2 * Copyright © 2012 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eugeni Dodonov <eugeni.dodonov@intel.com>
28 #include <linux/cpufreq.h>
30 #include "intel_drv.h"
31 #include "../../../platform/x86/intel_ips.h"
32 #include <linux/module.h>
35 * RC6 is a special power stage which allows the GPU to enter an very
36 * low-voltage mode when idle, using down to 0V while at this stage. This
37 * stage is entered automatically when the GPU is idle when RC6 support is
38 * enabled, and as soon as new workload arises GPU wakes up automatically as well.
40 * There are different RC6 modes available in Intel GPU, which differentiate
41 * among each other with the latency required to enter and leave RC6 and
42 * voltage consumed by the GPU in different states.
44 * The combination of the following flags define which states GPU is allowed
45 * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
46 * RC6pp is deepest RC6. Their support by hardware varies according to the
47 * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
48 * which brings the most power savings; deeper states save more power, but
49 * require higher latency to switch to and wake up.
51 #define INTEL_RC6_ENABLE (1<<0)
52 #define INTEL_RC6p_ENABLE (1<<1)
53 #define INTEL_RC6pp_ENABLE (1<<2)
55 static void gen9_init_clock_gating(struct drm_device
*dev
)
57 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
59 /* WaEnableLbsSlaRetryTimerDecrement:skl */
60 I915_WRITE(BDW_SCRATCH1
, I915_READ(BDW_SCRATCH1
) |
61 GEN9_LBS_SLA_RETRY_TIMER_DECREMENT_ENABLE
);
64 static void skl_init_clock_gating(struct drm_device
*dev
)
66 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
68 gen9_init_clock_gating(dev
);
70 if (INTEL_REVID(dev
) <= SKL_REVID_B0
) {
72 * WaDisableSDEUnitClockGating:skl
73 * WaSetGAPSunitClckGateDisable:skl
75 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
76 GEN8_GAPSUNIT_CLOCK_GATE_DISABLE
|
77 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
79 /* WaDisableVFUnitClockGating:skl */
80 I915_WRITE(GEN6_UCGCTL2
, I915_READ(GEN6_UCGCTL2
) |
81 GEN6_VFUNIT_CLOCK_GATE_DISABLE
);
84 if (INTEL_REVID(dev
) <= SKL_REVID_D0
) {
85 /* WaDisableHDCInvalidation:skl */
86 I915_WRITE(GAM_ECOCHK
, I915_READ(GAM_ECOCHK
) |
87 BDW_DISABLE_HDC_INVALIDATION
);
89 /* WaDisableChickenBitTSGBarrierAckForFFSliceCS:skl */
90 I915_WRITE(FF_SLICE_CS_CHICKEN2
,
91 _MASKED_BIT_ENABLE(GEN9_TSG_BARRIER_ACK_DISABLE
));
94 if (INTEL_REVID(dev
) <= SKL_REVID_E0
)
95 /* WaDisableLSQCROPERFforOCL:skl */
96 I915_WRITE(GEN8_L3SQCREG4
, I915_READ(GEN8_L3SQCREG4
) |
97 GEN8_LQSC_RO_PERF_DIS
);
100 static void bxt_init_clock_gating(struct drm_device
*dev
)
102 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
104 gen9_init_clock_gating(dev
);
108 * GEN8_SDEUNIT_CLOCK_GATE_DISABLE applies on A0 only.
109 * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
111 /* WaDisableSDEUnitClockGating:bxt */
112 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
113 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
|
114 GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ
);
116 /* FIXME: apply on A0 only */
117 I915_WRITE(TILECTL
, I915_READ(TILECTL
) | TILECTL_TLBPF
);
120 static void i915_pineview_get_mem_freq(struct drm_device
*dev
)
122 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
125 tmp
= I915_READ(CLKCFG
);
127 switch (tmp
& CLKCFG_FSB_MASK
) {
129 dev_priv
->fsb_freq
= 533; /* 133*4 */
132 dev_priv
->fsb_freq
= 800; /* 200*4 */
135 dev_priv
->fsb_freq
= 667; /* 167*4 */
138 dev_priv
->fsb_freq
= 400; /* 100*4 */
142 switch (tmp
& CLKCFG_MEM_MASK
) {
144 dev_priv
->mem_freq
= 533;
147 dev_priv
->mem_freq
= 667;
150 dev_priv
->mem_freq
= 800;
154 /* detect pineview DDR3 setting */
155 tmp
= I915_READ(CSHRDDR3CTL
);
156 dev_priv
->is_ddr3
= (tmp
& CSHRDDR3CTL_DDR3
) ? 1 : 0;
159 static void i915_ironlake_get_mem_freq(struct drm_device
*dev
)
161 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
164 ddrpll
= I915_READ16(DDRMPLL1
);
165 csipll
= I915_READ16(CSIPLL0
);
167 switch (ddrpll
& 0xff) {
169 dev_priv
->mem_freq
= 800;
172 dev_priv
->mem_freq
= 1066;
175 dev_priv
->mem_freq
= 1333;
178 dev_priv
->mem_freq
= 1600;
181 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
183 dev_priv
->mem_freq
= 0;
187 dev_priv
->ips
.r_t
= dev_priv
->mem_freq
;
189 switch (csipll
& 0x3ff) {
191 dev_priv
->fsb_freq
= 3200;
194 dev_priv
->fsb_freq
= 3733;
197 dev_priv
->fsb_freq
= 4266;
200 dev_priv
->fsb_freq
= 4800;
203 dev_priv
->fsb_freq
= 5333;
206 dev_priv
->fsb_freq
= 5866;
209 dev_priv
->fsb_freq
= 6400;
212 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
214 dev_priv
->fsb_freq
= 0;
218 if (dev_priv
->fsb_freq
== 3200) {
219 dev_priv
->ips
.c_m
= 0;
220 } else if (dev_priv
->fsb_freq
> 3200 && dev_priv
->fsb_freq
<= 4800) {
221 dev_priv
->ips
.c_m
= 1;
223 dev_priv
->ips
.c_m
= 2;
227 static const struct cxsr_latency cxsr_latency_table
[] = {
228 {1, 0, 800, 400, 3382, 33382, 3983, 33983}, /* DDR2-400 SC */
229 {1, 0, 800, 667, 3354, 33354, 3807, 33807}, /* DDR2-667 SC */
230 {1, 0, 800, 800, 3347, 33347, 3763, 33763}, /* DDR2-800 SC */
231 {1, 1, 800, 667, 6420, 36420, 6873, 36873}, /* DDR3-667 SC */
232 {1, 1, 800, 800, 5902, 35902, 6318, 36318}, /* DDR3-800 SC */
234 {1, 0, 667, 400, 3400, 33400, 4021, 34021}, /* DDR2-400 SC */
235 {1, 0, 667, 667, 3372, 33372, 3845, 33845}, /* DDR2-667 SC */
236 {1, 0, 667, 800, 3386, 33386, 3822, 33822}, /* DDR2-800 SC */
237 {1, 1, 667, 667, 6438, 36438, 6911, 36911}, /* DDR3-667 SC */
238 {1, 1, 667, 800, 5941, 35941, 6377, 36377}, /* DDR3-800 SC */
240 {1, 0, 400, 400, 3472, 33472, 4173, 34173}, /* DDR2-400 SC */
241 {1, 0, 400, 667, 3443, 33443, 3996, 33996}, /* DDR2-667 SC */
242 {1, 0, 400, 800, 3430, 33430, 3946, 33946}, /* DDR2-800 SC */
243 {1, 1, 400, 667, 6509, 36509, 7062, 37062}, /* DDR3-667 SC */
244 {1, 1, 400, 800, 5985, 35985, 6501, 36501}, /* DDR3-800 SC */
246 {0, 0, 800, 400, 3438, 33438, 4065, 34065}, /* DDR2-400 SC */
247 {0, 0, 800, 667, 3410, 33410, 3889, 33889}, /* DDR2-667 SC */
248 {0, 0, 800, 800, 3403, 33403, 3845, 33845}, /* DDR2-800 SC */
249 {0, 1, 800, 667, 6476, 36476, 6955, 36955}, /* DDR3-667 SC */
250 {0, 1, 800, 800, 5958, 35958, 6400, 36400}, /* DDR3-800 SC */
252 {0, 0, 667, 400, 3456, 33456, 4103, 34106}, /* DDR2-400 SC */
253 {0, 0, 667, 667, 3428, 33428, 3927, 33927}, /* DDR2-667 SC */
254 {0, 0, 667, 800, 3443, 33443, 3905, 33905}, /* DDR2-800 SC */
255 {0, 1, 667, 667, 6494, 36494, 6993, 36993}, /* DDR3-667 SC */
256 {0, 1, 667, 800, 5998, 35998, 6460, 36460}, /* DDR3-800 SC */
258 {0, 0, 400, 400, 3528, 33528, 4255, 34255}, /* DDR2-400 SC */
259 {0, 0, 400, 667, 3500, 33500, 4079, 34079}, /* DDR2-667 SC */
260 {0, 0, 400, 800, 3487, 33487, 4029, 34029}, /* DDR2-800 SC */
261 {0, 1, 400, 667, 6566, 36566, 7145, 37145}, /* DDR3-667 SC */
262 {0, 1, 400, 800, 6042, 36042, 6584, 36584}, /* DDR3-800 SC */
265 static const struct cxsr_latency
*intel_get_cxsr_latency(int is_desktop
,
270 const struct cxsr_latency
*latency
;
273 if (fsb
== 0 || mem
== 0)
276 for (i
= 0; i
< ARRAY_SIZE(cxsr_latency_table
); i
++) {
277 latency
= &cxsr_latency_table
[i
];
278 if (is_desktop
== latency
->is_desktop
&&
279 is_ddr3
== latency
->is_ddr3
&&
280 fsb
== latency
->fsb_freq
&& mem
== latency
->mem_freq
)
284 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
289 static void chv_set_memory_dvfs(struct drm_i915_private
*dev_priv
, bool enable
)
293 mutex_lock(&dev_priv
->rps
.hw_lock
);
295 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
);
297 val
&= ~FORCE_DDR_HIGH_FREQ
;
299 val
|= FORCE_DDR_HIGH_FREQ
;
300 val
&= ~FORCE_DDR_LOW_FREQ
;
301 val
|= FORCE_DDR_FREQ_REQ_ACK
;
302 vlv_punit_write(dev_priv
, PUNIT_REG_DDR_SETUP2
, val
);
304 if (wait_for((vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
) &
305 FORCE_DDR_FREQ_REQ_ACK
) == 0, 3))
306 DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
308 mutex_unlock(&dev_priv
->rps
.hw_lock
);
311 static void chv_set_memory_pm5(struct drm_i915_private
*dev_priv
, bool enable
)
315 mutex_lock(&dev_priv
->rps
.hw_lock
);
317 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DSPFREQ
);
319 val
|= DSP_MAXFIFO_PM5_ENABLE
;
321 val
&= ~DSP_MAXFIFO_PM5_ENABLE
;
322 vlv_punit_write(dev_priv
, PUNIT_REG_DSPFREQ
, val
);
324 mutex_unlock(&dev_priv
->rps
.hw_lock
);
327 #define FW_WM(value, plane) \
328 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
330 void intel_set_memory_cxsr(struct drm_i915_private
*dev_priv
, bool enable
)
332 struct drm_device
*dev
= dev_priv
->dev
;
335 if (IS_VALLEYVIEW(dev
)) {
336 I915_WRITE(FW_BLC_SELF_VLV
, enable
? FW_CSPWRDWNEN
: 0);
337 if (IS_CHERRYVIEW(dev
))
338 chv_set_memory_pm5(dev_priv
, enable
);
339 } else if (IS_G4X(dev
) || IS_CRESTLINE(dev
)) {
340 I915_WRITE(FW_BLC_SELF
, enable
? FW_BLC_SELF_EN
: 0);
341 } else if (IS_PINEVIEW(dev
)) {
342 val
= I915_READ(DSPFW3
) & ~PINEVIEW_SELF_REFRESH_EN
;
343 val
|= enable
? PINEVIEW_SELF_REFRESH_EN
: 0;
344 I915_WRITE(DSPFW3
, val
);
345 } else if (IS_I945G(dev
) || IS_I945GM(dev
)) {
346 val
= enable
? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN
) :
347 _MASKED_BIT_DISABLE(FW_BLC_SELF_EN
);
348 I915_WRITE(FW_BLC_SELF
, val
);
349 } else if (IS_I915GM(dev
)) {
350 val
= enable
? _MASKED_BIT_ENABLE(INSTPM_SELF_EN
) :
351 _MASKED_BIT_DISABLE(INSTPM_SELF_EN
);
352 I915_WRITE(INSTPM
, val
);
357 DRM_DEBUG_KMS("memory self-refresh is %s\n",
358 enable
? "enabled" : "disabled");
363 * Latency for FIFO fetches is dependent on several factors:
364 * - memory configuration (speed, channels)
366 * - current MCH state
367 * It can be fairly high in some situations, so here we assume a fairly
368 * pessimal value. It's a tradeoff between extra memory fetches (if we
369 * set this value too high, the FIFO will fetch frequently to stay full)
370 * and power consumption (set it too low to save power and we might see
371 * FIFO underruns and display "flicker").
373 * A value of 5us seems to be a good balance; safe for very low end
374 * platforms but not overly aggressive on lower latency configs.
376 static const int pessimal_latency_ns
= 5000;
378 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
379 ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
381 static int vlv_get_fifo_size(struct drm_device
*dev
,
382 enum pipe pipe
, int plane
)
384 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
385 int sprite0_start
, sprite1_start
, size
;
388 uint32_t dsparb
, dsparb2
, dsparb3
;
390 dsparb
= I915_READ(DSPARB
);
391 dsparb2
= I915_READ(DSPARB2
);
392 sprite0_start
= VLV_FIFO_START(dsparb
, dsparb2
, 0, 0);
393 sprite1_start
= VLV_FIFO_START(dsparb
, dsparb2
, 8, 4);
396 dsparb
= I915_READ(DSPARB
);
397 dsparb2
= I915_READ(DSPARB2
);
398 sprite0_start
= VLV_FIFO_START(dsparb
, dsparb2
, 16, 8);
399 sprite1_start
= VLV_FIFO_START(dsparb
, dsparb2
, 24, 12);
402 dsparb2
= I915_READ(DSPARB2
);
403 dsparb3
= I915_READ(DSPARB3
);
404 sprite0_start
= VLV_FIFO_START(dsparb3
, dsparb2
, 0, 16);
405 sprite1_start
= VLV_FIFO_START(dsparb3
, dsparb2
, 8, 20);
413 size
= sprite0_start
;
416 size
= sprite1_start
- sprite0_start
;
419 size
= 512 - 1 - sprite1_start
;
425 DRM_DEBUG_KMS("Pipe %c %s %c FIFO size: %d\n",
426 pipe_name(pipe
), plane
== 0 ? "primary" : "sprite",
427 plane
== 0 ? plane_name(pipe
) : sprite_name(pipe
, plane
- 1),
433 static int i9xx_get_fifo_size(struct drm_device
*dev
, int plane
)
435 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
436 uint32_t dsparb
= I915_READ(DSPARB
);
439 size
= dsparb
& 0x7f;
441 size
= ((dsparb
>> DSPARB_CSTART_SHIFT
) & 0x7f) - size
;
443 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb
,
444 plane
? "B" : "A", size
);
449 static int i830_get_fifo_size(struct drm_device
*dev
, int plane
)
451 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
452 uint32_t dsparb
= I915_READ(DSPARB
);
455 size
= dsparb
& 0x1ff;
457 size
= ((dsparb
>> DSPARB_BEND_SHIFT
) & 0x1ff) - size
;
458 size
>>= 1; /* Convert to cachelines */
460 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb
,
461 plane
? "B" : "A", size
);
466 static int i845_get_fifo_size(struct drm_device
*dev
, int plane
)
468 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
469 uint32_t dsparb
= I915_READ(DSPARB
);
472 size
= dsparb
& 0x7f;
473 size
>>= 2; /* Convert to cachelines */
475 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb
,
482 /* Pineview has different values for various configs */
483 static const struct intel_watermark_params pineview_display_wm
= {
484 .fifo_size
= PINEVIEW_DISPLAY_FIFO
,
485 .max_wm
= PINEVIEW_MAX_WM
,
486 .default_wm
= PINEVIEW_DFT_WM
,
487 .guard_size
= PINEVIEW_GUARD_WM
,
488 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
490 static const struct intel_watermark_params pineview_display_hplloff_wm
= {
491 .fifo_size
= PINEVIEW_DISPLAY_FIFO
,
492 .max_wm
= PINEVIEW_MAX_WM
,
493 .default_wm
= PINEVIEW_DFT_HPLLOFF_WM
,
494 .guard_size
= PINEVIEW_GUARD_WM
,
495 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
497 static const struct intel_watermark_params pineview_cursor_wm
= {
498 .fifo_size
= PINEVIEW_CURSOR_FIFO
,
499 .max_wm
= PINEVIEW_CURSOR_MAX_WM
,
500 .default_wm
= PINEVIEW_CURSOR_DFT_WM
,
501 .guard_size
= PINEVIEW_CURSOR_GUARD_WM
,
502 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
504 static const struct intel_watermark_params pineview_cursor_hplloff_wm
= {
505 .fifo_size
= PINEVIEW_CURSOR_FIFO
,
506 .max_wm
= PINEVIEW_CURSOR_MAX_WM
,
507 .default_wm
= PINEVIEW_CURSOR_DFT_WM
,
508 .guard_size
= PINEVIEW_CURSOR_GUARD_WM
,
509 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
511 static const struct intel_watermark_params g4x_wm_info
= {
512 .fifo_size
= G4X_FIFO_SIZE
,
513 .max_wm
= G4X_MAX_WM
,
514 .default_wm
= G4X_MAX_WM
,
516 .cacheline_size
= G4X_FIFO_LINE_SIZE
,
518 static const struct intel_watermark_params g4x_cursor_wm_info
= {
519 .fifo_size
= I965_CURSOR_FIFO
,
520 .max_wm
= I965_CURSOR_MAX_WM
,
521 .default_wm
= I965_CURSOR_DFT_WM
,
523 .cacheline_size
= G4X_FIFO_LINE_SIZE
,
525 static const struct intel_watermark_params valleyview_wm_info
= {
526 .fifo_size
= VALLEYVIEW_FIFO_SIZE
,
527 .max_wm
= VALLEYVIEW_MAX_WM
,
528 .default_wm
= VALLEYVIEW_MAX_WM
,
530 .cacheline_size
= G4X_FIFO_LINE_SIZE
,
532 static const struct intel_watermark_params valleyview_cursor_wm_info
= {
533 .fifo_size
= I965_CURSOR_FIFO
,
534 .max_wm
= VALLEYVIEW_CURSOR_MAX_WM
,
535 .default_wm
= I965_CURSOR_DFT_WM
,
537 .cacheline_size
= G4X_FIFO_LINE_SIZE
,
539 static const struct intel_watermark_params i965_cursor_wm_info
= {
540 .fifo_size
= I965_CURSOR_FIFO
,
541 .max_wm
= I965_CURSOR_MAX_WM
,
542 .default_wm
= I965_CURSOR_DFT_WM
,
544 .cacheline_size
= I915_FIFO_LINE_SIZE
,
546 static const struct intel_watermark_params i945_wm_info
= {
547 .fifo_size
= I945_FIFO_SIZE
,
548 .max_wm
= I915_MAX_WM
,
551 .cacheline_size
= I915_FIFO_LINE_SIZE
,
553 static const struct intel_watermark_params i915_wm_info
= {
554 .fifo_size
= I915_FIFO_SIZE
,
555 .max_wm
= I915_MAX_WM
,
558 .cacheline_size
= I915_FIFO_LINE_SIZE
,
560 static const struct intel_watermark_params i830_a_wm_info
= {
561 .fifo_size
= I855GM_FIFO_SIZE
,
562 .max_wm
= I915_MAX_WM
,
565 .cacheline_size
= I830_FIFO_LINE_SIZE
,
567 static const struct intel_watermark_params i830_bc_wm_info
= {
568 .fifo_size
= I855GM_FIFO_SIZE
,
569 .max_wm
= I915_MAX_WM
/2,
572 .cacheline_size
= I830_FIFO_LINE_SIZE
,
574 static const struct intel_watermark_params i845_wm_info
= {
575 .fifo_size
= I830_FIFO_SIZE
,
576 .max_wm
= I915_MAX_WM
,
579 .cacheline_size
= I830_FIFO_LINE_SIZE
,
583 * intel_calculate_wm - calculate watermark level
584 * @clock_in_khz: pixel clock
585 * @wm: chip FIFO params
586 * @pixel_size: display pixel size
587 * @latency_ns: memory latency for the platform
589 * Calculate the watermark level (the level at which the display plane will
590 * start fetching from memory again). Each chip has a different display
591 * FIFO size and allocation, so the caller needs to figure that out and pass
592 * in the correct intel_watermark_params structure.
594 * As the pixel clock runs, the FIFO will be drained at a rate that depends
595 * on the pixel size. When it reaches the watermark level, it'll start
596 * fetching FIFO line sized based chunks from memory until the FIFO fills
597 * past the watermark point. If the FIFO drains completely, a FIFO underrun
598 * will occur, and a display engine hang could result.
600 static unsigned long intel_calculate_wm(unsigned long clock_in_khz
,
601 const struct intel_watermark_params
*wm
,
604 unsigned long latency_ns
)
606 long entries_required
, wm_size
;
609 * Note: we need to make sure we don't overflow for various clock &
611 * clocks go from a few thousand to several hundred thousand.
612 * latency is usually a few thousand
614 entries_required
= ((clock_in_khz
/ 1000) * pixel_size
* latency_ns
) /
616 entries_required
= DIV_ROUND_UP(entries_required
, wm
->cacheline_size
);
618 DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required
);
620 wm_size
= fifo_size
- (entries_required
+ wm
->guard_size
);
622 DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size
);
624 /* Don't promote wm_size to unsigned... */
625 if (wm_size
> (long)wm
->max_wm
)
626 wm_size
= wm
->max_wm
;
628 wm_size
= wm
->default_wm
;
631 * Bspec seems to indicate that the value shouldn't be lower than
632 * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
633 * Lets go for 8 which is the burst size since certain platforms
634 * already use a hardcoded 8 (which is what the spec says should be
643 static struct drm_crtc
*single_enabled_crtc(struct drm_device
*dev
)
645 struct drm_crtc
*crtc
, *enabled
= NULL
;
647 for_each_crtc(dev
, crtc
) {
648 if (intel_crtc_active(crtc
)) {
658 static void pineview_update_wm(struct drm_crtc
*unused_crtc
)
660 struct drm_device
*dev
= unused_crtc
->dev
;
661 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
662 struct drm_crtc
*crtc
;
663 const struct cxsr_latency
*latency
;
667 latency
= intel_get_cxsr_latency(IS_PINEVIEW_G(dev
), dev_priv
->is_ddr3
,
668 dev_priv
->fsb_freq
, dev_priv
->mem_freq
);
670 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
671 intel_set_memory_cxsr(dev_priv
, false);
675 crtc
= single_enabled_crtc(dev
);
677 const struct drm_display_mode
*adjusted_mode
;
678 int pixel_size
= crtc
->primary
->state
->fb
->bits_per_pixel
/ 8;
681 adjusted_mode
= &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
682 clock
= adjusted_mode
->crtc_clock
;
685 wm
= intel_calculate_wm(clock
, &pineview_display_wm
,
686 pineview_display_wm
.fifo_size
,
687 pixel_size
, latency
->display_sr
);
688 reg
= I915_READ(DSPFW1
);
689 reg
&= ~DSPFW_SR_MASK
;
690 reg
|= FW_WM(wm
, SR
);
691 I915_WRITE(DSPFW1
, reg
);
692 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg
);
695 wm
= intel_calculate_wm(clock
, &pineview_cursor_wm
,
696 pineview_display_wm
.fifo_size
,
697 pixel_size
, latency
->cursor_sr
);
698 reg
= I915_READ(DSPFW3
);
699 reg
&= ~DSPFW_CURSOR_SR_MASK
;
700 reg
|= FW_WM(wm
, CURSOR_SR
);
701 I915_WRITE(DSPFW3
, reg
);
703 /* Display HPLL off SR */
704 wm
= intel_calculate_wm(clock
, &pineview_display_hplloff_wm
,
705 pineview_display_hplloff_wm
.fifo_size
,
706 pixel_size
, latency
->display_hpll_disable
);
707 reg
= I915_READ(DSPFW3
);
708 reg
&= ~DSPFW_HPLL_SR_MASK
;
709 reg
|= FW_WM(wm
, HPLL_SR
);
710 I915_WRITE(DSPFW3
, reg
);
712 /* cursor HPLL off SR */
713 wm
= intel_calculate_wm(clock
, &pineview_cursor_hplloff_wm
,
714 pineview_display_hplloff_wm
.fifo_size
,
715 pixel_size
, latency
->cursor_hpll_disable
);
716 reg
= I915_READ(DSPFW3
);
717 reg
&= ~DSPFW_HPLL_CURSOR_MASK
;
718 reg
|= FW_WM(wm
, HPLL_CURSOR
);
719 I915_WRITE(DSPFW3
, reg
);
720 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg
);
722 intel_set_memory_cxsr(dev_priv
, true);
724 intel_set_memory_cxsr(dev_priv
, false);
728 static bool g4x_compute_wm0(struct drm_device
*dev
,
730 const struct intel_watermark_params
*display
,
731 int display_latency_ns
,
732 const struct intel_watermark_params
*cursor
,
733 int cursor_latency_ns
,
737 struct drm_crtc
*crtc
;
738 const struct drm_display_mode
*adjusted_mode
;
739 int htotal
, hdisplay
, clock
, pixel_size
;
740 int line_time_us
, line_count
;
741 int entries
, tlb_miss
;
743 crtc
= intel_get_crtc_for_plane(dev
, plane
);
744 if (!intel_crtc_active(crtc
)) {
745 *cursor_wm
= cursor
->guard_size
;
746 *plane_wm
= display
->guard_size
;
750 adjusted_mode
= &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
751 clock
= adjusted_mode
->crtc_clock
;
752 htotal
= adjusted_mode
->crtc_htotal
;
753 hdisplay
= to_intel_crtc(crtc
)->config
->pipe_src_w
;
754 pixel_size
= crtc
->primary
->state
->fb
->bits_per_pixel
/ 8;
756 /* Use the small buffer method to calculate plane watermark */
757 entries
= ((clock
* pixel_size
/ 1000) * display_latency_ns
) / 1000;
758 tlb_miss
= display
->fifo_size
*display
->cacheline_size
- hdisplay
* 8;
761 entries
= DIV_ROUND_UP(entries
, display
->cacheline_size
);
762 *plane_wm
= entries
+ display
->guard_size
;
763 if (*plane_wm
> (int)display
->max_wm
)
764 *plane_wm
= display
->max_wm
;
766 /* Use the large buffer method to calculate cursor watermark */
767 line_time_us
= max(htotal
* 1000 / clock
, 1);
768 line_count
= (cursor_latency_ns
/ line_time_us
+ 1000) / 1000;
769 entries
= line_count
* crtc
->cursor
->state
->crtc_w
* pixel_size
;
770 tlb_miss
= cursor
->fifo_size
*cursor
->cacheline_size
- hdisplay
* 8;
773 entries
= DIV_ROUND_UP(entries
, cursor
->cacheline_size
);
774 *cursor_wm
= entries
+ cursor
->guard_size
;
775 if (*cursor_wm
> (int)cursor
->max_wm
)
776 *cursor_wm
= (int)cursor
->max_wm
;
782 * Check the wm result.
