projects / deliverable / linux.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/egtvedt...
[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 if (set_no_fbc_reason(dev_priv, FBC_CHIP_DEFAULT))
516 DRM_DEBUG_KMS("disabled per chip default\n");
517 goto out_disable;
518 }
519 if (!i915.enable_fbc) {
520 if (set_no_fbc_reason(dev_priv, FBC_MODULE_PARAM))
521 DRM_DEBUG_KMS("fbc disabled per module param\n");
522 goto out_disable;
523 }
524 if ((adjusted_mode->flags & DRM_MODE_FLAG_INTERLACE) ||
525 (adjusted_mode->flags & DRM_MODE_FLAG_DBLSCAN)) {
526 if (set_no_fbc_reason(dev_priv, FBC_UNSUPPORTED_MODE))
527 DRM_DEBUG_KMS("mode incompatible with compression, "
528 "disabling\n");
529 goto out_disable;
530 }
531
532 if (IS_G4X(dev) || INTEL_INFO(dev)->gen >= 5) {
533 max_width = 4096;
534 max_height = 2048;
535 } else {
536 max_width = 2048;
537 max_height = 1536;
538 }
539 if (intel_crtc->config.pipe_src_w > max_width ||
540 intel_crtc->config.pipe_src_h > max_height) {
541 if (set_no_fbc_reason(dev_priv, FBC_MODE_TOO_LARGE))
542 DRM_DEBUG_KMS("mode too large for compression, disabling\n");
543 goto out_disable;
544 }
545 if ((INTEL_INFO(dev)->gen < 4 || HAS_DDI(dev)) &&
546 intel_crtc->plane != PLANE_A) {
547 if (set_no_fbc_reason(dev_priv, FBC_BAD_PLANE))
548 DRM_DEBUG_KMS("plane not A, disabling compression\n");
549 goto out_disable;
550 }
551
552 /* The use of a CPU fence is mandatory in order to detect writes
553 * by the CPU to the scanout and trigger updates to the FBC.
554 */
555 if (obj->tiling_mode != I915_TILING_X ||
556 obj->fence_reg == I915_FENCE_REG_NONE) {
557 if (set_no_fbc_reason(dev_priv, FBC_NOT_TILED))
558 DRM_DEBUG_KMS("framebuffer not tiled or fenced, disabling compression\n");
559 goto out_disable;
560 }
561
562 /* If the kernel debugger is active, always disable compression */
563 if (in_dbg_master())
564 goto out_disable;
565
566 if (i915_gem_stolen_setup_compression(dev, intel_fb->obj->base.size)) {
567 if (set_no_fbc_reason(dev_priv, FBC_STOLEN_TOO_SMALL))
568 DRM_DEBUG_KMS("framebuffer too large, disabling compression\n");
569 goto out_disable;
570 }
571
572 /* If the scanout has not changed, don't modify the FBC settings.
573 * Note that we make the fundamental assumption that the fb->obj
574 * cannot be unpinned (and have its GTT offset and fence revoked)
575 * without first being decoupled from the scanout and FBC disabled.
576 */
577 if (dev_priv->fbc.plane == intel_crtc->plane &&
578 dev_priv->fbc.fb_id == fb->base.id &&
579 dev_priv->fbc.y == crtc->y)
580 return;
581
582 if (intel_fbc_enabled(dev)) {
583 /* We update FBC along two paths, after changing fb/crtc
584 * configuration (modeswitching) and after page-flipping
585 * finishes. For the latter, we know that not only did
586 * we disable the FBC at the start of the page-flip
587 * sequence, but also more than one vblank has passed.
588 *
589 * For the former case of modeswitching, it is possible
590 * to switch between two FBC valid configurations
591 * instantaneously so we do need to disable the FBC
592 * before we can modify its control registers. We also
593 * have to wait for the next vblank for that to take
594 * effect. However, since we delay enabling FBC we can
595 * assume that a vblank has passed since disabling and
596 * that we can safely alter the registers in the deferred
597 * callback.
598 *
599 * In the scenario that we go from a valid to invalid
600 * and then back to valid FBC configuration we have
601 * no strict enforcement that a vblank occurred since
602 * disabling the FBC. However, along all current pipe
603 * disabling paths we do need to wait for a vblank at
604 * some point. And we wait before enabling FBC anyway.
605 */
606 DRM_DEBUG_KMS("disabling active FBC for update\n");
607 intel_disable_fbc(dev);
608 }
609
610 intel_enable_fbc(crtc);
611 dev_priv->fbc.no_fbc_reason = FBC_OK;
612 return;
613
614 out_disable:
615 /* Multiple disables should be harmless */
616 if (intel_fbc_enabled(dev)) {
617 DRM_DEBUG_KMS("unsupported config, disabling FBC\n");
618 intel_disable_fbc(dev);
619 }
620 i915_gem_stolen_cleanup_compression(dev);
621 }
622
623 static void i915_pineview_get_mem_freq(struct drm_device *dev)
624 {
625 struct drm_i915_private *dev_priv = dev->dev_private;
626 u32 tmp;
627
628 tmp = I915_READ(CLKCFG);
629
630 switch (tmp & CLKCFG_FSB_MASK) {
631 case CLKCFG_FSB_533:
632 dev_priv->fsb_freq = 533; /* 133*4 */
633 break;
634 case CLKCFG_FSB_800:
635 dev_priv->fsb_freq = 800; /* 200*4 */
636 break;
637 case CLKCFG_FSB_667:
638 dev_priv->fsb_freq = 667; /* 167*4 */
639 break;
640 case CLKCFG_FSB_400:
641 dev_priv->fsb_freq = 400; /* 100*4 */
642 break;
643 }
644
645 switch (tmp & CLKCFG_MEM_MASK) {
646 case CLKCFG_MEM_533:
647 dev_priv->mem_freq = 533;
648 break;
649 case CLKCFG_MEM_667:
650 dev_priv->mem_freq = 667;
651 break;
652 case CLKCFG_MEM_800:
653 dev_priv->mem_freq = 800;
654 break;
655 }
656
657 /* detect pineview DDR3 setting */
658 tmp = I915_READ(CSHRDDR3CTL);
659 dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
660 }
661
662 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
663 {
664 struct drm_i915_private *dev_priv = dev->dev_private;
665 u16 ddrpll, csipll;
666
667 ddrpll = I915_READ16(DDRMPLL1);
668 csipll = I915_READ16(CSIPLL0);
669
670 switch (ddrpll & 0xff) {
671 case 0xc:
672 dev_priv->mem_freq = 800;
673 break;
674 case 0x10:
675 dev_priv->mem_freq = 1066;
676 break;
677 case 0x14:
678 dev_priv->mem_freq = 1333;
679 break;
680 case 0x18:
681 dev_priv->mem_freq = 1600;
682 break;
683 default:
684 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
685 ddrpll & 0xff);
686 dev_priv->mem_freq = 0;
687 break;
688 }
689
690 dev_priv->ips.r_t = dev_priv->mem_freq;
691
692 switch (csipll & 0x3ff) {
693 case 0x00c:
694 dev_priv->fsb_freq = 3200;
695 break;
696 case 0x00e:
697 dev_priv->fsb_freq = 3733;
698 break;
699 case 0x010:
700 dev_priv->fsb_freq = 4266;
701 break;
702 case 0x012:
703 dev_priv->fsb_freq = 4800;
704 break;
705 case 0x014:
706 dev_priv->fsb_freq = 5333;
707 break;
708 case 0x016:
709 dev_priv->fsb_freq = 5866;
710 break;
711 case 0x018:
712 dev_priv->fsb_freq = 6400;
713 break;
714 default:
715 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
716 csipll & 0x3ff);
717 dev_priv->fsb_freq = 0;
718 break;
719 }
720
721 if (dev_priv->fsb_freq == 3200) {
722 dev_priv->ips.c_m = 0;
723 } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
724 dev_priv->ips.c_m = 1;
725 } else {
726 dev_priv->ips.c_m = 2;
727 }
728 }
729
730 static const struct cxsr_latency cxsr_latency_table[] = {
731 {1, 0, 800, 400, 3382, 33382, 3983, 33983}, /* DDR2-400 SC */
732 {1, 0, 800, 667, 3354, 33354, 3807, 33807}, /* DDR2-667 SC */
733 {1, 0, 800, 800, 3347, 33347, 3763, 33763}, /* DDR2-800 SC */
734 {1, 1, 800, 667, 6420, 36420, 6873, 36873}, /* DDR3-667 SC */
735 {1, 1, 800, 800, 5902, 35902, 6318, 36318}, /* DDR3-800 SC */
736
737 {1, 0, 667, 400, 3400, 33400, 4021, 34021}, /* DDR2-400 SC */
738 {1, 0, 667, 667, 3372, 33372, 3845, 33845}, /* DDR2-667 SC */
739 {1, 0, 667, 800, 3386, 33386, 3822, 33822}, /* DDR2-800 SC */
740 {1, 1, 667, 667, 6438, 36438, 6911, 36911}, /* DDR3-667 SC */
741 {1, 1, 667, 800, 5941, 35941, 6377, 36377}, /* DDR3-800 SC */
742
743 {1, 0, 400, 400, 3472, 33472, 4173, 34173}, /* DDR2-400 SC */
744 {1, 0, 400, 667, 3443, 33443, 3996, 33996}, /* DDR2-667 SC */
745 {1, 0, 400, 800, 3430, 33430, 3946, 33946}, /* DDR2-800 SC */
746 {1, 1, 400, 667, 6509, 36509, 7062, 37062}, /* DDR3-667 SC */
747 {1, 1, 400, 800, 5985, 35985, 6501, 36501}, /* DDR3-800 SC */
748
749 {0, 0, 800, 400, 3438, 33438, 4065, 34065}, /* DDR2-400 SC */
750 {0, 0, 800, 667, 3410, 33410, 3889, 33889}, /* DDR2-667 SC */
751 {0, 0, 800, 800, 3403, 33403, 3845, 33845}, /* DDR2-800 SC */
752 {0, 1, 800, 667, 6476, 36476, 6955, 36955}, /* DDR3-667 SC */
753 {0, 1, 800, 800, 5958, 35958, 6400, 36400}, /* DDR3-800 SC */
754
755 {0, 0, 667, 400, 3456, 33456, 4103, 34106}, /* DDR2-400 SC */
756 {0, 0, 667, 667, 3428, 33428, 3927, 33927}, /* DDR2-667 SC */
757 {0, 0, 667, 800, 3443, 33443, 3905, 33905}, /* DDR2-800 SC */
758 {0, 1, 667, 667, 6494, 36494, 6993, 36993}, /* DDR3-667 SC */
759 {0, 1, 667, 800, 5998, 35998, 6460, 36460}, /* DDR3-800 SC */
760
761 {0, 0, 400, 400, 3528, 33528, 4255, 34255}, /* DDR2-400 SC */
762 {0, 0, 400, 667, 3500, 33500, 4079, 34079}, /* DDR2-667 SC */
763 {0, 0, 400, 800, 3487, 33487, 4029, 34029}, /* DDR2-800 SC */
764 {0, 1, 400, 667, 6566, 36566, 7145, 37145}, /* DDR3-667 SC */
765 {0, 1, 400, 800, 6042, 36042, 6584, 36584}, /* DDR3-800 SC */
766 };
767
768 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
769 int is_ddr3,
770 int fsb,
771 int mem)
772 {
773 const struct cxsr_latency *latency;
774 int i;
775
776 if (fsb == 0 || mem == 0)
777 return NULL;
778
779 for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
780 latency = &cxsr_latency_table[i];
781 if (is_desktop == latency->is_desktop &&
782 is_ddr3 == latency->is_ddr3 &&
783 fsb == latency->fsb_freq && mem == latency->mem_freq)
784 return latency;
785 }
786
787 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
788
789 return NULL;
790 }
791
792 static void pineview_disable_cxsr(struct drm_device *dev)
793 {
794 struct drm_i915_private *dev_priv = dev->dev_private;
795
796 /* deactivate cxsr */
797 I915_WRITE(DSPFW3, I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN);
798 }
799
800 /*
801 * Latency for FIFO fetches is dependent on several factors:
802 * - memory configuration (speed, channels)
803 * - chipset
804 * - current MCH state
805 * It can be fairly high in some situations, so here we assume a fairly
806 * pessimal value. It's a tradeoff between extra memory fetches (if we
807 * set this value too high, the FIFO will fetch frequently to stay full)
808 * and power consumption (set it too low to save power and we might see
809 * FIFO underruns and display "flicker").
810 *
811 * A value of 5us seems to be a good balance; safe for very low end
812 * platforms but not overly aggressive on lower latency configs.
813 */
814 static const int latency_ns = 5000;
815
816 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
817 {
818 struct drm_i915_private *dev_priv = dev->dev_private;
819 uint32_t dsparb = I915_READ(DSPARB);
820 int size;
821
822 size = dsparb & 0x7f;
823 if (plane)
824 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
825
826 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
827 plane ? "B" : "A", size);
828
829 return size;
830 }
831
832 static int i830_get_fifo_size(struct drm_device *dev, int plane)
833 {
834 struct drm_i915_private *dev_priv = dev->dev_private;
835 uint32_t dsparb = I915_READ(DSPARB);
836 int size;
837
838 size = dsparb & 0x1ff;
839 if (plane)
840 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
841 size >>= 1; /* Convert to cachelines */
842
843 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
844 plane ? "B" : "A", size);
845
846 return size;
847 }
848
849 static int i845_get_fifo_size(struct drm_device *dev, int plane)
850 {
851 struct drm_i915_private *dev_priv = dev->dev_private;
852 uint32_t dsparb = I915_READ(DSPARB);
853 int size;
854
855 size = dsparb & 0x7f;
856 size >>= 2; /* Convert to cachelines */
857
858 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
859 plane ? "B" : "A",
860 size);
861
862 return size;
863 }
864
865 /* Pineview has different values for various configs */
866 static const struct intel_watermark_params pineview_display_wm = {
867 PINEVIEW_DISPLAY_FIFO,
868 PINEVIEW_MAX_WM,
869 PINEVIEW_DFT_WM,
870 PINEVIEW_GUARD_WM,
871 PINEVIEW_FIFO_LINE_SIZE
872 };
873 static const struct intel_watermark_params pineview_display_hplloff_wm = {
874 PINEVIEW_DISPLAY_FIFO,
875 PINEVIEW_MAX_WM,
876 PINEVIEW_DFT_HPLLOFF_WM,
877 PINEVIEW_GUARD_WM,
878 PINEVIEW_FIFO_LINE_SIZE
879 };
880 static const struct intel_watermark_params pineview_cursor_wm = {
881 PINEVIEW_CURSOR_FIFO,
882 PINEVIEW_CURSOR_MAX_WM,
883 PINEVIEW_CURSOR_DFT_WM,
884 PINEVIEW_CURSOR_GUARD_WM,
885 PINEVIEW_FIFO_LINE_SIZE,
886 };
887 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
888 PINEVIEW_CURSOR_FIFO,
889 PINEVIEW_CURSOR_MAX_WM,
890 PINEVIEW_CURSOR_DFT_WM,
891 PINEVIEW_CURSOR_GUARD_WM,
892 PINEVIEW_FIFO_LINE_SIZE
893 };
894 static const struct intel_watermark_params g4x_wm_info = {
895 G4X_FIFO_SIZE,
896 G4X_MAX_WM,
897 G4X_MAX_WM,
898 2,
899 G4X_FIFO_LINE_SIZE,
900 };
901 static const struct intel_watermark_params g4x_cursor_wm_info = {
902 I965_CURSOR_FIFO,
903 I965_CURSOR_MAX_WM,
904 I965_CURSOR_DFT_WM,
905 2,
906 G4X_FIFO_LINE_SIZE,
907 };
908 static const struct intel_watermark_params valleyview_wm_info = {
909 VALLEYVIEW_FIFO_SIZE,
910 VALLEYVIEW_MAX_WM,
911 VALLEYVIEW_MAX_WM,
912 2,
913 G4X_FIFO_LINE_SIZE,
914 };
915 static const struct intel_watermark_params valleyview_cursor_wm_info = {
916 I965_CURSOR_FIFO,
917 VALLEYVIEW_CURSOR_MAX_WM,
918 I965_CURSOR_DFT_WM,
919 2,
920 G4X_FIFO_LINE_SIZE,
921 };
922 static const struct intel_watermark_params i965_cursor_wm_info = {
923 I965_CURSOR_FIFO,
924 I965_CURSOR_MAX_WM,
925 I965_CURSOR_DFT_WM,
926 2,
927 I915_FIFO_LINE_SIZE,
928 };
929 static const struct intel_watermark_params i945_wm_info = {
930 I945_FIFO_SIZE,
931 I915_MAX_WM,
932 1,
933 2,
934 I915_FIFO_LINE_SIZE
935 };
936 static const struct intel_watermark_params i915_wm_info = {
937 I915_FIFO_SIZE,
938 I915_MAX_WM,
939 1,
940 2,
941 I915_FIFO_LINE_SIZE
942 };
943 static const struct intel_watermark_params i830_wm_info = {
944 I855GM_FIFO_SIZE,
945 I915_MAX_WM,
946 1,
947 2,
948 I830_FIFO_LINE_SIZE
949 };
950 static const struct intel_watermark_params i845_wm_info = {
951 I830_FIFO_SIZE,
952 I915_MAX_WM,
953 1,
954 2,
955 I830_FIFO_LINE_SIZE
956 };
957
958 /**
959 * intel_calculate_wm - calculate watermark level
960 * @clock_in_khz: pixel clock
961 * @wm: chip FIFO params
962 * @pixel_size: display pixel size
963 * @latency_ns: memory latency for the platform
964 *
965 * Calculate the watermark level (the level at which the display plane will
966 * start fetching from memory again). Each chip has a different display
967 * FIFO size and allocation, so the caller needs to figure that out and pass
968 * in the correct intel_watermark_params structure.
969 *
970 * As the pixel clock runs, the FIFO will be drained at a rate that depends
971 * on the pixel size. When it reaches the watermark level, it'll start
972 * fetching FIFO line sized based chunks from memory until the FIFO fills
973 * past the watermark point. If the FIFO drains completely, a FIFO underrun
974 * will occur, and a display engine hang could result.
975 */
976 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
977 const struct intel_watermark_params *wm,
978 int fifo_size,
979 int pixel_size,
980 unsigned long latency_ns)
981 {
982 long entries_required, wm_size;
983
984 /*
985 * Note: we need to make sure we don't overflow for various clock &
986 * latency values.
987 * clocks go from a few thousand to several hundred thousand.
988 * latency is usually a few thousand
989 */
990 entries_required = ((clock_in_khz / 1000) * pixel_size * latency_ns) /
991 1000;
992 entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
993
994 DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
995
996 wm_size = fifo_size - (entries_required + wm->guard_size);
997
998 DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
999
1000 /* Don't promote wm_size to unsigned... */
1001 if (wm_size > (long)wm->max_wm)
1002 wm_size = wm->max_wm;
1003 if (wm_size <= 0)
1004 wm_size = wm->default_wm;
1005 return wm_size;
1006 }
1007
1008 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
1009 {
1010 struct drm_crtc *crtc, *enabled = NULL;
1011
1012 for_each_crtc(dev, crtc) {
1013 if (intel_crtc_active(crtc)) {
1014 if (enabled)
1015 return NULL;
1016 enabled = crtc;
1017 }
1018 }
1019
1020 return enabled;
1021 }
1022
1023 static void pineview_update_wm(struct drm_crtc *unused_crtc)
1024 {
1025 struct drm_device *dev = unused_crtc->dev;
1026 struct drm_i915_private *dev_priv = dev->dev_private;
1027 struct drm_crtc *crtc;
1028 const struct cxsr_latency *latency;
1029 u32 reg;
1030 unsigned long wm;
1031
1032 latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
1033 dev_priv->fsb_freq, dev_priv->mem_freq);
1034 if (!latency) {
1035 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
1036 pineview_disable_cxsr(dev);
1037 return;
1038 }
1039
1040 crtc = single_enabled_crtc(dev);
1041 if (crtc) {
1042 const struct drm_display_mode *adjusted_mode;
1043 int pixel_size = crtc->primary->fb->bits_per_pixel / 8;
1044 int clock;
1045
1046 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1047 clock = adjusted_mode->crtc_clock;
1048
1049 /* Display SR */
1050 wm = intel_calculate_wm(clock, &pineview_display_wm,
1051 pineview_display_wm.fifo_size,
1052 pixel_size, latency->display_sr);
1053 reg = I915_READ(DSPFW1);
1054 reg &= ~DSPFW_SR_MASK;
1055 reg |= wm << DSPFW_SR_SHIFT;
1056 I915_WRITE(DSPFW1, reg);
1057 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
1058
1059 /* cursor SR */
1060 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
1061 pineview_display_wm.fifo_size,
1062 pixel_size, latency->cursor_sr);
1063 reg = I915_READ(DSPFW3);
1064 reg &= ~DSPFW_CURSOR_SR_MASK;
1065 reg |= (wm & 0x3f) << DSPFW_CURSOR_SR_SHIFT;
1066 I915_WRITE(DSPFW3, reg);
1067
1068 /* Display HPLL off SR */
1069 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
1070 pineview_display_hplloff_wm.fifo_size,
1071 pixel_size, latency->display_hpll_disable);
1072 reg = I915_READ(DSPFW3);
1073 reg &= ~DSPFW_HPLL_SR_MASK;
1074 reg |= wm & DSPFW_HPLL_SR_MASK;
1075 I915_WRITE(DSPFW3, reg);
1076
1077 /* cursor HPLL off SR */
1078 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
1079 pineview_display_hplloff_wm.fifo_size,
1080 pixel_size, latency->cursor_hpll_disable);
1081 reg = I915_READ(DSPFW3);
1082 reg &= ~DSPFW_HPLL_CURSOR_MASK;
1083 reg |= (wm & 0x3f) << DSPFW_HPLL_CURSOR_SHIFT;
1084 I915_WRITE(DSPFW3, reg);
1085 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
1086
1087 /* activate cxsr */
1088 I915_WRITE(DSPFW3,
1089 I915_READ(DSPFW3) | PINEVIEW_SELF_REFRESH_EN);
1090 DRM_DEBUG_KMS("Self-refresh is enabled\n");
1091 } else {
1092 pineview_disable_cxsr(dev);
1093 DRM_DEBUG_KMS("Self-refresh is disabled\n");
1094 }
1095 }
1096
1097 static bool g4x_compute_wm0(struct drm_device *dev,
1098 int plane,
1099 const struct intel_watermark_params *display,
1100 int display_latency_ns,
1101 const struct intel_watermark_params *cursor,
1102 int cursor_latency_ns,
1103 int *plane_wm,
1104 int *cursor_wm)
1105 {
1106 struct drm_crtc *crtc;
1107 const struct drm_display_mode *adjusted_mode;
1108 int htotal, hdisplay, clock, pixel_size;
1109 int line_time_us, line_count;
1110 int entries, tlb_miss;
1111
1112 crtc = intel_get_crtc_for_plane(dev, plane);
1113 if (!intel_crtc_active(crtc)) {
1114 *cursor_wm = cursor->guard_size;
1115 *plane_wm = display->guard_size;
1116 return false;
1117 }
1118
1119 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1120 clock = adjusted_mode->crtc_clock;
1121 htotal = adjusted_mode->crtc_htotal;
1122 hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1123 pixel_size = crtc->primary->fb->bits_per_pixel / 8;
1124
1125 /* Use the small buffer method to calculate plane watermark */
1126 entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
1127 tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
1128 if (tlb_miss > 0)
1129 entries += tlb_miss;
1130 entries = DIV_ROUND_UP(entries, display->cacheline_size);
1131 *plane_wm = entries + display->guard_size;
1132 if (*plane_wm > (int)display->max_wm)
1133 *plane_wm = display->max_wm;
1134
1135 /* Use the large buffer method to calculate cursor watermark */
1136 line_time_us = max(htotal * 1000 / clock, 1);
1137 line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
1138 entries = line_count * to_intel_crtc(crtc)->cursor_width * pixel_size;
1139 tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
1140 if (tlb_miss > 0)
1141 entries += tlb_miss;
1142 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1143 *cursor_wm = entries + cursor->guard_size;
1144 if (*cursor_wm > (int)cursor->max_wm)
1145 *cursor_wm = (int)cursor->max_wm;
1146
1147 return true;
1148 }
1149
1150 /*
1151 * Check the wm result.
1152 *
1153 * If any calculated watermark values is larger than the maximum value that
1154 * can be programmed into the associated watermark register, that watermark
1155 * must be disabled.
1156 */
1157 static bool g4x_check_srwm(struct drm_device *dev,
1158 int display_wm, int cursor_wm,
1159 const struct intel_watermark_params *display,
1160 const struct intel_watermark_params *cursor)
1161 {
1162 DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
1163 display_wm, cursor_wm);
1164
1165 if (display_wm > display->max_wm) {
1166 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
1167 display_wm, display->max_wm);
1168 return false;
1169 }
1170
1171 if (cursor_wm > cursor->max_wm) {
1172 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
1173 cursor_wm, cursor->max_wm);
1174 return false;
1175 }
1176
1177 if (!(display_wm || cursor_wm)) {
1178 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
1179 return false;
1180 }
1181
1182 return true;
1183 }
1184
1185 static bool g4x_compute_srwm(struct drm_device *dev,
1186 int plane,
1187 int latency_ns,
1188 const struct intel_watermark_params *display,
1189 const struct intel_watermark_params *cursor,
1190 int *display_wm, int *cursor_wm)
1191 {
1192 struct drm_crtc *crtc;
1193 const struct drm_display_mode *adjusted_mode;
1194 int hdisplay, htotal, pixel_size, clock;
1195 unsigned long line_time_us;
1196 int line_count, line_size;
1197 int small, large;
1198 int entries;
1199
1200 if (!latency_ns) {
1201 *display_wm = *cursor_wm = 0;
1202 return false;
1203 }
1204
1205 crtc = intel_get_crtc_for_plane(dev, plane);
1206 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1207 clock = adjusted_mode->crtc_clock;
1208 htotal = adjusted_mode->crtc_htotal;
1209 hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1210 pixel_size = crtc->primary->fb->bits_per_pixel / 8;
1211
1212 line_time_us = max(htotal * 1000 / clock, 1);
1213 line_count = (latency_ns / line_time_us + 1000) / 1000;
1214 line_size = hdisplay * pixel_size;
1215
1216 /* Use the minimum of the small and large buffer method for primary */
1217 small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1218 large = line_count * line_size;
1219
1220 entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1221 *display_wm = entries + display->guard_size;
1222
1223 /* calculate the self-refresh watermark for display cursor */
1224 entries = line_count * pixel_size * to_intel_crtc(crtc)->cursor_width;
1225 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1226 *cursor_wm = entries + cursor->guard_size;
1227
1228 return g4x_check_srwm(dev,
1229 *display_wm, *cursor_wm,
1230 display, cursor);
1231 }
1232
1233 static bool vlv_compute_drain_latency(struct drm_device *dev,
1234 int plane,
1235 int *plane_prec_mult,
1236 int *plane_dl,
1237 int *cursor_prec_mult,
1238 int *cursor_dl)
1239 {
1240 struct drm_crtc *crtc;
1241 int clock, pixel_size;
1242 int entries;
1243
1244 crtc = intel_get_crtc_for_plane(dev, plane);
1245 if (!intel_crtc_active(crtc))
1246 return false;
1247
1248 clock = to_intel_crtc(crtc)->config.adjusted_mode.crtc_clock;
1249 pixel_size = crtc->primary->fb->bits_per_pixel / 8; /* BPP */
1250
1251 entries = (clock / 1000) * pixel_size;
1252 *plane_prec_mult = (entries > 256) ?
