drivers/gpu/drm/i915/intel_pm.c

   1 /*
   2  * Copyright © 2012 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Eugeni Dodonov <eugeni.dodonov@intel.com>
  25  *
  26  */
  27
  28 #include <linux/cpufreq.h>
  29 #include "i915_drv.h"
  30 #include "intel_drv.h"
  31 #include "../../../platform/x86/intel_ips.h"
  32 #include <linux/module.h>
  33 #include <linux/vgaarb.h>
  34 #include <drm/i915_powerwell.h>
  35 #include <linux/pm_runtime.h>
  36
  37 /**
  38  * RC6 is a special power stage which allows the GPU to enter an very
  39  * low-voltage mode when idle, using down to 0V while at this stage.  This
  40  * stage is entered automatically when the GPU is idle when RC6 support is
  41  * enabled, and as soon as new workload arises GPU wakes up automatically as well.
  42  *
  43  * There are different RC6 modes available in Intel GPU, which differentiate
  44  * among each other with the latency required to enter and leave RC6 and
  45  * voltage consumed by the GPU in different states.
  46  *
  47  * The combination of the following flags define which states GPU is allowed
  48  * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
  49  * RC6pp is deepest RC6. Their support by hardware varies according to the
  50  * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
  51  * which brings the most power savings; deeper states save more power, but
  52  * require higher latency to switch to and wake up.
  53  */
  54 #define INTEL_RC6_ENABLE                        (1<<0)
  55 #define INTEL_RC6p_ENABLE                       (1<<1)
  56 #define INTEL_RC6pp_ENABLE                      (1<<2)
  57
  58 /* FBC, or Frame Buffer Compression, is a technique employed to compress the
  59  * framebuffer contents in-memory, aiming at reducing the required bandwidth
  60  * during in-memory transfers and, therefore, reduce the power packet.
  61  *
  62  * The benefits of FBC are mostly visible with solid backgrounds and
  63  * variation-less patterns.
  64  *
  65  * FBC-related functionality can be enabled by the means of the
  66  * i915.i915_enable_fbc parameter
  67  */
  68
  69 static void i8xx_disable_fbc(struct drm_device *dev)
  70 {
  71         struct drm_i915_private *dev_priv = dev->dev_private;
  72         u32 fbc_ctl;
  73
  74         /* Disable compression */
  75         fbc_ctl = I915_READ(FBC_CONTROL);
  76         if ((fbc_ctl & FBC_CTL_EN) == 0)
  77                 return;
  78
  79         fbc_ctl &= ~FBC_CTL_EN;
  80         I915_WRITE(FBC_CONTROL, fbc_ctl);
  81
  82         /* Wait for compressing bit to clear */
  83         if (wait_for((I915_READ(FBC_STATUS) & FBC_STAT_COMPRESSING) == 0, 10)) {
  84                 DRM_DEBUG_KMS("FBC idle timed out\n");
  85                 return;
  86         }
  87
  88         DRM_DEBUG_KMS("disabled FBC\n");
  89 }
  90
  91 static void i8xx_enable_fbc(struct drm_crtc *crtc)
  92 {
  93         struct drm_device *dev = crtc->dev;
  94         struct drm_i915_private *dev_priv = dev->dev_private;
  95         struct drm_framebuffer *fb = crtc->primary->fb;
  96         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
  97         struct drm_i915_gem_object *obj = intel_fb->obj;
  98         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
  99         int cfb_pitch;
 100         int i;
 101         u32 fbc_ctl;
 102
 103         cfb_pitch = dev_priv->fbc.size / FBC_LL_SIZE;
 104         if (fb->pitches[0] < cfb_pitch)
 105                 cfb_pitch = fb->pitches[0];
 106
 107         /* FBC_CTL wants 32B or 64B units */
 108         if (IS_GEN2(dev))
 109                 cfb_pitch = (cfb_pitch / 32) - 1;
 110         else
 111                 cfb_pitch = (cfb_pitch / 64) - 1;
 112
 113         /* Clear old tags */
 114         for (i = 0; i < (FBC_LL_SIZE / 32) + 1; i++)
 115                 I915_WRITE(FBC_TAG + (i * 4), 0);
 116
 117         if (IS_GEN4(dev)) {
 118                 u32 fbc_ctl2;
 119
 120                 /* Set it up... */
 121                 fbc_ctl2 = FBC_CTL_FENCE_DBL | FBC_CTL_IDLE_IMM | FBC_CTL_CPU_FENCE;
 122                 fbc_ctl2 |= FBC_CTL_PLANE(intel_crtc->plane);
 123                 I915_WRITE(FBC_CONTROL2, fbc_ctl2);
 124                 I915_WRITE(FBC_FENCE_OFF, crtc->y);
 125         }
 126
 127         /* enable it... */
 128         fbc_ctl = I915_READ(FBC_CONTROL);
 129         fbc_ctl &= 0x3fff << FBC_CTL_INTERVAL_SHIFT;
 130         fbc_ctl |= FBC_CTL_EN | FBC_CTL_PERIODIC;
 131         if (IS_I945GM(dev))
 132                 fbc_ctl |= FBC_CTL_C3_IDLE; /* 945 needs special SR handling */
 133         fbc_ctl |= (cfb_pitch & 0xff) << FBC_CTL_STRIDE_SHIFT;
 134         fbc_ctl |= obj->fence_reg;
 135         I915_WRITE(FBC_CONTROL, fbc_ctl);
 136
 137         DRM_DEBUG_KMS("enabled FBC, pitch %d, yoff %d, plane %c\n",
 138                       cfb_pitch, crtc->y, plane_name(intel_crtc->plane));
 139 }
 140
 141 static bool i8xx_fbc_enabled(struct drm_device *dev)
 142 {
 143         struct drm_i915_private *dev_priv = dev->dev_private;
 144
 145         return I915_READ(FBC_CONTROL) & FBC_CTL_EN;
 146 }
 147
 148 static void g4x_enable_fbc(struct drm_crtc *crtc)
 149 {
 150         struct drm_device *dev = crtc->dev;
 151         struct drm_i915_private *dev_priv = dev->dev_private;
 152         struct drm_framebuffer *fb = crtc->primary->fb;
 153         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 154         struct drm_i915_gem_object *obj = intel_fb->obj;
 155         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 156         u32 dpfc_ctl;
 157
 158         dpfc_ctl = DPFC_CTL_PLANE(intel_crtc->plane) | DPFC_SR_EN;
 159         if (drm_format_plane_cpp(fb->pixel_format, 0) == 2)
 160                 dpfc_ctl |= DPFC_CTL_LIMIT_2X;
 161         else
 162                 dpfc_ctl |= DPFC_CTL_LIMIT_1X;
 163         dpfc_ctl |= DPFC_CTL_FENCE_EN | obj->fence_reg;
 164
 165         I915_WRITE(DPFC_FENCE_YOFF, crtc->y);
 166
 167         /* enable it... */
 168         I915_WRITE(DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
 169
 170         DRM_DEBUG_KMS("enabled fbc on plane %c\n", plane_name(intel_crtc->plane));
 171 }
 172
 173 static void g4x_disable_fbc(struct drm_device *dev)
 174 {
 175         struct drm_i915_private *dev_priv = dev->dev_private;
 176         u32 dpfc_ctl;
 177
 178         /* Disable compression */
 179         dpfc_ctl = I915_READ(DPFC_CONTROL);
 180         if (dpfc_ctl & DPFC_CTL_EN) {
 181                 dpfc_ctl &= ~DPFC_CTL_EN;
 182                 I915_WRITE(DPFC_CONTROL, dpfc_ctl);
 183
 184                 DRM_DEBUG_KMS("disabled FBC\n");
 185         }
 186 }
 187
 188 static bool g4x_fbc_enabled(struct drm_device *dev)
 189 {
 190         struct drm_i915_private *dev_priv = dev->dev_private;
 191
 192         return I915_READ(DPFC_CONTROL) & DPFC_CTL_EN;
 193 }
 194
 195 static void sandybridge_blit_fbc_update(struct drm_device *dev)
 196 {
 197         struct drm_i915_private *dev_priv = dev->dev_private;
 198         u32 blt_ecoskpd;
 199
 200         /* Make sure blitter notifies FBC of writes */
 201
 202         /* Blitter is part of Media powerwell on VLV. No impact of
 203          * his param in other platforms for now */
 204         gen6_gt_force_wake_get(dev_priv, FORCEWAKE_MEDIA);
 205
 206         blt_ecoskpd = I915_READ(GEN6_BLITTER_ECOSKPD);
 207         blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY <<
 208                 GEN6_BLITTER_LOCK_SHIFT;
 209         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 210         blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY;
 211         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 212         blt_ecoskpd &= ~(GEN6_BLITTER_FBC_NOTIFY <<
 213                          GEN6_BLITTER_LOCK_SHIFT);
 214         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 215         POSTING_READ(GEN6_BLITTER_ECOSKPD);
 216
 217         gen6_gt_force_wake_put(dev_priv, FORCEWAKE_MEDIA);
 218 }
 219
 220 static void ironlake_enable_fbc(struct drm_crtc *crtc)
 221 {
 222         struct drm_device *dev = crtc->dev;
 223         struct drm_i915_private *dev_priv = dev->dev_private;
 224         struct drm_framebuffer *fb = crtc->primary->fb;
 225         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 226         struct drm_i915_gem_object *obj = intel_fb->obj;
 227         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 228         u32 dpfc_ctl;
 229
 230         dpfc_ctl = DPFC_CTL_PLANE(intel_crtc->plane);
 231         if (drm_format_plane_cpp(fb->pixel_format, 0) == 2)
 232                 dpfc_ctl |= DPFC_CTL_LIMIT_2X;
 233         else
 234                 dpfc_ctl |= DPFC_CTL_LIMIT_1X;
 235         dpfc_ctl |= DPFC_CTL_FENCE_EN;
 236         if (IS_GEN5(dev))
 237                 dpfc_ctl |= obj->fence_reg;
 238
 239         I915_WRITE(ILK_DPFC_FENCE_YOFF, crtc->y);
 240         I915_WRITE(ILK_FBC_RT_BASE, i915_gem_obj_ggtt_offset(obj) | ILK_FBC_RT_VALID);
 241         /* enable it... */
 242         I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
 243
 244         if (IS_GEN6(dev)) {
 245                 I915_WRITE(SNB_DPFC_CTL_SA,
 246                            SNB_CPU_FENCE_ENABLE | obj->fence_reg);
 247                 I915_WRITE(DPFC_CPU_FENCE_OFFSET, crtc->y);
 248                 sandybridge_blit_fbc_update(dev);
 249         }
 250
 251         DRM_DEBUG_KMS("enabled fbc on plane %c\n", plane_name(intel_crtc->plane));
 252 }
 253
 254 static void ironlake_disable_fbc(struct drm_device *dev)
 255 {
 256         struct drm_i915_private *dev_priv = dev->dev_private;
 257         u32 dpfc_ctl;
 258
 259         /* Disable compression */
 260         dpfc_ctl = I915_READ(ILK_DPFC_CONTROL);
 261         if (dpfc_ctl & DPFC_CTL_EN) {
 262                 dpfc_ctl &= ~DPFC_CTL_EN;
 263                 I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl);
 264
 265                 DRM_DEBUG_KMS("disabled FBC\n");
 266         }
 267 }
 268
 269 static bool ironlake_fbc_enabled(struct drm_device *dev)
 270 {
 271         struct drm_i915_private *dev_priv = dev->dev_private;
 272
 273         return I915_READ(ILK_DPFC_CONTROL) & DPFC_CTL_EN;
 274 }
 275
 276 static void gen7_enable_fbc(struct drm_crtc *crtc)
 277 {
 278         struct drm_device *dev = crtc->dev;
 279         struct drm_i915_private *dev_priv = dev->dev_private;
 280         struct drm_framebuffer *fb = crtc->primary->fb;
 281         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 282         struct drm_i915_gem_object *obj = intel_fb->obj;
 283         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 284         u32 dpfc_ctl;
 285
 286         dpfc_ctl = IVB_DPFC_CTL_PLANE(intel_crtc->plane);
 287         if (drm_format_plane_cpp(fb->pixel_format, 0) == 2)
 288                 dpfc_ctl |= DPFC_CTL_LIMIT_2X;
 289         else
 290                 dpfc_ctl |= DPFC_CTL_LIMIT_1X;
 291         dpfc_ctl |= IVB_DPFC_CTL_FENCE_EN;
 292
 293         I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
 294
 295         if (IS_IVYBRIDGE(dev)) {
 296                 /* WaFbcAsynchFlipDisableFbcQueue:ivb */
 297                 I915_WRITE(ILK_DISPLAY_CHICKEN1,
 298                            I915_READ(ILK_DISPLAY_CHICKEN1) |
 299                            ILK_FBCQ_DIS);
 300         } else {
 301                 /* WaFbcAsynchFlipDisableFbcQueue:hsw,bdw */
 302                 I915_WRITE(CHICKEN_PIPESL_1(intel_crtc->pipe),
 303                            I915_READ(CHICKEN_PIPESL_1(intel_crtc->pipe)) |
 304                            HSW_FBCQ_DIS);
 305         }
 306
 307         I915_WRITE(SNB_DPFC_CTL_SA,
 308                    SNB_CPU_FENCE_ENABLE | obj->fence_reg);
 309         I915_WRITE(DPFC_CPU_FENCE_OFFSET, crtc->y);
 310
 311         sandybridge_blit_fbc_update(dev);
 312
 313         DRM_DEBUG_KMS("enabled fbc on plane %c\n", plane_name(intel_crtc->plane));
 314 }
 315
 316 bool intel_fbc_enabled(struct drm_device *dev)
 317 {
 318         struct drm_i915_private *dev_priv = dev->dev_private;
 319
 320         if (!dev_priv->display.fbc_enabled)
 321                 return false;
 322
 323         return dev_priv->display.fbc_enabled(dev);
 324 }
 325
 326 static void intel_fbc_work_fn(struct work_struct *__work)
 327 {
 328         struct intel_fbc_work *work =
 329                 container_of(to_delayed_work(__work),
 330                              struct intel_fbc_work, work);
 331         struct drm_device *dev = work->crtc->dev;
 332         struct drm_i915_private *dev_priv = dev->dev_private;
 333
 334         mutex_lock(&dev->struct_mutex);
 335         if (work == dev_priv->fbc.fbc_work) {
 336                 /* Double check that we haven't switched fb without cancelling
 337                  * the prior work.
 338                  */
 339                 if (work->crtc->primary->fb == work->fb) {
 340                         dev_priv->display.enable_fbc(work->crtc);
 341
 342                         dev_priv->fbc.plane = to_intel_crtc(work->crtc)->plane;
 343                         dev_priv->fbc.fb_id = work->crtc->primary->fb->base.id;
 344                         dev_priv->fbc.y = work->crtc->y;
 345                 }
 346
 347                 dev_priv->fbc.fbc_work = NULL;
 348         }
 349         mutex_unlock(&dev->struct_mutex);
 350
 351         kfree(work);
 352 }
 353
 354 static void intel_cancel_fbc_work(struct drm_i915_private *dev_priv)
 355 {
 356         if (dev_priv->fbc.fbc_work == NULL)
 357                 return;
 358
 359         DRM_DEBUG_KMS("cancelling pending FBC enable\n");
 360
 361         /* Synchronisation is provided by struct_mutex and checking of
 362          * dev_priv->fbc.fbc_work, so we can perform the cancellation
 363          * entirely asynchronously.
 364          */
 365         if (cancel_delayed_work(&dev_priv->fbc.fbc_work->work))
 366                 /* tasklet was killed before being run, clean up */
 367                 kfree(dev_priv->fbc.fbc_work);
 368
 369         /* Mark the work as no longer wanted so that if it does
 370          * wake-up (because the work was already running and waiting
 371          * for our mutex), it will discover that is no longer
 372          * necessary to run.
 373          */
 374         dev_priv->fbc.fbc_work = NULL;
 375 }
 376
 377 static void intel_enable_fbc(struct drm_crtc *crtc)
 378 {
 379         struct intel_fbc_work *work;
 380         struct drm_device *dev = crtc->dev;
 381         struct drm_i915_private *dev_priv = dev->dev_private;
 382
 383         if (!dev_priv->display.enable_fbc)
 384                 return;
 385
 386         intel_cancel_fbc_work(dev_priv);
 387
 388         work = kzalloc(sizeof(*work), GFP_KERNEL);
 389         if (work == NULL) {
 390                 DRM_ERROR("Failed to allocate FBC work structure\n");
 391                 dev_priv->display.enable_fbc(crtc);
 392                 return;
 393         }
 394
 395         work->crtc = crtc;
 396         work->fb = crtc->primary->fb;
 397         INIT_DELAYED_WORK(&work->work, intel_fbc_work_fn);
 398
 399         dev_priv->fbc.fbc_work = work;
 400
 401         /* Delay the actual enabling to let pageflipping cease and the
 402          * display to settle before starting the compression. Note that
 403          * this delay also serves a second purpose: it allows for a
 404          * vblank to pass after disabling the FBC before we attempt
 405          * to modify the control registers.
 406          *
 407          * A more complicated solution would involve tracking vblanks
 408          * following the termination of the page-flipping sequence
 409          * and indeed performing the enable as a co-routine and not
 410          * waiting synchronously upon the vblank.
 411          *
 412          * WaFbcWaitForVBlankBeforeEnable:ilk,snb
 413          */
 414         schedule_delayed_work(&work->work, msecs_to_jiffies(50));
 415 }
 416
 417 void intel_disable_fbc(struct drm_device *dev)
 418 {
 419         struct drm_i915_private *dev_priv = dev->dev_private;
 420
 421         intel_cancel_fbc_work(dev_priv);
 422
 423         if (!dev_priv->display.disable_fbc)
 424                 return;
 425
 426         dev_priv->display.disable_fbc(dev);
 427         dev_priv->fbc.plane = -1;
 428 }
 429
 430 static bool set_no_fbc_reason(struct drm_i915_private *dev_priv,
 431                               enum no_fbc_reason reason)
 432 {
 433         if (dev_priv->fbc.no_fbc_reason == reason)
 434                 return false;
 435
 436         dev_priv->fbc.no_fbc_reason = reason;
 437         return true;
 438 }
 439
 440 /**
 441  * intel_update_fbc - enable/disable FBC as needed
 442  * @dev: the drm_device
 443  *
 444  * Set up the framebuffer compression hardware at mode set time.  We
 445  * enable it if possible:
 446  *   - plane A only (on pre-965)
 447  *   - no pixel mulitply/line duplication
 448  *   - no alpha buffer discard
 449  *   - no dual wide
 450  *   - framebuffer <= max_hdisplay in width, max_vdisplay in height
 451  *
 452  * We can't assume that any compression will take place (worst case),
 453  * so the compressed buffer has to be the same size as the uncompressed
 454  * one.  It also must reside (along with the line length buffer) in
 455  * stolen memory.
 456  *
 457  * We need to enable/disable FBC on a global basis.
 458  */
 459 void intel_update_fbc(struct drm_device *dev)
 460 {
 461         struct drm_i915_private *dev_priv = dev->dev_private;
 462         struct drm_crtc *crtc = NULL, *tmp_crtc;
 463         struct intel_crtc *intel_crtc;
 464         struct drm_framebuffer *fb;
 465         struct intel_framebuffer *intel_fb;
 466         struct drm_i915_gem_object *obj;
 467         const struct drm_display_mode *adjusted_mode;
 468         unsigned int max_width, max_height;
 469
 470         if (!HAS_FBC(dev)) {
 471                 set_no_fbc_reason(dev_priv, FBC_UNSUPPORTED);
 472                 return;
 473         }
 474
 475         if (!i915.powersave) {
 476                 if (set_no_fbc_reason(dev_priv, FBC_MODULE_PARAM))
 477                         DRM_DEBUG_KMS("fbc disabled per module param\n");
 478                 return;
 479         }
 480
 481         /*
 482          * If FBC is already on, we just have to verify that we can
 483          * keep it that way...
 484          * Need to disable if:
 485          *   - more than one pipe is active
 486          *   - changing FBC params (stride, fence, mode)
 487          *   - new fb is too large to fit in compressed buffer
 488          *   - going to an unsupported config (interlace, pixel multiply, etc.)
 489          */
 490         for_each_crtc(dev, tmp_crtc) {
 491                 if (intel_crtc_active(tmp_crtc) &&
 492                     to_intel_crtc(tmp_crtc)->primary_enabled) {
 493                         if (crtc) {
 494                                 if (set_no_fbc_reason(dev_priv, FBC_MULTIPLE_PIPES))
 495                                         DRM_DEBUG_KMS("more than one pipe active, disabling compression\n");
 496                                 goto out_disable;
 497                         }
 498                         crtc = tmp_crtc;
 499                 }
 500         }
 501
 502         if (!crtc || crtc->primary->fb == NULL) {
 503                 if (set_no_fbc_reason(dev_priv, FBC_NO_OUTPUT))
 504                         DRM_DEBUG_KMS("no output, disabling\n");
 505                 goto out_disable;
 506         }
 507
 508         intel_crtc = to_intel_crtc(crtc);
 509         fb = crtc->primary->fb;
 510         intel_fb = to_intel_framebuffer(fb);
 511         obj = intel_fb->obj;
 512         adjusted_mode = &intel_crtc->config.adjusted_mode;
 513
 514         if (i915.enable_fbc < 0 &&
 515             INTEL_INFO(dev)->gen <= 7 && !IS_HASWELL(dev)) {
 516                 if (set_no_fbc_reason(dev_priv, FBC_CHIP_DEFAULT))
 517                         DRM_DEBUG_KMS("disabled per chip default\n");
 518                 goto out_disable;
 519         }
 520         if (!i915.enable_fbc) {
 521                 if (set_no_fbc_reason(dev_priv, FBC_MODULE_PARAM))
 522                         DRM_DEBUG_KMS("fbc disabled per module param\n");
 523                 goto out_disable;
 524         }
 525         if ((adjusted_mode->flags & DRM_MODE_FLAG_INTERLACE) ||
 526             (adjusted_mode->flags & DRM_MODE_FLAG_DBLSCAN)) {
 527                 if (set_no_fbc_reason(dev_priv, FBC_UNSUPPORTED_MODE))
 528                         DRM_DEBUG_KMS("mode incompatible with compression, "
 529                                       "disabling\n");
 530                 goto out_disable;
 531         }
 532
 533         if (IS_G4X(dev) || INTEL_INFO(dev)->gen >= 5) {
 534                 max_width = 4096;
 535                 max_height = 2048;
 536         } else {
 537                 max_width = 2048;
 538                 max_height = 1536;
 539         }
 540         if (intel_crtc->config.pipe_src_w > max_width ||
 541             intel_crtc->config.pipe_src_h > max_height) {
 542                 if (set_no_fbc_reason(dev_priv, FBC_MODE_TOO_LARGE))
 543                         DRM_DEBUG_KMS("mode too large for compression, disabling\n");
 544                 goto out_disable;
 545         }
 546         if ((INTEL_INFO(dev)->gen < 4 || HAS_DDI(dev)) &&
 547             intel_crtc->plane != PLANE_A) {
 548                 if (set_no_fbc_reason(dev_priv, FBC_BAD_PLANE))
 549                         DRM_DEBUG_KMS("plane not A, disabling compression\n");
 550                 goto out_disable;
 551         }
 552
 553         /* The use of a CPU fence is mandatory in order to detect writes
 554          * by the CPU to the scanout and trigger updates to the FBC.
 555          */
 556         if (obj->tiling_mode != I915_TILING_X ||
 557             obj->fence_reg == I915_FENCE_REG_NONE) {
 558                 if (set_no_fbc_reason(dev_priv, FBC_NOT_TILED))
 559                         DRM_DEBUG_KMS("framebuffer not tiled or fenced, disabling compression\n");
 560                 goto out_disable;
 561         }
 562
 563         /* If the kernel debugger is active, always disable compression */
 564         if (in_dbg_master())
 565                 goto out_disable;
 566
 567         if (i915_gem_stolen_setup_compression(dev, intel_fb->obj->base.size)) {
 568                 if (set_no_fbc_reason(dev_priv, FBC_STOLEN_TOO_SMALL))
 569                         DRM_DEBUG_KMS("framebuffer too large, disabling compression\n");
 570                 goto out_disable;
 571         }
 572
 573         /* If the scanout has not changed, don't modify the FBC settings.
 574          * Note that we make the fundamental assumption that the fb->obj
 575          * cannot be unpinned (and have its GTT offset and fence revoked)
 576          * without first being decoupled from the scanout and FBC disabled.
 577          */
 578         if (dev_priv->fbc.plane == intel_crtc->plane &&
 579             dev_priv->fbc.fb_id == fb->base.id &&
 580             dev_priv->fbc.y == crtc->y)
 581                 return;
 582
 583         if (intel_fbc_enabled(dev)) {
 584                 /* We update FBC along two paths, after changing fb/crtc
 585                  * configuration (modeswitching) and after page-flipping
 586                  * finishes. For the latter, we know that not only did
 587                  * we disable the FBC at the start of the page-flip
 588                  * sequence, but also more than one vblank has passed.
 589                  *
 590                  * For the former case of modeswitching, it is possible
 591                  * to switch between two FBC valid configurations
 592                  * instantaneously so we do need to disable the FBC
 593                  * before we can modify its control registers. We also
 594                  * have to wait for the next vblank for that to take
 595                  * effect. However, since we delay enabling FBC we can
 596                  * assume that a vblank has passed since disabling and
 597                  * that we can safely alter the registers in the deferred
 598                  * callback.
 599                  *
 600                  * In the scenario that we go from a valid to invalid
 601                  * and then back to valid FBC configuration we have
 602                  * no strict enforcement that a vblank occurred since
 603                  * disabling the FBC. However, along all current pipe
 604                  * disabling paths we do need to wait for a vblank at
 605                  * some point. And we wait before enabling FBC anyway.
 606                  */
 607                 DRM_DEBUG_KMS("disabling active FBC for update\n");
 608                 intel_disable_fbc(dev);
 609         }
 610
 611         intel_enable_fbc(crtc);
 612         dev_priv->fbc.no_fbc_reason = FBC_OK;
 613         return;
 614
 615 out_disable:
 616         /* Multiple disables should be harmless */
 617         if (intel_fbc_enabled(dev)) {
 618                 DRM_DEBUG_KMS("unsupported config, disabling FBC\n");
 619                 intel_disable_fbc(dev);
 620         }
 621         i915_gem_stolen_cleanup_compression(dev);
 622 }
 623
 624 static void i915_pineview_get_mem_freq(struct drm_device *dev)
 625 {
 626         struct drm_i915_private *dev_priv = dev->dev_private;
 627         u32 tmp;
 628
 629         tmp = I915_READ(CLKCFG);
 630
 631         switch (tmp & CLKCFG_FSB_MASK) {
 632         case CLKCFG_FSB_533:
 633                 dev_priv->fsb_freq = 533; /* 133*4 */
 634                 break;
 635         case CLKCFG_FSB_800:
 636                 dev_priv->fsb_freq = 800; /* 200*4 */
 637                 break;
 638         case CLKCFG_FSB_667:
 639                 dev_priv->fsb_freq =  667; /* 167*4 */
 640                 break;
 641         case CLKCFG_FSB_400:
 642                 dev_priv->fsb_freq = 400; /* 100*4 */
 643                 break;
 644         }
 645
 646         switch (tmp & CLKCFG_MEM_MASK) {
 647         case CLKCFG_MEM_533:
 648                 dev_priv->mem_freq = 533;
 649                 break;
 650         case CLKCFG_MEM_667:
 651                 dev_priv->mem_freq = 667;
 652                 break;
 653         case CLKCFG_MEM_800:
 654                 dev_priv->mem_freq = 800;
 655                 break;
 656         }
 657
 658         /* detect pineview DDR3 setting */
 659         tmp = I915_READ(CSHRDDR3CTL);
 660         dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
 661 }
 662
 663 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
 664 {
 665         struct drm_i915_private *dev_priv = dev->dev_private;
 666         u16 ddrpll, csipll;
 667
 668         ddrpll = I915_READ16(DDRMPLL1);
 669         csipll = I915_READ16(CSIPLL0);
 670
 671         switch (ddrpll & 0xff) {
 672         case 0xc:
 673                 dev_priv->mem_freq = 800;
 674                 break;
 675         case 0x10:
 676                 dev_priv->mem_freq = 1066;
 677                 break;
 678         case 0x14:
 679                 dev_priv->mem_freq = 1333;
 680                 break;
 681         case 0x18:
 682                 dev_priv->mem_freq = 1600;
 683                 break;
 684         default:
 685                 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
 686                                  ddrpll & 0xff);
 687                 dev_priv->mem_freq = 0;
 688                 break;
 689         }
 690
 691         dev_priv->ips.r_t = dev_priv->mem_freq;
 692
 693         switch (csipll & 0x3ff) {
 694         case 0x00c:
 695                 dev_priv->fsb_freq = 3200;
 696                 break;
 697         case 0x00e:
 698                 dev_priv->fsb_freq = 3733;
 699                 break;
 700         case 0x010:
 701                 dev_priv->fsb_freq = 4266;
 702                 break;
 703         case 0x012:
 704                 dev_priv->fsb_freq = 4800;
 705                 break;
 706         case 0x014:
 707                 dev_priv->fsb_freq = 5333;
 708                 break;
 709         case 0x016:
 710                 dev_priv->fsb_freq = 5866;
 711                 break;
 712         case 0x018:
 713                 dev_priv->fsb_freq = 6400;
 714                 break;
 715         default:
 716                 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
 717                                  csipll & 0x3ff);
 718                 dev_priv->fsb_freq = 0;
 719                 break;
 720         }
 721
 722         if (dev_priv->fsb_freq == 3200) {
 723                 dev_priv->ips.c_m = 0;
 724         } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
 725                 dev_priv->ips.c_m = 1;
 726         } else {
 727                 dev_priv->ips.c_m = 2;
 728         }
 729 }
 730
 731 static const struct cxsr_latency cxsr_latency_table[] = {
 732         {1, 0, 800, 400, 3382, 33382, 3983, 33983},    /* DDR2-400 SC */
 733         {1, 0, 800, 667, 3354, 33354, 3807, 33807},    /* DDR2-667 SC */
 734         {1, 0, 800, 800, 3347, 33347, 3763, 33763},    /* DDR2-800 SC */
 735         {1, 1, 800, 667, 6420, 36420, 6873, 36873},    /* DDR3-667 SC */
 736         {1, 1, 800, 800, 5902, 35902, 6318, 36318},    /* DDR3-800 SC */
 737
 738         {1, 0, 667, 400, 3400, 33400, 4021, 34021},    /* DDR2-400 SC */
 739         {1, 0, 667, 667, 3372, 33372, 3845, 33845},    /* DDR2-667 SC */
 740         {1, 0, 667, 800, 3386, 33386, 3822, 33822},    /* DDR2-800 SC */
 741         {1, 1, 667, 667, 6438, 36438, 6911, 36911},    /* DDR3-667 SC */
 742         {1, 1, 667, 800, 5941, 35941, 6377, 36377},    /* DDR3-800 SC */
 743
 744         {1, 0, 400, 400, 3472, 33472, 4173, 34173},    /* DDR2-400 SC */
 745         {1, 0, 400, 667, 3443, 33443, 3996, 33996},    /* DDR2-667 SC */
 746         {1, 0, 400, 800, 3430, 33430, 3946, 33946},    /* DDR2-800 SC */
 747         {1, 1, 400, 667, 6509, 36509, 7062, 37062},    /* DDR3-667 SC */
 748         {1, 1, 400, 800, 5985, 35985, 6501, 36501},    /* DDR3-800 SC */
 749
 750         {0, 0, 800, 400, 3438, 33438, 4065, 34065},    /* DDR2-400 SC */
 751         {0, 0, 800, 667, 3410, 33410, 3889, 33889},    /* DDR2-667 SC */
 752         {0, 0, 800, 800, 3403, 33403, 3845, 33845},    /* DDR2-800 SC */
 753         {0, 1, 800, 667, 6476, 36476, 6955, 36955},    /* DDR3-667 SC */
 754         {0, 1, 800, 800, 5958, 35958, 6400, 36400},    /* DDR3-800 SC */
 755
 756         {0, 0, 667, 400, 3456, 33456, 4103, 34106},    /* DDR2-400 SC */
 757         {0, 0, 667, 667, 3428, 33428, 3927, 33927},    /* DDR2-667 SC */
 758         {0, 0, 667, 800, 3443, 33443, 3905, 33905},    /* DDR2-800 SC */
 759         {0, 1, 667, 667, 6494, 36494, 6993, 36993},    /* DDR3-667 SC */
 760         {0, 1, 667, 800, 5998, 35998, 6460, 36460},    /* DDR3-800 SC */
 761
 762         {0, 0, 400, 400, 3528, 33528, 4255, 34255},    /* DDR2-400 SC */
 763         {0, 0, 400, 667, 3500, 33500, 4079, 34079},    /* DDR2-667 SC */
 764         {0, 0, 400, 800, 3487, 33487, 4029, 34029},    /* DDR2-800 SC */
 765         {0, 1, 400, 667, 6566, 36566, 7145, 37145},    /* DDR3-667 SC */
 766         {0, 1, 400, 800, 6042, 36042, 6584, 36584},    /* DDR3-800 SC */
 767 };
 768
 769 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
 770                                                          int is_ddr3,
 771                                                          int fsb,
 772                                                          int mem)
 773 {
 774         const struct cxsr_latency *latency;
 775         int i;
 776
 777         if (fsb == 0 || mem == 0)
 778                 return NULL;
 779
 780         for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
 781                 latency = &cxsr_latency_table[i];
 782                 if (is_desktop == latency->is_desktop &&
 783                     is_ddr3 == latency->is_ddr3 &&
 784                     fsb == latency->fsb_freq && mem == latency->mem_freq)
 785                         return latency;
 786         }
 787
 788         DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 789
 790         return NULL;
 791 }
 792
 793 static void pineview_disable_cxsr(struct drm_device *dev)
 794 {
 795         struct drm_i915_private *dev_priv = dev->dev_private;
 796
 797         /* deactivate cxsr */
 798         I915_WRITE(DSPFW3, I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN);
 799 }
 800
 801 /*
 802  * Latency for FIFO fetches is dependent on several factors:
 803  *   - memory configuration (speed, channels)
 804  *   - chipset
 805  *   - current MCH state
 806  * It can be fairly high in some situations, so here we assume a fairly
 807  * pessimal value.  It's a tradeoff between extra memory fetches (if we
 808  * set this value too high, the FIFO will fetch frequently to stay full)
 809  * and power consumption (set it too low to save power and we might see
 810  * FIFO underruns and display "flicker").
