Pileus Git - ~andy/linux/blob - 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
  34 #define FORCEWAKE_ACK_TIMEOUT_MS 2
  35
  36 /* FBC, or Frame Buffer Compression, is a technique employed to compress the
  37  * framebuffer contents in-memory, aiming at reducing the required bandwidth
  38  * during in-memory transfers and, therefore, reduce the power packet.
  39  *
  40  * The benefits of FBC are mostly visible with solid backgrounds and
  41  * variation-less patterns.
  42  *
  43  * FBC-related functionality can be enabled by the means of the
  44  * i915.i915_enable_fbc parameter
  45  */
  46
  47 static void i8xx_disable_fbc(struct drm_device *dev)
  48 {
  49         struct drm_i915_private *dev_priv = dev->dev_private;
  50         u32 fbc_ctl;
  51
  52         /* Disable compression */
  53         fbc_ctl = I915_READ(FBC_CONTROL);
  54         if ((fbc_ctl & FBC_CTL_EN) == 0)
  55                 return;
  56
  57         fbc_ctl &= ~FBC_CTL_EN;
  58         I915_WRITE(FBC_CONTROL, fbc_ctl);
  59
  60         /* Wait for compressing bit to clear */
  61         if (wait_for((I915_READ(FBC_STATUS) & FBC_STAT_COMPRESSING) == 0, 10)) {
  62                 DRM_DEBUG_KMS("FBC idle timed out\n");
  63                 return;
  64         }
  65
  66         DRM_DEBUG_KMS("disabled FBC\n");
  67 }
  68
  69 static void i8xx_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
  70 {
  71         struct drm_device *dev = crtc->dev;
  72         struct drm_i915_private *dev_priv = dev->dev_private;
  73         struct drm_framebuffer *fb = crtc->fb;
  74         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
  75         struct drm_i915_gem_object *obj = intel_fb->obj;
  76         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
  77         int cfb_pitch;
  78         int plane, i;
  79         u32 fbc_ctl, fbc_ctl2;
  80
  81         cfb_pitch = dev_priv->cfb_size / FBC_LL_SIZE;
  82         if (fb->pitches[0] < cfb_pitch)
  83                 cfb_pitch = fb->pitches[0];
  84
  85         /* FBC_CTL wants 64B units */
  86         cfb_pitch = (cfb_pitch / 64) - 1;
  87         plane = intel_crtc->plane == 0 ? FBC_CTL_PLANEA : FBC_CTL_PLANEB;
  88
  89         /* Clear old tags */
  90         for (i = 0; i < (FBC_LL_SIZE / 32) + 1; i++)
  91                 I915_WRITE(FBC_TAG + (i * 4), 0);
  92
  93         /* Set it up... */
  94         fbc_ctl2 = FBC_CTL_FENCE_DBL | FBC_CTL_IDLE_IMM | FBC_CTL_CPU_FENCE;
  95         fbc_ctl2 |= plane;
  96         I915_WRITE(FBC_CONTROL2, fbc_ctl2);
  97         I915_WRITE(FBC_FENCE_OFF, crtc->y);
  98
  99         /* enable it... */
 100         fbc_ctl = FBC_CTL_EN | FBC_CTL_PERIODIC;
 101         if (IS_I945GM(dev))
 102                 fbc_ctl |= FBC_CTL_C3_IDLE; /* 945 needs special SR handling */
 103         fbc_ctl |= (cfb_pitch & 0xff) << FBC_CTL_STRIDE_SHIFT;
 104         fbc_ctl |= (interval & 0x2fff) << FBC_CTL_INTERVAL_SHIFT;
 105         fbc_ctl |= obj->fence_reg;
 106         I915_WRITE(FBC_CONTROL, fbc_ctl);
 107
 108         DRM_DEBUG_KMS("enabled FBC, pitch %d, yoff %d, plane %d, ",
 109                       cfb_pitch, crtc->y, intel_crtc->plane);
 110 }
 111
 112 static bool i8xx_fbc_enabled(struct drm_device *dev)
 113 {
 114         struct drm_i915_private *dev_priv = dev->dev_private;
 115
 116         return I915_READ(FBC_CONTROL) & FBC_CTL_EN;
 117 }
 118
 119 static void g4x_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 120 {
 121         struct drm_device *dev = crtc->dev;
 122         struct drm_i915_private *dev_priv = dev->dev_private;
 123         struct drm_framebuffer *fb = crtc->fb;
 124         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 125         struct drm_i915_gem_object *obj = intel_fb->obj;
 126         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 127         int plane = intel_crtc->plane == 0 ? DPFC_CTL_PLANEA : DPFC_CTL_PLANEB;
 128         unsigned long stall_watermark = 200;
 129         u32 dpfc_ctl;
 130
 131         dpfc_ctl = plane | DPFC_SR_EN | DPFC_CTL_LIMIT_1X;
 132         dpfc_ctl |= DPFC_CTL_FENCE_EN | obj->fence_reg;
 133         I915_WRITE(DPFC_CHICKEN, DPFC_HT_MODIFY);
 134
 135         I915_WRITE(DPFC_RECOMP_CTL, DPFC_RECOMP_STALL_EN |
 136                    (stall_watermark << DPFC_RECOMP_STALL_WM_SHIFT) |
 137                    (interval << DPFC_RECOMP_TIMER_COUNT_SHIFT));
 138         I915_WRITE(DPFC_FENCE_YOFF, crtc->y);
 139
 140         /* enable it... */
 141         I915_WRITE(DPFC_CONTROL, I915_READ(DPFC_CONTROL) | DPFC_CTL_EN);
 142
 143         DRM_DEBUG_KMS("enabled fbc on plane %d\n", intel_crtc->plane);
 144 }
 145
 146 static void g4x_disable_fbc(struct drm_device *dev)
 147 {
 148         struct drm_i915_private *dev_priv = dev->dev_private;
 149         u32 dpfc_ctl;
 150
 151         /* Disable compression */
 152         dpfc_ctl = I915_READ(DPFC_CONTROL);
 153         if (dpfc_ctl & DPFC_CTL_EN) {
 154                 dpfc_ctl &= ~DPFC_CTL_EN;
 155                 I915_WRITE(DPFC_CONTROL, dpfc_ctl);
 156
 157                 DRM_DEBUG_KMS("disabled FBC\n");
 158         }
 159 }
 160
 161 static bool g4x_fbc_enabled(struct drm_device *dev)
 162 {
 163         struct drm_i915_private *dev_priv = dev->dev_private;
 164
 165         return I915_READ(DPFC_CONTROL) & DPFC_CTL_EN;
 166 }
 167
 168 static void sandybridge_blit_fbc_update(struct drm_device *dev)
 169 {
 170         struct drm_i915_private *dev_priv = dev->dev_private;
 171         u32 blt_ecoskpd;
 172
 173         /* Make sure blitter notifies FBC of writes */
 174         gen6_gt_force_wake_get(dev_priv);
 175         blt_ecoskpd = I915_READ(GEN6_BLITTER_ECOSKPD);
 176         blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY <<
 177                 GEN6_BLITTER_LOCK_SHIFT;
 178         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 179         blt_ecoskpd |= GEN6_BLITTER_FBC_NOTIFY;
 180         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 181         blt_ecoskpd &= ~(GEN6_BLITTER_FBC_NOTIFY <<
 182                          GEN6_BLITTER_LOCK_SHIFT);
 183         I915_WRITE(GEN6_BLITTER_ECOSKPD, blt_ecoskpd);
 184         POSTING_READ(GEN6_BLITTER_ECOSKPD);
 185         gen6_gt_force_wake_put(dev_priv);
