE_TYPE_CURSOR
)
2951 alloc_size
-= minimum
[id
];
2952 y_minimum
[id
] = (fb
->pixel_format
== DRM_FORMAT_NV12
) ? 8 : 0;
2953 alloc_size
-= y_minimum
[id
];
2957 * 2. Distribute the remaining space in proportion to the amount of
2958 * data each plane needs to fetch from memory.
2960 * FIXME: we may not allocate every single block here.
2962 total_data_rate
= skl_get_total_relative_data_rate(cstate
);
2964 start
= alloc
->start
;
2965 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
2966 struct drm_plane
*plane
= &intel_plane
->base
;
2967 struct drm_plane_state
*pstate
= intel_plane
->base
.state
;
2968 unsigned int data_rate
, y_data_rate
;
2969 uint16_t plane_blocks
, y_plane_blocks
= 0;
2970 int id
= skl_wm_plane_id(intel_plane
);
2972 if (pstate
->fb
== NULL
)
2974 if (plane
->type
== DRM_PLANE_TYPE_CURSOR
)
2977 data_rate
= skl_plane_relative_data_rate(cstate
, pstate
, 0);
2980 * allocation for (packed formats) or (uv-plane part of planar format):
2981 * promote the expression to 64 bits to avoid overflowing, the
2982 * result is < available as data_rate / total_data_rate < 1
2984 plane_blocks
= minimum
[id
];
2985 plane_blocks
+= div_u64((uint64_t)alloc_size
* data_rate
,
2988 ddb
->plane
[pipe
][id
].start
= start
;
2989 ddb
->plane
[pipe
][id
].end
= start
+ plane_blocks
;
2991 start
+= plane_blocks
;
2994 * allocation for y_plane part of planar format:
2996 if (pstate
->fb
->pixel_format
== DRM_FORMAT_NV12
) {
2997 y_data_rate
= skl_plane_relative_data_rate(cstate
,
3000 y_plane_blocks
= y_minimum
[id
];
3001 y_plane_blocks
+= div_u64((uint64_t)alloc_size
* y_data_rate
,
3004 ddb
->y_plane
[pipe
][id
].start
= start
;
3005 ddb
->y_plane
[pipe
][id
].end
= start
+ y_plane_blocks
;
3007 start
+= y_plane_blocks
;
3014 static uint32_t skl_pipe_pixel_rate(const struct intel_crtc_state
*config
)
3016 /* TODO: Take into account the scalers once we support them */
3017 return config
->base
.adjusted_mode
.crtc_clock
;
3021 * The max latency should be 257 (max the punit can code is 255 and we add 2us
3022 * for the read latency) and bytes_per_pixel should always be <= 8, so that
3023 * should allow pixel_rate up to ~2 GHz which seems sufficient since max
3024 * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
3026 static uint32_t skl_wm_method1(uint32_t pixel_rate
, uint8_t bytes_per_pixel
,
3029 uint32_t wm_intermediate_val
, ret
;
3034 wm_intermediate_val
= latency
* pixel_rate
* bytes_per_pixel
/ 512;
3035 ret
= DIV_ROUND_UP(wm_intermediate_val
, 1000);
3040 static uint32_t skl_wm_method2(uint32_t pixel_rate
, uint32_t pipe_htotal
,
3041 uint32_t horiz_pixels
, uint8_t bytes_per_pixel
,
3042 uint64_t tiling
, uint32_t latency
)
3045 uint32_t plane_bytes_per_line
, plane_blocks_per_line
;
3046 uint32_t wm_intermediate_val
;
3051 plane_bytes_per_line
= horiz_pixels
* bytes_per_pixel
;
3053 if (tiling
== I915_FORMAT_MOD_Y_TILED
||
3054 tiling
== I915_FORMAT_MOD_Yf_TILED
) {
3055 plane_bytes_per_line
*= 4;
3056 plane_blocks_per_line
= DIV_ROUND_UP(plane_bytes_per_line
, 512);
3057 plane_blocks_per_line
/= 4;
3059 plane_blocks_per_line
= DIV_ROUND_UP(plane_bytes_per_line
, 512);
3062 wm_intermediate_val
= latency
* pixel_rate
;
3063 ret
= DIV_ROUND_UP(wm_intermediate_val
, pipe_htotal
* 1000) *
3064 plane_blocks_per_line
;
3069 static bool skl_ddb_allocation_changed(const struct skl_ddb_allocation
*new_ddb
,
3070 const struct intel_crtc
*intel_crtc
)
3072 struct drm_device
*dev
= intel_crtc
->base
.dev
;
3073 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3074 const struct skl_ddb_allocation
*cur_ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
3077 * If ddb allocation of pipes changed, it may require recalculation of
3080 if (memcmp(new_ddb
->pipe
, cur_ddb
->pipe
, sizeof(new_ddb
->pipe
)))
3086 static bool skl_compute_plane_wm(const struct drm_i915_private
*dev_priv
,
3087 struct intel_crtc_state
*cstate
,
3088 struct intel_plane
*intel_plane
,
3089 uint16_t ddb_allocation
,
3091 uint16_t *out_blocks
, /* out */
3092 uint8_t *out_lines
/* out */)
3094 struct drm_plane
*plane
= &intel_plane
->base
;
3095 struct drm_framebuffer
*fb
= plane
->state
->fb
;
3096 uint32_t latency
= dev_priv
->wm
.skl_latency
[level
];
3097 uint32_t method1
, method2
;
3098 uint32_t plane_bytes_per_line
, plane_blocks_per_line
;
3099 uint32_t res_blocks
, res_lines
;
3100 uint32_t selected_result
;
3101 uint8_t bytes_per_pixel
;
3103 if (latency
== 0 || !cstate
->base
.active
|| !fb
)
3106 bytes_per_pixel
= drm_format_plane_cpp(fb
->pixel_format
, 0);
3107 method1
= skl_wm_method1(skl_pipe_pixel_rate(cstate
),
3110 method2
= skl_wm_method2(skl_pipe_pixel_rate(cstate
),
3111 cstate
->base
.adjusted_mode
.crtc_htotal
,
3117 plane_bytes_per_line
= cstate
->pipe_src_w
* bytes_per_pixel
;
3118 plane_blocks_per_line
= DIV_ROUND_UP(plane_bytes_per_line
, 512);
3120 if (fb
->modifier
[0] == I915_FORMAT_MOD_Y_TILED
||
3121 fb
->modifier
[0] == I915_FORMAT_MOD_Yf_TILED
) {
3122 uint32_t min_scanlines
= 4;
3123 uint32_t y_tile_minimum
;
3124 if (intel_rotation_90_or_270(plane
->state
->rotation
)) {
3125 int bpp
= (fb
->pixel_format
== DRM_FORMAT_NV12
) ?
3126 drm_format_plane_cpp(fb
->pixel_format
, 1) :
3127 drm_format_plane_cpp(fb
->pixel_format
, 0);
3137 WARN(1, "Unsupported pixel depth for rotation");
3140 y_tile_minimum
= plane_blocks_per_line
* min_scanlines
;
3141 selected_result
= max(method2
, y_tile_minimum
);
3143 if ((ddb_allocation
/ plane_blocks_per_line
) >= 1)
3144 selected_result
= min(method1
, method2
);
3146 selected_result
= method1
;
3149 res_blocks
= selected_result
+ 1;
3150 res_lines
= DIV_ROUND_UP(selected_result
, plane_blocks_per_line
);
3152 if (level
>= 1 && level
<= 7) {
3153 if (fb
->modifier
[0] == I915_FORMAT_MOD_Y_TILED
||
3154 fb
->modifier
[0] == I915_FORMAT_MOD_Yf_TILED
)
3160 if (res_blocks
>= ddb_allocation
|| res_lines
> 31)
3163 *out_blocks
= res_blocks
;
3164 *out_lines
= res_lines
;
3169 static void skl_compute_wm_level(const struct drm_i915_private
*dev_priv
,
3170 struct skl_ddb_allocation
*ddb
,
3171 struct intel_crtc_state
*cstate
,
3173 struct skl_wm_level
*result
)
3175 struct drm_device
*dev
= dev_priv
->dev
;
3176 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
3177 struct intel_plane
*intel_plane
;
3178 uint16_t ddb_blocks
;
3179 enum pipe pipe
= intel_crtc
->pipe
;
3181 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
3182 int i
= skl_wm_plane_id(intel_plane
);
3184 ddb_blocks
= skl_ddb_entry_size(&ddb
->plane
[pipe
][i
]);
3186 result
->plane_en
[i
] = skl_compute_plane_wm(dev_priv
,
3191 &result
->plane_res_b
[i
],
3192 &result
->plane_res_l
[i
]);
3197 skl_compute_linetime_wm(struct intel_crtc_state
*cstate
)
3199 if (!cstate
->base
.active
)
3202 if (WARN_ON(skl_pipe_pixel_rate(cstate
) == 0))
3205 return DIV_ROUND_UP(8 * cstate
->base
.adjusted_mode
.crtc_htotal
* 1000,
3206 skl_pipe_pixel_rate(cstate
));
3209 static void skl_compute_transition_wm(struct intel_crtc_state
*cstate
,
3210 struct skl_wm_level
*trans_wm
/* out */)
3212 struct drm_crtc
*crtc
= cstate
->base
.crtc
;
3213 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3214 struct intel_plane
*intel_plane
;
3216 if (!cstate
->base
.active
)
3219 /* Until we know more, just disable transition WMs */
3220 for_each_intel_plane_on_crtc(crtc
->dev
, intel_crtc
, intel_plane
) {
3221 int i
= skl_wm_plane_id(intel_plane
);
3223 trans_wm
->plane_en
[i
] = false;
3227 static void skl_compute_pipe_wm(struct intel_crtc_state
*cstate
,
3228 struct skl_ddb_allocation
*ddb
,
3229 struct skl_pipe_wm
*pipe_wm
)
3231 struct drm_device
*dev
= cstate
->base
.crtc
->dev
;
3232 const struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3233 int level
, max_level
= ilk_wm_max_level(dev
);
3235 for (level
= 0; level
<= max_level
; level
++) {
3236 skl_compute_wm_level(dev_priv
, ddb
, cstate
,
3237 level
, &pipe_wm
->wm
[level
]);
3239 pipe_wm
->linetime
= skl_compute_linetime_wm(cstate
);
3241 skl_compute_transition_wm(cstate
, &pipe_wm
->trans_wm
);
3244 static void skl_compute_wm_results(struct drm_device
*dev
,
3245 struct skl_pipe_wm
*p_wm
,
3246 struct skl_wm_values
*r
,
3247 struct intel_crtc
*intel_crtc
)
3249 int level
, max_level
= ilk_wm_max_level(dev
);
3250 enum pipe pipe
= intel_crtc
->pipe
;
3254 for (level
= 0; level
<= max_level
; level
++) {
3255 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++) {
3258 temp
|= p_wm
->wm
[level
].plane_res_l
[i
] <<
3259 PLANE_WM_LINES_SHIFT
;
3260 temp
|= p_wm
->wm
[level
].plane_res_b
[i
];
3261 if (p_wm
->wm
[level
].plane_en
[i
])
3262 temp
|= PLANE_WM_EN
;
3264 r
->plane
[pipe
][i
][level
] = temp
;
3269 temp
|= p_wm
->wm
[level
].plane_res_l
[PLANE_CURSOR
] << PLANE_WM_LINES_SHIFT
;
3270 temp
|= p_wm
->wm
[level
].plane_res_b
[PLANE_CURSOR
];
3272 if (p_wm
->wm
[level
].plane_en
[PLANE_CURSOR
])
3273 temp
|= PLANE_WM_EN
;
3275 r
->plane
[pipe
][PLANE_CURSOR
][level
] = temp
;
3279 /* transition WMs */
3280 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++) {
3282 temp
|= p_wm
->trans_wm
.plane_res_l
[i
] << PLANE_WM_LINES_SHIFT
;
3283 temp
|= p_wm
->trans_wm
.plane_res_b
[i
];
3284 if (p_wm
->trans_wm
.plane_en
[i
])
3285 temp
|= PLANE_WM_EN
;
3287 r
->plane_trans
[pipe
][i
] = temp
;
3291 temp
|= p_wm
->trans_wm
.plane_res_l
[PLANE_CURSOR
] << PLANE_WM_LINES_SHIFT
;
3292 temp
|= p_wm
->trans_wm
.plane_res_b
[PLANE_CURSOR
];
3293 if (p_wm
->trans_wm
.plane_en
[PLANE_CURSOR
])
3294 temp
|= PLANE_WM_EN
;
3296 r
->plane_trans
[pipe
][PLANE_CURSOR
] = temp
;
3298 r
->wm_linetime
[pipe
] = p_wm
->linetime
;
3301 static void skl_ddb_entry_write(struct drm_i915_private
*dev_priv
,
3303 const struct skl_ddb_entry
*entry
)
3306 I915_WRITE(reg
, (entry
->end
- 1) << 16 | entry
->start
);
3311 static void skl_write_wm_values(struct drm_i915_private
*dev_priv
,
3312 const struct skl_wm_values
*new)
3314 struct drm_device
*dev
= dev_priv
->dev
;
3315 struct intel_crtc
*crtc
;
3317 for_each_intel_crtc(dev
, crtc
) {
3318 int i
, level
, max_level
= ilk_wm_max_level(dev
);
3319 enum pipe pipe
= crtc
->pipe
;
3321 if (!new->dirty
[pipe
])
3324 I915_WRITE(PIPE_WM_LINETIME(pipe
), new->wm_linetime
[pipe
]);
3326 for (level
= 0; level
<= max_level
; level
++) {
3327 for (i
= 0; i
< intel_num_planes(crtc
); i
++)
3328 I915_WRITE(PLANE_WM(pipe
, i
, level
),
3329 new->plane
[pipe
][i
][level
]);
3330 I915_WRITE(CUR_WM(pipe
, level
),
3331 new->plane
[pipe
][PLANE_CURSOR
][level
]);
3333 for (i
= 0; i
< intel_num_planes(crtc
); i
++)
3334 I915_WRITE(PLANE_WM_TRANS(pipe
, i
),
3335 new->plane_trans
[pipe
][i
]);
3336 I915_WRITE(CUR_WM_TRANS(pipe
),
3337 new->plane_trans
[pipe
][PLANE_CURSOR
]);
3339 for (i
= 0; i
< intel_num_planes(crtc
); i
++) {
3340 skl_ddb_entry_write(dev_priv
,
3341 PLANE_BUF_CFG(pipe
, i
),
3342 &new->ddb
.plane
[pipe
][i
]);
3343 skl_ddb_entry_write(dev_priv
,
3344 PLANE_NV12_BUF_CFG(pipe
, i
),
3345 &new->ddb
.y_plane
[pipe
][i
]);
3348 skl_ddb_entry_write(dev_priv
, CUR_BUF_CFG(pipe
),
3349 &new->ddb
.plane
[pipe
][PLANE_CURSOR
]);
3354 * When setting up a new DDB allocation arrangement, we need to correctly
3355 * sequence the times at which the new allocations for the pipes are taken into
3356 * account or we'll have pipes fetching from space previously allocated to
3359 * Roughly the sequence looks like:
3360 * 1. re-allocate the pipe(s) with the allocation being reduced and not
3361 * overlapping with a previous light-up pipe (another way to put it is:
3362 * pipes with their new allocation strickly included into their old ones).
3363 * 2. re-allocate the other pipes that get their allocation reduced
3364 * 3. allocate the pipes having their allocation increased
3366 * Steps 1. and 2. are here to take care of the following case:
3367 * - Initially DDB looks like this:
3370 * - pipe B has a reduced DDB allocation that overlaps with the old pipe C
3374 * We need to sequence the re-allocation: C, B, A (and not B, C, A).
3378 skl_wm_flush_pipe(struct drm_i915_private
*dev_priv
, enum pipe pipe
, int pass
)
3382 DRM_DEBUG_KMS("flush pipe %c (pass %d)\n", pipe_name(pipe
), pass
);
3384 for_each_plane(dev_priv
, pipe
, plane
) {
3385 I915_WRITE(PLANE_SURF(pipe
, plane
),
3386 I915_READ(PLANE_SURF(pipe
, plane
)));
3388 I915_WRITE(CURBASE(pipe
), I915_READ(CURBASE(pipe
)));
3392 skl_ddb_allocation_included(const struct skl_ddb_allocation
*old
,
3393 const struct skl_ddb_allocation
*new,
3396 uint16_t old_size
, new_size
;
3398 old_size
= skl_ddb_entry_size(&old
->pipe
[pipe
]);
3399 new_size
= skl_ddb_entry_size(&new->pipe
[pipe
]);
3401 return old_size
!= new_size
&&
3402 new->pipe
[pipe
].start
>= old
->pipe
[pipe
].start
&&
3403 new->pipe
[pipe
].end
<= old
->pipe
[pipe
].end
;
3406 static void skl_flush_wm_values(struct drm_i915_private
*dev_priv
,
3407 struct skl_wm_values
*new_values
)
3409 struct drm_device
*dev
= dev_priv
->dev
;
3410 struct skl_ddb_allocation
*cur_ddb
, *new_ddb
;
3411 bool reallocated
[I915_MAX_PIPES
] = {};
3412 struct intel_crtc
*crtc
;
3415 new_ddb
= &new_values
->ddb
;
3416 cur_ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
3419 * First pass: flush the pipes with the new allocation contained into
3422 * We'll wait for the vblank on those pipes to ensure we can safely
3423 * re-allocate the freed space without this pipe fetching from it.
3425 for_each_intel_crtc(dev
, crtc
) {
3431 if (!skl_ddb_allocation_included(cur_ddb
, new_ddb
, pipe
))
3434 skl_wm_flush_pipe(dev_priv
, pipe
, 1);
3435 intel_wait_for_vblank(dev
, pipe
);
3437 reallocated
[pipe
] = true;
3442 * Second pass: flush the pipes that are having their allocation
3443 * reduced, but overlapping with a previous allocation.
3445 * Here as well we need to wait for the vblank to make sure the freed
3446 * space is not used anymore.
3448 for_each_intel_crtc(dev
, crtc
) {
3454 if (reallocated
[pipe
])
3457 if (skl_ddb_entry_size(&new_ddb
->pipe
[pipe
]) <
3458 skl_ddb_entry_size(&cur_ddb
->pipe
[pipe
])) {
3459 skl_wm_flush_pipe(dev_priv
, pipe
, 2);
3460 intel_wait_for_vblank(dev
, pipe
);
3461 reallocated
[pipe
] = true;
3466 * Third pass: flush the pipes that got more space allocated.
3468 * We don't need to actively wait for the update here, next vblank
3469 * will just get more DDB space with the correct WM values.
3471 for_each_intel_crtc(dev
, crtc
) {
3478 * At this point, only the pipes more space than before are
3479 * left to re-allocate.
3481 if (reallocated
[pipe
])
3484 skl_wm_flush_pipe(dev_priv
, pipe
, 3);
3488 static bool skl_update_pipe_wm(struct drm_crtc
*crtc
,
3489 struct skl_ddb_allocation
*ddb
, /* out */
3490 struct skl_pipe_wm
*pipe_wm
/* out */)
3492 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3493 struct intel_crtc_state
*cstate
= to_intel_crtc_state(crtc
->state
);
3495 skl_allocate_pipe_ddb(cstate
, ddb
);
3496 skl_compute_pipe_wm(cstate
, ddb
, pipe_wm
);
3498 if (!memcmp(&intel_crtc
->wm
.active
.skl
, pipe_wm
, sizeof(*pipe_wm
)))
3501 intel_crtc
->wm
.active
.skl
= *pipe_wm
;
3506 static void skl_update_other_pipe_wm(struct drm_device
*dev
,
3507 struct drm_crtc
*crtc
,
3508 struct skl_wm_values
*r
)
3510 struct intel_crtc
*intel_crtc
;
3511 struct intel_crtc
*this_crtc
= to_intel_crtc(crtc
);
3514 * If the WM update hasn't changed the allocation for this_crtc (the
3515 * crtc we are currently computing the new WM values for), other
3516 * enabled crtcs will keep the same allocation and we don't need to
3517 * recompute anything for them.