784 * If any calculated watermark values is larger than the maximum value that
785 * can be programmed into the associated watermark register, that watermark
788 static bool g4x_check_srwm(struct drm_device
*dev
,
789 int display_wm
, int cursor_wm
,
790 const struct intel_watermark_params
*display
,
791 const struct intel_watermark_params
*cursor
)
793 DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
794 display_wm
, cursor_wm
);
796 if (display_wm
> display
->max_wm
) {
797 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
798 display_wm
, display
->max_wm
);
802 if (cursor_wm
> cursor
->max_wm
) {
803 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
804 cursor_wm
, cursor
->max_wm
);
808 if (!(display_wm
|| cursor_wm
)) {
809 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
816 static bool g4x_compute_srwm(struct drm_device
*dev
,
819 const struct intel_watermark_params
*display
,
820 const struct intel_watermark_params
*cursor
,
821 int *display_wm
, int *cursor_wm
)
823 struct drm_crtc
*crtc
;
824 const struct drm_display_mode
*adjusted_mode
;
825 int hdisplay
, htotal
, pixel_size
, clock
;
826 unsigned long line_time_us
;
827 int line_count
, line_size
;
832 *display_wm
= *cursor_wm
= 0;
836 crtc
= intel_get_crtc_for_plane(dev
, plane
);
837 adjusted_mode
= &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
838 clock
= adjusted_mode
->crtc_clock
;
839 htotal
= adjusted_mode
->crtc_htotal
;
840 hdisplay
= to_intel_crtc(crtc
)->config
->pipe_src_w
;
841 pixel_size
= crtc
->primary
->state
->fb
->bits_per_pixel
/ 8;
843 line_time_us
= max(htotal
* 1000 / clock
, 1);
844 line_count
= (latency_ns
/ line_time_us
+ 1000) / 1000;
845 line_size
= hdisplay
* pixel_size
;
847 /* Use the minimum of the small and large buffer method for primary */
848 small
= ((clock
* pixel_size
/ 1000) * latency_ns
) / 1000;
849 large
= line_count
* line_size
;
851 entries
= DIV_ROUND_UP(min(small
, large
), display
->cacheline_size
);
852 *display_wm
= entries
+ display
->guard_size
;
854 /* calculate the self-refresh watermark for display cursor */
855 entries
= line_count
* pixel_size
* crtc
->cursor
->state
->crtc_w
;
856 entries
= DIV_ROUND_UP(entries
, cursor
->cacheline_size
);
857 *cursor_wm
= entries
+ cursor
->guard_size
;
859 return g4x_check_srwm(dev
,
860 *display_wm
, *cursor_wm
,
864 #define FW_WM_VLV(value, plane) \
865 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
867 static void vlv_write_wm_values(struct intel_crtc
*crtc
,
868 const struct vlv_wm_values
*wm
)
870 struct drm_i915_private
*dev_priv
= to_i915(crtc
->base
.dev
);
871 enum pipe pipe
= crtc
->pipe
;
873 I915_WRITE(VLV_DDL(pipe
),
874 (wm
->ddl
[pipe
].cursor
<< DDL_CURSOR_SHIFT
) |
875 (wm
->ddl
[pipe
].sprite
[1] << DDL_SPRITE_SHIFT(1)) |
876 (wm
->ddl
[pipe
].sprite
[0] << DDL_SPRITE_SHIFT(0)) |
877 (wm
->ddl
[pipe
].primary
<< DDL_PLANE_SHIFT
));
880 FW_WM(wm
->sr
.plane
, SR
) |
881 FW_WM(wm
->pipe
[PIPE_B
].cursor
, CURSORB
) |
882 FW_WM_VLV(wm
->pipe
[PIPE_B
].primary
, PLANEB
) |
883 FW_WM_VLV(wm
->pipe
[PIPE_A
].primary
, PLANEA
));
885 FW_WM_VLV(wm
->pipe
[PIPE_A
].sprite
[1], SPRITEB
) |
886 FW_WM(wm
->pipe
[PIPE_A
].cursor
, CURSORA
) |
887 FW_WM_VLV(wm
->pipe
[PIPE_A
].sprite
[0], SPRITEA
));
889 FW_WM(wm
->sr
.cursor
, CURSOR_SR
));
891 if (IS_CHERRYVIEW(dev_priv
)) {
892 I915_WRITE(DSPFW7_CHV
,
893 FW_WM_VLV(wm
->pipe
[PIPE_B
].sprite
[1], SPRITED
) |
894 FW_WM_VLV(wm
->pipe
[PIPE_B
].sprite
[0], SPRITEC
));
895 I915_WRITE(DSPFW8_CHV
,
896 FW_WM_VLV(wm
->pipe
[PIPE_C
].sprite
[1], SPRITEF
) |
897 FW_WM_VLV(wm
->pipe
[PIPE_C
].sprite
[0], SPRITEE
));
898 I915_WRITE(DSPFW9_CHV
,
899 FW_WM_VLV(wm
->pipe
[PIPE_C
].primary
, PLANEC
) |
900 FW_WM(wm
->pipe
[PIPE_C
].cursor
, CURSORC
));
902 FW_WM(wm
->sr
.plane
>> 9, SR_HI
) |
903 FW_WM(wm
->pipe
[PIPE_C
].sprite
[1] >> 8, SPRITEF_HI
) |
904 FW_WM(wm
->pipe
[PIPE_C
].sprite
[0] >> 8, SPRITEE_HI
) |
905 FW_WM(wm
->pipe
[PIPE_C
].primary
>> 8, PLANEC_HI
) |
906 FW_WM(wm
->pipe
[PIPE_B
].sprite
[1] >> 8, SPRITED_HI
) |
907 FW_WM(wm
->pipe
[PIPE_B
].sprite
[0] >> 8, SPRITEC_HI
) |
908 FW_WM(wm
->pipe
[PIPE_B
].primary
>> 8, PLANEB_HI
) |
909 FW_WM(wm
->pipe
[PIPE_A
].sprite
[1] >> 8, SPRITEB_HI
) |
910 FW_WM(wm
->pipe
[PIPE_A
].sprite
[0] >> 8, SPRITEA_HI
) |
911 FW_WM(wm
->pipe
[PIPE_A
].primary
>> 8, PLANEA_HI
));
914 FW_WM_VLV(wm
->pipe
[PIPE_B
].sprite
[1], SPRITED
) |
915 FW_WM_VLV(wm
->pipe
[PIPE_B
].sprite
[0], SPRITEC
));
917 FW_WM(wm
->sr
.plane
>> 9, SR_HI
) |
918 FW_WM(wm
->pipe
[PIPE_B
].sprite
[1] >> 8, SPRITED_HI
) |
919 FW_WM(wm
->pipe
[PIPE_B
].sprite
[0] >> 8, SPRITEC_HI
) |
920 FW_WM(wm
->pipe
[PIPE_B
].primary
>> 8, PLANEB_HI
) |
921 FW_WM(wm
->pipe
[PIPE_A
].sprite
[1] >> 8, SPRITEB_HI
) |
922 FW_WM(wm
->pipe
[PIPE_A
].sprite
[0] >> 8, SPRITEA_HI
) |
923 FW_WM(wm
->pipe
[PIPE_A
].primary
>> 8, PLANEA_HI
));
926 POSTING_READ(DSPFW1
);
928 dev_priv
->wm
.vlv
= *wm
;
933 static uint8_t vlv_compute_drain_latency(struct drm_crtc
*crtc
,
934 struct drm_plane
*plane
)
936 struct drm_device
*dev
= crtc
->dev
;
937 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
938 int entries
, prec_mult
, drain_latency
, pixel_size
;
939 int clock
= intel_crtc
->config
->base
.adjusted_mode
.crtc_clock
;
940 const int high_precision
= IS_CHERRYVIEW(dev
) ? 16 : 64;
943 * FIXME the plane might have an fb
944 * but be invisible (eg. due to clipping)
946 if (!intel_crtc
->active
|| !plane
->state
->fb
)
949 if (WARN(clock
== 0, "Pixel clock is zero!\n"))
952 pixel_size
= drm_format_plane_cpp(plane
->state
->fb
->pixel_format
, 0);
954 if (WARN(pixel_size
== 0, "Pixel size is zero!\n"))
957 entries
= DIV_ROUND_UP(clock
, 1000) * pixel_size
;
959 prec_mult
= high_precision
;
960 drain_latency
= 64 * prec_mult
* 4 / entries
;
962 if (drain_latency
> DRAIN_LATENCY_MASK
) {
964 drain_latency
= 64 * prec_mult
* 4 / entries
;
967 if (drain_latency
> DRAIN_LATENCY_MASK
)
968 drain_latency
= DRAIN_LATENCY_MASK
;
970 return drain_latency
| (prec_mult
== high_precision
?
971 DDL_PRECISION_HIGH
: DDL_PRECISION_LOW
);
974 static int vlv_compute_wm(struct intel_crtc
*crtc
,
975 struct intel_plane
*plane
,
978 int clock
, entries
, pixel_size
;
981 * FIXME the plane might have an fb
982 * but be invisible (eg. due to clipping)
984 if (!crtc
->active
|| !plane
->base
.state
->fb
)
987 pixel_size
= drm_format_plane_cpp(plane
->base
.state
->fb
->pixel_format
, 0);
988 clock
= crtc
->config
->base
.adjusted_mode
.crtc_clock
;
990 entries
= DIV_ROUND_UP(clock
, 1000) * pixel_size
;
993 * Set up the watermark such that we don't start issuing memory
994 * requests until we are within PND's max deadline value (256us).
995 * Idea being to be idle as long as possible while still taking
996 * advatange of PND's deadline scheduling. The limit of 8
997 * cachelines (used when the FIFO will anyway drain in less time
998 * than 256us) should match what we would be done if trickle
1001 return fifo_size
- clamp(DIV_ROUND_UP(256 * entries
, 64), 0, fifo_size
- 8);
1004 static bool vlv_compute_sr_wm(struct drm_device
*dev
,
1005 struct vlv_wm_values
*wm
)
1007 struct drm_i915_private
*dev_priv
= to_i915(dev
);
1008 struct drm_crtc
*crtc
;
1009 enum pipe pipe
= INVALID_PIPE
;
1012 struct intel_plane
*plane
;
1014 wm
->sr
.cursor
= wm
->sr
.plane
= 0;
1016 crtc
= single_enabled_crtc(dev
);
1017 /* maxfifo not supported on pipe C */
1018 if (crtc
&& to_intel_crtc(crtc
)->pipe
!= PIPE_C
) {
1019 pipe
= to_intel_crtc(crtc
)->pipe
;
1020 num_planes
= !!wm
->pipe
[pipe
].primary
+
1021 !!wm
->pipe
[pipe
].sprite
[0] +
1022 !!wm
->pipe
[pipe
].sprite
[1];
1023 fifo_size
= INTEL_INFO(dev_priv
)->num_pipes
* 512 - 1;
1026 if (fifo_size
== 0 || num_planes
> 1)
1029 wm
->sr
.cursor
= vlv_compute_wm(to_intel_crtc(crtc
),
1030 to_intel_plane(crtc
->cursor
), 0x3f);
1032 list_for_each_entry(plane
, &dev
->mode_config
.plane_list
, base
.head
) {
1033 if (plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
)
1036 if (plane
->pipe
!= pipe
)
1039 wm
->sr
.plane
= vlv_compute_wm(to_intel_crtc(crtc
),
1041 if (wm
->sr
.plane
!= 0)
1048 static void valleyview_update_wm(struct drm_crtc
*crtc
)
1050 struct drm_device
*dev
= crtc
->dev
;
1051 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1052 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
1053 enum pipe pipe
= intel_crtc
->pipe
;
1055 struct vlv_wm_values wm
= dev_priv
->wm
.vlv
;
1057 wm
.ddl
[pipe
].primary
= vlv_compute_drain_latency(crtc
, crtc
->primary
);
1058 wm
.pipe
[pipe
].primary
= vlv_compute_wm(intel_crtc
,
1059 to_intel_plane(crtc
->primary
),
1060 vlv_get_fifo_size(dev
, pipe
, 0));
1062 wm
.ddl
[pipe
].cursor
= vlv_compute_drain_latency(crtc
, crtc
->cursor
);
1063 wm
.pipe
[pipe
].cursor
= vlv_compute_wm(intel_crtc
,
1064 to_intel_plane(crtc
->cursor
),
1067 cxsr_enabled
= vlv_compute_sr_wm(dev
, &wm
);
1069 if (memcmp(&wm
, &dev_priv
->wm
.vlv
, sizeof(wm
)) == 0)
1072 DRM_DEBUG_KMS("Setting FIFO watermarks - %c: plane=%d, cursor=%d, "
1073 "SR: plane=%d, cursor=%d\n", pipe_name(pipe
),
1074 wm
.pipe
[pipe
].primary
, wm
.pipe
[pipe
].cursor
,
1075 wm
.sr
.plane
, wm
.sr
.cursor
);
1078 * FIXME DDR DVFS introduces massive memory latencies which
1079 * are not known to system agent so any deadline specified
1080 * by the display may not be respected. To support DDR DVFS
1081 * the watermark code needs to be rewritten to essentially
1082 * bypass deadline mechanism and rely solely on the
1083 * watermarks. For now disable DDR DVFS.
1085 if (IS_CHERRYVIEW(dev_priv
))
1086 chv_set_memory_dvfs(dev_priv
, false);
1089 intel_set_memory_cxsr(dev_priv
, false);
1091 vlv_write_wm_values(intel_crtc
, &wm
);
1094 intel_set_memory_cxsr(dev_priv
, true);
1097 static void valleyview_update_sprite_wm(struct drm_plane
*plane
,
1098 struct drm_crtc
*crtc
,
1099 uint32_t sprite_width
,
1100 uint32_t sprite_height
,
1102 bool enabled
, bool scaled
)
1104 struct drm_device
*dev
= crtc
->dev
;
1105 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1106 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
1107 enum pipe pipe
= intel_crtc
->pipe
;
1108 int sprite
= to_intel_plane(plane
)->plane
;
1110 struct vlv_wm_values wm
= dev_priv
->wm
.vlv
;
1113 wm
.ddl
[pipe
].sprite
[sprite
] =
1114 vlv_compute_drain_latency(crtc
, plane
);
1116 wm
.pipe
[pipe
].sprite
[sprite
] =
1117 vlv_compute_wm(intel_crtc
,
1118 to_intel_plane(plane
),
1119 vlv_get_fifo_size(dev
, pipe
, sprite
+1));
1121 wm
.ddl
[pipe
].sprite
[sprite
] = 0;
1122 wm
.pipe
[pipe
].sprite
[sprite
] = 0;
1125 cxsr_enabled
= vlv_compute_sr_wm(dev
, &wm
);
1127 if (memcmp(&wm
, &dev_priv
->wm
.vlv
, sizeof(wm
)) == 0)
1130 DRM_DEBUG_KMS("Setting FIFO watermarks - %c: sprite %c=%d, "
1131 "SR: plane=%d, cursor=%d\n", pipe_name(pipe
),
1132 sprite_name(pipe
, sprite
),
1133 wm
.pipe
[pipe
].sprite
[sprite
],
1134 wm
.sr
.plane
, wm
.sr
.cursor
);
1137 intel_set_memory_cxsr(dev_priv
, false);
1139 vlv_write_wm_values(intel_crtc
, &wm
);
1142 intel_set_memory_cxsr(dev_priv
, true);
1145 #define single_plane_enabled(mask) is_power_of_2(mask)
1147 static void g4x_update_wm(struct drm_crtc
*crtc
)
1149 struct drm_device
*dev
= crtc
->dev
;
1150 static const int sr_latency_ns
= 12000;
1151 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1152 int planea_wm
, planeb_wm
, cursora_wm
, cursorb_wm
;
1153 int plane_sr
, cursor_sr
;
1154 unsigned int enabled
= 0;
1157 if (g4x_compute_wm0(dev
, PIPE_A
,
1158 &g4x_wm_info
, pessimal_latency_ns
,
1159 &g4x_cursor_wm_info
, pessimal_latency_ns
,
1160 &planea_wm
, &cursora_wm
))
1161 enabled
|= 1 << PIPE_A
;
1163 if (g4x_compute_wm0(dev
, PIPE_B
,
1164 &g4x_wm_info
, pessimal_latency_ns
,
1165 &g4x_cursor_wm_info
, pessimal_latency_ns
,
1166 &planeb_wm
, &cursorb_wm
))
1167 enabled
|= 1 << PIPE_B
;
1169 if (single_plane_enabled(enabled
) &&
1170 g4x_compute_srwm(dev
, ffs(enabled
) - 1,
1173 &g4x_cursor_wm_info
,
1174 &plane_sr
, &cursor_sr
)) {
1175 cxsr_enabled
= true;
1177 cxsr_enabled
= false;
1178 intel_set_memory_cxsr(dev_priv
, false);
1179 plane_sr
= cursor_sr
= 0;
1182 DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, "
1183 "B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1184 planea_wm
, cursora_wm
,
1185 planeb_wm
, cursorb_wm
,
1186 plane_sr
, cursor_sr
);
1189 FW_WM(plane_sr
, SR
) |
1190 FW_WM(cursorb_wm
, CURSORB
) |
1191 FW_WM(planeb_wm
, PLANEB
) |
1192 FW_WM(planea_wm
, PLANEA
));
1194 (I915_READ(DSPFW2
) & ~DSPFW_CURSORA_MASK
) |
1195 FW_WM(cursora_wm
, CURSORA
));
1196 /* HPLL off in SR has some issues on G4x... disable it */
1198 (I915_READ(DSPFW3
) & ~(DSPFW_HPLL_SR_EN
| DSPFW_CURSOR_SR_MASK
)) |
1199 FW_WM(cursor_sr
, CURSOR_SR
));
1202 intel_set_memory_cxsr(dev_priv
, true);
1205 static void i965_update_wm(struct drm_crtc
*unused_crtc
)
1207 struct drm_device
*dev
= unused_crtc
->dev
;
1208 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1209 struct drm_crtc
*crtc
;
1214 /* Calc sr entries for one plane configs */
1215 crtc
= single_enabled_crtc(dev
);
1217 /* self-refresh has much higher latency */
1218 static const int sr_latency_ns
= 12000;
1219 const struct drm_display_mode
*adjusted_mode
=
1220 &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
1221 int clock
= adjusted_mode
->crtc_clock
;
1222 int htotal
= adjusted_mode
->crtc_htotal
;
1223 int hdisplay
= to_intel_crtc(crtc
)->config
->pipe_src_w
;
1224 int pixel_size
= crtc
->primary
->state
->fb
->bits_per_pixel
/ 8;
1225 unsigned long line_time_us
;
1228 line_time_us
= max(htotal
* 1000 / clock
, 1);
1230 /* Use ns/us then divide to preserve precision */
1231 entries
= (((sr_latency_ns
/ line_time_us
) + 1000) / 1000) *
1232 pixel_size
* hdisplay
;
1233 entries
= DIV_ROUND_UP(entries
, I915_FIFO_LINE_SIZE
);
1234 srwm
= I965_FIFO_SIZE
- entries
;
1238 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1241 entries
= (((sr_latency_ns
/ line_time_us
) + 1000) / 1000) *
1242 pixel_size
* crtc
->cursor
->state
->crtc_w
;
1243 entries
= DIV_ROUND_UP(entries
,
1244 i965_cursor_wm_info
.cacheline_size
);
1245 cursor_sr
= i965_cursor_wm_info
.fifo_size
-
1246 (entries
+ i965_cursor_wm_info
.guard_size
);
1248 if (cursor_sr
> i965_cursor_wm_info
.max_wm
)
1249 cursor_sr
= i965_cursor_wm_info
.max_wm
;
1251 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1252 "cursor %d\n", srwm
, cursor_sr
);
1254 cxsr_enabled
= true;
1256 cxsr_enabled
= false;
1257 /* Turn off self refresh if both pipes are enabled */
1258 intel_set_memory_cxsr(dev_priv
, false);
1261 DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1264 /* 965 has limitations... */
1265 I915_WRITE(DSPFW1
, FW_WM(srwm
, SR
) |
1269 I915_WRITE(DSPFW2
, FW_WM(8, CURSORA
) |
1270 FW_WM(8, PLANEC_OLD
));
1271 /* update cursor SR watermark */
1272 I915_WRITE(DSPFW3
, FW_WM(cursor_sr
, CURSOR_SR
));
1275 intel_set_memory_cxsr(dev_priv
, true);
1280 static void i9xx_update_wm(struct drm_crtc
*unused_crtc
)
1282 struct drm_device
*dev
= unused_crtc
->dev
;
1283 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1284 const struct intel_watermark_params
*wm_info
;
1289 int planea_wm
, planeb_wm
;
1290 struct drm_crtc
*crtc
, *enabled
= NULL
;
1293 wm_info
= &i945_wm_info
;
1294 else if (!IS_GEN2(dev
))
1295 wm_info
= &i915_wm_info
;
1297 wm_info
= &i830_a_wm_info
;
1299 fifo_size
= dev_priv
->display
.get_fifo_size(dev
, 0);
1300 crtc
= intel_get_crtc_for_plane(dev
, 0);
1301 if (intel_crtc_active(crtc
)) {
1302 const struct drm_display_mode
*adjusted_mode
;
1303 int cpp
= crtc
->primary
->state
->fb
->bits_per_pixel
/ 8;
1307 adjusted_mode
= &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
1308 planea_wm
= intel_calculate_wm(adjusted_mode
->crtc_clock
,
1309 wm_info
, fifo_size
, cpp
,
1310 pessimal_latency_ns
);
1313 planea_wm
= fifo_size
- wm_info
->guard_size
;
1314 if (planea_wm
> (long)wm_info
->max_wm
)
1315 planea_wm
= wm_info
->max_wm
;
1319 wm_info
= &i830_bc_wm_info
;
1321 fifo_size
= dev_priv
->display
.get_fifo_size(dev
, 1);
1322 crtc
= intel_get_crtc_for_plane(dev
, 1);
1323 if (intel_crtc_active(crtc
)) {
1324 const struct drm_display_mode
*adjusted_mode
;
1325 int cpp
= crtc
->primary
->state
->fb
->bits_per_pixel
/ 8;
1329 adjusted_mode
= &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
1330 planeb_wm
= intel_calculate_wm(adjusted_mode
->crtc_clock
,
1331 wm_info
, fifo_size
, cpp
,
1332 pessimal_latency_ns
);
1333 if (enabled
== NULL
)
1338 planeb_wm
= fifo_size
- wm_info
->guard_size
;
1339 if (planeb_wm
> (long)wm_info
->max_wm
)
1340 planeb_wm
= wm_info
->max_wm
;
1343 DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm
, planeb_wm
);
1345 if (IS_I915GM(dev
) && enabled
) {
1346 struct drm_i915_gem_object
*obj
;
1348 obj
= intel_fb_obj(enabled
->primary
->state
->fb
);
1350 /* self-refresh seems busted with untiled */
1351 if (obj
->tiling_mode
== I915_TILING_NONE
)
1356 * Overlay gets an aggressive default since video jitter is bad.
1360 /* Play safe and disable self-refresh before adjusting watermarks. */
1361 intel_set_memory_cxsr(dev_priv
, false);
1363 /* Calc sr entries for one plane configs */
1364 if (HAS_FW_BLC(dev
) && enabled
) {
1365 /* self-refresh has much higher latency */
1366 static const int sr_latency_ns
= 6000;
1367 const struct drm_display_mode
*adjusted_mode
=
1368 &to_intel_crtc(enabled
)->config
->base
.adjusted_mode
;
1369 int clock
= adjusted_mode
->crtc_clock
;
1370 int htotal
= adjusted_mode
->crtc_htotal
;
1371 int hdisplay
= to_intel_crtc(enabled
)->config
->pipe_src_w
;
1372 int pixel_size
= enabled
->primary
->state
->fb
->bits_per_pixel
/ 8;
1373 unsigned long line_time_us
;
1376 line_time_us
= max(htotal
* 1000 / clock
, 1);
1378 /* Use ns/us then divide to preserve precision */
1379 entries
= (((sr_latency_ns
/ line_time_us
) + 1000) / 1000) *
1380 pixel_size
* hdisplay
;
1381 entries
= DIV_ROUND_UP(entries
, wm_info
->cacheline_size
);
1382 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries
);
1383 srwm
= wm_info
->fifo_size
- entries
;
1387 if (IS_I945G(dev
) || IS_I945GM(dev
))
1388 I915_WRITE(FW_BLC_SELF
,
1389 FW_BLC_SELF_FIFO_MASK
| (srwm
& 0xff));
1390 else if (IS_I915GM(dev
))
1391 I915_WRITE(FW_BLC_SELF
, srwm
& 0x3f);
1394 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1395 planea_wm
, planeb_wm
, cwm
, srwm
);
1397 fwater_lo
= ((planeb_wm
& 0x3f) << 16) | (planea_wm
& 0x3f);
1398 fwater_hi
= (cwm
& 0x1f);
1400 /* Set request length to 8 cachelines per fetch */
1401 fwater_lo
= fwater_lo
| (1 << 24) | (1 << 8);
1402 fwater_hi
= fwater_hi
| (1 << 8);
1404 I915_WRITE(FW_BLC
, fwater_lo
);
1405 I915_WRITE(FW_BLC2
, fwater_hi
);
1408 intel_set_memory_cxsr(dev_priv
, true);
1411 static void i845_update_wm(struct drm_crtc
*unused_crtc
)
1413 struct drm_device
*dev
= unused_crtc
->dev
;
1414 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1415 struct drm_crtc
*crtc
;
1416 const struct drm_display_mode
*adjusted_mode
;
1420 crtc
= single_enabled_crtc(dev
);
1424 adjusted_mode
= &to_intel_crtc(crtc
)->config
->base
.adjusted_mode
;
1425 planea_wm
= intel_calculate_wm(adjusted_mode
->crtc_clock
,
1427 dev_priv
->display
.get_fifo_size(dev
, 0),
1428 4, pessimal_latency_ns
);
1429 fwater_lo
= I915_READ(FW_BLC
) & ~0xfff;
1430 fwater_lo
|= (3<<8) | planea_wm
;
1432 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm
);
1434 I915_WRITE(FW_BLC
, fwater_lo
);
1437 static uint32_t ilk_pipe_pixel_rate(struct drm_device
*dev
,
1438 struct drm_crtc
*crtc
)
1440 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
1441 uint32_t pixel_rate
;
1443 pixel_rate
= intel_crtc
->config
->base
.adjusted_mode
.crtc_clock
;
1445 /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
1446 * adjust the pixel_rate here. */
1448 if (intel_crtc
->config
->pch_pfit
.enabled
) {
1449 uint64_t pipe_w
, pipe_h
, pfit_w
, pfit_h
;
1450 uint32_t pfit_size
= intel_crtc
->config
->pch_pfit
.size
;
1452 pipe_w
= intel_crtc
->config
->pipe_src_w
;
1453 pipe_h
= intel_crtc
->config
->pipe_src_h
;
1454 pfit_w
= (pfit_size
>> 16) & 0xFFFF;
1455 pfit_h
= pfit_size
& 0xFFFF;
1456 if (pipe_w
< pfit_w
)
1458 if (pipe_h
< pfit_h
)
1461 pixel_rate
= div_u64((uint64_t) pixel_rate
* pipe_w
* pipe_h
,
1468 /* latency must be in 0.1us units. */
1469 static uint32_t ilk_wm_method1(uint32_t pixel_rate
, uint8_t bytes_per_pixel
,
1474 if (WARN(latency
== 0, "Latency value missing\n"))
1477 ret
= (uint64_t) pixel_rate
* bytes_per_pixel
* latency
;
1478 ret
= DIV_ROUND_UP_ULL(ret
, 64 * 10000) + 2;
1483 /* latency must be in 0.1us units. */
1484 static uint32_t ilk_wm_method2(uint32_t pixel_rate
, uint32_t pipe_htotal
,
1485 uint32_t horiz_pixels
, uint8_t bytes_per_pixel
,
1490 if (WARN(latency
== 0, "Latency value missing\n"))
1493 ret
= (latency
* pixel_rate
) / (pipe_htotal
* 10000);
1494 ret
= (ret
+ 1) * horiz_pixels
* bytes_per_pixel
;
1495 ret
= DIV_ROUND_UP(ret
, 64) + 2;
1499 static uint32_t ilk_wm_fbc(uint32_t pri_val
, uint32_t horiz_pixels
,
1500 uint8_t bytes_per_pixel
)
1502 return DIV_ROUND_UP(pri_val
* 64, horiz_pixels
* bytes_per_pixel
) + 2;
1505 struct skl_pipe_wm_parameters
{
1507 uint32_t pipe_htotal
;
1508 uint32_t pixel_rate
; /* in KHz */
1509 struct intel_plane_wm_parameters plane
[I915_MAX_PLANES
];
1510 struct intel_plane_wm_parameters cursor
;
1513 struct ilk_pipe_wm_parameters
{
1515 uint32_t pipe_htotal
;
1516 uint32_t pixel_rate
;
1517 struct intel_plane_wm_parameters pri
;
1518 struct intel_plane_wm_parameters spr
;
1519 struct intel_plane_wm_parameters cur
;
1522 struct ilk_wm_maximums
{
1529 /* used in computing the new watermarks state */
1530 struct intel_wm_config
{
1531 unsigned int num_pipes_active
;
1532 bool sprites_enabled
;
1533 bool sprites_scaled
;
1537 * For both WM_PIPE and WM_LP.
1538 * mem_value must be in 0.1us units.
1540 static uint32_t ilk_compute_pri_wm(const struct ilk_pipe_wm_parameters
*params
,
1544 uint32_t method1
, method2
;
1546 if (!params
->active
|| !params
->pri
.enabled
)
1549 method1
= ilk_wm_method1(params
->pixel_rate
,
1550 params
->pri
.bytes_per_pixel
,
1556 method2
= ilk_wm_method2(params
->pixel_rate
,
1557 params
->pipe_htotal
,
1558 params
->pri
.horiz_pixels
,
1559 params
->pri
.bytes_per_pixel
,
1562 return min(method1
, method2
);
1566 * For both WM_PIPE and WM_LP.
1567 * mem_value must be in 0.1us units.
1569 static uint32_t ilk_compute_spr_wm(const struct ilk_pipe_wm_parameters
*params
,
1572 uint32_t method1
, method2
;
1574 if (!params
->active
|| !params
->spr
.enabled
)
1577 method1
= ilk_wm_method1(params
->pixel_rate
,
1578 params
->spr
.bytes_per_pixel
,
1580 method2
= ilk_wm_method2(params
->pixel_rate
,
1581 params
->pipe_htotal
,
1582 params
->spr
.horiz_pixels
,
1583 params
->spr
.bytes_per_pixel
,
1585 return min(method1
, method2
);
1589 * For both WM_PIPE and WM_LP.
1590 * mem_value must be in 0.1us units.
1592 static uint32_t ilk_compute_cur_wm(const struct ilk_pipe_wm_parameters
*params
,
1595 if (!params
->active
|| !params
->cur
.enabled
)
1598 return ilk_wm_method2(params
->pixel_rate
,
1599 params
->pipe_htotal
,
1600 params
->cur
.horiz_pixels
,
1601 params
->cur
.bytes_per_pixel
,
1605 /* Only for WM_LP. */
1606 static uint32_t ilk_compute_fbc_wm(const struct ilk_pipe_wm_parameters
*params
,
1609 if (!params
->active
|| !params
->pri
.enabled
)
1612 return ilk_wm_fbc(pri_val
,
1613 params
->pri
.horiz_pixels
,
1614 params
->pri
.bytes_per_pixel
);
1617 static unsigned int ilk_display_fifo_size(const struct drm_device
*dev
)
1619 if (INTEL_INFO(dev
)->gen
>= 8)
1621 else if (INTEL_INFO(dev
)->gen
>= 7)
1627 static unsigned int ilk_plane_wm_reg_max(const struct drm_device
*dev
,
1628 int level
, bool is_sprite
)
1630 if (INTEL_INFO(dev
)->gen
>= 8)
1631 /* BDW primary/sprite plane watermarks */
1632 return level
== 0 ? 255 : 2047;
1633 else if (INTEL_INFO(dev
)->gen
>= 7)
1634 /* IVB/HSW primary/sprite plane watermarks */
1635 return level
== 0 ? 127 : 1023;
1636 else if (!is_sprite
)
1637 /* ILK/SNB primary plane watermarks */
1638 return level
== 0 ? 127 : 511;
1640 /* ILK/SNB sprite plane watermarks */
1641 return level
== 0 ? 63 : 255;
1644 static unsigned int ilk_cursor_wm_reg_max(const struct drm_device
*dev
,
1647 if (INTEL_INFO(dev
)->gen
>= 7)
1648 return level
== 0 ? 63 : 255;
1650 return level
== 0 ? 31 : 63;
1653 static unsigned int ilk_fbc_wm_reg_max(const struct drm_device
*dev
)
1655 if (INTEL_INFO(dev
)->gen
>= 8)
1661 /* Calculate the maximum primary/sprite plane watermark */
1662 static unsigned int ilk_plane_wm_max(const struct drm_device
*dev
,
1664 const struct intel_wm_config
*config
,
1665 enum intel_ddb_partitioning ddb_partitioning
,
1668 unsigned int fifo_size
= ilk_display_fifo_size(dev
);
1670 /* if sprites aren't enabled, sprites get nothing */
1671 if (is_sprite
&& !config
->sprites_enabled
)
1674 /* HSW allows LP1+ watermarks even with multiple pipes */
1675 if (level
== 0 || config
->num_pipes_active
> 1) {
1676 fifo_size
/= INTEL_INFO(dev
)->num_pipes
;
1679 * For some reason the non self refresh
1680 * FIFO size is only half of the self
1681 * refresh FIFO size on ILK/SNB.
