1 /*
2 * Copyright © 2012 Intel Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Eugeni Dodonov <eugeni.dodonov@intel.com>
25 *
26 */
27
28 #include <linux/cpufreq.h>
29 #include <drm/drm_plane_helper.h>
30 #include "i915_drv.h"
31 #include "intel_drv.h"
32 #include <linux/module.h>
33 #include <drm/drm_atomic_helper.h>
34
35 /**
36 * DOC: RC6
37 *
38 * RC6 is a special power stage which allows the GPU to enter an very
39 * low-voltage mode when idle, using down to 0V while at this stage. This
40 * stage is entered automatically when the GPU is idle when RC6 support is
41 * enabled, and as soon as new workload arises GPU wakes up automatically as well.
42 *
43 * There are different RC6 modes available in Intel GPU, which differentiate
44 * among each other with the latency required to enter and leave RC6 and
45 * voltage consumed by the GPU in different states.
46 *
47 * The combination of the following flags define which states GPU is allowed
48 * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
49 * RC6pp is deepest RC6. Their support by hardware varies according to the
50 * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
51 * which brings the most power savings; deeper states save more power, but
52 * require higher latency to switch to and wake up.
53 */
54 #define INTEL_RC6_ENABLE (1<<0)
55 #define INTEL_RC6p_ENABLE (1<<1)
56 #define INTEL_RC6pp_ENABLE (1<<2)
57
gen9_init_clock_gating(struct drm_i915_private * dev_priv)58 static void gen9_init_clock_gating(struct drm_i915_private *dev_priv)
59 {
60 if (HAS_LLC(dev_priv)) {
61 /*
62 * WaCompressedResourceDisplayNewHashMode:skl,kbl
63 * Display WA#0390: skl,kbl
64 *
65 * Must match Sampler, Pixel Back End, and Media. See
66 * WaCompressedResourceSamplerPbeMediaNewHashMode.
67 */
68 I915_WRITE(CHICKEN_PAR1_1,
69 I915_READ(CHICKEN_PAR1_1) |
70 SKL_DE_COMPRESSED_HASH_MODE);
71 }
72
73 /* See Bspec note for PSR2_CTL bit 31, Wa#828:skl,bxt,kbl,cfl */
74 I915_WRITE(CHICKEN_PAR1_1,
75 I915_READ(CHICKEN_PAR1_1) | SKL_EDP_PSR_FIX_RDWRAP);
76
77 I915_WRITE(GEN8_CONFIG0,
78 I915_READ(GEN8_CONFIG0) | GEN9_DEFAULT_FIXES);
79
80 /* WaEnableChickenDCPR:skl,bxt,kbl,glk,cfl */
81 I915_WRITE(GEN8_CHICKEN_DCPR_1,
82 I915_READ(GEN8_CHICKEN_DCPR_1) | MASK_WAKEMEM);
83
84 /* WaFbcTurnOffFbcWatermark:skl,bxt,kbl,cfl */
85 /* WaFbcWakeMemOn:skl,bxt,kbl,glk,cfl */
86 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
87 DISP_FBC_WM_DIS |
88 DISP_FBC_MEMORY_WAKE);
89
90 /* WaFbcHighMemBwCorruptionAvoidance:skl,bxt,kbl,cfl */
91 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
92 ILK_DPFC_DISABLE_DUMMY0);
93
94 if (IS_SKYLAKE(dev_priv)) {
95 /* WaDisableDopClockGating */
96 I915_WRITE(GEN7_MISCCPCTL, I915_READ(GEN7_MISCCPCTL)
97 & ~GEN7_DOP_CLOCK_GATE_ENABLE);
98 }
99 }
100
bxt_init_clock_gating(struct drm_i915_private * dev_priv)101 static void bxt_init_clock_gating(struct drm_i915_private *dev_priv)
102 {
103 gen9_init_clock_gating(dev_priv);
104
105 /* WaDisableSDEUnitClockGating:bxt */
106 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
107 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
108
109 /*
110 * FIXME:
111 * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
112 */
113 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
114 GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ);
115
116 /*
117 * Wa: Backlight PWM may stop in the asserted state, causing backlight
118 * to stay fully on.
119 */
120 I915_WRITE(GEN9_CLKGATE_DIS_0, I915_READ(GEN9_CLKGATE_DIS_0) |
121 PWM1_GATING_DIS | PWM2_GATING_DIS);
122 }
123
glk_init_clock_gating(struct drm_i915_private * dev_priv)124 static void glk_init_clock_gating(struct drm_i915_private *dev_priv)
125 {
126 gen9_init_clock_gating(dev_priv);
127
128 /*
129 * WaDisablePWMClockGating:glk
130 * Backlight PWM may stop in the asserted state, causing backlight
131 * to stay fully on.
132 */
133 I915_WRITE(GEN9_CLKGATE_DIS_0, I915_READ(GEN9_CLKGATE_DIS_0) |
134 PWM1_GATING_DIS | PWM2_GATING_DIS);
135
136 /* WaDDIIOTimeout:glk */
137 if (IS_GLK_REVID(dev_priv, 0, GLK_REVID_A1)) {
138 u32 val = I915_READ(CHICKEN_MISC_2);
139 val &= ~(GLK_CL0_PWR_DOWN |
140 GLK_CL1_PWR_DOWN |
141 GLK_CL2_PWR_DOWN);
142 I915_WRITE(CHICKEN_MISC_2, val);
143 }
144
145 }
146
i915_pineview_get_mem_freq(struct drm_i915_private * dev_priv)147 static void i915_pineview_get_mem_freq(struct drm_i915_private *dev_priv)
148 {
149 u32 tmp;
150
151 tmp = I915_READ(CLKCFG);
152
153 switch (tmp & CLKCFG_FSB_MASK) {
154 case CLKCFG_FSB_533:
155 dev_priv->fsb_freq = 533; /* 133*4 */
156 break;
157 case CLKCFG_FSB_800:
158 dev_priv->fsb_freq = 800; /* 200*4 */
159 break;
160 case CLKCFG_FSB_667:
161 dev_priv->fsb_freq = 667; /* 167*4 */
162 break;
163 case CLKCFG_FSB_400:
164 dev_priv->fsb_freq = 400; /* 100*4 */
165 break;
166 }
167
168 switch (tmp & CLKCFG_MEM_MASK) {
169 case CLKCFG_MEM_533:
170 dev_priv->mem_freq = 533;
171 break;
172 case CLKCFG_MEM_667:
173 dev_priv->mem_freq = 667;
174 break;
175 case CLKCFG_MEM_800:
176 dev_priv->mem_freq = 800;
177 break;
178 }
179
180 /* detect pineview DDR3 setting */
181 tmp = I915_READ(CSHRDDR3CTL);
182 dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
183 }
184
i915_ironlake_get_mem_freq(struct drm_i915_private * dev_priv)185 static void i915_ironlake_get_mem_freq(struct drm_i915_private *dev_priv)
186 {
187 u16 ddrpll, csipll;
188
189 ddrpll = I915_READ16(DDRMPLL1);
190 csipll = I915_READ16(CSIPLL0);
191
192 switch (ddrpll & 0xff) {
193 case 0xc:
194 dev_priv->mem_freq = 800;
195 break;
196 case 0x10:
197 dev_priv->mem_freq = 1066;
198 break;
199 case 0x14:
200 dev_priv->mem_freq = 1333;
201 break;
202 case 0x18:
203 dev_priv->mem_freq = 1600;
204 break;
205 default:
206 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
207 ddrpll & 0xff);
208 dev_priv->mem_freq = 0;
209 break;
210 }
211
212 dev_priv->ips.r_t = dev_priv->mem_freq;
213
214 switch (csipll & 0x3ff) {
215 case 0x00c:
216 dev_priv->fsb_freq = 3200;
217 break;
218 case 0x00e:
219 dev_priv->fsb_freq = 3733;
220 break;
221 case 0x010:
222 dev_priv->fsb_freq = 4266;
223 break;
224 case 0x012:
225 dev_priv->fsb_freq = 4800;
226 break;
227 case 0x014:
228 dev_priv->fsb_freq = 5333;
229 break;
230 case 0x016:
231 dev_priv->fsb_freq = 5866;
232 break;
233 case 0x018:
234 dev_priv->fsb_freq = 6400;
235 break;
236 default:
237 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
238 csipll & 0x3ff);
239 dev_priv->fsb_freq = 0;
240 break;
241 }
242
243 if (dev_priv->fsb_freq == 3200) {
244 dev_priv->ips.c_m = 0;
245 } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
246 dev_priv->ips.c_m = 1;
247 } else {
248 dev_priv->ips.c_m = 2;
249 }
250 }
251
252 static const struct cxsr_latency cxsr_latency_table[] = {
253 {1, 0, 800, 400, 3382, 33382, 3983, 33983}, /* DDR2-400 SC */
254 {1, 0, 800, 667, 3354, 33354, 3807, 33807}, /* DDR2-667 SC */
255 {1, 0, 800, 800, 3347, 33347, 3763, 33763}, /* DDR2-800 SC */
256 {1, 1, 800, 667, 6420, 36420, 6873, 36873}, /* DDR3-667 SC */
257 {1, 1, 800, 800, 5902, 35902, 6318, 36318}, /* DDR3-800 SC */
258
259 {1, 0, 667, 400, 3400, 33400, 4021, 34021}, /* DDR2-400 SC */
260 {1, 0, 667, 667, 3372, 33372, 3845, 33845}, /* DDR2-667 SC */
261 {1, 0, 667, 800, 3386, 33386, 3822, 33822}, /* DDR2-800 SC */
262 {1, 1, 667, 667, 6438, 36438, 6911, 36911}, /* DDR3-667 SC */
263 {1, 1, 667, 800, 5941, 35941, 6377, 36377}, /* DDR3-800 SC */
264
265 {1, 0, 400, 400, 3472, 33472, 4173, 34173}, /* DDR2-400 SC */
266 {1, 0, 400, 667, 3443, 33443, 3996, 33996}, /* DDR2-667 SC */
267 {1, 0, 400, 800, 3430, 33430, 3946, 33946}, /* DDR2-800 SC */
268 {1, 1, 400, 667, 6509, 36509, 7062, 37062}, /* DDR3-667 SC */
269 {1, 1, 400, 800, 5985, 35985, 6501, 36501}, /* DDR3-800 SC */
270
271 {0, 0, 800, 400, 3438, 33438, 4065, 34065}, /* DDR2-400 SC */
272 {0, 0, 800, 667, 3410, 33410, 3889, 33889}, /* DDR2-667 SC */
273 {0, 0, 800, 800, 3403, 33403, 3845, 33845}, /* DDR2-800 SC */
274 {0, 1, 800, 667, 6476, 36476, 6955, 36955}, /* DDR3-667 SC */
275 {0, 1, 800, 800, 5958, 35958, 6400, 36400}, /* DDR3-800 SC */
276
277 {0, 0, 667, 400, 3456, 33456, 4103, 34106}, /* DDR2-400 SC */
278 {0, 0, 667, 667, 3428, 33428, 3927, 33927}, /* DDR2-667 SC */
279 {0, 0, 667, 800, 3443, 33443, 3905, 33905}, /* DDR2-800 SC */
280 {0, 1, 667, 667, 6494, 36494, 6993, 36993}, /* DDR3-667 SC */
281 {0, 1, 667, 800, 5998, 35998, 6460, 36460}, /* DDR3-800 SC */
282
283 {0, 0, 400, 400, 3528, 33528, 4255, 34255}, /* DDR2-400 SC */
284 {0, 0, 400, 667, 3500, 33500, 4079, 34079}, /* DDR2-667 SC */
285 {0, 0, 400, 800, 3487, 33487, 4029, 34029}, /* DDR2-800 SC */
286 {0, 1, 400, 667, 6566, 36566, 7145, 37145}, /* DDR3-667 SC */
287 {0, 1, 400, 800, 6042, 36042, 6584, 36584}, /* DDR3-800 SC */
288 };
289
intel_get_cxsr_latency(bool is_desktop,bool is_ddr3,int fsb,int mem)290 static const struct cxsr_latency *intel_get_cxsr_latency(bool is_desktop,
291 bool is_ddr3,
292 int fsb,
293 int mem)
294 {
295 const struct cxsr_latency *latency;
296 int i;
297
298 if (fsb == 0 || mem == 0)
299 return NULL;
300
301 for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
302 latency = &cxsr_latency_table[i];
303 if (is_desktop == latency->is_desktop &&
304 is_ddr3 == latency->is_ddr3 &&
305 fsb == latency->fsb_freq && mem == latency->mem_freq)
306 return latency;
307 }
308
309 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
310
311 return NULL;
312 }
313
chv_set_memory_dvfs(struct drm_i915_private * dev_priv,bool enable)314 static void chv_set_memory_dvfs(struct drm_i915_private *dev_priv, bool enable)
315 {
316 u32 val;
317
318 mutex_lock(&dev_priv->pcu_lock);
319
320 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
321 if (enable)
322 val &= ~FORCE_DDR_HIGH_FREQ;
323 else
324 val |= FORCE_DDR_HIGH_FREQ;
325 val &= ~FORCE_DDR_LOW_FREQ;
326 val |= FORCE_DDR_FREQ_REQ_ACK;
327 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
328
329 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
330 FORCE_DDR_FREQ_REQ_ACK) == 0, 3))
331 DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
332
333 mutex_unlock(&dev_priv->pcu_lock);
334 }
335
chv_set_memory_pm5(struct drm_i915_private * dev_priv,bool enable)336 static void chv_set_memory_pm5(struct drm_i915_private *dev_priv, bool enable)
337 {
338 u32 val;
339
340 mutex_lock(&dev_priv->pcu_lock);
341
342 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
343 if (enable)
344 val |= DSP_MAXFIFO_PM5_ENABLE;
345 else
346 val &= ~DSP_MAXFIFO_PM5_ENABLE;
347 vlv_punit_write(dev_priv, PUNIT_REG_DSPFREQ, val);
348
349 mutex_unlock(&dev_priv->pcu_lock);
350 }
351
352 #define FW_WM(value, plane) \
353 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
354
_intel_set_memory_cxsr(struct drm_i915_private * dev_priv,bool enable)355 static bool _intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
356 {
357 bool was_enabled;
358 u32 val;
359
360 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
361 was_enabled = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
362 I915_WRITE(FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
363 POSTING_READ(FW_BLC_SELF_VLV);
364 } else if (IS_G4X(dev_priv) || IS_I965GM(dev_priv)) {
365 was_enabled = I915_READ(FW_BLC_SELF) & FW_BLC_SELF_EN;
366 I915_WRITE(FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
367 POSTING_READ(FW_BLC_SELF);
368 } else if (IS_PINEVIEW(dev_priv)) {
369 val = I915_READ(DSPFW3);
370 was_enabled = val & PINEVIEW_SELF_REFRESH_EN;
371 if (enable)
372 val |= PINEVIEW_SELF_REFRESH_EN;
373 else
374 val &= ~PINEVIEW_SELF_REFRESH_EN;
375 I915_WRITE(DSPFW3, val);
376 POSTING_READ(DSPFW3);
377 } else if (IS_I945G(dev_priv) || IS_I945GM(dev_priv)) {
378 was_enabled = I915_READ(FW_BLC_SELF) & FW_BLC_SELF_EN;
379 val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
380 _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
381 I915_WRITE(FW_BLC_SELF, val);
382 POSTING_READ(FW_BLC_SELF);
383 } else if (IS_I915GM(dev_priv)) {
384 /*
385 * FIXME can't find a bit like this for 915G, and
386 * and yet it does have the related watermark in
387 * FW_BLC_SELF. What's going on?
388 */
389 was_enabled = I915_READ(INSTPM) & INSTPM_SELF_EN;
390 val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
391 _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
392 I915_WRITE(INSTPM, val);
393 POSTING_READ(INSTPM);
394 } else {
395 return false;
396 }
397
398 trace_intel_memory_cxsr(dev_priv, was_enabled, enable);
399
400 DRM_DEBUG_KMS("memory self-refresh is %s (was %s)\n",
401 enableddisabled(enable),
402 enableddisabled(was_enabled));
403
404 return was_enabled;
405 }
406
407 /**
408 * intel_set_memory_cxsr - Configure CxSR state
409 * @dev_priv: i915 device
410 * @enable: Allow vs. disallow CxSR
411 *
412 * Allow or disallow the system to enter a special CxSR
413 * (C-state self refresh) state. What typically happens in CxSR mode
414 * is that several display FIFOs may get combined into a single larger
415 * FIFO for a particular plane (so called max FIFO mode) to allow the
416 * system to defer memory fetches longer, and the memory will enter
417 * self refresh.
418 *
419 * Note that enabling CxSR does not guarantee that the system enter
420 * this special mode, nor does it guarantee that the system stays
421 * in that mode once entered. So this just allows/disallows the system
422 * to autonomously utilize the CxSR mode. Other factors such as core
423 * C-states will affect when/if the system actually enters/exits the
424 * CxSR mode.
425 *
426 * Note that on VLV/CHV this actually only controls the max FIFO mode,
427 * and the system is free to enter/exit memory self refresh at any time
428 * even when the use of CxSR has been disallowed.
429 *
430 * While the system is actually in the CxSR/max FIFO mode, some plane
431 * control registers will not get latched on vblank. Thus in order to
432 * guarantee the system will respond to changes in the plane registers
433 * we must always disallow CxSR prior to making changes to those registers.
434 * Unfortunately the system will re-evaluate the CxSR conditions at
435 * frame start which happens after vblank start (which is when the plane
436 * registers would get latched), so we can't proceed with the plane update
437 * during the same frame where we disallowed CxSR.
438 *
439 * Certain platforms also have a deeper HPLL SR mode. Fortunately the
440 * HPLL SR mode depends on CxSR itself, so we don't have to hand hold
441 * the hardware w.r.t. HPLL SR when writing to plane registers.
442 * Disallowing just CxSR is sufficient.
443 */
intel_set_memory_cxsr(struct drm_i915_private * dev_priv,bool enable)444 bool intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
445 {
446 bool ret;
447
448 mutex_lock(&dev_priv->wm.wm_mutex);
449 ret = _intel_set_memory_cxsr(dev_priv, enable);
450 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
451 dev_priv->wm.vlv.cxsr = enable;
452 else if (IS_G4X(dev_priv))
453 dev_priv->wm.g4x.cxsr = enable;
454 mutex_unlock(&dev_priv->wm.wm_mutex);
455
456 return ret;
457 }
458
459 /*
460 * Latency for FIFO fetches is dependent on several factors:
461 * - memory configuration (speed, channels)
462 * - chipset
463 * - current MCH state
464 * It can be fairly high in some situations, so here we assume a fairly
465 * pessimal value. It's a tradeoff between extra memory fetches (if we
466 * set this value too high, the FIFO will fetch frequently to stay full)
467 * and power consumption (set it too low to save power and we might see
468 * FIFO underruns and display "flicker").
469 *
470 * A value of 5us seems to be a good balance; safe for very low end
471 * platforms but not overly aggressive on lower latency configs.
472 */
473 static const int pessimal_latency_ns = 5000;
474
475 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
476 ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
477
vlv_get_fifo_size(struct intel_crtc_state * crtc_state)478 static void vlv_get_fifo_size(struct intel_crtc_state *crtc_state)
479 {
480 struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
481 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
482 struct vlv_fifo_state *fifo_state = &crtc_state->wm.vlv.fifo_state;
483 enum i915_pipe pipe = crtc->pipe;
484 int sprite0_start, sprite1_start;
485
486 switch (pipe) {
487 uint32_t dsparb, dsparb2, dsparb3;
488 case PIPE_A:
489 dsparb = I915_READ(DSPARB);
490 dsparb2 = I915_READ(DSPARB2);
491 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 0, 0);
492 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 8, 4);
493 break;
494 case PIPE_B:
495 dsparb = I915_READ(DSPARB);
496 dsparb2 = I915_READ(DSPARB2);
497 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 16, 8);
498 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 24, 12);
499 break;
500 case PIPE_C:
501 dsparb2 = I915_READ(DSPARB2);
502 dsparb3 = I915_READ(DSPARB3);
503 sprite0_start = VLV_FIFO_START(dsparb3, dsparb2, 0, 16);
504 sprite1_start = VLV_FIFO_START(dsparb3, dsparb2, 8, 20);
505 break;
506 default:
507 MISSING_CASE(pipe);
508 return;
509 }
510
511 fifo_state->plane[PLANE_PRIMARY] = sprite0_start;
512 fifo_state->plane[PLANE_SPRITE0] = sprite1_start - sprite0_start;
513 fifo_state->plane[PLANE_SPRITE1] = 511 - sprite1_start;
514 fifo_state->plane[PLANE_CURSOR] = 63;
515 }
516
i9xx_get_fifo_size(struct drm_i915_private * dev_priv,int plane)517 static int i9xx_get_fifo_size(struct drm_i915_private *dev_priv, int plane)
518 {
519 uint32_t dsparb = I915_READ(DSPARB);
520 int size;
521
522 size = dsparb & 0x7f;
523 if (plane)
524 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
525
526 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
527 plane ? "B" : "A", size);
528
529 return size;
530 }
531
i830_get_fifo_size(struct drm_i915_private * dev_priv,int plane)532 static int i830_get_fifo_size(struct drm_i915_private *dev_priv, int plane)
533 {
534 uint32_t dsparb = I915_READ(DSPARB);
535 int size;
536
537 size = dsparb & 0x1ff;
538 if (plane)
539 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
540 size >>= 1; /* Convert to cachelines */
541
542 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
543 plane ? "B" : "A", size);
544
545 return size;
546 }
547
i845_get_fifo_size(struct drm_i915_private * dev_priv,int plane)548 static int i845_get_fifo_size(struct drm_i915_private *dev_priv, int plane)
549 {
550 uint32_t dsparb = I915_READ(DSPARB);
551 int size;
552
553 size = dsparb & 0x7f;
554 size >>= 2; /* Convert to cachelines */
555
556 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
557 plane ? "B" : "A",
558 size);
559
560 return size;
561 }
562
563 /* Pineview has different values for various configs */
564 static const struct intel_watermark_params pineview_display_wm = {
565 .fifo_size = PINEVIEW_DISPLAY_FIFO,
566 .max_wm = PINEVIEW_MAX_WM,
567 .default_wm = PINEVIEW_DFT_WM,
568 .guard_size = PINEVIEW_GUARD_WM,
569 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
570 };
571 static const struct intel_watermark_params pineview_display_hplloff_wm = {
572 .fifo_size = PINEVIEW_DISPLAY_FIFO,
573 .max_wm = PINEVIEW_MAX_WM,
574 .default_wm = PINEVIEW_DFT_HPLLOFF_WM,
575 .guard_size = PINEVIEW_GUARD_WM,
576 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
577 };
578 static const struct intel_watermark_params pineview_cursor_wm = {
579 .fifo_size = PINEVIEW_CURSOR_FIFO,
580 .max_wm = PINEVIEW_CURSOR_MAX_WM,
581 .default_wm = PINEVIEW_CURSOR_DFT_WM,
582 .guard_size = PINEVIEW_CURSOR_GUARD_WM,
583 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
584 };
585 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
586 .fifo_size = PINEVIEW_CURSOR_FIFO,
587 .max_wm = PINEVIEW_CURSOR_MAX_WM,
588 .default_wm = PINEVIEW_CURSOR_DFT_WM,
589 .guard_size = PINEVIEW_CURSOR_GUARD_WM,
590 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
591 };
592 static const struct intel_watermark_params i965_cursor_wm_info = {
593 .fifo_size = I965_CURSOR_FIFO,
594 .max_wm = I965_CURSOR_MAX_WM,
595 .default_wm = I965_CURSOR_DFT_WM,
596 .guard_size = 2,
597 .cacheline_size = I915_FIFO_LINE_SIZE,
598 };
599 static const struct intel_watermark_params i945_wm_info = {
600 .fifo_size = I945_FIFO_SIZE,
601 .max_wm = I915_MAX_WM,
602 .default_wm = 1,
603 .guard_size = 2,
604 .cacheline_size = I915_FIFO_LINE_SIZE,
605 };
606 static const struct intel_watermark_params i915_wm_info = {
607 .fifo_size = I915_FIFO_SIZE,
608 .max_wm = I915_MAX_WM,
609 .default_wm = 1,
610 .guard_size = 2,
611 .cacheline_size = I915_FIFO_LINE_SIZE,
612 };
613 static const struct intel_watermark_params i830_a_wm_info = {
614 .fifo_size = I855GM_FIFO_SIZE,
615 .max_wm = I915_MAX_WM,
616 .default_wm = 1,
617 .guard_size = 2,
618 .cacheline_size = I830_FIFO_LINE_SIZE,
619 };
620 static const struct intel_watermark_params i830_bc_wm_info = {
621 .fifo_size = I855GM_FIFO_SIZE,
622 .max_wm = I915_MAX_WM/2,
623 .default_wm = 1,
624 .guard_size = 2,
625 .cacheline_size = I830_FIFO_LINE_SIZE,
626 };
627 static const struct intel_watermark_params i845_wm_info = {
628 .fifo_size = I830_FIFO_SIZE,
629 .max_wm = I915_MAX_WM,
630 .default_wm = 1,
631 .guard_size = 2,
632 .cacheline_size = I830_FIFO_LINE_SIZE,
633 };
634
635 /**
636 * intel_wm_method1 - Method 1 / "small buffer" watermark formula
637 * @pixel_rate: Pipe pixel rate in kHz
638 * @cpp: Plane bytes per pixel
639 * @latency: Memory wakeup latency in 0.1us units
640 *
641 * Compute the watermark using the method 1 or "small buffer"
642 * formula. The caller may additonally add extra cachelines
643 * to account for TLB misses and clock crossings.
644 *
645 * This method is concerned with the short term drain rate
646 * of the FIFO, ie. it does not account for blanking periods
647 * which would effectively reduce the average drain rate across
648 * a longer period. The name "small" refers to the fact the
649 * FIFO is relatively small compared to the amount of data
650 * fetched.
651 *
652 * The FIFO level vs. time graph might look something like:
653 *
654 * |\ |\
655 * | \ | \
656 * __---__---__ (- plane active, _ blanking)
657 * -> time
658 *
659 * or perhaps like this:
660 *
661 * |\|\ |\|\
662 * __----__----__ (- plane active, _ blanking)
663 * -> time
664 *
665 * Returns:
666 * The watermark in bytes
667 */
intel_wm_method1(unsigned int pixel_rate,unsigned int cpp,unsigned int latency)668 static unsigned int intel_wm_method1(unsigned int pixel_rate,
669 unsigned int cpp,
670 unsigned int latency)
671 {
672 uint64_t ret;
673
674 ret = (uint64_t) pixel_rate * cpp * latency;
675 ret = DIV_ROUND_UP_ULL(ret, 10000);
676
677 return ret;
678 }
679
680 /**
681 * intel_wm_method2 - Method 2 / "large buffer" watermark formula
682 * @pixel_rate: Pipe pixel rate in kHz
683 * @htotal: Pipe horizontal total
684 * @width: Plane width in pixels
685 * @cpp: Plane bytes per pixel
686 * @latency: Memory wakeup latency in 0.1us units
687 *
688 * Compute the watermark using the method 2 or "large buffer"
689 * formula. The caller may additonally add extra cachelines
690 * to account for TLB misses and clock crossings.
691 *
692 * This method is concerned with the long term drain rate
693 * of the FIFO, ie. it does account for blanking periods
694 * which effectively reduce the average drain rate across
695 * a longer period. The name "large" refers to the fact the
696 * FIFO is relatively large compared to the amount of data
697 * fetched.
698 *
699 * The FIFO level vs. time graph might look something like:
700 *
701 * |\___ |\___
702 * | \___ | \___
703 * | \ | \
704 * __ --__--__--__--__--__--__ (- plane active, _ blanking)
705 * -> time
706 *
707 * Returns:
708 * The watermark in bytes
709 */
intel_wm_method2(unsigned int pixel_rate,unsigned int htotal,unsigned int width,unsigned int cpp,unsigned int latency)710 static unsigned int intel_wm_method2(unsigned int pixel_rate,
711 unsigned int htotal,
712 unsigned int width,
713 unsigned int cpp,
714 unsigned int latency)
715 {
716 unsigned int ret;
717
718 /*
719 * FIXME remove once all users are computing
720 * watermarks in the correct place.
721 */
722 if (WARN_ON_ONCE(htotal == 0))
723 htotal = 1;
724
725 ret = (latency * pixel_rate) / (htotal * 10000);
726 ret = (ret + 1) * width * cpp;
727
728 return ret;
729 }
730
731 /**
732 * intel_calculate_wm - calculate watermark level
733 * @pixel_rate: pixel clock
734 * @wm: chip FIFO params
735 * @cpp: bytes per pixel
736 * @latency_ns: memory latency for the platform
737 *
738 * Calculate the watermark level (the level at which the display plane will
739 * start fetching from memory again). Each chip has a different display
740 * FIFO size and allocation, so the caller needs to figure that out and pass
741 * in the correct intel_watermark_params structure.
742 *
743 * As the pixel clock runs, the FIFO will be drained at a rate that depends
744 * on the pixel size. When it reaches the watermark level, it'll start
745 * fetching FIFO line sized based chunks from memory until the FIFO fills
746 * past the watermark point. If the FIFO drains completely, a FIFO underrun
747 * will occur, and a display engine hang could result.
748 */
intel_calculate_wm(int pixel_rate,const struct intel_watermark_params * wm,int fifo_size,int cpp,unsigned int latency_ns)749 static unsigned int intel_calculate_wm(int pixel_rate,
750 const struct intel_watermark_params *wm,
751 int fifo_size, int cpp,
752 unsigned int latency_ns)
753 {
754 int entries, wm_size;
755
756 /*
757 * Note: we need to make sure we don't overflow for various clock &
758 * latency values.
759 * clocks go from a few thousand to several hundred thousand.
760 * latency is usually a few thousand
761 */
762 entries = intel_wm_method1(pixel_rate, cpp,
763 latency_ns / 100);
764 entries = DIV_ROUND_UP(entries, wm->cacheline_size) +
765 wm->guard_size;
766 DRM_DEBUG_KMS("FIFO entries required for mode: %d\n", entries);
767
768 wm_size = fifo_size - entries;
769 DRM_DEBUG_KMS("FIFO watermark level: %d\n", wm_size);
770
771 /* Don't promote wm_size to unsigned... */
772 if (wm_size > wm->max_wm)
773 wm_size = wm->max_wm;
774 if (wm_size <= 0)
775 wm_size = wm->default_wm;
776
777 /*
778 * Bspec seems to indicate that the value shouldn't be lower than
779 * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
780 * Lets go for 8 which is the burst size since certain platforms
781 * already use a hardcoded 8 (which is what the spec says should be
782 * done).
783 */
784 if (wm_size <= 8)
785 wm_size = 8;
786
787 return wm_size;
788 }
789
is_disabling(int old,int new,int threshold)790 static bool is_disabling(int old, int new, int threshold)
791 {
792 return old >= threshold && new < threshold;
793 }
794
is_enabling(int old,int new,int threshold)795 static bool is_enabling(int old, int new, int threshold)
796 {
797 return old < threshold && new >= threshold;
798 }
799
intel_wm_num_levels(struct drm_i915_private * dev_priv)800 static int intel_wm_num_levels(struct drm_i915_private *dev_priv)
801 {
802 return dev_priv->wm.max_level + 1;
803 }
804
intel_wm_plane_visible(const struct intel_crtc_state * crtc_state,const struct intel_plane_state * plane_state)805 static bool intel_wm_plane_visible(const struct intel_crtc_state *crtc_state,
806 const struct intel_plane_state *plane_state)
807 {
808 struct intel_plane *plane = to_intel_plane(plane_state->base.plane);
809
810 /* FIXME check the 'enable' instead */
811 if (!crtc_state->base.active)
812 return false;
813
814 /*
815 * Treat cursor with fb as always visible since cursor updates
816 * can happen faster than the vrefresh rate, and the current
817 * watermark code doesn't handle that correctly. Cursor updates
818 * which set/clear the fb or change the cursor size are going
819 * to get throttled by intel_legacy_cursor_update() to work
820 * around this problem with the watermark code.
821 */
822 if (plane->id == PLANE_CURSOR)
823 return plane_state->base.fb != NULL;
824 else
825 return plane_state->base.visible;
826 }
827
single_enabled_crtc(struct drm_i915_private * dev_priv)828 static struct intel_crtc *single_enabled_crtc(struct drm_i915_private *dev_priv)
829 {
830 struct intel_crtc *crtc, *enabled = NULL;
831
832 for_each_intel_crtc(&dev_priv->drm, crtc) {
833 if (intel_crtc_active(crtc)) {
834 if (enabled)
835 return NULL;
836 enabled = crtc;
837 }
838 }
839
840 return enabled;
841 }
842
pineview_update_wm(struct intel_crtc * unused_crtc)843 static void pineview_update_wm(struct intel_crtc *unused_crtc)
844 {
845 struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
846 struct intel_crtc *crtc;
847 const struct cxsr_latency *latency;
848 u32 reg;
849 unsigned int wm;
850
851 latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv),
852 dev_priv->is_ddr3,
853 dev_priv->fsb_freq,
854 dev_priv->mem_freq);
855 if (!latency) {
856 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
857 intel_set_memory_cxsr(dev_priv, false);
858 return;
859 }
860
861 crtc = single_enabled_crtc(dev_priv);
862 if (crtc) {
863 const struct drm_display_mode *adjusted_mode =
864 &crtc->config->base.adjusted_mode;
865 const struct drm_framebuffer *fb =
866 crtc->base.primary->state->fb;
867 int cpp = fb->format->cpp[0];
868 int clock = adjusted_mode->crtc_clock;
869
870 /* Display SR */
871 wm = intel_calculate_wm(clock, &pineview_display_wm,
872 pineview_display_wm.fifo_size,
873 cpp, latency->display_sr);
874 reg = I915_READ(DSPFW1);
875 reg &= ~DSPFW_SR_MASK;
876 reg |= FW_WM(wm, SR);
877 I915_WRITE(DSPFW1, reg);
878 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
879
880 /* cursor SR */
881 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
882 pineview_display_wm.fifo_size,
883 4, latency->cursor_sr);
884 reg = I915_READ(DSPFW3);
885 reg &= ~DSPFW_CURSOR_SR_MASK;
886 reg |= FW_WM(wm, CURSOR_SR);
887 I915_WRITE(DSPFW3, reg);
888
889 /* Display HPLL off SR */
890 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
891 pineview_display_hplloff_wm.fifo_size,
892 cpp, latency->display_hpll_disable);
893 reg = I915_READ(DSPFW3);
894 reg &= ~DSPFW_HPLL_SR_MASK;
895 reg |= FW_WM(wm, HPLL_SR);
896 I915_WRITE(DSPFW3, reg);
897
898 /* cursor HPLL off SR */
899 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
900 pineview_display_hplloff_wm.fifo_size,
901 4, latency->cursor_hpll_disable);
902 reg = I915_READ(DSPFW3);
903 reg &= ~DSPFW_HPLL_CURSOR_MASK;
904 reg |= FW_WM(wm, HPLL_CURSOR);
905 I915_WRITE(DSPFW3, reg);
906 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
907
908 intel_set_memory_cxsr(dev_priv, true);
909 } else {
910 intel_set_memory_cxsr(dev_priv, false);
911 }
912 }
913
914 /*
915 * Documentation says:
916 * "If the line size is small, the TLB fetches can get in the way of the
917 * data fetches, causing some lag in the pixel data return which is not
918 * accounted for in the above formulas. The following adjustment only
919 * needs to be applied if eight whole lines fit in the buffer at once.
920 * The WM is adjusted upwards by the difference between the FIFO size
921 * and the size of 8 whole lines. This adjustment is always performed
922 * in the actual pixel depth regardless of whether FBC is enabled or not."
923 */
g4x_tlb_miss_wa(int fifo_size,int width,int cpp)924 static int g4x_tlb_miss_wa(int fifo_size, int width, int cpp)
925 {
926 int tlb_miss = fifo_size * 64 - width * cpp * 8;
927
928 return max(0, tlb_miss);
929 }
930
g4x_write_wm_values(struct drm_i915_private * dev_priv,const struct g4x_wm_values * wm)931 static void g4x_write_wm_values(struct drm_i915_private *dev_priv,
932 const struct g4x_wm_values *wm)
933 {
934 enum i915_pipe pipe;
935
936 for_each_pipe(dev_priv, pipe)
937 trace_g4x_wm(intel_get_crtc_for_pipe(dev_priv, pipe), wm);
938
939 I915_WRITE(DSPFW1,
940 FW_WM(wm->sr.plane, SR) |
941 FW_WM(wm->pipe[PIPE_B].plane[PLANE_CURSOR], CURSORB) |
942 FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY], PLANEB) |
943 FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY], PLANEA));
944 I915_WRITE(DSPFW2,
945 (wm->fbc_en ? DSPFW_FBC_SR_EN : 0) |
946 FW_WM(wm->sr.fbc, FBC_SR) |
947 FW_WM(wm->hpll.fbc, FBC_HPLL_SR) |
948 FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEB) |
949 FW_WM(wm->pipe[PIPE_A].plane[PLANE_CURSOR], CURSORA) |
950 FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0], SPRITEA));
951 I915_WRITE(DSPFW3,
952 (wm->hpll_en ? DSPFW_HPLL_SR_EN : 0) |
953 FW_WM(wm->sr.cursor, CURSOR_SR) |
954 FW_WM(wm->hpll.cursor, HPLL_CURSOR) |
955 FW_WM(wm->hpll.plane, HPLL_SR));
956
957 POSTING_READ(DSPFW1);
958 }
959
960 #define FW_WM_VLV(value, plane) \
961 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
962
vlv_write_wm_values(struct drm_i915_private * dev_priv,const struct vlv_wm_values * wm)963 static void vlv_write_wm_values(struct drm_i915_private *dev_priv,
964 const struct vlv_wm_values *wm)
965 {
966 enum i915_pipe pipe;
967
968 for_each_pipe(dev_priv, pipe) {
969 trace_vlv_wm(intel_get_crtc_for_pipe(dev_priv, pipe), wm);
970
971 I915_WRITE(VLV_DDL(pipe),
972 (wm->ddl[pipe].plane[PLANE_CURSOR] << DDL_CURSOR_SHIFT) |
973 (wm->ddl[pipe].plane[PLANE_SPRITE1] << DDL_SPRITE_SHIFT(1)) |
974 (wm->ddl[pipe].plane[PLANE_SPRITE0] << DDL_SPRITE_SHIFT(0)) |
975 (wm->ddl[pipe].plane[PLANE_PRIMARY] << DDL_PLANE_SHIFT));
976 }
977
978 /*
979 * Zero the (unused) WM1 watermarks, and also clear all the
980 * high order bits so that there are no out of bounds values
981 * present in the registers during the reprogramming.
982 */
983 I915_WRITE(DSPHOWM, 0);
984 I915_WRITE(DSPHOWM1, 0);
985 I915_WRITE(DSPFW4, 0);
986 I915_WRITE(DSPFW5, 0);
987 I915_WRITE(DSPFW6, 0);
988
989 I915_WRITE(DSPFW1,
990 FW_WM(wm->sr.plane, SR) |
991 FW_WM(wm->pipe[PIPE_B].plane[PLANE_CURSOR], CURSORB) |
992 FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_PRIMARY], PLANEB) |
993 FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_PRIMARY], PLANEA));
994 I915_WRITE(DSPFW2,
995 FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_SPRITE1], SPRITEB) |
996 FW_WM(wm->pipe[PIPE_A].plane[PLANE_CURSOR], CURSORA) |
997 FW_WM_VLV(wm->pipe[PIPE_A].plane[PLANE_SPRITE0], SPRITEA));
998 I915_WRITE(DSPFW3,
999 FW_WM(wm->sr.cursor, CURSOR_SR));
1000
1001 if (IS_CHERRYVIEW(dev_priv)) {
1002 I915_WRITE(DSPFW7_CHV,
1003 FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE1], SPRITED) |
1004 FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEC));
1005 I915_WRITE(DSPFW8_CHV,
1006 FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_SPRITE1], SPRITEF) |
1007 FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_SPRITE0], SPRITEE));
1008 I915_WRITE(DSPFW9_CHV,
1009 FW_WM_VLV(wm->pipe[PIPE_C].plane[PLANE_PRIMARY], PLANEC) |
1010 FW_WM(wm->pipe[PIPE_C].plane[PLANE_CURSOR], CURSORC));
1011 I915_WRITE(DSPHOWM,
1012 FW_WM(wm->sr.plane >> 9, SR_HI) |
1013 FW_WM(wm->pipe[PIPE_C].plane[PLANE_SPRITE1] >> 8, SPRITEF_HI) |
1014 FW_WM(wm->pipe[PIPE_C].plane[PLANE_SPRITE0] >> 8, SPRITEE_HI) |
1015 FW_WM(wm->pipe[PIPE_C].plane[PLANE_PRIMARY] >> 8, PLANEC_HI) |
1016 FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE1] >> 8, SPRITED_HI) |
1017 FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0] >> 8, SPRITEC_HI) |
1018 FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY] >> 8, PLANEB_HI) |
1019 FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE1] >> 8, SPRITEB_HI) |
1020 FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0] >> 8, SPRITEA_HI) |
1021 FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY] >> 8, PLANEA_HI));
1022 } else {
1023 I915_WRITE(DSPFW7,
1024 FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE1], SPRITED) |
1025 FW_WM_VLV(wm->pipe[PIPE_B].plane[PLANE_SPRITE0], SPRITEC));
1026 I915_WRITE(DSPHOWM,
1027 FW_WM(wm->sr.plane >> 9, SR_HI) |
1028 FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE1] >> 8, SPRITED_HI) |
1029 FW_WM(wm->pipe[PIPE_B].plane[PLANE_SPRITE0] >> 8, SPRITEC_HI) |
1030 FW_WM(wm->pipe[PIPE_B].plane[PLANE_PRIMARY] >> 8, PLANEB_HI) |
1031 FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE1] >> 8, SPRITEB_HI) |
1032 FW_WM(wm->pipe[PIPE_A].plane[PLANE_SPRITE0] >> 8, SPRITEA_HI) |
1033 FW_WM(wm->pipe[PIPE_A].plane[PLANE_PRIMARY] >> 8, PLANEA_HI));
1034 }
1035
1036 POSTING_READ(DSPFW1);
1037 }
1038
1039 #undef FW_WM_VLV
1040
g4x_setup_wm_latency(struct drm_i915_private * dev_priv)1041 static void g4x_setup_wm_latency(struct drm_i915_private *dev_priv)
1042 {
1043 /* all latencies in usec */
1044 dev_priv->wm.pri_latency[G4X_WM_LEVEL_NORMAL] = 5;
1045 dev_priv->wm.pri_latency[G4X_WM_LEVEL_SR] = 12;
1046 dev_priv->wm.pri_latency[G4X_WM_LEVEL_HPLL] = 35;
1047
1048 dev_priv->wm.max_level = G4X_WM_LEVEL_HPLL;
1049 }
1050
g4x_plane_fifo_size(enum plane_id plane_id,int level)1051 static int g4x_plane_fifo_size(enum plane_id plane_id, int level)
1052 {
1053 /*
1054 * DSPCNTR[13] supposedly controls whether the
1055 * primary plane can use the FIFO space otherwise
1056 * reserved for the sprite plane. It's not 100% clear
1057 * what the actual FIFO size is, but it looks like we
1058 * can happily set both primary and sprite watermarks
1059 * up to 127 cachelines. So that would seem to mean
1060 * that either DSPCNTR[13] doesn't do anything, or that
1061 * the total FIFO is >= 256 cachelines in size. Either
1062 * way, we don't seem to have to worry about this
1063 * repartitioning as the maximum watermark value the
1064 * register can hold for each plane is lower than the
1065 * minimum FIFO size.
1066 */
1067 switch (plane_id) {
1068 case PLANE_CURSOR:
1069 return 63;
1070 case PLANE_PRIMARY:
1071 return level == G4X_WM_LEVEL_NORMAL ? 127 : 511;
1072 case PLANE_SPRITE0:
1073 return level == G4X_WM_LEVEL_NORMAL ? 127 : 0;
1074 default:
1075 MISSING_CASE(plane_id);
1076 return 0;
1077 }
1078 }
1079
g4x_fbc_fifo_size(int level)1080 static int g4x_fbc_fifo_size(int level)
1081 {
1082 switch (level) {
1083 case G4X_WM_LEVEL_SR:
1084 return 7;
1085 case G4X_WM_LEVEL_HPLL:
1086 return 15;
1087 default:
1088 MISSING_CASE(level);
1089 return 0;
1090 }
1091 }
1092
g4x_compute_wm(const struct intel_crtc_state * crtc_state,const struct intel_plane_state * plane_state,int level)1093 static uint16_t g4x_compute_wm(const struct intel_crtc_state *crtc_state,
1094 const struct intel_plane_state *plane_state,
1095 int level)
1096 {
1097 struct intel_plane *plane = to_intel_plane(plane_state->base.plane);
1098 struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
1099 const struct drm_display_mode *adjusted_mode =
1100 &crtc_state->base.adjusted_mode;
1101 int clock, htotal, cpp, width, wm;
1102 int latency = dev_priv->wm.pri_latency[level] * 10;
1103
1104 if (latency == 0)
1105 return USHRT_MAX;
1106
1107 if (!intel_wm_plane_visible(crtc_state, plane_state))
1108 return 0;
1109
1110 /*
1111 * Not 100% sure which way ELK should go here as the
1112 * spec only says CL/CTG should assume 32bpp and BW
1113 * doesn't need to. But as these things followed the
1114 * mobile vs. desktop lines on gen3 as well, let's
1115 * assume ELK doesn't need this.
1116 *
1117 * The spec also fails to list such a restriction for
1118 * the HPLL watermark, which seems a little strange.
1119 * Let's use 32bpp for the HPLL watermark as well.
