2 * Copyright © 2012 Intel Corporation
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
24 * Eugeni Dodonov <eugeni.dodonov@intel.com>
28 #include <linux/cpufreq.h>
30 #include "intel_drv.h"
31 #include "../../../platform/x86/intel_ips.h"
32 #include <linux/module.h>
34 /* FBC, or Frame Buffer Compression, is a technique employed to compress the
35 * framebuffer contents in-memory, aiming at reducing the required bandwidth
36 * during in-memory transfers and, therefore, reduce the power packet.
38 * The benefits of FBC are mostly visible with solid backgrounds and
39 * variation-less patterns.
41 * FBC-related functionality can be enabled by the means of the
42 * i915.i915_enable_fbc parameter
45 static void i8xx_disable_fbc(struct drm_device *dev)
47 struct drm_i915_private *dev_priv = dev->dev_private;
50 /* Disable compression */
51 fbc_ctl = I915_READ(FBC_CONTROL);
52 if ((fbc_ctl & FBC_CTL_EN) == 0)
55 fbc_ctl &= ~FBC_CTL_EN;
56 I915_WRITE(FBC_CONTROL, fbc_ctl);
58 /* Wait for compressing bit to clear */
59 if (wait_for((I915_READ(FBC_STATUS) & FBC_STAT_COMPRESSING) == 0, 10)) {
60 DRM_DEBUG_KMS("FBC idle timed out\n");
64 DRM_DEBUG_KMS("disabled FBC\n");
67 static void i8xx_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
69 struct drm_device *dev = crtc->dev;
70 struct drm_i915_private *dev_priv = dev->dev_private;
71 struct drm_framebuffer *fb = crtc->fb;
72 struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
73 struct drm_i915_gem_object *obj = intel_fb->obj;
74 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
77 u32 fbc_ctl, fbc_ctl2;
79 cfb_pitch = dev_priv->cfb_size / FBC_LL_SIZE;
80 if (fb->pitches[0] < cfb_pitch)
81 cfb_pitch = fb->pitches[0];
83 /* FBC_CTL wants 64B units */
84 cfb_pitch = (cfb_pitch / 64) - 1;
85 plane = intel_crtc->plane == 0 ? FBC_CTL_PLANEA : FBC_CTL_PLANEB;
88 for (i = 0; i < (FBC_LL_SIZE / 32) + 1; i++)
89 I915_WRITE(FBC_TAG + (i * 4), 0);
92 fbc_ctl2 = FBC_CTL_FENCE_DBL | FBC_CTL_IDLE_IMM | FBC_CTL_CPU_FENCE;
94 I915_WRITE(FBC_CONTROL2, fbc_ctl2);
95 I915_WRITE(FBC_FENCE_OFF, crtc->y);
98 fbc_ctl = FBC_CTL_EN | FBC_CTL_PERIODIC;
100 fbc_ctl |= FBC_CTL_C3_IDLE; /* 945 needs special SR handling */
101 fbc_ctl |= (cfb_pitch & 0xff) << FBC_CTL_STRIDE_SHIFT;
102 fbc_ctl |= (interval & 0x2fff) << FBC_CTL_INTERVAL_SHIFT;
103 fbc_ctl |= obj->fence_reg;
104 I915_WRITE(FBC_CONTROL, fbc_ctl);
106 DRM_DEBUG_KMS("enabled FBC, pitch %d, yoff %d, plane %d, ",
107 cfb_pitch, crtc->y, intel_crtc->plane);
110 static bool i8xx_fbc_enabled(struct drm_device *dev)
112 struct drm_i915_private *dev_priv = dev->dev_private;
114 return I915_READ(FBC_CONTROL) & FBC_CTL_EN;
117 static void g4x_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
119 struct drm_device *dev = crtc->dev;
120 struct drm_i915_private *dev_priv = dev->dev_private;
121 struct drm_framebuffer *fb = crtc->fb;
122 struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
123 struct drm_i915_gem_object *obj = intel_fb->obj;
124 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
125 int plane = intel_crtc->plane == 0 ? DPFC_CTL_PLANEA : DPFC_CTL_PLANEB;
126 unsigned long stall_watermark = 200;
129 dpfc_ctl = plane | DPFC_SR_EN | DPFC_CTL_LIMIT_1X;
130 dpfc_ctl |= DPFC_CTL_FENCE_EN | obj->fence_reg;
131 I915_WRITE(DPFC_CHICKEN, DPFC_HT_MODIFY);
133 I915_WRITE(DPFC_RECOMP_CTL, DPFC_RECOMP_STALL_EN |
134 (stall_watermark << DPFC_RECOMP_STALL_WM_SHIFT) |
135 (interval << DPFC_RECOMP_TIMER_COUNT_SHIFT));
136 I915_WRITE(DPFC_FENCE_YOFF, crtc->y);
139 I915_WRITE(DPFC_CONTROL, I915_READ(DPFC_CONTROL) | DPFC_CTL_EN);
141 DRM_DEBUG_KMS("enabled fbc on plane %d\n", intel_crtc->plane);
144 static void g4x_disable_fbc(struct drm_device *dev)
146 struct drm_i915_private *dev_priv = dev->dev_private;
149 /* Disable compression */
150 dpfc_ctl = I915_READ(DPFC_CONTROL);
151 if (dpfc_ctl & DPFC_CTL_EN) {
152 dpfc_ctl &= ~DPFC_CTL_EN;
153 I915_WRITE(DPFC_CONTROL, dpfc_ctl);
155 DRM_DEBUG_KMS("disabled FBC\n");
159 static bool g4x_fbc_enabled(struct drm_device *dev)
161 struct drm_i915_private *dev_priv = dev->dev_private;
163 return I915_READ(DPFC_CONTROL) & DPFC_CTL_EN;
166 static void sandybridge_blit_fbc_update(struct drm_device *dev)
168 struct drm_i915_private *dev_priv = dev->dev_private;
171 /* Make sure blitter notifies FBC of writes */
172 gen6_gt_force_wake_get(dev_priv);
173 blt_ecoskpd = I915_READ(GEN6_BLITTER_ECOSKPD);
174 blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY <<
175 GEN6_BLITTER_LOCK_SHIFT;
176 I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
177 blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY;
178 I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
179 blt_ecoskpd &= ~(GEN6_BLITTER_FBC_NOTIFY <<
180 GEN6_BLITTER_LOCK_SHIFT);
181 I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
182 POSTING_READ(GEN6_BLITTER_ECOSKPD);
183 gen6_gt_force_wake_put(dev_priv);
186 static void ironlake_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
188 struct drm_device *dev = crtc->dev;
189 struct drm_i915_private *dev_priv = dev->dev_private;
190 struct drm_framebuffer *fb = crtc->fb;
191 struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
192 struct drm_i915_gem_object *obj = intel_fb->obj;
193 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
194 int plane = intel_crtc->plane == 0 ? DPFC_CTL_PLANEA : DPFC_CTL_PLANEB;
195 unsigned long stall_watermark = 200;
198 dpfc_ctl = I915_READ(ILK_DPFC_CONTROL);
199 dpfc_ctl &= DPFC_RESERVED;
200 dpfc_ctl |= (plane | DPFC_CTL_LIMIT_1X);
201 /* Set persistent mode for front-buffer rendering, ala X. */
202 dpfc_ctl |= DPFC_CTL_PERSISTENT_MODE;
203 dpfc_ctl |= (DPFC_CTL_FENCE_EN | obj->fence_reg);
204 I915_WRITE(ILK_DPFC_CHICKEN, DPFC_HT_MODIFY);
206 I915_WRITE(ILK_DPFC_RECOMP_CTL, DPFC_RECOMP_STALL_EN |
207 (stall_watermark << DPFC_RECOMP_STALL_WM_SHIFT) |
208 (interval << DPFC_RECOMP_TIMER_COUNT_SHIFT));
209 I915_WRITE(ILK_DPFC_FENCE_YOFF, crtc->y);
210 I915_WRITE(ILK_FBC_RT_BASE, obj->gtt_offset | ILK_FBC_RT_VALID);
212 I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
215 I915_WRITE(SNB_DPFC_CTL_SA,
216 SNB_CPU_FENCE_ENABLE | obj->fence_reg);
217 I915_WRITE(DPFC_CPU_FENCE_OFFSET, crtc->y);
218 sandybridge_blit_fbc_update(dev);
221 DRM_DEBUG_KMS("enabled fbc on plane %d\n", intel_crtc->plane);
224 static void ironlake_disable_fbc(struct drm_device *dev)
226 struct drm_i915_private *dev_priv = dev->dev_private;
229 /* Disable compression */
230 dpfc_ctl = I915_READ(ILK_DPFC_CONTROL);
231 if (dpfc_ctl & DPFC_CTL_EN) {
232 dpfc_ctl &= ~DPFC_CTL_EN;
233 I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl);
235 DRM_DEBUG_KMS("disabled FBC\n");
239 static bool ironlake_fbc_enabled(struct drm_device *dev)
241 struct drm_i915_private *dev_priv = dev->dev_private;
243 return I915_READ(ILK_DPFC_CONTROL) & DPFC_CTL_EN;
246 bool intel_fbc_enabled(struct drm_device *dev)
248 struct drm_i915_private *dev_priv = dev->dev_private;
250 if (!dev_priv->display.fbc_enabled)
253 return dev_priv->display.fbc_enabled(dev);
256 static void intel_fbc_work_fn(struct work_struct *__work)
258 struct intel_fbc_work *work =
259 container_of(to_delayed_work(__work),
260 struct intel_fbc_work, work);
261 struct drm_device *dev = work->crtc->dev;
262 struct drm_i915_private *dev_priv = dev->dev_private;
264 mutex_lock(&dev->struct_mutex);
265 if (work == dev_priv->fbc_work) {
266 /* Double check that we haven't switched fb without cancelling
269 if (work->crtc->fb == work->fb) {
270 dev_priv->display.enable_fbc(work->crtc,
273 dev_priv->cfb_plane = to_intel_crtc(work->crtc)->plane;
274 dev_priv->cfb_fb = work->crtc->fb->base.id;
275 dev_priv->cfb_y = work->crtc->y;
278 dev_priv->fbc_work = NULL;
280 mutex_unlock(&dev->struct_mutex);
285 static void intel_cancel_fbc_work(struct drm_i915_private *dev_priv)
287 if (dev_priv->fbc_work == NULL)
290 DRM_DEBUG_KMS("cancelling pending FBC enable\n");
292 /* Synchronisation is provided by struct_mutex and checking of
293 * dev_priv->fbc_work, so we can perform the cancellation
294 * entirely asynchronously.
296 if (cancel_delayed_work(&dev_priv->fbc_work->work))
297 /* tasklet was killed before being run, clean up */
298 kfree(dev_priv->fbc_work);
300 /* Mark the work as no longer wanted so that if it does
301 * wake-up (because the work was already running and waiting
302 * for our mutex), it will discover that is no longer
305 dev_priv->fbc_work = NULL;
308 void intel_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
310 struct intel_fbc_work *work;
311 struct drm_device *dev = crtc->dev;
312 struct drm_i915_private *dev_priv = dev->dev_private;
314 if (!dev_priv->display.enable_fbc)
317 intel_cancel_fbc_work(dev_priv);
319 work = kzalloc(sizeof *work, GFP_KERNEL);
321 dev_priv->display.enable_fbc(crtc, interval);
327 work->interval = interval;
328 INIT_DELAYED_WORK(&work->work, intel_fbc_work_fn);
330 dev_priv->fbc_work = work;
332 DRM_DEBUG_KMS("scheduling delayed FBC enable\n");
334 /* Delay the actual enabling to let pageflipping cease and the
335 * display to settle before starting the compression. Note that
336 * this delay also serves a second purpose: it allows for a
337 * vblank to pass after disabling the FBC before we attempt
338 * to modify the control registers.
340 * A more complicated solution would involve tracking vblanks
341 * following the termination of the page-flipping sequence
342 * and indeed performing the enable as a co-routine and not
343 * waiting synchronously upon the vblank.
345 schedule_delayed_work(&work->work, msecs_to_jiffies(50));
348 void intel_disable_fbc(struct drm_device *dev)
350 struct drm_i915_private *dev_priv = dev->dev_private;
352 intel_cancel_fbc_work(dev_priv);
354 if (!dev_priv->display.disable_fbc)
357 dev_priv->display.disable_fbc(dev);
358 dev_priv->cfb_plane = -1;
362 * intel_update_fbc - enable/disable FBC as needed
363 * @dev: the drm_device
365 * Set up the framebuffer compression hardware at mode set time. We
366 * enable it if possible:
367 * - plane A only (on pre-965)
368 * - no pixel mulitply/line duplication
369 * - no alpha buffer discard
371 * - framebuffer <= 2048 in width, 1536 in height
373 * We can't assume that any compression will take place (worst case),
374 * so the compressed buffer has to be the same size as the uncompressed
375 * one. It also must reside (along with the line length buffer) in
378 * We need to enable/disable FBC on a global basis.
380 void intel_update_fbc(struct drm_device *dev)
382 struct drm_i915_private *dev_priv = dev->dev_private;
383 struct drm_crtc *crtc = NULL, *tmp_crtc;
384 struct intel_crtc *intel_crtc;
385 struct drm_framebuffer *fb;
386 struct intel_framebuffer *intel_fb;
387 struct drm_i915_gem_object *obj;
393 if (!I915_HAS_FBC(dev))
397 * If FBC is already on, we just have to verify that we can
398 * keep it that way...
399 * Need to disable if:
400 * - more than one pipe is active
401 * - changing FBC params (stride, fence, mode)
402 * - new fb is too large to fit in compressed buffer
403 * - going to an unsupported config (interlace, pixel multiply, etc.)
405 list_for_each_entry(tmp_crtc, &dev->mode_config.crtc_list, head) {
406 if (tmp_crtc->enabled &&
407 !to_intel_crtc(tmp_crtc)->primary_disabled &&
410 DRM_DEBUG_KMS("more than one pipe active, disabling compression\n");
411 dev_priv->no_fbc_reason = FBC_MULTIPLE_PIPES;
418 if (!crtc || crtc->fb == NULL) {
419 DRM_DEBUG_KMS("no output, disabling\n");
420 dev_priv->no_fbc_reason = FBC_NO_OUTPUT;
424 intel_crtc = to_intel_crtc(crtc);
426 intel_fb = to_intel_framebuffer(fb);
429 enable_fbc = i915_enable_fbc;
430 if (enable_fbc < 0) {
431 DRM_DEBUG_KMS("fbc set to per-chip default\n");
433 if (INTEL_INFO(dev)->gen <= 6)
437 DRM_DEBUG_KMS("fbc disabled per module param\n");
438 dev_priv->no_fbc_reason = FBC_MODULE_PARAM;
441 if (intel_fb->obj->base.size > dev_priv->cfb_size) {
442 DRM_DEBUG_KMS("framebuffer too large, disabling "
444 dev_priv->no_fbc_reason = FBC_STOLEN_TOO_SMALL;
447 if ((crtc->mode.flags & DRM_MODE_FLAG_INTERLACE) ||
448 (crtc->mode.flags & DRM_MODE_FLAG_DBLSCAN)) {
449 DRM_DEBUG_KMS("mode incompatible with compression, "
451 dev_priv->no_fbc_reason = FBC_UNSUPPORTED_MODE;
454 if ((crtc->mode.hdisplay > 2048) ||
455 (crtc->mode.vdisplay > 1536)) {
456 DRM_DEBUG_KMS("mode too large for compression, disabling\n");
457 dev_priv->no_fbc_reason = FBC_MODE_TOO_LARGE;
460 if ((IS_I915GM(dev) || IS_I945GM(dev)) && intel_crtc->plane != 0) {
461 DRM_DEBUG_KMS("plane not 0, disabling compression\n");
462 dev_priv->no_fbc_reason = FBC_BAD_PLANE;
466 /* The use of a CPU fence is mandatory in order to detect writes
467 * by the CPU to the scanout and trigger updates to the FBC.
469 if (obj->tiling_mode != I915_TILING_X ||
470 obj->fence_reg == I915_FENCE_REG_NONE) {
471 DRM_DEBUG_KMS("framebuffer not tiled or fenced, disabling compression\n");
472 dev_priv->no_fbc_reason = FBC_NOT_TILED;
476 /* If the kernel debugger is active, always disable compression */
480 /* If the scanout has not changed, don't modify the FBC settings.