1683 if (INTEL_INFO(dev
)->gen
<= 6)
1687 if (config
->sprites_enabled
) {
1688 /* level 0 is always calculated with 1:1 split */
1689 if (level
> 0 && ddb_partitioning
== INTEL_DDB_PART_5_6
) {
1698 /* clamp to max that the registers can hold */
1699 return min(fifo_size
, ilk_plane_wm_reg_max(dev
, level
, is_sprite
));
1702 /* Calculate the maximum cursor plane watermark */
1703 static unsigned int ilk_cursor_wm_max(const struct drm_device
*dev
,
1705 const struct intel_wm_config
*config
)
1707 /* HSW LP1+ watermarks w/ multiple pipes */
1708 if (level
> 0 && config
->num_pipes_active
> 1)
1711 /* otherwise just report max that registers can hold */
1712 return ilk_cursor_wm_reg_max(dev
, level
);
1715 static void ilk_compute_wm_maximums(const struct drm_device
*dev
,
1717 const struct intel_wm_config
*config
,
1718 enum intel_ddb_partitioning ddb_partitioning
,
1719 struct ilk_wm_maximums
*max
)
1721 max
->pri
= ilk_plane_wm_max(dev
, level
, config
, ddb_partitioning
, false);
1722 max
->spr
= ilk_plane_wm_max(dev
, level
, config
, ddb_partitioning
, true);
1723 max
->cur
= ilk_cursor_wm_max(dev
, level
, config
);
1724 max
->fbc
= ilk_fbc_wm_reg_max(dev
);
1727 static void ilk_compute_wm_reg_maximums(struct drm_device
*dev
,
1729 struct ilk_wm_maximums
*max
)
1731 max
->pri
= ilk_plane_wm_reg_max(dev
, level
, false);
1732 max
->spr
= ilk_plane_wm_reg_max(dev
, level
, true);
1733 max
->cur
= ilk_cursor_wm_reg_max(dev
, level
);
1734 max
->fbc
= ilk_fbc_wm_reg_max(dev
);
1737 static bool ilk_validate_wm_level(int level
,
1738 const struct ilk_wm_maximums
*max
,
1739 struct intel_wm_level
*result
)
1743 /* already determined to be invalid? */
1744 if (!result
->enable
)
1747 result
->enable
= result
->pri_val
<= max
->pri
&&
1748 result
->spr_val
<= max
->spr
&&
1749 result
->cur_val
<= max
->cur
;
1751 ret
= result
->enable
;
1754 * HACK until we can pre-compute everything,
1755 * and thus fail gracefully if LP0 watermarks
1758 if (level
== 0 && !result
->enable
) {
1759 if (result
->pri_val
> max
->pri
)
1760 DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
1761 level
, result
->pri_val
, max
->pri
);
1762 if (result
->spr_val
> max
->spr
)
1763 DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
1764 level
, result
->spr_val
, max
->spr
);
1765 if (result
->cur_val
> max
->cur
)
1766 DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
1767 level
, result
->cur_val
, max
->cur
);
1769 result
->pri_val
= min_t(uint32_t, result
->pri_val
, max
->pri
);
1770 result
->spr_val
= min_t(uint32_t, result
->spr_val
, max
->spr
);
1771 result
->cur_val
= min_t(uint32_t, result
->cur_val
, max
->cur
);
1772 result
->enable
= true;
1778 static void ilk_compute_wm_level(const struct drm_i915_private
*dev_priv
,
1780 const struct ilk_pipe_wm_parameters
*p
,
1781 struct intel_wm_level
*result
)
1783 uint16_t pri_latency
= dev_priv
->wm
.pri_latency
[level
];
1784 uint16_t spr_latency
= dev_priv
->wm
.spr_latency
[level
];
1785 uint16_t cur_latency
= dev_priv
->wm
.cur_latency
[level
];
1787 /* WM1+ latency values stored in 0.5us units */
1794 result
->pri_val
= ilk_compute_pri_wm(p
, pri_latency
, level
);
1795 result
->spr_val
= ilk_compute_spr_wm(p
, spr_latency
);
1796 result
->cur_val
= ilk_compute_cur_wm(p
, cur_latency
);
1797 result
->fbc_val
= ilk_compute_fbc_wm(p
, result
->pri_val
);
1798 result
->enable
= true;
1802 hsw_compute_linetime_wm(struct drm_device
*dev
, struct drm_crtc
*crtc
)
1804 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1805 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
1806 struct drm_display_mode
*mode
= &intel_crtc
->config
->base
.adjusted_mode
;
1807 u32 linetime
, ips_linetime
;
1809 if (!intel_crtc
->active
)
1812 /* The WM are computed with base on how long it takes to fill a single
1813 * row at the given clock rate, multiplied by 8.
1815 linetime
= DIV_ROUND_CLOSEST(mode
->crtc_htotal
* 1000 * 8,
1817 ips_linetime
= DIV_ROUND_CLOSEST(mode
->crtc_htotal
* 1000 * 8,
1818 dev_priv
->display
.get_display_clock_speed(dev_priv
->dev
));
1820 return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime
) |
1821 PIPE_WM_LINETIME_TIME(linetime
);
1824 static void intel_read_wm_latency(struct drm_device
*dev
, uint16_t wm
[8])
1826 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
1831 int level
, max_level
= ilk_wm_max_level(dev
);
1833 /* read the first set of memory latencies[0:3] */
1834 val
= 0; /* data0 to be programmed to 0 for first set */
1835 mutex_lock(&dev_priv
->rps
.hw_lock
);
1836 ret
= sandybridge_pcode_read(dev_priv
,
1837 GEN9_PCODE_READ_MEM_LATENCY
,
1839 mutex_unlock(&dev_priv
->rps
.hw_lock
);
1842 DRM_ERROR("SKL Mailbox read error = %d\n", ret
);
1846 wm
[0] = val
& GEN9_MEM_LATENCY_LEVEL_MASK
;
1847 wm
[1] = (val
>> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT
) &
1848 GEN9_MEM_LATENCY_LEVEL_MASK
;
1849 wm
[2] = (val
>> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT
) &
1850 GEN9_MEM_LATENCY_LEVEL_MASK
;
1851 wm
[3] = (val
>> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT
) &
1852 GEN9_MEM_LATENCY_LEVEL_MASK
;
1854 /* read the second set of memory latencies[4:7] */
1855 val
= 1; /* data0 to be programmed to 1 for second set */
1856 mutex_lock(&dev_priv
->rps
.hw_lock
);
1857 ret
= sandybridge_pcode_read(dev_priv
,
1858 GEN9_PCODE_READ_MEM_LATENCY
,
1860 mutex_unlock(&dev_priv
->rps
.hw_lock
);
1862 DRM_ERROR("SKL Mailbox read error = %d\n", ret
);
1866 wm
[4] = val
& GEN9_MEM_LATENCY_LEVEL_MASK
;
1867 wm
[5] = (val
>> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT
) &
1868 GEN9_MEM_LATENCY_LEVEL_MASK
;
1869 wm
[6] = (val
>> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT
) &
1870 GEN9_MEM_LATENCY_LEVEL_MASK
;
1871 wm
[7] = (val
>> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT
) &
1872 GEN9_MEM_LATENCY_LEVEL_MASK
;
1875 * WaWmMemoryReadLatency:skl
1877 * punit doesn't take into account the read latency so we need
1878 * to add 2us to the various latency levels we retrieve from
1880 * - W0 is a bit special in that it's the only level that
1881 * can't be disabled if we want to have display working, so
1882 * we always add 2us there.
1883 * - For levels >=1, punit returns 0us latency when they are
1884 * disabled, so we respect that and don't add 2us then
1886 * Additionally, if a level n (n > 1) has a 0us latency, all
1887 * levels m (m >= n) need to be disabled. We make sure to
1888 * sanitize the values out of the punit to satisfy this
1892 for (level
= 1; level
<= max_level
; level
++)
1896 for (i
= level
+ 1; i
<= max_level
; i
++)
1901 } else if (IS_HASWELL(dev
) || IS_BROADWELL(dev
)) {
1902 uint64_t sskpd
= I915_READ64(MCH_SSKPD
);
1904 wm
[0] = (sskpd
>> 56) & 0xFF;
1906 wm
[0] = sskpd
& 0xF;
1907 wm
[1] = (sskpd
>> 4) & 0xFF;
1908 wm
[2] = (sskpd
>> 12) & 0xFF;
1909 wm
[3] = (sskpd
>> 20) & 0x1FF;
1910 wm
[4] = (sskpd
>> 32) & 0x1FF;
1911 } else if (INTEL_INFO(dev
)->gen
>= 6) {
1912 uint32_t sskpd
= I915_READ(MCH_SSKPD
);
1914 wm
[0] = (sskpd
>> SSKPD_WM0_SHIFT
) & SSKPD_WM_MASK
;
1915 wm
[1] = (sskpd
>> SSKPD_WM1_SHIFT
) & SSKPD_WM_MASK
;
1916 wm
[2] = (sskpd
>> SSKPD_WM2_SHIFT
) & SSKPD_WM_MASK
;
1917 wm
[3] = (sskpd
>> SSKPD_WM3_SHIFT
) & SSKPD_WM_MASK
;
1918 } else if (INTEL_INFO(dev
)->gen
>= 5) {
1919 uint32_t mltr
= I915_READ(MLTR_ILK
);
1921 /* ILK primary LP0 latency is 700 ns */
1923 wm
[1] = (mltr
>> MLTR_WM1_SHIFT
) & ILK_SRLT_MASK
;
1924 wm
[2] = (mltr
>> MLTR_WM2_SHIFT
) & ILK_SRLT_MASK
;
1928 static void intel_fixup_spr_wm_latency(struct drm_device
*dev
, uint16_t wm
[5])
1930 /* ILK sprite LP0 latency is 1300 ns */
1931 if (INTEL_INFO(dev
)->gen
== 5)
1935 static void intel_fixup_cur_wm_latency(struct drm_device
*dev
, uint16_t wm
[5])
1937 /* ILK cursor LP0 latency is 1300 ns */
1938 if (INTEL_INFO(dev
)->gen
== 5)
1941 /* WaDoubleCursorLP3Latency:ivb */
1942 if (IS_IVYBRIDGE(dev
))
1946 int ilk_wm_max_level(const struct drm_device
*dev
)
1948 /* how many WM levels are we expecting */
1949 if (INTEL_INFO(dev
)->gen
>= 9)
1951 else if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
1953 else if (INTEL_INFO(dev
)->gen
>= 6)
1959 static void intel_print_wm_latency(struct drm_device
*dev
,
1961 const uint16_t wm
[8])
1963 int level
, max_level
= ilk_wm_max_level(dev
);
1965 for (level
= 0; level
<= max_level
; level
++) {
1966 unsigned int latency
= wm
[level
];
1969 DRM_ERROR("%s WM%d latency not provided\n",
1975 * - latencies are in us on gen9.
1976 * - before then, WM1+ latency values are in 0.5us units
1983 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
1984 name
, level
, wm
[level
],
1985 latency
/ 10, latency
% 10);
1989 static bool ilk_increase_wm_latency(struct drm_i915_private
*dev_priv
,
1990 uint16_t wm
[5], uint16_t min
)
1992 int level
, max_level
= ilk_wm_max_level(dev_priv
->dev
);
1997 wm
[0] = max(wm
[0], min
);
1998 for (level
= 1; level
<= max_level
; level
++)
1999 wm
[level
] = max_t(uint16_t, wm
[level
], DIV_ROUND_UP(min
, 5));
2004 static void snb_wm_latency_quirk(struct drm_device
*dev
)
2006 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2010 * The BIOS provided WM memory latency values are often
2011 * inadequate for high resolution displays. Adjust them.
2013 changed
= ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.pri_latency
, 12) |
2014 ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.spr_latency
, 12) |
2015 ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.cur_latency
, 12);
2020 DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2021 intel_print_wm_latency(dev
, "Primary", dev_priv
->wm
.pri_latency
);
2022 intel_print_wm_latency(dev
, "Sprite", dev_priv
->wm
.spr_latency
);
2023 intel_print_wm_latency(dev
, "Cursor", dev_priv
->wm
.cur_latency
);
2026 static void ilk_setup_wm_latency(struct drm_device
*dev
)
2028 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2030 intel_read_wm_latency(dev
, dev_priv
->wm
.pri_latency
);
2032 memcpy(dev_priv
->wm
.spr_latency
, dev_priv
->wm
.pri_latency
,
2033 sizeof(dev_priv
->wm
.pri_latency
));
2034 memcpy(dev_priv
->wm
.cur_latency
, dev_priv
->wm
.pri_latency
,
2035 sizeof(dev_priv
->wm
.pri_latency
));
2037 intel_fixup_spr_wm_latency(dev
, dev_priv
->wm
.spr_latency
);
2038 intel_fixup_cur_wm_latency(dev
, dev_priv
->wm
.cur_latency
);
2040 intel_print_wm_latency(dev
, "Primary", dev_priv
->wm
.pri_latency
);
2041 intel_print_wm_latency(dev
, "Sprite", dev_priv
->wm
.spr_latency
);
2042 intel_print_wm_latency(dev
, "Cursor", dev_priv
->wm
.cur_latency
);
2045 snb_wm_latency_quirk(dev
);
2048 static void skl_setup_wm_latency(struct drm_device
*dev
)
2050 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2052 intel_read_wm_latency(dev
, dev_priv
->wm
.skl_latency
);
2053 intel_print_wm_latency(dev
, "Gen9 Plane", dev_priv
->wm
.skl_latency
);
2056 static void ilk_compute_wm_parameters(struct drm_crtc
*crtc
,
2057 struct ilk_pipe_wm_parameters
*p
)
2059 struct drm_device
*dev
= crtc
->dev
;
2060 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
2061 enum pipe pipe
= intel_crtc
->pipe
;
2062 struct drm_plane
*plane
;
2064 if (!intel_crtc
->active
)
2068 p
->pipe_htotal
= intel_crtc
->config
->base
.adjusted_mode
.crtc_htotal
;
2069 p
->pixel_rate
= ilk_pipe_pixel_rate(dev
, crtc
);
2071 if (crtc
->primary
->state
->fb
) {
2072 p
->pri
.enabled
= true;
2073 p
->pri
.bytes_per_pixel
=
2074 crtc
->primary
->state
->fb
->bits_per_pixel
/ 8;
2076 p
->pri
.enabled
= false;
2077 p
->pri
.bytes_per_pixel
= 0;
2080 if (crtc
->cursor
->state
->fb
) {
2081 p
->cur
.enabled
= true;
2082 p
->cur
.bytes_per_pixel
= 4;
2084 p
->cur
.enabled
= false;
2085 p
->cur
.bytes_per_pixel
= 0;
2087 p
->pri
.horiz_pixels
= intel_crtc
->config
->pipe_src_w
;
2088 p
->cur
.horiz_pixels
= intel_crtc
->base
.cursor
->state
->crtc_w
;
2090 drm_for_each_legacy_plane(plane
, &dev
->mode_config
.plane_list
) {
2091 struct intel_plane
*intel_plane
= to_intel_plane(plane
);
2093 if (intel_plane
->pipe
== pipe
) {
2094 p
->spr
= intel_plane
->wm
;
2100 static void ilk_compute_wm_config(struct drm_device
*dev
,
2101 struct intel_wm_config
*config
)
2103 struct intel_crtc
*intel_crtc
;
2105 /* Compute the currently _active_ config */
2106 for_each_intel_crtc(dev
, intel_crtc
) {
2107 const struct intel_pipe_wm
*wm
= &intel_crtc
->wm
.active
;
2109 if (!wm
->pipe_enabled
)
2112 config
->sprites_enabled
|= wm
->sprites_enabled
;
2113 config
->sprites_scaled
|= wm
->sprites_scaled
;
2114 config
->num_pipes_active
++;
2118 /* Compute new watermarks for the pipe */
2119 static bool intel_compute_pipe_wm(struct drm_crtc
*crtc
,
2120 const struct ilk_pipe_wm_parameters
*params
,
2121 struct intel_pipe_wm
*pipe_wm
)
2123 struct drm_device
*dev
= crtc
->dev
;
2124 const struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2125 int level
, max_level
= ilk_wm_max_level(dev
);
2126 /* LP0 watermark maximums depend on this pipe alone */
2127 struct intel_wm_config config
= {
2128 .num_pipes_active
= 1,
2129 .sprites_enabled
= params
->spr
.enabled
,
2130 .sprites_scaled
= params
->spr
.scaled
,
2132 struct ilk_wm_maximums max
;
2134 pipe_wm
->pipe_enabled
= params
->active
;
2135 pipe_wm
->sprites_enabled
= params
->spr
.enabled
;
2136 pipe_wm
->sprites_scaled
= params
->spr
.scaled
;
2138 /* ILK/SNB: LP2+ watermarks only w/o sprites */
2139 if (INTEL_INFO(dev
)->gen
<= 6 && params
->spr
.enabled
)
2142 /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2143 if (params
->spr
.scaled
)
2146 ilk_compute_wm_level(dev_priv
, 0, params
, &pipe_wm
->wm
[0]);
2148 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
2149 pipe_wm
->linetime
= hsw_compute_linetime_wm(dev
, crtc
);
2151 /* LP0 watermarks always use 1/2 DDB partitioning */
2152 ilk_compute_wm_maximums(dev
, 0, &config
, INTEL_DDB_PART_1_2
, &max
);
2154 /* At least LP0 must be valid */
2155 if (!ilk_validate_wm_level(0, &max
, &pipe_wm
->wm
[0]))
2158 ilk_compute_wm_reg_maximums(dev
, 1, &max
);
2160 for (level
= 1; level
<= max_level
; level
++) {
2161 struct intel_wm_level wm
= {};
2163 ilk_compute_wm_level(dev_priv
, level
, params
, &wm
);
2166 * Disable any watermark level that exceeds the
2167 * register maximums since such watermarks are
2170 if (!ilk_validate_wm_level(level
, &max
, &wm
))
2173 pipe_wm
->wm
[level
] = wm
;
2180 * Merge the watermarks from all active pipes for a specific level.
2182 static void ilk_merge_wm_level(struct drm_device
*dev
,
2184 struct intel_wm_level
*ret_wm
)
2186 const struct intel_crtc
*intel_crtc
;
2188 ret_wm
->enable
= true;
2190 for_each_intel_crtc(dev
, intel_crtc
) {
2191 const struct intel_pipe_wm
*active
= &intel_crtc
->wm
.active
;
2192 const struct intel_wm_level
*wm
= &active
->wm
[level
];
2194 if (!active
->pipe_enabled
)
2198 * The watermark values may have been used in the past,
2199 * so we must maintain them in the registers for some
2200 * time even if the level is now disabled.
2203 ret_wm
->enable
= false;
2205 ret_wm
->pri_val
= max(ret_wm
->pri_val
, wm
->pri_val
);
2206 ret_wm
->spr_val
= max(ret_wm
->spr_val
, wm
->spr_val
);
2207 ret_wm
->cur_val
= max(ret_wm
->cur_val
, wm
->cur_val
);
2208 ret_wm
->fbc_val
= max(ret_wm
->fbc_val
, wm
->fbc_val
);
2213 * Merge all low power watermarks for all active pipes.
2215 static void ilk_wm_merge(struct drm_device
*dev
,
2216 const struct intel_wm_config
*config
,
2217 const struct ilk_wm_maximums
*max
,
2218 struct intel_pipe_wm
*merged
)
2220 int level
, max_level
= ilk_wm_max_level(dev
);
2221 int last_enabled_level
= max_level
;
2223 /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2224 if ((INTEL_INFO(dev
)->gen
<= 6 || IS_IVYBRIDGE(dev
)) &&
2225 config
->num_pipes_active
> 1)
2228 /* ILK: FBC WM must be disabled always */
2229 merged
->fbc_wm_enabled
= INTEL_INFO(dev
)->gen
>= 6;
2231 /* merge each WM1+ level */
2232 for (level
= 1; level
<= max_level
; level
++) {
2233 struct intel_wm_level
*wm
= &merged
->wm
[level
];
2235 ilk_merge_wm_level(dev
, level
, wm
);
2237 if (level
> last_enabled_level
)
2239 else if (!ilk_validate_wm_level(level
, max
, wm
))
2240 /* make sure all following levels get disabled */
2241 last_enabled_level
= level
- 1;
2244 * The spec says it is preferred to disable
2245 * FBC WMs instead of disabling a WM level.
2247 if (wm
->fbc_val
> max
->fbc
) {
2249 merged
->fbc_wm_enabled
= false;
2254 /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2256 * FIXME this is racy. FBC might get enabled later.
2257 * What we should check here is whether FBC can be
2258 * enabled sometime later.
2260 if (IS_GEN5(dev
) && !merged
->fbc_wm_enabled
&& intel_fbc_enabled(dev
)) {
2261 for (level
= 2; level
<= max_level
; level
++) {
2262 struct intel_wm_level
*wm
= &merged
->wm
[level
];
2269 static int ilk_wm_lp_to_level(int wm_lp
, const struct intel_pipe_wm
*pipe_wm
)
2271 /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2272 return wm_lp
+ (wm_lp
>= 2 && pipe_wm
->wm
[4].enable
);
2275 /* The value we need to program into the WM_LPx latency field */
2276 static unsigned int ilk_wm_lp_latency(struct drm_device
*dev
, int level
)
2278 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2280 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
2283 return dev_priv
->wm
.pri_latency
[level
];
2286 static void ilk_compute_wm_results(struct drm_device
*dev
,
2287 const struct intel_pipe_wm
*merged
,
2288 enum intel_ddb_partitioning partitioning
,
2289 struct ilk_wm_values
*results
)
2291 struct intel_crtc
*intel_crtc
;
2294 results
->enable_fbc_wm
= merged
->fbc_wm_enabled
;
2295 results
->partitioning
= partitioning
;
2297 /* LP1+ register values */
2298 for (wm_lp
= 1; wm_lp
<= 3; wm_lp
++) {
2299 const struct intel_wm_level
*r
;
2301 level
= ilk_wm_lp_to_level(wm_lp
, merged
);
2303 r
= &merged
->wm
[level
];
2306 * Maintain the watermark values even if the level is
2307 * disabled. Doing otherwise could cause underruns.
2309 results
->wm_lp
[wm_lp
- 1] =
2310 (ilk_wm_lp_latency(dev
, level
) << WM1_LP_LATENCY_SHIFT
) |
2311 (r
->pri_val
<< WM1_LP_SR_SHIFT
) |
2315 results
->wm_lp
[wm_lp
- 1] |= WM1_LP_SR_EN
;
2317 if (INTEL_INFO(dev
)->gen
>= 8)
2318 results
->wm_lp
[wm_lp
- 1] |=
2319 r
->fbc_val
<< WM1_LP_FBC_SHIFT_BDW
;
2321 results
->wm_lp
[wm_lp
- 1] |=
2322 r
->fbc_val
<< WM1_LP_FBC_SHIFT
;
2325 * Always set WM1S_LP_EN when spr_val != 0, even if the
2326 * level is disabled. Doing otherwise could cause underruns.
2328 if (INTEL_INFO(dev
)->gen
<= 6 && r
->spr_val
) {
2329 WARN_ON(wm_lp
!= 1);
2330 results
->wm_lp_spr
[wm_lp
- 1] = WM1S_LP_EN
| r
->spr_val
;
2332 results
->wm_lp_spr
[wm_lp
- 1] = r
->spr_val
;
2335 /* LP0 register values */
2336 for_each_intel_crtc(dev
, intel_crtc
) {
2337 enum pipe pipe
= intel_crtc
->pipe
;
2338 const struct intel_wm_level
*r
=
2339 &intel_crtc
->wm
.active
.wm
[0];
2341 if (WARN_ON(!r
->enable
))
2344 results
->wm_linetime
[pipe
] = intel_crtc
->wm
.active
.linetime
;
2346 results
->wm_pipe
[pipe
] =
2347 (r
->pri_val
<< WM0_PIPE_PLANE_SHIFT
) |
2348 (r
->spr_val
<< WM0_PIPE_SPRITE_SHIFT
) |
2353 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2354 * case both are at the same level. Prefer r1 in case they're the same. */
2355 static struct intel_pipe_wm
*ilk_find_best_result(struct drm_device
*dev
,
2356 struct intel_pipe_wm
*r1
,
2357 struct intel_pipe_wm
*r2
)
2359 int level
, max_level
= ilk_wm_max_level(dev
);
2360 int level1
= 0, level2
= 0;
2362 for (level
= 1; level
<= max_level
; level
++) {
2363 if (r1
->wm
[level
].enable
)
2365 if (r2
->wm
[level
].enable
)
2369 if (level1
== level2
) {
2370 if (r2
->fbc_wm_enabled
&& !r1
->fbc_wm_enabled
)
2374 } else if (level1
> level2
) {
2381 /* dirty bits used to track which watermarks need changes */
2382 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2383 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2384 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2385 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2386 #define WM_DIRTY_FBC (1 << 24)
2387 #define WM_DIRTY_DDB (1 << 25)
2389 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private
*dev_priv
,
2390 const struct ilk_wm_values
*old
,
2391 const struct ilk_wm_values
*new)
2393 unsigned int dirty
= 0;
2397 for_each_pipe(dev_priv
, pipe
) {
2398 if (old
->wm_linetime
[pipe
] != new->wm_linetime
[pipe
]) {
2399 dirty
|= WM_DIRTY_LINETIME(pipe
);
2400 /* Must disable LP1+ watermarks too */
2401 dirty
|= WM_DIRTY_LP_ALL
;
2404 if (old
->wm_pipe
[pipe
] != new->wm_pipe
[pipe
]) {
2405 dirty
|= WM_DIRTY_PIPE(pipe
);
2406 /* Must disable LP1+ watermarks too */
2407 dirty
|= WM_DIRTY_LP_ALL
;
2411 if (old
->enable_fbc_wm
!= new->enable_fbc_wm
) {
2412 dirty
|= WM_DIRTY_FBC
;
2413 /* Must disable LP1+ watermarks too */
2414 dirty
|= WM_DIRTY_LP_ALL
;
2417 if (old
->partitioning
!= new->partitioning
) {
2418 dirty
|= WM_DIRTY_DDB
;
2419 /* Must disable LP1+ watermarks too */
2420 dirty
|= WM_DIRTY_LP_ALL
;
2423 /* LP1+ watermarks already deemed dirty, no need to continue */
2424 if (dirty
& WM_DIRTY_LP_ALL
)
2427 /* Find the lowest numbered LP1+ watermark in need of an update... */
2428 for (wm_lp
= 1; wm_lp
<= 3; wm_lp
++) {
2429 if (old
->wm_lp
[wm_lp
- 1] != new->wm_lp
[wm_lp
- 1] ||
2430 old
->wm_lp_spr
[wm_lp
- 1] != new->wm_lp_spr
[wm_lp
- 1])
2434 /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2435 for (; wm_lp
<= 3; wm_lp
++)
2436 dirty
|= WM_DIRTY_LP(wm_lp
);
2441 static bool _ilk_disable_lp_wm(struct drm_i915_private
*dev_priv
,
2444 struct ilk_wm_values
*previous
= &dev_priv
->wm
.hw
;
2445 bool changed
= false;
2447 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp
[2] & WM1_LP_SR_EN
) {
2448 previous
->wm_lp
[2] &= ~WM1_LP_SR_EN
;
2449 I915_WRITE(WM3_LP_ILK
, previous
->wm_lp
[2]);
2452 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp
[1] & WM1_LP_SR_EN
) {
2453 previous
->wm_lp
[1] &= ~WM1_LP_SR_EN
;
2454 I915_WRITE(WM2_LP_ILK
, previous
->wm_lp
[1]);
2457 if (dirty
& WM_DIRTY_LP(1) && previous
->wm_lp
[0] & WM1_LP_SR_EN
) {
2458 previous
->wm_lp
[0] &= ~WM1_LP_SR_EN
;
2459 I915_WRITE(WM1_LP_ILK
, previous
->wm_lp
[0]);
2464 * Don't touch WM1S_LP_EN here.
2465 * Doing so could cause underruns.
2472 * The spec says we shouldn't write when we don't need, because every write
2473 * causes WMs to be re-evaluated, expending some power.
2475 static void ilk_write_wm_values(struct drm_i915_private
*dev_priv
,
2476 struct ilk_wm_values
*results
)
2478 struct drm_device
*dev
= dev_priv
->dev
;
2479 struct ilk_wm_values
*previous
= &dev_priv
->wm
.hw
;
2483 dirty
= ilk_compute_wm_dirty(dev_priv
, previous
, results
);
2487 _ilk_disable_lp_wm(dev_priv
, dirty
);
2489 if (dirty
& WM_DIRTY_PIPE(PIPE_A
))
2490 I915_WRITE(WM0_PIPEA_ILK
, results
->wm_pipe
[0]);
2491 if (dirty
& WM_DIRTY_PIPE(PIPE_B
))
2492 I915_WRITE(WM0_PIPEB_ILK
, results
->wm_pipe
[1]);
2493 if (dirty
& WM_DIRTY_PIPE(PIPE_C
))
2494 I915_WRITE(WM0_PIPEC_IVB
, results
->wm_pipe
[2]);
2496 if (dirty
& WM_DIRTY_LINETIME(PIPE_A
))
2497 I915_WRITE(PIPE_WM_LINETIME(PIPE_A
), results
->wm_linetime
[0]);
2498 if (dirty
& WM_DIRTY_LINETIME(PIPE_B
))
2499 I915_WRITE(PIPE_WM_LINETIME(PIPE_B
), results
->wm_linetime
[1]);
2500 if (dirty
& WM_DIRTY_LINETIME(PIPE_C
))
2501 I915_WRITE(PIPE_WM_LINETIME(PIPE_C
), results
->wm_linetime
[2]);
2503 if (dirty
& WM_DIRTY_DDB
) {
2504 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
)) {
2505 val
= I915_READ(WM_MISC
);
2506 if (results
->partitioning
== INTEL_DDB_PART_1_2
)
2507 val
&= ~WM_MISC_DATA_PARTITION_5_6
;
2509 val
|= WM_MISC_DATA_PARTITION_5_6
;
2510 I915_WRITE(WM_MISC
, val
);
2512 val
= I915_READ(DISP_ARB_CTL2
);
2513 if (results
->partitioning
== INTEL_DDB_PART_1_2
)
2514 val
&= ~DISP_DATA_PARTITION_5_6
;
2516 val
|= DISP_DATA_PARTITION_5_6
;
2517 I915_WRITE(DISP_ARB_CTL2
, val
);
2521 if (dirty
& WM_DIRTY_FBC
) {
2522 val
= I915_READ(DISP_ARB_CTL
);
2523 if (results
->enable_fbc_wm
)
2524 val
&= ~DISP_FBC_WM_DIS
;
2526 val
|= DISP_FBC_WM_DIS
;
2527 I915_WRITE(DISP_ARB_CTL
, val
);
2530 if (dirty
& WM_DIRTY_LP(1) &&
2531 previous
->wm_lp_spr
[0] != results
->wm_lp_spr
[0])
2532 I915_WRITE(WM1S_LP_ILK
, results
->wm_lp_spr
[0]);
2534 if (INTEL_INFO(dev
)->gen
>= 7) {
2535 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp_spr
[1] != results
->wm_lp_spr
[1])
2536 I915_WRITE(WM2S_LP_IVB
, results
->wm_lp_spr
[1]);
2537 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp_spr
[2] != results
->wm_lp_spr
[2])
2538 I915_WRITE(WM3S_LP_IVB
, results
->wm_lp_spr
[2]);
2541 if (dirty
& WM_DIRTY_LP(1) && previous
->wm_lp
[0] != results
->wm_lp
[0])
2542 I915_WRITE(WM1_LP_ILK
, results
->wm_lp
[0]);
2543 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp
[1] != results
->wm_lp
[1])
2544 I915_WRITE(WM2_LP_ILK
, results
->wm_lp
[1]);
2545 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp
[2] != results
->wm_lp
[2])
2546 I915_WRITE(WM3_LP_ILK
, results
->wm_lp
[2]);
2548 dev_priv
->wm
.hw
= *results
;
2551 static bool ilk_disable_lp_wm(struct drm_device
*dev
)
2553 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2555 return _ilk_disable_lp_wm(dev_priv
, WM_DIRTY_LP_ALL
);
2559 * On gen9, we need to allocate Display Data Buffer (DDB) portions to the
2560 * different active planes.