1253 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1254 *plane_dl = (64 * (*plane_prec_mult) * 4) / ((clock / 1000) *
1255 pixel_size);
1256
1257 entries = (clock / 1000) * 4; /* BPP is always 4 for cursor */
1258 *cursor_prec_mult = (entries > 256) ?
1259 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1260 *cursor_dl = (64 * (*cursor_prec_mult) * 4) / ((clock / 1000) * 4);
1261
1262 return true;
1263 }
1264
1265 /*
1266 * Update drain latency registers of memory arbiter
1267 *
1268 * Valleyview SoC has a new memory arbiter and needs drain latency registers
1269 * to be programmed. Each plane has a drain latency multiplier and a drain
1270 * latency value.
1271 */
1272
1273 static void vlv_update_drain_latency(struct drm_device *dev)
1274 {
1275 struct drm_i915_private *dev_priv = dev->dev_private;
1276 int planea_prec, planea_dl, planeb_prec, planeb_dl;
1277 int cursora_prec, cursora_dl, cursorb_prec, cursorb_dl;
1278 int plane_prec_mult, cursor_prec_mult; /* Precision multiplier is
1279 either 16 or 32 */
1280
1281 /* For plane A, Cursor A */
1282 if (vlv_compute_drain_latency(dev, 0, &plane_prec_mult, &planea_dl,
1283 &cursor_prec_mult, &cursora_dl)) {
1284 cursora_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1285 DDL_CURSORA_PRECISION_32 : DDL_CURSORA_PRECISION_16;
1286 planea_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1287 DDL_PLANEA_PRECISION_32 : DDL_PLANEA_PRECISION_16;
1288
1289 I915_WRITE(VLV_DDL1, cursora_prec |
1290 (cursora_dl << DDL_CURSORA_SHIFT) |
1291 planea_prec | planea_dl);
1292 }
1293
1294 /* For plane B, Cursor B */
1295 if (vlv_compute_drain_latency(dev, 1, &plane_prec_mult, &planeb_dl,
1296 &cursor_prec_mult, &cursorb_dl)) {
1297 cursorb_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1298 DDL_CURSORB_PRECISION_32 : DDL_CURSORB_PRECISION_16;
1299 planeb_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1300 DDL_PLANEB_PRECISION_32 : DDL_PLANEB_PRECISION_16;
1301
1302 I915_WRITE(VLV_DDL2, cursorb_prec |
1303 (cursorb_dl << DDL_CURSORB_SHIFT) |
1304 planeb_prec | planeb_dl);
1305 }
1306 }
1307
1308 #define single_plane_enabled(mask) is_power_of_2(mask)
1309
1310 static void valleyview_update_wm(struct drm_crtc *crtc)
1311 {
1312 struct drm_device *dev = crtc->dev;
1313 static const int sr_latency_ns = 12000;
1314 struct drm_i915_private *dev_priv = dev->dev_private;
1315 int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1316 int plane_sr, cursor_sr;
1317 int ignore_plane_sr, ignore_cursor_sr;
1318 unsigned int enabled = 0;
1319
1320 vlv_update_drain_latency(dev);
1321
1322 if (g4x_compute_wm0(dev, PIPE_A,
1323 &valleyview_wm_info, latency_ns,
1324 &valleyview_cursor_wm_info, latency_ns,
1325 &planea_wm, &cursora_wm))
1326 enabled |= 1 << PIPE_A;
1327
1328 if (g4x_compute_wm0(dev, PIPE_B,
1329 &valleyview_wm_info, latency_ns,
1330 &valleyview_cursor_wm_info, latency_ns,
1331 &planeb_wm, &cursorb_wm))
1332 enabled |= 1 << PIPE_B;
1333
1334 if (single_plane_enabled(enabled) &&
1335 g4x_compute_srwm(dev, ffs(enabled) - 1,
1336 sr_latency_ns,
1337 &valleyview_wm_info,
1338 &valleyview_cursor_wm_info,
1339 &plane_sr, &ignore_cursor_sr) &&
1340 g4x_compute_srwm(dev, ffs(enabled) - 1,
1341 2*sr_latency_ns,
1342 &valleyview_wm_info,
1343 &valleyview_cursor_wm_info,
1344 &ignore_plane_sr, &cursor_sr)) {
1345 I915_WRITE(FW_BLC_SELF_VLV, FW_CSPWRDWNEN);
1346 } else {
1347 I915_WRITE(FW_BLC_SELF_VLV,
1348 I915_READ(FW_BLC_SELF_VLV) & ~FW_CSPWRDWNEN);
1349 plane_sr = cursor_sr = 0;
1350 }
1351
1352 DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1353 planea_wm, cursora_wm,
1354 planeb_wm, cursorb_wm,
1355 plane_sr, cursor_sr);
1356
1357 I915_WRITE(DSPFW1,
1358 (plane_sr << DSPFW_SR_SHIFT) |
1359 (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1360 (planeb_wm << DSPFW_PLANEB_SHIFT) |
1361 planea_wm);
1362 I915_WRITE(DSPFW2,
1363 (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1364 (cursora_wm << DSPFW_CURSORA_SHIFT));
1365 I915_WRITE(DSPFW3,
1366 (I915_READ(DSPFW3) & ~DSPFW_CURSOR_SR_MASK) |
1367 (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1368 }
1369
1370 static void g4x_update_wm(struct drm_crtc *crtc)
1371 {
1372 struct drm_device *dev = crtc->dev;
1373 static const int sr_latency_ns = 12000;
1374 struct drm_i915_private *dev_priv = dev->dev_private;
1375 int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1376 int plane_sr, cursor_sr;
1377 unsigned int enabled = 0;
1378
1379 if (g4x_compute_wm0(dev, PIPE_A,
1380 &g4x_wm_info, latency_ns,
1381 &g4x_cursor_wm_info, latency_ns,
1382 &planea_wm, &cursora_wm))
1383 enabled |= 1 << PIPE_A;
1384
1385 if (g4x_compute_wm0(dev, PIPE_B,
1386 &g4x_wm_info, latency_ns,
1387 &g4x_cursor_wm_info, latency_ns,
1388 &planeb_wm, &cursorb_wm))
1389 enabled |= 1 << PIPE_B;
1390
1391 if (single_plane_enabled(enabled) &&
1392 g4x_compute_srwm(dev, ffs(enabled) - 1,
1393 sr_latency_ns,
1394 &g4x_wm_info,
1395 &g4x_cursor_wm_info,
1396 &plane_sr, &cursor_sr)) {
1397 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1398 } else {
1399 I915_WRITE(FW_BLC_SELF,
1400 I915_READ(FW_BLC_SELF) & ~FW_BLC_SELF_EN);
1401 plane_sr = cursor_sr = 0;
1402 }
1403
1404 DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1405 planea_wm, cursora_wm,
1406 planeb_wm, cursorb_wm,
1407 plane_sr, cursor_sr);
1408
1409 I915_WRITE(DSPFW1,
1410 (plane_sr << DSPFW_SR_SHIFT) |
1411 (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1412 (planeb_wm << DSPFW_PLANEB_SHIFT) |
1413 planea_wm);
1414 I915_WRITE(DSPFW2,
1415 (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1416 (cursora_wm << DSPFW_CURSORA_SHIFT));
1417 /* HPLL off in SR has some issues on G4x... disable it */
1418 I915_WRITE(DSPFW3,
1419 (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1420 (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1421 }
1422
1423 static void i965_update_wm(struct drm_crtc *unused_crtc)
1424 {
1425 struct drm_device *dev = unused_crtc->dev;
1426 struct drm_i915_private *dev_priv = dev->dev_private;
1427 struct drm_crtc *crtc;
1428 int srwm = 1;
1429 int cursor_sr = 16;
1430
1431 /* Calc sr entries for one plane configs */
1432 crtc = single_enabled_crtc(dev);
1433 if (crtc) {
1434 /* self-refresh has much higher latency */
1435 static const int sr_latency_ns = 12000;
1436 const struct drm_display_mode *adjusted_mode =
1437 &to_intel_crtc(crtc)->config.adjusted_mode;
1438 int clock = adjusted_mode->crtc_clock;
1439 int htotal = adjusted_mode->crtc_htotal;
1440 int hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1441 int pixel_size = crtc->primary->fb->bits_per_pixel / 8;
1442 unsigned long line_time_us;
1443 int entries;
1444
1445 line_time_us = max(htotal * 1000 / clock, 1);
1446
1447 /* Use ns/us then divide to preserve precision */
1448 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1449 pixel_size * hdisplay;
1450 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1451 srwm = I965_FIFO_SIZE - entries;
1452 if (srwm < 0)
1453 srwm = 1;
1454 srwm &= 0x1ff;
1455 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1456 entries, srwm);
1457
1458 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1459 pixel_size * to_intel_crtc(crtc)->cursor_width;
1460 entries = DIV_ROUND_UP(entries,
1461 i965_cursor_wm_info.cacheline_size);
1462 cursor_sr = i965_cursor_wm_info.fifo_size -
1463 (entries + i965_cursor_wm_info.guard_size);
1464
1465 if (cursor_sr > i965_cursor_wm_info.max_wm)
1466 cursor_sr = i965_cursor_wm_info.max_wm;
1467
1468 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1469 "cursor %d\n", srwm, cursor_sr);
1470
1471 if (IS_CRESTLINE(dev))
1472 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1473 } else {
1474 /* Turn off self refresh if both pipes are enabled */
1475 if (IS_CRESTLINE(dev))
1476 I915_WRITE(FW_BLC_SELF, I915_READ(FW_BLC_SELF)
1477 & ~FW_BLC_SELF_EN);
1478 }
1479
1480 DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1481 srwm);
1482
1483 /* 965 has limitations... */
1484 I915_WRITE(DSPFW1, (srwm << DSPFW_SR_SHIFT) |
1485 (8 << 16) | (8 << 8) | (8 << 0));
1486 I915_WRITE(DSPFW2, (8 << 8) | (8 << 0));
1487 /* update cursor SR watermark */
1488 I915_WRITE(DSPFW3, (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1489 }
1490
1491 static void i9xx_update_wm(struct drm_crtc *unused_crtc)
1492 {
1493 struct drm_device *dev = unused_crtc->dev;
1494 struct drm_i915_private *dev_priv = dev->dev_private;
1495 const struct intel_watermark_params *wm_info;
1496 uint32_t fwater_lo;
1497 uint32_t fwater_hi;
1498 int cwm, srwm = 1;
1499 int fifo_size;
1500 int planea_wm, planeb_wm;
1501 struct drm_crtc *crtc, *enabled = NULL;
1502
1503 if (IS_I945GM(dev))
1504 wm_info = &i945_wm_info;
1505 else if (!IS_GEN2(dev))
1506 wm_info = &i915_wm_info;
1507 else
1508 wm_info = &i830_wm_info;
1509
1510 fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1511 crtc = intel_get_crtc_for_plane(dev, 0);
1512 if (intel_crtc_active(crtc)) {
1513 const struct drm_display_mode *adjusted_mode;
1514 int cpp = crtc->primary->fb->bits_per_pixel / 8;
1515 if (IS_GEN2(dev))
1516 cpp = 4;
1517
1518 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1519 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1520 wm_info, fifo_size, cpp,
1521 latency_ns);
1522 enabled = crtc;
1523 } else
1524 planea_wm = fifo_size - wm_info->guard_size;
1525
1526 fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1527 crtc = intel_get_crtc_for_plane(dev, 1);
1528 if (intel_crtc_active(crtc)) {
1529 const struct drm_display_mode *adjusted_mode;
1530 int cpp = crtc->primary->fb->bits_per_pixel / 8;
1531 if (IS_GEN2(dev))
1532 cpp = 4;
1533
1534 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1535 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1536 wm_info, fifo_size, cpp,
1537 latency_ns);
1538 if (enabled == NULL)
1539 enabled = crtc;
1540 else
1541 enabled = NULL;
1542 } else
1543 planeb_wm = fifo_size - wm_info->guard_size;
1544
1545 DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1546
1547 if (IS_I915GM(dev) && enabled) {
1548 struct intel_framebuffer *fb;
1549
1550 fb = to_intel_framebuffer(enabled->primary->fb);
1551
1552 /* self-refresh seems busted with untiled */
1553 if (fb->obj->tiling_mode == I915_TILING_NONE)
1554 enabled = NULL;
1555 }
1556
1557 /*
1558 * Overlay gets an aggressive default since video jitter is bad.
1559 */
1560 cwm = 2;
1561
1562 /* Play safe and disable self-refresh before adjusting watermarks. */
1563 if (IS_I945G(dev) || IS_I945GM(dev))
1564 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN_MASK | 0);
1565 else if (IS_I915GM(dev))
1566 I915_WRITE(INSTPM, _MASKED_BIT_DISABLE(INSTPM_SELF_EN));
1567
1568 /* Calc sr entries for one plane configs */
1569 if (HAS_FW_BLC(dev) && enabled) {
1570 /* self-refresh has much higher latency */
1571 static const int sr_latency_ns = 6000;
1572 const struct drm_display_mode *adjusted_mode =
1573 &to_intel_crtc(enabled)->config.adjusted_mode;
1574 int clock = adjusted_mode->crtc_clock;
1575 int htotal = adjusted_mode->crtc_htotal;
1576 int hdisplay = to_intel_crtc(enabled)->config.pipe_src_w;
1577 int pixel_size = enabled->primary->fb->bits_per_pixel / 8;
1578 unsigned long line_time_us;
1579 int entries;
1580
1581 line_time_us = max(htotal * 1000 / clock, 1);
1582
1583 /* Use ns/us then divide to preserve precision */
1584 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1585 pixel_size * hdisplay;
1586 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1587 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1588 srwm = wm_info->fifo_size - entries;
1589 if (srwm < 0)
1590 srwm = 1;
1591
1592 if (IS_I945G(dev) || IS_I945GM(dev))
1593 I915_WRITE(FW_BLC_SELF,
1594 FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1595 else if (IS_I915GM(dev))
1596 I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1597 }
1598
1599 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1600 planea_wm, planeb_wm, cwm, srwm);
1601
1602 fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1603 fwater_hi = (cwm & 0x1f);
1604
1605 /* Set request length to 8 cachelines per fetch */
1606 fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1607 fwater_hi = fwater_hi | (1 << 8);
1608
1609 I915_WRITE(FW_BLC, fwater_lo);
1610 I915_WRITE(FW_BLC2, fwater_hi);
1611
1612 if (HAS_FW_BLC(dev)) {
1613 if (enabled) {
1614 if (IS_I945G(dev) || IS_I945GM(dev))
1615 I915_WRITE(FW_BLC_SELF,
1616 FW_BLC_SELF_EN_MASK | FW_BLC_SELF_EN);
1617 else if (IS_I915GM(dev))
1618 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_SELF_EN));
1619 DRM_DEBUG_KMS("memory self refresh enabled\n");
1620 } else
1621 DRM_DEBUG_KMS("memory self refresh disabled\n");
1622 }
1623 }
1624
1625 static void i845_update_wm(struct drm_crtc *unused_crtc)
1626 {
1627 struct drm_device *dev = unused_crtc->dev;
1628 struct drm_i915_private *dev_priv = dev->dev_private;
1629 struct drm_crtc *crtc;
1630 const struct drm_display_mode *adjusted_mode;
1631 uint32_t fwater_lo;
1632 int planea_wm;
1633
1634 crtc = single_enabled_crtc(dev);
1635 if (crtc == NULL)
1636 return;
1637
1638 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1639 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1640 &i845_wm_info,
1641 dev_priv->display.get_fifo_size(dev, 0),
1642 4, latency_ns);
1643 fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1644 fwater_lo |= (3<<8) | planea_wm;
1645
1646 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1647
1648 I915_WRITE(FW_BLC, fwater_lo);
1649 }
1650
1651 static uint32_t ilk_pipe_pixel_rate(struct drm_device *dev,
1652 struct drm_crtc *crtc)
1653 {
1654 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1655 uint32_t pixel_rate;
1656
1657 pixel_rate = intel_crtc->config.adjusted_mode.crtc_clock;
1658
1659 /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
1660 * adjust the pixel_rate here. */
1661
1662 if (intel_crtc->config.pch_pfit.enabled) {
1663 uint64_t pipe_w, pipe_h, pfit_w, pfit_h;
1664 uint32_t pfit_size = intel_crtc->config.pch_pfit.size;
1665
1666 pipe_w = intel_crtc->config.pipe_src_w;
1667 pipe_h = intel_crtc->config.pipe_src_h;
1668 pfit_w = (pfit_size >> 16) & 0xFFFF;
1669 pfit_h = pfit_size & 0xFFFF;
1670 if (pipe_w < pfit_w)
1671 pipe_w = pfit_w;
1672 if (pipe_h < pfit_h)
1673 pipe_h = pfit_h;
1674
1675 pixel_rate = div_u64((uint64_t) pixel_rate * pipe_w * pipe_h,
1676 pfit_w * pfit_h);
1677 }
1678
1679 return pixel_rate;
1680 }
1681
1682 /* latency must be in 0.1us units. */
1683 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
1684 uint32_t latency)
1685 {
1686 uint64_t ret;
1687
1688 if (WARN(latency == 0, "Latency value missing\n"))
1689 return UINT_MAX;
1690
1691 ret = (uint64_t) pixel_rate * bytes_per_pixel * latency;
1692 ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
1693
1694 return ret;
1695 }
1696
1697 /* latency must be in 0.1us units. */
1698 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
1699 uint32_t horiz_pixels, uint8_t bytes_per_pixel,
1700 uint32_t latency)
1701 {
1702 uint32_t ret;
1703
1704 if (WARN(latency == 0, "Latency value missing\n"))
1705 return UINT_MAX;
1706
1707 ret = (latency * pixel_rate) / (pipe_htotal * 10000);
1708 ret = (ret + 1) * horiz_pixels * bytes_per_pixel;
1709 ret = DIV_ROUND_UP(ret, 64) + 2;
1710 return ret;
1711 }
1712
1713 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
1714 uint8_t bytes_per_pixel)
1715 {
1716 return DIV_ROUND_UP(pri_val * 64, horiz_pixels * bytes_per_pixel) + 2;
1717 }
1718
1719 struct ilk_pipe_wm_parameters {
1720 bool active;
1721 uint32_t pipe_htotal;
1722 uint32_t pixel_rate;
1723 struct intel_plane_wm_parameters pri;
1724 struct intel_plane_wm_parameters spr;
1725 struct intel_plane_wm_parameters cur;
1726 };
1727
1728 struct ilk_wm_maximums {
1729 uint16_t pri;
1730 uint16_t spr;
1731 uint16_t cur;
1732 uint16_t fbc;
1733 };
1734
1735 /* used in computing the new watermarks state */
1736 struct intel_wm_config {
1737 unsigned int num_pipes_active;
1738 bool sprites_enabled;
1739 bool sprites_scaled;
1740 };
1741
1742 /*
1743 * For both WM_PIPE and WM_LP.
1744 * mem_value must be in 0.1us units.
1745 */
1746 static uint32_t ilk_compute_pri_wm(const struct ilk_pipe_wm_parameters *params,
1747 uint32_t mem_value,
1748 bool is_lp)
1749 {
1750 uint32_t method1, method2;
1751
1752 if (!params->active || !params->pri.enabled)
1753 return 0;
1754
1755 method1 = ilk_wm_method1(params->pixel_rate,
1756 params->pri.bytes_per_pixel,
1757 mem_value);
1758
1759 if (!is_lp)
1760 return method1;
1761
1762 method2 = ilk_wm_method2(params->pixel_rate,
1763 params->pipe_htotal,
1764 params->pri.horiz_pixels,
1765 params->pri.bytes_per_pixel,
1766 mem_value);
1767
1768 return min(method1, method2);
1769 }
1770
1771 /*
1772 * For both WM_PIPE and WM_LP.
1773 * mem_value must be in 0.1us units.
1774 */
1775 static uint32_t ilk_compute_spr_wm(const struct ilk_pipe_wm_parameters *params,
1776 uint32_t mem_value)
1777 {
1778 uint32_t method1, method2;
1779
1780 if (!params->active || !params->spr.enabled)
1781 return 0;
1782
1783 method1 = ilk_wm_method1(params->pixel_rate,
1784 params->spr.bytes_per_pixel,
1785 mem_value);
1786 method2 = ilk_wm_method2(params->pixel_rate,
1787 params->pipe_htotal,
1788 params->spr.horiz_pixels,
1789 params->spr.bytes_per_pixel,
1790 mem_value);
1791 return min(method1, method2);
1792 }
1793
1794 /*
1795 * For both WM_PIPE and WM_LP.
1796 * mem_value must be in 0.1us units.
1797 */
1798 static uint32_t ilk_compute_cur_wm(const struct ilk_pipe_wm_parameters *params,
1799 uint32_t mem_value)
1800 {
1801 if (!params->active || !params->cur.enabled)
1802 return 0;
1803
1804 return ilk_wm_method2(params->pixel_rate,
1805 params->pipe_htotal,
1806 params->cur.horiz_pixels,
1807 params->cur.bytes_per_pixel,
1808 mem_value);
1809 }
1810
1811 /* Only for WM_LP. */
1812 static uint32_t ilk_compute_fbc_wm(const struct ilk_pipe_wm_parameters *params,
1813 uint32_t pri_val)
1814 {
1815 if (!params->active || !params->pri.enabled)
1816 return 0;
1817
1818 return ilk_wm_fbc(pri_val,
1819 params->pri.horiz_pixels,
1820 params->pri.bytes_per_pixel);
1821 }
1822
1823 static unsigned int ilk_display_fifo_size(const struct drm_device *dev)
1824 {
1825 if (INTEL_INFO(dev)->gen >= 8)
1826 return 3072;
1827 else if (INTEL_INFO(dev)->gen >= 7)
1828 return 768;
1829 else
1830 return 512;
1831 }
1832
1833 static unsigned int ilk_plane_wm_reg_max(const struct drm_device *dev,
1834 int level, bool is_sprite)
1835 {
1836 if (INTEL_INFO(dev)->gen >= 8)
1837 /* BDW primary/sprite plane watermarks */
1838 return level == 0 ? 255 : 2047;
1839 else if (INTEL_INFO(dev)->gen >= 7)
1840 /* IVB/HSW primary/sprite plane watermarks */
1841 return level == 0 ? 127 : 1023;
1842 else if (!is_sprite)
1843 /* ILK/SNB primary plane watermarks */
1844 return level == 0 ? 127 : 511;
1845 else
1846 /* ILK/SNB sprite plane watermarks */
1847 return level == 0 ? 63 : 255;
1848 }
1849
1850 static unsigned int ilk_cursor_wm_reg_max(const struct drm_device *dev,
1851 int level)
1852 {
1853 if (INTEL_INFO(dev)->gen >= 7)
1854 return level == 0 ? 63 : 255;
1855 else
1856 return level == 0 ? 31 : 63;
1857 }
1858
1859 static unsigned int ilk_fbc_wm_reg_max(const struct drm_device *dev)
1860 {
1861 if (INTEL_INFO(dev)->gen >= 8)
1862 return 31;
1863 else
1864 return 15;
1865 }
1866
1867 /* Calculate the maximum primary/sprite plane watermark */
1868 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
1869 int level,
1870 const struct intel_wm_config *config,
1871 enum intel_ddb_partitioning ddb_partitioning,
1872 bool is_sprite)
1873 {
1874 unsigned int fifo_size = ilk_display_fifo_size(dev);
1875
1876 /* if sprites aren't enabled, sprites get nothing */
1877 if (is_sprite && !config->sprites_enabled)
1878 return 0;
1879
1880 /* HSW allows LP1+ watermarks even with multiple pipes */
1881 if (level == 0 || config->num_pipes_active > 1) {
1882 fifo_size /= INTEL_INFO(dev)->num_pipes;
1883
1884 /*
1885 * For some reason the non self refresh
1886 * FIFO size is only half of the self
1887 * refresh FIFO size on ILK/SNB.
1888 */
1889 if (INTEL_INFO(dev)->gen <= 6)
1890 fifo_size /= 2;
1891 }
1892
1893 if (config->sprites_enabled) {
1894 /* level 0 is always calculated with 1:1 split */
1895 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
1896 if (is_sprite)
1897 fifo_size *= 5;
1898 fifo_size /= 6;
1899 } else {
1900 fifo_size /= 2;
1901 }
1902 }
1903
1904 /* clamp to max that the registers can hold */
1905 return min(fifo_size, ilk_plane_wm_reg_max(dev, level, is_sprite));
1906 }
1907
1908 /* Calculate the maximum cursor plane watermark */
1909 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
1910 int level,
1911 const struct intel_wm_config *config)
1912 {
1913 /* HSW LP1+ watermarks w/ multiple pipes */
1914 if (level > 0 && config->num_pipes_active > 1)
1915 return 64;
1916
1917 /* otherwise just report max that registers can hold */
1918 return ilk_cursor_wm_reg_max(dev, level);
1919 }
1920
1921 static void ilk_compute_wm_maximums(const struct drm_device *dev,
1922 int level,
1923 const struct intel_wm_config *config,
1924 enum intel_ddb_partitioning ddb_partitioning,
1925 struct ilk_wm_maximums *max)
1926 {
1927 max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
1928 max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
1929 max->cur = ilk_cursor_wm_max(dev, level, config);
1930 max->fbc = ilk_fbc_wm_reg_max(dev);
1931 }
1932
1933 static void ilk_compute_wm_reg_maximums(struct drm_device *dev,
1934 int level,
1935 struct ilk_wm_maximums *max)
1936 {
1937 max->pri = ilk_plane_wm_reg_max(dev, level, false);
1938 max->spr = ilk_plane_wm_reg_max(dev, level, true);
1939 max->cur = ilk_cursor_wm_reg_max(dev, level);
1940 max->fbc = ilk_fbc_wm_reg_max(dev);
1941 }
1942
1943 static bool ilk_validate_wm_level(int level,
1944 const struct ilk_wm_maximums *max,
1945 struct intel_wm_level *result)
1946 {
1947 bool ret;
1948
1949 /* already determined to be invalid? */
1950 if (!result->enable)
1951 return false;
1952
1953 result->enable = result->pri_val <= max->pri &&
1954 result->spr_val <= max->spr &&
1955 result->cur_val <= max->cur;
1956
1957 ret = result->enable;
1958
1959 /*
1960 * HACK until we can pre-compute everything,
1961 * and thus fail gracefully if LP0 watermarks
1962 * are exceeded...
1963 */
1964 if (level == 0 && !result->enable) {
1965 if (result->pri_val > max->pri)
1966 DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
1967 level, result->pri_val, max->pri);
1968 if (result->spr_val > max->spr)
1969 DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
1970 level, result->spr_val, max->spr);
1971 if (result->cur_val > max->cur)
1972 DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
1973 level, result->cur_val, max->cur);
1974
1975 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
1976 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
1977 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
1978 result->enable = true;
1979 }
1980
1981 return ret;
1982 }
1983
1984 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
1985 int level,
1986 const struct ilk_pipe_wm_parameters *p,
1987 struct intel_wm_level *result)
1988 {
1989 uint16_t pri_latency = dev_priv->wm.pri_latency[level];
1990 uint16_t spr_latency = dev_priv->wm.spr_latency[level];
1991 uint16_t cur_latency = dev_priv->wm.cur_latency[level];
1992
1993 /* WM1+ latency values stored in 0.5us units */
1994 if (level > 0) {
1995 pri_latency *= 5;
1996 spr_latency *= 5;
1997 cur_latency *= 5;
1998 }
1999
2000 result->pri_val = ilk_compute_pri_wm(p, pri_latency, level);
2001 result->spr_val = ilk_compute_spr_wm(p, spr_latency);
2002 result->cur_val = ilk_compute_cur_wm(p, cur_latency);
2003 result->fbc_val = ilk_compute_fbc_wm(p, result->pri_val);
2004 result->enable = true;
2005 }
2006
2007 static uint32_t
2008 hsw_compute_linetime_wm(struct drm_device *dev, struct drm_crtc *crtc)
2009 {
2010 struct drm_i915_private *dev_priv = dev->dev_private;
2011 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2012 struct drm_display_mode *mode = &intel_crtc->config.adjusted_mode;
2013 u32 linetime, ips_linetime;
2014
2015 if (!intel_crtc_active(crtc))
2016 return 0;
2017
2018 /* The WM are computed with base on how long it takes to fill a single
2019 * row at the given clock rate, multiplied by 8.
