projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
powerpc: use _GLOBAL_TOC for memmove
[deliverable/linux.git] / drivers / gpu / drm / i915 / intel_pm.c
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 #include <linux/vgaarb.h>
34 #include <drm/i915_powerwell.h>
35 #include <linux/pm_runtime.h>
36
37 /**
38 * RC6 is a special power stage which allows the GPU to enter an very
39 * low-voltage mode when idle, using down to 0V while at this stage. This
40 * stage is entered automatically when the GPU is idle when RC6 support is
41 * enabled, and as soon as new workload arises GPU wakes up automatically as well.
42 *
43 * There are different RC6 modes available in Intel GPU, which differentiate
44 * among each other with the latency required to enter and leave RC6 and
45 * voltage consumed by the GPU in different states.
46 *
47 * The combination of the following flags define which states GPU is allowed
48 * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
49 * RC6pp is deepest RC6. Their support by hardware varies according to the
50 * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
51 * which brings the most power savings; deeper states save more power, but
52 * require higher latency to switch to and wake up.
53 */
54 #define INTEL_RC6_ENABLE (1<<0)
55 #define INTEL_RC6p_ENABLE (1<<1)
56 #define INTEL_RC6pp_ENABLE (1<<2)
57
58 /* FBC, or Frame Buffer Compression, is a technique employed to compress the
59 * framebuffer contents in-memory, aiming at reducing the required bandwidth
60 * during in-memory transfers and, therefore, reduce the power packet.
61 *
62 * The benefits of FBC are mostly visible with solid backgrounds and
63 * variation-less patterns.
64 *
65 * FBC-related functionality can be enabled by the means of the
66 * i915.i915_enable_fbc parameter
67 */
68
69 static void i8xx_disable_fbc(struct drm_device *dev)
70 {
71 struct drm_i915_private *dev_priv = dev->dev_private;
72 u32 fbc_ctl;
73
74 /* Disable compression */
75 fbc_ctl = I915_READ(FBC_CONTROL);
76 if ((fbc_ctl & FBC_CTL_EN) == 0)
77 return;
78
79 fbc_ctl &= ~FBC_CTL_EN;
80 I915_WRITE(FBC_CONTROL, fbc_ctl);
81
82 /* Wait for compressing bit to clear */
83 if (wait_for((I915_READ(FBC_STATUS) & FBC_STAT_COMPRESSING) == 0, 10)) {
84 DRM_DEBUG_KMS("FBC idle timed out\n");
85 return;
86 }
87
88 DRM_DEBUG_KMS("disabled FBC\n");
89 }
90
91 static void i8xx_enable_fbc(struct drm_crtc *crtc)
92 {
93 struct drm_device *dev = crtc->dev;
94 struct drm_i915_private *dev_priv = dev->dev_private;
95 struct drm_framebuffer *fb = crtc->primary->fb;
96 struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
97 struct drm_i915_gem_object *obj = intel_fb->obj;
98 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
99 int cfb_pitch;
100 int i;
101 u32 fbc_ctl;
102
103 cfb_pitch = dev_priv->fbc.size / FBC_LL_SIZE;
104 if (fb->pitches[0] < cfb_pitch)
105 cfb_pitch = fb->pitches[0];
106
107 /* FBC_CTL wants 32B or 64B units */
108 if (IS_GEN2(dev))
109 cfb_pitch = (cfb_pitch / 32) - 1;
110 else
111 cfb_pitch = (cfb_pitch / 64) - 1;
112
113 /* Clear old tags */
114 for (i = 0; i < (FBC_LL_SIZE / 32) + 1; i++)
115 I915_WRITE(FBC_TAG + (i * 4), 0);
116
117 if (IS_GEN4(dev)) {
118 u32 fbc_ctl2;
119
120 /* Set it up... */
121 fbc_ctl2 = FBC_CTL_FENCE_DBL | FBC_CTL_IDLE_IMM | FBC_CTL_CPU_FENCE;
122 fbc_ctl2 |= FBC_CTL_PLANE(intel_crtc->plane);
123 I915_WRITE(FBC_CONTROL2, fbc_ctl2);
124 I915_WRITE(FBC_FENCE_OFF, crtc->y);
125 }
126
127 /* enable it... */
128 fbc_ctl = I915_READ(FBC_CONTROL);
129 fbc_ctl &= 0x3fff << FBC_CTL_INTERVAL_SHIFT;
130 fbc_ctl |= FBC_CTL_EN | FBC_CTL_PERIODIC;
131 if (IS_I945GM(dev))
132 fbc_ctl |= FBC_CTL_C3_IDLE; /* 945 needs special SR handling */
133 fbc_ctl |= (cfb_pitch & 0xff) << FBC_CTL_STRIDE_SHIFT;
134 fbc_ctl |= obj->fence_reg;
135 I915_WRITE(FBC_CONTROL, fbc_ctl);
136
137 DRM_DEBUG_KMS("enabled FBC, pitch %d, yoff %d, plane %c\n",
138 cfb_pitch, crtc->y, plane_name(intel_crtc->plane));
139 }
140
141 static bool i8xx_fbc_enabled(struct drm_device *dev)
142 {
143 struct drm_i915_private *dev_priv = dev->dev_private;
144
145 return I915_READ(FBC_CONTROL) & FBC_CTL_EN;
146 }
147
148 static void g4x_enable_fbc(struct drm_crtc *crtc)
149 {
150 struct drm_device *dev = crtc->dev;
151 struct drm_i915_private *dev_priv = dev->dev_private;
152 struct drm_framebuffer *fb = crtc->primary->fb;
153 struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
154 struct drm_i915_gem_object *obj = intel_fb->obj;
155 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
156 u32 dpfc_ctl;
157
158 dpfc_ctl = DPFC_CTL_PLANE(intel_crtc->plane) | DPFC_SR_EN;
159 if (drm_format_plane_cpp(fb->pixel_format, 0) == 2)
160 dpfc_ctl |= DPFC_CTL_LIMIT_2X;
161 else
162 dpfc_ctl |= DPFC_CTL_LIMIT_1X;
163 dpfc_ctl |= DPFC_CTL_FENCE_EN | obj->fence_reg;
164
165 I915_WRITE(DPFC_FENCE_YOFF, crtc->y);
166
167 /* enable it... */
168 I915_WRITE(DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
169
170 DRM_DEBUG_KMS("enabled fbc on plane %c\n", plane_name(intel_crtc->plane));
171 }
172
173 static void g4x_disable_fbc(struct drm_device *dev)
174 {
175 struct drm_i915_private *dev_priv = dev->dev_private;
176 u32 dpfc_ctl;
177
178 /* Disable compression */
179 dpfc_ctl = I915_READ(DPFC_CONTROL);
180 if (dpfc_ctl & DPFC_CTL_EN) {
181 dpfc_ctl &= ~DPFC_CTL_EN;
182 I915_WRITE(DPFC_CONTROL, dpfc_ctl);
183
184 DRM_DEBUG_KMS("disabled FBC\n");
185 }
186 }
187
188 static bool g4x_fbc_enabled(struct drm_device *dev)
189 {
190 struct drm_i915_private *dev_priv = dev->dev_private;
191
192 return I915_READ(DPFC_CONTROL) & DPFC_CTL_EN;
193 }
194
195 static void sandybridge_blit_fbc_update(struct drm_device *dev)
196 {
197 struct drm_i915_private *dev_priv = dev->dev_private;
198 u32 blt_ecoskpd;
199
200 /* Make sure blitter notifies FBC of writes */
201
202 /* Blitter is part of Media powerwell on VLV. No impact of
203 * his param in other platforms for now */
204 gen6_gt_force_wake_get(dev_priv, FORCEWAKE_MEDIA);
205
206 blt_ecoskpd = I915_READ(GEN6_BLITTER_ECOSKPD);
207 blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY <<
208 GEN6_BLITTER_LOCK_SHIFT;
209 I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
210 blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY;
211 I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
212 blt_ecoskpd &= ~(GEN6_BLITTER_FBC_NOTIFY <<
213 GEN6_BLITTER_LOCK_SHIFT);
214 I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
215 POSTING_READ(GEN6_BLITTER_ECOSKPD);
216
217 gen6_gt_force_wake_put(dev_priv, FORCEWAKE_MEDIA);
218 }
219
220 static void ironlake_enable_fbc(struct drm_crtc *crtc)
221 {
222 struct drm_device *dev = crtc->dev;
223 struct drm_i915_private *dev_priv = dev->dev_private;
224 struct drm_framebuffer *fb = crtc->primary->fb;
225 struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
226 struct drm_i915_gem_object *obj = intel_fb->obj;
227 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
228 u32 dpfc_ctl;
229
230 dpfc_ctl = DPFC_CTL_PLANE(intel_crtc->plane);
231 if (drm_format_plane_cpp(fb->pixel_format, 0) == 2)
232 dpfc_ctl |= DPFC_CTL_LIMIT_2X;
233 else
234 dpfc_ctl |= DPFC_CTL_LIMIT_1X;
235 dpfc_ctl |= DPFC_CTL_FENCE_EN;
236 if (IS_GEN5(dev))
237 dpfc_ctl |= obj->fence_reg;
238
239 I915_WRITE(ILK_DPFC_FENCE_YOFF, crtc->y);
240 I915_WRITE(ILK_FBC_RT_BASE, i915_gem_obj_ggtt_offset(obj) | ILK_FBC_RT_VALID);
241 /* enable it... */
242 I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
243
244 if (IS_GEN6(dev)) {
245 I915_WRITE(SNB_DPFC_CTL_SA,
246 SNB_CPU_FENCE_ENABLE | obj->fence_reg);
247 I915_WRITE(DPFC_CPU_FENCE_OFFSET, crtc->y);
248 sandybridge_blit_fbc_update(dev);
249 }
250
251 DRM_DEBUG_KMS("enabled fbc on plane %c\n", plane_name(intel_crtc->plane));
252 }
253
254 static void ironlake_disable_fbc(struct drm_device *dev)
255 {
256 struct drm_i915_private *dev_priv = dev->dev_private;
257 u32 dpfc_ctl;
258
259 /* Disable compression */
260 dpfc_ctl = I915_READ(ILK_DPFC_CONTROL);
261 if (dpfc_ctl & DPFC_CTL_EN) {
262 dpfc_ctl &= ~DPFC_CTL_EN;
263 I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl);
264
265 DRM_DEBUG_KMS("disabled FBC\n");
266 }
267 }
268
269 static bool ironlake_fbc_enabled(struct drm_device *dev)
270 {
271 struct drm_i915_private *dev_priv = dev->dev_private;
272
273 return I915_READ(ILK_DPFC_CONTROL) & DPFC_CTL_EN;
274 }
275
276 static void gen7_enable_fbc(struct drm_crtc *crtc)
277 {
278 struct drm_device *dev = crtc->dev;
279 struct drm_i915_private *dev_priv = dev->dev_private;
280 struct drm_framebuffer *fb = crtc->primary->fb;
281 struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
282 struct drm_i915_gem_object *obj = intel_fb->obj;
283 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
284 u32 dpfc_ctl;
285
286 dpfc_ctl = IVB_DPFC_CTL_PLANE(intel_crtc->plane);
287 if (drm_format_plane_cpp(fb->pixel_format, 0) == 2)
288 dpfc_ctl |= DPFC_CTL_LIMIT_2X;
289 else
290 dpfc_ctl |= DPFC_CTL_LIMIT_1X;
291 dpfc_ctl |= IVB_DPFC_CTL_FENCE_EN;
292
293 I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
294
295 if (IS_IVYBRIDGE(dev)) {
296 /* WaFbcAsynchFlipDisableFbcQueue:ivb */
297 I915_WRITE(ILK_DISPLAY_CHICKEN1,
298 I915_READ(ILK_DISPLAY_CHICKEN1) |
299 ILK_FBCQ_DIS);
300 } else {
301 /* WaFbcAsynchFlipDisableFbcQueue:hsw,bdw */
302 I915_WRITE(CHICKEN_PIPESL_1(intel_crtc->pipe),
303 I915_READ(CHICKEN_PIPESL_1(intel_crtc->pipe)) |
304 HSW_FBCQ_DIS);
305 }
306
307 I915_WRITE(SNB_DPFC_CTL_SA,
308 SNB_CPU_FENCE_ENABLE | obj->fence_reg);
309 I915_WRITE(DPFC_CPU_FENCE_OFFSET, crtc->y);
310
311 sandybridge_blit_fbc_update(dev);
312
313 DRM_DEBUG_KMS("enabled fbc on plane %c\n", plane_name(intel_crtc->plane));
314 }
315
316 bool intel_fbc_enabled(struct drm_device *dev)
317 {
318 struct drm_i915_private *dev_priv = dev->dev_private;
319
320 if (!dev_priv->display.fbc_enabled)
321 return false;
322
323 return dev_priv->display.fbc_enabled(dev);
324 }
325
326 static void intel_fbc_work_fn(struct work_struct *__work)
327 {
328 struct intel_fbc_work *work =
329 container_of(to_delayed_work(__work),
330 struct intel_fbc_work, work);
331 struct drm_device *dev = work->crtc->dev;
332 struct drm_i915_private *dev_priv = dev->dev_private;
333
334 mutex_lock(&dev->struct_mutex);
335 if (work == dev_priv->fbc.fbc_work) {
336 /* Double check that we haven't switched fb without cancelling
337 * the prior work.
338 */
339 if (work->crtc->primary->fb == work->fb) {
340 dev_priv->display.enable_fbc(work->crtc);
341
342 dev_priv->fbc.plane = to_intel_crtc(work->crtc)->plane;
343 dev_priv->fbc.fb_id = work->crtc->primary->fb->base.id;
344 dev_priv->fbc.y = work->crtc->y;
345 }
346
347 dev_priv->fbc.fbc_work = NULL;
348 }
349 mutex_unlock(&dev->struct_mutex);
350
351 kfree(work);
352 }
353
354 static void intel_cancel_fbc_work(struct drm_i915_private *dev_priv)
355 {
356 if (dev_priv->fbc.fbc_work == NULL)
357 return;
358
359 DRM_DEBUG_KMS("cancelling pending FBC enable\n");
360
361 /* Synchronisation is provided by struct_mutex and checking of
362 * dev_priv->fbc.fbc_work, so we can perform the cancellation
363 * entirely asynchronously.
364 */
365 if (cancel_delayed_work(&dev_priv->fbc.fbc_work->work))
366 /* tasklet was killed before being run, clean up */
367 kfree(dev_priv->fbc.fbc_work);
368
369 /* Mark the work as no longer wanted so that if it does
370 * wake-up (because the work was already running and waiting
371 * for our mutex), it will discover that is no longer
372 * necessary to run.
373 */
374 dev_priv->fbc.fbc_work = NULL;
375 }
376
377 static void intel_enable_fbc(struct drm_crtc *crtc)
378 {
379 struct intel_fbc_work *work;
380 struct drm_device *dev = crtc->dev;
381 struct drm_i915_private *dev_priv = dev->dev_private;
382
383 if (!dev_priv->display.enable_fbc)
384 return;
385
386 intel_cancel_fbc_work(dev_priv);
387
388 work = kzalloc(sizeof(*work), GFP_KERNEL);
389 if (work == NULL) {
390 DRM_ERROR("Failed to allocate FBC work structure\n");
391 dev_priv->display.enable_fbc(crtc);
392 return;
393 }
394
395 work->crtc = crtc;
396 work->fb = crtc->primary->fb;
397 INIT_DELAYED_WORK(&work->work, intel_fbc_work_fn);
398
399 dev_priv->fbc.fbc_work = work;
400
401 /* Delay the actual enabling to let pageflipping cease and the
402 * display to settle before starting the compression. Note that
403 * this delay also serves a second purpose: it allows for a
404 * vblank to pass after disabling the FBC before we attempt
405 * to modify the control registers.
406 *
407 * A more complicated solution would involve tracking vblanks
408 * following the termination of the page-flipping sequence
409 * and indeed performing the enable as a co-routine and not
410 * waiting synchronously upon the vblank.
411 *
412 * WaFbcWaitForVBlankBeforeEnable:ilk,snb
413 */
414 schedule_delayed_work(&work->work, msecs_to_jiffies(50));
415 }
416
417 void intel_disable_fbc(struct drm_device *dev)
418 {
419 struct drm_i915_private *dev_priv = dev->dev_private;
420
421 intel_cancel_fbc_work(dev_priv);
422
423 if (!dev_priv->display.disable_fbc)
424 return;
425
426 dev_priv->display.disable_fbc(dev);
427 dev_priv->fbc.plane = -1;
428 }
429
430 static bool set_no_fbc_reason(struct drm_i915_private *dev_priv,
431 enum no_fbc_reason reason)
432 {
433 if (dev_priv->fbc.no_fbc_reason == reason)
434 return false;
435
436 dev_priv->fbc.no_fbc_reason = reason;
437 return true;
438 }
439
440 /**
441 * intel_update_fbc - enable/disable FBC as needed
442 * @dev: the drm_device
443 *
444 * Set up the framebuffer compression hardware at mode set time. We
445 * enable it if possible:
446 * - plane A only (on pre-965)
447 * - no pixel mulitply/line duplication
448 * - no alpha buffer discard
449 * - no dual wide
450 * - framebuffer <= max_hdisplay in width, max_vdisplay in height
451 *
452 * We can't assume that any compression will take place (worst case),
453 * so the compressed buffer has to be the same size as the uncompressed
454 * one. It also must reside (along with the line length buffer) in
455 * stolen memory.
456 *
457 * We need to enable/disable FBC on a global basis.
458 */
459 void intel_update_fbc(struct drm_device *dev)
460 {
461 struct drm_i915_private *dev_priv = dev->dev_private;
462 struct drm_crtc *crtc = NULL, *tmp_crtc;
463 struct intel_crtc *intel_crtc;
464 struct drm_framebuffer *fb;
465 struct intel_framebuffer *intel_fb;
466 struct drm_i915_gem_object *obj;
467 const struct drm_display_mode *adjusted_mode;
468 unsigned int max_width, max_height;
469
470 if (!HAS_FBC(dev)) {
471 set_no_fbc_reason(dev_priv, FBC_UNSUPPORTED);
472 return;
473 }
474
475 if (!i915.powersave) {
476 if (set_no_fbc_reason(dev_priv, FBC_MODULE_PARAM))
477 DRM_DEBUG_KMS("fbc disabled per module param\n");
478 return;
479 }
480
481 /*
482 * If FBC is already on, we just have to verify that we can
483 * keep it that way...
484 * Need to disable if:
485 * - more than one pipe is active
486 * - changing FBC params (stride, fence, mode)
487 * - new fb is too large to fit in compressed buffer
488 * - going to an unsupported config (interlace, pixel multiply, etc.)
489 */
490 for_each_crtc(dev, tmp_crtc) {
491 if (intel_crtc_active(tmp_crtc) &&
492 to_intel_crtc(tmp_crtc)->primary_enabled) {
493 if (crtc) {
494 if (set_no_fbc_reason(dev_priv, FBC_MULTIPLE_PIPES))
495 DRM_DEBUG_KMS("more than one pipe active, disabling compression\n");
496 goto out_disable;
497 }
498 crtc = tmp_crtc;
499 }
500 }
501
502 if (!crtc || crtc->primary->fb == NULL) {
503 if (set_no_fbc_reason(dev_priv, FBC_NO_OUTPUT))
504 DRM_DEBUG_KMS("no output, disabling\n");
505 goto out_disable;
506 }
507
508 intel_crtc = to_intel_crtc(crtc);
509 fb = crtc->primary->fb;
510 intel_fb = to_intel_framebuffer(fb);
511 obj = intel_fb->obj;
512 adjusted_mode = &intel_crtc->config.adjusted_mode;
513
514 if (i915.enable_fbc < 0 &&
515 INTEL_INFO(dev)->gen <= 7 && !IS_HASWELL(dev)) {
516 if (set_no_fbc_reason(dev_priv, FBC_CHIP_DEFAULT))
517 DRM_DEBUG_KMS("disabled per chip default\n");
518 goto out_disable;
519 }
520 if (!i915.enable_fbc) {
521 if (set_no_fbc_reason(dev_priv, FBC_MODULE_PARAM))
522 DRM_DEBUG_KMS("fbc disabled per module param\n");
523 goto out_disable;
524 }
525 if ((adjusted_mode->flags & DRM_MODE_FLAG_INTERLACE) ||
526 (adjusted_mode->flags & DRM_MODE_FLAG_DBLSCAN)) {
527 if (set_no_fbc_reason(dev_priv, FBC_UNSUPPORTED_MODE))
528 DRM_DEBUG_KMS("mode incompatible with compression, "
529 "disabling\n");
530 goto out_disable;
531 }
532
533 if (IS_G4X(dev) || INTEL_INFO(dev)->gen >= 5) {
534 max_width = 4096;
535 max_height = 2048;
536 } else {
537 max_width = 2048;
538 max_height = 1536;
539 }
540 if (intel_crtc->config.pipe_src_w > max_width ||
541 intel_crtc->config.pipe_src_h > max_height) {
542 if (set_no_fbc_reason(dev_priv, FBC_MODE_TOO_LARGE))
543 DRM_DEBUG_KMS("mode too large for compression, disabling\n");
544 goto out_disable;
545 }
546 if ((INTEL_INFO(dev)->gen < 4 || HAS_DDI(dev)) &&
547 intel_crtc->plane != PLANE_A) {
548 if (set_no_fbc_reason(dev_priv, FBC_BAD_PLANE))
549 DRM_DEBUG_KMS("plane not A, disabling compression\n");
550 goto out_disable;
551 }
552
553 /* The use of a CPU fence is mandatory in order to detect writes
554 * by the CPU to the scanout and trigger updates to the FBC.
555 */
556 if (obj->tiling_mode != I915_TILING_X ||
557 obj->fence_reg == I915_FENCE_REG_NONE) {
558 if (set_no_fbc_reason(dev_priv, FBC_NOT_TILED))
559 DRM_DEBUG_KMS("framebuffer not tiled or fenced, disabling compression\n");
560 goto out_disable;
561 }
562
563 /* If the kernel debugger is active, always disable compression */
564 if (in_dbg_master())
565 goto out_disable;
566
567 if (i915_gem_stolen_setup_compression(dev, intel_fb->obj->base.size)) {
568 if (set_no_fbc_reason(dev_priv, FBC_STOLEN_TOO_SMALL))
569 DRM_DEBUG_KMS("framebuffer too large, disabling compression\n");
570 goto out_disable;
571 }
572
573 /* If the scanout has not changed, don't modify the FBC settings.
574 * Note that we make the fundamental assumption that the fb->obj
575 * cannot be unpinned (and have its GTT offset and fence revoked)
576 * without first being decoupled from the scanout and FBC disabled.
577 */
578 if (dev_priv->fbc.plane == intel_crtc->plane &&
579 dev_priv->fbc.fb_id == fb->base.id &&
580 dev_priv->fbc.y == crtc->y)
581 return;
582
583 if (intel_fbc_enabled(dev)) {
584 /* We update FBC along two paths, after changing fb/crtc
585 * configuration (modeswitching) and after page-flipping
586 * finishes. For the latter, we know that not only did
587 * we disable the FBC at the start of the page-flip
588 * sequence, but also more than one vblank has passed.
589 *
590 * For the former case of modeswitching, it is possible
591 * to switch between two FBC valid configurations
592 * instantaneously so we do need to disable the FBC
593 * before we can modify its control registers. We also
594 * have to wait for the next vblank for that to take
595 * effect. However, since we delay enabling FBC we can
596 * assume that a vblank has passed since disabling and
597 * that we can safely alter the registers in the deferred
598 * callback.
599 *
600 * In the scenario that we go from a valid to invalid
601 * and then back to valid FBC configuration we have
602 * no strict enforcement that a vblank occurred since
603 * disabling the FBC. However, along all current pipe
604 * disabling paths we do need to wait for a vblank at
605 * some point. And we wait before enabling FBC anyway.
606 */
607 DRM_DEBUG_KMS("disabling active FBC for update\n");
608 intel_disable_fbc(dev);
609 }
610
611 intel_enable_fbc(crtc);
612 dev_priv->fbc.no_fbc_reason = FBC_OK;
613 return;
614
615 out_disable:
616 /* Multiple disables should be harmless */
617 if (intel_fbc_enabled(dev)) {
618 DRM_DEBUG_KMS("unsupported config, disabling FBC\n");
619 intel_disable_fbc(dev);
620 }
621 i915_gem_stolen_cleanup_compression(dev);
622 }
623
624 static void i915_pineview_get_mem_freq(struct drm_device *dev)
625 {
626 struct drm_i915_private *dev_priv = dev->dev_private;
627 u32 tmp;
628
629 tmp = I915_READ(CLKCFG);
630
631 switch (tmp & CLKCFG_FSB_MASK) {
632 case CLKCFG_FSB_533:
633 dev_priv->fsb_freq = 533; /* 133*4 */
634 break;
635 case CLKCFG_FSB_800:
636 dev_priv->fsb_freq = 800; /* 200*4 */
637 break;
638 case CLKCFG_FSB_667:
639 dev_priv->fsb_freq = 667; /* 167*4 */
640 break;
641 case CLKCFG_FSB_400:
642 dev_priv->fsb_freq = 400; /* 100*4 */
643 break;
644 }
645
646 switch (tmp & CLKCFG_MEM_MASK) {
647 case CLKCFG_MEM_533:
648 dev_priv->mem_freq = 533;
649 break;
650 case CLKCFG_MEM_667:
651 dev_priv->mem_freq = 667;
652 break;
653 case CLKCFG_MEM_800:
654 dev_priv->mem_freq = 800;
655 break;
656 }
657
658 /* detect pineview DDR3 setting */
659 tmp = I915_READ(CSHRDDR3CTL);
660 dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
661 }
662
663 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
664 {
665 struct drm_i915_private *dev_priv = dev->dev_private;
666 u16 ddrpll, csipll;
667
668 ddrpll = I915_READ16(DDRMPLL1);
669 csipll = I915_READ16(CSIPLL0);
670
671 switch (ddrpll & 0xff) {
672 case 0xc:
673 dev_priv->mem_freq = 800;
674 break;
675 case 0x10:
676 dev_priv->mem_freq = 1066;
677 break;
678 case 0x14:
679 dev_priv->mem_freq = 1333;
680 break;
681 case 0x18:
682 dev_priv->mem_freq = 1600;
683 break;
684 default:
685 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
686 ddrpll & 0xff);
687 dev_priv->mem_freq = 0;
688 break;
689 }
690
691 dev_priv->ips.r_t = dev_priv->mem_freq;
692
693 switch (csipll & 0x3ff) {
694 case 0x00c:
695 dev_priv->fsb_freq = 3200;
696 break;
697 case 0x00e:
698 dev_priv->fsb_freq = 3733;
699 break;
700 case 0x010:
701 dev_priv->fsb_freq = 4266;
702 break;
703 case 0x012:
704 dev_priv->fsb_freq = 4800;
705 break;
706 case 0x014:
707 dev_priv->fsb_freq = 5333;
708 break;
709 case 0x016:
710 dev_priv->fsb_freq = 5866;
711 break;
712 case 0x018:
713 dev_priv->fsb_freq = 6400;
714 break;
715 default:
716 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
717 csipll & 0x3ff);
718 dev_priv->fsb_freq = 0;
719 break;
720 }
721
722 if (dev_priv->fsb_freq == 3200) {
723 dev_priv->ips.c_m = 0;
724 } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
725 dev_priv->ips.c_m = 1;
726 } else {
727 dev_priv->ips.c_m = 2;
728 }
729 }
730
731 static const struct cxsr_latency cxsr_latency_table[] = {
732 {1, 0, 800, 400, 3382, 33382, 3983, 33983}, /* DDR2-400 SC */
733 {1, 0, 800, 667, 3354, 33354, 3807, 33807}, /* DDR2-667 SC */
734 {1, 0, 800, 800, 3347, 33347, 3763, 33763}, /* DDR2-800 SC */
735 {1, 1, 800, 667, 6420, 36420, 6873, 36873}, /* DDR3-667 SC */
736 {1, 1, 800, 800, 5902, 35902, 6318, 36318}, /* DDR3-800 SC */
737
738 {1, 0, 667, 400, 3400, 33400, 4021, 34021}, /* DDR2-400 SC */
739 {1, 0, 667, 667, 3372, 33372, 3845, 33845}, /* DDR2-667 SC */
740 {1, 0, 667, 800, 3386, 33386, 3822, 33822}, /* DDR2-800 SC */
741 {1, 1, 667, 667, 6438, 36438, 6911, 36911}, /* DDR3-667 SC */
742 {1, 1, 667, 800, 5941, 35941, 6377, 36377}, /* DDR3-800 SC */
743
744 {1, 0, 400, 400, 3472, 33472, 4173, 34173}, /* DDR2-400 SC */
745 {1, 0, 400, 667, 3443, 33443, 3996, 33996}, /* DDR2-667 SC */
746 {1, 0, 400, 800, 3430, 33430, 3946, 33946}, /* DDR2-800 SC */
747 {1, 1, 400, 667, 6509, 36509, 7062, 37062}, /* DDR3-667 SC */
748 {1, 1, 400, 800, 5985, 35985, 6501, 36501}, /* DDR3-800 SC */
749
750 {0, 0, 800, 400, 3438, 33438, 4065, 34065}, /* DDR2-400 SC */
751 {0, 0, 800, 667, 3410, 33410, 3889, 33889}, /* DDR2-667 SC */
752 {0, 0, 800, 800, 3403, 33403, 3845, 33845}, /* DDR2-800 SC */
753 {0, 1, 800, 667, 6476, 36476, 6955, 36955}, /* DDR3-667 SC */
754 {0, 1, 800, 800, 5958, 35958, 6400, 36400}, /* DDR3-800 SC */
755
756 {0, 0, 667, 400, 3456, 33456, 4103, 34106}, /* DDR2-400 SC */
757 {0, 0, 667, 667, 3428, 33428, 3927, 33927}, /* DDR2-667 SC */
758 {0, 0, 667, 800, 3443, 33443, 3905, 33905}, /* DDR2-800 SC */
759 {0, 1, 667, 667, 6494, 36494, 6993, 36993}, /* DDR3-667 SC */
760 {0, 1, 667, 800, 5998, 35998, 6460, 36460}, /* DDR3-800 SC */
761
762 {0, 0, 400, 400, 3528, 33528, 4255, 34255}, /* DDR2-400 SC */
763 {0, 0, 400, 667, 3500, 33500, 4079, 34079}, /* DDR2-667 SC */
764 {0, 0, 400, 800, 3487, 33487, 4029, 34029}, /* DDR2-800 SC */
765 {0, 1, 400, 667, 6566, 36566, 7145, 37145}, /* DDR3-667 SC */
766 {0, 1, 400, 800, 6042, 36042, 6584, 36584}, /* DDR3-800 SC */
767 };
768
769 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
770 int is_ddr3,
771 int fsb,
772 int mem)
773 {
774 const struct cxsr_latency *latency;
775 int i;
776
777 if (fsb == 0 || mem == 0)
778 return NULL;
779
780 for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
781 latency = &cxsr_latency_table[i];
782 if (is_desktop == latency->is_desktop &&
783 is_ddr3 == latency->is_ddr3 &&
784 fsb == latency->fsb_freq && mem == latency->mem_freq)
785 return latency;
786 }
787
788 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
789
790 return NULL;
791 }
792
793 static void pineview_disable_cxsr(struct drm_device *dev)
794 {
795 struct drm_i915_private *dev_priv = dev->dev_private;
796
797 /* deactivate cxsr */
798 I915_WRITE(DSPFW3, I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN);
799 }
800
801 /*
802 * Latency for FIFO fetches is dependent on several factors:
803 * - memory configuration (speed, channels)
804 * - chipset
805 * - current MCH state
806 * It can be fairly high in some situations, so here we assume a fairly
807 * pessimal value. It's a tradeoff between extra memory fetches (if we
808 * set this value too high, the FIFO will fetch frequently to stay full)
809 * and power consumption (set it too low to save power and we might see
810 * FIFO underruns and display "flicker").
