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 /* FBC, or Frame Buffer Compression, is a technique employed to compress the
  36  * framebuffer contents in-memory, aiming at reducing the required bandwidth
  37  * during in-memory transfers and, therefore, reduce the power packet.
  38  *
  39  * The benefits of FBC are mostly visible with solid backgrounds and
  40  * variation-less patterns.
  41  *
  42  * FBC-related functionality can be enabled by the means of the
  43  * i915.i915_enable_fbc parameter
  44  */
  45
  46 static void i8xx_disable_fbc(struct drm_device *dev)
  47 {
  48         struct drm_i915_private *dev_priv = dev->dev_private;
  49         u32 fbc_ctl;
  50
  51         /* Disable compression */
  52         fbc_ctl = I915_READ(FBC_CONTROL);
  53         if ((fbc_ctl & FBC_CTL_EN) == 0)
  54                 return;
  55
  56         fbc_ctl &= ~FBC_CTL_EN;
  57         I915_WRITE(FBC_CONTROL, fbc_ctl);
  58
  59         /* Wait for compressing bit to clear */
  60         if (wait_for((I915_READ(FBC_STATUS) & FBC_STAT_COMPRESSING) == 0, 10)) {
  61                 DRM_DEBUG_KMS("FBC idle timed out\n");
  62                 return;
  63         }
  64
  65         DRM_DEBUG_KMS("disabled FBC\n");
  66 }
  67
  68 static void i8xx_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
  69 {
  70         struct drm_device *dev = crtc->dev;
  71         struct drm_i915_private *dev_priv = dev->dev_private;
  72         struct drm_framebuffer *fb = crtc->fb;
  73         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
  74         struct drm_i915_gem_object *obj = intel_fb->obj;
  75         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
  76         int cfb_pitch;
  77         int plane, i;
  78         u32 fbc_ctl, fbc_ctl2;
  79
  80         cfb_pitch = dev_priv->fbc.size / FBC_LL_SIZE;
  81         if (fb->pitches[0] < cfb_pitch)
  82                 cfb_pitch = fb->pitches[0];
  83
  84         /* FBC_CTL wants 64B units */
  85         cfb_pitch = (cfb_pitch / 64) - 1;
  86         plane = intel_crtc->plane == 0 ? FBC_CTL_PLANEA : FBC_CTL_PLANEB;
  87
  88         /* Clear old tags */
  89         for (i = 0; i < (FBC_LL_SIZE / 32) + 1; i++)
  90                 I915_WRITE(FBC_TAG + (i * 4), 0);
  91
  92         /* Set it up... */
  93         fbc_ctl2 = FBC_CTL_FENCE_DBL | FBC_CTL_IDLE_IMM | FBC_CTL_CPU_FENCE;
  94         fbc_ctl2 |= plane;
  95         I915_WRITE(FBC_CONTROL2, fbc_ctl2);
  96         I915_WRITE(FBC_FENCE_OFF, crtc->y);
  97
  98         /* enable it... */
  99         fbc_ctl = FBC_CTL_EN | FBC_CTL_PERIODIC;
 100         if (IS_I945GM(dev))
 101                 fbc_ctl |= FBC_CTL_C3_IDLE; /* 945 needs special SR handling */
 102         fbc_ctl |= (cfb_pitch & 0xff) << FBC_CTL_STRIDE_SHIFT;
 103         fbc_ctl |= (interval & 0x2fff) << FBC_CTL_INTERVAL_SHIFT;
 104         fbc_ctl |= obj->fence_reg;
 105         I915_WRITE(FBC_CONTROL, fbc_ctl);
 106
 107         DRM_DEBUG_KMS("enabled FBC, pitch %d, yoff %d, plane %c, ",
 108                       cfb_pitch, crtc->y, plane_name(intel_crtc->plane));
 109 }
 110
 111 static bool i8xx_fbc_enabled(struct drm_device *dev)
 112 {
 113         struct drm_i915_private *dev_priv = dev->dev_private;
 114
 115         return I915_READ(FBC_CONTROL) & FBC_CTL_EN;
 116 }
 117
 118 static void g4x_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 119 {
 120         struct drm_device *dev = crtc->dev;
 121         struct drm_i915_private *dev_priv = dev->dev_private;
 122         struct drm_framebuffer *fb = crtc->fb;
 123         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 124         struct drm_i915_gem_object *obj = intel_fb->obj;
 125         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 126         int plane = intel_crtc->plane == 0 ? DPFC_CTL_PLANEA : DPFC_CTL_PLANEB;
 127         unsigned long stall_watermark = 200;
 128         u32 dpfc_ctl;
 129
 130         dpfc_ctl = plane | DPFC_SR_EN | DPFC_CTL_LIMIT_1X;
 131         dpfc_ctl |= DPFC_CTL_FENCE_EN | obj->fence_reg;
 132         I915_WRITE(DPFC_CHICKEN, DPFC_HT_MODIFY);
 133
 134         I915_WRITE(DPFC_RECOMP_CTL, DPFC_RECOMP_STALL_EN |
 135                    (stall_watermark << DPFC_RECOMP_STALL_WM_SHIFT) |
 136                    (interval << DPFC_RECOMP_TIMER_COUNT_SHIFT));
 137         I915_WRITE(DPFC_FENCE_YOFF, crtc->y);
 138
 139         /* enable it... */
 140         I915_WRITE(DPFC_CONTROL, I915_READ(DPFC_CONTROL) | DPFC_CTL_EN);
 141
 142         DRM_DEBUG_KMS("enabled fbc on plane %c\n", plane_name(intel_crtc->plane));
 143 }
 144
 145 static void g4x_disable_fbc(struct drm_device *dev)
 146 {
 147         struct drm_i915_private *dev_priv = dev->dev_private;
 148         u32 dpfc_ctl;
 149
 150         /* Disable compression */
 151         dpfc_ctl = I915_READ(DPFC_CONTROL);
 152         if (dpfc_ctl & DPFC_CTL_EN) {
 153                 dpfc_ctl &= ~DPFC_CTL_EN;
 154                 I915_WRITE(DPFC_CONTROL, dpfc_ctl);
 155
 156                 DRM_DEBUG_KMS("disabled FBC\n");
 157         }
 158 }
 159
 160 static bool g4x_fbc_enabled(struct drm_device *dev)
 161 {
 162         struct drm_i915_private *dev_priv = dev->dev_private;
 163
 164         return I915_READ(DPFC_CONTROL) & DPFC_CTL_EN;
 165 }
 166
 167 static void sandybridge_blit_fbc_update(struct drm_device *dev)
 168 {
 169         struct drm_i915_private *dev_priv = dev->dev_private;
 170         u32 blt_ecoskpd;
 171
 172         /* Make sure blitter notifies FBC of writes */
 173         gen6_gt_force_wake_get(dev_priv);
 174         blt_ecoskpd = I915_READ(GEN6_BLITTER_ECOSKPD);
 175         blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY <<
 176                 GEN6_BLITTER_LOCK_SHIFT;
 177         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 178         blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY;
 179         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 180         blt_ecoskpd &= ~(GEN6_BLITTER_FBC_NOTIFY <<
 181                          GEN6_BLITTER_LOCK_SHIFT);
 182         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 183         POSTING_READ(GEN6_BLITTER_ECOSKPD);
 184         gen6_gt_force_wake_put(dev_priv);
 185 }
 186
 187 static void ironlake_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 188 {
 189         struct drm_device *dev = crtc->dev;
 190         struct drm_i915_private *dev_priv = dev->dev_private;
 191         struct drm_framebuffer *fb = crtc->fb;
 192         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 193         struct drm_i915_gem_object *obj = intel_fb->obj;
 194         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 195         int plane = intel_crtc->plane == 0 ? DPFC_CTL_PLANEA : DPFC_CTL_PLANEB;
 196         unsigned long stall_watermark = 200;
 197         u32 dpfc_ctl;
 198
 199         dpfc_ctl = I915_READ(ILK_DPFC_CONTROL);
 200         dpfc_ctl &= DPFC_RESERVED;
 201         dpfc_ctl |= (plane | DPFC_CTL_LIMIT_1X);
 202         /* Set persistent mode for front-buffer rendering, ala X. */
 203         dpfc_ctl |= DPFC_CTL_PERSISTENT_MODE;
 204         dpfc_ctl |= (DPFC_CTL_FENCE_EN | obj->fence_reg);
 205         I915_WRITE(ILK_DPFC_CHICKEN, DPFC_HT_MODIFY);
 206
 207         I915_WRITE(ILK_DPFC_RECOMP_CTL, DPFC_RECOMP_STALL_EN |
 208                    (stall_watermark << DPFC_RECOMP_STALL_WM_SHIFT) |
 209                    (interval << DPFC_RECOMP_TIMER_COUNT_SHIFT));
 210         I915_WRITE(ILK_DPFC_FENCE_YOFF, crtc->y);
 211         I915_WRITE(ILK_FBC_RT_BASE, i915_gem_obj_ggtt_offset(obj) | ILK_FBC_RT_VALID);
 212         /* enable it... */
 213         I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
 214
 215         if (IS_GEN6(dev)) {
 216                 I915_WRITE(SNB_DPFC_CTL_SA,
 217                            SNB_CPU_FENCE_ENABLE | obj->fence_reg);
 218                 I915_WRITE(DPFC_CPU_FENCE_OFFSET, crtc->y);
 219                 sandybridge_blit_fbc_update(dev);
 220         }
 221
 222         DRM_DEBUG_KMS("enabled fbc on plane %c\n", plane_name(intel_crtc->plane));
 223 }
 224
 225 static void ironlake_disable_fbc(struct drm_device *dev)
 226 {
 227         struct drm_i915_private *dev_priv = dev->dev_private;
 228         u32 dpfc_ctl;
 229
 230         /* Disable compression */
 231         dpfc_ctl = I915_READ(ILK_DPFC_CONTROL);
 232         if (dpfc_ctl & DPFC_CTL_EN) {
 233                 dpfc_ctl &= ~DPFC_CTL_EN;
 234                 I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl);
 235
 236                 if (IS_IVYBRIDGE(dev))
 237                         /* WaFbcDisableDpfcClockGating:ivb */
 238                         I915_WRITE(ILK_DSPCLK_GATE_D,
 239                                    I915_READ(ILK_DSPCLK_GATE_D) &
 240                                    ~ILK_DPFCUNIT_CLOCK_GATE_DISABLE);
 241
 242                 if (IS_HASWELL(dev))
 243                         /* WaFbcDisableDpfcClockGating:hsw */
 244                         I915_WRITE(HSW_CLKGATE_DISABLE_PART_1,
 245                                    I915_READ(HSW_CLKGATE_DISABLE_PART_1) &
 246                                    ~HSW_DPFC_GATING_DISABLE);
 247
 248                 DRM_DEBUG_KMS("disabled FBC\n");
 249         }
 250 }
 251
 252 static bool ironlake_fbc_enabled(struct drm_device *dev)
 253 {
 254         struct drm_i915_private *dev_priv = dev->dev_private;
 255
 256         return I915_READ(ILK_DPFC_CONTROL) & DPFC_CTL_EN;
 257 }
 258
 259 static void gen7_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 260 {
 261         struct drm_device *dev = crtc->dev;
 262         struct drm_i915_private *dev_priv = dev->dev_private;
 263         struct drm_framebuffer *fb = crtc->fb;
 264         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 265         struct drm_i915_gem_object *obj = intel_fb->obj;
 266         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 267
 268         I915_WRITE(IVB_FBC_RT_BASE, i915_gem_obj_ggtt_offset(obj));
 269
 270         I915_WRITE(ILK_DPFC_CONTROL, DPFC_CTL_EN | DPFC_CTL_LIMIT_1X |
 271                    IVB_DPFC_CTL_FENCE_EN |
 272                    intel_crtc->plane << IVB_DPFC_CTL_PLANE_SHIFT);
 273
 274         if (IS_IVYBRIDGE(dev)) {
 275                 /* WaFbcAsynchFlipDisableFbcQueue:ivb */
 276                 I915_WRITE(ILK_DISPLAY_CHICKEN1, ILK_FBCQ_DIS);
 277                 /* WaFbcDisableDpfcClockGating:ivb */
 278                 I915_WRITE(ILK_DSPCLK_GATE_D,
 279                            I915_READ(ILK_DSPCLK_GATE_D) |
 280                            ILK_DPFCUNIT_CLOCK_GATE_DISABLE);
 281         } else {
 282                 /* WaFbcAsynchFlipDisableFbcQueue:hsw */
 283                 I915_WRITE(HSW_PIPE_SLICE_CHICKEN_1(intel_crtc->pipe),
 284                            HSW_BYPASS_FBC_QUEUE);
 285                 /* WaFbcDisableDpfcClockGating:hsw */
 286                 I915_WRITE(HSW_CLKGATE_DISABLE_PART_1,
 287                            I915_READ(HSW_CLKGATE_DISABLE_PART_1) |
 288                            HSW_DPFC_GATING_DISABLE);
 289         }
 290
 291         I915_WRITE(SNB_DPFC_CTL_SA,
 292                    SNB_CPU_FENCE_ENABLE | obj->fence_reg);
 293         I915_WRITE(DPFC_CPU_FENCE_OFFSET, crtc->y);
 294
 295         sandybridge_blit_fbc_update(dev);
 296
 297         DRM_DEBUG_KMS("enabled fbc on plane %d\n", intel_crtc->plane);
 298 }
 299
 300 bool intel_fbc_enabled(struct drm_device *dev)
 301 {
 302         struct drm_i915_private *dev_priv = dev->dev_private;
 303
 304         if (!dev_priv->display.fbc_enabled)
 305                 return false;
 306
 307         return dev_priv->display.fbc_enabled(dev);
 308 }
 309
 310 static void intel_fbc_work_fn(struct work_struct *__work)
 311 {
 312         struct intel_fbc_work *work =
 313                 container_of(to_delayed_work(__work),
 314                              struct intel_fbc_work, work);
 315         struct drm_device *dev = work->crtc->dev;
 316         struct drm_i915_private *dev_priv = dev->dev_private;
 317
 318         mutex_lock(&dev->struct_mutex);
 319         if (work == dev_priv->fbc.fbc_work) {
 320                 /* Double check that we haven't switched fb without cancelling
 321                  * the prior work.
 322                  */
 323                 if (work->crtc->fb == work->fb) {
 324                         dev_priv->display.enable_fbc(work->crtc,
 325                                                      work->interval);
 326
 327                         dev_priv->fbc.plane = to_intel_crtc(work->crtc)->plane;
 328                         dev_priv->fbc.fb_id = work->crtc->fb->base.id;
 329                         dev_priv->fbc.y = work->crtc->y;
 330                 }
 331
 332                 dev_priv->fbc.fbc_work = NULL;
 333         }
 334         mutex_unlock(&dev->struct_mutex);
 335
 336         kfree(work);
 337 }
 338
 339 static void intel_cancel_fbc_work(struct drm_i915_private *dev_priv)
 340 {
 341         if (dev_priv->fbc.fbc_work == NULL)
 342                 return;
 343
 344         DRM_DEBUG_KMS("cancelling pending FBC enable\n");
 345
 346         /* Synchronisation is provided by struct_mutex and checking of
 347          * dev_priv->fbc.fbc_work, so we can perform the cancellation
 348          * entirely asynchronously.
 349          */
 350         if (cancel_delayed_work(&dev_priv->fbc.fbc_work->work))
 351                 /* tasklet was killed before being run, clean up */
 352                 kfree(dev_priv->fbc.fbc_work);
 353
 354         /* Mark the work as no longer wanted so that if it does
 355          * wake-up (because the work was already running and waiting
 356          * for our mutex), it will discover that is no longer
 357          * necessary to run.
 358          */
 359         dev_priv->fbc.fbc_work = NULL;
 360 }
 361
 362 static void intel_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 363 {
 364         struct intel_fbc_work *work;
 365         struct drm_device *dev = crtc->dev;
 366         struct drm_i915_private *dev_priv = dev->dev_private;
 367
 368         if (!dev_priv->display.enable_fbc)
 369                 return;
 370
 371         intel_cancel_fbc_work(dev_priv);
 372
 373         work = kzalloc(sizeof(*work), GFP_KERNEL);
 374         if (work == NULL) {
 375                 DRM_ERROR("Failed to allocate FBC work structure\n");
 376                 dev_priv->display.enable_fbc(crtc, interval);
 377                 return;
 378         }
 379
 380         work->crtc = crtc;
 381         work->fb = crtc->fb;
 382         work->interval = interval;
 383         INIT_DELAYED_WORK(&work->work, intel_fbc_work_fn);
 384
 385         dev_priv->fbc.fbc_work = work;
 386
 387         /* Delay the actual enabling to let pageflipping cease and the
 388          * display to settle before starting the compression. Note that
 389          * this delay also serves a second purpose: it allows for a
 390          * vblank to pass after disabling the FBC before we attempt
 391          * to modify the control registers.
 392          *
 393          * A more complicated solution would involve tracking vblanks
 394          * following the termination of the page-flipping sequence
 395          * and indeed performing the enable as a co-routine and not
 396          * waiting synchronously upon the vblank.
 397          *
 398          * WaFbcWaitForVBlankBeforeEnable:ilk,snb
 399          */
 400         schedule_delayed_work(&work->work, msecs_to_jiffies(50));
 401 }
 402
 403 void intel_disable_fbc(struct drm_device *dev)
 404 {
 405         struct drm_i915_private *dev_priv = dev->dev_private;
 406
 407         intel_cancel_fbc_work(dev_priv);
 408
 409         if (!dev_priv->display.disable_fbc)
 410                 return;
 411
 412         dev_priv->display.disable_fbc(dev);
 413         dev_priv->fbc.plane = -1;
 414 }
 415
 416 static bool set_no_fbc_reason(struct drm_i915_private *dev_priv,
 417                               enum no_fbc_reason reason)
 418 {
 419         if (dev_priv->fbc.no_fbc_reason == reason)
 420                 return false;
 421
 422         dev_priv->fbc.no_fbc_reason = reason;
 423         return true;
 424 }
 425
 426 /**
 427  * intel_update_fbc - enable/disable FBC as needed
 428  * @dev: the drm_device
 429  *
 430  * Set up the framebuffer compression hardware at mode set time.  We
 431  * enable it if possible:
 432  *   - plane A only (on pre-965)
 433  *   - no pixel mulitply/line duplication
 434  *   - no alpha buffer discard
 435  *   - no dual wide
 436  *   - framebuffer <= max_hdisplay in width, max_vdisplay in height
 437  *
 438  * We can't assume that any compression will take place (worst case),
 439  * so the compressed buffer has to be the same size as the uncompressed
 440  * one.  It also must reside (along with the line length buffer) in
 441  * stolen memory.
 442  *
 443  * We need to enable/disable FBC on a global basis.
 444  */
 445 void intel_update_fbc(struct drm_device *dev)
 446 {
 447         struct drm_i915_private *dev_priv = dev->dev_private;
 448         struct drm_crtc *crtc = NULL, *tmp_crtc;
 449         struct intel_crtc *intel_crtc;
 450         struct drm_framebuffer *fb;
 451         struct intel_framebuffer *intel_fb;
 452         struct drm_i915_gem_object *obj;
 453         const struct drm_display_mode *adjusted_mode;
 454         unsigned int max_width, max_height;
 455
 456         if (!I915_HAS_FBC(dev)) {
 457                 set_no_fbc_reason(dev_priv, FBC_UNSUPPORTED);
 458                 return;
 459         }
 460
 461         if (!i915_powersave) {
 462                 if (set_no_fbc_reason(dev_priv, FBC_MODULE_PARAM))
 463                         DRM_DEBUG_KMS("fbc disabled per module param\n");
 464                 return;
 465         }
 466
 467         /*
 468          * If FBC is already on, we just have to verify that we can
 469          * keep it that way...
 470          * Need to disable if:
 471          *   - more than one pipe is active
 472          *   - changing FBC params (stride, fence, mode)
 473          *   - new fb is too large to fit in compressed buffer
 474          *   - going to an unsupported config (interlace, pixel multiply, etc.)
 475          */
 476         list_for_each_entry(tmp_crtc, &dev->mode_config.crtc_list, head) {
 477                 if (intel_crtc_active(tmp_crtc) &&
 478                     !to_intel_crtc(tmp_crtc)->primary_disabled) {
 479                         if (crtc) {
 480                                 if (set_no_fbc_reason(dev_priv, FBC_MULTIPLE_PIPES))
 481                                         DRM_DEBUG_KMS("more than one pipe active, disabling compression\n");
 482                                 goto out_disable;
 483                         }
 484                         crtc = tmp_crtc;
 485                 }
 486         }
 487
 488         if (!crtc || crtc->fb == NULL) {
 489                 if (set_no_fbc_reason(dev_priv, FBC_NO_OUTPUT))
 490                         DRM_DEBUG_KMS("no output, disabling\n");
 491                 goto out_disable;
 492         }
 493
 494         intel_crtc = to_intel_crtc(crtc);
 495         fb = crtc->fb;
 496         intel_fb = to_intel_framebuffer(fb);
 497         obj = intel_fb->obj;
 498         adjusted_mode = &intel_crtc->config.adjusted_mode;
 499
 500         if (i915_enable_fbc < 0 &&
 501             INTEL_INFO(dev)->gen <= 7 && !IS_HASWELL(dev)) {
 502                 if (set_no_fbc_reason(dev_priv, FBC_CHIP_DEFAULT))
 503                         DRM_DEBUG_KMS("disabled per chip default\n");
 504                 goto out_disable;
 505         }
 506         if (!i915_enable_fbc) {
 507                 if (set_no_fbc_reason(dev_priv, FBC_MODULE_PARAM))
 508                         DRM_DEBUG_KMS("fbc disabled per module param\n");
 509                 goto out_disable;
 510         }
 511         if ((adjusted_mode->flags & DRM_MODE_FLAG_INTERLACE) ||
 512             (adjusted_mode->flags & DRM_MODE_FLAG_DBLSCAN)) {
 513                 if (set_no_fbc_reason(dev_priv, FBC_UNSUPPORTED_MODE))
 514                         DRM_DEBUG_KMS("mode incompatible with compression, "
 515                                       "disabling\n");
 516                 goto out_disable;
 517         }
 518
 519         if (IS_G4X(dev) || INTEL_INFO(dev)->gen >= 5) {
 520                 max_width = 4096;
 521                 max_height = 2048;
 522         } else {
 523                 max_width = 2048;
 524                 max_height = 1536;
 525         }
 526         if (intel_crtc->config.pipe_src_w > max_width ||
 527             intel_crtc->config.pipe_src_h > max_height) {
 528                 if (set_no_fbc_reason(dev_priv, FBC_MODE_TOO_LARGE))
 529                         DRM_DEBUG_KMS("mode too large for compression, disabling\n");
 530                 goto out_disable;
 531         }
 532         if ((IS_I915GM(dev) || IS_I945GM(dev) || IS_HASWELL(dev)) &&
 533             intel_crtc->plane != 0) {
 534                 if (set_no_fbc_reason(dev_priv, FBC_BAD_PLANE))
 535                         DRM_DEBUG_KMS("plane not 0, disabling compression\n");
 536                 goto out_disable;
 537         }
 538
 539         /* The use of a CPU fence is mandatory in order to detect writes
 540          * by the CPU to the scanout and trigger updates to the FBC.
 541          */
 542         if (obj->tiling_mode != I915_TILING_X ||
 543             obj->fence_reg == I915_FENCE_REG_NONE) {
 544                 if (set_no_fbc_reason(dev_priv, FBC_NOT_TILED))
 545                         DRM_DEBUG_KMS("framebuffer not tiled or fenced, disabling compression\n");
 546                 goto out_disable;
 547         }
 548
 549         /* If the kernel debugger is active, always disable compression */
 550         if (in_dbg_master())
 551                 goto out_disable;
 552
 553         if (i915_gem_stolen_setup_compression(dev, intel_fb->obj->base.size)) {
 554                 if (set_no_fbc_reason(dev_priv, FBC_STOLEN_TOO_SMALL))
 555                         DRM_DEBUG_KMS("framebuffer too large, disabling compression\n");
 556                 goto out_disable;
 557         }
 558
 559         /* If the scanout has not changed, don't modify the FBC settings.
 560          * Note that we make the fundamental assumption that the fb->obj
 561          * cannot be unpinned (and have its GTT offset and fence revoked)
 562          * without first being decoupled from the scanout and FBC disabled.
 563          */
 564         if (dev_priv->fbc.plane == intel_crtc->plane &&
 565             dev_priv->fbc.fb_id == fb->base.id &&
 566             dev_priv->fbc.y == crtc->y)
 567                 return;
 568
 569         if (intel_fbc_enabled(dev)) {
 570                 /* We update FBC along two paths, after changing fb/crtc
 571                  * configuration (modeswitching) and after page-flipping
 572                  * finishes. For the latter, we know that not only did
 573                  * we disable the FBC at the start of the page-flip
 574                  * sequence, but also more than one vblank has passed.
 575                  *
 576                  * For the former case of modeswitching, it is possible
 577                  * to switch between two FBC valid configurations
 578                  * instantaneously so we do need to disable the FBC
 579                  * before we can modify its control registers. We also
 580                  * have to wait for the next vblank for that to take
 581                  * effect. However, since we delay enabling FBC we can
 582                  * assume that a vblank has passed since disabling and
 583                  * that we can safely alter the registers in the deferred
 584                  * callback.
 585                  *
 586                  * In the scenario that we go from a valid to invalid
 587                  * and then back to valid FBC configuration we have
 588                  * no strict enforcement that a vblank occurred since
 589                  * disabling the FBC. However, along all current pipe
 590                  * disabling paths we do need to wait for a vblank at
 591                  * some point. And we wait before enabling FBC anyway.
