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,cfl */
62 I915_WRITE(CHICKEN_PAR1_1
,
63 I915_READ(CHICKEN_PAR1_1
) | SKL_EDP_PSR_FIX_RDWRAP
);
66 * Display WA#0390: skl,bxt,kbl,glk
68 * Must match Sampler, Pixel Back End, and Media
69 * (0xE194 bit 8, 0x7014 bit 13, 0x4DDC bits 27 and 31).
71 * Including bits outside the page in the hash would
72 * require 2 (or 4?) MiB alignment of resources. Just
73 * assume the defaul hashing mode which only uses bits
76 I915_WRITE(CHICKEN_PAR1_1
,
77 I915_READ(CHICKEN_PAR1_1
) & ~SKL_RC_HASH_OUTSIDE
);
79 I915_WRITE(GEN8_CONFIG0
,
80 I915_READ(GEN8_CONFIG0
) | GEN9_DEFAULT_FIXES
);
82 /* WaEnableChickenDCPR:skl,bxt,kbl,glk,cfl */
83 I915_WRITE(GEN8_CHICKEN_DCPR_1
,
84 I915_READ(GEN8_CHICKEN_DCPR_1
) | MASK_WAKEMEM
);
86 /* WaFbcTurnOffFbcWatermark:skl,bxt,kbl,cfl */
87 /* WaFbcWakeMemOn:skl,bxt,kbl,glk,cfl */
88 I915_WRITE(DISP_ARB_CTL
, I915_READ(DISP_ARB_CTL
) |
90 DISP_FBC_MEMORY_WAKE
);
92 /* WaFbcHighMemBwCorruptionAvoidance:skl,bxt,kbl,cfl */
93 I915_WRITE(ILK_DPFC_CHICKEN
, I915_READ(ILK_DPFC_CHICKEN
) |
94 ILK_DPFC_DISABLE_DUMMY0
);
96 if (IS_SKYLAKE(dev_priv
)) {
97 /* WaDisableDopClockGating */
98 I915_WRITE(GEN7_MISCCPCTL
, I915_READ(GEN7_MISCCPCTL
)
99 & ~GEN7_DOP_CLOCK_GATE_ENABLE
);
103 static void bxt_init_clock_gating(struct drm_i915_private
*dev_priv
)
105 gen9_init_clock_gating(dev_priv
);
107 /* WaDisableSDEUnitClockGating:bxt */
108 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
109 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
113 * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
115 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
116 GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ
);
119 * Wa: Backlight PWM may stop in the asserted state, causing backlight
122 I915_WRITE(GEN9_CLKGATE_DIS_0
, I915_READ(GEN9_CLKGATE_DIS_0
) |
123 PWM1_GATING_DIS
| PWM2_GATING_DIS
);
126 static void glk_init_clock_gating(struct drm_i915_private
*dev_priv
)
128 gen9_init_clock_gating(dev_priv
);
131 * WaDisablePWMClockGating:glk
132 * Backlight PWM may stop in the asserted state, causing backlight
135 I915_WRITE(GEN9_CLKGATE_DIS_0
, I915_READ(GEN9_CLKGATE_DIS_0
) |
136 PWM1_GATING_DIS
| PWM2_GATING_DIS
);
138 /* WaDDIIOTimeout:glk */
139 if (IS_GLK_REVID(dev_priv
, 0, GLK_REVID_A1
)) {
140 u32 val
= I915_READ(CHICKEN_MISC_2
);
141 val
&= ~(GLK_CL0_PWR_DOWN
|
144 I915_WRITE(CHICKEN_MISC_2
, val
);
149 static void i915_pineview_get_mem_freq(struct drm_i915_private
*dev_priv
)
153 tmp
= I915_READ(CLKCFG
);
155 switch (tmp
& CLKCFG_FSB_MASK
) {
157 dev_priv
->fsb_freq
= 533; /* 133*4 */
160 dev_priv
->fsb_freq
= 800; /* 200*4 */
163 dev_priv
->fsb_freq
= 667; /* 167*4 */
166 dev_priv
->fsb_freq
= 400; /* 100*4 */
170 switch (tmp
& CLKCFG_MEM_MASK
) {
172 dev_priv
->mem_freq
= 533;
175 dev_priv
->mem_freq
= 667;
178 dev_priv
->mem_freq
= 800;
182 /* detect pineview DDR3 setting */
183 tmp
= I915_READ(CSHRDDR3CTL
);
184 dev_priv
->is_ddr3
= (tmp
& CSHRDDR3CTL_DDR3
) ? 1 : 0;
187 static void i915_ironlake_get_mem_freq(struct drm_i915_private
*dev_priv
)
191 ddrpll
= I915_READ16(DDRMPLL1
);
192 csipll
= I915_READ16(CSIPLL0
);
194 switch (ddrpll
& 0xff) {
196 dev_priv
->mem_freq
= 800;
199 dev_priv
->mem_freq
= 1066;
202 dev_priv
->mem_freq
= 1333;
205 dev_priv
->mem_freq
= 1600;
208 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
210 dev_priv
->mem_freq
= 0;
214 dev_priv
->ips
.r_t
= dev_priv
->mem_freq
;
216 switch (csipll
& 0x3ff) {
218 dev_priv
->fsb_freq
= 3200;
221 dev_priv
->fsb_freq
= 3733;
224 dev_priv
->fsb_freq
= 4266;
227 dev_priv
->fsb_freq
= 4800;
230 dev_priv
->fsb_freq
= 5333;
233 dev_priv
->fsb_freq
= 5866;
236 dev_priv
->fsb_freq
= 6400;
239 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
241 dev_priv
->fsb_freq
= 0;
245 if (dev_priv
->fsb_freq
== 3200) {
246 dev_priv
->ips
.c_m
= 0;
247 } else if (dev_priv
->fsb_freq
> 3200 && dev_priv
->fsb_freq
<= 4800) {
248 dev_priv
->ips
.c_m
= 1;
250 dev_priv
->ips
.c_m
= 2;
254 static const struct cxsr_latency cxsr_latency_table
[] = {
255 {1, 0, 800, 400, 3382, 33382, 3983, 33983}, /* DDR2-400 SC */
256 {1, 0, 800, 667, 3354, 33354, 3807, 33807}, /* DDR2-667 SC */
257 {1, 0, 800, 800, 3347, 33347, 3763, 33763}, /* DDR2-800 SC */
258 {1, 1, 800, 667, 6420, 36420, 6873, 36873}, /* DDR3-667 SC */
259 {1, 1, 800, 800, 5902, 35902, 6318, 36318}, /* DDR3-800 SC */
261 {1, 0, 667, 400, 3400, 33400, 4021, 34021}, /* DDR2-400 SC */
262 {1, 0, 667, 667, 3372, 33372, 3845, 33845}, /* DDR2-667 SC */
263 {1, 0, 667, 800, 3386, 33386, 3822, 33822}, /* DDR2-800 SC */
264 {1, 1, 667, 667, 6438, 36438, 6911, 36911}, /* DDR3-667 SC */
265 {1, 1, 667, 800, 5941, 35941, 6377, 36377}, /* DDR3-800 SC */
267 {1, 0, 400, 400, 3472, 33472, 4173, 34173}, /* DDR2-400 SC */
268 {1, 0, 400, 667, 3443, 33443, 3996, 33996}, /* DDR2-667 SC */
269 {1, 0, 400, 800, 3430, 33430, 3946, 33946}, /* DDR2-800 SC */
270 {1, 1, 400, 667, 6509, 36509, 7062, 37062}, /* DDR3-667 SC */
271 {1, 1, 400, 800, 5985, 35985, 6501, 36501}, /* DDR3-800 SC */
273 {0, 0, 800, 400, 3438, 33438, 4065, 34065}, /* DDR2-400 SC */
274 {0, 0, 800, 667, 3410, 33410, 3889, 33889}, /* DDR2-667 SC */
275 {0, 0, 800, 800, 3403, 33403, 3845, 33845}, /* DDR2-800 SC */
276 {0, 1, 800, 667, 6476, 36476, 6955, 36955}, /* DDR3-667 SC */
277 {0, 1, 800, 800, 5958, 35958, 6400, 36400}, /* DDR3-800 SC */
279 {0, 0, 667, 400, 3456, 33456, 4103, 34106}, /* DDR2-400 SC */
280 {0, 0, 667, 667, 3428, 33428, 3927, 33927}, /* DDR2-667 SC */
281 {0, 0, 667, 800, 3443, 33443, 3905, 33905}, /* DDR2-800 SC */
282 {0, 1, 667, 667, 6494, 36494, 6993, 36993}, /* DDR3-667 SC */
283 {0, 1, 667, 800, 5998, 35998, 6460, 36460}, /* DDR3-800 SC */
285 {0, 0, 400, 400, 3528, 33528, 4255, 34255}, /* DDR2-400 SC */
286 {0, 0, 400, 667, 3500, 33500, 4079, 34079}, /* DDR2-667 SC */
287 {0, 0, 400, 800, 3487, 33487, 4029, 34029}, /* DDR2-800 SC */
288 {0, 1, 400, 667, 6566, 36566, 7145, 37145}, /* DDR3-667 SC */
289 {0, 1, 400, 800, 6042, 36042, 6584, 36584}, /* DDR3-800 SC */
292 static const struct cxsr_latency
*intel_get_cxsr_latency(bool is_desktop
,
297 const struct cxsr_latency
*latency
;
300 if (fsb
== 0 || mem
== 0)
303 for (i
= 0; i
< ARRAY_SIZE(cxsr_latency_table
); i
++) {
304 latency
= &cxsr_latency_table
[i
];
305 if (is_desktop
== latency
->is_desktop
&&
306 is_ddr3
== latency
->is_ddr3
&&
307 fsb
== latency
->fsb_freq
&& mem
== latency
->mem_freq
)
311 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
316 static void chv_set_memory_dvfs(struct drm_i915_private
*dev_priv
, bool enable
)
320 mutex_lock(&dev_priv
->rps
.hw_lock
);
322 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
);
324 val
&= ~FORCE_DDR_HIGH_FREQ
;
326 val
|= FORCE_DDR_HIGH_FREQ
;
327 val
&= ~FORCE_DDR_LOW_FREQ
;
328 val
|= FORCE_DDR_FREQ_REQ_ACK
;
329 vlv_punit_write(dev_priv
, PUNIT_REG_DDR_SETUP2
, val
);
331 if (wait_for((vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
) &
332 FORCE_DDR_FREQ_REQ_ACK
) == 0, 3))
333 DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
335 mutex_unlock(&dev_priv
->rps
.hw_lock
);
338 static void chv_set_memory_pm5(struct drm_i915_private
*dev_priv
, bool enable
)
342 mutex_lock(&dev_priv
->rps
.hw_lock
);
344 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DSPFREQ
);
346 val
|= DSP_MAXFIFO_PM5_ENABLE
;
348 val
&= ~DSP_MAXFIFO_PM5_ENABLE
;
349 vlv_punit_write(dev_priv
, PUNIT_REG_DSPFREQ
, val
);
351 mutex_unlock(&dev_priv
->rps
.hw_lock
);
354 #define FW_WM(value, plane) \
355 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
357 static bool _intel_set_memory_cxsr(struct drm_i915_private
*dev_priv
, bool enable
)
362 if (IS_VALLEYVIEW(dev_priv
) || IS_CHERRYVIEW(dev_priv
)) {
363 was_enabled
= I915_READ(FW_BLC_SELF_VLV
) & FW_CSPWRDWNEN
;
364 I915_WRITE(FW_BLC_SELF_VLV
, enable
? FW_CSPWRDWNEN
: 0);
365 POSTING_READ(FW_BLC_SELF_VLV
);
366 } else if (IS_G4X(dev_priv
) || IS_I965GM(dev_priv
)) {
367 was_enabled
= I915_READ(FW_BLC_SELF
) & FW_BLC_SELF_EN
;
368 I915_WRITE(FW_BLC_SELF
, enable
? FW_BLC_SELF_EN
: 0);
369 POSTING_READ(FW_BLC_SELF
);
370 } else if (IS_PINEVIEW(dev_priv
)) {
371 val
= I915_READ(DSPFW3
);
372 was_enabled
= val
& PINEVIEW_SELF_REFRESH_EN
;
374 val
|= PINEVIEW_SELF_REFRESH_EN
;
376 val
&= ~PINEVIEW_SELF_REFRESH_EN
;
377 I915_WRITE(DSPFW3
, val
);
378 POSTING_READ(DSPFW3
);
379 } else if (IS_I945G(dev_priv
) || IS_I945GM(dev_priv
)) {
380 was_enabled
= I915_READ(FW_BLC_SELF
) & FW_BLC_SELF_EN
;
381 val
= enable
? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN
) :
382 _MASKED_BIT_DISABLE(FW_BLC_SELF_EN
);
383 I915_WRITE(FW_BLC_SELF
, val
);
384 POSTING_READ(FW_BLC_SELF
);
385 } else if (IS_I915GM(dev_priv
)) {
387 * FIXME can't find a bit like this for 915G, and
388 * and yet it does have the related watermark in
389 * FW_BLC_SELF. What's going on?
391 was_enabled
= I915_READ(INSTPM
) & INSTPM_SELF_EN
;
392 val
= enable
? _MASKED_BIT_ENABLE(INSTPM_SELF_EN
) :
393 _MASKED_BIT_DISABLE(INSTPM_SELF_EN
);
394 I915_WRITE(INSTPM
, val
);
395 POSTING_READ(INSTPM
);
400 trace_intel_memory_cxsr(dev_priv
, was_enabled
, enable
);
402 DRM_DEBUG_KMS("memory self-refresh is %s (was %s)\n",
403 enableddisabled(enable
),
404 enableddisabled(was_enabled
));
410 * intel_set_memory_cxsr - Configure CxSR state
411 * @dev_priv: i915 device
412 * @enable: Allow vs. disallow CxSR
414 * Allow or disallow the system to enter a special CxSR
415 * (C-state self refresh) state. What typically happens in CxSR mode
416 * is that several display FIFOs may get combined into a single larger
417 * FIFO for a particular plane (so called max FIFO mode) to allow the
418 * system to defer memory fetches longer, and the memory will enter
421 * Note that enabling CxSR does not guarantee that the system enter
422 * this special mode, nor does it guarantee that the system stays
423 * in that mode once entered. So this just allows/disallows the system
424 * to autonomously utilize the CxSR mode. Other factors such as core
425 * C-states will affect when/if the system actually enters/exits the
428 * Note that on VLV/CHV this actually only controls the max FIFO mode,
429 * and the system is free to enter/exit memory self refresh at any time
430 * even when the use of CxSR has been disallowed.
432 * While the system is actually in the CxSR/max FIFO mode, some plane
433 * control registers will not get latched on vblank. Thus in order to
434 * guarantee the system will respond to changes in the plane registers
435 * we must always disallow CxSR prior to making changes to those registers.
436 * Unfortunately the system will re-evaluate the CxSR conditions at
437 * frame start which happens after vblank start (which is when the plane
438 * registers would get latched), so we can't proceed with the plane update
439 * during the same frame where we disallowed CxSR.
441 * Certain platforms also have a deeper HPLL SR mode. Fortunately the
442 * HPLL SR mode depends on CxSR itself, so we don't have to hand hold
443 * the hardware w.r.t. HPLL SR when writing to plane registers.
444 * Disallowing just CxSR is sufficient.
446 bool intel_set_memory_cxsr(struct drm_i915_private
*dev_priv
, bool enable
)
450 mutex_lock(&dev_priv
->wm
.wm_mutex
);
451 ret
= _intel_set_memory_cxsr(dev_priv
, enable
);
452 if (IS_VALLEYVIEW(dev_priv
) || IS_CHERRYVIEW(dev_priv
))
453 dev_priv
->wm
.vlv
.cxsr
= enable
;
454 else if (IS_G4X(dev_priv
))
455 dev_priv
->wm
.g4x
.cxsr
= enable
;
456 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
462 * Latency for FIFO fetches is dependent on several factors:
463 * - memory configuration (speed, channels)
465 * - current MCH state
466 * It can be fairly high in some situations, so here we assume a fairly
467 * pessimal value. It's a tradeoff between extra memory fetches (if we
468 * set this value too high, the FIFO will fetch frequently to stay full)
469 * and power consumption (set it too low to save power and we might see
470 * FIFO underruns and display "flicker").
472 * A value of 5us seems to be a good balance; safe for very low end
473 * platforms but not overly aggressive on lower latency configs.
475 static const int pessimal_latency_ns
= 5000;
477 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
478 ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
480 static void vlv_get_fifo_size(struct intel_crtc_state
*crtc_state
)
482 struct intel_crtc
*crtc
= to_intel_crtc(crtc_state
->base
.crtc
);
483 struct drm_i915_private
*dev_priv
= to_i915(crtc
->base
.dev
);
484 struct vlv_fifo_state
*fifo_state
= &crtc_state
->wm
.vlv
.fifo_state
;
485 enum pipe pipe
= crtc
->pipe
;
486 int sprite0_start
, sprite1_start
;
489 uint32_t dsparb
, dsparb2
, dsparb3
;
491 dsparb
= I915_READ(DSPARB
);
492 dsparb2
= I915_READ(DSPARB2
);
493 sprite0_start
= VLV_FIFO_START(dsparb
, dsparb2
, 0, 0);
494 sprite1_start
= VLV_FIFO_START(dsparb
, dsparb2
, 8, 4);
497 dsparb
= I915_READ(DSPARB
);
498 dsparb2
= I915_READ(DSPARB2
);
499 sprite0_start
= VLV_FIFO_START(dsparb
, dsparb2
, 16, 8);
500 sprite1_start
= VLV_FIFO_START(dsparb
, dsparb2
, 24, 12);
503 dsparb2
= I915_READ(DSPARB2
);
504 dsparb3
= I915_READ(DSPARB3
);
505 sprite0_start
= VLV_FIFO_START(dsparb3
, dsparb2
, 0, 16);
506 sprite1_start
= VLV_FIFO_START(dsparb3
, dsparb2
, 8, 20);
513 fifo_state
->plane
[PLANE_PRIMARY
] = sprite0_start
;
514 fifo_state
->plane
[PLANE_SPRITE0
] = sprite1_start
- sprite0_start
;
515 fifo_state
->plane
[PLANE_SPRITE1
] = 511 - sprite1_start
;
516 fifo_state
->plane
[PLANE_CURSOR
] = 63;
519 static int i9xx_get_fifo_size(struct drm_i915_private
*dev_priv
, int plane
)
521 uint32_t dsparb
= I915_READ(DSPARB
);
524 size
= dsparb
& 0x7f;
526 size
= ((dsparb
>> DSPARB_CSTART_SHIFT
) & 0x7f) - size
;
528 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb
,
529 plane
? "B" : "A", size
);
534 static int i830_get_fifo_size(struct drm_i915_private
*dev_priv
, int plane
)
536 uint32_t dsparb
= I915_READ(DSPARB
);
539 size
= dsparb
& 0x1ff;
541 size
= ((dsparb
>> DSPARB_BEND_SHIFT
) & 0x1ff) - size
;
542 size
>>= 1; /* Convert to cachelines */
544 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb
,
545 plane
? "B" : "A", size
);
550 static int i845_get_fifo_size(struct drm_i915_private
*dev_priv
, int plane
)
552 uint32_t dsparb
= I915_READ(DSPARB
);
555 size
= dsparb
& 0x7f;
556 size
>>= 2; /* Convert to cachelines */
558 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb
,
565 /* Pineview has different values for various configs */
566 static const struct intel_watermark_params pineview_display_wm
= {
567 .fifo_size
= PINEVIEW_DISPLAY_FIFO
,
568 .max_wm
= PINEVIEW_MAX_WM
,
569 .default_wm
= PINEVIEW_DFT_WM
,
570 .guard_size
= PINEVIEW_GUARD_WM
,
571 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
573 static const struct intel_watermark_params pineview_display_hplloff_wm
= {
574 .fifo_size
= PINEVIEW_DISPLAY_FIFO
,
575 .max_wm
= PINEVIEW_MAX_WM
,
576 .default_wm
= PINEVIEW_DFT_HPLLOFF_WM
,
577 .guard_size
= PINEVIEW_GUARD_WM
,
578 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
580 static const struct intel_watermark_params pineview_cursor_wm
= {
581 .fifo_size
= PINEVIEW_CURSOR_FIFO
,
582 .max_wm
= PINEVIEW_CURSOR_MAX_WM
,
583 .default_wm
= PINEVIEW_CURSOR_DFT_WM
,
584 .guard_size
= PINEVIEW_CURSOR_GUARD_WM
,
585 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
587 static const struct intel_watermark_params pineview_cursor_hplloff_wm
= {
588 .fifo_size
= PINEVIEW_CURSOR_FIFO
,
589 .max_wm
= PINEVIEW_CURSOR_MAX_WM
,
590 .default_wm
= PINEVIEW_CURSOR_DFT_WM
,
591 .guard_size
= PINEVIEW_CURSOR_GUARD_WM
,
592 .cacheline_size
= PINEVIEW_FIFO_LINE_SIZE
,
594 static const struct intel_watermark_params i965_cursor_wm_info
= {
595 .fifo_size
= I965_CURSOR_FIFO
,
596 .max_wm
= I965_CURSOR_MAX_WM
,
597 .default_wm
= I965_CURSOR_DFT_WM
,
599 .cacheline_size
= I915_FIFO_LINE_SIZE
,
601 static const struct intel_watermark_params i945_wm_info
= {
602 .fifo_size
= I945_FIFO_SIZE
,
603 .max_wm
= I915_MAX_WM
,
606 .cacheline_size
= I915_FIFO_LINE_SIZE
,
608 static const struct intel_watermark_params i915_wm_info
= {
609 .fifo_size
= I915_FIFO_SIZE
,
610 .max_wm
= I915_MAX_WM
,
613 .cacheline_size
= I915_FIFO_LINE_SIZE
,
615 static const struct intel_watermark_params i830_a_wm_info
= {
616 .fifo_size
= I855GM_FIFO_SIZE
,
617 .max_wm
= I915_MAX_WM
,
620 .cacheline_size
= I830_FIFO_LINE_SIZE
,
622 static const struct intel_watermark_params i830_bc_wm_info
= {
623 .fifo_size
= I855GM_FIFO_SIZE
,
624 .max_wm
= I915_MAX_WM
/2,
627 .cacheline_size
= I830_FIFO_LINE_SIZE
,
629 static const struct intel_watermark_params i845_wm_info
= {
630 .fifo_size
= I830_FIFO_SIZE
,
631 .max_wm
= I915_MAX_WM
,
634 .cacheline_size
= I830_FIFO_LINE_SIZE
,
638 * intel_wm_method1 - Method 1 / "small buffer" watermark formula
639 * @pixel_rate: Pipe pixel rate in kHz
640 * @cpp: Plane bytes per pixel
641 * @latency: Memory wakeup latency in 0.1us units
643 * Compute the watermark using the method 1 or "small buffer"
644 * formula. The caller may additonally add extra cachelines
645 * to account for TLB misses and clock crossings.
647 * This method is concerned with the short term drain rate
648 * of the FIFO, ie. it does not account for blanking periods
649 * which would effectively reduce the average drain rate across
650 * a longer period. The name "small" refers to the fact the
651 * FIFO is relatively small compared to the amount of data
654 * The FIFO level vs. time graph might look something like:
658 * __---__---__ (- plane active, _ blanking)
661 * or perhaps like this:
664 * __----__----__ (- plane active, _ blanking)
668 * The watermark in bytes
670 static unsigned int intel_wm_method1(unsigned int pixel_rate
,
672 unsigned int latency
)
676 ret
= (uint64_t) pixel_rate
* cpp
* latency
;
677 ret
= DIV_ROUND_UP_ULL(ret
, 10000);
683 * intel_wm_method2 - Method 2 / "large buffer" watermark formula
684 * @pixel_rate: Pipe pixel rate in kHz
685 * @htotal: Pipe horizontal total
686 * @width: Plane width in pixels
687 * @cpp: Plane bytes per pixel
688 * @latency: Memory wakeup latency in 0.1us units
690 * Compute the watermark using the method 2 or "large buffer"
691 * formula. The caller may additonally add extra cachelines
692 * to account for TLB misses and clock crossings.
694 * This method is concerned with the long term drain rate
695 * of the FIFO, ie. it does account for blanking periods
696 * which effectively reduce the average drain rate across
697 * a longer period. The name "large" refers to the fact the
698 * FIFO is relatively large compared to the amount of data
701 * The FIFO level vs. time graph might look something like:
706 * __ --__--__--__--__--__--__ (- plane active, _ blanking)
710 * The watermark in bytes
712 static unsigned int intel_wm_method2(unsigned int pixel_rate
,
716 unsigned int latency
)
721 * FIXME remove once all users are computing
722 * watermarks in the correct place.
724 if (WARN_ON_ONCE(htotal
== 0))
727 ret
= (latency
* pixel_rate
) / (htotal
* 10000);
728 ret
= (ret
+ 1) * width
* cpp
;
734 * intel_calculate_wm - calculate watermark level
735 * @pixel_rate: pixel clock
736 * @wm: chip FIFO params
737 * @cpp: bytes per pixel
738 * @latency_ns: memory latency for the platform
740 * Calculate the watermark level (the level at which the display plane will
741 * start fetching from memory again). Each chip has a different display
742 * FIFO size and allocation, so the caller needs to figure that out and pass
743 * in the correct intel_watermark_params structure.
745 * As the pixel clock runs, the FIFO will be drained at a rate that depends
746 * on the pixel size. When it reaches the watermark level, it'll start
747 * fetching FIFO line sized based chunks from memory until the FIFO fills
748 * past the watermark point. If the FIFO drains completely, a FIFO underrun
749 * will occur, and a display engine hang could result.
751 static unsigned int intel_calculate_wm(int pixel_rate
,
752 const struct intel_watermark_params
*wm
,
753 int fifo_size
, int cpp
,
754 unsigned int latency_ns
)
756 int entries
, wm_size
;
759 * Note: we need to make sure we don't overflow for various clock &
761 * clocks go from a few thousand to several hundred thousand.
762 * latency is usually a few thousand
764 entries
= intel_wm_method1(pixel_rate
, cpp
,
766 entries
= DIV_ROUND_UP(entries
, wm
->cacheline_size
) +
768 DRM_DEBUG_KMS("FIFO entries required for mode: %d\n", entries
);
770 wm_size
= fifo_size
- entries
;
771 DRM_DEBUG_KMS("FIFO watermark level: %d\n", wm_size
);
773 /* Don't promote wm_size to unsigned... */
774 if (wm_size
> wm
->max_wm
)
775 wm_size
= wm
->max_wm
;
777 wm_size
= wm
->default_wm
;
780 * Bspec seems to indicate that the value shouldn't be lower than
781 * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
782 * Lets go for 8 which is the burst size since certain platforms
783 * already use a hardcoded 8 (which is what the spec says should be
792 static bool is_disabling(int old
, int new, int threshold
)
794 return old
>= threshold
&& new < threshold
;
797 static bool is_enabling(int old
, int new, int threshold
)
799 return old
< threshold
&& new >= threshold
;
802 static int intel_wm_num_levels(struct drm_i915_private
*dev_priv
)
804 return dev_priv
->wm
.max_level
+ 1;
807 static bool intel_wm_plane_visible(const struct intel_crtc_state
*crtc_state
,
808 const struct intel_plane_state
*plane_state
)
810 struct intel_plane
*plane
= to_intel_plane(plane_state
->base
.plane
);
812 /* FIXME check the 'enable' instead */
813 if (!crtc_state
->base
.active
)
817 * Treat cursor with fb as always visible since cursor updates
818 * can happen faster than the vrefresh rate, and the current
819 * watermark code doesn't handle that correctly. Cursor updates
820 * which set/clear the fb or change the cursor size are going
821 * to get throttled by intel_legacy_cursor_update() to work
822 * around this problem with the watermark code.
824 if (plane
->id
== PLANE_CURSOR
)
825 return plane_state
->base
.fb
!= NULL
;
827 return plane_state
->base
.visible
;
830 static struct intel_crtc
*single_enabled_crtc(struct drm_i915_private
*dev_priv
)
832 struct intel_crtc
*crtc
, *enabled
= NULL
;
834 for_each_intel_crtc(&dev_priv
->drm
, crtc
) {
835 if (intel_crtc_active(crtc
)) {
845 static void pineview_update_wm(struct intel_crtc
*unused_crtc
)
847 struct drm_i915_private
*dev_priv
= to_i915(unused_crtc
->base
.dev
);
848 struct intel_crtc
*crtc
;
849 const struct cxsr_latency
*latency
;
853 latency
= intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv
),
858 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
859 intel_set_memory_cxsr(dev_priv
, false);
863 crtc
= single_enabled_crtc(dev_priv
);
865 const struct drm_display_mode
*adjusted_mode
=
866 &crtc
->config
->base
.adjusted_mode
;
867 const struct drm_framebuffer
*fb
=
868 crtc
->base
.primary
->state
->fb
;
869 int cpp
= fb
->format
->cpp
[0];
870 int clock
= adjusted_mode
->crtc_clock
;
873 wm
= intel_calculate_wm(clock
, &pineview_display_wm
,
874 pineview_display_wm
.fifo_size
,
875 cpp
, latency
->display_sr
);
876 reg
= I915_READ(DSPFW1
);
877 reg
&= ~DSPFW_SR_MASK
;
878 reg
|= FW_WM(wm
, SR
);
879 I915_WRITE(DSPFW1
, reg
);
880 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg
);
883 wm
= intel_calculate_wm(clock
, &pineview_cursor_wm
,
884 pineview_display_wm
.fifo_size
,
885 4, latency
->cursor_sr
);
886 reg
= I915_READ(DSPFW3
);
887 reg
&= ~DSPFW_CURSOR_SR_MASK
;
888 reg
|= FW_WM(wm
, CURSOR_SR
);
889 I915_WRITE(DSPFW3
, reg
);
891 /* Display HPLL off SR */
892 wm
= intel_calculate_wm(clock
, &pineview_display_hplloff_wm
,
893 pineview_display_hplloff_wm
.fifo_size
,
894 cpp
, latency
->display_hpll_disable
);
895 reg
= I915_READ(DSPFW3
);
896 reg
&= ~DSPFW_HPLL_SR_MASK
;
897 reg
|= FW_WM(wm
, HPLL_SR
);
898 I915_WRITE(DSPFW3
, reg
);
900 /* cursor HPLL off SR */
901 wm
= intel_calculate_wm(clock
, &pineview_cursor_hplloff_wm
,
902 pineview_display_hplloff_wm
.fifo_size
,
903 4, latency
->cursor_hpll_disable
);
904 reg
= I915_READ(DSPFW3
);
905 reg
&= ~DSPFW_HPLL_CURSOR_MASK
;
906 reg
|= FW_WM(wm
, HPLL_CURSOR
);
907 I915_WRITE(DSPFW3
, reg
);
908 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg
);
910 intel_set_memory_cxsr(dev_priv
, true);
912 intel_set_memory_cxsr(dev_priv
, false);
917 * Documentation says:
918 * "If the line size is small, the TLB fetches can get in the way of the
919 * data fetches, causing some lag in the pixel data return which is not
920 * accounted for in the above formulas. The following adjustment only
921 * needs to be applied if eight whole lines fit in the buffer at once.
922 * The WM is adjusted upwards by the difference between the FIFO size
923 * and the size of 8 whole lines. This adjustment is always performed
924 * in the actual pixel depth regardless of whether FBC is enabled or not."
926 static int g4x_tlb_miss_wa(int fifo_size
, int width
, int cpp
)
928 int tlb_miss
= fifo_size
* 64 - width
* cpp
* 8;
930 return max(0, tlb_miss
);
933 static void g4x_write_wm_values(struct drm_i915_private
*dev_priv
,
934 const struct g4x_wm_values
*wm
)
938 for_each_pipe(dev_priv
, pipe
)
939 trace_g4x_wm(intel_get_crtc_for_pipe(dev_priv
, pipe
), wm
);
942 FW_WM(wm
->sr
.plane
, SR
) |
943 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_CURSOR
], CURSORB
) |
944 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
], PLANEB
) |
945 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
], PLANEA
));
947 (wm
->fbc_en
? DSPFW_FBC_SR_EN
: 0) |
948 FW_WM(wm
->sr
.fbc
, FBC_SR
) |
949 FW_WM(wm
->hpll
.fbc
, FBC_HPLL_SR
) |
950 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
], SPRITEB
) |
951 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_CURSOR
], CURSORA
) |
952 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
], SPRITEA
));
954 (wm
->hpll_en
? DSPFW_HPLL_SR_EN
: 0) |
955 FW_WM(wm
->sr
.cursor
, CURSOR_SR
) |
956 FW_WM(wm
->hpll
.cursor
, HPLL_CURSOR
) |
957 FW_WM(wm
->hpll
.plane
, HPLL_SR
));
959 POSTING_READ(DSPFW1
);
962 #define FW_WM_VLV(value, plane) \
963 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
965 static void vlv_write_wm_values(struct drm_i915_private
*dev_priv
,
966 const struct vlv_wm_values
*wm
)
970 for_each_pipe(dev_priv
, pipe
) {
971 trace_vlv_wm(intel_get_crtc_for_pipe(dev_priv
, pipe
), wm
);
973 I915_WRITE(VLV_DDL(pipe
),
974 (wm
->ddl
[pipe
].plane
[PLANE_CURSOR
] << DDL_CURSOR_SHIFT
) |
975 (wm
->ddl
[pipe
].plane
[PLANE_SPRITE1
] << DDL_SPRITE_SHIFT(1)) |
976 (wm
->ddl
[pipe
].plane
[PLANE_SPRITE0
] << DDL_SPRITE_SHIFT(0)) |
977 (wm
->ddl
[pipe
].plane
[PLANE_PRIMARY
] << DDL_PLANE_SHIFT
));
981 * Zero the (unused) WM1 watermarks, and also clear all the
982 * high order bits so that there are no out of bounds values
983 * present in the registers during the reprogramming.
985 I915_WRITE(DSPHOWM
, 0);
986 I915_WRITE(DSPHOWM1
, 0);
987 I915_WRITE(DSPFW4
, 0);
988 I915_WRITE(DSPFW5
, 0);
989 I915_WRITE(DSPFW6
, 0);
992 FW_WM(wm
->sr
.plane
, SR
) |
993 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_CURSOR
], CURSORB
) |
994 FW_WM_VLV(wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
], PLANEB
) |
995 FW_WM_VLV(wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
], PLANEA
));
997 FW_WM_VLV(wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE1
], SPRITEB
) |
998 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_CURSOR
], CURSORA
) |
999 FW_WM_VLV(wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
], SPRITEA
));
1001 FW_WM(wm
->sr
.cursor
, CURSOR_SR
));
1003 if (IS_CHERRYVIEW(dev_priv
)) {
1004 I915_WRITE(DSPFW7_CHV
,
1005 FW_WM_VLV(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
], SPRITED
) |
1006 FW_WM_VLV(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
], SPRITEC
));
1007 I915_WRITE(DSPFW8_CHV
,
1008 FW_WM_VLV(wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE1
], SPRITEF
) |
1009 FW_WM_VLV(wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE0
], SPRITEE
));
1010 I915_WRITE(DSPFW9_CHV
,
1011 FW_WM_VLV(wm
->pipe
[PIPE_C
].plane
[PLANE_PRIMARY
], PLANEC
) |
1012 FW_WM(wm
->pipe
[PIPE_C
].plane
[PLANE_CURSOR
], CURSORC
));
1014 FW_WM(wm
->sr
.plane
>> 9, SR_HI
) |
1015 FW_WM(wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE1
] >> 8, SPRITEF_HI
) |
1016 FW_WM(wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE0
] >> 8, SPRITEE_HI
) |
1017 FW_WM(wm
->pipe
[PIPE_C
].plane
[PLANE_PRIMARY
] >> 8, PLANEC_HI
) |
1018 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
] >> 8, SPRITED_HI
) |
1019 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
] >> 8, SPRITEC_HI
) |
1020 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
] >> 8, PLANEB_HI
) |
1021 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE1
] >> 8, SPRITEB_HI
) |
1022 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
] >> 8, SPRITEA_HI
) |
1023 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
] >> 8, PLANEA_HI
));
1026 FW_WM_VLV(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
], SPRITED
) |
1027 FW_WM_VLV(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
], SPRITEC
));
1029 FW_WM(wm
->sr
.plane
>> 9, SR_HI
) |
1030 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
] >> 8, SPRITED_HI
) |
1031 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
] >> 8, SPRITEC_HI
) |
1032 FW_WM(wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
] >> 8, PLANEB_HI
) |
1033 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE1
] >> 8, SPRITEB_HI
) |
1034 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
] >> 8, SPRITEA_HI
) |
1035 FW_WM(wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
] >> 8, PLANEA_HI
));
1038 POSTING_READ(DSPFW1
);
1043 static void g4x_setup_wm_latency(struct drm_i915_private
*dev_priv
)
1045 /* all latencies in usec */
1046 dev_priv
->wm
.pri_latency
[G4X_WM_LEVEL_NORMAL
] = 5;
1047 dev_priv
->wm
.pri_latency
[G4X_WM_LEVEL_SR
] = 12;
1048 dev_priv
->wm
.pri_latency
[G4X_WM_LEVEL_HPLL
] = 35;
1050 dev_priv
->wm
.max_level
= G4X_WM_LEVEL_HPLL
;
1053 static int g4x_plane_fifo_size(enum plane_id plane_id
, int level
)
1056 * DSPCNTR[13] supposedly controls whether the
1057 * primary plane can use the FIFO space otherwise
1058 * reserved for the sprite plane. It's not 100% clear
1059 * what the actual FIFO size is, but it looks like we
1060 * can happily set both primary and sprite watermarks
1061 * up to 127 cachelines. So that would seem to mean
1062 * that either DSPCNTR[13] doesn't do anything, or that
1063 * the total FIFO is >= 256 cachelines in size. Either
1064 * way, we don't seem to have to worry about this
1065 * repartitioning as the maximum watermark value the
1066 * register can hold for each plane is lower than the
1067 * minimum FIFO size.
1073 return level
== G4X_WM_LEVEL_NORMAL
? 127 : 511;
1075 return level
== G4X_WM_LEVEL_NORMAL
? 127 : 0;
1077 MISSING_CASE(plane_id
);
1082 static int g4x_fbc_fifo_size(int level
)
1085 case G4X_WM_LEVEL_SR
:
1087 case G4X_WM_LEVEL_HPLL
:
1090 MISSING_CASE(level
);
1095 static uint16_t g4x_compute_wm(const struct intel_crtc_state
*crtc_state
,
1096 const struct intel_plane_state
*plane_state
,
1099 struct intel_plane
*plane
= to_intel_plane(plane_state
->base
.plane
);
1100 struct drm_i915_private
*dev_priv
= to_i915(plane
->base
.dev
);
1101 const struct drm_display_mode
*adjusted_mode
=
1102 &crtc_state
->base
.adjusted_mode
;
1103 int clock
, htotal
, cpp
, width
, wm
;
1104 int latency
= dev_priv
->wm
.pri_latency
[level
] * 10;
1109 if (!intel_wm_plane_visible(crtc_state
, plane_state
))
1113 * Not 100% sure which way ELK should go here as the
1114 * spec only says CL/CTG should assume 32bpp and BW
1115 * doesn't need to. But as these things followed the
1116 * mobile vs. desktop lines on gen3 as well, let's
1117 * assume ELK doesn't need this.
1119 * The spec also fails to list such a restriction for
1120 * the HPLL watermark, which seems a little strange.
1121 * Let's use 32bpp for the HPLL watermark as well.
1123 if (IS_GM45(dev_priv
) && plane
->id
== PLANE_PRIMARY
&&
1124 level
!= G4X_WM_LEVEL_NORMAL
)
1127 cpp
= plane_state
->base
.fb
->format
->cpp
[0];
1129 clock
= adjusted_mode
->crtc_clock
;
1130 htotal
= adjusted_mode
->crtc_htotal
;
1132 if (plane
->id
== PLANE_CURSOR
)
1133 width
= plane_state
->base
.crtc_w
;
1135 width
= drm_rect_width(&plane_state
->base
.dst
);
1137 if (plane
->id
== PLANE_CURSOR
) {
1138 wm
= intel_wm_method2(clock
, htotal
, width
, cpp
, latency
);
1139 } else if (plane
->id
== PLANE_PRIMARY
&&
1140 level
== G4X_WM_LEVEL_NORMAL
) {
1141 wm
= intel_wm_method1(clock
, cpp
, latency
);
1145 small
= intel_wm_method1(clock
, cpp
, latency
);
1146 large
= intel_wm_method2(clock
, htotal
, width
, cpp
, latency
);
1148 wm
= min(small
, large
);
1151 wm
+= g4x_tlb_miss_wa(g4x_plane_fifo_size(plane
->id
, level
),
1154 wm
= DIV_ROUND_UP(wm
, 64) + 2;
1156 return min_t(int, wm
, USHRT_MAX
);
1159 static bool g4x_raw_plane_wm_set(struct intel_crtc_state
*crtc_state
,
1160 int level
, enum plane_id plane_id
, u16 value
)
1162 struct drm_i915_private
*dev_priv
= to_i915(crtc_state
->base
.crtc
->dev
);
1165 for (; level
< intel_wm_num_levels(dev_priv
); level
++) {
1166 struct g4x_pipe_wm
*raw
= &crtc_state
->wm
.g4x
.raw
[level
];
1168 dirty
|= raw
->plane
[plane_id
] != value
;
1169 raw
->plane
[plane_id
] = value
;
1175 static bool g4x_raw_fbc_wm_set(struct intel_crtc_state
*crtc_state
,
1176 int level
, u16 value
)
1178 struct drm_i915_private
*dev_priv
= to_i915(crtc_state
->base
.crtc
->dev
);
1181 /* NORMAL level doesn't have an FBC watermark */
1182 level
= max(level
, G4X_WM_LEVEL_SR
);
1184 for (; level
< intel_wm_num_levels(dev_priv
); level
++) {
1185 struct g4x_pipe_wm
*raw
= &crtc_state
->wm
.g4x
.raw
[level
];
1187 dirty
|= raw
->fbc
!= value
;
1194 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state
*cstate
,
1195 const struct intel_plane_state
*pstate
,
1198 static bool g4x_raw_plane_wm_compute(struct intel_crtc_state
*crtc_state
,
1199 const struct intel_plane_state
*plane_state
)
1201 struct intel_plane
*plane
= to_intel_plane(plane_state
->base
.plane
);
1202 int num_levels
= intel_wm_num_levels(to_i915(plane
->base
.dev
));
1203 enum plane_id plane_id
= plane
->id
;
1207 if (!intel_wm_plane_visible(crtc_state
, plane_state
)) {
1208 dirty
|= g4x_raw_plane_wm_set(crtc_state
, 0, plane_id
, 0);
1209 if (plane_id
== PLANE_PRIMARY
)
1210 dirty
|= g4x_raw_fbc_wm_set(crtc_state
, 0, 0);
1214 for (level
= 0; level
< num_levels
; level
++) {
1215 struct g4x_pipe_wm
*raw
= &crtc_state
->wm
.g4x
.raw
[level
];
1218 wm
= g4x_compute_wm(crtc_state
, plane_state
, level
);
1219 max_wm
= g4x_plane_fifo_size(plane_id
, level
);
1224 dirty
|= raw
->plane
[plane_id
] != wm
;
1225 raw
->plane
[plane_id
] = wm
;
1227 if (plane_id
!= PLANE_PRIMARY
||
1228 level
== G4X_WM_LEVEL_NORMAL
)
1231 wm
= ilk_compute_fbc_wm(crtc_state
, plane_state
,
1232 raw
->plane
[plane_id
]);
1233 max_wm
= g4x_fbc_fifo_size(level
);
1236 * FBC wm is not mandatory as we
1237 * can always just disable its use.