2563 #define SKL_DDB_SIZE 896 /* in blocks */
2564 #define BXT_DDB_SIZE 512
2567 skl_ddb_get_pipe_allocation_limits(struct drm_device
*dev
,
2568 struct drm_crtc
*for_crtc
,
2569 const struct intel_wm_config
*config
,
2570 const struct skl_pipe_wm_parameters
*params
,
2571 struct skl_ddb_entry
*alloc
/* out */)
2573 struct drm_crtc
*crtc
;
2574 unsigned int pipe_size
, ddb_size
;
2575 int nth_active_pipe
;
2577 if (!params
->active
) {
2583 if (IS_BROXTON(dev
))
2584 ddb_size
= BXT_DDB_SIZE
;
2586 ddb_size
= SKL_DDB_SIZE
;
2588 ddb_size
-= 4; /* 4 blocks for bypass path allocation */
2590 nth_active_pipe
= 0;
2591 for_each_crtc(dev
, crtc
) {
2592 if (!to_intel_crtc(crtc
)->active
)
2595 if (crtc
== for_crtc
)
2601 pipe_size
= ddb_size
/ config
->num_pipes_active
;
2602 alloc
->start
= nth_active_pipe
* ddb_size
/ config
->num_pipes_active
;
2603 alloc
->end
= alloc
->start
+ pipe_size
;
2606 static unsigned int skl_cursor_allocation(const struct intel_wm_config
*config
)
2608 if (config
->num_pipes_active
== 1)
2614 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry
*entry
, u32 reg
)
2616 entry
->start
= reg
& 0x3ff;
2617 entry
->end
= (reg
>> 16) & 0x3ff;
2622 void skl_ddb_get_hw_state(struct drm_i915_private
*dev_priv
,
2623 struct skl_ddb_allocation
*ddb
/* out */)
2629 for_each_pipe(dev_priv
, pipe
) {
2630 for_each_plane(dev_priv
, pipe
, plane
) {
2631 val
= I915_READ(PLANE_BUF_CFG(pipe
, plane
));
2632 skl_ddb_entry_init_from_hw(&ddb
->plane
[pipe
][plane
],
2636 val
= I915_READ(CUR_BUF_CFG(pipe
));
2637 skl_ddb_entry_init_from_hw(&ddb
->cursor
[pipe
], val
);
2642 skl_plane_relative_data_rate(const struct intel_plane_wm_parameters
*p
, int y
)
2645 /* for planar format */
2646 if (p
->y_bytes_per_pixel
) {
2647 if (y
) /* y-plane data rate */
2648 return p
->horiz_pixels
* p
->vert_pixels
* p
->y_bytes_per_pixel
;
2649 else /* uv-plane data rate */
2650 return (p
->horiz_pixels
/2) * (p
->vert_pixels
/2) * p
->bytes_per_pixel
;
2653 /* for packed formats */
2654 return p
->horiz_pixels
* p
->vert_pixels
* p
->bytes_per_pixel
;
2658 * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
2659 * a 8192x4096@32bpp framebuffer:
2660 * 3 * 4096 * 8192 * 4 < 2^32
2663 skl_get_total_relative_data_rate(struct intel_crtc
*intel_crtc
,
2664 const struct skl_pipe_wm_parameters
*params
)
2666 unsigned int total_data_rate
= 0;
2669 for (plane
= 0; plane
< intel_num_planes(intel_crtc
); plane
++) {
2670 const struct intel_plane_wm_parameters
*p
;
2672 p
= ¶ms
->plane
[plane
];
2676 total_data_rate
+= skl_plane_relative_data_rate(p
, 0); /* packed/uv */
2677 if (p
->y_bytes_per_pixel
) {
2678 total_data_rate
+= skl_plane_relative_data_rate(p
, 1); /* y-plane */
2682 return total_data_rate
;
2686 skl_allocate_pipe_ddb(struct drm_crtc
*crtc
,
2687 const struct intel_wm_config
*config
,
2688 const struct skl_pipe_wm_parameters
*params
,
2689 struct skl_ddb_allocation
*ddb
/* out */)
2691 struct drm_device
*dev
= crtc
->dev
;
2692 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2693 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
2694 enum pipe pipe
= intel_crtc
->pipe
;
2695 struct skl_ddb_entry
*alloc
= &ddb
->pipe
[pipe
];
2696 uint16_t alloc_size
, start
, cursor_blocks
;
2697 uint16_t minimum
[I915_MAX_PLANES
];
2698 uint16_t y_minimum
[I915_MAX_PLANES
];
2699 unsigned int total_data_rate
;
2702 skl_ddb_get_pipe_allocation_limits(dev
, crtc
, config
, params
, alloc
);
2703 alloc_size
= skl_ddb_entry_size(alloc
);
2704 if (alloc_size
== 0) {
2705 memset(ddb
->plane
[pipe
], 0, sizeof(ddb
->plane
[pipe
]));
2706 memset(&ddb
->cursor
[pipe
], 0, sizeof(ddb
->cursor
[pipe
]));
2710 cursor_blocks
= skl_cursor_allocation(config
);
2711 ddb
->cursor
[pipe
].start
= alloc
->end
- cursor_blocks
;
2712 ddb
->cursor
[pipe
].end
= alloc
->end
;
2714 alloc_size
-= cursor_blocks
;
2715 alloc
->end
-= cursor_blocks
;
2717 /* 1. Allocate the mininum required blocks for each active plane */
2718 for_each_plane(dev_priv
, pipe
, plane
) {
2719 const struct intel_plane_wm_parameters
*p
;
2721 p
= ¶ms
->plane
[plane
];
2726 alloc_size
-= minimum
[plane
];
2727 y_minimum
[plane
] = p
->y_bytes_per_pixel
? 8 : 0;
2728 alloc_size
-= y_minimum
[plane
];
2732 * 2. Distribute the remaining space in proportion to the amount of
2733 * data each plane needs to fetch from memory.
2735 * FIXME: we may not allocate every single block here.
2737 total_data_rate
= skl_get_total_relative_data_rate(intel_crtc
, params
);
2739 start
= alloc
->start
;
2740 for (plane
= 0; plane
< intel_num_planes(intel_crtc
); plane
++) {
2741 const struct intel_plane_wm_parameters
*p
;
2742 unsigned int data_rate
, y_data_rate
;
2743 uint16_t plane_blocks
, y_plane_blocks
= 0;
2745 p
= ¶ms
->plane
[plane
];
2749 data_rate
= skl_plane_relative_data_rate(p
, 0);
2752 * allocation for (packed formats) or (uv-plane part of planar format):
2753 * promote the expression to 64 bits to avoid overflowing, the
2754 * result is < available as data_rate / total_data_rate < 1
2756 plane_blocks
= minimum
[plane
];
2757 plane_blocks
+= div_u64((uint64_t)alloc_size
* data_rate
,
2760 ddb
->plane
[pipe
][plane
].start
= start
;
2761 ddb
->plane
[pipe
][plane
].end
= start
+ plane_blocks
;
2763 start
+= plane_blocks
;
2766 * allocation for y_plane part of planar format:
2768 if (p
->y_bytes_per_pixel
) {
2769 y_data_rate
= skl_plane_relative_data_rate(p
, 1);
2770 y_plane_blocks
= y_minimum
[plane
];
2771 y_plane_blocks
+= div_u64((uint64_t)alloc_size
* y_data_rate
,
2774 ddb
->y_plane
[pipe
][plane
].start
= start
;
2775 ddb
->y_plane
[pipe
][plane
].end
= start
+ y_plane_blocks
;
2777 start
+= y_plane_blocks
;
2784 static uint32_t skl_pipe_pixel_rate(const struct intel_crtc_state
*config
)
2786 /* TODO: Take into account the scalers once we support them */
2787 return config
->base
.adjusted_mode
.crtc_clock
;
2791 * The max latency should be 257 (max the punit can code is 255 and we add 2us
2792 * for the read latency) and bytes_per_pixel should always be <= 8, so that
2793 * should allow pixel_rate up to ~2 GHz which seems sufficient since max
2794 * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
2796 static uint32_t skl_wm_method1(uint32_t pixel_rate
, uint8_t bytes_per_pixel
,
2799 uint32_t wm_intermediate_val
, ret
;
2804 wm_intermediate_val
= latency
* pixel_rate
* bytes_per_pixel
/ 512;
2805 ret
= DIV_ROUND_UP(wm_intermediate_val
, 1000);
2810 static uint32_t skl_wm_method2(uint32_t pixel_rate
, uint32_t pipe_htotal
,
2811 uint32_t horiz_pixels
, uint8_t bytes_per_pixel
,
2812 uint64_t tiling
, uint32_t latency
)
2815 uint32_t plane_bytes_per_line
, plane_blocks_per_line
;
2816 uint32_t wm_intermediate_val
;
2821 plane_bytes_per_line
= horiz_pixels
* bytes_per_pixel
;
2823 if (tiling
== I915_FORMAT_MOD_Y_TILED
||
2824 tiling
== I915_FORMAT_MOD_Yf_TILED
) {
2825 plane_bytes_per_line
*= 4;
2826 plane_blocks_per_line
= DIV_ROUND_UP(plane_bytes_per_line
, 512);
2827 plane_blocks_per_line
/= 4;
2829 plane_blocks_per_line
= DIV_ROUND_UP(plane_bytes_per_line
, 512);
2832 wm_intermediate_val
= latency
* pixel_rate
;
2833 ret
= DIV_ROUND_UP(wm_intermediate_val
, pipe_htotal
* 1000) *
2834 plane_blocks_per_line
;
2839 static bool skl_ddb_allocation_changed(const struct skl_ddb_allocation
*new_ddb
,
2840 const struct intel_crtc
*intel_crtc
)
2842 struct drm_device
*dev
= intel_crtc
->base
.dev
;
2843 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
2844 const struct skl_ddb_allocation
*cur_ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
2845 enum pipe pipe
= intel_crtc
->pipe
;
2847 if (memcmp(new_ddb
->plane
[pipe
], cur_ddb
->plane
[pipe
],
2848 sizeof(new_ddb
->plane
[pipe
])))
2851 if (memcmp(&new_ddb
->cursor
[pipe
], &cur_ddb
->cursor
[pipe
],
2852 sizeof(new_ddb
->cursor
[pipe
])))
2858 static void skl_compute_wm_global_parameters(struct drm_device
*dev
,
2859 struct intel_wm_config
*config
)
2861 struct drm_crtc
*crtc
;
2862 struct drm_plane
*plane
;
2864 list_for_each_entry(crtc
, &dev
->mode_config
.crtc_list
, head
)
2865 config
->num_pipes_active
+= to_intel_crtc(crtc
)->active
;
2867 /* FIXME: I don't think we need those two global parameters on SKL */
2868 list_for_each_entry(plane
, &dev
->mode_config
.plane_list
, head
) {
2869 struct intel_plane
*intel_plane
= to_intel_plane(plane
);
2871 config
->sprites_enabled
|= intel_plane
->wm
.enabled
;
2872 config
->sprites_scaled
|= intel_plane
->wm
.scaled
;
2876 static void skl_compute_wm_pipe_parameters(struct drm_crtc
*crtc
,
2877 struct skl_pipe_wm_parameters
*p
)
2879 struct drm_device
*dev
= crtc
->dev
;
2880 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
2881 enum pipe pipe
= intel_crtc
->pipe
;
2882 struct drm_plane
*plane
;
2883 struct drm_framebuffer
*fb
;
2884 int i
= 1; /* Index for sprite planes start */
2886 p
->active
= intel_crtc
->active
;
2888 p
->pipe_htotal
= intel_crtc
->config
->base
.adjusted_mode
.crtc_htotal
;
2889 p
->pixel_rate
= skl_pipe_pixel_rate(intel_crtc
->config
);
2891 fb
= crtc
->primary
->state
->fb
;
2892 /* For planar: Bpp is for uv plane, y_Bpp is for y plane */
2894 p
->plane
[0].enabled
= true;
2895 p
->plane
[0].bytes_per_pixel
= fb
->pixel_format
== DRM_FORMAT_NV12
?
2896 drm_format_plane_cpp(fb
->pixel_format
, 1) : fb
->bits_per_pixel
/ 8;
2897 p
->plane
[0].y_bytes_per_pixel
= fb
->pixel_format
== DRM_FORMAT_NV12
?
2898 drm_format_plane_cpp(fb
->pixel_format
, 0) : 0;
2899 p
->plane
[0].tiling
= fb
->modifier
[0];
2901 p
->plane
[0].enabled
= false;
2902 p
->plane
[0].bytes_per_pixel
= 0;
2903 p
->plane
[0].y_bytes_per_pixel
= 0;
2904 p
->plane
[0].tiling
= DRM_FORMAT_MOD_NONE
;
2906 p
->plane
[0].horiz_pixels
= intel_crtc
->config
->pipe_src_w
;
2907 p
->plane
[0].vert_pixels
= intel_crtc
->config
->pipe_src_h
;
2908 p
->plane
[0].rotation
= crtc
->primary
->state
->rotation
;
2910 fb
= crtc
->cursor
->state
->fb
;
2911 p
->cursor
.y_bytes_per_pixel
= 0;
2913 p
->cursor
.enabled
= true;
2914 p
->cursor
.bytes_per_pixel
= fb
->bits_per_pixel
/ 8;
2915 p
->cursor
.horiz_pixels
= crtc
->cursor
->state
->crtc_w
;
2916 p
->cursor
.vert_pixels
= crtc
->cursor
->state
->crtc_h
;
2918 p
->cursor
.enabled
= false;
2919 p
->cursor
.bytes_per_pixel
= 0;
2920 p
->cursor
.horiz_pixels
= 64;
2921 p
->cursor
.vert_pixels
= 64;
2925 list_for_each_entry(plane
, &dev
->mode_config
.plane_list
, head
) {
2926 struct intel_plane
*intel_plane
= to_intel_plane(plane
);
2928 if (intel_plane
->pipe
== pipe
&&
2929 plane
->type
== DRM_PLANE_TYPE_OVERLAY
)
2930 p
->plane
[i
++] = intel_plane
->wm
;
2934 static bool skl_compute_plane_wm(const struct drm_i915_private
*dev_priv
,
2935 struct skl_pipe_wm_parameters
*p
,
2936 struct intel_plane_wm_parameters
*p_params
,
2937 uint16_t ddb_allocation
,
2939 uint16_t *out_blocks
, /* out */
2940 uint8_t *out_lines
/* out */)
2942 uint32_t latency
= dev_priv
->wm
.skl_latency
[level
];
2943 uint32_t method1
, method2
;
2944 uint32_t plane_bytes_per_line
, plane_blocks_per_line
;
2945 uint32_t res_blocks
, res_lines
;
2946 uint32_t selected_result
;
2947 uint8_t bytes_per_pixel
;
2949 if (latency
== 0 || !p
->active
|| !p_params
->enabled
)
2952 bytes_per_pixel
= p_params
->y_bytes_per_pixel
?
2953 p_params
->y_bytes_per_pixel
:
2954 p_params
->bytes_per_pixel
;
2955 method1
= skl_wm_method1(p
->pixel_rate
,
2958 method2
= skl_wm_method2(p
->pixel_rate
,
2960 p_params
->horiz_pixels
,
2965 plane_bytes_per_line
= p_params
->horiz_pixels
* bytes_per_pixel
;
2966 plane_blocks_per_line
= DIV_ROUND_UP(plane_bytes_per_line
, 512);
2968 if (p_params
->tiling
== I915_FORMAT_MOD_Y_TILED
||
2969 p_params
->tiling
== I915_FORMAT_MOD_Yf_TILED
) {
2970 uint32_t min_scanlines
= 4;
2971 uint32_t y_tile_minimum
;
2972 if (intel_rotation_90_or_270(p_params
->rotation
)) {
2973 switch (p_params
->bytes_per_pixel
) {
2981 WARN(1, "Unsupported pixel depth for rotation");
2984 y_tile_minimum
= plane_blocks_per_line
* min_scanlines
;
2985 selected_result
= max(method2
, y_tile_minimum
);
2987 if ((ddb_allocation
/ plane_blocks_per_line
) >= 1)
2988 selected_result
= min(method1
, method2
);
2990 selected_result
= method1
;
2993 res_blocks
= selected_result
+ 1;
2994 res_lines
= DIV_ROUND_UP(selected_result
, plane_blocks_per_line
);
2996 if (level
>= 1 && level
<= 7) {
2997 if (p_params
->tiling
== I915_FORMAT_MOD_Y_TILED
||
2998 p_params
->tiling
== I915_FORMAT_MOD_Yf_TILED
)
3004 if (res_blocks
>= ddb_allocation
|| res_lines
> 31)
3007 *out_blocks
= res_blocks
;
3008 *out_lines
= res_lines
;
3013 static void skl_compute_wm_level(const struct drm_i915_private
*dev_priv
,
3014 struct skl_ddb_allocation
*ddb
,
3015 struct skl_pipe_wm_parameters
*p
,
3019 struct skl_wm_level
*result
)
3021 uint16_t ddb_blocks
;
3024 for (i
= 0; i
< num_planes
; i
++) {
3025 ddb_blocks
= skl_ddb_entry_size(&ddb
->plane
[pipe
][i
]);
3027 result
->plane_en
[i
] = skl_compute_plane_wm(dev_priv
,
3031 &result
->plane_res_b
[i
],
3032 &result
->plane_res_l
[i
]);
3035 ddb_blocks
= skl_ddb_entry_size(&ddb
->cursor
[pipe
]);
3036 result
->cursor_en
= skl_compute_plane_wm(dev_priv
, p
, &p
->cursor
,
3038 &result
->cursor_res_b
,
3039 &result
->cursor_res_l
);
3043 skl_compute_linetime_wm(struct drm_crtc
*crtc
, struct skl_pipe_wm_parameters
*p
)
3045 if (!to_intel_crtc(crtc
)->active
)
3048 return DIV_ROUND_UP(8 * p
->pipe_htotal
* 1000, p
->pixel_rate
);
3052 static void skl_compute_transition_wm(struct drm_crtc
*crtc
,
3053 struct skl_pipe_wm_parameters
*params
,
3054 struct skl_wm_level
*trans_wm
/* out */)
3056 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3059 if (!params
->active
)
3062 /* Until we know more, just disable transition WMs */
3063 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++)
3064 trans_wm
->plane_en
[i
] = false;
3065 trans_wm
->cursor_en
= false;
3068 static void skl_compute_pipe_wm(struct drm_crtc
*crtc
,
3069 struct skl_ddb_allocation
*ddb
,
3070 struct skl_pipe_wm_parameters
*params
,
3071 struct skl_pipe_wm
*pipe_wm
)
3073 struct drm_device
*dev
= crtc
->dev
;
3074 const struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3075 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3076 int level
, max_level
= ilk_wm_max_level(dev
);
3078 for (level
= 0; level
<= max_level
; level
++) {
3079 skl_compute_wm_level(dev_priv
, ddb
, params
, intel_crtc
->pipe
,
3080 level
, intel_num_planes(intel_crtc
),
3081 &pipe_wm
->wm
[level
]);
3083 pipe_wm
->linetime
= skl_compute_linetime_wm(crtc
, params
);
3085 skl_compute_transition_wm(crtc
, params
, &pipe_wm
->trans_wm
);
3088 static void skl_compute_wm_results(struct drm_device
*dev
,
3089 struct skl_pipe_wm_parameters
*p
,
3090 struct skl_pipe_wm
*p_wm
,
3091 struct skl_wm_values
*r
,
3092 struct intel_crtc
*intel_crtc
)
3094 int level
, max_level
= ilk_wm_max_level(dev
);
3095 enum pipe pipe
= intel_crtc
->pipe
;
3099 for (level
= 0; level
<= max_level
; level
++) {
3100 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++) {
3103 temp
|= p_wm
->wm
[level
].plane_res_l
[i
] <<
3104 PLANE_WM_LINES_SHIFT
;
3105 temp
|= p_wm
->wm
[level
].plane_res_b
[i
];
3106 if (p_wm
->wm
[level
].plane_en
[i
])
3107 temp
|= PLANE_WM_EN
;
3109 r
->plane
[pipe
][i
][level
] = temp
;
3114 temp
|= p_wm
->wm
[level
].cursor_res_l
<< PLANE_WM_LINES_SHIFT
;
3115 temp
|= p_wm
->wm
[level
].cursor_res_b
;
3117 if (p_wm
->wm
[level
].cursor_en
)
3118 temp
|= PLANE_WM_EN
;
3120 r
->cursor
[pipe
][level
] = temp
;
3124 /* transition WMs */
3125 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++) {
3127 temp
|= p_wm
->trans_wm
.plane_res_l
[i
] << PLANE_WM_LINES_SHIFT
;
3128 temp
|= p_wm
->trans_wm
.plane_res_b
[i
];
3129 if (p_wm
->trans_wm
.plane_en
[i
])
3130 temp
|= PLANE_WM_EN
;
3132 r
->plane_trans
[pipe
][i
] = temp
;
3136 temp
|= p_wm
->trans_wm
.cursor_res_l
<< PLANE_WM_LINES_SHIFT
;
3137 temp
|= p_wm
->trans_wm
.cursor_res_b
;
3138 if (p_wm
->trans_wm
.cursor_en
)
3139 temp
|= PLANE_WM_EN
;
3141 r
->cursor_trans
[pipe
] = temp
;
3143 r
->wm_linetime
[pipe
] = p_wm
->linetime
;
3146 static void skl_ddb_entry_write(struct drm_i915_private
*dev_priv
, uint32_t reg
,
3147 const struct skl_ddb_entry
*entry
)
3150 I915_WRITE(reg
, (entry
->end
- 1) << 16 | entry
->start
);
3155 static void skl_write_wm_values(struct drm_i915_private
*dev_priv
,
3156 const struct skl_wm_values
*new)
3158 struct drm_device
*dev
= dev_priv
->dev
;
3159 struct intel_crtc
*crtc
;
3161 list_for_each_entry(crtc
, &dev
->mode_config
.crtc_list
, base
.head
) {
3162 int i
, level
, max_level
= ilk_wm_max_level(dev
);
3163 enum pipe pipe
= crtc
->pipe
;
3165 if (!new->dirty
[pipe
])
3168 I915_WRITE(PIPE_WM_LINETIME(pipe
), new->wm_linetime
[pipe
]);
3170 for (level
= 0; level
<= max_level
; level
++) {
3171 for (i
= 0; i
< intel_num_planes(crtc
); i
++)
3172 I915_WRITE(PLANE_WM(pipe
, i
, level
),
3173 new->plane
[pipe
][i
][level
]);
3174 I915_WRITE(CUR_WM(pipe
, level
),
3175 new->cursor
[pipe
][level
]);
3177 for (i
= 0; i
< intel_num_planes(crtc
); i
++)
3178 I915_WRITE(PLANE_WM_TRANS(pipe
, i
),
3179 new->plane_trans
[pipe
][i
]);
3180 I915_WRITE(CUR_WM_TRANS(pipe
), new->cursor_trans
[pipe
]);
3182 for (i
= 0; i
< intel_num_planes(crtc
); i
++) {
3183 skl_ddb_entry_write(dev_priv
,
3184 PLANE_BUF_CFG(pipe
, i
),
3185 &new->ddb
.plane
[pipe
][i
]);
3186 skl_ddb_entry_write(dev_priv
,
3187 PLANE_NV12_BUF_CFG(pipe
, i
),
3188 &new->ddb
.y_plane
[pipe
][i
]);
3191 skl_ddb_entry_write(dev_priv
, CUR_BUF_CFG(pipe
),
3192 &new->ddb
.cursor
[pipe
]);
3197 * When setting up a new DDB allocation arrangement, we need to correctly
3198 * sequence the times at which the new allocations for the pipes are taken into
3199 * account or we'll have pipes fetching from space previously allocated to
3202 * Roughly the sequence looks like:
3203 * 1. re-allocate the pipe(s) with the allocation being reduced and not
3204 * overlapping with a previous light-up pipe (another way to put it is:
3205 * pipes with their new allocation strickly included into their old ones).
3206 * 2. re-allocate the other pipes that get their allocation reduced
3207 * 3. allocate the pipes having their allocation increased
3209 * Steps 1. and 2. are here to take care of the following case:
3210 * - Initially DDB looks like this:
3213 * - pipe B has a reduced DDB allocation that overlaps with the old pipe C
3217 * We need to sequence the re-allocation: C, B, A (and not B, C, A).
3221 skl_wm_flush_pipe(struct drm_i915_private
*dev_priv
, enum pipe pipe
, int pass
)
3225 DRM_DEBUG_KMS("flush pipe %c (pass %d)\n", pipe_name(pipe
), pass
);
3227 for_each_plane(dev_priv
, pipe
, plane
) {
3228 I915_WRITE(PLANE_SURF(pipe
, plane
),
3229 I915_READ(PLANE_SURF(pipe
, plane
)));
3231 I915_WRITE(CURBASE(pipe
), I915_READ(CURBASE(pipe
)));
3235 skl_ddb_allocation_included(const struct skl_ddb_allocation
*old
,
3236 const struct skl_ddb_allocation
*new,
3239 uint16_t old_size
, new_size
;
3241 old_size
= skl_ddb_entry_size(&old
->pipe
[pipe
]);
3242 new_size
= skl_ddb_entry_size(&new->pipe
[pipe
]);
3244 return old_size
!= new_size
&&
3245 new->pipe
[pipe
].start
>= old
->pipe
[pipe
].start
&&
3246 new->pipe
[pipe
].end
<= old
->pipe
[pipe
].end
;
3249 static void skl_flush_wm_values(struct drm_i915_private
*dev_priv
,
3250 struct skl_wm_values
*new_values
)
3252 struct drm_device
*dev
= dev_priv
->dev
;
3253 struct skl_ddb_allocation
*cur_ddb
, *new_ddb
;
3254 bool reallocated
[I915_MAX_PIPES
] = {};
3255 struct intel_crtc
*crtc
;
3258 new_ddb
= &new_values
->ddb
;
3259 cur_ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
3262 * First pass: flush the pipes with the new allocation contained into
3265 * We'll wait for the vblank on those pipes to ensure we can safely
3266 * re-allocate the freed space without this pipe fetching from it.
3268 for_each_intel_crtc(dev
, crtc
) {
3274 if (!skl_ddb_allocation_included(cur_ddb
, new_ddb
, pipe
))
3277 skl_wm_flush_pipe(dev_priv
, pipe
, 1);
3278 intel_wait_for_vblank(dev
, pipe
);
3280 reallocated
[pipe
] = true;
3285 * Second pass: flush the pipes that are having their allocation
3286 * reduced, but overlapping with a previous allocation.
3288 * Here as well we need to wait for the vblank to make sure the freed
3289 * space is not used anymore.
3291 for_each_intel_crtc(dev
, crtc
) {
3297 if (reallocated
[pipe
])
3300 if (skl_ddb_entry_size(&new_ddb
->pipe
[pipe
]) <
3301 skl_ddb_entry_size(&cur_ddb
->pipe
[pipe
])) {
3302 skl_wm_flush_pipe(dev_priv
, pipe
, 2);
3303 intel_wait_for_vblank(dev
, pipe
);
3304 reallocated
[pipe
] = true;
3309 * Third pass: flush the pipes that got more space allocated.
3311 * We don't need to actively wait for the update here, next vblank
3312 * will just get more DDB space with the correct WM values.
3314 for_each_intel_crtc(dev
, crtc
) {
3321 * At this point, only the pipes more space than before are
3322 * left to re-allocate.
3324 if (reallocated
[pipe
])
3327 skl_wm_flush_pipe(dev_priv
, pipe
, 3);
3331 static bool skl_update_pipe_wm(struct drm_crtc
*crtc
,
3332 struct skl_pipe_wm_parameters
*params
,
3333 struct intel_wm_config
*config
,
3334 struct skl_ddb_allocation
*ddb
, /* out */
3335 struct skl_pipe_wm
*pipe_wm
/* out */)
3337 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3339 skl_compute_wm_pipe_parameters(crtc
, params
);
3340 skl_allocate_pipe_ddb(crtc
, config
, params
, ddb
);
3341 skl_compute_pipe_wm(crtc
, ddb
, params
, pipe_wm
);
3343 if (!memcmp(&intel_crtc
->wm
.skl_active
, pipe_wm
, sizeof(*pipe_wm
)))
3346 intel_crtc
->wm
.skl_active
= *pipe_wm
;
3351 static void skl_update_other_pipe_wm(struct drm_device
*dev
,
3352 struct drm_crtc
*crtc
,
3353 struct intel_wm_config
*config
,
3354 struct skl_wm_values
*r
)
3356 struct intel_crtc
*intel_crtc
;
3357 struct intel_crtc
*this_crtc
= to_intel_crtc(crtc
);
3360 * If the WM update hasn't changed the allocation for this_crtc (the
3361 * crtc we are currently computing the new WM values for), other
3362 * enabled crtcs will keep the same allocation and we don't need to
3363 * recompute anything for them.
3365 if (!skl_ddb_allocation_changed(&r
->ddb
, this_crtc
))
3369 * Otherwise, because of this_crtc being freshly enabled/disabled, the
3370 * other active pipes need new DDB allocation and WM values.
3372 list_for_each_entry(intel_crtc
, &dev
->mode_config
.crtc_list
,
3374 struct skl_pipe_wm_parameters params
= {};
3375 struct skl_pipe_wm pipe_wm
= {};
3378 if (this_crtc
->pipe
== intel_crtc
->pipe
)
3381 if (!intel_crtc
->active
)
3384 wm_changed
= skl_update_pipe_wm(&intel_crtc
->base
,
3389 * If we end up re-computing the other pipe WM values, it's
3390 * because it was really needed, so we expect the WM values to
3393 WARN_ON(!wm_changed
);
3395 skl_compute_wm_results(dev
, ¶ms
, &pipe_wm
, r
, intel_crtc
);
3396 r
->dirty
[intel_crtc
->pipe
] = true;
3400 static void skl_update_wm(struct drm_crtc
*crtc
)
3402 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3403 struct drm_device
*dev
= crtc
->dev
;
3404 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3405 struct skl_pipe_wm_parameters params
= {};
3406 struct skl_wm_values
*results
= &dev_priv
->wm
.skl_results
;
3407 struct skl_pipe_wm pipe_wm
= {};
3408 struct intel_wm_config config
= {};
3410 memset(results
, 0, sizeof(*results
));
3412 skl_compute_wm_global_parameters(dev
, &config
);
3414 if (!skl_update_pipe_wm(crtc
, ¶ms
, &config
,
3415 &results
->ddb
, &pipe_wm
))
3418 skl_compute_wm_results(dev
, ¶ms
, &pipe_wm
, results
, intel_crtc
);
3419 results
->dirty
[intel_crtc
->pipe
] = true;
3421 skl_update_other_pipe_wm(dev
, crtc
, &config
, results
);
3422 skl_write_wm_values(dev_priv
, results
);
3423 skl_flush_wm_values(dev_priv
, results
);
3425 /* store the new configuration */
3426 dev_priv
->wm
.skl_hw
= *results
;
3430 skl_update_sprite_wm(struct drm_plane
*plane
, struct drm_crtc
*crtc
,
3431 uint32_t sprite_width
, uint32_t sprite_height
,
3432 int pixel_size
, bool enabled
, bool scaled
)
3434 struct intel_plane
*intel_plane
= to_intel_plane(plane
);
3435 struct drm_framebuffer
*fb
= plane
->state
->fb
;
3437 intel_plane
->wm
.enabled
= enabled
;
3438 intel_plane
->wm
.scaled
= scaled
;
3439 intel_plane
->wm
.horiz_pixels
= sprite_width
;
3440 intel_plane
->wm
.vert_pixels
= sprite_height
;
3441 intel_plane
->wm
.tiling
= DRM_FORMAT_MOD_NONE
;
3443 /* For planar: Bpp is for UV plane, y_Bpp is for Y plane */
3444 intel_plane
->wm
.bytes_per_pixel
=
3445 (fb
&& fb
->pixel_format
== DRM_FORMAT_NV12
) ?
