2 * Copyright © 2012 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eugeni Dodonov <eugeni.dodonov@intel.com>
28 #include <linux/cpufreq.h>
29 #include <drm/drm_plane_helper.h>
31 #include "intel_drv.h"
32 #include "../../../platform/x86/intel_ips.h"
33 #include <linux/module.h>
34 #include <drm/drm_atomic_helper.h>
39 * RC6 is a special power stage which allows the GPU to enter an very
40 * low-voltage mode when idle, using down to 0V while at this stage. This
41 * stage is entered automatically when the GPU is idle when RC6 support is
42 * enabled, and as soon as new workload arises GPU wakes up automatically as well.
44 * There are different RC6 modes available in Intel GPU, which differentiate
45 * among each other with the latency required to enter and leave RC6 and
46 * voltage consumed by the GPU in different states.
48 * The combination of the following flags define which states GPU is allowed
49 * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
50 * RC6pp is deepest RC6. Their support by hardware varies according to the
51 * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
52 * which brings the most power savings; deeper states save more power, but
53 * require higher latency to switch to and wake up.
55 #define INTEL_RC6_ENABLE (1<<0)
56 #define INTEL_RC6p_ENABLE (1<<1)
57 #define INTEL_RC6pp_ENABLE (1<<2)
59 static void gen9_init_clock_gating(struct drm_i915_private
*dev_priv
)
61 /* See Bspec note for PSR2_CTL bit 31, Wa#828:skl,bxt,kbl */
62 I915_WRITE(CHICKEN_PAR1_1
,
63 I915_READ(CHICKEN_PAR1_1
) | SKL_EDP_PSR_FIX_RDWRAP
);
65 I915_WRITE(GEN8_CONFIG0
,
66 I915_READ(GEN8_CONFIG0
) | GEN9_DEFAULT_FIXES
);
68 /* WaEnableChickenDCPR:skl,bxt,kbl,glk */
69 I915_WRITE(GEN8_CHICKEN_DCPR_1
,
70 I915_READ(GEN8_CHICKEN_DCPR_1
) | MASK_WAKEMEM
);
72 /* WaFbcTurnOffFbcWatermark:skl,bxt,kbl */
73 /* WaFbcWakeMemOn:skl,bxt,kbl,glk */
74 I915_WRITE(DISP_ARB_CTL
, I915_READ(DISP_ARB_CTL
) |
76 DISP_FBC_MEMORY_WAKE
);
78 /* WaFbcHighMemBwCorruptionAvoidance:skl,bxt,kbl */
79 I915_WRITE(ILK_DPFC_CHICKEN
, I915_READ(ILK_DPFC_CHICKEN
) |
80 ILK_DPFC_DISABLE_DUMMY0
);
83 static void bxt_init_clock_gating(struct drm_i915_private
*dev_priv
)
85 gen9_init_clock_gating(dev_priv
);
87 /* WaDisableSDEUnitClockGating:bxt */
88 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
89 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
93 * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
95 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
96 GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ
);
99 * Wa: Backlight PWM may stop in the asserted state, causing backlight
102 I915_WRITE(GEN9_CLKGATE_DIS_0
, I915_READ(GEN9_CLKGATE_DIS_0
) |
103 PWM1_GATING_DIS
| PWM2_GATING_DIS
);
106 static void glk_init_clock_gating(struct drm_i915_private
*dev_priv
)
108 gen9_init_clock_gating(dev_priv
);
111 * WaDisablePWMClockGating:glk
112 * Backlight PWM may stop in the asserted state, causing backlight
115 I915_WRITE(GEN9_CLKGATE_DIS_0
, I915_READ(GEN9_CLKGATE_DIS_0
) |
116 PWM1_GATING_DIS
| PWM2_GATING_DIS
);
119 static void i915_pineview_get_mem_freq(struct drm_i915_private
*dev_priv
)
123 tmp
= I915_READ(CLKCFG
);
125 switch (tmp
& CLKCFG_FSB_MASK
) {
127 dev_priv
->fsb_freq
= 533; /* 133*4 */
130 dev_priv
->fsb_freq
= 800; /* 200*4 */
133 dev_priv
->fsb_freq
= 667; /* 167*4 */
136 dev_priv
->fsb_freq
= 400; /* 100*4 */
140 switch (tmp
& CLKCFG_MEM_MASK
) {
142 dev_priv
->mem_freq
= 533;
145 dev_priv
->mem_freq
= 667;
148 dev_priv
->mem_freq
= 800;
152 /* detect pineview DDR3 setting */
153 tmp
= I915_READ(CSHRDDR3CTL
);
154 dev_priv
->is_ddr3
= (tmp
& CSHRDDR3CTL_DDR3
) ? 1 : 0;
157 static void i915_ironlake_get_mem_freq(struct drm_i915_private
*dev_priv
)
161 ddrpll
= I915_READ16(DDRMPLL1
);
162 csipll
= I915_READ16(CSIPLL0
);
164 switch (ddrpll
& 0xff) {
166 dev_priv
->mem_freq
= 800;
169 dev_priv
->mem_freq
= 1066;
172 dev_priv
->mem_freq
= 1333;
175 dev_priv
->mem_freq
= 1600;
178 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
180 dev_priv
->mem_freq
= 0;
184 dev_priv
->ips
.r_t
= dev_priv
->mem_freq
;
186 switch (csipll
& 0x3ff) {
188 dev_priv
->fsb_freq
= 3200;
191 dev_priv
->fsb_freq
= 3733;
194 dev_priv
->fsb_freq
= 4266;
197 dev_priv
->fsb_freq
= 4800;
200 dev_priv
->fsb_freq
= 5333;
203 dev_priv
->fsb_freq
= 5866;
206 dev_priv
->fsb_freq
= 6400;
209 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
211 dev_priv
->fsb_freq
= 0;
215 if (dev_priv
->fsb_freq
== 3200) {
216 dev_priv
->ips
.c_m
= 0;
217 } else if (dev_priv
->fsb_freq
> 3200 && dev_priv
->fsb_freq
<= 4800) {
218 dev_priv
->ips
.c_m
= 1;
220 dev_priv
->ips
.c_m
= 2;
224 static const struct cxsr_latency cxsr_latency_table
[] = {
225 {1, 0, 800, 400, 3382, 33382, 3983, 33983}, /* DDR2-400 SC */
226 {1, 0, 800, 667, 3354, 33354, 3807, 33807}, /* DDR2-667 SC */
227 {1, 0, 800, 800, 3347, 33347, 3763, 33763}, /* DDR2-800 SC */
228 {1, 1, 800, 667, 6420, 36420, 6873, 36873}, /* DDR3-667 SC */
229 {1, 1, 800, 800, 5902, 35902, 6318, 36318}, /* DDR3-800 SC */
231 {1, 0, 667, 400, 3400, 33400, 4021, 34021}, /* DDR2-400 SC */
232 {1, 0, 667, 667, 3372, 33372, 3845, 33845}, /* DDR2-667 SC */
233 {1, 0, 667, 800, 3386, 33386, 3822, 33822}, /* DDR2-800 SC */
234 {1, 1, 667, 667, 6438, 36438, 6911, 36911}, /* DDR3-667 SC */
235 {1, 1, 667, 800, 5941, 35941, 6377, 36377}, /* DDR3-800 SC */
237 {1, 0, 400, 400, 3472, 33472, 4173, 34173}, /* DDR2-400 SC */
238 {1, 0, 400, 667, 3443, 33443, 3996, 33996}, /* DDR2-667 SC */
239 {1, 0, 400, 800, 3430, 33430, 3946, 33946}, /* DDR2-800 SC */
240 {1, 1, 400, 667, 6509, 36509, 7062, 37062}, /* DDR3-667 SC */
241 {1, 1, 400, 800, 5985, 35985, 6501, 36501}, /* DDR3-800 SC */
243 {0, 0, 800, 400, 3438, 33438, 4065, 34065}, /* DDR2-400 SC */
244 {0, 0, 800, 667, 3410, 33410, 3889, 33889}, /* DDR2-667 SC */
245 {0, 0, 800, 800, 3403, 33403, 3845, 33845}, /* DDR2-800 SC */
246 {0, 1, 800, 667, 6476, 36476, 6955, 36955}, /* DDR3-667 SC */
247 {0, 1, 800, 800, 5958, 35958, 6400, 36400}, /* DDR3-800 SC */
249 {0, 0, 667, 400, 3456, 33456, 4103, 34106}, /* DDR2-400 SC */
250 {0, 0, 667, 667, 3428, 33428, 3927, 33927}, /* DDR2-667 SC */
251 {0, 0, 667, 800, 3443, 33443, 3905, 33905}, /* DDR2-800 SC */
252 {0, 1, 667, 667, 6494, 36494, 6993, 36993}, /* DDR3-667 SC */
253 {0, 1, 667, 800, 5998, 35998, 6460, 36460}, /* DDR3-800 SC */
255 {0, 0, 400, 400, 3528, 33528, 4255, 34255}, /* DDR2-400 SC */
256 {0, 0, 400, 667, 3500, 33500, 4079, 34079}, /* DDR2-667 SC */
257 {0, 0, 400, 800, 3487, 33487, 4029, 34029}, /* DDR2-800 SC */
258 {0, 1, 400, 667, 6566, 36566, 7145, 37145}, /* DDR3-667 SC */
259 {0, 1, 400, 800, 6042, 36042, 6584, 36584}, /* DDR3-800 SC */
262 static const struct cxsr_latency
*intel_get_cxsr_latency(bool is_desktop
,
267 const struct cxsr_latency
*latency
;
270 if (fsb
== 0 || mem
== 0)
273 for (i
= 0; i
< ARRAY_SIZE(cxsr_latency_table
); i
++) {
274 latency
= &cxsr_latency_table
[i
];
275 if (is_desktop
== latency
->is_desktop
&&
276 is_ddr3
== latency
->is_ddr3
&&
277 fsb
== latency
->fsb_freq
&& mem
== latency
->mem_freq
)
281 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
286 static void chv_set_memory_dvfs(struct drm_i915_private
*dev_priv
, bool enable
)
290 mutex_lock(&dev_priv
->rps
.hw_lock
);
292 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
);
294 val
&= ~FORCE_DDR_HIGH_FREQ
;
296 val
|= FORCE_DDR_HIGH_FREQ
;
297 val
&= ~FORCE_DDR_LOW_FREQ
;
298 val
|= FORCE_DDR_FREQ_REQ_ACK
;
299 vlv_punit_write(dev_priv
, PUNIT_REG_DDR_SETUP2
, val
);
301 if (wait_for((vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
) &
302 FORCE_DDR_FREQ_REQ_ACK
) == 0, 3))
303 DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
305 mutex_unlock(&dev_priv
->rps
.hw_lock
);
308 static void chv_set_memory_pm5(struct drm_i915_private
*dev_priv
, bool enable
)
312 mutex_lock(&dev_priv
->rps
.hw_lock
);
314 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DSPFREQ
);
316 val
|= DSP_MAXFIFO_PM5_ENABLE
;
318 val
&= ~DSP_MAXFIFO_PM5_ENABLE
;
319 vlv_punit_write(dev_priv
, PUNIT_REG_DSPFREQ
, val
);
321 mutex_unlock(&dev_priv
->rps
.hw_lock
);
324 #define FW_WM(value, plane) \
325 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
327 static bool _intel_set_memory_cxsr(struct drm_i915_private
*dev_priv
, bool enable
)
332 if (IS_VALLEYVIEW(dev_priv
) || IS_CHERRYVIEW(dev_priv
)) {
333 was_enabled
= I915_READ(FW_BLC_SELF_VLV
) & FW_CSPWRDWNEN
;
334 I915_WRITE(FW_BLC_SELF_VLV
, enable
? FW_CSPWRDWNEN
: 0);
335 POSTING_READ(FW_BLC_SELF_VLV
);
336 } else if (IS_G4X(dev_priv
) || IS_I965GM(dev_priv
)) {
337 was_enabled
= I915_READ(FW_BLC_SELF
) & FW_BLC_SELF_EN
;
338 I915_WRITE(FW_BLC_SELF
, enable
? FW_BLC_SELF_EN
: 0);
339 POSTING_READ(FW_BLC_SELF
);
340 } else if (IS_PINEVIEW(dev_priv
)) {
341 val
= I915_READ(DSPFW3
);
342 was_enabled
= val
& PINEVIEW_SELF_REFRESH_EN
;
344 val
|= PINEVIEW_SELF_REFRESH_EN
;
346 val
&= ~PINEVIEW_SELF_REFRESH_EN
;
347 I915_WRITE(DSPFW3
, val
);
348 POSTING_READ(DSPFW3
);
349 } else if (IS_I945G(dev_priv
) || IS_I945GM(dev_priv
)) {
350 was_enabled
= I915_READ(FW_BLC_SELF
) & FW_BLC_SELF_EN
;
351 val
= enable
? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN
) :
352 _MASKED_BIT_DISABLE(FW_BLC_SELF_EN
);
353 I915_WRITE(FW_BLC_SELF
, val
);
354 POSTING_READ(FW_BLC_SELF
);
355 } else if (IS_I915GM(dev_priv
)) {
357 * FIXME can't find a bit like this for 915G, and
358 * and yet it does have the related watermark in
359 * FW_BLC_SELF. What's going on?
361 was_enabled
= I915_READ(INSTPM
) & INSTPM_SELF_EN
;
362 val
= enable
? _MASKED_BIT_ENABLE(INSTPM_SELF_EN
) :
363 _MASKED_BIT_DISABLE(INSTPM_SELF_EN
);
364 I915_WRITE(INSTPM
, val
);
365 POSTING_READ(INSTPM
);
370 DRM_DEBUG_KMS("memory self-refresh is %s (was %s)\n",
371 enableddisabled(enable
),
372 enableddisabled(was_enabled
));
377 bool intel_set_memory_cxsr(struct drm_i915_private
*dev_priv
, bool enable
)
381 mutex_lock(&dev_priv
->wm
.wm_mutex
);
382 ret
= _intel_set_memory_cxsr(dev_priv
, enable
);
383 dev_priv
->wm
.vlv
.cxsr
= enable
;
384 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
390 * Latency for FIFO fetches is dependent on several factors:
391 * - memory configuration (speed, channels)
393 * - current MCH state
394 * It can be fairly high in some situations, so here we assume a fairly
395 * pessimal value. It's a tradeoff between extra memory fetches (if we
396 * set this value too high, the FIFO will fetch frequently to stay full)
397 * and power consumption (set it too low to save power and we might see
398 * FIFO underruns and display "flicker").
400 * A value of 5us seems to be a good balance; safe for very low end
401 * platforms but not overly aggressive on lower latency configs.
403 static const int pessimal_latency_ns
= 5000;
405 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
406 ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
408 static int vlv_get_fifo_size(struct intel_plane
*plane
)
410 struct drm_i915_private
*dev_priv
= to_i915(plane
->base
.dev
);
411 int sprite0_start
, sprite1_start
, size
;
413 if (plane
->id
== PLANE_CURSOR
)
416 switch (plane
->pipe
) {
417 uint32_t dsparb
, dsparb2
, dsparb3
;
419 dsparb
= I915_READ(DSPARB
);
420 dsparb2
= I915_READ(DSPARB2
);
421 sprite0_start
= VLV_FIFO_START(dsparb
, dsparb2
, 0, 0);
422 sprite1_start
= VLV_FIFO_START(dsparb
, dsparb2
, 8, 4);
425 dsparb
= I915_READ(DSPARB
);
426 dsparb2
= I915_READ(DSPARB2
);
427 sprite0_start
= VLV_FIFO_START(dsparb
, dsparb2
, 16, 8);
428 sprite1_start
= VLV_FIFO_START(dsparb
, dsparb2
, 24, 12);
431 dsparb2
= I915_READ(DSPARB2
);
432 dsparb3
= I915_READ(DSPARB3
);
433 sprite0_start
= VLV_FIFO_START(dsparb3
, dsparb2
, 0, 16);
434 sprite1_start
= VLV_FIFO_START(dsparb3
, dsparb2
, 8, 20);
442 size
= sprite0_start
;
445 size
= sprite1_start
- sprite0_start
;
448 size
= 512 - 1 - sprite1_start
;
454 DRM_DEBUG_KMS("%s FIFO size: %d\n", plane
->base
.name
, size
);
459 static int i9xx_get_fifo_size(struct drm_i915_private
*dev_priv
, int plane
)
461 uint32_t dsparb
= I915_READ(DSPARB
);
464 size
= dsparb
& 0x7f;
466 size
= ((dsparb
>> DSPARB_CSTART_SHIFT
) & 0x7f) - size
;
468 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb
,
469 plane
? "B" : "A", size
);
474 static int i830_get_fifo_size(struct drm_i915_private
*dev_priv
, int plane
)
476 uint32_t dsparb
= I915_READ(DSPARB
);
479 size
= dsparb
& 0x1ff;
481 size
= ((dsparb
>> DSPARB_BEND_SHIFT
) & 0x1ff) - size
;
482 size
>>= 1; /* Convert to cachelines */
484 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb
,
485 plane
? "B" : "A", size
);
490 static int i845_get_fifo_size(struct drm_i915_private
*dev_priv
, int plane
)
492 uint32_t dsparb
= I915_READ(DSPARB
);
495 size
= dsparb
& 0x7f;
496 size
>>= 2; /* Convert to cachelines */
498 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb
,
505 /* Pineview has different values for various configs */
506 static const struct intel_watermark_params pineview_display_wm
= {
507 .fifo_size
= PINEVIEW_DISPLAY_FIFO
,
508 .max_wm
= PINEVIEW_MAX_WM
,
509 .default_wm
= PINEVIEW_DFT_WM
,
510 .guard_size
= PINEVIEW_GUARD_WM
,
511 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
513 static const struct intel_watermark_params pineview_display_hplloff_wm
= {
514 .fifo_size
= PINEVIEW_DISPLAY_FIFO
,
515 .max_wm
= PINEVIEW_MAX_WM
,
516 .default_wm
= PINEVIEW_DFT_HPLLOFF_WM
,
517 .guard_size
= PINEVIEW_GUARD_WM
,
518 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
520 static const struct intel_watermark_params pineview_cursor_wm
= {
521 .fifo_size
= PINEVIEW_CURSOR_FIFO
,
522 .max_wm
= PINEVIEW_CURSOR_MAX_WM
,
523 .default_wm
= PINEVIEW_CURSOR_DFT_WM
,
524 .guard_size
= PINEVIEW_CURSOR_GUARD_WM
,
525 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
527 static const struct intel_watermark_params pineview_cursor_hplloff_wm
= {
528 .fifo_size
= PINEVIEW_CURSOR_FIFO
,
529 .max_wm
= PINEVIEW_CURSOR_MAX_WM
,
530 .default_wm
= PINEVIEW_CURSOR_DFT_WM
,
531 .guard_size
= PINEVIEW_CURSOR_GUARD_WM
,
532 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
534 static const struct intel_watermark_params g4x_wm_info
= {
535 .fifo_size
= G4X_FIFO_SIZE
,
536 .max_wm
= G4X_MAX_WM
,
537 .default_wm
= G4X_MAX_WM
,
539 .cacheline_size
= G4X_FIFO_LINE_SIZE
,
541 static const struct intel_watermark_params g4x_cursor_wm_info
= {
542 .fifo_size
= I965_CURSOR_FIFO
,
543 .max_wm
= I965_CURSOR_MAX_WM
,
544 .default_wm
= I965_CURSOR_DFT_WM
,
546 .cacheline_size
= G4X_FIFO_LINE_SIZE
,
548 static const struct intel_watermark_params i965_cursor_wm_info
= {
549 .fifo_size
= I965_CURSOR_FIFO
,
550 .max_wm
= I965_CURSOR_MAX_WM
,
551 .default_wm
= I965_CURSOR_DFT_WM
,
553 .cacheline_size
= I915_FIFO_LINE_SIZE
,
555 static const struct intel_watermark_params i945_wm_info
= {
556 .fifo_size
= I945_FIFO_SIZE
,
557 .max_wm
= I915_MAX_WM
,
560 .cacheline_size
= I915_FIFO_LINE_SIZE
,
562 static const struct intel_watermark_params i915_wm_info
= {
563 .fifo_size
= I915_FIFO_SIZE
,
564 .max_wm
= I915_MAX_WM
,
567 .cacheline_size
= I915_FIFO_LINE_SIZE
,
569 static const struct intel_watermark_params i830_a_wm_info
= {
570 .fifo_size
= I855GM_FIFO_SIZE
,
571 .max_wm
= I915_MAX_WM
,
574 .cacheline_size
= I830_FIFO_LINE_SIZE
,
576 static const struct intel_watermark_params i830_bc_wm_info
= {
577 .fifo_size
= I855GM_FIFO_SIZE
,
578 .max_wm
= I915_MAX_WM
/2,
581 .cacheline_size
= I830_FIFO_LINE_SIZE
,
583 static const struct intel_watermark_params i845_wm_info
= {
584 .fifo_size
= I830_FIFO_SIZE
,
585 .max_wm
= I915_MAX_WM
,
588 .cacheline_size
= I830_FIFO_LINE_SIZE
,
592 * intel_calculate_wm - calculate watermark level
593 * @clock_in_khz: pixel clock
594 * @wm: chip FIFO params
595 * @cpp: bytes per pixel
596 * @latency_ns: memory latency for the platform
598 * Calculate the watermark level (the level at which the display plane will
599 * start fetching from memory again). Each chip has a different display
600 * FIFO size and allocation, so the caller needs to figure that out and pass
601 * in the correct intel_watermark_params structure.
603 * As the pixel clock runs, the FIFO will be drained at a rate that depends
604 * on the pixel size. When it reaches the watermark level, it'll start
605 * fetching FIFO line sized based chunks from memory until the FIFO fills
606 * past the watermark point. If the FIFO drains completely, a FIFO underrun
607 * will occur, and a display engine hang could result.
609 static unsigned long intel_calculate_wm(unsigned long clock_in_khz
,
610 const struct intel_watermark_params
*wm
,
611 int fifo_size
, int cpp
,
612 unsigned long latency_ns
)
614 long entries_required
, wm_size
;
617 * Note: we need to make sure we don't overflow for various clock &
619 * clocks go from a few thousand to several hundred thousand.
620 * latency is usually a few thousand
622 entries_required
= ((clock_in_khz
/ 1000) * cpp
* latency_ns
) /
624 entries_required
= DIV_ROUND_UP(entries_required
, wm
->cacheline_size
);
626 DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required
);
628 wm_size
= fifo_size
- (entries_required
+ wm
->guard_size
);
630 DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size
);
632 /* Don't promote wm_size to unsigned... */
633 if (wm_size
> (long)wm
->max_wm
)
634 wm_size
= wm
->max_wm
;
636 wm_size
= wm
->default_wm
;
639 * Bspec seems to indicate that the value shouldn't be lower than
640 * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
641 * Lets go for 8 which is the burst size since certain platforms
642 * already use a hardcoded 8 (which is what the spec says should be
651 static struct intel_crtc
*single_enabled_crtc(struct drm_i915_private
*dev_priv
)
653 struct intel_crtc
*crtc
, *enabled
= NULL
;
655 for_each_intel_crtc(&dev_priv
->drm
, crtc
) {
656 if (intel_crtc_active(crtc
)) {
666 static void pineview_update_wm(struct intel_crtc
*unused_crtc
)
668 struct drm_i915_private
*dev_priv
= to_i915(unused_crtc
->base
.dev
);
669 struct intel_crtc
*crtc
;
670 const struct cxsr_latency
*latency
;
674 latency
= intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv
),
679 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
680 intel_set_memory_cxsr(dev_priv
, false);
684 crtc
= single_enabled_crtc(dev_priv
);
686 const struct drm_display_mode
*adjusted_mode
=
687 &crtc
->config
->base
.adjusted_mode
;
688 const struct drm_framebuffer
*fb
=
689 crtc
->base
.primary
->state
->fb
;
690 int cpp
= fb
->format
->cpp
[0];
691 int clock
= adjusted_mode
->crtc_clock
;
694 wm
= intel_calculate_wm(clock
, &pineview_display_wm
,
695 pineview_display_wm
.fifo_size
,
696 cpp
, latency
->display_sr
);
697 reg
= I915_READ(DSPFW1
);
698 reg
&= ~DSPFW_SR_MASK
;
699 reg
|= FW_WM(wm
, SR
);
700 I915_WRITE(DSPFW1
, reg
);
701 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg
);
704 wm
= intel_calculate_wm(clock
, &pineview_cursor_wm
,
705 pineview_display_wm
.fifo_size
,
706 cpp
, latency
->cursor_sr
);
707 reg
= I915_READ(DSPFW3
);
708 reg
&= ~DSPFW_CURSOR_SR_MASK
;
709 reg
|= FW_WM(wm
, CURSOR_SR
);
710 I915_WRITE(DSPFW3
, reg
);
712 /* Display HPLL off SR */
713 wm
= intel_calculate_wm(clock
, &pineview_display_hplloff_wm
,
714 pineview_display_hplloff_wm
.fifo_size
,
715 cpp
, latency
->display_hpll_disable
);
716 reg
= I915_READ(DSPFW3
);
717 reg
&= ~DSPFW_HPLL_SR_MASK
;
718 reg
|= FW_WM(wm
, HPLL_SR
);
719 I915_WRITE(DSPFW3
, reg
);
721 /* cursor HPLL off SR */
722 wm
= intel_calculate_wm(clock
, &pineview_cursor_hplloff_wm
,
723 pineview_display_hplloff_wm
.fifo_size
,
724 cpp
, latency
->cursor_hpll_disable
);
725 reg
= I915_READ(DSPFW3
);
726 reg
&= ~DSPFW_HPLL_CURSOR_MASK
;
727 reg
|= FW_WM(wm
, HPLL_CURSOR
);
728 I915_WRITE(DSPFW3
, reg
);
729 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg
);
731 intel_set_memory_cxsr(dev_priv
, true);
733 intel_set_memory_cxsr(dev_priv
, false);
737 static bool g4x_compute_wm0(struct drm_i915_private
*dev_priv
,
739 const struct intel_watermark_params
*display
,
740 int display_latency_ns
,
741 const struct intel_watermark_params
*cursor
,
742 int cursor_latency_ns
,
746 struct intel_crtc
*crtc
;
747 const struct drm_display_mode
*adjusted_mode
;
748 const struct drm_framebuffer
*fb
;
749 int htotal
, hdisplay
, clock
, cpp
;
750 int line_time_us
, line_count
;
751 int entries
, tlb_miss
;
753 crtc
= intel_get_crtc_for_plane(dev_priv
, plane
);
754 if (!intel_crtc_active(crtc
)) {
755 *cursor_wm
= cursor
->guard_size
;
756 *plane_wm
= display
->guard_size
;
760 adjusted_mode
= &crtc
->config
->base
.adjusted_mode
;
761 fb
= crtc
->base
.primary
->state
->fb
;
762 clock
= adjusted_mode
->crtc_clock
;
763 htotal
= adjusted_mode
->crtc_htotal
;
764 hdisplay
= crtc
->config
->pipe_src_w
;
765 cpp
= fb
->format
->cpp
[0];
767 /* Use the small buffer method to calculate plane watermark */
768 entries
= ((clock
* cpp
/ 1000) * display_latency_ns
) / 1000;
769 tlb_miss
= display
->fifo_size
*display
->cacheline_size
- hdisplay
* 8;
772 entries
= DIV_ROUND_UP(entries
, display
->cacheline_size
);
773 *plane_wm
= entries
+ display
->guard_size
;
774 if (*plane_wm
> (int)display
->max_wm
)
775 *plane_wm
= display
->max_wm
;
777 /* Use the large buffer method to calculate cursor watermark */
778 line_time_us
= max(htotal
* 1000 / clock
, 1);
779 line_count
= (cursor_latency_ns
/ line_time_us
+ 1000) / 1000;
780 entries
= line_count
* crtc
->base
.cursor
->state
->crtc_w
* cpp
;
781 tlb_miss
= cursor
->fifo_size
*cursor
->cacheline_size
- hdisplay
* 8;
784 entries
= DIV_ROUND_UP(entries
, cursor
->cacheline_size
);
785 *cursor_wm
= entries
+ cursor
->guard_size
;
786 if (*cursor_wm
> (int)cursor
->max_wm
)
787 *cursor_wm
= (int)cursor
->max_wm
;
793 * Check the wm result.
795 * If any calculated watermark values is larger than the maximum value that
796 * can be programmed into the associated watermark register, that watermark
799 static bool g4x_check_srwm(struct drm_i915_private
*dev_priv
,
800 int display_wm
, int cursor_wm
,
801 const struct intel_watermark_params
*display
,
802 const struct intel_watermark_params
*cursor
)
804 DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
805 display_wm
, cursor_wm
);
807 if (display_wm
> display
->max_wm
) {
808 DRM_DEBUG_KMS("display watermark is too large(%d/%u), disabling\n",
809 display_wm
, display
->max_wm
);
813 if (cursor_wm
> cursor
->max_wm
) {
814 DRM_DEBUG_KMS("cursor watermark is too large(%d/%u), disabling\n",
815 cursor_wm
, cursor
->max_wm
);
819 if (!(display_wm
|| cursor_wm
)) {
820 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
827 static bool g4x_compute_srwm(struct drm_i915_private
*dev_priv
,
830 const struct intel_watermark_params
*display
,
831 const struct intel_watermark_params
*cursor
,
832 int *display_wm
, int *cursor_wm
)
834 struct intel_crtc
*crtc
;
835 const struct drm_display_mode
*adjusted_mode
;
836 const struct drm_framebuffer
*fb
;
837 int hdisplay
, htotal
, cpp
, clock
;
838 unsigned long line_time_us
;
839 int line_count
, line_size
;
844 *display_wm
= *cursor_wm
= 0;
848 crtc
= intel_get_crtc_for_plane(dev_priv
, plane
);
849 adjusted_mode
= &crtc
->config
->base
.adjusted_mode
;
850 fb
= crtc
->base
.primary
->state
->fb
;
851 clock
= adjusted_mode
->crtc_clock
;
852 htotal
= adjusted_mode
->crtc_htotal
;
853 hdisplay
= crtc
->config
->pipe_src_w
;
854 cpp
= fb
->format
->cpp
[0];
856 line_time_us
= max(htotal
* 1000 / clock
, 1);
857 line_count
= (latency_ns
/ line_time_us
+ 1000) / 1000;
858 line_size
= hdisplay
* cpp
;
860 /* Use the minimum of the small and large buffer method for primary */
861 small
= ((clock
* cpp
/ 1000) * latency_ns
) / 1000;
862 large
= line_count
* line_size
;
864 entries
= DIV_ROUND_UP(min(small
, large
), display
->cacheline_size
);
865 *display_wm
= entries
+ display
->guard_size
;
867 /* calculate the self-refresh watermark for display cursor */
868 entries
= line_count
* cpp
* crtc
->base
.cursor
->state
->crtc_w
;
869 entries
= DIV_ROUND_UP(entries
, cursor
->cacheline_size
);
870 *cursor_wm
= entries
+ cursor
->guard_size
;
872 return g4x_check_srwm(dev_priv
,
873 *display_wm
, *cursor_wm
,
877 #define FW_WM_VLV(value, plane) \
878 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
880 static void vlv_write_wm_values(struct drm_i915_private
*dev_priv
,
881 const struct vlv_wm_values
*wm
)
885 for_each_pipe(dev_priv
, pipe
) {
886 I915_WRITE(VLV_DDL(pipe
),
887 (wm
->ddl
[pipe
].plane
[PLANE_CURSOR
] << DDL_CURSOR_SHIFT
) |
888 (wm
->ddl
[pipe
].plane
[PLANE_SPRITE1
] << DDL_SPRITE_SHIFT(1)) |
889 (wm
->ddl
[pipe
].plane
[PLANE_SPRITE0
] << DDL_SPRITE_SHIFT(0)) |
890 (wm
->ddl
[pipe
].plane
[PLANE_PRIMARY
] << DDL_PLANE_SHIFT
));
894 * Zero the (unused) WM1 watermarks, and also clear all the
895 * high order bits so that there are no out of bounds values
896 * present in the registers during the reprogramming.
898 I915_WRITE(DSPHOWM
, 0);
899 I915_WRITE(DSPHOWM1
, 0);
900 I915_WRITE(DSPFW4
, 0);
901 I915_WRITE(DSPFW5
, 0);
902 I915_WRITE(DSPFW6
, 0);
905 FW_WM(wm
->sr
.plane
, SR
) |
906 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_CURSOR
], CURSORB
) |
907 FW_WM_VLV(wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
], PLANEB
) |
908 FW_WM_VLV(wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
], PLANEA
));
910 FW_WM_VLV(wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE1
], SPRITEB
) |
911 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_CURSOR
], CURSORA
) |
912 FW_WM_VLV(wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
], SPRITEA
));
914 FW_WM(wm
->sr
.cursor
, CURSOR_SR
));
916 if (IS_CHERRYVIEW(dev_priv
)) {
917 I915_WRITE(DSPFW7_CHV
,
918 FW_WM_VLV(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
], SPRITED
) |
919 FW_WM_VLV(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
], SPRITEC
));
920 I915_WRITE(DSPFW8_CHV
,
921 FW_WM_VLV(wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE1
], SPRITEF
) |
922 FW_WM_VLV(wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE0
], SPRITEE
));
923 I915_WRITE(DSPFW9_CHV
,
924 FW_WM_VLV(wm
->pipe
[PIPE_C
].plane
[PLANE_PRIMARY
], PLANEC
) |
925 FW_WM(wm
->pipe
[PIPE_C
].plane
[PLANE_CURSOR
], CURSORC
));
927 FW_WM(wm
->sr
.plane
>> 9, SR_HI
) |
928 FW_WM(wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE1
] >> 8, SPRITEF_HI
) |
929 FW_WM(wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE0
] >> 8, SPRITEE_HI
) |
930 FW_WM(wm
->pipe
[PIPE_C
].plane
[PLANE_PRIMARY
] >> 8, PLANEC_HI
) |
931 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
] >> 8, SPRITED_HI
) |
932 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
] >> 8, SPRITEC_HI
) |
933 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
] >> 8, PLANEB_HI
) |
934 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE1
] >> 8, SPRITEB_HI
) |
935 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
] >> 8, SPRITEA_HI
) |
936 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
] >> 8, PLANEA_HI
));
939 FW_WM_VLV(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
], SPRITED
) |
940 FW_WM_VLV(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
], SPRITEC
));
942 FW_WM(wm
->sr
.plane
>> 9, SR_HI
) |
943 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
] >> 8, SPRITED_HI
) |
944 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
] >> 8, SPRITEC_HI
) |
945 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
] >> 8, PLANEB_HI
) |
946 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE1
] >> 8, SPRITEB_HI
) |
947 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
] >> 8, SPRITEA_HI
) |
948 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
] >> 8, PLANEA_HI
));
951 POSTING_READ(DSPFW1
);
959 VLV_WM_LEVEL_DDR_DVFS
,
962 /* latency must be in 0.1us units. */
963 static unsigned int vlv_wm_method2(unsigned int pixel_rate
,
964 unsigned int pipe_htotal
,
965 unsigned int horiz_pixels
,
967 unsigned int latency
)
971 ret
= (latency
* pixel_rate
) / (pipe_htotal
* 10000);
972 ret
= (ret
+ 1) * horiz_pixels
* cpp
;
973 ret
= DIV_ROUND_UP(ret
, 64);
978 static void vlv_setup_wm_latency(struct drm_i915_private
*dev_priv
)
980 /* all latencies in usec */
981 dev_priv
->wm
.pri_latency
[VLV_WM_LEVEL_PM2
] = 3;
983 dev_priv
->wm
.max_level
= VLV_WM_LEVEL_PM2
;
985 if (IS_CHERRYVIEW(dev_priv
)) {
986 dev_priv
->wm
.pri_latency
[VLV_WM_LEVEL_PM5
] = 12;
987 dev_priv
->wm
.pri_latency
[VLV_WM_LEVEL_DDR_DVFS
] = 33;
989 dev_priv
->wm
.max_level
= VLV_WM_LEVEL_DDR_DVFS
;
993 static uint16_t vlv_compute_wm_level(const struct intel_crtc_state
*crtc_state
,
994 const struct intel_plane_state
*plane_state
,
997 struct intel_plane
*plane
= to_intel_plane(plane_state
->base
.plane
);
998 struct drm_i915_private
*dev_priv
= to_i915(plane
->base
.dev
);
999 const struct drm_display_mode
*adjusted_mode
=
1000 &crtc_state
->base
.adjusted_mode
;
1001 int clock
, htotal
, cpp
, width
, wm
;
1003 if (dev_priv
->wm
.pri_latency
[level
] == 0)
1006 if (!plane_state
->base
.visible
)
1009 cpp
= plane_state
->base
.fb
->format
->cpp
[0];
1010 clock
= adjusted_mode
->crtc_clock
;
1011 htotal
= adjusted_mode
->crtc_htotal
;
1012 width
= crtc_state
->pipe_src_w
;
1013 if (WARN_ON(htotal
== 0))
1016 if (plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
) {
1018 * FIXME the formula gives values that are
1019 * too big for the cursor FIFO, and hence we
1020 * would never be able to use cursors. For
1021 * now just hardcode the watermark.