2020 * */
2021 linetime = DIV_ROUND_CLOSEST(mode->crtc_htotal * 1000 * 8,
2022 mode->crtc_clock);
2023 ips_linetime = DIV_ROUND_CLOSEST(mode->crtc_htotal * 1000 * 8,
2024 intel_ddi_get_cdclk_freq(dev_priv));
2025
2026 return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2027 PIPE_WM_LINETIME_TIME(linetime);
2028 }
2029
2030 static void intel_read_wm_latency(struct drm_device *dev, uint16_t wm[5])
2031 {
2032 struct drm_i915_private *dev_priv = dev->dev_private;
2033
2034 if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2035 uint64_t sskpd = I915_READ64(MCH_SSKPD);
2036
2037 wm[0] = (sskpd >> 56) & 0xFF;
2038 if (wm[0] == 0)
2039 wm[0] = sskpd & 0xF;
2040 wm[1] = (sskpd >> 4) & 0xFF;
2041 wm[2] = (sskpd >> 12) & 0xFF;
2042 wm[3] = (sskpd >> 20) & 0x1FF;
2043 wm[4] = (sskpd >> 32) & 0x1FF;
2044 } else if (INTEL_INFO(dev)->gen >= 6) {
2045 uint32_t sskpd = I915_READ(MCH_SSKPD);
2046
2047 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2048 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2049 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2050 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2051 } else if (INTEL_INFO(dev)->gen >= 5) {
2052 uint32_t mltr = I915_READ(MLTR_ILK);
2053
2054 /* ILK primary LP0 latency is 700 ns */
2055 wm[0] = 7;
2056 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2057 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2058 }
2059 }
2060
2061 static void intel_fixup_spr_wm_latency(struct drm_device *dev, uint16_t wm[5])
2062 {
2063 /* ILK sprite LP0 latency is 1300 ns */
2064 if (INTEL_INFO(dev)->gen == 5)
2065 wm[0] = 13;
2066 }
2067
2068 static void intel_fixup_cur_wm_latency(struct drm_device *dev, uint16_t wm[5])
2069 {
2070 /* ILK cursor LP0 latency is 1300 ns */
2071 if (INTEL_INFO(dev)->gen == 5)
2072 wm[0] = 13;
2073
2074 /* WaDoubleCursorLP3Latency:ivb */
2075 if (IS_IVYBRIDGE(dev))
2076 wm[3] *= 2;
2077 }
2078
2079 int ilk_wm_max_level(const struct drm_device *dev)
2080 {
2081 /* how many WM levels are we expecting */
2082 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2083 return 4;
2084 else if (INTEL_INFO(dev)->gen >= 6)
2085 return 3;
2086 else
2087 return 2;
2088 }
2089
2090 static void intel_print_wm_latency(struct drm_device *dev,
2091 const char *name,
2092 const uint16_t wm[5])
2093 {
2094 int level, max_level = ilk_wm_max_level(dev);
2095
2096 for (level = 0; level <= max_level; level++) {
2097 unsigned int latency = wm[level];
2098
2099 if (latency == 0) {
2100 DRM_ERROR("%s WM%d latency not provided\n",
2101 name, level);
2102 continue;
2103 }
2104
2105 /* WM1+ latency values in 0.5us units */
2106 if (level > 0)
2107 latency *= 5;
2108
2109 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2110 name, level, wm[level],
2111 latency / 10, latency % 10);
2112 }
2113 }
2114
2115 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
2116 uint16_t wm[5], uint16_t min)
2117 {
2118 int level, max_level = ilk_wm_max_level(dev_priv->dev);
2119
2120 if (wm[0] >= min)
2121 return false;
2122
2123 wm[0] = max(wm[0], min);
2124 for (level = 1; level <= max_level; level++)
2125 wm[level] = max_t(uint16_t, wm[level], DIV_ROUND_UP(min, 5));
2126
2127 return true;
2128 }
2129
2130 static void snb_wm_latency_quirk(struct drm_device *dev)
2131 {
2132 struct drm_i915_private *dev_priv = dev->dev_private;
2133 bool changed;
2134
2135 /*
2136 * The BIOS provided WM memory latency values are often
2137 * inadequate for high resolution displays. Adjust them.
2138 */
2139 changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
2140 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
2141 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
2142
2143 if (!changed)
2144 return;
2145
2146 DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2147 intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2148 intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2149 intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2150 }
2151
2152 static void ilk_setup_wm_latency(struct drm_device *dev)
2153 {
2154 struct drm_i915_private *dev_priv = dev->dev_private;
2155
2156 intel_read_wm_latency(dev, dev_priv->wm.pri_latency);
2157
2158 memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2159 sizeof(dev_priv->wm.pri_latency));
2160 memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2161 sizeof(dev_priv->wm.pri_latency));
2162
2163 intel_fixup_spr_wm_latency(dev, dev_priv->wm.spr_latency);
2164 intel_fixup_cur_wm_latency(dev, dev_priv->wm.cur_latency);
2165
2166 intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2167 intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2168 intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2169
2170 if (IS_GEN6(dev))
2171 snb_wm_latency_quirk(dev);
2172 }
2173
2174 static void ilk_compute_wm_parameters(struct drm_crtc *crtc,
2175 struct ilk_pipe_wm_parameters *p)
2176 {
2177 struct drm_device *dev = crtc->dev;
2178 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2179 enum pipe pipe = intel_crtc->pipe;
2180 struct drm_plane *plane;
2181
2182 if (!intel_crtc_active(crtc))
2183 return;
2184
2185 p->active = true;
2186 p->pipe_htotal = intel_crtc->config.adjusted_mode.crtc_htotal;
2187 p->pixel_rate = ilk_pipe_pixel_rate(dev, crtc);
2188 p->pri.bytes_per_pixel = crtc->primary->fb->bits_per_pixel / 8;
2189 p->cur.bytes_per_pixel = 4;
2190 p->pri.horiz_pixels = intel_crtc->config.pipe_src_w;
2191 p->cur.horiz_pixels = intel_crtc->cursor_width;
2192 /* TODO: for now, assume primary and cursor planes are always enabled. */
2193 p->pri.enabled = true;
2194 p->cur.enabled = true;
2195
2196 drm_for_each_legacy_plane(plane, &dev->mode_config.plane_list) {
2197 struct intel_plane *intel_plane = to_intel_plane(plane);
2198
2199 if (intel_plane->pipe == pipe) {
2200 p->spr = intel_plane->wm;
2201 break;
2202 }
2203 }
2204 }
2205
2206 static void ilk_compute_wm_config(struct drm_device *dev,
2207 struct intel_wm_config *config)
2208 {
2209 struct intel_crtc *intel_crtc;
2210
2211 /* Compute the currently _active_ config */
2212 for_each_intel_crtc(dev, intel_crtc) {
2213 const struct intel_pipe_wm *wm = &intel_crtc->wm.active;
2214
2215 if (!wm->pipe_enabled)
2216 continue;
2217
2218 config->sprites_enabled |= wm->sprites_enabled;
2219 config->sprites_scaled |= wm->sprites_scaled;
2220 config->num_pipes_active++;
2221 }
2222 }
2223
2224 /* Compute new watermarks for the pipe */
2225 static bool intel_compute_pipe_wm(struct drm_crtc *crtc,
2226 const struct ilk_pipe_wm_parameters *params,
2227 struct intel_pipe_wm *pipe_wm)
2228 {
2229 struct drm_device *dev = crtc->dev;
2230 const struct drm_i915_private *dev_priv = dev->dev_private;
2231 int level, max_level = ilk_wm_max_level(dev);
2232 /* LP0 watermark maximums depend on this pipe alone */
2233 struct intel_wm_config config = {
2234 .num_pipes_active = 1,
2235 .sprites_enabled = params->spr.enabled,
2236 .sprites_scaled = params->spr.scaled,
2237 };
2238 struct ilk_wm_maximums max;
2239
2240 pipe_wm->pipe_enabled = params->active;
2241 pipe_wm->sprites_enabled = params->spr.enabled;
2242 pipe_wm->sprites_scaled = params->spr.scaled;
2243
2244 /* ILK/SNB: LP2+ watermarks only w/o sprites */
2245 if (INTEL_INFO(dev)->gen <= 6 && params->spr.enabled)
2246 max_level = 1;
2247
2248 /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2249 if (params->spr.scaled)
2250 max_level = 0;
2251
2252 ilk_compute_wm_level(dev_priv, 0, params, &pipe_wm->wm[0]);
2253
2254 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2255 pipe_wm->linetime = hsw_compute_linetime_wm(dev, crtc);
2256
2257 /* LP0 watermarks always use 1/2 DDB partitioning */
2258 ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
2259
2260 /* At least LP0 must be valid */
2261 if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0]))
2262 return false;
2263
2264 ilk_compute_wm_reg_maximums(dev, 1, &max);
2265
2266 for (level = 1; level <= max_level; level++) {
2267 struct intel_wm_level wm = {};
2268
2269 ilk_compute_wm_level(dev_priv, level, params, &wm);
2270
2271 /*
2272 * Disable any watermark level that exceeds the
2273 * register maximums since such watermarks are
2274 * always invalid.
2275 */
2276 if (!ilk_validate_wm_level(level, &max, &wm))
2277 break;
2278
2279 pipe_wm->wm[level] = wm;
2280 }
2281
2282 return true;
2283 }
2284
2285 /*
2286 * Merge the watermarks from all active pipes for a specific level.
2287 */
2288 static void ilk_merge_wm_level(struct drm_device *dev,
2289 int level,
2290 struct intel_wm_level *ret_wm)
2291 {
2292 const struct intel_crtc *intel_crtc;
2293
2294 ret_wm->enable = true;
2295
2296 for_each_intel_crtc(dev, intel_crtc) {
2297 const struct intel_pipe_wm *active = &intel_crtc->wm.active;
2298 const struct intel_wm_level *wm = &active->wm[level];
2299
2300 if (!active->pipe_enabled)
2301 continue;
2302
2303 /*
2304 * The watermark values may have been used in the past,
2305 * so we must maintain them in the registers for some
2306 * time even if the level is now disabled.
2307 */
2308 if (!wm->enable)
2309 ret_wm->enable = false;
2310
2311 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
2312 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
2313 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
2314 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
2315 }
2316 }
2317
2318 /*
2319 * Merge all low power watermarks for all active pipes.
2320 */
2321 static void ilk_wm_merge(struct drm_device *dev,
2322 const struct intel_wm_config *config,
2323 const struct ilk_wm_maximums *max,
2324 struct intel_pipe_wm *merged)
2325 {
2326 int level, max_level = ilk_wm_max_level(dev);
2327 int last_enabled_level = max_level;
2328
2329 /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2330 if ((INTEL_INFO(dev)->gen <= 6 || IS_IVYBRIDGE(dev)) &&
2331 config->num_pipes_active > 1)
2332 return;
2333
2334 /* ILK: FBC WM must be disabled always */
2335 merged->fbc_wm_enabled = INTEL_INFO(dev)->gen >= 6;
2336
2337 /* merge each WM1+ level */
2338 for (level = 1; level <= max_level; level++) {
2339 struct intel_wm_level *wm = &merged->wm[level];
2340
2341 ilk_merge_wm_level(dev, level, wm);
2342
2343 if (level > last_enabled_level)
2344 wm->enable = false;
2345 else if (!ilk_validate_wm_level(level, max, wm))
2346 /* make sure all following levels get disabled */
2347 last_enabled_level = level - 1;
2348
2349 /*
2350 * The spec says it is preferred to disable
2351 * FBC WMs instead of disabling a WM level.
2352 */
2353 if (wm->fbc_val > max->fbc) {
2354 if (wm->enable)
2355 merged->fbc_wm_enabled = false;
2356 wm->fbc_val = 0;
2357 }
2358 }
2359
2360 /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2361 /*
2362 * FIXME this is racy. FBC might get enabled later.
2363 * What we should check here is whether FBC can be
2364 * enabled sometime later.
2365 */
2366 if (IS_GEN5(dev) && !merged->fbc_wm_enabled && intel_fbc_enabled(dev)) {
2367 for (level = 2; level <= max_level; level++) {
2368 struct intel_wm_level *wm = &merged->wm[level];
2369
2370 wm->enable = false;
2371 }
2372 }
2373 }
2374
2375 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
2376 {
2377 /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2378 return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
2379 }
2380
2381 /* The value we need to program into the WM_LPx latency field */
2382 static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
2383 {
2384 struct drm_i915_private *dev_priv = dev->dev_private;
2385
2386 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2387 return 2 * level;
2388 else
2389 return dev_priv->wm.pri_latency[level];
2390 }
2391
2392 static void ilk_compute_wm_results(struct drm_device *dev,
2393 const struct intel_pipe_wm *merged,
2394 enum intel_ddb_partitioning partitioning,
2395 struct ilk_wm_values *results)
2396 {
2397 struct intel_crtc *intel_crtc;
2398 int level, wm_lp;
2399
2400 results->enable_fbc_wm = merged->fbc_wm_enabled;
2401 results->partitioning = partitioning;
2402
2403 /* LP1+ register values */
2404 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2405 const struct intel_wm_level *r;
2406
2407 level = ilk_wm_lp_to_level(wm_lp, merged);
2408
2409 r = &merged->wm[level];
2410
2411 /*
2412 * Maintain the watermark values even if the level is
2413 * disabled. Doing otherwise could cause underruns.
2414 */
2415 results->wm_lp[wm_lp - 1] =
2416 (ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
2417 (r->pri_val << WM1_LP_SR_SHIFT) |
2418 r->cur_val;
2419
2420 if (r->enable)
2421 results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
2422
2423 if (INTEL_INFO(dev)->gen >= 8)
2424 results->wm_lp[wm_lp - 1] |=
2425 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
2426 else
2427 results->wm_lp[wm_lp - 1] |=
2428 r->fbc_val << WM1_LP_FBC_SHIFT;
2429
2430 /*
2431 * Always set WM1S_LP_EN when spr_val != 0, even if the
2432 * level is disabled. Doing otherwise could cause underruns.
2433 */
2434 if (INTEL_INFO(dev)->gen <= 6 && r->spr_val) {
2435 WARN_ON(wm_lp != 1);
2436 results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
2437 } else
2438 results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2439 }
2440
2441 /* LP0 register values */
2442 for_each_intel_crtc(dev, intel_crtc) {
2443 enum pipe pipe = intel_crtc->pipe;
2444 const struct intel_wm_level *r =
2445 &intel_crtc->wm.active.wm[0];
2446
2447 if (WARN_ON(!r->enable))
2448 continue;
2449
2450 results->wm_linetime[pipe] = intel_crtc->wm.active.linetime;
2451
2452 results->wm_pipe[pipe] =
2453 (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
2454 (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
2455 r->cur_val;
2456 }
2457 }
2458
2459 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2460 * case both are at the same level. Prefer r1 in case they're the same. */
2461 static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
2462 struct intel_pipe_wm *r1,
2463 struct intel_pipe_wm *r2)
2464 {
2465 int level, max_level = ilk_wm_max_level(dev);
2466 int level1 = 0, level2 = 0;
2467
2468 for (level = 1; level <= max_level; level++) {
2469 if (r1->wm[level].enable)
2470 level1 = level;
2471 if (r2->wm[level].enable)
2472 level2 = level;
2473 }
2474
2475 if (level1 == level2) {
2476 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
2477 return r2;
2478 else
2479 return r1;
2480 } else if (level1 > level2) {
2481 return r1;
2482 } else {
2483 return r2;
2484 }
2485 }
2486
2487 /* dirty bits used to track which watermarks need changes */
2488 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2489 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2490 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2491 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2492 #define WM_DIRTY_FBC (1 << 24)
2493 #define WM_DIRTY_DDB (1 << 25)
2494
2495 static unsigned int ilk_compute_wm_dirty(struct drm_device *dev,
2496 const struct ilk_wm_values *old,
2497 const struct ilk_wm_values *new)
2498 {
2499 unsigned int dirty = 0;
2500 enum pipe pipe;
2501 int wm_lp;
2502
2503 for_each_pipe(pipe) {
2504 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
2505 dirty |= WM_DIRTY_LINETIME(pipe);
2506 /* Must disable LP1+ watermarks too */
2507 dirty |= WM_DIRTY_LP_ALL;
2508 }
2509
2510 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
2511 dirty |= WM_DIRTY_PIPE(pipe);
2512 /* Must disable LP1+ watermarks too */
2513 dirty |= WM_DIRTY_LP_ALL;
2514 }
2515 }
2516
2517 if (old->enable_fbc_wm != new->enable_fbc_wm) {
2518 dirty |= WM_DIRTY_FBC;
2519 /* Must disable LP1+ watermarks too */
2520 dirty |= WM_DIRTY_LP_ALL;
2521 }
2522
2523 if (old->partitioning != new->partitioning) {
2524 dirty |= WM_DIRTY_DDB;
2525 /* Must disable LP1+ watermarks too */
2526 dirty |= WM_DIRTY_LP_ALL;
2527 }
2528
2529 /* LP1+ watermarks already deemed dirty, no need to continue */
2530 if (dirty & WM_DIRTY_LP_ALL)
2531 return dirty;
2532
2533 /* Find the lowest numbered LP1+ watermark in need of an update... */
2534 for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2535 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
2536 old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
2537 break;
2538 }
2539
2540 /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2541 for (; wm_lp <= 3; wm_lp++)
2542 dirty |= WM_DIRTY_LP(wm_lp);
2543
2544 return dirty;
2545 }
2546
2547 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
2548 unsigned int dirty)
2549 {
2550 struct ilk_wm_values *previous = &dev_priv->wm.hw;
2551 bool changed = false;
2552
2553 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
2554 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
2555 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
2556 changed = true;
2557 }
2558 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
2559 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
2560 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
2561 changed = true;
2562 }
2563 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
2564 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
2565 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
2566 changed = true;
2567 }
2568
2569 /*
2570 * Don't touch WM1S_LP_EN here.
2571 * Doing so could cause underruns.
2572 */
2573
2574 return changed;
2575 }
2576
2577 /*
2578 * The spec says we shouldn't write when we don't need, because every write
2579 * causes WMs to be re-evaluated, expending some power.
2580 */
2581 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
2582 struct ilk_wm_values *results)
2583 {
2584 struct drm_device *dev = dev_priv->dev;
2585 struct ilk_wm_values *previous = &dev_priv->wm.hw;
2586 unsigned int dirty;
2587 uint32_t val;
2588
2589 dirty = ilk_compute_wm_dirty(dev, previous, results);
2590 if (!dirty)
2591 return;
2592
2593 _ilk_disable_lp_wm(dev_priv, dirty);
2594
2595 if (dirty & WM_DIRTY_PIPE(PIPE_A))
2596 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2597 if (dirty & WM_DIRTY_PIPE(PIPE_B))
2598 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
2599 if (dirty & WM_DIRTY_PIPE(PIPE_C))
2600 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
2601
2602 if (dirty & WM_DIRTY_LINETIME(PIPE_A))
2603 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
2604 if (dirty & WM_DIRTY_LINETIME(PIPE_B))
2605 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
2606 if (dirty & WM_DIRTY_LINETIME(PIPE_C))
2607 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
2608
2609 if (dirty & WM_DIRTY_DDB) {
2610 if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2611 val = I915_READ(WM_MISC);
2612 if (results->partitioning == INTEL_DDB_PART_1_2)
2613 val &= ~WM_MISC_DATA_PARTITION_5_6;
2614 else
2615 val |= WM_MISC_DATA_PARTITION_5_6;
2616 I915_WRITE(WM_MISC, val);
2617 } else {
2618 val = I915_READ(DISP_ARB_CTL2);
2619 if (results->partitioning == INTEL_DDB_PART_1_2)
2620 val &= ~DISP_DATA_PARTITION_5_6;
2621 else
2622 val |= DISP_DATA_PARTITION_5_6;
2623 I915_WRITE(DISP_ARB_CTL2, val);
2624 }
2625 }
2626
2627 if (dirty & WM_DIRTY_FBC) {
2628 val = I915_READ(DISP_ARB_CTL);
2629 if (results->enable_fbc_wm)
2630 val &= ~DISP_FBC_WM_DIS;
2631 else
2632 val |= DISP_FBC_WM_DIS;
2633 I915_WRITE(DISP_ARB_CTL, val);
2634 }
2635
2636 if (dirty & WM_DIRTY_LP(1) &&
2637 previous->wm_lp_spr[0] != results->wm_lp_spr[0])
2638 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
2639
2640 if (INTEL_INFO(dev)->gen >= 7) {
2641 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
2642 I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
2643 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
2644 I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
2645 }
2646
2647 if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
2648 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
2649 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
2650 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
2651 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
2652 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
2653
2654 dev_priv->wm.hw = *results;
2655 }
2656
2657 static bool ilk_disable_lp_wm(struct drm_device *dev)
2658 {
2659 struct drm_i915_private *dev_priv = dev->dev_private;
2660
2661 return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
2662 }
2663
2664 static void ilk_update_wm(struct drm_crtc *crtc)
2665 {
2666 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2667 struct drm_device *dev = crtc->dev;
2668 struct drm_i915_private *dev_priv = dev->dev_private;
2669 struct ilk_wm_maximums max;
2670 struct ilk_pipe_wm_parameters params = {};
2671 struct ilk_wm_values results = {};
2672 enum intel_ddb_partitioning partitioning;
2673 struct intel_pipe_wm pipe_wm = {};
2674 struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
2675 struct intel_wm_config config = {};
2676
2677 ilk_compute_wm_parameters(crtc, &params);
2678
2679 intel_compute_pipe_wm(crtc, &params, &pipe_wm);
2680
2681 if (!memcmp(&intel_crtc->wm.active, &pipe_wm, sizeof(pipe_wm)))
2682 return;
2683
2684 intel_crtc->wm.active = pipe_wm;
2685
2686 ilk_compute_wm_config(dev, &config);
2687
2688 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
2689 ilk_wm_merge(dev, &config, &max, &lp_wm_1_2);
2690
2691 /* 5/6 split only in single pipe config on IVB+ */
2692 if (INTEL_INFO(dev)->gen >= 7 &&
2693 config.num_pipes_active == 1 && config.sprites_enabled) {
2694 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
2695 ilk_wm_merge(dev, &config, &max, &lp_wm_5_6);
2696
2697 best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
2698 } else {
2699 best_lp_wm = &lp_wm_1_2;
2700 }
2701
2702 partitioning = (best_lp_wm == &lp_wm_1_2) ?
2703 INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
2704
2705 ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
2706
2707 ilk_write_wm_values(dev_priv, &results);
2708 }
2709
2710 static void ilk_update_sprite_wm(struct drm_plane *plane,
2711 struct drm_crtc *crtc,
2712 uint32_t sprite_width, int pixel_size,
2713 bool enabled, bool scaled)
2714 {
2715 struct drm_device *dev = plane->dev;
2716 struct intel_plane *intel_plane = to_intel_plane(plane);
2717
2718 intel_plane->wm.enabled = enabled;
2719 intel_plane->wm.scaled = scaled;
2720 intel_plane->wm.horiz_pixels = sprite_width;
2721 intel_plane->wm.bytes_per_pixel = pixel_size;
2722
2723 /*
2724 * IVB workaround: must disable low power watermarks for at least
2725 * one frame before enabling scaling. LP watermarks can be re-enabled
2726 * when scaling is disabled.
2727 *
2728 * WaCxSRDisabledForSpriteScaling:ivb
2729 */
2730 if (IS_IVYBRIDGE(dev) && scaled && ilk_disable_lp_wm(dev))
2731 intel_wait_for_vblank(dev, intel_plane->pipe);
2732
2733 ilk_update_wm(crtc);
2734 }
2735
2736 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
2737 {
2738 struct drm_device *dev = crtc->dev;
2739 struct drm_i915_private *dev_priv = dev->dev_private;
2740 struct ilk_wm_values *hw = &dev_priv->wm.hw;
2741 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2742 struct intel_pipe_wm *active = &intel_crtc->wm.active;
2743 enum pipe pipe = intel_crtc->pipe;
2744 static const unsigned int wm0_pipe_reg[] = {
2745 [PIPE_A] = WM0_PIPEA_ILK,
2746 [PIPE_B] = WM0_PIPEB_ILK,
2747 [PIPE_C] = WM0_PIPEC_IVB,
2748 };
2749
2750 hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
2751 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2752 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
2753
2754 active->pipe_enabled = intel_crtc_active(crtc);
2755
2756 if (active->pipe_enabled) {
2757 u32 tmp = hw->wm_pipe[pipe];
2758
2759 /*
2760 * For active pipes LP0 watermark is marked as
2761 * enabled, and LP1+ watermaks as disabled since
2762 * we can't really reverse compute them in case
2763 * multiple pipes are active.
2764 */
2765 active->wm[0].enable = true;
2766 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
2767 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
2768 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
2769 active->linetime = hw->wm_linetime[pipe];
2770 } else {
2771 int level, max_level = ilk_wm_max_level(dev);
2772
2773 /*
2774 * For inactive pipes, all watermark levels
2775 * should be marked as enabled but zeroed,
2776 * which is what we'd compute them to.
2777 */
2778 for (level = 0; level <= max_level; level++)
2779 active->wm[level].enable = true;
2780 }
2781 }
2782
2783 void ilk_wm_get_hw_state(struct drm_device *dev)
2784 {
2785 struct drm_i915_private *dev_priv = dev->dev_private;
2786 struct ilk_wm_values *hw = &dev_priv->wm.hw;
2787 struct drm_crtc *crtc;
2788
2789 for_each_crtc(dev, crtc)
2790 ilk_pipe_wm_get_hw_state(crtc);
2791
2792 hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
2793 hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
2794 hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
2795
2796 hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
2797 if (INTEL_INFO(dev)->gen >= 7) {
2798 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
2799 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
2800 }
2801
2802 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2803 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
2804 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
2805 else if (IS_IVYBRIDGE(dev))
2806 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
2807 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
2808
2809 hw->enable_fbc_wm =
2810 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
2811 }
2812
2813 /**
2814 * intel_update_watermarks - update FIFO watermark values based on current modes
2815 *
2816 * Calculate watermark values for the various WM regs based on current mode
2817 * and plane configuration.
2818 *
2819 * There are several cases to deal with here:
2820 * - normal (i.e. non-self-refresh)
2821 * - self-refresh (SR) mode
2822 * - lines are large relative to FIFO size (buffer can hold up to 2)
2823 * - lines are small relative to FIFO size (buffer can hold more than 2
2824 * lines), so need to account for TLB latency
2825 *
2826 * The normal calculation is:
2827 * watermark = dotclock * bytes per pixel * latency
2828 * where latency is platform & configuration dependent (we assume pessimal
2829 * values here).
2830 *
2831 * The SR calculation is:
2832 * watermark = (trunc(latency/line time)+1) * surface width *
2833 * bytes per pixel
2834 * where
2835 * line time = htotal / dotclock
2836 * surface width = hdisplay for normal plane and 64 for cursor
2837 * and latency is assumed to be high, as above.
2838 *
2839 * The final value programmed to the register should always be rounded up,
2840 * and include an extra 2 entries to account for clock crossings.
2841 *
2842 * We don't use the sprite, so we can ignore that. And on Crestline we have
2843 * to set the non-SR watermarks to 8.