3446 drm_format_plane_cpp(plane
->state
->fb
->pixel_format
, 1) : pixel_size
;
3447 intel_plane
->wm
.y_bytes_per_pixel
=
3448 (fb
&& fb
->pixel_format
== DRM_FORMAT_NV12
) ?
3449 drm_format_plane_cpp(plane
->state
->fb
->pixel_format
, 0) : 0;
3452 * Framebuffer can be NULL on plane disable, but it does not
3453 * matter for watermarks if we assume no tiling in that case.
3456 intel_plane
->wm
.tiling
= fb
->modifier
[0];
3457 intel_plane
->wm
.rotation
= plane
->state
->rotation
;
3459 skl_update_wm(crtc
);
3462 static void ilk_update_wm(struct drm_crtc
*crtc
)
3464 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3465 struct drm_device
*dev
= crtc
->dev
;
3466 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3467 struct ilk_wm_maximums max
;
3468 struct ilk_pipe_wm_parameters params
= {};
3469 struct ilk_wm_values results
= {};
3470 enum intel_ddb_partitioning partitioning
;
3471 struct intel_pipe_wm pipe_wm
= {};
3472 struct intel_pipe_wm lp_wm_1_2
= {}, lp_wm_5_6
= {}, *best_lp_wm
;
3473 struct intel_wm_config config
= {};
3475 ilk_compute_wm_parameters(crtc
, ¶ms
);
3477 intel_compute_pipe_wm(crtc
, ¶ms
, &pipe_wm
);
3479 if (!memcmp(&intel_crtc
->wm
.active
, &pipe_wm
, sizeof(pipe_wm
)))
3482 intel_crtc
->wm
.active
= pipe_wm
;
3484 ilk_compute_wm_config(dev
, &config
);
3486 ilk_compute_wm_maximums(dev
, 1, &config
, INTEL_DDB_PART_1_2
, &max
);
3487 ilk_wm_merge(dev
, &config
, &max
, &lp_wm_1_2
);
3489 /* 5/6 split only in single pipe config on IVB+ */
3490 if (INTEL_INFO(dev
)->gen
>= 7 &&
3491 config
.num_pipes_active
== 1 && config
.sprites_enabled
) {
3492 ilk_compute_wm_maximums(dev
, 1, &config
, INTEL_DDB_PART_5_6
, &max
);
3493 ilk_wm_merge(dev
, &config
, &max
, &lp_wm_5_6
);
3495 best_lp_wm
= ilk_find_best_result(dev
, &lp_wm_1_2
, &lp_wm_5_6
);
3497 best_lp_wm
= &lp_wm_1_2
;
3500 partitioning
= (best_lp_wm
== &lp_wm_1_2
) ?
3501 INTEL_DDB_PART_1_2
: INTEL_DDB_PART_5_6
;
3503 ilk_compute_wm_results(dev
, best_lp_wm
, partitioning
, &results
);
3505 ilk_write_wm_values(dev_priv
, &results
);
3509 ilk_update_sprite_wm(struct drm_plane
*plane
,
3510 struct drm_crtc
*crtc
,
3511 uint32_t sprite_width
, uint32_t sprite_height
,
3512 int pixel_size
, bool enabled
, bool scaled
)
3514 struct drm_device
*dev
= plane
->dev
;
3515 struct intel_plane
*intel_plane
= to_intel_plane(plane
);
3517 intel_plane
->wm
.enabled
= enabled
;
3518 intel_plane
->wm
.scaled
= scaled
;
3519 intel_plane
->wm
.horiz_pixels
= sprite_width
;
3520 intel_plane
->wm
.vert_pixels
= sprite_width
;
3521 intel_plane
->wm
.bytes_per_pixel
= pixel_size
;
3524 * IVB workaround: must disable low power watermarks for at least
3525 * one frame before enabling scaling. LP watermarks can be re-enabled
3526 * when scaling is disabled.
3528 * WaCxSRDisabledForSpriteScaling:ivb
3530 if (IS_IVYBRIDGE(dev
) && scaled
&& ilk_disable_lp_wm(dev
))
3531 intel_wait_for_vblank(dev
, intel_plane
->pipe
);
3533 ilk_update_wm(crtc
);
3536 static void skl_pipe_wm_active_state(uint32_t val
,
3537 struct skl_pipe_wm
*active
,
3543 bool is_enabled
= (val
& PLANE_WM_EN
) != 0;
3547 active
->wm
[level
].plane_en
[i
] = is_enabled
;
3548 active
->wm
[level
].plane_res_b
[i
] =
3549 val
& PLANE_WM_BLOCKS_MASK
;
3550 active
->wm
[level
].plane_res_l
[i
] =
3551 (val
>> PLANE_WM_LINES_SHIFT
) &
3552 PLANE_WM_LINES_MASK
;
3554 active
->wm
[level
].cursor_en
= is_enabled
;
3555 active
->wm
[level
].cursor_res_b
=
3556 val
& PLANE_WM_BLOCKS_MASK
;
3557 active
->wm
[level
].cursor_res_l
=
3558 (val
>> PLANE_WM_LINES_SHIFT
) &
3559 PLANE_WM_LINES_MASK
;
3563 active
->trans_wm
.plane_en
[i
] = is_enabled
;
3564 active
->trans_wm
.plane_res_b
[i
] =
3565 val
& PLANE_WM_BLOCKS_MASK
;
3566 active
->trans_wm
.plane_res_l
[i
] =
3567 (val
>> PLANE_WM_LINES_SHIFT
) &
3568 PLANE_WM_LINES_MASK
;
3570 active
->trans_wm
.cursor_en
= is_enabled
;
3571 active
->trans_wm
.cursor_res_b
=
3572 val
& PLANE_WM_BLOCKS_MASK
;
3573 active
->trans_wm
.cursor_res_l
=
3574 (val
>> PLANE_WM_LINES_SHIFT
) &
3575 PLANE_WM_LINES_MASK
;
3580 static void skl_pipe_wm_get_hw_state(struct drm_crtc
*crtc
)
3582 struct drm_device
*dev
= crtc
->dev
;
3583 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3584 struct skl_wm_values
*hw
= &dev_priv
->wm
.skl_hw
;
3585 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3586 struct skl_pipe_wm
*active
= &intel_crtc
->wm
.skl_active
;
3587 enum pipe pipe
= intel_crtc
->pipe
;
3588 int level
, i
, max_level
;
3591 max_level
= ilk_wm_max_level(dev
);
3593 hw
->wm_linetime
[pipe
] = I915_READ(PIPE_WM_LINETIME(pipe
));
3595 for (level
= 0; level
<= max_level
; level
++) {
3596 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++)
3597 hw
->plane
[pipe
][i
][level
] =
3598 I915_READ(PLANE_WM(pipe
, i
, level
));
3599 hw
->cursor
[pipe
][level
] = I915_READ(CUR_WM(pipe
, level
));
3602 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++)
3603 hw
->plane_trans
[pipe
][i
] = I915_READ(PLANE_WM_TRANS(pipe
, i
));
3604 hw
->cursor_trans
[pipe
] = I915_READ(CUR_WM_TRANS(pipe
));
3606 if (!intel_crtc
->active
)
3609 hw
->dirty
[pipe
] = true;
3611 active
->linetime
= hw
->wm_linetime
[pipe
];
3613 for (level
= 0; level
<= max_level
; level
++) {
3614 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++) {
3615 temp
= hw
->plane
[pipe
][i
][level
];
3616 skl_pipe_wm_active_state(temp
, active
, false,
3619 temp
= hw
->cursor
[pipe
][level
];
3620 skl_pipe_wm_active_state(temp
, active
, false, true, i
, level
);
3623 for (i
= 0; i
< intel_num_planes(intel_crtc
); i
++) {
3624 temp
= hw
->plane_trans
[pipe
][i
];
3625 skl_pipe_wm_active_state(temp
, active
, true, false, i
, 0);
3628 temp
= hw
->cursor_trans
[pipe
];
3629 skl_pipe_wm_active_state(temp
, active
, true, true, i
, 0);
3632 void skl_wm_get_hw_state(struct drm_device
*dev
)
3634 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3635 struct skl_ddb_allocation
*ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
3636 struct drm_crtc
*crtc
;
3638 skl_ddb_get_hw_state(dev_priv
, ddb
);
3639 list_for_each_entry(crtc
, &dev
->mode_config
.crtc_list
, head
)
3640 skl_pipe_wm_get_hw_state(crtc
);
3643 static void ilk_pipe_wm_get_hw_state(struct drm_crtc
*crtc
)
3645 struct drm_device
*dev
= crtc
->dev
;
3646 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3647 struct ilk_wm_values
*hw
= &dev_priv
->wm
.hw
;
3648 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3649 struct intel_pipe_wm
*active
= &intel_crtc
->wm
.active
;
3650 enum pipe pipe
= intel_crtc
->pipe
;
3651 static const unsigned int wm0_pipe_reg
[] = {
3652 [PIPE_A
] = WM0_PIPEA_ILK
,
3653 [PIPE_B
] = WM0_PIPEB_ILK
,
3654 [PIPE_C
] = WM0_PIPEC_IVB
,
3657 hw
->wm_pipe
[pipe
] = I915_READ(wm0_pipe_reg
[pipe
]);
3658 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
3659 hw
->wm_linetime
[pipe
] = I915_READ(PIPE_WM_LINETIME(pipe
));
3661 active
->pipe_enabled
= intel_crtc
->active
;
3663 if (active
->pipe_enabled
) {
3664 u32 tmp
= hw
->wm_pipe
[pipe
];
3667 * For active pipes LP0 watermark is marked as
3668 * enabled, and LP1+ watermaks as disabled since
3669 * we can't really reverse compute them in case
3670 * multiple pipes are active.
3672 active
->wm
[0].enable
= true;
3673 active
->wm
[0].pri_val
= (tmp
& WM0_PIPE_PLANE_MASK
) >> WM0_PIPE_PLANE_SHIFT
;
3674 active
->wm
[0].spr_val
= (tmp
& WM0_PIPE_SPRITE_MASK
) >> WM0_PIPE_SPRITE_SHIFT
;
3675 active
->wm
[0].cur_val
= tmp
& WM0_PIPE_CURSOR_MASK
;
3676 active
->linetime
= hw
->wm_linetime
[pipe
];
3678 int level
, max_level
= ilk_wm_max_level(dev
);
3681 * For inactive pipes, all watermark levels
3682 * should be marked as enabled but zeroed,
3683 * which is what we'd compute them to.
3685 for (level
= 0; level
<= max_level
; level
++)
3686 active
->wm
[level
].enable
= true;
3690 void ilk_wm_get_hw_state(struct drm_device
*dev
)
3692 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3693 struct ilk_wm_values
*hw
= &dev_priv
->wm
.hw
;
3694 struct drm_crtc
*crtc
;
3696 for_each_crtc(dev
, crtc
)
3697 ilk_pipe_wm_get_hw_state(crtc
);
3699 hw
->wm_lp
[0] = I915_READ(WM1_LP_ILK
);
3700 hw
->wm_lp
[1] = I915_READ(WM2_LP_ILK
);
3701 hw
->wm_lp
[2] = I915_READ(WM3_LP_ILK
);
3703 hw
->wm_lp_spr
[0] = I915_READ(WM1S_LP_ILK
);
3704 if (INTEL_INFO(dev
)->gen
>= 7) {
3705 hw
->wm_lp_spr
[1] = I915_READ(WM2S_LP_IVB
);
3706 hw
->wm_lp_spr
[2] = I915_READ(WM3S_LP_IVB
);
3709 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
3710 hw
->partitioning
= (I915_READ(WM_MISC
) & WM_MISC_DATA_PARTITION_5_6
) ?
3711 INTEL_DDB_PART_5_6
: INTEL_DDB_PART_1_2
;
3712 else if (IS_IVYBRIDGE(dev
))
3713 hw
->partitioning
= (I915_READ(DISP_ARB_CTL2
) & DISP_DATA_PARTITION_5_6
) ?
3714 INTEL_DDB_PART_5_6
: INTEL_DDB_PART_1_2
;
3717 !(I915_READ(DISP_ARB_CTL
) & DISP_FBC_WM_DIS
);
3721 * intel_update_watermarks - update FIFO watermark values based on current modes
3723 * Calculate watermark values for the various WM regs based on current mode
3724 * and plane configuration.
3726 * There are several cases to deal with here:
3727 * - normal (i.e. non-self-refresh)
3728 * - self-refresh (SR) mode
3729 * - lines are large relative to FIFO size (buffer can hold up to 2)
3730 * - lines are small relative to FIFO size (buffer can hold more than 2
3731 * lines), so need to account for TLB latency
3733 * The normal calculation is:
3734 * watermark = dotclock * bytes per pixel * latency
3735 * where latency is platform & configuration dependent (we assume pessimal
3738 * The SR calculation is:
3739 * watermark = (trunc(latency/line time)+1) * surface width *
3742 * line time = htotal / dotclock
3743 * surface width = hdisplay for normal plane and 64 for cursor
3744 * and latency is assumed to be high, as above.
3746 * The final value programmed to the register should always be rounded up,
3747 * and include an extra 2 entries to account for clock crossings.
3749 * We don't use the sprite, so we can ignore that. And on Crestline we have
3750 * to set the non-SR watermarks to 8.
3752 void intel_update_watermarks(struct drm_crtc
*crtc
)
3754 struct drm_i915_private
*dev_priv
= crtc
->dev
->dev_private
;
3756 if (dev_priv
->display
.update_wm
)
3757 dev_priv
->display
.update_wm(crtc
);
3760 void intel_update_sprite_watermarks(struct drm_plane
*plane
,
3761 struct drm_crtc
*crtc
,
3762 uint32_t sprite_width
,
3763 uint32_t sprite_height
,
3765 bool enabled
, bool scaled
)
3767 struct drm_i915_private
*dev_priv
= plane
->dev
->dev_private
;
3769 if (dev_priv
->display
.update_sprite_wm
)
3770 dev_priv
->display
.update_sprite_wm(plane
, crtc
,
3771 sprite_width
, sprite_height
,
3772 pixel_size
, enabled
, scaled
);
3776 * Lock protecting IPS related data structures
3778 DEFINE_SPINLOCK(mchdev_lock
);
3780 /* Global for IPS driver to get at the current i915 device. Protected by
3782 static struct drm_i915_private
*i915_mch_dev
;
3784 bool ironlake_set_drps(struct drm_device
*dev
, u8 val
)
3786 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3789 assert_spin_locked(&mchdev_lock
);
3791 rgvswctl
= I915_READ16(MEMSWCTL
);
3792 if (rgvswctl
& MEMCTL_CMD_STS
) {
3793 DRM_DEBUG("gpu busy, RCS change rejected\n");
3794 return false; /* still busy with another command */
3797 rgvswctl
= (MEMCTL_CMD_CHFREQ
<< MEMCTL_CMD_SHIFT
) |
3798 (val
<< MEMCTL_FREQ_SHIFT
) | MEMCTL_SFCAVM
;
3799 I915_WRITE16(MEMSWCTL
, rgvswctl
);
3800 POSTING_READ16(MEMSWCTL
);
3802 rgvswctl
|= MEMCTL_CMD_STS
;
3803 I915_WRITE16(MEMSWCTL
, rgvswctl
);
3808 static void ironlake_enable_drps(struct drm_device
*dev
)
3810 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3811 u32 rgvmodectl
= I915_READ(MEMMODECTL
);
3812 u8 fmax
, fmin
, fstart
, vstart
;
3814 spin_lock_irq(&mchdev_lock
);
3816 /* Enable temp reporting */
3817 I915_WRITE16(PMMISC
, I915_READ(PMMISC
) | MCPPCE_EN
);
3818 I915_WRITE16(TSC1
, I915_READ(TSC1
) | TSE
);
3820 /* 100ms RC evaluation intervals */
3821 I915_WRITE(RCUPEI
, 100000);
3822 I915_WRITE(RCDNEI
, 100000);
3824 /* Set max/min thresholds to 90ms and 80ms respectively */
3825 I915_WRITE(RCBMAXAVG
, 90000);
3826 I915_WRITE(RCBMINAVG
, 80000);
3828 I915_WRITE(MEMIHYST
, 1);
3830 /* Set up min, max, and cur for interrupt handling */
3831 fmax
= (rgvmodectl
& MEMMODE_FMAX_MASK
) >> MEMMODE_FMAX_SHIFT
;
3832 fmin
= (rgvmodectl
& MEMMODE_FMIN_MASK
);
3833 fstart
= (rgvmodectl
& MEMMODE_FSTART_MASK
) >>
3834 MEMMODE_FSTART_SHIFT
;
3836 vstart
= (I915_READ(PXVFREQ_BASE
+ (fstart
* 4)) & PXVFREQ_PX_MASK
) >>
3839 dev_priv
->ips
.fmax
= fmax
; /* IPS callback will increase this */
3840 dev_priv
->ips
.fstart
= fstart
;
3842 dev_priv
->ips
.max_delay
= fstart
;
3843 dev_priv
->ips
.min_delay
= fmin
;
3844 dev_priv
->ips
.cur_delay
= fstart
;
3846 DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
3847 fmax
, fmin
, fstart
);
3849 I915_WRITE(MEMINTREN
, MEMINT_CX_SUPR_EN
| MEMINT_EVAL_CHG_EN
);
3852 * Interrupts will be enabled in ironlake_irq_postinstall
3855 I915_WRITE(VIDSTART
, vstart
);
3856 POSTING_READ(VIDSTART
);
3858 rgvmodectl
|= MEMMODE_SWMODE_EN
;
3859 I915_WRITE(MEMMODECTL
, rgvmodectl
);
3861 if (wait_for_atomic((I915_READ(MEMSWCTL
) & MEMCTL_CMD_STS
) == 0, 10))
3862 DRM_ERROR("stuck trying to change perf mode\n");
3865 ironlake_set_drps(dev
, fstart
);
3867 dev_priv
->ips
.last_count1
= I915_READ(0x112e4) + I915_READ(0x112e8) +
3869 dev_priv
->ips
.last_time1
= jiffies_to_msecs(jiffies
);
3870 dev_priv
->ips
.last_count2
= I915_READ(0x112f4);
3871 dev_priv
->ips
.last_time2
= ktime_get_raw_ns();
3873 spin_unlock_irq(&mchdev_lock
);
3876 static void ironlake_disable_drps(struct drm_device
*dev
)
3878 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
3881 spin_lock_irq(&mchdev_lock
);
3883 rgvswctl
= I915_READ16(MEMSWCTL
);
3885 /* Ack interrupts, disable EFC interrupt */
3886 I915_WRITE(MEMINTREN
, I915_READ(MEMINTREN
) & ~MEMINT_EVAL_CHG_EN
);
3887 I915_WRITE(MEMINTRSTS
, MEMINT_EVAL_CHG
);
3888 I915_WRITE(DEIER
, I915_READ(DEIER
) & ~DE_PCU_EVENT
);
3889 I915_WRITE(DEIIR
, DE_PCU_EVENT
);
3890 I915_WRITE(DEIMR
, I915_READ(DEIMR
) | DE_PCU_EVENT
);
3892 /* Go back to the starting frequency */
3893 ironlake_set_drps(dev
, dev_priv
->ips
.fstart
);
3895 rgvswctl
|= MEMCTL_CMD_STS
;
3896 I915_WRITE(MEMSWCTL
, rgvswctl
);
3899 spin_unlock_irq(&mchdev_lock
);
3902 /* There's a funny hw issue where the hw returns all 0 when reading from
3903 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
3904 * ourselves, instead of doing a rmw cycle (which might result in us clearing
3905 * all limits and the gpu stuck at whatever frequency it is at atm).
3907 static u32
intel_rps_limits(struct drm_i915_private
*dev_priv
, u8 val
)
3911 /* Only set the down limit when we've reached the lowest level to avoid
3912 * getting more interrupts, otherwise leave this clear. This prevents a
3913 * race in the hw when coming out of rc6: There's a tiny window where
3914 * the hw runs at the minimal clock before selecting the desired
3915 * frequency, if the down threshold expires in that window we will not
3916 * receive a down interrupt. */
3917 if (IS_GEN9(dev_priv
->dev
)) {
3918 limits
= (dev_priv
->rps
.max_freq_softlimit
) << 23;
3919 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
3920 limits
|= (dev_priv
->rps
.min_freq_softlimit
) << 14;
3922 limits
= dev_priv
->rps
.max_freq_softlimit
<< 24;
3923 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
3924 limits
|= dev_priv
->rps
.min_freq_softlimit
<< 16;
3930 static void gen6_set_rps_thresholds(struct drm_i915_private
*dev_priv
, u8 val
)
3933 u32 threshold_up
= 0, threshold_down
= 0; /* in % */
3934 u32 ei_up
= 0, ei_down
= 0;
3936 new_power
= dev_priv
->rps
.power
;
3937 switch (dev_priv
->rps
.power
) {
3939 if (val
> dev_priv
->rps
.efficient_freq
+ 1 && val
> dev_priv
->rps
.cur_freq
)
3940 new_power
= BETWEEN
;
3944 if (val
<= dev_priv
->rps
.efficient_freq
&& val
< dev_priv
->rps
.cur_freq
)
3945 new_power
= LOW_POWER
;
3946 else if (val
>= dev_priv
->rps
.rp0_freq
&& val
> dev_priv
->rps
.cur_freq
)
3947 new_power
= HIGH_POWER
;
3951 if (val
< (dev_priv
->rps
.rp1_freq
+ dev_priv
->rps
.rp0_freq
) >> 1 && val
< dev_priv
->rps
.cur_freq
)
3952 new_power
= BETWEEN
;
3955 /* Max/min bins are special */
3956 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
3957 new_power
= LOW_POWER
;
3958 if (val
>= dev_priv
->rps
.max_freq_softlimit
)
3959 new_power
= HIGH_POWER
;
3960 if (new_power
== dev_priv
->rps
.power
)
3963 /* Note the units here are not exactly 1us, but 1280ns. */
3964 switch (new_power
) {
3966 /* Upclock if more than 95% busy over 16ms */
3970 /* Downclock if less than 85% busy over 32ms */
3972 threshold_down
= 85;
3976 /* Upclock if more than 90% busy over 13ms */
3980 /* Downclock if less than 75% busy over 32ms */
3982 threshold_down
= 75;
3986 /* Upclock if more than 85% busy over 10ms */
3990 /* Downclock if less than 60% busy over 32ms */
3992 threshold_down
= 60;
3996 I915_WRITE(GEN6_RP_UP_EI
,
3997 GT_INTERVAL_FROM_US(dev_priv
, ei_up
));
3998 I915_WRITE(GEN6_RP_UP_THRESHOLD
,
3999 GT_INTERVAL_FROM_US(dev_priv
, (ei_up
* threshold_up
/ 100)));
4001 I915_WRITE(GEN6_RP_DOWN_EI
,
4002 GT_INTERVAL_FROM_US(dev_priv
, ei_down
));
4003 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
,
4004 GT_INTERVAL_FROM_US(dev_priv
, (ei_down
* threshold_down
/ 100)));
4006 I915_WRITE(GEN6_RP_CONTROL
,
4007 GEN6_RP_MEDIA_TURBO
|
4008 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
4009 GEN6_RP_MEDIA_IS_GFX
|
4011 GEN6_RP_UP_BUSY_AVG
|
4012 GEN6_RP_DOWN_IDLE_AVG
);
4014 dev_priv
->rps
.power
= new_power
;
4015 dev_priv
->rps
.up_threshold
= threshold_up
;
4016 dev_priv
->rps
.down_threshold
= threshold_down
;
4017 dev_priv
->rps
.last_adj
= 0;
4020 static u32
gen6_rps_pm_mask(struct drm_i915_private
*dev_priv
, u8 val
)
4024 if (val
> dev_priv
->rps
.min_freq_softlimit
)
4025 mask
|= GEN6_PM_RP_DOWN_EI_EXPIRED
| GEN6_PM_RP_DOWN_THRESHOLD
| GEN6_PM_RP_DOWN_TIMEOUT
;
4026 if (val
< dev_priv
->rps
.max_freq_softlimit
)
4027 mask
|= GEN6_PM_RP_UP_EI_EXPIRED
| GEN6_PM_RP_UP_THRESHOLD
;
4029 mask
&= dev_priv
->pm_rps_events
;
4031 return gen6_sanitize_rps_pm_mask(dev_priv
, ~mask
);
4034 /* gen6_set_rps is called to update the frequency request, but should also be
4035 * called when the range (min_delay and max_delay) is modified so that we can
4036 * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
4037 static void gen6_set_rps(struct drm_device
*dev
, u8 val
)
4039 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4041 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
4042 WARN_ON(val
> dev_priv
->rps
.max_freq
);
4043 WARN_ON(val
< dev_priv
->rps
.min_freq
);
4045 /* min/max delay may still have been modified so be sure to
4046 * write the limits value.
4048 if (val
!= dev_priv
->rps
.cur_freq
) {
4049 gen6_set_rps_thresholds(dev_priv
, val
);
4052 I915_WRITE(GEN6_RPNSWREQ
,
4053 GEN9_FREQUENCY(val
));
4054 else if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
4055 I915_WRITE(GEN6_RPNSWREQ
,
4056 HSW_FREQUENCY(val
));
4058 I915_WRITE(GEN6_RPNSWREQ
,
4059 GEN6_FREQUENCY(val
) |
4061 GEN6_AGGRESSIVE_TURBO
);
4064 /* Make sure we continue to get interrupts
4065 * until we hit the minimum or maximum frequencies.
4067 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS
, intel_rps_limits(dev_priv
, val
));
4068 I915_WRITE(GEN6_PMINTRMSK
, gen6_rps_pm_mask(dev_priv
, val
));
4070 POSTING_READ(GEN6_RPNSWREQ
);
4072 dev_priv
->rps
.cur_freq
= val
;
4073 trace_intel_gpu_freq_change(val
* 50);
4076 static void valleyview_set_rps(struct drm_device
*dev
, u8 val
)
4078 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4080 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
4081 WARN_ON(val
> dev_priv
->rps
.max_freq
);
4082 WARN_ON(val
< dev_priv
->rps
.min_freq
);
4084 if (WARN_ONCE(IS_CHERRYVIEW(dev
) && (val
& 1),
4085 "Odd GPU freq value\n"))
4088 if (val
!= dev_priv
->rps
.cur_freq
) {
4089 vlv_punit_write(dev_priv
, PUNIT_REG_GPU_FREQ_REQ
, val
);
4090 if (!IS_CHERRYVIEW(dev_priv
))
4091 gen6_set_rps_thresholds(dev_priv
, val
);
4094 I915_WRITE(GEN6_PMINTRMSK
, gen6_rps_pm_mask(dev_priv
, val
));
4096 dev_priv
->rps
.cur_freq
= val
;
4097 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv
, val
));
4100 /* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
4102 * * If Gfx is Idle, then
4103 * 1. Forcewake Media well.
4104 * 2. Request idle freq.
4105 * 3. Release Forcewake of Media well.
4107 static void vlv_set_rps_idle(struct drm_i915_private
*dev_priv
)
4109 u32 val
= dev_priv
->rps
.idle_freq
;
4111 if (dev_priv
->rps
.cur_freq
<= val
)
4114 /* Wake up the media well, as that takes a lot less
4115 * power than the Render well. */
4116 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_MEDIA
);
4117 valleyview_set_rps(dev_priv
->dev
, val
);
4118 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_MEDIA
);
4121 void gen6_rps_busy(struct drm_i915_private
*dev_priv
)
4123 mutex_lock(&dev_priv
->rps
.hw_lock
);
4124 if (dev_priv
->rps
.enabled
) {
4125 if (dev_priv
->pm_rps_events
& (GEN6_PM_RP_DOWN_EI_EXPIRED
| GEN6_PM_RP_UP_EI_EXPIRED
))
4126 gen6_rps_reset_ei(dev_priv
);
4127 I915_WRITE(GEN6_PMINTRMSK
,
4128 gen6_rps_pm_mask(dev_priv
, dev_priv
->rps
.cur_freq
));
4130 mutex_unlock(&dev_priv
->rps
.hw_lock
);
4133 void gen6_rps_idle(struct drm_i915_private
*dev_priv
)
4135 struct drm_device
*dev
= dev_priv
->dev
;
4137 mutex_lock(&dev_priv
->rps
.hw_lock
);
4138 if (dev_priv
->rps
.enabled
) {
4139 if (IS_VALLEYVIEW(dev
))
4140 vlv_set_rps_idle(dev_priv
);
4142 gen6_set_rps(dev_priv
->dev
, dev_priv
->rps
.idle_freq
);
4143 dev_priv
->rps
.last_adj
= 0;
4144 I915_WRITE(GEN6_PMINTRMSK
, 0xffffffff);
4147 while (!list_empty(&dev_priv
->rps
.clients
))
4148 list_del_init(dev_priv
->rps
.clients
.next
);
4149 mutex_unlock(&dev_priv
->rps
.hw_lock
);
4152 void gen6_rps_boost(struct drm_i915_private
*dev_priv
,
4153 struct intel_rps_client
*rps
,
4154 unsigned long submitted
)
4158 /* Force a RPS boost (and don't count it against the client) if
4159 * the GPU is severely congested.