1025 wm
= vlv_wm_method2(clock
, htotal
, width
, cpp
,
1026 dev_priv
->wm
.pri_latency
[level
] * 10);
1029 return min_t(int, wm
, USHRT_MAX
);
1032 static void vlv_compute_fifo(struct intel_crtc
*crtc
)
1034 struct drm_device
*dev
= crtc
->base
.dev
;
1035 struct vlv_wm_state
*wm_state
= &crtc
->wm_state
;
1036 struct intel_plane
*plane
;
1037 unsigned int total_rate
= 0;
1038 const int fifo_size
= 512 - 1;
1039 int fifo_extra
, fifo_left
= fifo_size
;
1041 for_each_intel_plane_on_crtc(dev
, crtc
, plane
) {
1042 struct intel_plane_state
*state
=
1043 to_intel_plane_state(plane
->base
.state
);
1045 if (plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
)
1048 if (state
->base
.visible
) {
1049 wm_state
->num_active_planes
++;
1050 total_rate
+= state
->base
.fb
->format
->cpp
[0];
1054 for_each_intel_plane_on_crtc(dev
, crtc
, plane
) {
1055 struct intel_plane_state
*state
=
1056 to_intel_plane_state(plane
->base
.state
);
1059 if (plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
) {
1060 plane
->wm
.fifo_size
= 63;
1064 if (!state
->base
.visible
) {
1065 plane
->wm
.fifo_size
= 0;
1069 rate
= state
->base
.fb
->format
->cpp
[0];
1070 plane
->wm
.fifo_size
= fifo_size
* rate
/ total_rate
;
1071 fifo_left
-= plane
->wm
.fifo_size
;
1074 fifo_extra
= DIV_ROUND_UP(fifo_left
, wm_state
->num_active_planes
?: 1);
1076 /* spread the remainder evenly */
1077 for_each_intel_plane_on_crtc(dev
, crtc
, plane
) {
1083 if (plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
)
1086 /* give it all to the first plane if none are active */
1087 if (plane
->wm
.fifo_size
== 0 &&
1088 wm_state
->num_active_planes
)
1091 plane_extra
= min(fifo_extra
, fifo_left
);
1092 plane
->wm
.fifo_size
+= plane_extra
;
1093 fifo_left
-= plane_extra
;
1096 WARN_ON(fifo_left
!= 0);
1099 static u16
vlv_invert_wm_value(u16 wm
, u16 fifo_size
)
1104 return fifo_size
- wm
;
1107 static void vlv_invert_wms(struct intel_crtc
*crtc
)
1109 struct vlv_wm_state
*wm_state
= &crtc
->wm_state
;
1112 for (level
= 0; level
< wm_state
->num_levels
; level
++) {
1113 struct drm_i915_private
*dev_priv
= to_i915(crtc
->base
.dev
);
1114 const int sr_fifo_size
=
1115 INTEL_INFO(dev_priv
)->num_pipes
* 512 - 1;
1116 struct intel_plane
*plane
;
1118 wm_state
->sr
[level
].plane
=
1119 vlv_invert_wm_value(wm_state
->sr
[level
].plane
,
1121 wm_state
->sr
[level
].cursor
=
1122 vlv_invert_wm_value(wm_state
->sr
[level
].cursor
,
1125 for_each_intel_plane_on_crtc(&dev_priv
->drm
, crtc
, plane
) {
1126 wm_state
->wm
[level
].plane
[plane
->id
] =
1127 vlv_invert_wm_value(wm_state
->wm
[level
].plane
[plane
->id
],
1128 plane
->wm
.fifo_size
);
1133 static void vlv_compute_wm(struct intel_crtc
*crtc
)
1135 struct drm_i915_private
*dev_priv
= to_i915(crtc
->base
.dev
);
1136 struct vlv_wm_state
*wm_state
= &crtc
->wm_state
;
1137 struct intel_plane
*plane
;
1140 memset(wm_state
, 0, sizeof(*wm_state
));
1142 wm_state
->cxsr
= crtc
->pipe
!= PIPE_C
&& crtc
->wm
.cxsr_allowed
;
1143 wm_state
->num_levels
= dev_priv
->wm
.max_level
+ 1;
1145 wm_state
->num_active_planes
= 0;
1147 vlv_compute_fifo(crtc
);
1149 if (wm_state
->num_active_planes
!= 1)
1150 wm_state
->cxsr
= false;
1152 for_each_intel_plane_on_crtc(&dev_priv
->drm
, crtc
, plane
) {
1153 struct intel_plane_state
*state
=
1154 to_intel_plane_state(plane
->base
.state
);
1157 if (!state
->base
.visible
)
1160 /* normal watermarks */
1161 for (level
= 0; level
< wm_state
->num_levels
; level
++) {
1162 int wm
= vlv_compute_wm_level(crtc
->config
, state
, level
);
1163 int max_wm
= plane
->wm
.fifo_size
;
1166 if (WARN_ON(level
== 0 && wm
> max_wm
))
1172 wm_state
->wm
[level
].plane
[plane
->id
] = wm
;
1175 wm_state
->num_levels
= level
;
1177 if (!wm_state
->cxsr
)
1180 /* maxfifo watermarks */
1181 if (plane
->id
== PLANE_CURSOR
) {
1182 for (level
= 0; level
< wm_state
->num_levels
; level
++)
1183 wm_state
->sr
[level
].cursor
=
1184 wm_state
->wm
[level
].plane
[PLANE_CURSOR
];
1186 for (level
= 0; level
< wm_state
->num_levels
; level
++)
1187 wm_state
->sr
[level
].plane
=
1188 max(wm_state
->sr
[level
].plane
,
1189 wm_state
->wm
[level
].plane
[plane
->id
]);
1193 /* clear any (partially) filled invalid levels */
1194 for (level
= wm_state
->num_levels
; level
< dev_priv
->wm
.max_level
+ 1; level
++) {
1195 memset(&wm_state
->wm
[level
], 0, sizeof(wm_state
->wm
[level
]));
1196 memset(&wm_state
->sr
[level
], 0, sizeof(wm_state
->sr
[level
]));
1199 vlv_invert_wms(crtc
);
1202 #define VLV_FIFO(plane, value) \
1203 (((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1205 static void vlv_pipe_set_fifo_size(struct intel_crtc
*crtc
)
1207 struct drm_device
*dev
= crtc
->base
.dev
;
1208 struct drm_i915_private
*dev_priv
= to_i915(dev
);
1209 struct intel_plane
*plane
;
1210 int sprite0_start
= 0, sprite1_start
= 0, fifo_size
= 0;
1212 for_each_intel_plane_on_crtc(dev
, crtc
, plane
) {
1213 switch (plane
->id
) {
1215 sprite0_start
= plane
->wm
.fifo_size
;
1218 sprite1_start
= sprite0_start
+ plane
->wm
.fifo_size
;
1221 fifo_size
= sprite1_start
+ plane
->wm
.fifo_size
;
1224 WARN_ON(plane
->wm
.fifo_size
!= 63);
1227 MISSING_CASE(plane
->id
);
1232 WARN_ON(fifo_size
!= 512 - 1);
1234 DRM_DEBUG_KMS("Pipe %c FIFO split %d / %d / %d\n",
1235 pipe_name(crtc
->pipe
), sprite0_start
,
1236 sprite1_start
, fifo_size
);
1238 spin_lock(&dev_priv
->wm
.dsparb_lock
);
1240 switch (crtc
->pipe
) {
1241 uint32_t dsparb
, dsparb2
, dsparb3
;
1243 dsparb
= I915_READ(DSPARB
);
1244 dsparb2
= I915_READ(DSPARB2
);
1246 dsparb
&= ~(VLV_FIFO(SPRITEA
, 0xff) |
1247 VLV_FIFO(SPRITEB
, 0xff));
1248 dsparb
|= (VLV_FIFO(SPRITEA
, sprite0_start
) |
1249 VLV_FIFO(SPRITEB
, sprite1_start
));
1251 dsparb2
&= ~(VLV_FIFO(SPRITEA_HI
, 0x1) |
1252 VLV_FIFO(SPRITEB_HI
, 0x1));
1253 dsparb2
|= (VLV_FIFO(SPRITEA_HI
, sprite0_start
>> 8) |
1254 VLV_FIFO(SPRITEB_HI
, sprite1_start
>> 8));
1256 I915_WRITE(DSPARB
, dsparb
);
1257 I915_WRITE(DSPARB2
, dsparb2
);
1260 dsparb
= I915_READ(DSPARB
);
1261 dsparb2
= I915_READ(DSPARB2
);
1263 dsparb
&= ~(VLV_FIFO(SPRITEC
, 0xff) |
1264 VLV_FIFO(SPRITED
, 0xff));
1265 dsparb
|= (VLV_FIFO(SPRITEC
, sprite0_start
) |
1266 VLV_FIFO(SPRITED
, sprite1_start
));
1268 dsparb2
&= ~(VLV_FIFO(SPRITEC_HI
, 0xff) |
1269 VLV_FIFO(SPRITED_HI
, 0xff));
1270 dsparb2
|= (VLV_FIFO(SPRITEC_HI
, sprite0_start
>> 8) |
1271 VLV_FIFO(SPRITED_HI
, sprite1_start
>> 8));
1273 I915_WRITE(DSPARB
, dsparb
);
1274 I915_WRITE(DSPARB2
, dsparb2
);
1277 dsparb3
= I915_READ(DSPARB3
);
1278 dsparb2
= I915_READ(DSPARB2
);
1280 dsparb3
&= ~(VLV_FIFO(SPRITEE
, 0xff) |
1281 VLV_FIFO(SPRITEF
, 0xff));
1282 dsparb3
|= (VLV_FIFO(SPRITEE
, sprite0_start
) |
1283 VLV_FIFO(SPRITEF
, sprite1_start
));
1285 dsparb2
&= ~(VLV_FIFO(SPRITEE_HI
, 0xff) |
1286 VLV_FIFO(SPRITEF_HI
, 0xff));
1287 dsparb2
|= (VLV_FIFO(SPRITEE_HI
, sprite0_start
>> 8) |
1288 VLV_FIFO(SPRITEF_HI
, sprite1_start
>> 8));
1290 I915_WRITE(DSPARB3
, dsparb3
);
1291 I915_WRITE(DSPARB2
, dsparb2
);
1297 POSTING_READ(DSPARB
);
1299 spin_unlock(&dev_priv
->wm
.dsparb_lock
);
1304 static void vlv_merge_wm(struct drm_i915_private
*dev_priv
,
1305 struct vlv_wm_values
*wm
)
1307 struct intel_crtc
*crtc
;
1308 int num_active_crtcs
= 0;
1310 wm
->level
= dev_priv
->wm
.max_level
;
1313 for_each_intel_crtc(&dev_priv
->drm
, crtc
) {
1314 const struct vlv_wm_state
*wm_state
= &crtc
->wm_state
;
1319 if (!wm_state
->cxsr
)
1323 wm
->level
= min_t(int, wm
->level
, wm_state
->num_levels
- 1);
1326 if (num_active_crtcs
!= 1)
1329 if (num_active_crtcs
> 1)
1330 wm
->level
= VLV_WM_LEVEL_PM2
;
1332 for_each_intel_crtc(&dev_priv
->drm
, crtc
) {
1333 struct vlv_wm_state
*wm_state
= &crtc
->wm_state
;
1334 enum pipe pipe
= crtc
->pipe
;
1339 wm
->pipe
[pipe
] = wm_state
->wm
[wm
->level
];
1341 wm
->sr
= wm_state
->sr
[wm
->level
];
1343 wm
->ddl
[pipe
].plane
[PLANE_PRIMARY
] = DDL_PRECISION_HIGH
| 2;
1344 wm
->ddl
[pipe
].plane
[PLANE_SPRITE0
] = DDL_PRECISION_HIGH
| 2;
1345 wm
->ddl
[pipe
].plane
[PLANE_SPRITE1
] = DDL_PRECISION_HIGH
| 2;
1346 wm
->ddl
[pipe
].plane
[PLANE_CURSOR
] = DDL_PRECISION_HIGH
| 2;
1350 static bool is_disabling(int old
, int new, int threshold
)
1352 return old
>= threshold
&& new < threshold
;
1355 static bool is_enabling(int old
, int new, int threshold
)
1357 return old
< threshold
&& new >= threshold
;
1360 static void vlv_update_wm(struct intel_crtc
*crtc
)
1362 struct drm_i915_private
*dev_priv
= to_i915(crtc
->base
.dev
);
1363 enum pipe pipe
= crtc
->pipe
;
1364 struct vlv_wm_values
*old_wm
= &dev_priv
->wm
.vlv
;
1365 struct vlv_wm_values new_wm
= {};
1367 vlv_compute_wm(crtc
);
1368 vlv_merge_wm(dev_priv
, &new_wm
);
1370 if (memcmp(old_wm
, &new_wm
, sizeof(new_wm
)) == 0) {
1371 /* FIXME should be part of crtc atomic commit */
1372 vlv_pipe_set_fifo_size(crtc
);
1377 if (is_disabling(old_wm
->level
, new_wm
.level
, VLV_WM_LEVEL_DDR_DVFS
))
1378 chv_set_memory_dvfs(dev_priv
, false);
1380 if (is_disabling(old_wm
->level
, new_wm
.level
, VLV_WM_LEVEL_PM5
))
1381 chv_set_memory_pm5(dev_priv
, false);
1383 if (is_disabling(old_wm
->cxsr
, new_wm
.cxsr
, true))
1384 _intel_set_memory_cxsr(dev_priv
, false);
1386 /* FIXME should be part of crtc atomic commit */
1387 vlv_pipe_set_fifo_size(crtc
);
1389 vlv_write_wm_values(dev_priv
, &new_wm
);
1391 DRM_DEBUG_KMS("Setting FIFO watermarks - %c: plane=%d, cursor=%d, "
1392 "sprite0=%d, sprite1=%d, SR: plane=%d, cursor=%d level=%d cxsr=%d\n",
1393 pipe_name(pipe
), new_wm
.pipe
[pipe
].plane
[PLANE_PRIMARY
], new_wm
.pipe
[pipe
].plane
[PLANE_CURSOR
],
1394 new_wm
.pipe
[pipe
].plane
[PLANE_SPRITE0
], new_wm
.pipe
[pipe
].plane
[PLANE_SPRITE1
],
1395 new_wm
.sr
.plane
, new_wm
.sr
.cursor
, new_wm
.level
, new_wm
.cxsr
);
1397 if (is_enabling(old_wm
->cxsr
, new_wm
.cxsr
, true))
1398 _intel_set_memory_cxsr(dev_priv
, true);
1400 if (is_enabling(old_wm
->level
, new_wm
.level
, VLV_WM_LEVEL_PM5
))
1401 chv_set_memory_pm5(dev_priv
, true);
1403 if (is_enabling(old_wm
->level
, new_wm
.level
, VLV_WM_LEVEL_DDR_DVFS
))
1404 chv_set_memory_dvfs(dev_priv
, true);
1409 #define single_plane_enabled(mask) is_power_of_2(mask)
1411 static void g4x_update_wm(struct intel_crtc
*crtc
)
1413 struct drm_i915_private
*dev_priv
= to_i915(crtc
->base
.dev
);
1414 static const int sr_latency_ns
= 12000;
1415 int planea_wm
, planeb_wm
, cursora_wm
, cursorb_wm
;
1416 int plane_sr
, cursor_sr
;
1417 unsigned int enabled
= 0;
1420 if (g4x_compute_wm0(dev_priv
, PIPE_A
,
1421 &g4x_wm_info
, pessimal_latency_ns
,
1422 &g4x_cursor_wm_info
, pessimal_latency_ns
,
1423 &planea_wm
, &cursora_wm
))
1424 enabled
|= 1 << PIPE_A
;
1426 if (g4x_compute_wm0(dev_priv
, PIPE_B
,
1427 &g4x_wm_info
, pessimal_latency_ns
,
1428 &g4x_cursor_wm_info
, pessimal_latency_ns
,
1429 &planeb_wm
, &cursorb_wm
))
1430 enabled
|= 1 << PIPE_B
;
1432 if (single_plane_enabled(enabled
) &&
1433 g4x_compute_srwm(dev_priv
, ffs(enabled
) - 1,
1436 &g4x_cursor_wm_info
,
1437 &plane_sr
, &cursor_sr
)) {
1438 cxsr_enabled
= true;
1440 cxsr_enabled
= false;
1441 intel_set_memory_cxsr(dev_priv
, false);
1442 plane_sr
= cursor_sr
= 0;
1445 DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, "
1446 "B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1447 planea_wm
, cursora_wm
,
1448 planeb_wm
, cursorb_wm
,
1449 plane_sr
, cursor_sr
);
1452 FW_WM(plane_sr
, SR
) |
1453 FW_WM(cursorb_wm
, CURSORB
) |
1454 FW_WM(planeb_wm
, PLANEB
) |
1455 FW_WM(planea_wm
, PLANEA
));
1457 (I915_READ(DSPFW2
) & ~DSPFW_CURSORA_MASK
) |
1458 FW_WM(cursora_wm
, CURSORA
));
1459 /* HPLL off in SR has some issues on G4x... disable it */
1461 (I915_READ(DSPFW3
) & ~(DSPFW_HPLL_SR_EN
| DSPFW_CURSOR_SR_MASK
)) |
1462 FW_WM(cursor_sr
, CURSOR_SR
));
1465 intel_set_memory_cxsr(dev_priv
, true);
1468 static void i965_update_wm(struct intel_crtc
*unused_crtc
)
1470 struct drm_i915_private
*dev_priv
= to_i915(unused_crtc
->base
.dev
);
1471 struct intel_crtc
*crtc
;
1476 /* Calc sr entries for one plane configs */
1477 crtc
= single_enabled_crtc(dev_priv
);
1479 /* self-refresh has much higher latency */
1480 static const int sr_latency_ns
= 12000;
1481 const struct drm_display_mode
*adjusted_mode
=
1482 &crtc
->config
->base
.adjusted_mode
;
1483 const struct drm_framebuffer
*fb
=
1484 crtc
->base
.primary
->state
->fb
;
1485 int clock
= adjusted_mode
->crtc_clock
;
1486 int htotal
= adjusted_mode
->crtc_htotal
;
1487 int hdisplay
= crtc
->config
->pipe_src_w
;
1488 int cpp
= fb
->format
->cpp
[0];
1489 unsigned long line_time_us
;
1492 line_time_us
= max(htotal
* 1000 / clock
, 1);
1494 /* Use ns/us then divide to preserve precision */
1495 entries
= (((sr_latency_ns
/ line_time_us
) + 1000) / 1000) *
1497 entries
= DIV_ROUND_UP(entries
, I915_FIFO_LINE_SIZE
);
1498 srwm
= I965_FIFO_SIZE
- entries
;
1502 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1505 entries
= (((sr_latency_ns
/ line_time_us
) + 1000) / 1000) *
1506 cpp
* crtc
->base
.cursor
->state
->crtc_w
;
1507 entries
= DIV_ROUND_UP(entries
,
1508 i965_cursor_wm_info
.cacheline_size
);
1509 cursor_sr
= i965_cursor_wm_info
.fifo_size
-
1510 (entries
+ i965_cursor_wm_info
.guard_size
);
1512 if (cursor_sr
> i965_cursor_wm_info
.max_wm
)
1513 cursor_sr
= i965_cursor_wm_info
.max_wm
;
1515 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1516 "cursor %d\n", srwm
, cursor_sr
);
1518 cxsr_enabled
= true;
1520 cxsr_enabled
= false;
1521 /* Turn off self refresh if both pipes are enabled */
1522 intel_set_memory_cxsr(dev_priv
, false);
1525 DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1528 /* 965 has limitations... */
1529 I915_WRITE(DSPFW1
, FW_WM(srwm
, SR
) |
1533 I915_WRITE(DSPFW2
, FW_WM(8, CURSORA
) |
1534 FW_WM(8, PLANEC_OLD
));
1535 /* update cursor SR watermark */
1536 I915_WRITE(DSPFW3
, FW_WM(cursor_sr
, CURSOR_SR
));
1539 intel_set_memory_cxsr(dev_priv
, true);
1544 static void i9xx_update_wm(struct intel_crtc
*unused_crtc
)
1546 struct drm_i915_private
*dev_priv
= to_i915(unused_crtc
->base
.dev
);
1547 const struct intel_watermark_params
*wm_info
;
1552 int planea_wm
, planeb_wm
;
1553 struct intel_crtc
*crtc
, *enabled
= NULL
;
1555 if (IS_I945GM(dev_priv
))
1556 wm_info
= &i945_wm_info
;
1557 else if (!IS_GEN2(dev_priv
))
1558 wm_info
= &i915_wm_info
;
1560 wm_info
= &i830_a_wm_info
;
1562 fifo_size
= dev_priv
->display
.get_fifo_size(dev_priv
, 0);
1563 crtc
= intel_get_crtc_for_plane(dev_priv
, 0);
1564 if (intel_crtc_active(crtc
)) {
1565 const struct drm_display_mode
*adjusted_mode
=
1566 &crtc
->config
->base
.adjusted_mode
;
1567 const struct drm_framebuffer
*fb
=
1568 crtc
->base
.primary
->state
->fb
;
1571 if (IS_GEN2(dev_priv
))
1574 cpp
= fb
->format
->cpp
[0];
1576 planea_wm
= intel_calculate_wm(adjusted_mode
->crtc_clock
,
1577 wm_info
, fifo_size
, cpp
,
1578 pessimal_latency_ns
);
1581 planea_wm
= fifo_size
- wm_info
->guard_size
;
1582 if (planea_wm
> (long)wm_info
->max_wm
)
1583 planea_wm
= wm_info
->max_wm
;
1586 if (IS_GEN2(dev_priv
))
1587 wm_info
= &i830_bc_wm_info
;
1589 fifo_size
= dev_priv
->display
.get_fifo_size(dev_priv
, 1);
1590 crtc
= intel_get_crtc_for_plane(dev_priv
, 1);
1591 if (intel_crtc_active(crtc
)) {
1592 const struct drm_display_mode
*adjusted_mode
=
1593 &crtc
->config
->base
.adjusted_mode
;
1594 const struct drm_framebuffer
*fb
=
1595 crtc
->base
.primary
->state
->fb
;
1598 if (IS_GEN2(dev_priv
))
1601 cpp
= fb
->format
->cpp
[0];
1603 planeb_wm
= intel_calculate_wm(adjusted_mode
->crtc_clock
,
1604 wm_info
, fifo_size
, cpp
,
1605 pessimal_latency_ns
);
1606 if (enabled
== NULL
)
1611 planeb_wm
= fifo_size
- wm_info
->guard_size
;
1612 if (planeb_wm
> (long)wm_info
->max_wm
)
1613 planeb_wm
= wm_info
->max_wm
;
1616 DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm
, planeb_wm
);
1618 if (IS_I915GM(dev_priv
) && enabled
) {
1619 struct drm_i915_gem_object
*obj
;
1621 obj
= intel_fb_obj(enabled
->base
.primary
->state
->fb
);
1623 /* self-refresh seems busted with untiled */
1624 if (!i915_gem_object_is_tiled(obj
))
1629 * Overlay gets an aggressive default since video jitter is bad.
1633 /* Play safe and disable self-refresh before adjusting watermarks. */
1634 intel_set_memory_cxsr(dev_priv
, false);
1636 /* Calc sr entries for one plane configs */
1637 if (HAS_FW_BLC(dev_priv
) && enabled
) {
1638 /* self-refresh has much higher latency */
1639 static const int sr_latency_ns
= 6000;
1640 const struct drm_display_mode
*adjusted_mode
=
1641 &enabled
->config
->base
.adjusted_mode
;
1642 const struct drm_framebuffer
*fb
=
1643 enabled
->base
.primary
->state
->fb
;
1644 int clock
= adjusted_mode
->crtc_clock
;
1645 int htotal
= adjusted_mode
->crtc_htotal
;
1646 int hdisplay
= enabled
->config
->pipe_src_w
;
1648 unsigned long line_time_us
;
1651 if (IS_I915GM(dev_priv
) || IS_I945GM(dev_priv
))
1654 cpp
= fb
->format
->cpp
[0];
1656 line_time_us
= max(htotal
* 1000 / clock
, 1);
1658 /* Use ns/us then divide to preserve precision */
1659 entries
= (((sr_latency_ns
/ line_time_us
) + 1000) / 1000) *
1661 entries
= DIV_ROUND_UP(entries
, wm_info
->cacheline_size
);
1662 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries
);
1663 srwm
= wm_info
->fifo_size
- entries
;
1667 if (IS_I945G(dev_priv
) || IS_I945GM(dev_priv
))
1668 I915_WRITE(FW_BLC_SELF
,
1669 FW_BLC_SELF_FIFO_MASK
| (srwm
& 0xff));
1671 I915_WRITE(FW_BLC_SELF
, srwm
& 0x3f);
1674 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1675 planea_wm
, planeb_wm
, cwm
, srwm
);
1677 fwater_lo
= ((planeb_wm
& 0x3f) << 16) | (planea_wm
& 0x3f);
1678 fwater_hi
= (cwm
& 0x1f);
1680 /* Set request length to 8 cachelines per fetch */
1681 fwater_lo
= fwater_lo
| (1 << 24) | (1 << 8);
1682 fwater_hi
= fwater_hi
| (1 << 8);
1684 I915_WRITE(FW_BLC
, fwater_lo
);
1685 I915_WRITE(FW_BLC2
, fwater_hi
);
1688 intel_set_memory_cxsr(dev_priv
, true);
1691 static void i845_update_wm(struct intel_crtc
*unused_crtc
)
1693 struct drm_i915_private
*dev_priv
= to_i915(unused_crtc
->base
.dev
);
1694 struct intel_crtc
*crtc
;
1695 const struct drm_display_mode
*adjusted_mode
;
1699 crtc
= single_enabled_crtc(dev_priv
);
1703 adjusted_mode
= &crtc
->config
->base
.adjusted_mode
;
1704 planea_wm
= intel_calculate_wm(adjusted_mode
->crtc_clock
,
1706 dev_priv
->display
.get_fifo_size(dev_priv
, 0),
1707 4, pessimal_latency_ns
);
1708 fwater_lo
= I915_READ(FW_BLC
) & ~0xfff;
1709 fwater_lo
|= (3<<8) | planea_wm
;
1711 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm
);
1713 I915_WRITE(FW_BLC
, fwater_lo
);
1716 /* latency must be in 0.1us units. */
1717 static uint32_t ilk_wm_method1(uint32_t pixel_rate
, uint8_t cpp
, uint32_t latency
)
1721 if (WARN(latency
== 0, "Latency value missing\n"))
1724 ret
= (uint64_t) pixel_rate
* cpp
* latency
;
1725 ret
= DIV_ROUND_UP_ULL(ret
, 64 * 10000) + 2;
1730 /* latency must be in 0.1us units. */
1731 static uint32_t ilk_wm_method2(uint32_t pixel_rate
, uint32_t pipe_htotal
,
1732 uint32_t horiz_pixels
, uint8_t cpp
,
1737 if (WARN(latency
== 0, "Latency value missing\n"))
1739 if (WARN_ON(!pipe_htotal
))
1742 ret
= (latency
* pixel_rate
) / (pipe_htotal
* 10000);
1743 ret
= (ret
+ 1) * horiz_pixels
* cpp
;
1744 ret
= DIV_ROUND_UP(ret
, 64) + 2;
1748 static uint32_t ilk_wm_fbc(uint32_t pri_val
, uint32_t horiz_pixels
,
1752 * Neither of these should be possible since this function shouldn't be
1753 * called if the CRTC is off or the plane is invisible. But let's be
1754 * extra paranoid to avoid a potential divide-by-zero if we screw up
1755 * elsewhere in the driver.
1759 if (WARN_ON(!horiz_pixels
))
1762 return DIV_ROUND_UP(pri_val
* 64, horiz_pixels
* cpp
) + 2;
1765 struct ilk_wm_maximums
{
1773 * For both WM_PIPE and WM_LP.
1774 * mem_value must be in 0.1us units.
1776 static uint32_t ilk_compute_pri_wm(const struct intel_crtc_state
*cstate
,
1777 const struct intel_plane_state
*pstate
,
1781 uint32_t method1
, method2
;
1784 if (!cstate
->base
.active
|| !pstate
->base
.visible
)
1787 cpp
= pstate
->base
.fb
->format
->cpp
[0];
1789 method1
= ilk_wm_method1(cstate
->pixel_rate
, cpp
, mem_value
);
1794 method2
= ilk_wm_method2(cstate
->pixel_rate
,
1795 cstate
->base
.adjusted_mode
.crtc_htotal
,
1796 drm_rect_width(&pstate
->base
.dst
),
1799 return min(method1
, method2
);
1803 * For both WM_PIPE and WM_LP.
1804 * mem_value must be in 0.1us units.
1806 static uint32_t ilk_compute_spr_wm(const struct intel_crtc_state
*cstate
,
1807 const struct intel_plane_state
*pstate
,
1810 uint32_t method1
, method2
;
1813 if (!cstate
->base
.active
|| !pstate
->base
.visible
)
1816 cpp
= pstate
->base
.fb
->format
->cpp
[0];
1818 method1
= ilk_wm_method1(cstate
->pixel_rate
, cpp
, mem_value
);
1819 method2
= ilk_wm_method2(cstate
->pixel_rate
,
1820 cstate
->base
.adjusted_mode
.crtc_htotal
,
1821 drm_rect_width(&pstate
->base
.dst
),
1823 return min(method1
, method2
);
1827 * For both WM_PIPE and WM_LP.
1828 * mem_value must be in 0.1us units.
1830 static uint32_t ilk_compute_cur_wm(const struct intel_crtc_state
*cstate
,
1831 const struct intel_plane_state
*pstate
,
1835 * We treat the cursor plane as always-on for the purposes of watermark
1836 * calculation. Until we have two-stage watermark programming merged,
1837 * this is necessary to avoid flickering.
1840 int width
= pstate
->base
.visible
? pstate
->base
.crtc_w
: 64;
1842 if (!cstate
->base
.active
)
1845 return ilk_wm_method2(cstate
->pixel_rate
,
1846 cstate
->base
.adjusted_mode
.crtc_htotal
,
1847 width
, cpp
, mem_value
);
1850 /* Only for WM_LP. */
1851 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state
*cstate
,
1852 const struct intel_plane_state
*pstate
,
1857 if (!cstate
->base
.active
|| !pstate
->base
.visible
)
1860 cpp
= pstate
->base
.fb
->format
->cpp
[0];
1862 return ilk_wm_fbc(pri_val
, drm_rect_width(&pstate
->base
.dst
), cpp
);
1866 ilk_display_fifo_size(const struct drm_i915_private
*dev_priv
)
1868 if (INTEL_GEN(dev_priv
) >= 8)
1870 else if (INTEL_GEN(dev_priv
) >= 7)
1877 ilk_plane_wm_reg_max(const struct drm_i915_private
*dev_priv
,
1878 int level
, bool is_sprite
)
1880 if (INTEL_GEN(dev_priv
) >= 8)
1881 /* BDW primary/sprite plane watermarks */
1882 return level
== 0 ? 255 : 2047;
1883 else if (INTEL_GEN(dev_priv
) >= 7)
1884 /* IVB/HSW primary/sprite plane watermarks */
1885 return level
== 0 ? 127 : 1023;
1886 else if (!is_sprite
)
1887 /* ILK/SNB primary plane watermarks */
1888 return level
== 0 ? 127 : 511;
1890 /* ILK/SNB sprite plane watermarks */
1891 return level
== 0 ? 63 : 255;
1895 ilk_cursor_wm_reg_max(const struct drm_i915_private
*dev_priv
, int level
)
1897 if (INTEL_GEN(dev_priv
) >= 7)
1898 return level
== 0 ? 63 : 255;
1900 return level
== 0 ? 31 : 63;
1903 static unsigned int ilk_fbc_wm_reg_max(const struct drm_i915_private
*dev_priv
)
1905 if (INTEL_GEN(dev_priv
) >= 8)
1911 /* Calculate the maximum primary/sprite plane watermark */
1912 static unsigned int ilk_plane_wm_max(const struct drm_device
*dev
,
1914 const struct intel_wm_config
*config
,
1915 enum intel_ddb_partitioning ddb_partitioning
,
1918 struct drm_i915_private
*dev_priv
= to_i915(dev
);
1919 unsigned int fifo_size
= ilk_display_fifo_size(dev_priv
);
1921 /* if sprites aren't enabled, sprites get nothing */
1922 if (is_sprite
&& !config
->sprites_enabled
)
1925 /* HSW allows LP1+ watermarks even with multiple pipes */
1926 if (level
== 0 || config
->num_pipes_active
> 1) {
1927 fifo_size
/= INTEL_INFO(dev_priv
)->num_pipes
;
1930 * For some reason the non self refresh
1931 * FIFO size is only half of the self
1932 * refresh FIFO size on ILK/SNB.
1934 if (INTEL_GEN(dev_priv
) <= 6)
1938 if (config
->sprites_enabled
) {
1939 /* level 0 is always calculated with 1:1 split */
1940 if (level
> 0 && ddb_partitioning
== INTEL_DDB_PART_5_6
) {
1949 /* clamp to max that the registers can hold */
1950 return min(fifo_size
, ilk_plane_wm_reg_max(dev_priv
, level
, is_sprite
));
1953 /* Calculate the maximum cursor plane watermark */
1954 static unsigned int ilk_cursor_wm_max(const struct drm_device
*dev
,
1956 const struct intel_wm_config
*config
)
1958 /* HSW LP1+ watermarks w/ multiple pipes */
1959 if (level
> 0 && config
->num_pipes_active
> 1)
1962 /* otherwise just report max that registers can hold */
1963 return ilk_cursor_wm_reg_max(to_i915(dev
), level
);
1966 static void ilk_compute_wm_maximums(const struct drm_device
*dev
,
1968 const struct intel_wm_config
*config
,
1969 enum intel_ddb_partitioning ddb_partitioning
,
1970 struct ilk_wm_maximums
*max
)
1972 max
->pri
= ilk_plane_wm_max(dev
, level
, config
, ddb_partitioning
, false);
1973 max
->spr
= ilk_plane_wm_max(dev
, level
, config
, ddb_partitioning
, true);
1974 max
->cur
= ilk_cursor_wm_max(dev
, level
, config
);
1975 max
->fbc
= ilk_fbc_wm_reg_max(to_i915(dev
));
1978 static void ilk_compute_wm_reg_maximums(const struct drm_i915_private
*dev_priv
,
1980 struct ilk_wm_maximums
*max
)
1982 max
->pri
= ilk_plane_wm_reg_max(dev_priv
, level
, false);
1983 max
->spr
= ilk_plane_wm_reg_max(dev_priv
, level
, true);
1984 max
->cur
= ilk_cursor_wm_reg_max(dev_priv
, level
);
1985 max
->fbc
= ilk_fbc_wm_reg_max(dev_priv
);
1988 static bool ilk_validate_wm_level(int level
,
1989 const struct ilk_wm_maximums
*max
,
1990 struct intel_wm_level
*result
)
1994 /* already determined to be invalid? */
1995 if (!result
->enable
)
1998 result
->enable
= result
->pri_val
<= max
->pri
&&
1999 result
->spr_val
<= max
->spr
&&
2000 result
->cur_val
<= max
->cur
;
2002 ret
= result
->enable
;
2005 * HACK until we can pre-compute everything,
2006 * and thus fail gracefully if LP0 watermarks
2009 if (level
== 0 && !result
->enable
) {
2010 if (result
->pri_val
> max
->pri
)
2011 DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
2012 level
, result
->pri_val
, max
->pri
);
2013 if (result
->spr_val
> max
->spr
)
2014 DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
2015 level
, result
->spr_val
, max
->spr
);
2016 if (result
->cur_val
> max
->cur
)
2017 DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
2018 level
, result
->cur_val
, max
->cur
);
2020 result
->pri_val
= min_t(uint32_t, result
->pri_val
, max
->pri
);
2021 result
->spr_val
= min_t(uint32_t, result
->spr_val
, max
->spr
);
2022 result
->cur_val
= min_t(uint32_t, result
->cur_val
, max
->cur
);
2023 result
->enable
= true;
2029 static void ilk_compute_wm_level(const struct drm_i915_private
*dev_priv
,
2030 const struct intel_crtc
*intel_crtc
,
2032 struct intel_crtc_state
*cstate
,
2033 struct intel_plane_state
*pristate
,
2034 struct intel_plane_state
*sprstate
,
2035 struct intel_plane_state
*curstate
,
2036 struct intel_wm_level
*result
)
2038 uint16_t pri_latency
= dev_priv
->wm
.pri_latency
[level
];
2039 uint16_t spr_latency
= dev_priv
->wm
.spr_latency
[level
];
2040 uint16_t cur_latency
= dev_priv
->wm
.cur_latency
[level
];
2042 /* WM1+ latency values stored in 0.5us units */
2050 result
->pri_val
= ilk_compute_pri_wm(cstate
, pristate
,
2051 pri_latency
, level
);
2052 result
->fbc_val
= ilk_compute_fbc_wm(cstate
, pristate
, result
->pri_val
);
2056 result
->spr_val
= ilk_compute_spr_wm(cstate
, sprstate
, spr_latency
);
2059 result
->cur_val
= ilk_compute_cur_wm(cstate
, curstate
, cur_latency
);
2061 result
->enable
= true;
2065 hsw_compute_linetime_wm(const struct intel_crtc_state
*cstate
)
2067 const struct intel_atomic_state
*intel_state
=
2068 to_intel_atomic_state(cstate
->base
.state
);
2069 const struct drm_display_mode
*adjusted_mode
=
2070 &cstate
->base
.adjusted_mode
;
2071 u32 linetime
, ips_linetime
;
2073 if (!cstate
->base
.active
)
2075 if (WARN_ON(adjusted_mode
->crtc_clock
== 0))
2077 if (WARN_ON(intel_state
->cdclk
.logical
.cdclk
== 0))
2080 /* The WM are computed with base on how long it takes to fill a single
2081 * row at the given clock rate, multiplied by 8.