3519 if (!skl_ddb_allocation_changed(&r
->ddb
, this_crtc
))
3523 * Otherwise, because of this_crtc being freshly enabled/disabled, the
3524 * other active pipes need new DDB allocation and WM values.
3526 for_each_intel_crtc(dev
, intel_crtc
) {
3527 struct skl_pipe_wm pipe_wm
= {};
3530 if (this_crtc
->pipe
== intel_crtc
->pipe
)
3533 if (!intel_crtc
->active
)
3536 wm_changed
= skl_update_pipe_wm(&intel_crtc
->base
,
3540 * If we end up re-computing the other pipe WM values, it's
3541 * because it was really needed, so we expect the WM values to
3544 WARN_ON(!wm_changed
);
3546 skl_compute_wm_results(dev
, &pipe_wm
, r
, intel_crtc
);
3547 r
->dirty
[intel_crtc
->pipe
] = true;
3551 static void skl_clear_wm(struct skl_wm_values
*watermarks
, enum pipe pipe
)
3553 watermarks
->wm_linetime
[pipe
] = 0;
3554 memset(watermarks
->plane
[pipe
], 0,
3555 sizeof(uint32_t) * 8 * I915_MAX_PLANES
);
3556 memset(watermarks
->plane_trans
[pipe
],
3557 0, sizeof(uint32_t) * I915_MAX_PLANES
);
3558 watermarks
->plane_trans
[pipe
][PLANE_CURSOR
] = 0;
3560 /* Clear ddb entries for pipe */
3561 memset(&watermarks
->ddb
.pipe
[pipe
], 0, sizeof(struct skl_ddb_entry
));
3562 memset(&watermarks
->ddb
.plane
[pipe
], 0,
3563 sizeof(struct skl_ddb_entry
) * I915_MAX_PLANES
);
3564 memset(&watermarks
->ddb
.y_plane
[pipe
], 0,
3565 sizeof(struct skl_ddb_entry
) * I915_MAX_PLANES
);
3566 memset(&watermarks
->ddb
.plane
[pipe
][PLANE_CURSOR
], 0,
3567 sizeof(struct skl_ddb_entry
));
3571 static void skl_update_wm(struct drm_crtc
*crtc
)
3573 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3574 struct drm_device
*dev
= crtc
->dev
;
3575 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3576 struct skl_wm_values
*results
= &dev_priv
->wm
.skl_results
;
3577 struct intel_crtc_state
*cstate
= to_intel_crtc_state(crtc
->state
);
3578 struct skl_pipe_wm
*pipe_wm
= &cstate
->wm
.optimal
.skl
;
3581 /* Clear all dirty flags */
3582 memset(results
->dirty
, 0, sizeof(bool) * I915_MAX_PIPES
);
3584 skl_clear_wm(results
, intel_crtc
->pipe
);
3586 if (!skl_update_pipe_wm(crtc
, &results
->ddb
, pipe_wm
))
3589 skl_compute_wm_results(dev
, pipe_wm
, results
, intel_crtc
);
3590 results
->dirty
[intel_crtc
->pipe
] = true;
3592 skl_update_other_pipe_wm(dev
, crtc
, results
);
3593 skl_write_wm_values(dev_priv
, results
);
3594 skl_flush_wm_values(dev_priv
, results
);
3596 /* store the new configuration */
3597 dev_priv
->wm
.skl_hw
= *results
;
3600 static void ilk_program_watermarks(struct drm_i915_private
*dev_priv
)
3602 struct drm_device
*dev
= dev_priv
->dev
;
3603 struct intel_pipe_wm lp_wm_1_2
= {}, lp_wm_5_6
= {}, *best_lp_wm
;
3604 struct ilk_wm_maximums max
;
3605 struct intel_wm_config
*config
= &dev_priv
->wm
.config
;
3606 struct ilk_wm_values results
= {};
3607 enum intel_ddb_partitioning partitioning
;
3609 ilk_compute_wm_maximums(dev
, 1, config
, INTEL_DDB_PART_1_2
, &max
);
3610 ilk_wm_merge(dev
, config
, &max
, &lp_wm_1_2
);
3612 /* 5/6 split only in single pipe config on IVB+ */
3613 if (INTEL_INFO(dev
)->gen
>= 7 &&
3614 config
->num_pipes_active
== 1 && config
->sprites_enabled
) {
3615 ilk_compute_wm_maximums(dev
, 1, config
, INTEL_DDB_PART_5_6
, &max
);
3616 ilk_wm_merge(dev
, config
, &max
, &lp_wm_5_6
);
3618 best_lp_wm
= ilk_find_best_result(dev
, &lp_wm_1_2
, &lp_wm_5_6
);
3620 best_lp_wm
= &lp_wm_1_2
;
3623 partitioning
= (best_lp_wm
== &lp_wm_1_2
) ?
3624 INTEL_DDB_PART_1_2
: INTEL_DDB_PART_5_6
;
3626 ilk_compute_wm_results(dev
, best_lp_wm
, partitioning
, &results
);
3628 ilk_write_wm_values(dev_priv
, &results
);
3631 static void ilk_update_wm(struct drm_crtc
*crtc
)
3633 struct drm_i915_private
*dev_priv
= to_i915(crtc
->dev
);
3634 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3635 struct intel_crtc_state
*cstate
= to_intel_crtc_state(crtc
->state
);
3637 WARN_ON(cstate
->base
.active
!= intel_crtc
->active
);
3640 * IVB workaround: must disable low power watermarks for at least
3641 * one frame before enabling scaling. LP watermarks can be re-enabled
3642 * when scaling is disabled.
3644 * WaCxSRDisabledForSpriteScaling:ivb
3646 if (cstate
->disable_lp_wm
) {
3647 ilk_disable_lp_wm(crtc
->dev
);
3648 intel_wait_for_vblank(crtc
->dev
, intel_crtc
->pipe
);
3651 intel_crtc
->wm
.active
.ilk
= cstate
->wm
.optimal
.ilk
;
3653 ilk_program_watermarks(dev_priv
);
3656 static void skl_pipe_wm_active_state(uint32_t val
,
3657 struct skl_pipe_wm
*active
,
3663 bool is_enabled
= (val
& PLANE_WM_EN
) != 0;
3667 active
->wm
[level
].plane_en
[i
] = is_enabled
;
3668 active
->wm
[level
].plane_res_b
[i
] =
3669 val
& PLANE_WM_BLOCKS_MASK
;
3670 active
->wm
[level
].plane_res_l
[i
] =
3671 (val
>> PLANE_WM_LINES_SHIFT
) &
3672 PLANE_WM_LINES_MASK
;
3674 active
->wm
[level
].plane_en
[PLANE_CURSOR
] = is_enabled
;
3675 active
->wm
[level
].plane_res_b
[PLANE_CURSOR
] =
3676 val
& PLANE_WM_BLOCKS_MASK
;
3677 active
->wm
[level
].plane_res_l
[PLANE_CURSOR
] =
3678 (val
>> PLANE_WM_LINES_SHIFT
) &
3679 PLANE_WM_LINES_MASK
;
3683 active
->trans_wm
.plane_en
[i
] = is_enabled
;
3684 active
->trans_wm
.plane_res_b
[i
] =
3685 val
& PLANE_WM_BLOCKS_MASK
;
3686 active
->trans_wm
.plane_res_l
[i
] =
3687 (val
>> PLANE_WM_LINES_SHIFT
) &
3688 PLANE_WM_LINES_MASK
;
3690 active
->trans_wm
.plane_en
[PLANE_CURSOR
] = is_enabled
;
3691 active
->trans_wm
.plane_res_b
[PLANE_CURSOR
] =
3692 val
& PLANE_WM_BLOCKS_MASK
;
3693 active
->trans_wm
.plane_res_l
[PLANE_CURSOR
] =
3694 (val
>> PLANE_WM_LINES_SHIFT
) &
3695 PLANE_WM_LINES_MASK
;
3700 static void skl_pipe_wm_get_hw_state(struct drm_crtc
*crtc
)
3702 struct drm_device
*dev
= crtc
->dev
;
3703 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3704 struct skl_wm_values
*hw
= &dev_priv
->wm
.skl_hw
;
3705 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3706 struct intel_crtc_state
*cstate
= to_intel_crtc_state(crtc
->state
);
3707 struct skl_pipe_wm
*active
= &cstate
->wm
.optimal
.skl
;
3708 enum pipe pipe
= intel_crtc
->pipe
;
3709 int level
, i
, max_level
;
3712 max_level
= ilk_wm_max_level(dev
);
3714 hw
->wm_linetime
[pipe
] = I915_READ(PIPE_WM_LINETIME(pipe
));
3716 for (level
= 0; level
<= max_level
; level
++) {
3717 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++)
3718 hw
->plane
[pipe
][i
][level
] =
3719 I915_READ(PLANE_WM(pipe
, i
, level
));
3720 hw
->plane
[pipe
][PLANE_CURSOR
][level
] = I915_READ(CUR_WM(pipe
, level
));
3723 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++)
3724 hw
->plane_trans
[pipe
][i
] = I915_READ(PLANE_WM_TRANS(pipe
, i
));
3725 hw
->plane_trans
[pipe
][PLANE_CURSOR
] = I915_READ(CUR_WM_TRANS(pipe
));
3727 if (!intel_crtc
->active
)
3730 hw
->dirty
[pipe
] = true;
3732 active
->linetime
= hw
->wm_linetime
[pipe
];
3734 for (level
= 0; level
<= max_level
; level
++) {
3735 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++) {
3736 temp
= hw
->plane
[pipe
][i
][level
];
3737 skl_pipe_wm_active_state(temp
, active
, false,
3740 temp
= hw
->plane
[pipe
][PLANE_CURSOR
][level
];
3741 skl_pipe_wm_active_state(temp
, active
, false, true, i
, level
);
3744 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++) {
3745 temp
= hw
->plane_trans
[pipe
][i
];
3746 skl_pipe_wm_active_state(temp
, active
, true, false, i
, 0);
3749 temp
= hw
->plane_trans
[pipe
][PLANE_CURSOR
];
3750 skl_pipe_wm_active_state(temp
, active
, true, true, i
, 0);
3752 intel_crtc
->wm
.active
.skl
= *active
;
3755 void skl_wm_get_hw_state(struct drm_device
*dev
)
3757 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3758 struct skl_ddb_allocation
*ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
3759 struct drm_crtc
*crtc
;
3761 skl_ddb_get_hw_state(dev_priv
, ddb
);
3762 list_for_each_entry(crtc
, &dev
->mode_config
.crtc_list
, head
)
3763 skl_pipe_wm_get_hw_state(crtc
);
3766 static void ilk_pipe_wm_get_hw_state(struct drm_crtc
*crtc
)
3768 struct drm_device
*dev
= crtc
->dev
;
3769 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3770 struct ilk_wm_values
*hw
= &dev_priv
->wm
.hw
;
3771 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3772 struct intel_crtc_state
*cstate
= to_intel_crtc_state(crtc
->state
);
3773 struct intel_pipe_wm
*active
= &cstate
->wm
.optimal
.ilk
;
3774 enum pipe pipe
= intel_crtc
->pipe
;
3775 static const i915_reg_t wm0_pipe_reg
[] = {
3776 [PIPE_A
] = WM0_PIPEA_ILK
,
3777 [PIPE_B
] = WM0_PIPEB_ILK
,
3778 [PIPE_C
] = WM0_PIPEC_IVB
,
3781 hw
->wm_pipe
[pipe
] = I915_READ(wm0_pipe_reg
[pipe
]);
3782 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
3783 hw
->wm_linetime
[pipe
] = I915_READ(PIPE_WM_LINETIME(pipe
));
3785 active
->pipe_enabled
= intel_crtc
->active
;
3787 if (active
->pipe_enabled
) {
3788 u32 tmp
= hw
->wm_pipe
[pipe
];
3791 * For active pipes LP0 watermark is marked as
3792 * enabled, and LP1+ watermaks as disabled since
3793 * we can't really reverse compute them in case
3794 * multiple pipes are active.
3796 active
->wm
[0].enable
= true;
3797 active
->wm
[0].pri_val
= (tmp
& WM0_PIPE_PLANE_MASK
) >> WM0_PIPE_PLANE_SHIFT
;
3798 active
->wm
[0].spr_val
= (tmp
& WM0_PIPE_SPRITE_MASK
) >> WM0_PIPE_SPRITE_SHIFT
;
3799 active
->wm
[0].cur_val
= tmp
& WM0_PIPE_CURSOR_MASK
;
3800 active
->linetime
= hw
->wm_linetime
[pipe
];
3802 int level
, max_level
= ilk_wm_max_level(dev
);
3805 * For inactive pipes, all watermark levels
3806 * should be marked as enabled but zeroed,
3807 * which is what we'd compute them to.
3809 for (level
= 0; level
<= max_level
; level
++)
3810 active
->wm
[level
].enable
= true;
3813 intel_crtc
->wm
.active
.ilk
= *active
;
3816 #define _FW_WM(value, plane) \
3817 (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
3818 #define _FW_WM_VLV(value, plane) \
3819 (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
3821 static void vlv_read_wm_values(struct drm_i915_private
*dev_priv
,
3822 struct vlv_wm_values
*wm
)
3827 for_each_pipe(dev_priv
, pipe
) {
3828 tmp
= I915_READ(VLV_DDL(pipe
));
3830 wm
->ddl
[pipe
].primary
=
3831 (tmp
>> DDL_PLANE_SHIFT
) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
3832 wm
->ddl
[pipe
].cursor
=
3833 (tmp
>> DDL_CURSOR_SHIFT
) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
3834 wm
->ddl
[pipe
].sprite
[0] =
3835 (tmp
>> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
3836 wm
->ddl
[pipe
].sprite
[1] =
3837 (tmp
>> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
3840 tmp
= I915_READ(DSPFW1
);
3841 wm
->sr
.plane
= _FW_WM(tmp
, SR
);
3842 wm
->pipe
[PIPE_B
].cursor
= _FW_WM(tmp
, CURSORB
);
3843 wm
->pipe
[PIPE_B
].primary
= _FW_WM_VLV(tmp
, PLANEB
);
3844 wm
->pipe
[PIPE_A
].primary
= _FW_WM_VLV(tmp
, PLANEA
);
3846 tmp
= I915_READ(DSPFW2
);
3847 wm
->pipe
[PIPE_A
].sprite
[1] = _FW_WM_VLV(tmp
, SPRITEB
);
3848 wm
->pipe
[PIPE_A
].cursor
= _FW_WM(tmp
, CURSORA
);
3849 wm
->pipe
[PIPE_A
].sprite
[0] = _FW_WM_VLV(tmp
, SPRITEA
);
3851 tmp
= I915_READ(DSPFW3
);
3852 wm
->sr
.cursor
= _FW_WM(tmp
, CURSOR_SR
);
3854 if (IS_CHERRYVIEW(dev_priv
)) {
3855 tmp
= I915_READ(DSPFW7_CHV
);
3856 wm
->pipe
[PIPE_B
].sprite
[1] = _FW_WM_VLV(tmp
, SPRITED
);
3857 wm
->pipe
[PIPE_B
].sprite
[0] = _FW_WM_VLV(tmp
, SPRITEC
);
3859 tmp
= I915_READ(DSPFW8_CHV
);
3860 wm
->pipe
[PIPE_C
].sprite
[1] = _FW_WM_VLV(tmp
, SPRITEF
);
3861 wm
->pipe
[PIPE_C
].sprite
[0] = _FW_WM_VLV(tmp
, SPRITEE
);
3863 tmp
= I915_READ(DSPFW9_CHV
);
3864 wm
->pipe
[PIPE_C
].primary
= _FW_WM_VLV(tmp
, PLANEC
);
3865 wm
->pipe
[PIPE_C
].cursor
= _FW_WM(tmp
, CURSORC
);
3867 tmp
= I915_READ(DSPHOWM
);
3868 wm
->sr
.plane
|= _FW_WM(tmp
, SR_HI
) << 9;
3869 wm
->pipe
[PIPE_C
].sprite
[1] |= _FW_WM(tmp
, SPRITEF_HI
) << 8;
3870 wm
->pipe
[PIPE_C
].sprite
[0] |= _FW_WM(tmp
, SPRITEE_HI
) << 8;
3871 wm
->pipe
[PIPE_C
].primary
|= _FW_WM(tmp
, PLANEC_HI
) << 8;
3872 wm
->pipe
[PIPE_B
].sprite
[1] |= _FW_WM(tmp
, SPRITED_HI
) << 8;
3873 wm
->pipe
[PIPE_B
].sprite
[0] |= _FW_WM(tmp
, SPRITEC_HI
) << 8;
3874 wm
->pipe
[PIPE_B
].primary
|= _FW_WM(tmp
, PLANEB_HI
) << 8;
3875 wm
->pipe
[PIPE_A
].sprite
[1] |= _FW_WM(tmp
, SPRITEB_HI
) << 8;
3876 wm
->pipe
[PIPE_A
].sprite
[0] |= _FW_WM(tmp
, SPRITEA_HI
) << 8;
3877 wm
->pipe
[PIPE_A
].primary
|= _FW_WM(tmp
, PLANEA_HI
) << 8;
3879 tmp
= I915_READ(DSPFW7
);
3880 wm
->pipe
[PIPE_B
].sprite
[1] = _FW_WM_VLV(tmp
, SPRITED
);
3881 wm
->pipe
[PIPE_B
].sprite
[0] = _FW_WM_VLV(tmp
, SPRITEC
);
3883 tmp
= I915_READ(DSPHOWM
);
3884 wm
->sr
.plane
|= _FW_WM(tmp
, SR_HI
) << 9;
3885 wm
->pipe
[PIPE_B
].sprite
[1] |= _FW_WM(tmp
, SPRITED_HI
) << 8;
3886 wm
->pipe
[PIPE_B
].sprite
[0] |= _FW_WM(tmp
, SPRITEC_HI
) << 8;
3887 wm
->pipe
[PIPE_B
].primary
|= _FW_WM(tmp
, PLANEB_HI
) << 8;
3888 wm
->pipe
[PIPE_A
].sprite
[1] |= _FW_WM(tmp
, SPRITEB_HI
) << 8;
3889 wm
->pipe
[PIPE_A
].sprite
[0] |= _FW_WM(tmp
, SPRITEA_HI
) << 8;
3890 wm
->pipe
[PIPE_A
].primary
|= _FW_WM(tmp
, PLANEA_HI
) << 8;
3897 void vlv_wm_get_hw_state(struct drm_device
*dev
)
3899 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3900 struct vlv_wm_values
*wm
= &dev_priv
->wm
.vlv
;
3901 struct intel_plane
*plane
;
3905 vlv_read_wm_values(dev_priv
, wm
);
3907 for_each_intel_plane(dev
, plane
) {
3908 switch (plane
->base
.type
) {
3910 case DRM_PLANE_TYPE_CURSOR
:
3911 plane
->wm
.fifo_size
= 63;
3913 case DRM_PLANE_TYPE_PRIMARY
:
3914 plane
->wm
.fifo_size
= vlv_get_fifo_size(dev
, plane
->pipe
, 0);
3916 case DRM_PLANE_TYPE_OVERLAY
:
3917 sprite
= plane
->plane
;
3918 plane
->wm
.fifo_size
= vlv_get_fifo_size(dev
, plane
->pipe
, sprite
+ 1);
3923 wm
->cxsr
= I915_READ(FW_BLC_SELF_VLV
) & FW_CSPWRDWNEN
;
3924 wm
->level
= VLV_WM_LEVEL_PM2
;
3926 if (IS_CHERRYVIEW(dev_priv
)) {
3927 mutex_lock(&dev_priv
->rps
.hw_lock
);
3929 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DSPFREQ
);
3930 if (val
& DSP_MAXFIFO_PM5_ENABLE
)
3931 wm
->level
= VLV_WM_LEVEL_PM5
;
3934 * If DDR DVFS is disabled in the BIOS, Punit
3935 * will never ack the request. So if that happens
3936 * assume we don't have to enable/disable DDR DVFS
3937 * dynamically. To test that just set the REQ_ACK
3938 * bit to poke the Punit, but don't change the
3939 * HIGH/LOW bits so that we don't actually change
3940 * the current state.