481 * Note that we make the fundamental assumption that the fb->obj
482 * cannot be unpinned (and have its GTT offset and fence revoked)
483 * without first being decoupled from the scanout and FBC disabled.
485 if (dev_priv->cfb_plane == intel_crtc->plane &&
486 dev_priv->cfb_fb == fb->base.id &&
487 dev_priv->cfb_y == crtc->y)
490 if (intel_fbc_enabled(dev)) {
491 /* We update FBC along two paths, after changing fb/crtc
492 * configuration (modeswitching) and after page-flipping
493 * finishes. For the latter, we know that not only did
494 * we disable the FBC at the start of the page-flip
495 * sequence, but also more than one vblank has passed.
497 * For the former case of modeswitching, it is possible
498 * to switch between two FBC valid configurations
499 * instantaneously so we do need to disable the FBC
500 * before we can modify its control registers. We also
501 * have to wait for the next vblank for that to take
502 * effect. However, since we delay enabling FBC we can
503 * assume that a vblank has passed since disabling and
504 * that we can safely alter the registers in the deferred
507 * In the scenario that we go from a valid to invalid
508 * and then back to valid FBC configuration we have
509 * no strict enforcement that a vblank occurred since
510 * disabling the FBC. However, along all current pipe
511 * disabling paths we do need to wait for a vblank at
512 * some point. And we wait before enabling FBC anyway.
514 DRM_DEBUG_KMS("disabling active FBC for update\n");
515 intel_disable_fbc(dev);
518 intel_enable_fbc(crtc, 500);
522 /* Multiple disables should be harmless */
523 if (intel_fbc_enabled(dev)) {
524 DRM_DEBUG_KMS("unsupported config, disabling FBC\n");
525 intel_disable_fbc(dev);
529 static void i915_pineview_get_mem_freq(struct drm_device *dev)
531 drm_i915_private_t *dev_priv = dev->dev_private;
534 tmp = I915_READ(CLKCFG);
536 switch (tmp & CLKCFG_FSB_MASK) {
538 dev_priv->fsb_freq = 533; /* 133*4 */
541 dev_priv->fsb_freq = 800; /* 200*4 */
544 dev_priv->fsb_freq = 667; /* 167*4 */
547 dev_priv->fsb_freq = 400; /* 100*4 */
551 switch (tmp & CLKCFG_MEM_MASK) {
553 dev_priv->mem_freq = 533;
556 dev_priv->mem_freq = 667;
559 dev_priv->mem_freq = 800;
563 /* detect pineview DDR3 setting */
564 tmp = I915_READ(CSHRDDR3CTL);
565 dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
568 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
570 drm_i915_private_t *dev_priv = dev->dev_private;
573 ddrpll = I915_READ16(DDRMPLL1);
574 csipll = I915_READ16(CSIPLL0);
576 switch (ddrpll & 0xff) {
578 dev_priv->mem_freq = 800;
581 dev_priv->mem_freq = 1066;
584 dev_priv->mem_freq = 1333;
587 dev_priv->mem_freq = 1600;
590 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
592 dev_priv->mem_freq = 0;
596 dev_priv->ips.r_t = dev_priv->mem_freq;
598 switch (csipll & 0x3ff) {
600 dev_priv->fsb_freq = 3200;
603 dev_priv->fsb_freq = 3733;
606 dev_priv->fsb_freq = 4266;
609 dev_priv->fsb_freq = 4800;
612 dev_priv->fsb_freq = 5333;
615 dev_priv->fsb_freq = 5866;
618 dev_priv->fsb_freq = 6400;
621 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
623 dev_priv->fsb_freq = 0;
627 if (dev_priv->fsb_freq == 3200) {
628 dev_priv->ips.c_m = 0;
629 } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
630 dev_priv->ips.c_m = 1;
632 dev_priv->ips.c_m = 2;
636 static const struct cxsr_latency cxsr_latency_table[] = {
637 {1, 0, 800, 400, 3382, 33382, 3983, 33983}, /* DDR2-400 SC */
638 {1, 0, 800, 667, 3354, 33354, 3807, 33807}, /* DDR2-667 SC */
639 {1, 0, 800, 800, 3347, 33347, 3763, 33763}, /* DDR2-800 SC */
640 {1, 1, 800, 667, 6420, 36420, 6873, 36873}, /* DDR3-667 SC */
641 {1, 1, 800, 800, 5902, 35902, 6318, 36318}, /* DDR3-800 SC */
643 {1, 0, 667, 400, 3400, 33400, 4021, 34021}, /* DDR2-400 SC */
644 {1, 0, 667, 667, 3372, 33372, 3845, 33845}, /* DDR2-667 SC */
645 {1, 0, 667, 800, 3386, 33386, 3822, 33822}, /* DDR2-800 SC */
646 {1, 1, 667, 667, 6438, 36438, 6911, 36911}, /* DDR3-667 SC */
647 {1, 1, 667, 800, 5941, 35941, 6377, 36377}, /* DDR3-800 SC */
649 {1, 0, 400, 400, 3472, 33472, 4173, 34173}, /* DDR2-400 SC */
650 {1, 0, 400, 667, 3443, 33443, 3996, 33996}, /* DDR2-667 SC */
651 {1, 0, 400, 800, 3430, 33430, 3946, 33946}, /* DDR2-800 SC */
652 {1, 1, 400, 667, 6509, 36509, 7062, 37062}, /* DDR3-667 SC */
653 {1, 1, 400, 800, 5985, 35985, 6501, 36501}, /* DDR3-800 SC */
655 {0, 0, 800, 400, 3438, 33438, 4065, 34065}, /* DDR2-400 SC */
656 {0, 0, 800, 667, 3410, 33410, 3889, 33889}, /* DDR2-667 SC */
657 {0, 0, 800, 800, 3403, 33403, 3845, 33845}, /* DDR2-800 SC */
658 {0, 1, 800, 667, 6476, 36476, 6955, 36955}, /* DDR3-667 SC */
659 {0, 1, 800, 800, 5958, 35958, 6400, 36400}, /* DDR3-800 SC */
661 {0, 0, 667, 400, 3456, 33456, 4103, 34106}, /* DDR2-400 SC */
662 {0, 0, 667, 667, 3428, 33428, 3927, 33927}, /* DDR2-667 SC */
663 {0, 0, 667, 800, 3443, 33443, 3905, 33905}, /* DDR2-800 SC */
664 {0, 1, 667, 667, 6494, 36494, 6993, 36993}, /* DDR3-667 SC */
665 {0, 1, 667, 800, 5998, 35998, 6460, 36460}, /* DDR3-800 SC */
667 {0, 0, 400, 400, 3528, 33528, 4255, 34255}, /* DDR2-400 SC */
668 {0, 0, 400, 667, 3500, 33500, 4079, 34079}, /* DDR2-667 SC */
669 {0, 0, 400, 800, 3487, 33487, 4029, 34029}, /* DDR2-800 SC */
670 {0, 1, 400, 667, 6566, 36566, 7145, 37145}, /* DDR3-667 SC */
671 {0, 1, 400, 800, 6042, 36042, 6584, 36584}, /* DDR3-800 SC */
674 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
679 const struct cxsr_latency *latency;
682 if (fsb == 0 || mem == 0)
685 for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
686 latency = &cxsr_latency_table[i];
687 if (is_desktop == latency->is_desktop &&
688 is_ddr3 == latency->is_ddr3 &&
689 fsb == latency->fsb_freq && mem == latency->mem_freq)
693 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
698 static void pineview_disable_cxsr(struct drm_device *dev)
700 struct drm_i915_private *dev_priv = dev->dev_private;
702 /* deactivate cxsr */
703 I915_WRITE(DSPFW3, I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN);
707 * Latency for FIFO fetches is dependent on several factors:
708 * - memory configuration (speed, channels)
710 * - current MCH state
711 * It can be fairly high in some situations, so here we assume a fairly
712 * pessimal value. It's a tradeoff between extra memory fetches (if we
713 * set this value too high, the FIFO will fetch frequently to stay full)
714 * and power consumption (set it too low to save power and we might see
715 * FIFO underruns and display "flicker").
717 * A value of 5us seems to be a good balance; safe for very low end
718 * platforms but not overly aggressive on lower latency configs.
720 static const int latency_ns = 5000;
722 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
724 struct drm_i915_private *dev_priv = dev->dev_private;
725 uint32_t dsparb = I915_READ(DSPARB);
728 size = dsparb & 0x7f;
730 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
732 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
733 plane ? "B" : "A", size);
738 static int i85x_get_fifo_size(struct drm_device *dev, int plane)
740 struct drm_i915_private *dev_priv = dev->dev_private;
741 uint32_t dsparb = I915_READ(DSPARB);
744 size = dsparb & 0x1ff;
746 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
747 size >>= 1; /* Convert to cachelines */
749 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
750 plane ? "B" : "A", size);
755 static int i845_get_fifo_size(struct drm_device *dev, int plane)
757 struct drm_i915_private *dev_priv = dev->dev_private;
758 uint32_t dsparb = I915_READ(DSPARB);
761 size = dsparb & 0x7f;
762 size >>= 2; /* Convert to cachelines */
764 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
771 static int i830_get_fifo_size(struct drm_device *dev, int plane)
773 struct drm_i915_private *dev_priv = dev->dev_private;
774 uint32_t dsparb = I915_READ(DSPARB);
777 size = dsparb & 0x7f;
778 size >>= 1; /* Convert to cachelines */
780 DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
781 plane ? "B" : "A", size);
786 /* Pineview has different values for various configs */
787 static const struct intel_watermark_params pineview_display_wm = {
788 PINEVIEW_DISPLAY_FIFO,
792 PINEVIEW_FIFO_LINE_SIZE
794 static const struct intel_watermark_params pineview_display_hplloff_wm = {
795 PINEVIEW_DISPLAY_FIFO,
797 PINEVIEW_DFT_HPLLOFF_WM,
799 PINEVIEW_FIFO_LINE_SIZE
801 static const struct intel_watermark_params pineview_cursor_wm = {
802 PINEVIEW_CURSOR_FIFO,
803 PINEVIEW_CURSOR_MAX_WM,
804 PINEVIEW_CURSOR_DFT_WM,
805 PINEVIEW_CURSOR_GUARD_WM,
806 PINEVIEW_FIFO_LINE_SIZE,
808 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
809 PINEVIEW_CURSOR_FIFO,
810 PINEVIEW_CURSOR_MAX_WM,
811 PINEVIEW_CURSOR_DFT_WM,
812 PINEVIEW_CURSOR_GUARD_WM,
813 PINEVIEW_FIFO_LINE_SIZE
815 static const struct intel_watermark_params g4x_wm_info = {
822 static const struct intel_watermark_params g4x_cursor_wm_info = {
829 static const struct intel_watermark_params valleyview_wm_info = {
830 VALLEYVIEW_FIFO_SIZE,
836 static const struct intel_watermark_params valleyview_cursor_wm_info = {
838 VALLEYVIEW_CURSOR_MAX_WM,
843 static const struct intel_watermark_params i965_cursor_wm_info = {
850 static const struct intel_watermark_params i945_wm_info = {
857 static const struct intel_watermark_params i915_wm_info = {
864 static const struct intel_watermark_params i855_wm_info = {
871 static const struct intel_watermark_params i830_wm_info = {
879 static const struct intel_watermark_params ironlake_display_wm_info = {
886 static const struct intel_watermark_params ironlake_cursor_wm_info = {
893 static const struct intel_watermark_params ironlake_display_srwm_info = {
895 ILK_DISPLAY_MAX_SRWM,
896 ILK_DISPLAY_DFT_SRWM,
900 static const struct intel_watermark_params ironlake_cursor_srwm_info = {
908 static const struct intel_watermark_params sandybridge_display_wm_info = {
915 static const struct intel_watermark_params sandybridge_cursor_wm_info = {
922 static const struct intel_watermark_params sandybridge_display_srwm_info = {
924 SNB_DISPLAY_MAX_SRWM,
925 SNB_DISPLAY_DFT_SRWM,
929 static const struct intel_watermark_params sandybridge_cursor_srwm_info = {
939 * intel_calculate_wm - calculate watermark level
940 * @clock_in_khz: pixel clock
941 * @wm: chip FIFO params
942 * @pixel_size: display pixel size
943 * @latency_ns: memory latency for the platform
945 * Calculate the watermark level (the level at which the display plane will
946 * start fetching from memory again). Each chip has a different display
947 * FIFO size and allocation, so the caller needs to figure that out and pass
948 * in the correct intel_watermark_params structure.
950 * As the pixel clock runs, the FIFO will be drained at a rate that depends
951 * on the pixel size. When it reaches the watermark level, it'll start
952 * fetching FIFO line sized based chunks from memory until the FIFO fills
953 * past the watermark point. If the FIFO drains completely, a FIFO underrun
954 * will occur, and a display engine hang could result.
956 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
957 const struct intel_watermark_params *wm,
960 unsigned long latency_ns)
962 long entries_required, wm_size;
965 * Note: we need to make sure we don't overflow for various clock &
967 * clocks go from a few thousand to several hundred thousand.
968 * latency is usually a few thousand
970 entries_required = ((clock_in_khz / 1000) * pixel_size * latency_ns) /
972 entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
974 DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
976 wm_size = fifo_size - (entries_required + wm->guard_size);
978 DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
980 /* Don't promote wm_size to unsigned... */
981 if (wm_size > (long)wm->max_wm)
982 wm_size = wm->max_wm;
984 wm_size = wm->default_wm;
988 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
990 struct drm_crtc *crtc, *enabled = NULL;
992 list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
993 if (crtc->enabled && crtc->fb) {
1003 static void pineview_update_wm(struct drm_device *dev)
1005 struct drm_i915_private *dev_priv = dev->dev_private;
1006 struct drm_crtc *crtc;
1007 const struct cxsr_latency *latency;
1011 latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
1012 dev_priv->fsb_freq, dev_priv->mem_freq);
1014 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
1015 pineview_disable_cxsr(dev);
1019 crtc = single_enabled_crtc(dev);
1021 int clock = crtc->mode.clock;
1022 int pixel_size = crtc->fb->bits_per_pixel / 8;
1025 wm = intel_calculate_wm(clock, &pineview_display_wm,
1026 pineview_display_wm.fifo_size,
1027 pixel_size, latency->display_sr);
1028 reg = I915_READ(DSPFW1);
1029 reg &= ~DSPFW_SR_MASK;
1030 reg |= wm << DSPFW_SR_SHIFT;
1031 I915_WRITE(DSPFW1, reg);
1032 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
1035 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
1036 pineview_display_wm.fifo_size,
1037 pixel_size, latency->cursor_sr);
1038 reg = I915_READ(DSPFW3);
1039 reg &= ~DSPFW_CURSOR_SR_MASK;
1040 reg |= (wm & 0x3f) << DSPFW_CURSOR_SR_SHIFT;
1041 I915_WRITE(DSPFW3, reg);
1043 /* Display HPLL off SR */
1044 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
1045 pineview_display_hplloff_wm.fifo_size,
1046 pixel_size, latency->display_hpll_disable);
1047 reg = I915_READ(DSPFW3);
1048 reg &= ~DSPFW_HPLL_SR_MASK;
1049 reg |= wm & DSPFW_HPLL_SR_MASK;
1050 I915_WRITE(DSPFW3, reg);
1052 /* cursor HPLL off SR */
1053 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
1054 pineview_display_hplloff_wm.fifo_size,
1055 pixel_size, latency->cursor_hpll_disable);
1056 reg = I915_READ(DSPFW3);
1057 reg &= ~DSPFW_HPLL_CURSOR_MASK;
1058 reg |= (wm & 0x3f) << DSPFW_HPLL_CURSOR_SHIFT;
1059 I915_WRITE(DSPFW3, reg);
1060 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
1064 I915_READ(DSPFW3) | PINEVIEW_SELF_REFRESH_EN);
1065 DRM_DEBUG_KMS("Self-refresh is enabled\n");
1067 pineview_disable_cxsr(dev);
1068 DRM_DEBUG_KMS("Self-refresh is disabled\n");
1072 static bool g4x_compute_wm0(struct drm_device *dev,
1074 const struct intel_watermark_params *display,
1075 int display_latency_ns,
1076 const struct intel_watermark_params *cursor,
1077 int cursor_latency_ns,
1081 struct drm_crtc *crtc;
1082 int htotal, hdisplay, clock, pixel_size;
1083 int line_time_us, line_count;
1084 int entries, tlb_miss;
1086 crtc = intel_get_crtc_for_plane(dev, plane);
1087 if (crtc->fb == NULL || !crtc->enabled) {
1088 *cursor_wm = cursor->guard_size;
1089 *plane_wm = display->guard_size;
1093 htotal = crtc->mode.htotal;
1094 hdisplay = crtc->mode.hdisplay;
1095 clock = crtc->mode.clock;
1096 pixel_size = crtc->fb->bits_per_pixel / 8;
1098 /* Use the small buffer method to calculate plane watermark */
1099 entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
1100 tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
1102 entries += tlb_miss;
1103 entries = DIV_ROUND_UP(entries, display->cacheline_size);
1104 *plane_wm = entries + display->guard_size;
1105 if (*plane_wm > (int)display->max_wm)
1106 *plane_wm = display->max_wm;
1108 /* Use the large buffer method to calculate cursor watermark */
1109 line_time_us = ((htotal * 1000) / clock);
1110 line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
1111 entries = line_count * 64 * pixel_size;
1112 tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
1114 entries += tlb_miss;
1115 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1116 *cursor_wm = entries + cursor->guard_size;
1117 if (*cursor_wm > (int)cursor->max_wm)
1118 *cursor_wm = (int)cursor->max_wm;
1124 * Check the wm result.