 811  *
 812  * A value of 5us seems to be a good balance; safe for very low end
 813  * platforms but not overly aggressive on lower latency configs.
 814  */
 815 static const int latency_ns = 5000;
 816
 817 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
 818 {
 819         struct drm_i915_private *dev_priv = dev->dev_private;
 820         uint32_t dsparb = I915_READ(DSPARB);
 821         int size;
 822
 823         size = dsparb & 0x7f;
 824         if (plane)
 825                 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
 826
 827         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 828                       plane ? "B" : "A", size);
 829
 830         return size;
 831 }
 832
 833 static int i830_get_fifo_size(struct drm_device *dev, int plane)
 834 {
 835         struct drm_i915_private *dev_priv = dev->dev_private;
 836         uint32_t dsparb = I915_READ(DSPARB);
 837         int size;
 838
 839         size = dsparb & 0x1ff;
 840         if (plane)
 841                 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
 842         size >>= 1; /* Convert to cachelines */
 843
 844         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 845                       plane ? "B" : "A", size);
 846
 847         return size;
 848 }
 849
 850 static int i845_get_fifo_size(struct drm_device *dev, int plane)
 851 {
 852         struct drm_i915_private *dev_priv = dev->dev_private;
 853         uint32_t dsparb = I915_READ(DSPARB);
 854         int size;
 855
 856         size = dsparb & 0x7f;
 857         size >>= 2; /* Convert to cachelines */
 858
 859         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 860                       plane ? "B" : "A",
 861                       size);
 862
 863         return size;
 864 }
 865
 866 /* Pineview has different values for various configs */
 867 static const struct intel_watermark_params pineview_display_wm = {
 868         PINEVIEW_DISPLAY_FIFO,
 869         PINEVIEW_MAX_WM,
 870         PINEVIEW_DFT_WM,
 871         PINEVIEW_GUARD_WM,
 872         PINEVIEW_FIFO_LINE_SIZE
 873 };
 874 static const struct intel_watermark_params pineview_display_hplloff_wm = {
 875         PINEVIEW_DISPLAY_FIFO,
 876         PINEVIEW_MAX_WM,
 877         PINEVIEW_DFT_HPLLOFF_WM,
 878         PINEVIEW_GUARD_WM,
 879         PINEVIEW_FIFO_LINE_SIZE
 880 };
 881 static const struct intel_watermark_params pineview_cursor_wm = {
 882         PINEVIEW_CURSOR_FIFO,
 883         PINEVIEW_CURSOR_MAX_WM,
 884         PINEVIEW_CURSOR_DFT_WM,
 885         PINEVIEW_CURSOR_GUARD_WM,
 886         PINEVIEW_FIFO_LINE_SIZE,
 887 };
 888 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
 889         PINEVIEW_CURSOR_FIFO,
 890         PINEVIEW_CURSOR_MAX_WM,
 891         PINEVIEW_CURSOR_DFT_WM,
 892         PINEVIEW_CURSOR_GUARD_WM,
 893         PINEVIEW_FIFO_LINE_SIZE
 894 };
 895 static const struct intel_watermark_params g4x_wm_info = {
 896         G4X_FIFO_SIZE,
 897         G4X_MAX_WM,
 898         G4X_MAX_WM,
 899         2,
 900         G4X_FIFO_LINE_SIZE,
 901 };
 902 static const struct intel_watermark_params g4x_cursor_wm_info = {
 903         I965_CURSOR_FIFO,
 904         I965_CURSOR_MAX_WM,
 905         I965_CURSOR_DFT_WM,
 906         2,
 907         G4X_FIFO_LINE_SIZE,
 908 };
 909 static const struct intel_watermark_params valleyview_wm_info = {
 910         VALLEYVIEW_FIFO_SIZE,
 911         VALLEYVIEW_MAX_WM,
 912         VALLEYVIEW_MAX_WM,
 913         2,
 914         G4X_FIFO_LINE_SIZE,
 915 };
 916 static const struct intel_watermark_params valleyview_cursor_wm_info = {
 917         I965_CURSOR_FIFO,
 918         VALLEYVIEW_CURSOR_MAX_WM,
 919         I965_CURSOR_DFT_WM,
 920         2,
 921         G4X_FIFO_LINE_SIZE,
 922 };
 923 static const struct intel_watermark_params i965_cursor_wm_info = {
 924         I965_CURSOR_FIFO,
 925         I965_CURSOR_MAX_WM,
 926         I965_CURSOR_DFT_WM,
 927         2,
 928         I915_FIFO_LINE_SIZE,
 929 };
 930 static const struct intel_watermark_params i945_wm_info = {
 931         I945_FIFO_SIZE,
 932         I915_MAX_WM,
 933         1,
 934         2,
 935         I915_FIFO_LINE_SIZE
 936 };
 937 static const struct intel_watermark_params i915_wm_info = {
 938         I915_FIFO_SIZE,
 939         I915_MAX_WM,
 940         1,
 941         2,
 942         I915_FIFO_LINE_SIZE
 943 };
 944 static const struct intel_watermark_params i830_wm_info = {
 945         I855GM_FIFO_SIZE,
 946         I915_MAX_WM,
 947         1,
 948         2,
 949         I830_FIFO_LINE_SIZE
 950 };
 951 static const struct intel_watermark_params i845_wm_info = {
 952         I830_FIFO_SIZE,
 953         I915_MAX_WM,
 954         1,
 955         2,
 956         I830_FIFO_LINE_SIZE
 957 };
 958
 959 /**
 960  * intel_calculate_wm - calculate watermark level
 961  * @clock_in_khz: pixel clock
 962  * @wm: chip FIFO params
 963  * @pixel_size: display pixel size
 964  * @latency_ns: memory latency for the platform
 965  *
 966  * Calculate the watermark level (the level at which the display plane will
 967  * start fetching from memory again).  Each chip has a different display
 968  * FIFO size and allocation, so the caller needs to figure that out and pass
 969  * in the correct intel_watermark_params structure.
 970  *
 971  * As the pixel clock runs, the FIFO will be drained at a rate that depends
 972  * on the pixel size.  When it reaches the watermark level, it'll start
 973  * fetching FIFO line sized based chunks from memory until the FIFO fills
 974  * past the watermark point.  If the FIFO drains completely, a FIFO underrun
 975  * will occur, and a display engine hang could result.
 976  */
 977 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
 978                                         const struct intel_watermark_params *wm,
 979                                         int fifo_size,
 980                                         int pixel_size,
 981                                         unsigned long latency_ns)
 982 {
 983         long entries_required, wm_size;
 984
 985         /*
 986          * Note: we need to make sure we don't overflow for various clock &
 987          * latency values.
 988          * clocks go from a few thousand to several hundred thousand.
 989          * latency is usually a few thousand
 990          */
 991         entries_required = ((clock_in_khz / 1000) * pixel_size * latency_ns) /
 992                 1000;
 993         entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
 994
 995         DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
 996
 997         wm_size = fifo_size - (entries_required + wm->guard_size);
 998
 999         DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
1000
1001         /* Don't promote wm_size to unsigned... */
1002         if (wm_size > (long)wm->max_wm)
1003                 wm_size = wm->max_wm;
1004         if (wm_size <= 0)
1005                 wm_size = wm->default_wm;
1006         return wm_size;
1007 }
1008
1009 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
1010 {
1011         struct drm_crtc *crtc, *enabled = NULL;
1012
1013         for_each_crtc(dev, crtc) {
1014                 if (intel_crtc_active(crtc)) {
1015                         if (enabled)
1016                                 return NULL;
1017                         enabled = crtc;
1018                 }
1019         }
1020
1021         return enabled;
1022 }
1023
1024 static void pineview_update_wm(struct drm_crtc *unused_crtc)
1025 {
1026         struct drm_device *dev = unused_crtc->dev;
1027         struct drm_i915_private *dev_priv = dev->dev_private;
1028         struct drm_crtc *crtc;
1029         const struct cxsr_latency *latency;
1030         u32 reg;
1031         unsigned long wm;
1032
1033         latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
1034                                          dev_priv->fsb_freq, dev_priv->mem_freq);
1035         if (!latency) {
1036                 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
1037                 pineview_disable_cxsr(dev);
1038                 return;
1039         }
1040
1041         crtc = single_enabled_crtc(dev);
1042         if (crtc) {
1043                 const struct drm_display_mode *adjusted_mode;
1044                 int pixel_size = crtc->primary->fb->bits_per_pixel / 8;
1045                 int clock;
1046
1047                 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1048                 clock = adjusted_mode->crtc_clock;
1049
1050                 /* Display SR */
1051                 wm = intel_calculate_wm(clock, &pineview_display_wm,
1052                                         pineview_display_wm.fifo_size,
1053                                         pixel_size, latency->display_sr);
1054                 reg = I915_READ(DSPFW1);
1055                 reg &= ~DSPFW_SR_MASK;
1056                 reg |= wm << DSPFW_SR_SHIFT;
1057                 I915_WRITE(DSPFW1, reg);
1058                 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
1059
1060                 /* cursor SR */
1061                 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
1062                                         pineview_display_wm.fifo_size,
1063                                         pixel_size, latency->cursor_sr);
1064                 reg = I915_READ(DSPFW3);
1065                 reg &= ~DSPFW_CURSOR_SR_MASK;
1066                 reg |= (wm & 0x3f) << DSPFW_CURSOR_SR_SHIFT;
1067                 I915_WRITE(DSPFW3, reg);
1068
1069                 /* Display HPLL off SR */
1070                 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
1071                                         pineview_display_hplloff_wm.fifo_size,
1072                                         pixel_size, latency->display_hpll_disable);
1073                 reg = I915_READ(DSPFW3);
1074                 reg &= ~DSPFW_HPLL_SR_MASK;
1075                 reg |= wm & DSPFW_HPLL_SR_MASK;
1076                 I915_WRITE(DSPFW3, reg);
1077
1078                 /* cursor HPLL off SR */
1079                 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
1080                                         pineview_display_hplloff_wm.fifo_size,
1081                                         pixel_size, latency->cursor_hpll_disable);
1082                 reg = I915_READ(DSPFW3);
1083                 reg &= ~DSPFW_HPLL_CURSOR_MASK;
1084                 reg |= (wm & 0x3f) << DSPFW_HPLL_CURSOR_SHIFT;
1085                 I915_WRITE(DSPFW3, reg);
1086                 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
1087
1088                 /* activate cxsr */
1089                 I915_WRITE(DSPFW3,
1090                            I915_READ(DSPFW3) | PINEVIEW_SELF_REFRESH_EN);
1091                 DRM_DEBUG_KMS("Self-refresh is enabled\n");
1092         } else {
1093                 pineview_disable_cxsr(dev);
1094                 DRM_DEBUG_KMS("Self-refresh is disabled\n");
1095         }
1096 }
1097
1098 static bool g4x_compute_wm0(struct drm_device *dev,
1099                             int plane,
1100                             const struct intel_watermark_params *display,
1101                             int display_latency_ns,
1102                             const struct intel_watermark_params *cursor,
1103                             int cursor_latency_ns,
1104                             int *plane_wm,
1105                             int *cursor_wm)
1106 {
1107         struct drm_crtc *crtc;
1108         const struct drm_display_mode *adjusted_mode;
1109         int htotal, hdisplay, clock, pixel_size;
1110         int line_time_us, line_count;
1111         int entries, tlb_miss;
1112
1113         crtc = intel_get_crtc_for_plane(dev, plane);
1114         if (!intel_crtc_active(crtc)) {
1115                 *cursor_wm = cursor->guard_size;
1116                 *plane_wm = display->guard_size;
1117                 return false;
1118         }
1119
1120         adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1121         clock = adjusted_mode->crtc_clock;
1122         htotal = adjusted_mode->crtc_htotal;
1123         hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1124         pixel_size = crtc->primary->fb->bits_per_pixel / 8;
1125
1126         /* Use the small buffer method to calculate plane watermark */
1127         entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
1128         tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
1129         if (tlb_miss > 0)
1130                 entries += tlb_miss;
1131         entries = DIV_ROUND_UP(entries, display->cacheline_size);
1132         *plane_wm = entries + display->guard_size;
1133         if (*plane_wm > (int)display->max_wm)
1134                 *plane_wm = display->max_wm;
1135
1136         /* Use the large buffer method to calculate cursor watermark */
1137         line_time_us = max(htotal * 1000 / clock, 1);
1138         line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
1139         entries = line_count * to_intel_crtc(crtc)->cursor_width * pixel_size;
1140         tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
1141         if (tlb_miss > 0)
1142                 entries += tlb_miss;
1143         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1144         *cursor_wm = entries + cursor->guard_size;
1145         if (*cursor_wm > (int)cursor->max_wm)
1146                 *cursor_wm = (int)cursor->max_wm;
1147
1148         return true;
1149 }
1150
1151 /*
1152  * Check the wm result.
1153  *
1154  * If any calculated watermark values is larger than the maximum value that
1155  * can be programmed into the associated watermark register, that watermark
1156  * must be disabled.
1157  */
1158 static bool g4x_check_srwm(struct drm_device *dev,
1159                            int display_wm, int cursor_wm,
1160                            const struct intel_watermark_params *display,
1161                            const struct intel_watermark_params *cursor)
1162 {
1163         DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
1164                       display_wm, cursor_wm);
1165
1166         if (display_wm > display->max_wm) {
1167                 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
1168                               display_wm, display->max_wm);
1169                 return false;
1170         }
1171
1172         if (cursor_wm > cursor->max_wm) {
1173                 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
1174                               cursor_wm, cursor->max_wm);
1175                 return false;
1176         }
1177
1178         if (!(display_wm || cursor_wm)) {
1179                 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
1180                 return false;
1181         }
1182
1183         return true;
1184 }
1185
1186 static bool g4x_compute_srwm(struct drm_device *dev,
1187                              int plane,
1188                              int latency_ns,
1189                              const struct intel_watermark_params *display,
1190                              const struct intel_watermark_params *cursor,
1191                              int *display_wm, int *cursor_wm)
1192 {
1193         struct drm_crtc *crtc;
1194         const struct drm_display_mode *adjusted_mode;
1195         int hdisplay, htotal, pixel_size, clock;
1196         unsigned long line_time_us;
1197         int line_count, line_size;
1198         int small, large;
1199         int entries;
1200
1201         if (!latency_ns) {
1202                 *display_wm = *cursor_wm = 0;
1203                 return false;
1204         }
1205
1206         crtc = intel_get_crtc_for_plane(dev, plane);
1207         adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1208         clock = adjusted_mode->crtc_clock;
1209         htotal = adjusted_mode->crtc_htotal;
1210         hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1211         pixel_size = crtc->primary->fb->bits_per_pixel / 8;
1212
1213         line_time_us = max(htotal * 1000 / clock, 1);
1214         line_count = (latency_ns / line_time_us + 1000) / 1000;
1215         line_size = hdisplay * pixel_size;
1216
1217         /* Use the minimum of the small and large buffer method for primary */
1218         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1219         large = line_count * line_size;
1220
1221         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1222         *display_wm = entries + display->guard_size;
1223
1224         /* calculate the self-refresh watermark for display cursor */
1225         entries = line_count * pixel_size * to_intel_crtc(crtc)->cursor_width;
1226         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1227         *cursor_wm = entries + cursor->guard_size;
1228
1229         return g4x_check_srwm(dev,
1230                               *display_wm, *cursor_wm,
1231                               display, cursor);
1232 }
1233
1234 static bool vlv_compute_drain_latency(struct drm_device *dev,
1235                                      int plane,
1236                                      int *plane_prec_mult,
1237                                      int *plane_dl,
1238                                      int *cursor_prec_mult,
1239                                      int *cursor_dl)
1240 {
1241         struct drm_crtc *crtc;
1242         int clock, pixel_size;
1243         int entries;
1244
1245         crtc = intel_get_crtc_for_plane(dev, plane);
1246         if (!intel_crtc_active(crtc))
1247                 return false;
1248
1249         clock = to_intel_crtc(crtc)->config.adjusted_mode.crtc_clock;
1250         pixel_size = crtc->primary->fb->bits_per_pixel / 8;     /* BPP */
1251
1252         entries = (clock / 1000) * pixel_size;
1253         *plane_prec_mult = (entries > 256) ?
1254                 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1255         *plane_dl = (64 * (*plane_prec_mult) * 4) / ((clock / 1000) *
1256                                                      pixel_size);
1257
1258         entries = (clock / 1000) * 4;   /* BPP is always 4 for cursor */
1259         *cursor_prec_mult = (entries > 256) ?
1260                 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1261         *cursor_dl = (64 * (*cursor_prec_mult) * 4) / ((clock / 1000) * 4);
1262
1263         return true;
1264 }
1265
1266 /*
1267  * Update drain latency registers of memory arbiter
1268  *
1269  * Valleyview SoC has a new memory arbiter and needs drain latency registers
1270  * to be programmed. Each plane has a drain latency multiplier and a drain
1271  * latency value.
1272  */
1273
1274 static void vlv_update_drain_latency(struct drm_device *dev)
1275 {
1276         struct drm_i915_private *dev_priv = dev->dev_private;
1277         int planea_prec, planea_dl, planeb_prec, planeb_dl;
1278         int cursora_prec, cursora_dl, cursorb_prec, cursorb_dl;
1279         int plane_prec_mult, cursor_prec_mult; /* Precision multiplier is
1280                                                         either 16 or 32 */
1281
1282         /* For plane A, Cursor A */
1283         if (vlv_compute_drain_latency(dev, 0, &plane_prec_mult, &planea_dl,
1284                                       &cursor_prec_mult, &cursora_dl)) {
1285                 cursora_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1286                         DDL_CURSORA_PRECISION_32 : DDL_CURSORA_PRECISION_16;
1287                 planea_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1288                         DDL_PLANEA_PRECISION_32 : DDL_PLANEA_PRECISION_16;
1289
1290                 I915_WRITE(VLV_DDL1, cursora_prec |
1291                                 (cursora_dl << DDL_CURSORA_SHIFT) |
1292                                 planea_prec | planea_dl);
1293         }
1294
1295         /* For plane B, Cursor B */
1296         if (vlv_compute_drain_latency(dev, 1, &plane_prec_mult, &planeb_dl,
1297                                       &cursor_prec_mult, &cursorb_dl)) {
1298                 cursorb_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1299                         DDL_CURSORB_PRECISION_32 : DDL_CURSORB_PRECISION_16;
1300                 planeb_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1301                         DDL_PLANEB_PRECISION_32 : DDL_PLANEB_PRECISION_16;
1302
1303                 I915_WRITE(VLV_DDL2, cursorb_prec |
1304                                 (cursorb_dl << DDL_CURSORB_SHIFT) |
1305                                 planeb_prec | planeb_dl);
1306         }
1307 }
1308
1309 #define single_plane_enabled(mask) is_power_of_2(mask)
1310
1311 static void valleyview_update_wm(struct drm_crtc *crtc)
1312 {
1313         struct drm_device *dev = crtc->dev;
1314         static const int sr_latency_ns = 12000;
1315         struct drm_i915_private *dev_priv = dev->dev_private;
1316         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1317         int plane_sr, cursor_sr;
1318         int ignore_plane_sr, ignore_cursor_sr;
1319         unsigned int enabled = 0;
1320
1321         vlv_update_drain_latency(dev);
1322
1323         if (g4x_compute_wm0(dev, PIPE_A,
1324                             &valleyview_wm_info, latency_ns,
1325                             &valleyview_cursor_wm_info, latency_ns,
1326                             &planea_wm, &cursora_wm))
1327                 enabled |= 1 << PIPE_A;
1328
1329         if (g4x_compute_wm0(dev, PIPE_B,
1330                             &valleyview_wm_info, latency_ns,
1331                             &valleyview_cursor_wm_info, latency_ns,
1332                             &planeb_wm, &cursorb_wm))
1333                 enabled |= 1 << PIPE_B;
1334
1335         if (single_plane_enabled(enabled) &&
1336             g4x_compute_srwm(dev, ffs(enabled) - 1,
1337                              sr_latency_ns,
1338                              &valleyview_wm_info,
1339                              &valleyview_cursor_wm_info,
1340                              &plane_sr, &ignore_cursor_sr) &&
1341             g4x_compute_srwm(dev, ffs(enabled) - 1,
1342                              2*sr_latency_ns,
1343                              &valleyview_wm_info,
1344                              &valleyview_cursor_wm_info,
1345                              &ignore_plane_sr, &cursor_sr)) {
1346                 I915_WRITE(FW_BLC_SELF_VLV, FW_CSPWRDWNEN);
1347         } else {
1348                 I915_WRITE(FW_BLC_SELF_VLV,
1349                            I915_READ(FW_BLC_SELF_VLV) & ~FW_CSPWRDWNEN);
1350                 plane_sr = cursor_sr = 0;
1351         }
1352
1353         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1354                       planea_wm, cursora_wm,
1355                       planeb_wm, cursorb_wm,
1356                       plane_sr, cursor_sr);
1357
1358         I915_WRITE(DSPFW1,
1359                    (plane_sr << DSPFW_SR_SHIFT) |
1360                    (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1361                    (planeb_wm << DSPFW_PLANEB_SHIFT) |
1362                    planea_wm);
1363         I915_WRITE(DSPFW2,
1364                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1365                    (cursora_wm << DSPFW_CURSORA_SHIFT));
1366         I915_WRITE(DSPFW3,
1367                    (I915_READ(DSPFW3) & ~DSPFW_CURSOR_SR_MASK) |
1368                    (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1369 }
1370
1371 static void g4x_update_wm(struct drm_crtc *crtc)
1372 {
1373         struct drm_device *dev = crtc->dev;
1374         static const int sr_latency_ns = 12000;
1375         struct drm_i915_private *dev_priv = dev->dev_private;
1376         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1377         int plane_sr, cursor_sr;
1378         unsigned int enabled = 0;
1379
1380         if (g4x_compute_wm0(dev, PIPE_A,
1381                             &g4x_wm_info, latency_ns,
1382                             &g4x_cursor_wm_info, latency_ns,
1383                             &planea_wm, &cursora_wm))
1384                 enabled |= 1 << PIPE_A;
1385
1386         if (g4x_compute_wm0(dev, PIPE_B,
1387                             &g4x_wm_info, latency_ns,
1388                             &g4x_cursor_wm_info, latency_ns,
1389                             &planeb_wm, &cursorb_wm))
1390                 enabled |= 1 << PIPE_B;
1391
1392         if (single_plane_enabled(enabled) &&
1393             g4x_compute_srwm(dev, ffs(enabled) - 1,
1394                              sr_latency_ns,
1395                              &g4x_wm_info,
1396                              &g4x_cursor_wm_info,
1397                              &plane_sr, &cursor_sr)) {
1398                 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1399         } else {
1400                 I915_WRITE(FW_BLC_SELF,
1401                            I915_READ(FW_BLC_SELF) & ~FW_BLC_SELF_EN);
1402                 plane_sr = cursor_sr = 0;
1403         }
1404
1405         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1406                       planea_wm, cursora_wm,
1407                       planeb_wm, cursorb_wm,
1408                       plane_sr, cursor_sr);
1409
1410         I915_WRITE(DSPFW1,
1411                    (plane_sr << DSPFW_SR_SHIFT) |
1412                    (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1413                    (planeb_wm << DSPFW_PLANEB_SHIFT) |
1414                    planea_wm);
1415         I915_WRITE(DSPFW2,
1416                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1417                    (cursora_wm << DSPFW_CURSORA_SHIFT));
1418         /* HPLL off in SR has some issues on G4x... disable it */
1419         I915_WRITE(DSPFW3,
1420                    (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1421                    (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1422 }
1423
1424 static void i965_update_wm(struct drm_crtc *unused_crtc)
1425 {
1426         struct drm_device *dev = unused_crtc->dev;
1427         struct drm_i915_private *dev_priv = dev->dev_private;
1428         struct drm_crtc *crtc;
1429         int srwm = 1;
1430         int cursor_sr = 16;
1431
1432         /* Calc sr entries for one plane configs */
1433         crtc = single_enabled_crtc(dev);
1434         if (crtc) {
1435                 /* self-refresh has much higher latency */
1436                 static const int sr_latency_ns = 12000;
1437                 const struct drm_display_mode *adjusted_mode =
1438                         &to_intel_crtc(crtc)->config.adjusted_mode;
1439                 int clock = adjusted_mode->crtc_clock;
1440                 int htotal = adjusted_mode->crtc_htotal;
1441                 int hdisplay = to_intel_crtc(crtc)->config.pipe_src_w;
1442                 int pixel_size = crtc->primary->fb->bits_per_pixel / 8;
1443                 unsigned long line_time_us;
1444                 int entries;
1445
1446                 line_time_us = max(htotal * 1000 / clock, 1);
1447
1448                 /* Use ns/us then divide to preserve precision */
1449                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1450                         pixel_size * hdisplay;
1451                 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1452                 srwm = I965_FIFO_SIZE - entries;
1453                 if (srwm < 0)
1454                         srwm = 1;
1455                 srwm &= 0x1ff;
1456                 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1457                               entries, srwm);
1458
1459                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1460                         pixel_size * to_intel_crtc(crtc)->cursor_width;
1461                 entries = DIV_ROUND_UP(entries,
1462                                           i965_cursor_wm_info.cacheline_size);
1463                 cursor_sr = i965_cursor_wm_info.fifo_size -
1464                         (entries + i965_cursor_wm_info.guard_size);
1465
1466                 if (cursor_sr > i965_cursor_wm_info.max_wm)
1467                         cursor_sr = i965_cursor_wm_info.max_wm;
1468
1469                 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1470                               "cursor %d\n", srwm, cursor_sr);
1471
1472                 if (IS_CRESTLINE(dev))
1473                         I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1474         } else {
1475                 /* Turn off self refresh if both pipes are enabled */
1476                 if (IS_CRESTLINE(dev))
1477                         I915_WRITE(FW_BLC_SELF, I915_READ(FW_BLC_SELF)
1478                                    & ~FW_BLC_SELF_EN);
1479         }
1480
1481         DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1482                       srwm);
1483
1484         /* 965 has limitations... */
1485         I915_WRITE(DSPFW1, (srwm << DSPFW_SR_SHIFT) |
1486                    (8 << 16) | (8 << 8) | (8 << 0));
1487         I915_WRITE(DSPFW2, (8 << 8) | (8 << 0));
1488         /* update cursor SR watermark */
1489         I915_WRITE(DSPFW3, (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1490 }
1491
1492 static void i9xx_update_wm(struct drm_crtc *unused_crtc)
1493 {
1494         struct drm_device *dev = unused_crtc->dev;
1495         struct drm_i915_private *dev_priv = dev->dev_private;
1496         const struct intel_watermark_params *wm_info;
1497         uint32_t fwater_lo;
1498         uint32_t fwater_hi;
1499         int cwm, srwm = 1;
1500         int fifo_size;
1501         int planea_wm, planeb_wm;
1502         struct drm_crtc *crtc, *enabled = NULL;
1503
1504         if (IS_I945GM(dev))
1505                 wm_info = &i945_wm_info;
1506         else if (!IS_GEN2(dev))
1507                 wm_info = &i915_wm_info;
1508         else
1509                 wm_info = &i830_wm_info;
1510
1511         fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1512         crtc = intel_get_crtc_for_plane(dev, 0);
1513         if (intel_crtc_active(crtc)) {
1514                 const struct drm_display_mode *adjusted_mode;
1515                 int cpp = crtc->primary->fb->bits_per_pixel / 8;
1516                 if (IS_GEN2(dev))
1517                         cpp = 4;
1518
1519                 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1520                 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1521                                                wm_info, fifo_size, cpp,
1522                                                latency_ns);
1523                 enabled = crtc;
1524         } else
1525                 planea_wm = fifo_size - wm_info->guard_size;
1526
1527         fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1528         crtc = intel_get_crtc_for_plane(dev, 1);
1529         if (intel_crtc_active(crtc)) {
1530                 const struct drm_display_mode *adjusted_mode;
1531                 int cpp = crtc->primary->fb->bits_per_pixel / 8;
1532                 if (IS_GEN2(dev))
1533                         cpp = 4;
1534
1535                 adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1536                 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1537                                                wm_info, fifo_size, cpp,
1538                                                latency_ns);
1539                 if (enabled == NULL)
1540                         enabled = crtc;
1541                 else
1542                         enabled = NULL;
1543         } else
1544                 planeb_wm = fifo_size - wm_info->guard_size;
1545
1546         DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1547
1548         if (IS_I915GM(dev) && enabled) {
1549                 struct intel_framebuffer *fb;
1550
1551                 fb = to_intel_framebuffer(enabled->primary->fb);
1552
1553                 /* self-refresh seems busted with untiled */
1554                 if (fb->obj->tiling_mode == I915_TILING_NONE)
1555                         enabled = NULL;
1556         }
1557
1558         /*
1559          * Overlay gets an aggressive default since video jitter is bad.