1120 */
1121 if (IS_GM45(dev_priv) && plane->id == PLANE_PRIMARY &&
1122 level != G4X_WM_LEVEL_NORMAL)
1123 cpp = 4;
1124 else
1125 cpp = plane_state->base.fb->format->cpp[0];
1126
1127 clock = adjusted_mode->crtc_clock;
1128 htotal = adjusted_mode->crtc_htotal;
1129
1130 if (plane->id == PLANE_CURSOR)
1131 width = plane_state->base.crtc_w;
1132 else
1133 width = drm_rect_width(&plane_state->base.dst);
1134
1135 if (plane->id == PLANE_CURSOR) {
1136 wm = intel_wm_method2(clock, htotal, width, cpp, latency);
1137 } else if (plane->id == PLANE_PRIMARY &&
1138 level == G4X_WM_LEVEL_NORMAL) {
1139 wm = intel_wm_method1(clock, cpp, latency);
1140 } else {
1141 int small, large;
1142
1143 small = intel_wm_method1(clock, cpp, latency);
1144 large = intel_wm_method2(clock, htotal, width, cpp, latency);
1145
1146 wm = min(small, large);
1147 }
1148
1149 wm += g4x_tlb_miss_wa(g4x_plane_fifo_size(plane->id, level),
1150 width, cpp);
1151
1152 wm = DIV_ROUND_UP(wm, 64) + 2;
1153
1154 return min_t(int, wm, USHRT_MAX);
1155 }
1156
g4x_raw_plane_wm_set(struct intel_crtc_state * crtc_state,int level,enum plane_id plane_id,u16 value)1157 static bool g4x_raw_plane_wm_set(struct intel_crtc_state *crtc_state,
1158 int level, enum plane_id plane_id, u16 value)
1159 {
1160 struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1161 bool dirty = false;
1162
1163 for (; level < intel_wm_num_levels(dev_priv); level++) {
1164 struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1165
1166 dirty |= raw->plane[plane_id] != value;
1167 raw->plane[plane_id] = value;
1168 }
1169
1170 return dirty;
1171 }
1172
g4x_raw_fbc_wm_set(struct intel_crtc_state * crtc_state,int level,u16 value)1173 static bool g4x_raw_fbc_wm_set(struct intel_crtc_state *crtc_state,
1174 int level, u16 value)
1175 {
1176 struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1177 bool dirty = false;
1178
1179 /* NORMAL level doesn't have an FBC watermark */
1180 level = max(level, G4X_WM_LEVEL_SR);
1181
1182 for (; level < intel_wm_num_levels(dev_priv); level++) {
1183 struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1184
1185 dirty |= raw->fbc != value;
1186 raw->fbc = value;
1187 }
1188
1189 return dirty;
1190 }
1191
1192 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state *cstate,
1193 const struct intel_plane_state *pstate,
1194 uint32_t pri_val);
1195
g4x_raw_plane_wm_compute(struct intel_crtc_state * crtc_state,const struct intel_plane_state * plane_state)1196 static bool g4x_raw_plane_wm_compute(struct intel_crtc_state *crtc_state,
1197 const struct intel_plane_state *plane_state)
1198 {
1199 struct intel_plane *plane = to_intel_plane(plane_state->base.plane);
1200 int num_levels = intel_wm_num_levels(to_i915(plane->base.dev));
1201 enum plane_id plane_id = plane->id;
1202 bool dirty = false;
1203 int level;
1204
1205 if (!intel_wm_plane_visible(crtc_state, plane_state)) {
1206 dirty |= g4x_raw_plane_wm_set(crtc_state, 0, plane_id, 0);
1207 if (plane_id == PLANE_PRIMARY)
1208 dirty |= g4x_raw_fbc_wm_set(crtc_state, 0, 0);
1209 goto out;
1210 }
1211
1212 for (level = 0; level < num_levels; level++) {
1213 struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1214 int wm, max_wm;
1215
1216 wm = g4x_compute_wm(crtc_state, plane_state, level);
1217 max_wm = g4x_plane_fifo_size(plane_id, level);
1218
1219 if (wm > max_wm)
1220 break;
1221
1222 dirty |= raw->plane[plane_id] != wm;
1223 raw->plane[plane_id] = wm;
1224
1225 if (plane_id != PLANE_PRIMARY ||
1226 level == G4X_WM_LEVEL_NORMAL)
1227 continue;
1228
1229 wm = ilk_compute_fbc_wm(crtc_state, plane_state,
1230 raw->plane[plane_id]);
1231 max_wm = g4x_fbc_fifo_size(level);
1232
1233 /*
1234 * FBC wm is not mandatory as we
1235 * can always just disable its use.
1236 */
1237 if (wm > max_wm)
1238 wm = USHRT_MAX;
1239
1240 dirty |= raw->fbc != wm;
1241 raw->fbc = wm;
1242 }
1243
1244 /* mark watermarks as invalid */
1245 dirty |= g4x_raw_plane_wm_set(crtc_state, level, plane_id, USHRT_MAX);
1246
1247 if (plane_id == PLANE_PRIMARY)
1248 dirty |= g4x_raw_fbc_wm_set(crtc_state, level, USHRT_MAX);
1249
1250 out:
1251 if (dirty) {
1252 DRM_DEBUG_KMS("%s watermarks: normal=%d, SR=%d, HPLL=%d\n",
1253 plane->base.name,
1254 crtc_state->wm.g4x.raw[G4X_WM_LEVEL_NORMAL].plane[plane_id],
1255 crtc_state->wm.g4x.raw[G4X_WM_LEVEL_SR].plane[plane_id],
1256 crtc_state->wm.g4x.raw[G4X_WM_LEVEL_HPLL].plane[plane_id]);
1257
1258 if (plane_id == PLANE_PRIMARY)
1259 DRM_DEBUG_KMS("FBC watermarks: SR=%d, HPLL=%d\n",
1260 crtc_state->wm.g4x.raw[G4X_WM_LEVEL_SR].fbc,
1261 crtc_state->wm.g4x.raw[G4X_WM_LEVEL_HPLL].fbc);
1262 }
1263
1264 return dirty;
1265 }
1266
g4x_raw_plane_wm_is_valid(const struct intel_crtc_state * crtc_state,enum plane_id plane_id,int level)1267 static bool g4x_raw_plane_wm_is_valid(const struct intel_crtc_state *crtc_state,
1268 enum plane_id plane_id, int level)
1269 {
1270 const struct g4x_pipe_wm *raw = &crtc_state->wm.g4x.raw[level];
1271
1272 return raw->plane[plane_id] <= g4x_plane_fifo_size(plane_id, level);
1273 }
1274
g4x_raw_crtc_wm_is_valid(const struct intel_crtc_state * crtc_state,int level)1275 static bool g4x_raw_crtc_wm_is_valid(const struct intel_crtc_state *crtc_state,
1276 int level)
1277 {
1278 struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1279
1280 if (level > dev_priv->wm.max_level)
1281 return false;
1282
1283 return g4x_raw_plane_wm_is_valid(crtc_state, PLANE_PRIMARY, level) &&
1284 g4x_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE0, level) &&
1285 g4x_raw_plane_wm_is_valid(crtc_state, PLANE_CURSOR, level);
1286 }
1287
1288 /* mark all levels starting from 'level' as invalid */
g4x_invalidate_wms(struct intel_crtc * crtc,struct g4x_wm_state * wm_state,int level)1289 static void g4x_invalidate_wms(struct intel_crtc *crtc,
1290 struct g4x_wm_state *wm_state, int level)
1291 {
1292 if (level <= G4X_WM_LEVEL_NORMAL) {
1293 enum plane_id plane_id;
1294
1295 for_each_plane_id_on_crtc(crtc, plane_id)
1296 wm_state->wm.plane[plane_id] = USHRT_MAX;
1297 }
1298
1299 if (level <= G4X_WM_LEVEL_SR) {
1300 wm_state->cxsr = false;
1301 wm_state->sr.cursor = USHRT_MAX;
1302 wm_state->sr.plane = USHRT_MAX;
1303 wm_state->sr.fbc = USHRT_MAX;
1304 }
1305
1306 if (level <= G4X_WM_LEVEL_HPLL) {
1307 wm_state->hpll_en = false;
1308 wm_state->hpll.cursor = USHRT_MAX;
1309 wm_state->hpll.plane = USHRT_MAX;
1310 wm_state->hpll.fbc = USHRT_MAX;
1311 }
1312 }
1313
g4x_compute_pipe_wm(struct intel_crtc_state * crtc_state)1314 static int g4x_compute_pipe_wm(struct intel_crtc_state *crtc_state)
1315 {
1316 struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1317 struct intel_atomic_state *state =
1318 to_intel_atomic_state(crtc_state->base.state);
1319 struct g4x_wm_state *wm_state = &crtc_state->wm.g4x.optimal;
1320 int num_active_planes = hweight32(crtc_state->active_planes &
1321 ~BIT(PLANE_CURSOR));
1322 const struct g4x_pipe_wm *raw;
1323 const struct intel_plane_state *old_plane_state;
1324 const struct intel_plane_state *new_plane_state;
1325 struct intel_plane *plane;
1326 enum plane_id plane_id;
1327 int i, level;
1328 unsigned int dirty = 0;
1329
1330 for_each_oldnew_intel_plane_in_state(state, plane,
1331 old_plane_state,
1332 new_plane_state, i) {
1333 if (new_plane_state->base.crtc != &crtc->base &&
1334 old_plane_state->base.crtc != &crtc->base)
1335 continue;
1336
1337 if (g4x_raw_plane_wm_compute(crtc_state, new_plane_state))
1338 dirty |= BIT(plane->id);
1339 }
1340
1341 if (!dirty)
1342 return 0;
1343
1344 level = G4X_WM_LEVEL_NORMAL;
1345 if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1346 goto out;
1347
1348 raw = &crtc_state->wm.g4x.raw[level];
1349 for_each_plane_id_on_crtc(crtc, plane_id)
1350 wm_state->wm.plane[plane_id] = raw->plane[plane_id];
1351
1352 level = G4X_WM_LEVEL_SR;
1353
1354 if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1355 goto out;
1356
1357 raw = &crtc_state->wm.g4x.raw[level];
1358 wm_state->sr.plane = raw->plane[PLANE_PRIMARY];
1359 wm_state->sr.cursor = raw->plane[PLANE_CURSOR];
1360 wm_state->sr.fbc = raw->fbc;
1361
1362 wm_state->cxsr = num_active_planes == BIT(PLANE_PRIMARY);
1363
1364 level = G4X_WM_LEVEL_HPLL;
1365
1366 if (!g4x_raw_crtc_wm_is_valid(crtc_state, level))
1367 goto out;
1368
1369 raw = &crtc_state->wm.g4x.raw[level];
1370 wm_state->hpll.plane = raw->plane[PLANE_PRIMARY];
1371 wm_state->hpll.cursor = raw->plane[PLANE_CURSOR];
1372 wm_state->hpll.fbc = raw->fbc;
1373
1374 wm_state->hpll_en = wm_state->cxsr;
1375
1376 level++;
1377
1378 out:
1379 if (level == G4X_WM_LEVEL_NORMAL)
1380 return -EINVAL;
1381
1382 /* invalidate the higher levels */
1383 g4x_invalidate_wms(crtc, wm_state, level);
1384
1385 /*
1386 * Determine if the FBC watermark(s) can be used. IF
1387 * this isn't the case we prefer to disable the FBC
1388 ( watermark(s) rather than disable the SR/HPLL
1389 * level(s) entirely.
1390 */
1391 wm_state->fbc_en = level > G4X_WM_LEVEL_NORMAL;
1392
1393 if (level >= G4X_WM_LEVEL_SR &&
1394 wm_state->sr.fbc > g4x_fbc_fifo_size(G4X_WM_LEVEL_SR))
1395 wm_state->fbc_en = false;
1396 else if (level >= G4X_WM_LEVEL_HPLL &&
1397 wm_state->hpll.fbc > g4x_fbc_fifo_size(G4X_WM_LEVEL_HPLL))
1398 wm_state->fbc_en = false;
1399
1400 return 0;
1401 }
1402
g4x_compute_intermediate_wm(struct drm_device * dev,struct intel_crtc * crtc,struct intel_crtc_state * crtc_state)1403 static int g4x_compute_intermediate_wm(struct drm_device *dev,
1404 struct intel_crtc *crtc,
1405 struct intel_crtc_state *crtc_state)
1406 {
1407 struct g4x_wm_state *intermediate = &crtc_state->wm.g4x.intermediate;
1408 const struct g4x_wm_state *optimal = &crtc_state->wm.g4x.optimal;
1409 const struct g4x_wm_state *active = &crtc->wm.active.g4x;
1410 enum plane_id plane_id;
1411
1412 intermediate->cxsr = optimal->cxsr && active->cxsr &&
1413 !crtc_state->disable_cxsr;
1414 intermediate->hpll_en = optimal->hpll_en && active->hpll_en &&
1415 !crtc_state->disable_cxsr;
1416 intermediate->fbc_en = optimal->fbc_en && active->fbc_en;
1417
1418 for_each_plane_id_on_crtc(crtc, plane_id) {
1419 intermediate->wm.plane[plane_id] =
1420 max(optimal->wm.plane[plane_id],
1421 active->wm.plane[plane_id]);
1422
1423 WARN_ON(intermediate->wm.plane[plane_id] >
1424 g4x_plane_fifo_size(plane_id, G4X_WM_LEVEL_NORMAL));
1425 }
1426
1427 intermediate->sr.plane = max(optimal->sr.plane,
1428 active->sr.plane);
1429 intermediate->sr.cursor = max(optimal->sr.cursor,
1430 active->sr.cursor);
1431 intermediate->sr.fbc = max(optimal->sr.fbc,
1432 active->sr.fbc);
1433
1434 intermediate->hpll.plane = max(optimal->hpll.plane,
1435 active->hpll.plane);
1436 intermediate->hpll.cursor = max(optimal->hpll.cursor,
1437 active->hpll.cursor);
1438 intermediate->hpll.fbc = max(optimal->hpll.fbc,
1439 active->hpll.fbc);
1440
1441 WARN_ON((intermediate->sr.plane >
1442 g4x_plane_fifo_size(PLANE_PRIMARY, G4X_WM_LEVEL_SR) ||
1443 intermediate->sr.cursor >
1444 g4x_plane_fifo_size(PLANE_CURSOR, G4X_WM_LEVEL_SR)) &&
1445 intermediate->cxsr);
1446 WARN_ON((intermediate->sr.plane >
1447 g4x_plane_fifo_size(PLANE_PRIMARY, G4X_WM_LEVEL_HPLL) ||
1448 intermediate->sr.cursor >
1449 g4x_plane_fifo_size(PLANE_CURSOR, G4X_WM_LEVEL_HPLL)) &&
1450 intermediate->hpll_en);
1451
1452 WARN_ON(intermediate->sr.fbc > g4x_fbc_fifo_size(1) &&
1453 intermediate->fbc_en && intermediate->cxsr);
1454 WARN_ON(intermediate->hpll.fbc > g4x_fbc_fifo_size(2) &&
1455 intermediate->fbc_en && intermediate->hpll_en);
1456
1457 /*
1458 * If our intermediate WM are identical to the final WM, then we can
1459 * omit the post-vblank programming; only update if it's different.
1460 */
1461 if (memcmp(intermediate, optimal, sizeof(*intermediate)) != 0)
1462 crtc_state->wm.need_postvbl_update = true;
1463
1464 return 0;
1465 }
1466
g4x_merge_wm(struct drm_i915_private * dev_priv,struct g4x_wm_values * wm)1467 static void g4x_merge_wm(struct drm_i915_private *dev_priv,
1468 struct g4x_wm_values *wm)
1469 {
1470 struct intel_crtc *crtc;
1471 int num_active_crtcs = 0;
1472
1473 wm->cxsr = true;
1474 wm->hpll_en = true;
1475 wm->fbc_en = true;
1476
1477 for_each_intel_crtc(&dev_priv->drm, crtc) {
1478 const struct g4x_wm_state *wm_state = &crtc->wm.active.g4x;
1479
1480 if (!crtc->active)
1481 continue;
1482
1483 if (!wm_state->cxsr)
1484 wm->cxsr = false;
1485 if (!wm_state->hpll_en)
1486 wm->hpll_en = false;
1487 if (!wm_state->fbc_en)
1488 wm->fbc_en = false;
1489
1490 num_active_crtcs++;
1491 }
1492
1493 if (num_active_crtcs != 1) {
1494 wm->cxsr = false;
1495 wm->hpll_en = false;
1496 wm->fbc_en = false;
1497 }
1498
1499 for_each_intel_crtc(&dev_priv->drm, crtc) {
1500 const struct g4x_wm_state *wm_state = &crtc->wm.active.g4x;
1501 enum i915_pipe pipe = crtc->pipe;
1502
1503 wm->pipe[pipe] = wm_state->wm;
1504 if (crtc->active && wm->cxsr)
1505 wm->sr = wm_state->sr;
1506 if (crtc->active && wm->hpll_en)
1507 wm->hpll = wm_state->hpll;
1508 }
1509 }
1510
g4x_program_watermarks(struct drm_i915_private * dev_priv)1511 static void g4x_program_watermarks(struct drm_i915_private *dev_priv)
1512 {
1513 struct g4x_wm_values *old_wm = &dev_priv->wm.g4x;
1514 struct g4x_wm_values new_wm = {};
1515
1516 g4x_merge_wm(dev_priv, &new_wm);
1517
1518 if (memcmp(old_wm, &new_wm, sizeof(new_wm)) == 0)
1519 return;
1520
1521 if (is_disabling(old_wm->cxsr, new_wm.cxsr, true))
1522 _intel_set_memory_cxsr(dev_priv, false);
1523
1524 g4x_write_wm_values(dev_priv, &new_wm);
1525
1526 if (is_enabling(old_wm->cxsr, new_wm.cxsr, true))
1527 _intel_set_memory_cxsr(dev_priv, true);
1528
1529 *old_wm = new_wm;
1530 }
1531
g4x_initial_watermarks(struct intel_atomic_state * state,struct intel_crtc_state * crtc_state)1532 static void g4x_initial_watermarks(struct intel_atomic_state *state,
1533 struct intel_crtc_state *crtc_state)
1534 {
1535 struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1536 struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1537
1538 mutex_lock(&dev_priv->wm.wm_mutex);
1539 crtc->wm.active.g4x = crtc_state->wm.g4x.intermediate;
1540 g4x_program_watermarks(dev_priv);
1541 mutex_unlock(&dev_priv->wm.wm_mutex);
1542 }
1543
g4x_optimize_watermarks(struct intel_atomic_state * state,struct intel_crtc_state * crtc_state)1544 static void g4x_optimize_watermarks(struct intel_atomic_state *state,
1545 struct intel_crtc_state *crtc_state)
1546 {
1547 struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1548 struct intel_crtc *intel_crtc = to_intel_crtc(crtc_state->base.crtc);
1549
1550 if (!crtc_state->wm.need_postvbl_update)
1551 return;
1552
1553 mutex_lock(&dev_priv->wm.wm_mutex);
1554 intel_crtc->wm.active.g4x = crtc_state->wm.g4x.optimal;
1555 g4x_program_watermarks(dev_priv);
1556 mutex_unlock(&dev_priv->wm.wm_mutex);
1557 }
1558
1559 /* latency must be in 0.1us units. */
vlv_wm_method2(unsigned int pixel_rate,unsigned int htotal,unsigned int width,unsigned int cpp,unsigned int latency)1560 static unsigned int vlv_wm_method2(unsigned int pixel_rate,
1561 unsigned int htotal,
1562 unsigned int width,
1563 unsigned int cpp,
1564 unsigned int latency)
1565 {
1566 unsigned int ret;
1567
1568 ret = intel_wm_method2(pixel_rate, htotal,
1569 width, cpp, latency);
1570 ret = DIV_ROUND_UP(ret, 64);
1571
1572 return ret;
1573 }
1574
vlv_setup_wm_latency(struct drm_i915_private * dev_priv)1575 static void vlv_setup_wm_latency(struct drm_i915_private *dev_priv)
1576 {
1577 /* all latencies in usec */
1578 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM2] = 3;
1579
1580 dev_priv->wm.max_level = VLV_WM_LEVEL_PM2;
1581
1582 if (IS_CHERRYVIEW(dev_priv)) {
1583 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM5] = 12;
1584 dev_priv->wm.pri_latency[VLV_WM_LEVEL_DDR_DVFS] = 33;
1585
1586 dev_priv->wm.max_level = VLV_WM_LEVEL_DDR_DVFS;
1587 }
1588 }
1589
vlv_compute_wm_level(const struct intel_crtc_state * crtc_state,const struct intel_plane_state * plane_state,int level)1590 static uint16_t vlv_compute_wm_level(const struct intel_crtc_state *crtc_state,
1591 const struct intel_plane_state *plane_state,
1592 int level)
1593 {
1594 struct intel_plane *plane = to_intel_plane(plane_state->base.plane);
1595 struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
1596 const struct drm_display_mode *adjusted_mode =
1597 &crtc_state->base.adjusted_mode;
1598 int clock, htotal, cpp, width, wm;
1599
1600 if (dev_priv->wm.pri_latency[level] == 0)
1601 return USHRT_MAX;
1602
1603 if (!intel_wm_plane_visible(crtc_state, plane_state))
1604 return 0;
1605
1606 cpp = plane_state->base.fb->format->cpp[0];
1607 clock = adjusted_mode->crtc_clock;
1608 htotal = adjusted_mode->crtc_htotal;
1609 width = crtc_state->pipe_src_w;
1610
1611 if (plane->id == PLANE_CURSOR) {
1612 /*
1613 * FIXME the formula gives values that are
1614 * too big for the cursor FIFO, and hence we
1615 * would never be able to use cursors. For
1616 * now just hardcode the watermark.
1617 */
1618 wm = 63;
1619 } else {
1620 wm = vlv_wm_method2(clock, htotal, width, cpp,
1621 dev_priv->wm.pri_latency[level] * 10);
1622 }
1623
1624 return min_t(int, wm, USHRT_MAX);
1625 }
1626
vlv_need_sprite0_fifo_workaround(unsigned int active_planes)1627 static bool vlv_need_sprite0_fifo_workaround(unsigned int active_planes)
1628 {
1629 return (active_planes & (BIT(PLANE_SPRITE0) |
1630 BIT(PLANE_SPRITE1))) == BIT(PLANE_SPRITE1);
1631 }
1632
vlv_compute_fifo(struct intel_crtc_state * crtc_state)1633 static int vlv_compute_fifo(struct intel_crtc_state *crtc_state)
1634 {
1635 struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1636 const struct g4x_pipe_wm *raw =
1637 &crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM2];
1638 struct vlv_fifo_state *fifo_state = &crtc_state->wm.vlv.fifo_state;
1639 unsigned int active_planes = crtc_state->active_planes & ~BIT(PLANE_CURSOR);
1640 int num_active_planes = hweight32(active_planes);
1641 const int fifo_size = 511;
1642 int fifo_extra, fifo_left = fifo_size;
1643 int sprite0_fifo_extra = 0;
1644 unsigned int total_rate;
1645 enum plane_id plane_id;
1646
1647 /*
1648 * When enabling sprite0 after sprite1 has already been enabled
1649 * we tend to get an underrun unless sprite0 already has some
1650 * FIFO space allcoated. Hence we always allocate at least one
1651 * cacheline for sprite0 whenever sprite1 is enabled.
1652 *
1653 * All other plane enable sequences appear immune to this problem.
1654 */
1655 if (vlv_need_sprite0_fifo_workaround(active_planes))
1656 sprite0_fifo_extra = 1;
1657
1658 total_rate = raw->plane[PLANE_PRIMARY] +
1659 raw->plane[PLANE_SPRITE0] +
1660 raw->plane[PLANE_SPRITE1] +
1661 sprite0_fifo_extra;
1662
1663 if (total_rate > fifo_size)
1664 return -EINVAL;
1665
1666 if (total_rate == 0)
1667 total_rate = 1;
1668
1669 for_each_plane_id_on_crtc(crtc, plane_id) {
1670 unsigned int rate;
1671
1672 if ((active_planes & BIT(plane_id)) == 0) {
1673 fifo_state->plane[plane_id] = 0;
1674 continue;
1675 }
1676
1677 rate = raw->plane[plane_id];
1678 fifo_state->plane[plane_id] = fifo_size * rate / total_rate;
1679 fifo_left -= fifo_state->plane[plane_id];
1680 }
1681
1682 fifo_state->plane[PLANE_SPRITE0] += sprite0_fifo_extra;
1683 fifo_left -= sprite0_fifo_extra;
1684
1685 fifo_state->plane[PLANE_CURSOR] = 63;
1686
1687 fifo_extra = DIV_ROUND_UP(fifo_left, num_active_planes ?: 1);
1688
1689 /* spread the remainder evenly */
1690 for_each_plane_id_on_crtc(crtc, plane_id) {
1691 int plane_extra;
1692
1693 if (fifo_left == 0)
1694 break;
1695
1696 if ((active_planes & BIT(plane_id)) == 0)
1697 continue;
1698
1699 plane_extra = min(fifo_extra, fifo_left);
1700 fifo_state->plane[plane_id] += plane_extra;
1701 fifo_left -= plane_extra;
1702 }
1703
1704 WARN_ON(active_planes != 0 && fifo_left != 0);
1705
1706 /* give it all to the first plane if none are active */
1707 if (active_planes == 0) {
1708 WARN_ON(fifo_left != fifo_size);
1709 fifo_state->plane[PLANE_PRIMARY] = fifo_left;
1710 }
1711
1712 return 0;
1713 }
1714
1715 /* mark all levels starting from 'level' as invalid */
vlv_invalidate_wms(struct intel_crtc * crtc,struct vlv_wm_state * wm_state,int level)1716 static void vlv_invalidate_wms(struct intel_crtc *crtc,
1717 struct vlv_wm_state *wm_state, int level)
1718 {
1719 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1720
1721 for (; level < intel_wm_num_levels(dev_priv); level++) {
1722 enum plane_id plane_id;
1723
1724 for_each_plane_id_on_crtc(crtc, plane_id)
1725 wm_state->wm[level].plane[plane_id] = USHRT_MAX;
1726
1727 wm_state->sr[level].cursor = USHRT_MAX;
1728 wm_state->sr[level].plane = USHRT_MAX;
1729 }
1730 }
1731
vlv_invert_wm_value(u16 wm,u16 fifo_size)1732 static u16 vlv_invert_wm_value(u16 wm, u16 fifo_size)
1733 {
1734 if (wm > fifo_size)
1735 return USHRT_MAX;
1736 else
1737 return fifo_size - wm;
1738 }
1739
1740 /*
1741 * Starting from 'level' set all higher
1742 * levels to 'value' in the "raw" watermarks.
1743 */
vlv_raw_plane_wm_set(struct intel_crtc_state * crtc_state,int level,enum plane_id plane_id,u16 value)1744 static bool vlv_raw_plane_wm_set(struct intel_crtc_state *crtc_state,
1745 int level, enum plane_id plane_id, u16 value)
1746 {
1747 struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
1748 int num_levels = intel_wm_num_levels(dev_priv);
1749 bool dirty = false;
1750
1751 for (; level < num_levels; level++) {
1752 struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1753
1754 dirty |= raw->plane[plane_id] != value;
1755 raw->plane[plane_id] = value;
1756 }
1757
1758 return dirty;
1759 }
1760
vlv_raw_plane_wm_compute(struct intel_crtc_state * crtc_state,const struct intel_plane_state * plane_state)1761 static bool vlv_raw_plane_wm_compute(struct intel_crtc_state *crtc_state,
1762 const struct intel_plane_state *plane_state)
1763 {
1764 struct intel_plane *plane = to_intel_plane(plane_state->base.plane);
1765 enum plane_id plane_id = plane->id;
1766 int num_levels = intel_wm_num_levels(to_i915(plane->base.dev));
1767 int level;
1768 bool dirty = false;
1769
1770 if (!intel_wm_plane_visible(crtc_state, plane_state)) {
1771 dirty |= vlv_raw_plane_wm_set(crtc_state, 0, plane_id, 0);
1772 goto out;
1773 }
1774
1775 for (level = 0; level < num_levels; level++) {
1776 struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1777 int wm = vlv_compute_wm_level(crtc_state, plane_state, level);
1778 int max_wm = plane_id == PLANE_CURSOR ? 63 : 511;
1779
1780 if (wm > max_wm)
1781 break;
1782
1783 dirty |= raw->plane[plane_id] != wm;
1784 raw->plane[plane_id] = wm;
1785 }
1786
1787 /* mark all higher levels as invalid */
1788 dirty |= vlv_raw_plane_wm_set(crtc_state, level, plane_id, USHRT_MAX);
1789
1790 out:
1791 if (dirty)
1792 DRM_DEBUG_KMS("%s watermarks: PM2=%d, PM5=%d, DDR DVFS=%d\n",
1793 plane->base.name,
1794 crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM2].plane[plane_id],
1795 crtc_state->wm.vlv.raw[VLV_WM_LEVEL_PM5].plane[plane_id],
1796 crtc_state->wm.vlv.raw[VLV_WM_LEVEL_DDR_DVFS].plane[plane_id]);
1797
1798 return dirty;
1799 }
1800
vlv_raw_plane_wm_is_valid(const struct intel_crtc_state * crtc_state,enum plane_id plane_id,int level)1801 static bool vlv_raw_plane_wm_is_valid(const struct intel_crtc_state *crtc_state,
1802 enum plane_id plane_id, int level)
1803 {
1804 const struct g4x_pipe_wm *raw =
1805 &crtc_state->wm.vlv.raw[level];
1806 const struct vlv_fifo_state *fifo_state =
1807 &crtc_state->wm.vlv.fifo_state;
1808
1809 return raw->plane[plane_id] <= fifo_state->plane[plane_id];
1810 }
1811
vlv_raw_crtc_wm_is_valid(const struct intel_crtc_state * crtc_state,int level)1812 static bool vlv_raw_crtc_wm_is_valid(const struct intel_crtc_state *crtc_state, int level)
1813 {
1814 return vlv_raw_plane_wm_is_valid(crtc_state, PLANE_PRIMARY, level) &&
1815 vlv_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE0, level) &&
1816 vlv_raw_plane_wm_is_valid(crtc_state, PLANE_SPRITE1, level) &&
1817 vlv_raw_plane_wm_is_valid(crtc_state, PLANE_CURSOR, level);
1818 }
1819
vlv_compute_pipe_wm(struct intel_crtc_state * crtc_state)1820 static int vlv_compute_pipe_wm(struct intel_crtc_state *crtc_state)
1821 {
1822 struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1823 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1824 struct intel_atomic_state *state =
1825 to_intel_atomic_state(crtc_state->base.state);
1826 struct vlv_wm_state *wm_state = &crtc_state->wm.vlv.optimal;
1827 const struct vlv_fifo_state *fifo_state =
1828 &crtc_state->wm.vlv.fifo_state;
1829 int num_active_planes = hweight32(crtc_state->active_planes &
1830 ~BIT(PLANE_CURSOR));
1831 bool needs_modeset = drm_atomic_crtc_needs_modeset(&crtc_state->base);
1832 const struct intel_plane_state *old_plane_state;
1833 const struct intel_plane_state *new_plane_state;
1834 struct intel_plane *plane;
1835 enum plane_id plane_id;
1836 int level, ret, i;
1837 unsigned int dirty = 0;
1838
1839 for_each_oldnew_intel_plane_in_state(state, plane,
1840 old_plane_state,
1841 new_plane_state, i) {
1842 if (new_plane_state->base.crtc != &crtc->base &&
1843 old_plane_state->base.crtc != &crtc->base)
1844 continue;
1845
1846 if (vlv_raw_plane_wm_compute(crtc_state, new_plane_state))
1847 dirty |= BIT(plane->id);
1848 }
1849
1850 /*
1851 * DSPARB registers may have been reset due to the
1852 * power well being turned off. Make sure we restore
1853 * them to a consistent state even if no primary/sprite
1854 * planes are initially active.
1855 */
1856 if (needs_modeset)
1857 crtc_state->fifo_changed = true;
1858
1859 if (!dirty)
1860 return 0;
1861
1862 /* cursor changes don't warrant a FIFO recompute */
1863 if (dirty & ~BIT(PLANE_CURSOR)) {
1864 const struct intel_crtc_state *old_crtc_state =
1865 intel_atomic_get_old_crtc_state(state, crtc);
1866 const struct vlv_fifo_state *old_fifo_state =
1867 &old_crtc_state->wm.vlv.fifo_state;
1868
1869 ret = vlv_compute_fifo(crtc_state);
1870 if (ret)
1871 return ret;
1872
1873 if (needs_modeset ||
1874 memcmp(old_fifo_state, fifo_state,
1875 sizeof(*fifo_state)) != 0)
1876 crtc_state->fifo_changed = true;
1877 }
1878
1879 /* initially allow all levels */
1880 wm_state->num_levels = intel_wm_num_levels(dev_priv);
1881 /*
1882 * Note that enabling cxsr with no primary/sprite planes
1883 * enabled can wedge the pipe. Hence we only allow cxsr
1884 * with exactly one enabled primary/sprite plane.
1885 */
1886 wm_state->cxsr = crtc->pipe != PIPE_C && num_active_planes == 1;
1887
1888 for (level = 0; level < wm_state->num_levels; level++) {
1889 const struct g4x_pipe_wm *raw = &crtc_state->wm.vlv.raw[level];
1890 const int sr_fifo_size = INTEL_INFO(dev_priv)->num_pipes * 512 - 1;
1891
1892 if (!vlv_raw_crtc_wm_is_valid(crtc_state, level))
1893 break;
1894
1895 for_each_plane_id_on_crtc(crtc, plane_id) {
1896 wm_state->wm[level].plane[plane_id] =
1897 vlv_invert_wm_value(raw->plane[plane_id],
1898 fifo_state->plane[plane_id]);
1899 }
1900
1901 wm_state->sr[level].plane =
1902 vlv_invert_wm_value(max3(raw->plane[PLANE_PRIMARY],
1903 raw->plane[PLANE_SPRITE0],
1904 raw->plane[PLANE_SPRITE1]),
1905 sr_fifo_size);
1906
1907 wm_state->sr[level].cursor =
1908 vlv_invert_wm_value(raw->plane[PLANE_CURSOR],
1909 63);
1910 }
1911
1912 if (level == 0)
1913 return -EINVAL;
1914
1915 /* limit to only levels we can actually handle */
1916 wm_state->num_levels = level;
1917
1918 /* invalidate the higher levels */
1919 vlv_invalidate_wms(crtc, wm_state, level);
1920
1921 return 0;
1922 }
1923
1924 #define VLV_FIFO(plane, value) \
1925 (((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1926
vlv_atomic_update_fifo(struct intel_atomic_state * state,struct intel_crtc_state * crtc_state)1927 static void vlv_atomic_update_fifo(struct intel_atomic_state *state,
1928 struct intel_crtc_state *crtc_state)
1929 {
1930 struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
1931 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1932 const struct vlv_fifo_state *fifo_state =
1933 &crtc_state->wm.vlv.fifo_state;
1934 int sprite0_start, sprite1_start, fifo_size;
1935
1936 if (!crtc_state->fifo_changed)
1937 return;
1938
1939 sprite0_start = fifo_state->plane[PLANE_PRIMARY];
1940 sprite1_start = fifo_state->plane[PLANE_SPRITE0] + sprite0_start;
1941 fifo_size = fifo_state->plane[PLANE_SPRITE1] + sprite1_start;
1942
1943 WARN_ON(fifo_state->plane[PLANE_CURSOR] != 63);
1944 WARN_ON(fifo_size != 511);
1945
1946 trace_vlv_fifo_size(crtc, sprite0_start, sprite1_start, fifo_size);
1947
1948 /*
1949 * uncore.lock serves a double purpose here. It allows us to
1950 * use the less expensive I915_{READ,WRITE}_FW() functions, and
1951 * it protects the DSPARB registers from getting clobbered by
1952 * parallel updates from multiple pipes.
1953 *
1954 * intel_pipe_update_start() has already disabled interrupts
1955 * for us, so a plain spin_lock() is sufficient here.
1956 */
1957 lockmgr(&dev_priv->uncore.lock, LK_EXCLUSIVE);
1958
1959 switch (crtc->pipe) {
1960 uint32_t dsparb, dsparb2, dsparb3;
1961 case PIPE_A:
1962 dsparb = I915_READ_FW(DSPARB);
1963 dsparb2 = I915_READ_FW(DSPARB2);
1964
1965 dsparb &= ~(VLV_FIFO(SPRITEA, 0xff) |
1966 VLV_FIFO(SPRITEB, 0xff));
1967 dsparb |= (VLV_FIFO(SPRITEA, sprite0_start) |
1968 VLV_FIFO(SPRITEB, sprite1_start));
1969
1970 dsparb2 &= ~(VLV_FIFO(SPRITEA_HI, 0x1) |
1971 VLV_FIFO(SPRITEB_HI, 0x1));
1972 dsparb2 |= (VLV_FIFO(SPRITEA_HI, sprite0_start >> 8) |
1973 VLV_FIFO(SPRITEB_HI, sprite1_start >> 8));
1974
1975 I915_WRITE_FW(DSPARB, dsparb);
1976 I915_WRITE_FW(DSPARB2, dsparb2);
1977 break;
1978 case PIPE_B:
1979 dsparb = I915_READ_FW(DSPARB);
1980 dsparb2 = I915_READ_FW(DSPARB2);
1981
1982 dsparb &= ~(VLV_FIFO(SPRITEC, 0xff) |
1983 VLV_FIFO(SPRITED, 0xff));
1984 dsparb |= (VLV_FIFO(SPRITEC, sprite0_start) |
1985 VLV_FIFO(SPRITED, sprite1_start));
1986
1987 dsparb2 &= ~(VLV_FIFO(SPRITEC_HI, 0xff) |
1988 VLV_FIFO(SPRITED_HI, 0xff));
1989 dsparb2 |= (VLV_FIFO(SPRITEC_HI, sprite0_start >> 8) |
1990 VLV_FIFO(SPRITED_HI, sprite1_start >> 8));
1991
1992 I915_WRITE_FW(DSPARB, dsparb);
1993 I915_WRITE_FW(DSPARB2, dsparb2);
1994 break;
1995 case PIPE_C:
1996 dsparb3 = I915_READ_FW(DSPARB3);
1997 dsparb2 = I915_READ_FW(DSPARB2);
1998
1999 dsparb3 &= ~(VLV_FIFO(SPRITEE, 0xff) |
2000 VLV_FIFO(SPRITEF, 0xff));
2001 dsparb3 |= (VLV_FIFO(SPRITEE, sprite0_start) |
2002 VLV_FIFO(SPRITEF, sprite1_start));
2003
2004 dsparb2 &= ~(VLV_FIFO(SPRITEE_HI, 0xff) |
2005 VLV_FIFO(SPRITEF_HI, 0xff));
2006 dsparb2 |= (VLV_FIFO(SPRITEE_HI, sprite0_start >> 8) |
2007 VLV_FIFO(SPRITEF_HI, sprite1_start >> 8));
2008
2009 I915_WRITE_FW(DSPARB3, dsparb3);
2010 I915_WRITE_FW(DSPARB2, dsparb2);
2011 break;
2012 default:
2013 break;
2014 }
2015
2016 POSTING_READ_FW(DSPARB);
2017
2018 lockmgr(&dev_priv->uncore.lock, LK_RELEASE);
2019 }
2020
2021 #undef VLV_FIFO
2022
vlv_compute_intermediate_wm(struct drm_device * dev,struct intel_crtc * crtc,struct intel_crtc_state * crtc_state)2023 static int vlv_compute_intermediate_wm(struct drm_device *dev,
2024 struct intel_crtc *crtc,
2025 struct intel_crtc_state *crtc_state)
2026 {
2027 struct vlv_wm_state *intermediate = &crtc_state->wm.vlv.intermediate;
2028 const struct vlv_wm_state *optimal = &crtc_state->wm.vlv.optimal;
2029 const struct vlv_wm_state *active = &crtc->wm.active.vlv;
2030 int level;
2031
2032 intermediate->num_levels = min(optimal->num_levels, active->num_levels);
2033 intermediate->cxsr = optimal->cxsr && active->cxsr &&
2034 !crtc_state->disable_cxsr;
2035
2036 for (level = 0; level < intermediate->num_levels; level++) {
2037 enum plane_id plane_id;
2038
2039 for_each_plane_id_on_crtc(crtc, plane_id) {
2040 intermediate->wm[level].plane[plane_id] =
2041 min(optimal->wm[level].plane[plane_id],
2042 active->wm[level].plane[plane_id]);
2043 }
2044
2045 intermediate->sr[level].plane = min(optimal->sr[level].plane,
2046 active->sr[level].plane);
2047 intermediate->sr[level].cursor = min(optimal->sr[level].cursor,
2048 active->sr[level].cursor);
2049 }
2050
2051 vlv_invalidate_wms(crtc, intermediate, level);
2052
2053 /*
2054 * If our intermediate WM are identical to the final WM, then we can
2055 * omit the post-vblank programming; only update if it's different.
2056 */
2057 if (memcmp(intermediate, optimal, sizeof(*intermediate)) != 0)
2058 crtc_state->wm.need_postvbl_update = true;
2059
2060 return 0;
2061 }
2062
vlv_merge_wm(struct drm_i915_private * dev_priv,struct vlv_wm_values * wm)2063 static void vlv_merge_wm(struct drm_i915_private *dev_priv,
2064 struct vlv_wm_values *wm)
2065 {
2066 struct intel_crtc *crtc;
2067 int num_active_crtcs = 0;
2068
2069 wm->level = dev_priv->wm.max_level;
2070 wm->cxsr = true;
2071
2072 for_each_intel_crtc(&dev_priv->drm, crtc) {
2073 const struct vlv_wm_state *wm_state = &crtc->wm.active.vlv;
2074
2075 if (!crtc->active)
2076 continue;
2077
2078 if (!wm_state->cxsr)
2079 wm->cxsr = false;
2080
2081 num_active_crtcs++;
2082 wm->level = min_t(int, wm->level, wm_state->num_levels - 1);
2083 }
2084
2085 if (num_active_crtcs != 1)
2086 wm->cxsr = false;
2087
2088 if (num_active_crtcs > 1)
2089 wm->level = VLV_WM_LEVEL_PM2;
2090
2091 for_each_intel_crtc(&dev_priv->drm, crtc) {
2092 const struct vlv_wm_state *wm_state = &crtc->wm.active.vlv;
2093 enum i915_pipe pipe = crtc->pipe;
2094
2095 wm->pipe[pipe] = wm_state->wm[wm->level];
2096 if (crtc->active && wm->cxsr)
2097 wm->sr = wm_state->sr[wm->level];
2098
2099 wm->ddl[pipe].plane[PLANE_PRIMARY] = DDL_PRECISION_HIGH | 2;
2100 wm->ddl[pipe].plane[PLANE_SPRITE0] = DDL_PRECISION_HIGH | 2;
2101 wm->ddl[pipe].plane[PLANE_SPRITE1] = DDL_PRECISION_HIGH | 2;
2102 wm->ddl[pipe].plane[PLANE_CURSOR] = DDL_PRECISION_HIGH | 2;
2103 }
2104 }
2105
vlv_program_watermarks(struct drm_i915_private * dev_priv)2106 static void vlv_program_watermarks(struct drm_i915_private *dev_priv)
2107 {
2108 struct vlv_wm_values *old_wm = &dev_priv->wm.vlv;
2109 struct vlv_wm_values new_wm = {};
2110
2111 vlv_merge_wm(dev_priv, &new_wm);
2112
2113 if (memcmp(old_wm, &new_wm, sizeof(new_wm)) == 0)
2114 return;
2115
2116 if (is_disabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_DDR_DVFS))
2117 chv_set_memory_dvfs(dev_priv, false);
2118
2119 if (is_disabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_PM5))
2120 chv_set_memory_pm5(dev_priv, false);
2121
2122 if (is_disabling(old_wm->cxsr, new_wm.cxsr, true))
2123 _intel_set_memory_cxsr(dev_priv, false);
2124
2125 vlv_write_wm_values(dev_priv, &new_wm);
2126
2127 if (is_enabling(old_wm->cxsr, new_wm.cxsr, true))
2128 _intel_set_memory_cxsr(dev_priv, true);
2129
2130 if (is_enabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_PM5))
2131 chv_set_memory_pm5(dev_priv, true);
2132
2133 if (is_enabling(old_wm->level, new_wm.level, VLV_WM_LEVEL_DDR_DVFS))
2134 chv_set_memory_dvfs(dev_priv, true);
2135
2136 *old_wm = new_wm;
2137 }
2138
vlv_initial_watermarks(struct intel_atomic_state * state,struct intel_crtc_state * crtc_state)2139 static void vlv_initial_watermarks(struct intel_atomic_state *state,
2140 struct intel_crtc_state *crtc_state)
2141 {
2142 struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
2143 struct intel_crtc *crtc = to_intel_crtc(crtc_state->base.crtc);
2144
2145 mutex_lock(&dev_priv->wm.wm_mutex);
2146 crtc->wm.active.vlv = crtc_state->wm.vlv.intermediate;
2147 vlv_program_watermarks(dev_priv);
2148 mutex_unlock(&dev_priv->wm.wm_mutex);
2149 }
2150
vlv_optimize_watermarks(struct intel_atomic_state * state,struct intel_crtc_state * crtc_state)2151 static void vlv_optimize_watermarks(struct intel_atomic_state *state,
2152 struct intel_crtc_state *crtc_state)
2153 {
2154 struct drm_i915_private *dev_priv = to_i915(crtc_state->base.crtc->dev);
2155 struct intel_crtc *intel_crtc = to_intel_crtc(crtc_state->base.crtc);
2156
2157 if (!crtc_state->wm.need_postvbl_update)
2158 return;
2159
2160 mutex_lock(&dev_priv->wm.wm_mutex);
2161 intel_crtc->wm.active.vlv = crtc_state->wm.vlv.optimal;
2162 vlv_program_watermarks(dev_priv);
2163 mutex_unlock(&dev_priv->wm.wm_mutex);
2164 }
2165
i965_update_wm(struct intel_crtc * unused_crtc)2166 static void i965_update_wm(struct intel_crtc *unused_crtc)
2167 {
2168 struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
2169 struct intel_crtc *crtc;
2170 int srwm = 1;
2171 int cursor_sr = 16;
2172 bool cxsr_enabled;
2173
2174 /* Calc sr entries for one plane configs */
2175 crtc = single_enabled_crtc(dev_priv);
2176 if (crtc) {
2177 /* self-refresh has much higher latency */
2178 static const int sr_latency_ns = 12000;
2179 const struct drm_display_mode *adjusted_mode =
2180 &crtc->config->base.adjusted_mode;
2181 const struct drm_framebuffer *fb =
2182 crtc->base.primary->state->fb;
2183 int clock = adjusted_mode->crtc_clock;
2184 int htotal = adjusted_mode->crtc_htotal;
2185 int hdisplay = crtc->config->pipe_src_w;
2186 int cpp = fb->format->cpp[0];
2187 int entries;
2188
2189 entries = intel_wm_method2(clock, htotal,
2190 hdisplay, cpp, sr_latency_ns / 100);
2191 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
2192 srwm = I965_FIFO_SIZE - entries;
2193 if (srwm < 0)
2194 srwm = 1;
2195 srwm &= 0x1ff;
2196 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
2197 entries, srwm);
2198
2199 entries = intel_wm_method2(clock, htotal,
2200 crtc->base.cursor->state->crtc_w, 4,
2201 sr_latency_ns / 100);
2202 entries = DIV_ROUND_UP(entries,
2203 i965_cursor_wm_info.cacheline_size) +
2204 i965_cursor_wm_info.guard_size;
2205
2206 cursor_sr = i965_cursor_wm_info.fifo_size - entries;
2207 if (cursor_sr > i965_cursor_wm_info.max_wm)
2208 cursor_sr = i965_cursor_wm_info.max_wm;
2209
2210 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
2211 "cursor %d\n", srwm, cursor_sr);
2212
2213 cxsr_enabled = true;
2214 } else {
2215 cxsr_enabled = false;
2216 /* Turn off self refresh if both pipes are enabled */
2217 intel_set_memory_cxsr(dev_priv, false);
2218 }
2219
2220 DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
2221 srwm);
2222
2223 /* 965 has limitations... */
2224 I915_WRITE(DSPFW1, FW_WM(srwm, SR) |
2225 FW_WM(8, CURSORB) |
2226 FW_WM(8, PLANEB) |
2227 FW_WM(8, PLANEA));
2228 I915_WRITE(DSPFW2, FW_WM(8, CURSORA) |
2229 FW_WM(8, PLANEC_OLD));
2230 /* update cursor SR watermark */
2231 I915_WRITE(DSPFW3, FW_WM(cursor_sr, CURSOR_SR));
2232
2233 if (cxsr_enabled)
2234 intel_set_memory_cxsr(dev_priv, true);
2235 }
2236
2237 #undef FW_WM
2238
i9xx_update_wm(struct intel_crtc * unused_crtc)2239 static void i9xx_update_wm(struct intel_crtc *unused_crtc)
2240 {
2241 struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
2242 const struct intel_watermark_params *wm_info;
2243 uint32_t fwater_lo;
2244 uint32_t fwater_hi;
2245 int cwm, srwm = 1;
2246 int fifo_size;
2247 int planea_wm, planeb_wm;
2248 struct intel_crtc *crtc, *enabled = NULL;
2249
2250 if (IS_I945GM(dev_priv))
2251 wm_info = &i945_wm_info;
2252 else if (!IS_GEN2(dev_priv))
2253 wm_info = &i915_wm_info;
2254 else
2255 wm_info = &i830_a_wm_info;
2256
2257 fifo_size = dev_priv->display.get_fifo_size(dev_priv, 0);
2258 crtc = intel_get_crtc_for_plane(dev_priv, 0);
2259 if (intel_crtc_active(crtc)) {
2260 const struct drm_display_mode *adjusted_mode =
2261 &crtc->config->base.adjusted_mode;
2262 const struct drm_framebuffer *fb =
2263 crtc->base.primary->state->fb;
2264 int cpp;
2265
2266 if (IS_GEN2(dev_priv))
2267 cpp = 4;
2268 else
2269 cpp = fb->format->cpp[0];
2270
2271 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
2272 wm_info, fifo_size, cpp,
2273 pessimal_latency_ns);
2274 enabled = crtc;
2275 } else {
2276 planea_wm = fifo_size - wm_info->guard_size;
2277 if (planea_wm > (long)wm_info->max_wm)
2278 planea_wm = wm_info->max_wm;
2279 }
2280
2281 if (IS_GEN2(dev_priv))
2282 wm_info = &i830_bc_wm_info;
2283
2284 fifo_size = dev_priv->display.get_fifo_size(dev_priv, 1);
2285 crtc = intel_get_crtc_for_plane(dev_priv, 1);
2286 if (intel_crtc_active(crtc)) {
2287 const struct drm_display_mode *adjusted_mode =
2288 &crtc->config->base.adjusted_mode;
2289 const struct drm_framebuffer *fb =
2290 crtc->base.primary->state->fb;
2291 int cpp;
2292
2293 if (IS_GEN2(dev_priv))
2294 cpp = 4;
2295 else
2296 cpp = fb->format->cpp[0];
2297
2298 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
2299 wm_info, fifo_size, cpp,
2300 pessimal_latency_ns);
2301 if (enabled == NULL)
2302 enabled = crtc;
2303 else
2304 enabled = NULL;
2305 } else {
2306 planeb_wm = fifo_size - wm_info->guard_size;
2307 if (planeb_wm > (long)wm_info->max_wm)
2308 planeb_wm = wm_info->max_wm;
2309 }
2310
2311 DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
2312
2313 if (IS_I915GM(dev_priv) && enabled) {
2314 struct drm_i915_gem_object *obj;
2315
2316 obj = intel_fb_obj(enabled->base.primary->state->fb);
2317
2318 /* self-refresh seems busted with untiled */
2319 if (!i915_gem_object_is_tiled(obj))
2320 enabled = NULL;
2321 }
2322
2323 /*
2324 * Overlay gets an aggressive default since video jitter is bad.