 186 }
 187
 188 static void ironlake_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 189 {
 190         struct drm_device *dev = crtc->dev;
 191         struct drm_i915_private *dev_priv = dev->dev_private;
 192         struct drm_framebuffer *fb = crtc->fb;
 193         struct intel_framebuffer *intel_fb = to_intel_framebuffer(fb);
 194         struct drm_i915_gem_object *obj = intel_fb->obj;
 195         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 196         int plane = intel_crtc->plane == 0 ? DPFC_CTL_PLANEA : DPFC_CTL_PLANEB;
 197         unsigned long stall_watermark = 200;
 198         u32 dpfc_ctl;
 199
 200         dpfc_ctl = I915_READ(ILK_DPFC_CONTROL);
 201         dpfc_ctl &= DPFC_RESERVED;
 202         dpfc_ctl |= (plane | DPFC_CTL_LIMIT_1X);
 203         /* Set persistent mode for front-buffer rendering, ala X. */
 204         dpfc_ctl |= DPFC_CTL_PERSISTENT_MODE;
 205         dpfc_ctl |= (DPFC_CTL_FENCE_EN | obj->fence_reg);
 206         I915_WRITE(ILK_DPFC_CHICKEN, DPFC_HT_MODIFY);
 207
 208         I915_WRITE(ILK_DPFC_RECOMP_CTL, DPFC_RECOMP_STALL_EN |
 209                    (stall_watermark << DPFC_RECOMP_STALL_WM_SHIFT) |
 210                    (interval << DPFC_RECOMP_TIMER_COUNT_SHIFT));
 211         I915_WRITE(ILK_DPFC_FENCE_YOFF, crtc->y);
 212         I915_WRITE(ILK_FBC_RT_BASE, obj->gtt_offset | ILK_FBC_RT_VALID);
 213         /* enable it... */
 214         I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl | DPFC_CTL_EN);
 215
 216         if (IS_GEN6(dev)) {
 217                 I915_WRITE(SNB_DPFC_CTL_SA,
 218                            SNB_CPU_FENCE_ENABLE | obj->fence_reg);
 219                 I915_WRITE(DPFC_CPU_FENCE_OFFSET, crtc->y);
 220                 sandybridge_blit_fbc_update(dev);
 221         }
 222
 223         DRM_DEBUG_KMS("enabled fbc on plane %d\n", intel_crtc->plane);
 224 }
 225
 226 static void ironlake_disable_fbc(struct drm_device *dev)
 227 {
 228         struct drm_i915_private *dev_priv = dev->dev_private;
 229         u32 dpfc_ctl;
 230
 231         /* Disable compression */
 232         dpfc_ctl = I915_READ(ILK_DPFC_CONTROL);
 233         if (dpfc_ctl & DPFC_CTL_EN) {
 234                 dpfc_ctl &= ~DPFC_CTL_EN;
 235                 I915_WRITE(ILK_DPFC_CONTROL, dpfc_ctl);
 236
 237                 DRM_DEBUG_KMS("disabled FBC\n");
 238         }
 239 }
 240
 241 static bool ironlake_fbc_enabled(struct drm_device *dev)
 242 {
 243         struct drm_i915_private *dev_priv = dev->dev_private;
 244
 245         return I915_READ(ILK_DPFC_CONTROL) & DPFC_CTL_EN;
 246 }
 247
 248 bool intel_fbc_enabled(struct drm_device *dev)
 249 {
 250         struct drm_i915_private *dev_priv = dev->dev_private;
 251
 252         if (!dev_priv->display.fbc_enabled)
 253                 return false;
 254
 255         return dev_priv->display.fbc_enabled(dev);
 256 }
 257
 258 static void intel_fbc_work_fn(struct work_struct *__work)
 259 {
 260         struct intel_fbc_work *work =
 261                 container_of(to_delayed_work(__work),
 262                              struct intel_fbc_work, work);
 263         struct drm_device *dev = work->crtc->dev;
 264         struct drm_i915_private *dev_priv = dev->dev_private;
 265
 266         mutex_lock(&dev->struct_mutex);
 267         if (work == dev_priv->fbc_work) {
 268                 /* Double check that we haven't switched fb without cancelling
 269                  * the prior work.
 270                  */
 271                 if (work->crtc->fb == work->fb) {
 272                         dev_priv->display.enable_fbc(work->crtc,
 273                                                      work->interval);
 274
 275                         dev_priv->cfb_plane = to_intel_crtc(work->crtc)->plane;
 276                         dev_priv->cfb_fb = work->crtc->fb->base.id;
 277                         dev_priv->cfb_y = work->crtc->y;
 278                 }
 279
 280                 dev_priv->fbc_work = NULL;
 281         }
 282         mutex_unlock(&dev->struct_mutex);
 283
 284         kfree(work);
 285 }
 286
 287 static void intel_cancel_fbc_work(struct drm_i915_private *dev_priv)
 288 {
 289         if (dev_priv->fbc_work == NULL)
 290                 return;
 291
 292         DRM_DEBUG_KMS("cancelling pending FBC enable\n");
 293
 294         /* Synchronisation is provided by struct_mutex and checking of
 295          * dev_priv->fbc_work, so we can perform the cancellation
 296          * entirely asynchronously.
 297          */
 298         if (cancel_delayed_work(&dev_priv->fbc_work->work))
 299                 /* tasklet was killed before being run, clean up */
 300                 kfree(dev_priv->fbc_work);
 301
 302         /* Mark the work as no longer wanted so that if it does
 303          * wake-up (because the work was already running and waiting
 304          * for our mutex), it will discover that is no longer
 305          * necessary to run.
 306          */
 307         dev_priv->fbc_work = NULL;
 308 }
 309
 310 void intel_enable_fbc(struct drm_crtc *crtc, unsigned long interval)
 311 {
 312         struct intel_fbc_work *work;
 313         struct drm_device *dev = crtc->dev;
 314         struct drm_i915_private *dev_priv = dev->dev_private;
 315
 316         if (!dev_priv->display.enable_fbc)
 317                 return;
 318
 319         intel_cancel_fbc_work(dev_priv);
 320
 321         work = kzalloc(sizeof *work, GFP_KERNEL);
 322         if (work == NULL) {
 323                 dev_priv->display.enable_fbc(crtc, interval);
 324                 return;
 325         }
 326
 327         work->crtc = crtc;
 328         work->fb = crtc->fb;
 329         work->interval = interval;
 330         INIT_DELAYED_WORK(&work->work, intel_fbc_work_fn);
 331
 332         dev_priv->fbc_work = work;
 333
 334         DRM_DEBUG_KMS("scheduling delayed FBC enable\n");
 335
 336         /* Delay the actual enabling to let pageflipping cease and the
 337          * display to settle before starting the compression. Note that
 338          * this delay also serves a second purpose: it allows for a
 339          * vblank to pass after disabling the FBC before we attempt
 340          * to modify the control registers.