2844 */
2845 void intel_update_watermarks(struct drm_crtc *crtc)
2846 {
2847 struct drm_i915_private *dev_priv = crtc->dev->dev_private;
2848
2849 if (dev_priv->display.update_wm)
2850 dev_priv->display.update_wm(crtc);
2851 }
2852
2853 void intel_update_sprite_watermarks(struct drm_plane *plane,
2854 struct drm_crtc *crtc,
2855 uint32_t sprite_width, int pixel_size,
2856 bool enabled, bool scaled)
2857 {
2858 struct drm_i915_private *dev_priv = plane->dev->dev_private;
2859
2860 if (dev_priv->display.update_sprite_wm)
2861 dev_priv->display.update_sprite_wm(plane, crtc, sprite_width,
2862 pixel_size, enabled, scaled);
2863 }
2864
2865 static struct drm_i915_gem_object *
2866 intel_alloc_context_page(struct drm_device *dev)
2867 {
2868 struct drm_i915_gem_object *ctx;
2869 int ret;
2870
2871 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2872
2873 ctx = i915_gem_alloc_object(dev, 4096);
2874 if (!ctx) {
2875 DRM_DEBUG("failed to alloc power context, RC6 disabled\n");
2876 return NULL;
2877 }
2878
2879 ret = i915_gem_obj_ggtt_pin(ctx, 4096, 0);
2880 if (ret) {
2881 DRM_ERROR("failed to pin power context: %d\n", ret);
2882 goto err_unref;
2883 }
2884
2885 ret = i915_gem_object_set_to_gtt_domain(ctx, 1);
2886 if (ret) {
2887 DRM_ERROR("failed to set-domain on power context: %d\n", ret);
2888 goto err_unpin;
2889 }
2890
2891 return ctx;
2892
2893 err_unpin:
2894 i915_gem_object_ggtt_unpin(ctx);
2895 err_unref:
2896 drm_gem_object_unreference(&ctx->base);
2897 return NULL;
2898 }
2899
2900 /**
2901 * Lock protecting IPS related data structures
2902 */
2903 DEFINE_SPINLOCK(mchdev_lock);
2904
2905 /* Global for IPS driver to get at the current i915 device. Protected by
2906 * mchdev_lock. */
2907 static struct drm_i915_private *i915_mch_dev;
2908
2909 bool ironlake_set_drps(struct drm_device *dev, u8 val)
2910 {
2911 struct drm_i915_private *dev_priv = dev->dev_private;
2912 u16 rgvswctl;
2913
2914 assert_spin_locked(&mchdev_lock);
2915
2916 rgvswctl = I915_READ16(MEMSWCTL);
2917 if (rgvswctl & MEMCTL_CMD_STS) {
2918 DRM_DEBUG("gpu busy, RCS change rejected\n");
2919 return false; /* still busy with another command */
2920 }
2921
2922 rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
2923 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
2924 I915_WRITE16(MEMSWCTL, rgvswctl);
2925 POSTING_READ16(MEMSWCTL);
2926
2927 rgvswctl |= MEMCTL_CMD_STS;
2928 I915_WRITE16(MEMSWCTL, rgvswctl);
2929
2930 return true;
2931 }
2932
2933 static void ironlake_enable_drps(struct drm_device *dev)
2934 {
2935 struct drm_i915_private *dev_priv = dev->dev_private;
2936 u32 rgvmodectl = I915_READ(MEMMODECTL);
2937 u8 fmax, fmin, fstart, vstart;
2938
2939 spin_lock_irq(&mchdev_lock);
2940
2941 /* Enable temp reporting */
2942 I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
2943 I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
2944
2945 /* 100ms RC evaluation intervals */
2946 I915_WRITE(RCUPEI, 100000);
2947 I915_WRITE(RCDNEI, 100000);
2948
2949 /* Set max/min thresholds to 90ms and 80ms respectively */
2950 I915_WRITE(RCBMAXAVG, 90000);
2951 I915_WRITE(RCBMINAVG, 80000);
2952
2953 I915_WRITE(MEMIHYST, 1);
2954
2955 /* Set up min, max, and cur for interrupt handling */
2956 fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
2957 fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
2958 fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
2959 MEMMODE_FSTART_SHIFT;
2960
2961 vstart = (I915_READ(PXVFREQ_BASE + (fstart * 4)) & PXVFREQ_PX_MASK) >>
2962 PXVFREQ_PX_SHIFT;
2963
2964 dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
2965 dev_priv->ips.fstart = fstart;
2966
2967 dev_priv->ips.max_delay = fstart;
2968 dev_priv->ips.min_delay = fmin;
2969 dev_priv->ips.cur_delay = fstart;
2970
2971 DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
2972 fmax, fmin, fstart);
2973
2974 I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
2975
2976 /*
2977 * Interrupts will be enabled in ironlake_irq_postinstall
2978 */
2979
2980 I915_WRITE(VIDSTART, vstart);
2981 POSTING_READ(VIDSTART);
2982
2983 rgvmodectl |= MEMMODE_SWMODE_EN;
2984 I915_WRITE(MEMMODECTL, rgvmodectl);
2985
2986 if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
2987 DRM_ERROR("stuck trying to change perf mode\n");
2988 mdelay(1);
2989
2990 ironlake_set_drps(dev, fstart);
2991
2992 dev_priv->ips.last_count1 = I915_READ(0x112e4) + I915_READ(0x112e8) +
2993 I915_READ(0x112e0);
2994 dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
2995 dev_priv->ips.last_count2 = I915_READ(0x112f4);
2996 dev_priv->ips.last_time2 = ktime_get_raw_ns();
2997
2998 spin_unlock_irq(&mchdev_lock);
2999 }
3000
3001 static void ironlake_disable_drps(struct drm_device *dev)
3002 {
3003 struct drm_i915_private *dev_priv = dev->dev_private;
3004 u16 rgvswctl;
3005
3006 spin_lock_irq(&mchdev_lock);
3007
3008 rgvswctl = I915_READ16(MEMSWCTL);
3009
3010 /* Ack interrupts, disable EFC interrupt */
3011 I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
3012 I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
3013 I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
3014 I915_WRITE(DEIIR, DE_PCU_EVENT);
3015 I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
3016
3017 /* Go back to the starting frequency */
3018 ironlake_set_drps(dev, dev_priv->ips.fstart);
3019 mdelay(1);
3020 rgvswctl |= MEMCTL_CMD_STS;
3021 I915_WRITE(MEMSWCTL, rgvswctl);
3022 mdelay(1);
3023
3024 spin_unlock_irq(&mchdev_lock);
3025 }
3026
3027 /* There's a funny hw issue where the hw returns all 0 when reading from
3028 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
3029 * ourselves, instead of doing a rmw cycle (which might result in us clearing
3030 * all limits and the gpu stuck at whatever frequency it is at atm).
3031 */
3032 static u32 gen6_rps_limits(struct drm_i915_private *dev_priv, u8 val)
3033 {
3034 u32 limits;
3035
3036 /* Only set the down limit when we've reached the lowest level to avoid
3037 * getting more interrupts, otherwise leave this clear. This prevents a
3038 * race in the hw when coming out of rc6: There's a tiny window where
3039 * the hw runs at the minimal clock before selecting the desired
3040 * frequency, if the down threshold expires in that window we will not
3041 * receive a down interrupt. */
3042 limits = dev_priv->rps.max_freq_softlimit << 24;
3043 if (val <= dev_priv->rps.min_freq_softlimit)
3044 limits |= dev_priv->rps.min_freq_softlimit << 16;
3045
3046 return limits;
3047 }
3048
3049 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
3050 {
3051 int new_power;
3052
3053 new_power = dev_priv->rps.power;
3054 switch (dev_priv->rps.power) {
3055 case LOW_POWER:
3056 if (val > dev_priv->rps.efficient_freq + 1 && val > dev_priv->rps.cur_freq)
3057 new_power = BETWEEN;
3058 break;
3059
3060 case BETWEEN:
3061 if (val <= dev_priv->rps.efficient_freq && val < dev_priv->rps.cur_freq)
3062 new_power = LOW_POWER;
3063 else if (val >= dev_priv->rps.rp0_freq && val > dev_priv->rps.cur_freq)
3064 new_power = HIGH_POWER;
3065 break;
3066
3067 case HIGH_POWER:
3068 if (val < (dev_priv->rps.rp1_freq + dev_priv->rps.rp0_freq) >> 1 && val < dev_priv->rps.cur_freq)
3069 new_power = BETWEEN;
3070 break;
3071 }
3072 /* Max/min bins are special */
3073 if (val == dev_priv->rps.min_freq_softlimit)
3074 new_power = LOW_POWER;
3075 if (val == dev_priv->rps.max_freq_softlimit)
3076 new_power = HIGH_POWER;
3077 if (new_power == dev_priv->rps.power)
3078 return;
3079
3080 /* Note the units here are not exactly 1us, but 1280ns. */
3081 switch (new_power) {
3082 case LOW_POWER:
3083 /* Upclock if more than 95% busy over 16ms */
3084 I915_WRITE(GEN6_RP_UP_EI, 12500);
3085 I915_WRITE(GEN6_RP_UP_THRESHOLD, 11800);
3086
3087 /* Downclock if less than 85% busy over 32ms */
3088 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3089 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 21250);
3090
3091 I915_WRITE(GEN6_RP_CONTROL,
3092 GEN6_RP_MEDIA_TURBO |
3093 GEN6_RP_MEDIA_HW_NORMAL_MODE |
3094 GEN6_RP_MEDIA_IS_GFX |
3095 GEN6_RP_ENABLE |
3096 GEN6_RP_UP_BUSY_AVG |
3097 GEN6_RP_DOWN_IDLE_AVG);
3098 break;
3099
3100 case BETWEEN:
3101 /* Upclock if more than 90% busy over 13ms */
3102 I915_WRITE(GEN6_RP_UP_EI, 10250);
3103 I915_WRITE(GEN6_RP_UP_THRESHOLD, 9225);
3104
3105 /* Downclock if less than 75% busy over 32ms */
3106 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3107 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 18750);
3108
3109 I915_WRITE(GEN6_RP_CONTROL,
3110 GEN6_RP_MEDIA_TURBO |
3111 GEN6_RP_MEDIA_HW_NORMAL_MODE |
3112 GEN6_RP_MEDIA_IS_GFX |
3113 GEN6_RP_ENABLE |
3114 GEN6_RP_UP_BUSY_AVG |
3115 GEN6_RP_DOWN_IDLE_AVG);
3116 break;
3117
3118 case HIGH_POWER:
3119 /* Upclock if more than 85% busy over 10ms */
3120 I915_WRITE(GEN6_RP_UP_EI, 8000);
3121 I915_WRITE(GEN6_RP_UP_THRESHOLD, 6800);
3122
3123 /* Downclock if less than 60% busy over 32ms */
3124 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3125 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 15000);
3126
3127 I915_WRITE(GEN6_RP_CONTROL,
3128 GEN6_RP_MEDIA_TURBO |
3129 GEN6_RP_MEDIA_HW_NORMAL_MODE |
3130 GEN6_RP_MEDIA_IS_GFX |
3131 GEN6_RP_ENABLE |
3132 GEN6_RP_UP_BUSY_AVG |
3133 GEN6_RP_DOWN_IDLE_AVG);
3134 break;
3135 }
3136
3137 dev_priv->rps.power = new_power;
3138 dev_priv->rps.last_adj = 0;
3139 }
3140
3141 static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
3142 {
3143 u32 mask = 0;
3144
3145 if (val > dev_priv->rps.min_freq_softlimit)
3146 mask |= GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
3147 if (val < dev_priv->rps.max_freq_softlimit)
3148 mask |= GEN6_PM_RP_UP_THRESHOLD;
3149
3150 /* IVB and SNB hard hangs on looping batchbuffer
3151 * if GEN6_PM_UP_EI_EXPIRED is masked.
3152 */
3153 if (INTEL_INFO(dev_priv->dev)->gen <= 7 && !IS_HASWELL(dev_priv->dev))
3154 mask |= GEN6_PM_RP_UP_EI_EXPIRED;
3155
3156 if (IS_GEN8(dev_priv->dev))
3157 mask |= GEN8_PMINTR_REDIRECT_TO_NON_DISP;
3158
3159 return ~mask;
3160 }
3161
3162 /* gen6_set_rps is called to update the frequency request, but should also be
3163 * called when the range (min_delay and max_delay) is modified so that we can
3164 * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
3165 void gen6_set_rps(struct drm_device *dev, u8 val)
3166 {
3167 struct drm_i915_private *dev_priv = dev->dev_private;
3168
3169 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3170 WARN_ON(val > dev_priv->rps.max_freq_softlimit);
3171 WARN_ON(val < dev_priv->rps.min_freq_softlimit);
3172
3173 /* min/max delay may still have been modified so be sure to
3174 * write the limits value.
3175 */
3176 if (val != dev_priv->rps.cur_freq) {
3177 gen6_set_rps_thresholds(dev_priv, val);
3178
3179 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
3180 I915_WRITE(GEN6_RPNSWREQ,
3181 HSW_FREQUENCY(val));
3182 else
3183 I915_WRITE(GEN6_RPNSWREQ,
3184 GEN6_FREQUENCY(val) |
3185 GEN6_OFFSET(0) |
3186 GEN6_AGGRESSIVE_TURBO);
3187 }
3188
3189 /* Make sure we continue to get interrupts
3190 * until we hit the minimum or maximum frequencies.
3191 */
3192 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, gen6_rps_limits(dev_priv, val));
3193 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
3194
3195 POSTING_READ(GEN6_RPNSWREQ);
3196
3197 dev_priv->rps.cur_freq = val;
3198 trace_intel_gpu_freq_change(val * 50);
3199 }
3200
3201 /* vlv_set_rps_idle: Set the frequency to Rpn if Gfx clocks are down
3202 *
3203 * * If Gfx is Idle, then
3204 * 1. Mask Turbo interrupts
3205 * 2. Bring up Gfx clock
3206 * 3. Change the freq to Rpn and wait till P-Unit updates freq
3207 * 4. Clear the Force GFX CLK ON bit so that Gfx can down
3208 * 5. Unmask Turbo interrupts
3209 */
3210 static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
3211 {
3212 struct drm_device *dev = dev_priv->dev;
3213
3214 /* Latest VLV doesn't need to force the gfx clock */
3215 if (dev->pdev->revision >= 0xd) {
3216 valleyview_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
3217 return;
3218 }
3219
3220 /*
3221 * When we are idle. Drop to min voltage state.
3222 */
3223
3224 if (dev_priv->rps.cur_freq <= dev_priv->rps.min_freq_softlimit)
3225 return;
3226
3227 /* Mask turbo interrupt so that they will not come in between */
3228 I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
3229
3230 vlv_force_gfx_clock(dev_priv, true);
3231
3232 dev_priv->rps.cur_freq = dev_priv->rps.min_freq_softlimit;
3233
3234 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ,
3235 dev_priv->rps.min_freq_softlimit);
3236
3237 if (wait_for(((vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS))
3238 & GENFREQSTATUS) == 0, 5))
3239 DRM_ERROR("timed out waiting for Punit\n");
3240
3241 vlv_force_gfx_clock(dev_priv, false);
3242
3243 I915_WRITE(GEN6_PMINTRMSK,
3244 gen6_rps_pm_mask(dev_priv, dev_priv->rps.cur_freq));
3245 }
3246
3247 void gen6_rps_idle(struct drm_i915_private *dev_priv)
3248 {
3249 struct drm_device *dev = dev_priv->dev;
3250
3251 mutex_lock(&dev_priv->rps.hw_lock);
3252 if (dev_priv->rps.enabled) {
3253 if (IS_VALLEYVIEW(dev))
3254 vlv_set_rps_idle(dev_priv);
3255 else
3256 gen6_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
3257 dev_priv->rps.last_adj = 0;
3258 }
3259 mutex_unlock(&dev_priv->rps.hw_lock);
3260 }
3261
3262 void gen6_rps_boost(struct drm_i915_private *dev_priv)
3263 {
3264 struct drm_device *dev = dev_priv->dev;
3265
3266 mutex_lock(&dev_priv->rps.hw_lock);
3267 if (dev_priv->rps.enabled) {
3268 if (IS_VALLEYVIEW(dev))
3269 valleyview_set_rps(dev_priv->dev, dev_priv->rps.max_freq_softlimit);
3270 else
3271 gen6_set_rps(dev_priv->dev, dev_priv->rps.max_freq_softlimit);
3272 dev_priv->rps.last_adj = 0;
3273 }
3274 mutex_unlock(&dev_priv->rps.hw_lock);
3275 }
3276
3277 void valleyview_set_rps(struct drm_device *dev, u8 val)
3278 {
3279 struct drm_i915_private *dev_priv = dev->dev_private;
3280
3281 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3282 WARN_ON(val > dev_priv->rps.max_freq_softlimit);
3283 WARN_ON(val < dev_priv->rps.min_freq_softlimit);
3284
3285 DRM_DEBUG_DRIVER("GPU freq request from %d MHz (%u) to %d MHz (%u)\n",
3286 vlv_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
3287 dev_priv->rps.cur_freq,
3288 vlv_gpu_freq(dev_priv, val), val);
3289
3290 if (val != dev_priv->rps.cur_freq)
3291 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
3292
3293 I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
3294
3295 dev_priv->rps.cur_freq = val;
3296 trace_intel_gpu_freq_change(vlv_gpu_freq(dev_priv, val));
3297 }
3298
3299 static void gen8_disable_rps_interrupts(struct drm_device *dev)
3300 {
3301 struct drm_i915_private *dev_priv = dev->dev_private;
3302
3303 I915_WRITE(GEN6_PMINTRMSK, ~GEN8_PMINTR_REDIRECT_TO_NON_DISP);
3304 I915_WRITE(GEN8_GT_IER(2), I915_READ(GEN8_GT_IER(2)) &
3305 ~dev_priv->pm_rps_events);
3306 /* Complete PM interrupt masking here doesn't race with the rps work
3307 * item again unmasking PM interrupts because that is using a different
3308 * register (GEN8_GT_IMR(2)) to mask PM interrupts. The only risk is in
3309 * leaving stale bits in GEN8_GT_IIR(2) and GEN8_GT_IMR(2) which
3310 * gen8_enable_rps will clean up. */
3311
3312 spin_lock_irq(&dev_priv->irq_lock);
3313 dev_priv->rps.pm_iir = 0;
3314 spin_unlock_irq(&dev_priv->irq_lock);
3315
3316 I915_WRITE(GEN8_GT_IIR(2), dev_priv->pm_rps_events);
3317 }
3318
3319 static void gen6_disable_rps_interrupts(struct drm_device *dev)
3320 {
3321 struct drm_i915_private *dev_priv = dev->dev_private;
3322
3323 I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
3324 I915_WRITE(GEN6_PMIER, I915_READ(GEN6_PMIER) &
3325 ~dev_priv->pm_rps_events);
3326 /* Complete PM interrupt masking here doesn't race with the rps work
3327 * item again unmasking PM interrupts because that is using a different
3328 * register (PMIMR) to mask PM interrupts. The only risk is in leaving
3329 * stale bits in PMIIR and PMIMR which gen6_enable_rps will clean up. */
3330
3331 spin_lock_irq(&dev_priv->irq_lock);
3332 dev_priv->rps.pm_iir = 0;
3333 spin_unlock_irq(&dev_priv->irq_lock);
3334
3335 I915_WRITE(GEN6_PMIIR, dev_priv->pm_rps_events);
3336 }
3337
3338 static void gen6_disable_rps(struct drm_device *dev)
3339 {
3340 struct drm_i915_private *dev_priv = dev->dev_private;
3341
3342 I915_WRITE(GEN6_RC_CONTROL, 0);
3343 I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
3344
3345 if (IS_BROADWELL(dev))
3346 gen8_disable_rps_interrupts(dev);
3347 else
3348 gen6_disable_rps_interrupts(dev);
3349 }
3350
3351 static void valleyview_disable_rps(struct drm_device *dev)
3352 {
3353 struct drm_i915_private *dev_priv = dev->dev_private;
3354
3355 I915_WRITE(GEN6_RC_CONTROL, 0);
3356
3357 gen6_disable_rps_interrupts(dev);
3358 }
3359
3360 static void intel_print_rc6_info(struct drm_device *dev, u32 mode)
3361 {
3362 if (IS_VALLEYVIEW(dev)) {
3363 if (mode & (GEN7_RC_CTL_TO_MODE | GEN6_RC_CTL_EI_MODE(1)))
3364 mode = GEN6_RC_CTL_RC6_ENABLE;
3365 else
3366 mode = 0;
3367 }
3368 DRM_INFO("Enabling RC6 states: RC6 %s, RC6p %s, RC6pp %s\n",
3369 (mode & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off",
3370 (mode & GEN6_RC_CTL_RC6p_ENABLE) ? "on" : "off",
3371 (mode & GEN6_RC_CTL_RC6pp_ENABLE) ? "on" : "off");
3372 }
3373
3374 static int sanitize_rc6_option(const struct drm_device *dev, int enable_rc6)
3375 {
3376 /* No RC6 before Ironlake */
3377 if (INTEL_INFO(dev)->gen < 5)
3378 return 0;
3379
3380 /* RC6 is only on Ironlake mobile not on desktop */
3381 if (INTEL_INFO(dev)->gen == 5 && !IS_IRONLAKE_M(dev))
3382 return 0;
3383
3384 /* Respect the kernel parameter if it is set */
3385 if (enable_rc6 >= 0) {
3386 int mask;
3387
3388 if (INTEL_INFO(dev)->gen == 6 || IS_IVYBRIDGE(dev))
3389 mask = INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE |
3390 INTEL_RC6pp_ENABLE;
3391 else
3392 mask = INTEL_RC6_ENABLE;
3393
3394 if ((enable_rc6 & mask) != enable_rc6)
3395 DRM_INFO("Adjusting RC6 mask to %d (requested %d, valid %d)\n",
3396 enable_rc6 & mask, enable_rc6, mask);
3397
3398 return enable_rc6 & mask;
3399 }
3400
3401 /* Disable RC6 on Ironlake */
3402 if (INTEL_INFO(dev)->gen == 5)
3403 return 0;
3404
3405 if (IS_IVYBRIDGE(dev))
3406 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
3407
3408 return INTEL_RC6_ENABLE;
3409 }
3410
3411 int intel_enable_rc6(const struct drm_device *dev)
3412 {
3413 return i915.enable_rc6;
3414 }
3415
3416 static void gen8_enable_rps_interrupts(struct drm_device *dev)
3417 {
3418 struct drm_i915_private *dev_priv = dev->dev_private;
3419
3420 spin_lock_irq(&dev_priv->irq_lock);
3421 WARN_ON(dev_priv->rps.pm_iir);
3422 bdw_enable_pm_irq(dev_priv, dev_priv->pm_rps_events);
3423 I915_WRITE(GEN8_GT_IIR(2), dev_priv->pm_rps_events);
3424 spin_unlock_irq(&dev_priv->irq_lock);
3425 }
3426
3427 static void gen6_enable_rps_interrupts(struct drm_device *dev)
3428 {
3429 struct drm_i915_private *dev_priv = dev->dev_private;
3430
3431 spin_lock_irq(&dev_priv->irq_lock);
3432 WARN_ON(dev_priv->rps.pm_iir);
3433 snb_enable_pm_irq(dev_priv, dev_priv->pm_rps_events);
3434 I915_WRITE(GEN6_PMIIR, dev_priv->pm_rps_events);
3435 spin_unlock_irq(&dev_priv->irq_lock);
3436 }
3437
3438 static void parse_rp_state_cap(struct drm_i915_private *dev_priv, u32 rp_state_cap)
3439 {
3440 /* All of these values are in units of 50MHz */
3441 dev_priv->rps.cur_freq = 0;
3442 /* static values from HW: RP0 < RPe < RP1 < RPn (min_freq) */
3443 dev_priv->rps.rp1_freq = (rp_state_cap >> 8) & 0xff;
3444 dev_priv->rps.rp0_freq = (rp_state_cap >> 0) & 0xff;
3445 dev_priv->rps.min_freq = (rp_state_cap >> 16) & 0xff;
3446 /* XXX: only BYT has a special efficient freq */
3447 dev_priv->rps.efficient_freq = dev_priv->rps.rp1_freq;
3448 /* hw_max = RP0 until we check for overclocking */
3449 dev_priv->rps.max_freq = dev_priv->rps.rp0_freq;
3450
3451 /* Preserve min/max settings in case of re-init */
3452 if (dev_priv->rps.max_freq_softlimit == 0)
3453 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
3454
3455 if (dev_priv->rps.min_freq_softlimit == 0)
3456 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
3457 }
3458
3459 static void gen8_enable_rps(struct drm_device *dev)
3460 {
3461 struct drm_i915_private *dev_priv = dev->dev_private;
3462 struct intel_engine_cs *ring;
3463 uint32_t rc6_mask = 0, rp_state_cap;
3464 int unused;
3465
3466 /* 1a: Software RC state - RC0 */
3467 I915_WRITE(GEN6_RC_STATE, 0);
3468
3469 /* 1c & 1d: Get forcewake during program sequence. Although the driver
3470 * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
3471 gen6_gt_force_wake_get(dev_priv, FORCEWAKE_ALL);
3472
3473 /* 2a: Disable RC states. */
3474 I915_WRITE(GEN6_RC_CONTROL, 0);
3475
3476 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
3477 parse_rp_state_cap(dev_priv, rp_state_cap);
3478
3479 /* 2b: Program RC6 thresholds.*/
3480 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
3481 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
3482 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
3483 for_each_ring(ring, dev_priv, unused)
3484 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
3485 I915_WRITE(GEN6_RC_SLEEP, 0);
3486 I915_WRITE(GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
3487
3488 /* 3: Enable RC6 */
3489 if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
3490 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
3491 intel_print_rc6_info(dev, rc6_mask);
3492 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
3493 GEN6_RC_CTL_EI_MODE(1) |
3494 rc6_mask);
3495
3496 /* 4 Program defaults and thresholds for RPS*/
3497 I915_WRITE(GEN6_RPNSWREQ,
3498 HSW_FREQUENCY(dev_priv->rps.rp1_freq));
3499 I915_WRITE(GEN6_RC_VIDEO_FREQ,
3500 HSW_FREQUENCY(dev_priv->rps.rp1_freq));
3501 /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
3502 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
3503
3504 /* Docs recommend 900MHz, and 300 MHz respectively */
3505 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
3506 dev_priv->rps.max_freq_softlimit << 24 |
3507 dev_priv->rps.min_freq_softlimit << 16);
3508
3509 I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
3510 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
3511 I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
3512 I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
3513
3514 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
3515
3516 /* 5: Enable RPS */
3517 I915_WRITE(GEN6_RP_CONTROL,
3518 GEN6_RP_MEDIA_TURBO |
3519 GEN6_RP_MEDIA_HW_NORMAL_MODE |
3520 GEN6_RP_MEDIA_IS_GFX |
3521 GEN6_RP_ENABLE |
3522 GEN6_RP_UP_BUSY_AVG |
3523 GEN6_RP_DOWN_IDLE_AVG);
3524
3525 /* 6: Ring frequency + overclocking (our driver does this later */
3526
3527 gen6_set_rps(dev, (I915_READ(GEN6_GT_PERF_STATUS) & 0xff00) >> 8);
3528
3529 gen8_enable_rps_interrupts(dev);
3530
3531 gen6_gt_force_wake_put(dev_priv, FORCEWAKE_ALL);
3532 }
3533
3534 static void gen6_enable_rps(struct drm_device *dev)
3535 {
3536 struct drm_i915_private *dev_priv = dev->dev_private;
3537 struct intel_engine_cs *ring;
3538 u32 rp_state_cap;
3539 u32 gt_perf_status;
3540 u32 rc6vids, pcu_mbox = 0, rc6_mask = 0;
3541 u32 gtfifodbg;
3542 int rc6_mode;
3543 int i, ret;
3544
3545 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3546
3547 /* Here begins a magic sequence of register writes to enable
3548 * auto-downclocking.