2325 */
2326 cwm = 2;
2327
2328 /* Play safe and disable self-refresh before adjusting watermarks. */
2329 intel_set_memory_cxsr(dev_priv, false);
2330
2331 /* Calc sr entries for one plane configs */
2332 if (HAS_FW_BLC(dev_priv) && enabled) {
2333 /* self-refresh has much higher latency */
2334 static const int sr_latency_ns = 6000;
2335 const struct drm_display_mode *adjusted_mode =
2336 &enabled->config->base.adjusted_mode;
2337 const struct drm_framebuffer *fb =
2338 enabled->base.primary->state->fb;
2339 int clock = adjusted_mode->crtc_clock;
2340 int htotal = adjusted_mode->crtc_htotal;
2341 int hdisplay = enabled->config->pipe_src_w;
2342 int cpp;
2343 int entries;
2344
2345 if (IS_I915GM(dev_priv) || IS_I945GM(dev_priv))
2346 cpp = 4;
2347 else
2348 cpp = fb->format->cpp[0];
2349
2350 entries = intel_wm_method2(clock, htotal, hdisplay, cpp,
2351 sr_latency_ns / 100);
2352 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
2353 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
2354 srwm = wm_info->fifo_size - entries;
2355 if (srwm < 0)
2356 srwm = 1;
2357
2358 if (IS_I945G(dev_priv) || IS_I945GM(dev_priv))
2359 I915_WRITE(FW_BLC_SELF,
2360 FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
2361 else
2362 I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
2363 }
2364
2365 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
2366 planea_wm, planeb_wm, cwm, srwm);
2367
2368 fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
2369 fwater_hi = (cwm & 0x1f);
2370
2371 /* Set request length to 8 cachelines per fetch */
2372 fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
2373 fwater_hi = fwater_hi | (1 << 8);
2374
2375 I915_WRITE(FW_BLC, fwater_lo);
2376 I915_WRITE(FW_BLC2, fwater_hi);
2377
2378 if (enabled)
2379 intel_set_memory_cxsr(dev_priv, true);
2380 }
2381
i845_update_wm(struct intel_crtc * unused_crtc)2382 static void i845_update_wm(struct intel_crtc *unused_crtc)
2383 {
2384 struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
2385 struct intel_crtc *crtc;
2386 const struct drm_display_mode *adjusted_mode;
2387 uint32_t fwater_lo;
2388 int planea_wm;
2389
2390 crtc = single_enabled_crtc(dev_priv);
2391 if (crtc == NULL)
2392 return;
2393
2394 adjusted_mode = &crtc->config->base.adjusted_mode;
2395 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
2396 &i845_wm_info,
2397 dev_priv->display.get_fifo_size(dev_priv, 0),
2398 4, pessimal_latency_ns);
2399 fwater_lo = I915_READ(FW_BLC) & ~0xfff;
2400 fwater_lo |= (3<<8) | planea_wm;
2401
2402 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
2403
2404 I915_WRITE(FW_BLC, fwater_lo);
2405 }
2406
2407 /* latency must be in 0.1us units. */
ilk_wm_method1(unsigned int pixel_rate,unsigned int cpp,unsigned int latency)2408 static unsigned int ilk_wm_method1(unsigned int pixel_rate,
2409 unsigned int cpp,
2410 unsigned int latency)
2411 {
2412 unsigned int ret;
2413
2414 ret = intel_wm_method1(pixel_rate, cpp, latency);
2415 ret = DIV_ROUND_UP(ret, 64) + 2;
2416
2417 return ret;
2418 }
2419
2420 /* latency must be in 0.1us units. */
ilk_wm_method2(unsigned int pixel_rate,unsigned int htotal,unsigned int width,unsigned int cpp,unsigned int latency)2421 static unsigned int ilk_wm_method2(unsigned int pixel_rate,
2422 unsigned int htotal,
2423 unsigned int width,
2424 unsigned int cpp,
2425 unsigned int latency)
2426 {
2427 unsigned int ret;
2428
2429 ret = intel_wm_method2(pixel_rate, htotal,
2430 width, cpp, latency);
2431 ret = DIV_ROUND_UP(ret, 64) + 2;
2432
2433 return ret;
2434 }
2435
ilk_wm_fbc(uint32_t pri_val,uint32_t horiz_pixels,uint8_t cpp)2436 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
2437 uint8_t cpp)
2438 {
2439 /*
2440 * Neither of these should be possible since this function shouldn't be
2441 * called if the CRTC is off or the plane is invisible. But let's be
2442 * extra paranoid to avoid a potential divide-by-zero if we screw up
2443 * elsewhere in the driver.
2444 */
2445 if (WARN_ON(!cpp))
2446 return 0;
2447 if (WARN_ON(!horiz_pixels))
2448 return 0;
2449
2450 return DIV_ROUND_UP(pri_val * 64, horiz_pixels * cpp) + 2;
2451 }
2452
2453 struct ilk_wm_maximums {
2454 uint16_t pri;
2455 uint16_t spr;
2456 uint16_t cur;
2457 uint16_t fbc;
2458 };
2459
2460 /*
2461 * For both WM_PIPE and WM_LP.
2462 * mem_value must be in 0.1us units.
2463 */
ilk_compute_pri_wm(const struct intel_crtc_state * cstate,const struct intel_plane_state * pstate,uint32_t mem_value,bool is_lp)2464 static uint32_t ilk_compute_pri_wm(const struct intel_crtc_state *cstate,
2465 const struct intel_plane_state *pstate,
2466 uint32_t mem_value,
2467 bool is_lp)
2468 {
2469 uint32_t method1, method2;
2470 int cpp;
2471
2472 if (!intel_wm_plane_visible(cstate, pstate))
2473 return 0;
2474
2475 cpp = pstate->base.fb->format->cpp[0];
2476
2477 method1 = ilk_wm_method1(cstate->pixel_rate, cpp, mem_value);
2478
2479 if (!is_lp)
2480 return method1;
2481
2482 method2 = ilk_wm_method2(cstate->pixel_rate,
2483 cstate->base.adjusted_mode.crtc_htotal,
2484 drm_rect_width(&pstate->base.dst),
2485 cpp, mem_value);
2486
2487 return min(method1, method2);
2488 }
2489
2490 /*
2491 * For both WM_PIPE and WM_LP.
2492 * mem_value must be in 0.1us units.
2493 */
ilk_compute_spr_wm(const struct intel_crtc_state * cstate,const struct intel_plane_state * pstate,uint32_t mem_value)2494 static uint32_t ilk_compute_spr_wm(const struct intel_crtc_state *cstate,
2495 const struct intel_plane_state *pstate,
2496 uint32_t mem_value)
2497 {
2498 uint32_t method1, method2;
2499 int cpp;
2500
2501 if (!intel_wm_plane_visible(cstate, pstate))
2502 return 0;
2503
2504 cpp = pstate->base.fb->format->cpp[0];
2505
2506 method1 = ilk_wm_method1(cstate->pixel_rate, cpp, mem_value);
2507 method2 = ilk_wm_method2(cstate->pixel_rate,
2508 cstate->base.adjusted_mode.crtc_htotal,
2509 drm_rect_width(&pstate->base.dst),
2510 cpp, mem_value);
2511 return min(method1, method2);
2512 }
2513
2514 /*
2515 * For both WM_PIPE and WM_LP.
2516 * mem_value must be in 0.1us units.
2517 */
ilk_compute_cur_wm(const struct intel_crtc_state * cstate,const struct intel_plane_state * pstate,uint32_t mem_value)2518 static uint32_t ilk_compute_cur_wm(const struct intel_crtc_state *cstate,
2519 const struct intel_plane_state *pstate,
2520 uint32_t mem_value)
2521 {
2522 int cpp;
2523
2524 if (!intel_wm_plane_visible(cstate, pstate))
2525 return 0;
2526
2527 cpp = pstate->base.fb->format->cpp[0];
2528
2529 return ilk_wm_method2(cstate->pixel_rate,
2530 cstate->base.adjusted_mode.crtc_htotal,
2531 pstate->base.crtc_w, cpp, mem_value);
2532 }
2533
2534 /* Only for WM_LP. */
ilk_compute_fbc_wm(const struct intel_crtc_state * cstate,const struct intel_plane_state * pstate,uint32_t pri_val)2535 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state *cstate,
2536 const struct intel_plane_state *pstate,
2537 uint32_t pri_val)
2538 {
2539 int cpp;
2540
2541 if (!intel_wm_plane_visible(cstate, pstate))
2542 return 0;
2543
2544 cpp = pstate->base.fb->format->cpp[0];
2545
2546 return ilk_wm_fbc(pri_val, drm_rect_width(&pstate->base.dst), cpp);
2547 }
2548
2549 static unsigned int
ilk_display_fifo_size(const struct drm_i915_private * dev_priv)2550 ilk_display_fifo_size(const struct drm_i915_private *dev_priv)
2551 {
2552 if (INTEL_GEN(dev_priv) >= 8)
2553 return 3072;
2554 else if (INTEL_GEN(dev_priv) >= 7)
2555 return 768;
2556 else
2557 return 512;
2558 }
2559
2560 static unsigned int
ilk_plane_wm_reg_max(const struct drm_i915_private * dev_priv,int level,bool is_sprite)2561 ilk_plane_wm_reg_max(const struct drm_i915_private *dev_priv,
2562 int level, bool is_sprite)
2563 {
2564 if (INTEL_GEN(dev_priv) >= 8)
2565 /* BDW primary/sprite plane watermarks */
2566 return level == 0 ? 255 : 2047;
2567 else if (INTEL_GEN(dev_priv) >= 7)
2568 /* IVB/HSW primary/sprite plane watermarks */
2569 return level == 0 ? 127 : 1023;
2570 else if (!is_sprite)
2571 /* ILK/SNB primary plane watermarks */
2572 return level == 0 ? 127 : 511;
2573 else
2574 /* ILK/SNB sprite plane watermarks */
2575 return level == 0 ? 63 : 255;
2576 }
2577
2578 static unsigned int
ilk_cursor_wm_reg_max(const struct drm_i915_private * dev_priv,int level)2579 ilk_cursor_wm_reg_max(const struct drm_i915_private *dev_priv, int level)
2580 {
2581 if (INTEL_GEN(dev_priv) >= 7)
2582 return level == 0 ? 63 : 255;
2583 else
2584 return level == 0 ? 31 : 63;
2585 }
2586
ilk_fbc_wm_reg_max(const struct drm_i915_private * dev_priv)2587 static unsigned int ilk_fbc_wm_reg_max(const struct drm_i915_private *dev_priv)
2588 {
2589 if (INTEL_GEN(dev_priv) >= 8)
2590 return 31;
2591 else
2592 return 15;
2593 }
2594
2595 /* Calculate the maximum primary/sprite plane watermark */
ilk_plane_wm_max(const struct drm_device * dev,int level,const struct intel_wm_config * config,enum intel_ddb_partitioning ddb_partitioning,bool is_sprite)2596 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
2597 int level,
2598 const struct intel_wm_config *config,
2599 enum intel_ddb_partitioning ddb_partitioning,
2600 bool is_sprite)
2601 {
2602 struct drm_i915_private *dev_priv = to_i915(dev);
2603 unsigned int fifo_size = ilk_display_fifo_size(dev_priv);
2604
2605 /* if sprites aren't enabled, sprites get nothing */
2606 if (is_sprite && !config->sprites_enabled)
2607 return 0;
2608
2609 /* HSW allows LP1+ watermarks even with multiple pipes */
2610 if (level == 0 || config->num_pipes_active > 1) {
2611 fifo_size /= INTEL_INFO(dev_priv)->num_pipes;
2612
2613 /*
2614 * For some reason the non self refresh
2615 * FIFO size is only half of the self
2616 * refresh FIFO size on ILK/SNB.
2617 */
2618 if (INTEL_GEN(dev_priv) <= 6)
2619 fifo_size /= 2;
2620 }
2621
2622 if (config->sprites_enabled) {
2623 /* level 0 is always calculated with 1:1 split */
2624 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
2625 if (is_sprite)
2626 fifo_size *= 5;
2627 fifo_size /= 6;
2628 } else {
2629 fifo_size /= 2;
2630 }
2631 }
2632
2633 /* clamp to max that the registers can hold */
2634 return min(fifo_size, ilk_plane_wm_reg_max(dev_priv, level, is_sprite));
2635 }
2636
2637 /* Calculate the maximum cursor plane watermark */
ilk_cursor_wm_max(const struct drm_device * dev,int level,const struct intel_wm_config * config)2638 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
2639 int level,
2640 const struct intel_wm_config *config)
2641 {
2642 /* HSW LP1+ watermarks w/ multiple pipes */
2643 if (level > 0 && config->num_pipes_active > 1)
2644 return 64;
2645
2646 /* otherwise just report max that registers can hold */
2647 return ilk_cursor_wm_reg_max(to_i915(dev), level);
2648 }
2649
ilk_compute_wm_maximums(const struct drm_device * dev,int level,const struct intel_wm_config * config,enum intel_ddb_partitioning ddb_partitioning,struct ilk_wm_maximums * max)2650 static void ilk_compute_wm_maximums(const struct drm_device *dev,
2651 int level,
2652 const struct intel_wm_config *config,
2653 enum intel_ddb_partitioning ddb_partitioning,
2654 struct ilk_wm_maximums *max)
2655 {
2656 max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
2657 max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
2658 max->cur = ilk_cursor_wm_max(dev, level, config);
2659 max->fbc = ilk_fbc_wm_reg_max(to_i915(dev));
2660 }
2661
ilk_compute_wm_reg_maximums(const struct drm_i915_private * dev_priv,int level,struct ilk_wm_maximums * max)2662 static void ilk_compute_wm_reg_maximums(const struct drm_i915_private *dev_priv,
2663 int level,
2664 struct ilk_wm_maximums *max)
2665 {
2666 max->pri = ilk_plane_wm_reg_max(dev_priv, level, false);
2667 max->spr = ilk_plane_wm_reg_max(dev_priv, level, true);
2668 max->cur = ilk_cursor_wm_reg_max(dev_priv, level);
2669 max->fbc = ilk_fbc_wm_reg_max(dev_priv);
2670 }
2671
ilk_validate_wm_level(int level,const struct ilk_wm_maximums * max,struct intel_wm_level * result)2672 static bool ilk_validate_wm_level(int level,
2673 const struct ilk_wm_maximums *max,
2674 struct intel_wm_level *result)
2675 {
2676 bool ret;
2677
2678 /* already determined to be invalid? */
2679 if (!result->enable)
2680 return false;
2681
2682 result->enable = result->pri_val <= max->pri &&
2683 result->spr_val <= max->spr &&
2684 result->cur_val <= max->cur;
2685
2686 ret = result->enable;
2687
2688 /*
2689 * HACK until we can pre-compute everything,
2690 * and thus fail gracefully if LP0 watermarks
2691 * are exceeded...
2692 */
2693 if (level == 0 && !result->enable) {
2694 if (result->pri_val > max->pri)
2695 DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
2696 level, result->pri_val, max->pri);
2697 if (result->spr_val > max->spr)
2698 DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
2699 level, result->spr_val, max->spr);
2700 if (result->cur_val > max->cur)
2701 DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
2702 level, result->cur_val, max->cur);
2703
2704 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
2705 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
2706 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
2707 result->enable = true;
2708 }
2709
2710 return ret;
2711 }
2712
ilk_compute_wm_level(const struct drm_i915_private * dev_priv,const struct intel_crtc * intel_crtc,int level,struct intel_crtc_state * cstate,const struct intel_plane_state * pristate,const struct intel_plane_state * sprstate,const struct intel_plane_state * curstate,struct intel_wm_level * result)2713 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
2714 const struct intel_crtc *intel_crtc,
2715 int level,
2716 struct intel_crtc_state *cstate,
2717 const struct intel_plane_state *pristate,
2718 const struct intel_plane_state *sprstate,
2719 const struct intel_plane_state *curstate,
2720 struct intel_wm_level *result)
2721 {
2722 uint16_t pri_latency = dev_priv->wm.pri_latency[level];
2723 uint16_t spr_latency = dev_priv->wm.spr_latency[level];
2724 uint16_t cur_latency = dev_priv->wm.cur_latency[level];
2725
2726 /* WM1+ latency values stored in 0.5us units */
2727 if (level > 0) {
2728 pri_latency *= 5;
2729 spr_latency *= 5;
2730 cur_latency *= 5;
2731 }
2732
2733 if (pristate) {
2734 result->pri_val = ilk_compute_pri_wm(cstate, pristate,
2735 pri_latency, level);
2736 result->fbc_val = ilk_compute_fbc_wm(cstate, pristate, result->pri_val);
2737 }
2738
2739 if (sprstate)
2740 result->spr_val = ilk_compute_spr_wm(cstate, sprstate, spr_latency);
2741
2742 if (curstate)
2743 result->cur_val = ilk_compute_cur_wm(cstate, curstate, cur_latency);
2744
2745 result->enable = true;
2746 }
2747
2748 static uint32_t
hsw_compute_linetime_wm(const struct intel_crtc_state * cstate)2749 hsw_compute_linetime_wm(const struct intel_crtc_state *cstate)
2750 {
2751 const struct intel_atomic_state *intel_state =
2752 to_intel_atomic_state(cstate->base.state);
2753 const struct drm_display_mode *adjusted_mode =
2754 &cstate->base.adjusted_mode;
2755 u32 linetime, ips_linetime;
2756
2757 if (!cstate->base.active)
2758 return 0;
2759 if (WARN_ON(adjusted_mode->crtc_clock == 0))
2760 return 0;
2761 if (WARN_ON(intel_state->cdclk.logical.cdclk == 0))
2762 return 0;
2763
2764 /* The WM are computed with base on how long it takes to fill a single
2765 * row at the given clock rate, multiplied by 8.
2766 * */
2767 linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2768 adjusted_mode->crtc_clock);
2769 ips_linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2770 intel_state->cdclk.logical.cdclk);
2771
2772 return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2773 PIPE_WM_LINETIME_TIME(linetime);
2774 }
2775
intel_read_wm_latency(struct drm_i915_private * dev_priv,uint16_t wm[8])2776 static void intel_read_wm_latency(struct drm_i915_private *dev_priv,
2777 uint16_t wm[8])
2778 {
2779 if (INTEL_GEN(dev_priv) >= 9) {
2780 uint32_t val;
2781 int ret, i;
2782 int level, max_level = ilk_wm_max_level(dev_priv);
2783
2784 /* read the first set of memory latencies[0:3] */
2785 val = 0; /* data0 to be programmed to 0 for first set */
2786 mutex_lock(&dev_priv->pcu_lock);
2787 ret = sandybridge_pcode_read(dev_priv,
2788 GEN9_PCODE_READ_MEM_LATENCY,
2789 &val);
2790 mutex_unlock(&dev_priv->pcu_lock);
2791
2792 if (ret) {
2793 DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2794 return;
2795 }
2796
2797 wm[0] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2798 wm[1] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2799 GEN9_MEM_LATENCY_LEVEL_MASK;
2800 wm[2] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2801 GEN9_MEM_LATENCY_LEVEL_MASK;
2802 wm[3] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2803 GEN9_MEM_LATENCY_LEVEL_MASK;
2804
2805 /* read the second set of memory latencies[4:7] */
2806 val = 1; /* data0 to be programmed to 1 for second set */
2807 mutex_lock(&dev_priv->pcu_lock);
2808 ret = sandybridge_pcode_read(dev_priv,
2809 GEN9_PCODE_READ_MEM_LATENCY,
2810 &val);
2811 mutex_unlock(&dev_priv->pcu_lock);
2812 if (ret) {
2813 DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2814 return;
2815 }
2816
2817 wm[4] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2818 wm[5] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2819 GEN9_MEM_LATENCY_LEVEL_MASK;
2820 wm[6] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2821 GEN9_MEM_LATENCY_LEVEL_MASK;
2822 wm[7] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2823 GEN9_MEM_LATENCY_LEVEL_MASK;
2824
2825 /*
2826 * If a level n (n > 1) has a 0us latency, all levels m (m >= n)
2827 * need to be disabled. We make sure to sanitize the values out
2828 * of the punit to satisfy this requirement.
2829 */
2830 for (level = 1; level <= max_level; level++) {
2831 if (wm[level] == 0) {
2832 for (i = level + 1; i <= max_level; i++)
2833 wm[i] = 0;
2834 break;
2835 }
2836 }
2837
2838 /*
2839 * WaWmMemoryReadLatency:skl+,glk
2840 *
2841 * punit doesn't take into account the read latency so we need
2842 * to add 2us to the various latency levels we retrieve from the
2843 * punit when level 0 response data us 0us.
2844 */
2845 if (wm[0] == 0) {
2846 wm[0] += 2;
2847 for (level = 1; level <= max_level; level++) {
2848 if (wm[level] == 0)
2849 break;
2850 wm[level] += 2;
2851 }
2852 }
2853
2854 } else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
2855 uint64_t sskpd = I915_READ64(MCH_SSKPD);
2856
2857 wm[0] = (sskpd >> 56) & 0xFF;
2858 if (wm[0] == 0)
2859 wm[0] = sskpd & 0xF;
2860 wm[1] = (sskpd >> 4) & 0xFF;
2861 wm[2] = (sskpd >> 12) & 0xFF;
2862 wm[3] = (sskpd >> 20) & 0x1FF;
2863 wm[4] = (sskpd >> 32) & 0x1FF;
2864 } else if (INTEL_GEN(dev_priv) >= 6) {
2865 uint32_t sskpd = I915_READ(MCH_SSKPD);
2866
2867 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2868 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2869 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2870 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2871 } else if (INTEL_GEN(dev_priv) >= 5) {
2872 uint32_t mltr = I915_READ(MLTR_ILK);
2873
2874 /* ILK primary LP0 latency is 700 ns */
2875 wm[0] = 7;
2876 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2877 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2878 } else {
2879 MISSING_CASE(INTEL_DEVID(dev_priv));
2880 }
2881 }
2882
intel_fixup_spr_wm_latency(struct drm_i915_private * dev_priv,uint16_t wm[5])2883 static void intel_fixup_spr_wm_latency(struct drm_i915_private *dev_priv,
2884 uint16_t wm[5])
2885 {
2886 /* ILK sprite LP0 latency is 1300 ns */
2887 if (IS_GEN5(dev_priv))
2888 wm[0] = 13;
2889 }
2890
intel_fixup_cur_wm_latency(struct drm_i915_private * dev_priv,uint16_t wm[5])2891 static void intel_fixup_cur_wm_latency(struct drm_i915_private *dev_priv,
2892 uint16_t wm[5])
2893 {
2894 /* ILK cursor LP0 latency is 1300 ns */
2895 if (IS_GEN5(dev_priv))
2896 wm[0] = 13;
2897
2898 /* WaDoubleCursorLP3Latency:ivb */
2899 if (IS_IVYBRIDGE(dev_priv))
2900 wm[3] *= 2;
2901 }
2902
ilk_wm_max_level(const struct drm_i915_private * dev_priv)2903 int ilk_wm_max_level(const struct drm_i915_private *dev_priv)
2904 {
2905 /* how many WM levels are we expecting */
2906 if (INTEL_GEN(dev_priv) >= 9)
2907 return 7;
2908 else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
2909 return 4;
2910 else if (INTEL_GEN(dev_priv) >= 6)
2911 return 3;
2912 else
2913 return 2;
2914 }
2915
intel_print_wm_latency(struct drm_i915_private * dev_priv,const char * name,const uint16_t wm[8])2916 static void intel_print_wm_latency(struct drm_i915_private *dev_priv,
2917 const char *name,
2918 const uint16_t wm[8])
2919 {
2920 int level, max_level = ilk_wm_max_level(dev_priv);
2921
2922 for (level = 0; level <= max_level; level++) {
2923 unsigned int latency = wm[level];
2924
2925 if (latency == 0) {
2926 DRM_ERROR("%s WM%d latency not provided\n",
2927 name, level);
2928 continue;
2929 }
2930
2931 /*
2932 * - latencies are in us on gen9.
2933 * - before then, WM1+ latency values are in 0.5us units
2934 */
2935 if (INTEL_GEN(dev_priv) >= 9)
2936 latency *= 10;
2937 else if (level > 0)
2938 latency *= 5;
2939
2940 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2941 name, level, wm[level],
2942 latency / 10, latency % 10);
2943 }
2944 }
2945
ilk_increase_wm_latency(struct drm_i915_private * dev_priv,uint16_t wm[5],uint16_t min)2946 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
2947 uint16_t wm[5], uint16_t min)
2948 {
2949 int level, max_level = ilk_wm_max_level(dev_priv);
2950
2951 if (wm[0] >= min)
2952 return false;
2953
2954 wm[0] = max(wm[0], min);
2955 for (level = 1; level <= max_level; level++)
2956 wm[level] = max_t(uint16_t, wm[level], DIV_ROUND_UP(min, 5));
2957
2958 return true;
2959 }
2960
snb_wm_latency_quirk(struct drm_i915_private * dev_priv)2961 static void snb_wm_latency_quirk(struct drm_i915_private *dev_priv)
2962 {
2963 bool changed;
2964
2965 /*
2966 * The BIOS provided WM memory latency values are often
2967 * inadequate for high resolution displays. Adjust them.
2968 */
2969 changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
2970 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
2971 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
2972
2973 if (!changed)
2974 return;
2975
2976 DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2977 intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
2978 intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
2979 intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
2980 }
2981
ilk_setup_wm_latency(struct drm_i915_private * dev_priv)2982 static void ilk_setup_wm_latency(struct drm_i915_private *dev_priv)
2983 {
2984 intel_read_wm_latency(dev_priv, dev_priv->wm.pri_latency);
2985
2986 memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2987 sizeof(dev_priv->wm.pri_latency));
2988 memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2989 sizeof(dev_priv->wm.pri_latency));
2990
2991 intel_fixup_spr_wm_latency(dev_priv, dev_priv->wm.spr_latency);
2992 intel_fixup_cur_wm_latency(dev_priv, dev_priv->wm.cur_latency);
2993
2994 intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
2995 intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
2996 intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
2997
2998 if (IS_GEN6(dev_priv))
2999 snb_wm_latency_quirk(dev_priv);
3000 }
3001
skl_setup_wm_latency(struct drm_i915_private * dev_priv)3002 static void skl_setup_wm_latency(struct drm_i915_private *dev_priv)
3003 {
3004 intel_read_wm_latency(dev_priv, dev_priv->wm.skl_latency);
3005 intel_print_wm_latency(dev_priv, "Gen9 Plane", dev_priv->wm.skl_latency);
3006 }
3007
ilk_validate_pipe_wm(struct drm_device * dev,struct intel_pipe_wm * pipe_wm)3008 static bool ilk_validate_pipe_wm(struct drm_device *dev,
3009 struct intel_pipe_wm *pipe_wm)
3010 {
3011 /* LP0 watermark maximums depend on this pipe alone */
3012 const struct intel_wm_config config = {
3013 .num_pipes_active = 1,
3014 .sprites_enabled = pipe_wm->sprites_enabled,
3015 .sprites_scaled = pipe_wm->sprites_scaled,
3016 };
3017 struct ilk_wm_maximums max;
3018
3019 /* LP0 watermarks always use 1/2 DDB partitioning */
3020 ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
3021
3022 /* At least LP0 must be valid */
3023 if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0])) {
3024 DRM_DEBUG_KMS("LP0 watermark invalid\n");
3025 return false;
3026 }
3027
3028 return true;
3029 }
3030
3031 /* Compute new watermarks for the pipe */
ilk_compute_pipe_wm(struct intel_crtc_state * cstate)3032 static int ilk_compute_pipe_wm(struct intel_crtc_state *cstate)
3033 {
3034 struct drm_atomic_state *state = cstate->base.state;
3035 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
3036 struct intel_pipe_wm *pipe_wm;
3037 struct drm_device *dev = state->dev;
3038 const struct drm_i915_private *dev_priv = to_i915(dev);
3039 struct drm_plane *plane;
3040 const struct drm_plane_state *plane_state;
3041 const struct intel_plane_state *pristate = NULL;
3042 const struct intel_plane_state *sprstate = NULL;
3043 const struct intel_plane_state *curstate = NULL;
3044 int level, max_level = ilk_wm_max_level(dev_priv), usable_level;
3045 struct ilk_wm_maximums max;
3046
3047 pipe_wm = &cstate->wm.ilk.optimal;
3048
3049 drm_atomic_crtc_state_for_each_plane_state(plane, plane_state, &cstate->base) {
3050 const struct intel_plane_state *ps = to_intel_plane_state(plane_state);
3051
3052 if (plane->type == DRM_PLANE_TYPE_PRIMARY)
3053 pristate = ps;
3054 else if (plane->type == DRM_PLANE_TYPE_OVERLAY)
3055 sprstate = ps;
3056 else if (plane->type == DRM_PLANE_TYPE_CURSOR)
3057 curstate = ps;
3058 }
3059
3060 pipe_wm->pipe_enabled = cstate->base.active;
3061 if (sprstate) {
3062 pipe_wm->sprites_enabled = sprstate->base.visible;
3063 pipe_wm->sprites_scaled = sprstate->base.visible &&
3064 (drm_rect_width(&sprstate->base.dst) != drm_rect_width(&sprstate->base.src) >> 16 ||
3065 drm_rect_height(&sprstate->base.dst) != drm_rect_height(&sprstate->base.src) >> 16);
3066 }
3067
3068 usable_level = max_level;
3069
3070 /* ILK/SNB: LP2+ watermarks only w/o sprites */
3071 if (INTEL_GEN(dev_priv) <= 6 && pipe_wm->sprites_enabled)
3072 usable_level = 1;
3073
3074 /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
3075 if (pipe_wm->sprites_scaled)
3076 usable_level = 0;
3077
3078 memset(&pipe_wm->wm, 0, sizeof(pipe_wm->wm));
3079 ilk_compute_wm_level(dev_priv, intel_crtc, 0, cstate,
3080 pristate, sprstate, curstate, &pipe_wm->wm[0]);
3081
3082 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
3083 pipe_wm->linetime = hsw_compute_linetime_wm(cstate);
3084
3085 if (!ilk_validate_pipe_wm(dev, pipe_wm))
3086 return -EINVAL;
3087
3088 ilk_compute_wm_reg_maximums(dev_priv, 1, &max);
3089
3090 for (level = 1; level <= usable_level; level++) {
3091 struct intel_wm_level *wm = &pipe_wm->wm[level];
3092
3093 ilk_compute_wm_level(dev_priv, intel_crtc, level, cstate,
3094 pristate, sprstate, curstate, wm);
3095
3096 /*
3097 * Disable any watermark level that exceeds the
3098 * register maximums since such watermarks are
3099 * always invalid.
3100 */
3101 if (!ilk_validate_wm_level(level, &max, wm)) {
3102 memset(wm, 0, sizeof(*wm));
3103 break;
3104 }
3105 }
3106
3107 return 0;
3108 }
3109
3110 /*
3111 * Build a set of 'intermediate' watermark values that satisfy both the old
3112 * state and the new state. These can be programmed to the hardware
3113 * immediately.
3114 */
ilk_compute_intermediate_wm(struct drm_device * dev,struct intel_crtc * intel_crtc,struct intel_crtc_state * newstate)3115 static int ilk_compute_intermediate_wm(struct drm_device *dev,
3116 struct intel_crtc *intel_crtc,
3117 struct intel_crtc_state *newstate)
3118 {
3119 struct intel_pipe_wm *a = &newstate->wm.ilk.intermediate;
3120 struct intel_atomic_state *intel_state =
3121 to_intel_atomic_state(newstate->base.state);
3122 const struct intel_crtc_state *oldstate =
3123 intel_atomic_get_old_crtc_state(intel_state, intel_crtc);
3124 const struct intel_pipe_wm *b = &oldstate->wm.ilk.optimal;
3125 int level, max_level = ilk_wm_max_level(to_i915(dev));
3126
3127 /*
3128 * Start with the final, target watermarks, then combine with the
3129 * currently active watermarks to get values that are safe both before
3130 * and after the vblank.
3131 */
3132 *a = newstate->wm.ilk.optimal;
3133 if (!newstate->base.active || drm_atomic_crtc_needs_modeset(&newstate->base))
3134 return 0;
3135
3136 a->pipe_enabled |= b->pipe_enabled;
3137 a->sprites_enabled |= b->sprites_enabled;
3138 a->sprites_scaled |= b->sprites_scaled;
3139
3140 for (level = 0; level <= max_level; level++) {
3141 struct intel_wm_level *a_wm = &a->wm[level];
3142 const struct intel_wm_level *b_wm = &b->wm[level];
3143
3144 a_wm->enable &= b_wm->enable;
3145 a_wm->pri_val = max(a_wm->pri_val, b_wm->pri_val);
3146 a_wm->spr_val = max(a_wm->spr_val, b_wm->spr_val);
3147 a_wm->cur_val = max(a_wm->cur_val, b_wm->cur_val);
3148 a_wm->fbc_val = max(a_wm->fbc_val, b_wm->fbc_val);
3149 }
3150
3151 /*
3152 * We need to make sure that these merged watermark values are
3153 * actually a valid configuration themselves. If they're not,
3154 * there's no safe way to transition from the old state to
3155 * the new state, so we need to fail the atomic transaction.
3156 */
3157 if (!ilk_validate_pipe_wm(dev, a))
3158 return -EINVAL;
3159
3160 /*
3161 * If our intermediate WM are identical to the final WM, then we can
3162 * omit the post-vblank programming; only update if it's different.
3163 */
3164 if (memcmp(a, &newstate->wm.ilk.optimal, sizeof(*a)) != 0)
3165 newstate->wm.need_postvbl_update = true;
3166
3167 return 0;
3168 }
3169
3170 /*
3171 * Merge the watermarks from all active pipes for a specific level.
3172 */
ilk_merge_wm_level(struct drm_device * dev,int level,struct intel_wm_level * ret_wm)3173 static void ilk_merge_wm_level(struct drm_device *dev,
3174 int level,
3175 struct intel_wm_level *ret_wm)
3176 {
3177 const struct intel_crtc *intel_crtc;
3178
3179 ret_wm->enable = true;
3180
3181 for_each_intel_crtc(dev, intel_crtc) {
3182 const struct intel_pipe_wm *active = &intel_crtc->wm.active.ilk;
3183 const struct intel_wm_level *wm = &active->wm[level];
3184
3185 if (!active->pipe_enabled)
3186 continue;
3187
3188 /*
3189 * The watermark values may have been used in the past,
3190 * so we must maintain them in the registers for some
3191 * time even if the level is now disabled.
3192 */
3193 if (!wm->enable)
3194 ret_wm->enable = false;
3195
3196 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
3197 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
3198 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
3199 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
3200 }
3201 }
3202
3203 /*
3204 * Merge all low power watermarks for all active pipes.
3205 */
ilk_wm_merge(struct drm_device * dev,const struct intel_wm_config * config,const struct ilk_wm_maximums * max,struct intel_pipe_wm * merged)3206 static void ilk_wm_merge(struct drm_device *dev,
3207 const struct intel_wm_config *config,
3208 const struct ilk_wm_maximums *max,
3209 struct intel_pipe_wm *merged)
3210 {
3211 struct drm_i915_private *dev_priv = to_i915(dev);
3212 int level, max_level = ilk_wm_max_level(dev_priv);
3213 int last_enabled_level = max_level;
3214
3215 /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
3216 if ((INTEL_GEN(dev_priv) <= 6 || IS_IVYBRIDGE(dev_priv)) &&
3217 config->num_pipes_active > 1)
3218 last_enabled_level = 0;
3219
3220 /* ILK: FBC WM must be disabled always */
3221 merged->fbc_wm_enabled = INTEL_GEN(dev_priv) >= 6;
3222
3223 /* merge each WM1+ level */
3224 for (level = 1; level <= max_level; level++) {
3225 struct intel_wm_level *wm = &merged->wm[level];
3226
3227 ilk_merge_wm_level(dev, level, wm);
3228
3229 if (level > last_enabled_level)
3230 wm->enable = false;
3231 else if (!ilk_validate_wm_level(level, max, wm))
3232 /* make sure all following levels get disabled */
3233 last_enabled_level = level - 1;
3234
3235 /*
3236 * The spec says it is preferred to disable
3237 * FBC WMs instead of disabling a WM level.
3238 */
3239 if (wm->fbc_val > max->fbc) {
3240 if (wm->enable)
3241 merged->fbc_wm_enabled = false;
3242 wm->fbc_val = 0;
3243 }
3244 }
3245
3246 /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
3247 /*
3248 * FIXME this is racy. FBC might get enabled later.
3249 * What we should check here is whether FBC can be
3250 * enabled sometime later.
3251 */
3252 if (IS_GEN5(dev_priv) && !merged->fbc_wm_enabled &&
3253 intel_fbc_is_active(dev_priv)) {
3254 for (level = 2; level <= max_level; level++) {
3255 struct intel_wm_level *wm = &merged->wm[level];
3256
3257 wm->enable = false;
3258 }
3259 }
3260 }
3261
ilk_wm_lp_to_level(int wm_lp,const struct intel_pipe_wm * pipe_wm)3262 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
3263 {
3264 /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
3265 return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
3266 }
3267
3268 /* The value we need to program into the WM_LPx latency field */
ilk_wm_lp_latency(struct drm_device * dev,int level)3269 static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
3270 {
3271 struct drm_i915_private *dev_priv = to_i915(dev);
3272
3273 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
3274 return 2 * level;
3275 else
3276 return dev_priv->wm.pri_latency[level];
3277 }
3278
ilk_compute_wm_results(struct drm_device * dev,const struct intel_pipe_wm * merged,enum intel_ddb_partitioning partitioning,struct ilk_wm_values * results)3279 static void ilk_compute_wm_results(struct drm_device *dev,
3280 const struct intel_pipe_wm *merged,
3281 enum intel_ddb_partitioning partitioning,
3282 struct ilk_wm_values *results)
3283 {
3284 struct drm_i915_private *dev_priv = to_i915(dev);
3285 struct intel_crtc *intel_crtc;
3286 int level, wm_lp;
3287
3288 results->enable_fbc_wm = merged->fbc_wm_enabled;
3289 results->partitioning = partitioning;
3290
3291 /* LP1+ register values */
3292 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
3293 const struct intel_wm_level *r;
3294
3295 level = ilk_wm_lp_to_level(wm_lp, merged);
3296
3297 r = &merged->wm[level];
3298
3299 /*
3300 * Maintain the watermark values even if the level is
3301 * disabled. Doing otherwise could cause underruns.
3302 */
3303 results->wm_lp[wm_lp - 1] =
3304 (ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
3305 (r->pri_val << WM1_LP_SR_SHIFT) |
3306 r->cur_val;
3307
3308 if (r->enable)
3309 results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
3310
3311 if (INTEL_GEN(dev_priv) >= 8)
3312 results->wm_lp[wm_lp - 1] |=
3313 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
3314 else
3315 results->wm_lp[wm_lp - 1] |=
3316 r->fbc_val << WM1_LP_FBC_SHIFT;
3317
3318 /*
3319 * Always set WM1S_LP_EN when spr_val != 0, even if the
3320 * level is disabled. Doing otherwise could cause underruns.
3321 */
3322 if (INTEL_GEN(dev_priv) <= 6 && r->spr_val) {
3323 WARN_ON(wm_lp != 1);
3324 results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
3325 } else
3326 results->wm_lp_spr[wm_lp - 1] = r->spr_val;
3327 }
3328
3329 /* LP0 register values */
3330 for_each_intel_crtc(dev, intel_crtc) {
3331 enum i915_pipe pipe = intel_crtc->pipe;
3332 const struct intel_wm_level *r =
3333 &intel_crtc->wm.active.ilk.wm[0];
3334
3335 if (WARN_ON(!r->enable))
3336 continue;
3337
3338 results->wm_linetime[pipe] = intel_crtc->wm.active.ilk.linetime;
3339
3340 results->wm_pipe[pipe] =
3341 (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
3342 (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
3343 r->cur_val;
3344 }
3345 }
3346
3347 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
3348 * case both are at the same level. Prefer r1 in case they're the same. */
ilk_find_best_result(struct drm_device * dev,struct intel_pipe_wm * r1,struct intel_pipe_wm * r2)3349 static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
3350 struct intel_pipe_wm *r1,
3351 struct intel_pipe_wm *r2)
3352 {
3353 int level, max_level = ilk_wm_max_level(to_i915(dev));
3354 int level1 = 0, level2 = 0;
3355
3356 for (level = 1; level <= max_level; level++) {
3357 if (r1->wm[level].enable)
3358 level1 = level;
3359 if (r2->wm[level].enable)
3360 level2 = level;
3361 }
3362
3363 if (level1 == level2) {
3364 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
3365 return r2;
3366 else
3367 return r1;
3368 } else if (level1 > level2) {
3369 return r1;
3370 } else {
3371 return r2;
3372 }
3373 }
3374
3375 /* dirty bits used to track which watermarks need changes */
3376 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
3377 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
3378 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
3379 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
3380 #define WM_DIRTY_FBC (1 << 24)
3381 #define WM_DIRTY_DDB (1 << 25)
3382
ilk_compute_wm_dirty(struct drm_i915_private * dev_priv,const struct ilk_wm_values * old,const struct ilk_wm_values * new)3383 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private *dev_priv,
3384 const struct ilk_wm_values *old,
3385 const struct ilk_wm_values *new)
3386 {
3387 unsigned int dirty = 0;
3388 enum i915_pipe pipe;
3389 int wm_lp;
3390
3391 for_each_pipe(dev_priv, pipe) {
3392 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
3393 dirty |= WM_DIRTY_LINETIME(pipe);
3394 /* Must disable LP1+ watermarks too */
3395 dirty |= WM_DIRTY_LP_ALL;
3396 }
3397
3398 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
3399 dirty |= WM_DIRTY_PIPE(pipe);
3400 /* Must disable LP1+ watermarks too */
3401 dirty |= WM_DIRTY_LP_ALL;
3402 }
3403 }
3404
3405 if (old->enable_fbc_wm != new->enable_fbc_wm) {
3406 dirty |= WM_DIRTY_FBC;
3407 /* Must disable LP1+ watermarks too */
3408 dirty |= WM_DIRTY_LP_ALL;
3409 }
3410
3411 if (old->partitioning != new->partitioning) {
3412 dirty |= WM_DIRTY_DDB;
3413 /* Must disable LP1+ watermarks too */
3414 dirty |= WM_DIRTY_LP_ALL;
3415 }
3416
3417 /* LP1+ watermarks already deemed dirty, no need to continue */
3418 if (dirty & WM_DIRTY_LP_ALL)
3419 return dirty;
3420
3421 /* Find the lowest numbered LP1+ watermark in need of an update... */
3422 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
3423 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
3424 old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
3425 break;
3426 }
3427
3428 /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
3429 for (; wm_lp <= 3; wm_lp++)
3430 dirty |= WM_DIRTY_LP(wm_lp);
3431
3432 return dirty;
3433 }
3434
_ilk_disable_lp_wm(struct drm_i915_private * dev_priv,unsigned int dirty)3435 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
3436 unsigned int dirty)
3437 {
3438 struct ilk_wm_values *previous = &dev_priv->wm.hw;
3439 bool changed = false;
3440
3441 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
3442 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
3443 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
3444 changed = true;
3445 }
3446 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
3447 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
3448 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
3449 changed = true;
3450 }
3451 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
3452 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
3453 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
3454 changed = true;
3455 }
3456
3457 /*
3458 * Don't touch WM1S_LP_EN here.
3459 * Doing so could cause underruns.
3460 */
3461
3462 return changed;
3463 }
3464
3465 /*
3466 * The spec says we shouldn't write when we don't need, because every write
3467 * causes WMs to be re-evaluated, expending some power.
3468 */
ilk_write_wm_values(struct drm_i915_private * dev_priv,struct ilk_wm_values * results)3469 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
3470 struct ilk_wm_values *results)
3471 {
3472 struct ilk_wm_values *previous = &dev_priv->wm.hw;
3473 unsigned int dirty;
3474 uint32_t val;
3475
3476 dirty = ilk_compute_wm_dirty(dev_priv, previous, results);
3477 if (!dirty)
3478 return;
3479
3480 _ilk_disable_lp_wm(dev_priv, dirty);
3481
3482 if (dirty & WM_DIRTY_PIPE(PIPE_A))
3483 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
3484 if (dirty & WM_DIRTY_PIPE(PIPE_B))
3485 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
3486 if (dirty & WM_DIRTY_PIPE(PIPE_C))
3487 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
3488
3489 if (dirty & WM_DIRTY_LINETIME(PIPE_A))
3490 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
3491 if (dirty & WM_DIRTY_LINETIME(PIPE_B))
3492 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
3493 if (dirty & WM_DIRTY_LINETIME(PIPE_C))
3494 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
3495
3496 if (dirty & WM_DIRTY_DDB) {
3497 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
3498 val = I915_READ(WM_MISC);
3499 if (results->partitioning == INTEL_DDB_PART_1_2)
3500 val &= ~WM_MISC_DATA_PARTITION_5_6;
3501 else
3502 val |= WM_MISC_DATA_PARTITION_5_6;
3503 I915_WRITE(WM_MISC, val);
3504 } else {
3505 val = I915_READ(DISP_ARB_CTL2);
3506 if (results->partitioning == INTEL_DDB_PART_1_2)
3507 val &= ~DISP_DATA_PARTITION_5_6;
3508 else
3509 val |= DISP_DATA_PARTITION_5_6;
3510 I915_WRITE(DISP_ARB_CTL2, val);
3511 }
3512 }
3513
3514 if (dirty & WM_DIRTY_FBC) {
3515 val = I915_READ(DISP_ARB_CTL);
3516 if (results->enable_fbc_wm)
3517 val &= ~DISP_FBC_WM_DIS;
3518 else
3519 val |= DISP_FBC_WM_DIS;
3520 I915_WRITE(DISP_ARB_CTL, val);
3521 }
3522
3523 if (dirty & WM_DIRTY_LP(1) &&
3524 previous->wm_lp_spr[0] != results->wm_lp_spr[0])
3525 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
3526
3527 if (INTEL_GEN(dev_priv) >= 7) {
3528 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
3529 I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
3530 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
3531 I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
3532 }
3533
3534 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
3535 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
3536 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
3537 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
3538 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
3539 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
3540
3541 dev_priv->wm.hw = *results;
3542 }
3543
ilk_disable_lp_wm(struct drm_device * dev)3544 bool ilk_disable_lp_wm(struct drm_device *dev)
3545 {
3546 struct drm_i915_private *dev_priv = to_i915(dev);
3547
3548 return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
3549 }
3550
3551 /*
3552 * FIXME: We still don't have the proper code detect if we need to apply the WA,
3553 * so assume we'll always need it in order to avoid underruns.