1126 * If any calculated watermark values is larger than the maximum value that
1127 * can be programmed into the associated watermark register, that watermark
1130 static bool g4x_check_srwm(struct drm_device *dev,
1131 int display_wm, int cursor_wm,
1132 const struct intel_watermark_params *display,
1133 const struct intel_watermark_params *cursor)
1135 DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
1136 display_wm, cursor_wm);
1138 if (display_wm > display->max_wm) {
1139 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
1140 display_wm, display->max_wm);
1144 if (cursor_wm > cursor->max_wm) {
1145 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
1146 cursor_wm, cursor->max_wm);
1150 if (!(display_wm || cursor_wm)) {
1151 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
1158 static bool g4x_compute_srwm(struct drm_device *dev,
1161 const struct intel_watermark_params *display,
1162 const struct intel_watermark_params *cursor,
1163 int *display_wm, int *cursor_wm)
1165 struct drm_crtc *crtc;
1166 int hdisplay, htotal, pixel_size, clock;
1167 unsigned long line_time_us;
1168 int line_count, line_size;
1173 *display_wm = *cursor_wm = 0;
1177 crtc = intel_get_crtc_for_plane(dev, plane);
1178 hdisplay = crtc->mode.hdisplay;
1179 htotal = crtc->mode.htotal;
1180 clock = crtc->mode.clock;
1181 pixel_size = crtc->fb->bits_per_pixel / 8;
1183 line_time_us = (htotal * 1000) / clock;
1184 line_count = (latency_ns / line_time_us + 1000) / 1000;
1185 line_size = hdisplay * pixel_size;
1187 /* Use the minimum of the small and large buffer method for primary */
1188 small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1189 large = line_count * line_size;
1191 entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1192 *display_wm = entries + display->guard_size;
1194 /* calculate the self-refresh watermark for display cursor */
1195 entries = line_count * pixel_size * 64;
1196 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1197 *cursor_wm = entries + cursor->guard_size;
1199 return g4x_check_srwm(dev,
1200 *display_wm, *cursor_wm,
1204 static bool vlv_compute_drain_latency(struct drm_device *dev,
1206 int *plane_prec_mult,
1208 int *cursor_prec_mult,
1211 struct drm_crtc *crtc;
1212 int clock, pixel_size;
1215 crtc = intel_get_crtc_for_plane(dev, plane);
1216 if (crtc->fb == NULL || !crtc->enabled)
1219 clock = crtc->mode.clock; /* VESA DOT Clock */
1220 pixel_size = crtc->fb->bits_per_pixel / 8; /* BPP */
1222 entries = (clock / 1000) * pixel_size;
1223 *plane_prec_mult = (entries > 256) ?
1224 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1225 *plane_dl = (64 * (*plane_prec_mult) * 4) / ((clock / 1000) *
1228 entries = (clock / 1000) * 4; /* BPP is always 4 for cursor */
1229 *cursor_prec_mult = (entries > 256) ?
1230 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1231 *cursor_dl = (64 * (*cursor_prec_mult) * 4) / ((clock / 1000) * 4);
1237 * Update drain latency registers of memory arbiter
1239 * Valleyview SoC has a new memory arbiter and needs drain latency registers
1240 * to be programmed. Each plane has a drain latency multiplier and a drain
1244 static void vlv_update_drain_latency(struct drm_device *dev)
1246 struct drm_i915_private *dev_priv = dev->dev_private;
1247 int planea_prec, planea_dl, planeb_prec, planeb_dl;
1248 int cursora_prec, cursora_dl, cursorb_prec, cursorb_dl;
1249 int plane_prec_mult, cursor_prec_mult; /* Precision multiplier is
1252 /* For plane A, Cursor A */
1253 if (vlv_compute_drain_latency(dev, 0, &plane_prec_mult, &planea_dl,
1254 &cursor_prec_mult, &cursora_dl)) {
1255 cursora_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1256 DDL_CURSORA_PRECISION_32 : DDL_CURSORA_PRECISION_16;
1257 planea_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1258 DDL_PLANEA_PRECISION_32 : DDL_PLANEA_PRECISION_16;
1260 I915_WRITE(VLV_DDL1, cursora_prec |
1261 (cursora_dl << DDL_CURSORA_SHIFT) |
1262 planea_prec | planea_dl);
1265 /* For plane B, Cursor B */
1266 if (vlv_compute_drain_latency(dev, 1, &plane_prec_mult, &planeb_dl,
1267 &cursor_prec_mult, &cursorb_dl)) {
1268 cursorb_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1269 DDL_CURSORB_PRECISION_32 : DDL_CURSORB_PRECISION_16;
1270 planeb_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1271 DDL_PLANEB_PRECISION_32 : DDL_PLANEB_PRECISION_16;
1273 I915_WRITE(VLV_DDL2, cursorb_prec |
1274 (cursorb_dl << DDL_CURSORB_SHIFT) |
1275 planeb_prec | planeb_dl);
1279 #define single_plane_enabled(mask) is_power_of_2(mask)
1281 static void valleyview_update_wm(struct drm_device *dev)
1283 static const int sr_latency_ns = 12000;
1284 struct drm_i915_private *dev_priv = dev->dev_private;
1285 int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1286 int plane_sr, cursor_sr;
1287 unsigned int enabled = 0;
1289 vlv_update_drain_latency(dev);
1291 if (g4x_compute_wm0(dev, 0,
1292 &valleyview_wm_info, latency_ns,
1293 &valleyview_cursor_wm_info, latency_ns,
1294 &planea_wm, &cursora_wm))
1297 if (g4x_compute_wm0(dev, 1,
1298 &valleyview_wm_info, latency_ns,
1299 &valleyview_cursor_wm_info, latency_ns,
1300 &planeb_wm, &cursorb_wm))
1303 plane_sr = cursor_sr = 0;
1304 if (single_plane_enabled(enabled) &&
1305 g4x_compute_srwm(dev, ffs(enabled) - 1,
1307 &valleyview_wm_info,
1308 &valleyview_cursor_wm_info,
1309 &plane_sr, &cursor_sr))
1310 I915_WRITE(FW_BLC_SELF_VLV, FW_CSPWRDWNEN);
1312 I915_WRITE(FW_BLC_SELF_VLV,
1313 I915_READ(FW_BLC_SELF_VLV) & ~FW_CSPWRDWNEN);
1315 DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1316 planea_wm, cursora_wm,
1317 planeb_wm, cursorb_wm,
1318 plane_sr, cursor_sr);
1321 (plane_sr << DSPFW_SR_SHIFT) |
1322 (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1323 (planeb_wm << DSPFW_PLANEB_SHIFT) |
1326 (I915_READ(DSPFW2) & DSPFW_CURSORA_MASK) |
1327 (cursora_wm << DSPFW_CURSORA_SHIFT));
1329 (I915_READ(DSPFW3) | (cursor_sr << DSPFW_CURSOR_SR_SHIFT)));
1332 static void g4x_update_wm(struct drm_device *dev)
1334 static const int sr_latency_ns = 12000;
1335 struct drm_i915_private *dev_priv = dev->dev_private;
1336 int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1337 int plane_sr, cursor_sr;
1338 unsigned int enabled = 0;
1340 if (g4x_compute_wm0(dev, 0,
1341 &g4x_wm_info, latency_ns,
1342 &g4x_cursor_wm_info, latency_ns,
1343 &planea_wm, &cursora_wm))
1346 if (g4x_compute_wm0(dev, 1,
1347 &g4x_wm_info, latency_ns,
1348 &g4x_cursor_wm_info, latency_ns,
1349 &planeb_wm, &cursorb_wm))
1352 plane_sr = cursor_sr = 0;
1353 if (single_plane_enabled(enabled) &&
1354 g4x_compute_srwm(dev, ffs(enabled) - 1,
1357 &g4x_cursor_wm_info,
1358 &plane_sr, &cursor_sr))
1359 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1361 I915_WRITE(FW_BLC_SELF,
1362 I915_READ(FW_BLC_SELF) & ~FW_BLC_SELF_EN);
1364 DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1365 planea_wm, cursora_wm,
1366 planeb_wm, cursorb_wm,
1367 plane_sr, cursor_sr);
1370 (plane_sr << DSPFW_SR_SHIFT) |
1371 (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1372 (planeb_wm << DSPFW_PLANEB_SHIFT) |
1375 (I915_READ(DSPFW2) & DSPFW_CURSORA_MASK) |
1376 (cursora_wm << DSPFW_CURSORA_SHIFT));
1377 /* HPLL off in SR has some issues on G4x... disable it */
1379 (I915_READ(DSPFW3) & ~DSPFW_HPLL_SR_EN) |
1380 (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1383 static void i965_update_wm(struct drm_device *dev)
1385 struct drm_i915_private *dev_priv = dev->dev_private;
1386 struct drm_crtc *crtc;
1390 /* Calc sr entries for one plane configs */
1391 crtc = single_enabled_crtc(dev);
1393 /* self-refresh has much higher latency */
1394 static const int sr_latency_ns = 12000;
1395 int clock = crtc->mode.clock;
1396 int htotal = crtc->mode.htotal;
1397 int hdisplay = crtc->mode.hdisplay;
1398 int pixel_size = crtc->fb->bits_per_pixel / 8;
1399 unsigned long line_time_us;
1402 line_time_us = ((htotal * 1000) / clock);
1404 /* Use ns/us then divide to preserve precision */
1405 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1406 pixel_size * hdisplay;
1407 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1408 srwm = I965_FIFO_SIZE - entries;
1412 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1415 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1417 entries = DIV_ROUND_UP(entries,
1418 i965_cursor_wm_info.cacheline_size);
1419 cursor_sr = i965_cursor_wm_info.fifo_size -
1420 (entries + i965_cursor_wm_info.guard_size);
1422 if (cursor_sr > i965_cursor_wm_info.max_wm)
1423 cursor_sr = i965_cursor_wm_info.max_wm;
1425 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1426 "cursor %d\n", srwm, cursor_sr);
1428 if (IS_CRESTLINE(dev))
1429 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1431 /* Turn off self refresh if both pipes are enabled */
1432 if (IS_CRESTLINE(dev))
1433 I915_WRITE(FW_BLC_SELF, I915_READ(FW_BLC_SELF)
1437 DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1440 /* 965 has limitations... */
1441 I915_WRITE(DSPFW1, (srwm << DSPFW_SR_SHIFT) |
1442 (8 << 16) | (8 << 8) | (8 << 0));
1443 I915_WRITE(DSPFW2, (8 << 8) | (8 << 0));
1444 /* update cursor SR watermark */
1445 I915_WRITE(DSPFW3, (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1448 static void i9xx_update_wm(struct drm_device *dev)
1450 struct drm_i915_private *dev_priv = dev->dev_private;
1451 const struct intel_watermark_params *wm_info;
1456 int planea_wm, planeb_wm;
1457 struct drm_crtc *crtc, *enabled = NULL;
1460 wm_info = &i945_wm_info;
1461 else if (!IS_GEN2(dev))
1462 wm_info = &i915_wm_info;
1464 wm_info = &i855_wm_info;
1466 fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1467 crtc = intel_get_crtc_for_plane(dev, 0);
1468 if (crtc->enabled && crtc->fb) {
1469 planea_wm = intel_calculate_wm(crtc->mode.clock,
1471 crtc->fb->bits_per_pixel / 8,
1475 planea_wm = fifo_size - wm_info->guard_size;
1477 fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1478 crtc = intel_get_crtc_for_plane(dev, 1);
1479 if (crtc->enabled && crtc->fb) {
1480 planeb_wm = intel_calculate_wm(crtc->mode.clock,
1482 crtc->fb->bits_per_pixel / 8,
1484 if (enabled == NULL)
1489 planeb_wm = fifo_size - wm_info->guard_size;
1491 DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1494 * Overlay gets an aggressive default since video jitter is bad.
1498 /* Play safe and disable self-refresh before adjusting watermarks. */
1499 if (IS_I945G(dev) || IS_I945GM(dev))
1500 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN_MASK | 0);
1501 else if (IS_I915GM(dev))
1502 I915_WRITE(INSTPM, I915_READ(INSTPM) & ~INSTPM_SELF_EN);
1504 /* Calc sr entries for one plane configs */
1505 if (HAS_FW_BLC(dev) && enabled) {
1506 /* self-refresh has much higher latency */
1507 static const int sr_latency_ns = 6000;
1508 int clock = enabled->mode.clock;
1509 int htotal = enabled->mode.htotal;
1510 int hdisplay = enabled->mode.hdisplay;
1511 int pixel_size = enabled->fb->bits_per_pixel / 8;
1512 unsigned long line_time_us;
1515 line_time_us = (htotal * 1000) / clock;
1517 /* Use ns/us then divide to preserve precision */
1518 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1519 pixel_size * hdisplay;
1520 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1521 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1522 srwm = wm_info->fifo_size - entries;
1526 if (IS_I945G(dev) || IS_I945GM(dev))
1527 I915_WRITE(FW_BLC_SELF,
1528 FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1529 else if (IS_I915GM(dev))
1530 I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1533 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1534 planea_wm, planeb_wm, cwm, srwm);
1536 fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1537 fwater_hi = (cwm & 0x1f);
1539 /* Set request length to 8 cachelines per fetch */
1540 fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1541 fwater_hi = fwater_hi | (1 << 8);
1543 I915_WRITE(FW_BLC, fwater_lo);
1544 I915_WRITE(FW_BLC2, fwater_hi);
1546 if (HAS_FW_BLC(dev)) {
1548 if (IS_I945G(dev) || IS_I945GM(dev))
1549 I915_WRITE(FW_BLC_SELF,
1550 FW_BLC_SELF_EN_MASK | FW_BLC_SELF_EN);
1551 else if (IS_I915GM(dev))
1552 I915_WRITE(INSTPM, I915_READ(INSTPM) | INSTPM_SELF_EN);
1553 DRM_DEBUG_KMS("memory self refresh enabled\n");
1555 DRM_DEBUG_KMS("memory self refresh disabled\n");
1559 static void i830_update_wm(struct drm_device *dev)
1561 struct drm_i915_private *dev_priv = dev->dev_private;
1562 struct drm_crtc *crtc;
1566 crtc = single_enabled_crtc(dev);
1570 planea_wm = intel_calculate_wm(crtc->mode.clock, &i830_wm_info,
1571 dev_priv->display.get_fifo_size(dev, 0),
1572 crtc->fb->bits_per_pixel / 8,
1574 fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1575 fwater_lo |= (3<<8) | planea_wm;
1577 DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1579 I915_WRITE(FW_BLC, fwater_lo);
1582 #define ILK_LP0_PLANE_LATENCY 700
1583 #define ILK_LP0_CURSOR_LATENCY 1300
1586 * Check the wm result.