1242 dirty
|= raw
->fbc
!= wm
;
1246 /* mark watermarks as invalid */
1247 dirty
|= g4x_raw_plane_wm_set(crtc_state
, level
, plane_id
, USHRT_MAX
);
1249 if (plane_id
== PLANE_PRIMARY
)
1250 dirty
|= g4x_raw_fbc_wm_set(crtc_state
, level
, USHRT_MAX
);
1254 DRM_DEBUG_KMS("%s watermarks: normal=%d, SR=%d, HPLL=%d\n",
1256 crtc_state
->wm
.g4x
.raw
[G4X_WM_LEVEL_NORMAL
].plane
[plane_id
],
1257 crtc_state
->wm
.g4x
.raw
[G4X_WM_LEVEL_SR
].plane
[plane_id
],
1258 crtc_state
->wm
.g4x
.raw
[G4X_WM_LEVEL_HPLL
].plane
[plane_id
]);
1260 if (plane_id
== PLANE_PRIMARY
)
1261 DRM_DEBUG_KMS("FBC watermarks: SR=%d, HPLL=%d\n",
1262 crtc_state
->wm
.g4x
.raw
[G4X_WM_LEVEL_SR
].fbc
,
1263 crtc_state
->wm
.g4x
.raw
[G4X_WM_LEVEL_HPLL
].fbc
);
1269 static bool g4x_raw_plane_wm_is_valid(const struct intel_crtc_state
*crtc_state
,
1270 enum plane_id plane_id
, int level
)
1272 const struct g4x_pipe_wm
*raw
= &crtc_state
->wm
.g4x
.raw
[level
];
1274 return raw
->plane
[plane_id
] <= g4x_plane_fifo_size(plane_id
, level
);
1277 static bool g4x_raw_crtc_wm_is_valid(const struct intel_crtc_state
*crtc_state
,
1280 struct drm_i915_private
*dev_priv
= to_i915(crtc_state
->base
.crtc
->dev
);
1282 if (level
> dev_priv
->wm
.max_level
)
1285 return g4x_raw_plane_wm_is_valid(crtc_state
, PLANE_PRIMARY
, level
) &&
1286 g4x_raw_plane_wm_is_valid(crtc_state
, PLANE_SPRITE0
, level
) &&
1287 g4x_raw_plane_wm_is_valid(crtc_state
, PLANE_CURSOR
, level
);
1290 /* mark all levels starting from 'level' as invalid */
1291 static void g4x_invalidate_wms(struct intel_crtc
*crtc
,
1292 struct g4x_wm_state
*wm_state
, int level
)
1294 if (level
<= G4X_WM_LEVEL_NORMAL
) {
1295 enum plane_id plane_id
;
1297 for_each_plane_id_on_crtc(crtc
, plane_id
)
1298 wm_state
->wm
.plane
[plane_id
] = USHRT_MAX
;
1301 if (level
<= G4X_WM_LEVEL_SR
) {
1302 wm_state
->cxsr
= false;
1303 wm_state
->sr
.cursor
= USHRT_MAX
;
1304 wm_state
->sr
.plane
= USHRT_MAX
;
1305 wm_state
->sr
.fbc
= USHRT_MAX
;
1308 if (level
<= G4X_WM_LEVEL_HPLL
) {
1309 wm_state
->hpll_en
= false;
1310 wm_state
->hpll
.cursor
= USHRT_MAX
;
1311 wm_state
->hpll
.plane
= USHRT_MAX
;
1312 wm_state
->hpll
.fbc
= USHRT_MAX
;
1316 static int g4x_compute_pipe_wm(struct intel_crtc_state
*crtc_state
)
1318 struct intel_crtc
*crtc
= to_intel_crtc(crtc_state
->base
.crtc
);
1319 struct intel_atomic_state
*state
=
1320 to_intel_atomic_state(crtc_state
->base
.state
);
1321 struct g4x_wm_state
*wm_state
= &crtc_state
->wm
.g4x
.optimal
;
1322 int num_active_planes
= hweight32(crtc_state
->active_planes
&
1323 ~BIT(PLANE_CURSOR
));
1324 const struct g4x_pipe_wm
*raw
;
1325 struct intel_plane_state
*plane_state
;
1326 struct intel_plane
*plane
;
1327 enum plane_id plane_id
;
1329 unsigned int dirty
= 0;
1331 for_each_intel_plane_in_state(state
, plane
, plane_state
, i
) {
1332 const struct intel_plane_state
*old_plane_state
=
1333 to_intel_plane_state(plane
->base
.state
);
1335 if (plane_state
->base
.crtc
!= &crtc
->base
&&
1336 old_plane_state
->base
.crtc
!= &crtc
->base
)
1339 if (g4x_raw_plane_wm_compute(crtc_state
, plane_state
))
1340 dirty
|= BIT(plane
->id
);
1346 level
= G4X_WM_LEVEL_NORMAL
;
1347 if (!g4x_raw_crtc_wm_is_valid(crtc_state
, level
))
1350 raw
= &crtc_state
->wm
.g4x
.raw
[level
];
1351 for_each_plane_id_on_crtc(crtc
, plane_id
)
1352 wm_state
->wm
.plane
[plane_id
] = raw
->plane
[plane_id
];
1354 level
= G4X_WM_LEVEL_SR
;
1356 if (!g4x_raw_crtc_wm_is_valid(crtc_state
, level
))
1359 raw
= &crtc_state
->wm
.g4x
.raw
[level
];
1360 wm_state
->sr
.plane
= raw
->plane
[PLANE_PRIMARY
];
1361 wm_state
->sr
.cursor
= raw
->plane
[PLANE_CURSOR
];
1362 wm_state
->sr
.fbc
= raw
->fbc
;
1364 wm_state
->cxsr
= num_active_planes
== BIT(PLANE_PRIMARY
);
1366 level
= G4X_WM_LEVEL_HPLL
;
1368 if (!g4x_raw_crtc_wm_is_valid(crtc_state
, level
))
1371 raw
= &crtc_state
->wm
.g4x
.raw
[level
];
1372 wm_state
->hpll
.plane
= raw
->plane
[PLANE_PRIMARY
];
1373 wm_state
->hpll
.cursor
= raw
->plane
[PLANE_CURSOR
];
1374 wm_state
->hpll
.fbc
= raw
->fbc
;
1376 wm_state
->hpll_en
= wm_state
->cxsr
;
1381 if (level
== G4X_WM_LEVEL_NORMAL
)
1384 /* invalidate the higher levels */
1385 g4x_invalidate_wms(crtc
, wm_state
, level
);
1388 * Determine if the FBC watermark(s) can be used. IF
1389 * this isn't the case we prefer to disable the FBC
1390 ( watermark(s) rather than disable the SR/HPLL
1391 * level(s) entirely.
1393 wm_state
->fbc_en
= level
> G4X_WM_LEVEL_NORMAL
;
1395 if (level
>= G4X_WM_LEVEL_SR
&&
1396 wm_state
->sr
.fbc
> g4x_fbc_fifo_size(G4X_WM_LEVEL_SR
))
1397 wm_state
->fbc_en
= false;
1398 else if (level
>= G4X_WM_LEVEL_HPLL
&&
1399 wm_state
->hpll
.fbc
> g4x_fbc_fifo_size(G4X_WM_LEVEL_HPLL
))
1400 wm_state
->fbc_en
= false;
1405 static int g4x_compute_intermediate_wm(struct drm_device
*dev
,
1406 struct intel_crtc
*crtc
,
1407 struct intel_crtc_state
*crtc_state
)
1409 struct g4x_wm_state
*intermediate
= &crtc_state
->wm
.g4x
.intermediate
;
1410 const struct g4x_wm_state
*optimal
= &crtc_state
->wm
.g4x
.optimal
;
1411 const struct g4x_wm_state
*active
= &crtc
->wm
.active
.g4x
;
1412 enum plane_id plane_id
;
1414 intermediate
->cxsr
= optimal
->cxsr
&& active
->cxsr
&&
1415 !crtc_state
->disable_cxsr
;
1416 intermediate
->hpll_en
= optimal
->hpll_en
&& active
->hpll_en
&&
1417 !crtc_state
->disable_cxsr
;
1418 intermediate
->fbc_en
= optimal
->fbc_en
&& active
->fbc_en
;
1420 for_each_plane_id_on_crtc(crtc
, plane_id
) {
1421 intermediate
->wm
.plane
[plane_id
] =
1422 max(optimal
->wm
.plane
[plane_id
],
1423 active
->wm
.plane
[plane_id
]);
1425 WARN_ON(intermediate
->wm
.plane
[plane_id
] >
1426 g4x_plane_fifo_size(plane_id
, G4X_WM_LEVEL_NORMAL
));
1429 intermediate
->sr
.plane
= max(optimal
->sr
.plane
,
1431 intermediate
->sr
.cursor
= max(optimal
->sr
.cursor
,
1433 intermediate
->sr
.fbc
= max(optimal
->sr
.fbc
,
1436 intermediate
->hpll
.plane
= max(optimal
->hpll
.plane
,
1437 active
->hpll
.plane
);
1438 intermediate
->hpll
.cursor
= max(optimal
->hpll
.cursor
,
1439 active
->hpll
.cursor
);
1440 intermediate
->hpll
.fbc
= max(optimal
->hpll
.fbc
,
1443 WARN_ON((intermediate
->sr
.plane
>
1444 g4x_plane_fifo_size(PLANE_PRIMARY
, G4X_WM_LEVEL_SR
) ||
1445 intermediate
->sr
.cursor
>
1446 g4x_plane_fifo_size(PLANE_CURSOR
, G4X_WM_LEVEL_SR
)) &&
1447 intermediate
->cxsr
);
1448 WARN_ON((intermediate
->sr
.plane
>
1449 g4x_plane_fifo_size(PLANE_PRIMARY
, G4X_WM_LEVEL_HPLL
) ||
1450 intermediate
->sr
.cursor
>
1451 g4x_plane_fifo_size(PLANE_CURSOR
, G4X_WM_LEVEL_HPLL
)) &&
1452 intermediate
->hpll_en
);
1454 WARN_ON(intermediate
->sr
.fbc
> g4x_fbc_fifo_size(1) &&
1455 intermediate
->fbc_en
&& intermediate
->cxsr
);
1456 WARN_ON(intermediate
->hpll
.fbc
> g4x_fbc_fifo_size(2) &&
1457 intermediate
->fbc_en
&& intermediate
->hpll_en
);
1460 * If our intermediate WM are identical to the final WM, then we can
1461 * omit the post-vblank programming; only update if it's different.
1463 if (memcmp(intermediate
, optimal
, sizeof(*intermediate
)) != 0)
1464 crtc_state
->wm
.need_postvbl_update
= true;
1469 static void g4x_merge_wm(struct drm_i915_private
*dev_priv
,
1470 struct g4x_wm_values
*wm
)
1472 struct intel_crtc
*crtc
;
1473 int num_active_crtcs
= 0;
1479 for_each_intel_crtc(&dev_priv
->drm
, crtc
) {
1480 const struct g4x_wm_state
*wm_state
= &crtc
->wm
.active
.g4x
;
1485 if (!wm_state
->cxsr
)
1487 if (!wm_state
->hpll_en
)
1488 wm
->hpll_en
= false;
1489 if (!wm_state
->fbc_en
)
1495 if (num_active_crtcs
!= 1) {
1497 wm
->hpll_en
= false;
1501 for_each_intel_crtc(&dev_priv
->drm
, crtc
) {
1502 const struct g4x_wm_state
*wm_state
= &crtc
->wm
.active
.g4x
;
1503 enum pipe pipe
= crtc
->pipe
;
1505 wm
->pipe
[pipe
] = wm_state
->wm
;
1506 if (crtc
->active
&& wm
->cxsr
)
1507 wm
->sr
= wm_state
->sr
;
1508 if (crtc
->active
&& wm
->hpll_en
)
1509 wm
->hpll
= wm_state
->hpll
;
1513 static void g4x_program_watermarks(struct drm_i915_private
*dev_priv
)
1515 struct g4x_wm_values
*old_wm
= &dev_priv
->wm
.g4x
;
1516 struct g4x_wm_values new_wm
= {};
1518 g4x_merge_wm(dev_priv
, &new_wm
);
1520 if (memcmp(old_wm
, &new_wm
, sizeof(new_wm
)) == 0)
1523 if (is_disabling(old_wm
->cxsr
, new_wm
.cxsr
, true))
1524 _intel_set_memory_cxsr(dev_priv
, false);
1526 g4x_write_wm_values(dev_priv
, &new_wm
);
1528 if (is_enabling(old_wm
->cxsr
, new_wm
.cxsr
, true))
1529 _intel_set_memory_cxsr(dev_priv
, true);
1534 static void g4x_initial_watermarks(struct intel_atomic_state
*state
,
1535 struct intel_crtc_state
*crtc_state
)
1537 struct drm_i915_private
*dev_priv
= to_i915(crtc_state
->base
.crtc
->dev
);
1538 struct intel_crtc
*crtc
= to_intel_crtc(crtc_state
->base
.crtc
);
1540 mutex_lock(&dev_priv
->wm
.wm_mutex
);
1541 crtc
->wm
.active
.g4x
= crtc_state
->wm
.g4x
.intermediate
;
1542 g4x_program_watermarks(dev_priv
);
1543 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
1546 static void g4x_optimize_watermarks(struct intel_atomic_state
*state
,
1547 struct intel_crtc_state
*crtc_state
)
1549 struct drm_i915_private
*dev_priv
= to_i915(crtc_state
->base
.crtc
->dev
);
1550 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc_state
->base
.crtc
);
1552 if (!crtc_state
->wm
.need_postvbl_update
)
1555 mutex_lock(&dev_priv
->wm
.wm_mutex
);
1556 intel_crtc
->wm
.active
.g4x
= crtc_state
->wm
.g4x
.optimal
;
1557 g4x_program_watermarks(dev_priv
);
1558 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
1561 /* latency must be in 0.1us units. */
1562 static unsigned int vlv_wm_method2(unsigned int pixel_rate
,
1563 unsigned int htotal
,
1566 unsigned int latency
)
1570 ret
= intel_wm_method2(pixel_rate
, htotal
,
1571 width
, cpp
, latency
);
1572 ret
= DIV_ROUND_UP(ret
, 64);
1577 static void vlv_setup_wm_latency(struct drm_i915_private
*dev_priv
)
1579 /* all latencies in usec */
1580 dev_priv
->wm
.pri_latency
[VLV_WM_LEVEL_PM2
] = 3;
1582 dev_priv
->wm
.max_level
= VLV_WM_LEVEL_PM2
;
1584 if (IS_CHERRYVIEW(dev_priv
)) {
1585 dev_priv
->wm
.pri_latency
[VLV_WM_LEVEL_PM5
] = 12;
1586 dev_priv
->wm
.pri_latency
[VLV_WM_LEVEL_DDR_DVFS
] = 33;
1588 dev_priv
->wm
.max_level
= VLV_WM_LEVEL_DDR_DVFS
;
1592 static uint16_t vlv_compute_wm_level(const struct intel_crtc_state
*crtc_state
,
1593 const struct intel_plane_state
*plane_state
,
1596 struct intel_plane
*plane
= to_intel_plane(plane_state
->base
.plane
);
1597 struct drm_i915_private
*dev_priv
= to_i915(plane
->base
.dev
);
1598 const struct drm_display_mode
*adjusted_mode
=
1599 &crtc_state
->base
.adjusted_mode
;
1600 int clock
, htotal
, cpp
, width
, wm
;
1602 if (dev_priv
->wm
.pri_latency
[level
] == 0)
1605 if (!intel_wm_plane_visible(crtc_state
, plane_state
))
1608 cpp
= plane_state
->base
.fb
->format
->cpp
[0];
1609 clock
= adjusted_mode
->crtc_clock
;
1610 htotal
= adjusted_mode
->crtc_htotal
;
1611 width
= crtc_state
->pipe_src_w
;
1613 if (plane
->id
== PLANE_CURSOR
) {
1615 * FIXME the formula gives values that are
1616 * too big for the cursor FIFO, and hence we
1617 * would never be able to use cursors. For
1618 * now just hardcode the watermark.
1622 wm
= vlv_wm_method2(clock
, htotal
, width
, cpp
,
1623 dev_priv
->wm
.pri_latency
[level
] * 10);
1626 return min_t(int, wm
, USHRT_MAX
);
1629 static bool vlv_need_sprite0_fifo_workaround(unsigned int active_planes
)
1631 return (active_planes
& (BIT(PLANE_SPRITE0
) |
1632 BIT(PLANE_SPRITE1
))) == BIT(PLANE_SPRITE1
);
1635 static int vlv_compute_fifo(struct intel_crtc_state
*crtc_state
)
1637 struct intel_crtc
*crtc
= to_intel_crtc(crtc_state
->base
.crtc
);
1638 const struct g4x_pipe_wm
*raw
=
1639 &crtc_state
->wm
.vlv
.raw
[VLV_WM_LEVEL_PM2
];
1640 struct vlv_fifo_state
*fifo_state
= &crtc_state
->wm
.vlv
.fifo_state
;
1641 unsigned int active_planes
= crtc_state
->active_planes
& ~BIT(PLANE_CURSOR
);
1642 int num_active_planes
= hweight32(active_planes
);
1643 const int fifo_size
= 511;
1644 int fifo_extra
, fifo_left
= fifo_size
;
1645 int sprite0_fifo_extra
= 0;
1646 unsigned int total_rate
;
1647 enum plane_id plane_id
;
1650 * When enabling sprite0 after sprite1 has already been enabled
1651 * we tend to get an underrun unless sprite0 already has some
1652 * FIFO space allcoated. Hence we always allocate at least one
1653 * cacheline for sprite0 whenever sprite1 is enabled.
1655 * All other plane enable sequences appear immune to this problem.
1657 if (vlv_need_sprite0_fifo_workaround(active_planes
))
1658 sprite0_fifo_extra
= 1;
1660 total_rate
= raw
->plane
[PLANE_PRIMARY
] +
1661 raw
->plane
[PLANE_SPRITE0
] +
1662 raw
->plane
[PLANE_SPRITE1
] +
1665 if (total_rate
> fifo_size
)
1668 if (total_rate
== 0)
1671 for_each_plane_id_on_crtc(crtc
, plane_id
) {
1674 if ((active_planes
& BIT(plane_id
)) == 0) {
1675 fifo_state
->plane
[plane_id
] = 0;
1679 rate
= raw
->plane
[plane_id
];
1680 fifo_state
->plane
[plane_id
] = fifo_size
* rate
/ total_rate
;
1681 fifo_left
-= fifo_state
->plane
[plane_id
];
1684 fifo_state
->plane
[PLANE_SPRITE0
] += sprite0_fifo_extra
;
1685 fifo_left
-= sprite0_fifo_extra
;
1687 fifo_state
->plane
[PLANE_CURSOR
] = 63;
1689 fifo_extra
= DIV_ROUND_UP(fifo_left
, num_active_planes
?: 1);
1691 /* spread the remainder evenly */
1692 for_each_plane_id_on_crtc(crtc
, plane_id
) {
1698 if ((active_planes
& BIT(plane_id
)) == 0)
1701 plane_extra
= min(fifo_extra
, fifo_left
);
1702 fifo_state
->plane
[plane_id
] += plane_extra
;
1703 fifo_left
-= plane_extra
;
1706 WARN_ON(active_planes
!= 0 && fifo_left
!= 0);
1708 /* give it all to the first plane if none are active */
1709 if (active_planes
== 0) {
1710 WARN_ON(fifo_left
!= fifo_size
);
1711 fifo_state
->plane
[PLANE_PRIMARY
] = fifo_left
;
1717 /* mark all levels starting from 'level' as invalid */
1718 static void vlv_invalidate_wms(struct intel_crtc
*crtc
,
1719 struct vlv_wm_state
*wm_state
, int level
)
1721 struct drm_i915_private
*dev_priv
= to_i915(crtc
->base
.dev
);
1723 for (; level
< intel_wm_num_levels(dev_priv
); level
++) {
1724 enum plane_id plane_id
;
1726 for_each_plane_id_on_crtc(crtc
, plane_id
)
1727 wm_state
->wm
[level
].plane
[plane_id
] = USHRT_MAX
;
1729 wm_state
->sr
[level
].cursor
= USHRT_MAX
;
1730 wm_state
->sr
[level
].plane
= USHRT_MAX
;
1734 static u16
vlv_invert_wm_value(u16 wm
, u16 fifo_size
)
1739 return fifo_size
- wm
;
1743 * Starting from 'level' set all higher
1744 * levels to 'value' in the "raw" watermarks.
1746 static bool vlv_raw_plane_wm_set(struct intel_crtc_state
*crtc_state
,
1747 int level
, enum plane_id plane_id
, u16 value
)
1749 struct drm_i915_private
*dev_priv
= to_i915(crtc_state
->base
.crtc
->dev
);
1750 int num_levels
= intel_wm_num_levels(dev_priv
);
1753 for (; level
< num_levels
; level
++) {
1754 struct g4x_pipe_wm
*raw
= &crtc_state
->wm
.vlv
.raw
[level
];
1756 dirty
|= raw
->plane
[plane_id
] != value
;
1757 raw
->plane
[plane_id
] = value
;
1763 static bool vlv_raw_plane_wm_compute(struct intel_crtc_state
*crtc_state
,
1764 const struct intel_plane_state
*plane_state
)
1766 struct intel_plane
*plane
= to_intel_plane(plane_state
->base
.plane
);
1767 enum plane_id plane_id
= plane
->id
;
1768 int num_levels
= intel_wm_num_levels(to_i915(plane
->base
.dev
));
1772 if (!intel_wm_plane_visible(crtc_state
, plane_state
)) {
1773 dirty
|= vlv_raw_plane_wm_set(crtc_state
, 0, plane_id
, 0);
1777 for (level
= 0; level
< num_levels
; level
++) {
1778 struct g4x_pipe_wm
*raw
= &crtc_state
->wm
.vlv
.raw
[level
];
1779 int wm
= vlv_compute_wm_level(crtc_state
, plane_state
, level
);
1780 int max_wm
= plane_id
== PLANE_CURSOR
? 63 : 511;
1785 dirty
|= raw
->plane
[plane_id
] != wm
;
1786 raw
->plane
[plane_id
] = wm
;
1789 /* mark all higher levels as invalid */
1790 dirty
|= vlv_raw_plane_wm_set(crtc_state
, level
, plane_id
, USHRT_MAX
);
1794 DRM_DEBUG_KMS("%s watermarks: PM2=%d, PM5=%d, DDR DVFS=%d\n",
1796 crtc_state
->wm
.vlv
.raw
[VLV_WM_LEVEL_PM2
].plane
[plane_id
],
1797 crtc_state
->wm
.vlv
.raw
[VLV_WM_LEVEL_PM5
].plane
[plane_id
],
1798 crtc_state
->wm
.vlv
.raw
[VLV_WM_LEVEL_DDR_DVFS
].plane
[plane_id
]);
1803 static bool vlv_raw_plane_wm_is_valid(const struct intel_crtc_state
*crtc_state
,
1804 enum plane_id plane_id
, int level
)
1806 const struct g4x_pipe_wm
*raw
=
1807 &crtc_state
->wm
.vlv
.raw
[level
];
1808 const struct vlv_fifo_state
*fifo_state
=
1809 &crtc_state
->wm
.vlv
.fifo_state
;
1811 return raw
->plane
[plane_id
] <= fifo_state
->plane
[plane_id
];
1814 static bool vlv_raw_crtc_wm_is_valid(const struct intel_crtc_state
*crtc_state
, int level
)
1816 return vlv_raw_plane_wm_is_valid(crtc_state
, PLANE_PRIMARY
, level
) &&
1817 vlv_raw_plane_wm_is_valid(crtc_state
, PLANE_SPRITE0
, level
) &&
1818 vlv_raw_plane_wm_is_valid(crtc_state
, PLANE_SPRITE1
, level
) &&
1819 vlv_raw_plane_wm_is_valid(crtc_state
, PLANE_CURSOR
, level
);
1822 static int vlv_compute_pipe_wm(struct intel_crtc_state
*crtc_state
)
1824 struct intel_crtc
*crtc
= to_intel_crtc(crtc_state
->base
.crtc
);
1825 struct drm_i915_private
*dev_priv
= to_i915(crtc
->base
.dev
);
1826 struct intel_atomic_state
*state
=
1827 to_intel_atomic_state(crtc_state
->base
.state
);
1828 struct vlv_wm_state
*wm_state
= &crtc_state
->wm
.vlv
.optimal
;
1829 const struct vlv_fifo_state
*fifo_state
=
1830 &crtc_state
->wm
.vlv
.fifo_state
;
1831 int num_active_planes
= hweight32(crtc_state
->active_planes
&
1832 ~BIT(PLANE_CURSOR
));
1833 bool needs_modeset
= drm_atomic_crtc_needs_modeset(&crtc_state
->base
);
1834 struct intel_plane_state
*plane_state
;
1835 struct intel_plane
*plane
;
1836 enum plane_id plane_id
;
1838 unsigned int dirty
= 0;
1840 for_each_intel_plane_in_state(state
, plane
, plane_state
, i
) {
1841 const struct intel_plane_state
*old_plane_state
=
1842 to_intel_plane_state(plane
->base
.state
);
1844 if (plane_state
->base
.crtc
!= &crtc
->base
&&
1845 old_plane_state
->base
.crtc
!= &crtc
->base
)
1848 if (vlv_raw_plane_wm_compute(crtc_state
, plane_state
))
1849 dirty
|= BIT(plane
->id
);
1853 * DSPARB registers may have been reset due to the
1854 * power well being turned off. Make sure we restore
1855 * them to a consistent state even if no primary/sprite
1856 * planes are initially active.
1859 crtc_state
->fifo_changed
= true;
1864 /* cursor changes don't warrant a FIFO recompute */
1865 if (dirty
& ~BIT(PLANE_CURSOR
)) {
1866 const struct intel_crtc_state
*old_crtc_state
=
1867 to_intel_crtc_state(crtc
->base
.state
);
1868 const struct vlv_fifo_state
*old_fifo_state
=
1869 &old_crtc_state
->wm
.vlv
.fifo_state
;
1871 ret
= vlv_compute_fifo(crtc_state
);
1875 if (needs_modeset
||
1876 memcmp(old_fifo_state
, fifo_state
,
1877 sizeof(*fifo_state
)) != 0)
1878 crtc_state
->fifo_changed
= true;
1881 /* initially allow all levels */
1882 wm_state
->num_levels
= intel_wm_num_levels(dev_priv
);
1884 * Note that enabling cxsr with no primary/sprite planes
1885 * enabled can wedge the pipe. Hence we only allow cxsr
1886 * with exactly one enabled primary/sprite plane.
1888 wm_state
->cxsr
= crtc
->pipe
!= PIPE_C
&& num_active_planes
== 1;
1890 for (level
= 0; level
< wm_state
->num_levels
; level
++) {
1891 const struct g4x_pipe_wm
*raw
= &crtc_state
->wm
.vlv
.raw
[level
];
1892 const int sr_fifo_size
= INTEL_INFO(dev_priv
)->num_pipes
* 512 - 1;
1894 if (!vlv_raw_crtc_wm_is_valid(crtc_state
, level
))
1897 for_each_plane_id_on_crtc(crtc
, plane_id
) {
1898 wm_state
->wm
[level
].plane
[plane_id
] =
1899 vlv_invert_wm_value(raw
->plane
[plane_id
],
1900 fifo_state
->plane
[plane_id
]);
1903 wm_state
->sr
[level
].plane
=
1904 vlv_invert_wm_value(max3(raw
->plane
[PLANE_PRIMARY
],
1905 raw
->plane
[PLANE_SPRITE0
],
1906 raw
->plane
[PLANE_SPRITE1
]),
1909 wm_state
->sr
[level
].cursor
=
1910 vlv_invert_wm_value(raw
->plane
[PLANE_CURSOR
],
1917 /* limit to only levels we can actually handle */
1918 wm_state
->num_levels
= level
;
1920 /* invalidate the higher levels */
1921 vlv_invalidate_wms(crtc
, wm_state
, level
);
1926 #define VLV_FIFO(plane, value) \
1927 (((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1929 static void vlv_atomic_update_fifo(struct intel_atomic_state
*state
,
1930 struct intel_crtc_state
*crtc_state
)
1932 struct intel_crtc
*crtc
= to_intel_crtc(crtc_state
->base
.crtc
);
1933 struct drm_i915_private
*dev_priv
= to_i915(crtc
->base
.dev
);
1934 const struct vlv_fifo_state
*fifo_state
=
1935 &crtc_state
->wm
.vlv
.fifo_state
;
1936 int sprite0_start
, sprite1_start
, fifo_size
;
1938 if (!crtc_state
->fifo_changed
)
1941 sprite0_start
= fifo_state
->plane
[PLANE_PRIMARY
];
1942 sprite1_start
= fifo_state
->plane
[PLANE_SPRITE0
] + sprite0_start
;
1943 fifo_size
= fifo_state
->plane
[PLANE_SPRITE1
] + sprite1_start
;
1945 WARN_ON(fifo_state
->plane
[PLANE_CURSOR
] != 63);
1946 WARN_ON(fifo_size
!= 511);
1948 trace_vlv_fifo_size(crtc
, sprite0_start
, sprite1_start
, fifo_size
);
1951 * uncore.lock serves a double purpose here. It allows us to
1952 * use the less expensive I915_{READ,WRITE}_FW() functions, and
1953 * it protects the DSPARB registers from getting clobbered by
1954 * parallel updates from multiple pipes.
1956 * intel_pipe_update_start() has already disabled interrupts
1957 * for us, so a plain spin_lock() is sufficient here.
1959 spin_lock(&dev_priv
->uncore
.lock
);
1961 switch (crtc
->pipe
) {
1962 uint32_t dsparb
, dsparb2
, dsparb3
;
1964 dsparb
= I915_READ_FW(DSPARB
);
1965 dsparb2
= I915_READ_FW(DSPARB2
);
1967 dsparb
&= ~(VLV_FIFO(SPRITEA
, 0xff) |
1968 VLV_FIFO(SPRITEB
, 0xff));
1969 dsparb
|= (VLV_FIFO(SPRITEA
, sprite0_start
) |
1970 VLV_FIFO(SPRITEB
, sprite1_start
));
1972 dsparb2
&= ~(VLV_FIFO(SPRITEA_HI
, 0x1) |
1973 VLV_FIFO(SPRITEB_HI
, 0x1));
1974 dsparb2
|= (VLV_FIFO(SPRITEA_HI
, sprite0_start
>> 8) |
1975 VLV_FIFO(SPRITEB_HI
, sprite1_start
>> 8));
1977 I915_WRITE_FW(DSPARB
, dsparb
);
1978 I915_WRITE_FW(DSPARB2
, dsparb2
);
1981 dsparb
= I915_READ_FW(DSPARB
);
1982 dsparb2
= I915_READ_FW(DSPARB2
);
1984 dsparb
&= ~(VLV_FIFO(SPRITEC
, 0xff) |
1985 VLV_FIFO(SPRITED
, 0xff));
1986 dsparb
|= (VLV_FIFO(SPRITEC
, sprite0_start
) |
1987 VLV_FIFO(SPRITED
, sprite1_start
));
1989 dsparb2
&= ~(VLV_FIFO(SPRITEC_HI
, 0xff) |
1990 VLV_FIFO(SPRITED_HI
, 0xff));
1991 dsparb2
|= (VLV_FIFO(SPRITEC_HI
, sprite0_start
>> 8) |
1992 VLV_FIFO(SPRITED_HI
, sprite1_start
>> 8));
1994 I915_WRITE_FW(DSPARB
, dsparb
);
1995 I915_WRITE_FW(DSPARB2
, dsparb2
);
1998 dsparb3
= I915_READ_FW(DSPARB3
);
1999 dsparb2
= I915_READ_FW(DSPARB2
);
2001 dsparb3
&= ~(VLV_FIFO(SPRITEE
, 0xff) |
2002 VLV_FIFO(SPRITEF
, 0xff));
2003 dsparb3
|= (VLV_FIFO(SPRITEE
, sprite0_start
) |
2004 VLV_FIFO(SPRITEF
, sprite1_start
));
2006 dsparb2
&= ~(VLV_FIFO(SPRITEE_HI
, 0xff) |
2007 VLV_FIFO(SPRITEF_HI
, 0xff));
2008 dsparb2
|= (VLV_FIFO(SPRITEE_HI
, sprite0_start
>> 8) |
2009 VLV_FIFO(SPRITEF_HI
, sprite1_start
>> 8));
2011 I915_WRITE_FW(DSPARB3
, dsparb3
);
2012 I915_WRITE_FW(DSPARB2
, dsparb2
);
2018 POSTING_READ_FW(DSPARB
);
2020 spin_unlock(&dev_priv
->uncore
.lock
);
2025 static int vlv_compute_intermediate_wm(struct drm_device
*dev
,
2026 struct intel_crtc
*crtc
,
2027 struct intel_crtc_state
*crtc_state
)
2029 struct vlv_wm_state
*intermediate
= &crtc_state
->wm
.vlv
.intermediate
;
2030 const struct vlv_wm_state
*optimal
= &crtc_state
->wm
.vlv
.optimal
;
2031 const struct vlv_wm_state
*active
= &crtc
->wm
.active
.vlv
;
2034 intermediate
->num_levels
= min(optimal
->num_levels
, active
->num_levels
);
2035 intermediate
->cxsr
= optimal
->cxsr
&& active
->cxsr
&&
2036 !crtc_state
->disable_cxsr
;
2038 for (level
= 0; level
< intermediate
->num_levels
; level
++) {
2039 enum plane_id plane_id
;
2041 for_each_plane_id_on_crtc(crtc
, plane_id
) {
2042 intermediate
->wm
[level
].plane
[plane_id
] =
2043 min(optimal
->wm
[level
].plane
[plane_id
],
2044 active
->wm
[level
].plane
[plane_id
]);
2047 intermediate
->sr
[level
].plane
= min(optimal
->sr
[level
].plane
,
2048 active
->sr
[level
].plane
);
2049 intermediate
->sr
[level
].cursor
= min(optimal
->sr
[level
].cursor
,
2050 active
->sr
[level
].cursor
);
2053 vlv_invalidate_wms(crtc
, intermediate
, level
);
2056 * If our intermediate WM are identical to the final WM, then we can
2057 * omit the post-vblank programming; only update if it's different.
2059 if (memcmp(intermediate
, optimal
, sizeof(*intermediate
)) != 0)
2060 crtc_state
->wm
.need_postvbl_update
= true;
2065 static void vlv_merge_wm(struct drm_i915_private
*dev_priv
,
2066 struct vlv_wm_values
*wm
)
2068 struct intel_crtc
*crtc
;
2069 int num_active_crtcs
= 0;
2071 wm
->level
= dev_priv
->wm
.max_level
;
2074 for_each_intel_crtc(&dev_priv
->drm
, crtc
) {
2075 const struct vlv_wm_state
*wm_state
= &crtc
->wm
.active
.vlv
;
2080 if (!wm_state
->cxsr
)
2084 wm
->level
= min_t(int, wm
->level
, wm_state
->num_levels
- 1);
2087 if (num_active_crtcs
!= 1)
2090 if (num_active_crtcs
> 1)
2091 wm
->level
= VLV_WM_LEVEL_PM2
;
2093 for_each_intel_crtc(&dev_priv
->drm
, crtc
) {
2094 const struct vlv_wm_state
*wm_state
= &crtc
->wm
.active
.vlv
;
2095 enum pipe pipe
= crtc
->pipe
;
2097 wm
->pipe
[pipe
] = wm_state
->wm
[wm
->level
];
2098 if (crtc
->active
&& wm
->cxsr
)
2099 wm
->sr
= wm_state
->sr
[wm
->level
];
2101 wm
->ddl
[pipe
].plane
[PLANE_PRIMARY
] = DDL_PRECISION_HIGH
| 2;
2102 wm
->ddl
[pipe
].plane
[PLANE_SPRITE0
] = DDL_PRECISION_HIGH
| 2;
2103 wm
->ddl
[pipe
].plane
[PLANE_SPRITE1
] = DDL_PRECISION_HIGH
| 2;
2104 wm
->ddl
[pipe
].plane
[PLANE_CURSOR
] = DDL_PRECISION_HIGH
| 2;
2108 static void vlv_program_watermarks(struct drm_i915_private
*dev_priv
)
2110 struct vlv_wm_values
*old_wm
= &dev_priv
->wm
.vlv
;
2111 struct vlv_wm_values new_wm
= {};
2113 vlv_merge_wm(dev_priv
, &new_wm
);
2115 if (memcmp(old_wm
, &new_wm
, sizeof(new_wm
)) == 0)
2118 if (is_disabling(old_wm
->level
, new_wm
.level
, VLV_WM_LEVEL_DDR_DVFS
))
2119 chv_set_memory_dvfs(dev_priv
, false);
2121 if (is_disabling(old_wm
->level
, new_wm
.level
, VLV_WM_LEVEL_PM5
))
2122 chv_set_memory_pm5(dev_priv
, false);
2124 if (is_disabling(old_wm
->cxsr
, new_wm
.cxsr
, true))
2125 _intel_set_memory_cxsr(dev_priv
, false);
2127 vlv_write_wm_values(dev_priv
, &new_wm
);
2129 if (is_enabling(old_wm
->cxsr
, new_wm
.cxsr
, true))
2130 _intel_set_memory_cxsr(dev_priv
, true);
2132 if (is_enabling(old_wm
->level
, new_wm
.level
, VLV_WM_LEVEL_PM5
))
2133 chv_set_memory_pm5(dev_priv
, true);
2135 if (is_enabling(old_wm
->level
, new_wm
.level
, VLV_WM_LEVEL_DDR_DVFS
))
2136 chv_set_memory_dvfs(dev_priv
, true);
2141 static void vlv_initial_watermarks(struct intel_atomic_state
*state
,
2142 struct intel_crtc_state
*crtc_state
)
2144 struct drm_i915_private
*dev_priv
= to_i915(crtc_state
->base
.crtc
->dev
);
2145 struct intel_crtc
*crtc
= to_intel_crtc(crtc_state
->base
.crtc
);
2147 mutex_lock(&dev_priv
->wm
.wm_mutex
);
2148 crtc
->wm
.active
.vlv
= crtc_state
->wm
.vlv
.intermediate
;
2149 vlv_program_watermarks(dev_priv
);
2150 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
2153 static void vlv_optimize_watermarks(struct intel_atomic_state
*state
,
2154 struct intel_crtc_state
*crtc_state
)
2156 struct drm_i915_private
*dev_priv
= to_i915(crtc_state
->base
.crtc
->dev
);
2157 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc_state
->base
.crtc
);
2159 if (!crtc_state
->wm
.need_postvbl_update
)
2162 mutex_lock(&dev_priv
->wm
.wm_mutex
);
2163 intel_crtc
->wm
.active
.vlv
= crtc_state
->wm
.vlv
.optimal
;
2164 vlv_program_watermarks(dev_priv
);
2165 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
2168 static void i965_update_wm(struct intel_crtc
*unused_crtc
)
2170 struct drm_i915_private
*dev_priv
= to_i915(unused_crtc
->base
.dev
);
2171 struct intel_crtc
*crtc
;
2176 /* Calc sr entries for one plane configs */
2177 crtc
= single_enabled_crtc(dev_priv
);
2179 /* self-refresh has much higher latency */
2180 static const int sr_latency_ns
= 12000;
2181 const struct drm_display_mode
*adjusted_mode
=
2182 &crtc
->config
->base
.adjusted_mode
;
2183 const struct drm_framebuffer
*fb
=
2184 crtc
->base
.primary
->state
->fb
;
2185 int clock
= adjusted_mode
->crtc_clock
;
2186 int htotal
= adjusted_mode
->crtc_htotal
;
2187 int hdisplay
= crtc
->config
->pipe_src_w
;
2188 int cpp
= fb
->format
->cpp
[0];
2191 entries
= intel_wm_method2(clock
, htotal
,
2192 hdisplay
, cpp
, sr_latency_ns
/ 100);
2193 entries
= DIV_ROUND_UP(entries
, I915_FIFO_LINE_SIZE
);
2194 srwm
= I965_FIFO_SIZE
- entries
;
2198 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
2201 entries
= intel_wm_method2(clock
, htotal
,
2202 crtc
->base
.cursor
->state
->crtc_w
, 4,
2203 sr_latency_ns
/ 100);
2204 entries
= DIV_ROUND_UP(entries
,
2205 i965_cursor_wm_info
.cacheline_size
) +
2206 i965_cursor_wm_info
.guard_size
;
2208 cursor_sr
= i965_cursor_wm_info
.fifo_size
- entries
;
2209 if (cursor_sr
> i965_cursor_wm_info
.max_wm
)
2210 cursor_sr
= i965_cursor_wm_info
.max_wm
;
2212 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
2213 "cursor %d\n", srwm
, cursor_sr
);
2215 cxsr_enabled
= true;
2217 cxsr_enabled
= false;
2218 /* Turn off self refresh if both pipes are enabled */
2219 intel_set_memory_cxsr(dev_priv
, false);
2222 DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
2225 /* 965 has limitations... */
2226 I915_WRITE(DSPFW1
, FW_WM(srwm
, SR
) |
2230 I915_WRITE(DSPFW2
, FW_WM(8, CURSORA
) |
2231 FW_WM(8, PLANEC_OLD
));
2232 /* update cursor SR watermark */
2233 I915_WRITE(DSPFW3
, FW_WM(cursor_sr
, CURSOR_SR
));
2236 intel_set_memory_cxsr(dev_priv
, true);
2241 static void i9xx_update_wm(struct intel_crtc
*unused_crtc
)
2243 struct drm_i915_private
*dev_priv
= to_i915(unused_crtc
->base
.dev
);
2244 const struct intel_watermark_params
*wm_info
;
2249 int planea_wm
, planeb_wm
;
2250 struct intel_crtc
*crtc
, *enabled
= NULL
;
2252 if (IS_I945GM(dev_priv
))
2253 wm_info
= &i945_wm_info
;
2254 else if (!IS_GEN2(dev_priv
))
2255 wm_info
= &i915_wm_info
;
2257 wm_info
= &i830_a_wm_info
;
2259 fifo_size
= dev_priv
->display
.get_fifo_size(dev_priv
, 0);
2260 crtc
= intel_get_crtc_for_plane(dev_priv
, 0);
2261 if (intel_crtc_active(crtc
)) {
2262 const struct drm_display_mode
*adjusted_mode
=
2263 &crtc
->config
->base
.adjusted_mode
;
2264 const struct drm_framebuffer
*fb
=
2265 crtc
->base
.primary
->state
->fb
;
2268 if (IS_GEN2(dev_priv
))
2271 cpp
= fb
->format
->cpp
[0];
2273 planea_wm
= intel_calculate_wm(adjusted_mode
->crtc_clock
,
2274 wm_info
, fifo_size
, cpp
,
2275 pessimal_latency_ns
);
2278 planea_wm
= fifo_size
- wm_info
->guard_size
;
2279 if (planea_wm
> (long)wm_info
->max_wm
)
2280 planea_wm
= wm_info
->max_wm
;
2283 if (IS_GEN2(dev_priv
))
2284 wm_info
= &i830_bc_wm_info
;
2286 fifo_size
= dev_priv
->display
.get_fifo_size(dev_priv
, 1);
2287 crtc
= intel_get_crtc_for_plane(dev_priv
, 1);
2288 if (intel_crtc_active(crtc
)) {
2289 const struct drm_display_mode
*adjusted_mode
=
2290 &crtc
->config
->base
.adjusted_mode
;
2291 const struct drm_framebuffer
*fb
=
2292 crtc
->base
.primary
->state
->fb
;
2295 if (IS_GEN2(dev_priv
))
2298 cpp
= fb
->format
->cpp
[0];
2300 planeb_wm
= intel_calculate_wm(adjusted_mode
->crtc_clock
,
2301 wm_info
, fifo_size
, cpp
,
2302 pessimal_latency_ns
);
2303 if (enabled
== NULL
)
2308 planeb_wm
= fifo_size
- wm_info
->guard_size
;
2309 if (planeb_wm
> (long)wm_info
->max_wm
)
2310 planeb_wm
= wm_info
->max_wm
;
2313 DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm
, planeb_wm
);
2315 if (IS_I915GM(dev_priv
) && enabled
) {
2316 struct drm_i915_gem_object
*obj
;
2318 obj
= intel_fb_obj(enabled
->base
.primary
->state
->fb
);
2320 /* self-refresh seems busted with untiled */
2321 if (!i915_gem_object_is_tiled(obj
))
2326 * Overlay gets an aggressive default since video jitter is bad.