4161 if (rps
&& time_after(jiffies
, submitted
+ msecs_to_jiffies(20)))
4164 mutex_lock(&dev_priv
->rps
.hw_lock
);
4165 val
= dev_priv
->rps
.max_freq_softlimit
;
4166 if (dev_priv
->rps
.enabled
&&
4167 dev_priv
->mm
.busy
&&
4168 dev_priv
->rps
.cur_freq
< val
&&
4169 (rps
== NULL
|| list_empty(&rps
->link
))) {
4170 intel_set_rps(dev_priv
->dev
, val
);
4171 dev_priv
->rps
.last_adj
= 0;
4174 list_add(&rps
->link
, &dev_priv
->rps
.clients
);
4177 dev_priv
->rps
.boosts
++;
4179 mutex_unlock(&dev_priv
->rps
.hw_lock
);
4182 void intel_set_rps(struct drm_device
*dev
, u8 val
)
4184 if (IS_VALLEYVIEW(dev
))
4185 valleyview_set_rps(dev
, val
);
4187 gen6_set_rps(dev
, val
);
4190 static void gen9_disable_rps(struct drm_device
*dev
)
4192 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4194 I915_WRITE(GEN6_RC_CONTROL
, 0);
4195 I915_WRITE(GEN9_PG_ENABLE
, 0);
4198 static void gen6_disable_rps(struct drm_device
*dev
)
4200 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4202 I915_WRITE(GEN6_RC_CONTROL
, 0);
4203 I915_WRITE(GEN6_RPNSWREQ
, 1 << 31);
4206 static void cherryview_disable_rps(struct drm_device
*dev
)
4208 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4210 I915_WRITE(GEN6_RC_CONTROL
, 0);
4213 static void valleyview_disable_rps(struct drm_device
*dev
)
4215 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4217 /* we're doing forcewake before Disabling RC6,
4218 * This what the BIOS expects when going into suspend */
4219 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4221 I915_WRITE(GEN6_RC_CONTROL
, 0);
4223 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4226 static void intel_print_rc6_info(struct drm_device
*dev
, u32 mode
)
4228 if (IS_VALLEYVIEW(dev
)) {
4229 if (mode
& (GEN7_RC_CTL_TO_MODE
| GEN6_RC_CTL_EI_MODE(1)))
4230 mode
= GEN6_RC_CTL_RC6_ENABLE
;
4235 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s RC6p %s RC6pp %s\n",
4236 (mode
& GEN6_RC_CTL_RC6_ENABLE
) ? "on" : "off",
4237 (mode
& GEN6_RC_CTL_RC6p_ENABLE
) ? "on" : "off",
4238 (mode
& GEN6_RC_CTL_RC6pp_ENABLE
) ? "on" : "off");
4241 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s\n",
4242 (mode
& GEN6_RC_CTL_RC6_ENABLE
) ? "on" : "off");
4245 static int sanitize_rc6_option(const struct drm_device
*dev
, int enable_rc6
)
4247 /* No RC6 before Ironlake */
4248 if (INTEL_INFO(dev
)->gen
< 5)
4251 /* RC6 is only on Ironlake mobile not on desktop */
4252 if (INTEL_INFO(dev
)->gen
== 5 && !IS_IRONLAKE_M(dev
))
4255 /* Respect the kernel parameter if it is set */
4256 if (enable_rc6
>= 0) {
4260 mask
= INTEL_RC6_ENABLE
| INTEL_RC6p_ENABLE
|
4263 mask
= INTEL_RC6_ENABLE
;
4265 if ((enable_rc6
& mask
) != enable_rc6
)
4266 DRM_DEBUG_KMS("Adjusting RC6 mask to %d (requested %d, valid %d)\n",
4267 enable_rc6
& mask
, enable_rc6
, mask
);
4269 return enable_rc6
& mask
;
4272 /* Disable RC6 on Ironlake */
4273 if (INTEL_INFO(dev
)->gen
== 5)
4276 if (IS_IVYBRIDGE(dev
))
4277 return (INTEL_RC6_ENABLE
| INTEL_RC6p_ENABLE
);
4279 return INTEL_RC6_ENABLE
;
4282 int intel_enable_rc6(const struct drm_device
*dev
)
4284 return i915
.enable_rc6
;
4287 static void gen6_init_rps_frequencies(struct drm_device
*dev
)
4289 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4290 uint32_t rp_state_cap
;
4291 u32 ddcc_status
= 0;
4294 rp_state_cap
= I915_READ(GEN6_RP_STATE_CAP
);
4295 /* All of these values are in units of 50MHz */
4296 dev_priv
->rps
.cur_freq
= 0;
4297 /* static values from HW: RP0 > RP1 > RPn (min_freq) */
4298 dev_priv
->rps
.rp0_freq
= (rp_state_cap
>> 0) & 0xff;
4299 dev_priv
->rps
.rp1_freq
= (rp_state_cap
>> 8) & 0xff;
4300 dev_priv
->rps
.min_freq
= (rp_state_cap
>> 16) & 0xff;
4301 if (IS_SKYLAKE(dev
)) {
4302 /* Store the frequency values in 16.66 MHZ units, which is
4303 the natural hardware unit for SKL */
4304 dev_priv
->rps
.rp0_freq
*= GEN9_FREQ_SCALER
;
4305 dev_priv
->rps
.rp1_freq
*= GEN9_FREQ_SCALER
;
4306 dev_priv
->rps
.min_freq
*= GEN9_FREQ_SCALER
;
4308 /* hw_max = RP0 until we check for overclocking */
4309 dev_priv
->rps
.max_freq
= dev_priv
->rps
.rp0_freq
;
4311 dev_priv
->rps
.efficient_freq
= dev_priv
->rps
.rp1_freq
;
4312 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
)) {
4313 ret
= sandybridge_pcode_read(dev_priv
,
4314 HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL
,
4317 dev_priv
->rps
.efficient_freq
=
4319 ((ddcc_status
>> 8) & 0xff),
4320 dev_priv
->rps
.min_freq
,
4321 dev_priv
->rps
.max_freq
);
4324 dev_priv
->rps
.idle_freq
= dev_priv
->rps
.min_freq
;
4326 /* Preserve min/max settings in case of re-init */
4327 if (dev_priv
->rps
.max_freq_softlimit
== 0)
4328 dev_priv
->rps
.max_freq_softlimit
= dev_priv
->rps
.max_freq
;
4330 if (dev_priv
->rps
.min_freq_softlimit
== 0) {
4331 if (IS_HASWELL(dev
) || IS_BROADWELL(dev
))
4332 dev_priv
->rps
.min_freq_softlimit
=
4333 max_t(int, dev_priv
->rps
.efficient_freq
,
4334 intel_freq_opcode(dev_priv
, 450));
4336 dev_priv
->rps
.min_freq_softlimit
=
4337 dev_priv
->rps
.min_freq
;
4341 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
4342 static void gen9_enable_rps(struct drm_device
*dev
)
4344 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4346 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4348 gen6_init_rps_frequencies(dev
);
4350 /* Program defaults and thresholds for RPS*/
4351 I915_WRITE(GEN6_RC_VIDEO_FREQ
,
4352 GEN9_FREQUENCY(dev_priv
->rps
.rp1_freq
));
4354 /* 1 second timeout*/
4355 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
,
4356 GT_INTERVAL_FROM_US(dev_priv
, 1000000));
4358 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 0xa);
4360 /* Leaning on the below call to gen6_set_rps to program/setup the
4361 * Up/Down EI & threshold registers, as well as the RP_CONTROL,
4362 * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
4363 dev_priv
->rps
.power
= HIGH_POWER
; /* force a reset */
4364 gen6_set_rps(dev_priv
->dev
, dev_priv
->rps
.min_freq_softlimit
);
4366 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4369 static void gen9_enable_rc6(struct drm_device
*dev
)
4371 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4372 struct intel_engine_cs
*ring
;
4373 uint32_t rc6_mask
= 0;
4376 /* 1a: Software RC state - RC0 */
4377 I915_WRITE(GEN6_RC_STATE
, 0);
4379 /* 1b: Get forcewake during program sequence. Although the driver
4380 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4381 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4383 /* 2a: Disable RC states. */
4384 I915_WRITE(GEN6_RC_CONTROL
, 0);
4386 /* 2b: Program RC6 thresholds.*/
4387 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 54 << 16);
4388 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
4389 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
4390 for_each_ring(ring
, dev_priv
, unused
)
4391 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
4392 I915_WRITE(GEN6_RC_SLEEP
, 0);
4393 I915_WRITE(GEN6_RC6_THRESHOLD
, 37500); /* 37.5/125ms per EI */
4395 /* 2c: Program Coarse Power Gating Policies. */
4396 I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS
, 25);
4397 I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS
, 25);
4399 /* 3a: Enable RC6 */
4400 if (intel_enable_rc6(dev
) & INTEL_RC6_ENABLE
)
4401 rc6_mask
= GEN6_RC_CTL_RC6_ENABLE
;
4402 DRM_INFO("RC6 %s\n", (rc6_mask
& GEN6_RC_CTL_RC6_ENABLE
) ?
4404 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
4405 GEN6_RC_CTL_EI_MODE(1) |
4409 * 3b: Enable Coarse Power Gating only when RC6 is enabled.
4410 * WaDisableRenderPowerGating:skl,bxt - Render PG need to be disabled with RC6.
4412 I915_WRITE(GEN9_PG_ENABLE
, (rc6_mask
& GEN6_RC_CTL_RC6_ENABLE
) ?
4413 GEN9_MEDIA_PG_ENABLE
: 0);
4416 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4420 static void gen8_enable_rps(struct drm_device
*dev
)
4422 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4423 struct intel_engine_cs
*ring
;
4424 uint32_t rc6_mask
= 0;
4427 /* 1a: Software RC state - RC0 */
4428 I915_WRITE(GEN6_RC_STATE
, 0);
4430 /* 1c & 1d: Get forcewake during program sequence. Although the driver
4431 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4432 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4434 /* 2a: Disable RC states. */
4435 I915_WRITE(GEN6_RC_CONTROL
, 0);
4437 /* Initialize rps frequencies */
4438 gen6_init_rps_frequencies(dev
);
4440 /* 2b: Program RC6 thresholds.*/
4441 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16);
4442 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
4443 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
4444 for_each_ring(ring
, dev_priv
, unused
)
4445 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
4446 I915_WRITE(GEN6_RC_SLEEP
, 0);
4447 if (IS_BROADWELL(dev
))
4448 I915_WRITE(GEN6_RC6_THRESHOLD
, 625); /* 800us/1.28 for TO */
4450 I915_WRITE(GEN6_RC6_THRESHOLD
, 50000); /* 50/125ms per EI */
4453 if (intel_enable_rc6(dev
) & INTEL_RC6_ENABLE
)
4454 rc6_mask
= GEN6_RC_CTL_RC6_ENABLE
;
4455 intel_print_rc6_info(dev
, rc6_mask
);
4456 if (IS_BROADWELL(dev
))
4457 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
4458 GEN7_RC_CTL_TO_MODE
|
4461 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
4462 GEN6_RC_CTL_EI_MODE(1) |
4465 /* 4 Program defaults and thresholds for RPS*/
4466 I915_WRITE(GEN6_RPNSWREQ
,
4467 HSW_FREQUENCY(dev_priv
->rps
.rp1_freq
));
4468 I915_WRITE(GEN6_RC_VIDEO_FREQ
,
4469 HSW_FREQUENCY(dev_priv
->rps
.rp1_freq
));
4470 /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
4471 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 100000000 / 128); /* 1 second timeout */
4473 /* Docs recommend 900MHz, and 300 MHz respectively */
4474 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS
,
4475 dev_priv
->rps
.max_freq_softlimit
<< 24 |
4476 dev_priv
->rps
.min_freq_softlimit
<< 16);
4478 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 7600000 / 128); /* 76ms busyness per EI, 90% */
4479 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 31300000 / 128); /* 313ms busyness per EI, 70%*/
4480 I915_WRITE(GEN6_RP_UP_EI
, 66000); /* 84.48ms, XXX: random? */
4481 I915_WRITE(GEN6_RP_DOWN_EI
, 350000); /* 448ms, XXX: random? */
4483 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
4486 I915_WRITE(GEN6_RP_CONTROL
,
4487 GEN6_RP_MEDIA_TURBO
|
4488 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
4489 GEN6_RP_MEDIA_IS_GFX
|
4491 GEN6_RP_UP_BUSY_AVG
|
4492 GEN6_RP_DOWN_IDLE_AVG
);
4494 /* 6: Ring frequency + overclocking (our driver does this later */
4496 dev_priv
->rps
.power
= HIGH_POWER
; /* force a reset */
4497 gen6_set_rps(dev_priv
->dev
, dev_priv
->rps
.idle_freq
);
4499 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4502 static void gen6_enable_rps(struct drm_device
*dev
)
4504 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4505 struct intel_engine_cs
*ring
;
4506 u32 rc6vids
, pcu_mbox
= 0, rc6_mask
= 0;
4511 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
4513 /* Here begins a magic sequence of register writes to enable
4514 * auto-downclocking.
4516 * Perhaps there might be some value in exposing these to
4519 I915_WRITE(GEN6_RC_STATE
, 0);
4521 /* Clear the DBG now so we don't confuse earlier errors */
4522 if ((gtfifodbg
= I915_READ(GTFIFODBG
))) {
4523 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg
);
4524 I915_WRITE(GTFIFODBG
, gtfifodbg
);
4527 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
4529 /* Initialize rps frequencies */
4530 gen6_init_rps_frequencies(dev
);
4532 /* disable the counters and set deterministic thresholds */
4533 I915_WRITE(GEN6_RC_CONTROL
, 0);
4535 I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT
, 1000 << 16);
4536 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16 | 30);
4537 I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT
, 30);
4538 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000);
4539 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25);
4541 for_each_ring(ring
, dev_priv
, i
)
4542 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
4544 I915_WRITE(GEN6_RC_SLEEP
, 0);
4545 I915_WRITE(GEN6_RC1e_THRESHOLD
, 1000);
4546 if (IS_IVYBRIDGE(dev
))
4547 I915_WRITE(GEN6_RC6_THRESHOLD
, 125000);
4549 I915_WRITE(GEN6_RC6_THRESHOLD
, 50000);
4550 I915_WRITE(GEN6_RC6p_THRESHOLD
, 150000);
4551 I915_WRITE(GEN6_RC6pp_THRESHOLD
, 64000); /* unused */
4553 /* Check if we are enabling RC6 */
4554 rc6_mode
= intel_enable_rc6(dev_priv
->dev
);
4555 if (rc6_mode
& INTEL_RC6_ENABLE
)
4556 rc6_mask
|= GEN6_RC_CTL_RC6_ENABLE
;
4558 /* We don't use those on Haswell */
4559 if (!IS_HASWELL(dev
)) {
4560 if (rc6_mode
& INTEL_RC6p_ENABLE
)
4561 rc6_mask
|= GEN6_RC_CTL_RC6p_ENABLE
;
4563 if (rc6_mode
& INTEL_RC6pp_ENABLE
)
4564 rc6_mask
|= GEN6_RC_CTL_RC6pp_ENABLE
;
4567 intel_print_rc6_info(dev
, rc6_mask
);
4569 I915_WRITE(GEN6_RC_CONTROL
,
4571 GEN6_RC_CTL_EI_MODE(1) |
4572 GEN6_RC_CTL_HW_ENABLE
);
4574 /* Power down if completely idle for over 50ms */
4575 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 50000);
4576 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
4578 ret
= sandybridge_pcode_write(dev_priv
, GEN6_PCODE_WRITE_MIN_FREQ_TABLE
, 0);
4580 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
4582 ret
= sandybridge_pcode_read(dev_priv
, GEN6_READ_OC_PARAMS
, &pcu_mbox
);
4583 if (!ret
&& (pcu_mbox
& (1<<31))) { /* OC supported */
4584 DRM_DEBUG_DRIVER("Overclocking supported. Max: %dMHz, Overclock max: %dMHz\n",
4585 (dev_priv
->rps
.max_freq_softlimit
& 0xff) * 50,
4586 (pcu_mbox
& 0xff) * 50);
4587 dev_priv
->rps
.max_freq
= pcu_mbox
& 0xff;
4590 dev_priv
->rps
.power
= HIGH_POWER
; /* force a reset */
4591 gen6_set_rps(dev_priv
->dev
, dev_priv
->rps
.idle_freq
);
4594 ret
= sandybridge_pcode_read(dev_priv
, GEN6_PCODE_READ_RC6VIDS
, &rc6vids
);
4595 if (IS_GEN6(dev
) && ret
) {
4596 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
4597 } else if (IS_GEN6(dev
) && (GEN6_DECODE_RC6_VID(rc6vids
& 0xff) < 450)) {
4598 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
4599 GEN6_DECODE_RC6_VID(rc6vids
& 0xff), 450);
4600 rc6vids
&= 0xffff00;
4601 rc6vids
|= GEN6_ENCODE_RC6_VID(450);
4602 ret
= sandybridge_pcode_write(dev_priv
, GEN6_PCODE_WRITE_RC6VIDS
, rc6vids
);
4604 DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
4607 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
4610 static void __gen6_update_ring_freq(struct drm_device
*dev
)
4612 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4614 unsigned int gpu_freq
;
4615 unsigned int max_ia_freq
, min_ring_freq
;
4616 int scaling_factor
= 180;
4617 struct cpufreq_policy
*policy
;
4619 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
4621 policy
= cpufreq_cpu_get(0);
4623 max_ia_freq
= policy
->cpuinfo
.max_freq
;
4624 cpufreq_cpu_put(policy
);
4627 * Default to measured freq if none found, PCU will ensure we
4630 max_ia_freq
= tsc_khz
;
4633 /* Convert from kHz to MHz */
4634 max_ia_freq
/= 1000;
4636 min_ring_freq
= I915_READ(DCLK
) & 0xf;
4637 /* convert DDR frequency from units of 266.6MHz to bandwidth */
4638 min_ring_freq
= mult_frac(min_ring_freq
, 8, 3);
4641 * For each potential GPU frequency, load a ring frequency we'd like
4642 * to use for memory access. We do this by specifying the IA frequency
4643 * the PCU should use as a reference to determine the ring frequency.
4645 for (gpu_freq
= dev_priv
->rps
.max_freq
; gpu_freq
>= dev_priv
->rps
.min_freq
;
4647 int diff
= dev_priv
->rps
.max_freq
- gpu_freq
;
4648 unsigned int ia_freq
= 0, ring_freq
= 0;
4650 if (INTEL_INFO(dev
)->gen
>= 8) {
4651 /* max(2 * GT, DDR). NB: GT is 50MHz units */
4652 ring_freq
= max(min_ring_freq
, gpu_freq
);
4653 } else if (IS_HASWELL(dev
)) {
4654 ring_freq
= mult_frac(gpu_freq
, 5, 4);
4655 ring_freq
= max(min_ring_freq
, ring_freq
);
4656 /* leave ia_freq as the default, chosen by cpufreq */
4658 /* On older processors, there is no separate ring
4659 * clock domain, so in order to boost the bandwidth
4660 * of the ring, we need to upclock the CPU (ia_freq).
4662 * For GPU frequencies less than 750MHz,
4663 * just use the lowest ring freq.
4665 if (gpu_freq
< min_freq
)
4668 ia_freq
= max_ia_freq
- ((diff
* scaling_factor
) / 2);
4669 ia_freq
= DIV_ROUND_CLOSEST(ia_freq
, 100);
4672 sandybridge_pcode_write(dev_priv
,
4673 GEN6_PCODE_WRITE_MIN_FREQ_TABLE
,
4674 ia_freq
<< GEN6_PCODE_FREQ_IA_RATIO_SHIFT
|
4675 ring_freq
<< GEN6_PCODE_FREQ_RING_RATIO_SHIFT
|
4680 void gen6_update_ring_freq(struct drm_device
*dev
)
4682 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4684 if (INTEL_INFO(dev
)->gen
< 6 || IS_VALLEYVIEW(dev
))
4687 mutex_lock(&dev_priv
->rps
.hw_lock
);
4688 __gen6_update_ring_freq(dev
);
4689 mutex_unlock(&dev_priv
->rps
.hw_lock
);
4692 static int cherryview_rps_max_freq(struct drm_i915_private
*dev_priv
)
4694 struct drm_device
*dev
= dev_priv
->dev
;
4697 if (dev
->pdev
->revision
>= 0x20) {
4698 val
= vlv_punit_read(dev_priv
, FB_GFX_FMAX_AT_VMAX_FUSE
);
4700 switch (INTEL_INFO(dev
)->eu_total
) {
4702 /* (2 * 4) config */
4703 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT
);
4706 /* (2 * 6) config */
4707 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT
);
4710 /* (2 * 8) config */
4712 /* Setting (2 * 8) Min RP0 for any other combination */
4713 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT
);
4716 rp0
= (rp0
& FB_GFX_FREQ_FUSE_MASK
);
4718 /* For pre-production hardware */
4719 val
= vlv_punit_read(dev_priv
, PUNIT_GPU_STATUS_REG
);
4720 rp0
= (val
>> PUNIT_GPU_STATUS_MAX_FREQ_SHIFT
) &
4721 PUNIT_GPU_STATUS_MAX_FREQ_MASK
;
4726 static int cherryview_rps_rpe_freq(struct drm_i915_private
*dev_priv
)
4730 val
= vlv_punit_read(dev_priv
, PUNIT_GPU_DUTYCYCLE_REG
);
4731 rpe
= (val
>> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT
) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK
;
4736 static int cherryview_rps_guar_freq(struct drm_i915_private
*dev_priv
)
4738 struct drm_device
*dev
= dev_priv
->dev
;
4741 if (dev
->pdev
->revision
>= 0x20) {
4742 val
= vlv_punit_read(dev_priv
, FB_GFX_FMAX_AT_VMAX_FUSE
);
4743 rp1
= (val
& FB_GFX_FREQ_FUSE_MASK
);
4745 /* For pre-production hardware */
4746 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
4747 rp1
= ((val
>> PUNIT_GPU_STATUS_MAX_FREQ_SHIFT
) &
4748 PUNIT_GPU_STATUS_MAX_FREQ_MASK
);
4753 static int valleyview_rps_guar_freq(struct drm_i915_private
*dev_priv
)
4757 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FREQ_FUSE
);
4759 rp1
= (val
& FB_GFX_FGUARANTEED_FREQ_FUSE_MASK
) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT
;
4764 static int valleyview_rps_max_freq(struct drm_i915_private
*dev_priv
)
4768 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FREQ_FUSE
);
4770 rp0
= (val
& FB_GFX_MAX_FREQ_FUSE_MASK
) >> FB_GFX_MAX_FREQ_FUSE_SHIFT
;
4772 rp0
= min_t(u32
, rp0
, 0xea);
4777 static int valleyview_rps_rpe_freq(struct drm_i915_private
*dev_priv
)
4781 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FMAX_FUSE_LO
);
4782 rpe
= (val
& FB_FMAX_VMIN_FREQ_LO_MASK
) >> FB_FMAX_VMIN_FREQ_LO_SHIFT
;
4783 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FMAX_FUSE_HI
);
4784 rpe
|= (val
& FB_FMAX_VMIN_FREQ_HI_MASK
) << 5;
4789 static int valleyview_rps_min_freq(struct drm_i915_private
*dev_priv
)
4791 return vlv_punit_read(dev_priv
, PUNIT_REG_GPU_LFM
) & 0xff;
4794 /* Check that the pctx buffer wasn't move under us. */
4795 static void valleyview_check_pctx(struct drm_i915_private
*dev_priv
)
4797 unsigned long pctx_addr
= I915_READ(VLV_PCBR
) & ~4095;
4799 WARN_ON(pctx_addr
!= dev_priv
->mm
.stolen_base
+
4800 dev_priv
->vlv_pctx
->stolen
->start
);
4804 /* Check that the pcbr address is not empty. */
4805 static void cherryview_check_pctx(struct drm_i915_private
*dev_priv
)
4807 unsigned long pctx_addr
= I915_READ(VLV_PCBR
) & ~4095;
4809 WARN_ON((pctx_addr
>> VLV_PCBR_ADDR_SHIFT
) == 0);
4812 static void cherryview_setup_pctx(struct drm_device
*dev
)
4814 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4815 unsigned long pctx_paddr
, paddr
;
4816 struct i915_gtt
*gtt
= &dev_priv
->gtt
;
4818 int pctx_size
= 32*1024;
4820 WARN_ON(!mutex_is_locked(&dev
->struct_mutex
));
4822 pcbr
= I915_READ(VLV_PCBR
);
4823 if ((pcbr
>> VLV_PCBR_ADDR_SHIFT
) == 0) {
4824 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
4825 paddr
= (dev_priv
->mm
.stolen_base
+
4826 (gtt
->stolen_size
- pctx_size
));
4828 pctx_paddr
= (paddr
& (~4095));
4829 I915_WRITE(VLV_PCBR
, pctx_paddr
);
4832 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR
));
4835 static void valleyview_setup_pctx(struct drm_device
*dev
)
4837 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4838 struct drm_i915_gem_object
*pctx
;
4839 unsigned long pctx_paddr
;
4841 int pctx_size
= 24*1024;
4843 WARN_ON(!mutex_is_locked(&dev
->struct_mutex
));
4845 pcbr
= I915_READ(VLV_PCBR
);
4847 /* BIOS set it up already, grab the pre-alloc'd space */
4850 pcbr_offset
= (pcbr
& (~4095)) - dev_priv
->mm
.stolen_base
;
4851 pctx
= i915_gem_object_create_stolen_for_preallocated(dev_priv
->dev
,
4853 I915_GTT_OFFSET_NONE
,
4858 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
4861 * From the Gunit register HAS:
4862 * The Gfx driver is expected to program this register and ensure
4863 * proper allocation within Gfx stolen memory. For example, this
4864 * register should be programmed such than the PCBR range does not
4865 * overlap with other ranges, such as the frame buffer, protected
4866 * memory, or any other relevant ranges.
4868 pctx
= i915_gem_object_create_stolen(dev
, pctx_size
);
4870 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
4874 pctx_paddr
= dev_priv
->mm
.stolen_base
+ pctx
->stolen
->start
;
4875 I915_WRITE(VLV_PCBR
, pctx_paddr
);
4878 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR
));
4879 dev_priv
->vlv_pctx
= pctx
;
4882 static void valleyview_cleanup_pctx(struct drm_device
*dev
)
4884 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4886 if (WARN_ON(!dev_priv
->vlv_pctx
))
4889 drm_gem_object_unreference(&dev_priv
->vlv_pctx
->base
);
4890 dev_priv
->vlv_pctx
= NULL
;
4893 static void valleyview_init_gt_powersave(struct drm_device
*dev
)
4895 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4898 valleyview_setup_pctx(dev
);
4900 mutex_lock(&dev_priv
->rps
.hw_lock
);
4902 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
4903 switch ((val
>> 6) & 3) {
4906 dev_priv
->mem_freq
= 800;
4909 dev_priv
->mem_freq
= 1066;
4912 dev_priv
->mem_freq
= 1333;
4915 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv
->mem_freq
);
4917 dev_priv
->rps
.max_freq
= valleyview_rps_max_freq(dev_priv
);
4918 dev_priv
->rps
.rp0_freq
= dev_priv
->rps
.max_freq
;
4919 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
4920 intel_gpu_freq(dev_priv
, dev_priv
->rps
.max_freq
),
4921 dev_priv
->rps
.max_freq
);
4923 dev_priv
->rps
.efficient_freq
= valleyview_rps_rpe_freq(dev_priv
);
4924 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
4925 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
4926 dev_priv
->rps
.efficient_freq
);
4928 dev_priv
->rps
.rp1_freq
= valleyview_rps_guar_freq(dev_priv
);
4929 DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
4930 intel_gpu_freq(dev_priv
, dev_priv
->rps
.rp1_freq
),
4931 dev_priv
->rps
.rp1_freq
);
4933 dev_priv
->rps
.min_freq
= valleyview_rps_min_freq(dev_priv
);
4934 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
4935 intel_gpu_freq(dev_priv
, dev_priv
->rps
.min_freq
),
4936 dev_priv
->rps
.min_freq
);
4938 dev_priv
->rps
.idle_freq
= dev_priv
->rps
.min_freq
;
4940 /* Preserve min/max settings in case of re-init */
4941 if (dev_priv
->rps
.max_freq_softlimit
== 0)
4942 dev_priv
->rps
.max_freq_softlimit
= dev_priv
->rps
.max_freq
;
4944 if (dev_priv
->rps
.min_freq_softlimit
== 0)
4945 dev_priv
->rps
.min_freq_softlimit
= dev_priv
->rps
.min_freq
;
4947 mutex_unlock(&dev_priv
->rps
.hw_lock
);
4950 static void cherryview_init_gt_powersave(struct drm_device
*dev
)
4952 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
4955 cherryview_setup_pctx(dev
);
4957 mutex_lock(&dev_priv
->rps
.hw_lock
);
4959 mutex_lock(&dev_priv
->dpio_lock
);
4960 val
= vlv_cck_read(dev_priv
, CCK_FUSE_REG
);
4961 mutex_unlock(&dev_priv
->dpio_lock
);
4963 switch ((val
>> 2) & 0x7) {
4966 dev_priv
->rps
.cz_freq
= 200;
4967 dev_priv
->mem_freq
= 1600;
4970 dev_priv
->rps
.cz_freq
= 267;
4971 dev_priv
->mem_freq
= 1600;
4974 dev_priv
->rps
.cz_freq
= 333;
4975 dev_priv
->mem_freq
= 2000;
4978 dev_priv
->rps
.cz_freq
= 320;
4979 dev_priv
->mem_freq
= 1600;
4982 dev_priv
->rps
.cz_freq
= 400;
4983 dev_priv
->mem_freq
= 1600;
4986 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv
->mem_freq
);
4988 dev_priv
->rps
.max_freq
= cherryview_rps_max_freq(dev_priv
);
4989 dev_priv
->rps
.rp0_freq
= dev_priv
->rps
.max_freq
;
4990 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
4991 intel_gpu_freq(dev_priv
, dev_priv
->rps
.max_freq
),
4992 dev_priv
->rps
.max_freq
);
4994 dev_priv
->rps
.efficient_freq
= cherryview_rps_rpe_freq(dev_priv
);
4995 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
4996 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
4997 dev_priv
->rps
.efficient_freq
);
4999 dev_priv
->rps
.rp1_freq
= cherryview_rps_guar_freq(dev_priv
);
5000 DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
5001 intel_gpu_freq(dev_priv
, dev_priv
->rps
.rp1_freq
),
5002 dev_priv
->rps
.rp1_freq
);
5004 /* PUnit validated range is only [RPe, RP0] */
5005 dev_priv
->rps
.min_freq
= dev_priv
->rps
.efficient_freq
;
5006 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5007 intel_gpu_freq(dev_priv
, dev_priv
->rps
.min_freq
),
5008 dev_priv
->rps
.min_freq
);
5010 WARN_ONCE((dev_priv
->rps
.max_freq
|
5011 dev_priv
->rps
.efficient_freq
|
5012 dev_priv
->rps
.rp1_freq
|
5013 dev_priv
->rps
.min_freq
) & 1,
5014 "Odd GPU freq values\n");
5016 dev_priv
->rps
.idle_freq
= dev_priv
->rps
.min_freq
;
5018 /* Preserve min/max settings in case of re-init */
5019 if (dev_priv
->rps
.max_freq_softlimit
== 0)
5020 dev_priv
->rps
.max_freq_softlimit
= dev_priv
->rps
.max_freq
;
5022 if (dev_priv
->rps
.min_freq_softlimit
== 0)
5023 dev_priv
->rps
.min_freq_softlimit
= dev_priv
->rps
.min_freq
;
5025 mutex_unlock(&dev_priv
->rps
.hw_lock
);
5028 static void valleyview_cleanup_gt_powersave(struct drm_device
*dev
)
5030 valleyview_cleanup_pctx(dev
);
5033 static void cherryview_enable_rps(struct drm_device
*dev
)
5035 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5036 struct intel_engine_cs
*ring
;
5037 u32 gtfifodbg
, val
, rc6_mode
= 0, pcbr
;
5040 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
5042 gtfifodbg
= I915_READ(GTFIFODBG
);
5044 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5046 I915_WRITE(GTFIFODBG
, gtfifodbg
);
5049 cherryview_check_pctx(dev_priv
);
5051 /* 1a & 1b: Get forcewake during program sequence. Although the driver
5052 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5053 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
5055 /* Disable RC states. */
5056 I915_WRITE(GEN6_RC_CONTROL
, 0);
5058 /* 2a: Program RC6 thresholds.*/
5059 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16);
5060 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
5061 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
5063 for_each_ring(ring
, dev_priv
, i
)
5064 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
5065 I915_WRITE(GEN6_RC_SLEEP
, 0);
5067 /* TO threshold set to 500 us ( 0x186 * 1.28 us) */
5068 I915_WRITE(GEN6_RC6_THRESHOLD
, 0x186);
5070 /* allows RC6 residency counter to work */
5071 I915_WRITE(VLV_COUNTER_CONTROL
,
5072 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH
|
5073 VLV_MEDIA_RC6_COUNT_EN
|
5074 VLV_RENDER_RC6_COUNT_EN
));
5076 /* For now we assume BIOS is allocating and populating the PCBR */
5077 pcbr
= I915_READ(VLV_PCBR
);
5080 if ((intel_enable_rc6(dev
) & INTEL_RC6_ENABLE
) &&
5081 (pcbr
>> VLV_PCBR_ADDR_SHIFT
))
5082 rc6_mode
= GEN7_RC_CTL_TO_MODE
;
5084 I915_WRITE(GEN6_RC_CONTROL
, rc6_mode
);
5086 /* 4 Program defaults and thresholds for RPS*/
5087 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 1000000);
5088 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 59400);
5089 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 245000);
5090 I915_WRITE(GEN6_RP_UP_EI
, 66000);
5091 I915_WRITE(GEN6_RP_DOWN_EI
, 350000);
5093 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
5096 I915_WRITE(GEN6_RP_CONTROL
,
5097 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
5098 GEN6_RP_MEDIA_IS_GFX
|
5100 GEN6_RP_UP_BUSY_AVG
|
5101 GEN6_RP_DOWN_IDLE_AVG
);
5103 /* Setting Fixed Bias */
5104 val
= VLV_OVERRIDE_EN
|
5106 CHV_BIAS_CPU_50_SOC_50
;
5107 vlv_punit_write(dev_priv
, VLV_TURBO_SOC_OVERRIDE
, val
);
5109 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
5111 /* RPS code assumes GPLL is used */
5112 WARN_ONCE((val
& GPLLENABLE
) == 0, "GPLL not enabled\n");
5114 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", val
& GPLLENABLE
? "yes" : "no");
5115 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val
);
5117 dev_priv
->rps
.cur_freq
= (val
>> 8) & 0xff;
5118 DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
5119 intel_gpu_freq(dev_priv
, dev_priv
->rps
.cur_freq
),
5120 dev_priv
->rps
.cur_freq
);
5122 DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
5123 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
5124 dev_priv
->rps
.efficient_freq
);
5126 valleyview_set_rps(dev_priv
->dev
, dev_priv
->rps
.efficient_freq
);
5128 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
5131 static void valleyview_enable_rps(struct drm_device
*dev
)
5133 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5134 struct intel_engine_cs
*ring
;
5135 u32 gtfifodbg
, val
, rc6_mode
= 0;
5138 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
5140 valleyview_check_pctx(dev_priv
);
5142 if ((gtfifodbg
= I915_READ(GTFIFODBG
))) {
5143 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5145 I915_WRITE(GTFIFODBG
, gtfifodbg
);
5148 /* If VLV, Forcewake all wells, else re-direct to regular path */
5149 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
5151 /* Disable RC states. */
5152 I915_WRITE(GEN6_RC_CONTROL
, 0);
5154 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 1000000);
5155 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 59400);
5156 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 245000);
5157 I915_WRITE(GEN6_RP_UP_EI
, 66000);
5158 I915_WRITE(GEN6_RP_DOWN_EI
, 350000);
5160 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
5162 I915_WRITE(GEN6_RP_CONTROL
,
5163 GEN6_RP_MEDIA_TURBO
|
5164 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
5165 GEN6_RP_MEDIA_IS_GFX
|
5167 GEN6_RP_UP_BUSY_AVG
|
5168 GEN6_RP_DOWN_IDLE_CONT
);
5170 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 0x00280000);
5171 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000);
5172 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25);
5174 for_each_ring(ring
, dev_priv
, i
)
5175 I915_WRITE(RING_MAX_IDLE(ring
->mmio_base
), 10);
5177 I915_WRITE(GEN6_RC6_THRESHOLD
, 0x557);
5179 /* allows RC6 residency counter to work */
5180 I915_WRITE(VLV_COUNTER_CONTROL
,
5181 _MASKED_BIT_ENABLE(VLV_MEDIA_RC0_COUNT_EN
|
5182 VLV_RENDER_RC0_COUNT_EN
|
5183 VLV_MEDIA_RC6_COUNT_EN
|
5184 VLV_RENDER_RC6_COUNT_EN
));
5186 if (intel_enable_rc6(dev
) & INTEL_RC6_ENABLE
)
5187 rc6_mode
= GEN7_RC_CTL_TO_MODE
| VLV_RC_CTL_CTX_RST_PARALLEL
;
5189 intel_print_rc6_info(dev
, rc6_mode
);
5191 I915_WRITE(GEN6_RC_CONTROL
, rc6_mode
);
5193 /* Setting Fixed Bias */
5194 val
= VLV_OVERRIDE_EN
|
5196 VLV_BIAS_CPU_125_SOC_875
;
5197 vlv_punit_write(dev_priv
, VLV_TURBO_SOC_OVERRIDE
, val
);
5199 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
5201 /* RPS code assumes GPLL is used */
5202 WARN_ONCE((val
& GPLLENABLE
) == 0, "GPLL not enabled\n");
5204 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", val
& GPLLENABLE
? "yes" : "no");
5205 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val
);
5207 dev_priv
->rps
.cur_freq
= (val
>> 8) & 0xff;
5208 DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
5209 intel_gpu_freq(dev_priv
, dev_priv
->rps
.cur_freq
),
5210 dev_priv
->rps
.cur_freq
);
5212 DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
5213 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
5214 dev_priv
->rps
.efficient_freq
);
5216 valleyview_set_rps(dev_priv
->dev
, dev_priv
->rps
.efficient_freq
);
5218 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
5221 static unsigned long intel_pxfreq(u32 vidfreq
)
5224 int div
= (vidfreq
& 0x3f0000) >> 16;
5225 int post
= (vidfreq
& 0x3000) >> 12;
5226 int pre
= (vidfreq
& 0x7);
5231 freq
= ((div
* 133333) / ((1<<post
) * pre
));
5236 static const struct cparams
{
5242 { 1, 1333, 301, 28664 },
5243 { 1, 1066, 294, 24460 },
5244 { 1, 800, 294, 25192 },
5245 { 0, 1333, 276, 27605 },
5246 { 0, 1066, 276, 27605 },
5247 { 0, 800, 231, 23784 },
5250 static unsigned long __i915_chipset_val(struct drm_i915_private
*dev_priv
)
5252 u64 total_count
, diff
, ret
;
5253 u32 count1
, count2
, count3
, m
= 0, c
= 0;
5254 unsigned long now
= jiffies_to_msecs(jiffies
), diff1
;
5257 assert_spin_locked(&mchdev_lock
);
5259 diff1
= now
- dev_priv
->ips
.last_time1
;
5261 /* Prevent division-by-zero if we are asking too fast.