3942 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
);
3943 val
|= FORCE_DDR_FREQ_REQ_ACK
;
3944 vlv_punit_write(dev_priv
, PUNIT_REG_DDR_SETUP2
, val
);
3946 if (wait_for((vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
) &
3947 FORCE_DDR_FREQ_REQ_ACK
) == 0, 3)) {
3948 DRM_DEBUG_KMS("Punit not acking DDR DVFS request, "
3949 "assuming DDR DVFS is disabled\n");
3950 dev_priv
->wm
.max_level
= VLV_WM_LEVEL_PM5
;
3952 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
);
3953 if ((val
& FORCE_DDR_HIGH_FREQ
) == 0)
3954 wm
->level
= VLV_WM_LEVEL_DDR_DVFS
;
3957 mutex_unlock(&dev_priv
->rps
.hw_lock
);
3960 for_each_pipe(dev_priv
, pipe
)
3961 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
3962 pipe_name(pipe
), wm
->pipe
[pipe
].primary
, wm
->pipe
[pipe
].cursor
,
3963 wm
->pipe
[pipe
].sprite
[0], wm
->pipe
[pipe
].sprite
[1]);
3965 DRM_DEBUG_KMS("Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
3966 wm
->sr
.plane
, wm
->sr
.cursor
, wm
->level
, wm
->cxsr
);
3969 void ilk_wm_get_hw_state(struct drm_device
*dev
)
3971 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3972 struct ilk_wm_values
*hw
= &dev_priv
->wm
.hw
;
3973 struct drm_crtc
*crtc
;
3975 for_each_crtc(dev
, crtc
)
3976 ilk_pipe_wm_get_hw_state(crtc
);
3978 hw
->wm_lp
[0] = I915_READ(WM1_LP_ILK
);
3979 hw
->wm_lp
[1] = I915_READ(WM2_LP_ILK
);
3980 hw
->wm_lp
[2] = I915_READ(WM3_LP_ILK
);
3982 hw
->wm_lp_spr
[0] = I915_READ(WM1S_LP_ILK
);
3983 if (INTEL_INFO(dev
)->gen
>= 7) {
3984 hw
->wm_lp_spr
[1] = I915_READ(WM2S_LP_IVB
);
3985 hw
->wm_lp_spr
[2] = I915_READ(WM3S_LP_IVB
);
3988 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
3989 hw
->partitioning
= (I915_READ(WM_MISC
) & WM_MISC_DATA_PARTITION_5_6
) ?
3990 INTEL_DDB_PART_5_6
: INTEL_DDB_PART_1_2
;
3991 else if (IS_IVYBRIDGE(dev
))
3992 hw
->partitioning
= (I915_READ(DISP_ARB_CTL2
) & DISP_DATA_PARTITION_5_6
) ?
3993 INTEL_DDB_PART_5_6
: INTEL_DDB_PART_1_2
;
3996 !(I915_READ(DISP_ARB_CTL
) & DISP_FBC_WM_DIS
);
4000 * intel_update_watermarks - update FIFO watermark values based on current modes
4002 * Calculate watermark values for the various WM regs based on current mode
4003 * and plane configuration.
4005 * There are several cases to deal with here:
4006 * - normal (i.e. non-self-refresh)
4007 * - self-refresh (SR) mode
4008 * - lines are large relative to FIFO size (buffer can hold up to 2)
4009 * - lines are small relative to FIFO size (buffer can hold more than 2
4010 * lines), so need to account for TLB latency
4012 * The normal calculation is:
4013 * watermark = dotclock * bytes per pixel * latency
4014 * where latency is platform & configuration dependent (we assume pessimal
4017 * The SR calculation is:
4018 * watermark = (trunc(latency/line time)+1) * surface width *
4021 * line time = htotal / dotclock
4022 * surface width = hdisplay for normal plane and 64 for cursor
4023 * and latency is assumed to be high, as above.
4025 * The final value programmed to the register should always be rounded up,
4026 * and include an extra 2 entries to account for clock crossings.
4028 * We don't use the sprite, so we can ignore that. And on Crestline we have
4029 * to set the non-SR watermarks to 8.
4031 void intel_update_watermarks(struct drm_crtc
*crtc
)
4033 struct drm_i915_private
*dev_priv
= crtc
->dev
->dev_private
;
4035 if (dev_priv
->display
.update_wm
)
4036 dev_priv
->display
.update_wm(crtc
);
4040 * Lock protecting IPS related data structures
4042 DEFINE_SPINLOCK(mchdev_lock
);
4044 /* Global for IPS driver to get at the current i915 device. Protected by
4046 static struct drm_i915_private
*i915_mch_dev
;
4048 bool ironlake_set_drps(struct drm_device
*dev
, u8 val
)
4050 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4053 assert_spin_locked(&mchdev_lock
);
4055 rgvswctl
= I915_READ16(MEMSWCTL
);
4056 if (rgvswctl
& MEMCTL_CMD_STS
) {
4057 DRM_DEBUG("gpu busy, RCS change rejected\n");
4058 return false; /* still busy with another command */
4061 rgvswctl
= (MEMCTL_CMD_CHFREQ
<< MEMCTL_CMD_SHIFT
) |
4062 (val
<< MEMCTL_FREQ_SHIFT
) | MEMCTL_SFCAVM
;
4063 I915_WRITE16(MEMSWCTL
, rgvswctl
);
4064 POSTING_READ16(MEMSWCTL
);
4066 rgvswctl
|= MEMCTL_CMD_STS
;
4067 I915_WRITE16(MEMSWCTL
, rgvswctl
);
4072 static void ironlake_enable_drps(struct drm_device
*dev
)
4074 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4075 u32 rgvmodectl
= I915_READ(MEMMODECTL
);
4076 u8 fmax
, fmin
, fstart
, vstart
;
4078 spin_lock_irq(&mchdev_lock
);
4080 /* Enable temp reporting */
4081 I915_WRITE16(PMMISC
, I915_READ(PMMISC
) | MCPPCE_EN
);
4082 I915_WRITE16(TSC1
, I915_READ(TSC1
) | TSE
);
4084 /* 100ms RC evaluation intervals */
4085 I915_WRITE(RCUPEI
, 100000);
4086 I915_WRITE(RCDNEI
, 100000);
4088 /* Set max/min thresholds to 90ms and 80ms respectively */
4089 I915_WRITE(RCBMAXAVG
, 90000);
4090 I915_WRITE(RCBMINAVG
, 80000);
4092 I915_WRITE(MEMIHYST
, 1);
4094 /* Set up min, max, and cur for interrupt handling */
4095 fmax
= (rgvmodectl
& MEMMODE_FMAX_MASK
) >> MEMMODE_FMAX_SHIFT
;
4096 fmin
= (rgvmodectl
& MEMMODE_FMIN_MASK
);
4097 fstart
= (rgvmodectl
& MEMMODE_FSTART_MASK
) >>
4098 MEMMODE_FSTART_SHIFT
;
4100 vstart
= (I915_READ(PXVFREQ(fstart
)) & PXVFREQ_PX_MASK
) >>
4103 dev_priv
->ips
.fmax
= fmax
; /* IPS callback will increase this */
4104 dev_priv
->ips
.fstart
= fstart
;
4106 dev_priv
->ips
.max_delay
= fstart
;
4107 dev_priv
->ips
.min_delay
= fmin
;
4108 dev_priv
->ips
.cur_delay
= fstart
;
4110 DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
4111 fmax
, fmin
, fstart
);
4113 I915_WRITE(MEMINTREN
, MEMINT_CX_SUPR_EN
| MEMINT_EVAL_CHG_EN
);
4116 * Interrupts will be enabled in ironlake_irq_postinstall
4119 I915_WRITE(VIDSTART
, vstart
);
4120 POSTING_READ(VIDSTART
);
4122 rgvmodectl
|= MEMMODE_SWMODE_EN
;
4123 I915_WRITE(MEMMODECTL
, rgvmodectl
);
4125 if (wait_for_atomic((I915_READ(MEMSWCTL
) & MEMCTL_CMD_STS
) == 0, 10))
4126 DRM_ERROR("stuck trying to change perf mode\n");
4129 ironlake_set_drps(dev
, fstart
);
4131 dev_priv
->ips
.last_count1
= I915_READ(DMIEC
) +
4132 I915_READ(DDREC
) + I915_READ(CSIEC
);
4133 dev_priv
->ips
.last_time1
= jiffies_to_msecs(jiffies
);
4134 dev_priv
->ips
.last_count2
= I915_READ(GFXEC
);
4135 dev_priv
->ips
.last_time2
= ktime_get_raw_ns();
4137 spin_unlock_irq(&mchdev_lock
);
4140 static void ironlake_disable_drps(struct drm_device
*dev
)
4142 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4145 spin_lock_irq(&mchdev_lock
);
4147 rgvswctl
= I915_READ16(MEMSWCTL
);
4149 /* Ack interrupts, disable EFC interrupt */
4150 I915_WRITE(MEMINTREN
, I915_READ(MEMINTREN
) & ~MEMINT_EVAL_CHG_EN
);
4151 I915_WRITE(MEMINTRSTS
, MEMINT_EVAL_CHG
);
4152 I915_WRITE(DEIER
, I915_READ(DEIER
) & ~DE_PCU_EVENT
);
4153 I915_WRITE(DEIIR
, DE_PCU_EVENT
);
4154 I915_WRITE(DEIMR
, I915_READ(DEIMR
) | DE_PCU_EVENT
);
4156 /* Go back to the starting frequency */
4157 ironlake_set_drps(dev
, dev_priv
->ips
.fstart
);
4159 rgvswctl
|= MEMCTL_CMD_STS
;
4160 I915_WRITE(MEMSWCTL
, rgvswctl
);
4163 spin_unlock_irq(&mchdev_lock
);
4166 /* There's a funny hw issue where the hw returns all 0 when reading from
4167 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
4168 * ourselves, instead of doing a rmw cycle (which might result in us clearing
4169 * all limits and the gpu stuck at whatever frequency it is at atm).
4171 static u32
intel_rps_limits(struct drm_i915_private
*dev_priv
, u8 val
)
4175 /* Only set the down limit when we've reached the lowest level to avoid
4176 * getting more interrupts, otherwise leave this clear. This prevents a
4177 * race in the hw when coming out of rc6: There's a tiny window where
4178 * the hw runs at the minimal clock before selecting the desired
4179 * frequency, if the down threshold expires in that window we will not
4180 * receive a down interrupt. */
4181 if (IS_GEN9(dev_priv
->dev
)) {
4182 limits
= (dev_priv
->rps
.max_freq_softlimit
) << 23;
4183 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
4184 limits
|= (dev_priv
->rps
.min_freq_softlimit
) << 14;
4186 limits
= dev_priv
->rps
.max_freq_softlimit
<< 24;
4187 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
4188 limits
|= dev_priv
->rps
.min_freq_softlimit
<< 16;
4194 static void gen6_set_rps_thresholds(struct drm_i915_private
*dev_priv
, u8 val
)
4197 u32 threshold_up
= 0, threshold_down
= 0; /* in % */
4198 u32 ei_up
= 0, ei_down
= 0;
4200 new_power
= dev_priv
->rps
.power
;
4201 switch (dev_priv
->rps
.power
) {
4203 if (val
> dev_priv
->rps
.efficient_freq
+ 1 && val
> dev_priv
->rps
.cur_freq
)
4204 new_power
= BETWEEN
;
4208 if (val
<= dev_priv
->rps
.efficient_freq
&& val
< dev_priv
->rps
.cur_freq
)
4209 new_power
= LOW_POWER
;
4210 else if (val
>= dev_priv
->rps
.rp0_freq
&& val
> dev_priv
->rps
.cur_freq
)
4211 new_power
= HIGH_POWER
;
4215 if (val
< (dev_priv
->rps
.rp1_freq
+ dev_priv
->rps
.rp0_freq
) >> 1 && val
< dev_priv
->rps
.cur_freq
)
4216 new_power
= BETWEEN
;
4219 /* Max/min bins are special */
4220 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
4221 new_power
= LOW_POWER
;
4222 if (val
>= dev_priv
->rps
.max_freq_softlimit
)
4223 new_power
= HIGH_POWER
;
4224 if (new_power
== dev_priv
->rps
.power
)
4227 /* Note the units here are not exactly 1us, but 1280ns. */
4228 switch (new_power
) {
4230 /* Upclock if more than 95% busy over 16ms */
4234 /* Downclock if less than 85% busy over 32ms */
4236 threshold_down
= 85;
4240 /* Upclock if more than 90% busy over 13ms */
4244 /* Downclock if less than 75% busy over 32ms */
4246 threshold_down
= 75;
4250 /* Upclock if more than 85% busy over 10ms */
4254 /* Downclock if less than 60% busy over 32ms */
4256 threshold_down
= 60;
4260 I915_WRITE(GEN6_RP_UP_EI
,
4261 GT_INTERVAL_FROM_US(dev_priv
, ei_up
));
4262 I915_WRITE(GEN6_RP_UP_THRESHOLD
,
4263 GT_INTERVAL_FROM_US(dev_priv
, (ei_up
* threshold_up
/ 100)));
4265 I915_WRITE(GEN6_RP_DOWN_EI
,
4266 GT_INTERVAL_FROM_US(dev_priv
, ei_down
));
4267 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
,
4268 GT_INTERVAL_FROM_US(dev_priv
, (ei_down
* threshold_down
/ 100)));
4270 I915_WRITE(GEN6_RP_CONTROL
,
4271 GEN6_RP_MEDIA_TURBO
|
4272 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
4273 GEN6_RP_MEDIA_IS_GFX
|
4275 GEN6_RP_UP_BUSY_AVG
|
4276 GEN6_RP_DOWN_IDLE_AVG
);
4278 dev_priv
->rps
.power
= new_power
;
4279 dev_priv
->rps
.up_threshold
= threshold_up
;
4280 dev_priv
->rps
.down_threshold
= threshold_down
;
4281 dev_priv
->rps
.last_adj
= 0;
4284 static u32
gen6_rps_pm_mask(struct drm_i915_private
*dev_priv
, u8 val
)
4288 if (val
> dev_priv
->rps
.min_freq_softlimit
)
4289 mask
|= GEN6_PM_RP_DOWN_EI_EXPIRED
| GEN6_PM_RP_DOWN_THRESHOLD
| GEN6_PM_RP_DOWN_TIMEOUT
;
4290 if (val
< dev_priv
->rps
.max_freq_softlimit
)
4291 mask
|= GEN6_PM_RP_UP_EI_EXPIRED
| GEN6_PM_RP_UP_THRESHOLD
;
4293 mask
&= dev_priv
->pm_rps_events
;
4295 return gen6_sanitize_rps_pm_mask(dev_priv
, ~mask
);
4298 /* gen6_set_rps is called to update the frequency request, but should also be
4299 * called when the range (min_delay and max_delay) is modified so that we can
4300 * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
4301 static void gen6_set_rps(struct drm_device
*dev
, u8 val
)
4303 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4305 /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
4306 if (IS_BXT_REVID(dev
, 0, BXT_REVID_A1
))
4309 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
4310 WARN_ON(val
> dev_priv
->rps
.max_freq
);
4311 WARN_ON(val
< dev_priv
->rps
.min_freq
);
4313 /* min/max delay may still have been modified so be sure to
4314 * write the limits value.
4316 if (val
!= dev_priv
->rps
.cur_freq
) {
4317 gen6_set_rps_thresholds(dev_priv
, val
);
4320 I915_WRITE(GEN6_RPNSWREQ
,
4321 GEN9_FREQUENCY(val
));
4322 else if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
4323 I915_WRITE(GEN6_RPNSWREQ
,
4324 HSW_FREQUENCY(val
));
4326 I915_WRITE(GEN6_RPNSWREQ
,
4327 GEN6_FREQUENCY(val
) |
4329 GEN6_AGGRESSIVE_TURBO
);
4332 /* Make sure we continue to get interrupts
4333 * until we hit the minimum or maximum frequencies.
4335 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS
, intel_rps_limits(dev_priv
, val
));
4336 I915_WRITE(GEN6_PMINTRMSK
, gen6_rps_pm_mask(dev_priv
, val
));
4338 POSTING_READ(GEN6_RPNSWREQ
);
4340 dev_priv
->rps
.cur_freq
= val
;
4341 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv
, val
));
4344 static void valleyview_set_rps(struct drm_device
*dev
, u8 val
)
4346 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4348 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
4349 WARN_ON(val
> dev_priv
->rps
.max_freq
);
4350 WARN_ON(val
< dev_priv
->rps
.min_freq
);
4352 if (WARN_ONCE(IS_CHERRYVIEW(dev
) && (val
& 1),
4353 "Odd GPU freq value\n"))
4356 I915_WRITE(GEN6_PMINTRMSK
, gen6_rps_pm_mask(dev_priv
, val
));
4358 if (val
!= dev_priv
->rps
.cur_freq
) {
4359 vlv_punit_write(dev_priv
, PUNIT_REG_GPU_FREQ_REQ
, val
);
4360 if (!IS_CHERRYVIEW(dev_priv
))
4361 gen6_set_rps_thresholds(dev_priv
, val
);
4364 dev_priv
->rps
.cur_freq
= val
;
4365 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv
, val
));
4368 /* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
4370 * * If Gfx is Idle, then
4371 * 1. Forcewake Media well.
4372 * 2. Request idle freq.
4373 * 3. Release Forcewake of Media well.
4375 static void vlv_set_rps_idle(struct drm_i915_private
*dev_priv
)
4377 u32 val
= dev_priv
->rps
.idle_freq
;
4379 if (dev_priv
->rps
.cur_freq
<= val
)
4382 /* Wake up the media well, as that takes a lot less
4383 * power than the Render well. */
4384 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_MEDIA
);
4385 valleyview_set_rps(dev_priv
->dev
, val
);
4386 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_MEDIA
);
4389 void gen6_rps_busy(struct drm_i915_private
*dev_priv
)
4391 mutex_lock(&dev_priv
->rps
.hw_lock
);
4392 if (dev_priv
->rps
.enabled
) {
4393 if (dev_priv
->pm_rps_events
& (GEN6_PM_RP_DOWN_EI_EXPIRED
| GEN6_PM_RP_UP_EI_EXPIRED
))
4394 gen6_rps_reset_ei(dev_priv
);
4395 I915_WRITE(GEN6_PMINTRMSK
,
4396 gen6_rps_pm_mask(dev_priv
, dev_priv
->rps
.cur_freq
));
4398 mutex_unlock(&dev_priv
->rps
.hw_lock
);
4401 void gen6_rps_idle(struct drm_i915_private
*dev_priv
)
4403 struct drm_device
*dev
= dev_priv
->dev
;
4405 mutex_lock(&dev_priv
->rps
.hw_lock
);
4406 if (dev_priv
->rps
.enabled
) {
4407 if (IS_VALLEYVIEW(dev
) || IS_CHERRYVIEW(dev
))
4408 vlv_set_rps_idle(dev_priv
);
4410 gen6_set_rps(dev_priv
->dev
, dev_priv
->rps
.idle_freq
);
4411 dev_priv
->rps
.last_adj
= 0;
4412 I915_WRITE(GEN6_PMINTRMSK
, 0xffffffff);
4414 mutex_unlock(&dev_priv
->rps
.hw_lock
);
4416 spin_lock(&dev_priv
->rps
.client_lock
);
4417 while (!list_empty(&dev_priv
->rps
.clients
))
4418 list_del_init(dev_priv
->rps
.clients
.next
);
4419 spin_unlock(&dev_priv
->rps
.client_lock
);
4422 void gen6_rps_boost(struct drm_i915_private
*dev_priv
,
4423 struct intel_rps_client
*rps
,
4424 unsigned long submitted
)
4426 /* This is intentionally racy! We peek at the state here, then
4427 * validate inside the RPS worker.