2330 /* Play safe and disable self-refresh before adjusting watermarks. */
2331 intel_set_memory_cxsr(dev_priv
, false);
2333 /* Calc sr entries for one plane configs */
2334 if (HAS_FW_BLC(dev_priv
) && enabled
) {
2335 /* self-refresh has much higher latency */
2336 static const int sr_latency_ns
= 6000;
2337 const struct drm_display_mode
*adjusted_mode
=
2338 &enabled
->config
->base
.adjusted_mode
;
2339 const struct drm_framebuffer
*fb
=
2340 enabled
->base
.primary
->state
->fb
;
2341 int clock
= adjusted_mode
->crtc_clock
;
2342 int htotal
= adjusted_mode
->crtc_htotal
;
2343 int hdisplay
= enabled
->config
->pipe_src_w
;
2347 if (IS_I915GM(dev_priv
) || IS_I945GM(dev_priv
))
2350 cpp
= fb
->format
->cpp
[0];
2352 entries
= intel_wm_method2(clock
, htotal
, hdisplay
, cpp
,
2353 sr_latency_ns
/ 100);
2354 entries
= DIV_ROUND_UP(entries
, wm_info
->cacheline_size
);
2355 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries
);
2356 srwm
= wm_info
->fifo_size
- entries
;
2360 if (IS_I945G(dev_priv
) || IS_I945GM(dev_priv
))
2361 I915_WRITE(FW_BLC_SELF
,
2362 FW_BLC_SELF_FIFO_MASK
| (srwm
& 0xff));
2364 I915_WRITE(FW_BLC_SELF
, srwm
& 0x3f);
2367 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
2368 planea_wm
, planeb_wm
, cwm
, srwm
);
2370 fwater_lo
= ((planeb_wm
& 0x3f) << 16) | (planea_wm
& 0x3f);
2371 fwater_hi
= (cwm
& 0x1f);
2373 /* Set request length to 8 cachelines per fetch */
2374 fwater_lo
= fwater_lo
| (1 << 24) | (1 << 8);
2375 fwater_hi
= fwater_hi
| (1 << 8);
2377 I915_WRITE(FW_BLC
, fwater_lo
);
2378 I915_WRITE(FW_BLC2
, fwater_hi
);
2381 intel_set_memory_cxsr(dev_priv
, true);
2384 static void i845_update_wm(struct intel_crtc
*unused_crtc
)
2386 struct drm_i915_private
*dev_priv
= to_i915(unused_crtc
->base
.dev
);
2387 struct intel_crtc
*crtc
;
2388 const struct drm_display_mode
*adjusted_mode
;
2392 crtc
= single_enabled_crtc(dev_priv
);
2396 adjusted_mode
= &crtc
->config
->base
.adjusted_mode
;
2397 planea_wm
= intel_calculate_wm(adjusted_mode
->crtc_clock
,
2399 dev_priv
->display
.get_fifo_size(dev_priv
, 0),
2400 4, pessimal_latency_ns
);
2401 fwater_lo
= I915_READ(FW_BLC
) & ~0xfff;
2402 fwater_lo
|= (3<<8) | planea_wm
;
2404 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm
);
2406 I915_WRITE(FW_BLC
, fwater_lo
);
2409 /* latency must be in 0.1us units. */
2410 static unsigned int ilk_wm_method1(unsigned int pixel_rate
,
2412 unsigned int latency
)
2416 ret
= intel_wm_method1(pixel_rate
, cpp
, latency
);
2417 ret
= DIV_ROUND_UP(ret
, 64) + 2;
2422 /* latency must be in 0.1us units. */
2423 static unsigned int ilk_wm_method2(unsigned int pixel_rate
,
2424 unsigned int htotal
,
2427 unsigned int latency
)
2431 ret
= intel_wm_method2(pixel_rate
, htotal
,
2432 width
, cpp
, latency
);
2433 ret
= DIV_ROUND_UP(ret
, 64) + 2;
2438 static uint32_t ilk_wm_fbc(uint32_t pri_val
, uint32_t horiz_pixels
,
2442 * Neither of these should be possible since this function shouldn't be
2443 * called if the CRTC is off or the plane is invisible. But let's be
2444 * extra paranoid to avoid a potential divide-by-zero if we screw up
2445 * elsewhere in the driver.
2449 if (WARN_ON(!horiz_pixels
))
2452 return DIV_ROUND_UP(pri_val
* 64, horiz_pixels
* cpp
) + 2;
2455 struct ilk_wm_maximums
{
2463 * For both WM_PIPE and WM_LP.
2464 * mem_value must be in 0.1us units.
2466 static uint32_t ilk_compute_pri_wm(const struct intel_crtc_state
*cstate
,
2467 const struct intel_plane_state
*pstate
,
2471 uint32_t method1
, method2
;
2474 if (!intel_wm_plane_visible(cstate
, pstate
))
2477 cpp
= pstate
->base
.fb
->format
->cpp
[0];
2479 method1
= ilk_wm_method1(cstate
->pixel_rate
, cpp
, mem_value
);
2484 method2
= ilk_wm_method2(cstate
->pixel_rate
,
2485 cstate
->base
.adjusted_mode
.crtc_htotal
,
2486 drm_rect_width(&pstate
->base
.dst
),
2489 return min(method1
, method2
);
2493 * For both WM_PIPE and WM_LP.
2494 * mem_value must be in 0.1us units.
2496 static uint32_t ilk_compute_spr_wm(const struct intel_crtc_state
*cstate
,
2497 const struct intel_plane_state
*pstate
,
2500 uint32_t method1
, method2
;
2503 if (!intel_wm_plane_visible(cstate
, pstate
))
2506 cpp
= pstate
->base
.fb
->format
->cpp
[0];
2508 method1
= ilk_wm_method1(cstate
->pixel_rate
, cpp
, mem_value
);
2509 method2
= ilk_wm_method2(cstate
->pixel_rate
,
2510 cstate
->base
.adjusted_mode
.crtc_htotal
,
2511 drm_rect_width(&pstate
->base
.dst
),
2513 return min(method1
, method2
);
2517 * For both WM_PIPE and WM_LP.
2518 * mem_value must be in 0.1us units.
2520 static uint32_t ilk_compute_cur_wm(const struct intel_crtc_state
*cstate
,
2521 const struct intel_plane_state
*pstate
,
2526 if (!intel_wm_plane_visible(cstate
, pstate
))
2529 cpp
= pstate
->base
.fb
->format
->cpp
[0];
2531 return ilk_wm_method2(cstate
->pixel_rate
,
2532 cstate
->base
.adjusted_mode
.crtc_htotal
,
2533 pstate
->base
.crtc_w
, cpp
, mem_value
);
2536 /* Only for WM_LP. */
2537 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state
*cstate
,
2538 const struct intel_plane_state
*pstate
,
2543 if (!intel_wm_plane_visible(cstate
, pstate
))
2546 cpp
= pstate
->base
.fb
->format
->cpp
[0];
2548 return ilk_wm_fbc(pri_val
, drm_rect_width(&pstate
->base
.dst
), cpp
);
2552 ilk_display_fifo_size(const struct drm_i915_private
*dev_priv
)
2554 if (INTEL_GEN(dev_priv
) >= 8)
2556 else if (INTEL_GEN(dev_priv
) >= 7)
2563 ilk_plane_wm_reg_max(const struct drm_i915_private
*dev_priv
,
2564 int level
, bool is_sprite
)
2566 if (INTEL_GEN(dev_priv
) >= 8)
2567 /* BDW primary/sprite plane watermarks */
2568 return level
== 0 ? 255 : 2047;
2569 else if (INTEL_GEN(dev_priv
) >= 7)
2570 /* IVB/HSW primary/sprite plane watermarks */
2571 return level
== 0 ? 127 : 1023;
2572 else if (!is_sprite
)
2573 /* ILK/SNB primary plane watermarks */
2574 return level
== 0 ? 127 : 511;
2576 /* ILK/SNB sprite plane watermarks */
2577 return level
== 0 ? 63 : 255;
2581 ilk_cursor_wm_reg_max(const struct drm_i915_private
*dev_priv
, int level
)
2583 if (INTEL_GEN(dev_priv
) >= 7)
2584 return level
== 0 ? 63 : 255;
2586 return level
== 0 ? 31 : 63;
2589 static unsigned int ilk_fbc_wm_reg_max(const struct drm_i915_private
*dev_priv
)
2591 if (INTEL_GEN(dev_priv
) >= 8)
2597 /* Calculate the maximum primary/sprite plane watermark */
2598 static unsigned int ilk_plane_wm_max(const struct drm_device
*dev
,
2600 const struct intel_wm_config
*config
,
2601 enum intel_ddb_partitioning ddb_partitioning
,
2604 struct drm_i915_private
*dev_priv
= to_i915(dev
);
2605 unsigned int fifo_size
= ilk_display_fifo_size(dev_priv
);
2607 /* if sprites aren't enabled, sprites get nothing */
2608 if (is_sprite
&& !config
->sprites_enabled
)
2611 /* HSW allows LP1+ watermarks even with multiple pipes */
2612 if (level
== 0 || config
->num_pipes_active
> 1) {
2613 fifo_size
/= INTEL_INFO(dev_priv
)->num_pipes
;
2616 * For some reason the non self refresh
2617 * FIFO size is only half of the self
2618 * refresh FIFO size on ILK/SNB.
2620 if (INTEL_GEN(dev_priv
) <= 6)
2624 if (config
->sprites_enabled
) {
2625 /* level 0 is always calculated with 1:1 split */
2626 if (level
> 0 && ddb_partitioning
== INTEL_DDB_PART_5_6
) {
2635 /* clamp to max that the registers can hold */
2636 return min(fifo_size
, ilk_plane_wm_reg_max(dev_priv
, level
, is_sprite
));
2639 /* Calculate the maximum cursor plane watermark */
2640 static unsigned int ilk_cursor_wm_max(const struct drm_device
*dev
,
2642 const struct intel_wm_config
*config
)
2644 /* HSW LP1+ watermarks w/ multiple pipes */
2645 if (level
> 0 && config
->num_pipes_active
> 1)
2648 /* otherwise just report max that registers can hold */
2649 return ilk_cursor_wm_reg_max(to_i915(dev
), level
);
2652 static void ilk_compute_wm_maximums(const struct drm_device
*dev
,
2654 const struct intel_wm_config
*config
,
2655 enum intel_ddb_partitioning ddb_partitioning
,
2656 struct ilk_wm_maximums
*max
)
2658 max
->pri
= ilk_plane_wm_max(dev
, level
, config
, ddb_partitioning
, false);
2659 max
->spr
= ilk_plane_wm_max(dev
, level
, config
, ddb_partitioning
, true);
2660 max
->cur
= ilk_cursor_wm_max(dev
, level
, config
);
2661 max
->fbc
= ilk_fbc_wm_reg_max(to_i915(dev
));
2664 static void ilk_compute_wm_reg_maximums(const struct drm_i915_private
*dev_priv
,
2666 struct ilk_wm_maximums
*max
)
2668 max
->pri
= ilk_plane_wm_reg_max(dev_priv
, level
, false);
2669 max
->spr
= ilk_plane_wm_reg_max(dev_priv
, level
, true);
2670 max
->cur
= ilk_cursor_wm_reg_max(dev_priv
, level
);
2671 max
->fbc
= ilk_fbc_wm_reg_max(dev_priv
);
2674 static bool ilk_validate_wm_level(int level
,
2675 const struct ilk_wm_maximums
*max
,
2676 struct intel_wm_level
*result
)
2680 /* already determined to be invalid? */
2681 if (!result
->enable
)
2684 result
->enable
= result
->pri_val
<= max
->pri
&&
2685 result
->spr_val
<= max
->spr
&&
2686 result
->cur_val
<= max
->cur
;
2688 ret
= result
->enable
;
2691 * HACK until we can pre-compute everything,
2692 * and thus fail gracefully if LP0 watermarks
2695 if (level
== 0 && !result
->enable
) {
2696 if (result
->pri_val
> max
->pri
)
2697 DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
2698 level
, result
->pri_val
, max
->pri
);
2699 if (result
->spr_val
> max
->spr
)
2700 DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
2701 level
, result
->spr_val
, max
->spr
);
2702 if (result
->cur_val
> max
->cur
)
2703 DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
2704 level
, result
->cur_val
, max
->cur
);
2706 result
->pri_val
= min_t(uint32_t, result
->pri_val
, max
->pri
);
2707 result
->spr_val
= min_t(uint32_t, result
->spr_val
, max
->spr
);
2708 result
->cur_val
= min_t(uint32_t, result
->cur_val
, max
->cur
);
2709 result
->enable
= true;
2715 static void ilk_compute_wm_level(const struct drm_i915_private
*dev_priv
,
2716 const struct intel_crtc
*intel_crtc
,
2718 struct intel_crtc_state
*cstate
,
2719 struct intel_plane_state
*pristate
,
2720 struct intel_plane_state
*sprstate
,
2721 struct intel_plane_state
*curstate
,
2722 struct intel_wm_level
*result
)
2724 uint16_t pri_latency
= dev_priv
->wm
.pri_latency
[level
];
2725 uint16_t spr_latency
= dev_priv
->wm
.spr_latency
[level
];
2726 uint16_t cur_latency
= dev_priv
->wm
.cur_latency
[level
];
2728 /* WM1+ latency values stored in 0.5us units */
2736 result
->pri_val
= ilk_compute_pri_wm(cstate
, pristate
,
2737 pri_latency
, level
);
2738 result
->fbc_val
= ilk_compute_fbc_wm(cstate
, pristate
, result
->pri_val
);
2742 result
->spr_val
= ilk_compute_spr_wm(cstate
, sprstate
, spr_latency
);
2745 result
->cur_val
= ilk_compute_cur_wm(cstate
, curstate
, cur_latency
);
2747 result
->enable
= true;
2751 hsw_compute_linetime_wm(const struct intel_crtc_state
*cstate
)
2753 const struct intel_atomic_state
*intel_state
=
2754 to_intel_atomic_state(cstate
->base
.state
);
2755 const struct drm_display_mode
*adjusted_mode
=
2756 &cstate
->base
.adjusted_mode
;
2757 u32 linetime
, ips_linetime
;
2759 if (!cstate
->base
.active
)
2761 if (WARN_ON(adjusted_mode
->crtc_clock
== 0))
2763 if (WARN_ON(intel_state
->cdclk
.logical
.cdclk
== 0))
2766 /* The WM are computed with base on how long it takes to fill a single
2767 * row at the given clock rate, multiplied by 8.
2769 linetime
= DIV_ROUND_CLOSEST(adjusted_mode
->crtc_htotal
* 1000 * 8,
2770 adjusted_mode
->crtc_clock
);
2771 ips_linetime
= DIV_ROUND_CLOSEST(adjusted_mode
->crtc_htotal
* 1000 * 8,
2772 intel_state
->cdclk
.logical
.cdclk
);
2774 return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime
) |
2775 PIPE_WM_LINETIME_TIME(linetime
);
2778 static void intel_read_wm_latency(struct drm_i915_private
*dev_priv
,
2781 if (INTEL_GEN(dev_priv
) >= 9) {
2784 int level
, max_level
= ilk_wm_max_level(dev_priv
);
2786 /* read the first set of memory latencies[0:3] */
2787 val
= 0; /* data0 to be programmed to 0 for first set */
2788 mutex_lock(&dev_priv
->rps
.hw_lock
);
2789 ret
= sandybridge_pcode_read(dev_priv
,
2790 GEN9_PCODE_READ_MEM_LATENCY
,
2792 mutex_unlock(&dev_priv
->rps
.hw_lock
);
2795 DRM_ERROR("SKL Mailbox read error = %d\n", ret
);
2799 wm
[0] = val
& GEN9_MEM_LATENCY_LEVEL_MASK
;
2800 wm
[1] = (val
>> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT
) &
2801 GEN9_MEM_LATENCY_LEVEL_MASK
;
2802 wm
[2] = (val
>> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT
) &
2803 GEN9_MEM_LATENCY_LEVEL_MASK
;
2804 wm
[3] = (val
>> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT
) &
2805 GEN9_MEM_LATENCY_LEVEL_MASK
;
2807 /* read the second set of memory latencies[4:7] */
2808 val
= 1; /* data0 to be programmed to 1 for second set */
2809 mutex_lock(&dev_priv
->rps
.hw_lock
);
2810 ret
= sandybridge_pcode_read(dev_priv
,
2811 GEN9_PCODE_READ_MEM_LATENCY
,
2813 mutex_unlock(&dev_priv
->rps
.hw_lock
);
2815 DRM_ERROR("SKL Mailbox read error = %d\n", ret
);
2819 wm
[4] = val
& GEN9_MEM_LATENCY_LEVEL_MASK
;
2820 wm
[5] = (val
>> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT
) &
2821 GEN9_MEM_LATENCY_LEVEL_MASK
;
2822 wm
[6] = (val
>> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT
) &
2823 GEN9_MEM_LATENCY_LEVEL_MASK
;
2824 wm
[7] = (val
>> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT
) &
2825 GEN9_MEM_LATENCY_LEVEL_MASK
;
2828 * If a level n (n > 1) has a 0us latency, all levels m (m >= n)
2829 * need to be disabled. We make sure to sanitize the values out
2830 * of the punit to satisfy this requirement.
2832 for (level
= 1; level
<= max_level
; level
++) {
2833 if (wm
[level
] == 0) {
2834 for (i
= level
+ 1; i
<= max_level
; i
++)
2841 * WaWmMemoryReadLatency:skl+,glk
2843 * punit doesn't take into account the read latency so we need
2844 * to add 2us to the various latency levels we retrieve from the
2845 * punit when level 0 response data us 0us.
2849 for (level
= 1; level
<= max_level
; level
++) {
2856 } else if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
)) {
2857 uint64_t sskpd
= I915_READ64(MCH_SSKPD
);
2859 wm
[0] = (sskpd
>> 56) & 0xFF;
2861 wm
[0] = sskpd
& 0xF;
2862 wm
[1] = (sskpd
>> 4) & 0xFF;
2863 wm
[2] = (sskpd
>> 12) & 0xFF;
2864 wm
[3] = (sskpd
>> 20) & 0x1FF;
2865 wm
[4] = (sskpd
>> 32) & 0x1FF;
2866 } else if (INTEL_GEN(dev_priv
) >= 6) {
2867 uint32_t sskpd
= I915_READ(MCH_SSKPD
);
2869 wm
[0] = (sskpd
>> SSKPD_WM0_SHIFT
) & SSKPD_WM_MASK
;
2870 wm
[1] = (sskpd
>> SSKPD_WM1_SHIFT
) & SSKPD_WM_MASK
;
2871 wm
[2] = (sskpd
>> SSKPD_WM2_SHIFT
) & SSKPD_WM_MASK
;
2872 wm
[3] = (sskpd
>> SSKPD_WM3_SHIFT
) & SSKPD_WM_MASK
;
2873 } else if (INTEL_GEN(dev_priv
) >= 5) {
2874 uint32_t mltr
= I915_READ(MLTR_ILK
);
2876 /* ILK primary LP0 latency is 700 ns */
2878 wm
[1] = (mltr
>> MLTR_WM1_SHIFT
) & ILK_SRLT_MASK
;
2879 wm
[2] = (mltr
>> MLTR_WM2_SHIFT
) & ILK_SRLT_MASK
;
2881 MISSING_CASE(INTEL_DEVID(dev_priv
));
2885 static void intel_fixup_spr_wm_latency(struct drm_i915_private
*dev_priv
,
2888 /* ILK sprite LP0 latency is 1300 ns */
2889 if (IS_GEN5(dev_priv
))
2893 static void intel_fixup_cur_wm_latency(struct drm_i915_private
*dev_priv
,
2896 /* ILK cursor LP0 latency is 1300 ns */
2897 if (IS_GEN5(dev_priv
))
2900 /* WaDoubleCursorLP3Latency:ivb */
2901 if (IS_IVYBRIDGE(dev_priv
))
2905 int ilk_wm_max_level(const struct drm_i915_private
*dev_priv
)
2907 /* how many WM levels are we expecting */
2908 if (INTEL_GEN(dev_priv
) >= 9)
2910 else if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
2912 else if (INTEL_GEN(dev_priv
) >= 6)
2918 static void intel_print_wm_latency(struct drm_i915_private
*dev_priv
,
2920 const uint16_t wm
[8])
2922 int level
, max_level
= ilk_wm_max_level(dev_priv
);
2924 for (level
= 0; level
<= max_level
; level
++) {
2925 unsigned int latency
= wm
[level
];
2928 DRM_ERROR("%s WM%d latency not provided\n",
2934 * - latencies are in us on gen9.
2935 * - before then, WM1+ latency values are in 0.5us units
2937 if (INTEL_GEN(dev_priv
) >= 9)
2942 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2943 name
, level
, wm
[level
],
2944 latency
/ 10, latency
% 10);
2948 static bool ilk_increase_wm_latency(struct drm_i915_private
*dev_priv
,
2949 uint16_t wm
[5], uint16_t min
)
2951 int level
, max_level
= ilk_wm_max_level(dev_priv
);
2956 wm
[0] = max(wm
[0], min
);
2957 for (level
= 1; level
<= max_level
; level
++)
2958 wm
[level
] = max_t(uint16_t, wm
[level
], DIV_ROUND_UP(min
, 5));
2963 static void snb_wm_latency_quirk(struct drm_i915_private
*dev_priv
)
2968 * The BIOS provided WM memory latency values are often
2969 * inadequate for high resolution displays. Adjust them.
2971 changed
= ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.pri_latency
, 12) |
2972 ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.spr_latency
, 12) |
2973 ilk_increase_wm_latency(dev_priv
, dev_priv
->wm
.cur_latency
, 12);
2978 DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2979 intel_print_wm_latency(dev_priv
, "Primary", dev_priv
->wm
.pri_latency
);
2980 intel_print_wm_latency(dev_priv
, "Sprite", dev_priv
->wm
.spr_latency
);
2981 intel_print_wm_latency(dev_priv
, "Cursor", dev_priv
->wm
.cur_latency
);
2984 static void ilk_setup_wm_latency(struct drm_i915_private
*dev_priv
)
2986 intel_read_wm_latency(dev_priv
, dev_priv
->wm
.pri_latency
);
2988 memcpy(dev_priv
->wm
.spr_latency
, dev_priv
->wm
.pri_latency
,
2989 sizeof(dev_priv
->wm
.pri_latency
));
2990 memcpy(dev_priv
->wm
.cur_latency
, dev_priv
->wm
.pri_latency
,
2991 sizeof(dev_priv
->wm
.pri_latency
));
2993 intel_fixup_spr_wm_latency(dev_priv
, dev_priv
->wm
.spr_latency
);
2994 intel_fixup_cur_wm_latency(dev_priv
, dev_priv
->wm
.cur_latency
);
2996 intel_print_wm_latency(dev_priv
, "Primary", dev_priv
->wm
.pri_latency
);
2997 intel_print_wm_latency(dev_priv
, "Sprite", dev_priv
->wm
.spr_latency
);
2998 intel_print_wm_latency(dev_priv
, "Cursor", dev_priv
->wm
.cur_latency
);
3000 if (IS_GEN6(dev_priv
))
3001 snb_wm_latency_quirk(dev_priv
);
3004 static void skl_setup_wm_latency(struct drm_i915_private
*dev_priv
)
3006 intel_read_wm_latency(dev_priv
, dev_priv
->wm
.skl_latency
);
3007 intel_print_wm_latency(dev_priv
, "Gen9 Plane", dev_priv
->wm
.skl_latency
);
3010 static bool ilk_validate_pipe_wm(struct drm_device
*dev
,
3011 struct intel_pipe_wm
*pipe_wm
)
3013 /* LP0 watermark maximums depend on this pipe alone */
3014 const struct intel_wm_config config
= {
3015 .num_pipes_active
= 1,
3016 .sprites_enabled
= pipe_wm
->sprites_enabled
,
3017 .sprites_scaled
= pipe_wm
->sprites_scaled
,
3019 struct ilk_wm_maximums max
;
3021 /* LP0 watermarks always use 1/2 DDB partitioning */
3022 ilk_compute_wm_maximums(dev
, 0, &config
, INTEL_DDB_PART_1_2
, &max
);
3024 /* At least LP0 must be valid */
3025 if (!ilk_validate_wm_level(0, &max
, &pipe_wm
->wm
[0])) {
3026 DRM_DEBUG_KMS("LP0 watermark invalid\n");
3033 /* Compute new watermarks for the pipe */
3034 static int ilk_compute_pipe_wm(struct intel_crtc_state
*cstate
)
3036 struct drm_atomic_state
*state
= cstate
->base
.state
;
3037 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
3038 struct intel_pipe_wm
*pipe_wm
;
3039 struct drm_device
*dev
= state
->dev
;
3040 const struct drm_i915_private
*dev_priv
= to_i915(dev
);
3041 struct intel_plane
*intel_plane
;
3042 struct intel_plane_state
*pristate
= NULL
;
3043 struct intel_plane_state
*sprstate
= NULL
;
3044 struct intel_plane_state
*curstate
= NULL
;
3045 int level
, max_level
= ilk_wm_max_level(dev_priv
), usable_level
;
3046 struct ilk_wm_maximums max
;
3048 pipe_wm
= &cstate
->wm
.ilk
.optimal
;
3050 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
3051 struct intel_plane_state
*ps
;
3053 ps
= intel_atomic_get_existing_plane_state(state
,
3058 if (intel_plane
->base
.type
== DRM_PLANE_TYPE_PRIMARY
)
3060 else if (intel_plane
->base
.type
== DRM_PLANE_TYPE_OVERLAY
)
3062 else if (intel_plane
->base
.type
== DRM_PLANE_TYPE_CURSOR
)
3066 pipe_wm
->pipe_enabled
= cstate
->base
.active
;
3068 pipe_wm
->sprites_enabled
= sprstate
->base
.visible
;
3069 pipe_wm
->sprites_scaled
= sprstate
->base
.visible
&&
3070 (drm_rect_width(&sprstate
->base
.dst
) != drm_rect_width(&sprstate
->base
.src
) >> 16 ||
3071 drm_rect_height(&sprstate
->base
.dst
) != drm_rect_height(&sprstate
->base
.src
) >> 16);
3074 usable_level
= max_level
;
3076 /* ILK/SNB: LP2+ watermarks only w/o sprites */
3077 if (INTEL_GEN(dev_priv
) <= 6 && pipe_wm
->sprites_enabled
)
3080 /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
3081 if (pipe_wm
->sprites_scaled
)
3084 ilk_compute_wm_level(dev_priv
, intel_crtc
, 0, cstate
,
3085 pristate
, sprstate
, curstate
, &pipe_wm
->raw_wm
[0]);
3087 memset(&pipe_wm
->wm
, 0, sizeof(pipe_wm
->wm
));
3088 pipe_wm
->wm
[0] = pipe_wm
->raw_wm
[0];
3090 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
3091 pipe_wm
->linetime
= hsw_compute_linetime_wm(cstate
);
3093 if (!ilk_validate_pipe_wm(dev
, pipe_wm
))
3096 ilk_compute_wm_reg_maximums(dev_priv
, 1, &max
);
3098 for (level
= 1; level
<= max_level
; level
++) {
3099 struct intel_wm_level
*wm
= &pipe_wm
->raw_wm
[level
];
3101 ilk_compute_wm_level(dev_priv
, intel_crtc
, level
, cstate
,
3102 pristate
, sprstate
, curstate
, wm
);
3105 * Disable any watermark level that exceeds the
3106 * register maximums since such watermarks are
3109 if (level
> usable_level
)
3112 if (ilk_validate_wm_level(level
, &max
, wm
))
3113 pipe_wm
->wm
[level
] = *wm
;
3115 usable_level
= level
;
3122 * Build a set of 'intermediate' watermark values that satisfy both the old
3123 * state and the new state. These can be programmed to the hardware
3126 static int ilk_compute_intermediate_wm(struct drm_device
*dev
,
3127 struct intel_crtc
*intel_crtc
,
3128 struct intel_crtc_state
*newstate
)
3130 struct intel_pipe_wm
*a
= &newstate
->wm
.ilk
.intermediate
;
3131 struct intel_pipe_wm
*b
= &intel_crtc
->wm
.active
.ilk
;
3132 int level
, max_level
= ilk_wm_max_level(to_i915(dev
));
3135 * Start with the final, target watermarks, then combine with the
3136 * currently active watermarks to get values that are safe both before
3137 * and after the vblank.
3139 *a
= newstate
->wm
.ilk
.optimal
;
3140 a
->pipe_enabled
|= b
->pipe_enabled
;
3141 a
->sprites_enabled
|= b
->sprites_enabled
;
3142 a
->sprites_scaled
|= b
->sprites_scaled
;
3144 for (level
= 0; level
<= max_level
; level
++) {
3145 struct intel_wm_level
*a_wm
= &a
->wm
[level
];
3146 const struct intel_wm_level
*b_wm
= &b
->wm
[level
];
3148 a_wm
->enable
&= b_wm
->enable
;
3149 a_wm
->pri_val
= max(a_wm
->pri_val
, b_wm
->pri_val
);
3150 a_wm
->spr_val
= max(a_wm
->spr_val
, b_wm
->spr_val
);
3151 a_wm
->cur_val
= max(a_wm
->cur_val
, b_wm
->cur_val
);
3152 a_wm
->fbc_val
= max(a_wm
->fbc_val
, b_wm
->fbc_val
);
3156 * We need to make sure that these merged watermark values are
3157 * actually a valid configuration themselves. If they're not,
3158 * there's no safe way to transition from the old state to
3159 * the new state, so we need to fail the atomic transaction.
3161 if (!ilk_validate_pipe_wm(dev
, a
))
3165 * If our intermediate WM are identical to the final WM, then we can
3166 * omit the post-vblank programming; only update if it's different.
3168 if (memcmp(a
, &newstate
->wm
.ilk
.optimal
, sizeof(*a
)) != 0)
3169 newstate
->wm
.need_postvbl_update
= true;
3175 * Merge the watermarks from all active pipes for a specific level.
3177 static void ilk_merge_wm_level(struct drm_device
*dev
,
3179 struct intel_wm_level
*ret_wm
)
3181 const struct intel_crtc
*intel_crtc
;
3183 ret_wm
->enable
= true;
3185 for_each_intel_crtc(dev
, intel_crtc
) {
3186 const struct intel_pipe_wm
*active
= &intel_crtc
->wm
.active
.ilk
;
3187 const struct intel_wm_level
*wm
= &active
->wm
[level
];
3189 if (!active
->pipe_enabled
)
3193 * The watermark values may have been used in the past,
3194 * so we must maintain them in the registers for some
3195 * time even if the level is now disabled.
3198 ret_wm
->enable
= false;
3200 ret_wm
->pri_val
= max(ret_wm
->pri_val
, wm
->pri_val
);
3201 ret_wm
->spr_val
= max(ret_wm
->spr_val
, wm
->spr_val
);
3202 ret_wm
->cur_val
= max(ret_wm
->cur_val
, wm
->cur_val
);
3203 ret_wm
->fbc_val
= max(ret_wm
->fbc_val
, wm
->fbc_val
);
3208 * Merge all low power watermarks for all active pipes.
3210 static void ilk_wm_merge(struct drm_device
*dev
,
3211 const struct intel_wm_config
*config
,
3212 const struct ilk_wm_maximums
*max
,
3213 struct intel_pipe_wm
*merged
)
3215 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3216 int level
, max_level
= ilk_wm_max_level(dev_priv
);
3217 int last_enabled_level
= max_level
;
3219 /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
3220 if ((INTEL_GEN(dev_priv
) <= 6 || IS_IVYBRIDGE(dev_priv
)) &&
3221 config
->num_pipes_active
> 1)
3222 last_enabled_level
= 0;
3224 /* ILK: FBC WM must be disabled always */
3225 merged
->fbc_wm_enabled
= INTEL_GEN(dev_priv
) >= 6;
3227 /* merge each WM1+ level */
3228 for (level
= 1; level
<= max_level
; level
++) {
3229 struct intel_wm_level
*wm
= &merged
->wm
[level
];
3231 ilk_merge_wm_level(dev
, level
, wm
);
3233 if (level
> last_enabled_level
)
3235 else if (!ilk_validate_wm_level(level
, max
, wm
))
3236 /* make sure all following levels get disabled */
3237 last_enabled_level
= level
- 1;
3240 * The spec says it is preferred to disable
3241 * FBC WMs instead of disabling a WM level.
3243 if (wm
->fbc_val
> max
->fbc
) {
3245 merged
->fbc_wm_enabled
= false;
3250 /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
3252 * FIXME this is racy. FBC might get enabled later.
3253 * What we should check here is whether FBC can be
3254 * enabled sometime later.
3256 if (IS_GEN5(dev_priv
) && !merged
->fbc_wm_enabled
&&
3257 intel_fbc_is_active(dev_priv
)) {
3258 for (level
= 2; level
<= max_level
; level
++) {
3259 struct intel_wm_level
*wm
= &merged
->wm
[level
];
3266 static int ilk_wm_lp_to_level(int wm_lp
, const struct intel_pipe_wm
*pipe_wm
)
3268 /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
3269 return wm_lp
+ (wm_lp
>= 2 && pipe_wm
->wm
[4].enable
);
3272 /* The value we need to program into the WM_LPx latency field */
3273 static unsigned int ilk_wm_lp_latency(struct drm_device
*dev
, int level
)
3275 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3277 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
3280 return dev_priv
->wm
.pri_latency
[level
];
3283 static void ilk_compute_wm_results(struct drm_device
*dev
,
3284 const struct intel_pipe_wm
*merged
,
3285 enum intel_ddb_partitioning partitioning
,
3286 struct ilk_wm_values
*results
)
3288 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3289 struct intel_crtc
*intel_crtc
;
3292 results
->enable_fbc_wm
= merged
->fbc_wm_enabled
;
3293 results
->partitioning
= partitioning
;
3295 /* LP1+ register values */
3296 for (wm_lp
= 1; wm_lp
<= 3; wm_lp
++) {
3297 const struct intel_wm_level
*r
;
3299 level
= ilk_wm_lp_to_level(wm_lp
, merged
);
3301 r
= &merged
->wm
[level
];
3304 * Maintain the watermark values even if the level is
3305 * disabled. Doing otherwise could cause underruns.
3307 results
->wm_lp
[wm_lp
- 1] =
3308 (ilk_wm_lp_latency(dev
, level
) << WM1_LP_LATENCY_SHIFT
) |
3309 (r
->pri_val
<< WM1_LP_SR_SHIFT
) |
3313 results
->wm_lp
[wm_lp
- 1] |= WM1_LP_SR_EN
;
3315 if (INTEL_GEN(dev_priv
) >= 8)
3316 results
->wm_lp
[wm_lp
- 1] |=
3317 r
->fbc_val
<< WM1_LP_FBC_SHIFT_BDW
;
3319 results
->wm_lp
[wm_lp
- 1] |=
3320 r
->fbc_val
<< WM1_LP_FBC_SHIFT
;
3323 * Always set WM1S_LP_EN when spr_val != 0, even if the
3324 * level is disabled. Doing otherwise could cause underruns.
3326 if (INTEL_GEN(dev_priv
) <= 6 && r
->spr_val
) {
3327 WARN_ON(wm_lp
!= 1);
3328 results
->wm_lp_spr
[wm_lp
- 1] = WM1S_LP_EN
| r
->spr_val
;
3330 results
->wm_lp_spr
[wm_lp
- 1] = r
->spr_val
;
3333 /* LP0 register values */
3334 for_each_intel_crtc(dev
, intel_crtc
) {
3335 enum pipe pipe
= intel_crtc
->pipe
;
3336 const struct intel_wm_level
*r
=
3337 &intel_crtc
->wm
.active
.ilk
.wm
[0];
3339 if (WARN_ON(!r
->enable
))
3342 results
->wm_linetime
[pipe
] = intel_crtc
->wm
.active
.ilk
.linetime
;
3344 results
->wm_pipe
[pipe
] =
3345 (r
->pri_val
<< WM0_PIPE_PLANE_SHIFT
) |
3346 (r
->spr_val
<< WM0_PIPE_SPRITE_SHIFT
) |
3351 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
3352 * case both are at the same level. Prefer r1 in case they're the same. */
3353 static struct intel_pipe_wm
*ilk_find_best_result(struct drm_device
*dev
,
3354 struct intel_pipe_wm
*r1
,
3355 struct intel_pipe_wm
*r2
)
3357 int level
, max_level
= ilk_wm_max_level(to_i915(dev
));
3358 int level1
= 0, level2
= 0;
3360 for (level
= 1; level
<= max_level
; level
++) {
3361 if (r1
->wm
[level
].enable
)
3363 if (r2
->wm
[level
].enable
)
3367 if (level1
== level2
) {
3368 if (r2
->fbc_wm_enabled
&& !r1
->fbc_wm_enabled
)
3372 } else if (level1
> level2
) {
3379 /* dirty bits used to track which watermarks need changes */
3380 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
3381 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
3382 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
3383 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
3384 #define WM_DIRTY_FBC (1 << 24)
3385 #define WM_DIRTY_DDB (1 << 25)
3387 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private
*dev_priv
,
3388 const struct ilk_wm_values
*old
,
3389 const struct ilk_wm_values
*new)
3391 unsigned int dirty
= 0;
3395 for_each_pipe(dev_priv
, pipe
) {
3396 if (old
->wm_linetime
[pipe
] != new->wm_linetime
[pipe
]) {
3397 dirty
|= WM_DIRTY_LINETIME(pipe
);
3398 /* Must disable LP1+ watermarks too */
3399 dirty
|= WM_DIRTY_LP_ALL
;
3402 if (old
->wm_pipe
[pipe
] != new->wm_pipe
[pipe
]) {
3403 dirty
|= WM_DIRTY_PIPE(pipe
);
3404 /* Must disable LP1+ watermarks too */
3405 dirty
|= WM_DIRTY_LP_ALL
;
3409 if (old
->enable_fbc_wm
!= new->enable_fbc_wm
) {
3410 dirty
|= WM_DIRTY_FBC
;
3411 /* Must disable LP1+ watermarks too */
3412 dirty
|= WM_DIRTY_LP_ALL
;
3415 if (old
->partitioning
!= new->partitioning
) {
3416 dirty
|= WM_DIRTY_DDB
;
3417 /* Must disable LP1+ watermarks too */
3418 dirty
|= WM_DIRTY_LP_ALL
;
3421 /* LP1+ watermarks already deemed dirty, no need to continue */
3422 if (dirty
& WM_DIRTY_LP_ALL
)
3425 /* Find the lowest numbered LP1+ watermark in need of an update... */
3426 for (wm_lp
= 1; wm_lp
<= 3; wm_lp
++) {
3427 if (old
->wm_lp
[wm_lp
- 1] != new->wm_lp
[wm_lp
- 1] ||
3428 old
->wm_lp_spr
[wm_lp
- 1] != new->wm_lp_spr
[wm_lp
- 1])
3432 /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
3433 for (; wm_lp
<= 3; wm_lp
++)
3434 dirty
|= WM_DIRTY_LP(wm_lp
);
3439 static bool _ilk_disable_lp_wm(struct drm_i915_private
*dev_priv
,
3442 struct ilk_wm_values
*previous
= &dev_priv
->wm
.hw
;
3443 bool changed
= false;
3445 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp
[2] & WM1_LP_SR_EN
) {
3446 previous
->wm_lp
[2] &= ~WM1_LP_SR_EN
;
3447 I915_WRITE(WM3_LP_ILK
, previous
->wm_lp
[2]);
3450 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp
[1] & WM1_LP_SR_EN
) {
3451 previous
->wm_lp
[1] &= ~WM1_LP_SR_EN
;
3452 I915_WRITE(WM2_LP_ILK
, previous
->wm_lp
[1]);
3455 if (dirty
& WM_DIRTY_LP(1) && previous
->wm_lp
[0] & WM1_LP_SR_EN
) {
3456 previous
->wm_lp
[0] &= ~WM1_LP_SR_EN
;
3457 I915_WRITE(WM1_LP_ILK
, previous
->wm_lp
[0]);
3462 * Don't touch WM1S_LP_EN here.
3463 * Doing so could cause underruns.
3470 * The spec says we shouldn't write when we don't need, because every write
3471 * causes WMs to be re-evaluated, expending some power.
3473 static void ilk_write_wm_values(struct drm_i915_private
*dev_priv
,
3474 struct ilk_wm_values
*results
)
3476 struct ilk_wm_values
*previous
= &dev_priv
->wm
.hw
;
3480 dirty
= ilk_compute_wm_dirty(dev_priv
, previous
, results
);
3484 _ilk_disable_lp_wm(dev_priv
, dirty
);
3486 if (dirty
& WM_DIRTY_PIPE(PIPE_A
))
3487 I915_WRITE(WM0_PIPEA_ILK
, results
->wm_pipe
[0]);
3488 if (dirty
& WM_DIRTY_PIPE(PIPE_B
))
3489 I915_WRITE(WM0_PIPEB_ILK
, results
->wm_pipe
[1]);
3490 if (dirty
& WM_DIRTY_PIPE(PIPE_C
))
3491 I915_WRITE(WM0_PIPEC_IVB
, results
->wm_pipe
[2]);
3493 if (dirty
& WM_DIRTY_LINETIME(PIPE_A
))
3494 I915_WRITE(PIPE_WM_LINETIME(PIPE_A
), results
->wm_linetime
[0]);
3495 if (dirty
& WM_DIRTY_LINETIME(PIPE_B
))
3496 I915_WRITE(PIPE_WM_LINETIME(PIPE_B
), results
->wm_linetime
[1]);
3497 if (dirty
& WM_DIRTY_LINETIME(PIPE_C
))
3498 I915_WRITE(PIPE_WM_LINETIME(PIPE_C
), results
->wm_linetime
[2]);
3500 if (dirty
& WM_DIRTY_DDB
) {
3501 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
)) {
3502 val
= I915_READ(WM_MISC
);
3503 if (results
->partitioning
== INTEL_DDB_PART_1_2
)
3504 val
&= ~WM_MISC_DATA_PARTITION_5_6
;
3506 val
|= WM_MISC_DATA_PARTITION_5_6
;
3507 I915_WRITE(WM_MISC
, val
);
3509 val
= I915_READ(DISP_ARB_CTL2
);
3510 if (results
->partitioning
== INTEL_DDB_PART_1_2
)
3511 val
&= ~DISP_DATA_PARTITION_5_6
;
3513 val
|= DISP_DATA_PARTITION_5_6
;
3514 I915_WRITE(DISP_ARB_CTL2
, val
);
3518 if (dirty
& WM_DIRTY_FBC
) {
3519 val
= I915_READ(DISP_ARB_CTL
);
3520 if (results
->enable_fbc_wm
)
3521 val
&= ~DISP_FBC_WM_DIS
;
3523 val
|= DISP_FBC_WM_DIS
;
3524 I915_WRITE(DISP_ARB_CTL
, val
);
3527 if (dirty
& WM_DIRTY_LP(1) &&
3528 previous
->wm_lp_spr
[0] != results
->wm_lp_spr
[0])
3529 I915_WRITE(WM1S_LP_ILK
, results
->wm_lp_spr
[0]);
3531 if (INTEL_GEN(dev_priv
) >= 7) {
3532 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp_spr
[1] != results
->wm_lp_spr
[1])
3533 I915_WRITE(WM2S_LP_IVB
, results
->wm_lp_spr
[1]);
3534 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp_spr
[2] != results
->wm_lp_spr
[2])
3535 I915_WRITE(WM3S_LP_IVB
, results
->wm_lp_spr
[2]);
3538 if (dirty
& WM_DIRTY_LP(1) && previous
->wm_lp
[0] != results
->wm_lp
[0])
3539 I915_WRITE(WM1_LP_ILK
, results
->wm_lp
[0]);
3540 if (dirty
& WM_DIRTY_LP(2) && previous
->wm_lp
[1] != results
->wm_lp
[1])
3541 I915_WRITE(WM2_LP_ILK
, results
->wm_lp
[1]);
3542 if (dirty
& WM_DIRTY_LP(3) && previous
->wm_lp
[2] != results
->wm_lp
[2])
3543 I915_WRITE(WM3_LP_ILK
, results
->wm_lp
[2]);
3545 dev_priv
->wm
.hw
= *results
;
3548 bool ilk_disable_lp_wm(struct drm_device
*dev
)
3550 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3552 return _ilk_disable_lp_wm(dev_priv
, WM_DIRTY_LP_ALL
);
3556 * FIXME: We still don't have the proper code detect if we need to apply the WA,
3557 * so assume we'll always need it in order to avoid underruns.