 341          *
 342          * A more complicated solution would involve tracking vblanks
 343          * following the termination of the page-flipping sequence
 344          * and indeed performing the enable as a co-routine and not
 345          * waiting synchronously upon the vblank.
 346          */
 347         schedule_delayed_work(&work->work, msecs_to_jiffies(50));
 348 }
 349
 350 void intel_disable_fbc(struct drm_device *dev)
 351 {
 352         struct drm_i915_private *dev_priv = dev->dev_private;
 353
 354         intel_cancel_fbc_work(dev_priv);
 355
 356         if (!dev_priv->display.disable_fbc)
 357                 return;
 358
 359         dev_priv->display.disable_fbc(dev);
 360         dev_priv->cfb_plane = -1;
 361 }
 362
 363 /**
 364  * intel_update_fbc - enable/disable FBC as needed
 365  * @dev: the drm_device
 366  *
 367  * Set up the framebuffer compression hardware at mode set time.  We
 368  * enable it if possible:
 369  *   - plane A only (on pre-965)
 370  *   - no pixel mulitply/line duplication
 371  *   - no alpha buffer discard
 372  *   - no dual wide
 373  *   - framebuffer <= 2048 in width, 1536 in height
 374  *
 375  * We can't assume that any compression will take place (worst case),
 376  * so the compressed buffer has to be the same size as the uncompressed
 377  * one.  It also must reside (along with the line length buffer) in
 378  * stolen memory.
 379  *
 380  * We need to enable/disable FBC on a global basis.
 381  */
 382 void intel_update_fbc(struct drm_device *dev)
 383 {
 384         struct drm_i915_private *dev_priv = dev->dev_private;
 385         struct drm_crtc *crtc = NULL, *tmp_crtc;
 386         struct intel_crtc *intel_crtc;
 387         struct drm_framebuffer *fb;
 388         struct intel_framebuffer *intel_fb;
 389         struct drm_i915_gem_object *obj;
 390         int enable_fbc;
 391
 392         if (!i915_powersave)
 393                 return;
 394
 395         if (!I915_HAS_FBC(dev))
 396                 return;
 397
 398         /*
 399          * If FBC is already on, we just have to verify that we can
 400          * keep it that way...
 401          * Need to disable if:
 402          *   - more than one pipe is active
 403          *   - changing FBC params (stride, fence, mode)
 404          *   - new fb is too large to fit in compressed buffer
 405          *   - going to an unsupported config (interlace, pixel multiply, etc.)
 406          */
 407         list_for_each_entry(tmp_crtc, &dev->mode_config.crtc_list, head) {
 408                 if (tmp_crtc->enabled &&
 409                     !to_intel_crtc(tmp_crtc)->primary_disabled &&
 410                     tmp_crtc->fb) {
 411                         if (crtc) {
 412                                 DRM_DEBUG_KMS("more than one pipe active, disabling compression\n");
 413                                 dev_priv->no_fbc_reason = FBC_MULTIPLE_PIPES;
 414                                 goto out_disable;
 415                         }
 416                         crtc = tmp_crtc;
 417                 }
 418         }
 419
 420         if (!crtc || crtc->fb == NULL) {
 421                 DRM_DEBUG_KMS("no output, disabling\n");
 422                 dev_priv->no_fbc_reason = FBC_NO_OUTPUT;
 423                 goto out_disable;
 424         }
 425
 426         intel_crtc = to_intel_crtc(crtc);
 427         fb = crtc->fb;
 428         intel_fb = to_intel_framebuffer(fb);
 429         obj = intel_fb->obj;
 430
 431         enable_fbc = i915_enable_fbc;
 432         if (enable_fbc < 0) {
 433                 DRM_DEBUG_KMS("fbc set to per-chip default\n");
 434                 enable_fbc = 1;
 435                 if (INTEL_INFO(dev)->gen <= 6)
 436                         enable_fbc = 0;
 437         }
 438         if (!enable_fbc) {
 439                 DRM_DEBUG_KMS("fbc disabled per module param\n");
 440                 dev_priv->no_fbc_reason = FBC_MODULE_PARAM;
 441                 goto out_disable;
 442         }
 443         if (intel_fb->obj->base.size > dev_priv->cfb_size) {
 444                 DRM_DEBUG_KMS("framebuffer too large, disabling "
 445                               "compression\n");
 446                 dev_priv->no_fbc_reason = FBC_STOLEN_TOO_SMALL;
 447                 goto out_disable;
 448         }
 449         if ((crtc->mode.flags & DRM_MODE_FLAG_INTERLACE) ||
 450             (crtc->mode.flags & DRM_MODE_FLAG_DBLSCAN)) {
 451                 DRM_DEBUG_KMS("mode incompatible with compression, "
 452                               "disabling\n");
 453                 dev_priv->no_fbc_reason = FBC_UNSUPPORTED_MODE;
 454                 goto out_disable;
 455         }
 456         if ((crtc->mode.hdisplay > 2048) ||
 457             (crtc->mode.vdisplay > 1536)) {
 458                 DRM_DEBUG_KMS("mode too large for compression, disabling\n");
 459                 dev_priv->no_fbc_reason = FBC_MODE_TOO_LARGE;
 460                 goto out_disable;
 461         }
 462         if ((IS_I915GM(dev) || IS_I945GM(dev)) && intel_crtc->plane != 0) {
 463                 DRM_DEBUG_KMS("plane not 0, disabling compression\n");
 464                 dev_priv->no_fbc_reason = FBC_BAD_PLANE;
 465                 goto out_disable;
 466         }
 467
 468         /* The use of a CPU fence is mandatory in order to detect writes
 469          * by the CPU to the scanout and trigger updates to the FBC.
 470          */
 471         if (obj->tiling_mode != I915_TILING_X ||
 472             obj->fence_reg == I915_FENCE_REG_NONE) {
 473                 DRM_DEBUG_KMS("framebuffer not tiled or fenced, disabling compression\n");
 474                 dev_priv->no_fbc_reason = FBC_NOT_TILED;
 475                 goto out_disable;
 476         }
 477
 478         /* If the kernel debugger is active, always disable compression */
 479         if (in_dbg_master())
 480                 goto out_disable;
 481
 482         /* If the scanout has not changed, don't modify the FBC settings.
 483          * Note that we make the fundamental assumption that the fb->obj
 484          * cannot be unpinned (and have its GTT offset and fence revoked)
 485          * without first being decoupled from the scanout and FBC disabled.
 486          */
 487         if (dev_priv->cfb_plane == intel_crtc->plane &&
 488             dev_priv->cfb_fb == fb->base.id &&
 489             dev_priv->cfb_y == crtc->y)
 490                 return;
 491
 492         if (intel_fbc_enabled(dev)) {
 493                 /* We update FBC along two paths, after changing fb/crtc
 494                  * configuration (modeswitching) and after page-flipping
 495                  * finishes. For the latter, we know that not only did
 496                  * we disable the FBC at the start of the page-flip
 497                  * sequence, but also more than one vblank has passed.