4429 if (!(dev_priv
->mm
.busy
&&
4430 dev_priv
->rps
.enabled
&&
4431 dev_priv
->rps
.cur_freq
< dev_priv
->rps
.max_freq_softlimit
))
4434 /* Force a RPS boost (and don't count it against the client) if
4435 * the GPU is severely congested.
4437 if (rps
&& time_after(jiffies
, submitted
+ DRM_I915_THROTTLE_JIFFIES
))
4440 spin_lock(&dev_priv
->rps
.client_lock
);
4441 if (rps
== NULL
|| list_empty(&rps
->link
)) {
4442 spin_lock_irq(&dev_priv
->irq_lock
);
4443 if (dev_priv
->rps
.interrupts_enabled
) {
4444 dev_priv
->rps
.client_boost
= true;
4445 queue_work(dev_priv
->wq
, &dev_priv
->rps
.work
);
4447 spin_unlock_irq(&dev_priv
->irq_lock
);
4450 list_add(&rps
->link
, &dev_priv
->rps
.clients
);
4453 dev_priv
->rps
.boosts
++;
4455 spin_unlock(&dev_priv
->rps
.client_lock
);
4458 void intel_set_rps(struct drm_device
*dev
, u8 val
)
4460 if (IS_VALLEYVIEW(dev
) || IS_CHERRYVIEW(dev
))
4461 valleyview_set_rps(dev
, val
);
4463 gen6_set_rps(dev
, val
);
4466 static void gen9_disable_rps(struct drm_device
*dev
)
4468 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4470 I915_WRITE(GEN6_RC_CONTROL
, 0);
4471 I915_WRITE(GEN9_PG_ENABLE
, 0);
4474 static void gen6_disable_rps(struct drm_device
*dev
)
4476 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4478 I915_WRITE(GEN6_RC_CONTROL
, 0);
4479 I915_WRITE(GEN6_RPNSWREQ
, 1 << 31);
4482 static void cherryview_disable_rps(struct drm_device
*dev
)
4484 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4486 I915_WRITE(GEN6_RC_CONTROL
, 0);
4489 static void valleyview_disable_rps(struct drm_device
*dev
)
4491 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4493 /* we're doing forcewake before Disabling RC6,
4494 * This what the BIOS expects when going into suspend */
4495 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4497 I915_WRITE(GEN6_RC_CONTROL
, 0);
4499 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4502 static void intel_print_rc6_info(struct drm_device
*dev
, u32 mode
)
4504 if (IS_VALLEYVIEW(dev
) || IS_CHERRYVIEW(dev
)) {
4505 if (mode
& (GEN7_RC_CTL_TO_MODE
| GEN6_RC_CTL_EI_MODE(1)))
4506 mode
= GEN6_RC_CTL_RC6_ENABLE
;
4511 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s RC6p %s RC6pp %s\n",
4512 (mode
& GEN6_RC_CTL_RC6_ENABLE
) ? "on" : "off",
4513 (mode
& GEN6_RC_CTL_RC6p_ENABLE
) ? "on" : "off",
4514 (mode
& GEN6_RC_CTL_RC6pp_ENABLE
) ? "on" : "off");
4517 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s\n",
4518 (mode
& GEN6_RC_CTL_RC6_ENABLE
) ? "on" : "off");
4521 static int sanitize_rc6_option(const struct drm_device
*dev
, int enable_rc6
)
4523 /* No RC6 before Ironlake and code is gone for ilk. */
4524 if (INTEL_INFO(dev
)->gen
< 6)
4527 /* Respect the kernel parameter if it is set */
4528 if (enable_rc6
>= 0) {
4532 mask
= INTEL_RC6_ENABLE
| INTEL_RC6p_ENABLE
|
4535 mask
= INTEL_RC6_ENABLE
;
4537 if ((enable_rc6
& mask
) != enable_rc6
)
4538 DRM_DEBUG_KMS("Adjusting RC6 mask to %d (requested %d, valid %d)\n",
4539 enable_rc6
& mask
, enable_rc6
, mask
);
4541 return enable_rc6
& mask
;
4544 if (IS_IVYBRIDGE(dev
))
4545 return (INTEL_RC6_ENABLE
| INTEL_RC6p_ENABLE
);
4547 return INTEL_RC6_ENABLE
;
4550 int intel_enable_rc6(const struct drm_device
*dev
)
4552 return i915
.enable_rc6
;
4555 static void gen6_init_rps_frequencies(struct drm_device
*dev
)
4557 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4558 uint32_t rp_state_cap
;
4559 u32 ddcc_status
= 0;
4562 /* All of these values are in units of 50MHz */
4563 dev_priv
->rps
.cur_freq
= 0;
4564 /* static values from HW: RP0 > RP1 > RPn (min_freq) */
4565 if (IS_BROXTON(dev
)) {
4566 rp_state_cap
= I915_READ(BXT_RP_STATE_CAP
);
4567 dev_priv
->rps
.rp0_freq
= (rp_state_cap
>> 16) & 0xff;
4568 dev_priv
->rps
.rp1_freq
= (rp_state_cap
>> 8) & 0xff;
4569 dev_priv
->rps
.min_freq
= (rp_state_cap
>> 0) & 0xff;
4571 rp_state_cap
= I915_READ(GEN6_RP_STATE_CAP
);
4572 dev_priv
->rps
.rp0_freq
= (rp_state_cap
>> 0) & 0xff;
4573 dev_priv
->rps
.rp1_freq
= (rp_state_cap
>> 8) & 0xff;
4574 dev_priv
->rps
.min_freq
= (rp_state_cap
>> 16) & 0xff;
4577 /* hw_max = RP0 until we check for overclocking */
4578 dev_priv
->rps
.max_freq
= dev_priv
->rps
.rp0_freq
;
4580 dev_priv
->rps
.efficient_freq
= dev_priv
->rps
.rp1_freq
;
4581 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
) ||
4582 IS_SKYLAKE(dev
) || IS_KABYLAKE(dev
)) {
4583 ret
= sandybridge_pcode_read(dev_priv
,
4584 HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL
,
4587 dev_priv
->rps
.efficient_freq
=
4589 ((ddcc_status
>> 8) & 0xff),
4590 dev_priv
->rps
.min_freq
,
4591 dev_priv
->rps
.max_freq
);
4594 if (IS_SKYLAKE(dev
) || IS_KABYLAKE(dev
)) {
4595 /* Store the frequency values in 16.66 MHZ units, which is
4596 the natural hardware unit for SKL */
4597 dev_priv
->rps
.rp0_freq
*= GEN9_FREQ_SCALER
;
4598 dev_priv
->rps
.rp1_freq
*= GEN9_FREQ_SCALER
;
4599 dev_priv
->rps
.min_freq
*= GEN9_FREQ_SCALER
;
4600 dev_priv
->rps
.max_freq
*= GEN9_FREQ_SCALER
;
4601 dev_priv
->rps
.efficient_freq
*= GEN9_FREQ_SCALER
;
4604 dev_priv
->rps
.idle_freq
= dev_priv
->rps
.min_freq
;
4606 /* Preserve min/max settings in case of re-init */
4607 if (dev_priv
->rps
.max_freq_softlimit
== 0)
4608 dev_priv
->rps
.max_freq_softlimit
= dev_priv
->rps
.max_freq
;
4610 if (dev_priv
->rps
.min_freq_softlimit
== 0) {
4611 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
4612 dev_priv
->rps
.min_freq_softlimit
=
4613 max_t(int, dev_priv
->rps
.efficient_freq
,
4614 intel_freq_opcode(dev_priv
, 450));
4616 dev_priv
->rps
.min_freq_softlimit
=
4617 dev_priv
->rps
.min_freq
;
4621 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
4622 static void gen9_enable_rps(struct drm_device
*dev
)
4624 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4626 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4628 gen6_init_rps_frequencies(dev
);
4630 /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
4631 if (IS_BXT_REVID(dev
, 0, BXT_REVID_A1
)) {
4632 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4636 /* Program defaults and thresholds for RPS*/
4637 I915_WRITE(GEN6_RC_VIDEO_FREQ
,
4638 GEN9_FREQUENCY(dev_priv
->rps
.rp1_freq
));
4640 /* 1 second timeout*/
4641 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
,
4642 GT_INTERVAL_FROM_US(dev_priv
, 1000000));
4644 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 0xa);
4646 /* Leaning on the below call to gen6_set_rps to program/setup the
4647 * Up/Down EI & threshold registers, as well as the RP_CONTROL,
4648 * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
4649 dev_priv
->rps
.power
= HIGH_POWER
; /* force a reset */
4650 gen6_set_rps(dev_priv
->dev
, dev_priv
->rps
.min_freq_softlimit
);
4652 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4655 static void gen9_enable_rc6(struct drm_device
*dev
)
4657 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4658 struct intel_engine_cs
*ring
;
4659 uint32_t rc6_mask
= 0;
4662 /* 1a: Software RC state - RC0 */
4663 I915_WRITE(GEN6_RC_STATE
, 0);
4665 /* 1b: Get forcewake during program sequence. Although the driver
4666 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4667 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4669 /* 2a: Disable RC states. */
4670 I915_WRITE(GEN6_RC_CONTROL
, 0);
4672 /* 2b: Program RC6 thresholds.*/
4674 /* WaRsDoubleRc6WrlWithCoarsePowerGating: Doubling WRL only when CPG is enabled */
4675 if (IS_SKYLAKE(dev
))
4676 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 108 << 16);
4678 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 54 << 16);
4679 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
4680 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
4681 for_each_ring(ring
, dev_priv
, unused
)
4682 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
4684 if (HAS_GUC_UCODE(dev
))
4685 I915_WRITE(GUC_MAX_IDLE_COUNT
, 0xA);
4687 I915_WRITE(GEN6_RC_SLEEP
, 0);
4689 /* 2c: Program Coarse Power Gating Policies. */
4690 I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS
, 25);
4691 I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS
, 25);
4693 /* 3a: Enable RC6 */
4694 if (intel_enable_rc6(dev
) & INTEL_RC6_ENABLE
)
4695 rc6_mask
= GEN6_RC_CTL_RC6_ENABLE
;
4696 DRM_INFO("RC6 %s\n", (rc6_mask
& GEN6_RC_CTL_RC6_ENABLE
) ?
4698 /* WaRsUseTimeoutMode */
4699 if (IS_SKL_REVID(dev
, 0, SKL_REVID_D0
) ||
4700 IS_BXT_REVID(dev
, 0, BXT_REVID_A1
)) {
4701 I915_WRITE(GEN6_RC6_THRESHOLD
, 625); /* 800us */
4702 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
4703 GEN7_RC_CTL_TO_MODE
|
4706 I915_WRITE(GEN6_RC6_THRESHOLD
, 37500); /* 37.5/125ms per EI */
4707 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
4708 GEN6_RC_CTL_EI_MODE(1) |
4713 * 3b: Enable Coarse Power Gating only when RC6 is enabled.
4714 * WaRsDisableCoarsePowerGating:skl,bxt - Render/Media PG need to be disabled with RC6.
4716 if (IS_BXT_REVID(dev
, 0, BXT_REVID_A1
) ||
4717 ((IS_SKL_GT3(dev
) || IS_SKL_GT4(dev
)) &&
4718 IS_SKL_REVID(dev
, 0, SKL_REVID_F0
)))
4719 I915_WRITE(GEN9_PG_ENABLE
, 0);
4721 I915_WRITE(GEN9_PG_ENABLE
, (rc6_mask
& GEN6_RC_CTL_RC6_ENABLE
) ?
4722 (GEN9_RENDER_PG_ENABLE
| GEN9_MEDIA_PG_ENABLE
) : 0);
4724 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4728 static void gen8_enable_rps(struct drm_device
*dev
)
4730 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4731 struct intel_engine_cs
*ring
;
4732 uint32_t rc6_mask
= 0;
4735 /* 1a: Software RC state - RC0 */
4736 I915_WRITE(GEN6_RC_STATE
, 0);
4738 /* 1c & 1d: Get forcewake during program sequence. Although the driver
4739 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4740 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4742 /* 2a: Disable RC states. */
4743 I915_WRITE(GEN6_RC_CONTROL
, 0);
4745 /* Initialize rps frequencies */
4746 gen6_init_rps_frequencies(dev
);
4748 /* 2b: Program RC6 thresholds.*/
4749 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16);
4750 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
4751 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
4752 for_each_ring(ring
, dev_priv
, unused
)
4753 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
4754 I915_WRITE(GEN6_RC_SLEEP
, 0);
4755 if (IS_BROADWELL(dev
))
4756 I915_WRITE(GEN6_RC6_THRESHOLD
, 625); /* 800us/1.28 for TO */
4758 I915_WRITE(GEN6_RC6_THRESHOLD
, 50000); /* 50/125ms per EI */
4761 if (intel_enable_rc6(dev
) & INTEL_RC6_ENABLE
)
4762 rc6_mask
= GEN6_RC_CTL_RC6_ENABLE
;
4763 intel_print_rc6_info(dev
, rc6_mask
);
4764 if (IS_BROADWELL(dev
))
4765 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
4766 GEN7_RC_CTL_TO_MODE
|
4769 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
4770 GEN6_RC_CTL_EI_MODE(1) |
4773 /* 4 Program defaults and thresholds for RPS*/
4774 I915_WRITE(GEN6_RPNSWREQ
,
4775 HSW_FREQUENCY(dev_priv
->rps
.rp1_freq
));
4776 I915_WRITE(GEN6_RC_VIDEO_FREQ
,
4777 HSW_FREQUENCY(dev_priv
->rps
.rp1_freq
));
4778 /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
4779 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 100000000 / 128); /* 1 second timeout */
4781 /* Docs recommend 900MHz, and 300 MHz respectively */
4782 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS
,
4783 dev_priv
->rps
.max_freq_softlimit
<< 24 |
4784 dev_priv
->rps
.min_freq_softlimit
<< 16);
4786 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 7600000 / 128); /* 76ms busyness per EI, 90% */
4787 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 31300000 / 128); /* 313ms busyness per EI, 70%*/
4788 I915_WRITE(GEN6_RP_UP_EI
, 66000); /* 84.48ms, XXX: random? */
4789 I915_WRITE(GEN6_RP_DOWN_EI
, 350000); /* 448ms, XXX: random? */
4791 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
4794 I915_WRITE(GEN6_RP_CONTROL
,
4795 GEN6_RP_MEDIA_TURBO
|
4796 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
4797 GEN6_RP_MEDIA_IS_GFX
|
4799 GEN6_RP_UP_BUSY_AVG
|
4800 GEN6_RP_DOWN_IDLE_AVG
);
4802 /* 6: Ring frequency + overclocking (our driver does this later */
4804 dev_priv
->rps
.power
= HIGH_POWER
; /* force a reset */
4805 gen6_set_rps(dev_priv
->dev
, dev_priv
->rps
.idle_freq
);
4807 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4810 static void gen6_enable_rps(struct drm_device
*dev
)
4812 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4813 struct intel_engine_cs
*ring
;
4814 u32 rc6vids
, pcu_mbox
= 0, rc6_mask
= 0;
4819 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
4821 /* Here begins a magic sequence of register writes to enable
4822 * auto-downclocking.
4824 * Perhaps there might be some value in exposing these to
4827 I915_WRITE(GEN6_RC_STATE
, 0);
4829 /* Clear the DBG now so we don't confuse earlier errors */
4830 if ((gtfifodbg
= I915_READ(GTFIFODBG
))) {
4831 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg
);
4832 I915_WRITE(GTFIFODBG
, gtfifodbg
);
4835 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4837 /* Initialize rps frequencies */
4838 gen6_init_rps_frequencies(dev
);
4840 /* disable the counters and set deterministic thresholds */
4841 I915_WRITE(GEN6_RC_CONTROL
, 0);
4843 I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT
, 1000 << 16);
4844 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16 | 30);
4845 I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT
, 30);
4846 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000);
4847 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25);
4849 for_each_ring(ring
, dev_priv
, i
)
4850 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
4852 I915_WRITE(GEN6_RC_SLEEP
, 0);
4853 I915_WRITE(GEN6_RC1e_THRESHOLD
, 1000);
4854 if (IS_IVYBRIDGE(dev
))
4855 I915_WRITE(GEN6_RC6_THRESHOLD
, 125000);
4857 I915_WRITE(GEN6_RC6_THRESHOLD
, 50000);
4858 I915_WRITE(GEN6_RC6p_THRESHOLD
, 150000);
4859 I915_WRITE(GEN6_RC6pp_THRESHOLD
, 64000); /* unused */
4861 /* Check if we are enabling RC6 */
4862 rc6_mode
= intel_enable_rc6(dev_priv
->dev
);
4863 if (rc6_mode
& INTEL_RC6_ENABLE
)
4864 rc6_mask
|= GEN6_RC_CTL_RC6_ENABLE
;
4866 /* We don't use those on Haswell */
4867 if (!IS_HASWELL(dev
)) {
4868 if (rc6_mode
& INTEL_RC6p_ENABLE
)
4869 rc6_mask
|= GEN6_RC_CTL_RC6p_ENABLE
;
4871 if (rc6_mode
& INTEL_RC6pp_ENABLE
)
4872 rc6_mask
|= GEN6_RC_CTL_RC6pp_ENABLE
;
4875 intel_print_rc6_info(dev
, rc6_mask
);
4877 I915_WRITE(GEN6_RC_CONTROL
,
4879 GEN6_RC_CTL_EI_MODE(1) |
4880 GEN6_RC_CTL_HW_ENABLE
);
4882 /* Power down if completely idle for over 50ms */
4883 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 50000);
4884 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
4886 ret
= sandybridge_pcode_write(dev_priv
, GEN6_PCODE_WRITE_MIN_FREQ_TABLE
, 0);
4888 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
4890 ret
= sandybridge_pcode_read(dev_priv
, GEN6_READ_OC_PARAMS
, &pcu_mbox
);
4891 if (!ret
&& (pcu_mbox
& (1<<31))) { /* OC supported */
4892 DRM_DEBUG_DRIVER("Overclocking supported. Max: %dMHz, Overclock max: %dMHz\n",
4893 (dev_priv
->rps
.max_freq_softlimit
& 0xff) * 50,
4894 (pcu_mbox
& 0xff) * 50);
4895 dev_priv
->rps
.max_freq
= pcu_mbox
& 0xff;
4898 dev_priv
->rps
.power
= HIGH_POWER
; /* force a reset */
4899 gen6_set_rps(dev_priv
->dev
, dev_priv
->rps
.idle_freq
);
4902 ret
= sandybridge_pcode_read(dev_priv
, GEN6_PCODE_READ_RC6VIDS
, &rc6vids
);
4903 if (IS_GEN6(dev
) && ret
) {
4904 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
4905 } else if (IS_GEN6(dev
) && (GEN6_DECODE_RC6_VID(rc6vids
& 0xff) < 450)) {
4906 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
4907 GEN6_DECODE_RC6_VID(rc6vids
& 0xff), 450);
4908 rc6vids
&= 0xffff00;
4909 rc6vids
|= GEN6_ENCODE_RC6_VID(450);
4910 ret
= sandybridge_pcode_write(dev_priv
, GEN6_PCODE_WRITE_RC6VIDS
, rc6vids
);
4912 DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
4915 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4918 static void __gen6_update_ring_freq(struct drm_device
*dev
)
4920 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4922 unsigned int gpu_freq
;
4923 unsigned int max_ia_freq
, min_ring_freq
;
4924 unsigned int max_gpu_freq
, min_gpu_freq
;
4925 int scaling_factor
= 180;
4926 struct cpufreq_policy
*policy
;
4928 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
4930 policy
= cpufreq_cpu_get(0);
4932 max_ia_freq
= policy
->cpuinfo
.max_freq
;
4933 cpufreq_cpu_put(policy
);
4936 * Default to measured freq if none found, PCU will ensure we
4939 max_ia_freq
= tsc_khz
;
4942 /* Convert from kHz to MHz */
4943 max_ia_freq
/= 1000;
4945 min_ring_freq
= I915_READ(DCLK
) & 0xf;
4946 /* convert DDR frequency from units of 266.6MHz to bandwidth */
4947 min_ring_freq
= mult_frac(min_ring_freq
, 8, 3);
4949 if (IS_SKYLAKE(dev
) || IS_KABYLAKE(dev
)) {
4950 /* Convert GT frequency to 50 HZ units */
4951 min_gpu_freq
= dev_priv
->rps
.min_freq
/ GEN9_FREQ_SCALER
;
4952 max_gpu_freq
= dev_priv
->rps
.max_freq
/ GEN9_FREQ_SCALER
;
4954 min_gpu_freq
= dev_priv
->rps
.min_freq
;
4955 max_gpu_freq
= dev_priv
->rps
.max_freq
;
4959 * For each potential GPU frequency, load a ring frequency we'd like
4960 * to use for memory access. We do this by specifying the IA frequency
4961 * the PCU should use as a reference to determine the ring frequency.
