drivers/gpu/drm/i915/intel_pm.c

   1 /*
   2  * Copyright © 2012 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Eugeni Dodonov <eugeni.dodonov@intel.com>
  25  *
  26  */
  27
  28 #include <linux/cpufreq.h>
  29 #include "i915_drv.h"
  30 #include "intel_drv.h"
  31 #include "../../../platform/x86/intel_ips.h"
  32 #include <linux/module.h>
  33 #include <drm/i915_powerwell.h>
  34
  35 /**
  36  * RC6 is a special power stage which allows the GPU to enter an very
  37  * low-voltage mode when idle, using down to 0V while at this stage.  This
  38  * stage is entered automatically when the GPU is idle when RC6 support is
  39  * enabled, and as soon as new workload arises GPU wakes up automatically as well.
  40  *
  41  * There are different RC6 modes available in Intel GPU, which differentiate
  42  * among each other with the latency required to enter and leave RC6 and
  43  * voltage consumed by the GPU in different states.
  44  *
  45  * The combination of the following flags define which states GPU is allowed
  46  * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
  47  * RC6pp is deepest RC6. Their support by hardware varies according to the
  48  * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
  49  * which brings the most power savings; deeper states save more power, but
  50  * require higher latency to switch to and wake up.
  51  */
  52 #define INTEL_RC6_ENABLE                        (1<<0)
  53 #define INTEL_RC6p_ENABLE                       (1<<1)
  54 #define INTEL_RC6pp_ENABLE                      (1<<2)
  55
  56 /* FBC, or Frame Buffer Compression, is a technique employed to compress the
  57  * framebuffer contents in-memory, aiming at reducing the required bandwidth
  58  * during in-memory transfers and, therefore, reduce the power packet.
  59  *
  60  * The benefits of FBC are mostly visible with solid backgrounds and
  61  * variation-less patterns.
  62  *
  63  * FBC-related functionality can be enabled by the means of the
  64  * i915.i915_enable_fbc parameter
  65  */
  66
  67 static void i8xx_disable_fbc(struct drm_device *dev)
  68 {
  69         struct drm_i915_private *dev_priv = dev->dev_private;
  70         u32 fbc_ctl;
  71
  72         /* Disable compression */
  73         fbc_ctl = I915_READ(FBC_CONTROL);
  74         if ((fbc_ctl & FBC_CTL_EN) == 0)
  75                 return;
  76
  77         fbc_ctl &= ~FBC_CTL_EN;
  78         I915_WRITE(FBC_CONTROL, fbc_ctl);
  79
  80         /* Wait for compressing bit to clear */
  81         if (wait_for((I915_READ(FBC_STATUS) & FBC_STAT_COMPRESSING) == 0, 10)) {
  82                 DRM_DEBUG_KMS("FBC idle timed out\n");
  83                 return;
  84         }
  85
  86         DRM_DEBUG_KMS("disabled FBC\n");
  87 }
  88
  89 static void i8xx_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
  90 {
  91         struct drm_device *dev = crtc->dev;
  92         struct drm_i915_private *dev_priv = dev->dev_private;
  93         struct drm_framebuffer *fb = crtc->fb;
  94         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
  95         struct drm_i915_gem_object *obj = intel_fb->obj;
  96         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
  97         int cfb_pitch;
  98         int plane, i;
  99         u32 fbc_ctl, fbc_ctl2;
 100
 101         cfb_pitch = dev_priv->fbc.size / FBC_LL_SIZE;
 102         if (fb->pitches[0] < cfb_pitch)
 103                 cfb_pitch = fb->pitches[0];
 104
 105         /* FBC_CTL wants 64B units */
 106         cfb_pitch = (cfb_pitch / 64) - 1;
 107         plane = intel_crtc->plane == 0 ? FBC_CTL_PLANEA : FBC_CTL_PLANEB;
 108
 109         /* Clear old tags */
 110         for (i = 0; i < (FBC_LL_SIZE / 32) + 1; i++)
 111                 I915_WRITE(FBC_TAG + (i * 4), 0);
 112
 113         /* Set it up... */
 114         fbc_ctl2 = FBC_CTL_FENCE_DBL | FBC_CTL_IDLE_IMM | FBC_CTL_CPU_FENCE;
 115         fbc_ctl2 |= plane;
 116         I915_WRITE(FBC_CONTROL2, fbc_ctl2);
 117         I915_WRITE(FBC_FENCE_OFF, crtc->y);
 118
 119         /* enable it... */
 120         fbc_ctl = FBC_CTL_EN | FBC_CTL_PERIODIC;
 121         if (IS_I945GM(dev))
 122                 fbc_ctl |= FBC_CTL_C3_IDLE; /* 945 needs special SR handling */
 123         fbc_ctl |= (cfb_pitch & 0xff) << FBC_CTL_STRIDE_SHIFT;
 124         fbc_ctl |= (interval & 0x2fff) << FBC_CTL_INTERVAL_SHIFT;
 125         fbc_ctl |= obj->fence_reg;
 126         I915_WRITE(FBC_CONTROL, fbc_ctl);
 127
 128         DRM_DEBUG_KMS("enabled FBC, pitch %d, yoff %d, plane %c, ",
 129                       cfb_pitch, crtc->y, plane_name(intel_crtc->plane));
 130 }
 131
 132 static bool i8xx_fbc_enabled(struct drm_device *dev)
 133 {
 134         struct drm_i915_private *dev_priv = dev->dev_private;
 135
 136         return I915_READ(FBC_CONTROL) & FBC_CTL_EN;
 137 }
 138
 139 static void g4x_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 140 {
 141         struct drm_device *dev = crtc->dev;
 142         struct drm_i915_private *dev_priv = dev->dev_private;
 143         struct drm_framebuffer *fb = crtc->fb;
 144         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 145         struct drm_i915_gem_object *obj = intel_fb->obj;
 146         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 147         int plane = intel_crtc->plane == 0 ? DPFC_CTL_PLANEA : DPFC_CTL_PLANEB;
 148         unsigned long stall_watermark = 200;
 149         u32 dpfc_ctl;
 150
 151         dpfc_ctl = plane | DPFC_SR_EN | DPFC_CTL_LIMIT_1X;
 152         dpfc_ctl |= DPFC_CTL_FENCE_EN | obj->fence_reg;
 153         I915_WRITE(DPFC_CHICKEN, DPFC_HT_MODIFY);
 154
 155         I915_WRITE(DPFC_RECOMP_CTL, DPFC_RECOMP_STALL_EN |
 156                    (stall_watermark << DPFC_RECOMP_STALL_WM_SHIFT) |
 157                    (interval << DPFC_RECOMP_TIMER_COUNT_SHIFT));
 158         I915_WRITE(DPFC_FENCE_YOFF, crtc->y);
 159
 160         /* enable it... */
 161         I915_WRITE(DPFC_CONTROL, I915_READ(DPFC_CONTROL) | DPFC_CTL_EN);
 162
 163         DRM_DEBUG_KMS("enabled fbc on plane %c\n", plane_name(intel_crtc->plane));
 164 }
 165
 166 static void g4x_disable_fbc(struct drm_device *dev)
 167 {
 168         struct drm_i915_private *dev_priv = dev->dev_private;
 169         u32 dpfc_ctl;
 170
 171         /* Disable compression */
 172         dpfc_ctl = I915_READ(DPFC_CONTROL);
 173         if (dpfc_ctl & DPFC_CTL_EN) {
 174                 dpfc_ctl &= ~DPFC_CTL_EN;
 175                 I915_WRITE(DPFC_CONTROL, dpfc_ctl);
 176
 177                 DRM_DEBUG_KMS("disabled FBC\n");
 178         }
 179 }
 180
 181 static bool g4x_fbc_enabled(struct drm_device *dev)
 182 {
 183         struct drm_i915_private *dev_priv = dev->dev_private;
 184
 185         return I915_READ(DPFC_CONTROL) & DPFC_CTL_EN;
 186 }
 187
 188 static void sandybridge_blit_fbc_update(struct drm_device *dev)
 189 {
 190         struct drm_i915_private *dev_priv = dev->dev_private;
 191         u32 blt_ecoskpd;
 192
 193         /* Make sure blitter notifies FBC of writes */
 194         gen6_gt_force_wake_get(dev_priv);
 195         blt_ecoskpd = I915_READ(GEN6_BLITTER_ECOSKPD);
 196         blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY <<
 197                 GEN6_BLITTER_LOCK_SHIFT;
 198         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 199         blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY;
 200         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 201         blt_ecoskpd &= ~(GEN6_BLITTER_FBC_NOTIFY <<
 202                          GEN6_BLITTER_LOCK_SHIFT);
 203         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 204         POSTING_READ(GEN6_BLITTER_ECOSKPD);
 205         gen6_gt_force_wake_put(dev_priv);
 206 }
 207
 208 static void ironlake_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 209 {
 210         struct drm_device *dev = crtc->dev;
 211         struct drm_i915_private *dev_priv = dev->dev_private;
 212         struct drm_framebuffer *fb = crtc->fb;
 213         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 214         struct drm_i915_gem_object *obj = intel_fb->obj;
 215         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 216         int plane = intel_crtc->plane == 0 ? DPFC_CTL_PLANEA : DPFC_CTL_PLANEB;
 217         unsigned long stall_watermark = 200;
 218         u32 dpfc_ctl;
 219
 220         dpfc_ctl = I915_READ(ILK_DPFC_CONTROL);
 221         dpfc_ctl &= DPFC_RESERVED;
 222         dpfc_ctl |= (plane | DPFC_CTL_LIMIT_1X);
 223         /* Set persistent mode for front-buffer rendering, ala X. */
 224         dpfc_ctl |= DPFC_CTL_PERSISTENT_MODE;
 225         dpfc_ctl |= (DPFC_CTL_FENCE_EN | obj->fence_reg);
 226         I915_WRITE(ILK_DPFC_CHICKEN, DPFC_HT_MODIFY);
 227
 228         I915_WRITE(ILK_DPFC_RECOMP_CTL, DPFC_RECOMP_STALL_EN |
 229                    (stall_watermark << DPFC_RECOMP_STALL_WM_SHIFT) |
 230                    (interval << DPFC_RECOMP_TIMER_COUNT_SHIFT));
 231         I915_WRITE(ILK_DPFC_FENCE_YOFF, crtc->y);
 232         I915_WRITE(ILK_FBC_RT_BASE, i915_gem_obj_ggtt_offset(obj) | ILK_FBC_RT_VALID);
 233         /* enable it... */
 234         I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
 235
 236         if (IS_GEN6(dev)) {
 237                 I915_WRITE(SNB_DPFC_CTL_SA,
 238                            SNB_CPU_FENCE_ENABLE | obj->fence_reg);
 239                 I915_WRITE(DPFC_CPU_FENCE_OFFSET, crtc->y);
 240                 sandybridge_blit_fbc_update(dev);
 241         }
 242
 243         DRM_DEBUG_KMS("enabled fbc on plane %c\n", plane_name(intel_crtc->plane));
 244 }
 245
 246 static void ironlake_disable_fbc(struct drm_device *dev)
 247 {
 248         struct drm_i915_private *dev_priv = dev->dev_private;
 249         u32 dpfc_ctl;
 250
 251         /* Disable compression */
 252         dpfc_ctl = I915_READ(ILK_DPFC_CONTROL);
 253         if (dpfc_ctl & DPFC_CTL_EN) {
 254                 dpfc_ctl &= ~DPFC_CTL_EN;
 255                 I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl);
 256
 257                 DRM_DEBUG_KMS("disabled FBC\n");
 258         }
 259 }
 260
 261 static bool ironlake_fbc_enabled(struct drm_device *dev)
 262 {
 263         struct drm_i915_private *dev_priv = dev->dev_private;
 264
 265         return I915_READ(ILK_DPFC_CONTROL) & DPFC_CTL_EN;
 266 }
 267
 268 static void gen7_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 269 {
 270         struct drm_device *dev = crtc->dev;
 271         struct drm_i915_private *dev_priv = dev->dev_private;
 272         struct drm_framebuffer *fb = crtc->fb;
 273         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 274         struct drm_i915_gem_object *obj = intel_fb->obj;
 275         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 276
 277         I915_WRITE(IVB_FBC_RT_BASE, i915_gem_obj_ggtt_offset(obj));
 278
 279         I915_WRITE(ILK_DPFC_CONTROL, DPFC_CTL_EN | DPFC_CTL_LIMIT_1X |
 280                    IVB_DPFC_CTL_FENCE_EN |
 281                    intel_crtc->plane << IVB_DPFC_CTL_PLANE_SHIFT);
 282
 283         if (IS_IVYBRIDGE(dev)) {
 284                 /* WaFbcAsynchFlipDisableFbcQueue:ivb */
 285                 I915_WRITE(ILK_DISPLAY_CHICKEN1, ILK_FBCQ_DIS);
 286         } else {
 287                 /* WaFbcAsynchFlipDisableFbcQueue:hsw */
 288                 I915_WRITE(HSW_PIPE_SLICE_CHICKEN_1(intel_crtc->pipe),
 289                            HSW_BYPASS_FBC_QUEUE);
 290         }
 291
 292         I915_WRITE(SNB_DPFC_CTL_SA,
 293                    SNB_CPU_FENCE_ENABLE | obj->fence_reg);
 294         I915_WRITE(DPFC_CPU_FENCE_OFFSET, crtc->y);
 295
 296         sandybridge_blit_fbc_update(dev);
 297
 298         DRM_DEBUG_KMS("enabled fbc on plane %d\n", intel_crtc->plane);
 299 }
 300
 301 bool intel_fbc_enabled(struct drm_device *dev)
 302 {
 303         struct drm_i915_private *dev_priv = dev->dev_private;
 304
 305         if (!dev_priv->display.fbc_enabled)
 306                 return false;
 307
 308         return dev_priv->display.fbc_enabled(dev);
 309 }
 310
 311 static void intel_fbc_work_fn(struct work_struct *__work)
 312 {
 313         struct intel_fbc_work *work =
 314                 container_of(to_delayed_work(__work),
 315                              struct intel_fbc_work, work);
 316         struct drm_device *dev = work->crtc->dev;
 317         struct drm_i915_private *dev_priv = dev->dev_private;
 318
 319         mutex_lock(&dev->struct_mutex);
 320         if (work == dev_priv->fbc.fbc_work) {
 321                 /* Double check that we haven't switched fb without cancelling
 322                  * the prior work.
 323                  */
 324                 if (work->crtc->fb == work->fb) {
 325                         dev_priv->display.enable_fbc(work->crtc,
 326                                                      work->interval);
 327
 328                         dev_priv->fbc.plane = to_intel_crtc(work->crtc)->plane;
 329                         dev_priv->fbc.fb_id = work->crtc->fb->base.id;
 330                         dev_priv->fbc.y = work->crtc->y;
 331                 }
 332
 333                 dev_priv->fbc.fbc_work = NULL;
 334         }
 335         mutex_unlock(&dev->struct_mutex);
 336
 337         kfree(work);
 338 }
 339
 340 static void intel_cancel_fbc_work(struct drm_i915_private *dev_priv)
 341 {
 342         if (dev_priv->fbc.fbc_work == NULL)
 343                 return;
 344
 345         DRM_DEBUG_KMS("cancelling pending FBC enable\n");
 346
 347         /* Synchronisation is provided by struct_mutex and checking of
 348          * dev_priv->fbc.fbc_work, so we can perform the cancellation
 349          * entirely asynchronously.
 350          */
 351         if (cancel_delayed_work(&dev_priv->fbc.fbc_work->work))
 352                 /* tasklet was killed before being run, clean up */
 353                 kfree(dev_priv->fbc.fbc_work);
 354
 355         /* Mark the work as no longer wanted so that if it does
 356          * wake-up (because the work was already running and waiting
 357          * for our mutex), it will discover that is no longer
 358          * necessary to run.
 359          */
 360         dev_priv->fbc.fbc_work = NULL;
 361 }
 362
 363 static void intel_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 364 {
 365         struct intel_fbc_work *work;
 366         struct drm_device *dev = crtc->dev;
 367         struct drm_i915_private *dev_priv = dev->dev_private;
 368
 369         if (!dev_priv->display.enable_fbc)
 370                 return;
 371
 372         intel_cancel_fbc_work(dev_priv);
 373
 374         work = kzalloc(sizeof(*work), GFP_KERNEL);
 375         if (work == NULL) {
 376                 DRM_ERROR("Failed to allocate FBC work structure\n");
 377                 dev_priv->display.enable_fbc(crtc, interval);
 378                 return;
 379         }
 380
 381         work->crtc = crtc;
 382         work->fb = crtc->fb;
 383         work->interval = interval;
 384         INIT_DELAYED_WORK(&work->work, intel_fbc_work_fn);
 385
 386         dev_priv->fbc.fbc_work = work;
 387
 388         /* Delay the actual enabling to let pageflipping cease and the
 389          * display to settle before starting the compression. Note that
 390          * this delay also serves a second purpose: it allows for a
 391          * vblank to pass after disabling the FBC before we attempt
 392          * to modify the control registers.
 393          *
 394          * A more complicated solution would involve tracking vblanks
 395          * following the termination of the page-flipping sequence
 396          * and indeed performing the enable as a co-routine and not
 397          * waiting synchronously upon the vblank.
 398          *
 399          * WaFbcWaitForVBlankBeforeEnable:ilk,snb
 400          */
 401         schedule_delayed_work(&work->work, msecs_to_jiffies(50));
 402 }
 403
 404 void intel_disable_fbc(struct drm_device *dev)
 405 {
 406         struct drm_i915_private *dev_priv = dev->dev_private;
 407
 408         intel_cancel_fbc_work(dev_priv);
 409
 410         if (!dev_priv->display.disable_fbc)
 411                 return;
 412
 413         dev_priv->display.disable_fbc(dev);
 414         dev_priv->fbc.plane = -1;
 415 }
 416
 417 static bool set_no_fbc_reason(struct drm_i915_private *dev_priv,
 418                               enum no_fbc_reason reason)
 419 {
 420         if (dev_priv->fbc.no_fbc_reason == reason)
 421                 return false;
 422
 423         dev_priv->fbc.no_fbc_reason = reason;
 424         return true;
 425 }
 426
 427 /**
 428  * intel_update_fbc - enable/disable FBC as needed
 429  * @dev: the drm_device
 430  *
 431  * Set up the framebuffer compression hardware at mode set time.  We
 432  * enable it if possible:
 433  *   - plane A only (on pre-965)
 434  *   - no pixel mulitply/line duplication
 435  *   - no alpha buffer discard
 436  *   - no dual wide
 437  *   - framebuffer <= max_hdisplay in width, max_vdisplay in height
 438  *
 439  * We can't assume that any compression will take place (worst case),
 440  * so the compressed buffer has to be the same size as the uncompressed
 441  * one.  It also must reside (along with the line length buffer) in
 442  * stolen memory.
 443  *
 444  * We need to enable/disable FBC on a global basis.
 445  */
 446 void intel_update_fbc(struct drm_device *dev)
 447 {
 448         struct drm_i915_private *dev_priv = dev->dev_private;
 449         struct drm_crtc *crtc = NULL, *tmp_crtc;
 450         struct intel_crtc *intel_crtc;
 451         struct drm_framebuffer *fb;
 452         struct intel_framebuffer *intel_fb;
 453         struct drm_i915_gem_object *obj;
 454         const struct drm_display_mode *adjusted_mode;
 455         unsigned int max_width, max_height;
 456
 457         if (!I915_HAS_FBC(dev)) {
 458                 set_no_fbc_reason(dev_priv, FBC_UNSUPPORTED);
 459                 return;
 460         }
 461
 462         if (!i915_powersave) {
 463                 if (set_no_fbc_reason(dev_priv, FBC_MODULE_PARAM))
 464                         DRM_DEBUG_KMS("fbc disabled per module param\n");
 465                 return;
 466         }
 467
 468         /*
 469          * If FBC is already on, we just have to verify that we can
 470          * keep it that way...
 471          * Need to disable if:
 472          *   - more than one pipe is active
 473          *   - changing FBC params (stride, fence, mode)
 474          *   - new fb is too large to fit in compressed buffer
 475          *   - going to an unsupported config (interlace, pixel multiply, etc.)
 476          */
 477         list_for_each_entry(tmp_crtc, &dev->mode_config.crtc_list, head) {
 478                 if (intel_crtc_active(tmp_crtc) &&
 479                     to_intel_crtc(tmp_crtc)->primary_enabled) {
 480                         if (crtc) {
 481                                 if (set_no_fbc_reason(dev_priv, FBC_MULTIPLE_PIPES))
 482                                         DRM_DEBUG_KMS("more than one pipe active, disabling compression\n");
 483                                 goto out_disable;
 484                         }
 485                         crtc = tmp_crtc;
 486                 }
 487         }
 488
 489         if (!crtc || crtc->fb == NULL) {
 490                 if (set_no_fbc_reason(dev_priv, FBC_NO_OUTPUT))
 491                         DRM_DEBUG_KMS("no output, disabling\n");
 492                 goto out_disable;
 493         }
 494
 495         intel_crtc = to_intel_crtc(crtc);
 496         fb = crtc->fb;
 497         intel_fb = to_intel_framebuffer(fb);
 498         obj = intel_fb->obj;
 499         adjusted_mode = &intel_crtc->config.adjusted_mode;
 500
 501         if (i915_enable_fbc < 0 &&
 502             INTEL_INFO(dev)->gen <= 7 && !IS_HASWELL(dev)) {
 503                 if (set_no_fbc_reason(dev_priv, FBC_CHIP_DEFAULT))
 504                         DRM_DEBUG_KMS("disabled per chip default\n");
 505                 goto out_disable;
 506         }
 507         if (!i915_enable_fbc) {
 508                 if (set_no_fbc_reason(dev_priv, FBC_MODULE_PARAM))
 509                         DRM_DEBUG_KMS("fbc disabled per module param\n");
 510                 goto out_disable;
 511         }
 512         if ((adjusted_mode->flags & DRM_MODE_FLAG_INTERLACE) ||
 513             (adjusted_mode->flags & DRM_MODE_FLAG_DBLSCAN)) {
 514                 if (set_no_fbc_reason(dev_priv, FBC_UNSUPPORTED_MODE))
 515                         DRM_DEBUG_KMS("mode incompatible with compression, "
 516                                       "disabling\n");
 517                 goto out_disable;
 518         }
 519
 520         if (IS_G4X(dev) || INTEL_INFO(dev)->gen >= 5) {
 521                 max_width = 4096;
 522                 max_height = 2048;
 523         } else {
 524                 max_width = 2048;
 525                 max_height = 1536;
 526         }
 527         if (intel_crtc->config.pipe_src_w > max_width ||
 528             intel_crtc->config.pipe_src_h > max_height) {
 529                 if (set_no_fbc_reason(dev_priv, FBC_MODE_TOO_LARGE))
 530                         DRM_DEBUG_KMS("mode too large for compression, disabling\n");
 531                 goto out_disable;
 532         }
 533         if ((IS_I915GM(dev) || IS_I945GM(dev) || IS_HASWELL(dev)) &&
 534             intel_crtc->plane != 0) {
 535                 if (set_no_fbc_reason(dev_priv, FBC_BAD_PLANE))
 536                         DRM_DEBUG_KMS("plane not 0, disabling compression\n");
 537                 goto out_disable;
 538         }
 539
 540         /* The use of a CPU fence is mandatory in order to detect writes
 541          * by the CPU to the scanout and trigger updates to the FBC.
 542          */
 543         if (obj->tiling_mode != I915_TILING_X ||
 544             obj->fence_reg == I915_FENCE_REG_NONE) {
 545                 if (set_no_fbc_reason(dev_priv, FBC_NOT_TILED))
 546                         DRM_DEBUG_KMS("framebuffer not tiled or fenced, disabling compression\n");
 547                 goto out_disable;
 548         }
 549
 550         /* If the kernel debugger is active, always disable compression */
 551         if (in_dbg_master())
 552                 goto out_disable;
 553
 554         if (i915_gem_stolen_setup_compression(dev, intel_fb->obj->base.size)) {
 555                 if (set_no_fbc_reason(dev_priv, FBC_STOLEN_TOO_SMALL))
 556                         DRM_DEBUG_KMS("framebuffer too large, disabling compression\n");
 557                 goto out_disable;
 558         }
 559
 560         /* If the scanout has not changed, don't modify the FBC settings.
 561          * Note that we make the fundamental assumption that the fb->obj
 562          * cannot be unpinned (and have its GTT offset and fence revoked)
 563          * without first being decoupled from the scanout and FBC disabled.
 564          */
 565         if (dev_priv->fbc.plane == intel_crtc->plane &&
 566             dev_priv->fbc.fb_id == fb->base.id &&
 567             dev_priv->fbc.y == crtc->y)
 568                 return;
 569
 570         if (intel_fbc_enabled(dev)) {
 571                 /* We update FBC along two paths, after changing fb/crtc
 572                  * configuration (modeswitching) and after page-flipping
 573                  * finishes. For the latter, we know that not only did
 574                  * we disable the FBC at the start of the page-flip
 575                  * sequence, but also more than one vblank has passed.
 576                  *
 577                  * For the former case of modeswitching, it is possible
 578                  * to switch between two FBC valid configurations
 579                  * instantaneously so we do need to disable the FBC
 580                  * before we can modify its control registers. We also
 581                  * have to wait for the next vblank for that to take
 582                  * effect. However, since we delay enabling FBC we can
 583                  * assume that a vblank has passed since disabling and
 584                  * that we can safely alter the registers in the deferred
 585                  * callback.
 586                  *
 587                  * In the scenario that we go from a valid to invalid
 588                  * and then back to valid FBC configuration we have
 589                  * no strict enforcement that a vblank occurred since
 590                  * disabling the FBC. However, along all current pipe
 591                  * disabling paths we do need to wait for a vblank at
 592                  * some point. And we wait before enabling FBC anyway.