 592                  */
 593                 DRM_DEBUG_KMS("disabling active FBC for update\n");
 594                 intel_disable_fbc(dev);
 595         }
 596
 597         intel_enable_fbc(crtc, 500);
 598         dev_priv->fbc.no_fbc_reason = FBC_OK;
 599         return;
 600
 601 out_disable:
 602         /* Multiple disables should be harmless */
 603         if (intel_fbc_enabled(dev)) {
 604                 DRM_DEBUG_KMS("unsupported config, disabling FBC\n");
 605                 intel_disable_fbc(dev);
 606         }
 607         i915_gem_stolen_cleanup_compression(dev);
 608 }
 609
 610 static void i915_pineview_get_mem_freq(struct drm_device *dev)
 611 {
 612         drm_i915_private_t *dev_priv = dev->dev_private;
 613         u32 tmp;
 614
 615         tmp = I915_READ(CLKCFG);
 616
 617         switch (tmp & CLKCFG_FSB_MASK) {
 618         case CLKCFG_FSB_533:
 619                 dev_priv->fsb_freq = 533; /* 133*4 */
 620                 break;
 621         case CLKCFG_FSB_800:
 622                 dev_priv->fsb_freq = 800; /* 200*4 */
 623                 break;
 624         case CLKCFG_FSB_667:
 625                 dev_priv->fsb_freq =  667; /* 167*4 */
 626                 break;
 627         case CLKCFG_FSB_400:
 628                 dev_priv->fsb_freq = 400; /* 100*4 */
 629                 break;
 630         }
 631
 632         switch (tmp & CLKCFG_MEM_MASK) {
 633         case CLKCFG_MEM_533:
 634                 dev_priv->mem_freq = 533;
 635                 break;
 636         case CLKCFG_MEM_667:
 637                 dev_priv->mem_freq = 667;
 638                 break;
 639         case CLKCFG_MEM_800:
 640                 dev_priv->mem_freq = 800;
 641                 break;
 642         }
 643
 644         /* detect pineview DDR3 setting */
 645         tmp = I915_READ(CSHRDDR3CTL);
 646         dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
 647 }
 648
 649 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
 650 {
 651         drm_i915_private_t *dev_priv = dev->dev_private;
 652         u16 ddrpll, csipll;
 653
 654         ddrpll = I915_READ16(DDRMPLL1);
 655         csipll = I915_READ16(CSIPLL0);
 656
 657         switch (ddrpll & 0xff) {
 658         case 0xc:
 659                 dev_priv->mem_freq = 800;
 660                 break;
 661         case 0x10:
 662                 dev_priv->mem_freq = 1066;
 663                 break;
 664         case 0x14:
 665                 dev_priv->mem_freq = 1333;
 666                 break;
 667         case 0x18:
 668                 dev_priv->mem_freq = 1600;
 669                 break;
 670         default:
 671                 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
 672                                  ddrpll & 0xff);
 673                 dev_priv->mem_freq = 0;
 674                 break;
 675         }
 676
 677         dev_priv->ips.r_t = dev_priv->mem_freq;
 678
 679         switch (csipll & 0x3ff) {
 680         case 0x00c:
 681                 dev_priv->fsb_freq = 3200;
 682                 break;
 683         case 0x00e:
 684                 dev_priv->fsb_freq = 3733;
 685                 break;
 686         case 0x010:
 687                 dev_priv->fsb_freq = 4266;
 688                 break;
 689         case 0x012:
 690                 dev_priv->fsb_freq = 4800;
 691                 break;
 692         case 0x014:
 693                 dev_priv->fsb_freq = 5333;
 694                 break;
 695         case 0x016:
 696                 dev_priv->fsb_freq = 5866;
 697                 break;
 698         case 0x018:
 699                 dev_priv->fsb_freq = 6400;
 700                 break;
 701         default:
 702                 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
 703                                  csipll & 0x3ff);
 704                 dev_priv->fsb_freq = 0;
 705                 break;
 706         }
 707
 708         if (dev_priv->fsb_freq == 3200) {
 709                 dev_priv->ips.c_m = 0;
 710         } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
 711                 dev_priv->ips.c_m = 1;
 712         } else {
 713                 dev_priv->ips.c_m = 2;
 714         }
 715 }
 716
 717 static const struct cxsr_latency cxsr_latency_table[] = {
 718         {1, 0, 800, 400, 3382, 33382, 3983, 33983},    /* DDR2-400 SC */
 719         {1, 0, 800, 667, 3354, 33354, 3807, 33807},    /* DDR2-667 SC */
 720         {1, 0, 800, 800, 3347, 33347, 3763, 33763},    /* DDR2-800 SC */
 721         {1, 1, 800, 667, 6420, 36420, 6873, 36873},    /* DDR3-667 SC */
 722         {1, 1, 800, 800, 5902, 35902, 6318, 36318},    /* DDR3-800 SC */
 723
 724         {1, 0, 667, 400, 3400, 33400, 4021, 34021},    /* DDR2-400 SC */
 725         {1, 0, 667, 667, 3372, 33372, 3845, 33845},    /* DDR2-667 SC */
 726         {1, 0, 667, 800, 3386, 33386, 3822, 33822},    /* DDR2-800 SC */
 727         {1, 1, 667, 667, 6438, 36438, 6911, 36911},    /* DDR3-667 SC */
 728         {1, 1, 667, 800, 5941, 35941, 6377, 36377},    /* DDR3-800 SC */
 729
 730         {1, 0, 400, 400, 3472, 33472, 4173, 34173},    /* DDR2-400 SC */
 731         {1, 0, 400, 667, 3443, 33443, 3996, 33996},    /* DDR2-667 SC */
 732         {1, 0, 400, 800, 3430, 33430, 3946, 33946},    /* DDR2-800 SC */
 733         {1, 1, 400, 667, 6509, 36509, 7062, 37062},    /* DDR3-667 SC */
 734         {1, 1, 400, 800, 5985, 35985, 6501, 36501},    /* DDR3-800 SC */
 735
 736         {0, 0, 800, 400, 3438, 33438, 4065, 34065},    /* DDR2-400 SC */
 737         {0, 0, 800, 667, 3410, 33410, 3889, 33889},    /* DDR2-667 SC */
 738         {0, 0, 800, 800, 3403, 33403, 3845, 33845},    /* DDR2-800 SC */
 739         {0, 1, 800, 667, 6476, 36476, 6955, 36955},    /* DDR3-667 SC */
 740         {0, 1, 800, 800, 5958, 35958, 6400, 36400},    /* DDR3-800 SC */
 741
 742         {0, 0, 667, 400, 3456, 33456, 4103, 34106},    /* DDR2-400 SC */
 743         {0, 0, 667, 667, 3428, 33428, 3927, 33927},    /* DDR2-667 SC */
 744         {0, 0, 667, 800, 3443, 33443, 3905, 33905},    /* DDR2-800 SC */
 745         {0, 1, 667, 667, 6494, 36494, 6993, 36993},    /* DDR3-667 SC */
 746         {0, 1, 667, 800, 5998, 35998, 6460, 36460},    /* DDR3-800 SC */
 747
 748         {0, 0, 400, 400, 3528, 33528, 4255, 34255},    /* DDR2-400 SC */
 749         {0, 0, 400, 667, 3500, 33500, 4079, 34079},    /* DDR2-667 SC */
 750         {0, 0, 400, 800, 3487, 33487, 4029, 34029},    /* DDR2-800 SC */
 751         {0, 1, 400, 667, 6566, 36566, 7145, 37145},    /* DDR3-667 SC */
 752         {0, 1, 400, 800, 6042, 36042, 6584, 36584},    /* DDR3-800 SC */
 753 };
 754
 755 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
 756                                                          int is_ddr3,
 757                                                          int fsb,
 758                                                          int mem)
 759 {
 760         const struct cxsr_latency *latency;
 761         int i;
 762
 763         if (fsb == 0 || mem == 0)
 764                 return NULL;
 765
 766         for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
 767                 latency = &cxsr_latency_table[i];
 768                 if (is_desktop == latency->is_desktop &&
 769                     is_ddr3 == latency->is_ddr3 &&
 770                     fsb == latency->fsb_freq && mem == latency->mem_freq)
 771                         return latency;
 772         }
 773
 774         DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 775
 776         return NULL;
 777 }
 778
 779 static void pineview_disable_cxsr(struct drm_device *dev)
 780 {
 781         struct drm_i915_private *dev_priv = dev->dev_private;
 782
 783         /* deactivate cxsr */
 784         I915_WRITE(DSPFW3, I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN);
 785 }
 786
 787 /*
 788  * Latency for FIFO fetches is dependent on several factors:
 789  *   - memory configuration (speed, channels)
 790  *   - chipset
 791  *   - current MCH state
 792  * It can be fairly high in some situations, so here we assume a fairly
 793  * pessimal value.  It's a tradeoff between extra memory fetches (if we
 794  * set this value too high, the FIFO will fetch frequently to stay full)
 795  * and power consumption (set it too low to save power and we might see
 796  * FIFO underruns and display "flicker").
 797  *
 798  * A value of 5us seems to be a good balance; safe for very low end
 799  * platforms but not overly aggressive on lower latency configs.
 800  */
 801 static const int latency_ns = 5000;
 802
 803 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
 804 {
 805         struct drm_i915_private *dev_priv = dev->dev_private;
 806         uint32_t dsparb = I915_READ(DSPARB);
 807         int size;
 808
 809         size = dsparb & 0x7f;
 810         if (plane)
 811                 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
 812
 813         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 814                       plane ? "B" : "A", size);
 815
 816         return size;
 817 }
 818
 819 static int i85x_get_fifo_size(struct drm_device *dev, int plane)
 820 {
 821         struct drm_i915_private *dev_priv = dev->dev_private;
 822         uint32_t dsparb = I915_READ(DSPARB);
 823         int size;
 824
 825         size = dsparb & 0x1ff;
 826         if (plane)
 827                 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
 828         size >>= 1; /* Convert to cachelines */
 829
 830         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 831                       plane ? "B" : "A", size);
 832
 833         return size;
 834 }
 835
 836 static int i845_get_fifo_size(struct drm_device *dev, int plane)
 837 {
 838         struct drm_i915_private *dev_priv = dev->dev_private;
 839         uint32_t dsparb = I915_READ(DSPARB);
 840         int size;
 841
 842         size = dsparb & 0x7f;
 843         size >>= 2; /* Convert to cachelines */
 844
 845         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 846                       plane ? "B" : "A",
 847                       size);
 848
 849         return size;
 850 }
 851
 852 static int i830_get_fifo_size(struct drm_device *dev, int plane)
 853 {
 854         struct drm_i915_private *dev_priv = dev->dev_private;
 855         uint32_t dsparb = I915_READ(DSPARB);
 856         int size;
 857
 858         size = dsparb & 0x7f;
 859         size >>= 1; /* Convert to cachelines */
 860
 861         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 862                       plane ? "B" : "A", size);
 863
 864         return size;
 865 }
 866
 867 /* Pineview has different values for various configs */
 868 static const struct intel_watermark_params pineview_display_wm = {
 869         PINEVIEW_DISPLAY_FIFO,
 870         PINEVIEW_MAX_WM,
 871         PINEVIEW_DFT_WM,
 872         PINEVIEW_GUARD_WM,
 873         PINEVIEW_FIFO_LINE_SIZE
 874 };
 875 static const struct intel_watermark_params pineview_display_hplloff_wm = {
 876         PINEVIEW_DISPLAY_FIFO,
 877         PINEVIEW_MAX_WM,
 878         PINEVIEW_DFT_HPLLOFF_WM,
 879         PINEVIEW_GUARD_WM,
 880         PINEVIEW_FIFO_LINE_SIZE
 881 };
 882 static const struct intel_watermark_params pineview_cursor_wm = {
 883         PINEVIEW_CURSOR_FIFO,
 884         PINEVIEW_CURSOR_MAX_WM,
 885         PINEVIEW_CURSOR_DFT_WM,
 886         PINEVIEW_CURSOR_GUARD_WM,
 887         PINEVIEW_FIFO_LINE_SIZE,
 888 };
 889 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
 890         PINEVIEW_CURSOR_FIFO,
 891         PINEVIEW_CURSOR_MAX_WM,
 892         PINEVIEW_CURSOR_DFT_WM,
 893         PINEVIEW_CURSOR_GUARD_WM,
 894         PINEVIEW_FIFO_LINE_SIZE
 895 };
 896 static const struct intel_watermark_params g4x_wm_info = {
 897         G4X_FIFO_SIZE,
 898         G4X_MAX_WM,
 899         G4X_MAX_WM,
 900         2,
 901         G4X_FIFO_LINE_SIZE,
 902 };
 903 static const struct intel_watermark_params g4x_cursor_wm_info = {
 904         I965_CURSOR_FIFO,
 905         I965_CURSOR_MAX_WM,
 906         I965_CURSOR_DFT_WM,
 907         2,
 908         G4X_FIFO_LINE_SIZE,
 909 };
 910 static const struct intel_watermark_params valleyview_wm_info = {
 911         VALLEYVIEW_FIFO_SIZE,
 912         VALLEYVIEW_MAX_WM,
 913         VALLEYVIEW_MAX_WM,
 914         2,
 915         G4X_FIFO_LINE_SIZE,
 916 };
 917 static const struct intel_watermark_params valleyview_cursor_wm_info = {
 918         I965_CURSOR_FIFO,
 919         VALLEYVIEW_CURSOR_MAX_WM,
 920         I965_CURSOR_DFT_WM,
 921         2,
 922         G4X_FIFO_LINE_SIZE,
 923 };
 924 static const struct intel_watermark_params i965_cursor_wm_info = {
 925         I965_CURSOR_FIFO,
 926         I965_CURSOR_MAX_WM,
 927         I965_CURSOR_DFT_WM,
 928         2,
 929         I915_FIFO_LINE_SIZE,
 930 };
 931 static const struct intel_watermark_params i945_wm_info = {
 932         I945_FIFO_SIZE,
 933         I915_MAX_WM,
 934         1,
 935         2,
 936         I915_FIFO_LINE_SIZE
 937 };
 938 static const struct intel_watermark_params i915_wm_info = {
 939         I915_FIFO_SIZE,
 940         I915_MAX_WM,
 941         1,
 942         2,
 943         I915_FIFO_LINE_SIZE
 944 };
 945 static const struct intel_watermark_params i855_wm_info = {
 946         I855GM_FIFO_SIZE,
 947         I915_MAX_WM,
 948         1,
 949         2,
 950         I830_FIFO_LINE_SIZE
 951 };
 952 static const struct intel_watermark_params i830_wm_info = {
 953         I830_FIFO_SIZE,
 954         I915_MAX_WM,
 955         1,
 956         2,
 957         I830_FIFO_LINE_SIZE
 958 };
 959
 960 static const struct intel_watermark_params ironlake_display_wm_info = {
 961         ILK_DISPLAY_FIFO,
 962         ILK_DISPLAY_MAXWM,
 963         ILK_DISPLAY_DFTWM,
 964         2,
 965         ILK_FIFO_LINE_SIZE
 966 };
 967 static const struct intel_watermark_params ironlake_cursor_wm_info = {
 968         ILK_CURSOR_FIFO,
 969         ILK_CURSOR_MAXWM,
 970         ILK_CURSOR_DFTWM,
 971         2,
 972         ILK_FIFO_LINE_SIZE
 973 };
 974 static const struct intel_watermark_params ironlake_display_srwm_info = {
 975         ILK_DISPLAY_SR_FIFO,
 976         ILK_DISPLAY_MAX_SRWM,
 977         ILK_DISPLAY_DFT_SRWM,
 978         2,
 979         ILK_FIFO_LINE_SIZE
 980 };
 981 static const struct intel_watermark_params ironlake_cursor_srwm_info = {
 982         ILK_CURSOR_SR_FIFO,
 983         ILK_CURSOR_MAX_SRWM,
 984         ILK_CURSOR_DFT_SRWM,
 985         2,
 986         ILK_FIFO_LINE_SIZE
 987 };
 988
 989 static const struct intel_watermark_params sandybridge_display_wm_info = {
 990         SNB_DISPLAY_FIFO,
 991         SNB_DISPLAY_MAXWM,
 992         SNB_DISPLAY_DFTWM,
 993         2,
 994         SNB_FIFO_LINE_SIZE
 995 };
 996 static const struct intel_watermark_params sandybridge_cursor_wm_info = {
 997         SNB_CURSOR_FIFO,
 998         SNB_CURSOR_MAXWM,
 999         SNB_CURSOR_DFTWM,
1000         2,
1001         SNB_FIFO_LINE_SIZE
1002 };
1003 static const struct intel_watermark_params sandybridge_display_srwm_info = {
1004         SNB_DISPLAY_SR_FIFO,
1005         SNB_DISPLAY_MAX_SRWM,
1006         SNB_DISPLAY_DFT_SRWM,
1007         2,
1008         SNB_FIFO_LINE_SIZE
1009 };
1010 static const struct intel_watermark_params sandybridge_cursor_srwm_info = {
1011         SNB_CURSOR_SR_FIFO,
1012         SNB_CURSOR_MAX_SRWM,
1013         SNB_CURSOR_DFT_SRWM,
1014         2,
1015         SNB_FIFO_LINE_SIZE
1016 };
1017
1018
1019 /**
1020  * intel_calculate_wm - calculate watermark level
1021  * @clock_in_khz: pixel clock
1022  * @wm: chip FIFO params
1023  * @pixel_size: display pixel size
1024  * @latency_ns: memory latency for the platform
1025  *
1026  * Calculate the watermark level (the level at which the display plane will
1027  * start fetching from memory again).  Each chip has a different display
1028  * FIFO size and allocation, so the caller needs to figure that out and pass
1029  * in the correct intel_watermark_params structure.
1030  *
1031  * As the pixel clock runs, the FIFO will be drained at a rate that depends
1032  * on the pixel size.  When it reaches the watermark level, it'll start
1033  * fetching FIFO line sized based chunks from memory until the FIFO fills
1034  * past the watermark point.  If the FIFO drains completely, a FIFO underrun
1035  * will occur, and a display engine hang could result.
1036  */
1037 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
1038                                         const struct intel_watermark_params *wm,
1039                                         int fifo_size,
1040                                         int pixel_size,
1041                                         unsigned long latency_ns)
1042 {
1043         long entries_required, wm_size;
1044
1045         /*
1046          * Note: we need to make sure we don't overflow for various clock &
1047          * latency values.
1048          * clocks go from a few thousand to several hundred thousand.
1049          * latency is usually a few thousand
1050          */
1051         entries_required = ((clock_in_khz / 1000) * pixel_size * latency_ns) /
1052                 1000;
1053         entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
1054
1055         DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
1056
1057         wm_size = fifo_size - (entries_required + wm->guard_size);
1058
1059         DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
1060
1061         /* Don't promote wm_size to unsigned... */
1062         if (wm_size > (long)wm->max_wm)
1063                 wm_size = wm->max_wm;
1064         if (wm_size <= 0)
1065                 wm_size = wm->default_wm;
1066         return wm_size;
1067 }
1068
1069 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
1070 {
1071         struct drm_crtc *crtc, *enabled = NULL;
1072
1073         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
1074                 if (intel_crtc_active(crtc)) {
1075                         if (enabled)
1076                                 return NULL;
1077                         enabled = crtc;
1078                 }
1079         }
1080
1081         return enabled;
1082 }
1083
1084 static void pineview_update_wm(struct drm_crtc *unused_crtc)
1085 {
1086         struct drm_device *dev = unused_crtc->dev;
1087         struct drm_i915_private *dev_priv = dev->dev_private;
1088         struct drm_crtc *crtc;
1089         const struct cxsr_latency *latency;
1090         u32 reg;
1091         unsigned long wm;
1092
1093         latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
1094                                          dev_priv->fsb_freq, dev_priv->mem_freq);
1095         if (!latency) {
1096                 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
1097                 pineview_disable_cxsr(dev);
1098                 return;
1099         }
1100
1101         crtc = single_enabled_crtc(dev);
1102         if (crtc) {
1103                 const struct drm_display_mode *adjusted_mode;
1104                 int pixel_size = crtc->fb->bits_per_pixel / 8;
1105                 int clock;
1106
1107                 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1108                 clock = adjusted_mode->crtc_clock;
1109
1110                 /* Display SR */
1111                 wm = intel_calculate_wm(clock, &pineview_display_wm,
1112                                         pineview_display_wm.fifo_size,
1113                                         pixel_size, latency->display_sr);
1114                 reg = I915_READ(DSPFW1);
1115                 reg &= ~DSPFW_SR_MASK;
1116                 reg |= wm << DSPFW_SR_SHIFT;
1117                 I915_WRITE(DSPFW1, reg);
1118                 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
1119
1120                 /* cursor SR */
1121                 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
1122                                         pineview_display_wm.fifo_size,
1123                                         pixel_size, latency->cursor_sr);
1124                 reg = I915_READ(DSPFW3);
1125                 reg &= ~DSPFW_CURSOR_SR_MASK;
1126                 reg |= (wm & 0x3f) << DSPFW_CURSOR_SR_SHIFT;
1127                 I915_WRITE(DSPFW3, reg);
1128
1129                 /* Display HPLL off SR */
1130                 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
1131                                         pineview_display_hplloff_wm.fifo_size,
1132                                         pixel_size, latency->display_hpll_disable);
1133                 reg = I915_READ(DSPFW3);
1134                 reg &= ~DSPFW_HPLL_SR_MASK;
1135                 reg |= wm & DSPFW_HPLL_SR_MASK;
1136                 I915_WRITE(DSPFW3, reg);
1137
1138                 /* cursor HPLL off SR */
1139                 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
1140                                         pineview_display_hplloff_wm.fifo_size,
1141                                         pixel_size, latency->cursor_hpll_disable);
1142                 reg = I915_READ(DSPFW3);
1143                 reg &= ~DSPFW_HPLL_CURSOR_MASK;
1144                 reg |= (wm & 0x3f) << DSPFW_HPLL_CURSOR_SHIFT;
1145                 I915_WRITE(DSPFW3, reg);
1146                 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
1147
1148                 /* activate cxsr */
1149                 I915_WRITE(DSPFW3,
1150                            I915_READ(DSPFW3) | PINEVIEW_SELF_REFRESH_EN);
1151                 DRM_DEBUG_KMS("Self-refresh is enabled\n");
1152         } else {
1153                 pineview_disable_cxsr(dev);
1154                 DRM_DEBUG_KMS("Self-refresh is disabled\n");
1155         }
1156 }
1157
1158 static bool g4x_compute_wm0(struct drm_device *dev,
1159                             int plane,
1160                             const struct intel_watermark_params *display,
1161                             int display_latency_ns,
1162                             const struct intel_watermark_params *cursor,
1163                             int cursor_latency_ns,
1164                             int *plane_wm,
1165                             int *cursor_wm)
1166 {
1167         struct drm_crtc *crtc;
1168         const struct drm_display_mode *adjusted_mode;
1169         int htotal, hdisplay, clock, pixel_size;
1170         int line_time_us, line_count;
1171         int entries, tlb_miss;
1172
1173         crtc = intel_get_crtc_for_plane(dev, plane);
1174         if (!intel_crtc_active(crtc)) {
1175                 *cursor_wm = cursor->guard_size;
1176                 *plane_wm = display->guard_size;
1177                 return false;
1178         }
1179
1180         adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1181         clock = adjusted_mode->crtc_clock;
1182         htotal = adjusted_mode->htotal;
1183         hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1184         pixel_size = crtc->fb->bits_per_pixel / 8;
1185
1186         /* Use the small buffer method to calculate plane watermark */
1187         entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
1188         tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
1189         if (tlb_miss > 0)
1190                 entries += tlb_miss;
1191         entries = DIV_ROUND_UP(entries, display->cacheline_size);
1192         *plane_wm = entries + display->guard_size;
1193         if (*plane_wm > (int)display->max_wm)
1194                 *plane_wm = display->max_wm;
1195
1196         /* Use the large buffer method to calculate cursor watermark */
1197         line_time_us = ((htotal * 1000) / clock);
1198         line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
1199         entries = line_count * 64 * pixel_size;
1200         tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
1201         if (tlb_miss > 0)
1202                 entries += tlb_miss;
1203         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1204         *cursor_wm = entries + cursor->guard_size;
1205         if (*cursor_wm > (int)cursor->max_wm)
1206                 *cursor_wm = (int)cursor->max_wm;
1207
1208         return true;
1209 }
1210
1211 /*
1212  * Check the wm result.
1213  *
1214  * If any calculated watermark values is larger than the maximum value that
1215  * can be programmed into the associated watermark register, that watermark
1216  * must be disabled.
1217  */
1218 static bool g4x_check_srwm(struct drm_device *dev,
1219                            int display_wm, int cursor_wm,
1220                            const struct intel_watermark_params *display,
1221                            const struct intel_watermark_params *cursor)
1222 {
1223         DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
1224                       display_wm, cursor_wm);
1225
1226         if (display_wm > display->max_wm) {
1227                 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
1228                               display_wm, display->max_wm);
1229                 return false;
1230         }
1231
1232         if (cursor_wm > cursor->max_wm) {
1233                 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
1234                               cursor_wm, cursor->max_wm);
1235                 return false;
1236         }
1237
1238         if (!(display_wm || cursor_wm)) {
1239                 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
1240                 return false;
1241         }
1242
1243         return true;
1244 }
1245
1246 static bool g4x_compute_srwm(struct drm_device *dev,
1247                              int plane,
1248                              int latency_ns,
1249                              const struct intel_watermark_params *display,
1250                              const struct intel_watermark_params *cursor,
1251                              int *display_wm, int *cursor_wm)
1252 {
1253         struct drm_crtc *crtc;
1254         const struct drm_display_mode *adjusted_mode;
1255         int hdisplay, htotal, pixel_size, clock;
1256         unsigned long line_time_us;
1257         int line_count, line_size;
1258         int small, large;
1259         int entries;
1260
1261         if (!latency_ns) {
1262                 *display_wm = *cursor_wm = 0;
1263                 return false;
1264         }
1265
1266         crtc = intel_get_crtc_for_plane(dev, plane);
1267         adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1268         clock = adjusted_mode->crtc_clock;
1269         htotal = adjusted_mode->htotal;
1270         hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1271         pixel_size = crtc->fb->bits_per_pixel / 8;
1272
1273         line_time_us = (htotal * 1000) / clock;
1274         line_count = (latency_ns / line_time_us + 1000) / 1000;
1275         line_size = hdisplay * pixel_size;
1276
1277         /* Use the minimum of the small and large buffer method for primary */
1278         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1279         large = line_count * line_size;
1280
1281         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1282         *display_wm = entries + display->guard_size;
1283
1284         /* calculate the self-refresh watermark for display cursor */
1285         entries = line_count * pixel_size * 64;
1286         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1287         *cursor_wm = entries + cursor->guard_size;
1288
1289         return g4x_check_srwm(dev,
1290                               *display_wm, *cursor_wm,
1291                               display, cursor);
1292 }
1293
1294 static bool vlv_compute_drain_latency(struct drm_device *dev,
1295                                      int plane,
1296                                      int *plane_prec_mult,
1297                                      int *plane_dl,
1298                                      int *cursor_prec_mult,
1299                                      int *cursor_dl)
1300 {
1301         struct drm_crtc *crtc;
1302         int clock, pixel_size;
1303         int entries;
1304
1305         crtc = intel_get_crtc_for_plane(dev, plane);
1306         if (!intel_crtc_active(crtc))
1307                 return false;
1308
1309         clock = to_intel_crtc(crtc)->config.adjusted_mode.crtc_clock;
1310         pixel_size = crtc->fb->bits_per_pixel / 8;      /* BPP */
1311
1312         entries = (clock / 1000) * pixel_size;
1313         *plane_prec_mult = (entries > 256) ?
1314                 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1315         *plane_dl = (64 * (*plane_prec_mult) * 4) / ((clock / 1000) *
1316                                                      pixel_size);
1317
1318         entries = (clock / 1000) * 4;   /* BPP is always 4 for cursor */
1319         *cursor_prec_mult = (entries > 256) ?
1320                 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1321         *cursor_dl = (64 * (*cursor_prec_mult) * 4) / ((clock / 1000) * 4);
1322
1323         return true;
1324 }
1325
1326 /*
1327  * Update drain latency registers of memory arbiter
1328  *
1329  * Valleyview SoC has a new memory arbiter and needs drain latency registers
1330  * to be programmed. Each plane has a drain latency multiplier and a drain
1331  * latency value.