4963 for (gpu_freq
= max_gpu_freq
; gpu_freq
>= min_gpu_freq
; gpu_freq
--) {
4964 int diff
= max_gpu_freq
- gpu_freq
;
4965 unsigned int ia_freq
= 0, ring_freq
= 0;
4967 if (IS_SKYLAKE(dev
) || IS_KABYLAKE(dev
)) {
4969 * ring_freq = 2 * GT. ring_freq is in 100MHz units
4970 * No floor required for ring frequency on SKL.
4972 ring_freq
= gpu_freq
;
4973 } else if (INTEL_INFO(dev
)->gen
>= 8) {
4974 /* max(2 * GT, DDR). NB: GT is 50MHz units */
4975 ring_freq
= max(min_ring_freq
, gpu_freq
);
4976 } else if (IS_HASWELL(dev
)) {
4977 ring_freq
= mult_frac(gpu_freq
, 5, 4);
4978 ring_freq
= max(min_ring_freq
, ring_freq
);
4979 /* leave ia_freq as the default, chosen by cpufreq */
4981 /* On older processors, there is no separate ring
4982 * clock domain, so in order to boost the bandwidth
4983 * of the ring, we need to upclock the CPU (ia_freq).
4985 * For GPU frequencies less than 750MHz,
4986 * just use the lowest ring freq.
4988 if (gpu_freq
< min_freq
)
4991 ia_freq
= max_ia_freq
- ((diff
* scaling_factor
) / 2);
4992 ia_freq
= DIV_ROUND_CLOSEST(ia_freq
, 100);
4995 sandybridge_pcode_write(dev_priv
,
4996 GEN6_PCODE_WRITE_MIN_FREQ_TABLE
,
4997 ia_freq
<< GEN6_PCODE_FREQ_IA_RATIO_SHIFT
|
4998 ring_freq
<< GEN6_PCODE_FREQ_RING_RATIO_SHIFT
|
5003 void gen6_update_ring_freq(struct drm_device
*dev
)
5005 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5007 if (!HAS_CORE_RING_FREQ(dev
))
5010 mutex_lock(&dev_priv
->rps
.hw_lock
);
5011 __gen6_update_ring_freq(dev
);
5012 mutex_unlock(&dev_priv
->rps
.hw_lock
);
5015 static int cherryview_rps_max_freq(struct drm_i915_private
*dev_priv
)
5017 struct drm_device
*dev
= dev_priv
->dev
;
5020 val
= vlv_punit_read(dev_priv
, FB_GFX_FMAX_AT_VMAX_FUSE
);
5022 switch (INTEL_INFO(dev
)->eu_total
) {
5024 /* (2 * 4) config */
5025 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT
);
5028 /* (2 * 6) config */
5029 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT
);
5032 /* (2 * 8) config */
5034 /* Setting (2 * 8) Min RP0 for any other combination */
5035 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT
);
5039 rp0
= (rp0
& FB_GFX_FREQ_FUSE_MASK
);
5044 static int cherryview_rps_rpe_freq(struct drm_i915_private
*dev_priv
)
5048 val
= vlv_punit_read(dev_priv
, PUNIT_GPU_DUTYCYCLE_REG
);
5049 rpe
= (val
>> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT
) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK
;
5054 static int cherryview_rps_guar_freq(struct drm_i915_private
*dev_priv
)
5058 val
= vlv_punit_read(dev_priv
, FB_GFX_FMAX_AT_VMAX_FUSE
);
5059 rp1
= (val
& FB_GFX_FREQ_FUSE_MASK
);
5064 static int valleyview_rps_guar_freq(struct drm_i915_private
*dev_priv
)
5068 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FREQ_FUSE
);
5070 rp1
= (val
& FB_GFX_FGUARANTEED_FREQ_FUSE_MASK
) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT
;
5075 static int valleyview_rps_max_freq(struct drm_i915_private
*dev_priv
)
5079 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FREQ_FUSE
);
5081 rp0
= (val
& FB_GFX_MAX_FREQ_FUSE_MASK
) >> FB_GFX_MAX_FREQ_FUSE_SHIFT
;
5083 rp0
= min_t(u32
, rp0
, 0xea);
5088 static int valleyview_rps_rpe_freq(struct drm_i915_private
*dev_priv
)
5092 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FMAX_FUSE_LO
);
5093 rpe
= (val
& FB_FMAX_VMIN_FREQ_LO_MASK
) >> FB_FMAX_VMIN_FREQ_LO_SHIFT
;
5094 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FMAX_FUSE_HI
);
5095 rpe
|= (val
& FB_FMAX_VMIN_FREQ_HI_MASK
) << 5;
5100 static int valleyview_rps_min_freq(struct drm_i915_private
*dev_priv
)
5104 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_LFM
) & 0xff;
5106 * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
5107 * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
5108 * a BYT-M B0 the above register contains 0xbf. Moreover when setting
5109 * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
5110 * to make sure it matches what Punit accepts.
5112 return max_t(u32
, val
, 0xc0);
5115 /* Check that the pctx buffer wasn't move under us. */
5116 static void valleyview_check_pctx(struct drm_i915_private
*dev_priv
)
5118 unsigned long pctx_addr
= I915_READ(VLV_PCBR
) & ~4095;
5120 WARN_ON(pctx_addr
!= dev_priv
->mm
.stolen_base
+
5121 dev_priv
->vlv_pctx
->stolen
->start
);
5125 /* Check that the pcbr address is not empty. */
5126 static void cherryview_check_pctx(struct drm_i915_private
*dev_priv
)
5128 unsigned long pctx_addr
= I915_READ(VLV_PCBR
) & ~4095;
5130 WARN_ON((pctx_addr
>> VLV_PCBR_ADDR_SHIFT
) == 0);
5133 static void cherryview_setup_pctx(struct drm_device
*dev
)
5135 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5136 unsigned long pctx_paddr
, paddr
;
5137 struct i915_gtt
*gtt
= &dev_priv
->gtt
;
5139 int pctx_size
= 32*1024;
5141 WARN_ON(!mutex_is_locked(&dev
->struct_mutex
));
5143 pcbr
= I915_READ(VLV_PCBR
);
5144 if ((pcbr
>> VLV_PCBR_ADDR_SHIFT
) == 0) {
5145 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5146 paddr
= (dev_priv
->mm
.stolen_base
+
5147 (gtt
->stolen_size
- pctx_size
));
5149 pctx_paddr
= (paddr
& (~4095));
5150 I915_WRITE(VLV_PCBR
, pctx_paddr
);
5153 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR
));
5156 static void valleyview_setup_pctx(struct drm_device
*dev
)
5158 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5159 struct drm_i915_gem_object
*pctx
;
5160 unsigned long pctx_paddr
;
5162 int pctx_size
= 24*1024;
5164 WARN_ON(!mutex_is_locked(&dev
->struct_mutex
));
5166 pcbr
= I915_READ(VLV_PCBR
);
5168 /* BIOS set it up already, grab the pre-alloc'd space */
5171 pcbr_offset
= (pcbr
& (~4095)) - dev_priv
->mm
.stolen_base
;
5172 pctx
= i915_gem_object_create_stolen_for_preallocated(dev_priv
->dev
,
5174 I915_GTT_OFFSET_NONE
,
5179 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5182 * From the Gunit register HAS:
5183 * The Gfx driver is expected to program this register and ensure
5184 * proper allocation within Gfx stolen memory. For example, this
5185 * register should be programmed such than the PCBR range does not
5186 * overlap with other ranges, such as the frame buffer, protected
5187 * memory, or any other relevant ranges.
5189 pctx
= i915_gem_object_create_stolen(dev
, pctx_size
);
5191 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
5195 pctx_paddr
= dev_priv
->mm
.stolen_base
+ pctx
->stolen
->start
;
5196 I915_WRITE(VLV_PCBR
, pctx_paddr
);
5199 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR
));
5200 dev_priv
->vlv_pctx
= pctx
;
5203 static void valleyview_cleanup_pctx(struct drm_device
*dev
)
5205 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5207 if (WARN_ON(!dev_priv
->vlv_pctx
))
5210 drm_gem_object_unreference(&dev_priv
->vlv_pctx
->base
);
5211 dev_priv
->vlv_pctx
= NULL
;
5214 static void valleyview_init_gt_powersave(struct drm_device
*dev
)
5216 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5219 valleyview_setup_pctx(dev
);
5221 mutex_lock(&dev_priv
->rps
.hw_lock
);
5223 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
5224 switch ((val
>> 6) & 3) {
5227 dev_priv
->mem_freq
= 800;
5230 dev_priv
->mem_freq
= 1066;
5233 dev_priv
->mem_freq
= 1333;
5236 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv
->mem_freq
);
5238 dev_priv
->rps
.max_freq
= valleyview_rps_max_freq(dev_priv
);
5239 dev_priv
->rps
.rp0_freq
= dev_priv
->rps
.max_freq
;
5240 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5241 intel_gpu_freq(dev_priv
, dev_priv
->rps
.max_freq
),
5242 dev_priv
->rps
.max_freq
);
5244 dev_priv
->rps
.efficient_freq
= valleyview_rps_rpe_freq(dev_priv
);
5245 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5246 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
5247 dev_priv
->rps
.efficient_freq
);
5249 dev_priv
->rps
.rp1_freq
= valleyview_rps_guar_freq(dev_priv
);
5250 DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
5251 intel_gpu_freq(dev_priv
, dev_priv
->rps
.rp1_freq
),
5252 dev_priv
->rps
.rp1_freq
);
5254 dev_priv
->rps
.min_freq
= valleyview_rps_min_freq(dev_priv
);
5255 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5256 intel_gpu_freq(dev_priv
, dev_priv
->rps
.min_freq
),
5257 dev_priv
->rps
.min_freq
);
5259 dev_priv
->rps
.idle_freq
= dev_priv
->rps
.min_freq
;
5261 /* Preserve min/max settings in case of re-init */
5262 if (dev_priv
->rps
.max_freq_softlimit
== 0)
5263 dev_priv
->rps
.max_freq_softlimit
= dev_priv
->rps
.max_freq
;
5265 if (dev_priv
->rps
.min_freq_softlimit
== 0)
5266 dev_priv
->rps
.min_freq_softlimit
= dev_priv
->rps
.min_freq
;
5268 mutex_unlock(&dev_priv
->rps
.hw_lock
);
5271 static void cherryview_init_gt_powersave(struct drm_device
*dev
)
5273 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5276 cherryview_setup_pctx(dev
);
5278 mutex_lock(&dev_priv
->rps
.hw_lock
);
5280 mutex_lock(&dev_priv
->sb_lock
);
5281 val
= vlv_cck_read(dev_priv
, CCK_FUSE_REG
);
5282 mutex_unlock(&dev_priv
->sb_lock
);
5284 switch ((val
>> 2) & 0x7) {
5286 dev_priv
->mem_freq
= 2000;
5289 dev_priv
->mem_freq
= 1600;
5292 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv
->mem_freq
);
5294 dev_priv
->rps
.max_freq
= cherryview_rps_max_freq(dev_priv
);
5295 dev_priv
->rps
.rp0_freq
= dev_priv
->rps
.max_freq
;
5296 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5297 intel_gpu_freq(dev_priv
, dev_priv
->rps
.max_freq
),
5298 dev_priv
->rps
.max_freq
);
5300 dev_priv
->rps
.efficient_freq
= cherryview_rps_rpe_freq(dev_priv
);
5301 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5302 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
5303 dev_priv
->rps
.efficient_freq
);
5305 dev_priv
->rps
.rp1_freq
= cherryview_rps_guar_freq(dev_priv
);
5306 DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
5307 intel_gpu_freq(dev_priv
, dev_priv
->rps
.rp1_freq
),
5308 dev_priv
->rps
.rp1_freq
);
5310 /* PUnit validated range is only [RPe, RP0] */
5311 dev_priv
->rps
.min_freq
= dev_priv
->rps
.efficient_freq
;
5312 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5313 intel_gpu_freq(dev_priv
, dev_priv
->rps
.min_freq
),
5314 dev_priv
->rps
.min_freq
);
5316 WARN_ONCE((dev_priv
->rps
.max_freq
|
5317 dev_priv
->rps
.efficient_freq
|
5318 dev_priv
->rps
.rp1_freq
|
5319 dev_priv
->rps
.min_freq
) & 1,
5320 "Odd GPU freq values\n");
5322 dev_priv
->rps
.idle_freq
= dev_priv
->rps
.min_freq
;
5324 /* Preserve min/max settings in case of re-init */
5325 if (dev_priv
->rps
.max_freq_softlimit
== 0)
5326 dev_priv
->rps
.max_freq_softlimit
= dev_priv
->rps
.max_freq
;
5328 if (dev_priv
->rps
.min_freq_softlimit
== 0)
5329 dev_priv
->rps
.min_freq_softlimit
= dev_priv
->rps
.min_freq
;
5331 mutex_unlock(&dev_priv
->rps
.hw_lock
);
5334 static void valleyview_cleanup_gt_powersave(struct drm_device
*dev
)
5336 valleyview_cleanup_pctx(dev
);
5339 static void cherryview_enable_rps(struct drm_device
*dev
)
5341 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5342 struct intel_engine_cs
*ring
;
5343 u32 gtfifodbg
, val
, rc6_mode
= 0, pcbr
;
5346 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
5348 gtfifodbg
= I915_READ(GTFIFODBG
);
5350 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5352 I915_WRITE(GTFIFODBG
, gtfifodbg
);
5355 cherryview_check_pctx(dev_priv
);
5357 /* 1a & 1b: Get forcewake during program sequence. Although the driver
5358 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5359 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
5361 /* Disable RC states. */
5362 I915_WRITE(GEN6_RC_CONTROL
, 0);
5364 /* 2a: Program RC6 thresholds.*/
5365 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16);
5366 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
5367 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
5369 for_each_ring(ring
, dev_priv
, i
)
5370 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
5371 I915_WRITE(GEN6_RC_SLEEP
, 0);
5373 /* TO threshold set to 500 us ( 0x186 * 1.28 us) */
5374 I915_WRITE(GEN6_RC6_THRESHOLD
, 0x186);
5376 /* allows RC6 residency counter to work */
5377 I915_WRITE(VLV_COUNTER_CONTROL
,
5378 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH
|
5379 VLV_MEDIA_RC6_COUNT_EN
|
5380 VLV_RENDER_RC6_COUNT_EN
));
5382 /* For now we assume BIOS is allocating and populating the PCBR */
5383 pcbr
= I915_READ(VLV_PCBR
);
5386 if ((intel_enable_rc6(dev
) & INTEL_RC6_ENABLE
) &&
5387 (pcbr
>> VLV_PCBR_ADDR_SHIFT
))
5388 rc6_mode
= GEN7_RC_CTL_TO_MODE
;
5390 I915_WRITE(GEN6_RC_CONTROL
, rc6_mode
);
5392 /* 4 Program defaults and thresholds for RPS*/
5393 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 1000000);
5394 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 59400);
5395 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 245000);
5396 I915_WRITE(GEN6_RP_UP_EI
, 66000);
5397 I915_WRITE(GEN6_RP_DOWN_EI
, 350000);
5399 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
5402 I915_WRITE(GEN6_RP_CONTROL
,
5403 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
5404 GEN6_RP_MEDIA_IS_GFX
|
5406 GEN6_RP_UP_BUSY_AVG
|
5407 GEN6_RP_DOWN_IDLE_AVG
);
5409 /* Setting Fixed Bias */
5410 val
= VLV_OVERRIDE_EN
|
5412 CHV_BIAS_CPU_50_SOC_50
;
5413 vlv_punit_write(dev_priv
, VLV_TURBO_SOC_OVERRIDE
, val
);
5415 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
5417 /* RPS code assumes GPLL is used */
5418 WARN_ONCE((val
& GPLLENABLE
) == 0, "GPLL not enabled\n");
5420 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val
& GPLLENABLE
));
5421 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val
);
5423 dev_priv
->rps
.cur_freq
= (val
>> 8) & 0xff;
5424 DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
5425 intel_gpu_freq(dev_priv
, dev_priv
->rps
.cur_freq
),
5426 dev_priv
->rps
.cur_freq
);
5428 DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
5429 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
5430 dev_priv
->rps
.efficient_freq
);
5432 valleyview_set_rps(dev_priv
->dev
, dev_priv
->rps
.efficient_freq
);
5434 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
5437 static void valleyview_enable_rps(struct drm_device
*dev
)
5439 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5440 struct intel_engine_cs
*ring
;
5441 u32 gtfifodbg
, val
, rc6_mode
= 0;
5444 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
5446 valleyview_check_pctx(dev_priv
);
5448 if ((gtfifodbg
= I915_READ(GTFIFODBG
))) {
5449 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5451 I915_WRITE(GTFIFODBG
, gtfifodbg
);
5454 /* If VLV, Forcewake all wells, else re-direct to regular path */
5455 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
5457 /* Disable RC states. */
5458 I915_WRITE(GEN6_RC_CONTROL
, 0);
5460 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 1000000);
5461 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 59400);
5462 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 245000);
5463 I915_WRITE(GEN6_RP_UP_EI
, 66000);
5464 I915_WRITE(GEN6_RP_DOWN_EI
, 350000);
5466 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
5468 I915_WRITE(GEN6_RP_CONTROL
,
5469 GEN6_RP_MEDIA_TURBO
|
5470 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
5471 GEN6_RP_MEDIA_IS_GFX
|
5473 GEN6_RP_UP_BUSY_AVG
|
5474 GEN6_RP_DOWN_IDLE_CONT
);
5476 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 0x00280000);
5477 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000);
5478 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25);
5480 for_each_ring(ring
, dev_priv
, i
)
5481 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
5483 I915_WRITE(GEN6_RC6_THRESHOLD
, 0x557);
5485 /* allows RC6 residency counter to work */
5486 I915_WRITE(VLV_COUNTER_CONTROL
,
5487 _MASKED_BIT_ENABLE(VLV_MEDIA_RC0_COUNT_EN
|
5488 VLV_RENDER_RC0_COUNT_EN
|
5489 VLV_MEDIA_RC6_COUNT_EN
|
5490 VLV_RENDER_RC6_COUNT_EN
));
5492 if (intel_enable_rc6(dev
) & INTEL_RC6_ENABLE
)
5493 rc6_mode
= GEN7_RC_CTL_TO_MODE
| VLV_RC_CTL_CTX_RST_PARALLEL
;
5495 intel_print_rc6_info(dev
, rc6_mode
);
5497 I915_WRITE(GEN6_RC_CONTROL
, rc6_mode
);
5499 /* Setting Fixed Bias */
5500 val
= VLV_OVERRIDE_EN
|
5502 VLV_BIAS_CPU_125_SOC_875
;
5503 vlv_punit_write(dev_priv
, VLV_TURBO_SOC_OVERRIDE
, val
);
5505 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
5507 /* RPS code assumes GPLL is used */
5508 WARN_ONCE((val
& GPLLENABLE
) == 0, "GPLL not enabled\n");
5510 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val
& GPLLENABLE
));
5511 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val
);
5513 dev_priv
->rps
.cur_freq
= (val
>> 8) & 0xff;
5514 DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
5515 intel_gpu_freq(dev_priv
, dev_priv
->rps
.cur_freq
),
5516 dev_priv
->rps
.cur_freq
);
5518 DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
5519 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
5520 dev_priv
->rps
.efficient_freq
);
5522 valleyview_set_rps(dev_priv
->dev
, dev_priv
->rps
.efficient_freq
);
5524 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
5527 static unsigned long intel_pxfreq(u32 vidfreq
)
5530 int div
= (vidfreq
& 0x3f0000) >> 16;
5531 int post
= (vidfreq
& 0x3000) >> 12;
5532 int pre
= (vidfreq
& 0x7);
5537 freq
= ((div
* 133333) / ((1<<post
) * pre
));
5542 static const struct cparams
{
5548 { 1, 1333, 301, 28664 },
5549 { 1, 1066, 294, 24460 },
5550 { 1, 800, 294, 25192 },
5551 { 0, 1333, 276, 27605 },
5552 { 0, 1066, 276, 27605 },
5553 { 0, 800, 231, 23784 },
5556 static unsigned long __i915_chipset_val(struct drm_i915_private
*dev_priv
)
5558 u64 total_count
, diff
, ret
;
5559 u32 count1
, count2
, count3
, m
= 0, c
= 0;
5560 unsigned long now
= jiffies_to_msecs(jiffies
), diff1
;
5563 assert_spin_locked(&mchdev_lock
);
5565 diff1
= now
- dev_priv
->ips
.last_time1
;
5567 /* Prevent division-by-zero if we are asking too fast.