 498                  *
 499                  * For the former case of modeswitching, it is possible
 500                  * to switch between two FBC valid configurations
 501                  * instantaneously so we do need to disable the FBC
 502                  * before we can modify its control registers. We also
 503                  * have to wait for the next vblank for that to take
 504                  * effect. However, since we delay enabling FBC we can
 505                  * assume that a vblank has passed since disabling and
 506                  * that we can safely alter the registers in the deferred
 507                  * callback.
 508                  *
 509                  * In the scenario that we go from a valid to invalid
 510                  * and then back to valid FBC configuration we have
 511                  * no strict enforcement that a vblank occurred since
 512                  * disabling the FBC. However, along all current pipe
 513                  * disabling paths we do need to wait for a vblank at
 514                  * some point. And we wait before enabling FBC anyway.
 515                  */
 516                 DRM_DEBUG_KMS("disabling active FBC for update\n");
 517                 intel_disable_fbc(dev);
 518         }
 519
 520         intel_enable_fbc(crtc, 500);
 521         return;
 522
 523 out_disable:
 524         /* Multiple disables should be harmless */
 525         if (intel_fbc_enabled(dev)) {
 526                 DRM_DEBUG_KMS("unsupported config, disabling FBC\n");
 527                 intel_disable_fbc(dev);
 528         }
 529 }
 530
 531 static void i915_pineview_get_mem_freq(struct drm_device *dev)
 532 {
 533         drm_i915_private_t *dev_priv = dev->dev_private;
 534         u32 tmp;
 535
 536         tmp = I915_READ(CLKCFG);
 537
 538         switch (tmp & CLKCFG_FSB_MASK) {
 539         case CLKCFG_FSB_533:
 540                 dev_priv->fsb_freq = 533; /* 133*4 */
 541                 break;
 542         case CLKCFG_FSB_800:
 543                 dev_priv->fsb_freq = 800; /* 200*4 */
 544                 break;
 545         case CLKCFG_FSB_667:
 546                 dev_priv->fsb_freq =  667; /* 167*4 */
 547                 break;
 548         case CLKCFG_FSB_400:
 549                 dev_priv->fsb_freq = 400; /* 100*4 */
 550                 break;
 551         }
 552
 553         switch (tmp & CLKCFG_MEM_MASK) {
 554         case CLKCFG_MEM_533:
 555                 dev_priv->mem_freq = 533;
 556                 break;
 557         case CLKCFG_MEM_667:
 558                 dev_priv->mem_freq = 667;
 559                 break;
 560         case CLKCFG_MEM_800:
 561                 dev_priv->mem_freq = 800;
 562                 break;
 563         }
 564
 565         /* detect pineview DDR3 setting */
 566         tmp = I915_READ(CSHRDDR3CTL);
 567         dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
 568 }
 569
 570 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
 571 {
 572         drm_i915_private_t *dev_priv = dev->dev_private;
 573         u16 ddrpll, csipll;
 574
 575         ddrpll = I915_READ16(DDRMPLL1);
 576         csipll = I915_READ16(CSIPLL0);
 577
 578         switch (ddrpll & 0xff) {
 579         case 0xc:
 580                 dev_priv->mem_freq = 800;
 581                 break;
 582         case 0x10:
 583                 dev_priv->mem_freq = 1066;
 584                 break;
 585         case 0x14:
 586                 dev_priv->mem_freq = 1333;
 587                 break;
 588         case 0x18:
 589                 dev_priv->mem_freq = 1600;
 590                 break;
 591         default:
 592                 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
 593                                  ddrpll & 0xff);
 594                 dev_priv->mem_freq = 0;
 595                 break;
 596         }
 597
 598         dev_priv->ips.r_t = dev_priv->mem_freq;
 599
 600         switch (csipll & 0x3ff) {
 601         case 0x00c:
 602                 dev_priv->fsb_freq = 3200;
 603                 break;
 604         case 0x00e:
 605                 dev_priv->fsb_freq = 3733;
 606                 break;
 607         case 0x010:
 608                 dev_priv->fsb_freq = 4266;
 609                 break;
 610         case 0x012:
 611                 dev_priv->fsb_freq = 4800;
 612                 break;
 613         case 0x014:
 614                 dev_priv->fsb_freq = 5333;
 615                 break;
 616         case 0x016:
 617                 dev_priv->fsb_freq = 5866;
 618                 break;
 619         case 0x018:
 620                 dev_priv->fsb_freq = 6400;
 621                 break;
 622         default:
 623                 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
 624                                  csipll & 0x3ff);
 625                 dev_priv->fsb_freq = 0;
 626                 break;
 627         }
 628
 629         if (dev_priv->fsb_freq == 3200) {
 630                 dev_priv->ips.c_m = 0;
 631         } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
 632                 dev_priv->ips.c_m = 1;
 633         } else {
 634                 dev_priv->ips.c_m = 2;
 635         }
 636 }
 637
 638 static const struct cxsr_latency cxsr_latency_table[] = {
 639         {1, 0, 800, 400, 3382, 33382, 3983, 33983},    /* DDR2-400 SC */
 640         {1, 0, 800, 667, 3354, 33354, 3807, 33807},    /* DDR2-667 SC */
 641         {1, 0, 800, 800, 3347, 33347, 3763, 33763},    /* DDR2-800 SC */
 642         {1, 1, 800, 667, 6420, 36420, 6873, 36873},    /* DDR3-667 SC */
 643         {1, 1, 800, 800, 5902, 35902, 6318, 36318},    /* DDR3-800 SC */
 644
 645         {1, 0, 667, 400, 3400, 33400, 4021, 34021},    /* DDR2-400 SC */
 646         {1, 0, 667, 667, 3372, 33372, 3845, 33845},    /* DDR2-667 SC */
 647         {1, 0, 667, 800, 3386, 33386, 3822, 33822},    /* DDR2-800 SC */
 648         {1, 1, 667, 667, 6438, 36438, 6911, 36911},    /* DDR3-667 SC */
