?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
posix-clock: Fix return code on the poll method's error path
[linux/fpc-iii.git] / drivers / gpu / drm / i915 / intel_pm.c
blobf091ad12d694a43ebb0dce69b31089ffd4abe366
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 "i915_drv.h"
30 #include "intel_drv.h"
31 #include "../../../platform/x86/intel_ips.h"
32 #include <linux/module.h>
33
34 /**
35 * RC6 is a special power stage which allows the GPU to enter an very
36 * low-voltage mode when idle, using down to 0V while at this stage. This
37 * stage is entered automatically when the GPU is idle when RC6 support is
38 * enabled, and as soon as new workload arises GPU wakes up automatically as well.
39 *
40 * There are different RC6 modes available in Intel GPU, which differentiate
41 * among each other with the latency required to enter and leave RC6 and
42 * voltage consumed by the GPU in different states.
43 *
44 * The combination of the following flags define which states GPU is allowed
45 * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
46 * RC6pp is deepest RC6. Their support by hardware varies according to the
47 * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
48 * which brings the most power savings; deeper states save more power, but
49 * require higher latency to switch to and wake up.
50 */
51 #define INTEL_RC6_ENABLE (1<<0)
52 #define INTEL_RC6p_ENABLE (1<<1)
53 #define INTEL_RC6pp_ENABLE (1<<2)
54
55 static void bxt_init_clock_gating(struct drm_device *dev)
56 {
57 struct drm_i915_private *dev_priv = dev->dev_private;
58
59 /* WaDisableSDEUnitClockGating:bxt */
60 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
61 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
62
63 /*
64 * FIXME:
65 * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
66 */
67 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
68 GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ);
69 }
70
71 static void i915_pineview_get_mem_freq(struct drm_device *dev)
72 {
73 struct drm_i915_private *dev_priv = dev->dev_private;
74 u32 tmp;
75
76 tmp = I915_READ(CLKCFG);
77
78 switch (tmp & CLKCFG_FSB_MASK) {
79 case CLKCFG_FSB_533:
80 dev_priv->fsb_freq = 533; /* 133*4 */
81 break;
82 case CLKCFG_FSB_800:
83 dev_priv->fsb_freq = 800; /* 200*4 */
84 break;
85 case CLKCFG_FSB_667:
86 dev_priv->fsb_freq = 667; /* 167*4 */
87 break;
88 case CLKCFG_FSB_400:
89 dev_priv->fsb_freq = 400; /* 100*4 */
90 break;
91 }
92
93 switch (tmp & CLKCFG_MEM_MASK) {
94 case CLKCFG_MEM_533:
95 dev_priv->mem_freq = 533;
96 break;
97 case CLKCFG_MEM_667:
98 dev_priv->mem_freq = 667;
99 break;
100 case CLKCFG_MEM_800:
101 dev_priv->mem_freq = 800;
102 break;
103 }
104
105 /* detect pineview DDR3 setting */
106 tmp = I915_READ(CSHRDDR3CTL);
107 dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
108 }
109
110 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
111 {
112 struct drm_i915_private *dev_priv = dev->dev_private;
113 u16 ddrpll, csipll;
114
115 ddrpll = I915_READ16(DDRMPLL1);
116 csipll = I915_READ16(CSIPLL0);
117
118 switch (ddrpll & 0xff) {
119 case 0xc:
120 dev_priv->mem_freq = 800;
121 break;
122 case 0x10:
123 dev_priv->mem_freq = 1066;
124 break;
125 case 0x14:
126 dev_priv->mem_freq = 1333;
127 break;
128 case 0x18:
129 dev_priv->mem_freq = 1600;
130 break;
131 default:
132 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
133 ddrpll & 0xff);
134 dev_priv->mem_freq = 0;
135 break;
136 }
137
138 dev_priv->ips.r_t = dev_priv->mem_freq;
139
140 switch (csipll & 0x3ff) {
141 case 0x00c:
142 dev_priv->fsb_freq = 3200;
143 break;
144 case 0x00e:
145 dev_priv->fsb_freq = 3733;
146 break;
147 case 0x010:
148 dev_priv->fsb_freq = 4266;
149 break;
150 case 0x012:
151 dev_priv->fsb_freq = 4800;
152 break;
153 case 0x014:
154 dev_priv->fsb_freq = 5333;
155 break;
156 case 0x016:
157 dev_priv->fsb_freq = 5866;
158 break;
159 case 0x018:
160 dev_priv->fsb_freq = 6400;
161 break;
162 default:
163 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
164 csipll & 0x3ff);
165 dev_priv->fsb_freq = 0;
166 break;
167 }
168
169 if (dev_priv->fsb_freq == 3200) {
170 dev_priv->ips.c_m = 0;
171 } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
172 dev_priv->ips.c_m = 1;
173 } else {
174 dev_priv->ips.c_m = 2;
175 }
176 }
177
178 static const struct cxsr_latency cxsr_latency_table[] = {
179 {1, 0, 800, 400, 3382, 33382, 3983, 33983}, /* DDR2-400 SC */
180 {1, 0, 800, 667, 3354, 33354, 3807, 33807}, /* DDR2-667 SC */
181 {1, 0, 800, 800, 3347, 33347, 3763, 33763}, /* DDR2-800 SC */
182 {1, 1, 800, 667, 6420, 36420, 6873, 36873}, /* DDR3-667 SC */
183 {1, 1, 800, 800, 5902, 35902, 6318, 36318}, /* DDR3-800 SC */
184
185 {1, 0, 667, 400, 3400, 33400, 4021, 34021}, /* DDR2-400 SC */
186 {1, 0, 667, 667, 3372, 33372, 3845, 33845}, /* DDR2-667 SC */
187 {1, 0, 667, 800, 3386, 33386, 3822, 33822}, /* DDR2-800 SC */
188 {1, 1, 667, 667, 6438, 36438, 6911, 36911}, /* DDR3-667 SC */
189 {1, 1, 667, 800, 5941, 35941, 6377, 36377}, /* DDR3-800 SC */
190
191 {1, 0, 400, 400, 3472, 33472, 4173, 34173}, /* DDR2-400 SC */
192 {1, 0, 400, 667, 3443, 33443, 3996, 33996}, /* DDR2-667 SC */
193 {1, 0, 400, 800, 3430, 33430, 3946, 33946}, /* DDR2-800 SC */
194 {1, 1, 400, 667, 6509, 36509, 7062, 37062}, /* DDR3-667 SC */
195 {1, 1, 400, 800, 5985, 35985, 6501, 36501}, /* DDR3-800 SC */
196
197 {0, 0, 800, 400, 3438, 33438, 4065, 34065}, /* DDR2-400 SC */
198 {0, 0, 800, 667, 3410, 33410, 3889, 33889}, /* DDR2-667 SC */
199 {0, 0, 800, 800, 3403, 33403, 3845, 33845}, /* DDR2-800 SC */
200 {0, 1, 800, 667, 6476, 36476, 6955, 36955}, /* DDR3-667 SC */
201 {0, 1, 800, 800, 5958, 35958, 6400, 36400}, /* DDR3-800 SC */
202
203 {0, 0, 667, 400, 3456, 33456, 4103, 34106}, /* DDR2-400 SC */
204 {0, 0, 667, 667, 3428, 33428, 3927, 33927}, /* DDR2-667 SC */
205 {0, 0, 667, 800, 3443, 33443, 3905, 33905}, /* DDR2-800 SC */
206 {0, 1, 667, 667, 6494, 36494, 6993, 36993}, /* DDR3-667 SC */
207 {0, 1, 667, 800, 5998, 35998, 6460, 36460}, /* DDR3-800 SC */
208
209 {0, 0, 400, 400, 3528, 33528, 4255, 34255}, /* DDR2-400 SC */
210 {0, 0, 400, 667, 3500, 33500, 4079, 34079}, /* DDR2-667 SC */
211 {0, 0, 400, 800, 3487, 33487, 4029, 34029}, /* DDR2-800 SC */
212 {0, 1, 400, 667, 6566, 36566, 7145, 37145}, /* DDR3-667 SC */
213 {0, 1, 400, 800, 6042, 36042, 6584, 36584}, /* DDR3-800 SC */
214 };
215
216 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
217 int is_ddr3,
218 int fsb,
219 int mem)
220 {
221 const struct cxsr_latency *latency;
222 int i;
223
224 if (fsb == 0 || mem == 0)
225 return NULL;
226
227 for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
228 latency = &cxsr_latency_table[i];
229 if (is_desktop == latency->is_desktop &&
230 is_ddr3 == latency->is_ddr3 &&
231 fsb == latency->fsb_freq && mem == latency->mem_freq)
232 return latency;
233 }
234
235 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
236
237 return NULL;
238 }
239
240 static void chv_set_memory_dvfs(struct drm_i915_private *dev_priv, bool enable)
241 {
242 u32 val;
243
244 mutex_lock(&dev_priv->rps.hw_lock);
245
246 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
247 if (enable)
248 val &= ~FORCE_DDR_HIGH_FREQ;
249 else
250 val |= FORCE_DDR_HIGH_FREQ;
251 val &= ~FORCE_DDR_LOW_FREQ;
252 val |= FORCE_DDR_FREQ_REQ_ACK;
253 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
254
255 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
256 FORCE_DDR_FREQ_REQ_ACK) == 0, 3))
257 DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
258
259 mutex_unlock(&dev_priv->rps.hw_lock);
260 }
261
262 static void chv_set_memory_pm5(struct drm_i915_private *dev_priv, bool enable)
263 {
264 u32 val;
265
266 mutex_lock(&dev_priv->rps.hw_lock);
267
268 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
269 if (enable)
270 val |= DSP_MAXFIFO_PM5_ENABLE;
271 else
272 val &= ~DSP_MAXFIFO_PM5_ENABLE;
273 vlv_punit_write(dev_priv, PUNIT_REG_DSPFREQ, val);
274
275 mutex_unlock(&dev_priv->rps.hw_lock);
276 }
277
278 #define FW_WM(value, plane) \
279 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
280
281 void intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
282 {
283 struct drm_device *dev = dev_priv->dev;
284 u32 val;
285
286 if (IS_VALLEYVIEW(dev)) {
287 I915_WRITE(FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
288 POSTING_READ(FW_BLC_SELF_VLV);
289 dev_priv->wm.vlv.cxsr = enable;
290 } else if (IS_G4X(dev) || IS_CRESTLINE(dev)) {
291 I915_WRITE(FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
292 POSTING_READ(FW_BLC_SELF);
293 } else if (IS_PINEVIEW(dev)) {
294 val = I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN;
295 val |= enable ? PINEVIEW_SELF_REFRESH_EN : 0;
296 I915_WRITE(DSPFW3, val);
297 POSTING_READ(DSPFW3);
298 } else if (IS_I945G(dev) || IS_I945GM(dev)) {
299 val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
300 _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
301 I915_WRITE(FW_BLC_SELF, val);
302 POSTING_READ(FW_BLC_SELF);
303 } else if (IS_I915GM(dev)) {
304 val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
305 _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
306 I915_WRITE(INSTPM, val);
307 POSTING_READ(INSTPM);
308 } else {
309 return;
310 }
311
312 DRM_DEBUG_KMS("memory self-refresh is %s\n",
313 enable ? "enabled" : "disabled");
314 }
315
316
317 /*
318 * Latency for FIFO fetches is dependent on several factors:
319 * - memory configuration (speed, channels)
320 * - chipset
321 * - current MCH state
322 * It can be fairly high in some situations, so here we assume a fairly
323 * pessimal value. It's a tradeoff between extra memory fetches (if we
324 * set this value too high, the FIFO will fetch frequently to stay full)
325 * and power consumption (set it too low to save power and we might see
326 * FIFO underruns and display "flicker").
327 *
328 * A value of 5us seems to be a good balance; safe for very low end
329 * platforms but not overly aggressive on lower latency configs.
330 */
331 static const int pessimal_latency_ns = 5000;
332
333 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
334 ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
335
336 static int vlv_get_fifo_size(struct drm_device *dev,
337 enum pipe pipe, int plane)
338 {
339 struct drm_i915_private *dev_priv = dev->dev_private;
340 int sprite0_start, sprite1_start, size;
341
342 switch (pipe) {
343 uint32_t dsparb, dsparb2, dsparb3;
344 case PIPE_A:
345 dsparb = I915_READ(DSPARB);
346 dsparb2 = I915_READ(DSPARB2);
347 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 0, 0);
348 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 8, 4);
349 break;
350 case PIPE_B:
351 dsparb = I915_READ(DSPARB);
352 dsparb2 = I915_READ(DSPARB2);
353 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 16, 8);
354 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 24, 12);
355 break;
356 case PIPE_C:
357 dsparb2 = I915_READ(DSPARB2);
358 dsparb3 = I915_READ(DSPARB3);
359 sprite0_start = VLV_FIFO_START(dsparb3, dsparb2, 0, 16);
360 sprite1_start = VLV_FIFO_START(dsparb3, dsparb2, 8, 20);
361 break;
362 default:
363 return 0;
364 }
365
366 switch (plane) {
367 case 0:
368 size = sprite0_start;
369 break;
370 case 1:
371 size = sprite1_start - sprite0_start;
372 break;
373 case 2:
374 size = 512 - 1 - sprite1_start;
375 break;
376 default:
377 return 0;
378 }
379
380 DRM_DEBUG_KMS("Pipe %c %s %c FIFO size: %d\n",
381 pipe_name(pipe), plane == 0 ? "primary" : "sprite",
382 plane == 0 ? plane_name(pipe) : sprite_name(pipe, plane - 1),
383 size);
384
385 return size;
386 }
387
388 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
389 {
390 struct drm_i915_private *dev_priv = dev->dev_private;
391 uint32_t dsparb = I915_READ(DSPARB);
392 int size;
393
394 size = dsparb & 0x7f;
395 if (plane)
396 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
397
398 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
399 plane ? "B" : "A", size);
400
401 return size;
402 }
403
404 static int i830_get_fifo_size(struct drm_device *dev, int plane)
405 {
406 struct drm_i915_private *dev_priv = dev->dev_private;
407 uint32_t dsparb = I915_READ(DSPARB);
408 int size;
409
410 size = dsparb & 0x1ff;
411 if (plane)
412 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
413 size >>= 1; /* Convert to cachelines */
414
415 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
416 plane ? "B" : "A", size);
417
418 return size;
419 }
420
421 static int i845_get_fifo_size(struct drm_device *dev, int plane)
422 {
423 struct drm_i915_private *dev_priv = dev->dev_private;
424 uint32_t dsparb = I915_READ(DSPARB);
425 int size;
426
427 size = dsparb & 0x7f;
428 size >>= 2; /* Convert to cachelines */
429
430 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
431 plane ? "B" : "A",
432 size);
433
434 return size;
435 }
436
437 /* Pineview has different values for various configs */
438 static const struct intel_watermark_params pineview_display_wm = {
439 .fifo_size = PINEVIEW_DISPLAY_FIFO,
440 .max_wm = PINEVIEW_MAX_WM,
441 .default_wm = PINEVIEW_DFT_WM,
442 .guard_size = PINEVIEW_GUARD_WM,
443 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
444 };
445 static const struct intel_watermark_params pineview_display_hplloff_wm = {
446 .fifo_size = PINEVIEW_DISPLAY_FIFO,
447 .max_wm = PINEVIEW_MAX_WM,
448 .default_wm = PINEVIEW_DFT_HPLLOFF_WM,
449 .guard_size = PINEVIEW_GUARD_WM,
450 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
451 };
452 static const struct intel_watermark_params pineview_cursor_wm = {
453 .fifo_size = PINEVIEW_CURSOR_FIFO,
454 .max_wm = PINEVIEW_CURSOR_MAX_WM,
455 .default_wm = PINEVIEW_CURSOR_DFT_WM,
456 .guard_size = PINEVIEW_CURSOR_GUARD_WM,
457 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
458 };
459 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
460 .fifo_size = PINEVIEW_CURSOR_FIFO,
461 .max_wm = PINEVIEW_CURSOR_MAX_WM,
462 .default_wm = PINEVIEW_CURSOR_DFT_WM,
463 .guard_size = PINEVIEW_CURSOR_GUARD_WM,
464 .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
465 };
466 static const struct intel_watermark_params g4x_wm_info = {
467 .fifo_size = G4X_FIFO_SIZE,
468 .max_wm = G4X_MAX_WM,
469 .default_wm = G4X_MAX_WM,
470 .guard_size = 2,
471 .cacheline_size = G4X_FIFO_LINE_SIZE,
472 };
473 static const struct intel_watermark_params g4x_cursor_wm_info = {
474 .fifo_size = I965_CURSOR_FIFO,
475 .max_wm = I965_CURSOR_MAX_WM,
476 .default_wm = I965_CURSOR_DFT_WM,
477 .guard_size = 2,
478 .cacheline_size = G4X_FIFO_LINE_SIZE,
479 };
480 static const struct intel_watermark_params valleyview_wm_info = {
481 .fifo_size = VALLEYVIEW_FIFO_SIZE,
482 .max_wm = VALLEYVIEW_MAX_WM,
483 .default_wm = VALLEYVIEW_MAX_WM,
484 .guard_size = 2,
485 .cacheline_size = G4X_FIFO_LINE_SIZE,
486 };
487 static const struct intel_watermark_params valleyview_cursor_wm_info = {
488 .fifo_size = I965_CURSOR_FIFO,
489 .max_wm = VALLEYVIEW_CURSOR_MAX_WM,
490 .default_wm = I965_CURSOR_DFT_WM,
491 .guard_size = 2,
492 .cacheline_size = G4X_FIFO_LINE_SIZE,
493 };
494 static const struct intel_watermark_params i965_cursor_wm_info = {
495 .fifo_size = I965_CURSOR_FIFO,
496 .max_wm = I965_CURSOR_MAX_WM,
497 .default_wm = I965_CURSOR_DFT_WM,
498 .guard_size = 2,
499 .cacheline_size = I915_FIFO_LINE_SIZE,
500 };
501 static const struct intel_watermark_params i945_wm_info = {
502 .fifo_size = I945_FIFO_SIZE,
503 .max_wm = I915_MAX_WM,
504 .default_wm = 1,
505 .guard_size = 2,
506 .cacheline_size = I915_FIFO_LINE_SIZE,
507 };
508 static const struct intel_watermark_params i915_wm_info = {
509 .fifo_size = I915_FIFO_SIZE,
510 .max_wm = I915_MAX_WM,
511 .default_wm = 1,
512 .guard_size = 2,
513 .cacheline_size = I915_FIFO_LINE_SIZE,
514 };
515 static const struct intel_watermark_params i830_a_wm_info = {
516 .fifo_size = I855GM_FIFO_SIZE,
517 .max_wm = I915_MAX_WM,
518 .default_wm = 1,
519 .guard_size = 2,
520 .cacheline_size = I830_FIFO_LINE_SIZE,
521 };
522 static const struct intel_watermark_params i830_bc_wm_info = {
523 .fifo_size = I855GM_FIFO_SIZE,
524 .max_wm = I915_MAX_WM/2,
525 .default_wm = 1,
526 .guard_size = 2,
527 .cacheline_size = I830_FIFO_LINE_SIZE,
528 };
529 static const struct intel_watermark_params i845_wm_info = {
530 .fifo_size = I830_FIFO_SIZE,
531 .max_wm = I915_MAX_WM,
532 .default_wm = 1,
533 .guard_size = 2,
534 .cacheline_size = I830_FIFO_LINE_SIZE,
535 };
536
537 /**
538 * intel_calculate_wm - calculate watermark level
539 * @clock_in_khz: pixel clock
540 * @wm: chip FIFO params
541 * @pixel_size: display pixel size
542 * @latency_ns: memory latency for the platform
543 *
544 * Calculate the watermark level (the level at which the display plane will
545 * start fetching from memory again). Each chip has a different display
546 * FIFO size and allocation, so the caller needs to figure that out and pass
547 * in the correct intel_watermark_params structure.
548 *
549 * As the pixel clock runs, the FIFO will be drained at a rate that depends
550 * on the pixel size. When it reaches the watermark level, it'll start
551 * fetching FIFO line sized based chunks from memory until the FIFO fills
552 * past the watermark point. If the FIFO drains completely, a FIFO underrun
553 * will occur, and a display engine hang could result.
554 */
555 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
556 const struct intel_watermark_params *wm,
557 int fifo_size,
558 int pixel_size,
559 unsigned long latency_ns)
560 {
561 long entries_required, wm_size;
562
563 /*
564 * Note: we need to make sure we don't overflow for various clock &
565 * latency values.
566 * clocks go from a few thousand to several hundred thousand.
567 * latency is usually a few thousand
568 */
569 entries_required = ((clock_in_khz / 1000) * pixel_size * latency_ns) /
570 1000;
571 entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
572
573 DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
574
575 wm_size = fifo_size - (entries_required + wm->guard_size);
576
577 DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
578
579 /* Don't promote wm_size to unsigned... */
580 if (wm_size > (long)wm->max_wm)
581 wm_size = wm->max_wm;
582 if (wm_size <= 0)
583 wm_size = wm->default_wm;
584
585 /*
586 * Bspec seems to indicate that the value shouldn't be lower than
587 * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
588 * Lets go for 8 which is the burst size since certain platforms
589 * already use a hardcoded 8 (which is what the spec says should be
590 * done).
591 */
592 if (wm_size <= 8)
593 wm_size = 8;
594
595 return wm_size;
596 }
597
598 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
599 {
600 struct drm_crtc *crtc, *enabled = NULL;
601
602 for_each_crtc(dev, crtc) {
603 if (intel_crtc_active(crtc)) {
604 if (enabled)
605 return NULL;
606 enabled = crtc;
607 }
608 }
609
610 return enabled;
611 }
612
613 static void pineview_update_wm(struct drm_crtc *unused_crtc)
614 {
615 struct drm_device *dev = unused_crtc->dev;
616 struct drm_i915_private *dev_priv = dev->dev_private;
617 struct drm_crtc *crtc;
618 const struct cxsr_latency *latency;
619 u32 reg;
620 unsigned long wm;
621
622 latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
623 dev_priv->fsb_freq, dev_priv->mem_freq);
624 if (!latency) {
625 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
626 intel_set_memory_cxsr(dev_priv, false);
627 return;
628 }
629
630 crtc = single_enabled_crtc(dev);
631 if (crtc) {
632 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
633 int pixel_size = crtc->primary->state->fb->bits_per_pixel / 8;
634 int clock = adjusted_mode->crtc_clock;
635
636 /* Display SR */
637 wm = intel_calculate_wm(clock, &pineview_display_wm,
638 pineview_display_wm.fifo_size,
639 pixel_size, latency->display_sr);
640 reg = I915_READ(DSPFW1);
641 reg &= ~DSPFW_SR_MASK;
642 reg |= FW_WM(wm, SR);
643 I915_WRITE(DSPFW1, reg);
644 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
645
646 /* cursor SR */
647 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
648 pineview_display_wm.fifo_size,
649 pixel_size, latency->cursor_sr);
650 reg = I915_READ(DSPFW3);
651 reg &= ~DSPFW_CURSOR_SR_MASK;
652 reg |= FW_WM(wm, CURSOR_SR);
653 I915_WRITE(DSPFW3, reg);
654
655 /* Display HPLL off SR */
656 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
657 pineview_display_hplloff_wm.fifo_size,
658 pixel_size, latency->display_hpll_disable);
659 reg = I915_READ(DSPFW3);
660 reg &= ~DSPFW_HPLL_SR_MASK;
661 reg |= FW_WM(wm, HPLL_SR);
662 I915_WRITE(DSPFW3, reg);
663
664 /* cursor HPLL off SR */
665 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
666 pineview_display_hplloff_wm.fifo_size,
667 pixel_size, latency->cursor_hpll_disable);
668 reg = I915_READ(DSPFW3);
669 reg &= ~DSPFW_HPLL_CURSOR_MASK;
670 reg |= FW_WM(wm, HPLL_CURSOR);
671 I915_WRITE(DSPFW3, reg);
672 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
673
674 intel_set_memory_cxsr(dev_priv, true);
675 } else {
676 intel_set_memory_cxsr(dev_priv, false);
677 }
678 }
679
680 static bool g4x_compute_wm0(struct drm_device *dev,
681 int plane,
682 const struct intel_watermark_params *display,
683 int display_latency_ns,
684 const struct intel_watermark_params *cursor,
685 int cursor_latency_ns,
686 int *plane_wm,
687 int *cursor_wm)
688 {
689 struct drm_crtc *crtc;
690 const struct drm_display_mode *adjusted_mode;
691 int htotal, hdisplay, clock, pixel_size;
692 int line_time_us, line_count;
693 int entries, tlb_miss;
694
695 crtc = intel_get_crtc_for_plane(dev, plane);
696 if (!intel_crtc_active(crtc)) {
697 *cursor_wm = cursor->guard_size;
698 *plane_wm = display->guard_size;
699 return false;
700 }
701
702 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
703 clock = adjusted_mode->crtc_clock;
704 htotal = adjusted_mode->crtc_htotal;
705 hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
706 pixel_size = crtc->primary->state->fb->bits_per_pixel / 8;
707
708 /* Use the small buffer method to calculate plane watermark */
709 entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
710 tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
711 if (tlb_miss > 0)
712 entries += tlb_miss;
713 entries = DIV_ROUND_UP(entries, display->cacheline_size);
714 *plane_wm = entries + display->guard_size;
715 if (*plane_wm > (int)display->max_wm)
716 *plane_wm = display->max_wm;
717
718 /* Use the large buffer method to calculate cursor watermark */
719 line_time_us = max(htotal * 1000 / clock, 1);
720 line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
721 entries = line_count * crtc->cursor->state->crtc_w * pixel_size;
722 tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
723 if (tlb_miss > 0)
724 entries += tlb_miss;
725 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
726 *cursor_wm = entries + cursor->guard_size;
727 if (*cursor_wm > (int)cursor->max_wm)
728 *cursor_wm = (int)cursor->max_wm;
729
730 return true;
731 }
732
733 /*
734 * Check the wm result.
735 *
736 * If any calculated watermark values is larger than the maximum value that
737 * can be programmed into the associated watermark register, that watermark
738 * must be disabled.
739 */
740 static bool g4x_check_srwm(struct drm_device *dev,
741 int display_wm, int cursor_wm,
742 const struct intel_watermark_params *display,
743 const struct intel_watermark_params *cursor)
744 {
745 DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
746 display_wm, cursor_wm);
747
748 if (display_wm > display->max_wm) {
749 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
750 display_wm, display->max_wm);
751 return false;
752 }
753
754 if (cursor_wm > cursor->max_wm) {
755 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
756 cursor_wm, cursor->max_wm);
757 return false;
758 }
759
760 if (!(display_wm || cursor_wm)) {
761 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
762 return false;
763 }
764
765 return true;
766 }
767
768 static bool g4x_compute_srwm(struct drm_device *dev,
769 int plane,
770 int latency_ns,
771 const struct intel_watermark_params *display,
772 const struct intel_watermark_params *cursor,
773 int *display_wm, int *cursor_wm)
774 {
775 struct drm_crtc *crtc;
776 const struct drm_display_mode *adjusted_mode;
777 int hdisplay, htotal, pixel_size, clock;
778 unsigned long line_time_us;
779 int line_count, line_size;
780 int small, large;
781 int entries;
782
783 if (!latency_ns) {
784 *display_wm = *cursor_wm = 0;
785 return false;
786 }
787
788 crtc = intel_get_crtc_for_plane(dev, plane);
789 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
790 clock = adjusted_mode->crtc_clock;
791 htotal = adjusted_mode->crtc_htotal;
792 hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
793 pixel_size = crtc->primary->state->fb->bits_per_pixel / 8;
794
795 line_time_us = max(htotal * 1000 / clock, 1);
796 line_count = (latency_ns / line_time_us + 1000) / 1000;
797 line_size = hdisplay * pixel_size;
798
799 /* Use the minimum of the small and large buffer method for primary */
800 small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
801 large = line_count * line_size;
802
803 entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
804 *display_wm = entries + display->guard_size;
805
806 /* calculate the self-refresh watermark for display cursor */
807 entries = line_count * pixel_size * crtc->cursor->state->crtc_w;
808 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
809 *cursor_wm = entries + cursor->guard_size;
810
811 return g4x_check_srwm(dev,
812 *display_wm, *cursor_wm,
813 display, cursor);
814 }
815
816 #define FW_WM_VLV(value, plane) \
817 (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
818
819 static void vlv_write_wm_values(struct intel_crtc *crtc,
820 const struct vlv_wm_values *wm)
821 {
822 struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
823 enum pipe pipe = crtc->pipe;
824
825 I915_WRITE(VLV_DDL(pipe),
826 (wm->ddl[pipe].cursor << DDL_CURSOR_SHIFT) |
827 (wm->ddl[pipe].sprite[1] << DDL_SPRITE_SHIFT(1)) |
828 (wm->ddl[pipe].sprite[0] << DDL_SPRITE_SHIFT(0)) |
829 (wm->ddl[pipe].primary << DDL_PLANE_SHIFT));
830
831 I915_WRITE(DSPFW1,
832 FW_WM(wm->sr.plane, SR) |
833 FW_WM(wm->pipe[PIPE_B].cursor, CURSORB) |
834 FW_WM_VLV(wm->pipe[PIPE_B].primary, PLANEB) |
835 FW_WM_VLV(wm->pipe[PIPE_A].primary, PLANEA));
836 I915_WRITE(DSPFW2,
837 FW_WM_VLV(wm->pipe[PIPE_A].sprite[1], SPRITEB) |
838 FW_WM(wm->pipe[PIPE_A].cursor, CURSORA) |
839 FW_WM_VLV(wm->pipe[PIPE_A].sprite[0], SPRITEA));
840 I915_WRITE(DSPFW3,
841 FW_WM(wm->sr.cursor, CURSOR_SR));
842
843 if (IS_CHERRYVIEW(dev_priv)) {
844 I915_WRITE(DSPFW7_CHV,
845 FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
846 FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
847 I915_WRITE(DSPFW8_CHV,
848 FW_WM_VLV(wm->pipe[PIPE_C].sprite[1], SPRITEF) |
849 FW_WM_VLV(wm->pipe[PIPE_C].sprite[0], SPRITEE));
850 I915_WRITE(DSPFW9_CHV,
851 FW_WM_VLV(wm->pipe[PIPE_C].primary, PLANEC) |
852 FW_WM(wm->pipe[PIPE_C].cursor, CURSORC));
853 I915_WRITE(DSPHOWM,
854 FW_WM(wm->sr.plane >> 9, SR_HI) |
855 FW_WM(wm->pipe[PIPE_C].sprite[1] >> 8, SPRITEF_HI) |
856 FW_WM(wm->pipe[PIPE_C].sprite[0] >> 8, SPRITEE_HI) |
857 FW_WM(wm->pipe[PIPE_C].primary >> 8, PLANEC_HI) |
858 FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
859 FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
860 FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
861 FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
862 FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
863 FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
864 } else {
865 I915_WRITE(DSPFW7,
866 FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
867 FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
868 I915_WRITE(DSPHOWM,
869 FW_WM(wm->sr.plane >> 9, SR_HI) |
870 FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
871 FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
872 FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
873 FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
874 FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
875 FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
876 }
877
878 /* zero (unused) WM1 watermarks */
879 I915_WRITE(DSPFW4, 0);
880 I915_WRITE(DSPFW5, 0);
881 I915_WRITE(DSPFW6, 0);
882 I915_WRITE(DSPHOWM1, 0);
883
884 POSTING_READ(DSPFW1);
885 }
886
887 #undef FW_WM_VLV
888
889 enum vlv_wm_level {
890 VLV_WM_LEVEL_PM2,
891 VLV_WM_LEVEL_PM5,
892 VLV_WM_LEVEL_DDR_DVFS,
893 };
894
895 /* latency must be in 0.1us units. */
896 static unsigned int vlv_wm_method2(unsigned int pixel_rate,
897 unsigned int pipe_htotal,
898 unsigned int horiz_pixels,
899 unsigned int bytes_per_pixel,
900 unsigned int latency)
901 {
902 unsigned int ret;
903
904 ret = (latency * pixel_rate) / (pipe_htotal * 10000);
905 ret = (ret + 1) * horiz_pixels * bytes_per_pixel;
906 ret = DIV_ROUND_UP(ret, 64);
907
908 return ret;
909 }
910
911 static void vlv_setup_wm_latency(struct drm_device *dev)
912 {
913 struct drm_i915_private *dev_priv = dev->dev_private;
914
915 /* all latencies in usec */
916 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM2] = 3;
917
918 dev_priv->wm.max_level = VLV_WM_LEVEL_PM2;
919
920 if (IS_CHERRYVIEW(dev_priv)) {
921 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM5] = 12;
922 dev_priv->wm.pri_latency[VLV_WM_LEVEL_DDR_DVFS] = 33;
923
924 dev_priv->wm.max_level = VLV_WM_LEVEL_DDR_DVFS;
925 }
926 }
927
928 static uint16_t vlv_compute_wm_level(struct intel_plane *plane,
929 struct intel_crtc *crtc,
930 const struct intel_plane_state *state,
931 int level)
932 {
933 struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
934 int clock, htotal, pixel_size, width, wm;
935
936 if (dev_priv->wm.pri_latency[level] == 0)
937 return USHRT_MAX;
938
939 if (!state->visible)
940 return 0;
941
942 pixel_size = drm_format_plane_cpp(state->base.fb->pixel_format, 0);
943 clock = crtc->config->base.adjusted_mode.crtc_clock;
944 htotal = crtc->config->base.adjusted_mode.crtc_htotal;
945 width = crtc->config->pipe_src_w;
946 if (WARN_ON(htotal == 0))
947 htotal = 1;
948
949 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
950 /*
951 * FIXME the formula gives values that are
952 * too big for the cursor FIFO, and hence we
953 * would never be able to use cursors. For
954 * now just hardcode the watermark.