3549 *
3550 * Perhaps there might be some value in exposing these to
3551 * userspace...
3552 */
3553 I915_WRITE(GEN6_RC_STATE, 0);
3554
3555 /* Clear the DBG now so we don't confuse earlier errors */
3556 if ((gtfifodbg = I915_READ(GTFIFODBG))) {
3557 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
3558 I915_WRITE(GTFIFODBG, gtfifodbg);
3559 }
3560
3561 gen6_gt_force_wake_get(dev_priv, FORCEWAKE_ALL);
3562
3563 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
3564 gt_perf_status = I915_READ(GEN6_GT_PERF_STATUS);
3565
3566 parse_rp_state_cap(dev_priv, rp_state_cap);
3567
3568 /* disable the counters and set deterministic thresholds */
3569 I915_WRITE(GEN6_RC_CONTROL, 0);
3570
3571 I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
3572 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
3573 I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
3574 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
3575 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
3576
3577 for_each_ring(ring, dev_priv, i)
3578 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
3579
3580 I915_WRITE(GEN6_RC_SLEEP, 0);
3581 I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
3582 if (IS_IVYBRIDGE(dev))
3583 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
3584 else
3585 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
3586 I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
3587 I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
3588
3589 /* Check if we are enabling RC6 */
3590 rc6_mode = intel_enable_rc6(dev_priv->dev);
3591 if (rc6_mode & INTEL_RC6_ENABLE)
3592 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
3593
3594 /* We don't use those on Haswell */
3595 if (!IS_HASWELL(dev)) {
3596 if (rc6_mode & INTEL_RC6p_ENABLE)
3597 rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
3598
3599 if (rc6_mode & INTEL_RC6pp_ENABLE)
3600 rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
3601 }
3602
3603 intel_print_rc6_info(dev, rc6_mask);
3604
3605 I915_WRITE(GEN6_RC_CONTROL,
3606 rc6_mask |
3607 GEN6_RC_CTL_EI_MODE(1) |
3608 GEN6_RC_CTL_HW_ENABLE);
3609
3610 /* Power down if completely idle for over 50ms */
3611 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
3612 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
3613
3614 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_MIN_FREQ_TABLE, 0);
3615 if (ret)
3616 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
3617
3618 ret = sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &pcu_mbox);
3619 if (!ret && (pcu_mbox & (1<<31))) { /* OC supported */
3620 DRM_DEBUG_DRIVER("Overclocking supported. Max: %dMHz, Overclock max: %dMHz\n",
3621 (dev_priv->rps.max_freq_softlimit & 0xff) * 50,
3622 (pcu_mbox & 0xff) * 50);
3623 dev_priv->rps.max_freq = pcu_mbox & 0xff;
3624 }
3625
3626 dev_priv->rps.power = HIGH_POWER; /* force a reset */
3627 gen6_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
3628
3629 gen6_enable_rps_interrupts(dev);
3630
3631 rc6vids = 0;
3632 ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
3633 if (IS_GEN6(dev) && ret) {
3634 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
3635 } else if (IS_GEN6(dev) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
3636 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
3637 GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
3638 rc6vids &= 0xffff00;
3639 rc6vids |= GEN6_ENCODE_RC6_VID(450);
3640 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
3641 if (ret)
3642 DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
3643 }
3644
3645 gen6_gt_force_wake_put(dev_priv, FORCEWAKE_ALL);
3646 }
3647
3648 static void __gen6_update_ring_freq(struct drm_device *dev)
3649 {
3650 struct drm_i915_private *dev_priv = dev->dev_private;
3651 int min_freq = 15;
3652 unsigned int gpu_freq;
3653 unsigned int max_ia_freq, min_ring_freq;
3654 int scaling_factor = 180;
3655 struct cpufreq_policy *policy;
3656
3657 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3658
3659 policy = cpufreq_cpu_get(0);
3660 if (policy) {
3661 max_ia_freq = policy->cpuinfo.max_freq;
3662 cpufreq_cpu_put(policy);
3663 } else {
3664 /*
3665 * Default to measured freq if none found, PCU will ensure we
3666 * don't go over
3667 */
3668 max_ia_freq = tsc_khz;
3669 }
3670
3671 /* Convert from kHz to MHz */
3672 max_ia_freq /= 1000;
3673
3674 min_ring_freq = I915_READ(DCLK) & 0xf;
3675 /* convert DDR frequency from units of 266.6MHz to bandwidth */
3676 min_ring_freq = mult_frac(min_ring_freq, 8, 3);
3677
3678 /*
3679 * For each potential GPU frequency, load a ring frequency we'd like
3680 * to use for memory access. We do this by specifying the IA frequency
3681 * the PCU should use as a reference to determine the ring frequency.
3682 */
3683 for (gpu_freq = dev_priv->rps.max_freq_softlimit; gpu_freq >= dev_priv->rps.min_freq_softlimit;
3684 gpu_freq--) {
3685 int diff = dev_priv->rps.max_freq_softlimit - gpu_freq;
3686 unsigned int ia_freq = 0, ring_freq = 0;
3687
3688 if (INTEL_INFO(dev)->gen >= 8) {
3689 /* max(2 * GT, DDR). NB: GT is 50MHz units */
3690 ring_freq = max(min_ring_freq, gpu_freq);
3691 } else if (IS_HASWELL(dev)) {
3692 ring_freq = mult_frac(gpu_freq, 5, 4);
3693 ring_freq = max(min_ring_freq, ring_freq);
3694 /* leave ia_freq as the default, chosen by cpufreq */
3695 } else {
3696 /* On older processors, there is no separate ring
3697 * clock domain, so in order to boost the bandwidth
3698 * of the ring, we need to upclock the CPU (ia_freq).
3699 *
3700 * For GPU frequencies less than 750MHz,
3701 * just use the lowest ring freq.
3702 */
3703 if (gpu_freq < min_freq)
3704 ia_freq = 800;
3705 else
3706 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
3707 ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
3708 }
3709
3710 sandybridge_pcode_write(dev_priv,
3711 GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
3712 ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
3713 ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
3714 gpu_freq);
3715 }
3716 }
3717
3718 void gen6_update_ring_freq(struct drm_device *dev)
3719 {
3720 struct drm_i915_private *dev_priv = dev->dev_private;
3721
3722 if (INTEL_INFO(dev)->gen < 6 || IS_VALLEYVIEW(dev))
3723 return;
3724
3725 mutex_lock(&dev_priv->rps.hw_lock);
3726 __gen6_update_ring_freq(dev);
3727 mutex_unlock(&dev_priv->rps.hw_lock);
3728 }
3729
3730 int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
3731 {
3732 u32 val, rp0;
3733
3734 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
3735
3736 rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
3737 /* Clamp to max */
3738 rp0 = min_t(u32, rp0, 0xea);
3739
3740 return rp0;
3741 }
3742
3743 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
3744 {
3745 u32 val, rpe;
3746
3747 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
3748 rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
3749 val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
3750 rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
3751
3752 return rpe;
3753 }
3754
3755 int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
3756 {
3757 return vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
3758 }
3759
3760 /* Check that the pctx buffer wasn't move under us. */
3761 static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
3762 {
3763 unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
3764
3765 WARN_ON(pctx_addr != dev_priv->mm.stolen_base +
3766 dev_priv->vlv_pctx->stolen->start);
3767 }
3768
3769 static void valleyview_setup_pctx(struct drm_device *dev)
3770 {
3771 struct drm_i915_private *dev_priv = dev->dev_private;
3772 struct drm_i915_gem_object *pctx;
3773 unsigned long pctx_paddr;
3774 u32 pcbr;
3775 int pctx_size = 24*1024;
3776
3777 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
3778
3779 pcbr = I915_READ(VLV_PCBR);
3780 if (pcbr) {
3781 /* BIOS set it up already, grab the pre-alloc'd space */
3782 int pcbr_offset;
3783
3784 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
3785 pctx = i915_gem_object_create_stolen_for_preallocated(dev_priv->dev,
3786 pcbr_offset,
3787 I915_GTT_OFFSET_NONE,
3788 pctx_size);
3789 goto out;
3790 }
3791
3792 /*
3793 * From the Gunit register HAS:
3794 * The Gfx driver is expected to program this register and ensure
3795 * proper allocation within Gfx stolen memory. For example, this
3796 * register should be programmed such than the PCBR range does not
3797 * overlap with other ranges, such as the frame buffer, protected
3798 * memory, or any other relevant ranges.
3799 */
3800 pctx = i915_gem_object_create_stolen(dev, pctx_size);
3801 if (!pctx) {
3802 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
3803 return;
3804 }
3805
3806 pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
3807 I915_WRITE(VLV_PCBR, pctx_paddr);
3808
3809 out:
3810 dev_priv->vlv_pctx = pctx;
3811 }
3812
3813 static void valleyview_cleanup_pctx(struct drm_device *dev)
3814 {
3815 struct drm_i915_private *dev_priv = dev->dev_private;
3816
3817 if (WARN_ON(!dev_priv->vlv_pctx))
3818 return;
3819
3820 drm_gem_object_unreference(&dev_priv->vlv_pctx->base);
3821 dev_priv->vlv_pctx = NULL;
3822 }
3823
3824 static void valleyview_init_gt_powersave(struct drm_device *dev)
3825 {
3826 struct drm_i915_private *dev_priv = dev->dev_private;
3827
3828 valleyview_setup_pctx(dev);
3829
3830 mutex_lock(&dev_priv->rps.hw_lock);
3831
3832 dev_priv->rps.max_freq = valleyview_rps_max_freq(dev_priv);
3833 dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
3834 DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
3835 vlv_gpu_freq(dev_priv, dev_priv->rps.max_freq),
3836 dev_priv->rps.max_freq);
3837
3838 dev_priv->rps.efficient_freq = valleyview_rps_rpe_freq(dev_priv);
3839 DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
3840 vlv_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
3841 dev_priv->rps.efficient_freq);
3842
3843 dev_priv->rps.min_freq = valleyview_rps_min_freq(dev_priv);
3844 DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
3845 vlv_gpu_freq(dev_priv, dev_priv->rps.min_freq),
3846 dev_priv->rps.min_freq);
3847
3848 /* Preserve min/max settings in case of re-init */
3849 if (dev_priv->rps.max_freq_softlimit == 0)
3850 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
3851
3852 if (dev_priv->rps.min_freq_softlimit == 0)
3853 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
3854
3855 mutex_unlock(&dev_priv->rps.hw_lock);
3856 }
3857
3858 static void valleyview_cleanup_gt_powersave(struct drm_device *dev)
3859 {
3860 valleyview_cleanup_pctx(dev);
3861 }
3862
3863 static void valleyview_enable_rps(struct drm_device *dev)
3864 {
3865 struct drm_i915_private *dev_priv = dev->dev_private;
3866 struct intel_engine_cs *ring;
3867 u32 gtfifodbg, val, rc6_mode = 0;
3868 int i;
3869
3870 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3871
3872 valleyview_check_pctx(dev_priv);
3873
3874 if ((gtfifodbg = I915_READ(GTFIFODBG))) {
3875 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
3876 gtfifodbg);
3877 I915_WRITE(GTFIFODBG, gtfifodbg);
3878 }
3879
3880 /* If VLV, Forcewake all wells, else re-direct to regular path */
3881 gen6_gt_force_wake_get(dev_priv, FORCEWAKE_ALL);
3882
3883 I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
3884 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
3885 I915_WRITE(GEN6_RP_UP_EI, 66000);
3886 I915_WRITE(GEN6_RP_DOWN_EI, 350000);
3887
3888 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
3889
3890 I915_WRITE(GEN6_RP_CONTROL,
3891 GEN6_RP_MEDIA_TURBO |
3892 GEN6_RP_MEDIA_HW_NORMAL_MODE |
3893 GEN6_RP_MEDIA_IS_GFX |
3894 GEN6_RP_ENABLE |
3895 GEN6_RP_UP_BUSY_AVG |
3896 GEN6_RP_DOWN_IDLE_CONT);
3897
3898 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
3899 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
3900 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
3901
3902 for_each_ring(ring, dev_priv, i)
3903 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
3904
3905 I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
3906
3907 /* allows RC6 residency counter to work */
3908 I915_WRITE(VLV_COUNTER_CONTROL,
3909 _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
3910 VLV_MEDIA_RC6_COUNT_EN |
3911 VLV_RENDER_RC6_COUNT_EN));
3912 if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
3913 rc6_mode = GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
3914
3915 intel_print_rc6_info(dev, rc6_mode);
3916
3917 I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
3918
3919 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
3920
3921 DRM_DEBUG_DRIVER("GPLL enabled? %s\n", val & 0x10 ? "yes" : "no");
3922 DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
3923
3924 dev_priv->rps.cur_freq = (val >> 8) & 0xff;
3925 DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
3926 vlv_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
3927 dev_priv->rps.cur_freq);
3928
3929 DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
3930 vlv_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
3931 dev_priv->rps.efficient_freq);
3932
3933 valleyview_set_rps(dev_priv->dev, dev_priv->rps.efficient_freq);
3934
3935 gen6_enable_rps_interrupts(dev);
3936
3937 gen6_gt_force_wake_put(dev_priv, FORCEWAKE_ALL);
3938 }
3939
3940 void ironlake_teardown_rc6(struct drm_device *dev)
3941 {
3942 struct drm_i915_private *dev_priv = dev->dev_private;
3943
3944 if (dev_priv->ips.renderctx) {
3945 i915_gem_object_ggtt_unpin(dev_priv->ips.renderctx);
3946 drm_gem_object_unreference(&dev_priv->ips.renderctx->base);
3947 dev_priv->ips.renderctx = NULL;
3948 }
3949
3950 if (dev_priv->ips.pwrctx) {
3951 i915_gem_object_ggtt_unpin(dev_priv->ips.pwrctx);
3952 drm_gem_object_unreference(&dev_priv->ips.pwrctx->base);
3953 dev_priv->ips.pwrctx = NULL;
3954 }
3955 }
3956
3957 static void ironlake_disable_rc6(struct drm_device *dev)
3958 {
3959 struct drm_i915_private *dev_priv = dev->dev_private;
3960
3961 if (I915_READ(PWRCTXA)) {
3962 /* Wake the GPU, prevent RC6, then restore RSTDBYCTL */
3963 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) | RCX_SW_EXIT);
3964 wait_for(((I915_READ(RSTDBYCTL) & RSX_STATUS_MASK) == RSX_STATUS_ON),
3965 50);
3966
3967 I915_WRITE(PWRCTXA, 0);
3968 POSTING_READ(PWRCTXA);
3969
3970 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
3971 POSTING_READ(RSTDBYCTL);
3972 }
3973 }
3974
3975 static int ironlake_setup_rc6(struct drm_device *dev)
3976 {
3977 struct drm_i915_private *dev_priv = dev->dev_private;
3978
3979 if (dev_priv->ips.renderctx == NULL)
3980 dev_priv->ips.renderctx = intel_alloc_context_page(dev);
3981 if (!dev_priv->ips.renderctx)
3982 return -ENOMEM;
3983
3984 if (dev_priv->ips.pwrctx == NULL)
3985 dev_priv->ips.pwrctx = intel_alloc_context_page(dev);
3986 if (!dev_priv->ips.pwrctx) {
3987 ironlake_teardown_rc6(dev);
3988 return -ENOMEM;
3989 }
3990
3991 return 0;
3992 }
3993
3994 static void ironlake_enable_rc6(struct drm_device *dev)
3995 {
3996 struct drm_i915_private *dev_priv = dev->dev_private;
3997 struct intel_engine_cs *ring = &dev_priv->ring[RCS];
3998 bool was_interruptible;
3999 int ret;
4000
4001 /* rc6 disabled by default due to repeated reports of hanging during
4002 * boot and resume.
4003 */
4004 if (!intel_enable_rc6(dev))
4005 return;
4006
4007 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
4008
4009 ret = ironlake_setup_rc6(dev);
4010 if (ret)
4011 return;
4012
4013 was_interruptible = dev_priv->mm.interruptible;
4014 dev_priv->mm.interruptible = false;
4015
4016 /*
4017 * GPU can automatically power down the render unit if given a page
4018 * to save state.
4019 */
4020 ret = intel_ring_begin(ring, 6);
4021 if (ret) {
4022 ironlake_teardown_rc6(dev);
4023 dev_priv->mm.interruptible = was_interruptible;
4024 return;
4025 }
4026
4027 intel_ring_emit(ring, MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN);
4028 intel_ring_emit(ring, MI_SET_CONTEXT);
4029 intel_ring_emit(ring, i915_gem_obj_ggtt_offset(dev_priv->ips.renderctx) |
4030 MI_MM_SPACE_GTT |
4031 MI_SAVE_EXT_STATE_EN |
4032 MI_RESTORE_EXT_STATE_EN |
4033 MI_RESTORE_INHIBIT);
4034 intel_ring_emit(ring, MI_SUSPEND_FLUSH);
4035 intel_ring_emit(ring, MI_NOOP);
4036 intel_ring_emit(ring, MI_FLUSH);
4037 intel_ring_advance(ring);
4038
4039 /*
4040 * Wait for the command parser to advance past MI_SET_CONTEXT. The HW
4041 * does an implicit flush, combined with MI_FLUSH above, it should be
4042 * safe to assume that renderctx is valid
4043 */
4044 ret = intel_ring_idle(ring);
4045 dev_priv->mm.interruptible = was_interruptible;
4046 if (ret) {
4047 DRM_ERROR("failed to enable ironlake power savings\n");
4048 ironlake_teardown_rc6(dev);
4049 return;
4050 }
4051
4052 I915_WRITE(PWRCTXA, i915_gem_obj_ggtt_offset(dev_priv->ips.pwrctx) | PWRCTX_EN);
4053 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
4054
4055 intel_print_rc6_info(dev, GEN6_RC_CTL_RC6_ENABLE);
4056 }
4057
4058 static unsigned long intel_pxfreq(u32 vidfreq)
4059 {
4060 unsigned long freq;
4061 int div = (vidfreq & 0x3f0000) >> 16;
4062 int post = (vidfreq & 0x3000) >> 12;
4063 int pre = (vidfreq & 0x7);
4064
4065 if (!pre)
4066 return 0;
4067
4068 freq = ((div * 133333) / ((1<<post) * pre));
4069
4070 return freq;
4071 }
4072
4073 static const struct cparams {
4074 u16 i;
4075 u16 t;
4076 u16 m;
4077 u16 c;
4078 } cparams[] = {
4079 { 1, 1333, 301, 28664 },
4080 { 1, 1066, 294, 24460 },
4081 { 1, 800, 294, 25192 },
4082 { 0, 1333, 276, 27605 },
4083 { 0, 1066, 276, 27605 },
4084 { 0, 800, 231, 23784 },
4085 };
4086
4087 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
4088 {
4089 u64 total_count, diff, ret;
4090 u32 count1, count2, count3, m = 0, c = 0;
4091 unsigned long now = jiffies_to_msecs(jiffies), diff1;
4092 int i;
4093
4094 assert_spin_locked(&mchdev_lock);
4095
4096 diff1 = now - dev_priv->ips.last_time1;
4097
4098 /* Prevent division-by-zero if we are asking too fast.
4099 * Also, we don't get interesting results if we are polling
4100 * faster than once in 10ms, so just return the saved value
4101 * in such cases.
4102 */
4103 if (diff1 <= 10)
4104 return dev_priv->ips.chipset_power;
4105
4106 count1 = I915_READ(DMIEC);
4107 count2 = I915_READ(DDREC);
4108 count3 = I915_READ(CSIEC);
4109
4110 total_count = count1 + count2 + count3;
4111
4112 /* FIXME: handle per-counter overflow */
4113 if (total_count < dev_priv->ips.last_count1) {
4114 diff = ~0UL - dev_priv->ips.last_count1;
4115 diff += total_count;
4116 } else {
4117 diff = total_count - dev_priv->ips.last_count1;
4118 }
4119
4120 for (i = 0; i < ARRAY_SIZE(cparams); i++) {
4121 if (cparams[i].i == dev_priv->ips.c_m &&
4122 cparams[i].t == dev_priv->ips.r_t) {
4123 m = cparams[i].m;
4124 c = cparams[i].c;
4125 break;
4126 }
4127 }
4128
4129 diff = div_u64(diff, diff1);
4130 ret = ((m * diff) + c);
4131 ret = div_u64(ret, 10);
4132
4133 dev_priv->ips.last_count1 = total_count;
4134 dev_priv->ips.last_time1 = now;
4135
4136 dev_priv->ips.chipset_power = ret;
4137
4138 return ret;
4139 }
4140
4141 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
4142 {
4143 struct drm_device *dev = dev_priv->dev;
4144 unsigned long val;
4145
4146 if (INTEL_INFO(dev)->gen != 5)
4147 return 0;
4148
4149 spin_lock_irq(&mchdev_lock);
4150
4151 val = __i915_chipset_val(dev_priv);
4152
4153 spin_unlock_irq(&mchdev_lock);
4154
4155 return val;
4156 }
4157
4158 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
4159 {
4160 unsigned long m, x, b;
4161 u32 tsfs;
4162
4163 tsfs = I915_READ(TSFS);
4164
4165 m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
4166 x = I915_READ8(TR1);
4167
4168 b = tsfs & TSFS_INTR_MASK;
4169
4170 return ((m * x) / 127) - b;
4171 }
4172
4173 static u16 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
4174 {
4175 struct drm_device *dev = dev_priv->dev;
4176 static const struct v_table {
4177 u16 vd; /* in .1 mil */
4178 u16 vm; /* in .1 mil */
4179 } v_table[] = {
4180 { 0, 0, },
4181 { 375, 0, },
4182 { 500, 0, },
4183 { 625, 0, },
4184 { 750, 0, },
4185 { 875, 0, },
4186 { 1000, 0, },
4187 { 1125, 0, },
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 { 4125, 3000, },
4210 { 4125, 3000, },
4211 { 4125, 3000, },
4212 { 4250, 3125, },
4213 { 4375, 3250, },
4214 { 4500, 3375, },
4215 { 4625, 3500, },
4216 { 4750, 3625, },
4217 { 4875, 3750, },
4218 { 5000, 3875, },
4219 { 5125, 4000, },
4220 { 5250, 4125, },
4221 { 5375, 4250, },
4222 { 5500, 4375, },
4223 { 5625, 4500, },
4224 { 5750, 4625, },
4225 { 5875, 4750, },
4226 { 6000, 4875, },
4227 { 6125, 5000, },
4228 { 6250, 5125, },
4229 { 6375, 5250, },
4230 { 6500, 5375, },
4231 { 6625, 5500, },
4232 { 6750, 5625, },
4233 { 6875, 5750, },
4234 { 7000, 5875, },
4235 { 7125, 6000, },
4236 { 7250, 6125, },
4237 { 7375, 6250, },
4238 { 7500, 6375, },
4239 { 7625, 6500, },
4240 { 7750, 6625, },
4241 { 7875, 6750, },
4242 { 8000, 6875, },
4243 { 8125, 7000, },
4244 { 8250, 7125, },
4245 { 8375, 7250, },
4246 { 8500, 7375, },
4247 { 8625, 7500, },
4248 { 8750, 7625, },
4249 { 8875, 7750, },
4250 { 9000, 7875, },
4251 { 9125, 8000, },
4252 { 9250, 8125, },
4253 { 9375, 8250, },
4254 { 9500, 8375, },
4255 { 9625, 8500, },
4256 { 9750, 8625, },
4257 { 9875, 8750, },
4258 { 10000, 8875, },
4259 { 10125, 9000, },
4260 { 10250, 9125, },
4261 { 10375, 9250, },
4262 { 10500, 9375, },
4263 { 10625, 9500, },
4264 { 10750, 9625, },
4265 { 10875, 9750, },
4266 { 11000, 9875, },
4267 { 11125, 10000, },
4268 { 11250, 10125, },
4269 { 11375, 10250, },
4270 { 11500, 10375, },
4271 { 11625, 10500, },
4272 { 11750, 10625, },
4273 { 11875, 10750, },
4274 { 12000, 10875, },
4275 { 12125, 11000, },
4276 { 12250, 11125, },
4277 { 12375, 11250, },
4278 { 12500, 11375, },
4279 { 12625, 11500, },
4280 { 12750, 11625, },
4281 { 12875, 11750, },
4282 { 13000, 11875, },
4283 { 13125, 12000, },
4284 { 13250, 12125, },
4285 { 13375, 12250, },
4286 { 13500, 12375, },
4287 { 13625, 12500, },
4288 { 13750, 12625, },
4289 { 13875, 12750, },
4290 { 14000, 12875, },
4291 { 14125, 13000, },
4292 { 14250, 13125, },
4293 { 14375, 13250, },
4294 { 14500, 13375, },
4295 { 14625, 13500, },
4296 { 14750, 13625, },
4297 { 14875, 13750, },
4298 { 15000, 13875, },
4299 { 15125, 14000, },
4300 { 15250, 14125, },
4301 { 15375, 14250, },
4302 { 15500, 14375, },
4303 { 15625, 14500, },
4304 { 15750, 14625, },
4305 { 15875, 14750, },
4306 { 16000, 14875, },
4307 { 16125, 15000, },
4308 };
4309 if (INTEL_INFO(dev)->is_mobile)
4310 return v_table[pxvid].vm;
4311 else
4312 return v_table[pxvid].vd;
4313 }
4314
4315 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
4316 {
4317 u64 now, diff, diffms;
4318 u32 count;
4319
4320 assert_spin_locked(&mchdev_lock);
4321
4322 now = ktime_get_raw_ns();
4323 diffms = now - dev_priv->ips.last_time2;
4324 do_div(diffms, NSEC_PER_MSEC);
4325
4326 /* Don't divide by 0 */
4327 if (!diffms)
4328 return;
4329
4330 count = I915_READ(GFXEC);
4331
4332 if (count < dev_priv->ips.last_count2) {
4333 diff = ~0UL - dev_priv->ips.last_count2;
4334 diff += count;
4335 } else {
4336 diff = count - dev_priv->ips.last_count2;
4337 }
4338
4339 dev_priv->ips.last_count2 = count;
4340 dev_priv->ips.last_time2 = now;
4341
4342 /* More magic constants... */
4343 diff = diff * 1181;
4344 diff = div_u64(diff, diffms * 10);
4345 dev_priv->ips.gfx_power = diff;
4346 }
4347
4348 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
4349 {
4350 struct drm_device *dev = dev_priv->dev;
4351
4352 if (INTEL_INFO(dev)->gen != 5)
4353 return;
4354
4355 spin_lock_irq(&mchdev_lock);
4356
4357 __i915_update_gfx_val(dev_priv);
4358
4359 spin_unlock_irq(&mchdev_lock);
4360 }
4361
4362 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
4363 {
4364 unsigned long t, corr, state1, corr2, state2;
4365 u32 pxvid, ext_v;
4366
4367 assert_spin_locked(&mchdev_lock);
4368
4369 pxvid = I915_READ(PXVFREQ_BASE + (dev_priv->rps.cur_freq * 4));
4370 pxvid = (pxvid >> 24) & 0x7f;
4371 ext_v = pvid_to_extvid(dev_priv, pxvid);
4372
4373 state1 = ext_v;
4374
4375 t = i915_mch_val(dev_priv);
4376
4377 /* Revel in the empirically derived constants */
4378
4379 /* Correction factor in 1/100000 units */
4380 if (t > 80)
4381 corr = ((t * 2349) + 135940);
4382 else if (t >= 50)
4383 corr = ((t * 964) + 29317);
4384 else /* < 50 */
4385 corr = ((t * 301) + 1004);
4386
4387 corr = corr * ((150142 * state1) / 10000 - 78642);
4388 corr /= 100000;
4389 corr2 = (corr * dev_priv->ips.corr);
4390
4391 state2 = (corr2 * state1) / 10000;
4392 state2 /= 100; /* convert to mW */
4393
4394 __i915_update_gfx_val(dev_priv);
4395
4396 return dev_priv->ips.gfx_power + state2;
4397 }
4398
4399 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
4400 {
4401 struct drm_device *dev = dev_priv->dev;
4402 unsigned long val;
4403
4404 if (INTEL_INFO(dev)->gen != 5)
4405 return 0;
4406
4407 spin_lock_irq(&mchdev_lock);
4408
4409 val = __i915_gfx_val(dev_priv);
4410
4411 spin_unlock_irq(&mchdev_lock);
4412
4413 return val;
4414 }
4415
4416 /**
4417 * i915_read_mch_val - return value for IPS use
4418 *
4419 * Calculate and return a value for the IPS driver to use when deciding whether
4420 * we have thermal and power headroom to increase CPU or GPU power budget.