 649         {1, 1, 667, 800, 5941, 35941, 6377, 36377},    /* DDR3-800 SC */
 650
 651         {1, 0, 400, 400, 3472, 33472, 4173, 34173},    /* DDR2-400 SC */
 652         {1, 0, 400, 667, 3443, 33443, 3996, 33996},    /* DDR2-667 SC */
 653         {1, 0, 400, 800, 3430, 33430, 3946, 33946},    /* DDR2-800 SC */
 654         {1, 1, 400, 667, 6509, 36509, 7062, 37062},    /* DDR3-667 SC */
 655         {1, 1, 400, 800, 5985, 35985, 6501, 36501},    /* DDR3-800 SC */
 656
 657         {0, 0, 800, 400, 3438, 33438, 4065, 34065},    /* DDR2-400 SC */
 658         {0, 0, 800, 667, 3410, 33410, 3889, 33889},    /* DDR2-667 SC */
 659         {0, 0, 800, 800, 3403, 33403, 3845, 33845},    /* DDR2-800 SC */
 660         {0, 1, 800, 667, 6476, 36476, 6955, 36955},    /* DDR3-667 SC */
 661         {0, 1, 800, 800, 5958, 35958, 6400, 36400},    /* DDR3-800 SC */
 662
 663         {0, 0, 667, 400, 3456, 33456, 4103, 34106},    /* DDR2-400 SC */
 664         {0, 0, 667, 667, 3428, 33428, 3927, 33927},    /* DDR2-667 SC */
 665         {0, 0, 667, 800, 3443, 33443, 3905, 33905},    /* DDR2-800 SC */
 666         {0, 1, 667, 667, 6494, 36494, 6993, 36993},    /* DDR3-667 SC */
 667         {0, 1, 667, 800, 5998, 35998, 6460, 36460},    /* DDR3-800 SC */
 668
 669         {0, 0, 400, 400, 3528, 33528, 4255, 34255},    /* DDR2-400 SC */
 670         {0, 0, 400, 667, 3500, 33500, 4079, 34079},    /* DDR2-667 SC */
 671         {0, 0, 400, 800, 3487, 33487, 4029, 34029},    /* DDR2-800 SC */
 672         {0, 1, 400, 667, 6566, 36566, 7145, 37145},    /* DDR3-667 SC */
 673         {0, 1, 400, 800, 6042, 36042, 6584, 36584},    /* DDR3-800 SC */
 674 };
 675
 676 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
 677                                                          int is_ddr3,
 678                                                          int fsb,
 679                                                          int mem)
 680 {
 681         const struct cxsr_latency *latency;
 682         int i;
 683
 684         if (fsb == 0 || mem == 0)
 685                 return NULL;
 686
 687         for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
 688                 latency = &cxsr_latency_table[i];
 689                 if (is_desktop == latency->is_desktop &&
 690                     is_ddr3 == latency->is_ddr3 &&
 691                     fsb == latency->fsb_freq && mem == latency->mem_freq)
 692                         return latency;
 693         }
 694
 695         DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 696
 697         return NULL;
 698 }
 699
 700 static void pineview_disable_cxsr(struct drm_device *dev)
 701 {
 702         struct drm_i915_private *dev_priv = dev->dev_private;
 703
 704         /* deactivate cxsr */
 705         I915_WRITE(DSPFW3, I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN);
 706 }
 707
 708 /*
 709  * Latency for FIFO fetches is dependent on several factors:
 710  *   - memory configuration (speed, channels)
 711  *   - chipset
 712  *   - current MCH state
 713  * It can be fairly high in some situations, so here we assume a fairly
 714  * pessimal value.  It's a tradeoff between extra memory fetches (if we
 715  * set this value too high, the FIFO will fetch frequently to stay full)
 716  * and power consumption (set it too low to save power and we might see
 717  * FIFO underruns and display "flicker").
 718  *
 719  * A value of 5us seems to be a good balance; safe for very low end
 720  * platforms but not overly aggressive on lower latency configs.
 721  */
 722 static const int latency_ns = 5000;
 723
 724 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
 725 {
 726         struct drm_i915_private *dev_priv = dev->dev_private;
 727         uint32_t dsparb = I915_READ(DSPARB);
 728         int size;
 729
 730         size = dsparb & 0x7f;
 731         if (plane)
 732                 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
 733
 734         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 735                       plane ? "B" : "A", size);
 736
 737         return size;
 738 }
 739
 740 static int i85x_get_fifo_size(struct drm_device *dev, int plane)
 741 {
 742         struct drm_i915_private *dev_priv = dev->dev_private;
 743         uint32_t dsparb = I915_READ(DSPARB);
 744         int size;
 745
 746         size = dsparb & 0x1ff;
 747         if (plane)
 748                 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
 749         size >>= 1; /* Convert to cachelines */
 750
 751         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 752                       plane ? "B" : "A", size);
 753
 754         return size;
 755 }
 756
 757 static int i845_get_fifo_size(struct drm_device *dev, int plane)
 758 {
 759         struct drm_i915_private *dev_priv = dev->dev_private;
 760         uint32_t dsparb = I915_READ(DSPARB);
 761         int size;
 762
 763         size = dsparb & 0x7f;
 764         size >>= 2; /* Convert to cachelines */
 765
 766         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 767                       plane ? "B" : "A",
 768                       size);
 769
 770         return size;
 771 }
 772
 773 static int i830_get_fifo_size(struct drm_device *dev, int plane)
 774 {
 775         struct drm_i915_private *dev_priv = dev->dev_private;
 776         uint32_t dsparb = I915_READ(DSPARB);
 777         int size;
 778
 779         size = dsparb & 0x7f;
 780         size >>= 1; /* Convert to cachelines */
 781
 782         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 783                       plane ? "B" : "A", size);
 784
 785         return size;
 786 }
 787
 788 /* Pineview has different values for various configs */
 789 static const struct intel_watermark_params pineview_display_wm = {
 790         PINEVIEW_DISPLAY_FIFO,
 791         PINEVIEW_MAX_WM,
 792         PINEVIEW_DFT_WM,
 793         PINEVIEW_GUARD_WM,
 794         PINEVIEW_FIFO_LINE_SIZE
 795 };
 796 static const struct intel_watermark_params pineview_display_hplloff_wm = {