 593                  */
 594                 DRM_DEBUG_KMS("disabling active FBC for update\n");
 595                 intel_disable_fbc(dev);
 596         }
 597
 598         intel_enable_fbc(crtc, 500);
 599         dev_priv->fbc.no_fbc_reason = FBC_OK;
 600         return;
 601
 602 out_disable:
 603         /* Multiple disables should be harmless */
 604         if (intel_fbc_enabled(dev)) {
 605                 DRM_DEBUG_KMS("unsupported config, disabling FBC\n");
 606                 intel_disable_fbc(dev);
 607         }
 608         i915_gem_stolen_cleanup_compression(dev);
 609 }
 610
 611 static void i915_pineview_get_mem_freq(struct drm_device *dev)
 612 {
 613         drm_i915_private_t *dev_priv = dev->dev_private;
 614         u32 tmp;
 615
 616         tmp = I915_READ(CLKCFG);
 617
 618         switch (tmp & CLKCFG_FSB_MASK) {
 619         case CLKCFG_FSB_533:
 620                 dev_priv->fsb_freq = 533; /* 133*4 */
 621                 break;
 622         case CLKCFG_FSB_800:
 623                 dev_priv->fsb_freq = 800; /* 200*4 */
 624                 break;
 625         case CLKCFG_FSB_667:
 626                 dev_priv->fsb_freq =  667; /* 167*4 */
 627                 break;
 628         case CLKCFG_FSB_400:
 629                 dev_priv->fsb_freq = 400; /* 100*4 */
 630                 break;
 631         }
 632
 633         switch (tmp & CLKCFG_MEM_MASK) {
 634         case CLKCFG_MEM_533:
 635                 dev_priv->mem_freq = 533;
 636                 break;
 637         case CLKCFG_MEM_667:
 638                 dev_priv->mem_freq = 667;
 639                 break;
 640         case CLKCFG_MEM_800:
 641                 dev_priv->mem_freq = 800;
 642                 break;
 643         }
 644
 645         /* detect pineview DDR3 setting */
 646         tmp = I915_READ(CSHRDDR3CTL);
 647         dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
 648 }
 649
 650 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
 651 {
 652         drm_i915_private_t *dev_priv = dev->dev_private;
 653         u16 ddrpll, csipll;
 654
 655         ddrpll = I915_READ16(DDRMPLL1);
 656         csipll = I915_READ16(CSIPLL0);
 657
 658         switch (ddrpll & 0xff) {
 659         case 0xc:
 660                 dev_priv->mem_freq = 800;
 661                 break;
 662         case 0x10:
 663                 dev_priv->mem_freq = 1066;
 664                 break;
 665         case 0x14:
 666                 dev_priv->mem_freq = 1333;
 667                 break;
 668         case 0x18:
 669                 dev_priv->mem_freq = 1600;
 670                 break;
 671         default:
 672                 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
 673                                  ddrpll & 0xff);
 674                 dev_priv->mem_freq = 0;
 675                 break;
 676         }
 677
 678         dev_priv->ips.r_t = dev_priv->mem_freq;
 679
 680         switch (csipll & 0x3ff) {
 681         case 0x00c:
 682                 dev_priv->fsb_freq = 3200;
 683                 break;
 684         case 0x00e:
 685                 dev_priv->fsb_freq = 3733;
 686                 break;
 687         case 0x010:
 688                 dev_priv->fsb_freq = 4266;
 689                 break;
 690         case 0x012:
 691                 dev_priv->fsb_freq = 4800;
 692                 break;
 693         case 0x014:
 694                 dev_priv->fsb_freq = 5333;
 695                 break;
 696         case 0x016:
 697                 dev_priv->fsb_freq = 5866;
 698                 break;
 699         case 0x018:
 700                 dev_priv->fsb_freq = 6400;
 701                 break;
 702         default:
 703                 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
 704                                  csipll & 0x3ff);
 705                 dev_priv->fsb_freq = 0;
 706                 break;
 707         }
 708
 709         if (dev_priv->fsb_freq == 3200) {
 710                 dev_priv->ips.c_m = 0;
 711         } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
 712                 dev_priv->ips.c_m = 1;
 713         } else {
 714                 dev_priv->ips.c_m = 2;
 715         }
 716 }
 717
 718 static const struct cxsr_latency cxsr_latency_table[] = {
 719         {1, 0, 800, 400, 3382, 33382, 3983, 33983},    /* DDR2-400 SC */
 720         {1, 0, 800, 667, 3354, 33354, 3807, 33807},    /* DDR2-667 SC */
 721         {1, 0, 800, 800, 3347, 33347, 3763, 33763},    /* DDR2-800 SC */
 722         {1, 1, 800, 667, 6420, 36420, 6873, 36873},    /* DDR3-667 SC */
 723         {1, 1, 800, 800, 5902, 35902, 6318, 36318},    /* DDR3-800 SC */
 724
 725         {1, 0, 667, 400, 3400, 33400, 4021, 34021},    /* DDR2-400 SC */
 726         {1, 0, 667, 667, 3372, 33372, 3845, 33845},    /* DDR2-667 SC */
 727         {1, 0, 667, 800, 3386, 33386, 3822, 33822},    /* DDR2-800 SC */
 728         {1, 1, 667, 667, 6438, 36438, 6911, 36911},    /* DDR3-667 SC */
 729         {1, 1, 667, 800, 5941, 35941, 6377, 36377},    /* DDR3-800 SC */
 730
 731         {1, 0, 400, 400, 3472, 33472, 4173, 34173},    /* DDR2-400 SC */
 732         {1, 0, 400, 667, 3443, 33443, 3996, 33996},    /* DDR2-667 SC */
 733         {1, 0, 400, 800, 3430, 33430, 3946, 33946},    /* DDR2-800 SC */
 734         {1, 1, 400, 667, 6509, 36509, 7062, 37062},    /* DDR3-667 SC */
 735         {1, 1, 400, 800, 5985, 35985, 6501, 36501},    /* DDR3-800 SC */
 736
 737         {0, 0, 800, 400, 3438, 33438, 4065, 34065},    /* DDR2-400 SC */
 738         {0, 0, 800, 667, 3410, 33410, 3889, 33889},    /* DDR2-667 SC */
 739         {0, 0, 800, 800, 3403, 33403, 3845, 33845},    /* DDR2-800 SC */
 740         {0, 1, 800, 667, 6476, 36476, 6955, 36955},    /* DDR3-667 SC */
 741         {0, 1, 800, 800, 5958, 35958, 6400, 36400},    /* DDR3-800 SC */
 742
 743         {0, 0, 667, 400, 3456, 33456, 4103, 34106},    /* DDR2-400 SC */
 744         {0, 0, 667, 667, 3428, 33428, 3927, 33927},    /* DDR2-667 SC */
 745         {0, 0, 667, 800, 3443, 33443, 3905, 33905},    /* DDR2-800 SC */
 746         {0, 1, 667, 667, 6494, 36494, 6993, 36993},    /* DDR3-667 SC */
 747         {0, 1, 667, 800, 5998, 35998, 6460, 36460},    /* DDR3-800 SC */
 748
 749         {0, 0, 400, 400, 3528, 33528, 4255, 34255},    /* DDR2-400 SC */
 750         {0, 0, 400, 667, 3500, 33500, 4079, 34079},    /* DDR2-667 SC */
 751         {0, 0, 400, 800, 3487, 33487, 4029, 34029},    /* DDR2-800 SC */
 752         {0, 1, 400, 667, 6566, 36566, 7145, 37145},    /* DDR3-667 SC */
 753         {0, 1, 400, 800, 6042, 36042, 6584, 36584},    /* DDR3-800 SC */
 754 };
 755
 756 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
 757                                                          int is_ddr3,
 758                                                          int fsb,
 759                                                          int mem)
 760 {
 761         const struct cxsr_latency *latency;
 762         int i;
 763
 764         if (fsb == 0 || mem == 0)
 765                 return NULL;
 766
 767         for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
 768                 latency = &cxsr_latency_table[i];
 769                 if (is_desktop == latency->is_desktop &&
 770                     is_ddr3 == latency->is_ddr3 &&
 771                     fsb == latency->fsb_freq && mem == latency->mem_freq)
 772                         return latency;
 773         }
 774
 775         DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 776
 777         return NULL;
 778 }
 779
 780 static void pineview_disable_cxsr(struct drm_device *dev)
 781 {
 782         struct drm_i915_private *dev_priv = dev->dev_private;
 783
 784         /* deactivate cxsr */
 785         I915_WRITE(DSPFW3, I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN);
 786 }
 787
 788 /*
 789  * Latency for FIFO fetches is dependent on several factors:
 790  *   - memory configuration (speed, channels)
 791  *   - chipset
 792  *   - current MCH state
 793  * It can be fairly high in some situations, so here we assume a fairly
 794  * pessimal value.  It's a tradeoff between extra memory fetches (if we
 795  * set this value too high, the FIFO will fetch frequently to stay full)
 796  * and power consumption (set it too low to save power and we might see
 797  * FIFO underruns and display "flicker").
 798  *
 799  * A value of 5us seems to be a good balance; safe for very low end
 800  * platforms but not overly aggressive on lower latency configs.
 801  */
 802 static const int latency_ns = 5000;
 803
 804 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
 805 {
 806         struct drm_i915_private *dev_priv = dev->dev_private;
 807         uint32_t dsparb = I915_READ(DSPARB);
 808         int size;
 809
 810         size = dsparb & 0x7f;
 811         if (plane)
 812                 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
 813
 814         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 815                       plane ? "B" : "A", size);
 816
 817         return size;
 818 }
 819
 820 static int i85x_get_fifo_size(struct drm_device *dev, int plane)
 821 {
 822         struct drm_i915_private *dev_priv = dev->dev_private;
 823         uint32_t dsparb = I915_READ(DSPARB);
 824         int size;
 825
 826         size = dsparb & 0x1ff;
 827         if (plane)
 828                 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
 829         size >>= 1; /* Convert to cachelines */
 830
 831         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 832                       plane ? "B" : "A", size);
 833
 834         return size;
 835 }
 836
 837 static int i845_get_fifo_size(struct drm_device *dev, int plane)
 838 {
 839         struct drm_i915_private *dev_priv = dev->dev_private;
 840         uint32_t dsparb = I915_READ(DSPARB);
 841         int size;
 842
 843         size = dsparb & 0x7f;
 844         size >>= 2; /* Convert to cachelines */
 845
 846         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 847                       plane ? "B" : "A",
 848                       size);
 849
 850         return size;
 851 }
 852
 853 static int i830_get_fifo_size(struct drm_device *dev, int plane)
 854 {
 855         struct drm_i915_private *dev_priv = dev->dev_private;
 856         uint32_t dsparb = I915_READ(DSPARB);
 857         int size;
 858
 859         size = dsparb & 0x7f;
 860         size >>= 1; /* Convert to cachelines */
 861
 862         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 863                       plane ? "B" : "A", size);
 864
 865         return size;
 866 }
 867
 868 /* Pineview has different values for various configs */
 869 static const struct intel_watermark_params pineview_display_wm = {
 870         PINEVIEW_DISPLAY_FIFO,
 871         PINEVIEW_MAX_WM,
 872         PINEVIEW_DFT_WM,
 873         PINEVIEW_GUARD_WM,
 874         PINEVIEW_FIFO_LINE_SIZE
 875 };
 876 static const struct intel_watermark_params pineview_display_hplloff_wm = {
 877         PINEVIEW_DISPLAY_FIFO,
 878         PINEVIEW_MAX_WM,
 879         PINEVIEW_DFT_HPLLOFF_WM,
 880         PINEVIEW_GUARD_WM,
 881         PINEVIEW_FIFO_LINE_SIZE
 882 };
 883 static const struct intel_watermark_params pineview_cursor_wm = {
 884         PINEVIEW_CURSOR_FIFO,
 885         PINEVIEW_CURSOR_MAX_WM,
 886         PINEVIEW_CURSOR_DFT_WM,
 887         PINEVIEW_CURSOR_GUARD_WM,
 888         PINEVIEW_FIFO_LINE_SIZE,
 889 };
 890 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
 891         PINEVIEW_CURSOR_FIFO,
 892         PINEVIEW_CURSOR_MAX_WM,
 893         PINEVIEW_CURSOR_DFT_WM,
 894         PINEVIEW_CURSOR_GUARD_WM,
 895         PINEVIEW_FIFO_LINE_SIZE
 896 };
 897 static const struct intel_watermark_params g4x_wm_info = {
 898         G4X_FIFO_SIZE,
 899         G4X_MAX_WM,
 900         G4X_MAX_WM,
 901         2,
 902         G4X_FIFO_LINE_SIZE,
 903 };
 904 static const struct intel_watermark_params g4x_cursor_wm_info = {
 905         I965_CURSOR_FIFO,
 906         I965_CURSOR_MAX_WM,
 907         I965_CURSOR_DFT_WM,
 908         2,
 909         G4X_FIFO_LINE_SIZE,
 910 };
 911 static const struct intel_watermark_params valleyview_wm_info = {
 912         VALLEYVIEW_FIFO_SIZE,
 913         VALLEYVIEW_MAX_WM,
 914         VALLEYVIEW_MAX_WM,
 915         2,
 916         G4X_FIFO_LINE_SIZE,
 917 };
 918 static const struct intel_watermark_params valleyview_cursor_wm_info = {
 919         I965_CURSOR_FIFO,
 920         VALLEYVIEW_CURSOR_MAX_WM,
 921         I965_CURSOR_DFT_WM,
 922         2,
 923         G4X_FIFO_LINE_SIZE,
 924 };
 925 static const struct intel_watermark_params i965_cursor_wm_info = {
 926         I965_CURSOR_FIFO,
 927         I965_CURSOR_MAX_WM,
 928         I965_CURSOR_DFT_WM,
 929         2,
 930         I915_FIFO_LINE_SIZE,
 931 };
 932 static const struct intel_watermark_params i945_wm_info = {
 933         I945_FIFO_SIZE,
 934         I915_MAX_WM,
 935         1,
 936         2,
 937         I915_FIFO_LINE_SIZE
 938 };
 939 static const struct intel_watermark_params i915_wm_info = {
 940         I915_FIFO_SIZE,
 941         I915_MAX_WM,
 942         1,
 943         2,
 944         I915_FIFO_LINE_SIZE
 945 };
 946 static const struct intel_watermark_params i855_wm_info = {
 947         I855GM_FIFO_SIZE,
 948         I915_MAX_WM,
 949         1,
 950         2,
 951         I830_FIFO_LINE_SIZE
 952 };
 953 static const struct intel_watermark_params i830_wm_info = {
 954         I830_FIFO_SIZE,
 955         I915_MAX_WM,
 956         1,
 957         2,
 958         I830_FIFO_LINE_SIZE
 959 };
 960
 961 static const struct intel_watermark_params ironlake_display_wm_info = {
 962         ILK_DISPLAY_FIFO,
 963         ILK_DISPLAY_MAXWM,
 964         ILK_DISPLAY_DFTWM,
 965         2,
 966         ILK_FIFO_LINE_SIZE
 967 };
 968 static const struct intel_watermark_params ironlake_cursor_wm_info = {
 969         ILK_CURSOR_FIFO,
 970         ILK_CURSOR_MAXWM,
 971         ILK_CURSOR_DFTWM,
 972         2,
 973         ILK_FIFO_LINE_SIZE
 974 };
 975 static const struct intel_watermark_params ironlake_display_srwm_info = {
 976         ILK_DISPLAY_SR_FIFO,
 977         ILK_DISPLAY_MAX_SRWM,
 978         ILK_DISPLAY_DFT_SRWM,
 979         2,
 980         ILK_FIFO_LINE_SIZE
 981 };
 982 static const struct intel_watermark_params ironlake_cursor_srwm_info = {
 983         ILK_CURSOR_SR_FIFO,
 984         ILK_CURSOR_MAX_SRWM,
 985         ILK_CURSOR_DFT_SRWM,
 986         2,
 987         ILK_FIFO_LINE_SIZE
 988 };
 989
 990 static const struct intel_watermark_params sandybridge_display_wm_info = {
 991         SNB_DISPLAY_FIFO,
 992         SNB_DISPLAY_MAXWM,
 993         SNB_DISPLAY_DFTWM,
 994         2,
 995         SNB_FIFO_LINE_SIZE
 996 };
 997 static const struct intel_watermark_params sandybridge_cursor_wm_info = {
 998         SNB_CURSOR_FIFO,
 999         SNB_CURSOR_MAXWM,
1000         SNB_CURSOR_DFTWM,
1001         2,
1002         SNB_FIFO_LINE_SIZE
1003 };
1004 static const struct intel_watermark_params sandybridge_display_srwm_info = {
1005         SNB_DISPLAY_SR_FIFO,
1006         SNB_DISPLAY_MAX_SRWM,
1007         SNB_DISPLAY_DFT_SRWM,
1008         2,
1009         SNB_FIFO_LINE_SIZE
1010 };
1011 static const struct intel_watermark_params sandybridge_cursor_srwm_info = {
1012         SNB_CURSOR_SR_FIFO,
1013         SNB_CURSOR_MAX_SRWM,
1014         SNB_CURSOR_DFT_SRWM,
1015         2,
1016         SNB_FIFO_LINE_SIZE
1017 };
1018
1019
1020 /**
1021  * intel_calculate_wm - calculate watermark level
1022  * @clock_in_khz: pixel clock
1023  * @wm: chip FIFO params
1024  * @pixel_size: display pixel size
1025  * @latency_ns: memory latency for the platform
1026  *
1027  * Calculate the watermark level (the level at which the display plane will
1028  * start fetching from memory again).  Each chip has a different display
1029  * FIFO size and allocation, so the caller needs to figure that out and pass
1030  * in the correct intel_watermark_params structure.
1031  *
1032  * As the pixel clock runs, the FIFO will be drained at a rate that depends
1033  * on the pixel size.  When it reaches the watermark level, it'll start
1034  * fetching FIFO line sized based chunks from memory until the FIFO fills
1035  * past the watermark point.  If the FIFO drains completely, a FIFO underrun
1036  * will occur, and a display engine hang could result.
1037  */
1038 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
1039                                         const struct intel_watermark_params *wm,
1040                                         int fifo_size,
1041                                         int pixel_size,
1042                                         unsigned long latency_ns)
1043 {
1044         long entries_required, wm_size;
1045
1046         /*
1047          * Note: we need to make sure we don't overflow for various clock &
1048          * latency values.
1049          * clocks go from a few thousand to several hundred thousand.
1050          * latency is usually a few thousand
1051          */
1052         entries_required = ((clock_in_khz / 1000) * pixel_size * latency_ns) /
1053                 1000;
1054         entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
1055
1056         DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
1057
1058         wm_size = fifo_size - (entries_required + wm->guard_size);
1059
1060         DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
1061
1062         /* Don't promote wm_size to unsigned... */
1063         if (wm_size > (long)wm->max_wm)
1064                 wm_size = wm->max_wm;
1065         if (wm_size <= 0)
1066                 wm_size = wm->default_wm;
1067         return wm_size;
1068 }
1069
1070 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
1071 {
1072         struct drm_crtc *crtc, *enabled = NULL;
1073
1074         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
1075                 if (intel_crtc_active(crtc)) {
1076                         if (enabled)
1077                                 return NULL;
1078                         enabled = crtc;
1079                 }
1080         }
1081
1082         return enabled;
1083 }
1084
1085 static void pineview_update_wm(struct drm_crtc *unused_crtc)
1086 {
1087         struct drm_device *dev = unused_crtc->dev;
1088         struct drm_i915_private *dev_priv = dev->dev_private;
1089         struct drm_crtc *crtc;
1090         const struct cxsr_latency *latency;
1091         u32 reg;
1092         unsigned long wm;
1093
1094         latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
1095                                          dev_priv->fsb_freq, dev_priv->mem_freq);
1096         if (!latency) {
1097                 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
1098                 pineview_disable_cxsr(dev);
1099                 return;
1100         }
1101
1102         crtc = single_enabled_crtc(dev);
1103         if (crtc) {
1104                 const struct drm_display_mode *adjusted_mode;
1105                 int pixel_size = crtc->fb->bits_per_pixel / 8;
1106                 int clock;
1107
1108                 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1109                 clock = adjusted_mode->crtc_clock;
1110
1111                 /* Display SR */
1112                 wm = intel_calculate_wm(clock, &pineview_display_wm,
1113                                         pineview_display_wm.fifo_size,
1114                                         pixel_size, latency->display_sr);
1115                 reg = I915_READ(DSPFW1);
1116                 reg &= ~DSPFW_SR_MASK;
1117                 reg |= wm << DSPFW_SR_SHIFT;
1118                 I915_WRITE(DSPFW1, reg);
1119                 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
1120
1121                 /* cursor SR */
1122                 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
1123                                         pineview_display_wm.fifo_size,
1124                                         pixel_size, latency->cursor_sr);
1125                 reg = I915_READ(DSPFW3);
1126                 reg &= ~DSPFW_CURSOR_SR_MASK;
1127                 reg |= (wm & 0x3f) << DSPFW_CURSOR_SR_SHIFT;
1128                 I915_WRITE(DSPFW3, reg);
1129
1130                 /* Display HPLL off SR */
1131                 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
1132                                         pineview_display_hplloff_wm.fifo_size,
1133                                         pixel_size, latency->display_hpll_disable);
1134                 reg = I915_READ(DSPFW3);
1135                 reg &= ~DSPFW_HPLL_SR_MASK;
1136                 reg |= wm & DSPFW_HPLL_SR_MASK;
1137                 I915_WRITE(DSPFW3, reg);
1138
1139                 /* cursor HPLL off SR */
1140                 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
1141                                         pineview_display_hplloff_wm.fifo_size,
1142                                         pixel_size, latency->cursor_hpll_disable);
1143                 reg = I915_READ(DSPFW3);
1144                 reg &= ~DSPFW_HPLL_CURSOR_MASK;
1145                 reg |= (wm & 0x3f) << DSPFW_HPLL_CURSOR_SHIFT;
1146                 I915_WRITE(DSPFW3, reg);
1147                 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
1148
1149                 /* activate cxsr */
1150                 I915_WRITE(DSPFW3,
1151                            I915_READ(DSPFW3) | PINEVIEW_SELF_REFRESH_EN);
1152                 DRM_DEBUG_KMS("Self-refresh is enabled\n");
1153         } else {
1154                 pineview_disable_cxsr(dev);
1155                 DRM_DEBUG_KMS("Self-refresh is disabled\n");
1156         }
1157 }
1158
1159 static bool g4x_compute_wm0(struct drm_device *dev,
1160                             int plane,
1161                             const struct intel_watermark_params *display,
1162                             int display_latency_ns,
1163                             const struct intel_watermark_params *cursor,
1164                             int cursor_latency_ns,
1165                             int *plane_wm,
1166                             int *cursor_wm)
1167 {
1168         struct drm_crtc *crtc;
1169         const struct drm_display_mode *adjusted_mode;
1170         int htotal, hdisplay, clock, pixel_size;
1171         int line_time_us, line_count;
1172         int entries, tlb_miss;
1173
1174         crtc = intel_get_crtc_for_plane(dev, plane);
1175         if (!intel_crtc_active(crtc)) {
1176                 *cursor_wm = cursor->guard_size;
1177                 *plane_wm = display->guard_size;
1178                 return false;
1179         }
1180
1181         adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1182         clock = adjusted_mode->crtc_clock;
1183         htotal = adjusted_mode->htotal;
1184         hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1185         pixel_size = crtc->fb->bits_per_pixel / 8;
1186
1187         /* Use the small buffer method to calculate plane watermark */
1188         entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
1189         tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
1190         if (tlb_miss > 0)
1191                 entries += tlb_miss;
1192         entries = DIV_ROUND_UP(entries, display->cacheline_size);
1193         *plane_wm = entries + display->guard_size;
1194         if (*plane_wm > (int)display->max_wm)
1195                 *plane_wm = display->max_wm;
1196
1197         /* Use the large buffer method to calculate cursor watermark */
1198         line_time_us = ((htotal * 1000) / clock);
1199         line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
1200         entries = line_count * 64 * pixel_size;
1201         tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
1202         if (tlb_miss > 0)
1203                 entries += tlb_miss;
1204         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1205         *cursor_wm = entries + cursor->guard_size;
1206         if (*cursor_wm > (int)cursor->max_wm)
1207                 *cursor_wm = (int)cursor->max_wm;
1208
1209         return true;
1210 }
1211
1212 /*
1213  * Check the wm result.
1214  *
1215  * If any calculated watermark values is larger than the maximum value that
1216  * can be programmed into the associated watermark register, that watermark
1217  * must be disabled.
1218  */
1219 static bool g4x_check_srwm(struct drm_device *dev,
1220                            int display_wm, int cursor_wm,
1221                            const struct intel_watermark_params *display,
1222                            const struct intel_watermark_params *cursor)
1223 {
1224         DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
1225                       display_wm, cursor_wm);
1226
1227         if (display_wm > display->max_wm) {
1228                 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
1229                               display_wm, display->max_wm);
1230                 return false;
1231         }
1232
1233         if (cursor_wm > cursor->max_wm) {
1234                 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
1235                               cursor_wm, cursor->max_wm);
1236                 return false;
1237         }
1238
1239         if (!(display_wm || cursor_wm)) {
1240                 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
1241                 return false;
1242         }
1243
1244         return true;
1245 }
1246
1247 static bool g4x_compute_srwm(struct drm_device *dev,
1248                              int plane,
1249                              int latency_ns,
1250                              const struct intel_watermark_params *display,
1251                              const struct intel_watermark_params *cursor,
1252                              int *display_wm, int *cursor_wm)
1253 {
1254         struct drm_crtc *crtc;
1255         const struct drm_display_mode *adjusted_mode;
1256         int hdisplay, htotal, pixel_size, clock;
1257         unsigned long line_time_us;
1258         int line_count, line_size;
1259         int small, large;
1260         int entries;
1261
1262         if (!latency_ns) {
1263                 *display_wm = *cursor_wm = 0;
1264                 return false;
1265         }
1266
1267         crtc = intel_get_crtc_for_plane(dev, plane);
1268         adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1269         clock = adjusted_mode->crtc_clock;
1270         htotal = adjusted_mode->htotal;
1271         hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1272         pixel_size = crtc->fb->bits_per_pixel / 8;
1273
1274         line_time_us = (htotal * 1000) / clock;
1275         line_count = (latency_ns / line_time_us + 1000) / 1000;
1276         line_size = hdisplay * pixel_size;
1277
1278         /* Use the minimum of the small and large buffer method for primary */
1279         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1280         large = line_count * line_size;
1281
1282         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1283         *display_wm = entries + display->guard_size;
1284
1285         /* calculate the self-refresh watermark for display cursor */
1286         entries = line_count * pixel_size * 64;
1287         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1288         *cursor_wm = entries + cursor->guard_size;
1289
1290         return g4x_check_srwm(dev,
1291                               *display_wm, *cursor_wm,
1292                               display, cursor);
1293 }
1294
1295 static bool vlv_compute_drain_latency(struct drm_device *dev,
1296                                      int plane,
1297                                      int *plane_prec_mult,
1298                                      int *plane_dl,
1299                                      int *cursor_prec_mult,
1300                                      int *cursor_dl)
1301 {
1302         struct drm_crtc *crtc;
1303         int clock, pixel_size;
1304         int entries;
1305
1306         crtc = intel_get_crtc_for_plane(dev, plane);
1307         if (!intel_crtc_active(crtc))
1308                 return false;
1309
1310         clock = to_intel_crtc(crtc)->config.adjusted_mode.crtc_clock;
1311         pixel_size = crtc->fb->bits_per_pixel / 8;      /* BPP */
1312
1313         entries = (clock / 1000) * pixel_size;
1314         *plane_prec_mult = (entries > 256) ?
1315                 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1316         *plane_dl = (64 * (*plane_prec_mult) * 4) / ((clock / 1000) *
1317                                                      pixel_size);
1318
1319         entries = (clock / 1000) * 4;   /* BPP is always 4 for cursor */
1320         *cursor_prec_mult = (entries > 256) ?
1321                 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1322         *cursor_dl = (64 * (*cursor_prec_mult) * 4) / ((clock / 1000) * 4);
1323
1324         return true;
1325 }
1326
1327 /*
1328  * Update drain latency registers of memory arbiter
1329  *
1330  * Valleyview SoC has a new memory arbiter and needs drain latency registers
1331  * to be programmed. Each plane has a drain latency multiplier and a drain
1332  * latency value.