3559 static bool skl_needs_memory_bw_wa(struct intel_atomic_state
*state
)
3561 struct drm_i915_private
*dev_priv
= to_i915(state
->base
.dev
);
3563 if (IS_GEN9_BC(dev_priv
) || IS_BROXTON(dev_priv
))
3570 intel_has_sagv(struct drm_i915_private
*dev_priv
)
3572 if (IS_KABYLAKE(dev_priv
) || IS_COFFEELAKE(dev_priv
) ||
3573 IS_CANNONLAKE(dev_priv
))
3576 if (IS_SKYLAKE(dev_priv
) &&
3577 dev_priv
->sagv_status
!= I915_SAGV_NOT_CONTROLLED
)
3584 * SAGV dynamically adjusts the system agent voltage and clock frequencies
3585 * depending on power and performance requirements. The display engine access
3586 * to system memory is blocked during the adjustment time. Because of the
3587 * blocking time, having this enabled can cause full system hangs and/or pipe
3588 * underruns if we don't meet all of the following requirements:
3590 * - <= 1 pipe enabled
3591 * - All planes can enable watermarks for latencies >= SAGV engine block time
3592 * - We're not using an interlaced display configuration
3595 intel_enable_sagv(struct drm_i915_private
*dev_priv
)
3599 if (!intel_has_sagv(dev_priv
))
3602 if (dev_priv
->sagv_status
== I915_SAGV_ENABLED
)
3605 DRM_DEBUG_KMS("Enabling the SAGV\n");
3606 mutex_lock(&dev_priv
->rps
.hw_lock
);
3608 ret
= sandybridge_pcode_write(dev_priv
, GEN9_PCODE_SAGV_CONTROL
,
3611 /* We don't need to wait for the SAGV when enabling */
3612 mutex_unlock(&dev_priv
->rps
.hw_lock
);
3615 * Some skl systems, pre-release machines in particular,
3616 * don't actually have an SAGV.
3618 if (IS_SKYLAKE(dev_priv
) && ret
== -ENXIO
) {
3619 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
3620 dev_priv
->sagv_status
= I915_SAGV_NOT_CONTROLLED
;
3622 } else if (ret
< 0) {
3623 DRM_ERROR("Failed to enable the SAGV\n");
3627 dev_priv
->sagv_status
= I915_SAGV_ENABLED
;
3632 intel_disable_sagv(struct drm_i915_private
*dev_priv
)
3636 if (!intel_has_sagv(dev_priv
))
3639 if (dev_priv
->sagv_status
== I915_SAGV_DISABLED
)
3642 DRM_DEBUG_KMS("Disabling the SAGV\n");
3643 mutex_lock(&dev_priv
->rps
.hw_lock
);
3645 /* bspec says to keep retrying for at least 1 ms */
3646 ret
= skl_pcode_request(dev_priv
, GEN9_PCODE_SAGV_CONTROL
,
3648 GEN9_SAGV_IS_DISABLED
, GEN9_SAGV_IS_DISABLED
,
3650 mutex_unlock(&dev_priv
->rps
.hw_lock
);
3653 * Some skl systems, pre-release machines in particular,
3654 * don't actually have an SAGV.
3656 if (IS_SKYLAKE(dev_priv
) && ret
== -ENXIO
) {
3657 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
3658 dev_priv
->sagv_status
= I915_SAGV_NOT_CONTROLLED
;
3660 } else if (ret
< 0) {
3661 DRM_ERROR("Failed to disable the SAGV (%d)\n", ret
);
3665 dev_priv
->sagv_status
= I915_SAGV_DISABLED
;
3669 bool intel_can_enable_sagv(struct drm_atomic_state
*state
)
3671 struct drm_device
*dev
= state
->dev
;
3672 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3673 struct intel_atomic_state
*intel_state
= to_intel_atomic_state(state
);
3674 struct intel_crtc
*crtc
;
3675 struct intel_plane
*plane
;
3676 struct intel_crtc_state
*cstate
;
3679 int sagv_block_time_us
= IS_GEN9(dev_priv
) ? 30 : 20;
3681 if (!intel_has_sagv(dev_priv
))
3685 * SKL+ workaround: bspec recommends we disable the SAGV when we have
3686 * more then one pipe enabled
3688 * If there are no active CRTCs, no additional checks need be performed
3690 if (hweight32(intel_state
->active_crtcs
) == 0)
3692 else if (hweight32(intel_state
->active_crtcs
) > 1)
3695 /* Since we're now guaranteed to only have one active CRTC... */
3696 pipe
= ffs(intel_state
->active_crtcs
) - 1;
3697 crtc
= intel_get_crtc_for_pipe(dev_priv
, pipe
);
3698 cstate
= to_intel_crtc_state(crtc
->base
.state
);
3700 if (crtc
->base
.state
->adjusted_mode
.flags
& DRM_MODE_FLAG_INTERLACE
)
3703 for_each_intel_plane_on_crtc(dev
, crtc
, plane
) {
3704 struct skl_plane_wm
*wm
=
3705 &cstate
->wm
.skl
.optimal
.planes
[plane
->id
];
3707 /* Skip this plane if it's not enabled */
3708 if (!wm
->wm
[0].plane_en
)
3711 /* Find the highest enabled wm level for this plane */
3712 for (level
= ilk_wm_max_level(dev_priv
);
3713 !wm
->wm
[level
].plane_en
; --level
)
3716 latency
= dev_priv
->wm
.skl_latency
[level
];
3718 if (skl_needs_memory_bw_wa(intel_state
) &&
3719 plane
->base
.state
->fb
->modifier
==
3720 I915_FORMAT_MOD_X_TILED
)
3724 * If any of the planes on this pipe don't enable wm levels that
3725 * incur memory latencies higher than sagv_block_time_us we
3726 * can't enable the SAGV.
3728 if (latency
< sagv_block_time_us
)
3736 skl_ddb_get_pipe_allocation_limits(struct drm_device
*dev
,
3737 const struct intel_crtc_state
*cstate
,
3738 struct skl_ddb_entry
*alloc
, /* out */
3739 int *num_active
/* out */)
3741 struct drm_atomic_state
*state
= cstate
->base
.state
;
3742 struct intel_atomic_state
*intel_state
= to_intel_atomic_state(state
);
3743 struct drm_i915_private
*dev_priv
= to_i915(dev
);
3744 struct drm_crtc
*for_crtc
= cstate
->base
.crtc
;
3745 unsigned int pipe_size
, ddb_size
;
3746 int nth_active_pipe
;
3748 if (WARN_ON(!state
) || !cstate
->base
.active
) {
3751 *num_active
= hweight32(dev_priv
->active_crtcs
);
3755 if (intel_state
->active_pipe_changes
)
3756 *num_active
= hweight32(intel_state
->active_crtcs
);
3758 *num_active
= hweight32(dev_priv
->active_crtcs
);
3760 ddb_size
= INTEL_INFO(dev_priv
)->ddb_size
;
3761 WARN_ON(ddb_size
== 0);
3763 ddb_size
-= 4; /* 4 blocks for bypass path allocation */
3766 * If the state doesn't change the active CRTC's, then there's
3767 * no need to recalculate; the existing pipe allocation limits
3768 * should remain unchanged. Note that we're safe from racing
3769 * commits since any racing commit that changes the active CRTC
3770 * list would need to grab _all_ crtc locks, including the one
3771 * we currently hold.
3773 if (!intel_state
->active_pipe_changes
) {
3775 * alloc may be cleared by clear_intel_crtc_state,
3776 * copy from old state to be sure
3778 *alloc
= to_intel_crtc_state(for_crtc
->state
)->wm
.skl
.ddb
;
3782 nth_active_pipe
= hweight32(intel_state
->active_crtcs
&
3783 (drm_crtc_mask(for_crtc
) - 1));
3784 pipe_size
= ddb_size
/ hweight32(intel_state
->active_crtcs
);
3785 alloc
->start
= nth_active_pipe
* ddb_size
/ *num_active
;
3786 alloc
->end
= alloc
->start
+ pipe_size
;
3789 static unsigned int skl_cursor_allocation(int num_active
)
3791 if (num_active
== 1)
3797 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry
*entry
, u32 reg
)
3799 entry
->start
= reg
& 0x3ff;
3800 entry
->end
= (reg
>> 16) & 0x3ff;
3805 void skl_ddb_get_hw_state(struct drm_i915_private
*dev_priv
,
3806 struct skl_ddb_allocation
*ddb
/* out */)
3808 struct intel_crtc
*crtc
;
3810 memset(ddb
, 0, sizeof(*ddb
));
3812 for_each_intel_crtc(&dev_priv
->drm
, crtc
) {
3813 enum intel_display_power_domain power_domain
;
3814 enum plane_id plane_id
;
3815 enum pipe pipe
= crtc
->pipe
;
3817 power_domain
= POWER_DOMAIN_PIPE(pipe
);
3818 if (!intel_display_power_get_if_enabled(dev_priv
, power_domain
))
3821 for_each_plane_id_on_crtc(crtc
, plane_id
) {
3824 if (plane_id
!= PLANE_CURSOR
)
3825 val
= I915_READ(PLANE_BUF_CFG(pipe
, plane_id
));
3827 val
= I915_READ(CUR_BUF_CFG(pipe
));
3829 skl_ddb_entry_init_from_hw(&ddb
->plane
[pipe
][plane_id
], val
);
3832 intel_display_power_put(dev_priv
, power_domain
);
3837 * Determines the downscale amount of a plane for the purposes of watermark calculations.
3838 * The bspec defines downscale amount as:
3841 * Horizontal down scale amount = maximum[1, Horizontal source size /
3842 * Horizontal destination size]
3843 * Vertical down scale amount = maximum[1, Vertical source size /
3844 * Vertical destination size]
3845 * Total down scale amount = Horizontal down scale amount *
3846 * Vertical down scale amount
3849 * Return value is provided in 16.16 fixed point form to retain fractional part.
3850 * Caller should take care of dividing & rounding off the value.
3852 static uint_fixed_16_16_t
3853 skl_plane_downscale_amount(const struct intel_crtc_state
*cstate
,
3854 const struct intel_plane_state
*pstate
)
3856 struct intel_plane
*plane
= to_intel_plane(pstate
->base
.plane
);
3857 uint32_t src_w
, src_h
, dst_w
, dst_h
;
3858 uint_fixed_16_16_t fp_w_ratio
, fp_h_ratio
;
3859 uint_fixed_16_16_t downscale_h
, downscale_w
;
3861 if (WARN_ON(!intel_wm_plane_visible(cstate
, pstate
)))
3862 return u32_to_fixed16(0);
3864 /* n.b., src is 16.16 fixed point, dst is whole integer */
3865 if (plane
->id
== PLANE_CURSOR
) {
3867 * Cursors only support 0/180 degree rotation,
3868 * hence no need to account for rotation here.
3870 src_w
= pstate
->base
.src_w
>> 16;
3871 src_h
= pstate
->base
.src_h
>> 16;
3872 dst_w
= pstate
->base
.crtc_w
;
3873 dst_h
= pstate
->base
.crtc_h
;
3876 * Src coordinates are already rotated by 270 degrees for
3877 * the 90/270 degree plane rotation cases (to match the
3878 * GTT mapping), hence no need to account for rotation here.
3880 src_w
= drm_rect_width(&pstate
->base
.src
) >> 16;
3881 src_h
= drm_rect_height(&pstate
->base
.src
) >> 16;
3882 dst_w
= drm_rect_width(&pstate
->base
.dst
);
3883 dst_h
= drm_rect_height(&pstate
->base
.dst
);
3886 fp_w_ratio
= div_fixed16(src_w
, dst_w
);
3887 fp_h_ratio
= div_fixed16(src_h
, dst_h
);
3888 downscale_w
= max_fixed16(fp_w_ratio
, u32_to_fixed16(1));
3889 downscale_h
= max_fixed16(fp_h_ratio
, u32_to_fixed16(1));
3891 return mul_fixed16(downscale_w
, downscale_h
);
3894 static uint_fixed_16_16_t
3895 skl_pipe_downscale_amount(const struct intel_crtc_state
*crtc_state
)
3897 uint_fixed_16_16_t pipe_downscale
= u32_to_fixed16(1);
3899 if (!crtc_state
->base
.enable
)
3900 return pipe_downscale
;
3902 if (crtc_state
->pch_pfit
.enabled
) {
3903 uint32_t src_w
, src_h
, dst_w
, dst_h
;
3904 uint32_t pfit_size
= crtc_state
->pch_pfit
.size
;
3905 uint_fixed_16_16_t fp_w_ratio
, fp_h_ratio
;
3906 uint_fixed_16_16_t downscale_h
, downscale_w
;
3908 src_w
= crtc_state
->pipe_src_w
;
3909 src_h
= crtc_state
->pipe_src_h
;
3910 dst_w
= pfit_size
>> 16;
3911 dst_h
= pfit_size
& 0xffff;
3913 if (!dst_w
|| !dst_h
)
3914 return pipe_downscale
;
3916 fp_w_ratio
= div_fixed16(src_w
, dst_w
);
3917 fp_h_ratio
= div_fixed16(src_h
, dst_h
);
3918 downscale_w
= max_fixed16(fp_w_ratio
, u32_to_fixed16(1));
3919 downscale_h
= max_fixed16(fp_h_ratio
, u32_to_fixed16(1));
3921 pipe_downscale
= mul_fixed16(downscale_w
, downscale_h
);
3924 return pipe_downscale
;
3927 int skl_check_pipe_max_pixel_rate(struct intel_crtc
*intel_crtc
,
3928 struct intel_crtc_state
*cstate
)
3930 struct drm_crtc_state
*crtc_state
= &cstate
->base
;
3931 struct drm_atomic_state
*state
= crtc_state
->state
;
3932 struct drm_plane
*plane
;
3933 const struct drm_plane_state
*pstate
;
3934 struct intel_plane_state
*intel_pstate
;
3935 int crtc_clock
, dotclk
;
3936 uint32_t pipe_max_pixel_rate
;
3937 uint_fixed_16_16_t pipe_downscale
;
3938 uint_fixed_16_16_t max_downscale
= u32_to_fixed16(1);
3940 if (!cstate
->base
.enable
)
3943 drm_atomic_crtc_state_for_each_plane_state(plane
, pstate
, crtc_state
) {
3944 uint_fixed_16_16_t plane_downscale
;
3945 uint_fixed_16_16_t fp_9_div_8
= div_fixed16(9, 8);
3948 if (!intel_wm_plane_visible(cstate
,
3949 to_intel_plane_state(pstate
)))
3952 if (WARN_ON(!pstate
->fb
))
3955 intel_pstate
= to_intel_plane_state(pstate
);
3956 plane_downscale
= skl_plane_downscale_amount(cstate
,
3958 bpp
= pstate
->fb
->format
->cpp
[0] * 8;
3960 plane_downscale
= mul_fixed16(plane_downscale
,
3963 max_downscale
= max_fixed16(plane_downscale
, max_downscale
);
3965 pipe_downscale
= skl_pipe_downscale_amount(cstate
);
3967 pipe_downscale
= mul_fixed16(pipe_downscale
, max_downscale
);
3969 crtc_clock
= crtc_state
->adjusted_mode
.crtc_clock
;
3970 dotclk
= to_intel_atomic_state(state
)->cdclk
.logical
.cdclk
;
3972 if (IS_GEMINILAKE(to_i915(intel_crtc
->base
.dev
)))
3975 pipe_max_pixel_rate
= div_round_up_u32_fixed16(dotclk
, pipe_downscale
);
3977 if (pipe_max_pixel_rate
< crtc_clock
) {
3978 DRM_DEBUG_KMS("Max supported pixel clock with scaling exceeded\n");
3986 skl_plane_relative_data_rate(const struct intel_crtc_state
*cstate
,
3987 const struct drm_plane_state
*pstate
,
3990 struct intel_plane
*plane
= to_intel_plane(pstate
->plane
);
3991 struct intel_plane_state
*intel_pstate
= to_intel_plane_state(pstate
);
3993 uint32_t width
= 0, height
= 0;
3994 struct drm_framebuffer
*fb
;
3996 uint_fixed_16_16_t down_scale_amount
;
3998 if (!intel_pstate
->base
.visible
)
4002 format
= fb
->format
->format
;
4004 if (plane
->id
== PLANE_CURSOR
)
4006 if (y
&& format
!= DRM_FORMAT_NV12
)
4010 * Src coordinates are already rotated by 270 degrees for
4011 * the 90/270 degree plane rotation cases (to match the
4012 * GTT mapping), hence no need to account for rotation here.
4014 width
= drm_rect_width(&intel_pstate
->base
.src
) >> 16;
4015 height
= drm_rect_height(&intel_pstate
->base
.src
) >> 16;
4017 /* for planar format */
4018 if (format
== DRM_FORMAT_NV12
) {
4019 if (y
) /* y-plane data rate */
4020 data_rate
= width
* height
*
4022 else /* uv-plane data rate */
4023 data_rate
= (width
/ 2) * (height
/ 2) *
4026 /* for packed formats */
4027 data_rate
= width
* height
* fb
->format
->cpp
[0];
4030 down_scale_amount
= skl_plane_downscale_amount(cstate
, intel_pstate
);
4032 return mul_round_up_u32_fixed16(data_rate
, down_scale_amount
);
4036 * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
4037 * a 8192x4096@32bpp framebuffer:
4038 * 3 * 4096 * 8192 * 4 < 2^32
4041 skl_get_total_relative_data_rate(struct intel_crtc_state
*intel_cstate
,
4042 unsigned *plane_data_rate
,
4043 unsigned *plane_y_data_rate
)
4045 struct drm_crtc_state
*cstate
= &intel_cstate
->base
;
4046 struct drm_atomic_state
*state
= cstate
->state
;
4047 struct drm_plane
*plane
;
4048 const struct drm_plane_state
*pstate
;
4049 unsigned int total_data_rate
= 0;
4051 if (WARN_ON(!state
))
4054 /* Calculate and cache data rate for each plane */
4055 drm_atomic_crtc_state_for_each_plane_state(plane
, pstate
, cstate
) {
4056 enum plane_id plane_id
= to_intel_plane(plane
)->id
;
4060 rate
= skl_plane_relative_data_rate(intel_cstate
,
4062 plane_data_rate
[plane_id
] = rate
;
4064 total_data_rate
+= rate
;
4067 rate
= skl_plane_relative_data_rate(intel_cstate
,
4069 plane_y_data_rate
[plane_id
] = rate
;
4071 total_data_rate
+= rate
;
4074 return total_data_rate
;
4078 skl_ddb_min_alloc(const struct drm_plane_state
*pstate
,
4081 struct drm_framebuffer
*fb
= pstate
->fb
;
4082 struct intel_plane_state
*intel_pstate
= to_intel_plane_state(pstate
);
4083 uint32_t src_w
, src_h
;
4084 uint32_t min_scanlines
= 8;
4090 /* For packed formats, no y-plane, return 0 */
4091 if (y
&& fb
->format
->format
!= DRM_FORMAT_NV12
)
4094 /* For Non Y-tile return 8-blocks */
4095 if (fb
->modifier
!= I915_FORMAT_MOD_Y_TILED
&&
4096 fb
->modifier
!= I915_FORMAT_MOD_Yf_TILED
&&
4097 fb
->modifier
!= I915_FORMAT_MOD_Y_TILED_CCS
&&
4098 fb
->modifier
!= I915_FORMAT_MOD_Yf_TILED_CCS
)
4102 * Src coordinates are already rotated by 270 degrees for
4103 * the 90/270 degree plane rotation cases (to match the
4104 * GTT mapping), hence no need to account for rotation here.
4106 src_w
= drm_rect_width(&intel_pstate
->base
.src
) >> 16;
4107 src_h
= drm_rect_height(&intel_pstate
->base
.src
) >> 16;
4109 /* Halve UV plane width and height for NV12 */
4110 if (fb
->format
->format
== DRM_FORMAT_NV12
&& !y
) {
4115 if (fb
->format
->format
== DRM_FORMAT_NV12
&& !y
)
4116 plane_bpp
= fb
->format
->cpp
[1];
4118 plane_bpp
= fb
->format
->cpp
[0];
4120 if (drm_rotation_90_or_270(pstate
->rotation
)) {
4121 switch (plane_bpp
) {
4135 WARN(1, "Unsupported pixel depth %u for rotation",
4141 return DIV_ROUND_UP((4 * src_w
* plane_bpp
), 512) * min_scanlines
/4 + 3;
4145 skl_ddb_calc_min(const struct intel_crtc_state
*cstate
, int num_active
,
4146 uint16_t *minimum
, uint16_t *y_minimum
)
4148 const struct drm_plane_state
*pstate
;
4149 struct drm_plane
*plane
;
4151 drm_atomic_crtc_state_for_each_plane_state(plane
, pstate
, &cstate
->base
) {
4152 enum plane_id plane_id
= to_intel_plane(plane
)->id
;
4154 if (plane_id
== PLANE_CURSOR
)
4157 if (!pstate
->visible
)
4160 minimum
[plane_id
] = skl_ddb_min_alloc(pstate
, 0);
4161 y_minimum
[plane_id
] = skl_ddb_min_alloc(pstate
, 1);
4164 minimum
[PLANE_CURSOR
] = skl_cursor_allocation(num_active
);
4168 skl_allocate_pipe_ddb(struct intel_crtc_state
*cstate
,
4169 struct skl_ddb_allocation
*ddb
/* out */)
4171 struct drm_atomic_state
*state
= cstate
->base
.state
;
4172 struct drm_crtc
*crtc
= cstate
->base
.crtc
;
4173 struct drm_device
*dev
= crtc
->dev
;
4174 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
4175 enum pipe pipe
= intel_crtc
->pipe
;
4176 struct skl_ddb_entry
*alloc
= &cstate
->wm
.skl
.ddb
;
4177 uint16_t alloc_size
, start
;
4178 uint16_t minimum
[I915_MAX_PLANES
] = {};
4179 uint16_t y_minimum
[I915_MAX_PLANES
] = {};
4180 unsigned int total_data_rate
;
4181 enum plane_id plane_id
;
4183 unsigned plane_data_rate
[I915_MAX_PLANES
] = {};
4184 unsigned plane_y_data_rate
[I915_MAX_PLANES
] = {};
4185 uint16_t total_min_blocks
= 0;
4187 /* Clear the partitioning for disabled planes. */
4188 memset(ddb
->plane
[pipe
], 0, sizeof(ddb
->plane
[pipe
]));
4189 memset(ddb
->y_plane
[pipe
], 0, sizeof(ddb
->y_plane
[pipe
]));
4191 if (WARN_ON(!state
))
4194 if (!cstate
->base
.active
) {
4195 alloc
->start
= alloc
->end
= 0;
4199 skl_ddb_get_pipe_allocation_limits(dev
, cstate
, alloc
, &num_active
);
4200 alloc_size
= skl_ddb_entry_size(alloc
);
4201 if (alloc_size
== 0)
4204 skl_ddb_calc_min(cstate
, num_active
, minimum
, y_minimum
);
4207 * 1. Allocate the mininum required blocks for each active plane
4208 * and allocate the cursor, it doesn't require extra allocation
4209 * proportional to the data rate.
4212 for_each_plane_id_on_crtc(intel_crtc
, plane_id
) {
4213 total_min_blocks
+= minimum
[plane_id
];
4214 total_min_blocks
+= y_minimum
[plane_id
];
4217 if (total_min_blocks
> alloc_size
) {
4218 DRM_DEBUG_KMS("Requested display configuration exceeds system DDB limitations");
4219 DRM_DEBUG_KMS("minimum required %d/%d\n", total_min_blocks
,
4224 alloc_size
-= total_min_blocks
;
4225 ddb
->plane
[pipe
][PLANE_CURSOR
].start
= alloc
->end
- minimum
[PLANE_CURSOR
];
4226 ddb
->plane
[pipe
][PLANE_CURSOR
].end
= alloc
->end
;
4229 * 2. Distribute the remaining space in proportion to the amount of
4230 * data each plane needs to fetch from memory.
4232 * FIXME: we may not allocate every single block here.
4234 total_data_rate
= skl_get_total_relative_data_rate(cstate
,
4237 if (total_data_rate
== 0)
4240 start
= alloc
->start
;
4241 for_each_plane_id_on_crtc(intel_crtc
, plane_id
) {
4242 unsigned int data_rate
, y_data_rate
;
4243 uint16_t plane_blocks
, y_plane_blocks
= 0;
4245 if (plane_id
== PLANE_CURSOR
)
4248 data_rate
= plane_data_rate
[plane_id
];
4251 * allocation for (packed formats) or (uv-plane part of planar format):
4252 * promote the expression to 64 bits to avoid overflowing, the
4253 * result is < available as data_rate / total_data_rate < 1
4255 plane_blocks
= minimum
[plane_id
];
4256 plane_blocks
+= div_u64((uint64_t)alloc_size
* data_rate
,
4259 /* Leave disabled planes at (0,0) */
4261 ddb
->plane
[pipe
][plane_id
].start
= start
;
4262 ddb
->plane
[pipe
][plane_id
].end
= start
+ plane_blocks
;
4265 start
+= plane_blocks
;
4268 * allocation for y_plane part of planar format:
4270 y_data_rate
= plane_y_data_rate
[plane_id
];
4272 y_plane_blocks
= y_minimum
[plane_id
];
4273 y_plane_blocks
+= div_u64((uint64_t)alloc_size
* y_data_rate
,
4277 ddb
->y_plane
[pipe
][plane_id
].start
= start
;
4278 ddb
->y_plane
[pipe
][plane_id
].end
= start
+ y_plane_blocks
;
4281 start
+= y_plane_blocks
;
4288 * The max latency should be 257 (max the punit can code is 255 and we add 2us
4289 * for the read latency) and cpp should always be <= 8, so that
4290 * should allow pixel_rate up to ~2 GHz which seems sufficient since max
4291 * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
4293 static uint_fixed_16_16_t
4294 skl_wm_method1(const struct drm_i915_private
*dev_priv
, uint32_t pixel_rate
,
4295 uint8_t cpp
, uint32_t latency
)
4297 uint32_t wm_intermediate_val
;
4298 uint_fixed_16_16_t ret
;
4301 return FP_16_16_MAX
;
4303 wm_intermediate_val
= latency
* pixel_rate
* cpp
;
4304 ret
= div_fixed16(wm_intermediate_val
, 1000 * 512);
4306 if (INTEL_GEN(dev_priv
) >= 10)
4307 ret
= add_fixed16_u32(ret
, 1);
4312 static uint_fixed_16_16_t
skl_wm_method2(uint32_t pixel_rate
,
4313 uint32_t pipe_htotal
,
4315 uint_fixed_16_16_t plane_blocks_per_line
)
4317 uint32_t wm_intermediate_val
;
4318 uint_fixed_16_16_t ret
;
4321 return FP_16_16_MAX
;
4323 wm_intermediate_val
= latency
* pixel_rate
;
4324 wm_intermediate_val
= DIV_ROUND_UP(wm_intermediate_val
,
4325 pipe_htotal
* 1000);
4326 ret
= mul_u32_fixed16(wm_intermediate_val
, plane_blocks_per_line
);
4330 static uint_fixed_16_16_t
4331 intel_get_linetime_us(struct intel_crtc_state
*cstate
)
4333 uint32_t pixel_rate
;
4334 uint32_t crtc_htotal
;
4335 uint_fixed_16_16_t linetime_us
;
4337 if (!cstate
->base
.active
)
4338 return u32_to_fixed16(0);
4340 pixel_rate
= cstate
->pixel_rate
;
4342 if (WARN_ON(pixel_rate
== 0))
4343 return u32_to_fixed16(0);
4345 crtc_htotal
= cstate
->base
.adjusted_mode
.crtc_htotal
;
4346 linetime_us
= div_fixed16(crtc_htotal
* 1000, pixel_rate
);
4352 skl_adjusted_plane_pixel_rate(const struct intel_crtc_state
*cstate
,
4353 const struct intel_plane_state
*pstate
)
4355 uint64_t adjusted_pixel_rate
;
4356 uint_fixed_16_16_t downscale_amount
;
4358 /* Shouldn't reach here on disabled planes... */
4359 if (WARN_ON(!intel_wm_plane_visible(cstate
, pstate
)))
4363 * Adjusted plane pixel rate is just the pipe's adjusted pixel rate
4364 * with additional adjustments for plane-specific scaling.
4366 adjusted_pixel_rate
= cstate
->pixel_rate
;
4367 downscale_amount
= skl_plane_downscale_amount(cstate
, pstate
);
4369 return mul_round_up_u32_fixed16(adjusted_pixel_rate
,
4373 static int skl_compute_plane_wm(const struct drm_i915_private
*dev_priv
,
4374 struct intel_crtc_state
*cstate
,
4375 const struct intel_plane_state
*intel_pstate
,
4376 uint16_t ddb_allocation
,
4378 uint16_t *out_blocks
, /* out */
4379 uint8_t *out_lines
, /* out */
4380 bool *enabled
/* out */)
4382 struct intel_plane
*plane
= to_intel_plane(intel_pstate
->base
.plane
);
4383 const struct drm_plane_state
*pstate
= &intel_pstate
->base
;
4384 const struct drm_framebuffer
*fb
= pstate
->fb
;
4385 uint32_t latency
= dev_priv
->wm
.skl_latency
[level
];
4386 uint_fixed_16_16_t method1
, method2
;
4387 uint_fixed_16_16_t plane_blocks_per_line
;
4388 uint_fixed_16_16_t selected_result
;
4389 uint32_t interm_pbpl
;
4390 uint32_t plane_bytes_per_line
;
4391 uint32_t res_blocks
, res_lines
;
4394 uint32_t plane_pixel_rate
;
4395 uint_fixed_16_16_t y_tile_minimum
;
4396 uint32_t y_min_scanlines
;
4397 struct intel_atomic_state
*state
=
4398 to_intel_atomic_state(cstate
->base
.state
);
4399 bool apply_memory_bw_wa
= skl_needs_memory_bw_wa(state
);
4400 bool y_tiled
, x_tiled
;
4403 !intel_wm_plane_visible(cstate
, intel_pstate
)) {
4408 y_tiled
= fb
->modifier
== I915_FORMAT_MOD_Y_TILED
||
4409 fb
->modifier
== I915_FORMAT_MOD_Yf_TILED
||
4410 fb
->modifier
== I915_FORMAT_MOD_Y_TILED_CCS
||
4411 fb
->modifier
== I915_FORMAT_MOD_Yf_TILED_CCS
;
4412 x_tiled
= fb
->modifier
== I915_FORMAT_MOD_X_TILED
;
4414 /* Display WA #1141: kbl,cfl */
4415 if ((IS_KABYLAKE(dev_priv
) || IS_COFFEELAKE(dev_priv
)) &&
4416 dev_priv
->ipc_enabled
)
4419 if (apply_memory_bw_wa
&& x_tiled
)
4422 if (plane
->id
== PLANE_CURSOR
) {
4423 width
= intel_pstate
->base
.crtc_w
;
4426 * Src coordinates are already rotated by 270 degrees for
4427 * the 90/270 degree plane rotation cases (to match the
4428 * GTT mapping), hence no need to account for rotation here.
4430 width
= drm_rect_width(&intel_pstate
->base
.src
) >> 16;
4433 cpp
= (fb
->format
->format
== DRM_FORMAT_NV12
) ? fb
->format
->cpp
[1] :
4435 plane_pixel_rate
= skl_adjusted_plane_pixel_rate(cstate
, intel_pstate
);
4437 if (drm_rotation_90_or_270(pstate
->rotation
)) {
4441 y_min_scanlines
= 16;
4444 y_min_scanlines
= 8;
4447 y_min_scanlines
= 4;
4454 y_min_scanlines
= 4;
4457 if (apply_memory_bw_wa
)
4458 y_min_scanlines
*= 2;
4460 plane_bytes_per_line
= width
* cpp
;
4462 interm_pbpl
= DIV_ROUND_UP(plane_bytes_per_line
*
4463 y_min_scanlines
, 512);
4465 if (INTEL_GEN(dev_priv
) >= 10)
4468 plane_blocks_per_line
= div_fixed16(interm_pbpl
,
4470 } else if (x_tiled
&& INTEL_GEN(dev_priv
) == 9) {
4471 interm_pbpl
= DIV_ROUND_UP(plane_bytes_per_line
, 512);
4472 plane_blocks_per_line
= u32_to_fixed16(interm_pbpl
);
4474 interm_pbpl
= DIV_ROUND_UP(plane_bytes_per_line
, 512) + 1;
4475 plane_blocks_per_line
= u32_to_fixed16(interm_pbpl
);
4478 method1
= skl_wm_method1(dev_priv
, plane_pixel_rate
, cpp
, latency
);
4479 method2
= skl_wm_method2(plane_pixel_rate
,
4480 cstate
->base
.adjusted_mode
.crtc_htotal
,
4482 plane_blocks_per_line
);
4484 y_tile_minimum
= mul_u32_fixed16(y_min_scanlines
,
4485 plane_blocks_per_line
);
4488 selected_result
= max_fixed16(method2
, y_tile_minimum
);
4490 uint32_t linetime_us
;
4492 linetime_us
= fixed16_to_u32_round_up(
4493 intel_get_linetime_us(cstate
));
4494 if ((cpp
* cstate
->base
.adjusted_mode
.crtc_htotal
/ 512 < 1) &&
4495 (plane_bytes_per_line
/ 512 < 1))
4496 selected_result
= method2
;
4497 else if (ddb_allocation
>=
4498 fixed16_to_u32_round_up(plane_blocks_per_line
))
4499 selected_result
= min_fixed16(method1
, method2
);
4500 else if (latency
>= linetime_us
)
4501 selected_result
= min_fixed16(method1
, method2
);
4503 selected_result
= method1
;
4506 res_blocks
= fixed16_to_u32_round_up(selected_result
) + 1;
4507 res_lines
= div_round_up_fixed16(selected_result
,
4508 plane_blocks_per_line
);
4510 /* Display WA #1125: skl,bxt,kbl,glk */
4512 (fb
->modifier
== I915_FORMAT_MOD_Y_TILED_CCS
||
4513 fb
->modifier
== I915_FORMAT_MOD_Yf_TILED_CCS
))
4514 res_blocks
+= fixed16_to_u32_round_up(y_tile_minimum
);
4516 /* Display WA #1126: skl,bxt,kbl,glk */
4517 if (level
>= 1 && level
<= 7) {
4519 res_blocks
+= fixed16_to_u32_round_up(y_tile_minimum
);
4520 res_lines
+= y_min_scanlines
;
4526 if (res_blocks
>= ddb_allocation
|| res_lines
> 31) {
4530 * If there are no valid level 0 watermarks, then we can't
4531 * support this display configuration.
4536 struct drm_plane
*plane
= pstate
->plane
;
4538 DRM_DEBUG_KMS("Requested display configuration exceeds system watermark limitations\n");
4539 DRM_DEBUG_KMS("[PLANE:%d:%s] blocks required = %u/%u, lines required = %u/31\n",
4540 plane
->base
.id
, plane
->name
,
4541 res_blocks
, ddb_allocation
, res_lines
);
4546 *out_blocks
= res_blocks
;
4547 *out_lines
= res_lines
;
4554 skl_compute_wm_levels(const struct drm_i915_private
*dev_priv
,
4555 struct skl_ddb_allocation
*ddb
,
4556 struct intel_crtc_state
*cstate
,
4557 const struct intel_plane_state
*intel_pstate
,
4558 struct skl_plane_wm
*wm
)
4560 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
4561 struct drm_plane
*plane
= intel_pstate
->base
.plane
;
4562 struct intel_plane
*intel_plane
= to_intel_plane(plane
);
4563 uint16_t ddb_blocks
;
4564 enum pipe pipe
= intel_crtc
->pipe
;
4565 int level
, max_level
= ilk_wm_max_level(dev_priv
);
4568 if (WARN_ON(!intel_pstate
->base
.fb
))
4571 ddb_blocks
= skl_ddb_entry_size(&ddb
->plane
[pipe
][intel_plane
->id
]);
4573 for (level
= 0; level
<= max_level
; level
++) {
4574 struct skl_wm_level
*result
= &wm
->wm
[level
];
4576 ret
= skl_compute_plane_wm(dev_priv
,
4581 &result
->plane_res_b
,
4582 &result
->plane_res_l
,
4592 skl_compute_linetime_wm(struct intel_crtc_state
*cstate
)
4594 struct drm_atomic_state
*state
= cstate
->base
.state
;
4595 struct drm_i915_private
*dev_priv
= to_i915(state
->dev
);
4596 uint_fixed_16_16_t linetime_us
;
4597 uint32_t linetime_wm
;
4599 linetime_us
= intel_get_linetime_us(cstate
);
4601 if (is_fixed16_zero(linetime_us
))
4604 linetime_wm
= fixed16_to_u32_round_up(mul_u32_fixed16(8, linetime_us
));
4606 /* Display WA #1135: bxt. */
4607 if (IS_BROXTON(dev_priv
) && dev_priv
->ipc_enabled
)
4608 linetime_wm
= DIV_ROUND_UP(linetime_wm
, 2);
4613 static void skl_compute_transition_wm(struct intel_crtc_state
*cstate
,
4614 struct skl_wm_level
*trans_wm
/* out */)
4616 if (!cstate
->base
.active
)
4619 /* Until we know more, just disable transition WMs */
4620 trans_wm
->plane_en
= false;
4623 static int skl_build_pipe_wm(struct intel_crtc_state
*cstate
,
4624 struct skl_ddb_allocation
*ddb
,
4625 struct skl_pipe_wm
*pipe_wm
)
4627 struct drm_device
*dev
= cstate
->base
.crtc
->dev
;
4628 struct drm_crtc_state
*crtc_state
= &cstate
->base
;
4629 const struct drm_i915_private
*dev_priv
= to_i915(dev
);
4630 struct drm_plane
*plane
;
4631 const struct drm_plane_state
*pstate
;
4632 struct skl_plane_wm
*wm
;
4636 * We'll only calculate watermarks for planes that are actually
4637 * enabled, so make sure all other planes are set as disabled.
4639 memset(pipe_wm
->planes
, 0, sizeof(pipe_wm
->planes
));
4641 drm_atomic_crtc_state_for_each_plane_state(plane
, pstate
, crtc_state
) {
4642 const struct intel_plane_state
*intel_pstate
=
4643 to_intel_plane_state(pstate
);
4644 enum plane_id plane_id
= to_intel_plane(plane
)->id
;
4646 wm
= &pipe_wm
->planes
[plane_id
];
4648 ret
= skl_compute_wm_levels(dev_priv
, ddb
, cstate
,
4652 skl_compute_transition_wm(cstate
, &wm
->trans_wm
);
4654 pipe_wm
->linetime
= skl_compute_linetime_wm(cstate
);
4659 static void skl_ddb_entry_write(struct drm_i915_private
*dev_priv
,
4661 const struct skl_ddb_entry
*entry
)
4664 I915_WRITE(reg
, (entry
->end
- 1) << 16 | entry
->start
);
4669 static void skl_write_wm_level(struct drm_i915_private
*dev_priv
,
4671 const struct skl_wm_level
*level
)
4675 if (level
->plane_en
) {
4677 val
|= level
->plane_res_b
;
4678 val
|= level
->plane_res_l
<< PLANE_WM_LINES_SHIFT
;
4681 I915_WRITE(reg
, val
);
4684 static void skl_write_plane_wm(struct intel_crtc
*intel_crtc
,
4685 const struct skl_plane_wm
*wm
,
4686 const struct skl_ddb_allocation
*ddb
,
4687 enum plane_id plane_id
)
4689 struct drm_crtc
*crtc
= &intel_crtc
->base
;
4690 struct drm_device
*dev
= crtc
->dev
;
4691 struct drm_i915_private
*dev_priv
= to_i915(dev
);
4692 int level
, max_level
= ilk_wm_max_level(dev_priv
);
4693 enum pipe pipe
= intel_crtc
->pipe
;
4695 for (level
= 0; level
<= max_level
; level
++) {
4696 skl_write_wm_level(dev_priv
, PLANE_WM(pipe
, plane_id
, level
),
4699 skl_write_wm_level(dev_priv
, PLANE_WM_TRANS(pipe
, plane_id
),
4702 skl_ddb_entry_write(dev_priv
, PLANE_BUF_CFG(pipe
, plane_id
),
4703 &ddb
->plane
[pipe
][plane_id
]);
4704 skl_ddb_entry_write(dev_priv
, PLANE_NV12_BUF_CFG(pipe
, plane_id
),
4705 &ddb
->y_plane
[pipe
][plane_id
]);
4708 static void skl_write_cursor_wm(struct intel_crtc
*intel_crtc
,
4709 const struct skl_plane_wm
*wm
,
4710 const struct skl_ddb_allocation
*ddb
)
4712 struct drm_crtc
*crtc
= &intel_crtc
->base
;
4713 struct drm_device
*dev
= crtc
->dev
;
4714 struct drm_i915_private
*dev_priv
= to_i915(dev
);
4715 int level
, max_level
= ilk_wm_max_level(dev_priv
);
4716 enum pipe pipe
= intel_crtc
->pipe
;
4718 for (level
= 0; level
<= max_level
; level
++) {
4719 skl_write_wm_level(dev_priv
, CUR_WM(pipe
, level
),
4722 skl_write_wm_level(dev_priv
, CUR_WM_TRANS(pipe
), &wm
->trans_wm
);
4724 skl_ddb_entry_write(dev_priv
, CUR_BUF_CFG(pipe
),
4725 &ddb
->plane
[pipe
][PLANE_CURSOR
]);
4728 bool skl_wm_level_equals(const struct skl_wm_level
*l1
,
4729 const struct skl_wm_level
*l2
)
4731 if (l1
->plane_en
!= l2
->plane_en
)
4734 /* If both planes aren't enabled, the rest shouldn't matter */
4738 return (l1
->plane_res_l
== l2
->plane_res_l
&&
4739 l1
->plane_res_b
== l2
->plane_res_b
);
4742 static inline bool skl_ddb_entries_overlap(const struct skl_ddb_entry
*a
,
4743 const struct skl_ddb_entry
*b
)
4745 return a
->start
< b
->end
&& b
->start
< a
->end
;
4748 bool skl_ddb_allocation_overlaps(const struct skl_ddb_entry
**entries
,
4749 const struct skl_ddb_entry
*ddb
,
4754 for (i
= 0; i
< I915_MAX_PIPES
; i
++)
4755 if (i
!= ignore
&& entries
[i
] &&
4756 skl_ddb_entries_overlap(ddb
, entries
[i
]))
4762 static int skl_update_pipe_wm(struct drm_crtc_state
*cstate
,
4763 const struct skl_pipe_wm
*old_pipe_wm
,
4764 struct skl_pipe_wm
*pipe_wm
, /* out */
4765 struct skl_ddb_allocation
*ddb
, /* out */
4766 bool *changed
/* out */)
4768 struct intel_crtc_state
*intel_cstate
= to_intel_crtc_state(cstate
);
4771 ret
= skl_build_pipe_wm(intel_cstate
, ddb
, pipe_wm
);
4775 if (!memcmp(old_pipe_wm
, pipe_wm
, sizeof(*pipe_wm
)))
4784 pipes_modified(struct drm_atomic_state
*state
)
4786 struct drm_crtc
*crtc
;
4787 struct drm_crtc_state
*cstate
;
4788 uint32_t i
, ret
= 0;
4790 for_each_new_crtc_in_state(state
, crtc
, cstate
, i
)
4791 ret
|= drm_crtc_mask(crtc
);
4797 skl_ddb_add_affected_planes(struct intel_crtc_state
*cstate
)
4799 struct drm_atomic_state
*state
= cstate
->base
.state
;
4800 struct drm_device
*dev
= state
->dev
;
4801 struct drm_crtc
*crtc
= cstate
->base
.crtc
;
4802 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
4803 struct drm_i915_private
*dev_priv
= to_i915(dev
);
4804 struct intel_atomic_state
*intel_state
= to_intel_atomic_state(state
);
4805 struct skl_ddb_allocation
*new_ddb
= &intel_state
->wm_results
.ddb
;
4806 struct skl_ddb_allocation
*cur_ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
4807 struct drm_plane_state
*plane_state
;
4808 struct drm_plane
*plane
;
4809 enum pipe pipe
= intel_crtc
->pipe
;
4811 WARN_ON(!drm_atomic_get_existing_crtc_state(state
, crtc
));
4813 drm_for_each_plane_mask(plane
, dev
, cstate
->base
.plane_mask
) {
4814 enum plane_id plane_id
= to_intel_plane(plane
)->id
;
4816 if (skl_ddb_entry_equal(&cur_ddb
->plane
[pipe
][plane_id
],
4817 &new_ddb
->plane
[pipe
][plane_id
]) &&
4818 skl_ddb_entry_equal(&cur_ddb
->y_plane
[pipe
][plane_id
],
4819 &new_ddb
->y_plane
[pipe
][plane_id
]))
4822 plane_state
= drm_atomic_get_plane_state(state
, plane
);
4823 if (IS_ERR(plane_state
))
4824 return PTR_ERR(plane_state
);
4831 skl_compute_ddb(struct drm_atomic_state
*state
)
4833 struct drm_device
*dev
= state
->dev
;
4834 struct drm_i915_private
*dev_priv
= to_i915(dev
);
4835 struct intel_atomic_state
*intel_state
= to_intel_atomic_state(state
);
4836 struct intel_crtc
*intel_crtc
;
4837 struct skl_ddb_allocation
*ddb
= &intel_state
->wm_results
.ddb
;
4838 uint32_t realloc_pipes
= pipes_modified(state
);
4842 * If this is our first atomic update following hardware readout,
4843 * we can't trust the DDB that the BIOS programmed for us. Let's
4844 * pretend that all pipes switched active status so that we'll
4845 * ensure a full DDB recompute.