3554 */
skl_needs_memory_bw_wa(struct intel_atomic_state * state)3555 static bool skl_needs_memory_bw_wa(struct intel_atomic_state *state)
3556 {
3557 struct drm_i915_private *dev_priv = to_i915(state->base.dev);
3558
3559 if (IS_GEN9_BC(dev_priv) || IS_BROXTON(dev_priv))
3560 return true;
3561
3562 return false;
3563 }
3564
3565 static bool
intel_has_sagv(struct drm_i915_private * dev_priv)3566 intel_has_sagv(struct drm_i915_private *dev_priv)
3567 {
3568 if (IS_KABYLAKE(dev_priv) || IS_COFFEELAKE(dev_priv) ||
3569 IS_CANNONLAKE(dev_priv))
3570 return true;
3571
3572 if (IS_SKYLAKE(dev_priv) &&
3573 dev_priv->sagv_status != I915_SAGV_NOT_CONTROLLED)
3574 return true;
3575
3576 return false;
3577 }
3578
3579 /*
3580 * SAGV dynamically adjusts the system agent voltage and clock frequencies
3581 * depending on power and performance requirements. The display engine access
3582 * to system memory is blocked during the adjustment time. Because of the
3583 * blocking time, having this enabled can cause full system hangs and/or pipe
3584 * underruns if we don't meet all of the following requirements:
3585 *
3586 * - <= 1 pipe enabled
3587 * - All planes can enable watermarks for latencies >= SAGV engine block time
3588 * - We're not using an interlaced display configuration
3589 */
3590 int
intel_enable_sagv(struct drm_i915_private * dev_priv)3591 intel_enable_sagv(struct drm_i915_private *dev_priv)
3592 {
3593 int ret;
3594
3595 if (!intel_has_sagv(dev_priv))
3596 return 0;
3597
3598 if (dev_priv->sagv_status == I915_SAGV_ENABLED)
3599 return 0;
3600
3601 DRM_DEBUG_KMS("Enabling the SAGV\n");
3602 mutex_lock(&dev_priv->pcu_lock);
3603
3604 ret = sandybridge_pcode_write(dev_priv, GEN9_PCODE_SAGV_CONTROL,
3605 GEN9_SAGV_ENABLE);
3606
3607 /* We don't need to wait for the SAGV when enabling */
3608 mutex_unlock(&dev_priv->pcu_lock);
3609
3610 /*
3611 * Some skl systems, pre-release machines in particular,
3612 * don't actually have an SAGV.
3613 */
3614 if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
3615 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
3616 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
3617 return 0;
3618 } else if (ret < 0) {
3619 DRM_ERROR("Failed to enable the SAGV\n");
3620 return ret;
3621 }
3622
3623 dev_priv->sagv_status = I915_SAGV_ENABLED;
3624 return 0;
3625 }
3626
3627 int
intel_disable_sagv(struct drm_i915_private * dev_priv)3628 intel_disable_sagv(struct drm_i915_private *dev_priv)
3629 {
3630 int ret;
3631
3632 if (!intel_has_sagv(dev_priv))
3633 return 0;
3634
3635 if (dev_priv->sagv_status == I915_SAGV_DISABLED)
3636 return 0;
3637
3638 DRM_DEBUG_KMS("Disabling the SAGV\n");
3639 mutex_lock(&dev_priv->pcu_lock);
3640
3641 /* bspec says to keep retrying for at least 1 ms */
3642 ret = skl_pcode_request(dev_priv, GEN9_PCODE_SAGV_CONTROL,
3643 GEN9_SAGV_DISABLE,
3644 GEN9_SAGV_IS_DISABLED, GEN9_SAGV_IS_DISABLED,
3645 1);
3646 mutex_unlock(&dev_priv->pcu_lock);
3647
3648 /*
3649 * Some skl systems, pre-release machines in particular,
3650 * don't actually have an SAGV.
3651 */
3652 if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
3653 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
3654 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
3655 return 0;
3656 } else if (ret < 0) {
3657 DRM_ERROR("Failed to disable the SAGV (%d)\n", ret);
3658 return ret;
3659 }
3660
3661 dev_priv->sagv_status = I915_SAGV_DISABLED;
3662 return 0;
3663 }
3664
intel_can_enable_sagv(struct drm_atomic_state * state)3665 bool intel_can_enable_sagv(struct drm_atomic_state *state)
3666 {
3667 struct drm_device *dev = state->dev;
3668 struct drm_i915_private *dev_priv = to_i915(dev);
3669 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3670 struct intel_crtc *crtc;
3671 struct intel_plane *plane;
3672 struct intel_crtc_state *cstate;
3673 enum i915_pipe pipe;
3674 int level, latency;
3675 int sagv_block_time_us = IS_GEN9(dev_priv) ? 30 : 20;
3676
3677 if (!intel_has_sagv(dev_priv))
3678 return false;
3679
3680 /*
3681 * SKL+ workaround: bspec recommends we disable the SAGV when we have
3682 * more then one pipe enabled
3683 *
3684 * If there are no active CRTCs, no additional checks need be performed
3685 */
3686 if (hweight32(intel_state->active_crtcs) == 0)
3687 return true;
3688 else if (hweight32(intel_state->active_crtcs) > 1)
3689 return false;
3690
3691 /* Since we're now guaranteed to only have one active CRTC... */
3692 pipe = ffs(intel_state->active_crtcs) - 1;
3693 crtc = intel_get_crtc_for_pipe(dev_priv, pipe);
3694 cstate = to_intel_crtc_state(crtc->base.state);
3695
3696 if (crtc->base.state->adjusted_mode.flags & DRM_MODE_FLAG_INTERLACE)
3697 return false;
3698
3699 for_each_intel_plane_on_crtc(dev, crtc, plane) {
3700 struct skl_plane_wm *wm =
3701 &cstate->wm.skl.optimal.planes[plane->id];
3702
3703 /* Skip this plane if it's not enabled */
3704 if (!wm->wm[0].plane_en)
3705 continue;
3706
3707 /* Find the highest enabled wm level for this plane */
3708 for (level = ilk_wm_max_level(dev_priv);
3709 !wm->wm[level].plane_en; --level)
3710 { }
3711
3712 latency = dev_priv->wm.skl_latency[level];
3713
3714 if (skl_needs_memory_bw_wa(intel_state) &&
3715 plane->base.state->fb->modifier ==
3716 I915_FORMAT_MOD_X_TILED)
3717 latency += 15;
3718
3719 /*
3720 * If any of the planes on this pipe don't enable wm levels that
3721 * incur memory latencies higher than sagv_block_time_us we
3722 * can't enable the SAGV.
3723 */
3724 if (latency < sagv_block_time_us)
3725 return false;
3726 }
3727
3728 return true;
3729 }
3730
3731 static void
skl_ddb_get_pipe_allocation_limits(struct drm_device * dev,const struct intel_crtc_state * cstate,struct skl_ddb_entry * alloc,int * num_active)3732 skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
3733 const struct intel_crtc_state *cstate,
3734 struct skl_ddb_entry *alloc, /* out */
3735 int *num_active /* out */)
3736 {
3737 struct drm_atomic_state *state = cstate->base.state;
3738 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3739 struct drm_i915_private *dev_priv = to_i915(dev);
3740 struct drm_crtc *for_crtc = cstate->base.crtc;
3741 unsigned int pipe_size, ddb_size;
3742 int nth_active_pipe;
3743
3744 if (WARN_ON(!state) || !cstate->base.active) {
3745 alloc->start = 0;
3746 alloc->end = 0;
3747 *num_active = hweight32(dev_priv->active_crtcs);
3748 return;
3749 }
3750
3751 if (intel_state->active_pipe_changes)
3752 *num_active = hweight32(intel_state->active_crtcs);
3753 else
3754 *num_active = hweight32(dev_priv->active_crtcs);
3755
3756 ddb_size = INTEL_INFO(dev_priv)->ddb_size;
3757 WARN_ON(ddb_size == 0);
3758
3759 ddb_size -= 4; /* 4 blocks for bypass path allocation */
3760
3761 /*
3762 * If the state doesn't change the active CRTC's, then there's
3763 * no need to recalculate; the existing pipe allocation limits
3764 * should remain unchanged. Note that we're safe from racing
3765 * commits since any racing commit that changes the active CRTC
3766 * list would need to grab _all_ crtc locks, including the one
3767 * we currently hold.
3768 */
3769 if (!intel_state->active_pipe_changes) {
3770 /*
3771 * alloc may be cleared by clear_intel_crtc_state,
3772 * copy from old state to be sure
3773 */
3774 *alloc = to_intel_crtc_state(for_crtc->state)->wm.skl.ddb;
3775 return;
3776 }
3777
3778 nth_active_pipe = hweight32(intel_state->active_crtcs &
3779 (drm_crtc_mask(for_crtc) - 1));
3780 pipe_size = ddb_size / hweight32(intel_state->active_crtcs);
3781 alloc->start = nth_active_pipe * ddb_size / *num_active;
3782 alloc->end = alloc->start + pipe_size;
3783 }
3784
skl_cursor_allocation(int num_active)3785 static unsigned int skl_cursor_allocation(int num_active)
3786 {
3787 if (num_active == 1)
3788 return 32;
3789
3790 return 8;
3791 }
3792
skl_ddb_entry_init_from_hw(struct skl_ddb_entry * entry,u32 reg)3793 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry *entry, u32 reg)
3794 {
3795 entry->start = reg & 0x3ff;
3796 entry->end = (reg >> 16) & 0x3ff;
3797 if (entry->end)
3798 entry->end += 1;
3799 }
3800
skl_ddb_get_hw_state(struct drm_i915_private * dev_priv,struct skl_ddb_allocation * ddb)3801 void skl_ddb_get_hw_state(struct drm_i915_private *dev_priv,
3802 struct skl_ddb_allocation *ddb /* out */)
3803 {
3804 struct intel_crtc *crtc;
3805
3806 memset(ddb, 0, sizeof(*ddb));
3807
3808 for_each_intel_crtc(&dev_priv->drm, crtc) {
3809 enum intel_display_power_domain power_domain;
3810 enum plane_id plane_id;
3811 enum i915_pipe pipe = crtc->pipe;
3812
3813 power_domain = POWER_DOMAIN_PIPE(pipe);
3814 if (!intel_display_power_get_if_enabled(dev_priv, power_domain))
3815 continue;
3816
3817 for_each_plane_id_on_crtc(crtc, plane_id) {
3818 u32 val;
3819
3820 if (plane_id != PLANE_CURSOR)
3821 val = I915_READ(PLANE_BUF_CFG(pipe, plane_id));
3822 else
3823 val = I915_READ(CUR_BUF_CFG(pipe));
3824
3825 skl_ddb_entry_init_from_hw(&ddb->plane[pipe][plane_id], val);
3826 }
3827
3828 intel_display_power_put(dev_priv, power_domain);
3829 }
3830 }
3831
3832 /*
3833 * Determines the downscale amount of a plane for the purposes of watermark calculations.
3834 * The bspec defines downscale amount as:
3835 *
3836 * """
3837 * Horizontal down scale amount = maximum[1, Horizontal source size /
3838 * Horizontal destination size]
3839 * Vertical down scale amount = maximum[1, Vertical source size /
3840 * Vertical destination size]
3841 * Total down scale amount = Horizontal down scale amount *
3842 * Vertical down scale amount
3843 * """
3844 *
3845 * Return value is provided in 16.16 fixed point form to retain fractional part.
3846 * Caller should take care of dividing & rounding off the value.
3847 */
3848 static uint_fixed_16_16_t
skl_plane_downscale_amount(const struct intel_crtc_state * cstate,const struct intel_plane_state * pstate)3849 skl_plane_downscale_amount(const struct intel_crtc_state *cstate,
3850 const struct intel_plane_state *pstate)
3851 {
3852 struct intel_plane *plane = to_intel_plane(pstate->base.plane);
3853 uint32_t src_w, src_h, dst_w, dst_h;
3854 uint_fixed_16_16_t fp_w_ratio, fp_h_ratio;
3855 uint_fixed_16_16_t downscale_h, downscale_w;
3856
3857 if (WARN_ON(!intel_wm_plane_visible(cstate, pstate)))
3858 return u32_to_fixed16(0);
3859
3860 /* n.b., src is 16.16 fixed point, dst is whole integer */
3861 if (plane->id == PLANE_CURSOR) {
3862 /*
3863 * Cursors only support 0/180 degree rotation,
3864 * hence no need to account for rotation here.
3865 */
3866 src_w = pstate->base.src_w >> 16;
3867 src_h = pstate->base.src_h >> 16;
3868 dst_w = pstate->base.crtc_w;
3869 dst_h = pstate->base.crtc_h;
3870 } else {
3871 /*
3872 * Src coordinates are already rotated by 270 degrees for
3873 * the 90/270 degree plane rotation cases (to match the
3874 * GTT mapping), hence no need to account for rotation here.
3875 */
3876 src_w = drm_rect_width(&pstate->base.src) >> 16;
3877 src_h = drm_rect_height(&pstate->base.src) >> 16;
3878 dst_w = drm_rect_width(&pstate->base.dst);
3879 dst_h = drm_rect_height(&pstate->base.dst);
3880 }
3881
3882 fp_w_ratio = div_fixed16(src_w, dst_w);
3883 fp_h_ratio = div_fixed16(src_h, dst_h);
3884 downscale_w = max_fixed16(fp_w_ratio, u32_to_fixed16(1));
3885 downscale_h = max_fixed16(fp_h_ratio, u32_to_fixed16(1));
3886
3887 return mul_fixed16(downscale_w, downscale_h);
3888 }
3889
3890 static uint_fixed_16_16_t
skl_pipe_downscale_amount(const struct intel_crtc_state * crtc_state)3891 skl_pipe_downscale_amount(const struct intel_crtc_state *crtc_state)
3892 {
3893 uint_fixed_16_16_t pipe_downscale = u32_to_fixed16(1);
3894
3895 if (!crtc_state->base.enable)
3896 return pipe_downscale;
3897
3898 if (crtc_state->pch_pfit.enabled) {
3899 uint32_t src_w, src_h, dst_w, dst_h;
3900 uint32_t pfit_size = crtc_state->pch_pfit.size;
3901 uint_fixed_16_16_t fp_w_ratio, fp_h_ratio;
3902 uint_fixed_16_16_t downscale_h, downscale_w;
3903
3904 src_w = crtc_state->pipe_src_w;
3905 src_h = crtc_state->pipe_src_h;
3906 dst_w = pfit_size >> 16;
3907 dst_h = pfit_size & 0xffff;
3908
3909 if (!dst_w || !dst_h)
3910 return pipe_downscale;
3911
3912 fp_w_ratio = div_fixed16(src_w, dst_w);
3913 fp_h_ratio = div_fixed16(src_h, dst_h);
3914 downscale_w = max_fixed16(fp_w_ratio, u32_to_fixed16(1));
3915 downscale_h = max_fixed16(fp_h_ratio, u32_to_fixed16(1));
3916
3917 pipe_downscale = mul_fixed16(downscale_w, downscale_h);
3918 }
3919
3920 return pipe_downscale;
3921 }
3922
skl_check_pipe_max_pixel_rate(struct intel_crtc * intel_crtc,struct intel_crtc_state * cstate)3923 int skl_check_pipe_max_pixel_rate(struct intel_crtc *intel_crtc,
3924 struct intel_crtc_state *cstate)
3925 {
3926 struct drm_crtc_state *crtc_state = &cstate->base;
3927 struct drm_atomic_state *state = crtc_state->state;
3928 struct drm_plane *plane;
3929 const struct drm_plane_state *pstate;
3930 struct intel_plane_state *intel_pstate;
3931 int crtc_clock, dotclk;
3932 uint32_t pipe_max_pixel_rate;
3933 uint_fixed_16_16_t pipe_downscale;
3934 uint_fixed_16_16_t max_downscale = u32_to_fixed16(1);
3935
3936 if (!cstate->base.enable)
3937 return 0;
3938
3939 drm_atomic_crtc_state_for_each_plane_state(plane, pstate, crtc_state) {
3940 uint_fixed_16_16_t plane_downscale;
3941 uint_fixed_16_16_t fp_9_div_8 = div_fixed16(9, 8);
3942 int bpp;
3943
3944 if (!intel_wm_plane_visible(cstate,
3945 to_intel_plane_state(pstate)))
3946 continue;
3947
3948 if (WARN_ON(!pstate->fb))
3949 return -EINVAL;
3950
3951 intel_pstate = to_intel_plane_state(pstate);
3952 plane_downscale = skl_plane_downscale_amount(cstate,
3953 intel_pstate);
3954 bpp = pstate->fb->format->cpp[0] * 8;
3955 if (bpp == 64)
3956 plane_downscale = mul_fixed16(plane_downscale,
3957 fp_9_div_8);
3958
3959 max_downscale = max_fixed16(plane_downscale, max_downscale);
3960 }
3961 pipe_downscale = skl_pipe_downscale_amount(cstate);
3962
3963 pipe_downscale = mul_fixed16(pipe_downscale, max_downscale);
3964
3965 crtc_clock = crtc_state->adjusted_mode.crtc_clock;
3966 dotclk = to_intel_atomic_state(state)->cdclk.logical.cdclk;
3967
3968 if (IS_GEMINILAKE(to_i915(intel_crtc->base.dev)))
3969 dotclk *= 2;
3970
3971 pipe_max_pixel_rate = div_round_up_u32_fixed16(dotclk, pipe_downscale);
3972
3973 if (pipe_max_pixel_rate < crtc_clock) {
3974 DRM_DEBUG_KMS("Max supported pixel clock with scaling exceeded\n");
3975 return -EINVAL;
3976 }
3977
3978 return 0;
3979 }
3980
3981 static unsigned int
skl_plane_relative_data_rate(const struct intel_crtc_state * cstate,const struct drm_plane_state * pstate,int y)3982 skl_plane_relative_data_rate(const struct intel_crtc_state *cstate,
3983 const struct drm_plane_state *pstate,
3984 int y)
3985 {
3986 struct intel_plane *plane = to_intel_plane(pstate->plane);
3987 struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
3988 uint32_t data_rate;
3989 uint32_t width = 0, height = 0;
3990 struct drm_framebuffer *fb;
3991 u32 format;
3992 uint_fixed_16_16_t down_scale_amount;
3993
3994 if (!intel_pstate->base.visible)
3995 return 0;
3996
3997 fb = pstate->fb;
3998 format = fb->format->format;
3999
4000 if (plane->id == PLANE_CURSOR)
4001 return 0;
4002 if (y && format != DRM_FORMAT_NV12)
4003 return 0;
4004
4005 /*
4006 * Src coordinates are already rotated by 270 degrees for
4007 * the 90/270 degree plane rotation cases (to match the
4008 * GTT mapping), hence no need to account for rotation here.
4009 */
4010 width = drm_rect_width(&intel_pstate->base.src) >> 16;
4011 height = drm_rect_height(&intel_pstate->base.src) >> 16;
4012
4013 /* for planar format */
4014 if (format == DRM_FORMAT_NV12) {
4015 if (y) /* y-plane data rate */
4016 data_rate = width * height *
4017 fb->format->cpp[0];
4018 else /* uv-plane data rate */
4019 data_rate = (width / 2) * (height / 2) *
4020 fb->format->cpp[1];
4021 } else {
4022 /* for packed formats */
4023 data_rate = width * height * fb->format->cpp[0];
4024 }
4025
4026 down_scale_amount = skl_plane_downscale_amount(cstate, intel_pstate);
4027
4028 return mul_round_up_u32_fixed16(data_rate, down_scale_amount);
4029 }
4030
4031 /*
4032 * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
4033 * a 8192x4096@32bpp framebuffer:
4034 * 3 * 4096 * 8192 * 4 < 2^32
4035 */
4036 static unsigned int
skl_get_total_relative_data_rate(struct intel_crtc_state * intel_cstate,unsigned * plane_data_rate,unsigned * plane_y_data_rate)4037 skl_get_total_relative_data_rate(struct intel_crtc_state *intel_cstate,
4038 unsigned *plane_data_rate,
4039 unsigned *plane_y_data_rate)
4040 {
4041 struct drm_crtc_state *cstate = &intel_cstate->base;
4042 struct drm_atomic_state *state = cstate->state;
4043 struct drm_plane *plane;
4044 const struct drm_plane_state *pstate;
4045 unsigned int total_data_rate = 0;
4046
4047 if (WARN_ON(!state))
4048 return 0;
4049
4050 /* Calculate and cache data rate for each plane */
4051 drm_atomic_crtc_state_for_each_plane_state(plane, pstate, cstate) {
4052 enum plane_id plane_id = to_intel_plane(plane)->id;
4053 unsigned int rate;
4054
4055 /* packed/uv */
4056 rate = skl_plane_relative_data_rate(intel_cstate,
4057 pstate, 0);
4058 plane_data_rate[plane_id] = rate;
4059
4060 total_data_rate += rate;
4061
4062 /* y-plane */
4063 rate = skl_plane_relative_data_rate(intel_cstate,
4064 pstate, 1);
4065 plane_y_data_rate[plane_id] = rate;
4066
4067 total_data_rate += rate;
4068 }
4069
4070 return total_data_rate;
4071 }
4072
4073 static uint16_t
skl_ddb_min_alloc(const struct drm_plane_state * pstate,const int y)4074 skl_ddb_min_alloc(const struct drm_plane_state *pstate,
4075 const int y)
4076 {
4077 struct drm_framebuffer *fb = pstate->fb;
4078 struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
4079 uint32_t src_w, src_h;
4080 uint32_t min_scanlines = 8;
4081 uint8_t plane_bpp;
4082
4083 if (WARN_ON(!fb))
4084 return 0;
4085
4086 /* For packed formats, no y-plane, return 0 */
4087 if (y && fb->format->format != DRM_FORMAT_NV12)
4088 return 0;
4089
4090 /* For Non Y-tile return 8-blocks */
4091 if (fb->modifier != I915_FORMAT_MOD_Y_TILED &&
4092 fb->modifier != I915_FORMAT_MOD_Yf_TILED &&
4093 fb->modifier != I915_FORMAT_MOD_Y_TILED_CCS &&
4094 fb->modifier != I915_FORMAT_MOD_Yf_TILED_CCS)
4095 return 8;
4096
4097 /*
4098 * Src coordinates are already rotated by 270 degrees for
4099 * the 90/270 degree plane rotation cases (to match the
4100 * GTT mapping), hence no need to account for rotation here.
4101 */
4102 src_w = drm_rect_width(&intel_pstate->base.src) >> 16;
4103 src_h = drm_rect_height(&intel_pstate->base.src) >> 16;
4104
4105 /* Halve UV plane width and height for NV12 */
4106 if (fb->format->format == DRM_FORMAT_NV12 && !y) {
4107 src_w /= 2;
4108 src_h /= 2;
4109 }
4110
4111 if (fb->format->format == DRM_FORMAT_NV12 && !y)
4112 plane_bpp = fb->format->cpp[1];
4113 else
4114 plane_bpp = fb->format->cpp[0];
4115
4116 if (drm_rotation_90_or_270(pstate->rotation)) {
4117 switch (plane_bpp) {
4118 case 1:
4119 min_scanlines = 32;
4120 break;
4121 case 2:
4122 min_scanlines = 16;
4123 break;
4124 case 4:
4125 min_scanlines = 8;
4126 break;
4127 case 8:
4128 min_scanlines = 4;
4129 break;
4130 default:
4131 WARN(1, "Unsupported pixel depth %u for rotation",
4132 plane_bpp);
4133 min_scanlines = 32;
4134 }
4135 }
4136
4137 return DIV_ROUND_UP((4 * src_w * plane_bpp), 512) * min_scanlines/4 + 3;
4138 }
4139
4140 static void
skl_ddb_calc_min(const struct intel_crtc_state * cstate,int num_active,uint16_t * minimum,uint16_t * y_minimum)4141 skl_ddb_calc_min(const struct intel_crtc_state *cstate, int num_active,
4142 uint16_t *minimum, uint16_t *y_minimum)
4143 {
4144 const struct drm_plane_state *pstate;
4145 struct drm_plane *plane;
4146
4147 drm_atomic_crtc_state_for_each_plane_state(plane, pstate, &cstate->base) {
4148 enum plane_id plane_id = to_intel_plane(plane)->id;
4149
4150 if (plane_id == PLANE_CURSOR)
4151 continue;
4152
4153 if (!pstate->visible)
4154 continue;
4155
4156 minimum[plane_id] = skl_ddb_min_alloc(pstate, 0);
4157 y_minimum[plane_id] = skl_ddb_min_alloc(pstate, 1);
4158 }
4159
4160 minimum[PLANE_CURSOR] = skl_cursor_allocation(num_active);
4161 }
4162
4163 static int
skl_allocate_pipe_ddb(struct intel_crtc_state * cstate,struct skl_ddb_allocation * ddb)4164 skl_allocate_pipe_ddb(struct intel_crtc_state *cstate,
4165 struct skl_ddb_allocation *ddb /* out */)
4166 {
4167 struct drm_atomic_state *state = cstate->base.state;
4168 struct drm_crtc *crtc = cstate->base.crtc;
4169 struct drm_device *dev = crtc->dev;
4170 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4171 enum i915_pipe pipe = intel_crtc->pipe;
4172 struct skl_ddb_entry *alloc = &cstate->wm.skl.ddb;
4173 uint16_t alloc_size, start;
4174 uint16_t minimum[I915_MAX_PLANES] = {};
4175 uint16_t y_minimum[I915_MAX_PLANES] = {};
4176 unsigned int total_data_rate;
4177 enum plane_id plane_id;
4178 int num_active;
4179 unsigned plane_data_rate[I915_MAX_PLANES] = {};
4180 unsigned plane_y_data_rate[I915_MAX_PLANES] = {};
4181 uint16_t total_min_blocks = 0;
4182
4183 /* Clear the partitioning for disabled planes. */
4184 memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
4185 memset(ddb->y_plane[pipe], 0, sizeof(ddb->y_plane[pipe]));
4186
4187 if (WARN_ON(!state))
4188 return 0;
4189
4190 if (!cstate->base.active) {
4191 alloc->start = alloc->end = 0;
4192 return 0;
4193 }
4194
4195 skl_ddb_get_pipe_allocation_limits(dev, cstate, alloc, &num_active);
4196 alloc_size = skl_ddb_entry_size(alloc);
4197 if (alloc_size == 0)
4198 return 0;
4199
4200 skl_ddb_calc_min(cstate, num_active, minimum, y_minimum);
4201
4202 /*
4203 * 1. Allocate the mininum required blocks for each active plane
4204 * and allocate the cursor, it doesn't require extra allocation
4205 * proportional to the data rate.
4206 */
4207
4208 for_each_plane_id_on_crtc(intel_crtc, plane_id) {
4209 total_min_blocks += minimum[plane_id];
4210 total_min_blocks += y_minimum[plane_id];
4211 }
4212
4213 if (total_min_blocks > alloc_size) {
4214 DRM_DEBUG_KMS("Requested display configuration exceeds system DDB limitations");
4215 DRM_DEBUG_KMS("minimum required %d/%d\n", total_min_blocks,
4216 alloc_size);
4217 return -EINVAL;
4218 }
4219
4220 alloc_size -= total_min_blocks;
4221 ddb->plane[pipe][PLANE_CURSOR].start = alloc->end - minimum[PLANE_CURSOR];
4222 ddb->plane[pipe][PLANE_CURSOR].end = alloc->end;
4223
4224 /*
4225 * 2. Distribute the remaining space in proportion to the amount of
4226 * data each plane needs to fetch from memory.
4227 *
4228 * FIXME: we may not allocate every single block here.
4229 */
4230 total_data_rate = skl_get_total_relative_data_rate(cstate,
4231 plane_data_rate,
4232 plane_y_data_rate);
4233 if (total_data_rate == 0)
4234 return 0;
4235
4236 start = alloc->start;
4237 for_each_plane_id_on_crtc(intel_crtc, plane_id) {
4238 unsigned int data_rate, y_data_rate;
4239 uint16_t plane_blocks, y_plane_blocks = 0;
4240
4241 if (plane_id == PLANE_CURSOR)
4242 continue;
4243
4244 data_rate = plane_data_rate[plane_id];
4245
4246 /*
4247 * allocation for (packed formats) or (uv-plane part of planar format):
4248 * promote the expression to 64 bits to avoid overflowing, the
4249 * result is < available as data_rate / total_data_rate < 1
4250 */
4251 plane_blocks = minimum[plane_id];
4252 plane_blocks += div_u64((uint64_t)alloc_size * data_rate,
4253 total_data_rate);
4254
4255 /* Leave disabled planes at (0,0) */
4256 if (data_rate) {
4257 ddb->plane[pipe][plane_id].start = start;
4258 ddb->plane[pipe][plane_id].end = start + plane_blocks;
4259 }
4260
4261 start += plane_blocks;
4262
4263 /*
4264 * allocation for y_plane part of planar format:
4265 */
4266 y_data_rate = plane_y_data_rate[plane_id];
4267
4268 y_plane_blocks = y_minimum[plane_id];
4269 y_plane_blocks += div_u64((uint64_t)alloc_size * y_data_rate,
4270 total_data_rate);
4271
4272 if (y_data_rate) {
4273 ddb->y_plane[pipe][plane_id].start = start;
4274 ddb->y_plane[pipe][plane_id].end = start + y_plane_blocks;
4275 }
4276
4277 start += y_plane_blocks;
4278 }
4279
4280 return 0;
4281 }
4282
4283 /*
4284 * The max latency should be 257 (max the punit can code is 255 and we add 2us
4285 * for the read latency) and cpp should always be <= 8, so that
4286 * should allow pixel_rate up to ~2 GHz which seems sufficient since max
4287 * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
4288 */
4289 static uint_fixed_16_16_t
skl_wm_method1(const struct drm_i915_private * dev_priv,uint32_t pixel_rate,uint8_t cpp,uint32_t latency)4290 skl_wm_method1(const struct drm_i915_private *dev_priv, uint32_t pixel_rate,
4291 uint8_t cpp, uint32_t latency)
4292 {
4293 uint32_t wm_intermediate_val;
4294 uint_fixed_16_16_t ret;
4295
4296 if (latency == 0)
4297 return FP_16_16_MAX;
4298
4299 wm_intermediate_val = latency * pixel_rate * cpp;
4300 ret = div_fixed16(wm_intermediate_val, 1000 * 512);
4301
4302 if (INTEL_GEN(dev_priv) >= 10)
4303 ret = add_fixed16_u32(ret, 1);
4304
4305 return ret;
4306 }
4307
skl_wm_method2(uint32_t pixel_rate,uint32_t pipe_htotal,uint32_t latency,uint_fixed_16_16_t plane_blocks_per_line)4308 static uint_fixed_16_16_t skl_wm_method2(uint32_t pixel_rate,
4309 uint32_t pipe_htotal,
4310 uint32_t latency,
4311 uint_fixed_16_16_t plane_blocks_per_line)
4312 {
4313 uint32_t wm_intermediate_val;
4314 uint_fixed_16_16_t ret;
4315
4316 if (latency == 0)
4317 return FP_16_16_MAX;
4318
4319 wm_intermediate_val = latency * pixel_rate;
4320 wm_intermediate_val = DIV_ROUND_UP(wm_intermediate_val,
4321 pipe_htotal * 1000);
4322 ret = mul_u32_fixed16(wm_intermediate_val, plane_blocks_per_line);
4323 return ret;
4324 }
4325
4326 static uint_fixed_16_16_t
intel_get_linetime_us(struct intel_crtc_state * cstate)4327 intel_get_linetime_us(struct intel_crtc_state *cstate)
4328 {
4329 uint32_t pixel_rate;
4330 uint32_t crtc_htotal;
4331 uint_fixed_16_16_t linetime_us;
4332
4333 if (!cstate->base.active)
4334 return u32_to_fixed16(0);
4335
4336 pixel_rate = cstate->pixel_rate;
4337
4338 if (WARN_ON(pixel_rate == 0))
4339 return u32_to_fixed16(0);
4340
4341 crtc_htotal = cstate->base.adjusted_mode.crtc_htotal;
4342 linetime_us = div_fixed16(crtc_htotal * 1000, pixel_rate);
4343
4344 return linetime_us;
4345 }
4346
4347 static uint32_t
skl_adjusted_plane_pixel_rate(const struct intel_crtc_state * cstate,const struct intel_plane_state * pstate)4348 skl_adjusted_plane_pixel_rate(const struct intel_crtc_state *cstate,
4349 const struct intel_plane_state *pstate)
4350 {
4351 uint64_t adjusted_pixel_rate;
4352 uint_fixed_16_16_t downscale_amount;
4353
4354 /* Shouldn't reach here on disabled planes... */
4355 if (WARN_ON(!intel_wm_plane_visible(cstate, pstate)))
4356 return 0;
4357
4358 /*
4359 * Adjusted plane pixel rate is just the pipe's adjusted pixel rate
4360 * with additional adjustments for plane-specific scaling.
4361 */
4362 adjusted_pixel_rate = cstate->pixel_rate;
4363 downscale_amount = skl_plane_downscale_amount(cstate, pstate);
4364
4365 return mul_round_up_u32_fixed16(adjusted_pixel_rate,
4366 downscale_amount);
4367 }
4368
4369 static int
skl_compute_plane_wm_params(const struct drm_i915_private * dev_priv,struct intel_crtc_state * cstate,const struct intel_plane_state * intel_pstate,struct skl_wm_params * wp)4370 skl_compute_plane_wm_params(const struct drm_i915_private *dev_priv,
4371 struct intel_crtc_state *cstate,
4372 const struct intel_plane_state *intel_pstate,
4373 struct skl_wm_params *wp)
4374 {
4375 struct intel_plane *plane = to_intel_plane(intel_pstate->base.plane);
4376 const struct drm_plane_state *pstate = &intel_pstate->base;
4377 const struct drm_framebuffer *fb = pstate->fb;
4378 uint32_t interm_pbpl;
4379 struct intel_atomic_state *state =
4380 to_intel_atomic_state(cstate->base.state);
4381 bool apply_memory_bw_wa = skl_needs_memory_bw_wa(state);
4382
4383 if (!intel_wm_plane_visible(cstate, intel_pstate))
4384 return 0;
4385
4386 wp->y_tiled = fb->modifier == I915_FORMAT_MOD_Y_TILED ||
4387 fb->modifier == I915_FORMAT_MOD_Yf_TILED ||
4388 fb->modifier == I915_FORMAT_MOD_Y_TILED_CCS ||
4389 fb->modifier == I915_FORMAT_MOD_Yf_TILED_CCS;
4390 wp->x_tiled = fb->modifier == I915_FORMAT_MOD_X_TILED;
4391 wp->rc_surface = fb->modifier == I915_FORMAT_MOD_Y_TILED_CCS ||
4392 fb->modifier == I915_FORMAT_MOD_Yf_TILED_CCS;
4393
4394 if (plane->id == PLANE_CURSOR) {
4395 wp->width = intel_pstate->base.crtc_w;
4396 } else {
4397 /*
4398 * Src coordinates are already rotated by 270 degrees for
4399 * the 90/270 degree plane rotation cases (to match the
4400 * GTT mapping), hence no need to account for rotation here.
4401 */
4402 wp->width = drm_rect_width(&intel_pstate->base.src) >> 16;
4403 }
4404
4405 wp->cpp = (fb->format->format == DRM_FORMAT_NV12) ? fb->format->cpp[1] :
4406 fb->format->cpp[0];
4407 wp->plane_pixel_rate = skl_adjusted_plane_pixel_rate(cstate,
4408 intel_pstate);
4409
4410 if (drm_rotation_90_or_270(pstate->rotation)) {
4411
4412 switch (wp->cpp) {
4413 case 1:
4414 wp->y_min_scanlines = 16;
4415 break;
4416 case 2:
4417 wp->y_min_scanlines = 8;
4418 break;
4419 case 4:
4420 wp->y_min_scanlines = 4;
4421 break;
4422 default:
4423 MISSING_CASE(wp->cpp);
4424 return -EINVAL;
4425 }
4426 } else {
4427 wp->y_min_scanlines = 4;
4428 }
4429
4430 if (apply_memory_bw_wa)
4431 wp->y_min_scanlines *= 2;
4432
4433 wp->plane_bytes_per_line = wp->width * wp->cpp;
4434 if (wp->y_tiled) {
4435 interm_pbpl = DIV_ROUND_UP(wp->plane_bytes_per_line *
4436 wp->y_min_scanlines, 512);
4437
4438 if (INTEL_GEN(dev_priv) >= 10)
4439 interm_pbpl++;
4440
4441 wp->plane_blocks_per_line = div_fixed16(interm_pbpl,
4442 wp->y_min_scanlines);
4443 } else if (wp->x_tiled && IS_GEN9(dev_priv)) {
4444 interm_pbpl = DIV_ROUND_UP(wp->plane_bytes_per_line, 512);
4445 wp->plane_blocks_per_line = u32_to_fixed16(interm_pbpl);
4446 } else {
4447 interm_pbpl = DIV_ROUND_UP(wp->plane_bytes_per_line, 512) + 1;
4448 wp->plane_blocks_per_line = u32_to_fixed16(interm_pbpl);
4449 }
4450
4451 wp->y_tile_minimum = mul_u32_fixed16(wp->y_min_scanlines,
4452 wp->plane_blocks_per_line);
4453 wp->linetime_us = fixed16_to_u32_round_up(
4454 intel_get_linetime_us(cstate));
4455
4456 return 0;
4457 }
4458
skl_compute_plane_wm(const struct drm_i915_private * dev_priv,struct intel_crtc_state * cstate,const struct intel_plane_state * intel_pstate,uint16_t ddb_allocation,int level,const struct skl_wm_params * wp,uint16_t * out_blocks,uint8_t * out_lines,bool * enabled)4459 static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv,
4460 struct intel_crtc_state *cstate,
4461 const struct intel_plane_state *intel_pstate,
4462 uint16_t ddb_allocation,
4463 int level,
4464 const struct skl_wm_params *wp,
4465 uint16_t *out_blocks, /* out */
4466 uint8_t *out_lines, /* out */
4467 bool *enabled /* out */)
4468 {
4469 const struct drm_plane_state *pstate = &intel_pstate->base;
4470 uint32_t latency = dev_priv->wm.skl_latency[level];
4471 uint_fixed_16_16_t method1, method2;
4472 uint_fixed_16_16_t selected_result;
4473 uint32_t res_blocks, res_lines;
4474 struct intel_atomic_state *state =
4475 to_intel_atomic_state(cstate->base.state);
4476 bool apply_memory_bw_wa = skl_needs_memory_bw_wa(state);
4477
4478 if (latency == 0 ||
4479 !intel_wm_plane_visible(cstate, intel_pstate)) {
4480 *enabled = false;
4481 return 0;
4482 }
4483
4484 /* Display WA #1141: kbl,cfl */
4485 if ((IS_KABYLAKE(dev_priv) || IS_COFFEELAKE(dev_priv) ||
4486 IS_CNL_REVID(dev_priv, CNL_REVID_A0, CNL_REVID_B0)) &&
4487 dev_priv->ipc_enabled)
4488 latency += 4;
4489
4490 if (apply_memory_bw_wa && wp->x_tiled)
4491 latency += 15;
4492
4493 method1 = skl_wm_method1(dev_priv, wp->plane_pixel_rate,
4494 wp->cpp, latency);
4495 method2 = skl_wm_method2(wp->plane_pixel_rate,
4496 cstate->base.adjusted_mode.crtc_htotal,
4497 latency,
4498 wp->plane_blocks_per_line);
4499
4500 if (wp->y_tiled) {
4501 selected_result = max_fixed16(method2, wp->y_tile_minimum);
4502 } else {
4503 if ((wp->cpp * cstate->base.adjusted_mode.crtc_htotal /
4504 512 < 1) && (wp->plane_bytes_per_line / 512 < 1))
4505 selected_result = method2;
4506 else if (ddb_allocation >=
4507 fixed16_to_u32_round_up(wp->plane_blocks_per_line))
4508 selected_result = min_fixed16(method1, method2);
4509 else if (latency >= wp->linetime_us)
4510 selected_result = min_fixed16(method1, method2);
4511 else
4512 selected_result = method1;
4513 }
4514
4515 res_blocks = fixed16_to_u32_round_up(selected_result) + 1;
4516 res_lines = div_round_up_fixed16(selected_result,
4517 wp->plane_blocks_per_line);
4518
4519 /* Display WA #1125: skl,bxt,kbl,glk */
4520 if (level == 0 && wp->rc_surface)
4521 res_blocks += fixed16_to_u32_round_up(wp->y_tile_minimum);
4522
4523 /* Display WA #1126: skl,bxt,kbl,glk */
4524 if (level >= 1 && level <= 7) {
4525 if (wp->y_tiled) {
4526 res_blocks += fixed16_to_u32_round_up(
4527 wp->y_tile_minimum);
4528 res_lines += wp->y_min_scanlines;
4529 } else {
4530 res_blocks++;
4531 }
4532 }
4533
4534 if (res_blocks >= ddb_allocation || res_lines > 31) {
4535 *enabled = false;
4536
4537 /*
4538 * If there are no valid level 0 watermarks, then we can't
4539 * support this display configuration.
4540 */
4541 if (level) {
4542 return 0;
4543 } else {
4544 struct drm_plane *plane = pstate->plane;
4545
4546 DRM_DEBUG_KMS("Requested display configuration exceeds system watermark limitations\n");
4547 DRM_DEBUG_KMS("[PLANE:%d:%s] blocks required = %u/%u, lines required = %u/31\n",
4548 plane->base.id, plane->name,
4549 res_blocks, ddb_allocation, res_lines);
4550 return -EINVAL;
4551 }
4552 }
4553
4554 *out_blocks = res_blocks;
4555 *out_lines = res_lines;
4556 *enabled = true;
4557
4558 return 0;
4559 }
4560
4561 static int
skl_compute_wm_levels(const struct drm_i915_private * dev_priv,struct skl_ddb_allocation * ddb,struct intel_crtc_state * cstate,const struct intel_plane_state * intel_pstate,const struct skl_wm_params * wm_params,struct skl_plane_wm * wm)4562 skl_compute_wm_levels(const struct drm_i915_private *dev_priv,
4563 struct skl_ddb_allocation *ddb,
4564 struct intel_crtc_state *cstate,
4565 const struct intel_plane_state *intel_pstate,
4566 const struct skl_wm_params *wm_params,
4567 struct skl_plane_wm *wm)
4568 {
4569 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4570 struct drm_plane *plane = intel_pstate->base.plane;
4571 struct intel_plane *intel_plane = to_intel_plane(plane);
4572 uint16_t ddb_blocks;
4573 enum i915_pipe pipe = intel_crtc->pipe;
4574 int level, max_level = ilk_wm_max_level(dev_priv);
4575 int ret;
4576
4577 if (WARN_ON(!intel_pstate->base.fb))
4578 return -EINVAL;
4579
4580 ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][intel_plane->id]);
4581
4582 for (level = 0; level <= max_level; level++) {
4583 struct skl_wm_level *result = &wm->wm[level];
4584
4585 ret = skl_compute_plane_wm(dev_priv,
4586 cstate,
4587 intel_pstate,
4588 ddb_blocks,
4589 level,
4590 wm_params,
4591 &result->plane_res_b,
4592 &result->plane_res_l,
4593 &result->plane_en);
4594 if (ret)
4595 return ret;
4596 }
4597
4598 return 0;
4599 }
4600
4601 static uint32_t
skl_compute_linetime_wm(struct intel_crtc_state * cstate)4602 skl_compute_linetime_wm(struct intel_crtc_state *cstate)
4603 {
4604 struct drm_atomic_state *state = cstate->base.state;
4605 struct drm_i915_private *dev_priv = to_i915(state->dev);
4606 uint_fixed_16_16_t linetime_us;
4607 uint32_t linetime_wm;
4608
4609 linetime_us = intel_get_linetime_us(cstate);
4610
4611 if (is_fixed16_zero(linetime_us))
4612 return 0;
4613
4614 linetime_wm = fixed16_to_u32_round_up(mul_u32_fixed16(8, linetime_us));
4615
4616 /* Display WA #1135: bxt:ALL GLK:ALL */
4617 if ((IS_BROXTON(dev_priv) || IS_GEMINILAKE(dev_priv)) &&
4618 dev_priv->ipc_enabled)
4619 linetime_wm /= 2;
4620
4621 return linetime_wm;
4622 }
4623
skl_compute_transition_wm(struct intel_crtc_state * cstate,struct skl_wm_params * wp,struct skl_wm_level * wm_l0,uint16_t ddb_allocation,struct skl_wm_level * trans_wm)4624 static void skl_compute_transition_wm(struct intel_crtc_state *cstate,
4625 struct skl_wm_params *wp,
4626 struct skl_wm_level *wm_l0,
4627 uint16_t ddb_allocation,
4628 struct skl_wm_level *trans_wm /* out */)
4629 {
4630 struct drm_device *dev = cstate->base.crtc->dev;
4631 const struct drm_i915_private *dev_priv = to_i915(dev);
4632 uint16_t trans_min, trans_y_tile_min;
4633 const uint16_t trans_amount = 10; /* This is configurable amount */
4634 uint16_t trans_offset_b, res_blocks;
4635
4636 if (!cstate->base.active)
4637 goto exit;
4638
4639 /* Transition WM are not recommended by HW team for GEN9 */
4640 if (INTEL_GEN(dev_priv) <= 9)
4641 goto exit;
4642
4643 /* Transition WM don't make any sense if ipc is disabled */
4644 if (!dev_priv->ipc_enabled)
4645 goto exit;
4646
4647 if (INTEL_GEN(dev_priv) >= 10)
4648 trans_min = 4;
4649
4650 trans_offset_b = trans_min + trans_amount;
4651
4652 if (wp->y_tiled) {
4653 trans_y_tile_min = (uint16_t) mul_round_up_u32_fixed16(2,
4654 wp->y_tile_minimum);
4655 res_blocks = max(wm_l0->plane_res_b, trans_y_tile_min) +
4656 trans_offset_b;
4657 } else {
4658 res_blocks = wm_l0->plane_res_b + trans_offset_b;
4659
4660 /* WA BUG:1938466 add one block for non y-tile planes */
4661 if (IS_CNL_REVID(dev_priv, CNL_REVID_A0, CNL_REVID_A0))
4662 res_blocks += 1;
4663
4664 }
4665
4666 res_blocks += 1;
4667
4668 if (res_blocks < ddb_allocation) {
4669 trans_wm->plane_res_b = res_blocks;
4670 trans_wm->plane_en = true;
4671 return;
4672 }
4673
4674 exit:
4675 trans_wm->plane_en = false;
4676 }
4677
skl_build_pipe_wm(struct intel_crtc_state * cstate,struct skl_ddb_allocation * ddb,struct skl_pipe_wm * pipe_wm)4678 static int skl_build_pipe_wm(struct intel_crtc_state *cstate,
4679 struct skl_ddb_allocation *ddb,
4680 struct skl_pipe_wm *pipe_wm)
4681 {
4682 struct drm_device *dev = cstate->base.crtc->dev;
4683 struct drm_crtc_state *crtc_state = &cstate->base;
4684 const struct drm_i915_private *dev_priv = to_i915(dev);
4685 struct drm_plane *plane;
4686 const struct drm_plane_state *pstate;
4687 struct skl_plane_wm *wm;
4688 int ret;
4689
4690 /*
4691 * We'll only calculate watermarks for planes that are actually
4692 * enabled, so make sure all other planes are set as disabled.