1333  */
1334
1335 static void vlv_update_drain_latency(struct drm_device *dev)
1336 {
1337         struct drm_i915_private *dev_priv = dev->dev_private;
1338         int planea_prec, planea_dl, planeb_prec, planeb_dl;
1339         int cursora_prec, cursora_dl, cursorb_prec, cursorb_dl;
1340         int plane_prec_mult, cursor_prec_mult; /* Precision multiplier is
1341                                                         either 16 or 32 */
1342
1343         /* For plane A, Cursor A */
1344         if (vlv_compute_drain_latency(dev, 0, &plane_prec_mult, &planea_dl,
1345                                       &cursor_prec_mult, &cursora_dl)) {
1346                 cursora_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1347                         DDL_CURSORA_PRECISION_32 : DDL_CURSORA_PRECISION_16;
1348                 planea_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1349                         DDL_PLANEA_PRECISION_32 : DDL_PLANEA_PRECISION_16;
1350
1351                 I915_WRITE(VLV_DDL1, cursora_prec |
1352                                 (cursora_dl << DDL_CURSORA_SHIFT) |
1353                                 planea_prec | planea_dl);
1354         }
1355
1356         /* For plane B, Cursor B */
1357         if (vlv_compute_drain_latency(dev, 1, &plane_prec_mult, &planeb_dl,
1358                                       &cursor_prec_mult, &cursorb_dl)) {
1359                 cursorb_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1360                         DDL_CURSORB_PRECISION_32 : DDL_CURSORB_PRECISION_16;
1361                 planeb_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1362                         DDL_PLANEB_PRECISION_32 : DDL_PLANEB_PRECISION_16;
1363
1364                 I915_WRITE(VLV_DDL2, cursorb_prec |
1365                                 (cursorb_dl << DDL_CURSORB_SHIFT) |
1366                                 planeb_prec | planeb_dl);
1367         }
1368 }
1369
1370 #define single_plane_enabled(mask) is_power_of_2(mask)
1371
1372 static void valleyview_update_wm(struct drm_crtc *crtc)
1373 {
1374         struct drm_device *dev = crtc->dev;
1375         static const int sr_latency_ns = 12000;
1376         struct drm_i915_private *dev_priv = dev->dev_private;
1377         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1378         int plane_sr, cursor_sr;
1379         int ignore_plane_sr, ignore_cursor_sr;
1380         unsigned int enabled = 0;
1381
1382         vlv_update_drain_latency(dev);
1383
1384         if (g4x_compute_wm0(dev, PIPE_A,
1385                             &valleyview_wm_info, latency_ns,
1386                             &valleyview_cursor_wm_info, latency_ns,
1387                             &planea_wm, &cursora_wm))
1388                 enabled |= 1 << PIPE_A;
1389
1390         if (g4x_compute_wm0(dev, PIPE_B,
1391                             &valleyview_wm_info, latency_ns,
1392                             &valleyview_cursor_wm_info, latency_ns,
1393                             &planeb_wm, &cursorb_wm))
1394                 enabled |= 1 << PIPE_B;
1395
1396         if (single_plane_enabled(enabled) &&
1397             g4x_compute_srwm(dev, ffs(enabled) - 1,
1398                              sr_latency_ns,
1399                              &valleyview_wm_info,
1400                              &valleyview_cursor_wm_info,
1401                              &plane_sr, &ignore_cursor_sr) &&
1402             g4x_compute_srwm(dev, ffs(enabled) - 1,
1403                              2*sr_latency_ns,
1404                              &valleyview_wm_info,
1405                              &valleyview_cursor_wm_info,
1406                              &ignore_plane_sr, &cursor_sr)) {
1407                 I915_WRITE(FW_BLC_SELF_VLV, FW_CSPWRDWNEN);
1408         } else {
1409                 I915_WRITE(FW_BLC_SELF_VLV,
1410                            I915_READ(FW_BLC_SELF_VLV) & ~FW_CSPWRDWNEN);
1411                 plane_sr = cursor_sr = 0;
1412         }
1413
1414         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1415                       planea_wm, cursora_wm,
1416                       planeb_wm, cursorb_wm,
1417                       plane_sr, cursor_sr);
1418
1419         I915_WRITE(DSPFW1,
1420                    (plane_sr << DSPFW_SR_SHIFT) |
1421                    (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1422                    (planeb_wm << DSPFW_PLANEB_SHIFT) |
1423                    planea_wm);
1424         I915_WRITE(DSPFW2,
1425                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1426                    (cursora_wm << DSPFW_CURSORA_SHIFT));
1427         I915_WRITE(DSPFW3,
1428                    (I915_READ(DSPFW3) & ~DSPFW_CURSOR_SR_MASK) |
1429                    (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1430 }
1431
1432 static void g4x_update_wm(struct drm_crtc *crtc)
1433 {
1434         struct drm_device *dev = crtc->dev;
1435         static const int sr_latency_ns = 12000;
1436         struct drm_i915_private *dev_priv = dev->dev_private;
1437         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1438         int plane_sr, cursor_sr;
1439         unsigned int enabled = 0;
1440
1441         if (g4x_compute_wm0(dev, PIPE_A,
1442                             &g4x_wm_info, latency_ns,
1443                             &g4x_cursor_wm_info, latency_ns,
1444                             &planea_wm, &cursora_wm))
1445                 enabled |= 1 << PIPE_A;
1446
1447         if (g4x_compute_wm0(dev, PIPE_B,
1448                             &g4x_wm_info, latency_ns,
1449                             &g4x_cursor_wm_info, latency_ns,
1450                             &planeb_wm, &cursorb_wm))
1451                 enabled |= 1 << PIPE_B;
1452
1453         if (single_plane_enabled(enabled) &&
1454             g4x_compute_srwm(dev, ffs(enabled) - 1,
1455                              sr_latency_ns,
1456                              &g4x_wm_info,
1457                              &g4x_cursor_wm_info,
1458                              &plane_sr, &cursor_sr)) {
1459                 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1460         } else {
1461                 I915_WRITE(FW_BLC_SELF,
1462                            I915_READ(FW_BLC_SELF) & ~FW_BLC_SELF_EN);
1463                 plane_sr = cursor_sr = 0;
1464         }
1465
1466         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1467                       planea_wm, cursora_wm,
1468                       planeb_wm, cursorb_wm,
1469                       plane_sr, cursor_sr);
1470
1471         I915_WRITE(DSPFW1,
1472                    (plane_sr << DSPFW_SR_SHIFT) |
1473                    (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1474                    (planeb_wm << DSPFW_PLANEB_SHIFT) |
1475                    planea_wm);
1476         I915_WRITE(DSPFW2,
1477                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1478                    (cursora_wm << DSPFW_CURSORA_SHIFT));
1479         /* HPLL off in SR has some issues on G4x... disable it */
1480         I915_WRITE(DSPFW3,
1481                    (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1482                    (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1483 }
1484
1485 static void i965_update_wm(struct drm_crtc *unused_crtc)
1486 {
1487         struct drm_device *dev = unused_crtc->dev;
1488         struct drm_i915_private *dev_priv = dev->dev_private;
1489         struct drm_crtc *crtc;
1490         int srwm = 1;
1491         int cursor_sr = 16;
1492
1493         /* Calc sr entries for one plane configs */
1494         crtc = single_enabled_crtc(dev);
1495         if (crtc) {
1496                 /* self-refresh has much higher latency */
1497                 static const int sr_latency_ns = 12000;
1498                 const struct drm_display_mode *adjusted_mode =
1499                         &to_intel_crtc(crtc)->config.adjusted_mode;
1500                 int clock = adjusted_mode->crtc_clock;
1501                 int htotal = adjusted_mode->htotal;
1502                 int hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1503                 int pixel_size = crtc->fb->bits_per_pixel / 8;
1504                 unsigned long line_time_us;
1505                 int entries;
1506
1507                 line_time_us = ((htotal * 1000) / clock);
1508
1509                 /* Use ns/us then divide to preserve precision */
1510                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1511                         pixel_size * hdisplay;
1512                 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1513                 srwm = I965_FIFO_SIZE - entries;
1514                 if (srwm < 0)
1515                         srwm = 1;
1516                 srwm &= 0x1ff;
1517                 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1518                               entries, srwm);
1519
1520                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1521                         pixel_size * 64;
1522                 entries = DIV_ROUND_UP(entries,
1523                                           i965_cursor_wm_info.cacheline_size);
1524                 cursor_sr = i965_cursor_wm_info.fifo_size -
1525                         (entries + i965_cursor_wm_info.guard_size);
1526
1527                 if (cursor_sr > i965_cursor_wm_info.max_wm)
1528                         cursor_sr = i965_cursor_wm_info.max_wm;
1529
1530                 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1531                               "cursor %d\n", srwm, cursor_sr);
1532
1533                 if (IS_CRESTLINE(dev))
1534                         I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1535         } else {
1536                 /* Turn off self refresh if both pipes are enabled */
1537                 if (IS_CRESTLINE(dev))
1538                         I915_WRITE(FW_BLC_SELF, I915_READ(FW_BLC_SELF)
1539                                    & ~FW_BLC_SELF_EN);
1540         }
1541
1542         DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1543                       srwm);
1544
1545         /* 965 has limitations... */
1546         I915_WRITE(DSPFW1, (srwm << DSPFW_SR_SHIFT) |
1547                    (8 << 16) | (8 << 8) | (8 << 0));
1548         I915_WRITE(DSPFW2, (8 << 8) | (8 << 0));
1549         /* update cursor SR watermark */
1550         I915_WRITE(DSPFW3, (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1551 }
1552
1553 static void i9xx_update_wm(struct drm_crtc *unused_crtc)
1554 {
1555         struct drm_device *dev = unused_crtc->dev;
1556         struct drm_i915_private *dev_priv = dev->dev_private;
1557         const struct intel_watermark_params *wm_info;
1558         uint32_t fwater_lo;
1559         uint32_t fwater_hi;
1560         int cwm, srwm = 1;
1561         int fifo_size;
1562         int planea_wm, planeb_wm;
1563         struct drm_crtc *crtc, *enabled = NULL;
1564
1565         if (IS_I945GM(dev))
1566                 wm_info = &i945_wm_info;
1567         else if (!IS_GEN2(dev))
1568                 wm_info = &i915_wm_info;
1569         else
1570                 wm_info = &i855_wm_info;
1571
1572         fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1573         crtc = intel_get_crtc_for_plane(dev, 0);
1574         if (intel_crtc_active(crtc)) {
1575                 const struct drm_display_mode *adjusted_mode;
1576                 int cpp = crtc->fb->bits_per_pixel / 8;
1577                 if (IS_GEN2(dev))
1578                         cpp = 4;
1579
1580                 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1581                 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1582                                                wm_info, fifo_size, cpp,
1583                                                latency_ns);
1584                 enabled = crtc;
1585         } else
1586                 planea_wm = fifo_size - wm_info->guard_size;
1587
1588         fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1589         crtc = intel_get_crtc_for_plane(dev, 1);
1590         if (intel_crtc_active(crtc)) {
1591                 const struct drm_display_mode *adjusted_mode;
1592                 int cpp = crtc->fb->bits_per_pixel / 8;
1593                 if (IS_GEN2(dev))
1594                         cpp = 4;
1595
1596                 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1597                 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1598                                                wm_info, fifo_size, cpp,
1599                                                latency_ns);
1600                 if (enabled == NULL)
1601                         enabled = crtc;
1602                 else
1603                         enabled = NULL;
1604         } else
1605                 planeb_wm = fifo_size - wm_info->guard_size;
1606
1607         DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1608
1609         /*
1610          * Overlay gets an aggressive default since video jitter is bad.
1611          */
1612         cwm = 2;
1613
1614         /* Play safe and disable self-refresh before adjusting watermarks. */
1615         if (IS_I945G(dev) || IS_I945GM(dev))
1616                 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN_MASK | 0);
1617         else if (IS_I915GM(dev))
1618                 I915_WRITE(INSTPM, I915_READ(INSTPM) & ~INSTPM_SELF_EN);
1619
1620         /* Calc sr entries for one plane configs */
1621         if (HAS_FW_BLC(dev) && enabled) {
1622                 /* self-refresh has much higher latency */
1623                 static const int sr_latency_ns = 6000;
1624                 const struct drm_display_mode *adjusted_mode =
1625                         &to_intel_crtc(enabled)->config.adjusted_mode;
1626                 int clock = adjusted_mode->crtc_clock;
1627                 int htotal = adjusted_mode->htotal;
1628                 int hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1629                 int pixel_size = enabled->fb->bits_per_pixel / 8;
1630                 unsigned long line_time_us;
1631                 int entries;
1632
1633                 line_time_us = (htotal * 1000) / clock;
1634
1635                 /* Use ns/us then divide to preserve precision */
1636                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1637                         pixel_size * hdisplay;
1638                 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1639                 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1640                 srwm = wm_info->fifo_size - entries;
1641                 if (srwm < 0)
1642                         srwm = 1;
1643
1644                 if (IS_I945G(dev) || IS_I945GM(dev))
1645                         I915_WRITE(FW_BLC_SELF,
1646                                    FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1647                 else if (IS_I915GM(dev))
1648                         I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1649         }
1650
1651         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1652                       planea_wm, planeb_wm, cwm, srwm);
1653
1654         fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1655         fwater_hi = (cwm & 0x1f);
1656
1657         /* Set request length to 8 cachelines per fetch */
1658         fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1659         fwater_hi = fwater_hi | (1 << 8);
1660
1661         I915_WRITE(FW_BLC, fwater_lo);
1662         I915_WRITE(FW_BLC2, fwater_hi);
1663
1664         if (HAS_FW_BLC(dev)) {
1665                 if (enabled) {
1666                         if (IS_I945G(dev) || IS_I945GM(dev))
1667                                 I915_WRITE(FW_BLC_SELF,
1668                                            FW_BLC_SELF_EN_MASK | FW_BLC_SELF_EN);
1669                         else if (IS_I915GM(dev))
1670                                 I915_WRITE(INSTPM, I915_READ(INSTPM) | INSTPM_SELF_EN);
1671                         DRM_DEBUG_KMS("memory self refresh enabled\n");
1672                 } else
1673                         DRM_DEBUG_KMS("memory self refresh disabled\n");
1674         }
1675 }
1676
1677 static void i830_update_wm(struct drm_crtc *unused_crtc)
1678 {
1679         struct drm_device *dev = unused_crtc->dev;
1680         struct drm_i915_private *dev_priv = dev->dev_private;
1681         struct drm_crtc *crtc;
1682         const struct drm_display_mode *adjusted_mode;
1683         uint32_t fwater_lo;
1684         int planea_wm;
1685
1686         crtc = single_enabled_crtc(dev);
1687         if (crtc == NULL)
1688                 return;
1689
1690         adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1691         planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1692                                        &i830_wm_info,
1693                                        dev_priv->display.get_fifo_size(dev, 0),
1694                                        4, latency_ns);
1695         fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1696         fwater_lo |= (3<<8) | planea_wm;
1697
1698         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1699
1700         I915_WRITE(FW_BLC, fwater_lo);
1701 }
1702
1703 /*
1704  * Check the wm result.
1705  *
1706  * If any calculated watermark values is larger than the maximum value that
1707  * can be programmed into the associated watermark register, that watermark
1708  * must be disabled.
1709  */
1710 static bool ironlake_check_srwm(struct drm_device *dev, int level,
1711                                 int fbc_wm, int display_wm, int cursor_wm,
1712                                 const struct intel_watermark_params *display,
1713                                 const struct intel_watermark_params *cursor)
1714 {
1715         struct drm_i915_private *dev_priv = dev->dev_private;
1716
1717         DRM_DEBUG_KMS("watermark %d: display plane %d, fbc lines %d,"
1718                       " cursor %d\n", level, display_wm, fbc_wm, cursor_wm);
1719
1720         if (fbc_wm > SNB_FBC_MAX_SRWM) {
1721                 DRM_DEBUG_KMS("fbc watermark(%d) is too large(%d), disabling wm%d+\n",
1722                               fbc_wm, SNB_FBC_MAX_SRWM, level);
1723
1724                 /* fbc has it's own way to disable FBC WM */
1725                 I915_WRITE(DISP_ARB_CTL,
1726                            I915_READ(DISP_ARB_CTL) | DISP_FBC_WM_DIS);
1727                 return false;
1728         } else if (INTEL_INFO(dev)->gen >= 6) {
1729                 /* enable FBC WM (except on ILK, where it must remain off) */
1730                 I915_WRITE(DISP_ARB_CTL,
1731                            I915_READ(DISP_ARB_CTL) & ~DISP_FBC_WM_DIS);
1732         }
1733
1734         if (display_wm > display->max_wm) {
1735                 DRM_DEBUG_KMS("display watermark(%d) is too large(%d), disabling wm%d+\n",
1736                               display_wm, SNB_DISPLAY_MAX_SRWM, level);
1737                 return false;
1738         }
1739
1740         if (cursor_wm > cursor->max_wm) {
1741                 DRM_DEBUG_KMS("cursor watermark(%d) is too large(%d), disabling wm%d+\n",
1742                               cursor_wm, SNB_CURSOR_MAX_SRWM, level);
1743                 return false;
1744         }
1745
1746         if (!(fbc_wm || display_wm || cursor_wm)) {
1747                 DRM_DEBUG_KMS("latency %d is 0, disabling wm%d+\n", level, level);
1748                 return false;
1749         }
1750
1751         return true;
1752 }
1753
1754 /*
1755  * Compute watermark values of WM[1-3],
1756  */
1757 static bool ironlake_compute_srwm(struct drm_device *dev, int level, int plane,
1758                                   int latency_ns,
1759                                   const struct intel_watermark_params *display,
1760                                   const struct intel_watermark_params *cursor,
1761                                   int *fbc_wm, int *display_wm, int *cursor_wm)
1762 {
1763         struct drm_crtc *crtc;
1764         const struct drm_display_mode *adjusted_mode;
1765         unsigned long line_time_us;
1766         int hdisplay, htotal, pixel_size, clock;
1767         int line_count, line_size;
1768         int small, large;
1769         int entries;
1770
1771         if (!latency_ns) {
1772                 *fbc_wm = *display_wm = *cursor_wm = 0;
1773                 return false;
1774         }
1775
1776         crtc = intel_get_crtc_for_plane(dev, plane);
1777         adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1778         clock = adjusted_mode->crtc_clock;
1779         htotal = adjusted_mode->htotal;
1780         hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1781         pixel_size = crtc->fb->bits_per_pixel / 8;
1782
1783         line_time_us = (htotal * 1000) / clock;
1784         line_count = (latency_ns / line_time_us + 1000) / 1000;
1785         line_size = hdisplay * pixel_size;
1786
1787         /* Use the minimum of the small and large buffer method for primary */
1788         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1789         large = line_count * line_size;
1790
1791         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1792         *display_wm = entries + display->guard_size;
1793
1794         /*
1795          * Spec says:
1796          * FBC WM = ((Final Primary WM * 64) / number of bytes per line) + 2
1797          */
1798         *fbc_wm = DIV_ROUND_UP(*display_wm * 64, line_size) + 2;
1799
1800         /* calculate the self-refresh watermark for display cursor */
1801         entries = line_count * pixel_size * 64;
1802         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1803         *cursor_wm = entries + cursor->guard_size;
1804
1805         return ironlake_check_srwm(dev, level,
1806                                    *fbc_wm, *display_wm, *cursor_wm,
1807                                    display, cursor);
1808 }
1809
1810 static void ironlake_update_wm(struct drm_crtc *crtc)
1811 {
1812         struct drm_device *dev = crtc->dev;
1813         struct drm_i915_private *dev_priv = dev->dev_private;
1814         int fbc_wm, plane_wm, cursor_wm;
1815         unsigned int enabled;
1816
1817         enabled = 0;
1818         if (g4x_compute_wm0(dev, PIPE_A,
1819                             &ironlake_display_wm_info,
1820                             dev_priv->wm.pri_latency[0] * 100,
1821                             &ironlake_cursor_wm_info,
1822                             dev_priv->wm.cur_latency[0] * 100,
1823                             &plane_wm, &cursor_wm)) {
1824                 I915_WRITE(WM0_PIPEA_ILK,
1825                            (plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm);
1826                 DRM_DEBUG_KMS("FIFO watermarks For pipe A -"
1827                               " plane %d, " "cursor: %d\n",
1828                               plane_wm, cursor_wm);
1829                 enabled |= 1 << PIPE_A;
1830         }
1831
1832         if (g4x_compute_wm0(dev, PIPE_B,
1833                             &ironlake_display_wm_info,
1834                             dev_priv->wm.pri_latency[0] * 100,
1835                             &ironlake_cursor_wm_info,
1836                             dev_priv->wm.cur_latency[0] * 100,
1837                             &plane_wm, &cursor_wm)) {
1838                 I915_WRITE(WM0_PIPEB_ILK,
1839                            (plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm);
1840                 DRM_DEBUG_KMS("FIFO watermarks For pipe B -"
1841                               " plane %d, cursor: %d\n",
1842                               plane_wm, cursor_wm);
1843                 enabled |= 1 << PIPE_B;
1844         }
1845
1846         /*
1847          * Calculate and update the self-refresh watermark only when one
1848          * display plane is used.
1849          */
1850         I915_WRITE(WM3_LP_ILK, 0);
1851         I915_WRITE(WM2_LP_ILK, 0);
1852         I915_WRITE(WM1_LP_ILK, 0);
1853
1854         if (!single_plane_enabled(enabled))
1855                 return;
1856         enabled = ffs(enabled) - 1;
1857
1858         /* WM1 */
1859         if (!ironlake_compute_srwm(dev, 1, enabled,
1860                                    dev_priv->wm.pri_latency[1] * 500,
1861                                    &ironlake_display_srwm_info,
1862                                    &ironlake_cursor_srwm_info,
1863                                    &fbc_wm, &plane_wm, &cursor_wm))
1864                 return;
1865
1866         I915_WRITE(WM1_LP_ILK,
1867                    WM1_LP_SR_EN |
1868                    (dev_priv->wm.pri_latency[1] << WM1_LP_LATENCY_SHIFT) |
1869                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1870                    (plane_wm << WM1_LP_SR_SHIFT) |
1871                    cursor_wm);
1872
1873         /* WM2 */
1874         if (!ironlake_compute_srwm(dev, 2, enabled,
1875                                    dev_priv->wm.pri_latency[2] * 500,
1876                                    &ironlake_display_srwm_info,
1877                                    &ironlake_cursor_srwm_info,
1878                                    &fbc_wm, &plane_wm, &cursor_wm))
1879                 return;
1880
1881         I915_WRITE(WM2_LP_ILK,
1882                    WM2_LP_EN |
1883                    (dev_priv->wm.pri_latency[2] << WM1_LP_LATENCY_SHIFT) |
1884                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1885                    (plane_wm << WM1_LP_SR_SHIFT) |
1886                    cursor_wm);
1887
1888         /*
1889          * WM3 is unsupported on ILK, probably because we don't have latency
1890          * data for that power state
1891          */
1892 }
1893
1894 static void sandybridge_update_wm(struct drm_crtc *crtc)
1895 {
1896         struct drm_device *dev = crtc->dev;
1897         struct drm_i915_private *dev_priv = dev->dev_private;
1898         int latency = dev_priv->wm.pri_latency[0] * 100;        /* In unit 0.1us */
1899         u32 val;
1900         int fbc_wm, plane_wm, cursor_wm;
1901         unsigned int enabled;
1902
1903         enabled = 0;
1904         if (g4x_compute_wm0(dev, PIPE_A,
1905                             &sandybridge_display_wm_info, latency,
1906                             &sandybridge_cursor_wm_info, latency,
1907                             &plane_wm, &cursor_wm)) {
1908                 val = I915_READ(WM0_PIPEA_ILK);
1909                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
1910                 I915_WRITE(WM0_PIPEA_ILK, val |
1911                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
1912                 DRM_DEBUG_KMS("FIFO watermarks For pipe A -"
1913                               " plane %d, " "cursor: %d\n",
1914                               plane_wm, cursor_wm);
1915                 enabled |= 1 << PIPE_A;
1916         }
1917
1918         if (g4x_compute_wm0(dev, PIPE_B,
1919                             &sandybridge_display_wm_info, latency,
1920                             &sandybridge_cursor_wm_info, latency,
1921                             &plane_wm, &cursor_wm)) {
1922                 val = I915_READ(WM0_PIPEB_ILK);
1923                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
1924                 I915_WRITE(WM0_PIPEB_ILK, val |
1925                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
1926                 DRM_DEBUG_KMS("FIFO watermarks For pipe B -"
1927                               " plane %d, cursor: %d\n",
1928                               plane_wm, cursor_wm);
1929                 enabled |= 1 << PIPE_B;
1930         }
1931
1932         /*
1933          * Calculate and update the self-refresh watermark only when one
1934          * display plane is used.
1935          *
1936          * SNB support 3 levels of watermark.
1937          *
1938          * WM1/WM2/WM2 watermarks have to be enabled in the ascending order,
1939          * and disabled in the descending order
1940          *
1941          */
1942         I915_WRITE(WM3_LP_ILK, 0);
1943         I915_WRITE(WM2_LP_ILK, 0);
1944         I915_WRITE(WM1_LP_ILK, 0);
1945
1946         if (!single_plane_enabled(enabled) ||
1947             dev_priv->sprite_scaling_enabled)
1948                 return;
1949         enabled = ffs(enabled) - 1;
1950
1951         /* WM1 */
1952         if (!ironlake_compute_srwm(dev, 1, enabled,
1953                                    dev_priv->wm.pri_latency[1] * 500,
1954                                    &sandybridge_display_srwm_info,
1955                                    &sandybridge_cursor_srwm_info,
1956                                    &fbc_wm, &plane_wm, &cursor_wm))
1957                 return;
1958
1959         I915_WRITE(WM1_LP_ILK,
1960                    WM1_LP_SR_EN |
1961                    (dev_priv->wm.pri_latency[1] << WM1_LP_LATENCY_SHIFT) |
1962                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1963                    (plane_wm << WM1_LP_SR_SHIFT) |
1964                    cursor_wm);
1965
1966         /* WM2 */
1967         if (!ironlake_compute_srwm(dev, 2, enabled,
1968                                    dev_priv->wm.pri_latency[2] * 500,
1969                                    &sandybridge_display_srwm_info,
1970                                    &sandybridge_cursor_srwm_info,
1971                                    &fbc_wm, &plane_wm, &cursor_wm))
1972                 return;
1973
1974         I915_WRITE(WM2_LP_ILK,
1975                    WM2_LP_EN |
1976                    (dev_priv->wm.pri_latency[2] << WM1_LP_LATENCY_SHIFT) |
1977                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1978                    (plane_wm << WM1_LP_SR_SHIFT) |
1979                    cursor_wm);
1980
1981         /* WM3 */
1982         if (!ironlake_compute_srwm(dev, 3, enabled,
1983                                    dev_priv->wm.pri_latency[3] * 500,
1984                                    &sandybridge_display_srwm_info,
1985                                    &sandybridge_cursor_srwm_info,
1986                                    &fbc_wm, &plane_wm, &cursor_wm))
1987                 return;
1988
1989         I915_WRITE(WM3_LP_ILK,
1990                    WM3_LP_EN |
1991                    (dev_priv->wm.pri_latency[3] << WM1_LP_LATENCY_SHIFT) |
1992                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1993                    (plane_wm << WM1_LP_SR_SHIFT) |
1994                    cursor_wm);
1995 }
1996
1997 static void ivybridge_update_wm(struct drm_crtc *crtc)
1998 {
1999         struct drm_device *dev = crtc->dev;
2000         struct drm_i915_private *dev_priv = dev->dev_private;
2001         int latency = dev_priv->wm.pri_latency[0] * 100;        /* In unit 0.1us */
2002         u32 val;
2003         int fbc_wm, plane_wm, cursor_wm;
2004         int ignore_fbc_wm, ignore_plane_wm, ignore_cursor_wm;
2005         unsigned int enabled;
2006
2007         enabled = 0;
2008         if (g4x_compute_wm0(dev, PIPE_A,
2009                             &sandybridge_display_wm_info, latency,
2010                             &sandybridge_cursor_wm_info, latency,
2011                             &plane_wm, &cursor_wm)) {
2012                 val = I915_READ(WM0_PIPEA_ILK);
2013                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
2014                 I915_WRITE(WM0_PIPEA_ILK, val |
2015                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
2016                 DRM_DEBUG_KMS("FIFO watermarks For pipe A -"
2017                               " plane %d, " "cursor: %d\n",
2018                               plane_wm, cursor_wm);
2019                 enabled |= 1 << PIPE_A;
2020         }
2021
2022         if (g4x_compute_wm0(dev, PIPE_B,
2023                             &sandybridge_display_wm_info, latency,
2024                             &sandybridge_cursor_wm_info, latency,
2025                             &plane_wm, &cursor_wm)) {
2026                 val = I915_READ(WM0_PIPEB_ILK);
2027                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
2028                 I915_WRITE(WM0_PIPEB_ILK, val |
2029                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
2030                 DRM_DEBUG_KMS("FIFO watermarks For pipe B -"
2031                               " plane %d, cursor: %d\n",
2032                               plane_wm, cursor_wm);
2033                 enabled |= 1 << PIPE_B;
2034         }
2035
2036         if (g4x_compute_wm0(dev, PIPE_C,
2037                             &sandybridge_display_wm_info, latency,
2038                             &sandybridge_cursor_wm_info, latency,
2039                             &plane_wm, &cursor_wm)) {
2040                 val = I915_READ(WM0_PIPEC_IVB);
2041                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
2042                 I915_WRITE(WM0_PIPEC_IVB, val |
2043                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
2044                 DRM_DEBUG_KMS("FIFO watermarks For pipe C -"
2045                               " plane %d, cursor: %d\n",
2046                               plane_wm, cursor_wm);
2047                 enabled |= 1 << PIPE_C;
2048         }
2049
2050         /*
2051          * Calculate and update the self-refresh watermark only when one
2052          * display plane is used.
2053          *
2054          * SNB support 3 levels of watermark.