4847 if (dev_priv
->wm
.distrust_bios_wm
) {
4848 ret
= drm_modeset_lock(&dev
->mode_config
.connection_mutex
,
4849 state
->acquire_ctx
);
4853 intel_state
->active_pipe_changes
= ~0;
4856 * We usually only initialize intel_state->active_crtcs if we
4857 * we're doing a modeset; make sure this field is always
4858 * initialized during the sanitization process that happens
4859 * on the first commit too.
4861 if (!intel_state
->modeset
)
4862 intel_state
->active_crtcs
= dev_priv
->active_crtcs
;
4866 * If the modeset changes which CRTC's are active, we need to
4867 * recompute the DDB allocation for *all* active pipes, even
4868 * those that weren't otherwise being modified in any way by this
4869 * atomic commit. Due to the shrinking of the per-pipe allocations
4870 * when new active CRTC's are added, it's possible for a pipe that
4871 * we were already using and aren't changing at all here to suddenly
4872 * become invalid if its DDB needs exceeds its new allocation.
4874 * Note that if we wind up doing a full DDB recompute, we can't let
4875 * any other display updates race with this transaction, so we need
4876 * to grab the lock on *all* CRTC's.
4878 if (intel_state
->active_pipe_changes
) {
4880 intel_state
->wm_results
.dirty_pipes
= ~0;
4884 * We're not recomputing for the pipes not included in the commit, so
4885 * make sure we start with the current state.
4887 memcpy(ddb
, &dev_priv
->wm
.skl_hw
.ddb
, sizeof(*ddb
));
4889 for_each_intel_crtc_mask(dev
, intel_crtc
, realloc_pipes
) {
4890 struct intel_crtc_state
*cstate
;
4892 cstate
= intel_atomic_get_crtc_state(state
, intel_crtc
);
4894 return PTR_ERR(cstate
);
4896 ret
= skl_allocate_pipe_ddb(cstate
, ddb
);
4900 ret
= skl_ddb_add_affected_planes(cstate
);
4909 skl_copy_wm_for_pipe(struct skl_wm_values
*dst
,
4910 struct skl_wm_values
*src
,
4913 memcpy(dst
->ddb
.y_plane
[pipe
], src
->ddb
.y_plane
[pipe
],
4914 sizeof(dst
->ddb
.y_plane
[pipe
]));
4915 memcpy(dst
->ddb
.plane
[pipe
], src
->ddb
.plane
[pipe
],
4916 sizeof(dst
->ddb
.plane
[pipe
]));
4920 skl_print_wm_changes(const struct drm_atomic_state
*state
)
4922 const struct drm_device
*dev
= state
->dev
;
4923 const struct drm_i915_private
*dev_priv
= to_i915(dev
);
4924 const struct intel_atomic_state
*intel_state
=
4925 to_intel_atomic_state(state
);
4926 const struct drm_crtc
*crtc
;
4927 const struct drm_crtc_state
*cstate
;
4928 const struct intel_plane
*intel_plane
;
4929 const struct skl_ddb_allocation
*old_ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
4930 const struct skl_ddb_allocation
*new_ddb
= &intel_state
->wm_results
.ddb
;
4933 for_each_new_crtc_in_state(state
, crtc
, cstate
, i
) {
4934 const struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
4935 enum pipe pipe
= intel_crtc
->pipe
;
4937 for_each_intel_plane_on_crtc(dev
, intel_crtc
, intel_plane
) {
4938 enum plane_id plane_id
= intel_plane
->id
;
4939 const struct skl_ddb_entry
*old
, *new;
4941 old
= &old_ddb
->plane
[pipe
][plane_id
];
4942 new = &new_ddb
->plane
[pipe
][plane_id
];
4944 if (skl_ddb_entry_equal(old
, new))
4947 DRM_DEBUG_ATOMIC("[PLANE:%d:%s] ddb (%d - %d) -> (%d - %d)\n",
4948 intel_plane
->base
.base
.id
,
4949 intel_plane
->base
.name
,
4950 old
->start
, old
->end
,
4951 new->start
, new->end
);
4957 skl_compute_wm(struct drm_atomic_state
*state
)
4959 struct drm_crtc
*crtc
;
4960 struct drm_crtc_state
*cstate
;
4961 struct intel_atomic_state
*intel_state
= to_intel_atomic_state(state
);
4962 struct skl_wm_values
*results
= &intel_state
->wm_results
;
4963 struct drm_device
*dev
= state
->dev
;
4964 struct skl_pipe_wm
*pipe_wm
;
4965 bool changed
= false;
4969 * When we distrust bios wm we always need to recompute to set the
4970 * expected DDB allocations for each CRTC.
4972 if (to_i915(dev
)->wm
.distrust_bios_wm
)
4976 * If this transaction isn't actually touching any CRTC's, don't
4977 * bother with watermark calculation. Note that if we pass this
4978 * test, we're guaranteed to hold at least one CRTC state mutex,
4979 * which means we can safely use values like dev_priv->active_crtcs
4980 * since any racing commits that want to update them would need to
4981 * hold _all_ CRTC state mutexes.
4983 for_each_new_crtc_in_state(state
, crtc
, cstate
, i
)
4989 /* Clear all dirty flags */
4990 results
->dirty_pipes
= 0;
4992 ret
= skl_compute_ddb(state
);
4997 * Calculate WM's for all pipes that are part of this transaction.
4998 * Note that the DDB allocation above may have added more CRTC's that
4999 * weren't otherwise being modified (and set bits in dirty_pipes) if
5000 * pipe allocations had to change.
5002 * FIXME: Now that we're doing this in the atomic check phase, we
5003 * should allow skl_update_pipe_wm() to return failure in cases where
5004 * no suitable watermark values can be found.
5006 for_each_new_crtc_in_state(state
, crtc
, cstate
, i
) {
5007 struct intel_crtc_state
*intel_cstate
=
5008 to_intel_crtc_state(cstate
);
5009 const struct skl_pipe_wm
*old_pipe_wm
=
5010 &to_intel_crtc_state(crtc
->state
)->wm
.skl
.optimal
;
5012 pipe_wm
= &intel_cstate
->wm
.skl
.optimal
;
5013 ret
= skl_update_pipe_wm(cstate
, old_pipe_wm
, pipe_wm
,
5014 &results
->ddb
, &changed
);
5019 results
->dirty_pipes
|= drm_crtc_mask(crtc
);
5021 if ((results
->dirty_pipes
& drm_crtc_mask(crtc
)) == 0)
5022 /* This pipe's WM's did not change */
5025 intel_cstate
->update_wm_pre
= true;
5028 skl_print_wm_changes(state
);
5033 static void skl_atomic_update_crtc_wm(struct intel_atomic_state
*state
,
5034 struct intel_crtc_state
*cstate
)
5036 struct intel_crtc
*crtc
= to_intel_crtc(cstate
->base
.crtc
);
5037 struct drm_i915_private
*dev_priv
= to_i915(state
->base
.dev
);
5038 struct skl_pipe_wm
*pipe_wm
= &cstate
->wm
.skl
.optimal
;
5039 const struct skl_ddb_allocation
*ddb
= &state
->wm_results
.ddb
;
5040 enum pipe pipe
= crtc
->pipe
;
5041 enum plane_id plane_id
;
5043 if (!(state
->wm_results
.dirty_pipes
& drm_crtc_mask(&crtc
->base
)))
5046 I915_WRITE(PIPE_WM_LINETIME(pipe
), pipe_wm
->linetime
);
5048 for_each_plane_id_on_crtc(crtc
, plane_id
) {
5049 if (plane_id
!= PLANE_CURSOR
)
5050 skl_write_plane_wm(crtc
, &pipe_wm
->planes
[plane_id
],
5053 skl_write_cursor_wm(crtc
, &pipe_wm
->planes
[plane_id
],
5058 static void skl_initial_wm(struct intel_atomic_state
*state
,
5059 struct intel_crtc_state
*cstate
)
5061 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
5062 struct drm_device
*dev
= intel_crtc
->base
.dev
;
5063 struct drm_i915_private
*dev_priv
= to_i915(dev
);
5064 struct skl_wm_values
*results
= &state
->wm_results
;
5065 struct skl_wm_values
*hw_vals
= &dev_priv
->wm
.skl_hw
;
5066 enum pipe pipe
= intel_crtc
->pipe
;
5068 if ((results
->dirty_pipes
& drm_crtc_mask(&intel_crtc
->base
)) == 0)
5071 mutex_lock(&dev_priv
->wm
.wm_mutex
);
5073 if (cstate
->base
.active_changed
)
5074 skl_atomic_update_crtc_wm(state
, cstate
);
5076 skl_copy_wm_for_pipe(hw_vals
, results
, pipe
);
5078 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
5081 static void ilk_compute_wm_config(struct drm_device
*dev
,
5082 struct intel_wm_config
*config
)
5084 struct intel_crtc
*crtc
;
5086 /* Compute the currently _active_ config */
5087 for_each_intel_crtc(dev
, crtc
) {
5088 const struct intel_pipe_wm
*wm
= &crtc
->wm
.active
.ilk
;
5090 if (!wm
->pipe_enabled
)
5093 config
->sprites_enabled
|= wm
->sprites_enabled
;
5094 config
->sprites_scaled
|= wm
->sprites_scaled
;
5095 config
->num_pipes_active
++;
5099 static void ilk_program_watermarks(struct drm_i915_private
*dev_priv
)
5101 struct drm_device
*dev
= &dev_priv
->drm
;
5102 struct intel_pipe_wm lp_wm_1_2
= {}, lp_wm_5_6
= {}, *best_lp_wm
;
5103 struct ilk_wm_maximums max
;
5104 struct intel_wm_config config
= {};
5105 struct ilk_wm_values results
= {};
5106 enum intel_ddb_partitioning partitioning
;
5108 ilk_compute_wm_config(dev
, &config
);
5110 ilk_compute_wm_maximums(dev
, 1, &config
, INTEL_DDB_PART_1_2
, &max
);
5111 ilk_wm_merge(dev
, &config
, &max
, &lp_wm_1_2
);
5113 /* 5/6 split only in single pipe config on IVB+ */
5114 if (INTEL_GEN(dev_priv
) >= 7 &&
5115 config
.num_pipes_active
== 1 && config
.sprites_enabled
) {
5116 ilk_compute_wm_maximums(dev
, 1, &config
, INTEL_DDB_PART_5_6
, &max
);
5117 ilk_wm_merge(dev
, &config
, &max
, &lp_wm_5_6
);
5119 best_lp_wm
= ilk_find_best_result(dev
, &lp_wm_1_2
, &lp_wm_5_6
);
5121 best_lp_wm
= &lp_wm_1_2
;
5124 partitioning
= (best_lp_wm
== &lp_wm_1_2
) ?
5125 INTEL_DDB_PART_1_2
: INTEL_DDB_PART_5_6
;
5127 ilk_compute_wm_results(dev
, best_lp_wm
, partitioning
, &results
);
5129 ilk_write_wm_values(dev_priv
, &results
);
5132 static void ilk_initial_watermarks(struct intel_atomic_state
*state
,
5133 struct intel_crtc_state
*cstate
)
5135 struct drm_i915_private
*dev_priv
= to_i915(cstate
->base
.crtc
->dev
);
5136 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
5138 mutex_lock(&dev_priv
->wm
.wm_mutex
);
5139 intel_crtc
->wm
.active
.ilk
= cstate
->wm
.ilk
.intermediate
;
5140 ilk_program_watermarks(dev_priv
);
5141 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
5144 static void ilk_optimize_watermarks(struct intel_atomic_state
*state
,
5145 struct intel_crtc_state
*cstate
)
5147 struct drm_i915_private
*dev_priv
= to_i915(cstate
->base
.crtc
->dev
);
5148 struct intel_crtc
*intel_crtc
= to_intel_crtc(cstate
->base
.crtc
);
5150 mutex_lock(&dev_priv
->wm
.wm_mutex
);
5151 if (cstate
->wm
.need_postvbl_update
) {
5152 intel_crtc
->wm
.active
.ilk
= cstate
->wm
.ilk
.optimal
;
5153 ilk_program_watermarks(dev_priv
);
5155 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
5158 static inline void skl_wm_level_from_reg_val(uint32_t val
,
5159 struct skl_wm_level
*level
)
5161 level
->plane_en
= val
& PLANE_WM_EN
;
5162 level
->plane_res_b
= val
& PLANE_WM_BLOCKS_MASK
;
5163 level
->plane_res_l
= (val
>> PLANE_WM_LINES_SHIFT
) &
5164 PLANE_WM_LINES_MASK
;
5167 void skl_pipe_wm_get_hw_state(struct drm_crtc
*crtc
,
5168 struct skl_pipe_wm
*out
)
5170 struct drm_i915_private
*dev_priv
= to_i915(crtc
->dev
);
5171 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
5172 enum pipe pipe
= intel_crtc
->pipe
;
5173 int level
, max_level
;
5174 enum plane_id plane_id
;
5177 max_level
= ilk_wm_max_level(dev_priv
);
5179 for_each_plane_id_on_crtc(intel_crtc
, plane_id
) {
5180 struct skl_plane_wm
*wm
= &out
->planes
[plane_id
];
5182 for (level
= 0; level
<= max_level
; level
++) {
5183 if (plane_id
!= PLANE_CURSOR
)
5184 val
= I915_READ(PLANE_WM(pipe
, plane_id
, level
));
5186 val
= I915_READ(CUR_WM(pipe
, level
));
5188 skl_wm_level_from_reg_val(val
, &wm
->wm
[level
]);
5191 if (plane_id
!= PLANE_CURSOR
)
5192 val
= I915_READ(PLANE_WM_TRANS(pipe
, plane_id
));
5194 val
= I915_READ(CUR_WM_TRANS(pipe
));
5196 skl_wm_level_from_reg_val(val
, &wm
->trans_wm
);
5199 if (!intel_crtc
->active
)
5202 out
->linetime
= I915_READ(PIPE_WM_LINETIME(pipe
));
5205 void skl_wm_get_hw_state(struct drm_device
*dev
)
5207 struct drm_i915_private
*dev_priv
= to_i915(dev
);
5208 struct skl_wm_values
*hw
= &dev_priv
->wm
.skl_hw
;
5209 struct skl_ddb_allocation
*ddb
= &dev_priv
->wm
.skl_hw
.ddb
;
5210 struct drm_crtc
*crtc
;
5211 struct intel_crtc
*intel_crtc
;
5212 struct intel_crtc_state
*cstate
;
5214 skl_ddb_get_hw_state(dev_priv
, ddb
);
5215 list_for_each_entry(crtc
, &dev
->mode_config
.crtc_list
, head
) {
5216 intel_crtc
= to_intel_crtc(crtc
);
5217 cstate
= to_intel_crtc_state(crtc
->state
);
5219 skl_pipe_wm_get_hw_state(crtc
, &cstate
->wm
.skl
.optimal
);
5221 if (intel_crtc
->active
)
5222 hw
->dirty_pipes
|= drm_crtc_mask(crtc
);
5225 if (dev_priv
->active_crtcs
) {
5226 /* Fully recompute DDB on first atomic commit */
5227 dev_priv
->wm
.distrust_bios_wm
= true;
5229 /* Easy/common case; just sanitize DDB now if everything off */
5230 memset(ddb
, 0, sizeof(*ddb
));
5234 static void ilk_pipe_wm_get_hw_state(struct drm_crtc
*crtc
)
5236 struct drm_device
*dev
= crtc
->dev
;
5237 struct drm_i915_private
*dev_priv
= to_i915(dev
);
5238 struct ilk_wm_values
*hw
= &dev_priv
->wm
.hw
;
5239 struct intel_crtc
*intel_crtc
= to_intel_crtc(crtc
);
5240 struct intel_crtc_state
*cstate
= to_intel_crtc_state(crtc
->state
);
5241 struct intel_pipe_wm
*active
= &cstate
->wm
.ilk
.optimal
;
5242 enum pipe pipe
= intel_crtc
->pipe
;
5243 static const i915_reg_t wm0_pipe_reg
[] = {
5244 [PIPE_A
] = WM0_PIPEA_ILK
,
5245 [PIPE_B
] = WM0_PIPEB_ILK
,
5246 [PIPE_C
] = WM0_PIPEC_IVB
,
5249 hw
->wm_pipe
[pipe
] = I915_READ(wm0_pipe_reg
[pipe
]);
5250 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
5251 hw
->wm_linetime
[pipe
] = I915_READ(PIPE_WM_LINETIME(pipe
));
5253 memset(active
, 0, sizeof(*active
));
5255 active
->pipe_enabled
= intel_crtc
->active
;
5257 if (active
->pipe_enabled
) {
5258 u32 tmp
= hw
->wm_pipe
[pipe
];
5261 * For active pipes LP0 watermark is marked as
5262 * enabled, and LP1+ watermaks as disabled since
5263 * we can't really reverse compute them in case
5264 * multiple pipes are active.
5266 active
->wm
[0].enable
= true;
5267 active
->wm
[0].pri_val
= (tmp
& WM0_PIPE_PLANE_MASK
) >> WM0_PIPE_PLANE_SHIFT
;
5268 active
->wm
[0].spr_val
= (tmp
& WM0_PIPE_SPRITE_MASK
) >> WM0_PIPE_SPRITE_SHIFT
;
5269 active
->wm
[0].cur_val
= tmp
& WM0_PIPE_CURSOR_MASK
;
5270 active
->linetime
= hw
->wm_linetime
[pipe
];
5272 int level
, max_level
= ilk_wm_max_level(dev_priv
);
5275 * For inactive pipes, all watermark levels
5276 * should be marked as enabled but zeroed,
5277 * which is what we'd compute them to.
5279 for (level
= 0; level
<= max_level
; level
++)
5280 active
->wm
[level
].enable
= true;
5283 intel_crtc
->wm
.active
.ilk
= *active
;
5286 #define _FW_WM(value, plane) \
5287 (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
5288 #define _FW_WM_VLV(value, plane) \
5289 (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
5291 static void g4x_read_wm_values(struct drm_i915_private
*dev_priv
,
5292 struct g4x_wm_values
*wm
)
5296 tmp
= I915_READ(DSPFW1
);
5297 wm
->sr
.plane
= _FW_WM(tmp
, SR
);
5298 wm
->pipe
[PIPE_B
].plane
[PLANE_CURSOR
] = _FW_WM(tmp
, CURSORB
);
5299 wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
] = _FW_WM(tmp
, PLANEB
);
5300 wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
] = _FW_WM(tmp
, PLANEA
);
5302 tmp
= I915_READ(DSPFW2
);
5303 wm
->fbc_en
= tmp
& DSPFW_FBC_SR_EN
;
5304 wm
->sr
.fbc
= _FW_WM(tmp
, FBC_SR
);
5305 wm
->hpll
.fbc
= _FW_WM(tmp
, FBC_HPLL_SR
);
5306 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
] = _FW_WM(tmp
, SPRITEB
);
5307 wm
->pipe
[PIPE_A
].plane
[PLANE_CURSOR
] = _FW_WM(tmp
, CURSORA
);
5308 wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
] = _FW_WM(tmp
, SPRITEA
);
5310 tmp
= I915_READ(DSPFW3
);
5311 wm
->hpll_en
= tmp
& DSPFW_HPLL_SR_EN
;
5312 wm
->sr
.cursor
= _FW_WM(tmp
, CURSOR_SR
);
5313 wm
->hpll
.cursor
= _FW_WM(tmp
, HPLL_CURSOR
);
5314 wm
->hpll
.plane
= _FW_WM(tmp
, HPLL_SR
);
5317 static void vlv_read_wm_values(struct drm_i915_private
*dev_priv
,
5318 struct vlv_wm_values
*wm
)
5323 for_each_pipe(dev_priv
, pipe
) {
5324 tmp
= I915_READ(VLV_DDL(pipe
));
5326 wm
->ddl
[pipe
].plane
[PLANE_PRIMARY
] =
5327 (tmp
>> DDL_PLANE_SHIFT
) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
5328 wm
->ddl
[pipe
].plane
[PLANE_CURSOR
] =
5329 (tmp
>> DDL_CURSOR_SHIFT
) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
5330 wm
->ddl
[pipe
].plane
[PLANE_SPRITE0
] =
5331 (tmp
>> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
5332 wm
->ddl
[pipe
].plane
[PLANE_SPRITE1
] =
5333 (tmp
>> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH
| DRAIN_LATENCY_MASK
);
5336 tmp
= I915_READ(DSPFW1
);
5337 wm
->sr
.plane
= _FW_WM(tmp
, SR
);
5338 wm
->pipe
[PIPE_B
].plane
[PLANE_CURSOR
] = _FW_WM(tmp
, CURSORB
);
5339 wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
] = _FW_WM_VLV(tmp
, PLANEB
);
5340 wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
] = _FW_WM_VLV(tmp
, PLANEA
);
5342 tmp
= I915_READ(DSPFW2
);
5343 wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE1
] = _FW_WM_VLV(tmp
, SPRITEB
);
5344 wm
->pipe
[PIPE_A
].plane
[PLANE_CURSOR
] = _FW_WM(tmp
, CURSORA
);
5345 wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
] = _FW_WM_VLV(tmp
, SPRITEA
);
5347 tmp
= I915_READ(DSPFW3
);
5348 wm
->sr
.cursor
= _FW_WM(tmp
, CURSOR_SR
);
5350 if (IS_CHERRYVIEW(dev_priv
)) {
5351 tmp
= I915_READ(DSPFW7_CHV
);
5352 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
] = _FW_WM_VLV(tmp
, SPRITED
);
5353 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
] = _FW_WM_VLV(tmp
, SPRITEC
);
5355 tmp
= I915_READ(DSPFW8_CHV
);
5356 wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE1
] = _FW_WM_VLV(tmp
, SPRITEF
);
5357 wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE0
] = _FW_WM_VLV(tmp
, SPRITEE
);
5359 tmp
= I915_READ(DSPFW9_CHV
);
5360 wm
->pipe
[PIPE_C
].plane
[PLANE_PRIMARY
] = _FW_WM_VLV(tmp
, PLANEC
);
5361 wm
->pipe
[PIPE_C
].plane
[PLANE_CURSOR
] = _FW_WM(tmp
, CURSORC
);
5363 tmp
= I915_READ(DSPHOWM
);
5364 wm
->sr
.plane
|= _FW_WM(tmp
, SR_HI
) << 9;
5365 wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE1
] |= _FW_WM(tmp
, SPRITEF_HI
) << 8;
5366 wm
->pipe
[PIPE_C
].plane
[PLANE_SPRITE0
] |= _FW_WM(tmp
, SPRITEE_HI
) << 8;
5367 wm
->pipe
[PIPE_C
].plane
[PLANE_PRIMARY
] |= _FW_WM(tmp
, PLANEC_HI
) << 8;
5368 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
] |= _FW_WM(tmp
, SPRITED_HI
) << 8;
5369 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
] |= _FW_WM(tmp
, SPRITEC_HI
) << 8;
5370 wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
] |= _FW_WM(tmp
, PLANEB_HI
) << 8;
5371 wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE1
] |= _FW_WM(tmp
, SPRITEB_HI
) << 8;
5372 wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
] |= _FW_WM(tmp
, SPRITEA_HI
) << 8;
5373 wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
] |= _FW_WM(tmp
, PLANEA_HI
) << 8;
5375 tmp
= I915_READ(DSPFW7
);
5376 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
] = _FW_WM_VLV(tmp
, SPRITED
);
5377 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
] = _FW_WM_VLV(tmp
, SPRITEC
);
5379 tmp
= I915_READ(DSPHOWM
);
5380 wm
->sr
.plane
|= _FW_WM(tmp
, SR_HI
) << 9;
5381 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE1
] |= _FW_WM(tmp
, SPRITED_HI
) << 8;
5382 wm
->pipe
[PIPE_B
].plane
[PLANE_SPRITE0
] |= _FW_WM(tmp
, SPRITEC_HI
) << 8;
5383 wm
->pipe
[PIPE_B
].plane
[PLANE_PRIMARY
] |= _FW_WM(tmp
, PLANEB_HI
) << 8;
5384 wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE1
] |= _FW_WM(tmp
, SPRITEB_HI
) << 8;
5385 wm
->pipe
[PIPE_A
].plane
[PLANE_SPRITE0
] |= _FW_WM(tmp
, SPRITEA_HI
) << 8;
5386 wm
->pipe
[PIPE_A
].plane
[PLANE_PRIMARY
] |= _FW_WM(tmp
, PLANEA_HI
) << 8;
5393 void g4x_wm_get_hw_state(struct drm_device
*dev
)
5395 struct drm_i915_private
*dev_priv
= to_i915(dev
);
5396 struct g4x_wm_values
*wm
= &dev_priv
->wm
.g4x
;
5397 struct intel_crtc
*crtc
;
5399 g4x_read_wm_values(dev_priv
, wm
);
5401 wm
->cxsr
= I915_READ(FW_BLC_SELF
) & FW_BLC_SELF_EN
;
5403 for_each_intel_crtc(dev
, crtc
) {
5404 struct intel_crtc_state
*crtc_state
=
5405 to_intel_crtc_state(crtc
->base
.state
);
5406 struct g4x_wm_state
*active
= &crtc
->wm
.active
.g4x
;
5407 struct g4x_pipe_wm
*raw
;
5408 enum pipe pipe
= crtc
->pipe
;
5409 enum plane_id plane_id
;
5410 int level
, max_level
;
5412 active
->cxsr
= wm
->cxsr
;
5413 active
->hpll_en
= wm
->hpll_en
;
5414 active
->fbc_en
= wm
->fbc_en
;
5416 active
->sr
= wm
->sr
;
5417 active
->hpll
= wm
->hpll
;
5419 for_each_plane_id_on_crtc(crtc
, plane_id
) {
5420 active
->wm
.plane
[plane_id
] =
5421 wm
->pipe
[pipe
].plane
[plane_id
];
5424 if (wm
->cxsr
&& wm
->hpll_en
)
5425 max_level
= G4X_WM_LEVEL_HPLL
;
5427 max_level
= G4X_WM_LEVEL_SR
;
5429 max_level
= G4X_WM_LEVEL_NORMAL
;
5431 level
= G4X_WM_LEVEL_NORMAL
;
5432 raw
= &crtc_state
->wm
.g4x
.raw
[level
];
5433 for_each_plane_id_on_crtc(crtc
, plane_id
)
5434 raw
->plane
[plane_id
] = active
->wm
.plane
[plane_id
];
5436 if (++level
> max_level
)
5439 raw
= &crtc_state
->wm
.g4x
.raw
[level
];
5440 raw
->plane
[PLANE_PRIMARY
] = active
->sr
.plane
;
5441 raw
->plane
[PLANE_CURSOR
] = active
->sr
.cursor
;
5442 raw
->plane
[PLANE_SPRITE0
] = 0;
5443 raw
->fbc
= active
->sr
.fbc
;
5445 if (++level
> max_level
)
5448 raw
= &crtc_state
->wm
.g4x
.raw
[level
];
5449 raw
->plane
[PLANE_PRIMARY
] = active
->hpll
.plane
;
5450 raw
->plane
[PLANE_CURSOR
] = active
->hpll
.cursor
;
5451 raw
->plane
[PLANE_SPRITE0
] = 0;
5452 raw
->fbc
= active
->hpll
.fbc
;
5455 for_each_plane_id_on_crtc(crtc
, plane_id
)
5456 g4x_raw_plane_wm_set(crtc_state
, level
,
5457 plane_id
, USHRT_MAX
);
5458 g4x_raw_fbc_wm_set(crtc_state
, level
, USHRT_MAX
);
5460 crtc_state
->wm
.g4x
.optimal
= *active
;
5461 crtc_state
->wm
.g4x
.intermediate
= *active
;
5463 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite=%d\n",
5465 wm
->pipe
[pipe
].plane
[PLANE_PRIMARY
],
5466 wm
->pipe
[pipe
].plane
[PLANE_CURSOR
],
5467 wm
->pipe
[pipe
].plane
[PLANE_SPRITE0
]);
5470 DRM_DEBUG_KMS("Initial SR watermarks: plane=%d, cursor=%d fbc=%d\n",
5471 wm
->sr
.plane
, wm
->sr
.cursor
, wm
->sr
.fbc
);
5472 DRM_DEBUG_KMS("Initial HPLL watermarks: plane=%d, SR cursor=%d fbc=%d\n",
5473 wm
->hpll
.plane
, wm
->hpll
.cursor
, wm
->hpll
.fbc
);
5474 DRM_DEBUG_KMS("Initial SR=%s HPLL=%s FBC=%s\n",
5475 yesno(wm
->cxsr
), yesno(wm
->hpll_en
), yesno(wm
->fbc_en
));
5478 void g4x_wm_sanitize(struct drm_i915_private
*dev_priv
)
5480 struct intel_plane
*plane
;
5481 struct intel_crtc
*crtc
;
5483 mutex_lock(&dev_priv
->wm
.wm_mutex
);
5485 for_each_intel_plane(&dev_priv
->drm
, plane
) {
5486 struct intel_crtc
*crtc
=
5487 intel_get_crtc_for_pipe(dev_priv
, plane
->pipe
);
5488 struct intel_crtc_state
*crtc_state
=
5489 to_intel_crtc_state(crtc
->base
.state
);
5490 struct intel_plane_state
*plane_state
=
5491 to_intel_plane_state(plane
->base
.state
);
5492 struct g4x_wm_state
*wm_state
= &crtc_state
->wm
.g4x
.optimal
;
5493 enum plane_id plane_id
= plane
->id
;
5496 if (plane_state
->base
.visible
)
5499 for (level
= 0; level
< 3; level
++) {
5500 struct g4x_pipe_wm
*raw
=
5501 &crtc_state
->wm
.g4x
.raw
[level
];
5503 raw
->plane
[plane_id
] = 0;
5504 wm_state
->wm
.plane
[plane_id
] = 0;
5507 if (plane_id
== PLANE_PRIMARY
) {
5508 for (level
= 0; level
< 3; level
++) {
5509 struct g4x_pipe_wm
*raw
=
5510 &crtc_state
->wm
.g4x
.raw
[level
];
5514 wm_state
->sr
.fbc
= 0;
5515 wm_state
->hpll
.fbc
= 0;
5516 wm_state
->fbc_en
= false;
5520 for_each_intel_crtc(&dev_priv
->drm
, crtc
) {
5521 struct intel_crtc_state
*crtc_state
=
5522 to_intel_crtc_state(crtc
->base
.state
);
5524 crtc_state
->wm
.g4x
.intermediate
=
5525 crtc_state
->wm
.g4x
.optimal
;
5526 crtc
->wm
.active
.g4x
= crtc_state
->wm
.g4x
.optimal
;
5529 g4x_program_watermarks(dev_priv
);
5531 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
5534 void vlv_wm_get_hw_state(struct drm_device
*dev
)
5536 struct drm_i915_private
*dev_priv
= to_i915(dev
);
5537 struct vlv_wm_values
*wm
= &dev_priv
->wm
.vlv
;
5538 struct intel_crtc
*crtc
;
5541 vlv_read_wm_values(dev_priv
, wm
);
5543 wm
->cxsr
= I915_READ(FW_BLC_SELF_VLV
) & FW_CSPWRDWNEN
;
5544 wm
->level
= VLV_WM_LEVEL_PM2
;
5546 if (IS_CHERRYVIEW(dev_priv
)) {
5547 mutex_lock(&dev_priv
->rps
.hw_lock
);
5549 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DSPFREQ
);
5550 if (val
& DSP_MAXFIFO_PM5_ENABLE
)
5551 wm
->level
= VLV_WM_LEVEL_PM5
;
5554 * If DDR DVFS is disabled in the BIOS, Punit
5555 * will never ack the request. So if that happens
5556 * assume we don't have to enable/disable DDR DVFS
5557 * dynamically. To test that just set the REQ_ACK
5558 * bit to poke the Punit, but don't change the
5559 * HIGH/LOW bits so that we don't actually change
5560 * the current state.
5562 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
);
5563 val
|= FORCE_DDR_FREQ_REQ_ACK
;
5564 vlv_punit_write(dev_priv
, PUNIT_REG_DDR_SETUP2
, val
);
5566 if (wait_for((vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
) &
5567 FORCE_DDR_FREQ_REQ_ACK
) == 0, 3)) {
5568 DRM_DEBUG_KMS("Punit not acking DDR DVFS request, "
5569 "assuming DDR DVFS is disabled\n");
5570 dev_priv
->wm
.max_level
= VLV_WM_LEVEL_PM5
;
5572 val
= vlv_punit_read(dev_priv
, PUNIT_REG_DDR_SETUP2
);
5573 if ((val
& FORCE_DDR_HIGH_FREQ
) == 0)
5574 wm
->level
= VLV_WM_LEVEL_DDR_DVFS
;
5577 mutex_unlock(&dev_priv
->rps
.hw_lock
);
5580 for_each_intel_crtc(dev
, crtc
) {
5581 struct intel_crtc_state
*crtc_state
=
5582 to_intel_crtc_state(crtc
->base
.state
);
5583 struct vlv_wm_state
*active
= &crtc
->wm
.active
.vlv
;
5584 const struct vlv_fifo_state
*fifo_state
=
5585 &crtc_state
->wm
.vlv
.fifo_state
;
5586 enum pipe pipe
= crtc
->pipe
;
5587 enum plane_id plane_id
;
5590 vlv_get_fifo_size(crtc_state
);
5592 active
->num_levels
= wm
->level
+ 1;
5593 active
->cxsr
= wm
->cxsr
;
5595 for (level
= 0; level
< active
->num_levels
; level
++) {
5596 struct g4x_pipe_wm
*raw
=
5597 &crtc_state
->wm
.vlv
.raw
[level
];
5599 active
->sr
[level
].plane
= wm
->sr
.plane
;
5600 active
->sr
[level
].cursor
= wm
->sr
.cursor
;
5602 for_each_plane_id_on_crtc(crtc
, plane_id
) {
5603 active
->wm
[level
].plane
[plane_id
] =
5604 wm
->pipe
[pipe
].plane
[plane_id
];
5606 raw
->plane
[plane_id
] =
5607 vlv_invert_wm_value(active
->wm
[level
].plane
[plane_id
],
5608 fifo_state
->plane
[plane_id
]);
5612 for_each_plane_id_on_crtc(crtc
, plane_id
)
5613 vlv_raw_plane_wm_set(crtc_state
, level
,
5614 plane_id
, USHRT_MAX
);
5615 vlv_invalidate_wms(crtc
, active
, level
);
5617 crtc_state
->wm
.vlv
.optimal
= *active
;
5618 crtc_state
->wm
.vlv
.intermediate
= *active
;
5620 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
5622 wm
->pipe
[pipe
].plane
[PLANE_PRIMARY
],
5623 wm
->pipe
[pipe
].plane
[PLANE_CURSOR
],
5624 wm
->pipe
[pipe
].plane
[PLANE_SPRITE0
],
5625 wm
->pipe
[pipe
].plane
[PLANE_SPRITE1
]);
5628 DRM_DEBUG_KMS("Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
5629 wm
->sr
.plane
, wm
->sr
.cursor
, wm
->level
, wm
->cxsr
);
5632 void vlv_wm_sanitize(struct drm_i915_private
*dev_priv
)
5634 struct intel_plane
*plane
;
5635 struct intel_crtc
*crtc
;
5637 mutex_lock(&dev_priv
->wm
.wm_mutex
);
5639 for_each_intel_plane(&dev_priv
->drm
, plane
) {
5640 struct intel_crtc
*crtc
=
5641 intel_get_crtc_for_pipe(dev_priv
, plane
->pipe
);
5642 struct intel_crtc_state
*crtc_state
=
5643 to_intel_crtc_state(crtc
->base
.state
);
5644 struct intel_plane_state
*plane_state
=
5645 to_intel_plane_state(plane
->base
.state
);
5646 struct vlv_wm_state
*wm_state
= &crtc_state
->wm
.vlv
.optimal
;
5647 const struct vlv_fifo_state
*fifo_state
=
5648 &crtc_state
->wm
.vlv
.fifo_state
;
5649 enum plane_id plane_id
= plane
->id
;
5652 if (plane_state
->base
.visible
)
5655 for (level
= 0; level
< wm_state
->num_levels
; level
++) {
5656 struct g4x_pipe_wm
*raw
=
5657 &crtc_state
->wm
.vlv
.raw
[level
];
5659 raw
->plane
[plane_id
] = 0;
5661 wm_state
->wm
[level
].plane
[plane_id
] =
5662 vlv_invert_wm_value(raw
->plane
[plane_id
],
5663 fifo_state
->plane
[plane_id
]);
5667 for_each_intel_crtc(&dev_priv
->drm
, crtc
) {
5668 struct intel_crtc_state
*crtc_state
=
5669 to_intel_crtc_state(crtc
->base
.state
);
5671 crtc_state
->wm
.vlv
.intermediate
=
5672 crtc_state
->wm
.vlv
.optimal
;
5673 crtc
->wm
.active
.vlv
= crtc_state
->wm
.vlv
.optimal
;
5676 vlv_program_watermarks(dev_priv
);
5678 mutex_unlock(&dev_priv
->wm
.wm_mutex
);
5681 void ilk_wm_get_hw_state(struct drm_device
*dev
)
5683 struct drm_i915_private
*dev_priv
= to_i915(dev
);
5684 struct ilk_wm_values
*hw
= &dev_priv
->wm
.hw
;
5685 struct drm_crtc
*crtc
;
5687 for_each_crtc(dev
, crtc
)
5688 ilk_pipe_wm_get_hw_state(crtc
);
5690 hw
->wm_lp
[0] = I915_READ(WM1_LP_ILK
);
5691 hw
->wm_lp
[1] = I915_READ(WM2_LP_ILK
);
5692 hw
->wm_lp
[2] = I915_READ(WM3_LP_ILK
);
5694 hw
->wm_lp_spr
[0] = I915_READ(WM1S_LP_ILK
);
5695 if (INTEL_GEN(dev_priv
) >= 7) {
5696 hw
->wm_lp_spr
[1] = I915_READ(WM2S_LP_IVB
);
5697 hw
->wm_lp_spr
[2] = I915_READ(WM3S_LP_IVB
);
5700 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
5701 hw
->partitioning
= (I915_READ(WM_MISC
) & WM_MISC_DATA_PARTITION_5_6
) ?
5702 INTEL_DDB_PART_5_6
: INTEL_DDB_PART_1_2
;
5703 else if (IS_IVYBRIDGE(dev_priv
))
5704 hw
->partitioning
= (I915_READ(DISP_ARB_CTL2
) & DISP_DATA_PARTITION_5_6
) ?
5705 INTEL_DDB_PART_5_6
: INTEL_DDB_PART_1_2
;
5708 !(I915_READ(DISP_ARB_CTL
) & DISP_FBC_WM_DIS
);
5712 * intel_update_watermarks - update FIFO watermark values based on current modes
5714 * Calculate watermark values for the various WM regs based on current mode
5715 * and plane configuration.
5717 * There are several cases to deal with here:
5718 * - normal (i.e. non-self-refresh)
5719 * - self-refresh (SR) mode
5720 * - lines are large relative to FIFO size (buffer can hold up to 2)
5721 * - lines are small relative to FIFO size (buffer can hold more than 2
5722 * lines), so need to account for TLB latency
5724 * The normal calculation is:
5725 * watermark = dotclock * bytes per pixel * latency
5726 * where latency is platform & configuration dependent (we assume pessimal
5729 * The SR calculation is:
5730 * watermark = (trunc(latency/line time)+1) * surface width *
5733 * line time = htotal / dotclock
5734 * surface width = hdisplay for normal plane and 64 for cursor
5735 * and latency is assumed to be high, as above.
5737 * The final value programmed to the register should always be rounded up,
5738 * and include an extra 2 entries to account for clock crossings.
5740 * We don't use the sprite, so we can ignore that. And on Crestline we have
5741 * to set the non-SR watermarks to 8.