1332  */
1333
1334 static void vlv_update_drain_latency(struct drm_device *dev)
1335 {
1336         struct drm_i915_private *dev_priv = dev->dev_private;
1337         int planea_prec, planea_dl, planeb_prec, planeb_dl;
1338         int cursora_prec, cursora_dl, cursorb_prec, cursorb_dl;
1339         int plane_prec_mult, cursor_prec_mult; /* Precision multiplier is
1340                                                         either 16 or 32 */
1341
1342         /* For plane A, Cursor A */
1343         if (vlv_compute_drain_latency(dev, 0, &plane_prec_mult, &planea_dl,
1344                                       &cursor_prec_mult, &cursora_dl)) {
1345                 cursora_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1346                         DDL_CURSORA_PRECISION_32 : DDL_CURSORA_PRECISION_16;
1347                 planea_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1348                         DDL_PLANEA_PRECISION_32 : DDL_PLANEA_PRECISION_16;
1349
1350                 I915_WRITE(VLV_DDL1, cursora_prec |
1351                                 (cursora_dl << DDL_CURSORA_SHIFT) |
1352                                 planea_prec | planea_dl);
1353         }
1354
1355         /* For plane B, Cursor B */
1356         if (vlv_compute_drain_latency(dev, 1, &plane_prec_mult, &planeb_dl,
1357                                       &cursor_prec_mult, &cursorb_dl)) {
1358                 cursorb_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1359                         DDL_CURSORB_PRECISION_32 : DDL_CURSORB_PRECISION_16;
1360                 planeb_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1361                         DDL_PLANEB_PRECISION_32 : DDL_PLANEB_PRECISION_16;
1362
1363                 I915_WRITE(VLV_DDL2, cursorb_prec |
1364                                 (cursorb_dl << DDL_CURSORB_SHIFT) |
1365                                 planeb_prec | planeb_dl);
1366         }
1367 }
1368
1369 #define single_plane_enabled(mask) is_power_of_2(mask)
1370
1371 static void valleyview_update_wm(struct drm_crtc *crtc)
1372 {
1373         struct drm_device *dev = crtc->dev;
1374         static const int sr_latency_ns = 12000;
1375         struct drm_i915_private *dev_priv = dev->dev_private;
1376         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1377         int plane_sr, cursor_sr;
1378         int ignore_plane_sr, ignore_cursor_sr;
1379         unsigned int enabled = 0;
1380
1381         vlv_update_drain_latency(dev);
1382
1383         if (g4x_compute_wm0(dev, PIPE_A,
1384                             &valleyview_wm_info, latency_ns,
1385                             &valleyview_cursor_wm_info, latency_ns,
1386                             &planea_wm, &cursora_wm))
1387                 enabled |= 1 << PIPE_A;
1388
1389         if (g4x_compute_wm0(dev, PIPE_B,
1390                             &valleyview_wm_info, latency_ns,
1391                             &valleyview_cursor_wm_info, latency_ns,
1392                             &planeb_wm, &cursorb_wm))
1393                 enabled |= 1 << PIPE_B;
1394
1395         if (single_plane_enabled(enabled) &&
1396             g4x_compute_srwm(dev, ffs(enabled) - 1,
1397                              sr_latency_ns,
1398                              &valleyview_wm_info,
1399                              &valleyview_cursor_wm_info,
1400                              &plane_sr, &ignore_cursor_sr) &&
1401             g4x_compute_srwm(dev, ffs(enabled) - 1,
1402                              2*sr_latency_ns,
1403                              &valleyview_wm_info,
1404                              &valleyview_cursor_wm_info,
1405                              &ignore_plane_sr, &cursor_sr)) {
1406                 I915_WRITE(FW_BLC_SELF_VLV, FW_CSPWRDWNEN);
1407         } else {
1408                 I915_WRITE(FW_BLC_SELF_VLV,
1409                            I915_READ(FW_BLC_SELF_VLV) & ~FW_CSPWRDWNEN);
1410                 plane_sr = cursor_sr = 0;
1411         }
1412
1413         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1414                       planea_wm, cursora_wm,
1415                       planeb_wm, cursorb_wm,
1416                       plane_sr, cursor_sr);
1417
1418         I915_WRITE(DSPFW1,
1419                    (plane_sr << DSPFW_SR_SHIFT) |
1420                    (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1421                    (planeb_wm << DSPFW_PLANEB_SHIFT) |
1422                    planea_wm);
1423         I915_WRITE(DSPFW2,
1424                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1425                    (cursora_wm << DSPFW_CURSORA_SHIFT));
1426         I915_WRITE(DSPFW3,
1427                    (I915_READ(DSPFW3) & ~DSPFW_CURSOR_SR_MASK) |
1428                    (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1429 }
1430
1431 static void g4x_update_wm(struct drm_crtc *crtc)
1432 {
1433         struct drm_device *dev = crtc->dev;
1434         static const int sr_latency_ns = 12000;
1435         struct drm_i915_private *dev_priv = dev->dev_private;
1436         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1437         int plane_sr, cursor_sr;
1438         unsigned int enabled = 0;
1439
1440         if (g4x_compute_wm0(dev, PIPE_A,
1441                             &g4x_wm_info, latency_ns,
1442                             &g4x_cursor_wm_info, latency_ns,
1443                             &planea_wm, &cursora_wm))
1444                 enabled |= 1 << PIPE_A;
1445
1446         if (g4x_compute_wm0(dev, PIPE_B,
1447                             &g4x_wm_info, latency_ns,
1448                             &g4x_cursor_wm_info, latency_ns,
1449                             &planeb_wm, &cursorb_wm))
1450                 enabled |= 1 << PIPE_B;
1451
1452         if (single_plane_enabled(enabled) &&
1453             g4x_compute_srwm(dev, ffs(enabled) - 1,
1454                              sr_latency_ns,
1455                              &g4x_wm_info,
1456                              &g4x_cursor_wm_info,
1457                              &plane_sr, &cursor_sr)) {
1458                 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1459         } else {
1460                 I915_WRITE(FW_BLC_SELF,
1461                            I915_READ(FW_BLC_SELF) & ~FW_BLC_SELF_EN);
1462                 plane_sr = cursor_sr = 0;
1463         }
1464
1465         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1466                       planea_wm, cursora_wm,
1467                       planeb_wm, cursorb_wm,
1468                       plane_sr, cursor_sr);
1469
1470         I915_WRITE(DSPFW1,
1471                    (plane_sr << DSPFW_SR_SHIFT) |
1472                    (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1473                    (planeb_wm << DSPFW_PLANEB_SHIFT) |
1474                    planea_wm);
1475         I915_WRITE(DSPFW2,
1476                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1477                    (cursora_wm << DSPFW_CURSORA_SHIFT));
1478         /* HPLL off in SR has some issues on G4x... disable it */
1479         I915_WRITE(DSPFW3,
1480                    (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1481                    (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1482 }
1483
1484 static void i965_update_wm(struct drm_crtc *unused_crtc)
1485 {
1486         struct drm_device *dev = unused_crtc->dev;
1487         struct drm_i915_private *dev_priv = dev->dev_private;
1488         struct drm_crtc *crtc;
1489         int srwm = 1;
1490         int cursor_sr = 16;
1491
1492         /* Calc sr entries for one plane configs */
1493         crtc = single_enabled_crtc(dev);
1494         if (crtc) {
1495                 /* self-refresh has much higher latency */
1496                 static const int sr_latency_ns = 12000;
1497                 const struct drm_display_mode *adjusted_mode =
1498                         &to_intel_crtc(crtc)->config.adjusted_mode;
1499                 int clock = adjusted_mode->crtc_clock;
1500                 int htotal = adjusted_mode->htotal;
1501                 int hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1502                 int pixel_size = crtc->fb->bits_per_pixel / 8;
1503                 unsigned long line_time_us;
1504                 int entries;
1505
1506                 line_time_us = ((htotal * 1000) / clock);
1507
1508                 /* Use ns/us then divide to preserve precision */
1509                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1510                         pixel_size * hdisplay;
1511                 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1512                 srwm = I965_FIFO_SIZE - entries;
1513                 if (srwm < 0)
1514                         srwm = 1;
1515                 srwm &= 0x1ff;
1516                 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1517                               entries, srwm);
1518
1519                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1520                         pixel_size * 64;
1521                 entries = DIV_ROUND_UP(entries,
1522                                           i965_cursor_wm_info.cacheline_size);
1523                 cursor_sr = i965_cursor_wm_info.fifo_size -
1524                         (entries + i965_cursor_wm_info.guard_size);
1525
1526                 if (cursor_sr > i965_cursor_wm_info.max_wm)
1527                         cursor_sr = i965_cursor_wm_info.max_wm;
1528
1529                 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1530                               "cursor %d\n", srwm, cursor_sr);
1531
1532                 if (IS_CRESTLINE(dev))
1533                         I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1534         } else {
1535                 /* Turn off self refresh if both pipes are enabled */
1536                 if (IS_CRESTLINE(dev))
1537                         I915_WRITE(FW_BLC_SELF, I915_READ(FW_BLC_SELF)
1538                                    & ~FW_BLC_SELF_EN);
1539         }
1540
1541         DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1542                       srwm);
1543
1544         /* 965 has limitations... */
1545         I915_WRITE(DSPFW1, (srwm << DSPFW_SR_SHIFT) |
1546                    (8 << 16) | (8 << 8) | (8 << 0));
1547         I915_WRITE(DSPFW2, (8 << 8) | (8 << 0));
1548         /* update cursor SR watermark */
1549         I915_WRITE(DSPFW3, (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1550 }
1551
1552 static void i9xx_update_wm(struct drm_crtc *unused_crtc)
1553 {
1554         struct drm_device *dev = unused_crtc->dev;
1555         struct drm_i915_private *dev_priv = dev->dev_private;
1556         const struct intel_watermark_params *wm_info;
1557         uint32_t fwater_lo;
1558         uint32_t fwater_hi;
1559         int cwm, srwm = 1;
1560         int fifo_size;
1561         int planea_wm, planeb_wm;
1562         struct drm_crtc *crtc, *enabled = NULL;
1563
1564         if (IS_I945GM(dev))
1565                 wm_info = &i945_wm_info;
1566         else if (!IS_GEN2(dev))
1567                 wm_info = &i915_wm_info;
1568         else
1569                 wm_info = &i855_wm_info;
1570
1571         fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1572         crtc = intel_get_crtc_for_plane(dev, 0);
1573         if (intel_crtc_active(crtc)) {
1574                 const struct drm_display_mode *adjusted_mode;
1575                 int cpp = crtc->fb->bits_per_pixel / 8;
1576                 if (IS_GEN2(dev))
1577                         cpp = 4;
1578
1579                 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1580                 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1581                                                wm_info, fifo_size, cpp,
1582                                                latency_ns);
1583                 enabled = crtc;
1584         } else
1585                 planea_wm = fifo_size - wm_info->guard_size;
1586
1587         fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1588         crtc = intel_get_crtc_for_plane(dev, 1);
1589         if (intel_crtc_active(crtc)) {
1590                 const struct drm_display_mode *adjusted_mode;
1591                 int cpp = crtc->fb->bits_per_pixel / 8;
1592                 if (IS_GEN2(dev))
1593                         cpp = 4;
1594
1595                 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1596                 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1597                                                wm_info, fifo_size, cpp,
1598                                                latency_ns);
1599                 if (enabled == NULL)
1600                         enabled = crtc;
1601                 else
1602                         enabled = NULL;
1603         } else
1604                 planeb_wm = fifo_size - wm_info->guard_size;
1605
1606         DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1607
1608         /*
1609          * Overlay gets an aggressive default since video jitter is bad.
1610          */
1611         cwm = 2;
1612
1613         /* Play safe and disable self-refresh before adjusting watermarks. */
1614         if (IS_I945G(dev) || IS_I945GM(dev))
1615                 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN_MASK | 0);
1616         else if (IS_I915GM(dev))
1617                 I915_WRITE(INSTPM, I915_READ(INSTPM) & ~INSTPM_SELF_EN);
1618
1619         /* Calc sr entries for one plane configs */
1620         if (HAS_FW_BLC(dev) && enabled) {
1621                 /* self-refresh has much higher latency */
1622                 static const int sr_latency_ns = 6000;
1623                 const struct drm_display_mode *adjusted_mode =
1624                         &to_intel_crtc(enabled)->config.adjusted_mode;
1625                 int clock = adjusted_mode->crtc_clock;
1626                 int htotal = adjusted_mode->htotal;
1627                 int hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1628                 int pixel_size = enabled->fb->bits_per_pixel / 8;
1629                 unsigned long line_time_us;
1630                 int entries;
1631
1632                 line_time_us = (htotal * 1000) / clock;
1633
1634                 /* Use ns/us then divide to preserve precision */
1635                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1636                         pixel_size * hdisplay;
1637                 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1638                 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1639                 srwm = wm_info->fifo_size - entries;
1640                 if (srwm < 0)
1641                         srwm = 1;
1642
1643                 if (IS_I945G(dev) || IS_I945GM(dev))
1644                         I915_WRITE(FW_BLC_SELF,
1645                                    FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1646                 else if (IS_I915GM(dev))
1647                         I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1648         }
1649
1650         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1651                       planea_wm, planeb_wm, cwm, srwm);
1652
1653         fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1654         fwater_hi = (cwm & 0x1f);
1655
1656         /* Set request length to 8 cachelines per fetch */
1657         fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1658         fwater_hi = fwater_hi | (1 << 8);
1659
1660         I915_WRITE(FW_BLC, fwater_lo);
1661         I915_WRITE(FW_BLC2, fwater_hi);
1662
1663         if (HAS_FW_BLC(dev)) {
1664                 if (enabled) {
1665                         if (IS_I945G(dev) || IS_I945GM(dev))
1666                                 I915_WRITE(FW_BLC_SELF,
1667                                            FW_BLC_SELF_EN_MASK | FW_BLC_SELF_EN);
1668                         else if (IS_I915GM(dev))
1669                                 I915_WRITE(INSTPM, I915_READ(INSTPM) | INSTPM_SELF_EN);
1670                         DRM_DEBUG_KMS("memory self refresh enabled\n");
1671                 } else
1672                         DRM_DEBUG_KMS("memory self refresh disabled\n");
1673         }
1674 }
1675
1676 static void i830_update_wm(struct drm_crtc *unused_crtc)
1677 {
1678         struct drm_device *dev = unused_crtc->dev;
1679         struct drm_i915_private *dev_priv = dev->dev_private;
1680         struct drm_crtc *crtc;
1681         const struct drm_display_mode *adjusted_mode;
1682         uint32_t fwater_lo;
1683         int planea_wm;
1684
1685         crtc = single_enabled_crtc(dev);
1686         if (crtc == NULL)
1687                 return;
1688
1689         adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1690         planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1691                                        &i830_wm_info,
1692                                        dev_priv->display.get_fifo_size(dev, 0),
1693                                        4, latency_ns);
1694         fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1695         fwater_lo |= (3<<8) | planea_wm;
1696
1697         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1698
1699         I915_WRITE(FW_BLC, fwater_lo);
1700 }
1701
1702 /*
1703  * Check the wm result.
1704  *
1705  * If any calculated watermark values is larger than the maximum value that
1706  * can be programmed into the associated watermark register, that watermark
1707  * must be disabled.
1708  */
1709 static bool ironlake_check_srwm(struct drm_device *dev, int level,
1710                                 int fbc_wm, int display_wm, int cursor_wm,
1711                                 const struct intel_watermark_params *display,
1712                                 const struct intel_watermark_params *cursor)
1713 {
1714         struct drm_i915_private *dev_priv = dev->dev_private;
1715
1716         DRM_DEBUG_KMS("watermark %d: display plane %d, fbc lines %d,"
1717                       " cursor %d\n", level, display_wm, fbc_wm, cursor_wm);
1718
1719         if (fbc_wm > SNB_FBC_MAX_SRWM) {
1720                 DRM_DEBUG_KMS("fbc watermark(%d) is too large(%d), disabling wm%d+\n",
1721                               fbc_wm, SNB_FBC_MAX_SRWM, level);
1722
1723                 /* fbc has it's own way to disable FBC WM */
1724                 I915_WRITE(DISP_ARB_CTL,
1725                            I915_READ(DISP_ARB_CTL) | DISP_FBC_WM_DIS);
1726                 return false;
1727         } else if (INTEL_INFO(dev)->gen >= 6) {
1728                 /* enable FBC WM (except on ILK, where it must remain off) */
1729                 I915_WRITE(DISP_ARB_CTL,
1730                            I915_READ(DISP_ARB_CTL) & ~DISP_FBC_WM_DIS);
1731         }
1732
1733         if (display_wm > display->max_wm) {
1734                 DRM_DEBUG_KMS("display watermark(%d) is too large(%d), disabling wm%d+\n",
1735                               display_wm, SNB_DISPLAY_MAX_SRWM, level);
1736                 return false;
1737         }
1738
1739         if (cursor_wm > cursor->max_wm) {
1740                 DRM_DEBUG_KMS("cursor watermark(%d) is too large(%d), disabling wm%d+\n",
1741                               cursor_wm, SNB_CURSOR_MAX_SRWM, level);
1742                 return false;
1743         }
1744
1745         if (!(fbc_wm || display_wm || cursor_wm)) {
1746                 DRM_DEBUG_KMS("latency %d is 0, disabling wm%d+\n", level, level);
1747                 return false;
1748         }
1749
1750         return true;
1751 }
1752
1753 /*
1754  * Compute watermark values of WM[1-3],
1755  */
1756 static bool ironlake_compute_srwm(struct drm_device *dev, int level, int plane,
1757                                   int latency_ns,
1758                                   const struct intel_watermark_params *display,
1759                                   const struct intel_watermark_params *cursor,
1760                                   int *fbc_wm, int *display_wm, int *cursor_wm)
1761 {
1762         struct drm_crtc *crtc;
1763         const struct drm_display_mode *adjusted_mode;
1764         unsigned long line_time_us;
1765         int hdisplay, htotal, pixel_size, clock;
1766         int line_count, line_size;
1767         int small, large;
1768         int entries;
1769
1770         if (!latency_ns) {
1771                 *fbc_wm = *display_wm = *cursor_wm = 0;
1772                 return false;
1773         }
1774
1775         crtc = intel_get_crtc_for_plane(dev, plane);
1776         adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1777         clock = adjusted_mode->crtc_clock;
1778         htotal = adjusted_mode->htotal;
1779         hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1780         pixel_size = crtc->fb->bits_per_pixel / 8;
1781
1782         line_time_us = (htotal * 1000) / clock;
1783         line_count = (latency_ns / line_time_us + 1000) / 1000;
1784         line_size = hdisplay * pixel_size;
1785
1786         /* Use the minimum of the small and large buffer method for primary */
1787         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1788         large = line_count * line_size;
1789
1790         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1791         *display_wm = entries + display->guard_size;
1792
1793         /*
1794          * Spec says:
1795          * FBC WM = ((Final Primary WM * 64) / number of bytes per line) + 2
1796          */
1797         *fbc_wm = DIV_ROUND_UP(*display_wm * 64, line_size) + 2;
1798
1799         /* calculate the self-refresh watermark for display cursor */
1800         entries = line_count * pixel_size * 64;
1801         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1802         *cursor_wm = entries + cursor->guard_size;
1803
1804         return ironlake_check_srwm(dev, level,
1805                                    *fbc_wm, *display_wm, *cursor_wm,
1806                                    display, cursor);
1807 }
1808
1809 static void ironlake_update_wm(struct drm_crtc *crtc)
1810 {
1811         struct drm_device *dev = crtc->dev;
1812         struct drm_i915_private *dev_priv = dev->dev_private;
1813         int fbc_wm, plane_wm, cursor_wm;
1814         unsigned int enabled;
1815
1816         enabled = 0;
1817         if (g4x_compute_wm0(dev, PIPE_A,
1818                             &ironlake_display_wm_info,
1819                             dev_priv->wm.pri_latency[0] * 100,
1820                             &ironlake_cursor_wm_info,
1821                             dev_priv->wm.cur_latency[0] * 100,
1822                             &plane_wm, &cursor_wm)) {
1823                 I915_WRITE(WM0_PIPEA_ILK,
1824                            (plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm);
1825                 DRM_DEBUG_KMS("FIFO watermarks For pipe A -"
1826                               " plane %d, " "cursor: %d\n",
1827                               plane_wm, cursor_wm);
1828                 enabled |= 1 << PIPE_A;
1829         }
1830
1831         if (g4x_compute_wm0(dev, PIPE_B,
1832                             &ironlake_display_wm_info,
1833                             dev_priv->wm.pri_latency[0] * 100,
1834                             &ironlake_cursor_wm_info,
1835                             dev_priv->wm.cur_latency[0] * 100,
1836                             &plane_wm, &cursor_wm)) {
1837                 I915_WRITE(WM0_PIPEB_ILK,
1838                            (plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm);
1839                 DRM_DEBUG_KMS("FIFO watermarks For pipe B -"
1840                               " plane %d, cursor: %d\n",
1841                               plane_wm, cursor_wm);
1842                 enabled |= 1 << PIPE_B;
1843         }
1844
1845         /*
1846          * Calculate and update the self-refresh watermark only when one
1847          * display plane is used.
1848          */
1849         I915_WRITE(WM3_LP_ILK, 0);
1850         I915_WRITE(WM2_LP_ILK, 0);
1851         I915_WRITE(WM1_LP_ILK, 0);
1852
1853         if (!single_plane_enabled(enabled))
1854                 return;
1855         enabled = ffs(enabled) - 1;
1856
1857         /* WM1 */
1858         if (!ironlake_compute_srwm(dev, 1, enabled,
1859                                    dev_priv->wm.pri_latency[1] * 500,
1860                                    &ironlake_display_srwm_info,
1861                                    &ironlake_cursor_srwm_info,
1862                                    &fbc_wm, &plane_wm, &cursor_wm))
1863                 return;
1864
1865         I915_WRITE(WM1_LP_ILK,
1866                    WM1_LP_SR_EN |
1867                    (dev_priv->wm.pri_latency[1] << WM1_LP_LATENCY_SHIFT) |
1868                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1869                    (plane_wm << WM1_LP_SR_SHIFT) |
1870                    cursor_wm);
1871
1872         /* WM2 */
1873         if (!ironlake_compute_srwm(dev, 2, enabled,
1874                                    dev_priv->wm.pri_latency[2] * 500,
1875                                    &ironlake_display_srwm_info,
1876                                    &ironlake_cursor_srwm_info,
1877                                    &fbc_wm, &plane_wm, &cursor_wm))
1878                 return;
1879
1880         I915_WRITE(WM2_LP_ILK,
1881                    WM2_LP_EN |
1882                    (dev_priv->wm.pri_latency[2] << WM1_LP_LATENCY_SHIFT) |
1883                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1884                    (plane_wm << WM1_LP_SR_SHIFT) |
1885                    cursor_wm);
1886
1887         /*
1888          * WM3 is unsupported on ILK, probably because we don't have latency
1889          * data for that power state
1890          */
1891 }
1892
1893 static void sandybridge_update_wm(struct drm_crtc *crtc)
1894 {
1895         struct drm_device *dev = crtc->dev;
1896         struct drm_i915_private *dev_priv = dev->dev_private;
1897         int latency = dev_priv->wm.pri_latency[0] * 100;        /* In unit 0.1us */
1898         u32 val;
1899         int fbc_wm, plane_wm, cursor_wm;
1900         unsigned int enabled;
1901
1902         enabled = 0;
1903         if (g4x_compute_wm0(dev, PIPE_A,
1904                             &sandybridge_display_wm_info, latency,
1905                             &sandybridge_cursor_wm_info, latency,
1906                             &plane_wm, &cursor_wm)) {
1907                 val = I915_READ(WM0_PIPEA_ILK);
1908                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
1909                 I915_WRITE(WM0_PIPEA_ILK, val |
1910                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
1911                 DRM_DEBUG_KMS("FIFO watermarks For pipe A -"
1912                               " plane %d, " "cursor: %d\n",
1913                               plane_wm, cursor_wm);
1914                 enabled |= 1 << PIPE_A;
1915         }
1916
1917         if (g4x_compute_wm0(dev, PIPE_B,
1918                             &sandybridge_display_wm_info, latency,
1919                             &sandybridge_cursor_wm_info, latency,
1920                             &plane_wm, &cursor_wm)) {
1921                 val = I915_READ(WM0_PIPEB_ILK);
1922                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
1923                 I915_WRITE(WM0_PIPEB_ILK, val |
1924                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
1925                 DRM_DEBUG_KMS("FIFO watermarks For pipe B -"
1926                               " plane %d, cursor: %d\n",
1927                               plane_wm, cursor_wm);
1928                 enabled |= 1 << PIPE_B;
1929         }
1930
1931         /*
1932          * Calculate and update the self-refresh watermark only when one
1933          * display plane is used.
1934          *
1935          * SNB support 3 levels of watermark.
1936          *
1937          * WM1/WM2/WM2 watermarks have to be enabled in the ascending order,
1938          * and disabled in the descending order
1939          *
1940          */
1941         I915_WRITE(WM3_LP_ILK, 0);
1942         I915_WRITE(WM2_LP_ILK, 0);
1943         I915_WRITE(WM1_LP_ILK, 0);
1944
1945         if (!single_plane_enabled(enabled) ||
1946             dev_priv->sprite_scaling_enabled)
1947                 return;
1948         enabled = ffs(enabled) - 1;
1949
1950         /* WM1 */
1951         if (!ironlake_compute_srwm(dev, 1, enabled,
1952                                    dev_priv->wm.pri_latency[1] * 500,
1953                                    &sandybridge_display_srwm_info,
1954                                    &sandybridge_cursor_srwm_info,
1955                                    &fbc_wm, &plane_wm, &cursor_wm))
1956                 return;
1957
1958         I915_WRITE(WM1_LP_ILK,
1959                    WM1_LP_SR_EN |
1960                    (dev_priv->wm.pri_latency[1] << WM1_LP_LATENCY_SHIFT) |
1961                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1962                    (plane_wm << WM1_LP_SR_SHIFT) |
1963                    cursor_wm);
1964
1965         /* WM2 */
1966         if (!ironlake_compute_srwm(dev, 2, enabled,
1967                                    dev_priv->wm.pri_latency[2] * 500,
1968                                    &sandybridge_display_srwm_info,
1969                                    &sandybridge_cursor_srwm_info,
1970                                    &fbc_wm, &plane_wm, &cursor_wm))
1971                 return;
1972
1973         I915_WRITE(WM2_LP_ILK,
1974                    WM2_LP_EN |
1975                    (dev_priv->wm.pri_latency[2] << WM1_LP_LATENCY_SHIFT) |
1976                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1977                    (plane_wm << WM1_LP_SR_SHIFT) |
1978                    cursor_wm);
1979
1980         /* WM3 */
1981         if (!ironlake_compute_srwm(dev, 3, enabled,
1982                                    dev_priv->wm.pri_latency[3] * 500,
1983                                    &sandybridge_display_srwm_info,
1984                                    &sandybridge_cursor_srwm_info,
1985                                    &fbc_wm, &plane_wm, &cursor_wm))
1986                 return;
1987
1988         I915_WRITE(WM3_LP_ILK,
1989                    WM3_LP_EN |
1990                    (dev_priv->wm.pri_latency[3] << WM1_LP_LATENCY_SHIFT) |
1991                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1992                    (plane_wm << WM1_LP_SR_SHIFT) |
1993                    cursor_wm);
1994 }
1995
1996 static void ivybridge_update_wm(struct drm_crtc *crtc)
1997 {
1998         struct drm_device *dev = crtc->dev;
1999         struct drm_i915_private *dev_priv = dev->dev_private;
2000         int latency = dev_priv->wm.pri_latency[0] * 100;        /* In unit 0.1us */
2001         u32 val;
2002         int fbc_wm, plane_wm, cursor_wm;
2003         int ignore_fbc_wm, ignore_plane_wm, ignore_cursor_wm;
2004         unsigned int enabled;
2005
2006         enabled = 0;
2007         if (g4x_compute_wm0(dev, PIPE_A,
2008                             &sandybridge_display_wm_info, latency,
2009                             &sandybridge_cursor_wm_info, latency,
2010                             &plane_wm, &cursor_wm)) {
2011                 val = I915_READ(WM0_PIPEA_ILK);
2012                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
2013                 I915_WRITE(WM0_PIPEA_ILK, val |
2014                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
2015                 DRM_DEBUG_KMS("FIFO watermarks For pipe A -"
2016                               " plane %d, " "cursor: %d\n",
2017                               plane_wm, cursor_wm);
2018                 enabled |= 1 << PIPE_A;
2019         }
2020
2021         if (g4x_compute_wm0(dev, PIPE_B,
2022                             &sandybridge_display_wm_info, latency,
2023                             &sandybridge_cursor_wm_info, latency,
2024                             &plane_wm, &cursor_wm)) {
2025                 val = I915_READ(WM0_PIPEB_ILK);
2026                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
2027                 I915_WRITE(WM0_PIPEB_ILK, val |
2028                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
2029                 DRM_DEBUG_KMS("FIFO watermarks For pipe B -"
2030                               " plane %d, cursor: %d\n",
2031                               plane_wm, cursor_wm);
2032                 enabled |= 1 << PIPE_B;
2033         }
2034
2035         if (g4x_compute_wm0(dev, PIPE_C,
2036                             &sandybridge_display_wm_info, latency,
2037                             &sandybridge_cursor_wm_info, latency,
2038                             &plane_wm, &cursor_wm)) {
2039                 val = I915_READ(WM0_PIPEC_IVB);
2040                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
2041                 I915_WRITE(WM0_PIPEC_IVB, val |
2042                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
2043                 DRM_DEBUG_KMS("FIFO watermarks For pipe C -"
2044                               " plane %d, cursor: %d\n",
2045                               plane_wm, cursor_wm);
2046                 enabled |= 1 << PIPE_C;
2047         }
2048
2049         /*
2050          * Calculate and update the self-refresh watermark only when one
2051          * display plane is used.
2052          *
2053          * SNB support 3 levels of watermark.