4421 */
4422 unsigned long i915_read_mch_val(void)
4423 {
4424 struct drm_i915_private *dev_priv;
4425 unsigned long chipset_val, graphics_val, ret = 0;
4426
4427 spin_lock_irq(&mchdev_lock);
4428 if (!i915_mch_dev)
4429 goto out_unlock;
4430 dev_priv = i915_mch_dev;
4431
4432 chipset_val = __i915_chipset_val(dev_priv);
4433 graphics_val = __i915_gfx_val(dev_priv);
4434
4435 ret = chipset_val + graphics_val;
4436
4437 out_unlock:
4438 spin_unlock_irq(&mchdev_lock);
4439
4440 return ret;
4441 }
4442 EXPORT_SYMBOL_GPL(i915_read_mch_val);
4443
4444 /**
4445 * i915_gpu_raise - raise GPU frequency limit
4446 *
4447 * Raise the limit; IPS indicates we have thermal headroom.
4448 */
4449 bool i915_gpu_raise(void)
4450 {
4451 struct drm_i915_private *dev_priv;
4452 bool ret = true;
4453
4454 spin_lock_irq(&mchdev_lock);
4455 if (!i915_mch_dev) {
4456 ret = false;
4457 goto out_unlock;
4458 }
4459 dev_priv = i915_mch_dev;
4460
4461 if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
4462 dev_priv->ips.max_delay--;
4463
4464 out_unlock:
4465 spin_unlock_irq(&mchdev_lock);
4466
4467 return ret;
4468 }
4469 EXPORT_SYMBOL_GPL(i915_gpu_raise);
4470
4471 /**
4472 * i915_gpu_lower - lower GPU frequency limit
4473 *
4474 * IPS indicates we're close to a thermal limit, so throttle back the GPU
4475 * frequency maximum.
4476 */
4477 bool i915_gpu_lower(void)
4478 {
4479 struct drm_i915_private *dev_priv;
4480 bool ret = true;
4481
4482 spin_lock_irq(&mchdev_lock);
4483 if (!i915_mch_dev) {
4484 ret = false;
4485 goto out_unlock;
4486 }
4487 dev_priv = i915_mch_dev;
4488
4489 if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
4490 dev_priv->ips.max_delay++;
4491
4492 out_unlock:
4493 spin_unlock_irq(&mchdev_lock);
4494
4495 return ret;
4496 }
4497 EXPORT_SYMBOL_GPL(i915_gpu_lower);
4498
4499 /**
4500 * i915_gpu_busy - indicate GPU business to IPS
4501 *
4502 * Tell the IPS driver whether or not the GPU is busy.
4503 */
4504 bool i915_gpu_busy(void)
4505 {
4506 struct drm_i915_private *dev_priv;
4507 struct intel_engine_cs *ring;
4508 bool ret = false;
4509 int i;
4510
4511 spin_lock_irq(&mchdev_lock);
4512 if (!i915_mch_dev)
4513 goto out_unlock;
4514 dev_priv = i915_mch_dev;
4515
4516 for_each_ring(ring, dev_priv, i)
4517 ret |= !list_empty(&ring->request_list);
4518
4519 out_unlock:
4520 spin_unlock_irq(&mchdev_lock);
4521
4522 return ret;
4523 }
4524 EXPORT_SYMBOL_GPL(i915_gpu_busy);
4525
4526 /**
4527 * i915_gpu_turbo_disable - disable graphics turbo
4528 *
4529 * Disable graphics turbo by resetting the max frequency and setting the
4530 * current frequency to the default.
4531 */
4532 bool i915_gpu_turbo_disable(void)
4533 {
4534 struct drm_i915_private *dev_priv;
4535 bool ret = true;
4536
4537 spin_lock_irq(&mchdev_lock);
4538 if (!i915_mch_dev) {
4539 ret = false;
4540 goto out_unlock;
4541 }
4542 dev_priv = i915_mch_dev;
4543
4544 dev_priv->ips.max_delay = dev_priv->ips.fstart;
4545
4546 if (!ironlake_set_drps(dev_priv->dev, dev_priv->ips.fstart))
4547 ret = false;
4548
4549 out_unlock:
4550 spin_unlock_irq(&mchdev_lock);
4551
4552 return ret;
4553 }
4554 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
4555
4556 /**
4557 * Tells the intel_ips driver that the i915 driver is now loaded, if
4558 * IPS got loaded first.
4559 *
4560 * This awkward dance is so that neither module has to depend on the
4561 * other in order for IPS to do the appropriate communication of
4562 * GPU turbo limits to i915.
4563 */
4564 static void
4565 ips_ping_for_i915_load(void)
4566 {
4567 void (*link)(void);
4568
4569 link = symbol_get(ips_link_to_i915_driver);
4570 if (link) {
4571 link();
4572 symbol_put(ips_link_to_i915_driver);
4573 }
4574 }
4575
4576 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
4577 {
4578 /* We only register the i915 ips part with intel-ips once everything is
4579 * set up, to avoid intel-ips sneaking in and reading bogus values. */
4580 spin_lock_irq(&mchdev_lock);
4581 i915_mch_dev = dev_priv;
4582 spin_unlock_irq(&mchdev_lock);
4583
4584 ips_ping_for_i915_load();
4585 }
4586
4587 void intel_gpu_ips_teardown(void)
4588 {
4589 spin_lock_irq(&mchdev_lock);
4590 i915_mch_dev = NULL;
4591 spin_unlock_irq(&mchdev_lock);
4592 }
4593
4594 static void intel_init_emon(struct drm_device *dev)
4595 {
4596 struct drm_i915_private *dev_priv = dev->dev_private;
4597 u32 lcfuse;
4598 u8 pxw[16];
4599 int i;
4600
4601 /* Disable to program */
4602 I915_WRITE(ECR, 0);
4603 POSTING_READ(ECR);
4604
4605 /* Program energy weights for various events */
4606 I915_WRITE(SDEW, 0x15040d00);
4607 I915_WRITE(CSIEW0, 0x007f0000);
4608 I915_WRITE(CSIEW1, 0x1e220004);
4609 I915_WRITE(CSIEW2, 0x04000004);
4610
4611 for (i = 0; i < 5; i++)
4612 I915_WRITE(PEW + (i * 4), 0);
4613 for (i = 0; i < 3; i++)
4614 I915_WRITE(DEW + (i * 4), 0);
4615
4616 /* Program P-state weights to account for frequency power adjustment */
4617 for (i = 0; i < 16; i++) {
4618 u32 pxvidfreq = I915_READ(PXVFREQ_BASE + (i * 4));
4619 unsigned long freq = intel_pxfreq(pxvidfreq);
4620 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
4621 PXVFREQ_PX_SHIFT;
4622 unsigned long val;
4623
4624 val = vid * vid;
4625 val *= (freq / 1000);
4626 val *= 255;
4627 val /= (127*127*900);
4628 if (val > 0xff)
4629 DRM_ERROR("bad pxval: %ld\n", val);
4630 pxw[i] = val;
4631 }
4632 /* Render standby states get 0 weight */
4633 pxw[14] = 0;
4634 pxw[15] = 0;
4635
4636 for (i = 0; i < 4; i++) {
4637 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
4638 (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
4639 I915_WRITE(PXW + (i * 4), val);
4640 }
4641
4642 /* Adjust magic regs to magic values (more experimental results) */
4643 I915_WRITE(OGW0, 0);
4644 I915_WRITE(OGW1, 0);
4645 I915_WRITE(EG0, 0x00007f00);
4646 I915_WRITE(EG1, 0x0000000e);
4647 I915_WRITE(EG2, 0x000e0000);
4648 I915_WRITE(EG3, 0x68000300);
4649 I915_WRITE(EG4, 0x42000000);
4650 I915_WRITE(EG5, 0x00140031);
4651 I915_WRITE(EG6, 0);
4652 I915_WRITE(EG7, 0);
4653
4654 for (i = 0; i < 8; i++)
4655 I915_WRITE(PXWL + (i * 4), 0);
4656
4657 /* Enable PMON + select events */
4658 I915_WRITE(ECR, 0x80000019);
4659
4660 lcfuse = I915_READ(LCFUSE02);
4661
4662 dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
4663 }
4664
4665 void intel_init_gt_powersave(struct drm_device *dev)
4666 {
4667 i915.enable_rc6 = sanitize_rc6_option(dev, i915.enable_rc6);
4668
4669 if (IS_VALLEYVIEW(dev))
4670 valleyview_init_gt_powersave(dev);
4671 }
4672
4673 void intel_cleanup_gt_powersave(struct drm_device *dev)
4674 {
4675 if (IS_VALLEYVIEW(dev))
4676 valleyview_cleanup_gt_powersave(dev);
4677 }
4678
4679 void intel_disable_gt_powersave(struct drm_device *dev)
4680 {
4681 struct drm_i915_private *dev_priv = dev->dev_private;
4682
4683 /* Interrupts should be disabled already to avoid re-arming. */
4684 WARN_ON(dev->irq_enabled);
4685
4686 if (IS_IRONLAKE_M(dev)) {
4687 ironlake_disable_drps(dev);
4688 ironlake_disable_rc6(dev);
4689 } else if (IS_GEN6(dev) || IS_GEN7(dev) || IS_BROADWELL(dev)) {
4690 if (cancel_delayed_work_sync(&dev_priv->rps.delayed_resume_work))
4691 intel_runtime_pm_put(dev_priv);
4692
4693 cancel_work_sync(&dev_priv->rps.work);
4694 mutex_lock(&dev_priv->rps.hw_lock);
4695 if (IS_VALLEYVIEW(dev))
4696 valleyview_disable_rps(dev);
4697 else
4698 gen6_disable_rps(dev);
4699 dev_priv->rps.enabled = false;
4700 mutex_unlock(&dev_priv->rps.hw_lock);
4701 }
4702 }
4703
4704 static void intel_gen6_powersave_work(struct work_struct *work)
4705 {
4706 struct drm_i915_private *dev_priv =
4707 container_of(work, struct drm_i915_private,
4708 rps.delayed_resume_work.work);
4709 struct drm_device *dev = dev_priv->dev;
4710
4711 mutex_lock(&dev_priv->rps.hw_lock);
4712
4713 if (IS_VALLEYVIEW(dev)) {
4714 valleyview_enable_rps(dev);
4715 } else if (IS_BROADWELL(dev)) {
4716 gen8_enable_rps(dev);
4717 __gen6_update_ring_freq(dev);
4718 } else {
4719 gen6_enable_rps(dev);
4720 __gen6_update_ring_freq(dev);
4721 }
4722 dev_priv->rps.enabled = true;
4723 mutex_unlock(&dev_priv->rps.hw_lock);
4724
4725 intel_runtime_pm_put(dev_priv);
4726 }
4727
4728 void intel_enable_gt_powersave(struct drm_device *dev)
4729 {
4730 struct drm_i915_private *dev_priv = dev->dev_private;
4731
4732 if (IS_IRONLAKE_M(dev)) {
4733 mutex_lock(&dev->struct_mutex);
4734 ironlake_enable_drps(dev);
4735 ironlake_enable_rc6(dev);
4736 intel_init_emon(dev);
4737 mutex_unlock(&dev->struct_mutex);
4738 } else if (IS_GEN6(dev) || IS_GEN7(dev) || IS_BROADWELL(dev)) {
4739 /*
4740 * PCU communication is slow and this doesn't need to be
4741 * done at any specific time, so do this out of our fast path
4742 * to make resume and init faster.
4743 *
4744 * We depend on the HW RC6 power context save/restore
4745 * mechanism when entering D3 through runtime PM suspend. So
4746 * disable RPM until RPS/RC6 is properly setup. We can only
4747 * get here via the driver load/system resume/runtime resume
4748 * paths, so the _noresume version is enough (and in case of
4749 * runtime resume it's necessary).
4750 */
4751 if (schedule_delayed_work(&dev_priv->rps.delayed_resume_work,
4752 round_jiffies_up_relative(HZ)))
4753 intel_runtime_pm_get_noresume(dev_priv);
4754 }
4755 }
4756
4757 void intel_reset_gt_powersave(struct drm_device *dev)
4758 {
4759 struct drm_i915_private *dev_priv = dev->dev_private;
4760
4761 dev_priv->rps.enabled = false;
4762 intel_enable_gt_powersave(dev);
4763 }
4764
4765 static void ibx_init_clock_gating(struct drm_device *dev)
4766 {
4767 struct drm_i915_private *dev_priv = dev->dev_private;
4768
4769 /*
4770 * On Ibex Peak and Cougar Point, we need to disable clock
4771 * gating for the panel power sequencer or it will fail to
4772 * start up when no ports are active.
4773 */
4774 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
4775 }
4776
4777 static void g4x_disable_trickle_feed(struct drm_device *dev)
4778 {
4779 struct drm_i915_private *dev_priv = dev->dev_private;
4780 int pipe;
4781
4782 for_each_pipe(pipe) {
4783 I915_WRITE(DSPCNTR(pipe),
4784 I915_READ(DSPCNTR(pipe)) |
4785 DISPPLANE_TRICKLE_FEED_DISABLE);
4786 intel_flush_primary_plane(dev_priv, pipe);
4787 }
4788 }
4789
4790 static void ilk_init_lp_watermarks(struct drm_device *dev)
4791 {
4792 struct drm_i915_private *dev_priv = dev->dev_private;
4793
4794 I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
4795 I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
4796 I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
4797
4798 /*
4799 * Don't touch WM1S_LP_EN here.
4800 * Doing so could cause underruns.
4801 */
4802 }
4803
4804 static void ironlake_init_clock_gating(struct drm_device *dev)
4805 {
4806 struct drm_i915_private *dev_priv = dev->dev_private;
4807 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
4808
4809 /*
4810 * Required for FBC
4811 * WaFbcDisableDpfcClockGating:ilk
4812 */
4813 dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
4814 ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
4815 ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
4816
4817 I915_WRITE(PCH_3DCGDIS0,
4818 MARIUNIT_CLOCK_GATE_DISABLE |
4819 SVSMUNIT_CLOCK_GATE_DISABLE);
4820 I915_WRITE(PCH_3DCGDIS1,
4821 VFMUNIT_CLOCK_GATE_DISABLE);
4822
4823 /*
4824 * According to the spec the following bits should be set in
4825 * order to enable memory self-refresh
4826 * The bit 22/21 of 0x42004
4827 * The bit 5 of 0x42020
4828 * The bit 15 of 0x45000
4829 */
4830 I915_WRITE(ILK_DISPLAY_CHICKEN2,
4831 (I915_READ(ILK_DISPLAY_CHICKEN2) |
4832 ILK_DPARB_GATE | ILK_VSDPFD_FULL));
4833 dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
4834 I915_WRITE(DISP_ARB_CTL,
4835 (I915_READ(DISP_ARB_CTL) |
4836 DISP_FBC_WM_DIS));
4837
4838 ilk_init_lp_watermarks(dev);
4839
4840 /*
4841 * Based on the document from hardware guys the following bits
4842 * should be set unconditionally in order to enable FBC.
4843 * The bit 22 of 0x42000
4844 * The bit 22 of 0x42004
4845 * The bit 7,8,9 of 0x42020.
4846 */
4847 if (IS_IRONLAKE_M(dev)) {
4848 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
4849 I915_WRITE(ILK_DISPLAY_CHICKEN1,
4850 I915_READ(ILK_DISPLAY_CHICKEN1) |
4851 ILK_FBCQ_DIS);
4852 I915_WRITE(ILK_DISPLAY_CHICKEN2,
4853 I915_READ(ILK_DISPLAY_CHICKEN2) |
4854 ILK_DPARB_GATE);
4855 }
4856
4857 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
4858
4859 I915_WRITE(ILK_DISPLAY_CHICKEN2,
4860 I915_READ(ILK_DISPLAY_CHICKEN2) |
4861 ILK_ELPIN_409_SELECT);
4862 I915_WRITE(_3D_CHICKEN2,
4863 _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
4864 _3D_CHICKEN2_WM_READ_PIPELINED);
4865
4866 /* WaDisableRenderCachePipelinedFlush:ilk */
4867 I915_WRITE(CACHE_MODE_0,
4868 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
4869
4870 /* WaDisable_RenderCache_OperationalFlush:ilk */
4871 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
4872
4873 g4x_disable_trickle_feed(dev);
4874
4875 ibx_init_clock_gating(dev);
4876 }
4877
4878 static void cpt_init_clock_gating(struct drm_device *dev)
4879 {
4880 struct drm_i915_private *dev_priv = dev->dev_private;
4881 int pipe;
4882 uint32_t val;
4883
4884 /*
4885 * On Ibex Peak and Cougar Point, we need to disable clock
4886 * gating for the panel power sequencer or it will fail to
4887 * start up when no ports are active.
4888 */
4889 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
4890 PCH_DPLUNIT_CLOCK_GATE_DISABLE |
4891 PCH_CPUNIT_CLOCK_GATE_DISABLE);
4892 I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
4893 DPLS_EDP_PPS_FIX_DIS);
4894 /* The below fixes the weird display corruption, a few pixels shifted
4895 * downward, on (only) LVDS of some HP laptops with IVY.
4896 */
4897 for_each_pipe(pipe) {
4898 val = I915_READ(TRANS_CHICKEN2(pipe));
4899 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
4900 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
4901 if (dev_priv->vbt.fdi_rx_polarity_inverted)
4902 val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
4903 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
4904 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
4905 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
4906 I915_WRITE(TRANS_CHICKEN2(pipe), val);
4907 }
4908 /* WADP0ClockGatingDisable */
4909 for_each_pipe(pipe) {
4910 I915_WRITE(TRANS_CHICKEN1(pipe),
4911 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
4912 }
4913 }
4914
4915 static void gen6_check_mch_setup(struct drm_device *dev)
4916 {
4917 struct drm_i915_private *dev_priv = dev->dev_private;
4918 uint32_t tmp;
4919
4920 tmp = I915_READ(MCH_SSKPD);
4921 if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL) {
4922 DRM_INFO("Wrong MCH_SSKPD value: 0x%08x\n", tmp);
4923 DRM_INFO("This can cause pipe underruns and display issues.\n");
4924 DRM_INFO("Please upgrade your BIOS to fix this.\n");
4925 }
4926 }
4927
4928 static void gen6_init_clock_gating(struct drm_device *dev)
4929 {
4930 struct drm_i915_private *dev_priv = dev->dev_private;
4931 uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
4932
4933 I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
4934
4935 I915_WRITE(ILK_DISPLAY_CHICKEN2,
4936 I915_READ(ILK_DISPLAY_CHICKEN2) |
4937 ILK_ELPIN_409_SELECT);
4938
4939 /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
4940 I915_WRITE(_3D_CHICKEN,
4941 _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
4942
4943 /* WaSetupGtModeTdRowDispatch:snb */
4944 if (IS_SNB_GT1(dev))
4945 I915_WRITE(GEN6_GT_MODE,
4946 _MASKED_BIT_ENABLE(GEN6_TD_FOUR_ROW_DISPATCH_DISABLE));
4947
4948 /* WaDisable_RenderCache_OperationalFlush:snb */
4949 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
4950
4951 /*
4952 * BSpec recoomends 8x4 when MSAA is used,
4953 * however in practice 16x4 seems fastest.
4954 *
4955 * Note that PS/WM thread counts depend on the WIZ hashing
4956 * disable bit, which we don't touch here, but it's good
4957 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
4958 */
4959 I915_WRITE(GEN6_GT_MODE,
4960 GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
4961
4962 ilk_init_lp_watermarks(dev);
4963
4964 I915_WRITE(CACHE_MODE_0,
4965 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
4966
4967 I915_WRITE(GEN6_UCGCTL1,
4968 I915_READ(GEN6_UCGCTL1) |
4969 GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
4970 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
4971
4972 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
4973 * gating disable must be set. Failure to set it results in
4974 * flickering pixels due to Z write ordering failures after
4975 * some amount of runtime in the Mesa "fire" demo, and Unigine
4976 * Sanctuary and Tropics, and apparently anything else with
4977 * alpha test or pixel discard.
4978 *
4979 * According to the spec, bit 11 (RCCUNIT) must also be set,
4980 * but we didn't debug actual testcases to find it out.
4981 *
4982 * WaDisableRCCUnitClockGating:snb
4983 * WaDisableRCPBUnitClockGating:snb
4984 */
4985 I915_WRITE(GEN6_UCGCTL2,
4986 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
4987 GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
4988
4989 /* WaStripsFansDisableFastClipPerformanceFix:snb */
4990 I915_WRITE(_3D_CHICKEN3,
4991 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
4992
4993 /*
4994 * Bspec says:
4995 * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
4996 * 3DSTATE_SF number of SF output attributes is more than 16."
4997 */
4998 I915_WRITE(_3D_CHICKEN3,
4999 _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
5000
5001 /*
5002 * According to the spec the following bits should be
5003 * set in order to enable memory self-refresh and fbc:
5004 * The bit21 and bit22 of 0x42000
5005 * The bit21 and bit22 of 0x42004
5006 * The bit5 and bit7 of 0x42020
5007 * The bit14 of 0x70180
5008 * The bit14 of 0x71180
5009 *
5010 * WaFbcAsynchFlipDisableFbcQueue:snb
5011 */
5012 I915_WRITE(ILK_DISPLAY_CHICKEN1,
5013 I915_READ(ILK_DISPLAY_CHICKEN1) |
5014 ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
5015 I915_WRITE(ILK_DISPLAY_CHICKEN2,
5016 I915_READ(ILK_DISPLAY_CHICKEN2) |
5017 ILK_DPARB_GATE | ILK_VSDPFD_FULL);
5018 I915_WRITE(ILK_DSPCLK_GATE_D,
5019 I915_READ(ILK_DSPCLK_GATE_D) |
5020 ILK_DPARBUNIT_CLOCK_GATE_ENABLE |
5021 ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
5022
5023 g4x_disable_trickle_feed(dev);
5024
5025 cpt_init_clock_gating(dev);
5026
5027 gen6_check_mch_setup(dev);
5028 }
5029
5030 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
5031 {
5032 uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
5033
5034 /*
5035 * WaVSThreadDispatchOverride:ivb,vlv
5036 *
5037 * This actually overrides the dispatch
5038 * mode for all thread types.
5039 */
5040 reg &= ~GEN7_FF_SCHED_MASK;
5041 reg |= GEN7_FF_TS_SCHED_HW;
5042 reg |= GEN7_FF_VS_SCHED_HW;
5043 reg |= GEN7_FF_DS_SCHED_HW;
5044
5045 I915_WRITE(GEN7_FF_THREAD_MODE, reg);
5046 }
5047
5048 static void lpt_init_clock_gating(struct drm_device *dev)
5049 {
5050 struct drm_i915_private *dev_priv = dev->dev_private;
5051
5052 /*
5053 * TODO: this bit should only be enabled when really needed, then
5054 * disabled when not needed anymore in order to save power.
5055 */
5056 if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE)
5057 I915_WRITE(SOUTH_DSPCLK_GATE_D,
5058 I915_READ(SOUTH_DSPCLK_GATE_D) |
5059 PCH_LP_PARTITION_LEVEL_DISABLE);
5060
5061 /* WADPOClockGatingDisable:hsw */
5062 I915_WRITE(_TRANSA_CHICKEN1,
5063 I915_READ(_TRANSA_CHICKEN1) |
5064 TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
5065 }
5066
5067 static void lpt_suspend_hw(struct drm_device *dev)
5068 {
5069 struct drm_i915_private *dev_priv = dev->dev_private;
5070
5071 if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE) {
5072 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
5073
5074 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
5075 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
5076 }
5077 }
5078
5079 static void gen8_init_clock_gating(struct drm_device *dev)
5080 {
5081 struct drm_i915_private *dev_priv = dev->dev_private;
5082 enum pipe pipe;
5083
5084 I915_WRITE(WM3_LP_ILK, 0);
5085 I915_WRITE(WM2_LP_ILK, 0);
5086 I915_WRITE(WM1_LP_ILK, 0);
5087
5088 /* FIXME(BDW): Check all the w/a, some might only apply to
5089 * pre-production hw. */
5090
5091 /* WaDisablePartialInstShootdown:bdw */
5092 I915_WRITE(GEN8_ROW_CHICKEN,
5093 _MASKED_BIT_ENABLE(PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE));
5094
5095 /* WaDisableThreadStallDopClockGating:bdw */
5096 /* FIXME: Unclear whether we really need this on production bdw. */
5097 I915_WRITE(GEN8_ROW_CHICKEN,
5098 _MASKED_BIT_ENABLE(STALL_DOP_GATING_DISABLE));
5099
5100 /*
5101 * This GEN8_CENTROID_PIXEL_OPT_DIS W/A is only needed for
5102 * pre-production hardware
5103 */
5104 I915_WRITE(HALF_SLICE_CHICKEN3,
5105 _MASKED_BIT_ENABLE(GEN8_CENTROID_PIXEL_OPT_DIS));
5106 I915_WRITE(HALF_SLICE_CHICKEN3,
5107 _MASKED_BIT_ENABLE(GEN8_SAMPLER_POWER_BYPASS_DIS));
5108 I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_BWGTLB_DISABLE));
5109
5110 I915_WRITE(_3D_CHICKEN3,
5111 _3D_CHICKEN_SDE_LIMIT_FIFO_POLY_DEPTH(2));
5112
5113 I915_WRITE(COMMON_SLICE_CHICKEN2,
5114 _MASKED_BIT_ENABLE(GEN8_CSC2_SBE_VUE_CACHE_CONSERVATIVE));
5115
5116 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
5117 _MASKED_BIT_ENABLE(GEN7_SINGLE_SUBSCAN_DISPATCH_ENABLE));
5118
5119 /* WaDisableDopClockGating:bdw May not be needed for production */
5120 I915_WRITE(GEN7_ROW_CHICKEN2,
5121 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5122
5123 /* WaSwitchSolVfFArbitrationPriority:bdw */
5124 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
5125
5126 /* WaPsrDPAMaskVBlankInSRD:bdw */
5127 I915_WRITE(CHICKEN_PAR1_1,
5128 I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
5129
5130 /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
5131 for_each_pipe(pipe) {
5132 I915_WRITE(CHICKEN_PIPESL_1(pipe),
5133 I915_READ(CHICKEN_PIPESL_1(pipe)) |
5134 BDW_DPRS_MASK_VBLANK_SRD);
5135 }
5136
5137 /* Use Force Non-Coherent whenever executing a 3D context. This is a
5138 * workaround for for a possible hang in the unlikely event a TLB
5139 * invalidation occurs during a PSD flush.