5743 void intel_update_watermarks(struct intel_crtc
*crtc
)
5745 struct drm_i915_private
*dev_priv
= to_i915(crtc
->base
.dev
);
5747 if (dev_priv
->display
.update_wm
)
5748 dev_priv
->display
.update_wm(crtc
);
5752 * Lock protecting IPS related data structures
5754 DEFINE_SPINLOCK(mchdev_lock
);
5756 /* Global for IPS driver to get at the current i915 device. Protected by
5758 static struct drm_i915_private
*i915_mch_dev
;
5760 bool ironlake_set_drps(struct drm_i915_private
*dev_priv
, u8 val
)
5764 lockdep_assert_held(&mchdev_lock
);
5766 rgvswctl
= I915_READ16(MEMSWCTL
);
5767 if (rgvswctl
& MEMCTL_CMD_STS
) {
5768 DRM_DEBUG("gpu busy, RCS change rejected\n");
5769 return false; /* still busy with another command */
5772 rgvswctl
= (MEMCTL_CMD_CHFREQ
<< MEMCTL_CMD_SHIFT
) |
5773 (val
<< MEMCTL_FREQ_SHIFT
) | MEMCTL_SFCAVM
;
5774 I915_WRITE16(MEMSWCTL
, rgvswctl
);
5775 POSTING_READ16(MEMSWCTL
);
5777 rgvswctl
|= MEMCTL_CMD_STS
;
5778 I915_WRITE16(MEMSWCTL
, rgvswctl
);
5783 static void ironlake_enable_drps(struct drm_i915_private
*dev_priv
)
5786 u8 fmax
, fmin
, fstart
, vstart
;
5788 spin_lock_irq(&mchdev_lock
);
5790 rgvmodectl
= I915_READ(MEMMODECTL
);
5792 /* Enable temp reporting */
5793 I915_WRITE16(PMMISC
, I915_READ(PMMISC
) | MCPPCE_EN
);
5794 I915_WRITE16(TSC1
, I915_READ(TSC1
) | TSE
);
5796 /* 100ms RC evaluation intervals */
5797 I915_WRITE(RCUPEI
, 100000);
5798 I915_WRITE(RCDNEI
, 100000);
5800 /* Set max/min thresholds to 90ms and 80ms respectively */
5801 I915_WRITE(RCBMAXAVG
, 90000);
5802 I915_WRITE(RCBMINAVG
, 80000);
5804 I915_WRITE(MEMIHYST
, 1);
5806 /* Set up min, max, and cur for interrupt handling */
5807 fmax
= (rgvmodectl
& MEMMODE_FMAX_MASK
) >> MEMMODE_FMAX_SHIFT
;
5808 fmin
= (rgvmodectl
& MEMMODE_FMIN_MASK
);
5809 fstart
= (rgvmodectl
& MEMMODE_FSTART_MASK
) >>
5810 MEMMODE_FSTART_SHIFT
;
5812 vstart
= (I915_READ(PXVFREQ(fstart
)) & PXVFREQ_PX_MASK
) >>
5815 dev_priv
->ips
.fmax
= fmax
; /* IPS callback will increase this */
5816 dev_priv
->ips
.fstart
= fstart
;
5818 dev_priv
->ips
.max_delay
= fstart
;
5819 dev_priv
->ips
.min_delay
= fmin
;
5820 dev_priv
->ips
.cur_delay
= fstart
;
5822 DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
5823 fmax
, fmin
, fstart
);
5825 I915_WRITE(MEMINTREN
, MEMINT_CX_SUPR_EN
| MEMINT_EVAL_CHG_EN
);
5828 * Interrupts will be enabled in ironlake_irq_postinstall
5831 I915_WRITE(VIDSTART
, vstart
);
5832 POSTING_READ(VIDSTART
);
5834 rgvmodectl
|= MEMMODE_SWMODE_EN
;
5835 I915_WRITE(MEMMODECTL
, rgvmodectl
);
5837 if (wait_for_atomic((I915_READ(MEMSWCTL
) & MEMCTL_CMD_STS
) == 0, 10))
5838 DRM_ERROR("stuck trying to change perf mode\n");
5841 ironlake_set_drps(dev_priv
, fstart
);
5843 dev_priv
->ips
.last_count1
= I915_READ(DMIEC
) +
5844 I915_READ(DDREC
) + I915_READ(CSIEC
);
5845 dev_priv
->ips
.last_time1
= jiffies_to_msecs(jiffies
);
5846 dev_priv
->ips
.last_count2
= I915_READ(GFXEC
);
5847 dev_priv
->ips
.last_time2
= ktime_get_raw_ns();
5849 spin_unlock_irq(&mchdev_lock
);
5852 static void ironlake_disable_drps(struct drm_i915_private
*dev_priv
)
5856 spin_lock_irq(&mchdev_lock
);
5858 rgvswctl
= I915_READ16(MEMSWCTL
);
5860 /* Ack interrupts, disable EFC interrupt */
5861 I915_WRITE(MEMINTREN
, I915_READ(MEMINTREN
) & ~MEMINT_EVAL_CHG_EN
);
5862 I915_WRITE(MEMINTRSTS
, MEMINT_EVAL_CHG
);
5863 I915_WRITE(DEIER
, I915_READ(DEIER
) & ~DE_PCU_EVENT
);
5864 I915_WRITE(DEIIR
, DE_PCU_EVENT
);
5865 I915_WRITE(DEIMR
, I915_READ(DEIMR
) | DE_PCU_EVENT
);
5867 /* Go back to the starting frequency */
5868 ironlake_set_drps(dev_priv
, dev_priv
->ips
.fstart
);
5870 rgvswctl
|= MEMCTL_CMD_STS
;
5871 I915_WRITE(MEMSWCTL
, rgvswctl
);
5874 spin_unlock_irq(&mchdev_lock
);
5877 /* There's a funny hw issue where the hw returns all 0 when reading from
5878 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
5879 * ourselves, instead of doing a rmw cycle (which might result in us clearing
5880 * all limits and the gpu stuck at whatever frequency it is at atm).
5882 static u32
intel_rps_limits(struct drm_i915_private
*dev_priv
, u8 val
)
5886 /* Only set the down limit when we've reached the lowest level to avoid
5887 * getting more interrupts, otherwise leave this clear. This prevents a
5888 * race in the hw when coming out of rc6: There's a tiny window where
5889 * the hw runs at the minimal clock before selecting the desired
5890 * frequency, if the down threshold expires in that window we will not
5891 * receive a down interrupt. */
5892 if (INTEL_GEN(dev_priv
) >= 9) {
5893 limits
= (dev_priv
->rps
.max_freq_softlimit
) << 23;
5894 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
5895 limits
|= (dev_priv
->rps
.min_freq_softlimit
) << 14;
5897 limits
= dev_priv
->rps
.max_freq_softlimit
<< 24;
5898 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
5899 limits
|= dev_priv
->rps
.min_freq_softlimit
<< 16;
5905 static void gen6_set_rps_thresholds(struct drm_i915_private
*dev_priv
, u8 val
)
5908 u32 threshold_up
= 0, threshold_down
= 0; /* in % */
5909 u32 ei_up
= 0, ei_down
= 0;
5911 new_power
= dev_priv
->rps
.power
;
5912 switch (dev_priv
->rps
.power
) {
5914 if (val
> dev_priv
->rps
.efficient_freq
+ 1 &&
5915 val
> dev_priv
->rps
.cur_freq
)
5916 new_power
= BETWEEN
;
5920 if (val
<= dev_priv
->rps
.efficient_freq
&&
5921 val
< dev_priv
->rps
.cur_freq
)
5922 new_power
= LOW_POWER
;
5923 else if (val
>= dev_priv
->rps
.rp0_freq
&&
5924 val
> dev_priv
->rps
.cur_freq
)
5925 new_power
= HIGH_POWER
;
5929 if (val
< (dev_priv
->rps
.rp1_freq
+ dev_priv
->rps
.rp0_freq
) >> 1 &&
5930 val
< dev_priv
->rps
.cur_freq
)
5931 new_power
= BETWEEN
;
5934 /* Max/min bins are special */
5935 if (val
<= dev_priv
->rps
.min_freq_softlimit
)
5936 new_power
= LOW_POWER
;
5937 if (val
>= dev_priv
->rps
.max_freq_softlimit
)
5938 new_power
= HIGH_POWER
;
5939 if (new_power
== dev_priv
->rps
.power
)
5942 /* Note the units here are not exactly 1us, but 1280ns. */
5943 switch (new_power
) {
5945 /* Upclock if more than 95% busy over 16ms */
5949 /* Downclock if less than 85% busy over 32ms */
5951 threshold_down
= 85;
5955 /* Upclock if more than 90% busy over 13ms */
5959 /* Downclock if less than 75% busy over 32ms */
5961 threshold_down
= 75;
5965 /* Upclock if more than 85% busy over 10ms */
5969 /* Downclock if less than 60% busy over 32ms */
5971 threshold_down
= 60;
5975 /* When byt can survive without system hang with dynamic
5976 * sw freq adjustments, this restriction can be lifted.
5978 if (IS_VALLEYVIEW(dev_priv
))
5981 I915_WRITE(GEN6_RP_UP_EI
,
5982 GT_INTERVAL_FROM_US(dev_priv
, ei_up
));
5983 I915_WRITE(GEN6_RP_UP_THRESHOLD
,
5984 GT_INTERVAL_FROM_US(dev_priv
,
5985 ei_up
* threshold_up
/ 100));
5987 I915_WRITE(GEN6_RP_DOWN_EI
,
5988 GT_INTERVAL_FROM_US(dev_priv
, ei_down
));
5989 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
,
5990 GT_INTERVAL_FROM_US(dev_priv
,
5991 ei_down
* threshold_down
/ 100));
5993 I915_WRITE(GEN6_RP_CONTROL
,
5994 GEN6_RP_MEDIA_TURBO
|
5995 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
5996 GEN6_RP_MEDIA_IS_GFX
|
5998 GEN6_RP_UP_BUSY_AVG
|
5999 GEN6_RP_DOWN_IDLE_AVG
);
6002 dev_priv
->rps
.power
= new_power
;
6003 dev_priv
->rps
.up_threshold
= threshold_up
;
6004 dev_priv
->rps
.down_threshold
= threshold_down
;
6005 dev_priv
->rps
.last_adj
= 0;
6008 static u32
gen6_rps_pm_mask(struct drm_i915_private
*dev_priv
, u8 val
)
6012 /* We use UP_EI_EXPIRED interupts for both up/down in manual mode */
6013 if (val
> dev_priv
->rps
.min_freq_softlimit
)
6014 mask
|= GEN6_PM_RP_UP_EI_EXPIRED
| GEN6_PM_RP_DOWN_THRESHOLD
| GEN6_PM_RP_DOWN_TIMEOUT
;
6015 if (val
< dev_priv
->rps
.max_freq_softlimit
)
6016 mask
|= GEN6_PM_RP_UP_EI_EXPIRED
| GEN6_PM_RP_UP_THRESHOLD
;
6018 mask
&= dev_priv
->pm_rps_events
;
6020 return gen6_sanitize_rps_pm_mask(dev_priv
, ~mask
);
6023 /* gen6_set_rps is called to update the frequency request, but should also be
6024 * called when the range (min_delay and max_delay) is modified so that we can
6025 * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
6026 static int gen6_set_rps(struct drm_i915_private
*dev_priv
, u8 val
)
6028 /* min/max delay may still have been modified so be sure to
6029 * write the limits value.
6031 if (val
!= dev_priv
->rps
.cur_freq
) {
6032 gen6_set_rps_thresholds(dev_priv
, val
);
6034 if (INTEL_GEN(dev_priv
) >= 9)
6035 I915_WRITE(GEN6_RPNSWREQ
,
6036 GEN9_FREQUENCY(val
));
6037 else if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
6038 I915_WRITE(GEN6_RPNSWREQ
,
6039 HSW_FREQUENCY(val
));
6041 I915_WRITE(GEN6_RPNSWREQ
,
6042 GEN6_FREQUENCY(val
) |
6044 GEN6_AGGRESSIVE_TURBO
);
6047 /* Make sure we continue to get interrupts
6048 * until we hit the minimum or maximum frequencies.
6050 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS
, intel_rps_limits(dev_priv
, val
));
6051 I915_WRITE(GEN6_PMINTRMSK
, gen6_rps_pm_mask(dev_priv
, val
));
6053 dev_priv
->rps
.cur_freq
= val
;
6054 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv
, val
));
6059 static int valleyview_set_rps(struct drm_i915_private
*dev_priv
, u8 val
)
6063 if (WARN_ONCE(IS_CHERRYVIEW(dev_priv
) && (val
& 1),
6064 "Odd GPU freq value\n"))
6067 I915_WRITE(GEN6_PMINTRMSK
, gen6_rps_pm_mask(dev_priv
, val
));
6069 if (val
!= dev_priv
->rps
.cur_freq
) {
6070 err
= vlv_punit_write(dev_priv
, PUNIT_REG_GPU_FREQ_REQ
, val
);
6074 gen6_set_rps_thresholds(dev_priv
, val
);
6077 dev_priv
->rps
.cur_freq
= val
;
6078 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv
, val
));
6083 /* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
6085 * * If Gfx is Idle, then
6086 * 1. Forcewake Media well.
6087 * 2. Request idle freq.
6088 * 3. Release Forcewake of Media well.
6090 static void vlv_set_rps_idle(struct drm_i915_private
*dev_priv
)
6092 u32 val
= dev_priv
->rps
.idle_freq
;
6095 if (dev_priv
->rps
.cur_freq
<= val
)
6098 /* The punit delays the write of the frequency and voltage until it
6099 * determines the GPU is awake. During normal usage we don't want to
6100 * waste power changing the frequency if the GPU is sleeping (rc6).
6101 * However, the GPU and driver is now idle and we do not want to delay
6102 * switching to minimum voltage (reducing power whilst idle) as we do
6103 * not expect to be woken in the near future and so must flush the
6104 * change by waking the device.
6106 * We choose to take the media powerwell (either would do to trick the
6107 * punit into committing the voltage change) as that takes a lot less
6108 * power than the render powerwell.
6110 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_MEDIA
);
6111 err
= valleyview_set_rps(dev_priv
, val
);
6112 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_MEDIA
);
6115 DRM_ERROR("Failed to set RPS for idle\n");
6118 void gen6_rps_busy(struct drm_i915_private
*dev_priv
)
6120 mutex_lock(&dev_priv
->rps
.hw_lock
);
6121 if (dev_priv
->rps
.enabled
) {
6124 if (dev_priv
->pm_rps_events
& GEN6_PM_RP_UP_EI_EXPIRED
)
6125 gen6_rps_reset_ei(dev_priv
);
6126 I915_WRITE(GEN6_PMINTRMSK
,
6127 gen6_rps_pm_mask(dev_priv
, dev_priv
->rps
.cur_freq
));
6129 gen6_enable_rps_interrupts(dev_priv
);
6131 /* Use the user's desired frequency as a guide, but for better
6132 * performance, jump directly to RPe as our starting frequency.
6134 freq
= max(dev_priv
->rps
.cur_freq
,
6135 dev_priv
->rps
.efficient_freq
);
6137 if (intel_set_rps(dev_priv
,
6139 dev_priv
->rps
.min_freq_softlimit
,
6140 dev_priv
->rps
.max_freq_softlimit
)))
6141 DRM_DEBUG_DRIVER("Failed to set idle frequency\n");
6143 mutex_unlock(&dev_priv
->rps
.hw_lock
);
6146 void gen6_rps_idle(struct drm_i915_private
*dev_priv
)
6148 /* Flush our bottom-half so that it does not race with us
6149 * setting the idle frequency and so that it is bounded by
6150 * our rpm wakeref. And then disable the interrupts to stop any
6151 * futher RPS reclocking whilst we are asleep.
6153 gen6_disable_rps_interrupts(dev_priv
);
6155 mutex_lock(&dev_priv
->rps
.hw_lock
);
6156 if (dev_priv
->rps
.enabled
) {
6157 if (IS_VALLEYVIEW(dev_priv
) || IS_CHERRYVIEW(dev_priv
))
6158 vlv_set_rps_idle(dev_priv
);
6160 gen6_set_rps(dev_priv
, dev_priv
->rps
.idle_freq
);
6161 dev_priv
->rps
.last_adj
= 0;
6162 I915_WRITE(GEN6_PMINTRMSK
,
6163 gen6_sanitize_rps_pm_mask(dev_priv
, ~0));
6165 mutex_unlock(&dev_priv
->rps
.hw_lock
);
6168 void gen6_rps_boost(struct drm_i915_gem_request
*rq
,
6169 struct intel_rps_client
*rps
)
6171 struct drm_i915_private
*i915
= rq
->i915
;
6172 unsigned long flags
;
6175 /* This is intentionally racy! We peek at the state here, then
6176 * validate inside the RPS worker.
6178 if (!i915
->rps
.enabled
)
6182 spin_lock_irqsave(&rq
->lock
, flags
);
6183 if (!rq
->waitboost
&& !i915_gem_request_completed(rq
)) {
6184 atomic_inc(&i915
->rps
.num_waiters
);
6185 rq
->waitboost
= true;
6188 spin_unlock_irqrestore(&rq
->lock
, flags
);
6192 if (READ_ONCE(i915
->rps
.cur_freq
) < i915
->rps
.boost_freq
)
6193 schedule_work(&i915
->rps
.work
);
6195 atomic_inc(rps
? &rps
->boosts
: &i915
->rps
.boosts
);
6198 int intel_set_rps(struct drm_i915_private
*dev_priv
, u8 val
)
6202 lockdep_assert_held(&dev_priv
->rps
.hw_lock
);
6203 GEM_BUG_ON(val
> dev_priv
->rps
.max_freq
);
6204 GEM_BUG_ON(val
< dev_priv
->rps
.min_freq
);
6206 if (!dev_priv
->rps
.enabled
) {
6207 dev_priv
->rps
.cur_freq
= val
;
6211 if (IS_VALLEYVIEW(dev_priv
) || IS_CHERRYVIEW(dev_priv
))
6212 err
= valleyview_set_rps(dev_priv
, val
);
6214 err
= gen6_set_rps(dev_priv
, val
);
6219 static void gen9_disable_rc6(struct drm_i915_private
*dev_priv
)
6221 I915_WRITE(GEN6_RC_CONTROL
, 0);
6222 I915_WRITE(GEN9_PG_ENABLE
, 0);
6225 static void gen9_disable_rps(struct drm_i915_private
*dev_priv
)
6227 I915_WRITE(GEN6_RP_CONTROL
, 0);
6230 static void gen6_disable_rps(struct drm_i915_private
*dev_priv
)
6232 I915_WRITE(GEN6_RC_CONTROL
, 0);
6233 I915_WRITE(GEN6_RPNSWREQ
, 1 << 31);
6234 I915_WRITE(GEN6_RP_CONTROL
, 0);
6237 static void cherryview_disable_rps(struct drm_i915_private
*dev_priv
)
6239 I915_WRITE(GEN6_RC_CONTROL
, 0);
6242 static void valleyview_disable_rps(struct drm_i915_private
*dev_priv
)
6244 /* we're doing forcewake before Disabling RC6,
6245 * This what the BIOS expects when going into suspend */
6246 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
6248 I915_WRITE(GEN6_RC_CONTROL
, 0);
6250 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
6253 static void intel_print_rc6_info(struct drm_i915_private
*dev_priv
, u32 mode
)
6255 if (IS_VALLEYVIEW(dev_priv
) || IS_CHERRYVIEW(dev_priv
)) {
6256 if (mode
& (GEN7_RC_CTL_TO_MODE
| GEN6_RC_CTL_EI_MODE(1)))
6257 mode
= GEN6_RC_CTL_RC6_ENABLE
;
6261 if (HAS_RC6p(dev_priv
))
6262 DRM_DEBUG_DRIVER("Enabling RC6 states: "
6263 "RC6 %s RC6p %s RC6pp %s\n",
6264 onoff(mode
& GEN6_RC_CTL_RC6_ENABLE
),
6265 onoff(mode
& GEN6_RC_CTL_RC6p_ENABLE
),
6266 onoff(mode
& GEN6_RC_CTL_RC6pp_ENABLE
));
6269 DRM_DEBUG_DRIVER("Enabling RC6 states: RC6 %s\n",
6270 onoff(mode
& GEN6_RC_CTL_RC6_ENABLE
));
6273 static bool bxt_check_bios_rc6_setup(struct drm_i915_private
*dev_priv
)
6275 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
6276 bool enable_rc6
= true;
6277 unsigned long rc6_ctx_base
;
6281 rc_ctl
= I915_READ(GEN6_RC_CONTROL
);
6282 rc_sw_target
= (I915_READ(GEN6_RC_STATE
) & RC_SW_TARGET_STATE_MASK
) >>
6283 RC_SW_TARGET_STATE_SHIFT
;
6284 DRM_DEBUG_DRIVER("BIOS enabled RC states: "
6285 "HW_CTRL %s HW_RC6 %s SW_TARGET_STATE %x\n",
6286 onoff(rc_ctl
& GEN6_RC_CTL_HW_ENABLE
),
6287 onoff(rc_ctl
& GEN6_RC_CTL_RC6_ENABLE
),
6290 if (!(I915_READ(RC6_LOCATION
) & RC6_CTX_IN_DRAM
)) {
6291 DRM_DEBUG_DRIVER("RC6 Base location not set properly.\n");
6296 * The exact context size is not known for BXT, so assume a page size
6299 rc6_ctx_base
= I915_READ(RC6_CTX_BASE
) & RC6_CTX_BASE_MASK
;
6300 if (!((rc6_ctx_base
>= ggtt
->stolen_reserved_base
) &&
6301 (rc6_ctx_base
+ PAGE_SIZE
<= ggtt
->stolen_reserved_base
+
6302 ggtt
->stolen_reserved_size
))) {
6303 DRM_DEBUG_DRIVER("RC6 Base address not as expected.\n");
6307 if (!(((I915_READ(PWRCTX_MAXCNT_RCSUNIT
) & IDLE_TIME_MASK
) > 1) &&
6308 ((I915_READ(PWRCTX_MAXCNT_VCSUNIT0
) & IDLE_TIME_MASK
) > 1) &&
6309 ((I915_READ(PWRCTX_MAXCNT_BCSUNIT
) & IDLE_TIME_MASK
) > 1) &&
6310 ((I915_READ(PWRCTX_MAXCNT_VECSUNIT
) & IDLE_TIME_MASK
) > 1))) {
6311 DRM_DEBUG_DRIVER("Engine Idle wait time not set properly.\n");
6315 if (!I915_READ(GEN8_PUSHBUS_CONTROL
) ||
6316 !I915_READ(GEN8_PUSHBUS_ENABLE
) ||
6317 !I915_READ(GEN8_PUSHBUS_SHIFT
)) {
6318 DRM_DEBUG_DRIVER("Pushbus not setup properly.\n");
6322 if (!I915_READ(GEN6_GFXPAUSE
)) {
6323 DRM_DEBUG_DRIVER("GFX pause not setup properly.\n");
6327 if (!I915_READ(GEN8_MISC_CTRL0
)) {
6328 DRM_DEBUG_DRIVER("GPM control not setup properly.\n");
6335 int sanitize_rc6_option(struct drm_i915_private
*dev_priv
, int enable_rc6
)
6337 /* No RC6 before Ironlake and code is gone for ilk. */
6338 if (INTEL_INFO(dev_priv
)->gen
< 6)
6344 if (IS_GEN9_LP(dev_priv
) && !bxt_check_bios_rc6_setup(dev_priv
)) {
6345 DRM_INFO("RC6 disabled by BIOS\n");
6349 /* Respect the kernel parameter if it is set */
6350 if (enable_rc6
>= 0) {
6353 if (HAS_RC6p(dev_priv
))
6354 mask
= INTEL_RC6_ENABLE
| INTEL_RC6p_ENABLE
|
6357 mask
= INTEL_RC6_ENABLE
;
6359 if ((enable_rc6
& mask
) != enable_rc6
)
6360 DRM_DEBUG_DRIVER("Adjusting RC6 mask to %d "
6361 "(requested %d, valid %d)\n",
6362 enable_rc6
& mask
, enable_rc6
, mask
);
6364 return enable_rc6
& mask
;
6367 if (IS_IVYBRIDGE(dev_priv
))
6368 return (INTEL_RC6_ENABLE
| INTEL_RC6p_ENABLE
);
6370 return INTEL_RC6_ENABLE
;
6373 static void gen6_init_rps_frequencies(struct drm_i915_private
*dev_priv
)
6375 /* All of these values are in units of 50MHz */
6377 /* static values from HW: RP0 > RP1 > RPn (min_freq) */
6378 if (IS_GEN9_LP(dev_priv
)) {
6379 u32 rp_state_cap
= I915_READ(BXT_RP_STATE_CAP
);
6380 dev_priv
->rps
.rp0_freq
= (rp_state_cap
>> 16) & 0xff;
6381 dev_priv
->rps
.rp1_freq
= (rp_state_cap
>> 8) & 0xff;
6382 dev_priv
->rps
.min_freq
= (rp_state_cap
>> 0) & 0xff;
6384 u32 rp_state_cap
= I915_READ(GEN6_RP_STATE_CAP
);
6385 dev_priv
->rps
.rp0_freq
= (rp_state_cap
>> 0) & 0xff;
6386 dev_priv
->rps
.rp1_freq
= (rp_state_cap
>> 8) & 0xff;
6387 dev_priv
->rps
.min_freq
= (rp_state_cap
>> 16) & 0xff;
6389 /* hw_max = RP0 until we check for overclocking */
6390 dev_priv
->rps
.max_freq
= dev_priv
->rps
.rp0_freq
;
6392 dev_priv
->rps
.efficient_freq
= dev_priv
->rps
.rp1_freq
;
6393 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
) ||
6394 IS_GEN9_BC(dev_priv
) || IS_CANNONLAKE(dev_priv
)) {
6395 u32 ddcc_status
= 0;
6397 if (sandybridge_pcode_read(dev_priv
,
6398 HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL
,
6400 dev_priv
->rps
.efficient_freq
=
6402 ((ddcc_status
>> 8) & 0xff),
6403 dev_priv
->rps
.min_freq
,
6404 dev_priv
->rps
.max_freq
);
6407 if (IS_GEN9_BC(dev_priv
) || IS_CANNONLAKE(dev_priv
)) {
6408 /* Store the frequency values in 16.66 MHZ units, which is
6409 * the natural hardware unit for SKL
6411 dev_priv
->rps
.rp0_freq
*= GEN9_FREQ_SCALER
;
6412 dev_priv
->rps
.rp1_freq
*= GEN9_FREQ_SCALER
;
6413 dev_priv
->rps
.min_freq
*= GEN9_FREQ_SCALER
;
6414 dev_priv
->rps
.max_freq
*= GEN9_FREQ_SCALER
;
6415 dev_priv
->rps
.efficient_freq
*= GEN9_FREQ_SCALER
;
6419 static void reset_rps(struct drm_i915_private
*dev_priv
,
6420 int (*set
)(struct drm_i915_private
*, u8
))
6422 u8 freq
= dev_priv
->rps
.cur_freq
;
6425 dev_priv
->rps
.power
= -1;
6426 dev_priv
->rps
.cur_freq
= -1;
6428 if (set(dev_priv
, freq
))
6429 DRM_ERROR("Failed to reset RPS to initial values\n");
6432 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
6433 static void gen9_enable_rps(struct drm_i915_private
*dev_priv
)
6435 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
6437 /* Program defaults and thresholds for RPS*/
6438 I915_WRITE(GEN6_RC_VIDEO_FREQ
,
6439 GEN9_FREQUENCY(dev_priv
->rps
.rp1_freq
));
6441 /* 1 second timeout*/
6442 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
,
6443 GT_INTERVAL_FROM_US(dev_priv
, 1000000));
6445 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 0xa);
6447 /* Leaning on the below call to gen6_set_rps to program/setup the
6448 * Up/Down EI & threshold registers, as well as the RP_CONTROL,
6449 * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
6450 reset_rps(dev_priv
, gen6_set_rps
);
6452 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
6455 static void gen9_enable_rc6(struct drm_i915_private
*dev_priv
)
6457 struct intel_engine_cs
*engine
;
6458 enum intel_engine_id id
;
6459 uint32_t rc6_mask
= 0;
6461 /* 1a: Software RC state - RC0 */
6462 I915_WRITE(GEN6_RC_STATE
, 0);
6464 /* 1b: Get forcewake during program sequence. Although the driver
6465 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
6466 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
6468 /* 2a: Disable RC states. */
6469 I915_WRITE(GEN6_RC_CONTROL
, 0);
6471 /* 2b: Program RC6 thresholds.*/
6473 /* WaRsDoubleRc6WrlWithCoarsePowerGating: Doubling WRL only when CPG is enabled */
6474 if (IS_SKYLAKE(dev_priv
))
6475 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 108 << 16);
6477 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 54 << 16);
6478 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
6479 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
6480 for_each_engine(engine
, dev_priv
, id
)
6481 I915_WRITE(RING_MAX_IDLE(engine
->mmio_base
), 10);
6483 if (HAS_GUC(dev_priv
))
6484 I915_WRITE(GUC_MAX_IDLE_COUNT
, 0xA);
6486 I915_WRITE(GEN6_RC_SLEEP
, 0);
6488 /* 2c: Program Coarse Power Gating Policies. */
6489 I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS
, 25);
6490 I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS
, 25);
6492 /* 3a: Enable RC6 */
6493 if (intel_enable_rc6() & INTEL_RC6_ENABLE
)
6494 rc6_mask
= GEN6_RC_CTL_RC6_ENABLE
;
6495 DRM_INFO("RC6 %s\n", onoff(rc6_mask
& GEN6_RC_CTL_RC6_ENABLE
));
6496 I915_WRITE(GEN6_RC6_THRESHOLD
, 37500); /* 37.5/125ms per EI */
6497 I915_WRITE(GEN6_RC_CONTROL
,
6498 GEN6_RC_CTL_HW_ENABLE
| GEN6_RC_CTL_EI_MODE(1) | rc6_mask
);
6501 * 3b: Enable Coarse Power Gating only when RC6 is enabled.
6502 * WaRsDisableCoarsePowerGating:skl,bxt - Render/Media PG need to be disabled with RC6.
6504 if (NEEDS_WaRsDisableCoarsePowerGating(dev_priv
))
6505 I915_WRITE(GEN9_PG_ENABLE
, 0);
6507 I915_WRITE(GEN9_PG_ENABLE
, (rc6_mask
& GEN6_RC_CTL_RC6_ENABLE
) ?
6508 (GEN9_RENDER_PG_ENABLE
| GEN9_MEDIA_PG_ENABLE
) : 0);
6510 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
6513 static void gen8_enable_rps(struct drm_i915_private
*dev_priv
)
6515 struct intel_engine_cs
*engine
;
6516 enum intel_engine_id id
;
6517 uint32_t rc6_mask
= 0;
6519 /* 1a: Software RC state - RC0 */
6520 I915_WRITE(GEN6_RC_STATE
, 0);
6522 /* 1c & 1d: Get forcewake during program sequence. Although the driver
6523 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
6524 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
6526 /* 2a: Disable RC states. */
6527 I915_WRITE(GEN6_RC_CONTROL
, 0);
6529 /* 2b: Program RC6 thresholds.*/
6530 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16);
6531 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
6532 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
6533 for_each_engine(engine
, dev_priv
, id
)
6534 I915_WRITE(RING_MAX_IDLE(engine
->mmio_base
), 10);
6535 I915_WRITE(GEN6_RC_SLEEP
, 0);
6536 if (IS_BROADWELL(dev_priv
))
6537 I915_WRITE(GEN6_RC6_THRESHOLD
, 625); /* 800us/1.28 for TO */
6539 I915_WRITE(GEN6_RC6_THRESHOLD
, 50000); /* 50/125ms per EI */
6542 if (intel_enable_rc6() & INTEL_RC6_ENABLE
)
6543 rc6_mask
= GEN6_RC_CTL_RC6_ENABLE
;
6544 intel_print_rc6_info(dev_priv
, rc6_mask
);
6545 if (IS_BROADWELL(dev_priv
))
6546 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
6547 GEN7_RC_CTL_TO_MODE
|
6550 I915_WRITE(GEN6_RC_CONTROL
, GEN6_RC_CTL_HW_ENABLE
|
6551 GEN6_RC_CTL_EI_MODE(1) |
6554 /* 4 Program defaults and thresholds for RPS*/
6555 I915_WRITE(GEN6_RPNSWREQ
,
6556 HSW_FREQUENCY(dev_priv
->rps
.rp1_freq
));
6557 I915_WRITE(GEN6_RC_VIDEO_FREQ
,
6558 HSW_FREQUENCY(dev_priv
->rps
.rp1_freq
));
6559 /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
6560 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 100000000 / 128); /* 1 second timeout */
6562 /* Docs recommend 900MHz, and 300 MHz respectively */
6563 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS
,
6564 dev_priv
->rps
.max_freq_softlimit
<< 24 |
6565 dev_priv
->rps
.min_freq_softlimit
<< 16);
6567 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 7600000 / 128); /* 76ms busyness per EI, 90% */
6568 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 31300000 / 128); /* 313ms busyness per EI, 70%*/
6569 I915_WRITE(GEN6_RP_UP_EI
, 66000); /* 84.48ms, XXX: random? */
6570 I915_WRITE(GEN6_RP_DOWN_EI
, 350000); /* 448ms, XXX: random? */
6572 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
6575 I915_WRITE(GEN6_RP_CONTROL
,
6576 GEN6_RP_MEDIA_TURBO
|
6577 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
6578 GEN6_RP_MEDIA_IS_GFX
|
6580 GEN6_RP_UP_BUSY_AVG
|
6581 GEN6_RP_DOWN_IDLE_AVG
);
6583 /* 6: Ring frequency + overclocking (our driver does this later */
6585 reset_rps(dev_priv
, gen6_set_rps
);
6587 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
6590 static void gen6_enable_rps(struct drm_i915_private
*dev_priv
)
6592 struct intel_engine_cs
*engine
;
6593 enum intel_engine_id id
;
6594 u32 rc6vids
, rc6_mask
= 0;
6599 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
6601 /* Here begins a magic sequence of register writes to enable
6602 * auto-downclocking.
6604 * Perhaps there might be some value in exposing these to
6607 I915_WRITE(GEN6_RC_STATE
, 0);
6609 /* Clear the DBG now so we don't confuse earlier errors */
6610 gtfifodbg
= I915_READ(GTFIFODBG
);
6612 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg
);
6613 I915_WRITE(GTFIFODBG
, gtfifodbg
);
6616 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
6618 /* disable the counters and set deterministic thresholds */
6619 I915_WRITE(GEN6_RC_CONTROL
, 0);
6621 I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT
, 1000 << 16);
6622 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16 | 30);
6623 I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT
, 30);
6624 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000);
6625 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25);
6627 for_each_engine(engine
, dev_priv
, id
)
6628 I915_WRITE(RING_MAX_IDLE(engine
->mmio_base
), 10);
6630 I915_WRITE(GEN6_RC_SLEEP
, 0);
6631 I915_WRITE(GEN6_RC1e_THRESHOLD
, 1000);
6632 if (IS_IVYBRIDGE(dev_priv
))
6633 I915_WRITE(GEN6_RC6_THRESHOLD
, 125000);
6635 I915_WRITE(GEN6_RC6_THRESHOLD
, 50000);
6636 I915_WRITE(GEN6_RC6p_THRESHOLD
, 150000);
6637 I915_WRITE(GEN6_RC6pp_THRESHOLD
, 64000); /* unused */
6639 /* Check if we are enabling RC6 */
6640 rc6_mode
= intel_enable_rc6();
6641 if (rc6_mode
& INTEL_RC6_ENABLE
)
6642 rc6_mask
|= GEN6_RC_CTL_RC6_ENABLE
;
6644 /* We don't use those on Haswell */
6645 if (!IS_HASWELL(dev_priv
)) {
6646 if (rc6_mode
& INTEL_RC6p_ENABLE
)
6647 rc6_mask
|= GEN6_RC_CTL_RC6p_ENABLE
;
6649 if (rc6_mode
& INTEL_RC6pp_ENABLE
)
6650 rc6_mask
|= GEN6_RC_CTL_RC6pp_ENABLE
;
6653 intel_print_rc6_info(dev_priv
, rc6_mask
);
6655 I915_WRITE(GEN6_RC_CONTROL
,
6657 GEN6_RC_CTL_EI_MODE(1) |
6658 GEN6_RC_CTL_HW_ENABLE
);
6660 /* Power down if completely idle for over 50ms */
6661 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 50000);
6662 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
6664 reset_rps(dev_priv
, gen6_set_rps
);
6667 ret
= sandybridge_pcode_read(dev_priv
, GEN6_PCODE_READ_RC6VIDS
, &rc6vids
);
6668 if (IS_GEN6(dev_priv
) && ret
) {
6669 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
6670 } else if (IS_GEN6(dev_priv
) && (GEN6_DECODE_RC6_VID(rc6vids
& 0xff) < 450)) {
6671 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
6672 GEN6_DECODE_RC6_VID(rc6vids
& 0xff), 450);
6673 rc6vids
&= 0xffff00;
6674 rc6vids
|= GEN6_ENCODE_RC6_VID(450);
6675 ret
= sandybridge_pcode_write(dev_priv
, GEN6_PCODE_WRITE_RC6VIDS
, rc6vids
);
6677 DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
6680 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
6683 static void gen6_update_ring_freq(struct drm_i915_private
*dev_priv
)
6686 unsigned int gpu_freq
;
6687 unsigned int max_ia_freq
, min_ring_freq
;
6688 unsigned int max_gpu_freq
, min_gpu_freq
;
6689 int scaling_factor
= 180;
6690 struct cpufreq_policy
*policy
;
6692 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
6694 policy
= cpufreq_cpu_get(0);
6696 max_ia_freq
= policy
->cpuinfo
.max_freq
;
6697 cpufreq_cpu_put(policy
);
6700 * Default to measured freq if none found, PCU will ensure we
6703 max_ia_freq
= tsc_khz
;
6706 /* Convert from kHz to MHz */
6707 max_ia_freq
/= 1000;
6709 min_ring_freq
= I915_READ(DCLK
) & 0xf;
6710 /* convert DDR frequency from units of 266.6MHz to bandwidth */
6711 min_ring_freq
= mult_frac(min_ring_freq
, 8, 3);
6713 if (IS_GEN9_BC(dev_priv
) || IS_CANNONLAKE(dev_priv
)) {
6714 /* Convert GT frequency to 50 HZ units */
6715 min_gpu_freq
= dev_priv
->rps
.min_freq
/ GEN9_FREQ_SCALER
;
6716 max_gpu_freq
= dev_priv
->rps
.max_freq
/ GEN9_FREQ_SCALER
;
6718 min_gpu_freq
= dev_priv
->rps
.min_freq
;
6719 max_gpu_freq
= dev_priv
->rps
.max_freq
;
6723 * For each potential GPU frequency, load a ring frequency we'd like
6724 * to use for memory access. We do this by specifying the IA frequency
6725 * the PCU should use as a reference to determine the ring frequency.
6727 for (gpu_freq
= max_gpu_freq
; gpu_freq
>= min_gpu_freq
; gpu_freq
--) {
6728 int diff
= max_gpu_freq
- gpu_freq
;
6729 unsigned int ia_freq
= 0, ring_freq
= 0;
6731 if (IS_GEN9_BC(dev_priv
) || IS_CANNONLAKE(dev_priv
)) {
6733 * ring_freq = 2 * GT. ring_freq is in 100MHz units
6734 * No floor required for ring frequency on SKL.
6736 ring_freq
= gpu_freq
;
6737 } else if (INTEL_INFO(dev_priv
)->gen
>= 8) {
6738 /* max(2 * GT, DDR). NB: GT is 50MHz units */
6739 ring_freq
= max(min_ring_freq
, gpu_freq
);
6740 } else if (IS_HASWELL(dev_priv
)) {
6741 ring_freq
= mult_frac(gpu_freq
, 5, 4);
6742 ring_freq
= max(min_ring_freq
, ring_freq
);
6743 /* leave ia_freq as the default, chosen by cpufreq */
6745 /* On older processors, there is no separate ring
6746 * clock domain, so in order to boost the bandwidth
6747 * of the ring, we need to upclock the CPU (ia_freq).
6749 * For GPU frequencies less than 750MHz,
6750 * just use the lowest ring freq.
6752 if (gpu_freq
< min_freq
)
6755 ia_freq
= max_ia_freq
- ((diff
* scaling_factor
) / 2);
6756 ia_freq
= DIV_ROUND_CLOSEST(ia_freq
, 100);
6759 sandybridge_pcode_write(dev_priv
,
6760 GEN6_PCODE_WRITE_MIN_FREQ_TABLE
,
6761 ia_freq
<< GEN6_PCODE_FREQ_IA_RATIO_SHIFT
|
6762 ring_freq
<< GEN6_PCODE_FREQ_RING_RATIO_SHIFT
|
6767 static int cherryview_rps_max_freq(struct drm_i915_private
*dev_priv
)
6771 val
= vlv_punit_read(dev_priv
, FB_GFX_FMAX_AT_VMAX_FUSE
);
6773 switch (INTEL_INFO(dev_priv
)->sseu
.eu_total
) {
6775 /* (2 * 4) config */
6776 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT
);
6779 /* (2 * 6) config */
6780 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT
);
6783 /* (2 * 8) config */
6785 /* Setting (2 * 8) Min RP0 for any other combination */
6786 rp0
= (val
>> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT
);
6790 rp0
= (rp0
& FB_GFX_FREQ_FUSE_MASK
);
6795 static int cherryview_rps_rpe_freq(struct drm_i915_private
*dev_priv
)
6799 val
= vlv_punit_read(dev_priv
, PUNIT_GPU_DUTYCYCLE_REG
);
6800 rpe
= (val
>> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT
) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK
;
6805 static int cherryview_rps_guar_freq(struct drm_i915_private
*dev_priv
)
6809 val
= vlv_punit_read(dev_priv
, FB_GFX_FMAX_AT_VMAX_FUSE
);
6810 rp1
= (val
& FB_GFX_FREQ_FUSE_MASK
);
6815 static u32
cherryview_rps_min_freq(struct drm_i915_private
*dev_priv
)
6819 val
= vlv_punit_read(dev_priv
, FB_GFX_FMIN_AT_VMIN_FUSE
);
6820 rpn
= ((val
>> FB_GFX_FMIN_AT_VMIN_FUSE_SHIFT
) &
6821 FB_GFX_FREQ_FUSE_MASK
);
6826 static int valleyview_rps_guar_freq(struct drm_i915_private
*dev_priv
)
6830 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FREQ_FUSE
);
6832 rp1
= (val
& FB_GFX_FGUARANTEED_FREQ_FUSE_MASK
) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT
;
6837 static int valleyview_rps_max_freq(struct drm_i915_private
*dev_priv
)
6841 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FREQ_FUSE
);
6843 rp0
= (val
& FB_GFX_MAX_FREQ_FUSE_MASK
) >> FB_GFX_MAX_FREQ_FUSE_SHIFT
;
6845 rp0
= min_t(u32
, rp0
, 0xea);
6850 static int valleyview_rps_rpe_freq(struct drm_i915_private
*dev_priv
)
6854 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FMAX_FUSE_LO
);
6855 rpe
= (val
& FB_FMAX_VMIN_FREQ_LO_MASK
) >> FB_FMAX_VMIN_FREQ_LO_SHIFT
;
6856 val
= vlv_nc_read(dev_priv
, IOSF_NC_FB_GFX_FMAX_FUSE_HI
);
6857 rpe
|= (val
& FB_FMAX_VMIN_FREQ_HI_MASK
) << 5;
6862 static int valleyview_rps_min_freq(struct drm_i915_private
*dev_priv
)
6866 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_LFM
) & 0xff;
6868 * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
6869 * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
6870 * a BYT-M B0 the above register contains 0xbf. Moreover when setting
6871 * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
6872 * to make sure it matches what Punit accepts.