955 */
956 wm = 63;
957 } else {
958 wm = vlv_wm_method2(clock, htotal, width, pixel_size,
959 dev_priv->wm.pri_latency[level] * 10);
960 }
961
962 return min_t(int, wm, USHRT_MAX);
963 }
964
965 static void vlv_compute_fifo(struct intel_crtc *crtc)
966 {
967 struct drm_device *dev = crtc->base.dev;
968 struct vlv_wm_state *wm_state = &crtc->wm_state;
969 struct intel_plane *plane;
970 unsigned int total_rate = 0;
971 const int fifo_size = 512 - 1;
972 int fifo_extra, fifo_left = fifo_size;
973
974 for_each_intel_plane_on_crtc(dev, crtc, plane) {
975 struct intel_plane_state *state =
976 to_intel_plane_state(plane->base.state);
977
978 if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
979 continue;
980
981 if (state->visible) {
982 wm_state->num_active_planes++;
983 total_rate += drm_format_plane_cpp(state->base.fb->pixel_format, 0);
984 }
985 }
986
987 for_each_intel_plane_on_crtc(dev, crtc, plane) {
988 struct intel_plane_state *state =
989 to_intel_plane_state(plane->base.state);
990 unsigned int rate;
991
992 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
993 plane->wm.fifo_size = 63;
994 continue;
995 }
996
997 if (!state->visible) {
998 plane->wm.fifo_size = 0;
999 continue;
1000 }
1001
1002 rate = drm_format_plane_cpp(state->base.fb->pixel_format, 0);
1003 plane->wm.fifo_size = fifo_size * rate / total_rate;
1004 fifo_left -= plane->wm.fifo_size;
1005 }
1006
1007 fifo_extra = DIV_ROUND_UP(fifo_left, wm_state->num_active_planes ?: 1);
1008
1009 /* spread the remainder evenly */
1010 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1011 int plane_extra;
1012
1013 if (fifo_left == 0)
1014 break;
1015
1016 if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
1017 continue;
1018
1019 /* give it all to the first plane if none are active */
1020 if (plane->wm.fifo_size == 0 &&
1021 wm_state->num_active_planes)
1022 continue;
1023
1024 plane_extra = min(fifo_extra, fifo_left);
1025 plane->wm.fifo_size += plane_extra;
1026 fifo_left -= plane_extra;
1027 }
1028
1029 WARN_ON(fifo_left != 0);
1030 }
1031
1032 static void vlv_invert_wms(struct intel_crtc *crtc)
1033 {
1034 struct vlv_wm_state *wm_state = &crtc->wm_state;
1035 int level;
1036
1037 for (level = 0; level < wm_state->num_levels; level++) {
1038 struct drm_device *dev = crtc->base.dev;
1039 const int sr_fifo_size = INTEL_INFO(dev)->num_pipes * 512 - 1;
1040 struct intel_plane *plane;
1041
1042 wm_state->sr[level].plane = sr_fifo_size - wm_state->sr[level].plane;
1043 wm_state->sr[level].cursor = 63 - wm_state->sr[level].cursor;
1044
1045 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1046 switch (plane->base.type) {
1047 int sprite;
1048 case DRM_PLANE_TYPE_CURSOR:
1049 wm_state->wm[level].cursor = plane->wm.fifo_size -
1050 wm_state->wm[level].cursor;
1051 break;
1052 case DRM_PLANE_TYPE_PRIMARY:
1053 wm_state->wm[level].primary = plane->wm.fifo_size -
1054 wm_state->wm[level].primary;
1055 break;
1056 case DRM_PLANE_TYPE_OVERLAY:
1057 sprite = plane->plane;
1058 wm_state->wm[level].sprite[sprite] = plane->wm.fifo_size -
1059 wm_state->wm[level].sprite[sprite];
1060 break;
1061 }
1062 }
1063 }
1064 }
1065
1066 static void vlv_compute_wm(struct intel_crtc *crtc)
1067 {
1068 struct drm_device *dev = crtc->base.dev;
1069 struct vlv_wm_state *wm_state = &crtc->wm_state;
1070 struct intel_plane *plane;
1071 int sr_fifo_size = INTEL_INFO(dev)->num_pipes * 512 - 1;
1072 int level;
1073
1074 memset(wm_state, 0, sizeof(*wm_state));
1075
1076 wm_state->cxsr = crtc->pipe != PIPE_C && crtc->wm.cxsr_allowed;
1077 wm_state->num_levels = to_i915(dev)->wm.max_level + 1;
1078
1079 wm_state->num_active_planes = 0;
1080
1081 vlv_compute_fifo(crtc);
1082
1083 if (wm_state->num_active_planes != 1)
1084 wm_state->cxsr = false;
1085
1086 if (wm_state->cxsr) {
1087 for (level = 0; level < wm_state->num_levels; level++) {
1088 wm_state->sr[level].plane = sr_fifo_size;
1089 wm_state->sr[level].cursor = 63;
1090 }
1091 }
1092
1093 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1094 struct intel_plane_state *state =
1095 to_intel_plane_state(plane->base.state);
1096
1097 if (!state->visible)
1098 continue;
1099
1100 /* normal watermarks */
1101 for (level = 0; level < wm_state->num_levels; level++) {
1102 int wm = vlv_compute_wm_level(plane, crtc, state, level);
1103 int max_wm = plane->base.type == DRM_PLANE_TYPE_CURSOR ? 63 : 511;
1104
1105 /* hack */
1106 if (WARN_ON(level == 0 && wm > max_wm))
1107 wm = max_wm;
1108
1109 if (wm > plane->wm.fifo_size)
1110 break;
1111
1112 switch (plane->base.type) {
1113 int sprite;
1114 case DRM_PLANE_TYPE_CURSOR:
1115 wm_state->wm[level].cursor = wm;
1116 break;
1117 case DRM_PLANE_TYPE_PRIMARY:
1118 wm_state->wm[level].primary = wm;
1119 break;
1120 case DRM_PLANE_TYPE_OVERLAY:
1121 sprite = plane->plane;
1122 wm_state->wm[level].sprite[sprite] = wm;
1123 break;
1124 }
1125 }
1126
1127 wm_state->num_levels = level;
1128
1129 if (!wm_state->cxsr)
1130 continue;
1131
1132 /* maxfifo watermarks */
1133 switch (plane->base.type) {
1134 int sprite, level;
1135 case DRM_PLANE_TYPE_CURSOR:
1136 for (level = 0; level < wm_state->num_levels; level++)
1137 wm_state->sr[level].cursor =
1138 wm_state->wm[level].cursor;
1139 break;
1140 case DRM_PLANE_TYPE_PRIMARY:
1141 for (level = 0; level < wm_state->num_levels; level++)
1142 wm_state->sr[level].plane =
1143 min(wm_state->sr[level].plane,
1144 wm_state->wm[level].primary);
1145 break;
1146 case DRM_PLANE_TYPE_OVERLAY:
1147 sprite = plane->plane;
1148 for (level = 0; level < wm_state->num_levels; level++)
1149 wm_state->sr[level].plane =
1150 min(wm_state->sr[level].plane,
1151 wm_state->wm[level].sprite[sprite]);
1152 break;
1153 }
1154 }
1155
1156 /* clear any (partially) filled invalid levels */
1157 for (level = wm_state->num_levels; level < to_i915(dev)->wm.max_level + 1; level++) {
1158 memset(&wm_state->wm[level], 0, sizeof(wm_state->wm[level]));
1159 memset(&wm_state->sr[level], 0, sizeof(wm_state->sr[level]));
1160 }
1161
1162 vlv_invert_wms(crtc);
1163 }
1164
1165 #define VLV_FIFO(plane, value) \
1166 (((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1167
1168 static void vlv_pipe_set_fifo_size(struct intel_crtc *crtc)
1169 {
1170 struct drm_device *dev = crtc->base.dev;
1171 struct drm_i915_private *dev_priv = to_i915(dev);
1172 struct intel_plane *plane;
1173 int sprite0_start = 0, sprite1_start = 0, fifo_size = 0;
1174
1175 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1176 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
1177 WARN_ON(plane->wm.fifo_size != 63);
1178 continue;
1179 }
1180
1181 if (plane->base.type == DRM_PLANE_TYPE_PRIMARY)
1182 sprite0_start = plane->wm.fifo_size;
1183 else if (plane->plane == 0)
1184 sprite1_start = sprite0_start + plane->wm.fifo_size;
1185 else
1186 fifo_size = sprite1_start + plane->wm.fifo_size;
1187 }
1188
1189 WARN_ON(fifo_size != 512 - 1);
1190
1191 DRM_DEBUG_KMS("Pipe %c FIFO split %d / %d / %d\n",
1192 pipe_name(crtc->pipe), sprite0_start,
1193 sprite1_start, fifo_size);
1194
1195 switch (crtc->pipe) {
1196 uint32_t dsparb, dsparb2, dsparb3;
1197 case PIPE_A:
1198 dsparb = I915_READ(DSPARB);
1199 dsparb2 = I915_READ(DSPARB2);
1200
1201 dsparb &= ~(VLV_FIFO(SPRITEA, 0xff) |
1202 VLV_FIFO(SPRITEB, 0xff));
1203 dsparb |= (VLV_FIFO(SPRITEA, sprite0_start) |
1204 VLV_FIFO(SPRITEB, sprite1_start));
1205
1206 dsparb2 &= ~(VLV_FIFO(SPRITEA_HI, 0x1) |
1207 VLV_FIFO(SPRITEB_HI, 0x1));
1208 dsparb2 |= (VLV_FIFO(SPRITEA_HI, sprite0_start >> 8) |
1209 VLV_FIFO(SPRITEB_HI, sprite1_start >> 8));
1210
1211 I915_WRITE(DSPARB, dsparb);
1212 I915_WRITE(DSPARB2, dsparb2);
1213 break;
1214 case PIPE_B:
1215 dsparb = I915_READ(DSPARB);
1216 dsparb2 = I915_READ(DSPARB2);
1217
1218 dsparb &= ~(VLV_FIFO(SPRITEC, 0xff) |
1219 VLV_FIFO(SPRITED, 0xff));
1220 dsparb |= (VLV_FIFO(SPRITEC, sprite0_start) |
1221 VLV_FIFO(SPRITED, sprite1_start));
1222
1223 dsparb2 &= ~(VLV_FIFO(SPRITEC_HI, 0xff) |
1224 VLV_FIFO(SPRITED_HI, 0xff));
1225 dsparb2 |= (VLV_FIFO(SPRITEC_HI, sprite0_start >> 8) |
1226 VLV_FIFO(SPRITED_HI, sprite1_start >> 8));
1227
1228 I915_WRITE(DSPARB, dsparb);
1229 I915_WRITE(DSPARB2, dsparb2);
1230 break;
1231 case PIPE_C:
1232 dsparb3 = I915_READ(DSPARB3);
1233 dsparb2 = I915_READ(DSPARB2);
1234
1235 dsparb3 &= ~(VLV_FIFO(SPRITEE, 0xff) |
1236 VLV_FIFO(SPRITEF, 0xff));
1237 dsparb3 |= (VLV_FIFO(SPRITEE, sprite0_start) |
1238 VLV_FIFO(SPRITEF, sprite1_start));
1239
1240 dsparb2 &= ~(VLV_FIFO(SPRITEE_HI, 0xff) |
1241 VLV_FIFO(SPRITEF_HI, 0xff));
1242 dsparb2 |= (VLV_FIFO(SPRITEE_HI, sprite0_start >> 8) |
1243 VLV_FIFO(SPRITEF_HI, sprite1_start >> 8));
1244
1245 I915_WRITE(DSPARB3, dsparb3);
1246 I915_WRITE(DSPARB2, dsparb2);
1247 break;
1248 default:
1249 break;
1250 }
1251 }
1252
1253 #undef VLV_FIFO
1254
1255 static void vlv_merge_wm(struct drm_device *dev,
1256 struct vlv_wm_values *wm)
1257 {
1258 struct intel_crtc *crtc;
1259 int num_active_crtcs = 0;
1260
1261 wm->level = to_i915(dev)->wm.max_level;
1262 wm->cxsr = true;
1263
1264 for_each_intel_crtc(dev, crtc) {
1265 const struct vlv_wm_state *wm_state = &crtc->wm_state;
1266
1267 if (!crtc->active)
1268 continue;
1269
1270 if (!wm_state->cxsr)
1271 wm->cxsr = false;
1272
1273 num_active_crtcs++;
1274 wm->level = min_t(int, wm->level, wm_state->num_levels - 1);
1275 }
1276
1277 if (num_active_crtcs != 1)
1278 wm->cxsr = false;
1279
1280 if (num_active_crtcs > 1)
1281 wm->level = VLV_WM_LEVEL_PM2;
1282
1283 for_each_intel_crtc(dev, crtc) {
1284 struct vlv_wm_state *wm_state = &crtc->wm_state;
1285 enum pipe pipe = crtc->pipe;
1286
1287 if (!crtc->active)
1288 continue;
1289
1290 wm->pipe[pipe] = wm_state->wm[wm->level];
1291 if (wm->cxsr)
1292 wm->sr = wm_state->sr[wm->level];
1293
1294 wm->ddl[pipe].primary = DDL_PRECISION_HIGH | 2;
1295 wm->ddl[pipe].sprite[0] = DDL_PRECISION_HIGH | 2;
1296 wm->ddl[pipe].sprite[1] = DDL_PRECISION_HIGH | 2;
1297 wm->ddl[pipe].cursor = DDL_PRECISION_HIGH | 2;
1298 }
1299 }
1300
1301 static void vlv_update_wm(struct drm_crtc *crtc)
1302 {
1303 struct drm_device *dev = crtc->dev;
1304 struct drm_i915_private *dev_priv = dev->dev_private;
1305 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1306 enum pipe pipe = intel_crtc->pipe;
1307 struct vlv_wm_values wm = {};
1308
1309 vlv_compute_wm(intel_crtc);
1310 vlv_merge_wm(dev, &wm);
1311
1312 if (memcmp(&dev_priv->wm.vlv, &wm, sizeof(wm)) == 0) {
1313 /* FIXME should be part of crtc atomic commit */
1314 vlv_pipe_set_fifo_size(intel_crtc);
1315 return;
1316 }
1317
1318 if (wm.level < VLV_WM_LEVEL_DDR_DVFS &&
1319 dev_priv->wm.vlv.level >= VLV_WM_LEVEL_DDR_DVFS)
1320 chv_set_memory_dvfs(dev_priv, false);
1321
1322 if (wm.level < VLV_WM_LEVEL_PM5 &&
1323 dev_priv->wm.vlv.level >= VLV_WM_LEVEL_PM5)
1324 chv_set_memory_pm5(dev_priv, false);
1325
1326 if (!wm.cxsr && dev_priv->wm.vlv.cxsr)
1327 intel_set_memory_cxsr(dev_priv, false);
1328
1329 /* FIXME should be part of crtc atomic commit */
1330 vlv_pipe_set_fifo_size(intel_crtc);
1331
1332 vlv_write_wm_values(intel_crtc, &wm);
1333
1334 DRM_DEBUG_KMS("Setting FIFO watermarks - %c: plane=%d, cursor=%d, "
1335 "sprite0=%d, sprite1=%d, SR: plane=%d, cursor=%d level=%d cxsr=%d\n",
1336 pipe_name(pipe), wm.pipe[pipe].primary, wm.pipe[pipe].cursor,
1337 wm.pipe[pipe].sprite[0], wm.pipe[pipe].sprite[1],
1338 wm.sr.plane, wm.sr.cursor, wm.level, wm.cxsr);
1339
1340 if (wm.cxsr && !dev_priv->wm.vlv.cxsr)
1341 intel_set_memory_cxsr(dev_priv, true);
1342
1343 if (wm.level >= VLV_WM_LEVEL_PM5 &&
1344 dev_priv->wm.vlv.level < VLV_WM_LEVEL_PM5)
1345 chv_set_memory_pm5(dev_priv, true);
1346
1347 if (wm.level >= VLV_WM_LEVEL_DDR_DVFS &&
1348 dev_priv->wm.vlv.level < VLV_WM_LEVEL_DDR_DVFS)
1349 chv_set_memory_dvfs(dev_priv, true);
1350
1351 dev_priv->wm.vlv = wm;
1352 }
1353
1354 #define single_plane_enabled(mask) is_power_of_2(mask)
1355
1356 static void g4x_update_wm(struct drm_crtc *crtc)
1357 {
1358 struct drm_device *dev = crtc->dev;
1359 static const int sr_latency_ns = 12000;
1360 struct drm_i915_private *dev_priv = dev->dev_private;
1361 int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1362 int plane_sr, cursor_sr;
1363 unsigned int enabled = 0;
1364 bool cxsr_enabled;
1365
1366 if (g4x_compute_wm0(dev, PIPE_A,
1367 &g4x_wm_info, pessimal_latency_ns,
1368 &g4x_cursor_wm_info, pessimal_latency_ns,
1369 &planea_wm, &cursora_wm))
1370 enabled |= 1 << PIPE_A;
1371
1372 if (g4x_compute_wm0(dev, PIPE_B,
1373 &g4x_wm_info, pessimal_latency_ns,
1374 &g4x_cursor_wm_info, pessimal_latency_ns,
1375 &planeb_wm, &cursorb_wm))
1376 enabled |= 1 << PIPE_B;
1377
1378 if (single_plane_enabled(enabled) &&
1379 g4x_compute_srwm(dev, ffs(enabled) - 1,
1380 sr_latency_ns,
1381 &g4x_wm_info,
1382 &g4x_cursor_wm_info,
1383 &plane_sr, &cursor_sr)) {
1384 cxsr_enabled = true;
1385 } else {
1386 cxsr_enabled = false;
1387 intel_set_memory_cxsr(dev_priv, false);
1388 plane_sr = cursor_sr = 0;
1389 }
1390
1391 DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, "
1392 "B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1393 planea_wm, cursora_wm,
1394 planeb_wm, cursorb_wm,
1395 plane_sr, cursor_sr);
1396
1397 I915_WRITE(DSPFW1,
1398 FW_WM(plane_sr, SR) |
1399 FW_WM(cursorb_wm, CURSORB) |
1400 FW_WM(planeb_wm, PLANEB) |
1401 FW_WM(planea_wm, PLANEA));
1402 I915_WRITE(DSPFW2,
1403 (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1404 FW_WM(cursora_wm, CURSORA));
1405 /* HPLL off in SR has some issues on G4x... disable it */
1406 I915_WRITE(DSPFW3,
1407 (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1408 FW_WM(cursor_sr, CURSOR_SR));
1409
1410 if (cxsr_enabled)
1411 intel_set_memory_cxsr(dev_priv, true);
1412 }
1413
1414 static void i965_update_wm(struct drm_crtc *unused_crtc)
1415 {
1416 struct drm_device *dev = unused_crtc->dev;
1417 struct drm_i915_private *dev_priv = dev->dev_private;
1418 struct drm_crtc *crtc;
1419 int srwm = 1;
1420 int cursor_sr = 16;
1421 bool cxsr_enabled;
1422
1423 /* Calc sr entries for one plane configs */
1424 crtc = single_enabled_crtc(dev);
1425 if (crtc) {
1426 /* self-refresh has much higher latency */
1427 static const int sr_latency_ns = 12000;
1428 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1429 int clock = adjusted_mode->crtc_clock;
1430 int htotal = adjusted_mode->crtc_htotal;
1431 int hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
1432 int pixel_size = crtc->primary->state->fb->bits_per_pixel / 8;
1433 unsigned long line_time_us;
1434 int entries;
1435
1436 line_time_us = max(htotal * 1000 / clock, 1);
1437
1438 /* Use ns/us then divide to preserve precision */
1439 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1440 pixel_size * hdisplay;
1441 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1442 srwm = I965_FIFO_SIZE - entries;
1443 if (srwm < 0)
1444 srwm = 1;
1445 srwm &= 0x1ff;
1446 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1447 entries, srwm);
1448
1449 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1450 pixel_size * crtc->cursor->state->crtc_w;
1451 entries = DIV_ROUND_UP(entries,
1452 i965_cursor_wm_info.cacheline_size);
1453 cursor_sr = i965_cursor_wm_info.fifo_size -
1454 (entries + i965_cursor_wm_info.guard_size);
1455
1456 if (cursor_sr > i965_cursor_wm_info.max_wm)
1457 cursor_sr = i965_cursor_wm_info.max_wm;
1458
1459 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1460 "cursor %d\n", srwm, cursor_sr);
1461
1462 cxsr_enabled = true;
1463 } else {
1464 cxsr_enabled = false;
1465 /* Turn off self refresh if both pipes are enabled */
1466 intel_set_memory_cxsr(dev_priv, false);
1467 }
1468
1469 DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1470 srwm);
1471
1472 /* 965 has limitations... */
1473 I915_WRITE(DSPFW1, FW_WM(srwm, SR) |
1474 FW_WM(8, CURSORB) |
1475 FW_WM(8, PLANEB) |
1476 FW_WM(8, PLANEA));
1477 I915_WRITE(DSPFW2, FW_WM(8, CURSORA) |
1478 FW_WM(8, PLANEC_OLD));
1479 /* update cursor SR watermark */
1480 I915_WRITE(DSPFW3, FW_WM(cursor_sr, CURSOR_SR));
1481
1482 if (cxsr_enabled)
1483 intel_set_memory_cxsr(dev_priv, true);
1484 }
1485
1486 #undef FW_WM
1487
1488 static void i9xx_update_wm(struct drm_crtc *unused_crtc)
1489 {
1490 struct drm_device *dev = unused_crtc->dev;
1491 struct drm_i915_private *dev_priv = dev->dev_private;
1492 const struct intel_watermark_params *wm_info;
1493 uint32_t fwater_lo;
1494 uint32_t fwater_hi;
1495 int cwm, srwm = 1;
1496 int fifo_size;
1497 int planea_wm, planeb_wm;
1498 struct drm_crtc *crtc, *enabled = NULL;
1499
1500 if (IS_I945GM(dev))
1501 wm_info = &i945_wm_info;
1502 else if (!IS_GEN2(dev))
1503 wm_info = &i915_wm_info;
1504 else
1505 wm_info = &i830_a_wm_info;
1506
1507 fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1508 crtc = intel_get_crtc_for_plane(dev, 0);
1509 if (intel_crtc_active(crtc)) {
1510 const struct drm_display_mode *adjusted_mode;
1511 int cpp = crtc->primary->state->fb->bits_per_pixel / 8;
1512 if (IS_GEN2(dev))
1513 cpp = 4;
1514
1515 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1516 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1517 wm_info, fifo_size, cpp,
1518 pessimal_latency_ns);
1519 enabled = crtc;
1520 } else {
1521 planea_wm = fifo_size - wm_info->guard_size;
1522 if (planea_wm > (long)wm_info->max_wm)
1523 planea_wm = wm_info->max_wm;
1524 }
1525
1526 if (IS_GEN2(dev))
1527 wm_info = &i830_bc_wm_info;
1528
1529 fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1530 crtc = intel_get_crtc_for_plane(dev, 1);
1531 if (intel_crtc_active(crtc)) {
1532 const struct drm_display_mode *adjusted_mode;
1533 int cpp = crtc->primary->state->fb->bits_per_pixel / 8;
1534 if (IS_GEN2(dev))
1535 cpp = 4;
1536
1537 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1538 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1539 wm_info, fifo_size, cpp,
1540 pessimal_latency_ns);
1541 if (enabled == NULL)
1542 enabled = crtc;
1543 else
1544 enabled = NULL;
1545 } else {
1546 planeb_wm = fifo_size - wm_info->guard_size;
1547 if (planeb_wm > (long)wm_info->max_wm)
1548 planeb_wm = wm_info->max_wm;
1549 }
1550
1551 DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1552
1553 if (IS_I915GM(dev) && enabled) {
1554 struct drm_i915_gem_object *obj;
1555
1556 obj = intel_fb_obj(enabled->primary->state->fb);
1557
1558 /* self-refresh seems busted with untiled */
1559 if (obj->tiling_mode == I915_TILING_NONE)
1560 enabled = NULL;
1561 }
1562
1563 /*
1564 * Overlay gets an aggressive default since video jitter is bad.
1565 */
1566 cwm = 2;
1567
1568 /* Play safe and disable self-refresh before adjusting watermarks. */
1569 intel_set_memory_cxsr(dev_priv, false);
1570
1571 /* Calc sr entries for one plane configs */
1572 if (HAS_FW_BLC(dev) && enabled) {
1573 /* self-refresh has much higher latency */
1574 static const int sr_latency_ns = 6000;
1575 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(enabled)->config->base.adjusted_mode;
1576 int clock = adjusted_mode->crtc_clock;
1577 int htotal = adjusted_mode->crtc_htotal;
1578 int hdisplay = to_intel_crtc(enabled)->config->pipe_src_w;
1579 int pixel_size = enabled->primary->state->fb->bits_per_pixel / 8;
1580 unsigned long line_time_us;
1581 int entries;
1582
1583 line_time_us = max(htotal * 1000 / clock, 1);
1584
1585 /* Use ns/us then divide to preserve precision */
1586 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1587 pixel_size * hdisplay;
1588 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1589 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1590 srwm = wm_info->fifo_size - entries;
1591 if (srwm < 0)
1592 srwm = 1;
1593
1594 if (IS_I945G(dev) || IS_I945GM(dev))
1595 I915_WRITE(FW_BLC_SELF,
1596 FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1597 else if (IS_I915GM(dev))
1598 I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1599 }
1600
1601 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1602 planea_wm, planeb_wm, cwm, srwm);
1603
1604 fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1605 fwater_hi = (cwm & 0x1f);
1606
1607 /* Set request length to 8 cachelines per fetch */
1608 fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1609 fwater_hi = fwater_hi | (1 << 8);
1610
1611 I915_WRITE(FW_BLC, fwater_lo);
1612 I915_WRITE(FW_BLC2, fwater_hi);
1613
1614 if (enabled)
1615 intel_set_memory_cxsr(dev_priv, true);
1616 }
1617
1618 static void i845_update_wm(struct drm_crtc *unused_crtc)
1619 {
1620 struct drm_device *dev = unused_crtc->dev;
1621 struct drm_i915_private *dev_priv = dev->dev_private;
1622 struct drm_crtc *crtc;
1623 const struct drm_display_mode *adjusted_mode;
1624 uint32_t fwater_lo;
1625 int planea_wm;
1626
1627 crtc = single_enabled_crtc(dev);
1628 if (crtc == NULL)
1629 return;
1630
1631 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1632 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1633 &i845_wm_info,
1634 dev_priv->display.get_fifo_size(dev, 0),
1635 4, pessimal_latency_ns);
1636 fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1637 fwater_lo |= (3<<8) | planea_wm;
1638
1639 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1640
1641 I915_WRITE(FW_BLC, fwater_lo);
1642 }
1643
1644 uint32_t ilk_pipe_pixel_rate(const struct intel_crtc_state *pipe_config)
1645 {
1646 uint32_t pixel_rate;
1647
1648 pixel_rate = pipe_config->base.adjusted_mode.crtc_clock;
1649
1650 /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
1651 * adjust the pixel_rate here. */
1652
1653 if (pipe_config->pch_pfit.enabled) {
1654 uint64_t pipe_w, pipe_h, pfit_w, pfit_h;
1655 uint32_t pfit_size = pipe_config->pch_pfit.size;
1656
1657 pipe_w = pipe_config->pipe_src_w;
1658 pipe_h = pipe_config->pipe_src_h;
1659
1660 pfit_w = (pfit_size >> 16) & 0xFFFF;
1661 pfit_h = pfit_size & 0xFFFF;
1662 if (pipe_w < pfit_w)
1663 pipe_w = pfit_w;
1664 if (pipe_h < pfit_h)
1665 pipe_h = pfit_h;
1666
1667 pixel_rate = div_u64((uint64_t) pixel_rate * pipe_w * pipe_h,
1668 pfit_w * pfit_h);
1669 }
1670
1671 return pixel_rate;
1672 }
1673
1674 /* latency must be in 0.1us units. */
1675 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
1676 uint32_t latency)
1677 {
1678 uint64_t ret;
1679
1680 if (WARN(latency == 0, "Latency value missing\n"))
1681 return UINT_MAX;
1682
1683 ret = (uint64_t) pixel_rate * bytes_per_pixel * latency;
1684 ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
1685
1686 return ret;
1687 }
1688
1689 /* latency must be in 0.1us units. */
1690 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
1691 uint32_t horiz_pixels, uint8_t bytes_per_pixel,
1692 uint32_t latency)
1693 {
1694 uint32_t ret;
1695
1696 if (WARN(latency == 0, "Latency value missing\n"))
1697 return UINT_MAX;
1698
1699 ret = (latency * pixel_rate) / (pipe_htotal * 10000);
1700 ret = (ret + 1) * horiz_pixels * bytes_per_pixel;
1701 ret = DIV_ROUND_UP(ret, 64) + 2;
1702 return ret;
1703 }
1704
1705 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
1706 uint8_t bytes_per_pixel)
1707 {
1708 return DIV_ROUND_UP(pri_val * 64, horiz_pixels * bytes_per_pixel) + 2;
1709 }
1710
1711 struct skl_pipe_wm_parameters {
1712 bool active;
1713 uint32_t pipe_htotal;
1714 uint32_t pixel_rate; /* in KHz */
1715 struct intel_plane_wm_parameters plane[I915_MAX_PLANES];
1716 };
1717
1718 struct ilk_wm_maximums {
1719 uint16_t pri;
1720 uint16_t spr;
1721 uint16_t cur;
1722 uint16_t fbc;
1723 };
1724
1725 /* used in computing the new watermarks state */
1726 struct intel_wm_config {
1727 unsigned int num_pipes_active;
1728 bool sprites_enabled;
1729 bool sprites_scaled;
1730 };
1731
1732 /*
1733 * For both WM_PIPE and WM_LP.
1734 * mem_value must be in 0.1us units.
1735 */
1736 static uint32_t ilk_compute_pri_wm(const struct intel_crtc_state *cstate,
1737 const struct intel_plane_state *pstate,
1738 uint32_t mem_value,
1739 bool is_lp)
1740 {
1741 int bpp = pstate->base.fb ? pstate->base.fb->bits_per_pixel / 8 : 0;
1742 uint32_t method1, method2;
1743
1744 if (!cstate->base.active || !pstate->visible)
1745 return 0;
1746
1747 method1 = ilk_wm_method1(ilk_pipe_pixel_rate(cstate), bpp, mem_value);
1748
1749 if (!is_lp)
1750 return method1;
1751
1752 method2 = ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1753 cstate->base.adjusted_mode.crtc_htotal,
1754 drm_rect_width(&pstate->dst),
1755 bpp,
1756 mem_value);
1757
1758 return min(method1, method2);
1759 }
1760
1761 /*
1762 * For both WM_PIPE and WM_LP.
1763 * mem_value must be in 0.1us units.
1764 */
1765 static uint32_t ilk_compute_spr_wm(const struct intel_crtc_state *cstate,
1766 const struct intel_plane_state *pstate,
1767 uint32_t mem_value)
1768 {
1769 int bpp = pstate->base.fb ? pstate->base.fb->bits_per_pixel / 8 : 0;
1770 uint32_t method1, method2;
1771
1772 if (!cstate->base.active || !pstate->visible)
1773 return 0;
1774
1775 method1 = ilk_wm_method1(ilk_pipe_pixel_rate(cstate), bpp, mem_value);
1776 method2 = ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1777 cstate->base.adjusted_mode.crtc_htotal,
1778 drm_rect_width(&pstate->dst),
1779 bpp,
1780 mem_value);
1781 return min(method1, method2);
1782 }
1783
1784 /*
1785 * For both WM_PIPE and WM_LP.
1786 * mem_value must be in 0.1us units.