2083 linetime
= DIV_ROUND_CLOSEST(adjusted_mode
->crtc_htotal
* 1000 * 8,
2084 adjusted_mode
->crtc_clock
);
2085 ips_linetime
= DIV_ROUND_CLOSEST(adjusted_mode
->crtc_htotal
* 1000 * 8,
2086 intel_state
->cdclk
.logical
.cdclk
);
2088 return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime
) |
2089 PIPE_WM_LINETIME_TIME(linetime
);
2092 static void intel_read_wm_latency(struct drm_i915_private
*dev_priv
,
2095 if (IS_GEN9(dev_priv
)) {
2098 int level
, max_level
= ilk_wm_max_level(dev_priv
);
2100 /* read the first set of memory latencies[0:3] */
2101 val
= 0; /* data0 to be programmed to 0 for first set */
2102 mutex_lock(&dev_priv
->rps
.hw_lock
);
2103 ret
= sandybridge_pcode_read(dev_priv
,
2104 GEN9_PCODE_READ_MEM_LATENCY
,
2106 mutex_unlock(&dev_priv
->rps
.hw_lock
);
2109 DRM_ERROR("SKL Mailbox read error = %d\n", ret
);
2113 wm
[0] = val
& GEN9_MEM_LATENCY_LEVEL_MASK
;
2114 wm
[1] = (val
>> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT
) &
2115 GEN9_MEM_LATENCY_LEVEL_MASK
;
2116 wm
[2] = (val
>> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT
) &
2117 GEN9_MEM_LATENCY_LEVEL_MASK
;
2118 wm
[3] = (val
>> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT
) &
2119 GEN9_MEM_LATENCY_LEVEL_MASK
;
2121 /* read the second set of memory latencies[4:7] */
2122 val
= 1; /* data0 to be programmed to 1 for second set */
2123 mutex_lock(&dev_priv
->rps
.hw_lock
);
2124 ret
= sandybridge_pcode_read(dev_priv
,
2125 GEN9_PCODE_READ_MEM_LATENCY
,
2127 mutex_unlock(&dev_priv
->rps
.hw_lock
);
2129 DRM_ERROR("SKL Mailbox read error = %d\n", ret
);
2133 wm
[4] = val
& GEN9_MEM_LATENCY_LEVEL_MASK
;
2134 wm
[5] = (val
>> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT
) &
2135 GEN9_MEM_LATENCY_LEVEL_MASK
;
2136 wm
[6] = (val
>> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT
) &
2137 GEN9_MEM_LATENCY_LEVEL_MASK
;
2138 wm
[7] = (val
>> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT
) &
2139 GEN9_MEM_LATENCY_LEVEL_MASK
;
2142 * If a level n (n > 1) has a 0us latency, all levels m (m >= n)
2143 * need to be disabled. We make sure to sanitize the values out
2144 * of the punit to satisfy this requirement.
2146 for (level
= 1; level
<= max_level
; level
++) {
2147 if (wm
[level
] == 0) {
2148 for (i
= level
+ 1; i
<= max_level
; i
++)
2155 * WaWmMemoryReadLatency:skl,glk
2157 * punit doesn't take into account the read latency so we need
2158 * to add 2us to the various latency levels we retrieve from the
2159 * punit when level 0 response data us 0us.
2163 for (level
= 1; level
<= max_level
; level
++) {
2170 } else if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
)) {
2171 uint64_t sskpd
= I915_READ64(MCH_SSKPD
);
2173 wm
[0] = (sskpd
>> 56) & 0xFF;
2175 wm
[0] = sskpd
& 0xF;
2176 wm
[1] = (sskpd
>> 4) & 0xFF;
2177 wm
[2] = (sskpd
>> 12) & 0xFF;
2178 wm
[3] = (sskpd
>> 20) & 0x1FF;
2179 wm
[4] = (sskpd
>> 32) & 0x1FF;
2180 } else if (INTEL_GEN(dev_priv
) >= 6) {
2181 uint32_t sskpd
= I915_READ(MCH_SSKPD
);
2183 wm
[0] = (sskpd
>> SSKPD_WM0_SHIFT
) & SSKPD_WM_MASK
;
2184 wm
[1] = (sskpd
>> SSKPD_WM1_SHIFT
) & SSKPD_WM_MASK
;
2185 wm
[2] = (sskpd
>> SSKPD_WM2_SHIFT
) & SSKPD_WM_MASK
;
2186 wm
[3] = (sskpd
>> SSKPD_WM3_SHIFT
) & SSKPD_WM_MASK
;
2187 } else if (INTEL_GEN(dev_priv
) >= 5) {
2188 uint32_t mltr
= I915_READ(MLTR_ILK
);
2190 /* ILK primary LP0 latency is 700 ns */
2192 wm
[1] = (mltr
>> MLTR_WM1_SHIFT
) & ILK_SRLT_MASK
;
2193 wm
[2] = (mltr
>> MLTR_WM2_SHIFT
) & ILK_SRLT_MASK
;
2197 static void intel_fixup_spr_wm_latency(struct drm_i915_private
*dev_priv
,
2200 /* ILK sprite LP0 latency is 1300 ns */
2201 if (IS_GEN5(dev_priv
))
2205 static void intel_fixup_cur_wm_latency(struct drm_i915_private
*dev_priv
,
2208 /* ILK cursor LP0 latency is 1300 ns */
2209 if (IS_GEN5(dev_priv
))
2212 /* WaDoubleCursorLP3Latency:ivb */
2213 if (IS_IVYBRIDGE(dev_priv
))
2217 int ilk_wm_max_level(const struct drm_i915_private
*dev_priv
)
2219 /* how many WM levels are we expecting */
2220 if (INTEL_GEN(dev_priv
) >= 9)
2222 else if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
2224 else if (INTEL_GEN(dev_priv
) >= 6)
2230 static void intel_print_wm_latency(struct drm_i915_private
*dev_priv
,
2232 const uint16_t wm
[8])
2234 int level
, max_level
= ilk_wm_max_level(dev_priv
);
2236 for (level
= 0; level
<= max_level
; level
++) {
2237 unsigned int latency
= wm
[level
];
2240 DRM_ERROR("%s WM%d latency not provided\n",
2246 * - latencies are in us on gen9.
2247 * - before then, WM1+ latency values are in 0.5us units
2249 if (IS_GEN9(dev_priv
))
2254 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2255 name
, level
, wm
[level
],
2256 latency
/ 10, latency
% 10);
2260 static bool ilk_increase_wm_latency(struct drm_i915_private
*dev_priv
,
2261 uint16_t wm
[5], uint16_t min
)
2263 int level
, max_level
= ilk_wm_max_level(dev_priv
);
2268 wm
[0] = max(wm
[0], min
);
2269 for (level
= 1; level
<= max_level
; level
++)
2270 wm
[level
] = max_t(uint16_t, wm
[level
], DIV_ROUND_UP(min
, 5));
2275 static void snb_wm_latency_quirk(struct drm_i915_private
*dev_priv
)
2280 * The BIOS provided WM memory latency values are often
2281 * inadequate for high resolution displays. Adjust them.
2283 changed
= ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.pri_latency
, 12) |
2284 ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.spr_latency
, 12) |
2285 ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.cur_latency
, 12);
2290 DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2291 intel_print_wm_latency(dev_priv
, "Primary", dev_priv
->wm
.pri_latency
);
2292 intel_print_wm_latency(dev_priv
, "Sprite", dev_priv
->wm
.spr_latency
);
2293 intel_print_wm_latency(dev_priv
, "Cursor", dev_priv
->wm
.cur_latency
);
2296 static void ilk_setup_wm_latency(struct drm_i915_private
*dev_priv
)
2298 intel_read_wm_latency(dev_priv
, dev_priv
->wm
.pri_latency
);
2300 memcpy(dev_priv
->wm
.spr_latency
, dev_priv
->wm
.pri_latency
,
2301 sizeof(dev_priv
->wm
.pri_latency
));
2302 memcpy(dev_priv
->wm
.cur_latency
, dev_priv
->wm
.pri_latency
,
2303 sizeof(dev_priv
->wm
.pri_latency
));
2305 intel_fixup_spr_wm_latency(dev_priv
, dev_priv
->wm
.spr_latency
);
2306 intel_fixup_cur_wm_latency(dev_priv
, dev_priv
->wm
.cur_latency
);
2308 intel_print_wm_latency(dev_priv
, "Primary", dev_priv
->wm
.pri_latency
);
2309 intel_print_wm_latency(dev_priv
, "Sprite", dev_priv
->wm
.spr_latency
);
2310 intel_print_wm_latency(dev_priv
, "Cursor", dev_priv
->wm
.cur_latency
);
2312 if (IS_GEN6(dev_priv
))
2313 snb_wm_latency_quirk(dev_priv
);
2316 static void skl_setup_wm_latency(struct drm_i915_private
*dev_priv
)
2318 intel_read_wm_latency(dev_priv
, dev_priv
->wm
.skl_latency
);
2319 intel_print_wm_latency(dev_priv
, "Gen9 Plane", dev_priv
->wm
.skl_latency
);
2322 static bool ilk_validate_pipe_wm(struct drm_device
*dev
,
2323 struct intel_pipe_wm
*pipe_wm
)
2325 /* LP0 watermark maximums depend on this pipe alone */
2326 const struct intel_wm_config config
= {
2327 .num_pipes_active
= 1,
2328 .sprites_enabled
= pipe_wm
->sprites_enabled
,
2329 .sprites_scaled
= pipe_wm
->sprites_scaled
,
2331 struct ilk_wm_maximums max
;
2333 /* LP0 watermarks always use 1/2 DDB partitioning */
2334 ilk_compute_wm_maximums(dev
, 0, &config
, INTEL_DDB_PART_1_2
, &max
);
2336 /* At least LP0 must be valid */
2337 if (!ilk_validate_wm_level(0, &max
, &pipe_wm
->wm
[0])) {
2338 DRM_DEBUG_KMS("LP0 watermark invalid\n");
2345 /* Compute new watermarks for the pipe */
2346 static int ilk_compute_pipe_wm(struct intel_crtc_state
*cstate
)
2348 struct drm_atomic_state
*state
= cstate
->base
.state
;
2349 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
2350 struct intel_pipe_wm
*pipe_wm
;
2351 struct drm_device
*dev
= state
->dev
;
2352 const struct drm_i915_private
*dev_priv
= to_i915(dev
);
2353 struct intel_plane
*intel_plane
;
2354 struct intel_plane_state
*pristate
= NULL
;
2355 struct intel_plane_state
*sprstate
= NULL
;
2356 struct intel_plane_state
*curstate
= NULL
;
2357 int level
, max_level
= ilk_wm_max_level(dev_priv
), usable_level
;
2358 struct ilk_wm_maximums max
;
2360 pipe_wm
= &cstate
->wm
.ilk
.optimal
;
2362 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
2363 struct intel_plane_state
*ps
;
2365 ps
= intel_atomic_get_existing_plane_state(state
,
2370 if (intel_plane
->base
.type
== DRM_PLANE_TYPE_PRIMARY
)
2372 else if (intel_plane
->base
.type
== DRM_PLANE_TYPE_OVERLAY
)
2374 else if (intel_plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
)
2378 pipe_wm
->pipe_enabled
= cstate
->base
.active
;
2380 pipe_wm
->sprites_enabled
= sprstate
->base
.visible
;
2381 pipe_wm
->sprites_scaled
= sprstate
->base
.visible
&&
2382 (drm_rect_width(&sprstate
->base
.dst
) != drm_rect_width(&sprstate
->base
.src
) >> 16 ||
2383 drm_rect_height(&sprstate
->base
.dst
) != drm_rect_height(&sprstate
->base
.src
) >> 16);
2386 usable_level
= max_level
;
2388 /* ILK/SNB: LP2+ watermarks only w/o sprites */
2389 if (INTEL_GEN(dev_priv
) <= 6 && pipe_wm
->sprites_enabled
)
2392 /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2393 if (pipe_wm
->sprites_scaled
)
2396 ilk_compute_wm_level(dev_priv
, intel_crtc
, 0, cstate
,
2397 pristate
, sprstate
, curstate
, &pipe_wm
->raw_wm
[0]);
2399 memset(&pipe_wm
->wm
, 0, sizeof(pipe_wm
->wm
));
2400 pipe_wm
->wm
[0] = pipe_wm
->raw_wm
[0];
2402 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
2403 pipe_wm
->linetime
= hsw_compute_linetime_wm(cstate
);
2405 if (!ilk_validate_pipe_wm(dev
, pipe_wm
))
2408 ilk_compute_wm_reg_maximums(dev_priv
, 1, &max
);
2410 for (level
= 1; level
<= max_level
; level
++) {
2411 struct intel_wm_level
*wm
= &pipe_wm
->raw_wm
[level
];
2413 ilk_compute_wm_level(dev_priv
, intel_crtc
, level
, cstate
,
2414 pristate
, sprstate
, curstate
, wm
);
2417 * Disable any watermark level that exceeds the
2418 * register maximums since such watermarks are
2421 if (level
> usable_level
)
2424 if (ilk_validate_wm_level(level
, &max
, wm
))
2425 pipe_wm
->wm
[level
] = *wm
;
2427 usable_level
= level
;
2434 * Build a set of 'intermediate' watermark values that satisfy both the old
2435 * state and the new state. These can be programmed to the hardware
2438 static int ilk_compute_intermediate_wm(struct drm_device
*dev
,
2439 struct intel_crtc
*intel_crtc
,
2440 struct intel_crtc_state
*newstate
)
2442 struct intel_pipe_wm
*a
= &newstate
->wm
.ilk
.intermediate
;
2443 struct intel_pipe_wm
*b
= &intel_crtc
->wm
.active
.ilk
;
2444 int level
, max_level
= ilk_wm_max_level(to_i915(dev
));
2447 * Start with the final, target watermarks, then combine with the
2448 * currently active watermarks to get values that are safe both before
2449 * and after the vblank.
2451 *a
= newstate
->wm
.ilk
.optimal
;
2452 a
->pipe_enabled
|= b
->pipe_enabled
;
2453 a
->sprites_enabled
|= b
->sprites_enabled
;
2454 a
->sprites_scaled
|= b
->sprites_scaled
;
2456 for (level
= 0; level
<= max_level
; level
++) {
2457 struct intel_wm_level
*a_wm
= &a
->wm
[level
];
2458 const struct intel_wm_level
*b_wm
= &b
->wm
[level
];
2460 a_wm
->enable
&= b_wm
->enable
;
2461 a_wm
->pri_val
= max(a_wm
->pri_val
, b_wm
->pri_val
);
2462 a_wm
->spr_val
= max(a_wm
->spr_val
, b_wm
->spr_val
);
2463 a_wm
->cur_val
= max(a_wm
->cur_val
, b_wm
->cur_val
);
2464 a_wm
->fbc_val
= max(a_wm
->fbc_val
, b_wm
->fbc_val
);
2468 * We need to make sure that these merged watermark values are
2469 * actually a valid configuration themselves. If they're not,
2470 * there's no safe way to transition from the old state to
2471 * the new state, so we need to fail the atomic transaction.
2473 if (!ilk_validate_pipe_wm(dev
, a
))
2477 * If our intermediate WM are identical to the final WM, then we can
2478 * omit the post-vblank programming; only update if it's different.
2480 if (memcmp(a
, &newstate
->wm
.ilk
.optimal
, sizeof(*a
)) == 0)
2481 newstate
->wm
.need_postvbl_update
= false;
2487 * Merge the watermarks from all active pipes for a specific level.
2489 static void ilk_merge_wm_level(struct drm_device
*dev
,
2491 struct intel_wm_level
*ret_wm
)
2493 const struct intel_crtc
*intel_crtc
;
2495 ret_wm
->enable
= true;
2497 for_each_intel_crtc(dev
, intel_crtc
) {
2498 const struct intel_pipe_wm
*active
= &intel_crtc
->wm
.active
.ilk
;
2499 const struct intel_wm_level
*wm
= &active
->wm
[level
];
2501 if (!active
->pipe_enabled
)
2505 * The watermark values may have been used in the past,
2506 * so we must maintain them in the registers for some
2507 * time even if the level is now disabled.
2510 ret_wm
->enable
= false;
2512 ret_wm
->pri_val
= max(ret_wm
->pri_val
, wm
->pri_val
);
2513 ret_wm
->spr_val
= max(ret_wm
->spr_val
, wm
->spr_val
);
2514 ret_wm
->cur_val
= max(ret_wm
->cur_val
, wm
->cur_val
);
2515 ret_wm
->fbc_val
= max(ret_wm
->fbc_val
, wm
->fbc_val
);
2520 * Merge all low power watermarks for all active pipes.
2522 static void ilk_wm_merge(struct drm_device
*dev
,
2523 const struct intel_wm_config
*config
,
2524 const struct ilk_wm_maximums
*max
,
2525 struct intel_pipe_wm
*merged
)
2527 struct drm_i915_private
*dev_priv
= to_i915(dev
);
2528 int level
, max_level
= ilk_wm_max_level(dev_priv
);
2529 int last_enabled_level
= max_level
;
2531 /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2532 if ((INTEL_GEN(dev_priv
) <= 6 || IS_IVYBRIDGE(dev_priv
)) &&
2533 config
->num_pipes_active
> 1)
2534 last_enabled_level
= 0;
2536 /* ILK: FBC WM must be disabled always */
2537 merged
->fbc_wm_enabled
= INTEL_GEN(dev_priv
) >= 6;
2539 /* merge each WM1+ level */
2540 for (level
= 1; level
<= max_level
; level
++) {
2541 struct intel_wm_level
*wm
= &merged
->wm
[level
];
2543 ilk_merge_wm_level(dev
, level
, wm
);
2545 if (level
> last_enabled_level
)
2547 else if (!ilk_validate_wm_level(level
, max
, wm
))
2548 /* make sure all following levels get disabled */
2549 last_enabled_level
= level
- 1;
2552 * The spec says it is preferred to disable
2553 * FBC WMs instead of disabling a WM level.
2555 if (wm
->fbc_val
> max
->fbc
) {
2557 merged
->fbc_wm_enabled
= false;
2562 /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2564 * FIXME this is racy. FBC might get enabled later.
2565 * What we should check here is whether FBC can be
2566 * enabled sometime later.
2568 if (IS_GEN5(dev_priv
) && !merged
->fbc_wm_enabled
&&
2569 intel_fbc_is_active(dev_priv
)) {
2570 for (level
= 2; level
<= max_level
; level
++) {
2571 struct intel_wm_level
*wm
= &merged
->wm
[level
];
2578 static int ilk_wm_lp_to_level(int wm_lp
, const struct intel_pipe_wm
*pipe_wm
)
2580 /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2581 return wm_lp
+ (wm_lp
>= 2 && pipe_wm
->wm
[4].enable
);
2584 /* The value we need to program into the WM_LPx latency field */
2585 static unsigned int ilk_wm_lp_latency(struct drm_device
*dev
, int level
)
2587 struct drm_i915_private
*dev_priv
= to_i915(dev
);
2589 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
2592 return dev_priv
->wm
.pri_latency
[level
];
2595 static void ilk_compute_wm_results(struct drm_device
*dev
,
2596 const struct intel_pipe_wm
*merged
,
2597 enum intel_ddb_partitioning partitioning
,
2598 struct ilk_wm_values
*results
)
2600 struct drm_i915_private
*dev_priv
= to_i915(dev
);
2601 struct intel_crtc
*intel_crtc
;
2604 results
->enable_fbc_wm
= merged
->fbc_wm_enabled
;
2605 results
->partitioning
= partitioning
;
2607 /* LP1+ register values */
2608 for (wm_lp
= 1; wm_lp
<= 3; wm_lp
++) {
2609 const struct intel_wm_level
*r
;
2611 level
= ilk_wm_lp_to_level(wm_lp
, merged
);
2613 r
= &merged
->wm
[level
];
2616 * Maintain the watermark values even if the level is
2617 * disabled. Doing otherwise could cause underruns.
2619 results
->wm_lp
[wm_lp
- 1] =
2620 (ilk_wm_lp_latency(dev
, level
) << WM1_LP_LATENCY_SHIFT
) |
2621 (r
->pri_val
<< WM1_LP_SR_SHIFT
) |
2625 results
->wm_lp
[wm_lp
- 1] |= WM1_LP_SR_EN
;
2627 if (INTEL_GEN(dev_priv
) >= 8)
2628 results
->wm_lp
[wm_lp
- 1] |=
2629 r
->fbc_val
<< WM1_LP_FBC_SHIFT_BDW
;
2631 results
->wm_lp
[wm_lp
- 1] |=
2632 r
->fbc_val
<< WM1_LP_FBC_SHIFT
;
2635 * Always set WM1S_LP_EN when spr_val != 0, even if the
2636 * level is disabled. Doing otherwise could cause underruns.
2638 if (INTEL_GEN(dev_priv
) <= 6 && r
->spr_val
) {
2639 WARN_ON(wm_lp
!= 1);
2640 results
->wm_lp_spr
[wm_lp
- 1] = WM1S_LP_EN
| r
->spr_val
;
2642 results
->wm_lp_spr
[wm_lp
- 1] = r
->spr_val
;
2645 /* LP0 register values */
2646 for_each_intel_crtc(dev
, intel_crtc
) {
2647 enum pipe pipe
= intel_crtc
->pipe
;
2648 const struct intel_wm_level
*r
=
2649 &intel_crtc
->wm
.active
.ilk
.wm
[0];
2651 if (WARN_ON(!r
->enable
))
2654 results
->wm_linetime
[pipe
] = intel_crtc
->wm
.active
.ilk
.linetime
;
2656 results
->wm_pipe
[pipe
] =
2657 (r
->pri_val
<< WM0_PIPE_PLANE_SHIFT
) |
2658 (r
->spr_val
<< WM0_PIPE_SPRITE_SHIFT
) |
2663 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2664 * case both are at the same level. Prefer r1 in case they're the same. */
2665 static struct intel_pipe_wm
*ilk_find_best_result(struct drm_device
*dev
,
2666 struct intel_pipe_wm
*r1
,
2667 struct intel_pipe_wm
*r2
)
2669 int level
, max_level
= ilk_wm_max_level(to_i915(dev
));
2670 int level1
= 0, level2
= 0;
2672 for (level
= 1; level
<= max_level
; level
++) {
2673 if (r1
->wm
[level
].enable
)
2675 if (r2
->wm
[level
].enable
)
2679 if (level1
== level2
) {
2680 if (r2
->fbc_wm_enabled
&& !r1
->fbc_wm_enabled
)
2684 } else if (level1
> level2
) {
2691 /* dirty bits used to track which watermarks need changes */
2692 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2693 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2694 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2695 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2696 #define WM_DIRTY_FBC (1 << 24)
2697 #define WM_DIRTY_DDB (1 << 25)
2699 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private
*dev_priv
,
2700 const struct ilk_wm_values
*old
,
2701 const struct ilk_wm_values
*new)
2703 unsigned int dirty
= 0;
2707 for_each_pipe(dev_priv
, pipe
) {
2708 if (old
->wm_linetime
[pipe
] != new->wm_linetime
[pipe
]) {
2709 dirty
|= WM_DIRTY_LINETIME(pipe
);
2710 /* Must disable LP1+ watermarks too */
2711 dirty
|= WM_DIRTY_LP_ALL
;
2714 if (old
->wm_pipe
[pipe
] != new->wm_pipe
[pipe
]) {
2715 dirty
|= WM_DIRTY_PIPE(pipe
);
2716 /* Must disable LP1+ watermarks too */
2717 dirty
|= WM_DIRTY_LP_ALL
;
2721 if (old
->enable_fbc_wm
!= new->enable_fbc_wm
) {
2722 dirty
|= WM_DIRTY_FBC
;
2723 /* Must disable LP1+ watermarks too */
2724 dirty
|= WM_DIRTY_LP_ALL
;
2727 if (old
->partitioning
!= new->partitioning
) {
2728 dirty
|= WM_DIRTY_DDB
;
2729 /* Must disable LP1+ watermarks too */
2730 dirty
|= WM_DIRTY_LP_ALL
;
2733 /* LP1+ watermarks already deemed dirty, no need to continue */
2734 if (dirty
& WM_DIRTY_LP_ALL
)
2737 /* Find the lowest numbered LP1+ watermark in need of an update... */
2738 for (wm_lp
= 1; wm_lp
<= 3; wm_lp
++) {
2739 if (old
->wm_lp
[wm_lp
- 1] != new->wm_lp
[wm_lp
- 1] ||
2740 old
->wm_lp_spr
[wm_lp
- 1] != new->wm_lp_spr
[wm_lp
- 1])
2744 /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2745 for (; wm_lp
<= 3; wm_lp
++)
2746 dirty
|= WM_DIRTY_LP(wm_lp
);
2751 static bool _ilk_disable_lp_wm(struct drm_i915_private
*dev_priv
,
2754 struct ilk_wm_values
*previous
= &dev_priv
->wm
.hw
;
2755 bool changed
= false;
2757 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp
[2] & WM1_LP_SR_EN
) {
2758 previous
->wm_lp
[2] &= ~WM1_LP_SR_EN
;
2759 I915_WRITE(WM3_LP_ILK
, previous
->wm_lp
[2]);
2762 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp
[1] & WM1_LP_SR_EN
) {
2763 previous
->wm_lp
[1] &= ~WM1_LP_SR_EN
;
2764 I915_WRITE(WM2_LP_ILK
, previous
->wm_lp
[1]);
2767 if (dirty
& WM_DIRTY_LP(1) && previous
->wm_lp
[0] & WM1_LP_SR_EN
) {
2768 previous
->wm_lp
[0] &= ~WM1_LP_SR_EN
;
2769 I915_WRITE(WM1_LP_ILK
, previous
->wm_lp
[0]);
2774 * Don't touch WM1S_LP_EN here.
2775 * Doing so could cause underruns.
2782 * The spec says we shouldn't write when we don't need, because every write
2783 * causes WMs to be re-evaluated, expending some power.
2785 static void ilk_write_wm_values(struct drm_i915_private
*dev_priv
,
2786 struct ilk_wm_values
*results
)
2788 struct ilk_wm_values
*previous
= &dev_priv
->wm
.hw
;
2792 dirty
= ilk_compute_wm_dirty(dev_priv
, previous
, results
);
2796 _ilk_disable_lp_wm(dev_priv
, dirty
);
2798 if (dirty
& WM_DIRTY_PIPE(PIPE_A
))
2799 I915_WRITE(WM0_PIPEA_ILK
, results
->wm_pipe
[0]);
2800 if (dirty
& WM_DIRTY_PIPE(PIPE_B
))
2801 I915_WRITE(WM0_PIPEB_ILK
, results
->wm_pipe
[1]);
2802 if (dirty
& WM_DIRTY_PIPE(PIPE_C
))
2803 I915_WRITE(WM0_PIPEC_IVB
, results
->wm_pipe
[2]);
2805 if (dirty
& WM_DIRTY_LINETIME(PIPE_A
))
2806 I915_WRITE(PIPE_WM_LINETIME(PIPE_A
), results
->wm_linetime
[0]);
2807 if (dirty
& WM_DIRTY_LINETIME(PIPE_B
))
2808 I915_WRITE(PIPE_WM_LINETIME(PIPE_B
), results
->wm_linetime
[1]);
2809 if (dirty
& WM_DIRTY_LINETIME(PIPE_C
))
2810 I915_WRITE(PIPE_WM_LINETIME(PIPE_C
), results
->wm_linetime
[2]);
2812 if (dirty
& WM_DIRTY_DDB
) {
2813 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
)) {
2814 val
= I915_READ(WM_MISC
);
2815 if (results
->partitioning
== INTEL_DDB_PART_1_2
)
2816 val
&= ~WM_MISC_DATA_PARTITION_5_6
;
2818 val
|= WM_MISC_DATA_PARTITION_5_6
;
2819 I915_WRITE(WM_MISC
, val
);
2821 val
= I915_READ(DISP_ARB_CTL2
);
2822 if (results
->partitioning
== INTEL_DDB_PART_1_2
)
2823 val
&= ~DISP_DATA_PARTITION_5_6
;
2825 val
|= DISP_DATA_PARTITION_5_6
;
2826 I915_WRITE(DISP_ARB_CTL2
, val
);
2830 if (dirty
& WM_DIRTY_FBC
) {
2831 val
= I915_READ(DISP_ARB_CTL
);
2832 if (results
->enable_fbc_wm
)
2833 val
&= ~DISP_FBC_WM_DIS
;
2835 val
|= DISP_FBC_WM_DIS
;
2836 I915_WRITE(DISP_ARB_CTL
, val
);
2839 if (dirty
& WM_DIRTY_LP(1) &&
2840 previous
->wm_lp_spr
[0] != results
->wm_lp_spr
[0])
2841 I915_WRITE(WM1S_LP_ILK
, results
->wm_lp_spr
[0]);
2843 if (INTEL_GEN(dev_priv
) >= 7) {
2844 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp_spr
[1] != results
->wm_lp_spr
[1])
2845 I915_WRITE(WM2S_LP_IVB
, results
->wm_lp_spr
[1]);
2846 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp_spr
[2] != results
->wm_lp_spr
[2])
2847 I915_WRITE(WM3S_LP_IVB
, results
->wm_lp_spr
[2]);
2850 if (dirty
& WM_DIRTY_LP(1) && previous
->wm_lp
[0] != results
->wm_lp
[0])
2851 I915_WRITE(WM1_LP_ILK
, results
->wm_lp
[0]);
2852 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp
[1] != results
->wm_lp
[1])
2853 I915_WRITE(WM2_LP_ILK
, results
->wm_lp
[1]);
2854 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp
[2] != results
->wm_lp
[2])
2855 I915_WRITE(WM3_LP_ILK
, results
->wm_lp
[2]);
2857 dev_priv
->wm
.hw
= *results
;
2860 bool ilk_disable_lp_wm(struct drm_device
*dev
)
2862 struct drm_i915_private
*dev_priv
= to_i915(dev
);
2864 return _ilk_disable_lp_wm(dev_priv
, WM_DIRTY_LP_ALL
);
2867 #define SKL_SAGV_BLOCK_TIME 30 /* µs */
2870 * FIXME: We still don't have the proper code detect if we need to apply the WA,
2871 * so assume we'll always need it in order to avoid underruns.
2873 static bool skl_needs_memory_bw_wa(struct intel_atomic_state
*state
)
2875 struct drm_i915_private
*dev_priv
= to_i915(state
->base
.dev
);
2877 if (IS_GEN9_BC(dev_priv
) || IS_BROXTON(dev_priv
))
2884 intel_has_sagv(struct drm_i915_private
*dev_priv
)
2886 if (IS_KABYLAKE(dev_priv
))
2889 if (IS_SKYLAKE(dev_priv
) &&
2890 dev_priv
->sagv_status
!= I915_SAGV_NOT_CONTROLLED
)
2897 * SAGV dynamically adjusts the system agent voltage and clock frequencies
2898 * depending on power and performance requirements. The display engine access
2899 * to system memory is blocked during the adjustment time. Because of the
2900 * blocking time, having this enabled can cause full system hangs and/or pipe
2901 * underruns if we don't meet all of the following requirements:
2903 * - <= 1 pipe enabled
2904 * - All planes can enable watermarks for latencies >= SAGV engine block time
2905 * - We're not using an interlaced display configuration
2908 intel_enable_sagv(struct drm_i915_private
*dev_priv
)
2912 if (!intel_has_sagv(dev_priv
))
2915 if (dev_priv
->sagv_status
== I915_SAGV_ENABLED
)
2918 DRM_DEBUG_KMS("Enabling the SAGV\n");
2919 mutex_lock(&dev_priv
->rps
.hw_lock
);
2921 ret
= sandybridge_pcode_write(dev_priv
, GEN9_PCODE_SAGV_CONTROL
,
2924 /* We don't need to wait for the SAGV when enabling */
2925 mutex_unlock(&dev_priv
->rps
.hw_lock
);
2928 * Some skl systems, pre-release machines in particular,
2929 * don't actually have an SAGV.
2931 if (IS_SKYLAKE(dev_priv
) && ret
== -ENXIO
) {
2932 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
2933 dev_priv
->sagv_status
= I915_SAGV_NOT_CONTROLLED
;
2935 } else if (ret
< 0) {
2936 DRM_ERROR("Failed to enable the SAGV\n");
2940 dev_priv
->sagv_status
= I915_SAGV_ENABLED
;
2945 intel_disable_sagv(struct drm_i915_private
*dev_priv
)
2949 if (!intel_has_sagv(dev_priv
))
2952 if (dev_priv
->sagv_status
== I915_SAGV_DISABLED
)
2955 DRM_DEBUG_KMS("Disabling the SAGV\n");
2956 mutex_lock(&dev_priv
->rps
.hw_lock
);
2958 /* bspec says to keep retrying for at least 1 ms */
2959 ret
= skl_pcode_request(dev_priv
, GEN9_PCODE_SAGV_CONTROL
,
2961 GEN9_SAGV_IS_DISABLED
, GEN9_SAGV_IS_DISABLED
,
2963 mutex_unlock(&dev_priv
->rps
.hw_lock
);
2966 * Some skl systems, pre-release machines in particular,
2967 * don't actually have an SAGV.
2969 if (IS_SKYLAKE(dev_priv
) && ret
== -ENXIO
) {
2970 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
2971 dev_priv
->sagv_status
= I915_SAGV_NOT_CONTROLLED
;
2973 } else if (ret
< 0) {
2974 DRM_ERROR("Failed to disable the SAGV (%d)\n", ret
);
2978 dev_priv
->sagv_status
= I915_SAGV_DISABLED
;
2982 bool intel_can_enable_sagv(struct drm_atomic_state
*state
)
2984 struct drm_device
*dev
= state
->dev
;
2985 struct drm_i915_private
*dev_priv
= to_i915(dev
);
2986 struct intel_atomic_state
*intel_state
= to_intel_atomic_state(state
);
2987 struct intel_crtc
*crtc
;
2988 struct intel_plane
*plane
;
2989 struct intel_crtc_state
*cstate
;
2993 if (!intel_has_sagv(dev_priv
))
2997 * SKL workaround: bspec recommends we disable the SAGV when we have
2998 * more then one pipe enabled
3000 * If there are no active CRTCs, no additional checks need be performed
3002 if (hweight32(intel_state
->active_crtcs
) == 0)
3004 else if (hweight32(intel_state
->active_crtcs
) > 1)
3007 /* Since we're now guaranteed to only have one active CRTC... */
3008 pipe
= ffs(intel_state
->active_crtcs
) - 1;
3009 crtc
= intel_get_crtc_for_pipe(dev_priv
, pipe
);
3010 cstate
= to_intel_crtc_state(crtc
->base
.state
);
3012 if (crtc
->base
.state
->adjusted_mode
.flags
& DRM_MODE_FLAG_INTERLACE
)
3015 for_each_intel_plane_on_crtc(dev
, crtc
, plane
) {
3016 struct skl_plane_wm
*wm
=
3017 &cstate
->wm
.skl
.optimal
.planes
[plane
->id
];
3019 /* Skip this plane if it's not enabled */
3020 if (!wm
->wm
[0].plane_en
)
3023 /* Find the highest enabled wm level for this plane */
3024 for (level
= ilk_wm_max_level(dev_priv
);
3025 !wm
->wm
[level
].plane_en
; --level
)
3028 latency
= dev_priv
->wm
.skl_latency
[level
];
3030 if (skl_needs_memory_bw_wa(intel_state
) &&
3031 plane
->base
.state
->fb
->modifier
==
3032 I915_FORMAT_MOD_X_TILED
)
3036 * If any of the planes on this pipe don't enable wm levels
3037 * that incur memory latencies higher then 30µs we can't enable
3040 if (latency
< SKL_SAGV_BLOCK_TIME
)
3048 skl_ddb_get_pipe_allocation_limits(struct drm_device
*dev
,
3049 const struct intel_crtc_state
*cstate
,
3050 struct skl_ddb_entry
*alloc
, /* out */
3051 int *num_active
/* out */)
3053 struct drm_atomic_state
*state
= cstate
->base
.state
;
3054 struct intel_atomic_state
*intel_state
= to_intel_atomic_state(state
);
3055 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3056 struct drm_crtc
*for_crtc
= cstate
->base
.crtc
;
3057 unsigned int pipe_size
, ddb_size
;
3058 int nth_active_pipe
;
3060 if (WARN_ON(!state
) || !cstate
->base
.active
) {
3063 *num_active
= hweight32(dev_priv
->active_crtcs
);
3067 if (intel_state
->active_pipe_changes
)
3068 *num_active
= hweight32(intel_state
->active_crtcs
);
3070 *num_active
= hweight32(dev_priv
->active_crtcs
);
3072 ddb_size
= INTEL_INFO(dev_priv
)->ddb_size
;
3073 WARN_ON(ddb_size
== 0);
3075 ddb_size
-= 4; /* 4 blocks for bypass path allocation */
3078 * If the state doesn't change the active CRTC's, then there's
3079 * no need to recalculate; the existing pipe allocation limits
3080 * should remain unchanged. Note that we're safe from racing
3081 * commits since any racing commit that changes the active CRTC
3082 * list would need to grab _all_ crtc locks, including the one
3083 * we currently hold.