811 *
812 * A value of 5us seems to be a good balance; safe for very low end
813 * platforms but not overly aggressive on lower latency configs.
814 */
815 static const int latency_ns = 5000;
816
817 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
818 {
819 struct drm_i915_private *dev_priv = dev->dev_private;
820 uint32_t dsparb = I915_READ(DSPARB);
821 int size;
822
823 size = dsparb & 0x7f;
824 if (plane)
825 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
826
827 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
828 plane ? "B" : "A", size);
829
830 return size;
831 }
832
833 static int i830_get_fifo_size(struct drm_device *dev, int plane)
834 {
835 struct drm_i915_private *dev_priv = dev->dev_private;
836 uint32_t dsparb = I915_READ(DSPARB);
837 int size;
838
839 size = dsparb & 0x1ff;
840 if (plane)
841 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
842 size >>= 1; /* Convert to cachelines */
843
844 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
845 plane ? "B" : "A", size);
846
847 return size;
848 }
849
850 static int i845_get_fifo_size(struct drm_device *dev, int plane)
851 {
852 struct drm_i915_private *dev_priv = dev->dev_private;
853 uint32_t dsparb = I915_READ(DSPARB);
854 int size;
855
856 size = dsparb & 0x7f;
857 size >>= 2; /* Convert to cachelines */
858
859 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
860 plane ? "B" : "A",
861 size);
862
863 return size;
864 }
865
866 /* Pineview has different values for various configs */
867 static const struct intel_watermark_params pineview_display_wm = {
868 PINEVIEW_DISPLAY_FIFO,
869 PINEVIEW_MAX_WM,
870 PINEVIEW_DFT_WM,
871 PINEVIEW_GUARD_WM,
872 PINEVIEW_FIFO_LINE_SIZE
873 };
874 static const struct intel_watermark_params pineview_display_hplloff_wm = {
875 PINEVIEW_DISPLAY_FIFO,
876 PINEVIEW_MAX_WM,
877 PINEVIEW_DFT_HPLLOFF_WM,
878 PINEVIEW_GUARD_WM,
879 PINEVIEW_FIFO_LINE_SIZE
880 };
881 static const struct intel_watermark_params pineview_cursor_wm = {
882 PINEVIEW_CURSOR_FIFO,
883 PINEVIEW_CURSOR_MAX_WM,
884 PINEVIEW_CURSOR_DFT_WM,
885 PINEVIEW_CURSOR_GUARD_WM,
886 PINEVIEW_FIFO_LINE_SIZE,
887 };
888 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
889 PINEVIEW_CURSOR_FIFO,
890 PINEVIEW_CURSOR_MAX_WM,
891 PINEVIEW_CURSOR_DFT_WM,
892 PINEVIEW_CURSOR_GUARD_WM,
893 PINEVIEW_FIFO_LINE_SIZE
894 };
895 static const struct intel_watermark_params g4x_wm_info = {
896 G4X_FIFO_SIZE,
897 G4X_MAX_WM,
898 G4X_MAX_WM,
899 2,
900 G4X_FIFO_LINE_SIZE,
901 };
902 static const struct intel_watermark_params g4x_cursor_wm_info = {
903 I965_CURSOR_FIFO,
904 I965_CURSOR_MAX_WM,
905 I965_CURSOR_DFT_WM,
906 2,
907 G4X_FIFO_LINE_SIZE,
908 };
909 static const struct intel_watermark_params valleyview_wm_info = {
910 VALLEYVIEW_FIFO_SIZE,
911 VALLEYVIEW_MAX_WM,
912 VALLEYVIEW_MAX_WM,
913 2,
914 G4X_FIFO_LINE_SIZE,
915 };
916 static const struct intel_watermark_params valleyview_cursor_wm_info = {
917 I965_CURSOR_FIFO,
918 VALLEYVIEW_CURSOR_MAX_WM,
919 I965_CURSOR_DFT_WM,
920 2,
921 G4X_FIFO_LINE_SIZE,
922 };
923 static const struct intel_watermark_params i965_cursor_wm_info = {
924 I965_CURSOR_FIFO,
925 I965_CURSOR_MAX_WM,
926 I965_CURSOR_DFT_WM,
927 2,
928 I915_FIFO_LINE_SIZE,
929 };
930 static const struct intel_watermark_params i945_wm_info = {
931 I945_FIFO_SIZE,
932 I915_MAX_WM,
933 1,
934 2,
935 I915_FIFO_LINE_SIZE
936 };
937 static const struct intel_watermark_params i915_wm_info = {
938 I915_FIFO_SIZE,
939 I915_MAX_WM,
940 1,
941 2,
942 I915_FIFO_LINE_SIZE
943 };
944 static const struct intel_watermark_params i830_wm_info = {
945 I855GM_FIFO_SIZE,
946 I915_MAX_WM,
947 1,
948 2,
949 I830_FIFO_LINE_SIZE
950 };
951 static const struct intel_watermark_params i845_wm_info = {
952 I830_FIFO_SIZE,
953 I915_MAX_WM,
954 1,
955 2,
956 I830_FIFO_LINE_SIZE
957 };
958
959 /**
960 * intel_calculate_wm - calculate watermark level
961 * @clock_in_khz: pixel clock
962 * @wm: chip FIFO params
963 * @pixel_size: display pixel size
964 * @latency_ns: memory latency for the platform
965 *
966 * Calculate the watermark level (the level at which the display plane will
967 * start fetching from memory again). Each chip has a different display
968 * FIFO size and allocation, so the caller needs to figure that out and pass
969 * in the correct intel_watermark_params structure.
970 *
971 * As the pixel clock runs, the FIFO will be drained at a rate that depends
972 * on the pixel size. When it reaches the watermark level, it'll start
973 * fetching FIFO line sized based chunks from memory until the FIFO fills
974 * past the watermark point. If the FIFO drains completely, a FIFO underrun
975 * will occur, and a display engine hang could result.
976 */
977 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
978 const struct intel_watermark_params *wm,
979 int fifo_size,
980 int pixel_size,
981 unsigned long latency_ns)
982 {
983 long entries_required, wm_size;
984
985 /*
986 * Note: we need to make sure we don't overflow for various clock &
987 * latency values.
988 * clocks go from a few thousand to several hundred thousand.
989 * latency is usually a few thousand
990 */
991 entries_required = ((clock_in_khz / 1000) * pixel_size * latency_ns) /
992 1000;
993 entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
994
995 DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
996
997 wm_size = fifo_size - (entries_required + wm->guard_size);
998
999 DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
1000
1001 /* Don't promote wm_size to unsigned... */
1002 if (wm_size > (long)wm->max_wm)
1003 wm_size = wm->max_wm;
1004 if (wm_size <= 0)
1005 wm_size = wm->default_wm;
1006 return wm_size;
1007 }
1008
1009 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
1010 {
1011 struct drm_crtc *crtc, *enabled = NULL;
1012
1013 for_each_crtc(dev, crtc) {
1014 if (intel_crtc_active(crtc)) {
1015 if (enabled)
1016 return NULL;
1017 enabled = crtc;
1018 }
1019 }
1020
1021 return enabled;
1022 }
1023
1024 static void pineview_update_wm(struct drm_crtc *unused_crtc)
1025 {
1026 struct drm_device *dev = unused_crtc->dev;
1027 struct drm_i915_private *dev_priv = dev->dev_private;
1028 struct drm_crtc *crtc;
1029 const struct cxsr_latency *latency;
1030 u32 reg;
1031 unsigned long wm;
1032
1033 latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
1034 dev_priv->fsb_freq, dev_priv->mem_freq);
1035 if (!latency) {
1036 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
1037 pineview_disable_cxsr(dev);
1038 return;
1039 }
1040
1041 crtc = single_enabled_crtc(dev);
1042 if (crtc) {
1043 const struct drm_display_mode *adjusted_mode;
1044 int pixel_size = crtc->primary->fb->bits_per_pixel / 8;
1045 int clock;
1046
1047 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1048 clock = adjusted_mode->crtc_clock;
1049
1050 /* Display SR */
1051 wm = intel_calculate_wm(clock, &pineview_display_wm,
1052 pineview_display_wm.fifo_size,
1053 pixel_size, latency->display_sr);
1054 reg = I915_READ(DSPFW1);
1055 reg &= ~DSPFW_SR_MASK;
1056 reg |= wm << DSPFW_SR_SHIFT;
1057 I915_WRITE(DSPFW1, reg);
1058 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
1059
1060 /* cursor SR */
1061 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
1062 pineview_display_wm.fifo_size,
1063 pixel_size, latency->cursor_sr);
1064 reg = I915_READ(DSPFW3);
1065 reg &= ~DSPFW_CURSOR_SR_MASK;
1066 reg |= (wm & 0x3f) << DSPFW_CURSOR_SR_SHIFT;
1067 I915_WRITE(DSPFW3, reg);
1068
1069 /* Display HPLL off SR */
1070 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
1071 pineview_display_hplloff_wm.fifo_size,
1072 pixel_size, latency->display_hpll_disable);
1073 reg = I915_READ(DSPFW3);
1074 reg &= ~DSPFW_HPLL_SR_MASK;
1075 reg |= wm & DSPFW_HPLL_SR_MASK;
1076 I915_WRITE(DSPFW3, reg);
1077
1078 /* cursor HPLL off SR */
1079 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
1080 pineview_display_hplloff_wm.fifo_size,
1081 pixel_size, latency->cursor_hpll_disable);
1082 reg = I915_READ(DSPFW3);
1083 reg &= ~DSPFW_HPLL_CURSOR_MASK;
1084 reg |= (wm & 0x3f) << DSPFW_HPLL_CURSOR_SHIFT;
1085 I915_WRITE(DSPFW3, reg);
1086 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
1087
1088 /* activate cxsr */
1089 I915_WRITE(DSPFW3,
1090 I915_READ(DSPFW3) | PINEVIEW_SELF_REFRESH_EN);
1091 DRM_DEBUG_KMS("Self-refresh is enabled\n");
1092 } else {
1093 pineview_disable_cxsr(dev);
1094 DRM_DEBUG_KMS("Self-refresh is disabled\n");
1095 }
1096 }
1097
1098 static bool g4x_compute_wm0(struct drm_device *dev,
1099 int plane,
1100 const struct intel_watermark_params *display,
1101 int display_latency_ns,
1102 const struct intel_watermark_params *cursor,
1103 int cursor_latency_ns,
1104 int *plane_wm,
1105 int *cursor_wm)
1106 {
1107 struct drm_crtc *crtc;
1108 const struct drm_display_mode *adjusted_mode;
1109 int htotal, hdisplay, clock, pixel_size;
1110 int line_time_us, line_count;
1111 int entries, tlb_miss;
1112
1113 crtc = intel_get_crtc_for_plane(dev, plane);
1114 if (!intel_crtc_active(crtc)) {
1115 *cursor_wm = cursor->guard_size;
1116 *plane_wm = display->guard_size;
1117 return false;
1118 }
1119
1120 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1121 clock = adjusted_mode->crtc_clock;
1122 htotal = adjusted_mode->crtc_htotal;
1123 hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1124 pixel_size = crtc->primary->fb->bits_per_pixel / 8;
1125
1126 /* Use the small buffer method to calculate plane watermark */
1127 entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
1128 tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
1129 if (tlb_miss > 0)
1130 entries += tlb_miss;
1131 entries = DIV_ROUND_UP(entries, display->cacheline_size);
1132 *plane_wm = entries + display->guard_size;
1133 if (*plane_wm > (int)display->max_wm)
1134 *plane_wm = display->max_wm;
1135
1136 /* Use the large buffer method to calculate cursor watermark */
1137 line_time_us = max(htotal * 1000 / clock, 1);
1138 line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
1139 entries = line_count * to_intel_crtc(crtc)->cursor_width * pixel_size;
1140 tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
1141 if (tlb_miss > 0)
1142 entries += tlb_miss;
1143 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1144 *cursor_wm = entries + cursor->guard_size;
1145 if (*cursor_wm > (int)cursor->max_wm)
1146 *cursor_wm = (int)cursor->max_wm;
1147
1148 return true;
1149 }
1150
1151 /*
1152 * Check the wm result.
1153 *
1154 * If any calculated watermark values is larger than the maximum value that
1155 * can be programmed into the associated watermark register, that watermark
1156 * must be disabled.
1157 */
1158 static bool g4x_check_srwm(struct drm_device *dev,
1159 int display_wm, int cursor_wm,
1160 const struct intel_watermark_params *display,
1161 const struct intel_watermark_params *cursor)
1162 {
1163 DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
1164 display_wm, cursor_wm);
1165
1166 if (display_wm > display->max_wm) {
1167 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
1168 display_wm, display->max_wm);
1169 return false;
1170 }
1171
1172 if (cursor_wm > cursor->max_wm) {
1173 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
1174 cursor_wm, cursor->max_wm);
1175 return false;
1176 }
1177
1178 if (!(display_wm || cursor_wm)) {
1179 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
1180 return false;
1181 }
1182
1183 return true;
1184 }
1185
1186 static bool g4x_compute_srwm(struct drm_device *dev,
1187 int plane,
1188 int latency_ns,
1189 const struct intel_watermark_params *display,
1190 const struct intel_watermark_params *cursor,
1191 int *display_wm, int *cursor_wm)
1192 {
1193 struct drm_crtc *crtc;
1194 const struct drm_display_mode *adjusted_mode;
1195 int hdisplay, htotal, pixel_size, clock;
1196 unsigned long line_time_us;
1197 int line_count, line_size;
1198 int small, large;
1199 int entries;
1200
1201 if (!latency_ns) {
1202 *display_wm = *cursor_wm = 0;
1203 return false;
1204 }
1205
1206 crtc = intel_get_crtc_for_plane(dev, plane);
1207 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1208 clock = adjusted_mode->crtc_clock;
1209 htotal = adjusted_mode->crtc_htotal;
1210 hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1211 pixel_size = crtc->primary->fb->bits_per_pixel / 8;
1212
1213 line_time_us = max(htotal * 1000 / clock, 1);
1214 line_count = (latency_ns / line_time_us + 1000) / 1000;
1215 line_size = hdisplay * pixel_size;
1216
1217 /* Use the minimum of the small and large buffer method for primary */
1218 small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1219 large = line_count * line_size;
1220
1221 entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1222 *display_wm = entries + display->guard_size;
1223
1224 /* calculate the self-refresh watermark for display cursor */
1225 entries = line_count * pixel_size * to_intel_crtc(crtc)->cursor_width;
1226 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1227 *cursor_wm = entries + cursor->guard_size;
1228
1229 return g4x_check_srwm(dev,
1230 *display_wm, *cursor_wm,
1231 display, cursor);
1232 }
1233
1234 static bool vlv_compute_drain_latency(struct drm_device *dev,
1235 int plane,
1236 int *plane_prec_mult,
1237 int *plane_dl,
1238 int *cursor_prec_mult,
1239 int *cursor_dl)
1240 {
1241 struct drm_crtc *crtc;
1242 int clock, pixel_size;
1243 int entries;
1244
1245 crtc = intel_get_crtc_for_plane(dev, plane);
1246 if (!intel_crtc_active(crtc))
1247 return false;
1248
1249 clock = to_intel_crtc(crtc)->config.adjusted_mode.crtc_clock;
1250 pixel_size = crtc->primary->fb->bits_per_pixel / 8; /* BPP */
1251
1252 entries = (clock / 1000) * pixel_size;
1253 *plane_prec_mult = (entries > 256) ?
1254 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1255 *plane_dl = (64 * (*plane_prec_mult) * 4) / ((clock / 1000) *
1256 pixel_size);
1257
1258 entries = (clock / 1000) * 4; /* BPP is always 4 for cursor */
1259 *cursor_prec_mult = (entries > 256) ?
1260 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1261 *cursor_dl = (64 * (*cursor_prec_mult) * 4) / ((clock / 1000) * 4);
1262
1263 return true;
1264 }
1265
1266 /*
1267 * Update drain latency registers of memory arbiter
1268 *
1269 * Valleyview SoC has a new memory arbiter and needs drain latency registers
1270 * to be programmed. Each plane has a drain latency multiplier and a drain
1271 * latency value.
1272 */
1273
1274 static void vlv_update_drain_latency(struct drm_device *dev)
1275 {
1276 struct drm_i915_private *dev_priv = dev->dev_private;
1277 int planea_prec, planea_dl, planeb_prec, planeb_dl;
1278 int cursora_prec, cursora_dl, cursorb_prec, cursorb_dl;
1279 int plane_prec_mult, cursor_prec_mult; /* Precision multiplier is
1280 either 16 or 32 */
1281
1282 /* For plane A, Cursor A */
1283 if (vlv_compute_drain_latency(dev, 0, &plane_prec_mult, &planea_dl,
1284 &cursor_prec_mult, &cursora_dl)) {
1285 cursora_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1286 DDL_CURSORA_PRECISION_32 : DDL_CURSORA_PRECISION_16;
1287 planea_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1288 DDL_PLANEA_PRECISION_32 : DDL_PLANEA_PRECISION_16;
1289
1290 I915_WRITE(VLV_DDL1, cursora_prec |
1291 (cursora_dl << DDL_CURSORA_SHIFT) |
1292 planea_prec | planea_dl);
1293 }
1294
1295 /* For plane B, Cursor B */
1296 if (vlv_compute_drain_latency(dev, 1, &plane_prec_mult, &planeb_dl,
1297 &cursor_prec_mult, &cursorb_dl)) {
1298 cursorb_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1299 DDL_CURSORB_PRECISION_32 : DDL_CURSORB_PRECISION_16;
1300 planeb_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1301 DDL_PLANEB_PRECISION_32 : DDL_PLANEB_PRECISION_16;
1302
1303 I915_WRITE(VLV_DDL2, cursorb_prec |
1304 (cursorb_dl << DDL_CURSORB_SHIFT) |
1305 planeb_prec | planeb_dl);
1306 }
1307 }
1308
1309 #define single_plane_enabled(mask) is_power_of_2(mask)
1310
1311 static void valleyview_update_wm(struct drm_crtc *crtc)
1312 {
1313 struct drm_device *dev = crtc->dev;
1314 static const int sr_latency_ns = 12000;
1315 struct drm_i915_private *dev_priv = dev->dev_private;
1316 int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1317 int plane_sr, cursor_sr;
1318 int ignore_plane_sr, ignore_cursor_sr;
1319 unsigned int enabled = 0;
1320
1321 vlv_update_drain_latency(dev);
1322
1323 if (g4x_compute_wm0(dev, PIPE_A,
1324 &valleyview_wm_info, latency_ns,
1325 &valleyview_cursor_wm_info, latency_ns,
1326 &planea_wm, &cursora_wm))
1327 enabled |= 1 << PIPE_A;
1328
1329 if (g4x_compute_wm0(dev, PIPE_B,
1330 &valleyview_wm_info, latency_ns,
1331 &valleyview_cursor_wm_info, latency_ns,
1332 &planeb_wm, &cursorb_wm))
1333 enabled |= 1 << PIPE_B;
1334
1335 if (single_plane_enabled(enabled) &&
1336 g4x_compute_srwm(dev, ffs(enabled) - 1,
1337 sr_latency_ns,
1338 &valleyview_wm_info,
1339 &valleyview_cursor_wm_info,
1340 &plane_sr, &ignore_cursor_sr) &&
1341 g4x_compute_srwm(dev, ffs(enabled) - 1,
1342 2*sr_latency_ns,
1343 &valleyview_wm_info,
1344 &valleyview_cursor_wm_info,
1345 &ignore_plane_sr, &cursor_sr)) {
1346 I915_WRITE(FW_BLC_SELF_VLV, FW_CSPWRDWNEN);
1347 } else {
1348 I915_WRITE(FW_BLC_SELF_VLV,
1349 I915_READ(FW_BLC_SELF_VLV) & ~FW_CSPWRDWNEN);
1350 plane_sr = cursor_sr = 0;
1351 }
1352
1353 DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1354 planea_wm, cursora_wm,
1355 planeb_wm, cursorb_wm,
1356 plane_sr, cursor_sr);
1357
1358 I915_WRITE(DSPFW1,
1359 (plane_sr << DSPFW_SR_SHIFT) |
1360 (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1361 (planeb_wm << DSPFW_PLANEB_SHIFT) |
1362 planea_wm);
1363 I915_WRITE(DSPFW2,
1364 (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1365 (cursora_wm << DSPFW_CURSORA_SHIFT));
1366 I915_WRITE(DSPFW3,
1367 (I915_READ(DSPFW3) & ~DSPFW_CURSOR_SR_MASK) |
1368 (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1369 }
1370
1371 static void g4x_update_wm(struct drm_crtc *crtc)
1372 {
1373 struct drm_device *dev = crtc->dev;
1374 static const int sr_latency_ns = 12000;
1375 struct drm_i915_private *dev_priv = dev->dev_private;
1376 int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1377 int plane_sr, cursor_sr;
1378 unsigned int enabled = 0;
1379
1380 if (g4x_compute_wm0(dev, PIPE_A,
1381 &g4x_wm_info, latency_ns,
1382 &g4x_cursor_wm_info, latency_ns,
1383 &planea_wm, &cursora_wm))
1384 enabled |= 1 << PIPE_A;
1385
1386 if (g4x_compute_wm0(dev, PIPE_B,
1387 &g4x_wm_info, latency_ns,
1388 &g4x_cursor_wm_info, latency_ns,
1389 &planeb_wm, &cursorb_wm))
1390 enabled |= 1 << PIPE_B;
1391
1392 if (single_plane_enabled(enabled) &&
1393 g4x_compute_srwm(dev, ffs(enabled) - 1,
1394 sr_latency_ns,
1395 &g4x_wm_info,
1396 &g4x_cursor_wm_info,
1397 &plane_sr, &cursor_sr)) {
1398 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1399 } else {
1400 I915_WRITE(FW_BLC_SELF,
1401 I915_READ(FW_BLC_SELF) & ~FW_BLC_SELF_EN);
1402 plane_sr = cursor_sr = 0;
1403 }
1404
1405 DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1406 planea_wm, cursora_wm,
1407 planeb_wm, cursorb_wm,
1408 plane_sr, cursor_sr);
1409
1410 I915_WRITE(DSPFW1,
1411 (plane_sr << DSPFW_SR_SHIFT) |
1412 (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1413 (planeb_wm << DSPFW_PLANEB_SHIFT) |
1414 planea_wm);
1415 I915_WRITE(DSPFW2,
1416 (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1417 (cursora_wm << DSPFW_CURSORA_SHIFT));
1418 /* HPLL off in SR has some issues on G4x... disable it */
1419 I915_WRITE(DSPFW3,
1420 (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1421 (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1422 }
1423
1424 static void i965_update_wm(struct drm_crtc *unused_crtc)
1425 {
1426 struct drm_device *dev = unused_crtc->dev;
1427 struct drm_i915_private *dev_priv = dev->dev_private;
1428 struct drm_crtc *crtc;
1429 int srwm = 1;
1430 int cursor_sr = 16;
1431
1432 /* Calc sr entries for one plane configs */
1433 crtc = single_enabled_crtc(dev);
1434 if (crtc) {
1435 /* self-refresh has much higher latency */
1436 static const int sr_latency_ns = 12000;
1437 const struct drm_display_mode *adjusted_mode =
1438 &to_intel_crtc(crtc)->config.adjusted_mode;
1439 int clock = adjusted_mode->crtc_clock;
1440 int htotal = adjusted_mode->crtc_htotal;
1441 int hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1442 int pixel_size = crtc->primary->fb->bits_per_pixel / 8;
1443 unsigned long line_time_us;
1444 int entries;
1445
1446 line_time_us = max(htotal * 1000 / clock, 1);
1447
1448 /* Use ns/us then divide to preserve precision */
1449 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1450 pixel_size * hdisplay;
1451 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1452 srwm = I965_FIFO_SIZE - entries;
1453 if (srwm < 0)
1454 srwm = 1;
1455 srwm &= 0x1ff;
1456 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1457 entries, srwm);
1458
1459 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1460 pixel_size * to_intel_crtc(crtc)->cursor_width;
1461 entries = DIV_ROUND_UP(entries,
1462 i965_cursor_wm_info.cacheline_size);
1463 cursor_sr = i965_cursor_wm_info.fifo_size -
1464 (entries + i965_cursor_wm_info.guard_size);
1465
1466 if (cursor_sr > i965_cursor_wm_info.max_wm)
1467 cursor_sr = i965_cursor_wm_info.max_wm;
1468
1469 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1470 "cursor %d\n", srwm, cursor_sr);
1471
1472 if (IS_CRESTLINE(dev))
1473 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1474 } else {
1475 /* Turn off self refresh if both pipes are enabled */
1476 if (IS_CRESTLINE(dev))
1477 I915_WRITE(FW_BLC_SELF, I915_READ(FW_BLC_SELF)
1478 & ~FW_BLC_SELF_EN);
1479 }
1480
1481 DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1482 srwm);
1483
1484 /* 965 has limitations... */
1485 I915_WRITE(DSPFW1, (srwm << DSPFW_SR_SHIFT) |
1486 (8 << 16) | (8 << 8) | (8 << 0));
1487 I915_WRITE(DSPFW2, (8 << 8) | (8 << 0));
1488 /* update cursor SR watermark */
1489 I915_WRITE(DSPFW3, (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1490 }
1491
1492 static void i9xx_update_wm(struct drm_crtc *unused_crtc)
1493 {
1494 struct drm_device *dev = unused_crtc->dev;
1495 struct drm_i915_private *dev_priv = dev->dev_private;
1496 const struct intel_watermark_params *wm_info;
1497 uint32_t fwater_lo;
1498 uint32_t fwater_hi;
1499 int cwm, srwm = 1;
1500 int fifo_size;
1501 int planea_wm, planeb_wm;
1502 struct drm_crtc *crtc, *enabled = NULL;
1503
1504 if (IS_I945GM(dev))
1505 wm_info = &i945_wm_info;
1506 else if (!IS_GEN2(dev))
1507 wm_info = &i915_wm_info;
1508 else
1509 wm_info = &i830_wm_info;
1510
1511 fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1512 crtc = intel_get_crtc_for_plane(dev, 0);
1513 if (intel_crtc_active(crtc)) {
1514 const struct drm_display_mode *adjusted_mode;
1515 int cpp = crtc->primary->fb->bits_per_pixel / 8;
1516 if (IS_GEN2(dev))
1517 cpp = 4;
1518
1519 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1520 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1521 wm_info, fifo_size, cpp,
1522 latency_ns);
1523 enabled = crtc;
1524 } else
1525 planea_wm = fifo_size - wm_info->guard_size;
1526
1527 fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1528 crtc = intel_get_crtc_for_plane(dev, 1);
1529 if (intel_crtc_active(crtc)) {
1530 const struct drm_display_mode *adjusted_mode;
1531 int cpp = crtc->primary->fb->bits_per_pixel / 8;
1532 if (IS_GEN2(dev))
1533 cpp = 4;
1534
1535 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1536 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1537 wm_info, fifo_size, cpp,
1538 latency_ns);
1539 if (enabled == NULL)
1540 enabled = crtc;
1541 else
1542 enabled = NULL;
1543 } else
1544 planeb_wm = fifo_size - wm_info->guard_size;
1545
1546 DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1547
1548 if (IS_I915GM(dev) && enabled) {
1549 struct intel_framebuffer *fb;
1550
1551 fb = to_intel_framebuffer(enabled->primary->fb);
1552
1553 /* self-refresh seems busted with untiled */
1554 if (fb->obj->tiling_mode == I915_TILING_NONE)
1555 enabled = NULL;
1556 }
1557
1558 /*
1559 * Overlay gets an aggressive default since video jitter is bad.