2055          *
2056          * WM1/WM2/WM2 watermarks have to be enabled in the ascending order,
2057          * and disabled in the descending order
2058          *
2059          */
2060         I915_WRITE(WM3_LP_ILK, 0);
2061         I915_WRITE(WM2_LP_ILK, 0);
2062         I915_WRITE(WM1_LP_ILK, 0);
2063
2064         if (!single_plane_enabled(enabled) ||
2065             dev_priv->sprite_scaling_enabled)
2066                 return;
2067         enabled = ffs(enabled) - 1;
2068
2069         /* WM1 */
2070         if (!ironlake_compute_srwm(dev, 1, enabled,
2071                                    dev_priv->wm.pri_latency[1] * 500,
2072                                    &sandybridge_display_srwm_info,
2073                                    &sandybridge_cursor_srwm_info,
2074                                    &fbc_wm, &plane_wm, &cursor_wm))
2075                 return;
2076
2077         I915_WRITE(WM1_LP_ILK,
2078                    WM1_LP_SR_EN |
2079                    (dev_priv->wm.pri_latency[1] << WM1_LP_LATENCY_SHIFT) |
2080                    (fbc_wm << WM1_LP_FBC_SHIFT) |
2081                    (plane_wm << WM1_LP_SR_SHIFT) |
2082                    cursor_wm);
2083
2084         /* WM2 */
2085         if (!ironlake_compute_srwm(dev, 2, enabled,
2086                                    dev_priv->wm.pri_latency[2] * 500,
2087                                    &sandybridge_display_srwm_info,
2088                                    &sandybridge_cursor_srwm_info,
2089                                    &fbc_wm, &plane_wm, &cursor_wm))
2090                 return;
2091
2092         I915_WRITE(WM2_LP_ILK,
2093                    WM2_LP_EN |
2094                    (dev_priv->wm.pri_latency[2] << WM1_LP_LATENCY_SHIFT) |
2095                    (fbc_wm << WM1_LP_FBC_SHIFT) |
2096                    (plane_wm << WM1_LP_SR_SHIFT) |
2097                    cursor_wm);
2098
2099         /* WM3, note we have to correct the cursor latency */
2100         if (!ironlake_compute_srwm(dev, 3, enabled,
2101                                    dev_priv->wm.pri_latency[3] * 500,
2102                                    &sandybridge_display_srwm_info,
2103                                    &sandybridge_cursor_srwm_info,
2104                                    &fbc_wm, &plane_wm, &ignore_cursor_wm) ||
2105             !ironlake_compute_srwm(dev, 3, enabled,
2106                                    dev_priv->wm.cur_latency[3] * 500,
2107                                    &sandybridge_display_srwm_info,
2108                                    &sandybridge_cursor_srwm_info,
2109                                    &ignore_fbc_wm, &ignore_plane_wm, &cursor_wm))
2110                 return;
2111
2112         I915_WRITE(WM3_LP_ILK,
2113                    WM3_LP_EN |
2114                    (dev_priv->wm.pri_latency[3] << WM1_LP_LATENCY_SHIFT) |
2115                    (fbc_wm << WM1_LP_FBC_SHIFT) |
2116                    (plane_wm << WM1_LP_SR_SHIFT) |
2117                    cursor_wm);
2118 }
2119
2120 static uint32_t ilk_pipe_pixel_rate(struct drm_device *dev,
2121                                     struct drm_crtc *crtc)
2122 {
2123         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2124         uint32_t pixel_rate;
2125
2126         pixel_rate = intel_crtc->config.adjusted_mode.crtc_clock;
2127
2128         /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
2129          * adjust the pixel_rate here. */
2130
2131         if (intel_crtc->config.pch_pfit.enabled) {
2132                 uint64_t pipe_w, pipe_h, pfit_w, pfit_h;
2133                 uint32_t pfit_size = intel_crtc->config.pch_pfit.size;
2134
2135                 pipe_w = intel_crtc->config.pipe_src_w;
2136                 pipe_h = intel_crtc->config.pipe_src_h;
2137                 pfit_w = (pfit_size >> 16) & 0xFFFF;
2138                 pfit_h = pfit_size & 0xFFFF;
2139                 if (pipe_w < pfit_w)
2140                         pipe_w = pfit_w;
2141                 if (pipe_h < pfit_h)
2142                         pipe_h = pfit_h;
2143
2144                 pixel_rate = div_u64((uint64_t) pixel_rate * pipe_w * pipe_h,
2145                                      pfit_w * pfit_h);
2146         }
2147
2148         return pixel_rate;
2149 }
2150
2151 /* latency must be in 0.1us units. */
2152 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
2153                                uint32_t latency)
2154 {
2155         uint64_t ret;
2156
2157         if (WARN(latency == 0, "Latency value missing\n"))
2158                 return UINT_MAX;
2159
2160         ret = (uint64_t) pixel_rate * bytes_per_pixel * latency;
2161         ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
2162
2163         return ret;
2164 }
2165
2166 /* latency must be in 0.1us units. */
2167 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
2168                                uint32_t horiz_pixels, uint8_t bytes_per_pixel,
2169                                uint32_t latency)
2170 {
2171         uint32_t ret;
2172
2173         if (WARN(latency == 0, "Latency value missing\n"))
2174                 return UINT_MAX;
2175
2176         ret = (latency * pixel_rate) / (pipe_htotal * 10000);
2177         ret = (ret + 1) * horiz_pixels * bytes_per_pixel;
2178         ret = DIV_ROUND_UP(ret, 64) + 2;
2179         return ret;
2180 }
2181
2182 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
2183                            uint8_t bytes_per_pixel)
2184 {
2185         return DIV_ROUND_UP(pri_val * 64, horiz_pixels * bytes_per_pixel) + 2;
2186 }
2187
2188 struct hsw_pipe_wm_parameters {
2189         bool active;
2190         uint32_t pipe_htotal;
2191         uint32_t pixel_rate;
2192         struct intel_plane_wm_parameters pri;
2193         struct intel_plane_wm_parameters spr;
2194         struct intel_plane_wm_parameters cur;
2195 };
2196
2197 struct hsw_wm_maximums {
2198         uint16_t pri;
2199         uint16_t spr;
2200         uint16_t cur;
2201         uint16_t fbc;
2202 };
2203
2204 /* used in computing the new watermarks state */
2205 struct intel_wm_config {
2206         unsigned int num_pipes_active;
2207         bool sprites_enabled;
2208         bool sprites_scaled;
2209 };
2210
2211 /*
2212  * For both WM_PIPE and WM_LP.
2213  * mem_value must be in 0.1us units.
2214  */
2215 static uint32_t ilk_compute_pri_wm(const struct hsw_pipe_wm_parameters *params,
2216                                    uint32_t mem_value,
2217                                    bool is_lp)
2218 {
2219         uint32_t method1, method2;
2220
2221         if (!params->active || !params->pri.enabled)
2222                 return 0;
2223
2224         method1 = ilk_wm_method1(params->pixel_rate,
2225                                  params->pri.bytes_per_pixel,
2226                                  mem_value);
2227
2228         if (!is_lp)
2229                 return method1;
2230
2231         method2 = ilk_wm_method2(params->pixel_rate,
2232                                  params->pipe_htotal,
2233                                  params->pri.horiz_pixels,
2234                                  params->pri.bytes_per_pixel,
2235                                  mem_value);
2236
2237         return min(method1, method2);
2238 }
2239
2240 /*
2241  * For both WM_PIPE and WM_LP.
2242  * mem_value must be in 0.1us units.
2243  */
2244 static uint32_t ilk_compute_spr_wm(const struct hsw_pipe_wm_parameters *params,
2245                                    uint32_t mem_value)
2246 {
2247         uint32_t method1, method2;
2248
2249         if (!params->active || !params->spr.enabled)
2250                 return 0;
2251
2252         method1 = ilk_wm_method1(params->pixel_rate,
2253                                  params->spr.bytes_per_pixel,
2254                                  mem_value);
2255         method2 = ilk_wm_method2(params->pixel_rate,
2256                                  params->pipe_htotal,
2257                                  params->spr.horiz_pixels,
2258                                  params->spr.bytes_per_pixel,
2259                                  mem_value);
2260         return min(method1, method2);
2261 }
2262
2263 /*
2264  * For both WM_PIPE and WM_LP.
2265  * mem_value must be in 0.1us units.
2266  */
2267 static uint32_t ilk_compute_cur_wm(const struct hsw_pipe_wm_parameters *params,
2268                                    uint32_t mem_value)
2269 {
2270         if (!params->active || !params->cur.enabled)
2271                 return 0;
2272
2273         return ilk_wm_method2(params->pixel_rate,
2274                               params->pipe_htotal,
2275                               params->cur.horiz_pixels,
2276                               params->cur.bytes_per_pixel,
2277                               mem_value);
2278 }
2279
2280 /* Only for WM_LP. */
2281 static uint32_t ilk_compute_fbc_wm(const struct hsw_pipe_wm_parameters *params,
2282                                    uint32_t pri_val)
2283 {
2284         if (!params->active || !params->pri.enabled)
2285                 return 0;
2286
2287         return ilk_wm_fbc(pri_val,
2288                           params->pri.horiz_pixels,
2289                           params->pri.bytes_per_pixel);
2290 }
2291
2292 static unsigned int ilk_display_fifo_size(const struct drm_device *dev)
2293 {
2294         if (INTEL_INFO(dev)->gen >= 7)
2295                 return 768;
2296         else
2297                 return 512;
2298 }
2299
2300 /* Calculate the maximum primary/sprite plane watermark */
2301 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
2302                                      int level,
2303                                      const struct intel_wm_config *config,
2304                                      enum intel_ddb_partitioning ddb_partitioning,
2305                                      bool is_sprite)
2306 {
2307         unsigned int fifo_size = ilk_display_fifo_size(dev);
2308         unsigned int max;
2309
2310         /* if sprites aren't enabled, sprites get nothing */
2311         if (is_sprite && !config->sprites_enabled)
2312                 return 0;
2313
2314         /* HSW allows LP1+ watermarks even with multiple pipes */
2315         if (level == 0 || config->num_pipes_active > 1) {
2316                 fifo_size /= INTEL_INFO(dev)->num_pipes;
2317
2318                 /*
2319                  * For some reason the non self refresh
2320                  * FIFO size is only half of the self
2321                  * refresh FIFO size on ILK/SNB.
2322                  */
2323                 if (INTEL_INFO(dev)->gen <= 6)
2324                         fifo_size /= 2;
2325         }
2326
2327         if (config->sprites_enabled) {
2328                 /* level 0 is always calculated with 1:1 split */
2329                 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
2330                         if (is_sprite)
2331                                 fifo_size *= 5;
2332                         fifo_size /= 6;
2333                 } else {
2334                         fifo_size /= 2;
2335                 }
2336         }
2337
2338         /* clamp to max that the registers can hold */
2339         if (INTEL_INFO(dev)->gen >= 7)
2340                 /* IVB/HSW primary/sprite plane watermarks */
2341                 max = level == 0 ? 127 : 1023;
2342         else if (!is_sprite)
2343                 /* ILK/SNB primary plane watermarks */
2344                 max = level == 0 ? 127 : 511;
2345         else
2346                 /* ILK/SNB sprite plane watermarks */
2347                 max = level == 0 ? 63 : 255;
2348
2349         return min(fifo_size, max);
2350 }
2351
2352 /* Calculate the maximum cursor plane watermark */
2353 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
2354                                       int level,
2355                                       const struct intel_wm_config *config)
2356 {
2357         /* HSW LP1+ watermarks w/ multiple pipes */
2358         if (level > 0 && config->num_pipes_active > 1)
2359                 return 64;
2360
2361         /* otherwise just report max that registers can hold */
2362         if (INTEL_INFO(dev)->gen >= 7)
2363                 return level == 0 ? 63 : 255;
2364         else
2365                 return level == 0 ? 31 : 63;
2366 }
2367
2368 /* Calculate the maximum FBC watermark */
2369 static unsigned int ilk_fbc_wm_max(void)
2370 {
2371         /* max that registers can hold */
2372         return 15;
2373 }
2374
2375 static void ilk_compute_wm_maximums(struct drm_device *dev,
2376                                     int level,
2377                                     const struct intel_wm_config *config,
2378                                     enum intel_ddb_partitioning ddb_partitioning,
2379                                     struct hsw_wm_maximums *max)
2380 {
2381         max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
2382         max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
2383         max->cur = ilk_cursor_wm_max(dev, level, config);
2384         max->fbc = ilk_fbc_wm_max();
2385 }
2386
2387 static bool ilk_validate_wm_level(int level,
2388                                   const struct hsw_wm_maximums *max,
2389                                   struct intel_wm_level *result)
2390 {
2391         bool ret;
2392
2393         /* already determined to be invalid? */
2394         if (!result->enable)
2395                 return false;
2396
2397         result->enable = result->pri_val <= max->pri &&
2398                          result->spr_val <= max->spr &&
2399                          result->cur_val <= max->cur;
2400
2401         ret = result->enable;
2402
2403         /*
2404          * HACK until we can pre-compute everything,
2405          * and thus fail gracefully if LP0 watermarks
2406          * are exceeded...
2407          */
2408         if (level == 0 && !result->enable) {
2409                 if (result->pri_val > max->pri)
2410                         DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
2411                                       level, result->pri_val, max->pri);
2412                 if (result->spr_val > max->spr)
2413                         DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
2414                                       level, result->spr_val, max->spr);
2415                 if (result->cur_val > max->cur)
2416                         DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
2417                                       level, result->cur_val, max->cur);
2418
2419                 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
2420                 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
2421                 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
2422                 result->enable = true;
2423         }
2424
2425         return ret;
2426 }
2427
2428 static void ilk_compute_wm_level(struct drm_i915_private *dev_priv,
2429                                  int level,
2430                                  const struct hsw_pipe_wm_parameters *p,
2431                                  struct intel_wm_level *result)
2432 {
2433         uint16_t pri_latency = dev_priv->wm.pri_latency[level];
2434         uint16_t spr_latency = dev_priv->wm.spr_latency[level];
2435         uint16_t cur_latency = dev_priv->wm.cur_latency[level];
2436
2437         /* WM1+ latency values stored in 0.5us units */
2438         if (level > 0) {
2439                 pri_latency *= 5;
2440                 spr_latency *= 5;
2441                 cur_latency *= 5;
2442         }
2443
2444         result->pri_val = ilk_compute_pri_wm(p, pri_latency, level);
2445         result->spr_val = ilk_compute_spr_wm(p, spr_latency);
2446         result->cur_val = ilk_compute_cur_wm(p, cur_latency);
2447         result->fbc_val = ilk_compute_fbc_wm(p, result->pri_val);
2448         result->enable = true;
2449 }
2450
2451 static uint32_t
2452 hsw_compute_linetime_wm(struct drm_device *dev, struct drm_crtc *crtc)
2453 {
2454         struct drm_i915_private *dev_priv = dev->dev_private;
2455         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2456         struct drm_display_mode *mode = &intel_crtc->config.adjusted_mode;
2457         u32 linetime, ips_linetime;
2458
2459         if (!intel_crtc_active(crtc))
2460                 return 0;
2461
2462         /* The WM are computed with base on how long it takes to fill a single
2463          * row at the given clock rate, multiplied by 8.
2464          * */
2465         linetime = DIV_ROUND_CLOSEST(mode->htotal * 1000 * 8, mode->clock);
2466         ips_linetime = DIV_ROUND_CLOSEST(mode->htotal * 1000 * 8,
2467                                          intel_ddi_get_cdclk_freq(dev_priv));
2468
2469         return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2470                PIPE_WM_LINETIME_TIME(linetime);
2471 }
2472
2473 static void intel_read_wm_latency(struct drm_device *dev, uint16_t wm[5])
2474 {
2475         struct drm_i915_private *dev_priv = dev->dev_private;
2476
2477         if (IS_HASWELL(dev)) {
2478                 uint64_t sskpd = I915_READ64(MCH_SSKPD);
2479
2480                 wm[0] = (sskpd >> 56) & 0xFF;
2481                 if (wm[0] == 0)
2482                         wm[0] = sskpd & 0xF;
2483                 wm[1] = (sskpd >> 4) & 0xFF;
2484                 wm[2] = (sskpd >> 12) & 0xFF;
2485                 wm[3] = (sskpd >> 20) & 0x1FF;
2486                 wm[4] = (sskpd >> 32) & 0x1FF;
2487         } else if (INTEL_INFO(dev)->gen >= 6) {
2488                 uint32_t sskpd = I915_READ(MCH_SSKPD);
2489
2490                 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2491                 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2492                 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2493                 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2494         } else if (INTEL_INFO(dev)->gen >= 5) {
2495                 uint32_t mltr = I915_READ(MLTR_ILK);
2496
2497                 /* ILK primary LP0 latency is 700 ns */
2498                 wm[0] = 7;
2499                 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2500                 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2501         }
2502 }
2503
2504 static void intel_fixup_spr_wm_latency(struct drm_device *dev, uint16_t wm[5])
2505 {
2506         /* ILK sprite LP0 latency is 1300 ns */
2507         if (INTEL_INFO(dev)->gen == 5)
2508                 wm[0] = 13;
2509 }
2510
2511 static void intel_fixup_cur_wm_latency(struct drm_device *dev, uint16_t wm[5])
2512 {
2513         /* ILK cursor LP0 latency is 1300 ns */
2514         if (INTEL_INFO(dev)->gen == 5)
2515                 wm[0] = 13;
2516
2517         /* WaDoubleCursorLP3Latency:ivb */
2518         if (IS_IVYBRIDGE(dev))
2519                 wm[3] *= 2;
2520 }
2521
2522 static int ilk_wm_max_level(const struct drm_device *dev)
2523 {
2524         /* how many WM levels are we expecting */
2525         if (IS_HASWELL(dev))
2526                 return 4;
2527         else if (INTEL_INFO(dev)->gen >= 6)
2528                 return 3;
2529         else
2530                 return 2;
2531 }
2532
2533 static void intel_print_wm_latency(struct drm_device *dev,
2534                                    const char *name,
2535                                    const uint16_t wm[5])
2536 {
2537         int level, max_level = ilk_wm_max_level(dev);
2538
2539         for (level = 0; level <= max_level; level++) {
2540                 unsigned int latency = wm[level];
2541
2542                 if (latency == 0) {
2543                         DRM_ERROR("%s WM%d latency not provided\n",
2544                                   name, level);
2545                         continue;
2546                 }
2547
2548                 /* WM1+ latency values in 0.5us units */
2549                 if (level > 0)
2550                         latency *= 5;
2551
2552                 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2553                               name, level, wm[level],
2554                               latency / 10, latency % 10);
2555         }
2556 }
2557
2558 static void intel_setup_wm_latency(struct drm_device *dev)
2559 {
2560         struct drm_i915_private *dev_priv = dev->dev_private;
2561
2562         intel_read_wm_latency(dev, dev_priv->wm.pri_latency);
2563
2564         memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2565                sizeof(dev_priv->wm.pri_latency));
2566         memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2567                sizeof(dev_priv->wm.pri_latency));
2568
2569         intel_fixup_spr_wm_latency(dev, dev_priv->wm.spr_latency);
2570         intel_fixup_cur_wm_latency(dev, dev_priv->wm.cur_latency);
2571
2572         intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2573         intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2574         intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2575 }
2576
2577 static void hsw_compute_wm_parameters(struct drm_crtc *crtc,
2578                                       struct hsw_pipe_wm_parameters *p,
2579                                       struct intel_wm_config *config)
2580 {
2581         struct drm_device *dev = crtc->dev;
2582         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2583         enum pipe pipe = intel_crtc->pipe;
2584         struct drm_plane *plane;
2585
2586         p->active = intel_crtc_active(crtc);
2587         if (p->active) {
2588                 p->pipe_htotal = intel_crtc->config.adjusted_mode.htotal;
2589                 p->pixel_rate = ilk_pipe_pixel_rate(dev, crtc);
2590                 p->pri.bytes_per_pixel = crtc->fb->bits_per_pixel / 8;
2591                 p->cur.bytes_per_pixel = 4;
2592                 p->pri.horiz_pixels = intel_crtc->config.pipe_src_w;
2593                 p->cur.horiz_pixels = 64;
2594                 /* TODO: for now, assume primary and cursor planes are always enabled. */
2595                 p->pri.enabled = true;
2596                 p->cur.enabled = true;
2597         }
2598
2599         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
2600                 config->num_pipes_active += intel_crtc_active(crtc);
2601
2602         list_for_each_entry(plane, &dev->mode_config.plane_list, head) {
2603                 struct intel_plane *intel_plane = to_intel_plane(plane);
2604
2605                 if (intel_plane->pipe == pipe)
2606                         p->spr = intel_plane->wm;
2607
2608                 config->sprites_enabled |= intel_plane->wm.enabled;
2609                 config->sprites_scaled |= intel_plane->wm.scaled;
2610         }
2611 }
2612
2613 /* Compute new watermarks for the pipe */
2614 static bool intel_compute_pipe_wm(struct drm_crtc *crtc,
2615                                   const struct hsw_pipe_wm_parameters *params,
2616                                   struct intel_pipe_wm *pipe_wm)
2617 {
2618         struct drm_device *dev = crtc->dev;
2619         struct drm_i915_private *dev_priv = dev->dev_private;
2620         int level, max_level = ilk_wm_max_level(dev);
2621         /* LP0 watermark maximums depend on this pipe alone */
2622         struct intel_wm_config config = {
2623                 .num_pipes_active = 1,
2624                 .sprites_enabled = params->spr.enabled,
2625                 .sprites_scaled = params->spr.scaled,
2626         };
2627         struct hsw_wm_maximums max;
2628
2629         /* LP0 watermarks always use 1/2 DDB partitioning */
2630         ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
2631
2632         for (level = 0; level <= max_level; level++)
2633                 ilk_compute_wm_level(dev_priv, level, params,
2634                                      &pipe_wm->wm[level]);
2635
2636         pipe_wm->linetime = hsw_compute_linetime_wm(dev, crtc);
2637
2638         /* At least LP0 must be valid */
2639         return ilk_validate_wm_level(0, &max, &pipe_wm->wm[0]);
2640 }
2641
2642 /*
2643  * Merge the watermarks from all active pipes for a specific level.
2644  */
2645 static void ilk_merge_wm_level(struct drm_device *dev,
2646                                int level,
2647                                struct intel_wm_level *ret_wm)
2648 {
2649         const struct intel_crtc *intel_crtc;
2650
2651         list_for_each_entry(intel_crtc, &dev->mode_config.crtc_list, base.head) {
2652                 const struct intel_wm_level *wm =
2653                         &intel_crtc->wm.active.wm[level];
2654
2655                 if (!wm->enable)
2656                         return;
2657
2658                 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
2659                 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
2660                 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
2661                 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
2662         }
2663
2664         ret_wm->enable = true;
2665 }
2666
2667 /*
2668  * Merge all low power watermarks for all active pipes.
2669  */
2670 static void ilk_wm_merge(struct drm_device *dev,
2671                          const struct hsw_wm_maximums *max,
2672                          struct intel_pipe_wm *merged)
2673 {
2674         int level, max_level = ilk_wm_max_level(dev);
2675
2676         merged->fbc_wm_enabled = true;
2677
2678         /* merge each WM1+ level */
2679         for (level = 1; level <= max_level; level++) {
2680                 struct intel_wm_level *wm = &merged->wm[level];
2681
2682                 ilk_merge_wm_level(dev, level, wm);
2683
2684                 if (!ilk_validate_wm_level(level, max, wm))
2685                         break;
2686
2687                 /*
2688                  * The spec says it is preferred to disable
2689                  * FBC WMs instead of disabling a WM level.
2690                  */
2691                 if (wm->fbc_val > max->fbc) {
2692                         merged->fbc_wm_enabled = false;
2693                         wm->fbc_val = 0;
2694                 }
2695         }
2696 }
2697
2698 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
2699 {
2700         /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2701         return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
2702 }
2703
2704 static void hsw_compute_wm_results(struct drm_device *dev,
2705                                    const struct intel_pipe_wm *merged,
2706                                    enum intel_ddb_partitioning partitioning,
2707                                    struct hsw_wm_values *results)
2708 {
2709         struct intel_crtc *intel_crtc;
2710         int level, wm_lp;
2711
2712         results->enable_fbc_wm = merged->fbc_wm_enabled;
2713         results->partitioning = partitioning;
2714
2715         /* LP1+ register values */
2716         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2717                 const struct intel_wm_level *r;
2718
2719                 level = ilk_wm_lp_to_level(wm_lp, merged);
2720
2721                 r = &merged->wm[level];
2722                 if (!r->enable)
2723                         break;
2724
2725                 results->wm_lp[wm_lp - 1] = HSW_WM_LP_VAL(level * 2,
2726                                                           r->fbc_val,
2727                                                           r->pri_val,
2728                                                           r->cur_val);
2729                 results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2730         }
2731
2732         /* LP0 register values */
2733         list_for_each_entry(intel_crtc, &dev->mode_config.crtc_list, base.head) {
2734                 enum pipe pipe = intel_crtc->pipe;
2735                 const struct intel_wm_level *r =
2736                         &intel_crtc->wm.active.wm[0];
2737
2738                 if (WARN_ON(!r->enable))
2739                         continue;
2740
2741                 results->wm_linetime[pipe] = intel_crtc->wm.active.linetime;
2742
2743                 results->wm_pipe[pipe] =
2744                         (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
2745                         (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
2746                         r->cur_val;
2747         }
2748 }
2749
2750 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2751  * case both are at the same level. Prefer r1 in case they're the same. */
2752 static struct intel_pipe_wm *hsw_find_best_result(struct drm_device *dev,
2753                                                   struct intel_pipe_wm *r1,
2754                                                   struct intel_pipe_wm *r2)
2755 {
2756         int level, max_level = ilk_wm_max_level(dev);
2757         int level1 = 0, level2 = 0;
2758
2759         for (level = 1; level <= max_level; level++) {
2760                 if (r1->wm[level].enable)
2761                         level1 = level;
2762                 if (r2->wm[level].enable)
2763                         level2 = level;
2764         }
2765
2766         if (level1 == level2) {
2767                 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
2768                         return r2;
2769                 else
2770                         return r1;
2771         } else if (level1 > level2) {
2772                 return r1;
2773         } else {
2774                 return r2;
2775         }
2776 }
2777
2778 /* dirty bits used to track which watermarks need changes */
2779 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2780 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2781 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2782 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2783 #define WM_DIRTY_FBC (1 << 24)
2784 #define WM_DIRTY_DDB (1 << 25)
2785
2786 static unsigned int ilk_compute_wm_dirty(struct drm_device *dev,
2787                                          const struct hsw_wm_values *old,
2788                                          const struct hsw_wm_values *new)
2789 {
2790         unsigned int dirty = 0;
2791         enum pipe pipe;
2792         int wm_lp;
2793
2794         for_each_pipe(pipe) {
2795                 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
2796                         dirty |= WM_DIRTY_LINETIME(pipe);
2797                         /* Must disable LP1+ watermarks too */
2798                         dirty |= WM_DIRTY_LP_ALL;
2799                 }
2800
2801                 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
2802                         dirty |= WM_DIRTY_PIPE(pipe);
2803                         /* Must disable LP1+ watermarks too */
2804                         dirty |= WM_DIRTY_LP_ALL;
2805                 }
2806         }
2807
2808         if (old->enable_fbc_wm != new->enable_fbc_wm) {
2809                 dirty |= WM_DIRTY_FBC;
2810                 /* Must disable LP1+ watermarks too */
2811                 dirty |= WM_DIRTY_LP_ALL;
2812         }
2813
2814         if (old->partitioning != new->partitioning) {
2815                 dirty |= WM_DIRTY_DDB;
2816                 /* Must disable LP1+ watermarks too */
2817                 dirty |= WM_DIRTY_LP_ALL;
2818         }
2819
2820         /* LP1+ watermarks already deemed dirty, no need to continue */
2821         if (dirty & WM_DIRTY_LP_ALL)
2822                 return dirty;
2823
2824         /* Find the lowest numbered LP1+ watermark in need of an update... */
2825         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2826                 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
2827                     old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
2828                         break;
2829         }
2830
2831         /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2832         for (; wm_lp <= 3; wm_lp++)
2833                 dirty |= WM_DIRTY_LP(wm_lp);
2834
2835         return dirty;
2836 }
2837
2838 /*
2839  * The spec says we shouldn't write when we don't need, because every write
2840  * causes WMs to be re-evaluated, expending some power.