2054          *
2055          * WM1/WM2/WM2 watermarks have to be enabled in the ascending order,
2056          * and disabled in the descending order
2057          *
2058          */
2059         I915_WRITE(WM3_LP_ILK, 0);
2060         I915_WRITE(WM2_LP_ILK, 0);
2061         I915_WRITE(WM1_LP_ILK, 0);
2062
2063         if (!single_plane_enabled(enabled) ||
2064             dev_priv->sprite_scaling_enabled)
2065                 return;
2066         enabled = ffs(enabled) - 1;
2067
2068         /* WM1 */
2069         if (!ironlake_compute_srwm(dev, 1, enabled,
2070                                    dev_priv->wm.pri_latency[1] * 500,
2071                                    &sandybridge_display_srwm_info,
2072                                    &sandybridge_cursor_srwm_info,
2073                                    &fbc_wm, &plane_wm, &cursor_wm))
2074                 return;
2075
2076         I915_WRITE(WM1_LP_ILK,
2077                    WM1_LP_SR_EN |
2078                    (dev_priv->wm.pri_latency[1] << WM1_LP_LATENCY_SHIFT) |
2079                    (fbc_wm << WM1_LP_FBC_SHIFT) |
2080                    (plane_wm << WM1_LP_SR_SHIFT) |
2081                    cursor_wm);
2082
2083         /* WM2 */
2084         if (!ironlake_compute_srwm(dev, 2, enabled,
2085                                    dev_priv->wm.pri_latency[2] * 500,
2086                                    &sandybridge_display_srwm_info,
2087                                    &sandybridge_cursor_srwm_info,
2088                                    &fbc_wm, &plane_wm, &cursor_wm))
2089                 return;
2090
2091         I915_WRITE(WM2_LP_ILK,
2092                    WM2_LP_EN |
2093                    (dev_priv->wm.pri_latency[2] << WM1_LP_LATENCY_SHIFT) |
2094                    (fbc_wm << WM1_LP_FBC_SHIFT) |
2095                    (plane_wm << WM1_LP_SR_SHIFT) |
2096                    cursor_wm);
2097
2098         /* WM3, note we have to correct the cursor latency */
2099         if (!ironlake_compute_srwm(dev, 3, enabled,
2100                                    dev_priv->wm.pri_latency[3] * 500,
2101                                    &sandybridge_display_srwm_info,
2102                                    &sandybridge_cursor_srwm_info,
2103                                    &fbc_wm, &plane_wm, &ignore_cursor_wm) ||
2104             !ironlake_compute_srwm(dev, 3, enabled,
2105                                    dev_priv->wm.cur_latency[3] * 500,
2106                                    &sandybridge_display_srwm_info,
2107                                    &sandybridge_cursor_srwm_info,
2108                                    &ignore_fbc_wm, &ignore_plane_wm, &cursor_wm))
2109                 return;
2110
2111         I915_WRITE(WM3_LP_ILK,
2112                    WM3_LP_EN |
2113                    (dev_priv->wm.pri_latency[3] << WM1_LP_LATENCY_SHIFT) |
2114                    (fbc_wm << WM1_LP_FBC_SHIFT) |
2115                    (plane_wm << WM1_LP_SR_SHIFT) |
2116                    cursor_wm);
2117 }
2118
2119 static uint32_t ilk_pipe_pixel_rate(struct drm_device *dev,
2120                                     struct drm_crtc *crtc)
2121 {
2122         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2123         uint32_t pixel_rate;
2124
2125         pixel_rate = intel_crtc->config.adjusted_mode.crtc_clock;
2126
2127         /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
2128          * adjust the pixel_rate here. */
2129
2130         if (intel_crtc->config.pch_pfit.enabled) {
2131                 uint64_t pipe_w, pipe_h, pfit_w, pfit_h;
2132                 uint32_t pfit_size = intel_crtc->config.pch_pfit.size;
2133
2134                 pipe_w = intel_crtc->config.pipe_src_w;
2135                 pipe_h = intel_crtc->config.pipe_src_h;
2136                 pfit_w = (pfit_size >> 16) & 0xFFFF;
2137                 pfit_h = pfit_size & 0xFFFF;
2138                 if (pipe_w < pfit_w)
2139                         pipe_w = pfit_w;
2140                 if (pipe_h < pfit_h)
2141                         pipe_h = pfit_h;
2142
2143                 pixel_rate = div_u64((uint64_t) pixel_rate * pipe_w * pipe_h,
2144                                      pfit_w * pfit_h);
2145         }
2146
2147         return pixel_rate;
2148 }
2149
2150 /* latency must be in 0.1us units. */
2151 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
2152                                uint32_t latency)
2153 {
2154         uint64_t ret;
2155
2156         if (WARN(latency == 0, "Latency value missing\n"))
2157                 return UINT_MAX;
2158
2159         ret = (uint64_t) pixel_rate * bytes_per_pixel * latency;
2160         ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
2161
2162         return ret;
2163 }
2164
2165 /* latency must be in 0.1us units. */
2166 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
2167                                uint32_t horiz_pixels, uint8_t bytes_per_pixel,
2168                                uint32_t latency)
2169 {
2170         uint32_t ret;
2171
2172         if (WARN(latency == 0, "Latency value missing\n"))
2173                 return UINT_MAX;
2174
2175         ret = (latency * pixel_rate) / (pipe_htotal * 10000);
2176         ret = (ret + 1) * horiz_pixels * bytes_per_pixel;
2177         ret = DIV_ROUND_UP(ret, 64) + 2;
2178         return ret;
2179 }
2180
2181 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
2182                            uint8_t bytes_per_pixel)
2183 {
2184         return DIV_ROUND_UP(pri_val * 64, horiz_pixels * bytes_per_pixel) + 2;
2185 }
2186
2187 struct hsw_pipe_wm_parameters {
2188         bool active;
2189         uint32_t pipe_htotal;
2190         uint32_t pixel_rate;
2191         struct intel_plane_wm_parameters pri;
2192         struct intel_plane_wm_parameters spr;
2193         struct intel_plane_wm_parameters cur;
2194 };
2195
2196 struct hsw_wm_maximums {
2197         uint16_t pri;
2198         uint16_t spr;
2199         uint16_t cur;
2200         uint16_t fbc;
2201 };
2202
2203 struct hsw_wm_values {
2204         uint32_t wm_pipe[3];
2205         uint32_t wm_lp[3];
2206         uint32_t wm_lp_spr[3];
2207         uint32_t wm_linetime[3];
2208         bool enable_fbc_wm;
2209 };
2210
2211 /* used in computing the new watermarks state */
2212 struct intel_wm_config {
2213         unsigned int num_pipes_active;
2214         bool sprites_enabled;
2215         bool sprites_scaled;
2216         bool fbc_wm_enabled;
2217 };
2218
2219 /*
2220  * For both WM_PIPE and WM_LP.
2221  * mem_value must be in 0.1us units.
2222  */
2223 static uint32_t ilk_compute_pri_wm(const struct hsw_pipe_wm_parameters *params,
2224                                    uint32_t mem_value,
2225                                    bool is_lp)
2226 {
2227         uint32_t method1, method2;
2228
2229         if (!params->active || !params->pri.enabled)
2230                 return 0;
2231
2232         method1 = ilk_wm_method1(params->pixel_rate,
2233                                  params->pri.bytes_per_pixel,
2234                                  mem_value);
2235
2236         if (!is_lp)
2237                 return method1;
2238
2239         method2 = ilk_wm_method2(params->pixel_rate,
2240                                  params->pipe_htotal,
2241                                  params->pri.horiz_pixels,
2242                                  params->pri.bytes_per_pixel,
2243                                  mem_value);
2244
2245         return min(method1, method2);
2246 }
2247
2248 /*
2249  * For both WM_PIPE and WM_LP.
2250  * mem_value must be in 0.1us units.
2251  */
2252 static uint32_t ilk_compute_spr_wm(const struct hsw_pipe_wm_parameters *params,
2253                                    uint32_t mem_value)
2254 {
2255         uint32_t method1, method2;
2256
2257         if (!params->active || !params->spr.enabled)
2258                 return 0;
2259
2260         method1 = ilk_wm_method1(params->pixel_rate,
2261                                  params->spr.bytes_per_pixel,
2262                                  mem_value);
2263         method2 = ilk_wm_method2(params->pixel_rate,
2264                                  params->pipe_htotal,
2265                                  params->spr.horiz_pixels,
2266                                  params->spr.bytes_per_pixel,
2267                                  mem_value);
2268         return min(method1, method2);
2269 }
2270
2271 /*
2272  * For both WM_PIPE and WM_LP.
2273  * mem_value must be in 0.1us units.
2274  */
2275 static uint32_t ilk_compute_cur_wm(const struct hsw_pipe_wm_parameters *params,
2276                                    uint32_t mem_value)
2277 {
2278         if (!params->active || !params->cur.enabled)
2279                 return 0;
2280
2281         return ilk_wm_method2(params->pixel_rate,
2282                               params->pipe_htotal,
2283                               params->cur.horiz_pixels,
2284                               params->cur.bytes_per_pixel,
2285                               mem_value);
2286 }
2287
2288 /* Only for WM_LP. */
2289 static uint32_t ilk_compute_fbc_wm(const struct hsw_pipe_wm_parameters *params,
2290                                    uint32_t pri_val)
2291 {
2292         if (!params->active || !params->pri.enabled)
2293                 return 0;
2294
2295         return ilk_wm_fbc(pri_val,
2296                           params->pri.horiz_pixels,
2297                           params->pri.bytes_per_pixel);
2298 }
2299
2300 static unsigned int ilk_display_fifo_size(const struct drm_device *dev)
2301 {
2302         if (INTEL_INFO(dev)->gen >= 7)
2303                 return 768;
2304         else
2305                 return 512;
2306 }
2307
2308 /* Calculate the maximum primary/sprite plane watermark */
2309 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
2310                                      int level,
2311                                      const struct intel_wm_config *config,
2312                                      enum intel_ddb_partitioning ddb_partitioning,
2313                                      bool is_sprite)
2314 {
2315         unsigned int fifo_size = ilk_display_fifo_size(dev);
2316         unsigned int max;
2317
2318         /* if sprites aren't enabled, sprites get nothing */
2319         if (is_sprite && !config->sprites_enabled)
2320                 return 0;
2321
2322         /* HSW allows LP1+ watermarks even with multiple pipes */
2323         if (level == 0 || config->num_pipes_active > 1) {
2324                 fifo_size /= INTEL_INFO(dev)->num_pipes;
2325
2326                 /*
2327                  * For some reason the non self refresh
2328                  * FIFO size is only half of the self
2329                  * refresh FIFO size on ILK/SNB.
2330                  */
2331                 if (INTEL_INFO(dev)->gen <= 6)
2332                         fifo_size /= 2;
2333         }
2334
2335         if (config->sprites_enabled) {
2336                 /* level 0 is always calculated with 1:1 split */
2337                 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
2338                         if (is_sprite)
2339                                 fifo_size *= 5;
2340                         fifo_size /= 6;
2341                 } else {
2342                         fifo_size /= 2;
2343                 }
2344         }
2345
2346         /* clamp to max that the registers can hold */
2347         if (INTEL_INFO(dev)->gen >= 7)
2348                 /* IVB/HSW primary/sprite plane watermarks */
2349                 max = level == 0 ? 127 : 1023;
2350         else if (!is_sprite)
2351                 /* ILK/SNB primary plane watermarks */
2352                 max = level == 0 ? 127 : 511;
2353         else
2354                 /* ILK/SNB sprite plane watermarks */
2355                 max = level == 0 ? 63 : 255;
2356
2357         return min(fifo_size, max);
2358 }
2359
2360 /* Calculate the maximum cursor plane watermark */
2361 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
2362                                       int level,
2363                                       const struct intel_wm_config *config)
2364 {
2365         /* HSW LP1+ watermarks w/ multiple pipes */
2366         if (level > 0 && config->num_pipes_active > 1)
2367                 return 64;
2368
2369         /* otherwise just report max that registers can hold */
2370         if (INTEL_INFO(dev)->gen >= 7)
2371                 return level == 0 ? 63 : 255;
2372         else
2373                 return level == 0 ? 31 : 63;
2374 }
2375
2376 /* Calculate the maximum FBC watermark */
2377 static unsigned int ilk_fbc_wm_max(void)
2378 {
2379         /* max that registers can hold */
2380         return 15;
2381 }
2382
2383 static void ilk_wm_max(struct drm_device *dev,
2384                        int level,
2385                        const struct intel_wm_config *config,
2386                        enum intel_ddb_partitioning ddb_partitioning,
2387                        struct hsw_wm_maximums *max)
2388 {
2389         max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
2390         max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
2391         max->cur = ilk_cursor_wm_max(dev, level, config);
2392         max->fbc = ilk_fbc_wm_max();
2393 }
2394
2395 static bool ilk_check_wm(int level,
2396                          const struct hsw_wm_maximums *max,
2397                          struct intel_wm_level *result)
2398 {
2399         bool ret;
2400
2401         /* already determined to be invalid? */
2402         if (!result->enable)
2403                 return false;
2404
2405         result->enable = result->pri_val <= max->pri &&
2406                          result->spr_val <= max->spr &&
2407                          result->cur_val <= max->cur;
2408
2409         ret = result->enable;
2410
2411         /*
2412          * HACK until we can pre-compute everything,
2413          * and thus fail gracefully if LP0 watermarks
2414          * are exceeded...
2415          */
2416         if (level == 0 && !result->enable) {
2417                 if (result->pri_val > max->pri)
2418                         DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
2419                                       level, result->pri_val, max->pri);
2420                 if (result->spr_val > max->spr)
2421                         DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
2422                                       level, result->spr_val, max->spr);
2423                 if (result->cur_val > max->cur)
2424                         DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
2425                                       level, result->cur_val, max->cur);
2426
2427                 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
2428                 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
2429                 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
2430                 result->enable = true;
2431         }
2432
2433         DRM_DEBUG_KMS("WM%d: %sabled\n", level, result->enable ? "en" : "dis");
2434
2435         return ret;
2436 }
2437
2438 static void ilk_compute_wm_level(struct drm_i915_private *dev_priv,
2439                                  int level,
2440                                  const struct hsw_pipe_wm_parameters *p,
2441                                  struct intel_wm_level *result)
2442 {
2443         uint16_t pri_latency = dev_priv->wm.pri_latency[level];
2444         uint16_t spr_latency = dev_priv->wm.spr_latency[level];
2445         uint16_t cur_latency = dev_priv->wm.cur_latency[level];
2446
2447         /* WM1+ latency values stored in 0.5us units */
2448         if (level > 0) {
2449                 pri_latency *= 5;
2450                 spr_latency *= 5;
2451                 cur_latency *= 5;
2452         }
2453
2454         result->pri_val = ilk_compute_pri_wm(p, pri_latency, level);
2455         result->spr_val = ilk_compute_spr_wm(p, spr_latency);
2456         result->cur_val = ilk_compute_cur_wm(p, cur_latency);
2457         result->fbc_val = ilk_compute_fbc_wm(p, result->pri_val);
2458         result->enable = true;
2459 }
2460
2461 static bool hsw_compute_lp_wm(struct drm_i915_private *dev_priv,
2462                               int level, const struct hsw_wm_maximums *max,
2463                               const struct hsw_pipe_wm_parameters *params,
2464                               struct intel_wm_level *result)
2465 {
2466         enum pipe pipe;
2467         struct intel_wm_level res[3];
2468
2469         for (pipe = PIPE_A; pipe <= PIPE_C; pipe++)
2470                 ilk_compute_wm_level(dev_priv, level, &params[pipe], &res[pipe]);
2471
2472         result->pri_val = max3(res[0].pri_val, res[1].pri_val, res[2].pri_val);
2473         result->spr_val = max3(res[0].spr_val, res[1].spr_val, res[2].spr_val);
2474         result->cur_val = max3(res[0].cur_val, res[1].cur_val, res[2].cur_val);
2475         result->fbc_val = max3(res[0].fbc_val, res[1].fbc_val, res[2].fbc_val);
2476         result->enable = true;
2477
2478         return ilk_check_wm(level, max, result);
2479 }
2480
2481
2482 static uint32_t hsw_compute_wm_pipe(struct drm_device *dev,
2483                                     const struct hsw_pipe_wm_parameters *params)
2484 {
2485         struct drm_i915_private *dev_priv = dev->dev_private;
2486         struct intel_wm_config config = {
2487                 .num_pipes_active = 1,
2488                 .sprites_enabled = params->spr.enabled,
2489                 .sprites_scaled = params->spr.scaled,
2490         };
2491         struct hsw_wm_maximums max;
2492         struct intel_wm_level res;
2493
2494         if (!params->active)
2495                 return 0;
2496
2497         ilk_wm_max(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
2498
2499         ilk_compute_wm_level(dev_priv, 0, params, &res);
2500
2501         ilk_check_wm(0, &max, &res);
2502
2503         return (res.pri_val << WM0_PIPE_PLANE_SHIFT) |
2504                (res.spr_val << WM0_PIPE_SPRITE_SHIFT) |
2505                res.cur_val;
2506 }
2507
2508 static uint32_t
2509 hsw_compute_linetime_wm(struct drm_device *dev, struct drm_crtc *crtc)
2510 {
2511         struct drm_i915_private *dev_priv = dev->dev_private;
2512         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2513         struct drm_display_mode *mode = &intel_crtc->config.adjusted_mode;
2514         u32 linetime, ips_linetime;
2515
2516         if (!intel_crtc_active(crtc))
2517                 return 0;
2518
2519         /* The WM are computed with base on how long it takes to fill a single
2520          * row at the given clock rate, multiplied by 8.
2521          * */
2522         linetime = DIV_ROUND_CLOSEST(mode->htotal * 1000 * 8, mode->clock);
2523         ips_linetime = DIV_ROUND_CLOSEST(mode->htotal * 1000 * 8,
2524                                          intel_ddi_get_cdclk_freq(dev_priv));
2525
2526         return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2527                PIPE_WM_LINETIME_TIME(linetime);
2528 }
2529
2530 static void intel_read_wm_latency(struct drm_device *dev, uint16_t wm[5])
2531 {
2532         struct drm_i915_private *dev_priv = dev->dev_private;
2533
2534         if (IS_HASWELL(dev)) {
2535                 uint64_t sskpd = I915_READ64(MCH_SSKPD);
2536
2537                 wm[0] = (sskpd >> 56) & 0xFF;
2538                 if (wm[0] == 0)
2539                         wm[0] = sskpd & 0xF;
2540                 wm[1] = (sskpd >> 4) & 0xFF;
2541                 wm[2] = (sskpd >> 12) & 0xFF;
2542                 wm[3] = (sskpd >> 20) & 0x1FF;
2543                 wm[4] = (sskpd >> 32) & 0x1FF;
2544         } else if (INTEL_INFO(dev)->gen >= 6) {
2545                 uint32_t sskpd = I915_READ(MCH_SSKPD);
2546
2547                 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2548                 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2549                 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2550                 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2551         } else if (INTEL_INFO(dev)->gen >= 5) {
2552                 uint32_t mltr = I915_READ(MLTR_ILK);
2553
2554                 /* ILK primary LP0 latency is 700 ns */
2555                 wm[0] = 7;
2556                 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2557                 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2558         }
2559 }
2560
2561 static void intel_fixup_spr_wm_latency(struct drm_device *dev, uint16_t wm[5])
2562 {
2563         /* ILK sprite LP0 latency is 1300 ns */
2564         if (INTEL_INFO(dev)->gen == 5)
2565                 wm[0] = 13;
2566 }
2567
2568 static void intel_fixup_cur_wm_latency(struct drm_device *dev, uint16_t wm[5])
2569 {
2570         /* ILK cursor LP0 latency is 1300 ns */
2571         if (INTEL_INFO(dev)->gen == 5)
2572                 wm[0] = 13;
2573
2574         /* WaDoubleCursorLP3Latency:ivb */
2575         if (IS_IVYBRIDGE(dev))
2576                 wm[3] *= 2;
2577 }
2578
2579 static int ilk_wm_max_level(const struct drm_device *dev)
2580 {
2581         /* how many WM levels are we expecting */
2582         if (IS_HASWELL(dev))
2583                 return 4;
2584         else if (INTEL_INFO(dev)->gen >= 6)
2585                 return 3;
2586         else
2587                 return 2;
2588 }
2589
2590 static void intel_print_wm_latency(struct drm_device *dev,
2591                                    const char *name,
2592                                    const uint16_t wm[5])
2593 {
2594         int level, max_level = ilk_wm_max_level(dev);
2595
2596         for (level = 0; level <= max_level; level++) {
2597                 unsigned int latency = wm[level];
2598
2599                 if (latency == 0) {
2600                         DRM_ERROR("%s WM%d latency not provided\n",
2601                                   name, level);
2602                         continue;
2603                 }
2604
2605                 /* WM1+ latency values in 0.5us units */
2606                 if (level > 0)
2607                         latency *= 5;
2608
2609                 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2610                               name, level, wm[level],
2611                               latency / 10, latency % 10);
2612         }
2613 }
2614
2615 static void intel_setup_wm_latency(struct drm_device *dev)
2616 {
2617         struct drm_i915_private *dev_priv = dev->dev_private;
2618
2619         intel_read_wm_latency(dev, dev_priv->wm.pri_latency);
2620
2621         memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2622                sizeof(dev_priv->wm.pri_latency));
2623         memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2624                sizeof(dev_priv->wm.pri_latency));
2625
2626         intel_fixup_spr_wm_latency(dev, dev_priv->wm.spr_latency);
2627         intel_fixup_cur_wm_latency(dev, dev_priv->wm.cur_latency);
2628
2629         intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2630         intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2631         intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2632 }
2633
2634 static void hsw_compute_wm_parameters(struct drm_device *dev,
2635                                       struct hsw_pipe_wm_parameters *params,
2636                                       struct hsw_wm_maximums *lp_max_1_2,
2637                                       struct hsw_wm_maximums *lp_max_5_6)
2638 {
2639         struct drm_crtc *crtc;
2640         struct drm_plane *plane;
2641         enum pipe pipe;
2642         struct intel_wm_config config = {};
2643
2644         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
2645                 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2646                 struct hsw_pipe_wm_parameters *p;
2647
2648                 pipe = intel_crtc->pipe;
2649                 p = &params[pipe];
2650
2651                 p->active = intel_crtc_active(crtc);
2652                 if (!p->active)
2653                         continue;
2654
2655                 config.num_pipes_active++;
2656
2657                 p->pipe_htotal = intel_crtc->config.adjusted_mode.htotal;
2658                 p->pixel_rate = ilk_pipe_pixel_rate(dev, crtc);
2659                 p->pri.bytes_per_pixel = crtc->fb->bits_per_pixel / 8;
2660                 p->cur.bytes_per_pixel = 4;
2661                 p->pri.horiz_pixels = intel_crtc->config.pipe_src_w;
2662                 p->cur.horiz_pixels = 64;
2663                 /* TODO: for now, assume primary and cursor planes are always enabled. */
2664                 p->pri.enabled = true;
2665                 p->cur.enabled = true;
2666         }
2667
2668         list_for_each_entry(plane, &dev->mode_config.plane_list, head) {
2669                 struct intel_plane *intel_plane = to_intel_plane(plane);
2670                 struct hsw_pipe_wm_parameters *p;
2671
2672                 pipe = intel_plane->pipe;
2673                 p = &params[pipe];
2674
2675                 p->spr = intel_plane->wm;
2676
2677                 config.sprites_enabled |= p->spr.enabled;
2678                 config.sprites_scaled |= p->spr.scaled;
2679         }
2680
2681         ilk_wm_max(dev, 1, &config, INTEL_DDB_PART_1_2, lp_max_1_2);
2682
2683         /* 5/6 split only in single pipe config on IVB+ */
2684         if (INTEL_INFO(dev)->gen >= 7 && config.num_pipes_active <= 1)
2685                 ilk_wm_max(dev, 1, &config, INTEL_DDB_PART_5_6, lp_max_5_6);
2686         else
2687                 *lp_max_5_6 = *lp_max_1_2;
2688 }
2689
2690 static void hsw_compute_wm_results(struct drm_device *dev,
2691                                    const struct hsw_pipe_wm_parameters *params,
2692                                    const struct hsw_wm_maximums *lp_maximums,
2693                                    struct hsw_wm_values *results)
2694 {
2695         struct drm_i915_private *dev_priv = dev->dev_private;
2696         struct drm_crtc *crtc;
2697         struct intel_wm_level lp_results[4] = {};
2698         enum pipe pipe;
2699         int level, max_level, wm_lp;
2700
2701         for (level = 1; level <= 4; level++)
2702                 if (!hsw_compute_lp_wm(dev_priv, level,
2703                                        lp_maximums, params,
2704                                        &lp_results[level - 1]))
2705                         break;
2706         max_level = level - 1;
2707
2708         memset(results, 0, sizeof(*results));
2709
2710         /* The spec says it is preferred to disable FBC WMs instead of disabling
2711          * a WM level. */
2712         results->enable_fbc_wm = true;
2713         for (level = 1; level <= max_level; level++) {
2714                 if (lp_results[level - 1].fbc_val > lp_maximums->fbc) {
2715                         results->enable_fbc_wm = false;
2716                         lp_results[level - 1].fbc_val = 0;
2717                 }
2718         }
2719
2720         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2721                 const struct intel_wm_level *r;
2722
2723                 level = (max_level == 4 && wm_lp > 1) ? wm_lp + 1 : wm_lp;
2724                 if (level > max_level)
2725                         break;
2726
2727                 r = &lp_results[level - 1];
2728                 results->wm_lp[wm_lp - 1] = HSW_WM_LP_VAL(level * 2,
2729                                                           r->fbc_val,
2730                                                           r->pri_val,
2731                                                           r->cur_val);
2732                 results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2733         }
2734
2735         for_each_pipe(pipe)
2736                 results->wm_pipe[pipe] = hsw_compute_wm_pipe(dev,
2737                                                              &params[pipe]);
2738
2739         for_each_pipe(pipe) {
2740                 crtc = dev_priv->pipe_to_crtc_mapping[pipe];
2741                 results->wm_linetime[pipe] = hsw_compute_linetime_wm(dev, crtc);
2742         }
2743 }
2744
2745 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2746  * case both are at the same level. Prefer r1 in case they're the same. */
2747 static struct hsw_wm_values *hsw_find_best_result(struct hsw_wm_values *r1,
2748                                                   struct hsw_wm_values *r2)
2749 {
2750         int i, val_r1 = 0, val_r2 = 0;
2751
2752         for (i = 0; i < 3; i++) {
2753                 if (r1->wm_lp[i] & WM3_LP_EN)
2754                         val_r1 = r1->wm_lp[i] & WM1_LP_LATENCY_MASK;
2755                 if (r2->wm_lp[i] & WM3_LP_EN)
2756                         val_r2 = r2->wm_lp[i] & WM1_LP_LATENCY_MASK;
2757         }
2758
2759         if (val_r1 == val_r2) {
2760                 if (r2->enable_fbc_wm && !r1->enable_fbc_wm)
2761                         return r2;
2762                 else
2763                         return r1;
2764         } else if (val_r1 > val_r2) {
2765                 return r1;
2766         } else {
2767                 return r2;
2768         }
2769 }
2770
2771 /*
2772  * The spec says we shouldn't write when we don't need, because every write
2773  * causes WMs to be re-evaluated, expending some power.