1560 */
1561 cwm = 2;
1562
1563 /* Play safe and disable self-refresh before adjusting watermarks. */
1564 if (IS_I945G(dev) || IS_I945GM(dev))
1565 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN_MASK | 0);
1566 else if (IS_I915GM(dev))
1567 I915_WRITE(INSTPM, _MASKED_BIT_DISABLE(INSTPM_SELF_EN));
1568
1569 /* Calc sr entries for one plane configs */
1570 if (HAS_FW_BLC(dev) && enabled) {
1571 /* self-refresh has much higher latency */
1572 static const int sr_latency_ns = 6000;
1573 const struct drm_display_mode *adjusted_mode =
1574 &to_intel_crtc(enabled)->config.adjusted_mode;
1575 int clock = adjusted_mode->crtc_clock;
1576 int htotal = adjusted_mode->crtc_htotal;
1577 int hdisplay = to_intel_crtc(enabled)->config.pipe_src_w;
1578 int pixel_size = enabled->primary->fb->bits_per_pixel / 8;
1579 unsigned long line_time_us;
1580 int entries;
1581
1582 line_time_us = max(htotal * 1000 / clock, 1);
1583
1584 /* Use ns/us then divide to preserve precision */
1585 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1586 pixel_size * hdisplay;
1587 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1588 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1589 srwm = wm_info->fifo_size - entries;
1590 if (srwm < 0)
1591 srwm = 1;
1592
1593 if (IS_I945G(dev) || IS_I945GM(dev))
1594 I915_WRITE(FW_BLC_SELF,
1595 FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1596 else if (IS_I915GM(dev))
1597 I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1598 }
1599
1600 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1601 planea_wm, planeb_wm, cwm, srwm);
1602
1603 fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1604 fwater_hi = (cwm & 0x1f);
1605
1606 /* Set request length to 8 cachelines per fetch */
1607 fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1608 fwater_hi = fwater_hi | (1 << 8);
1609
1610 I915_WRITE(FW_BLC, fwater_lo);
1611 I915_WRITE(FW_BLC2, fwater_hi);
1612
1613 if (HAS_FW_BLC(dev)) {
1614 if (enabled) {
1615 if (IS_I945G(dev) || IS_I945GM(dev))
1616 I915_WRITE(FW_BLC_SELF,
1617 FW_BLC_SELF_EN_MASK | FW_BLC_SELF_EN);
1618 else if (IS_I915GM(dev))
1619 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_SELF_EN));
1620 DRM_DEBUG_KMS("memory self refresh enabled\n");
1621 } else
1622 DRM_DEBUG_KMS("memory self refresh disabled\n");
1623 }
1624 }
1625
1626 static void i845_update_wm(struct drm_crtc *unused_crtc)
1627 {
1628 struct drm_device *dev = unused_crtc->dev;
1629 struct drm_i915_private *dev_priv = dev->dev_private;
1630 struct drm_crtc *crtc;
1631 const struct drm_display_mode *adjusted_mode;
1632 uint32_t fwater_lo;
1633 int planea_wm;
1634
1635 crtc = single_enabled_crtc(dev);
1636 if (crtc == NULL)
1637 return;
1638
1639 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1640 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1641 &i845_wm_info,
1642 dev_priv->display.get_fifo_size(dev, 0),
1643 4, latency_ns);
1644 fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1645 fwater_lo |= (3<<8) | planea_wm;
1646
1647 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1648
1649 I915_WRITE(FW_BLC, fwater_lo);
1650 }
1651
1652 static uint32_t ilk_pipe_pixel_rate(struct drm_device *dev,
1653 struct drm_crtc *crtc)
1654 {
1655 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1656 uint32_t pixel_rate;
1657
1658 pixel_rate = intel_crtc->config.adjusted_mode.crtc_clock;
1659
1660 /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
1661 * adjust the pixel_rate here. */
1662
1663 if (intel_crtc->config.pch_pfit.enabled) {
1664 uint64_t pipe_w, pipe_h, pfit_w, pfit_h;
1665 uint32_t pfit_size = intel_crtc->config.pch_pfit.size;
1666
1667 pipe_w = intel_crtc->config.pipe_src_w;
1668 pipe_h = intel_crtc->config.pipe_src_h;
1669 pfit_w = (pfit_size >> 16) & 0xFFFF;
1670 pfit_h = pfit_size & 0xFFFF;
1671 if (pipe_w < pfit_w)
1672 pipe_w = pfit_w;
1673 if (pipe_h < pfit_h)
1674 pipe_h = pfit_h;
1675
1676 pixel_rate = div_u64((uint64_t) pixel_rate * pipe_w * pipe_h,
1677 pfit_w * pfit_h);
1678 }
1679
1680 return pixel_rate;
1681 }
1682
1683 /* latency must be in 0.1us units. */
1684 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
1685 uint32_t latency)
1686 {
1687 uint64_t ret;
1688
1689 if (WARN(latency == 0, "Latency value missing\n"))
1690 return UINT_MAX;
1691
1692 ret = (uint64_t) pixel_rate * bytes_per_pixel * latency;
1693 ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
1694
1695 return ret;
1696 }
1697
1698 /* latency must be in 0.1us units. */
1699 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
1700 uint32_t horiz_pixels, uint8_t bytes_per_pixel,
1701 uint32_t latency)
1702 {
1703 uint32_t ret;
1704
1705 if (WARN(latency == 0, "Latency value missing\n"))
1706 return UINT_MAX;
1707
1708 ret = (latency * pixel_rate) / (pipe_htotal * 10000);
1709 ret = (ret + 1) * horiz_pixels * bytes_per_pixel;
1710 ret = DIV_ROUND_UP(ret, 64) + 2;
1711 return ret;
1712 }
1713
1714 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
1715 uint8_t bytes_per_pixel)
1716 {
1717 return DIV_ROUND_UP(pri_val * 64, horiz_pixels * bytes_per_pixel) + 2;
1718 }
1719
1720 struct ilk_pipe_wm_parameters {
1721 bool active;
1722 uint32_t pipe_htotal;
1723 uint32_t pixel_rate;
1724 struct intel_plane_wm_parameters pri;
1725 struct intel_plane_wm_parameters spr;
1726 struct intel_plane_wm_parameters cur;
1727 };
1728
1729 struct ilk_wm_maximums {
1730 uint16_t pri;
1731 uint16_t spr;
1732 uint16_t cur;
1733 uint16_t fbc;
1734 };
1735
1736 /* used in computing the new watermarks state */
1737 struct intel_wm_config {
1738 unsigned int num_pipes_active;
1739 bool sprites_enabled;
1740 bool sprites_scaled;
1741 };
1742
1743 /*
1744 * For both WM_PIPE and WM_LP.
1745 * mem_value must be in 0.1us units.
1746 */
1747 static uint32_t ilk_compute_pri_wm(const struct ilk_pipe_wm_parameters *params,
1748 uint32_t mem_value,
1749 bool is_lp)
1750 {
1751 uint32_t method1, method2;
1752
1753 if (!params->active || !params->pri.enabled)
1754 return 0;
1755
1756 method1 = ilk_wm_method1(params->pixel_rate,
1757 params->pri.bytes_per_pixel,
1758 mem_value);
1759
1760 if (!is_lp)
1761 return method1;
1762
1763 method2 = ilk_wm_method2(params->pixel_rate,
1764 params->pipe_htotal,
1765 params->pri.horiz_pixels,
1766 params->pri.bytes_per_pixel,
1767 mem_value);
1768
1769 return min(method1, method2);
1770 }
1771
1772 /*
1773 * For both WM_PIPE and WM_LP.
1774 * mem_value must be in 0.1us units.
1775 */
1776 static uint32_t ilk_compute_spr_wm(const struct ilk_pipe_wm_parameters *params,
1777 uint32_t mem_value)
1778 {
1779 uint32_t method1, method2;
1780
1781 if (!params->active || !params->spr.enabled)
1782 return 0;
1783
1784 method1 = ilk_wm_method1(params->pixel_rate,
1785 params->spr.bytes_per_pixel,
1786 mem_value);
1787 method2 = ilk_wm_method2(params->pixel_rate,
1788 params->pipe_htotal,
1789 params->spr.horiz_pixels,
1790 params->spr.bytes_per_pixel,
1791 mem_value);
1792 return min(method1, method2);
1793 }
1794
1795 /*
1796 * For both WM_PIPE and WM_LP.
1797 * mem_value must be in 0.1us units.
1798 */
1799 static uint32_t ilk_compute_cur_wm(const struct ilk_pipe_wm_parameters *params,
1800 uint32_t mem_value)
1801 {
1802 if (!params->active || !params->cur.enabled)
1803 return 0;
1804
1805 return ilk_wm_method2(params->pixel_rate,
1806 params->pipe_htotal,
1807 params->cur.horiz_pixels,
1808 params->cur.bytes_per_pixel,
1809 mem_value);
1810 }
1811
1812 /* Only for WM_LP. */
1813 static uint32_t ilk_compute_fbc_wm(const struct ilk_pipe_wm_parameters *params,
1814 uint32_t pri_val)
1815 {
1816 if (!params->active || !params->pri.enabled)
1817 return 0;
1818
1819 return ilk_wm_fbc(pri_val,
1820 params->pri.horiz_pixels,
1821 params->pri.bytes_per_pixel);
1822 }
1823
1824 static unsigned int ilk_display_fifo_size(const struct drm_device *dev)
1825 {
1826 if (INTEL_INFO(dev)->gen >= 8)
1827 return 3072;
1828 else if (INTEL_INFO(dev)->gen >= 7)
1829 return 768;
1830 else
1831 return 512;
1832 }
1833
1834 static unsigned int ilk_plane_wm_reg_max(const struct drm_device *dev,
1835 int level, bool is_sprite)
1836 {
1837 if (INTEL_INFO(dev)->gen >= 8)
1838 /* BDW primary/sprite plane watermarks */
1839 return level == 0 ? 255 : 2047;
1840 else if (INTEL_INFO(dev)->gen >= 7)
1841 /* IVB/HSW primary/sprite plane watermarks */
1842 return level == 0 ? 127 : 1023;
1843 else if (!is_sprite)
1844 /* ILK/SNB primary plane watermarks */
1845 return level == 0 ? 127 : 511;
1846 else
1847 /* ILK/SNB sprite plane watermarks */
1848 return level == 0 ? 63 : 255;
1849 }
1850
1851 static unsigned int ilk_cursor_wm_reg_max(const struct drm_device *dev,
1852 int level)
1853 {
1854 if (INTEL_INFO(dev)->gen >= 7)
1855 return level == 0 ? 63 : 255;
1856 else
1857 return level == 0 ? 31 : 63;
1858 }
1859
1860 static unsigned int ilk_fbc_wm_reg_max(const struct drm_device *dev)
1861 {
1862 if (INTEL_INFO(dev)->gen >= 8)
1863 return 31;
1864 else
1865 return 15;
1866 }
1867
1868 /* Calculate the maximum primary/sprite plane watermark */
1869 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
1870 int level,
1871 const struct intel_wm_config *config,
1872 enum intel_ddb_partitioning ddb_partitioning,
1873 bool is_sprite)
1874 {
1875 unsigned int fifo_size = ilk_display_fifo_size(dev);
1876
1877 /* if sprites aren't enabled, sprites get nothing */
1878 if (is_sprite && !config->sprites_enabled)
1879 return 0;
1880
1881 /* HSW allows LP1+ watermarks even with multiple pipes */
1882 if (level == 0 || config->num_pipes_active > 1) {
1883 fifo_size /= INTEL_INFO(dev)->num_pipes;
1884
1885 /*
1886 * For some reason the non self refresh
1887 * FIFO size is only half of the self
1888 * refresh FIFO size on ILK/SNB.
1889 */
1890 if (INTEL_INFO(dev)->gen <= 6)
1891 fifo_size /= 2;
1892 }
1893
1894 if (config->sprites_enabled) {
1895 /* level 0 is always calculated with 1:1 split */
1896 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
1897 if (is_sprite)
1898 fifo_size *= 5;
1899 fifo_size /= 6;
1900 } else {
1901 fifo_size /= 2;
1902 }
1903 }
1904
1905 /* clamp to max that the registers can hold */
1906 return min(fifo_size, ilk_plane_wm_reg_max(dev, level, is_sprite));
1907 }
1908
1909 /* Calculate the maximum cursor plane watermark */
1910 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
1911 int level,
1912 const struct intel_wm_config *config)
1913 {
1914 /* HSW LP1+ watermarks w/ multiple pipes */
1915 if (level > 0 && config->num_pipes_active > 1)
1916 return 64;
1917
1918 /* otherwise just report max that registers can hold */
1919 return ilk_cursor_wm_reg_max(dev, level);
1920 }
1921
1922 static void ilk_compute_wm_maximums(const struct drm_device *dev,
1923 int level,
1924 const struct intel_wm_config *config,
1925 enum intel_ddb_partitioning ddb_partitioning,
1926 struct ilk_wm_maximums *max)
1927 {
1928 max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
1929 max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
1930 max->cur = ilk_cursor_wm_max(dev, level, config);
1931 max->fbc = ilk_fbc_wm_reg_max(dev);
1932 }
1933
1934 static void ilk_compute_wm_reg_maximums(struct drm_device *dev,
1935 int level,
1936 struct ilk_wm_maximums *max)
1937 {
1938 max->pri = ilk_plane_wm_reg_max(dev, level, false);
1939 max->spr = ilk_plane_wm_reg_max(dev, level, true);
1940 max->cur = ilk_cursor_wm_reg_max(dev, level);
1941 max->fbc = ilk_fbc_wm_reg_max(dev);
1942 }
1943
1944 static bool ilk_validate_wm_level(int level,
1945 const struct ilk_wm_maximums *max,
1946 struct intel_wm_level *result)
1947 {
1948 bool ret;
1949
1950 /* already determined to be invalid? */
1951 if (!result->enable)
1952 return false;
1953
1954 result->enable = result->pri_val <= max->pri &&
1955 result->spr_val <= max->spr &&
1956 result->cur_val <= max->cur;
1957
1958 ret = result->enable;
1959
1960 /*
1961 * HACK until we can pre-compute everything,
1962 * and thus fail gracefully if LP0 watermarks
1963 * are exceeded...
1964 */
1965 if (level == 0 && !result->enable) {
1966 if (result->pri_val > max->pri)
1967 DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
1968 level, result->pri_val, max->pri);
1969 if (result->spr_val > max->spr)
1970 DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
1971 level, result->spr_val, max->spr);
1972 if (result->cur_val > max->cur)
1973 DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
1974 level, result->cur_val, max->cur);
1975
1976 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
1977 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
1978 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
1979 result->enable = true;
1980 }
1981
1982 return ret;
1983 }
1984
1985 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
1986 int level,
1987 const struct ilk_pipe_wm_parameters *p,
1988 struct intel_wm_level *result)
1989 {
1990 uint16_t pri_latency = dev_priv->wm.pri_latency[level];
1991 uint16_t spr_latency = dev_priv->wm.spr_latency[level];
1992 uint16_t cur_latency = dev_priv->wm.cur_latency[level];
1993
1994 /* WM1+ latency values stored in 0.5us units */
1995 if (level > 0) {
1996 pri_latency *= 5;
1997 spr_latency *= 5;
1998 cur_latency *= 5;
1999 }
2000
2001 result->pri_val = ilk_compute_pri_wm(p, pri_latency, level);
2002 result->spr_val = ilk_compute_spr_wm(p, spr_latency);
2003 result->cur_val = ilk_compute_cur_wm(p, cur_latency);
2004 result->fbc_val = ilk_compute_fbc_wm(p, result->pri_val);
2005 result->enable = true;
2006 }
2007
2008 static uint32_t
2009 hsw_compute_linetime_wm(struct drm_device *dev, struct drm_crtc *crtc)
2010 {
2011 struct drm_i915_private *dev_priv = dev->dev_private;
2012 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2013 struct drm_display_mode *mode = &intel_crtc->config.adjusted_mode;
2014 u32 linetime, ips_linetime;
2015
2016 if (!intel_crtc_active(crtc))
2017 return 0;
2018
2019 /* The WM are computed with base on how long it takes to fill a single
2020 * row at the given clock rate, multiplied by 8.
2021 * */
2022 linetime = DIV_ROUND_CLOSEST(mode->crtc_htotal * 1000 * 8,
2023 mode->crtc_clock);
2024 ips_linetime = DIV_ROUND_CLOSEST(mode->crtc_htotal * 1000 * 8,
2025 intel_ddi_get_cdclk_freq(dev_priv));
2026
2027 return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2028 PIPE_WM_LINETIME_TIME(linetime);
2029 }
2030
2031 static void intel_read_wm_latency(struct drm_device *dev, uint16_t wm[5])
2032 {
2033 struct drm_i915_private *dev_priv = dev->dev_private;
2034
2035 if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2036 uint64_t sskpd = I915_READ64(MCH_SSKPD);
2037
2038 wm[0] = (sskpd >> 56) & 0xFF;
2039 if (wm[0] == 0)
2040 wm[0] = sskpd & 0xF;
2041 wm[1] = (sskpd >> 4) & 0xFF;
2042 wm[2] = (sskpd >> 12) & 0xFF;
2043 wm[3] = (sskpd >> 20) & 0x1FF;
2044 wm[4] = (sskpd >> 32) & 0x1FF;
2045 } else if (INTEL_INFO(dev)->gen >= 6) {
2046 uint32_t sskpd = I915_READ(MCH_SSKPD);
2047
2048 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2049 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2050 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2051 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2052 } else if (INTEL_INFO(dev)->gen >= 5) {
2053 uint32_t mltr = I915_READ(MLTR_ILK);
2054
2055 /* ILK primary LP0 latency is 700 ns */
2056 wm[0] = 7;
2057 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2058 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2059 }
2060 }
2061
2062 static void intel_fixup_spr_wm_latency(struct drm_device *dev, uint16_t wm[5])
2063 {
2064 /* ILK sprite LP0 latency is 1300 ns */
2065 if (INTEL_INFO(dev)->gen == 5)
2066 wm[0] = 13;
2067 }
2068
2069 static void intel_fixup_cur_wm_latency(struct drm_device *dev, uint16_t wm[5])
2070 {
2071 /* ILK cursor LP0 latency is 1300 ns */
2072 if (INTEL_INFO(dev)->gen == 5)
2073 wm[0] = 13;
2074
2075 /* WaDoubleCursorLP3Latency:ivb */
2076 if (IS_IVYBRIDGE(dev))
2077 wm[3] *= 2;
2078 }
2079
2080 int ilk_wm_max_level(const struct drm_device *dev)
2081 {
2082 /* how many WM levels are we expecting */
2083 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2084 return 4;
2085 else if (INTEL_INFO(dev)->gen >= 6)
2086 return 3;
2087 else
2088 return 2;
2089 }
2090
2091 static void intel_print_wm_latency(struct drm_device *dev,
2092 const char *name,
2093 const uint16_t wm[5])
2094 {
2095 int level, max_level = ilk_wm_max_level(dev);
2096
2097 for (level = 0; level <= max_level; level++) {
2098 unsigned int latency = wm[level];
2099
2100 if (latency == 0) {
2101 DRM_ERROR("%s WM%d latency not provided\n",
2102 name, level);
2103 continue;
2104 }
2105
2106 /* WM1+ latency values in 0.5us units */
2107 if (level > 0)
2108 latency *= 5;
2109
2110 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2111 name, level, wm[level],
2112 latency / 10, latency % 10);
2113 }
2114 }
2115
2116 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
2117 uint16_t wm[5], uint16_t min)
2118 {
2119 int level, max_level = ilk_wm_max_level(dev_priv->dev);
2120
2121 if (wm[0] >= min)
2122 return false;
2123
2124 wm[0] = max(wm[0], min);
2125 for (level = 1; level <= max_level; level++)
2126 wm[level] = max_t(uint16_t, wm[level], DIV_ROUND_UP(min, 5));
2127
2128 return true;
2129 }
2130
2131 static void snb_wm_latency_quirk(struct drm_device *dev)
2132 {
2133 struct drm_i915_private *dev_priv = dev->dev_private;
2134 bool changed;
2135
2136 /*
2137 * The BIOS provided WM memory latency values are often
2138 * inadequate for high resolution displays. Adjust them.
2139 */
2140 changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
2141 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
2142 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
2143
2144 if (!changed)
2145 return;
2146
2147 DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2148 intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2149 intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2150 intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2151 }
2152
2153 static void ilk_setup_wm_latency(struct drm_device *dev)
2154 {
2155 struct drm_i915_private *dev_priv = dev->dev_private;
2156
2157 intel_read_wm_latency(dev, dev_priv->wm.pri_latency);
2158
2159 memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2160 sizeof(dev_priv->wm.pri_latency));
2161 memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2162 sizeof(dev_priv->wm.pri_latency));
2163
2164 intel_fixup_spr_wm_latency(dev, dev_priv->wm.spr_latency);
2165 intel_fixup_cur_wm_latency(dev, dev_priv->wm.cur_latency);
2166
2167 intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2168 intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2169 intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2170
2171 if (IS_GEN6(dev))
2172 snb_wm_latency_quirk(dev);
2173 }
2174
2175 static void ilk_compute_wm_parameters(struct drm_crtc *crtc,
2176 struct ilk_pipe_wm_parameters *p)
2177 {
2178 struct drm_device *dev = crtc->dev;
2179 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2180 enum pipe pipe = intel_crtc->pipe;
2181 struct drm_plane *plane;
2182
2183 if (!intel_crtc_active(crtc))
2184 return;
2185
2186 p->active = true;
2187 p->pipe_htotal = intel_crtc->config.adjusted_mode.crtc_htotal;
2188 p->pixel_rate = ilk_pipe_pixel_rate(dev, crtc);
2189 p->pri.bytes_per_pixel = crtc->primary->fb->bits_per_pixel / 8;
2190 p->cur.bytes_per_pixel = 4;
2191 p->pri.horiz_pixels = intel_crtc->config.pipe_src_w;
2192 p->cur.horiz_pixels = intel_crtc->cursor_width;
2193 /* TODO: for now, assume primary and cursor planes are always enabled. */
2194 p->pri.enabled = true;
2195 p->cur.enabled = true;
2196
2197 drm_for_each_legacy_plane(plane, &dev->mode_config.plane_list) {
2198 struct intel_plane *intel_plane = to_intel_plane(plane);
2199
2200 if (intel_plane->pipe == pipe) {
2201 p->spr = intel_plane->wm;
2202 break;
2203 }
2204 }
2205 }
2206
2207 static void ilk_compute_wm_config(struct drm_device *dev,
2208 struct intel_wm_config *config)
2209 {
2210 struct intel_crtc *intel_crtc;
2211
2212 /* Compute the currently _active_ config */
2213 for_each_intel_crtc(dev, intel_crtc) {
2214 const struct intel_pipe_wm *wm = &intel_crtc->wm.active;
2215
2216 if (!wm->pipe_enabled)
2217 continue;
2218
2219 config->sprites_enabled |= wm->sprites_enabled;
2220 config->sprites_scaled |= wm->sprites_scaled;
2221 config->num_pipes_active++;
2222 }
2223 }
2224
2225 /* Compute new watermarks for the pipe */
2226 static bool intel_compute_pipe_wm(struct drm_crtc *crtc,
2227 const struct ilk_pipe_wm_parameters *params,
2228 struct intel_pipe_wm *pipe_wm)
2229 {
2230 struct drm_device *dev = crtc->dev;
2231 const struct drm_i915_private *dev_priv = dev->dev_private;
2232 int level, max_level = ilk_wm_max_level(dev);
2233 /* LP0 watermark maximums depend on this pipe alone */
2234 struct intel_wm_config config = {
2235 .num_pipes_active = 1,
2236 .sprites_enabled = params->spr.enabled,
2237 .sprites_scaled = params->spr.scaled,
2238 };
2239 struct ilk_wm_maximums max;
2240
2241 pipe_wm->pipe_enabled = params->active;
2242 pipe_wm->sprites_enabled = params->spr.enabled;
2243 pipe_wm->sprites_scaled = params->spr.scaled;
2244
2245 /* ILK/SNB: LP2+ watermarks only w/o sprites */
2246 if (INTEL_INFO(dev)->gen <= 6 && params->spr.enabled)
2247 max_level = 1;
2248
2249 /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2250 if (params->spr.scaled)
2251 max_level = 0;
2252
2253 ilk_compute_wm_level(dev_priv, 0, params, &pipe_wm->wm[0]);
2254
2255 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2256 pipe_wm->linetime = hsw_compute_linetime_wm(dev, crtc);
2257
2258 /* LP0 watermarks always use 1/2 DDB partitioning */
2259 ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
2260
2261 /* At least LP0 must be valid */
2262 if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0]))
2263 return false;
2264
2265 ilk_compute_wm_reg_maximums(dev, 1, &max);
2266
2267 for (level = 1; level <= max_level; level++) {
2268 struct intel_wm_level wm = {};
2269
2270 ilk_compute_wm_level(dev_priv, level, params, &wm);
2271
2272 /*
2273 * Disable any watermark level that exceeds the
2274 * register maximums since such watermarks are
2275 * always invalid.
2276 */
2277 if (!ilk_validate_wm_level(level, &max, &wm))
2278 break;
2279
2280 pipe_wm->wm[level] = wm;
2281 }
2282
2283 return true;
2284 }
2285
2286 /*
2287 * Merge the watermarks from all active pipes for a specific level.
2288 */
2289 static void ilk_merge_wm_level(struct drm_device *dev,
2290 int level,
2291 struct intel_wm_level *ret_wm)
2292 {
2293 const struct intel_crtc *intel_crtc;
2294
2295 ret_wm->enable = true;
2296
2297 for_each_intel_crtc(dev, intel_crtc) {
2298 const struct intel_pipe_wm *active = &intel_crtc->wm.active;
2299 const struct intel_wm_level *wm = &active->wm[level];
2300
2301 if (!active->pipe_enabled)
2302 continue;
2303
2304 /*
2305 * The watermark values may have been used in the past,
2306 * so we must maintain them in the registers for some
2307 * time even if the level is now disabled.
2308 */
2309 if (!wm->enable)
2310 ret_wm->enable = false;
2311
2312 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
2313 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
2314 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
2315 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
2316 }
2317 }
2318
2319 /*
2320 * Merge all low power watermarks for all active pipes.
2321 */
2322 static void ilk_wm_merge(struct drm_device *dev,
2323 const struct intel_wm_config *config,
2324 const struct ilk_wm_maximums *max,
2325 struct intel_pipe_wm *merged)
2326 {
2327 int level, max_level = ilk_wm_max_level(dev);
2328 int last_enabled_level = max_level;
2329
2330 /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2331 if ((INTEL_INFO(dev)->gen <= 6 || IS_IVYBRIDGE(dev)) &&
2332 config->num_pipes_active > 1)
2333 return;
2334
2335 /* ILK: FBC WM must be disabled always */
2336 merged->fbc_wm_enabled = INTEL_INFO(dev)->gen >= 6;
2337
2338 /* merge each WM1+ level */
2339 for (level = 1; level <= max_level; level++) {
2340 struct intel_wm_level *wm = &merged->wm[level];
2341
2342 ilk_merge_wm_level(dev, level, wm);
2343
2344 if (level > last_enabled_level)
2345 wm->enable = false;
2346 else if (!ilk_validate_wm_level(level, max, wm))
2347 /* make sure all following levels get disabled */
2348 last_enabled_level = level - 1;
2349
2350 /*
2351 * The spec says it is preferred to disable
2352 * FBC WMs instead of disabling a WM level.
2353 */
2354 if (wm->fbc_val > max->fbc) {
2355 if (wm->enable)
2356 merged->fbc_wm_enabled = false;
2357 wm->fbc_val = 0;
2358 }
2359 }
2360
2361 /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2362 /*
2363 * FIXME this is racy. FBC might get enabled later.
2364 * What we should check here is whether FBC can be
2365 * enabled sometime later.
2366 */
2367 if (IS_GEN5(dev) && !merged->fbc_wm_enabled && intel_fbc_enabled(dev)) {
2368 for (level = 2; level <= max_level; level++) {
2369 struct intel_wm_level *wm = &merged->wm[level];
2370
2371 wm->enable = false;
2372 }
2373 }
2374 }
2375
2376 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
2377 {
2378 /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2379 return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
2380 }
2381
2382 /* The value we need to program into the WM_LPx latency field */
2383 static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
2384 {
2385 struct drm_i915_private *dev_priv = dev->dev_private;
2386
2387 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2388 return 2 * level;
2389 else
2390 return dev_priv->wm.pri_latency[level];
2391 }
2392
2393 static void ilk_compute_wm_results(struct drm_device *dev,
2394 const struct intel_pipe_wm *merged,
2395 enum intel_ddb_partitioning partitioning,
2396 struct ilk_wm_values *results)
2397 {
2398 struct intel_crtc *intel_crtc;
2399 int level, wm_lp;
2400
2401 results->enable_fbc_wm = merged->fbc_wm_enabled;
2402 results->partitioning = partitioning;
2403
2404 /* LP1+ register values */
2405 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2406 const struct intel_wm_level *r;
2407
2408 level = ilk_wm_lp_to_level(wm_lp, merged);
2409
2410 r = &merged->wm[level];
2411
2412 /*
2413 * Maintain the watermark values even if the level is
2414 * disabled. Doing otherwise could cause underruns.
2415 */
2416 results->wm_lp[wm_lp - 1] =
2417 (ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
2418 (r->pri_val << WM1_LP_SR_SHIFT) |
2419 r->cur_val;
2420
2421 if (r->enable)
2422 results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
2423
2424 if (INTEL_INFO(dev)->gen >= 8)
2425 results->wm_lp[wm_lp - 1] |=
2426 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
2427 else
2428 results->wm_lp[wm_lp - 1] |=
2429 r->fbc_val << WM1_LP_FBC_SHIFT;
2430
2431 /*
2432 * Always set WM1S_LP_EN when spr_val != 0, even if the
2433 * level is disabled. Doing otherwise could cause underruns.