2841  */
2842 static void hsw_write_wm_values(struct drm_i915_private *dev_priv,
2843                                 struct hsw_wm_values *results)
2844 {
2845         struct hsw_wm_values *previous = &dev_priv->wm.hw;
2846         unsigned int dirty;
2847         uint32_t val;
2848
2849         dirty = ilk_compute_wm_dirty(dev_priv->dev, previous, results);
2850         if (!dirty)
2851                 return;
2852
2853         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != 0)
2854                 I915_WRITE(WM3_LP_ILK, 0);
2855         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != 0)
2856                 I915_WRITE(WM2_LP_ILK, 0);
2857         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != 0)
2858                 I915_WRITE(WM1_LP_ILK, 0);
2859
2860         if (dirty & WM_DIRTY_PIPE(PIPE_A))
2861                 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2862         if (dirty & WM_DIRTY_PIPE(PIPE_B))
2863                 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
2864         if (dirty & WM_DIRTY_PIPE(PIPE_C))
2865                 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
2866
2867         if (dirty & WM_DIRTY_LINETIME(PIPE_A))
2868                 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
2869         if (dirty & WM_DIRTY_LINETIME(PIPE_B))
2870                 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
2871         if (dirty & WM_DIRTY_LINETIME(PIPE_C))
2872                 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
2873
2874         if (dirty & WM_DIRTY_DDB) {
2875                 val = I915_READ(WM_MISC);
2876                 if (results->partitioning == INTEL_DDB_PART_1_2)
2877                         val &= ~WM_MISC_DATA_PARTITION_5_6;
2878                 else
2879                         val |= WM_MISC_DATA_PARTITION_5_6;
2880                 I915_WRITE(WM_MISC, val);
2881         }
2882
2883         if (dirty & WM_DIRTY_FBC) {
2884                 val = I915_READ(DISP_ARB_CTL);
2885                 if (results->enable_fbc_wm)
2886                         val &= ~DISP_FBC_WM_DIS;
2887                 else
2888                         val |= DISP_FBC_WM_DIS;
2889                 I915_WRITE(DISP_ARB_CTL, val);
2890         }
2891
2892         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp_spr[0] != results->wm_lp_spr[0])
2893                 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
2894         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
2895                 I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
2896         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
2897                 I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
2898
2899         if (dirty & WM_DIRTY_LP(1) && results->wm_lp[0] != 0)
2900                 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
2901         if (dirty & WM_DIRTY_LP(2) && results->wm_lp[1] != 0)
2902                 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
2903         if (dirty & WM_DIRTY_LP(3) && results->wm_lp[2] != 0)
2904                 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
2905
2906         dev_priv->wm.hw = *results;
2907 }
2908
2909 static void haswell_update_wm(struct drm_crtc *crtc)
2910 {
2911         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2912         struct drm_device *dev = crtc->dev;
2913         struct drm_i915_private *dev_priv = dev->dev_private;
2914         struct hsw_wm_maximums max;
2915         struct hsw_pipe_wm_parameters params = {};
2916         struct hsw_wm_values results = {};
2917         enum intel_ddb_partitioning partitioning;
2918         struct intel_pipe_wm pipe_wm = {};
2919         struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
2920         struct intel_wm_config config = {};
2921
2922         hsw_compute_wm_parameters(crtc, &params, &config);
2923
2924         intel_compute_pipe_wm(crtc, &params, &pipe_wm);
2925
2926         if (!memcmp(&intel_crtc->wm.active, &pipe_wm, sizeof(pipe_wm)))
2927                 return;
2928
2929         intel_crtc->wm.active = pipe_wm;
2930
2931         ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
2932         ilk_wm_merge(dev, &max, &lp_wm_1_2);
2933
2934         /* 5/6 split only in single pipe config on IVB+ */
2935         if (INTEL_INFO(dev)->gen >= 7 &&
2936             config.num_pipes_active == 1 && config.sprites_enabled) {
2937                 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
2938                 ilk_wm_merge(dev, &max, &lp_wm_5_6);
2939
2940                 best_lp_wm = hsw_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
2941         } else {
2942                 best_lp_wm = &lp_wm_1_2;
2943         }
2944
2945         partitioning = (best_lp_wm == &lp_wm_1_2) ?
2946                        INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
2947
2948         hsw_compute_wm_results(dev, best_lp_wm, partitioning, &results);
2949
2950         hsw_write_wm_values(dev_priv, &results);
2951 }
2952
2953 static void haswell_update_sprite_wm(struct drm_plane *plane,
2954                                      struct drm_crtc *crtc,
2955                                      uint32_t sprite_width, int pixel_size,
2956                                      bool enabled, bool scaled)
2957 {
2958         struct intel_plane *intel_plane = to_intel_plane(plane);
2959
2960         intel_plane->wm.enabled = enabled;
2961         intel_plane->wm.scaled = scaled;
2962         intel_plane->wm.horiz_pixels = sprite_width;
2963         intel_plane->wm.bytes_per_pixel = pixel_size;
2964
2965         haswell_update_wm(crtc);
2966 }
2967
2968 static bool
2969 sandybridge_compute_sprite_wm(struct drm_device *dev, int plane,
2970                               uint32_t sprite_width, int pixel_size,
2971                               const struct intel_watermark_params *display,
2972                               int display_latency_ns, int *sprite_wm)
2973 {
2974         struct drm_crtc *crtc;
2975         int clock;
2976         int entries, tlb_miss;
2977
2978         crtc = intel_get_crtc_for_plane(dev, plane);
2979         if (!intel_crtc_active(crtc)) {
2980                 *sprite_wm = display->guard_size;
2981                 return false;
2982         }
2983
2984         clock = to_intel_crtc(crtc)->config.adjusted_mode.crtc_clock;
2985
2986         /* Use the small buffer method to calculate the sprite watermark */
2987         entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
2988         tlb_miss = display->fifo_size*display->cacheline_size -
2989                 sprite_width * 8;
2990         if (tlb_miss > 0)
2991                 entries += tlb_miss;
2992         entries = DIV_ROUND_UP(entries, display->cacheline_size);
2993         *sprite_wm = entries + display->guard_size;
2994         if (*sprite_wm > (int)display->max_wm)
2995                 *sprite_wm = display->max_wm;
2996
2997         return true;
2998 }
2999
3000 static bool
3001 sandybridge_compute_sprite_srwm(struct drm_device *dev, int plane,
3002                                 uint32_t sprite_width, int pixel_size,
3003                                 const struct intel_watermark_params *display,
3004                                 int latency_ns, int *sprite_wm)
3005 {
3006         struct drm_crtc *crtc;
3007         unsigned long line_time_us;
3008         int clock;
3009         int line_count, line_size;
3010         int small, large;
3011         int entries;
3012
3013         if (!latency_ns) {
3014                 *sprite_wm = 0;
3015                 return false;
3016         }
3017
3018         crtc = intel_get_crtc_for_plane(dev, plane);
3019         clock = to_intel_crtc(crtc)->config.adjusted_mode.crtc_clock;
3020         if (!clock) {
3021                 *sprite_wm = 0;
3022                 return false;
3023         }
3024
3025         line_time_us = (sprite_width * 1000) / clock;
3026         if (!line_time_us) {
3027                 *sprite_wm = 0;
3028                 return false;
3029         }
3030
3031         line_count = (latency_ns / line_time_us + 1000) / 1000;
3032         line_size = sprite_width * pixel_size;
3033
3034         /* Use the minimum of the small and large buffer method for primary */
3035         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
3036         large = line_count * line_size;
3037
3038         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
3039         *sprite_wm = entries + display->guard_size;
3040
3041         return *sprite_wm > 0x3ff ? false : true;
3042 }
3043
3044 static void sandybridge_update_sprite_wm(struct drm_plane *plane,
3045                                          struct drm_crtc *crtc,
3046                                          uint32_t sprite_width, int pixel_size,
3047                                          bool enabled, bool scaled)
3048 {
3049         struct drm_device *dev = plane->dev;
3050         struct drm_i915_private *dev_priv = dev->dev_private;
3051         int pipe = to_intel_plane(plane)->pipe;
3052         int latency = dev_priv->wm.spr_latency[0] * 100;        /* In unit 0.1us */
3053         u32 val;
3054         int sprite_wm, reg;
3055         int ret;
3056
3057         if (!enabled)
3058                 return;
3059
3060         switch (pipe) {
3061         case 0:
3062                 reg = WM0_PIPEA_ILK;
3063                 break;
3064         case 1:
3065                 reg = WM0_PIPEB_ILK;
3066                 break;
3067         case 2:
3068                 reg = WM0_PIPEC_IVB;
3069                 break;
3070         default:
3071                 return; /* bad pipe */
3072         }
3073
3074         ret = sandybridge_compute_sprite_wm(dev, pipe, sprite_width, pixel_size,
3075                                             &sandybridge_display_wm_info,
3076                                             latency, &sprite_wm);
3077         if (!ret) {
3078                 DRM_DEBUG_KMS("failed to compute sprite wm for pipe %c\n",
3079                               pipe_name(pipe));
3080                 return;
3081         }
3082
3083         val = I915_READ(reg);
3084         val &= ~WM0_PIPE_SPRITE_MASK;
3085         I915_WRITE(reg, val | (sprite_wm << WM0_PIPE_SPRITE_SHIFT));
3086         DRM_DEBUG_KMS("sprite watermarks For pipe %c - %d\n", pipe_name(pipe), sprite_wm);
3087
3088
3089         ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
3090                                               pixel_size,
3091                                               &sandybridge_display_srwm_info,
3092                                               dev_priv->wm.spr_latency[1] * 500,
3093                                               &sprite_wm);
3094         if (!ret) {
3095                 DRM_DEBUG_KMS("failed to compute sprite lp1 wm on pipe %c\n",
3096                               pipe_name(pipe));
3097                 return;
3098         }
3099         I915_WRITE(WM1S_LP_ILK, sprite_wm);
3100
3101         /* Only IVB has two more LP watermarks for sprite */
3102         if (!IS_IVYBRIDGE(dev))
3103                 return;
3104
3105         ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
3106                                               pixel_size,
3107                                               &sandybridge_display_srwm_info,
3108                                               dev_priv->wm.spr_latency[2] * 500,
3109                                               &sprite_wm);
3110         if (!ret) {
3111                 DRM_DEBUG_KMS("failed to compute sprite lp2 wm on pipe %c\n",
3112                               pipe_name(pipe));
3113                 return;
3114         }
3115         I915_WRITE(WM2S_LP_IVB, sprite_wm);
3116
3117         ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
3118                                               pixel_size,
3119                                               &sandybridge_display_srwm_info,
3120                                               dev_priv->wm.spr_latency[3] * 500,
3121                                               &sprite_wm);
3122         if (!ret) {
3123                 DRM_DEBUG_KMS("failed to compute sprite lp3 wm on pipe %c\n",
3124                               pipe_name(pipe));
3125                 return;
3126         }
3127         I915_WRITE(WM3S_LP_IVB, sprite_wm);
3128 }
3129
3130 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
3131 {
3132         struct drm_device *dev = crtc->dev;
3133         struct drm_i915_private *dev_priv = dev->dev_private;
3134         struct hsw_wm_values *hw = &dev_priv->wm.hw;
3135         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3136         struct intel_pipe_wm *active = &intel_crtc->wm.active;
3137         enum pipe pipe = intel_crtc->pipe;
3138         static const unsigned int wm0_pipe_reg[] = {
3139                 [PIPE_A] = WM0_PIPEA_ILK,
3140                 [PIPE_B] = WM0_PIPEB_ILK,
3141                 [PIPE_C] = WM0_PIPEC_IVB,
3142         };
3143
3144         hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
3145         hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
3146
3147         if (intel_crtc_active(crtc)) {
3148                 u32 tmp = hw->wm_pipe[pipe];
3149
3150                 /*
3151                  * For active pipes LP0 watermark is marked as
3152                  * enabled, and LP1+ watermaks as disabled since
3153                  * we can't really reverse compute them in case
3154                  * multiple pipes are active.
3155                  */
3156                 active->wm[0].enable = true;
3157                 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
3158                 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
3159                 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
3160                 active->linetime = hw->wm_linetime[pipe];
3161         } else {
3162                 int level, max_level = ilk_wm_max_level(dev);
3163
3164                 /*
3165                  * For inactive pipes, all watermark levels
3166                  * should be marked as enabled but zeroed,
3167                  * which is what we'd compute them to.
3168                  */
3169                 for (level = 0; level <= max_level; level++)
3170                         active->wm[level].enable = true;
3171         }
3172 }
3173
3174 void ilk_wm_get_hw_state(struct drm_device *dev)
3175 {
3176         struct drm_i915_private *dev_priv = dev->dev_private;
3177         struct hsw_wm_values *hw = &dev_priv->wm.hw;
3178         struct drm_crtc *crtc;
3179
3180         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
3181                 ilk_pipe_wm_get_hw_state(crtc);
3182
3183         hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
3184         hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
3185         hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
3186
3187         hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
3188         hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
3189         hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
3190
3191         hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
3192                 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
3193
3194         hw->enable_fbc_wm =
3195                 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
3196 }
3197
3198 /**
3199  * intel_update_watermarks - update FIFO watermark values based on current modes
3200  *
3201  * Calculate watermark values for the various WM regs based on current mode
3202  * and plane configuration.
3203  *
3204  * There are several cases to deal with here:
3205  *   - normal (i.e. non-self-refresh)
3206  *   - self-refresh (SR) mode
3207  *   - lines are large relative to FIFO size (buffer can hold up to 2)
3208  *   - lines are small relative to FIFO size (buffer can hold more than 2
3209  *     lines), so need to account for TLB latency
3210  *
3211  *   The normal calculation is:
3212  *     watermark = dotclock * bytes per pixel * latency
3213  *   where latency is platform & configuration dependent (we assume pessimal
3214  *   values here).
3215  *
3216  *   The SR calculation is:
3217  *     watermark = (trunc(latency/line time)+1) * surface width *
3218  *       bytes per pixel
3219  *   where
3220  *     line time = htotal / dotclock
3221  *     surface width = hdisplay for normal plane and 64 for cursor
3222  *   and latency is assumed to be high, as above.
3223  *
3224  * The final value programmed to the register should always be rounded up,
3225  * and include an extra 2 entries to account for clock crossings.
3226  *
3227  * We don't use the sprite, so we can ignore that.  And on Crestline we have
3228  * to set the non-SR watermarks to 8.
3229  */
3230 void intel_update_watermarks(struct drm_crtc *crtc)
3231 {
3232         struct drm_i915_private *dev_priv = crtc->dev->dev_private;
3233
3234         if (dev_priv->display.update_wm)
3235                 dev_priv->display.update_wm(crtc);
3236 }
3237
3238 void intel_update_sprite_watermarks(struct drm_plane *plane,
3239                                     struct drm_crtc *crtc,
3240                                     uint32_t sprite_width, int pixel_size,
3241                                     bool enabled, bool scaled)
3242 {
3243         struct drm_i915_private *dev_priv = plane->dev->dev_private;
3244
3245         if (dev_priv->display.update_sprite_wm)
3246                 dev_priv->display.update_sprite_wm(plane, crtc, sprite_width,
3247                                                    pixel_size, enabled, scaled);
3248 }
3249
3250 static struct drm_i915_gem_object *
3251 intel_alloc_context_page(struct drm_device *dev)
3252 {
3253         struct drm_i915_gem_object *ctx;
3254         int ret;
3255
3256         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
3257
3258         ctx = i915_gem_alloc_object(dev, 4096);
3259         if (!ctx) {
3260                 DRM_DEBUG("failed to alloc power context, RC6 disabled\n");
3261                 return NULL;
3262         }
3263
3264         ret = i915_gem_obj_ggtt_pin(ctx, 4096, true, false);
3265         if (ret) {
3266                 DRM_ERROR("failed to pin power context: %d\n", ret);
3267                 goto err_unref;
3268         }
3269
3270         ret = i915_gem_object_set_to_gtt_domain(ctx, 1);
3271         if (ret) {
3272                 DRM_ERROR("failed to set-domain on power context: %d\n", ret);
3273                 goto err_unpin;
3274         }
3275
3276         return ctx;
3277
3278 err_unpin:
3279         i915_gem_object_unpin(ctx);
3280 err_unref:
3281         drm_gem_object_unreference(&ctx->base);
3282         return NULL;
3283 }
3284
3285 /**
3286  * Lock protecting IPS related data structures
3287  */
3288 DEFINE_SPINLOCK(mchdev_lock);
3289
3290 /* Global for IPS driver to get at the current i915 device. Protected by
3291  * mchdev_lock. */
3292 static struct drm_i915_private *i915_mch_dev;
3293
3294 bool ironlake_set_drps(struct drm_device *dev, u8 val)
3295 {
3296         struct drm_i915_private *dev_priv = dev->dev_private;
3297         u16 rgvswctl;
3298
3299         assert_spin_locked(&mchdev_lock);
3300
3301         rgvswctl = I915_READ16(MEMSWCTL);
3302         if (rgvswctl & MEMCTL_CMD_STS) {
3303                 DRM_DEBUG("gpu busy, RCS change rejected\n");
3304                 return false; /* still busy with another command */
3305         }
3306
3307         rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
3308                 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
3309         I915_WRITE16(MEMSWCTL, rgvswctl);
3310         POSTING_READ16(MEMSWCTL);
3311
3312         rgvswctl |= MEMCTL_CMD_STS;
3313         I915_WRITE16(MEMSWCTL, rgvswctl);
3314
3315         return true;
3316 }
3317
3318 static void ironlake_enable_drps(struct drm_device *dev)
3319 {
3320         struct drm_i915_private *dev_priv = dev->dev_private;
3321         u32 rgvmodectl = I915_READ(MEMMODECTL);
3322         u8 fmax, fmin, fstart, vstart;
3323
3324         spin_lock_irq(&mchdev_lock);
3325
3326         /* Enable temp reporting */
3327         I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
3328         I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
3329
3330         /* 100ms RC evaluation intervals */
3331         I915_WRITE(RCUPEI, 100000);
3332         I915_WRITE(RCDNEI, 100000);
3333
3334         /* Set max/min thresholds to 90ms and 80ms respectively */
3335         I915_WRITE(RCBMAXAVG, 90000);
3336         I915_WRITE(RCBMINAVG, 80000);
3337
3338         I915_WRITE(MEMIHYST, 1);
3339
3340         /* Set up min, max, and cur for interrupt handling */
3341         fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
3342         fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
3343         fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
3344                 MEMMODE_FSTART_SHIFT;
3345
3346         vstart = (I915_READ(PXVFREQ_BASE + (fstart * 4)) & PXVFREQ_PX_MASK) >>
3347                 PXVFREQ_PX_SHIFT;
3348
3349         dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
3350         dev_priv->ips.fstart = fstart;
3351
3352         dev_priv->ips.max_delay = fstart;
3353         dev_priv->ips.min_delay = fmin;
3354         dev_priv->ips.cur_delay = fstart;
3355
3356         DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
3357                          fmax, fmin, fstart);
3358
3359         I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
3360
3361         /*
3362          * Interrupts will be enabled in ironlake_irq_postinstall
3363          */
3364
3365         I915_WRITE(VIDSTART, vstart);
3366         POSTING_READ(VIDSTART);
3367
3368         rgvmodectl |= MEMMODE_SWMODE_EN;
3369         I915_WRITE(MEMMODECTL, rgvmodectl);
3370
3371         if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
3372                 DRM_ERROR("stuck trying to change perf mode\n");
3373         mdelay(1);
3374
3375         ironlake_set_drps(dev, fstart);
3376
3377         dev_priv->ips.last_count1 = I915_READ(0x112e4) + I915_READ(0x112e8) +
3378                 I915_READ(0x112e0);
3379         dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
3380         dev_priv->ips.last_count2 = I915_READ(0x112f4);
3381         getrawmonotonic(&dev_priv->ips.last_time2);
3382
3383         spin_unlock_irq(&mchdev_lock);
3384 }
3385
3386 static void ironlake_disable_drps(struct drm_device *dev)
3387 {
3388         struct drm_i915_private *dev_priv = dev->dev_private;
3389         u16 rgvswctl;
3390
3391         spin_lock_irq(&mchdev_lock);
3392
3393         rgvswctl = I915_READ16(MEMSWCTL);
3394
3395         /* Ack interrupts, disable EFC interrupt */
3396         I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
3397         I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
3398         I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
3399         I915_WRITE(DEIIR, DE_PCU_EVENT);
3400         I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
3401
3402         /* Go back to the starting frequency */
3403         ironlake_set_drps(dev, dev_priv->ips.fstart);
3404         mdelay(1);
3405         rgvswctl |= MEMCTL_CMD_STS;
3406         I915_WRITE(MEMSWCTL, rgvswctl);
3407         mdelay(1);
3408
3409         spin_unlock_irq(&mchdev_lock);
3410 }
3411
3412 /* There's a funny hw issue where the hw returns all 0 when reading from
3413  * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
3414  * ourselves, instead of doing a rmw cycle (which might result in us clearing
3415  * all limits and the gpu stuck at whatever frequency it is at atm).
3416  */
3417 static u32 gen6_rps_limits(struct drm_i915_private *dev_priv, u8 *val)
3418 {
3419         u32 limits;
3420
3421         limits = 0;
3422
3423         if (*val >= dev_priv->rps.max_delay)
3424                 *val = dev_priv->rps.max_delay;
3425         limits |= dev_priv->rps.max_delay << 24;
3426
3427         /* Only set the down limit when we've reached the lowest level to avoid
3428          * getting more interrupts, otherwise leave this clear. This prevents a
3429          * race in the hw when coming out of rc6: There's a tiny window where
3430          * the hw runs at the minimal clock before selecting the desired
3431          * frequency, if the down threshold expires in that window we will not
3432          * receive a down interrupt. */
3433         if (*val <= dev_priv->rps.min_delay) {
3434                 *val = dev_priv->rps.min_delay;
3435                 limits |= dev_priv->rps.min_delay << 16;
3436         }
3437
3438         return limits;
3439 }
3440
3441 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
3442 {
3443         int new_power;
3444
3445         new_power = dev_priv->rps.power;
3446         switch (dev_priv->rps.power) {
3447         case LOW_POWER:
3448                 if (val > dev_priv->rps.rpe_delay + 1 && val > dev_priv->rps.cur_delay)
3449                         new_power = BETWEEN;
3450                 break;
3451
3452         case BETWEEN:
3453                 if (val <= dev_priv->rps.rpe_delay && val < dev_priv->rps.cur_delay)
3454                         new_power = LOW_POWER;
3455                 else if (val >= dev_priv->rps.rp0_delay && val > dev_priv->rps.cur_delay)
3456                         new_power = HIGH_POWER;
3457                 break;
3458
3459         case HIGH_POWER:
3460                 if (val < (dev_priv->rps.rp1_delay + dev_priv->rps.rp0_delay) >> 1 && val < dev_priv->rps.cur_delay)
3461                         new_power = BETWEEN;
3462                 break;
3463         }
3464         /* Max/min bins are special */
3465         if (val == dev_priv->rps.min_delay)
3466                 new_power = LOW_POWER;
3467         if (val == dev_priv->rps.max_delay)
3468                 new_power = HIGH_POWER;
3469         if (new_power == dev_priv->rps.power)
3470                 return;
3471
3472         /* Note the units here are not exactly 1us, but 1280ns. */
3473         switch (new_power) {
3474         case LOW_POWER:
3475                 /* Upclock if more than 95% busy over 16ms */
3476                 I915_WRITE(GEN6_RP_UP_EI, 12500);
3477                 I915_WRITE(GEN6_RP_UP_THRESHOLD, 11800);
3478
3479                 /* Downclock if less than 85% busy over 32ms */
3480                 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3481                 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 21250);
3482
3483                 I915_WRITE(GEN6_RP_CONTROL,
3484                            GEN6_RP_MEDIA_TURBO |
3485                            GEN6_RP_MEDIA_HW_NORMAL_MODE |
3486                            GEN6_RP_MEDIA_IS_GFX |
3487                            GEN6_RP_ENABLE |
3488                            GEN6_RP_UP_BUSY_AVG |
3489                            GEN6_RP_DOWN_IDLE_AVG);
3490                 break;
3491
3492         case BETWEEN:
3493                 /* Upclock if more than 90% busy over 13ms */
3494                 I915_WRITE(GEN6_RP_UP_EI, 10250);
3495                 I915_WRITE(GEN6_RP_UP_THRESHOLD, 9225);
3496
3497                 /* Downclock if less than 75% busy over 32ms */
3498                 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3499                 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 18750);
3500
3501                 I915_WRITE(GEN6_RP_CONTROL,
3502                            GEN6_RP_MEDIA_TURBO |
3503                            GEN6_RP_MEDIA_HW_NORMAL_MODE |
3504                            GEN6_RP_MEDIA_IS_GFX |
3505                            GEN6_RP_ENABLE |
3506                            GEN6_RP_UP_BUSY_AVG |
3507                            GEN6_RP_DOWN_IDLE_AVG);
3508                 break;
3509
3510         case HIGH_POWER:
3511                 /* Upclock if more than 85% busy over 10ms */
3512                 I915_WRITE(GEN6_RP_UP_EI, 8000);
3513                 I915_WRITE(GEN6_RP_UP_THRESHOLD, 6800);
3514
3515                 /* Downclock if less than 60% busy over 32ms */
3516                 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3517                 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 15000);
3518
3519                 I915_WRITE(GEN6_RP_CONTROL,
3520                            GEN6_RP_MEDIA_TURBO |
3521                            GEN6_RP_MEDIA_HW_NORMAL_MODE |
3522                            GEN6_RP_MEDIA_IS_GFX |
3523                            GEN6_RP_ENABLE |
3524                            GEN6_RP_UP_BUSY_AVG |
3525                            GEN6_RP_DOWN_IDLE_AVG);
3526                 break;
3527         }
3528
3529         dev_priv->rps.power = new_power;
3530         dev_priv->rps.last_adj = 0;
3531 }
3532
3533 void gen6_set_rps(struct drm_device *dev, u8 val)
3534 {
3535         struct drm_i915_private *dev_priv = dev->dev_private;
3536         u32 limits = gen6_rps_limits(dev_priv, &val);
3537
3538         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3539         WARN_ON(val > dev_priv->rps.max_delay);
3540         WARN_ON(val < dev_priv->rps.min_delay);
3541
3542         if (val == dev_priv->rps.cur_delay)
3543                 return;
3544
3545         gen6_set_rps_thresholds(dev_priv, val);
3546
3547         if (IS_HASWELL(dev))
3548                 I915_WRITE(GEN6_RPNSWREQ,
3549                            HSW_FREQUENCY(val));
3550         else
3551                 I915_WRITE(GEN6_RPNSWREQ,
3552                            GEN6_FREQUENCY(val) |
3553                            GEN6_OFFSET(0) |
3554                            GEN6_AGGRESSIVE_TURBO);
3555
3556         /* Make sure we continue to get interrupts
3557          * until we hit the minimum or maximum frequencies.
3558          */
3559         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, limits);
3560
3561         POSTING_READ(GEN6_RPNSWREQ);
3562
3563         dev_priv->rps.cur_delay = val;
3564
3565         trace_intel_gpu_freq_change(val * 50);
3566 }
3567
3568 void gen6_rps_idle(struct drm_i915_private *dev_priv)
3569 {
3570         mutex_lock(&dev_priv->rps.hw_lock);
3571         if (dev_priv->rps.enabled) {
3572                 if (dev_priv->info->is_valleyview)
3573                         valleyview_set_rps(dev_priv->dev, dev_priv->rps.min_delay);
3574                 else
3575                         gen6_set_rps(dev_priv->dev, dev_priv->rps.min_delay);
3576                 dev_priv->rps.last_adj = 0;
3577         }
3578         mutex_unlock(&dev_priv->rps.hw_lock);
3579 }
3580
3581 void gen6_rps_boost(struct drm_i915_private *dev_priv)
3582 {
3583         mutex_lock(&dev_priv->rps.hw_lock);
3584         if (dev_priv->rps.enabled) {
3585                 if (dev_priv->info->is_valleyview)
3586                         valleyview_set_rps(dev_priv->dev, dev_priv->rps.max_delay);
3587                 else
3588                         gen6_set_rps(dev_priv->dev, dev_priv->rps.max_delay);
3589                 dev_priv->rps.last_adj = 0;
3590         }
3591         mutex_unlock(&dev_priv->rps.hw_lock);
3592 }
3593
3594 /*
3595  * Wait until the previous freq change has completed,
3596  * or the timeout elapsed, and then update our notion
3597  * of the current GPU frequency.