2774  */
2775 static void hsw_write_wm_values(struct drm_i915_private *dev_priv,
2776                                 struct hsw_wm_values *results,
2777                                 enum intel_ddb_partitioning partitioning)
2778 {
2779         struct hsw_wm_values previous;
2780         uint32_t val;
2781         enum intel_ddb_partitioning prev_partitioning;
2782         bool prev_enable_fbc_wm;
2783
2784         previous.wm_pipe[0] = I915_READ(WM0_PIPEA_ILK);
2785         previous.wm_pipe[1] = I915_READ(WM0_PIPEB_ILK);
2786         previous.wm_pipe[2] = I915_READ(WM0_PIPEC_IVB);
2787         previous.wm_lp[0] = I915_READ(WM1_LP_ILK);
2788         previous.wm_lp[1] = I915_READ(WM2_LP_ILK);
2789         previous.wm_lp[2] = I915_READ(WM3_LP_ILK);
2790         previous.wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
2791         previous.wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
2792         previous.wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
2793         previous.wm_linetime[0] = I915_READ(PIPE_WM_LINETIME(PIPE_A));
2794         previous.wm_linetime[1] = I915_READ(PIPE_WM_LINETIME(PIPE_B));
2795         previous.wm_linetime[2] = I915_READ(PIPE_WM_LINETIME(PIPE_C));
2796
2797         prev_partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
2798                                 INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
2799
2800         prev_enable_fbc_wm = !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
2801
2802         if (memcmp(results->wm_pipe, previous.wm_pipe,
2803                    sizeof(results->wm_pipe)) == 0 &&
2804             memcmp(results->wm_lp, previous.wm_lp,
2805                    sizeof(results->wm_lp)) == 0 &&
2806             memcmp(results->wm_lp_spr, previous.wm_lp_spr,
2807                    sizeof(results->wm_lp_spr)) == 0 &&
2808             memcmp(results->wm_linetime, previous.wm_linetime,
2809                    sizeof(results->wm_linetime)) == 0 &&
2810             partitioning == prev_partitioning &&
2811             results->enable_fbc_wm == prev_enable_fbc_wm)
2812                 return;
2813
2814         if (previous.wm_lp[2] != 0)
2815                 I915_WRITE(WM3_LP_ILK, 0);
2816         if (previous.wm_lp[1] != 0)
2817                 I915_WRITE(WM2_LP_ILK, 0);
2818         if (previous.wm_lp[0] != 0)
2819                 I915_WRITE(WM1_LP_ILK, 0);
2820
2821         if (previous.wm_pipe[0] != results->wm_pipe[0])
2822                 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2823         if (previous.wm_pipe[1] != results->wm_pipe[1])
2824                 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
2825         if (previous.wm_pipe[2] != results->wm_pipe[2])
2826                 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
2827
2828         if (previous.wm_linetime[0] != results->wm_linetime[0])
2829                 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
2830         if (previous.wm_linetime[1] != results->wm_linetime[1])
2831                 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
2832         if (previous.wm_linetime[2] != results->wm_linetime[2])
2833                 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
2834
2835         if (prev_partitioning != partitioning) {
2836                 val = I915_READ(WM_MISC);
2837                 if (partitioning == INTEL_DDB_PART_1_2)
2838                         val &= ~WM_MISC_DATA_PARTITION_5_6;
2839                 else
2840                         val |= WM_MISC_DATA_PARTITION_5_6;
2841                 I915_WRITE(WM_MISC, val);
2842         }
2843
2844         if (prev_enable_fbc_wm != results->enable_fbc_wm) {
2845                 val = I915_READ(DISP_ARB_CTL);
2846                 if (results->enable_fbc_wm)
2847                         val &= ~DISP_FBC_WM_DIS;
2848                 else
2849                         val |= DISP_FBC_WM_DIS;
2850                 I915_WRITE(DISP_ARB_CTL, val);
2851         }
2852
2853         if (previous.wm_lp_spr[0] != results->wm_lp_spr[0])
2854                 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
2855         if (previous.wm_lp_spr[1] != results->wm_lp_spr[1])
2856                 I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
2857         if (previous.wm_lp_spr[2] != results->wm_lp_spr[2])
2858                 I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
2859
2860         if (results->wm_lp[0] != 0)
2861                 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
2862         if (results->wm_lp[1] != 0)
2863                 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
2864         if (results->wm_lp[2] != 0)
2865                 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
2866 }
2867
2868 static void haswell_update_wm(struct drm_crtc *crtc)
2869 {
2870         struct drm_device *dev = crtc->dev;
2871         struct drm_i915_private *dev_priv = dev->dev_private;
2872         struct hsw_wm_maximums lp_max_1_2, lp_max_5_6;
2873         struct hsw_pipe_wm_parameters params[3];
2874         struct hsw_wm_values results_1_2, results_5_6, *best_results;
2875         enum intel_ddb_partitioning partitioning;
2876
2877         hsw_compute_wm_parameters(dev, params, &lp_max_1_2, &lp_max_5_6);
2878
2879         hsw_compute_wm_results(dev, params,
2880                                &lp_max_1_2, &results_1_2);
2881         if (lp_max_1_2.pri != lp_max_5_6.pri) {
2882                 hsw_compute_wm_results(dev, params,
2883                                        &lp_max_5_6, &results_5_6);
2884                 best_results = hsw_find_best_result(&results_1_2, &results_5_6);
2885         } else {
2886                 best_results = &results_1_2;
2887         }
2888
2889         partitioning = (best_results == &results_1_2) ?
2890                        INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
2891
2892         hsw_write_wm_values(dev_priv, best_results, partitioning);
2893 }
2894
2895 static void haswell_update_sprite_wm(struct drm_plane *plane,
2896                                      struct drm_crtc *crtc,
2897                                      uint32_t sprite_width, int pixel_size,
2898                                      bool enabled, bool scaled)
2899 {
2900         struct intel_plane *intel_plane = to_intel_plane(plane);
2901
2902         intel_plane->wm.enabled = enabled;
2903         intel_plane->wm.scaled = scaled;
2904         intel_plane->wm.horiz_pixels = sprite_width;
2905         intel_plane->wm.bytes_per_pixel = pixel_size;
2906
2907         haswell_update_wm(crtc);
2908 }
2909
2910 static bool
2911 sandybridge_compute_sprite_wm(struct drm_device *dev, int plane,
2912                               uint32_t sprite_width, int pixel_size,
2913                               const struct intel_watermark_params *display,
2914                               int display_latency_ns, int *sprite_wm)
2915 {
2916         struct drm_crtc *crtc;
2917         int clock;
2918         int entries, tlb_miss;
2919
2920         crtc = intel_get_crtc_for_plane(dev, plane);
2921         if (!intel_crtc_active(crtc)) {
2922                 *sprite_wm = display->guard_size;
2923                 return false;
2924         }
2925
2926         clock = to_intel_crtc(crtc)->config.adjusted_mode.crtc_clock;
2927
2928         /* Use the small buffer method to calculate the sprite watermark */
2929         entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
2930         tlb_miss = display->fifo_size*display->cacheline_size -
2931                 sprite_width * 8;
2932         if (tlb_miss > 0)
2933                 entries += tlb_miss;
2934         entries = DIV_ROUND_UP(entries, display->cacheline_size);
2935         *sprite_wm = entries + display->guard_size;
2936         if (*sprite_wm > (int)display->max_wm)
2937                 *sprite_wm = display->max_wm;
2938
2939         return true;
2940 }
2941
2942 static bool
2943 sandybridge_compute_sprite_srwm(struct drm_device *dev, int plane,
2944                                 uint32_t sprite_width, int pixel_size,
2945                                 const struct intel_watermark_params *display,
2946                                 int latency_ns, int *sprite_wm)
2947 {
2948         struct drm_crtc *crtc;
2949         unsigned long line_time_us;
2950         int clock;
2951         int line_count, line_size;
2952         int small, large;
2953         int entries;
2954
2955         if (!latency_ns) {
2956                 *sprite_wm = 0;
2957                 return false;
2958         }
2959
2960         crtc = intel_get_crtc_for_plane(dev, plane);
2961         clock = to_intel_crtc(crtc)->config.adjusted_mode.crtc_clock;
2962         if (!clock) {
2963                 *sprite_wm = 0;
2964                 return false;
2965         }
2966
2967         line_time_us = (sprite_width * 1000) / clock;
2968         if (!line_time_us) {
2969                 *sprite_wm = 0;
2970                 return false;
2971         }
2972
2973         line_count = (latency_ns / line_time_us + 1000) / 1000;
2974         line_size = sprite_width * pixel_size;
2975
2976         /* Use the minimum of the small and large buffer method for primary */
2977         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
2978         large = line_count * line_size;
2979
2980         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
2981         *sprite_wm = entries + display->guard_size;
2982
2983         return *sprite_wm > 0x3ff ? false : true;
2984 }
2985
2986 static void sandybridge_update_sprite_wm(struct drm_plane *plane,
2987                                          struct drm_crtc *crtc,
2988                                          uint32_t sprite_width, int pixel_size,
2989                                          bool enabled, bool scaled)
2990 {
2991         struct drm_device *dev = plane->dev;
2992         struct drm_i915_private *dev_priv = dev->dev_private;
2993         int pipe = to_intel_plane(plane)->pipe;
2994         int latency = dev_priv->wm.spr_latency[0] * 100;        /* In unit 0.1us */
2995         u32 val;
2996         int sprite_wm, reg;
2997         int ret;
2998
2999         if (!enabled)
3000                 return;
3001
3002         switch (pipe) {
3003         case 0:
3004                 reg = WM0_PIPEA_ILK;
3005                 break;
3006         case 1:
3007                 reg = WM0_PIPEB_ILK;
3008                 break;
3009         case 2:
3010                 reg = WM0_PIPEC_IVB;
3011                 break;
3012         default:
3013                 return; /* bad pipe */
3014         }
3015
3016         ret = sandybridge_compute_sprite_wm(dev, pipe, sprite_width, pixel_size,
3017                                             &sandybridge_display_wm_info,
3018                                             latency, &sprite_wm);
3019         if (!ret) {
3020                 DRM_DEBUG_KMS("failed to compute sprite wm for pipe %c\n",
3021                               pipe_name(pipe));
3022                 return;
3023         }
3024
3025         val = I915_READ(reg);
3026         val &= ~WM0_PIPE_SPRITE_MASK;
3027         I915_WRITE(reg, val | (sprite_wm << WM0_PIPE_SPRITE_SHIFT));
3028         DRM_DEBUG_KMS("sprite watermarks For pipe %c - %d\n", pipe_name(pipe), sprite_wm);
3029
3030
3031         ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
3032                                               pixel_size,
3033                                               &sandybridge_display_srwm_info,
3034                                               dev_priv->wm.spr_latency[1] * 500,
3035                                               &sprite_wm);
3036         if (!ret) {
3037                 DRM_DEBUG_KMS("failed to compute sprite lp1 wm on pipe %c\n",
3038                               pipe_name(pipe));
3039                 return;
3040         }
3041         I915_WRITE(WM1S_LP_ILK, sprite_wm);
3042
3043         /* Only IVB has two more LP watermarks for sprite */
3044         if (!IS_IVYBRIDGE(dev))
3045                 return;
3046
3047         ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
3048                                               pixel_size,
3049                                               &sandybridge_display_srwm_info,
3050                                               dev_priv->wm.spr_latency[2] * 500,
3051                                               &sprite_wm);
3052         if (!ret) {
3053                 DRM_DEBUG_KMS("failed to compute sprite lp2 wm on pipe %c\n",
3054                               pipe_name(pipe));
3055                 return;
3056         }
3057         I915_WRITE(WM2S_LP_IVB, sprite_wm);
3058
3059         ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
3060                                               pixel_size,
3061                                               &sandybridge_display_srwm_info,
3062                                               dev_priv->wm.spr_latency[3] * 500,
3063                                               &sprite_wm);
3064         if (!ret) {
3065                 DRM_DEBUG_KMS("failed to compute sprite lp3 wm on pipe %c\n",
3066                               pipe_name(pipe));
3067                 return;
3068         }
3069         I915_WRITE(WM3S_LP_IVB, sprite_wm);
3070 }
3071
3072 /**
3073  * intel_update_watermarks - update FIFO watermark values based on current modes
3074  *
3075  * Calculate watermark values for the various WM regs based on current mode
3076  * and plane configuration.
3077  *
3078  * There are several cases to deal with here:
3079  *   - normal (i.e. non-self-refresh)
3080  *   - self-refresh (SR) mode
3081  *   - lines are large relative to FIFO size (buffer can hold up to 2)
3082  *   - lines are small relative to FIFO size (buffer can hold more than 2
3083  *     lines), so need to account for TLB latency
3084  *
3085  *   The normal calculation is:
3086  *     watermark = dotclock * bytes per pixel * latency
3087  *   where latency is platform & configuration dependent (we assume pessimal
3088  *   values here).
3089  *
3090  *   The SR calculation is:
3091  *     watermark = (trunc(latency/line time)+1) * surface width *
3092  *       bytes per pixel
3093  *   where
3094  *     line time = htotal / dotclock
3095  *     surface width = hdisplay for normal plane and 64 for cursor
3096  *   and latency is assumed to be high, as above.
3097  *
3098  * The final value programmed to the register should always be rounded up,
3099  * and include an extra 2 entries to account for clock crossings.
3100  *
3101  * We don't use the sprite, so we can ignore that.  And on Crestline we have
3102  * to set the non-SR watermarks to 8.
3103  */
3104 void intel_update_watermarks(struct drm_crtc *crtc)
3105 {
3106         struct drm_i915_private *dev_priv = crtc->dev->dev_private;
3107
3108         if (dev_priv->display.update_wm)
3109                 dev_priv->display.update_wm(crtc);
3110 }
3111
3112 void intel_update_sprite_watermarks(struct drm_plane *plane,
3113                                     struct drm_crtc *crtc,
3114                                     uint32_t sprite_width, int pixel_size,
3115                                     bool enabled, bool scaled)
3116 {
3117         struct drm_i915_private *dev_priv = plane->dev->dev_private;
3118
3119         if (dev_priv->display.update_sprite_wm)
3120                 dev_priv->display.update_sprite_wm(plane, crtc, sprite_width,
3121                                                    pixel_size, enabled, scaled);
3122 }
3123
3124 static struct drm_i915_gem_object *
3125 intel_alloc_context_page(struct drm_device *dev)
3126 {
3127         struct drm_i915_gem_object *ctx;
3128         int ret;
3129
3130         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
3131
3132         ctx = i915_gem_alloc_object(dev, 4096);
3133         if (!ctx) {
3134                 DRM_DEBUG("failed to alloc power context, RC6 disabled\n");
3135                 return NULL;
3136         }
3137
3138         ret = i915_gem_obj_ggtt_pin(ctx, 4096, true, false);
3139         if (ret) {
3140                 DRM_ERROR("failed to pin power context: %d\n", ret);
3141                 goto err_unref;
3142         }
3143
3144         ret = i915_gem_object_set_to_gtt_domain(ctx, 1);
3145         if (ret) {
3146                 DRM_ERROR("failed to set-domain on power context: %d\n", ret);
3147                 goto err_unpin;
3148         }
3149
3150         return ctx;
3151
3152 err_unpin:
3153         i915_gem_object_unpin(ctx);
3154 err_unref:
3155         drm_gem_object_unreference(&ctx->base);
3156         return NULL;
3157 }
3158
3159 /**
3160  * Lock protecting IPS related data structures
3161  */
3162 DEFINE_SPINLOCK(mchdev_lock);
3163
3164 /* Global for IPS driver to get at the current i915 device. Protected by
3165  * mchdev_lock. */
3166 static struct drm_i915_private *i915_mch_dev;
3167
3168 bool ironlake_set_drps(struct drm_device *dev, u8 val)
3169 {
3170         struct drm_i915_private *dev_priv = dev->dev_private;
3171         u16 rgvswctl;
3172
3173         assert_spin_locked(&mchdev_lock);
3174
3175         rgvswctl = I915_READ16(MEMSWCTL);
3176         if (rgvswctl & MEMCTL_CMD_STS) {
3177                 DRM_DEBUG("gpu busy, RCS change rejected\n");
3178                 return false; /* still busy with another command */
3179         }
3180
3181         rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
3182                 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
3183         I915_WRITE16(MEMSWCTL, rgvswctl);
3184         POSTING_READ16(MEMSWCTL);
3185
3186         rgvswctl |= MEMCTL_CMD_STS;
3187         I915_WRITE16(MEMSWCTL, rgvswctl);
3188
3189         return true;
3190 }
3191
3192 static void ironlake_enable_drps(struct drm_device *dev)
3193 {
3194         struct drm_i915_private *dev_priv = dev->dev_private;
3195         u32 rgvmodectl = I915_READ(MEMMODECTL);
3196         u8 fmax, fmin, fstart, vstart;
3197
3198         spin_lock_irq(&mchdev_lock);
3199
3200         /* Enable temp reporting */
3201         I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
3202         I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
3203
3204         /* 100ms RC evaluation intervals */
3205         I915_WRITE(RCUPEI, 100000);
3206         I915_WRITE(RCDNEI, 100000);
3207
3208         /* Set max/min thresholds to 90ms and 80ms respectively */
3209         I915_WRITE(RCBMAXAVG, 90000);
3210         I915_WRITE(RCBMINAVG, 80000);
3211
3212         I915_WRITE(MEMIHYST, 1);
3213
3214         /* Set up min, max, and cur for interrupt handling */
3215         fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
3216         fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
3217         fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
3218                 MEMMODE_FSTART_SHIFT;
3219
3220         vstart = (I915_READ(PXVFREQ_BASE + (fstart * 4)) & PXVFREQ_PX_MASK) >>
3221                 PXVFREQ_PX_SHIFT;
3222
3223         dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
3224         dev_priv->ips.fstart = fstart;
3225
3226         dev_priv->ips.max_delay = fstart;
3227         dev_priv->ips.min_delay = fmin;
3228         dev_priv->ips.cur_delay = fstart;
3229
3230         DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
3231                          fmax, fmin, fstart);
3232
3233         I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
3234
3235         /*
3236          * Interrupts will be enabled in ironlake_irq_postinstall
3237          */
3238
3239         I915_WRITE(VIDSTART, vstart);
3240         POSTING_READ(VIDSTART);
3241
3242         rgvmodectl |= MEMMODE_SWMODE_EN;
3243         I915_WRITE(MEMMODECTL, rgvmodectl);
3244
3245         if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
3246                 DRM_ERROR("stuck trying to change perf mode\n");
3247         mdelay(1);
3248
3249         ironlake_set_drps(dev, fstart);
3250
3251         dev_priv->ips.last_count1 = I915_READ(0x112e4) + I915_READ(0x112e8) +
3252                 I915_READ(0x112e0);
3253         dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
3254         dev_priv->ips.last_count2 = I915_READ(0x112f4);
3255         getrawmonotonic(&dev_priv->ips.last_time2);
3256
3257         spin_unlock_irq(&mchdev_lock);
3258 }
3259
3260 static void ironlake_disable_drps(struct drm_device *dev)
3261 {
3262         struct drm_i915_private *dev_priv = dev->dev_private;
3263         u16 rgvswctl;
3264
3265         spin_lock_irq(&mchdev_lock);
3266
3267         rgvswctl = I915_READ16(MEMSWCTL);
3268
3269         /* Ack interrupts, disable EFC interrupt */
3270         I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
3271         I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
3272         I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
3273         I915_WRITE(DEIIR, DE_PCU_EVENT);
3274         I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
3275
3276         /* Go back to the starting frequency */
3277         ironlake_set_drps(dev, dev_priv->ips.fstart);
3278         mdelay(1);
3279         rgvswctl |= MEMCTL_CMD_STS;
3280         I915_WRITE(MEMSWCTL, rgvswctl);
3281         mdelay(1);
3282
3283         spin_unlock_irq(&mchdev_lock);
3284 }
3285
3286 /* There's a funny hw issue where the hw returns all 0 when reading from
3287  * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
3288  * ourselves, instead of doing a rmw cycle (which might result in us clearing
3289  * all limits and the gpu stuck at whatever frequency it is at atm).
3290  */
3291 static u32 gen6_rps_limits(struct drm_i915_private *dev_priv, u8 *val)
3292 {
3293         u32 limits;
3294
3295         limits = 0;
3296
3297         if (*val >= dev_priv->rps.max_delay)
3298                 *val = dev_priv->rps.max_delay;
3299         limits |= dev_priv->rps.max_delay << 24;
3300
3301         /* Only set the down limit when we've reached the lowest level to avoid
3302          * getting more interrupts, otherwise leave this clear. This prevents a
3303          * race in the hw when coming out of rc6: There's a tiny window where
3304          * the hw runs at the minimal clock before selecting the desired
3305          * frequency, if the down threshold expires in that window we will not
3306          * receive a down interrupt. */
3307         if (*val <= dev_priv->rps.min_delay) {
3308                 *val = dev_priv->rps.min_delay;
3309                 limits |= dev_priv->rps.min_delay << 16;
3310         }
3311
3312         return limits;
3313 }
3314
3315 void gen6_set_rps(struct drm_device *dev, u8 val)
3316 {
3317         struct drm_i915_private *dev_priv = dev->dev_private;
3318         u32 limits = gen6_rps_limits(dev_priv, &val);
3319
3320         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3321         WARN_ON(val > dev_priv->rps.max_delay);
3322         WARN_ON(val < dev_priv->rps.min_delay);
3323
3324         if (val == dev_priv->rps.cur_delay)
3325                 return;
3326
3327         if (IS_HASWELL(dev))
3328                 I915_WRITE(GEN6_RPNSWREQ,
3329                            HSW_FREQUENCY(val));
3330         else
3331                 I915_WRITE(GEN6_RPNSWREQ,
3332                            GEN6_FREQUENCY(val) |
3333                            GEN6_OFFSET(0) |
3334                            GEN6_AGGRESSIVE_TURBO);
3335
3336         /* Make sure we continue to get interrupts
3337          * until we hit the minimum or maximum frequencies.
3338          */
3339         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, limits);
3340
3341         POSTING_READ(GEN6_RPNSWREQ);
3342
3343         dev_priv->rps.cur_delay = val;
3344
3345         trace_intel_gpu_freq_change(val * 50);
3346 }
3347
3348 /*
3349  * Wait until the previous freq change has completed,
3350  * or the timeout elapsed, and then update our notion
3351  * of the current GPU frequency.
3352  */
3353 static void vlv_update_rps_cur_delay(struct drm_i915_private *dev_priv)
3354 {
3355         u32 pval;
3356
3357         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3358
3359         if (wait_for(((pval = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS)) & GENFREQSTATUS) == 0, 10))
3360                 DRM_DEBUG_DRIVER("timed out waiting for Punit\n");
3361
3362         pval >>= 8;
3363
3364         if (pval != dev_priv->rps.cur_delay)
3365                 DRM_DEBUG_DRIVER("Punit overrode GPU freq: %d MHz (%u) requested, but got %d Mhz (%u)\n",
3366                                  vlv_gpu_freq(dev_priv->mem_freq, dev_priv->rps.cur_delay),
3367                                  dev_priv->rps.cur_delay,
3368                                  vlv_gpu_freq(dev_priv->mem_freq, pval), pval);
3369
3370         dev_priv->rps.cur_delay = pval;
3371 }
3372
3373 void valleyview_set_rps(struct drm_device *dev, u8 val)
3374 {
3375         struct drm_i915_private *dev_priv = dev->dev_private;
3376
3377         gen6_rps_limits(dev_priv, &val);
3378
3379         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3380         WARN_ON(val > dev_priv->rps.max_delay);
3381         WARN_ON(val < dev_priv->rps.min_delay);
3382
3383         vlv_update_rps_cur_delay(dev_priv);
3384
3385         DRM_DEBUG_DRIVER("GPU freq request from %d MHz (%u) to %d MHz (%u)\n",
3386                          vlv_gpu_freq(dev_priv->mem_freq,
3387                                       dev_priv->rps.cur_delay),
3388                          dev_priv->rps.cur_delay,
3389                          vlv_gpu_freq(dev_priv->mem_freq, val), val);
3390
3391         if (val == dev_priv->rps.cur_delay)
3392                 return;
3393
3394         vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
3395
3396         dev_priv->rps.cur_delay = val;
3397
3398         trace_intel_gpu_freq_change(vlv_gpu_freq(dev_priv->mem_freq, val));
3399 }
3400
3401 static void gen6_disable_rps_interrupts(struct drm_device *dev)
3402 {
3403         struct drm_i915_private *dev_priv = dev->dev_private;
3404
3405         I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
3406         I915_WRITE(GEN6_PMIER, I915_READ(GEN6_PMIER) & ~GEN6_PM_RPS_EVENTS);
3407         /* Complete PM interrupt masking here doesn't race with the rps work
3408          * item again unmasking PM interrupts because that is using a different
3409          * register (PMIMR) to mask PM interrupts. The only risk is in leaving
3410          * stale bits in PMIIR and PMIMR which gen6_enable_rps will clean up. */
3411
3412         spin_lock_irq(&dev_priv->irq_lock);
3413         dev_priv->rps.pm_iir = 0;
3414         spin_unlock_irq(&dev_priv->irq_lock);
3415
3416         I915_WRITE(GEN6_PMIIR, GEN6_PM_RPS_EVENTS);
3417 }
3418
3419 static void gen6_disable_rps(struct drm_device *dev)
3420 {
3421         struct drm_i915_private *dev_priv = dev->dev_private;
3422
3423         I915_WRITE(GEN6_RC_CONTROL, 0);
3424         I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
3425
3426         gen6_disable_rps_interrupts(dev);
3427 }
3428
3429 static void valleyview_disable_rps(struct drm_device *dev)
3430 {
3431         struct drm_i915_private *dev_priv = dev->dev_private;
3432
3433         I915_WRITE(GEN6_RC_CONTROL, 0);
3434
3435         gen6_disable_rps_interrupts(dev);
3436
3437         if (dev_priv->vlv_pctx) {
3438                 drm_gem_object_unreference(&dev_priv->vlv_pctx->base);
3439                 dev_priv->vlv_pctx = NULL;
3440         }
3441 }
3442
3443 int intel_enable_rc6(const struct drm_device *dev)
3444 {
3445         /* No RC6 before Ironlake */
3446         if (INTEL_INFO(dev)->gen < 5)
3447                 return 0;
3448
3449         /* Respect the kernel parameter if it is set */
3450         if (i915_enable_rc6 >= 0)
3451                 return i915_enable_rc6;
3452
3453         /* Disable RC6 on Ironlake */
3454         if (INTEL_INFO(dev)->gen == 5)
3455                 return 0;
3456
3457         if (IS_HASWELL(dev)) {
3458                 DRM_DEBUG_DRIVER("Haswell: only RC6 available\n");
3459                 return INTEL_RC6_ENABLE;
3460         }
3461
3462         /* snb/ivb have more than one rc6 state. */
3463         if (INTEL_INFO(dev)->gen == 6) {
3464                 DRM_DEBUG_DRIVER("Sandybridge: deep RC6 disabled\n");
3465                 return INTEL_RC6_ENABLE;
3466         }
3467
3468         DRM_DEBUG_DRIVER("RC6 and deep RC6 enabled\n");
3469         return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
3470 }
3471
3472 static void gen6_enable_rps_interrupts(struct drm_device *dev)
3473 {
3474         struct drm_i915_private *dev_priv = dev->dev_private;
3475         u32 enabled_intrs;
3476
3477         spin_lock_irq(&dev_priv->irq_lock);
3478         WARN_ON(dev_priv->rps.pm_iir);
3479         snb_enable_pm_irq(dev_priv, GEN6_PM_RPS_EVENTS);
3480         I915_WRITE(GEN6_PMIIR, GEN6_PM_RPS_EVENTS);
3481         spin_unlock_irq(&dev_priv->irq_lock);
3482
3483         /* only unmask PM interrupts we need. Mask all others. */
3484         enabled_intrs = GEN6_PM_RPS_EVENTS;
3485
3486         /* IVB and SNB hard hangs on looping batchbuffer
3487          * if GEN6_PM_UP_EI_EXPIRED is masked.
3488          */
3489         if (INTEL_INFO(dev)->gen <= 7 && !IS_HASWELL(dev))
3490                 enabled_intrs |= GEN6_PM_RP_UP_EI_EXPIRED;
3491
3492         I915_WRITE(GEN6_PMINTRMSK, ~enabled_intrs);
3493 }
3494
3495 static void gen6_enable_rps(struct drm_device *dev)
3496 {
3497         struct drm_i915_private *dev_priv = dev->dev_private;
3498         struct intel_ring_buffer *ring;
3499         u32 rp_state_cap;
3500         u32 gt_perf_status;
3501         u32 rc6vids, pcu_mbox, rc6_mask = 0;
3502         u32 gtfifodbg;
3503         int rc6_mode;
3504         int i, ret;
3505
3506         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3507
3508         /* Here begins a magic sequence of register writes to enable
3509          * auto-downclocking.