1588 * If any calculated watermark values is larger than the maximum value that
1589 * can be programmed into the associated watermark register, that watermark
1592 static bool ironlake_check_srwm(struct drm_device *dev, int level,
1593 int fbc_wm, int display_wm, int cursor_wm,
1594 const struct intel_watermark_params *display,
1595 const struct intel_watermark_params *cursor)
1597 struct drm_i915_private *dev_priv = dev->dev_private;
1599 DRM_DEBUG_KMS("watermark %d: display plane %d, fbc lines %d,"
1600 " cursor %d\n", level, display_wm, fbc_wm, cursor_wm);
1602 if (fbc_wm > SNB_FBC_MAX_SRWM) {
1603 DRM_DEBUG_KMS("fbc watermark(%d) is too large(%d), disabling wm%d+\n",
1604 fbc_wm, SNB_FBC_MAX_SRWM, level);
1606 /* fbc has it's own way to disable FBC WM */
1607 I915_WRITE(DISP_ARB_CTL,
1608 I915_READ(DISP_ARB_CTL) | DISP_FBC_WM_DIS);
1612 if (display_wm > display->max_wm) {
1613 DRM_DEBUG_KMS("display watermark(%d) is too large(%d), disabling wm%d+\n",
1614 display_wm, SNB_DISPLAY_MAX_SRWM, level);
1618 if (cursor_wm > cursor->max_wm) {
1619 DRM_DEBUG_KMS("cursor watermark(%d) is too large(%d), disabling wm%d+\n",
1620 cursor_wm, SNB_CURSOR_MAX_SRWM, level);
1624 if (!(fbc_wm || display_wm || cursor_wm)) {
1625 DRM_DEBUG_KMS("latency %d is 0, disabling wm%d+\n", level, level);
1633 * Compute watermark values of WM[1-3],
1635 static bool ironlake_compute_srwm(struct drm_device *dev, int level, int plane,
1637 const struct intel_watermark_params *display,
1638 const struct intel_watermark_params *cursor,
1639 int *fbc_wm, int *display_wm, int *cursor_wm)
1641 struct drm_crtc *crtc;
1642 unsigned long line_time_us;
1643 int hdisplay, htotal, pixel_size, clock;
1644 int line_count, line_size;
1649 *fbc_wm = *display_wm = *cursor_wm = 0;
1653 crtc = intel_get_crtc_for_plane(dev, plane);
1654 hdisplay = crtc->mode.hdisplay;
1655 htotal = crtc->mode.htotal;
1656 clock = crtc->mode.clock;
1657 pixel_size = crtc->fb->bits_per_pixel / 8;
1659 line_time_us = (htotal * 1000) / clock;
1660 line_count = (latency_ns / line_time_us + 1000) / 1000;
1661 line_size = hdisplay * pixel_size;
1663 /* Use the minimum of the small and large buffer method for primary */
1664 small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1665 large = line_count * line_size;
1667 entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1668 *display_wm = entries + display->guard_size;
1672 * FBC WM = ((Final Primary WM * 64) / number of bytes per line) + 2
1674 *fbc_wm = DIV_ROUND_UP(*display_wm * 64, line_size) + 2;
1676 /* calculate the self-refresh watermark for display cursor */
1677 entries = line_count * pixel_size * 64;
1678 entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1679 *cursor_wm = entries + cursor->guard_size;
1681 return ironlake_check_srwm(dev, level,
1682 *fbc_wm, *display_wm, *cursor_wm,
1686 static void ironlake_update_wm(struct drm_device *dev)
1688 struct drm_i915_private *dev_priv = dev->dev_private;
1689 int fbc_wm, plane_wm, cursor_wm;
1690 unsigned int enabled;
1693 if (g4x_compute_wm0(dev, 0,
1694 &ironlake_display_wm_info,
1695 ILK_LP0_PLANE_LATENCY,
1696 &ironlake_cursor_wm_info,
1697 ILK_LP0_CURSOR_LATENCY,
1698 &plane_wm, &cursor_wm)) {
1699 I915_WRITE(WM0_PIPEA_ILK,
1700 (plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm);
1701 DRM_DEBUG_KMS("FIFO watermarks For pipe A -"
1702 " plane %d, " "cursor: %d\n",
1703 plane_wm, cursor_wm);
1707 if (g4x_compute_wm0(dev, 1,
1708 &ironlake_display_wm_info,
1709 ILK_LP0_PLANE_LATENCY,
1710 &ironlake_cursor_wm_info,
1711 ILK_LP0_CURSOR_LATENCY,
1712 &plane_wm, &cursor_wm)) {
1713 I915_WRITE(WM0_PIPEB_ILK,
1714 (plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm);
1715 DRM_DEBUG_KMS("FIFO watermarks For pipe B -"
1716 " plane %d, cursor: %d\n",
1717 plane_wm, cursor_wm);
1722 * Calculate and update the self-refresh watermark only when one
1723 * display plane is used.
1725 I915_WRITE(WM3_LP_ILK, 0);
1726 I915_WRITE(WM2_LP_ILK, 0);
1727 I915_WRITE(WM1_LP_ILK, 0);
1729 if (!single_plane_enabled(enabled))
1731 enabled = ffs(enabled) - 1;
1734 if (!ironlake_compute_srwm(dev, 1, enabled,
1735 ILK_READ_WM1_LATENCY() * 500,
1736 &ironlake_display_srwm_info,
1737 &ironlake_cursor_srwm_info,
1738 &fbc_wm, &plane_wm, &cursor_wm))
1741 I915_WRITE(WM1_LP_ILK,
1743 (ILK_READ_WM1_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1744 (fbc_wm << WM1_LP_FBC_SHIFT) |
1745 (plane_wm << WM1_LP_SR_SHIFT) |
1749 if (!ironlake_compute_srwm(dev, 2, enabled,
1750 ILK_READ_WM2_LATENCY() * 500,
1751 &ironlake_display_srwm_info,
1752 &ironlake_cursor_srwm_info,
1753 &fbc_wm, &plane_wm, &cursor_wm))
1756 I915_WRITE(WM2_LP_ILK,
1758 (ILK_READ_WM2_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1759 (fbc_wm << WM1_LP_FBC_SHIFT) |
1760 (plane_wm << WM1_LP_SR_SHIFT) |
1764 * WM3 is unsupported on ILK, probably because we don't have latency
1765 * data for that power state
1769 static void sandybridge_update_wm(struct drm_device *dev)
1771 struct drm_i915_private *dev_priv = dev->dev_private;
1772 int latency = SNB_READ_WM0_LATENCY() * 100; /* In unit 0.1us */
1774 int fbc_wm, plane_wm, cursor_wm;
1775 unsigned int enabled;
1778 if (g4x_compute_wm0(dev, 0,
1779 &sandybridge_display_wm_info, latency,
1780 &sandybridge_cursor_wm_info, latency,
1781 &plane_wm, &cursor_wm)) {
1782 val = I915_READ(WM0_PIPEA_ILK);
1783 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
1784 I915_WRITE(WM0_PIPEA_ILK, val |
1785 ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
1786 DRM_DEBUG_KMS("FIFO watermarks For pipe A -"
1787 " plane %d, " "cursor: %d\n",
1788 plane_wm, cursor_wm);
1792 if (g4x_compute_wm0(dev, 1,
1793 &sandybridge_display_wm_info, latency,
1794 &sandybridge_cursor_wm_info, latency,
1795 &plane_wm, &cursor_wm)) {
1796 val = I915_READ(WM0_PIPEB_ILK);
1797 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
1798 I915_WRITE(WM0_PIPEB_ILK, val |
1799 ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
1800 DRM_DEBUG_KMS("FIFO watermarks For pipe B -"
1801 " plane %d, cursor: %d\n",
1802 plane_wm, cursor_wm);
1806 if ((dev_priv->num_pipe == 3) &&
1807 g4x_compute_wm0(dev, 2,
1808 &sandybridge_display_wm_info, latency,
1809 &sandybridge_cursor_wm_info, latency,
1810 &plane_wm, &cursor_wm)) {
1811 val = I915_READ(WM0_PIPEC_IVB);
1812 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
1813 I915_WRITE(WM0_PIPEC_IVB, val |
1814 ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
1815 DRM_DEBUG_KMS("FIFO watermarks For pipe C -"
1816 " plane %d, cursor: %d\n",
1817 plane_wm, cursor_wm);
1822 * Calculate and update the self-refresh watermark only when one
1823 * display plane is used.
1825 * SNB support 3 levels of watermark.
1827 * WM1/WM2/WM2 watermarks have to be enabled in the ascending order,
1828 * and disabled in the descending order
1831 I915_WRITE(WM3_LP_ILK, 0);
1832 I915_WRITE(WM2_LP_ILK, 0);
1833 I915_WRITE(WM1_LP_ILK, 0);
1835 if (!single_plane_enabled(enabled) ||
1836 dev_priv->sprite_scaling_enabled)
1838 enabled = ffs(enabled) - 1;
1841 if (!ironlake_compute_srwm(dev, 1, enabled,
1842 SNB_READ_WM1_LATENCY() * 500,
1843 &sandybridge_display_srwm_info,
1844 &sandybridge_cursor_srwm_info,
1845 &fbc_wm, &plane_wm, &cursor_wm))
1848 I915_WRITE(WM1_LP_ILK,
1850 (SNB_READ_WM1_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1851 (fbc_wm << WM1_LP_FBC_SHIFT) |
1852 (plane_wm << WM1_LP_SR_SHIFT) |
1856 if (!ironlake_compute_srwm(dev, 2, enabled,
1857 SNB_READ_WM2_LATENCY() * 500,
1858 &sandybridge_display_srwm_info,
1859 &sandybridge_cursor_srwm_info,
1860 &fbc_wm, &plane_wm, &cursor_wm))
1863 I915_WRITE(WM2_LP_ILK,
1865 (SNB_READ_WM2_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1866 (fbc_wm << WM1_LP_FBC_SHIFT) |
1867 (plane_wm << WM1_LP_SR_SHIFT) |
1871 if (!ironlake_compute_srwm(dev, 3, enabled,
1872 SNB_READ_WM3_LATENCY() * 500,
1873 &sandybridge_display_srwm_info,
1874 &sandybridge_cursor_srwm_info,
1875 &fbc_wm, &plane_wm, &cursor_wm))
1878 I915_WRITE(WM3_LP_ILK,
1880 (SNB_READ_WM3_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1881 (fbc_wm << WM1_LP_FBC_SHIFT) |
1882 (plane_wm << WM1_LP_SR_SHIFT) |
1887 haswell_update_linetime_wm(struct drm_device *dev, int pipe,
1888 struct drm_display_mode *mode)
1890 struct drm_i915_private *dev_priv = dev->dev_private;
1893 temp = I915_READ(PIPE_WM_LINETIME(pipe));
1894 temp &= ~PIPE_WM_LINETIME_MASK;
1896 /* The WM are computed with base on how long it takes to fill a single
1897 * row at the given clock rate, multiplied by 8.
1899 temp |= PIPE_WM_LINETIME_TIME(
1900 ((mode->crtc_hdisplay * 1000) / mode->clock) * 8);
1902 /* IPS watermarks are only used by pipe A, and are ignored by
1903 * pipes B and C. They are calculated similarly to the common
1904 * linetime values, except that we are using CD clock frequency
1905 * in MHz instead of pixel rate for the division.
1907 * This is a placeholder for the IPS watermark calculation code.
1910 I915_WRITE(PIPE_WM_LINETIME(pipe), temp);
1914 sandybridge_compute_sprite_wm(struct drm_device *dev, int plane,
1915 uint32_t sprite_width, int pixel_size,
1916 const struct intel_watermark_params *display,
1917 int display_latency_ns, int *sprite_wm)
1919 struct drm_crtc *crtc;
1921 int entries, tlb_miss;
1923 crtc = intel_get_crtc_for_plane(dev, plane);
1924 if (crtc->fb == NULL || !crtc->enabled) {
1925 *sprite_wm = display->guard_size;
1929 clock = crtc->mode.clock;
1931 /* Use the small buffer method to calculate the sprite watermark */
1932 entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
1933 tlb_miss = display->fifo_size*display->cacheline_size -
1936 entries += tlb_miss;
1937 entries = DIV_ROUND_UP(entries, display->cacheline_size);
1938 *sprite_wm = entries + display->guard_size;
1939 if (*sprite_wm > (int)display->max_wm)
1940 *sprite_wm = display->max_wm;
1946 sandybridge_compute_sprite_srwm(struct drm_device *dev, int plane,
1947 uint32_t sprite_width, int pixel_size,
1948 const struct intel_watermark_params *display,
1949 int latency_ns, int *sprite_wm)
1951 struct drm_crtc *crtc;
1952 unsigned long line_time_us;
1954 int line_count, line_size;
1963 crtc = intel_get_crtc_for_plane(dev, plane);
1964 clock = crtc->mode.clock;
1970 line_time_us = (sprite_width * 1000) / clock;
1971 if (!line_time_us) {
1976 line_count = (latency_ns / line_time_us + 1000) / 1000;
1977 line_size = sprite_width * pixel_size;
1979 /* Use the minimum of the small and large buffer method for primary */
1980 small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1981 large = line_count * line_size;
1983 entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1984 *sprite_wm = entries + display->guard_size;
1986 return *sprite_wm > 0x3ff ? false : true;
1989 static void sandybridge_update_sprite_wm(struct drm_device *dev, int pipe,
1990 uint32_t sprite_width, int pixel_size)
1992 struct drm_i915_private *dev_priv = dev->dev_private;
1993 int latency = SNB_READ_WM0_LATENCY() * 100; /* In unit 0.1us */
2000 reg = WM0_PIPEA_ILK;
2003 reg = WM0_PIPEB_ILK;
2006 reg = WM0_PIPEC_IVB;
2009 return; /* bad pipe */
2012 ret = sandybridge_compute_sprite_wm(dev, pipe, sprite_width, pixel_size,
2013 &sandybridge_display_wm_info,
2014 latency, &sprite_wm);
2016 DRM_DEBUG_KMS("failed to compute sprite wm for pipe %d\n",
2021 val = I915_READ(reg);
2022 val &= ~WM0_PIPE_SPRITE_MASK;
2023 I915_WRITE(reg, val | (sprite_wm << WM0_PIPE_SPRITE_SHIFT));
2024 DRM_DEBUG_KMS("sprite watermarks For pipe %d - %d\n", pipe, sprite_wm);
2027 ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
2029 &sandybridge_display_srwm_info,
2030 SNB_READ_WM1_LATENCY() * 500,
2033 DRM_DEBUG_KMS("failed to compute sprite lp1 wm on pipe %d\n",
2037 I915_WRITE(WM1S_LP_ILK, sprite_wm);
2039 /* Only IVB has two more LP watermarks for sprite */
2040 if (!IS_IVYBRIDGE(dev))
2043 ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
2045 &sandybridge_display_srwm_info,
2046 SNB_READ_WM2_LATENCY() * 500,
2049 DRM_DEBUG_KMS("failed to compute sprite lp2 wm on pipe %d\n",
2053 I915_WRITE(WM2S_LP_IVB, sprite_wm);
2055 ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
2057 &sandybridge_display_srwm_info,
2058 SNB_READ_WM3_LATENCY() * 500,
2061 DRM_DEBUG_KMS("failed to compute sprite lp3 wm on pipe %d\n",
2065 I915_WRITE(WM3S_LP_IVB, sprite_wm);
2069 * intel_update_watermarks - update FIFO watermark values based on current modes
2071 * Calculate watermark values for the various WM regs based on current mode
2072 * and plane configuration.
2074 * There are several cases to deal with here:
2075 * - normal (i.e. non-self-refresh)
2076 * - self-refresh (SR) mode
2077 * - lines are large relative to FIFO size (buffer can hold up to 2)
2078 * - lines are small relative to FIFO size (buffer can hold more than 2
2079 * lines), so need to account for TLB latency
2081 * The normal calculation is:
2082 * watermark = dotclock * bytes per pixel * latency
2083 * where latency is platform & configuration dependent (we assume pessimal
2086 * The SR calculation is:
2087 * watermark = (trunc(latency/line time)+1) * surface width *
2090 * line time = htotal / dotclock
2091 * surface width = hdisplay for normal plane and 64 for cursor
2092 * and latency is assumed to be high, as above.