 797         PINEVIEW_DISPLAY_FIFO,
 798         PINEVIEW_MAX_WM,
 799         PINEVIEW_DFT_HPLLOFF_WM,
 800         PINEVIEW_GUARD_WM,
 801         PINEVIEW_FIFO_LINE_SIZE
 802 };
 803 static const struct intel_watermark_params pineview_cursor_wm = {
 804         PINEVIEW_CURSOR_FIFO,
 805         PINEVIEW_CURSOR_MAX_WM,
 806         PINEVIEW_CURSOR_DFT_WM,
 807         PINEVIEW_CURSOR_GUARD_WM,
 808         PINEVIEW_FIFO_LINE_SIZE,
 809 };
 810 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
 811         PINEVIEW_CURSOR_FIFO,
 812         PINEVIEW_CURSOR_MAX_WM,
 813         PINEVIEW_CURSOR_DFT_WM,
 814         PINEVIEW_CURSOR_GUARD_WM,
 815         PINEVIEW_FIFO_LINE_SIZE
 816 };
 817 static const struct intel_watermark_params g4x_wm_info = {
 818         G4X_FIFO_SIZE,
 819         G4X_MAX_WM,
 820         G4X_MAX_WM,
 821         2,
 822         G4X_FIFO_LINE_SIZE,
 823 };
 824 static const struct intel_watermark_params g4x_cursor_wm_info = {
 825         I965_CURSOR_FIFO,
 826         I965_CURSOR_MAX_WM,
 827         I965_CURSOR_DFT_WM,
 828         2,
 829         G4X_FIFO_LINE_SIZE,
 830 };
 831 static const struct intel_watermark_params valleyview_wm_info = {
 832         VALLEYVIEW_FIFO_SIZE,
 833         VALLEYVIEW_MAX_WM,
 834         VALLEYVIEW_MAX_WM,
 835         2,
 836         G4X_FIFO_LINE_SIZE,
 837 };
 838 static const struct intel_watermark_params valleyview_cursor_wm_info = {
 839         I965_CURSOR_FIFO,
 840         VALLEYVIEW_CURSOR_MAX_WM,
 841         I965_CURSOR_DFT_WM,
 842         2,
 843         G4X_FIFO_LINE_SIZE,
 844 };
 845 static const struct intel_watermark_params i965_cursor_wm_info = {
 846         I965_CURSOR_FIFO,
 847         I965_CURSOR_MAX_WM,
 848         I965_CURSOR_DFT_WM,
 849         2,
 850         I915_FIFO_LINE_SIZE,
 851 };
 852 static const struct intel_watermark_params i945_wm_info = {
 853         I945_FIFO_SIZE,
 854         I915_MAX_WM,
 855         1,
 856         2,
 857         I915_FIFO_LINE_SIZE
 858 };
 859 static const struct intel_watermark_params i915_wm_info = {
 860         I915_FIFO_SIZE,
 861         I915_MAX_WM,
 862         1,
 863         2,
 864         I915_FIFO_LINE_SIZE
 865 };
 866 static const struct intel_watermark_params i855_wm_info = {
 867         I855GM_FIFO_SIZE,
 868         I915_MAX_WM,
 869         1,
 870         2,
 871         I830_FIFO_LINE_SIZE
 872 };
 873 static const struct intel_watermark_params i830_wm_info = {
 874         I830_FIFO_SIZE,
 875         I915_MAX_WM,
 876         1,
 877         2,
 878         I830_FIFO_LINE_SIZE
 879 };
 880
 881 static const struct intel_watermark_params ironlake_display_wm_info = {
 882         ILK_DISPLAY_FIFO,
 883         ILK_DISPLAY_MAXWM,
 884         ILK_DISPLAY_DFTWM,
 885         2,
 886         ILK_FIFO_LINE_SIZE
 887 };
 888 static const struct intel_watermark_params ironlake_cursor_wm_info = {
 889         ILK_CURSOR_FIFO,
 890         ILK_CURSOR_MAXWM,
 891         ILK_CURSOR_DFTWM,
 892         2,
 893         ILK_FIFO_LINE_SIZE
 894 };
 895 static const struct intel_watermark_params ironlake_display_srwm_info = {
 896         ILK_DISPLAY_SR_FIFO,
 897         ILK_DISPLAY_MAX_SRWM,
 898         ILK_DISPLAY_DFT_SRWM,
 899         2,
 900         ILK_FIFO_LINE_SIZE
 901 };
 902 static const struct intel_watermark_params ironlake_cursor_srwm_info = {
 903         ILK_CURSOR_SR_FIFO,
 904         ILK_CURSOR_MAX_SRWM,
 905         ILK_CURSOR_DFT_SRWM,
 906         2,
 907         ILK_FIFO_LINE_SIZE
 908 };
 909
 910 static const struct intel_watermark_params sandybridge_display_wm_info = {
 911         SNB_DISPLAY_FIFO,
 912         SNB_DISPLAY_MAXWM,
 913         SNB_DISPLAY_DFTWM,
 914         2,
 915         SNB_FIFO_LINE_SIZE
 916 };
 917 static const struct intel_watermark_params sandybridge_cursor_wm_info = {
 918         SNB_CURSOR_FIFO,
 919         SNB_CURSOR_MAXWM,
 920         SNB_CURSOR_DFTWM,
 921         2,
 922         SNB_FIFO_LINE_SIZE
 923 };
 924 static const struct intel_watermark_params sandybridge_display_srwm_info = {
 925         SNB_DISPLAY_SR_FIFO,
 926         SNB_DISPLAY_MAX_SRWM,
 927         SNB_DISPLAY_DFT_SRWM,
 928         2,
 929         SNB_FIFO_LINE_SIZE
 930 };
 931 static const struct intel_watermark_params sandybridge_cursor_srwm_info = {
 932         SNB_CURSOR_SR_FIFO,
 933         SNB_CURSOR_MAX_SRWM,
 934         SNB_CURSOR_DFT_SRWM,
 935         2,
 936         SNB_FIFO_LINE_SIZE
 937 };
 938
 939
 940 /**
 941  * intel_calculate_wm - calculate watermark level
 942  * @clock_in_khz: pixel clock
 943  * @wm: chip FIFO params
 944  * @pixel_size: display pixel size
 945  * @latency_ns: memory latency for the platform
 946  *
 947  * Calculate the watermark level (the level at which the display plane will
 948  * start fetching from memory again).  Each chip has a different display
 949  * FIFO size and allocation, so the caller needs to figure that out and pass
 950  * in the correct intel_watermark_params structure.
 951  *
 952  * As the pixel clock runs, the FIFO will be drained at a rate that depends
 953  * on the pixel size.  When it reaches the watermark level, it'll start
 954  * fetching FIFO line sized based chunks from memory until the FIFO fills
 955  * past the watermark point.  If the FIFO drains completely, a FIFO underrun
 956  * will occur, and a display engine hang could result.
 957  */
 958 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
 959                                         const struct intel_watermark_params *wm,
 960                                         int fifo_size,
 961                                         int pixel_size,
 962                                         unsigned long latency_ns)
 963 {
 964         long entries_required, wm_size;
 965
 966         /*
 967          * Note: we need to make sure we don't overflow for various clock &
 968          * latency values.
 969          * clocks go from a few thousand to several hundred thousand.