3510          *
3511          * Perhaps there might be some value in exposing these to
3512          * userspace...
3513          */
3514         I915_WRITE(GEN6_RC_STATE, 0);
3515
3516         /* Clear the DBG now so we don't confuse earlier errors */
3517         if ((gtfifodbg = I915_READ(GTFIFODBG))) {
3518                 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
3519                 I915_WRITE(GTFIFODBG, gtfifodbg);
3520         }
3521
3522         gen6_gt_force_wake_get(dev_priv);
3523
3524         rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
3525         gt_perf_status = I915_READ(GEN6_GT_PERF_STATUS);
3526
3527         /* In units of 50MHz */
3528         dev_priv->rps.hw_max = dev_priv->rps.max_delay = rp_state_cap & 0xff;
3529         dev_priv->rps.min_delay = (rp_state_cap & 0xff0000) >> 16;
3530         dev_priv->rps.cur_delay = 0;
3531
3532         /* disable the counters and set deterministic thresholds */
3533         I915_WRITE(GEN6_RC_CONTROL, 0);
3534
3535         I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
3536         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
3537         I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
3538         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
3539         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
3540
3541         for_each_ring(ring, dev_priv, i)
3542                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
3543
3544         I915_WRITE(GEN6_RC_SLEEP, 0);
3545         I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
3546         if (INTEL_INFO(dev)->gen <= 6 || IS_IVYBRIDGE(dev))
3547                 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
3548         else
3549                 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
3550         I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
3551         I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
3552
3553         /* Check if we are enabling RC6 */
3554         rc6_mode = intel_enable_rc6(dev_priv->dev);
3555         if (rc6_mode & INTEL_RC6_ENABLE)
3556                 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
3557
3558         /* We don't use those on Haswell */
3559         if (!IS_HASWELL(dev)) {
3560                 if (rc6_mode & INTEL_RC6p_ENABLE)
3561                         rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
3562
3563                 if (rc6_mode & INTEL_RC6pp_ENABLE)
3564                         rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
3565         }
3566
3567         DRM_INFO("Enabling RC6 states: RC6 %s, RC6p %s, RC6pp %s\n",
3568                         (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off",
3569                         (rc6_mask & GEN6_RC_CTL_RC6p_ENABLE) ? "on" : "off",
3570                         (rc6_mask & GEN6_RC_CTL_RC6pp_ENABLE) ? "on" : "off");
3571
3572         I915_WRITE(GEN6_RC_CONTROL,
3573                    rc6_mask |
3574                    GEN6_RC_CTL_EI_MODE(1) |
3575                    GEN6_RC_CTL_HW_ENABLE);
3576
3577         if (IS_HASWELL(dev)) {
3578                 I915_WRITE(GEN6_RPNSWREQ,
3579                            HSW_FREQUENCY(10));
3580                 I915_WRITE(GEN6_RC_VIDEO_FREQ,
3581                            HSW_FREQUENCY(12));
3582         } else {
3583                 I915_WRITE(GEN6_RPNSWREQ,
3584                            GEN6_FREQUENCY(10) |
3585                            GEN6_OFFSET(0) |
3586                            GEN6_AGGRESSIVE_TURBO);
3587                 I915_WRITE(GEN6_RC_VIDEO_FREQ,
3588                            GEN6_FREQUENCY(12));
3589         }
3590
3591         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
3592         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
3593                    dev_priv->rps.max_delay << 24 |
3594                    dev_priv->rps.min_delay << 16);
3595
3596         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
3597         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
3598         I915_WRITE(GEN6_RP_UP_EI, 66000);
3599         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
3600
3601         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
3602         I915_WRITE(GEN6_RP_CONTROL,
3603                    GEN6_RP_MEDIA_TURBO |
3604                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
3605                    GEN6_RP_MEDIA_IS_GFX |
3606                    GEN6_RP_ENABLE |
3607                    GEN6_RP_UP_BUSY_AVG |
3608                    (IS_HASWELL(dev) ? GEN7_RP_DOWN_IDLE_AVG : GEN6_RP_DOWN_IDLE_CONT));
3609
3610         ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_MIN_FREQ_TABLE, 0);
3611         if (!ret) {
3612                 pcu_mbox = 0;
3613                 ret = sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &pcu_mbox);
3614                 if (!ret && (pcu_mbox & (1<<31))) { /* OC supported */
3615                         DRM_DEBUG_DRIVER("Overclocking supported. Max: %dMHz, Overclock max: %dMHz\n",
3616                                          (dev_priv->rps.max_delay & 0xff) * 50,
3617                                          (pcu_mbox & 0xff) * 50);
3618                         dev_priv->rps.hw_max = pcu_mbox & 0xff;
3619                 }
3620         } else {
3621                 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
3622         }
3623
3624         gen6_set_rps(dev_priv->dev, (gt_perf_status & 0xff00) >> 8);
3625
3626         gen6_enable_rps_interrupts(dev);
3627
3628         rc6vids = 0;
3629         ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
3630         if (IS_GEN6(dev) && ret) {
3631                 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
3632         } else if (IS_GEN6(dev) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
3633                 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
3634                           GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
3635                 rc6vids &= 0xffff00;
3636                 rc6vids |= GEN6_ENCODE_RC6_VID(450);
3637                 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
3638                 if (ret)
3639                         DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
3640         }
3641
3642         gen6_gt_force_wake_put(dev_priv);
3643 }
3644
3645 void gen6_update_ring_freq(struct drm_device *dev)
3646 {
3647         struct drm_i915_private *dev_priv = dev->dev_private;
3648         int min_freq = 15;
3649         unsigned int gpu_freq;
3650         unsigned int max_ia_freq, min_ring_freq;
3651         int scaling_factor = 180;
3652
3653         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3654
3655         max_ia_freq = cpufreq_quick_get_max(0);
3656         /*
3657          * Default to measured freq if none found, PCU will ensure we don't go
3658          * over
3659          */
3660         if (!max_ia_freq)
3661                 max_ia_freq = tsc_khz;
3662
3663         /* Convert from kHz to MHz */
3664         max_ia_freq /= 1000;
3665
3666         min_ring_freq = I915_READ(MCHBAR_MIRROR_BASE_SNB + DCLK);
3667         /* convert DDR frequency from units of 133.3MHz to bandwidth */
3668         min_ring_freq = (2 * 4 * min_ring_freq + 2) / 3;
3669
3670         /*
3671          * For each potential GPU frequency, load a ring frequency we'd like
3672          * to use for memory access.  We do this by specifying the IA frequency
3673          * the PCU should use as a reference to determine the ring frequency.
3674          */
3675         for (gpu_freq = dev_priv->rps.max_delay; gpu_freq >= dev_priv->rps.min_delay;
3676              gpu_freq--) {
3677                 int diff = dev_priv->rps.max_delay - gpu_freq;
3678                 unsigned int ia_freq = 0, ring_freq = 0;
3679
3680                 if (IS_HASWELL(dev)) {
3681                         ring_freq = (gpu_freq * 5 + 3) / 4;
3682                         ring_freq = max(min_ring_freq, ring_freq);
3683                         /* leave ia_freq as the default, chosen by cpufreq */
3684                 } else {
3685                         /* On older processors, there is no separate ring
3686                          * clock domain, so in order to boost the bandwidth
3687                          * of the ring, we need to upclock the CPU (ia_freq).
3688                          *
3689                          * For GPU frequencies less than 750MHz,
3690                          * just use the lowest ring freq.
3691                          */
3692                         if (gpu_freq < min_freq)
3693                                 ia_freq = 800;
3694                         else
3695                                 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
3696                         ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
3697                 }
3698
3699                 sandybridge_pcode_write(dev_priv,
3700                                         GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
3701                                         ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
3702                                         ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
3703                                         gpu_freq);
3704         }
3705 }
3706
3707 int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
3708 {
3709         u32 val, rp0;
3710
3711         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
3712
3713         rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
3714         /* Clamp to max */
3715         rp0 = min_t(u32, rp0, 0xea);
3716
3717         return rp0;
3718 }
3719
3720 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
3721 {
3722         u32 val, rpe;
3723
3724         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
3725         rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
3726         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
3727         rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
3728
3729         return rpe;
3730 }
3731
3732 int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
3733 {
3734         return vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
3735 }
3736
3737 static void vlv_rps_timer_work(struct work_struct *work)
3738 {
3739         drm_i915_private_t *dev_priv = container_of(work, drm_i915_private_t,
3740                                                     rps.vlv_work.work);
3741
3742         /*
3743          * Timer fired, we must be idle.  Drop to min voltage state.
3744          * Note: we use RPe here since it should match the
3745          * Vmin we were shooting for.  That should give us better
3746          * perf when we come back out of RC6 than if we used the
3747          * min freq available.
3748          */
3749         mutex_lock(&dev_priv->rps.hw_lock);
3750         if (dev_priv->rps.cur_delay > dev_priv->rps.rpe_delay)
3751                 valleyview_set_rps(dev_priv->dev, dev_priv->rps.rpe_delay);
3752         mutex_unlock(&dev_priv->rps.hw_lock);
3753 }
3754
3755 static void valleyview_setup_pctx(struct drm_device *dev)
3756 {
3757         struct drm_i915_private *dev_priv = dev->dev_private;
3758         struct drm_i915_gem_object *pctx;
3759         unsigned long pctx_paddr;
3760         u32 pcbr;
3761         int pctx_size = 24*1024;
3762
3763         pcbr = I915_READ(VLV_PCBR);
3764         if (pcbr) {
3765                 /* BIOS set it up already, grab the pre-alloc'd space */
3766                 int pcbr_offset;
3767
3768                 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
3769                 pctx = i915_gem_object_create_stolen_for_preallocated(dev_priv->dev,
3770                                                                       pcbr_offset,
3771                                                                       I915_GTT_OFFSET_NONE,
3772                                                                       pctx_size);
3773                 goto out;
3774         }
3775
3776         /*
3777          * From the Gunit register HAS:
3778          * The Gfx driver is expected to program this register and ensure
3779          * proper allocation within Gfx stolen memory.  For example, this
3780          * register should be programmed such than the PCBR range does not
3781          * overlap with other ranges, such as the frame buffer, protected
3782          * memory, or any other relevant ranges.
3783          */
3784         pctx = i915_gem_object_create_stolen(dev, pctx_size);
3785         if (!pctx) {
3786                 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
3787                 return;
3788         }
3789
3790         pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
3791         I915_WRITE(VLV_PCBR, pctx_paddr);
3792
3793 out:
3794         dev_priv->vlv_pctx = pctx;
3795 }
3796
3797 static void valleyview_enable_rps(struct drm_device *dev)
3798 {
3799         struct drm_i915_private *dev_priv = dev->dev_private;
3800         struct intel_ring_buffer *ring;
3801         u32 gtfifodbg, val, rc6_mode = 0;
3802         int i;
3803
3804         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3805
3806         if ((gtfifodbg = I915_READ(GTFIFODBG))) {
3807                 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
3808                                  gtfifodbg);
3809                 I915_WRITE(GTFIFODBG, gtfifodbg);
3810         }
3811
3812         valleyview_setup_pctx(dev);
3813
3814         gen6_gt_force_wake_get(dev_priv);
3815
3816         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
3817         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
3818         I915_WRITE(GEN6_RP_UP_EI, 66000);
3819         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
3820
3821         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
3822
3823         I915_WRITE(GEN6_RP_CONTROL,
3824                    GEN6_RP_MEDIA_TURBO |
3825                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
3826                    GEN6_RP_MEDIA_IS_GFX |
3827                    GEN6_RP_ENABLE |
3828                    GEN6_RP_UP_BUSY_AVG |
3829                    GEN6_RP_DOWN_IDLE_CONT);
3830
3831         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
3832         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
3833         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
3834
3835         for_each_ring(ring, dev_priv, i)
3836                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
3837
3838         I915_WRITE(GEN6_RC6_THRESHOLD, 0xc350);
3839
3840         /* allows RC6 residency counter to work */
3841         I915_WRITE(VLV_COUNTER_CONTROL,
3842                    _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
3843                                       VLV_MEDIA_RC6_COUNT_EN |
3844                                       VLV_RENDER_RC6_COUNT_EN));
3845         if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
3846                 rc6_mode = GEN7_RC_CTL_TO_MODE;
3847         I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
3848
3849         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
3850         switch ((val >> 6) & 3) {
3851         case 0:
3852         case 1:
3853                 dev_priv->mem_freq = 800;
3854                 break;
3855         case 2:
3856                 dev_priv->mem_freq = 1066;
3857                 break;
3858         case 3:
3859                 dev_priv->mem_freq = 1333;
3860                 break;
3861         }
3862         DRM_DEBUG_DRIVER("DDR speed: %d MHz", dev_priv->mem_freq);
3863
3864         DRM_DEBUG_DRIVER("GPLL enabled? %s\n", val & 0x10 ? "yes" : "no");
3865         DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
3866
3867         dev_priv->rps.cur_delay = (val >> 8) & 0xff;
3868         DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
3869                          vlv_gpu_freq(dev_priv->mem_freq,
3870                                       dev_priv->rps.cur_delay),
3871                          dev_priv->rps.cur_delay);
3872
3873         dev_priv->rps.max_delay = valleyview_rps_max_freq(dev_priv);
3874         dev_priv->rps.hw_max = dev_priv->rps.max_delay;
3875         DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
3876                          vlv_gpu_freq(dev_priv->mem_freq,
3877                                       dev_priv->rps.max_delay),
3878                          dev_priv->rps.max_delay);
3879
3880         dev_priv->rps.rpe_delay = valleyview_rps_rpe_freq(dev_priv);
3881         DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
3882                          vlv_gpu_freq(dev_priv->mem_freq,
3883                                       dev_priv->rps.rpe_delay),
3884                          dev_priv->rps.rpe_delay);
3885
3886         dev_priv->rps.min_delay = valleyview_rps_min_freq(dev_priv);
3887         DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
3888                          vlv_gpu_freq(dev_priv->mem_freq,
3889                                       dev_priv->rps.min_delay),
3890                          dev_priv->rps.min_delay);
3891
3892         DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
3893                          vlv_gpu_freq(dev_priv->mem_freq,
3894                                       dev_priv->rps.rpe_delay),
3895                          dev_priv->rps.rpe_delay);
3896
3897         INIT_DELAYED_WORK(&dev_priv->rps.vlv_work, vlv_rps_timer_work);
3898
3899         valleyview_set_rps(dev_priv->dev, dev_priv->rps.rpe_delay);
3900
3901         gen6_enable_rps_interrupts(dev);
3902
3903         gen6_gt_force_wake_put(dev_priv);
3904 }
3905
3906 void ironlake_teardown_rc6(struct drm_device *dev)
3907 {
3908         struct drm_i915_private *dev_priv = dev->dev_private;
3909
3910         if (dev_priv->ips.renderctx) {
3911                 i915_gem_object_unpin(dev_priv->ips.renderctx);
3912                 drm_gem_object_unreference(&dev_priv->ips.renderctx->base);
3913                 dev_priv->ips.renderctx = NULL;
3914         }
3915
3916         if (dev_priv->ips.pwrctx) {
3917                 i915_gem_object_unpin(dev_priv->ips.pwrctx);
3918                 drm_gem_object_unreference(&dev_priv->ips.pwrctx->base);
3919                 dev_priv->ips.pwrctx = NULL;
3920         }
3921 }
3922
3923 static void ironlake_disable_rc6(struct drm_device *dev)
3924 {
3925         struct drm_i915_private *dev_priv = dev->dev_private;
3926
3927         if (I915_READ(PWRCTXA)) {
3928                 /* Wake the GPU, prevent RC6, then restore RSTDBYCTL */
3929                 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) | RCX_SW_EXIT);
3930                 wait_for(((I915_READ(RSTDBYCTL) & RSX_STATUS_MASK) == RSX_STATUS_ON),
3931                          50);
3932
3933                 I915_WRITE(PWRCTXA, 0);
3934                 POSTING_READ(PWRCTXA);
3935
3936                 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
3937                 POSTING_READ(RSTDBYCTL);
3938         }
3939 }
3940
3941 static int ironlake_setup_rc6(struct drm_device *dev)
3942 {
3943         struct drm_i915_private *dev_priv = dev->dev_private;
3944
3945         if (dev_priv->ips.renderctx == NULL)
3946                 dev_priv->ips.renderctx = intel_alloc_context_page(dev);
3947         if (!dev_priv->ips.renderctx)
3948                 return -ENOMEM;
3949
3950         if (dev_priv->ips.pwrctx == NULL)
3951                 dev_priv->ips.pwrctx = intel_alloc_context_page(dev);
3952         if (!dev_priv->ips.pwrctx) {
3953                 ironlake_teardown_rc6(dev);
3954                 return -ENOMEM;
3955         }
3956
3957         return 0;
3958 }
3959
3960 static void ironlake_enable_rc6(struct drm_device *dev)
3961 {
3962         struct drm_i915_private *dev_priv = dev->dev_private;
3963         struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
3964         bool was_interruptible;
3965         int ret;
3966
3967         /* rc6 disabled by default due to repeated reports of hanging during
3968          * boot and resume.
3969          */
3970         if (!intel_enable_rc6(dev))
3971                 return;
3972
3973         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
3974
3975         ret = ironlake_setup_rc6(dev);
3976         if (ret)
3977                 return;
3978
3979         was_interruptible = dev_priv->mm.interruptible;
3980         dev_priv->mm.interruptible = false;
3981
3982         /*
3983          * GPU can automatically power down the render unit if given a page
3984          * to save state.
3985          */
3986         ret = intel_ring_begin(ring, 6);
3987         if (ret) {
3988                 ironlake_teardown_rc6(dev);
3989                 dev_priv->mm.interruptible = was_interruptible;
3990                 return;
3991         }
3992
3993         intel_ring_emit(ring, MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN);
3994         intel_ring_emit(ring, MI_SET_CONTEXT);
3995         intel_ring_emit(ring, i915_gem_obj_ggtt_offset(dev_priv->ips.renderctx) |
3996                         MI_MM_SPACE_GTT |
3997                         MI_SAVE_EXT_STATE_EN |
3998                         MI_RESTORE_EXT_STATE_EN |
3999                         MI_RESTORE_INHIBIT);
4000         intel_ring_emit(ring, MI_SUSPEND_FLUSH);
4001         intel_ring_emit(ring, MI_NOOP);
4002         intel_ring_emit(ring, MI_FLUSH);
4003         intel_ring_advance(ring);
4004
4005         /*
4006          * Wait for the command parser to advance past MI_SET_CONTEXT. The HW
4007          * does an implicit flush, combined with MI_FLUSH above, it should be
4008          * safe to assume that renderctx is valid
4009          */
4010         ret = intel_ring_idle(ring);
4011         dev_priv->mm.interruptible = was_interruptible;
4012         if (ret) {
4013                 DRM_ERROR("failed to enable ironlake power savings\n");
4014                 ironlake_teardown_rc6(dev);
4015                 return;
4016         }
4017
4018         I915_WRITE(PWRCTXA, i915_gem_obj_ggtt_offset(dev_priv->ips.pwrctx) | PWRCTX_EN);
4019         I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
4020 }
4021
4022 static unsigned long intel_pxfreq(u32 vidfreq)
4023 {
4024         unsigned long freq;
4025         int div = (vidfreq & 0x3f0000) >> 16;
4026         int post = (vidfreq & 0x3000) >> 12;
4027         int pre = (vidfreq & 0x7);
4028
4029         if (!pre)
4030                 return 0;
4031
4032         freq = ((div * 133333) / ((1<<post) * pre));
4033
4034         return freq;
4035 }
4036
4037 static const struct cparams {
4038         u16 i;
4039         u16 t;
4040         u16 m;
4041         u16 c;
4042 } cparams[] = {
4043         { 1, 1333, 301, 28664 },
4044         { 1, 1066, 294, 24460 },
4045         { 1, 800, 294, 25192 },
4046         { 0, 1333, 276, 27605 },
4047         { 0, 1066, 276, 27605 },
4048         { 0, 800, 231, 23784 },
4049 };
4050
4051 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
4052 {
4053         u64 total_count, diff, ret;
4054         u32 count1, count2, count3, m = 0, c = 0;
4055         unsigned long now = jiffies_to_msecs(jiffies), diff1;
4056         int i;
4057
4058         assert_spin_locked(&mchdev_lock);
4059
4060         diff1 = now - dev_priv->ips.last_time1;
4061
4062         /* Prevent division-by-zero if we are asking too fast.
4063          * Also, we don't get interesting results if we are polling
4064          * faster than once in 10ms, so just return the saved value
4065          * in such cases.
4066          */
4067         if (diff1 <= 10)
4068                 return dev_priv->ips.chipset_power;
4069
4070         count1 = I915_READ(DMIEC);
4071         count2 = I915_READ(DDREC);
4072         count3 = I915_READ(CSIEC);
4073
4074         total_count = count1 + count2 + count3;
4075
4076         /* FIXME: handle per-counter overflow */
4077         if (total_count < dev_priv->ips.last_count1) {
4078                 diff = ~0UL - dev_priv->ips.last_count1;
4079                 diff += total_count;
4080         } else {
4081                 diff = total_count - dev_priv->ips.last_count1;
4082         }
4083
4084         for (i = 0; i < ARRAY_SIZE(cparams); i++) {
4085                 if (cparams[i].i == dev_priv->ips.c_m &&
4086                     cparams[i].t == dev_priv->ips.r_t) {
4087                         m = cparams[i].m;
4088                         c = cparams[i].c;
4089                         break;
4090                 }
4091         }
4092
4093         diff = div_u64(diff, diff1);
4094         ret = ((m * diff) + c);
4095         ret = div_u64(ret, 10);
4096
4097         dev_priv->ips.last_count1 = total_count;
4098         dev_priv->ips.last_time1 = now;
4099
4100         dev_priv->ips.chipset_power = ret;
4101
4102         return ret;
4103 }
4104
4105 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
4106 {
4107         unsigned long val;
4108
4109         if (dev_priv->info->gen != 5)
4110                 return 0;
4111
4112         spin_lock_irq(&mchdev_lock);
4113
4114         val = __i915_chipset_val(dev_priv);
4115
4116         spin_unlock_irq(&mchdev_lock);
4117
4118         return val;
4119 }
4120
4121 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
4122 {
4123         unsigned long m, x, b;
4124         u32 tsfs;
4125
4126         tsfs = I915_READ(TSFS);
4127
4128         m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
4129         x = I915_READ8(TR1);
4130
4131         b = tsfs & TSFS_INTR_MASK;
4132
4133         return ((m * x) / 127) - b;
4134 }
4135
4136 static u16 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
4137 {
4138         static const struct v_table {
4139                 u16 vd; /* in .1 mil */
4140                 u16 vm; /* in .1 mil */
4141         } v_table[] = {
4142                 { 0, 0, },
4143                 { 375, 0, },
4144                 { 500, 0, },
4145                 { 625, 0, },
4146                 { 750, 0, },
4147                 { 875, 0, },
4148                 { 1000, 0, },
4149                 { 1125, 0, },
4150                 { 4125, 3000, },
4151                 { 4125, 3000, },
4152                 { 4125, 3000, },
4153                 { 4125, 3000, },
4154                 { 4125, 3000, },
4155                 { 4125, 3000, },
4156                 { 4125, 3000, },
4157                 { 4125, 3000, },
4158                 { 4125, 3000, },
4159                 { 4125, 3000, },
4160                 { 4125, 3000, },
4161                 { 4125, 3000, },
4162                 { 4125, 3000, },
4163                 { 4125, 3000, },
4164                 { 4125, 3000, },
4165                 { 4125, 3000, },
4166                 { 4125, 3000, },
4167                 { 4125, 3000, },
4168                 { 4125, 3000, },
4169                 { 4125, 3000, },
4170                 { 4125, 3000, },
4171                 { 4125, 3000, },
4172                 { 4125, 3000, },
4173                 { 4125, 3000, },
4174                 { 4250, 3125, },
4175                 { 4375, 3250, },
4176                 { 4500, 3375, },
4177                 { 4625, 3500, },
4178                 { 4750, 3625, },
4179                 { 4875, 3750, },
4180                 { 5000, 3875, },
4181                 { 5125, 4000, },
4182                 { 5250, 4125, },
4183                 { 5375, 4250, },
4184                 { 5500, 4375, },
4185                 { 5625, 4500, },
4186                 { 5750, 4625, },
4187                 { 5875, 4750, },
4188                 { 6000, 4875, },
4189                 { 6125, 5000, },
4190                 { 6250, 5125, },
4191                 { 6375, 5250, },
4192                 { 6500, 5375, },
4193                 { 6625, 5500, },
4194                 { 6750, 5625, },
4195                 { 6875, 5750, },
4196                 { 7000, 5875, },
4197                 { 7125, 6000, },
4198                 { 7250, 6125, },
4199                 { 7375, 6250, },
4200                 { 7500, 6375, },
4201                 { 7625, 6500, },
4202                 { 7750, 6625, },
4203                 { 7875, 6750, },
4204                 { 8000, 6875, },
4205                 { 8125, 7000, },
4206                 { 8250, 7125, },
4207                 { 8375, 7250, },
4208                 { 8500, 7375, },
4209                 { 8625, 7500, },
4210                 { 8750, 7625, },
4211                 { 8875, 7750, },
4212                 { 9000, 7875, },
4213                 { 9125, 8000, },
4214                 { 9250, 8125, },
4215                 { 9375, 8250, },
4216                 { 9500, 8375, },
4217                 { 9625, 8500, },
4218                 { 9750, 8625, },
4219                 { 9875, 8750, },
4220                 { 10000, 8875, },
4221                 { 10125, 9000, },
4222                 { 10250, 9125, },
4223                 { 10375, 9250, },
4224                 { 10500, 9375, },
4225                 { 10625, 9500, },
4226                 { 10750, 9625, },
4227                 { 10875, 9750, },
4228                 { 11000, 9875, },
4229                 { 11125, 10000, },
4230                 { 11250, 10125, },
4231                 { 11375, 10250, },
4232                 { 11500, 10375, },
4233                 { 11625, 10500, },
4234                 { 11750, 10625, },
4235                 { 11875, 10750, },
4236                 { 12000, 10875, },
4237                 { 12125, 11000, },
4238                 { 12250, 11125, },
4239                 { 12375, 11250, },
4240                 { 12500, 11375, },
4241                 { 12625, 11500, },
4242                 { 12750, 11625, },
4243                 { 12875, 11750, },
4244                 { 13000, 11875, },
4245                 { 13125, 12000, },
4246                 { 13250, 12125, },
4247                 { 13375, 12250, },
4248                 { 13500, 12375, },
4249                 { 13625, 12500, },
4250                 { 13750, 12625, },
4251                 { 13875, 12750, },
4252                 { 14000, 12875, },
4253                 { 14125, 13000, },
4254                 { 14250, 13125, },
4255                 { 14375, 13250, },
4256                 { 14500, 13375, },
4257                 { 14625, 13500, },
4258                 { 14750, 13625, },
4259                 { 14875, 13750, },
4260                 { 15000, 13875, },
4261                 { 15125, 14000, },
4262                 { 15250, 14125, },
4263                 { 15375, 14250, },
4264                 { 15500, 14375, },
4265                 { 15625, 14500, },
4266                 { 15750, 14625, },
4267                 { 15875, 14750, },
4268                 { 16000, 14875, },
4269                 { 16125, 15000, },
4270         };
4271         if (dev_priv->info->is_mobile)
4272                 return v_table[pxvid].vm;
4273         else
4274                 return v_table[pxvid].vd;
4275 }
4276
4277 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
4278 {
4279         struct timespec now, diff1;
4280         u64 diff;
4281         unsigned long diffms;
4282         u32 count;
4283
4284         assert_spin_locked(&mchdev_lock);
4285
4286         getrawmonotonic(&now);
4287         diff1 = timespec_sub(now, dev_priv->ips.last_time2);
4288
4289         /* Don't divide by 0 */
4290         diffms = diff1.tv_sec * 1000 + diff1.tv_nsec / 1000000;
4291         if (!diffms)
4292                 return;
4293
4294         count = I915_READ(GFXEC);
4295
4296         if (count < dev_priv->ips.last_count2) {
4297                 diff = ~0UL - dev_priv->ips.last_count2;
4298                 diff += count;
4299         } else {
4300                 diff = count - dev_priv->ips.last_count2;
4301         }
4302
4303         dev_priv->ips.last_count2 = count;
4304         dev_priv->ips.last_time2 = now;
4305
4306         /* More magic constants... */
4307         diff = diff * 1181;
4308         diff = div_u64(diff, diffms * 10);
4309         dev_priv->ips.gfx_power = diff;
4310 }
4311
4312 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
4313 {
4314         if (dev_priv->info->gen != 5)
4315                 return;
4316
4317         spin_lock_irq(&mchdev_lock);
4318
4319         __i915_update_gfx_val(dev_priv);
4320
4321         spin_unlock_irq(&mchdev_lock);
4322 }
4323
4324 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
4325 {
4326         unsigned long t, corr, state1, corr2, state2;
4327         u32 pxvid, ext_v;
4328
4329         assert_spin_locked(&mchdev_lock);
4330
4331         pxvid = I915_READ(PXVFREQ_BASE + (dev_priv->rps.cur_delay * 4));
4332         pxvid = (pxvid >> 24) & 0x7f;
4333         ext_v = pvid_to_extvid(dev_priv, pxvid);
4334
4335         state1 = ext_v;
4336
4337         t = i915_mch_val(dev_priv);
4338
4339         /* Revel in the empirically derived constants */
4340
4341         /* Correction factor in 1/100000 units */
4342         if (t > 80)
4343                 corr = ((t * 2349) + 135940);
4344         else if (t >= 50)
4345                 corr = ((t * 964) + 29317);
4346         else /* < 50 */
4347                 corr = ((t * 301) + 1004);
4348
4349         corr = corr * ((150142 * state1) / 10000 - 78642);
4350         corr /= 100000;
4351         corr2 = (corr * dev_priv->ips.corr);
4352
4353         state2 = (corr2 * state1) / 10000;
4354         state2 /= 100; /* convert to mW */
4355
4356         __i915_update_gfx_val(dev_priv);
4357
4358         return dev_priv->ips.gfx_power + state2;
4359 }
4360
4361 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
4362 {
4363         unsigned long val;
4364
4365         if (dev_priv->info->gen != 5)
4366                 return 0;
4367
4368         spin_lock_irq(&mchdev_lock);
4369
4370         val = __i915_gfx_val(dev_priv);
4371
4372         spin_unlock_irq(&mchdev_lock);
4373
4374         return val;
4375 }
4376
4377 /**
4378  * i915_read_mch_val - return value for IPS use
4379  *
4380  * Calculate and return a value for the IPS driver to use when deciding whether
4381  * we have thermal and power headroom to increase CPU or GPU power budget.