2094 * The final value programmed to the register should always be rounded up,
2095 * and include an extra 2 entries to account for clock crossings.
2097 * We don't use the sprite, so we can ignore that. And on Crestline we have
2098 * to set the non-SR watermarks to 8.
2100 void intel_update_watermarks(struct drm_device *dev)
2102 struct drm_i915_private *dev_priv = dev->dev_private;
2104 if (dev_priv->display.update_wm)
2105 dev_priv->display.update_wm(dev);
2108 void intel_update_linetime_watermarks(struct drm_device *dev,
2109 int pipe, struct drm_display_mode *mode)
2111 struct drm_i915_private *dev_priv = dev->dev_private;
2113 if (dev_priv->display.update_linetime_wm)
2114 dev_priv->display.update_linetime_wm(dev, pipe, mode);
2117 void intel_update_sprite_watermarks(struct drm_device *dev, int pipe,
2118 uint32_t sprite_width, int pixel_size)
2120 struct drm_i915_private *dev_priv = dev->dev_private;
2122 if (dev_priv->display.update_sprite_wm)
2123 dev_priv->display.update_sprite_wm(dev, pipe, sprite_width,
2127 static struct drm_i915_gem_object *
2128 intel_alloc_context_page(struct drm_device *dev)
2130 struct drm_i915_gem_object *ctx;
2133 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2135 ctx = i915_gem_alloc_object(dev, 4096);
2137 DRM_DEBUG("failed to alloc power context, RC6 disabled\n");
2141 ret = i915_gem_object_pin(ctx, 4096, true, false);
2143 DRM_ERROR("failed to pin power context: %d\n", ret);
2147 ret = i915_gem_object_set_to_gtt_domain(ctx, 1);
2149 DRM_ERROR("failed to set-domain on power context: %d\n", ret);
2156 i915_gem_object_unpin(ctx);
2158 drm_gem_object_unreference(&ctx->base);
2159 mutex_unlock(&dev->struct_mutex);
2164 * Lock protecting IPS related data structures
2166 DEFINE_SPINLOCK(mchdev_lock);
2168 /* Global for IPS driver to get at the current i915 device. Protected by
2170 static struct drm_i915_private *i915_mch_dev;
2172 bool ironlake_set_drps(struct drm_device *dev, u8 val)
2174 struct drm_i915_private *dev_priv = dev->dev_private;
2177 assert_spin_locked(&mchdev_lock);
2179 rgvswctl = I915_READ16(MEMSWCTL);
2180 if (rgvswctl & MEMCTL_CMD_STS) {
2181 DRM_DEBUG("gpu busy, RCS change rejected\n");
2182 return false; /* still busy with another command */
2185 rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
2186 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
2187 I915_WRITE16(MEMSWCTL, rgvswctl);
2188 POSTING_READ16(MEMSWCTL);
2190 rgvswctl |= MEMCTL_CMD_STS;
2191 I915_WRITE16(MEMSWCTL, rgvswctl);
2196 static void ironlake_enable_drps(struct drm_device *dev)
2198 struct drm_i915_private *dev_priv = dev->dev_private;
2199 u32 rgvmodectl = I915_READ(MEMMODECTL);
2200 u8 fmax, fmin, fstart, vstart;
2202 spin_lock_irq(&mchdev_lock);
2204 /* Enable temp reporting */
2205 I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
2206 I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
2208 /* 100ms RC evaluation intervals */
2209 I915_WRITE(RCUPEI, 100000);
2210 I915_WRITE(RCDNEI, 100000);
2212 /* Set max/min thresholds to 90ms and 80ms respectively */
2213 I915_WRITE(RCBMAXAVG, 90000);
2214 I915_WRITE(RCBMINAVG, 80000);
2216 I915_WRITE(MEMIHYST, 1);
2218 /* Set up min, max, and cur for interrupt handling */
2219 fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
2220 fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
2221 fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
2222 MEMMODE_FSTART_SHIFT;
2224 vstart = (I915_READ(PXVFREQ_BASE + (fstart * 4)) & PXVFREQ_PX_MASK) >>
2227 dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
2228 dev_priv->ips.fstart = fstart;
2230 dev_priv->ips.max_delay = fstart;
2231 dev_priv->ips.min_delay = fmin;
2232 dev_priv->ips.cur_delay = fstart;
2234 DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
2235 fmax, fmin, fstart);
2237 I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
2240 * Interrupts will be enabled in ironlake_irq_postinstall
2243 I915_WRITE(VIDSTART, vstart);
2244 POSTING_READ(VIDSTART);
2246 rgvmodectl |= MEMMODE_SWMODE_EN;
2247 I915_WRITE(MEMMODECTL, rgvmodectl);
2249 if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
2250 DRM_ERROR("stuck trying to change perf mode\n");
2253 ironlake_set_drps(dev, fstart);
2255 dev_priv->ips.last_count1 = I915_READ(0x112e4) + I915_READ(0x112e8) +
2257 dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
2258 dev_priv->ips.last_count2 = I915_READ(0x112f4);
2259 getrawmonotonic(&dev_priv->ips.last_time2);
2261 spin_unlock_irq(&mchdev_lock);
2264 static void ironlake_disable_drps(struct drm_device *dev)
2266 struct drm_i915_private *dev_priv = dev->dev_private;
2269 spin_lock_irq(&mchdev_lock);
2271 rgvswctl = I915_READ16(MEMSWCTL);
2273 /* Ack interrupts, disable EFC interrupt */
2274 I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
2275 I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
2276 I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
2277 I915_WRITE(DEIIR, DE_PCU_EVENT);
2278 I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
2280 /* Go back to the starting frequency */
2281 ironlake_set_drps(dev, dev_priv->ips.fstart);
2283 rgvswctl |= MEMCTL_CMD_STS;
2284 I915_WRITE(MEMSWCTL, rgvswctl);
2287 spin_unlock_irq(&mchdev_lock);
2290 /* There's a funny hw issue where the hw returns all 0 when reading from
2291 * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
2292 * ourselves, instead of doing a rmw cycle (which might result in us clearing
2293 * all limits and the gpu stuck at whatever frequency it is at atm).
2295 static u32 gen6_rps_limits(struct drm_i915_private *dev_priv, u8 *val)
2301 if (*val >= dev_priv->rps.max_delay)
2302 *val = dev_priv->rps.max_delay;
2303 limits |= dev_priv->rps.max_delay << 24;
2305 /* Only set the down limit when we've reached the lowest level to avoid
2306 * getting more interrupts, otherwise leave this clear. This prevents a
2307 * race in the hw when coming out of rc6: There's a tiny window where
2308 * the hw runs at the minimal clock before selecting the desired
2309 * frequency, if the down threshold expires in that window we will not
2310 * receive a down interrupt. */
2311 if (*val <= dev_priv->rps.min_delay) {
2312 *val = dev_priv->rps.min_delay;
2313 limits |= dev_priv->rps.min_delay << 16;
2319 void gen6_set_rps(struct drm_device *dev, u8 val)
2321 struct drm_i915_private *dev_priv = dev->dev_private;
2322 u32 limits = gen6_rps_limits(dev_priv, &val);
2324 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2326 if (val == dev_priv->rps.cur_delay)
2329 I915_WRITE(GEN6_RPNSWREQ,
2330 GEN6_FREQUENCY(val) |
2332 GEN6_AGGRESSIVE_TURBO);
2334 /* Make sure we continue to get interrupts
2335 * until we hit the minimum or maximum frequencies.
2337 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, limits);
2339 dev_priv->rps.cur_delay = val;
2342 static void gen6_disable_rps(struct drm_device *dev)
2344 struct drm_i915_private *dev_priv = dev->dev_private;
2346 I915_WRITE(GEN6_RC_CONTROL, 0);
2347 I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
2348 I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
2349 I915_WRITE(GEN6_PMIER, 0);
2350 /* Complete PM interrupt masking here doesn't race with the rps work
2351 * item again unmasking PM interrupts because that is using a different
2352 * register (PMIMR) to mask PM interrupts. The only risk is in leaving
2353 * stale bits in PMIIR and PMIMR which gen6_enable_rps will clean up. */
2355 spin_lock_irq(&dev_priv->rps.lock);
2356 dev_priv->rps.pm_iir = 0;
2357 spin_unlock_irq(&dev_priv->rps.lock);
2359 I915_WRITE(GEN6_PMIIR, I915_READ(GEN6_PMIIR));
2362 int intel_enable_rc6(const struct drm_device *dev)
2364 /* Respect the kernel parameter if it is set */
2365 if (i915_enable_rc6 >= 0)
2366 return i915_enable_rc6;
2368 if (INTEL_INFO(dev)->gen == 5) {
2369 #ifdef CONFIG_INTEL_IOMMU
2370 /* Disable rc6 on ilk if VT-d is on. */
2371 if (intel_iommu_gfx_mapped)
2374 DRM_DEBUG_DRIVER("Ironlake: only RC6 available\n");
2375 return INTEL_RC6_ENABLE;
2378 if (IS_HASWELL(dev)) {
2379 DRM_DEBUG_DRIVER("Haswell: only RC6 available\n");
2380 return INTEL_RC6_ENABLE;
2383 /* snb/ivb have more than one rc6 state. */
2384 if (INTEL_INFO(dev)->gen == 6) {
2385 DRM_DEBUG_DRIVER("Sandybridge: deep RC6 disabled\n");
2386 return INTEL_RC6_ENABLE;
2389 DRM_DEBUG_DRIVER("RC6 and deep RC6 enabled\n");
2390 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
2393 static void gen6_enable_rps(struct drm_device *dev)
2395 struct drm_i915_private *dev_priv = dev->dev_private;
2396 struct intel_ring_buffer *ring;
2399 u32 pcu_mbox, rc6_mask = 0;
2404 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2406 /* Here begins a magic sequence of register writes to enable
2407 * auto-downclocking.
2409 * Perhaps there might be some value in exposing these to
2412 I915_WRITE(GEN6_RC_STATE, 0);
2414 /* Clear the DBG now so we don't confuse earlier errors */
2415 if ((gtfifodbg = I915_READ(GTFIFODBG))) {
2416 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
2417 I915_WRITE(GTFIFODBG, gtfifodbg);
2420 gen6_gt_force_wake_get(dev_priv);
2422 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
2423 gt_perf_status = I915_READ(GEN6_GT_PERF_STATUS);
2425 /* In units of 100MHz */
2426 dev_priv->rps.max_delay = rp_state_cap & 0xff;
2427 dev_priv->rps.min_delay = (rp_state_cap & 0xff0000) >> 16;
2428 dev_priv->rps.cur_delay = 0;
2430 /* disable the counters and set deterministic thresholds */
2431 I915_WRITE(GEN6_RC_CONTROL, 0);
2433 I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
2434 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
2435 I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
2436 I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
2437 I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
2439 for_each_ring(ring, dev_priv, i)
2440 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
2442 I915_WRITE(GEN6_RC_SLEEP, 0);
2443 I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
2444 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
2445 I915_WRITE(GEN6_RC6p_THRESHOLD, 100000);
2446 I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
2448 /* Check if we are enabling RC6 */
2449 rc6_mode = intel_enable_rc6(dev_priv->dev);
2450 if (rc6_mode & INTEL_RC6_ENABLE)
2451 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
2453 /* We don't use those on Haswell */
2454 if (!IS_HASWELL(dev)) {
2455 if (rc6_mode & INTEL_RC6p_ENABLE)
2456 rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
2458 if (rc6_mode & INTEL_RC6pp_ENABLE)
2459 rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
2462 DRM_INFO("Enabling RC6 states: RC6 %s, RC6p %s, RC6pp %s\n",
2463 (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off",
2464 (rc6_mask & GEN6_RC_CTL_RC6p_ENABLE) ? "on" : "off",
2465 (rc6_mask & GEN6_RC_CTL_RC6pp_ENABLE) ? "on" : "off");
2467 I915_WRITE(GEN6_RC_CONTROL,
2469 GEN6_RC_CTL_EI_MODE(1) |
2470 GEN6_RC_CTL_HW_ENABLE);
2472 I915_WRITE(GEN6_RPNSWREQ,
2473 GEN6_FREQUENCY(10) |
2475 GEN6_AGGRESSIVE_TURBO);
2476 I915_WRITE(GEN6_RC_VIDEO_FREQ,
2477 GEN6_FREQUENCY(12));
2479 I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
2480 I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
2481 dev_priv->rps.max_delay << 24 |
2482 dev_priv->rps.min_delay << 16);
2484 I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
2485 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
2486 I915_WRITE(GEN6_RP_UP_EI, 66000);
2487 I915_WRITE(GEN6_RP_DOWN_EI, 350000);
2489 I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
2490 I915_WRITE(GEN6_RP_CONTROL,
2491 GEN6_RP_MEDIA_TURBO |
2492 GEN6_RP_MEDIA_HW_NORMAL_MODE |
2493 GEN6_RP_MEDIA_IS_GFX |
2495 GEN6_RP_UP_BUSY_AVG |
2496 (IS_HASWELL(dev) ? GEN7_RP_DOWN_IDLE_AVG : GEN6_RP_DOWN_IDLE_CONT));
2498 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
2500 DRM_ERROR("timeout waiting for pcode mailbox to become idle\n");
2502 I915_WRITE(GEN6_PCODE_DATA, 0);
2503 I915_WRITE(GEN6_PCODE_MAILBOX,
2505 GEN6_PCODE_WRITE_MIN_FREQ_TABLE);
2506 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
2508 DRM_ERROR("timeout waiting for pcode mailbox to finish\n");
2510 /* Check for overclock support */
2511 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
2513 DRM_ERROR("timeout waiting for pcode mailbox to become idle\n");
2514 I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_READ_OC_PARAMS);
2515 pcu_mbox = I915_READ(GEN6_PCODE_DATA);
2516 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
2518 DRM_ERROR("timeout waiting for pcode mailbox to finish\n");
2519 if (pcu_mbox & (1<<31)) { /* OC supported */
2520 dev_priv->rps.max_delay = pcu_mbox & 0xff;
2521 DRM_DEBUG_DRIVER("overclocking supported, adjusting frequency max to %dMHz\n", pcu_mbox * 50);
2524 gen6_set_rps(dev_priv->dev, (gt_perf_status & 0xff00) >> 8);
2526 /* requires MSI enabled */
2527 I915_WRITE(GEN6_PMIER, GEN6_PM_DEFERRED_EVENTS);
2528 spin_lock_irq(&dev_priv->rps.lock);
2529 WARN_ON(dev_priv->rps.pm_iir != 0);
2530 I915_WRITE(GEN6_PMIMR, 0);
2531 spin_unlock_irq(&dev_priv->rps.lock);
2532 /* enable all PM interrupts */
2533 I915_WRITE(GEN6_PMINTRMSK, 0);
2535 gen6_gt_force_wake_put(dev_priv);
2538 static void gen6_update_ring_freq(struct drm_device *dev)
2540 struct drm_i915_private *dev_priv = dev->dev_private;
2542 int gpu_freq, ia_freq, max_ia_freq;
2543 int scaling_factor = 180;
2545 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2547 max_ia_freq = cpufreq_quick_get_max(0);
2549 * Default to measured freq if none found, PCU will ensure we don't go
2553 max_ia_freq = tsc_khz;
2555 /* Convert from kHz to MHz */
2556 max_ia_freq /= 1000;
2559 * For each potential GPU frequency, load a ring frequency we'd like
2560 * to use for memory access. We do this by specifying the IA frequency
2561 * the PCU should use as a reference to determine the ring frequency.