2434 */
2435 if (INTEL_INFO(dev)->gen <= 6 && r->spr_val) {
2436 WARN_ON(wm_lp != 1);
2437 results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
2438 } else
2439 results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2440 }
2441
2442 /* LP0 register values */
2443 for_each_intel_crtc(dev, intel_crtc) {
2444 enum pipe pipe = intel_crtc->pipe;
2445 const struct intel_wm_level *r =
2446 &intel_crtc->wm.active.wm[0];
2447
2448 if (WARN_ON(!r->enable))
2449 continue;
2450
2451 results->wm_linetime[pipe] = intel_crtc->wm.active.linetime;
2452
2453 results->wm_pipe[pipe] =
2454 (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
2455 (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
2456 r->cur_val;
2457 }
2458 }
2459
2460 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2461 * case both are at the same level. Prefer r1 in case they're the same. */
2462 static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
2463 struct intel_pipe_wm *r1,
2464 struct intel_pipe_wm *r2)
2465 {
2466 int level, max_level = ilk_wm_max_level(dev);
2467 int level1 = 0, level2 = 0;
2468
2469 for (level = 1; level <= max_level; level++) {
2470 if (r1->wm[level].enable)
2471 level1 = level;
2472 if (r2->wm[level].enable)
2473 level2 = level;
2474 }
2475
2476 if (level1 == level2) {
2477 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
2478 return r2;
2479 else
2480 return r1;
2481 } else if (level1 > level2) {
2482 return r1;
2483 } else {
2484 return r2;
2485 }
2486 }
2487
2488 /* dirty bits used to track which watermarks need changes */
2489 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2490 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2491 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2492 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2493 #define WM_DIRTY_FBC (1 << 24)
2494 #define WM_DIRTY_DDB (1 << 25)
2495
2496 static unsigned int ilk_compute_wm_dirty(struct drm_device *dev,
2497 const struct ilk_wm_values *old,
2498 const struct ilk_wm_values *new)
2499 {
2500 unsigned int dirty = 0;
2501 enum pipe pipe;
2502 int wm_lp;
2503
2504 for_each_pipe(pipe) {
2505 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
2506 dirty |= WM_DIRTY_LINETIME(pipe);
2507 /* Must disable LP1+ watermarks too */
2508 dirty |= WM_DIRTY_LP_ALL;
2509 }
2510
2511 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
2512 dirty |= WM_DIRTY_PIPE(pipe);
2513 /* Must disable LP1+ watermarks too */
2514 dirty |= WM_DIRTY_LP_ALL;
2515 }
2516 }
2517
2518 if (old->enable_fbc_wm != new->enable_fbc_wm) {
2519 dirty |= WM_DIRTY_FBC;
2520 /* Must disable LP1+ watermarks too */
2521 dirty |= WM_DIRTY_LP_ALL;
2522 }
2523
2524 if (old->partitioning != new->partitioning) {
2525 dirty |= WM_DIRTY_DDB;
2526 /* Must disable LP1+ watermarks too */
2527 dirty |= WM_DIRTY_LP_ALL;
2528 }
2529
2530 /* LP1+ watermarks already deemed dirty, no need to continue */
2531 if (dirty & WM_DIRTY_LP_ALL)
2532 return dirty;
2533
2534 /* Find the lowest numbered LP1+ watermark in need of an update... */
2535 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2536 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
2537 old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
2538 break;
2539 }
2540
2541 /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2542 for (; wm_lp <= 3; wm_lp++)
2543 dirty |= WM_DIRTY_LP(wm_lp);
2544
2545 return dirty;
2546 }
2547
2548 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
2549 unsigned int dirty)
2550 {
2551 struct ilk_wm_values *previous = &dev_priv->wm.hw;
2552 bool changed = false;
2553
2554 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
2555 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
2556 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
2557 changed = true;
2558 }
2559 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
2560 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
2561 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
2562 changed = true;
2563 }
2564 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
2565 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
2566 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
2567 changed = true;
2568 }
2569
2570 /*
2571 * Don't touch WM1S_LP_EN here.
2572 * Doing so could cause underruns.
2573 */
2574
2575 return changed;
2576 }
2577
2578 /*
2579 * The spec says we shouldn't write when we don't need, because every write
2580 * causes WMs to be re-evaluated, expending some power.
2581 */
2582 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
2583 struct ilk_wm_values *results)
2584 {
2585 struct drm_device *dev = dev_priv->dev;
2586 struct ilk_wm_values *previous = &dev_priv->wm.hw;
2587 unsigned int dirty;
2588 uint32_t val;
2589
2590 dirty = ilk_compute_wm_dirty(dev, previous, results);
2591 if (!dirty)
2592 return;
2593
2594 _ilk_disable_lp_wm(dev_priv, dirty);
2595
2596 if (dirty & WM_DIRTY_PIPE(PIPE_A))
2597 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2598 if (dirty & WM_DIRTY_PIPE(PIPE_B))
2599 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
2600 if (dirty & WM_DIRTY_PIPE(PIPE_C))
2601 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
2602
2603 if (dirty & WM_DIRTY_LINETIME(PIPE_A))
2604 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
2605 if (dirty & WM_DIRTY_LINETIME(PIPE_B))
2606 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
2607 if (dirty & WM_DIRTY_LINETIME(PIPE_C))
2608 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
2609
2610 if (dirty & WM_DIRTY_DDB) {
2611 if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2612 val = I915_READ(WM_MISC);
2613 if (results->partitioning == INTEL_DDB_PART_1_2)
2614 val &= ~WM_MISC_DATA_PARTITION_5_6;
2615 else
2616 val |= WM_MISC_DATA_PARTITION_5_6;
2617 I915_WRITE(WM_MISC, val);
2618 } else {
2619 val = I915_READ(DISP_ARB_CTL2);
2620 if (results->partitioning == INTEL_DDB_PART_1_2)
2621 val &= ~DISP_DATA_PARTITION_5_6;
2622 else
2623 val |= DISP_DATA_PARTITION_5_6;
2624 I915_WRITE(DISP_ARB_CTL2, val);
2625 }
2626 }
2627
2628 if (dirty & WM_DIRTY_FBC) {
2629 val = I915_READ(DISP_ARB_CTL);
2630 if (results->enable_fbc_wm)
2631 val &= ~DISP_FBC_WM_DIS;
2632 else
2633 val |= DISP_FBC_WM_DIS;
2634 I915_WRITE(DISP_ARB_CTL, val);
2635 }
2636
2637 if (dirty & WM_DIRTY_LP(1) &&
2638 previous->wm_lp_spr[0] != results->wm_lp_spr[0])
2639 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
2640
2641 if (INTEL_INFO(dev)->gen >= 7) {
2642 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
2643 I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
2644 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
2645 I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
2646 }
2647
2648 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
2649 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
2650 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
2651 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
2652 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
2653 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
2654
2655 dev_priv->wm.hw = *results;
2656 }
2657
2658 static bool ilk_disable_lp_wm(struct drm_device *dev)
2659 {
2660 struct drm_i915_private *dev_priv = dev->dev_private;
2661
2662 return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
2663 }
2664
2665 static void ilk_update_wm(struct drm_crtc *crtc)
2666 {
2667 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2668 struct drm_device *dev = crtc->dev;
2669 struct drm_i915_private *dev_priv = dev->dev_private;
2670 struct ilk_wm_maximums max;
2671 struct ilk_pipe_wm_parameters params = {};
2672 struct ilk_wm_values results = {};
2673 enum intel_ddb_partitioning partitioning;
2674 struct intel_pipe_wm pipe_wm = {};
2675 struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
2676 struct intel_wm_config config = {};
2677
2678 ilk_compute_wm_parameters(crtc, &params);
2679
2680 intel_compute_pipe_wm(crtc, &params, &pipe_wm);
2681
2682 if (!memcmp(&intel_crtc->wm.active, &pipe_wm, sizeof(pipe_wm)))
2683 return;
2684
2685 intel_crtc->wm.active = pipe_wm;
2686
2687 ilk_compute_wm_config(dev, &config);
2688
2689 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
2690 ilk_wm_merge(dev, &config, &max, &lp_wm_1_2);
2691
2692 /* 5/6 split only in single pipe config on IVB+ */
2693 if (INTEL_INFO(dev)->gen >= 7 &&
2694 config.num_pipes_active == 1 && config.sprites_enabled) {
2695 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
2696 ilk_wm_merge(dev, &config, &max, &lp_wm_5_6);
2697
2698 best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
2699 } else {
2700 best_lp_wm = &lp_wm_1_2;
2701 }
2702
2703 partitioning = (best_lp_wm == &lp_wm_1_2) ?
2704 INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
2705
2706 ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
2707
2708 ilk_write_wm_values(dev_priv, &results);
2709 }
2710
2711 static void ilk_update_sprite_wm(struct drm_plane *plane,
2712 struct drm_crtc *crtc,
2713 uint32_t sprite_width, int pixel_size,
2714 bool enabled, bool scaled)
2715 {
2716 struct drm_device *dev = plane->dev;
2717 struct intel_plane *intel_plane = to_intel_plane(plane);
2718
2719 intel_plane->wm.enabled = enabled;
2720 intel_plane->wm.scaled = scaled;
2721 intel_plane->wm.horiz_pixels = sprite_width;
2722 intel_plane->wm.bytes_per_pixel = pixel_size;
2723
2724 /*
2725 * IVB workaround: must disable low power watermarks for at least
2726 * one frame before enabling scaling. LP watermarks can be re-enabled
2727 * when scaling is disabled.
2728 *
2729 * WaCxSRDisabledForSpriteScaling:ivb
2730 */
2731 if (IS_IVYBRIDGE(dev) && scaled && ilk_disable_lp_wm(dev))
2732 intel_wait_for_vblank(dev, intel_plane->pipe);
2733
2734 ilk_update_wm(crtc);
2735 }
2736
2737 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
2738 {
2739 struct drm_device *dev = crtc->dev;
2740 struct drm_i915_private *dev_priv = dev->dev_private;
2741 struct ilk_wm_values *hw = &dev_priv->wm.hw;
2742 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2743 struct intel_pipe_wm *active = &intel_crtc->wm.active;
2744 enum pipe pipe = intel_crtc->pipe;
2745 static const unsigned int wm0_pipe_reg[] = {
2746 [PIPE_A] = WM0_PIPEA_ILK,
2747 [PIPE_B] = WM0_PIPEB_ILK,
2748 [PIPE_C] = WM0_PIPEC_IVB,
2749 };
2750
2751 hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
2752 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2753 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
2754
2755 active->pipe_enabled = intel_crtc_active(crtc);
2756
2757 if (active->pipe_enabled) {
2758 u32 tmp = hw->wm_pipe[pipe];
2759
2760 /*
2761 * For active pipes LP0 watermark is marked as
2762 * enabled, and LP1+ watermaks as disabled since
2763 * we can't really reverse compute them in case
2764 * multiple pipes are active.
2765 */
2766 active->wm[0].enable = true;
2767 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
2768 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
2769 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
2770 active->linetime = hw->wm_linetime[pipe];
2771 } else {
2772 int level, max_level = ilk_wm_max_level(dev);
2773
2774 /*
2775 * For inactive pipes, all watermark levels
2776 * should be marked as enabled but zeroed,
2777 * which is what we'd compute them to.
2778 */
2779 for (level = 0; level <= max_level; level++)
2780 active->wm[level].enable = true;
2781 }
2782 }
2783
2784 void ilk_wm_get_hw_state(struct drm_device *dev)
2785 {
2786 struct drm_i915_private *dev_priv = dev->dev_private;
2787 struct ilk_wm_values *hw = &dev_priv->wm.hw;
2788 struct drm_crtc *crtc;
2789
2790 for_each_crtc(dev, crtc)
2791 ilk_pipe_wm_get_hw_state(crtc);
2792
2793 hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
2794 hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
2795 hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
2796
2797 hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
2798 if (INTEL_INFO(dev)->gen >= 7) {
2799 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
2800 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
2801 }
2802
2803 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2804 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
2805 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
2806 else if (IS_IVYBRIDGE(dev))
2807 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
2808 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
2809
2810 hw->enable_fbc_wm =
2811 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
2812 }
2813
2814 /**
2815 * intel_update_watermarks - update FIFO watermark values based on current modes
2816 *
2817 * Calculate watermark values for the various WM regs based on current mode
2818 * and plane configuration.
2819 *
2820 * There are several cases to deal with here:
2821 * - normal (i.e. non-self-refresh)
2822 * - self-refresh (SR) mode
2823 * - lines are large relative to FIFO size (buffer can hold up to 2)
2824 * - lines are small relative to FIFO size (buffer can hold more than 2
2825 * lines), so need to account for TLB latency
2826 *
2827 * The normal calculation is:
2828 * watermark = dotclock * bytes per pixel * latency
2829 * where latency is platform & configuration dependent (we assume pessimal
2830 * values here).
2831 *
2832 * The SR calculation is:
2833 * watermark = (trunc(latency/line time)+1) * surface width *
2834 * bytes per pixel
2835 * where
2836 * line time = htotal / dotclock
2837 * surface width = hdisplay for normal plane and 64 for cursor
2838 * and latency is assumed to be high, as above.
2839 *
2840 * The final value programmed to the register should always be rounded up,
2841 * and include an extra 2 entries to account for clock crossings.
2842 *
2843 * We don't use the sprite, so we can ignore that. And on Crestline we have
2844 * to set the non-SR watermarks to 8.
2845 */
2846 void intel_update_watermarks(struct drm_crtc *crtc)
2847 {
2848 struct drm_i915_private *dev_priv = crtc->dev->dev_private;
2849
2850 if (dev_priv->display.update_wm)
2851 dev_priv->display.update_wm(crtc);
2852 }
2853
2854 void intel_update_sprite_watermarks(struct drm_plane *plane,
2855 struct drm_crtc *crtc,
2856 uint32_t sprite_width, int pixel_size,
2857 bool enabled, bool scaled)
2858 {
2859 struct drm_i915_private *dev_priv = plane->dev->dev_private;
2860
2861 if (dev_priv->display.update_sprite_wm)
2862 dev_priv->display.update_sprite_wm(plane, crtc, sprite_width,
2863 pixel_size, enabled, scaled);
2864 }
2865
2866 static struct drm_i915_gem_object *
2867 intel_alloc_context_page(struct drm_device *dev)
2868 {
2869 struct drm_i915_gem_object *ctx;
2870 int ret;
2871
2872 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2873
2874 ctx = i915_gem_alloc_object(dev, 4096);
2875 if (!ctx) {
2876 DRM_DEBUG("failed to alloc power context, RC6 disabled\n");
2877 return NULL;
2878 }
2879
2880 ret = i915_gem_obj_ggtt_pin(ctx, 4096, 0);
2881 if (ret) {
2882 DRM_ERROR("failed to pin power context: %d\n", ret);
2883 goto err_unref;
2884 }
2885
2886 ret = i915_gem_object_set_to_gtt_domain(ctx, 1);
2887 if (ret) {
2888 DRM_ERROR("failed to set-domain on power context: %d\n", ret);
2889 goto err_unpin;
2890 }
2891
2892 return ctx;
2893
2894 err_unpin:
2895 i915_gem_object_ggtt_unpin(ctx);
2896 err_unref:
2897 drm_gem_object_unreference(&ctx->base);
2898 return NULL;
2899 }
2900
2901 /**
2902 * Lock protecting IPS related data structures
2903 */
2904 DEFINE_SPINLOCK(mchdev_lock);
2905
2906 /* Global for IPS driver to get at the current i915 device. Protected by
2907 * mchdev_lock. */
2908 static struct drm_i915_private *i915_mch_dev;
2909
2910 bool ironlake_set_drps(struct drm_device *dev, u8 val)
2911 {
2912 struct drm_i915_private *dev_priv = dev->dev_private;
2913 u16 rgvswctl;
2914
2915 assert_spin_locked(&mchdev_lock);
2916
2917 rgvswctl = I915_READ16(MEMSWCTL);
2918 if (rgvswctl & MEMCTL_CMD_STS) {
2919 DRM_DEBUG("gpu busy, RCS change rejected\n");
2920 return false; /* still busy with another command */
2921 }
2922
2923 rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
2924 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
2925 I915_WRITE16(MEMSWCTL, rgvswctl);
2926 POSTING_READ16(MEMSWCTL);
2927
2928 rgvswctl |= MEMCTL_CMD_STS;
2929 I915_WRITE16(MEMSWCTL, rgvswctl);
2930
2931 return true;
2932 }
2933
2934 static void ironlake_enable_drps(struct drm_device *dev)
2935 {
2936 struct drm_i915_private *dev_priv = dev->dev_private;
2937 u32 rgvmodectl = I915_READ(MEMMODECTL);
2938 u8 fmax, fmin, fstart, vstart;
2939
2940 spin_lock_irq(&mchdev_lock);
2941
2942 /* Enable temp reporting */
2943 I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
2944 I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
2945
2946 /* 100ms RC evaluation intervals */
2947 I915_WRITE(RCUPEI, 100000);
2948 I915_WRITE(RCDNEI, 100000);
2949
2950 /* Set max/min thresholds to 90ms and 80ms respectively */
2951 I915_WRITE(RCBMAXAVG, 90000);
2952 I915_WRITE(RCBMINAVG, 80000);
2953
2954 I915_WRITE(MEMIHYST, 1);
2955
2956 /* Set up min, max, and cur for interrupt handling */
2957 fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
2958 fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
2959 fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
2960 MEMMODE_FSTART_SHIFT;
2961
2962 vstart = (I915_READ(PXVFREQ_BASE + (fstart * 4)) & PXVFREQ_PX_MASK) >>
2963 PXVFREQ_PX_SHIFT;
2964
2965 dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
2966 dev_priv->ips.fstart = fstart;
2967
2968 dev_priv->ips.max_delay = fstart;
2969 dev_priv->ips.min_delay = fmin;
2970 dev_priv->ips.cur_delay = fstart;
2971
2972 DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
2973 fmax, fmin, fstart);
2974
2975 I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
2976
2977 /*
2978 * Interrupts will be enabled in ironlake_irq_postinstall
2979 */
2980
2981 I915_WRITE(VIDSTART, vstart);
2982 POSTING_READ(VIDSTART);
2983
2984 rgvmodectl |= MEMMODE_SWMODE_EN;
2985 I915_WRITE(MEMMODECTL, rgvmodectl);
2986
2987 if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
2988 DRM_ERROR("stuck trying to change perf mode\n");
2989 mdelay(1);
2990
2991 ironlake_set_drps(dev, fstart);
2992
2993 dev_priv->ips.last_count1 = I915_READ(0x112e4) + I915_READ(0x112e8) +
2994 I915_READ(0x112e0);
2995 dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
2996 dev_priv->ips.last_count2 = I915_READ(0x112f4);
2997 getrawmonotonic(&dev_priv->ips.last_time2);
2998
2999 spin_unlock_irq(&mchdev_lock);
3000 }
3001
3002 static void ironlake_disable_drps(struct drm_device *dev)
3003 {
3004 struct drm_i915_private *dev_priv = dev->dev_private;
3005 u16 rgvswctl;
3006
3007 spin_lock_irq(&mchdev_lock);
3008
3009 rgvswctl = I915_READ16(MEMSWCTL);
3010
3011 /* Ack interrupts, disable EFC interrupt */
3012 I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
3013 I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
3014 I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
3015 I915_WRITE(DEIIR, DE_PCU_EVENT);
3016 I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
3017
3018 /* Go back to the starting frequency */
3019 ironlake_set_drps(dev, dev_priv->ips.fstart);
3020 mdelay(1);
3021 rgvswctl |= MEMCTL_CMD_STS;
3022 I915_WRITE(MEMSWCTL, rgvswctl);
3023 mdelay(1);
3024
3025 spin_unlock_irq(&mchdev_lock);
3026 }
3027
3028 /* There's a funny hw issue where the hw returns all 0 when reading from
3029 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
3030 * ourselves, instead of doing a rmw cycle (which might result in us clearing
3031 * all limits and the gpu stuck at whatever frequency it is at atm).
3032 */
3033 static u32 gen6_rps_limits(struct drm_i915_private *dev_priv, u8 val)
3034 {
3035 u32 limits;
3036
3037 /* Only set the down limit when we've reached the lowest level to avoid
3038 * getting more interrupts, otherwise leave this clear. This prevents a
3039 * race in the hw when coming out of rc6: There's a tiny window where
3040 * the hw runs at the minimal clock before selecting the desired
3041 * frequency, if the down threshold expires in that window we will not
3042 * receive a down interrupt. */
3043 limits = dev_priv->rps.max_freq_softlimit << 24;
3044 if (val <= dev_priv->rps.min_freq_softlimit)
3045 limits |= dev_priv->rps.min_freq_softlimit << 16;
3046
3047 return limits;
3048 }
3049
3050 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
3051 {
3052 int new_power;
3053
3054 new_power = dev_priv->rps.power;
3055 switch (dev_priv->rps.power) {
3056 case LOW_POWER:
3057 if (val > dev_priv->rps.efficient_freq + 1 && val > dev_priv->rps.cur_freq)
3058 new_power = BETWEEN;
3059 break;
3060
3061 case BETWEEN:
3062 if (val <= dev_priv->rps.efficient_freq && val < dev_priv->rps.cur_freq)
3063 new_power = LOW_POWER;
3064 else if (val >= dev_priv->rps.rp0_freq && val > dev_priv->rps.cur_freq)
3065 new_power = HIGH_POWER;
3066 break;
3067
3068 case HIGH_POWER:
3069 if (val < (dev_priv->rps.rp1_freq + dev_priv->rps.rp0_freq) >> 1 && val < dev_priv->rps.cur_freq)
3070 new_power = BETWEEN;
3071 break;
3072 }
3073 /* Max/min bins are special */
3074 if (val == dev_priv->rps.min_freq_softlimit)
3075 new_power = LOW_POWER;
3076 if (val == dev_priv->rps.max_freq_softlimit)
3077 new_power = HIGH_POWER;
3078 if (new_power == dev_priv->rps.power)
3079 return;
3080
3081 /* Note the units here are not exactly 1us, but 1280ns. */
3082 switch (new_power) {
3083 case LOW_POWER:
3084 /* Upclock if more than 95% busy over 16ms */
3085 I915_WRITE(GEN6_RP_UP_EI, 12500);
3086 I915_WRITE(GEN6_RP_UP_THRESHOLD, 11800);
3087
3088 /* Downclock if less than 85% busy over 32ms */
3089 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3090 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 21250);
3091
3092 I915_WRITE(GEN6_RP_CONTROL,
3093 GEN6_RP_MEDIA_TURBO |
3094 GEN6_RP_MEDIA_HW_NORMAL_MODE |
3095 GEN6_RP_MEDIA_IS_GFX |
3096 GEN6_RP_ENABLE |
3097 GEN6_RP_UP_BUSY_AVG |
3098 GEN6_RP_DOWN_IDLE_AVG);
3099 break;
3100
3101 case BETWEEN:
3102 /* Upclock if more than 90% busy over 13ms */
3103 I915_WRITE(GEN6_RP_UP_EI, 10250);
3104 I915_WRITE(GEN6_RP_UP_THRESHOLD, 9225);
3105
3106 /* Downclock if less than 75% busy over 32ms */
3107 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3108 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 18750);
3109
3110 I915_WRITE(GEN6_RP_CONTROL,
3111 GEN6_RP_MEDIA_TURBO |
3112 GEN6_RP_MEDIA_HW_NORMAL_MODE |
3113 GEN6_RP_MEDIA_IS_GFX |
3114 GEN6_RP_ENABLE |
3115 GEN6_RP_UP_BUSY_AVG |
3116 GEN6_RP_DOWN_IDLE_AVG);
3117 break;
3118
3119 case HIGH_POWER:
3120 /* Upclock if more than 85% busy over 10ms */
3121 I915_WRITE(GEN6_RP_UP_EI, 8000);
3122 I915_WRITE(GEN6_RP_UP_THRESHOLD, 6800);
3123
3124 /* Downclock if less than 60% busy over 32ms */
3125 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3126 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 15000);
3127
3128 I915_WRITE(GEN6_RP_CONTROL,
3129 GEN6_RP_MEDIA_TURBO |
3130 GEN6_RP_MEDIA_HW_NORMAL_MODE |
3131 GEN6_RP_MEDIA_IS_GFX |
3132 GEN6_RP_ENABLE |
3133 GEN6_RP_UP_BUSY_AVG |
3134 GEN6_RP_DOWN_IDLE_AVG);
3135 break;
3136 }
3137
3138 dev_priv->rps.power = new_power;
3139 dev_priv->rps.last_adj = 0;
3140 }
3141
3142 static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
3143 {
3144 u32 mask = 0;
3145
3146 if (val > dev_priv->rps.min_freq_softlimit)
3147 mask |= GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
3148 if (val < dev_priv->rps.max_freq_softlimit)
3149 mask |= GEN6_PM_RP_UP_THRESHOLD;
3150
3151 /* IVB and SNB hard hangs on looping batchbuffer
3152 * if GEN6_PM_UP_EI_EXPIRED is masked.
3153 */
3154 if (INTEL_INFO(dev_priv->dev)->gen <= 7 && !IS_HASWELL(dev_priv->dev))
3155 mask |= GEN6_PM_RP_UP_EI_EXPIRED;
3156
3157 if (IS_GEN8(dev_priv->dev))
3158 mask |= GEN8_PMINTR_REDIRECT_TO_NON_DISP;
3159
3160 return ~mask;
3161 }
3162
3163 /* gen6_set_rps is called to update the frequency request, but should also be
3164 * called when the range (min_delay and max_delay) is modified so that we can
3165 * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
3166 void gen6_set_rps(struct drm_device *dev, u8 val)
3167 {
3168 struct drm_i915_private *dev_priv = dev->dev_private;
3169
3170 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3171 WARN_ON(val > dev_priv->rps.max_freq_softlimit);
3172 WARN_ON(val < dev_priv->rps.min_freq_softlimit);
3173
3174 /* min/max delay may still have been modified so be sure to
3175 * write the limits value.
3176 */
3177 if (val != dev_priv->rps.cur_freq) {
3178 gen6_set_rps_thresholds(dev_priv, val);
3179
3180 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
3181 I915_WRITE(GEN6_RPNSWREQ,
3182 HSW_FREQUENCY(val));
3183 else
3184 I915_WRITE(GEN6_RPNSWREQ,
3185 GEN6_FREQUENCY(val) |
3186 GEN6_OFFSET(0) |
3187 GEN6_AGGRESSIVE_TURBO);
3188 }
3189
3190 /* Make sure we continue to get interrupts
3191 * until we hit the minimum or maximum frequencies.
3192 */
3193 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, gen6_rps_limits(dev_priv, val));
3194 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
3195
3196 POSTING_READ(GEN6_RPNSWREQ);
3197
3198 dev_priv->rps.cur_freq = val;
3199 trace_intel_gpu_freq_change(val * 50);
3200 }
3201
3202 /* vlv_set_rps_idle: Set the frequency to Rpn if Gfx clocks are down
3203 *
3204 * * If Gfx is Idle, then
3205 * 1. Mask Turbo interrupts
3206 * 2. Bring up Gfx clock
3207 * 3. Change the freq to Rpn and wait till P-Unit updates freq
3208 * 4. Clear the Force GFX CLK ON bit so that Gfx can down
3209 * 5. Unmask Turbo interrupts
3210 */
3211 static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
3212 {
3213 /*
3214 * When we are idle. Drop to min voltage state.