1560          */
1561         cwm = 2;
1562
1563         /* Play safe and disable self-refresh before adjusting watermarks. */
1564         if (IS_I945G(dev) || IS_I945GM(dev))
1565                 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN_MASK | 0);
1566         else if (IS_I915GM(dev))
1567                 I915_WRITE(INSTPM, _MASKED_BIT_DISABLE(INSTPM_SELF_EN));
1568
1569         /* Calc sr entries for one plane configs */
1570         if (HAS_FW_BLC(dev) && enabled) {
1571                 /* self-refresh has much higher latency */
1572                 static const int sr_latency_ns = 6000;
1573                 const struct drm_display_mode *adjusted_mode =
1574                         &to_intel_crtc(enabled)->config.adjusted_mode;
1575                 int clock = adjusted_mode->crtc_clock;
1576                 int htotal = adjusted_mode->crtc_htotal;
1577                 int hdisplay = to_intel_crtc(enabled)->config.pipe_src_w;
1578                 int pixel_size = enabled->primary->fb->bits_per_pixel / 8;
1579                 unsigned long line_time_us;
1580                 int entries;
1581
1582                 line_time_us = max(htotal * 1000 / clock, 1);
1583
1584                 /* Use ns/us then divide to preserve precision */
1585                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1586                         pixel_size * hdisplay;
1587                 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1588                 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1589                 srwm = wm_info->fifo_size - entries;
1590                 if (srwm < 0)
1591                         srwm = 1;
1592
1593                 if (IS_I945G(dev) || IS_I945GM(dev))
1594                         I915_WRITE(FW_BLC_SELF,
1595                                    FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1596                 else if (IS_I915GM(dev))
1597                         I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1598         }
1599
1600         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1601                       planea_wm, planeb_wm, cwm, srwm);
1602
1603         fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1604         fwater_hi = (cwm & 0x1f);
1605
1606         /* Set request length to 8 cachelines per fetch */
1607         fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1608         fwater_hi = fwater_hi | (1 << 8);
1609
1610         I915_WRITE(FW_BLC, fwater_lo);
1611         I915_WRITE(FW_BLC2, fwater_hi);
1612
1613         if (HAS_FW_BLC(dev)) {
1614                 if (enabled) {
1615                         if (IS_I945G(dev) || IS_I945GM(dev))
1616                                 I915_WRITE(FW_BLC_SELF,
1617                                            FW_BLC_SELF_EN_MASK | FW_BLC_SELF_EN);
1618                         else if (IS_I915GM(dev))
1619                                 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_SELF_EN));
1620                         DRM_DEBUG_KMS("memory self refresh enabled\n");
1621                 } else
1622                         DRM_DEBUG_KMS("memory self refresh disabled\n");
1623         }
1624 }
1625
1626 static void i845_update_wm(struct drm_crtc *unused_crtc)
1627 {
1628         struct drm_device *dev = unused_crtc->dev;
1629         struct drm_i915_private *dev_priv = dev->dev_private;
1630         struct drm_crtc *crtc;
1631         const struct drm_display_mode *adjusted_mode;
1632         uint32_t fwater_lo;
1633         int planea_wm;
1634
1635         crtc = single_enabled_crtc(dev);
1636         if (crtc == NULL)
1637                 return;
1638
1639         adjusted_mode = &to_intel_crtc(crtc)->config.adjusted_mode;
1640         planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1641                                        &i845_wm_info,
1642                                        dev_priv->display.get_fifo_size(dev, 0),
1643                                        4, latency_ns);
1644         fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1645         fwater_lo |= (3<<8) | planea_wm;
1646
1647         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1648
1649         I915_WRITE(FW_BLC, fwater_lo);
1650 }
1651
1652 static uint32_t ilk_pipe_pixel_rate(struct drm_device *dev,
1653                                     struct drm_crtc *crtc)
1654 {
1655         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1656         uint32_t pixel_rate;
1657
1658         pixel_rate = intel_crtc->config.adjusted_mode.crtc_clock;
1659
1660         /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
1661          * adjust the pixel_rate here. */
1662
1663         if (intel_crtc->config.pch_pfit.enabled) {
1664                 uint64_t pipe_w, pipe_h, pfit_w, pfit_h;
1665                 uint32_t pfit_size = intel_crtc->config.pch_pfit.size;
1666
1667                 pipe_w = intel_crtc->config.pipe_src_w;
1668                 pipe_h = intel_crtc->config.pipe_src_h;
1669                 pfit_w = (pfit_size >> 16) & 0xFFFF;
1670                 pfit_h = pfit_size & 0xFFFF;
1671                 if (pipe_w < pfit_w)
1672                         pipe_w = pfit_w;
1673                 if (pipe_h < pfit_h)
1674                         pipe_h = pfit_h;
1675
1676                 pixel_rate = div_u64((uint64_t) pixel_rate * pipe_w * pipe_h,
1677                                      pfit_w * pfit_h);
1678         }
1679
1680         return pixel_rate;
1681 }
1682
1683 /* latency must be in 0.1us units. */
1684 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
1685                                uint32_t latency)
1686 {
1687         uint64_t ret;
1688
1689         if (WARN(latency == 0, "Latency value missing\n"))
1690                 return UINT_MAX;
1691
1692         ret = (uint64_t) pixel_rate * bytes_per_pixel * latency;
1693         ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
1694
1695         return ret;
1696 }
1697
1698 /* latency must be in 0.1us units. */
1699 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
1700                                uint32_t horiz_pixels, uint8_t bytes_per_pixel,
1701                                uint32_t latency)
1702 {
1703         uint32_t ret;
1704
1705         if (WARN(latency == 0, "Latency value missing\n"))
1706                 return UINT_MAX;
1707
1708         ret = (latency * pixel_rate) / (pipe_htotal * 10000);
1709         ret = (ret + 1) * horiz_pixels * bytes_per_pixel;
1710         ret = DIV_ROUND_UP(ret, 64) + 2;
1711         return ret;
1712 }
1713
1714 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
1715                            uint8_t bytes_per_pixel)
1716 {
1717         return DIV_ROUND_UP(pri_val * 64, horiz_pixels * bytes_per_pixel) + 2;
1718 }
1719
1720 struct ilk_pipe_wm_parameters {
1721         bool active;
1722         uint32_t pipe_htotal;
1723         uint32_t pixel_rate;
1724         struct intel_plane_wm_parameters pri;
1725         struct intel_plane_wm_parameters spr;
1726         struct intel_plane_wm_parameters cur;
1727 };
1728
1729 struct ilk_wm_maximums {
1730         uint16_t pri;
1731         uint16_t spr;
1732         uint16_t cur;
1733         uint16_t fbc;
1734 };
1735
1736 /* used in computing the new watermarks state */
1737 struct intel_wm_config {
1738         unsigned int num_pipes_active;
1739         bool sprites_enabled;
1740         bool sprites_scaled;
1741 };
1742
1743 /*
1744  * For both WM_PIPE and WM_LP.
1745  * mem_value must be in 0.1us units.
1746  */
1747 static uint32_t ilk_compute_pri_wm(const struct ilk_pipe_wm_parameters *params,
1748                                    uint32_t mem_value,
1749                                    bool is_lp)
1750 {
1751         uint32_t method1, method2;
1752
1753         if (!params->active || !params->pri.enabled)
1754                 return 0;
1755
1756         method1 = ilk_wm_method1(params->pixel_rate,
1757                                  params->pri.bytes_per_pixel,
1758                                  mem_value);
1759
1760         if (!is_lp)
1761                 return method1;
1762
1763         method2 = ilk_wm_method2(params->pixel_rate,
1764                                  params->pipe_htotal,
1765                                  params->pri.horiz_pixels,
1766                                  params->pri.bytes_per_pixel,
1767                                  mem_value);
1768
1769         return min(method1, method2);
1770 }
1771
1772 /*
1773  * For both WM_PIPE and WM_LP.
1774  * mem_value must be in 0.1us units.
1775  */
1776 static uint32_t ilk_compute_spr_wm(const struct ilk_pipe_wm_parameters *params,
1777                                    uint32_t mem_value)
1778 {
1779         uint32_t method1, method2;
1780
1781         if (!params->active || !params->spr.enabled)
1782                 return 0;
1783
1784         method1 = ilk_wm_method1(params->pixel_rate,
1785                                  params->spr.bytes_per_pixel,
1786                                  mem_value);
1787         method2 = ilk_wm_method2(params->pixel_rate,
1788                                  params->pipe_htotal,
1789                                  params->spr.horiz_pixels,
1790                                  params->spr.bytes_per_pixel,
1791                                  mem_value);
1792         return min(method1, method2);
1793 }
1794
1795 /*
1796  * For both WM_PIPE and WM_LP.
1797  * mem_value must be in 0.1us units.
1798  */
1799 static uint32_t ilk_compute_cur_wm(const struct ilk_pipe_wm_parameters *params,
1800                                    uint32_t mem_value)
1801 {
1802         if (!params->active || !params->cur.enabled)
1803                 return 0;
1804
1805         return ilk_wm_method2(params->pixel_rate,
1806                               params->pipe_htotal,
1807                               params->cur.horiz_pixels,
1808                               params->cur.bytes_per_pixel,
1809                               mem_value);
1810 }
1811
1812 /* Only for WM_LP. */
1813 static uint32_t ilk_compute_fbc_wm(const struct ilk_pipe_wm_parameters *params,
1814                                    uint32_t pri_val)
1815 {
1816         if (!params->active || !params->pri.enabled)
1817                 return 0;
1818
1819         return ilk_wm_fbc(pri_val,
1820                           params->pri.horiz_pixels,
1821                           params->pri.bytes_per_pixel);
1822 }
1823
1824 static unsigned int ilk_display_fifo_size(const struct drm_device *dev)
1825 {
1826         if (INTEL_INFO(dev)->gen >= 8)
1827                 return 3072;
1828         else if (INTEL_INFO(dev)->gen >= 7)
1829                 return 768;
1830         else
1831                 return 512;
1832 }
1833
1834 static unsigned int ilk_plane_wm_reg_max(const struct drm_device *dev,
1835                                          int level, bool is_sprite)
1836 {
1837         if (INTEL_INFO(dev)->gen >= 8)
1838                 /* BDW primary/sprite plane watermarks */
1839                 return level == 0 ? 255 : 2047;
1840         else if (INTEL_INFO(dev)->gen >= 7)
1841                 /* IVB/HSW primary/sprite plane watermarks */
1842                 return level == 0 ? 127 : 1023;
1843         else if (!is_sprite)
1844                 /* ILK/SNB primary plane watermarks */
1845                 return level == 0 ? 127 : 511;
1846         else
1847                 /* ILK/SNB sprite plane watermarks */
1848                 return level == 0 ? 63 : 255;
1849 }
1850
1851 static unsigned int ilk_cursor_wm_reg_max(const struct drm_device *dev,
1852                                           int level)
1853 {
1854         if (INTEL_INFO(dev)->gen >= 7)
1855                 return level == 0 ? 63 : 255;
1856         else
1857                 return level == 0 ? 31 : 63;
1858 }
1859
1860 static unsigned int ilk_fbc_wm_reg_max(const struct drm_device *dev)
1861 {
1862         if (INTEL_INFO(dev)->gen >= 8)
1863                 return 31;
1864         else
1865                 return 15;
1866 }
1867
1868 /* Calculate the maximum primary/sprite plane watermark */
1869 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
1870                                      int level,
1871                                      const struct intel_wm_config *config,
1872                                      enum intel_ddb_partitioning ddb_partitioning,
1873                                      bool is_sprite)
1874 {
1875         unsigned int fifo_size = ilk_display_fifo_size(dev);
1876
1877         /* if sprites aren't enabled, sprites get nothing */
1878         if (is_sprite && !config->sprites_enabled)
1879                 return 0;
1880
1881         /* HSW allows LP1+ watermarks even with multiple pipes */
1882         if (level == 0 || config->num_pipes_active > 1) {
1883                 fifo_size /= INTEL_INFO(dev)->num_pipes;
1884
1885                 /*
1886                  * For some reason the non self refresh
1887                  * FIFO size is only half of the self
1888                  * refresh FIFO size on ILK/SNB.
1889                  */
1890                 if (INTEL_INFO(dev)->gen <= 6)
1891                         fifo_size /= 2;
1892         }
1893
1894         if (config->sprites_enabled) {
1895                 /* level 0 is always calculated with 1:1 split */
1896                 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
1897                         if (is_sprite)
1898                                 fifo_size *= 5;
1899                         fifo_size /= 6;
1900                 } else {
1901                         fifo_size /= 2;
1902                 }
1903         }
1904
1905         /* clamp to max that the registers can hold */
1906         return min(fifo_size, ilk_plane_wm_reg_max(dev, level, is_sprite));
1907 }
1908
1909 /* Calculate the maximum cursor plane watermark */
1910 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
1911                                       int level,
1912                                       const struct intel_wm_config *config)
1913 {
1914         /* HSW LP1+ watermarks w/ multiple pipes */
1915         if (level > 0 && config->num_pipes_active > 1)
1916                 return 64;
1917
1918         /* otherwise just report max that registers can hold */
1919         return ilk_cursor_wm_reg_max(dev, level);
1920 }
1921
1922 static void ilk_compute_wm_maximums(const struct drm_device *dev,
1923                                     int level,
1924                                     const struct intel_wm_config *config,
1925                                     enum intel_ddb_partitioning ddb_partitioning,
1926                                     struct ilk_wm_maximums *max)
1927 {
1928         max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
1929         max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
1930         max->cur = ilk_cursor_wm_max(dev, level, config);
1931         max->fbc = ilk_fbc_wm_reg_max(dev);
1932 }
1933
1934 static void ilk_compute_wm_reg_maximums(struct drm_device *dev,
1935                                         int level,
1936                                         struct ilk_wm_maximums *max)
1937 {
1938         max->pri = ilk_plane_wm_reg_max(dev, level, false);
1939         max->spr = ilk_plane_wm_reg_max(dev, level, true);
1940         max->cur = ilk_cursor_wm_reg_max(dev, level);
1941         max->fbc = ilk_fbc_wm_reg_max(dev);
1942 }
1943
1944 static bool ilk_validate_wm_level(int level,
1945                                   const struct ilk_wm_maximums *max,
1946                                   struct intel_wm_level *result)
1947 {
1948         bool ret;
1949
1950         /* already determined to be invalid? */
1951         if (!result->enable)
1952                 return false;
1953
1954         result->enable = result->pri_val <= max->pri &&
1955                          result->spr_val <= max->spr &&
1956                          result->cur_val <= max->cur;
1957
1958         ret = result->enable;
1959
1960         /*
1961          * HACK until we can pre-compute everything,
1962          * and thus fail gracefully if LP0 watermarks
1963          * are exceeded...
1964          */
1965         if (level == 0 && !result->enable) {
1966                 if (result->pri_val > max->pri)
1967                         DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
1968                                       level, result->pri_val, max->pri);
1969                 if (result->spr_val > max->spr)
1970                         DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
1971                                       level, result->spr_val, max->spr);
1972                 if (result->cur_val > max->cur)
1973                         DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
1974                                       level, result->cur_val, max->cur);
1975
1976                 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
1977                 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
1978                 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
1979                 result->enable = true;
1980         }
1981
1982         return ret;
1983 }
1984
1985 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
1986                                  int level,
1987                                  const struct ilk_pipe_wm_parameters *p,
1988                                  struct intel_wm_level *result)
1989 {
1990         uint16_t pri_latency = dev_priv->wm.pri_latency[level];
1991         uint16_t spr_latency = dev_priv->wm.spr_latency[level];
1992         uint16_t cur_latency = dev_priv->wm.cur_latency[level];
1993
1994         /* WM1+ latency values stored in 0.5us units */
1995         if (level > 0) {
1996                 pri_latency *= 5;
1997                 spr_latency *= 5;
1998                 cur_latency *= 5;
1999         }
2000
2001         result->pri_val = ilk_compute_pri_wm(p, pri_latency, level);
2002         result->spr_val = ilk_compute_spr_wm(p, spr_latency);
2003         result->cur_val = ilk_compute_cur_wm(p, cur_latency);
2004         result->fbc_val = ilk_compute_fbc_wm(p, result->pri_val);
2005         result->enable = true;
2006 }
2007
2008 static uint32_t
2009 hsw_compute_linetime_wm(struct drm_device *dev, struct drm_crtc *crtc)
2010 {
2011         struct drm_i915_private *dev_priv = dev->dev_private;
2012         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2013         struct drm_display_mode *mode = &intel_crtc->config.adjusted_mode;
2014         u32 linetime, ips_linetime;
2015
2016         if (!intel_crtc_active(crtc))
2017                 return 0;
2018
2019         /* The WM are computed with base on how long it takes to fill a single
2020          * row at the given clock rate, multiplied by 8.
2021          * */
2022         linetime = DIV_ROUND_CLOSEST(mode->crtc_htotal * 1000 * 8,
2023                                      mode->crtc_clock);
2024         ips_linetime = DIV_ROUND_CLOSEST(mode->crtc_htotal * 1000 * 8,
2025                                          intel_ddi_get_cdclk_freq(dev_priv));
2026
2027         return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2028                PIPE_WM_LINETIME_TIME(linetime);
2029 }
2030
2031 static void intel_read_wm_latency(struct drm_device *dev, uint16_t wm[5])
2032 {
2033         struct drm_i915_private *dev_priv = dev->dev_private;
2034
2035         if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2036                 uint64_t sskpd = I915_READ64(MCH_SSKPD);
2037
2038                 wm[0] = (sskpd >> 56) & 0xFF;
2039                 if (wm[0] == 0)
2040                         wm[0] = sskpd & 0xF;
2041                 wm[1] = (sskpd >> 4) & 0xFF;
2042                 wm[2] = (sskpd >> 12) & 0xFF;
2043                 wm[3] = (sskpd >> 20) & 0x1FF;
2044                 wm[4] = (sskpd >> 32) & 0x1FF;
2045         } else if (INTEL_INFO(dev)->gen >= 6) {
2046                 uint32_t sskpd = I915_READ(MCH_SSKPD);
2047
2048                 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2049                 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2050                 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2051                 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2052         } else if (INTEL_INFO(dev)->gen >= 5) {
2053                 uint32_t mltr = I915_READ(MLTR_ILK);
2054
2055                 /* ILK primary LP0 latency is 700 ns */
2056                 wm[0] = 7;
2057                 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2058                 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2059         }
2060 }
2061
2062 static void intel_fixup_spr_wm_latency(struct drm_device *dev, uint16_t wm[5])
2063 {
2064         /* ILK sprite LP0 latency is 1300 ns */
2065         if (INTEL_INFO(dev)->gen == 5)
2066                 wm[0] = 13;
2067 }
2068
2069 static void intel_fixup_cur_wm_latency(struct drm_device *dev, uint16_t wm[5])
2070 {
2071         /* ILK cursor LP0 latency is 1300 ns */
2072         if (INTEL_INFO(dev)->gen == 5)
2073                 wm[0] = 13;
2074
2075         /* WaDoubleCursorLP3Latency:ivb */
2076         if (IS_IVYBRIDGE(dev))
2077                 wm[3] *= 2;
2078 }
2079
2080 int ilk_wm_max_level(const struct drm_device *dev)
2081 {
2082         /* how many WM levels are we expecting */
2083         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2084                 return 4;
2085         else if (INTEL_INFO(dev)->gen >= 6)
2086                 return 3;
2087         else
2088                 return 2;
2089 }
2090
2091 static void intel_print_wm_latency(struct drm_device *dev,
2092                                    const char *name,
2093                                    const uint16_t wm[5])
2094 {
2095         int level, max_level = ilk_wm_max_level(dev);
2096
2097         for (level = 0; level <= max_level; level++) {
2098                 unsigned int latency = wm[level];
2099
2100                 if (latency == 0) {
2101                         DRM_ERROR("%s WM%d latency not provided\n",
2102                                   name, level);
2103                         continue;
2104                 }
2105
2106                 /* WM1+ latency values in 0.5us units */
2107                 if (level > 0)
2108                         latency *= 5;
2109
2110                 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2111                               name, level, wm[level],
2112                               latency / 10, latency % 10);
2113         }
2114 }
2115
2116 static void ilk_setup_wm_latency(struct drm_device *dev)
2117 {
2118         struct drm_i915_private *dev_priv = dev->dev_private;
2119
2120         intel_read_wm_latency(dev, dev_priv->wm.pri_latency);
2121
2122         memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2123                sizeof(dev_priv->wm.pri_latency));
2124         memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2125                sizeof(dev_priv->wm.pri_latency));
2126
2127         intel_fixup_spr_wm_latency(dev, dev_priv->wm.spr_latency);
2128         intel_fixup_cur_wm_latency(dev, dev_priv->wm.cur_latency);
2129
2130         intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2131         intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2132         intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2133 }
2134
2135 static void ilk_compute_wm_parameters(struct drm_crtc *crtc,
2136                                       struct ilk_pipe_wm_parameters *p)
2137 {
2138         struct drm_device *dev = crtc->dev;
2139         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2140         enum pipe pipe = intel_crtc->pipe;
2141         struct drm_plane *plane;
2142
2143         if (!intel_crtc_active(crtc))
2144                 return;
2145
2146         p->active = true;
2147         p->pipe_htotal = intel_crtc->config.adjusted_mode.crtc_htotal;
2148         p->pixel_rate = ilk_pipe_pixel_rate(dev, crtc);
2149         p->pri.bytes_per_pixel = crtc->primary->fb->bits_per_pixel / 8;
2150         p->cur.bytes_per_pixel = 4;
2151         p->pri.horiz_pixels = intel_crtc->config.pipe_src_w;
2152         p->cur.horiz_pixels = intel_crtc->cursor_width;
2153         /* TODO: for now, assume primary and cursor planes are always enabled. */
2154         p->pri.enabled = true;
2155         p->cur.enabled = true;
2156
2157         drm_for_each_legacy_plane(plane, &dev->mode_config.plane_list) {
2158                 struct intel_plane *intel_plane = to_intel_plane(plane);
2159
2160                 if (intel_plane->pipe == pipe) {
2161                         p->spr = intel_plane->wm;
2162                         break;
2163                 }
2164         }
2165 }
2166
2167 static void ilk_compute_wm_config(struct drm_device *dev,
2168                                   struct intel_wm_config *config)
2169 {
2170         struct intel_crtc *intel_crtc;
2171
2172         /* Compute the currently _active_ config */
2173         for_each_intel_crtc(dev, intel_crtc) {
2174                 const struct intel_pipe_wm *wm = &intel_crtc->wm.active;
2175
2176                 if (!wm->pipe_enabled)
2177                         continue;
2178
2179                 config->sprites_enabled |= wm->sprites_enabled;
2180                 config->sprites_scaled |= wm->sprites_scaled;
2181                 config->num_pipes_active++;
2182         }
2183 }
2184
2185 /* Compute new watermarks for the pipe */
2186 static bool intel_compute_pipe_wm(struct drm_crtc *crtc,
2187                                   const struct ilk_pipe_wm_parameters *params,
2188                                   struct intel_pipe_wm *pipe_wm)
2189 {
2190         struct drm_device *dev = crtc->dev;
2191         const struct drm_i915_private *dev_priv = dev->dev_private;
2192         int level, max_level = ilk_wm_max_level(dev);
2193         /* LP0 watermark maximums depend on this pipe alone */
2194         struct intel_wm_config config = {
2195                 .num_pipes_active = 1,
2196                 .sprites_enabled = params->spr.enabled,
2197                 .sprites_scaled = params->spr.scaled,
2198         };
2199         struct ilk_wm_maximums max;
2200
2201         pipe_wm->pipe_enabled = params->active;
2202         pipe_wm->sprites_enabled = params->spr.enabled;
2203         pipe_wm->sprites_scaled = params->spr.scaled;
2204
2205         /* ILK/SNB: LP2+ watermarks only w/o sprites */
2206         if (INTEL_INFO(dev)->gen <= 6 && params->spr.enabled)
2207                 max_level = 1;
2208
2209         /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2210         if (params->spr.scaled)
2211                 max_level = 0;
2212
2213         ilk_compute_wm_level(dev_priv, 0, params, &pipe_wm->wm[0]);
2214
2215         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2216                 pipe_wm->linetime = hsw_compute_linetime_wm(dev, crtc);
2217
2218         /* LP0 watermarks always use 1/2 DDB partitioning */
2219         ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
2220
2221         /* At least LP0 must be valid */
2222         if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0]))
2223                 return false;
2224
2225         ilk_compute_wm_reg_maximums(dev, 1, &max);
2226
2227         for (level = 1; level <= max_level; level++) {
2228                 struct intel_wm_level wm = {};
2229
2230                 ilk_compute_wm_level(dev_priv, level, params, &wm);
2231
2232                 /*
2233                  * Disable any watermark level that exceeds the
2234                  * register maximums since such watermarks are
2235                  * always invalid.
2236                  */
2237                 if (!ilk_validate_wm_level(level, &max, &wm))
2238                         break;
2239
2240                 pipe_wm->wm[level] = wm;
2241         }
2242
2243         return true;
2244 }
2245
2246 /*
2247  * Merge the watermarks from all active pipes for a specific level.
2248  */
2249 static void ilk_merge_wm_level(struct drm_device *dev,
2250                                int level,
2251                                struct intel_wm_level *ret_wm)
2252 {
2253         const struct intel_crtc *intel_crtc;
2254
2255         ret_wm->enable = true;
2256
2257         for_each_intel_crtc(dev, intel_crtc) {
2258                 const struct intel_pipe_wm *active = &intel_crtc->wm.active;
2259                 const struct intel_wm_level *wm = &active->wm[level];
2260
2261                 if (!active->pipe_enabled)
2262                         continue;
2263
2264                 /*
2265                  * The watermark values may have been used in the past,
2266                  * so we must maintain them in the registers for some
2267                  * time even if the level is now disabled.
2268                  */
2269                 if (!wm->enable)
2270                         ret_wm->enable = false;
2271
2272                 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
2273                 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
2274                 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
2275                 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
2276         }
2277 }
2278
2279 /*
2280  * Merge all low power watermarks for all active pipes.
2281  */
2282 static void ilk_wm_merge(struct drm_device *dev,
2283                          const struct intel_wm_config *config,
2284                          const struct ilk_wm_maximums *max,
2285                          struct intel_pipe_wm *merged)
2286 {
2287         int level, max_level = ilk_wm_max_level(dev);
2288         int last_enabled_level = max_level;
2289
2290         /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2291         if ((INTEL_INFO(dev)->gen <= 6 || IS_IVYBRIDGE(dev)) &&
2292             config->num_pipes_active > 1)
2293                 return;
2294
2295         /* ILK: FBC WM must be disabled always */
2296         merged->fbc_wm_enabled = INTEL_INFO(dev)->gen >= 6;
2297
2298         /* merge each WM1+ level */
2299         for (level = 1; level <= max_level; level++) {
2300                 struct intel_wm_level *wm = &merged->wm[level];
2301
2302                 ilk_merge_wm_level(dev, level, wm);
2303
2304                 if (level > last_enabled_level)
2305                         wm->enable = false;
2306                 else if (!ilk_validate_wm_level(level, max, wm))
2307                         /* make sure all following levels get disabled */
2308                         last_enabled_level = level - 1;
2309
2310                 /*
2311                  * The spec says it is preferred to disable
2312                  * FBC WMs instead of disabling a WM level.
2313                  */
2314                 if (wm->fbc_val > max->fbc) {
2315                         if (wm->enable)
2316                                 merged->fbc_wm_enabled = false;
2317                         wm->fbc_val = 0;
2318                 }
2319         }
2320
2321         /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2322         /*
2323          * FIXME this is racy. FBC might get enabled later.
2324          * What we should check here is whether FBC can be
2325          * enabled sometime later.
2326          */
2327         if (IS_GEN5(dev) && !merged->fbc_wm_enabled && intel_fbc_enabled(dev)) {
2328                 for (level = 2; level <= max_level; level++) {
2329                         struct intel_wm_level *wm = &merged->wm[level];
2330
2331                         wm->enable = false;
2332                 }
2333         }
2334 }
2335
2336 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
2337 {
2338         /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2339         return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
2340 }
2341
2342 /* The value we need to program into the WM_LPx latency field */
2343 static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
2344 {
2345         struct drm_i915_private *dev_priv = dev->dev_private;
2346
2347         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2348                 return 2 * level;
2349         else
2350                 return dev_priv->wm.pri_latency[level];
2351 }
2352
2353 static void ilk_compute_wm_results(struct drm_device *dev,
2354                                    const struct intel_pipe_wm *merged,
2355                                    enum intel_ddb_partitioning partitioning,
2356                                    struct ilk_wm_values *results)
2357 {
2358         struct intel_crtc *intel_crtc;
2359         int level, wm_lp;
2360
2361         results->enable_fbc_wm = merged->fbc_wm_enabled;
2362         results->partitioning = partitioning;
2363
2364         /* LP1+ register values */
2365         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2366                 const struct intel_wm_level *r;
2367
2368                 level = ilk_wm_lp_to_level(wm_lp, merged);
2369
2370                 r = &merged->wm[level];
2371
2372                 /*
2373                  * Maintain the watermark values even if the level is
2374                  * disabled. Doing otherwise could cause underruns.