4693 */
4694 memset(pipe_wm->planes, 0, sizeof(pipe_wm->planes));
4695
4696 drm_atomic_crtc_state_for_each_plane_state(plane, pstate, crtc_state) {
4697 const struct intel_plane_state *intel_pstate =
4698 to_intel_plane_state(pstate);
4699 enum plane_id plane_id = to_intel_plane(plane)->id;
4700 struct skl_wm_params wm_params;
4701 enum i915_pipe pipe = to_intel_crtc(cstate->base.crtc)->pipe;
4702 uint16_t ddb_blocks;
4703
4704 wm = &pipe_wm->planes[plane_id];
4705 ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][plane_id]);
4706 memset(&wm_params, 0, sizeof(struct skl_wm_params));
4707
4708 ret = skl_compute_plane_wm_params(dev_priv, cstate,
4709 intel_pstate, &wm_params);
4710 if (ret)
4711 return ret;
4712
4713 ret = skl_compute_wm_levels(dev_priv, ddb, cstate,
4714 intel_pstate, &wm_params, wm);
4715 if (ret)
4716 return ret;
4717 skl_compute_transition_wm(cstate, &wm_params, &wm->wm[0],
4718 ddb_blocks, &wm->trans_wm);
4719 }
4720 pipe_wm->linetime = skl_compute_linetime_wm(cstate);
4721
4722 return 0;
4723 }
4724
skl_ddb_entry_write(struct drm_i915_private * dev_priv,i915_reg_t reg,const struct skl_ddb_entry * entry)4725 static void skl_ddb_entry_write(struct drm_i915_private *dev_priv,
4726 i915_reg_t reg,
4727 const struct skl_ddb_entry *entry)
4728 {
4729 if (entry->end)
4730 I915_WRITE(reg, (entry->end - 1) << 16 | entry->start);
4731 else
4732 I915_WRITE(reg, 0);
4733 }
4734
skl_write_wm_level(struct drm_i915_private * dev_priv,i915_reg_t reg,const struct skl_wm_level * level)4735 static void skl_write_wm_level(struct drm_i915_private *dev_priv,
4736 i915_reg_t reg,
4737 const struct skl_wm_level *level)
4738 {
4739 uint32_t val = 0;
4740
4741 if (level->plane_en) {
4742 val |= PLANE_WM_EN;
4743 val |= level->plane_res_b;
4744 val |= level->plane_res_l << PLANE_WM_LINES_SHIFT;
4745 }
4746
4747 I915_WRITE(reg, val);
4748 }
4749
skl_write_plane_wm(struct intel_crtc * intel_crtc,const struct skl_plane_wm * wm,const struct skl_ddb_allocation * ddb,enum plane_id plane_id)4750 static void skl_write_plane_wm(struct intel_crtc *intel_crtc,
4751 const struct skl_plane_wm *wm,
4752 const struct skl_ddb_allocation *ddb,
4753 enum plane_id plane_id)
4754 {
4755 struct drm_crtc *crtc = &intel_crtc->base;
4756 struct drm_device *dev = crtc->dev;
4757 struct drm_i915_private *dev_priv = to_i915(dev);
4758 int level, max_level = ilk_wm_max_level(dev_priv);
4759 enum i915_pipe pipe = intel_crtc->pipe;
4760
4761 for (level = 0; level <= max_level; level++) {
4762 skl_write_wm_level(dev_priv, PLANE_WM(pipe, plane_id, level),
4763 &wm->wm[level]);
4764 }
4765 skl_write_wm_level(dev_priv, PLANE_WM_TRANS(pipe, plane_id),
4766 &wm->trans_wm);
4767
4768 skl_ddb_entry_write(dev_priv, PLANE_BUF_CFG(pipe, plane_id),
4769 &ddb->plane[pipe][plane_id]);
4770 skl_ddb_entry_write(dev_priv, PLANE_NV12_BUF_CFG(pipe, plane_id),
4771 &ddb->y_plane[pipe][plane_id]);
4772 }
4773
skl_write_cursor_wm(struct intel_crtc * intel_crtc,const struct skl_plane_wm * wm,const struct skl_ddb_allocation * ddb)4774 static void skl_write_cursor_wm(struct intel_crtc *intel_crtc,
4775 const struct skl_plane_wm *wm,
4776 const struct skl_ddb_allocation *ddb)
4777 {
4778 struct drm_crtc *crtc = &intel_crtc->base;
4779 struct drm_device *dev = crtc->dev;
4780 struct drm_i915_private *dev_priv = to_i915(dev);
4781 int level, max_level = ilk_wm_max_level(dev_priv);
4782 enum i915_pipe pipe = intel_crtc->pipe;
4783
4784 for (level = 0; level <= max_level; level++) {
4785 skl_write_wm_level(dev_priv, CUR_WM(pipe, level),
4786 &wm->wm[level]);
4787 }
4788 skl_write_wm_level(dev_priv, CUR_WM_TRANS(pipe), &wm->trans_wm);
4789
4790 skl_ddb_entry_write(dev_priv, CUR_BUF_CFG(pipe),
4791 &ddb->plane[pipe][PLANE_CURSOR]);
4792 }
4793
skl_wm_level_equals(const struct skl_wm_level * l1,const struct skl_wm_level * l2)4794 bool skl_wm_level_equals(const struct skl_wm_level *l1,
4795 const struct skl_wm_level *l2)
4796 {
4797 if (l1->plane_en != l2->plane_en)
4798 return false;
4799
4800 /* If both planes aren't enabled, the rest shouldn't matter */
4801 if (!l1->plane_en)
4802 return true;
4803
4804 return (l1->plane_res_l == l2->plane_res_l &&
4805 l1->plane_res_b == l2->plane_res_b);
4806 }
4807
skl_ddb_entries_overlap(const struct skl_ddb_entry * a,const struct skl_ddb_entry * b)4808 static inline bool skl_ddb_entries_overlap(const struct skl_ddb_entry *a,
4809 const struct skl_ddb_entry *b)
4810 {
4811 return a->start < b->end && b->start < a->end;
4812 }
4813
skl_ddb_allocation_overlaps(struct drm_i915_private * dev_priv,const struct skl_ddb_entry ** entries,const struct skl_ddb_entry * ddb,int ignore)4814 bool skl_ddb_allocation_overlaps(struct drm_i915_private *dev_priv,
4815 const struct skl_ddb_entry **entries,
4816 const struct skl_ddb_entry *ddb,
4817 int ignore)
4818 {
4819 enum i915_pipe pipe;
4820
4821 for_each_pipe(dev_priv, pipe) {
4822 if (pipe != ignore && entries[pipe] &&
4823 skl_ddb_entries_overlap(ddb, entries[pipe]))
4824 return true;
4825 }
4826
4827 return false;
4828 }
4829
skl_update_pipe_wm(struct drm_crtc_state * cstate,const struct skl_pipe_wm * old_pipe_wm,struct skl_pipe_wm * pipe_wm,struct skl_ddb_allocation * ddb,bool * changed)4830 static int skl_update_pipe_wm(struct drm_crtc_state *cstate,
4831 const struct skl_pipe_wm *old_pipe_wm,
4832 struct skl_pipe_wm *pipe_wm, /* out */
4833 struct skl_ddb_allocation *ddb, /* out */
4834 bool *changed /* out */)
4835 {
4836 struct intel_crtc_state *intel_cstate = to_intel_crtc_state(cstate);
4837 int ret;
4838
4839 ret = skl_build_pipe_wm(intel_cstate, ddb, pipe_wm);
4840 if (ret)
4841 return ret;
4842
4843 if (!memcmp(old_pipe_wm, pipe_wm, sizeof(*pipe_wm)))
4844 *changed = false;
4845 else
4846 *changed = true;
4847
4848 return 0;
4849 }
4850
4851 static uint32_t
pipes_modified(struct drm_atomic_state * state)4852 pipes_modified(struct drm_atomic_state *state)
4853 {
4854 struct drm_crtc *crtc;
4855 struct drm_crtc_state *cstate;
4856 uint32_t i, ret = 0;
4857
4858 for_each_new_crtc_in_state(state, crtc, cstate, i)
4859 ret |= drm_crtc_mask(crtc);
4860
4861 return ret;
4862 }
4863
4864 static int
skl_ddb_add_affected_planes(struct intel_crtc_state * cstate)4865 skl_ddb_add_affected_planes(struct intel_crtc_state *cstate)
4866 {
4867 struct drm_atomic_state *state = cstate->base.state;
4868 struct drm_device *dev = state->dev;
4869 struct drm_crtc *crtc = cstate->base.crtc;
4870 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4871 struct drm_i915_private *dev_priv = to_i915(dev);
4872 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
4873 struct skl_ddb_allocation *new_ddb = &intel_state->wm_results.ddb;
4874 struct skl_ddb_allocation *cur_ddb = &dev_priv->wm.skl_hw.ddb;
4875 struct drm_plane_state *plane_state;
4876 struct drm_plane *plane;
4877 enum i915_pipe pipe = intel_crtc->pipe;
4878
4879 WARN_ON(!drm_atomic_get_existing_crtc_state(state, crtc));
4880
4881 drm_for_each_plane_mask(plane, dev, cstate->base.plane_mask) {
4882 enum plane_id plane_id = to_intel_plane(plane)->id;
4883
4884 if (skl_ddb_entry_equal(&cur_ddb->plane[pipe][plane_id],
4885 &new_ddb->plane[pipe][plane_id]) &&
4886 skl_ddb_entry_equal(&cur_ddb->y_plane[pipe][plane_id],
4887 &new_ddb->y_plane[pipe][plane_id]))
4888 continue;
4889
4890 plane_state = drm_atomic_get_plane_state(state, plane);
4891 if (IS_ERR(plane_state))
4892 return PTR_ERR(plane_state);
4893 }
4894
4895 return 0;
4896 }
4897
4898 static int
skl_compute_ddb(struct drm_atomic_state * state)4899 skl_compute_ddb(struct drm_atomic_state *state)
4900 {
4901 struct drm_device *dev = state->dev;
4902 struct drm_i915_private *dev_priv = to_i915(dev);
4903 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
4904 struct intel_crtc *intel_crtc;
4905 struct skl_ddb_allocation *ddb = &intel_state->wm_results.ddb;
4906 uint32_t realloc_pipes = pipes_modified(state);
4907 int ret;
4908
4909 /*
4910 * If this is our first atomic update following hardware readout,
4911 * we can't trust the DDB that the BIOS programmed for us. Let's
4912 * pretend that all pipes switched active status so that we'll
4913 * ensure a full DDB recompute.
4914 */
4915 if (dev_priv->wm.distrust_bios_wm) {
4916 ret = drm_modeset_lock(&dev->mode_config.connection_mutex,
4917 state->acquire_ctx);
4918 if (ret)
4919 return ret;
4920
4921 intel_state->active_pipe_changes = ~0;
4922
4923 /*
4924 * We usually only initialize intel_state->active_crtcs if we
4925 * we're doing a modeset; make sure this field is always
4926 * initialized during the sanitization process that happens
4927 * on the first commit too.
4928 */
4929 if (!intel_state->modeset)
4930 intel_state->active_crtcs = dev_priv->active_crtcs;
4931 }
4932
4933 /*
4934 * If the modeset changes which CRTC's are active, we need to
4935 * recompute the DDB allocation for *all* active pipes, even
4936 * those that weren't otherwise being modified in any way by this
4937 * atomic commit. Due to the shrinking of the per-pipe allocations
4938 * when new active CRTC's are added, it's possible for a pipe that
4939 * we were already using and aren't changing at all here to suddenly
4940 * become invalid if its DDB needs exceeds its new allocation.
4941 *
4942 * Note that if we wind up doing a full DDB recompute, we can't let
4943 * any other display updates race with this transaction, so we need
4944 * to grab the lock on *all* CRTC's.
4945 */
4946 if (intel_state->active_pipe_changes) {
4947 realloc_pipes = ~0;
4948 intel_state->wm_results.dirty_pipes = ~0;
4949 }
4950
4951 /*
4952 * We're not recomputing for the pipes not included in the commit, so
4953 * make sure we start with the current state.
4954 */
4955 memcpy(ddb, &dev_priv->wm.skl_hw.ddb, sizeof(*ddb));
4956
4957 for_each_intel_crtc_mask(dev, intel_crtc, realloc_pipes) {
4958 struct intel_crtc_state *cstate;
4959
4960 cstate = intel_atomic_get_crtc_state(state, intel_crtc);
4961 if (IS_ERR(cstate))
4962 return PTR_ERR(cstate);
4963
4964 ret = skl_allocate_pipe_ddb(cstate, ddb);
4965 if (ret)
4966 return ret;
4967
4968 ret = skl_ddb_add_affected_planes(cstate);
4969 if (ret)
4970 return ret;
4971 }
4972
4973 return 0;
4974 }
4975
4976 static void
skl_copy_wm_for_pipe(struct skl_wm_values * dst,struct skl_wm_values * src,enum i915_pipe pipe)4977 skl_copy_wm_for_pipe(struct skl_wm_values *dst,
4978 struct skl_wm_values *src,
4979 enum i915_pipe pipe)
4980 {
4981 memcpy(dst->ddb.y_plane[pipe], src->ddb.y_plane[pipe],
4982 sizeof(dst->ddb.y_plane[pipe]));
4983 memcpy(dst->ddb.plane[pipe], src->ddb.plane[pipe],
4984 sizeof(dst->ddb.plane[pipe]));
4985 }
4986
4987 static void
skl_print_wm_changes(const struct drm_atomic_state * state)4988 skl_print_wm_changes(const struct drm_atomic_state *state)
4989 {
4990 const struct drm_device *dev = state->dev;
4991 const struct drm_i915_private *dev_priv = to_i915(dev);
4992 const struct intel_atomic_state *intel_state =
4993 to_intel_atomic_state(state);
4994 const struct drm_crtc *crtc;
4995 const struct drm_crtc_state *cstate;
4996 const struct intel_plane *intel_plane;
4997 const struct skl_ddb_allocation *old_ddb = &dev_priv->wm.skl_hw.ddb;
4998 const struct skl_ddb_allocation *new_ddb = &intel_state->wm_results.ddb;
4999 int i;
5000
5001 for_each_new_crtc_in_state(state, crtc, cstate, i) {
5002 const struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
5003 enum i915_pipe pipe = intel_crtc->pipe;
5004
5005 for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
5006 enum plane_id plane_id = intel_plane->id;
5007 const struct skl_ddb_entry *old, *new;
5008
5009 old = &old_ddb->plane[pipe][plane_id];
5010 new = &new_ddb->plane[pipe][plane_id];
5011
5012 if (skl_ddb_entry_equal(old, new))
5013 continue;
5014
5015 DRM_DEBUG_ATOMIC("[PLANE:%d:%s] ddb (%d - %d) -> (%d - %d)\n",
5016 intel_plane->base.base.id,
5017 intel_plane->base.name,
5018 old->start, old->end,
5019 new->start, new->end);
5020 }
5021 }
5022 }
5023
5024 static int
skl_compute_wm(struct drm_atomic_state * state)5025 skl_compute_wm(struct drm_atomic_state *state)
5026 {
5027 struct drm_crtc *crtc;
5028 struct drm_crtc_state *cstate;
5029 struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
5030 struct skl_wm_values *results = &intel_state->wm_results;
5031 struct drm_device *dev = state->dev;
5032 struct skl_pipe_wm *pipe_wm;
5033 bool changed = false;
5034 int ret, i;
5035
5036 /*
5037 * When we distrust bios wm we always need to recompute to set the
5038 * expected DDB allocations for each CRTC.
5039 */
5040 if (to_i915(dev)->wm.distrust_bios_wm)
5041 changed = true;
5042
5043 /*
5044 * If this transaction isn't actually touching any CRTC's, don't
5045 * bother with watermark calculation. Note that if we pass this
5046 * test, we're guaranteed to hold at least one CRTC state mutex,
5047 * which means we can safely use values like dev_priv->active_crtcs
5048 * since any racing commits that want to update them would need to
5049 * hold _all_ CRTC state mutexes.
5050 */
5051 for_each_new_crtc_in_state(state, crtc, cstate, i)
5052 changed = true;
5053
5054 if (!changed)
5055 return 0;
5056
5057 /* Clear all dirty flags */
5058 results->dirty_pipes = 0;
5059
5060 ret = skl_compute_ddb(state);
5061 if (ret)
5062 return ret;
5063
5064 /*
5065 * Calculate WM's for all pipes that are part of this transaction.
5066 * Note that the DDB allocation above may have added more CRTC's that
5067 * weren't otherwise being modified (and set bits in dirty_pipes) if
5068 * pipe allocations had to change.
5069 *
5070 * FIXME: Now that we're doing this in the atomic check phase, we
5071 * should allow skl_update_pipe_wm() to return failure in cases where
5072 * no suitable watermark values can be found.
5073 */
5074 for_each_new_crtc_in_state(state, crtc, cstate, i) {
5075 struct intel_crtc_state *intel_cstate =
5076 to_intel_crtc_state(cstate);
5077 const struct skl_pipe_wm *old_pipe_wm =
5078 &to_intel_crtc_state(crtc->state)->wm.skl.optimal;
5079
5080 pipe_wm = &intel_cstate->wm.skl.optimal;
5081 ret = skl_update_pipe_wm(cstate, old_pipe_wm, pipe_wm,
5082 &results->ddb, &changed);
5083 if (ret)
5084 return ret;
5085
5086 if (changed)
5087 results->dirty_pipes |= drm_crtc_mask(crtc);
5088
5089 if ((results->dirty_pipes & drm_crtc_mask(crtc)) == 0)
5090 /* This pipe's WM's did not change */
5091 continue;
5092
5093 intel_cstate->update_wm_pre = true;
5094 }
5095
5096 skl_print_wm_changes(state);
5097
5098 return 0;
5099 }
5100
skl_atomic_update_crtc_wm(struct intel_atomic_state * state,struct intel_crtc_state * cstate)5101 static void skl_atomic_update_crtc_wm(struct intel_atomic_state *state,
5102 struct intel_crtc_state *cstate)
5103 {
5104 struct intel_crtc *crtc = to_intel_crtc(cstate->base.crtc);
5105 struct drm_i915_private *dev_priv = to_i915(state->base.dev);
5106 struct skl_pipe_wm *pipe_wm = &cstate->wm.skl.optimal;
5107 const struct skl_ddb_allocation *ddb = &state->wm_results.ddb;
5108 enum i915_pipe pipe = crtc->pipe;
5109 enum plane_id plane_id;
5110
5111 if (!(state->wm_results.dirty_pipes & drm_crtc_mask(&crtc->base)))
5112 return;
5113
5114 I915_WRITE(PIPE_WM_LINETIME(pipe), pipe_wm->linetime);
5115
5116 for_each_plane_id_on_crtc(crtc, plane_id) {
5117 if (plane_id != PLANE_CURSOR)
5118 skl_write_plane_wm(crtc, &pipe_wm->planes[plane_id],
5119 ddb, plane_id);
5120 else
5121 skl_write_cursor_wm(crtc, &pipe_wm->planes[plane_id],
5122 ddb);
5123 }
5124 }
5125
skl_initial_wm(struct intel_atomic_state * state,struct intel_crtc_state * cstate)5126 static void skl_initial_wm(struct intel_atomic_state *state,
5127 struct intel_crtc_state *cstate)
5128 {
5129 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
5130 struct drm_device *dev = intel_crtc->base.dev;
5131 struct drm_i915_private *dev_priv = to_i915(dev);
5132 struct skl_wm_values *results = &state->wm_results;
5133 struct skl_wm_values *hw_vals = &dev_priv->wm.skl_hw;
5134 enum i915_pipe pipe = intel_crtc->pipe;
5135
5136 if ((results->dirty_pipes & drm_crtc_mask(&intel_crtc->base)) == 0)
5137 return;
5138
5139 mutex_lock(&dev_priv->wm.wm_mutex);
5140
5141 if (cstate->base.active_changed)
5142 skl_atomic_update_crtc_wm(state, cstate);
5143
5144 skl_copy_wm_for_pipe(hw_vals, results, pipe);
5145
5146 mutex_unlock(&dev_priv->wm.wm_mutex);
5147 }
5148
ilk_compute_wm_config(struct drm_device * dev,struct intel_wm_config * config)5149 static void ilk_compute_wm_config(struct drm_device *dev,
5150 struct intel_wm_config *config)
5151 {
5152 struct intel_crtc *crtc;
5153
5154 /* Compute the currently _active_ config */
5155 for_each_intel_crtc(dev, crtc) {
5156 const struct intel_pipe_wm *wm = &crtc->wm.active.ilk;
5157
5158 if (!wm->pipe_enabled)
5159 continue;
5160
5161 config->sprites_enabled |= wm->sprites_enabled;
5162 config->sprites_scaled |= wm->sprites_scaled;
5163 config->num_pipes_active++;
5164 }
5165 }
5166
ilk_program_watermarks(struct drm_i915_private * dev_priv)5167 static void ilk_program_watermarks(struct drm_i915_private *dev_priv)
5168 {
5169 struct drm_device *dev = &dev_priv->drm;
5170 struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
5171 struct ilk_wm_maximums max;
5172 struct intel_wm_config config = {};
5173 struct ilk_wm_values results = {};
5174 enum intel_ddb_partitioning partitioning;
5175
5176 ilk_compute_wm_config(dev, &config);
5177
5178 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
5179 ilk_wm_merge(dev, &config, &max, &lp_wm_1_2);
5180
5181 /* 5/6 split only in single pipe config on IVB+ */
5182 if (INTEL_GEN(dev_priv) >= 7 &&
5183 config.num_pipes_active == 1 && config.sprites_enabled) {
5184 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
5185 ilk_wm_merge(dev, &config, &max, &lp_wm_5_6);
5186
5187 best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
5188 } else {
5189 best_lp_wm = &lp_wm_1_2;
5190 }
5191
5192 partitioning = (best_lp_wm == &lp_wm_1_2) ?
5193 INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
5194
5195 ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
5196
5197 ilk_write_wm_values(dev_priv, &results);
5198 }
5199
ilk_initial_watermarks(struct intel_atomic_state * state,struct intel_crtc_state * cstate)5200 static void ilk_initial_watermarks(struct intel_atomic_state *state,
5201 struct intel_crtc_state *cstate)
5202 {
5203 struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
5204 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
5205
5206 mutex_lock(&dev_priv->wm.wm_mutex);
5207 intel_crtc->wm.active.ilk = cstate->wm.ilk.intermediate;
5208 ilk_program_watermarks(dev_priv);
5209 mutex_unlock(&dev_priv->wm.wm_mutex);
5210 }
5211
ilk_optimize_watermarks(struct intel_atomic_state * state,struct intel_crtc_state * cstate)5212 static void ilk_optimize_watermarks(struct intel_atomic_state *state,
5213 struct intel_crtc_state *cstate)
5214 {
5215 struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
5216 struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
5217
5218 mutex_lock(&dev_priv->wm.wm_mutex);
5219 if (cstate->wm.need_postvbl_update) {
5220 intel_crtc->wm.active.ilk = cstate->wm.ilk.optimal;
5221 ilk_program_watermarks(dev_priv);
5222 }
5223 mutex_unlock(&dev_priv->wm.wm_mutex);
5224 }
5225
skl_wm_level_from_reg_val(uint32_t val,struct skl_wm_level * level)5226 static inline void skl_wm_level_from_reg_val(uint32_t val,
5227 struct skl_wm_level *level)
5228 {
5229 level->plane_en = val & PLANE_WM_EN;
5230 level->plane_res_b = val & PLANE_WM_BLOCKS_MASK;
5231 level->plane_res_l = (val >> PLANE_WM_LINES_SHIFT) &
5232 PLANE_WM_LINES_MASK;
5233 }
5234
skl_pipe_wm_get_hw_state(struct drm_crtc * crtc,struct skl_pipe_wm * out)5235 void skl_pipe_wm_get_hw_state(struct drm_crtc *crtc,
5236 struct skl_pipe_wm *out)
5237 {
5238 struct drm_i915_private *dev_priv = to_i915(crtc->dev);
5239 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
5240 enum i915_pipe pipe = intel_crtc->pipe;
5241 int level, max_level;
5242 enum plane_id plane_id;
5243 uint32_t val;
5244
5245 max_level = ilk_wm_max_level(dev_priv);
5246
5247 for_each_plane_id_on_crtc(intel_crtc, plane_id) {
5248 struct skl_plane_wm *wm = &out->planes[plane_id];
5249
5250 for (level = 0; level <= max_level; level++) {
5251 if (plane_id != PLANE_CURSOR)
5252 val = I915_READ(PLANE_WM(pipe, plane_id, level));
5253 else
5254 val = I915_READ(CUR_WM(pipe, level));
5255
5256 skl_wm_level_from_reg_val(val, &wm->wm[level]);
5257 }
5258
5259 if (plane_id != PLANE_CURSOR)
5260 val = I915_READ(PLANE_WM_TRANS(pipe, plane_id));
5261 else
5262 val = I915_READ(CUR_WM_TRANS(pipe));
5263
5264 skl_wm_level_from_reg_val(val, &wm->trans_wm);
5265 }
5266
5267 if (!intel_crtc->active)
5268 return;
5269
5270 out->linetime = I915_READ(PIPE_WM_LINETIME(pipe));
5271 }
5272
skl_wm_get_hw_state(struct drm_device * dev)5273 void skl_wm_get_hw_state(struct drm_device *dev)
5274 {
5275 struct drm_i915_private *dev_priv = to_i915(dev);
5276 struct skl_wm_values *hw = &dev_priv->wm.skl_hw;
5277 struct skl_ddb_allocation *ddb = &dev_priv->wm.skl_hw.ddb;
5278 struct drm_crtc *crtc;
5279 struct intel_crtc *intel_crtc;
5280 struct intel_crtc_state *cstate;
5281
5282 skl_ddb_get_hw_state(dev_priv, ddb);
5283 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
5284 intel_crtc = to_intel_crtc(crtc);
5285 cstate = to_intel_crtc_state(crtc->state);
5286
5287 skl_pipe_wm_get_hw_state(crtc, &cstate->wm.skl.optimal);
5288
5289 if (intel_crtc->active)
5290 hw->dirty_pipes |= drm_crtc_mask(crtc);
5291 }
5292
5293 if (dev_priv->active_crtcs) {
5294 /* Fully recompute DDB on first atomic commit */
5295 dev_priv->wm.distrust_bios_wm = true;
5296 } else {
5297 /* Easy/common case; just sanitize DDB now if everything off */
5298 memset(ddb, 0, sizeof(*ddb));
5299 }
5300 }
5301
ilk_pipe_wm_get_hw_state(struct drm_crtc * crtc)5302 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
5303 {
5304 struct drm_device *dev = crtc->dev;
5305 struct drm_i915_private *dev_priv = to_i915(dev);
5306 struct ilk_wm_values *hw = &dev_priv->wm.hw;
5307 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
5308 struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
5309 struct intel_pipe_wm *active = &cstate->wm.ilk.optimal;
5310 enum i915_pipe pipe = intel_crtc->pipe;
5311 static const i915_reg_t wm0_pipe_reg[] = {
5312 [PIPE_A] = WM0_PIPEA_ILK,
5313 [PIPE_B] = WM0_PIPEB_ILK,
5314 [PIPE_C] = WM0_PIPEC_IVB,
5315 };
5316
5317 hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
5318 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
5319 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
5320
5321 memset(active, 0, sizeof(*active));
5322
5323 active->pipe_enabled = intel_crtc->active;
5324
5325 if (active->pipe_enabled) {
5326 u32 tmp = hw->wm_pipe[pipe];
5327
5328 /*
5329 * For active pipes LP0 watermark is marked as
5330 * enabled, and LP1+ watermaks as disabled since
5331 * we can't really reverse compute them in case
5332 * multiple pipes are active.
5333 */
5334 active->wm[0].enable = true;
5335 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
5336 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
5337 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
5338 active->linetime = hw->wm_linetime[pipe];
5339 } else {
5340 int level, max_level = ilk_wm_max_level(dev_priv);
5341
5342 /*
5343 * For inactive pipes, all watermark levels
5344 * should be marked as enabled but zeroed,
5345 * which is what we'd compute them to.
5346 */
5347 for (level = 0; level <= max_level; level++)
5348 active->wm[level].enable = true;
5349 }
5350
5351 intel_crtc->wm.active.ilk = *active;
5352 }
5353
5354 #define _FW_WM(value, plane) \
5355 (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
5356 #define _FW_WM_VLV(value, plane) \
5357 (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
5358
g4x_read_wm_values(struct drm_i915_private * dev_priv,struct g4x_wm_values * wm)5359 static void g4x_read_wm_values(struct drm_i915_private *dev_priv,
5360 struct g4x_wm_values *wm)
5361 {
5362 uint32_t tmp;
5363
5364 tmp = I915_READ(DSPFW1);
5365 wm->sr.plane = _FW_WM(tmp, SR);
5366 wm->pipe[PIPE_B].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORB);
5367 wm->pipe[PIPE_B].plane[PLANE_PRIMARY] = _FW_WM(tmp, PLANEB);
5368 wm->pipe[PIPE_A].plane[PLANE_PRIMARY] = _FW_WM(tmp, PLANEA);
5369
5370 tmp = I915_READ(DSPFW2);
5371 wm->fbc_en = tmp & DSPFW_FBC_SR_EN;
5372 wm->sr.fbc = _FW_WM(tmp, FBC_SR);
5373 wm->hpll.fbc = _FW_WM(tmp, FBC_HPLL_SR);
5374 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM(tmp, SPRITEB);
5375 wm->pipe[PIPE_A].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORA);
5376 wm->pipe[PIPE_A].plane[PLANE_SPRITE0] = _FW_WM(tmp, SPRITEA);
5377
5378 tmp = I915_READ(DSPFW3);
5379 wm->hpll_en = tmp & DSPFW_HPLL_SR_EN;
5380 wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
5381 wm->hpll.cursor = _FW_WM(tmp, HPLL_CURSOR);
5382 wm->hpll.plane = _FW_WM(tmp, HPLL_SR);
5383 }
5384
vlv_read_wm_values(struct drm_i915_private * dev_priv,struct vlv_wm_values * wm)5385 static void vlv_read_wm_values(struct drm_i915_private *dev_priv,
5386 struct vlv_wm_values *wm)
5387 {
5388 enum i915_pipe pipe;
5389 uint32_t tmp;
5390
5391 for_each_pipe(dev_priv, pipe) {
5392 tmp = I915_READ(VLV_DDL(pipe));
5393
5394 wm->ddl[pipe].plane[PLANE_PRIMARY] =
5395 (tmp >> DDL_PLANE_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
5396 wm->ddl[pipe].plane[PLANE_CURSOR] =
5397 (tmp >> DDL_CURSOR_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
5398 wm->ddl[pipe].plane[PLANE_SPRITE0] =
5399 (tmp >> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
5400 wm->ddl[pipe].plane[PLANE_SPRITE1] =
5401 (tmp >> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
5402 }
5403
5404 tmp = I915_READ(DSPFW1);
5405 wm->sr.plane = _FW_WM(tmp, SR);
5406 wm->pipe[PIPE_B].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORB);
5407 wm->pipe[PIPE_B].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEB);
5408 wm->pipe[PIPE_A].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEA);
5409
5410 tmp = I915_READ(DSPFW2);
5411 wm->pipe[PIPE_A].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITEB);
5412 wm->pipe[PIPE_A].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORA);
5413 wm->pipe[PIPE_A].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEA);
5414
5415 tmp = I915_READ(DSPFW3);
5416 wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
5417
5418 if (IS_CHERRYVIEW(dev_priv)) {
5419 tmp = I915_READ(DSPFW7_CHV);
5420 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITED);
5421 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEC);
5422
5423 tmp = I915_READ(DSPFW8_CHV);
5424 wm->pipe[PIPE_C].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITEF);
5425 wm->pipe[PIPE_C].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEE);
5426
5427 tmp = I915_READ(DSPFW9_CHV);
5428 wm->pipe[PIPE_C].plane[PLANE_PRIMARY] = _FW_WM_VLV(tmp, PLANEC);
5429 wm->pipe[PIPE_C].plane[PLANE_CURSOR] = _FW_WM(tmp, CURSORC);
5430
5431 tmp = I915_READ(DSPHOWM);
5432 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
5433 wm->pipe[PIPE_C].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEF_HI) << 8;
5434 wm->pipe[PIPE_C].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEE_HI) << 8;
5435 wm->pipe[PIPE_C].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEC_HI) << 8;
5436 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITED_HI) << 8;
5437 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
5438 wm->pipe[PIPE_B].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEB_HI) << 8;
5439 wm->pipe[PIPE_A].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
5440 wm->pipe[PIPE_A].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
5441 wm->pipe[PIPE_A].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEA_HI) << 8;
5442 } else {
5443 tmp = I915_READ(DSPFW7);
5444 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] = _FW_WM_VLV(tmp, SPRITED);
5445 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] = _FW_WM_VLV(tmp, SPRITEC);
5446
5447 tmp = I915_READ(DSPHOWM);
5448 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
5449 wm->pipe[PIPE_B].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITED_HI) << 8;
5450 wm->pipe[PIPE_B].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
5451 wm->pipe[PIPE_B].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEB_HI) << 8;
5452 wm->pipe[PIPE_A].plane[PLANE_SPRITE1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
5453 wm->pipe[PIPE_A].plane[PLANE_SPRITE0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
5454 wm->pipe[PIPE_A].plane[PLANE_PRIMARY] |= _FW_WM(tmp, PLANEA_HI) << 8;
5455 }
5456 }
5457
5458 #undef _FW_WM
5459 #undef _FW_WM_VLV
5460
g4x_wm_get_hw_state(struct drm_device * dev)5461 void g4x_wm_get_hw_state(struct drm_device *dev)
5462 {
5463 struct drm_i915_private *dev_priv = to_i915(dev);
5464 struct g4x_wm_values *wm = &dev_priv->wm.g4x;
5465 struct intel_crtc *crtc;
5466
5467 g4x_read_wm_values(dev_priv, wm);
5468
5469 wm->cxsr = I915_READ(FW_BLC_SELF) & FW_BLC_SELF_EN;
5470
5471 for_each_intel_crtc(dev, crtc) {
5472 struct intel_crtc_state *crtc_state =
5473 to_intel_crtc_state(crtc->base.state);
5474 struct g4x_wm_state *active = &crtc->wm.active.g4x;
5475 struct g4x_pipe_wm *raw;
5476 enum i915_pipe pipe = crtc->pipe;
5477 enum plane_id plane_id;
5478 int level, max_level;
5479
5480 active->cxsr = wm->cxsr;
5481 active->hpll_en = wm->hpll_en;
5482 active->fbc_en = wm->fbc_en;
5483
5484 active->sr = wm->sr;
5485 active->hpll = wm->hpll;
5486
5487 for_each_plane_id_on_crtc(crtc, plane_id) {
5488 active->wm.plane[plane_id] =
5489 wm->pipe[pipe].plane[plane_id];
5490 }
5491
5492 if (wm->cxsr && wm->hpll_en)
5493 max_level = G4X_WM_LEVEL_HPLL;
5494 else if (wm->cxsr)
5495 max_level = G4X_WM_LEVEL_SR;
5496 else
5497 max_level = G4X_WM_LEVEL_NORMAL;
5498
5499 level = G4X_WM_LEVEL_NORMAL;
5500 raw = &crtc_state->wm.g4x.raw[level];
5501 for_each_plane_id_on_crtc(crtc, plane_id)
5502 raw->plane[plane_id] = active->wm.plane[plane_id];
5503
5504 if (++level > max_level)
5505 goto out;
5506
5507 raw = &crtc_state->wm.g4x.raw[level];
5508 raw->plane[PLANE_PRIMARY] = active->sr.plane;
5509 raw->plane[PLANE_CURSOR] = active->sr.cursor;
5510 raw->plane[PLANE_SPRITE0] = 0;
5511 raw->fbc = active->sr.fbc;
5512
5513 if (++level > max_level)
5514 goto out;
5515
5516 raw = &crtc_state->wm.g4x.raw[level];
5517 raw->plane[PLANE_PRIMARY] = active->hpll.plane;
5518 raw->plane[PLANE_CURSOR] = active->hpll.cursor;
5519 raw->plane[PLANE_SPRITE0] = 0;
5520 raw->fbc = active->hpll.fbc;
5521
5522 out:
5523 for_each_plane_id_on_crtc(crtc, plane_id)
5524 g4x_raw_plane_wm_set(crtc_state, level,
5525 plane_id, USHRT_MAX);
5526 g4x_raw_fbc_wm_set(crtc_state, level, USHRT_MAX);
5527
5528 crtc_state->wm.g4x.optimal = *active;
5529 crtc_state->wm.g4x.intermediate = *active;
5530
5531 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite=%d\n",
5532 pipe_name(pipe),
5533 wm->pipe[pipe].plane[PLANE_PRIMARY],
5534 wm->pipe[pipe].plane[PLANE_CURSOR],
5535 wm->pipe[pipe].plane[PLANE_SPRITE0]);
5536 }
5537
5538 DRM_DEBUG_KMS("Initial SR watermarks: plane=%d, cursor=%d fbc=%d\n",
5539 wm->sr.plane, wm->sr.cursor, wm->sr.fbc);
5540 DRM_DEBUG_KMS("Initial HPLL watermarks: plane=%d, SR cursor=%d fbc=%d\n",
5541 wm->hpll.plane, wm->hpll.cursor, wm->hpll.fbc);
5542 DRM_DEBUG_KMS("Initial SR=%s HPLL=%s FBC=%s\n",
5543 yesno(wm->cxsr), yesno(wm->hpll_en), yesno(wm->fbc_en));
5544 }
5545
g4x_wm_sanitize(struct drm_i915_private * dev_priv)5546 void g4x_wm_sanitize(struct drm_i915_private *dev_priv)
5547 {
5548 struct intel_plane *plane;
5549 struct intel_crtc *crtc;
5550
5551 mutex_lock(&dev_priv->wm.wm_mutex);
5552
5553 for_each_intel_plane(&dev_priv->drm, plane) {
5554 struct intel_crtc *crtc =
5555 intel_get_crtc_for_pipe(dev_priv, plane->pipe);
5556 struct intel_crtc_state *crtc_state =
5557 to_intel_crtc_state(crtc->base.state);
5558 struct intel_plane_state *plane_state =
5559 to_intel_plane_state(plane->base.state);
5560 struct g4x_wm_state *wm_state = &crtc_state->wm.g4x.optimal;
5561 enum plane_id plane_id = plane->id;
5562 int level;
5563
5564 if (plane_state->base.visible)
5565 continue;
5566
5567 for (level = 0; level < 3; level++) {
5568 struct g4x_pipe_wm *raw =
5569 &crtc_state->wm.g4x.raw[level];
5570
5571 raw->plane[plane_id] = 0;
5572 wm_state->wm.plane[plane_id] = 0;
5573 }
5574
5575 if (plane_id == PLANE_PRIMARY) {
5576 for (level = 0; level < 3; level++) {
5577 struct g4x_pipe_wm *raw =
5578 &crtc_state->wm.g4x.raw[level];
5579 raw->fbc = 0;
5580 }
5581
5582 wm_state->sr.fbc = 0;
5583 wm_state->hpll.fbc = 0;
5584 wm_state->fbc_en = false;
5585 }
5586 }
5587
5588 for_each_intel_crtc(&dev_priv->drm, crtc) {
5589 struct intel_crtc_state *crtc_state =
5590 to_intel_crtc_state(crtc->base.state);
5591
5592 crtc_state->wm.g4x.intermediate =
5593 crtc_state->wm.g4x.optimal;
5594 crtc->wm.active.g4x = crtc_state->wm.g4x.optimal;
5595 }
5596
5597 g4x_program_watermarks(dev_priv);
5598
5599 mutex_unlock(&dev_priv->wm.wm_mutex);
5600 }
5601
vlv_wm_get_hw_state(struct drm_device * dev)5602 void vlv_wm_get_hw_state(struct drm_device *dev)
5603 {
5604 struct drm_i915_private *dev_priv = to_i915(dev);
5605 struct vlv_wm_values *wm = &dev_priv->wm.vlv;
5606 struct intel_crtc *crtc;
5607 u32 val;
5608
5609 vlv_read_wm_values(dev_priv, wm);
5610
5611 wm->cxsr = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
5612 wm->level = VLV_WM_LEVEL_PM2;
5613
5614 if (IS_CHERRYVIEW(dev_priv)) {
5615 mutex_lock(&dev_priv->pcu_lock);
5616
5617 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
5618 if (val & DSP_MAXFIFO_PM5_ENABLE)
5619 wm->level = VLV_WM_LEVEL_PM5;
5620
5621 /*
5622 * If DDR DVFS is disabled in the BIOS, Punit
5623 * will never ack the request. So if that happens
5624 * assume we don't have to enable/disable DDR DVFS
5625 * dynamically. To test that just set the REQ_ACK
5626 * bit to poke the Punit, but don't change the
5627 * HIGH/LOW bits so that we don't actually change
5628 * the current state.
5629 */
5630 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
5631 val |= FORCE_DDR_FREQ_REQ_ACK;
5632 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
5633
5634 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
5635 FORCE_DDR_FREQ_REQ_ACK) == 0, 3)) {
5636 DRM_DEBUG_KMS("Punit not acking DDR DVFS request, "
5637 "assuming DDR DVFS is disabled\n");
5638 dev_priv->wm.max_level = VLV_WM_LEVEL_PM5;
5639 } else {
5640 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
5641 if ((val & FORCE_DDR_HIGH_FREQ) == 0)
5642 wm->level = VLV_WM_LEVEL_DDR_DVFS;
5643 }
5644
5645 mutex_unlock(&dev_priv->pcu_lock);
5646 }
5647
5648 for_each_intel_crtc(dev, crtc) {
5649 struct intel_crtc_state *crtc_state =
5650 to_intel_crtc_state(crtc->base.state);
5651 struct vlv_wm_state *active = &crtc->wm.active.vlv;
5652 const struct vlv_fifo_state *fifo_state =
5653 &crtc_state->wm.vlv.fifo_state;
5654 enum i915_pipe pipe = crtc->pipe;
5655 enum plane_id plane_id;
5656 int level;
5657
5658 vlv_get_fifo_size(crtc_state);
5659
5660 active->num_levels = wm->level + 1;
5661 active->cxsr = wm->cxsr;
5662
5663 for (level = 0; level < active->num_levels; level++) {
5664 struct g4x_pipe_wm *raw =
5665 &crtc_state->wm.vlv.raw[level];
5666
5667 active->sr[level].plane = wm->sr.plane;
5668 active->sr[level].cursor = wm->sr.cursor;
5669
5670 for_each_plane_id_on_crtc(crtc, plane_id) {
5671 active->wm[level].plane[plane_id] =
5672 wm->pipe[pipe].plane[plane_id];
5673
5674 raw->plane[plane_id] =
5675 vlv_invert_wm_value(active->wm[level].plane[plane_id],
5676 fifo_state->plane[plane_id]);
5677 }
5678 }
5679
5680 for_each_plane_id_on_crtc(crtc, plane_id)
5681 vlv_raw_plane_wm_set(crtc_state, level,
5682 plane_id, USHRT_MAX);
5683 vlv_invalidate_wms(crtc, active, level);
5684
5685 crtc_state->wm.vlv.optimal = *active;
5686 crtc_state->wm.vlv.intermediate = *active;
5687
5688 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
5689 pipe_name(pipe),
5690 wm->pipe[pipe].plane[PLANE_PRIMARY],
5691 wm->pipe[pipe].plane[PLANE_CURSOR],
5692 wm->pipe[pipe].plane[PLANE_SPRITE0],
5693 wm->pipe[pipe].plane[PLANE_SPRITE1]);
5694 }
5695
5696 DRM_DEBUG_KMS("Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
5697 wm->sr.plane, wm->sr.cursor, wm->level, wm->cxsr);
5698 }
5699
vlv_wm_sanitize(struct drm_i915_private * dev_priv)5700 void vlv_wm_sanitize(struct drm_i915_private *dev_priv)
5701 {
5702 struct intel_plane *plane;
5703 struct intel_crtc *crtc;
5704
5705 mutex_lock(&dev_priv->wm.wm_mutex);
5706
5707 for_each_intel_plane(&dev_priv->drm, plane) {
5708 struct intel_crtc *crtc =
5709 intel_get_crtc_for_pipe(dev_priv, plane->pipe);
5710 struct intel_crtc_state *crtc_state =
5711 to_intel_crtc_state(crtc->base.state);
5712 struct intel_plane_state *plane_state =
5713 to_intel_plane_state(plane->base.state);
5714 struct vlv_wm_state *wm_state = &crtc_state->wm.vlv.optimal;
5715 const struct vlv_fifo_state *fifo_state =
5716 &crtc_state->wm.vlv.fifo_state;
5717 enum plane_id plane_id = plane->id;
5718 int level;
5719
5720 if (plane_state->base.visible)
5721 continue;
5722
5723 for (level = 0; level < wm_state->num_levels; level++) {
5724 struct g4x_pipe_wm *raw =
5725 &crtc_state->wm.vlv.raw[level];
5726
5727 raw->plane[plane_id] = 0;
5728
5729 wm_state->wm[level].plane[plane_id] =
5730 vlv_invert_wm_value(raw->plane[plane_id],
5731 fifo_state->plane[plane_id]);
5732 }
5733 }
5734
5735 for_each_intel_crtc(&dev_priv->drm, crtc) {
5736 struct intel_crtc_state *crtc_state =
5737 to_intel_crtc_state(crtc->base.state);
5738
5739 crtc_state->wm.vlv.intermediate =
5740 crtc_state->wm.vlv.optimal;
5741 crtc->wm.active.vlv = crtc_state->wm.vlv.optimal;
5742 }
5743
5744 vlv_program_watermarks(dev_priv);
5745
5746 mutex_unlock(&dev_priv->wm.wm_mutex);
5747 }
5748
5749 /*
5750 * FIXME should probably kill this and improve
5751 * the real watermark readout/sanitation instead
5752 */
ilk_init_lp_watermarks(struct drm_i915_private * dev_priv)5753 static void ilk_init_lp_watermarks(struct drm_i915_private *dev_priv)
5754 {
5755 I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
5756 I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
5757 I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
5758
5759 /*
5760 * Don't touch WM1S_LP_EN here.
5761 * Doing so could cause underruns.
5762 */
5763 }
5764
ilk_wm_get_hw_state(struct drm_device * dev)5765 void ilk_wm_get_hw_state(struct drm_device *dev)
5766 {
5767 struct drm_i915_private *dev_priv = to_i915(dev);
5768 struct ilk_wm_values *hw = &dev_priv->wm.hw;
5769 struct drm_crtc *crtc;
5770
5771 ilk_init_lp_watermarks(dev_priv);
5772
5773 for_each_crtc(dev, crtc)
5774 ilk_pipe_wm_get_hw_state(crtc);
5775
5776 hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
5777 hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
5778 hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
5779
5780 hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
5781 if (INTEL_GEN(dev_priv) >= 7) {
5782 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
5783 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
5784 }
5785
5786 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
5787 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
5788 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
5789 else if (IS_IVYBRIDGE(dev_priv))
5790 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
5791 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
5792
5793 hw->enable_fbc_wm =
5794 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
5795 }
5796
5797 /**
5798 * intel_update_watermarks - update FIFO watermark values based on current modes
5799 *
5800 * Calculate watermark values for the various WM regs based on current mode
5801 * and plane configuration.