2563 for (gpu_freq = dev_priv->rps.max_delay; gpu_freq >= dev_priv->rps.min_delay;
2565 int diff = dev_priv->rps.max_delay - gpu_freq;
2568 * For GPU frequencies less than 750MHz, just use the lowest
2571 if (gpu_freq < min_freq)
2574 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
2575 ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
2577 I915_WRITE(GEN6_PCODE_DATA,
2578 (ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT) |
2580 I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY |
2581 GEN6_PCODE_WRITE_MIN_FREQ_TABLE);
2582 if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) &
2583 GEN6_PCODE_READY) == 0, 10)) {
2584 DRM_ERROR("pcode write of freq table timed out\n");
2590 void ironlake_teardown_rc6(struct drm_device *dev)
2592 struct drm_i915_private *dev_priv = dev->dev_private;
2594 if (dev_priv->renderctx) {
2595 i915_gem_object_unpin(dev_priv->renderctx);
2596 drm_gem_object_unreference(&dev_priv->renderctx->base);
2597 dev_priv->renderctx = NULL;
2600 if (dev_priv->pwrctx) {
2601 i915_gem_object_unpin(dev_priv->pwrctx);
2602 drm_gem_object_unreference(&dev_priv->pwrctx->base);
2603 dev_priv->pwrctx = NULL;
2607 static void ironlake_disable_rc6(struct drm_device *dev)
2609 struct drm_i915_private *dev_priv = dev->dev_private;
2611 if (I915_READ(PWRCTXA)) {
2612 /* Wake the GPU, prevent RC6, then restore RSTDBYCTL */
2613 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) | RCX_SW_EXIT);
2614 wait_for(((I915_READ(RSTDBYCTL) & RSX_STATUS_MASK) == RSX_STATUS_ON),
2617 I915_WRITE(PWRCTXA, 0);
2618 POSTING_READ(PWRCTXA);
2620 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
2621 POSTING_READ(RSTDBYCTL);
2625 static int ironlake_setup_rc6(struct drm_device *dev)
2627 struct drm_i915_private *dev_priv = dev->dev_private;
2629 if (dev_priv->renderctx == NULL)
2630 dev_priv->renderctx = intel_alloc_context_page(dev);
2631 if (!dev_priv->renderctx)
2634 if (dev_priv->pwrctx == NULL)
2635 dev_priv->pwrctx = intel_alloc_context_page(dev);
2636 if (!dev_priv->pwrctx) {
2637 ironlake_teardown_rc6(dev);
2644 static void ironlake_enable_rc6(struct drm_device *dev)
2646 struct drm_i915_private *dev_priv = dev->dev_private;
2647 struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
2650 /* rc6 disabled by default due to repeated reports of hanging during
2653 if (!intel_enable_rc6(dev))
2656 WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2658 ret = ironlake_setup_rc6(dev);
2663 * GPU can automatically power down the render unit if given a page
2666 ret = intel_ring_begin(ring, 6);
2668 ironlake_teardown_rc6(dev);
2672 intel_ring_emit(ring, MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN);
2673 intel_ring_emit(ring, MI_SET_CONTEXT);
2674 intel_ring_emit(ring, dev_priv->renderctx->gtt_offset |
2676 MI_SAVE_EXT_STATE_EN |
2677 MI_RESTORE_EXT_STATE_EN |
2678 MI_RESTORE_INHIBIT);
2679 intel_ring_emit(ring, MI_SUSPEND_FLUSH);
2680 intel_ring_emit(ring, MI_NOOP);
2681 intel_ring_emit(ring, MI_FLUSH);
2682 intel_ring_advance(ring);
2685 * Wait for the command parser to advance past MI_SET_CONTEXT. The HW
2686 * does an implicit flush, combined with MI_FLUSH above, it should be
2687 * safe to assume that renderctx is valid
2689 ret = intel_wait_ring_idle(ring);
2691 DRM_ERROR("failed to enable ironlake power power savings\n");
2692 ironlake_teardown_rc6(dev);
2696 I915_WRITE(PWRCTXA, dev_priv->pwrctx->gtt_offset | PWRCTX_EN);
2697 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
2700 static unsigned long intel_pxfreq(u32 vidfreq)
2703 int div = (vidfreq & 0x3f0000) >> 16;
2704 int post = (vidfreq & 0x3000) >> 12;
2705 int pre = (vidfreq & 0x7);
2710 freq = ((div * 133333) / ((1<<post) * pre));
2715 static const struct cparams {
2721 { 1, 1333, 301, 28664 },
2722 { 1, 1066, 294, 24460 },
2723 { 1, 800, 294, 25192 },
2724 { 0, 1333, 276, 27605 },
2725 { 0, 1066, 276, 27605 },
2726 { 0, 800, 231, 23784 },
2729 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
2731 u64 total_count, diff, ret;
2732 u32 count1, count2, count3, m = 0, c = 0;
2733 unsigned long now = jiffies_to_msecs(jiffies), diff1;
2736 assert_spin_locked(&mchdev_lock);
2738 diff1 = now - dev_priv->ips.last_time1;
2740 /* Prevent division-by-zero if we are asking too fast.
2741 * Also, we don't get interesting results if we are polling
2742 * faster than once in 10ms, so just return the saved value
2746 return dev_priv->ips.chipset_power;
2748 count1 = I915_READ(DMIEC);
2749 count2 = I915_READ(DDREC);
2750 count3 = I915_READ(CSIEC);
2752 total_count = count1 + count2 + count3;
2754 /* FIXME: handle per-counter overflow */
2755 if (total_count < dev_priv->ips.last_count1) {
2756 diff = ~0UL - dev_priv->ips.last_count1;
2757 diff += total_count;
2759 diff = total_count - dev_priv->ips.last_count1;
2762 for (i = 0; i < ARRAY_SIZE(cparams); i++) {
2763 if (cparams[i].i == dev_priv->ips.c_m &&
2764 cparams[i].t == dev_priv->ips.r_t) {
2771 diff = div_u64(diff, diff1);
2772 ret = ((m * diff) + c);
2773 ret = div_u64(ret, 10);
2775 dev_priv->ips.last_count1 = total_count;
2776 dev_priv->ips.last_time1 = now;
2778 dev_priv->ips.chipset_power = ret;
2783 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
2785 unsigned long m, x, b;
2788 tsfs = I915_READ(TSFS);
2790 m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
2791 x = I915_READ8(TR1);
2793 b = tsfs & TSFS_INTR_MASK;
2795 return ((m * x) / 127) - b;
2798 static u16 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
2800 static const struct v_table {
2801 u16 vd; /* in .1 mil */
2802 u16 vm; /* in .1 mil */
2933 if (dev_priv->info->is_mobile)
2934 return v_table[pxvid].vm;
2936 return v_table[pxvid].vd;
2939 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
2941 struct timespec now, diff1;
2943 unsigned long diffms;
2946 assert_spin_locked(&mchdev_lock);
2948 getrawmonotonic(&now);
2949 diff1 = timespec_sub(now, dev_priv->ips.last_time2);
2951 /* Don't divide by 0 */
2952 diffms = diff1.tv_sec * 1000 + diff1.tv_nsec / 1000000;
2956 count = I915_READ(GFXEC);
2958 if (count < dev_priv->ips.last_count2) {
2959 diff = ~0UL - dev_priv->ips.last_count2;
2962 diff = count - dev_priv->ips.last_count2;
2965 dev_priv->ips.last_count2 = count;
2966 dev_priv->ips.last_time2 = now;
2968 /* More magic constants... */
2970 diff = div_u64(diff, diffms * 10);
2971 dev_priv->ips.gfx_power = diff;
2974 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
2976 if (dev_priv->info->gen != 5)
2979 spin_lock_irq(&mchdev_lock);
2981 __i915_update_gfx_val(dev_priv);
2983 spin_unlock_irq(&mchdev_lock);
2986 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
2988 unsigned long t, corr, state1, corr2, state2;
2991 assert_spin_locked(&mchdev_lock);
2993 pxvid = I915_READ(PXVFREQ_BASE + (dev_priv->rps.cur_delay * 4));
2994 pxvid = (pxvid >> 24) & 0x7f;
2995 ext_v = pvid_to_extvid(dev_priv, pxvid);
2999 t = i915_mch_val(dev_priv);
3001 /* Revel in the empirically derived constants */
3003 /* Correction factor in 1/100000 units */
3005 corr = ((t * 2349) + 135940);
3007 corr = ((t * 964) + 29317);
3009 corr = ((t * 301) + 1004);
3011 corr = corr * ((150142 * state1) / 10000 - 78642);
3013 corr2 = (corr * dev_priv->ips.corr);
3015 state2 = (corr2 * state1) / 10000;
3016 state2 /= 100; /* convert to mW */
3018 __i915_update_gfx_val(dev_priv);
3020 return dev_priv->ips.gfx_power + state2;
3024 * i915_read_mch_val - return value for IPS use
3026 * Calculate and return a value for the IPS driver to use when deciding whether
3027 * we have thermal and power headroom to increase CPU or GPU power budget.
3029 unsigned long i915_read_mch_val(void)
3031 struct drm_i915_private *dev_priv;
3032 unsigned long chipset_val, graphics_val, ret = 0;
3034 spin_lock_irq(&mchdev_lock);
3037 dev_priv = i915_mch_dev;
3039 chipset_val = i915_chipset_val(dev_priv);
3040 graphics_val = i915_gfx_val(dev_priv);
3042 ret = chipset_val + graphics_val;
3045 spin_unlock_irq(&mchdev_lock);
3049 EXPORT_SYMBOL_GPL(i915_read_mch_val);
3052 * i915_gpu_raise - raise GPU frequency limit
3054 * Raise the limit; IPS indicates we have thermal headroom.
3056 bool i915_gpu_raise(void)
3058 struct drm_i915_private *dev_priv;
3061 spin_lock_irq(&mchdev_lock);
3062 if (!i915_mch_dev) {
3066 dev_priv = i915_mch_dev;
3068 if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
3069 dev_priv->ips.max_delay--;
3072 spin_unlock_irq(&mchdev_lock);
3076 EXPORT_SYMBOL_GPL(i915_gpu_raise);
3079 * i915_gpu_lower - lower GPU frequency limit
3081 * IPS indicates we're close to a thermal limit, so throttle back the GPU
3082 * frequency maximum.
3084 bool i915_gpu_lower(void)
3086 struct drm_i915_private *dev_priv;
3089 spin_lock_irq(&mchdev_lock);
3090 if (!i915_mch_dev) {
3094 dev_priv = i915_mch_dev;
3096 if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
3097 dev_priv->ips.max_delay++;
3100 spin_unlock_irq(&mchdev_lock);
3104 EXPORT_SYMBOL_GPL(i915_gpu_lower);
3107 * i915_gpu_busy - indicate GPU business to IPS
3109 * Tell the IPS driver whether or not the GPU is busy.
3111 bool i915_gpu_busy(void)
3113 struct drm_i915_private *dev_priv;
3114 struct intel_ring_buffer *ring;
3118 spin_lock_irq(&mchdev_lock);
3121 dev_priv = i915_mch_dev;
3123 for_each_ring(ring, dev_priv, i)
3124 ret |= !list_empty(&ring->request_list);
3127 spin_unlock_irq(&mchdev_lock);
3131 EXPORT_SYMBOL_GPL(i915_gpu_busy);
3134 * i915_gpu_turbo_disable - disable graphics turbo
3136 * Disable graphics turbo by resetting the max frequency and setting the
3137 * current frequency to the default.
3139 bool i915_gpu_turbo_disable(void)
3141 struct drm_i915_private *dev_priv;
3144 spin_lock_irq(&mchdev_lock);
3145 if (!i915_mch_dev) {
3149 dev_priv = i915_mch_dev;
3151 dev_priv->ips.max_delay = dev_priv->ips.fstart;
3153 if (!ironlake_set_drps(dev_priv->dev, dev_priv->ips.fstart))
3157 spin_unlock_irq(&mchdev_lock);
3161 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
3164 * Tells the intel_ips driver that the i915 driver is now loaded, if
3165 * IPS got loaded first.
3167 * This awkward dance is so that neither module has to depend on the
3168 * other in order for IPS to do the appropriate communication of
3169 * GPU turbo limits to i915.
3172 ips_ping_for_i915_load(void)
3176 link = symbol_get(ips_link_to_i915_driver);
3179 symbol_put(ips_link_to_i915_driver);
3183 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
3185 /* We only register the i915 ips part with intel-ips once everything is
3186 * set up, to avoid intel-ips sneaking in and reading bogus values. */
3187 spin_lock_irq(&mchdev_lock);
3188 i915_mch_dev = dev_priv;
3189 spin_unlock_irq(&mchdev_lock);
3191 ips_ping_for_i915_load();
3194 void intel_gpu_ips_teardown(void)
3196 spin_lock_irq(&mchdev_lock);
3197 i915_mch_dev = NULL;
3198 spin_unlock_irq(&mchdev_lock);
3200 static void intel_init_emon(struct drm_device *dev)
3202 struct drm_i915_private *dev_priv = dev->dev_private;
3207 /* Disable to program */
3211 /* Program energy weights for various events */
3212 I915_WRITE(SDEW, 0x15040d00);
3213 I915_WRITE(CSIEW0, 0x007f0000);
3214 I915_WRITE(CSIEW1, 0x1e220004);
3215 I915_WRITE(CSIEW2, 0x04000004);
3217 for (i = 0; i < 5; i++)
3218 I915_WRITE(PEW + (i * 4), 0);
3219 for (i = 0; i < 3; i++)
3220 I915_WRITE(DEW + (i * 4), 0);
3222 /* Program P-state weights to account for frequency power adjustment */
3223 for (i = 0; i < 16; i++) {
3224 u32 pxvidfreq = I915_READ(PXVFREQ_BASE + (i * 4));
3225 unsigned long freq = intel_pxfreq(pxvidfreq);
3226 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
3231 val *= (freq / 1000);
3233 val /= (127*127*900);
3235 DRM_ERROR("bad pxval: %ld\n", val);
3238 /* Render standby states get 0 weight */
3242 for (i = 0; i < 4; i++) {
3243 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
3244 (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
3245 I915_WRITE(PXW + (i * 4), val);
3248 /* Adjust magic regs to magic values (more experimental results) */
3249 I915_WRITE(OGW0, 0);
3250 I915_WRITE(OGW1, 0);
3251 I915_WRITE(EG0, 0x00007f00);
3252 I915_WRITE(EG1, 0x0000000e);
3253 I915_WRITE(EG2, 0x000e0000);
3254 I915_WRITE(EG3, 0x68000300);
3255 I915_WRITE(EG4, 0x42000000);
3256 I915_WRITE(EG5, 0x00140031);
3260 for (i = 0; i < 8; i++)
3261 I915_WRITE(PXWL + (i * 4), 0);
3263 /* Enable PMON + select events */
3264 I915_WRITE(ECR, 0x80000019);
3266 lcfuse = I915_READ(LCFUSE02);
3268 dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
3271 void intel_disable_gt_powersave(struct drm_device *dev)
3273 if (IS_IRONLAKE_M(dev)) {
3274 ironlake_disable_drps(dev);
3275 ironlake_disable_rc6(dev);
3276 } else if (INTEL_INFO(dev)->gen >= 6 && !IS_VALLEYVIEW(dev)) {
3277 gen6_disable_rps(dev);
3281 void intel_enable_gt_powersave(struct drm_device *dev)
3283 if (IS_IRONLAKE_M(dev)) {
3284 ironlake_enable_drps(dev);
3285 ironlake_enable_rc6(dev);
3286 intel_init_emon(dev);
3287 } else if ((IS_GEN6(dev) || IS_GEN7(dev)) && !IS_VALLEYVIEW(dev)) {
3288 gen6_enable_rps(dev);
3289 gen6_update_ring_freq(dev);
3293 static void ironlake_init_clock_gating(struct drm_device *dev)
3295 struct drm_i915_private *dev_priv = dev->dev_private;
3296 uint32_t dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE;
3298 /* Required for FBC */
3299 dspclk_gate |= DPFCUNIT_CLOCK_GATE_DISABLE |
3300 DPFCRUNIT_CLOCK_GATE_DISABLE |
3301 DPFDUNIT_CLOCK_GATE_DISABLE;
3302 /* Required for CxSR */
3303 dspclk_gate |= DPARBUNIT_CLOCK_GATE_DISABLE;
3305 I915_WRITE(PCH_3DCGDIS0,
3306 MARIUNIT_CLOCK_GATE_DISABLE |
3307 SVSMUNIT_CLOCK_GATE_DISABLE);
3308 I915_WRITE(PCH_3DCGDIS1,
3309 VFMUNIT_CLOCK_GATE_DISABLE);
3311 I915_WRITE(PCH_DSPCLK_GATE_D, dspclk_gate);
3314 * According to the spec the following bits should be set in
3315 * order to enable memory self-refresh
3316 * The bit 22/21 of 0x42004
3317 * The bit 5 of 0x42020
3318 * The bit 15 of 0x45000
3320 I915_WRITE(ILK_DISPLAY_CHICKEN2,
3321 (I915_READ(ILK_DISPLAY_CHICKEN2) |
3322 ILK_DPARB_GATE | ILK_VSDPFD_FULL));
3323 I915_WRITE(ILK_DSPCLK_GATE,
3324 (I915_READ(ILK_DSPCLK_GATE) |
3325 ILK_DPARB_CLK_GATE));
3326 I915_WRITE(DISP_ARB_CTL,
3327 (I915_READ(DISP_ARB_CTL) |
3329 I915_WRITE(WM3_LP_ILK, 0);
3330 I915_WRITE(WM2_LP_ILK, 0);
3331 I915_WRITE(WM1_LP_ILK, 0);
3334 * Based on the document from hardware guys the following bits
3335 * should be set unconditionally in order to enable FBC.