 970          * latency is usually a few thousand
 971          */
 972         entries_required = ((clock_in_khz / 1000) * pixel_size * latency_ns) /
 973                 1000;
 974         entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
 975
 976         DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
 977
 978         wm_size = fifo_size - (entries_required + wm->guard_size);
 979
 980         DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
 981
 982         /* Don't promote wm_size to unsigned... */
 983         if (wm_size > (long)wm->max_wm)
 984                 wm_size = wm->max_wm;
 985         if (wm_size <= 0)
 986                 wm_size = wm->default_wm;
 987         return wm_size;
 988 }
 989
 990 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
 991 {
 992         struct drm_crtc *crtc, *enabled = NULL;
 993
 994         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
 995                 if (crtc->enabled && crtc->fb) {
 996                         if (enabled)
 997                                 return NULL;
 998                         enabled = crtc;
 999                 }
1000         }
1001
1002         return enabled;
1003 }
1004
1005 static void pineview_update_wm(struct drm_device *dev)
1006 {
1007         struct drm_i915_private *dev_priv = dev->dev_private;
1008         struct drm_crtc *crtc;
1009         const struct cxsr_latency *latency;
1010         u32 reg;
1011         unsigned long wm;
1012
1013         latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
1014                                          dev_priv->fsb_freq, dev_priv->mem_freq);
1015         if (!latency) {
1016                 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
1017                 pineview_disable_cxsr(dev);
1018                 return;
1019         }
1020
1021         crtc = single_enabled_crtc(dev);
1022         if (crtc) {
1023                 int clock = crtc->mode.clock;
1024                 int pixel_size = crtc->fb->bits_per_pixel / 8;
1025
1026                 /* Display SR */
1027                 wm = intel_calculate_wm(clock, &pineview_display_wm,
1028                                         pineview_display_wm.fifo_size,
1029                                         pixel_size, latency->display_sr);
1030                 reg = I915_READ(DSPFW1);
1031                 reg &= ~DSPFW_SR_MASK;
1032                 reg |= wm << DSPFW_SR_SHIFT;
1033                 I915_WRITE(DSPFW1, reg);
1034                 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
1035
1036                 /* cursor SR */
1037                 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
1038                                         pineview_display_wm.fifo_size,
1039                                         pixel_size, latency->cursor_sr);
1040                 reg = I915_READ(DSPFW3);
1041                 reg &= ~DSPFW_CURSOR_SR_MASK;
1042                 reg |= (wm & 0x3f) << DSPFW_CURSOR_SR_SHIFT;
1043                 I915_WRITE(DSPFW3, reg);
1044
1045                 /* Display HPLL off SR */
1046                 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
1047                                         pineview_display_hplloff_wm.fifo_size,
1048                                         pixel_size, latency->display_hpll_disable);
1049                 reg = I915_READ(DSPFW3);
1050                 reg &= ~DSPFW_HPLL_SR_MASK;
1051                 reg |= wm & DSPFW_HPLL_SR_MASK;
1052                 I915_WRITE(DSPFW3, reg);
1053
1054                 /* cursor HPLL off SR */
1055                 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
1056                                         pineview_display_hplloff_wm.fifo_size,
1057                                         pixel_size, latency->cursor_hpll_disable);
1058                 reg = I915_READ(DSPFW3);
1059                 reg &= ~DSPFW_HPLL_CURSOR_MASK;
1060                 reg |= (wm & 0x3f) << DSPFW_HPLL_CURSOR_SHIFT;
1061                 I915_WRITE(DSPFW3, reg);
1062                 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
1063
1064                 /* activate cxsr */
1065                 I915_WRITE(DSPFW3,
1066                            I915_READ(DSPFW3) | PINEVIEW_SELF_REFRESH_EN);
1067                 DRM_DEBUG_KMS("Self-refresh is enabled\n");
1068         } else {
1069                 pineview_disable_cxsr(dev);
1070                 DRM_DEBUG_KMS("Self-refresh is disabled\n");
1071         }
1072 }
1073
1074 static bool g4x_compute_wm0(struct drm_device *dev,
1075                             int plane,
1076                             const struct intel_watermark_params *display,
1077                             int display_latency_ns,
1078                             const struct intel_watermark_params *cursor,
1079                             int cursor_latency_ns,
1080                             int *plane_wm,
1081                             int *cursor_wm)
1082 {
1083         struct drm_crtc *crtc;
1084         int htotal, hdisplay, clock, pixel_size;
1085         int line_time_us, line_count;
1086         int entries, tlb_miss;
1087
1088         crtc = intel_get_crtc_for_plane(dev, plane);
1089         if (crtc->fb == NULL || !crtc->enabled) {
1090                 *cursor_wm = cursor->guard_size;
1091                 *plane_wm = display->guard_size;
1092                 return false;
1093         }
1094
1095         htotal = crtc->mode.htotal;
1096         hdisplay = crtc->mode.hdisplay;
1097         clock = crtc->mode.clock;
1098         pixel_size = crtc->fb->bits_per_pixel / 8;
1099
1100         /* Use the small buffer method to calculate plane watermark */
1101         entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
1102         tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
1103         if (tlb_miss > 0)
1104                 entries += tlb_miss;
1105         entries = DIV_ROUND_UP(entries, display->cacheline_size);
1106         *plane_wm = entries + display->guard_size;
1107         if (*plane_wm > (int)display->max_wm)
1108                 *plane_wm = display->max_wm;
1109
1110         /* Use the large buffer method to calculate cursor watermark */
1111         line_time_us = ((htotal * 1000) / clock);
1112         line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
1113         entries = line_count * 64 * pixel_size;
1114         tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
1115         if (tlb_miss > 0)
1116                 entries += tlb_miss;
1117         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1118         *cursor_wm = entries + cursor->guard_size;
1119         if (*cursor_wm > (int)cursor->max_wm)
1120                 *cursor_wm = (int)cursor->max_wm;
1121
1122         return true;
1123 }
1124
1125 /*
1126  * Check the wm result.
1127  *
1128  * If any calculated watermark values is larger than the maximum value that
1129  * can be programmed into the associated watermark register, that watermark
1130  * must be disabled.
1131  */
1132 static bool g4x_check_srwm(struct drm_device *dev,
1133                            int display_wm, int cursor_wm,
1134                            const struct intel_watermark_params *display,
1135                            const struct intel_watermark_params *cursor)
1136 {
1137         DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
1138                       display_wm, cursor_wm);
1139
1140         if (display_wm > display->max_wm) {
1141                 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
1142                               display_wm, display->max_wm);
1143                 return false;
1144         }
1145
1146         if (cursor_wm > cursor->max_wm) {
1147                 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
1148                               cursor_wm, cursor->max_wm);
1149                 return false;
1150         }
1151
1152         if (!(display_wm || cursor_wm)) {
1153                 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
1154                 return false;
1155         }
1156
1157         return true;
1158 }
1159
1160 static bool g4x_compute_srwm(struct drm_device *dev,
1161                              int plane,
1162                              int latency_ns,
1163                              const struct intel_watermark_params *display,
1164                              const struct intel_watermark_params *cursor,
1165                              int *display_wm, int *cursor_wm)
1166 {
1167         struct drm_crtc *crtc;
1168         int hdisplay, htotal, pixel_size, clock;
1169         unsigned long line_time_us;
1170         int line_count, line_size;
1171         int small, large;
1172         int entries;
1173
1174         if (!latency_ns) {
1175                 *display_wm = *cursor_wm = 0;
1176                 return false;
1177         }
1178
1179         crtc = intel_get_crtc_for_plane(dev, plane);
1180         hdisplay = crtc->mode.hdisplay;
1181         htotal = crtc->mode.htotal;
1182         clock = crtc->mode.clock;
1183         pixel_size = crtc->fb->bits_per_pixel / 8;
1184
1185         line_time_us = (htotal * 1000) / clock;
1186         line_count = (latency_ns / line_time_us + 1000) / 1000;
1187         line_size = hdisplay * pixel_size;
1188
1189         /* Use the minimum of the small and large buffer method for primary */
1190         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1191         large = line_count * line_size;
1192
1193         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1194         *display_wm = entries + display->guard_size;
1195
1196         /* calculate the self-refresh watermark for display cursor */
1197         entries = line_count * pixel_size * 64;
1198         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1199         *cursor_wm = entries + cursor->guard_size;
1200
1201         return g4x_check_srwm(dev,
1202                               *display_wm, *cursor_wm,
1203                               display, cursor);
1204 }
1205
1206 static bool vlv_compute_drain_latency(struct drm_device *dev,
1207                                      int plane,
1208                                      int *plane_prec_mult,
1209                                      int *plane_dl,
1210                                      int *cursor_prec_mult,
1211                                      int *cursor_dl)
1212 {
1213         struct drm_crtc *crtc;
1214         int clock, pixel_size;
1215         int entries;
1216
1217         crtc = intel_get_crtc_for_plane(dev, plane);
1218         if (crtc->fb == NULL || !crtc->enabled)
1219                 return false;
1220
1221         clock = crtc->mode.clock;       /* VESA DOT Clock */
1222         pixel_size = crtc->fb->bits_per_pixel / 8;      /* BPP */
1223
1224         entries = (clock / 1000) * pixel_size;
1225         *plane_prec_mult = (entries > 256) ?