2375                  */
2376                 results->wm_lp[wm_lp - 1] =
2377                         (ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
2378                         (r->pri_val << WM1_LP_SR_SHIFT) |
2379                         r->cur_val;
2380
2381                 if (r->enable)
2382                         results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
2383
2384                 if (INTEL_INFO(dev)->gen >= 8)
2385                         results->wm_lp[wm_lp - 1] |=
2386                                 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
2387                 else
2388                         results->wm_lp[wm_lp - 1] |=
2389                                 r->fbc_val << WM1_LP_FBC_SHIFT;
2390
2391                 /*
2392                  * Always set WM1S_LP_EN when spr_val != 0, even if the
2393                  * level is disabled. Doing otherwise could cause underruns.
2394                  */
2395                 if (INTEL_INFO(dev)->gen <= 6 && r->spr_val) {
2396                         WARN_ON(wm_lp != 1);
2397                         results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
2398                 } else
2399                         results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2400         }
2401
2402         /* LP0 register values */
2403         for_each_intel_crtc(dev, intel_crtc) {
2404                 enum pipe pipe = intel_crtc->pipe;
2405                 const struct intel_wm_level *r =
2406                         &intel_crtc->wm.active.wm[0];
2407
2408                 if (WARN_ON(!r->enable))
2409                         continue;
2410
2411                 results->wm_linetime[pipe] = intel_crtc->wm.active.linetime;
2412
2413                 results->wm_pipe[pipe] =
2414                         (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
2415                         (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
2416                         r->cur_val;
2417         }
2418 }
2419
2420 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2421  * case both are at the same level. Prefer r1 in case they're the same. */
2422 static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
2423                                                   struct intel_pipe_wm *r1,
2424                                                   struct intel_pipe_wm *r2)
2425 {
2426         int level, max_level = ilk_wm_max_level(dev);
2427         int level1 = 0, level2 = 0;
2428
2429         for (level = 1; level <= max_level; level++) {
2430                 if (r1->wm[level].enable)
2431                         level1 = level;
2432                 if (r2->wm[level].enable)
2433                         level2 = level;
2434         }
2435
2436         if (level1 == level2) {
2437                 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
2438                         return r2;
2439                 else
2440                         return r1;
2441         } else if (level1 > level2) {
2442                 return r1;
2443         } else {
2444                 return r2;
2445         }
2446 }
2447
2448 /* dirty bits used to track which watermarks need changes */
2449 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2450 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2451 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2452 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2453 #define WM_DIRTY_FBC (1 << 24)
2454 #define WM_DIRTY_DDB (1 << 25)
2455
2456 static unsigned int ilk_compute_wm_dirty(struct drm_device *dev,
2457                                          const struct ilk_wm_values *old,
2458                                          const struct ilk_wm_values *new)
2459 {
2460         unsigned int dirty = 0;
2461         enum pipe pipe;
2462         int wm_lp;
2463
2464         for_each_pipe(pipe) {
2465                 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
2466                         dirty |= WM_DIRTY_LINETIME(pipe);
2467                         /* Must disable LP1+ watermarks too */
2468                         dirty |= WM_DIRTY_LP_ALL;
2469                 }
2470
2471                 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
2472                         dirty |= WM_DIRTY_PIPE(pipe);
2473                         /* Must disable LP1+ watermarks too */
2474                         dirty |= WM_DIRTY_LP_ALL;
2475                 }
2476         }
2477
2478         if (old->enable_fbc_wm != new->enable_fbc_wm) {
2479                 dirty |= WM_DIRTY_FBC;
2480                 /* Must disable LP1+ watermarks too */
2481                 dirty |= WM_DIRTY_LP_ALL;
2482         }
2483
2484         if (old->partitioning != new->partitioning) {
2485                 dirty |= WM_DIRTY_DDB;
2486                 /* Must disable LP1+ watermarks too */
2487                 dirty |= WM_DIRTY_LP_ALL;
2488         }
2489
2490         /* LP1+ watermarks already deemed dirty, no need to continue */
2491         if (dirty & WM_DIRTY_LP_ALL)
2492                 return dirty;
2493
2494         /* Find the lowest numbered LP1+ watermark in need of an update... */
2495         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2496                 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
2497                     old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
2498                         break;
2499         }
2500
2501         /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2502         for (; wm_lp <= 3; wm_lp++)
2503                 dirty |= WM_DIRTY_LP(wm_lp);
2504
2505         return dirty;
2506 }
2507
2508 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
2509                                unsigned int dirty)
2510 {
2511         struct ilk_wm_values *previous = &dev_priv->wm.hw;
2512         bool changed = false;
2513
2514         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
2515                 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
2516                 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
2517                 changed = true;
2518         }
2519         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
2520                 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
2521                 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
2522                 changed = true;
2523         }
2524         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
2525                 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
2526                 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
2527                 changed = true;
2528         }
2529
2530         /*
2531          * Don't touch WM1S_LP_EN here.
2532          * Doing so could cause underruns.
2533          */
2534
2535         return changed;
2536 }
2537
2538 /*
2539  * The spec says we shouldn't write when we don't need, because every write
2540  * causes WMs to be re-evaluated, expending some power.
2541  */
2542 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
2543                                 struct ilk_wm_values *results)
2544 {
2545         struct drm_device *dev = dev_priv->dev;
2546         struct ilk_wm_values *previous = &dev_priv->wm.hw;
2547         unsigned int dirty;
2548         uint32_t val;
2549
2550         dirty = ilk_compute_wm_dirty(dev, previous, results);
2551         if (!dirty)
2552                 return;
2553
2554         _ilk_disable_lp_wm(dev_priv, dirty);
2555
2556         if (dirty & WM_DIRTY_PIPE(PIPE_A))
2557                 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2558         if (dirty & WM_DIRTY_PIPE(PIPE_B))
2559                 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
2560         if (dirty & WM_DIRTY_PIPE(PIPE_C))
2561                 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
2562
2563         if (dirty & WM_DIRTY_LINETIME(PIPE_A))
2564                 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
2565         if (dirty & WM_DIRTY_LINETIME(PIPE_B))
2566                 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
2567         if (dirty & WM_DIRTY_LINETIME(PIPE_C))
2568                 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
2569
2570         if (dirty & WM_DIRTY_DDB) {
2571                 if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2572                         val = I915_READ(WM_MISC);
2573                         if (results->partitioning == INTEL_DDB_PART_1_2)
2574                                 val &= ~WM_MISC_DATA_PARTITION_5_6;
2575                         else
2576                                 val |= WM_MISC_DATA_PARTITION_5_6;
2577                         I915_WRITE(WM_MISC, val);
2578                 } else {
2579                         val = I915_READ(DISP_ARB_CTL2);
2580                         if (results->partitioning == INTEL_DDB_PART_1_2)
2581                                 val &= ~DISP_DATA_PARTITION_5_6;
2582                         else
2583                                 val |= DISP_DATA_PARTITION_5_6;
2584                         I915_WRITE(DISP_ARB_CTL2, val);
2585                 }
2586         }
2587
2588         if (dirty & WM_DIRTY_FBC) {
2589                 val = I915_READ(DISP_ARB_CTL);
2590                 if (results->enable_fbc_wm)
2591                         val &= ~DISP_FBC_WM_DIS;
2592                 else
2593                         val |= DISP_FBC_WM_DIS;
2594                 I915_WRITE(DISP_ARB_CTL, val);
2595         }
2596
2597         if (dirty & WM_DIRTY_LP(1) &&
2598             previous->wm_lp_spr[0] != results->wm_lp_spr[0])
2599                 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
2600
2601         if (INTEL_INFO(dev)->gen >= 7) {
2602                 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
2603                         I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
2604                 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
2605                         I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
2606         }
2607
2608         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
2609                 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
2610         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
2611                 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
2612         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
2613                 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
2614
2615         dev_priv->wm.hw = *results;
2616 }
2617
2618 static bool ilk_disable_lp_wm(struct drm_device *dev)
2619 {
2620         struct drm_i915_private *dev_priv = dev->dev_private;
2621
2622         return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
2623 }
2624
2625 static void ilk_update_wm(struct drm_crtc *crtc)
2626 {
2627         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2628         struct drm_device *dev = crtc->dev;
2629         struct drm_i915_private *dev_priv = dev->dev_private;
2630         struct ilk_wm_maximums max;
2631         struct ilk_pipe_wm_parameters params = {};
2632         struct ilk_wm_values results = {};
2633         enum intel_ddb_partitioning partitioning;
2634         struct intel_pipe_wm pipe_wm = {};
2635         struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
2636         struct intel_wm_config config = {};
2637
2638         ilk_compute_wm_parameters(crtc, &params);
2639
2640         intel_compute_pipe_wm(crtc, &params, &pipe_wm);
2641
2642         if (!memcmp(&intel_crtc->wm.active, &pipe_wm, sizeof(pipe_wm)))
2643                 return;
2644
2645         intel_crtc->wm.active = pipe_wm;
2646
2647         ilk_compute_wm_config(dev, &config);
2648
2649         ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
2650         ilk_wm_merge(dev, &config, &max, &lp_wm_1_2);
2651
2652         /* 5/6 split only in single pipe config on IVB+ */
2653         if (INTEL_INFO(dev)->gen >= 7 &&
2654             config.num_pipes_active == 1 && config.sprites_enabled) {
2655                 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
2656                 ilk_wm_merge(dev, &config, &max, &lp_wm_5_6);
2657
2658                 best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
2659         } else {
2660                 best_lp_wm = &lp_wm_1_2;
2661         }
2662
2663         partitioning = (best_lp_wm == &lp_wm_1_2) ?
2664                        INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
2665
2666         ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
2667
2668         ilk_write_wm_values(dev_priv, &results);
2669 }
2670
2671 static void ilk_update_sprite_wm(struct drm_plane *plane,
2672                                      struct drm_crtc *crtc,
2673                                      uint32_t sprite_width, int pixel_size,
2674                                      bool enabled, bool scaled)
2675 {
2676         struct drm_device *dev = plane->dev;
2677         struct intel_plane *intel_plane = to_intel_plane(plane);
2678
2679         intel_plane->wm.enabled = enabled;
2680         intel_plane->wm.scaled = scaled;
2681         intel_plane->wm.horiz_pixels = sprite_width;
2682         intel_plane->wm.bytes_per_pixel = pixel_size;
2683
2684         /*
2685          * IVB workaround: must disable low power watermarks for at least
2686          * one frame before enabling scaling.  LP watermarks can be re-enabled
2687          * when scaling is disabled.
2688          *
2689          * WaCxSRDisabledForSpriteScaling:ivb
2690          */
2691         if (IS_IVYBRIDGE(dev) && scaled && ilk_disable_lp_wm(dev))
2692                 intel_wait_for_vblank(dev, intel_plane->pipe);
2693
2694         ilk_update_wm(crtc);
2695 }
2696
2697 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
2698 {
2699         struct drm_device *dev = crtc->dev;
2700         struct drm_i915_private *dev_priv = dev->dev_private;
2701         struct ilk_wm_values *hw = &dev_priv->wm.hw;
2702         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2703         struct intel_pipe_wm *active = &intel_crtc->wm.active;
2704         enum pipe pipe = intel_crtc->pipe;
2705         static const unsigned int wm0_pipe_reg[] = {
2706                 [PIPE_A] = WM0_PIPEA_ILK,
2707                 [PIPE_B] = WM0_PIPEB_ILK,
2708                 [PIPE_C] = WM0_PIPEC_IVB,
2709         };
2710
2711         hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
2712         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2713                 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
2714
2715         active->pipe_enabled = intel_crtc_active(crtc);
2716
2717         if (active->pipe_enabled) {
2718                 u32 tmp = hw->wm_pipe[pipe];
2719
2720                 /*
2721                  * For active pipes LP0 watermark is marked as
2722                  * enabled, and LP1+ watermaks as disabled since
2723                  * we can't really reverse compute them in case
2724                  * multiple pipes are active.
2725                  */
2726                 active->wm[0].enable = true;
2727                 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
2728                 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
2729                 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
2730                 active->linetime = hw->wm_linetime[pipe];
2731         } else {
2732                 int level, max_level = ilk_wm_max_level(dev);
2733
2734                 /*
2735                  * For inactive pipes, all watermark levels
2736                  * should be marked as enabled but zeroed,
2737                  * which is what we'd compute them to.
2738                  */
2739                 for (level = 0; level <= max_level; level++)
2740                         active->wm[level].enable = true;
2741         }
2742 }
2743
2744 void ilk_wm_get_hw_state(struct drm_device *dev)
2745 {
2746         struct drm_i915_private *dev_priv = dev->dev_private;
2747         struct ilk_wm_values *hw = &dev_priv->wm.hw;
2748         struct drm_crtc *crtc;
2749
2750         for_each_crtc(dev, crtc)
2751                 ilk_pipe_wm_get_hw_state(crtc);
2752
2753         hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
2754         hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
2755         hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
2756
2757         hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
2758         if (INTEL_INFO(dev)->gen >= 7) {
2759                 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
2760                 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
2761         }
2762
2763         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2764                 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
2765                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
2766         else if (IS_IVYBRIDGE(dev))
2767                 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
2768                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
2769
2770         hw->enable_fbc_wm =
2771                 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
2772 }
2773
2774 /**
2775  * intel_update_watermarks - update FIFO watermark values based on current modes
2776  *
2777  * Calculate watermark values for the various WM regs based on current mode
2778  * and plane configuration.
2779  *
2780  * There are several cases to deal with here:
2781  *   - normal (i.e. non-self-refresh)
2782  *   - self-refresh (SR) mode
2783  *   - lines are large relative to FIFO size (buffer can hold up to 2)
2784  *   - lines are small relative to FIFO size (buffer can hold more than 2
2785  *     lines), so need to account for TLB latency
2786  *
2787  *   The normal calculation is:
2788  *     watermark = dotclock * bytes per pixel * latency
2789  *   where latency is platform & configuration dependent (we assume pessimal
2790  *   values here).
2791  *
2792  *   The SR calculation is:
2793  *     watermark = (trunc(latency/line time)+1) * surface width *
2794  *       bytes per pixel
2795  *   where
2796  *     line time = htotal / dotclock
2797  *     surface width = hdisplay for normal plane and 64 for cursor
2798  *   and latency is assumed to be high, as above.
2799  *
2800  * The final value programmed to the register should always be rounded up,
2801  * and include an extra 2 entries to account for clock crossings.
2802  *
2803  * We don't use the sprite, so we can ignore that.  And on Crestline we have
2804  * to set the non-SR watermarks to 8.
2805  */
2806 void intel_update_watermarks(struct drm_crtc *crtc)
2807 {
2808         struct drm_i915_private *dev_priv = crtc->dev->dev_private;
2809
2810         if (dev_priv->display.update_wm)
2811                 dev_priv->display.update_wm(crtc);
2812 }
2813
2814 void intel_update_sprite_watermarks(struct drm_plane *plane,
2815                                     struct drm_crtc *crtc,
2816                                     uint32_t sprite_width, int pixel_size,
2817                                     bool enabled, bool scaled)
2818 {
2819         struct drm_i915_private *dev_priv = plane->dev->dev_private;
2820
2821         if (dev_priv->display.update_sprite_wm)
2822                 dev_priv->display.update_sprite_wm(plane, crtc, sprite_width,
2823                                                    pixel_size, enabled, scaled);
2824 }
2825
2826 static struct drm_i915_gem_object *
2827 intel_alloc_context_page(struct drm_device *dev)
2828 {
2829         struct drm_i915_gem_object *ctx;
2830         int ret;
2831
2832         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2833
2834         ctx = i915_gem_alloc_object(dev, 4096);
2835         if (!ctx) {
2836                 DRM_DEBUG("failed to alloc power context, RC6 disabled\n");
2837                 return NULL;
2838         }
2839
2840         ret = i915_gem_obj_ggtt_pin(ctx, 4096, 0);
2841         if (ret) {
2842                 DRM_ERROR("failed to pin power context: %d\n", ret);
2843                 goto err_unref;
2844         }
2845
2846         ret = i915_gem_object_set_to_gtt_domain(ctx, 1);
2847         if (ret) {
2848                 DRM_ERROR("failed to set-domain on power context: %d\n", ret);
2849                 goto err_unpin;
2850         }
2851
2852         return ctx;
2853
2854 err_unpin:
2855         i915_gem_object_ggtt_unpin(ctx);
2856 err_unref:
2857         drm_gem_object_unreference(&ctx->base);
2858         return NULL;
2859 }
2860
2861 /**
2862  * Lock protecting IPS related data structures
2863  */
2864 DEFINE_SPINLOCK(mchdev_lock);
2865
2866 /* Global for IPS driver to get at the current i915 device. Protected by
2867  * mchdev_lock. */
2868 static struct drm_i915_private *i915_mch_dev;
2869
2870 bool ironlake_set_drps(struct drm_device *dev, u8 val)
2871 {
2872         struct drm_i915_private *dev_priv = dev->dev_private;
2873         u16 rgvswctl;
2874
2875         assert_spin_locked(&mchdev_lock);
2876
2877         rgvswctl = I915_READ16(MEMSWCTL);
2878         if (rgvswctl & MEMCTL_CMD_STS) {
2879                 DRM_DEBUG("gpu busy, RCS change rejected\n");
2880                 return false; /* still busy with another command */
2881         }
2882
2883         rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
2884                 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
2885         I915_WRITE16(MEMSWCTL, rgvswctl);
2886         POSTING_READ16(MEMSWCTL);
2887
2888         rgvswctl |= MEMCTL_CMD_STS;
2889         I915_WRITE16(MEMSWCTL, rgvswctl);
2890
2891         return true;
2892 }
2893
2894 static void ironlake_enable_drps(struct drm_device *dev)
2895 {
2896         struct drm_i915_private *dev_priv = dev->dev_private;
2897         u32 rgvmodectl = I915_READ(MEMMODECTL);
2898         u8 fmax, fmin, fstart, vstart;
2899
2900         spin_lock_irq(&mchdev_lock);
2901
2902         /* Enable temp reporting */
2903         I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
2904         I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
2905
2906         /* 100ms RC evaluation intervals */
2907         I915_WRITE(RCUPEI, 100000);
2908         I915_WRITE(RCDNEI, 100000);
2909
2910         /* Set max/min thresholds to 90ms and 80ms respectively */
2911         I915_WRITE(RCBMAXAVG, 90000);
2912         I915_WRITE(RCBMINAVG, 80000);
2913
2914         I915_WRITE(MEMIHYST, 1);
2915
2916         /* Set up min, max, and cur for interrupt handling */
2917         fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
2918         fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
2919         fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
2920                 MEMMODE_FSTART_SHIFT;
2921
2922         vstart = (I915_READ(PXVFREQ_BASE + (fstart * 4)) & PXVFREQ_PX_MASK) >>
2923                 PXVFREQ_PX_SHIFT;
2924
2925         dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
2926         dev_priv->ips.fstart = fstart;
2927
2928         dev_priv->ips.max_delay = fstart;
2929         dev_priv->ips.min_delay = fmin;
2930         dev_priv->ips.cur_delay = fstart;
2931
2932         DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
2933                          fmax, fmin, fstart);
2934
2935         I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
2936
2937         /*
2938          * Interrupts will be enabled in ironlake_irq_postinstall
2939          */
2940
2941         I915_WRITE(VIDSTART, vstart);
2942         POSTING_READ(VIDSTART);
2943
2944         rgvmodectl |= MEMMODE_SWMODE_EN;
2945         I915_WRITE(MEMMODECTL, rgvmodectl);
2946
2947         if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
2948                 DRM_ERROR("stuck trying to change perf mode\n");
2949         mdelay(1);
2950
2951         ironlake_set_drps(dev, fstart);
2952
2953         dev_priv->ips.last_count1 = I915_READ(0x112e4) + I915_READ(0x112e8) +
2954                 I915_READ(0x112e0);
2955         dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
2956         dev_priv->ips.last_count2 = I915_READ(0x112f4);
2957         getrawmonotonic(&dev_priv->ips.last_time2);
2958
2959         spin_unlock_irq(&mchdev_lock);
2960 }
2961
2962 static void ironlake_disable_drps(struct drm_device *dev)
2963 {
2964         struct drm_i915_private *dev_priv = dev->dev_private;
2965         u16 rgvswctl;
2966
2967         spin_lock_irq(&mchdev_lock);
2968
2969         rgvswctl = I915_READ16(MEMSWCTL);
2970
2971         /* Ack interrupts, disable EFC interrupt */
2972         I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
2973         I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
2974         I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
2975         I915_WRITE(DEIIR, DE_PCU_EVENT);
2976         I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
2977
2978         /* Go back to the starting frequency */
2979         ironlake_set_drps(dev, dev_priv->ips.fstart);
2980         mdelay(1);
2981         rgvswctl |= MEMCTL_CMD_STS;
2982         I915_WRITE(MEMSWCTL, rgvswctl);
2983         mdelay(1);
2984
2985         spin_unlock_irq(&mchdev_lock);
2986 }
2987
2988 /* There's a funny hw issue where the hw returns all 0 when reading from
2989  * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
2990  * ourselves, instead of doing a rmw cycle (which might result in us clearing
2991  * all limits and the gpu stuck at whatever frequency it is at atm).
2992  */
2993 static u32 gen6_rps_limits(struct drm_i915_private *dev_priv, u8 val)
2994 {
2995         u32 limits;
2996
2997         /* Only set the down limit when we've reached the lowest level to avoid
2998          * getting more interrupts, otherwise leave this clear. This prevents a
2999          * race in the hw when coming out of rc6: There's a tiny window where
3000          * the hw runs at the minimal clock before selecting the desired
3001          * frequency, if the down threshold expires in that window we will not
3002          * receive a down interrupt. */
3003         limits = dev_priv->rps.max_freq_softlimit << 24;
3004         if (val <= dev_priv->rps.min_freq_softlimit)
3005                 limits |= dev_priv->rps.min_freq_softlimit << 16;
3006
3007         return limits;
3008 }
3009
3010 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
3011 {
3012         int new_power;
3013
3014         new_power = dev_priv->rps.power;
3015         switch (dev_priv->rps.power) {
3016         case LOW_POWER:
3017                 if (val > dev_priv->rps.efficient_freq + 1 && val > dev_priv->rps.cur_freq)
3018                         new_power = BETWEEN;
3019                 break;
3020
3021         case BETWEEN:
3022                 if (val <= dev_priv->rps.efficient_freq && val < dev_priv->rps.cur_freq)
3023                         new_power = LOW_POWER;
3024                 else if (val >= dev_priv->rps.rp0_freq && val > dev_priv->rps.cur_freq)
3025                         new_power = HIGH_POWER;
3026                 break;
3027
3028         case HIGH_POWER:
3029                 if (val < (dev_priv->rps.rp1_freq + dev_priv->rps.rp0_freq) >> 1 && val < dev_priv->rps.cur_freq)
3030                         new_power = BETWEEN;
3031                 break;
3032         }
3033         /* Max/min bins are special */
3034         if (val == dev_priv->rps.min_freq_softlimit)
3035                 new_power = LOW_POWER;
3036         if (val == dev_priv->rps.max_freq_softlimit)
3037                 new_power = HIGH_POWER;
3038         if (new_power == dev_priv->rps.power)
3039                 return;
3040
3041         /* Note the units here are not exactly 1us, but 1280ns. */
3042         switch (new_power) {
3043         case LOW_POWER:
3044                 /* Upclock if more than 95% busy over 16ms */
3045                 I915_WRITE(GEN6_RP_UP_EI, 12500);
3046                 I915_WRITE(GEN6_RP_UP_THRESHOLD, 11800);
3047
3048                 /* Downclock if less than 85% busy over 32ms */
3049                 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3050                 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 21250);
3051
3052                 I915_WRITE(GEN6_RP_CONTROL,
3053                            GEN6_RP_MEDIA_TURBO |
3054                            GEN6_RP_MEDIA_HW_NORMAL_MODE |
3055                            GEN6_RP_MEDIA_IS_GFX |
3056                            GEN6_RP_ENABLE |
3057                            GEN6_RP_UP_BUSY_AVG |
3058                            GEN6_RP_DOWN_IDLE_AVG);
3059                 break;
3060
3061         case BETWEEN:
3062                 /* Upclock if more than 90% busy over 13ms */
3063                 I915_WRITE(GEN6_RP_UP_EI, 10250);
3064                 I915_WRITE(GEN6_RP_UP_THRESHOLD, 9225);
3065
3066                 /* Downclock if less than 75% busy over 32ms */
3067                 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3068                 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 18750);
3069
3070                 I915_WRITE(GEN6_RP_CONTROL,
3071                            GEN6_RP_MEDIA_TURBO |
3072                            GEN6_RP_MEDIA_HW_NORMAL_MODE |
3073                            GEN6_RP_MEDIA_IS_GFX |
3074                            GEN6_RP_ENABLE |
3075                            GEN6_RP_UP_BUSY_AVG |
3076                            GEN6_RP_DOWN_IDLE_AVG);
3077                 break;
3078
3079         case HIGH_POWER:
3080                 /* Upclock if more than 85% busy over 10ms */
3081                 I915_WRITE(GEN6_RP_UP_EI, 8000);
3082                 I915_WRITE(GEN6_RP_UP_THRESHOLD, 6800);
3083
3084                 /* Downclock if less than 60% busy over 32ms */
3085                 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3086                 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 15000);
3087
3088                 I915_WRITE(GEN6_RP_CONTROL,
3089                            GEN6_RP_MEDIA_TURBO |
3090                            GEN6_RP_MEDIA_HW_NORMAL_MODE |
3091                            GEN6_RP_MEDIA_IS_GFX |
3092                            GEN6_RP_ENABLE |
3093                            GEN6_RP_UP_BUSY_AVG |
3094                            GEN6_RP_DOWN_IDLE_AVG);
3095                 break;
3096         }
3097
3098         dev_priv->rps.power = new_power;
3099         dev_priv->rps.last_adj = 0;
3100 }
3101
3102 static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
3103 {
3104         u32 mask = 0;
3105
3106         if (val > dev_priv->rps.min_freq_softlimit)
3107                 mask |= GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
3108         if (val < dev_priv->rps.max_freq_softlimit)
3109                 mask |= GEN6_PM_RP_UP_THRESHOLD;
3110
3111         /* IVB and SNB hard hangs on looping batchbuffer
3112          * if GEN6_PM_UP_EI_EXPIRED is masked.
3113          */
3114         if (INTEL_INFO(dev_priv->dev)->gen <= 7 && !IS_HASWELL(dev_priv->dev))
3115                 mask |= GEN6_PM_RP_UP_EI_EXPIRED;
3116
3117         if (IS_GEN8(dev_priv->dev))
3118                 mask |= GEN8_PMINTR_REDIRECT_TO_NON_DISP;
3119
3120         return ~mask;
3121 }
3122
3123 /* gen6_set_rps is called to update the frequency request, but should also be
3124  * called when the range (min_delay and max_delay) is modified so that we can
3125  * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
3126 void gen6_set_rps(struct drm_device *dev, u8 val)
3127 {
3128         struct drm_i915_private *dev_priv = dev->dev_private;
3129
3130         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3131         WARN_ON(val > dev_priv->rps.max_freq_softlimit);
3132         WARN_ON(val < dev_priv->rps.min_freq_softlimit);
3133
3134         /* min/max delay may still have been modified so be sure to
3135          * write the limits value.
3136          */
3137         if (val != dev_priv->rps.cur_freq) {
3138                 gen6_set_rps_thresholds(dev_priv, val);
3139
3140                 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
3141                         I915_WRITE(GEN6_RPNSWREQ,
3142                                    HSW_FREQUENCY(val));
3143                 else
3144                         I915_WRITE(GEN6_RPNSWREQ,
3145                                    GEN6_FREQUENCY(val) |
3146                                    GEN6_OFFSET(0) |
3147                                    GEN6_AGGRESSIVE_TURBO);
3148         }
3149
3150         /* Make sure we continue to get interrupts
3151          * until we hit the minimum or maximum frequencies.
3152          */
3153         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, gen6_rps_limits(dev_priv, val));
3154         I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
3155
3156         POSTING_READ(GEN6_RPNSWREQ);
3157
3158         dev_priv->rps.cur_freq = val;
3159         trace_intel_gpu_freq_change(val * 50);
3160 }
3161
3162 /* vlv_set_rps_idle: Set the frequency to Rpn if Gfx clocks are down
3163  *
3164  * * If Gfx is Idle, then
3165  * 1. Mask Turbo interrupts
3166  * 2. Bring up Gfx clock
3167  * 3. Change the freq to Rpn and wait till P-Unit updates freq
3168  * 4. Clear the Force GFX CLK ON bit so that Gfx can down
3169  * 5. Unmask Turbo interrupts
3170 */
3171 static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
3172 {
3173         /*
3174          * When we are idle.  Drop to min voltage state.