3215 */
3216
3217 if (dev_priv->rps.cur_freq <= dev_priv->rps.min_freq_softlimit)
3218 return;
3219
3220 /* Mask turbo interrupt so that they will not come in between */
3221 I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
3222
3223 vlv_force_gfx_clock(dev_priv, true);
3224
3225 dev_priv->rps.cur_freq = dev_priv->rps.min_freq_softlimit;
3226
3227 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ,
3228 dev_priv->rps.min_freq_softlimit);
3229
3230 if (wait_for(((vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS))
3231 & GENFREQSTATUS) == 0, 5))
3232 DRM_ERROR("timed out waiting for Punit\n");
3233
3234 vlv_force_gfx_clock(dev_priv, false);
3235
3236 I915_WRITE(GEN6_PMINTRMSK,
3237 gen6_rps_pm_mask(dev_priv, dev_priv->rps.cur_freq));
3238 }
3239
3240 void gen6_rps_idle(struct drm_i915_private *dev_priv)
3241 {
3242 struct drm_device *dev = dev_priv->dev;
3243
3244 mutex_lock(&dev_priv->rps.hw_lock);
3245 if (dev_priv->rps.enabled) {
3246 if (IS_VALLEYVIEW(dev))
3247 vlv_set_rps_idle(dev_priv);
3248 else
3249 gen6_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
3250 dev_priv->rps.last_adj = 0;
3251 }
3252 mutex_unlock(&dev_priv->rps.hw_lock);
3253 }
3254
3255 void gen6_rps_boost(struct drm_i915_private *dev_priv)
3256 {
3257 struct drm_device *dev = dev_priv->dev;
3258
3259 mutex_lock(&dev_priv->rps.hw_lock);
3260 if (dev_priv->rps.enabled) {
3261 if (IS_VALLEYVIEW(dev))
3262 valleyview_set_rps(dev_priv->dev, dev_priv->rps.max_freq_softlimit);
3263 else
3264 gen6_set_rps(dev_priv->dev, dev_priv->rps.max_freq_softlimit);
3265 dev_priv->rps.last_adj = 0;
3266 }
3267 mutex_unlock(&dev_priv->rps.hw_lock);
3268 }
3269
3270 void valleyview_set_rps(struct drm_device *dev, u8 val)
3271 {
3272 struct drm_i915_private *dev_priv = dev->dev_private;
3273
3274 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3275 WARN_ON(val > dev_priv->rps.max_freq_softlimit);
3276 WARN_ON(val < dev_priv->rps.min_freq_softlimit);
3277
3278 DRM_DEBUG_DRIVER("GPU freq request from %d MHz (%u) to %d MHz (%u)\n",
3279 vlv_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
3280 dev_priv->rps.cur_freq,
3281 vlv_gpu_freq(dev_priv, val), val);
3282
3283 if (val != dev_priv->rps.cur_freq)
3284 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
3285
3286 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
3287
3288 dev_priv->rps.cur_freq = val;
3289 trace_intel_gpu_freq_change(vlv_gpu_freq(dev_priv, val));
3290 }
3291
3292 static void gen8_disable_rps_interrupts(struct drm_device *dev)
3293 {
3294 struct drm_i915_private *dev_priv = dev->dev_private;
3295
3296 I915_WRITE(GEN6_PMINTRMSK, ~GEN8_PMINTR_REDIRECT_TO_NON_DISP);
3297 I915_WRITE(GEN8_GT_IER(2), I915_READ(GEN8_GT_IER(2)) &
3298 ~dev_priv->pm_rps_events);
3299 /* Complete PM interrupt masking here doesn't race with the rps work
3300 * item again unmasking PM interrupts because that is using a different
3301 * register (GEN8_GT_IMR(2)) to mask PM interrupts. The only risk is in
3302 * leaving stale bits in GEN8_GT_IIR(2) and GEN8_GT_IMR(2) which
3303 * gen8_enable_rps will clean up. */
3304
3305 spin_lock_irq(&dev_priv->irq_lock);
3306 dev_priv->rps.pm_iir = 0;
3307 spin_unlock_irq(&dev_priv->irq_lock);
3308
3309 I915_WRITE(GEN8_GT_IIR(2), dev_priv->pm_rps_events);
3310 }
3311
3312 static void gen6_disable_rps_interrupts(struct drm_device *dev)
3313 {
3314 struct drm_i915_private *dev_priv = dev->dev_private;
3315
3316 I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
3317 I915_WRITE(GEN6_PMIER, I915_READ(GEN6_PMIER) &
3318 ~dev_priv->pm_rps_events);
3319 /* Complete PM interrupt masking here doesn't race with the rps work
3320 * item again unmasking PM interrupts because that is using a different
3321 * register (PMIMR) to mask PM interrupts. The only risk is in leaving
3322 * stale bits in PMIIR and PMIMR which gen6_enable_rps will clean up. */
3323
3324 spin_lock_irq(&dev_priv->irq_lock);
3325 dev_priv->rps.pm_iir = 0;
3326 spin_unlock_irq(&dev_priv->irq_lock);
3327
3328 I915_WRITE(GEN6_PMIIR, dev_priv->pm_rps_events);
3329 }
3330
3331 static void gen6_disable_rps(struct drm_device *dev)
3332 {
3333 struct drm_i915_private *dev_priv = dev->dev_private;
3334
3335 I915_WRITE(GEN6_RC_CONTROL, 0);
3336 I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
3337
3338 if (IS_BROADWELL(dev))
3339 gen8_disable_rps_interrupts(dev);
3340 else
3341 gen6_disable_rps_interrupts(dev);
3342 }
3343
3344 static void valleyview_disable_rps(struct drm_device *dev)
3345 {
3346 struct drm_i915_private *dev_priv = dev->dev_private;
3347
3348 I915_WRITE(GEN6_RC_CONTROL, 0);
3349
3350 gen6_disable_rps_interrupts(dev);
3351 }
3352
3353 static void intel_print_rc6_info(struct drm_device *dev, u32 mode)
3354 {
3355 if (IS_VALLEYVIEW(dev)) {
3356 if (mode & (GEN7_RC_CTL_TO_MODE | GEN6_RC_CTL_EI_MODE(1)))
3357 mode = GEN6_RC_CTL_RC6_ENABLE;
3358 else
3359 mode = 0;
3360 }
3361 DRM_INFO("Enabling RC6 states: RC6 %s, RC6p %s, RC6pp %s\n",
3362 (mode & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off",
3363 (mode & GEN6_RC_CTL_RC6p_ENABLE) ? "on" : "off",
3364 (mode & GEN6_RC_CTL_RC6pp_ENABLE) ? "on" : "off");
3365 }
3366
3367 static int sanitize_rc6_option(const struct drm_device *dev, int enable_rc6)
3368 {
3369 /* No RC6 before Ironlake */
3370 if (INTEL_INFO(dev)->gen < 5)
3371 return 0;
3372
3373 /* RC6 is only on Ironlake mobile not on desktop */
3374 if (INTEL_INFO(dev)->gen == 5 && !IS_IRONLAKE_M(dev))
3375 return 0;
3376
3377 /* Respect the kernel parameter if it is set */
3378 if (enable_rc6 >= 0) {
3379 int mask;
3380
3381 if (INTEL_INFO(dev)->gen == 6 || IS_IVYBRIDGE(dev))
3382 mask = INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE |
3383 INTEL_RC6pp_ENABLE;
3384 else
3385 mask = INTEL_RC6_ENABLE;
3386
3387 if ((enable_rc6 & mask) != enable_rc6)
3388 DRM_INFO("Adjusting RC6 mask to %d (requested %d, valid %d)\n",
3389 enable_rc6 & mask, enable_rc6, mask);
3390
3391 return enable_rc6 & mask;
3392 }
3393
3394 /* Disable RC6 on Ironlake */
3395 if (INTEL_INFO(dev)->gen == 5)
3396 return 0;
3397
3398 if (IS_IVYBRIDGE(dev))
3399 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
3400
3401 return INTEL_RC6_ENABLE;
3402 }
3403
3404 int intel_enable_rc6(const struct drm_device *dev)
3405 {
3406 return i915.enable_rc6;
3407 }
3408
3409 static void gen8_enable_rps_interrupts(struct drm_device *dev)
3410 {
3411 struct drm_i915_private *dev_priv = dev->dev_private;
3412
3413 spin_lock_irq(&dev_priv->irq_lock);
3414 WARN_ON(dev_priv->rps.pm_iir);
3415 bdw_enable_pm_irq(dev_priv, dev_priv->pm_rps_events);
3416 I915_WRITE(GEN8_GT_IIR(2), dev_priv->pm_rps_events);
3417 spin_unlock_irq(&dev_priv->irq_lock);
3418 }
3419
3420 static void gen6_enable_rps_interrupts(struct drm_device *dev)
3421 {
3422 struct drm_i915_private *dev_priv = dev->dev_private;
3423
3424 spin_lock_irq(&dev_priv->irq_lock);
3425 WARN_ON(dev_priv->rps.pm_iir);
3426 snb_enable_pm_irq(dev_priv, dev_priv->pm_rps_events);
3427 I915_WRITE(GEN6_PMIIR, dev_priv->pm_rps_events);
3428 spin_unlock_irq(&dev_priv->irq_lock);
3429 }
3430
3431 static void parse_rp_state_cap(struct drm_i915_private *dev_priv, u32 rp_state_cap)
3432 {
3433 /* All of these values are in units of 50MHz */
3434 dev_priv->rps.cur_freq = 0;
3435 /* static values from HW: RP0 < RPe < RP1 < RPn (min_freq) */
3436 dev_priv->rps.rp1_freq = (rp_state_cap >> 8) & 0xff;
3437 dev_priv->rps.rp0_freq = (rp_state_cap >> 0) & 0xff;
3438 dev_priv->rps.min_freq = (rp_state_cap >> 16) & 0xff;
3439 /* XXX: only BYT has a special efficient freq */
3440 dev_priv->rps.efficient_freq = dev_priv->rps.rp1_freq;
3441 /* hw_max = RP0 until we check for overclocking */
3442 dev_priv->rps.max_freq = dev_priv->rps.rp0_freq;
3443
3444 /* Preserve min/max settings in case of re-init */
3445 if (dev_priv->rps.max_freq_softlimit == 0)
3446 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
3447
3448 if (dev_priv->rps.min_freq_softlimit == 0)
3449 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
3450 }
3451
3452 static void gen8_enable_rps(struct drm_device *dev)
3453 {
3454 struct drm_i915_private *dev_priv = dev->dev_private;
3455 struct intel_engine_cs *ring;
3456 uint32_t rc6_mask = 0, rp_state_cap;
3457 int unused;
3458
3459 /* 1a: Software RC state - RC0 */
3460 I915_WRITE(GEN6_RC_STATE, 0);
3461
3462 /* 1c & 1d: Get forcewake during program sequence. Although the driver
3463 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
3464 gen6_gt_force_wake_get(dev_priv, FORCEWAKE_ALL);
3465
3466 /* 2a: Disable RC states. */
3467 I915_WRITE(GEN6_RC_CONTROL, 0);
3468
3469 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
3470 parse_rp_state_cap(dev_priv, rp_state_cap);
3471
3472 /* 2b: Program RC6 thresholds.*/
3473 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
3474 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
3475 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
3476 for_each_ring(ring, dev_priv, unused)
3477 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
3478 I915_WRITE(GEN6_RC_SLEEP, 0);
3479 I915_WRITE(GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
3480
3481 /* 3: Enable RC6 */
3482 if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
3483 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
3484 intel_print_rc6_info(dev, rc6_mask);
3485 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
3486 GEN6_RC_CTL_EI_MODE(1) |
3487 rc6_mask);
3488
3489 /* 4 Program defaults and thresholds for RPS*/
3490 I915_WRITE(GEN6_RPNSWREQ,
3491 HSW_FREQUENCY(dev_priv->rps.rp1_freq));
3492 I915_WRITE(GEN6_RC_VIDEO_FREQ,
3493 HSW_FREQUENCY(dev_priv->rps.rp1_freq));
3494 /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
3495 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
3496
3497 /* Docs recommend 900MHz, and 300 MHz respectively */
3498 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
3499 dev_priv->rps.max_freq_softlimit << 24 |
3500 dev_priv->rps.min_freq_softlimit << 16);
3501
3502 I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
3503 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
3504 I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
3505 I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
3506
3507 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
3508
3509 /* WaDisablePwrmtrEvent:chv (pre-production hw) */
3510 I915_WRITE(0xA80C, I915_READ(0xA80C) & 0x00ffffff);
3511 I915_WRITE(0xA810, I915_READ(0xA810) & 0xffffff00);
3512
3513 /* 5: Enable RPS */
3514 I915_WRITE(GEN6_RP_CONTROL,
3515 GEN6_RP_MEDIA_TURBO |
3516 GEN6_RP_MEDIA_HW_NORMAL_MODE |
3517 GEN6_RP_MEDIA_IS_GFX | /* WaSetMaskForGfxBusyness:chv (pre-production hw ?) */
3518 GEN6_RP_ENABLE |
3519 GEN6_RP_UP_BUSY_AVG |
3520 GEN6_RP_DOWN_IDLE_AVG);
3521
3522 /* 6: Ring frequency + overclocking (our driver does this later */
3523
3524 gen6_set_rps(dev, (I915_READ(GEN6_GT_PERF_STATUS) & 0xff00) >> 8);
3525
3526 gen8_enable_rps_interrupts(dev);
3527
3528 gen6_gt_force_wake_put(dev_priv, FORCEWAKE_ALL);
3529 }
3530
3531 static void gen6_enable_rps(struct drm_device *dev)
3532 {
3533 struct drm_i915_private *dev_priv = dev->dev_private;
3534 struct intel_engine_cs *ring;
3535 u32 rp_state_cap;
3536 u32 gt_perf_status;
3537 u32 rc6vids, pcu_mbox = 0, rc6_mask = 0;
3538 u32 gtfifodbg;
3539 int rc6_mode;
3540 int i, ret;
3541
3542 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3543
3544 /* Here begins a magic sequence of register writes to enable
3545 * auto-downclocking.
3546 *
3547 * Perhaps there might be some value in exposing these to
3548 * userspace...
3549 */
3550 I915_WRITE(GEN6_RC_STATE, 0);
3551
3552 /* Clear the DBG now so we don't confuse earlier errors */
3553 if ((gtfifodbg = I915_READ(GTFIFODBG))) {
3554 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
3555 I915_WRITE(GTFIFODBG, gtfifodbg);
3556 }
3557
3558 gen6_gt_force_wake_get(dev_priv, FORCEWAKE_ALL);
3559
3560 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
3561 gt_perf_status = I915_READ(GEN6_GT_PERF_STATUS);
3562
3563 parse_rp_state_cap(dev_priv, rp_state_cap);
3564
3565 /* disable the counters and set deterministic thresholds */
3566 I915_WRITE(GEN6_RC_CONTROL, 0);
3567
3568 I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
3569 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
3570 I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
3571 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
3572 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
3573
3574 for_each_ring(ring, dev_priv, i)
3575 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
3576
3577 I915_WRITE(GEN6_RC_SLEEP, 0);
3578 I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
3579 if (IS_IVYBRIDGE(dev))
3580 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
3581 else
3582 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
3583 I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
3584 I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
3585
3586 /* Check if we are enabling RC6 */
3587 rc6_mode = intel_enable_rc6(dev_priv->dev);
3588 if (rc6_mode & INTEL_RC6_ENABLE)
3589 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
3590
3591 /* We don't use those on Haswell */
3592 if (!IS_HASWELL(dev)) {
3593 if (rc6_mode & INTEL_RC6p_ENABLE)
3594 rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
3595
3596 if (rc6_mode & INTEL_RC6pp_ENABLE)
3597 rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
3598 }
3599
3600 intel_print_rc6_info(dev, rc6_mask);
3601
3602 I915_WRITE(GEN6_RC_CONTROL,
3603 rc6_mask |
3604 GEN6_RC_CTL_EI_MODE(1) |
3605 GEN6_RC_CTL_HW_ENABLE);
3606
3607 /* Power down if completely idle for over 50ms */
3608 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
3609 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
3610
3611 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_MIN_FREQ_TABLE, 0);
3612 if (ret)
3613 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
3614
3615 ret = sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &pcu_mbox);
3616 if (!ret && (pcu_mbox & (1<<31))) { /* OC supported */
3617 DRM_DEBUG_DRIVER("Overclocking supported. Max: %dMHz, Overclock max: %dMHz\n",
3618 (dev_priv->rps.max_freq_softlimit & 0xff) * 50,
3619 (pcu_mbox & 0xff) * 50);
3620 dev_priv->rps.max_freq = pcu_mbox & 0xff;
3621 }
3622
3623 dev_priv->rps.power = HIGH_POWER; /* force a reset */
3624 gen6_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
3625
3626 gen6_enable_rps_interrupts(dev);
3627
3628 rc6vids = 0;
3629 ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
3630 if (IS_GEN6(dev) && ret) {
3631 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
3632 } else if (IS_GEN6(dev) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
3633 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
3634 GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
3635 rc6vids &= 0xffff00;
3636 rc6vids |= GEN6_ENCODE_RC6_VID(450);
3637 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
3638 if (ret)
3639 DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
3640 }
3641
3642 gen6_gt_force_wake_put(dev_priv, FORCEWAKE_ALL);
3643 }
3644
3645 static void __gen6_update_ring_freq(struct drm_device *dev)
3646 {
3647 struct drm_i915_private *dev_priv = dev->dev_private;
3648 int min_freq = 15;
3649 unsigned int gpu_freq;
3650 unsigned int max_ia_freq, min_ring_freq;
3651 int scaling_factor = 180;
3652 struct cpufreq_policy *policy;
3653
3654 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3655
3656 policy = cpufreq_cpu_get(0);
3657 if (policy) {
3658 max_ia_freq = policy->cpuinfo.max_freq;
3659 cpufreq_cpu_put(policy);
3660 } else {
3661 /*
3662 * Default to measured freq if none found, PCU will ensure we
3663 * don't go over
3664 */
3665 max_ia_freq = tsc_khz;
3666 }
3667
3668 /* Convert from kHz to MHz */
3669 max_ia_freq /= 1000;
3670
3671 min_ring_freq = I915_READ(DCLK) & 0xf;
3672 /* convert DDR frequency from units of 266.6MHz to bandwidth */
3673 min_ring_freq = mult_frac(min_ring_freq, 8, 3);
3674
3675 /*
3676 * For each potential GPU frequency, load a ring frequency we'd like
3677 * to use for memory access. We do this by specifying the IA frequency
3678 * the PCU should use as a reference to determine the ring frequency.
3679 */
3680 for (gpu_freq = dev_priv->rps.max_freq_softlimit; gpu_freq >= dev_priv->rps.min_freq_softlimit;
3681 gpu_freq--) {
3682 int diff = dev_priv->rps.max_freq_softlimit - gpu_freq;
3683 unsigned int ia_freq = 0, ring_freq = 0;
3684
3685 if (INTEL_INFO(dev)->gen >= 8) {
3686 /* max(2 * GT, DDR). NB: GT is 50MHz units */
3687 ring_freq = max(min_ring_freq, gpu_freq);
3688 } else if (IS_HASWELL(dev)) {
3689 ring_freq = mult_frac(gpu_freq, 5, 4);
3690 ring_freq = max(min_ring_freq, ring_freq);
3691 /* leave ia_freq as the default, chosen by cpufreq */
3692 } else {
3693 /* On older processors, there is no separate ring
3694 * clock domain, so in order to boost the bandwidth
3695 * of the ring, we need to upclock the CPU (ia_freq).
3696 *
3697 * For GPU frequencies less than 750MHz,
3698 * just use the lowest ring freq.
3699 */
3700 if (gpu_freq < min_freq)
3701 ia_freq = 800;
3702 else
3703 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
3704 ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
3705 }
3706
3707 sandybridge_pcode_write(dev_priv,
3708 GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
3709 ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
3710 ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
3711 gpu_freq);
3712 }
3713 }
3714
3715 void gen6_update_ring_freq(struct drm_device *dev)
3716 {
3717 struct drm_i915_private *dev_priv = dev->dev_private;
3718
3719 if (INTEL_INFO(dev)->gen < 6 || IS_VALLEYVIEW(dev))
3720 return;
3721
3722 mutex_lock(&dev_priv->rps.hw_lock);
3723 __gen6_update_ring_freq(dev);
3724 mutex_unlock(&dev_priv->rps.hw_lock);
3725 }
3726
3727 int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
3728 {
3729 u32 val, rp0;
3730
3731 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
3732
3733 rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
3734 /* Clamp to max */
3735 rp0 = min_t(u32, rp0, 0xea);
3736
3737 return rp0;
3738 }
3739
3740 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
3741 {
3742 u32 val, rpe;
3743
3744 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
3745 rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
3746 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
3747 rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
3748
3749 return rpe;
3750 }
3751
3752 int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
3753 {
3754 return vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
3755 }
3756
3757 /* Check that the pctx buffer wasn't move under us. */
3758 static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
3759 {
3760 unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
3761
3762 WARN_ON(pctx_addr != dev_priv->mm.stolen_base +
3763 dev_priv->vlv_pctx->stolen->start);
3764 }
3765
3766 static void valleyview_setup_pctx(struct drm_device *dev)
3767 {
3768 struct drm_i915_private *dev_priv = dev->dev_private;
3769 struct drm_i915_gem_object *pctx;
3770 unsigned long pctx_paddr;
3771 u32 pcbr;
3772 int pctx_size = 24*1024;
3773
3774 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
3775
3776 pcbr = I915_READ(VLV_PCBR);
3777 if (pcbr) {
3778 /* BIOS set it up already, grab the pre-alloc'd space */
3779 int pcbr_offset;
3780
3781 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
3782 pctx = i915_gem_object_create_stolen_for_preallocated(dev_priv->dev,
3783 pcbr_offset,
3784 I915_GTT_OFFSET_NONE,
3785 pctx_size);
3786 goto out;
3787 }
3788
3789 /*
3790 * From the Gunit register HAS:
3791 * The Gfx driver is expected to program this register and ensure
3792 * proper allocation within Gfx stolen memory. For example, this
3793 * register should be programmed such than the PCBR range does not
3794 * overlap with other ranges, such as the frame buffer, protected
3795 * memory, or any other relevant ranges.
3796 */
3797 pctx = i915_gem_object_create_stolen(dev, pctx_size);
3798 if (!pctx) {
3799 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
3800 return;
3801 }
3802
3803 pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
3804 I915_WRITE(VLV_PCBR, pctx_paddr);
3805
3806 out:
3807 dev_priv->vlv_pctx = pctx;
3808 }
3809
3810 static void valleyview_cleanup_pctx(struct drm_device *dev)
3811 {
3812 struct drm_i915_private *dev_priv = dev->dev_private;
3813
3814 if (WARN_ON(!dev_priv->vlv_pctx))
3815 return;
3816
3817 drm_gem_object_unreference(&dev_priv->vlv_pctx->base);
3818 dev_priv->vlv_pctx = NULL;
3819 }
3820
3821 static void valleyview_init_gt_powersave(struct drm_device *dev)
3822 {
3823 struct drm_i915_private *dev_priv = dev->dev_private;
3824
3825 valleyview_setup_pctx(dev);
3826
3827 mutex_lock(&dev_priv->rps.hw_lock);
3828
3829 dev_priv->rps.max_freq = valleyview_rps_max_freq(dev_priv);
3830 dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
3831 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
3832 vlv_gpu_freq(dev_priv, dev_priv->rps.max_freq),
3833 dev_priv->rps.max_freq);
3834
3835 dev_priv->rps.efficient_freq = valleyview_rps_rpe_freq(dev_priv);
3836 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
3837 vlv_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
3838 dev_priv->rps.efficient_freq);
3839
3840 dev_priv->rps.min_freq = valleyview_rps_min_freq(dev_priv);
3841 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
3842 vlv_gpu_freq(dev_priv, dev_priv->rps.min_freq),
3843 dev_priv->rps.min_freq);
3844
3845 /* Preserve min/max settings in case of re-init */
3846 if (dev_priv->rps.max_freq_softlimit == 0)
3847 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
3848
3849 if (dev_priv->rps.min_freq_softlimit == 0)
3850 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
3851
3852 mutex_unlock(&dev_priv->rps.hw_lock);
3853 }
3854
3855 static void valleyview_cleanup_gt_powersave(struct drm_device *dev)
3856 {
3857 valleyview_cleanup_pctx(dev);
3858 }
3859
3860 static void valleyview_enable_rps(struct drm_device *dev)
3861 {
3862 struct drm_i915_private *dev_priv = dev->dev_private;
3863 struct intel_engine_cs *ring;
3864 u32 gtfifodbg, val, rc6_mode = 0;
3865 int i;
3866
3867 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3868
3869 valleyview_check_pctx(dev_priv);
3870
3871 if ((gtfifodbg = I915_READ(GTFIFODBG))) {
3872 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
3873 gtfifodbg);
3874 I915_WRITE(GTFIFODBG, gtfifodbg);
3875 }
3876
3877 /* If VLV, Forcewake all wells, else re-direct to regular path */
3878 gen6_gt_force_wake_get(dev_priv, FORCEWAKE_ALL);
3879
3880 I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
3881 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
3882 I915_WRITE(GEN6_RP_UP_EI, 66000);
3883 I915_WRITE(GEN6_RP_DOWN_EI, 350000);
3884
3885 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
3886
3887 I915_WRITE(GEN6_RP_CONTROL,
3888 GEN6_RP_MEDIA_TURBO |
3889 GEN6_RP_MEDIA_HW_NORMAL_MODE |
3890 GEN6_RP_MEDIA_IS_GFX |
3891 GEN6_RP_ENABLE |
3892 GEN6_RP_UP_BUSY_AVG |
3893 GEN6_RP_DOWN_IDLE_CONT);
3894
3895 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
3896 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
3897 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
3898
3899 for_each_ring(ring, dev_priv, i)
3900 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
3901
3902 I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
3903
3904 /* allows RC6 residency counter to work */
3905 I915_WRITE(VLV_COUNTER_CONTROL,
3906 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
3907 VLV_MEDIA_RC6_COUNT_EN |
3908 VLV_RENDER_RC6_COUNT_EN));
3909 if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
3910 rc6_mode = GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
3911
3912 intel_print_rc6_info(dev, rc6_mode);
3913
3914 I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
3915
3916 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
3917
3918 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", val & 0x10 ? "yes" : "no");
3919 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
3920
3921 dev_priv->rps.cur_freq = (val >> 8) & 0xff;
3922 DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
3923 vlv_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
3924 dev_priv->rps.cur_freq);
3925
3926 DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
3927 vlv_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
3928 dev_priv->rps.efficient_freq);
3929
3930 valleyview_set_rps(dev_priv->dev, dev_priv->rps.efficient_freq);
3931
3932 gen6_enable_rps_interrupts(dev);
3933
3934 gen6_gt_force_wake_put(dev_priv, FORCEWAKE_ALL);
3935 }
3936
3937 void ironlake_teardown_rc6(struct drm_device *dev)
3938 {
3939 struct drm_i915_private *dev_priv = dev->dev_private;
3940
3941 if (dev_priv->ips.renderctx) {
3942 i915_gem_object_ggtt_unpin(dev_priv->ips.renderctx);
3943 drm_gem_object_unreference(&dev_priv->ips.renderctx->base);
3944 dev_priv->ips.renderctx = NULL;
3945 }
3946
3947 if (dev_priv->ips.pwrctx) {
3948 i915_gem_object_ggtt_unpin(dev_priv->ips.pwrctx);
3949 drm_gem_object_unreference(&dev_priv->ips.pwrctx->base);
3950 dev_priv->ips.pwrctx = NULL;
3951 }
3952 }
3953
3954 static void ironlake_disable_rc6(struct drm_device *dev)
3955 {
3956 struct drm_i915_private *dev_priv = dev->dev_private;
3957
3958 if (I915_READ(PWRCTXA)) {
3959 /* Wake the GPU, prevent RC6, then restore RSTDBYCTL */
3960 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) | RCX_SW_EXIT);
3961 wait_for(((I915_READ(RSTDBYCTL) & RSX_STATUS_MASK) == RSX_STATUS_ON),
3962 50);
3963
3964 I915_WRITE(PWRCTXA, 0);
3965 POSTING_READ(PWRCTXA);
3966
3967 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
3968 POSTING_READ(RSTDBYCTL);
3969 }
3970 }
3971
3972 static int ironlake_setup_rc6(struct drm_device *dev)
3973 {
3974 struct drm_i915_private *dev_priv = dev->dev_private;
3975
3976 if (dev_priv->ips.renderctx == NULL)
3977 dev_priv->ips.renderctx = intel_alloc_context_page(dev);
3978 if (!dev_priv->ips.renderctx)
3979 return -ENOMEM;
3980
3981 if (dev_priv->ips.pwrctx == NULL)
3982 dev_priv->ips.pwrctx = intel_alloc_context_page(dev);
3983 if (!dev_priv->ips.pwrctx) {
3984 ironlake_teardown_rc6(dev);
3985 return -ENOMEM;
3986 }
3987
3988 return 0;
3989 }
3990
3991 static void ironlake_enable_rc6(struct drm_device *dev)
3992 {
3993 struct drm_i915_private *dev_priv = dev->dev_private;
3994 struct intel_engine_cs *ring = &dev_priv->ring[RCS];
3995 bool was_interruptible;
3996 int ret;
3997
3998 /* rc6 disabled by default due to repeated reports of hanging during
3999 * boot and resume.
4000 */
4001 if (!intel_enable_rc6(dev))
4002 return;
4003
4004 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
4005
4006 ret = ironlake_setup_rc6(dev);
4007 if (ret)
4008 return;
4009
4010 was_interruptible = dev_priv->mm.interruptible;
4011 dev_priv->mm.interruptible = false;
4012
4013 /*
4014 * GPU can automatically power down the render unit if given a page
4015 * to save state.
4016 */
4017 ret = intel_ring_begin(ring, 6);
4018 if (ret) {
4019 ironlake_teardown_rc6(dev);
4020 dev_priv->mm.interruptible = was_interruptible;
4021 return;
4022 }
4023
4024 intel_ring_emit(ring, MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN);
4025 intel_ring_emit(ring, MI_SET_CONTEXT);
4026 intel_ring_emit(ring, i915_gem_obj_ggtt_offset(dev_priv->ips.renderctx) |
4027 MI_MM_SPACE_GTT |
4028 MI_SAVE_EXT_STATE_EN |
4029 MI_RESTORE_EXT_STATE_EN |
4030 MI_RESTORE_INHIBIT);
4031 intel_ring_emit(ring, MI_SUSPEND_FLUSH);
4032 intel_ring_emit(ring, MI_NOOP);
4033 intel_ring_emit(ring, MI_FLUSH);
4034 intel_ring_advance(ring);
4035
4036 /*
4037 * Wait for the command parser to advance past MI_SET_CONTEXT. The HW
4038 * does an implicit flush, combined with MI_FLUSH above, it should be
4039 * safe to assume that renderctx is valid
4040 */
4041 ret = intel_ring_idle(ring);
4042 dev_priv->mm.interruptible = was_interruptible;
4043 if (ret) {
4044 DRM_ERROR("failed to enable ironlake power savings\n");
4045 ironlake_teardown_rc6(dev);
4046 return;
4047 }
4048
4049 I915_WRITE(PWRCTXA, i915_gem_obj_ggtt_offset(dev_priv->ips.pwrctx) | PWRCTX_EN);
4050 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
4051
4052 intel_print_rc6_info(dev, GEN6_RC_CTL_RC6_ENABLE);
4053 }
4054
4055 static unsigned long intel_pxfreq(u32 vidfreq)
4056 {
4057 unsigned long freq;
4058 int div = (vidfreq & 0x3f0000) >> 16;
4059 int post = (vidfreq & 0x3000) >> 12;
4060 int pre = (vidfreq & 0x7);
4061
4062 if (!pre)
4063 return 0;
4064
4065 freq = ((div * 133333) / ((1<<post) * pre));
4066
4067 return freq;
4068 }
4069
4070 static const struct cparams {
4071 u16 i;
4072 u16 t;
4073 u16 m;
4074 u16 c;
4075 } cparams[] = {
4076 { 1, 1333, 301, 28664 },
4077 { 1, 1066, 294, 24460 },
4078 { 1, 800, 294, 25192 },
4079 { 0, 1333, 276, 27605 },
4080 { 0, 1066, 276, 27605 },
4081 { 0, 800, 231, 23784 },
4082 };
4083
4084 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
4085 {
4086 u64 total_count, diff, ret;
4087 u32 count1, count2, count3, m = 0, c = 0;
4088 unsigned long now = jiffies_to_msecs(jiffies), diff1;
4089 int i;
4090
4091 assert_spin_locked(&mchdev_lock);
4092
4093 diff1 = now - dev_priv->ips.last_time1;
4094
4095 /* Prevent division-by-zero if we are asking too fast.
4096 * Also, we don't get interesting results if we are polling
4097 * faster than once in 10ms, so just return the saved value
4098 * in such cases.