1787 */
1788 static uint32_t ilk_compute_cur_wm(const struct intel_crtc_state *cstate,
1789 const struct intel_plane_state *pstate,
1790 uint32_t mem_value)
1791 {
1792 int bpp = pstate->base.fb ? pstate->base.fb->bits_per_pixel / 8 : 0;
1793
1794 if (!cstate->base.active || !pstate->visible)
1795 return 0;
1796
1797 return ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1798 cstate->base.adjusted_mode.crtc_htotal,
1799 drm_rect_width(&pstate->dst),
1800 bpp,
1801 mem_value);
1802 }
1803
1804 /* Only for WM_LP. */
1805 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state *cstate,
1806 const struct intel_plane_state *pstate,
1807 uint32_t pri_val)
1808 {
1809 int bpp = pstate->base.fb ? pstate->base.fb->bits_per_pixel / 8 : 0;
1810
1811 if (!cstate->base.active || !pstate->visible)
1812 return 0;
1813
1814 return ilk_wm_fbc(pri_val, drm_rect_width(&pstate->dst), bpp);
1815 }
1816
1817 static unsigned int ilk_display_fifo_size(const struct drm_device *dev)
1818 {
1819 if (INTEL_INFO(dev)->gen >= 8)
1820 return 3072;
1821 else if (INTEL_INFO(dev)->gen >= 7)
1822 return 768;
1823 else
1824 return 512;
1825 }
1826
1827 static unsigned int ilk_plane_wm_reg_max(const struct drm_device *dev,
1828 int level, bool is_sprite)
1829 {
1830 if (INTEL_INFO(dev)->gen >= 8)
1831 /* BDW primary/sprite plane watermarks */
1832 return level == 0 ? 255 : 2047;
1833 else if (INTEL_INFO(dev)->gen >= 7)
1834 /* IVB/HSW primary/sprite plane watermarks */
1835 return level == 0 ? 127 : 1023;
1836 else if (!is_sprite)
1837 /* ILK/SNB primary plane watermarks */
1838 return level == 0 ? 127 : 511;
1839 else
1840 /* ILK/SNB sprite plane watermarks */
1841 return level == 0 ? 63 : 255;
1842 }
1843
1844 static unsigned int ilk_cursor_wm_reg_max(const struct drm_device *dev,
1845 int level)
1846 {
1847 if (INTEL_INFO(dev)->gen >= 7)
1848 return level == 0 ? 63 : 255;
1849 else
1850 return level == 0 ? 31 : 63;
1851 }
1852
1853 static unsigned int ilk_fbc_wm_reg_max(const struct drm_device *dev)
1854 {
1855 if (INTEL_INFO(dev)->gen >= 8)
1856 return 31;
1857 else
1858 return 15;
1859 }
1860
1861 /* Calculate the maximum primary/sprite plane watermark */
1862 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
1863 int level,
1864 const struct intel_wm_config *config,
1865 enum intel_ddb_partitioning ddb_partitioning,
1866 bool is_sprite)
1867 {
1868 unsigned int fifo_size = ilk_display_fifo_size(dev);
1869
1870 /* if sprites aren't enabled, sprites get nothing */
1871 if (is_sprite && !config->sprites_enabled)
1872 return 0;
1873
1874 /* HSW allows LP1+ watermarks even with multiple pipes */
1875 if (level == 0 || config->num_pipes_active > 1) {
1876 fifo_size /= INTEL_INFO(dev)->num_pipes;
1877
1878 /*
1879 * For some reason the non self refresh
1880 * FIFO size is only half of the self
1881 * refresh FIFO size on ILK/SNB.
1882 */
1883 if (INTEL_INFO(dev)->gen <= 6)
1884 fifo_size /= 2;
1885 }
1886
1887 if (config->sprites_enabled) {
1888 /* level 0 is always calculated with 1:1 split */
1889 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
1890 if (is_sprite)
1891 fifo_size *= 5;
1892 fifo_size /= 6;
1893 } else {
1894 fifo_size /= 2;
1895 }
1896 }
1897
1898 /* clamp to max that the registers can hold */
1899 return min(fifo_size, ilk_plane_wm_reg_max(dev, level, is_sprite));
1900 }
1901
1902 /* Calculate the maximum cursor plane watermark */
1903 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
1904 int level,
1905 const struct intel_wm_config *config)
1906 {
1907 /* HSW LP1+ watermarks w/ multiple pipes */
1908 if (level > 0 && config->num_pipes_active > 1)
1909 return 64;
1910
1911 /* otherwise just report max that registers can hold */
1912 return ilk_cursor_wm_reg_max(dev, level);
1913 }
1914
1915 static void ilk_compute_wm_maximums(const struct drm_device *dev,
1916 int level,
1917 const struct intel_wm_config *config,
1918 enum intel_ddb_partitioning ddb_partitioning,
1919 struct ilk_wm_maximums *max)
1920 {
1921 max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
1922 max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
1923 max->cur = ilk_cursor_wm_max(dev, level, config);
1924 max->fbc = ilk_fbc_wm_reg_max(dev);
1925 }
1926
1927 static void ilk_compute_wm_reg_maximums(struct drm_device *dev,
1928 int level,
1929 struct ilk_wm_maximums *max)
1930 {
1931 max->pri = ilk_plane_wm_reg_max(dev, level, false);
1932 max->spr = ilk_plane_wm_reg_max(dev, level, true);
1933 max->cur = ilk_cursor_wm_reg_max(dev, level);
1934 max->fbc = ilk_fbc_wm_reg_max(dev);
1935 }
1936
1937 static bool ilk_validate_wm_level(int level,
1938 const struct ilk_wm_maximums *max,
1939 struct intel_wm_level *result)
1940 {
1941 bool ret;
1942
1943 /* already determined to be invalid? */
1944 if (!result->enable)
1945 return false;
1946
1947 result->enable = result->pri_val <= max->pri &&
1948 result->spr_val <= max->spr &&
1949 result->cur_val <= max->cur;
1950
1951 ret = result->enable;
1952
1953 /*
1954 * HACK until we can pre-compute everything,
1955 * and thus fail gracefully if LP0 watermarks
1956 * are exceeded...
1957 */
1958 if (level == 0 && !result->enable) {
1959 if (result->pri_val > max->pri)
1960 DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
1961 level, result->pri_val, max->pri);
1962 if (result->spr_val > max->spr)
1963 DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
1964 level, result->spr_val, max->spr);
1965 if (result->cur_val > max->cur)
1966 DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
1967 level, result->cur_val, max->cur);
1968
1969 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
1970 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
1971 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
1972 result->enable = true;
1973 }
1974
1975 return ret;
1976 }
1977
1978 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
1979 const struct intel_crtc *intel_crtc,
1980 int level,
1981 struct intel_crtc_state *cstate,
1982 struct intel_wm_level *result)
1983 {
1984 struct intel_plane *intel_plane;
1985 uint16_t pri_latency = dev_priv->wm.pri_latency[level];
1986 uint16_t spr_latency = dev_priv->wm.spr_latency[level];
1987 uint16_t cur_latency = dev_priv->wm.cur_latency[level];
1988
1989 /* WM1+ latency values stored in 0.5us units */
1990 if (level > 0) {
1991 pri_latency *= 5;
1992 spr_latency *= 5;
1993 cur_latency *= 5;
1994 }
1995
1996 for_each_intel_plane_on_crtc(dev_priv->dev, intel_crtc, intel_plane) {
1997 struct intel_plane_state *pstate =
1998 to_intel_plane_state(intel_plane->base.state);
1999
2000 switch (intel_plane->base.type) {
2001 case DRM_PLANE_TYPE_PRIMARY:
2002 result->pri_val = ilk_compute_pri_wm(cstate, pstate,
2003 pri_latency,
2004 level);
2005 result->fbc_val = ilk_compute_fbc_wm(cstate, pstate,
2006 result->pri_val);
2007 break;
2008 case DRM_PLANE_TYPE_OVERLAY:
2009 result->spr_val = ilk_compute_spr_wm(cstate, pstate,
2010 spr_latency);
2011 break;
2012 case DRM_PLANE_TYPE_CURSOR:
2013 result->cur_val = ilk_compute_cur_wm(cstate, pstate,
2014 cur_latency);
2015 break;
2016 }
2017 }
2018
2019 result->enable = true;
2020 }
2021
2022 static uint32_t
2023 hsw_compute_linetime_wm(struct drm_device *dev, struct drm_crtc *crtc)
2024 {
2025 struct drm_i915_private *dev_priv = dev->dev_private;
2026 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2027 const struct drm_display_mode *adjusted_mode = &intel_crtc->config->base.adjusted_mode;
2028 u32 linetime, ips_linetime;
2029
2030 if (!intel_crtc->active)
2031 return 0;
2032
2033 /* The WM are computed with base on how long it takes to fill a single
2034 * row at the given clock rate, multiplied by 8.
2035 * */
2036 linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2037 adjusted_mode->crtc_clock);
2038 ips_linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2039 dev_priv->cdclk_freq);
2040
2041 return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2042 PIPE_WM_LINETIME_TIME(linetime);
2043 }
2044
2045 static void intel_read_wm_latency(struct drm_device *dev, uint16_t wm[8])
2046 {
2047 struct drm_i915_private *dev_priv = dev->dev_private;
2048
2049 if (IS_GEN9(dev)) {
2050 uint32_t val;
2051 int ret, i;
2052 int level, max_level = ilk_wm_max_level(dev);
2053
2054 /* read the first set of memory latencies[0:3] */
2055 val = 0; /* data0 to be programmed to 0 for first set */
2056 mutex_lock(&dev_priv->rps.hw_lock);
2057 ret = sandybridge_pcode_read(dev_priv,
2058 GEN9_PCODE_READ_MEM_LATENCY,
2059 &val);
2060 mutex_unlock(&dev_priv->rps.hw_lock);
2061
2062 if (ret) {
2063 DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2064 return;
2065 }
2066
2067 wm[0] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2068 wm[1] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2069 GEN9_MEM_LATENCY_LEVEL_MASK;
2070 wm[2] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2071 GEN9_MEM_LATENCY_LEVEL_MASK;
2072 wm[3] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2073 GEN9_MEM_LATENCY_LEVEL_MASK;
2074
2075 /* read the second set of memory latencies[4:7] */
2076 val = 1; /* data0 to be programmed to 1 for second set */
2077 mutex_lock(&dev_priv->rps.hw_lock);
2078 ret = sandybridge_pcode_read(dev_priv,
2079 GEN9_PCODE_READ_MEM_LATENCY,
2080 &val);
2081 mutex_unlock(&dev_priv->rps.hw_lock);
2082 if (ret) {
2083 DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2084 return;
2085 }
2086
2087 wm[4] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2088 wm[5] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2089 GEN9_MEM_LATENCY_LEVEL_MASK;
2090 wm[6] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2091 GEN9_MEM_LATENCY_LEVEL_MASK;
2092 wm[7] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2093 GEN9_MEM_LATENCY_LEVEL_MASK;
2094
2095 /*
2096 * WaWmMemoryReadLatency:skl
2097 *
2098 * punit doesn't take into account the read latency so we need
2099 * to add 2us to the various latency levels we retrieve from
2100 * the punit.
2101 * - W0 is a bit special in that it's the only level that
2102 * can't be disabled if we want to have display working, so
2103 * we always add 2us there.
2104 * - For levels >=1, punit returns 0us latency when they are
2105 * disabled, so we respect that and don't add 2us then
2106 *
2107 * Additionally, if a level n (n > 1) has a 0us latency, all
2108 * levels m (m >= n) need to be disabled. We make sure to
2109 * sanitize the values out of the punit to satisfy this
2110 * requirement.
2111 */
2112 wm[0] += 2;
2113 for (level = 1; level <= max_level; level++)
2114 if (wm[level] != 0)
2115 wm[level] += 2;
2116 else {
2117 for (i = level + 1; i <= max_level; i++)
2118 wm[i] = 0;
2119
2120 break;
2121 }
2122 } else if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2123 uint64_t sskpd = I915_READ64(MCH_SSKPD);
2124
2125 wm[0] = (sskpd >> 56) & 0xFF;
2126 if (wm[0] == 0)
2127 wm[0] = sskpd & 0xF;
2128 wm[1] = (sskpd >> 4) & 0xFF;
2129 wm[2] = (sskpd >> 12) & 0xFF;
2130 wm[3] = (sskpd >> 20) & 0x1FF;
2131 wm[4] = (sskpd >> 32) & 0x1FF;
2132 } else if (INTEL_INFO(dev)->gen >= 6) {
2133 uint32_t sskpd = I915_READ(MCH_SSKPD);
2134
2135 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2136 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2137 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2138 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2139 } else if (INTEL_INFO(dev)->gen >= 5) {
2140 uint32_t mltr = I915_READ(MLTR_ILK);
2141
2142 /* ILK primary LP0 latency is 700 ns */
2143 wm[0] = 7;
2144 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2145 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2146 }
2147 }
2148
2149 static void intel_fixup_spr_wm_latency(struct drm_device *dev, uint16_t wm[5])
2150 {
2151 /* ILK sprite LP0 latency is 1300 ns */
2152 if (INTEL_INFO(dev)->gen == 5)
2153 wm[0] = 13;
2154 }
2155
2156 static void intel_fixup_cur_wm_latency(struct drm_device *dev, uint16_t wm[5])
2157 {
2158 /* ILK cursor LP0 latency is 1300 ns */
2159 if (INTEL_INFO(dev)->gen == 5)
2160 wm[0] = 13;
2161
2162 /* WaDoubleCursorLP3Latency:ivb */
2163 if (IS_IVYBRIDGE(dev))
2164 wm[3] *= 2;
2165 }
2166
2167 int ilk_wm_max_level(const struct drm_device *dev)
2168 {
2169 /* how many WM levels are we expecting */
2170 if (INTEL_INFO(dev)->gen >= 9)
2171 return 7;
2172 else if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2173 return 4;
2174 else if (INTEL_INFO(dev)->gen >= 6)
2175 return 3;
2176 else
2177 return 2;
2178 }
2179
2180 static void intel_print_wm_latency(struct drm_device *dev,
2181 const char *name,
2182 const uint16_t wm[8])
2183 {
2184 int level, max_level = ilk_wm_max_level(dev);
2185
2186 for (level = 0; level <= max_level; level++) {
2187 unsigned int latency = wm[level];
2188
2189 if (latency == 0) {
2190 DRM_ERROR("%s WM%d latency not provided\n",
2191 name, level);
2192 continue;
2193 }
2194
2195 /*
2196 * - latencies are in us on gen9.
2197 * - before then, WM1+ latency values are in 0.5us units
2198 */
2199 if (IS_GEN9(dev))
2200 latency *= 10;
2201 else if (level > 0)
2202 latency *= 5;
2203
2204 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2205 name, level, wm[level],
2206 latency / 10, latency % 10);
2207 }
2208 }
2209
2210 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
2211 uint16_t wm[5], uint16_t min)
2212 {
2213 int level, max_level = ilk_wm_max_level(dev_priv->dev);
2214
2215 if (wm[0] >= min)
2216 return false;
2217
2218 wm[0] = max(wm[0], min);
2219 for (level = 1; level <= max_level; level++)
2220 wm[level] = max_t(uint16_t, wm[level], DIV_ROUND_UP(min, 5));
2221
2222 return true;
2223 }
2224
2225 static void snb_wm_latency_quirk(struct drm_device *dev)
2226 {
2227 struct drm_i915_private *dev_priv = dev->dev_private;
2228 bool changed;
2229
2230 /*
2231 * The BIOS provided WM memory latency values are often
2232 * inadequate for high resolution displays. Adjust them.
2233 */
2234 changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
2235 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
2236 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
2237
2238 if (!changed)
2239 return;
2240
2241 DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2242 intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2243 intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2244 intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2245 }
2246
2247 static void ilk_setup_wm_latency(struct drm_device *dev)
2248 {
2249 struct drm_i915_private *dev_priv = dev->dev_private;
2250
2251 intel_read_wm_latency(dev, dev_priv->wm.pri_latency);
2252
2253 memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2254 sizeof(dev_priv->wm.pri_latency));
2255 memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2256 sizeof(dev_priv->wm.pri_latency));
2257
2258 intel_fixup_spr_wm_latency(dev, dev_priv->wm.spr_latency);
2259 intel_fixup_cur_wm_latency(dev, dev_priv->wm.cur_latency);
2260
2261 intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2262 intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2263 intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2264
2265 if (IS_GEN6(dev))
2266 snb_wm_latency_quirk(dev);
2267 }
2268
2269 static void skl_setup_wm_latency(struct drm_device *dev)
2270 {
2271 struct drm_i915_private *dev_priv = dev->dev_private;
2272
2273 intel_read_wm_latency(dev, dev_priv->wm.skl_latency);
2274 intel_print_wm_latency(dev, "Gen9 Plane", dev_priv->wm.skl_latency);
2275 }
2276
2277 static void ilk_compute_wm_config(struct drm_device *dev,
2278 struct intel_wm_config *config)
2279 {
2280 struct intel_crtc *intel_crtc;
2281
2282 /* Compute the currently _active_ config */
2283 for_each_intel_crtc(dev, intel_crtc) {
2284 const struct intel_pipe_wm *wm = &intel_crtc->wm.active;
2285
2286 if (!wm->pipe_enabled)
2287 continue;
2288
2289 config->sprites_enabled |= wm->sprites_enabled;
2290 config->sprites_scaled |= wm->sprites_scaled;
2291 config->num_pipes_active++;
2292 }
2293 }
2294
2295 /* Compute new watermarks for the pipe */
2296 static bool intel_compute_pipe_wm(struct intel_crtc_state *cstate,
2297 struct intel_pipe_wm *pipe_wm)
2298 {
2299 struct drm_crtc *crtc = cstate->base.crtc;
2300 struct drm_device *dev = crtc->dev;
2301 const struct drm_i915_private *dev_priv = dev->dev_private;
2302 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2303 struct intel_plane *intel_plane;
2304 struct intel_plane_state *sprstate = NULL;
2305 int level, max_level = ilk_wm_max_level(dev);
2306 /* LP0 watermark maximums depend on this pipe alone */
2307 struct intel_wm_config config = {
2308 .num_pipes_active = 1,
2309 };
2310 struct ilk_wm_maximums max;
2311
2312 for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
2313 if (intel_plane->base.type == DRM_PLANE_TYPE_OVERLAY) {
2314 sprstate = to_intel_plane_state(intel_plane->base.state);
2315 break;
2316 }
2317 }
2318
2319 config.sprites_enabled = sprstate->visible;
2320 config.sprites_scaled = sprstate->visible &&
2321 (drm_rect_width(&sprstate->dst) != drm_rect_width(&sprstate->src) >> 16 ||
2322 drm_rect_height(&sprstate->dst) != drm_rect_height(&sprstate->src) >> 16);
2323
2324 pipe_wm->pipe_enabled = cstate->base.active;
2325 pipe_wm->sprites_enabled = sprstate->visible;
2326 pipe_wm->sprites_scaled = config.sprites_scaled;
2327
2328 /* ILK/SNB: LP2+ watermarks only w/o sprites */
2329 if (INTEL_INFO(dev)->gen <= 6 && sprstate->visible)
2330 max_level = 1;
2331
2332 /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2333 if (config.sprites_scaled)
2334 max_level = 0;
2335
2336 ilk_compute_wm_level(dev_priv, intel_crtc, 0, cstate, &pipe_wm->wm[0]);
2337
2338 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2339 pipe_wm->linetime = hsw_compute_linetime_wm(dev, crtc);
2340
2341 /* LP0 watermarks always use 1/2 DDB partitioning */
2342 ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
2343
2344 /* At least LP0 must be valid */
2345 if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0]))
2346 return false;
2347
2348 ilk_compute_wm_reg_maximums(dev, 1, &max);
2349
2350 for (level = 1; level <= max_level; level++) {
2351 struct intel_wm_level wm = {};
2352
2353 ilk_compute_wm_level(dev_priv, intel_crtc, level, cstate, &wm);
2354
2355 /*
2356 * Disable any watermark level that exceeds the
2357 * register maximums since such watermarks are
2358 * always invalid.
2359 */
2360 if (!ilk_validate_wm_level(level, &max, &wm))
2361 break;
2362
2363 pipe_wm->wm[level] = wm;
2364 }
2365
2366 return true;
2367 }
2368
2369 /*
2370 * Merge the watermarks from all active pipes for a specific level.
2371 */
2372 static void ilk_merge_wm_level(struct drm_device *dev,
2373 int level,
2374 struct intel_wm_level *ret_wm)
2375 {
2376 const struct intel_crtc *intel_crtc;
2377
2378 ret_wm->enable = true;
2379
2380 for_each_intel_crtc(dev, intel_crtc) {
2381 const struct intel_pipe_wm *active = &intel_crtc->wm.active;
2382 const struct intel_wm_level *wm = &active->wm[level];
2383
2384 if (!active->pipe_enabled)
2385 continue;
2386
2387 /*
2388 * The watermark values may have been used in the past,
2389 * so we must maintain them in the registers for some
2390 * time even if the level is now disabled.
2391 */
2392 if (!wm->enable)
2393 ret_wm->enable = false;
2394
2395 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
2396 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
2397 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
2398 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
2399 }
2400 }
2401
2402 /*
2403 * Merge all low power watermarks for all active pipes.
2404 */
2405 static void ilk_wm_merge(struct drm_device *dev,
2406 const struct intel_wm_config *config,
2407 const struct ilk_wm_maximums *max,
2408 struct intel_pipe_wm *merged)
2409 {
2410 struct drm_i915_private *dev_priv = dev->dev_private;
2411 int level, max_level = ilk_wm_max_level(dev);
2412 int last_enabled_level = max_level;
2413
2414 /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2415 if ((INTEL_INFO(dev)->gen <= 6 || IS_IVYBRIDGE(dev)) &&
2416 config->num_pipes_active > 1)
2417 return;
2418
2419 /* ILK: FBC WM must be disabled always */
2420 merged->fbc_wm_enabled = INTEL_INFO(dev)->gen >= 6;
2421
2422 /* merge each WM1+ level */
2423 for (level = 1; level <= max_level; level++) {
2424 struct intel_wm_level *wm = &merged->wm[level];
2425
2426 ilk_merge_wm_level(dev, level, wm);
2427
2428 if (level > last_enabled_level)
2429 wm->enable = false;
2430 else if (!ilk_validate_wm_level(level, max, wm))
2431 /* make sure all following levels get disabled */
2432 last_enabled_level = level - 1;
2433
2434 /*
2435 * The spec says it is preferred to disable
2436 * FBC WMs instead of disabling a WM level.
2437 */
2438 if (wm->fbc_val > max->fbc) {
2439 if (wm->enable)
2440 merged->fbc_wm_enabled = false;
2441 wm->fbc_val = 0;
2442 }
2443 }
2444
2445 /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2446 /*
2447 * FIXME this is racy. FBC might get enabled later.
2448 * What we should check here is whether FBC can be
2449 * enabled sometime later.
2450 */
2451 if (IS_GEN5(dev) && !merged->fbc_wm_enabled &&
2452 intel_fbc_enabled(dev_priv)) {
2453 for (level = 2; level <= max_level; level++) {
2454 struct intel_wm_level *wm = &merged->wm[level];
2455
2456 wm->enable = false;
2457 }
2458 }
2459 }
2460
2461 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
2462 {
2463 /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2464 return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
2465 }
2466
2467 /* The value we need to program into the WM_LPx latency field */
2468 static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
2469 {
2470 struct drm_i915_private *dev_priv = dev->dev_private;
2471
2472 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2473 return 2 * level;
2474 else
2475 return dev_priv->wm.pri_latency[level];
2476 }
2477
2478 static void ilk_compute_wm_results(struct drm_device *dev,
2479 const struct intel_pipe_wm *merged,
2480 enum intel_ddb_partitioning partitioning,
2481 struct ilk_wm_values *results)
2482 {
2483 struct intel_crtc *intel_crtc;
2484 int level, wm_lp;
2485
2486 results->enable_fbc_wm = merged->fbc_wm_enabled;
2487 results->partitioning = partitioning;
2488
2489 /* LP1+ register values */
2490 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2491 const struct intel_wm_level *r;
2492
2493 level = ilk_wm_lp_to_level(wm_lp, merged);
2494
2495 r = &merged->wm[level];
2496
2497 /*
2498 * Maintain the watermark values even if the level is
2499 * disabled. Doing otherwise could cause underruns.
2500 */
2501 results->wm_lp[wm_lp - 1] =
2502 (ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
2503 (r->pri_val << WM1_LP_SR_SHIFT) |
2504 r->cur_val;
2505
2506 if (r->enable)
2507 results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
2508
2509 if (INTEL_INFO(dev)->gen >= 8)
2510 results->wm_lp[wm_lp - 1] |=
2511 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
2512 else
2513 results->wm_lp[wm_lp - 1] |=
2514 r->fbc_val << WM1_LP_FBC_SHIFT;
2515
2516 /*
2517 * Always set WM1S_LP_EN when spr_val != 0, even if the
2518 * level is disabled. Doing otherwise could cause underruns.
2519 */
2520 if (INTEL_INFO(dev)->gen <= 6 && r->spr_val) {
2521 WARN_ON(wm_lp != 1);
2522 results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
2523 } else
2524 results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2525 }
2526
2527 /* LP0 register values */
2528 for_each_intel_crtc(dev, intel_crtc) {
2529 enum pipe pipe = intel_crtc->pipe;
2530 const struct intel_wm_level *r =
2531 &intel_crtc->wm.active.wm[0];
2532
2533 if (WARN_ON(!r->enable))
2534 continue;
2535
2536 results->wm_linetime[pipe] = intel_crtc->wm.active.linetime;
2537
2538 results->wm_pipe[pipe] =
2539 (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
2540 (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
2541 r->cur_val;
2542 }
2543 }
2544
2545 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2546 * case both are at the same level. Prefer r1 in case they're the same. */
2547 static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
2548 struct intel_pipe_wm *r1,
2549 struct intel_pipe_wm *r2)
2550 {
2551 int level, max_level = ilk_wm_max_level(dev);
2552 int level1 = 0, level2 = 0;
2553
2554 for (level = 1; level <= max_level; level++) {
2555 if (r1->wm[level].enable)
2556 level1 = level;
2557 if (r2->wm[level].enable)
2558 level2 = level;
2559 }
2560
2561 if (level1 == level2) {
2562 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
2563 return r2;
2564 else
2565 return r1;
2566 } else if (level1 > level2) {
2567 return r1;
2568 } else {
2569 return r2;
2570 }
2571 }
2572
2573 /* dirty bits used to track which watermarks need changes */
2574 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2575 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2576 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2577 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2578 #define WM_DIRTY_FBC (1 << 24)
2579 #define WM_DIRTY_DDB (1 << 25)
2580
2581 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private *dev_priv,
2582 const struct ilk_wm_values *old,
2583 const struct ilk_wm_values *new)
2584 {
2585 unsigned int dirty = 0;
2586 enum pipe pipe;
2587 int wm_lp;
2588
2589 for_each_pipe(dev_priv, pipe) {
2590 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
2591 dirty |= WM_DIRTY_LINETIME(pipe);
2592 /* Must disable LP1+ watermarks too */
2593 dirty |= WM_DIRTY_LP_ALL;
2594 }
2595
2596 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
2597 dirty |= WM_DIRTY_PIPE(pipe);
2598 /* Must disable LP1+ watermarks too */
2599 dirty |= WM_DIRTY_LP_ALL;
2600 }
2601 }
2602
2603 if (old->enable_fbc_wm != new->enable_fbc_wm) {
2604 dirty |= WM_DIRTY_FBC;
2605 /* Must disable LP1+ watermarks too */
2606 dirty |= WM_DIRTY_LP_ALL;
2607 }
2608
2609 if (old->partitioning != new->partitioning) {
2610 dirty |= WM_DIRTY_DDB;
2611 /* Must disable LP1+ watermarks too */
2612 dirty |= WM_DIRTY_LP_ALL;
2613 }
2614
2615 /* LP1+ watermarks already deemed dirty, no need to continue */
2616 if (dirty & WM_DIRTY_LP_ALL)
2617 return dirty;
2618
2619 /* Find the lowest numbered LP1+ watermark in need of an update... */
2620 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2621 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
2622 old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
2623 break;
2624 }
2625
2626 /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2627 for (; wm_lp <= 3; wm_lp++)
2628 dirty |= WM_DIRTY_LP(wm_lp);
2629
2630 return dirty;
2631 }
2632
2633 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
2634 unsigned int dirty)
2635 {
2636 struct ilk_wm_values *previous = &dev_priv->wm.hw;
2637 bool changed = false;
2638
2639 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
2640 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
2641 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
2642 changed = true;
2643 }
2644 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
2645 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
2646 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
2647 changed = true;
2648 }
2649 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
2650 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
2651 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
2652 changed = true;
2653 }
2654
2655 /*
2656 * Don't touch WM1S_LP_EN here.
2657 * Doing so could cause underruns.
2658 */
2659
2660 return changed;
2661 }
2662
2663 /*
2664 * The spec says we shouldn't write when we don't need, because every write
2665 * causes WMs to be re-evaluated, expending some power.
2666 */
2667 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
2668 struct ilk_wm_values *results)
2669 {
2670 struct drm_device *dev = dev_priv->dev;
2671 struct ilk_wm_values *previous = &dev_priv->wm.hw;
2672 unsigned int dirty;
2673 uint32_t val;
2674
2675 dirty = ilk_compute_wm_dirty(dev_priv, previous, results);
2676 if (!dirty)
2677 return;
2678
2679 _ilk_disable_lp_wm(dev_priv, dirty);
2680
2681 if (dirty & WM_DIRTY_PIPE(PIPE_A))
2682 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2683 if (dirty & WM_DIRTY_PIPE(PIPE_B))
2684 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
2685 if (dirty & WM_DIRTY_PIPE(PIPE_C))
2686 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
2687
2688 if (dirty & WM_DIRTY_LINETIME(PIPE_A))
2689 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
2690 if (dirty & WM_DIRTY_LINETIME(PIPE_B))
2691 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
2692 if (dirty & WM_DIRTY_LINETIME(PIPE_C))
2693 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
2694
2695 if (dirty & WM_DIRTY_DDB) {
2696 if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2697 val = I915_READ(WM_MISC);
2698 if (results->partitioning == INTEL_DDB_PART_1_2)
2699 val &= ~WM_MISC_DATA_PARTITION_5_6;
2700 else
2701 val |= WM_MISC_DATA_PARTITION_5_6;
2702 I915_WRITE(WM_MISC, val);
2703 } else {
2704 val = I915_READ(DISP_ARB_CTL2);
2705 if (results->partitioning == INTEL_DDB_PART_1_2)
2706 val &= ~DISP_DATA_PARTITION_5_6;
2707 else
2708 val |= DISP_DATA_PARTITION_5_6;
2709 I915_WRITE(DISP_ARB_CTL2, val);
2710 }
2711 }
2712
2713 if (dirty & WM_DIRTY_FBC) {
2714 val = I915_READ(DISP_ARB_CTL);
2715 if (results->enable_fbc_wm)
2716 val &= ~DISP_FBC_WM_DIS;
2717 else
2718 val |= DISP_FBC_WM_DIS;
2719 I915_WRITE(DISP_ARB_CTL, val);
2720 }
2721
2722 if (dirty & WM_DIRTY_LP(1) &&
2723 previous->wm_lp_spr[0] != results->wm_lp_spr[0])
2724 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
2725
2726 if (INTEL_INFO(dev)->gen >= 7) {
2727 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
2728 I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
2729 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
2730 I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
2731 }
2732
2733 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
2734 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
2735 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
2736 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
2737 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
2738 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
2739
2740 dev_priv->wm.hw = *results;
2741 }
2742
2743 static bool ilk_disable_lp_wm(struct drm_device *dev)
2744 {
2745 struct drm_i915_private *dev_priv = dev->dev_private;
2746
2747 return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
2748 }
2749
2750 /*
2751 * On gen9, we need to allocate Display Data Buffer (DDB) portions to the
2752 * different active planes.