6874 return max_t(u32
, val
, 0xc0);
6877 /* Check that the pctx buffer wasn't move under us. */
6878 static void valleyview_check_pctx(struct drm_i915_private
*dev_priv
)
6880 unsigned long pctx_addr
= I915_READ(VLV_PCBR
) & ~4095;
6882 WARN_ON(pctx_addr
!= dev_priv
->mm
.stolen_base
+
6883 dev_priv
->vlv_pctx
->stolen
->start
);
6887 /* Check that the pcbr address is not empty. */
6888 static void cherryview_check_pctx(struct drm_i915_private
*dev_priv
)
6890 unsigned long pctx_addr
= I915_READ(VLV_PCBR
) & ~4095;
6892 WARN_ON((pctx_addr
>> VLV_PCBR_ADDR_SHIFT
) == 0);
6895 static void cherryview_setup_pctx(struct drm_i915_private
*dev_priv
)
6897 struct i915_ggtt
*ggtt
= &dev_priv
->ggtt
;
6898 unsigned long pctx_paddr
, paddr
;
6900 int pctx_size
= 32*1024;
6902 pcbr
= I915_READ(VLV_PCBR
);
6903 if ((pcbr
>> VLV_PCBR_ADDR_SHIFT
) == 0) {
6904 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
6905 paddr
= (dev_priv
->mm
.stolen_base
+
6906 (ggtt
->stolen_size
- pctx_size
));
6908 pctx_paddr
= (paddr
& (~4095));
6909 I915_WRITE(VLV_PCBR
, pctx_paddr
);
6912 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR
));
6915 static void valleyview_setup_pctx(struct drm_i915_private
*dev_priv
)
6917 struct drm_i915_gem_object
*pctx
;
6918 unsigned long pctx_paddr
;
6920 int pctx_size
= 24*1024;
6922 pcbr
= I915_READ(VLV_PCBR
);
6924 /* BIOS set it up already, grab the pre-alloc'd space */
6927 pcbr_offset
= (pcbr
& (~4095)) - dev_priv
->mm
.stolen_base
;
6928 pctx
= i915_gem_object_create_stolen_for_preallocated(dev_priv
,
6930 I915_GTT_OFFSET_NONE
,
6935 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
6938 * From the Gunit register HAS:
6939 * The Gfx driver is expected to program this register and ensure
6940 * proper allocation within Gfx stolen memory. For example, this
6941 * register should be programmed such than the PCBR range does not
6942 * overlap with other ranges, such as the frame buffer, protected
6943 * memory, or any other relevant ranges.
6945 pctx
= i915_gem_object_create_stolen(dev_priv
, pctx_size
);
6947 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
6951 pctx_paddr
= dev_priv
->mm
.stolen_base
+ pctx
->stolen
->start
;
6952 I915_WRITE(VLV_PCBR
, pctx_paddr
);
6955 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR
));
6956 dev_priv
->vlv_pctx
= pctx
;
6959 static void valleyview_cleanup_pctx(struct drm_i915_private
*dev_priv
)
6961 if (WARN_ON(!dev_priv
->vlv_pctx
))
6964 i915_gem_object_put(dev_priv
->vlv_pctx
);
6965 dev_priv
->vlv_pctx
= NULL
;
6968 static void vlv_init_gpll_ref_freq(struct drm_i915_private
*dev_priv
)
6970 dev_priv
->rps
.gpll_ref_freq
=
6971 vlv_get_cck_clock(dev_priv
, "GPLL ref",
6972 CCK_GPLL_CLOCK_CONTROL
,
6973 dev_priv
->czclk_freq
);
6975 DRM_DEBUG_DRIVER("GPLL reference freq: %d kHz\n",
6976 dev_priv
->rps
.gpll_ref_freq
);
6979 static void valleyview_init_gt_powersave(struct drm_i915_private
*dev_priv
)
6983 valleyview_setup_pctx(dev_priv
);
6985 vlv_init_gpll_ref_freq(dev_priv
);
6987 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
6988 switch ((val
>> 6) & 3) {
6991 dev_priv
->mem_freq
= 800;
6994 dev_priv
->mem_freq
= 1066;
6997 dev_priv
->mem_freq
= 1333;
7000 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv
->mem_freq
);
7002 dev_priv
->rps
.max_freq
= valleyview_rps_max_freq(dev_priv
);
7003 dev_priv
->rps
.rp0_freq
= dev_priv
->rps
.max_freq
;
7004 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
7005 intel_gpu_freq(dev_priv
, dev_priv
->rps
.max_freq
),
7006 dev_priv
->rps
.max_freq
);
7008 dev_priv
->rps
.efficient_freq
= valleyview_rps_rpe_freq(dev_priv
);
7009 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
7010 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
7011 dev_priv
->rps
.efficient_freq
);
7013 dev_priv
->rps
.rp1_freq
= valleyview_rps_guar_freq(dev_priv
);
7014 DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
7015 intel_gpu_freq(dev_priv
, dev_priv
->rps
.rp1_freq
),
7016 dev_priv
->rps
.rp1_freq
);
7018 dev_priv
->rps
.min_freq
= valleyview_rps_min_freq(dev_priv
);
7019 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
7020 intel_gpu_freq(dev_priv
, dev_priv
->rps
.min_freq
),
7021 dev_priv
->rps
.min_freq
);
7024 static void cherryview_init_gt_powersave(struct drm_i915_private
*dev_priv
)
7028 cherryview_setup_pctx(dev_priv
);
7030 vlv_init_gpll_ref_freq(dev_priv
);
7032 mutex_lock(&dev_priv
->sb_lock
);
7033 val
= vlv_cck_read(dev_priv
, CCK_FUSE_REG
);
7034 mutex_unlock(&dev_priv
->sb_lock
);
7036 switch ((val
>> 2) & 0x7) {
7038 dev_priv
->mem_freq
= 2000;
7041 dev_priv
->mem_freq
= 1600;
7044 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv
->mem_freq
);
7046 dev_priv
->rps
.max_freq
= cherryview_rps_max_freq(dev_priv
);
7047 dev_priv
->rps
.rp0_freq
= dev_priv
->rps
.max_freq
;
7048 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
7049 intel_gpu_freq(dev_priv
, dev_priv
->rps
.max_freq
),
7050 dev_priv
->rps
.max_freq
);
7052 dev_priv
->rps
.efficient_freq
= cherryview_rps_rpe_freq(dev_priv
);
7053 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
7054 intel_gpu_freq(dev_priv
, dev_priv
->rps
.efficient_freq
),
7055 dev_priv
->rps
.efficient_freq
);
7057 dev_priv
->rps
.rp1_freq
= cherryview_rps_guar_freq(dev_priv
);
7058 DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
7059 intel_gpu_freq(dev_priv
, dev_priv
->rps
.rp1_freq
),
7060 dev_priv
->rps
.rp1_freq
);
7062 dev_priv
->rps
.min_freq
= cherryview_rps_min_freq(dev_priv
);
7063 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
7064 intel_gpu_freq(dev_priv
, dev_priv
->rps
.min_freq
),
7065 dev_priv
->rps
.min_freq
);
7067 WARN_ONCE((dev_priv
->rps
.max_freq
|
7068 dev_priv
->rps
.efficient_freq
|
7069 dev_priv
->rps
.rp1_freq
|
7070 dev_priv
->rps
.min_freq
) & 1,
7071 "Odd GPU freq values\n");
7074 static void valleyview_cleanup_gt_powersave(struct drm_i915_private
*dev_priv
)
7076 valleyview_cleanup_pctx(dev_priv
);
7079 static void cherryview_enable_rps(struct drm_i915_private
*dev_priv
)
7081 struct intel_engine_cs
*engine
;
7082 enum intel_engine_id id
;
7083 u32 gtfifodbg
, val
, rc6_mode
= 0, pcbr
;
7085 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
7087 gtfifodbg
= I915_READ(GTFIFODBG
) & ~(GT_FIFO_SBDEDICATE_FREE_ENTRY_CHV
|
7088 GT_FIFO_FREE_ENTRIES_CHV
);
7090 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
7092 I915_WRITE(GTFIFODBG
, gtfifodbg
);
7095 cherryview_check_pctx(dev_priv
);
7097 /* 1a & 1b: Get forcewake during program sequence. Although the driver
7098 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
7099 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
7101 /* Disable RC states. */
7102 I915_WRITE(GEN6_RC_CONTROL
, 0);
7104 /* 2a: Program RC6 thresholds.*/
7105 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 40 << 16);
7106 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000); /* 12500 * 1280ns */
7107 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25); /* 25 * 1280ns */
7109 for_each_engine(engine
, dev_priv
, id
)
7110 I915_WRITE(RING_MAX_IDLE(engine
->mmio_base
), 10);
7111 I915_WRITE(GEN6_RC_SLEEP
, 0);
7113 /* TO threshold set to 500 us ( 0x186 * 1.28 us) */
7114 I915_WRITE(GEN6_RC6_THRESHOLD
, 0x186);
7116 /* allows RC6 residency counter to work */
7117 I915_WRITE(VLV_COUNTER_CONTROL
,
7118 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH
|
7119 VLV_MEDIA_RC6_COUNT_EN
|
7120 VLV_RENDER_RC6_COUNT_EN
));
7122 /* For now we assume BIOS is allocating and populating the PCBR */
7123 pcbr
= I915_READ(VLV_PCBR
);
7126 if ((intel_enable_rc6() & INTEL_RC6_ENABLE
) &&
7127 (pcbr
>> VLV_PCBR_ADDR_SHIFT
))
7128 rc6_mode
= GEN7_RC_CTL_TO_MODE
;
7130 I915_WRITE(GEN6_RC_CONTROL
, rc6_mode
);
7132 /* 4 Program defaults and thresholds for RPS*/
7133 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 1000000);
7134 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 59400);
7135 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 245000);
7136 I915_WRITE(GEN6_RP_UP_EI
, 66000);
7137 I915_WRITE(GEN6_RP_DOWN_EI
, 350000);
7139 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
7142 I915_WRITE(GEN6_RP_CONTROL
,
7143 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
7144 GEN6_RP_MEDIA_IS_GFX
|
7146 GEN6_RP_UP_BUSY_AVG
|
7147 GEN6_RP_DOWN_IDLE_AVG
);
7149 /* Setting Fixed Bias */
7150 val
= VLV_OVERRIDE_EN
|
7152 CHV_BIAS_CPU_50_SOC_50
;
7153 vlv_punit_write(dev_priv
, VLV_TURBO_SOC_OVERRIDE
, val
);
7155 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
7157 /* RPS code assumes GPLL is used */
7158 WARN_ONCE((val
& GPLLENABLE
) == 0, "GPLL not enabled\n");
7160 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val
& GPLLENABLE
));
7161 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val
);
7163 reset_rps(dev_priv
, valleyview_set_rps
);
7165 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
7168 static void valleyview_enable_rps(struct drm_i915_private
*dev_priv
)
7170 struct intel_engine_cs
*engine
;
7171 enum intel_engine_id id
;
7172 u32 gtfifodbg
, val
, rc6_mode
= 0;
7174 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
7176 valleyview_check_pctx(dev_priv
);
7178 gtfifodbg
= I915_READ(GTFIFODBG
);
7180 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
7182 I915_WRITE(GTFIFODBG
, gtfifodbg
);
7185 /* If VLV, Forcewake all wells, else re-direct to regular path */
7186 intel_uncore_forcewake_get(dev_priv
, FORCEWAKE_ALL
);
7188 /* Disable RC states. */
7189 I915_WRITE(GEN6_RC_CONTROL
, 0);
7191 I915_WRITE(GEN6_RP_DOWN_TIMEOUT
, 1000000);
7192 I915_WRITE(GEN6_RP_UP_THRESHOLD
, 59400);
7193 I915_WRITE(GEN6_RP_DOWN_THRESHOLD
, 245000);
7194 I915_WRITE(GEN6_RP_UP_EI
, 66000);
7195 I915_WRITE(GEN6_RP_DOWN_EI
, 350000);
7197 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS
, 10);
7199 I915_WRITE(GEN6_RP_CONTROL
,
7200 GEN6_RP_MEDIA_TURBO
|
7201 GEN6_RP_MEDIA_HW_NORMAL_MODE
|
7202 GEN6_RP_MEDIA_IS_GFX
|
7204 GEN6_RP_UP_BUSY_AVG
|
7205 GEN6_RP_DOWN_IDLE_CONT
);
7207 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT
, 0x00280000);
7208 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL
, 125000);
7209 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS
, 25);
7211 for_each_engine(engine
, dev_priv
, id
)
7212 I915_WRITE(RING_MAX_IDLE(engine
->mmio_base
), 10);
7214 I915_WRITE(GEN6_RC6_THRESHOLD
, 0x557);
7216 /* allows RC6 residency counter to work */
7217 I915_WRITE(VLV_COUNTER_CONTROL
,
7218 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH
|
7219 VLV_MEDIA_RC0_COUNT_EN
|
7220 VLV_RENDER_RC0_COUNT_EN
|
7221 VLV_MEDIA_RC6_COUNT_EN
|
7222 VLV_RENDER_RC6_COUNT_EN
));
7224 if (intel_enable_rc6() & INTEL_RC6_ENABLE
)
7225 rc6_mode
= GEN7_RC_CTL_TO_MODE
| VLV_RC_CTL_CTX_RST_PARALLEL
;
7227 intel_print_rc6_info(dev_priv
, rc6_mode
);
7229 I915_WRITE(GEN6_RC_CONTROL
, rc6_mode
);
7231 /* Setting Fixed Bias */
7232 val
= VLV_OVERRIDE_EN
|
7234 VLV_BIAS_CPU_125_SOC_875
;
7235 vlv_punit_write(dev_priv
, VLV_TURBO_SOC_OVERRIDE
, val
);
7237 val
= vlv_punit_read(dev_priv
, PUNIT_REG_GPU_FREQ_STS
);
7239 /* RPS code assumes GPLL is used */
7240 WARN_ONCE((val
& GPLLENABLE
) == 0, "GPLL not enabled\n");
7242 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val
& GPLLENABLE
));
7243 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val
);
7245 reset_rps(dev_priv
, valleyview_set_rps
);
7247 intel_uncore_forcewake_put(dev_priv
, FORCEWAKE_ALL
);
7250 static unsigned long intel_pxfreq(u32 vidfreq
)
7253 int div
= (vidfreq
& 0x3f0000) >> 16;
7254 int post
= (vidfreq
& 0x3000) >> 12;
7255 int pre
= (vidfreq
& 0x7);
7260 freq
= ((div
* 133333) / ((1<<post
) * pre
));
7265 static const struct cparams
{
7271 { 1, 1333, 301, 28664 },
7272 { 1, 1066, 294, 24460 },
7273 { 1, 800, 294, 25192 },
7274 { 0, 1333, 276, 27605 },
7275 { 0, 1066, 276, 27605 },
7276 { 0, 800, 231, 23784 },
7279 static unsigned long __i915_chipset_val(struct drm_i915_private
*dev_priv
)
7281 u64 total_count
, diff
, ret
;
7282 u32 count1
, count2
, count3
, m
= 0, c
= 0;
7283 unsigned long now
= jiffies_to_msecs(jiffies
), diff1
;
7286 lockdep_assert_held(&mchdev_lock
);
7288 diff1
= now
- dev_priv
->ips
.last_time1
;
7290 /* Prevent division-by-zero if we are asking too fast.
7291 * Also, we don't get interesting results if we are polling
7292 * faster than once in 10ms, so just return the saved value
7296 return dev_priv
->ips
.chipset_power
;
7298 count1
= I915_READ(DMIEC
);
7299 count2
= I915_READ(DDREC
);
7300 count3
= I915_READ(CSIEC
);
7302 total_count
= count1
+ count2
+ count3
;
7304 /* FIXME: handle per-counter overflow */
7305 if (total_count
< dev_priv
->ips
.last_count1
) {
7306 diff
= ~0UL - dev_priv
->ips
.last_count1
;
7307 diff
+= total_count
;
7309 diff
= total_count
- dev_priv
->ips
.last_count1
;
7312 for (i
= 0; i
< ARRAY_SIZE(cparams
); i
++) {
7313 if (cparams
[i
].i
== dev_priv
->ips
.c_m
&&
7314 cparams
[i
].t
== dev_priv
->ips
.r_t
) {
7321 diff
= div_u64(diff
, diff1
);
7322 ret
= ((m
* diff
) + c
);
7323 ret
= div_u64(ret
, 10);
7325 dev_priv
->ips
.last_count1
= total_count
;
7326 dev_priv
->ips
.last_time1
= now
;
7328 dev_priv
->ips
.chipset_power
= ret
;
7333 unsigned long i915_chipset_val(struct drm_i915_private
*dev_priv
)
7337 if (INTEL_INFO(dev_priv
)->gen
!= 5)
7340 spin_lock_irq(&mchdev_lock
);
7342 val
= __i915_chipset_val(dev_priv
);
7344 spin_unlock_irq(&mchdev_lock
);
7349 unsigned long i915_mch_val(struct drm_i915_private
*dev_priv
)
7351 unsigned long m
, x
, b
;
7354 tsfs
= I915_READ(TSFS
);
7356 m
= ((tsfs
& TSFS_SLOPE_MASK
) >> TSFS_SLOPE_SHIFT
);
7357 x
= I915_READ8(TR1
);
7359 b
= tsfs
& TSFS_INTR_MASK
;
7361 return ((m
* x
) / 127) - b
;
7364 static int _pxvid_to_vd(u8 pxvid
)
7369 if (pxvid
>= 8 && pxvid
< 31)
7372 return (pxvid
+ 2) * 125;
7375 static u32
pvid_to_extvid(struct drm_i915_private
*dev_priv
, u8 pxvid
)
7377 const int vd
= _pxvid_to_vd(pxvid
);
7378 const int vm
= vd
- 1125;
7380 if (INTEL_INFO(dev_priv
)->is_mobile
)
7381 return vm
> 0 ? vm
: 0;
7386 static void __i915_update_gfx_val(struct drm_i915_private
*dev_priv
)
7388 u64 now
, diff
, diffms
;
7391 lockdep_assert_held(&mchdev_lock
);
7393 now
= ktime_get_raw_ns();
7394 diffms
= now
- dev_priv
->ips
.last_time2
;
7395 do_div(diffms
, NSEC_PER_MSEC
);
7397 /* Don't divide by 0 */
7401 count
= I915_READ(GFXEC
);
7403 if (count
< dev_priv
->ips
.last_count2
) {
7404 diff
= ~0UL - dev_priv
->ips
.last_count2
;
7407 diff
= count
- dev_priv
->ips
.last_count2
;
7410 dev_priv
->ips
.last_count2
= count
;
7411 dev_priv
->ips
.last_time2
= now
;
7413 /* More magic constants... */
7415 diff
= div_u64(diff
, diffms
* 10);
7416 dev_priv
->ips
.gfx_power
= diff
;
7419 void i915_update_gfx_val(struct drm_i915_private
*dev_priv
)
7421 if (INTEL_INFO(dev_priv
)->gen
!= 5)
7424 spin_lock_irq(&mchdev_lock
);
7426 __i915_update_gfx_val(dev_priv
);
7428 spin_unlock_irq(&mchdev_lock
);
7431 static unsigned long __i915_gfx_val(struct drm_i915_private
*dev_priv
)
7433 unsigned long t
, corr
, state1
, corr2
, state2
;
7436 lockdep_assert_held(&mchdev_lock
);
7438 pxvid
= I915_READ(PXVFREQ(dev_priv
->rps
.cur_freq
));
7439 pxvid
= (pxvid
>> 24) & 0x7f;
7440 ext_v
= pvid_to_extvid(dev_priv
, pxvid
);
7444 t
= i915_mch_val(dev_priv
);
7446 /* Revel in the empirically derived constants */
7448 /* Correction factor in 1/100000 units */
7450 corr
= ((t
* 2349) + 135940);
7452 corr
= ((t
* 964) + 29317);
7454 corr
= ((t
* 301) + 1004);
7456 corr
= corr
* ((150142 * state1
) / 10000 - 78642);
7458 corr2
= (corr
* dev_priv
->ips
.corr
);
7460 state2
= (corr2
* state1
) / 10000;
7461 state2
/= 100; /* convert to mW */
7463 __i915_update_gfx_val(dev_priv
);
7465 return dev_priv
->ips
.gfx_power
+ state2
;
7468 unsigned long i915_gfx_val(struct drm_i915_private
*dev_priv
)
7472 if (INTEL_INFO(dev_priv
)->gen
!= 5)
7475 spin_lock_irq(&mchdev_lock
);
7477 val
= __i915_gfx_val(dev_priv
);
7479 spin_unlock_irq(&mchdev_lock
);
7485 * i915_read_mch_val - return value for IPS use
7487 * Calculate and return a value for the IPS driver to use when deciding whether
7488 * we have thermal and power headroom to increase CPU or GPU power budget.
7490 unsigned long i915_read_mch_val(void)
7492 struct drm_i915_private
*dev_priv
;
7493 unsigned long chipset_val
, graphics_val
, ret
= 0;
7495 spin_lock_irq(&mchdev_lock
);
7498 dev_priv
= i915_mch_dev
;
7500 chipset_val
= __i915_chipset_val(dev_priv
);
7501 graphics_val
= __i915_gfx_val(dev_priv
);
7503 ret
= chipset_val
+ graphics_val
;
7506 spin_unlock_irq(&mchdev_lock
);
7510 EXPORT_SYMBOL_GPL(i915_read_mch_val
);
7513 * i915_gpu_raise - raise GPU frequency limit
7515 * Raise the limit; IPS indicates we have thermal headroom.
7517 bool i915_gpu_raise(void)
7519 struct drm_i915_private
*dev_priv
;
7522 spin_lock_irq(&mchdev_lock
);
7523 if (!i915_mch_dev
) {
7527 dev_priv
= i915_mch_dev
;
7529 if (dev_priv
->ips
.max_delay
> dev_priv
->ips
.fmax
)
7530 dev_priv
->ips
.max_delay
--;
7533 spin_unlock_irq(&mchdev_lock
);
7537 EXPORT_SYMBOL_GPL(i915_gpu_raise
);
7540 * i915_gpu_lower - lower GPU frequency limit
7542 * IPS indicates we're close to a thermal limit, so throttle back the GPU
7543 * frequency maximum.
7545 bool i915_gpu_lower(void)
7547 struct drm_i915_private
*dev_priv
;
7550 spin_lock_irq(&mchdev_lock
);
7551 if (!i915_mch_dev
) {
7555 dev_priv
= i915_mch_dev
;
7557 if (dev_priv
->ips
.max_delay
< dev_priv
->ips
.min_delay
)
7558 dev_priv
->ips
.max_delay
++;
7561 spin_unlock_irq(&mchdev_lock
);
7565 EXPORT_SYMBOL_GPL(i915_gpu_lower
);
7568 * i915_gpu_busy - indicate GPU business to IPS
7570 * Tell the IPS driver whether or not the GPU is busy.
7572 bool i915_gpu_busy(void)
7576 spin_lock_irq(&mchdev_lock
);
7578 ret
= i915_mch_dev
->gt
.awake
;
7579 spin_unlock_irq(&mchdev_lock
);
7583 EXPORT_SYMBOL_GPL(i915_gpu_busy
);
7586 * i915_gpu_turbo_disable - disable graphics turbo
7588 * Disable graphics turbo by resetting the max frequency and setting the
7589 * current frequency to the default.
7591 bool i915_gpu_turbo_disable(void)
7593 struct drm_i915_private
*dev_priv
;
7596 spin_lock_irq(&mchdev_lock
);
7597 if (!i915_mch_dev
) {
7601 dev_priv
= i915_mch_dev
;
7603 dev_priv
->ips
.max_delay
= dev_priv
->ips
.fstart
;
7605 if (!ironlake_set_drps(dev_priv
, dev_priv
->ips
.fstart
))
7609 spin_unlock_irq(&mchdev_lock
);
7613 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable
);
7616 * Tells the intel_ips driver that the i915 driver is now loaded, if
7617 * IPS got loaded first.
7619 * This awkward dance is so that neither module has to depend on the
7620 * other in order for IPS to do the appropriate communication of
7621 * GPU turbo limits to i915.
7624 ips_ping_for_i915_load(void)
7628 link
= symbol_get(ips_link_to_i915_driver
);
7631 symbol_put(ips_link_to_i915_driver
);
7635 void intel_gpu_ips_init(struct drm_i915_private
*dev_priv
)
7637 /* We only register the i915 ips part with intel-ips once everything is
7638 * set up, to avoid intel-ips sneaking in and reading bogus values. */
7639 spin_lock_irq(&mchdev_lock
);
7640 i915_mch_dev
= dev_priv
;
7641 spin_unlock_irq(&mchdev_lock
);
7643 ips_ping_for_i915_load();
7646 void intel_gpu_ips_teardown(void)
7648 spin_lock_irq(&mchdev_lock
);
7649 i915_mch_dev
= NULL
;
7650 spin_unlock_irq(&mchdev_lock
);
7653 static void intel_init_emon(struct drm_i915_private
*dev_priv
)
7659 /* Disable to program */
7663 /* Program energy weights for various events */
7664 I915_WRITE(SDEW
, 0x15040d00);
7665 I915_WRITE(CSIEW0
, 0x007f0000);
7666 I915_WRITE(CSIEW1
, 0x1e220004);
7667 I915_WRITE(CSIEW2
, 0x04000004);
7669 for (i
= 0; i
< 5; i
++)
7670 I915_WRITE(PEW(i
), 0);
7671 for (i
= 0; i
< 3; i
++)
7672 I915_WRITE(DEW(i
), 0);
7674 /* Program P-state weights to account for frequency power adjustment */
7675 for (i
= 0; i
< 16; i
++) {
7676 u32 pxvidfreq
= I915_READ(PXVFREQ(i
));
7677 unsigned long freq
= intel_pxfreq(pxvidfreq
);
7678 unsigned long vid
= (pxvidfreq
& PXVFREQ_PX_MASK
) >>
7683 val
*= (freq
/ 1000);
7685 val
/= (127*127*900);
7687 DRM_ERROR("bad pxval: %ld\n", val
);
7690 /* Render standby states get 0 weight */
7694 for (i
= 0; i
< 4; i
++) {
7695 u32 val
= (pxw
[i
*4] << 24) | (pxw
[(i
*4)+1] << 16) |
7696 (pxw
[(i
*4)+2] << 8) | (pxw
[(i
*4)+3]);
7697 I915_WRITE(PXW(i
), val
);
7700 /* Adjust magic regs to magic values (more experimental results) */
7701 I915_WRITE(OGW0
, 0);
7702 I915_WRITE(OGW1
, 0);
7703 I915_WRITE(EG0
, 0x00007f00);
7704 I915_WRITE(EG1
, 0x0000000e);
7705 I915_WRITE(EG2
, 0x000e0000);
7706 I915_WRITE(EG3
, 0x68000300);
7707 I915_WRITE(EG4
, 0x42000000);
7708 I915_WRITE(EG5
, 0x00140031);
7712 for (i
= 0; i
< 8; i
++)
7713 I915_WRITE(PXWL(i
), 0);
7715 /* Enable PMON + select events */
7716 I915_WRITE(ECR
, 0x80000019);
7718 lcfuse
= I915_READ(LCFUSE02
);
7720 dev_priv
->ips
.corr
= (lcfuse
& LCFUSE_HIV_MASK
);
7723 void intel_init_gt_powersave(struct drm_i915_private
*dev_priv
)
7726 * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
7729 if (!i915
.enable_rc6
) {
7730 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
7731 intel_runtime_pm_get(dev_priv
);
7734 mutex_lock(&dev_priv
->drm
.struct_mutex
);
7735 mutex_lock(&dev_priv
->rps
.hw_lock
);
7737 /* Initialize RPS limits (for userspace) */
7738 if (IS_CHERRYVIEW(dev_priv
))
7739 cherryview_init_gt_powersave(dev_priv
);
7740 else if (IS_VALLEYVIEW(dev_priv
))
7741 valleyview_init_gt_powersave(dev_priv
);
7742 else if (INTEL_GEN(dev_priv
) >= 6)
7743 gen6_init_rps_frequencies(dev_priv
);
7745 /* Derive initial user preferences/limits from the hardware limits */
7746 dev_priv
->rps
.idle_freq
= dev_priv
->rps
.min_freq
;
7747 dev_priv
->rps
.cur_freq
= dev_priv
->rps
.idle_freq
;
7749 dev_priv
->rps
.max_freq_softlimit
= dev_priv
->rps
.max_freq
;
7750 dev_priv
->rps
.min_freq_softlimit
= dev_priv
->rps
.min_freq
;
7752 if (IS_HASWELL(dev_priv
) || IS_BROADWELL(dev_priv
))
7753 dev_priv
->rps
.min_freq_softlimit
=
7755 dev_priv
->rps
.efficient_freq
,
7756 intel_freq_opcode(dev_priv
, 450));
7758 /* After setting max-softlimit, find the overclock max freq */
7759 if (IS_GEN6(dev_priv
) ||
7760 IS_IVYBRIDGE(dev_priv
) || IS_HASWELL(dev_priv
)) {
7763 sandybridge_pcode_read(dev_priv
, GEN6_READ_OC_PARAMS
, ¶ms
);
7764 if (params
& BIT(31)) { /* OC supported */
7765 DRM_DEBUG_DRIVER("Overclocking supported, max: %dMHz, overclock: %dMHz\n",
7766 (dev_priv
->rps
.max_freq
& 0xff) * 50,
7767 (params
& 0xff) * 50);
7768 dev_priv
->rps
.max_freq
= params
& 0xff;
7772 /* Finally allow us to boost to max by default */
7773 dev_priv
->rps
.boost_freq
= dev_priv
->rps
.max_freq
;
7775 mutex_unlock(&dev_priv
->rps
.hw_lock
);
7776 mutex_unlock(&dev_priv
->drm
.struct_mutex
);
7778 intel_autoenable_gt_powersave(dev_priv
);
7781 void intel_cleanup_gt_powersave(struct drm_i915_private
*dev_priv
)
7783 if (IS_VALLEYVIEW(dev_priv
))
7784 valleyview_cleanup_gt_powersave(dev_priv
);
7786 if (!i915
.enable_rc6
)
7787 intel_runtime_pm_put(dev_priv
);
7791 * intel_suspend_gt_powersave - suspend PM work and helper threads
7792 * @dev_priv: i915 device
7794 * We don't want to disable RC6 or other features here, we just want
7795 * to make sure any work we've queued has finished and won't bother
7796 * us while we're suspended.
7798 void intel_suspend_gt_powersave(struct drm_i915_private
*dev_priv
)
7800 if (INTEL_GEN(dev_priv
) < 6)
7803 if (cancel_delayed_work_sync(&dev_priv
->rps
.autoenable_work
))
7804 intel_runtime_pm_put(dev_priv
);
7806 /* gen6_rps_idle() will be called later to disable interrupts */
7809 void intel_sanitize_gt_powersave(struct drm_i915_private
*dev_priv
)
7811 dev_priv
->rps
.enabled
= true; /* force disabling */
7812 intel_disable_gt_powersave(dev_priv
);
7814 gen6_reset_rps_interrupts(dev_priv
);
7817 void intel_disable_gt_powersave(struct drm_i915_private
*dev_priv
)
7819 if (!READ_ONCE(dev_priv
->rps
.enabled
))
7822 mutex_lock(&dev_priv
->rps
.hw_lock
);
7824 if (INTEL_GEN(dev_priv
) >= 9) {
7825 gen9_disable_rc6(dev_priv
);
7826 gen9_disable_rps(dev_priv
);
7827 } else if (IS_CHERRYVIEW(dev_priv
)) {
7828 cherryview_disable_rps(dev_priv
);
7829 } else if (IS_VALLEYVIEW(dev_priv
)) {
7830 valleyview_disable_rps(dev_priv
);
7831 } else if (INTEL_GEN(dev_priv
) >= 6) {
7832 gen6_disable_rps(dev_priv
);
7833 } else if (IS_IRONLAKE_M(dev_priv
)) {
7834 ironlake_disable_drps(dev_priv
);
7837 dev_priv
->rps
.enabled
= false;
7838 mutex_unlock(&dev_priv
->rps
.hw_lock
);
7841 void intel_enable_gt_powersave(struct drm_i915_private
*dev_priv
)
7843 /* We shouldn't be disabling as we submit, so this should be less
7844 * racy than it appears!
7846 if (READ_ONCE(dev_priv
->rps
.enabled
))
7849 /* Powersaving is controlled by the host when inside a VM */
7850 if (intel_vgpu_active(dev_priv
))
7853 mutex_lock(&dev_priv
->rps
.hw_lock
);
7855 if (IS_CHERRYVIEW(dev_priv
)) {
7856 cherryview_enable_rps(dev_priv
);
7857 } else if (IS_VALLEYVIEW(dev_priv
)) {
7858 valleyview_enable_rps(dev_priv
);
7859 } else if (INTEL_GEN(dev_priv
) >= 9) {
7860 gen9_enable_rc6(dev_priv
);
7861 gen9_enable_rps(dev_priv
);
7862 if (IS_GEN9_BC(dev_priv
) || IS_CANNONLAKE(dev_priv
))
7863 gen6_update_ring_freq(dev_priv
);
7864 } else if (IS_BROADWELL(dev_priv
)) {
7865 gen8_enable_rps(dev_priv
);
7866 gen6_update_ring_freq(dev_priv
);
7867 } else if (INTEL_GEN(dev_priv
) >= 6) {
7868 gen6_enable_rps(dev_priv
);
7869 gen6_update_ring_freq(dev_priv
);
7870 } else if (IS_IRONLAKE_M(dev_priv
)) {
7871 ironlake_enable_drps(dev_priv
);
7872 intel_init_emon(dev_priv
);
7875 WARN_ON(dev_priv
->rps
.max_freq
< dev_priv
->rps
.min_freq
);
7876 WARN_ON(dev_priv
->rps
.idle_freq
> dev_priv
->rps
.max_freq
);
7878 WARN_ON(dev_priv
->rps
.efficient_freq
< dev_priv
->rps
.min_freq
);
7879 WARN_ON(dev_priv
->rps
.efficient_freq
> dev_priv
->rps
.max_freq
);
7881 dev_priv
->rps
.enabled
= true;
7882 mutex_unlock(&dev_priv
->rps
.hw_lock
);
7885 static void __intel_autoenable_gt_powersave(struct work_struct
*work
)
7887 struct drm_i915_private
*dev_priv
=
7888 container_of(work
, typeof(*dev_priv
), rps
.autoenable_work
.work
);
7889 struct intel_engine_cs
*rcs
;
7890 struct drm_i915_gem_request
*req
;
7892 if (READ_ONCE(dev_priv
->rps
.enabled
))
7895 rcs
= dev_priv
->engine
[RCS
];
7896 if (rcs
->last_retired_context
)
7899 if (!rcs
->init_context
)
7902 mutex_lock(&dev_priv
->drm
.struct_mutex
);
7904 req
= i915_gem_request_alloc(rcs
, dev_priv
->kernel_context
);
7908 if (!i915
.enable_execlists
&& i915_switch_context(req
) == 0)
7909 rcs
->init_context(req
);
7911 /* Mark the device busy, calling intel_enable_gt_powersave() */
7912 i915_add_request(req
);
7915 mutex_unlock(&dev_priv
->drm
.struct_mutex
);
7917 intel_runtime_pm_put(dev_priv
);
7920 void intel_autoenable_gt_powersave(struct drm_i915_private
*dev_priv
)
7922 if (READ_ONCE(dev_priv
->rps
.enabled
))
7925 if (IS_IRONLAKE_M(dev_priv
)) {
7926 ironlake_enable_drps(dev_priv
);
7927 intel_init_emon(dev_priv
);
7928 } else if (INTEL_INFO(dev_priv
)->gen
>= 6) {
7930 * PCU communication is slow and this doesn't need to be
7931 * done at any specific time, so do this out of our fast path
7932 * to make resume and init faster.
7934 * We depend on the HW RC6 power context save/restore
7935 * mechanism when entering D3 through runtime PM suspend. So
7936 * disable RPM until RPS/RC6 is properly setup. We can only
7937 * get here via the driver load/system resume/runtime resume
7938 * paths, so the _noresume version is enough (and in case of
7939 * runtime resume it's necessary).
7941 if (queue_delayed_work(dev_priv
->wq
,
7942 &dev_priv
->rps
.autoenable_work
,
7943 round_jiffies_up_relative(HZ
)))
7944 intel_runtime_pm_get_noresume(dev_priv
);
7948 static void ibx_init_clock_gating(struct drm_i915_private
*dev_priv
)
7951 * On Ibex Peak and Cougar Point, we need to disable clock
7952 * gating for the panel power sequencer or it will fail to
7953 * start up when no ports are active.
7955 I915_WRITE(SOUTH_DSPCLK_GATE_D
, PCH_DPLSUNIT_CLOCK_GATE_DISABLE
);
7958 static void g4x_disable_trickle_feed(struct drm_i915_private
*dev_priv
)
7962 for_each_pipe(dev_priv
, pipe
) {
7963 I915_WRITE(DSPCNTR(pipe
),
7964 I915_READ(DSPCNTR(pipe
)) |
7965 DISPPLANE_TRICKLE_FEED_DISABLE
);
7967 I915_WRITE(DSPSURF(pipe
), I915_READ(DSPSURF(pipe
)));
7968 POSTING_READ(DSPSURF(pipe
));
7972 static void ilk_init_lp_watermarks(struct drm_i915_private
*dev_priv
)
7974 I915_WRITE(WM3_LP_ILK
, I915_READ(WM3_LP_ILK
) & ~WM1_LP_SR_EN
);
7975 I915_WRITE(WM2_LP_ILK
, I915_READ(WM2_LP_ILK
) & ~WM1_LP_SR_EN
);
7976 I915_WRITE(WM1_LP_ILK
, I915_READ(WM1_LP_ILK
) & ~WM1_LP_SR_EN
);
7979 * Don't touch WM1S_LP_EN here.
7980 * Doing so could cause underruns.
7984 static void ilk_init_clock_gating(struct drm_i915_private
*dev_priv
)
7986 uint32_t dspclk_gate
= ILK_VRHUNIT_CLOCK_GATE_DISABLE
;
7990 * WaFbcDisableDpfcClockGating:ilk
7992 dspclk_gate
|= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE
|
7993 ILK_DPFCUNIT_CLOCK_GATE_DISABLE
|
7994 ILK_DPFDUNIT_CLOCK_GATE_ENABLE
;
7996 I915_WRITE(PCH_3DCGDIS0
,
7997 MARIUNIT_CLOCK_GATE_DISABLE
|
7998 SVSMUNIT_CLOCK_GATE_DISABLE
);
7999 I915_WRITE(PCH_3DCGDIS1
,
8000 VFMUNIT_CLOCK_GATE_DISABLE
);
8003 * According to the spec the following bits should be set in
8004 * order to enable memory self-refresh
8005 * The bit 22/21 of 0x42004
8006 * The bit 5 of 0x42020
8007 * The bit 15 of 0x45000
8009 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
8010 (I915_READ(ILK_DISPLAY_CHICKEN2
) |
8011 ILK_DPARB_GATE
| ILK_VSDPFD_FULL
));
8012 dspclk_gate
|= ILK_DPARBUNIT_CLOCK_GATE_ENABLE
;
8013 I915_WRITE(DISP_ARB_CTL
,
8014 (I915_READ(DISP_ARB_CTL
) |
8017 ilk_init_lp_watermarks(dev_priv
);
8020 * Based on the document from hardware guys the following bits
8021 * should be set unconditionally in order to enable FBC.
8022 * The bit 22 of 0x42000
8023 * The bit 22 of 0x42004
8024 * The bit 7,8,9 of 0x42020.
8026 if (IS_IRONLAKE_M(dev_priv
)) {
8027 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
8028 I915_WRITE(ILK_DISPLAY_CHICKEN1
,
8029 I915_READ(ILK_DISPLAY_CHICKEN1
) |
8031 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
8032 I915_READ(ILK_DISPLAY_CHICKEN2
) |
8036 I915_WRITE(ILK_DSPCLK_GATE_D
, dspclk_gate
);
8038 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
8039 I915_READ(ILK_DISPLAY_CHICKEN2
) |
8040 ILK_ELPIN_409_SELECT
);
8041 I915_WRITE(_3D_CHICKEN2
,
8042 _3D_CHICKEN2_WM_READ_PIPELINED
<< 16 |
8043 _3D_CHICKEN2_WM_READ_PIPELINED
);
8045 /* WaDisableRenderCachePipelinedFlush:ilk */
8046 I915_WRITE(CACHE_MODE_0
,
8047 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE
));
8049 /* WaDisable_RenderCache_OperationalFlush:ilk */
8050 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
8052 g4x_disable_trickle_feed(dev_priv
);
8054 ibx_init_clock_gating(dev_priv
);
8057 static void cpt_init_clock_gating(struct drm_i915_private
*dev_priv
)
8063 * On Ibex Peak and Cougar Point, we need to disable clock
8064 * gating for the panel power sequencer or it will fail to
8065 * start up when no ports are active.
8067 I915_WRITE(SOUTH_DSPCLK_GATE_D
, PCH_DPLSUNIT_CLOCK_GATE_DISABLE
|
8068 PCH_DPLUNIT_CLOCK_GATE_DISABLE
|
8069 PCH_CPUNIT_CLOCK_GATE_DISABLE
);
8070 I915_WRITE(SOUTH_CHICKEN2
, I915_READ(SOUTH_CHICKEN2
) |
8071 DPLS_EDP_PPS_FIX_DIS
);
8072 /* The below fixes the weird display corruption, a few pixels shifted
8073 * downward, on (only) LVDS of some HP laptops with IVY.
8075 for_each_pipe(dev_priv
, pipe
) {
8076 val
= I915_READ(TRANS_CHICKEN2(pipe
));
8077 val
|= TRANS_CHICKEN2_TIMING_OVERRIDE
;
8078 val
&= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED
;
8079 if (dev_priv
->vbt
.fdi_rx_polarity_inverted
)
8080 val
|= TRANS_CHICKEN2_FDI_POLARITY_REVERSED
;
8081 val
&= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK
;
8082 val
&= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER
;
8083 val
&= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH
;
8084 I915_WRITE(TRANS_CHICKEN2(pipe
), val
);
8086 /* WADP0ClockGatingDisable */
8087 for_each_pipe(dev_priv
, pipe
) {
8088 I915_WRITE(TRANS_CHICKEN1(pipe
),
8089 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE
);
8093 static void gen6_check_mch_setup(struct drm_i915_private
*dev_priv
)
8097 tmp
= I915_READ(MCH_SSKPD
);
8098 if ((tmp
& MCH_SSKPD_WM0_MASK
) != MCH_SSKPD_WM0_VAL
)
8099 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
8103 static void gen6_init_clock_gating(struct drm_i915_private
*dev_priv
)
8105 uint32_t dspclk_gate
= ILK_VRHUNIT_CLOCK_GATE_DISABLE
;
8107 I915_WRITE(ILK_DSPCLK_GATE_D
, dspclk_gate
);
8109 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
8110 I915_READ(ILK_DISPLAY_CHICKEN2
) |
8111 ILK_ELPIN_409_SELECT
);
8113 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
8114 I915_WRITE(_3D_CHICKEN
,
8115 _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB
));
8117 /* WaDisable_RenderCache_OperationalFlush:snb */
8118 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
8121 * BSpec recoomends 8x4 when MSAA is used,
8122 * however in practice 16x4 seems fastest.
8124 * Note that PS/WM thread counts depend on the WIZ hashing
8125 * disable bit, which we don't touch here, but it's good
8126 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
8128 I915_WRITE(GEN6_GT_MODE
,
8129 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
8131 ilk_init_lp_watermarks(dev_priv
);
8133 I915_WRITE(CACHE_MODE_0
,
8134 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB
));
8136 I915_WRITE(GEN6_UCGCTL1
,
8137 I915_READ(GEN6_UCGCTL1
) |
8138 GEN6_BLBUNIT_CLOCK_GATE_DISABLE
|
8139 GEN6_CSUNIT_CLOCK_GATE_DISABLE
);
8141 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
8142 * gating disable must be set. Failure to set it results in
8143 * flickering pixels due to Z write ordering failures after
8144 * some amount of runtime in the Mesa "fire" demo, and Unigine
8145 * Sanctuary and Tropics, and apparently anything else with
8146 * alpha test or pixel discard.
8148 * According to the spec, bit 11 (RCCUNIT) must also be set,
8149 * but we didn't debug actual testcases to find it out.
8151 * WaDisableRCCUnitClockGating:snb
8152 * WaDisableRCPBUnitClockGating:snb
8154 I915_WRITE(GEN6_UCGCTL2
,
8155 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE
|
8156 GEN6_RCCUNIT_CLOCK_GATE_DISABLE
);
8158 /* WaStripsFansDisableFastClipPerformanceFix:snb */
8159 I915_WRITE(_3D_CHICKEN3
,
8160 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL
));
8164 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
8165 * 3DSTATE_SF number of SF output attributes is more than 16."