5802 *
5803 * There are several cases to deal with here:
5804 * - normal (i.e. non-self-refresh)
5805 * - self-refresh (SR) mode
5806 * - lines are large relative to FIFO size (buffer can hold up to 2)
5807 * - lines are small relative to FIFO size (buffer can hold more than 2
5808 * lines), so need to account for TLB latency
5809 *
5810 * The normal calculation is:
5811 * watermark = dotclock * bytes per pixel * latency
5812 * where latency is platform & configuration dependent (we assume pessimal
5813 * values here).
5814 *
5815 * The SR calculation is:
5816 * watermark = (trunc(latency/line time)+1) * surface width *
5817 * bytes per pixel
5818 * where
5819 * line time = htotal / dotclock
5820 * surface width = hdisplay for normal plane and 64 for cursor
5821 * and latency is assumed to be high, as above.
5822 *
5823 * The final value programmed to the register should always be rounded up,
5824 * and include an extra 2 entries to account for clock crossings.
5825 *
5826 * We don't use the sprite, so we can ignore that. And on Crestline we have
5827 * to set the non-SR watermarks to 8.
5828 */
intel_update_watermarks(struct intel_crtc * crtc)5829 void intel_update_watermarks(struct intel_crtc *crtc)
5830 {
5831 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
5832
5833 if (dev_priv->display.update_wm)
5834 dev_priv->display.update_wm(crtc);
5835 }
5836
intel_enable_ipc(struct drm_i915_private * dev_priv)5837 void intel_enable_ipc(struct drm_i915_private *dev_priv)
5838 {
5839 u32 val;
5840
5841 /* Display WA #0477 WaDisableIPC: skl */
5842 if (IS_SKYLAKE(dev_priv)) {
5843 dev_priv->ipc_enabled = false;
5844 return;
5845 }
5846
5847 val = I915_READ(DISP_ARB_CTL2);
5848
5849 if (dev_priv->ipc_enabled)
5850 val |= DISP_IPC_ENABLE;
5851 else
5852 val &= ~DISP_IPC_ENABLE;
5853
5854 I915_WRITE(DISP_ARB_CTL2, val);
5855 }
5856
intel_init_ipc(struct drm_i915_private * dev_priv)5857 void intel_init_ipc(struct drm_i915_private *dev_priv)
5858 {
5859 dev_priv->ipc_enabled = false;
5860 if (!HAS_IPC(dev_priv))
5861 return;
5862
5863 dev_priv->ipc_enabled = true;
5864 intel_enable_ipc(dev_priv);
5865 }
5866
5867 /*
5868 * Lock protecting IPS related data structures
5869 */
5870 DEFINE_SPINLOCK(mchdev_lock);
5871
5872 /* Global for IPS driver to get at the current i915 device. Protected by
5873 * mchdev_lock. */
5874 static struct drm_i915_private *i915_mch_dev;
5875
ironlake_set_drps(struct drm_i915_private * dev_priv,u8 val)5876 bool ironlake_set_drps(struct drm_i915_private *dev_priv, u8 val)
5877 {
5878 u16 rgvswctl;
5879
5880 lockdep_assert_held(&mchdev_lock);
5881
5882 rgvswctl = I915_READ16(MEMSWCTL);
5883 if (rgvswctl & MEMCTL_CMD_STS) {
5884 DRM_DEBUG("gpu busy, RCS change rejected\n");
5885 return false; /* still busy with another command */
5886 }
5887
5888 rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
5889 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
5890 I915_WRITE16(MEMSWCTL, rgvswctl);
5891 POSTING_READ16(MEMSWCTL);
5892
5893 rgvswctl |= MEMCTL_CMD_STS;
5894 I915_WRITE16(MEMSWCTL, rgvswctl);
5895
5896 return true;
5897 }
5898
ironlake_enable_drps(struct drm_i915_private * dev_priv)5899 static void ironlake_enable_drps(struct drm_i915_private *dev_priv)
5900 {
5901 u32 rgvmodectl;
5902 u8 fmax, fmin, fstart, vstart;
5903
5904 spin_lock_irq(&mchdev_lock);
5905
5906 rgvmodectl = I915_READ(MEMMODECTL);
5907
5908 /* Enable temp reporting */
5909 I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
5910 I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
5911
5912 /* 100ms RC evaluation intervals */
5913 I915_WRITE(RCUPEI, 100000);
5914 I915_WRITE(RCDNEI, 100000);
5915
5916 /* Set max/min thresholds to 90ms and 80ms respectively */
5917 I915_WRITE(RCBMAXAVG, 90000);
5918 I915_WRITE(RCBMINAVG, 80000);
5919
5920 I915_WRITE(MEMIHYST, 1);
5921
5922 /* Set up min, max, and cur for interrupt handling */
5923 fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
5924 fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
5925 fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
5926 MEMMODE_FSTART_SHIFT;
5927
5928 vstart = (I915_READ(PXVFREQ(fstart)) & PXVFREQ_PX_MASK) >>
5929 PXVFREQ_PX_SHIFT;
5930
5931 dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
5932 dev_priv->ips.fstart = fstart;
5933
5934 dev_priv->ips.max_delay = fstart;
5935 dev_priv->ips.min_delay = fmin;
5936 dev_priv->ips.cur_delay = fstart;
5937
5938 DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
5939 fmax, fmin, fstart);
5940
5941 I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
5942
5943 /*
5944 * Interrupts will be enabled in ironlake_irq_postinstall
5945 */
5946
5947 I915_WRITE(VIDSTART, vstart);
5948 POSTING_READ(VIDSTART);
5949
5950 rgvmodectl |= MEMMODE_SWMODE_EN;
5951 I915_WRITE(MEMMODECTL, rgvmodectl);
5952
5953 if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
5954 DRM_ERROR("stuck trying to change perf mode\n");
5955 mdelay(1);
5956
5957 ironlake_set_drps(dev_priv, fstart);
5958
5959 dev_priv->ips.last_count1 = I915_READ(DMIEC) +
5960 I915_READ(DDREC) + I915_READ(CSIEC);
5961 dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
5962 dev_priv->ips.last_count2 = I915_READ(GFXEC);
5963 dev_priv->ips.last_time2 = ktime_get_raw_ns();
5964
5965 spin_unlock_irq(&mchdev_lock);
5966 }
5967
ironlake_disable_drps(struct drm_i915_private * dev_priv)5968 static void ironlake_disable_drps(struct drm_i915_private *dev_priv)
5969 {
5970 u16 rgvswctl;
5971
5972 spin_lock_irq(&mchdev_lock);
5973
5974 rgvswctl = I915_READ16(MEMSWCTL);
5975
5976 /* Ack interrupts, disable EFC interrupt */
5977 I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
5978 I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
5979 I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
5980 I915_WRITE(DEIIR, DE_PCU_EVENT);
5981 I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
5982
5983 /* Go back to the starting frequency */
5984 ironlake_set_drps(dev_priv, dev_priv->ips.fstart);
5985 mdelay(1);
5986 rgvswctl |= MEMCTL_CMD_STS;
5987 I915_WRITE(MEMSWCTL, rgvswctl);
5988 mdelay(1);
5989
5990 spin_unlock_irq(&mchdev_lock);
5991 }
5992
5993 /* There's a funny hw issue where the hw returns all 0 when reading from
5994 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
5995 * ourselves, instead of doing a rmw cycle (which might result in us clearing
5996 * all limits and the gpu stuck at whatever frequency it is at atm).
5997 */
intel_rps_limits(struct drm_i915_private * dev_priv,u8 val)5998 static u32 intel_rps_limits(struct drm_i915_private *dev_priv, u8 val)
5999 {
6000 struct intel_rps *rps = &dev_priv->gt_pm.rps;
6001 u32 limits;
6002
6003 /* Only set the down limit when we've reached the lowest level to avoid
6004 * getting more interrupts, otherwise leave this clear. This prevents a
6005 * race in the hw when coming out of rc6: There's a tiny window where
6006 * the hw runs at the minimal clock before selecting the desired
6007 * frequency, if the down threshold expires in that window we will not
6008 * receive a down interrupt. */
6009 if (INTEL_GEN(dev_priv) >= 9) {
6010 limits = (rps->max_freq_softlimit) << 23;
6011 if (val <= rps->min_freq_softlimit)
6012 limits |= (rps->min_freq_softlimit) << 14;
6013 } else {
6014 limits = rps->max_freq_softlimit << 24;
6015 if (val <= rps->min_freq_softlimit)
6016 limits |= rps->min_freq_softlimit << 16;
6017 }
6018
6019 return limits;
6020 }
6021
gen6_set_rps_thresholds(struct drm_i915_private * dev_priv,u8 val)6022 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
6023 {
6024 struct intel_rps *rps = &dev_priv->gt_pm.rps;
6025 int new_power;
6026 u32 threshold_up = 0, threshold_down = 0; /* in % */
6027 u32 ei_up = 0, ei_down = 0;
6028
6029 new_power = rps->power;
6030 switch (rps->power) {
6031 case LOW_POWER:
6032 if (val > rps->efficient_freq + 1 &&
6033 val > rps->cur_freq)
6034 new_power = BETWEEN;
6035 break;
6036
6037 case BETWEEN:
6038 if (val <= rps->efficient_freq &&
6039 val < rps->cur_freq)
6040 new_power = LOW_POWER;
6041 else if (val >= rps->rp0_freq &&
6042 val > rps->cur_freq)
6043 new_power = HIGH_POWER;
6044 break;
6045
6046 case HIGH_POWER:
6047 if (val < (rps->rp1_freq + rps->rp0_freq) >> 1 &&
6048 val < rps->cur_freq)
6049 new_power = BETWEEN;
6050 break;
6051 }
6052 /* Max/min bins are special */
6053 if (val <= rps->min_freq_softlimit)
6054 new_power = LOW_POWER;
6055 if (val >= rps->max_freq_softlimit)
6056 new_power = HIGH_POWER;
6057 if (new_power == rps->power)
6058 return;
6059
6060 /* Note the units here are not exactly 1us, but 1280ns. */
6061 switch (new_power) {
6062 case LOW_POWER:
6063 /* Upclock if more than 95% busy over 16ms */
6064 ei_up = 16000;
6065 threshold_up = 95;
6066
6067 /* Downclock if less than 85% busy over 32ms */
6068 ei_down = 32000;
6069 threshold_down = 85;
6070 break;
6071
6072 case BETWEEN:
6073 /* Upclock if more than 90% busy over 13ms */
6074 ei_up = 13000;
6075 threshold_up = 90;
6076
6077 /* Downclock if less than 75% busy over 32ms */
6078 ei_down = 32000;
6079 threshold_down = 75;
6080 break;
6081
6082 case HIGH_POWER:
6083 /* Upclock if more than 85% busy over 10ms */
6084 ei_up = 10000;
6085 threshold_up = 85;
6086
6087 /* Downclock if less than 60% busy over 32ms */
6088 ei_down = 32000;
6089 threshold_down = 60;
6090 break;
6091 }
6092
6093 /* When byt can survive without system hang with dynamic
6094 * sw freq adjustments, this restriction can be lifted.
6095 */
6096 if (IS_VALLEYVIEW(dev_priv))
6097 goto skip_hw_write;
6098
6099 I915_WRITE(GEN6_RP_UP_EI,
6100 GT_INTERVAL_FROM_US(dev_priv, ei_up));
6101 I915_WRITE(GEN6_RP_UP_THRESHOLD,
6102 GT_INTERVAL_FROM_US(dev_priv,
6103 ei_up * threshold_up / 100));
6104
6105 I915_WRITE(GEN6_RP_DOWN_EI,
6106 GT_INTERVAL_FROM_US(dev_priv, ei_down));
6107 I915_WRITE(GEN6_RP_DOWN_THRESHOLD,
6108 GT_INTERVAL_FROM_US(dev_priv,
6109 ei_down * threshold_down / 100));
6110
6111 I915_WRITE(GEN6_RP_CONTROL,
6112 GEN6_RP_MEDIA_TURBO |
6113 GEN6_RP_MEDIA_HW_NORMAL_MODE |
6114 GEN6_RP_MEDIA_IS_GFX |
6115 GEN6_RP_ENABLE |
6116 GEN6_RP_UP_BUSY_AVG |
6117 GEN6_RP_DOWN_IDLE_AVG);
6118
6119 skip_hw_write:
6120 rps->power = new_power;
6121 rps->up_threshold = threshold_up;
6122 rps->down_threshold = threshold_down;
6123 rps->last_adj = 0;
6124 }
6125
gen6_rps_pm_mask(struct drm_i915_private * dev_priv,u8 val)6126 static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
6127 {
6128 struct intel_rps *rps = &dev_priv->gt_pm.rps;
6129 u32 mask = 0;
6130
6131 /* We use UP_EI_EXPIRED interupts for both up/down in manual mode */
6132 if (val > rps->min_freq_softlimit)
6133 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
6134 if (val < rps->max_freq_softlimit)
6135 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD;
6136
6137 mask &= dev_priv->pm_rps_events;
6138
6139 return gen6_sanitize_rps_pm_mask(dev_priv, ~mask);
6140 }
6141
6142 /* gen6_set_rps is called to update the frequency request, but should also be
6143 * called when the range (min_delay and max_delay) is modified so that we can
6144 * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
gen6_set_rps(struct drm_i915_private * dev_priv,u8 val)6145 static int gen6_set_rps(struct drm_i915_private *dev_priv, u8 val)
6146 {
6147 struct intel_rps *rps = &dev_priv->gt_pm.rps;
6148
6149 /* min/max delay may still have been modified so be sure to
6150 * write the limits value.
6151 */
6152 if (val != rps->cur_freq) {
6153 gen6_set_rps_thresholds(dev_priv, val);
6154
6155 if (INTEL_GEN(dev_priv) >= 9)
6156 I915_WRITE(GEN6_RPNSWREQ,
6157 GEN9_FREQUENCY(val));
6158 else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
6159 I915_WRITE(GEN6_RPNSWREQ,
6160 HSW_FREQUENCY(val));
6161 else
6162 I915_WRITE(GEN6_RPNSWREQ,
6163 GEN6_FREQUENCY(val) |
6164 GEN6_OFFSET(0) |
6165 GEN6_AGGRESSIVE_TURBO);
6166 }
6167
6168 /* Make sure we continue to get interrupts
6169 * until we hit the minimum or maximum frequencies.
6170 */
6171 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, intel_rps_limits(dev_priv, val));
6172 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
6173
6174 rps->cur_freq = val;
6175 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
6176
6177 return 0;
6178 }
6179
valleyview_set_rps(struct drm_i915_private * dev_priv,u8 val)6180 static int valleyview_set_rps(struct drm_i915_private *dev_priv, u8 val)
6181 {
6182 int err;
6183
6184 if (WARN_ONCE(IS_CHERRYVIEW(dev_priv) && (val & 1),
6185 "Odd GPU freq value\n"))
6186 val &= ~1;
6187
6188 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
6189
6190 if (val != dev_priv->gt_pm.rps.cur_freq) {
6191 err = vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
6192 if (err)
6193 return err;
6194
6195 gen6_set_rps_thresholds(dev_priv, val);
6196 }
6197
6198 dev_priv->gt_pm.rps.cur_freq = val;
6199 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
6200
6201 return 0;
6202 }
6203
6204 /* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
6205 *
6206 * * If Gfx is Idle, then
6207 * 1. Forcewake Media well.
6208 * 2. Request idle freq.
6209 * 3. Release Forcewake of Media well.
6210 */
vlv_set_rps_idle(struct drm_i915_private * dev_priv)6211 static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
6212 {
6213 struct intel_rps *rps = &dev_priv->gt_pm.rps;
6214 u32 val = rps->idle_freq;
6215 int err;
6216
6217 if (rps->cur_freq <= val)
6218 return;
6219
6220 /* The punit delays the write of the frequency and voltage until it
6221 * determines the GPU is awake. During normal usage we don't want to
6222 * waste power changing the frequency if the GPU is sleeping (rc6).
6223 * However, the GPU and driver is now idle and we do not want to delay
6224 * switching to minimum voltage (reducing power whilst idle) as we do
6225 * not expect to be woken in the near future and so must flush the
6226 * change by waking the device.
6227 *
6228 * We choose to take the media powerwell (either would do to trick the
6229 * punit into committing the voltage change) as that takes a lot less
6230 * power than the render powerwell.
6231 */
6232 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_MEDIA);
6233 err = valleyview_set_rps(dev_priv, val);
6234 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_MEDIA);
6235
6236 if (err)
6237 DRM_ERROR("Failed to set RPS for idle\n");
6238 }
6239
gen6_rps_busy(struct drm_i915_private * dev_priv)6240 void gen6_rps_busy(struct drm_i915_private *dev_priv)
6241 {
6242 struct intel_rps *rps = &dev_priv->gt_pm.rps;
6243
6244 mutex_lock(&dev_priv->pcu_lock);
6245 if (rps->enabled) {
6246 u8 freq;
6247
6248 if (dev_priv->pm_rps_events & GEN6_PM_RP_UP_EI_EXPIRED)
6249 gen6_rps_reset_ei(dev_priv);
6250 I915_WRITE(GEN6_PMINTRMSK,
6251 gen6_rps_pm_mask(dev_priv, rps->cur_freq));
6252
6253 gen6_enable_rps_interrupts(dev_priv);
6254
6255 /* Use the user's desired frequency as a guide, but for better
6256 * performance, jump directly to RPe as our starting frequency.
6257 */
6258 freq = max(rps->cur_freq,
6259 rps->efficient_freq);
6260
6261 if (intel_set_rps(dev_priv,
6262 clamp(freq,
6263 rps->min_freq_softlimit,
6264 rps->max_freq_softlimit)))
6265 DRM_DEBUG_DRIVER("Failed to set idle frequency\n");
6266 }
6267 mutex_unlock(&dev_priv->pcu_lock);
6268 }
6269
gen6_rps_idle(struct drm_i915_private * dev_priv)6270 void gen6_rps_idle(struct drm_i915_private *dev_priv)
6271 {
6272 struct intel_rps *rps = &dev_priv->gt_pm.rps;
6273
6274 /* Flush our bottom-half so that it does not race with us
6275 * setting the idle frequency and so that it is bounded by
6276 * our rpm wakeref. And then disable the interrupts to stop any
6277 * futher RPS reclocking whilst we are asleep.
6278 */
6279 gen6_disable_rps_interrupts(dev_priv);
6280
6281 mutex_lock(&dev_priv->pcu_lock);
6282 if (rps->enabled) {
6283 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
6284 vlv_set_rps_idle(dev_priv);
6285 else
6286 gen6_set_rps(dev_priv, rps->idle_freq);
6287 rps->last_adj = 0;
6288 I915_WRITE(GEN6_PMINTRMSK,
6289 gen6_sanitize_rps_pm_mask(dev_priv, ~0));
6290 }
6291 mutex_unlock(&dev_priv->pcu_lock);
6292 }
6293
gen6_rps_boost(struct drm_i915_gem_request * rq,struct intel_rps_client * rps_client)6294 void gen6_rps_boost(struct drm_i915_gem_request *rq,
6295 struct intel_rps_client *rps_client)
6296 {
6297 struct intel_rps *rps = &rq->i915->gt_pm.rps;
6298 unsigned long flags;
6299 bool boost;
6300
6301 /* This is intentionally racy! We peek at the state here, then
6302 * validate inside the RPS worker.
6303 */
6304 if (!rps->enabled)
6305 return;
6306
6307 boost = false;
6308 spin_lock_irqsave(&rq->lock, flags);
6309 if (!rq->waitboost && !i915_gem_request_completed(rq)) {
6310 atomic_inc(&rps->num_waiters);
6311 rq->waitboost = true;
6312 boost = true;
6313 }
6314 spin_unlock_irqrestore(&rq->lock, flags);
6315 if (!boost)
6316 return;
6317
6318 if (READ_ONCE(rps->cur_freq) < rps->boost_freq)
6319 schedule_work(&rps->work);
6320
6321 atomic_inc(rps_client ? &rps_client->boosts : &rps->boosts);
6322 }
6323
intel_set_rps(struct drm_i915_private * dev_priv,u8 val)6324 int intel_set_rps(struct drm_i915_private *dev_priv, u8 val)
6325 {
6326 struct intel_rps *rps = &dev_priv->gt_pm.rps;
6327 int err;
6328
6329 lockdep_assert_held(&dev_priv->pcu_lock);
6330 GEM_BUG_ON(val > rps->max_freq);
6331 GEM_BUG_ON(val < rps->min_freq);
6332
6333 if (!rps->enabled) {
6334 rps->cur_freq = val;
6335 return 0;
6336 }
6337
6338 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
6339 err = valleyview_set_rps(dev_priv, val);
6340 else
6341 err = gen6_set_rps(dev_priv, val);
6342
6343 return err;
6344 }
6345
gen9_disable_rc6(struct drm_i915_private * dev_priv)6346 static void gen9_disable_rc6(struct drm_i915_private *dev_priv)
6347 {
6348 I915_WRITE(GEN6_RC_CONTROL, 0);
6349 I915_WRITE(GEN9_PG_ENABLE, 0);
6350 }
6351
gen9_disable_rps(struct drm_i915_private * dev_priv)6352 static void gen9_disable_rps(struct drm_i915_private *dev_priv)
6353 {
6354 I915_WRITE(GEN6_RP_CONTROL, 0);
6355 }
6356
gen6_disable_rc6(struct drm_i915_private * dev_priv)6357 static void gen6_disable_rc6(struct drm_i915_private *dev_priv)
6358 {
6359 I915_WRITE(GEN6_RC_CONTROL, 0);
6360 }
6361
gen6_disable_rps(struct drm_i915_private * dev_priv)6362 static void gen6_disable_rps(struct drm_i915_private *dev_priv)
6363 {
6364 I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
6365 I915_WRITE(GEN6_RP_CONTROL, 0);
6366 }
6367
cherryview_disable_rc6(struct drm_i915_private * dev_priv)6368 static void cherryview_disable_rc6(struct drm_i915_private *dev_priv)
6369 {
6370 I915_WRITE(GEN6_RC_CONTROL, 0);
6371 }
6372
cherryview_disable_rps(struct drm_i915_private * dev_priv)6373 static void cherryview_disable_rps(struct drm_i915_private *dev_priv)
6374 {
6375 I915_WRITE(GEN6_RP_CONTROL, 0);
6376 }
6377
valleyview_disable_rc6(struct drm_i915_private * dev_priv)6378 static void valleyview_disable_rc6(struct drm_i915_private *dev_priv)
6379 {
6380 /* We're doing forcewake before Disabling RC6,
6381 * This what the BIOS expects when going into suspend */
6382 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6383
6384 I915_WRITE(GEN6_RC_CONTROL, 0);
6385
6386 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6387 }
6388
valleyview_disable_rps(struct drm_i915_private * dev_priv)6389 static void valleyview_disable_rps(struct drm_i915_private *dev_priv)
6390 {
6391 I915_WRITE(GEN6_RP_CONTROL, 0);
6392 }
6393
intel_print_rc6_info(struct drm_i915_private * dev_priv,u32 mode)6394 static void intel_print_rc6_info(struct drm_i915_private *dev_priv, u32 mode)
6395 {
6396 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
6397 if (mode & (GEN7_RC_CTL_TO_MODE | GEN6_RC_CTL_EI_MODE(1)))
6398 mode = GEN6_RC_CTL_RC6_ENABLE;
6399 else
6400 mode = 0;
6401 }
6402 if (HAS_RC6p(dev_priv))
6403 DRM_DEBUG_DRIVER("Enabling RC6 states: "
6404 "RC6 %s RC6p %s RC6pp %s\n",
6405 onoff(mode & GEN6_RC_CTL_RC6_ENABLE),
6406 onoff(mode & GEN6_RC_CTL_RC6p_ENABLE),
6407 onoff(mode & GEN6_RC_CTL_RC6pp_ENABLE));
6408
6409 else
6410 DRM_DEBUG_DRIVER("Enabling RC6 states: RC6 %s\n",
6411 onoff(mode & GEN6_RC_CTL_RC6_ENABLE));
6412 }
6413
bxt_check_bios_rc6_setup(struct drm_i915_private * dev_priv)6414 static bool bxt_check_bios_rc6_setup(struct drm_i915_private *dev_priv)
6415 {
6416 struct i915_ggtt *ggtt = &dev_priv->ggtt;
6417 bool enable_rc6 = true;
6418 unsigned long rc6_ctx_base;
6419 u32 rc_ctl;
6420 int rc_sw_target;
6421
6422 rc_ctl = I915_READ(GEN6_RC_CONTROL);
6423 rc_sw_target = (I915_READ(GEN6_RC_STATE) & RC_SW_TARGET_STATE_MASK) >>
6424 RC_SW_TARGET_STATE_SHIFT;
6425 DRM_DEBUG_DRIVER("BIOS enabled RC states: "
6426 "HW_CTRL %s HW_RC6 %s SW_TARGET_STATE %x\n",
6427 onoff(rc_ctl & GEN6_RC_CTL_HW_ENABLE),
6428 onoff(rc_ctl & GEN6_RC_CTL_RC6_ENABLE),
6429 rc_sw_target);
6430
6431 if (!(I915_READ(RC6_LOCATION) & RC6_CTX_IN_DRAM)) {
6432 DRM_DEBUG_DRIVER("RC6 Base location not set properly.\n");
6433 enable_rc6 = false;
6434 }
6435
6436 /*
6437 * The exact context size is not known for BXT, so assume a page size
6438 * for this check.
6439 */
6440 rc6_ctx_base = I915_READ(RC6_CTX_BASE) & RC6_CTX_BASE_MASK;
6441 if (!((rc6_ctx_base >= ggtt->stolen_reserved_base) &&
6442 (rc6_ctx_base + PAGE_SIZE <= ggtt->stolen_reserved_base +
6443 ggtt->stolen_reserved_size))) {
6444 DRM_DEBUG_DRIVER("RC6 Base address not as expected.\n");
6445 enable_rc6 = false;
6446 }
6447
6448 if (!(((I915_READ(PWRCTX_MAXCNT_RCSUNIT) & IDLE_TIME_MASK) > 1) &&
6449 ((I915_READ(PWRCTX_MAXCNT_VCSUNIT0) & IDLE_TIME_MASK) > 1) &&
6450 ((I915_READ(PWRCTX_MAXCNT_BCSUNIT) & IDLE_TIME_MASK) > 1) &&
6451 ((I915_READ(PWRCTX_MAXCNT_VECSUNIT) & IDLE_TIME_MASK) > 1))) {
6452 DRM_DEBUG_DRIVER("Engine Idle wait time not set properly.\n");
6453 enable_rc6 = false;
6454 }
6455
6456 if (!I915_READ(GEN8_PUSHBUS_CONTROL) ||
6457 !I915_READ(GEN8_PUSHBUS_ENABLE) ||
6458 !I915_READ(GEN8_PUSHBUS_SHIFT)) {
6459 DRM_DEBUG_DRIVER("Pushbus not setup properly.\n");
6460 enable_rc6 = false;
6461 }
6462
6463 if (!I915_READ(GEN6_GFXPAUSE)) {
6464 DRM_DEBUG_DRIVER("GFX pause not setup properly.\n");
6465 enable_rc6 = false;
6466 }
6467
6468 if (!I915_READ(GEN8_MISC_CTRL0)) {
6469 DRM_DEBUG_DRIVER("GPM control not setup properly.\n");
6470 enable_rc6 = false;
6471 }
6472
6473 return enable_rc6;
6474 }
6475
sanitize_rc6_option(struct drm_i915_private * dev_priv,int enable_rc6)6476 int sanitize_rc6_option(struct drm_i915_private *dev_priv, int enable_rc6)
6477 {
6478 /* No RC6 before Ironlake and code is gone for ilk. */
6479 if (INTEL_INFO(dev_priv)->gen < 6)
6480 return 0;
6481
6482 if (!enable_rc6)
6483 return 0;
6484
6485 if (IS_GEN9_LP(dev_priv) && !bxt_check_bios_rc6_setup(dev_priv)) {
6486 DRM_INFO("RC6 disabled by BIOS\n");
6487 return 0;
6488 }
6489
6490 /* Respect the kernel parameter if it is set */
6491 if (enable_rc6 >= 0) {
6492 int mask;
6493
6494 if (HAS_RC6p(dev_priv))
6495 mask = INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE |
6496 INTEL_RC6pp_ENABLE;
6497 else
6498 mask = INTEL_RC6_ENABLE;
6499
6500 if ((enable_rc6 & mask) != enable_rc6)
6501 DRM_DEBUG_DRIVER("Adjusting RC6 mask to %d "
6502 "(requested %d, valid %d)\n",
6503 enable_rc6 & mask, enable_rc6, mask);
6504
6505 return enable_rc6 & mask;
6506 }
6507
6508 if (IS_IVYBRIDGE(dev_priv))
6509 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
6510
6511 return INTEL_RC6_ENABLE;
6512 }
6513
gen6_init_rps_frequencies(struct drm_i915_private * dev_priv)6514 static void gen6_init_rps_frequencies(struct drm_i915_private *dev_priv)
6515 {
6516 struct intel_rps *rps = &dev_priv->gt_pm.rps;
6517
6518 /* All of these values are in units of 50MHz */
6519
6520 /* static values from HW: RP0 > RP1 > RPn (min_freq) */
6521 if (IS_GEN9_LP(dev_priv)) {
6522 u32 rp_state_cap = I915_READ(BXT_RP_STATE_CAP);
6523 rps->rp0_freq = (rp_state_cap >> 16) & 0xff;
6524 rps->rp1_freq = (rp_state_cap >> 8) & 0xff;
6525 rps->min_freq = (rp_state_cap >> 0) & 0xff;
6526 } else {
6527 u32 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
6528 rps->rp0_freq = (rp_state_cap >> 0) & 0xff;
6529 rps->rp1_freq = (rp_state_cap >> 8) & 0xff;
6530 rps->min_freq = (rp_state_cap >> 16) & 0xff;
6531 }
6532 /* hw_max = RP0 until we check for overclocking */
6533 rps->max_freq = rps->rp0_freq;
6534
6535 rps->efficient_freq = rps->rp1_freq;
6536 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv) ||
6537 IS_GEN9_BC(dev_priv) || IS_CANNONLAKE(dev_priv)) {
6538 u32 ddcc_status = 0;
6539
6540 if (sandybridge_pcode_read(dev_priv,
6541 HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL,
6542 &ddcc_status) == 0)
6543 rps->efficient_freq =
6544 clamp_t(u8,
6545 ((ddcc_status >> 8) & 0xff),
6546 rps->min_freq,
6547 rps->max_freq);
6548 }
6549
6550 if (IS_GEN9_BC(dev_priv) || IS_CANNONLAKE(dev_priv)) {
6551 /* Store the frequency values in 16.66 MHZ units, which is
6552 * the natural hardware unit for SKL
6553 */
6554 rps->rp0_freq *= GEN9_FREQ_SCALER;
6555 rps->rp1_freq *= GEN9_FREQ_SCALER;
6556 rps->min_freq *= GEN9_FREQ_SCALER;
6557 rps->max_freq *= GEN9_FREQ_SCALER;
6558 rps->efficient_freq *= GEN9_FREQ_SCALER;
6559 }
6560 }
6561
reset_rps(struct drm_i915_private * dev_priv,int (* set)(struct drm_i915_private *,u8))6562 static void reset_rps(struct drm_i915_private *dev_priv,
6563 int (*set)(struct drm_i915_private *, u8))
6564 {
6565 struct intel_rps *rps = &dev_priv->gt_pm.rps;
6566 u8 freq = rps->cur_freq;
6567
6568 /* force a reset */
6569 rps->power = -1;
6570 rps->cur_freq = -1;
6571
6572 if (set(dev_priv, freq))
6573 DRM_ERROR("Failed to reset RPS to initial values\n");
6574 }
6575
6576 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
gen9_enable_rps(struct drm_i915_private * dev_priv)6577 static void gen9_enable_rps(struct drm_i915_private *dev_priv)
6578 {
6579 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6580
6581 /* Program defaults and thresholds for RPS*/
6582 I915_WRITE(GEN6_RC_VIDEO_FREQ,
6583 GEN9_FREQUENCY(dev_priv->gt_pm.rps.rp1_freq));
6584
6585 /* 1 second timeout*/
6586 I915_WRITE(GEN6_RP_DOWN_TIMEOUT,
6587 GT_INTERVAL_FROM_US(dev_priv, 1000000));
6588
6589 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 0xa);
6590
6591 /* Leaning on the below call to gen6_set_rps to program/setup the
6592 * Up/Down EI & threshold registers, as well as the RP_CONTROL,
6593 * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
6594 reset_rps(dev_priv, gen6_set_rps);
6595
6596 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6597 }
6598
gen9_enable_rc6(struct drm_i915_private * dev_priv)6599 static void gen9_enable_rc6(struct drm_i915_private *dev_priv)
6600 {
6601 struct intel_engine_cs *engine;
6602 enum intel_engine_id id;
6603 u32 rc6_mode, rc6_mask = 0;
6604
6605 /* 1a: Software RC state - RC0 */
6606 I915_WRITE(GEN6_RC_STATE, 0);
6607
6608 /* 1b: Get forcewake during program sequence. Although the driver
6609 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
6610 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6611
6612 /* 2a: Disable RC states. */
6613 I915_WRITE(GEN6_RC_CONTROL, 0);
6614
6615 /* 2b: Program RC6 thresholds.*/
6616
6617 /* WaRsDoubleRc6WrlWithCoarsePowerGating: Doubling WRL only when CPG is enabled */
6618 if (IS_SKYLAKE(dev_priv))
6619 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 108 << 16);
6620 else
6621 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16);
6622 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
6623 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
6624 for_each_engine(engine, dev_priv, id)
6625 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6626
6627 if (HAS_GUC(dev_priv))
6628 I915_WRITE(GUC_MAX_IDLE_COUNT, 0xA);
6629
6630 I915_WRITE(GEN6_RC_SLEEP, 0);
6631
6632 /* 2c: Program Coarse Power Gating Policies. */
6633 I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS, 25);
6634 I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS, 25);
6635
6636 /* 3a: Enable RC6 */
6637 if (intel_rc6_enabled() & INTEL_RC6_ENABLE)
6638 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
6639 DRM_INFO("RC6 %s\n", onoff(rc6_mask & GEN6_RC_CTL_RC6_ENABLE));
6640 I915_WRITE(GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */
6641
6642 /* WaRsUseTimeoutMode:cnl (pre-prod) */
6643 if (IS_CNL_REVID(dev_priv, CNL_REVID_A0, CNL_REVID_C0))
6644 rc6_mode = GEN7_RC_CTL_TO_MODE;
6645 else
6646 rc6_mode = GEN6_RC_CTL_EI_MODE(1);
6647
6648 I915_WRITE(GEN6_RC_CONTROL,
6649 GEN6_RC_CTL_HW_ENABLE | rc6_mode | rc6_mask);
6650
6651 /*
6652 * 3b: Enable Coarse Power Gating only when RC6 is enabled.
6653 * WaRsDisableCoarsePowerGating:skl,bxt - Render/Media PG need to be disabled with RC6.
6654 */
6655 if (NEEDS_WaRsDisableCoarsePowerGating(dev_priv))
6656 I915_WRITE(GEN9_PG_ENABLE, 0);
6657 else
6658 I915_WRITE(GEN9_PG_ENABLE, (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ?
6659 (GEN9_RENDER_PG_ENABLE | GEN9_MEDIA_PG_ENABLE) : 0);
6660
6661 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6662 }
6663
gen8_enable_rc6(struct drm_i915_private * dev_priv)6664 static void gen8_enable_rc6(struct drm_i915_private *dev_priv)
6665 {
6666 struct intel_engine_cs *engine;
6667 enum intel_engine_id id;
6668 uint32_t rc6_mask = 0;
6669
6670 /* 1a: Software RC state - RC0 */
6671 I915_WRITE(GEN6_RC_STATE, 0);
6672
6673 /* 1b: Get forcewake during program sequence. Although the driver
6674 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
6675 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6676
6677 /* 2a: Disable RC states. */
6678 I915_WRITE(GEN6_RC_CONTROL, 0);
6679
6680 /* 2b: Program RC6 thresholds.*/
6681 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
6682 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
6683 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
6684 for_each_engine(engine, dev_priv, id)
6685 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6686 I915_WRITE(GEN6_RC_SLEEP, 0);
6687 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */
6688
6689 /* 3: Enable RC6 */
6690 if (intel_rc6_enabled() & INTEL_RC6_ENABLE)
6691 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
6692 intel_print_rc6_info(dev_priv, rc6_mask);
6693
6694 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
6695 GEN7_RC_CTL_TO_MODE |
6696 rc6_mask);
6697
6698 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6699 }
6700
gen8_enable_rps(struct drm_i915_private * dev_priv)6701 static void gen8_enable_rps(struct drm_i915_private *dev_priv)
6702 {
6703 struct intel_rps *rps = &dev_priv->gt_pm.rps;
6704
6705 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6706
6707 /* 1 Program defaults and thresholds for RPS*/
6708 I915_WRITE(GEN6_RPNSWREQ,
6709 HSW_FREQUENCY(rps->rp1_freq));
6710 I915_WRITE(GEN6_RC_VIDEO_FREQ,
6711 HSW_FREQUENCY(rps->rp1_freq));
6712 /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
6713 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
6714
6715 /* Docs recommend 900MHz, and 300 MHz respectively */
6716 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
6717 rps->max_freq_softlimit << 24 |
6718 rps->min_freq_softlimit << 16);
6719
6720 I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
6721 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
6722 I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
6723 I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
6724
6725 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
6726
6727 /* 2: Enable RPS */
6728 I915_WRITE(GEN6_RP_CONTROL,
6729 GEN6_RP_MEDIA_TURBO |
6730 GEN6_RP_MEDIA_HW_NORMAL_MODE |
6731 GEN6_RP_MEDIA_IS_GFX |
6732 GEN6_RP_ENABLE |
6733 GEN6_RP_UP_BUSY_AVG |
6734 GEN6_RP_DOWN_IDLE_AVG);
6735
6736 reset_rps(dev_priv, gen6_set_rps);
6737
6738 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6739 }
6740
gen6_enable_rc6(struct drm_i915_private * dev_priv)6741 static void gen6_enable_rc6(struct drm_i915_private *dev_priv)
6742 {
6743 struct intel_engine_cs *engine;
6744 enum intel_engine_id id;
6745 u32 rc6vids, rc6_mask = 0;
6746 u32 gtfifodbg;
6747 int rc6_mode;
6748 int ret;
6749
6750 I915_WRITE(GEN6_RC_STATE, 0);
6751
6752 /* Clear the DBG now so we don't confuse earlier errors */
6753 gtfifodbg = I915_READ(GTFIFODBG);
6754 if (gtfifodbg) {
6755 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
6756 I915_WRITE(GTFIFODBG, gtfifodbg);
6757 }
6758
6759 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6760
6761 /* disable the counters and set deterministic thresholds */
6762 I915_WRITE(GEN6_RC_CONTROL, 0);
6763
6764 I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
6765 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
6766 I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
6767 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
6768 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
6769
6770 for_each_engine(engine, dev_priv, id)
6771 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6772
6773 I915_WRITE(GEN6_RC_SLEEP, 0);
6774 I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
6775 if (IS_IVYBRIDGE(dev_priv))
6776 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
6777 else
6778 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
6779 I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
6780 I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
6781
6782 /* Check if we are enabling RC6 */
6783 rc6_mode = intel_rc6_enabled();
6784 if (rc6_mode & INTEL_RC6_ENABLE)
6785 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
6786
6787 /* We don't use those on Haswell */
6788 if (!IS_HASWELL(dev_priv)) {
6789 if (rc6_mode & INTEL_RC6p_ENABLE)
6790 rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
6791
6792 if (rc6_mode & INTEL_RC6pp_ENABLE)
6793 rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
6794 }
6795
6796 intel_print_rc6_info(dev_priv, rc6_mask);
6797
6798 I915_WRITE(GEN6_RC_CONTROL,
6799 rc6_mask |
6800 GEN6_RC_CTL_EI_MODE(1) |
6801 GEN6_RC_CTL_HW_ENABLE);
6802
6803 rc6vids = 0;
6804 ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
6805 if (IS_GEN6(dev_priv) && ret) {
6806 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
6807 } else if (IS_GEN6(dev_priv) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
6808 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
6809 GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
6810 rc6vids &= 0xffff00;
6811 rc6vids |= GEN6_ENCODE_RC6_VID(450);
6812 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
6813 if (ret)
6814 DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
6815 }
6816
6817 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6818 }
6819
gen6_enable_rps(struct drm_i915_private * dev_priv)6820 static void gen6_enable_rps(struct drm_i915_private *dev_priv)
6821 {
6822 /* Here begins a magic sequence of register writes to enable
6823 * auto-downclocking.
6824 *
6825 * Perhaps there might be some value in exposing these to
6826 * userspace...
6827 */
6828 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6829
6830 /* Power down if completely idle for over 50ms */
6831 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
6832 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
6833
6834 reset_rps(dev_priv, gen6_set_rps);
6835
6836 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6837 }
6838
gen6_update_ring_freq(struct drm_i915_private * dev_priv)6839 static void gen6_update_ring_freq(struct drm_i915_private *dev_priv)
6840 {
6841 struct intel_rps *rps = &dev_priv->gt_pm.rps;
6842 int min_freq = 15;
6843 unsigned int gpu_freq;
6844 unsigned int max_ia_freq, min_ring_freq;
6845 unsigned int max_gpu_freq, min_gpu_freq;
6846 int scaling_factor = 180;
6847
6848 WARN_ON(!mutex_is_locked(&dev_priv->pcu_lock));
6849
6850 #if 0
6851 policy = cpufreq_cpu_get(0);
6852 if (policy) {
6853 max_ia_freq = policy->cpuinfo.max_freq;
6854 cpufreq_cpu_put(policy);
6855 } else {
6856 /*
6857 * Default to measured freq if none found, PCU will ensure we
6858 * don't go over
6859 */
6860 max_ia_freq = tsc_khz;
6861 }
6862 #else
6863 max_ia_freq = tsc_frequency / 1000;
6864 #endif
6865
6866 /* Convert from kHz to MHz */
6867 max_ia_freq /= 1000;
6868
6869 min_ring_freq = I915_READ(DCLK) & 0xf;
6870 /* convert DDR frequency from units of 266.6MHz to bandwidth */
6871 min_ring_freq = mult_frac(min_ring_freq, 8, 3);
6872
6873 if (IS_GEN9_BC(dev_priv) || IS_CANNONLAKE(dev_priv)) {
6874 /* Convert GT frequency to 50 HZ units */
6875 min_gpu_freq = rps->min_freq / GEN9_FREQ_SCALER;
6876 max_gpu_freq = rps->max_freq / GEN9_FREQ_SCALER;
6877 } else {
6878 min_gpu_freq = rps->min_freq;
6879 max_gpu_freq = rps->max_freq;
6880 }
6881
6882 /*
6883 * For each potential GPU frequency, load a ring frequency we'd like
6884 * to use for memory access. We do this by specifying the IA frequency
6885 * the PCU should use as a reference to determine the ring frequency.
6886 */
6887 for (gpu_freq = max_gpu_freq; gpu_freq >= min_gpu_freq; gpu_freq--) {
6888 int diff = max_gpu_freq - gpu_freq;
6889 unsigned int ia_freq = 0, ring_freq = 0;
6890
6891 if (IS_GEN9_BC(dev_priv) || IS_CANNONLAKE(dev_priv)) {
6892 /*
6893 * ring_freq = 2 * GT. ring_freq is in 100MHz units
6894 * No floor required for ring frequency on SKL.
6895 */
6896 ring_freq = gpu_freq;
6897 } else if (INTEL_INFO(dev_priv)->gen >= 8) {
6898 /* max(2 * GT, DDR). NB: GT is 50MHz units */
6899 ring_freq = max(min_ring_freq, gpu_freq);
6900 } else if (IS_HASWELL(dev_priv)) {
6901 ring_freq = mult_frac(gpu_freq, 5, 4);
6902 ring_freq = max(min_ring_freq, ring_freq);
6903 /* leave ia_freq as the default, chosen by cpufreq */
6904 } else {
6905 /* On older processors, there is no separate ring
6906 * clock domain, so in order to boost the bandwidth
6907 * of the ring, we need to upclock the CPU (ia_freq).
6908 *
6909 * For GPU frequencies less than 750MHz,
6910 * just use the lowest ring freq.
6911 */
6912 if (gpu_freq < min_freq)
6913 ia_freq = 800;
6914 else
6915 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
6916 ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
6917 }
6918
6919 sandybridge_pcode_write(dev_priv,
6920 GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
6921 ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
6922 ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
6923 gpu_freq);
6924 }
6925 }
6926
cherryview_rps_max_freq(struct drm_i915_private * dev_priv)6927 static int cherryview_rps_max_freq(struct drm_i915_private *dev_priv)
6928 {
6929 u32 val, rp0;
6930
6931 val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
6932
6933 switch (INTEL_INFO(dev_priv)->sseu.eu_total) {
6934 case 8:
6935 /* (2 * 4) config */
6936 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT);
6937 break;
6938 case 12:
6939 /* (2 * 6) config */
6940 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT);
6941 break;
6942 case 16:
6943 /* (2 * 8) config */
6944 default:
6945 /* Setting (2 * 8) Min RP0 for any other combination */
6946 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT);
6947 break;
6948 }
6949
6950 rp0 = (rp0 & FB_GFX_FREQ_FUSE_MASK);
6951
6952 return rp0;
6953 }
6954
cherryview_rps_rpe_freq(struct drm_i915_private * dev_priv)6955 static int cherryview_rps_rpe_freq(struct drm_i915_private *dev_priv)
6956 {
6957 u32 val, rpe;
6958
6959 val = vlv_punit_read(dev_priv, PUNIT_GPU_DUTYCYCLE_REG);
6960 rpe = (val >> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK;
6961
6962 return rpe;
6963 }
6964
cherryview_rps_guar_freq(struct drm_i915_private * dev_priv)6965 static int cherryview_rps_guar_freq(struct drm_i915_private *dev_priv)
6966 {
6967 u32 val, rp1;
6968
6969 val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
6970 rp1 = (val & FB_GFX_FREQ_FUSE_MASK);
6971
6972 return rp1;
6973 }
6974
cherryview_rps_min_freq(struct drm_i915_private * dev_priv)6975 static u32 cherryview_rps_min_freq(struct drm_i915_private *dev_priv)
6976 {
6977 u32 val, rpn;
6978
6979 val = vlv_punit_read(dev_priv, FB_GFX_FMIN_AT_VMIN_FUSE);
6980 rpn = ((val >> FB_GFX_FMIN_AT_VMIN_FUSE_SHIFT) &
6981 FB_GFX_FREQ_FUSE_MASK);
6982
6983 return rpn;
6984 }
6985
valleyview_rps_guar_freq(struct drm_i915_private * dev_priv)6986 static int valleyview_rps_guar_freq(struct drm_i915_private *dev_priv)
6987 {
6988 u32 val, rp1;
6989
6990 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
6991
6992 rp1 = (val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT;
6993
6994 return rp1;
6995 }
6996
valleyview_rps_max_freq(struct drm_i915_private * dev_priv)6997 static int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
6998 {
6999 u32 val, rp0;
7000
7001 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
7002
7003 rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
7004 /* Clamp to max */
7005 rp0 = min_t(u32, rp0, 0xea);
7006
7007 return rp0;
7008 }
7009
valleyview_rps_rpe_freq(struct drm_i915_private * dev_priv)7010 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
7011 {
7012 u32 val, rpe;
7013
7014 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
7015 rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
7016 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
7017 rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
7018
7019 return rpe;
7020 }
7021
valleyview_rps_min_freq(struct drm_i915_private * dev_priv)7022 static int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
7023 {
7024 u32 val;
7025
7026 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
7027 /*
7028 * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
7029 * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
7030 * a BYT-M B0 the above register contains 0xbf. Moreover when setting
7031 * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
7032 * to make sure it matches what Punit accepts.