2753 */
2754
2755 #define SKL_DDB_SIZE 896 /* in blocks */
2756 #define BXT_DDB_SIZE 512
2757
2758 static void
2759 skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
2760 struct drm_crtc *for_crtc,
2761 const struct intel_wm_config *config,
2762 const struct skl_pipe_wm_parameters *params,
2763 struct skl_ddb_entry *alloc /* out */)
2764 {
2765 struct drm_crtc *crtc;
2766 unsigned int pipe_size, ddb_size;
2767 int nth_active_pipe;
2768
2769 if (!params->active) {
2770 alloc->start = 0;
2771 alloc->end = 0;
2772 return;
2773 }
2774
2775 if (IS_BROXTON(dev))
2776 ddb_size = BXT_DDB_SIZE;
2777 else
2778 ddb_size = SKL_DDB_SIZE;
2779
2780 ddb_size -= 4; /* 4 blocks for bypass path allocation */
2781
2782 nth_active_pipe = 0;
2783 for_each_crtc(dev, crtc) {
2784 if (!to_intel_crtc(crtc)->active)
2785 continue;
2786
2787 if (crtc == for_crtc)
2788 break;
2789
2790 nth_active_pipe++;
2791 }
2792
2793 pipe_size = ddb_size / config->num_pipes_active;
2794 alloc->start = nth_active_pipe * ddb_size / config->num_pipes_active;
2795 alloc->end = alloc->start + pipe_size;
2796 }
2797
2798 static unsigned int skl_cursor_allocation(const struct intel_wm_config *config)
2799 {
2800 if (config->num_pipes_active == 1)
2801 return 32;
2802
2803 return 8;
2804 }
2805
2806 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry *entry, u32 reg)
2807 {
2808 entry->start = reg & 0x3ff;
2809 entry->end = (reg >> 16) & 0x3ff;
2810 if (entry->end)
2811 entry->end += 1;
2812 }
2813
2814 void skl_ddb_get_hw_state(struct drm_i915_private *dev_priv,
2815 struct skl_ddb_allocation *ddb /* out */)
2816 {
2817 enum pipe pipe;
2818 int plane;
2819 u32 val;
2820
2821 memset(ddb, 0, sizeof(*ddb));
2822
2823 for_each_pipe(dev_priv, pipe) {
2824 if (!intel_display_power_is_enabled(dev_priv, POWER_DOMAIN_PIPE(pipe)))
2825 continue;
2826
2827 for_each_plane(dev_priv, pipe, plane) {
2828 val = I915_READ(PLANE_BUF_CFG(pipe, plane));
2829 skl_ddb_entry_init_from_hw(&ddb->plane[pipe][plane],
2830 val);
2831 }
2832
2833 val = I915_READ(CUR_BUF_CFG(pipe));
2834 skl_ddb_entry_init_from_hw(&ddb->plane[pipe][PLANE_CURSOR],
2835 val);
2836 }
2837 }
2838
2839 static unsigned int
2840 skl_plane_relative_data_rate(const struct intel_plane_wm_parameters *p, int y)
2841 {
2842
2843 /* for planar format */
2844 if (p->y_bytes_per_pixel) {
2845 if (y) /* y-plane data rate */
2846 return p->horiz_pixels * p->vert_pixels * p->y_bytes_per_pixel;
2847 else /* uv-plane data rate */
2848 return (p->horiz_pixels/2) * (p->vert_pixels/2) * p->bytes_per_pixel;
2849 }
2850
2851 /* for packed formats */
2852 return p->horiz_pixels * p->vert_pixels * p->bytes_per_pixel;
2853 }
2854
2855 /*
2856 * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
2857 * a 8192x4096@32bpp framebuffer:
2858 * 3 * 4096 * 8192 * 4 < 2^32
2859 */
2860 static unsigned int
2861 skl_get_total_relative_data_rate(struct intel_crtc *intel_crtc,
2862 const struct skl_pipe_wm_parameters *params)
2863 {
2864 unsigned int total_data_rate = 0;
2865 int plane;
2866
2867 for (plane = 0; plane < intel_num_planes(intel_crtc); plane++) {
2868 const struct intel_plane_wm_parameters *p;
2869
2870 p = &params->plane[plane];
2871 if (!p->enabled)
2872 continue;
2873
2874 total_data_rate += skl_plane_relative_data_rate(p, 0); /* packed/uv */
2875 if (p->y_bytes_per_pixel) {
2876 total_data_rate += skl_plane_relative_data_rate(p, 1); /* y-plane */
2877 }
2878 }
2879
2880 return total_data_rate;
2881 }
2882
2883 static void
2884 skl_allocate_pipe_ddb(struct drm_crtc *crtc,
2885 const struct intel_wm_config *config,
2886 const struct skl_pipe_wm_parameters *params,
2887 struct skl_ddb_allocation *ddb /* out */)
2888 {
2889 struct drm_device *dev = crtc->dev;
2890 struct drm_i915_private *dev_priv = dev->dev_private;
2891 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2892 enum pipe pipe = intel_crtc->pipe;
2893 struct skl_ddb_entry *alloc = &ddb->pipe[pipe];
2894 uint16_t alloc_size, start, cursor_blocks;
2895 uint16_t minimum[I915_MAX_PLANES];
2896 uint16_t y_minimum[I915_MAX_PLANES];
2897 unsigned int total_data_rate;
2898 int plane;
2899
2900 skl_ddb_get_pipe_allocation_limits(dev, crtc, config, params, alloc);
2901 alloc_size = skl_ddb_entry_size(alloc);
2902 if (alloc_size == 0) {
2903 memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
2904 memset(&ddb->plane[pipe][PLANE_CURSOR], 0,
2905 sizeof(ddb->plane[pipe][PLANE_CURSOR]));
2906 return;
2907 }
2908
2909 cursor_blocks = skl_cursor_allocation(config);
2910 ddb->plane[pipe][PLANE_CURSOR].start = alloc->end - cursor_blocks;
2911 ddb->plane[pipe][PLANE_CURSOR].end = alloc->end;
2912
2913 alloc_size -= cursor_blocks;
2914 alloc->end -= cursor_blocks;
2915
2916 /* 1. Allocate the mininum required blocks for each active plane */
2917 for_each_plane(dev_priv, pipe, plane) {
2918 const struct intel_plane_wm_parameters *p;
2919
2920 p = &params->plane[plane];
2921 if (!p->enabled)
2922 continue;
2923
2924 minimum[plane] = 8;
2925 alloc_size -= minimum[plane];
2926 y_minimum[plane] = p->y_bytes_per_pixel ? 8 : 0;
2927 alloc_size -= y_minimum[plane];
2928 }
2929
2930 /*
2931 * 2. Distribute the remaining space in proportion to the amount of
2932 * data each plane needs to fetch from memory.
2933 *
2934 * FIXME: we may not allocate every single block here.
2935 */
2936 total_data_rate = skl_get_total_relative_data_rate(intel_crtc, params);
2937
2938 start = alloc->start;
2939 for (plane = 0; plane < intel_num_planes(intel_crtc); plane++) {
2940 const struct intel_plane_wm_parameters *p;
2941 unsigned int data_rate, y_data_rate;
2942 uint16_t plane_blocks, y_plane_blocks = 0;
2943
2944 p = &params->plane[plane];
2945 if (!p->enabled)
2946 continue;
2947
2948 data_rate = skl_plane_relative_data_rate(p, 0);
2949
2950 /*
2951 * allocation for (packed formats) or (uv-plane part of planar format):
2952 * promote the expression to 64 bits to avoid overflowing, the
2953 * result is < available as data_rate / total_data_rate < 1
2954 */
2955 plane_blocks = minimum[plane];
2956 plane_blocks += div_u64((uint64_t)alloc_size * data_rate,
2957 total_data_rate);
2958
2959 ddb->plane[pipe][plane].start = start;
2960 ddb->plane[pipe][plane].end = start + plane_blocks;
2961
2962 start += plane_blocks;
2963
2964 /*
2965 * allocation for y_plane part of planar format:
2966 */
2967 if (p->y_bytes_per_pixel) {
2968 y_data_rate = skl_plane_relative_data_rate(p, 1);
2969 y_plane_blocks = y_minimum[plane];
2970 y_plane_blocks += div_u64((uint64_t)alloc_size * y_data_rate,
2971 total_data_rate);
2972
2973 ddb->y_plane[pipe][plane].start = start;
2974 ddb->y_plane[pipe][plane].end = start + y_plane_blocks;
2975
2976 start += y_plane_blocks;
2977 }
2978
2979 }
2980
2981 }
2982
2983 static uint32_t skl_pipe_pixel_rate(const struct intel_crtc_state *config)
2984 {
2985 /* TODO: Take into account the scalers once we support them */
2986 return config->base.adjusted_mode.crtc_clock;
2987 }
2988
2989 /*
2990 * The max latency should be 257 (max the punit can code is 255 and we add 2us
2991 * for the read latency) and bytes_per_pixel should always be <= 8, so that
2992 * should allow pixel_rate up to ~2 GHz which seems sufficient since max
2993 * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
2994 */
2995 static uint32_t skl_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
2996 uint32_t latency)
2997 {
2998 uint32_t wm_intermediate_val, ret;
2999
3000 if (latency == 0)
3001 return UINT_MAX;
3002
3003 wm_intermediate_val = latency * pixel_rate * bytes_per_pixel / 512;
3004 ret = DIV_ROUND_UP(wm_intermediate_val, 1000);
3005
3006 return ret;
3007 }
3008
3009 static uint32_t skl_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
3010 uint32_t horiz_pixels, uint8_t bytes_per_pixel,
3011 uint64_t tiling, uint32_t latency)
3012 {
3013 uint32_t ret;
3014 uint32_t plane_bytes_per_line, plane_blocks_per_line;
3015 uint32_t wm_intermediate_val;
3016
3017 if (latency == 0)
3018 return UINT_MAX;
3019
3020 plane_bytes_per_line = horiz_pixels * bytes_per_pixel;
3021
3022 if (tiling == I915_FORMAT_MOD_Y_TILED ||
3023 tiling == I915_FORMAT_MOD_Yf_TILED) {
3024 plane_bytes_per_line *= 4;
3025 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
3026 plane_blocks_per_line /= 4;
3027 } else {
3028 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
3029 }
3030
3031 wm_intermediate_val = latency * pixel_rate;
3032 ret = DIV_ROUND_UP(wm_intermediate_val, pipe_htotal * 1000) *
3033 plane_blocks_per_line;
3034
3035 return ret;
3036 }
3037
3038 static bool skl_ddb_allocation_changed(const struct skl_ddb_allocation *new_ddb,
3039 const struct intel_crtc *intel_crtc)
3040 {
3041 struct drm_device *dev = intel_crtc->base.dev;
3042 struct drm_i915_private *dev_priv = dev->dev_private;
3043 const struct skl_ddb_allocation *cur_ddb = &dev_priv->wm.skl_hw.ddb;
3044 enum pipe pipe = intel_crtc->pipe;
3045
3046 if (memcmp(new_ddb->plane[pipe], cur_ddb->plane[pipe],
3047 sizeof(new_ddb->plane[pipe])))
3048 return true;
3049
3050 if (memcmp(&new_ddb->plane[pipe][PLANE_CURSOR], &cur_ddb->plane[pipe][PLANE_CURSOR],
3051 sizeof(new_ddb->plane[pipe][PLANE_CURSOR])))
3052 return true;
3053
3054 return false;
3055 }
3056
3057 static void skl_compute_wm_global_parameters(struct drm_device *dev,
3058 struct intel_wm_config *config)
3059 {
3060 struct drm_crtc *crtc;
3061 struct drm_plane *plane;
3062
3063 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
3064 config->num_pipes_active += to_intel_crtc(crtc)->active;
3065
3066 /* FIXME: I don't think we need those two global parameters on SKL */
3067 list_for_each_entry(plane, &dev->mode_config.plane_list, head) {
3068 struct intel_plane *intel_plane = to_intel_plane(plane);
3069
3070 config->sprites_enabled |= intel_plane->wm.enabled;
3071 config->sprites_scaled |= intel_plane->wm.scaled;
3072 }
3073 }
3074
3075 static void skl_compute_wm_pipe_parameters(struct drm_crtc *crtc,
3076 struct skl_pipe_wm_parameters *p)
3077 {
3078 struct drm_device *dev = crtc->dev;
3079 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3080 enum pipe pipe = intel_crtc->pipe;
3081 struct drm_plane *plane;
3082 struct drm_framebuffer *fb;
3083 int i = 1; /* Index for sprite planes start */
3084
3085 p->active = intel_crtc->active;
3086 if (p->active) {
3087 p->pipe_htotal = intel_crtc->config->base.adjusted_mode.crtc_htotal;
3088 p->pixel_rate = skl_pipe_pixel_rate(intel_crtc->config);
3089
3090 fb = crtc->primary->state->fb;
3091 /* For planar: Bpp is for uv plane, y_Bpp is for y plane */
3092 if (fb) {
3093 p->plane[0].enabled = true;
3094 p->plane[0].bytes_per_pixel = fb->pixel_format == DRM_FORMAT_NV12 ?
3095 drm_format_plane_cpp(fb->pixel_format, 1) :
3096 drm_format_plane_cpp(fb->pixel_format, 0);
3097 p->plane[0].y_bytes_per_pixel = fb->pixel_format == DRM_FORMAT_NV12 ?
3098 drm_format_plane_cpp(fb->pixel_format, 0) : 0;
3099 p->plane[0].tiling = fb->modifier[0];
3100 } else {
3101 p->plane[0].enabled = false;
3102 p->plane[0].bytes_per_pixel = 0;
3103 p->plane[0].y_bytes_per_pixel = 0;
3104 p->plane[0].tiling = DRM_FORMAT_MOD_NONE;
3105 }
3106 p->plane[0].horiz_pixels = intel_crtc->config->pipe_src_w;
3107 p->plane[0].vert_pixels = intel_crtc->config->pipe_src_h;
3108 p->plane[0].rotation = crtc->primary->state->rotation;
3109
3110 fb = crtc->cursor->state->fb;
3111 p->plane[PLANE_CURSOR].y_bytes_per_pixel = 0;
3112 if (fb) {
3113 p->plane[PLANE_CURSOR].enabled = true;
3114 p->plane[PLANE_CURSOR].bytes_per_pixel = fb->bits_per_pixel / 8;
3115 p->plane[PLANE_CURSOR].horiz_pixels = crtc->cursor->state->crtc_w;
3116 p->plane[PLANE_CURSOR].vert_pixels = crtc->cursor->state->crtc_h;
3117 } else {
3118 p->plane[PLANE_CURSOR].enabled = false;
3119 p->plane[PLANE_CURSOR].bytes_per_pixel = 0;
3120 p->plane[PLANE_CURSOR].horiz_pixels = 64;
3121 p->plane[PLANE_CURSOR].vert_pixels = 64;
3122 }
3123 }
3124
3125 list_for_each_entry(plane, &dev->mode_config.plane_list, head) {
3126 struct intel_plane *intel_plane = to_intel_plane(plane);
3127
3128 if (intel_plane->pipe == pipe &&
3129 plane->type == DRM_PLANE_TYPE_OVERLAY)
3130 p->plane[i++] = intel_plane->wm;
3131 }
3132 }
3133
3134 static bool skl_compute_plane_wm(const struct drm_i915_private *dev_priv,
3135 struct skl_pipe_wm_parameters *p,
3136 struct intel_plane_wm_parameters *p_params,
3137 uint16_t ddb_allocation,
3138 int level,
3139 uint16_t *out_blocks, /* out */
3140 uint8_t *out_lines /* out */)
3141 {
3142 uint32_t latency = dev_priv->wm.skl_latency[level];
3143 uint32_t method1, method2;
3144 uint32_t plane_bytes_per_line, plane_blocks_per_line;
3145 uint32_t res_blocks, res_lines;
3146 uint32_t selected_result;
3147 uint8_t bytes_per_pixel;
3148
3149 if (latency == 0 || !p->active || !p_params->enabled)
3150 return false;
3151
3152 bytes_per_pixel = p_params->y_bytes_per_pixel ?
3153 p_params->y_bytes_per_pixel :
3154 p_params->bytes_per_pixel;
3155 method1 = skl_wm_method1(p->pixel_rate,
3156 bytes_per_pixel,
3157 latency);
3158 method2 = skl_wm_method2(p->pixel_rate,
3159 p->pipe_htotal,
3160 p_params->horiz_pixels,
3161 bytes_per_pixel,
3162 p_params->tiling,
3163 latency);
3164
3165 plane_bytes_per_line = p_params->horiz_pixels * bytes_per_pixel;
3166 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
3167
3168 if (p_params->tiling == I915_FORMAT_MOD_Y_TILED ||
3169 p_params->tiling == I915_FORMAT_MOD_Yf_TILED) {
3170 uint32_t min_scanlines = 4;
3171 uint32_t y_tile_minimum;
3172 if (intel_rotation_90_or_270(p_params->rotation)) {
3173 switch (p_params->bytes_per_pixel) {
3174 case 1:
3175 min_scanlines = 16;
3176 break;
3177 case 2:
3178 min_scanlines = 8;
3179 break;
3180 case 8:
3181 WARN(1, "Unsupported pixel depth for rotation");
3182 }
3183 }
3184 y_tile_minimum = plane_blocks_per_line * min_scanlines;
3185 selected_result = max(method2, y_tile_minimum);
3186 } else {
3187 if ((ddb_allocation / plane_blocks_per_line) >= 1)
3188 selected_result = min(method1, method2);
3189 else
3190 selected_result = method1;
3191 }
3192
3193 res_blocks = selected_result + 1;
3194 res_lines = DIV_ROUND_UP(selected_result, plane_blocks_per_line);
3195
3196 if (level >= 1 && level <= 7) {
3197 if (p_params->tiling == I915_FORMAT_MOD_Y_TILED ||
3198 p_params->tiling == I915_FORMAT_MOD_Yf_TILED)
3199 res_lines += 4;
3200 else
3201 res_blocks++;
3202 }
3203
3204 if (res_blocks >= ddb_allocation || res_lines > 31)
3205 return false;
3206
3207 *out_blocks = res_blocks;
3208 *out_lines = res_lines;
3209
3210 return true;
3211 }
3212
3213 static void skl_compute_wm_level(const struct drm_i915_private *dev_priv,
3214 struct skl_ddb_allocation *ddb,
3215 struct skl_pipe_wm_parameters *p,
3216 enum pipe pipe,
3217 int level,
3218 int num_planes,
3219 struct skl_wm_level *result)
3220 {
3221 uint16_t ddb_blocks;
3222 int i;
3223
3224 for (i = 0; i < num_planes; i++) {
3225 ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][i]);
3226
3227 result->plane_en[i] = skl_compute_plane_wm(dev_priv,
3228 p, &p->plane[i],
3229 ddb_blocks,
3230 level,
3231 &result->plane_res_b[i],
3232 &result->plane_res_l[i]);
3233 }
3234
3235 ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][PLANE_CURSOR]);
3236 result->plane_en[PLANE_CURSOR] = skl_compute_plane_wm(dev_priv, p,
3237 &p->plane[PLANE_CURSOR],
3238 ddb_blocks, level,
3239 &result->plane_res_b[PLANE_CURSOR],
3240 &result->plane_res_l[PLANE_CURSOR]);
3241 }
3242
3243 static uint32_t
3244 skl_compute_linetime_wm(struct drm_crtc *crtc, struct skl_pipe_wm_parameters *p)
3245 {
3246 if (!to_intel_crtc(crtc)->active)
3247 return 0;
3248
3249 if (WARN_ON(p->pixel_rate == 0))
3250 return 0;
3251
3252 return DIV_ROUND_UP(8 * p->pipe_htotal * 1000, p->pixel_rate);
3253 }
3254
3255 static void skl_compute_transition_wm(struct drm_crtc *crtc,
3256 struct skl_pipe_wm_parameters *params,
3257 struct skl_wm_level *trans_wm /* out */)
3258 {
3259 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3260 int i;
3261
3262 if (!params->active)
3263 return;
3264
3265 /* Until we know more, just disable transition WMs */
3266 for (i = 0; i < intel_num_planes(intel_crtc); i++)
3267 trans_wm->plane_en[i] = false;
3268 trans_wm->plane_en[PLANE_CURSOR] = false;
3269 }
3270
3271 static void skl_compute_pipe_wm(struct drm_crtc *crtc,
3272 struct skl_ddb_allocation *ddb,
3273 struct skl_pipe_wm_parameters *params,
3274 struct skl_pipe_wm *pipe_wm)
3275 {
3276 struct drm_device *dev = crtc->dev;
3277 const struct drm_i915_private *dev_priv = dev->dev_private;
3278 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3279 int level, max_level = ilk_wm_max_level(dev);
3280
3281 for (level = 0; level <= max_level; level++) {
3282 skl_compute_wm_level(dev_priv, ddb, params, intel_crtc->pipe,
3283 level, intel_num_planes(intel_crtc),
3284 &pipe_wm->wm[level]);
3285 }
3286 pipe_wm->linetime = skl_compute_linetime_wm(crtc, params);
3287
3288 skl_compute_transition_wm(crtc, params, &pipe_wm->trans_wm);
3289 }
3290
3291 static void skl_compute_wm_results(struct drm_device *dev,
3292 struct skl_pipe_wm_parameters *p,
3293 struct skl_pipe_wm *p_wm,
3294 struct skl_wm_values *r,
3295 struct intel_crtc *intel_crtc)
3296 {
3297 int level, max_level = ilk_wm_max_level(dev);
3298 enum pipe pipe = intel_crtc->pipe;
3299 uint32_t temp;
3300 int i;
3301
3302 for (level = 0; level <= max_level; level++) {
3303 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3304 temp = 0;
3305
3306 temp |= p_wm->wm[level].plane_res_l[i] <<
3307 PLANE_WM_LINES_SHIFT;
3308 temp |= p_wm->wm[level].plane_res_b[i];
3309 if (p_wm->wm[level].plane_en[i])
3310 temp |= PLANE_WM_EN;
3311
3312 r->plane[pipe][i][level] = temp;
3313 }
3314
3315 temp = 0;
3316
3317 temp |= p_wm->wm[level].plane_res_l[PLANE_CURSOR] << PLANE_WM_LINES_SHIFT;
3318 temp |= p_wm->wm[level].plane_res_b[PLANE_CURSOR];
3319
3320 if (p_wm->wm[level].plane_en[PLANE_CURSOR])
3321 temp |= PLANE_WM_EN;
3322
3323 r->plane[pipe][PLANE_CURSOR][level] = temp;
3324
3325 }
3326
3327 /* transition WMs */
3328 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3329 temp = 0;
3330 temp |= p_wm->trans_wm.plane_res_l[i] << PLANE_WM_LINES_SHIFT;
3331 temp |= p_wm->trans_wm.plane_res_b[i];
3332 if (p_wm->trans_wm.plane_en[i])
3333 temp |= PLANE_WM_EN;
3334
3335 r->plane_trans[pipe][i] = temp;
3336 }
3337
3338 temp = 0;
3339 temp |= p_wm->trans_wm.plane_res_l[PLANE_CURSOR] << PLANE_WM_LINES_SHIFT;
3340 temp |= p_wm->trans_wm.plane_res_b[PLANE_CURSOR];
3341 if (p_wm->trans_wm.plane_en[PLANE_CURSOR])
3342 temp |= PLANE_WM_EN;
3343
3344 r->plane_trans[pipe][PLANE_CURSOR] = temp;
3345
3346 r->wm_linetime[pipe] = p_wm->linetime;
3347 }
3348
3349 static void skl_ddb_entry_write(struct drm_i915_private *dev_priv, uint32_t reg,
3350 const struct skl_ddb_entry *entry)
3351 {
3352 if (entry->end)
3353 I915_WRITE(reg, (entry->end - 1) << 16 | entry->start);
3354 else
3355 I915_WRITE(reg, 0);
3356 }
3357
3358 static void skl_write_wm_values(struct drm_i915_private *dev_priv,
3359 const struct skl_wm_values *new)
3360 {
3361 struct drm_device *dev = dev_priv->dev;
3362 struct intel_crtc *crtc;
3363
3364 list_for_each_entry(crtc, &dev->mode_config.crtc_list, base.head) {
3365 int i, level, max_level = ilk_wm_max_level(dev);
3366 enum pipe pipe = crtc->pipe;
3367
3368 if (!new->dirty[pipe])
3369 continue;
3370
3371 I915_WRITE(PIPE_WM_LINETIME(pipe), new->wm_linetime[pipe]);
3372
3373 for (level = 0; level <= max_level; level++) {
3374 for (i = 0; i < intel_num_planes(crtc); i++)
3375 I915_WRITE(PLANE_WM(pipe, i, level),
3376 new->plane[pipe][i][level]);
3377 I915_WRITE(CUR_WM(pipe, level),
3378 new->plane[pipe][PLANE_CURSOR][level]);
3379 }
3380 for (i = 0; i < intel_num_planes(crtc); i++)
3381 I915_WRITE(PLANE_WM_TRANS(pipe, i),
3382 new->plane_trans[pipe][i]);
3383 I915_WRITE(CUR_WM_TRANS(pipe),
3384 new->plane_trans[pipe][PLANE_CURSOR]);
3385
3386 for (i = 0; i < intel_num_planes(crtc); i++) {
3387 skl_ddb_entry_write(dev_priv,
3388 PLANE_BUF_CFG(pipe, i),
3389 &new->ddb.plane[pipe][i]);
3390 skl_ddb_entry_write(dev_priv,
3391 PLANE_NV12_BUF_CFG(pipe, i),
3392 &new->ddb.y_plane[pipe][i]);
3393 }
3394
3395 skl_ddb_entry_write(dev_priv, CUR_BUF_CFG(pipe),
3396 &new->ddb.plane[pipe][PLANE_CURSOR]);
3397 }
3398 }
3399
3400 /*
3401 * When setting up a new DDB allocation arrangement, we need to correctly
3402 * sequence the times at which the new allocations for the pipes are taken into
3403 * account or we'll have pipes fetching from space previously allocated to
3404 * another pipe.
3405 *
3406 * Roughly the sequence looks like:
3407 * 1. re-allocate the pipe(s) with the allocation being reduced and not
3408 * overlapping with a previous light-up pipe (another way to put it is:
3409 * pipes with their new allocation strickly included into their old ones).
3410 * 2. re-allocate the other pipes that get their allocation reduced
3411 * 3. allocate the pipes having their allocation increased
3412 *
3413 * Steps 1. and 2. are here to take care of the following case:
3414 * - Initially DDB looks like this:
3415 * | B | C |
3416 * - enable pipe A.
3417 * - pipe B has a reduced DDB allocation that overlaps with the old pipe C
3418 * allocation
3419 * | A | B | C |
3420 *
3421 * We need to sequence the re-allocation: C, B, A (and not B, C, A).
3422 */
3423
3424 static void
3425 skl_wm_flush_pipe(struct drm_i915_private *dev_priv, enum pipe pipe, int pass)
3426 {
3427 int plane;
3428
3429 DRM_DEBUG_KMS("flush pipe %c (pass %d)\n", pipe_name(pipe), pass);
3430
3431 for_each_plane(dev_priv, pipe, plane) {
3432 I915_WRITE(PLANE_SURF(pipe, plane),
3433 I915_READ(PLANE_SURF(pipe, plane)));
3434 }
3435 I915_WRITE(CURBASE(pipe), I915_READ(CURBASE(pipe)));
3436 }
3437
3438 static bool
3439 skl_ddb_allocation_included(const struct skl_ddb_allocation *old,
3440 const struct skl_ddb_allocation *new,
3441 enum pipe pipe)
3442 {
3443 uint16_t old_size, new_size;
3444
3445 old_size = skl_ddb_entry_size(&old->pipe[pipe]);
3446 new_size = skl_ddb_entry_size(&new->pipe[pipe]);
3447
3448 return old_size != new_size &&
3449 new->pipe[pipe].start >= old->pipe[pipe].start &&
3450 new->pipe[pipe].end <= old->pipe[pipe].end;
3451 }
3452
3453 static void skl_flush_wm_values(struct drm_i915_private *dev_priv,
3454 struct skl_wm_values *new_values)
3455 {
3456 struct drm_device *dev = dev_priv->dev;
3457 struct skl_ddb_allocation *cur_ddb, *new_ddb;
3458 bool reallocated[I915_MAX_PIPES] = {};
3459 struct intel_crtc *crtc;
3460 enum pipe pipe;
3461
3462 new_ddb = &new_values->ddb;
3463 cur_ddb = &dev_priv->wm.skl_hw.ddb;
3464
3465 /*
3466 * First pass: flush the pipes with the new allocation contained into
3467 * the old space.
3468 *
3469 * We'll wait for the vblank on those pipes to ensure we can safely
3470 * re-allocate the freed space without this pipe fetching from it.
3471 */
3472 for_each_intel_crtc(dev, crtc) {
3473 if (!crtc->active)
3474 continue;
3475
3476 pipe = crtc->pipe;
3477
3478 if (!skl_ddb_allocation_included(cur_ddb, new_ddb, pipe))
3479 continue;
3480
3481 skl_wm_flush_pipe(dev_priv, pipe, 1);
3482 intel_wait_for_vblank(dev, pipe);
3483
3484 reallocated[pipe] = true;
3485 }
3486
3487
3488 /*
3489 * Second pass: flush the pipes that are having their allocation
3490 * reduced, but overlapping with a previous allocation.
3491 *
3492 * Here as well we need to wait for the vblank to make sure the freed
3493 * space is not used anymore.
3494 */
3495 for_each_intel_crtc(dev, crtc) {
3496 if (!crtc->active)
3497 continue;
3498
3499 pipe = crtc->pipe;
3500
3501 if (reallocated[pipe])
3502 continue;
3503
3504 if (skl_ddb_entry_size(&new_ddb->pipe[pipe]) <
3505 skl_ddb_entry_size(&cur_ddb->pipe[pipe])) {
3506 skl_wm_flush_pipe(dev_priv, pipe, 2);
3507 intel_wait_for_vblank(dev, pipe);
3508 reallocated[pipe] = true;
3509 }
3510 }
3511
3512 /*
3513 * Third pass: flush the pipes that got more space allocated.
3514 *
3515 * We don't need to actively wait for the update here, next vblank
3516 * will just get more DDB space with the correct WM values.
3517 */
3518 for_each_intel_crtc(dev, crtc) {
3519 if (!crtc->active)
3520 continue;
3521
3522 pipe = crtc->pipe;
3523
3524 /*
3525 * At this point, only the pipes more space than before are
3526 * left to re-allocate.
3527 */
3528 if (reallocated[pipe])
3529 continue;
3530
3531 skl_wm_flush_pipe(dev_priv, pipe, 3);
3532 }
3533 }
3534
3535 static bool skl_update_pipe_wm(struct drm_crtc *crtc,
3536 struct skl_pipe_wm_parameters *params,
3537 struct intel_wm_config *config,
3538 struct skl_ddb_allocation *ddb, /* out */
3539 struct skl_pipe_wm *pipe_wm /* out */)
3540 {
3541 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3542
3543 skl_compute_wm_pipe_parameters(crtc, params);
3544 skl_allocate_pipe_ddb(crtc, config, params, ddb);
3545 skl_compute_pipe_wm(crtc, ddb, params, pipe_wm);
3546
3547 if (!memcmp(&intel_crtc->wm.skl_active, pipe_wm, sizeof(*pipe_wm)))
3548 return false;
3549
3550 intel_crtc->wm.skl_active = *pipe_wm;
3551
3552 return true;
3553 }
3554
3555 static void skl_update_other_pipe_wm(struct drm_device *dev,
3556 struct drm_crtc *crtc,
3557 struct intel_wm_config *config,
3558 struct skl_wm_values *r)
3559 {
3560 struct intel_crtc *intel_crtc;
3561 struct intel_crtc *this_crtc = to_intel_crtc(crtc);
3562
3563 /*
3564 * If the WM update hasn't changed the allocation for this_crtc (the
3565 * crtc we are currently computing the new WM values for), other
3566 * enabled crtcs will keep the same allocation and we don't need to
3567 * recompute anything for them.
3568 */
3569 if (!skl_ddb_allocation_changed(&r->ddb, this_crtc))
3570 return;
3571
3572 /*
3573 * Otherwise, because of this_crtc being freshly enabled/disabled, the
3574 * other active pipes need new DDB allocation and WM values.
3575 */
3576 list_for_each_entry(intel_crtc, &dev->mode_config.crtc_list,
3577 base.head) {
3578 struct skl_pipe_wm_parameters params = {};
3579 struct skl_pipe_wm pipe_wm = {};
3580 bool wm_changed;
3581
3582 if (this_crtc->pipe == intel_crtc->pipe)
3583 continue;
3584
3585 if (!intel_crtc->active)
3586 continue;
3587
3588 wm_changed = skl_update_pipe_wm(&intel_crtc->base,
3589 &params, config,
3590 &r->ddb, &pipe_wm);
3591
3592 /*
3593 * If we end up re-computing the other pipe WM values, it's
3594 * because it was really needed, so we expect the WM values to
3595 * be different.