3085 if (!intel_state
->active_pipe_changes
) {
3087 * alloc may be cleared by clear_intel_crtc_state,
3088 * copy from old state to be sure
3090 *alloc
= to_intel_crtc_state(for_crtc
->state
)->wm
.skl
.ddb
;
3094 nth_active_pipe
= hweight32(intel_state
->active_crtcs
&
3095 (drm_crtc_mask(for_crtc
) - 1));
3096 pipe_size
= ddb_size
/ hweight32(intel_state
->active_crtcs
);
3097 alloc
->start
= nth_active_pipe
* ddb_size
/ *num_active
;
3098 alloc
->end
= alloc
->start
+ pipe_size
;
3101 static unsigned int skl_cursor_allocation(int num_active
)
3103 if (num_active
== 1)
3109 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry
*entry
, u32 reg
)
3111 entry
->start
= reg
& 0x3ff;
3112 entry
->end
= (reg
>> 16) & 0x3ff;
3117 void skl_ddb_get_hw_state(struct drm_i915_private
*dev_priv
,
3118 struct skl_ddb_allocation
*ddb
/* out */)
3120 struct intel_crtc
*crtc
;
3122 memset(ddb
, 0, sizeof(*ddb
));
3124 for_each_intel_crtc(&dev_priv
->drm
, crtc
) {
3125 enum intel_display_power_domain power_domain
;
3126 enum plane_id plane_id
;
3127 enum pipe pipe
= crtc
->pipe
;
3129 power_domain
= POWER_DOMAIN_PIPE(pipe
);
3130 if (!intel_display_power_get_if_enabled(dev_priv
, power_domain
))
3133 for_each_plane_id_on_crtc(crtc
, plane_id
) {
3136 if (plane_id
!= PLANE_CURSOR
)
3137 val
= I915_READ(PLANE_BUF_CFG(pipe
, plane_id
));
3139 val
= I915_READ(CUR_BUF_CFG(pipe
));
3141 skl_ddb_entry_init_from_hw(&ddb
->plane
[pipe
][plane_id
], val
);
3144 intel_display_power_put(dev_priv
, power_domain
);
3149 * Determines the downscale amount of a plane for the purposes of watermark calculations.
3150 * The bspec defines downscale amount as:
3153 * Horizontal down scale amount = maximum[1, Horizontal source size /
3154 * Horizontal destination size]
3155 * Vertical down scale amount = maximum[1, Vertical source size /
3156 * Vertical destination size]
3157 * Total down scale amount = Horizontal down scale amount *
3158 * Vertical down scale amount
3161 * Return value is provided in 16.16 fixed point form to retain fractional part.
3162 * Caller should take care of dividing & rounding off the value.
3165 skl_plane_downscale_amount(const struct intel_plane_state
*pstate
)
3167 uint32_t downscale_h
, downscale_w
;
3168 uint32_t src_w
, src_h
, dst_w
, dst_h
;
3170 if (WARN_ON(!pstate
->base
.visible
))
3171 return DRM_PLANE_HELPER_NO_SCALING
;
3173 /* n.b., src is 16.16 fixed point, dst is whole integer */
3174 src_w
= drm_rect_width(&pstate
->base
.src
);
3175 src_h
= drm_rect_height(&pstate
->base
.src
);
3176 dst_w
= drm_rect_width(&pstate
->base
.dst
);
3177 dst_h
= drm_rect_height(&pstate
->base
.dst
);
3178 if (drm_rotation_90_or_270(pstate
->base
.rotation
))
3181 downscale_h
= max(src_h
/ dst_h
, (uint32_t)DRM_PLANE_HELPER_NO_SCALING
);
3182 downscale_w
= max(src_w
/ dst_w
, (uint32_t)DRM_PLANE_HELPER_NO_SCALING
);
3184 /* Provide result in 16.16 fixed point */
3185 return (uint64_t)downscale_w
* downscale_h
>> 16;
3189 skl_plane_relative_data_rate(const struct intel_crtc_state
*cstate
,
3190 const struct drm_plane_state
*pstate
,
3193 struct intel_plane_state
*intel_pstate
= to_intel_plane_state(pstate
);
3194 uint32_t down_scale_amount
, data_rate
;
3195 uint32_t width
= 0, height
= 0;
3196 struct drm_framebuffer
*fb
;
3199 if (!intel_pstate
->base
.visible
)
3203 format
= fb
->format
->format
;
3205 if (pstate
->plane
->type
== DRM_PLANE_TYPE_CURSOR
)
3207 if (y
&& format
!= DRM_FORMAT_NV12
)
3210 width
= drm_rect_width(&intel_pstate
->base
.src
) >> 16;
3211 height
= drm_rect_height(&intel_pstate
->base
.src
) >> 16;
3213 if (drm_rotation_90_or_270(pstate
->rotation
))
3214 swap(width
, height
);
3216 /* for planar format */
3217 if (format
== DRM_FORMAT_NV12
) {
3218 if (y
) /* y-plane data rate */
3219 data_rate
= width
* height
*
3221 else /* uv-plane data rate */
3222 data_rate
= (width
/ 2) * (height
/ 2) *
3225 /* for packed formats */
3226 data_rate
= width
* height
* fb
->format
->cpp
[0];
3229 down_scale_amount
= skl_plane_downscale_amount(intel_pstate
);
3231 return (uint64_t)data_rate
* down_scale_amount
>> 16;
3235 * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
3236 * a 8192x4096@32bpp framebuffer:
3237 * 3 * 4096 * 8192 * 4 < 2^32
3240 skl_get_total_relative_data_rate(struct intel_crtc_state
*intel_cstate
,
3241 unsigned *plane_data_rate
,
3242 unsigned *plane_y_data_rate
)
3244 struct drm_crtc_state
*cstate
= &intel_cstate
->base
;
3245 struct drm_atomic_state
*state
= cstate
->state
;
3246 struct drm_plane
*plane
;
3247 const struct drm_plane_state
*pstate
;
3248 unsigned int total_data_rate
= 0;
3250 if (WARN_ON(!state
))
3253 /* Calculate and cache data rate for each plane */
3254 drm_atomic_crtc_state_for_each_plane_state(plane
, pstate
, cstate
) {
3255 enum plane_id plane_id
= to_intel_plane(plane
)->id
;
3259 rate
= skl_plane_relative_data_rate(intel_cstate
,
3261 plane_data_rate
[plane_id
] = rate
;
3263 total_data_rate
+= rate
;
3266 rate
= skl_plane_relative_data_rate(intel_cstate
,
3268 plane_y_data_rate
[plane_id
] = rate
;
3270 total_data_rate
+= rate
;
3273 return total_data_rate
;
3277 skl_ddb_min_alloc(const struct drm_plane_state
*pstate
,
3280 struct drm_framebuffer
*fb
= pstate
->fb
;
3281 struct intel_plane_state
*intel_pstate
= to_intel_plane_state(pstate
);
3282 uint32_t src_w
, src_h
;
3283 uint32_t min_scanlines
= 8;
3289 /* For packed formats, no y-plane, return 0 */
3290 if (y
&& fb
->format
->format
!= DRM_FORMAT_NV12
)
3293 /* For Non Y-tile return 8-blocks */
3294 if (fb
->modifier
!= I915_FORMAT_MOD_Y_TILED
&&
3295 fb
->modifier
!= I915_FORMAT_MOD_Yf_TILED
)
3298 src_w
= drm_rect_width(&intel_pstate
->base
.src
) >> 16;
3299 src_h
= drm_rect_height(&intel_pstate
->base
.src
) >> 16;
3301 if (drm_rotation_90_or_270(pstate
->rotation
))
3304 /* Halve UV plane width and height for NV12 */
3305 if (fb
->format
->format
== DRM_FORMAT_NV12
&& !y
) {
3310 if (fb
->format
->format
== DRM_FORMAT_NV12
&& !y
)
3311 plane_bpp
= fb
->format
->cpp
[1];
3313 plane_bpp
= fb
->format
->cpp
[0];
3315 if (drm_rotation_90_or_270(pstate
->rotation
)) {
3316 switch (plane_bpp
) {
3330 WARN(1, "Unsupported pixel depth %u for rotation",
3336 return DIV_ROUND_UP((4 * src_w
* plane_bpp
), 512) * min_scanlines
/4 + 3;
3340 skl_ddb_calc_min(const struct intel_crtc_state
*cstate
, int num_active
,
3341 uint16_t *minimum
, uint16_t *y_minimum
)
3343 const struct drm_plane_state
*pstate
;
3344 struct drm_plane
*plane
;
3346 drm_atomic_crtc_state_for_each_plane_state(plane
, pstate
, &cstate
->base
) {
3347 enum plane_id plane_id
= to_intel_plane(plane
)->id
;
3349 if (plane_id
== PLANE_CURSOR
)
3352 if (!pstate
->visible
)
3355 minimum
[plane_id
] = skl_ddb_min_alloc(pstate
, 0);
3356 y_minimum
[plane_id
] = skl_ddb_min_alloc(pstate
, 1);
3359 minimum
[PLANE_CURSOR
] = skl_cursor_allocation(num_active
);
3363 skl_allocate_pipe_ddb(struct intel_crtc_state
*cstate
,
3364 struct skl_ddb_allocation
*ddb
/* out */)
3366 struct drm_atomic_state
*state
= cstate
->base
.state
;
3367 struct drm_crtc
*crtc
= cstate
->base
.crtc
;
3368 struct drm_device
*dev
= crtc
->dev
;
3369 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3370 enum pipe pipe
= intel_crtc
->pipe
;
3371 struct skl_ddb_entry
*alloc
= &cstate
->wm
.skl
.ddb
;
3372 uint16_t alloc_size
, start
;
3373 uint16_t minimum
[I915_MAX_PLANES
] = {};
3374 uint16_t y_minimum
[I915_MAX_PLANES
] = {};
3375 unsigned int total_data_rate
;
3376 enum plane_id plane_id
;
3378 unsigned plane_data_rate
[I915_MAX_PLANES
] = {};
3379 unsigned plane_y_data_rate
[I915_MAX_PLANES
] = {};
3381 /* Clear the partitioning for disabled planes. */
3382 memset(ddb
->plane
[pipe
], 0, sizeof(ddb
->plane
[pipe
]));
3383 memset(ddb
->y_plane
[pipe
], 0, sizeof(ddb
->y_plane
[pipe
]));
3385 if (WARN_ON(!state
))
3388 if (!cstate
->base
.active
) {
3389 alloc
->start
= alloc
->end
= 0;
3393 skl_ddb_get_pipe_allocation_limits(dev
, cstate
, alloc
, &num_active
);
3394 alloc_size
= skl_ddb_entry_size(alloc
);
3395 if (alloc_size
== 0) {
3396 memset(ddb
->plane
[pipe
], 0, sizeof(ddb
->plane
[pipe
]));
3400 skl_ddb_calc_min(cstate
, num_active
, minimum
, y_minimum
);
3403 * 1. Allocate the mininum required blocks for each active plane
3404 * and allocate the cursor, it doesn't require extra allocation
3405 * proportional to the data rate.
3408 for_each_plane_id_on_crtc(intel_crtc
, plane_id
) {
3409 alloc_size
-= minimum
[plane_id
];
3410 alloc_size
-= y_minimum
[plane_id
];
3413 ddb
->plane
[pipe
][PLANE_CURSOR
].start
= alloc
->end
- minimum
[PLANE_CURSOR
];
3414 ddb
->plane
[pipe
][PLANE_CURSOR
].end
= alloc
->end
;
3417 * 2. Distribute the remaining space in proportion to the amount of
3418 * data each plane needs to fetch from memory.
3420 * FIXME: we may not allocate every single block here.
3422 total_data_rate
= skl_get_total_relative_data_rate(cstate
,
3425 if (total_data_rate
== 0)
3428 start
= alloc
->start
;
3429 for_each_plane_id_on_crtc(intel_crtc
, plane_id
) {
3430 unsigned int data_rate
, y_data_rate
;
3431 uint16_t plane_blocks
, y_plane_blocks
= 0;
3433 if (plane_id
== PLANE_CURSOR
)
3436 data_rate
= plane_data_rate
[plane_id
];
3439 * allocation for (packed formats) or (uv-plane part of planar format):
3440 * promote the expression to 64 bits to avoid overflowing, the
3441 * result is < available as data_rate / total_data_rate < 1
3443 plane_blocks
= minimum
[plane_id
];
3444 plane_blocks
+= div_u64((uint64_t)alloc_size
* data_rate
,
3447 /* Leave disabled planes at (0,0) */
3449 ddb
->plane
[pipe
][plane_id
].start
= start
;
3450 ddb
->plane
[pipe
][plane_id
].end
= start
+ plane_blocks
;
3453 start
+= plane_blocks
;
3456 * allocation for y_plane part of planar format:
3458 y_data_rate
= plane_y_data_rate
[plane_id
];
3460 y_plane_blocks
= y_minimum
[plane_id
];
3461 y_plane_blocks
+= div_u64((uint64_t)alloc_size
* y_data_rate
,
3465 ddb
->y_plane
[pipe
][plane_id
].start
= start
;
3466 ddb
->y_plane
[pipe
][plane_id
].end
= start
+ y_plane_blocks
;
3469 start
+= y_plane_blocks
;
3476 * The max latency should be 257 (max the punit can code is 255 and we add 2us
3477 * for the read latency) and cpp should always be <= 8, so that
3478 * should allow pixel_rate up to ~2 GHz which seems sufficient since max
3479 * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
3481 static uint_fixed_16_16_t
skl_wm_method1(uint32_t pixel_rate
, uint8_t cpp
,
3484 uint32_t wm_intermediate_val
;
3485 uint_fixed_16_16_t ret
;
3488 return FP_16_16_MAX
;
3490 wm_intermediate_val
= latency
* pixel_rate
* cpp
;
3491 ret
= fixed_16_16_div_round_up_u64(wm_intermediate_val
, 1000 * 512);
3495 static uint_fixed_16_16_t
skl_wm_method2(uint32_t pixel_rate
,
3496 uint32_t pipe_htotal
,
3498 uint_fixed_16_16_t plane_blocks_per_line
)
3500 uint32_t wm_intermediate_val
;
3501 uint_fixed_16_16_t ret
;
3504 return FP_16_16_MAX
;
3506 wm_intermediate_val
= latency
* pixel_rate
;
3507 wm_intermediate_val
= DIV_ROUND_UP(wm_intermediate_val
,
3508 pipe_htotal
* 1000);
3509 ret
= mul_u32_fixed_16_16(wm_intermediate_val
, plane_blocks_per_line
);
3513 static uint32_t skl_adjusted_plane_pixel_rate(const struct intel_crtc_state
*cstate
,
3514 struct intel_plane_state
*pstate
)
3516 uint64_t adjusted_pixel_rate
;
3517 uint64_t downscale_amount
;
3518 uint64_t pixel_rate
;
3520 /* Shouldn't reach here on disabled planes... */
3521 if (WARN_ON(!pstate
->base
.visible
))
3525 * Adjusted plane pixel rate is just the pipe's adjusted pixel rate
3526 * with additional adjustments for plane-specific scaling.
3528 adjusted_pixel_rate
= cstate
->pixel_rate
;
3529 downscale_amount
= skl_plane_downscale_amount(pstate
);
3531 pixel_rate
= adjusted_pixel_rate
* downscale_amount
>> 16;
3532 WARN_ON(pixel_rate
!= clamp_t(uint32_t, pixel_rate
, 0, ~0));
3537 static int skl_compute_plane_wm(const struct drm_i915_private
*dev_priv
,
3538 struct intel_crtc_state
*cstate
,
3539 struct intel_plane_state
*intel_pstate
,
3540 uint16_t ddb_allocation
,
3542 uint16_t *out_blocks
, /* out */
3543 uint8_t *out_lines
, /* out */
3544 bool *enabled
/* out */)
3546 struct drm_plane_state
*pstate
= &intel_pstate
->base
;
3547 struct drm_framebuffer
*fb
= pstate
->fb
;
3548 uint32_t latency
= dev_priv
->wm
.skl_latency
[level
];
3549 uint_fixed_16_16_t method1
, method2
;
3550 uint_fixed_16_16_t plane_blocks_per_line
;
3551 uint_fixed_16_16_t selected_result
;
3552 uint32_t interm_pbpl
;
3553 uint32_t plane_bytes_per_line
;
3554 uint32_t res_blocks
, res_lines
;
3556 uint32_t width
= 0, height
= 0;
3557 uint32_t plane_pixel_rate
;
3558 uint_fixed_16_16_t y_tile_minimum
;
3559 uint32_t y_min_scanlines
;
3560 struct intel_atomic_state
*state
=
3561 to_intel_atomic_state(cstate
->base
.state
);
3562 bool apply_memory_bw_wa
= skl_needs_memory_bw_wa(state
);
3563 bool y_tiled
, x_tiled
;
3565 if (latency
== 0 || !cstate
->base
.active
|| !intel_pstate
->base
.visible
) {
3570 y_tiled
= fb
->modifier
== I915_FORMAT_MOD_Y_TILED
||
3571 fb
->modifier
== I915_FORMAT_MOD_Yf_TILED
;
3572 x_tiled
= fb
->modifier
== I915_FORMAT_MOD_X_TILED
;
3574 /* Display WA #1141: kbl. */
3575 if (IS_KABYLAKE(dev_priv
) && dev_priv
->ipc_enabled
)
3578 if (apply_memory_bw_wa
&& x_tiled
)
3581 width
= drm_rect_width(&intel_pstate
->base
.src
) >> 16;
3582 height
= drm_rect_height(&intel_pstate
->base
.src
) >> 16;
3584 if (drm_rotation_90_or_270(pstate
->rotation
))
3585 swap(width
, height
);
3587 cpp
= fb
->format
->cpp
[0];
3588 plane_pixel_rate
= skl_adjusted_plane_pixel_rate(cstate
, intel_pstate
);
3590 if (drm_rotation_90_or_270(pstate
->rotation
)) {
3591 int cpp
= (fb
->format
->format
== DRM_FORMAT_NV12
) ?
3592 fb
->format
->cpp
[1] :
3597 y_min_scanlines
= 16;
3600 y_min_scanlines
= 8;
3603 y_min_scanlines
= 4;
3610 y_min_scanlines
= 4;
3613 if (apply_memory_bw_wa
)
3614 y_min_scanlines
*= 2;
3616 plane_bytes_per_line
= width
* cpp
;
3618 interm_pbpl
= DIV_ROUND_UP(plane_bytes_per_line
*
3619 y_min_scanlines
, 512);
3620 plane_blocks_per_line
=
3621 fixed_16_16_div_round_up(interm_pbpl
, y_min_scanlines
);
3622 } else if (x_tiled
) {
3623 interm_pbpl
= DIV_ROUND_UP(plane_bytes_per_line
, 512);
3624 plane_blocks_per_line
= u32_to_fixed_16_16(interm_pbpl
);
3626 interm_pbpl
= DIV_ROUND_UP(plane_bytes_per_line
, 512) + 1;
3627 plane_blocks_per_line
= u32_to_fixed_16_16(interm_pbpl
);
3630 method1
= skl_wm_method1(plane_pixel_rate
, cpp
, latency
);
3631 method2
= skl_wm_method2(plane_pixel_rate
,
3632 cstate
->base
.adjusted_mode
.crtc_htotal
,
3634 plane_blocks_per_line
);
3636 y_tile_minimum
= mul_u32_fixed_16_16(y_min_scanlines
,
3637 plane_blocks_per_line
);
3640 selected_result
= max_fixed_16_16(method2
, y_tile_minimum
);
3642 if ((cpp
* cstate
->base
.adjusted_mode
.crtc_htotal
/ 512 < 1) &&
3643 (plane_bytes_per_line
/ 512 < 1))
3644 selected_result
= method2
;
3645 else if ((ddb_allocation
/
3646 fixed_16_16_to_u32_round_up(plane_blocks_per_line
)) >= 1)
3647 selected_result
= min_fixed_16_16(method1
, method2
);
3649 selected_result
= method1
;
3652 res_blocks
= fixed_16_16_to_u32_round_up(selected_result
) + 1;
3653 res_lines
= DIV_ROUND_UP(selected_result
.val
,
3654 plane_blocks_per_line
.val
);
3656 if (level
>= 1 && level
<= 7) {
3658 res_blocks
+= fixed_16_16_to_u32_round_up(y_tile_minimum
);
3659 res_lines
+= y_min_scanlines
;
3665 if (res_blocks
>= ddb_allocation
|| res_lines
> 31) {
3669 * If there are no valid level 0 watermarks, then we can't
3670 * support this display configuration.
3675 struct drm_plane
*plane
= pstate
->plane
;
3677 DRM_DEBUG_KMS("Requested display configuration exceeds system watermark limitations\n");
3678 DRM_DEBUG_KMS("[PLANE:%d:%s] blocks required = %u/%u, lines required = %u/31\n",
3679 plane
->base
.id
, plane
->name
,
3680 res_blocks
, ddb_allocation
, res_lines
);
3685 *out_blocks
= res_blocks
;
3686 *out_lines
= res_lines
;
3693 skl_compute_wm_level(const struct drm_i915_private
*dev_priv
,
3694 struct skl_ddb_allocation
*ddb
,
3695 struct intel_crtc_state
*cstate
,
3696 struct intel_plane
*intel_plane
,
3698 struct skl_wm_level
*result
)
3700 struct drm_atomic_state
*state
= cstate
->base
.state
;
3701 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
3702 struct drm_plane
*plane
= &intel_plane
->base
;
3703 struct intel_plane_state
*intel_pstate
= NULL
;
3704 uint16_t ddb_blocks
;
3705 enum pipe pipe
= intel_crtc
->pipe
;
3710 intel_atomic_get_existing_plane_state(state
,
3714 * Note: If we start supporting multiple pending atomic commits against
3715 * the same planes/CRTC's in the future, plane->state will no longer be
3716 * the correct pre-state to use for the calculations here and we'll
3717 * need to change where we get the 'unchanged' plane data from.
3719 * For now this is fine because we only allow one queued commit against
3720 * a CRTC. Even if the plane isn't modified by this transaction and we
3721 * don't have a plane lock, we still have the CRTC's lock, so we know
3722 * that no other transactions are racing with us to update it.
3725 intel_pstate
= to_intel_plane_state(plane
->state
);
3727 WARN_ON(!intel_pstate
->base
.fb
);
3729 ddb_blocks
= skl_ddb_entry_size(&ddb
->plane
[pipe
][intel_plane
->id
]);
3731 ret
= skl_compute_plane_wm(dev_priv
,
3736 &result
->plane_res_b
,
3737 &result
->plane_res_l
,
3746 skl_compute_linetime_wm(struct intel_crtc_state
*cstate
)
3748 struct drm_atomic_state
*state
= cstate
->base
.state
;
3749 struct drm_i915_private
*dev_priv
= to_i915(state
->dev
);
3750 uint32_t pixel_rate
;
3751 uint32_t linetime_wm
;
3753 if (!cstate
->base
.active
)
3756 pixel_rate
= cstate
->pixel_rate
;
3758 if (WARN_ON(pixel_rate
== 0))
3761 linetime_wm
= DIV_ROUND_UP(8 * cstate
->base
.adjusted_mode
.crtc_htotal
*
3764 /* Display WA #1135: bxt. */
3765 if (IS_BROXTON(dev_priv
) && dev_priv
->ipc_enabled
)
3766 linetime_wm
= DIV_ROUND_UP(linetime_wm
, 2);
3771 static void skl_compute_transition_wm(struct intel_crtc_state
*cstate
,
3772 struct skl_wm_level
*trans_wm
/* out */)
3774 if (!cstate
->base
.active
)
3777 /* Until we know more, just disable transition WMs */
3778 trans_wm
->plane_en
= false;
3781 static int skl_build_pipe_wm(struct intel_crtc_state
*cstate
,
3782 struct skl_ddb_allocation
*ddb
,
3783 struct skl_pipe_wm
*pipe_wm
)
3785 struct drm_device
*dev
= cstate
->base
.crtc
->dev
;
3786 const struct drm_i915_private
*dev_priv
= to_i915(dev
);
3787 struct intel_plane
*intel_plane
;
3788 struct skl_plane_wm
*wm
;
3789 int level
, max_level
= ilk_wm_max_level(dev_priv
);
3793 * We'll only calculate watermarks for planes that are actually
3794 * enabled, so make sure all other planes are set as disabled.
3796 memset(pipe_wm
->planes
, 0, sizeof(pipe_wm
->planes
));
3798 for_each_intel_plane_mask(&dev_priv
->drm
,
3800 cstate
->base
.plane_mask
) {
3801 wm
= &pipe_wm
->planes
[intel_plane
->id
];
3803 for (level
= 0; level
<= max_level
; level
++) {
3804 ret
= skl_compute_wm_level(dev_priv
, ddb
, cstate
,
3810 skl_compute_transition_wm(cstate
, &wm
->trans_wm
);
3812 pipe_wm
->linetime
= skl_compute_linetime_wm(cstate
);
3817 static void skl_ddb_entry_write(struct drm_i915_private
*dev_priv
,
3819 const struct skl_ddb_entry
*entry
)
3822 I915_WRITE(reg
, (entry
->end
- 1) << 16 | entry
->start
);
3827 static void skl_write_wm_level(struct drm_i915_private
*dev_priv
,
3829 const struct skl_wm_level
*level
)
3833 if (level
->plane_en
) {
3835 val
|= level
->plane_res_b
;
3836 val
|= level
->plane_res_l
<< PLANE_WM_LINES_SHIFT
;
3839 I915_WRITE(reg
, val
);
3842 static void skl_write_plane_wm(struct intel_crtc
*intel_crtc
,
3843 const struct skl_plane_wm
*wm
,
3844 const struct skl_ddb_allocation
*ddb
,
3845 enum plane_id plane_id
)
3847 struct drm_crtc
*crtc
= &intel_crtc
->base
;
3848 struct drm_device
*dev
= crtc
->dev
;
3849 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3850 int level
, max_level
= ilk_wm_max_level(dev_priv
);
3851 enum pipe pipe
= intel_crtc
->pipe
;
3853 for (level
= 0; level
<= max_level
; level
++) {
3854 skl_write_wm_level(dev_priv
, PLANE_WM(pipe
, plane_id
, level
),
3857 skl_write_wm_level(dev_priv
, PLANE_WM_TRANS(pipe
, plane_id
),
3860 skl_ddb_entry_write(dev_priv
, PLANE_BUF_CFG(pipe
, plane_id
),
3861 &ddb
->plane
[pipe
][plane_id
]);
3862 skl_ddb_entry_write(dev_priv
, PLANE_NV12_BUF_CFG(pipe
, plane_id
),
3863 &ddb
->y_plane
[pipe
][plane_id
]);
3866 static void skl_write_cursor_wm(struct intel_crtc
*intel_crtc
,
3867 const struct skl_plane_wm
*wm
,
3868 const struct skl_ddb_allocation
*ddb
)
3870 struct drm_crtc
*crtc
= &intel_crtc
->base
;
3871 struct drm_device
*dev
= crtc
->dev
;
3872 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3873 int level
, max_level
= ilk_wm_max_level(dev_priv
);
3874 enum pipe pipe
= intel_crtc
->pipe
;
3876 for (level
= 0; level
<= max_level
; level
++) {
3877 skl_write_wm_level(dev_priv
, CUR_WM(pipe
, level
),
3880 skl_write_wm_level(dev_priv
, CUR_WM_TRANS(pipe
), &wm
->trans_wm
);
3882 skl_ddb_entry_write(dev_priv
, CUR_BUF_CFG(pipe
),
3883 &ddb
->plane
[pipe
][PLANE_CURSOR
]);
3886 bool skl_wm_level_equals(const struct skl_wm_level
*l1
,
3887 const struct skl_wm_level
*l2
)
3889 if (l1
->plane_en
!= l2
->plane_en
)
3892 /* If both planes aren't enabled, the rest shouldn't matter */
3896 return (l1
->plane_res_l
== l2
->plane_res_l
&&
3897 l1
->plane_res_b
== l2
->plane_res_b
);
3900 static inline bool skl_ddb_entries_overlap(const struct skl_ddb_entry
*a
,
3901 const struct skl_ddb_entry
*b
)
3903 return a
->start
< b
->end
&& b
->start
< a
->end
;
3906 bool skl_ddb_allocation_overlaps(const struct skl_ddb_entry
**entries
,
3907 const struct skl_ddb_entry
*ddb
,
3912 for (i
= 0; i
< I915_MAX_PIPES
; i
++)
3913 if (i
!= ignore
&& entries
[i
] &&
3914 skl_ddb_entries_overlap(ddb
, entries
[i
]))
3920 static int skl_update_pipe_wm(struct drm_crtc_state
*cstate
,
3921 const struct skl_pipe_wm
*old_pipe_wm
,
3922 struct skl_pipe_wm
*pipe_wm
, /* out */
3923 struct skl_ddb_allocation
*ddb
, /* out */
3924 bool *changed
/* out */)
3926 struct intel_crtc_state
*intel_cstate
= to_intel_crtc_state(cstate
);
3929 ret
= skl_build_pipe_wm(intel_cstate
, ddb
, pipe_wm
);
3933 if (!memcmp(old_pipe_wm
, pipe_wm
, sizeof(*pipe_wm
)))
3942 pipes_modified(struct drm_atomic_state
*state
)
3944 struct drm_crtc
*crtc
;
3945 struct drm_crtc_state
*cstate
;
3946 uint32_t i
, ret
= 0;
3948 for_each_crtc_in_state(state
, crtc
, cstate
, i
)
3949 ret
|= drm_crtc_mask(crtc
);
3955 skl_ddb_add_affected_planes(struct intel_crtc_state
*cstate
)
3957 struct drm_atomic_state
*state
= cstate
->base
.state
;
3958 struct drm_device
*dev
= state
->dev
;
3959 struct drm_crtc
*crtc
= cstate
->base
.crtc
;
3960 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
3961 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3962 struct intel_atomic_state
*intel_state
= to_intel_atomic_state(state
);
3963 struct skl_ddb_allocation
*new_ddb
= &intel_state
->wm_results
.ddb
;
3964 struct skl_ddb_allocation
*cur_ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
3965 struct drm_plane_state
*plane_state
;
3966 struct drm_plane
*plane
;
3967 enum pipe pipe
= intel_crtc
->pipe
;
3969 WARN_ON(!drm_atomic_get_existing_crtc_state(state
, crtc
));
3971 drm_for_each_plane_mask(plane
, dev
, cstate
->base
.plane_mask
) {
3972 enum plane_id plane_id
= to_intel_plane(plane
)->id
;
3974 if (skl_ddb_entry_equal(&cur_ddb
->plane
[pipe
][plane_id
],
3975 &new_ddb
->plane
[pipe
][plane_id
]) &&
3976 skl_ddb_entry_equal(&cur_ddb
->y_plane
[pipe
][plane_id
],
3977 &new_ddb
->y_plane
[pipe
][plane_id
]))
3980 plane_state
= drm_atomic_get_plane_state(state
, plane
);
3981 if (IS_ERR(plane_state
))
3982 return PTR_ERR(plane_state
);
3989 skl_compute_ddb(struct drm_atomic_state
*state
)
3991 struct drm_device
*dev
= state
->dev
;
3992 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3993 struct intel_atomic_state
*intel_state
= to_intel_atomic_state(state
);
3994 struct intel_crtc
*intel_crtc
;
3995 struct skl_ddb_allocation
*ddb
= &intel_state
->wm_results
.ddb
;
3996 uint32_t realloc_pipes
= pipes_modified(state
);
4000 * If this is our first atomic update following hardware readout,
4001 * we can't trust the DDB that the BIOS programmed for us. Let's
4002 * pretend that all pipes switched active status so that we'll
4003 * ensure a full DDB recompute.
4005 if (dev_priv
->wm
.distrust_bios_wm
) {
4006 ret
= drm_modeset_lock(&dev
->mode_config
.connection_mutex
,
4007 state
->acquire_ctx
);
4011 intel_state
->active_pipe_changes
= ~0;
4014 * We usually only initialize intel_state->active_crtcs if we
4015 * we're doing a modeset; make sure this field is always
4016 * initialized during the sanitization process that happens
4017 * on the first commit too.
4019 if (!intel_state
->modeset
)
4020 intel_state
->active_crtcs
= dev_priv
->active_crtcs
;
4024 * If the modeset changes which CRTC's are active, we need to
4025 * recompute the DDB allocation for *all* active pipes, even
4026 * those that weren't otherwise being modified in any way by this
4027 * atomic commit. Due to the shrinking of the per-pipe allocations
4028 * when new active CRTC's are added, it's possible for a pipe that
4029 * we were already using and aren't changing at all here to suddenly
4030 * become invalid if its DDB needs exceeds its new allocation.
4032 * Note that if we wind up doing a full DDB recompute, we can't let
4033 * any other display updates race with this transaction, so we need
4034 * to grab the lock on *all* CRTC's.
4036 if (intel_state
->active_pipe_changes
) {
4038 intel_state
->wm_results
.dirty_pipes
= ~0;
4042 * We're not recomputing for the pipes not included in the commit, so
4043 * make sure we start with the current state.
4045 memcpy(ddb
, &dev_priv
->wm
.skl_hw
.ddb
, sizeof(*ddb
));
4047 for_each_intel_crtc_mask(dev
, intel_crtc
, realloc_pipes
) {
4048 struct intel_crtc_state
*cstate
;
4050 cstate
= intel_atomic_get_crtc_state(state
, intel_crtc
);
4052 return PTR_ERR(cstate
);
4054 ret
= skl_allocate_pipe_ddb(cstate
, ddb
);
4058 ret
= skl_ddb_add_affected_planes(cstate
);
4067 skl_copy_wm_for_pipe(struct skl_wm_values
*dst
,
4068 struct skl_wm_values
*src
,
4071 memcpy(dst
->ddb
.y_plane
[pipe
], src
->ddb
.y_plane
[pipe
],
4072 sizeof(dst
->ddb
.y_plane
[pipe
]));
4073 memcpy(dst
->ddb
.plane
[pipe
], src
->ddb
.plane
[pipe
],
4074 sizeof(dst
->ddb
.plane
[pipe
]));
4078 skl_print_wm_changes(const struct drm_atomic_state
*state
)
4080 const struct drm_device
*dev
= state
->dev
;
4081 const struct drm_i915_private
*dev_priv
= to_i915(dev
);
4082 const struct intel_atomic_state
*intel_state
=
4083 to_intel_atomic_state(state
);
4084 const struct drm_crtc
*crtc
;
4085 const struct drm_crtc_state
*cstate
;
4086 const struct intel_plane
*intel_plane
;
4087 const struct skl_ddb_allocation
*old_ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
4088 const struct skl_ddb_allocation
*new_ddb
= &intel_state
->wm_results
.ddb
;
4091 for_each_crtc_in_state(state
, crtc
, cstate
, i
) {
4092 const struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
4093 enum pipe pipe
= intel_crtc
->pipe
;
4095 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
4096 enum plane_id plane_id
= intel_plane
->id
;
4097 const struct skl_ddb_entry
*old
, *new;
4099 old
= &old_ddb
->plane
[pipe
][plane_id
];
4100 new = &new_ddb
->plane
[pipe
][plane_id
];
4102 if (skl_ddb_entry_equal(old
, new))
4105 DRM_DEBUG_ATOMIC("[PLANE:%d:%s] ddb (%d - %d) -> (%d - %d)\n",
4106 intel_plane
->base
.base
.id
,
4107 intel_plane
->base
.name
,
4108 old
->start
, old
->end
,
4109 new->start
, new->end
);
4115 skl_compute_wm(struct drm_atomic_state
*state
)
4117 struct drm_crtc
*crtc
;
4118 struct drm_crtc_state
*cstate
;
4119 struct intel_atomic_state
*intel_state
= to_intel_atomic_state(state
);
4120 struct skl_wm_values
*results
= &intel_state
->wm_results
;
4121 struct skl_pipe_wm
*pipe_wm
;
4122 bool changed
= false;
4126 * If this transaction isn't actually touching any CRTC's, don't
4127 * bother with watermark calculation. Note that if we pass this
4128 * test, we're guaranteed to hold at least one CRTC state mutex,
4129 * which means we can safely use values like dev_priv->active_crtcs
4130 * since any racing commits that want to update them would need to
4131 * hold _all_ CRTC state mutexes.
4133 for_each_crtc_in_state(state
, crtc
, cstate
, i
)
4138 /* Clear all dirty flags */
4139 results
->dirty_pipes
= 0;
4141 ret
= skl_compute_ddb(state
);
4146 * Calculate WM's for all pipes that are part of this transaction.
4147 * Note that the DDB allocation above may have added more CRTC's that
4148 * weren't otherwise being modified (and set bits in dirty_pipes) if
4149 * pipe allocations had to change.