3175          */
3176
3177         if (dev_priv->rps.cur_freq <= dev_priv->rps.min_freq_softlimit)
3178                 return;
3179
3180         /* Mask turbo interrupt so that they will not come in between */
3181         I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
3182
3183         vlv_force_gfx_clock(dev_priv, true);
3184
3185         dev_priv->rps.cur_freq = dev_priv->rps.min_freq_softlimit;
3186
3187         vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ,
3188                                         dev_priv->rps.min_freq_softlimit);
3189
3190         if (wait_for(((vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS))
3191                                 & GENFREQSTATUS) == 0, 5))
3192                 DRM_ERROR("timed out waiting for Punit\n");
3193
3194         vlv_force_gfx_clock(dev_priv, false);
3195
3196         I915_WRITE(GEN6_PMINTRMSK,
3197                    gen6_rps_pm_mask(dev_priv, dev_priv->rps.cur_freq));
3198 }
3199
3200 void gen6_rps_idle(struct drm_i915_private *dev_priv)
3201 {
3202         struct drm_device *dev = dev_priv->dev;
3203
3204         mutex_lock(&dev_priv->rps.hw_lock);
3205         if (dev_priv->rps.enabled) {
3206                 if (IS_VALLEYVIEW(dev))
3207                         vlv_set_rps_idle(dev_priv);
3208                 else
3209                         gen6_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
3210                 dev_priv->rps.last_adj = 0;
3211         }
3212         mutex_unlock(&dev_priv->rps.hw_lock);
3213 }
3214
3215 void gen6_rps_boost(struct drm_i915_private *dev_priv)
3216 {
3217         struct drm_device *dev = dev_priv->dev;
3218
3219         mutex_lock(&dev_priv->rps.hw_lock);
3220         if (dev_priv->rps.enabled) {
3221                 if (IS_VALLEYVIEW(dev))
3222                         valleyview_set_rps(dev_priv->dev, dev_priv->rps.max_freq_softlimit);
3223                 else
3224                         gen6_set_rps(dev_priv->dev, dev_priv->rps.max_freq_softlimit);
3225                 dev_priv->rps.last_adj = 0;
3226         }
3227         mutex_unlock(&dev_priv->rps.hw_lock);
3228 }
3229
3230 void valleyview_set_rps(struct drm_device *dev, u8 val)
3231 {
3232         struct drm_i915_private *dev_priv = dev->dev_private;
3233
3234         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3235         WARN_ON(val > dev_priv->rps.max_freq_softlimit);
3236         WARN_ON(val < dev_priv->rps.min_freq_softlimit);
3237
3238         DRM_DEBUG_DRIVER("GPU freq request from %d MHz (%u) to %d MHz (%u)\n",
3239                          vlv_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
3240                          dev_priv->rps.cur_freq,
3241                          vlv_gpu_freq(dev_priv, val), val);
3242
3243         if (val != dev_priv->rps.cur_freq)
3244                 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
3245
3246         I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
3247
3248         dev_priv->rps.cur_freq = val;
3249         trace_intel_gpu_freq_change(vlv_gpu_freq(dev_priv, val));
3250 }
3251
3252 static void gen8_disable_rps_interrupts(struct drm_device *dev)
3253 {
3254         struct drm_i915_private *dev_priv = dev->dev_private;
3255
3256         I915_WRITE(GEN6_PMINTRMSK, ~GEN8_PMINTR_REDIRECT_TO_NON_DISP);
3257         I915_WRITE(GEN8_GT_IER(2), I915_READ(GEN8_GT_IER(2)) &
3258                                    ~dev_priv->pm_rps_events);
3259         /* Complete PM interrupt masking here doesn't race with the rps work
3260          * item again unmasking PM interrupts because that is using a different
3261          * register (GEN8_GT_IMR(2)) to mask PM interrupts. The only risk is in
3262          * leaving stale bits in GEN8_GT_IIR(2) and GEN8_GT_IMR(2) which
3263          * gen8_enable_rps will clean up. */
3264
3265         spin_lock_irq(&dev_priv->irq_lock);
3266         dev_priv->rps.pm_iir = 0;
3267         spin_unlock_irq(&dev_priv->irq_lock);
3268
3269         I915_WRITE(GEN8_GT_IIR(2), dev_priv->pm_rps_events);
3270 }
3271
3272 static void gen6_disable_rps_interrupts(struct drm_device *dev)
3273 {
3274         struct drm_i915_private *dev_priv = dev->dev_private;
3275
3276         I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
3277         I915_WRITE(GEN6_PMIER, I915_READ(GEN6_PMIER) &
3278                                 ~dev_priv->pm_rps_events);
3279         /* Complete PM interrupt masking here doesn't race with the rps work
3280          * item again unmasking PM interrupts because that is using a different
3281          * register (PMIMR) to mask PM interrupts. The only risk is in leaving
3282          * stale bits in PMIIR and PMIMR which gen6_enable_rps will clean up. */
3283
3284         spin_lock_irq(&dev_priv->irq_lock);
3285         dev_priv->rps.pm_iir = 0;
3286         spin_unlock_irq(&dev_priv->irq_lock);
3287
3288         I915_WRITE(GEN6_PMIIR, dev_priv->pm_rps_events);
3289 }
3290
3291 static void gen6_disable_rps(struct drm_device *dev)
3292 {
3293         struct drm_i915_private *dev_priv = dev->dev_private;
3294
3295         I915_WRITE(GEN6_RC_CONTROL, 0);
3296         I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
3297
3298         if (IS_BROADWELL(dev))
3299                 gen8_disable_rps_interrupts(dev);
3300         else
3301                 gen6_disable_rps_interrupts(dev);
3302 }
3303
3304 static void valleyview_disable_rps(struct drm_device *dev)
3305 {
3306         struct drm_i915_private *dev_priv = dev->dev_private;
3307
3308         I915_WRITE(GEN6_RC_CONTROL, 0);
3309
3310         gen6_disable_rps_interrupts(dev);
3311 }
3312
3313 static void intel_print_rc6_info(struct drm_device *dev, u32 mode)
3314 {
3315         if (IS_VALLEYVIEW(dev)) {
3316                 if (mode & (GEN7_RC_CTL_TO_MODE | GEN6_RC_CTL_EI_MODE(1)))
3317                         mode = GEN6_RC_CTL_RC6_ENABLE;
3318                 else
3319                         mode = 0;
3320         }
3321         DRM_INFO("Enabling RC6 states: RC6 %s, RC6p %s, RC6pp %s\n",
3322                  (mode & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off",
3323                  (mode & GEN6_RC_CTL_RC6p_ENABLE) ? "on" : "off",
3324                  (mode & GEN6_RC_CTL_RC6pp_ENABLE) ? "on" : "off");
3325 }
3326
3327 static int sanitize_rc6_option(const struct drm_device *dev, int enable_rc6)
3328 {
3329         /* No RC6 before Ironlake */
3330         if (INTEL_INFO(dev)->gen < 5)
3331                 return 0;
3332
3333         /* RC6 is only on Ironlake mobile not on desktop */
3334         if (INTEL_INFO(dev)->gen == 5 && !IS_IRONLAKE_M(dev))
3335                 return 0;
3336
3337         /* Respect the kernel parameter if it is set */
3338         if (enable_rc6 >= 0) {
3339                 int mask;
3340
3341                 if (INTEL_INFO(dev)->gen == 6 || IS_IVYBRIDGE(dev))
3342                         mask = INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE |
3343                                INTEL_RC6pp_ENABLE;
3344                 else
3345                         mask = INTEL_RC6_ENABLE;
3346
3347                 if ((enable_rc6 & mask) != enable_rc6)
3348                         DRM_INFO("Adjusting RC6 mask to %d (requested %d, valid %d)\n",
3349                                  enable_rc6 & mask, enable_rc6, mask);
3350
3351                 return enable_rc6 & mask;
3352         }
3353
3354         /* Disable RC6 on Ironlake */
3355         if (INTEL_INFO(dev)->gen == 5)
3356                 return 0;
3357
3358         if (IS_IVYBRIDGE(dev))
3359                 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
3360
3361         return INTEL_RC6_ENABLE;
3362 }
3363
3364 int intel_enable_rc6(const struct drm_device *dev)
3365 {
3366         return i915.enable_rc6;
3367 }
3368
3369 static void gen8_enable_rps_interrupts(struct drm_device *dev)
3370 {
3371         struct drm_i915_private *dev_priv = dev->dev_private;
3372
3373         spin_lock_irq(&dev_priv->irq_lock);
3374         WARN_ON(dev_priv->rps.pm_iir);
3375         bdw_enable_pm_irq(dev_priv, dev_priv->pm_rps_events);
3376         I915_WRITE(GEN8_GT_IIR(2), dev_priv->pm_rps_events);
3377         spin_unlock_irq(&dev_priv->irq_lock);
3378 }
3379
3380 static void gen6_enable_rps_interrupts(struct drm_device *dev)
3381 {
3382         struct drm_i915_private *dev_priv = dev->dev_private;
3383
3384         spin_lock_irq(&dev_priv->irq_lock);
3385         WARN_ON(dev_priv->rps.pm_iir);
3386         snb_enable_pm_irq(dev_priv, dev_priv->pm_rps_events);
3387         I915_WRITE(GEN6_PMIIR, dev_priv->pm_rps_events);
3388         spin_unlock_irq(&dev_priv->irq_lock);
3389 }
3390
3391 static void parse_rp_state_cap(struct drm_i915_private *dev_priv, u32 rp_state_cap)
3392 {
3393         /* All of these values are in units of 50MHz */
3394         dev_priv->rps.cur_freq          = 0;
3395         /* static values from HW: RP0 < RPe < RP1 < RPn (min_freq) */
3396         dev_priv->rps.rp1_freq          = (rp_state_cap >>  8) & 0xff;
3397         dev_priv->rps.rp0_freq          = (rp_state_cap >>  0) & 0xff;
3398         dev_priv->rps.min_freq          = (rp_state_cap >> 16) & 0xff;
3399         /* XXX: only BYT has a special efficient freq */
3400         dev_priv->rps.efficient_freq    = dev_priv->rps.rp1_freq;
3401         /* hw_max = RP0 until we check for overclocking */
3402         dev_priv->rps.max_freq          = dev_priv->rps.rp0_freq;
3403
3404         /* Preserve min/max settings in case of re-init */
3405         if (dev_priv->rps.max_freq_softlimit == 0)
3406                 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
3407
3408         if (dev_priv->rps.min_freq_softlimit == 0)
3409                 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
3410 }
3411
3412 static void gen8_enable_rps(struct drm_device *dev)
3413 {
3414         struct drm_i915_private *dev_priv = dev->dev_private;
3415         struct intel_ring_buffer *ring;
3416         uint32_t rc6_mask = 0, rp_state_cap;
3417         int unused;
3418
3419         /* 1a: Software RC state - RC0 */
3420         I915_WRITE(GEN6_RC_STATE, 0);
3421
3422         /* 1c & 1d: Get forcewake during program sequence. Although the driver
3423          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
3424         gen6_gt_force_wake_get(dev_priv, FORCEWAKE_ALL);
3425
3426         /* 2a: Disable RC states. */
3427         I915_WRITE(GEN6_RC_CONTROL, 0);
3428
3429         rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
3430         parse_rp_state_cap(dev_priv, rp_state_cap);
3431
3432         /* 2b: Program RC6 thresholds.*/
3433         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
3434         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
3435         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
3436         for_each_ring(ring, dev_priv, unused)
3437                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
3438         I915_WRITE(GEN6_RC_SLEEP, 0);
3439         I915_WRITE(GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
3440
3441         /* 3: Enable RC6 */
3442         if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
3443                 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
3444         intel_print_rc6_info(dev, rc6_mask);
3445         I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
3446                                     GEN6_RC_CTL_EI_MODE(1) |
3447                                     rc6_mask);
3448
3449         /* 4 Program defaults and thresholds for RPS*/
3450         I915_WRITE(GEN6_RPNSWREQ,
3451                    HSW_FREQUENCY(dev_priv->rps.rp1_freq));
3452         I915_WRITE(GEN6_RC_VIDEO_FREQ,
3453                    HSW_FREQUENCY(dev_priv->rps.rp1_freq));
3454         /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
3455         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
3456
3457         /* Docs recommend 900MHz, and 300 MHz respectively */
3458         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
3459                    dev_priv->rps.max_freq_softlimit << 24 |
3460                    dev_priv->rps.min_freq_softlimit << 16);
3461
3462         I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
3463         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
3464         I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
3465         I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
3466
3467         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
3468
3469         /* 5: Enable RPS */
3470         I915_WRITE(GEN6_RP_CONTROL,
3471                    GEN6_RP_MEDIA_TURBO |
3472                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
3473                    GEN6_RP_MEDIA_IS_GFX |
3474                    GEN6_RP_ENABLE |
3475                    GEN6_RP_UP_BUSY_AVG |
3476                    GEN6_RP_DOWN_IDLE_AVG);
3477
3478         /* 6: Ring frequency + overclocking (our driver does this later */
3479
3480         gen6_set_rps(dev, (I915_READ(GEN6_GT_PERF_STATUS) & 0xff00) >> 8);
3481
3482         gen8_enable_rps_interrupts(dev);
3483
3484         gen6_gt_force_wake_put(dev_priv, FORCEWAKE_ALL);
3485 }
3486
3487 static void gen6_enable_rps(struct drm_device *dev)
3488 {
3489         struct drm_i915_private *dev_priv = dev->dev_private;
3490         struct intel_ring_buffer *ring;
3491         u32 rp_state_cap;
3492         u32 gt_perf_status;
3493         u32 rc6vids, pcu_mbox = 0, rc6_mask = 0;
3494         u32 gtfifodbg;
3495         int rc6_mode;
3496         int i, ret;
3497
3498         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3499
3500         /* Here begins a magic sequence of register writes to enable
3501          * auto-downclocking.
3502          *
3503          * Perhaps there might be some value in exposing these to
3504          * userspace...
3505          */
3506         I915_WRITE(GEN6_RC_STATE, 0);
3507
3508         /* Clear the DBG now so we don't confuse earlier errors */
3509         if ((gtfifodbg = I915_READ(GTFIFODBG))) {
3510                 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
3511                 I915_WRITE(GTFIFODBG, gtfifodbg);
3512         }
3513
3514         gen6_gt_force_wake_get(dev_priv, FORCEWAKE_ALL);
3515
3516         rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
3517         gt_perf_status = I915_READ(GEN6_GT_PERF_STATUS);
3518
3519         parse_rp_state_cap(dev_priv, rp_state_cap);
3520
3521         /* disable the counters and set deterministic thresholds */
3522         I915_WRITE(GEN6_RC_CONTROL, 0);
3523
3524         I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
3525         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
3526         I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
3527         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
3528         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
3529
3530         for_each_ring(ring, dev_priv, i)
3531                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
3532
3533         I915_WRITE(GEN6_RC_SLEEP, 0);
3534         I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
3535         if (IS_IVYBRIDGE(dev))
3536                 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
3537         else
3538                 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
3539         I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
3540         I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
3541
3542         /* Check if we are enabling RC6 */
3543         rc6_mode = intel_enable_rc6(dev_priv->dev);
3544         if (rc6_mode & INTEL_RC6_ENABLE)
3545                 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
3546
3547         /* We don't use those on Haswell */
3548         if (!IS_HASWELL(dev)) {
3549                 if (rc6_mode & INTEL_RC6p_ENABLE)
3550                         rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
3551
3552                 if (rc6_mode & INTEL_RC6pp_ENABLE)
3553                         rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
3554         }
3555
3556         intel_print_rc6_info(dev, rc6_mask);
3557
3558         I915_WRITE(GEN6_RC_CONTROL,
3559                    rc6_mask |
3560                    GEN6_RC_CTL_EI_MODE(1) |
3561                    GEN6_RC_CTL_HW_ENABLE);
3562
3563         /* Power down if completely idle for over 50ms */
3564         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
3565         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
3566
3567         ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_MIN_FREQ_TABLE, 0);
3568         if (ret)
3569                 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
3570
3571         ret = sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &pcu_mbox);
3572         if (!ret && (pcu_mbox & (1<<31))) { /* OC supported */
3573                 DRM_DEBUG_DRIVER("Overclocking supported. Max: %dMHz, Overclock max: %dMHz\n",
3574                                  (dev_priv->rps.max_freq_softlimit & 0xff) * 50,
3575                                  (pcu_mbox & 0xff) * 50);
3576                 dev_priv->rps.max_freq = pcu_mbox & 0xff;
3577         }
3578
3579         dev_priv->rps.power = HIGH_POWER; /* force a reset */
3580         gen6_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
3581
3582         gen6_enable_rps_interrupts(dev);
3583
3584         rc6vids = 0;
3585         ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
3586         if (IS_GEN6(dev) && ret) {
3587                 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
3588         } else if (IS_GEN6(dev) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
3589                 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
3590                           GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
3591                 rc6vids &= 0xffff00;
3592                 rc6vids |= GEN6_ENCODE_RC6_VID(450);
3593                 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
3594                 if (ret)
3595                         DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
3596         }
3597
3598         gen6_gt_force_wake_put(dev_priv, FORCEWAKE_ALL);
3599 }
3600
3601 static void __gen6_update_ring_freq(struct drm_device *dev)
3602 {
3603         struct drm_i915_private *dev_priv = dev->dev_private;
3604         int min_freq = 15;
3605         unsigned int gpu_freq;
3606         unsigned int max_ia_freq, min_ring_freq;
3607         int scaling_factor = 180;
3608         struct cpufreq_policy *policy;
3609
3610         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3611
3612         policy = cpufreq_cpu_get(0);
3613         if (policy) {
3614                 max_ia_freq = policy->cpuinfo.max_freq;
3615                 cpufreq_cpu_put(policy);
3616         } else {
3617                 /*
3618                  * Default to measured freq if none found, PCU will ensure we
3619                  * don't go over
3620                  */
3621                 max_ia_freq = tsc_khz;
3622         }
3623
3624         /* Convert from kHz to MHz */
3625         max_ia_freq /= 1000;
3626
3627         min_ring_freq = I915_READ(DCLK) & 0xf;
3628         /* convert DDR frequency from units of 266.6MHz to bandwidth */
3629         min_ring_freq = mult_frac(min_ring_freq, 8, 3);
3630
3631         /*
3632          * For each potential GPU frequency, load a ring frequency we'd like
3633          * to use for memory access.  We do this by specifying the IA frequency
3634          * the PCU should use as a reference to determine the ring frequency.
3635          */
3636         for (gpu_freq = dev_priv->rps.max_freq_softlimit; gpu_freq >= dev_priv->rps.min_freq_softlimit;
3637              gpu_freq--) {
3638                 int diff = dev_priv->rps.max_freq_softlimit - gpu_freq;
3639                 unsigned int ia_freq = 0, ring_freq = 0;
3640
3641                 if (INTEL_INFO(dev)->gen >= 8) {
3642                         /* max(2 * GT, DDR). NB: GT is 50MHz units */
3643                         ring_freq = max(min_ring_freq, gpu_freq);
3644                 } else if (IS_HASWELL(dev)) {
3645                         ring_freq = mult_frac(gpu_freq, 5, 4);
3646                         ring_freq = max(min_ring_freq, ring_freq);
3647                         /* leave ia_freq as the default, chosen by cpufreq */
3648                 } else {
3649                         /* On older processors, there is no separate ring
3650                          * clock domain, so in order to boost the bandwidth
3651                          * of the ring, we need to upclock the CPU (ia_freq).
3652                          *
3653                          * For GPU frequencies less than 750MHz,
3654                          * just use the lowest ring freq.
3655                          */
3656                         if (gpu_freq < min_freq)
3657                                 ia_freq = 800;
3658                         else
3659                                 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
3660                         ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
3661                 }
3662
3663                 sandybridge_pcode_write(dev_priv,
3664                                         GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
3665                                         ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
3666                                         ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
3667                                         gpu_freq);
3668         }
3669 }
3670
3671 void gen6_update_ring_freq(struct drm_device *dev)
3672 {
3673         struct drm_i915_private *dev_priv = dev->dev_private;
3674
3675         if (INTEL_INFO(dev)->gen < 6 || IS_VALLEYVIEW(dev))
3676                 return;
3677
3678         mutex_lock(&dev_priv->rps.hw_lock);
3679         __gen6_update_ring_freq(dev);
3680         mutex_unlock(&dev_priv->rps.hw_lock);
3681 }
3682
3683 int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
3684 {
3685         u32 val, rp0;
3686
3687         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
3688
3689         rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
3690         /* Clamp to max */
3691         rp0 = min_t(u32, rp0, 0xea);
3692
3693         return rp0;
3694 }
3695
3696 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
3697 {
3698         u32 val, rpe;
3699
3700         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
3701         rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
3702         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
3703         rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
3704
3705         return rpe;
3706 }
3707
3708 int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
3709 {
3710         return vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
3711 }
3712
3713 /* Check that the pctx buffer wasn't move under us. */
3714 static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
3715 {
3716         unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
3717
3718         WARN_ON(pctx_addr != dev_priv->mm.stolen_base +
3719                              dev_priv->vlv_pctx->stolen->start);
3720 }
3721
3722 static void valleyview_setup_pctx(struct drm_device *dev)
3723 {
3724         struct drm_i915_private *dev_priv = dev->dev_private;
3725         struct drm_i915_gem_object *pctx;
3726         unsigned long pctx_paddr;
3727         u32 pcbr;
3728         int pctx_size = 24*1024;
3729
3730         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
3731
3732         pcbr = I915_READ(VLV_PCBR);
3733         if (pcbr) {
3734                 /* BIOS set it up already, grab the pre-alloc'd space */
3735                 int pcbr_offset;
3736
3737                 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
3738                 pctx = i915_gem_object_create_stolen_for_preallocated(dev_priv->dev,
3739                                                                       pcbr_offset,
3740                                                                       I915_GTT_OFFSET_NONE,
3741                                                                       pctx_size);
3742                 goto out;
3743         }
3744
3745         /*
3746          * From the Gunit register HAS:
3747          * The Gfx driver is expected to program this register and ensure
3748          * proper allocation within Gfx stolen memory.  For example, this
3749          * register should be programmed such than the PCBR range does not
3750          * overlap with other ranges, such as the frame buffer, protected
3751          * memory, or any other relevant ranges.
3752          */
3753         pctx = i915_gem_object_create_stolen(dev, pctx_size);
3754         if (!pctx) {
3755                 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
3756                 return;
3757         }
3758
3759         pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
3760         I915_WRITE(VLV_PCBR, pctx_paddr);
3761
3762 out:
3763         dev_priv->vlv_pctx = pctx;
3764 }
3765
3766 static void valleyview_cleanup_pctx(struct drm_device *dev)
3767 {
3768         struct drm_i915_private *dev_priv = dev->dev_private;
3769
3770         if (WARN_ON(!dev_priv->vlv_pctx))
3771                 return;
3772
3773         drm_gem_object_unreference(&dev_priv->vlv_pctx->base);
3774         dev_priv->vlv_pctx = NULL;
3775 }
3776
3777 static void valleyview_init_gt_powersave(struct drm_device *dev)
3778 {
3779         struct drm_i915_private *dev_priv = dev->dev_private;
3780
3781         valleyview_setup_pctx(dev);
3782
3783         mutex_lock(&dev_priv->rps.hw_lock);
3784
3785         dev_priv->rps.max_freq = valleyview_rps_max_freq(dev_priv);
3786         dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
3787         DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
3788                          vlv_gpu_freq(dev_priv, dev_priv->rps.max_freq),
3789                          dev_priv->rps.max_freq);
3790
3791         dev_priv->rps.efficient_freq = valleyview_rps_rpe_freq(dev_priv);
3792         DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
3793                          vlv_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
3794                          dev_priv->rps.efficient_freq);
3795
3796         dev_priv->rps.min_freq = valleyview_rps_min_freq(dev_priv);
3797         DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
3798                          vlv_gpu_freq(dev_priv, dev_priv->rps.min_freq),
3799                          dev_priv->rps.min_freq);
3800
3801         /* Preserve min/max settings in case of re-init */
3802         if (dev_priv->rps.max_freq_softlimit == 0)
3803                 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
3804
3805         if (dev_priv->rps.min_freq_softlimit == 0)
3806                 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
3807
3808         mutex_unlock(&dev_priv->rps.hw_lock);
3809 }
3810
3811 static void valleyview_cleanup_gt_powersave(struct drm_device *dev)
3812 {
3813         valleyview_cleanup_pctx(dev);
3814 }
3815
3816 static void valleyview_enable_rps(struct drm_device *dev)
3817 {
3818         struct drm_i915_private *dev_priv = dev->dev_private;
3819         struct intel_ring_buffer *ring;
3820         u32 gtfifodbg, val, rc6_mode = 0;
3821         int i;
3822
3823         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3824
3825         valleyview_check_pctx(dev_priv);
3826
3827         if ((gtfifodbg = I915_READ(GTFIFODBG))) {
3828                 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
3829                                  gtfifodbg);
3830                 I915_WRITE(GTFIFODBG, gtfifodbg);
3831         }
3832
3833         /* If VLV, Forcewake all wells, else re-direct to regular path */
3834         gen6_gt_force_wake_get(dev_priv, FORCEWAKE_ALL);
3835
3836         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
3837         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
3838         I915_WRITE(GEN6_RP_UP_EI, 66000);
3839         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
3840
3841         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
3842
3843         I915_WRITE(GEN6_RP_CONTROL,
3844                    GEN6_RP_MEDIA_TURBO |
3845                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
3846                    GEN6_RP_MEDIA_IS_GFX |
3847                    GEN6_RP_ENABLE |
3848                    GEN6_RP_UP_BUSY_AVG |
3849                    GEN6_RP_DOWN_IDLE_CONT);
3850
3851         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
3852         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
3853         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
3854
3855         for_each_ring(ring, dev_priv, i)
3856                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
3857
3858         I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
3859
3860         /* allows RC6 residency counter to work */
3861         I915_WRITE(VLV_COUNTER_CONTROL,
3862                    _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
3863                                       VLV_MEDIA_RC6_COUNT_EN |
3864                                       VLV_RENDER_RC6_COUNT_EN));
3865         if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
3866                 rc6_mode = GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
3867
3868         intel_print_rc6_info(dev, rc6_mode);
3869
3870         I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
3871
3872         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
3873
3874         DRM_DEBUG_DRIVER("GPLL enabled? %s\n", val & 0x10 ? "yes" : "no");
3875         DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
3876
3877         dev_priv->rps.cur_freq = (val >> 8) & 0xff;
3878         DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
3879                          vlv_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
3880                          dev_priv->rps.cur_freq);
3881
3882         DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
3883                          vlv_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
3884                          dev_priv->rps.efficient_freq);
3885
3886         valleyview_set_rps(dev_priv->dev, dev_priv->rps.efficient_freq);
3887
3888         gen6_enable_rps_interrupts(dev);
3889
3890         gen6_gt_force_wake_put(dev_priv, FORCEWAKE_ALL);
3891 }
3892
3893 void ironlake_teardown_rc6(struct drm_device *dev)
3894 {
3895         struct drm_i915_private *dev_priv = dev->dev_private;
3896
3897         if (dev_priv->ips.renderctx) {
3898                 i915_gem_object_ggtt_unpin(dev_priv->ips.renderctx);
3899                 drm_gem_object_unreference(&dev_priv->ips.renderctx->base);
3900                 dev_priv->ips.renderctx = NULL;
3901         }
3902
3903         if (dev_priv->ips.pwrctx) {
3904                 i915_gem_object_ggtt_unpin(dev_priv->ips.pwrctx);
3905                 drm_gem_object_unreference(&dev_priv->ips.pwrctx->base);
3906                 dev_priv->ips.pwrctx = NULL;
3907         }
3908 }
3909
3910 static void ironlake_disable_rc6(struct drm_device *dev)
3911 {
3912         struct drm_i915_private *dev_priv = dev->dev_private;
3913
3914         if (I915_READ(PWRCTXA)) {
3915                 /* Wake the GPU, prevent RC6, then restore RSTDBYCTL */
3916                 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) | RCX_SW_EXIT);
3917                 wait_for(((I915_READ(RSTDBYCTL) & RSX_STATUS_MASK) == RSX_STATUS_ON),
3918                          50);
3919
3920                 I915_WRITE(PWRCTXA, 0);
3921                 POSTING_READ(PWRCTXA);
3922
3923                 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
3924                 POSTING_READ(RSTDBYCTL);
3925         }
3926 }
3927
3928 static int ironlake_setup_rc6(struct drm_device *dev)
3929 {
3930         struct drm_i915_private *dev_priv = dev->dev_private;
3931
3932         if (dev_priv->ips.renderctx == NULL)
3933                 dev_priv->ips.renderctx = intel_alloc_context_page(dev);
3934         if (!dev_priv->ips.renderctx)
3935                 return -ENOMEM;
3936
3937         if (dev_priv->ips.pwrctx == NULL)
3938                 dev_priv->ips.pwrctx = intel_alloc_context_page(dev);
3939         if (!dev_priv->ips.pwrctx) {
3940                 ironlake_teardown_rc6(dev);
3941                 return -ENOMEM;
3942         }
3943
3944         return 0;
3945 }
3946
3947 static void ironlake_enable_rc6(struct drm_device *dev)
3948 {
3949         struct drm_i915_private *dev_priv = dev->dev_private;
3950         struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
3951         bool was_interruptible;
3952         int ret;
3953
3954         /* rc6 disabled by default due to repeated reports of hanging during
3955          * boot and resume.
3956          */
3957         if (!intel_enable_rc6(dev))
3958                 return;
3959
3960         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
3961
3962         ret = ironlake_setup_rc6(dev);
3963         if (ret)
3964                 return;
3965
3966         was_interruptible = dev_priv->mm.interruptible;
3967         dev_priv->mm.interruptible = false;
3968
3969         /*
3970          * GPU can automatically power down the render unit if given a page
3971          * to save state.
3972          */
3973         ret = intel_ring_begin(ring, 6);
3974         if (ret) {
3975                 ironlake_teardown_rc6(dev);
3976                 dev_priv->mm.interruptible = was_interruptible;
3977                 return;
3978         }
3979
3980         intel_ring_emit(ring, MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN);
3981         intel_ring_emit(ring, MI_SET_CONTEXT);
3982         intel_ring_emit(ring, i915_gem_obj_ggtt_offset(dev_priv->ips.renderctx) |
3983                         MI_MM_SPACE_GTT |
3984                         MI_SAVE_EXT_STATE_EN |
3985                         MI_RESTORE_EXT_STATE_EN |
3986                         MI_RESTORE_INHIBIT);
3987         intel_ring_emit(ring, MI_SUSPEND_FLUSH);
3988         intel_ring_emit(ring, MI_NOOP);
3989         intel_ring_emit(ring, MI_FLUSH);
3990         intel_ring_advance(ring);
3991
3992         /*
3993          * Wait for the command parser to advance past MI_SET_CONTEXT. The HW
3994          * does an implicit flush, combined with MI_FLUSH above, it should be
3995          * safe to assume that renderctx is valid
3996          */
3997         ret = intel_ring_idle(ring);
3998         dev_priv->mm.interruptible = was_interruptible;
3999         if (ret) {
4000                 DRM_ERROR("failed to enable ironlake power savings\n");
4001                 ironlake_teardown_rc6(dev);
4002                 return;
4003         }
4004
4005         I915_WRITE(PWRCTXA, i915_gem_obj_ggtt_offset(dev_priv->ips.pwrctx) | PWRCTX_EN);
4006         I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
4007
4008         intel_print_rc6_info(dev, GEN6_RC_CTL_RC6_ENABLE);
4009 }
4010
4011 static unsigned long intel_pxfreq(u32 vidfreq)
4012 {
4013         unsigned long freq;
4014         int div = (vidfreq & 0x3f0000) >> 16;
4015         int post = (vidfreq & 0x3000) >> 12;
4016         int pre = (vidfreq & 0x7);
4017
4018         if (!pre)
4019                 return 0;
4020
4021         freq = ((div * 133333) / ((1<<post) * pre));
4022
4023         return freq;
4024 }
4025
4026 static const struct cparams {
4027         u16 i;
4028         u16 t;
4029         u16 m;
4030         u16 c;
4031 } cparams[] = {
4032         { 1, 1333, 301, 28664 },
4033         { 1, 1066, 294, 24460 },
4034         { 1, 800, 294, 25192 },
4035         { 0, 1333, 276, 27605 },
4036         { 0, 1066, 276, 27605 },
4037         { 0, 800, 231, 23784 },
4038 };
4039
4040 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
4041 {
4042         u64 total_count, diff, ret;
4043         u32 count1, count2, count3, m = 0, c = 0;
4044         unsigned long now = jiffies_to_msecs(jiffies), diff1;
4045         int i;
4046
4047         assert_spin_locked(&mchdev_lock);
4048
4049         diff1 = now - dev_priv->ips.last_time1;
4050
4051         /* Prevent division-by-zero if we are asking too fast.