3598  */
3599 static void vlv_update_rps_cur_delay(struct drm_i915_private *dev_priv)
3600 {
3601         u32 pval;
3602
3603         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3604
3605         if (wait_for(((pval = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS)) & GENFREQSTATUS) == 0, 10))
3606                 DRM_DEBUG_DRIVER("timed out waiting for Punit\n");
3607
3608         pval >>= 8;
3609
3610         if (pval != dev_priv->rps.cur_delay)
3611                 DRM_DEBUG_DRIVER("Punit overrode GPU freq: %d MHz (%u) requested, but got %d Mhz (%u)\n",
3612                                  vlv_gpu_freq(dev_priv->mem_freq, dev_priv->rps.cur_delay),
3613                                  dev_priv->rps.cur_delay,
3614                                  vlv_gpu_freq(dev_priv->mem_freq, pval), pval);
3615
3616         dev_priv->rps.cur_delay = pval;
3617 }
3618
3619 void valleyview_set_rps(struct drm_device *dev, u8 val)
3620 {
3621         struct drm_i915_private *dev_priv = dev->dev_private;
3622
3623         gen6_rps_limits(dev_priv, &val);
3624
3625         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3626         WARN_ON(val > dev_priv->rps.max_delay);
3627         WARN_ON(val < dev_priv->rps.min_delay);
3628
3629         vlv_update_rps_cur_delay(dev_priv);
3630
3631         DRM_DEBUG_DRIVER("GPU freq request from %d MHz (%u) to %d MHz (%u)\n",
3632                          vlv_gpu_freq(dev_priv->mem_freq,
3633                                       dev_priv->rps.cur_delay),
3634                          dev_priv->rps.cur_delay,
3635                          vlv_gpu_freq(dev_priv->mem_freq, val), val);
3636
3637         if (val == dev_priv->rps.cur_delay)
3638                 return;
3639
3640         vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
3641
3642         dev_priv->rps.cur_delay = val;
3643
3644         trace_intel_gpu_freq_change(vlv_gpu_freq(dev_priv->mem_freq, val));
3645 }
3646
3647 static void gen6_disable_rps_interrupts(struct drm_device *dev)
3648 {
3649         struct drm_i915_private *dev_priv = dev->dev_private;
3650
3651         I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
3652         I915_WRITE(GEN6_PMIER, I915_READ(GEN6_PMIER) & ~GEN6_PM_RPS_EVENTS);
3653         /* Complete PM interrupt masking here doesn't race with the rps work
3654          * item again unmasking PM interrupts because that is using a different
3655          * register (PMIMR) to mask PM interrupts. The only risk is in leaving
3656          * stale bits in PMIIR and PMIMR which gen6_enable_rps will clean up. */
3657
3658         spin_lock_irq(&dev_priv->irq_lock);
3659         dev_priv->rps.pm_iir = 0;
3660         spin_unlock_irq(&dev_priv->irq_lock);
3661
3662         I915_WRITE(GEN6_PMIIR, GEN6_PM_RPS_EVENTS);
3663 }
3664
3665 static void gen6_disable_rps(struct drm_device *dev)
3666 {
3667         struct drm_i915_private *dev_priv = dev->dev_private;
3668
3669         I915_WRITE(GEN6_RC_CONTROL, 0);
3670         I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
3671
3672         gen6_disable_rps_interrupts(dev);
3673 }
3674
3675 static void valleyview_disable_rps(struct drm_device *dev)
3676 {
3677         struct drm_i915_private *dev_priv = dev->dev_private;
3678
3679         I915_WRITE(GEN6_RC_CONTROL, 0);
3680
3681         gen6_disable_rps_interrupts(dev);
3682
3683         if (dev_priv->vlv_pctx) {
3684                 drm_gem_object_unreference(&dev_priv->vlv_pctx->base);
3685                 dev_priv->vlv_pctx = NULL;
3686         }
3687 }
3688
3689 static void intel_print_rc6_info(struct drm_device *dev, u32 mode)
3690 {
3691         if (IS_GEN6(dev))
3692                 DRM_DEBUG_DRIVER("Sandybridge: deep RC6 disabled\n");
3693
3694         if (IS_HASWELL(dev))
3695                 DRM_DEBUG_DRIVER("Haswell: only RC6 available\n");
3696
3697         DRM_INFO("Enabling RC6 states: RC6 %s, RC6p %s, RC6pp %s\n",
3698                         (mode & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off",
3699                         (mode & GEN6_RC_CTL_RC6p_ENABLE) ? "on" : "off",
3700                         (mode & GEN6_RC_CTL_RC6pp_ENABLE) ? "on" : "off");
3701 }
3702
3703 int intel_enable_rc6(const struct drm_device *dev)
3704 {
3705         /* No RC6 before Ironlake */
3706         if (INTEL_INFO(dev)->gen < 5)
3707                 return 0;
3708
3709         /* Respect the kernel parameter if it is set */
3710         if (i915_enable_rc6 >= 0)
3711                 return i915_enable_rc6;
3712
3713         /* Disable RC6 on Ironlake */
3714         if (INTEL_INFO(dev)->gen == 5)
3715                 return 0;
3716
3717         if (IS_HASWELL(dev))
3718                 return INTEL_RC6_ENABLE;
3719
3720         /* snb/ivb have more than one rc6 state. */
3721         if (INTEL_INFO(dev)->gen == 6)
3722                 return INTEL_RC6_ENABLE;
3723
3724         return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
3725 }
3726
3727 static void gen6_enable_rps_interrupts(struct drm_device *dev)
3728 {
3729         struct drm_i915_private *dev_priv = dev->dev_private;
3730         u32 enabled_intrs;
3731
3732         spin_lock_irq(&dev_priv->irq_lock);
3733         WARN_ON(dev_priv->rps.pm_iir);
3734         snb_enable_pm_irq(dev_priv, GEN6_PM_RPS_EVENTS);
3735         I915_WRITE(GEN6_PMIIR, GEN6_PM_RPS_EVENTS);
3736         spin_unlock_irq(&dev_priv->irq_lock);
3737
3738         /* only unmask PM interrupts we need. Mask all others. */
3739         enabled_intrs = GEN6_PM_RPS_EVENTS;
3740
3741         /* IVB and SNB hard hangs on looping batchbuffer
3742          * if GEN6_PM_UP_EI_EXPIRED is masked.
3743          */
3744         if (INTEL_INFO(dev)->gen <= 7 && !IS_HASWELL(dev))
3745                 enabled_intrs |= GEN6_PM_RP_UP_EI_EXPIRED;
3746
3747         I915_WRITE(GEN6_PMINTRMSK, ~enabled_intrs);
3748 }
3749
3750 static void gen6_enable_rps(struct drm_device *dev)
3751 {
3752         struct drm_i915_private *dev_priv = dev->dev_private;
3753         struct intel_ring_buffer *ring;
3754         u32 rp_state_cap;
3755         u32 gt_perf_status;
3756         u32 rc6vids, pcu_mbox, rc6_mask = 0;
3757         u32 gtfifodbg;
3758         int rc6_mode;
3759         int i, ret;
3760
3761         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3762
3763         /* Here begins a magic sequence of register writes to enable
3764          * auto-downclocking.
3765          *
3766          * Perhaps there might be some value in exposing these to
3767          * userspace...
3768          */
3769         I915_WRITE(GEN6_RC_STATE, 0);
3770
3771         /* Clear the DBG now so we don't confuse earlier errors */
3772         if ((gtfifodbg = I915_READ(GTFIFODBG))) {
3773                 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
3774                 I915_WRITE(GTFIFODBG, gtfifodbg);
3775         }
3776
3777         gen6_gt_force_wake_get(dev_priv);
3778
3779         rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
3780         gt_perf_status = I915_READ(GEN6_GT_PERF_STATUS);
3781
3782         /* In units of 50MHz */
3783         dev_priv->rps.hw_max = dev_priv->rps.max_delay = rp_state_cap & 0xff;
3784         dev_priv->rps.min_delay = (rp_state_cap >> 16) & 0xff;
3785         dev_priv->rps.rp1_delay = (rp_state_cap >>  8) & 0xff;
3786         dev_priv->rps.rp0_delay = (rp_state_cap >>  0) & 0xff;
3787         dev_priv->rps.rpe_delay = dev_priv->rps.rp1_delay;
3788         dev_priv->rps.cur_delay = 0;
3789
3790         /* disable the counters and set deterministic thresholds */
3791         I915_WRITE(GEN6_RC_CONTROL, 0);
3792
3793         I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
3794         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
3795         I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
3796         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
3797         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
3798
3799         for_each_ring(ring, dev_priv, i)
3800                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
3801
3802         I915_WRITE(GEN6_RC_SLEEP, 0);
3803         I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
3804         if (INTEL_INFO(dev)->gen <= 6 || IS_IVYBRIDGE(dev))
3805                 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
3806         else
3807                 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
3808         I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
3809         I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
3810
3811         /* Check if we are enabling RC6 */
3812         rc6_mode = intel_enable_rc6(dev_priv->dev);
3813         if (rc6_mode & INTEL_RC6_ENABLE)
3814                 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
3815
3816         /* We don't use those on Haswell */
3817         if (!IS_HASWELL(dev)) {
3818                 if (rc6_mode & INTEL_RC6p_ENABLE)
3819                         rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
3820
3821                 if (rc6_mode & INTEL_RC6pp_ENABLE)
3822                         rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
3823         }
3824
3825         intel_print_rc6_info(dev, rc6_mask);
3826
3827         I915_WRITE(GEN6_RC_CONTROL,
3828                    rc6_mask |
3829                    GEN6_RC_CTL_EI_MODE(1) |
3830                    GEN6_RC_CTL_HW_ENABLE);
3831
3832         /* Power down if completely idle for over 50ms */
3833         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
3834         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
3835
3836         ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_MIN_FREQ_TABLE, 0);
3837         if (!ret) {
3838                 pcu_mbox = 0;
3839                 ret = sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &pcu_mbox);
3840                 if (!ret && (pcu_mbox & (1<<31))) { /* OC supported */
3841                         DRM_DEBUG_DRIVER("Overclocking supported. Max: %dMHz, Overclock max: %dMHz\n",
3842                                          (dev_priv->rps.max_delay & 0xff) * 50,
3843                                          (pcu_mbox & 0xff) * 50);
3844                         dev_priv->rps.hw_max = pcu_mbox & 0xff;
3845                 }
3846         } else {
3847                 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
3848         }
3849
3850         dev_priv->rps.power = HIGH_POWER; /* force a reset */
3851         gen6_set_rps(dev_priv->dev, dev_priv->rps.min_delay);
3852
3853         gen6_enable_rps_interrupts(dev);
3854
3855         rc6vids = 0;
3856         ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
3857         if (IS_GEN6(dev) && ret) {
3858                 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
3859         } else if (IS_GEN6(dev) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
3860                 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
3861                           GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
3862                 rc6vids &= 0xffff00;
3863                 rc6vids |= GEN6_ENCODE_RC6_VID(450);
3864                 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
3865                 if (ret)
3866                         DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
3867         }
3868
3869         gen6_gt_force_wake_put(dev_priv);
3870 }
3871
3872 void gen6_update_ring_freq(struct drm_device *dev)
3873 {
3874         struct drm_i915_private *dev_priv = dev->dev_private;
3875         int min_freq = 15;
3876         unsigned int gpu_freq;
3877         unsigned int max_ia_freq, min_ring_freq;
3878         int scaling_factor = 180;
3879         struct cpufreq_policy *policy;
3880
3881         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3882
3883         policy = cpufreq_cpu_get(0);
3884         if (policy) {
3885                 max_ia_freq = policy->cpuinfo.max_freq;
3886                 cpufreq_cpu_put(policy);
3887         } else {
3888                 /*
3889                  * Default to measured freq if none found, PCU will ensure we
3890                  * don't go over
3891                  */
3892                 max_ia_freq = tsc_khz;
3893         }
3894
3895         /* Convert from kHz to MHz */
3896         max_ia_freq /= 1000;
3897
3898         min_ring_freq = I915_READ(DCLK) & 0xf;
3899         /* convert DDR frequency from units of 266.6MHz to bandwidth */
3900         min_ring_freq = mult_frac(min_ring_freq, 8, 3);
3901
3902         /*
3903          * For each potential GPU frequency, load a ring frequency we'd like
3904          * to use for memory access.  We do this by specifying the IA frequency
3905          * the PCU should use as a reference to determine the ring frequency.
3906          */
3907         for (gpu_freq = dev_priv->rps.max_delay; gpu_freq >= dev_priv->rps.min_delay;
3908              gpu_freq--) {
3909                 int diff = dev_priv->rps.max_delay - gpu_freq;
3910                 unsigned int ia_freq = 0, ring_freq = 0;
3911
3912                 if (IS_HASWELL(dev)) {
3913                         ring_freq = mult_frac(gpu_freq, 5, 4);
3914                         ring_freq = max(min_ring_freq, ring_freq);
3915                         /* leave ia_freq as the default, chosen by cpufreq */
3916                 } else {
3917                         /* On older processors, there is no separate ring
3918                          * clock domain, so in order to boost the bandwidth
3919                          * of the ring, we need to upclock the CPU (ia_freq).
3920                          *
3921                          * For GPU frequencies less than 750MHz,
3922                          * just use the lowest ring freq.
3923                          */
3924                         if (gpu_freq < min_freq)
3925                                 ia_freq = 800;
3926                         else
3927                                 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
3928                         ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
3929                 }
3930
3931                 sandybridge_pcode_write(dev_priv,
3932                                         GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
3933                                         ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
3934                                         ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
3935                                         gpu_freq);
3936         }
3937 }
3938
3939 int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
3940 {
3941         u32 val, rp0;
3942
3943         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
3944
3945         rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
3946         /* Clamp to max */
3947         rp0 = min_t(u32, rp0, 0xea);
3948
3949         return rp0;
3950 }
3951
3952 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
3953 {
3954         u32 val, rpe;
3955
3956         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
3957         rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
3958         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
3959         rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
3960
3961         return rpe;
3962 }
3963
3964 int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
3965 {
3966         return vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
3967 }
3968
3969 static void valleyview_setup_pctx(struct drm_device *dev)
3970 {
3971         struct drm_i915_private *dev_priv = dev->dev_private;
3972         struct drm_i915_gem_object *pctx;
3973         unsigned long pctx_paddr;
3974         u32 pcbr;
3975         int pctx_size = 24*1024;
3976
3977         pcbr = I915_READ(VLV_PCBR);
3978         if (pcbr) {
3979                 /* BIOS set it up already, grab the pre-alloc'd space */
3980                 int pcbr_offset;
3981
3982                 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
3983                 pctx = i915_gem_object_create_stolen_for_preallocated(dev_priv->dev,
3984                                                                       pcbr_offset,
3985                                                                       I915_GTT_OFFSET_NONE,
3986                                                                       pctx_size);
3987                 goto out;
3988         }
3989
3990         /*
3991          * From the Gunit register HAS:
3992          * The Gfx driver is expected to program this register and ensure
3993          * proper allocation within Gfx stolen memory.  For example, this
3994          * register should be programmed such than the PCBR range does not
3995          * overlap with other ranges, such as the frame buffer, protected
3996          * memory, or any other relevant ranges.
3997          */
3998         pctx = i915_gem_object_create_stolen(dev, pctx_size);
3999         if (!pctx) {
4000                 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
4001                 return;
4002         }
4003
4004         pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
4005         I915_WRITE(VLV_PCBR, pctx_paddr);
4006
4007 out:
4008         dev_priv->vlv_pctx = pctx;
4009 }
4010
4011 static void valleyview_enable_rps(struct drm_device *dev)
4012 {
4013         struct drm_i915_private *dev_priv = dev->dev_private;
4014         struct intel_ring_buffer *ring;
4015         u32 gtfifodbg, val, rc6_mode = 0;
4016         int i;
4017
4018         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4019
4020         if ((gtfifodbg = I915_READ(GTFIFODBG))) {
4021                 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
4022                                  gtfifodbg);
4023                 I915_WRITE(GTFIFODBG, gtfifodbg);
4024         }
4025
4026         valleyview_setup_pctx(dev);
4027
4028         gen6_gt_force_wake_get(dev_priv);
4029
4030         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
4031         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
4032         I915_WRITE(GEN6_RP_UP_EI, 66000);
4033         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
4034
4035         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
4036
4037         I915_WRITE(GEN6_RP_CONTROL,
4038                    GEN6_RP_MEDIA_TURBO |
4039                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
4040                    GEN6_RP_MEDIA_IS_GFX |
4041                    GEN6_RP_ENABLE |
4042                    GEN6_RP_UP_BUSY_AVG |
4043                    GEN6_RP_DOWN_IDLE_CONT);
4044
4045         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
4046         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
4047         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
4048
4049         for_each_ring(ring, dev_priv, i)
4050                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4051
4052         I915_WRITE(GEN6_RC6_THRESHOLD, 0xc350);
4053
4054         /* allows RC6 residency counter to work */
4055         I915_WRITE(VLV_COUNTER_CONTROL,
4056                    _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
4057                                       VLV_MEDIA_RC6_COUNT_EN |
4058                                       VLV_RENDER_RC6_COUNT_EN));
4059         if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
4060                 rc6_mode = GEN7_RC_CTL_TO_MODE;
4061
4062         intel_print_rc6_info(dev, rc6_mode);
4063
4064         I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
4065
4066         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
4067
4068         DRM_DEBUG_DRIVER("GPLL enabled? %s\n", val & 0x10 ? "yes" : "no");
4069         DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
4070
4071         dev_priv->rps.cur_delay = (val >> 8) & 0xff;
4072         DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
4073                          vlv_gpu_freq(dev_priv->mem_freq,
4074                                       dev_priv->rps.cur_delay),
4075                          dev_priv->rps.cur_delay);
4076
4077         dev_priv->rps.max_delay = valleyview_rps_max_freq(dev_priv);
4078         dev_priv->rps.hw_max = dev_priv->rps.max_delay;
4079         DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
4080                          vlv_gpu_freq(dev_priv->mem_freq,
4081                                       dev_priv->rps.max_delay),
4082                          dev_priv->rps.max_delay);
4083
4084         dev_priv->rps.rpe_delay = valleyview_rps_rpe_freq(dev_priv);
4085         DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
4086                          vlv_gpu_freq(dev_priv->mem_freq,
4087                                       dev_priv->rps.rpe_delay),
4088                          dev_priv->rps.rpe_delay);
4089
4090         dev_priv->rps.min_delay = valleyview_rps_min_freq(dev_priv);
4091         DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
4092                          vlv_gpu_freq(dev_priv->mem_freq,
4093                                       dev_priv->rps.min_delay),
4094                          dev_priv->rps.min_delay);
4095
4096         DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
4097                          vlv_gpu_freq(dev_priv->mem_freq,
4098                                       dev_priv->rps.rpe_delay),
4099                          dev_priv->rps.rpe_delay);
4100
4101         valleyview_set_rps(dev_priv->dev, dev_priv->rps.rpe_delay);
4102
4103         gen6_enable_rps_interrupts(dev);
4104
4105         gen6_gt_force_wake_put(dev_priv);
4106 }
4107
4108 void ironlake_teardown_rc6(struct drm_device *dev)
4109 {
4110         struct drm_i915_private *dev_priv = dev->dev_private;
4111
4112         if (dev_priv->ips.renderctx) {
4113                 i915_gem_object_unpin(dev_priv->ips.renderctx);
4114                 drm_gem_object_unreference(&dev_priv->ips.renderctx->base);
4115                 dev_priv->ips.renderctx = NULL;
4116         }
4117
4118         if (dev_priv->ips.pwrctx) {
4119                 i915_gem_object_unpin(dev_priv->ips.pwrctx);
4120                 drm_gem_object_unreference(&dev_priv->ips.pwrctx->base);
4121                 dev_priv->ips.pwrctx = NULL;
4122         }
4123 }
4124
4125 static void ironlake_disable_rc6(struct drm_device *dev)
4126 {
4127         struct drm_i915_private *dev_priv = dev->dev_private;
4128
4129         if (I915_READ(PWRCTXA)) {
4130                 /* Wake the GPU, prevent RC6, then restore RSTDBYCTL */
4131                 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) | RCX_SW_EXIT);
4132                 wait_for(((I915_READ(RSTDBYCTL) & RSX_STATUS_MASK) == RSX_STATUS_ON),
4133                          50);
4134
4135                 I915_WRITE(PWRCTXA, 0);
4136                 POSTING_READ(PWRCTXA);
4137
4138                 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
4139                 POSTING_READ(RSTDBYCTL);
4140         }
4141 }
4142
4143 static int ironlake_setup_rc6(struct drm_device *dev)
4144 {
4145         struct drm_i915_private *dev_priv = dev->dev_private;
4146
4147         if (dev_priv->ips.renderctx == NULL)
4148                 dev_priv->ips.renderctx = intel_alloc_context_page(dev);
4149         if (!dev_priv->ips.renderctx)
4150                 return -ENOMEM;
4151
4152         if (dev_priv->ips.pwrctx == NULL)
4153                 dev_priv->ips.pwrctx = intel_alloc_context_page(dev);
4154         if (!dev_priv->ips.pwrctx) {
4155                 ironlake_teardown_rc6(dev);
4156                 return -ENOMEM;
4157         }
4158
4159         return 0;
4160 }
4161
4162 static void ironlake_enable_rc6(struct drm_device *dev)
4163 {
4164         struct drm_i915_private *dev_priv = dev->dev_private;
4165         struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
4166         bool was_interruptible;
4167         int ret;
4168
4169         /* rc6 disabled by default due to repeated reports of hanging during
4170          * boot and resume.
4171          */
4172         if (!intel_enable_rc6(dev))
4173                 return;
4174
4175         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
4176
4177         ret = ironlake_setup_rc6(dev);
4178         if (ret)
4179                 return;
4180
4181         was_interruptible = dev_priv->mm.interruptible;
4182         dev_priv->mm.interruptible = false;
4183
4184         /*
4185          * GPU can automatically power down the render unit if given a page
4186          * to save state.
4187          */
4188         ret = intel_ring_begin(ring, 6);
4189         if (ret) {
4190                 ironlake_teardown_rc6(dev);
4191                 dev_priv->mm.interruptible = was_interruptible;
4192                 return;
4193         }
4194
4195         intel_ring_emit(ring, MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN);
4196         intel_ring_emit(ring, MI_SET_CONTEXT);
4197         intel_ring_emit(ring, i915_gem_obj_ggtt_offset(dev_priv->ips.renderctx) |
4198                         MI_MM_SPACE_GTT |
4199                         MI_SAVE_EXT_STATE_EN |
4200                         MI_RESTORE_EXT_STATE_EN |
4201                         MI_RESTORE_INHIBIT);
4202         intel_ring_emit(ring, MI_SUSPEND_FLUSH);
4203         intel_ring_emit(ring, MI_NOOP);
4204         intel_ring_emit(ring, MI_FLUSH);
4205         intel_ring_advance(ring);
4206
4207         /*
4208          * Wait for the command parser to advance past MI_SET_CONTEXT. The HW
4209          * does an implicit flush, combined with MI_FLUSH above, it should be
4210          * safe to assume that renderctx is valid
4211          */
4212         ret = intel_ring_idle(ring);
4213         dev_priv->mm.interruptible = was_interruptible;
4214         if (ret) {
4215                 DRM_ERROR("failed to enable ironlake power savings\n");
4216                 ironlake_teardown_rc6(dev);
4217                 return;
4218         }
4219
4220         I915_WRITE(PWRCTXA, i915_gem_obj_ggtt_offset(dev_priv->ips.pwrctx) | PWRCTX_EN);
4221         I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
4222
4223         intel_print_rc6_info(dev, INTEL_RC6_ENABLE);
4224 }
4225
4226 static unsigned long intel_pxfreq(u32 vidfreq)
4227 {
4228         unsigned long freq;
4229         int div = (vidfreq & 0x3f0000) >> 16;
4230         int post = (vidfreq & 0x3000) >> 12;
4231         int pre = (vidfreq & 0x7);
4232
4233         if (!pre)
4234                 return 0;
4235
4236         freq = ((div * 133333) / ((1<<post) * pre));
4237
4238         return freq;
4239 }
4240
4241 static const struct cparams {
4242         u16 i;
4243         u16 t;
4244         u16 m;
4245         u16 c;
4246 } cparams[] = {
4247         { 1, 1333, 301, 28664 },
4248         { 1, 1066, 294, 24460 },
4249         { 1, 800, 294, 25192 },
4250         { 0, 1333, 276, 27605 },
4251         { 0, 1066, 276, 27605 },
4252         { 0, 800, 231, 23784 },
4253 };
4254
4255 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
4256 {
4257         u64 total_count, diff, ret;
4258         u32 count1, count2, count3, m = 0, c = 0;
4259         unsigned long now = jiffies_to_msecs(jiffies), diff1;
4260         int i;
4261
4262         assert_spin_locked(&mchdev_lock);
4263
4264         diff1 = now - dev_priv->ips.last_time1;
4265
4266         /* Prevent division-by-zero if we are asking too fast.
4267          * Also, we don't get interesting results if we are polling
4268          * faster than once in 10ms, so just return the saved value
4269          * in such cases.
4270          */
4271         if (diff1 <= 10)
4272                 return dev_priv->ips.chipset_power;
4273
4274         count1 = I915_READ(DMIEC);
4275         count2 = I915_READ(DDREC);
4276         count3 = I915_READ(CSIEC);
4277
4278         total_count = count1 + count2 + count3;
4279
4280         /* FIXME: handle per-counter overflow */
4281         if (total_count < dev_priv->ips.last_count1) {
4282                 diff = ~0UL - dev_priv->ips.last_count1;
4283                 diff += total_count;
4284         } else {
4285                 diff = total_count - dev_priv->ips.last_count1;
4286         }
4287
4288         for (i = 0; i < ARRAY_SIZE(cparams); i++) {
4289                 if (cparams[i].i == dev_priv->ips.c_m &&
4290                     cparams[i].t == dev_priv->ips.r_t) {
4291                         m = cparams[i].m;
4292                         c = cparams[i].c;
4293                         break;
4294                 }
4295         }
4296
4297         diff = div_u64(diff, diff1);
4298         ret = ((m * diff) + c);
4299         ret = div_u64(ret, 10);
4300
4301         dev_priv->ips.last_count1 = total_count;
4302         dev_priv->ips.last_time1 = now;
4303
4304         dev_priv->ips.chipset_power = ret;
4305
4306         return ret;
4307 }
4308
4309 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
4310 {
4311         unsigned long val;
4312
4313         if (dev_priv->info->gen != 5)
4314                 return 0;
4315
4316         spin_lock_irq(&mchdev_lock);
4317
4318         val = __i915_chipset_val(dev_priv);
4319
4320         spin_unlock_irq(&mchdev_lock);
4321
4322         return val;
4323 }
4324
4325 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
4326 {
4327         unsigned long m, x, b;
4328         u32 tsfs;
4329
4330         tsfs = I915_READ(TSFS);
4331
4332         m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
4333         x = I915_READ8(TR1);
4334
4335         b = tsfs & TSFS_INTR_MASK;
4336
4337         return ((m * x) / 127) - b;
4338 }
4339
4340 static u16 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
4341 {
4342         static const struct v_table {
4343                 u16 vd; /* in .1 mil */
4344                 u16 vm; /* in .1 mil */
4345         } v_table[] = {
4346                 { 0, 0, },
4347                 { 375, 0, },
4348                 { 500, 0, },
4349                 { 625, 0, },
4350                 { 750, 0, },
4351                 { 875, 0, },
4352                 { 1000, 0, },
4353                 { 1125, 0, },
4354                 { 4125, 3000, },
4355                 { 4125, 3000, },
4356                 { 4125, 3000, },
4357                 { 4125, 3000, },
4358                 { 4125, 3000, },
4359                 { 4125, 3000, },
4360                 { 4125, 3000, },
4361                 { 4125, 3000, },
4362                 { 4125, 3000, },
4363                 { 4125, 3000, },
4364                 { 4125, 3000, },
4365                 { 4125, 3000, },
4366                 { 4125, 3000, },
4367                 { 4125, 3000, },
4368                 { 4125, 3000, },
4369                 { 4125, 3000, },
4370                 { 4125, 3000, },
4371                 { 4125, 3000, },
4372                 { 4125, 3000, },
4373                 { 4125, 3000, },
4374                 { 4125, 3000, },
4375                 { 4125, 3000, },
4376                 { 4125, 3000, },
4377                 { 4125, 3000, },
4378                 { 4250, 3125, },
4379                 { 4375, 3250, },
4380                 { 4500, 3375, },
4381                 { 4625, 3500, },
4382                 { 4750, 3625, },
4383                 { 4875, 3750, },
4384                 { 5000, 3875, },
4385                 { 5125, 4000, },
4386                 { 5250, 4125, },
4387                 { 5375, 4250, },
4388                 { 5500, 4375, },
4389                 { 5625, 4500, },
4390                 { 5750, 4625, },
4391                 { 5875, 4750, },
4392                 { 6000, 4875, },
4393                 { 6125, 5000, },
4394                 { 6250, 5125, },
4395                 { 6375, 5250, },
4396                 { 6500, 5375, },
4397                 { 6625, 5500, },
4398                 { 6750, 5625, },
4399                 { 6875, 5750, },
4400                 { 7000, 5875, },
4401                 { 7125, 6000, },
4402                 { 7250, 6125, },
4403                 { 7375, 6250, },
4404                 { 7500, 6375, },
4405                 { 7625, 6500, },
4406                 { 7750, 6625, },
4407                 { 7875, 6750, },
4408                 { 8000, 6875, },
4409                 { 8125, 7000, },
4410                 { 8250, 7125, },
4411                 { 8375, 7250, },
4412                 { 8500, 7375, },
4413                 { 8625, 7500, },
4414                 { 8750, 7625, },
4415                 { 8875, 7750, },
4416                 { 9000, 7875, },
4417                 { 9125, 8000, },
4418                 { 9250, 8125, },
4419                 { 9375, 8250, },
4420                 { 9500, 8375, },
4421                 { 9625, 8500, },
4422                 { 9750, 8625, },
4423                 { 9875, 8750, },
4424                 { 10000, 8875, },
4425                 { 10125, 9000, },
4426                 { 10250, 9125, },
4427                 { 10375, 9250, },
4428                 { 10500, 9375, },
4429                 { 10625, 9500, },
4430                 { 10750, 9625, },
4431                 { 10875, 9750, },
4432                 { 11000, 9875, },
4433                 { 11125, 10000, },
4434                 { 11250, 10125, },
4435                 { 11375, 10250, },
4436                 { 11500, 10375, },
4437                 { 11625, 10500, },
4438                 { 11750, 10625, },
4439                 { 11875, 10750, },
4440                 { 12000, 10875, },
4441                 { 12125, 11000, },
4442                 { 12250, 11125, },
4443                 { 12375, 11250, },
4444                 { 12500, 11375, },
4445                 { 12625, 11500, },
4446                 { 12750, 11625, },
4447                 { 12875, 11750, },
4448                 { 13000, 11875, },
4449                 { 13125, 12000, },
4450                 { 13250, 12125, },
4451                 { 13375, 12250, },
4452                 { 13500, 12375, },
4453                 { 13625, 12500, },
4454                 { 13750, 12625, },
4455                 { 13875, 12750, },
4456                 { 14000, 12875, },
4457                 { 14125, 13000, },
4458                 { 14250, 13125, },
4459                 { 14375, 13250, },
4460                 { 14500, 13375, },
4461                 { 14625, 13500, },
4462                 { 14750, 13625, },
4463                 { 14875, 13750, },
4464                 { 15000, 13875, },
4465                 { 15125, 14000, },
4466                 { 15250, 14125, },
4467                 { 15375, 14250, },
4468                 { 15500, 14375, },
4469                 { 15625, 14500, },
4470                 { 15750, 14625, },
4471                 { 15875, 14750, },
4472                 { 16000, 14875, },
4473                 { 16125, 15000, },
4474         };
4475         if (dev_priv->info->is_mobile)
4476                 return v_table[pxvid].vm;
4477         else
4478                 return v_table[pxvid].vd;
4479 }
4480
4481 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
4482 {
4483         struct timespec now, diff1;
4484         u64 diff;
4485         unsigned long diffms;
4486         u32 count;
4487
4488         assert_spin_locked(&mchdev_lock);
4489
4490         getrawmonotonic(&now);
4491         diff1 = timespec_sub(now, dev_priv->ips.last_time2);
4492
4493         /* Don't divide by 0 */
4494         diffms = diff1.tv_sec * 1000 + diff1.tv_nsec / 1000000;
4495         if (!diffms)
4496                 return;
4497
4498         count = I915_READ(GFXEC);
4499
4500         if (count < dev_priv->ips.last_count2) {
4501                 diff = ~0UL - dev_priv->ips.last_count2;
4502                 diff += count;
4503         } else {
4504                 diff = count - dev_priv->ips.last_count2;
4505         }
4506
4507         dev_priv->ips.last_count2 = count;
4508         dev_priv->ips.last_time2 = now;
4509
4510         /* More magic constants... */
4511         diff = diff * 1181;
4512         diff = div_u64(diff, diffms * 10);
4513         dev_priv->ips.gfx_power = diff;
4514 }
4515
4516 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
4517 {
4518         if (dev_priv->info->gen != 5)
4519                 return;
4520
4521         spin_lock_irq(&mchdev_lock);
4522
4523         __i915_update_gfx_val(dev_priv);
4524
4525         spin_unlock_irq(&mchdev_lock);
4526 }
4527
4528 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
4529 {
4530         unsigned long t, corr, state1, corr2, state2;
4531         u32 pxvid, ext_v;
4532
4533         assert_spin_locked(&mchdev_lock);
4534
4535         pxvid = I915_READ(PXVFREQ_BASE + (dev_priv->rps.cur_delay * 4));
4536         pxvid = (pxvid >> 24) & 0x7f;
4537         ext_v = pvid_to_extvid(dev_priv, pxvid);
4538
4539         state1 = ext_v;
4540
4541         t = i915_mch_val(dev_priv);
4542
4543         /* Revel in the empirically derived constants */
4544
4545         /* Correction factor in 1/100000 units */
4546         if (t > 80)
4547                 corr = ((t * 2349) + 135940);
4548         else if (t >= 50)
4549                 corr = ((t * 964) + 29317);
4550         else /* < 50 */
4551                 corr = ((t * 301) + 1004);
4552
4553         corr = corr * ((150142 * state1) / 10000 - 78642);
4554         corr /= 100000;
4555         corr2 = (corr * dev_priv->ips.corr);
4556
4557         state2 = (corr2 * state1) / 10000;
4558         state2 /= 100; /* convert to mW */
4559
4560         __i915_update_gfx_val(dev_priv);
4561
4562         return dev_priv->ips.gfx_power + state2;
4563 }
4564
4565 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
4566 {
4567         unsigned long val;
4568
4569         if (dev_priv->info->gen != 5)
4570                 return 0;
4571
4572         spin_lock_irq(&mchdev_lock);
4573
4574         val = __i915_gfx_val(dev_priv);
4575
4576         spin_unlock_irq(&mchdev_lock);
4577
4578         return val;
4579 }
4580
4581 /**
4582  * i915_read_mch_val - return value for IPS use
4583  *
4584  * Calculate and return a value for the IPS driver to use when deciding whether
4585  * we have thermal and power headroom to increase CPU or GPU power budget.