5568 * Also, we don't get interesting results if we are polling
5569 * faster than once in 10ms, so just return the saved value
5573 return dev_priv
->ips
.chipset_power
;
5575 count1
= I915_READ(DMIEC
);
5576 count2
= I915_READ(DDREC
);
5577 count3
= I915_READ(CSIEC
);
5579 total_count
= count1
+ count2
+ count3
;
5581 /* FIXME: handle per-counter overflow */
5582 if (total_count
< dev_priv
->ips
.last_count1
) {
5583 diff
= ~0UL - dev_priv
->ips
.last_count1
;
5584 diff
+= total_count
;
5586 diff
= total_count
- dev_priv
->ips
.last_count1
;
5589 for (i
= 0; i
< ARRAY_SIZE(cparams
); i
++) {
5590 if (cparams
[i
].i
== dev_priv
->ips
.c_m
&&
5591 cparams
[i
].t
== dev_priv
->ips
.r_t
) {
5598 diff
= div_u64(diff
, diff1
);
5599 ret
= ((m
* diff
) + c
);
5600 ret
= div_u64(ret
, 10);
5602 dev_priv
->ips
.last_count1
= total_count
;
5603 dev_priv
->ips
.last_time1
= now
;
5605 dev_priv
->ips
.chipset_power
= ret
;
5610 unsigned long i915_chipset_val(struct drm_i915_private
*dev_priv
)
5612 struct drm_device
*dev
= dev_priv
->dev
;
5615 if (INTEL_INFO(dev
)->gen
!= 5)
5618 spin_lock_irq(&mchdev_lock
);
5620 val
= __i915_chipset_val(dev_priv
);
5622 spin_unlock_irq(&mchdev_lock
);
5627 unsigned long i915_mch_val(struct drm_i915_private
*dev_priv
)
5629 unsigned long m
, x
, b
;
5632 tsfs
= I915_READ(TSFS
);
5634 m
= ((tsfs
& TSFS_SLOPE_MASK
) >> TSFS_SLOPE_SHIFT
);
5635 x
= I915_READ8(TR1
);
5637 b
= tsfs
& TSFS_INTR_MASK
;
5639 return ((m
* x
) / 127) - b
;
5642 static int _pxvid_to_vd(u8 pxvid
)
5647 if (pxvid
>= 8 && pxvid
< 31)
5650 return (pxvid
+ 2) * 125;
5653 static u32
pvid_to_extvid(struct drm_i915_private
*dev_priv
, u8 pxvid
)
5655 struct drm_device
*dev
= dev_priv
->dev
;
5656 const int vd
= _pxvid_to_vd(pxvid
);
5657 const int vm
= vd
- 1125;
5659 if (INTEL_INFO(dev
)->is_mobile
)
5660 return vm
> 0 ? vm
: 0;
5665 static void __i915_update_gfx_val(struct drm_i915_private
*dev_priv
)
5667 u64 now
, diff
, diffms
;
5670 assert_spin_locked(&mchdev_lock
);
5672 now
= ktime_get_raw_ns();
5673 diffms
= now
- dev_priv
->ips
.last_time2
;
5674 do_div(diffms
, NSEC_PER_MSEC
);
5676 /* Don't divide by 0 */
5680 count
= I915_READ(GFXEC
);
5682 if (count
< dev_priv
->ips
.last_count2
) {
5683 diff
= ~0UL - dev_priv
->ips
.last_count2
;
5686 diff
= count
- dev_priv
->ips
.last_count2
;
5689 dev_priv
->ips
.last_count2
= count
;
5690 dev_priv
->ips
.last_time2
= now
;
5692 /* More magic constants... */
5694 diff
= div_u64(diff
, diffms
* 10);
5695 dev_priv
->ips
.gfx_power
= diff
;
5698 void i915_update_gfx_val(struct drm_i915_private
*dev_priv
)
5700 struct drm_device
*dev
= dev_priv
->dev
;
5702 if (INTEL_INFO(dev
)->gen
!= 5)
5705 spin_lock_irq(&mchdev_lock
);
5707 __i915_update_gfx_val(dev_priv
);
5709 spin_unlock_irq(&mchdev_lock
);
5712 static unsigned long __i915_gfx_val(struct drm_i915_private
*dev_priv
)
5714 unsigned long t
, corr
, state1
, corr2
, state2
;
5717 assert_spin_locked(&mchdev_lock
);
5719 pxvid
= I915_READ(PXVFREQ(dev_priv
->rps
.cur_freq
));
5720 pxvid
= (pxvid
>> 24) & 0x7f;
5721 ext_v
= pvid_to_extvid(dev_priv
, pxvid
);
5725 t
= i915_mch_val(dev_priv
);
5727 /* Revel in the empirically derived constants */
5729 /* Correction factor in 1/100000 units */
5731 corr
= ((t
* 2349) + 135940);
5733 corr
= ((t
* 964) + 29317);
5735 corr
= ((t
* 301) + 1004);
5737 corr
= corr
* ((150142 * state1
) / 10000 - 78642);
5739 corr2
= (corr
* dev_priv
->ips
.corr
);
5741 state2
= (corr2
* state1
) / 10000;
5742 state2
/= 100; /* convert to mW */
5744 __i915_update_gfx_val(dev_priv
);
5746 return dev_priv
->ips
.gfx_power
+ state2
;
5749 unsigned long i915_gfx_val(struct drm_i915_private
*dev_priv
)
5751 struct drm_device
*dev
= dev_priv
->dev
;
5754 if (INTEL_INFO(dev
)->gen
!= 5)
5757 spin_lock_irq(&mchdev_lock
);
5759 val
= __i915_gfx_val(dev_priv
);
5761 spin_unlock_irq(&mchdev_lock
);
5767 * i915_read_mch_val - return value for IPS use
5769 * Calculate and return a value for the IPS driver to use when deciding whether
5770 * we have thermal and power headroom to increase CPU or GPU power budget.
5772 unsigned long i915_read_mch_val(void)
5774 struct drm_i915_private
*dev_priv
;
5775 unsigned long chipset_val
, graphics_val
, ret
= 0;
5777 spin_lock_irq(&mchdev_lock
);
5780 dev_priv
= i915_mch_dev
;
5782 chipset_val
= __i915_chipset_val(dev_priv
);
5783 graphics_val
= __i915_gfx_val(dev_priv
);
5785 ret
= chipset_val
+ graphics_val
;
5788 spin_unlock_irq(&mchdev_lock
);
5792 EXPORT_SYMBOL_GPL(i915_read_mch_val
);
5795 * i915_gpu_raise - raise GPU frequency limit
5797 * Raise the limit; IPS indicates we have thermal headroom.
5799 bool i915_gpu_raise(void)
5801 struct drm_i915_private
*dev_priv
;
5804 spin_lock_irq(&mchdev_lock
);
5805 if (!i915_mch_dev
) {
5809 dev_priv
= i915_mch_dev
;
5811 if (dev_priv
->ips
.max_delay
> dev_priv
->ips
.fmax
)
5812 dev_priv
->ips
.max_delay
--;
5815 spin_unlock_irq(&mchdev_lock
);
5819 EXPORT_SYMBOL_GPL(i915_gpu_raise
);
5822 * i915_gpu_lower - lower GPU frequency limit
5824 * IPS indicates we're close to a thermal limit, so throttle back the GPU
5825 * frequency maximum.
5827 bool i915_gpu_lower(void)
5829 struct drm_i915_private
*dev_priv
;
5832 spin_lock_irq(&mchdev_lock
);
5833 if (!i915_mch_dev
) {
5837 dev_priv
= i915_mch_dev
;
5839 if (dev_priv
->ips
.max_delay
< dev_priv
->ips
.min_delay
)
5840 dev_priv
->ips
.max_delay
++;
5843 spin_unlock_irq(&mchdev_lock
);
5847 EXPORT_SYMBOL_GPL(i915_gpu_lower
);
5850 * i915_gpu_busy - indicate GPU business to IPS
5852 * Tell the IPS driver whether or not the GPU is busy.
5854 bool i915_gpu_busy(void)
5856 struct drm_i915_private
*dev_priv
;
5857 struct intel_engine_cs
*ring
;
5861 spin_lock_irq(&mchdev_lock
);
5864 dev_priv
= i915_mch_dev
;
5866 for_each_ring(ring
, dev_priv
, i
)
5867 ret
|= !list_empty(&ring
->request_list
);
5870 spin_unlock_irq(&mchdev_lock
);
5874 EXPORT_SYMBOL_GPL(i915_gpu_busy
);
5877 * i915_gpu_turbo_disable - disable graphics turbo
5879 * Disable graphics turbo by resetting the max frequency and setting the
5880 * current frequency to the default.
5882 bool i915_gpu_turbo_disable(void)
5884 struct drm_i915_private
*dev_priv
;
5887 spin_lock_irq(&mchdev_lock
);
5888 if (!i915_mch_dev
) {
5892 dev_priv
= i915_mch_dev
;
5894 dev_priv
->ips
.max_delay
= dev_priv
->ips
.fstart
;
5896 if (!ironlake_set_drps(dev_priv
->dev
, dev_priv
->ips
.fstart
))
5900 spin_unlock_irq(&mchdev_lock
);
5904 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable
);
5907 * Tells the intel_ips driver that the i915 driver is now loaded, if
5908 * IPS got loaded first.
5910 * This awkward dance is so that neither module has to depend on the
5911 * other in order for IPS to do the appropriate communication of
5912 * GPU turbo limits to i915.
5915 ips_ping_for_i915_load(void)
5919 link
= symbol_get(ips_link_to_i915_driver
);
5922 symbol_put(ips_link_to_i915_driver
);
5926 void intel_gpu_ips_init(struct drm_i915_private
*dev_priv
)
5928 /* We only register the i915 ips part with intel-ips once everything is
5929 * set up, to avoid intel-ips sneaking in and reading bogus values. */
5930 spin_lock_irq(&mchdev_lock
);
5931 i915_mch_dev
= dev_priv
;
5932 spin_unlock_irq(&mchdev_lock
);
5934 ips_ping_for_i915_load();
5937 void intel_gpu_ips_teardown(void)
5939 spin_lock_irq(&mchdev_lock
);
5940 i915_mch_dev
= NULL
;
5941 spin_unlock_irq(&mchdev_lock
);
5944 static void intel_init_emon(struct drm_device
*dev
)
5946 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5951 /* Disable to program */
5955 /* Program energy weights for various events */
5956 I915_WRITE(SDEW
, 0x15040d00);
5957 I915_WRITE(CSIEW0
, 0x007f0000);
5958 I915_WRITE(CSIEW1
, 0x1e220004);
5959 I915_WRITE(CSIEW2
, 0x04000004);
5961 for (i
= 0; i
< 5; i
++)
5962 I915_WRITE(PEW(i
), 0);
5963 for (i
= 0; i
< 3; i
++)
5964 I915_WRITE(DEW(i
), 0);
5966 /* Program P-state weights to account for frequency power adjustment */
5967 for (i
= 0; i
< 16; i
++) {
5968 u32 pxvidfreq
= I915_READ(PXVFREQ(i
));
5969 unsigned long freq
= intel_pxfreq(pxvidfreq
);
5970 unsigned long vid
= (pxvidfreq
& PXVFREQ_PX_MASK
) >>
5975 val
*= (freq
/ 1000);
5977 val
/= (127*127*900);
5979 DRM_ERROR("bad pxval: %ld\n", val
);
5982 /* Render standby states get 0 weight */
5986 for (i
= 0; i
< 4; i
++) {
5987 u32 val
= (pxw
[i
*4] << 24) | (pxw
[(i
*4)+1] << 16) |
5988 (pxw
[(i
*4)+2] << 8) | (pxw
[(i
*4)+3]);
5989 I915_WRITE(PXW(i
), val
);
5992 /* Adjust magic regs to magic values (more experimental results) */
5993 I915_WRITE(OGW0
, 0);
5994 I915_WRITE(OGW1
, 0);
5995 I915_WRITE(EG0
, 0x00007f00);
5996 I915_WRITE(EG1
, 0x0000000e);
5997 I915_WRITE(EG2
, 0x000e0000);
5998 I915_WRITE(EG3
, 0x68000300);
5999 I915_WRITE(EG4
, 0x42000000);
6000 I915_WRITE(EG5
, 0x00140031);
6004 for (i
= 0; i
< 8; i
++)
6005 I915_WRITE(PXWL(i
), 0);
6007 /* Enable PMON + select events */
6008 I915_WRITE(ECR
, 0x80000019);
6010 lcfuse
= I915_READ(LCFUSE02
);
6012 dev_priv
->ips
.corr
= (lcfuse
& LCFUSE_HIV_MASK
);
6015 void intel_init_gt_powersave(struct drm_device
*dev
)
6017 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6019 i915
.enable_rc6
= sanitize_rc6_option(dev
, i915
.enable_rc6
);
6021 * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
6024 if (!i915
.enable_rc6
) {
6025 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
6026 intel_runtime_pm_get(dev_priv
);
6029 if (IS_CHERRYVIEW(dev
))
6030 cherryview_init_gt_powersave(dev
);
6031 else if (IS_VALLEYVIEW(dev
))
6032 valleyview_init_gt_powersave(dev
);
6035 void intel_cleanup_gt_powersave(struct drm_device
*dev
)
6037 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6039 if (IS_CHERRYVIEW(dev
))
6041 else if (IS_VALLEYVIEW(dev
))
6042 valleyview_cleanup_gt_powersave(dev
);
6044 if (!i915
.enable_rc6
)
6045 intel_runtime_pm_put(dev_priv
);
6048 static void gen6_suspend_rps(struct drm_device
*dev
)
6050 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6052 flush_delayed_work(&dev_priv
->rps
.delayed_resume_work
);
6054 gen6_disable_rps_interrupts(dev
);
6058 * intel_suspend_gt_powersave - suspend PM work and helper threads
6061 * We don't want to disable RC6 or other features here, we just want
6062 * to make sure any work we've queued has finished and won't bother
6063 * us while we're suspended.
6065 void intel_suspend_gt_powersave(struct drm_device
*dev
)
6067 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6069 if (INTEL_INFO(dev
)->gen
< 6)
6072 gen6_suspend_rps(dev
);
6074 /* Force GPU to min freq during suspend */
6075 gen6_rps_idle(dev_priv
);
6078 void intel_disable_gt_powersave(struct drm_device
*dev
)
6080 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6082 if (IS_IRONLAKE_M(dev
)) {
6083 ironlake_disable_drps(dev
);
6084 } else if (INTEL_INFO(dev
)->gen
>= 6) {
6085 intel_suspend_gt_powersave(dev
);
6087 mutex_lock(&dev_priv
->rps
.hw_lock
);
6088 if (INTEL_INFO(dev
)->gen
>= 9)
6089 gen9_disable_rps(dev
);
6090 else if (IS_CHERRYVIEW(dev
))
6091 cherryview_disable_rps(dev
);
6092 else if (IS_VALLEYVIEW(dev
))
6093 valleyview_disable_rps(dev
);
6095 gen6_disable_rps(dev
);
6097 dev_priv
->rps
.enabled
= false;
6098 mutex_unlock(&dev_priv
->rps
.hw_lock
);
6102 static void intel_gen6_powersave_work(struct work_struct
*work
)
6104 struct drm_i915_private
*dev_priv
=
6105 container_of(work
, struct drm_i915_private
,
6106 rps
.delayed_resume_work
.work
);
6107 struct drm_device
*dev
= dev_priv
->dev
;
6109 mutex_lock(&dev_priv
->rps
.hw_lock
);
6111 gen6_reset_rps_interrupts(dev
);
6113 if (IS_CHERRYVIEW(dev
)) {
6114 cherryview_enable_rps(dev
);
6115 } else if (IS_VALLEYVIEW(dev
)) {
6116 valleyview_enable_rps(dev
);
6117 } else if (INTEL_INFO(dev
)->gen
>= 9) {
6118 gen9_enable_rc6(dev
);
6119 gen9_enable_rps(dev
);
6120 if (IS_SKYLAKE(dev
) || IS_KABYLAKE(dev
))
6121 __gen6_update_ring_freq(dev
);
6122 } else if (IS_BROADWELL(dev
)) {
6123 gen8_enable_rps(dev
);
6124 __gen6_update_ring_freq(dev
);
6126 gen6_enable_rps(dev
);
6127 __gen6_update_ring_freq(dev
);
6130 WARN_ON(dev_priv
->rps
.max_freq
< dev_priv
->rps
.min_freq
);
6131 WARN_ON(dev_priv
->rps
.idle_freq
> dev_priv
->rps
.max_freq
);
6133 WARN_ON(dev_priv
->rps
.efficient_freq
< dev_priv
->rps
.min_freq
);
6134 WARN_ON(dev_priv
->rps
.efficient_freq
> dev_priv
->rps
.max_freq
);
6136 dev_priv
->rps
.enabled
= true;
6138 gen6_enable_rps_interrupts(dev
);
6140 mutex_unlock(&dev_priv
->rps
.hw_lock
);
6142 intel_runtime_pm_put(dev_priv
);
6145 void intel_enable_gt_powersave(struct drm_device
*dev
)
6147 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6149 /* Powersaving is controlled by the host when inside a VM */
6150 if (intel_vgpu_active(dev
))
6153 if (IS_IRONLAKE_M(dev
)) {
6154 mutex_lock(&dev
->struct_mutex
);
6155 ironlake_enable_drps(dev
);
6156 intel_init_emon(dev
);
6157 mutex_unlock(&dev
->struct_mutex
);
6158 } else if (INTEL_INFO(dev
)->gen
>= 6) {
6160 * PCU communication is slow and this doesn't need to be
6161 * done at any specific time, so do this out of our fast path
6162 * to make resume and init faster.
6164 * We depend on the HW RC6 power context save/restore
6165 * mechanism when entering D3 through runtime PM suspend. So
6166 * disable RPM until RPS/RC6 is properly setup. We can only
6167 * get here via the driver load/system resume/runtime resume
6168 * paths, so the _noresume version is enough (and in case of
6169 * runtime resume it's necessary).