4151 * FIXME: Now that we're doing this in the atomic check phase, we
4152 * should allow skl_update_pipe_wm() to return failure in cases where
4153 * no suitable watermark values can be found.
4155 for_each_crtc_in_state(state
, crtc
, cstate
, i
) {
4156 struct intel_crtc_state
*intel_cstate
=
4157 to_intel_crtc_state(cstate
);
4158 const struct skl_pipe_wm
*old_pipe_wm
=
4159 &to_intel_crtc_state(crtc
->state
)->wm
.skl
.optimal
;
4161 pipe_wm
= &intel_cstate
->wm
.skl
.optimal
;
4162 ret
= skl_update_pipe_wm(cstate
, old_pipe_wm
, pipe_wm
,
4163 &results
->ddb
, &changed
);
4168 results
->dirty_pipes
|= drm_crtc_mask(crtc
);
4170 if ((results
->dirty_pipes
& drm_crtc_mask(crtc
)) == 0)
4171 /* This pipe's WM's did not change */
4174 intel_cstate
->update_wm_pre
= true;
4177 skl_print_wm_changes(state
);
4182 static void skl_atomic_update_crtc_wm(struct intel_atomic_state
*state
,
4183 struct intel_crtc_state
*cstate
)
4185 struct intel_crtc
*crtc
= to_intel_crtc(cstate
->base
.crtc
);
4186 struct drm_i915_private
*dev_priv
= to_i915(state
->base
.dev
);
4187 struct skl_pipe_wm
*pipe_wm
= &cstate
->wm
.skl
.optimal
;
4188 const struct skl_ddb_allocation
*ddb
= &state
->wm_results
.ddb
;
4189 enum pipe pipe
= crtc
->pipe
;
4190 enum plane_id plane_id
;
4192 if (!(state
->wm_results
.dirty_pipes
& drm_crtc_mask(&crtc
->base
)))
4195 I915_WRITE(PIPE_WM_LINETIME(pipe
), pipe_wm
->linetime
);
4197 for_each_plane_id_on_crtc(crtc
, plane_id
) {
4198 if (plane_id
!= PLANE_CURSOR
)
4199 skl_write_plane_wm(crtc
, &pipe_wm
->planes
[plane_id
],
4202 skl_write_cursor_wm(crtc
, &pipe_wm
->planes
[plane_id
],
4207 static void skl_initial_wm(struct intel_atomic_state
*state
,
4208 struct intel_crtc_state
*cstate
)
4210 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
4211 struct drm_device
*dev
= intel_crtc
->base
.dev
;
4212 struct drm_i915_private
*dev_priv
= to_i915(dev
);
4213 struct skl_wm_values
*results
= &state
->wm_results
;
4214 struct skl_wm_values
*hw_vals
= &dev_priv
->wm
.skl_hw
;
4215 enum pipe pipe
= intel_crtc
->pipe
;
4217 if ((results
->dirty_pipes
& drm_crtc_mask(&intel_crtc
->base
)) == 0)
4220 mutex_lock(&dev_priv
->wm
.wm_mutex
);
4222 if (cstate
->base
.active_changed
)
4223 skl_atomic_update_crtc_wm(state
, cstate
);
4225 skl_copy_wm_for_pipe(hw_vals
, results
, pipe
);
4227 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
4230 static void ilk_compute_wm_config(struct drm_device
*dev
,
4231 struct intel_wm_config
*config
)
4233 struct intel_crtc
*crtc
;
4235 /* Compute the currently _active_ config */
4236 for_each_intel_crtc(dev
, crtc
) {
4237 const struct intel_pipe_wm
*wm
= &crtc
->wm
.active
.ilk
;
4239 if (!wm
->pipe_enabled
)
4242 config
->sprites_enabled
|= wm
->sprites_enabled
;
4243 config
->sprites_scaled
|= wm
->sprites_scaled
;
4244 config
->num_pipes_active
++;
4248 static void ilk_program_watermarks(struct drm_i915_private
*dev_priv
)
4250 struct drm_device
*dev
= &dev_priv
->drm
;
4251 struct intel_pipe_wm lp_wm_1_2
= {}, lp_wm_5_6
= {}, *best_lp_wm
;
4252 struct ilk_wm_maximums max
;
4253 struct intel_wm_config config
= {};
4254 struct ilk_wm_values results
= {};
4255 enum intel_ddb_partitioning partitioning
;
4257 ilk_compute_wm_config(dev
, &config
);
4259 ilk_compute_wm_maximums(dev
, 1, &config
, INTEL_DDB_PART_1_2
, &max
);
4260 ilk_wm_merge(dev
, &config
, &max
, &lp_wm_1_2
);
4262 /* 5/6 split only in single pipe config on IVB+ */
4263 if (INTEL_GEN(dev_priv
) >= 7 &&
4264 config
.num_pipes_active
== 1 && config
.sprites_enabled
) {
4265 ilk_compute_wm_maximums(dev
, 1, &config
, INTEL_DDB_PART_5_6
, &max
);
4266 ilk_wm_merge(dev
, &config
, &max
, &lp_wm_5_6
);
4268 best_lp_wm
= ilk_find_best_result(dev
, &lp_wm_1_2
, &lp_wm_5_6
);
4270 best_lp_wm
= &lp_wm_1_2
;
4273 partitioning
= (best_lp_wm
== &lp_wm_1_2
) ?
4274 INTEL_DDB_PART_1_2
: INTEL_DDB_PART_5_6
;
4276 ilk_compute_wm_results(dev
, best_lp_wm
, partitioning
, &results
);
4278 ilk_write_wm_values(dev_priv
, &results
);
4281 static void ilk_initial_watermarks(struct intel_atomic_state
*state
,
4282 struct intel_crtc_state
*cstate
)
4284 struct drm_i915_private
*dev_priv
= to_i915(cstate
->base
.crtc
->dev
);
4285 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
4287 mutex_lock(&dev_priv
->wm
.wm_mutex
);
4288 intel_crtc
->wm
.active
.ilk
= cstate
->wm
.ilk
.intermediate
;
4289 ilk_program_watermarks(dev_priv
);
4290 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
4293 static void ilk_optimize_watermarks(struct intel_atomic_state
*state
,
4294 struct intel_crtc_state
*cstate
)
4296 struct drm_i915_private
*dev_priv
= to_i915(cstate
->base
.crtc
->dev
);
4297 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
4299 mutex_lock(&dev_priv
->wm
.wm_mutex
);
4300 if (cstate
->wm
.need_postvbl_update
) {
4301 intel_crtc
->wm
.active
.ilk
= cstate
->wm
.ilk
.optimal
;
4302 ilk_program_watermarks(dev_priv
);
4304 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
4307 static inline void skl_wm_level_from_reg_val(uint32_t val
,
4308 struct skl_wm_level
*level
)
4310 level
->plane_en
= val
& PLANE_WM_EN
;
4311 level
->plane_res_b
= val
& PLANE_WM_BLOCKS_MASK
;
4312 level
->plane_res_l
= (val
>> PLANE_WM_LINES_SHIFT
) &
4313 PLANE_WM_LINES_MASK
;
4316 void skl_pipe_wm_get_hw_state(struct drm_crtc
*crtc
,
4317 struct skl_pipe_wm
*out
)
4319 struct drm_i915_private
*dev_priv
= to_i915(crtc
->dev
);
4320 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
4321 enum pipe pipe
= intel_crtc
->pipe
;
4322 int level
, max_level
;
4323 enum plane_id plane_id
;
4326 max_level
= ilk_wm_max_level(dev_priv
);
4328 for_each_plane_id_on_crtc(intel_crtc
, plane_id
) {
4329 struct skl_plane_wm
*wm
= &out
->planes
[plane_id
];
4331 for (level
= 0; level
<= max_level
; level
++) {
4332 if (plane_id
!= PLANE_CURSOR
)
4333 val
= I915_READ(PLANE_WM(pipe
, plane_id
, level
));
4335 val
= I915_READ(CUR_WM(pipe
, level
));
4337 skl_wm_level_from_reg_val(val
, &wm
->wm
[level
]);
4340 if (plane_id
!= PLANE_CURSOR
)
4341 val
= I915_READ(PLANE_WM_TRANS(pipe
, plane_id
));
4343 val
= I915_READ(CUR_WM_TRANS(pipe
));
4345 skl_wm_level_from_reg_val(val
, &wm
->trans_wm
);
4348 if (!intel_crtc
->active
)
4351 out
->linetime
= I915_READ(PIPE_WM_LINETIME(pipe
));
4354 void skl_wm_get_hw_state(struct drm_device
*dev
)
4356 struct drm_i915_private
*dev_priv
= to_i915(dev
);
4357 struct skl_wm_values
*hw
= &dev_priv
->wm
.skl_hw
;
4358 struct skl_ddb_allocation
*ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
4359 struct drm_crtc
*crtc
;
4360 struct intel_crtc
*intel_crtc
;
4361 struct intel_crtc_state
*cstate
;
4363 skl_ddb_get_hw_state(dev_priv
, ddb
);
4364 list_for_each_entry(crtc
, &dev
->mode_config
.crtc_list
, head
) {
4365 intel_crtc
= to_intel_crtc(crtc
);
4366 cstate
= to_intel_crtc_state(crtc
->state
);
4368 skl_pipe_wm_get_hw_state(crtc
, &cstate
->wm
.skl
.optimal
);
4370 if (intel_crtc
->active
)
4371 hw
->dirty_pipes
|= drm_crtc_mask(crtc
);
4374 if (dev_priv
->active_crtcs
) {
4375 /* Fully recompute DDB on first atomic commit */
4376 dev_priv
->wm
.distrust_bios_wm
= true;
4378 /* Easy/common case; just sanitize DDB now if everything off */
4379 memset(ddb
, 0, sizeof(*ddb
));
4383 static void ilk_pipe_wm_get_hw_state(struct drm_crtc
*crtc
)
4385 struct drm_device
*dev
= crtc
->dev
;
4386 struct drm_i915_private
*dev_priv
= to_i915(dev
);
4387 struct ilk_wm_values
*hw
= &dev_priv
->wm
.hw
;
4388 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
4389 struct intel_crtc_state
*cstate
= to_intel_crtc_state(crtc
->state
);
4390 struct intel_pipe_wm
*active
= &cstate
->wm
.ilk
.optimal
;
4391 enum pipe pipe
= intel_crtc
->pipe
;
4392 static const i915_reg_t wm0_pipe_reg
[] = {
4393 [PIPE_A
] = WM0_PIPEA_ILK
,
4394 [PIPE_B
] = WM0_PIPEB_ILK
,
4395 [PIPE_C
] = WM0_PIPEC_IVB
,
4398 hw
->wm_pipe
[pipe
] = I915_READ(wm0_pipe_reg
[pipe
]);
4399 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
4400 hw
->wm_linetime
[pipe
] = I915_READ(PIPE_WM_LINETIME(pipe
));
4402 memset(active
, 0, sizeof(*active
));
4404 active
->pipe_enabled
= intel_crtc
->active
;
4406 if (active
->pipe_enabled
) {
4407 u32 tmp
= hw
->wm_pipe
[pipe
];
4410 * For active pipes LP0 watermark is marked as
4411 * enabled, and LP1+ watermaks as disabled since
4412 * we can't really reverse compute them in case
4413 * multiple pipes are active.
4415 active
->wm
[0].enable
= true;
4416 active
->wm
[0].pri_val
= (tmp
& WM0_PIPE_PLANE_MASK
) >> WM0_PIPE_PLANE_SHIFT
;
4417 active
->wm
[0].spr_val
= (tmp
& WM0_PIPE_SPRITE_MASK
) >> WM0_PIPE_SPRITE_SHIFT
;
4418 active
->wm
[0].cur_val
= tmp
& WM0_PIPE_CURSOR_MASK
;
4419 active
->linetime
= hw
->wm_linetime
[pipe
];
4421 int level
, max_level
= ilk_wm_max_level(dev_priv
);
4424 * For inactive pipes, all watermark levels
4425 * should be marked as enabled but zeroed,
4426 * which is what we'd compute them to.
4428 for (level
= 0; level
<= max_level
; level
++)
4429 active
->wm
[level
].enable
= true;
4432 intel_crtc
->wm
.active
.ilk
= *active
;
4435 #define _FW_WM(value, plane) \
4436 (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
4437 #define _FW_WM_VLV(value, plane) \
4438 (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
4440 static void vlv_read_wm_values(struct drm_i915_private
*dev_priv
,
4441 struct vlv_wm_values
*wm
)
4446 for_each_pipe(dev_priv
, pipe
) {
4447 tmp
= I915_READ(VLV_DDL(pipe
));
4449 wm
->ddl
[pipe
].plane
[PLANE_PRIMARY
] =
4450 (tmp
>> DDL_PLANE_SHIFT
) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
4451 wm
->ddl
[pipe
].plane
[PLANE_CURSOR
] =
4452 (tmp
>> DDL_CURSOR_SHIFT
) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
4453 wm
->ddl
[pipe
].plane
[PLANE_SPRITE0
] =
4454 (tmp
>> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
4455 wm
->ddl
[pipe
].plane
[PLANE_SPRITE1
] =
4456 (tmp
>> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
4459 tmp
= I915_READ(DSPFW1
);
4460 wm
->sr
.plane
= _FW_WM(tmp
, SR
);
4461 wm
->pipe
[PIPE_B
].plane
[PLANE_CURSOR
] = _FW_WM(tmp
, CURSORB
);
4462 wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
] = _FW_WM_VLV(tmp
, PLANEB
);
4463 wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
] = _FW_WM_VLV(tmp
, PLANEA
);
4465 tmp
= I915_READ(DSPFW2
);
4466 wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE1
] = _FW_WM_VLV(tmp
, SPRITEB
);
4467 wm
->pipe
[PIPE_A
].plane
[PLANE_CURSOR
] = _FW_WM(tmp
, CURSORA
);
4468 wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
] = _FW_WM_VLV(tmp
, SPRITEA
);
4470 tmp
= I915_READ(DSPFW3
);
4471 wm
->sr
.cursor
= _FW_WM(tmp
, CURSOR_SR
);
4473 if (IS_CHERRYVIEW(dev_priv
)) {
4474 tmp
= I915_READ(DSPFW7_CHV
);
4475 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
] = _FW_WM_VLV(tmp
, SPRITED
);
4476 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
] = _FW_WM_VLV(tmp
, SPRITEC
);
4478 tmp
= I915_READ(DSPFW8_CHV
);
4479 wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE1
] = _FW_WM_VLV(tmp
, SPRITEF
);
4480 wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE0
] = _FW_WM_VLV(tmp
, SPRITEE
);
4482 tmp
= I915_READ(DSPFW9_CHV
);
4483 wm
->pipe
[PIPE_C
].plane
[PLANE_PRIMARY
] = _FW_WM_VLV(tmp
, PLANEC
);
4484 wm
->pipe
[PIPE_C
].plane
[PLANE_CURSOR
] = _FW_WM(tmp
, CURSORC
);
4486 tmp
= I915_READ(DSPHOWM
);
4487 wm
->sr
.plane
|= _FW_WM(tmp
, SR_HI
) << 9;
4488 wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE1
] |= _FW_WM(tmp
, SPRITEF_HI
) << 8;
4489 wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE0
] |= _FW_WM(tmp
, SPRITEE_HI
) << 8;
4490 wm
->pipe
[PIPE_C
].plane
[PLANE_PRIMARY
] |= _FW_WM(tmp
, PLANEC_HI
) << 8;
4491 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
] |= _FW_WM(tmp
, SPRITED_HI
) << 8;
4492 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
] |= _FW_WM(tmp
, SPRITEC_HI
) << 8;
4493 wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
] |= _FW_WM(tmp
, PLANEB_HI
) << 8;
4494 wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE1
] |= _FW_WM(tmp
, SPRITEB_HI
) << 8;
4495 wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
] |= _FW_WM(tmp
, SPRITEA_HI
) << 8;
4496 wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
] |= _FW_WM(tmp
, PLANEA_HI
) << 8;
4498 tmp
= I915_READ(DSPFW7
);
4499 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
] = _FW_WM_VLV(tmp
, SPRITED
);
4500 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
] = _FW_WM_VLV(tmp
, SPRITEC
);
4502 tmp
= I915_READ(DSPHOWM
);
4503 wm
->sr
.plane
|= _FW_WM(tmp
, SR_HI
) << 9;
4504 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
] |= _FW_WM(tmp
, SPRITED_HI
) << 8;
4505 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
] |= _FW_WM(tmp
, SPRITEC_HI
) << 8;
4506 wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
] |= _FW_WM(tmp
, PLANEB_HI
) << 8;
4507 wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE1
] |= _FW_WM(tmp
, SPRITEB_HI
) << 8;
4508 wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
] |= _FW_WM(tmp
, SPRITEA_HI
) << 8;
4509 wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
] |= _FW_WM(tmp
, PLANEA_HI
) << 8;
4516 void vlv_wm_get_hw_state(struct drm_device
*dev
)
4518 struct drm_i915_private
*dev_priv
= to_i915(dev
);
4519 struct vlv_wm_values
*wm
= &dev_priv
->wm
.vlv
;
4520 struct intel_plane
*plane
;
4524 vlv_read_wm_values(dev_priv
, wm
);
4526 for_each_intel_plane(dev
, plane
)
4527 plane
->wm
.fifo_size
= vlv_get_fifo_size(plane
);
4529 wm
->cxsr
= I915_READ(FW_BLC_SELF_VLV
) & FW_CSPWRDWNEN
;
4530 wm
->level
= VLV_WM_LEVEL_PM2
;
4532 if (IS_CHERRYVIEW(dev_priv
)) {
4533 mutex_lock(&dev_priv
->rps
.hw_lock
);
4535 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DSPFREQ
);
4536 if (val
& DSP_MAXFIFO_PM5_ENABLE
)
4537 wm
->level
= VLV_WM_LEVEL_PM5
;
4540 * If DDR DVFS is disabled in the BIOS, Punit
4541 * will never ack the request. So if that happens
4542 * assume we don't have to enable/disable DDR DVFS
4543 * dynamically. To test that just set the REQ_ACK
4544 * bit to poke the Punit, but don't change the
4545 * HIGH/LOW bits so that we don't actually change
4546 * the current state.
4548 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
);
4549 val
|= FORCE_DDR_FREQ_REQ_ACK
;
4550 vlv_punit_write(dev_priv
, PUNIT_REG_DDR_SETUP2
, val
);
4552 if (wait_for((vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
) &
4553 FORCE_DDR_FREQ_REQ_ACK
) == 0, 3)) {
4554 DRM_DEBUG_KMS("Punit not acking DDR DVFS request, "
4555 "assuming DDR DVFS is disabled\n");
4556 dev_priv
->wm
.max_level
= VLV_WM_LEVEL_PM5
;
4558 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
);
4559 if ((val
& FORCE_DDR_HIGH_FREQ
) == 0)
4560 wm
->level
= VLV_WM_LEVEL_DDR_DVFS
;
4563 mutex_unlock(&dev_priv
->rps
.hw_lock
);
4566 for_each_pipe(dev_priv
, pipe
)
4567 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
4569 wm
->pipe
[pipe
].plane
[PLANE_PRIMARY
],
4570 wm
->pipe
[pipe
].plane
[PLANE_CURSOR
],
4571 wm
->pipe
[pipe
].plane
[PLANE_SPRITE0
],
4572 wm
->pipe
[pipe
].plane
[PLANE_SPRITE1
]);
4574 DRM_DEBUG_KMS("Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
4575 wm
->sr
.plane
, wm
->sr
.cursor
, wm
->level
, wm
->cxsr
);
4578 void ilk_wm_get_hw_state(struct drm_device
*dev
)
4580 struct drm_i915_private
*dev_priv
= to_i915(dev
);
4581 struct ilk_wm_values
*hw
= &dev_priv
->wm
.hw
;
4582 struct drm_crtc
*crtc
;
4584 for_each_crtc(dev
, crtc
)
4585 ilk_pipe_wm_get_hw_state(crtc
);
4587 hw
->wm_lp
[0] = I915_READ(WM1_LP_ILK
);
4588 hw
->wm_lp
[1] = I915_READ(WM2_LP_ILK
);
4589 hw
->wm_lp
[2] = I915_READ(WM3_LP_ILK
);
4591 hw
->wm_lp_spr
[0] = I915_READ(WM1S_LP_ILK
);
4592 if (INTEL_GEN(dev_priv
) >= 7) {
4593 hw
->wm_lp_spr
[1] = I915_READ(WM2S_LP_IVB
);
4594 hw
->wm_lp_spr
[2] = I915_READ(WM3S_LP_IVB
);
4597 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
4598 hw
->partitioning
= (I915_READ(WM_MISC
) & WM_MISC_DATA_PARTITION_5_6
) ?
4599 INTEL_DDB_PART_5_6
: INTEL_DDB_PART_1_2
;
4600 else if (IS_IVYBRIDGE(dev_priv
))
4601 hw
->partitioning
= (I915_READ(DISP_ARB_CTL2
) & DISP_DATA_PARTITION_5_6
) ?
4602 INTEL_DDB_PART_5_6
: INTEL_DDB_PART_1_2
;
4605 !(I915_READ(DISP_ARB_CTL
) & DISP_FBC_WM_DIS
);
4609 * intel_update_watermarks - update FIFO watermark values based on current modes
4611 * Calculate watermark values for the various WM regs based on current mode
4612 * and plane configuration.
4614 * There are several cases to deal with here:
4615 * - normal (i.e. non-self-refresh)
4616 * - self-refresh (SR) mode
4617 * - lines are large relative to FIFO size (buffer can hold up to 2)
4618 * - lines are small relative to FIFO size (buffer can hold more than 2
4619 * lines), so need to account for TLB latency
4621 * The normal calculation is:
4622 * watermark = dotclock * bytes per pixel * latency
4623 * where latency is platform & configuration dependent (we assume pessimal
4626 * The SR calculation is:
4627 * watermark = (trunc(latency/line time)+1) * surface width *
4630 * line time = htotal / dotclock
4631 * surface width = hdisplay for normal plane and 64 for cursor
4632 * and latency is assumed to be high, as above.
4634 * The final value programmed to the register should always be rounded up,
4635 * and include an extra 2 entries to account for clock crossings.
4637 * We don't use the sprite, so we can ignore that. And on Crestline we have
4638 * to set the non-SR watermarks to 8.
4640 void intel_update_watermarks(struct intel_crtc
*crtc
)
4642 struct drm_i915_private
*dev_priv
= to_i915(crtc
->base
.dev
);
4644 if (dev_priv
->display
.update_wm
)
4645 dev_priv
->display
.update_wm(crtc
);
4649 * Lock protecting IPS related data structures
4651 DEFINE_SPINLOCK(mchdev_lock
);
4653 /* Global for IPS driver to get at the current i915 device. Protected by
4655 static struct drm_i915_private
*i915_mch_dev
;
4657 bool ironlake_set_drps(struct drm_i915_private
*dev_priv
, u8 val
)
4661 assert_spin_locked(&mchdev_lock
);
4663 rgvswctl
= I915_READ16(MEMSWCTL
);
4664 if (rgvswctl
& MEMCTL_CMD_STS
) {
4665 DRM_DEBUG("gpu busy, RCS change rejected\n");
4666 return false; /* still busy with another command */
4669 rgvswctl
= (MEMCTL_CMD_CHFREQ
<< MEMCTL_CMD_SHIFT
) |
4670 (val
<< MEMCTL_FREQ_SHIFT
) | MEMCTL_SFCAVM
;
4671 I915_WRITE16(MEMSWCTL
, rgvswctl
);
4672 POSTING_READ16(MEMSWCTL
);
4674 rgvswctl
|= MEMCTL_CMD_STS
;
4675 I915_WRITE16(MEMSWCTL
, rgvswctl
);
4680 static void ironlake_enable_drps(struct drm_i915_private
*dev_priv
)
4683 u8 fmax
, fmin
, fstart
, vstart
;
4685 spin_lock_irq(&mchdev_lock
);
4687 rgvmodectl
= I915_READ(MEMMODECTL
);
4689 /* Enable temp reporting */
4690 I915_WRITE16(PMMISC
, I915_READ(PMMISC
) | MCPPCE_EN
);
4691 I915_WRITE16(TSC1
, I915_READ(TSC1
) | TSE
);
4693 /* 100ms RC evaluation intervals */
4694 I915_WRITE(RCUPEI
, 100000);
4695 I915_WRITE(RCDNEI
, 100000);
4697 /* Set max/min thresholds to 90ms and 80ms respectively */
4698 I915_WRITE(RCBMAXAVG
, 90000);
4699 I915_WRITE(RCBMINAVG
, 80000);
4701 I915_WRITE(MEMIHYST
, 1);
4703 /* Set up min, max, and cur for interrupt handling */
4704 fmax
= (rgvmodectl
& MEMMODE_FMAX_MASK
) >> MEMMODE_FMAX_SHIFT
;
4705 fmin
= (rgvmodectl
& MEMMODE_FMIN_MASK
);
4706 fstart
= (rgvmodectl
& MEMMODE_FSTART_MASK
) >>
4707 MEMMODE_FSTART_SHIFT
;
4709 vstart
= (I915_READ(PXVFREQ(fstart
)) & PXVFREQ_PX_MASK
) >>
4712 dev_priv
->ips
.fmax
= fmax
; /* IPS callback will increase this */
4713 dev_priv
->ips
.fstart
= fstart
;
4715 dev_priv
->ips
.max_delay
= fstart
;
4716 dev_priv
->ips
.min_delay
= fmin
;
4717 dev_priv
->ips
.cur_delay
= fstart
;
4719 DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
4720 fmax
, fmin
, fstart
);
4722 I915_WRITE(MEMINTREN
, MEMINT_CX_SUPR_EN
| MEMINT_EVAL_CHG_EN
);
4725 * Interrupts will be enabled in ironlake_irq_postinstall
4728 I915_WRITE(VIDSTART
, vstart
);
4729 POSTING_READ(VIDSTART
);
4731 rgvmodectl
|= MEMMODE_SWMODE_EN
;
4732 I915_WRITE(MEMMODECTL
, rgvmodectl
);
4734 if (wait_for_atomic((I915_READ(MEMSWCTL
) & MEMCTL_CMD_STS
) == 0, 10))
4735 DRM_ERROR("stuck trying to change perf mode\n");
4738 ironlake_set_drps(dev_priv
, fstart
);
4740 dev_priv
->ips
.last_count1
= I915_READ(DMIEC
) +
4741 I915_READ(DDREC
) + I915_READ(CSIEC
);
4742 dev_priv
->ips
.last_time1
= jiffies_to_msecs(jiffies
);
4743 dev_priv
->ips
.last_count2
= I915_READ(GFXEC
);
4744 dev_priv
->ips
.last_time2
= ktime_get_raw_ns();
4746 spin_unlock_irq(&mchdev_lock
);
4749 static void ironlake_disable_drps(struct drm_i915_private
*dev_priv
)
4753 spin_lock_irq(&mchdev_lock
);
4755 rgvswctl
= I915_READ16(MEMSWCTL
);
4757 /* Ack interrupts, disable EFC interrupt */
4758 I915_WRITE(MEMINTREN
, I915_READ(MEMINTREN
) & ~MEMINT_EVAL_CHG_EN
);
4759 I915_WRITE(MEMINTRSTS
, MEMINT_EVAL_CHG
);
4760 I915_WRITE(DEIER
, I915_READ(DEIER
) & ~DE_PCU_EVENT
);
4761 I915_WRITE(DEIIR
, DE_PCU_EVENT
);
4762 I915_WRITE(DEIMR
, I915_READ(DEIMR
) | DE_PCU_EVENT
);
4764 /* Go back to the starting frequency */
4765 ironlake_set_drps(dev_priv
, dev_priv
->ips
.fstart
);
4767 rgvswctl
|= MEMCTL_CMD_STS
;
4768 I915_WRITE(MEMSWCTL
, rgvswctl
);
4771 spin_unlock_irq(&mchdev_lock
);
4774 /* There's a funny hw issue where the hw returns all 0 when reading from
4775 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
4776 * ourselves, instead of doing a rmw cycle (which might result in us clearing
4777 * all limits and the gpu stuck at whatever frequency it is at atm).
4779 static u32
intel_rps_limits(struct drm_i915_private
*dev_priv
, u8 val
)
4783 /* Only set the down limit when we've reached the lowest level to avoid
4784 * getting more interrupts, otherwise leave this clear. This prevents a
4785 * race in the hw when coming out of rc6: There's a tiny window where
4786 * the hw runs at the minimal clock before selecting the desired
4787 * frequency, if the down threshold expires in that window we will not
4788 * receive a down interrupt. */
4789 if (IS_GEN9(dev_priv
)) {
4790 limits
= (dev_priv
->rps
.max_freq_softlimit
) << 23;
4791 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
4792 limits
|= (dev_priv
->rps
.min_freq_softlimit
) << 14;
4794 limits
= dev_priv
->rps
.max_freq_softlimit
<< 24;
4795 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
4796 limits
|= dev_priv
->rps
.min_freq_softlimit
<< 16;
4802 static void gen6_set_rps_thresholds(struct drm_i915_private
*dev_priv
, u8 val
)
4805 u32 threshold_up
= 0, threshold_down
= 0; /* in % */
4806 u32 ei_up
= 0, ei_down
= 0;
4808 new_power
= dev_priv
->rps
.power
;
4809 switch (dev_priv
->rps
.power
) {
4811 if (val
> dev_priv
->rps
.efficient_freq
+ 1 &&
4812 val
> dev_priv
->rps
.cur_freq
)
4813 new_power
= BETWEEN
;
4817 if (val
<= dev_priv
->rps
.efficient_freq
&&
4818 val
< dev_priv
->rps
.cur_freq
)
4819 new_power
= LOW_POWER
;
4820 else if (val
>= dev_priv
->rps
.rp0_freq
&&
4821 val
> dev_priv
->rps
.cur_freq
)
4822 new_power
= HIGH_POWER
;
4826 if (val
< (dev_priv
->rps
.rp1_freq
+ dev_priv
->rps
.rp0_freq
) >> 1 &&
4827 val
< dev_priv
->rps
.cur_freq
)
4828 new_power
= BETWEEN
;
4831 /* Max/min bins are special */
4832 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
4833 new_power
= LOW_POWER
;
4834 if (val
>= dev_priv
->rps
.max_freq_softlimit
)
4835 new_power
= HIGH_POWER
;
4836 if (new_power
== dev_priv
->rps
.power
)
4839 /* Note the units here are not exactly 1us, but 1280ns. */
4840 switch (new_power
) {
4842 /* Upclock if more than 95% busy over 16ms */
4846 /* Downclock if less than 85% busy over 32ms */
4848 threshold_down
= 85;
4852 /* Upclock if more than 90% busy over 13ms */
4856 /* Downclock if less than 75% busy over 32ms */
4858 threshold_down
= 75;
4862 /* Upclock if more than 85% busy over 10ms */
4866 /* Downclock if less than 60% busy over 32ms */
4868 threshold_down
= 60;
4872 I915_WRITE(GEN6_RP_UP_EI
,
4873 GT_INTERVAL_FROM_US(dev_priv
, ei_up
));
4874 I915_WRITE(GEN6_RP_UP_THRESHOLD
,
4875 GT_INTERVAL_FROM_US(dev_priv
,
4876 ei_up
* threshold_up
/ 100));
4878 I915_WRITE(GEN6_RP_DOWN_EI
,
4879 GT_INTERVAL_FROM_US(dev_priv
, ei_down
));
4880 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
,
4881 GT_INTERVAL_FROM_US(dev_priv
,
4882 ei_down
* threshold_down
/ 100));
4884 I915_WRITE(GEN6_RP_CONTROL
,
4885 GEN6_RP_MEDIA_TURBO
|
4886 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
4887 GEN6_RP_MEDIA_IS_GFX
|
4889 GEN6_RP_UP_BUSY_AVG
|
4890 GEN6_RP_DOWN_IDLE_AVG
);
4892 dev_priv
->rps
.power
= new_power
;
4893 dev_priv
->rps
.up_threshold
= threshold_up
;
4894 dev_priv
->rps
.down_threshold
= threshold_down
;
4895 dev_priv
->rps
.last_adj
= 0;
4898 static u32
gen6_rps_pm_mask(struct drm_i915_private
*dev_priv
, u8 val
)
4902 if (val
> dev_priv
->rps
.min_freq_softlimit
)
4903 mask
|= GEN6_PM_RP_DOWN_EI_EXPIRED
| GEN6_PM_RP_DOWN_THRESHOLD
| GEN6_PM_RP_DOWN_TIMEOUT
;
4904 if (val
< dev_priv
->rps
.max_freq_softlimit
)
4905 mask
|= GEN6_PM_RP_UP_EI_EXPIRED
| GEN6_PM_RP_UP_THRESHOLD
;
4907 mask
&= dev_priv
->pm_rps_events
;
4909 return gen6_sanitize_rps_pm_mask(dev_priv
, ~mask
);
4912 /* gen6_set_rps is called to update the frequency request, but should also be
4913 * called when the range (min_delay and max_delay) is modified so that we can
4914 * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
4915 static int gen6_set_rps(struct drm_i915_private
*dev_priv
, u8 val
)
4917 /* min/max delay may still have been modified so be sure to
4918 * write the limits value.
4920 if (val
!= dev_priv
->rps
.cur_freq
) {
4921 gen6_set_rps_thresholds(dev_priv
, val
);
4923 if (IS_GEN9(dev_priv
))
4924 I915_WRITE(GEN6_RPNSWREQ
,
4925 GEN9_FREQUENCY(val
));
4926 else if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
4927 I915_WRITE(GEN6_RPNSWREQ
,
4928 HSW_FREQUENCY(val
));
4930 I915_WRITE(GEN6_RPNSWREQ
,
4931 GEN6_FREQUENCY(val
) |
4933 GEN6_AGGRESSIVE_TURBO
);
4936 /* Make sure we continue to get interrupts
4937 * until we hit the minimum or maximum frequencies.
4939 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS
, intel_rps_limits(dev_priv
, val
));
4940 I915_WRITE(GEN6_PMINTRMSK
, gen6_rps_pm_mask(dev_priv
, val
));
4942 POSTING_READ(GEN6_RPNSWREQ
);
4944 dev_priv
->rps
.cur_freq
= val
;
4945 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv
, val
));
4950 static int valleyview_set_rps(struct drm_i915_private
*dev_priv
, u8 val
)
4954 if (WARN_ONCE(IS_CHERRYVIEW(dev_priv
) && (val
& 1),
4955 "Odd GPU freq value\n"))
4958 I915_WRITE(GEN6_PMINTRMSK
, gen6_rps_pm_mask(dev_priv
, val
));
4960 if (val
!= dev_priv
->rps
.cur_freq
) {
4961 err
= vlv_punit_write(dev_priv
, PUNIT_REG_GPU_FREQ_REQ
, val
);
4965 gen6_set_rps_thresholds(dev_priv
, val
);
4968 dev_priv
->rps
.cur_freq
= val
;
4969 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv
, val
));
4974 /* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
4976 * * If Gfx is Idle, then
4977 * 1. Forcewake Media well.
4978 * 2. Request idle freq.
4979 * 3. Release Forcewake of Media well.
4981 static void vlv_set_rps_idle(struct drm_i915_private
*dev_priv
)
4983 u32 val
= dev_priv
->rps
.idle_freq
;
4986 if (dev_priv
->rps
.cur_freq
<= val
)
4989 /* The punit delays the write of the frequency and voltage until it
4990 * determines the GPU is awake. During normal usage we don't want to
4991 * waste power changing the frequency if the GPU is sleeping (rc6).
4992 * However, the GPU and driver is now idle and we do not want to delay
4993 * switching to minimum voltage (reducing power whilst idle) as we do
4994 * not expect to be woken in the near future and so must flush the
4995 * change by waking the device.
4997 * We choose to take the media powerwell (either would do to trick the
4998 * punit into committing the voltage change) as that takes a lot less
4999 * power than the render powerwell.
5001 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_MEDIA
);
5002 err
= valleyview_set_rps(dev_priv
, val
);
5003 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_MEDIA
);
5006 DRM_ERROR("Failed to set RPS for idle\n");
5009 void gen6_rps_busy(struct drm_i915_private
*dev_priv
)
5011 mutex_lock(&dev_priv
->rps
.hw_lock
);
5012 if (dev_priv
->rps
.enabled
) {
5015 if (dev_priv
->pm_rps_events
& (GEN6_PM_RP_DOWN_EI_EXPIRED
| GEN6_PM_RP_UP_EI_EXPIRED
))
5016 gen6_rps_reset_ei(dev_priv
);
5017 I915_WRITE(GEN6_PMINTRMSK
,
5018 gen6_rps_pm_mask(dev_priv
, dev_priv
->rps
.cur_freq
));
5020 gen6_enable_rps_interrupts(dev_priv
);
5022 /* Use the user's desired frequency as a guide, but for better
5023 * performance, jump directly to RPe as our starting frequency.