7033 */
7034 return max_t(u32, val, 0xc0);
7035 }
7036
7037 /* Check that the pctx buffer wasn't move under us. */
valleyview_check_pctx(struct drm_i915_private * dev_priv)7038 static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
7039 {
7040 unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
7041
7042 WARN_ON(pctx_addr != dev_priv->mm.stolen_base +
7043 dev_priv->vlv_pctx->stolen->start);
7044 }
7045
7046
7047 /* Check that the pcbr address is not empty. */
cherryview_check_pctx(struct drm_i915_private * dev_priv)7048 static void cherryview_check_pctx(struct drm_i915_private *dev_priv)
7049 {
7050 unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
7051
7052 WARN_ON((pctx_addr >> VLV_PCBR_ADDR_SHIFT) == 0);
7053 }
7054
cherryview_setup_pctx(struct drm_i915_private * dev_priv)7055 static void cherryview_setup_pctx(struct drm_i915_private *dev_priv)
7056 {
7057 struct i915_ggtt *ggtt = &dev_priv->ggtt;
7058 unsigned long pctx_paddr, paddr;
7059 u32 pcbr;
7060 int pctx_size = 32*1024;
7061
7062 pcbr = I915_READ(VLV_PCBR);
7063 if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) {
7064 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
7065 paddr = (dev_priv->mm.stolen_base +
7066 (ggtt->stolen_size - pctx_size));
7067
7068 pctx_paddr = (paddr & (~4095));
7069 I915_WRITE(VLV_PCBR, pctx_paddr);
7070 }
7071
7072 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
7073 }
7074
valleyview_setup_pctx(struct drm_i915_private * dev_priv)7075 static void valleyview_setup_pctx(struct drm_i915_private *dev_priv)
7076 {
7077 struct drm_i915_gem_object *pctx;
7078 unsigned long pctx_paddr;
7079 u32 pcbr;
7080 int pctx_size = 24*1024;
7081
7082 pcbr = I915_READ(VLV_PCBR);
7083 if (pcbr) {
7084 /* BIOS set it up already, grab the pre-alloc'd space */
7085 int pcbr_offset;
7086
7087 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
7088 pctx = i915_gem_object_create_stolen_for_preallocated(dev_priv,
7089 pcbr_offset,
7090 I915_GTT_OFFSET_NONE,
7091 pctx_size);
7092 goto out;
7093 }
7094
7095 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
7096
7097 /*
7098 * From the Gunit register HAS:
7099 * The Gfx driver is expected to program this register and ensure
7100 * proper allocation within Gfx stolen memory. For example, this
7101 * register should be programmed such than the PCBR range does not
7102 * overlap with other ranges, such as the frame buffer, protected
7103 * memory, or any other relevant ranges.
7104 */
7105 pctx = i915_gem_object_create_stolen(dev_priv, pctx_size);
7106 if (!pctx) {
7107 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
7108 goto out;
7109 }
7110
7111 pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
7112 I915_WRITE(VLV_PCBR, pctx_paddr);
7113
7114 out:
7115 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
7116 dev_priv->vlv_pctx = pctx;
7117 }
7118
valleyview_cleanup_pctx(struct drm_i915_private * dev_priv)7119 static void valleyview_cleanup_pctx(struct drm_i915_private *dev_priv)
7120 {
7121 if (WARN_ON(!dev_priv->vlv_pctx))
7122 return;
7123
7124 i915_gem_object_put(dev_priv->vlv_pctx);
7125 dev_priv->vlv_pctx = NULL;
7126 }
7127
vlv_init_gpll_ref_freq(struct drm_i915_private * dev_priv)7128 static void vlv_init_gpll_ref_freq(struct drm_i915_private *dev_priv)
7129 {
7130 dev_priv->gt_pm.rps.gpll_ref_freq =
7131 vlv_get_cck_clock(dev_priv, "GPLL ref",
7132 CCK_GPLL_CLOCK_CONTROL,
7133 dev_priv->czclk_freq);
7134
7135 DRM_DEBUG_DRIVER("GPLL reference freq: %d kHz\n",
7136 dev_priv->gt_pm.rps.gpll_ref_freq);
7137 }
7138
valleyview_init_gt_powersave(struct drm_i915_private * dev_priv)7139 static void valleyview_init_gt_powersave(struct drm_i915_private *dev_priv)
7140 {
7141 struct intel_rps *rps = &dev_priv->gt_pm.rps;
7142 u32 val;
7143
7144 valleyview_setup_pctx(dev_priv);
7145
7146 vlv_init_gpll_ref_freq(dev_priv);
7147
7148 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
7149 switch ((val >> 6) & 3) {
7150 case 0:
7151 case 1:
7152 dev_priv->mem_freq = 800;
7153 break;
7154 case 2:
7155 dev_priv->mem_freq = 1066;
7156 break;
7157 case 3:
7158 dev_priv->mem_freq = 1333;
7159 break;
7160 }
7161 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
7162
7163 rps->max_freq = valleyview_rps_max_freq(dev_priv);
7164 rps->rp0_freq = rps->max_freq;
7165 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
7166 intel_gpu_freq(dev_priv, rps->max_freq),
7167 rps->max_freq);
7168
7169 rps->efficient_freq = valleyview_rps_rpe_freq(dev_priv);
7170 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
7171 intel_gpu_freq(dev_priv, rps->efficient_freq),
7172 rps->efficient_freq);
7173
7174 rps->rp1_freq = valleyview_rps_guar_freq(dev_priv);
7175 DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
7176 intel_gpu_freq(dev_priv, rps->rp1_freq),
7177 rps->rp1_freq);
7178
7179 rps->min_freq = valleyview_rps_min_freq(dev_priv);
7180 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
7181 intel_gpu_freq(dev_priv, rps->min_freq),
7182 rps->min_freq);
7183 }
7184
cherryview_init_gt_powersave(struct drm_i915_private * dev_priv)7185 static void cherryview_init_gt_powersave(struct drm_i915_private *dev_priv)
7186 {
7187 struct intel_rps *rps = &dev_priv->gt_pm.rps;
7188 u32 val;
7189
7190 cherryview_setup_pctx(dev_priv);
7191
7192 vlv_init_gpll_ref_freq(dev_priv);
7193
7194 mutex_lock(&dev_priv->sb_lock);
7195 val = vlv_cck_read(dev_priv, CCK_FUSE_REG);
7196 mutex_unlock(&dev_priv->sb_lock);
7197
7198 switch ((val >> 2) & 0x7) {
7199 case 3:
7200 dev_priv->mem_freq = 2000;
7201 break;
7202 default:
7203 dev_priv->mem_freq = 1600;
7204 break;
7205 }
7206 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
7207
7208 rps->max_freq = cherryview_rps_max_freq(dev_priv);
7209 rps->rp0_freq = rps->max_freq;
7210 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
7211 intel_gpu_freq(dev_priv, rps->max_freq),
7212 rps->max_freq);
7213
7214 rps->efficient_freq = cherryview_rps_rpe_freq(dev_priv);
7215 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
7216 intel_gpu_freq(dev_priv, rps->efficient_freq),
7217 rps->efficient_freq);
7218
7219 rps->rp1_freq = cherryview_rps_guar_freq(dev_priv);
7220 DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
7221 intel_gpu_freq(dev_priv, rps->rp1_freq),
7222 rps->rp1_freq);
7223
7224 rps->min_freq = cherryview_rps_min_freq(dev_priv);
7225 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
7226 intel_gpu_freq(dev_priv, rps->min_freq),
7227 rps->min_freq);
7228
7229 WARN_ONCE((rps->max_freq | rps->efficient_freq | rps->rp1_freq |
7230 rps->min_freq) & 1,
7231 "Odd GPU freq values\n");
7232 }
7233
valleyview_cleanup_gt_powersave(struct drm_i915_private * dev_priv)7234 static void valleyview_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
7235 {
7236 valleyview_cleanup_pctx(dev_priv);
7237 }
7238
cherryview_enable_rc6(struct drm_i915_private * dev_priv)7239 static void cherryview_enable_rc6(struct drm_i915_private *dev_priv)
7240 {
7241 struct intel_engine_cs *engine;
7242 enum intel_engine_id id;
7243 u32 gtfifodbg, rc6_mode = 0, pcbr;
7244
7245 gtfifodbg = I915_READ(GTFIFODBG) & ~(GT_FIFO_SBDEDICATE_FREE_ENTRY_CHV |
7246 GT_FIFO_FREE_ENTRIES_CHV);
7247 if (gtfifodbg) {
7248 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
7249 gtfifodbg);
7250 I915_WRITE(GTFIFODBG, gtfifodbg);
7251 }
7252
7253 cherryview_check_pctx(dev_priv);
7254
7255 /* 1a & 1b: Get forcewake during program sequence. Although the driver
7256 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
7257 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
7258
7259 /* Disable RC states. */
7260 I915_WRITE(GEN6_RC_CONTROL, 0);
7261
7262 /* 2a: Program RC6 thresholds.*/
7263 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
7264 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
7265 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
7266
7267 for_each_engine(engine, dev_priv, id)
7268 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
7269 I915_WRITE(GEN6_RC_SLEEP, 0);
7270
7271 /* TO threshold set to 500 us ( 0x186 * 1.28 us) */
7272 I915_WRITE(GEN6_RC6_THRESHOLD, 0x186);
7273
7274 /* Allows RC6 residency counter to work */
7275 I915_WRITE(VLV_COUNTER_CONTROL,
7276 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
7277 VLV_MEDIA_RC6_COUNT_EN |
7278 VLV_RENDER_RC6_COUNT_EN));
7279
7280 /* For now we assume BIOS is allocating and populating the PCBR */
7281 pcbr = I915_READ(VLV_PCBR);
7282
7283 /* 3: Enable RC6 */
7284 if ((intel_rc6_enabled() & INTEL_RC6_ENABLE) &&
7285 (pcbr >> VLV_PCBR_ADDR_SHIFT))
7286 rc6_mode = GEN7_RC_CTL_TO_MODE;
7287
7288 I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
7289
7290 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
7291 }
7292
cherryview_enable_rps(struct drm_i915_private * dev_priv)7293 static void cherryview_enable_rps(struct drm_i915_private *dev_priv)
7294 {
7295 u32 val;
7296
7297 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
7298
7299 /* 1: Program defaults and thresholds for RPS*/
7300 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
7301 I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
7302 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
7303 I915_WRITE(GEN6_RP_UP_EI, 66000);
7304 I915_WRITE(GEN6_RP_DOWN_EI, 350000);
7305
7306 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
7307
7308 /* 2: Enable RPS */
7309 I915_WRITE(GEN6_RP_CONTROL,
7310 GEN6_RP_MEDIA_HW_NORMAL_MODE |
7311 GEN6_RP_MEDIA_IS_GFX |
7312 GEN6_RP_ENABLE |
7313 GEN6_RP_UP_BUSY_AVG |
7314 GEN6_RP_DOWN_IDLE_AVG);
7315
7316 /* Setting Fixed Bias */
7317 val = VLV_OVERRIDE_EN |
7318 VLV_SOC_TDP_EN |
7319 CHV_BIAS_CPU_50_SOC_50;
7320 vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
7321
7322 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
7323
7324 /* RPS code assumes GPLL is used */
7325 WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
7326
7327 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
7328 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
7329
7330 reset_rps(dev_priv, valleyview_set_rps);
7331
7332 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
7333 }
7334
valleyview_enable_rc6(struct drm_i915_private * dev_priv)7335 static void valleyview_enable_rc6(struct drm_i915_private *dev_priv)
7336 {
7337 struct intel_engine_cs *engine;
7338 enum intel_engine_id id;
7339 u32 gtfifodbg, rc6_mode = 0;
7340
7341 valleyview_check_pctx(dev_priv);
7342
7343 gtfifodbg = I915_READ(GTFIFODBG);
7344 if (gtfifodbg) {
7345 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
7346 gtfifodbg);
7347 I915_WRITE(GTFIFODBG, gtfifodbg);
7348 }
7349
7350 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
7351
7352 /* Disable RC states. */
7353 I915_WRITE(GEN6_RC_CONTROL, 0);
7354
7355 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
7356 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
7357 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
7358
7359 for_each_engine(engine, dev_priv, id)
7360 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
7361
7362 I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
7363
7364 /* Allows RC6 residency counter to work */
7365 I915_WRITE(VLV_COUNTER_CONTROL,
7366 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
7367 VLV_MEDIA_RC0_COUNT_EN |
7368 VLV_RENDER_RC0_COUNT_EN |
7369 VLV_MEDIA_RC6_COUNT_EN |
7370 VLV_RENDER_RC6_COUNT_EN));
7371
7372 if (intel_rc6_enabled() & INTEL_RC6_ENABLE)
7373 rc6_mode = GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
7374
7375 intel_print_rc6_info(dev_priv, rc6_mode);
7376
7377 I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
7378
7379 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
7380 }
7381
valleyview_enable_rps(struct drm_i915_private * dev_priv)7382 static void valleyview_enable_rps(struct drm_i915_private *dev_priv)
7383 {
7384 u32 val;
7385
7386 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
7387
7388 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
7389 I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
7390 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
7391 I915_WRITE(GEN6_RP_UP_EI, 66000);
7392 I915_WRITE(GEN6_RP_DOWN_EI, 350000);
7393
7394 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
7395
7396 I915_WRITE(GEN6_RP_CONTROL,
7397 GEN6_RP_MEDIA_TURBO |
7398 GEN6_RP_MEDIA_HW_NORMAL_MODE |
7399 GEN6_RP_MEDIA_IS_GFX |
7400 GEN6_RP_ENABLE |
7401 GEN6_RP_UP_BUSY_AVG |
7402 GEN6_RP_DOWN_IDLE_CONT);
7403
7404 /* Setting Fixed Bias */
7405 val = VLV_OVERRIDE_EN |
7406 VLV_SOC_TDP_EN |
7407 VLV_BIAS_CPU_125_SOC_875;
7408 vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
7409
7410 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
7411
7412 /* RPS code assumes GPLL is used */
7413 WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
7414
7415 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
7416 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
7417
7418 reset_rps(dev_priv, valleyview_set_rps);
7419
7420 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
7421 }
7422
intel_pxfreq(u32 vidfreq)7423 static unsigned long intel_pxfreq(u32 vidfreq)
7424 {
7425 unsigned long freq;
7426 int div = (vidfreq & 0x3f0000) >> 16;
7427 int post = (vidfreq & 0x3000) >> 12;
7428 int pre = (vidfreq & 0x7);
7429
7430 if (!pre)
7431 return 0;
7432
7433 freq = ((div * 133333) / ((1<<post) * pre));
7434
7435 return freq;
7436 }
7437
7438 static const struct cparams {
7439 u16 i;
7440 u16 t;
7441 u16 m;
7442 u16 c;
7443 } cparams[] = {
7444 { 1, 1333, 301, 28664 },
7445 { 1, 1066, 294, 24460 },
7446 { 1, 800, 294, 25192 },
7447 { 0, 1333, 276, 27605 },
7448 { 0, 1066, 276, 27605 },
7449 { 0, 800, 231, 23784 },
7450 };
7451
__i915_chipset_val(struct drm_i915_private * dev_priv)7452 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
7453 {
7454 u64 total_count, diff, ret;
7455 u32 count1, count2, count3, m = 0, c = 0;
7456 unsigned long now = jiffies_to_msecs(jiffies), diff1;
7457 int i;
7458
7459 lockdep_assert_held(&mchdev_lock);
7460
7461 diff1 = now - dev_priv->ips.last_time1;
7462
7463 /* Prevent division-by-zero if we are asking too fast.
7464 * Also, we don't get interesting results if we are polling
7465 * faster than once in 10ms, so just return the saved value
7466 * in such cases.
7467 */
7468 if (diff1 <= 10)
7469 return dev_priv->ips.chipset_power;
7470
7471 count1 = I915_READ(DMIEC);
7472 count2 = I915_READ(DDREC);
7473 count3 = I915_READ(CSIEC);
7474
7475 total_count = count1 + count2 + count3;
7476
7477 /* FIXME: handle per-counter overflow */
7478 if (total_count < dev_priv->ips.last_count1) {
7479 diff = ~0UL - dev_priv->ips.last_count1;
7480 diff += total_count;
7481 } else {
7482 diff = total_count - dev_priv->ips.last_count1;
7483 }
7484
7485 for (i = 0; i < ARRAY_SIZE(cparams); i++) {
7486 if (cparams[i].i == dev_priv->ips.c_m &&
7487 cparams[i].t == dev_priv->ips.r_t) {
7488 m = cparams[i].m;
7489 c = cparams[i].c;
7490 break;
7491 }
7492 }
7493
7494 diff = div_u64(diff, diff1);
7495 ret = ((m * diff) + c);
7496 ret = div_u64(ret, 10);
7497
7498 dev_priv->ips.last_count1 = total_count;
7499 dev_priv->ips.last_time1 = now;
7500
7501 dev_priv->ips.chipset_power = ret;
7502
7503 return ret;
7504 }
7505
i915_chipset_val(struct drm_i915_private * dev_priv)7506 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
7507 {
7508 unsigned long val;
7509
7510 if (INTEL_INFO(dev_priv)->gen != 5)
7511 return 0;
7512
7513 spin_lock_irq(&mchdev_lock);
7514
7515 val = __i915_chipset_val(dev_priv);
7516
7517 spin_unlock_irq(&mchdev_lock);
7518
7519 return val;
7520 }
7521
i915_mch_val(struct drm_i915_private * dev_priv)7522 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
7523 {
7524 unsigned long m, x, b;
7525 u32 tsfs;
7526
7527 tsfs = I915_READ(TSFS);
7528
7529 m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
7530 x = I915_READ8(TR1);
7531
7532 b = tsfs & TSFS_INTR_MASK;
7533
7534 return ((m * x) / 127) - b;
7535 }
7536
_pxvid_to_vd(u8 pxvid)7537 static int _pxvid_to_vd(u8 pxvid)
7538 {
7539 if (pxvid == 0)
7540 return 0;
7541
7542 if (pxvid >= 8 && pxvid < 31)
7543 pxvid = 31;
7544
7545 return (pxvid + 2) * 125;
7546 }
7547
pvid_to_extvid(struct drm_i915_private * dev_priv,u8 pxvid)7548 static u32 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
7549 {
7550 const int vd = _pxvid_to_vd(pxvid);
7551 const int vm = vd - 1125;
7552
7553 if (INTEL_INFO(dev_priv)->is_mobile)
7554 return vm > 0 ? vm : 0;
7555
7556 return vd;
7557 }
7558
__i915_update_gfx_val(struct drm_i915_private * dev_priv)7559 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
7560 {
7561 u64 now, diff, diffms;
7562 u32 count;
7563
7564 lockdep_assert_held(&mchdev_lock);
7565
7566 now = ktime_get_raw_ns();
7567 diffms = now - dev_priv->ips.last_time2;
7568 do_div(diffms, NSEC_PER_MSEC);
7569
7570 /* Don't divide by 0 */
7571 if (!diffms)
7572 return;
7573
7574 count = I915_READ(GFXEC);
7575
7576 if (count < dev_priv->ips.last_count2) {
7577 diff = ~0UL - dev_priv->ips.last_count2;
7578 diff += count;
7579 } else {
7580 diff = count - dev_priv->ips.last_count2;
7581 }
7582
7583 dev_priv->ips.last_count2 = count;
7584 dev_priv->ips.last_time2 = now;
7585
7586 /* More magic constants... */
7587 diff = diff * 1181;
7588 diff = div_u64(diff, diffms * 10);
7589 dev_priv->ips.gfx_power = diff;
7590 }
7591
i915_update_gfx_val(struct drm_i915_private * dev_priv)7592 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
7593 {
7594 if (INTEL_INFO(dev_priv)->gen != 5)
7595 return;
7596
7597 spin_lock_irq(&mchdev_lock);
7598
7599 __i915_update_gfx_val(dev_priv);
7600
7601 spin_unlock_irq(&mchdev_lock);
7602 }
7603
__i915_gfx_val(struct drm_i915_private * dev_priv)7604 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
7605 {
7606 unsigned long t, corr, state1, corr2, state2;
7607 u32 pxvid, ext_v;
7608
7609 lockdep_assert_held(&mchdev_lock);
7610
7611 pxvid = I915_READ(PXVFREQ(dev_priv->gt_pm.rps.cur_freq));
7612 pxvid = (pxvid >> 24) & 0x7f;
7613 ext_v = pvid_to_extvid(dev_priv, pxvid);
7614
7615 state1 = ext_v;
7616
7617 t = i915_mch_val(dev_priv);
7618
7619 /* Revel in the empirically derived constants */
7620
7621 /* Correction factor in 1/100000 units */
7622 if (t > 80)
7623 corr = ((t * 2349) + 135940);
7624 else if (t >= 50)
7625 corr = ((t * 964) + 29317);
7626 else /* < 50 */
7627 corr = ((t * 301) + 1004);
7628
7629 corr = corr * ((150142 * state1) / 10000 - 78642);
7630 corr /= 100000;
7631 corr2 = (corr * dev_priv->ips.corr);
7632
7633 state2 = (corr2 * state1) / 10000;
7634 state2 /= 100; /* convert to mW */
7635
7636 __i915_update_gfx_val(dev_priv);
7637
7638 return dev_priv->ips.gfx_power + state2;
7639 }
7640
i915_gfx_val(struct drm_i915_private * dev_priv)7641 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
7642 {
7643 unsigned long val;
7644
7645 if (INTEL_INFO(dev_priv)->gen != 5)
7646 return 0;
7647
7648 spin_lock_irq(&mchdev_lock);
7649
7650 val = __i915_gfx_val(dev_priv);
7651
7652 spin_unlock_irq(&mchdev_lock);
7653
7654 return val;
7655 }
7656
7657 /**
7658 * i915_read_mch_val - return value for IPS use
7659 *
7660 * Calculate and return a value for the IPS driver to use when deciding whether
7661 * we have thermal and power headroom to increase CPU or GPU power budget.
7662 */
i915_read_mch_val(void)7663 unsigned long i915_read_mch_val(void)
7664 {
7665 struct drm_i915_private *dev_priv;
7666 unsigned long chipset_val, graphics_val, ret = 0;
7667
7668 spin_lock_irq(&mchdev_lock);
7669 if (!i915_mch_dev)
7670 goto out_unlock;
7671 dev_priv = i915_mch_dev;
7672
7673 chipset_val = __i915_chipset_val(dev_priv);
7674 graphics_val = __i915_gfx_val(dev_priv);
7675
7676 ret = chipset_val + graphics_val;
7677
7678 out_unlock:
7679 spin_unlock_irq(&mchdev_lock);
7680
7681 return ret;
7682 }
7683 EXPORT_SYMBOL_GPL(i915_read_mch_val);
7684
7685 /**
7686 * i915_gpu_raise - raise GPU frequency limit
7687 *
7688 * Raise the limit; IPS indicates we have thermal headroom.
7689 */
i915_gpu_raise(void)7690 bool i915_gpu_raise(void)
7691 {
7692 struct drm_i915_private *dev_priv;
7693 bool ret = true;
7694
7695 spin_lock_irq(&mchdev_lock);
7696 if (!i915_mch_dev) {
7697 ret = false;
7698 goto out_unlock;
7699 }
7700 dev_priv = i915_mch_dev;
7701
7702 if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
7703 dev_priv->ips.max_delay--;
7704
7705 out_unlock:
7706 spin_unlock_irq(&mchdev_lock);
7707
7708 return ret;
7709 }
7710 EXPORT_SYMBOL_GPL(i915_gpu_raise);
7711
7712 /**
7713 * i915_gpu_lower - lower GPU frequency limit
7714 *
7715 * IPS indicates we're close to a thermal limit, so throttle back the GPU
7716 * frequency maximum.
7717 */
i915_gpu_lower(void)7718 bool i915_gpu_lower(void)
7719 {
7720 struct drm_i915_private *dev_priv;
7721 bool ret = true;
7722
7723 spin_lock_irq(&mchdev_lock);
7724 if (!i915_mch_dev) {
7725 ret = false;
7726 goto out_unlock;
7727 }
7728 dev_priv = i915_mch_dev;
7729
7730 if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
7731 dev_priv->ips.max_delay++;
7732
7733 out_unlock:
7734 spin_unlock_irq(&mchdev_lock);
7735
7736 return ret;
7737 }
7738 EXPORT_SYMBOL_GPL(i915_gpu_lower);
7739
7740 /**
7741 * i915_gpu_busy - indicate GPU business to IPS
7742 *
7743 * Tell the IPS driver whether or not the GPU is busy.
7744 */
i915_gpu_busy(void)7745 bool i915_gpu_busy(void)
7746 {
7747 bool ret = false;
7748
7749 spin_lock_irq(&mchdev_lock);
7750 if (i915_mch_dev)
7751 ret = i915_mch_dev->gt.awake;
7752 spin_unlock_irq(&mchdev_lock);
7753
7754 return ret;
7755 }
7756 EXPORT_SYMBOL_GPL(i915_gpu_busy);
7757
7758 /**
7759 * i915_gpu_turbo_disable - disable graphics turbo
7760 *
7761 * Disable graphics turbo by resetting the max frequency and setting the
7762 * current frequency to the default.
7763 */
i915_gpu_turbo_disable(void)7764 bool i915_gpu_turbo_disable(void)
7765 {
7766 struct drm_i915_private *dev_priv;
7767 bool ret = true;
7768
7769 spin_lock_irq(&mchdev_lock);
7770 if (!i915_mch_dev) {
7771 ret = false;
7772 goto out_unlock;
7773 }
7774 dev_priv = i915_mch_dev;
7775
7776 dev_priv->ips.max_delay = dev_priv->ips.fstart;
7777
7778 if (!ironlake_set_drps(dev_priv, dev_priv->ips.fstart))
7779 ret = false;
7780
7781 out_unlock:
7782 spin_unlock_irq(&mchdev_lock);
7783
7784 return ret;
7785 }
7786 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
7787
7788 /**
7789 * Tells the intel_ips driver that the i915 driver is now loaded, if
7790 * IPS got loaded first.
7791 *
7792 * This awkward dance is so that neither module has to depend on the
7793 * other in order for IPS to do the appropriate communication of
7794 * GPU turbo limits to i915.
7795 */
7796 static void
ips_ping_for_i915_load(void)7797 ips_ping_for_i915_load(void)
7798 {
7799 #if 0
7800 void (*link)(void);
7801
7802 link = symbol_get(ips_link_to_i915_driver);
7803 if (link) {
7804 link();
7805 symbol_put(ips_link_to_i915_driver);
7806 }
7807 #endif
7808 }
7809
intel_gpu_ips_init(struct drm_i915_private * dev_priv)7810 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
7811 {
7812 /* We only register the i915 ips part with intel-ips once everything is
7813 * set up, to avoid intel-ips sneaking in and reading bogus values. */
7814 spin_lock_irq(&mchdev_lock);
7815 i915_mch_dev = dev_priv;
7816 spin_unlock_irq(&mchdev_lock);
7817
7818 ips_ping_for_i915_load();
7819 }
7820
intel_gpu_ips_teardown(void)7821 void intel_gpu_ips_teardown(void)
7822 {
7823 spin_lock_irq(&mchdev_lock);
7824 i915_mch_dev = NULL;
7825 spin_unlock_irq(&mchdev_lock);
7826 }
7827
intel_init_emon(struct drm_i915_private * dev_priv)7828 static void intel_init_emon(struct drm_i915_private *dev_priv)
7829 {
7830 u32 lcfuse;
7831 u8 pxw[16];
7832 int i;
7833
7834 /* Disable to program */
7835 I915_WRITE(ECR, 0);
7836 POSTING_READ(ECR);
7837
7838 /* Program energy weights for various events */
7839 I915_WRITE(SDEW, 0x15040d00);
7840 I915_WRITE(CSIEW0, 0x007f0000);
7841 I915_WRITE(CSIEW1, 0x1e220004);
7842 I915_WRITE(CSIEW2, 0x04000004);
7843
7844 for (i = 0; i < 5; i++)
7845 I915_WRITE(PEW(i), 0);
7846 for (i = 0; i < 3; i++)
7847 I915_WRITE(DEW(i), 0);
7848
7849 /* Program P-state weights to account for frequency power adjustment */
7850 for (i = 0; i < 16; i++) {
7851 u32 pxvidfreq = I915_READ(PXVFREQ(i));
7852 unsigned long freq = intel_pxfreq(pxvidfreq);
7853 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
7854 PXVFREQ_PX_SHIFT;
7855 unsigned long val;
7856
7857 val = vid * vid;
7858 val *= (freq / 1000);
7859 val *= 255;
7860 val /= (127*127*900);
7861 if (val > 0xff)
7862 DRM_ERROR("bad pxval: %ld\n", val);
7863 pxw[i] = val;
7864 }
7865 /* Render standby states get 0 weight */
7866 pxw[14] = 0;
7867 pxw[15] = 0;
7868
7869 for (i = 0; i < 4; i++) {
7870 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
7871 (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
7872 I915_WRITE(PXW(i), val);
7873 }
7874
7875 /* Adjust magic regs to magic values (more experimental results) */
7876 I915_WRITE(OGW0, 0);
7877 I915_WRITE(OGW1, 0);
7878 I915_WRITE(EG0, 0x00007f00);
7879 I915_WRITE(EG1, 0x0000000e);
7880 I915_WRITE(EG2, 0x000e0000);
7881 I915_WRITE(EG3, 0x68000300);
7882 I915_WRITE(EG4, 0x42000000);
7883 I915_WRITE(EG5, 0x00140031);
7884 I915_WRITE(EG6, 0);
7885 I915_WRITE(EG7, 0);
7886
7887 for (i = 0; i < 8; i++)
7888 I915_WRITE(PXWL(i), 0);
7889
7890 /* Enable PMON + select events */
7891 I915_WRITE(ECR, 0x80000019);
7892
7893 lcfuse = I915_READ(LCFUSE02);
7894
7895 dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
7896 }
7897
intel_init_gt_powersave(struct drm_i915_private * dev_priv)7898 void intel_init_gt_powersave(struct drm_i915_private *dev_priv)
7899 {
7900 struct intel_rps *rps = &dev_priv->gt_pm.rps;
7901
7902 /*
7903 * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
7904 * requirement.
7905 */
7906 if (!i915_modparams.enable_rc6) {
7907 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
7908 intel_runtime_pm_get(dev_priv);
7909 }
7910
7911 mutex_lock(&dev_priv->drm.struct_mutex);
7912 mutex_lock(&dev_priv->pcu_lock);
7913
7914 /* Initialize RPS limits (for userspace) */
7915 if (IS_CHERRYVIEW(dev_priv))
7916 cherryview_init_gt_powersave(dev_priv);
7917 else if (IS_VALLEYVIEW(dev_priv))
7918 valleyview_init_gt_powersave(dev_priv);
7919 else if (INTEL_GEN(dev_priv) >= 6)
7920 gen6_init_rps_frequencies(dev_priv);
7921
7922 /* Derive initial user preferences/limits from the hardware limits */
7923 rps->idle_freq = rps->min_freq;
7924 rps->cur_freq = rps->idle_freq;
7925
7926 rps->max_freq_softlimit = rps->max_freq;
7927 rps->min_freq_softlimit = rps->min_freq;
7928
7929 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
7930 rps->min_freq_softlimit =
7931 max_t(int,
7932 rps->efficient_freq,
7933 intel_freq_opcode(dev_priv, 450));
7934
7935 /* After setting max-softlimit, find the overclock max freq */
7936 if (IS_GEN6(dev_priv) ||
7937 IS_IVYBRIDGE(dev_priv) || IS_HASWELL(dev_priv)) {
7938 u32 params = 0;
7939
7940 sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, ¶ms);
7941 if (params & BIT(31)) { /* OC supported */
7942 DRM_DEBUG_DRIVER("Overclocking supported, max: %dMHz, overclock: %dMHz\n",
7943 (rps->max_freq & 0xff) * 50,
7944 (params & 0xff) * 50);
7945 rps->max_freq = params & 0xff;
7946 }
7947 }
7948
7949 /* Finally allow us to boost to max by default */
7950 rps->boost_freq = rps->max_freq;
7951
7952 mutex_unlock(&dev_priv->pcu_lock);
7953 mutex_unlock(&dev_priv->drm.struct_mutex);
7954
7955 intel_autoenable_gt_powersave(dev_priv);
7956 }
7957
intel_cleanup_gt_powersave(struct drm_i915_private * dev_priv)7958 void intel_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
7959 {
7960 if (IS_VALLEYVIEW(dev_priv))
7961 valleyview_cleanup_gt_powersave(dev_priv);
7962
7963 if (!i915_modparams.enable_rc6)
7964 intel_runtime_pm_put(dev_priv);
7965 }
7966
7967 /**
7968 * intel_suspend_gt_powersave - suspend PM work and helper threads
7969 * @dev_priv: i915 device
7970 *
7971 * We don't want to disable RC6 or other features here, we just want
7972 * to make sure any work we've queued has finished and won't bother
7973 * us while we're suspended.
7974 */
intel_suspend_gt_powersave(struct drm_i915_private * dev_priv)7975 void intel_suspend_gt_powersave(struct drm_i915_private *dev_priv)
7976 {
7977 if (INTEL_GEN(dev_priv) < 6)
7978 return;
7979
7980 if (cancel_delayed_work_sync(&dev_priv->gt_pm.autoenable_work))
7981 intel_runtime_pm_put(dev_priv);
7982
7983 /* gen6_rps_idle() will be called later to disable interrupts */
7984 }
7985
intel_sanitize_gt_powersave(struct drm_i915_private * dev_priv)7986 void intel_sanitize_gt_powersave(struct drm_i915_private *dev_priv)
7987 {
7988 dev_priv->gt_pm.rps.enabled = true; /* force RPS disabling */
7989 dev_priv->gt_pm.rc6.enabled = true; /* force RC6 disabling */
7990 intel_disable_gt_powersave(dev_priv);
7991
7992 gen6_reset_rps_interrupts(dev_priv);
7993 }
7994
intel_disable_llc_pstate(struct drm_i915_private * i915)7995 static inline void intel_disable_llc_pstate(struct drm_i915_private *i915)
7996 {
7997 lockdep_assert_held(&i915->pcu_lock);
7998
7999 if (!i915->gt_pm.llc_pstate.enabled)
8000 return;
8001
8002 /* Currently there is no HW configuration to be done to disable. */
8003
8004 i915->gt_pm.llc_pstate.enabled = false;
8005 }
8006
intel_disable_rc6(struct drm_i915_private * dev_priv)8007 static void intel_disable_rc6(struct drm_i915_private *dev_priv)
8008 {
8009 lockdep_assert_held(&dev_priv->pcu_lock);
8010
8011 if (!dev_priv->gt_pm.rc6.enabled)
8012 return;
8013
8014 if (INTEL_GEN(dev_priv) >= 9)
8015 gen9_disable_rc6(dev_priv);
8016 else if (IS_CHERRYVIEW(dev_priv))
8017 cherryview_disable_rc6(dev_priv);
8018 else if (IS_VALLEYVIEW(dev_priv))
8019 valleyview_disable_rc6(dev_priv);
8020 else if (INTEL_GEN(dev_priv) >= 6)
8021 gen6_disable_rc6(dev_priv);
8022
8023 dev_priv->gt_pm.rc6.enabled = false;
8024 }
8025
intel_disable_rps(struct drm_i915_private * dev_priv)8026 static void intel_disable_rps(struct drm_i915_private *dev_priv)
8027 {
8028 lockdep_assert_held(&dev_priv->pcu_lock);
8029
8030 if (!dev_priv->gt_pm.rps.enabled)
8031 return;
8032
8033 if (INTEL_GEN(dev_priv) >= 9)
8034 gen9_disable_rps(dev_priv);
8035 else if (IS_CHERRYVIEW(dev_priv))
8036 cherryview_disable_rps(dev_priv);
8037 else if (IS_VALLEYVIEW(dev_priv))
8038 valleyview_disable_rps(dev_priv);
8039 else if (INTEL_GEN(dev_priv) >= 6)
8040 gen6_disable_rps(dev_priv);
8041 else if (IS_IRONLAKE_M(dev_priv))
8042 ironlake_disable_drps(dev_priv);
8043
8044 dev_priv->gt_pm.rps.enabled = false;
8045 }
8046
intel_disable_gt_powersave(struct drm_i915_private * dev_priv)8047 void intel_disable_gt_powersave(struct drm_i915_private *dev_priv)
8048 {
8049 mutex_lock(&dev_priv->pcu_lock);
8050
8051 intel_disable_rc6(dev_priv);
8052 intel_disable_rps(dev_priv);
8053 if (HAS_LLC(dev_priv))
8054 intel_disable_llc_pstate(dev_priv);
8055
8056 mutex_unlock(&dev_priv->pcu_lock);
8057 }
8058
intel_enable_llc_pstate(struct drm_i915_private * i915)8059 static inline void intel_enable_llc_pstate(struct drm_i915_private *i915)
8060 {
8061 lockdep_assert_held(&i915->pcu_lock);
8062
8063 if (i915->gt_pm.llc_pstate.enabled)
8064 return;
8065
8066 gen6_update_ring_freq(i915);
8067
8068 i915->gt_pm.llc_pstate.enabled = true;
8069 }
8070
intel_enable_rc6(struct drm_i915_private * dev_priv)8071 static void intel_enable_rc6(struct drm_i915_private *dev_priv)
8072 {
8073 lockdep_assert_held(&dev_priv->pcu_lock);
8074
8075 if (dev_priv->gt_pm.rc6.enabled)
8076 return;
8077
8078 if (IS_CHERRYVIEW(dev_priv))
8079 cherryview_enable_rc6(dev_priv);
8080 else if (IS_VALLEYVIEW(dev_priv))
8081 valleyview_enable_rc6(dev_priv);
8082 else if (INTEL_GEN(dev_priv) >= 9)
8083 gen9_enable_rc6(dev_priv);
8084 else if (IS_BROADWELL(dev_priv))
8085 gen8_enable_rc6(dev_priv);
8086 else if (INTEL_GEN(dev_priv) >= 6)
8087 gen6_enable_rc6(dev_priv);
8088
8089 dev_priv->gt_pm.rc6.enabled = true;
8090 }
8091
intel_enable_rps(struct drm_i915_private * dev_priv)8092 static void intel_enable_rps(struct drm_i915_private *dev_priv)
8093 {
8094 struct intel_rps *rps = &dev_priv->gt_pm.rps;
8095
8096 lockdep_assert_held(&dev_priv->pcu_lock);
8097
8098 if (rps->enabled)
8099 return;
8100
8101 if (IS_CHERRYVIEW(dev_priv)) {
8102 cherryview_enable_rps(dev_priv);
8103 } else if (IS_VALLEYVIEW(dev_priv)) {
8104 valleyview_enable_rps(dev_priv);
8105 } else if (INTEL_GEN(dev_priv) >= 9) {
8106 gen9_enable_rps(dev_priv);
8107 } else if (IS_BROADWELL(dev_priv)) {
8108 gen8_enable_rps(dev_priv);
8109 } else if (INTEL_GEN(dev_priv) >= 6) {
8110 gen6_enable_rps(dev_priv);
8111 } else if (IS_IRONLAKE_M(dev_priv)) {
8112 ironlake_enable_drps(dev_priv);
8113 intel_init_emon(dev_priv);
8114 }
8115
8116 WARN_ON(rps->max_freq < rps->min_freq);
8117 WARN_ON(rps->idle_freq > rps->max_freq);
8118
8119 WARN_ON(rps->efficient_freq < rps->min_freq);
8120 WARN_ON(rps->efficient_freq > rps->max_freq);
8121
8122 rps->enabled = true;
8123 }
8124
intel_enable_gt_powersave(struct drm_i915_private * dev_priv)8125 void intel_enable_gt_powersave(struct drm_i915_private *dev_priv)
8126 {
8127 /* Powersaving is controlled by the host when inside a VM */
8128 if (intel_vgpu_active(dev_priv))
8129 return;
8130
8131 mutex_lock(&dev_priv->pcu_lock);
8132
8133 intel_enable_rc6(dev_priv);
8134 intel_enable_rps(dev_priv);
8135 if (HAS_LLC(dev_priv))
8136 intel_enable_llc_pstate(dev_priv);
8137
8138 mutex_unlock(&dev_priv->pcu_lock);
8139 }
8140
__intel_autoenable_gt_powersave(struct work_struct * work)8141 static void __intel_autoenable_gt_powersave(struct work_struct *work)
8142 {
8143 struct drm_i915_private *dev_priv =
8144 container_of(work,
8145 typeof(*dev_priv),
8146 gt_pm.autoenable_work.work);
8147 struct intel_engine_cs *rcs;
8148 struct drm_i915_gem_request *req;
8149
8150 rcs = dev_priv->engine[RCS];
8151 if (rcs->last_retired_context)
8152 goto out;
8153
8154 if (!rcs->init_context)
8155 goto out;
8156
8157 mutex_lock(&dev_priv->drm.struct_mutex);
8158
8159 req = i915_gem_request_alloc(rcs, dev_priv->kernel_context);
8160 if (IS_ERR(req))
8161 goto unlock;
8162
8163 if (!i915_modparams.enable_execlists && i915_switch_context(req) == 0)
8164 rcs->init_context(req);
8165
8166 /* Mark the device busy, calling intel_enable_gt_powersave() */
8167 i915_add_request(req);
8168
8169 unlock:
8170 mutex_unlock(&dev_priv->drm.struct_mutex);
8171 out:
8172 intel_runtime_pm_put(dev_priv);
8173 }
8174
intel_autoenable_gt_powersave(struct drm_i915_private * dev_priv)8175 void intel_autoenable_gt_powersave(struct drm_i915_private *dev_priv)
8176 {
8177 if (IS_IRONLAKE_M(dev_priv)) {
8178 ironlake_enable_drps(dev_priv);
8179 intel_init_emon(dev_priv);
8180 } else if (INTEL_INFO(dev_priv)->gen >= 6) {
8181 /*
8182 * PCU communication is slow and this doesn't need to be
8183 * done at any specific time, so do this out of our fast path
8184 * to make resume and init faster.
8185 *
8186 * We depend on the HW RC6 power context save/restore
8187 * mechanism when entering D3 through runtime PM suspend. So
8188 * disable RPM until RPS/RC6 is properly setup. We can only
8189 * get here via the driver load/system resume/runtime resume
8190 * paths, so the _noresume version is enough (and in case of
8191 * runtime resume it's necessary).
8192 */
8193 if (queue_delayed_work(dev_priv->wq,
8194 &dev_priv->gt_pm.autoenable_work,
8195 round_jiffies_up_relative(HZ)))
8196 intel_runtime_pm_get_noresume(dev_priv);
8197 }
8198 }
8199
ibx_init_clock_gating(struct drm_i915_private * dev_priv)8200 static void ibx_init_clock_gating(struct drm_i915_private *dev_priv)
8201 {
8202 /*
8203 * On Ibex Peak and Cougar Point, we need to disable clock
8204 * gating for the panel power sequencer or it will fail to
8205 * start up when no ports are active.
8206 */
8207 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
8208 }
8209
g4x_disable_trickle_feed(struct drm_i915_private * dev_priv)8210 static void g4x_disable_trickle_feed(struct drm_i915_private *dev_priv)
8211 {
8212 enum i915_pipe pipe;
8213
8214 for_each_pipe(dev_priv, pipe) {
8215 I915_WRITE(DSPCNTR(pipe),
8216 I915_READ(DSPCNTR(pipe)) |
8217 DISPPLANE_TRICKLE_FEED_DISABLE);
8218
8219 I915_WRITE(DSPSURF(pipe), I915_READ(DSPSURF(pipe)));
8220 POSTING_READ(DSPSURF(pipe));
8221 }
8222 }
8223
ilk_init_clock_gating(struct drm_i915_private * dev_priv)8224 static void ilk_init_clock_gating(struct drm_i915_private *dev_priv)
8225 {
8226 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
8227
8228 /*
8229 * Required for FBC
8230 * WaFbcDisableDpfcClockGating:ilk
8231 */
8232 dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
8233 ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
8234 ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
8235
8236 I915_WRITE(PCH_3DCGDIS0,
8237 MARIUNIT_CLOCK_GATE_DISABLE |
8238 SVSMUNIT_CLOCK_GATE_DISABLE);
8239 I915_WRITE(PCH_3DCGDIS1,
8240 VFMUNIT_CLOCK_GATE_DISABLE);
8241
8242 /*
8243 * According to the spec the following bits should be set in
8244 * order to enable memory self-refresh
8245 * The bit 22/21 of 0x42004
8246 * The bit 5 of 0x42020
8247 * The bit 15 of 0x45000
8248 */
8249 I915_WRITE(ILK_DISPLAY_CHICKEN2,
8250 (I915_READ(ILK_DISPLAY_CHICKEN2) |
8251 ILK_DPARB_GATE | ILK_VSDPFD_FULL));
8252 dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
8253 I915_WRITE(DISP_ARB_CTL,
8254 (I915_READ(DISP_ARB_CTL) |
8255 DISP_FBC_WM_DIS));
8256
8257 /*
8258 * Based on the document from hardware guys the following bits
8259 * should be set unconditionally in order to enable FBC.
8260 * The bit 22 of 0x42000
8261 * The bit 22 of 0x42004
8262 * The bit 7,8,9 of 0x42020.
8263 */
8264 if (IS_IRONLAKE_M(dev_priv)) {
8265 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
8266 I915_WRITE(ILK_DISPLAY_CHICKEN1,
8267 I915_READ(ILK_DISPLAY_CHICKEN1) |
8268 ILK_FBCQ_DIS);
8269 I915_WRITE(ILK_DISPLAY_CHICKEN2,
8270 I915_READ(ILK_DISPLAY_CHICKEN2) |
8271 ILK_DPARB_GATE);
8272 }
8273
8274 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
8275
8276 I915_WRITE(ILK_DISPLAY_CHICKEN2,
8277 I915_READ(ILK_DISPLAY_CHICKEN2) |
8278 ILK_ELPIN_409_SELECT);
8279 I915_WRITE(_3D_CHICKEN2,
8280 _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
8281 _3D_CHICKEN2_WM_READ_PIPELINED);
8282
8283 /* WaDisableRenderCachePipelinedFlush:ilk */
8284 I915_WRITE(CACHE_MODE_0,
8285 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
8286
8287 /* WaDisable_RenderCache_OperationalFlush:ilk */
8288 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8289
8290 g4x_disable_trickle_feed(dev_priv);
8291
8292 ibx_init_clock_gating(dev_priv);
8293 }
8294
cpt_init_clock_gating(struct drm_i915_private * dev_priv)8295 static void cpt_init_clock_gating(struct drm_i915_private *dev_priv)
8296 {
8297 int pipe;
8298 uint32_t val;
8299
8300 /*
8301 * On Ibex Peak and Cougar Point, we need to disable clock
8302 * gating for the panel power sequencer or it will fail to
8303 * start up when no ports are active.
8304 */
8305 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
8306 PCH_DPLUNIT_CLOCK_GATE_DISABLE |
8307 PCH_CPUNIT_CLOCK_GATE_DISABLE);
8308 I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
8309 DPLS_EDP_PPS_FIX_DIS);
8310 /* The below fixes the weird display corruption, a few pixels shifted
8311 * downward, on (only) LVDS of some HP laptops with IVY.
8312 */
8313 for_each_pipe(dev_priv, pipe) {
8314 val = I915_READ(TRANS_CHICKEN2(pipe));
8315 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
8316 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
8317 if (dev_priv->vbt.fdi_rx_polarity_inverted)
8318 val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
8319 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
8320 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
8321 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
8322 I915_WRITE(TRANS_CHICKEN2(pipe), val);
8323 }
8324 /* WADP0ClockGatingDisable */
8325 for_each_pipe(dev_priv, pipe) {
8326 I915_WRITE(TRANS_CHICKEN1(pipe),
8327 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
8328 }
8329 }
8330
gen6_check_mch_setup(struct drm_i915_private * dev_priv)8331 static void gen6_check_mch_setup(struct drm_i915_private *dev_priv)
8332 {
8333 uint32_t tmp;
8334
8335 tmp = I915_READ(MCH_SSKPD);
8336 if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL)
8337 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
8338 tmp);
8339 }
8340
gen6_init_clock_gating(struct drm_i915_private * dev_priv)8341 static void gen6_init_clock_gating(struct drm_i915_private *dev_priv)
8342 {
8343 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
8344
8345 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
8346
8347 I915_WRITE(ILK_DISPLAY_CHICKEN2,
8348 I915_READ(ILK_DISPLAY_CHICKEN2) |
8349 ILK_ELPIN_409_SELECT);
8350
8351 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
8352 I915_WRITE(_3D_CHICKEN,
8353 _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
8354
8355 /* WaDisable_RenderCache_OperationalFlush:snb */
8356 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8357
8358 /*
8359 * BSpec recoomends 8x4 when MSAA is used,
8360 * however in practice 16x4 seems fastest.