4052          * Also, we don't get interesting results if we are polling
4053          * faster than once in 10ms, so just return the saved value
4054          * in such cases.
4055          */
4056         if (diff1 <= 10)
4057                 return dev_priv->ips.chipset_power;
4058
4059         count1 = I915_READ(DMIEC);
4060         count2 = I915_READ(DDREC);
4061         count3 = I915_READ(CSIEC);
4062
4063         total_count = count1 + count2 + count3;
4064
4065         /* FIXME: handle per-counter overflow */
4066         if (total_count < dev_priv->ips.last_count1) {
4067                 diff = ~0UL - dev_priv->ips.last_count1;
4068                 diff += total_count;
4069         } else {
4070                 diff = total_count - dev_priv->ips.last_count1;
4071         }
4072
4073         for (i = 0; i < ARRAY_SIZE(cparams); i++) {
4074                 if (cparams[i].i == dev_priv->ips.c_m &&
4075                     cparams[i].t == dev_priv->ips.r_t) {
4076                         m = cparams[i].m;
4077                         c = cparams[i].c;
4078                         break;
4079                 }
4080         }
4081
4082         diff = div_u64(diff, diff1);
4083         ret = ((m * diff) + c);
4084         ret = div_u64(ret, 10);
4085
4086         dev_priv->ips.last_count1 = total_count;
4087         dev_priv->ips.last_time1 = now;
4088
4089         dev_priv->ips.chipset_power = ret;
4090
4091         return ret;
4092 }
4093
4094 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
4095 {
4096         struct drm_device *dev = dev_priv->dev;
4097         unsigned long val;
4098
4099         if (INTEL_INFO(dev)->gen != 5)
4100                 return 0;
4101
4102         spin_lock_irq(&mchdev_lock);
4103
4104         val = __i915_chipset_val(dev_priv);
4105
4106         spin_unlock_irq(&mchdev_lock);
4107
4108         return val;
4109 }
4110
4111 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
4112 {
4113         unsigned long m, x, b;
4114         u32 tsfs;
4115
4116         tsfs = I915_READ(TSFS);
4117
4118         m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
4119         x = I915_READ8(TR1);
4120
4121         b = tsfs & TSFS_INTR_MASK;
4122
4123         return ((m * x) / 127) - b;
4124 }
4125
4126 static u16 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
4127 {
4128         struct drm_device *dev = dev_priv->dev;
4129         static const struct v_table {
4130                 u16 vd; /* in .1 mil */
4131                 u16 vm; /* in .1 mil */
4132         } v_table[] = {
4133                 { 0, 0, },
4134                 { 375, 0, },
4135                 { 500, 0, },
4136                 { 625, 0, },
4137                 { 750, 0, },
4138                 { 875, 0, },
4139                 { 1000, 0, },
4140                 { 1125, 0, },
4141                 { 4125, 3000, },
4142                 { 4125, 3000, },
4143                 { 4125, 3000, },
4144                 { 4125, 3000, },
4145                 { 4125, 3000, },
4146                 { 4125, 3000, },
4147                 { 4125, 3000, },
4148                 { 4125, 3000, },
4149                 { 4125, 3000, },
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                 { 4250, 3125, },
4166                 { 4375, 3250, },
4167                 { 4500, 3375, },
4168                 { 4625, 3500, },
4169                 { 4750, 3625, },
4170                 { 4875, 3750, },
4171                 { 5000, 3875, },
4172                 { 5125, 4000, },
4173                 { 5250, 4125, },
4174                 { 5375, 4250, },
4175                 { 5500, 4375, },
4176                 { 5625, 4500, },
4177                 { 5750, 4625, },
4178                 { 5875, 4750, },
4179                 { 6000, 4875, },
4180                 { 6125, 5000, },
4181                 { 6250, 5125, },
4182                 { 6375, 5250, },
4183                 { 6500, 5375, },
4184                 { 6625, 5500, },
4185                 { 6750, 5625, },
4186                 { 6875, 5750, },
4187                 { 7000, 5875, },
4188                 { 7125, 6000, },
4189                 { 7250, 6125, },
4190                 { 7375, 6250, },
4191                 { 7500, 6375, },
4192                 { 7625, 6500, },
4193                 { 7750, 6625, },
4194                 { 7875, 6750, },
4195                 { 8000, 6875, },
4196                 { 8125, 7000, },
4197                 { 8250, 7125, },
4198                 { 8375, 7250, },
4199                 { 8500, 7375, },
4200                 { 8625, 7500, },
4201                 { 8750, 7625, },
4202                 { 8875, 7750, },
4203                 { 9000, 7875, },
4204                 { 9125, 8000, },
4205                 { 9250, 8125, },
4206                 { 9375, 8250, },
4207                 { 9500, 8375, },
4208                 { 9625, 8500, },
4209                 { 9750, 8625, },
4210                 { 9875, 8750, },
4211                 { 10000, 8875, },
4212                 { 10125, 9000, },
4213                 { 10250, 9125, },
4214                 { 10375, 9250, },
4215                 { 10500, 9375, },
4216                 { 10625, 9500, },
4217                 { 10750, 9625, },
4218                 { 10875, 9750, },
4219                 { 11000, 9875, },
4220                 { 11125, 10000, },
4221                 { 11250, 10125, },
4222                 { 11375, 10250, },
4223                 { 11500, 10375, },
4224                 { 11625, 10500, },
4225                 { 11750, 10625, },
4226                 { 11875, 10750, },
4227                 { 12000, 10875, },
4228                 { 12125, 11000, },
4229                 { 12250, 11125, },
4230                 { 12375, 11250, },
4231                 { 12500, 11375, },
4232                 { 12625, 11500, },
4233                 { 12750, 11625, },
4234                 { 12875, 11750, },
4235                 { 13000, 11875, },
4236                 { 13125, 12000, },
4237                 { 13250, 12125, },
4238                 { 13375, 12250, },
4239                 { 13500, 12375, },
4240                 { 13625, 12500, },
4241                 { 13750, 12625, },
4242                 { 13875, 12750, },
4243                 { 14000, 12875, },
4244                 { 14125, 13000, },
4245                 { 14250, 13125, },
4246                 { 14375, 13250, },
4247                 { 14500, 13375, },
4248                 { 14625, 13500, },
4249                 { 14750, 13625, },
4250                 { 14875, 13750, },
4251                 { 15000, 13875, },
4252                 { 15125, 14000, },
4253                 { 15250, 14125, },
4254                 { 15375, 14250, },
4255                 { 15500, 14375, },
4256                 { 15625, 14500, },
4257                 { 15750, 14625, },
4258                 { 15875, 14750, },
4259                 { 16000, 14875, },
4260                 { 16125, 15000, },
4261         };
4262         if (INTEL_INFO(dev)->is_mobile)
4263                 return v_table[pxvid].vm;
4264         else
4265                 return v_table[pxvid].vd;
4266 }
4267
4268 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
4269 {
4270         struct timespec now, diff1;
4271         u64 diff;
4272         unsigned long diffms;
4273         u32 count;
4274
4275         assert_spin_locked(&mchdev_lock);
4276
4277         getrawmonotonic(&now);
4278         diff1 = timespec_sub(now, dev_priv->ips.last_time2);
4279
4280         /* Don't divide by 0 */
4281         diffms = diff1.tv_sec * 1000 + diff1.tv_nsec / 1000000;
4282         if (!diffms)
4283                 return;
4284
4285         count = I915_READ(GFXEC);
4286
4287         if (count < dev_priv->ips.last_count2) {
4288                 diff = ~0UL - dev_priv->ips.last_count2;
4289                 diff += count;
4290         } else {
4291                 diff = count - dev_priv->ips.last_count2;
4292         }
4293
4294         dev_priv->ips.last_count2 = count;
4295         dev_priv->ips.last_time2 = now;
4296
4297         /* More magic constants... */
4298         diff = diff * 1181;
4299         diff = div_u64(diff, diffms * 10);
4300         dev_priv->ips.gfx_power = diff;
4301 }
4302
4303 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
4304 {
4305         struct drm_device *dev = dev_priv->dev;
4306
4307         if (INTEL_INFO(dev)->gen != 5)
4308                 return;
4309
4310         spin_lock_irq(&mchdev_lock);
4311
4312         __i915_update_gfx_val(dev_priv);
4313
4314         spin_unlock_irq(&mchdev_lock);
4315 }
4316
4317 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
4318 {
4319         unsigned long t, corr, state1, corr2, state2;
4320         u32 pxvid, ext_v;
4321
4322         assert_spin_locked(&mchdev_lock);
4323
4324         pxvid = I915_READ(PXVFREQ_BASE + (dev_priv->rps.cur_freq * 4));
4325         pxvid = (pxvid >> 24) & 0x7f;
4326         ext_v = pvid_to_extvid(dev_priv, pxvid);
4327
4328         state1 = ext_v;
4329
4330         t = i915_mch_val(dev_priv);
4331
4332         /* Revel in the empirically derived constants */
4333
4334         /* Correction factor in 1/100000 units */
4335         if (t > 80)
4336                 corr = ((t * 2349) + 135940);
4337         else if (t >= 50)
4338                 corr = ((t * 964) + 29317);
4339         else /* < 50 */
4340                 corr = ((t * 301) + 1004);
4341
4342         corr = corr * ((150142 * state1) / 10000 - 78642);
4343         corr /= 100000;
4344         corr2 = (corr * dev_priv->ips.corr);
4345
4346         state2 = (corr2 * state1) / 10000;
4347         state2 /= 100; /* convert to mW */
4348
4349         __i915_update_gfx_val(dev_priv);
4350
4351         return dev_priv->ips.gfx_power + state2;
4352 }
4353
4354 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
4355 {
4356         struct drm_device *dev = dev_priv->dev;
4357         unsigned long val;
4358
4359         if (INTEL_INFO(dev)->gen != 5)
4360                 return 0;
4361
4362         spin_lock_irq(&mchdev_lock);
4363
4364         val = __i915_gfx_val(dev_priv);
4365
4366         spin_unlock_irq(&mchdev_lock);
4367
4368         return val;
4369 }
4370
4371 /**
4372  * i915_read_mch_val - return value for IPS use
4373  *
4374  * Calculate and return a value for the IPS driver to use when deciding whether
4375  * we have thermal and power headroom to increase CPU or GPU power budget.
4376  */
4377 unsigned long i915_read_mch_val(void)
4378 {
4379         struct drm_i915_private *dev_priv;
4380         unsigned long chipset_val, graphics_val, ret = 0;
4381
4382         spin_lock_irq(&mchdev_lock);
4383         if (!i915_mch_dev)
4384                 goto out_unlock;
4385         dev_priv = i915_mch_dev;
4386
4387         chipset_val = __i915_chipset_val(dev_priv);
4388         graphics_val = __i915_gfx_val(dev_priv);
4389
4390         ret = chipset_val + graphics_val;
4391
4392 out_unlock:
4393         spin_unlock_irq(&mchdev_lock);
4394
4395         return ret;
4396 }
4397 EXPORT_SYMBOL_GPL(i915_read_mch_val);
4398
4399 /**
4400  * i915_gpu_raise - raise GPU frequency limit
4401  *
4402  * Raise the limit; IPS indicates we have thermal headroom.
4403  */
4404 bool i915_gpu_raise(void)
4405 {
4406         struct drm_i915_private *dev_priv;
4407         bool ret = true;
4408
4409         spin_lock_irq(&mchdev_lock);
4410         if (!i915_mch_dev) {
4411                 ret = false;
4412                 goto out_unlock;
4413         }
4414         dev_priv = i915_mch_dev;
4415
4416         if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
4417                 dev_priv->ips.max_delay--;
4418
4419 out_unlock:
4420         spin_unlock_irq(&mchdev_lock);
4421
4422         return ret;
4423 }
4424 EXPORT_SYMBOL_GPL(i915_gpu_raise);
4425
4426 /**
4427  * i915_gpu_lower - lower GPU frequency limit
4428  *
4429  * IPS indicates we're close to a thermal limit, so throttle back the GPU
4430  * frequency maximum.
4431  */
4432 bool i915_gpu_lower(void)
4433 {
4434         struct drm_i915_private *dev_priv;
4435         bool ret = true;
4436
4437         spin_lock_irq(&mchdev_lock);
4438         if (!i915_mch_dev) {
4439                 ret = false;
4440                 goto out_unlock;
4441         }
4442         dev_priv = i915_mch_dev;
4443
4444         if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
4445                 dev_priv->ips.max_delay++;
4446
4447 out_unlock:
4448         spin_unlock_irq(&mchdev_lock);
4449
4450         return ret;
4451 }
4452 EXPORT_SYMBOL_GPL(i915_gpu_lower);
4453
4454 /**
4455  * i915_gpu_busy - indicate GPU business to IPS
4456  *
4457  * Tell the IPS driver whether or not the GPU is busy.
4458  */
4459 bool i915_gpu_busy(void)
4460 {
4461         struct drm_i915_private *dev_priv;
4462         struct intel_ring_buffer *ring;
4463         bool ret = false;
4464         int i;
4465
4466         spin_lock_irq(&mchdev_lock);
4467         if (!i915_mch_dev)
4468                 goto out_unlock;
4469         dev_priv = i915_mch_dev;
4470
4471         for_each_ring(ring, dev_priv, i)
4472                 ret |= !list_empty(&ring->request_list);
4473
4474 out_unlock:
4475         spin_unlock_irq(&mchdev_lock);
4476
4477         return ret;
4478 }
4479 EXPORT_SYMBOL_GPL(i915_gpu_busy);
4480
4481 /**
4482  * i915_gpu_turbo_disable - disable graphics turbo
4483  *
4484  * Disable graphics turbo by resetting the max frequency and setting the
4485  * current frequency to the default.
4486  */
4487 bool i915_gpu_turbo_disable(void)
4488 {
4489         struct drm_i915_private *dev_priv;
4490         bool ret = true;
4491
4492         spin_lock_irq(&mchdev_lock);
4493         if (!i915_mch_dev) {
4494                 ret = false;
4495                 goto out_unlock;
4496         }
4497         dev_priv = i915_mch_dev;
4498
4499         dev_priv->ips.max_delay = dev_priv->ips.fstart;
4500
4501         if (!ironlake_set_drps(dev_priv->dev, dev_priv->ips.fstart))
4502                 ret = false;
4503
4504 out_unlock:
4505         spin_unlock_irq(&mchdev_lock);
4506
4507         return ret;
4508 }
4509 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
4510
4511 /**
4512  * Tells the intel_ips driver that the i915 driver is now loaded, if
4513  * IPS got loaded first.
4514  *
4515  * This awkward dance is so that neither module has to depend on the
4516  * other in order for IPS to do the appropriate communication of
4517  * GPU turbo limits to i915.
4518  */
4519 static void
4520 ips_ping_for_i915_load(void)
4521 {
4522         void (*link)(void);
4523
4524         link = symbol_get(ips_link_to_i915_driver);
4525         if (link) {
4526                 link();
4527                 symbol_put(ips_link_to_i915_driver);
4528         }
4529 }
4530
4531 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
4532 {
4533         /* We only register the i915 ips part with intel-ips once everything is
4534          * set up, to avoid intel-ips sneaking in and reading bogus values. */
4535         spin_lock_irq(&mchdev_lock);
4536         i915_mch_dev = dev_priv;
4537         spin_unlock_irq(&mchdev_lock);
4538
4539         ips_ping_for_i915_load();
4540 }
4541
4542 void intel_gpu_ips_teardown(void)
4543 {
4544         spin_lock_irq(&mchdev_lock);
4545         i915_mch_dev = NULL;
4546         spin_unlock_irq(&mchdev_lock);
4547 }
4548
4549 static void intel_init_emon(struct drm_device *dev)
4550 {
4551         struct drm_i915_private *dev_priv = dev->dev_private;
4552         u32 lcfuse;
4553         u8 pxw[16];
4554         int i;
4555
4556         /* Disable to program */
4557         I915_WRITE(ECR, 0);
4558         POSTING_READ(ECR);
4559
4560         /* Program energy weights for various events */
4561         I915_WRITE(SDEW, 0x15040d00);
4562         I915_WRITE(CSIEW0, 0x007f0000);
4563         I915_WRITE(CSIEW1, 0x1e220004);
4564         I915_WRITE(CSIEW2, 0x04000004);
4565
4566         for (i = 0; i < 5; i++)
4567                 I915_WRITE(PEW + (i * 4), 0);
4568         for (i = 0; i < 3; i++)
4569                 I915_WRITE(DEW + (i * 4), 0);
4570
4571         /* Program P-state weights to account for frequency power adjustment */
4572         for (i = 0; i < 16; i++) {
4573                 u32 pxvidfreq = I915_READ(PXVFREQ_BASE + (i * 4));
4574                 unsigned long freq = intel_pxfreq(pxvidfreq);
4575                 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
4576                         PXVFREQ_PX_SHIFT;
4577                 unsigned long val;
4578
4579                 val = vid * vid;
4580                 val *= (freq / 1000);
4581                 val *= 255;
4582                 val /= (127*127*900);
4583                 if (val > 0xff)
4584                         DRM_ERROR("bad pxval: %ld\n", val);
4585                 pxw[i] = val;
4586         }
4587         /* Render standby states get 0 weight */
4588         pxw[14] = 0;
4589         pxw[15] = 0;
4590
4591         for (i = 0; i < 4; i++) {
4592                 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
4593                         (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
4594                 I915_WRITE(PXW + (i * 4), val);
4595         }
4596
4597         /* Adjust magic regs to magic values (more experimental results) */
4598         I915_WRITE(OGW0, 0);
4599         I915_WRITE(OGW1, 0);
4600         I915_WRITE(EG0, 0x00007f00);
4601         I915_WRITE(EG1, 0x0000000e);
4602         I915_WRITE(EG2, 0x000e0000);
4603         I915_WRITE(EG3, 0x68000300);
4604         I915_WRITE(EG4, 0x42000000);
4605         I915_WRITE(EG5, 0x00140031);
4606         I915_WRITE(EG6, 0);
4607         I915_WRITE(EG7, 0);
4608
4609         for (i = 0; i < 8; i++)
4610                 I915_WRITE(PXWL + (i * 4), 0);
4611
4612         /* Enable PMON + select events */
4613         I915_WRITE(ECR, 0x80000019);
4614
4615         lcfuse = I915_READ(LCFUSE02);
4616
4617         dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
4618 }
4619
4620 void intel_init_gt_powersave(struct drm_device *dev)
4621 {
4622         i915.enable_rc6 = sanitize_rc6_option(dev, i915.enable_rc6);
4623
4624         if (IS_VALLEYVIEW(dev))
4625                 valleyview_init_gt_powersave(dev);
4626 }
4627
4628 void intel_cleanup_gt_powersave(struct drm_device *dev)
4629 {
4630         if (IS_VALLEYVIEW(dev))
4631                 valleyview_cleanup_gt_powersave(dev);
4632 }
4633
4634 void intel_disable_gt_powersave(struct drm_device *dev)
4635 {
4636         struct drm_i915_private *dev_priv = dev->dev_private;
4637
4638         /* Interrupts should be disabled already to avoid re-arming. */
4639         WARN_ON(dev->irq_enabled);
4640
4641         if (IS_IRONLAKE_M(dev)) {
4642                 ironlake_disable_drps(dev);
4643                 ironlake_disable_rc6(dev);
4644         } else if (IS_GEN6(dev) || IS_GEN7(dev) || IS_BROADWELL(dev)) {
4645                 cancel_delayed_work_sync(&dev_priv->rps.delayed_resume_work);
4646                 cancel_work_sync(&dev_priv->rps.work);
4647                 mutex_lock(&dev_priv->rps.hw_lock);
4648                 if (IS_VALLEYVIEW(dev))
4649                         valleyview_disable_rps(dev);
4650                 else
4651                         gen6_disable_rps(dev);
4652                 dev_priv->rps.enabled = false;
4653                 mutex_unlock(&dev_priv->rps.hw_lock);
4654         }
4655 }
4656
4657 static void intel_gen6_powersave_work(struct work_struct *work)
4658 {
4659         struct drm_i915_private *dev_priv =
4660                 container_of(work, struct drm_i915_private,
4661                              rps.delayed_resume_work.work);
4662         struct drm_device *dev = dev_priv->dev;
4663
4664         mutex_lock(&dev_priv->rps.hw_lock);
4665
4666         if (IS_VALLEYVIEW(dev)) {
4667                 valleyview_enable_rps(dev);
4668         } else if (IS_BROADWELL(dev)) {
4669                 gen8_enable_rps(dev);
4670                 __gen6_update_ring_freq(dev);
4671         } else {
4672                 gen6_enable_rps(dev);
4673                 __gen6_update_ring_freq(dev);
4674         }
4675         dev_priv->rps.enabled = true;
4676         mutex_unlock(&dev_priv->rps.hw_lock);
4677
4678         intel_runtime_pm_put(dev_priv);
4679 }
4680
4681 void intel_enable_gt_powersave(struct drm_device *dev)
4682 {
4683         struct drm_i915_private *dev_priv = dev->dev_private;
4684
4685         if (IS_IRONLAKE_M(dev)) {
4686                 mutex_lock(&dev->struct_mutex);
4687                 ironlake_enable_drps(dev);
4688                 ironlake_enable_rc6(dev);
4689                 intel_init_emon(dev);
4690                 mutex_unlock(&dev->struct_mutex);
4691         } else if (IS_GEN6(dev) || IS_GEN7(dev) || IS_BROADWELL(dev)) {
4692                 /*
4693                  * PCU communication is slow and this doesn't need to be
4694                  * done at any specific time, so do this out of our fast path
4695                  * to make resume and init faster.
4696                  *
4697                  * We depend on the HW RC6 power context save/restore
4698                  * mechanism when entering D3 through runtime PM suspend. So
4699                  * disable RPM until RPS/RC6 is properly setup. We can only
4700                  * get here via the driver load/system resume/runtime resume
4701                  * paths, so the _noresume version is enough (and in case of
4702                  * runtime resume it's necessary).
4703                  */
4704                 if (schedule_delayed_work(&dev_priv->rps.delayed_resume_work,
4705                                            round_jiffies_up_relative(HZ)))
4706                         intel_runtime_pm_get_noresume(dev_priv);
4707         }
4708 }
4709
4710 void intel_reset_gt_powersave(struct drm_device *dev)
4711 {
4712         struct drm_i915_private *dev_priv = dev->dev_private;
4713
4714         dev_priv->rps.enabled = false;
4715         intel_enable_gt_powersave(dev);
4716 }
4717
4718 static void ibx_init_clock_gating(struct drm_device *dev)
4719 {
4720         struct drm_i915_private *dev_priv = dev->dev_private;
4721
4722         /*
4723          * On Ibex Peak and Cougar Point, we need to disable clock
4724          * gating for the panel power sequencer or it will fail to
4725          * start up when no ports are active.
4726          */
4727         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
4728 }
4729
4730 static void g4x_disable_trickle_feed(struct drm_device *dev)
4731 {
4732         struct drm_i915_private *dev_priv = dev->dev_private;
4733         int pipe;
4734
4735         for_each_pipe(pipe) {
4736                 I915_WRITE(DSPCNTR(pipe),
4737                            I915_READ(DSPCNTR(pipe)) |
4738                            DISPPLANE_TRICKLE_FEED_DISABLE);
4739                 intel_flush_primary_plane(dev_priv, pipe);
4740         }
4741 }
4742
4743 static void ilk_init_lp_watermarks(struct drm_device *dev)
4744 {
4745         struct drm_i915_private *dev_priv = dev->dev_private;
4746
4747         I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
4748         I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
4749         I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
4750
4751         /*
4752          * Don't touch WM1S_LP_EN here.
4753          * Doing so could cause underruns.
4754          */
4755 }
4756
4757 static void ironlake_init_clock_gating(struct drm_device *dev)
4758 {
4759         struct drm_i915_private *dev_priv = dev->dev_private;
4760         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
4761
4762         /*
4763          * Required for FBC
4764          * WaFbcDisableDpfcClockGating:ilk
4765          */
4766         dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
4767                    ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
4768                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
4769
4770         I915_WRITE(PCH_3DCGDIS0,
4771                    MARIUNIT_CLOCK_GATE_DISABLE |
4772                    SVSMUNIT_CLOCK_GATE_DISABLE);
4773         I915_WRITE(PCH_3DCGDIS1,
4774                    VFMUNIT_CLOCK_GATE_DISABLE);
4775
4776         /*
4777          * According to the spec the following bits should be set in
4778          * order to enable memory self-refresh
4779          * The bit 22/21 of 0x42004
4780          * The bit 5 of 0x42020
4781          * The bit 15 of 0x45000
4782          */
4783         I915_WRITE(ILK_DISPLAY_CHICKEN2,
4784                    (I915_READ(ILK_DISPLAY_CHICKEN2) |
4785                     ILK_DPARB_GATE | ILK_VSDPFD_FULL));
4786         dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
4787         I915_WRITE(DISP_ARB_CTL,
4788                    (I915_READ(DISP_ARB_CTL) |
4789                     DISP_FBC_WM_DIS));
4790
4791         ilk_init_lp_watermarks(dev);
4792
4793         /*
4794          * Based on the document from hardware guys the following bits
4795          * should be set unconditionally in order to enable FBC.
4796          * The bit 22 of 0x42000
4797          * The bit 22 of 0x42004
4798          * The bit 7,8,9 of 0x42020.
4799          */
4800         if (IS_IRONLAKE_M(dev)) {
4801                 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
4802                 I915_WRITE(ILK_DISPLAY_CHICKEN1,
4803                            I915_READ(ILK_DISPLAY_CHICKEN1) |
4804                            ILK_FBCQ_DIS);
4805                 I915_WRITE(ILK_DISPLAY_CHICKEN2,
4806                            I915_READ(ILK_DISPLAY_CHICKEN2) |
4807                            ILK_DPARB_GATE);
4808         }
4809
4810         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
4811
4812         I915_WRITE(ILK_DISPLAY_CHICKEN2,
4813                    I915_READ(ILK_DISPLAY_CHICKEN2) |
4814                    ILK_ELPIN_409_SELECT);
4815         I915_WRITE(_3D_CHICKEN2,
4816                    _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
4817                    _3D_CHICKEN2_WM_READ_PIPELINED);
4818
4819         /* WaDisableRenderCachePipelinedFlush:ilk */
4820         I915_WRITE(CACHE_MODE_0,
4821                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
4822
4823         /* WaDisable_RenderCache_OperationalFlush:ilk */
4824         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
4825
4826         g4x_disable_trickle_feed(dev);
4827
4828         ibx_init_clock_gating(dev);
4829 }
4830
4831 static void cpt_init_clock_gating(struct drm_device *dev)
4832 {
4833         struct drm_i915_private *dev_priv = dev->dev_private;
4834         int pipe;
4835         uint32_t val;
4836
4837         /*
4838          * On Ibex Peak and Cougar Point, we need to disable clock
4839          * gating for the panel power sequencer or it will fail to
4840          * start up when no ports are active.
4841          */
4842         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
4843                    PCH_DPLUNIT_CLOCK_GATE_DISABLE |
4844                    PCH_CPUNIT_CLOCK_GATE_DISABLE);
4845         I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
4846                    DPLS_EDP_PPS_FIX_DIS);
4847         /* The below fixes the weird display corruption, a few pixels shifted
4848          * downward, on (only) LVDS of some HP laptops with IVY.
4849          */
4850         for_each_pipe(pipe) {
4851                 val = I915_READ(TRANS_CHICKEN2(pipe));
4852                 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
4853                 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
4854                 if (dev_priv->vbt.fdi_rx_polarity_inverted)
4855                         val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
4856                 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
4857                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
4858                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
4859                 I915_WRITE(TRANS_CHICKEN2(pipe), val);
4860         }
4861         /* WADP0ClockGatingDisable */
4862         for_each_pipe(pipe) {
4863                 I915_WRITE(TRANS_CHICKEN1(pipe),
4864                            TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
4865         }
4866 }
4867
4868 static void gen6_check_mch_setup(struct drm_device *dev)
4869 {
4870         struct drm_i915_private *dev_priv = dev->dev_private;
4871         uint32_t tmp;
4872
4873         tmp = I915_READ(MCH_SSKPD);
4874         if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL) {
4875                 DRM_INFO("Wrong MCH_SSKPD value: 0x%08x\n", tmp);
4876                 DRM_INFO("This can cause pipe underruns and display issues.\n");
4877                 DRM_INFO("Please upgrade your BIOS to fix this.\n");
4878         }
4879 }
4880
4881 static void gen6_init_clock_gating(struct drm_device *dev)
4882 {
4883         struct drm_i915_private *dev_priv = dev->dev_private;
4884         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
4885
4886         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
4887
4888         I915_WRITE(ILK_DISPLAY_CHICKEN2,
4889                    I915_READ(ILK_DISPLAY_CHICKEN2) |
4890                    ILK_ELPIN_409_SELECT);
4891
4892         /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
4893         I915_WRITE(_3D_CHICKEN,
4894                    _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
4895
4896         /* WaSetupGtModeTdRowDispatch:snb */
4897         if (IS_SNB_GT1(dev))
4898                 I915_WRITE(GEN6_GT_MODE,
4899                            _MASKED_BIT_ENABLE(GEN6_TD_FOUR_ROW_DISPATCH_DISABLE));
4900
4901         /* WaDisable_RenderCache_OperationalFlush:snb */
4902         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
4903
4904         /*
4905          * BSpec recoomends 8x4 when MSAA is used,
4906          * however in practice 16x4 seems fastest.
4907          *
4908          * Note that PS/WM thread counts depend on the WIZ hashing
4909          * disable bit, which we don't touch here, but it's good
4910          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
4911          */
4912         I915_WRITE(GEN6_GT_MODE,
4913                    GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
4914
4915         ilk_init_lp_watermarks(dev);
4916
4917         I915_WRITE(CACHE_MODE_0,
4918                    _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
4919
4920         I915_WRITE(GEN6_UCGCTL1,
4921                    I915_READ(GEN6_UCGCTL1) |
4922                    GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
4923                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
4924
4925         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
4926          * gating disable must be set.  Failure to set it results in
4927          * flickering pixels due to Z write ordering failures after
4928          * some amount of runtime in the Mesa "fire" demo, and Unigine
4929          * Sanctuary and Tropics, and apparently anything else with
4930          * alpha test or pixel discard.