3336 * The bit 22 of 0x42000
3337 * The bit 22 of 0x42004
3338 * The bit 7,8,9 of 0x42020.
3340 if (IS_IRONLAKE_M(dev)) {
3341 I915_WRITE(ILK_DISPLAY_CHICKEN1,
3342 I915_READ(ILK_DISPLAY_CHICKEN1) |
3344 I915_WRITE(ILK_DISPLAY_CHICKEN2,
3345 I915_READ(ILK_DISPLAY_CHICKEN2) |
3347 I915_WRITE(ILK_DSPCLK_GATE,
3348 I915_READ(ILK_DSPCLK_GATE) |
3354 I915_WRITE(ILK_DISPLAY_CHICKEN2,
3355 I915_READ(ILK_DISPLAY_CHICKEN2) |
3356 ILK_ELPIN_409_SELECT);
3357 I915_WRITE(_3D_CHICKEN2,
3358 _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
3359 _3D_CHICKEN2_WM_READ_PIPELINED);
3362 static void gen6_init_clock_gating(struct drm_device *dev)
3364 struct drm_i915_private *dev_priv = dev->dev_private;
3366 uint32_t dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE;
3368 I915_WRITE(PCH_DSPCLK_GATE_D, dspclk_gate);
3370 I915_WRITE(ILK_DISPLAY_CHICKEN2,
3371 I915_READ(ILK_DISPLAY_CHICKEN2) |
3372 ILK_ELPIN_409_SELECT);
3374 I915_WRITE(WM3_LP_ILK, 0);
3375 I915_WRITE(WM2_LP_ILK, 0);
3376 I915_WRITE(WM1_LP_ILK, 0);
3378 I915_WRITE(CACHE_MODE_0,
3379 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
3381 I915_WRITE(GEN6_UCGCTL1,
3382 I915_READ(GEN6_UCGCTL1) |
3383 GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
3384 GEN6_CSUNIT_CLOCK_GATE_DISABLE);
3386 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
3387 * gating disable must be set. Failure to set it results in
3388 * flickering pixels due to Z write ordering failures after
3389 * some amount of runtime in the Mesa "fire" demo, and Unigine
3390 * Sanctuary and Tropics, and apparently anything else with
3391 * alpha test or pixel discard.
3393 * According to the spec, bit 11 (RCCUNIT) must also be set,
3394 * but we didn't debug actual testcases to find it out.
3396 * Also apply WaDisableVDSUnitClockGating and
3397 * WaDisableRCPBUnitClockGating.
3399 I915_WRITE(GEN6_UCGCTL2,
3400 GEN7_VDSUNIT_CLOCK_GATE_DISABLE |
3401 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
3402 GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
3404 /* Bspec says we need to always set all mask bits. */
3405 I915_WRITE(_3D_CHICKEN, (0xFFFF << 16) |
3406 _3D_CHICKEN_SF_DISABLE_FASTCLIP_CULL);
3409 * According to the spec the following bits should be
3410 * set in order to enable memory self-refresh and fbc:
3411 * The bit21 and bit22 of 0x42000
3412 * The bit21 and bit22 of 0x42004
3413 * The bit5 and bit7 of 0x42020
3414 * The bit14 of 0x70180
3415 * The bit14 of 0x71180
3417 I915_WRITE(ILK_DISPLAY_CHICKEN1,
3418 I915_READ(ILK_DISPLAY_CHICKEN1) |
3419 ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
3420 I915_WRITE(ILK_DISPLAY_CHICKEN2,
3421 I915_READ(ILK_DISPLAY_CHICKEN2) |
3422 ILK_DPARB_GATE | ILK_VSDPFD_FULL);
3423 I915_WRITE(ILK_DSPCLK_GATE,
3424 I915_READ(ILK_DSPCLK_GATE) |
3425 ILK_DPARB_CLK_GATE |
3428 I915_WRITE(GEN6_MBCTL, I915_READ(GEN6_MBCTL) |
3429 GEN6_MBCTL_ENABLE_BOOT_FETCH);
3431 for_each_pipe(pipe) {
3432 I915_WRITE(DSPCNTR(pipe),
3433 I915_READ(DSPCNTR(pipe)) |
3434 DISPPLANE_TRICKLE_FEED_DISABLE);
3435 intel_flush_display_plane(dev_priv, pipe);
3439 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
3441 uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
3443 reg &= ~GEN7_FF_SCHED_MASK;
3444 reg |= GEN7_FF_TS_SCHED_HW;
3445 reg |= GEN7_FF_VS_SCHED_HW;
3446 reg |= GEN7_FF_DS_SCHED_HW;
3448 I915_WRITE(GEN7_FF_THREAD_MODE, reg);
3451 static void haswell_init_clock_gating(struct drm_device *dev)
3453 struct drm_i915_private *dev_priv = dev->dev_private;
3455 uint32_t dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE;
3457 I915_WRITE(PCH_DSPCLK_GATE_D, dspclk_gate);
3459 I915_WRITE(WM3_LP_ILK, 0);
3460 I915_WRITE(WM2_LP_ILK, 0);
3461 I915_WRITE(WM1_LP_ILK, 0);
3463 /* According to the spec, bit 13 (RCZUNIT) must be set on IVB.
3464 * This implements the WaDisableRCZUnitClockGating workaround.
3466 I915_WRITE(GEN6_UCGCTL2, GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
3468 I915_WRITE(ILK_DSPCLK_GATE, IVB_VRHUNIT_CLK_GATE);
3470 I915_WRITE(IVB_CHICKEN3,
3471 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
3472 CHICKEN3_DGMG_DONE_FIX_DISABLE);
3474 /* Apply the WaDisableRHWOOptimizationForRenderHang workaround. */
3475 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
3476 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
3478 /* WaApplyL3ControlAndL3ChickenMode requires those two on Ivy Bridge */
3479 I915_WRITE(GEN7_L3CNTLREG1,
3480 GEN7_WA_FOR_GEN7_L3_CONTROL);
3481 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
3482 GEN7_WA_L3_CHICKEN_MODE);
3484 /* This is required by WaCatErrorRejectionIssue */
3485 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
3486 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
3487 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
3489 for_each_pipe(pipe) {
3490 I915_WRITE(DSPCNTR(pipe),
3491 I915_READ(DSPCNTR(pipe)) |
3492 DISPPLANE_TRICKLE_FEED_DISABLE);
3493 intel_flush_display_plane(dev_priv, pipe);
3496 gen7_setup_fixed_func_scheduler(dev_priv);
3498 /* WaDisable4x2SubspanOptimization */
3499 I915_WRITE(CACHE_MODE_1,
3500 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
3502 /* XXX: This is a workaround for early silicon revisions and should be
3507 WM_DBG_DISALLOW_MULTIPLE_LP |
3508 WM_DBG_DISALLOW_SPRITE |
3509 WM_DBG_DISALLOW_MAXFIFO);
3513 static void ivybridge_init_clock_gating(struct drm_device *dev)
3515 struct drm_i915_private *dev_priv = dev->dev_private;
3517 uint32_t dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE;
3520 I915_WRITE(PCH_DSPCLK_GATE_D, dspclk_gate);
3522 I915_WRITE(WM3_LP_ILK, 0);
3523 I915_WRITE(WM2_LP_ILK, 0);
3524 I915_WRITE(WM1_LP_ILK, 0);
3526 I915_WRITE(ILK_DSPCLK_GATE, IVB_VRHUNIT_CLK_GATE);
3528 I915_WRITE(IVB_CHICKEN3,
3529 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
3530 CHICKEN3_DGMG_DONE_FIX_DISABLE);
3532 /* Apply the WaDisableRHWOOptimizationForRenderHang workaround. */
3533 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
3534 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
3536 /* WaApplyL3ControlAndL3ChickenMode requires those two on Ivy Bridge */
3537 I915_WRITE(GEN7_L3CNTLREG1,
3538 GEN7_WA_FOR_GEN7_L3_CONTROL);
3539 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
3540 GEN7_WA_L3_CHICKEN_MODE);
3542 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
3543 * gating disable must be set. Failure to set it results in
3544 * flickering pixels due to Z write ordering failures after
3545 * some amount of runtime in the Mesa "fire" demo, and Unigine
3546 * Sanctuary and Tropics, and apparently anything else with
3547 * alpha test or pixel discard.
3549 * According to the spec, bit 11 (RCCUNIT) must also be set,
3550 * but we didn't debug actual testcases to find it out.
3552 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
3553 * This implements the WaDisableRCZUnitClockGating workaround.
3555 I915_WRITE(GEN6_UCGCTL2,
3556 GEN6_RCZUNIT_CLOCK_GATE_DISABLE |
3557 GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
3559 /* This is required by WaCatErrorRejectionIssue */
3560 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
3561 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
3562 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
3564 for_each_pipe(pipe) {
3565 I915_WRITE(DSPCNTR(pipe),
3566 I915_READ(DSPCNTR(pipe)) |
3567 DISPPLANE_TRICKLE_FEED_DISABLE);
3568 intel_flush_display_plane(dev_priv, pipe);
3571 I915_WRITE(GEN6_MBCTL, I915_READ(GEN6_MBCTL) |
3572 GEN6_MBCTL_ENABLE_BOOT_FETCH);
3574 gen7_setup_fixed_func_scheduler(dev_priv);
3576 /* WaDisable4x2SubspanOptimization */
3577 I915_WRITE(CACHE_MODE_1,
3578 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
3580 snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
3581 snpcr &= ~GEN6_MBC_SNPCR_MASK;
3582 snpcr |= GEN6_MBC_SNPCR_MED;
3583 I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
3586 static void valleyview_init_clock_gating(struct drm_device *dev)
3588 struct drm_i915_private *dev_priv = dev->dev_private;
3590 uint32_t dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE;
3592 I915_WRITE(PCH_DSPCLK_GATE_D, dspclk_gate);
3594 I915_WRITE(WM3_LP_ILK, 0);
3595 I915_WRITE(WM2_LP_ILK, 0);
3596 I915_WRITE(WM1_LP_ILK, 0);
3598 I915_WRITE(ILK_DSPCLK_GATE, IVB_VRHUNIT_CLK_GATE);
3600 I915_WRITE(IVB_CHICKEN3,
3601 CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
3602 CHICKEN3_DGMG_DONE_FIX_DISABLE);
3604 /* Apply the WaDisableRHWOOptimizationForRenderHang workaround. */
3605 I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
3606 GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
3608 /* WaApplyL3ControlAndL3ChickenMode requires those two on Ivy Bridge */
3609 I915_WRITE(GEN7_L3CNTLREG1, GEN7_WA_FOR_GEN7_L3_CONTROL);
3610 I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER, GEN7_WA_L3_CHICKEN_MODE);
3612 /* This is required by WaCatErrorRejectionIssue */
3613 I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
3614 I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
3615 GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
3617 I915_WRITE(GEN6_MBCTL, I915_READ(GEN6_MBCTL) |
3618 GEN6_MBCTL_ENABLE_BOOT_FETCH);
3621 /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
3622 * gating disable must be set. Failure to set it results in
3623 * flickering pixels due to Z write ordering failures after
3624 * some amount of runtime in the Mesa "fire" demo, and Unigine
3625 * Sanctuary and Tropics, and apparently anything else with
3626 * alpha test or pixel discard.
3628 * According to the spec, bit 11 (RCCUNIT) must also be set,
3629 * but we didn't debug actual testcases to find it out.
3631 * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
3632 * This implements the WaDisableRCZUnitClockGating workaround.
3634 * Also apply WaDisableVDSUnitClockGating and
3635 * WaDisableRCPBUnitClockGating.
3637 I915_WRITE(GEN6_UCGCTL2,
3638 GEN7_VDSUNIT_CLOCK_GATE_DISABLE |
3639 GEN7_TDLUNIT_CLOCK_GATE_DISABLE |
3640 GEN6_RCZUNIT_CLOCK_GATE_DISABLE |
3641 GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
3642 GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
3644 I915_WRITE(GEN7_UCGCTL4, GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
3646 for_each_pipe(pipe) {
3647 I915_WRITE(DSPCNTR(pipe),
3648 I915_READ(DSPCNTR(pipe)) |
3649 DISPPLANE_TRICKLE_FEED_DISABLE);
3650 intel_flush_display_plane(dev_priv, pipe);
3653 I915_WRITE(CACHE_MODE_1,
3654 _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
3657 * On ValleyView, the GUnit needs to signal the GT
3658 * when flip and other events complete. So enable
3659 * all the GUnit->GT interrupts here
3661 I915_WRITE(VLV_DPFLIPSTAT, PIPEB_LINE_COMPARE_INT_EN |
3662 PIPEB_HLINE_INT_EN | PIPEB_VBLANK_INT_EN |
3663 SPRITED_FLIPDONE_INT_EN | SPRITEC_FLIPDONE_INT_EN |
3664 PLANEB_FLIPDONE_INT_EN | PIPEA_LINE_COMPARE_INT_EN |
3665 PIPEA_HLINE_INT_EN | PIPEA_VBLANK_INT_EN |
3666 SPRITEB_FLIPDONE_INT_EN | SPRITEA_FLIPDONE_INT_EN |
3667 PLANEA_FLIPDONE_INT_EN);
3670 static void g4x_init_clock_gating(struct drm_device *dev)
3672 struct drm_i915_private *dev_priv = dev->dev_private;
3673 uint32_t dspclk_gate;
3675 I915_WRITE(RENCLK_GATE_D1, 0);
3676 I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
3677 GS_UNIT_CLOCK_GATE_DISABLE |
3678 CL_UNIT_CLOCK_GATE_DISABLE);
3679 I915_WRITE(RAMCLK_GATE_D, 0);
3680 dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
3681 OVRUNIT_CLOCK_GATE_DISABLE |
3682 OVCUNIT_CLOCK_GATE_DISABLE;
3684 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
3685 I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
3688 static void crestline_init_clock_gating(struct drm_device *dev)
3690 struct drm_i915_private *dev_priv = dev->dev_private;
3692 I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
3693 I915_WRITE(RENCLK_GATE_D2, 0);
3694 I915_WRITE(DSPCLK_GATE_D, 0);
3695 I915_WRITE(RAMCLK_GATE_D, 0);
3696 I915_WRITE16(DEUC, 0);
3699 static void broadwater_init_clock_gating(struct drm_device *dev)
3701 struct drm_i915_private *dev_priv = dev->dev_private;
3703 I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
3704 I965_RCC_CLOCK_GATE_DISABLE |
3705 I965_RCPB_CLOCK_GATE_DISABLE |
3706 I965_ISC_CLOCK_GATE_DISABLE |
3707 I965_FBC_CLOCK_GATE_DISABLE);
3708 I915_WRITE(RENCLK_GATE_D2, 0);
3711 static void gen3_init_clock_gating(struct drm_device *dev)
3713 struct drm_i915_private *dev_priv = dev->dev_private;
3714 u32 dstate = I915_READ(D_STATE);
3716 dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
3717 DSTATE_DOT_CLOCK_GATING;
3718 I915_WRITE(D_STATE, dstate);
3720 if (IS_PINEVIEW(dev))
3721 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
3724 static void i85x_init_clock_gating(struct drm_device *dev)
3726 struct drm_i915_private *dev_priv = dev->dev_private;
3728 I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
3731 static void i830_init_clock_gating(struct drm_device *dev)
3733 struct drm_i915_private *dev_priv = dev->dev_private;
3735 I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
3738 static void ibx_init_clock_gating(struct drm_device *dev)
3740 struct drm_i915_private *dev_priv = dev->dev_private;
3743 * On Ibex Peak and Cougar Point, we need to disable clock
3744 * gating for the panel power sequencer or it will fail to
3745 * start up when no ports are active.