5262 * Also, we don't get interesting results if we are polling
5263 * faster than once in 10ms, so just return the saved value
5267 return dev_priv
->ips
.chipset_power
;
5269 count1
= I915_READ(DMIEC
);
5270 count2
= I915_READ(DDREC
);
5271 count3
= I915_READ(CSIEC
);
5273 total_count
= count1
+ count2
+ count3
;
5275 /* FIXME: handle per-counter overflow */
5276 if (total_count
< dev_priv
->ips
.last_count1
) {
5277 diff
= ~0UL - dev_priv
->ips
.last_count1
;
5278 diff
+= total_count
;
5280 diff
= total_count
- dev_priv
->ips
.last_count1
;
5283 for (i
= 0; i
< ARRAY_SIZE(cparams
); i
++) {
5284 if (cparams
[i
].i
== dev_priv
->ips
.c_m
&&
5285 cparams
[i
].t
== dev_priv
->ips
.r_t
) {
5292 diff
= div_u64(diff
, diff1
);
5293 ret
= ((m
* diff
) + c
);
5294 ret
= div_u64(ret
, 10);
5296 dev_priv
->ips
.last_count1
= total_count
;
5297 dev_priv
->ips
.last_time1
= now
;
5299 dev_priv
->ips
.chipset_power
= ret
;
5304 unsigned long i915_chipset_val(struct drm_i915_private
*dev_priv
)
5306 struct drm_device
*dev
= dev_priv
->dev
;
5309 if (INTEL_INFO(dev
)->gen
!= 5)
5312 spin_lock_irq(&mchdev_lock
);
5314 val
= __i915_chipset_val(dev_priv
);
5316 spin_unlock_irq(&mchdev_lock
);
5321 unsigned long i915_mch_val(struct drm_i915_private
*dev_priv
)
5323 unsigned long m
, x
, b
;
5326 tsfs
= I915_READ(TSFS
);
5328 m
= ((tsfs
& TSFS_SLOPE_MASK
) >> TSFS_SLOPE_SHIFT
);
5329 x
= I915_READ8(TR1
);
5331 b
= tsfs
& TSFS_INTR_MASK
;
5333 return ((m
* x
) / 127) - b
;
5336 static int _pxvid_to_vd(u8 pxvid
)
5341 if (pxvid
>= 8 && pxvid
< 31)
5344 return (pxvid
+ 2) * 125;
5347 static u32
pvid_to_extvid(struct drm_i915_private
*dev_priv
, u8 pxvid
)
5349 struct drm_device
*dev
= dev_priv
->dev
;
5350 const int vd
= _pxvid_to_vd(pxvid
);
5351 const int vm
= vd
- 1125;
5353 if (INTEL_INFO(dev
)->is_mobile
)
5354 return vm
> 0 ? vm
: 0;
5359 static void __i915_update_gfx_val(struct drm_i915_private
*dev_priv
)
5361 u64 now
, diff
, diffms
;
5364 assert_spin_locked(&mchdev_lock
);
5366 now
= ktime_get_raw_ns();
5367 diffms
= now
- dev_priv
->ips
.last_time2
;
5368 do_div(diffms
, NSEC_PER_MSEC
);
5370 /* Don't divide by 0 */
5374 count
= I915_READ(GFXEC
);
5376 if (count
< dev_priv
->ips
.last_count2
) {
5377 diff
= ~0UL - dev_priv
->ips
.last_count2
;
5380 diff
= count
- dev_priv
->ips
.last_count2
;
5383 dev_priv
->ips
.last_count2
= count
;
5384 dev_priv
->ips
.last_time2
= now
;
5386 /* More magic constants... */
5388 diff
= div_u64(diff
, diffms
* 10);
5389 dev_priv
->ips
.gfx_power
= diff
;
5392 void i915_update_gfx_val(struct drm_i915_private
*dev_priv
)
5394 struct drm_device
*dev
= dev_priv
->dev
;
5396 if (INTEL_INFO(dev
)->gen
!= 5)
5399 spin_lock_irq(&mchdev_lock
);
5401 __i915_update_gfx_val(dev_priv
);
5403 spin_unlock_irq(&mchdev_lock
);
5406 static unsigned long __i915_gfx_val(struct drm_i915_private
*dev_priv
)
5408 unsigned long t
, corr
, state1
, corr2
, state2
;
5411 assert_spin_locked(&mchdev_lock
);
5413 pxvid
= I915_READ(PXVFREQ_BASE
+ (dev_priv
->rps
.cur_freq
* 4));
5414 pxvid
= (pxvid
>> 24) & 0x7f;
5415 ext_v
= pvid_to_extvid(dev_priv
, pxvid
);
5419 t
= i915_mch_val(dev_priv
);
5421 /* Revel in the empirically derived constants */
5423 /* Correction factor in 1/100000 units */
5425 corr
= ((t
* 2349) + 135940);
5427 corr
= ((t
* 964) + 29317);
5429 corr
= ((t
* 301) + 1004);
5431 corr
= corr
* ((150142 * state1
) / 10000 - 78642);
5433 corr2
= (corr
* dev_priv
->ips
.corr
);
5435 state2
= (corr2
* state1
) / 10000;
5436 state2
/= 100; /* convert to mW */
5438 __i915_update_gfx_val(dev_priv
);
5440 return dev_priv
->ips
.gfx_power
+ state2
;
5443 unsigned long i915_gfx_val(struct drm_i915_private
*dev_priv
)
5445 struct drm_device
*dev
= dev_priv
->dev
;
5448 if (INTEL_INFO(dev
)->gen
!= 5)
5451 spin_lock_irq(&mchdev_lock
);
5453 val
= __i915_gfx_val(dev_priv
);
5455 spin_unlock_irq(&mchdev_lock
);
5461 * i915_read_mch_val - return value for IPS use
5463 * Calculate and return a value for the IPS driver to use when deciding whether
5464 * we have thermal and power headroom to increase CPU or GPU power budget.
5466 unsigned long i915_read_mch_val(void)
5468 struct drm_i915_private
*dev_priv
;
5469 unsigned long chipset_val
, graphics_val
, ret
= 0;
5471 spin_lock_irq(&mchdev_lock
);
5474 dev_priv
= i915_mch_dev
;
5476 chipset_val
= __i915_chipset_val(dev_priv
);
5477 graphics_val
= __i915_gfx_val(dev_priv
);
5479 ret
= chipset_val
+ graphics_val
;
5482 spin_unlock_irq(&mchdev_lock
);
5486 EXPORT_SYMBOL_GPL(i915_read_mch_val
);
5489 * i915_gpu_raise - raise GPU frequency limit
5491 * Raise the limit; IPS indicates we have thermal headroom.
5493 bool i915_gpu_raise(void)
5495 struct drm_i915_private
*dev_priv
;
5498 spin_lock_irq(&mchdev_lock
);
5499 if (!i915_mch_dev
) {
5503 dev_priv
= i915_mch_dev
;
5505 if (dev_priv
->ips
.max_delay
> dev_priv
->ips
.fmax
)
5506 dev_priv
->ips
.max_delay
--;
5509 spin_unlock_irq(&mchdev_lock
);
5513 EXPORT_SYMBOL_GPL(i915_gpu_raise
);
5516 * i915_gpu_lower - lower GPU frequency limit
5518 * IPS indicates we're close to a thermal limit, so throttle back the GPU
5519 * frequency maximum.
5521 bool i915_gpu_lower(void)
5523 struct drm_i915_private
*dev_priv
;
5526 spin_lock_irq(&mchdev_lock
);
5527 if (!i915_mch_dev
) {
5531 dev_priv
= i915_mch_dev
;
5533 if (dev_priv
->ips
.max_delay
< dev_priv
->ips
.min_delay
)
5534 dev_priv
->ips
.max_delay
++;
5537 spin_unlock_irq(&mchdev_lock
);
5541 EXPORT_SYMBOL_GPL(i915_gpu_lower
);
5544 * i915_gpu_busy - indicate GPU business to IPS
5546 * Tell the IPS driver whether or not the GPU is busy.
5548 bool i915_gpu_busy(void)
5550 struct drm_i915_private
*dev_priv
;
5551 struct intel_engine_cs
*ring
;
5555 spin_lock_irq(&mchdev_lock
);
5558 dev_priv
= i915_mch_dev
;
5560 for_each_ring(ring
, dev_priv
, i
)
5561 ret
|= !list_empty(&ring
->request_list
);
5564 spin_unlock_irq(&mchdev_lock
);
5568 EXPORT_SYMBOL_GPL(i915_gpu_busy
);
5571 * i915_gpu_turbo_disable - disable graphics turbo
5573 * Disable graphics turbo by resetting the max frequency and setting the
5574 * current frequency to the default.
5576 bool i915_gpu_turbo_disable(void)
5578 struct drm_i915_private
*dev_priv
;
5581 spin_lock_irq(&mchdev_lock
);
5582 if (!i915_mch_dev
) {
5586 dev_priv
= i915_mch_dev
;
5588 dev_priv
->ips
.max_delay
= dev_priv
->ips
.fstart
;
5590 if (!ironlake_set_drps(dev_priv
->dev
, dev_priv
->ips
.fstart
))
5594 spin_unlock_irq(&mchdev_lock
);
5598 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable
);
5601 * Tells the intel_ips driver that the i915 driver is now loaded, if
5602 * IPS got loaded first.
5604 * This awkward dance is so that neither module has to depend on the
5605 * other in order for IPS to do the appropriate communication of
5606 * GPU turbo limits to i915.
5609 ips_ping_for_i915_load(void)
5613 link
= symbol_get(ips_link_to_i915_driver
);
5616 symbol_put(ips_link_to_i915_driver
);
5620 void intel_gpu_ips_init(struct drm_i915_private
*dev_priv
)
5622 /* We only register the i915 ips part with intel-ips once everything is
5623 * set up, to avoid intel-ips sneaking in and reading bogus values. */
5624 spin_lock_irq(&mchdev_lock
);
5625 i915_mch_dev
= dev_priv
;
5626 spin_unlock_irq(&mchdev_lock
);
5628 ips_ping_for_i915_load();
5631 void intel_gpu_ips_teardown(void)
5633 spin_lock_irq(&mchdev_lock
);
5634 i915_mch_dev
= NULL
;
5635 spin_unlock_irq(&mchdev_lock
);
5638 static void intel_init_emon(struct drm_device
*dev
)
5640 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5645 /* Disable to program */
5649 /* Program energy weights for various events */
5650 I915_WRITE(SDEW
, 0x15040d00);
5651 I915_WRITE(CSIEW0
, 0x007f0000);
5652 I915_WRITE(CSIEW1
, 0x1e220004);
5653 I915_WRITE(CSIEW2
, 0x04000004);
5655 for (i
= 0; i
< 5; i
++)
5656 I915_WRITE(PEW
+ (i
* 4), 0);
5657 for (i
= 0; i
< 3; i
++)
5658 I915_WRITE(DEW
+ (i
* 4), 0);
5660 /* Program P-state weights to account for frequency power adjustment */
5661 for (i
= 0; i
< 16; i
++) {
5662 u32 pxvidfreq
= I915_READ(PXVFREQ_BASE
+ (i
* 4));
5663 unsigned long freq
= intel_pxfreq(pxvidfreq
);
5664 unsigned long vid
= (pxvidfreq
& PXVFREQ_PX_MASK
) >>
5669 val
*= (freq
/ 1000);
5671 val
/= (127*127*900);
5673 DRM_ERROR("bad pxval: %ld\n", val
);
5676 /* Render standby states get 0 weight */
5680 for (i
= 0; i
< 4; i
++) {
5681 u32 val
= (pxw
[i
*4] << 24) | (pxw
[(i
*4)+1] << 16) |
5682 (pxw
[(i
*4)+2] << 8) | (pxw
[(i
*4)+3]);
5683 I915_WRITE(PXW
+ (i
* 4), val
);
5686 /* Adjust magic regs to magic values (more experimental results) */
5687 I915_WRITE(OGW0
, 0);
5688 I915_WRITE(OGW1
, 0);
5689 I915_WRITE(EG0
, 0x00007f00);
5690 I915_WRITE(EG1
, 0x0000000e);
5691 I915_WRITE(EG2
, 0x000e0000);
5692 I915_WRITE(EG3
, 0x68000300);
5693 I915_WRITE(EG4
, 0x42000000);
5694 I915_WRITE(EG5
, 0x00140031);
5698 for (i
= 0; i
< 8; i
++)
5699 I915_WRITE(PXWL
+ (i
* 4), 0);
5701 /* Enable PMON + select events */
5702 I915_WRITE(ECR
, 0x80000019);
5704 lcfuse
= I915_READ(LCFUSE02
);
5706 dev_priv
->ips
.corr
= (lcfuse
& LCFUSE_HIV_MASK
);
5709 void intel_init_gt_powersave(struct drm_device
*dev
)
5711 i915
.enable_rc6
= sanitize_rc6_option(dev
, i915
.enable_rc6
);
5713 if (IS_CHERRYVIEW(dev
))
5714 cherryview_init_gt_powersave(dev
);
5715 else if (IS_VALLEYVIEW(dev
))
5716 valleyview_init_gt_powersave(dev
);
5719 void intel_cleanup_gt_powersave(struct drm_device
*dev
)
5721 if (IS_CHERRYVIEW(dev
))
5723 else if (IS_VALLEYVIEW(dev
))
5724 valleyview_cleanup_gt_powersave(dev
);
5727 static void gen6_suspend_rps(struct drm_device
*dev
)
5729 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5731 flush_delayed_work(&dev_priv
->rps
.delayed_resume_work
);
5733 gen6_disable_rps_interrupts(dev
);
5737 * intel_suspend_gt_powersave - suspend PM work and helper threads
5740 * We don't want to disable RC6 or other features here, we just want
5741 * to make sure any work we've queued has finished and won't bother
5742 * us while we're suspended.
5744 void intel_suspend_gt_powersave(struct drm_device
*dev
)
5746 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5748 if (INTEL_INFO(dev
)->gen
< 6)
5751 gen6_suspend_rps(dev
);
5753 /* Force GPU to min freq during suspend */
5754 gen6_rps_idle(dev_priv
);
5757 void intel_disable_gt_powersave(struct drm_device
*dev
)
5759 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5761 if (IS_IRONLAKE_M(dev
)) {
5762 ironlake_disable_drps(dev
);
5763 } else if (INTEL_INFO(dev
)->gen
>= 6) {
5764 intel_suspend_gt_powersave(dev
);
5766 mutex_lock(&dev_priv
->rps
.hw_lock
);
5767 if (INTEL_INFO(dev
)->gen
>= 9)
5768 gen9_disable_rps(dev
);
5769 else if (IS_CHERRYVIEW(dev
))
5770 cherryview_disable_rps(dev
);
5771 else if (IS_VALLEYVIEW(dev
))
5772 valleyview_disable_rps(dev
);
5774 gen6_disable_rps(dev
);
5776 dev_priv
->rps
.enabled
= false;
5777 mutex_unlock(&dev_priv
->rps
.hw_lock
);
5781 static void intel_gen6_powersave_work(struct work_struct
*work
)
5783 struct drm_i915_private
*dev_priv
=
5784 container_of(work
, struct drm_i915_private
,
5785 rps
.delayed_resume_work
.work
);
5786 struct drm_device
*dev
= dev_priv
->dev
;
5788 mutex_lock(&dev_priv
->rps
.hw_lock
);
5790 gen6_reset_rps_interrupts(dev
);
5792 if (IS_CHERRYVIEW(dev
)) {
5793 cherryview_enable_rps(dev
);
5794 } else if (IS_VALLEYVIEW(dev
)) {
5795 valleyview_enable_rps(dev
);
5796 } else if (INTEL_INFO(dev
)->gen
>= 9) {
5797 gen9_enable_rc6(dev
);
5798 gen9_enable_rps(dev
);
5799 __gen6_update_ring_freq(dev
);
5800 } else if (IS_BROADWELL(dev
)) {
5801 gen8_enable_rps(dev
);
5802 __gen6_update_ring_freq(dev
);
5804 gen6_enable_rps(dev
);
5805 __gen6_update_ring_freq(dev
);
5808 WARN_ON(dev_priv
->rps
.max_freq
< dev_priv
->rps
.min_freq
);
5809 WARN_ON(dev_priv
->rps
.idle_freq
> dev_priv
->rps
.max_freq
);
5811 WARN_ON(dev_priv
->rps
.efficient_freq
< dev_priv
->rps
.min_freq
);
5812 WARN_ON(dev_priv
->rps
.efficient_freq
> dev_priv
->rps
.max_freq
);
5814 dev_priv
->rps
.enabled
= true;
5816 gen6_enable_rps_interrupts(dev
);
5818 mutex_unlock(&dev_priv
->rps
.hw_lock
);
5820 intel_runtime_pm_put(dev_priv
);
5823 void intel_enable_gt_powersave(struct drm_device
*dev
)
5825 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5827 /* Powersaving is controlled by the host when inside a VM */
5828 if (intel_vgpu_active(dev
))
5831 if (IS_IRONLAKE_M(dev
)) {
5832 mutex_lock(&dev
->struct_mutex
);
5833 ironlake_enable_drps(dev
);
5834 intel_init_emon(dev
);
5835 mutex_unlock(&dev
->struct_mutex
);
5836 } else if (INTEL_INFO(dev
)->gen
>= 6) {
5838 * PCU communication is slow and this doesn't need to be
5839 * done at any specific time, so do this out of our fast path
5840 * to make resume and init faster.
5842 * We depend on the HW RC6 power context save/restore
5843 * mechanism when entering D3 through runtime PM suspend. So
5844 * disable RPM until RPS/RC6 is properly setup. We can only
5845 * get here via the driver load/system resume/runtime resume
5846 * paths, so the _noresume version is enough (and in case of
5847 * runtime resume it's necessary).
5849 if (schedule_delayed_work(&dev_priv
->rps
.delayed_resume_work
,
5850 round_jiffies_up_relative(HZ
)))
5851 intel_runtime_pm_get_noresume(dev_priv
);
5855 void intel_reset_gt_powersave(struct drm_device
*dev
)
5857 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5859 if (INTEL_INFO(dev
)->gen
< 6)
5862 gen6_suspend_rps(dev
);
5863 dev_priv
->rps
.enabled
= false;
5866 static void ibx_init_clock_gating(struct drm_device
*dev
)
5868 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5871 * On Ibex Peak and Cougar Point, we need to disable clock
5872 * gating for the panel power sequencer or it will fail to
5873 * start up when no ports are active.
5875 I915_WRITE(SOUTH_DSPCLK_GATE_D
, PCH_DPLSUNIT_CLOCK_GATE_DISABLE
);
5878 static void g4x_disable_trickle_feed(struct drm_device
*dev
)
5880 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5883 for_each_pipe(dev_priv
, pipe
) {
5884 I915_WRITE(DSPCNTR(pipe
),
5885 I915_READ(DSPCNTR(pipe
)) |
5886 DISPPLANE_TRICKLE_FEED_DISABLE
);
5887 intel_flush_primary_plane(dev_priv
, pipe
);
5891 static void ilk_init_lp_watermarks(struct drm_device
*dev
)
5893 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5895 I915_WRITE(WM3_LP_ILK
, I915_READ(WM3_LP_ILK
) & ~WM1_LP_SR_EN
);
5896 I915_WRITE(WM2_LP_ILK
, I915_READ(WM2_LP_ILK
) & ~WM1_LP_SR_EN
);
5897 I915_WRITE(WM1_LP_ILK
, I915_READ(WM1_LP_ILK
) & ~WM1_LP_SR_EN
);
5900 * Don't touch WM1S_LP_EN here.
5901 * Doing so could cause underruns.
5905 static void ironlake_init_clock_gating(struct drm_device
*dev
)
5907 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5908 uint32_t dspclk_gate
= ILK_VRHUNIT_CLOCK_GATE_DISABLE
;
5912 * WaFbcDisableDpfcClockGating:ilk
5914 dspclk_gate
|= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE
|
5915 ILK_DPFCUNIT_CLOCK_GATE_DISABLE
|
5916 ILK_DPFDUNIT_CLOCK_GATE_ENABLE
;
5918 I915_WRITE(PCH_3DCGDIS0
,
5919 MARIUNIT_CLOCK_GATE_DISABLE
|
5920 SVSMUNIT_CLOCK_GATE_DISABLE
);
5921 I915_WRITE(PCH_3DCGDIS1
,
5922 VFMUNIT_CLOCK_GATE_DISABLE
);
5925 * According to the spec the following bits should be set in
5926 * order to enable memory self-refresh
5927 * The bit 22/21 of 0x42004
5928 * The bit 5 of 0x42020
5929 * The bit 15 of 0x45000
5931 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
5932 (I915_READ(ILK_DISPLAY_CHICKEN2
) |
5933 ILK_DPARB_GATE
| ILK_VSDPFD_FULL
));
5934 dspclk_gate
|= ILK_DPARBUNIT_CLOCK_GATE_ENABLE
;
5935 I915_WRITE(DISP_ARB_CTL
,
5936 (I915_READ(DISP_ARB_CTL
) |
5939 ilk_init_lp_watermarks(dev
);
5942 * Based on the document from hardware guys the following bits
5943 * should be set unconditionally in order to enable FBC.
5944 * The bit 22 of 0x42000
5945 * The bit 22 of 0x42004
5946 * The bit 7,8,9 of 0x42020.
5948 if (IS_IRONLAKE_M(dev
)) {
5949 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
5950 I915_WRITE(ILK_DISPLAY_CHICKEN1
,
5951 I915_READ(ILK_DISPLAY_CHICKEN1
) |
5953 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
5954 I915_READ(ILK_DISPLAY_CHICKEN2
) |
5958 I915_WRITE(ILK_DSPCLK_GATE_D
, dspclk_gate
);
5960 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
5961 I915_READ(ILK_DISPLAY_CHICKEN2
) |
5962 ILK_ELPIN_409_SELECT
);
5963 I915_WRITE(_3D_CHICKEN2
,
5964 _3D_CHICKEN2_WM_READ_PIPELINED
<< 16 |
5965 _3D_CHICKEN2_WM_READ_PIPELINED
);
5967 /* WaDisableRenderCachePipelinedFlush:ilk */
5968 I915_WRITE(CACHE_MODE_0
,
5969 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE
));
5971 /* WaDisable_RenderCache_OperationalFlush:ilk */
5972 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
5974 g4x_disable_trickle_feed(dev
);
5976 ibx_init_clock_gating(dev
);
5979 static void cpt_init_clock_gating(struct drm_device
*dev
)
5981 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
5986 * On Ibex Peak and Cougar Point, we need to disable clock
5987 * gating for the panel power sequencer or it will fail to
5988 * start up when no ports are active.
5990 I915_WRITE(SOUTH_DSPCLK_GATE_D
, PCH_DPLSUNIT_CLOCK_GATE_DISABLE
|
5991 PCH_DPLUNIT_CLOCK_GATE_DISABLE
|
5992 PCH_CPUNIT_CLOCK_GATE_DISABLE
);
5993 I915_WRITE(SOUTH_CHICKEN2
, I915_READ(SOUTH_CHICKEN2
) |
5994 DPLS_EDP_PPS_FIX_DIS
);
5995 /* The below fixes the weird display corruption, a few pixels shifted
5996 * downward, on (only) LVDS of some HP laptops with IVY.
5998 for_each_pipe(dev_priv
, pipe
) {
5999 val
= I915_READ(TRANS_CHICKEN2(pipe
));
6000 val
|= TRANS_CHICKEN2_TIMING_OVERRIDE
;
6001 val
&= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED
;
6002 if (dev_priv
->vbt
.fdi_rx_polarity_inverted
)
6003 val
|= TRANS_CHICKEN2_FDI_POLARITY_REVERSED
;
6004 val
&= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK
;
6005 val
&= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER
;
6006 val
&= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH
;
6007 I915_WRITE(TRANS_CHICKEN2(pipe
), val
);
6009 /* WADP0ClockGatingDisable */
6010 for_each_pipe(dev_priv
, pipe
) {
6011 I915_WRITE(TRANS_CHICKEN1(pipe
),
6012 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE
);
6016 static void gen6_check_mch_setup(struct drm_device
*dev
)
6018 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6021 tmp
= I915_READ(MCH_SSKPD
);
6022 if ((tmp
& MCH_SSKPD_WM0_MASK
) != MCH_SSKPD_WM0_VAL
)
6023 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
6027 static void gen6_init_clock_gating(struct drm_device
*dev
)
6029 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6030 uint32_t dspclk_gate
= ILK_VRHUNIT_CLOCK_GATE_DISABLE
;
6032 I915_WRITE(ILK_DSPCLK_GATE_D
, dspclk_gate
);
6034 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6035 I915_READ(ILK_DISPLAY_CHICKEN2
) |
6036 ILK_ELPIN_409_SELECT
);
6038 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
6039 I915_WRITE(_3D_CHICKEN
,
6040 _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB
));
6042 /* WaDisable_RenderCache_OperationalFlush:snb */
6043 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6046 * BSpec recoomends 8x4 when MSAA is used,
6047 * however in practice 16x4 seems fastest.