5025 freq
= max(dev_priv
->rps
.cur_freq
,
5026 dev_priv
->rps
.efficient_freq
);
5028 if (intel_set_rps(dev_priv
,
5030 dev_priv
->rps
.min_freq_softlimit
,
5031 dev_priv
->rps
.max_freq_softlimit
)))
5032 DRM_DEBUG_DRIVER("Failed to set idle frequency\n");
5034 mutex_unlock(&dev_priv
->rps
.hw_lock
);
5037 void gen6_rps_idle(struct drm_i915_private
*dev_priv
)
5039 /* Flush our bottom-half so that it does not race with us
5040 * setting the idle frequency and so that it is bounded by
5041 * our rpm wakeref. And then disable the interrupts to stop any
5042 * futher RPS reclocking whilst we are asleep.
5044 gen6_disable_rps_interrupts(dev_priv
);
5046 mutex_lock(&dev_priv
->rps
.hw_lock
);
5047 if (dev_priv
->rps
.enabled
) {
5048 if (IS_VALLEYVIEW(dev_priv
) || IS_CHERRYVIEW(dev_priv
))
5049 vlv_set_rps_idle(dev_priv
);
5051 gen6_set_rps(dev_priv
, dev_priv
->rps
.idle_freq
);
5052 dev_priv
->rps
.last_adj
= 0;
5053 I915_WRITE(GEN6_PMINTRMSK
,
5054 gen6_sanitize_rps_pm_mask(dev_priv
, ~0));
5056 mutex_unlock(&dev_priv
->rps
.hw_lock
);
5058 spin_lock(&dev_priv
->rps
.client_lock
);
5059 while (!list_empty(&dev_priv
->rps
.clients
))
5060 list_del_init(dev_priv
->rps
.clients
.next
);
5061 spin_unlock(&dev_priv
->rps
.client_lock
);
5064 void gen6_rps_boost(struct drm_i915_private
*dev_priv
,
5065 struct intel_rps_client
*rps
,
5066 unsigned long submitted
)
5068 /* This is intentionally racy! We peek at the state here, then
5069 * validate inside the RPS worker.
5071 if (!(dev_priv
->gt
.awake
&&
5072 dev_priv
->rps
.enabled
&&
5073 dev_priv
->rps
.cur_freq
< dev_priv
->rps
.boost_freq
))
5076 /* Force a RPS boost (and don't count it against the client) if
5077 * the GPU is severely congested.
5079 if (rps
&& time_after(jiffies
, submitted
+ DRM_I915_THROTTLE_JIFFIES
))
5082 spin_lock(&dev_priv
->rps
.client_lock
);
5083 if (rps
== NULL
|| list_empty(&rps
->link
)) {
5084 spin_lock_irq(&dev_priv
->irq_lock
);
5085 if (dev_priv
->rps
.interrupts_enabled
) {
5086 dev_priv
->rps
.client_boost
= true;
5087 schedule_work(&dev_priv
->rps
.work
);
5089 spin_unlock_irq(&dev_priv
->irq_lock
);
5092 list_add(&rps
->link
, &dev_priv
->rps
.clients
);
5095 dev_priv
->rps
.boosts
++;
5097 spin_unlock(&dev_priv
->rps
.client_lock
);
5100 int intel_set_rps(struct drm_i915_private
*dev_priv
, u8 val
)
5104 lockdep_assert_held(&dev_priv
->rps
.hw_lock
);
5105 GEM_BUG_ON(val
> dev_priv
->rps
.max_freq
);
5106 GEM_BUG_ON(val
< dev_priv
->rps
.min_freq
);
5108 if (!dev_priv
->rps
.enabled
) {
5109 dev_priv
->rps
.cur_freq
= val
;
5113 if (IS_VALLEYVIEW(dev_priv
) || IS_CHERRYVIEW(dev_priv
))
5114 err
= valleyview_set_rps(dev_priv
, val
);
5116 err
= gen6_set_rps(dev_priv
, val
);
5121 static void gen9_disable_rc6(struct drm_i915_private
*dev_priv
)
5123 I915_WRITE(GEN6_RC_CONTROL
, 0);
5124 I915_WRITE(GEN9_PG_ENABLE
, 0);
5127 static void gen9_disable_rps(struct drm_i915_private
*dev_priv
)
5129 I915_WRITE(GEN6_RP_CONTROL
, 0);
5132 static void gen6_disable_rps(struct drm_i915_private
*dev_priv
)
5134 I915_WRITE(GEN6_RC_CONTROL
, 0);
5135 I915_WRITE(GEN6_RPNSWREQ
, 1 << 31);
5136 I915_WRITE(GEN6_RP_CONTROL
, 0);
5139 static void cherryview_disable_rps(struct drm_i915_private
*dev_priv
)
5141 I915_WRITE(GEN6_RC_CONTROL
, 0);
5144 static void valleyview_disable_rps(struct drm_i915_private
*dev_priv
)
5146 /* we're doing forcewake before Disabling RC6,
5147 * This what the BIOS expects when going into suspend */
5148 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
5150 I915_WRITE(GEN6_RC_CONTROL
, 0);
5152 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
5155 static void intel_print_rc6_info(struct drm_i915_private
*dev_priv
, u32 mode
)
5157 if (IS_VALLEYVIEW(dev_priv
) || IS_CHERRYVIEW(dev_priv
)) {
5158 if (mode
& (GEN7_RC_CTL_TO_MODE
| GEN6_RC_CTL_EI_MODE(1)))
5159 mode
= GEN6_RC_CTL_RC6_ENABLE
;
5163 if (HAS_RC6p(dev_priv
))
5164 DRM_DEBUG_DRIVER("Enabling RC6 states: "
5165 "RC6 %s RC6p %s RC6pp %s\n",
5166 onoff(mode
& GEN6_RC_CTL_RC6_ENABLE
),
5167 onoff(mode
& GEN6_RC_CTL_RC6p_ENABLE
),
5168 onoff(mode
& GEN6_RC_CTL_RC6pp_ENABLE
));
5171 DRM_DEBUG_DRIVER("Enabling RC6 states: RC6 %s\n",
5172 onoff(mode
& GEN6_RC_CTL_RC6_ENABLE
));
5175 static bool bxt_check_bios_rc6_setup(struct drm_i915_private
*dev_priv
)
5177 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
5178 bool enable_rc6
= true;
5179 unsigned long rc6_ctx_base
;
5183 rc_ctl
= I915_READ(GEN6_RC_CONTROL
);
5184 rc_sw_target
= (I915_READ(GEN6_RC_STATE
) & RC_SW_TARGET_STATE_MASK
) >>
5185 RC_SW_TARGET_STATE_SHIFT
;
5186 DRM_DEBUG_DRIVER("BIOS enabled RC states: "
5187 "HW_CTRL %s HW_RC6 %s SW_TARGET_STATE %x\n",
5188 onoff(rc_ctl
& GEN6_RC_CTL_HW_ENABLE
),
5189 onoff(rc_ctl
& GEN6_RC_CTL_RC6_ENABLE
),
5192 if (!(I915_READ(RC6_LOCATION
) & RC6_CTX_IN_DRAM
)) {
5193 DRM_DEBUG_DRIVER("RC6 Base location not set properly.\n");
5198 * The exact context size is not known for BXT, so assume a page size
5201 rc6_ctx_base
= I915_READ(RC6_CTX_BASE
) & RC6_CTX_BASE_MASK
;
5202 if (!((rc6_ctx_base
>= ggtt
->stolen_reserved_base
) &&
5203 (rc6_ctx_base
+ PAGE_SIZE
<= ggtt
->stolen_reserved_base
+
5204 ggtt
->stolen_reserved_size
))) {
5205 DRM_DEBUG_DRIVER("RC6 Base address not as expected.\n");
5209 if (!(((I915_READ(PWRCTX_MAXCNT_RCSUNIT
) & IDLE_TIME_MASK
) > 1) &&
5210 ((I915_READ(PWRCTX_MAXCNT_VCSUNIT0
) & IDLE_TIME_MASK
) > 1) &&
5211 ((I915_READ(PWRCTX_MAXCNT_BCSUNIT
) & IDLE_TIME_MASK
) > 1) &&
5212 ((I915_READ(PWRCTX_MAXCNT_VECSUNIT
) & IDLE_TIME_MASK
) > 1))) {
5213 DRM_DEBUG_DRIVER("Engine Idle wait time not set properly.\n");
5217 if (!I915_READ(GEN8_PUSHBUS_CONTROL
) ||
5218 !I915_READ(GEN8_PUSHBUS_ENABLE
) ||
5219 !I915_READ(GEN8_PUSHBUS_SHIFT
)) {
5220 DRM_DEBUG_DRIVER("Pushbus not setup properly.\n");
5224 if (!I915_READ(GEN6_GFXPAUSE
)) {
5225 DRM_DEBUG_DRIVER("GFX pause not setup properly.\n");
5229 if (!I915_READ(GEN8_MISC_CTRL0
)) {
5230 DRM_DEBUG_DRIVER("GPM control not setup properly.\n");
5237 int sanitize_rc6_option(struct drm_i915_private
*dev_priv
, int enable_rc6
)
5239 /* No RC6 before Ironlake and code is gone for ilk. */
5240 if (INTEL_INFO(dev_priv
)->gen
< 6)
5246 if (IS_GEN9_LP(dev_priv
) && !bxt_check_bios_rc6_setup(dev_priv
)) {
5247 DRM_INFO("RC6 disabled by BIOS\n");
5251 /* Respect the kernel parameter if it is set */
5252 if (enable_rc6
>= 0) {
5255 if (HAS_RC6p(dev_priv
))
5256 mask
= INTEL_RC6_ENABLE
| INTEL_RC6p_ENABLE
|
5259 mask
= INTEL_RC6_ENABLE
;
5261 if ((enable_rc6
& mask
) != enable_rc6
)
5262 DRM_DEBUG_DRIVER("Adjusting RC6 mask to %d "
5263 "(requested %d, valid %d)\n",
5264 enable_rc6
& mask
, enable_rc6
, mask
);
5266 return enable_rc6
& mask
;
5269 if (IS_IVYBRIDGE(dev_priv
))
5270 return (INTEL_RC6_ENABLE
| INTEL_RC6p_ENABLE
);
5272 return INTEL_RC6_ENABLE
;
5275 static void gen6_init_rps_frequencies(struct drm_i915_private
*dev_priv
)
5277 /* All of these values are in units of 50MHz */
5279 /* static values from HW: RP0 > RP1 > RPn (min_freq) */
5280 if (IS_GEN9_LP(dev_priv
)) {
5281 u32 rp_state_cap
= I915_READ(BXT_RP_STATE_CAP
);
5282 dev_priv
->rps
.rp0_freq
= (rp_state_cap
>> 16) & 0xff;
5283 dev_priv
->rps
.rp1_freq
= (rp_state_cap
>> 8) & 0xff;
5284 dev_priv
->rps
.min_freq
= (rp_state_cap
>> 0) & 0xff;
5286 u32 rp_state_cap
= I915_READ(GEN6_RP_STATE_CAP
);
5287 dev_priv
->rps
.rp0_freq
= (rp_state_cap
>> 0) & 0xff;
5288 dev_priv
->rps
.rp1_freq
= (rp_state_cap
>> 8) & 0xff;
5289 dev_priv
->rps
.min_freq
= (rp_state_cap
>> 16) & 0xff;
5291 /* hw_max = RP0 until we check for overclocking */
5292 dev_priv
->rps
.max_freq
= dev_priv
->rps
.rp0_freq
;
5294 dev_priv
->rps
.efficient_freq
= dev_priv
->rps
.rp1_freq
;
5295 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
) ||
5296 IS_GEN9_BC(dev_priv
)) {
5297 u32 ddcc_status
= 0;
5299 if (sandybridge_pcode_read(dev_priv
,
5300 HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL
,
5302 dev_priv
->rps
.efficient_freq
=
5304 ((ddcc_status
>> 8) & 0xff),
5305 dev_priv
->rps
.min_freq
,
5306 dev_priv
->rps
.max_freq
);
5309 if (IS_GEN9_BC(dev_priv
)) {
5310 /* Store the frequency values in 16.66 MHZ units, which is
5311 * the natural hardware unit for SKL
5313 dev_priv
->rps
.rp0_freq
*= GEN9_FREQ_SCALER
;
5314 dev_priv
->rps
.rp1_freq
*= GEN9_FREQ_SCALER
;
5315 dev_priv
->rps
.min_freq
*= GEN9_FREQ_SCALER
;
5316 dev_priv
->rps
.max_freq
*= GEN9_FREQ_SCALER
;
5317 dev_priv
->rps
.efficient_freq
*= GEN9_FREQ_SCALER
;
5321 static void reset_rps(struct drm_i915_private
*dev_priv
,
5322 int (*set
)(struct drm_i915_private
*, u8
))
5324 u8 freq
= dev_priv
->rps
.cur_freq
;
5327 dev_priv
->rps
.power
= -1;
5328 dev_priv
->rps
.cur_freq
= -1;
5330 if (set(dev_priv
, freq
))
5331 DRM_ERROR("Failed to reset RPS to initial values\n");
5334 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
5335 static void gen9_enable_rps(struct drm_i915_private
*dev_priv
)
5337 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
5339 /* Program defaults and thresholds for RPS*/
5340 I915_WRITE(GEN6_RC_VIDEO_FREQ
,
5341 GEN9_FREQUENCY(dev_priv
->rps
.rp1_freq
));
5343 /* 1 second timeout*/
5344 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
,
5345 GT_INTERVAL_FROM_US(dev_priv
, 1000000));
5347 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 0xa);
5349 /* Leaning on the below call to gen6_set_rps to program/setup the
5350 * Up/Down EI & threshold registers, as well as the RP_CONTROL,
5351 * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
5352 reset_rps(dev_priv
, gen6_set_rps
);
5354 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
5357 static void gen9_enable_rc6(struct drm_i915_private
*dev_priv
)
5359 struct intel_engine_cs
*engine
;
5360 enum intel_engine_id id
;
5361 uint32_t rc6_mask
= 0;
5363 /* 1a: Software RC state - RC0 */
5364 I915_WRITE(GEN6_RC_STATE
, 0);
5366 /* 1b: Get forcewake during program sequence. Although the driver
5367 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5368 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
5370 /* 2a: Disable RC states. */
5371 I915_WRITE(GEN6_RC_CONTROL
, 0);
5373 /* 2b: Program RC6 thresholds.*/
5375 /* WaRsDoubleRc6WrlWithCoarsePowerGating: Doubling WRL only when CPG is enabled */
5376 if (IS_SKYLAKE(dev_priv
))
5377 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 108 << 16);
5379 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 54 << 16);
5380 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
5381 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
5382 for_each_engine(engine
, dev_priv
, id
)
5383 I915_WRITE(RING_MAX_IDLE(engine
->mmio_base
), 10);
5385 if (HAS_GUC(dev_priv
))
5386 I915_WRITE(GUC_MAX_IDLE_COUNT
, 0xA);
5388 I915_WRITE(GEN6_RC_SLEEP
, 0);
5390 /* 2c: Program Coarse Power Gating Policies. */
5391 I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS
, 25);
5392 I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS
, 25);
5394 /* 3a: Enable RC6 */
5395 if (intel_enable_rc6() & INTEL_RC6_ENABLE
)
5396 rc6_mask
= GEN6_RC_CTL_RC6_ENABLE
;
5397 DRM_INFO("RC6 %s\n", onoff(rc6_mask
& GEN6_RC_CTL_RC6_ENABLE
));
5398 I915_WRITE(GEN6_RC6_THRESHOLD
, 37500); /* 37.5/125ms per EI */
5399 I915_WRITE(GEN6_RC_CONTROL
,
5400 GEN6_RC_CTL_HW_ENABLE
| GEN6_RC_CTL_EI_MODE(1) | rc6_mask
);
5403 * 3b: Enable Coarse Power Gating only when RC6 is enabled.
5404 * WaRsDisableCoarsePowerGating:skl,bxt - Render/Media PG need to be disabled with RC6.
5406 if (NEEDS_WaRsDisableCoarsePowerGating(dev_priv
))
5407 I915_WRITE(GEN9_PG_ENABLE
, 0);
5409 I915_WRITE(GEN9_PG_ENABLE
, (rc6_mask
& GEN6_RC_CTL_RC6_ENABLE
) ?
5410 (GEN9_RENDER_PG_ENABLE
| GEN9_MEDIA_PG_ENABLE
) : 0);
5412 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
5415 static void gen8_enable_rps(struct drm_i915_private
*dev_priv
)
5417 struct intel_engine_cs
*engine
;
5418 enum intel_engine_id id
;
5419 uint32_t rc6_mask
= 0;
5421 /* 1a: Software RC state - RC0 */
5422 I915_WRITE(GEN6_RC_STATE
, 0);
5424 /* 1c & 1d: Get forcewake during program sequence. Although the driver
5425 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5426 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
5428 /* 2a: Disable RC states. */
5429 I915_WRITE(GEN6_RC_CONTROL
, 0);
5431 /* 2b: Program RC6 thresholds.*/
5432 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16);
5433 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
5434 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
5435 for_each_engine(engine
, dev_priv
, id
)
5436 I915_WRITE(RING_MAX_IDLE(engine
->mmio_base
), 10);
5437 I915_WRITE(GEN6_RC_SLEEP
, 0);
5438 if (IS_BROADWELL(dev_priv
))
5439 I915_WRITE(GEN6_RC6_THRESHOLD
, 625); /* 800us/1.28 for TO */
5441 I915_WRITE(GEN6_RC6_THRESHOLD
, 50000); /* 50/125ms per EI */
5444 if (intel_enable_rc6() & INTEL_RC6_ENABLE
)
5445 rc6_mask
= GEN6_RC_CTL_RC6_ENABLE
;
5446 intel_print_rc6_info(dev_priv
, rc6_mask
);
5447 if (IS_BROADWELL(dev_priv
))
5448 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
5449 GEN7_RC_CTL_TO_MODE
|
5452 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
5453 GEN6_RC_CTL_EI_MODE(1) |
5456 /* 4 Program defaults and thresholds for RPS*/
5457 I915_WRITE(GEN6_RPNSWREQ
,
5458 HSW_FREQUENCY(dev_priv
->rps
.rp1_freq
));
5459 I915_WRITE(GEN6_RC_VIDEO_FREQ
,
5460 HSW_FREQUENCY(dev_priv
->rps
.rp1_freq
));
5461 /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
5462 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 100000000 / 128); /* 1 second timeout */
5464 /* Docs recommend 900MHz, and 300 MHz respectively */
5465 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS
,
5466 dev_priv
->rps
.max_freq_softlimit
<< 24 |
5467 dev_priv
->rps
.min_freq_softlimit
<< 16);
5469 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 7600000 / 128); /* 76ms busyness per EI, 90% */
5470 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 31300000 / 128); /* 313ms busyness per EI, 70%*/
5471 I915_WRITE(GEN6_RP_UP_EI
, 66000); /* 84.48ms, XXX: random? */
5472 I915_WRITE(GEN6_RP_DOWN_EI
, 350000); /* 448ms, XXX: random? */
5474 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
5477 I915_WRITE(GEN6_RP_CONTROL
,
5478 GEN6_RP_MEDIA_TURBO
|
5479 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
5480 GEN6_RP_MEDIA_IS_GFX
|
5482 GEN6_RP_UP_BUSY_AVG
|
5483 GEN6_RP_DOWN_IDLE_AVG
);
5485 /* 6: Ring frequency + overclocking (our driver does this later */
5487 reset_rps(dev_priv
, gen6_set_rps
);
5489 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
5492 static void gen6_enable_rps(struct drm_i915_private
*dev_priv
)
5494 struct intel_engine_cs
*engine
;
5495 enum intel_engine_id id
;
5496 u32 rc6vids
, rc6_mask
= 0;
5501 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
5503 /* Here begins a magic sequence of register writes to enable
5504 * auto-downclocking.
5506 * Perhaps there might be some value in exposing these to
5509 I915_WRITE(GEN6_RC_STATE
, 0);
5511 /* Clear the DBG now so we don't confuse earlier errors */
5512 gtfifodbg
= I915_READ(GTFIFODBG
);
5514 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg
);
5515 I915_WRITE(GTFIFODBG
, gtfifodbg
);
5518 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
5520 /* disable the counters and set deterministic thresholds */
5521 I915_WRITE(GEN6_RC_CONTROL
, 0);
5523 I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT
, 1000 << 16);
5524 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16 | 30);
5525 I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT
, 30);
5526 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000);
5527 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25);
5529 for_each_engine(engine
, dev_priv
, id
)
5530 I915_WRITE(RING_MAX_IDLE(engine
->mmio_base
), 10);
5532 I915_WRITE(GEN6_RC_SLEEP
, 0);
5533 I915_WRITE(GEN6_RC1e_THRESHOLD
, 1000);
5534 if (IS_IVYBRIDGE(dev_priv
))
5535 I915_WRITE(GEN6_RC6_THRESHOLD
, 125000);
5537 I915_WRITE(GEN6_RC6_THRESHOLD
, 50000);
5538 I915_WRITE(GEN6_RC6p_THRESHOLD
, 150000);
5539 I915_WRITE(GEN6_RC6pp_THRESHOLD
, 64000); /* unused */
5541 /* Check if we are enabling RC6 */
5542 rc6_mode
= intel_enable_rc6();
5543 if (rc6_mode
& INTEL_RC6_ENABLE
)
5544 rc6_mask
|= GEN6_RC_CTL_RC6_ENABLE
;
5546 /* We don't use those on Haswell */
5547 if (!IS_HASWELL(dev_priv
)) {
5548 if (rc6_mode
& INTEL_RC6p_ENABLE
)
5549 rc6_mask
|= GEN6_RC_CTL_RC6p_ENABLE
;
5551 if (rc6_mode
& INTEL_RC6pp_ENABLE
)
5552 rc6_mask
|= GEN6_RC_CTL_RC6pp_ENABLE
;
5555 intel_print_rc6_info(dev_priv
, rc6_mask
);
5557 I915_WRITE(GEN6_RC_CONTROL
,
5559 GEN6_RC_CTL_EI_MODE(1) |
5560 GEN6_RC_CTL_HW_ENABLE
);
5562 /* Power down if completely idle for over 50ms */
5563 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 50000);
5564 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
5566 reset_rps(dev_priv
, gen6_set_rps
);
5569 ret
= sandybridge_pcode_read(dev_priv
, GEN6_PCODE_READ_RC6VIDS
, &rc6vids
);
5570 if (IS_GEN6(dev_priv
) && ret
) {
5571 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
5572 } else if (IS_GEN6(dev_priv
) && (GEN6_DECODE_RC6_VID(rc6vids
& 0xff) < 450)) {
5573 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
5574 GEN6_DECODE_RC6_VID(rc6vids
& 0xff), 450);
5575 rc6vids
&= 0xffff00;
5576 rc6vids
|= GEN6_ENCODE_RC6_VID(450);
5577 ret
= sandybridge_pcode_write(dev_priv
, GEN6_PCODE_WRITE_RC6VIDS
, rc6vids
);
5579 DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
5582 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
5585 static void gen6_update_ring_freq(struct drm_i915_private
*dev_priv
)
5588 unsigned int gpu_freq
;
5589 unsigned int max_ia_freq
, min_ring_freq
;
5590 unsigned int max_gpu_freq
, min_gpu_freq
;
5591 int scaling_factor
= 180;
5592 struct cpufreq_policy
*policy
;
5594 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
5596 policy
= cpufreq_cpu_get(0);
5598 max_ia_freq
= policy
->cpuinfo
.max_freq
;
5599 cpufreq_cpu_put(policy
);
5602 * Default to measured freq if none found, PCU will ensure we
5605 max_ia_freq
= tsc_khz
;
5608 /* Convert from kHz to MHz */
5609 max_ia_freq
/= 1000;
5611 min_ring_freq
= I915_READ(DCLK
) & 0xf;
5612 /* convert DDR frequency from units of 266.6MHz to bandwidth */
5613 min_ring_freq
= mult_frac(min_ring_freq
, 8, 3);
5615 if (IS_GEN9_BC(dev_priv
)) {
5616 /* Convert GT frequency to 50 HZ units */
5617 min_gpu_freq
= dev_priv
->rps
.min_freq
/ GEN9_FREQ_SCALER
;
5618 max_gpu_freq
= dev_priv
->rps
.max_freq
/ GEN9_FREQ_SCALER
;
5620 min_gpu_freq
= dev_priv
->rps
.min_freq
;
5621 max_gpu_freq
= dev_priv
->rps
.max_freq
;
5625 * For each potential GPU frequency, load a ring frequency we'd like
5626 * to use for memory access. We do this by specifying the IA frequency
5627 * the PCU should use as a reference to determine the ring frequency.
5629 for (gpu_freq
= max_gpu_freq
; gpu_freq
>= min_gpu_freq
; gpu_freq
--) {
5630 int diff
= max_gpu_freq
- gpu_freq
;
5631 unsigned int ia_freq
= 0, ring_freq
= 0;
5633 if (IS_GEN9_BC(dev_priv
)) {
5635 * ring_freq = 2 * GT. ring_freq is in 100MHz units
5636 * No floor required for ring frequency on SKL.
5638 ring_freq
= gpu_freq
;
5639 } else if (INTEL_INFO(dev_priv
)->gen
>= 8) {
5640 /* max(2 * GT, DDR). NB: GT is 50MHz units */
5641 ring_freq
= max(min_ring_freq
, gpu_freq
);
5642 } else if (IS_HASWELL(dev_priv
)) {
5643 ring_freq
= mult_frac(gpu_freq
, 5, 4);
5644 ring_freq
= max(min_ring_freq
, ring_freq
);
5645 /* leave ia_freq as the default, chosen by cpufreq */
5647 /* On older processors, there is no separate ring
5648 * clock domain, so in order to boost the bandwidth
5649 * of the ring, we need to upclock the CPU (ia_freq).
5651 * For GPU frequencies less than 750MHz,
5652 * just use the lowest ring freq.
5654 if (gpu_freq
< min_freq
)
5657 ia_freq
= max_ia_freq
- ((diff
* scaling_factor
) / 2);
5658 ia_freq
= DIV_ROUND_CLOSEST(ia_freq
, 100);
5661 sandybridge_pcode_write(dev_priv
,
5662 GEN6_PCODE_WRITE_MIN_FREQ_TABLE
,
5663 ia_freq
<< GEN6_PCODE_FREQ_IA_RATIO_SHIFT
|
5664 ring_freq
<< GEN6_PCODE_FREQ_RING_RATIO_SHIFT
|
5669 static int cherryview_rps_max_freq(struct drm_i915_private
*dev_priv
)
5673 val
= vlv_punit_read(dev_priv
, FB_GFX_FMAX_AT_VMAX_FUSE
);
5675 switch (INTEL_INFO(dev_priv
)->sseu
.eu_total
) {
5677 /* (2 * 4) config */
5678 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT
);
5681 /* (2 * 6) config */
5682 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT
);
5685 /* (2 * 8) config */
5687 /* Setting (2 * 8) Min RP0 for any other combination */
5688 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT
);
5692 rp0
= (rp0
& FB_GFX_FREQ_FUSE_MASK
);
5697 static int cherryview_rps_rpe_freq(struct drm_i915_private
*dev_priv
)
5701 val
= vlv_punit_read(dev_priv
, PUNIT_GPU_DUTYCYCLE_REG
);
5702 rpe
= (val
>> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT
) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK
;
5707 static int cherryview_rps_guar_freq(struct drm_i915_private
*dev_priv
)
5711 val
= vlv_punit_read(dev_priv
, FB_GFX_FMAX_AT_VMAX_FUSE
);
5712 rp1
= (val
& FB_GFX_FREQ_FUSE_MASK
);
5717 static u32
cherryview_rps_min_freq(struct drm_i915_private
*dev_priv
)
5721 val
= vlv_punit_read(dev_priv
, FB_GFX_FMIN_AT_VMIN_FUSE
);
5722 rpn
= ((val
>> FB_GFX_FMIN_AT_VMIN_FUSE_SHIFT
) &
5723 FB_GFX_FREQ_FUSE_MASK
);
5728 static int valleyview_rps_guar_freq(struct drm_i915_private
*dev_priv
)
5732 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FREQ_FUSE
);
5734 rp1
= (val
& FB_GFX_FGUARANTEED_FREQ_FUSE_MASK
) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT
;
5739 static int valleyview_rps_max_freq(struct drm_i915_private
*dev_priv
)
5743 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FREQ_FUSE
);
5745 rp0
= (val
& FB_GFX_MAX_FREQ_FUSE_MASK
) >> FB_GFX_MAX_FREQ_FUSE_SHIFT
;
5747 rp0
= min_t(u32
, rp0
, 0xea);
5752 static int valleyview_rps_rpe_freq(struct drm_i915_private
*dev_priv
)
5756 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FMAX_FUSE_LO
);
5757 rpe
= (val
& FB_FMAX_VMIN_FREQ_LO_MASK
) >> FB_FMAX_VMIN_FREQ_LO_SHIFT
;
5758 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FMAX_FUSE_HI
);
5759 rpe
|= (val
& FB_FMAX_VMIN_FREQ_HI_MASK
) << 5;
5764 static int valleyview_rps_min_freq(struct drm_i915_private
*dev_priv
)
5768 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_LFM
) & 0xff;
5770 * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
5771 * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
5772 * a BYT-M B0 the above register contains 0xbf. Moreover when setting
5773 * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
5774 * to make sure it matches what Punit accepts.
5776 return max_t(u32
, val
, 0xc0);
5779 /* Check that the pctx buffer wasn't move under us. */
5780 static void valleyview_check_pctx(struct drm_i915_private
*dev_priv
)
5782 unsigned long pctx_addr
= I915_READ(VLV_PCBR
) & ~4095;
5784 WARN_ON(pctx_addr
!= dev_priv
->mm
.stolen_base
+
5785 dev_priv
->vlv_pctx
->stolen
->start
);
5789 /* Check that the pcbr address is not empty. */
5790 static void cherryview_check_pctx(struct drm_i915_private
*dev_priv
)
5792 unsigned long pctx_addr
= I915_READ(VLV_PCBR
) & ~4095;
5794 WARN_ON((pctx_addr
>> VLV_PCBR_ADDR_SHIFT
) == 0);
5797 static void cherryview_setup_pctx(struct drm_i915_private
*dev_priv
)
5799 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
5800 unsigned long pctx_paddr
, paddr
;
5802 int pctx_size
= 32*1024;
5804 pcbr
= I915_READ(VLV_PCBR
);
5805 if ((pcbr
>> VLV_PCBR_ADDR_SHIFT
) == 0) {
5806 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5807 paddr
= (dev_priv
->mm
.stolen_base
+
5808 (ggtt
->stolen_size
- pctx_size
));
5810 pctx_paddr
= (paddr
& (~4095));
5811 I915_WRITE(VLV_PCBR
, pctx_paddr
);
5814 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR
));
5817 static void valleyview_setup_pctx(struct drm_i915_private
*dev_priv
)
5819 struct drm_i915_gem_object
*pctx
;
5820 unsigned long pctx_paddr
;
5822 int pctx_size
= 24*1024;
5824 pcbr
= I915_READ(VLV_PCBR
);
5826 /* BIOS set it up already, grab the pre-alloc'd space */
5829 pcbr_offset
= (pcbr
& (~4095)) - dev_priv
->mm
.stolen_base
;
5830 pctx
= i915_gem_object_create_stolen_for_preallocated(dev_priv
,
5832 I915_GTT_OFFSET_NONE
,
5837 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5840 * From the Gunit register HAS:
5841 * The Gfx driver is expected to program this register and ensure
5842 * proper allocation within Gfx stolen memory. For example, this
5843 * register should be programmed such than the PCBR range does not
5844 * overlap with other ranges, such as the frame buffer, protected
5845 * memory, or any other relevant ranges.