4382  */
4383 unsigned long i915_read_mch_val(void)
4384 {
4385         struct drm_i915_private *dev_priv;
4386         unsigned long chipset_val, graphics_val, ret = 0;
4387
4388         spin_lock_irq(&mchdev_lock);
4389         if (!i915_mch_dev)
4390                 goto out_unlock;
4391         dev_priv = i915_mch_dev;
4392
4393         chipset_val = __i915_chipset_val(dev_priv);
4394         graphics_val = __i915_gfx_val(dev_priv);
4395
4396         ret = chipset_val + graphics_val;
4397
4398 out_unlock:
4399         spin_unlock_irq(&mchdev_lock);
4400
4401         return ret;
4402 }
4403 EXPORT_SYMBOL_GPL(i915_read_mch_val);
4404
4405 /**
4406  * i915_gpu_raise - raise GPU frequency limit
4407  *
4408  * Raise the limit; IPS indicates we have thermal headroom.
4409  */
4410 bool i915_gpu_raise(void)
4411 {
4412         struct drm_i915_private *dev_priv;
4413         bool ret = true;
4414
4415         spin_lock_irq(&mchdev_lock);
4416         if (!i915_mch_dev) {
4417                 ret = false;
4418                 goto out_unlock;
4419         }
4420         dev_priv = i915_mch_dev;
4421
4422         if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
4423                 dev_priv->ips.max_delay--;
4424
4425 out_unlock:
4426         spin_unlock_irq(&mchdev_lock);
4427
4428         return ret;
4429 }
4430 EXPORT_SYMBOL_GPL(i915_gpu_raise);
4431
4432 /**
4433  * i915_gpu_lower - lower GPU frequency limit
4434  *
4435  * IPS indicates we're close to a thermal limit, so throttle back the GPU
4436  * frequency maximum.
4437  */
4438 bool i915_gpu_lower(void)
4439 {
4440         struct drm_i915_private *dev_priv;
4441         bool ret = true;
4442
4443         spin_lock_irq(&mchdev_lock);
4444         if (!i915_mch_dev) {
4445                 ret = false;
4446                 goto out_unlock;
4447         }
4448         dev_priv = i915_mch_dev;
4449
4450         if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
4451                 dev_priv->ips.max_delay++;
4452
4453 out_unlock:
4454         spin_unlock_irq(&mchdev_lock);
4455
4456         return ret;
4457 }
4458 EXPORT_SYMBOL_GPL(i915_gpu_lower);
4459
4460 /**
4461  * i915_gpu_busy - indicate GPU business to IPS
4462  *
4463  * Tell the IPS driver whether or not the GPU is busy.
4464  */
4465 bool i915_gpu_busy(void)
4466 {
4467         struct drm_i915_private *dev_priv;
4468         struct intel_ring_buffer *ring;
4469         bool ret = false;
4470         int i;
4471
4472         spin_lock_irq(&mchdev_lock);
4473         if (!i915_mch_dev)
4474                 goto out_unlock;
4475         dev_priv = i915_mch_dev;
4476
4477         for_each_ring(ring, dev_priv, i)
4478                 ret |= !list_empty(&ring->request_list);
4479
4480 out_unlock:
4481         spin_unlock_irq(&mchdev_lock);
4482
4483         return ret;
4484 }
4485 EXPORT_SYMBOL_GPL(i915_gpu_busy);
4486
4487 /**
4488  * i915_gpu_turbo_disable - disable graphics turbo
4489  *
4490  * Disable graphics turbo by resetting the max frequency and setting the
4491  * current frequency to the default.
4492  */
4493 bool i915_gpu_turbo_disable(void)
4494 {
4495         struct drm_i915_private *dev_priv;
4496         bool ret = true;
4497
4498         spin_lock_irq(&mchdev_lock);
4499         if (!i915_mch_dev) {
4500                 ret = false;
4501                 goto out_unlock;
4502         }
4503         dev_priv = i915_mch_dev;
4504
4505         dev_priv->ips.max_delay = dev_priv->ips.fstart;
4506
4507         if (!ironlake_set_drps(dev_priv->dev, dev_priv->ips.fstart))
4508                 ret = false;
4509
4510 out_unlock:
4511         spin_unlock_irq(&mchdev_lock);
4512
4513         return ret;
4514 }
4515 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
4516
4517 /**
4518  * Tells the intel_ips driver that the i915 driver is now loaded, if
4519  * IPS got loaded first.
4520  *
4521  * This awkward dance is so that neither module has to depend on the
4522  * other in order for IPS to do the appropriate communication of
4523  * GPU turbo limits to i915.
4524  */
4525 static void
4526 ips_ping_for_i915_load(void)
4527 {
4528         void (*link)(void);
4529
4530         link = symbol_get(ips_link_to_i915_driver);
4531         if (link) {
4532                 link();
4533                 symbol_put(ips_link_to_i915_driver);
4534         }
4535 }
4536
4537 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
4538 {
4539         /* We only register the i915 ips part with intel-ips once everything is
4540          * set up, to avoid intel-ips sneaking in and reading bogus values. */
4541         spin_lock_irq(&mchdev_lock);
4542         i915_mch_dev = dev_priv;
4543         spin_unlock_irq(&mchdev_lock);
4544
4545         ips_ping_for_i915_load();
4546 }
4547
4548 void intel_gpu_ips_teardown(void)
4549 {
4550         spin_lock_irq(&mchdev_lock);
4551         i915_mch_dev = NULL;
4552         spin_unlock_irq(&mchdev_lock);
4553 }
4554 static void intel_init_emon(struct drm_device *dev)
4555 {
4556         struct drm_i915_private *dev_priv = dev->dev_private;
4557         u32 lcfuse;
4558         u8 pxw[16];
4559         int i;
4560
4561         /* Disable to program */
4562         I915_WRITE(ECR, 0);
4563         POSTING_READ(ECR);
4564
4565         /* Program energy weights for various events */
4566         I915_WRITE(SDEW, 0x15040d00);
4567         I915_WRITE(CSIEW0, 0x007f0000);
4568         I915_WRITE(CSIEW1, 0x1e220004);
4569         I915_WRITE(CSIEW2, 0x04000004);
4570
4571         for (i = 0; i < 5; i++)
4572                 I915_WRITE(PEW + (i * 4), 0);
4573         for (i = 0; i < 3; i++)
4574                 I915_WRITE(DEW + (i * 4), 0);
4575
4576         /* Program P-state weights to account for frequency power adjustment */
4577         for (i = 0; i < 16; i++) {
4578                 u32 pxvidfreq = I915_READ(PXVFREQ_BASE + (i * 4));
4579                 unsigned long freq = intel_pxfreq(pxvidfreq);
4580                 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
4581                         PXVFREQ_PX_SHIFT;
4582                 unsigned long val;
4583
4584                 val = vid * vid;
4585                 val *= (freq / 1000);
4586                 val *= 255;
4587                 val /= (127*127*900);
4588                 if (val > 0xff)
4589                         DRM_ERROR("bad pxval: %ld\n", val);
4590                 pxw[i] = val;
4591         }
4592         /* Render standby states get 0 weight */
4593         pxw[14] = 0;
4594         pxw[15] = 0;
4595
4596         for (i = 0; i < 4; i++) {
4597                 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
4598                         (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
4599                 I915_WRITE(PXW + (i * 4), val);
4600         }
4601
4602         /* Adjust magic regs to magic values (more experimental results) */
4603         I915_WRITE(OGW0, 0);
4604         I915_WRITE(OGW1, 0);
4605         I915_WRITE(EG0, 0x00007f00);
4606         I915_WRITE(EG1, 0x0000000e);
4607         I915_WRITE(EG2, 0x000e0000);
4608         I915_WRITE(EG3, 0x68000300);
4609         I915_WRITE(EG4, 0x42000000);
4610         I915_WRITE(EG5, 0x00140031);
4611         I915_WRITE(EG6, 0);
4612         I915_WRITE(EG7, 0);
4613
4614         for (i = 0; i < 8; i++)
4615                 I915_WRITE(PXWL + (i * 4), 0);
4616
4617         /* Enable PMON + select events */
4618         I915_WRITE(ECR, 0x80000019);
4619
4620         lcfuse = I915_READ(LCFUSE02);
4621
4622         dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
4623 }
4624
4625 void intel_disable_gt_powersave(struct drm_device *dev)
4626 {
4627         struct drm_i915_private *dev_priv = dev->dev_private;
4628
4629         /* Interrupts should be disabled already to avoid re-arming. */
4630         WARN_ON(dev->irq_enabled);
4631
4632         if (IS_IRONLAKE_M(dev)) {
4633                 ironlake_disable_drps(dev);
4634                 ironlake_disable_rc6(dev);
4635         } else if (INTEL_INFO(dev)->gen >= 6) {
4636                 cancel_delayed_work_sync(&dev_priv->rps.delayed_resume_work);
4637                 cancel_work_sync(&dev_priv->rps.work);
4638                 if (IS_VALLEYVIEW(dev))
4639                         cancel_delayed_work_sync(&dev_priv->rps.vlv_work);
4640                 mutex_lock(&dev_priv->rps.hw_lock);
4641                 if (IS_VALLEYVIEW(dev))
4642                         valleyview_disable_rps(dev);
4643                 else
4644                         gen6_disable_rps(dev);
4645                 mutex_unlock(&dev_priv->rps.hw_lock);
4646         }
4647 }
4648
4649 static void intel_gen6_powersave_work(struct work_struct *work)
4650 {
4651         struct drm_i915_private *dev_priv =
4652                 container_of(work, struct drm_i915_private,
4653                              rps.delayed_resume_work.work);
4654         struct drm_device *dev = dev_priv->dev;
4655
4656         mutex_lock(&dev_priv->rps.hw_lock);
4657
4658         if (IS_VALLEYVIEW(dev)) {
4659                 valleyview_enable_rps(dev);
4660         } else {
4661                 gen6_enable_rps(dev);
4662                 gen6_update_ring_freq(dev);
4663         }
4664         mutex_unlock(&dev_priv->rps.hw_lock);
4665 }
4666
4667 void intel_enable_gt_powersave(struct drm_device *dev)
4668 {
4669         struct drm_i915_private *dev_priv = dev->dev_private;
4670
4671         if (IS_IRONLAKE_M(dev)) {
4672                 ironlake_enable_drps(dev);
4673                 ironlake_enable_rc6(dev);
4674                 intel_init_emon(dev);
4675         } else if (IS_GEN6(dev) || IS_GEN7(dev)) {
4676                 /*
4677                  * PCU communication is slow and this doesn't need to be
4678                  * done at any specific time, so do this out of our fast path
4679                  * to make resume and init faster.
4680                  */
4681                 schedule_delayed_work(&dev_priv->rps.delayed_resume_work,
4682                                       round_jiffies_up_relative(HZ));
4683         }
4684 }
4685
4686 static void ibx_init_clock_gating(struct drm_device *dev)
4687 {
4688         struct drm_i915_private *dev_priv = dev->dev_private;
4689
4690         /*
4691          * On Ibex Peak and Cougar Point, we need to disable clock
4692          * gating for the panel power sequencer or it will fail to
4693          * start up when no ports are active.
4694          */
4695         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
4696 }
4697
4698 static void g4x_disable_trickle_feed(struct drm_device *dev)
4699 {
4700         struct drm_i915_private *dev_priv = dev->dev_private;
4701         int pipe;
4702
4703         for_each_pipe(pipe) {
4704                 I915_WRITE(DSPCNTR(pipe),
4705                            I915_READ(DSPCNTR(pipe)) |
4706                            DISPPLANE_TRICKLE_FEED_DISABLE);
4707                 intel_flush_display_plane(dev_priv, pipe);
4708         }
4709 }
4710
4711 static void ironlake_init_clock_gating(struct drm_device *dev)
4712 {
4713         struct drm_i915_private *dev_priv = dev->dev_private;
4714         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
4715
4716         /*
4717          * Required for FBC
4718          * WaFbcDisableDpfcClockGating:ilk
4719          */
4720         dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
4721                    ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
4722                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
4723
4724         I915_WRITE(PCH_3DCGDIS0,
4725                    MARIUNIT_CLOCK_GATE_DISABLE |
4726                    SVSMUNIT_CLOCK_GATE_DISABLE);
4727         I915_WRITE(PCH_3DCGDIS1,
4728                    VFMUNIT_CLOCK_GATE_DISABLE);
4729
4730         /*
4731          * According to the spec the following bits should be set in
4732          * order to enable memory self-refresh
4733          * The bit 22/21 of 0x42004
4734          * The bit 5 of 0x42020
4735          * The bit 15 of 0x45000
4736          */
4737         I915_WRITE(ILK_DISPLAY_CHICKEN2,
4738                    (I915_READ(ILK_DISPLAY_CHICKEN2) |
4739                     ILK_DPARB_GATE | ILK_VSDPFD_FULL));
4740         dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
4741         I915_WRITE(DISP_ARB_CTL,
4742                    (I915_READ(DISP_ARB_CTL) |
4743                     DISP_FBC_WM_DIS));
4744         I915_WRITE(WM3_LP_ILK, 0);
4745         I915_WRITE(WM2_LP_ILK, 0);
4746         I915_WRITE(WM1_LP_ILK, 0);
4747
4748         /*
4749          * Based on the document from hardware guys the following bits
4750          * should be set unconditionally in order to enable FBC.
4751          * The bit 22 of 0x42000
4752          * The bit 22 of 0x42004
4753          * The bit 7,8,9 of 0x42020.
4754          */
4755         if (IS_IRONLAKE_M(dev)) {
4756                 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
4757                 I915_WRITE(ILK_DISPLAY_CHICKEN1,
4758                            I915_READ(ILK_DISPLAY_CHICKEN1) |
4759                            ILK_FBCQ_DIS);
4760                 I915_WRITE(ILK_DISPLAY_CHICKEN2,
4761                            I915_READ(ILK_DISPLAY_CHICKEN2) |
4762                            ILK_DPARB_GATE);
4763         }
4764
4765         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
4766
4767         I915_WRITE(ILK_DISPLAY_CHICKEN2,
4768                    I915_READ(ILK_DISPLAY_CHICKEN2) |
4769                    ILK_ELPIN_409_SELECT);
4770         I915_WRITE(_3D_CHICKEN2,
4771                    _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
4772                    _3D_CHICKEN2_WM_READ_PIPELINED);
4773
4774         /* WaDisableRenderCachePipelinedFlush:ilk */
4775         I915_WRITE(CACHE_MODE_0,
4776                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
4777
4778         g4x_disable_trickle_feed(dev);
4779
4780         ibx_init_clock_gating(dev);
4781 }
4782
4783 static void cpt_init_clock_gating(struct drm_device *dev)
4784 {
4785         struct drm_i915_private *dev_priv = dev->dev_private;
4786         int pipe;
4787         uint32_t val;
4788
4789         /*
4790          * On Ibex Peak and Cougar Point, we need to disable clock
4791          * gating for the panel power sequencer or it will fail to
4792          * start up when no ports are active.
4793          */
4794         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
4795         I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
4796                    DPLS_EDP_PPS_FIX_DIS);
4797         /* The below fixes the weird display corruption, a few pixels shifted
4798          * downward, on (only) LVDS of some HP laptops with IVY.
4799          */
4800         for_each_pipe(pipe) {
4801                 val = I915_READ(TRANS_CHICKEN2(pipe));
4802                 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
4803                 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
4804                 if (dev_priv->vbt.fdi_rx_polarity_inverted)
4805                         val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
4806                 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
4807                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
4808                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
4809                 I915_WRITE(TRANS_CHICKEN2(pipe), val);
4810         }
4811         /* WADP0ClockGatingDisable */
4812         for_each_pipe(pipe) {
4813                 I915_WRITE(TRANS_CHICKEN1(pipe),
4814                            TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
4815         }
4816 }
4817
4818 static void gen6_check_mch_setup(struct drm_device *dev)
4819 {
4820         struct drm_i915_private *dev_priv = dev->dev_private;
4821         uint32_t tmp;
4822
4823         tmp = I915_READ(MCH_SSKPD);
4824         if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL) {
4825                 DRM_INFO("Wrong MCH_SSKPD value: 0x%08x\n", tmp);
4826                 DRM_INFO("This can cause pipe underruns and display issues.\n");
4827                 DRM_INFO("Please upgrade your BIOS to fix this.\n");
4828         }
4829 }
4830
4831 static void gen6_init_clock_gating(struct drm_device *dev)
4832 {
4833         struct drm_i915_private *dev_priv = dev->dev_private;
4834         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
4835
4836         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
4837
4838         I915_WRITE(ILK_DISPLAY_CHICKEN2,
4839                    I915_READ(ILK_DISPLAY_CHICKEN2) |
4840                    ILK_ELPIN_409_SELECT);
4841
4842         /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
4843         I915_WRITE(_3D_CHICKEN,
4844                    _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
4845
4846         /* WaSetupGtModeTdRowDispatch:snb */
4847         if (IS_SNB_GT1(dev))
4848                 I915_WRITE(GEN6_GT_MODE,
4849                            _MASKED_BIT_ENABLE(GEN6_TD_FOUR_ROW_DISPATCH_DISABLE));
4850
4851         I915_WRITE(WM3_LP_ILK, 0);
4852         I915_WRITE(WM2_LP_ILK, 0);
4853         I915_WRITE(WM1_LP_ILK, 0);
4854
4855         I915_WRITE(CACHE_MODE_0,
4856                    _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
4857
4858         I915_WRITE(GEN6_UCGCTL1,
4859                    I915_READ(GEN6_UCGCTL1) |
4860                    GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
4861                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
4862
4863         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
4864          * gating disable must be set.  Failure to set it results in
4865          * flickering pixels due to Z write ordering failures after
4866          * some amount of runtime in the Mesa "fire" demo, and Unigine
4867          * Sanctuary and Tropics, and apparently anything else with
4868          * alpha test or pixel discard.
4869          *
4870          * According to the spec, bit 11 (RCCUNIT) must also be set,
4871          * but we didn't debug actual testcases to find it out.
4872          *
4873          * Also apply WaDisableVDSUnitClockGating:snb and
4874          * WaDisableRCPBUnitClockGating:snb.
4875          */
4876         I915_WRITE(GEN6_UCGCTL2,
4877                    GEN7_VDSUNIT_CLOCK_GATE_DISABLE |
4878                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
4879                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
4880
4881         /* Bspec says we need to always set all mask bits. */
4882         I915_WRITE(_3D_CHICKEN3, (0xFFFF << 16) |
4883                    _3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL);
4884
4885         /*
4886          * According to the spec the following bits should be
4887          * set in order to enable memory self-refresh and fbc:
4888          * The bit21 and bit22 of 0x42000
4889          * The bit21 and bit22 of 0x42004
4890          * The bit5 and bit7 of 0x42020
4891          * The bit14 of 0x70180
4892          * The bit14 of 0x71180
4893          *
4894          * WaFbcAsynchFlipDisableFbcQueue:snb
4895          */
4896         I915_WRITE(ILK_DISPLAY_CHICKEN1,
4897                    I915_READ(ILK_DISPLAY_CHICKEN1) |
4898                    ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
4899         I915_WRITE(ILK_DISPLAY_CHICKEN2,
4900                    I915_READ(ILK_DISPLAY_CHICKEN2) |
4901                    ILK_DPARB_GATE | ILK_VSDPFD_FULL);
4902         I915_WRITE(ILK_DSPCLK_GATE_D,
4903                    I915_READ(ILK_DSPCLK_GATE_D) |
4904                    ILK_DPARBUNIT_CLOCK_GATE_ENABLE  |
4905                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
4906
4907         g4x_disable_trickle_feed(dev);
4908
4909         /* The default value should be 0x200 according to docs, but the two
4910          * platforms I checked have a 0 for this. (Maybe BIOS overrides?) */
4911         I915_WRITE(GEN6_GT_MODE, _MASKED_BIT_DISABLE(0xffff));
4912         I915_WRITE(GEN6_GT_MODE, _MASKED_BIT_ENABLE(GEN6_GT_MODE_HI));
4913
4914         cpt_init_clock_gating(dev);
4915
4916         gen6_check_mch_setup(dev);
4917 }
4918
4919 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
4920 {
4921         uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
4922
4923         reg &= ~GEN7_FF_SCHED_MASK;
4924         reg |= GEN7_FF_TS_SCHED_HW;
4925         reg |= GEN7_FF_VS_SCHED_HW;
4926         reg |= GEN7_FF_DS_SCHED_HW;
4927
4928         if (IS_HASWELL(dev_priv->dev))
4929                 reg &= ~GEN7_FF_VS_REF_CNT_FFME;
4930
4931         I915_WRITE(GEN7_FF_THREAD_MODE, reg);
4932 }
4933
4934 static void lpt_init_clock_gating(struct drm_device *dev)
4935 {
4936         struct drm_i915_private *dev_priv = dev->dev_private;
4937
4938         /*
4939          * TODO: this bit should only be enabled when really needed, then
4940          * disabled when not needed anymore in order to save power.
4941          */
4942         if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE)
4943                 I915_WRITE(SOUTH_DSPCLK_GATE_D,
4944                            I915_READ(SOUTH_DSPCLK_GATE_D) |
4945                            PCH_LP_PARTITION_LEVEL_DISABLE);
4946
4947         /* WADPOClockGatingDisable:hsw */
4948         I915_WRITE(_TRANSA_CHICKEN1,
4949                    I915_READ(_TRANSA_CHICKEN1) |
4950                    TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
4951 }
4952
4953 static void lpt_suspend_hw(struct drm_device *dev)
4954 {
4955         struct drm_i915_private *dev_priv = dev->dev_private;
4956
4957         if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE) {
4958                 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
4959
4960                 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
4961                 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
4962         }
4963 }
4964
4965 static void haswell_init_clock_gating(struct drm_device *dev)
4966 {
4967         struct drm_i915_private *dev_priv = dev->dev_private;
4968
4969         I915_WRITE(WM3_LP_ILK, 0);
4970         I915_WRITE(WM2_LP_ILK, 0);
4971         I915_WRITE(WM1_LP_ILK, 0);
4972
4973         /* According to the spec, bit 13 (RCZUNIT) must be set on IVB.
4974          * This implements the WaDisableRCZUnitClockGating:hsw workaround.
4975          */
4976         I915_WRITE(GEN6_UCGCTL2, GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
4977
4978         /* Apply the WaDisableRHWOOptimizationForRenderHang:hsw workaround. */
4979         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
4980                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
4981
4982         /* WaApplyL3ControlAndL3ChickenMode:hsw */
4983         I915_WRITE(GEN7_L3CNTLREG1,
4984                         GEN7_WA_FOR_GEN7_L3_CONTROL);
4985         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
4986                         GEN7_WA_L3_CHICKEN_MODE);
4987
4988         /* This is required by WaCatErrorRejectionIssue:hsw */
4989         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
4990                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
4991                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
4992
4993         /* WaVSRefCountFullforceMissDisable:hsw */
4994         gen7_setup_fixed_func_scheduler(dev_priv);
4995
4996         /* WaDisable4x2SubspanOptimization:hsw */
4997         I915_WRITE(CACHE_MODE_1,
4998                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
4999
5000         /* WaSwitchSolVfFArbitrationPriority:hsw */
5001         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
5002
5003         /* WaRsPkgCStateDisplayPMReq:hsw */
5004         I915_WRITE(CHICKEN_PAR1_1,
5005                    I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
5006
5007         lpt_init_clock_gating(dev);
5008 }
5009
5010 static void ivybridge_init_clock_gating(struct drm_device *dev)
5011 {
5012         struct drm_i915_private *dev_priv = dev->dev_private;
5013         uint32_t snpcr;
5014
5015         I915_WRITE(WM3_LP_ILK, 0);
5016         I915_WRITE(WM2_LP_ILK, 0);
5017         I915_WRITE(WM1_LP_ILK, 0);
5018
5019         I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
5020
5021         /* WaDisableEarlyCull:ivb */
5022         I915_WRITE(_3D_CHICKEN3,
5023                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
5024
5025         /* WaDisableBackToBackFlipFix:ivb */
5026         I915_WRITE(IVB_CHICKEN3,
5027                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
5028                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
5029
5030         /* WaDisablePSDDualDispatchEnable:ivb */
5031         if (IS_IVB_GT1(dev))
5032                 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
5033                            _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
5034         else
5035                 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1_GT2,
5036                            _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
5037
5038         /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
5039         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
5040                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
5041
5042         /* WaApplyL3ControlAndL3ChickenMode:ivb */
5043         I915_WRITE(GEN7_L3CNTLREG1,
5044                         GEN7_WA_FOR_GEN7_L3_CONTROL);
5045         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
5046                    GEN7_WA_L3_CHICKEN_MODE);
5047         if (IS_IVB_GT1(dev))
5048                 I915_WRITE(GEN7_ROW_CHICKEN2,
5049                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5050         else
5051                 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
5052                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5053
5054
5055         /* WaForceL3Serialization:ivb */
5056         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
5057                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
5058
5059         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
5060          * gating disable must be set.  Failure to set it results in
5061          * flickering pixels due to Z write ordering failures after
5062          * some amount of runtime in the Mesa "fire" demo, and Unigine
5063          * Sanctuary and Tropics, and apparently anything else with
5064          * alpha test or pixel discard.
5065          *
5066          * According to the spec, bit 11 (RCCUNIT) must also be set,
5067          * but we didn't debug actual testcases to find it out.