6171 if (schedule_delayed_work(&dev_priv
->rps
.delayed_resume_work
,
6172 round_jiffies_up_relative(HZ
)))
6173 intel_runtime_pm_get_noresume(dev_priv
);
6177 void intel_reset_gt_powersave(struct drm_device
*dev
)
6179 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6181 if (INTEL_INFO(dev
)->gen
< 6)
6184 gen6_suspend_rps(dev
);
6185 dev_priv
->rps
.enabled
= false;
6188 static void ibx_init_clock_gating(struct drm_device
*dev
)
6190 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6193 * On Ibex Peak and Cougar Point, we need to disable clock
6194 * gating for the panel power sequencer or it will fail to
6195 * start up when no ports are active.
6197 I915_WRITE(SOUTH_DSPCLK_GATE_D
, PCH_DPLSUNIT_CLOCK_GATE_DISABLE
);
6200 static void g4x_disable_trickle_feed(struct drm_device
*dev
)
6202 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6205 for_each_pipe(dev_priv
, pipe
) {
6206 I915_WRITE(DSPCNTR(pipe
),
6207 I915_READ(DSPCNTR(pipe
)) |
6208 DISPPLANE_TRICKLE_FEED_DISABLE
);
6210 I915_WRITE(DSPSURF(pipe
), I915_READ(DSPSURF(pipe
)));
6211 POSTING_READ(DSPSURF(pipe
));
6215 static void ilk_init_lp_watermarks(struct drm_device
*dev
)
6217 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6219 I915_WRITE(WM3_LP_ILK
, I915_READ(WM3_LP_ILK
) & ~WM1_LP_SR_EN
);
6220 I915_WRITE(WM2_LP_ILK
, I915_READ(WM2_LP_ILK
) & ~WM1_LP_SR_EN
);
6221 I915_WRITE(WM1_LP_ILK
, I915_READ(WM1_LP_ILK
) & ~WM1_LP_SR_EN
);
6224 * Don't touch WM1S_LP_EN here.
6225 * Doing so could cause underruns.
6229 static void ironlake_init_clock_gating(struct drm_device
*dev
)
6231 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6232 uint32_t dspclk_gate
= ILK_VRHUNIT_CLOCK_GATE_DISABLE
;
6236 * WaFbcDisableDpfcClockGating:ilk
6238 dspclk_gate
|= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE
|
6239 ILK_DPFCUNIT_CLOCK_GATE_DISABLE
|
6240 ILK_DPFDUNIT_CLOCK_GATE_ENABLE
;
6242 I915_WRITE(PCH_3DCGDIS0
,
6243 MARIUNIT_CLOCK_GATE_DISABLE
|
6244 SVSMUNIT_CLOCK_GATE_DISABLE
);
6245 I915_WRITE(PCH_3DCGDIS1
,
6246 VFMUNIT_CLOCK_GATE_DISABLE
);
6249 * According to the spec the following bits should be set in
6250 * order to enable memory self-refresh
6251 * The bit 22/21 of 0x42004
6252 * The bit 5 of 0x42020
6253 * The bit 15 of 0x45000
6255 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6256 (I915_READ(ILK_DISPLAY_CHICKEN2
) |
6257 ILK_DPARB_GATE
| ILK_VSDPFD_FULL
));
6258 dspclk_gate
|= ILK_DPARBUNIT_CLOCK_GATE_ENABLE
;
6259 I915_WRITE(DISP_ARB_CTL
,
6260 (I915_READ(DISP_ARB_CTL
) |
6263 ilk_init_lp_watermarks(dev
);
6266 * Based on the document from hardware guys the following bits
6267 * should be set unconditionally in order to enable FBC.
6268 * The bit 22 of 0x42000
6269 * The bit 22 of 0x42004
6270 * The bit 7,8,9 of 0x42020.
6272 if (IS_IRONLAKE_M(dev
)) {
6273 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
6274 I915_WRITE(ILK_DISPLAY_CHICKEN1
,
6275 I915_READ(ILK_DISPLAY_CHICKEN1
) |
6277 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6278 I915_READ(ILK_DISPLAY_CHICKEN2
) |
6282 I915_WRITE(ILK_DSPCLK_GATE_D
, dspclk_gate
);
6284 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6285 I915_READ(ILK_DISPLAY_CHICKEN2
) |
6286 ILK_ELPIN_409_SELECT
);
6287 I915_WRITE(_3D_CHICKEN2
,
6288 _3D_CHICKEN2_WM_READ_PIPELINED
<< 16 |
6289 _3D_CHICKEN2_WM_READ_PIPELINED
);
6291 /* WaDisableRenderCachePipelinedFlush:ilk */
6292 I915_WRITE(CACHE_MODE_0
,
6293 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE
));
6295 /* WaDisable_RenderCache_OperationalFlush:ilk */
6296 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6298 g4x_disable_trickle_feed(dev
);
6300 ibx_init_clock_gating(dev
);
6303 static void cpt_init_clock_gating(struct drm_device
*dev
)
6305 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6310 * On Ibex Peak and Cougar Point, we need to disable clock
6311 * gating for the panel power sequencer or it will fail to
6312 * start up when no ports are active.
6314 I915_WRITE(SOUTH_DSPCLK_GATE_D
, PCH_DPLSUNIT_CLOCK_GATE_DISABLE
|
6315 PCH_DPLUNIT_CLOCK_GATE_DISABLE
|
6316 PCH_CPUNIT_CLOCK_GATE_DISABLE
);
6317 I915_WRITE(SOUTH_CHICKEN2
, I915_READ(SOUTH_CHICKEN2
) |
6318 DPLS_EDP_PPS_FIX_DIS
);
6319 /* The below fixes the weird display corruption, a few pixels shifted
6320 * downward, on (only) LVDS of some HP laptops with IVY.
6322 for_each_pipe(dev_priv
, pipe
) {
6323 val
= I915_READ(TRANS_CHICKEN2(pipe
));
6324 val
|= TRANS_CHICKEN2_TIMING_OVERRIDE
;
6325 val
&= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED
;
6326 if (dev_priv
->vbt
.fdi_rx_polarity_inverted
)
6327 val
|= TRANS_CHICKEN2_FDI_POLARITY_REVERSED
;
6328 val
&= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK
;
6329 val
&= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER
;
6330 val
&= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH
;
6331 I915_WRITE(TRANS_CHICKEN2(pipe
), val
);
6333 /* WADP0ClockGatingDisable */
6334 for_each_pipe(dev_priv
, pipe
) {
6335 I915_WRITE(TRANS_CHICKEN1(pipe
),
6336 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE
);
6340 static void gen6_check_mch_setup(struct drm_device
*dev
)
6342 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6345 tmp
= I915_READ(MCH_SSKPD
);
6346 if ((tmp
& MCH_SSKPD_WM0_MASK
) != MCH_SSKPD_WM0_VAL
)
6347 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
6351 static void gen6_init_clock_gating(struct drm_device
*dev
)
6353 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6354 uint32_t dspclk_gate
= ILK_VRHUNIT_CLOCK_GATE_DISABLE
;
6356 I915_WRITE(ILK_DSPCLK_GATE_D
, dspclk_gate
);
6358 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6359 I915_READ(ILK_DISPLAY_CHICKEN2
) |
6360 ILK_ELPIN_409_SELECT
);
6362 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
6363 I915_WRITE(_3D_CHICKEN
,
6364 _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB
));
6366 /* WaDisable_RenderCache_OperationalFlush:snb */
6367 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6370 * BSpec recoomends 8x4 when MSAA is used,
6371 * however in practice 16x4 seems fastest.
6373 * Note that PS/WM thread counts depend on the WIZ hashing
6374 * disable bit, which we don't touch here, but it's good
6375 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6377 I915_WRITE(GEN6_GT_MODE
,
6378 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
6380 ilk_init_lp_watermarks(dev
);
6382 I915_WRITE(CACHE_MODE_0
,
6383 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB
));
6385 I915_WRITE(GEN6_UCGCTL1
,
6386 I915_READ(GEN6_UCGCTL1
) |
6387 GEN6_BLBUNIT_CLOCK_GATE_DISABLE
|
6388 GEN6_CSUNIT_CLOCK_GATE_DISABLE
);
6390 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
6391 * gating disable must be set. Failure to set it results in
6392 * flickering pixels due to Z write ordering failures after
6393 * some amount of runtime in the Mesa "fire" demo, and Unigine
6394 * Sanctuary and Tropics, and apparently anything else with
6395 * alpha test or pixel discard.
6397 * According to the spec, bit 11 (RCCUNIT) must also be set,
6398 * but we didn't debug actual testcases to find it out.
6400 * WaDisableRCCUnitClockGating:snb
6401 * WaDisableRCPBUnitClockGating:snb
6403 I915_WRITE(GEN6_UCGCTL2
,
6404 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE
|
6405 GEN6_RCCUNIT_CLOCK_GATE_DISABLE
);
6407 /* WaStripsFansDisableFastClipPerformanceFix:snb */
6408 I915_WRITE(_3D_CHICKEN3
,
6409 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL
));
6413 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
6414 * 3DSTATE_SF number of SF output attributes is more than 16."
6416 I915_WRITE(_3D_CHICKEN3
,
6417 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH
));
6420 * According to the spec the following bits should be
6421 * set in order to enable memory self-refresh and fbc:
6422 * The bit21 and bit22 of 0x42000
6423 * The bit21 and bit22 of 0x42004
6424 * The bit5 and bit7 of 0x42020
6425 * The bit14 of 0x70180
6426 * The bit14 of 0x71180
6428 * WaFbcAsynchFlipDisableFbcQueue:snb
6430 I915_WRITE(ILK_DISPLAY_CHICKEN1
,
6431 I915_READ(ILK_DISPLAY_CHICKEN1
) |
6432 ILK_FBCQ_DIS
| ILK_PABSTRETCH_DIS
);
6433 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6434 I915_READ(ILK_DISPLAY_CHICKEN2
) |
6435 ILK_DPARB_GATE
| ILK_VSDPFD_FULL
);
6436 I915_WRITE(ILK_DSPCLK_GATE_D
,
6437 I915_READ(ILK_DSPCLK_GATE_D
) |
6438 ILK_DPARBUNIT_CLOCK_GATE_ENABLE
|
6439 ILK_DPFDUNIT_CLOCK_GATE_ENABLE
);
6441 g4x_disable_trickle_feed(dev
);
6443 cpt_init_clock_gating(dev
);
6445 gen6_check_mch_setup(dev
);
6448 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private
*dev_priv
)
6450 uint32_t reg
= I915_READ(GEN7_FF_THREAD_MODE
);
6453 * WaVSThreadDispatchOverride:ivb,vlv
6455 * This actually overrides the dispatch
6456 * mode for all thread types.
6458 reg
&= ~GEN7_FF_SCHED_MASK
;
6459 reg
|= GEN7_FF_TS_SCHED_HW
;
6460 reg
|= GEN7_FF_VS_SCHED_HW
;
6461 reg
|= GEN7_FF_DS_SCHED_HW
;
6463 I915_WRITE(GEN7_FF_THREAD_MODE
, reg
);
6466 static void lpt_init_clock_gating(struct drm_device
*dev
)
6468 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6471 * TODO: this bit should only be enabled when really needed, then
6472 * disabled when not needed anymore in order to save power.
6474 if (HAS_PCH_LPT_LP(dev
))
6475 I915_WRITE(SOUTH_DSPCLK_GATE_D
,
6476 I915_READ(SOUTH_DSPCLK_GATE_D
) |
6477 PCH_LP_PARTITION_LEVEL_DISABLE
);
6479 /* WADPOClockGatingDisable:hsw */
6480 I915_WRITE(TRANS_CHICKEN1(PIPE_A
),
6481 I915_READ(TRANS_CHICKEN1(PIPE_A
)) |
6482 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE
);
6485 static void lpt_suspend_hw(struct drm_device
*dev
)
6487 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6489 if (HAS_PCH_LPT_LP(dev
)) {
6490 uint32_t val
= I915_READ(SOUTH_DSPCLK_GATE_D
);
6492 val
&= ~PCH_LP_PARTITION_LEVEL_DISABLE
;
6493 I915_WRITE(SOUTH_DSPCLK_GATE_D
, val
);
6497 static void broadwell_init_clock_gating(struct drm_device
*dev
)
6499 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6503 ilk_init_lp_watermarks(dev
);
6505 /* WaSwitchSolVfFArbitrationPriority:bdw */
6506 I915_WRITE(GAM_ECOCHK
, I915_READ(GAM_ECOCHK
) | HSW_ECOCHK_ARB_PRIO_SOL
);
6508 /* WaPsrDPAMaskVBlankInSRD:bdw */
6509 I915_WRITE(CHICKEN_PAR1_1
,
6510 I915_READ(CHICKEN_PAR1_1
) | DPA_MASK_VBLANK_SRD
);
6512 /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
6513 for_each_pipe(dev_priv
, pipe
) {
6514 I915_WRITE(CHICKEN_PIPESL_1(pipe
),
6515 I915_READ(CHICKEN_PIPESL_1(pipe
)) |
6516 BDW_DPRS_MASK_VBLANK_SRD
);
6519 /* WaVSRefCountFullforceMissDisable:bdw */
6520 /* WaDSRefCountFullforceMissDisable:bdw */
6521 I915_WRITE(GEN7_FF_THREAD_MODE
,
6522 I915_READ(GEN7_FF_THREAD_MODE
) &
6523 ~(GEN8_FF_DS_REF_CNT_FFME
| GEN7_FF_VS_REF_CNT_FFME
));
6525 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL
,
6526 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE
));
6528 /* WaDisableSDEUnitClockGating:bdw */
6529 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
6530 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
6533 * WaProgramL3SqcReg1Default:bdw
6534 * WaTempDisableDOPClkGating:bdw
6536 misccpctl
= I915_READ(GEN7_MISCCPCTL
);
6537 I915_WRITE(GEN7_MISCCPCTL
, misccpctl
& ~GEN7_DOP_CLOCK_GATE_ENABLE
);
6538 I915_WRITE(GEN8_L3SQCREG1
, BDW_WA_L3SQCREG1_DEFAULT
);
6539 I915_WRITE(GEN7_MISCCPCTL
, misccpctl
);
6542 * WaGttCachingOffByDefault:bdw
6543 * GTT cache may not work with big pages, so if those
6544 * are ever enabled GTT cache may need to be disabled.
6546 I915_WRITE(HSW_GTT_CACHE_EN
, GTT_CACHE_EN_ALL
);
6548 lpt_init_clock_gating(dev
);
6551 static void haswell_init_clock_gating(struct drm_device
*dev
)
6553 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6555 ilk_init_lp_watermarks(dev
);
6557 /* L3 caching of data atomics doesn't work -- disable it. */
6558 I915_WRITE(HSW_SCRATCH1
, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE
);
6559 I915_WRITE(HSW_ROW_CHICKEN3
,
6560 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE
));
6562 /* This is required by WaCatErrorRejectionIssue:hsw */
6563 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
6564 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
6565 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
6567 /* WaVSRefCountFullforceMissDisable:hsw */
6568 I915_WRITE(GEN7_FF_THREAD_MODE
,
6569 I915_READ(GEN7_FF_THREAD_MODE
) & ~GEN7_FF_VS_REF_CNT_FFME
);
6571 /* WaDisable_RenderCache_OperationalFlush:hsw */
6572 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6574 /* enable HiZ Raw Stall Optimization */
6575 I915_WRITE(CACHE_MODE_0_GEN7
,
6576 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE
));
6578 /* WaDisable4x2SubspanOptimization:hsw */
6579 I915_WRITE(CACHE_MODE_1
,
6580 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
6583 * BSpec recommends 8x4 when MSAA is used,
6584 * however in practice 16x4 seems fastest.
6586 * Note that PS/WM thread counts depend on the WIZ hashing
6587 * disable bit, which we don't touch here, but it's good
6588 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6590 I915_WRITE(GEN7_GT_MODE
,
6591 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
6593 /* WaSampleCChickenBitEnable:hsw */
6594 I915_WRITE(HALF_SLICE_CHICKEN3
,
6595 _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE
));
6597 /* WaSwitchSolVfFArbitrationPriority:hsw */
6598 I915_WRITE(GAM_ECOCHK
, I915_READ(GAM_ECOCHK
) | HSW_ECOCHK_ARB_PRIO_SOL
);
6600 /* WaRsPkgCStateDisplayPMReq:hsw */
6601 I915_WRITE(CHICKEN_PAR1_1
,
6602 I915_READ(CHICKEN_PAR1_1
) | FORCE_ARB_IDLE_PLANES
);
6604 lpt_init_clock_gating(dev
);
6607 static void ivybridge_init_clock_gating(struct drm_device
*dev
)
6609 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6612 ilk_init_lp_watermarks(dev
);
6614 I915_WRITE(ILK_DSPCLK_GATE_D
, ILK_VRHUNIT_CLOCK_GATE_DISABLE
);
6616 /* WaDisableEarlyCull:ivb */
6617 I915_WRITE(_3D_CHICKEN3
,
6618 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL
));
6620 /* WaDisableBackToBackFlipFix:ivb */
6621 I915_WRITE(IVB_CHICKEN3
,
6622 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE
|
6623 CHICKEN3_DGMG_DONE_FIX_DISABLE
);
6625 /* WaDisablePSDDualDispatchEnable:ivb */
6626 if (IS_IVB_GT1(dev
))
6627 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1
,
6628 _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE
));
6630 /* WaDisable_RenderCache_OperationalFlush:ivb */
6631 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6633 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
6634 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1
,
6635 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC
);
6637 /* WaApplyL3ControlAndL3ChickenMode:ivb */
6638 I915_WRITE(GEN7_L3CNTLREG1
,
6639 GEN7_WA_FOR_GEN7_L3_CONTROL
);
6640 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER
,
6641 GEN7_WA_L3_CHICKEN_MODE
);
6642 if (IS_IVB_GT1(dev
))
6643 I915_WRITE(GEN7_ROW_CHICKEN2
,
6644 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
6646 /* must write both registers */
6647 I915_WRITE(GEN7_ROW_CHICKEN2
,
6648 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
6649 I915_WRITE(GEN7_ROW_CHICKEN2_GT2
,
6650 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
6653 /* WaForceL3Serialization:ivb */
6654 I915_WRITE(GEN7_L3SQCREG4
, I915_READ(GEN7_L3SQCREG4
) &
6655 ~L3SQ_URB_READ_CAM_MATCH_DISABLE
);
6658 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6659 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
6661 I915_WRITE(GEN6_UCGCTL2
,
6662 GEN6_RCZUNIT_CLOCK_GATE_DISABLE
);
6664 /* This is required by WaCatErrorRejectionIssue:ivb */
6665 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
6666 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
6667 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
6669 g4x_disable_trickle_feed(dev
);
6671 gen7_setup_fixed_func_scheduler(dev_priv
);
6673 if (0) { /* causes HiZ corruption on ivb:gt1 */
6674 /* enable HiZ Raw Stall Optimization */
6675 I915_WRITE(CACHE_MODE_0_GEN7
,
6676 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE
));
6679 /* WaDisable4x2SubspanOptimization:ivb */
6680 I915_WRITE(CACHE_MODE_1
,
6681 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
6684 * BSpec recommends 8x4 when MSAA is used,
6685 * however in practice 16x4 seems fastest.