3747 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
3750 static void cpt_init_clock_gating(struct drm_device *dev)
3752 struct drm_i915_private *dev_priv = dev->dev_private;
3756 * On Ibex Peak and Cougar Point, we need to disable clock
3757 * gating for the panel power sequencer or it will fail to
3758 * start up when no ports are active.
3760 I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
3761 I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
3762 DPLS_EDP_PPS_FIX_DIS);
3763 /* Without this, mode sets may fail silently on FDI */
3765 I915_WRITE(TRANS_CHICKEN2(pipe), TRANS_AUTOTRAIN_GEN_STALL_DIS);
3768 void intel_init_clock_gating(struct drm_device *dev)
3770 struct drm_i915_private *dev_priv = dev->dev_private;
3772 dev_priv->display.init_clock_gating(dev);
3774 if (dev_priv->display.init_pch_clock_gating)
3775 dev_priv->display.init_pch_clock_gating(dev);
3778 /* Starting with Haswell, we have different power wells for
3779 * different parts of the GPU. This attempts to enable them all.
3781 void intel_init_power_wells(struct drm_device *dev)
3783 struct drm_i915_private *dev_priv = dev->dev_private;
3784 unsigned long power_wells[] = {
3791 if (!IS_HASWELL(dev))
3794 mutex_lock(&dev->struct_mutex);
3796 for (i = 0; i < ARRAY_SIZE(power_wells); i++) {
3797 int well = I915_READ(power_wells[i]);
3799 if ((well & HSW_PWR_WELL_STATE) == 0) {
3800 I915_WRITE(power_wells[i], well & HSW_PWR_WELL_ENABLE);
3801 if (wait_for(I915_READ(power_wells[i] & HSW_PWR_WELL_STATE), 20))
3802 DRM_ERROR("Error enabling power well %lx\n", power_wells[i]);
3806 mutex_unlock(&dev->struct_mutex);
3809 /* Set up chip specific power management-related functions */
3810 void intel_init_pm(struct drm_device *dev)
3812 struct drm_i915_private *dev_priv = dev->dev_private;
3814 if (I915_HAS_FBC(dev)) {
3815 if (HAS_PCH_SPLIT(dev)) {
3816 dev_priv->display.fbc_enabled = ironlake_fbc_enabled;
3817 dev_priv->display.enable_fbc = ironlake_enable_fbc;
3818 dev_priv->display.disable_fbc = ironlake_disable_fbc;
3819 } else if (IS_GM45(dev)) {
3820 dev_priv->display.fbc_enabled = g4x_fbc_enabled;
3821 dev_priv->display.enable_fbc = g4x_enable_fbc;
3822 dev_priv->display.disable_fbc = g4x_disable_fbc;
3823 } else if (IS_CRESTLINE(dev)) {
3824 dev_priv->display.fbc_enabled = i8xx_fbc_enabled;
3825 dev_priv->display.enable_fbc = i8xx_enable_fbc;
3826 dev_priv->display.disable_fbc = i8xx_disable_fbc;
3828 /* 855GM needs testing */
3832 if (IS_PINEVIEW(dev))
3833 i915_pineview_get_mem_freq(dev);
3834 else if (IS_GEN5(dev))
3835 i915_ironlake_get_mem_freq(dev);
3837 /* For FIFO watermark updates */
3838 if (HAS_PCH_SPLIT(dev)) {
3839 if (HAS_PCH_IBX(dev))
3840 dev_priv->display.init_pch_clock_gating = ibx_init_clock_gating;
3841 else if (HAS_PCH_CPT(dev))
3842 dev_priv->display.init_pch_clock_gating = cpt_init_clock_gating;
3845 if (I915_READ(MLTR_ILK) & ILK_SRLT_MASK)
3846 dev_priv->display.update_wm = ironlake_update_wm;
3848 DRM_DEBUG_KMS("Failed to get proper latency. "
3850 dev_priv->display.update_wm = NULL;
3852 dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
3853 } else if (IS_GEN6(dev)) {
3854 if (SNB_READ_WM0_LATENCY()) {
3855 dev_priv->display.update_wm = sandybridge_update_wm;
3856 dev_priv->display.update_sprite_wm = sandybridge_update_sprite_wm;
3858 DRM_DEBUG_KMS("Failed to read display plane latency. "
3860 dev_priv->display.update_wm = NULL;
3862 dev_priv->display.init_clock_gating = gen6_init_clock_gating;
3863 } else if (IS_IVYBRIDGE(dev)) {
3864 /* FIXME: detect B0+ stepping and use auto training */
3865 if (SNB_READ_WM0_LATENCY()) {
3866 dev_priv->display.update_wm = sandybridge_update_wm;
3867 dev_priv->display.update_sprite_wm = sandybridge_update_sprite_wm;
3869 DRM_DEBUG_KMS("Failed to read display plane latency. "
3871 dev_priv->display.update_wm = NULL;
3873 dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
3874 } else if (IS_HASWELL(dev)) {
3875 if (SNB_READ_WM0_LATENCY()) {
3876 dev_priv->display.update_wm = sandybridge_update_wm;
3877 dev_priv->display.update_sprite_wm = sandybridge_update_sprite_wm;
3878 dev_priv->display.update_linetime_wm = haswell_update_linetime_wm;
3880 DRM_DEBUG_KMS("Failed to read display plane latency. "
3882 dev_priv->display.update_wm = NULL;
3884 dev_priv->display.init_clock_gating = haswell_init_clock_gating;
3886 dev_priv->display.update_wm = NULL;
3887 } else if (IS_VALLEYVIEW(dev)) {
3888 dev_priv->display.update_wm = valleyview_update_wm;
3889 dev_priv->display.init_clock_gating =
3890 valleyview_init_clock_gating;
3891 } else if (IS_PINEVIEW(dev)) {
3892 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
3895 dev_priv->mem_freq)) {
3896 DRM_INFO("failed to find known CxSR latency "
3897 "(found ddr%s fsb freq %d, mem freq %d), "
3899 (dev_priv->is_ddr3 == 1) ? "3" : "2",
3900 dev_priv->fsb_freq, dev_priv->mem_freq);
3901 /* Disable CxSR and never update its watermark again */
3902 pineview_disable_cxsr(dev);
3903 dev_priv->display.update_wm = NULL;
3905 dev_priv->display.update_wm = pineview_update_wm;
3906 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
3907 } else if (IS_G4X(dev)) {
3908 dev_priv->display.update_wm = g4x_update_wm;
3909 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
3910 } else if (IS_GEN4(dev)) {
3911 dev_priv->display.update_wm = i965_update_wm;
3912 if (IS_CRESTLINE(dev))
3913 dev_priv->display.init_clock_gating = crestline_init_clock_gating;
3914 else if (IS_BROADWATER(dev))
3915 dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
3916 } else if (IS_GEN3(dev)) {
3917 dev_priv->display.update_wm = i9xx_update_wm;
3918 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
3919 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
3920 } else if (IS_I865G(dev)) {
3921 dev_priv->display.update_wm = i830_update_wm;
3922 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
3923 dev_priv->display.get_fifo_size = i830_get_fifo_size;
3924 } else if (IS_I85X(dev)) {
3925 dev_priv->display.update_wm = i9xx_update_wm;
3926 dev_priv->display.get_fifo_size = i85x_get_fifo_size;
3927 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
3929 dev_priv->display.update_wm = i830_update_wm;
3930 dev_priv->display.init_clock_gating = i830_init_clock_gating;
3932 dev_priv->display.get_fifo_size = i845_get_fifo_size;
3934 dev_priv->display.get_fifo_size = i830_get_fifo_size;
3938 static void __gen6_gt_wait_for_thread_c0(struct drm_i915_private *dev_priv)
3940 u32 gt_thread_status_mask;
3942 if (IS_HASWELL(dev_priv->dev))
3943 gt_thread_status_mask = GEN6_GT_THREAD_STATUS_CORE_MASK_HSW;
3945 gt_thread_status_mask = GEN6_GT_THREAD_STATUS_CORE_MASK;
3947 /* w/a for a sporadic read returning 0 by waiting for the GT
3948 * thread to wake up.
3950 if (wait_for_atomic_us((I915_READ_NOTRACE(GEN6_GT_THREAD_STATUS_REG) & gt_thread_status_mask) == 0, 500))
3951 DRM_ERROR("GT thread status wait timed out\n");
3954 static void __gen6_gt_force_wake_get(struct drm_i915_private *dev_priv)
3958 if (IS_HASWELL(dev_priv->dev))
3959 forcewake_ack = FORCEWAKE_ACK_HSW;
3961 forcewake_ack = FORCEWAKE_ACK;
3963 if (wait_for_atomic_us((I915_READ_NOTRACE(forcewake_ack) & 1) == 0, 500))
3964 DRM_ERROR("Force wake wait timed out\n");
3966 I915_WRITE_NOTRACE(FORCEWAKE, 1);
3967 POSTING_READ(FORCEWAKE);
3969 if (wait_for_atomic_us((I915_READ_NOTRACE(forcewake_ack) & 1), 500))
3970 DRM_ERROR("Force wake wait timed out\n");
3972 __gen6_gt_wait_for_thread_c0(dev_priv);
3975 static void __gen6_gt_force_wake_mt_get(struct drm_i915_private *dev_priv)
3979 if (IS_HASWELL(dev_priv->dev))
3980 forcewake_ack = FORCEWAKE_ACK_HSW;
3982 forcewake_ack = FORCEWAKE_MT_ACK;
3984 if (wait_for_atomic_us((I915_READ_NOTRACE(forcewake_ack) & 1) == 0, 500))
3985 DRM_ERROR("Force wake wait timed out\n");
3987 I915_WRITE_NOTRACE(FORCEWAKE_MT, _MASKED_BIT_ENABLE(1));
3988 POSTING_READ(FORCEWAKE_MT);
3990 if (wait_for_atomic_us((I915_READ_NOTRACE(forcewake_ack) & 1), 500))
3991 DRM_ERROR("Force wake wait timed out\n");
3993 __gen6_gt_wait_for_thread_c0(dev_priv);
3997 * Generally this is called implicitly by the register read function. However,
3998 * if some sequence requires the GT to not power down then this function should
3999 * be called at the beginning of the sequence followed by a call to
4000 * gen6_gt_force_wake_put() at the end of the sequence.
4002 void gen6_gt_force_wake_get(struct drm_i915_private *dev_priv)
4004 unsigned long irqflags;
4006 spin_lock_irqsave(&dev_priv->gt_lock, irqflags);
4007 if (dev_priv->forcewake_count++ == 0)
4008 dev_priv->gt.force_wake_get(dev_priv);
4009 spin_unlock_irqrestore(&dev_priv->gt_lock, irqflags);
4012 void gen6_gt_check_fifodbg(struct drm_i915_private *dev_priv)
4015 gtfifodbg = I915_READ_NOTRACE(GTFIFODBG);
4016 if (WARN(gtfifodbg & GT_FIFO_CPU_ERROR_MASK,
4017 "MMIO read or write has been dropped %x\n", gtfifodbg))
4018 I915_WRITE_NOTRACE(GTFIFODBG, GT_FIFO_CPU_ERROR_MASK);
4021 static void __gen6_gt_force_wake_put(struct drm_i915_private *dev_priv)
4023 I915_WRITE_NOTRACE(FORCEWAKE, 0);
4024 POSTING_READ(FORCEWAKE);
4025 gen6_gt_check_fifodbg(dev_priv);
4028 static void __gen6_gt_force_wake_mt_put(struct drm_i915_private *dev_priv)
4030 I915_WRITE_NOTRACE(FORCEWAKE_MT, _MASKED_BIT_DISABLE(1));
4031 POSTING_READ(FORCEWAKE_MT);
4032 gen6_gt_check_fifodbg(dev_priv);
4036 * see gen6_gt_force_wake_get()
4038 void gen6_gt_force_wake_put(struct drm_i915_private *dev_priv)
4040 unsigned long irqflags;
4042 spin_lock_irqsave(&dev_priv->gt_lock, irqflags);
4043 if (--dev_priv->forcewake_count == 0)
4044 dev_priv->gt.force_wake_put(dev_priv);
4045 spin_unlock_irqrestore(&dev_priv->gt_lock, irqflags);
4048 int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
4052 if (dev_priv->gt_fifo_count < GT_FIFO_NUM_RESERVED_ENTRIES) {
4054 u32 fifo = I915_READ_NOTRACE(GT_FIFO_FREE_ENTRIES);
4055 while (fifo <= GT_FIFO_NUM_RESERVED_ENTRIES && loop--) {
4057 fifo = I915_READ_NOTRACE(GT_FIFO_FREE_ENTRIES);
4059 if (WARN_ON(loop < 0 && fifo <= GT_FIFO_NUM_RESERVED_ENTRIES))
4061 dev_priv->gt_fifo_count = fifo;
4063 dev_priv->gt_fifo_count--;
4068 static void vlv_force_wake_get(struct drm_i915_private *dev_priv)
4070 /* Already awake? */
4071 if ((I915_READ(0x130094) & 0xa1) == 0xa1)
4074 I915_WRITE_NOTRACE(FORCEWAKE_VLV, 0xffffffff);
4075 POSTING_READ(FORCEWAKE_VLV);
4077 if (wait_for_atomic_us((I915_READ_NOTRACE(FORCEWAKE_ACK_VLV) & 1), 500))
4078 DRM_ERROR("Force wake wait timed out\n");
4080 __gen6_gt_wait_for_thread_c0(dev_priv);
4083 static void vlv_force_wake_put(struct drm_i915_private *dev_priv)
4085 I915_WRITE_NOTRACE(FORCEWAKE_VLV, 0xffff0000);
4086 /* FIXME: confirm VLV behavior with Punit folks */
4087 POSTING_READ(FORCEWAKE_VLV);
4090 void intel_gt_init(struct drm_device *dev)
4092 struct drm_i915_private *dev_priv = dev->dev_private;
4094 spin_lock_init(&dev_priv->gt_lock);
4096 if (IS_VALLEYVIEW(dev)) {
4097 dev_priv->gt.force_wake_get = vlv_force_wake_get;
4098 dev_priv->gt.force_wake_put = vlv_force_wake_put;
4099 } else if (INTEL_INFO(dev)->gen >= 6) {
4100 dev_priv->gt.force_wake_get = __gen6_gt_force_wake_get;
4101 dev_priv->gt.force_wake_put = __gen6_gt_force_wake_put;
4103 /* IVB configs may use multi-threaded forcewake */
4104 if (IS_IVYBRIDGE(dev) || IS_HASWELL(dev)) {
4107 /* A small trick here - if the bios hasn't configured
4108 * MT forcewake, and if the device is in RC6, then
4109 * force_wake_mt_get will not wake the device and the
4110 * ECOBUS read will return zero. Which will be
4111 * (correctly) interpreted by the test below as MT
4112 * forcewake being disabled.
4114 mutex_lock(&dev->struct_mutex);
4115 __gen6_gt_force_wake_mt_get(dev_priv);
4116 ecobus = I915_READ_NOTRACE(ECOBUS);
4117 __gen6_gt_force_wake_mt_put(dev_priv);
4118 mutex_unlock(&dev->struct_mutex);
4120 if (ecobus & FORCEWAKE_MT_ENABLE) {
4121 DRM_DEBUG_KMS("Using MT version of forcewake\n");
4122 dev_priv->gt.force_wake_get =
4123 __gen6_gt_force_wake_mt_get;
4124 dev_priv->gt.force_wake_put =
4125 __gen6_gt_force_wake_mt_put;