3596 */
3597 WARN_ON(!wm_changed);
3598
3599 skl_compute_wm_results(dev, &params, &pipe_wm, r, intel_crtc);
3600 r->dirty[intel_crtc->pipe] = true;
3601 }
3602 }
3603
3604 static void skl_clear_wm(struct skl_wm_values *watermarks, enum pipe pipe)
3605 {
3606 watermarks->wm_linetime[pipe] = 0;
3607 memset(watermarks->plane[pipe], 0,
3608 sizeof(uint32_t) * 8 * I915_MAX_PLANES);
3609 memset(watermarks->plane_trans[pipe],
3610 0, sizeof(uint32_t) * I915_MAX_PLANES);
3611 watermarks->plane_trans[pipe][PLANE_CURSOR] = 0;
3612
3613 /* Clear ddb entries for pipe */
3614 memset(&watermarks->ddb.pipe[pipe], 0, sizeof(struct skl_ddb_entry));
3615 memset(&watermarks->ddb.plane[pipe], 0,
3616 sizeof(struct skl_ddb_entry) * I915_MAX_PLANES);
3617 memset(&watermarks->ddb.y_plane[pipe], 0,
3618 sizeof(struct skl_ddb_entry) * I915_MAX_PLANES);
3619 memset(&watermarks->ddb.plane[pipe][PLANE_CURSOR], 0,
3620 sizeof(struct skl_ddb_entry));
3621
3622 }
3623
3624 static void skl_update_wm(struct drm_crtc *crtc)
3625 {
3626 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3627 struct drm_device *dev = crtc->dev;
3628 struct drm_i915_private *dev_priv = dev->dev_private;
3629 struct skl_pipe_wm_parameters params = {};
3630 struct skl_wm_values *results = &dev_priv->wm.skl_results;
3631 struct skl_pipe_wm pipe_wm = {};
3632 struct intel_wm_config config = {};
3633
3634
3635 /* Clear all dirty flags */
3636 memset(results->dirty, 0, sizeof(bool) * I915_MAX_PIPES);
3637
3638 skl_clear_wm(results, intel_crtc->pipe);
3639
3640 skl_compute_wm_global_parameters(dev, &config);
3641
3642 if (!skl_update_pipe_wm(crtc, &params, &config,
3643 &results->ddb, &pipe_wm))
3644 return;
3645
3646 skl_compute_wm_results(dev, &params, &pipe_wm, results, intel_crtc);
3647 results->dirty[intel_crtc->pipe] = true;
3648
3649 skl_update_other_pipe_wm(dev, crtc, &config, results);
3650 skl_write_wm_values(dev_priv, results);
3651 skl_flush_wm_values(dev_priv, results);
3652
3653 /* store the new configuration */
3654 dev_priv->wm.skl_hw = *results;
3655 }
3656
3657 static void
3658 skl_update_sprite_wm(struct drm_plane *plane, struct drm_crtc *crtc,
3659 uint32_t sprite_width, uint32_t sprite_height,
3660 int pixel_size, bool enabled, bool scaled)
3661 {
3662 struct intel_plane *intel_plane = to_intel_plane(plane);
3663 struct drm_framebuffer *fb = plane->state->fb;
3664
3665 intel_plane->wm.enabled = enabled;
3666 intel_plane->wm.scaled = scaled;
3667 intel_plane->wm.horiz_pixels = sprite_width;
3668 intel_plane->wm.vert_pixels = sprite_height;
3669 intel_plane->wm.tiling = DRM_FORMAT_MOD_NONE;
3670
3671 /* For planar: Bpp is for UV plane, y_Bpp is for Y plane */
3672 intel_plane->wm.bytes_per_pixel =
3673 (fb && fb->pixel_format == DRM_FORMAT_NV12) ?
3674 drm_format_plane_cpp(plane->state->fb->pixel_format, 1) : pixel_size;
3675 intel_plane->wm.y_bytes_per_pixel =
3676 (fb && fb->pixel_format == DRM_FORMAT_NV12) ?
3677 drm_format_plane_cpp(plane->state->fb->pixel_format, 0) : 0;
3678
3679 /*
3680 * Framebuffer can be NULL on plane disable, but it does not
3681 * matter for watermarks if we assume no tiling in that case.
3682 */
3683 if (fb)
3684 intel_plane->wm.tiling = fb->modifier[0];
3685 intel_plane->wm.rotation = plane->state->rotation;
3686
3687 skl_update_wm(crtc);
3688 }
3689
3690 static void ilk_update_wm(struct drm_crtc *crtc)
3691 {
3692 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3693 struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
3694 struct drm_device *dev = crtc->dev;
3695 struct drm_i915_private *dev_priv = dev->dev_private;
3696 struct ilk_wm_maximums max;
3697 struct ilk_wm_values results = {};
3698 enum intel_ddb_partitioning partitioning;
3699 struct intel_pipe_wm pipe_wm = {};
3700 struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
3701 struct intel_wm_config config = {};
3702
3703 WARN_ON(cstate->base.active != intel_crtc->active);
3704
3705 intel_compute_pipe_wm(cstate, &pipe_wm);
3706
3707 if (!memcmp(&intel_crtc->wm.active, &pipe_wm, sizeof(pipe_wm)))
3708 return;
3709
3710 intel_crtc->wm.active = pipe_wm;
3711
3712 ilk_compute_wm_config(dev, &config);
3713
3714 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
3715 ilk_wm_merge(dev, &config, &max, &lp_wm_1_2);
3716
3717 /* 5/6 split only in single pipe config on IVB+ */
3718 if (INTEL_INFO(dev)->gen >= 7 &&
3719 config.num_pipes_active == 1 && config.sprites_enabled) {
3720 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
3721 ilk_wm_merge(dev, &config, &max, &lp_wm_5_6);
3722
3723 best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
3724 } else {
3725 best_lp_wm = &lp_wm_1_2;
3726 }
3727
3728 partitioning = (best_lp_wm == &lp_wm_1_2) ?
3729 INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
3730
3731 ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
3732
3733 ilk_write_wm_values(dev_priv, &results);
3734 }
3735
3736 static void
3737 ilk_update_sprite_wm(struct drm_plane *plane,
3738 struct drm_crtc *crtc,
3739 uint32_t sprite_width, uint32_t sprite_height,
3740 int pixel_size, bool enabled, bool scaled)
3741 {
3742 struct drm_device *dev = plane->dev;
3743 struct intel_plane *intel_plane = to_intel_plane(plane);
3744
3745 /*
3746 * IVB workaround: must disable low power watermarks for at least
3747 * one frame before enabling scaling. LP watermarks can be re-enabled
3748 * when scaling is disabled.
3749 *
3750 * WaCxSRDisabledForSpriteScaling:ivb
3751 */
3752 if (IS_IVYBRIDGE(dev) && scaled && ilk_disable_lp_wm(dev))
3753 intel_wait_for_vblank(dev, intel_plane->pipe);
3754
3755 ilk_update_wm(crtc);
3756 }
3757
3758 static void skl_pipe_wm_active_state(uint32_t val,
3759 struct skl_pipe_wm *active,
3760 bool is_transwm,
3761 bool is_cursor,
3762 int i,
3763 int level)
3764 {
3765 bool is_enabled = (val & PLANE_WM_EN) != 0;
3766
3767 if (!is_transwm) {
3768 if (!is_cursor) {
3769 active->wm[level].plane_en[i] = is_enabled;
3770 active->wm[level].plane_res_b[i] =
3771 val & PLANE_WM_BLOCKS_MASK;
3772 active->wm[level].plane_res_l[i] =
3773 (val >> PLANE_WM_LINES_SHIFT) &
3774 PLANE_WM_LINES_MASK;
3775 } else {
3776 active->wm[level].plane_en[PLANE_CURSOR] = is_enabled;
3777 active->wm[level].plane_res_b[PLANE_CURSOR] =
3778 val & PLANE_WM_BLOCKS_MASK;
3779 active->wm[level].plane_res_l[PLANE_CURSOR] =
3780 (val >> PLANE_WM_LINES_SHIFT) &
3781 PLANE_WM_LINES_MASK;
3782 }
3783 } else {
3784 if (!is_cursor) {
3785 active->trans_wm.plane_en[i] = is_enabled;
3786 active->trans_wm.plane_res_b[i] =
3787 val & PLANE_WM_BLOCKS_MASK;
3788 active->trans_wm.plane_res_l[i] =
3789 (val >> PLANE_WM_LINES_SHIFT) &
3790 PLANE_WM_LINES_MASK;
3791 } else {
3792 active->trans_wm.plane_en[PLANE_CURSOR] = is_enabled;
3793 active->trans_wm.plane_res_b[PLANE_CURSOR] =
3794 val & PLANE_WM_BLOCKS_MASK;
3795 active->trans_wm.plane_res_l[PLANE_CURSOR] =
3796 (val >> PLANE_WM_LINES_SHIFT) &
3797 PLANE_WM_LINES_MASK;
3798 }
3799 }
3800 }
3801
3802 static void skl_pipe_wm_get_hw_state(struct drm_crtc *crtc)
3803 {
3804 struct drm_device *dev = crtc->dev;
3805 struct drm_i915_private *dev_priv = dev->dev_private;
3806 struct skl_wm_values *hw = &dev_priv->wm.skl_hw;
3807 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3808 struct skl_pipe_wm *active = &intel_crtc->wm.skl_active;
3809 enum pipe pipe = intel_crtc->pipe;
3810 int level, i, max_level;
3811 uint32_t temp;
3812
3813 max_level = ilk_wm_max_level(dev);
3814
3815 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
3816
3817 for (level = 0; level <= max_level; level++) {
3818 for (i = 0; i < intel_num_planes(intel_crtc); i++)
3819 hw->plane[pipe][i][level] =
3820 I915_READ(PLANE_WM(pipe, i, level));
3821 hw->plane[pipe][PLANE_CURSOR][level] = I915_READ(CUR_WM(pipe, level));
3822 }
3823
3824 for (i = 0; i < intel_num_planes(intel_crtc); i++)
3825 hw->plane_trans[pipe][i] = I915_READ(PLANE_WM_TRANS(pipe, i));
3826 hw->plane_trans[pipe][PLANE_CURSOR] = I915_READ(CUR_WM_TRANS(pipe));
3827
3828 if (!intel_crtc->active)
3829 return;
3830
3831 hw->dirty[pipe] = true;
3832
3833 active->linetime = hw->wm_linetime[pipe];
3834
3835 for (level = 0; level <= max_level; level++) {
3836 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3837 temp = hw->plane[pipe][i][level];
3838 skl_pipe_wm_active_state(temp, active, false,
3839 false, i, level);
3840 }
3841 temp = hw->plane[pipe][PLANE_CURSOR][level];
3842 skl_pipe_wm_active_state(temp, active, false, true, i, level);
3843 }
3844
3845 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3846 temp = hw->plane_trans[pipe][i];
3847 skl_pipe_wm_active_state(temp, active, true, false, i, 0);
3848 }
3849
3850 temp = hw->plane_trans[pipe][PLANE_CURSOR];
3851 skl_pipe_wm_active_state(temp, active, true, true, i, 0);
3852 }
3853
3854 void skl_wm_get_hw_state(struct drm_device *dev)
3855 {
3856 struct drm_i915_private *dev_priv = dev->dev_private;
3857 struct skl_ddb_allocation *ddb = &dev_priv->wm.skl_hw.ddb;
3858 struct drm_crtc *crtc;
3859
3860 skl_ddb_get_hw_state(dev_priv, ddb);
3861 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
3862 skl_pipe_wm_get_hw_state(crtc);
3863 }
3864
3865 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
3866 {
3867 struct drm_device *dev = crtc->dev;
3868 struct drm_i915_private *dev_priv = dev->dev_private;
3869 struct ilk_wm_values *hw = &dev_priv->wm.hw;
3870 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3871 struct intel_pipe_wm *active = &intel_crtc->wm.active;
3872 enum pipe pipe = intel_crtc->pipe;
3873 static const unsigned int wm0_pipe_reg[] = {
3874 [PIPE_A] = WM0_PIPEA_ILK,
3875 [PIPE_B] = WM0_PIPEB_ILK,
3876 [PIPE_C] = WM0_PIPEC_IVB,
3877 };
3878
3879 hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
3880 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
3881 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
3882
3883 active->pipe_enabled = intel_crtc->active;
3884
3885 if (active->pipe_enabled) {
3886 u32 tmp = hw->wm_pipe[pipe];
3887
3888 /*
3889 * For active pipes LP0 watermark is marked as
3890 * enabled, and LP1+ watermaks as disabled since
3891 * we can't really reverse compute them in case
3892 * multiple pipes are active.
3893 */
3894 active->wm[0].enable = true;
3895 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
3896 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
3897 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
3898 active->linetime = hw->wm_linetime[pipe];
3899 } else {
3900 int level, max_level = ilk_wm_max_level(dev);
3901
3902 /*
3903 * For inactive pipes, all watermark levels
3904 * should be marked as enabled but zeroed,
3905 * which is what we'd compute them to.
3906 */
3907 for (level = 0; level <= max_level; level++)
3908 active->wm[level].enable = true;
3909 }
3910 }
3911
3912 #define _FW_WM(value, plane) \
3913 (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
3914 #define _FW_WM_VLV(value, plane) \
3915 (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
3916
3917 static void vlv_read_wm_values(struct drm_i915_private *dev_priv,
3918 struct vlv_wm_values *wm)
3919 {
3920 enum pipe pipe;
3921 uint32_t tmp;
3922
3923 for_each_pipe(dev_priv, pipe) {
3924 tmp = I915_READ(VLV_DDL(pipe));
3925
3926 wm->ddl[pipe].primary =
3927 (tmp >> DDL_PLANE_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
3928 wm->ddl[pipe].cursor =
3929 (tmp >> DDL_CURSOR_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
3930 wm->ddl[pipe].sprite[0] =
3931 (tmp >> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
3932 wm->ddl[pipe].sprite[1] =
3933 (tmp >> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
3934 }
3935
3936 tmp = I915_READ(DSPFW1);
3937 wm->sr.plane = _FW_WM(tmp, SR);
3938 wm->pipe[PIPE_B].cursor = _FW_WM(tmp, CURSORB);
3939 wm->pipe[PIPE_B].primary = _FW_WM_VLV(tmp, PLANEB);
3940 wm->pipe[PIPE_A].primary = _FW_WM_VLV(tmp, PLANEA);
3941
3942 tmp = I915_READ(DSPFW2);
3943 wm->pipe[PIPE_A].sprite[1] = _FW_WM_VLV(tmp, SPRITEB);
3944 wm->pipe[PIPE_A].cursor = _FW_WM(tmp, CURSORA);
3945 wm->pipe[PIPE_A].sprite[0] = _FW_WM_VLV(tmp, SPRITEA);
3946
3947 tmp = I915_READ(DSPFW3);
3948 wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
3949
3950 if (IS_CHERRYVIEW(dev_priv)) {
3951 tmp = I915_READ(DSPFW7_CHV);
3952 wm->pipe[PIPE_B].sprite[1] = _FW_WM_VLV(tmp, SPRITED);
3953 wm->pipe[PIPE_B].sprite[0] = _FW_WM_VLV(tmp, SPRITEC);
3954
3955 tmp = I915_READ(DSPFW8_CHV);
3956 wm->pipe[PIPE_C].sprite[1] = _FW_WM_VLV(tmp, SPRITEF);
3957 wm->pipe[PIPE_C].sprite[0] = _FW_WM_VLV(tmp, SPRITEE);
3958
3959 tmp = I915_READ(DSPFW9_CHV);
3960 wm->pipe[PIPE_C].primary = _FW_WM_VLV(tmp, PLANEC);
3961 wm->pipe[PIPE_C].cursor = _FW_WM(tmp, CURSORC);
3962
3963 tmp = I915_READ(DSPHOWM);
3964 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
3965 wm->pipe[PIPE_C].sprite[1] |= _FW_WM(tmp, SPRITEF_HI) << 8;
3966 wm->pipe[PIPE_C].sprite[0] |= _FW_WM(tmp, SPRITEE_HI) << 8;
3967 wm->pipe[PIPE_C].primary |= _FW_WM(tmp, PLANEC_HI) << 8;
3968 wm->pipe[PIPE_B].sprite[1] |= _FW_WM(tmp, SPRITED_HI) << 8;
3969 wm->pipe[PIPE_B].sprite[0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
3970 wm->pipe[PIPE_B].primary |= _FW_WM(tmp, PLANEB_HI) << 8;
3971 wm->pipe[PIPE_A].sprite[1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
3972 wm->pipe[PIPE_A].sprite[0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
3973 wm->pipe[PIPE_A].primary |= _FW_WM(tmp, PLANEA_HI) << 8;
3974 } else {
3975 tmp = I915_READ(DSPFW7);
3976 wm->pipe[PIPE_B].sprite[1] = _FW_WM_VLV(tmp, SPRITED);
3977 wm->pipe[PIPE_B].sprite[0] = _FW_WM_VLV(tmp, SPRITEC);
3978
3979 tmp = I915_READ(DSPHOWM);
3980 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
3981 wm->pipe[PIPE_B].sprite[1] |= _FW_WM(tmp, SPRITED_HI) << 8;
3982 wm->pipe[PIPE_B].sprite[0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
3983 wm->pipe[PIPE_B].primary |= _FW_WM(tmp, PLANEB_HI) << 8;
3984 wm->pipe[PIPE_A].sprite[1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
3985 wm->pipe[PIPE_A].sprite[0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
3986 wm->pipe[PIPE_A].primary |= _FW_WM(tmp, PLANEA_HI) << 8;
3987 }
3988 }
3989
3990 #undef _FW_WM
3991 #undef _FW_WM_VLV
3992
3993 void vlv_wm_get_hw_state(struct drm_device *dev)
3994 {
3995 struct drm_i915_private *dev_priv = to_i915(dev);
3996 struct vlv_wm_values *wm = &dev_priv->wm.vlv;
3997 struct intel_plane *plane;
3998 enum pipe pipe;
3999 u32 val;
4000
4001 vlv_read_wm_values(dev_priv, wm);
4002
4003 for_each_intel_plane(dev, plane) {
4004 switch (plane->base.type) {
4005 int sprite;
4006 case DRM_PLANE_TYPE_CURSOR:
4007 plane->wm.fifo_size = 63;
4008 break;
4009 case DRM_PLANE_TYPE_PRIMARY:
4010 plane->wm.fifo_size = vlv_get_fifo_size(dev, plane->pipe, 0);
4011 break;
4012 case DRM_PLANE_TYPE_OVERLAY:
4013 sprite = plane->plane;
4014 plane->wm.fifo_size = vlv_get_fifo_size(dev, plane->pipe, sprite + 1);
4015 break;
4016 }
4017 }
4018
4019 wm->cxsr = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
4020 wm->level = VLV_WM_LEVEL_PM2;
4021
4022 if (IS_CHERRYVIEW(dev_priv)) {
4023 mutex_lock(&dev_priv->rps.hw_lock);
4024
4025 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
4026 if (val & DSP_MAXFIFO_PM5_ENABLE)
4027 wm->level = VLV_WM_LEVEL_PM5;
4028
4029 /*
4030 * If DDR DVFS is disabled in the BIOS, Punit
4031 * will never ack the request. So if that happens
4032 * assume we don't have to enable/disable DDR DVFS
4033 * dynamically. To test that just set the REQ_ACK
4034 * bit to poke the Punit, but don't change the
4035 * HIGH/LOW bits so that we don't actually change
4036 * the current state.
4037 */
4038 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
4039 val |= FORCE_DDR_FREQ_REQ_ACK;
4040 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
4041
4042 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
4043 FORCE_DDR_FREQ_REQ_ACK) == 0, 3)) {
4044 DRM_DEBUG_KMS("Punit not acking DDR DVFS request, "
4045 "assuming DDR DVFS is disabled\n");
4046 dev_priv->wm.max_level = VLV_WM_LEVEL_PM5;
4047 } else {
4048 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
4049 if ((val & FORCE_DDR_HIGH_FREQ) == 0)
4050 wm->level = VLV_WM_LEVEL_DDR_DVFS;
4051 }
4052
4053 mutex_unlock(&dev_priv->rps.hw_lock);
4054 }
4055
4056 for_each_pipe(dev_priv, pipe)
4057 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
4058 pipe_name(pipe), wm->pipe[pipe].primary, wm->pipe[pipe].cursor,
4059 wm->pipe[pipe].sprite[0], wm->pipe[pipe].sprite[1]);
4060
4061 DRM_DEBUG_KMS("Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
4062 wm->sr.plane, wm->sr.cursor, wm->level, wm->cxsr);
4063 }
4064
4065 void ilk_wm_get_hw_state(struct drm_device *dev)
4066 {
4067 struct drm_i915_private *dev_priv = dev->dev_private;
4068 struct ilk_wm_values *hw = &dev_priv->wm.hw;
4069 struct drm_crtc *crtc;
4070
4071 for_each_crtc(dev, crtc)
4072 ilk_pipe_wm_get_hw_state(crtc);
4073
4074 hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
4075 hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
4076 hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
4077
4078 hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
4079 if (INTEL_INFO(dev)->gen >= 7) {
4080 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
4081 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
4082 }
4083
4084 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
4085 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
4086 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4087 else if (IS_IVYBRIDGE(dev))
4088 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
4089 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4090
4091 hw->enable_fbc_wm =
4092 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
4093 }
4094
4095 /**
4096 * intel_update_watermarks - update FIFO watermark values based on current modes
4097 *
4098 * Calculate watermark values for the various WM regs based on current mode
4099 * and plane configuration.
4100 *
4101 * There are several cases to deal with here:
4102 * - normal (i.e. non-self-refresh)
4103 * - self-refresh (SR) mode
4104 * - lines are large relative to FIFO size (buffer can hold up to 2)
4105 * - lines are small relative to FIFO size (buffer can hold more than 2
4106 * lines), so need to account for TLB latency
4107 *
4108 * The normal calculation is:
4109 * watermark = dotclock * bytes per pixel * latency
4110 * where latency is platform & configuration dependent (we assume pessimal
4111 * values here).
4112 *
4113 * The SR calculation is:
4114 * watermark = (trunc(latency/line time)+1) * surface width *
4115 * bytes per pixel
4116 * where
4117 * line time = htotal / dotclock
4118 * surface width = hdisplay for normal plane and 64 for cursor
4119 * and latency is assumed to be high, as above.
4120 *
4121 * The final value programmed to the register should always be rounded up,
4122 * and include an extra 2 entries to account for clock crossings.
4123 *
4124 * We don't use the sprite, so we can ignore that. And on Crestline we have
4125 * to set the non-SR watermarks to 8.
4126 */
4127 void intel_update_watermarks(struct drm_crtc *crtc)
4128 {
4129 struct drm_i915_private *dev_priv = crtc->dev->dev_private;
4130
4131 if (dev_priv->display.update_wm)
4132 dev_priv->display.update_wm(crtc);
4133 }
4134
4135 void intel_update_sprite_watermarks(struct drm_plane *plane,
4136 struct drm_crtc *crtc,
4137 uint32_t sprite_width,
4138 uint32_t sprite_height,
4139 int pixel_size,
4140 bool enabled, bool scaled)
4141 {
4142 struct drm_i915_private *dev_priv = plane->dev->dev_private;
4143
4144 if (dev_priv->display.update_sprite_wm)
4145 dev_priv->display.update_sprite_wm(plane, crtc,
4146 sprite_width, sprite_height,
4147 pixel_size, enabled, scaled);
4148 }
4149
4150 /**
4151 * Lock protecting IPS related data structures
4152 */
4153 DEFINE_SPINLOCK(mchdev_lock);
4154
4155 /* Global for IPS driver to get at the current i915 device. Protected by
4156 * mchdev_lock. */
4157 static struct drm_i915_private *i915_mch_dev;
4158
4159 bool ironlake_set_drps(struct drm_device *dev, u8 val)
4160 {
4161 struct drm_i915_private *dev_priv = dev->dev_private;
4162 u16 rgvswctl;
4163
4164 assert_spin_locked(&mchdev_lock);
4165
4166 rgvswctl = I915_READ16(MEMSWCTL);
4167 if (rgvswctl & MEMCTL_CMD_STS) {
4168 DRM_DEBUG("gpu busy, RCS change rejected\n");
4169 return false; /* still busy with another command */
4170 }
4171
4172 rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
4173 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
4174 I915_WRITE16(MEMSWCTL, rgvswctl);
4175 POSTING_READ16(MEMSWCTL);
4176
4177 rgvswctl |= MEMCTL_CMD_STS;
4178 I915_WRITE16(MEMSWCTL, rgvswctl);
4179
4180 return true;
4181 }
4182
4183 static void ironlake_enable_drps(struct drm_device *dev)
4184 {
4185 struct drm_i915_private *dev_priv = dev->dev_private;
4186 u32 rgvmodectl = I915_READ(MEMMODECTL);
4187 u8 fmax, fmin, fstart, vstart;
4188
4189 spin_lock_irq(&mchdev_lock);
4190
4191 /* Enable temp reporting */
4192 I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
4193 I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
4194
4195 /* 100ms RC evaluation intervals */
4196 I915_WRITE(RCUPEI, 100000);
4197 I915_WRITE(RCDNEI, 100000);
4198
4199 /* Set max/min thresholds to 90ms and 80ms respectively */
4200 I915_WRITE(RCBMAXAVG, 90000);
4201 I915_WRITE(RCBMINAVG, 80000);
4202
4203 I915_WRITE(MEMIHYST, 1);
4204
4205 /* Set up min, max, and cur for interrupt handling */
4206 fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
4207 fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
4208 fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
4209 MEMMODE_FSTART_SHIFT;
4210
4211 vstart = (I915_READ(PXVFREQ(fstart)) & PXVFREQ_PX_MASK) >>
4212 PXVFREQ_PX_SHIFT;
4213
4214 dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
4215 dev_priv->ips.fstart = fstart;
4216
4217 dev_priv->ips.max_delay = fstart;
4218 dev_priv->ips.min_delay = fmin;
4219 dev_priv->ips.cur_delay = fstart;
4220
4221 DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
4222 fmax, fmin, fstart);
4223
4224 I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
4225
4226 /*
4227 * Interrupts will be enabled in ironlake_irq_postinstall
4228 */
4229
4230 I915_WRITE(VIDSTART, vstart);
4231 POSTING_READ(VIDSTART);
4232
4233 rgvmodectl |= MEMMODE_SWMODE_EN;
4234 I915_WRITE(MEMMODECTL, rgvmodectl);
4235
4236 if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
4237 DRM_ERROR("stuck trying to change perf mode\n");
4238 mdelay(1);
4239
4240 ironlake_set_drps(dev, fstart);
4241
4242 dev_priv->ips.last_count1 = I915_READ(DMIEC) +
4243 I915_READ(DDREC) + I915_READ(CSIEC);
4244 dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
4245 dev_priv->ips.last_count2 = I915_READ(GFXEC);
4246 dev_priv->ips.last_time2 = ktime_get_raw_ns();
4247
4248 spin_unlock_irq(&mchdev_lock);
4249 }
4250
4251 static void ironlake_disable_drps(struct drm_device *dev)
4252 {
4253 struct drm_i915_private *dev_priv = dev->dev_private;
4254 u16 rgvswctl;
4255
4256 spin_lock_irq(&mchdev_lock);
4257
4258 rgvswctl = I915_READ16(MEMSWCTL);
4259
4260 /* Ack interrupts, disable EFC interrupt */
4261 I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
4262 I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
4263 I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
4264 I915_WRITE(DEIIR, DE_PCU_EVENT);
4265 I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
4266
4267 /* Go back to the starting frequency */
4268 ironlake_set_drps(dev, dev_priv->ips.fstart);
4269 mdelay(1);
4270 rgvswctl |= MEMCTL_CMD_STS;
4271 I915_WRITE(MEMSWCTL, rgvswctl);
4272 mdelay(1);
4273
4274 spin_unlock_irq(&mchdev_lock);
4275 }
4276
4277 /* There's a funny hw issue where the hw returns all 0 when reading from
4278 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
4279 * ourselves, instead of doing a rmw cycle (which might result in us clearing
4280 * all limits and the gpu stuck at whatever frequency it is at atm).
4281 */
4282 static u32 intel_rps_limits(struct drm_i915_private *dev_priv, u8 val)
4283 {
4284 u32 limits;
4285
4286 /* Only set the down limit when we've reached the lowest level to avoid
4287 * getting more interrupts, otherwise leave this clear. This prevents a
4288 * race in the hw when coming out of rc6: There's a tiny window where
4289 * the hw runs at the minimal clock before selecting the desired
4290 * frequency, if the down threshold expires in that window we will not
4291 * receive a down interrupt. */
4292 if (IS_GEN9(dev_priv->dev)) {
4293 limits = (dev_priv->rps.max_freq_softlimit) << 23;
4294 if (val <= dev_priv->rps.min_freq_softlimit)
4295 limits |= (dev_priv->rps.min_freq_softlimit) << 14;
4296 } else {
4297 limits = dev_priv->rps.max_freq_softlimit << 24;
4298 if (val <= dev_priv->rps.min_freq_softlimit)
4299 limits |= dev_priv->rps.min_freq_softlimit << 16;
4300 }
4301
4302 return limits;
4303 }
4304
4305 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
4306 {
4307 int new_power;
4308 u32 threshold_up = 0, threshold_down = 0; /* in % */
4309 u32 ei_up = 0, ei_down = 0;
4310
4311 new_power = dev_priv->rps.power;
4312 switch (dev_priv->rps.power) {
4313 case LOW_POWER:
4314 if (val > dev_priv->rps.efficient_freq + 1 && val > dev_priv->rps.cur_freq)
4315 new_power = BETWEEN;
4316 break;
4317
4318 case BETWEEN:
4319 if (val <= dev_priv->rps.efficient_freq && val < dev_priv->rps.cur_freq)
4320 new_power = LOW_POWER;
4321 else if (val >= dev_priv->rps.rp0_freq && val > dev_priv->rps.cur_freq)
4322 new_power = HIGH_POWER;
4323 break;
4324
4325 case HIGH_POWER:
4326 if (val < (dev_priv->rps.rp1_freq + dev_priv->rps.rp0_freq) >> 1 && val < dev_priv->rps.cur_freq)
4327 new_power = BETWEEN;
4328 break;
4329 }
4330 /* Max/min bins are special */
4331 if (val <= dev_priv->rps.min_freq_softlimit)
4332 new_power = LOW_POWER;
4333 if (val >= dev_priv->rps.max_freq_softlimit)
4334 new_power = HIGH_POWER;
4335 if (new_power == dev_priv->rps.power)
4336 return;
4337
4338 /* Note the units here are not exactly 1us, but 1280ns. */
4339 switch (new_power) {
4340 case LOW_POWER:
4341 /* Upclock if more than 95% busy over 16ms */
4342 ei_up = 16000;
4343 threshold_up = 95;
4344
4345 /* Downclock if less than 85% busy over 32ms */
4346 ei_down = 32000;
4347 threshold_down = 85;
4348 break;
4349
4350 case BETWEEN:
4351 /* Upclock if more than 90% busy over 13ms */
4352 ei_up = 13000;
4353 threshold_up = 90;
4354
4355 /* Downclock if less than 75% busy over 32ms */
4356 ei_down = 32000;
4357 threshold_down = 75;
4358 break;
4359
4360 case HIGH_POWER:
4361 /* Upclock if more than 85% busy over 10ms */
4362 ei_up = 10000;
4363 threshold_up = 85;
4364
4365 /* Downclock if less than 60% busy over 32ms */
4366 ei_down = 32000;
4367 threshold_down = 60;
4368 break;
4369 }
4370
4371 I915_WRITE(GEN6_RP_UP_EI,
4372 GT_INTERVAL_FROM_US(dev_priv, ei_up));
4373 I915_WRITE(GEN6_RP_UP_THRESHOLD,
4374 GT_INTERVAL_FROM_US(dev_priv, (ei_up * threshold_up / 100)));
4375
4376 I915_WRITE(GEN6_RP_DOWN_EI,
4377 GT_INTERVAL_FROM_US(dev_priv, ei_down));
4378 I915_WRITE(GEN6_RP_DOWN_THRESHOLD,
4379 GT_INTERVAL_FROM_US(dev_priv, (ei_down * threshold_down / 100)));
4380
4381 I915_WRITE(GEN6_RP_CONTROL,
4382 GEN6_RP_MEDIA_TURBO |
4383 GEN6_RP_MEDIA_HW_NORMAL_MODE |
4384 GEN6_RP_MEDIA_IS_GFX |
4385 GEN6_RP_ENABLE |
4386 GEN6_RP_UP_BUSY_AVG |
4387 GEN6_RP_DOWN_IDLE_AVG);
4388
4389 dev_priv->rps.power = new_power;
4390 dev_priv->rps.up_threshold = threshold_up;
4391 dev_priv->rps.down_threshold = threshold_down;
4392 dev_priv->rps.last_adj = 0;
4393 }
4394
4395 static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
4396 {
4397 u32 mask = 0;
4398
4399 if (val > dev_priv->rps.min_freq_softlimit)
4400 mask |= GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
4401 if (val < dev_priv->rps.max_freq_softlimit)
4402 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD;
4403
4404 mask &= dev_priv->pm_rps_events;
4405
4406 return gen6_sanitize_rps_pm_mask(dev_priv, ~mask);
4407 }
4408
4409 /* gen6_set_rps is called to update the frequency request, but should also be
4410 * called when the range (min_delay and max_delay) is modified so that we can
4411 * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
4412 static void gen6_set_rps(struct drm_device *dev, u8 val)
4413 {
4414 struct drm_i915_private *dev_priv = dev->dev_private;
4415
4416 /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
4417 if (IS_BROXTON(dev) && (INTEL_REVID(dev) < BXT_REVID_B0))
4418 return;
4419
4420 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4421 WARN_ON(val > dev_priv->rps.max_freq);
4422 WARN_ON(val < dev_priv->rps.min_freq);
4423
4424 /* min/max delay may still have been modified so be sure to
4425 * write the limits value.
4426 */
4427 if (val != dev_priv->rps.cur_freq) {
4428 gen6_set_rps_thresholds(dev_priv, val);
4429
4430 if (IS_GEN9(dev))
4431 I915_WRITE(GEN6_RPNSWREQ,
4432 GEN9_FREQUENCY(val));
4433 else if (IS_HASWELL(dev) || IS_BROADWELL(dev))
4434 I915_WRITE(GEN6_RPNSWREQ,
4435 HSW_FREQUENCY(val));
4436 else
4437 I915_WRITE(GEN6_RPNSWREQ,
4438 GEN6_FREQUENCY(val) |
4439 GEN6_OFFSET(0) |
4440 GEN6_AGGRESSIVE_TURBO);
4441 }
4442
4443 /* Make sure we continue to get interrupts
4444 * until we hit the minimum or maximum frequencies.