4099 */
4100 if (diff1 <= 10)
4101 return dev_priv->ips.chipset_power;
4102
4103 count1 = I915_READ(DMIEC);
4104 count2 = I915_READ(DDREC);
4105 count3 = I915_READ(CSIEC);
4106
4107 total_count = count1 + count2 + count3;
4108
4109 /* FIXME: handle per-counter overflow */
4110 if (total_count < dev_priv->ips.last_count1) {
4111 diff = ~0UL - dev_priv->ips.last_count1;
4112 diff += total_count;
4113 } else {
4114 diff = total_count - dev_priv->ips.last_count1;
4115 }
4116
4117 for (i = 0; i < ARRAY_SIZE(cparams); i++) {
4118 if (cparams[i].i == dev_priv->ips.c_m &&
4119 cparams[i].t == dev_priv->ips.r_t) {
4120 m = cparams[i].m;
4121 c = cparams[i].c;
4122 break;
4123 }
4124 }
4125
4126 diff = div_u64(diff, diff1);
4127 ret = ((m * diff) + c);
4128 ret = div_u64(ret, 10);
4129
4130 dev_priv->ips.last_count1 = total_count;
4131 dev_priv->ips.last_time1 = now;
4132
4133 dev_priv->ips.chipset_power = ret;
4134
4135 return ret;
4136 }
4137
4138 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
4139 {
4140 struct drm_device *dev = dev_priv->dev;
4141 unsigned long val;
4142
4143 if (INTEL_INFO(dev)->gen != 5)
4144 return 0;
4145
4146 spin_lock_irq(&mchdev_lock);
4147
4148 val = __i915_chipset_val(dev_priv);
4149
4150 spin_unlock_irq(&mchdev_lock);
4151
4152 return val;
4153 }
4154
4155 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
4156 {
4157 unsigned long m, x, b;
4158 u32 tsfs;
4159
4160 tsfs = I915_READ(TSFS);
4161
4162 m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
4163 x = I915_READ8(TR1);
4164
4165 b = tsfs & TSFS_INTR_MASK;
4166
4167 return ((m * x) / 127) - b;
4168 }
4169
4170 static u16 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
4171 {
4172 struct drm_device *dev = dev_priv->dev;
4173 static const struct v_table {
4174 u16 vd; /* in .1 mil */
4175 u16 vm; /* in .1 mil */
4176 } v_table[] = {
4177 { 0, 0, },
4178 { 375, 0, },
4179 { 500, 0, },
4180 { 625, 0, },
4181 { 750, 0, },
4182 { 875, 0, },
4183 { 1000, 0, },
4184 { 1125, 0, },
4185 { 4125, 3000, },
4186 { 4125, 3000, },
4187 { 4125, 3000, },
4188 { 4125, 3000, },
4189 { 4125, 3000, },
4190 { 4125, 3000, },
4191 { 4125, 3000, },
4192 { 4125, 3000, },
4193 { 4125, 3000, },
4194 { 4125, 3000, },
4195 { 4125, 3000, },
4196 { 4125, 3000, },
4197 { 4125, 3000, },
4198 { 4125, 3000, },
4199 { 4125, 3000, },
4200 { 4125, 3000, },
4201 { 4125, 3000, },
4202 { 4125, 3000, },
4203 { 4125, 3000, },
4204 { 4125, 3000, },
4205 { 4125, 3000, },
4206 { 4125, 3000, },
4207 { 4125, 3000, },
4208 { 4125, 3000, },
4209 { 4250, 3125, },
4210 { 4375, 3250, },
4211 { 4500, 3375, },
4212 { 4625, 3500, },
4213 { 4750, 3625, },
4214 { 4875, 3750, },
4215 { 5000, 3875, },
4216 { 5125, 4000, },
4217 { 5250, 4125, },
4218 { 5375, 4250, },
4219 { 5500, 4375, },
4220 { 5625, 4500, },
4221 { 5750, 4625, },
4222 { 5875, 4750, },
4223 { 6000, 4875, },
4224 { 6125, 5000, },
4225 { 6250, 5125, },
4226 { 6375, 5250, },
4227 { 6500, 5375, },
4228 { 6625, 5500, },
4229 { 6750, 5625, },
4230 { 6875, 5750, },
4231 { 7000, 5875, },
4232 { 7125, 6000, },
4233 { 7250, 6125, },
4234 { 7375, 6250, },
4235 { 7500, 6375, },
4236 { 7625, 6500, },
4237 { 7750, 6625, },
4238 { 7875, 6750, },
4239 { 8000, 6875, },
4240 { 8125, 7000, },
4241 { 8250, 7125, },
4242 { 8375, 7250, },
4243 { 8500, 7375, },
4244 { 8625, 7500, },
4245 { 8750, 7625, },
4246 { 8875, 7750, },
4247 { 9000, 7875, },
4248 { 9125, 8000, },
4249 { 9250, 8125, },
4250 { 9375, 8250, },
4251 { 9500, 8375, },
4252 { 9625, 8500, },
4253 { 9750, 8625, },
4254 { 9875, 8750, },
4255 { 10000, 8875, },
4256 { 10125, 9000, },
4257 { 10250, 9125, },
4258 { 10375, 9250, },
4259 { 10500, 9375, },
4260 { 10625, 9500, },
4261 { 10750, 9625, },
4262 { 10875, 9750, },
4263 { 11000, 9875, },
4264 { 11125, 10000, },
4265 { 11250, 10125, },
4266 { 11375, 10250, },
4267 { 11500, 10375, },
4268 { 11625, 10500, },
4269 { 11750, 10625, },
4270 { 11875, 10750, },
4271 { 12000, 10875, },
4272 { 12125, 11000, },
4273 { 12250, 11125, },
4274 { 12375, 11250, },
4275 { 12500, 11375, },
4276 { 12625, 11500, },
4277 { 12750, 11625, },
4278 { 12875, 11750, },
4279 { 13000, 11875, },
4280 { 13125, 12000, },
4281 { 13250, 12125, },
4282 { 13375, 12250, },
4283 { 13500, 12375, },
4284 { 13625, 12500, },
4285 { 13750, 12625, },
4286 { 13875, 12750, },
4287 { 14000, 12875, },
4288 { 14125, 13000, },
4289 { 14250, 13125, },
4290 { 14375, 13250, },
4291 { 14500, 13375, },
4292 { 14625, 13500, },
4293 { 14750, 13625, },
4294 { 14875, 13750, },
4295 { 15000, 13875, },
4296 { 15125, 14000, },
4297 { 15250, 14125, },
4298 { 15375, 14250, },
4299 { 15500, 14375, },
4300 { 15625, 14500, },
4301 { 15750, 14625, },
4302 { 15875, 14750, },
4303 { 16000, 14875, },
4304 { 16125, 15000, },
4305 };
4306 if (INTEL_INFO(dev)->is_mobile)
4307 return v_table[pxvid].vm;
4308 else
4309 return v_table[pxvid].vd;
4310 }
4311
4312 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
4313 {
4314 struct timespec now, diff1;
4315 u64 diff;
4316 unsigned long diffms;
4317 u32 count;
4318
4319 assert_spin_locked(&mchdev_lock);
4320
4321 getrawmonotonic(&now);
4322 diff1 = timespec_sub(now, dev_priv->ips.last_time2);
4323
4324 /* Don't divide by 0 */
4325 diffms = diff1.tv_sec * 1000 + diff1.tv_nsec / 1000000;
4326 if (!diffms)
4327 return;
4328
4329 count = I915_READ(GFXEC);
4330
4331 if (count < dev_priv->ips.last_count2) {
4332 diff = ~0UL - dev_priv->ips.last_count2;
4333 diff += count;
4334 } else {
4335 diff = count - dev_priv->ips.last_count2;
4336 }
4337
4338 dev_priv->ips.last_count2 = count;
4339 dev_priv->ips.last_time2 = now;
4340
4341 /* More magic constants... */
4342 diff = diff * 1181;
4343 diff = div_u64(diff, diffms * 10);
4344 dev_priv->ips.gfx_power = diff;
4345 }
4346
4347 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
4348 {
4349 struct drm_device *dev = dev_priv->dev;
4350
4351 if (INTEL_INFO(dev)->gen != 5)
4352 return;
4353
4354 spin_lock_irq(&mchdev_lock);
4355
4356 __i915_update_gfx_val(dev_priv);
4357
4358 spin_unlock_irq(&mchdev_lock);
4359 }
4360
4361 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
4362 {
4363 unsigned long t, corr, state1, corr2, state2;
4364 u32 pxvid, ext_v;
4365
4366 assert_spin_locked(&mchdev_lock);
4367
4368 pxvid = I915_READ(PXVFREQ_BASE + (dev_priv->rps.cur_freq * 4));
4369 pxvid = (pxvid >> 24) & 0x7f;
4370 ext_v = pvid_to_extvid(dev_priv, pxvid);
4371
4372 state1 = ext_v;
4373
4374 t = i915_mch_val(dev_priv);
4375
4376 /* Revel in the empirically derived constants */
4377
4378 /* Correction factor in 1/100000 units */
4379 if (t > 80)
4380 corr = ((t * 2349) + 135940);
4381 else if (t >= 50)
4382 corr = ((t * 964) + 29317);
4383 else /* < 50 */
4384 corr = ((t * 301) + 1004);
4385
4386 corr = corr * ((150142 * state1) / 10000 - 78642);
4387 corr /= 100000;
4388 corr2 = (corr * dev_priv->ips.corr);
4389
4390 state2 = (corr2 * state1) / 10000;
4391 state2 /= 100; /* convert to mW */
4392
4393 __i915_update_gfx_val(dev_priv);
4394
4395 return dev_priv->ips.gfx_power + state2;
4396 }
4397
4398 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
4399 {
4400 struct drm_device *dev = dev_priv->dev;
4401 unsigned long val;
4402
4403 if (INTEL_INFO(dev)->gen != 5)
4404 return 0;
4405
4406 spin_lock_irq(&mchdev_lock);
4407
4408 val = __i915_gfx_val(dev_priv);
4409
4410 spin_unlock_irq(&mchdev_lock);
4411
4412 return val;
4413 }
4414
4415 /**
4416 * i915_read_mch_val - return value for IPS use
4417 *
4418 * Calculate and return a value for the IPS driver to use when deciding whether
4419 * we have thermal and power headroom to increase CPU or GPU power budget.
4420 */
4421 unsigned long i915_read_mch_val(void)
4422 {
4423 struct drm_i915_private *dev_priv;
4424 unsigned long chipset_val, graphics_val, ret = 0;
4425
4426 spin_lock_irq(&mchdev_lock);
4427 if (!i915_mch_dev)
4428 goto out_unlock;
4429 dev_priv = i915_mch_dev;
4430
4431 chipset_val = __i915_chipset_val(dev_priv);
4432 graphics_val = __i915_gfx_val(dev_priv);
4433
4434 ret = chipset_val + graphics_val;
4435
4436 out_unlock:
4437 spin_unlock_irq(&mchdev_lock);
4438
4439 return ret;
4440 }
4441 EXPORT_SYMBOL_GPL(i915_read_mch_val);
4442
4443 /**
4444 * i915_gpu_raise - raise GPU frequency limit
4445 *
4446 * Raise the limit; IPS indicates we have thermal headroom.
4447 */
4448 bool i915_gpu_raise(void)
4449 {
4450 struct drm_i915_private *dev_priv;
4451 bool ret = true;
4452
4453 spin_lock_irq(&mchdev_lock);
4454 if (!i915_mch_dev) {
4455 ret = false;
4456 goto out_unlock;
4457 }
4458 dev_priv = i915_mch_dev;
4459
4460 if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
4461 dev_priv->ips.max_delay--;
4462
4463 out_unlock:
4464 spin_unlock_irq(&mchdev_lock);
4465
4466 return ret;
4467 }
4468 EXPORT_SYMBOL_GPL(i915_gpu_raise);
4469
4470 /**
4471 * i915_gpu_lower - lower GPU frequency limit
4472 *
4473 * IPS indicates we're close to a thermal limit, so throttle back the GPU
4474 * frequency maximum.
4475 */
4476 bool i915_gpu_lower(void)
4477 {
4478 struct drm_i915_private *dev_priv;
4479 bool ret = true;
4480
4481 spin_lock_irq(&mchdev_lock);
4482 if (!i915_mch_dev) {
4483 ret = false;
4484 goto out_unlock;
4485 }
4486 dev_priv = i915_mch_dev;
4487
4488 if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
4489 dev_priv->ips.max_delay++;
4490
4491 out_unlock:
4492 spin_unlock_irq(&mchdev_lock);
4493
4494 return ret;
4495 }
4496 EXPORT_SYMBOL_GPL(i915_gpu_lower);
4497
4498 /**
4499 * i915_gpu_busy - indicate GPU business to IPS
4500 *
4501 * Tell the IPS driver whether or not the GPU is busy.
4502 */
4503 bool i915_gpu_busy(void)
4504 {
4505 struct drm_i915_private *dev_priv;
4506 struct intel_engine_cs *ring;
4507 bool ret = false;
4508 int i;
4509
4510 spin_lock_irq(&mchdev_lock);
4511 if (!i915_mch_dev)
4512 goto out_unlock;
4513 dev_priv = i915_mch_dev;
4514
4515 for_each_ring(ring, dev_priv, i)
4516 ret |= !list_empty(&ring->request_list);
4517
4518 out_unlock:
4519 spin_unlock_irq(&mchdev_lock);
4520
4521 return ret;
4522 }
4523 EXPORT_SYMBOL_GPL(i915_gpu_busy);
4524
4525 /**
4526 * i915_gpu_turbo_disable - disable graphics turbo
4527 *
4528 * Disable graphics turbo by resetting the max frequency and setting the
4529 * current frequency to the default.
4530 */
4531 bool i915_gpu_turbo_disable(void)
4532 {
4533 struct drm_i915_private *dev_priv;
4534 bool ret = true;
4535
4536 spin_lock_irq(&mchdev_lock);
4537 if (!i915_mch_dev) {
4538 ret = false;
4539 goto out_unlock;
4540 }
4541 dev_priv = i915_mch_dev;
4542
4543 dev_priv->ips.max_delay = dev_priv->ips.fstart;
4544
4545 if (!ironlake_set_drps(dev_priv->dev, dev_priv->ips.fstart))
4546 ret = false;
4547
4548 out_unlock:
4549 spin_unlock_irq(&mchdev_lock);
4550
4551 return ret;
4552 }
4553 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
4554
4555 /**
4556 * Tells the intel_ips driver that the i915 driver is now loaded, if
4557 * IPS got loaded first.
4558 *
4559 * This awkward dance is so that neither module has to depend on the
4560 * other in order for IPS to do the appropriate communication of
4561 * GPU turbo limits to i915.
4562 */
4563 static void
4564 ips_ping_for_i915_load(void)
4565 {
4566 void (*link)(void);
4567
4568 link = symbol_get(ips_link_to_i915_driver);
4569 if (link) {
4570 link();
4571 symbol_put(ips_link_to_i915_driver);
4572 }
4573 }
4574
4575 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
4576 {
4577 /* We only register the i915 ips part with intel-ips once everything is
4578 * set up, to avoid intel-ips sneaking in and reading bogus values. */
4579 spin_lock_irq(&mchdev_lock);
4580 i915_mch_dev = dev_priv;
4581 spin_unlock_irq(&mchdev_lock);
4582
4583 ips_ping_for_i915_load();
4584 }
4585
4586 void intel_gpu_ips_teardown(void)
4587 {
4588 spin_lock_irq(&mchdev_lock);
4589 i915_mch_dev = NULL;
4590 spin_unlock_irq(&mchdev_lock);
4591 }
4592
4593 static void intel_init_emon(struct drm_device *dev)
4594 {
4595 struct drm_i915_private *dev_priv = dev->dev_private;
4596 u32 lcfuse;
4597 u8 pxw[16];
4598 int i;
4599
4600 /* Disable to program */
4601 I915_WRITE(ECR, 0);
4602 POSTING_READ(ECR);
4603
4604 /* Program energy weights for various events */
4605 I915_WRITE(SDEW, 0x15040d00);
4606 I915_WRITE(CSIEW0, 0x007f0000);
4607 I915_WRITE(CSIEW1, 0x1e220004);
4608 I915_WRITE(CSIEW2, 0x04000004);
4609
4610 for (i = 0; i < 5; i++)
4611 I915_WRITE(PEW + (i * 4), 0);
4612 for (i = 0; i < 3; i++)
4613 I915_WRITE(DEW + (i * 4), 0);
4614
4615 /* Program P-state weights to account for frequency power adjustment */
4616 for (i = 0; i < 16; i++) {
4617 u32 pxvidfreq = I915_READ(PXVFREQ_BASE + (i * 4));
4618 unsigned long freq = intel_pxfreq(pxvidfreq);
4619 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
4620 PXVFREQ_PX_SHIFT;
4621 unsigned long val;
4622
4623 val = vid * vid;
4624 val *= (freq / 1000);
4625 val *= 255;
4626 val /= (127*127*900);
4627 if (val > 0xff)
4628 DRM_ERROR("bad pxval: %ld\n", val);
4629 pxw[i] = val;
4630 }
4631 /* Render standby states get 0 weight */
4632 pxw[14] = 0;
4633 pxw[15] = 0;
4634
4635 for (i = 0; i < 4; i++) {
4636 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
4637 (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
4638 I915_WRITE(PXW + (i * 4), val);
4639 }
4640
4641 /* Adjust magic regs to magic values (more experimental results) */
4642 I915_WRITE(OGW0, 0);
4643 I915_WRITE(OGW1, 0);
4644 I915_WRITE(EG0, 0x00007f00);
4645 I915_WRITE(EG1, 0x0000000e);
4646 I915_WRITE(EG2, 0x000e0000);
4647 I915_WRITE(EG3, 0x68000300);
4648 I915_WRITE(EG4, 0x42000000);
4649 I915_WRITE(EG5, 0x00140031);
4650 I915_WRITE(EG6, 0);
4651 I915_WRITE(EG7, 0);
4652
4653 for (i = 0; i < 8; i++)
4654 I915_WRITE(PXWL + (i * 4), 0);
4655
4656 /* Enable PMON + select events */
4657 I915_WRITE(ECR, 0x80000019);
4658
4659 lcfuse = I915_READ(LCFUSE02);
4660
4661 dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
4662 }
4663
4664 void intel_init_gt_powersave(struct drm_device *dev)
4665 {
4666 i915.enable_rc6 = sanitize_rc6_option(dev, i915.enable_rc6);
4667
4668 if (IS_VALLEYVIEW(dev))
4669 valleyview_init_gt_powersave(dev);
4670 }
4671
4672 void intel_cleanup_gt_powersave(struct drm_device *dev)
4673 {
4674 if (IS_VALLEYVIEW(dev))
4675 valleyview_cleanup_gt_powersave(dev);
4676 }
4677
4678 void intel_disable_gt_powersave(struct drm_device *dev)
4679 {
4680 struct drm_i915_private *dev_priv = dev->dev_private;
4681
4682 /* Interrupts should be disabled already to avoid re-arming. */
4683 WARN_ON(dev->irq_enabled);
4684
4685 if (IS_IRONLAKE_M(dev)) {
4686 ironlake_disable_drps(dev);
4687 ironlake_disable_rc6(dev);
4688 } else if (IS_GEN6(dev) || IS_GEN7(dev) || IS_BROADWELL(dev)) {
4689 if (cancel_delayed_work_sync(&dev_priv->rps.delayed_resume_work))
4690 intel_runtime_pm_put(dev_priv);
4691
4692 cancel_work_sync(&dev_priv->rps.work);
4693 mutex_lock(&dev_priv->rps.hw_lock);
4694 if (IS_VALLEYVIEW(dev))
4695 valleyview_disable_rps(dev);
4696 else
4697 gen6_disable_rps(dev);
4698 dev_priv->rps.enabled = false;
4699 mutex_unlock(&dev_priv->rps.hw_lock);
4700 }
4701 }
4702
4703 static void intel_gen6_powersave_work(struct work_struct *work)
4704 {
4705 struct drm_i915_private *dev_priv =
4706 container_of(work, struct drm_i915_private,
4707 rps.delayed_resume_work.work);
4708 struct drm_device *dev = dev_priv->dev;
4709
4710 mutex_lock(&dev_priv->rps.hw_lock);
4711
4712 if (IS_VALLEYVIEW(dev)) {
4713 valleyview_enable_rps(dev);
4714 } else if (IS_BROADWELL(dev)) {
4715 gen8_enable_rps(dev);
4716 __gen6_update_ring_freq(dev);
4717 } else {
4718 gen6_enable_rps(dev);
4719 __gen6_update_ring_freq(dev);
4720 }
4721 dev_priv->rps.enabled = true;
4722 mutex_unlock(&dev_priv->rps.hw_lock);
4723
4724 intel_runtime_pm_put(dev_priv);
4725 }
4726
4727 void intel_enable_gt_powersave(struct drm_device *dev)
4728 {
4729 struct drm_i915_private *dev_priv = dev->dev_private;
4730
4731 if (IS_IRONLAKE_M(dev)) {
4732 mutex_lock(&dev->struct_mutex);
4733 ironlake_enable_drps(dev);
4734 ironlake_enable_rc6(dev);
4735 intel_init_emon(dev);
4736 mutex_unlock(&dev->struct_mutex);
4737 } else if (IS_GEN6(dev) || IS_GEN7(dev) || IS_BROADWELL(dev)) {
4738 /*
4739 * PCU communication is slow and this doesn't need to be
4740 * done at any specific time, so do this out of our fast path
4741 * to make resume and init faster.
4742 *
4743 * We depend on the HW RC6 power context save/restore
4744 * mechanism when entering D3 through runtime PM suspend. So
4745 * disable RPM until RPS/RC6 is properly setup. We can only
4746 * get here via the driver load/system resume/runtime resume
4747 * paths, so the _noresume version is enough (and in case of
4748 * runtime resume it's necessary).
4749 */
4750 if (schedule_delayed_work(&dev_priv->rps.delayed_resume_work,
4751 round_jiffies_up_relative(HZ)))
4752 intel_runtime_pm_get_noresume(dev_priv);
4753 }
4754 }
4755
4756 void intel_reset_gt_powersave(struct drm_device *dev)
4757 {
4758 struct drm_i915_private *dev_priv = dev->dev_private;
4759
4760 dev_priv->rps.enabled = false;
4761 intel_enable_gt_powersave(dev);
4762 }
4763
4764 static void ibx_init_clock_gating(struct drm_device *dev)
4765 {
4766 struct drm_i915_private *dev_priv = dev->dev_private;
4767
4768 /*
4769 * On Ibex Peak and Cougar Point, we need to disable clock
4770 * gating for the panel power sequencer or it will fail to
4771 * start up when no ports are active.
4772 */
4773 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
4774 }
4775
4776 static void g4x_disable_trickle_feed(struct drm_device *dev)
4777 {
4778 struct drm_i915_private *dev_priv = dev->dev_private;
4779 int pipe;
4780
4781 for_each_pipe(pipe) {
4782 I915_WRITE(DSPCNTR(pipe),
4783 I915_READ(DSPCNTR(pipe)) |
4784 DISPPLANE_TRICKLE_FEED_DISABLE);
4785 intel_flush_primary_plane(dev_priv, pipe);
4786 }
4787 }
4788
4789 static void ilk_init_lp_watermarks(struct drm_device *dev)
4790 {
4791 struct drm_i915_private *dev_priv = dev->dev_private;
4792
4793 I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
4794 I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
4795 I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
4796
4797 /*
4798 * Don't touch WM1S_LP_EN here.
4799 * Doing so could cause underruns.
4800 */
4801 }
4802
4803 static void ironlake_init_clock_gating(struct drm_device *dev)
4804 {
4805 struct drm_i915_private *dev_priv = dev->dev_private;
4806 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
4807
4808 /*
4809 * Required for FBC
4810 * WaFbcDisableDpfcClockGating:ilk
4811 */
4812 dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
4813 ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
4814 ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
4815
4816 I915_WRITE(PCH_3DCGDIS0,
4817 MARIUNIT_CLOCK_GATE_DISABLE |
4818 SVSMUNIT_CLOCK_GATE_DISABLE);
4819 I915_WRITE(PCH_3DCGDIS1,
4820 VFMUNIT_CLOCK_GATE_DISABLE);
4821
4822 /*
4823 * According to the spec the following bits should be set in
4824 * order to enable memory self-refresh
4825 * The bit 22/21 of 0x42004
4826 * The bit 5 of 0x42020
4827 * The bit 15 of 0x45000
4828 */
4829 I915_WRITE(ILK_DISPLAY_CHICKEN2,
4830 (I915_READ(ILK_DISPLAY_CHICKEN2) |
4831 ILK_DPARB_GATE | ILK_VSDPFD_FULL));
4832 dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
4833 I915_WRITE(DISP_ARB_CTL,
4834 (I915_READ(DISP_ARB_CTL) |
4835 DISP_FBC_WM_DIS));
4836
4837 ilk_init_lp_watermarks(dev);
4838
4839 /*
4840 * Based on the document from hardware guys the following bits
4841 * should be set unconditionally in order to enable FBC.
4842 * The bit 22 of 0x42000
4843 * The bit 22 of 0x42004
4844 * The bit 7,8,9 of 0x42020.
4845 */
4846 if (IS_IRONLAKE_M(dev)) {
4847 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
4848 I915_WRITE(ILK_DISPLAY_CHICKEN1,
4849 I915_READ(ILK_DISPLAY_CHICKEN1) |
4850 ILK_FBCQ_DIS);
4851 I915_WRITE(ILK_DISPLAY_CHICKEN2,
4852 I915_READ(ILK_DISPLAY_CHICKEN2) |
4853 ILK_DPARB_GATE);
4854 }
4855
4856 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
4857
4858 I915_WRITE(ILK_DISPLAY_CHICKEN2,
4859 I915_READ(ILK_DISPLAY_CHICKEN2) |
4860 ILK_ELPIN_409_SELECT);
4861 I915_WRITE(_3D_CHICKEN2,
4862 _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
4863 _3D_CHICKEN2_WM_READ_PIPELINED);
4864
4865 /* WaDisableRenderCachePipelinedFlush:ilk */
4866 I915_WRITE(CACHE_MODE_0,
4867 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
4868
4869 /* WaDisable_RenderCache_OperationalFlush:ilk */
4870 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
4871
4872 g4x_disable_trickle_feed(dev);
4873
4874 ibx_init_clock_gating(dev);
4875 }
4876
4877 static void cpt_init_clock_gating(struct drm_device *dev)
4878 {
4879 struct drm_i915_private *dev_priv = dev->dev_private;
4880 int pipe;
4881 uint32_t val;
4882
4883 /*
4884 * On Ibex Peak and Cougar Point, we need to disable clock
4885 * gating for the panel power sequencer or it will fail to
4886 * start up when no ports are active.
4887 */
4888 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
4889 PCH_DPLUNIT_CLOCK_GATE_DISABLE |
4890 PCH_CPUNIT_CLOCK_GATE_DISABLE);
4891 I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
4892 DPLS_EDP_PPS_FIX_DIS);
4893 /* The below fixes the weird display corruption, a few pixels shifted
4894 * downward, on (only) LVDS of some HP laptops with IVY.
4895 */
4896 for_each_pipe(pipe) {
4897 val = I915_READ(TRANS_CHICKEN2(pipe));
4898 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
4899 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
4900 if (dev_priv->vbt.fdi_rx_polarity_inverted)
4901 val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
4902 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
4903 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
4904 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
4905 I915_WRITE(TRANS_CHICKEN2(pipe), val);
4906 }
4907 /* WADP0ClockGatingDisable */
4908 for_each_pipe(pipe) {
4909 I915_WRITE(TRANS_CHICKEN1(pipe),
4910 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
4911 }
4912 }
4913
4914 static void gen6_check_mch_setup(struct drm_device *dev)
4915 {
4916 struct drm_i915_private *dev_priv = dev->dev_private;
4917 uint32_t tmp;
4918
4919 tmp = I915_READ(MCH_SSKPD);
4920 if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL) {
4921 DRM_INFO("Wrong MCH_SSKPD value: 0x%08x\n", tmp);
4922 DRM_INFO("This can cause pipe underruns and display issues.\n");
4923 DRM_INFO("Please upgrade your BIOS to fix this.\n");
4924 }
4925 }
4926
4927 static void gen6_init_clock_gating(struct drm_device *dev)
4928 {
4929 struct drm_i915_private *dev_priv = dev->dev_private;
4930 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
4931
4932 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
4933
4934 I915_WRITE(ILK_DISPLAY_CHICKEN2,
4935 I915_READ(ILK_DISPLAY_CHICKEN2) |
4936 ILK_ELPIN_409_SELECT);
4937
4938 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
4939 I915_WRITE(_3D_CHICKEN,
4940 _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
4941
4942 /* WaSetupGtModeTdRowDispatch:snb */
4943 if (IS_SNB_GT1(dev))
4944 I915_WRITE(GEN6_GT_MODE,
4945 _MASKED_BIT_ENABLE(GEN6_TD_FOUR_ROW_DISPATCH_DISABLE));
4946
4947 /* WaDisable_RenderCache_OperationalFlush:snb */
4948 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
4949
4950 /*
4951 * BSpec recoomends 8x4 when MSAA is used,
4952 * however in practice 16x4 seems fastest.
4953 *
4954 * Note that PS/WM thread counts depend on the WIZ hashing
4955 * disable bit, which we don't touch here, but it's good
4956 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
4957 */
4958 I915_WRITE(GEN6_GT_MODE,
4959 GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
4960
4961 ilk_init_lp_watermarks(dev);
4962
4963 I915_WRITE(CACHE_MODE_0,
4964 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
4965
4966 I915_WRITE(GEN6_UCGCTL1,
4967 I915_READ(GEN6_UCGCTL1) |
4968 GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
4969 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
4970
4971 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
4972 * gating disable must be set. Failure to set it results in
4973 * flickering pixels due to Z write ordering failures after
4974 * some amount of runtime in the Mesa "fire" demo, and Unigine
4975 * Sanctuary and Tropics, and apparently anything else with
4976 * alpha test or pixel discard.
4977 *
4978 * According to the spec, bit 11 (RCCUNIT) must also be set,
4979 * but we didn't debug actual testcases to find it out.
4980 *
4981 * WaDisableRCCUnitClockGating:snb
4982 * WaDisableRCPBUnitClockGating:snb
4983 */
4984 I915_WRITE(GEN6_UCGCTL2,
4985 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
4986 GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
4987
4988 /* WaStripsFansDisableFastClipPerformanceFix:snb */
4989 I915_WRITE(_3D_CHICKEN3,
4990 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
4991
4992 /*
4993 * Bspec says:
4994 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
4995 * 3DSTATE_SF number of SF output attributes is more than 16."