5847 pctx
= i915_gem_object_create_stolen(dev_priv
, pctx_size
);
5849 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
5853 pctx_paddr
= dev_priv
->mm
.stolen_base
+ pctx
->stolen
->start
;
5854 I915_WRITE(VLV_PCBR
, pctx_paddr
);
5857 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR
));
5858 dev_priv
->vlv_pctx
= pctx
;
5861 static void valleyview_cleanup_pctx(struct drm_i915_private
*dev_priv
)
5863 if (WARN_ON(!dev_priv
->vlv_pctx
))
5866 i915_gem_object_put(dev_priv
->vlv_pctx
);
5867 dev_priv
->vlv_pctx
= NULL
;
5870 static void vlv_init_gpll_ref_freq(struct drm_i915_private
*dev_priv
)
5872 dev_priv
->rps
.gpll_ref_freq
=
5873 vlv_get_cck_clock(dev_priv
, "GPLL ref",
5874 CCK_GPLL_CLOCK_CONTROL
,
5875 dev_priv
->czclk_freq
);
5877 DRM_DEBUG_DRIVER("GPLL reference freq: %d kHz\n",
5878 dev_priv
->rps
.gpll_ref_freq
);
5881 static void valleyview_init_gt_powersave(struct drm_i915_private
*dev_priv
)
5885 valleyview_setup_pctx(dev_priv
);
5887 vlv_init_gpll_ref_freq(dev_priv
);
5889 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
5890 switch ((val
>> 6) & 3) {
5893 dev_priv
->mem_freq
= 800;
5896 dev_priv
->mem_freq
= 1066;
5899 dev_priv
->mem_freq
= 1333;
5902 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv
->mem_freq
);
5904 dev_priv
->rps
.max_freq
= valleyview_rps_max_freq(dev_priv
);
5905 dev_priv
->rps
.rp0_freq
= dev_priv
->rps
.max_freq
;
5906 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5907 intel_gpu_freq(dev_priv
, dev_priv
->rps
.max_freq
),
5908 dev_priv
->rps
.max_freq
);
5910 dev_priv
->rps
.efficient_freq
= valleyview_rps_rpe_freq(dev_priv
);
5911 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5912 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
5913 dev_priv
->rps
.efficient_freq
);
5915 dev_priv
->rps
.rp1_freq
= valleyview_rps_guar_freq(dev_priv
);
5916 DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
5917 intel_gpu_freq(dev_priv
, dev_priv
->rps
.rp1_freq
),
5918 dev_priv
->rps
.rp1_freq
);
5920 dev_priv
->rps
.min_freq
= valleyview_rps_min_freq(dev_priv
);
5921 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5922 intel_gpu_freq(dev_priv
, dev_priv
->rps
.min_freq
),
5923 dev_priv
->rps
.min_freq
);
5926 static void cherryview_init_gt_powersave(struct drm_i915_private
*dev_priv
)
5930 cherryview_setup_pctx(dev_priv
);
5932 vlv_init_gpll_ref_freq(dev_priv
);
5934 mutex_lock(&dev_priv
->sb_lock
);
5935 val
= vlv_cck_read(dev_priv
, CCK_FUSE_REG
);
5936 mutex_unlock(&dev_priv
->sb_lock
);
5938 switch ((val
>> 2) & 0x7) {
5940 dev_priv
->mem_freq
= 2000;
5943 dev_priv
->mem_freq
= 1600;
5946 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv
->mem_freq
);
5948 dev_priv
->rps
.max_freq
= cherryview_rps_max_freq(dev_priv
);
5949 dev_priv
->rps
.rp0_freq
= dev_priv
->rps
.max_freq
;
5950 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5951 intel_gpu_freq(dev_priv
, dev_priv
->rps
.max_freq
),
5952 dev_priv
->rps
.max_freq
);
5954 dev_priv
->rps
.efficient_freq
= cherryview_rps_rpe_freq(dev_priv
);
5955 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5956 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
5957 dev_priv
->rps
.efficient_freq
);
5959 dev_priv
->rps
.rp1_freq
= cherryview_rps_guar_freq(dev_priv
);
5960 DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
5961 intel_gpu_freq(dev_priv
, dev_priv
->rps
.rp1_freq
),
5962 dev_priv
->rps
.rp1_freq
);
5964 dev_priv
->rps
.min_freq
= cherryview_rps_min_freq(dev_priv
);
5965 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5966 intel_gpu_freq(dev_priv
, dev_priv
->rps
.min_freq
),
5967 dev_priv
->rps
.min_freq
);
5969 WARN_ONCE((dev_priv
->rps
.max_freq
|
5970 dev_priv
->rps
.efficient_freq
|
5971 dev_priv
->rps
.rp1_freq
|
5972 dev_priv
->rps
.min_freq
) & 1,
5973 "Odd GPU freq values\n");
5976 static void valleyview_cleanup_gt_powersave(struct drm_i915_private
*dev_priv
)
5978 valleyview_cleanup_pctx(dev_priv
);
5981 static void cherryview_enable_rps(struct drm_i915_private
*dev_priv
)
5983 struct intel_engine_cs
*engine
;
5984 enum intel_engine_id id
;
5985 u32 gtfifodbg
, val
, rc6_mode
= 0, pcbr
;
5987 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
5989 gtfifodbg
= I915_READ(GTFIFODBG
) & ~(GT_FIFO_SBDEDICATE_FREE_ENTRY_CHV
|
5990 GT_FIFO_FREE_ENTRIES_CHV
);
5992 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5994 I915_WRITE(GTFIFODBG
, gtfifodbg
);
5997 cherryview_check_pctx(dev_priv
);
5999 /* 1a & 1b: Get forcewake during program sequence. Although the driver
6000 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
6001 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
6003 /* Disable RC states. */
6004 I915_WRITE(GEN6_RC_CONTROL
, 0);
6006 /* 2a: Program RC6 thresholds.*/
6007 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16);
6008 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
6009 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
6011 for_each_engine(engine
, dev_priv
, id
)
6012 I915_WRITE(RING_MAX_IDLE(engine
->mmio_base
), 10);
6013 I915_WRITE(GEN6_RC_SLEEP
, 0);
6015 /* TO threshold set to 500 us ( 0x186 * 1.28 us) */
6016 I915_WRITE(GEN6_RC6_THRESHOLD
, 0x186);
6018 /* allows RC6 residency counter to work */
6019 I915_WRITE(VLV_COUNTER_CONTROL
,
6020 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH
|
6021 VLV_MEDIA_RC6_COUNT_EN
|
6022 VLV_RENDER_RC6_COUNT_EN
));
6024 /* For now we assume BIOS is allocating and populating the PCBR */
6025 pcbr
= I915_READ(VLV_PCBR
);
6028 if ((intel_enable_rc6() & INTEL_RC6_ENABLE
) &&
6029 (pcbr
>> VLV_PCBR_ADDR_SHIFT
))
6030 rc6_mode
= GEN7_RC_CTL_TO_MODE
;
6032 I915_WRITE(GEN6_RC_CONTROL
, rc6_mode
);
6034 /* 4 Program defaults and thresholds for RPS*/
6035 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 1000000);
6036 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 59400);
6037 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 245000);
6038 I915_WRITE(GEN6_RP_UP_EI
, 66000);
6039 I915_WRITE(GEN6_RP_DOWN_EI
, 350000);
6041 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
6044 I915_WRITE(GEN6_RP_CONTROL
,
6045 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
6046 GEN6_RP_MEDIA_IS_GFX
|
6048 GEN6_RP_UP_BUSY_AVG
|
6049 GEN6_RP_DOWN_IDLE_AVG
);
6051 /* Setting Fixed Bias */
6052 val
= VLV_OVERRIDE_EN
|
6054 CHV_BIAS_CPU_50_SOC_50
;
6055 vlv_punit_write(dev_priv
, VLV_TURBO_SOC_OVERRIDE
, val
);
6057 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
6059 /* RPS code assumes GPLL is used */
6060 WARN_ONCE((val
& GPLLENABLE
) == 0, "GPLL not enabled\n");
6062 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val
& GPLLENABLE
));
6063 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val
);
6065 reset_rps(dev_priv
, valleyview_set_rps
);
6067 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
6070 static void valleyview_enable_rps(struct drm_i915_private
*dev_priv
)
6072 struct intel_engine_cs
*engine
;
6073 enum intel_engine_id id
;
6074 u32 gtfifodbg
, val
, rc6_mode
= 0;
6076 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
6078 valleyview_check_pctx(dev_priv
);
6080 gtfifodbg
= I915_READ(GTFIFODBG
);
6082 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
6084 I915_WRITE(GTFIFODBG
, gtfifodbg
);
6087 /* If VLV, Forcewake all wells, else re-direct to regular path */
6088 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
6090 /* Disable RC states. */
6091 I915_WRITE(GEN6_RC_CONTROL
, 0);
6093 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 1000000);
6094 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 59400);
6095 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 245000);
6096 I915_WRITE(GEN6_RP_UP_EI
, 66000);
6097 I915_WRITE(GEN6_RP_DOWN_EI
, 350000);
6099 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
6101 I915_WRITE(GEN6_RP_CONTROL
,
6102 GEN6_RP_MEDIA_TURBO
|
6103 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
6104 GEN6_RP_MEDIA_IS_GFX
|
6106 GEN6_RP_UP_BUSY_AVG
|
6107 GEN6_RP_DOWN_IDLE_CONT
);
6109 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 0x00280000);
6110 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000);
6111 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25);
6113 for_each_engine(engine
, dev_priv
, id
)
6114 I915_WRITE(RING_MAX_IDLE(engine
->mmio_base
), 10);
6116 I915_WRITE(GEN6_RC6_THRESHOLD
, 0x557);
6118 /* allows RC6 residency counter to work */
6119 I915_WRITE(VLV_COUNTER_CONTROL
,
6120 _MASKED_BIT_ENABLE(VLV_MEDIA_RC0_COUNT_EN
|
6121 VLV_RENDER_RC0_COUNT_EN
|
6122 VLV_MEDIA_RC6_COUNT_EN
|
6123 VLV_RENDER_RC6_COUNT_EN
));
6125 if (intel_enable_rc6() & INTEL_RC6_ENABLE
)
6126 rc6_mode
= GEN7_RC_CTL_TO_MODE
| VLV_RC_CTL_CTX_RST_PARALLEL
;
6128 intel_print_rc6_info(dev_priv
, rc6_mode
);
6130 I915_WRITE(GEN6_RC_CONTROL
, rc6_mode
);
6132 /* Setting Fixed Bias */
6133 val
= VLV_OVERRIDE_EN
|
6135 VLV_BIAS_CPU_125_SOC_875
;
6136 vlv_punit_write(dev_priv
, VLV_TURBO_SOC_OVERRIDE
, val
);
6138 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
6140 /* RPS code assumes GPLL is used */
6141 WARN_ONCE((val
& GPLLENABLE
) == 0, "GPLL not enabled\n");
6143 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val
& GPLLENABLE
));
6144 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val
);
6146 reset_rps(dev_priv
, valleyview_set_rps
);
6148 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
6151 static unsigned long intel_pxfreq(u32 vidfreq
)
6154 int div
= (vidfreq
& 0x3f0000) >> 16;
6155 int post
= (vidfreq
& 0x3000) >> 12;
6156 int pre
= (vidfreq
& 0x7);
6161 freq
= ((div
* 133333) / ((1<<post
) * pre
));
6166 static const struct cparams
{
6172 { 1, 1333, 301, 28664 },
6173 { 1, 1066, 294, 24460 },
6174 { 1, 800, 294, 25192 },
6175 { 0, 1333, 276, 27605 },
6176 { 0, 1066, 276, 27605 },
6177 { 0, 800, 231, 23784 },
6180 static unsigned long __i915_chipset_val(struct drm_i915_private
*dev_priv
)
6182 u64 total_count
, diff
, ret
;
6183 u32 count1
, count2
, count3
, m
= 0, c
= 0;
6184 unsigned long now
= jiffies_to_msecs(jiffies
), diff1
;
6187 assert_spin_locked(&mchdev_lock
);
6189 diff1
= now
- dev_priv
->ips
.last_time1
;
6191 /* Prevent division-by-zero if we are asking too fast.
6192 * Also, we don't get interesting results if we are polling
6193 * faster than once in 10ms, so just return the saved value
6197 return dev_priv
->ips
.chipset_power
;
6199 count1
= I915_READ(DMIEC
);
6200 count2
= I915_READ(DDREC
);
6201 count3
= I915_READ(CSIEC
);
6203 total_count
= count1
+ count2
+ count3
;
6205 /* FIXME: handle per-counter overflow */
6206 if (total_count
< dev_priv
->ips
.last_count1
) {
6207 diff
= ~0UL - dev_priv
->ips
.last_count1
;
6208 diff
+= total_count
;
6210 diff
= total_count
- dev_priv
->ips
.last_count1
;
6213 for (i
= 0; i
< ARRAY_SIZE(cparams
); i
++) {
6214 if (cparams
[i
].i
== dev_priv
->ips
.c_m
&&
6215 cparams
[i
].t
== dev_priv
->ips
.r_t
) {
6222 diff
= div_u64(diff
, diff1
);
6223 ret
= ((m
* diff
) + c
);
6224 ret
= div_u64(ret
, 10);
6226 dev_priv
->ips
.last_count1
= total_count
;
6227 dev_priv
->ips
.last_time1
= now
;
6229 dev_priv
->ips
.chipset_power
= ret
;
6234 unsigned long i915_chipset_val(struct drm_i915_private
*dev_priv
)
6238 if (INTEL_INFO(dev_priv
)->gen
!= 5)
6241 spin_lock_irq(&mchdev_lock
);
6243 val
= __i915_chipset_val(dev_priv
);
6245 spin_unlock_irq(&mchdev_lock
);
6250 unsigned long i915_mch_val(struct drm_i915_private
*dev_priv
)
6252 unsigned long m
, x
, b
;
6255 tsfs
= I915_READ(TSFS
);
6257 m
= ((tsfs
& TSFS_SLOPE_MASK
) >> TSFS_SLOPE_SHIFT
);
6258 x
= I915_READ8(TR1
);
6260 b
= tsfs
& TSFS_INTR_MASK
;
6262 return ((m
* x
) / 127) - b
;
6265 static int _pxvid_to_vd(u8 pxvid
)
6270 if (pxvid
>= 8 && pxvid
< 31)
6273 return (pxvid
+ 2) * 125;
6276 static u32
pvid_to_extvid(struct drm_i915_private
*dev_priv
, u8 pxvid
)
6278 const int vd
= _pxvid_to_vd(pxvid
);
6279 const int vm
= vd
- 1125;
6281 if (INTEL_INFO(dev_priv
)->is_mobile
)
6282 return vm
> 0 ? vm
: 0;
6287 static void __i915_update_gfx_val(struct drm_i915_private
*dev_priv
)
6289 u64 now
, diff
, diffms
;
6292 assert_spin_locked(&mchdev_lock
);
6294 now
= ktime_get_raw_ns();
6295 diffms
= now
- dev_priv
->ips
.last_time2
;
6296 do_div(diffms
, NSEC_PER_MSEC
);
6298 /* Don't divide by 0 */
6302 count
= I915_READ(GFXEC
);
6304 if (count
< dev_priv
->ips
.last_count2
) {
6305 diff
= ~0UL - dev_priv
->ips
.last_count2
;
6308 diff
= count
- dev_priv
->ips
.last_count2
;
6311 dev_priv
->ips
.last_count2
= count
;
6312 dev_priv
->ips
.last_time2
= now
;
6314 /* More magic constants... */
6316 diff
= div_u64(diff
, diffms
* 10);
6317 dev_priv
->ips
.gfx_power
= diff
;
6320 void i915_update_gfx_val(struct drm_i915_private
*dev_priv
)
6322 if (INTEL_INFO(dev_priv
)->gen
!= 5)
6325 spin_lock_irq(&mchdev_lock
);
6327 __i915_update_gfx_val(dev_priv
);
6329 spin_unlock_irq(&mchdev_lock
);
6332 static unsigned long __i915_gfx_val(struct drm_i915_private
*dev_priv
)
6334 unsigned long t
, corr
, state1
, corr2
, state2
;
6337 assert_spin_locked(&mchdev_lock
);
6339 pxvid
= I915_READ(PXVFREQ(dev_priv
->rps
.cur_freq
));
6340 pxvid
= (pxvid
>> 24) & 0x7f;
6341 ext_v
= pvid_to_extvid(dev_priv
, pxvid
);
6345 t
= i915_mch_val(dev_priv
);
6347 /* Revel in the empirically derived constants */
6349 /* Correction factor in 1/100000 units */
6351 corr
= ((t
* 2349) + 135940);
6353 corr
= ((t
* 964) + 29317);
6355 corr
= ((t
* 301) + 1004);
6357 corr
= corr
* ((150142 * state1
) / 10000 - 78642);
6359 corr2
= (corr
* dev_priv
->ips
.corr
);
6361 state2
= (corr2
* state1
) / 10000;
6362 state2
/= 100; /* convert to mW */
6364 __i915_update_gfx_val(dev_priv
);
6366 return dev_priv
->ips
.gfx_power
+ state2
;
6369 unsigned long i915_gfx_val(struct drm_i915_private
*dev_priv
)
6373 if (INTEL_INFO(dev_priv
)->gen
!= 5)
6376 spin_lock_irq(&mchdev_lock
);
6378 val
= __i915_gfx_val(dev_priv
);
6380 spin_unlock_irq(&mchdev_lock
);
6386 * i915_read_mch_val - return value for IPS use
6388 * Calculate and return a value for the IPS driver to use when deciding whether
6389 * we have thermal and power headroom to increase CPU or GPU power budget.
6391 unsigned long i915_read_mch_val(void)
6393 struct drm_i915_private
*dev_priv
;
6394 unsigned long chipset_val
, graphics_val
, ret
= 0;
6396 spin_lock_irq(&mchdev_lock
);
6399 dev_priv
= i915_mch_dev
;
6401 chipset_val
= __i915_chipset_val(dev_priv
);
6402 graphics_val
= __i915_gfx_val(dev_priv
);
6404 ret
= chipset_val
+ graphics_val
;
6407 spin_unlock_irq(&mchdev_lock
);
6411 EXPORT_SYMBOL_GPL(i915_read_mch_val
);
6414 * i915_gpu_raise - raise GPU frequency limit
6416 * Raise the limit; IPS indicates we have thermal headroom.
6418 bool i915_gpu_raise(void)
6420 struct drm_i915_private
*dev_priv
;
6423 spin_lock_irq(&mchdev_lock
);
6424 if (!i915_mch_dev
) {
6428 dev_priv
= i915_mch_dev
;
6430 if (dev_priv
->ips
.max_delay
> dev_priv
->ips
.fmax
)
6431 dev_priv
->ips
.max_delay
--;
6434 spin_unlock_irq(&mchdev_lock
);
6438 EXPORT_SYMBOL_GPL(i915_gpu_raise
);
6441 * i915_gpu_lower - lower GPU frequency limit
6443 * IPS indicates we're close to a thermal limit, so throttle back the GPU
6444 * frequency maximum.
6446 bool i915_gpu_lower(void)
6448 struct drm_i915_private
*dev_priv
;
6451 spin_lock_irq(&mchdev_lock
);
6452 if (!i915_mch_dev
) {
6456 dev_priv
= i915_mch_dev
;
6458 if (dev_priv
->ips
.max_delay
< dev_priv
->ips
.min_delay
)
6459 dev_priv
->ips
.max_delay
++;
6462 spin_unlock_irq(&mchdev_lock
);
6466 EXPORT_SYMBOL_GPL(i915_gpu_lower
);
6469 * i915_gpu_busy - indicate GPU business to IPS
6471 * Tell the IPS driver whether or not the GPU is busy.
6473 bool i915_gpu_busy(void)
6477 spin_lock_irq(&mchdev_lock
);
6479 ret
= i915_mch_dev
->gt
.awake
;
6480 spin_unlock_irq(&mchdev_lock
);
6484 EXPORT_SYMBOL_GPL(i915_gpu_busy
);
6487 * i915_gpu_turbo_disable - disable graphics turbo
6489 * Disable graphics turbo by resetting the max frequency and setting the
6490 * current frequency to the default.
6492 bool i915_gpu_turbo_disable(void)
6494 struct drm_i915_private
*dev_priv
;
6497 spin_lock_irq(&mchdev_lock
);
6498 if (!i915_mch_dev
) {
6502 dev_priv
= i915_mch_dev
;
6504 dev_priv
->ips
.max_delay
= dev_priv
->ips
.fstart
;
6506 if (!ironlake_set_drps(dev_priv
, dev_priv
->ips
.fstart
))
6510 spin_unlock_irq(&mchdev_lock
);
6514 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable
);
6517 * Tells the intel_ips driver that the i915 driver is now loaded, if
6518 * IPS got loaded first.
6520 * This awkward dance is so that neither module has to depend on the
6521 * other in order for IPS to do the appropriate communication of
6522 * GPU turbo limits to i915.
6525 ips_ping_for_i915_load(void)
6529 link
= symbol_get(ips_link_to_i915_driver
);
6532 symbol_put(ips_link_to_i915_driver
);
6536 void intel_gpu_ips_init(struct drm_i915_private
*dev_priv
)
6538 /* We only register the i915 ips part with intel-ips once everything is
6539 * set up, to avoid intel-ips sneaking in and reading bogus values. */
6540 spin_lock_irq(&mchdev_lock
);
6541 i915_mch_dev
= dev_priv
;
6542 spin_unlock_irq(&mchdev_lock
);
6544 ips_ping_for_i915_load();
6547 void intel_gpu_ips_teardown(void)
6549 spin_lock_irq(&mchdev_lock
);
6550 i915_mch_dev
= NULL
;
6551 spin_unlock_irq(&mchdev_lock
);
6554 static void intel_init_emon(struct drm_i915_private
*dev_priv
)
6560 /* Disable to program */
6564 /* Program energy weights for various events */
6565 I915_WRITE(SDEW
, 0x15040d00);
6566 I915_WRITE(CSIEW0
, 0x007f0000);
6567 I915_WRITE(CSIEW1
, 0x1e220004);
6568 I915_WRITE(CSIEW2
, 0x04000004);
6570 for (i
= 0; i
< 5; i
++)
6571 I915_WRITE(PEW(i
), 0);
6572 for (i
= 0; i
< 3; i
++)
6573 I915_WRITE(DEW(i
), 0);
6575 /* Program P-state weights to account for frequency power adjustment */
6576 for (i
= 0; i
< 16; i
++) {
6577 u32 pxvidfreq
= I915_READ(PXVFREQ(i
));
6578 unsigned long freq
= intel_pxfreq(pxvidfreq
);
6579 unsigned long vid
= (pxvidfreq
& PXVFREQ_PX_MASK
) >>
6584 val
*= (freq
/ 1000);
6586 val
/= (127*127*900);
6588 DRM_ERROR("bad pxval: %ld\n", val
);
6591 /* Render standby states get 0 weight */
6595 for (i
= 0; i
< 4; i
++) {
6596 u32 val
= (pxw
[i
*4] << 24) | (pxw
[(i
*4)+1] << 16) |
6597 (pxw
[(i
*4)+2] << 8) | (pxw
[(i
*4)+3]);
6598 I915_WRITE(PXW(i
), val
);
6601 /* Adjust magic regs to magic values (more experimental results) */
6602 I915_WRITE(OGW0
, 0);
6603 I915_WRITE(OGW1
, 0);
6604 I915_WRITE(EG0
, 0x00007f00);
6605 I915_WRITE(EG1
, 0x0000000e);
6606 I915_WRITE(EG2
, 0x000e0000);
6607 I915_WRITE(EG3
, 0x68000300);
6608 I915_WRITE(EG4
, 0x42000000);
6609 I915_WRITE(EG5
, 0x00140031);
6613 for (i
= 0; i
< 8; i
++)
6614 I915_WRITE(PXWL(i
), 0);
6616 /* Enable PMON + select events */
6617 I915_WRITE(ECR
, 0x80000019);
6619 lcfuse
= I915_READ(LCFUSE02
);
6621 dev_priv
->ips
.corr
= (lcfuse
& LCFUSE_HIV_MASK
);
6624 void intel_init_gt_powersave(struct drm_i915_private
*dev_priv
)
6627 * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
6630 if (!i915
.enable_rc6
) {
6631 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
6632 intel_runtime_pm_get(dev_priv
);
6635 mutex_lock(&dev_priv
->drm
.struct_mutex
);
6636 mutex_lock(&dev_priv
->rps
.hw_lock
);
6638 /* Initialize RPS limits (for userspace) */
6639 if (IS_CHERRYVIEW(dev_priv
))
6640 cherryview_init_gt_powersave(dev_priv
);
6641 else if (IS_VALLEYVIEW(dev_priv
))
6642 valleyview_init_gt_powersave(dev_priv
);
6643 else if (INTEL_GEN(dev_priv
) >= 6)
6644 gen6_init_rps_frequencies(dev_priv
);
6646 /* Derive initial user preferences/limits from the hardware limits */
6647 dev_priv
->rps
.idle_freq
= dev_priv
->rps
.min_freq
;
6648 dev_priv
->rps
.cur_freq
= dev_priv
->rps
.idle_freq
;
6650 dev_priv
->rps
.max_freq_softlimit
= dev_priv
->rps
.max_freq
;
6651 dev_priv
->rps
.min_freq_softlimit
= dev_priv
->rps
.min_freq
;
6653 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
6654 dev_priv
->rps
.min_freq_softlimit
=
6656 dev_priv
->rps
.efficient_freq
,
6657 intel_freq_opcode(dev_priv
, 450));
6659 /* After setting max-softlimit, find the overclock max freq */
6660 if (IS_GEN6(dev_priv
) ||
6661 IS_IVYBRIDGE(dev_priv
) || IS_HASWELL(dev_priv
)) {
6664 sandybridge_pcode_read(dev_priv
, GEN6_READ_OC_PARAMS
, ¶ms
);
6665 if (params
& BIT(31)) { /* OC supported */
6666 DRM_DEBUG_DRIVER("Overclocking supported, max: %dMHz, overclock: %dMHz\n",
6667 (dev_priv
->rps
.max_freq
& 0xff) * 50,
6668 (params
& 0xff) * 50);
6669 dev_priv
->rps
.max_freq
= params
& 0xff;
6673 /* Finally allow us to boost to max by default */
6674 dev_priv
->rps
.boost_freq
= dev_priv
->rps
.max_freq
;
6676 mutex_unlock(&dev_priv
->rps
.hw_lock
);
6677 mutex_unlock(&dev_priv
->drm
.struct_mutex
);
6679 intel_autoenable_gt_powersave(dev_priv
);
6682 void intel_cleanup_gt_powersave(struct drm_i915_private
*dev_priv
)
6684 if (IS_VALLEYVIEW(dev_priv
))
6685 valleyview_cleanup_gt_powersave(dev_priv
);
6687 if (!i915
.enable_rc6
)
6688 intel_runtime_pm_put(dev_priv
);
6692 * intel_suspend_gt_powersave - suspend PM work and helper threads
6693 * @dev_priv: i915 device
6695 * We don't want to disable RC6 or other features here, we just want
6696 * to make sure any work we've queued has finished and won't bother
6697 * us while we're suspended.
6699 void intel_suspend_gt_powersave(struct drm_i915_private
*dev_priv
)
6701 if (INTEL_GEN(dev_priv
) < 6)
6704 if (cancel_delayed_work_sync(&dev_priv
->rps
.autoenable_work
))
6705 intel_runtime_pm_put(dev_priv
);
6707 /* gen6_rps_idle() will be called later to disable interrupts */
6710 void intel_sanitize_gt_powersave(struct drm_i915_private
*dev_priv
)
6712 dev_priv
->rps
.enabled
= true; /* force disabling */
6713 intel_disable_gt_powersave(dev_priv
);
6715 gen6_reset_rps_interrupts(dev_priv
);
6718 void intel_disable_gt_powersave(struct drm_i915_private
*dev_priv
)
6720 if (!READ_ONCE(dev_priv
->rps
.enabled
))
6723 mutex_lock(&dev_priv
->rps
.hw_lock
);
6725 if (INTEL_GEN(dev_priv
) >= 9) {
6726 gen9_disable_rc6(dev_priv
);
6727 gen9_disable_rps(dev_priv
);
6728 } else if (IS_CHERRYVIEW(dev_priv
)) {
6729 cherryview_disable_rps(dev_priv
);
6730 } else if (IS_VALLEYVIEW(dev_priv
)) {
6731 valleyview_disable_rps(dev_priv
);
6732 } else if (INTEL_GEN(dev_priv
) >= 6) {
6733 gen6_disable_rps(dev_priv
);
6734 } else if (IS_IRONLAKE_M(dev_priv
)) {
6735 ironlake_disable_drps(dev_priv
);
6738 dev_priv
->rps
.enabled
= false;
6739 mutex_unlock(&dev_priv
->rps
.hw_lock
);
6742 void intel_enable_gt_powersave(struct drm_i915_private
*dev_priv
)
6744 /* We shouldn't be disabling as we submit, so this should be less
6745 * racy than it appears!
6747 if (READ_ONCE(dev_priv
->rps
.enabled
))
6750 /* Powersaving is controlled by the host when inside a VM */
6751 if (intel_vgpu_active(dev_priv
))
6754 mutex_lock(&dev_priv
->rps
.hw_lock
);
6756 if (IS_CHERRYVIEW(dev_priv
)) {
6757 cherryview_enable_rps(dev_priv
);
6758 } else if (IS_VALLEYVIEW(dev_priv
)) {
6759 valleyview_enable_rps(dev_priv
);
6760 } else if (INTEL_GEN(dev_priv
) >= 9) {
6761 gen9_enable_rc6(dev_priv
);
6762 gen9_enable_rps(dev_priv
);
6763 if (IS_GEN9_BC(dev_priv
))
6764 gen6_update_ring_freq(dev_priv
);
6765 } else if (IS_BROADWELL(dev_priv
)) {
6766 gen8_enable_rps(dev_priv
);
6767 gen6_update_ring_freq(dev_priv
);
6768 } else if (INTEL_GEN(dev_priv
) >= 6) {
6769 gen6_enable_rps(dev_priv
);
6770 gen6_update_ring_freq(dev_priv
);
6771 } else if (IS_IRONLAKE_M(dev_priv
)) {
6772 ironlake_enable_drps(dev_priv
);
6773 intel_init_emon(dev_priv
);
6776 WARN_ON(dev_priv
->rps
.max_freq
< dev_priv
->rps
.min_freq
);
6777 WARN_ON(dev_priv
->rps
.idle_freq
> dev_priv
->rps
.max_freq
);
6779 WARN_ON(dev_priv
->rps
.efficient_freq
< dev_priv
->rps
.min_freq
);
6780 WARN_ON(dev_priv
->rps
.efficient_freq
> dev_priv
->rps
.max_freq
);
6782 dev_priv
->rps
.enabled
= true;
6783 mutex_unlock(&dev_priv
->rps
.hw_lock
);
6786 static void __intel_autoenable_gt_powersave(struct work_struct
*work
)
6788 struct drm_i915_private
*dev_priv
=
6789 container_of(work
, typeof(*dev_priv
), rps
.autoenable_work
.work
);
6790 struct intel_engine_cs
*rcs
;
6791 struct drm_i915_gem_request
*req
;
6793 if (READ_ONCE(dev_priv
->rps
.enabled
))
6796 rcs
= dev_priv
->engine
[RCS
];
6797 if (rcs
->last_retired_context
)
6800 if (!rcs
->init_context
)
6803 mutex_lock(&dev_priv
->drm
.struct_mutex
);
6805 req
= i915_gem_request_alloc(rcs
, dev_priv
->kernel_context
);
6809 if (!i915
.enable_execlists
&& i915_switch_context(req
) == 0)
6810 rcs
->init_context(req
);
6812 /* Mark the device busy, calling intel_enable_gt_powersave() */
6813 i915_add_request_no_flush(req
);
6816 mutex_unlock(&dev_priv
->drm
.struct_mutex
);
6818 intel_runtime_pm_put(dev_priv
);
6821 void intel_autoenable_gt_powersave(struct drm_i915_private
*dev_priv
)
6823 if (READ_ONCE(dev_priv
->rps
.enabled
))
6826 if (IS_IRONLAKE_M(dev_priv
)) {
6827 ironlake_enable_drps(dev_priv
);
6828 intel_init_emon(dev_priv
);
6829 } else if (INTEL_INFO(dev_priv
)->gen
>= 6) {
6831 * PCU communication is slow and this doesn't need to be
6832 * done at any specific time, so do this out of our fast path
6833 * to make resume and init faster.
6835 * We depend on the HW RC6 power context save/restore
6836 * mechanism when entering D3 through runtime PM suspend. So
6837 * disable RPM until RPS/RC6 is properly setup. We can only
6838 * get here via the driver load/system resume/runtime resume
6839 * paths, so the _noresume version is enough (and in case of
6840 * runtime resume it's necessary).
6842 if (queue_delayed_work(dev_priv
->wq
,
6843 &dev_priv
->rps
.autoenable_work
,
6844 round_jiffies_up_relative(HZ
)))
6845 intel_runtime_pm_get_noresume(dev_priv
);
6849 static void ibx_init_clock_gating(struct drm_i915_private
*dev_priv
)
6852 * On Ibex Peak and Cougar Point, we need to disable clock
6853 * gating for the panel power sequencer or it will fail to
6854 * start up when no ports are active.
6856 I915_WRITE(SOUTH_DSPCLK_GATE_D
, PCH_DPLSUNIT_CLOCK_GATE_DISABLE
);
6859 static void g4x_disable_trickle_feed(struct drm_i915_private
*dev_priv
)
6863 for_each_pipe(dev_priv
, pipe
) {
6864 I915_WRITE(DSPCNTR(pipe
),
6865 I915_READ(DSPCNTR(pipe
)) |
6866 DISPPLANE_TRICKLE_FEED_DISABLE
);
6868 I915_WRITE(DSPSURF(pipe
), I915_READ(DSPSURF(pipe
)));
6869 POSTING_READ(DSPSURF(pipe
));
6873 static void ilk_init_lp_watermarks(struct drm_i915_private
*dev_priv
)
6875 I915_WRITE(WM3_LP_ILK
, I915_READ(WM3_LP_ILK
) & ~WM1_LP_SR_EN
);
6876 I915_WRITE(WM2_LP_ILK
, I915_READ(WM2_LP_ILK
) & ~WM1_LP_SR_EN
);
6877 I915_WRITE(WM1_LP_ILK
, I915_READ(WM1_LP_ILK
) & ~WM1_LP_SR_EN
);
6880 * Don't touch WM1S_LP_EN here.
6881 * Doing so could cause underruns.
6885 static void ironlake_init_clock_gating(struct drm_i915_private
*dev_priv
)
6887 uint32_t dspclk_gate
= ILK_VRHUNIT_CLOCK_GATE_DISABLE
;
6891 * WaFbcDisableDpfcClockGating:ilk
6893 dspclk_gate
|= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE
|
6894 ILK_DPFCUNIT_CLOCK_GATE_DISABLE
|
6895 ILK_DPFDUNIT_CLOCK_GATE_ENABLE
;
6897 I915_WRITE(PCH_3DCGDIS0
,
6898 MARIUNIT_CLOCK_GATE_DISABLE
|
6899 SVSMUNIT_CLOCK_GATE_DISABLE
);
6900 I915_WRITE(PCH_3DCGDIS1
,
6901 VFMUNIT_CLOCK_GATE_DISABLE
);
6904 * According to the spec the following bits should be set in
6905 * order to enable memory self-refresh
6906 * The bit 22/21 of 0x42004
6907 * The bit 5 of 0x42020
6908 * The bit 15 of 0x45000
6910 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6911 (I915_READ(ILK_DISPLAY_CHICKEN2
) |
6912 ILK_DPARB_GATE
| ILK_VSDPFD_FULL
));
6913 dspclk_gate
|= ILK_DPARBUNIT_CLOCK_GATE_ENABLE
;
6914 I915_WRITE(DISP_ARB_CTL
,
6915 (I915_READ(DISP_ARB_CTL
) |
6918 ilk_init_lp_watermarks(dev_priv
);
6921 * Based on the document from hardware guys the following bits
6922 * should be set unconditionally in order to enable FBC.
6923 * The bit 22 of 0x42000
6924 * The bit 22 of 0x42004
6925 * The bit 7,8,9 of 0x42020.
6927 if (IS_IRONLAKE_M(dev_priv
)) {
6928 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
6929 I915_WRITE(ILK_DISPLAY_CHICKEN1
,
6930 I915_READ(ILK_DISPLAY_CHICKEN1
) |
6932 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6933 I915_READ(ILK_DISPLAY_CHICKEN2
) |
6937 I915_WRITE(ILK_DSPCLK_GATE_D
, dspclk_gate
);
6939 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
6940 I915_READ(ILK_DISPLAY_CHICKEN2
) |
6941 ILK_ELPIN_409_SELECT
);
6942 I915_WRITE(_3D_CHICKEN2
,
6943 _3D_CHICKEN2_WM_READ_PIPELINED
<< 16 |
6944 _3D_CHICKEN2_WM_READ_PIPELINED
);
6946 /* WaDisableRenderCachePipelinedFlush:ilk */
6947 I915_WRITE(CACHE_MODE_0
,
6948 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE
));
6950 /* WaDisable_RenderCache_OperationalFlush:ilk */
6951 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
6953 g4x_disable_trickle_feed(dev_priv
);
6955 ibx_init_clock_gating(dev_priv
);
6958 static void cpt_init_clock_gating(struct drm_i915_private
*dev_priv
)
6964 * On Ibex Peak and Cougar Point, we need to disable clock
6965 * gating for the panel power sequencer or it will fail to
6966 * start up when no ports are active.
6968 I915_WRITE(SOUTH_DSPCLK_GATE_D
, PCH_DPLSUNIT_CLOCK_GATE_DISABLE
|
6969 PCH_DPLUNIT_CLOCK_GATE_DISABLE
|
6970 PCH_CPUNIT_CLOCK_GATE_DISABLE
);
6971 I915_WRITE(SOUTH_CHICKEN2
, I915_READ(SOUTH_CHICKEN2
) |
6972 DPLS_EDP_PPS_FIX_DIS
);
6973 /* The below fixes the weird display corruption, a few pixels shifted
6974 * downward, on (only) LVDS of some HP laptops with IVY.
6976 for_each_pipe(dev_priv
, pipe
) {
6977 val
= I915_READ(TRANS_CHICKEN2(pipe
));
6978 val
|= TRANS_CHICKEN2_TIMING_OVERRIDE
;
6979 val
&= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED
;
6980 if (dev_priv
->vbt
.fdi_rx_polarity_inverted
)
6981 val
|= TRANS_CHICKEN2_FDI_POLARITY_REVERSED
;
6982 val
&= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK
;
6983 val
&= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER
;
6984 val
&= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH
;
6985 I915_WRITE(TRANS_CHICKEN2(pipe
), val
);
6987 /* WADP0ClockGatingDisable */
6988 for_each_pipe(dev_priv
, pipe
) {
6989 I915_WRITE(TRANS_CHICKEN1(pipe
),
6990 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE
);
6994 static void gen6_check_mch_setup(struct drm_i915_private
*dev_priv
)
6998 tmp
= I915_READ(MCH_SSKPD
);
6999 if ((tmp
& MCH_SSKPD_WM0_MASK
) != MCH_SSKPD_WM0_VAL
)
7000 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
7004 static void gen6_init_clock_gating(struct drm_i915_private
*dev_priv
)
7006 uint32_t dspclk_gate
= ILK_VRHUNIT_CLOCK_GATE_DISABLE
;
7008 I915_WRITE(ILK_DSPCLK_GATE_D
, dspclk_gate
);
7010 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
7011 I915_READ(ILK_DISPLAY_CHICKEN2
) |
7012 ILK_ELPIN_409_SELECT
);
7014 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
7015 I915_WRITE(_3D_CHICKEN
,
7016 _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB
));
7018 /* WaDisable_RenderCache_OperationalFlush:snb */
7019 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
7022 * BSpec recoomends 8x4 when MSAA is used,
7023 * however in practice 16x4 seems fastest.
7025 * Note that PS/WM thread counts depend on the WIZ hashing
7026 * disable bit, which we don't touch here, but it's good
7027 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7029 I915_WRITE(GEN6_GT_MODE
,
7030 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
7032 ilk_init_lp_watermarks(dev_priv
);
7034 I915_WRITE(CACHE_MODE_0
,
7035 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB
));
7037 I915_WRITE(GEN6_UCGCTL1
,
7038 I915_READ(GEN6_UCGCTL1
) |
7039 GEN6_BLBUNIT_CLOCK_GATE_DISABLE
|
7040 GEN6_CSUNIT_CLOCK_GATE_DISABLE
);
7042 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
7043 * gating disable must be set. Failure to set it results in
7044 * flickering pixels due to Z write ordering failures after
7045 * some amount of runtime in the Mesa "fire" demo, and Unigine
7046 * Sanctuary and Tropics, and apparently anything else with
7047 * alpha test or pixel discard.
7049 * According to the spec, bit 11 (RCCUNIT) must also be set,
7050 * but we didn't debug actual testcases to find it out.
7052 * WaDisableRCCUnitClockGating:snb
7053 * WaDisableRCPBUnitClockGating:snb
7055 I915_WRITE(GEN6_UCGCTL2
,
7056 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE
|
7057 GEN6_RCCUNIT_CLOCK_GATE_DISABLE
);
7059 /* WaStripsFansDisableFastClipPerformanceFix:snb */
7060 I915_WRITE(_3D_CHICKEN3
,
7061 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL
));
7065 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
7066 * 3DSTATE_SF number of SF output attributes is more than 16."
7068 I915_WRITE(_3D_CHICKEN3
,
7069 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH
));
7072 * According to the spec the following bits should be
7073 * set in order to enable memory self-refresh and fbc:
7074 * The bit21 and bit22 of 0x42000
7075 * The bit21 and bit22 of 0x42004
7076 * The bit5 and bit7 of 0x42020
7077 * The bit14 of 0x70180
7078 * The bit14 of 0x71180
7080 * WaFbcAsynchFlipDisableFbcQueue:snb
7082 I915_WRITE(ILK_DISPLAY_CHICKEN1
,
7083 I915_READ(ILK_DISPLAY_CHICKEN1
) |
7084 ILK_FBCQ_DIS
| ILK_PABSTRETCH_DIS
);
7085 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
7086 I915_READ(ILK_DISPLAY_CHICKEN2
) |
7087 ILK_DPARB_GATE
| ILK_VSDPFD_FULL
);
7088 I915_WRITE(ILK_DSPCLK_GATE_D
,
7089 I915_READ(ILK_DSPCLK_GATE_D
) |
7090 ILK_DPARBUNIT_CLOCK_GATE_ENABLE
|
7091 ILK_DPFDUNIT_CLOCK_GATE_ENABLE
);
7093 g4x_disable_trickle_feed(dev_priv
);
7095 cpt_init_clock_gating(dev_priv
);
7097 gen6_check_mch_setup(dev_priv
);
7100 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private
*dev_priv
)
7102 uint32_t reg
= I915_READ(GEN7_FF_THREAD_MODE
);
7105 * WaVSThreadDispatchOverride:ivb,vlv
7107 * This actually overrides the dispatch
7108 * mode for all thread types.