8167 I915_WRITE(_3D_CHICKEN3
,
8168 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH
));
8171 * According to the spec the following bits should be
8172 * set in order to enable memory self-refresh and fbc:
8173 * The bit21 and bit22 of 0x42000
8174 * The bit21 and bit22 of 0x42004
8175 * The bit5 and bit7 of 0x42020
8176 * The bit14 of 0x70180
8177 * The bit14 of 0x71180
8179 * WaFbcAsynchFlipDisableFbcQueue:snb
8181 I915_WRITE(ILK_DISPLAY_CHICKEN1
,
8182 I915_READ(ILK_DISPLAY_CHICKEN1
) |
8183 ILK_FBCQ_DIS
| ILK_PABSTRETCH_DIS
);
8184 I915_WRITE(ILK_DISPLAY_CHICKEN2
,
8185 I915_READ(ILK_DISPLAY_CHICKEN2
) |
8186 ILK_DPARB_GATE
| ILK_VSDPFD_FULL
);
8187 I915_WRITE(ILK_DSPCLK_GATE_D
,
8188 I915_READ(ILK_DSPCLK_GATE_D
) |
8189 ILK_DPARBUNIT_CLOCK_GATE_ENABLE
|
8190 ILK_DPFDUNIT_CLOCK_GATE_ENABLE
);
8192 g4x_disable_trickle_feed(dev_priv
);
8194 cpt_init_clock_gating(dev_priv
);
8196 gen6_check_mch_setup(dev_priv
);
8199 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private
*dev_priv
)
8201 uint32_t reg
= I915_READ(GEN7_FF_THREAD_MODE
);
8204 * WaVSThreadDispatchOverride:ivb,vlv
8206 * This actually overrides the dispatch
8207 * mode for all thread types.
8209 reg
&= ~GEN7_FF_SCHED_MASK
;
8210 reg
|= GEN7_FF_TS_SCHED_HW
;
8211 reg
|= GEN7_FF_VS_SCHED_HW
;
8212 reg
|= GEN7_FF_DS_SCHED_HW
;
8214 I915_WRITE(GEN7_FF_THREAD_MODE
, reg
);
8217 static void lpt_init_clock_gating(struct drm_i915_private
*dev_priv
)
8220 * TODO: this bit should only be enabled when really needed, then
8221 * disabled when not needed anymore in order to save power.
8223 if (HAS_PCH_LPT_LP(dev_priv
))
8224 I915_WRITE(SOUTH_DSPCLK_GATE_D
,
8225 I915_READ(SOUTH_DSPCLK_GATE_D
) |
8226 PCH_LP_PARTITION_LEVEL_DISABLE
);
8228 /* WADPOClockGatingDisable:hsw */
8229 I915_WRITE(TRANS_CHICKEN1(PIPE_A
),
8230 I915_READ(TRANS_CHICKEN1(PIPE_A
)) |
8231 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE
);
8234 static void lpt_suspend_hw(struct drm_i915_private
*dev_priv
)
8236 if (HAS_PCH_LPT_LP(dev_priv
)) {
8237 uint32_t val
= I915_READ(SOUTH_DSPCLK_GATE_D
);
8239 val
&= ~PCH_LP_PARTITION_LEVEL_DISABLE
;
8240 I915_WRITE(SOUTH_DSPCLK_GATE_D
, val
);
8244 static void gen8_set_l3sqc_credits(struct drm_i915_private
*dev_priv
,
8245 int general_prio_credits
,
8246 int high_prio_credits
)
8250 /* WaTempDisableDOPClkGating:bdw */
8251 misccpctl
= I915_READ(GEN7_MISCCPCTL
);
8252 I915_WRITE(GEN7_MISCCPCTL
, misccpctl
& ~GEN7_DOP_CLOCK_GATE_ENABLE
);
8254 I915_WRITE(GEN8_L3SQCREG1
,
8255 L3_GENERAL_PRIO_CREDITS(general_prio_credits
) |
8256 L3_HIGH_PRIO_CREDITS(high_prio_credits
));
8259 * Wait at least 100 clocks before re-enabling clock gating.
8260 * See the definition of L3SQCREG1 in BSpec.
8262 POSTING_READ(GEN8_L3SQCREG1
);
8264 I915_WRITE(GEN7_MISCCPCTL
, misccpctl
);
8267 static void cnl_init_clock_gating(struct drm_i915_private
*dev_priv
)
8269 /* This is not an Wa. Enable for better image quality */
8270 I915_WRITE(_3D_CHICKEN3
,
8271 _MASKED_BIT_ENABLE(_3D_CHICKEN3_AA_LINE_QUALITY_FIX_ENABLE
));
8273 /* WaEnableChickenDCPR:cnl */
8274 I915_WRITE(GEN8_CHICKEN_DCPR_1
,
8275 I915_READ(GEN8_CHICKEN_DCPR_1
) | MASK_WAKEMEM
);
8277 /* WaFbcWakeMemOn:cnl */
8278 I915_WRITE(DISP_ARB_CTL
, I915_READ(DISP_ARB_CTL
) |
8279 DISP_FBC_MEMORY_WAKE
);
8281 /* WaSarbUnitClockGatingDisable:cnl (pre-prod) */
8282 if (IS_CNL_REVID(dev_priv
, CNL_REVID_A0
, CNL_REVID_B0
))
8283 I915_WRITE(SLICE_UNIT_LEVEL_CLKGATE
,
8284 I915_READ(SLICE_UNIT_LEVEL_CLKGATE
) |
8285 SARBUNIT_CLKGATE_DIS
);
8288 static void kbl_init_clock_gating(struct drm_i915_private
*dev_priv
)
8290 gen9_init_clock_gating(dev_priv
);
8292 /* WaDisableSDEUnitClockGating:kbl */
8293 if (IS_KBL_REVID(dev_priv
, 0, KBL_REVID_B0
))
8294 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
8295 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
8297 /* WaDisableGamClockGating:kbl */
8298 if (IS_KBL_REVID(dev_priv
, 0, KBL_REVID_B0
))
8299 I915_WRITE(GEN6_UCGCTL1
, I915_READ(GEN6_UCGCTL1
) |
8300 GEN6_GAMUNIT_CLOCK_GATE_DISABLE
);
8302 /* WaFbcNukeOnHostModify:kbl,cfl */
8303 I915_WRITE(ILK_DPFC_CHICKEN
, I915_READ(ILK_DPFC_CHICKEN
) |
8304 ILK_DPFC_NUKE_ON_ANY_MODIFICATION
);
8307 static void skl_init_clock_gating(struct drm_i915_private
*dev_priv
)
8309 gen9_init_clock_gating(dev_priv
);
8311 /* WAC6entrylatency:skl */
8312 I915_WRITE(FBC_LLC_READ_CTRL
, I915_READ(FBC_LLC_READ_CTRL
) |
8313 FBC_LLC_FULLY_OPEN
);
8315 /* WaFbcNukeOnHostModify:skl */
8316 I915_WRITE(ILK_DPFC_CHICKEN
, I915_READ(ILK_DPFC_CHICKEN
) |
8317 ILK_DPFC_NUKE_ON_ANY_MODIFICATION
);
8320 static void bdw_init_clock_gating(struct drm_i915_private
*dev_priv
)
8324 ilk_init_lp_watermarks(dev_priv
);
8326 /* WaSwitchSolVfFArbitrationPriority:bdw */
8327 I915_WRITE(GAM_ECOCHK
, I915_READ(GAM_ECOCHK
) | HSW_ECOCHK_ARB_PRIO_SOL
);
8329 /* WaPsrDPAMaskVBlankInSRD:bdw */
8330 I915_WRITE(CHICKEN_PAR1_1
,
8331 I915_READ(CHICKEN_PAR1_1
) | DPA_MASK_VBLANK_SRD
);
8333 /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
8334 for_each_pipe(dev_priv
, pipe
) {
8335 I915_WRITE(CHICKEN_PIPESL_1(pipe
),
8336 I915_READ(CHICKEN_PIPESL_1(pipe
)) |
8337 BDW_DPRS_MASK_VBLANK_SRD
);
8340 /* WaVSRefCountFullforceMissDisable:bdw */
8341 /* WaDSRefCountFullforceMissDisable:bdw */
8342 I915_WRITE(GEN7_FF_THREAD_MODE
,
8343 I915_READ(GEN7_FF_THREAD_MODE
) &
8344 ~(GEN8_FF_DS_REF_CNT_FFME
| GEN7_FF_VS_REF_CNT_FFME
));
8346 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL
,
8347 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE
));
8349 /* WaDisableSDEUnitClockGating:bdw */
8350 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
8351 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
8353 /* WaProgramL3SqcReg1Default:bdw */
8354 gen8_set_l3sqc_credits(dev_priv
, 30, 2);
8357 * WaGttCachingOffByDefault:bdw
8358 * GTT cache may not work with big pages, so if those
8359 * are ever enabled GTT cache may need to be disabled.
8361 I915_WRITE(HSW_GTT_CACHE_EN
, GTT_CACHE_EN_ALL
);
8363 /* WaKVMNotificationOnConfigChange:bdw */
8364 I915_WRITE(CHICKEN_PAR2_1
, I915_READ(CHICKEN_PAR2_1
)
8365 | KVM_CONFIG_CHANGE_NOTIFICATION_SELECT
);
8367 lpt_init_clock_gating(dev_priv
);
8369 /* WaDisableDopClockGating:bdw
8371 * Also see the CHICKEN2 write in bdw_init_workarounds() to disable DOP
8374 I915_WRITE(GEN6_UCGCTL1
,
8375 I915_READ(GEN6_UCGCTL1
) | GEN6_EU_TCUNIT_CLOCK_GATE_DISABLE
);
8378 static void hsw_init_clock_gating(struct drm_i915_private
*dev_priv
)
8380 ilk_init_lp_watermarks(dev_priv
);
8382 /* L3 caching of data atomics doesn't work -- disable it. */
8383 I915_WRITE(HSW_SCRATCH1
, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE
);
8384 I915_WRITE(HSW_ROW_CHICKEN3
,
8385 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE
));
8387 /* This is required by WaCatErrorRejectionIssue:hsw */
8388 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
8389 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
8390 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
8392 /* WaVSRefCountFullforceMissDisable:hsw */
8393 I915_WRITE(GEN7_FF_THREAD_MODE
,
8394 I915_READ(GEN7_FF_THREAD_MODE
) & ~GEN7_FF_VS_REF_CNT_FFME
);
8396 /* WaDisable_RenderCache_OperationalFlush:hsw */
8397 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
8399 /* enable HiZ Raw Stall Optimization */
8400 I915_WRITE(CACHE_MODE_0_GEN7
,
8401 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE
));
8403 /* WaDisable4x2SubspanOptimization:hsw */
8404 I915_WRITE(CACHE_MODE_1
,
8405 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
8408 * BSpec recommends 8x4 when MSAA is used,
8409 * however in practice 16x4 seems fastest.
8411 * Note that PS/WM thread counts depend on the WIZ hashing
8412 * disable bit, which we don't touch here, but it's good
8413 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
8415 I915_WRITE(GEN7_GT_MODE
,
8416 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
8418 /* WaSampleCChickenBitEnable:hsw */
8419 I915_WRITE(HALF_SLICE_CHICKEN3
,
8420 _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE
));
8422 /* WaSwitchSolVfFArbitrationPriority:hsw */
8423 I915_WRITE(GAM_ECOCHK
, I915_READ(GAM_ECOCHK
) | HSW_ECOCHK_ARB_PRIO_SOL
);
8425 /* WaRsPkgCStateDisplayPMReq:hsw */
8426 I915_WRITE(CHICKEN_PAR1_1
,
8427 I915_READ(CHICKEN_PAR1_1
) | FORCE_ARB_IDLE_PLANES
);
8429 lpt_init_clock_gating(dev_priv
);
8432 static void ivb_init_clock_gating(struct drm_i915_private
*dev_priv
)
8436 ilk_init_lp_watermarks(dev_priv
);
8438 I915_WRITE(ILK_DSPCLK_GATE_D
, ILK_VRHUNIT_CLOCK_GATE_DISABLE
);
8440 /* WaDisableEarlyCull:ivb */
8441 I915_WRITE(_3D_CHICKEN3
,
8442 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL
));
8444 /* WaDisableBackToBackFlipFix:ivb */
8445 I915_WRITE(IVB_CHICKEN3
,
8446 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE
|
8447 CHICKEN3_DGMG_DONE_FIX_DISABLE
);
8449 /* WaDisablePSDDualDispatchEnable:ivb */
8450 if (IS_IVB_GT1(dev_priv
))
8451 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1
,
8452 _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE
));
8454 /* WaDisable_RenderCache_OperationalFlush:ivb */
8455 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
8457 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
8458 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1
,
8459 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC
);
8461 /* WaApplyL3ControlAndL3ChickenMode:ivb */
8462 I915_WRITE(GEN7_L3CNTLREG1
,
8463 GEN7_WA_FOR_GEN7_L3_CONTROL
);
8464 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER
,
8465 GEN7_WA_L3_CHICKEN_MODE
);
8466 if (IS_IVB_GT1(dev_priv
))
8467 I915_WRITE(GEN7_ROW_CHICKEN2
,
8468 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
8470 /* must write both registers */
8471 I915_WRITE(GEN7_ROW_CHICKEN2
,
8472 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
8473 I915_WRITE(GEN7_ROW_CHICKEN2_GT2
,
8474 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
8477 /* WaForceL3Serialization:ivb */
8478 I915_WRITE(GEN7_L3SQCREG4
, I915_READ(GEN7_L3SQCREG4
) &
8479 ~L3SQ_URB_READ_CAM_MATCH_DISABLE
);
8482 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
8483 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
8485 I915_WRITE(GEN6_UCGCTL2
,
8486 GEN6_RCZUNIT_CLOCK_GATE_DISABLE
);
8488 /* This is required by WaCatErrorRejectionIssue:ivb */
8489 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
8490 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
8491 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
8493 g4x_disable_trickle_feed(dev_priv
);
8495 gen7_setup_fixed_func_scheduler(dev_priv
);
8497 if (0) { /* causes HiZ corruption on ivb:gt1 */
8498 /* enable HiZ Raw Stall Optimization */
8499 I915_WRITE(CACHE_MODE_0_GEN7
,
8500 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE
));
8503 /* WaDisable4x2SubspanOptimization:ivb */
8504 I915_WRITE(CACHE_MODE_1
,
8505 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
8508 * BSpec recommends 8x4 when MSAA is used,
8509 * however in practice 16x4 seems fastest.
8511 * Note that PS/WM thread counts depend on the WIZ hashing
8512 * disable bit, which we don't touch here, but it's good
8513 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
8515 I915_WRITE(GEN7_GT_MODE
,
8516 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
8518 snpcr
= I915_READ(GEN6_MBCUNIT_SNPCR
);
8519 snpcr
&= ~GEN6_MBC_SNPCR_MASK
;
8520 snpcr
|= GEN6_MBC_SNPCR_MED
;
8521 I915_WRITE(GEN6_MBCUNIT_SNPCR
, snpcr
);
8523 if (!HAS_PCH_NOP(dev_priv
))
8524 cpt_init_clock_gating(dev_priv
);
8526 gen6_check_mch_setup(dev_priv
);
8529 static void vlv_init_clock_gating(struct drm_i915_private
*dev_priv
)
8531 /* WaDisableEarlyCull:vlv */
8532 I915_WRITE(_3D_CHICKEN3
,
8533 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL
));
8535 /* WaDisableBackToBackFlipFix:vlv */
8536 I915_WRITE(IVB_CHICKEN3
,
8537 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE
|
8538 CHICKEN3_DGMG_DONE_FIX_DISABLE
);
8540 /* WaPsdDispatchEnable:vlv */
8541 /* WaDisablePSDDualDispatchEnable:vlv */
8542 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1
,
8543 _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP
|
8544 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE
));
8546 /* WaDisable_RenderCache_OperationalFlush:vlv */
8547 I915_WRITE(CACHE_MODE_0_GEN7
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
8549 /* WaForceL3Serialization:vlv */
8550 I915_WRITE(GEN7_L3SQCREG4
, I915_READ(GEN7_L3SQCREG4
) &
8551 ~L3SQ_URB_READ_CAM_MATCH_DISABLE
);
8553 /* WaDisableDopClockGating:vlv */
8554 I915_WRITE(GEN7_ROW_CHICKEN2
,
8555 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE
));
8557 /* This is required by WaCatErrorRejectionIssue:vlv */
8558 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
,
8559 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG
) |
8560 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB
);
8562 gen7_setup_fixed_func_scheduler(dev_priv
);
8565 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
8566 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
8568 I915_WRITE(GEN6_UCGCTL2
,
8569 GEN6_RCZUNIT_CLOCK_GATE_DISABLE
);
8571 /* WaDisableL3Bank2xClockGate:vlv
8572 * Disabling L3 clock gating- MMIO 940c[25] = 1
8573 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
8574 I915_WRITE(GEN7_UCGCTL4
,
8575 I915_READ(GEN7_UCGCTL4
) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE
);
8578 * BSpec says this must be set, even though
8579 * WaDisable4x2SubspanOptimization isn't listed for VLV.
8581 I915_WRITE(CACHE_MODE_1
,
8582 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE
));
8585 * BSpec recommends 8x4 when MSAA is used,
8586 * however in practice 16x4 seems fastest.
8588 * Note that PS/WM thread counts depend on the WIZ hashing
8589 * disable bit, which we don't touch here, but it's good
8590 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
8592 I915_WRITE(GEN7_GT_MODE
,
8593 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK
, GEN6_WIZ_HASHING_16x4
));
8596 * WaIncreaseL3CreditsForVLVB0:vlv
8597 * This is the hardware default actually.
8599 I915_WRITE(GEN7_L3SQCREG1
, VLV_B0_WA_L3SQCREG1_VALUE
);
8602 * WaDisableVLVClockGating_VBIIssue:vlv
8603 * Disable clock gating on th GCFG unit to prevent a delay
8604 * in the reporting of vblank events.
8606 I915_WRITE(VLV_GUNIT_CLOCK_GATE
, GCFG_DIS
);
8609 static void chv_init_clock_gating(struct drm_i915_private
*dev_priv
)
8611 /* WaVSRefCountFullforceMissDisable:chv */
8612 /* WaDSRefCountFullforceMissDisable:chv */
8613 I915_WRITE(GEN7_FF_THREAD_MODE
,
8614 I915_READ(GEN7_FF_THREAD_MODE
) &
8615 ~(GEN8_FF_DS_REF_CNT_FFME
| GEN7_FF_VS_REF_CNT_FFME
));
8617 /* WaDisableSemaphoreAndSyncFlipWait:chv */
8618 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL
,
8619 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE
));
8621 /* WaDisableCSUnitClockGating:chv */
8622 I915_WRITE(GEN6_UCGCTL1
, I915_READ(GEN6_UCGCTL1
) |
8623 GEN6_CSUNIT_CLOCK_GATE_DISABLE
);
8625 /* WaDisableSDEUnitClockGating:chv */
8626 I915_WRITE(GEN8_UCGCTL6
, I915_READ(GEN8_UCGCTL6
) |
8627 GEN8_SDEUNIT_CLOCK_GATE_DISABLE
);
8630 * WaProgramL3SqcReg1Default:chv
8631 * See gfxspecs/Related Documents/Performance Guide/
8632 * LSQC Setting Recommendations.
8634 gen8_set_l3sqc_credits(dev_priv
, 38, 2);
8637 * GTT cache may not work with big pages, so if those
8638 * are ever enabled GTT cache may need to be disabled.
8640 I915_WRITE(HSW_GTT_CACHE_EN
, GTT_CACHE_EN_ALL
);
8643 static void g4x_init_clock_gating(struct drm_i915_private
*dev_priv
)
8645 uint32_t dspclk_gate
;
8647 I915_WRITE(RENCLK_GATE_D1
, 0);
8648 I915_WRITE(RENCLK_GATE_D2
, VF_UNIT_CLOCK_GATE_DISABLE
|
8649 GS_UNIT_CLOCK_GATE_DISABLE
|
8650 CL_UNIT_CLOCK_GATE_DISABLE
);
8651 I915_WRITE(RAMCLK_GATE_D
, 0);
8652 dspclk_gate
= VRHUNIT_CLOCK_GATE_DISABLE
|
8653 OVRUNIT_CLOCK_GATE_DISABLE
|
8654 OVCUNIT_CLOCK_GATE_DISABLE
;
8655 if (IS_GM45(dev_priv
))
8656 dspclk_gate
|= DSSUNIT_CLOCK_GATE_DISABLE
;
8657 I915_WRITE(DSPCLK_GATE_D
, dspclk_gate
);
8659 /* WaDisableRenderCachePipelinedFlush */
8660 I915_WRITE(CACHE_MODE_0
,
8661 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE
));
8663 /* WaDisable_RenderCache_OperationalFlush:g4x */
8664 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
8666 g4x_disable_trickle_feed(dev_priv
);
8669 static void i965gm_init_clock_gating(struct drm_i915_private
*dev_priv
)
8671 I915_WRITE(RENCLK_GATE_D1
, I965_RCC_CLOCK_GATE_DISABLE
);
8672 I915_WRITE(RENCLK_GATE_D2
, 0);
8673 I915_WRITE(DSPCLK_GATE_D
, 0);
8674 I915_WRITE(RAMCLK_GATE_D
, 0);
8675 I915_WRITE16(DEUC
, 0);
8676 I915_WRITE(MI_ARB_STATE
,
8677 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
8679 /* WaDisable_RenderCache_OperationalFlush:gen4 */
8680 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
8683 static void i965g_init_clock_gating(struct drm_i915_private
*dev_priv
)
8685 I915_WRITE(RENCLK_GATE_D1
, I965_RCZ_CLOCK_GATE_DISABLE
|
8686 I965_RCC_CLOCK_GATE_DISABLE
|
8687 I965_RCPB_CLOCK_GATE_DISABLE
|
8688 I965_ISC_CLOCK_GATE_DISABLE
|
8689 I965_FBC_CLOCK_GATE_DISABLE
);
8690 I915_WRITE(RENCLK_GATE_D2
, 0);
8691 I915_WRITE(MI_ARB_STATE
,
8692 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
8694 /* WaDisable_RenderCache_OperationalFlush:gen4 */
8695 I915_WRITE(CACHE_MODE_0
, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE
));
8698 static void gen3_init_clock_gating(struct drm_i915_private
*dev_priv
)
8700 u32 dstate
= I915_READ(D_STATE
);
8702 dstate
|= DSTATE_PLL_D3_OFF
| DSTATE_GFX_CLOCK_GATING
|
8703 DSTATE_DOT_CLOCK_GATING
;
8704 I915_WRITE(D_STATE
, dstate
);
8706 if (IS_PINEVIEW(dev_priv
))
8707 I915_WRITE(ECOSKPD
, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY
));
8709 /* IIR "flip pending" means done if this bit is set */
8710 I915_WRITE(ECOSKPD
, _MASKED_BIT_DISABLE(ECO_FLIP_DONE
));
8712 /* interrupts should cause a wake up from C3 */
8713 I915_WRITE(INSTPM
, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN
));
8715 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
8716 I915_WRITE(MI_ARB_STATE
, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE
));
8718 I915_WRITE(MI_ARB_STATE
,
8719 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE
));
8722 static void i85x_init_clock_gating(struct drm_i915_private
*dev_priv
)
8724 I915_WRITE(RENCLK_GATE_D1
, SV_CLOCK_GATE_DISABLE
);
8726 /* interrupts should cause a wake up from C3 */
8727 I915_WRITE(MI_STATE
, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN
) |
8728 _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE
));
8730 I915_WRITE(MEM_MODE
,
8731 _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE
));
8734 static void i830_init_clock_gating(struct drm_i915_private
*dev_priv
)
8736 I915_WRITE(MEM_MODE
,
8737 _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE
) |
8738 _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE
));
8741 void intel_init_clock_gating(struct drm_i915_private
*dev_priv
)
8743 dev_priv
->display
.init_clock_gating(dev_priv
);
8746 void intel_suspend_hw(struct drm_i915_private
*dev_priv
)
8748 if (HAS_PCH_LPT(dev_priv
))
8749 lpt_suspend_hw(dev_priv
);
8752 static void nop_init_clock_gating(struct drm_i915_private
*dev_priv
)
8754 DRM_DEBUG_KMS("No clock gating settings or workarounds applied.\n");
8758 * intel_init_clock_gating_hooks - setup the clock gating hooks
8759 * @dev_priv: device private
8761 * Setup the hooks that configure which clocks of a given platform can be
8762 * gated and also apply various GT and display specific workarounds for these
8763 * platforms. Note that some GT specific workarounds are applied separately
8764 * when GPU contexts or batchbuffers start their execution.
8766 void intel_init_clock_gating_hooks(struct drm_i915_private
*dev_priv
)
8768 if (IS_CANNONLAKE(dev_priv
))
8769 dev_priv
->display
.init_clock_gating
= cnl_init_clock_gating
;
8770 else if (IS_SKYLAKE(dev_priv
))
8771 dev_priv
->display
.init_clock_gating
= skl_init_clock_gating
;
8772 else if (IS_KABYLAKE(dev_priv
) || IS_COFFEELAKE(dev_priv
))
8773 dev_priv
->display
.init_clock_gating
= kbl_init_clock_gating
;
8774 else if (IS_BROXTON(dev_priv
))
8775 dev_priv
->display
.init_clock_gating
= bxt_init_clock_gating
;
8776 else if (IS_GEMINILAKE(dev_priv
))
8777 dev_priv
->display
.init_clock_gating
= glk_init_clock_gating
;
8778 else if (IS_BROADWELL(dev_priv
))
8779 dev_priv
->display
.init_clock_gating
= bdw_init_clock_gating
;
8780 else if (IS_CHERRYVIEW(dev_priv
))
8781 dev_priv
->display
.init_clock_gating
= chv_init_clock_gating
;
8782 else if (IS_HASWELL(dev_priv
))
8783 dev_priv
->display
.init_clock_gating
= hsw_init_clock_gating
;
8784 else if (IS_IVYBRIDGE(dev_priv
))
8785 dev_priv
->display
.init_clock_gating
= ivb_init_clock_gating
;
8786 else if (IS_VALLEYVIEW(dev_priv
))
8787 dev_priv
->display
.init_clock_gating
= vlv_init_clock_gating
;
8788 else if (IS_GEN6(dev_priv
))
8789 dev_priv
->display
.init_clock_gating
= gen6_init_clock_gating
;
8790 else if (IS_GEN5(dev_priv
))
8791 dev_priv
->display
.init_clock_gating
= ilk_init_clock_gating
;
8792 else if (IS_G4X(dev_priv
))
8793 dev_priv
->display
.init_clock_gating
= g4x_init_clock_gating
;
8794 else if (IS_I965GM(dev_priv
))
8795 dev_priv
->display
.init_clock_gating
= i965gm_init_clock_gating
;
8796 else if (IS_I965G(dev_priv
))
8797 dev_priv
->display
.init_clock_gating
= i965g_init_clock_gating
;
8798 else if (IS_GEN3(dev_priv
))
8799 dev_priv
->display
.init_clock_gating
= gen3_init_clock_gating
;
8800 else if (IS_I85X(dev_priv
) || IS_I865G(dev_priv
))
8801 dev_priv
->display
.init_clock_gating
= i85x_init_clock_gating
;
8802 else if (IS_GEN2(dev_priv
))
8803 dev_priv
->display
.init_clock_gating
= i830_init_clock_gating
;
8805 MISSING_CASE(INTEL_DEVID(dev_priv
));
8806 dev_priv
->display
.init_clock_gating
= nop_init_clock_gating
;
8810 /* Set up chip specific power management-related functions */
8811 void intel_init_pm(struct drm_i915_private
*dev_priv
)
8813 intel_fbc_init(dev_priv
);
8816 if (IS_PINEVIEW(dev_priv
))
8817 i915_pineview_get_mem_freq(dev_priv
);
8818 else if (IS_GEN5(dev_priv
))
8819 i915_ironlake_get_mem_freq(dev_priv
);
8821 /* For FIFO watermark updates */
8822 if (INTEL_GEN(dev_priv
) >= 9) {
8823 skl_setup_wm_latency(dev_priv
);
8824 dev_priv
->display
.initial_watermarks
= skl_initial_wm
;
8825 dev_priv
->display
.atomic_update_watermarks
= skl_atomic_update_crtc_wm
;
8826 dev_priv
->display
.compute_global_watermarks
= skl_compute_wm
;
8827 } else if (HAS_PCH_SPLIT(dev_priv
)) {
8828 ilk_setup_wm_latency(dev_priv
);
8830 if ((IS_GEN5(dev_priv
) && dev_priv
->wm
.pri_latency
[1] &&
8831 dev_priv
->wm
.spr_latency
[1] && dev_priv
->wm
.cur_latency
[1]) ||
8832 (!IS_GEN5(dev_priv
) && dev_priv
->wm
.pri_latency
[0] &&
8833 dev_priv
->wm
.spr_latency
[0] && dev_priv
->wm
.cur_latency
[0])) {
8834 dev_priv
->display
.compute_pipe_wm
= ilk_compute_pipe_wm
;
8835 dev_priv
->display
.compute_intermediate_wm
=
8836 ilk_compute_intermediate_wm
;
8837 dev_priv
->display
.initial_watermarks
=
8838 ilk_initial_watermarks
;
8839 dev_priv
->display
.optimize_watermarks
=
8840 ilk_optimize_watermarks
;
8842 DRM_DEBUG_KMS("Failed to read display plane latency. "
8845 } else if (IS_VALLEYVIEW(dev_priv
) || IS_CHERRYVIEW(dev_priv
)) {
8846 vlv_setup_wm_latency(dev_priv
);
8847 dev_priv
->display
.compute_pipe_wm
= vlv_compute_pipe_wm
;
8848 dev_priv
->display
.compute_intermediate_wm
= vlv_compute_intermediate_wm
;
8849 dev_priv
->display
.initial_watermarks
= vlv_initial_watermarks
;
8850 dev_priv
->display
.optimize_watermarks
= vlv_optimize_watermarks
;
8851 dev_priv
->display
.atomic_update_watermarks
= vlv_atomic_update_fifo
;
8852 } else if (IS_G4X(dev_priv
)) {
8853 g4x_setup_wm_latency(dev_priv
);
8854 dev_priv
->display
.compute_pipe_wm
= g4x_compute_pipe_wm
;
8855 dev_priv
->display
.compute_intermediate_wm
= g4x_compute_intermediate_wm
;
8856 dev_priv
->display
.initial_watermarks
= g4x_initial_watermarks
;
8857 dev_priv
->display
.optimize_watermarks
= g4x_optimize_watermarks
;
8858 } else if (IS_PINEVIEW(dev_priv
)) {
8859 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv
),
8862 dev_priv
->mem_freq
)) {
8863 DRM_INFO("failed to find known CxSR latency "
8864 "(found ddr%s fsb freq %d, mem freq %d), "
8866 (dev_priv
->is_ddr3
== 1) ? "3" : "2",
8867 dev_priv
->fsb_freq
, dev_priv
->mem_freq
);
8868 /* Disable CxSR and never update its watermark again */
8869 intel_set_memory_cxsr(dev_priv
, false);
8870 dev_priv
->display
.update_wm
= NULL
;
8872 dev_priv
->display
.update_wm
= pineview_update_wm
;
8873 } else if (IS_GEN4(dev_priv
)) {
8874 dev_priv
->display
.update_wm
= i965_update_wm
;
8875 } else if (IS_GEN3(dev_priv
)) {
8876 dev_priv
->display
.update_wm
= i9xx_update_wm
;
8877 dev_priv
->display
.get_fifo_size
= i9xx_get_fifo_size
;
8878 } else if (IS_GEN2(dev_priv
)) {
8879 if (INTEL_INFO(dev_priv
)->num_pipes
== 1) {
8880 dev_priv
->display
.update_wm
= i845_update_wm
;
8881 dev_priv
->display
.get_fifo_size
= i845_get_fifo_size
;
8883 dev_priv
->display
.update_wm
= i9xx_update_wm
;
8884 dev_priv
->display
.get_fifo_size
= i830_get_fifo_size
;
8887 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
8891 static inline int gen6_check_mailbox_status(struct drm_i915_private
*dev_priv
)
8894 I915_READ_FW(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_ERROR_MASK
;
8897 case GEN6_PCODE_SUCCESS
:
8899 case GEN6_PCODE_UNIMPLEMENTED_CMD
:
8901 case GEN6_PCODE_ILLEGAL_CMD
:
8903 case GEN6_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE
:
8904 case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE
:
8906 case GEN6_PCODE_TIMEOUT
:
8909 MISSING_CASE(flags
);
8914 static inline int gen7_check_mailbox_status(struct drm_i915_private
*dev_priv
)
8917 I915_READ_FW(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_ERROR_MASK
;
8920 case GEN6_PCODE_SUCCESS
:
8922 case GEN6_PCODE_ILLEGAL_CMD
:
8924 case GEN7_PCODE_TIMEOUT
:
8926 case GEN7_PCODE_ILLEGAL_DATA
:
8928 case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE
:
8931 MISSING_CASE(flags
);
8936 int sandybridge_pcode_read(struct drm_i915_private
*dev_priv
, u32 mbox
, u32
*val
)
8940 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
8942 /* GEN6_PCODE_* are outside of the forcewake domain, we can
8943 * use te fw I915_READ variants to reduce the amount of work
8944 * required when reading/writing.
8947 if (I915_READ_FW(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) {
8948 DRM_DEBUG_DRIVER("warning: pcode (read from mbox %x) mailbox access failed for %ps\n",
8949 mbox
, __builtin_return_address(0));
8953 I915_WRITE_FW(GEN6_PCODE_DATA
, *val
);
8954 I915_WRITE_FW(GEN6_PCODE_DATA1
, 0);
8955 I915_WRITE_FW(GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
| mbox
);
8957 if (__intel_wait_for_register_fw(dev_priv
,
8958 GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
, 0,
8960 DRM_ERROR("timeout waiting for pcode read (from mbox %x) to finish for %ps\n",
8961 mbox
, __builtin_return_address(0));
8965 *val
= I915_READ_FW(GEN6_PCODE_DATA
);
8966 I915_WRITE_FW(GEN6_PCODE_DATA
, 0);
8968 if (INTEL_GEN(dev_priv
) > 6)
8969 status
= gen7_check_mailbox_status(dev_priv
);
8971 status
= gen6_check_mailbox_status(dev_priv
);
8974 DRM_DEBUG_DRIVER("warning: pcode (read from mbox %x) mailbox access failed for %ps: %d\n",
8975 mbox
, __builtin_return_address(0), status
);
8982 int sandybridge_pcode_write(struct drm_i915_private
*dev_priv
,
8987 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
8989 /* GEN6_PCODE_* are outside of the forcewake domain, we can
8990 * use te fw I915_READ variants to reduce the amount of work
8991 * required when reading/writing.
8994 if (I915_READ_FW(GEN6_PCODE_MAILBOX
) & GEN6_PCODE_READY
) {
8995 DRM_DEBUG_DRIVER("warning: pcode (write of 0x%08x to mbox %x) mailbox access failed for %ps\n",
8996 val
, mbox
, __builtin_return_address(0));
9000 I915_WRITE_FW(GEN6_PCODE_DATA
, val
);
9001 I915_WRITE_FW(GEN6_PCODE_DATA1
, 0);
9002 I915_WRITE_FW(GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
| mbox
);
9004 if (__intel_wait_for_register_fw(dev_priv
,
9005 GEN6_PCODE_MAILBOX
, GEN6_PCODE_READY
, 0,
9007 DRM_ERROR("timeout waiting for pcode write of 0x%08x to mbox %x to finish for %ps\n",
9008 val
, mbox
, __builtin_return_address(0));
9012 I915_WRITE_FW(GEN6_PCODE_DATA
, 0);
9014 if (INTEL_GEN(dev_priv
) > 6)
9015 status
= gen7_check_mailbox_status(dev_priv
);
9017 status
= gen6_check_mailbox_status(dev_priv
);
9020 DRM_DEBUG_DRIVER("warning: pcode (write of 0x%08x to mbox %x) mailbox access failed for %ps: %d\n",
9021 val
, mbox
, __builtin_return_address(0), status
);
9028 static bool skl_pcode_try_request(struct drm_i915_private
*dev_priv
, u32 mbox
,
9029 u32 request
, u32 reply_mask
, u32 reply
,
9034 *status
= sandybridge_pcode_read(dev_priv
, mbox
, &val
);
9036 return *status
|| ((val
& reply_mask
) == reply
);
9040 * skl_pcode_request - send PCODE request until acknowledgment
9041 * @dev_priv: device private
9042 * @mbox: PCODE mailbox ID the request is targeted for
9043 * @request: request ID
9044 * @reply_mask: mask used to check for request acknowledgment
9045 * @reply: value used to check for request acknowledgment
9046 * @timeout_base_ms: timeout for polling with preemption enabled
9048 * Keep resending the @request to @mbox until PCODE acknowledges it, PCODE
9049 * reports an error or an overall timeout of @timeout_base_ms+50 ms expires.
9050 * The request is acknowledged once the PCODE reply dword equals @reply after
9051 * applying @reply_mask. Polling is first attempted with preemption enabled
9052 * for @timeout_base_ms and if this times out for another 50 ms with
9053 * preemption disabled.
9055 * Returns 0 on success, %-ETIMEDOUT in case of a timeout, <0 in case of some
9056 * other error as reported by PCODE.
9058 int skl_pcode_request(struct drm_i915_private
*dev_priv
, u32 mbox
, u32 request
,
9059 u32 reply_mask
, u32 reply
, int timeout_base_ms
)
9064 WARN_ON(!mutex_is_locked(&dev_priv
->rps
.hw_lock
));
9066 #define COND skl_pcode_try_request(dev_priv, mbox, request, reply_mask, reply, \
9070 * Prime the PCODE by doing a request first. Normally it guarantees
9071 * that a subsequent request, at most @timeout_base_ms later, succeeds.
9072 * _wait_for() doesn't guarantee when its passed condition is evaluated
9073 * first, so send the first request explicitly.
9079 ret
= _wait_for(COND
, timeout_base_ms
* 1000, 10);
9084 * The above can time out if the number of requests was low (2 in the
9085 * worst case) _and_ PCODE was busy for some reason even after a
9086 * (queued) request and @timeout_base_ms delay. As a workaround retry
9087 * the poll with preemption disabled to maximize the number of
9088 * requests. Increase the timeout from @timeout_base_ms to 50ms to
9089 * account for interrupts that could reduce the number of these
9090 * requests, and for any quirks of the PCODE firmware that delays
9091 * the request completion.
9093 DRM_DEBUG_KMS("PCODE timeout, retrying with preemption disabled\n");
9094 WARN_ON_ONCE(timeout_base_ms
> 3);
9096 ret
= wait_for_atomic(COND
, 50);
9100 return ret
? ret
: status
;
9104 static int byt_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
9108 * Slow = Fast = GPLL ref * N
9110 return DIV_ROUND_CLOSEST(dev_priv
->rps
.gpll_ref_freq
* (val
- 0xb7), 1000);
9113 static int byt_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
9115 return DIV_ROUND_CLOSEST(1000 * val
, dev_priv
->rps
.gpll_ref_freq
) + 0xb7;
9118 static int chv_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
9122 * CU (slow) = CU2x (fast) / 2 = GPLL ref * N / 2
9124 return DIV_ROUND_CLOSEST(dev_priv
->rps
.gpll_ref_freq
* val
, 2 * 2 * 1000);
9127 static int chv_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
9129 /* CHV needs even values */
9130 return DIV_ROUND_CLOSEST(2 * 1000 * val
, dev_priv
->rps
.gpll_ref_freq
) * 2;
9133 int intel_gpu_freq(struct drm_i915_private
*dev_priv
, int val
)
9135 if (INTEL_GEN(dev_priv
) >= 9)
9136 return DIV_ROUND_CLOSEST(val
* GT_FREQUENCY_MULTIPLIER
,
9138 else if (IS_CHERRYVIEW(dev_priv
))
9139 return chv_gpu_freq(dev_priv
, val
);
9140 else if (IS_VALLEYVIEW(dev_priv
))
9141 return byt_gpu_freq(dev_priv
, val
);
9143 return val
* GT_FREQUENCY_MULTIPLIER
;
9146 int intel_freq_opcode(struct drm_i915_private
*dev_priv
, int val
)
9148 if (INTEL_GEN(dev_priv
) >= 9)
9149 return DIV_ROUND_CLOSEST(val
* GEN9_FREQ_SCALER
,
9150 GT_FREQUENCY_MULTIPLIER
);
9151 else if (IS_CHERRYVIEW(dev_priv
))
9152 return chv_freq_opcode(dev_priv
, val
);
9153 else if (IS_VALLEYVIEW(dev_priv
))
9154 return byt_freq_opcode(dev_priv
, val
);
9156 return DIV_ROUND_CLOSEST(val
, GT_FREQUENCY_MULTIPLIER
);
9159 void intel_pm_setup(struct drm_i915_private
*dev_priv
)
9161 mutex_init(&dev_priv
->rps
.hw_lock
);
9163 INIT_DELAYED_WORK(&dev_priv
->rps
.autoenable_work
,
9164 __intel_autoenable_gt_powersave
);
9165 atomic_set(&dev_priv
->rps
.num_waiters
, 0);
9167 dev_priv
->pm
.suspended
= false;
9168 atomic_set(&dev_priv
->pm
.wakeref_count
, 0);
9171 static u64
vlv_residency_raw(struct drm_i915_private
*dev_priv
,
9172 const i915_reg_t reg
)
9174 u32 lower
, upper
, tmp
;
9177 /* The register accessed do not need forcewake. We borrow
9178 * uncore lock to prevent concurrent access to range reg.
9180 spin_lock_irq(&dev_priv
->uncore
.lock
);
9182 /* vlv and chv residency counters are 40 bits in width.
9183 * With a control bit, we can choose between upper or lower
9184 * 32bit window into this counter.
9186 * Although we always use the counter in high-range mode elsewhere,
9187 * userspace may attempt to read the value before rc6 is initialised,
9188 * before we have set the default VLV_COUNTER_CONTROL value. So always
9189 * set the high bit to be safe.
9191 I915_WRITE_FW(VLV_COUNTER_CONTROL
,
9192 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH
));
9193 upper
= I915_READ_FW(reg
);
9197 I915_WRITE_FW(VLV_COUNTER_CONTROL
,
9198 _MASKED_BIT_DISABLE(VLV_COUNT_RANGE_HIGH
));
9199 lower
= I915_READ_FW(reg
);
9201 I915_WRITE_FW(VLV_COUNTER_CONTROL
,
9202 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH
));
9203 upper
= I915_READ_FW(reg
);
9204 } while (upper
!= tmp
&& --loop
);
9206 /* Everywhere else we always use VLV_COUNTER_CONTROL with the
9207 * VLV_COUNT_RANGE_HIGH bit set - so it is safe to leave it set
9211 spin_unlock_irq(&dev_priv
->uncore
.lock
);
9213 return lower
| (u64
)upper
<< 8;
9216 u64
intel_rc6_residency_us(struct drm_i915_private
*dev_priv
,
9217 const i915_reg_t reg
)
9219 u64 time_hw
, units
, div
;
9221 if (!intel_enable_rc6())
9224 intel_runtime_pm_get(dev_priv
);
9226 /* On VLV and CHV, residency time is in CZ units rather than 1.28us */
9227 if (IS_VALLEYVIEW(dev_priv
) || IS_CHERRYVIEW(dev_priv
)) {
9229 div
= dev_priv
->czclk_freq
;
9231 time_hw
= vlv_residency_raw(dev_priv
, reg
);
9232 } else if (IS_GEN9_LP(dev_priv
)) {
9234 div
= 1200; /* 833.33ns */
9236 time_hw
= I915_READ(reg
);
9238 units
= 128000; /* 1.28us */
9241 time_hw
= I915_READ(reg
);
9244 intel_runtime_pm_put(dev_priv
);
9245 return DIV_ROUND_UP_ULL(time_hw
* units
, div
);