4996 */
4997 I915_WRITE(_3D_CHICKEN3,
4998 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
4999
5000 /*
5001 * According to the spec the following bits should be
5002 * set in order to enable memory self-refresh and fbc:
5003 * The bit21 and bit22 of 0x42000
5004 * The bit21 and bit22 of 0x42004
5005 * The bit5 and bit7 of 0x42020
5006 * The bit14 of 0x70180
5007 * The bit14 of 0x71180
5008 *
5009 * WaFbcAsynchFlipDisableFbcQueue:snb
5010 */
5011 I915_WRITE(ILK_DISPLAY_CHICKEN1,
5012 I915_READ(ILK_DISPLAY_CHICKEN1) |
5013 ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
5014 I915_WRITE(ILK_DISPLAY_CHICKEN2,
5015 I915_READ(ILK_DISPLAY_CHICKEN2) |
5016 ILK_DPARB_GATE | ILK_VSDPFD_FULL);
5017 I915_WRITE(ILK_DSPCLK_GATE_D,
5018 I915_READ(ILK_DSPCLK_GATE_D) |
5019 ILK_DPARBUNIT_CLOCK_GATE_ENABLE |
5020 ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
5021
5022 g4x_disable_trickle_feed(dev);
5023
5024 cpt_init_clock_gating(dev);
5025
5026 gen6_check_mch_setup(dev);
5027 }
5028
5029 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
5030 {
5031 uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
5032
5033 /*
5034 * WaVSThreadDispatchOverride:ivb,vlv
5035 *
5036 * This actually overrides the dispatch
5037 * mode for all thread types.
5038 */
5039 reg &= ~GEN7_FF_SCHED_MASK;
5040 reg |= GEN7_FF_TS_SCHED_HW;
5041 reg |= GEN7_FF_VS_SCHED_HW;
5042 reg |= GEN7_FF_DS_SCHED_HW;
5043
5044 I915_WRITE(GEN7_FF_THREAD_MODE, reg);
5045 }
5046
5047 static void lpt_init_clock_gating(struct drm_device *dev)
5048 {
5049 struct drm_i915_private *dev_priv = dev->dev_private;
5050
5051 /*
5052 * TODO: this bit should only be enabled when really needed, then
5053 * disabled when not needed anymore in order to save power.
5054 */
5055 if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE)
5056 I915_WRITE(SOUTH_DSPCLK_GATE_D,
5057 I915_READ(SOUTH_DSPCLK_GATE_D) |
5058 PCH_LP_PARTITION_LEVEL_DISABLE);
5059
5060 /* WADPOClockGatingDisable:hsw */
5061 I915_WRITE(_TRANSA_CHICKEN1,
5062 I915_READ(_TRANSA_CHICKEN1) |
5063 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
5064 }
5065
5066 static void lpt_suspend_hw(struct drm_device *dev)
5067 {
5068 struct drm_i915_private *dev_priv = dev->dev_private;
5069
5070 if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE) {
5071 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
5072
5073 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
5074 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
5075 }
5076 }
5077
5078 static void gen8_init_clock_gating(struct drm_device *dev)
5079 {
5080 struct drm_i915_private *dev_priv = dev->dev_private;
5081 enum pipe pipe;
5082
5083 I915_WRITE(WM3_LP_ILK, 0);
5084 I915_WRITE(WM2_LP_ILK, 0);
5085 I915_WRITE(WM1_LP_ILK, 0);
5086
5087 /* FIXME(BDW): Check all the w/a, some might only apply to
5088 * pre-production hw. */
5089
5090 /* WaDisablePartialInstShootdown:bdw */
5091 I915_WRITE(GEN8_ROW_CHICKEN,
5092 _MASKED_BIT_ENABLE(PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE));
5093
5094 /* WaDisableThreadStallDopClockGating:bdw */
5095 /* FIXME: Unclear whether we really need this on production bdw. */
5096 I915_WRITE(GEN8_ROW_CHICKEN,
5097 _MASKED_BIT_ENABLE(STALL_DOP_GATING_DISABLE));
5098
5099 /*
5100 * This GEN8_CENTROID_PIXEL_OPT_DIS W/A is only needed for
5101 * pre-production hardware
5102 */
5103 I915_WRITE(HALF_SLICE_CHICKEN3,
5104 _MASKED_BIT_ENABLE(GEN8_CENTROID_PIXEL_OPT_DIS));
5105 I915_WRITE(HALF_SLICE_CHICKEN3,
5106 _MASKED_BIT_ENABLE(GEN8_SAMPLER_POWER_BYPASS_DIS));
5107 I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_BWGTLB_DISABLE));
5108
5109 I915_WRITE(_3D_CHICKEN3,
5110 _3D_CHICKEN_SDE_LIMIT_FIFO_POLY_DEPTH(2));
5111
5112 I915_WRITE(COMMON_SLICE_CHICKEN2,
5113 _MASKED_BIT_ENABLE(GEN8_CSC2_SBE_VUE_CACHE_CONSERVATIVE));
5114
5115 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
5116 _MASKED_BIT_ENABLE(GEN7_SINGLE_SUBSCAN_DISPATCH_ENABLE));
5117
5118 /* WaDisableDopClockGating:bdw May not be needed for production */
5119 I915_WRITE(GEN7_ROW_CHICKEN2,
5120 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5121
5122 /* WaSwitchSolVfFArbitrationPriority:bdw */
5123 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
5124
5125 /* WaPsrDPAMaskVBlankInSRD:bdw */
5126 I915_WRITE(CHICKEN_PAR1_1,
5127 I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
5128
5129 /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
5130 for_each_pipe(pipe) {
5131 I915_WRITE(CHICKEN_PIPESL_1(pipe),
5132 I915_READ(CHICKEN_PIPESL_1(pipe)) |
5133 BDW_DPRS_MASK_VBLANK_SRD);
5134 }
5135
5136 /* Use Force Non-Coherent whenever executing a 3D context. This is a
5137 * workaround for for a possible hang in the unlikely event a TLB
5138 * invalidation occurs during a PSD flush.
5139 */
5140 I915_WRITE(HDC_CHICKEN0,
5141 I915_READ(HDC_CHICKEN0) |
5142 _MASKED_BIT_ENABLE(HDC_FORCE_NON_COHERENT));
5143
5144 /* WaVSRefCountFullforceMissDisable:bdw */
5145 /* WaDSRefCountFullforceMissDisable:bdw */
5146 I915_WRITE(GEN7_FF_THREAD_MODE,
5147 I915_READ(GEN7_FF_THREAD_MODE) &
5148 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
5149
5150 /*
5151 * BSpec recommends 8x4 when MSAA is used,
5152 * however in practice 16x4 seems fastest.
5153 *
5154 * Note that PS/WM thread counts depend on the WIZ hashing
5155 * disable bit, which we don't touch here, but it's good
5156 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
5157 */
5158 I915_WRITE(GEN7_GT_MODE,
5159 GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
5160
5161 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
5162 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
5163
5164 /* WaDisableSDEUnitClockGating:bdw */
5165 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
5166 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
5167
5168 /* Wa4x4STCOptimizationDisable:bdw */
5169 I915_WRITE(CACHE_MODE_1,
5170 _MASKED_BIT_ENABLE(GEN8_4x4_STC_OPTIMIZATION_DISABLE));
5171 }
5172
5173 static void haswell_init_clock_gating(struct drm_device *dev)
5174 {
5175 struct drm_i915_private *dev_priv = dev->dev_private;
5176
5177 ilk_init_lp_watermarks(dev);
5178
5179 /* L3 caching of data atomics doesn't work -- disable it. */
5180 I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
5181 I915_WRITE(HSW_ROW_CHICKEN3,
5182 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
5183
5184 /* This is required by WaCatErrorRejectionIssue:hsw */
5185 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5186 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5187 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5188
5189 /* WaVSRefCountFullforceMissDisable:hsw */
5190 I915_WRITE(GEN7_FF_THREAD_MODE,
5191 I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
5192
5193 /* WaDisable_RenderCache_OperationalFlush:hsw */
5194 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5195
5196 /* enable HiZ Raw Stall Optimization */
5197 I915_WRITE(CACHE_MODE_0_GEN7,
5198 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
5199
5200 /* WaDisable4x2SubspanOptimization:hsw */
5201 I915_WRITE(CACHE_MODE_1,
5202 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5203
5204 /*
5205 * BSpec recommends 8x4 when MSAA is used,
5206 * however in practice 16x4 seems fastest.
5207 *
5208 * Note that PS/WM thread counts depend on the WIZ hashing
5209 * disable bit, which we don't touch here, but it's good
5210 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
5211 */
5212 I915_WRITE(GEN7_GT_MODE,
5213 GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
5214
5215 /* WaSwitchSolVfFArbitrationPriority:hsw */
5216 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
5217
5218 /* WaRsPkgCStateDisplayPMReq:hsw */
5219 I915_WRITE(CHICKEN_PAR1_1,
5220 I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
5221
5222 lpt_init_clock_gating(dev);
5223 }
5224
5225 static void ivybridge_init_clock_gating(struct drm_device *dev)
5226 {
5227 struct drm_i915_private *dev_priv = dev->dev_private;
5228 uint32_t snpcr;
5229
5230 ilk_init_lp_watermarks(dev);
5231
5232 I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
5233
5234 /* WaDisableEarlyCull:ivb */
5235 I915_WRITE(_3D_CHICKEN3,
5236 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
5237
5238 /* WaDisableBackToBackFlipFix:ivb */
5239 I915_WRITE(IVB_CHICKEN3,
5240 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
5241 CHICKEN3_DGMG_DONE_FIX_DISABLE);
5242
5243 /* WaDisablePSDDualDispatchEnable:ivb */
5244 if (IS_IVB_GT1(dev))
5245 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
5246 _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
5247
5248 /* WaDisable_RenderCache_OperationalFlush:ivb */
5249 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5250
5251 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
5252 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
5253 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
5254
5255 /* WaApplyL3ControlAndL3ChickenMode:ivb */
5256 I915_WRITE(GEN7_L3CNTLREG1,
5257 GEN7_WA_FOR_GEN7_L3_CONTROL);
5258 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
5259 GEN7_WA_L3_CHICKEN_MODE);
5260 if (IS_IVB_GT1(dev))
5261 I915_WRITE(GEN7_ROW_CHICKEN2,
5262 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5263 else {
5264 /* must write both registers */
5265 I915_WRITE(GEN7_ROW_CHICKEN2,
5266 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5267 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
5268 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5269 }
5270
5271 /* WaForceL3Serialization:ivb */
5272 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
5273 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
5274
5275 /*
5276 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
5277 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
5278 */
5279 I915_WRITE(GEN6_UCGCTL2,
5280 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
5281
5282 /* This is required by WaCatErrorRejectionIssue:ivb */
5283 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5284 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5285 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5286
5287 g4x_disable_trickle_feed(dev);
5288
5289 gen7_setup_fixed_func_scheduler(dev_priv);
5290
5291 if (0) { /* causes HiZ corruption on ivb:gt1 */
5292 /* enable HiZ Raw Stall Optimization */
5293 I915_WRITE(CACHE_MODE_0_GEN7,
5294 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
5295 }
5296
5297 /* WaDisable4x2SubspanOptimization:ivb */
5298 I915_WRITE(CACHE_MODE_1,
5299 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5300
5301 /*
5302 * BSpec recommends 8x4 when MSAA is used,
5303 * however in practice 16x4 seems fastest.
5304 *
5305 * Note that PS/WM thread counts depend on the WIZ hashing
5306 * disable bit, which we don't touch here, but it's good
5307 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
5308 */
5309 I915_WRITE(GEN7_GT_MODE,
5310 GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
5311
5312 snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
5313 snpcr &= ~GEN6_MBC_SNPCR_MASK;
5314 snpcr |= GEN6_MBC_SNPCR_MED;
5315 I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
5316
5317 if (!HAS_PCH_NOP(dev))
5318 cpt_init_clock_gating(dev);
5319
5320 gen6_check_mch_setup(dev);
5321 }
5322
5323 static void valleyview_init_clock_gating(struct drm_device *dev)
5324 {
5325 struct drm_i915_private *dev_priv = dev->dev_private;
5326 u32 val;
5327
5328 mutex_lock(&dev_priv->rps.hw_lock);
5329 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5330 mutex_unlock(&dev_priv->rps.hw_lock);
5331 switch ((val >> 6) & 3) {
5332 case 0:
5333 case 1:
5334 dev_priv->mem_freq = 800;
5335 break;
5336 case 2:
5337 dev_priv->mem_freq = 1066;
5338 break;
5339 case 3:
5340 dev_priv->mem_freq = 1333;
5341 break;
5342 }
5343 DRM_DEBUG_DRIVER("DDR speed: %d MHz", dev_priv->mem_freq);
5344
5345 dev_priv->vlv_cdclk_freq = valleyview_cur_cdclk(dev_priv);
5346 DRM_DEBUG_DRIVER("Current CD clock rate: %d MHz",
5347 dev_priv->vlv_cdclk_freq);
5348
5349 I915_WRITE(DSPCLK_GATE_D, VRHUNIT_CLOCK_GATE_DISABLE);
5350
5351 /* WaDisableEarlyCull:vlv */
5352 I915_WRITE(_3D_CHICKEN3,
5353 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
5354
5355 /* WaDisableBackToBackFlipFix:vlv */
5356 I915_WRITE(IVB_CHICKEN3,
5357 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
5358 CHICKEN3_DGMG_DONE_FIX_DISABLE);
5359
5360 /* WaPsdDispatchEnable:vlv */
5361 /* WaDisablePSDDualDispatchEnable:vlv */
5362 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
5363 _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
5364 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
5365
5366 /* WaDisable_RenderCache_OperationalFlush:vlv */
5367 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5368
5369 /* WaForceL3Serialization:vlv */
5370 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
5371 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
5372
5373 /* WaDisableDopClockGating:vlv */
5374 I915_WRITE(GEN7_ROW_CHICKEN2,
5375 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5376
5377 /* This is required by WaCatErrorRejectionIssue:vlv */
5378 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5379 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5380 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5381
5382 gen7_setup_fixed_func_scheduler(dev_priv);
5383
5384 /*
5385 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
5386 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
5387 */
5388 I915_WRITE(GEN6_UCGCTL2,
5389 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
5390
5391 /* WaDisableL3Bank2xClockGate:vlv
5392 * Disabling L3 clock gating- MMIO 940c[25] = 1
5393 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
5394 I915_WRITE(GEN7_UCGCTL4,
5395 I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
5396
5397 I915_WRITE(MI_ARB_VLV, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE);
5398
5399 /*
5400 * BSpec says this must be set, even though
5401 * WaDisable4x2SubspanOptimization isn't listed for VLV.
5402 */
5403 I915_WRITE(CACHE_MODE_1,
5404 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5405
5406 /*
5407 * WaIncreaseL3CreditsForVLVB0:vlv
5408 * This is the hardware default actually.
5409 */
5410 I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
5411
5412 /*
5413 * WaDisableVLVClockGating_VBIIssue:vlv
5414 * Disable clock gating on th GCFG unit to prevent a delay
5415 * in the reporting of vblank events.
5416 */
5417 I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
5418 }
5419
5420 static void cherryview_init_clock_gating(struct drm_device *dev)
5421 {
5422 struct drm_i915_private *dev_priv = dev->dev_private;
5423
5424 I915_WRITE(DSPCLK_GATE_D, VRHUNIT_CLOCK_GATE_DISABLE);
5425
5426 I915_WRITE(MI_ARB_VLV, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE);
5427
5428 /* WaDisablePartialInstShootdown:chv */
5429 I915_WRITE(GEN8_ROW_CHICKEN,
5430 _MASKED_BIT_ENABLE(PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE));
5431
5432 /* WaDisableThreadStallDopClockGating:chv */
5433 I915_WRITE(GEN8_ROW_CHICKEN,
5434 _MASKED_BIT_ENABLE(STALL_DOP_GATING_DISABLE));
5435
5436 /* WaVSRefCountFullforceMissDisable:chv */
5437 /* WaDSRefCountFullforceMissDisable:chv */
5438 I915_WRITE(GEN7_FF_THREAD_MODE,
5439 I915_READ(GEN7_FF_THREAD_MODE) &
5440 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
5441
5442 /* WaDisableSemaphoreAndSyncFlipWait:chv */
5443 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
5444 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
5445
5446 /* WaDisableCSUnitClockGating:chv */
5447 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
5448 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
5449
5450 /* WaDisableSDEUnitClockGating:chv */
5451 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
5452 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
5453
5454 /* WaDisableSamplerPowerBypass:chv (pre-production hw) */
5455 I915_WRITE(HALF_SLICE_CHICKEN3,
5456 _MASKED_BIT_ENABLE(GEN8_SAMPLER_POWER_BYPASS_DIS));
5457
5458 /* WaDisableGunitClockGating:chv (pre-production hw) */
5459 I915_WRITE(VLV_GUNIT_CLOCK_GATE, I915_READ(VLV_GUNIT_CLOCK_GATE) |
5460 GINT_DIS);
5461
5462 /* WaDisableFfDopClockGating:chv (pre-production hw) */
5463 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
5464 _MASKED_BIT_ENABLE(GEN8_FF_DOP_CLOCK_GATE_DISABLE));
5465
5466 /* WaDisableDopClockGating:chv (pre-production hw) */
5467 I915_WRITE(GEN7_ROW_CHICKEN2,
5468 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5469 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
5470 GEN6_EU_TCUNIT_CLOCK_GATE_DISABLE);
5471 }
5472
5473 static void g4x_init_clock_gating(struct drm_device *dev)
5474 {
5475 struct drm_i915_private *dev_priv = dev->dev_private;
5476 uint32_t dspclk_gate;
5477
5478 I915_WRITE(RENCLK_GATE_D1, 0);
5479 I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
5480 GS_UNIT_CLOCK_GATE_DISABLE |
5481 CL_UNIT_CLOCK_GATE_DISABLE);
5482 I915_WRITE(RAMCLK_GATE_D, 0);
5483 dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
5484 OVRUNIT_CLOCK_GATE_DISABLE |
5485 OVCUNIT_CLOCK_GATE_DISABLE;
5486 if (IS_GM45(dev))
5487 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
5488 I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
5489
5490 /* WaDisableRenderCachePipelinedFlush */
5491 I915_WRITE(CACHE_MODE_0,
5492 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
5493
5494 /* WaDisable_RenderCache_OperationalFlush:g4x */
5495 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5496
5497 g4x_disable_trickle_feed(dev);
5498 }
5499
5500 static void crestline_init_clock_gating(struct drm_device *dev)
5501 {
5502 struct drm_i915_private *dev_priv = dev->dev_private;
5503
5504 I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
5505 I915_WRITE(RENCLK_GATE_D2, 0);
5506 I915_WRITE(DSPCLK_GATE_D, 0);
5507 I915_WRITE(RAMCLK_GATE_D, 0);
5508 I915_WRITE16(DEUC, 0);
5509 I915_WRITE(MI_ARB_STATE,
5510 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
5511
5512 /* WaDisable_RenderCache_OperationalFlush:gen4 */
5513 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5514 }
5515
5516 static void broadwater_init_clock_gating(struct drm_device *dev)
5517 {
5518 struct drm_i915_private *dev_priv = dev->dev_private;
5519
5520 I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
5521 I965_RCC_CLOCK_GATE_DISABLE |
5522 I965_RCPB_CLOCK_GATE_DISABLE |
5523 I965_ISC_CLOCK_GATE_DISABLE |
5524 I965_FBC_CLOCK_GATE_DISABLE);
5525 I915_WRITE(RENCLK_GATE_D2, 0);
5526 I915_WRITE(MI_ARB_STATE,
5527 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
5528
5529 /* WaDisable_RenderCache_OperationalFlush:gen4 */
5530 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5531 }
5532
5533 static void gen3_init_clock_gating(struct drm_device *dev)
5534 {
5535 struct drm_i915_private *dev_priv = dev->dev_private;
5536 u32 dstate = I915_READ(D_STATE);
5537
5538 dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
5539 DSTATE_DOT_CLOCK_GATING;
5540 I915_WRITE(D_STATE, dstate);
5541
5542 if (IS_PINEVIEW(dev))
5543 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
5544
5545 /* IIR "flip pending" means done if this bit is set */
5546 I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
5547
5548 /* interrupts should cause a wake up from C3 */
5549 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
5550
5551 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
5552 I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
5553 }
5554
5555 static void i85x_init_clock_gating(struct drm_device *dev)
5556 {
5557 struct drm_i915_private *dev_priv = dev->dev_private;
5558
5559 I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
5560
5561 /* interrupts should cause a wake up from C3 */
5562 I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
5563 _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
5564 }
5565
5566 static void i830_init_clock_gating(struct drm_device *dev)
5567 {
5568 struct drm_i915_private *dev_priv = dev->dev_private;
5569
5570 I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
5571 }
5572
5573 void intel_init_clock_gating(struct drm_device *dev)
5574 {
5575 struct drm_i915_private *dev_priv = dev->dev_private;
5576
5577 dev_priv->display.init_clock_gating(dev);
5578 }
5579
5580 void intel_suspend_hw(struct drm_device *dev)
5581 {
5582 if (HAS_PCH_LPT(dev))
5583 lpt_suspend_hw(dev);
5584 }
5585
5586 #define for_each_power_well(i, power_well, domain_mask, power_domains) \
5587 for (i = 0; \
5588 i < (power_domains)->power_well_count && \
5589 ((power_well) = &(power_domains)->power_wells[i]); \
5590 i++) \
5591 if ((power_well)->domains & (domain_mask))
5592
5593 #define for_each_power_well_rev(i, power_well, domain_mask, power_domains) \
5594 for (i = (power_domains)->power_well_count - 1; \
5595 i >= 0 && ((power_well) = &(power_domains)->power_wells[i]);\
5596 i--) \
5597 if ((power_well)->domains & (domain_mask))
5598
5599 /**
5600 * We should only use the power well if we explicitly asked the hardware to
5601 * enable it, so check if it's enabled and also check if we've requested it to
5602 * be enabled.
5603 */
5604 static bool hsw_power_well_enabled(struct drm_i915_private *dev_priv,
5605 struct i915_power_well *power_well)
5606 {
5607 return I915_READ(HSW_PWR_WELL_DRIVER) ==
5608 (HSW_PWR_WELL_ENABLE_REQUEST | HSW_PWR_WELL_STATE_ENABLED);
5609 }
5610
5611 bool intel_display_power_enabled_sw(struct drm_i915_private *dev_priv,
5612 enum intel_display_power_domain domain)
5613 {
5614 struct i915_power_domains *power_domains;
5615 struct i915_power_well *power_well;
5616 bool is_enabled;
5617 int i;
5618
5619 if (dev_priv->pm.suspended)
5620 return false;
5621
5622 power_domains = &dev_priv->power_domains;
5623 is_enabled = true;
5624 for_each_power_well_rev(i, power_well, BIT(domain), power_domains) {
5625 if (power_well->always_on)
5626 continue;
5627
5628 if (!power_well->count) {
5629 is_enabled = false;
5630 break;
5631 }
5632 }
5633 return is_enabled;
5634 }
5635
5636 bool intel_display_power_enabled(struct drm_i915_private *dev_priv,
5637 enum intel_display_power_domain domain)
5638 {
5639 struct i915_power_domains *power_domains;
5640 struct i915_power_well *power_well;
5641 bool is_enabled;
5642 int i;
5643
5644 if (dev_priv->pm.suspended)
5645 return false;
5646
5647 power_domains = &dev_priv->power_domains;
5648
5649 is_enabled = true;
5650
5651 mutex_lock(&power_domains->lock);
5652 for_each_power_well_rev(i, power_well, BIT(domain), power_domains) {
5653 if (power_well->always_on)
5654 continue;
5655
5656 if (!power_well->ops->is_enabled(dev_priv, power_well)) {
5657 is_enabled = false;
5658 break;
5659 }
5660 }
5661 mutex_unlock(&power_domains->lock);
5662
5663 return is_enabled;
5664 }
5665
5666 /*
5667 * Starting with Haswell, we have a "Power Down Well" that can be turned off
5668 * when not needed anymore. We have 4 registers that can request the power well
5669 * to be enabled, and it will only be disabled if none of the registers is
5670 * requesting it to be enabled.
5671 */
5672 static void hsw_power_well_post_enable(struct drm_i915_private *dev_priv)
5673 {
5674 struct drm_device *dev = dev_priv->dev;
5675 unsigned long irqflags;
5676
5677 /*
5678 * After we re-enable the power well, if we touch VGA register 0x3d5
5679 * we'll get unclaimed register interrupts. This stops after we write
5680 * anything to the VGA MSR register. The vgacon module uses this
5681 * register all the time, so if we unbind our driver and, as a
5682 * consequence, bind vgacon, we'll get stuck in an infinite loop at
5683 * console_unlock(). So make here we touch the VGA MSR register, making
5684 * sure vgacon can keep working normally without triggering interrupts
5685 * and error messages.
5686 */
5687 vga_get_uninterruptible(dev->pdev, VGA_RSRC_LEGACY_IO);
5688 outb(inb(VGA_MSR_READ), VGA_MSR_WRITE);
5689 vga_put(dev->pdev, VGA_RSRC_LEGACY_IO);
5690
5691 if (IS_BROADWELL(dev)) {
5692 spin_lock_irqsave(&dev_priv->irq_lock, irqflags);
5693 I915_WRITE(GEN8_DE_PIPE_IMR(PIPE_B),
5694 dev_priv->de_irq_mask[PIPE_B]);
5695 I915_WRITE(GEN8_DE_PIPE_IER(PIPE_B),
5696 ~dev_priv->de_irq_mask[PIPE_B] |
5697 GEN8_PIPE_VBLANK);
5698 I915_WRITE(GEN8_DE_PIPE_IMR(PIPE_C),
5699 dev_priv->de_irq_mask[PIPE_C]);
5700 I915_WRITE(GEN8_DE_PIPE_IER(PIPE_C),
5701 ~dev_priv->de_irq_mask[PIPE_C] |
5702 GEN8_PIPE_VBLANK);
5703 POSTING_READ(GEN8_DE_PIPE_IER(PIPE_C));
5704 spin_unlock_irqrestore(&dev_priv->irq_lock, irqflags);
5705 }
5706 }
5707
5708 static void hsw_set_power_well(struct drm_i915_private *dev_priv,
5709 struct i915_power_well *power_well, bool enable)
5710 {
5711 bool is_enabled, enable_requested;
5712 uint32_t tmp;
5713
5714 tmp = I915_READ(HSW_PWR_WELL_DRIVER);
5715 is_enabled = tmp & HSW_PWR_WELL_STATE_ENABLED;
5716 enable_requested = tmp & HSW_PWR_WELL_ENABLE_REQUEST;
5717
5718 if (enable) {
5719 if (!enable_requested)
5720 I915_WRITE(HSW_PWR_WELL_DRIVER,
5721 HSW_PWR_WELL_ENABLE_REQUEST);
5722
5723 if (!is_enabled) {
5724 DRM_DEBUG_KMS("Enabling power well\n");
5725 if (wait_for((I915_READ(HSW_PWR_WELL_DRIVER) &
5726 HSW_PWR_WELL_STATE_ENABLED), 20))
5727 DRM_ERROR("Timeout enabling power well\n");
5728 }
5729
5730 hsw_power_well_post_enable(dev_priv);
5731 } else {
5732 if (enable_requested) {
5733 I915_WRITE(HSW_PWR_WELL_DRIVER, 0);
5734 POSTING_READ(HSW_PWR_WELL_DRIVER);
5735 DRM_DEBUG_KMS("Requesting to disable the power well\n");
5736 }
5737 }
5738 }
5739
5740 static void hsw_power_well_sync_hw(struct drm_i915_private *dev_priv,
5741 struct i915_power_well *power_well)
5742 {
5743 hsw_set_power_well(dev_priv, power_well, power_well->count > 0);
5744
5745 /*
5746 * We're taking over the BIOS, so clear any requests made by it since
5747 * the driver is in charge now.
5748 */
5749 if (I915_READ(HSW_PWR_WELL_BIOS) & HSW_PWR_WELL_ENABLE_REQUEST)
5750 I915_WRITE(HSW_PWR_WELL_BIOS, 0);
5751 }
5752
5753 static void hsw_power_well_enable(struct drm_i915_private *dev_priv,
5754 struct i915_power_well *power_well)
5755 {
5756 hsw_set_power_well(dev_priv, power_well, true);
5757 }
5758
5759 static void hsw_power_well_disable(struct drm_i915_private *dev_priv,
5760 struct i915_power_well *power_well)
5761 {
5762 hsw_set_power_well(dev_priv, power_well, false);
5763 }
5764
5765 static void i9xx_always_on_power_well_noop(struct drm_i915_private *dev_priv,
5766 struct i915_power_well *power_well)
5767 {
5768 }
5769
5770 static bool i9xx_always_on_power_well_enabled(struct drm_i915_private *dev_priv,
5771 struct i915_power_well *power_well)
5772 {
5773 return true;
5774 }
5775
5776 void __vlv_set_power_well(struct drm_i915_private *dev_priv,
5777 enum punit_power_well power_well_id, bool enable)
5778 {
5779 struct drm_device *dev = dev_priv->dev;
5780 u32 mask;
5781 u32 state;
5782 u32 ctrl;
5783 enum pipe pipe;
5784
5785 if (power_well_id == PUNIT_POWER_WELL_DPIO_CMN_BC) {
5786 if (enable) {
5787 /*
5788 * Enable the CRI clock source so we can get at the
5789 * display and the reference clock for VGA
5790 * hotplug / manual detection.
5791 */
5792 I915_WRITE(DPLL(PIPE_B), I915_READ(DPLL(PIPE_B)) |
5793 DPLL_REFA_CLK_ENABLE_VLV |
5794 DPLL_INTEGRATED_CRI_CLK_VLV);
5795 udelay(1); /* >10ns for cmnreset, >0ns for sidereset */
5796 } else {
5797 for_each_pipe(pipe)
5798 assert_pll_disabled(dev_priv, pipe);
5799 /* Assert common reset */
5800 I915_WRITE(DPIO_CTL, I915_READ(DPIO_CTL) &
5801 ~DPIO_CMNRST);
5802 }
5803 }
5804
5805 mask = PUNIT_PWRGT_MASK(power_well_id);
5806 state = enable ? PUNIT_PWRGT_PWR_ON(power_well_id) :
5807 PUNIT_PWRGT_PWR_GATE(power_well_id);
5808
5809 mutex_lock(&dev_priv->rps.hw_lock);
5810
5811 #define COND \
5812 ((vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_STATUS) & mask) == state)
5813
5814 if (COND)
5815 goto out;
5816
5817 ctrl = vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_CTRL);
5818 ctrl &= ~mask;
5819 ctrl |= state;
5820 vlv_punit_write(dev_priv, PUNIT_REG_PWRGT_CTRL, ctrl);
5821
5822 if (wait_for(COND, 100))
5823 DRM_ERROR("timout setting power well state %08x (%08x)\n",
5824 state,
5825 vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_CTRL));
5826
5827 #undef COND
5828
5829 out:
5830 mutex_unlock(&dev_priv->rps.hw_lock);
5831
5832 /*
5833 * From VLV2A0_DP_eDP_DPIO_driver_vbios_notes_10.docx -
5834 * 6. De-assert cmn_reset/side_reset. Same as VLV X0.