4445 */
4446 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, intel_rps_limits(dev_priv, val));
4447 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
4448
4449 POSTING_READ(GEN6_RPNSWREQ);
4450
4451 dev_priv->rps.cur_freq = val;
4452 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
4453 }
4454
4455 static void valleyview_set_rps(struct drm_device *dev, u8 val)
4456 {
4457 struct drm_i915_private *dev_priv = dev->dev_private;
4458
4459 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4460 WARN_ON(val > dev_priv->rps.max_freq);
4461 WARN_ON(val < dev_priv->rps.min_freq);
4462
4463 if (WARN_ONCE(IS_CHERRYVIEW(dev) && (val & 1),
4464 "Odd GPU freq value\n"))
4465 val &= ~1;
4466
4467 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
4468
4469 if (val != dev_priv->rps.cur_freq) {
4470 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
4471 if (!IS_CHERRYVIEW(dev_priv))
4472 gen6_set_rps_thresholds(dev_priv, val);
4473 }
4474
4475 dev_priv->rps.cur_freq = val;
4476 trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
4477 }
4478
4479 /* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
4480 *
4481 * * If Gfx is Idle, then
4482 * 1. Forcewake Media well.
4483 * 2. Request idle freq.
4484 * 3. Release Forcewake of Media well.
4485 */
4486 static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
4487 {
4488 u32 val = dev_priv->rps.idle_freq;
4489
4490 if (dev_priv->rps.cur_freq <= val)
4491 return;
4492
4493 /* Wake up the media well, as that takes a lot less
4494 * power than the Render well. */
4495 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_MEDIA);
4496 valleyview_set_rps(dev_priv->dev, val);
4497 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_MEDIA);
4498 }
4499
4500 void gen6_rps_busy(struct drm_i915_private *dev_priv)
4501 {
4502 mutex_lock(&dev_priv->rps.hw_lock);
4503 if (dev_priv->rps.enabled) {
4504 if (dev_priv->pm_rps_events & (GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_UP_EI_EXPIRED))
4505 gen6_rps_reset_ei(dev_priv);
4506 I915_WRITE(GEN6_PMINTRMSK,
4507 gen6_rps_pm_mask(dev_priv, dev_priv->rps.cur_freq));
4508 }
4509 mutex_unlock(&dev_priv->rps.hw_lock);
4510 }
4511
4512 void gen6_rps_idle(struct drm_i915_private *dev_priv)
4513 {
4514 struct drm_device *dev = dev_priv->dev;
4515
4516 mutex_lock(&dev_priv->rps.hw_lock);
4517 if (dev_priv->rps.enabled) {
4518 if (IS_VALLEYVIEW(dev))
4519 vlv_set_rps_idle(dev_priv);
4520 else
4521 gen6_set_rps(dev_priv->dev, dev_priv->rps.idle_freq);
4522 dev_priv->rps.last_adj = 0;
4523 I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
4524 }
4525 mutex_unlock(&dev_priv->rps.hw_lock);
4526
4527 spin_lock(&dev_priv->rps.client_lock);
4528 while (!list_empty(&dev_priv->rps.clients))
4529 list_del_init(dev_priv->rps.clients.next);
4530 spin_unlock(&dev_priv->rps.client_lock);
4531 }
4532
4533 void gen6_rps_boost(struct drm_i915_private *dev_priv,
4534 struct intel_rps_client *rps,
4535 unsigned long submitted)
4536 {
4537 /* This is intentionally racy! We peek at the state here, then
4538 * validate inside the RPS worker.
4539 */
4540 if (!(dev_priv->mm.busy &&
4541 dev_priv->rps.enabled &&
4542 dev_priv->rps.cur_freq < dev_priv->rps.max_freq_softlimit))
4543 return;
4544
4545 /* Force a RPS boost (and don't count it against the client) if
4546 * the GPU is severely congested.
4547 */
4548 if (rps && time_after(jiffies, submitted + DRM_I915_THROTTLE_JIFFIES))
4549 rps = NULL;
4550
4551 spin_lock(&dev_priv->rps.client_lock);
4552 if (rps == NULL || list_empty(&rps->link)) {
4553 spin_lock_irq(&dev_priv->irq_lock);
4554 if (dev_priv->rps.interrupts_enabled) {
4555 dev_priv->rps.client_boost = true;
4556 queue_work(dev_priv->wq, &dev_priv->rps.work);
4557 }
4558 spin_unlock_irq(&dev_priv->irq_lock);
4559
4560 if (rps != NULL) {
4561 list_add(&rps->link, &dev_priv->rps.clients);
4562 rps->boosts++;
4563 } else
4564 dev_priv->rps.boosts++;
4565 }
4566 spin_unlock(&dev_priv->rps.client_lock);
4567 }
4568
4569 void intel_set_rps(struct drm_device *dev, u8 val)
4570 {
4571 if (IS_VALLEYVIEW(dev))
4572 valleyview_set_rps(dev, val);
4573 else
4574 gen6_set_rps(dev, val);
4575 }
4576
4577 static void gen9_disable_rps(struct drm_device *dev)
4578 {
4579 struct drm_i915_private *dev_priv = dev->dev_private;
4580
4581 I915_WRITE(GEN6_RC_CONTROL, 0);
4582 I915_WRITE(GEN9_PG_ENABLE, 0);
4583 }
4584
4585 static void gen6_disable_rps(struct drm_device *dev)
4586 {
4587 struct drm_i915_private *dev_priv = dev->dev_private;
4588
4589 I915_WRITE(GEN6_RC_CONTROL, 0);
4590 I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
4591 }
4592
4593 static void cherryview_disable_rps(struct drm_device *dev)
4594 {
4595 struct drm_i915_private *dev_priv = dev->dev_private;
4596
4597 I915_WRITE(GEN6_RC_CONTROL, 0);
4598 }
4599
4600 static void valleyview_disable_rps(struct drm_device *dev)
4601 {
4602 struct drm_i915_private *dev_priv = dev->dev_private;
4603
4604 /* we're doing forcewake before Disabling RC6,
4605 * This what the BIOS expects when going into suspend */
4606 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4607
4608 I915_WRITE(GEN6_RC_CONTROL, 0);
4609
4610 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4611 }
4612
4613 static void intel_print_rc6_info(struct drm_device *dev, u32 mode)
4614 {
4615 if (IS_VALLEYVIEW(dev)) {
4616 if (mode & (GEN7_RC_CTL_TO_MODE | GEN6_RC_CTL_EI_MODE(1)))
4617 mode = GEN6_RC_CTL_RC6_ENABLE;
4618 else
4619 mode = 0;
4620 }
4621 if (HAS_RC6p(dev))
4622 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s RC6p %s RC6pp %s\n",
4623 (mode & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off",
4624 (mode & GEN6_RC_CTL_RC6p_ENABLE) ? "on" : "off",
4625 (mode & GEN6_RC_CTL_RC6pp_ENABLE) ? "on" : "off");
4626
4627 else
4628 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s\n",
4629 (mode & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off");
4630 }
4631
4632 static int sanitize_rc6_option(const struct drm_device *dev, int enable_rc6)
4633 {
4634 /* No RC6 before Ironlake and code is gone for ilk. */
4635 if (INTEL_INFO(dev)->gen < 6)
4636 return 0;
4637
4638 /* Respect the kernel parameter if it is set */
4639 if (enable_rc6 >= 0) {
4640 int mask;
4641
4642 if (HAS_RC6p(dev))
4643 mask = INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE |
4644 INTEL_RC6pp_ENABLE;
4645 else
4646 mask = INTEL_RC6_ENABLE;
4647
4648 if ((enable_rc6 & mask) != enable_rc6)
4649 DRM_DEBUG_KMS("Adjusting RC6 mask to %d (requested %d, valid %d)\n",
4650 enable_rc6 & mask, enable_rc6, mask);
4651
4652 return enable_rc6 & mask;
4653 }
4654
4655 if (IS_IVYBRIDGE(dev))
4656 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
4657
4658 return INTEL_RC6_ENABLE;
4659 }
4660
4661 int intel_enable_rc6(const struct drm_device *dev)
4662 {
4663 return i915.enable_rc6;
4664 }
4665
4666 static void gen6_init_rps_frequencies(struct drm_device *dev)
4667 {
4668 struct drm_i915_private *dev_priv = dev->dev_private;
4669 uint32_t rp_state_cap;
4670 u32 ddcc_status = 0;
4671 int ret;
4672
4673 /* All of these values are in units of 50MHz */
4674 dev_priv->rps.cur_freq = 0;
4675 /* static values from HW: RP0 > RP1 > RPn (min_freq) */
4676 if (IS_BROXTON(dev)) {
4677 rp_state_cap = I915_READ(BXT_RP_STATE_CAP);
4678 dev_priv->rps.rp0_freq = (rp_state_cap >> 16) & 0xff;
4679 dev_priv->rps.rp1_freq = (rp_state_cap >> 8) & 0xff;
4680 dev_priv->rps.min_freq = (rp_state_cap >> 0) & 0xff;
4681 } else {
4682 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
4683 dev_priv->rps.rp0_freq = (rp_state_cap >> 0) & 0xff;
4684 dev_priv->rps.rp1_freq = (rp_state_cap >> 8) & 0xff;
4685 dev_priv->rps.min_freq = (rp_state_cap >> 16) & 0xff;
4686 }
4687
4688 /* hw_max = RP0 until we check for overclocking */
4689 dev_priv->rps.max_freq = dev_priv->rps.rp0_freq;
4690
4691 dev_priv->rps.efficient_freq = dev_priv->rps.rp1_freq;
4692 if (IS_HASWELL(dev) || IS_BROADWELL(dev) || IS_SKYLAKE(dev)) {
4693 ret = sandybridge_pcode_read(dev_priv,
4694 HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL,
4695 &ddcc_status);
4696 if (0 == ret)
4697 dev_priv->rps.efficient_freq =
4698 clamp_t(u8,
4699 ((ddcc_status >> 8) & 0xff),
4700 dev_priv->rps.min_freq,
4701 dev_priv->rps.max_freq);
4702 }
4703
4704 if (IS_SKYLAKE(dev)) {
4705 /* Store the frequency values in 16.66 MHZ units, which is
4706 the natural hardware unit for SKL */
4707 dev_priv->rps.rp0_freq *= GEN9_FREQ_SCALER;
4708 dev_priv->rps.rp1_freq *= GEN9_FREQ_SCALER;
4709 dev_priv->rps.min_freq *= GEN9_FREQ_SCALER;
4710 dev_priv->rps.max_freq *= GEN9_FREQ_SCALER;
4711 dev_priv->rps.efficient_freq *= GEN9_FREQ_SCALER;
4712 }
4713
4714 dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
4715
4716 /* Preserve min/max settings in case of re-init */
4717 if (dev_priv->rps.max_freq_softlimit == 0)
4718 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
4719
4720 if (dev_priv->rps.min_freq_softlimit == 0) {
4721 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
4722 dev_priv->rps.min_freq_softlimit =
4723 max_t(int, dev_priv->rps.efficient_freq,
4724 intel_freq_opcode(dev_priv, 450));
4725 else
4726 dev_priv->rps.min_freq_softlimit =
4727 dev_priv->rps.min_freq;
4728 }
4729 }
4730
4731 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
4732 static void gen9_enable_rps(struct drm_device *dev)
4733 {
4734 struct drm_i915_private *dev_priv = dev->dev_private;
4735
4736 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4737
4738 gen6_init_rps_frequencies(dev);
4739
4740 /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
4741 if (IS_BROXTON(dev) && (INTEL_REVID(dev) < BXT_REVID_B0)) {
4742 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4743 return;
4744 }
4745
4746 /* Program defaults and thresholds for RPS*/
4747 I915_WRITE(GEN6_RC_VIDEO_FREQ,
4748 GEN9_FREQUENCY(dev_priv->rps.rp1_freq));
4749
4750 /* 1 second timeout*/
4751 I915_WRITE(GEN6_RP_DOWN_TIMEOUT,
4752 GT_INTERVAL_FROM_US(dev_priv, 1000000));
4753
4754 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 0xa);
4755
4756 /* Leaning on the below call to gen6_set_rps to program/setup the
4757 * Up/Down EI & threshold registers, as well as the RP_CONTROL,
4758 * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
4759 dev_priv->rps.power = HIGH_POWER; /* force a reset */
4760 gen6_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
4761
4762 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4763 }
4764
4765 static void gen9_enable_rc6(struct drm_device *dev)
4766 {
4767 struct drm_i915_private *dev_priv = dev->dev_private;
4768 struct intel_engine_cs *ring;
4769 uint32_t rc6_mask = 0;
4770 int unused;
4771
4772 /* 1a: Software RC state - RC0 */
4773 I915_WRITE(GEN6_RC_STATE, 0);
4774
4775 /* 1b: Get forcewake during program sequence. Although the driver
4776 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4777 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4778
4779 /* 2a: Disable RC states. */
4780 I915_WRITE(GEN6_RC_CONTROL, 0);
4781
4782 /* 2b: Program RC6 thresholds.*/
4783
4784 /* WaRsDoubleRc6WrlWithCoarsePowerGating: Doubling WRL only when CPG is enabled */
4785 if (IS_SKYLAKE(dev))
4786 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 108 << 16);
4787 else
4788 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16);
4789 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
4790 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
4791 for_each_ring(ring, dev_priv, unused)
4792 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4793
4794 if (HAS_GUC_UCODE(dev))
4795 I915_WRITE(GUC_MAX_IDLE_COUNT, 0xA);
4796
4797 I915_WRITE(GEN6_RC_SLEEP, 0);
4798
4799 /* 2c: Program Coarse Power Gating Policies. */
4800 I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS, 25);
4801 I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS, 25);
4802
4803 /* 3a: Enable RC6 */
4804 if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
4805 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
4806 DRM_INFO("RC6 %s\n", (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ?
4807 "on" : "off");
4808 /* WaRsUseTimeoutMode */
4809 if ((IS_SKYLAKE(dev) && INTEL_REVID(dev) <= SKL_REVID_D0) ||
4810 (IS_BROXTON(dev) && INTEL_REVID(dev) <= BXT_REVID_A0)) {
4811 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us */
4812 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4813 GEN7_RC_CTL_TO_MODE |
4814 rc6_mask);
4815 } else {
4816 I915_WRITE(GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */
4817 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4818 GEN6_RC_CTL_EI_MODE(1) |
4819 rc6_mask);
4820 }
4821
4822 /*
4823 * 3b: Enable Coarse Power Gating only when RC6 is enabled.
4824 * WaRsDisableCoarsePowerGating:skl,bxt - Render/Media PG need to be disabled with RC6.
4825 */
4826 if ((IS_BROXTON(dev) && (INTEL_REVID(dev) < BXT_REVID_B0)) ||
4827 ((IS_SKL_GT3(dev) || IS_SKL_GT4(dev)) && (INTEL_REVID(dev) <= SKL_REVID_F0)))
4828 I915_WRITE(GEN9_PG_ENABLE, 0);
4829 else
4830 I915_WRITE(GEN9_PG_ENABLE, (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ?
4831 (GEN9_RENDER_PG_ENABLE | GEN9_MEDIA_PG_ENABLE) : 0);
4832
4833 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4834
4835 }
4836
4837 static void gen8_enable_rps(struct drm_device *dev)
4838 {
4839 struct drm_i915_private *dev_priv = dev->dev_private;
4840 struct intel_engine_cs *ring;
4841 uint32_t rc6_mask = 0;
4842 int unused;
4843
4844 /* 1a: Software RC state - RC0 */
4845 I915_WRITE(GEN6_RC_STATE, 0);
4846
4847 /* 1c & 1d: Get forcewake during program sequence. Although the driver
4848 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4849 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4850
4851 /* 2a: Disable RC states. */
4852 I915_WRITE(GEN6_RC_CONTROL, 0);
4853
4854 /* Initialize rps frequencies */
4855 gen6_init_rps_frequencies(dev);
4856
4857 /* 2b: Program RC6 thresholds.*/
4858 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
4859 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
4860 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
4861 for_each_ring(ring, dev_priv, unused)
4862 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4863 I915_WRITE(GEN6_RC_SLEEP, 0);
4864 if (IS_BROADWELL(dev))
4865 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */
4866 else
4867 I915_WRITE(GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
4868
4869 /* 3: Enable RC6 */
4870 if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
4871 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
4872 intel_print_rc6_info(dev, rc6_mask);
4873 if (IS_BROADWELL(dev))
4874 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4875 GEN7_RC_CTL_TO_MODE |
4876 rc6_mask);
4877 else
4878 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4879 GEN6_RC_CTL_EI_MODE(1) |
4880 rc6_mask);
4881
4882 /* 4 Program defaults and thresholds for RPS*/
4883 I915_WRITE(GEN6_RPNSWREQ,
4884 HSW_FREQUENCY(dev_priv->rps.rp1_freq));
4885 I915_WRITE(GEN6_RC_VIDEO_FREQ,
4886 HSW_FREQUENCY(dev_priv->rps.rp1_freq));
4887 /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
4888 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
4889
4890 /* Docs recommend 900MHz, and 300 MHz respectively */
4891 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
4892 dev_priv->rps.max_freq_softlimit << 24 |
4893 dev_priv->rps.min_freq_softlimit << 16);
4894
4895 I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
4896 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
4897 I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
4898 I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
4899
4900 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
4901
4902 /* 5: Enable RPS */
4903 I915_WRITE(GEN6_RP_CONTROL,
4904 GEN6_RP_MEDIA_TURBO |
4905 GEN6_RP_MEDIA_HW_NORMAL_MODE |
4906 GEN6_RP_MEDIA_IS_GFX |
4907 GEN6_RP_ENABLE |
4908 GEN6_RP_UP_BUSY_AVG |
4909 GEN6_RP_DOWN_IDLE_AVG);
4910
4911 /* 6: Ring frequency + overclocking (our driver does this later */
4912
4913 dev_priv->rps.power = HIGH_POWER; /* force a reset */
4914 gen6_set_rps(dev_priv->dev, dev_priv->rps.idle_freq);
4915
4916 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4917 }
4918
4919 static void gen6_enable_rps(struct drm_device *dev)
4920 {
4921 struct drm_i915_private *dev_priv = dev->dev_private;
4922 struct intel_engine_cs *ring;
4923 u32 rc6vids, pcu_mbox = 0, rc6_mask = 0;
4924 u32 gtfifodbg;
4925 int rc6_mode;
4926 int i, ret;
4927
4928 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4929
4930 /* Here begins a magic sequence of register writes to enable
4931 * auto-downclocking.
4932 *
4933 * Perhaps there might be some value in exposing these to
4934 * userspace...
4935 */
4936 I915_WRITE(GEN6_RC_STATE, 0);
4937
4938 /* Clear the DBG now so we don't confuse earlier errors */
4939 if ((gtfifodbg = I915_READ(GTFIFODBG))) {
4940 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
4941 I915_WRITE(GTFIFODBG, gtfifodbg);
4942 }
4943
4944 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4945
4946 /* Initialize rps frequencies */
4947 gen6_init_rps_frequencies(dev);
4948
4949 /* disable the counters and set deterministic thresholds */
4950 I915_WRITE(GEN6_RC_CONTROL, 0);
4951
4952 I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
4953 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
4954 I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
4955 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
4956 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
4957
4958 for_each_ring(ring, dev_priv, i)
4959 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4960
4961 I915_WRITE(GEN6_RC_SLEEP, 0);
4962 I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
4963 if (IS_IVYBRIDGE(dev))
4964 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
4965 else
4966 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
4967 I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
4968 I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
4969
4970 /* Check if we are enabling RC6 */
4971 rc6_mode = intel_enable_rc6(dev_priv->dev);
4972 if (rc6_mode & INTEL_RC6_ENABLE)
4973 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
4974
4975 /* We don't use those on Haswell */
4976 if (!IS_HASWELL(dev)) {
4977 if (rc6_mode & INTEL_RC6p_ENABLE)
4978 rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
4979
4980 if (rc6_mode & INTEL_RC6pp_ENABLE)
4981 rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
4982 }
4983
4984 intel_print_rc6_info(dev, rc6_mask);
4985
4986 I915_WRITE(GEN6_RC_CONTROL,
4987 rc6_mask |
4988 GEN6_RC_CTL_EI_MODE(1) |
4989 GEN6_RC_CTL_HW_ENABLE);
4990
4991 /* Power down if completely idle for over 50ms */
4992 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
4993 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
4994
4995 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_MIN_FREQ_TABLE, 0);
4996 if (ret)
4997 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
4998
4999 ret = sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &pcu_mbox);
5000 if (!ret && (pcu_mbox & (1<<31))) { /* OC supported */
5001 DRM_DEBUG_DRIVER("Overclocking supported. Max: %dMHz, Overclock max: %dMHz\n",
5002 (dev_priv->rps.max_freq_softlimit & 0xff) * 50,
5003 (pcu_mbox & 0xff) * 50);
5004 dev_priv->rps.max_freq = pcu_mbox & 0xff;
5005 }
5006
5007 dev_priv->rps.power = HIGH_POWER; /* force a reset */
5008 gen6_set_rps(dev_priv->dev, dev_priv->rps.idle_freq);
5009
5010 rc6vids = 0;
5011 ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
5012 if (IS_GEN6(dev) && ret) {
5013 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
5014 } else if (IS_GEN6(dev) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
5015 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
5016 GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
5017 rc6vids &= 0xffff00;
5018 rc6vids |= GEN6_ENCODE_RC6_VID(450);
5019 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
5020 if (ret)
5021 DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
5022 }
5023
5024 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5025 }
5026
5027 static void __gen6_update_ring_freq(struct drm_device *dev)
5028 {
5029 struct drm_i915_private *dev_priv = dev->dev_private;
5030 int min_freq = 15;
5031 unsigned int gpu_freq;
5032 unsigned int max_ia_freq, min_ring_freq;
5033 unsigned int max_gpu_freq, min_gpu_freq;
5034 int scaling_factor = 180;
5035 struct cpufreq_policy *policy;
5036
5037 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5038
5039 policy = cpufreq_cpu_get(0);
5040 if (policy) {
5041 max_ia_freq = policy->cpuinfo.max_freq;
5042 cpufreq_cpu_put(policy);
5043 } else {
5044 /*
5045 * Default to measured freq if none found, PCU will ensure we
5046 * don't go over
5047 */
5048 max_ia_freq = tsc_khz;
5049 }
5050
5051 /* Convert from kHz to MHz */
5052 max_ia_freq /= 1000;
5053
5054 min_ring_freq = I915_READ(DCLK) & 0xf;
5055 /* convert DDR frequency from units of 266.6MHz to bandwidth */
5056 min_ring_freq = mult_frac(min_ring_freq, 8, 3);
5057
5058 if (IS_SKYLAKE(dev)) {
5059 /* Convert GT frequency to 50 HZ units */
5060 min_gpu_freq = dev_priv->rps.min_freq / GEN9_FREQ_SCALER;
5061 max_gpu_freq = dev_priv->rps.max_freq / GEN9_FREQ_SCALER;
5062 } else {
5063 min_gpu_freq = dev_priv->rps.min_freq;
5064 max_gpu_freq = dev_priv->rps.max_freq;
5065 }
5066
5067 /*
5068 * For each potential GPU frequency, load a ring frequency we'd like
5069 * to use for memory access. We do this by specifying the IA frequency
5070 * the PCU should use as a reference to determine the ring frequency.
5071 */
5072 for (gpu_freq = max_gpu_freq; gpu_freq >= min_gpu_freq; gpu_freq--) {
5073 int diff = max_gpu_freq - gpu_freq;
5074 unsigned int ia_freq = 0, ring_freq = 0;
5075
5076 if (IS_SKYLAKE(dev)) {
5077 /*
5078 * ring_freq = 2 * GT. ring_freq is in 100MHz units
5079 * No floor required for ring frequency on SKL.
5080 */
5081 ring_freq = gpu_freq;
5082 } else if (INTEL_INFO(dev)->gen >= 8) {
5083 /* max(2 * GT, DDR). NB: GT is 50MHz units */
5084 ring_freq = max(min_ring_freq, gpu_freq);
5085 } else if (IS_HASWELL(dev)) {
5086 ring_freq = mult_frac(gpu_freq, 5, 4);
5087 ring_freq = max(min_ring_freq, ring_freq);
5088 /* leave ia_freq as the default, chosen by cpufreq */
5089 } else {
5090 /* On older processors, there is no separate ring
5091 * clock domain, so in order to boost the bandwidth
5092 * of the ring, we need to upclock the CPU (ia_freq).
5093 *
5094 * For GPU frequencies less than 750MHz,
5095 * just use the lowest ring freq.
5096 */
5097 if (gpu_freq < min_freq)
5098 ia_freq = 800;
5099 else
5100 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
5101 ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
5102 }
5103
5104 sandybridge_pcode_write(dev_priv,
5105 GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
5106 ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
5107 ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
5108 gpu_freq);
5109 }
5110 }
5111
5112 void gen6_update_ring_freq(struct drm_device *dev)
5113 {
5114 struct drm_i915_private *dev_priv = dev->dev_private;
5115
5116 if (!HAS_CORE_RING_FREQ(dev))
5117 return;
5118
5119 mutex_lock(&dev_priv->rps.hw_lock);
5120 __gen6_update_ring_freq(dev);
5121 mutex_unlock(&dev_priv->rps.hw_lock);
5122 }
5123
5124 static int cherryview_rps_max_freq(struct drm_i915_private *dev_priv)
5125 {
5126 struct drm_device *dev = dev_priv->dev;
5127 u32 val, rp0;
5128
5129 val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
5130
5131 switch (INTEL_INFO(dev)->eu_total) {
5132 case 8:
5133 /* (2 * 4) config */
5134 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT);
5135 break;
5136 case 12:
5137 /* (2 * 6) config */
5138 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT);
5139 break;
5140 case 16:
5141 /* (2 * 8) config */
5142 default:
5143 /* Setting (2 * 8) Min RP0 for any other combination */
5144 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT);
5145 break;
5146 }
5147
5148 rp0 = (rp0 & FB_GFX_FREQ_FUSE_MASK);
5149
5150 return rp0;
5151 }
5152
5153 static int cherryview_rps_rpe_freq(struct drm_i915_private *dev_priv)
5154 {
5155 u32 val, rpe;
5156
5157 val = vlv_punit_read(dev_priv, PUNIT_GPU_DUTYCYCLE_REG);
5158 rpe = (val >> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK;
5159
5160 return rpe;
5161 }
5162
5163 static int cherryview_rps_guar_freq(struct drm_i915_private *dev_priv)
5164 {
5165 u32 val, rp1;
5166
5167 val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
5168 rp1 = (val & FB_GFX_FREQ_FUSE_MASK);
5169
5170 return rp1;
5171 }
5172
5173 static int valleyview_rps_guar_freq(struct drm_i915_private *dev_priv)
5174 {
5175 u32 val, rp1;
5176
5177 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
5178
5179 rp1 = (val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT;
5180
5181 return rp1;
5182 }
5183
5184 static int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
5185 {
5186 u32 val, rp0;
5187
5188 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
5189
5190 rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
5191 /* Clamp to max */
5192 rp0 = min_t(u32, rp0, 0xea);
5193
5194 return rp0;
5195 }
5196
5197 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
5198 {
5199 u32 val, rpe;
5200
5201 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
5202 rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
5203 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
5204 rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
5205
5206 return rpe;
5207 }
5208
5209 static int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
5210 {
5211 return vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
5212 }
5213
5214 /* Check that the pctx buffer wasn't move under us. */
5215 static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
5216 {
5217 unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
5218
5219 WARN_ON(pctx_addr != dev_priv->mm.stolen_base +
5220 dev_priv->vlv_pctx->stolen->start);
5221 }
5222
5223
5224 /* Check that the pcbr address is not empty. */
5225 static void cherryview_check_pctx(struct drm_i915_private *dev_priv)
5226 {
5227 unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
5228
5229 WARN_ON((pctx_addr >> VLV_PCBR_ADDR_SHIFT) == 0);
5230 }
5231
5232 static void cherryview_setup_pctx(struct drm_device *dev)
5233 {
5234 struct drm_i915_private *dev_priv = dev->dev_private;
5235 unsigned long pctx_paddr, paddr;
5236 struct i915_gtt *gtt = &dev_priv->gtt;
5237 u32 pcbr;
5238 int pctx_size = 32*1024;
5239
5240 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
5241
5242 pcbr = I915_READ(VLV_PCBR);
5243 if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) {
5244 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5245 paddr = (dev_priv->mm.stolen_base +
5246 (gtt->stolen_size - pctx_size));
5247
5248 pctx_paddr = (paddr & (~4095));
5249 I915_WRITE(VLV_PCBR, pctx_paddr);
5250 }
5251
5252 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
5253 }
5254
5255 static void valleyview_setup_pctx(struct drm_device *dev)
5256 {
5257 struct drm_i915_private *dev_priv = dev->dev_private;
5258 struct drm_i915_gem_object *pctx;
5259 unsigned long pctx_paddr;
5260 u32 pcbr;
5261 int pctx_size = 24*1024;
5262
5263 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
5264
5265 pcbr = I915_READ(VLV_PCBR);
5266 if (pcbr) {
5267 /* BIOS set it up already, grab the pre-alloc'd space */
5268 int pcbr_offset;
5269
5270 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
5271 pctx = i915_gem_object_create_stolen_for_preallocated(dev_priv->dev,
5272 pcbr_offset,
5273 I915_GTT_OFFSET_NONE,
5274 pctx_size);
5275 goto out;
5276 }
5277
5278 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5279
5280 /*
5281 * From the Gunit register HAS:
5282 * The Gfx driver is expected to program this register and ensure
5283 * proper allocation within Gfx stolen memory. For example, this
5284 * register should be programmed such than the PCBR range does not
5285 * overlap with other ranges, such as the frame buffer, protected
5286 * memory, or any other relevant ranges.