6049 * Note that PS/WM thread counts depend on the WIZ hashing
6050 * disable bit, which we don't touch here, but it's good
6051 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6053 I915_WRITE(GEN6_GT_MODE
,
6054 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
6056 ilk_init_lp_watermarks(dev
);
6058 I915_WRITE(CACHE_MODE_0
,
6059 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB
));
6061 I915_WRITE(GEN6_UCGCTL1
,
6062 I915_READ(GEN6_UCGCTL1
) |
6063 GEN6_BLBUNIT_CLOCK_GATE_DISABLE
|
6064 GEN6_CSUNIT_CLOCK_GATE_DISABLE
);
6066 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
6067 * gating disable must be set. Failure to set it results in
6068 * flickering pixels due to Z write ordering failures after
6069 * some amount of runtime in the Mesa "fire" demo, and Unigine
6070 * Sanctuary and Tropics, and apparently anything else with
6071 * alpha test or pixel discard.
6073 * According to the spec, bit 11 (RCCUNIT) must also be set,
6074 * but we didn't debug actual testcases to find it out.
6076 * WaDisableRCCUnitClockGating:snb
6077 * WaDisableRCPBUnitClockGating:snb
6079 I915_WRITE(GEN6_UCGCTL2
,
6080 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE
|
6081 GEN6_RCCUNIT_CLOCK_GATE_DISABLE
);
6083 /* WaStripsFansDisableFastClipPerformanceFix:snb */
6084 I915_WRITE(_3D_CHICKEN3
,
6085 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL
));
6089 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
6090 * 3DSTATE_SF number of SF output attributes is more than 16."
6092 I915_WRITE(_3D_CHICKEN3
,
6093 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH
));
6096 * According to the spec the following bits should be
6097 * set in order to enable memory self-refresh and fbc:
6098 * The bit21 and bit22 of 0x42000
6099 * The bit21 and bit22 of 0x42004
6100 * The bit5 and bit7 of 0x42020
6101 * The bit14 of 0x70180
6102 * The bit14 of 0x71180
6104 * WaFbcAsynchFlipDisableFbcQueue:snb
6106 I915_WRITE(ILK_DISPLAY_CHICKEN1
,
6107 I915_READ(ILK_DISPLAY_CHICKEN1
) |
6108 ILK_FBCQ_DIS
| ILK_PABSTRETCH_DIS
);
6109 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6110 I915_READ(ILK_DISPLAY_CHICKEN2
) |
6111 ILK_DPARB_GATE
| ILK_VSDPFD_FULL
);
6112 I915_WRITE(ILK_DSPCLK_GATE_D
,
6113 I915_READ(ILK_DSPCLK_GATE_D
) |
6114 ILK_DPARBUNIT_CLOCK_GATE_ENABLE
|
6115 ILK_DPFDUNIT_CLOCK_GATE_ENABLE
);
6117 g4x_disable_trickle_feed(dev
);
6119 cpt_init_clock_gating(dev
);
6121 gen6_check_mch_setup(dev
);
6124 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private
*dev_priv
)
6126 uint32_t reg
= I915_READ(GEN7_FF_THREAD_MODE
);
6129 * WaVSThreadDispatchOverride:ivb,vlv
6131 * This actually overrides the dispatch
6132 * mode for all thread types.
6134 reg
&= ~GEN7_FF_SCHED_MASK
;
6135 reg
|= GEN7_FF_TS_SCHED_HW
;
6136 reg
|= GEN7_FF_VS_SCHED_HW
;
6137 reg
|= GEN7_FF_DS_SCHED_HW
;
6139 I915_WRITE(GEN7_FF_THREAD_MODE
, reg
);
6142 static void lpt_init_clock_gating(struct drm_device
*dev
)
6144 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6147 * TODO: this bit should only be enabled when really needed, then
6148 * disabled when not needed anymore in order to save power.
6150 if (dev_priv
->pch_id
== INTEL_PCH_LPT_LP_DEVICE_ID_TYPE
)
6151 I915_WRITE(SOUTH_DSPCLK_GATE_D
,
6152 I915_READ(SOUTH_DSPCLK_GATE_D
) |
6153 PCH_LP_PARTITION_LEVEL_DISABLE
);
6155 /* WADPOClockGatingDisable:hsw */
6156 I915_WRITE(_TRANSA_CHICKEN1
,
6157 I915_READ(_TRANSA_CHICKEN1
) |
6158 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE
);
6161 static void lpt_suspend_hw(struct drm_device
*dev
)
6163 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6165 if (dev_priv
->pch_id
== INTEL_PCH_LPT_LP_DEVICE_ID_TYPE
) {
6166 uint32_t val
= I915_READ(SOUTH_DSPCLK_GATE_D
);
6168 val
&= ~PCH_LP_PARTITION_LEVEL_DISABLE
;
6169 I915_WRITE(SOUTH_DSPCLK_GATE_D
, val
);
6173 static void broadwell_init_clock_gating(struct drm_device
*dev
)
6175 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6178 ilk_init_lp_watermarks(dev
);
6180 /* WaSwitchSolVfFArbitrationPriority:bdw */
6181 I915_WRITE(GAM_ECOCHK
, I915_READ(GAM_ECOCHK
) | HSW_ECOCHK_ARB_PRIO_SOL
);
6183 /* WaPsrDPAMaskVBlankInSRD:bdw */
6184 I915_WRITE(CHICKEN_PAR1_1
,
6185 I915_READ(CHICKEN_PAR1_1
) | DPA_MASK_VBLANK_SRD
);
6187 /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
6188 for_each_pipe(dev_priv
, pipe
) {
6189 I915_WRITE(CHICKEN_PIPESL_1(pipe
),
6190 I915_READ(CHICKEN_PIPESL_1(pipe
)) |
6191 BDW_DPRS_MASK_VBLANK_SRD
);
6194 /* WaVSRefCountFullforceMissDisable:bdw */
6195 /* WaDSRefCountFullforceMissDisable:bdw */
6196 I915_WRITE(GEN7_FF_THREAD_MODE
,
6197 I915_READ(GEN7_FF_THREAD_MODE
) &
6198 ~(GEN8_FF_DS_REF_CNT_FFME
| GEN7_FF_VS_REF_CNT_FFME
));
6200 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL
,
6201 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE
));
6203 /* WaDisableSDEUnitClockGating:bdw */
6204 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
6205 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
6207 lpt_init_clock_gating(dev
);
6210 static void haswell_init_clock_gating(struct drm_device
*dev
)
6212 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6214 ilk_init_lp_watermarks(dev
);
6216 /* L3 caching of data atomics doesn't work -- disable it. */
6217 I915_WRITE(HSW_SCRATCH1
, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE
);
6218 I915_WRITE(HSW_ROW_CHICKEN3
,
6219 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE
));
6221 /* This is required by WaCatErrorRejectionIssue:hsw */
6222 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
6223 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
6224 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
6226 /* WaVSRefCountFullforceMissDisable:hsw */
6227 I915_WRITE(GEN7_FF_THREAD_MODE
,
6228 I915_READ(GEN7_FF_THREAD_MODE
) & ~GEN7_FF_VS_REF_CNT_FFME
);
6230 /* WaDisable_RenderCache_OperationalFlush:hsw */
6231 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6233 /* enable HiZ Raw Stall Optimization */
6234 I915_WRITE(CACHE_MODE_0_GEN7
,
6235 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE
));
6237 /* WaDisable4x2SubspanOptimization:hsw */
6238 I915_WRITE(CACHE_MODE_1
,
6239 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
6242 * BSpec recommends 8x4 when MSAA is used,
6243 * however in practice 16x4 seems fastest.
6245 * Note that PS/WM thread counts depend on the WIZ hashing
6246 * disable bit, which we don't touch here, but it's good
6247 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6249 I915_WRITE(GEN7_GT_MODE
,
6250 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
6252 /* WaSampleCChickenBitEnable:hsw */
6253 I915_WRITE(HALF_SLICE_CHICKEN3
,
6254 _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE
));
6256 /* WaSwitchSolVfFArbitrationPriority:hsw */
6257 I915_WRITE(GAM_ECOCHK
, I915_READ(GAM_ECOCHK
) | HSW_ECOCHK_ARB_PRIO_SOL
);
6259 /* WaRsPkgCStateDisplayPMReq:hsw */
6260 I915_WRITE(CHICKEN_PAR1_1
,
6261 I915_READ(CHICKEN_PAR1_1
) | FORCE_ARB_IDLE_PLANES
);
6263 lpt_init_clock_gating(dev
);
6266 static void ivybridge_init_clock_gating(struct drm_device
*dev
)
6268 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6271 ilk_init_lp_watermarks(dev
);
6273 I915_WRITE(ILK_DSPCLK_GATE_D
, ILK_VRHUNIT_CLOCK_GATE_DISABLE
);
6275 /* WaDisableEarlyCull:ivb */
6276 I915_WRITE(_3D_CHICKEN3
,
6277 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL
));
6279 /* WaDisableBackToBackFlipFix:ivb */
6280 I915_WRITE(IVB_CHICKEN3
,
6281 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE
|
6282 CHICKEN3_DGMG_DONE_FIX_DISABLE
);
6284 /* WaDisablePSDDualDispatchEnable:ivb */
6285 if (IS_IVB_GT1(dev
))
6286 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1
,
6287 _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE
));
6289 /* WaDisable_RenderCache_OperationalFlush:ivb */
6290 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6292 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
6293 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1
,
6294 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC
);
6296 /* WaApplyL3ControlAndL3ChickenMode:ivb */
6297 I915_WRITE(GEN7_L3CNTLREG1
,
6298 GEN7_WA_FOR_GEN7_L3_CONTROL
);
6299 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER
,
6300 GEN7_WA_L3_CHICKEN_MODE
);
6301 if (IS_IVB_GT1(dev
))
6302 I915_WRITE(GEN7_ROW_CHICKEN2
,
6303 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
6305 /* must write both registers */
6306 I915_WRITE(GEN7_ROW_CHICKEN2
,
6307 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
6308 I915_WRITE(GEN7_ROW_CHICKEN2_GT2
,
6309 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
6312 /* WaForceL3Serialization:ivb */
6313 I915_WRITE(GEN7_L3SQCREG4
, I915_READ(GEN7_L3SQCREG4
) &
6314 ~L3SQ_URB_READ_CAM_MATCH_DISABLE
);
6317 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6318 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
6320 I915_WRITE(GEN6_UCGCTL2
,
6321 GEN6_RCZUNIT_CLOCK_GATE_DISABLE
);
6323 /* This is required by WaCatErrorRejectionIssue:ivb */
6324 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
6325 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
6326 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
6328 g4x_disable_trickle_feed(dev
);
6330 gen7_setup_fixed_func_scheduler(dev_priv
);
6332 if (0) { /* causes HiZ corruption on ivb:gt1 */
6333 /* enable HiZ Raw Stall Optimization */
6334 I915_WRITE(CACHE_MODE_0_GEN7
,
6335 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE
));
6338 /* WaDisable4x2SubspanOptimization:ivb */
6339 I915_WRITE(CACHE_MODE_1
,
6340 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
6343 * BSpec recommends 8x4 when MSAA is used,
6344 * however in practice 16x4 seems fastest.
6346 * Note that PS/WM thread counts depend on the WIZ hashing
6347 * disable bit, which we don't touch here, but it's good
6348 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6350 I915_WRITE(GEN7_GT_MODE
,
6351 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
6353 snpcr
= I915_READ(GEN6_MBCUNIT_SNPCR
);
6354 snpcr
&= ~GEN6_MBC_SNPCR_MASK
;
6355 snpcr
|= GEN6_MBC_SNPCR_MED
;
6356 I915_WRITE(GEN6_MBCUNIT_SNPCR
, snpcr
);
6358 if (!HAS_PCH_NOP(dev
))
6359 cpt_init_clock_gating(dev
);
6361 gen6_check_mch_setup(dev
);
6364 static void vlv_init_display_clock_gating(struct drm_i915_private
*dev_priv
)
6366 I915_WRITE(DSPCLK_GATE_D
, VRHUNIT_CLOCK_GATE_DISABLE
);
6369 * Disable trickle feed and enable pnd deadline calculation
6371 I915_WRITE(MI_ARB_VLV
, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
);
6372 I915_WRITE(CBR1_VLV
, 0);
6375 static void valleyview_init_clock_gating(struct drm_device
*dev
)
6377 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6379 vlv_init_display_clock_gating(dev_priv
);
6381 /* WaDisableEarlyCull:vlv */
6382 I915_WRITE(_3D_CHICKEN3
,
6383 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL
));
6385 /* WaDisableBackToBackFlipFix:vlv */
6386 I915_WRITE(IVB_CHICKEN3
,
6387 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE
|
6388 CHICKEN3_DGMG_DONE_FIX_DISABLE
);
6390 /* WaPsdDispatchEnable:vlv */
6391 /* WaDisablePSDDualDispatchEnable:vlv */
6392 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1
,
6393 _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP
|
6394 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE
));
6396 /* WaDisable_RenderCache_OperationalFlush:vlv */
6397 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6399 /* WaForceL3Serialization:vlv */
6400 I915_WRITE(GEN7_L3SQCREG4
, I915_READ(GEN7_L3SQCREG4
) &
6401 ~L3SQ_URB_READ_CAM_MATCH_DISABLE
);
6403 /* WaDisableDopClockGating:vlv */
6404 I915_WRITE(GEN7_ROW_CHICKEN2
,
6405 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
6407 /* This is required by WaCatErrorRejectionIssue:vlv */
6408 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
6409 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
6410 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
6412 gen7_setup_fixed_func_scheduler(dev_priv
);
6415 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6416 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
6418 I915_WRITE(GEN6_UCGCTL2
,
6419 GEN6_RCZUNIT_CLOCK_GATE_DISABLE
);
6421 /* WaDisableL3Bank2xClockGate:vlv
6422 * Disabling L3 clock gating- MMIO 940c[25] = 1
6423 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
6424 I915_WRITE(GEN7_UCGCTL4
,
6425 I915_READ(GEN7_UCGCTL4
) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE
);
6428 * BSpec says this must be set, even though
6429 * WaDisable4x2SubspanOptimization isn't listed for VLV.
6431 I915_WRITE(CACHE_MODE_1
,
6432 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
6435 * BSpec recommends 8x4 when MSAA is used,
6436 * however in practice 16x4 seems fastest.
6438 * Note that PS/WM thread counts depend on the WIZ hashing
6439 * disable bit, which we don't touch here, but it's good
6440 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6442 I915_WRITE(GEN7_GT_MODE
,
6443 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
6446 * WaIncreaseL3CreditsForVLVB0:vlv
6447 * This is the hardware default actually.
6449 I915_WRITE(GEN7_L3SQCREG1
, VLV_B0_WA_L3SQCREG1_VALUE
);
6452 * WaDisableVLVClockGating_VBIIssue:vlv
6453 * Disable clock gating on th GCFG unit to prevent a delay
6454 * in the reporting of vblank events.
6456 I915_WRITE(VLV_GUNIT_CLOCK_GATE
, GCFG_DIS
);
6459 static void cherryview_init_clock_gating(struct drm_device
*dev
)
6461 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6463 vlv_init_display_clock_gating(dev_priv
);
6465 /* WaVSRefCountFullforceMissDisable:chv */
6466 /* WaDSRefCountFullforceMissDisable:chv */
6467 I915_WRITE(GEN7_FF_THREAD_MODE
,
6468 I915_READ(GEN7_FF_THREAD_MODE
) &
6469 ~(GEN8_FF_DS_REF_CNT_FFME
| GEN7_FF_VS_REF_CNT_FFME
));
6471 /* WaDisableSemaphoreAndSyncFlipWait:chv */
6472 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL
,
6473 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE
));
6475 /* WaDisableCSUnitClockGating:chv */
6476 I915_WRITE(GEN6_UCGCTL1
, I915_READ(GEN6_UCGCTL1
) |
6477 GEN6_CSUNIT_CLOCK_GATE_DISABLE
);
6479 /* WaDisableSDEUnitClockGating:chv */
6480 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
6481 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
6484 static void g4x_init_clock_gating(struct drm_device
*dev
)
6486 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6487 uint32_t dspclk_gate
;
6489 I915_WRITE(RENCLK_GATE_D1
, 0);
6490 I915_WRITE(RENCLK_GATE_D2
, VF_UNIT_CLOCK_GATE_DISABLE
|
6491 GS_UNIT_CLOCK_GATE_DISABLE
|
6492 CL_UNIT_CLOCK_GATE_DISABLE
);
6493 I915_WRITE(RAMCLK_GATE_D
, 0);
6494 dspclk_gate
= VRHUNIT_CLOCK_GATE_DISABLE
|
6495 OVRUNIT_CLOCK_GATE_DISABLE
|
6496 OVCUNIT_CLOCK_GATE_DISABLE
;
6498 dspclk_gate
|= DSSUNIT_CLOCK_GATE_DISABLE
;
6499 I915_WRITE(DSPCLK_GATE_D
, dspclk_gate
);
6501 /* WaDisableRenderCachePipelinedFlush */
6502 I915_WRITE(CACHE_MODE_0
,
6503 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE
));
6505 /* WaDisable_RenderCache_OperationalFlush:g4x */
6506 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6508 g4x_disable_trickle_feed(dev
);
6511 static void crestline_init_clock_gating(struct drm_device
*dev
)
6513 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6515 I915_WRITE(RENCLK_GATE_D1
, I965_RCC_CLOCK_GATE_DISABLE
);
6516 I915_WRITE(RENCLK_GATE_D2
, 0);
6517 I915_WRITE(DSPCLK_GATE_D
, 0);
6518 I915_WRITE(RAMCLK_GATE_D
, 0);
6519 I915_WRITE16(DEUC
, 0);
6520 I915_WRITE(MI_ARB_STATE
,
6521 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
6523 /* WaDisable_RenderCache_OperationalFlush:gen4 */
6524 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6527 static void broadwater_init_clock_gating(struct drm_device
*dev
)
6529 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6531 I915_WRITE(RENCLK_GATE_D1
, I965_RCZ_CLOCK_GATE_DISABLE
|
6532 I965_RCC_CLOCK_GATE_DISABLE
|
6533 I965_RCPB_CLOCK_GATE_DISABLE
|
6534 I965_ISC_CLOCK_GATE_DISABLE
|
6535 I965_FBC_CLOCK_GATE_DISABLE
);
6536 I915_WRITE(RENCLK_GATE_D2
, 0);
6537 I915_WRITE(MI_ARB_STATE
,
6538 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
6540 /* WaDisable_RenderCache_OperationalFlush:gen4 */
6541 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6544 static void gen3_init_clock_gating(struct drm_device
*dev
)
6546 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6547 u32 dstate
= I915_READ(D_STATE
);
6549 dstate
|= DSTATE_PLL_D3_OFF
| DSTATE_GFX_CLOCK_GATING
|
6550 DSTATE_DOT_CLOCK_GATING
;
6551 I915_WRITE(D_STATE
, dstate
);
6553 if (IS_PINEVIEW(dev
))
6554 I915_WRITE(ECOSKPD
, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY
));
6556 /* IIR "flip pending" means done if this bit is set */
6557 I915_WRITE(ECOSKPD
, _MASKED_BIT_DISABLE(ECO_FLIP_DONE
));
6559 /* interrupts should cause a wake up from C3 */
6560 I915_WRITE(INSTPM
, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN
));
6562 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
6563 I915_WRITE(MI_ARB_STATE
, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE
));
6565 I915_WRITE(MI_ARB_STATE
,
6566 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
6569 static void i85x_init_clock_gating(struct drm_device
*dev
)
6571 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6573 I915_WRITE(RENCLK_GATE_D1
, SV_CLOCK_GATE_DISABLE
);
6575 /* interrupts should cause a wake up from C3 */
6576 I915_WRITE(MI_STATE
, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN
) |
6577 _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE
));
6579 I915_WRITE(MEM_MODE
,
6580 _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE
));
6583 static void i830_init_clock_gating(struct drm_device
*dev
)
6585 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6587 I915_WRITE(DSPCLK_GATE_D
, OVRUNIT_CLOCK_GATE_DISABLE
);
6589 I915_WRITE(MEM_MODE
,
6590 _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE
) |
6591 _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE
));
6594 void intel_init_clock_gating(struct drm_device
*dev
)
6596 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6598 if (dev_priv
->display
.init_clock_gating
)
6599 dev_priv
->display
.init_clock_gating(dev
);
6602 void intel_suspend_hw(struct drm_device
*dev
)
6604 if (HAS_PCH_LPT(dev
))
6605 lpt_suspend_hw(dev
);
6608 /* Set up chip specific power management-related functions */
6609 void intel_init_pm(struct drm_device
*dev
)
6611 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6613 intel_fbc_init(dev_priv
);
6616 if (IS_PINEVIEW(dev
))
6617 i915_pineview_get_mem_freq(dev
);
6618 else if (IS_GEN5(dev
))
6619 i915_ironlake_get_mem_freq(dev
);
6621 /* For FIFO watermark updates */
6622 if (INTEL_INFO(dev
)->gen
>= 9) {
6623 skl_setup_wm_latency(dev
);
6625 if (IS_BROXTON(dev
))
6626 dev_priv
->display
.init_clock_gating
=
6627 bxt_init_clock_gating
;
6628 else if (IS_SKYLAKE(dev
))
6629 dev_priv
->display
.init_clock_gating
=
6630 skl_init_clock_gating
;
6631 dev_priv
->display
.update_wm
= skl_update_wm
;
6632 dev_priv
->display
.update_sprite_wm
= skl_update_sprite_wm
;
6633 } else if (HAS_PCH_SPLIT(dev
)) {
6634 ilk_setup_wm_latency(dev
);
6636 if ((IS_GEN5(dev
) && dev_priv
->wm
.pri_latency
[1] &&
6637 dev_priv
->wm
.spr_latency
[1] && dev_priv
->wm
.cur_latency
[1]) ||
6638 (!IS_GEN5(dev
) && dev_priv
->wm
.pri_latency
[0] &&
6639 dev_priv
->wm
.spr_latency
[0] && dev_priv
->wm
.cur_latency
[0])) {
6640 dev_priv
->display
.update_wm
= ilk_update_wm
;
6641 dev_priv
->display
.update_sprite_wm
= ilk_update_sprite_wm
;
6643 DRM_DEBUG_KMS("Failed to read display plane latency. "
6648 dev_priv
->display
.init_clock_gating
= ironlake_init_clock_gating
;
6649 else if (IS_GEN6(dev
))
6650 dev_priv
->display
.init_clock_gating
= gen6_init_clock_gating
;
6651 else if (IS_IVYBRIDGE(dev
))
6652 dev_priv
->display
.init_clock_gating
= ivybridge_init_clock_gating
;
6653 else if (IS_HASWELL(dev
))
6654 dev_priv
->display
.init_clock_gating
= haswell_init_clock_gating
;
6655 else if (INTEL_INFO(dev
)->gen
== 8)
6656 dev_priv
->display
.init_clock_gating
= broadwell_init_clock_gating
;
6657 } else if (IS_CHERRYVIEW(dev
)) {
6658 dev_priv
->display
.update_wm
= valleyview_update_wm
;
6659 dev_priv
->display
.update_sprite_wm
= valleyview_update_sprite_wm
;
6660 dev_priv
->display
.init_clock_gating
=
6661 cherryview_init_clock_gating
;
6662 } else if (IS_VALLEYVIEW(dev
)) {
6663 dev_priv
->display
.update_wm
= valleyview_update_wm
;
6664 dev_priv
->display
.update_sprite_wm
= valleyview_update_sprite_wm
;
6665 dev_priv
->display
.init_clock_gating
=
6666 valleyview_init_clock_gating
;
6667 } else if (IS_PINEVIEW(dev
)) {
6668 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev
),
6671 dev_priv
->mem_freq
)) {
6672 DRM_INFO("failed to find known CxSR latency "
6673 "(found ddr%s fsb freq %d, mem freq %d), "
6675 (dev_priv
->is_ddr3
== 1) ? "3" : "2",
6676 dev_priv
->fsb_freq
, dev_priv
->mem_freq
);
6677 /* Disable CxSR and never update its watermark again */
6678 intel_set_memory_cxsr(dev_priv
, false);
6679 dev_priv
->display
.update_wm
= NULL
;
6681 dev_priv
->display
.update_wm
= pineview_update_wm
;
6682 dev_priv
->display
.init_clock_gating
= gen3_init_clock_gating
;
6683 } else if (IS_G4X(dev
)) {
6684 dev_priv
->display
.update_wm
= g4x_update_wm
;
6685 dev_priv
->display
.init_clock_gating
= g4x_init_clock_gating
;
6686 } else if (IS_GEN4(dev
)) {
6687 dev_priv
->display
.update_wm
= i965_update_wm
;
6688 if (IS_CRESTLINE(dev
))
6689 dev_priv
->display
.init_clock_gating
= crestline_init_clock_gating
;
6690 else if (IS_BROADWATER(dev
))
6691 dev_priv
->display
.init_clock_gating
= broadwater_init_clock_gating
;
6692 } else if (IS_GEN3(dev
)) {
6693 dev_priv
->display
.update_wm
= i9xx_update_wm
;
6694 dev_priv
->display
.get_fifo_size
= i9xx_get_fifo_size
;
6695 dev_priv
->display
.init_clock_gating
= gen3_init_clock_gating
;
6696 } else if (IS_GEN2(dev
)) {
6697 if (INTEL_INFO(dev
)->num_pipes
== 1) {
6698 dev_priv
->display
.update_wm
= i845_update_wm
;
6699 dev_priv
->display
.get_fifo_size
= i845_get_fifo_size
;
6701 dev_priv
->display
.update_wm
= i9xx_update_wm
;
6702 dev_priv
->display
.get_fifo_size
= i830_get_fifo_size
;
6705 if (IS_I85X(dev
) || IS_I865G(dev
))
6706 dev_priv
->display
.init_clock_gating
= i85x_init_clock_gating
;
6708 dev_priv
->display
.init_clock_gating
= i830_init_clock_gating
;
6710 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
6714 int sandybridge_pcode_read(struct drm_i915_private
*dev_priv
, u32 mbox
, u32
*val
)
6716 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
6718 if (I915_READ(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) {
6719 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
6723 I915_WRITE(GEN6_PCODE_DATA
, *val
);
6724 I915_WRITE(GEN6_PCODE_DATA1
, 0);
6725 I915_WRITE(GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
| mbox
);
6727 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) == 0,
6729 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox
);
6733 *val
= I915_READ(GEN6_PCODE_DATA
);
6734 I915_WRITE(GEN6_PCODE_DATA
, 0);
6739 int sandybridge_pcode_write(struct drm_i915_private
*dev_priv
, u32 mbox
, u32 val
)
6741 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
6743 if (I915_READ(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) {
6744 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
6748 I915_WRITE(GEN6_PCODE_DATA
, val
);
6749 I915_WRITE(GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
| mbox
);
6751 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) == 0,
6753 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox
);
6757 I915_WRITE(GEN6_PCODE_DATA
, 0);
6762 static int vlv_gpu_freq_div(unsigned int czclk_freq
)
6764 switch (czclk_freq
) {
6779 static int byt_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
6781 int div
, czclk_freq
= DIV_ROUND_CLOSEST(dev_priv
->mem_freq
, 4);
6783 div
= vlv_gpu_freq_div(czclk_freq
);
6787 return DIV_ROUND_CLOSEST(czclk_freq
* (val
+ 6 - 0xbd), div
);
6790 static int byt_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
6792 int mul
, czclk_freq
= DIV_ROUND_CLOSEST(dev_priv
->mem_freq
, 4);
6794 mul
= vlv_gpu_freq_div(czclk_freq
);
6798 return DIV_ROUND_CLOSEST(mul
* val
, czclk_freq
) + 0xbd - 6;
6801 static int chv_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
6803 int div
, czclk_freq
= dev_priv
->rps
.cz_freq
;
6805 div
= vlv_gpu_freq_div(czclk_freq
) / 2;
6809 return DIV_ROUND_CLOSEST(czclk_freq
* val
, 2 * div
) / 2;
6812 static int chv_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
6814 int mul
, czclk_freq
= dev_priv
->rps
.cz_freq
;
6816 mul
= vlv_gpu_freq_div(czclk_freq
) / 2;
6820 /* CHV needs even values */
6821 return DIV_ROUND_CLOSEST(val
* 2 * mul
, czclk_freq
) * 2;
6824 int intel_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
6826 if (IS_GEN9(dev_priv
->dev
))
6827 return (val
* GT_FREQUENCY_MULTIPLIER
) / GEN9_FREQ_SCALER
;
6828 else if (IS_CHERRYVIEW(dev_priv
->dev
))
6829 return chv_gpu_freq(dev_priv
, val
);
6830 else if (IS_VALLEYVIEW(dev_priv
->dev
))
6831 return byt_gpu_freq(dev_priv
, val
);
6833 return val
* GT_FREQUENCY_MULTIPLIER
;
6836 int intel_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
6838 if (IS_GEN9(dev_priv
->dev
))
6839 return (val
* GEN9_FREQ_SCALER
) / GT_FREQUENCY_MULTIPLIER
;
6840 else if (IS_CHERRYVIEW(dev_priv
->dev
))
6841 return chv_freq_opcode(dev_priv
, val
);
6842 else if (IS_VALLEYVIEW(dev_priv
->dev
))
6843 return byt_freq_opcode(dev_priv
, val
);
6845 return val
/ GT_FREQUENCY_MULTIPLIER
;
6848 struct request_boost
{
6849 struct work_struct work
;
6850 struct drm_i915_gem_request
*req
;
6853 static void __intel_rps_boost_work(struct work_struct
*work
)
6855 struct request_boost
*boost
= container_of(work
, struct request_boost
, work
);
6856 struct drm_i915_gem_request
*req
= boost
->req
;
6858 if (!i915_gem_request_completed(req
, true))
6859 gen6_rps_boost(to_i915(req
->ring
->dev
), NULL
,
6860 req
->emitted_jiffies
);
6862 i915_gem_request_unreference__unlocked(req
);
6866 void intel_queue_rps_boost_for_request(struct drm_device
*dev
,
6867 struct drm_i915_gem_request
*req
)
6869 struct request_boost
*boost
;
6871 if (req
== NULL
|| INTEL_INFO(dev
)->gen
< 6)
6874 if (i915_gem_request_completed(req
, true))
6877 boost
= kmalloc(sizeof(*boost
), GFP_ATOMIC
);
6881 i915_gem_request_reference(req
);
6884 INIT_WORK(&boost
->work
, __intel_rps_boost_work
);
6885 queue_work(to_i915(dev
)->wq
, &boost
->work
);
6888 void intel_pm_setup(struct drm_device
*dev
)
6890 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
6892 mutex_init(&dev_priv
->rps
.hw_lock
);
6894 INIT_DELAYED_WORK(&dev_priv
->rps
.delayed_resume_work
,
6895 intel_gen6_powersave_work
);
6896 INIT_LIST_HEAD(&dev_priv
->rps
.clients
);
6897 INIT_LIST_HEAD(&dev_priv
->rps
.semaphores
.link
);
6898 INIT_LIST_HEAD(&dev_priv
->rps
.mmioflips
.link
);
6900 dev_priv
->pm
.suspended
= false;