5835 * a. GUnit 0x2110 bit[0] set to 1 (def 0)
5836 * b. The other bits such as sfr settings / modesel may all
5837 * be set to 0.
5838 *
5839 * This should only be done on init and resume from S3 with
5840 * both PLLs disabled, or we risk losing DPIO and PLL
5841 * synchronization.
5842 */
5843 if (power_well_id == PUNIT_POWER_WELL_DPIO_CMN_BC && enable)
5844 I915_WRITE(DPIO_CTL, I915_READ(DPIO_CTL) | DPIO_CMNRST);
5845 }
5846
5847 static void vlv_set_power_well(struct drm_i915_private *dev_priv,
5848 struct i915_power_well *power_well, bool enable)
5849 {
5850 enum punit_power_well power_well_id = power_well->data;
5851
5852 __vlv_set_power_well(dev_priv, power_well_id, enable);
5853 }
5854
5855 static void vlv_power_well_sync_hw(struct drm_i915_private *dev_priv,
5856 struct i915_power_well *power_well)
5857 {
5858 vlv_set_power_well(dev_priv, power_well, power_well->count > 0);
5859 }
5860
5861 static void vlv_power_well_enable(struct drm_i915_private *dev_priv,
5862 struct i915_power_well *power_well)
5863 {
5864 vlv_set_power_well(dev_priv, power_well, true);
5865 }
5866
5867 static void vlv_power_well_disable(struct drm_i915_private *dev_priv,
5868 struct i915_power_well *power_well)
5869 {
5870 vlv_set_power_well(dev_priv, power_well, false);
5871 }
5872
5873 static bool vlv_power_well_enabled(struct drm_i915_private *dev_priv,
5874 struct i915_power_well *power_well)
5875 {
5876 int power_well_id = power_well->data;
5877 bool enabled = false;
5878 u32 mask;
5879 u32 state;
5880 u32 ctrl;
5881
5882 mask = PUNIT_PWRGT_MASK(power_well_id);
5883 ctrl = PUNIT_PWRGT_PWR_ON(power_well_id);
5884
5885 mutex_lock(&dev_priv->rps.hw_lock);
5886
5887 state = vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_STATUS) & mask;
5888 /*
5889 * We only ever set the power-on and power-gate states, anything
5890 * else is unexpected.
5891 */
5892 WARN_ON(state != PUNIT_PWRGT_PWR_ON(power_well_id) &&
5893 state != PUNIT_PWRGT_PWR_GATE(power_well_id));
5894 if (state == ctrl)
5895 enabled = true;
5896
5897 /*
5898 * A transient state at this point would mean some unexpected party
5899 * is poking at the power controls too.
5900 */
5901 ctrl = vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_CTRL) & mask;
5902 WARN_ON(ctrl != state);
5903
5904 mutex_unlock(&dev_priv->rps.hw_lock);
5905
5906 return enabled;
5907 }
5908
5909 static void vlv_display_power_well_enable(struct drm_i915_private *dev_priv,
5910 struct i915_power_well *power_well)
5911 {
5912 WARN_ON_ONCE(power_well->data != PUNIT_POWER_WELL_DISP2D);
5913
5914 vlv_set_power_well(dev_priv, power_well, true);
5915
5916 spin_lock_irq(&dev_priv->irq_lock);
5917 valleyview_enable_display_irqs(dev_priv);
5918 spin_unlock_irq(&dev_priv->irq_lock);
5919
5920 /*
5921 * During driver initialization/resume we can avoid restoring the
5922 * part of the HW/SW state that will be inited anyway explicitly.
5923 */
5924 if (dev_priv->power_domains.initializing)
5925 return;
5926
5927 intel_hpd_init(dev_priv->dev);
5928
5929 i915_redisable_vga_power_on(dev_priv->dev);
5930 }
5931
5932 static void vlv_display_power_well_disable(struct drm_i915_private *dev_priv,
5933 struct i915_power_well *power_well)
5934 {
5935 WARN_ON_ONCE(power_well->data != PUNIT_POWER_WELL_DISP2D);
5936
5937 spin_lock_irq(&dev_priv->irq_lock);
5938 valleyview_disable_display_irqs(dev_priv);
5939 spin_unlock_irq(&dev_priv->irq_lock);
5940
5941 vlv_set_power_well(dev_priv, power_well, false);
5942 }
5943
5944 static void check_power_well_state(struct drm_i915_private *dev_priv,
5945 struct i915_power_well *power_well)
5946 {
5947 bool enabled = power_well->ops->is_enabled(dev_priv, power_well);
5948
5949 if (power_well->always_on || !i915.disable_power_well) {
5950 if (!enabled)
5951 goto mismatch;
5952
5953 return;
5954 }
5955
5956 if (enabled != (power_well->count > 0))
5957 goto mismatch;
5958
5959 return;
5960
5961 mismatch:
5962 WARN(1, "state mismatch for '%s' (always_on %d hw state %d use-count %d disable_power_well %d\n",
5963 power_well->name, power_well->always_on, enabled,
5964 power_well->count, i915.disable_power_well);
5965 }
5966
5967 void intel_display_power_get(struct drm_i915_private *dev_priv,
5968 enum intel_display_power_domain domain)
5969 {
5970 struct i915_power_domains *power_domains;
5971 struct i915_power_well *power_well;
5972 int i;
5973
5974 intel_runtime_pm_get(dev_priv);
5975
5976 power_domains = &dev_priv->power_domains;
5977
5978 mutex_lock(&power_domains->lock);
5979
5980 for_each_power_well(i, power_well, BIT(domain), power_domains) {
5981 if (!power_well->count++) {
5982 DRM_DEBUG_KMS("enabling %s\n", power_well->name);
5983 power_well->ops->enable(dev_priv, power_well);
5984 }
5985
5986 check_power_well_state(dev_priv, power_well);
5987 }
5988
5989 power_domains->domain_use_count[domain]++;
5990
5991 mutex_unlock(&power_domains->lock);
5992 }
5993
5994 void intel_display_power_put(struct drm_i915_private *dev_priv,
5995 enum intel_display_power_domain domain)
5996 {
5997 struct i915_power_domains *power_domains;
5998 struct i915_power_well *power_well;
5999 int i;
6000
6001 power_domains = &dev_priv->power_domains;
6002
6003 mutex_lock(&power_domains->lock);
6004
6005 WARN_ON(!power_domains->domain_use_count[domain]);
6006 power_domains->domain_use_count[domain]--;
6007
6008 for_each_power_well_rev(i, power_well, BIT(domain), power_domains) {
6009 WARN_ON(!power_well->count);
6010
6011 if (!--power_well->count && i915.disable_power_well) {
6012 DRM_DEBUG_KMS("disabling %s\n", power_well->name);
6013 power_well->ops->disable(dev_priv, power_well);
6014 }
6015
6016 check_power_well_state(dev_priv, power_well);
6017 }
6018
6019 mutex_unlock(&power_domains->lock);
6020
6021 intel_runtime_pm_put(dev_priv);
6022 }
6023
6024 static struct i915_power_domains *hsw_pwr;
6025
6026 /* Display audio driver power well request */
6027 void i915_request_power_well(void)
6028 {
6029 struct drm_i915_private *dev_priv;
6030
6031 if (WARN_ON(!hsw_pwr))
6032 return;
6033
6034 dev_priv = container_of(hsw_pwr, struct drm_i915_private,
6035 power_domains);
6036 intel_display_power_get(dev_priv, POWER_DOMAIN_AUDIO);
6037 }
6038 EXPORT_SYMBOL_GPL(i915_request_power_well);
6039
6040 /* Display audio driver power well release */
6041 void i915_release_power_well(void)
6042 {
6043 struct drm_i915_private *dev_priv;
6044
6045 if (WARN_ON(!hsw_pwr))
6046 return;
6047
6048 dev_priv = container_of(hsw_pwr, struct drm_i915_private,
6049 power_domains);
6050 intel_display_power_put(dev_priv, POWER_DOMAIN_AUDIO);
6051 }
6052 EXPORT_SYMBOL_GPL(i915_release_power_well);
6053
6054 #define POWER_DOMAIN_MASK (BIT(POWER_DOMAIN_NUM) - 1)
6055
6056 #define HSW_ALWAYS_ON_POWER_DOMAINS ( \
6057 BIT(POWER_DOMAIN_PIPE_A) | \
6058 BIT(POWER_DOMAIN_TRANSCODER_EDP) | \
6059 BIT(POWER_DOMAIN_PORT_DDI_A_2_LANES) | \
6060 BIT(POWER_DOMAIN_PORT_DDI_A_4_LANES) | \
6061 BIT(POWER_DOMAIN_PORT_DDI_B_2_LANES) | \
6062 BIT(POWER_DOMAIN_PORT_DDI_B_4_LANES) | \
6063 BIT(POWER_DOMAIN_PORT_DDI_C_2_LANES) | \
6064 BIT(POWER_DOMAIN_PORT_DDI_C_4_LANES) | \
6065 BIT(POWER_DOMAIN_PORT_DDI_D_2_LANES) | \
6066 BIT(POWER_DOMAIN_PORT_DDI_D_4_LANES) | \
6067 BIT(POWER_DOMAIN_PORT_CRT) | \
6068 BIT(POWER_DOMAIN_INIT))
6069 #define HSW_DISPLAY_POWER_DOMAINS ( \
6070 (POWER_DOMAIN_MASK & ~HSW_ALWAYS_ON_POWER_DOMAINS) | \
6071 BIT(POWER_DOMAIN_INIT))
6072
6073 #define BDW_ALWAYS_ON_POWER_DOMAINS ( \
6074 HSW_ALWAYS_ON_POWER_DOMAINS | \
6075 BIT(POWER_DOMAIN_PIPE_A_PANEL_FITTER))
6076 #define BDW_DISPLAY_POWER_DOMAINS ( \
6077 (POWER_DOMAIN_MASK & ~BDW_ALWAYS_ON_POWER_DOMAINS) | \
6078 BIT(POWER_DOMAIN_INIT))
6079
6080 #define VLV_ALWAYS_ON_POWER_DOMAINS BIT(POWER_DOMAIN_INIT)
6081 #define VLV_DISPLAY_POWER_DOMAINS POWER_DOMAIN_MASK
6082
6083 #define VLV_DPIO_CMN_BC_POWER_DOMAINS ( \
6084 BIT(POWER_DOMAIN_PORT_DDI_B_2_LANES) | \
6085 BIT(POWER_DOMAIN_PORT_DDI_B_4_LANES) | \
6086 BIT(POWER_DOMAIN_PORT_DDI_C_2_LANES) | \
6087 BIT(POWER_DOMAIN_PORT_DDI_C_4_LANES) | \
6088 BIT(POWER_DOMAIN_PORT_CRT) | \
6089 BIT(POWER_DOMAIN_INIT))
6090
6091 #define VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS ( \
6092 BIT(POWER_DOMAIN_PORT_DDI_B_2_LANES) | \
6093 BIT(POWER_DOMAIN_PORT_DDI_B_4_LANES) | \
6094 BIT(POWER_DOMAIN_INIT))
6095
6096 #define VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS ( \
6097 BIT(POWER_DOMAIN_PORT_DDI_B_4_LANES) | \
6098 BIT(POWER_DOMAIN_INIT))
6099
6100 #define VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS ( \
6101 BIT(POWER_DOMAIN_PORT_DDI_C_2_LANES) | \
6102 BIT(POWER_DOMAIN_PORT_DDI_C_4_LANES) | \
6103 BIT(POWER_DOMAIN_INIT))
6104
6105 #define VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS ( \
6106 BIT(POWER_DOMAIN_PORT_DDI_C_4_LANES) | \
6107 BIT(POWER_DOMAIN_INIT))
6108
6109 static const struct i915_power_well_ops i9xx_always_on_power_well_ops = {
6110 .sync_hw = i9xx_always_on_power_well_noop,
6111 .enable = i9xx_always_on_power_well_noop,
6112 .disable = i9xx_always_on_power_well_noop,
6113 .is_enabled = i9xx_always_on_power_well_enabled,
6114 };
6115
6116 static struct i915_power_well i9xx_always_on_power_well[] = {
6117 {
6118 .name = "always-on",
6119 .always_on = 1,
6120 .domains = POWER_DOMAIN_MASK,
6121 .ops = &i9xx_always_on_power_well_ops,
6122 },
6123 };
6124
6125 static const struct i915_power_well_ops hsw_power_well_ops = {
6126 .sync_hw = hsw_power_well_sync_hw,
6127 .enable = hsw_power_well_enable,
6128 .disable = hsw_power_well_disable,
6129 .is_enabled = hsw_power_well_enabled,
6130 };
6131
6132 static struct i915_power_well hsw_power_wells[] = {
6133 {
6134 .name = "always-on",
6135 .always_on = 1,
6136 .domains = HSW_ALWAYS_ON_POWER_DOMAINS,
6137 .ops = &i9xx_always_on_power_well_ops,
6138 },
6139 {
6140 .name = "display",
6141 .domains = HSW_DISPLAY_POWER_DOMAINS,
6142 .ops = &hsw_power_well_ops,
6143 },
6144 };
6145
6146 static struct i915_power_well bdw_power_wells[] = {
6147 {
6148 .name = "always-on",
6149 .always_on = 1,
6150 .domains = BDW_ALWAYS_ON_POWER_DOMAINS,
6151 .ops = &i9xx_always_on_power_well_ops,
6152 },
6153 {
6154 .name = "display",
6155 .domains = BDW_DISPLAY_POWER_DOMAINS,
6156 .ops = &hsw_power_well_ops,
6157 },
6158 };
6159
6160 static const struct i915_power_well_ops vlv_display_power_well_ops = {
6161 .sync_hw = vlv_power_well_sync_hw,
6162 .enable = vlv_display_power_well_enable,
6163 .disable = vlv_display_power_well_disable,
6164 .is_enabled = vlv_power_well_enabled,
6165 };
6166
6167 static const struct i915_power_well_ops vlv_dpio_power_well_ops = {
6168 .sync_hw = vlv_power_well_sync_hw,
6169 .enable = vlv_power_well_enable,
6170 .disable = vlv_power_well_disable,
6171 .is_enabled = vlv_power_well_enabled,
6172 };
6173
6174 static struct i915_power_well vlv_power_wells[] = {
6175 {
6176 .name = "always-on",
6177 .always_on = 1,
6178 .domains = VLV_ALWAYS_ON_POWER_DOMAINS,
6179 .ops = &i9xx_always_on_power_well_ops,
6180 },
6181 {
6182 .name = "display",
6183 .domains = VLV_DISPLAY_POWER_DOMAINS,
6184 .data = PUNIT_POWER_WELL_DISP2D,
6185 .ops = &vlv_display_power_well_ops,
6186 },
6187 {
6188 .name = "dpio-tx-b-01",
6189 .domains = VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS |
6190 VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS |
6191 VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS |
6192 VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS,
6193 .ops = &vlv_dpio_power_well_ops,
6194 .data = PUNIT_POWER_WELL_DPIO_TX_B_LANES_01,
6195 },
6196 {
6197 .name = "dpio-tx-b-23",
6198 .domains = VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS |
6199 VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS |
6200 VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS |
6201 VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS,
6202 .ops = &vlv_dpio_power_well_ops,
6203 .data = PUNIT_POWER_WELL_DPIO_TX_B_LANES_23,
6204 },
6205 {
6206 .name = "dpio-tx-c-01",
6207 .domains = VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS |
6208 VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS |
6209 VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS |
6210 VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS,
6211 .ops = &vlv_dpio_power_well_ops,
6212 .data = PUNIT_POWER_WELL_DPIO_TX_C_LANES_01,
6213 },
6214 {
6215 .name = "dpio-tx-c-23",
6216 .domains = VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS |
6217 VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS |
6218 VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS |
6219 VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS,
6220 .ops = &vlv_dpio_power_well_ops,
6221 .data = PUNIT_POWER_WELL_DPIO_TX_C_LANES_23,
6222 },
6223 {
6224 .name = "dpio-common",
6225 .domains = VLV_DPIO_CMN_BC_POWER_DOMAINS,
6226 .data = PUNIT_POWER_WELL_DPIO_CMN_BC,
6227 .ops = &vlv_dpio_power_well_ops,
6228 },
6229 };
6230
6231 #define set_power_wells(power_domains, __power_wells) ({ \
6232 (power_domains)->power_wells = (__power_wells); \
6233 (power_domains)->power_well_count = ARRAY_SIZE(__power_wells); \
6234 })
6235
6236 int intel_power_domains_init(struct drm_i915_private *dev_priv)
6237 {
6238 struct i915_power_domains *power_domains = &dev_priv->power_domains;
6239
6240 mutex_init(&power_domains->lock);
6241
6242 /*
6243 * The enabling order will be from lower to higher indexed wells,
6244 * the disabling order is reversed.
6245 */
6246 if (IS_HASWELL(dev_priv->dev)) {
6247 set_power_wells(power_domains, hsw_power_wells);
6248 hsw_pwr = power_domains;
6249 } else if (IS_BROADWELL(dev_priv->dev)) {
6250 set_power_wells(power_domains, bdw_power_wells);
6251 hsw_pwr = power_domains;
6252 } else if (IS_VALLEYVIEW(dev_priv->dev)) {
6253 set_power_wells(power_domains, vlv_power_wells);
6254 } else {
6255 set_power_wells(power_domains, i9xx_always_on_power_well);
6256 }
6257
6258 return 0;
6259 }
6260
6261 void intel_power_domains_remove(struct drm_i915_private *dev_priv)
6262 {
6263 hsw_pwr = NULL;
6264 }
6265
6266 static void intel_power_domains_resume(struct drm_i915_private *dev_priv)
6267 {
6268 struct i915_power_domains *power_domains = &dev_priv->power_domains;
6269 struct i915_power_well *power_well;
6270 int i;
6271
6272 mutex_lock(&power_domains->lock);
6273 for_each_power_well(i, power_well, POWER_DOMAIN_MASK, power_domains)
6274 power_well->ops->sync_hw(dev_priv, power_well);
6275 mutex_unlock(&power_domains->lock);
6276 }
6277
6278 void intel_power_domains_init_hw(struct drm_i915_private *dev_priv)
6279 {
6280 struct i915_power_domains *power_domains = &dev_priv->power_domains;
6281
6282 power_domains->initializing = true;
6283 /* For now, we need the power well to be always enabled. */
6284 intel_display_set_init_power(dev_priv, true);
6285 intel_power_domains_resume(dev_priv);
6286 power_domains->initializing = false;
6287 }
6288
6289 void intel_aux_display_runtime_get(struct drm_i915_private *dev_priv)
6290 {
6291 intel_runtime_pm_get(dev_priv);
6292 }
6293
6294 void intel_aux_display_runtime_put(struct drm_i915_private *dev_priv)
6295 {
6296 intel_runtime_pm_put(dev_priv);
6297 }
6298
6299 void intel_runtime_pm_get(struct drm_i915_private *dev_priv)
6300 {
6301 struct drm_device *dev = dev_priv->dev;
6302 struct device *device = &dev->pdev->dev;
6303
6304 if (!HAS_RUNTIME_PM(dev))
6305 return;
6306
6307 pm_runtime_get_sync(device);
6308 WARN(dev_priv->pm.suspended, "Device still suspended.\n");
6309 }
6310
6311 void intel_runtime_pm_get_noresume(struct drm_i915_private *dev_priv)
6312 {
6313 struct drm_device *dev = dev_priv->dev;
6314 struct device *device = &dev->pdev->dev;
6315
6316 if (!HAS_RUNTIME_PM(dev))
6317 return;
6318
6319 WARN(dev_priv->pm.suspended, "Getting nosync-ref while suspended.\n");
6320 pm_runtime_get_noresume(device);
6321 }
6322
6323 void intel_runtime_pm_put(struct drm_i915_private *dev_priv)
6324 {
6325 struct drm_device *dev = dev_priv->dev;
6326 struct device *device = &dev->pdev->dev;
6327
6328 if (!HAS_RUNTIME_PM(dev))
6329 return;
6330
6331 pm_runtime_mark_last_busy(device);
6332 pm_runtime_put_autosuspend(device);
6333 }
6334
6335 void intel_init_runtime_pm(struct drm_i915_private *dev_priv)
6336 {
6337 struct drm_device *dev = dev_priv->dev;
6338 struct device *device = &dev->pdev->dev;
6339
6340 if (!HAS_RUNTIME_PM(dev))
6341 return;
6342
6343 pm_runtime_set_active(device);
6344
6345 /*
6346 * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
6347 * requirement.
6348 */
6349 if (!intel_enable_rc6(dev)) {
6350 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
6351 return;
6352 }
6353
6354 pm_runtime_set_autosuspend_delay(device, 10000); /* 10s */
6355 pm_runtime_mark_last_busy(device);
6356 pm_runtime_use_autosuspend(device);
6357
6358 pm_runtime_put_autosuspend(device);
6359 }
6360
6361 void intel_fini_runtime_pm(struct drm_i915_private *dev_priv)
6362 {
6363 struct drm_device *dev = dev_priv->dev;
6364 struct device *device = &dev->pdev->dev;
6365
6366 if (!HAS_RUNTIME_PM(dev))
6367 return;
6368
6369 if (!intel_enable_rc6(dev))
6370 return;
6371
6372 /* Make sure we're not suspended first. */
6373 pm_runtime_get_sync(device);
6374 pm_runtime_disable(device);
6375 }
6376
6377 /* Set up chip specific power management-related functions */
6378 void intel_init_pm(struct drm_device *dev)
6379 {
6380 struct drm_i915_private *dev_priv = dev->dev_private;
6381
6382 if (HAS_FBC(dev)) {
6383 if (INTEL_INFO(dev)->gen >= 7) {
6384 dev_priv->display.fbc_enabled = ironlake_fbc_enabled;
6385 dev_priv->display.enable_fbc = gen7_enable_fbc;
6386 dev_priv->display.disable_fbc = ironlake_disable_fbc;
6387 } else if (INTEL_INFO(dev)->gen >= 5) {
6388 dev_priv->display.fbc_enabled = ironlake_fbc_enabled;
6389 dev_priv->display.enable_fbc = ironlake_enable_fbc;
6390 dev_priv->display.disable_fbc = ironlake_disable_fbc;
6391 } else if (IS_GM45(dev)) {
6392 dev_priv->display.fbc_enabled = g4x_fbc_enabled;
6393 dev_priv->display.enable_fbc = g4x_enable_fbc;
6394 dev_priv->display.disable_fbc = g4x_disable_fbc;
6395 } else {
6396 dev_priv->display.fbc_enabled = i8xx_fbc_enabled;
6397 dev_priv->display.enable_fbc = i8xx_enable_fbc;
6398 dev_priv->display.disable_fbc = i8xx_disable_fbc;
6399
6400 /* This value was pulled out of someone's hat */
6401 I915_WRITE(FBC_CONTROL, 500 << FBC_CTL_INTERVAL_SHIFT);
6402 }
6403 }
6404
6405 /* For cxsr */
6406 if (IS_PINEVIEW(dev))
6407 i915_pineview_get_mem_freq(dev);
6408 else if (IS_GEN5(dev))
6409 i915_ironlake_get_mem_freq(dev);
6410
6411 /* For FIFO watermark updates */
6412 if (HAS_PCH_SPLIT(dev)) {
6413 ilk_setup_wm_latency(dev);
6414
6415 if ((IS_GEN5(dev) && dev_priv->wm.pri_latency[1] &&
6416 dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
6417 (!IS_GEN5(dev) && dev_priv->wm.pri_latency[0] &&
6418 dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
6419 dev_priv->display.update_wm = ilk_update_wm;
6420 dev_priv->display.update_sprite_wm = ilk_update_sprite_wm;
6421 } else {
6422 DRM_DEBUG_KMS("Failed to read display plane latency. "
6423 "Disable CxSR\n");
6424 }
6425
6426 if (IS_GEN5(dev))
6427 dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
6428 else if (IS_GEN6(dev))
6429 dev_priv->display.init_clock_gating = gen6_init_clock_gating;
6430 else if (IS_IVYBRIDGE(dev))
6431 dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
6432 else if (IS_HASWELL(dev))
6433 dev_priv->display.init_clock_gating = haswell_init_clock_gating;
6434 else if (INTEL_INFO(dev)->gen == 8)
6435 dev_priv->display.init_clock_gating = gen8_init_clock_gating;
6436 } else if (IS_CHERRYVIEW(dev)) {
6437 dev_priv->display.update_wm = valleyview_update_wm;
6438 dev_priv->display.init_clock_gating =
6439 cherryview_init_clock_gating;
6440 } else if (IS_VALLEYVIEW(dev)) {
6441 dev_priv->display.update_wm = valleyview_update_wm;
6442 dev_priv->display.init_clock_gating =
6443 valleyview_init_clock_gating;
6444 } else if (IS_PINEVIEW(dev)) {
6445 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
6446 dev_priv->is_ddr3,
6447 dev_priv->fsb_freq,
6448 dev_priv->mem_freq)) {
6449 DRM_INFO("failed to find known CxSR latency "
6450 "(found ddr%s fsb freq %d, mem freq %d), "
6451 "disabling CxSR\n",
6452 (dev_priv->is_ddr3 == 1) ? "3" : "2",
6453 dev_priv->fsb_freq, dev_priv->mem_freq);
6454 /* Disable CxSR and never update its watermark again */
6455 pineview_disable_cxsr(dev);
6456 dev_priv->display.update_wm = NULL;
6457 } else
6458 dev_priv->display.update_wm = pineview_update_wm;
6459 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
6460 } else if (IS_G4X(dev)) {
6461 dev_priv->display.update_wm = g4x_update_wm;
6462 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
6463 } else if (IS_GEN4(dev)) {
6464 dev_priv->display.update_wm = i965_update_wm;
6465 if (IS_CRESTLINE(dev))
6466 dev_priv->display.init_clock_gating = crestline_init_clock_gating;
6467 else if (IS_BROADWATER(dev))
6468 dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
6469 } else if (IS_GEN3(dev)) {
6470 dev_priv->display.update_wm = i9xx_update_wm;
6471 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
6472 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
6473 } else if (IS_GEN2(dev)) {
6474 if (INTEL_INFO(dev)->num_pipes == 1) {
6475 dev_priv->display.update_wm = i845_update_wm;
6476 dev_priv->display.get_fifo_size = i845_get_fifo_size;
6477 } else {
6478 dev_priv->display.update_wm = i9xx_update_wm;
6479 dev_priv->display.get_fifo_size = i830_get_fifo_size;
6480 }
6481
6482 if (IS_I85X(dev) || IS_I865G(dev))
6483 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
6484 else
6485 dev_priv->display.init_clock_gating = i830_init_clock_gating;
6486 } else {
6487 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
6488 }
6489 }
6490
6491 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u8 mbox, u32 *val)
6492 {
6493 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6494
6495 if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
6496 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
6497 return -EAGAIN;
6498 }
6499
6500 I915_WRITE(GEN6_PCODE_DATA, *val);
6501 I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
6502
6503 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
6504 500)) {
6505 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
6506 return -ETIMEDOUT;
6507 }
6508
6509 *val = I915_READ(GEN6_PCODE_DATA);
6510 I915_WRITE(GEN6_PCODE_DATA, 0);
6511
6512 return 0;
6513 }
6514
6515 int sandybridge_pcode_write(struct drm_i915_private *dev_priv, u8 mbox, u32 val)
6516 {
6517 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6518
6519 if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
6520 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
6521 return -EAGAIN;
6522 }
6523
6524 I915_WRITE(GEN6_PCODE_DATA, val);
6525 I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
6526
6527 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
6528 500)) {
6529 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
6530 return -ETIMEDOUT;
6531 }
6532
6533 I915_WRITE(GEN6_PCODE_DATA, 0);
6534
6535 return 0;
6536 }
6537
6538 int vlv_gpu_freq(struct drm_i915_private *dev_priv, int val)
6539 {
6540 int div;
6541
6542 /* 4 x czclk */
6543 switch (dev_priv->mem_freq) {
6544 case 800:
6545 div = 10;
6546 break;
6547 case 1066:
6548 div = 12;
6549 break;
6550 case 1333:
6551 div = 16;
6552 break;
6553 default:
6554 return -1;
6555 }
6556
6557 return DIV_ROUND_CLOSEST(dev_priv->mem_freq * (val + 6 - 0xbd), 4 * div);
6558 }
6559
6560 int vlv_freq_opcode(struct drm_i915_private *dev_priv, int val)
6561 {
6562 int mul;
6563
6564 /* 4 x czclk */
6565 switch (dev_priv->mem_freq) {
6566 case 800:
6567 mul = 10;
6568 break;
6569 case 1066:
6570 mul = 12;
6571 break;
6572 case 1333:
6573 mul = 16;
6574 break;
6575 default:
6576 return -1;
6577 }
6578
6579 return DIV_ROUND_CLOSEST(4 * mul * val, dev_priv->mem_freq) + 0xbd - 6;
6580 }
6581
6582 void intel_pm_setup(struct drm_device *dev)
6583 {
6584 struct drm_i915_private *dev_priv = dev->dev_private;
6585
6586 mutex_init(&dev_priv->rps.hw_lock);
6587
6588 INIT_DELAYED_WORK(&dev_priv->rps.delayed_resume_work,
6589 intel_gen6_powersave_work);
6590
6591 dev_priv->pm.suspended = false;
6592 dev_priv->pm.irqs_disabled = false;
6593 }
Linux kernel - Deliverables
This page took 0.178217 seconds and 5 git commands to generate.