5140 */
5141 I915_WRITE(HDC_CHICKEN0,
5142 I915_READ(HDC_CHICKEN0) |
5143 _MASKED_BIT_ENABLE(HDC_FORCE_NON_COHERENT));
5144
5145 /* WaVSRefCountFullforceMissDisable:bdw */
5146 /* WaDSRefCountFullforceMissDisable:bdw */
5147 I915_WRITE(GEN7_FF_THREAD_MODE,
5148 I915_READ(GEN7_FF_THREAD_MODE) &
5149 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
5150
5151 /*
5152 * BSpec recommends 8x4 when MSAA is used,
5153 * however in practice 16x4 seems fastest.
5154 *
5155 * Note that PS/WM thread counts depend on the WIZ hashing
5156 * disable bit, which we don't touch here, but it's good
5157 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
5158 */
5159 I915_WRITE(GEN7_GT_MODE,
5160 GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
5161
5162 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
5163 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
5164
5165 /* WaDisableSDEUnitClockGating:bdw */
5166 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
5167 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
5168
5169 /* Wa4x4STCOptimizationDisable:bdw */
5170 I915_WRITE(CACHE_MODE_1,
5171 _MASKED_BIT_ENABLE(GEN8_4x4_STC_OPTIMIZATION_DISABLE));
5172 }
5173
5174 static void haswell_init_clock_gating(struct drm_device *dev)
5175 {
5176 struct drm_i915_private *dev_priv = dev->dev_private;
5177
5178 ilk_init_lp_watermarks(dev);
5179
5180 /* L3 caching of data atomics doesn't work -- disable it. */
5181 I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
5182 I915_WRITE(HSW_ROW_CHICKEN3,
5183 _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
5184
5185 /* This is required by WaCatErrorRejectionIssue:hsw */
5186 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5187 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5188 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5189
5190 /* WaVSRefCountFullforceMissDisable:hsw */
5191 I915_WRITE(GEN7_FF_THREAD_MODE,
5192 I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
5193
5194 /* WaDisable_RenderCache_OperationalFlush:hsw */
5195 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5196
5197 /* enable HiZ Raw Stall Optimization */
5198 I915_WRITE(CACHE_MODE_0_GEN7,
5199 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
5200
5201 /* WaDisable4x2SubspanOptimization:hsw */
5202 I915_WRITE(CACHE_MODE_1,
5203 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5204
5205 /*
5206 * BSpec recommends 8x4 when MSAA is used,
5207 * however in practice 16x4 seems fastest.
5208 *
5209 * Note that PS/WM thread counts depend on the WIZ hashing
5210 * disable bit, which we don't touch here, but it's good
5211 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
5212 */
5213 I915_WRITE(GEN7_GT_MODE,
5214 GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
5215
5216 /* WaSwitchSolVfFArbitrationPriority:hsw */
5217 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
5218
5219 /* WaRsPkgCStateDisplayPMReq:hsw */
5220 I915_WRITE(CHICKEN_PAR1_1,
5221 I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
5222
5223 lpt_init_clock_gating(dev);
5224 }
5225
5226 static void ivybridge_init_clock_gating(struct drm_device *dev)
5227 {
5228 struct drm_i915_private *dev_priv = dev->dev_private;
5229 uint32_t snpcr;
5230
5231 ilk_init_lp_watermarks(dev);
5232
5233 I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
5234
5235 /* WaDisableEarlyCull:ivb */
5236 I915_WRITE(_3D_CHICKEN3,
5237 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
5238
5239 /* WaDisableBackToBackFlipFix:ivb */
5240 I915_WRITE(IVB_CHICKEN3,
5241 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
5242 CHICKEN3_DGMG_DONE_FIX_DISABLE);
5243
5244 /* WaDisablePSDDualDispatchEnable:ivb */
5245 if (IS_IVB_GT1(dev))
5246 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
5247 _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
5248
5249 /* WaDisable_RenderCache_OperationalFlush:ivb */
5250 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5251
5252 /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
5253 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
5254 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
5255
5256 /* WaApplyL3ControlAndL3ChickenMode:ivb */
5257 I915_WRITE(GEN7_L3CNTLREG1,
5258 GEN7_WA_FOR_GEN7_L3_CONTROL);
5259 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
5260 GEN7_WA_L3_CHICKEN_MODE);
5261 if (IS_IVB_GT1(dev))
5262 I915_WRITE(GEN7_ROW_CHICKEN2,
5263 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5264 else {
5265 /* must write both registers */
5266 I915_WRITE(GEN7_ROW_CHICKEN2,
5267 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5268 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
5269 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5270 }
5271
5272 /* WaForceL3Serialization:ivb */
5273 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
5274 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
5275
5276 /*
5277 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
5278 * This implements the WaDisableRCZUnitClockGating:ivb workaround.
5279 */
5280 I915_WRITE(GEN6_UCGCTL2,
5281 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
5282
5283 /* This is required by WaCatErrorRejectionIssue:ivb */
5284 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5285 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5286 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5287
5288 g4x_disable_trickle_feed(dev);
5289
5290 gen7_setup_fixed_func_scheduler(dev_priv);
5291
5292 if (0) { /* causes HiZ corruption on ivb:gt1 */
5293 /* enable HiZ Raw Stall Optimization */
5294 I915_WRITE(CACHE_MODE_0_GEN7,
5295 _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
5296 }
5297
5298 /* WaDisable4x2SubspanOptimization:ivb */
5299 I915_WRITE(CACHE_MODE_1,
5300 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5301
5302 /*
5303 * BSpec recommends 8x4 when MSAA is used,
5304 * however in practice 16x4 seems fastest.
5305 *
5306 * Note that PS/WM thread counts depend on the WIZ hashing
5307 * disable bit, which we don't touch here, but it's good
5308 * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
5309 */
5310 I915_WRITE(GEN7_GT_MODE,
5311 GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
5312
5313 snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
5314 snpcr &= ~GEN6_MBC_SNPCR_MASK;
5315 snpcr |= GEN6_MBC_SNPCR_MED;
5316 I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
5317
5318 if (!HAS_PCH_NOP(dev))
5319 cpt_init_clock_gating(dev);
5320
5321 gen6_check_mch_setup(dev);
5322 }
5323
5324 static void valleyview_init_clock_gating(struct drm_device *dev)
5325 {
5326 struct drm_i915_private *dev_priv = dev->dev_private;
5327 u32 val;
5328
5329 mutex_lock(&dev_priv->rps.hw_lock);
5330 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5331 mutex_unlock(&dev_priv->rps.hw_lock);
5332 switch ((val >> 6) & 3) {
5333 case 0:
5334 case 1:
5335 dev_priv->mem_freq = 800;
5336 break;
5337 case 2:
5338 dev_priv->mem_freq = 1066;
5339 break;
5340 case 3:
5341 dev_priv->mem_freq = 1333;
5342 break;
5343 }
5344 DRM_DEBUG_DRIVER("DDR speed: %d MHz", dev_priv->mem_freq);
5345
5346 dev_priv->vlv_cdclk_freq = valleyview_cur_cdclk(dev_priv);
5347 DRM_DEBUG_DRIVER("Current CD clock rate: %d MHz",
5348 dev_priv->vlv_cdclk_freq);
5349
5350 I915_WRITE(DSPCLK_GATE_D, VRHUNIT_CLOCK_GATE_DISABLE);
5351
5352 /* WaDisableEarlyCull:vlv */
5353 I915_WRITE(_3D_CHICKEN3,
5354 _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
5355
5356 /* WaDisableBackToBackFlipFix:vlv */
5357 I915_WRITE(IVB_CHICKEN3,
5358 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
5359 CHICKEN3_DGMG_DONE_FIX_DISABLE);
5360
5361 /* WaPsdDispatchEnable:vlv */
5362 /* WaDisablePSDDualDispatchEnable:vlv */
5363 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
5364 _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
5365 GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
5366
5367 /* WaDisable_RenderCache_OperationalFlush:vlv */
5368 I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5369
5370 /* WaForceL3Serialization:vlv */
5371 I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
5372 ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
5373
5374 /* WaDisableDopClockGating:vlv */
5375 I915_WRITE(GEN7_ROW_CHICKEN2,
5376 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5377
5378 /* This is required by WaCatErrorRejectionIssue:vlv */
5379 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5380 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5381 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5382
5383 gen7_setup_fixed_func_scheduler(dev_priv);
5384
5385 /*
5386 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
5387 * This implements the WaDisableRCZUnitClockGating:vlv workaround.
5388 */
5389 I915_WRITE(GEN6_UCGCTL2,
5390 GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
5391
5392 /* WaDisableL3Bank2xClockGate:vlv
5393 * Disabling L3 clock gating- MMIO 940c[25] = 1
5394 * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
5395 I915_WRITE(GEN7_UCGCTL4,
5396 I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
5397
5398 I915_WRITE(MI_ARB_VLV, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE);
5399
5400 /*
5401 * BSpec says this must be set, even though
5402 * WaDisable4x2SubspanOptimization isn't listed for VLV.
5403 */
5404 I915_WRITE(CACHE_MODE_1,
5405 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5406
5407 /*
5408 * WaIncreaseL3CreditsForVLVB0:vlv
5409 * This is the hardware default actually.
5410 */
5411 I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
5412
5413 /*
5414 * WaDisableVLVClockGating_VBIIssue:vlv
5415 * Disable clock gating on th GCFG unit to prevent a delay
5416 * in the reporting of vblank events.
5417 */
5418 I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
5419 }
5420
5421 static void cherryview_init_clock_gating(struct drm_device *dev)
5422 {
5423 struct drm_i915_private *dev_priv = dev->dev_private;
5424
5425 I915_WRITE(DSPCLK_GATE_D, VRHUNIT_CLOCK_GATE_DISABLE);
5426
5427 I915_WRITE(MI_ARB_VLV, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE);
5428
5429 /* WaDisablePartialInstShootdown:chv */
5430 I915_WRITE(GEN8_ROW_CHICKEN,
5431 _MASKED_BIT_ENABLE(PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE));
5432
5433 /* WaDisableThreadStallDopClockGating:chv */
5434 I915_WRITE(GEN8_ROW_CHICKEN,
5435 _MASKED_BIT_ENABLE(STALL_DOP_GATING_DISABLE));
5436
5437 /* WaVSRefCountFullforceMissDisable:chv */
5438 /* WaDSRefCountFullforceMissDisable:chv */
5439 I915_WRITE(GEN7_FF_THREAD_MODE,
5440 I915_READ(GEN7_FF_THREAD_MODE) &
5441 ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
5442
5443 /* WaDisableSemaphoreAndSyncFlipWait:chv */
5444 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
5445 _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
5446
5447 /* WaDisableCSUnitClockGating:chv */
5448 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
5449 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
5450
5451 /* WaDisableSDEUnitClockGating:chv */
5452 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
5453 GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
5454
5455 /* WaDisableSamplerPowerBypass:chv (pre-production hw) */
5456 I915_WRITE(HALF_SLICE_CHICKEN3,
5457 _MASKED_BIT_ENABLE(GEN8_SAMPLER_POWER_BYPASS_DIS));
5458
5459 /* WaDisableGunitClockGating:chv (pre-production hw) */
5460 I915_WRITE(VLV_GUNIT_CLOCK_GATE, I915_READ(VLV_GUNIT_CLOCK_GATE) |
5461 GINT_DIS);
5462
5463 /* WaDisableFfDopClockGating:chv (pre-production hw) */
5464 I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
5465 _MASKED_BIT_ENABLE(GEN8_FF_DOP_CLOCK_GATE_DISABLE));
5466
5467 /* WaDisableDopClockGating:chv (pre-production hw) */
5468 I915_WRITE(GEN7_ROW_CHICKEN2,
5469 _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5470 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
5471 GEN6_EU_TCUNIT_CLOCK_GATE_DISABLE);
5472 }
5473
5474 static void g4x_init_clock_gating(struct drm_device *dev)
5475 {
5476 struct drm_i915_private *dev_priv = dev->dev_private;
5477 uint32_t dspclk_gate;
5478
5479 I915_WRITE(RENCLK_GATE_D1, 0);
5480 I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
5481 GS_UNIT_CLOCK_GATE_DISABLE |
5482 CL_UNIT_CLOCK_GATE_DISABLE);
5483 I915_WRITE(RAMCLK_GATE_D, 0);
5484 dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
5485 OVRUNIT_CLOCK_GATE_DISABLE |
5486 OVCUNIT_CLOCK_GATE_DISABLE;
5487 if (IS_GM45(dev))
5488 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
5489 I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
5490
5491 /* WaDisableRenderCachePipelinedFlush */
5492 I915_WRITE(CACHE_MODE_0,
5493 _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
5494
5495 /* WaDisable_RenderCache_OperationalFlush:g4x */
5496 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5497
5498 g4x_disable_trickle_feed(dev);
5499 }
5500
5501 static void crestline_init_clock_gating(struct drm_device *dev)
5502 {
5503 struct drm_i915_private *dev_priv = dev->dev_private;
5504
5505 I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
5506 I915_WRITE(RENCLK_GATE_D2, 0);
5507 I915_WRITE(DSPCLK_GATE_D, 0);
5508 I915_WRITE(RAMCLK_GATE_D, 0);
5509 I915_WRITE16(DEUC, 0);
5510 I915_WRITE(MI_ARB_STATE,
5511 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
5512
5513 /* WaDisable_RenderCache_OperationalFlush:gen4 */
5514 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5515 }
5516
5517 static void broadwater_init_clock_gating(struct drm_device *dev)
5518 {
5519 struct drm_i915_private *dev_priv = dev->dev_private;
5520
5521 I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
5522 I965_RCC_CLOCK_GATE_DISABLE |
5523 I965_RCPB_CLOCK_GATE_DISABLE |
5524 I965_ISC_CLOCK_GATE_DISABLE |
5525 I965_FBC_CLOCK_GATE_DISABLE);
5526 I915_WRITE(RENCLK_GATE_D2, 0);
5527 I915_WRITE(MI_ARB_STATE,
5528 _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
5529
5530 /* WaDisable_RenderCache_OperationalFlush:gen4 */
5531 I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5532 }
5533
5534 static void gen3_init_clock_gating(struct drm_device *dev)
5535 {
5536 struct drm_i915_private *dev_priv = dev->dev_private;
5537 u32 dstate = I915_READ(D_STATE);
5538
5539 dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
5540 DSTATE_DOT_CLOCK_GATING;
5541 I915_WRITE(D_STATE, dstate);
5542
5543 if (IS_PINEVIEW(dev))
5544 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
5545
5546 /* IIR "flip pending" means done if this bit is set */
5547 I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
5548
5549 /* interrupts should cause a wake up from C3 */
5550 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
5551
5552 /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
5553 I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
5554 }
5555
5556 static void i85x_init_clock_gating(struct drm_device *dev)
5557 {
5558 struct drm_i915_private *dev_priv = dev->dev_private;
5559
5560 I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
5561
5562 /* interrupts should cause a wake up from C3 */
5563 I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
5564 _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
5565 }
5566
5567 static void i830_init_clock_gating(struct drm_device *dev)
5568 {
5569 struct drm_i915_private *dev_priv = dev->dev_private;
5570
5571 I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
5572 }
5573
5574 void intel_init_clock_gating(struct drm_device *dev)
5575 {
5576 struct drm_i915_private *dev_priv = dev->dev_private;
5577
5578 dev_priv->display.init_clock_gating(dev);
5579 }
5580
5581 void intel_suspend_hw(struct drm_device *dev)
5582 {
5583 if (HAS_PCH_LPT(dev))
5584 lpt_suspend_hw(dev);
5585 }
5586
5587 #define for_each_power_well(i, power_well, domain_mask, power_domains) \
5588 for (i = 0; \
5589 i < (power_domains)->power_well_count && \
5590 ((power_well) = &(power_domains)->power_wells[i]); \
5591 i++) \
5592 if ((power_well)->domains & (domain_mask))
5593
5594 #define for_each_power_well_rev(i, power_well, domain_mask, power_domains) \
5595 for (i = (power_domains)->power_well_count - 1; \
5596 i >= 0 && ((power_well) = &(power_domains)->power_wells[i]);\
5597 i--) \
5598 if ((power_well)->domains & (domain_mask))
5599
5600 /**
5601 * We should only use the power well if we explicitly asked the hardware to
5602 * enable it, so check if it's enabled and also check if we've requested it to
5603 * be enabled.
5604 */
5605 static bool hsw_power_well_enabled(struct drm_i915_private *dev_priv,
5606 struct i915_power_well *power_well)
5607 {
5608 return I915_READ(HSW_PWR_WELL_DRIVER) ==
5609 (HSW_PWR_WELL_ENABLE_REQUEST | HSW_PWR_WELL_STATE_ENABLED);
5610 }
5611
5612 bool intel_display_power_enabled_unlocked(struct drm_i915_private *dev_priv,
5613 enum intel_display_power_domain domain)
5614 {
5615 struct i915_power_domains *power_domains;
5616 struct i915_power_well *power_well;
5617 bool is_enabled;
5618 int i;
5619
5620 if (dev_priv->pm.suspended)
5621 return false;
5622
5623 power_domains = &dev_priv->power_domains;
5624
5625 is_enabled = true;
5626
5627 for_each_power_well_rev(i, power_well, BIT(domain), power_domains) {
5628 if (power_well->always_on)
5629 continue;
5630
5631 if (!power_well->hw_enabled) {
5632 is_enabled = false;
5633 break;
5634 }
5635 }
5636
5637 return is_enabled;
5638 }
5639
5640 bool intel_display_power_enabled(struct drm_i915_private *dev_priv,
5641 enum intel_display_power_domain domain)
5642 {
5643 struct i915_power_domains *power_domains;
5644 bool ret;
5645
5646 power_domains = &dev_priv->power_domains;
5647
5648 mutex_lock(&power_domains->lock);
5649 ret = intel_display_power_enabled_unlocked(dev_priv, domain);
5650 mutex_unlock(&power_domains->lock);
5651
5652 return ret;
5653 }
5654
5655 /*
5656 * Starting with Haswell, we have a "Power Down Well" that can be turned off
5657 * when not needed anymore. We have 4 registers that can request the power well
5658 * to be enabled, and it will only be disabled if none of the registers is
5659 * requesting it to be enabled.
5660 */
5661 static void hsw_power_well_post_enable(struct drm_i915_private *dev_priv)
5662 {
5663 struct drm_device *dev = dev_priv->dev;
5664 unsigned long irqflags;
5665
5666 /*
5667 * After we re-enable the power well, if we touch VGA register 0x3d5
5668 * we'll get unclaimed register interrupts. This stops after we write
5669 * anything to the VGA MSR register. The vgacon module uses this
5670 * register all the time, so if we unbind our driver and, as a
5671 * consequence, bind vgacon, we'll get stuck in an infinite loop at
5672 * console_unlock(). So make here we touch the VGA MSR register, making
5673 * sure vgacon can keep working normally without triggering interrupts
5674 * and error messages.
5675 */
5676 vga_get_uninterruptible(dev->pdev, VGA_RSRC_LEGACY_IO);
5677 outb(inb(VGA_MSR_READ), VGA_MSR_WRITE);
5678 vga_put(dev->pdev, VGA_RSRC_LEGACY_IO);
5679
5680 if (IS_BROADWELL(dev)) {
5681 spin_lock_irqsave(&dev_priv->irq_lock, irqflags);
5682 I915_WRITE(GEN8_DE_PIPE_IMR(PIPE_B),
5683 dev_priv->de_irq_mask[PIPE_B]);
5684 I915_WRITE(GEN8_DE_PIPE_IER(PIPE_B),
5685 ~dev_priv->de_irq_mask[PIPE_B] |
5686 GEN8_PIPE_VBLANK);
5687 I915_WRITE(GEN8_DE_PIPE_IMR(PIPE_C),
5688 dev_priv->de_irq_mask[PIPE_C]);
5689 I915_WRITE(GEN8_DE_PIPE_IER(PIPE_C),
5690 ~dev_priv->de_irq_mask[PIPE_C] |
5691 GEN8_PIPE_VBLANK);
5692 POSTING_READ(GEN8_DE_PIPE_IER(PIPE_C));
5693 spin_unlock_irqrestore(&dev_priv->irq_lock, irqflags);
5694 }
5695 }
5696
5697 static void hsw_set_power_well(struct drm_i915_private *dev_priv,
5698 struct i915_power_well *power_well, bool enable)
5699 {
5700 bool is_enabled, enable_requested;
5701 uint32_t tmp;
5702
5703 tmp = I915_READ(HSW_PWR_WELL_DRIVER);
5704 is_enabled = tmp & HSW_PWR_WELL_STATE_ENABLED;
5705 enable_requested = tmp & HSW_PWR_WELL_ENABLE_REQUEST;
5706
5707 if (enable) {
5708 if (!enable_requested)
5709 I915_WRITE(HSW_PWR_WELL_DRIVER,
5710 HSW_PWR_WELL_ENABLE_REQUEST);
5711
5712 if (!is_enabled) {
5713 DRM_DEBUG_KMS("Enabling power well\n");
5714 if (wait_for((I915_READ(HSW_PWR_WELL_DRIVER) &
5715 HSW_PWR_WELL_STATE_ENABLED), 20))
5716 DRM_ERROR("Timeout enabling power well\n");
5717 }
5718
5719 hsw_power_well_post_enable(dev_priv);
5720 } else {
5721 if (enable_requested) {
5722 I915_WRITE(HSW_PWR_WELL_DRIVER, 0);
5723 POSTING_READ(HSW_PWR_WELL_DRIVER);
5724 DRM_DEBUG_KMS("Requesting to disable the power well\n");
5725 }
5726 }
5727 }
5728
5729 static void hsw_power_well_sync_hw(struct drm_i915_private *dev_priv,
5730 struct i915_power_well *power_well)
5731 {
5732 hsw_set_power_well(dev_priv, power_well, power_well->count > 0);
5733
5734 /*
5735 * We're taking over the BIOS, so clear any requests made by it since
5736 * the driver is in charge now.
5737 */
5738 if (I915_READ(HSW_PWR_WELL_BIOS) & HSW_PWR_WELL_ENABLE_REQUEST)
5739 I915_WRITE(HSW_PWR_WELL_BIOS, 0);
5740 }
5741
5742 static void hsw_power_well_enable(struct drm_i915_private *dev_priv,
5743 struct i915_power_well *power_well)
5744 {
5745 hsw_set_power_well(dev_priv, power_well, true);
5746 }
5747
5748 static void hsw_power_well_disable(struct drm_i915_private *dev_priv,
5749 struct i915_power_well *power_well)
5750 {
5751 hsw_set_power_well(dev_priv, power_well, false);
5752 }
5753
5754 static void i9xx_always_on_power_well_noop(struct drm_i915_private *dev_priv,
5755 struct i915_power_well *power_well)
5756 {
5757 }
5758
5759 static bool i9xx_always_on_power_well_enabled(struct drm_i915_private *dev_priv,
5760 struct i915_power_well *power_well)
5761 {
5762 return true;
5763 }
5764
5765 void __vlv_set_power_well(struct drm_i915_private *dev_priv,
5766 enum punit_power_well power_well_id, bool enable)
5767 {
5768 struct drm_device *dev = dev_priv->dev;
5769 u32 mask;
5770 u32 state;
5771 u32 ctrl;
5772 enum pipe pipe;
5773
5774 if (power_well_id == PUNIT_POWER_WELL_DPIO_CMN_BC) {
5775 if (enable) {
5776 /*
5777 * Enable the CRI clock source so we can get at the
5778 * display and the reference clock for VGA
5779 * hotplug / manual detection.
5780 */
5781 I915_WRITE(DPLL(PIPE_B), I915_READ(DPLL(PIPE_B)) |
5782 DPLL_REFA_CLK_ENABLE_VLV |
5783 DPLL_INTEGRATED_CRI_CLK_VLV);
5784 udelay(1); /* >10ns for cmnreset, >0ns for sidereset */
5785 } else {
5786 for_each_pipe(pipe)
5787 assert_pll_disabled(dev_priv, pipe);
5788 /* Assert common reset */
5789 I915_WRITE(DPIO_CTL, I915_READ(DPIO_CTL) &
5790 ~DPIO_CMNRST);
5791 }
5792 }
5793
5794 mask = PUNIT_PWRGT_MASK(power_well_id);
5795 state = enable ? PUNIT_PWRGT_PWR_ON(power_well_id) :
5796 PUNIT_PWRGT_PWR_GATE(power_well_id);
5797
5798 mutex_lock(&dev_priv->rps.hw_lock);
5799
5800 #define COND \
5801 ((vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_STATUS) & mask) == state)
5802
5803 if (COND)
5804 goto out;
5805
5806 ctrl = vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_CTRL);
5807 ctrl &= ~mask;
5808 ctrl |= state;
5809 vlv_punit_write(dev_priv, PUNIT_REG_PWRGT_CTRL, ctrl);
5810
5811 if (wait_for(COND, 100))
5812 DRM_ERROR("timout setting power well state %08x (%08x)\n",
5813 state,
5814 vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_CTRL));
5815
5816 #undef COND
5817
5818 out:
5819 mutex_unlock(&dev_priv->rps.hw_lock);
5820
5821 /*
5822 * From VLV2A0_DP_eDP_DPIO_driver_vbios_notes_10.docx -
5823 * 6. De-assert cmn_reset/side_reset. Same as VLV X0.
5824 * a. GUnit 0x2110 bit[0] set to 1 (def 0)
5825 * b. The other bits such as sfr settings / modesel may all
5826 * be set to 0.
5827 *
5828 * This should only be done on init and resume from S3 with
5829 * both PLLs disabled, or we risk losing DPIO and PLL
5830 * synchronization.
5831 */
5832 if (power_well_id == PUNIT_POWER_WELL_DPIO_CMN_BC && enable)
5833 I915_WRITE(DPIO_CTL, I915_READ(DPIO_CTL) | DPIO_CMNRST);
5834 }
5835
5836 static void vlv_set_power_well(struct drm_i915_private *dev_priv,
5837 struct i915_power_well *power_well, bool enable)
5838 {
5839 enum punit_power_well power_well_id = power_well->data;
5840
5841 __vlv_set_power_well(dev_priv, power_well_id, enable);
5842 }
5843
5844 static void vlv_power_well_sync_hw(struct drm_i915_private *dev_priv,
5845 struct i915_power_well *power_well)
5846 {
5847 vlv_set_power_well(dev_priv, power_well, power_well->count > 0);
5848 }
5849
5850 static void vlv_power_well_enable(struct drm_i915_private *dev_priv,
5851 struct i915_power_well *power_well)
5852 {
5853 vlv_set_power_well(dev_priv, power_well, true);
5854 }
5855
5856 static void vlv_power_well_disable(struct drm_i915_private *dev_priv,
5857 struct i915_power_well *power_well)
5858 {
5859 vlv_set_power_well(dev_priv, power_well, false);
5860 }
5861
5862 static bool vlv_power_well_enabled(struct drm_i915_private *dev_priv,
5863 struct i915_power_well *power_well)
5864 {
5865 int power_well_id = power_well->data;
5866 bool enabled = false;
5867 u32 mask;
5868 u32 state;
5869 u32 ctrl;
5870
5871 mask = PUNIT_PWRGT_MASK(power_well_id);
5872 ctrl = PUNIT_PWRGT_PWR_ON(power_well_id);
5873
5874 mutex_lock(&dev_priv->rps.hw_lock);
5875
5876 state = vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_STATUS) & mask;
5877 /*
5878 * We only ever set the power-on and power-gate states, anything
5879 * else is unexpected.
5880 */
5881 WARN_ON(state != PUNIT_PWRGT_PWR_ON(power_well_id) &&
5882 state != PUNIT_PWRGT_PWR_GATE(power_well_id));
5883 if (state == ctrl)
5884 enabled = true;
5885
5886 /*
5887 * A transient state at this point would mean some unexpected party
5888 * is poking at the power controls too.