6687 * Note that PS/WM thread counts depend on the WIZ hashing
6688 * disable bit, which we don't touch here, but it's good
6689 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6691 I915_WRITE(GEN7_GT_MODE
,
6692 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
6694 snpcr
= I915_READ(GEN6_MBCUNIT_SNPCR
);
6695 snpcr
&= ~GEN6_MBC_SNPCR_MASK
;
6696 snpcr
|= GEN6_MBC_SNPCR_MED
;
6697 I915_WRITE(GEN6_MBCUNIT_SNPCR
, snpcr
);
6699 if (!HAS_PCH_NOP(dev
))
6700 cpt_init_clock_gating(dev
);
6702 gen6_check_mch_setup(dev
);
6705 static void vlv_init_display_clock_gating(struct drm_i915_private
*dev_priv
)
6707 I915_WRITE(DSPCLK_GATE_D
, VRHUNIT_CLOCK_GATE_DISABLE
);
6710 * Disable trickle feed and enable pnd deadline calculation
6712 I915_WRITE(MI_ARB_VLV
, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
);
6713 I915_WRITE(CBR1_VLV
, 0);
6716 static void valleyview_init_clock_gating(struct drm_device
*dev
)
6718 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6720 vlv_init_display_clock_gating(dev_priv
);
6722 /* WaDisableEarlyCull:vlv */
6723 I915_WRITE(_3D_CHICKEN3
,
6724 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL
));
6726 /* WaDisableBackToBackFlipFix:vlv */
6727 I915_WRITE(IVB_CHICKEN3
,
6728 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE
|
6729 CHICKEN3_DGMG_DONE_FIX_DISABLE
);
6731 /* WaPsdDispatchEnable:vlv */
6732 /* WaDisablePSDDualDispatchEnable:vlv */
6733 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1
,
6734 _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP
|
6735 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE
));
6737 /* WaDisable_RenderCache_OperationalFlush:vlv */
6738 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6740 /* WaForceL3Serialization:vlv */
6741 I915_WRITE(GEN7_L3SQCREG4
, I915_READ(GEN7_L3SQCREG4
) &
6742 ~L3SQ_URB_READ_CAM_MATCH_DISABLE
);
6744 /* WaDisableDopClockGating:vlv */
6745 I915_WRITE(GEN7_ROW_CHICKEN2
,
6746 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
6748 /* This is required by WaCatErrorRejectionIssue:vlv */
6749 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
6750 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
6751 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
6753 gen7_setup_fixed_func_scheduler(dev_priv
);
6756 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6757 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
6759 I915_WRITE(GEN6_UCGCTL2
,
6760 GEN6_RCZUNIT_CLOCK_GATE_DISABLE
);
6762 /* WaDisableL3Bank2xClockGate:vlv
6763 * Disabling L3 clock gating- MMIO 940c[25] = 1
6764 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
6765 I915_WRITE(GEN7_UCGCTL4
,
6766 I915_READ(GEN7_UCGCTL4
) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE
);
6769 * BSpec says this must be set, even though
6770 * WaDisable4x2SubspanOptimization isn't listed for VLV.
6772 I915_WRITE(CACHE_MODE_1
,
6773 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
6776 * BSpec recommends 8x4 when MSAA is used,
6777 * however in practice 16x4 seems fastest.
6779 * Note that PS/WM thread counts depend on the WIZ hashing
6780 * disable bit, which we don't touch here, but it's good
6781 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6783 I915_WRITE(GEN7_GT_MODE
,
6784 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
6787 * WaIncreaseL3CreditsForVLVB0:vlv
6788 * This is the hardware default actually.
6790 I915_WRITE(GEN7_L3SQCREG1
, VLV_B0_WA_L3SQCREG1_VALUE
);
6793 * WaDisableVLVClockGating_VBIIssue:vlv
6794 * Disable clock gating on th GCFG unit to prevent a delay
6795 * in the reporting of vblank events.
6797 I915_WRITE(VLV_GUNIT_CLOCK_GATE
, GCFG_DIS
);
6800 static void cherryview_init_clock_gating(struct drm_device
*dev
)
6802 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6804 vlv_init_display_clock_gating(dev_priv
);
6806 /* WaVSRefCountFullforceMissDisable:chv */
6807 /* WaDSRefCountFullforceMissDisable:chv */
6808 I915_WRITE(GEN7_FF_THREAD_MODE
,
6809 I915_READ(GEN7_FF_THREAD_MODE
) &
6810 ~(GEN8_FF_DS_REF_CNT_FFME
| GEN7_FF_VS_REF_CNT_FFME
));
6812 /* WaDisableSemaphoreAndSyncFlipWait:chv */
6813 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL
,
6814 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE
));
6816 /* WaDisableCSUnitClockGating:chv */
6817 I915_WRITE(GEN6_UCGCTL1
, I915_READ(GEN6_UCGCTL1
) |
6818 GEN6_CSUNIT_CLOCK_GATE_DISABLE
);
6820 /* WaDisableSDEUnitClockGating:chv */
6821 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
6822 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
6825 * GTT cache may not work with big pages, so if those
6826 * are ever enabled GTT cache may need to be disabled.
6828 I915_WRITE(HSW_GTT_CACHE_EN
, GTT_CACHE_EN_ALL
);
6831 static void g4x_init_clock_gating(struct drm_device
*dev
)
6833 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6834 uint32_t dspclk_gate
;
6836 I915_WRITE(RENCLK_GATE_D1
, 0);
6837 I915_WRITE(RENCLK_GATE_D2
, VF_UNIT_CLOCK_GATE_DISABLE
|
6838 GS_UNIT_CLOCK_GATE_DISABLE
|
6839 CL_UNIT_CLOCK_GATE_DISABLE
);
6840 I915_WRITE(RAMCLK_GATE_D
, 0);
6841 dspclk_gate
= VRHUNIT_CLOCK_GATE_DISABLE
|
6842 OVRUNIT_CLOCK_GATE_DISABLE
|
6843 OVCUNIT_CLOCK_GATE_DISABLE
;
6845 dspclk_gate
|= DSSUNIT_CLOCK_GATE_DISABLE
;
6846 I915_WRITE(DSPCLK_GATE_D
, dspclk_gate
);
6848 /* WaDisableRenderCachePipelinedFlush */
6849 I915_WRITE(CACHE_MODE_0
,
6850 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE
));
6852 /* WaDisable_RenderCache_OperationalFlush:g4x */
6853 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6855 g4x_disable_trickle_feed(dev
);
6858 static void crestline_init_clock_gating(struct drm_device
*dev
)
6860 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6862 I915_WRITE(RENCLK_GATE_D1
, I965_RCC_CLOCK_GATE_DISABLE
);
6863 I915_WRITE(RENCLK_GATE_D2
, 0);
6864 I915_WRITE(DSPCLK_GATE_D
, 0);
6865 I915_WRITE(RAMCLK_GATE_D
, 0);
6866 I915_WRITE16(DEUC
, 0);
6867 I915_WRITE(MI_ARB_STATE
,
6868 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
6870 /* WaDisable_RenderCache_OperationalFlush:gen4 */
6871 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6874 static void broadwater_init_clock_gating(struct drm_device
*dev
)
6876 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6878 I915_WRITE(RENCLK_GATE_D1
, I965_RCZ_CLOCK_GATE_DISABLE
|
6879 I965_RCC_CLOCK_GATE_DISABLE
|
6880 I965_RCPB_CLOCK_GATE_DISABLE
|
6881 I965_ISC_CLOCK_GATE_DISABLE
|
6882 I965_FBC_CLOCK_GATE_DISABLE
);
6883 I915_WRITE(RENCLK_GATE_D2
, 0);
6884 I915_WRITE(MI_ARB_STATE
,
6885 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
6887 /* WaDisable_RenderCache_OperationalFlush:gen4 */
6888 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6891 static void gen3_init_clock_gating(struct drm_device
*dev
)
6893 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6894 u32 dstate
= I915_READ(D_STATE
);
6896 dstate
|= DSTATE_PLL_D3_OFF
| DSTATE_GFX_CLOCK_GATING
|
6897 DSTATE_DOT_CLOCK_GATING
;
6898 I915_WRITE(D_STATE
, dstate
);
6900 if (IS_PINEVIEW(dev
))
6901 I915_WRITE(ECOSKPD
, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY
));
6903 /* IIR "flip pending" means done if this bit is set */
6904 I915_WRITE(ECOSKPD
, _MASKED_BIT_DISABLE(ECO_FLIP_DONE
));
6906 /* interrupts should cause a wake up from C3 */
6907 I915_WRITE(INSTPM
, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN
));
6909 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
6910 I915_WRITE(MI_ARB_STATE
, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE
));
6912 I915_WRITE(MI_ARB_STATE
,
6913 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
6916 static void i85x_init_clock_gating(struct drm_device
*dev
)
6918 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6920 I915_WRITE(RENCLK_GATE_D1
, SV_CLOCK_GATE_DISABLE
);
6922 /* interrupts should cause a wake up from C3 */
6923 I915_WRITE(MI_STATE
, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN
) |
6924 _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE
));
6926 I915_WRITE(MEM_MODE
,
6927 _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE
));
6930 static void i830_init_clock_gating(struct drm_device
*dev
)
6932 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6934 I915_WRITE(DSPCLK_GATE_D
, OVRUNIT_CLOCK_GATE_DISABLE
);
6936 I915_WRITE(MEM_MODE
,
6937 _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE
) |
6938 _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE
));
6941 void intel_init_clock_gating(struct drm_device
*dev
)
6943 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6945 if (dev_priv
->display
.init_clock_gating
)
6946 dev_priv
->display
.init_clock_gating(dev
);
6949 void intel_suspend_hw(struct drm_device
*dev
)
6951 if (HAS_PCH_LPT(dev
))
6952 lpt_suspend_hw(dev
);
6955 /* Set up chip specific power management-related functions */
6956 void intel_init_pm(struct drm_device
*dev
)
6958 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6960 intel_fbc_init(dev_priv
);
6963 if (IS_PINEVIEW(dev
))
6964 i915_pineview_get_mem_freq(dev
);
6965 else if (IS_GEN5(dev
))
6966 i915_ironlake_get_mem_freq(dev
);
6968 /* For FIFO watermark updates */
6969 if (INTEL_INFO(dev
)->gen
>= 9) {
6970 skl_setup_wm_latency(dev
);
6972 if (IS_BROXTON(dev
))
6973 dev_priv
->display
.init_clock_gating
=
6974 bxt_init_clock_gating
;
6975 dev_priv
->display
.update_wm
= skl_update_wm
;
6976 } else if (HAS_PCH_SPLIT(dev
)) {
6977 ilk_setup_wm_latency(dev
);
6979 if ((IS_GEN5(dev
) && dev_priv
->wm
.pri_latency
[1] &&
6980 dev_priv
->wm
.spr_latency
[1] && dev_priv
->wm
.cur_latency
[1]) ||
6981 (!IS_GEN5(dev
) && dev_priv
->wm
.pri_latency
[0] &&
6982 dev_priv
->wm
.spr_latency
[0] && dev_priv
->wm
.cur_latency
[0])) {
6983 dev_priv
->display
.update_wm
= ilk_update_wm
;
6984 dev_priv
->display
.compute_pipe_wm
= ilk_compute_pipe_wm
;
6986 DRM_DEBUG_KMS("Failed to read display plane latency. "
6991 dev_priv
->display
.init_clock_gating
= ironlake_init_clock_gating
;
6992 else if (IS_GEN6(dev
))
6993 dev_priv
->display
.init_clock_gating
= gen6_init_clock_gating
;
6994 else if (IS_IVYBRIDGE(dev
))
6995 dev_priv
->display
.init_clock_gating
= ivybridge_init_clock_gating
;
6996 else if (IS_HASWELL(dev
))
6997 dev_priv
->display
.init_clock_gating
= haswell_init_clock_gating
;
6998 else if (INTEL_INFO(dev
)->gen
== 8)
6999 dev_priv
->display
.init_clock_gating
= broadwell_init_clock_gating
;
7000 } else if (IS_CHERRYVIEW(dev
)) {
7001 vlv_setup_wm_latency(dev
);
7003 dev_priv
->display
.update_wm
= vlv_update_wm
;
7004 dev_priv
->display
.init_clock_gating
=
7005 cherryview_init_clock_gating
;
7006 } else if (IS_VALLEYVIEW(dev
)) {
7007 vlv_setup_wm_latency(dev
);
7009 dev_priv
->display
.update_wm
= vlv_update_wm
;
7010 dev_priv
->display
.init_clock_gating
=
7011 valleyview_init_clock_gating
;
7012 } else if (IS_PINEVIEW(dev
)) {
7013 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev
),
7016 dev_priv
->mem_freq
)) {
7017 DRM_INFO("failed to find known CxSR latency "
7018 "(found ddr%s fsb freq %d, mem freq %d), "
7020 (dev_priv
->is_ddr3
== 1) ? "3" : "2",
7021 dev_priv
->fsb_freq
, dev_priv
->mem_freq
);
7022 /* Disable CxSR and never update its watermark again */
7023 intel_set_memory_cxsr(dev_priv
, false);
7024 dev_priv
->display
.update_wm
= NULL
;
7026 dev_priv
->display
.update_wm
= pineview_update_wm
;
7027 dev_priv
->display
.init_clock_gating
= gen3_init_clock_gating
;
7028 } else if (IS_G4X(dev
)) {
7029 dev_priv
->display
.update_wm
= g4x_update_wm
;
7030 dev_priv
->display
.init_clock_gating
= g4x_init_clock_gating
;
7031 } else if (IS_GEN4(dev
)) {
7032 dev_priv
->display
.update_wm
= i965_update_wm
;
7033 if (IS_CRESTLINE(dev
))
7034 dev_priv
->display
.init_clock_gating
= crestline_init_clock_gating
;
7035 else if (IS_BROADWATER(dev
))
7036 dev_priv
->display
.init_clock_gating
= broadwater_init_clock_gating
;
7037 } else if (IS_GEN3(dev
)) {
7038 dev_priv
->display
.update_wm
= i9xx_update_wm
;
7039 dev_priv
->display
.get_fifo_size
= i9xx_get_fifo_size
;
7040 dev_priv
->display
.init_clock_gating
= gen3_init_clock_gating
;
7041 } else if (IS_GEN2(dev
)) {
7042 if (INTEL_INFO(dev
)->num_pipes
== 1) {
7043 dev_priv
->display
.update_wm
= i845_update_wm
;
7044 dev_priv
->display
.get_fifo_size
= i845_get_fifo_size
;
7046 dev_priv
->display
.update_wm
= i9xx_update_wm
;
7047 dev_priv
->display
.get_fifo_size
= i830_get_fifo_size
;
7050 if (IS_I85X(dev
) || IS_I865G(dev
))
7051 dev_priv
->display
.init_clock_gating
= i85x_init_clock_gating
;
7053 dev_priv
->display
.init_clock_gating
= i830_init_clock_gating
;
7055 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
7059 int sandybridge_pcode_read(struct drm_i915_private
*dev_priv
, u32 mbox
, u32
*val
)
7061 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
7063 if (I915_READ(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) {
7064 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
7068 I915_WRITE(GEN6_PCODE_DATA
, *val
);
7069 I915_WRITE(GEN6_PCODE_DATA1
, 0);
7070 I915_WRITE(GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
| mbox
);
7072 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) == 0,
7074 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox
);
7078 *val
= I915_READ(GEN6_PCODE_DATA
);
7079 I915_WRITE(GEN6_PCODE_DATA
, 0);
7084 int sandybridge_pcode_write(struct drm_i915_private
*dev_priv
, u32 mbox
, u32 val
)
7086 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
7088 if (I915_READ(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) {
7089 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
7093 I915_WRITE(GEN6_PCODE_DATA
, val
);
7094 I915_WRITE(GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
| mbox
);
7096 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) == 0,
7098 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox
);
7102 I915_WRITE(GEN6_PCODE_DATA
, 0);
7107 static int vlv_gpu_freq_div(unsigned int czclk_freq
)
7109 switch (czclk_freq
) {
7124 static int byt_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
7126 int div
, czclk_freq
= DIV_ROUND_CLOSEST(dev_priv
->czclk_freq
, 1000);
7128 div
= vlv_gpu_freq_div(czclk_freq
);
7132 return DIV_ROUND_CLOSEST(czclk_freq
* (val
+ 6 - 0xbd), div
);
7135 static int byt_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
7137 int mul
, czclk_freq
= DIV_ROUND_CLOSEST(dev_priv
->czclk_freq
, 1000);
7139 mul
= vlv_gpu_freq_div(czclk_freq
);
7143 return DIV_ROUND_CLOSEST(mul
* val
, czclk_freq
) + 0xbd - 6;
7146 static int chv_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
7148 int div
, czclk_freq
= DIV_ROUND_CLOSEST(dev_priv
->czclk_freq
, 1000);
7150 div
= vlv_gpu_freq_div(czclk_freq
) / 2;
7154 return DIV_ROUND_CLOSEST(czclk_freq
* val
, 2 * div
) / 2;
7157 static int chv_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
7159 int mul
, czclk_freq
= DIV_ROUND_CLOSEST(dev_priv
->czclk_freq
, 1000);
7161 mul
= vlv_gpu_freq_div(czclk_freq
) / 2;
7165 /* CHV needs even values */
7166 return DIV_ROUND_CLOSEST(val
* 2 * mul
, czclk_freq
) * 2;
7169 int intel_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
7171 if (IS_GEN9(dev_priv
->dev
))
7172 return DIV_ROUND_CLOSEST(val
* GT_FREQUENCY_MULTIPLIER
,
7174 else if (IS_CHERRYVIEW(dev_priv
->dev
))
7175 return chv_gpu_freq(dev_priv
, val
);
7176 else if (IS_VALLEYVIEW(dev_priv
->dev
))
7177 return byt_gpu_freq(dev_priv
, val
);
7179 return val
* GT_FREQUENCY_MULTIPLIER
;
7182 int intel_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
7184 if (IS_GEN9(dev_priv
->dev
))
7185 return DIV_ROUND_CLOSEST(val
* GEN9_FREQ_SCALER
,
7186 GT_FREQUENCY_MULTIPLIER
);
7187 else if (IS_CHERRYVIEW(dev_priv
->dev
))
7188 return chv_freq_opcode(dev_priv
, val
);
7189 else if (IS_VALLEYVIEW(dev_priv
->dev
))
7190 return byt_freq_opcode(dev_priv
, val
);
7192 return DIV_ROUND_CLOSEST(val
, GT_FREQUENCY_MULTIPLIER
);
7195 struct request_boost
{
7196 struct work_struct work
;
7197 struct drm_i915_gem_request
*req
;
7200 static void __intel_rps_boost_work(struct work_struct
*work
)
7202 struct request_boost
*boost
= container_of(work
, struct request_boost
, work
);
7203 struct drm_i915_gem_request
*req
= boost
->req
;
7205 if (!i915_gem_request_completed(req
, true))
7206 gen6_rps_boost(to_i915(req
->ring
->dev
), NULL
,
7207 req
->emitted_jiffies
);
7209 i915_gem_request_unreference__unlocked(req
);
7213 void intel_queue_rps_boost_for_request(struct drm_device
*dev
,
7214 struct drm_i915_gem_request
*req
)
7216 struct request_boost
*boost
;
7218 if (req
== NULL
|| INTEL_INFO(dev
)->gen
< 6)
7221 if (i915_gem_request_completed(req
, true))
7224 boost
= kmalloc(sizeof(*boost
), GFP_ATOMIC
);
7228 i915_gem_request_reference(req
);
7231 INIT_WORK(&boost
->work
, __intel_rps_boost_work
);
7232 queue_work(to_i915(dev
)->wq
, &boost
->work
);
7235 void intel_pm_setup(struct drm_device
*dev
)
7237 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
7239 mutex_init(&dev_priv
->rps
.hw_lock
);
7240 spin_lock_init(&dev_priv
->rps
.client_lock
);
7242 INIT_DELAYED_WORK(&dev_priv
->rps
.delayed_resume_work
,
7243 intel_gen6_powersave_work
);
7244 INIT_LIST_HEAD(&dev_priv
->rps
.clients
);
7245 INIT_LIST_HEAD(&dev_priv
->rps
.semaphores
.link
);
7246 INIT_LIST_HEAD(&dev_priv
->rps
.mmioflips
.link
);
7248 dev_priv
->pm
.suspended
= false;
7249 atomic_set(&dev_priv
->pm
.wakeref_count
, 0);