4931          *
4932          * According to the spec, bit 11 (RCCUNIT) must also be set,
4933          * but we didn't debug actual testcases to find it out.
4934          *
4935          * WaDisableRCCUnitClockGating:snb
4936          * WaDisableRCPBUnitClockGating:snb
4937          */
4938         I915_WRITE(GEN6_UCGCTL2,
4939                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
4940                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
4941
4942         /* WaStripsFansDisableFastClipPerformanceFix:snb */
4943         I915_WRITE(_3D_CHICKEN3,
4944                    _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
4945
4946         /*
4947          * Bspec says:
4948          * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
4949          * 3DSTATE_SF number of SF output attributes is more than 16."
4950          */
4951         I915_WRITE(_3D_CHICKEN3,
4952                    _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
4953
4954         /*
4955          * According to the spec the following bits should be
4956          * set in order to enable memory self-refresh and fbc:
4957          * The bit21 and bit22 of 0x42000
4958          * The bit21 and bit22 of 0x42004
4959          * The bit5 and bit7 of 0x42020
4960          * The bit14 of 0x70180
4961          * The bit14 of 0x71180
4962          *
4963          * WaFbcAsynchFlipDisableFbcQueue:snb
4964          */
4965         I915_WRITE(ILK_DISPLAY_CHICKEN1,
4966                    I915_READ(ILK_DISPLAY_CHICKEN1) |
4967                    ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
4968         I915_WRITE(ILK_DISPLAY_CHICKEN2,
4969                    I915_READ(ILK_DISPLAY_CHICKEN2) |
4970                    ILK_DPARB_GATE | ILK_VSDPFD_FULL);
4971         I915_WRITE(ILK_DSPCLK_GATE_D,
4972                    I915_READ(ILK_DSPCLK_GATE_D) |
4973                    ILK_DPARBUNIT_CLOCK_GATE_ENABLE  |
4974                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
4975
4976         g4x_disable_trickle_feed(dev);
4977
4978         cpt_init_clock_gating(dev);
4979
4980         gen6_check_mch_setup(dev);
4981 }
4982
4983 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
4984 {
4985         uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
4986
4987         /*
4988          * WaVSThreadDispatchOverride:ivb,vlv
4989          *
4990          * This actually overrides the dispatch
4991          * mode for all thread types.
4992          */
4993         reg &= ~GEN7_FF_SCHED_MASK;
4994         reg |= GEN7_FF_TS_SCHED_HW;
4995         reg |= GEN7_FF_VS_SCHED_HW;
4996         reg |= GEN7_FF_DS_SCHED_HW;
4997
4998         I915_WRITE(GEN7_FF_THREAD_MODE, reg);
4999 }
5000
5001 static void lpt_init_clock_gating(struct drm_device *dev)
5002 {
5003         struct drm_i915_private *dev_priv = dev->dev_private;
5004
5005         /*
5006          * TODO: this bit should only be enabled when really needed, then
5007          * disabled when not needed anymore in order to save power.
5008          */
5009         if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE)
5010                 I915_WRITE(SOUTH_DSPCLK_GATE_D,
5011                            I915_READ(SOUTH_DSPCLK_GATE_D) |
5012                            PCH_LP_PARTITION_LEVEL_DISABLE);
5013
5014         /* WADPOClockGatingDisable:hsw */
5015         I915_WRITE(_TRANSA_CHICKEN1,
5016                    I915_READ(_TRANSA_CHICKEN1) |
5017                    TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
5018 }
5019
5020 static void lpt_suspend_hw(struct drm_device *dev)
5021 {
5022         struct drm_i915_private *dev_priv = dev->dev_private;
5023
5024         if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE) {
5025                 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
5026
5027                 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
5028                 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
5029         }
5030 }
5031
5032 static void gen8_init_clock_gating(struct drm_device *dev)
5033 {
5034         struct drm_i915_private *dev_priv = dev->dev_private;
5035         enum pipe pipe;
5036
5037         I915_WRITE(WM3_LP_ILK, 0);
5038         I915_WRITE(WM2_LP_ILK, 0);
5039         I915_WRITE(WM1_LP_ILK, 0);
5040
5041         /* FIXME(BDW): Check all the w/a, some might only apply to
5042          * pre-production hw. */
5043
5044         /* WaDisablePartialInstShootdown:bdw */
5045         I915_WRITE(GEN8_ROW_CHICKEN,
5046                    _MASKED_BIT_ENABLE(PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE));
5047
5048         /* WaDisableThreadStallDopClockGating:bdw */
5049         /* FIXME: Unclear whether we really need this on production bdw. */
5050         I915_WRITE(GEN8_ROW_CHICKEN,
5051                    _MASKED_BIT_ENABLE(STALL_DOP_GATING_DISABLE));
5052
5053         /*
5054          * This GEN8_CENTROID_PIXEL_OPT_DIS W/A is only needed for
5055          * pre-production hardware
5056          */
5057         I915_WRITE(HALF_SLICE_CHICKEN3,
5058                    _MASKED_BIT_ENABLE(GEN8_CENTROID_PIXEL_OPT_DIS));
5059         I915_WRITE(HALF_SLICE_CHICKEN3,
5060                    _MASKED_BIT_ENABLE(GEN8_SAMPLER_POWER_BYPASS_DIS));
5061         I915_WRITE(GAMTARBMODE, _MASKED_BIT_ENABLE(ARB_MODE_BWGTLB_DISABLE));
5062
5063         I915_WRITE(_3D_CHICKEN3,
5064                    _3D_CHICKEN_SDE_LIMIT_FIFO_POLY_DEPTH(2));
5065
5066         I915_WRITE(COMMON_SLICE_CHICKEN2,
5067                    _MASKED_BIT_ENABLE(GEN8_CSC2_SBE_VUE_CACHE_CONSERVATIVE));
5068
5069         I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
5070                    _MASKED_BIT_ENABLE(GEN7_SINGLE_SUBSCAN_DISPATCH_ENABLE));
5071
5072         /* WaDisableDopClockGating:bdw May not be needed for production */
5073         I915_WRITE(GEN7_ROW_CHICKEN2,
5074                    _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5075
5076         /* WaSwitchSolVfFArbitrationPriority:bdw */
5077         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
5078
5079         /* WaPsrDPAMaskVBlankInSRD:bdw */
5080         I915_WRITE(CHICKEN_PAR1_1,
5081                    I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
5082
5083         /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
5084         for_each_pipe(pipe) {
5085                 I915_WRITE(CHICKEN_PIPESL_1(pipe),
5086                            I915_READ(CHICKEN_PIPESL_1(pipe)) |
5087                            BDW_DPRS_MASK_VBLANK_SRD);
5088         }
5089
5090         /* Use Force Non-Coherent whenever executing a 3D context. This is a
5091          * workaround for for a possible hang in the unlikely event a TLB
5092          * invalidation occurs during a PSD flush.
5093          */
5094         I915_WRITE(HDC_CHICKEN0,
5095                    I915_READ(HDC_CHICKEN0) |
5096                    _MASKED_BIT_ENABLE(HDC_FORCE_NON_COHERENT));
5097
5098         /* WaVSRefCountFullforceMissDisable:bdw */
5099         /* WaDSRefCountFullforceMissDisable:bdw */
5100         I915_WRITE(GEN7_FF_THREAD_MODE,
5101                    I915_READ(GEN7_FF_THREAD_MODE) &
5102                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
5103
5104         /*
5105          * BSpec recommends 8x4 when MSAA is used,
5106          * however in practice 16x4 seems fastest.
5107          *
5108          * Note that PS/WM thread counts depend on the WIZ hashing
5109          * disable bit, which we don't touch here, but it's good
5110          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
5111          */
5112         I915_WRITE(GEN7_GT_MODE,
5113                    GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
5114
5115         I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
5116                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
5117
5118         /* WaDisableSDEUnitClockGating:bdw */
5119         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
5120                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
5121
5122         /* Wa4x4STCOptimizationDisable:bdw */
5123         I915_WRITE(CACHE_MODE_1,
5124                    _MASKED_BIT_ENABLE(GEN8_4x4_STC_OPTIMIZATION_DISABLE));
5125 }
5126
5127 static void haswell_init_clock_gating(struct drm_device *dev)
5128 {
5129         struct drm_i915_private *dev_priv = dev->dev_private;
5130
5131         ilk_init_lp_watermarks(dev);
5132
5133         /* L3 caching of data atomics doesn't work -- disable it. */
5134         I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
5135         I915_WRITE(HSW_ROW_CHICKEN3,
5136                    _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
5137
5138         /* This is required by WaCatErrorRejectionIssue:hsw */
5139         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5140                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5141                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5142
5143         /* WaVSRefCountFullforceMissDisable:hsw */
5144         I915_WRITE(GEN7_FF_THREAD_MODE,
5145                    I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
5146
5147         /* WaDisable_RenderCache_OperationalFlush:hsw */
5148         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5149
5150         /* enable HiZ Raw Stall Optimization */
5151         I915_WRITE(CACHE_MODE_0_GEN7,
5152                    _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
5153
5154         /* WaDisable4x2SubspanOptimization:hsw */
5155         I915_WRITE(CACHE_MODE_1,
5156                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5157
5158         /*
5159          * BSpec recommends 8x4 when MSAA is used,
5160          * however in practice 16x4 seems fastest.
5161          *
5162          * Note that PS/WM thread counts depend on the WIZ hashing
5163          * disable bit, which we don't touch here, but it's good
5164          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
5165          */
5166         I915_WRITE(GEN7_GT_MODE,
5167                    GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
5168
5169         /* WaSwitchSolVfFArbitrationPriority:hsw */
5170         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
5171
5172         /* WaRsPkgCStateDisplayPMReq:hsw */
5173         I915_WRITE(CHICKEN_PAR1_1,
5174                    I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
5175
5176         lpt_init_clock_gating(dev);
5177 }
5178
5179 static void ivybridge_init_clock_gating(struct drm_device *dev)
5180 {
5181         struct drm_i915_private *dev_priv = dev->dev_private;
5182         uint32_t snpcr;
5183
5184         ilk_init_lp_watermarks(dev);
5185
5186         I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
5187
5188         /* WaDisableEarlyCull:ivb */
5189         I915_WRITE(_3D_CHICKEN3,
5190                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
5191
5192         /* WaDisableBackToBackFlipFix:ivb */
5193         I915_WRITE(IVB_CHICKEN3,
5194                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
5195                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
5196
5197         /* WaDisablePSDDualDispatchEnable:ivb */
5198         if (IS_IVB_GT1(dev))
5199                 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
5200                            _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
5201
5202         /* WaDisable_RenderCache_OperationalFlush:ivb */
5203         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5204
5205         /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
5206         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
5207                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
5208
5209         /* WaApplyL3ControlAndL3ChickenMode:ivb */
5210         I915_WRITE(GEN7_L3CNTLREG1,
5211                         GEN7_WA_FOR_GEN7_L3_CONTROL);
5212         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
5213                    GEN7_WA_L3_CHICKEN_MODE);
5214         if (IS_IVB_GT1(dev))
5215                 I915_WRITE(GEN7_ROW_CHICKEN2,
5216                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5217         else {
5218                 /* must write both registers */
5219                 I915_WRITE(GEN7_ROW_CHICKEN2,
5220                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5221                 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
5222                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5223         }
5224
5225         /* WaForceL3Serialization:ivb */
5226         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
5227                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
5228
5229         /*
5230          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
5231          * This implements the WaDisableRCZUnitClockGating:ivb workaround.
5232          */
5233         I915_WRITE(GEN6_UCGCTL2,
5234                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
5235
5236         /* This is required by WaCatErrorRejectionIssue:ivb */
5237         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5238                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5239                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5240
5241         g4x_disable_trickle_feed(dev);
5242
5243         gen7_setup_fixed_func_scheduler(dev_priv);
5244
5245         if (0) { /* causes HiZ corruption on ivb:gt1 */
5246                 /* enable HiZ Raw Stall Optimization */
5247                 I915_WRITE(CACHE_MODE_0_GEN7,
5248                            _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
5249         }
5250
5251         /* WaDisable4x2SubspanOptimization:ivb */
5252         I915_WRITE(CACHE_MODE_1,
5253                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5254
5255         /*
5256          * BSpec recommends 8x4 when MSAA is used,
5257          * however in practice 16x4 seems fastest.
5258          *
5259          * Note that PS/WM thread counts depend on the WIZ hashing
5260          * disable bit, which we don't touch here, but it's good
5261          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
5262          */
5263         I915_WRITE(GEN7_GT_MODE,
5264                    GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
5265
5266         snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
5267         snpcr &= ~GEN6_MBC_SNPCR_MASK;
5268         snpcr |= GEN6_MBC_SNPCR_MED;
5269         I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
5270
5271         if (!HAS_PCH_NOP(dev))
5272                 cpt_init_clock_gating(dev);
5273
5274         gen6_check_mch_setup(dev);
5275 }
5276
5277 static void valleyview_init_clock_gating(struct drm_device *dev)
5278 {
5279         struct drm_i915_private *dev_priv = dev->dev_private;
5280         u32 val;
5281
5282         mutex_lock(&dev_priv->rps.hw_lock);
5283         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5284         mutex_unlock(&dev_priv->rps.hw_lock);
5285         switch ((val >> 6) & 3) {
5286         case 0:
5287         case 1:
5288                 dev_priv->mem_freq = 800;
5289                 break;
5290         case 2:
5291                 dev_priv->mem_freq = 1066;
5292                 break;
5293         case 3:
5294                 dev_priv->mem_freq = 1333;
5295                 break;
5296         }
5297         DRM_DEBUG_DRIVER("DDR speed: %d MHz", dev_priv->mem_freq);
5298
5299         dev_priv->vlv_cdclk_freq = valleyview_cur_cdclk(dev_priv);
5300         DRM_DEBUG_DRIVER("Current CD clock rate: %d MHz",
5301                          dev_priv->vlv_cdclk_freq);
5302
5303         I915_WRITE(DSPCLK_GATE_D, VRHUNIT_CLOCK_GATE_DISABLE);
5304
5305         /* WaDisableEarlyCull:vlv */
5306         I915_WRITE(_3D_CHICKEN3,
5307                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
5308
5309         /* WaDisableBackToBackFlipFix:vlv */
5310         I915_WRITE(IVB_CHICKEN3,
5311                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
5312                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
5313
5314         /* WaPsdDispatchEnable:vlv */
5315         /* WaDisablePSDDualDispatchEnable:vlv */
5316         I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
5317                    _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
5318                                       GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
5319
5320         /* WaDisable_RenderCache_OperationalFlush:vlv */
5321         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5322
5323         /* WaForceL3Serialization:vlv */
5324         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
5325                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
5326
5327         /* WaDisableDopClockGating:vlv */
5328         I915_WRITE(GEN7_ROW_CHICKEN2,
5329                    _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
5330
5331         /* This is required by WaCatErrorRejectionIssue:vlv */
5332         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
5333                    I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
5334                    GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
5335
5336         gen7_setup_fixed_func_scheduler(dev_priv);
5337
5338         /*
5339          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
5340          * This implements the WaDisableRCZUnitClockGating:vlv workaround.
5341          */
5342         I915_WRITE(GEN6_UCGCTL2,
5343                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
5344
5345         /* WaDisableL3Bank2xClockGate:vlv */
5346         I915_WRITE(GEN7_UCGCTL4, GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
5347
5348         I915_WRITE(MI_ARB_VLV, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE);
5349
5350         /*
5351          * BSpec says this must be set, even though
5352          * WaDisable4x2SubspanOptimization isn't listed for VLV.
5353          */
5354         I915_WRITE(CACHE_MODE_1,
5355                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
5356
5357         /*
5358          * WaIncreaseL3CreditsForVLVB0:vlv
5359          * This is the hardware default actually.
5360          */
5361         I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
5362
5363         /*
5364          * WaDisableVLVClockGating_VBIIssue:vlv
5365          * Disable clock gating on th GCFG unit to prevent a delay
5366          * in the reporting of vblank events.
5367          */
5368         I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
5369 }
5370
5371 static void cherryview_init_clock_gating(struct drm_device *dev)
5372 {
5373         struct drm_i915_private *dev_priv = dev->dev_private;
5374
5375         I915_WRITE(DSPCLK_GATE_D, VRHUNIT_CLOCK_GATE_DISABLE);
5376
5377         I915_WRITE(MI_ARB_VLV, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE);
5378
5379         /* WaDisablePartialInstShootdown:chv */
5380         I915_WRITE(GEN8_ROW_CHICKEN,
5381                    _MASKED_BIT_ENABLE(PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE));
5382
5383         /* WaDisableThreadStallDopClockGating:chv */
5384         I915_WRITE(GEN8_ROW_CHICKEN,
5385                    _MASKED_BIT_ENABLE(STALL_DOP_GATING_DISABLE));
5386
5387         /* WaVSRefCountFullforceMissDisable:chv */
5388         /* WaDSRefCountFullforceMissDisable:chv */
5389         I915_WRITE(GEN7_FF_THREAD_MODE,
5390                    I915_READ(GEN7_FF_THREAD_MODE) &
5391                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
5392
5393         /* WaDisableSemaphoreAndSyncFlipWait:chv */
5394         I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
5395                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
5396
5397         /* WaDisableCSUnitClockGating:chv */
5398         I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
5399                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
5400
5401         /* WaDisableSDEUnitClockGating:chv */
5402         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
5403                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
5404
5405         /* WaDisableSamplerPowerBypass:chv (pre-production hw) */
5406         I915_WRITE(HALF_SLICE_CHICKEN3,
5407                    _MASKED_BIT_ENABLE(GEN8_SAMPLER_POWER_BYPASS_DIS));
5408 }
5409
5410 static void g4x_init_clock_gating(struct drm_device *dev)
5411 {
5412         struct drm_i915_private *dev_priv = dev->dev_private;
5413         uint32_t dspclk_gate;
5414
5415         I915_WRITE(RENCLK_GATE_D1, 0);
5416         I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
5417                    GS_UNIT_CLOCK_GATE_DISABLE |
5418                    CL_UNIT_CLOCK_GATE_DISABLE);
5419         I915_WRITE(RAMCLK_GATE_D, 0);
5420         dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
5421                 OVRUNIT_CLOCK_GATE_DISABLE |
5422                 OVCUNIT_CLOCK_GATE_DISABLE;
5423         if (IS_GM45(dev))
5424                 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
5425         I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
5426
5427         /* WaDisableRenderCachePipelinedFlush */
5428         I915_WRITE(CACHE_MODE_0,
5429                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
5430
5431         /* WaDisable_RenderCache_OperationalFlush:g4x */
5432         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5433
5434         g4x_disable_trickle_feed(dev);
5435 }
5436
5437 static void crestline_init_clock_gating(struct drm_device *dev)
5438 {
5439         struct drm_i915_private *dev_priv = dev->dev_private;
5440
5441         I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
5442         I915_WRITE(RENCLK_GATE_D2, 0);
5443         I915_WRITE(DSPCLK_GATE_D, 0);
5444         I915_WRITE(RAMCLK_GATE_D, 0);
5445         I915_WRITE16(DEUC, 0);
5446         I915_WRITE(MI_ARB_STATE,
5447                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
5448
5449         /* WaDisable_RenderCache_OperationalFlush:gen4 */
5450         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5451 }
5452
5453 static void broadwater_init_clock_gating(struct drm_device *dev)
5454 {
5455         struct drm_i915_private *dev_priv = dev->dev_private;
5456
5457         I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
5458                    I965_RCC_CLOCK_GATE_DISABLE |
5459                    I965_RCPB_CLOCK_GATE_DISABLE |
5460                    I965_ISC_CLOCK_GATE_DISABLE |
5461                    I965_FBC_CLOCK_GATE_DISABLE);
5462         I915_WRITE(RENCLK_GATE_D2, 0);
5463         I915_WRITE(MI_ARB_STATE,
5464                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
5465
5466         /* WaDisable_RenderCache_OperationalFlush:gen4 */
5467         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5468 }
5469
5470 static void gen3_init_clock_gating(struct drm_device *dev)
5471 {
5472         struct drm_i915_private *dev_priv = dev->dev_private;
5473         u32 dstate = I915_READ(D_STATE);
5474
5475         dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
5476                 DSTATE_DOT_CLOCK_GATING;
5477         I915_WRITE(D_STATE, dstate);
5478
5479         if (IS_PINEVIEW(dev))
5480                 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
5481
5482         /* IIR "flip pending" means done if this bit is set */
5483         I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
5484 }
5485
5486 static void i85x_init_clock_gating(struct drm_device *dev)
5487 {
5488         struct drm_i915_private *dev_priv = dev->dev_private;
5489
5490         I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
5491 }
5492
5493 static void i830_init_clock_gating(struct drm_device *dev)
5494 {
5495         struct drm_i915_private *dev_priv = dev->dev_private;
5496
5497         I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
5498 }
5499
5500 void intel_init_clock_gating(struct drm_device *dev)
5501 {
5502         struct drm_i915_private *dev_priv = dev->dev_private;
5503
5504         dev_priv->display.init_clock_gating(dev);
5505 }
5506
5507 void intel_suspend_hw(struct drm_device *dev)
5508 {
5509         if (HAS_PCH_LPT(dev))
5510                 lpt_suspend_hw(dev);
5511 }
5512
5513 #define for_each_power_well(i, power_well, domain_mask, power_domains)  \
5514         for (i = 0;                                                     \
5515              i < (power_domains)->power_well_count &&                   \
5516                  ((power_well) = &(power_domains)->power_wells[i]);     \
5517              i++)                                                       \
5518                 if ((power_well)->domains & (domain_mask))
5519
5520 #define for_each_power_well_rev(i, power_well, domain_mask, power_domains) \
5521         for (i = (power_domains)->power_well_count - 1;                  \
5522              i >= 0 && ((power_well) = &(power_domains)->power_wells[i]);\
5523              i--)                                                        \
5524                 if ((power_well)->domains & (domain_mask))
5525
5526 /**
5527  * We should only use the power well if we explicitly asked the hardware to
5528  * enable it, so check if it's enabled and also check if we've requested it to
5529  * be enabled.
5530  */
5531 static bool hsw_power_well_enabled(struct drm_i915_private *dev_priv,
5532                                    struct i915_power_well *power_well)
5533 {
5534         return I915_READ(HSW_PWR_WELL_DRIVER) ==
5535                      (HSW_PWR_WELL_ENABLE_REQUEST | HSW_PWR_WELL_STATE_ENABLED);
5536 }
5537
5538 bool intel_display_power_enabled_sw(struct drm_i915_private *dev_priv,
5539                                     enum intel_display_power_domain domain)
5540 {
5541         struct i915_power_domains *power_domains;
5542
5543         power_domains = &dev_priv->power_domains;
5544
5545         return power_domains->domain_use_count[domain];
5546 }
5547
5548 bool intel_display_power_enabled(struct drm_i915_private *dev_priv,
5549                                  enum intel_display_power_domain domain)
5550 {
5551         struct i915_power_domains *power_domains;
5552         struct i915_power_well *power_well;
5553         bool is_enabled;
5554         int i;
5555
5556         if (dev_priv->pm.suspended)
5557                 return false;
5558
5559         power_domains = &dev_priv->power_domains;
5560
5561         is_enabled = true;
5562
5563         mutex_lock(&power_domains->lock);
5564         for_each_power_well_rev(i, power_well, BIT(domain), power_domains) {
5565                 if (power_well->always_on)
5566                         continue;
5567
5568                 if (!power_well->ops->is_enabled(dev_priv, power_well)) {
5569                         is_enabled = false;
5570                         break;
5571                 }
5572         }
5573         mutex_unlock(&power_domains->lock);
5574
5575         return is_enabled;
5576 }
5577
5578 /*
5579  * Starting with Haswell, we have a "Power Down Well" that can be turned off
5580  * when not needed anymore. We have 4 registers that can request the power well
5581  * to be enabled, and it will only be disabled if none of the registers is
5582  * requesting it to be enabled.
5583  */
5584 static void hsw_power_well_post_enable(struct drm_i915_private *dev_priv)
5585 {
5586         struct drm_device *dev = dev_priv->dev;
5587         unsigned long irqflags;
5588
5589         /*
5590          * After we re-enable the power well, if we touch VGA register 0x3d5
5591          * we'll get unclaimed register interrupts. This stops after we write
5592          * anything to the VGA MSR register. The vgacon module uses this
5593          * register all the time, so if we unbind our driver and, as a
5594          * consequence, bind vgacon, we'll get stuck in an infinite loop at
5595          * console_unlock(). So make here we touch the VGA MSR register, making
5596          * sure vgacon can keep working normally without triggering interrupts
5597          * and error messages.
5598          */
5599         vga_get_uninterruptible(dev->pdev, VGA_RSRC_LEGACY_IO);
5600         outb(inb(VGA_MSR_READ), VGA_MSR_WRITE);
5601         vga_put(dev->pdev, VGA_RSRC_LEGACY_IO);
5602
5603         if (IS_BROADWELL(dev)) {
5604                 spin_lock_irqsave(&dev_priv->irq_lock, irqflags);
5605                 I915_WRITE(GEN8_DE_PIPE_IMR(PIPE_B),
5606                            dev_priv->de_irq_mask[PIPE_B]);
5607                 I915_WRITE(GEN8_DE_PIPE_IER(PIPE_B),
5608                            ~dev_priv->de_irq_mask[PIPE_B] |
5609                            GEN8_PIPE_VBLANK);
5610                 I915_WRITE(GEN8_DE_PIPE_IMR(PIPE_C),
5611                            dev_priv->de_irq_mask[PIPE_C]);
5612                 I915_WRITE(GEN8_DE_PIPE_IER(PIPE_C),
5613                            ~dev_priv->de_irq_mask[PIPE_C] |
5614                            GEN8_PIPE_VBLANK);
5615                 POSTING_READ(GEN8_DE_PIPE_IER(PIPE_C));
5616                 spin_unlock_irqrestore(&dev_priv->irq_lock, irqflags);
5617         }
5618 }
5619
5620 static void reset_vblank_counter(struct drm_device *dev, enum pipe pipe)
5621 {
5622         assert_spin_locked(&dev->vbl_lock);
5623
5624         dev->vblank[pipe].last = 0;
5625 }
5626
5627 static void hsw_power_well_post_disable(struct drm_i915_private *dev_priv)
5628 {
5629         struct drm_device *dev = dev_priv->dev;
5630         enum pipe pipe;
5631         unsigned long irqflags;
5632
5633         /*
5634          * After this, the registers on the pipes that are part of the power
5635          * well will become zero, so we have to adjust our counters according to
5636          * that.
5637          *
5638          * FIXME: Should we do this in general in drm_vblank_post_modeset?
5639          */
5640         spin_lock_irqsave(&dev->vbl_lock, irqflags);
5641         for_each_pipe(pipe)
5642                 if (pipe != PIPE_A)
5643                         reset_vblank_counter(dev, pipe);
5644         spin_unlock_irqrestore(&dev->vbl_lock, irqflags);
5645 }
5646
5647 static void hsw_set_power_well(struct drm_i915_private *dev_priv,
5648                                struct i915_power_well *power_well, bool enable)
5649 {
5650         bool is_enabled, enable_requested;
5651         uint32_t tmp;
5652
5653         tmp = I915_READ(HSW_PWR_WELL_DRIVER);
5654         is_enabled = tmp & HSW_PWR_WELL_STATE_ENABLED;
5655         enable_requested = tmp & HSW_PWR_WELL_ENABLE_REQUEST;
5656
5657         if (enable) {
5658                 if (!enable_requested)
5659                         I915_WRITE(HSW_PWR_WELL_DRIVER,
5660                                    HSW_PWR_WELL_ENABLE_REQUEST);
5661
5662                 if (!is_enabled) {
5663                         DRM_DEBUG_KMS("Enabling power well\n");
5664                         if (wait_for((I915_READ(HSW_PWR_WELL_DRIVER) &
5665                                       HSW_PWR_WELL_STATE_ENABLED), 20))
5666                                 DRM_ERROR("Timeout enabling power well\n");
5667                 }
5668
5669                 hsw_power_well_post_enable(dev_priv);
5670         } else {
5671                 if (enable_requested) {
5672                         I915_WRITE(HSW_PWR_WELL_DRIVER, 0);
5673                         POSTING_READ(HSW_PWR_WELL_DRIVER);
5674                         DRM_DEBUG_KMS("Requesting to disable the power well\n");
5675
5676                         hsw_power_well_post_disable(dev_priv);
5677                 }
5678         }
5679 }
5680
5681 static void hsw_power_well_sync_hw(struct drm_i915_private *dev_priv,
5682                                    struct i915_power_well *power_well)
5683 {
5684         hsw_set_power_well(dev_priv, power_well, power_well->count > 0);
5685
5686         /*
5687          * We're taking over the BIOS, so clear any requests made by it since
5688          * the driver is in charge now.