4586  */
4587 unsigned long i915_read_mch_val(void)
4588 {
4589         struct drm_i915_private *dev_priv;
4590         unsigned long chipset_val, graphics_val, ret = 0;
4591
4592         spin_lock_irq(&mchdev_lock);
4593         if (!i915_mch_dev)
4594                 goto out_unlock;
4595         dev_priv = i915_mch_dev;
4596
4597         chipset_val = __i915_chipset_val(dev_priv);
4598         graphics_val = __i915_gfx_val(dev_priv);
4599
4600         ret = chipset_val + graphics_val;
4601
4602 out_unlock:
4603         spin_unlock_irq(&mchdev_lock);
4604
4605         return ret;
4606 }
4607 EXPORT_SYMBOL_GPL(i915_read_mch_val);
4608
4609 /**
4610  * i915_gpu_raise - raise GPU frequency limit
4611  *
4612  * Raise the limit; IPS indicates we have thermal headroom.
4613  */
4614 bool i915_gpu_raise(void)
4615 {
4616         struct drm_i915_private *dev_priv;
4617         bool ret = true;
4618
4619         spin_lock_irq(&mchdev_lock);
4620         if (!i915_mch_dev) {
4621                 ret = false;
4622                 goto out_unlock;
4623         }
4624         dev_priv = i915_mch_dev;
4625
4626         if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
4627                 dev_priv->ips.max_delay--;
4628
4629 out_unlock:
4630         spin_unlock_irq(&mchdev_lock);
4631
4632         return ret;
4633 }
4634 EXPORT_SYMBOL_GPL(i915_gpu_raise);
4635
4636 /**
4637  * i915_gpu_lower - lower GPU frequency limit
4638  *
4639  * IPS indicates we're close to a thermal limit, so throttle back the GPU
4640  * frequency maximum.
4641  */
4642 bool i915_gpu_lower(void)
4643 {
4644         struct drm_i915_private *dev_priv;
4645         bool ret = true;
4646
4647         spin_lock_irq(&mchdev_lock);
4648         if (!i915_mch_dev) {
4649                 ret = false;
4650                 goto out_unlock;
4651         }
4652         dev_priv = i915_mch_dev;
4653
4654         if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
4655                 dev_priv->ips.max_delay++;
4656
4657 out_unlock:
4658         spin_unlock_irq(&mchdev_lock);
4659
4660         return ret;
4661 }
4662 EXPORT_SYMBOL_GPL(i915_gpu_lower);
4663
4664 /**
4665  * i915_gpu_busy - indicate GPU business to IPS
4666  *
4667  * Tell the IPS driver whether or not the GPU is busy.
4668  */
4669 bool i915_gpu_busy(void)
4670 {
4671         struct drm_i915_private *dev_priv;
4672         struct intel_ring_buffer *ring;
4673         bool ret = false;
4674         int i;
4675
4676         spin_lock_irq(&mchdev_lock);
4677         if (!i915_mch_dev)
4678                 goto out_unlock;
4679         dev_priv = i915_mch_dev;
4680
4681         for_each_ring(ring, dev_priv, i)
4682                 ret |= !list_empty(&ring->request_list);
4683
4684 out_unlock:
4685         spin_unlock_irq(&mchdev_lock);
4686
4687         return ret;
4688 }
4689 EXPORT_SYMBOL_GPL(i915_gpu_busy);
4690
4691 /**
4692  * i915_gpu_turbo_disable - disable graphics turbo
4693  *
4694  * Disable graphics turbo by resetting the max frequency and setting the
4695  * current frequency to the default.
4696  */
4697 bool i915_gpu_turbo_disable(void)
4698 {
4699         struct drm_i915_private *dev_priv;
4700         bool ret = true;
4701
4702         spin_lock_irq(&mchdev_lock);
4703         if (!i915_mch_dev) {
4704                 ret = false;
4705                 goto out_unlock;
4706         }
4707         dev_priv = i915_mch_dev;
4708
4709         dev_priv->ips.max_delay = dev_priv->ips.fstart;
4710
4711         if (!ironlake_set_drps(dev_priv->dev, dev_priv->ips.fstart))
4712                 ret = false;
4713
4714 out_unlock:
4715         spin_unlock_irq(&mchdev_lock);
4716
4717         return ret;
4718 }
4719 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
4720
4721 /**
4722  * Tells the intel_ips driver that the i915 driver is now loaded, if
4723  * IPS got loaded first.
4724  *
4725  * This awkward dance is so that neither module has to depend on the
4726  * other in order for IPS to do the appropriate communication of
4727  * GPU turbo limits to i915.
4728  */
4729 static void
4730 ips_ping_for_i915_load(void)
4731 {
4732         void (*link)(void);
4733
4734         link = symbol_get(ips_link_to_i915_driver);
4735         if (link) {
4736                 link();
4737                 symbol_put(ips_link_to_i915_driver);
4738         }
4739 }
4740
4741 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
4742 {
4743         /* We only register the i915 ips part with intel-ips once everything is
4744          * set up, to avoid intel-ips sneaking in and reading bogus values. */
4745         spin_lock_irq(&mchdev_lock);
4746         i915_mch_dev = dev_priv;
4747         spin_unlock_irq(&mchdev_lock);
4748
4749         ips_ping_for_i915_load();
4750 }
4751
4752 void intel_gpu_ips_teardown(void)
4753 {
4754         spin_lock_irq(&mchdev_lock);
4755         i915_mch_dev = NULL;
4756         spin_unlock_irq(&mchdev_lock);
4757 }
4758 static void intel_init_emon(struct drm_device *dev)
4759 {
4760         struct drm_i915_private *dev_priv = dev->dev_private;
4761         u32 lcfuse;
4762         u8 pxw[16];
4763         int i;
4764
4765         /* Disable to program */
4766         I915_WRITE(ECR, 0);
4767         POSTING_READ(ECR);
4768
4769         /* Program energy weights for various events */
4770         I915_WRITE(SDEW, 0x15040d00);
4771         I915_WRITE(CSIEW0, 0x007f0000);
4772         I915_WRITE(CSIEW1, 0x1e220004);
4773         I915_WRITE(CSIEW2, 0x04000004);
4774
4775         for (i = 0; i < 5; i++)
4776                 I915_WRITE(PEW + (i * 4), 0);
4777         for (i = 0; i < 3; i++)
4778                 I915_WRITE(DEW + (i * 4), 0);
4779
4780         /* Program P-state weights to account for frequency power adjustment */
4781         for (i = 0; i < 16; i++) {
4782                 u32 pxvidfreq = I915_READ(PXVFREQ_BASE + (i * 4));
4783                 unsigned long freq = intel_pxfreq(pxvidfreq);
4784                 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
4785                         PXVFREQ_PX_SHIFT;
4786                 unsigned long val;
4787
4788                 val = vid * vid;
4789                 val *= (freq / 1000);
4790                 val *= 255;
4791                 val /= (127*127*900);
4792                 if (val > 0xff)
4793                         DRM_ERROR("bad pxval: %ld\n", val);
4794                 pxw[i] = val;
4795         }
4796         /* Render standby states get 0 weight */
4797         pxw[14] = 0;
4798         pxw[15] = 0;
4799
4800         for (i = 0; i < 4; i++) {
4801                 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
4802                         (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
4803                 I915_WRITE(PXW + (i * 4), val);
4804         }
4805
4806         /* Adjust magic regs to magic values (more experimental results) */
4807         I915_WRITE(OGW0, 0);
4808         I915_WRITE(OGW1, 0);
4809         I915_WRITE(EG0, 0x00007f00);
4810         I915_WRITE(EG1, 0x0000000e);
4811         I915_WRITE(EG2, 0x000e0000);
4812         I915_WRITE(EG3, 0x68000300);
4813         I915_WRITE(EG4, 0x42000000);
4814         I915_WRITE(EG5, 0x00140031);
4815         I915_WRITE(EG6, 0);
4816         I915_WRITE(EG7, 0);
4817
4818         for (i = 0; i < 8; i++)
4819                 I915_WRITE(PXWL + (i * 4), 0);
4820
4821         /* Enable PMON + select events */
4822         I915_WRITE(ECR, 0x80000019);
4823
4824         lcfuse = I915_READ(LCFUSE02);
4825
4826         dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
4827 }
4828
4829 void intel_disable_gt_powersave(struct drm_device *dev)
4830 {
4831         struct drm_i915_private *dev_priv = dev->dev_private;
4832
4833         /* Interrupts should be disabled already to avoid re-arming. */
4834         WARN_ON(dev->irq_enabled);
4835
4836         if (IS_IRONLAKE_M(dev)) {
4837                 ironlake_disable_drps(dev);
4838                 ironlake_disable_rc6(dev);
4839         } else if (INTEL_INFO(dev)->gen >= 6) {
4840                 cancel_delayed_work_sync(&dev_priv->rps.delayed_resume_work);
4841                 cancel_work_sync(&dev_priv->rps.work);
4842                 mutex_lock(&dev_priv->rps.hw_lock);
4843                 if (IS_VALLEYVIEW(dev))
4844                         valleyview_disable_rps(dev);
4845                 else
4846                         gen6_disable_rps(dev);
4847                 dev_priv->rps.enabled = false;
4848                 mutex_unlock(&dev_priv->rps.hw_lock);
4849         }
4850 }
4851
4852 static void intel_gen6_powersave_work(struct work_struct *work)
4853 {
4854         struct drm_i915_private *dev_priv =
4855                 container_of(work, struct drm_i915_private,
4856                              rps.delayed_resume_work.work);
4857         struct drm_device *dev = dev_priv->dev;
4858
4859         mutex_lock(&dev_priv->rps.hw_lock);
4860
4861         if (IS_VALLEYVIEW(dev)) {
4862                 valleyview_enable_rps(dev);
4863         } else {
4864                 gen6_enable_rps(dev);
4865                 gen6_update_ring_freq(dev);
4866         }
4867         dev_priv->rps.enabled = true;
4868         mutex_unlock(&dev_priv->rps.hw_lock);
4869 }
4870
4871 void intel_enable_gt_powersave(struct drm_device *dev)
4872 {
4873         struct drm_i915_private *dev_priv = dev->dev_private;
4874
4875         if (IS_IRONLAKE_M(dev)) {
4876                 ironlake_enable_drps(dev);
4877                 ironlake_enable_rc6(dev);
4878                 intel_init_emon(dev);
4879         } else if (IS_GEN6(dev) || IS_GEN7(dev)) {
4880                 /*
4881                  * PCU communication is slow and this doesn't need to be
4882                  * done at any specific time, so do this out of our fast path
4883                  * to make resume and init faster.
4884                  */
4885                 schedule_delayed_work(&dev_priv->rps.delayed_resume_work,
4886                                       round_jiffies_up_relative(HZ));
4887         }
4888 }
4889
4890 static void ibx_init_clock_gating(struct drm_device *dev)
4891 {
4892         struct drm_i915_private *dev_priv = dev->dev_private;
4893
4894         /*
4895          * On Ibex Peak and Cougar Point, we need to disable clock
4896          * gating for the panel power sequencer or it will fail to
4897          * start up when no ports are active.
4898          */
4899         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
4900 }
4901
4902 static void g4x_disable_trickle_feed(struct drm_device *dev)
4903 {
4904         struct drm_i915_private *dev_priv = dev->dev_private;
4905         int pipe;
4906
4907         for_each_pipe(pipe) {
4908                 I915_WRITE(DSPCNTR(pipe),
4909                            I915_READ(DSPCNTR(pipe)) |
4910                            DISPPLANE_TRICKLE_FEED_DISABLE);
4911                 intel_flush_primary_plane(dev_priv, pipe);
4912         }
4913 }
4914
4915 static void ironlake_init_clock_gating(struct drm_device *dev)
4916 {
4917         struct drm_i915_private *dev_priv = dev->dev_private;
4918         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
4919
4920         /*
4921          * Required for FBC
4922          * WaFbcDisableDpfcClockGating:ilk
4923          */
4924         dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
4925                    ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
4926                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
4927
4928         I915_WRITE(PCH_3DCGDIS0,
4929                    MARIUNIT_CLOCK_GATE_DISABLE |
4930                    SVSMUNIT_CLOCK_GATE_DISABLE);
4931         I915_WRITE(PCH_3DCGDIS1,
4932                    VFMUNIT_CLOCK_GATE_DISABLE);
4933
4934         /*
4935          * According to the spec the following bits should be set in
4936          * order to enable memory self-refresh
4937          * The bit 22/21 of 0x42004
4938          * The bit 5 of 0x42020
4939          * The bit 15 of 0x45000
4940          */
4941         I915_WRITE(ILK_DISPLAY_CHICKEN2,
4942                    (I915_READ(ILK_DISPLAY_CHICKEN2) |
4943                     ILK_DPARB_GATE | ILK_VSDPFD_FULL));
4944         dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
4945         I915_WRITE(DISP_ARB_CTL,
4946                    (I915_READ(DISP_ARB_CTL) |
4947                     DISP_FBC_WM_DIS));
4948         I915_WRITE(WM3_LP_ILK, 0);
4949         I915_WRITE(WM2_LP_ILK, 0);
4950         I915_WRITE(WM1_LP_ILK, 0);
4951
4952         /*
4953          * Based on the document from hardware guys the following bits
4954          * should be set unconditionally in order to enable FBC.
4955          * The bit 22 of 0x42000
4956          * The bit 22 of 0x42004
4957          * The bit 7,8,9 of 0x42020.
4958          */
4959         if (IS_IRONLAKE_M(dev)) {
4960                 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
4961                 I915_WRITE(ILK_DISPLAY_CHICKEN1,
4962                            I915_READ(ILK_DISPLAY_CHICKEN1) |
4963                            ILK_FBCQ_DIS);
4964                 I915_WRITE(ILK_DISPLAY_CHICKEN2,
4965                            I915_READ(ILK_DISPLAY_CHICKEN2) |
4966                            ILK_DPARB_GATE);
4967         }
4968
4969         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
4970
4971         I915_WRITE(ILK_DISPLAY_CHICKEN2,
4972                    I915_READ(ILK_DISPLAY_CHICKEN2) |
4973                    ILK_ELPIN_409_SELECT);
4974         I915_WRITE(_3D_CHICKEN2,
4975                    _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
4976                    _3D_CHICKEN2_WM_READ_PIPELINED);
4977
4978         /* WaDisableRenderCachePipelinedFlush:ilk */
4979         I915_WRITE(CACHE_MODE_0,
4980                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
4981
4982         g4x_disable_trickle_feed(dev);
4983
4984         ibx_init_clock_gating(dev);
4985 }
4986
4987 static void cpt_init_clock_gating(struct drm_device *dev)
4988 {
4989         struct drm_i915_private *dev_priv = dev->dev_private;
4990         int pipe;
4991         uint32_t val;
4992
4993         /*
4994          * On Ibex Peak and Cougar Point, we need to disable clock
4995          * gating for the panel power sequencer or it will fail to
4996          * start up when no ports are active.
4997          */
4998         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
4999                    PCH_DPLUNIT_CLOCK_GATE_DISABLE |
5000                    PCH_CPUNIT_CLOCK_GATE_DISABLE);
5001         I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
5002                    DPLS_EDP_PPS_FIX_DIS);
5003         /* The below fixes the weird display corruption, a few pixels shifted
5004          * downward, on (only) LVDS of some HP laptops with IVY.
5005          */
5006         for_each_pipe(pipe) {
5007                 val = I915_READ(TRANS_CHICKEN2(pipe));
5008                 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
5009                 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
5010                 if (dev_priv->vbt.fdi_rx_polarity_inverted)
5011                         val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
5012                 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
5013                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
5014                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
5015                 I915_WRITE(TRANS_CHICKEN2(pipe), val);
5016         }
5017         /* WADP0ClockGatingDisable */
5018         for_each_pipe(pipe) {
5019                 I915_WRITE(TRANS_CHICKEN1(pipe),
5020                            TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
5021         }
5022 }
5023
5024 static void gen6_check_mch_setup(struct drm_device *dev)
5025 {
5026         struct drm_i915_private *dev_priv = dev->dev_private;
5027         uint32_t tmp;
5028
5029         tmp = I915_READ(MCH_SSKPD);
5030         if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL) {
5031                 DRM_INFO("Wrong MCH_SSKPD value: 0x%08x\n", tmp);
5032                 DRM_INFO("This can cause pipe underruns and display issues.\n");
5033                 DRM_INFO("Please upgrade your BIOS to fix this.\n");
5034         }
5035 }
5036
5037 static void gen6_init_clock_gating(struct drm_device *dev)
5038 {
5039         struct drm_i915_private *dev_priv = dev->dev_private;
5040         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
5041
5042         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
5043
5044         I915_WRITE(ILK_DISPLAY_CHICKEN2,
5045                    I915_READ(ILK_DISPLAY_CHICKEN2) |
5046                    ILK_ELPIN_409_SELECT);
5047
5048         /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
5049         I915_WRITE(_3D_CHICKEN,
5050                    _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
5051
5052         /* WaSetupGtModeTdRowDispatch:snb */
5053         if (IS_SNB_GT1(dev))
5054                 I915_WRITE(GEN6_GT_MODE,
5055                            _MASKED_BIT_ENABLE(GEN6_TD_FOUR_ROW_DISPATCH_DISABLE));
5056
5057         I915_WRITE(WM3_LP_ILK, 0);
5058         I915_WRITE(WM2_LP_ILK, 0);
5059         I915_WRITE(WM1_LP_ILK, 0);
5060
5061         I915_WRITE(CACHE_MODE_0,
5062                    _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
5063
5064         I915_WRITE(GEN6_UCGCTL1,
5065                    I915_READ(GEN6_UCGCTL1) |
5066                    GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
5067                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
5068
5069         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
5070          * gating disable must be set.  Failure to set it results in
5071          * flickering pixels due to Z write ordering failures after
5072          * some amount of runtime in the Mesa "fire" demo, and Unigine
5073          * Sanctuary and Tropics, and apparently anything else with
5074          * alpha test or pixel discard.
5075          *
5076          * According to the spec, bit 11 (RCCUNIT) must also be set,
5077          * but we didn't debug actual testcases to find it out.
5078          *
5079          * Also apply WaDisableVDSUnitClockGating:snb and
5080          * WaDisableRCPBUnitClockGating:snb.
5081          */
5082         I915_WRITE(GEN6_UCGCTL2,
5083                    GEN7_VDSUNIT_CLOCK_GATE_DISABLE |
5084                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
5085                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
5086
5087         /* Bspec says we need to always set all mask bits. */
5088         I915_WRITE(_3D_CHICKEN3, (0xFFFF << 16) |
5089                    _3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL);
5090
5091         /*
5092          * According to the spec the following bits should be
5093          * set in order to enable memory self-refresh and fbc:
5094          * The bit21 and bit22 of 0x42000
5095          * The bit21 and bit22 of 0x42004
5096          * The bit5 and bit7 of 0x42020
5097          * The bit14 of 0x70180
5098          * The bit14 of 0x71180
5099          *
5100          * WaFbcAsynchFlipDisableFbcQueue:snb
5101          */
5102         I915_WRITE(ILK_DISPLAY_CHICKEN1,
5103                    I915_READ(ILK_DISPLAY_CHICKEN1) |
5104                    ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
5105         I915_WRITE(ILK_DISPLAY_CHICKEN2,
5106                    I915_READ(ILK_DISPLAY_CHICKEN2) |
5107                    ILK_DPARB_GATE | ILK_VSDPFD_FULL);
5108         I915_WRITE(ILK_DSPCLK_GATE_D,
5109                    I915_READ(ILK_DSPCLK_GATE_D) |
5110                    ILK_DPARBUNIT_CLOCK_GATE_ENABLE  |
5111                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
5112
5113         g4x_disable_trickle_feed(dev);
5114
5115         /* The default value should be 0x200 according to docs, but the two
5116          * platforms I checked have a 0 for this. (Maybe BIOS overrides?) */
5117         I915_WRITE(GEN6_GT_MODE, _MASKED_BIT_DISABLE(0xffff));
5118         I915_WRITE(GEN6_GT_MODE, _MASKED_BIT_ENABLE(GEN6_GT_MODE_HI));
5119
5120         cpt_init_clock_gating(dev);
5121
5122         gen6_check_mch_setup(dev);
5123 }
5124
5125 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
5126 {
5127         uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
5128
5129         reg &= ~GEN7_FF_SCHED_MASK;
5130         reg |= GEN7_FF_TS_SCHED_HW;
5131         reg |= GEN7_FF_VS_SCHED_HW;
5132         reg |= GEN7_FF_DS_SCHED_HW;
5133
5134         if (IS_HASWELL(dev_priv->dev))
5135                 reg &= ~GEN7_FF_VS_REF_CNT_FFME;
5136
5137         I915_WRITE(GEN7_FF_THREAD_MODE, reg);
5138 }
5139
5140 static void lpt_init_clock_gating(struct drm_device *dev)
5141 {
5142         struct drm_i915_private *dev_priv = dev->dev_private;
5143
5144         /*
5145          * TODO: this bit should only be enabled when really needed, then
5146          * disabled when not needed anymore in order to save power.
5147          */
5148         if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE)
5149                 I915_WRITE(SOUTH_DSPCLK_GATE_D,
5150                            I915_READ(SOUTH_DSPCLK_GATE_D) |
5151                            PCH_LP_PARTITION_LEVEL_DISABLE);
5152
5153         /* WADPOClockGatingDisable:hsw */
5154         I915_WRITE(_TRANSA_CHICKEN1,
5155                    I915_READ(_TRANSA_CHICKEN1) |
5156                    TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
5157 }
5158
5159 static void lpt_suspend_hw(struct drm_device *dev)
5160 {
5161         struct drm_i915_private *dev_priv = dev->dev_private;
5162
5163         if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE) {
5164                 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
5165
5166                 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
5167                 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
5168         }
5169 }
5170
5171 static void haswell_init_clock_gating(struct drm_device *dev)
5172 {
5173         struct drm_i915_private *dev_priv = dev->dev_private;
5174
5175         I915_WRITE(WM3_LP_ILK, 0);
5176         I915_WRITE(WM2_LP_ILK, 0);
5177         I915_WRITE(WM1_LP_ILK, 0);
5178
5179         /* According to the spec, bit 13 (RCZUNIT) must be set on IVB.
5180          * This implements the WaDisableRCZUnitClockGating:hsw workaround.
5181          */
5182         I915_WRITE(GEN6_UCGCTL2, GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
5183
5184         /* Apply the WaDisableRHWOOptimizationForRenderHang:hsw workaround. */
5185         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
5186                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
5187
5188         /* WaApplyL3ControlAndL3ChickenMode:hsw */
5189         I915_WRITE(GEN7_L3CNTLREG1,
5190                         GEN7_WA_FOR_GEN7_L3_CONTROL);
5191         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
5192                         GEN7_WA_L3_CHICKEN_MODE);
5193
5194         /* L3 caching of data atomics doesn't work -- disable it. */
5195         I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
5196         I915_WRITE(HSW_ROW_CHICKEN3,
5197                    _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
5198
5199         /* This is required by WaCatErrorRejectionIssue:hsw */
5200         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5201                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5202                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5203
5204         /* WaVSRefCountFullforceMissDisable:hsw */
5205         gen7_setup_fixed_func_scheduler(dev_priv);
5206
5207         /* WaDisable4x2SubspanOptimization:hsw */
5208         I915_WRITE(CACHE_MODE_1,
5209                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5210
5211         /* WaSwitchSolVfFArbitrationPriority:hsw */
5212         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
5213
5214         /* WaRsPkgCStateDisplayPMReq:hsw */
5215         I915_WRITE(CHICKEN_PAR1_1,
5216                    I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
5217
5218         lpt_init_clock_gating(dev);
5219 }
5220
5221 static void ivybridge_init_clock_gating(struct drm_device *dev)
5222 {
5223         struct drm_i915_private *dev_priv = dev->dev_private;
5224         uint32_t snpcr;
5225
5226         I915_WRITE(WM3_LP_ILK, 0);
5227         I915_WRITE(WM2_LP_ILK, 0);
5228         I915_WRITE(WM1_LP_ILK, 0);
5229
5230         I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
5231
5232         /* WaDisableEarlyCull:ivb */
5233         I915_WRITE(_3D_CHICKEN3,
5234                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
5235
5236         /* WaDisableBackToBackFlipFix:ivb */
5237         I915_WRITE(IVB_CHICKEN3,
5238                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
5239                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
5240
5241         /* WaDisablePSDDualDispatchEnable:ivb */
5242         if (IS_IVB_GT1(dev))
5243                 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
5244                            _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
5245         else
5246                 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1_GT2,
5247                            _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
5248
5249         /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
5250         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
5251                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
5252
5253         /* WaApplyL3ControlAndL3ChickenMode:ivb */
5254         I915_WRITE(GEN7_L3CNTLREG1,
5255                         GEN7_WA_FOR_GEN7_L3_CONTROL);
5256         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
5257                    GEN7_WA_L3_CHICKEN_MODE);
5258         if (IS_IVB_GT1(dev))
5259                 I915_WRITE(GEN7_ROW_CHICKEN2,
5260                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5261         else
5262                 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
5263                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5264
5265
5266         /* WaForceL3Serialization:ivb */
5267         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
5268                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
5269
5270         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
5271          * gating disable must be set.  Failure to set it results in
5272          * flickering pixels due to Z write ordering failures after
5273          * some amount of runtime in the Mesa "fire" demo, and Unigine
5274          * Sanctuary and Tropics, and apparently anything else with
5275          * alpha test or pixel discard.