7110 reg
&= ~GEN7_FF_SCHED_MASK
;
7111 reg
|= GEN7_FF_TS_SCHED_HW
;
7112 reg
|= GEN7_FF_VS_SCHED_HW
;
7113 reg
|= GEN7_FF_DS_SCHED_HW
;
7115 I915_WRITE(GEN7_FF_THREAD_MODE
, reg
);
7118 static void lpt_init_clock_gating(struct drm_i915_private
*dev_priv
)
7121 * TODO: this bit should only be enabled when really needed, then
7122 * disabled when not needed anymore in order to save power.
7124 if (HAS_PCH_LPT_LP(dev_priv
))
7125 I915_WRITE(SOUTH_DSPCLK_GATE_D
,
7126 I915_READ(SOUTH_DSPCLK_GATE_D
) |
7127 PCH_LP_PARTITION_LEVEL_DISABLE
);
7129 /* WADPOClockGatingDisable:hsw */
7130 I915_WRITE(TRANS_CHICKEN1(PIPE_A
),
7131 I915_READ(TRANS_CHICKEN1(PIPE_A
)) |
7132 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE
);
7135 static void lpt_suspend_hw(struct drm_i915_private
*dev_priv
)
7137 if (HAS_PCH_LPT_LP(dev_priv
)) {
7138 uint32_t val
= I915_READ(SOUTH_DSPCLK_GATE_D
);
7140 val
&= ~PCH_LP_PARTITION_LEVEL_DISABLE
;
7141 I915_WRITE(SOUTH_DSPCLK_GATE_D
, val
);
7145 static void gen8_set_l3sqc_credits(struct drm_i915_private
*dev_priv
,
7146 int general_prio_credits
,
7147 int high_prio_credits
)
7151 /* WaTempDisableDOPClkGating:bdw */
7152 misccpctl
= I915_READ(GEN7_MISCCPCTL
);
7153 I915_WRITE(GEN7_MISCCPCTL
, misccpctl
& ~GEN7_DOP_CLOCK_GATE_ENABLE
);
7155 I915_WRITE(GEN8_L3SQCREG1
,
7156 L3_GENERAL_PRIO_CREDITS(general_prio_credits
) |
7157 L3_HIGH_PRIO_CREDITS(high_prio_credits
));
7160 * Wait at least 100 clocks before re-enabling clock gating.
7161 * See the definition of L3SQCREG1 in BSpec.
7163 POSTING_READ(GEN8_L3SQCREG1
);
7165 I915_WRITE(GEN7_MISCCPCTL
, misccpctl
);
7168 static void kabylake_init_clock_gating(struct drm_i915_private
*dev_priv
)
7170 gen9_init_clock_gating(dev_priv
);
7172 /* WaDisableSDEUnitClockGating:kbl */
7173 if (IS_KBL_REVID(dev_priv
, 0, KBL_REVID_B0
))
7174 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
7175 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
7177 /* WaDisableGamClockGating:kbl */
7178 if (IS_KBL_REVID(dev_priv
, 0, KBL_REVID_B0
))
7179 I915_WRITE(GEN6_UCGCTL1
, I915_READ(GEN6_UCGCTL1
) |
7180 GEN6_GAMUNIT_CLOCK_GATE_DISABLE
);
7182 /* WaFbcNukeOnHostModify:kbl */
7183 I915_WRITE(ILK_DPFC_CHICKEN
, I915_READ(ILK_DPFC_CHICKEN
) |
7184 ILK_DPFC_NUKE_ON_ANY_MODIFICATION
);
7187 static void skylake_init_clock_gating(struct drm_i915_private
*dev_priv
)
7189 gen9_init_clock_gating(dev_priv
);
7191 /* WAC6entrylatency:skl */
7192 I915_WRITE(FBC_LLC_READ_CTRL
, I915_READ(FBC_LLC_READ_CTRL
) |
7193 FBC_LLC_FULLY_OPEN
);
7195 /* WaFbcNukeOnHostModify:skl */
7196 I915_WRITE(ILK_DPFC_CHICKEN
, I915_READ(ILK_DPFC_CHICKEN
) |
7197 ILK_DPFC_NUKE_ON_ANY_MODIFICATION
);
7200 static void broadwell_init_clock_gating(struct drm_i915_private
*dev_priv
)
7204 ilk_init_lp_watermarks(dev_priv
);
7206 /* WaSwitchSolVfFArbitrationPriority:bdw */
7207 I915_WRITE(GAM_ECOCHK
, I915_READ(GAM_ECOCHK
) | HSW_ECOCHK_ARB_PRIO_SOL
);
7209 /* WaPsrDPAMaskVBlankInSRD:bdw */
7210 I915_WRITE(CHICKEN_PAR1_1
,
7211 I915_READ(CHICKEN_PAR1_1
) | DPA_MASK_VBLANK_SRD
);
7213 /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
7214 for_each_pipe(dev_priv
, pipe
) {
7215 I915_WRITE(CHICKEN_PIPESL_1(pipe
),
7216 I915_READ(CHICKEN_PIPESL_1(pipe
)) |
7217 BDW_DPRS_MASK_VBLANK_SRD
);
7220 /* WaVSRefCountFullforceMissDisable:bdw */
7221 /* WaDSRefCountFullforceMissDisable:bdw */
7222 I915_WRITE(GEN7_FF_THREAD_MODE
,
7223 I915_READ(GEN7_FF_THREAD_MODE
) &
7224 ~(GEN8_FF_DS_REF_CNT_FFME
| GEN7_FF_VS_REF_CNT_FFME
));
7226 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL
,
7227 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE
));
7229 /* WaDisableSDEUnitClockGating:bdw */
7230 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
7231 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
7233 /* WaProgramL3SqcReg1Default:bdw */
7234 gen8_set_l3sqc_credits(dev_priv
, 30, 2);
7237 * WaGttCachingOffByDefault:bdw
7238 * GTT cache may not work with big pages, so if those
7239 * are ever enabled GTT cache may need to be disabled.
7241 I915_WRITE(HSW_GTT_CACHE_EN
, GTT_CACHE_EN_ALL
);
7243 /* WaKVMNotificationOnConfigChange:bdw */
7244 I915_WRITE(CHICKEN_PAR2_1
, I915_READ(CHICKEN_PAR2_1
)
7245 | KVM_CONFIG_CHANGE_NOTIFICATION_SELECT
);
7247 lpt_init_clock_gating(dev_priv
);
7249 /* WaDisableDopClockGating:bdw
7251 * Also see the CHICKEN2 write in bdw_init_workarounds() to disable DOP
7254 I915_WRITE(GEN6_UCGCTL1
,
7255 I915_READ(GEN6_UCGCTL1
) | GEN6_EU_TCUNIT_CLOCK_GATE_DISABLE
);
7258 static void haswell_init_clock_gating(struct drm_i915_private
*dev_priv
)
7260 ilk_init_lp_watermarks(dev_priv
);
7262 /* L3 caching of data atomics doesn't work -- disable it. */
7263 I915_WRITE(HSW_SCRATCH1
, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE
);
7264 I915_WRITE(HSW_ROW_CHICKEN3
,
7265 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE
));
7267 /* This is required by WaCatErrorRejectionIssue:hsw */
7268 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
7269 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
7270 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
7272 /* WaVSRefCountFullforceMissDisable:hsw */
7273 I915_WRITE(GEN7_FF_THREAD_MODE
,
7274 I915_READ(GEN7_FF_THREAD_MODE
) & ~GEN7_FF_VS_REF_CNT_FFME
);
7276 /* WaDisable_RenderCache_OperationalFlush:hsw */
7277 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
7279 /* enable HiZ Raw Stall Optimization */
7280 I915_WRITE(CACHE_MODE_0_GEN7
,
7281 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE
));
7283 /* WaDisable4x2SubspanOptimization:hsw */
7284 I915_WRITE(CACHE_MODE_1
,
7285 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
7288 * BSpec recommends 8x4 when MSAA is used,
7289 * however in practice 16x4 seems fastest.
7291 * Note that PS/WM thread counts depend on the WIZ hashing
7292 * disable bit, which we don't touch here, but it's good
7293 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7295 I915_WRITE(GEN7_GT_MODE
,
7296 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
7298 /* WaSampleCChickenBitEnable:hsw */
7299 I915_WRITE(HALF_SLICE_CHICKEN3
,
7300 _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE
));
7302 /* WaSwitchSolVfFArbitrationPriority:hsw */
7303 I915_WRITE(GAM_ECOCHK
, I915_READ(GAM_ECOCHK
) | HSW_ECOCHK_ARB_PRIO_SOL
);
7305 /* WaRsPkgCStateDisplayPMReq:hsw */
7306 I915_WRITE(CHICKEN_PAR1_1
,
7307 I915_READ(CHICKEN_PAR1_1
) | FORCE_ARB_IDLE_PLANES
);
7309 lpt_init_clock_gating(dev_priv
);
7312 static void ivybridge_init_clock_gating(struct drm_i915_private
*dev_priv
)
7316 ilk_init_lp_watermarks(dev_priv
);
7318 I915_WRITE(ILK_DSPCLK_GATE_D
, ILK_VRHUNIT_CLOCK_GATE_DISABLE
);
7320 /* WaDisableEarlyCull:ivb */
7321 I915_WRITE(_3D_CHICKEN3
,
7322 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL
));
7324 /* WaDisableBackToBackFlipFix:ivb */
7325 I915_WRITE(IVB_CHICKEN3
,
7326 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE
|
7327 CHICKEN3_DGMG_DONE_FIX_DISABLE
);
7329 /* WaDisablePSDDualDispatchEnable:ivb */
7330 if (IS_IVB_GT1(dev_priv
))
7331 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1
,
7332 _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE
));
7334 /* WaDisable_RenderCache_OperationalFlush:ivb */
7335 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
7337 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
7338 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1
,
7339 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC
);
7341 /* WaApplyL3ControlAndL3ChickenMode:ivb */
7342 I915_WRITE(GEN7_L3CNTLREG1
,
7343 GEN7_WA_FOR_GEN7_L3_CONTROL
);
7344 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER
,
7345 GEN7_WA_L3_CHICKEN_MODE
);
7346 if (IS_IVB_GT1(dev_priv
))
7347 I915_WRITE(GEN7_ROW_CHICKEN2
,
7348 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
7350 /* must write both registers */
7351 I915_WRITE(GEN7_ROW_CHICKEN2
,
7352 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
7353 I915_WRITE(GEN7_ROW_CHICKEN2_GT2
,
7354 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
7357 /* WaForceL3Serialization:ivb */
7358 I915_WRITE(GEN7_L3SQCREG4
, I915_READ(GEN7_L3SQCREG4
) &
7359 ~L3SQ_URB_READ_CAM_MATCH_DISABLE
);
7362 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7363 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
7365 I915_WRITE(GEN6_UCGCTL2
,
7366 GEN6_RCZUNIT_CLOCK_GATE_DISABLE
);
7368 /* This is required by WaCatErrorRejectionIssue:ivb */
7369 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
7370 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
7371 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
7373 g4x_disable_trickle_feed(dev_priv
);
7375 gen7_setup_fixed_func_scheduler(dev_priv
);
7377 if (0) { /* causes HiZ corruption on ivb:gt1 */
7378 /* enable HiZ Raw Stall Optimization */
7379 I915_WRITE(CACHE_MODE_0_GEN7
,
7380 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE
));
7383 /* WaDisable4x2SubspanOptimization:ivb */
7384 I915_WRITE(CACHE_MODE_1
,
7385 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
7388 * BSpec recommends 8x4 when MSAA is used,
7389 * however in practice 16x4 seems fastest.
7391 * Note that PS/WM thread counts depend on the WIZ hashing
7392 * disable bit, which we don't touch here, but it's good
7393 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7395 I915_WRITE(GEN7_GT_MODE
,
7396 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
7398 snpcr
= I915_READ(GEN6_MBCUNIT_SNPCR
);
7399 snpcr
&= ~GEN6_MBC_SNPCR_MASK
;
7400 snpcr
|= GEN6_MBC_SNPCR_MED
;
7401 I915_WRITE(GEN6_MBCUNIT_SNPCR
, snpcr
);
7403 if (!HAS_PCH_NOP(dev_priv
))
7404 cpt_init_clock_gating(dev_priv
);
7406 gen6_check_mch_setup(dev_priv
);
7409 static void valleyview_init_clock_gating(struct drm_i915_private
*dev_priv
)
7411 /* WaDisableEarlyCull:vlv */
7412 I915_WRITE(_3D_CHICKEN3
,
7413 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL
));
7415 /* WaDisableBackToBackFlipFix:vlv */
7416 I915_WRITE(IVB_CHICKEN3
,
7417 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE
|
7418 CHICKEN3_DGMG_DONE_FIX_DISABLE
);
7420 /* WaPsdDispatchEnable:vlv */
7421 /* WaDisablePSDDualDispatchEnable:vlv */
7422 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1
,
7423 _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP
|
7424 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE
));
7426 /* WaDisable_RenderCache_OperationalFlush:vlv */
7427 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
7429 /* WaForceL3Serialization:vlv */
7430 I915_WRITE(GEN7_L3SQCREG4
, I915_READ(GEN7_L3SQCREG4
) &
7431 ~L3SQ_URB_READ_CAM_MATCH_DISABLE
);
7433 /* WaDisableDopClockGating:vlv */
7434 I915_WRITE(GEN7_ROW_CHICKEN2
,
7435 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
7437 /* This is required by WaCatErrorRejectionIssue:vlv */
7438 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
7439 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
7440 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
7442 gen7_setup_fixed_func_scheduler(dev_priv
);
7445 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7446 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
7448 I915_WRITE(GEN6_UCGCTL2
,
7449 GEN6_RCZUNIT_CLOCK_GATE_DISABLE
);
7451 /* WaDisableL3Bank2xClockGate:vlv
7452 * Disabling L3 clock gating- MMIO 940c[25] = 1
7453 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
7454 I915_WRITE(GEN7_UCGCTL4
,
7455 I915_READ(GEN7_UCGCTL4
) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE
);
7458 * BSpec says this must be set, even though
7459 * WaDisable4x2SubspanOptimization isn't listed for VLV.
7461 I915_WRITE(CACHE_MODE_1
,
7462 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
7465 * BSpec recommends 8x4 when MSAA is used,
7466 * however in practice 16x4 seems fastest.
7468 * Note that PS/WM thread counts depend on the WIZ hashing
7469 * disable bit, which we don't touch here, but it's good
7470 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7472 I915_WRITE(GEN7_GT_MODE
,
7473 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
7476 * WaIncreaseL3CreditsForVLVB0:vlv
7477 * This is the hardware default actually.
7479 I915_WRITE(GEN7_L3SQCREG1
, VLV_B0_WA_L3SQCREG1_VALUE
);
7482 * WaDisableVLVClockGating_VBIIssue:vlv
7483 * Disable clock gating on th GCFG unit to prevent a delay
7484 * in the reporting of vblank events.
7486 I915_WRITE(VLV_GUNIT_CLOCK_GATE
, GCFG_DIS
);
7489 static void cherryview_init_clock_gating(struct drm_i915_private
*dev_priv
)
7491 /* WaVSRefCountFullforceMissDisable:chv */
7492 /* WaDSRefCountFullforceMissDisable:chv */
7493 I915_WRITE(GEN7_FF_THREAD_MODE
,
7494 I915_READ(GEN7_FF_THREAD_MODE
) &
7495 ~(GEN8_FF_DS_REF_CNT_FFME
| GEN7_FF_VS_REF_CNT_FFME
));
7497 /* WaDisableSemaphoreAndSyncFlipWait:chv */
7498 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL
,
7499 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE
));
7501 /* WaDisableCSUnitClockGating:chv */
7502 I915_WRITE(GEN6_UCGCTL1
, I915_READ(GEN6_UCGCTL1
) |
7503 GEN6_CSUNIT_CLOCK_GATE_DISABLE
);
7505 /* WaDisableSDEUnitClockGating:chv */
7506 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
7507 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
7510 * WaProgramL3SqcReg1Default:chv
7511 * See gfxspecs/Related Documents/Performance Guide/
7512 * LSQC Setting Recommendations.
7514 gen8_set_l3sqc_credits(dev_priv
, 38, 2);
7517 * GTT cache may not work with big pages, so if those
7518 * are ever enabled GTT cache may need to be disabled.
7520 I915_WRITE(HSW_GTT_CACHE_EN
, GTT_CACHE_EN_ALL
);
7523 static void g4x_init_clock_gating(struct drm_i915_private
*dev_priv
)
7525 uint32_t dspclk_gate
;
7527 I915_WRITE(RENCLK_GATE_D1
, 0);
7528 I915_WRITE(RENCLK_GATE_D2
, VF_UNIT_CLOCK_GATE_DISABLE
|
7529 GS_UNIT_CLOCK_GATE_DISABLE
|
7530 CL_UNIT_CLOCK_GATE_DISABLE
);
7531 I915_WRITE(RAMCLK_GATE_D
, 0);
7532 dspclk_gate
= VRHUNIT_CLOCK_GATE_DISABLE
|
7533 OVRUNIT_CLOCK_GATE_DISABLE
|
7534 OVCUNIT_CLOCK_GATE_DISABLE
;
7535 if (IS_GM45(dev_priv
))
7536 dspclk_gate
|= DSSUNIT_CLOCK_GATE_DISABLE
;
7537 I915_WRITE(DSPCLK_GATE_D
, dspclk_gate
);
7539 /* WaDisableRenderCachePipelinedFlush */
7540 I915_WRITE(CACHE_MODE_0
,
7541 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE
));
7543 /* WaDisable_RenderCache_OperationalFlush:g4x */
7544 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
7546 g4x_disable_trickle_feed(dev_priv
);
7549 static void crestline_init_clock_gating(struct drm_i915_private
*dev_priv
)
7551 I915_WRITE(RENCLK_GATE_D1
, I965_RCC_CLOCK_GATE_DISABLE
);
7552 I915_WRITE(RENCLK_GATE_D2
, 0);
7553 I915_WRITE(DSPCLK_GATE_D
, 0);
7554 I915_WRITE(RAMCLK_GATE_D
, 0);
7555 I915_WRITE16(DEUC
, 0);
7556 I915_WRITE(MI_ARB_STATE
,
7557 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
7559 /* WaDisable_RenderCache_OperationalFlush:gen4 */
7560 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
7563 static void broadwater_init_clock_gating(struct drm_i915_private
*dev_priv
)
7565 I915_WRITE(RENCLK_GATE_D1
, I965_RCZ_CLOCK_GATE_DISABLE
|
7566 I965_RCC_CLOCK_GATE_DISABLE
|
7567 I965_RCPB_CLOCK_GATE_DISABLE
|
7568 I965_ISC_CLOCK_GATE_DISABLE
|
7569 I965_FBC_CLOCK_GATE_DISABLE
);
7570 I915_WRITE(RENCLK_GATE_D2
, 0);
7571 I915_WRITE(MI_ARB_STATE
,
7572 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
7574 /* WaDisable_RenderCache_OperationalFlush:gen4 */
7575 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
7578 static void gen3_init_clock_gating(struct drm_i915_private
*dev_priv
)
7580 u32 dstate
= I915_READ(D_STATE
);
7582 dstate
|= DSTATE_PLL_D3_OFF
| DSTATE_GFX_CLOCK_GATING
|
7583 DSTATE_DOT_CLOCK_GATING
;
7584 I915_WRITE(D_STATE
, dstate
);
7586 if (IS_PINEVIEW(dev_priv
))
7587 I915_WRITE(ECOSKPD
, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY
));
7589 /* IIR "flip pending" means done if this bit is set */
7590 I915_WRITE(ECOSKPD
, _MASKED_BIT_DISABLE(ECO_FLIP_DONE
));
7592 /* interrupts should cause a wake up from C3 */
7593 I915_WRITE(INSTPM
, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN
));
7595 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
7596 I915_WRITE(MI_ARB_STATE
, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE
));
7598 I915_WRITE(MI_ARB_STATE
,
7599 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
7602 static void i85x_init_clock_gating(struct drm_i915_private
*dev_priv
)
7604 I915_WRITE(RENCLK_GATE_D1
, SV_CLOCK_GATE_DISABLE
);
7606 /* interrupts should cause a wake up from C3 */
7607 I915_WRITE(MI_STATE
, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN
) |
7608 _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE
));
7610 I915_WRITE(MEM_MODE
,
7611 _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE
));
7614 static void i830_init_clock_gating(struct drm_i915_private
*dev_priv
)
7616 I915_WRITE(MEM_MODE
,
7617 _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE
) |
7618 _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE
));
7621 void intel_init_clock_gating(struct drm_i915_private
*dev_priv
)
7623 dev_priv
->display
.init_clock_gating(dev_priv
);
7626 void intel_suspend_hw(struct drm_i915_private
*dev_priv
)
7628 if (HAS_PCH_LPT(dev_priv
))
7629 lpt_suspend_hw(dev_priv
);
7632 static void nop_init_clock_gating(struct drm_i915_private
*dev_priv
)
7634 DRM_DEBUG_KMS("No clock gating settings or workarounds applied.\n");
7638 * intel_init_clock_gating_hooks - setup the clock gating hooks
7639 * @dev_priv: device private
7641 * Setup the hooks that configure which clocks of a given platform can be
7642 * gated and also apply various GT and display specific workarounds for these
7643 * platforms. Note that some GT specific workarounds are applied separately
7644 * when GPU contexts or batchbuffers start their execution.
7646 void intel_init_clock_gating_hooks(struct drm_i915_private
*dev_priv
)
7648 if (IS_SKYLAKE(dev_priv
))
7649 dev_priv
->display
.init_clock_gating
= skylake_init_clock_gating
;
7650 else if (IS_KABYLAKE(dev_priv
))
7651 dev_priv
->display
.init_clock_gating
= kabylake_init_clock_gating
;
7652 else if (IS_BROXTON(dev_priv
))
7653 dev_priv
->display
.init_clock_gating
= bxt_init_clock_gating
;
7654 else if (IS_GEMINILAKE(dev_priv
))
7655 dev_priv
->display
.init_clock_gating
= glk_init_clock_gating
;
7656 else if (IS_BROADWELL(dev_priv
))
7657 dev_priv
->display
.init_clock_gating
= broadwell_init_clock_gating
;
7658 else if (IS_CHERRYVIEW(dev_priv
))
7659 dev_priv
->display
.init_clock_gating
= cherryview_init_clock_gating
;
7660 else if (IS_HASWELL(dev_priv
))
7661 dev_priv
->display
.init_clock_gating
= haswell_init_clock_gating
;
7662 else if (IS_IVYBRIDGE(dev_priv
))
7663 dev_priv
->display
.init_clock_gating
= ivybridge_init_clock_gating
;
7664 else if (IS_VALLEYVIEW(dev_priv
))
7665 dev_priv
->display
.init_clock_gating
= valleyview_init_clock_gating
;
7666 else if (IS_GEN6(dev_priv
))
7667 dev_priv
->display
.init_clock_gating
= gen6_init_clock_gating
;
7668 else if (IS_GEN5(dev_priv
))
7669 dev_priv
->display
.init_clock_gating
= ironlake_init_clock_gating
;
7670 else if (IS_G4X(dev_priv
))
7671 dev_priv
->display
.init_clock_gating
= g4x_init_clock_gating
;
7672 else if (IS_I965GM(dev_priv
))
7673 dev_priv
->display
.init_clock_gating
= crestline_init_clock_gating
;
7674 else if (IS_I965G(dev_priv
))
7675 dev_priv
->display
.init_clock_gating
= broadwater_init_clock_gating
;
7676 else if (IS_GEN3(dev_priv
))
7677 dev_priv
->display
.init_clock_gating
= gen3_init_clock_gating
;
7678 else if (IS_I85X(dev_priv
) || IS_I865G(dev_priv
))
7679 dev_priv
->display
.init_clock_gating
= i85x_init_clock_gating
;
7680 else if (IS_GEN2(dev_priv
))
7681 dev_priv
->display
.init_clock_gating
= i830_init_clock_gating
;
7683 MISSING_CASE(INTEL_DEVID(dev_priv
));
7684 dev_priv
->display
.init_clock_gating
= nop_init_clock_gating
;
7688 /* Set up chip specific power management-related functions */
7689 void intel_init_pm(struct drm_i915_private
*dev_priv
)
7691 intel_fbc_init(dev_priv
);
7694 if (IS_PINEVIEW(dev_priv
))
7695 i915_pineview_get_mem_freq(dev_priv
);
7696 else if (IS_GEN5(dev_priv
))
7697 i915_ironlake_get_mem_freq(dev_priv
);
7699 /* For FIFO watermark updates */
7700 if (INTEL_GEN(dev_priv
) >= 9) {
7701 skl_setup_wm_latency(dev_priv
);
7702 dev_priv
->display
.initial_watermarks
= skl_initial_wm
;
7703 dev_priv
->display
.atomic_update_watermarks
= skl_atomic_update_crtc_wm
;
7704 dev_priv
->display
.compute_global_watermarks
= skl_compute_wm
;
7705 } else if (HAS_PCH_SPLIT(dev_priv
)) {
7706 ilk_setup_wm_latency(dev_priv
);
7708 if ((IS_GEN5(dev_priv
) && dev_priv
->wm
.pri_latency
[1] &&
7709 dev_priv
->wm
.spr_latency
[1] && dev_priv
->wm
.cur_latency
[1]) ||
7710 (!IS_GEN5(dev_priv
) && dev_priv
->wm
.pri_latency
[0] &&
7711 dev_priv
->wm
.spr_latency
[0] && dev_priv
->wm
.cur_latency
[0])) {
7712 dev_priv
->display
.compute_pipe_wm
= ilk_compute_pipe_wm
;
7713 dev_priv
->display
.compute_intermediate_wm
=
7714 ilk_compute_intermediate_wm
;
7715 dev_priv
->display
.initial_watermarks
=
7716 ilk_initial_watermarks
;
7717 dev_priv
->display
.optimize_watermarks
=
7718 ilk_optimize_watermarks
;
7720 DRM_DEBUG_KMS("Failed to read display plane latency. "
7723 } else if (IS_VALLEYVIEW(dev_priv
) || IS_CHERRYVIEW(dev_priv
)) {
7724 vlv_setup_wm_latency(dev_priv
);
7725 dev_priv
->display
.update_wm
= vlv_update_wm
;
7726 } else if (IS_PINEVIEW(dev_priv
)) {
7727 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv
),
7730 dev_priv
->mem_freq
)) {
7731 DRM_INFO("failed to find known CxSR latency "
7732 "(found ddr%s fsb freq %d, mem freq %d), "
7734 (dev_priv
->is_ddr3
== 1) ? "3" : "2",
7735 dev_priv
->fsb_freq
, dev_priv
->mem_freq
);
7736 /* Disable CxSR and never update its watermark again */
7737 intel_set_memory_cxsr(dev_priv
, false);
7738 dev_priv
->display
.update_wm
= NULL
;
7740 dev_priv
->display
.update_wm
= pineview_update_wm
;
7741 } else if (IS_G4X(dev_priv
)) {
7742 dev_priv
->display
.update_wm
= g4x_update_wm
;
7743 } else if (IS_GEN4(dev_priv
)) {
7744 dev_priv
->display
.update_wm
= i965_update_wm
;
7745 } else if (IS_GEN3(dev_priv
)) {
7746 dev_priv
->display
.update_wm
= i9xx_update_wm
;
7747 dev_priv
->display
.get_fifo_size
= i9xx_get_fifo_size
;
7748 } else if (IS_GEN2(dev_priv
)) {
7749 if (INTEL_INFO(dev_priv
)->num_pipes
== 1) {
7750 dev_priv
->display
.update_wm
= i845_update_wm
;
7751 dev_priv
->display
.get_fifo_size
= i845_get_fifo_size
;
7753 dev_priv
->display
.update_wm
= i9xx_update_wm
;
7754 dev_priv
->display
.get_fifo_size
= i830_get_fifo_size
;
7757 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
7761 static inline int gen6_check_mailbox_status(struct drm_i915_private
*dev_priv
)
7764 I915_READ_FW(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_ERROR_MASK
;
7767 case GEN6_PCODE_SUCCESS
:
7769 case GEN6_PCODE_UNIMPLEMENTED_CMD
:
7770 case GEN6_PCODE_ILLEGAL_CMD
:
7772 case GEN6_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE
:
7773 case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE
:
7775 case GEN6_PCODE_TIMEOUT
:
7783 static inline int gen7_check_mailbox_status(struct drm_i915_private
*dev_priv
)
7786 I915_READ_FW(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_ERROR_MASK
;
7789 case GEN6_PCODE_SUCCESS
:
7791 case GEN6_PCODE_ILLEGAL_CMD
:
7793 case GEN7_PCODE_TIMEOUT
:
7795 case GEN7_PCODE_ILLEGAL_DATA
:
7797 case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE
:
7800 MISSING_CASE(flags
);
7805 int sandybridge_pcode_read(struct drm_i915_private
*dev_priv
, u32 mbox
, u32
*val
)
7809 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
7811 /* GEN6_PCODE_* are outside of the forcewake domain, we can
7812 * use te fw I915_READ variants to reduce the amount of work
7813 * required when reading/writing.
7816 if (I915_READ_FW(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) {
7817 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
7821 I915_WRITE_FW(GEN6_PCODE_DATA
, *val
);
7822 I915_WRITE_FW(GEN6_PCODE_DATA1
, 0);
7823 I915_WRITE_FW(GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
| mbox
);
7825 if (intel_wait_for_register_fw(dev_priv
,
7826 GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
, 0,
7828 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox
);
7832 *val
= I915_READ_FW(GEN6_PCODE_DATA
);
7833 I915_WRITE_FW(GEN6_PCODE_DATA
, 0);
7835 if (INTEL_GEN(dev_priv
) > 6)
7836 status
= gen7_check_mailbox_status(dev_priv
);
7838 status
= gen6_check_mailbox_status(dev_priv
);
7841 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed: %d\n",
7849 int sandybridge_pcode_write(struct drm_i915_private
*dev_priv
,
7854 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
7856 /* GEN6_PCODE_* are outside of the forcewake domain, we can
7857 * use te fw I915_READ variants to reduce the amount of work
7858 * required when reading/writing.
7861 if (I915_READ_FW(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) {
7862 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
7866 I915_WRITE_FW(GEN6_PCODE_DATA
, val
);
7867 I915_WRITE_FW(GEN6_PCODE_DATA1
, 0);
7868 I915_WRITE_FW(GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
| mbox
);
7870 if (intel_wait_for_register_fw(dev_priv
,
7871 GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
, 0,
7873 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox
);
7877 I915_WRITE_FW(GEN6_PCODE_DATA
, 0);
7879 if (INTEL_GEN(dev_priv
) > 6)
7880 status
= gen7_check_mailbox_status(dev_priv
);
7882 status
= gen6_check_mailbox_status(dev_priv
);
7885 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed: %d\n",
7893 static bool skl_pcode_try_request(struct drm_i915_private
*dev_priv
, u32 mbox
,
7894 u32 request
, u32 reply_mask
, u32 reply
,
7899 *status
= sandybridge_pcode_read(dev_priv
, mbox
, &val
);
7901 return *status
|| ((val
& reply_mask
) == reply
);
7905 * skl_pcode_request - send PCODE request until acknowledgment
7906 * @dev_priv: device private
7907 * @mbox: PCODE mailbox ID the request is targeted for
7908 * @request: request ID
7909 * @reply_mask: mask used to check for request acknowledgment
7910 * @reply: value used to check for request acknowledgment
7911 * @timeout_base_ms: timeout for polling with preemption enabled
7913 * Keep resending the @request to @mbox until PCODE acknowledges it, PCODE
7914 * reports an error or an overall timeout of @timeout_base_ms+10 ms expires.
7915 * The request is acknowledged once the PCODE reply dword equals @reply after
7916 * applying @reply_mask. Polling is first attempted with preemption enabled
7917 * for @timeout_base_ms and if this times out for another 10 ms with
7918 * preemption disabled.
7920 * Returns 0 on success, %-ETIMEDOUT in case of a timeout, <0 in case of some
7921 * other error as reported by PCODE.
7923 int skl_pcode_request(struct drm_i915_private
*dev_priv
, u32 mbox
, u32 request
,
7924 u32 reply_mask
, u32 reply
, int timeout_base_ms
)
7929 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
7931 #define COND skl_pcode_try_request(dev_priv, mbox, request, reply_mask, reply, \
7935 * Prime the PCODE by doing a request first. Normally it guarantees
7936 * that a subsequent request, at most @timeout_base_ms later, succeeds.
7937 * _wait_for() doesn't guarantee when its passed condition is evaluated
7938 * first, so send the first request explicitly.
7944 ret
= _wait_for(COND
, timeout_base_ms
* 1000, 10);
7949 * The above can time out if the number of requests was low (2 in the
7950 * worst case) _and_ PCODE was busy for some reason even after a
7951 * (queued) request and @timeout_base_ms delay. As a workaround retry
7952 * the poll with preemption disabled to maximize the number of
7953 * requests. Increase the timeout from @timeout_base_ms to 10ms to
7954 * account for interrupts that could reduce the number of these
7957 DRM_DEBUG_KMS("PCODE timeout, retrying with preemption disabled\n");
7958 WARN_ON_ONCE(timeout_base_ms
> 3);
7960 ret
= wait_for_atomic(COND
, 10);
7964 return ret
? ret
: status
;
7968 static int byt_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
7972 * Slow = Fast = GPLL ref * N
7974 return DIV_ROUND_CLOSEST(dev_priv
->rps
.gpll_ref_freq
* (val
- 0xb7), 1000);
7977 static int byt_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
7979 return DIV_ROUND_CLOSEST(1000 * val
, dev_priv
->rps
.gpll_ref_freq
) + 0xb7;
7982 static int chv_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
7986 * CU (slow) = CU2x (fast) / 2 = GPLL ref * N / 2
7988 return DIV_ROUND_CLOSEST(dev_priv
->rps
.gpll_ref_freq
* val
, 2 * 2 * 1000);
7991 static int chv_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
7993 /* CHV needs even values */
7994 return DIV_ROUND_CLOSEST(2 * 1000 * val
, dev_priv
->rps
.gpll_ref_freq
) * 2;
7997 int intel_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
7999 if (IS_GEN9(dev_priv
))
8000 return DIV_ROUND_CLOSEST(val
* GT_FREQUENCY_MULTIPLIER
,
8002 else if (IS_CHERRYVIEW(dev_priv
))
8003 return chv_gpu_freq(dev_priv
, val
);
8004 else if (IS_VALLEYVIEW(dev_priv
))
8005 return byt_gpu_freq(dev_priv
, val
);
8007 return val
* GT_FREQUENCY_MULTIPLIER
;
8010 int intel_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
8012 if (IS_GEN9(dev_priv
))
8013 return DIV_ROUND_CLOSEST(val
* GEN9_FREQ_SCALER
,
8014 GT_FREQUENCY_MULTIPLIER
);
8015 else if (IS_CHERRYVIEW(dev_priv
))
8016 return chv_freq_opcode(dev_priv
, val
);
8017 else if (IS_VALLEYVIEW(dev_priv
))
8018 return byt_freq_opcode(dev_priv
, val
);
8020 return DIV_ROUND_CLOSEST(val
, GT_FREQUENCY_MULTIPLIER
);
8023 struct request_boost
{
8024 struct work_struct work
;
8025 struct drm_i915_gem_request
*req
;
8028 static void __intel_rps_boost_work(struct work_struct
*work
)
8030 struct request_boost
*boost
= container_of(work
, struct request_boost
, work
);
8031 struct drm_i915_gem_request
*req
= boost
->req
;
8033 if (!i915_gem_request_completed(req
))
8034 gen6_rps_boost(req
->i915
, NULL
, req
->emitted_jiffies
);
8036 i915_gem_request_put(req
);
8040 void intel_queue_rps_boost_for_request(struct drm_i915_gem_request
*req
)
8042 struct request_boost
*boost
;
8044 if (req
== NULL
|| INTEL_GEN(req
->i915
) < 6)
8047 if (i915_gem_request_completed(req
))
8050 boost
= kmalloc(sizeof(*boost
), GFP_ATOMIC
);
8054 boost
->req
= i915_gem_request_get(req
);
8056 INIT_WORK(&boost
->work
, __intel_rps_boost_work
);
8057 queue_work(req
->i915
->wq
, &boost
->work
);
8060 void intel_pm_setup(struct drm_i915_private
*dev_priv
)
8062 mutex_init(&dev_priv
->rps
.hw_lock
);
8063 spin_lock_init(&dev_priv
->rps
.client_lock
);
8065 INIT_DELAYED_WORK(&dev_priv
->rps
.autoenable_work
,
8066 __intel_autoenable_gt_powersave
);
8067 INIT_LIST_HEAD(&dev_priv
->rps
.clients
);
8069 dev_priv
->pm
.suspended
= false;
8070 atomic_set(&dev_priv
->pm
.wakeref_count
, 0);