5689          */
5690         if (I915_READ(HSW_PWR_WELL_BIOS) & HSW_PWR_WELL_ENABLE_REQUEST)
5691                 I915_WRITE(HSW_PWR_WELL_BIOS, 0);
5692 }
5693
5694 static void hsw_power_well_enable(struct drm_i915_private *dev_priv,
5695                                   struct i915_power_well *power_well)
5696 {
5697         hsw_set_power_well(dev_priv, power_well, true);
5698 }
5699
5700 static void hsw_power_well_disable(struct drm_i915_private *dev_priv,
5701                                    struct i915_power_well *power_well)
5702 {
5703         hsw_set_power_well(dev_priv, power_well, false);
5704 }
5705
5706 static void i9xx_always_on_power_well_noop(struct drm_i915_private *dev_priv,
5707                                            struct i915_power_well *power_well)
5708 {
5709 }
5710
5711 static bool i9xx_always_on_power_well_enabled(struct drm_i915_private *dev_priv,
5712                                              struct i915_power_well *power_well)
5713 {
5714         return true;
5715 }
5716
5717 static void vlv_set_power_well(struct drm_i915_private *dev_priv,
5718                                struct i915_power_well *power_well, bool enable)
5719 {
5720         enum punit_power_well power_well_id = power_well->data;
5721         u32 mask;
5722         u32 state;
5723         u32 ctrl;
5724
5725         mask = PUNIT_PWRGT_MASK(power_well_id);
5726         state = enable ? PUNIT_PWRGT_PWR_ON(power_well_id) :
5727                          PUNIT_PWRGT_PWR_GATE(power_well_id);
5728
5729         mutex_lock(&dev_priv->rps.hw_lock);
5730
5731 #define COND \
5732         ((vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_STATUS) & mask) == state)
5733
5734         if (COND)
5735                 goto out;
5736
5737         ctrl = vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_CTRL);
5738         ctrl &= ~mask;
5739         ctrl |= state;
5740         vlv_punit_write(dev_priv, PUNIT_REG_PWRGT_CTRL, ctrl);
5741
5742         if (wait_for(COND, 100))
5743                 DRM_ERROR("timout setting power well state %08x (%08x)\n",
5744                           state,
5745                           vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_CTRL));
5746
5747 #undef COND
5748
5749 out:
5750         mutex_unlock(&dev_priv->rps.hw_lock);
5751 }
5752
5753 static void vlv_power_well_sync_hw(struct drm_i915_private *dev_priv,
5754                                    struct i915_power_well *power_well)
5755 {
5756         vlv_set_power_well(dev_priv, power_well, power_well->count > 0);
5757 }
5758
5759 static void vlv_power_well_enable(struct drm_i915_private *dev_priv,
5760                                   struct i915_power_well *power_well)
5761 {
5762         vlv_set_power_well(dev_priv, power_well, true);
5763 }
5764
5765 static void vlv_power_well_disable(struct drm_i915_private *dev_priv,
5766                                    struct i915_power_well *power_well)
5767 {
5768         vlv_set_power_well(dev_priv, power_well, false);
5769 }
5770
5771 static bool vlv_power_well_enabled(struct drm_i915_private *dev_priv,
5772                                    struct i915_power_well *power_well)
5773 {
5774         int power_well_id = power_well->data;
5775         bool enabled = false;
5776         u32 mask;
5777         u32 state;
5778         u32 ctrl;
5779
5780         mask = PUNIT_PWRGT_MASK(power_well_id);
5781         ctrl = PUNIT_PWRGT_PWR_ON(power_well_id);
5782
5783         mutex_lock(&dev_priv->rps.hw_lock);
5784
5785         state = vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_STATUS) & mask;
5786         /*
5787          * We only ever set the power-on and power-gate states, anything
5788          * else is unexpected.
5789          */
5790         WARN_ON(state != PUNIT_PWRGT_PWR_ON(power_well_id) &&
5791                 state != PUNIT_PWRGT_PWR_GATE(power_well_id));
5792         if (state == ctrl)
5793                 enabled = true;
5794
5795         /*
5796          * A transient state at this point would mean some unexpected party
5797          * is poking at the power controls too.
5798          */
5799         ctrl = vlv_punit_read(dev_priv, PUNIT_REG_PWRGT_CTRL) & mask;
5800         WARN_ON(ctrl != state);
5801
5802         mutex_unlock(&dev_priv->rps.hw_lock);
5803
5804         return enabled;
5805 }
5806
5807 static void vlv_display_power_well_enable(struct drm_i915_private *dev_priv,
5808                                           struct i915_power_well *power_well)
5809 {
5810         WARN_ON_ONCE(power_well->data != PUNIT_POWER_WELL_DISP2D);
5811
5812         vlv_set_power_well(dev_priv, power_well, true);
5813
5814         spin_lock_irq(&dev_priv->irq_lock);
5815         valleyview_enable_display_irqs(dev_priv);
5816         spin_unlock_irq(&dev_priv->irq_lock);
5817
5818         /*
5819          * During driver initialization/resume we can avoid restoring the
5820          * part of the HW/SW state that will be inited anyway explicitly.
5821          */
5822         if (dev_priv->power_domains.initializing)
5823                 return;
5824
5825         intel_hpd_init(dev_priv->dev);
5826
5827         i915_redisable_vga_power_on(dev_priv->dev);
5828 }
5829
5830 static void vlv_display_power_well_disable(struct drm_i915_private *dev_priv,
5831                                            struct i915_power_well *power_well)
5832 {
5833         struct drm_device *dev = dev_priv->dev;
5834         enum pipe pipe;
5835
5836         WARN_ON_ONCE(power_well->data != PUNIT_POWER_WELL_DISP2D);
5837
5838         spin_lock_irq(&dev_priv->irq_lock);
5839         for_each_pipe(pipe)
5840                 __intel_set_cpu_fifo_underrun_reporting(dev, pipe, false);
5841
5842         valleyview_disable_display_irqs(dev_priv);
5843         spin_unlock_irq(&dev_priv->irq_lock);
5844
5845         spin_lock_irq(&dev->vbl_lock);
5846         for_each_pipe(pipe)
5847                 reset_vblank_counter(dev, pipe);
5848         spin_unlock_irq(&dev->vbl_lock);
5849
5850         vlv_set_power_well(dev_priv, power_well, false);
5851 }
5852
5853 static void check_power_well_state(struct drm_i915_private *dev_priv,
5854                                    struct i915_power_well *power_well)
5855 {
5856         bool enabled = power_well->ops->is_enabled(dev_priv, power_well);
5857
5858         if (power_well->always_on || !i915.disable_power_well) {
5859                 if (!enabled)
5860                         goto mismatch;
5861
5862                 return;
5863         }
5864
5865         if (enabled != (power_well->count > 0))
5866                 goto mismatch;
5867
5868         return;
5869
5870 mismatch:
5871         WARN(1, "state mismatch for '%s' (always_on %d hw state %d use-count %d disable_power_well %d\n",
5872                   power_well->name, power_well->always_on, enabled,
5873                   power_well->count, i915.disable_power_well);
5874 }
5875
5876 void intel_display_power_get(struct drm_i915_private *dev_priv,
5877                              enum intel_display_power_domain domain)
5878 {
5879         struct i915_power_domains *power_domains;
5880         struct i915_power_well *power_well;
5881         int i;
5882
5883         intel_runtime_pm_get(dev_priv);
5884
5885         power_domains = &dev_priv->power_domains;
5886
5887         mutex_lock(&power_domains->lock);
5888
5889         for_each_power_well(i, power_well, BIT(domain), power_domains) {
5890                 if (!power_well->count++) {
5891                         DRM_DEBUG_KMS("enabling %s\n", power_well->name);
5892                         power_well->ops->enable(dev_priv, power_well);
5893                 }
5894
5895                 check_power_well_state(dev_priv, power_well);
5896         }
5897
5898         power_domains->domain_use_count[domain]++;
5899
5900         mutex_unlock(&power_domains->lock);
5901 }
5902
5903 void intel_display_power_put(struct drm_i915_private *dev_priv,
5904                              enum intel_display_power_domain domain)
5905 {
5906         struct i915_power_domains *power_domains;
5907         struct i915_power_well *power_well;
5908         int i;
5909
5910         power_domains = &dev_priv->power_domains;
5911
5912         mutex_lock(&power_domains->lock);
5913
5914         WARN_ON(!power_domains->domain_use_count[domain]);
5915         power_domains->domain_use_count[domain]--;
5916
5917         for_each_power_well_rev(i, power_well, BIT(domain), power_domains) {
5918                 WARN_ON(!power_well->count);
5919
5920                 if (!--power_well->count && i915.disable_power_well) {
5921                         DRM_DEBUG_KMS("disabling %s\n", power_well->name);
5922                         power_well->ops->disable(dev_priv, power_well);
5923                 }
5924
5925                 check_power_well_state(dev_priv, power_well);
5926         }
5927
5928         mutex_unlock(&power_domains->lock);
5929
5930         intel_runtime_pm_put(dev_priv);
5931 }
5932
5933 static struct i915_power_domains *hsw_pwr;
5934
5935 /* Display audio driver power well request */
5936 void i915_request_power_well(void)
5937 {
5938         struct drm_i915_private *dev_priv;
5939
5940         if (WARN_ON(!hsw_pwr))
5941                 return;
5942
5943         dev_priv = container_of(hsw_pwr, struct drm_i915_private,
5944                                 power_domains);
5945         intel_display_power_get(dev_priv, POWER_DOMAIN_AUDIO);
5946 }
5947 EXPORT_SYMBOL_GPL(i915_request_power_well);
5948
5949 /* Display audio driver power well release */
5950 void i915_release_power_well(void)
5951 {
5952         struct drm_i915_private *dev_priv;
5953
5954         if (WARN_ON(!hsw_pwr))
5955                 return;
5956
5957         dev_priv = container_of(hsw_pwr, struct drm_i915_private,
5958                                 power_domains);
5959         intel_display_power_put(dev_priv, POWER_DOMAIN_AUDIO);
5960 }
5961 EXPORT_SYMBOL_GPL(i915_release_power_well);
5962
5963 #define POWER_DOMAIN_MASK (BIT(POWER_DOMAIN_NUM) - 1)
5964
5965 #define HSW_ALWAYS_ON_POWER_DOMAINS (                   \
5966         BIT(POWER_DOMAIN_PIPE_A) |                      \
5967         BIT(POWER_DOMAIN_TRANSCODER_EDP) |              \
5968         BIT(POWER_DOMAIN_PORT_DDI_A_2_LANES) |          \
5969         BIT(POWER_DOMAIN_PORT_DDI_A_4_LANES) |          \
5970         BIT(POWER_DOMAIN_PORT_DDI_B_2_LANES) |          \
5971         BIT(POWER_DOMAIN_PORT_DDI_B_4_LANES) |          \
5972         BIT(POWER_DOMAIN_PORT_DDI_C_2_LANES) |          \
5973         BIT(POWER_DOMAIN_PORT_DDI_C_4_LANES) |          \
5974         BIT(POWER_DOMAIN_PORT_DDI_D_2_LANES) |          \
5975         BIT(POWER_DOMAIN_PORT_DDI_D_4_LANES) |          \
5976         BIT(POWER_DOMAIN_PORT_CRT) |                    \
5977         BIT(POWER_DOMAIN_INIT))
5978 #define HSW_DISPLAY_POWER_DOMAINS (                             \
5979         (POWER_DOMAIN_MASK & ~HSW_ALWAYS_ON_POWER_DOMAINS) |    \
5980         BIT(POWER_DOMAIN_INIT))
5981
5982 #define BDW_ALWAYS_ON_POWER_DOMAINS (                   \
5983         HSW_ALWAYS_ON_POWER_DOMAINS |                   \
5984         BIT(POWER_DOMAIN_PIPE_A_PANEL_FITTER))
5985 #define BDW_DISPLAY_POWER_DOMAINS (                             \
5986         (POWER_DOMAIN_MASK & ~BDW_ALWAYS_ON_POWER_DOMAINS) |    \
5987         BIT(POWER_DOMAIN_INIT))
5988
5989 #define VLV_ALWAYS_ON_POWER_DOMAINS     BIT(POWER_DOMAIN_INIT)
5990 #define VLV_DISPLAY_POWER_DOMAINS       POWER_DOMAIN_MASK
5991
5992 #define VLV_DPIO_CMN_BC_POWER_DOMAINS (         \
5993         BIT(POWER_DOMAIN_PORT_DDI_B_2_LANES) |  \
5994         BIT(POWER_DOMAIN_PORT_DDI_B_4_LANES) |  \
5995         BIT(POWER_DOMAIN_PORT_DDI_C_2_LANES) |  \
5996         BIT(POWER_DOMAIN_PORT_DDI_C_4_LANES) |  \
5997         BIT(POWER_DOMAIN_PORT_CRT) |            \
5998         BIT(POWER_DOMAIN_INIT))
5999
6000 #define VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS (  \
6001         BIT(POWER_DOMAIN_PORT_DDI_B_2_LANES) |  \
6002         BIT(POWER_DOMAIN_PORT_DDI_B_4_LANES) |  \
6003         BIT(POWER_DOMAIN_INIT))
6004
6005 #define VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS (  \
6006         BIT(POWER_DOMAIN_PORT_DDI_B_4_LANES) |  \
6007         BIT(POWER_DOMAIN_INIT))
6008
6009 #define VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS (  \
6010         BIT(POWER_DOMAIN_PORT_DDI_C_2_LANES) |  \
6011         BIT(POWER_DOMAIN_PORT_DDI_C_4_LANES) |  \
6012         BIT(POWER_DOMAIN_INIT))
6013
6014 #define VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS (  \
6015         BIT(POWER_DOMAIN_PORT_DDI_C_4_LANES) |  \
6016         BIT(POWER_DOMAIN_INIT))
6017
6018 static const struct i915_power_well_ops i9xx_always_on_power_well_ops = {
6019         .sync_hw = i9xx_always_on_power_well_noop,
6020         .enable = i9xx_always_on_power_well_noop,
6021         .disable = i9xx_always_on_power_well_noop,
6022         .is_enabled = i9xx_always_on_power_well_enabled,
6023 };
6024
6025 static struct i915_power_well i9xx_always_on_power_well[] = {
6026         {
6027                 .name = "always-on",
6028                 .always_on = 1,
6029                 .domains = POWER_DOMAIN_MASK,
6030                 .ops = &i9xx_always_on_power_well_ops,
6031         },
6032 };
6033
6034 static const struct i915_power_well_ops hsw_power_well_ops = {
6035         .sync_hw = hsw_power_well_sync_hw,
6036         .enable = hsw_power_well_enable,
6037         .disable = hsw_power_well_disable,
6038         .is_enabled = hsw_power_well_enabled,
6039 };
6040
6041 static struct i915_power_well hsw_power_wells[] = {
6042         {
6043                 .name = "always-on",
6044                 .always_on = 1,
6045                 .domains = HSW_ALWAYS_ON_POWER_DOMAINS,
6046                 .ops = &i9xx_always_on_power_well_ops,
6047         },
6048         {
6049                 .name = "display",
6050                 .domains = HSW_DISPLAY_POWER_DOMAINS,
6051                 .ops = &hsw_power_well_ops,
6052         },
6053 };
6054
6055 static struct i915_power_well bdw_power_wells[] = {
6056         {
6057                 .name = "always-on",
6058                 .always_on = 1,
6059                 .domains = BDW_ALWAYS_ON_POWER_DOMAINS,
6060                 .ops = &i9xx_always_on_power_well_ops,
6061         },
6062         {
6063                 .name = "display",
6064                 .domains = BDW_DISPLAY_POWER_DOMAINS,
6065                 .ops = &hsw_power_well_ops,
6066         },
6067 };
6068
6069 static const struct i915_power_well_ops vlv_display_power_well_ops = {
6070         .sync_hw = vlv_power_well_sync_hw,
6071         .enable = vlv_display_power_well_enable,
6072         .disable = vlv_display_power_well_disable,
6073         .is_enabled = vlv_power_well_enabled,
6074 };
6075
6076 static const struct i915_power_well_ops vlv_dpio_power_well_ops = {
6077         .sync_hw = vlv_power_well_sync_hw,
6078         .enable = vlv_power_well_enable,
6079         .disable = vlv_power_well_disable,
6080         .is_enabled = vlv_power_well_enabled,
6081 };
6082
6083 static struct i915_power_well vlv_power_wells[] = {
6084         {
6085                 .name = "always-on",
6086                 .always_on = 1,
6087                 .domains = VLV_ALWAYS_ON_POWER_DOMAINS,
6088                 .ops = &i9xx_always_on_power_well_ops,
6089         },
6090         {
6091                 .name = "display",
6092                 .domains = VLV_DISPLAY_POWER_DOMAINS,
6093                 .data = PUNIT_POWER_WELL_DISP2D,
6094                 .ops = &vlv_display_power_well_ops,
6095         },
6096         {
6097                 .name = "dpio-common",
6098                 .domains = VLV_DPIO_CMN_BC_POWER_DOMAINS,
6099                 .data = PUNIT_POWER_WELL_DPIO_CMN_BC,
6100                 .ops = &vlv_dpio_power_well_ops,
6101         },
6102         {
6103                 .name = "dpio-tx-b-01",
6104                 .domains = VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS |
6105                            VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS |
6106                            VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS |
6107                            VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS,
6108                 .ops = &vlv_dpio_power_well_ops,
6109                 .data = PUNIT_POWER_WELL_DPIO_TX_B_LANES_01,
6110         },
6111         {
6112                 .name = "dpio-tx-b-23",
6113                 .domains = VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS |
6114                            VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS |
6115                            VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS |
6116                            VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS,
6117                 .ops = &vlv_dpio_power_well_ops,
6118                 .data = PUNIT_POWER_WELL_DPIO_TX_B_LANES_23,
6119         },
6120         {
6121                 .name = "dpio-tx-c-01",
6122                 .domains = VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS |
6123                            VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS |
6124                            VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS |
6125                            VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS,
6126                 .ops = &vlv_dpio_power_well_ops,
6127                 .data = PUNIT_POWER_WELL_DPIO_TX_C_LANES_01,
6128         },
6129         {
6130                 .name = "dpio-tx-c-23",
6131                 .domains = VLV_DPIO_TX_B_LANES_01_POWER_DOMAINS |
6132                            VLV_DPIO_TX_B_LANES_23_POWER_DOMAINS |
6133                            VLV_DPIO_TX_C_LANES_01_POWER_DOMAINS |
6134                            VLV_DPIO_TX_C_LANES_23_POWER_DOMAINS,
6135                 .ops = &vlv_dpio_power_well_ops,
6136                 .data = PUNIT_POWER_WELL_DPIO_TX_C_LANES_23,
6137         },
6138 };
6139
6140 #define set_power_wells(power_domains, __power_wells) ({                \
6141         (power_domains)->power_wells = (__power_wells);                 \
6142         (power_domains)->power_well_count = ARRAY_SIZE(__power_wells);  \
6143 })
6144
6145 int intel_power_domains_init(struct drm_i915_private *dev_priv)
6146 {
6147         struct i915_power_domains *power_domains = &dev_priv->power_domains;
6148
6149         mutex_init(&power_domains->lock);
6150
6151         /*
6152          * The enabling order will be from lower to higher indexed wells,
6153          * the disabling order is reversed.
6154          */
6155         if (IS_HASWELL(dev_priv->dev)) {
6156                 set_power_wells(power_domains, hsw_power_wells);
6157                 hsw_pwr = power_domains;
6158         } else if (IS_BROADWELL(dev_priv->dev)) {
6159                 set_power_wells(power_domains, bdw_power_wells);
6160                 hsw_pwr = power_domains;
6161         } else if (IS_VALLEYVIEW(dev_priv->dev)) {
6162                 set_power_wells(power_domains, vlv_power_wells);
6163         } else {
6164                 set_power_wells(power_domains, i9xx_always_on_power_well);
6165         }
6166
6167         return 0;
6168 }
6169
6170 void intel_power_domains_remove(struct drm_i915_private *dev_priv)
6171 {
6172         hsw_pwr = NULL;
6173 }
6174
6175 static void intel_power_domains_resume(struct drm_i915_private *dev_priv)
6176 {
6177         struct i915_power_domains *power_domains = &dev_priv->power_domains;
6178         struct i915_power_well *power_well;
6179         int i;
6180
6181         mutex_lock(&power_domains->lock);
6182         for_each_power_well(i, power_well, POWER_DOMAIN_MASK, power_domains)
6183                 power_well->ops->sync_hw(dev_priv, power_well);
6184         mutex_unlock(&power_domains->lock);
6185 }
6186
6187 void intel_power_domains_init_hw(struct drm_i915_private *dev_priv)
6188 {
6189         struct i915_power_domains *power_domains = &dev_priv->power_domains;
6190
6191         power_domains->initializing = true;
6192         /* For now, we need the power well to be always enabled. */
6193         intel_display_set_init_power(dev_priv, true);
6194         intel_power_domains_resume(dev_priv);
6195         power_domains->initializing = false;
6196 }
6197
6198 void intel_aux_display_runtime_get(struct drm_i915_private *dev_priv)
6199 {
6200         intel_runtime_pm_get(dev_priv);
6201 }
6202
6203 void intel_aux_display_runtime_put(struct drm_i915_private *dev_priv)
6204 {
6205         intel_runtime_pm_put(dev_priv);
6206 }
6207
6208 void intel_runtime_pm_get(struct drm_i915_private *dev_priv)
6209 {
6210         struct drm_device *dev = dev_priv->dev;
6211         struct device *device = &dev->pdev->dev;
6212
6213         if (!HAS_RUNTIME_PM(dev))
6214                 return;
6215
6216         pm_runtime_get_sync(device);
6217         WARN(dev_priv->pm.suspended, "Device still suspended.\n");
6218 }
6219
6220 void intel_runtime_pm_get_noresume(struct drm_i915_private *dev_priv)
6221 {
6222         struct drm_device *dev = dev_priv->dev;
6223         struct device *device = &dev->pdev->dev;
6224
6225         if (!HAS_RUNTIME_PM(dev))
6226                 return;
6227
6228         WARN(dev_priv->pm.suspended, "Getting nosync-ref while suspended.\n");
6229         pm_runtime_get_noresume(device);
6230 }
6231
6232 void intel_runtime_pm_put(struct drm_i915_private *dev_priv)
6233 {
6234         struct drm_device *dev = dev_priv->dev;
6235         struct device *device = &dev->pdev->dev;
6236
6237         if (!HAS_RUNTIME_PM(dev))
6238                 return;
6239
6240         pm_runtime_mark_last_busy(device);
6241         pm_runtime_put_autosuspend(device);
6242 }
6243
6244 void intel_init_runtime_pm(struct drm_i915_private *dev_priv)
6245 {
6246         struct drm_device *dev = dev_priv->dev;
6247         struct device *device = &dev->pdev->dev;
6248
6249         if (!HAS_RUNTIME_PM(dev))
6250                 return;
6251
6252         pm_runtime_set_active(device);
6253
6254         /*
6255          * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
6256          * requirement.
6257          */
6258         if (!intel_enable_rc6(dev)) {
6259                 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
6260                 return;
6261         }
6262
6263         pm_runtime_set_autosuspend_delay(device, 10000); /* 10s */
6264         pm_runtime_mark_last_busy(device);
6265         pm_runtime_use_autosuspend(device);
6266
6267         pm_runtime_put_autosuspend(device);
6268 }
6269
6270 void intel_fini_runtime_pm(struct drm_i915_private *dev_priv)
6271 {
6272         struct drm_device *dev = dev_priv->dev;
6273         struct device *device = &dev->pdev->dev;
6274
6275         if (!HAS_RUNTIME_PM(dev))
6276                 return;
6277
6278         if (!intel_enable_rc6(dev))
6279                 return;
6280
6281         /* Make sure we're not suspended first. */
6282         pm_runtime_get_sync(device);
6283         pm_runtime_disable(device);
6284 }
6285
6286 /* Set up chip specific power management-related functions */
6287 void intel_init_pm(struct drm_device *dev)
6288 {
6289         struct drm_i915_private *dev_priv = dev->dev_private;
6290
6291         if (HAS_FBC(dev)) {
6292                 if (INTEL_INFO(dev)->gen >= 7) {
6293                         dev_priv->display.fbc_enabled = ironlake_fbc_enabled;
6294                         dev_priv->display.enable_fbc = gen7_enable_fbc;
6295                         dev_priv->display.disable_fbc = ironlake_disable_fbc;
6296                 } else if (INTEL_INFO(dev)->gen >= 5) {
6297                         dev_priv->display.fbc_enabled = ironlake_fbc_enabled;
6298                         dev_priv->display.enable_fbc = ironlake_enable_fbc;
6299                         dev_priv->display.disable_fbc = ironlake_disable_fbc;
6300                 } else if (IS_GM45(dev)) {
6301                         dev_priv->display.fbc_enabled = g4x_fbc_enabled;
6302                         dev_priv->display.enable_fbc = g4x_enable_fbc;
6303                         dev_priv->display.disable_fbc = g4x_disable_fbc;
6304                 } else {
6305                         dev_priv->display.fbc_enabled = i8xx_fbc_enabled;
6306                         dev_priv->display.enable_fbc = i8xx_enable_fbc;
6307                         dev_priv->display.disable_fbc = i8xx_disable_fbc;
6308
6309                         /* This value was pulled out of someone's hat */
6310                         I915_WRITE(FBC_CONTROL, 500 << FBC_CTL_INTERVAL_SHIFT);
6311                 }
6312         }
6313
6314         /* For cxsr */
6315         if (IS_PINEVIEW(dev))
6316                 i915_pineview_get_mem_freq(dev);
6317         else if (IS_GEN5(dev))
6318                 i915_ironlake_get_mem_freq(dev);
6319
6320         /* For FIFO watermark updates */
6321         if (HAS_PCH_SPLIT(dev)) {
6322                 ilk_setup_wm_latency(dev);
6323
6324                 if ((IS_GEN5(dev) && dev_priv->wm.pri_latency[1] &&
6325                      dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
6326                     (!IS_GEN5(dev) && dev_priv->wm.pri_latency[0] &&
6327                      dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
6328                         dev_priv->display.update_wm = ilk_update_wm;
6329                         dev_priv->display.update_sprite_wm = ilk_update_sprite_wm;
6330                 } else {
6331                         DRM_DEBUG_KMS("Failed to read display plane latency. "
6332                                       "Disable CxSR\n");
6333                 }
6334
6335                 if (IS_GEN5(dev))
6336                         dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
6337                 else if (IS_GEN6(dev))
6338                         dev_priv->display.init_clock_gating = gen6_init_clock_gating;
6339                 else if (IS_IVYBRIDGE(dev))
6340                         dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
6341                 else if (IS_HASWELL(dev))
6342                         dev_priv->display.init_clock_gating = haswell_init_clock_gating;
6343                 else if (INTEL_INFO(dev)->gen == 8)
6344                         dev_priv->display.init_clock_gating = gen8_init_clock_gating;
6345         } else if (IS_CHERRYVIEW(dev)) {
6346                 dev_priv->display.update_wm = valleyview_update_wm;
6347                 dev_priv->display.init_clock_gating =
6348                         cherryview_init_clock_gating;
6349         } else if (IS_VALLEYVIEW(dev)) {
6350                 dev_priv->display.update_wm = valleyview_update_wm;
6351                 dev_priv->display.init_clock_gating =
6352                         valleyview_init_clock_gating;
6353         } else if (IS_PINEVIEW(dev)) {
6354                 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
6355                                             dev_priv->is_ddr3,
6356                                             dev_priv->fsb_freq,
6357                                             dev_priv->mem_freq)) {
6358                         DRM_INFO("failed to find known CxSR latency "
6359                                  "(found ddr%s fsb freq %d, mem freq %d), "
6360                                  "disabling CxSR\n",
6361                                  (dev_priv->is_ddr3 == 1) ? "3" : "2",
6362                                  dev_priv->fsb_freq, dev_priv->mem_freq);
6363                         /* Disable CxSR and never update its watermark again */
6364                         pineview_disable_cxsr(dev);
6365                         dev_priv->display.update_wm = NULL;
6366                 } else
6367                         dev_priv->display.update_wm = pineview_update_wm;
6368                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
6369         } else if (IS_G4X(dev)) {
6370                 dev_priv->display.update_wm = g4x_update_wm;
6371                 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
6372         } else if (IS_GEN4(dev)) {
6373                 dev_priv->display.update_wm = i965_update_wm;
6374                 if (IS_CRESTLINE(dev))
6375                         dev_priv->display.init_clock_gating = crestline_init_clock_gating;
6376                 else if (IS_BROADWATER(dev))
6377                         dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
6378         } else if (IS_GEN3(dev)) {
6379                 dev_priv->display.update_wm = i9xx_update_wm;
6380                 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
6381                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
6382         } else if (IS_GEN2(dev)) {
6383                 if (INTEL_INFO(dev)->num_pipes == 1) {
6384                         dev_priv->display.update_wm = i845_update_wm;
6385                         dev_priv->display.get_fifo_size = i845_get_fifo_size;
6386                 } else {
6387                         dev_priv->display.update_wm = i9xx_update_wm;
6388                         dev_priv->display.get_fifo_size = i830_get_fifo_size;
6389                 }
6390
6391                 if (IS_I85X(dev) || IS_I865G(dev))
6392                         dev_priv->display.init_clock_gating = i85x_init_clock_gating;
6393                 else
6394                         dev_priv->display.init_clock_gating = i830_init_clock_gating;
6395         } else {
6396                 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
6397         }
6398 }
6399
6400 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u8 mbox, u32 *val)
6401 {
6402         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6403
6404         if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
6405                 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
6406                 return -EAGAIN;
6407         }
6408
6409         I915_WRITE(GEN6_PCODE_DATA, *val);
6410         I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
6411
6412         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
6413                      500)) {
6414                 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
6415                 return -ETIMEDOUT;
6416         }
6417
6418         *val = I915_READ(GEN6_PCODE_DATA);
6419         I915_WRITE(GEN6_PCODE_DATA, 0);
6420
6421         return 0;
6422 }
6423
6424 int sandybridge_pcode_write(struct drm_i915_private *dev_priv, u8 mbox, u32 val)
6425 {
6426         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6427
6428         if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
6429                 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
6430                 return -EAGAIN;
6431         }
6432
6433         I915_WRITE(GEN6_PCODE_DATA, val);
6434         I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
6435
6436         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
6437                      500)) {
6438                 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
6439                 return -ETIMEDOUT;
6440         }
6441
6442         I915_WRITE(GEN6_PCODE_DATA, 0);
6443
6444         return 0;
6445 }
6446
6447 int vlv_gpu_freq(struct drm_i915_private *dev_priv, int val)
6448 {
6449         int div;
6450
6451         /* 4 x czclk */
6452         switch (dev_priv->mem_freq) {
6453         case 800:
6454                 div = 10;
6455                 break;
6456         case 1066:
6457                 div = 12;
6458                 break;
6459         case 1333:
6460                 div = 16;
6461                 break;
6462         default:
6463                 return -1;
6464         }
6465
6466         return DIV_ROUND_CLOSEST(dev_priv->mem_freq * (val + 6 - 0xbd), 4 * div);
6467 }
6468
6469 int vlv_freq_opcode(struct drm_i915_private *dev_priv, int val)
6470 {
6471         int mul;
6472
6473         /* 4 x czclk */
6474         switch (dev_priv->mem_freq) {
6475         case 800:
6476                 mul = 10;
6477                 break;
6478         case 1066:
6479                 mul = 12;
6480                 break;
6481         case 1333:
6482                 mul = 16;
6483                 break;
6484         default:
6485                 return -1;
6486         }
6487
6488         return DIV_ROUND_CLOSEST(4 * mul * val, dev_priv->mem_freq) + 0xbd - 6;
6489 }
6490
6491 void intel_pm_setup(struct drm_device *dev)
6492 {
6493         struct drm_i915_private *dev_priv = dev->dev_private;
6494
6495         mutex_init(&dev_priv->rps.hw_lock);
6496
6497         INIT_DELAYED_WORK(&dev_priv->rps.delayed_resume_work,
6498                           intel_gen6_powersave_work);
6499
6500         dev_priv->pm.suspended = false;
6501         dev_priv->pm.irqs_disabled = false;
6502 }