5276          *
5277          * According to the spec, bit 11 (RCCUNIT) must also be set,
5278          * but we didn't debug actual testcases to find it out.
5279          *
5280          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
5281          * This implements the WaDisableRCZUnitClockGating:ivb workaround.
5282          */
5283         I915_WRITE(GEN6_UCGCTL2,
5284                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE |
5285                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
5286
5287         /* This is required by WaCatErrorRejectionIssue:ivb */
5288         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5289                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5290                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5291
5292         g4x_disable_trickle_feed(dev);
5293
5294         /* WaVSRefCountFullforceMissDisable:ivb */
5295         gen7_setup_fixed_func_scheduler(dev_priv);
5296
5297         /* WaDisable4x2SubspanOptimization:ivb */
5298         I915_WRITE(CACHE_MODE_1,
5299                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5300
5301         snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
5302         snpcr &= ~GEN6_MBC_SNPCR_MASK;
5303         snpcr |= GEN6_MBC_SNPCR_MED;
5304         I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
5305
5306         if (!HAS_PCH_NOP(dev))
5307                 cpt_init_clock_gating(dev);
5308
5309         gen6_check_mch_setup(dev);
5310 }
5311
5312 static void valleyview_init_clock_gating(struct drm_device *dev)
5313 {
5314         struct drm_i915_private *dev_priv = dev->dev_private;
5315         u32 val;
5316
5317         mutex_lock(&dev_priv->rps.hw_lock);
5318         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5319         mutex_unlock(&dev_priv->rps.hw_lock);
5320         switch ((val >> 6) & 3) {
5321         case 0:
5322                 dev_priv->mem_freq = 800;
5323                 break;
5324         case 1:
5325                 dev_priv->mem_freq = 1066;
5326                 break;
5327         case 2:
5328                 dev_priv->mem_freq = 1333;
5329                 break;
5330         case 3:
5331                 /*
5332                  * Probably a BIOS/Punit bug, or a new platform we don't
5333                  * support yet.
5334                  */
5335                 WARN(1, "invalid DDR freq detected, assuming 800MHz\n");
5336                 dev_priv->mem_freq = 800;
5337                 break;
5338         }
5339         DRM_DEBUG_DRIVER("DDR speed: %d MHz", dev_priv->mem_freq);
5340
5341         I915_WRITE(DSPCLK_GATE_D, VRHUNIT_CLOCK_GATE_DISABLE);
5342
5343         /* WaDisableEarlyCull:vlv */
5344         I915_WRITE(_3D_CHICKEN3,
5345                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
5346
5347         /* WaDisableBackToBackFlipFix:vlv */
5348         I915_WRITE(IVB_CHICKEN3,
5349                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
5350                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
5351
5352         /* WaDisablePSDDualDispatchEnable:vlv */
5353         I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
5354                    _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
5355                                       GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
5356
5357         /* Apply the WaDisableRHWOOptimizationForRenderHang:vlv workaround. */
5358         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
5359                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
5360
5361         /* WaApplyL3ControlAndL3ChickenMode:vlv */
5362         I915_WRITE(GEN7_L3CNTLREG1, I915_READ(GEN7_L3CNTLREG1) | GEN7_L3AGDIS);
5363         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER, GEN7_WA_L3_CHICKEN_MODE);
5364
5365         /* WaForceL3Serialization:vlv */
5366         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
5367                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
5368
5369         /* WaDisableDopClockGating:vlv */
5370         I915_WRITE(GEN7_ROW_CHICKEN2,
5371                    _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5372
5373         /* This is required by WaCatErrorRejectionIssue:vlv */
5374         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5375                    I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5376                    GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5377
5378         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
5379          * gating disable must be set.  Failure to set it results in
5380          * flickering pixels due to Z write ordering failures after
5381          * some amount of runtime in the Mesa "fire" demo, and Unigine
5382          * Sanctuary and Tropics, and apparently anything else with
5383          * alpha test or pixel discard.
5384          *
5385          * According to the spec, bit 11 (RCCUNIT) must also be set,
5386          * but we didn't debug actual testcases to find it out.
5387          *
5388          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
5389          * This implements the WaDisableRCZUnitClockGating:vlv workaround.
5390          *
5391          * Also apply WaDisableVDSUnitClockGating:vlv and
5392          * WaDisableRCPBUnitClockGating:vlv.
5393          */
5394         I915_WRITE(GEN6_UCGCTL2,
5395                    GEN7_VDSUNIT_CLOCK_GATE_DISABLE |
5396                    GEN7_TDLUNIT_CLOCK_GATE_DISABLE |
5397                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE |
5398                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
5399                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
5400
5401         I915_WRITE(GEN7_UCGCTL4, GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
5402
5403         I915_WRITE(MI_ARB_VLV, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE);
5404
5405         I915_WRITE(CACHE_MODE_1,
5406                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5407
5408         /*
5409          * WaDisableVLVClockGating_VBIIssue:vlv
5410          * Disable clock gating on th GCFG unit to prevent a delay
5411          * in the reporting of vblank events.
5412          */
5413         I915_WRITE(VLV_GUNIT_CLOCK_GATE, 0xffffffff);
5414
5415         /* Conservative clock gating settings for now */
5416         I915_WRITE(0x9400, 0xffffffff);
5417         I915_WRITE(0x9404, 0xffffffff);
5418         I915_WRITE(0x9408, 0xffffffff);
5419         I915_WRITE(0x940c, 0xffffffff);
5420         I915_WRITE(0x9410, 0xffffffff);
5421         I915_WRITE(0x9414, 0xffffffff);
5422         I915_WRITE(0x9418, 0xffffffff);
5423 }
5424
5425 static void g4x_init_clock_gating(struct drm_device *dev)
5426 {
5427         struct drm_i915_private *dev_priv = dev->dev_private;
5428         uint32_t dspclk_gate;
5429
5430         I915_WRITE(RENCLK_GATE_D1, 0);
5431         I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
5432                    GS_UNIT_CLOCK_GATE_DISABLE |
5433                    CL_UNIT_CLOCK_GATE_DISABLE);
5434         I915_WRITE(RAMCLK_GATE_D, 0);
5435         dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
5436                 OVRUNIT_CLOCK_GATE_DISABLE |
5437                 OVCUNIT_CLOCK_GATE_DISABLE;
5438         if (IS_GM45(dev))
5439                 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
5440         I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
5441
5442         /* WaDisableRenderCachePipelinedFlush */
5443         I915_WRITE(CACHE_MODE_0,
5444                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
5445
5446         g4x_disable_trickle_feed(dev);
5447 }
5448
5449 static void crestline_init_clock_gating(struct drm_device *dev)
5450 {
5451         struct drm_i915_private *dev_priv = dev->dev_private;
5452
5453         I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
5454         I915_WRITE(RENCLK_GATE_D2, 0);
5455         I915_WRITE(DSPCLK_GATE_D, 0);
5456         I915_WRITE(RAMCLK_GATE_D, 0);
5457         I915_WRITE16(DEUC, 0);
5458         I915_WRITE(MI_ARB_STATE,
5459                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
5460 }
5461
5462 static void broadwater_init_clock_gating(struct drm_device *dev)
5463 {
5464         struct drm_i915_private *dev_priv = dev->dev_private;
5465
5466         I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
5467                    I965_RCC_CLOCK_GATE_DISABLE |
5468                    I965_RCPB_CLOCK_GATE_DISABLE |
5469                    I965_ISC_CLOCK_GATE_DISABLE |
5470                    I965_FBC_CLOCK_GATE_DISABLE);
5471         I915_WRITE(RENCLK_GATE_D2, 0);
5472         I915_WRITE(MI_ARB_STATE,
5473                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
5474 }
5475
5476 static void gen3_init_clock_gating(struct drm_device *dev)
5477 {
5478         struct drm_i915_private *dev_priv = dev->dev_private;
5479         u32 dstate = I915_READ(D_STATE);
5480
5481         dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
5482                 DSTATE_DOT_CLOCK_GATING;
5483         I915_WRITE(D_STATE, dstate);
5484
5485         if (IS_PINEVIEW(dev))
5486                 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
5487
5488         /* IIR "flip pending" means done if this bit is set */
5489         I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
5490 }
5491
5492 static void i85x_init_clock_gating(struct drm_device *dev)
5493 {
5494         struct drm_i915_private *dev_priv = dev->dev_private;
5495
5496         I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
5497 }
5498
5499 static void i830_init_clock_gating(struct drm_device *dev)
5500 {
5501         struct drm_i915_private *dev_priv = dev->dev_private;
5502
5503         I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
5504 }
5505
5506 void intel_init_clock_gating(struct drm_device *dev)
5507 {
5508         struct drm_i915_private *dev_priv = dev->dev_private;
5509
5510         dev_priv->display.init_clock_gating(dev);
5511 }
5512
5513 void intel_suspend_hw(struct drm_device *dev)
5514 {
5515         if (HAS_PCH_LPT(dev))
5516                 lpt_suspend_hw(dev);
5517 }
5518
5519 static bool is_always_on_power_domain(struct drm_device *dev,
5520                                       enum intel_display_power_domain domain)
5521 {
5522         unsigned long always_on_domains;
5523
5524         BUG_ON(BIT(domain) & ~POWER_DOMAIN_MASK);
5525
5526         if (IS_HASWELL(dev)) {
5527                 always_on_domains = HSW_ALWAYS_ON_POWER_DOMAINS;
5528         } else {
5529                 WARN_ON(1);
5530                 return true;
5531         }
5532
5533         return BIT(domain) & always_on_domains;
5534 }
5535
5536 /**
5537  * We should only use the power well if we explicitly asked the hardware to
5538  * enable it, so check if it's enabled and also check if we've requested it to
5539  * be enabled.
5540  */
5541 bool intel_display_power_enabled(struct drm_device *dev,
5542                                  enum intel_display_power_domain domain)
5543 {
5544         struct drm_i915_private *dev_priv = dev->dev_private;
5545
5546         if (!HAS_POWER_WELL(dev))
5547                 return true;
5548
5549         if (is_always_on_power_domain(dev, domain))
5550                 return true;
5551
5552         return I915_READ(HSW_PWR_WELL_DRIVER) ==
5553                      (HSW_PWR_WELL_ENABLE_REQUEST | HSW_PWR_WELL_STATE_ENABLED);
5554 }
5555
5556 static void __intel_set_power_well(struct drm_device *dev, bool enable)
5557 {
5558         struct drm_i915_private *dev_priv = dev->dev_private;
5559         bool is_enabled, enable_requested;
5560         uint32_t tmp;
5561
5562         tmp = I915_READ(HSW_PWR_WELL_DRIVER);
5563         is_enabled = tmp & HSW_PWR_WELL_STATE_ENABLED;
5564         enable_requested = tmp & HSW_PWR_WELL_ENABLE_REQUEST;
5565
5566         if (enable) {
5567                 if (!enable_requested)
5568                         I915_WRITE(HSW_PWR_WELL_DRIVER,
5569                                    HSW_PWR_WELL_ENABLE_REQUEST);
5570
5571                 if (!is_enabled) {
5572                         DRM_DEBUG_KMS("Enabling power well\n");
5573                         if (wait_for((I915_READ(HSW_PWR_WELL_DRIVER) &
5574                                       HSW_PWR_WELL_STATE_ENABLED), 20))
5575                                 DRM_ERROR("Timeout enabling power well\n");
5576                 }
5577         } else {
5578                 if (enable_requested) {
5579                         unsigned long irqflags;
5580                         enum pipe p;
5581
5582                         I915_WRITE(HSW_PWR_WELL_DRIVER, 0);
5583                         POSTING_READ(HSW_PWR_WELL_DRIVER);
5584                         DRM_DEBUG_KMS("Requesting to disable the power well\n");
5585
5586                         /*
5587                          * After this, the registers on the pipes that are part
5588                          * of the power well will become zero, so we have to
5589                          * adjust our counters according to that.
5590                          *
5591                          * FIXME: Should we do this in general in
5592                          * drm_vblank_post_modeset?
5593                          */
5594                         spin_lock_irqsave(&dev->vbl_lock, irqflags);
5595                         for_each_pipe(p)
5596                                 if (p != PIPE_A)
5597                                         dev->vblank[p].last = 0;
5598                         spin_unlock_irqrestore(&dev->vbl_lock, irqflags);
5599                 }
5600         }
5601 }
5602
5603 static void __intel_power_well_get(struct drm_device *dev,
5604                                    struct i915_power_well *power_well)
5605 {
5606         if (!power_well->count++)
5607                 __intel_set_power_well(dev, true);
5608 }
5609
5610 static void __intel_power_well_put(struct drm_device *dev,
5611                                    struct i915_power_well *power_well)
5612 {
5613         WARN_ON(!power_well->count);
5614         if (!--power_well->count && i915_disable_power_well)
5615                 __intel_set_power_well(dev, false);
5616 }
5617
5618 void intel_display_power_get(struct drm_device *dev,
5619                              enum intel_display_power_domain domain)
5620 {
5621         struct drm_i915_private *dev_priv = dev->dev_private;
5622         struct i915_power_domains *power_domains;
5623
5624         if (!HAS_POWER_WELL(dev))
5625                 return;
5626
5627         if (is_always_on_power_domain(dev, domain))
5628                 return;
5629
5630         power_domains = &dev_priv->power_domains;
5631
5632         mutex_lock(&power_domains->lock);
5633         __intel_power_well_get(dev, &power_domains->power_wells[0]);
5634         mutex_unlock(&power_domains->lock);
5635 }
5636
5637 void intel_display_power_put(struct drm_device *dev,
5638                              enum intel_display_power_domain domain)
5639 {
5640         struct drm_i915_private *dev_priv = dev->dev_private;
5641         struct i915_power_domains *power_domains;
5642
5643         if (!HAS_POWER_WELL(dev))
5644                 return;
5645
5646         if (is_always_on_power_domain(dev, domain))
5647                 return;
5648
5649         power_domains = &dev_priv->power_domains;
5650
5651         mutex_lock(&power_domains->lock);
5652         __intel_power_well_put(dev, &power_domains->power_wells[0]);
5653         mutex_unlock(&power_domains->lock);
5654 }
5655
5656 static struct i915_power_domains *hsw_pwr;
5657
5658 /* Display audio driver power well request */
5659 void i915_request_power_well(void)
5660 {
5661         struct drm_i915_private *dev_priv;
5662
5663         if (WARN_ON(!hsw_pwr))
5664                 return;
5665
5666         dev_priv = container_of(hsw_pwr, struct drm_i915_private,
5667                                 power_domains);
5668
5669         mutex_lock(&hsw_pwr->lock);
5670         __intel_power_well_get(dev_priv->dev, &hsw_pwr->power_wells[0]);
5671         mutex_unlock(&hsw_pwr->lock);
5672 }
5673 EXPORT_SYMBOL_GPL(i915_request_power_well);
5674
5675 /* Display audio driver power well release */
5676 void i915_release_power_well(void)
5677 {
5678         struct drm_i915_private *dev_priv;
5679
5680         if (WARN_ON(!hsw_pwr))
5681                 return;
5682
5683         dev_priv = container_of(hsw_pwr, struct drm_i915_private,
5684                                 power_domains);
5685
5686         mutex_lock(&hsw_pwr->lock);
5687         __intel_power_well_put(dev_priv->dev, &hsw_pwr->power_wells[0]);
5688         mutex_unlock(&hsw_pwr->lock);
5689 }
5690 EXPORT_SYMBOL_GPL(i915_release_power_well);
5691
5692 int intel_power_domains_init(struct drm_device *dev)
5693 {
5694         struct drm_i915_private *dev_priv = dev->dev_private;
5695         struct i915_power_domains *power_domains = &dev_priv->power_domains;
5696         struct i915_power_well *power_well;
5697
5698         mutex_init(&power_domains->lock);
5699         hsw_pwr = power_domains;
5700
5701         power_well = &power_domains->power_wells[0];
5702         power_well->count = 0;
5703
5704         return 0;
5705 }
5706
5707 void intel_power_domains_remove(struct drm_device *dev)
5708 {
5709         hsw_pwr = NULL;
5710 }
5711
5712 static void intel_power_domains_resume(struct drm_device *dev)
5713 {
5714         struct drm_i915_private *dev_priv = dev->dev_private;
5715         struct i915_power_domains *power_domains = &dev_priv->power_domains;
5716         struct i915_power_well *power_well;
5717
5718         if (!HAS_POWER_WELL(dev))
5719                 return;
5720
5721         mutex_lock(&power_domains->lock);
5722
5723         power_well = &power_domains->power_wells[0];
5724         __intel_set_power_well(dev, power_well->count > 0);
5725
5726         mutex_unlock(&power_domains->lock);
5727 }
5728
5729 /*
5730  * Starting with Haswell, we have a "Power Down Well" that can be turned off
5731  * when not needed anymore. We have 4 registers that can request the power well
5732  * to be enabled, and it will only be disabled if none of the registers is
5733  * requesting it to be enabled.
5734  */
5735 void intel_power_domains_init_hw(struct drm_device *dev)
5736 {
5737         struct drm_i915_private *dev_priv = dev->dev_private;
5738
5739         if (!HAS_POWER_WELL(dev))
5740                 return;
5741
5742         /* For now, we need the power well to be always enabled. */
5743         intel_display_set_init_power(dev, true);
5744         intel_power_domains_resume(dev);
5745
5746         /* We're taking over the BIOS, so clear any requests made by it since
5747          * the driver is in charge now. */
5748         if (I915_READ(HSW_PWR_WELL_BIOS) & HSW_PWR_WELL_ENABLE_REQUEST)
5749                 I915_WRITE(HSW_PWR_WELL_BIOS, 0);
5750 }
5751
5752 /* Disables PC8 so we can use the GMBUS and DP AUX interrupts. */
5753 void intel_aux_display_runtime_get(struct drm_i915_private *dev_priv)
5754 {
5755         hsw_disable_package_c8(dev_priv);
5756 }
5757
5758 void intel_aux_display_runtime_put(struct drm_i915_private *dev_priv)
5759 {
5760         hsw_enable_package_c8(dev_priv);
5761 }
5762
5763 /* Set up chip specific power management-related functions */
5764 void intel_init_pm(struct drm_device *dev)
5765 {
5766         struct drm_i915_private *dev_priv = dev->dev_private;
5767
5768         if (I915_HAS_FBC(dev)) {
5769                 if (HAS_PCH_SPLIT(dev)) {
5770                         dev_priv->display.fbc_enabled = ironlake_fbc_enabled;
5771                         if (IS_IVYBRIDGE(dev) || IS_HASWELL(dev))
5772                                 dev_priv->display.enable_fbc =
5773                                         gen7_enable_fbc;
5774                         else
5775                                 dev_priv->display.enable_fbc =
5776                                         ironlake_enable_fbc;
5777                         dev_priv->display.disable_fbc = ironlake_disable_fbc;
5778                 } else if (IS_GM45(dev)) {
5779                         dev_priv->display.fbc_enabled = g4x_fbc_enabled;
5780                         dev_priv->display.enable_fbc = g4x_enable_fbc;
5781                         dev_priv->display.disable_fbc = g4x_disable_fbc;
5782                 } else if (IS_CRESTLINE(dev)) {
5783                         dev_priv->display.fbc_enabled = i8xx_fbc_enabled;
5784                         dev_priv->display.enable_fbc = i8xx_enable_fbc;
5785                         dev_priv->display.disable_fbc = i8xx_disable_fbc;
5786                 }
5787                 /* 855GM needs testing */
5788         }
5789
5790         /* For cxsr */
5791         if (IS_PINEVIEW(dev))
5792                 i915_pineview_get_mem_freq(dev);
5793         else if (IS_GEN5(dev))
5794                 i915_ironlake_get_mem_freq(dev);
5795
5796         /* For FIFO watermark updates */
5797         if (HAS_PCH_SPLIT(dev)) {
5798                 intel_setup_wm_latency(dev);
5799
5800                 if (IS_GEN5(dev)) {
5801                         if (dev_priv->wm.pri_latency[1] &&
5802                             dev_priv->wm.spr_latency[1] &&
5803                             dev_priv->wm.cur_latency[1])
5804                                 dev_priv->display.update_wm = ironlake_update_wm;
5805                         else {
5806                                 DRM_DEBUG_KMS("Failed to get proper latency. "
5807                                               "Disable CxSR\n");
5808                                 dev_priv->display.update_wm = NULL;
5809                         }
5810                         dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
5811                 } else if (IS_GEN6(dev)) {
5812                         if (dev_priv->wm.pri_latency[0] &&
5813                             dev_priv->wm.spr_latency[0] &&
5814                             dev_priv->wm.cur_latency[0]) {
5815                                 dev_priv->display.update_wm = sandybridge_update_wm;
5816                                 dev_priv->display.update_sprite_wm = sandybridge_update_sprite_wm;
5817                         } else {
5818                                 DRM_DEBUG_KMS("Failed to read display plane latency. "
5819                                               "Disable CxSR\n");
5820                                 dev_priv->display.update_wm = NULL;
5821                         }
5822                         dev_priv->display.init_clock_gating = gen6_init_clock_gating;
5823                 } else if (IS_IVYBRIDGE(dev)) {
5824                         if (dev_priv->wm.pri_latency[0] &&
5825                             dev_priv->wm.spr_latency[0] &&
5826                             dev_priv->wm.cur_latency[0]) {
5827                                 dev_priv->display.update_wm = ivybridge_update_wm;
5828                                 dev_priv->display.update_sprite_wm = sandybridge_update_sprite_wm;
5829                         } else {
5830                                 DRM_DEBUG_KMS("Failed to read display plane latency. "
5831                                               "Disable CxSR\n");
5832                                 dev_priv->display.update_wm = NULL;
5833                         }
5834                         dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
5835                 } else if (IS_HASWELL(dev)) {
5836                         if (dev_priv->wm.pri_latency[0] &&
5837                             dev_priv->wm.spr_latency[0] &&
5838                             dev_priv->wm.cur_latency[0]) {
5839                                 dev_priv->display.update_wm = haswell_update_wm;
5840                                 dev_priv->display.update_sprite_wm =
5841                                         haswell_update_sprite_wm;
5842                         } else {
5843                                 DRM_DEBUG_KMS("Failed to read display plane latency. "
5844                                               "Disable CxSR\n");
5845                                 dev_priv->display.update_wm = NULL;
5846                         }
5847                         dev_priv->display.init_clock_gating = haswell_init_clock_gating;
5848                 } else
5849                         dev_priv->display.update_wm = NULL;
5850         } else if (IS_VALLEYVIEW(dev)) {
5851                 dev_priv->display.update_wm = valleyview_update_wm;
5852                 dev_priv->display.init_clock_gating =
5853                         valleyview_init_clock_gating;
5854         } else if (IS_PINEVIEW(dev)) {
5855                 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
5856                                             dev_priv->is_ddr3,
5857                                             dev_priv->fsb_freq,
5858                                             dev_priv->mem_freq)) {
5859                         DRM_INFO("failed to find known CxSR latency "
5860                                  "(found ddr%s fsb freq %d, mem freq %d), "
5861                                  "disabling CxSR\n",
5862                                  (dev_priv->is_ddr3 == 1) ? "3" : "2",
5863                                  dev_priv->fsb_freq, dev_priv->mem_freq);
5864                         /* Disable CxSR and never update its watermark again */
5865                         pineview_disable_cxsr(dev);
5866                         dev_priv->display.update_wm = NULL;
5867                 } else
5868                         dev_priv->display.update_wm = pineview_update_wm;
5869                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
5870         } else if (IS_G4X(dev)) {
5871                 dev_priv->display.update_wm = g4x_update_wm;
5872                 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
5873         } else if (IS_GEN4(dev)) {
5874                 dev_priv->display.update_wm = i965_update_wm;
5875                 if (IS_CRESTLINE(dev))
5876                         dev_priv->display.init_clock_gating = crestline_init_clock_gating;
5877                 else if (IS_BROADWATER(dev))
5878                         dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
5879         } else if (IS_GEN3(dev)) {
5880                 dev_priv->display.update_wm = i9xx_update_wm;
5881                 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
5882                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
5883         } else if (IS_I865G(dev)) {
5884                 dev_priv->display.update_wm = i830_update_wm;
5885                 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
5886                 dev_priv->display.get_fifo_size = i830_get_fifo_size;
5887         } else if (IS_I85X(dev)) {
5888                 dev_priv->display.update_wm = i9xx_update_wm;
5889                 dev_priv->display.get_fifo_size = i85x_get_fifo_size;
5890                 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
5891         } else {
5892                 dev_priv->display.update_wm = i830_update_wm;
5893                 dev_priv->display.init_clock_gating = i830_init_clock_gating;
5894                 if (IS_845G(dev))
5895                         dev_priv->display.get_fifo_size = i845_get_fifo_size;
5896                 else
5897                         dev_priv->display.get_fifo_size = i830_get_fifo_size;
5898         }
5899 }
5900
5901 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u8 mbox, u32 *val)
5902 {
5903         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5904
5905         if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
5906                 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
5907                 return -EAGAIN;
5908         }
5909
5910         I915_WRITE(GEN6_PCODE_DATA, *val);
5911         I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
5912
5913         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
5914                      500)) {
5915                 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
5916                 return -ETIMEDOUT;
5917         }
5918
5919         *val = I915_READ(GEN6_PCODE_DATA);
5920         I915_WRITE(GEN6_PCODE_DATA, 0);
5921
5922         return 0;
5923 }
5924
5925 int sandybridge_pcode_write(struct drm_i915_private *dev_priv, u8 mbox, u32 val)
5926 {
5927         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5928
5929         if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
5930                 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
5931                 return -EAGAIN;
5932         }
5933
5934         I915_WRITE(GEN6_PCODE_DATA, val);
5935         I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
5936
5937         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
5938                      500)) {
5939                 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
5940                 return -ETIMEDOUT;
5941         }
5942
5943         I915_WRITE(GEN6_PCODE_DATA, 0);
5944
5945         return 0;
5946 }
5947
5948 int vlv_gpu_freq(int ddr_freq, int val)
5949 {
5950         int mult, base;
5951
5952         switch (ddr_freq) {
5953         case 800:
5954                 mult = 20;
5955                 base = 120;
5956                 break;
5957         case 1066:
5958                 mult = 22;
5959                 base = 133;
5960                 break;
5961         case 1333:
5962                 mult = 21;
5963                 base = 125;
5964                 break;
5965         default:
5966                 return -1;
5967         }
5968
5969         return ((val - 0xbd) * mult) + base;
5970 }
5971
5972 int vlv_freq_opcode(int ddr_freq, int val)
5973 {
5974         int mult, base;
5975
5976         switch (ddr_freq) {
5977         case 800:
5978                 mult = 20;
5979                 base = 120;
5980                 break;
5981         case 1066:
5982                 mult = 22;
5983                 base = 133;
5984                 break;
5985         case 1333:
5986                 mult = 21;
5987                 base = 125;
5988                 break;
5989         default:
5990                 return -1;
5991         }
5992
5993         val /= mult;
5994         val -= base / mult;
5995         val += 0xbd;
5996
5997         if (val > 0xea)
5998                 val = 0xea;
5999
6000         return val;
6001 }
6002
6003 void intel_pm_init(struct drm_device *dev)
6004 {
6005         struct drm_i915_private *dev_priv = dev->dev_private;
6006
6007         INIT_DELAYED_WORK(&dev_priv->rps.delayed_resume_work,
6008                           intel_gen6_powersave_work);
6009 }
6010