5889 */
5890 ctrl = vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_CTRL) & mask;
5891 WARN_ON(ctrl != state);
5892
5893 mutex_unlock(&dev_priv->rps.hw_lock);
5894
5895 return enabled;
5896 }
5897
5898 static void vlv_display_power_well_enable(struct drm_i915_private *dev_priv,
5899 struct i915_power_well *power_well)
5900 {
5901 WARN_ON_ONCE(power_well->data != PUNIT_POWER_WELL_DISP2D);
5902
5903 vlv_set_power_well(dev_priv, power_well, true);
5904
5905 spin_lock_irq(&dev_priv->irq_lock);
5906 valleyview_enable_display_irqs(dev_priv);
5907 spin_unlock_irq(&dev_priv->irq_lock);
5908
5909 /*
5910 * During driver initialization/resume we can avoid restoring the
5911 * part of the HW/SW state that will be inited anyway explicitly.
5912 */
5913 if (dev_priv->power_domains.initializing)
5914 return;
5915
5916 intel_hpd_init(dev_priv->dev);
5917
5918 i915_redisable_vga_power_on(dev_priv->dev);
5919 }
5920
5921 static void vlv_display_power_well_disable(struct drm_i915_private *dev_priv,
5922 struct i915_power_well *power_well)
5923 {
5924 WARN_ON_ONCE(power_well->data != PUNIT_POWER_WELL_DISP2D);
5925
5926 spin_lock_irq(&dev_priv->irq_lock);
5927 valleyview_disable_display_irqs(dev_priv);
5928 spin_unlock_irq(&dev_priv->irq_lock);
5929
5930 vlv_set_power_well(dev_priv, power_well, false);
5931 }
5932
5933 static void check_power_well_state(struct drm_i915_private *dev_priv,
5934 struct i915_power_well *power_well)
5935 {
5936 bool enabled = power_well->ops->is_enabled(dev_priv, power_well);
5937
5938 if (power_well->always_on || !i915.disable_power_well) {
5939 if (!enabled)
5940 goto mismatch;
5941
5942 return;
5943 }
5944
5945 if (enabled != (power_well->count > 0))
5946 goto mismatch;
5947
5948 return;
5949
5950 mismatch:
5951 WARN(1, "state mismatch for '%s' (always_on %d hw state %d use-count %d disable_power_well %d\n",
5952 power_well->name, power_well->always_on, enabled,
5953 power_well->count, i915.disable_power_well);
5954 }
5955
5956 void intel_display_power_get(struct drm_i915_private *dev_priv,
5957 enum intel_display_power_domain domain)
5958 {
5959 struct i915_power_domains *power_domains;
5960 struct i915_power_well *power_well;
5961 int i;
5962
5963 intel_runtime_pm_get(dev_priv);
5964
5965 power_domains = &dev_priv->power_domains;
5966
5967 mutex_lock(&power_domains->lock);
5968
5969 for_each_power_well(i, power_well, BIT(domain), power_domains) {
5970 if (!power_well->count++) {
5971 DRM_DEBUG_KMS("enabling %s\n", power_well->name);
5972 power_well->ops->enable(dev_priv, power_well);
5973 power_well->hw_enabled = true;
5974 }
5975
5976 check_power_well_state(dev_priv, power_well);
5977 }
5978
5979 power_domains->domain_use_count[domain]++;
5980
5981 mutex_unlock(&power_domains->lock);
5982 }
5983
5984 void intel_display_power_put(struct drm_i915_private *dev_priv,
5985 enum intel_display_power_domain domain)
5986 {
5987 struct i915_power_domains *power_domains;
5988 struct i915_power_well *power_well;
5989 int i;
5990
5991 power_domains = &dev_priv->power_domains;
5992
5993 mutex_lock(&power_domains->lock);
5994
5995 WARN_ON(!power_domains->domain_use_count[domain]);
5996 power_domains->domain_use_count[domain]--;
5997
5998 for_each_power_well_rev(i, power_well, BIT(domain), power_domains) {
5999 WARN_ON(!power_well->count);
6000
6001 if (!--power_well->count && i915.disable_power_well) {
6002 DRM_DEBUG_KMS("disabling %s\n", power_well->name);
6003 power_well->hw_enabled = false;
6004 power_well->ops->disable(dev_priv, power_well);
6005 }
6006
6007 check_power_well_state(dev_priv, power_well);
6008 }
6009
6010 mutex_unlock(&power_domains->lock);
6011
6012 intel_runtime_pm_put(dev_priv);
6013 }
6014
6015 static struct i915_power_domains *hsw_pwr;
6016
6017 /* Display audio driver power well request */
6018 int i915_request_power_well(void)
6019 {
6020 struct drm_i915_private *dev_priv;
6021
6022 if (!hsw_pwr)
6023 return -ENODEV;
6024
6025 dev_priv = container_of(hsw_pwr, struct drm_i915_private,
6026 power_domains);
6027 intel_display_power_get(dev_priv, POWER_DOMAIN_AUDIO);
6028 return 0;
6029 }
6030 EXPORT_SYMBOL_GPL(i915_request_power_well);
6031
6032 /* Display audio driver power well release */
6033 int i915_release_power_well(void)
6034 {
6035 struct drm_i915_private *dev_priv;
6036
6037 if (!hsw_pwr)
6038 return -ENODEV;
6039
6040 dev_priv = container_of(hsw_pwr, struct drm_i915_private,
6041 power_domains);
6042 intel_display_power_put(dev_priv, POWER_DOMAIN_AUDIO);
6043 return 0;
6044 }
6045 EXPORT_SYMBOL_GPL(i915_release_power_well);
6046
6047 /*
6048 * Private interface for the audio driver to get CDCLK in kHz.
6049 *
6050 * Caller must request power well using i915_request_power_well() prior to
6051 * making the call.
6052 */
6053 int i915_get_cdclk_freq(void)
6054 {
6055 struct drm_i915_private *dev_priv;
6056
6057 if (!hsw_pwr)
6058 return -ENODEV;
6059
6060 dev_priv = container_of(hsw_pwr, struct drm_i915_private,
6061 power_domains);
6062
6063 return intel_ddi_get_cdclk_freq(dev_priv);
6064 }
6065 EXPORT_SYMBOL_GPL(i915_get_cdclk_freq);
6066
6067
6068 #define POWER_DOMAIN_MASK (BIT(POWER_DOMAIN_NUM) - 1)
6069
6070 #define HSW_ALWAYS_ON_POWER_DOMAINS ( \
6071 BIT(POWER_DOMAIN_PIPE_A) | \
6072 BIT(POWER_DOMAIN_TRANSCODER_EDP) | \
6073 BIT(POWER_DOMAIN_PORT_DDI_A_2_LANES) | \
6074 BIT(POWER_DOMAIN_PORT_DDI_A_4_LANES) | \
6075 BIT(POWER_DOMAIN_PORT_DDI_B_2_LANES) | \
6076 BIT(POWER_DOMAIN_PORT_DDI_B_4_LANES) | \
6077 BIT(POWER_DOMAIN_PORT_DDI_C_2_LANES) | \
6078 BIT(POWER_DOMAIN_PORT_DDI_C_4_LANES) | \
6079 BIT(POWER_DOMAIN_PORT_DDI_D_2_LANES) | \
6080 BIT(POWER_DOMAIN_PORT_DDI_D_4_LANES) | \
6081 BIT(POWER_DOMAIN_PORT_CRT) | \
6082 BIT(POWER_DOMAIN_INIT))
6083 #define HSW_DISPLAY_POWER_DOMAINS ( \
6084 (POWER_DOMAIN_MASK & ~HSW_ALWAYS_ON_POWER_DOMAINS) | \
6085 BIT(POWER_DOMAIN_INIT))
6086
6087 #define BDW_ALWAYS_ON_POWER_DOMAINS ( \
6088 HSW_ALWAYS_ON_POWER_DOMAINS | \
6089 BIT(POWER_DOMAIN_PIPE_A_PANEL_FITTER))
6090 #define BDW_DISPLAY_POWER_DOMAINS ( \
6091 (POWER_DOMAIN_MASK & ~BDW_ALWAYS_ON_POWER_DOMAINS) | \
6092 BIT(POWER_DOMAIN_INIT))
6093
6094 #define VLV_ALWAYS_ON_POWER_DOMAINS BIT(POWER_DOMAIN_INIT)
6095 #define VLV_DISPLAY_POWER_DOMAINS POWER_DOMAIN_MASK
6096
6097 #define VLV_DPIO_CMN_BC_POWER_DOMAINS ( \
6098 BIT(POWER_DOMAIN_PORT_DDI_B_2_LANES) | \
6099 BIT(POWER_DOMAIN_PORT_DDI_B_4_LANES) | \
6100 BIT(POWER_DOMAIN_PORT_DDI_C_2_LANES) | \
6101 BIT(POWER_DOMAIN_PORT_DDI_C_4_LANES) | \
6102 BIT(POWER_DOMAIN_PORT_CRT) | \
6103 BIT(POWER_DOMAIN_INIT))
6104
6105 #define VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS ( \
6106 BIT(POWER_DOMAIN_PORT_DDI_B_2_LANES) | \
6107 BIT(POWER_DOMAIN_PORT_DDI_B_4_LANES) | \
6108 BIT(POWER_DOMAIN_INIT))
6109
6110 #define VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS ( \
6111 BIT(POWER_DOMAIN_PORT_DDI_B_4_LANES) | \
6112 BIT(POWER_DOMAIN_INIT))
6113
6114 #define VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS ( \
6115 BIT(POWER_DOMAIN_PORT_DDI_C_2_LANES) | \
6116 BIT(POWER_DOMAIN_PORT_DDI_C_4_LANES) | \
6117 BIT(POWER_DOMAIN_INIT))
6118
6119 #define VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS ( \
6120 BIT(POWER_DOMAIN_PORT_DDI_C_4_LANES) | \
6121 BIT(POWER_DOMAIN_INIT))
6122
6123 static const struct i915_power_well_ops i9xx_always_on_power_well_ops = {
6124 .sync_hw = i9xx_always_on_power_well_noop,
6125 .enable = i9xx_always_on_power_well_noop,
6126 .disable = i9xx_always_on_power_well_noop,
6127 .is_enabled = i9xx_always_on_power_well_enabled,
6128 };
6129
6130 static struct i915_power_well i9xx_always_on_power_well[] = {
6131 {
6132 .name = "always-on",
6133 .always_on = 1,
6134 .domains = POWER_DOMAIN_MASK,
6135 .ops = &i9xx_always_on_power_well_ops,
6136 },
6137 };
6138
6139 static const struct i915_power_well_ops hsw_power_well_ops = {
6140 .sync_hw = hsw_power_well_sync_hw,
6141 .enable = hsw_power_well_enable,
6142 .disable = hsw_power_well_disable,
6143 .is_enabled = hsw_power_well_enabled,
6144 };
6145
6146 static struct i915_power_well hsw_power_wells[] = {
6147 {
6148 .name = "always-on",
6149 .always_on = 1,
6150 .domains = HSW_ALWAYS_ON_POWER_DOMAINS,
6151 .ops = &i9xx_always_on_power_well_ops,
6152 },
6153 {
6154 .name = "display",
6155 .domains = HSW_DISPLAY_POWER_DOMAINS,
6156 .ops = &hsw_power_well_ops,
6157 },
6158 };
6159
6160 static struct i915_power_well bdw_power_wells[] = {
6161 {
6162 .name = "always-on",
6163 .always_on = 1,
6164 .domains = BDW_ALWAYS_ON_POWER_DOMAINS,
6165 .ops = &i9xx_always_on_power_well_ops,
6166 },
6167 {
6168 .name = "display",
6169 .domains = BDW_DISPLAY_POWER_DOMAINS,
6170 .ops = &hsw_power_well_ops,
6171 },
6172 };
6173
6174 static const struct i915_power_well_ops vlv_display_power_well_ops = {
6175 .sync_hw = vlv_power_well_sync_hw,
6176 .enable = vlv_display_power_well_enable,
6177 .disable = vlv_display_power_well_disable,
6178 .is_enabled = vlv_power_well_enabled,
6179 };
6180
6181 static const struct i915_power_well_ops vlv_dpio_power_well_ops = {
6182 .sync_hw = vlv_power_well_sync_hw,
6183 .enable = vlv_power_well_enable,
6184 .disable = vlv_power_well_disable,
6185 .is_enabled = vlv_power_well_enabled,
6186 };
6187
6188 static struct i915_power_well vlv_power_wells[] = {
6189 {
6190 .name = "always-on",
6191 .always_on = 1,
6192 .domains = VLV_ALWAYS_ON_POWER_DOMAINS,
6193 .ops = &i9xx_always_on_power_well_ops,
6194 },
6195 {
6196 .name = "display",
6197 .domains = VLV_DISPLAY_POWER_DOMAINS,
6198 .data = PUNIT_POWER_WELL_DISP2D,
6199 .ops = &vlv_display_power_well_ops,
6200 },
6201 {
6202 .name = "dpio-tx-b-01",
6203 .domains = VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS |
6204 VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS |
6205 VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS |
6206 VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS,
6207 .ops = &vlv_dpio_power_well_ops,
6208 .data = PUNIT_POWER_WELL_DPIO_TX_B_LANES_01,
6209 },
6210 {
6211 .name = "dpio-tx-b-23",
6212 .domains = VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS |
6213 VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS |
6214 VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS |
6215 VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS,
6216 .ops = &vlv_dpio_power_well_ops,
6217 .data = PUNIT_POWER_WELL_DPIO_TX_B_LANES_23,
6218 },
6219 {
6220 .name = "dpio-tx-c-01",
6221 .domains = VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS |
6222 VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS |
6223 VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS |
6224 VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS,
6225 .ops = &vlv_dpio_power_well_ops,
6226 .data = PUNIT_POWER_WELL_DPIO_TX_C_LANES_01,
6227 },
6228 {
6229 .name = "dpio-tx-c-23",
6230 .domains = VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS |
6231 VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS |
6232 VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS |
6233 VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS,
6234 .ops = &vlv_dpio_power_well_ops,
6235 .data = PUNIT_POWER_WELL_DPIO_TX_C_LANES_23,
6236 },
6237 {
6238 .name = "dpio-common",
6239 .domains = VLV_DPIO_CMN_BC_POWER_DOMAINS,
6240 .data = PUNIT_POWER_WELL_DPIO_CMN_BC,
6241 .ops = &vlv_dpio_power_well_ops,
6242 },
6243 };
6244
6245 #define set_power_wells(power_domains, __power_wells) ({ \
6246 (power_domains)->power_wells = (__power_wells); \
6247 (power_domains)->power_well_count = ARRAY_SIZE(__power_wells); \
6248 })
6249
6250 int intel_power_domains_init(struct drm_i915_private *dev_priv)
6251 {
6252 struct i915_power_domains *power_domains = &dev_priv->power_domains;
6253
6254 mutex_init(&power_domains->lock);
6255
6256 /*
6257 * The enabling order will be from lower to higher indexed wells,
6258 * the disabling order is reversed.
6259 */
6260 if (IS_HASWELL(dev_priv->dev)) {
6261 set_power_wells(power_domains, hsw_power_wells);
6262 hsw_pwr = power_domains;
6263 } else if (IS_BROADWELL(dev_priv->dev)) {
6264 set_power_wells(power_domains, bdw_power_wells);
6265 hsw_pwr = power_domains;
6266 } else if (IS_VALLEYVIEW(dev_priv->dev)) {
6267 set_power_wells(power_domains, vlv_power_wells);
6268 } else {
6269 set_power_wells(power_domains, i9xx_always_on_power_well);
6270 }
6271
6272 return 0;
6273 }
6274
6275 void intel_power_domains_remove(struct drm_i915_private *dev_priv)
6276 {
6277 hsw_pwr = NULL;
6278 }
6279
6280 static void intel_power_domains_resume(struct drm_i915_private *dev_priv)
6281 {
6282 struct i915_power_domains *power_domains = &dev_priv->power_domains;
6283 struct i915_power_well *power_well;
6284 int i;
6285
6286 mutex_lock(&power_domains->lock);
6287 for_each_power_well(i, power_well, POWER_DOMAIN_MASK, power_domains) {
6288 power_well->ops->sync_hw(dev_priv, power_well);
6289 power_well->hw_enabled = power_well->ops->is_enabled(dev_priv,
6290 power_well);
6291 }
6292 mutex_unlock(&power_domains->lock);
6293 }
6294
6295 void intel_power_domains_init_hw(struct drm_i915_private *dev_priv)
6296 {
6297 struct i915_power_domains *power_domains = &dev_priv->power_domains;
6298
6299 power_domains->initializing = true;
6300 /* For now, we need the power well to be always enabled. */
6301 intel_display_set_init_power(dev_priv, true);
6302 intel_power_domains_resume(dev_priv);
6303 power_domains->initializing = false;
6304 }
6305
6306 void intel_aux_display_runtime_get(struct drm_i915_private *dev_priv)
6307 {
6308 intel_runtime_pm_get(dev_priv);
6309 }
6310
6311 void intel_aux_display_runtime_put(struct drm_i915_private *dev_priv)
6312 {
6313 intel_runtime_pm_put(dev_priv);
6314 }
6315
6316 void intel_runtime_pm_get(struct drm_i915_private *dev_priv)
6317 {
6318 struct drm_device *dev = dev_priv->dev;
6319 struct device *device = &dev->pdev->dev;
6320
6321 if (!HAS_RUNTIME_PM(dev))
6322 return;
6323
6324 pm_runtime_get_sync(device);
6325 WARN(dev_priv->pm.suspended, "Device still suspended.\n");
6326 }
6327
6328 void intel_runtime_pm_get_noresume(struct drm_i915_private *dev_priv)
6329 {
6330 struct drm_device *dev = dev_priv->dev;
6331 struct device *device = &dev->pdev->dev;
6332
6333 if (!HAS_RUNTIME_PM(dev))
6334 return;
6335
6336 WARN(dev_priv->pm.suspended, "Getting nosync-ref while suspended.\n");
6337 pm_runtime_get_noresume(device);
6338 }
6339
6340 void intel_runtime_pm_put(struct drm_i915_private *dev_priv)
6341 {
6342 struct drm_device *dev = dev_priv->dev;
6343 struct device *device = &dev->pdev->dev;
6344
6345 if (!HAS_RUNTIME_PM(dev))
6346 return;
6347
6348 pm_runtime_mark_last_busy(device);
6349 pm_runtime_put_autosuspend(device);
6350 }
6351
6352 void intel_init_runtime_pm(struct drm_i915_private *dev_priv)
6353 {
6354 struct drm_device *dev = dev_priv->dev;
6355 struct device *device = &dev->pdev->dev;
6356
6357 if (!HAS_RUNTIME_PM(dev))
6358 return;
6359
6360 pm_runtime_set_active(device);
6361
6362 /*
6363 * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
6364 * requirement.
6365 */
6366 if (!intel_enable_rc6(dev)) {
6367 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
6368 return;
6369 }
6370
6371 pm_runtime_set_autosuspend_delay(device, 10000); /* 10s */
6372 pm_runtime_mark_last_busy(device);
6373 pm_runtime_use_autosuspend(device);
6374
6375 pm_runtime_put_autosuspend(device);
6376 }
6377
6378 void intel_fini_runtime_pm(struct drm_i915_private *dev_priv)
6379 {
6380 struct drm_device *dev = dev_priv->dev;
6381 struct device *device = &dev->pdev->dev;
6382
6383 if (!HAS_RUNTIME_PM(dev))
6384 return;
6385
6386 if (!intel_enable_rc6(dev))
6387 return;
6388
6389 /* Make sure we're not suspended first. */
6390 pm_runtime_get_sync(device);
6391 pm_runtime_disable(device);
6392 }
6393
6394 /* Set up chip specific power management-related functions */
6395 void intel_init_pm(struct drm_device *dev)
6396 {
6397 struct drm_i915_private *dev_priv = dev->dev_private;
6398
6399 if (HAS_FBC(dev)) {
6400 if (INTEL_INFO(dev)->gen >= 7) {
6401 dev_priv->display.fbc_enabled = ironlake_fbc_enabled;
6402 dev_priv->display.enable_fbc = gen7_enable_fbc;
6403 dev_priv->display.disable_fbc = ironlake_disable_fbc;
6404 } else if (INTEL_INFO(dev)->gen >= 5) {
6405 dev_priv->display.fbc_enabled = ironlake_fbc_enabled;
6406 dev_priv->display.enable_fbc = ironlake_enable_fbc;
6407 dev_priv->display.disable_fbc = ironlake_disable_fbc;
6408 } else if (IS_GM45(dev)) {
6409 dev_priv->display.fbc_enabled = g4x_fbc_enabled;
6410 dev_priv->display.enable_fbc = g4x_enable_fbc;
6411 dev_priv->display.disable_fbc = g4x_disable_fbc;
6412 } else {
6413 dev_priv->display.fbc_enabled = i8xx_fbc_enabled;
6414 dev_priv->display.enable_fbc = i8xx_enable_fbc;
6415 dev_priv->display.disable_fbc = i8xx_disable_fbc;
6416
6417 /* This value was pulled out of someone's hat */
6418 I915_WRITE(FBC_CONTROL, 500 << FBC_CTL_INTERVAL_SHIFT);
6419 }
6420 }
6421
6422 /* For cxsr */
6423 if (IS_PINEVIEW(dev))
6424 i915_pineview_get_mem_freq(dev);
6425 else if (IS_GEN5(dev))
6426 i915_ironlake_get_mem_freq(dev);
6427
6428 /* For FIFO watermark updates */
6429 if (HAS_PCH_SPLIT(dev)) {
6430 ilk_setup_wm_latency(dev);
6431
6432 if ((IS_GEN5(dev) && dev_priv->wm.pri_latency[1] &&
6433 dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
6434 (!IS_GEN5(dev) && dev_priv->wm.pri_latency[0] &&
6435 dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
6436 dev_priv->display.update_wm = ilk_update_wm;
6437 dev_priv->display.update_sprite_wm = ilk_update_sprite_wm;
6438 } else {
6439 DRM_DEBUG_KMS("Failed to read display plane latency. "
6440 "Disable CxSR\n");
6441 }
6442
6443 if (IS_GEN5(dev))
6444 dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
6445 else if (IS_GEN6(dev))
6446 dev_priv->display.init_clock_gating = gen6_init_clock_gating;
6447 else if (IS_IVYBRIDGE(dev))
6448 dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
6449 else if (IS_HASWELL(dev))
6450 dev_priv->display.init_clock_gating = haswell_init_clock_gating;
6451 else if (INTEL_INFO(dev)->gen == 8)
6452 dev_priv->display.init_clock_gating = gen8_init_clock_gating;
6453 } else if (IS_CHERRYVIEW(dev)) {
6454 dev_priv->display.update_wm = valleyview_update_wm;
6455 dev_priv->display.init_clock_gating =
6456 cherryview_init_clock_gating;
6457 } else if (IS_VALLEYVIEW(dev)) {
6458 dev_priv->display.update_wm = valleyview_update_wm;
6459 dev_priv->display.init_clock_gating =
6460 valleyview_init_clock_gating;
6461 } else if (IS_PINEVIEW(dev)) {
6462 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
6463 dev_priv->is_ddr3,
6464 dev_priv->fsb_freq,
6465 dev_priv->mem_freq)) {
6466 DRM_INFO("failed to find known CxSR latency "
6467 "(found ddr%s fsb freq %d, mem freq %d), "
6468 "disabling CxSR\n",
6469 (dev_priv->is_ddr3 == 1) ? "3" : "2",
6470 dev_priv->fsb_freq, dev_priv->mem_freq);
6471 /* Disable CxSR and never update its watermark again */
6472 pineview_disable_cxsr(dev);
6473 dev_priv->display.update_wm = NULL;
6474 } else
6475 dev_priv->display.update_wm = pineview_update_wm;
6476 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
6477 } else if (IS_G4X(dev)) {
6478 dev_priv->display.update_wm = g4x_update_wm;
6479 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
6480 } else if (IS_GEN4(dev)) {
6481 dev_priv->display.update_wm = i965_update_wm;
6482 if (IS_CRESTLINE(dev))
6483 dev_priv->display.init_clock_gating = crestline_init_clock_gating;
6484 else if (IS_BROADWATER(dev))
6485 dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
6486 } else if (IS_GEN3(dev)) {
6487 dev_priv->display.update_wm = i9xx_update_wm;
6488 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
6489 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
6490 } else if (IS_GEN2(dev)) {
6491 if (INTEL_INFO(dev)->num_pipes == 1) {
6492 dev_priv->display.update_wm = i845_update_wm;
6493 dev_priv->display.get_fifo_size = i845_get_fifo_size;
6494 } else {
6495 dev_priv->display.update_wm = i9xx_update_wm;
6496 dev_priv->display.get_fifo_size = i830_get_fifo_size;
6497 }
6498
6499 if (IS_I85X(dev) || IS_I865G(dev))
6500 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
6501 else
6502 dev_priv->display.init_clock_gating = i830_init_clock_gating;
6503 } else {
6504 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
6505 }
6506 }
6507
6508 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u8 mbox, u32 *val)
6509 {
6510 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6511
6512 if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
6513 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
6514 return -EAGAIN;
6515 }
6516
6517 I915_WRITE(GEN6_PCODE_DATA, *val);
6518 I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
6519
6520 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
6521 500)) {
6522 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
6523 return -ETIMEDOUT;
6524 }
6525
6526 *val = I915_READ(GEN6_PCODE_DATA);
6527 I915_WRITE(GEN6_PCODE_DATA, 0);
6528
6529 return 0;
6530 }
6531
6532 int sandybridge_pcode_write(struct drm_i915_private *dev_priv, u8 mbox, u32 val)
6533 {
6534 WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6535
6536 if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
6537 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
6538 return -EAGAIN;
6539 }
6540
6541 I915_WRITE(GEN6_PCODE_DATA, val);
6542 I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
6543
6544 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
6545 500)) {
6546 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
6547 return -ETIMEDOUT;
6548 }
6549
6550 I915_WRITE(GEN6_PCODE_DATA, 0);
6551
6552 return 0;
6553 }
6554
6555 int vlv_gpu_freq(struct drm_i915_private *dev_priv, int val)
6556 {
6557 int div;
6558
6559 /* 4 x czclk */
6560 switch (dev_priv->mem_freq) {
6561 case 800:
6562 div = 10;
6563 break;
6564 case 1066:
6565 div = 12;
6566 break;
6567 case 1333:
6568 div = 16;
6569 break;
6570 default:
6571 return -1;
6572 }
6573
6574 return DIV_ROUND_CLOSEST(dev_priv->mem_freq * (val + 6 - 0xbd), 4 * div);
6575 }
6576
6577 int vlv_freq_opcode(struct drm_i915_private *dev_priv, int val)
6578 {
6579 int mul;
6580
6581 /* 4 x czclk */
6582 switch (dev_priv->mem_freq) {
6583 case 800:
6584 mul = 10;
6585 break;
6586 case 1066:
6587 mul = 12;
6588 break;
6589 case 1333:
6590 mul = 16;
6591 break;
6592 default:
6593 return -1;
6594 }
6595
6596 return DIV_ROUND_CLOSEST(4 * mul * val, dev_priv->mem_freq) + 0xbd - 6;
6597 }
6598
6599 void intel_pm_setup(struct drm_device *dev)
6600 {
6601 struct drm_i915_private *dev_priv = dev->dev_private;
6602
6603 mutex_init(&dev_priv->rps.hw_lock);
6604
6605 INIT_DELAYED_WORK(&dev_priv->rps.delayed_resume_work,
6606 intel_gen6_powersave_work);
6607
6608 dev_priv->pm.suspended = false;
6609 dev_priv->pm.irqs_disabled = false;
6610 }
Linux kernel - Deliverables
This page took 0.196202 seconds and 6 git commands to generate.