5287 */
5288 pctx = i915_gem_object_create_stolen(dev, pctx_size);
5289 if (!pctx) {
5290 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
5291 return;
5292 }
5293
5294 pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
5295 I915_WRITE(VLV_PCBR, pctx_paddr);
5296
5297 out:
5298 DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
5299 dev_priv->vlv_pctx = pctx;
5300 }
5301
5302 static void valleyview_cleanup_pctx(struct drm_device *dev)
5303 {
5304 struct drm_i915_private *dev_priv = dev->dev_private;
5305
5306 if (WARN_ON(!dev_priv->vlv_pctx))
5307 return;
5308
5309 drm_gem_object_unreference(&dev_priv->vlv_pctx->base);
5310 dev_priv->vlv_pctx = NULL;
5311 }
5312
5313 static void valleyview_init_gt_powersave(struct drm_device *dev)
5314 {
5315 struct drm_i915_private *dev_priv = dev->dev_private;
5316 u32 val;
5317
5318 valleyview_setup_pctx(dev);
5319
5320 mutex_lock(&dev_priv->rps.hw_lock);
5321
5322 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5323 switch ((val >> 6) & 3) {
5324 case 0:
5325 case 1:
5326 dev_priv->mem_freq = 800;
5327 break;
5328 case 2:
5329 dev_priv->mem_freq = 1066;
5330 break;
5331 case 3:
5332 dev_priv->mem_freq = 1333;
5333 break;
5334 }
5335 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
5336
5337 dev_priv->rps.max_freq = valleyview_rps_max_freq(dev_priv);
5338 dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
5339 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5340 intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
5341 dev_priv->rps.max_freq);
5342
5343 dev_priv->rps.efficient_freq = valleyview_rps_rpe_freq(dev_priv);
5344 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5345 intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5346 dev_priv->rps.efficient_freq);
5347
5348 dev_priv->rps.rp1_freq = valleyview_rps_guar_freq(dev_priv);
5349 DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
5350 intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
5351 dev_priv->rps.rp1_freq);
5352
5353 dev_priv->rps.min_freq = valleyview_rps_min_freq(dev_priv);
5354 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5355 intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
5356 dev_priv->rps.min_freq);
5357
5358 dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
5359
5360 /* Preserve min/max settings in case of re-init */
5361 if (dev_priv->rps.max_freq_softlimit == 0)
5362 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
5363
5364 if (dev_priv->rps.min_freq_softlimit == 0)
5365 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
5366
5367 mutex_unlock(&dev_priv->rps.hw_lock);
5368 }
5369
5370 static void cherryview_init_gt_powersave(struct drm_device *dev)
5371 {
5372 struct drm_i915_private *dev_priv = dev->dev_private;
5373 u32 val;
5374
5375 cherryview_setup_pctx(dev);
5376
5377 mutex_lock(&dev_priv->rps.hw_lock);
5378
5379 mutex_lock(&dev_priv->sb_lock);
5380 val = vlv_cck_read(dev_priv, CCK_FUSE_REG);
5381 mutex_unlock(&dev_priv->sb_lock);
5382
5383 switch ((val >> 2) & 0x7) {
5384 case 3:
5385 dev_priv->mem_freq = 2000;
5386 break;
5387 default:
5388 dev_priv->mem_freq = 1600;
5389 break;
5390 }
5391 DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
5392
5393 dev_priv->rps.max_freq = cherryview_rps_max_freq(dev_priv);
5394 dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
5395 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5396 intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
5397 dev_priv->rps.max_freq);
5398
5399 dev_priv->rps.efficient_freq = cherryview_rps_rpe_freq(dev_priv);
5400 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5401 intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5402 dev_priv->rps.efficient_freq);
5403
5404 dev_priv->rps.rp1_freq = cherryview_rps_guar_freq(dev_priv);
5405 DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
5406 intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
5407 dev_priv->rps.rp1_freq);
5408
5409 /* PUnit validated range is only [RPe, RP0] */
5410 dev_priv->rps.min_freq = dev_priv->rps.efficient_freq;
5411 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5412 intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
5413 dev_priv->rps.min_freq);
5414
5415 WARN_ONCE((dev_priv->rps.max_freq |
5416 dev_priv->rps.efficient_freq |
5417 dev_priv->rps.rp1_freq |
5418 dev_priv->rps.min_freq) & 1,
5419 "Odd GPU freq values\n");
5420
5421 dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
5422
5423 /* Preserve min/max settings in case of re-init */
5424 if (dev_priv->rps.max_freq_softlimit == 0)
5425 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
5426
5427 if (dev_priv->rps.min_freq_softlimit == 0)
5428 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
5429
5430 mutex_unlock(&dev_priv->rps.hw_lock);
5431 }
5432
5433 static void valleyview_cleanup_gt_powersave(struct drm_device *dev)
5434 {
5435 valleyview_cleanup_pctx(dev);
5436 }
5437
5438 static void cherryview_enable_rps(struct drm_device *dev)
5439 {
5440 struct drm_i915_private *dev_priv = dev->dev_private;
5441 struct intel_engine_cs *ring;
5442 u32 gtfifodbg, val, rc6_mode = 0, pcbr;
5443 int i;
5444
5445 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5446
5447 gtfifodbg = I915_READ(GTFIFODBG);
5448 if (gtfifodbg) {
5449 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5450 gtfifodbg);
5451 I915_WRITE(GTFIFODBG, gtfifodbg);
5452 }
5453
5454 cherryview_check_pctx(dev_priv);
5455
5456 /* 1a & 1b: Get forcewake during program sequence. Although the driver
5457 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5458 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5459
5460 /* Disable RC states. */
5461 I915_WRITE(GEN6_RC_CONTROL, 0);
5462
5463 /* 2a: Program RC6 thresholds.*/
5464 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
5465 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5466 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
5467
5468 for_each_ring(ring, dev_priv, i)
5469 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
5470 I915_WRITE(GEN6_RC_SLEEP, 0);
5471
5472 /* TO threshold set to 500 us ( 0x186 * 1.28 us) */
5473 I915_WRITE(GEN6_RC6_THRESHOLD, 0x186);
5474
5475 /* allows RC6 residency counter to work */
5476 I915_WRITE(VLV_COUNTER_CONTROL,
5477 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
5478 VLV_MEDIA_RC6_COUNT_EN |
5479 VLV_RENDER_RC6_COUNT_EN));
5480
5481 /* For now we assume BIOS is allocating and populating the PCBR */
5482 pcbr = I915_READ(VLV_PCBR);
5483
5484 /* 3: Enable RC6 */
5485 if ((intel_enable_rc6(dev) & INTEL_RC6_ENABLE) &&
5486 (pcbr >> VLV_PCBR_ADDR_SHIFT))
5487 rc6_mode = GEN7_RC_CTL_TO_MODE;
5488
5489 I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
5490
5491 /* 4 Program defaults and thresholds for RPS*/
5492 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
5493 I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
5494 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
5495 I915_WRITE(GEN6_RP_UP_EI, 66000);
5496 I915_WRITE(GEN6_RP_DOWN_EI, 350000);
5497
5498 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5499
5500 /* 5: Enable RPS */
5501 I915_WRITE(GEN6_RP_CONTROL,
5502 GEN6_RP_MEDIA_HW_NORMAL_MODE |
5503 GEN6_RP_MEDIA_IS_GFX |
5504 GEN6_RP_ENABLE |
5505 GEN6_RP_UP_BUSY_AVG |
5506 GEN6_RP_DOWN_IDLE_AVG);
5507
5508 /* Setting Fixed Bias */
5509 val = VLV_OVERRIDE_EN |
5510 VLV_SOC_TDP_EN |
5511 CHV_BIAS_CPU_50_SOC_50;
5512 vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
5513
5514 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5515
5516 /* RPS code assumes GPLL is used */
5517 WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
5518
5519 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
5520 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
5521
5522 dev_priv->rps.cur_freq = (val >> 8) & 0xff;
5523 DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
5524 intel_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
5525 dev_priv->rps.cur_freq);
5526
5527 DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
5528 intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5529 dev_priv->rps.efficient_freq);
5530
5531 valleyview_set_rps(dev_priv->dev, dev_priv->rps.efficient_freq);
5532
5533 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5534 }
5535
5536 static void valleyview_enable_rps(struct drm_device *dev)
5537 {
5538 struct drm_i915_private *dev_priv = dev->dev_private;
5539 struct intel_engine_cs *ring;
5540 u32 gtfifodbg, val, rc6_mode = 0;
5541 int i;
5542
5543 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5544
5545 valleyview_check_pctx(dev_priv);
5546
5547 if ((gtfifodbg = I915_READ(GTFIFODBG))) {
5548 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5549 gtfifodbg);
5550 I915_WRITE(GTFIFODBG, gtfifodbg);
5551 }
5552
5553 /* If VLV, Forcewake all wells, else re-direct to regular path */
5554 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5555
5556 /* Disable RC states. */
5557 I915_WRITE(GEN6_RC_CONTROL, 0);
5558
5559 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
5560 I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
5561 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
5562 I915_WRITE(GEN6_RP_UP_EI, 66000);
5563 I915_WRITE(GEN6_RP_DOWN_EI, 350000);
5564
5565 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5566
5567 I915_WRITE(GEN6_RP_CONTROL,
5568 GEN6_RP_MEDIA_TURBO |
5569 GEN6_RP_MEDIA_HW_NORMAL_MODE |
5570 GEN6_RP_MEDIA_IS_GFX |
5571 GEN6_RP_ENABLE |
5572 GEN6_RP_UP_BUSY_AVG |
5573 GEN6_RP_DOWN_IDLE_CONT);
5574
5575 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
5576 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
5577 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
5578
5579 for_each_ring(ring, dev_priv, i)
5580 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
5581
5582 I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
5583
5584 /* allows RC6 residency counter to work */
5585 I915_WRITE(VLV_COUNTER_CONTROL,
5586 _MASKED_BIT_ENABLE(VLV_MEDIA_RC0_COUNT_EN |
5587 VLV_RENDER_RC0_COUNT_EN |
5588 VLV_MEDIA_RC6_COUNT_EN |
5589 VLV_RENDER_RC6_COUNT_EN));
5590
5591 if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
5592 rc6_mode = GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
5593
5594 intel_print_rc6_info(dev, rc6_mode);
5595
5596 I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
5597
5598 /* Setting Fixed Bias */
5599 val = VLV_OVERRIDE_EN |
5600 VLV_SOC_TDP_EN |
5601 VLV_BIAS_CPU_125_SOC_875;
5602 vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
5603
5604 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5605
5606 /* RPS code assumes GPLL is used */
5607 WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
5608
5609 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
5610 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
5611
5612 dev_priv->rps.cur_freq = (val >> 8) & 0xff;
5613 DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
5614 intel_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
5615 dev_priv->rps.cur_freq);
5616
5617 DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
5618 intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5619 dev_priv->rps.efficient_freq);
5620
5621 valleyview_set_rps(dev_priv->dev, dev_priv->rps.efficient_freq);
5622
5623 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5624 }
5625
5626 static unsigned long intel_pxfreq(u32 vidfreq)
5627 {
5628 unsigned long freq;
5629 int div = (vidfreq & 0x3f0000) >> 16;
5630 int post = (vidfreq & 0x3000) >> 12;
5631 int pre = (vidfreq & 0x7);
5632
5633 if (!pre)
5634 return 0;
5635
5636 freq = ((div * 133333) / ((1<<post) * pre));
5637
5638 return freq;
5639 }
5640
5641 static const struct cparams {
5642 u16 i;
5643 u16 t;
5644 u16 m;
5645 u16 c;
5646 } cparams[] = {
5647 { 1, 1333, 301, 28664 },
5648 { 1, 1066, 294, 24460 },
5649 { 1, 800, 294, 25192 },
5650 { 0, 1333, 276, 27605 },
5651 { 0, 1066, 276, 27605 },
5652 { 0, 800, 231, 23784 },
5653 };
5654
5655 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
5656 {
5657 u64 total_count, diff, ret;
5658 u32 count1, count2, count3, m = 0, c = 0;
5659 unsigned long now = jiffies_to_msecs(jiffies), diff1;
5660 int i;
5661
5662 assert_spin_locked(&mchdev_lock);
5663
5664 diff1 = now - dev_priv->ips.last_time1;
5665
5666 /* Prevent division-by-zero if we are asking too fast.
5667 * Also, we don't get interesting results if we are polling
5668 * faster than once in 10ms, so just return the saved value
5669 * in such cases.
5670 */
5671 if (diff1 <= 10)
5672 return dev_priv->ips.chipset_power;
5673
5674 count1 = I915_READ(DMIEC);
5675 count2 = I915_READ(DDREC);
5676 count3 = I915_READ(CSIEC);
5677
5678 total_count = count1 + count2 + count3;
5679
5680 /* FIXME: handle per-counter overflow */
5681 if (total_count < dev_priv->ips.last_count1) {
5682 diff = ~0UL - dev_priv->ips.last_count1;
5683 diff += total_count;
5684 } else {
5685 diff = total_count - dev_priv->ips.last_count1;
5686 }
5687
5688 for (i = 0; i < ARRAY_SIZE(cparams); i++) {
5689 if (cparams[i].i == dev_priv->ips.c_m &&
5690 cparams[i].t == dev_priv->ips.r_t) {
5691 m = cparams[i].m;
5692 c = cparams[i].c;
5693 break;
5694 }
5695 }
5696
5697 diff = div_u64(diff, diff1);
5698 ret = ((m * diff) + c);
5699 ret = div_u64(ret, 10);
5700
5701 dev_priv->ips.last_count1 = total_count;
5702 dev_priv->ips.last_time1 = now;
5703
5704 dev_priv->ips.chipset_power = ret;
5705
5706 return ret;
5707 }
5708
5709 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
5710 {
5711 struct drm_device *dev = dev_priv->dev;
5712 unsigned long val;
5713
5714 if (INTEL_INFO(dev)->gen != 5)
5715 return 0;
5716
5717 spin_lock_irq(&mchdev_lock);
5718
5719 val = __i915_chipset_val(dev_priv);
5720
5721 spin_unlock_irq(&mchdev_lock);
5722
5723 return val;
5724 }
5725
5726 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
5727 {
5728 unsigned long m, x, b;
5729 u32 tsfs;
5730
5731 tsfs = I915_READ(TSFS);
5732
5733 m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
5734 x = I915_READ8(TR1);
5735
5736 b = tsfs & TSFS_INTR_MASK;
5737
5738 return ((m * x) / 127) - b;
5739 }
5740
5741 static int _pxvid_to_vd(u8 pxvid)
5742 {
5743 if (pxvid == 0)
5744 return 0;
5745
5746 if (pxvid >= 8 && pxvid < 31)
5747 pxvid = 31;
5748
5749 return (pxvid + 2) * 125;
5750 }
5751
5752 static u32 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
5753 {
5754 struct drm_device *dev = dev_priv->dev;
5755 const int vd = _pxvid_to_vd(pxvid);
5756 const int vm = vd - 1125;
5757
5758 if (INTEL_INFO(dev)->is_mobile)
5759 return vm > 0 ? vm : 0;
5760
5761 return vd;
5762 }
5763
5764 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
5765 {
5766 u64 now, diff, diffms;
5767 u32 count;
5768
5769 assert_spin_locked(&mchdev_lock);
5770
5771 now = ktime_get_raw_ns();
5772 diffms = now - dev_priv->ips.last_time2;
5773 do_div(diffms, NSEC_PER_MSEC);
5774
5775 /* Don't divide by 0 */
5776 if (!diffms)
5777 return;
5778
5779 count = I915_READ(GFXEC);
5780
5781 if (count < dev_priv->ips.last_count2) {
5782 diff = ~0UL - dev_priv->ips.last_count2;
5783 diff += count;
5784 } else {
5785 diff = count - dev_priv->ips.last_count2;
5786 }
5787
5788 dev_priv->ips.last_count2 = count;
5789 dev_priv->ips.last_time2 = now;
5790
5791 /* More magic constants... */
5792 diff = diff * 1181;
5793 diff = div_u64(diff, diffms * 10);
5794 dev_priv->ips.gfx_power = diff;
5795 }
5796
5797 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
5798 {
5799 struct drm_device *dev = dev_priv->dev;
5800
5801 if (INTEL_INFO(dev)->gen != 5)
5802 return;
5803
5804 spin_lock_irq(&mchdev_lock);
5805
5806 __i915_update_gfx_val(dev_priv);
5807
5808 spin_unlock_irq(&mchdev_lock);
5809 }
5810
5811 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
5812 {
5813 unsigned long t, corr, state1, corr2, state2;
5814 u32 pxvid, ext_v;
5815
5816 assert_spin_locked(&mchdev_lock);
5817
5818 pxvid = I915_READ(PXVFREQ(dev_priv->rps.cur_freq));
5819 pxvid = (pxvid >> 24) & 0x7f;
5820 ext_v = pvid_to_extvid(dev_priv, pxvid);
5821
5822 state1 = ext_v;
5823
5824 t = i915_mch_val(dev_priv);
5825
5826 /* Revel in the empirically derived constants */
5827
5828 /* Correction factor in 1/100000 units */
5829 if (t > 80)
5830 corr = ((t * 2349) + 135940);
5831 else if (t >= 50)
5832 corr = ((t * 964) + 29317);
5833 else /* < 50 */
5834 corr = ((t * 301) + 1004);
5835
5836 corr = corr * ((150142 * state1) / 10000 - 78642);
5837 corr /= 100000;
5838 corr2 = (corr * dev_priv->ips.corr);
5839
5840 state2 = (corr2 * state1) / 10000;
5841 state2 /= 100; /* convert to mW */
5842
5843 __i915_update_gfx_val(dev_priv);
5844
5845 return dev_priv->ips.gfx_power + state2;
5846 }
5847
5848 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
5849 {
5850 struct drm_device *dev = dev_priv->dev;
5851 unsigned long val;
5852
5853 if (INTEL_INFO(dev)->gen != 5)
5854 return 0;
5855
5856 spin_lock_irq(&mchdev_lock);
5857
5858 val = __i915_gfx_val(dev_priv);
5859
5860 spin_unlock_irq(&mchdev_lock);
5861
5862 return val;
5863 }
5864
5865 /**
5866 * i915_read_mch_val - return value for IPS use
5867 *
5868 * Calculate and return a value for the IPS driver to use when deciding whether
5869 * we have thermal and power headroom to increase CPU or GPU power budget.
5870 */
5871 unsigned long i915_read_mch_val(void)
5872 {
5873 struct drm_i915_private *dev_priv;
5874 unsigned long chipset_val, graphics_val, ret = 0;
5875
5876 spin_lock_irq(&mchdev_lock);
5877 if (!i915_mch_dev)
5878 goto out_unlock;
5879 dev_priv = i915_mch_dev;
5880
5881 chipset_val = __i915_chipset_val(dev_priv);
5882 graphics_val = __i915_gfx_val(dev_priv);
5883
5884 ret = chipset_val + graphics_val;
5885
5886 out_unlock:
5887 spin_unlock_irq(&mchdev_lock);
5888
5889 return ret;
5890 }
5891 EXPORT_SYMBOL_GPL(i915_read_mch_val);
5892
5893 /**
5894 * i915_gpu_raise - raise GPU frequency limit
5895 *
5896 * Raise the limit; IPS indicates we have thermal headroom.
5897 */
5898 bool i915_gpu_raise(void)
5899 {
5900 struct drm_i915_private *dev_priv;
5901 bool ret = true;
5902
5903 spin_lock_irq(&mchdev_lock);
5904 if (!i915_mch_dev) {
5905 ret = false;
5906 goto out_unlock;
5907 }
5908 dev_priv = i915_mch_dev;
5909
5910 if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
5911 dev_priv->ips.max_delay--;
5912
5913 out_unlock:
5914 spin_unlock_irq(&mchdev_lock);
5915
5916 return ret;
5917 }
5918 EXPORT_SYMBOL_GPL(i915_gpu_raise);
5919
5920 /**
5921 * i915_gpu_lower - lower GPU frequency limit
5922 *
5923 * IPS indicates we're close to a thermal limit, so throttle back the GPU
5924 * frequency maximum.
5925 */
5926 bool i915_gpu_lower(void)
5927 {
5928 struct drm_i915_private *dev_priv;
5929 bool ret = true;
5930
5931 spin_lock_irq(&mchdev_lock);
5932 if (!i915_mch_dev) {
5933 ret = false;
5934 goto out_unlock;
5935 }
5936 dev_priv = i915_mch_dev;
5937
5938 if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
5939 dev_priv->ips.max_delay++;
5940
5941 out_unlock:
5942 spin_unlock_irq(&mchdev_lock);
5943
5944 return ret;
5945 }
5946 EXPORT_SYMBOL_GPL(i915_gpu_lower);
5947
5948 /**
5949 * i915_gpu_busy - indicate GPU business to IPS
5950 *
5951 * Tell the IPS driver whether or not the GPU is busy.
5952 */
5953 bool i915_gpu_busy(void)
5954 {
5955 struct drm_i915_private *dev_priv;
5956 struct intel_engine_cs *ring;
5957 bool ret = false;
5958 int i;
5959
5960 spin_lock_irq(&mchdev_lock);
5961 if (!i915_mch_dev)
5962 goto out_unlock;
5963 dev_priv = i915_mch_dev;
5964
5965 for_each_ring(ring, dev_priv, i)
5966 ret |= !list_empty(&ring->request_list);
5967
5968 out_unlock:
5969 spin_unlock_irq(&mchdev_lock);
5970
5971 return ret;
5972 }
5973 EXPORT_SYMBOL_GPL(i915_gpu_busy);
5974
5975 /**
5976 * i915_gpu_turbo_disable - disable graphics turbo
5977 *
5978 * Disable graphics turbo by resetting the max frequency and setting the
5979 * current frequency to the default.
5980 */
5981 bool i915_gpu_turbo_disable(void)
5982 {
5983 struct drm_i915_private *dev_priv;
5984 bool ret = true;
5985
5986 spin_lock_irq(&mchdev_lock);
5987 if (!i915_mch_dev) {
5988 ret = false;
5989 goto out_unlock;
5990 }
5991 dev_priv = i915_mch_dev;
5992
5993 dev_priv->ips.max_delay = dev_priv->ips.fstart;
5994
5995 if (!ironlake_set_drps(dev_priv->dev, dev_priv->ips.fstart))
5996 ret = false;
5997
5998 out_unlock:
5999 spin_unlock_irq(&mchdev_lock);
6000
6001 return ret;
6002 }
6003 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
6004
6005 /**
6006 * Tells the intel_ips driver that the i915 driver is now loaded, if
6007 * IPS got loaded first.
6008 *
6009 * This awkward dance is so that neither module has to depend on the
6010 * other in order for IPS to do the appropriate communication of
6011 * GPU turbo limits to i915.
6012 */
6013 static void
6014 ips_ping_for_i915_load(void)
6015 {
6016 void (*link)(void);
6017
6018 link = symbol_get(ips_link_to_i915_driver);
6019 if (link) {
6020 link();
6021 symbol_put(ips_link_to_i915_driver);
6022 }
6023 }
6024
6025 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
6026 {
6027 /* We only register the i915 ips part with intel-ips once everything is
6028 * set up, to avoid intel-ips sneaking in and reading bogus values. */
6029 spin_lock_irq(&mchdev_lock);
6030 i915_mch_dev = dev_priv;
6031 spin_unlock_irq(&mchdev_lock);
6032
6033 ips_ping_for_i915_load();
6034 }
6035
6036 void intel_gpu_ips_teardown(void)
6037 {
6038 spin_lock_irq(&mchdev_lock);
6039 i915_mch_dev = NULL;
6040 spin_unlock_irq(&mchdev_lock);
6041 }
6042
6043 static void intel_init_emon(struct drm_device *dev)
6044 {
6045 struct drm_i915_private *dev_priv = dev->dev_private;
6046 u32 lcfuse;
6047 u8 pxw[16];
6048 int i;
6049
6050 /* Disable to program */
6051 I915_WRITE(ECR, 0);
6052 POSTING_READ(ECR);
6053
6054 /* Program energy weights for various events */
6055 I915_WRITE(SDEW, 0x15040d00);
6056 I915_WRITE(CSIEW0, 0x007f0000);
6057 I915_WRITE(CSIEW1, 0x1e220004);
6058 I915_WRITE(CSIEW2, 0x04000004);
6059
6060 for (i = 0; i < 5; i++)
6061 I915_WRITE(PEW(i), 0);
6062 for (i = 0; i < 3; i++)
6063 I915_WRITE(DEW(i), 0);
6064
6065 /* Program P-state weights to account for frequency power adjustment */
6066 for (i = 0; i < 16; i++) {
6067 u32 pxvidfreq = I915_READ(PXVFREQ(i));
6068 unsigned long freq = intel_pxfreq(pxvidfreq);
6069 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
6070 PXVFREQ_PX_SHIFT;
6071 unsigned long val;
6072
6073 val = vid * vid;
6074 val *= (freq / 1000);
6075 val *= 255;
6076 val /= (127*127*900);
6077 if (val > 0xff)
6078 DRM_ERROR("bad pxval: %ld\n", val);
6079 pxw[i] = val;
6080 }
6081 /* Render standby states get 0 weight */
6082 pxw[14] = 0;
6083 pxw[15] = 0;
6084
6085 for (i = 0; i < 4; i++) {
6086 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
6087 (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
6088 I915_WRITE(PXW(i), val);
6089 }
6090
6091 /* Adjust magic regs to magic values (more experimental results) */
6092 I915_WRITE(OGW0, 0);
6093 I915_WRITE(OGW1, 0);
6094 I915_WRITE(EG0, 0x00007f00);
6095 I915_WRITE(EG1, 0x0000000e);
6096 I915_WRITE(EG2, 0x000e0000);
6097 I915_WRITE(EG3, 0x68000300);
6098 I915_WRITE(EG4, 0x42000000);
6099 I915_WRITE(EG5, 0x00140031);
6100 I915_WRITE(EG6, 0);
6101 I915_WRITE(EG7, 0);
6102
6103 for (i = 0; i < 8; i++)
6104 I915_WRITE(PXWL(i), 0);
6105
6106 /* Enable PMON + select events */
6107 I915_WRITE(ECR, 0x80000019);
6108
6109 lcfuse = I915_READ(LCFUSE02);
6110
6111 dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
6112 }
6113
6114 void intel_init_gt_powersave(struct drm_device *dev)
6115 {
6116 i915.enable_rc6 = sanitize_rc6_option(dev, i915.enable_rc6);
6117
6118 if (IS_CHERRYVIEW(dev))
6119 cherryview_init_gt_powersave(dev);
6120 else if (IS_VALLEYVIEW(dev))
6121 valleyview_init_gt_powersave(dev);
6122 }
6123
6124 void intel_cleanup_gt_powersave(struct drm_device *dev)
6125 {
6126 if (IS_CHERRYVIEW(dev))
6127 return;
6128 else if (IS_VALLEYVIEW(dev))
6129 valleyview_cleanup_gt_powersave(dev);
6130 }
6131
6132 static void gen6_suspend_rps(struct drm_device *dev)
6133 {
6134 struct drm_i915_private *dev_priv = dev->dev_private;
6135
6136 flush_delayed_work(&dev_priv->rps.delayed_resume_work);
6137
6138 gen6_disable_rps_interrupts(dev);
6139 }
6140
6141 /**
6142 * intel_suspend_gt_powersave - suspend PM work and helper threads
6143 * @dev: drm device
6144 *
6145 * We don't want to disable RC6 or other features here, we just want
6146 * to make sure any work we've queued has finished and won't bother
6147 * us while we're suspended.
6148 */
6149 void intel_suspend_gt_powersave(struct drm_device *dev)
6150 {
6151 struct drm_i915_private *dev_priv = dev->dev_private;
6152
6153 if (INTEL_INFO(dev)->gen < 6)
6154 return;
6155
6156 gen6_suspend_rps(dev);
6157
6158 /* Force GPU to min freq during suspend */
6159 gen6_rps_idle(dev_priv);
6160 }
6161
6162 void intel_disable_gt_powersave(struct drm_device *dev)
6163 {
6164 struct drm_i915_private *dev_priv = dev->dev_private;
6165
6166 if (IS_IRONLAKE_M(dev)) {
6167 ironlake_disable_drps(dev);
6168 } else if (INTEL_INFO(dev)->gen >= 6) {
6169 intel_suspend_gt_powersave(dev);
6170
6171 mutex_lock(&dev_priv->rps.hw_lock);
6172 if (INTEL_INFO(dev)->gen >= 9)
6173 gen9_disable_rps(dev);
6174 else if (IS_CHERRYVIEW(dev))
6175 cherryview_disable_rps(dev);
6176 else if (IS_VALLEYVIEW(dev))
6177 valleyview_disable_rps(dev);
6178 else
6179 gen6_disable_rps(dev);
6180
6181 dev_priv->rps.enabled = false;
6182 mutex_unlock(&dev_priv->rps.hw_lock);
6183 }
6184 }
6185
6186 static void intel_gen6_powersave_work(struct work_struct *work)
6187 {
6188 struct drm_i915_private *dev_priv =
6189 container_of(work, struct drm_i915_private,
6190 rps.delayed_resume_work.work);
6191 struct drm_device *dev = dev_priv->dev;
6192
6193 mutex_lock(&dev_priv->rps.hw_lock);
6194
6195 gen6_reset_rps_interrupts(dev);
6196
6197 if (IS_CHERRYVIEW(dev)) {
6198 cherryview_enable_rps(dev);
6199 } else if (IS_VALLEYVIEW(dev)) {
6200 valleyview_enable_rps(dev);
6201 } else if (INTEL_INFO(dev)->gen >= 9) {
6202 gen9_enable_rc6(dev);
6203 gen9_enable_rps(dev);
6204 if (IS_SKYLAKE(dev))
6205 __gen6_update_ring_freq(dev);
6206 } else if (IS_BROADWELL(dev)) {
6207 gen8_enable_rps(dev);
6208 __gen6_update_ring_freq(dev);
6209 } else {
6210 gen6_enable_rps(dev);
6211 __gen6_update_ring_freq(dev);
6212 }
6213
6214 WARN_ON(dev_priv->rps.max_freq < dev_priv->rps.min_freq);
6215 WARN_ON(dev_priv->rps.idle_freq > dev_priv->rps.max_freq);
6216
6217 WARN_ON(dev_priv->rps.efficient_freq < dev_priv->rps.min_freq);
6218 WARN_ON(dev_priv->rps.efficient_freq > dev_priv->rps.max_freq);
6219
6220 dev_priv->rps.enabled = true;
6221
6222 gen6_enable_rps_interrupts(dev);
6223
6224 mutex_unlock(&dev_priv->rps.hw_lock);
6225
6226 intel_runtime_pm_put(dev_priv);
6227 }
6228
6229 void intel_enable_gt_powersave(struct drm_device *dev)
6230 {
6231 struct drm_i915_private *dev_priv = dev->dev_private;
6232
6233 /* Powersaving is controlled by the host when inside a VM */
6234 if (intel_vgpu_active(dev))
6235 return;
6236
6237 if (IS_IRONLAKE_M(dev)) {
6238 mutex_lock(&dev->struct_mutex);
6239 ironlake_enable_drps(dev);
6240 intel_init_emon(dev);
6241 mutex_unlock(&dev->struct_mutex);
6242 } else if (INTEL_INFO(dev)->gen >= 6) {
6243 /*
6244 * PCU communication is slow and this doesn't need to be
6245 * done at any specific time, so do this out of our fast path
6246 * to make resume and init faster.
6247 *
6248 * We depend on the HW RC6 power context save/restore
6249 * mechanism when entering D3 through runtime PM suspend. So
6250 * disable RPM until RPS/RC6 is properly setup. We can only
6251 * get here via the driver load/system resume/runtime resume
6252 * paths, so the _noresume version is enough (and in case of
6253 * runtime resume it's necessary).
6254 */
6255 if (schedule_delayed_work(&dev_priv->rps.delayed_resume_work,
6256 round_jiffies_up_relative(HZ)))
6257 intel_runtime_pm_get_noresume(dev_priv);
6258 }
6259 }
6260
6261 void intel_reset_gt_powersave(struct drm_device *dev)
6262 {
6263 struct drm_i915_private *dev_priv = dev->dev_private;
6264
6265 if (INTEL_INFO(dev)->gen < 6)
6266 return;
6267
6268 gen6_suspend_rps(dev);
6269 dev_priv->rps.enabled = false;
6270 }
6271
6272 static void ibx_init_clock_gating(struct drm_device *dev)
6273 {
6274 struct drm_i915_private *dev_priv = dev->dev_private;
6275
6276 /*
6277 * On Ibex Peak and Cougar Point, we need to disable clock
6278 * gating for the panel power sequencer or it will fail to
6279 * start up when no ports are active.
6280 */
6281 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
6282 }
6283
6284 static void g4x_disable_trickle_feed(struct drm_device *dev)
6285 {
6286 struct drm_i915_private *dev_priv = dev->dev_private;
6287 enum pipe pipe;
6288
6289 for_each_pipe(dev_priv, pipe) {
6290 I915_WRITE(DSPCNTR(pipe),
6291 I915_READ(DSPCNTR(pipe)) |
6292 DISPPLANE_TRICKLE_FEED_DISABLE);
6293
6294 I915_WRITE(DSPSURF(pipe), I915_READ(DSPSURF(pipe)));
6295 POSTING_READ(DSPSURF(pipe));
6296 }
6297 }
6298
6299 static void ilk_init_lp_watermarks(struct drm_device *dev)
6300 {
6301 struct drm_i915_private *dev_priv = dev->dev_private;
6302
6303 I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
6304 I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
6305 I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
6306
6307 /*
6308 * Don't touch WM1S_LP_EN here.
6309 * Doing so could cause underruns.
6310 */
6311 }
6312
6313 static void ironlake_init_clock_gating(struct drm_device *dev)
6314 {
6315 struct drm_i915_private *dev_priv = dev->dev_private;
6316 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6317
6318 /*
6319 * Required for FBC
6320 * WaFbcDisableDpfcClockGating:ilk
6321 */
6322 dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
6323 ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
6324 ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
6325
6326 I915_WRITE(PCH_3DCGDIS0,
6327 MARIUNIT_CLOCK_GATE_DISABLE |
6328 SVSMUNIT_CLOCK_GATE_DISABLE);
6329 I915_WRITE(PCH_3DCGDIS1,
6330 VFMUNIT_CLOCK_GATE_DISABLE);
6331
6332 /*
6333 * According to the spec the following bits should be set in
6334 * order to enable memory self-refresh
6335 * The bit 22/21 of 0x42004
6336 * The bit 5 of 0x42020
6337 * The bit 15 of 0x45000
6338 */
6339 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6340 (I915_READ(ILK_DISPLAY_CHICKEN2) |
6341 ILK_DPARB_GATE | ILK_VSDPFD_FULL));
6342 dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
6343 I915_WRITE(DISP_ARB_CTL,
6344 (I915_READ(DISP_ARB_CTL) |
6345 DISP_FBC_WM_DIS));
6346
6347 ilk_init_lp_watermarks(dev);
6348
6349 /*
6350 * Based on the document from hardware guys the following bits
6351 * should be set unconditionally in order to enable FBC.
6352 * The bit 22 of 0x42000
6353 * The bit 22 of 0x42004
6354 * The bit 7,8,9 of 0x42020.
6355 */
6356 if (IS_IRONLAKE_M(dev)) {
6357 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
6358 I915_WRITE(ILK_DISPLAY_CHICKEN1,
6359 I915_READ(ILK_DISPLAY_CHICKEN1) |
6360 ILK_FBCQ_DIS);
6361 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6362 I915_READ(ILK_DISPLAY_CHICKEN2) |
6363 ILK_DPARB_GATE);
6364 }
6365
6366 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
6367
6368 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6369 I915_READ(ILK_DISPLAY_CHICKEN2) |
6370 ILK_ELPIN_409_SELECT);
6371 I915_WRITE(_3D_CHICKEN2,
6372 _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
6373 _3D_CHICKEN2_WM_READ_PIPELINED);
6374
6375 /* WaDisableRenderCachePipelinedFlush:ilk */
6376 I915_WRITE(CACHE_MODE_0,
6377 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
6378
6379 /* WaDisable_RenderCache_OperationalFlush:ilk */
6380 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6381
6382 g4x_disable_trickle_feed(dev);
6383
6384 ibx_init_clock_gating(dev);
6385 }
6386
6387 static void cpt_init_clock_gating(struct drm_device *dev)
6388 {
6389 struct drm_i915_private *dev_priv = dev->dev_private;
6390 int pipe;
6391 uint32_t val;
6392
6393 /*
6394 * On Ibex Peak and Cougar Point, we need to disable clock
6395 * gating for the panel power sequencer or it will fail to
6396 * start up when no ports are active.