5068          *
5069          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
5070          * This implements the WaDisableRCZUnitClockGating:ivb workaround.
5071          */
5072         I915_WRITE(GEN6_UCGCTL2,
5073                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE |
5074                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
5075
5076         /* This is required by WaCatErrorRejectionIssue:ivb */
5077         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5078                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5079                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5080
5081         g4x_disable_trickle_feed(dev);
5082
5083         /* WaVSRefCountFullforceMissDisable:ivb */
5084         gen7_setup_fixed_func_scheduler(dev_priv);
5085
5086         /* WaDisable4x2SubspanOptimization:ivb */
5087         I915_WRITE(CACHE_MODE_1,
5088                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5089
5090         snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
5091         snpcr &= ~GEN6_MBC_SNPCR_MASK;
5092         snpcr |= GEN6_MBC_SNPCR_MED;
5093         I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
5094
5095         if (!HAS_PCH_NOP(dev))
5096                 cpt_init_clock_gating(dev);
5097
5098         gen6_check_mch_setup(dev);
5099 }
5100
5101 static void valleyview_init_clock_gating(struct drm_device *dev)
5102 {
5103         struct drm_i915_private *dev_priv = dev->dev_private;
5104
5105         I915_WRITE(DSPCLK_GATE_D, VRHUNIT_CLOCK_GATE_DISABLE);
5106
5107         /* WaDisableEarlyCull:vlv */
5108         I915_WRITE(_3D_CHICKEN3,
5109                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
5110
5111         /* WaDisableBackToBackFlipFix:vlv */
5112         I915_WRITE(IVB_CHICKEN3,
5113                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
5114                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
5115
5116         /* WaDisablePSDDualDispatchEnable:vlv */
5117         I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
5118                    _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
5119                                       GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
5120
5121         /* Apply the WaDisableRHWOOptimizationForRenderHang:vlv workaround. */
5122         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
5123                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
5124
5125         /* WaApplyL3ControlAndL3ChickenMode:vlv */
5126         I915_WRITE(GEN7_L3CNTLREG1, I915_READ(GEN7_L3CNTLREG1) | GEN7_L3AGDIS);
5127         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER, GEN7_WA_L3_CHICKEN_MODE);
5128
5129         /* WaForceL3Serialization:vlv */
5130         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
5131                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
5132
5133         /* WaDisableDopClockGating:vlv */
5134         I915_WRITE(GEN7_ROW_CHICKEN2,
5135                    _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5136
5137         /* This is required by WaCatErrorRejectionIssue:vlv */
5138         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5139                    I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5140                    GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5141
5142         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
5143          * gating disable must be set.  Failure to set it results in
5144          * flickering pixels due to Z write ordering failures after
5145          * some amount of runtime in the Mesa "fire" demo, and Unigine
5146          * Sanctuary and Tropics, and apparently anything else with
5147          * alpha test or pixel discard.
5148          *
5149          * According to the spec, bit 11 (RCCUNIT) must also be set,
5150          * but we didn't debug actual testcases to find it out.
5151          *
5152          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
5153          * This implements the WaDisableRCZUnitClockGating:vlv workaround.
5154          *
5155          * Also apply WaDisableVDSUnitClockGating:vlv and
5156          * WaDisableRCPBUnitClockGating:vlv.
5157          */
5158         I915_WRITE(GEN6_UCGCTL2,
5159                    GEN7_VDSUNIT_CLOCK_GATE_DISABLE |
5160                    GEN7_TDLUNIT_CLOCK_GATE_DISABLE |
5161                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE |
5162                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
5163                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
5164
5165         I915_WRITE(GEN7_UCGCTL4, GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
5166
5167         I915_WRITE(MI_ARB_VLV, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE);
5168
5169         I915_WRITE(CACHE_MODE_1,
5170                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5171
5172         /*
5173          * WaDisableVLVClockGating_VBIIssue:vlv
5174          * Disable clock gating on th GCFG unit to prevent a delay
5175          * in the reporting of vblank events.
5176          */
5177         I915_WRITE(VLV_GUNIT_CLOCK_GATE, 0xffffffff);
5178
5179         /* Conservative clock gating settings for now */
5180         I915_WRITE(0x9400, 0xffffffff);
5181         I915_WRITE(0x9404, 0xffffffff);
5182         I915_WRITE(0x9408, 0xffffffff);
5183         I915_WRITE(0x940c, 0xffffffff);
5184         I915_WRITE(0x9410, 0xffffffff);
5185         I915_WRITE(0x9414, 0xffffffff);
5186         I915_WRITE(0x9418, 0xffffffff);
5187 }
5188
5189 static void g4x_init_clock_gating(struct drm_device *dev)
5190 {
5191         struct drm_i915_private *dev_priv = dev->dev_private;
5192         uint32_t dspclk_gate;
5193
5194         I915_WRITE(RENCLK_GATE_D1, 0);
5195         I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
5196                    GS_UNIT_CLOCK_GATE_DISABLE |
5197                    CL_UNIT_CLOCK_GATE_DISABLE);
5198         I915_WRITE(RAMCLK_GATE_D, 0);
5199         dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
5200                 OVRUNIT_CLOCK_GATE_DISABLE |
5201                 OVCUNIT_CLOCK_GATE_DISABLE;
5202         if (IS_GM45(dev))
5203                 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
5204         I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
5205
5206         /* WaDisableRenderCachePipelinedFlush */
5207         I915_WRITE(CACHE_MODE_0,
5208                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
5209
5210         g4x_disable_trickle_feed(dev);
5211 }
5212
5213 static void crestline_init_clock_gating(struct drm_device *dev)
5214 {
5215         struct drm_i915_private *dev_priv = dev->dev_private;
5216
5217         I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
5218         I915_WRITE(RENCLK_GATE_D2, 0);
5219         I915_WRITE(DSPCLK_GATE_D, 0);
5220         I915_WRITE(RAMCLK_GATE_D, 0);
5221         I915_WRITE16(DEUC, 0);
5222         I915_WRITE(MI_ARB_STATE,
5223                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
5224 }
5225
5226 static void broadwater_init_clock_gating(struct drm_device *dev)
5227 {
5228         struct drm_i915_private *dev_priv = dev->dev_private;
5229
5230         I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
5231                    I965_RCC_CLOCK_GATE_DISABLE |
5232                    I965_RCPB_CLOCK_GATE_DISABLE |
5233                    I965_ISC_CLOCK_GATE_DISABLE |
5234                    I965_FBC_CLOCK_GATE_DISABLE);
5235         I915_WRITE(RENCLK_GATE_D2, 0);
5236         I915_WRITE(MI_ARB_STATE,
5237                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
5238 }
5239
5240 static void gen3_init_clock_gating(struct drm_device *dev)
5241 {
5242         struct drm_i915_private *dev_priv = dev->dev_private;
5243         u32 dstate = I915_READ(D_STATE);
5244
5245         dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
5246                 DSTATE_DOT_CLOCK_GATING;
5247         I915_WRITE(D_STATE, dstate);
5248
5249         if (IS_PINEVIEW(dev))
5250                 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
5251
5252         /* IIR "flip pending" means done if this bit is set */
5253         I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
5254 }
5255
5256 static void i85x_init_clock_gating(struct drm_device *dev)
5257 {
5258         struct drm_i915_private *dev_priv = dev->dev_private;
5259
5260         I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
5261 }
5262
5263 static void i830_init_clock_gating(struct drm_device *dev)
5264 {
5265         struct drm_i915_private *dev_priv = dev->dev_private;
5266
5267         I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
5268 }
5269
5270 void intel_init_clock_gating(struct drm_device *dev)
5271 {
5272         struct drm_i915_private *dev_priv = dev->dev_private;
5273
5274         dev_priv->display.init_clock_gating(dev);
5275 }
5276
5277 void intel_suspend_hw(struct drm_device *dev)
5278 {
5279         if (HAS_PCH_LPT(dev))
5280                 lpt_suspend_hw(dev);
5281 }
5282
5283 /**
5284  * We should only use the power well if we explicitly asked the hardware to
5285  * enable it, so check if it's enabled and also check if we've requested it to
5286  * be enabled.
5287  */
5288 bool intel_display_power_enabled(struct drm_device *dev,
5289                                  enum intel_display_power_domain domain)
5290 {
5291         struct drm_i915_private *dev_priv = dev->dev_private;
5292
5293         if (!HAS_POWER_WELL(dev))
5294                 return true;
5295
5296         switch (domain) {
5297         case POWER_DOMAIN_PIPE_A:
5298         case POWER_DOMAIN_TRANSCODER_EDP:
5299                 return true;
5300         case POWER_DOMAIN_VGA:
5301         case POWER_DOMAIN_PIPE_B:
5302         case POWER_DOMAIN_PIPE_C:
5303         case POWER_DOMAIN_PIPE_A_PANEL_FITTER:
5304         case POWER_DOMAIN_PIPE_B_PANEL_FITTER:
5305         case POWER_DOMAIN_PIPE_C_PANEL_FITTER:
5306         case POWER_DOMAIN_TRANSCODER_A:
5307         case POWER_DOMAIN_TRANSCODER_B:
5308         case POWER_DOMAIN_TRANSCODER_C:
5309                 return I915_READ(HSW_PWR_WELL_DRIVER) ==
5310                      (HSW_PWR_WELL_ENABLE_REQUEST | HSW_PWR_WELL_STATE_ENABLED);
5311         default:
5312                 BUG();
5313         }
5314 }
5315
5316 static void __intel_set_power_well(struct drm_device *dev, bool enable)
5317 {
5318         struct drm_i915_private *dev_priv = dev->dev_private;
5319         bool is_enabled, enable_requested;
5320         uint32_t tmp;
5321
5322         tmp = I915_READ(HSW_PWR_WELL_DRIVER);
5323         is_enabled = tmp & HSW_PWR_WELL_STATE_ENABLED;
5324         enable_requested = tmp & HSW_PWR_WELL_ENABLE_REQUEST;
5325
5326         if (enable) {
5327                 if (!enable_requested)
5328                         I915_WRITE(HSW_PWR_WELL_DRIVER,
5329                                    HSW_PWR_WELL_ENABLE_REQUEST);
5330
5331                 if (!is_enabled) {
5332                         DRM_DEBUG_KMS("Enabling power well\n");
5333                         if (wait_for((I915_READ(HSW_PWR_WELL_DRIVER) &
5334                                       HSW_PWR_WELL_STATE_ENABLED), 20))
5335                                 DRM_ERROR("Timeout enabling power well\n");
5336                 }
5337         } else {
5338                 if (enable_requested) {
5339                         unsigned long irqflags;
5340                         enum pipe p;
5341
5342                         I915_WRITE(HSW_PWR_WELL_DRIVER, 0);
5343                         POSTING_READ(HSW_PWR_WELL_DRIVER);
5344                         DRM_DEBUG_KMS("Requesting to disable the power well\n");
5345
5346                         /*
5347                          * After this, the registers on the pipes that are part
5348                          * of the power well will become zero, so we have to
5349                          * adjust our counters according to that.
5350                          *
5351                          * FIXME: Should we do this in general in
5352                          * drm_vblank_post_modeset?
5353                          */
5354                         spin_lock_irqsave(&dev->vbl_lock, irqflags);
5355                         for_each_pipe(p)
5356                                 if (p != PIPE_A)
5357                                         dev->last_vblank[p] = 0;
5358                         spin_unlock_irqrestore(&dev->vbl_lock, irqflags);
5359                 }
5360         }
5361 }
5362
5363 static void __intel_power_well_get(struct i915_power_well *power_well)
5364 {
5365         if (!power_well->count++)
5366                 __intel_set_power_well(power_well->device, true);
5367 }
5368
5369 static void __intel_power_well_put(struct i915_power_well *power_well)
5370 {
5371         WARN_ON(!power_well->count);
5372         if (!--power_well->count)
5373                 __intel_set_power_well(power_well->device, false);
5374 }
5375
5376 void intel_display_power_get(struct drm_device *dev,
5377                              enum intel_display_power_domain domain)
5378 {
5379         struct drm_i915_private *dev_priv = dev->dev_private;
5380         struct i915_power_well *power_well = &dev_priv->power_well;
5381
5382         if (!HAS_POWER_WELL(dev))
5383                 return;
5384
5385         switch (domain) {
5386         case POWER_DOMAIN_PIPE_A:
5387         case POWER_DOMAIN_TRANSCODER_EDP:
5388                 return;
5389         case POWER_DOMAIN_VGA:
5390         case POWER_DOMAIN_PIPE_B:
5391         case POWER_DOMAIN_PIPE_C:
5392         case POWER_DOMAIN_PIPE_A_PANEL_FITTER:
5393         case POWER_DOMAIN_PIPE_B_PANEL_FITTER:
5394         case POWER_DOMAIN_PIPE_C_PANEL_FITTER:
5395         case POWER_DOMAIN_TRANSCODER_A:
5396         case POWER_DOMAIN_TRANSCODER_B:
5397         case POWER_DOMAIN_TRANSCODER_C:
5398                 spin_lock_irq(&power_well->lock);
5399                 __intel_power_well_get(power_well);
5400                 spin_unlock_irq(&power_well->lock);
5401                 return;
5402         default:
5403                 BUG();
5404         }
5405 }
5406
5407 void intel_display_power_put(struct drm_device *dev,
5408                              enum intel_display_power_domain domain)
5409 {
5410         struct drm_i915_private *dev_priv = dev->dev_private;
5411         struct i915_power_well *power_well = &dev_priv->power_well;
5412
5413         if (!HAS_POWER_WELL(dev))
5414                 return;
5415
5416         switch (domain) {
5417         case POWER_DOMAIN_PIPE_A:
5418         case POWER_DOMAIN_TRANSCODER_EDP:
5419                 return;
5420         case POWER_DOMAIN_VGA:
5421         case POWER_DOMAIN_PIPE_B:
5422         case POWER_DOMAIN_PIPE_C:
5423         case POWER_DOMAIN_PIPE_A_PANEL_FITTER:
5424         case POWER_DOMAIN_PIPE_B_PANEL_FITTER:
5425         case POWER_DOMAIN_PIPE_C_PANEL_FITTER:
5426         case POWER_DOMAIN_TRANSCODER_A:
5427         case POWER_DOMAIN_TRANSCODER_B:
5428         case POWER_DOMAIN_TRANSCODER_C:
5429                 spin_lock_irq(&power_well->lock);
5430                 __intel_power_well_put(power_well);
5431                 spin_unlock_irq(&power_well->lock);
5432                 return;
5433         default:
5434                 BUG();
5435         }
5436 }
5437
5438 static struct i915_power_well *hsw_pwr;
5439
5440 /* Display audio driver power well request */
5441 void i915_request_power_well(void)
5442 {
5443         if (WARN_ON(!hsw_pwr))
5444                 return;
5445
5446         spin_lock_irq(&hsw_pwr->lock);
5447         __intel_power_well_get(hsw_pwr);
5448         spin_unlock_irq(&hsw_pwr->lock);
5449 }
5450 EXPORT_SYMBOL_GPL(i915_request_power_well);
5451
5452 /* Display audio driver power well release */
5453 void i915_release_power_well(void)
5454 {
5455         if (WARN_ON(!hsw_pwr))
5456                 return;
5457
5458         spin_lock_irq(&hsw_pwr->lock);
5459         __intel_power_well_put(hsw_pwr);
5460         spin_unlock_irq(&hsw_pwr->lock);
5461 }
5462 EXPORT_SYMBOL_GPL(i915_release_power_well);
5463
5464 int i915_init_power_well(struct drm_device *dev)
5465 {
5466         struct drm_i915_private *dev_priv = dev->dev_private;
5467
5468         hsw_pwr = &dev_priv->power_well;
5469
5470         hsw_pwr->device = dev;
5471         spin_lock_init(&hsw_pwr->lock);
5472         hsw_pwr->count = 0;
5473
5474         return 0;
5475 }
5476
5477 void i915_remove_power_well(struct drm_device *dev)
5478 {
5479         hsw_pwr = NULL;
5480 }
5481
5482 void intel_set_power_well(struct drm_device *dev, bool enable)
5483 {
5484         struct drm_i915_private *dev_priv = dev->dev_private;
5485         struct i915_power_well *power_well = &dev_priv->power_well;
5486
5487         if (!HAS_POWER_WELL(dev))
5488                 return;
5489
5490         if (!i915_disable_power_well && !enable)
5491                 return;
5492
5493         spin_lock_irq(&power_well->lock);
5494
5495         /*
5496          * This function will only ever contribute one
5497          * to the power well reference count. i915_request
5498          * is what tracks whether we have or have not
5499          * added the one to the reference count.
5500          */
5501         if (power_well->i915_request == enable)
5502                 goto out;
5503
5504         power_well->i915_request = enable;
5505
5506         if (enable)
5507                 __intel_power_well_get(power_well);
5508         else
5509                 __intel_power_well_put(power_well);
5510
5511  out:
5512         spin_unlock_irq(&power_well->lock);
5513 }
5514
5515 void intel_resume_power_well(struct drm_device *dev)
5516 {
5517         struct drm_i915_private *dev_priv = dev->dev_private;
5518         struct i915_power_well *power_well = &dev_priv->power_well;
5519
5520         if (!HAS_POWER_WELL(dev))
5521                 return;
5522
5523         spin_lock_irq(&power_well->lock);
5524         __intel_set_power_well(dev, power_well->count > 0);
5525         spin_unlock_irq(&power_well->lock);
5526 }
5527
5528 /*
5529  * Starting with Haswell, we have a "Power Down Well" that can be turned off
5530  * when not needed anymore. We have 4 registers that can request the power well
5531  * to be enabled, and it will only be disabled if none of the registers is
5532  * requesting it to be enabled.
5533  */
5534 void intel_init_power_well(struct drm_device *dev)
5535 {
5536         struct drm_i915_private *dev_priv = dev->dev_private;
5537
5538         if (!HAS_POWER_WELL(dev))
5539                 return;
5540
5541         /* For now, we need the power well to be always enabled. */
5542         intel_set_power_well(dev, true);
5543         intel_resume_power_well(dev);
5544
5545         /* We're taking over the BIOS, so clear any requests made by it since
5546          * the driver is in charge now. */
5547         if (I915_READ(HSW_PWR_WELL_BIOS) & HSW_PWR_WELL_ENABLE_REQUEST)
5548                 I915_WRITE(HSW_PWR_WELL_BIOS, 0);
5549 }
5550
5551 /* Disables PC8 so we can use the GMBUS and DP AUX interrupts. */
5552 void intel_aux_display_runtime_get(struct drm_i915_private *dev_priv)
5553 {
5554         hsw_disable_package_c8(dev_priv);
5555 }
5556
5557 void intel_aux_display_runtime_put(struct drm_i915_private *dev_priv)
5558 {
5559         hsw_enable_package_c8(dev_priv);
5560 }
5561
5562 /* Set up chip specific power management-related functions */
5563 void intel_init_pm(struct drm_device *dev)
5564 {
5565         struct drm_i915_private *dev_priv = dev->dev_private;
5566
5567         if (I915_HAS_FBC(dev)) {
5568                 if (HAS_PCH_SPLIT(dev)) {
5569                         dev_priv->display.fbc_enabled = ironlake_fbc_enabled;
5570                         if (IS_IVYBRIDGE(dev) || IS_HASWELL(dev))
5571                                 dev_priv->display.enable_fbc =
5572                                         gen7_enable_fbc;
5573                         else
5574                                 dev_priv->display.enable_fbc =
5575                                         ironlake_enable_fbc;
5576                         dev_priv->display.disable_fbc = ironlake_disable_fbc;
5577                 } else if (IS_GM45(dev)) {
5578                         dev_priv->display.fbc_enabled = g4x_fbc_enabled;
5579                         dev_priv->display.enable_fbc = g4x_enable_fbc;
5580                         dev_priv->display.disable_fbc = g4x_disable_fbc;
5581                 } else if (IS_CRESTLINE(dev)) {
5582                         dev_priv->display.fbc_enabled = i8xx_fbc_enabled;
5583                         dev_priv->display.enable_fbc = i8xx_enable_fbc;
5584                         dev_priv->display.disable_fbc = i8xx_disable_fbc;
5585                 }
5586                 /* 855GM needs testing */
5587         }
5588
5589         /* For cxsr */
5590         if (IS_PINEVIEW(dev))
5591                 i915_pineview_get_mem_freq(dev);
5592         else if (IS_GEN5(dev))
5593                 i915_ironlake_get_mem_freq(dev);
5594
5595         /* For FIFO watermark updates */
5596         if (HAS_PCH_SPLIT(dev)) {
5597                 intel_setup_wm_latency(dev);
5598
5599                 if (IS_GEN5(dev)) {
5600                         if (dev_priv->wm.pri_latency[1] &&
5601                             dev_priv->wm.spr_latency[1] &&
5602                             dev_priv->wm.cur_latency[1])
5603                                 dev_priv->display.update_wm = ironlake_update_wm;
5604                         else {
5605                                 DRM_DEBUG_KMS("Failed to get proper latency. "
5606                                               "Disable CxSR\n");
5607                                 dev_priv->display.update_wm = NULL;
5608                         }
5609                         dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
5610                 } else if (IS_GEN6(dev)) {
5611                         if (dev_priv->wm.pri_latency[0] &&
5612                             dev_priv->wm.spr_latency[0] &&
5613                             dev_priv->wm.cur_latency[0]) {
5614                                 dev_priv->display.update_wm = sandybridge_update_wm;
5615                                 dev_priv->display.update_sprite_wm = sandybridge_update_sprite_wm;
5616                         } else {
5617                                 DRM_DEBUG_KMS("Failed to read display plane latency. "
5618                                               "Disable CxSR\n");
5619                                 dev_priv->display.update_wm = NULL;
5620                         }
5621                         dev_priv->display.init_clock_gating = gen6_init_clock_gating;
5622                 } else if (IS_IVYBRIDGE(dev)) {
5623                         if (dev_priv->wm.pri_latency[0] &&
5624                             dev_priv->wm.spr_latency[0] &&
5625                             dev_priv->wm.cur_latency[0]) {
5626                                 dev_priv->display.update_wm = ivybridge_update_wm;
5627                                 dev_priv->display.update_sprite_wm = sandybridge_update_sprite_wm;
5628                         } else {
5629                                 DRM_DEBUG_KMS("Failed to read display plane latency. "
5630                                               "Disable CxSR\n");
5631                                 dev_priv->display.update_wm = NULL;
5632                         }
5633                         dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
5634                 } else if (IS_HASWELL(dev)) {
5635                         if (dev_priv->wm.pri_latency[0] &&
5636                             dev_priv->wm.spr_latency[0] &&
5637                             dev_priv->wm.cur_latency[0]) {
5638                                 dev_priv->display.update_wm = haswell_update_wm;
5639                                 dev_priv->display.update_sprite_wm =
5640                                         haswell_update_sprite_wm;
5641                         } else {
5642                                 DRM_DEBUG_KMS("Failed to read display plane latency. "
5643                                               "Disable CxSR\n");
5644                                 dev_priv->display.update_wm = NULL;
5645                         }
5646                         dev_priv->display.init_clock_gating = haswell_init_clock_gating;
5647                 } else
5648                         dev_priv->display.update_wm = NULL;
5649         } else if (IS_VALLEYVIEW(dev)) {
5650                 dev_priv->display.update_wm = valleyview_update_wm;
5651                 dev_priv->display.init_clock_gating =
5652                         valleyview_init_clock_gating;
5653         } else if (IS_PINEVIEW(dev)) {
5654                 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
5655                                             dev_priv->is_ddr3,
5656                                             dev_priv->fsb_freq,
5657                                             dev_priv->mem_freq)) {
5658                         DRM_INFO("failed to find known CxSR latency "
5659                                  "(found ddr%s fsb freq %d, mem freq %d), "
5660                                  "disabling CxSR\n",
5661                                  (dev_priv->is_ddr3 == 1) ? "3" : "2",
5662                                  dev_priv->fsb_freq, dev_priv->mem_freq);
5663                         /* Disable CxSR and never update its watermark again */
5664                         pineview_disable_cxsr(dev);
5665                         dev_priv->display.update_wm = NULL;
5666                 } else
5667                         dev_priv->display.update_wm = pineview_update_wm;
5668                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
5669         } else if (IS_G4X(dev)) {
5670                 dev_priv->display.update_wm = g4x_update_wm;
5671                 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
5672         } else if (IS_GEN4(dev)) {
5673                 dev_priv->display.update_wm = i965_update_wm;
5674                 if (IS_CRESTLINE(dev))
5675                         dev_priv->display.init_clock_gating = crestline_init_clock_gating;
5676                 else if (IS_BROADWATER(dev))
5677                         dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
5678         } else if (IS_GEN3(dev)) {
5679                 dev_priv->display.update_wm = i9xx_update_wm;
5680                 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
5681                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
5682         } else if (IS_I865G(dev)) {
5683                 dev_priv->display.update_wm = i830_update_wm;
5684                 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
5685                 dev_priv->display.get_fifo_size = i830_get_fifo_size;
5686         } else if (IS_I85X(dev)) {
5687                 dev_priv->display.update_wm = i9xx_update_wm;
5688                 dev_priv->display.get_fifo_size = i85x_get_fifo_size;
5689                 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
5690         } else {
5691                 dev_priv->display.update_wm = i830_update_wm;
5692                 dev_priv->display.init_clock_gating = i830_init_clock_gating;
5693                 if (IS_845G(dev))
5694                         dev_priv->display.get_fifo_size = i845_get_fifo_size;
5695                 else
5696                         dev_priv->display.get_fifo_size = i830_get_fifo_size;
5697         }
5698 }
5699
5700 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u8 mbox, u32 *val)
5701 {
5702         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5703
5704         if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
5705                 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
5706                 return -EAGAIN;
5707         }
5708
5709         I915_WRITE(GEN6_PCODE_DATA, *val);
5710         I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
5711
5712         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
5713                      500)) {
5714                 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
5715                 return -ETIMEDOUT;
5716         }
5717
5718         *val = I915_READ(GEN6_PCODE_DATA);
5719         I915_WRITE(GEN6_PCODE_DATA, 0);
5720
5721         return 0;
5722 }
5723
5724 int sandybridge_pcode_write(struct drm_i915_private *dev_priv, u8 mbox, u32 val)
5725 {
5726         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5727
5728         if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
5729                 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
5730                 return -EAGAIN;
5731         }
5732
5733         I915_WRITE(GEN6_PCODE_DATA, val);
5734         I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
5735
5736         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
5737                      500)) {
5738                 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
5739                 return -ETIMEDOUT;
5740         }
5741
5742         I915_WRITE(GEN6_PCODE_DATA, 0);
5743
5744         return 0;
5745 }
5746
5747 int vlv_gpu_freq(int ddr_freq, int val)
5748 {
5749         int mult, base;
5750
5751         switch (ddr_freq) {
5752         case 800:
5753                 mult = 20;
5754                 base = 120;
5755                 break;
5756         case 1066:
5757                 mult = 22;
5758                 base = 133;
5759                 break;
5760         case 1333:
5761                 mult = 21;
5762                 base = 125;
5763                 break;
5764         default:
5765                 return -1;
5766         }
5767
5768         return ((val - 0xbd) * mult) + base;
5769 }
5770
5771 int vlv_freq_opcode(int ddr_freq, int val)
5772 {
5773         int mult, base;
5774
5775         switch (ddr_freq) {
5776         case 800:
5777                 mult = 20;
5778                 base = 120;
5779                 break;
5780         case 1066:
5781                 mult = 22;
5782                 base = 133;
5783                 break;
5784         case 1333:
5785                 mult = 21;
5786                 base = 125;
5787                 break;
5788         default:
5789                 return -1;
5790         }
5791
5792         val /= mult;
5793         val -= base / mult;
5794         val += 0xbd;
5795
5796         if (val > 0xea)
5797                 val = 0xea;
5798
5799         return val;
5800 }
5801
5802 void intel_pm_init(struct drm_device *dev)
5803 {
5804         struct drm_i915_private *dev_priv = dev->dev_private;
5805
5806         INIT_DELAYED_WORK(&dev_priv->rps.delayed_resume_work,
5807                           intel_gen6_powersave_work);
5808 }
5809