8361 *
8362 * Note that PS/WM thread counts depend on the WIZ hashing
8363 * disable bit, which we don't touch here, but it's good
8364 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
8365 */
8366 I915_WRITE(GEN6_GT_MODE,
8367 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
8368
8369 I915_WRITE(CACHE_MODE_0,
8370 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
8371
8372 I915_WRITE(GEN6_UCGCTL1,
8373 I915_READ(GEN6_UCGCTL1) |
8374 GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
8375 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
8376
8377 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
8378 * gating disable must be set. Failure to set it results in
8379 * flickering pixels due to Z write ordering failures after
8380 * some amount of runtime in the Mesa "fire" demo, and Unigine
8381 * Sanctuary and Tropics, and apparently anything else with
8382 * alpha test or pixel discard.
8383 *
8384 * According to the spec, bit 11 (RCCUNIT) must also be set,
8385 * but we didn't debug actual testcases to find it out.
8386 *
8387 * WaDisableRCCUnitClockGating:snb
8388 * WaDisableRCPBUnitClockGating:snb
8389 */
8390 I915_WRITE(GEN6_UCGCTL2,
8391 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
8392 GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
8393
8394 /* WaStripsFansDisableFastClipPerformanceFix:snb */
8395 I915_WRITE(_3D_CHICKEN3,
8396 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
8397
8398 /*
8399 * Bspec says:
8400 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
8401 * 3DSTATE_SF number of SF output attributes is more than 16."
8402 */
8403 I915_WRITE(_3D_CHICKEN3,
8404 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
8405
8406 /*
8407 * According to the spec the following bits should be
8408 * set in order to enable memory self-refresh and fbc:
8409 * The bit21 and bit22 of 0x42000
8410 * The bit21 and bit22 of 0x42004
8411 * The bit5 and bit7 of 0x42020
8412 * The bit14 of 0x70180
8413 * The bit14 of 0x71180
8414 *
8415 * WaFbcAsynchFlipDisableFbcQueue:snb
8416 */
8417 I915_WRITE(ILK_DISPLAY_CHICKEN1,
8418 I915_READ(ILK_DISPLAY_CHICKEN1) |
8419 ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
8420 I915_WRITE(ILK_DISPLAY_CHICKEN2,
8421 I915_READ(ILK_DISPLAY_CHICKEN2) |
8422 ILK_DPARB_GATE | ILK_VSDPFD_FULL);
8423 I915_WRITE(ILK_DSPCLK_GATE_D,
8424 I915_READ(ILK_DSPCLK_GATE_D) |
8425 ILK_DPARBUNIT_CLOCK_GATE_ENABLE |
8426 ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
8427
8428 g4x_disable_trickle_feed(dev_priv);
8429
8430 cpt_init_clock_gating(dev_priv);
8431
8432 gen6_check_mch_setup(dev_priv);
8433 }
8434
gen7_setup_fixed_func_scheduler(struct drm_i915_private * dev_priv)8435 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
8436 {
8437 uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
8438
8439 /*
8440 * WaVSThreadDispatchOverride:ivb,vlv
8441 *
8442 * This actually overrides the dispatch
8443 * mode for all thread types.
8444 */
8445 reg &= ~GEN7_FF_SCHED_MASK;
8446 reg |= GEN7_FF_TS_SCHED_HW;
8447 reg |= GEN7_FF_VS_SCHED_HW;
8448 reg |= GEN7_FF_DS_SCHED_HW;
8449
8450 I915_WRITE(GEN7_FF_THREAD_MODE, reg);
8451 }
8452
lpt_init_clock_gating(struct drm_i915_private * dev_priv)8453 static void lpt_init_clock_gating(struct drm_i915_private *dev_priv)
8454 {
8455 /*
8456 * TODO: this bit should only be enabled when really needed, then
8457 * disabled when not needed anymore in order to save power.
8458 */
8459 if (HAS_PCH_LPT_LP(dev_priv))
8460 I915_WRITE(SOUTH_DSPCLK_GATE_D,
8461 I915_READ(SOUTH_DSPCLK_GATE_D) |
8462 PCH_LP_PARTITION_LEVEL_DISABLE);
8463
8464 /* WADPOClockGatingDisable:hsw */
8465 I915_WRITE(TRANS_CHICKEN1(PIPE_A),
8466 I915_READ(TRANS_CHICKEN1(PIPE_A)) |
8467 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
8468 }
8469
lpt_suspend_hw(struct drm_i915_private * dev_priv)8470 static void lpt_suspend_hw(struct drm_i915_private *dev_priv)
8471 {
8472 if (HAS_PCH_LPT_LP(dev_priv)) {
8473 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
8474
8475 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
8476 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
8477 }
8478 }
8479
gen8_set_l3sqc_credits(struct drm_i915_private * dev_priv,int general_prio_credits,int high_prio_credits)8480 static void gen8_set_l3sqc_credits(struct drm_i915_private *dev_priv,
8481 int general_prio_credits,
8482 int high_prio_credits)
8483 {
8484 u32 misccpctl;
8485 u32 val;
8486
8487 /* WaTempDisableDOPClkGating:bdw */
8488 misccpctl = I915_READ(GEN7_MISCCPCTL);
8489 I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
8490
8491 val = I915_READ(GEN8_L3SQCREG1);
8492 val &= ~L3_PRIO_CREDITS_MASK;
8493 val |= L3_GENERAL_PRIO_CREDITS(general_prio_credits);
8494 val |= L3_HIGH_PRIO_CREDITS(high_prio_credits);
8495 I915_WRITE(GEN8_L3SQCREG1, val);
8496
8497 /*
8498 * Wait at least 100 clocks before re-enabling clock gating.
8499 * See the definition of L3SQCREG1 in BSpec.
8500 */
8501 POSTING_READ(GEN8_L3SQCREG1);
8502 udelay(1);
8503 I915_WRITE(GEN7_MISCCPCTL, misccpctl);
8504 }
8505
cnp_init_clock_gating(struct drm_i915_private * dev_priv)8506 static void cnp_init_clock_gating(struct drm_i915_private *dev_priv)
8507 {
8508 if (!HAS_PCH_CNP(dev_priv))
8509 return;
8510
8511 /* Wa #1181 */
8512 I915_WRITE(SOUTH_DSPCLK_GATE_D, I915_READ(SOUTH_DSPCLK_GATE_D) |
8513 CNP_PWM_CGE_GATING_DISABLE);
8514 }
8515
cnl_init_clock_gating(struct drm_i915_private * dev_priv)8516 static void cnl_init_clock_gating(struct drm_i915_private *dev_priv)
8517 {
8518 cnp_init_clock_gating(dev_priv);
8519
8520 /* This is not an Wa. Enable for better image quality */
8521 I915_WRITE(_3D_CHICKEN3,
8522 _MASKED_BIT_ENABLE(_3D_CHICKEN3_AA_LINE_QUALITY_FIX_ENABLE));
8523
8524 /* WaEnableChickenDCPR:cnl */
8525 I915_WRITE(GEN8_CHICKEN_DCPR_1,
8526 I915_READ(GEN8_CHICKEN_DCPR_1) | MASK_WAKEMEM);
8527
8528 /* WaFbcWakeMemOn:cnl */
8529 I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
8530 DISP_FBC_MEMORY_WAKE);
8531
8532 /* WaSarbUnitClockGatingDisable:cnl (pre-prod) */
8533 if (IS_CNL_REVID(dev_priv, CNL_REVID_A0, CNL_REVID_B0))
8534 I915_WRITE(SLICE_UNIT_LEVEL_CLKGATE,
8535 I915_READ(SLICE_UNIT_LEVEL_CLKGATE) |
8536 SARBUNIT_CLKGATE_DIS);
8537 }
8538
cfl_init_clock_gating(struct drm_i915_private * dev_priv)8539 static void cfl_init_clock_gating(struct drm_i915_private *dev_priv)
8540 {
8541 cnp_init_clock_gating(dev_priv);
8542 gen9_init_clock_gating(dev_priv);
8543
8544 /* WaFbcNukeOnHostModify:cfl */
8545 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
8546 ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
8547 }
8548
kbl_init_clock_gating(struct drm_i915_private * dev_priv)8549 static void kbl_init_clock_gating(struct drm_i915_private *dev_priv)
8550 {
8551 gen9_init_clock_gating(dev_priv);
8552
8553 /* WaDisableSDEUnitClockGating:kbl */
8554 if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
8555 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
8556 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
8557
8558 /* WaDisableGamClockGating:kbl */
8559 if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
8560 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
8561 GEN6_GAMUNIT_CLOCK_GATE_DISABLE);
8562
8563 /* WaFbcNukeOnHostModify:kbl */
8564 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
8565 ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
8566 }
8567
skl_init_clock_gating(struct drm_i915_private * dev_priv)8568 static void skl_init_clock_gating(struct drm_i915_private *dev_priv)
8569 {
8570 gen9_init_clock_gating(dev_priv);
8571
8572 /* WAC6entrylatency:skl */
8573 I915_WRITE(FBC_LLC_READ_CTRL, I915_READ(FBC_LLC_READ_CTRL) |
8574 FBC_LLC_FULLY_OPEN);
8575
8576 /* WaFbcNukeOnHostModify:skl */
8577 I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
8578 ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
8579 }
8580
bdw_init_clock_gating(struct drm_i915_private * dev_priv)8581 static void bdw_init_clock_gating(struct drm_i915_private *dev_priv)
8582 {
8583 /* The GTT cache must be disabled if the system is using 2M pages. */
8584 bool can_use_gtt_cache = !HAS_PAGE_SIZES(dev_priv,
8585 I915_GTT_PAGE_SIZE_2M);
8586 enum i915_pipe pipe;
8587
8588 /* WaSwitchSolVfFArbitrationPriority:bdw */
8589 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
8590
8591 /* WaPsrDPAMaskVBlankInSRD:bdw */
8592 I915_WRITE(CHICKEN_PAR1_1,
8593 I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
8594
8595 /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
8596 for_each_pipe(dev_priv, pipe) {
8597 I915_WRITE(CHICKEN_PIPESL_1(pipe),
8598 I915_READ(CHICKEN_PIPESL_1(pipe)) |
8599 BDW_DPRS_MASK_VBLANK_SRD);
8600 }
8601
8602 /* WaVSRefCountFullforceMissDisable:bdw */
8603 /* WaDSRefCountFullforceMissDisable:bdw */
8604 I915_WRITE(GEN7_FF_THREAD_MODE,
8605 I915_READ(GEN7_FF_THREAD_MODE) &
8606 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
8607
8608 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
8609 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
8610
8611 /* WaDisableSDEUnitClockGating:bdw */
8612 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
8613 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
8614
8615 /* WaProgramL3SqcReg1Default:bdw */
8616 gen8_set_l3sqc_credits(dev_priv, 30, 2);
8617
8618 /* WaGttCachingOffByDefault:bdw */
8619 I915_WRITE(HSW_GTT_CACHE_EN, can_use_gtt_cache ? GTT_CACHE_EN_ALL : 0);
8620
8621 /* WaKVMNotificationOnConfigChange:bdw */
8622 I915_WRITE(CHICKEN_PAR2_1, I915_READ(CHICKEN_PAR2_1)
8623 | KVM_CONFIG_CHANGE_NOTIFICATION_SELECT);
8624
8625 lpt_init_clock_gating(dev_priv);
8626
8627 /* WaDisableDopClockGating:bdw
8628 *
8629 * Also see the CHICKEN2 write in bdw_init_workarounds() to disable DOP
8630 * clock gating.
8631 */
8632 I915_WRITE(GEN6_UCGCTL1,
8633 I915_READ(GEN6_UCGCTL1) | GEN6_EU_TCUNIT_CLOCK_GATE_DISABLE);
8634 }
8635
hsw_init_clock_gating(struct drm_i915_private * dev_priv)8636 static void hsw_init_clock_gating(struct drm_i915_private *dev_priv)
8637 {
8638 /* L3 caching of data atomics doesn't work -- disable it. */
8639 I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
8640 I915_WRITE(HSW_ROW_CHICKEN3,
8641 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
8642
8643 /* This is required by WaCatErrorRejectionIssue:hsw */
8644 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
8645 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
8646 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
8647
8648 /* WaVSRefCountFullforceMissDisable:hsw */
8649 I915_WRITE(GEN7_FF_THREAD_MODE,
8650 I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
8651
8652 /* WaDisable_RenderCache_OperationalFlush:hsw */
8653 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8654
8655 /* enable HiZ Raw Stall Optimization */
8656 I915_WRITE(CACHE_MODE_0_GEN7,
8657 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
8658
8659 /* WaDisable4x2SubspanOptimization:hsw */
8660 I915_WRITE(CACHE_MODE_1,
8661 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
8662
8663 /*
8664 * BSpec recommends 8x4 when MSAA is used,
8665 * however in practice 16x4 seems fastest.
8666 *
8667 * Note that PS/WM thread counts depend on the WIZ hashing
8668 * disable bit, which we don't touch here, but it's good
8669 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
8670 */
8671 I915_WRITE(GEN7_GT_MODE,
8672 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
8673
8674 /* WaSampleCChickenBitEnable:hsw */
8675 I915_WRITE(HALF_SLICE_CHICKEN3,
8676 _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE));
8677
8678 /* WaSwitchSolVfFArbitrationPriority:hsw */
8679 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
8680
8681 lpt_init_clock_gating(dev_priv);
8682 }
8683
ivb_init_clock_gating(struct drm_i915_private * dev_priv)8684 static void ivb_init_clock_gating(struct drm_i915_private *dev_priv)
8685 {
8686 uint32_t snpcr;
8687
8688 I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
8689
8690 /* WaDisableEarlyCull:ivb */
8691 I915_WRITE(_3D_CHICKEN3,
8692 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
8693
8694 /* WaDisableBackToBackFlipFix:ivb */
8695 I915_WRITE(IVB_CHICKEN3,
8696 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
8697 CHICKEN3_DGMG_DONE_FIX_DISABLE);
8698
8699 /* WaDisablePSDDualDispatchEnable:ivb */
8700 if (IS_IVB_GT1(dev_priv))
8701 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
8702 _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
8703
8704 /* WaDisable_RenderCache_OperationalFlush:ivb */
8705 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8706
8707 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
8708 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
8709 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
8710
8711 /* WaApplyL3ControlAndL3ChickenMode:ivb */
8712 I915_WRITE(GEN7_L3CNTLREG1,
8713 GEN7_WA_FOR_GEN7_L3_CONTROL);
8714 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
8715 GEN7_WA_L3_CHICKEN_MODE);
8716 if (IS_IVB_GT1(dev_priv))
8717 I915_WRITE(GEN7_ROW_CHICKEN2,
8718 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
8719 else {
8720 /* must write both registers */
8721 I915_WRITE(GEN7_ROW_CHICKEN2,
8722 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
8723 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
8724 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
8725 }
8726
8727 /* WaForceL3Serialization:ivb */
8728 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
8729 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
8730
8731 /*
8732 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
8733 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
8734 */
8735 I915_WRITE(GEN6_UCGCTL2,
8736 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
8737
8738 /* This is required by WaCatErrorRejectionIssue:ivb */
8739 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
8740 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
8741 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
8742
8743 g4x_disable_trickle_feed(dev_priv);
8744
8745 gen7_setup_fixed_func_scheduler(dev_priv);
8746
8747 if (0) { /* causes HiZ corruption on ivb:gt1 */
8748 /* enable HiZ Raw Stall Optimization */
8749 I915_WRITE(CACHE_MODE_0_GEN7,
8750 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
8751 }
8752
8753 /* WaDisable4x2SubspanOptimization:ivb */
8754 I915_WRITE(CACHE_MODE_1,
8755 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
8756
8757 /*
8758 * BSpec recommends 8x4 when MSAA is used,
8759 * however in practice 16x4 seems fastest.
8760 *
8761 * Note that PS/WM thread counts depend on the WIZ hashing
8762 * disable bit, which we don't touch here, but it's good
8763 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
8764 */
8765 I915_WRITE(GEN7_GT_MODE,
8766 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
8767
8768 snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
8769 snpcr &= ~GEN6_MBC_SNPCR_MASK;
8770 snpcr |= GEN6_MBC_SNPCR_MED;
8771 I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
8772
8773 if (!HAS_PCH_NOP(dev_priv))
8774 cpt_init_clock_gating(dev_priv);
8775
8776 gen6_check_mch_setup(dev_priv);
8777 }
8778
vlv_init_clock_gating(struct drm_i915_private * dev_priv)8779 static void vlv_init_clock_gating(struct drm_i915_private *dev_priv)
8780 {
8781 /* WaDisableEarlyCull:vlv */
8782 I915_WRITE(_3D_CHICKEN3,
8783 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
8784
8785 /* WaDisableBackToBackFlipFix:vlv */
8786 I915_WRITE(IVB_CHICKEN3,
8787 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
8788 CHICKEN3_DGMG_DONE_FIX_DISABLE);
8789
8790 /* WaPsdDispatchEnable:vlv */
8791 /* WaDisablePSDDualDispatchEnable:vlv */
8792 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
8793 _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
8794 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
8795
8796 /* WaDisable_RenderCache_OperationalFlush:vlv */
8797 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8798
8799 /* WaForceL3Serialization:vlv */
8800 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
8801 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
8802
8803 /* WaDisableDopClockGating:vlv */
8804 I915_WRITE(GEN7_ROW_CHICKEN2,
8805 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
8806
8807 /* This is required by WaCatErrorRejectionIssue:vlv */
8808 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
8809 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
8810 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
8811
8812 gen7_setup_fixed_func_scheduler(dev_priv);
8813
8814 /*
8815 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
8816 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
8817 */
8818 I915_WRITE(GEN6_UCGCTL2,
8819 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
8820
8821 /* WaDisableL3Bank2xClockGate:vlv
8822 * Disabling L3 clock gating- MMIO 940c[25] = 1
8823 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
8824 I915_WRITE(GEN7_UCGCTL4,
8825 I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
8826
8827 /*
8828 * BSpec says this must be set, even though
8829 * WaDisable4x2SubspanOptimization isn't listed for VLV.
8830 */
8831 I915_WRITE(CACHE_MODE_1,
8832 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
8833
8834 /*
8835 * BSpec recommends 8x4 when MSAA is used,
8836 * however in practice 16x4 seems fastest.
8837 *
8838 * Note that PS/WM thread counts depend on the WIZ hashing
8839 * disable bit, which we don't touch here, but it's good
8840 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
8841 */
8842 I915_WRITE(GEN7_GT_MODE,
8843 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
8844
8845 /*
8846 * WaIncreaseL3CreditsForVLVB0:vlv
8847 * This is the hardware default actually.
8848 */
8849 I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
8850
8851 /*
8852 * WaDisableVLVClockGating_VBIIssue:vlv
8853 * Disable clock gating on th GCFG unit to prevent a delay
8854 * in the reporting of vblank events.
8855 */
8856 I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
8857 }
8858
chv_init_clock_gating(struct drm_i915_private * dev_priv)8859 static void chv_init_clock_gating(struct drm_i915_private *dev_priv)
8860 {
8861 /* WaVSRefCountFullforceMissDisable:chv */
8862 /* WaDSRefCountFullforceMissDisable:chv */
8863 I915_WRITE(GEN7_FF_THREAD_MODE,
8864 I915_READ(GEN7_FF_THREAD_MODE) &
8865 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
8866
8867 /* WaDisableSemaphoreAndSyncFlipWait:chv */
8868 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
8869 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
8870
8871 /* WaDisableCSUnitClockGating:chv */
8872 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
8873 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
8874
8875 /* WaDisableSDEUnitClockGating:chv */
8876 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
8877 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
8878
8879 /*
8880 * WaProgramL3SqcReg1Default:chv
8881 * See gfxspecs/Related Documents/Performance Guide/
8882 * LSQC Setting Recommendations.
8883 */
8884 gen8_set_l3sqc_credits(dev_priv, 38, 2);
8885
8886 /*
8887 * GTT cache may not work with big pages, so if those
8888 * are ever enabled GTT cache may need to be disabled.
8889 */
8890 I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
8891 }
8892
g4x_init_clock_gating(struct drm_i915_private * dev_priv)8893 static void g4x_init_clock_gating(struct drm_i915_private *dev_priv)
8894 {
8895 uint32_t dspclk_gate;
8896
8897 I915_WRITE(RENCLK_GATE_D1, 0);
8898 I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
8899 GS_UNIT_CLOCK_GATE_DISABLE |
8900 CL_UNIT_CLOCK_GATE_DISABLE);
8901 I915_WRITE(RAMCLK_GATE_D, 0);
8902 dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
8903 OVRUNIT_CLOCK_GATE_DISABLE |
8904 OVCUNIT_CLOCK_GATE_DISABLE;
8905 if (IS_GM45(dev_priv))
8906 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
8907 I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
8908
8909 /* WaDisableRenderCachePipelinedFlush */
8910 I915_WRITE(CACHE_MODE_0,
8911 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
8912
8913 /* WaDisable_RenderCache_OperationalFlush:g4x */
8914 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8915
8916 g4x_disable_trickle_feed(dev_priv);
8917 }
8918
i965gm_init_clock_gating(struct drm_i915_private * dev_priv)8919 static void i965gm_init_clock_gating(struct drm_i915_private *dev_priv)
8920 {
8921 I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
8922 I915_WRITE(RENCLK_GATE_D2, 0);
8923 I915_WRITE(DSPCLK_GATE_D, 0);
8924 I915_WRITE(RAMCLK_GATE_D, 0);
8925 I915_WRITE16(DEUC, 0);
8926 I915_WRITE(MI_ARB_STATE,
8927 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
8928
8929 /* WaDisable_RenderCache_OperationalFlush:gen4 */
8930 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8931 }
8932
i965g_init_clock_gating(struct drm_i915_private * dev_priv)8933 static void i965g_init_clock_gating(struct drm_i915_private *dev_priv)
8934 {
8935 I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
8936 I965_RCC_CLOCK_GATE_DISABLE |
8937 I965_RCPB_CLOCK_GATE_DISABLE |
8938 I965_ISC_CLOCK_GATE_DISABLE |
8939 I965_FBC_CLOCK_GATE_DISABLE);
8940 I915_WRITE(RENCLK_GATE_D2, 0);
8941 I915_WRITE(MI_ARB_STATE,
8942 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
8943
8944 /* WaDisable_RenderCache_OperationalFlush:gen4 */
8945 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
8946 }
8947
gen3_init_clock_gating(struct drm_i915_private * dev_priv)8948 static void gen3_init_clock_gating(struct drm_i915_private *dev_priv)
8949 {
8950 u32 dstate = I915_READ(D_STATE);
8951
8952 dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
8953 DSTATE_DOT_CLOCK_GATING;
8954 I915_WRITE(D_STATE, dstate);
8955
8956 if (IS_PINEVIEW(dev_priv))
8957 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
8958
8959 /* IIR "flip pending" means done if this bit is set */
8960 I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
8961
8962 /* interrupts should cause a wake up from C3 */
8963 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
8964
8965 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
8966 I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
8967
8968 I915_WRITE(MI_ARB_STATE,
8969 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
8970 }
8971
i85x_init_clock_gating(struct drm_i915_private * dev_priv)8972 static void i85x_init_clock_gating(struct drm_i915_private *dev_priv)
8973 {
8974 I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
8975
8976 /* interrupts should cause a wake up from C3 */
8977 I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
8978 _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
8979
8980 I915_WRITE(MEM_MODE,
8981 _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE));
8982 }
8983
i830_init_clock_gating(struct drm_i915_private * dev_priv)8984 static void i830_init_clock_gating(struct drm_i915_private *dev_priv)
8985 {
8986 I915_WRITE(MEM_MODE,
8987 _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE) |
8988 _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE));
8989 }
8990
intel_init_clock_gating(struct drm_i915_private * dev_priv)8991 void intel_init_clock_gating(struct drm_i915_private *dev_priv)
8992 {
8993 dev_priv->display.init_clock_gating(dev_priv);
8994 }
8995
intel_suspend_hw(struct drm_i915_private * dev_priv)8996 void intel_suspend_hw(struct drm_i915_private *dev_priv)
8997 {
8998 if (HAS_PCH_LPT(dev_priv))
8999 lpt_suspend_hw(dev_priv);
9000 }
9001
nop_init_clock_gating(struct drm_i915_private * dev_priv)9002 static void nop_init_clock_gating(struct drm_i915_private *dev_priv)
9003 {
9004 DRM_DEBUG_KMS("No clock gating settings or workarounds applied.\n");
9005 }
9006
9007 /**
9008 * intel_init_clock_gating_hooks - setup the clock gating hooks
9009 * @dev_priv: device private
9010 *
9011 * Setup the hooks that configure which clocks of a given platform can be
9012 * gated and also apply various GT and display specific workarounds for these
9013 * platforms. Note that some GT specific workarounds are applied separately
9014 * when GPU contexts or batchbuffers start their execution.
9015 */
intel_init_clock_gating_hooks(struct drm_i915_private * dev_priv)9016 void intel_init_clock_gating_hooks(struct drm_i915_private *dev_priv)
9017 {
9018 if (IS_CANNONLAKE(dev_priv))
9019 dev_priv->display.init_clock_gating = cnl_init_clock_gating;
9020 else if (IS_COFFEELAKE(dev_priv))
9021 dev_priv->display.init_clock_gating = cfl_init_clock_gating;
9022 else if (IS_SKYLAKE(dev_priv))
9023 dev_priv->display.init_clock_gating = skl_init_clock_gating;
9024 else if (IS_KABYLAKE(dev_priv))
9025 dev_priv->display.init_clock_gating = kbl_init_clock_gating;
9026 else if (IS_BROXTON(dev_priv))
9027 dev_priv->display.init_clock_gating = bxt_init_clock_gating;
9028 else if (IS_GEMINILAKE(dev_priv))
9029 dev_priv->display.init_clock_gating = glk_init_clock_gating;
9030 else if (IS_BROADWELL(dev_priv))
9031 dev_priv->display.init_clock_gating = bdw_init_clock_gating;
9032 else if (IS_CHERRYVIEW(dev_priv))
9033 dev_priv->display.init_clock_gating = chv_init_clock_gating;
9034 else if (IS_HASWELL(dev_priv))
9035 dev_priv->display.init_clock_gating = hsw_init_clock_gating;
9036 else if (IS_IVYBRIDGE(dev_priv))
9037 dev_priv->display.init_clock_gating = ivb_init_clock_gating;
9038 else if (IS_VALLEYVIEW(dev_priv))
9039 dev_priv->display.init_clock_gating = vlv_init_clock_gating;
9040 else if (IS_GEN6(dev_priv))
9041 dev_priv->display.init_clock_gating = gen6_init_clock_gating;
9042 else if (IS_GEN5(dev_priv))
9043 dev_priv->display.init_clock_gating = ilk_init_clock_gating;
9044 else if (IS_G4X(dev_priv))
9045 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
9046 else if (IS_I965GM(dev_priv))
9047 dev_priv->display.init_clock_gating = i965gm_init_clock_gating;
9048 else if (IS_I965G(dev_priv))
9049 dev_priv->display.init_clock_gating = i965g_init_clock_gating;
9050 else if (IS_GEN3(dev_priv))
9051 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
9052 else if (IS_I85X(dev_priv) || IS_I865G(dev_priv))
9053 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
9054 else if (IS_GEN2(dev_priv))
9055 dev_priv->display.init_clock_gating = i830_init_clock_gating;
9056 else {
9057 MISSING_CASE(INTEL_DEVID(dev_priv));
9058 dev_priv->display.init_clock_gating = nop_init_clock_gating;
9059 }
9060 }
9061
9062 /* Set up chip specific power management-related functions */
intel_init_pm(struct drm_i915_private * dev_priv)9063 void intel_init_pm(struct drm_i915_private *dev_priv)
9064 {
9065 intel_fbc_init(dev_priv);
9066
9067 /* For cxsr */
9068 if (IS_PINEVIEW(dev_priv))
9069 i915_pineview_get_mem_freq(dev_priv);
9070 else if (IS_GEN5(dev_priv))
9071 i915_ironlake_get_mem_freq(dev_priv);
9072
9073 /* For FIFO watermark updates */
9074 if (INTEL_GEN(dev_priv) >= 9) {
9075 skl_setup_wm_latency(dev_priv);
9076 dev_priv->display.initial_watermarks = skl_initial_wm;
9077 dev_priv->display.atomic_update_watermarks = skl_atomic_update_crtc_wm;
9078 dev_priv->display.compute_global_watermarks = skl_compute_wm;
9079 } else if (HAS_PCH_SPLIT(dev_priv)) {
9080 ilk_setup_wm_latency(dev_priv);
9081
9082 if ((IS_GEN5(dev_priv) && dev_priv->wm.pri_latency[1] &&
9083 dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
9084 (!IS_GEN5(dev_priv) && dev_priv->wm.pri_latency[0] &&
9085 dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
9086 dev_priv->display.compute_pipe_wm = ilk_compute_pipe_wm;
9087 dev_priv->display.compute_intermediate_wm =
9088 ilk_compute_intermediate_wm;
9089 dev_priv->display.initial_watermarks =
9090 ilk_initial_watermarks;
9091 dev_priv->display.optimize_watermarks =
9092 ilk_optimize_watermarks;
9093 } else {
9094 DRM_DEBUG_KMS("Failed to read display plane latency. "
9095 "Disable CxSR\n");
9096 }
9097 } else if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
9098 vlv_setup_wm_latency(dev_priv);
9099 dev_priv->display.compute_pipe_wm = vlv_compute_pipe_wm;
9100 dev_priv->display.compute_intermediate_wm = vlv_compute_intermediate_wm;
9101 dev_priv->display.initial_watermarks = vlv_initial_watermarks;
9102 dev_priv->display.optimize_watermarks = vlv_optimize_watermarks;
9103 dev_priv->display.atomic_update_watermarks = vlv_atomic_update_fifo;
9104 } else if (IS_G4X(dev_priv)) {
9105 g4x_setup_wm_latency(dev_priv);
9106 dev_priv->display.compute_pipe_wm = g4x_compute_pipe_wm;
9107 dev_priv->display.compute_intermediate_wm = g4x_compute_intermediate_wm;
9108 dev_priv->display.initial_watermarks = g4x_initial_watermarks;
9109 dev_priv->display.optimize_watermarks = g4x_optimize_watermarks;
9110 } else if (IS_PINEVIEW(dev_priv)) {
9111 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv),
9112 dev_priv->is_ddr3,
9113 dev_priv->fsb_freq,
9114 dev_priv->mem_freq)) {
9115 DRM_INFO("failed to find known CxSR latency "
9116 "(found ddr%s fsb freq %d, mem freq %d), "
9117 "disabling CxSR\n",
9118 (dev_priv->is_ddr3 == 1) ? "3" : "2",
9119 dev_priv->fsb_freq, dev_priv->mem_freq);
9120 /* Disable CxSR and never update its watermark again */
9121 intel_set_memory_cxsr(dev_priv, false);
9122 dev_priv->display.update_wm = NULL;
9123 } else
9124 dev_priv->display.update_wm = pineview_update_wm;
9125 } else if (IS_GEN4(dev_priv)) {
9126 dev_priv->display.update_wm = i965_update_wm;
9127 } else if (IS_GEN3(dev_priv)) {
9128 dev_priv->display.update_wm = i9xx_update_wm;
9129 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
9130 } else if (IS_GEN2(dev_priv)) {
9131 if (INTEL_INFO(dev_priv)->num_pipes == 1) {
9132 dev_priv->display.update_wm = i845_update_wm;
9133 dev_priv->display.get_fifo_size = i845_get_fifo_size;
9134 } else {
9135 dev_priv->display.update_wm = i9xx_update_wm;
9136 dev_priv->display.get_fifo_size = i830_get_fifo_size;
9137 }
9138 } else {
9139 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
9140 }
9141 }
9142
gen6_check_mailbox_status(struct drm_i915_private * dev_priv)9143 static inline int gen6_check_mailbox_status(struct drm_i915_private *dev_priv)
9144 {
9145 uint32_t flags =
9146 I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_ERROR_MASK;
9147
9148 switch (flags) {
9149 case GEN6_PCODE_SUCCESS:
9150 return 0;
9151 case GEN6_PCODE_UNIMPLEMENTED_CMD:
9152 return -ENODEV;
9153 case GEN6_PCODE_ILLEGAL_CMD:
9154 return -ENXIO;
9155 case GEN6_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
9156 case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
9157 return -EOVERFLOW;
9158 case GEN6_PCODE_TIMEOUT:
9159 return -ETIMEDOUT;
9160 default:
9161 MISSING_CASE(flags);
9162 return 0;
9163 }
9164 }
9165
gen7_check_mailbox_status(struct drm_i915_private * dev_priv)9166 static inline int gen7_check_mailbox_status(struct drm_i915_private *dev_priv)
9167 {
9168 uint32_t flags =
9169 I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_ERROR_MASK;
9170
9171 switch (flags) {
9172 case GEN6_PCODE_SUCCESS:
9173 return 0;
9174 case GEN6_PCODE_ILLEGAL_CMD:
9175 return -ENXIO;
9176 case GEN7_PCODE_TIMEOUT:
9177 return -ETIMEDOUT;
9178 case GEN7_PCODE_ILLEGAL_DATA:
9179 return -EINVAL;
9180 case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
9181 return -EOVERFLOW;
9182 default:
9183 MISSING_CASE(flags);
9184 return 0;
9185 }
9186 }
9187
sandybridge_pcode_read(struct drm_i915_private * dev_priv,u32 mbox,u32 * val)9188 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u32 mbox, u32 *val)
9189 {
9190 int status;
9191
9192 WARN_ON(!mutex_is_locked(&dev_priv->pcu_lock));
9193
9194 /* GEN6_PCODE_* are outside of the forcewake domain, we can
9195 * use te fw I915_READ variants to reduce the amount of work
9196 * required when reading/writing.
9197 */
9198
9199 if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
9200 DRM_DEBUG_DRIVER("warning: pcode (read from mbox %x) mailbox access failed for %ps\n",
9201 mbox, __builtin_return_address(0));
9202 return -EAGAIN;
9203 }
9204
9205 I915_WRITE_FW(GEN6_PCODE_DATA, *val);
9206 I915_WRITE_FW(GEN6_PCODE_DATA1, 0);
9207 I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
9208
9209 if (__intel_wait_for_register_fw(dev_priv,
9210 GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
9211 500, 0, NULL)) {
9212 DRM_ERROR("timeout waiting for pcode read (from mbox %x) to finish for %ps\n",
9213 mbox, __builtin_return_address(0));
9214 return -ETIMEDOUT;
9215 }
9216
9217 *val = I915_READ_FW(GEN6_PCODE_DATA);
9218 I915_WRITE_FW(GEN6_PCODE_DATA, 0);
9219
9220 if (INTEL_GEN(dev_priv) > 6)
9221 status = gen7_check_mailbox_status(dev_priv);
9222 else
9223 status = gen6_check_mailbox_status(dev_priv);
9224
9225 if (status) {
9226 DRM_DEBUG_DRIVER("warning: pcode (read from mbox %x) mailbox access failed for %ps: %d\n",
9227 mbox, __builtin_return_address(0), status);
9228 return status;
9229 }
9230
9231 return 0;
9232 }
9233
sandybridge_pcode_write(struct drm_i915_private * dev_priv,u32 mbox,u32 val)9234 int sandybridge_pcode_write(struct drm_i915_private *dev_priv,
9235 u32 mbox, u32 val)
9236 {
9237 int status;
9238
9239 WARN_ON(!mutex_is_locked(&dev_priv->pcu_lock));
9240
9241 /* GEN6_PCODE_* are outside of the forcewake domain, we can
9242 * use te fw I915_READ variants to reduce the amount of work
9243 * required when reading/writing.
9244 */
9245
9246 if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
9247 DRM_DEBUG_DRIVER("warning: pcode (write of 0x%08x to mbox %x) mailbox access failed for %ps\n",
9248 val, mbox, __builtin_return_address(0));
9249 return -EAGAIN;
9250 }
9251
9252 I915_WRITE_FW(GEN6_PCODE_DATA, val);
9253 I915_WRITE_FW(GEN6_PCODE_DATA1, 0);
9254 I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
9255
9256 if (__intel_wait_for_register_fw(dev_priv,
9257 GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
9258 500, 0, NULL)) {
9259 DRM_ERROR("timeout waiting for pcode write of 0x%08x to mbox %x to finish for %ps\n",
9260 val, mbox, __builtin_return_address(0));
9261 return -ETIMEDOUT;
9262 }
9263
9264 I915_WRITE_FW(GEN6_PCODE_DATA, 0);
9265
9266 if (INTEL_GEN(dev_priv) > 6)
9267 status = gen7_check_mailbox_status(dev_priv);
9268 else
9269 status = gen6_check_mailbox_status(dev_priv);
9270
9271 if (status) {
9272 DRM_DEBUG_DRIVER("warning: pcode (write of 0x%08x to mbox %x) mailbox access failed for %ps: %d\n",
9273 val, mbox, __builtin_return_address(0), status);
9274 return status;
9275 }
9276
9277 return 0;
9278 }
9279
skl_pcode_try_request(struct drm_i915_private * dev_priv,u32 mbox,u32 request,u32 reply_mask,u32 reply,u32 * status)9280 static bool skl_pcode_try_request(struct drm_i915_private *dev_priv, u32 mbox,
9281 u32 request, u32 reply_mask, u32 reply,
9282 u32 *status)
9283 {
9284 u32 val = request;
9285
9286 *status = sandybridge_pcode_read(dev_priv, mbox, &val);
9287
9288 return *status || ((val & reply_mask) == reply);
9289 }
9290
9291 /**
9292 * skl_pcode_request - send PCODE request until acknowledgment
9293 * @dev_priv: device private
9294 * @mbox: PCODE mailbox ID the request is targeted for
9295 * @request: request ID
9296 * @reply_mask: mask used to check for request acknowledgment
9297 * @reply: value used to check for request acknowledgment
9298 * @timeout_base_ms: timeout for polling with preemption enabled
9299 *
9300 * Keep resending the @request to @mbox until PCODE acknowledges it, PCODE
9301 * reports an error or an overall timeout of @timeout_base_ms+50 ms expires.
9302 * The request is acknowledged once the PCODE reply dword equals @reply after
9303 * applying @reply_mask. Polling is first attempted with preemption enabled
9304 * for @timeout_base_ms and if this times out for another 50 ms with
9305 * preemption disabled.
9306 *
9307 * Returns 0 on success, %-ETIMEDOUT in case of a timeout, <0 in case of some
9308 * other error as reported by PCODE.
9309 */
skl_pcode_request(struct drm_i915_private * dev_priv,u32 mbox,u32 request,u32 reply_mask,u32 reply,int timeout_base_ms)9310 int skl_pcode_request(struct drm_i915_private *dev_priv, u32 mbox, u32 request,
9311 u32 reply_mask, u32 reply, int timeout_base_ms)
9312 {
9313 u32 status;
9314 int ret;
9315
9316 WARN_ON(!mutex_is_locked(&dev_priv->pcu_lock));
9317
9318 #define COND skl_pcode_try_request(dev_priv, mbox, request, reply_mask, reply, \
9319 &status)
9320
9321 /*
9322 * Prime the PCODE by doing a request first. Normally it guarantees
9323 * that a subsequent request, at most @timeout_base_ms later, succeeds.
9324 * _wait_for() doesn't guarantee when its passed condition is evaluated
9325 * first, so send the first request explicitly.
9326 */
9327 if (COND) {
9328 ret = 0;
9329 goto out;
9330 }
9331 ret = _wait_for(COND, timeout_base_ms * 1000, 10);
9332 if (!ret)
9333 goto out;
9334
9335 /*
9336 * The above can time out if the number of requests was low (2 in the
9337 * worst case) _and_ PCODE was busy for some reason even after a
9338 * (queued) request and @timeout_base_ms delay. As a workaround retry
9339 * the poll with preemption disabled to maximize the number of
9340 * requests. Increase the timeout from @timeout_base_ms to 50ms to
9341 * account for interrupts that could reduce the number of these
9342 * requests, and for any quirks of the PCODE firmware that delays
9343 * the request completion.
9344 */
9345 DRM_DEBUG_KMS("PCODE timeout, retrying with preemption disabled\n");
9346 WARN_ON_ONCE(timeout_base_ms > 3);
9347 preempt_disable();
9348 ret = wait_for_atomic(COND, 50);
9349 preempt_enable();
9350
9351 out:
9352 return ret ? ret : status;
9353 #undef COND
9354 }
9355
byt_gpu_freq(struct drm_i915_private * dev_priv,int val)9356 static int byt_gpu_freq(struct drm_i915_private *dev_priv, int val)
9357 {
9358 struct intel_rps *rps = &dev_priv->gt_pm.rps;
9359
9360 /*
9361 * N = val - 0xb7
9362 * Slow = Fast = GPLL ref * N
9363 */
9364 return DIV_ROUND_CLOSEST(rps->gpll_ref_freq * (val - 0xb7), 1000);
9365 }
9366
byt_freq_opcode(struct drm_i915_private * dev_priv,int val)9367 static int byt_freq_opcode(struct drm_i915_private *dev_priv, int val)
9368 {
9369 struct intel_rps *rps = &dev_priv->gt_pm.rps;
9370
9371 return DIV_ROUND_CLOSEST(1000 * val, rps->gpll_ref_freq) + 0xb7;
9372 }
9373
chv_gpu_freq(struct drm_i915_private * dev_priv,int val)9374 static int chv_gpu_freq(struct drm_i915_private *dev_priv, int val)
9375 {
9376 struct intel_rps *rps = &dev_priv->gt_pm.rps;
9377
9378 /*
9379 * N = val / 2
9380 * CU (slow) = CU2x (fast) / 2 = GPLL ref * N / 2
9381 */
9382 return DIV_ROUND_CLOSEST(rps->gpll_ref_freq * val, 2 * 2 * 1000);
9383 }
9384
chv_freq_opcode(struct drm_i915_private * dev_priv,int val)9385 static int chv_freq_opcode(struct drm_i915_private *dev_priv, int val)
9386 {
9387 struct intel_rps *rps = &dev_priv->gt_pm.rps;
9388
9389 /* CHV needs even values */
9390 return DIV_ROUND_CLOSEST(2 * 1000 * val, rps->gpll_ref_freq) * 2;
9391 }
9392
intel_gpu_freq(struct drm_i915_private * dev_priv,int val)9393 int intel_gpu_freq(struct drm_i915_private *dev_priv, int val)
9394 {
9395 if (INTEL_GEN(dev_priv) >= 9)
9396 return DIV_ROUND_CLOSEST(val * GT_FREQUENCY_MULTIPLIER,
9397 GEN9_FREQ_SCALER);
9398 else if (IS_CHERRYVIEW(dev_priv))
9399 return chv_gpu_freq(dev_priv, val);
9400 else if (IS_VALLEYVIEW(dev_priv))
9401 return byt_gpu_freq(dev_priv, val);
9402 else
9403 return val * GT_FREQUENCY_MULTIPLIER;
9404 }
9405
intel_freq_opcode(struct drm_i915_private * dev_priv,int val)9406 int intel_freq_opcode(struct drm_i915_private *dev_priv, int val)
9407 {
9408 if (INTEL_GEN(dev_priv) >= 9)
9409 return DIV_ROUND_CLOSEST(val * GEN9_FREQ_SCALER,
9410 GT_FREQUENCY_MULTIPLIER);
9411 else if (IS_CHERRYVIEW(dev_priv))
9412 return chv_freq_opcode(dev_priv, val);
9413 else if (IS_VALLEYVIEW(dev_priv))
9414 return byt_freq_opcode(dev_priv, val);
9415 else
9416 return DIV_ROUND_CLOSEST(val, GT_FREQUENCY_MULTIPLIER);
9417 }
9418
intel_pm_setup(struct drm_i915_private * dev_priv)9419 void intel_pm_setup(struct drm_i915_private *dev_priv)
9420 {
9421 lockinit(&dev_priv->pcu_lock, "i9pcul", 0, LK_CANRECURSE);
9422
9423 INIT_DELAYED_WORK(&dev_priv->gt_pm.autoenable_work,
9424 __intel_autoenable_gt_powersave);
9425 atomic_set(&dev_priv->gt_pm.rps.num_waiters, 0);
9426
9427 dev_priv->runtime_pm.suspended = false;
9428 atomic_set(&dev_priv->runtime_pm.wakeref_count, 0);
9429 }
9430
vlv_residency_raw(struct drm_i915_private * dev_priv,const i915_reg_t reg)9431 static u64 vlv_residency_raw(struct drm_i915_private *dev_priv,
9432 const i915_reg_t reg)
9433 {
9434 u32 lower, upper, tmp;
9435 int loop = 2;
9436
9437 /* The register accessed do not need forcewake. We borrow
9438 * uncore lock to prevent concurrent access to range reg.
9439 */
9440 spin_lock_irq(&dev_priv->uncore.lock);
9441
9442 /* vlv and chv residency counters are 40 bits in width.
9443 * With a control bit, we can choose between upper or lower
9444 * 32bit window into this counter.
9445 *
9446 * Although we always use the counter in high-range mode elsewhere,
9447 * userspace may attempt to read the value before rc6 is initialised,
9448 * before we have set the default VLV_COUNTER_CONTROL value. So always
9449 * set the high bit to be safe.
9450 */
9451 I915_WRITE_FW(VLV_COUNTER_CONTROL,
9452 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH));
9453 upper = I915_READ_FW(reg);
9454 do {
9455 tmp = upper;
9456
9457 I915_WRITE_FW(VLV_COUNTER_CONTROL,
9458 _MASKED_BIT_DISABLE(VLV_COUNT_RANGE_HIGH));
9459 lower = I915_READ_FW(reg);
9460
9461 I915_WRITE_FW(VLV_COUNTER_CONTROL,
9462 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH));
9463 upper = I915_READ_FW(reg);
9464 } while (upper != tmp && --loop);
9465
9466 /* Everywhere else we always use VLV_COUNTER_CONTROL with the
9467 * VLV_COUNT_RANGE_HIGH bit set - so it is safe to leave it set
9468 * now.
9469 */
9470
9471 spin_unlock_irq(&dev_priv->uncore.lock);
9472
9473 return lower | (u64)upper << 8;
9474 }
9475
intel_rc6_residency_us(struct drm_i915_private * dev_priv,const i915_reg_t reg)9476 u64 intel_rc6_residency_us(struct drm_i915_private *dev_priv,
9477 const i915_reg_t reg)
9478 {
9479 u64 time_hw, units, div;
9480
9481 if (!intel_rc6_enabled())
9482 return 0;
9483
9484 intel_runtime_pm_get(dev_priv);
9485
9486 /* On VLV and CHV, residency time is in CZ units rather than 1.28us */
9487 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
9488 units = 1000;
9489 div = dev_priv->czclk_freq;
9490
9491 time_hw = vlv_residency_raw(dev_priv, reg);
9492 } else if (IS_GEN9_LP(dev_priv)) {
9493 units = 1000;
9494 div = 1200; /* 833.33ns */
9495
9496 time_hw = I915_READ(reg);
9497 } else {
9498 units = 128000; /* 1.28us */
9499 div = 100000;
9500
9501 time_hw = I915_READ(reg);
9502 }
9503
9504 intel_runtime_pm_put(dev_priv);
9505 return DIV_ROUND_UP_ULL(time_hw * units, div);
9506 }
9507