6397 */
6398 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
6399 PCH_DPLUNIT_CLOCK_GATE_DISABLE |
6400 PCH_CPUNIT_CLOCK_GATE_DISABLE);
6401 I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
6402 DPLS_EDP_PPS_FIX_DIS);
6403 /* The below fixes the weird display corruption, a few pixels shifted
6404 * downward, on (only) LVDS of some HP laptops with IVY.
6405 */
6406 for_each_pipe(dev_priv, pipe) {
6407 val = I915_READ(TRANS_CHICKEN2(pipe));
6408 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
6409 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6410 if (dev_priv->vbt.fdi_rx_polarity_inverted)
6411 val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6412 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
6413 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
6414 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
6415 I915_WRITE(TRANS_CHICKEN2(pipe), val);
6416 }
6417 /* WADP0ClockGatingDisable */
6418 for_each_pipe(dev_priv, pipe) {
6419 I915_WRITE(TRANS_CHICKEN1(pipe),
6420 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
6421 }
6422 }
6423
6424 static void gen6_check_mch_setup(struct drm_device *dev)
6425 {
6426 struct drm_i915_private *dev_priv = dev->dev_private;
6427 uint32_t tmp;
6428
6429 tmp = I915_READ(MCH_SSKPD);
6430 if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL)
6431 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
6432 tmp);
6433 }
6434
6435 static void gen6_init_clock_gating(struct drm_device *dev)
6436 {
6437 struct drm_i915_private *dev_priv = dev->dev_private;
6438 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6439
6440 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
6441
6442 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6443 I915_READ(ILK_DISPLAY_CHICKEN2) |
6444 ILK_ELPIN_409_SELECT);
6445
6446 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
6447 I915_WRITE(_3D_CHICKEN,
6448 _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
6449
6450 /* WaDisable_RenderCache_OperationalFlush:snb */
6451 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6452
6453 /*
6454 * BSpec recoomends 8x4 when MSAA is used,
6455 * however in practice 16x4 seems fastest.
6456 *
6457 * Note that PS/WM thread counts depend on the WIZ hashing
6458 * disable bit, which we don't touch here, but it's good
6459 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6460 */
6461 I915_WRITE(GEN6_GT_MODE,
6462 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6463
6464 ilk_init_lp_watermarks(dev);
6465
6466 I915_WRITE(CACHE_MODE_0,
6467 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
6468
6469 I915_WRITE(GEN6_UCGCTL1,
6470 I915_READ(GEN6_UCGCTL1) |
6471 GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
6472 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
6473
6474 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
6475 * gating disable must be set. Failure to set it results in
6476 * flickering pixels due to Z write ordering failures after
6477 * some amount of runtime in the Mesa "fire" demo, and Unigine
6478 * Sanctuary and Tropics, and apparently anything else with
6479 * alpha test or pixel discard.
6480 *
6481 * According to the spec, bit 11 (RCCUNIT) must also be set,
6482 * but we didn't debug actual testcases to find it out.
6483 *
6484 * WaDisableRCCUnitClockGating:snb
6485 * WaDisableRCPBUnitClockGating:snb
6486 */
6487 I915_WRITE(GEN6_UCGCTL2,
6488 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
6489 GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
6490
6491 /* WaStripsFansDisableFastClipPerformanceFix:snb */
6492 I915_WRITE(_3D_CHICKEN3,
6493 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
6494
6495 /*
6496 * Bspec says:
6497 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
6498 * 3DSTATE_SF number of SF output attributes is more than 16."
6499 */
6500 I915_WRITE(_3D_CHICKEN3,
6501 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
6502
6503 /*
6504 * According to the spec the following bits should be
6505 * set in order to enable memory self-refresh and fbc:
6506 * The bit21 and bit22 of 0x42000
6507 * The bit21 and bit22 of 0x42004
6508 * The bit5 and bit7 of 0x42020
6509 * The bit14 of 0x70180
6510 * The bit14 of 0x71180
6511 *
6512 * WaFbcAsynchFlipDisableFbcQueue:snb
6513 */
6514 I915_WRITE(ILK_DISPLAY_CHICKEN1,
6515 I915_READ(ILK_DISPLAY_CHICKEN1) |
6516 ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
6517 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6518 I915_READ(ILK_DISPLAY_CHICKEN2) |
6519 ILK_DPARB_GATE | ILK_VSDPFD_FULL);
6520 I915_WRITE(ILK_DSPCLK_GATE_D,
6521 I915_READ(ILK_DSPCLK_GATE_D) |
6522 ILK_DPARBUNIT_CLOCK_GATE_ENABLE |
6523 ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
6524
6525 g4x_disable_trickle_feed(dev);
6526
6527 cpt_init_clock_gating(dev);
6528
6529 gen6_check_mch_setup(dev);
6530 }
6531
6532 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
6533 {
6534 uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
6535
6536 /*
6537 * WaVSThreadDispatchOverride:ivb,vlv
6538 *
6539 * This actually overrides the dispatch
6540 * mode for all thread types.
6541 */
6542 reg &= ~GEN7_FF_SCHED_MASK;
6543 reg |= GEN7_FF_TS_SCHED_HW;
6544 reg |= GEN7_FF_VS_SCHED_HW;
6545 reg |= GEN7_FF_DS_SCHED_HW;
6546
6547 I915_WRITE(GEN7_FF_THREAD_MODE, reg);
6548 }
6549
6550 static void lpt_init_clock_gating(struct drm_device *dev)
6551 {
6552 struct drm_i915_private *dev_priv = dev->dev_private;
6553
6554 /*
6555 * TODO: this bit should only be enabled when really needed, then
6556 * disabled when not needed anymore in order to save power.
6557 */
6558 if (HAS_PCH_LPT_LP(dev))
6559 I915_WRITE(SOUTH_DSPCLK_GATE_D,
6560 I915_READ(SOUTH_DSPCLK_GATE_D) |
6561 PCH_LP_PARTITION_LEVEL_DISABLE);
6562
6563 /* WADPOClockGatingDisable:hsw */
6564 I915_WRITE(TRANS_CHICKEN1(PIPE_A),
6565 I915_READ(TRANS_CHICKEN1(PIPE_A)) |
6566 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
6567 }
6568
6569 static void lpt_suspend_hw(struct drm_device *dev)
6570 {
6571 struct drm_i915_private *dev_priv = dev->dev_private;
6572
6573 if (HAS_PCH_LPT_LP(dev)) {
6574 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
6575
6576 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
6577 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
6578 }
6579 }
6580
6581 static void broadwell_init_clock_gating(struct drm_device *dev)
6582 {
6583 struct drm_i915_private *dev_priv = dev->dev_private;
6584 enum pipe pipe;
6585 uint32_t misccpctl;
6586
6587 ilk_init_lp_watermarks(dev);
6588
6589 /* WaSwitchSolVfFArbitrationPriority:bdw */
6590 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
6591
6592 /* WaPsrDPAMaskVBlankInSRD:bdw */
6593 I915_WRITE(CHICKEN_PAR1_1,
6594 I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
6595
6596 /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
6597 for_each_pipe(dev_priv, pipe) {
6598 I915_WRITE(CHICKEN_PIPESL_1(pipe),
6599 I915_READ(CHICKEN_PIPESL_1(pipe)) |
6600 BDW_DPRS_MASK_VBLANK_SRD);
6601 }
6602
6603 /* WaVSRefCountFullforceMissDisable:bdw */
6604 /* WaDSRefCountFullforceMissDisable:bdw */
6605 I915_WRITE(GEN7_FF_THREAD_MODE,
6606 I915_READ(GEN7_FF_THREAD_MODE) &
6607 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
6608
6609 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
6610 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
6611
6612 /* WaDisableSDEUnitClockGating:bdw */
6613 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
6614 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
6615
6616 /*
6617 * WaProgramL3SqcReg1Default:bdw
6618 * WaTempDisableDOPClkGating:bdw
6619 */
6620 misccpctl = I915_READ(GEN7_MISCCPCTL);
6621 I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
6622 I915_WRITE(GEN8_L3SQCREG1, BDW_WA_L3SQCREG1_DEFAULT);
6623 I915_WRITE(GEN7_MISCCPCTL, misccpctl);
6624
6625 /*
6626 * WaGttCachingOffByDefault:bdw
6627 * GTT cache may not work with big pages, so if those
6628 * are ever enabled GTT cache may need to be disabled.
6629 */
6630 I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
6631
6632 lpt_init_clock_gating(dev);
6633 }
6634
6635 static void haswell_init_clock_gating(struct drm_device *dev)
6636 {
6637 struct drm_i915_private *dev_priv = dev->dev_private;
6638
6639 ilk_init_lp_watermarks(dev);
6640
6641 /* L3 caching of data atomics doesn't work -- disable it. */
6642 I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
6643 I915_WRITE(HSW_ROW_CHICKEN3,
6644 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
6645
6646 /* This is required by WaCatErrorRejectionIssue:hsw */
6647 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
6648 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
6649 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
6650
6651 /* WaVSRefCountFullforceMissDisable:hsw */
6652 I915_WRITE(GEN7_FF_THREAD_MODE,
6653 I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
6654
6655 /* WaDisable_RenderCache_OperationalFlush:hsw */
6656 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6657
6658 /* enable HiZ Raw Stall Optimization */
6659 I915_WRITE(CACHE_MODE_0_GEN7,
6660 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
6661
6662 /* WaDisable4x2SubspanOptimization:hsw */
6663 I915_WRITE(CACHE_MODE_1,
6664 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
6665
6666 /*
6667 * BSpec recommends 8x4 when MSAA is used,
6668 * however in practice 16x4 seems fastest.
6669 *
6670 * Note that PS/WM thread counts depend on the WIZ hashing
6671 * disable bit, which we don't touch here, but it's good
6672 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6673 */
6674 I915_WRITE(GEN7_GT_MODE,
6675 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6676
6677 /* WaSampleCChickenBitEnable:hsw */
6678 I915_WRITE(HALF_SLICE_CHICKEN3,
6679 _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE));
6680
6681 /* WaSwitchSolVfFArbitrationPriority:hsw */
6682 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
6683
6684 /* WaRsPkgCStateDisplayPMReq:hsw */
6685 I915_WRITE(CHICKEN_PAR1_1,
6686 I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
6687
6688 lpt_init_clock_gating(dev);
6689 }
6690
6691 static void ivybridge_init_clock_gating(struct drm_device *dev)
6692 {
6693 struct drm_i915_private *dev_priv = dev->dev_private;
6694 uint32_t snpcr;
6695
6696 ilk_init_lp_watermarks(dev);
6697
6698 I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
6699
6700 /* WaDisableEarlyCull:ivb */
6701 I915_WRITE(_3D_CHICKEN3,
6702 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
6703
6704 /* WaDisableBackToBackFlipFix:ivb */
6705 I915_WRITE(IVB_CHICKEN3,
6706 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
6707 CHICKEN3_DGMG_DONE_FIX_DISABLE);
6708
6709 /* WaDisablePSDDualDispatchEnable:ivb */
6710 if (IS_IVB_GT1(dev))
6711 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
6712 _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
6713
6714 /* WaDisable_RenderCache_OperationalFlush:ivb */
6715 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6716
6717 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
6718 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
6719 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
6720
6721 /* WaApplyL3ControlAndL3ChickenMode:ivb */
6722 I915_WRITE(GEN7_L3CNTLREG1,
6723 GEN7_WA_FOR_GEN7_L3_CONTROL);
6724 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
6725 GEN7_WA_L3_CHICKEN_MODE);
6726 if (IS_IVB_GT1(dev))
6727 I915_WRITE(GEN7_ROW_CHICKEN2,
6728 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6729 else {
6730 /* must write both registers */
6731 I915_WRITE(GEN7_ROW_CHICKEN2,
6732 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6733 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
6734 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6735 }
6736
6737 /* WaForceL3Serialization:ivb */
6738 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
6739 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
6740
6741 /*
6742 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6743 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
6744 */
6745 I915_WRITE(GEN6_UCGCTL2,
6746 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
6747
6748 /* This is required by WaCatErrorRejectionIssue:ivb */
6749 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
6750 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
6751 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
6752
6753 g4x_disable_trickle_feed(dev);
6754
6755 gen7_setup_fixed_func_scheduler(dev_priv);
6756
6757 if (0) { /* causes HiZ corruption on ivb:gt1 */
6758 /* enable HiZ Raw Stall Optimization */
6759 I915_WRITE(CACHE_MODE_0_GEN7,
6760 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
6761 }
6762
6763 /* WaDisable4x2SubspanOptimization:ivb */
6764 I915_WRITE(CACHE_MODE_1,
6765 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
6766
6767 /*
6768 * BSpec recommends 8x4 when MSAA is used,
6769 * however in practice 16x4 seems fastest.
6770 *
6771 * Note that PS/WM thread counts depend on the WIZ hashing
6772 * disable bit, which we don't touch here, but it's good
6773 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6774 */
6775 I915_WRITE(GEN7_GT_MODE,
6776 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6777
6778 snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
6779 snpcr &= ~GEN6_MBC_SNPCR_MASK;
6780 snpcr |= GEN6_MBC_SNPCR_MED;
6781 I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
6782
6783 if (!HAS_PCH_NOP(dev))
6784 cpt_init_clock_gating(dev);
6785
6786 gen6_check_mch_setup(dev);
6787 }
6788
6789 static void vlv_init_display_clock_gating(struct drm_i915_private *dev_priv)
6790 {
6791 I915_WRITE(DSPCLK_GATE_D, VRHUNIT_CLOCK_GATE_DISABLE);
6792
6793 /*
6794 * Disable trickle feed and enable pnd deadline calculation
6795 */
6796 I915_WRITE(MI_ARB_VLV, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE);
6797 I915_WRITE(CBR1_VLV, 0);
6798 }
6799
6800 static void valleyview_init_clock_gating(struct drm_device *dev)
6801 {
6802 struct drm_i915_private *dev_priv = dev->dev_private;
6803
6804 vlv_init_display_clock_gating(dev_priv);
6805
6806 /* WaDisableEarlyCull:vlv */
6807 I915_WRITE(_3D_CHICKEN3,
6808 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
6809
6810 /* WaDisableBackToBackFlipFix:vlv */
6811 I915_WRITE(IVB_CHICKEN3,
6812 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
6813 CHICKEN3_DGMG_DONE_FIX_DISABLE);
6814
6815 /* WaPsdDispatchEnable:vlv */
6816 /* WaDisablePSDDualDispatchEnable:vlv */
6817 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
6818 _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
6819 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
6820
6821 /* WaDisable_RenderCache_OperationalFlush:vlv */
6822 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6823
6824 /* WaForceL3Serialization:vlv */
6825 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
6826 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
6827
6828 /* WaDisableDopClockGating:vlv */
6829 I915_WRITE(GEN7_ROW_CHICKEN2,
6830 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6831
6832 /* This is required by WaCatErrorRejectionIssue:vlv */
6833 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
6834 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
6835 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
6836
6837 gen7_setup_fixed_func_scheduler(dev_priv);
6838
6839 /*
6840 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6841 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
6842 */
6843 I915_WRITE(GEN6_UCGCTL2,
6844 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
6845
6846 /* WaDisableL3Bank2xClockGate:vlv
6847 * Disabling L3 clock gating- MMIO 940c[25] = 1
6848 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
6849 I915_WRITE(GEN7_UCGCTL4,
6850 I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
6851
6852 /*
6853 * BSpec says this must be set, even though
6854 * WaDisable4x2SubspanOptimization isn't listed for VLV.
6855 */
6856 I915_WRITE(CACHE_MODE_1,
6857 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
6858
6859 /*
6860 * BSpec recommends 8x4 when MSAA is used,
6861 * however in practice 16x4 seems fastest.
6862 *
6863 * Note that PS/WM thread counts depend on the WIZ hashing
6864 * disable bit, which we don't touch here, but it's good
6865 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6866 */
6867 I915_WRITE(GEN7_GT_MODE,
6868 _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6869
6870 /*
6871 * WaIncreaseL3CreditsForVLVB0:vlv
6872 * This is the hardware default actually.
6873 */
6874 I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
6875
6876 /*
6877 * WaDisableVLVClockGating_VBIIssue:vlv
6878 * Disable clock gating on th GCFG unit to prevent a delay
6879 * in the reporting of vblank events.
6880 */
6881 I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
6882 }
6883
6884 static void cherryview_init_clock_gating(struct drm_device *dev)
6885 {
6886 struct drm_i915_private *dev_priv = dev->dev_private;
6887
6888 vlv_init_display_clock_gating(dev_priv);
6889
6890 /* WaVSRefCountFullforceMissDisable:chv */
6891 /* WaDSRefCountFullforceMissDisable:chv */
6892 I915_WRITE(GEN7_FF_THREAD_MODE,
6893 I915_READ(GEN7_FF_THREAD_MODE) &
6894 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
6895
6896 /* WaDisableSemaphoreAndSyncFlipWait:chv */
6897 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
6898 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
6899
6900 /* WaDisableCSUnitClockGating:chv */
6901 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
6902 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
6903
6904 /* WaDisableSDEUnitClockGating:chv */
6905 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
6906 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
6907
6908 /*
6909 * GTT cache may not work with big pages, so if those
6910 * are ever enabled GTT cache may need to be disabled.
6911 */
6912 I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
6913 }
6914
6915 static void g4x_init_clock_gating(struct drm_device *dev)
6916 {
6917 struct drm_i915_private *dev_priv = dev->dev_private;
6918 uint32_t dspclk_gate;
6919
6920 I915_WRITE(RENCLK_GATE_D1, 0);
6921 I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
6922 GS_UNIT_CLOCK_GATE_DISABLE |
6923 CL_UNIT_CLOCK_GATE_DISABLE);
6924 I915_WRITE(RAMCLK_GATE_D, 0);
6925 dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
6926 OVRUNIT_CLOCK_GATE_DISABLE |
6927 OVCUNIT_CLOCK_GATE_DISABLE;
6928 if (IS_GM45(dev))
6929 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
6930 I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
6931
6932 /* WaDisableRenderCachePipelinedFlush */
6933 I915_WRITE(CACHE_MODE_0,
6934 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
6935
6936 /* WaDisable_RenderCache_OperationalFlush:g4x */
6937 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6938
6939 g4x_disable_trickle_feed(dev);
6940 }
6941
6942 static void crestline_init_clock_gating(struct drm_device *dev)
6943 {
6944 struct drm_i915_private *dev_priv = dev->dev_private;
6945
6946 I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
6947 I915_WRITE(RENCLK_GATE_D2, 0);
6948 I915_WRITE(DSPCLK_GATE_D, 0);
6949 I915_WRITE(RAMCLK_GATE_D, 0);
6950 I915_WRITE16(DEUC, 0);
6951 I915_WRITE(MI_ARB_STATE,
6952 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
6953
6954 /* WaDisable_RenderCache_OperationalFlush:gen4 */
6955 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6956 }
6957
6958 static void broadwater_init_clock_gating(struct drm_device *dev)
6959 {
6960 struct drm_i915_private *dev_priv = dev->dev_private;
6961
6962 I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
6963 I965_RCC_CLOCK_GATE_DISABLE |
6964 I965_RCPB_CLOCK_GATE_DISABLE |
6965 I965_ISC_CLOCK_GATE_DISABLE |
6966 I965_FBC_CLOCK_GATE_DISABLE);
6967 I915_WRITE(RENCLK_GATE_D2, 0);
6968 I915_WRITE(MI_ARB_STATE,
6969 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
6970
6971 /* WaDisable_RenderCache_OperationalFlush:gen4 */
6972 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6973 }
6974
6975 static void gen3_init_clock_gating(struct drm_device *dev)
6976 {
6977 struct drm_i915_private *dev_priv = dev->dev_private;
6978 u32 dstate = I915_READ(D_STATE);
6979
6980 dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
6981 DSTATE_DOT_CLOCK_GATING;
6982 I915_WRITE(D_STATE, dstate);
6983
6984 if (IS_PINEVIEW(dev))
6985 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
6986
6987 /* IIR "flip pending" means done if this bit is set */
6988 I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
6989
6990 /* interrupts should cause a wake up from C3 */
6991 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
6992
6993 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
6994 I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
6995
6996 I915_WRITE(MI_ARB_STATE,
6997 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
6998 }
6999
7000 static void i85x_init_clock_gating(struct drm_device *dev)
7001 {
7002 struct drm_i915_private *dev_priv = dev->dev_private;
7003
7004 I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
7005
7006 /* interrupts should cause a wake up from C3 */
7007 I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
7008 _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
7009
7010 I915_WRITE(MEM_MODE,
7011 _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE));
7012 }
7013
7014 static void i830_init_clock_gating(struct drm_device *dev)
7015 {
7016 struct drm_i915_private *dev_priv = dev->dev_private;
7017
7018 I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
7019
7020 I915_WRITE(MEM_MODE,
7021 _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE) |
7022 _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE));
7023 }
7024
7025 void intel_init_clock_gating(struct drm_device *dev)
7026 {
7027 struct drm_i915_private *dev_priv = dev->dev_private;
7028
7029 if (dev_priv->display.init_clock_gating)
7030 dev_priv->display.init_clock_gating(dev);
7031 }
7032
7033 void intel_suspend_hw(struct drm_device *dev)
7034 {
7035 if (HAS_PCH_LPT(dev))
7036 lpt_suspend_hw(dev);
7037 }
7038
7039 /* Set up chip specific power management-related functions */
7040 void intel_init_pm(struct drm_device *dev)
7041 {
7042 struct drm_i915_private *dev_priv = dev->dev_private;
7043
7044 intel_fbc_init(dev_priv);
7045
7046 /* For cxsr */
7047 if (IS_PINEVIEW(dev))
7048 i915_pineview_get_mem_freq(dev);
7049 else if (IS_GEN5(dev))
7050 i915_ironlake_get_mem_freq(dev);
7051
7052 /* For FIFO watermark updates */
7053 if (INTEL_INFO(dev)->gen >= 9) {
7054 skl_setup_wm_latency(dev);
7055
7056 if (IS_BROXTON(dev))
7057 dev_priv->display.init_clock_gating =
7058 bxt_init_clock_gating;
7059 dev_priv->display.update_wm = skl_update_wm;
7060 dev_priv->display.update_sprite_wm = skl_update_sprite_wm;
7061 } else if (HAS_PCH_SPLIT(dev)) {
7062 ilk_setup_wm_latency(dev);
7063
7064 if ((IS_GEN5(dev) && dev_priv->wm.pri_latency[1] &&
7065 dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
7066 (!IS_GEN5(dev) && dev_priv->wm.pri_latency[0] &&
7067 dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
7068 dev_priv->display.update_wm = ilk_update_wm;
7069 dev_priv->display.update_sprite_wm = ilk_update_sprite_wm;
7070 } else {
7071 DRM_DEBUG_KMS("Failed to read display plane latency. "
7072 "Disable CxSR\n");
7073 }
7074
7075 if (IS_GEN5(dev))
7076 dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
7077 else if (IS_GEN6(dev))
7078 dev_priv->display.init_clock_gating = gen6_init_clock_gating;
7079 else if (IS_IVYBRIDGE(dev))
7080 dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
7081 else if (IS_HASWELL(dev))
7082 dev_priv->display.init_clock_gating = haswell_init_clock_gating;
7083 else if (INTEL_INFO(dev)->gen == 8)
7084 dev_priv->display.init_clock_gating = broadwell_init_clock_gating;
7085 } else if (IS_CHERRYVIEW(dev)) {
7086 vlv_setup_wm_latency(dev);
7087
7088 dev_priv->display.update_wm = vlv_update_wm;
7089 dev_priv->display.init_clock_gating =
7090 cherryview_init_clock_gating;
7091 } else if (IS_VALLEYVIEW(dev)) {
7092 vlv_setup_wm_latency(dev);
7093
7094 dev_priv->display.update_wm = vlv_update_wm;
7095 dev_priv->display.init_clock_gating =
7096 valleyview_init_clock_gating;
7097 } else if (IS_PINEVIEW(dev)) {
7098 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
7099 dev_priv->is_ddr3,
7100 dev_priv->fsb_freq,
7101 dev_priv->mem_freq)) {
7102 DRM_INFO("failed to find known CxSR latency "
7103 "(found ddr%s fsb freq %d, mem freq %d), "
7104 "disabling CxSR\n",
7105 (dev_priv->is_ddr3 == 1) ? "3" : "2",
7106 dev_priv->fsb_freq, dev_priv->mem_freq);
7107 /* Disable CxSR and never update its watermark again */
7108 intel_set_memory_cxsr(dev_priv, false);
7109 dev_priv->display.update_wm = NULL;
7110 } else
7111 dev_priv->display.update_wm = pineview_update_wm;
7112 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
7113 } else if (IS_G4X(dev)) {
7114 dev_priv->display.update_wm = g4x_update_wm;
7115 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
7116 } else if (IS_GEN4(dev)) {
7117 dev_priv->display.update_wm = i965_update_wm;
7118 if (IS_CRESTLINE(dev))
7119 dev_priv->display.init_clock_gating = crestline_init_clock_gating;
7120 else if (IS_BROADWATER(dev))
7121 dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
7122 } else if (IS_GEN3(dev)) {
7123 dev_priv->display.update_wm = i9xx_update_wm;
7124 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
7125 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
7126 } else if (IS_GEN2(dev)) {
7127 if (INTEL_INFO(dev)->num_pipes == 1) {
7128 dev_priv->display.update_wm = i845_update_wm;
7129 dev_priv->display.get_fifo_size = i845_get_fifo_size;
7130 } else {
7131 dev_priv->display.update_wm = i9xx_update_wm;
7132 dev_priv->display.get_fifo_size = i830_get_fifo_size;
7133 }
7134
7135 if (IS_I85X(dev) || IS_I865G(dev))
7136 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
7137 else
7138 dev_priv->display.init_clock_gating = i830_init_clock_gating;
7139 } else {
7140 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
7141 }
7142 }
7143
7144 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u32 mbox, u32 *val)
7145 {
7146 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7147
7148 if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
7149 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
7150 return -EAGAIN;
7151 }
7152
7153 I915_WRITE(GEN6_PCODE_DATA, *val);
7154 I915_WRITE(GEN6_PCODE_DATA1, 0);
7155 I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
7156
7157 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
7158 500)) {
7159 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
7160 return -ETIMEDOUT;
7161 }
7162
7163 *val = I915_READ(GEN6_PCODE_DATA);
7164 I915_WRITE(GEN6_PCODE_DATA, 0);
7165
7166 return 0;
7167 }
7168
7169 int sandybridge_pcode_write(struct drm_i915_private *dev_priv, u32 mbox, u32 val)
7170 {
7171 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7172
7173 if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
7174 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
7175 return -EAGAIN;
7176 }
7177
7178 I915_WRITE(GEN6_PCODE_DATA, val);
7179 I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
7180
7181 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
7182 500)) {
7183 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
7184 return -ETIMEDOUT;
7185 }
7186
7187 I915_WRITE(GEN6_PCODE_DATA, 0);
7188
7189 return 0;
7190 }
7191
7192 static int vlv_gpu_freq_div(unsigned int czclk_freq)
7193 {
7194 switch (czclk_freq) {
7195 case 200:
7196 return 10;
7197 case 267:
7198 return 12;
7199 case 320:
7200 case 333:
7201 return 16;
7202 case 400:
7203 return 20;
7204 default:
7205 return -1;
7206 }
7207 }
7208
7209 static int byt_gpu_freq(struct drm_i915_private *dev_priv, int val)
7210 {
7211 int div, czclk_freq = DIV_ROUND_CLOSEST(dev_priv->czclk_freq, 1000);
7212
7213 div = vlv_gpu_freq_div(czclk_freq);
7214 if (div < 0)
7215 return div;
7216
7217 return DIV_ROUND_CLOSEST(czclk_freq * (val + 6 - 0xbd), div);
7218 }
7219
7220 static int byt_freq_opcode(struct drm_i915_private *dev_priv, int val)
7221 {
7222 int mul, czclk_freq = DIV_ROUND_CLOSEST(dev_priv->czclk_freq, 1000);
7223
7224 mul = vlv_gpu_freq_div(czclk_freq);
7225 if (mul < 0)
7226 return mul;
7227
7228 return DIV_ROUND_CLOSEST(mul * val, czclk_freq) + 0xbd - 6;
7229 }
7230
7231 static int chv_gpu_freq(struct drm_i915_private *dev_priv, int val)
7232 {
7233 int div, czclk_freq = DIV_ROUND_CLOSEST(dev_priv->czclk_freq, 1000);
7234
7235 div = vlv_gpu_freq_div(czclk_freq) / 2;
7236 if (div < 0)
7237 return div;
7238
7239 return DIV_ROUND_CLOSEST(czclk_freq * val, 2 * div) / 2;
7240 }
7241
7242 static int chv_freq_opcode(struct drm_i915_private *dev_priv, int val)
7243 {
7244 int mul, czclk_freq = DIV_ROUND_CLOSEST(dev_priv->czclk_freq, 1000);
7245
7246 mul = vlv_gpu_freq_div(czclk_freq) / 2;
7247 if (mul < 0)
7248 return mul;
7249
7250 /* CHV needs even values */
7251 return DIV_ROUND_CLOSEST(val * 2 * mul, czclk_freq) * 2;
7252 }
7253
7254 int intel_gpu_freq(struct drm_i915_private *dev_priv, int val)
7255 {
7256 if (IS_GEN9(dev_priv->dev))
7257 return DIV_ROUND_CLOSEST(val * GT_FREQUENCY_MULTIPLIER,
7258 GEN9_FREQ_SCALER);
7259 else if (IS_CHERRYVIEW(dev_priv->dev))
7260 return chv_gpu_freq(dev_priv, val);
7261 else if (IS_VALLEYVIEW(dev_priv->dev))
7262 return byt_gpu_freq(dev_priv, val);
7263 else
7264 return val * GT_FREQUENCY_MULTIPLIER;
7265 }
7266
7267 int intel_freq_opcode(struct drm_i915_private *dev_priv, int val)
7268 {
7269 if (IS_GEN9(dev_priv->dev))
7270 return DIV_ROUND_CLOSEST(val * GEN9_FREQ_SCALER,
7271 GT_FREQUENCY_MULTIPLIER);
7272 else if (IS_CHERRYVIEW(dev_priv->dev))
7273 return chv_freq_opcode(dev_priv, val);
7274 else if (IS_VALLEYVIEW(dev_priv->dev))
7275 return byt_freq_opcode(dev_priv, val);
7276 else
7277 return DIV_ROUND_CLOSEST(val, GT_FREQUENCY_MULTIPLIER);
7278 }
7279
7280 struct request_boost {
7281 struct work_struct work;
7282 struct drm_i915_gem_request *req;
7283 };
7284
7285 static void __intel_rps_boost_work(struct work_struct *work)
7286 {
7287 struct request_boost *boost = container_of(work, struct request_boost, work);
7288 struct drm_i915_gem_request *req = boost->req;
7289
7290 if (!i915_gem_request_completed(req, true))
7291 gen6_rps_boost(to_i915(req->ring->dev), NULL,
7292 req->emitted_jiffies);
7293
7294 i915_gem_request_unreference__unlocked(req);
7295 kfree(boost);
7296 }
7297
7298 void intel_queue_rps_boost_for_request(struct drm_device *dev,
7299 struct drm_i915_gem_request *req)
7300 {
7301 struct request_boost *boost;
7302
7303 if (req == NULL || INTEL_INFO(dev)->gen < 6)
7304 return;
7305
7306 if (i915_gem_request_completed(req, true))
7307 return;
7308
7309 boost = kmalloc(sizeof(*boost), GFP_ATOMIC);
7310 if (boost == NULL)
7311 return;
7312
7313 i915_gem_request_reference(req);
7314 boost->req = req;
7315
7316 INIT_WORK(&boost->work, __intel_rps_boost_work);
7317 queue_work(to_i915(dev)->wq, &boost->work);
7318 }
7319
7320 void intel_pm_setup(struct drm_device *dev)
7321 {
7322 struct drm_i915_private *dev_priv = dev->dev_private;
7323
7324 mutex_init(&dev_priv->rps.hw_lock);
7325 spin_lock_init(&dev_priv->rps.client_lock);
7326
7327 INIT_DELAYED_WORK(&dev_priv->rps.delayed_resume_work,
7328 intel_gen6_powersave_work);
7329 INIT_LIST_HEAD(&dev_priv->rps.clients);
7330 INIT_LIST_HEAD(&dev_priv->rps.semaphores.link);
7331 INIT_LIST_HEAD(&dev_priv->rps.mmioflips.link);
7332
7333 dev_priv->pm.suspended = false;
7334 }
Linux with added FPC-III support