1226                 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1227         *plane_dl = (64 * (*plane_prec_mult) * 4) / ((clock / 1000) *
1228                                                      pixel_size);
1229
1230         entries = (clock / 1000) * 4;   /* BPP is always 4 for cursor */
1231         *cursor_prec_mult = (entries > 256) ?
1232                 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_16;
1233         *cursor_dl = (64 * (*cursor_prec_mult) * 4) / ((clock / 1000) * 4);
1234
1235         return true;
1236 }
1237
1238 /*
1239  * Update drain latency registers of memory arbiter
1240  *
1241  * Valleyview SoC has a new memory arbiter and needs drain latency registers
1242  * to be programmed. Each plane has a drain latency multiplier and a drain
1243  * latency value.
1244  */
1245
1246 static void vlv_update_drain_latency(struct drm_device *dev)
1247 {
1248         struct drm_i915_private *dev_priv = dev->dev_private;
1249         int planea_prec, planea_dl, planeb_prec, planeb_dl;
1250         int cursora_prec, cursora_dl, cursorb_prec, cursorb_dl;
1251         int plane_prec_mult, cursor_prec_mult; /* Precision multiplier is
1252                                                         either 16 or 32 */
1253
1254         /* For plane A, Cursor A */
1255         if (vlv_compute_drain_latency(dev, 0, &plane_prec_mult, &planea_dl,
1256                                       &cursor_prec_mult, &cursora_dl)) {
1257                 cursora_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1258                         DDL_CURSORA_PRECISION_32 : DDL_CURSORA_PRECISION_16;
1259                 planea_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1260                         DDL_PLANEA_PRECISION_32 : DDL_PLANEA_PRECISION_16;
1261
1262                 I915_WRITE(VLV_DDL1, cursora_prec |
1263                                 (cursora_dl << DDL_CURSORA_SHIFT) |
1264                                 planea_prec | planea_dl);
1265         }
1266
1267         /* For plane B, Cursor B */
1268         if (vlv_compute_drain_latency(dev, 1, &plane_prec_mult, &planeb_dl,
1269                                       &cursor_prec_mult, &cursorb_dl)) {
1270                 cursorb_prec = (cursor_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1271                         DDL_CURSORB_PRECISION_32 : DDL_CURSORB_PRECISION_16;
1272                 planeb_prec = (plane_prec_mult == DRAIN_LATENCY_PRECISION_32) ?
1273                         DDL_PLANEB_PRECISION_32 : DDL_PLANEB_PRECISION_16;
1274
1275                 I915_WRITE(VLV_DDL2, cursorb_prec |
1276                                 (cursorb_dl << DDL_CURSORB_SHIFT) |
1277                                 planeb_prec | planeb_dl);
1278         }
1279 }
1280
1281 #define single_plane_enabled(mask) is_power_of_2(mask)
1282
1283 static void valleyview_update_wm(struct drm_device *dev)
1284 {
1285         static const int sr_latency_ns = 12000;
1286         struct drm_i915_private *dev_priv = dev->dev_private;
1287         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1288         int plane_sr, cursor_sr;
1289         unsigned int enabled = 0;
1290
1291         vlv_update_drain_latency(dev);
1292
1293         if (g4x_compute_wm0(dev, 0,
1294                             &valleyview_wm_info, latency_ns,
1295                             &valleyview_cursor_wm_info, latency_ns,
1296                             &planea_wm, &cursora_wm))
1297                 enabled |= 1;
1298
1299         if (g4x_compute_wm0(dev, 1,
1300                             &valleyview_wm_info, latency_ns,
1301                             &valleyview_cursor_wm_info, latency_ns,
1302                             &planeb_wm, &cursorb_wm))
1303                 enabled |= 2;
1304
1305         plane_sr = cursor_sr = 0;
1306         if (single_plane_enabled(enabled) &&
1307             g4x_compute_srwm(dev, ffs(enabled) - 1,
1308                              sr_latency_ns,
1309                              &valleyview_wm_info,
1310                              &valleyview_cursor_wm_info,
1311                              &plane_sr, &cursor_sr))
1312                 I915_WRITE(FW_BLC_SELF_VLV, FW_CSPWRDWNEN);
1313         else
1314                 I915_WRITE(FW_BLC_SELF_VLV,
1315                            I915_READ(FW_BLC_SELF_VLV) & ~FW_CSPWRDWNEN);
1316
1317         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1318                       planea_wm, cursora_wm,
1319                       planeb_wm, cursorb_wm,
1320                       plane_sr, cursor_sr);
1321
1322         I915_WRITE(DSPFW1,
1323                    (plane_sr << DSPFW_SR_SHIFT) |
1324                    (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1325                    (planeb_wm << DSPFW_PLANEB_SHIFT) |
1326                    planea_wm);
1327         I915_WRITE(DSPFW2,
1328                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1329                    (cursora_wm << DSPFW_CURSORA_SHIFT));
1330         I915_WRITE(DSPFW3,
1331                    (I915_READ(DSPFW3) & ~DSPFW_CURSOR_SR_MASK) |
1332                    (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1333 }
1334
1335 static void g4x_update_wm(struct drm_device *dev)
1336 {
1337         static const int sr_latency_ns = 12000;
1338         struct drm_i915_private *dev_priv = dev->dev_private;
1339         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1340         int plane_sr, cursor_sr;
1341         unsigned int enabled = 0;
1342
1343         if (g4x_compute_wm0(dev, 0,
1344                             &g4x_wm_info, latency_ns,
1345                             &g4x_cursor_wm_info, latency_ns,
1346                             &planea_wm, &cursora_wm))
1347                 enabled |= 1;
1348
1349         if (g4x_compute_wm0(dev, 1,
1350                             &g4x_wm_info, latency_ns,
1351                             &g4x_cursor_wm_info, latency_ns,
1352                             &planeb_wm, &cursorb_wm))
1353                 enabled |= 2;
1354
1355         plane_sr = cursor_sr = 0;
1356         if (single_plane_enabled(enabled) &&
1357             g4x_compute_srwm(dev, ffs(enabled) - 1,
1358                              sr_latency_ns,
1359                              &g4x_wm_info,
1360                              &g4x_cursor_wm_info,
1361                              &plane_sr, &cursor_sr))
1362                 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1363         else
1364                 I915_WRITE(FW_BLC_SELF,
1365                            I915_READ(FW_BLC_SELF) & ~FW_BLC_SELF_EN);
1366
1367         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1368                       planea_wm, cursora_wm,
1369                       planeb_wm, cursorb_wm,
1370                       plane_sr, cursor_sr);
1371
1372         I915_WRITE(DSPFW1,
1373                    (plane_sr << DSPFW_SR_SHIFT) |
1374                    (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1375                    (planeb_wm << DSPFW_PLANEB_SHIFT) |
1376                    planea_wm);
1377         I915_WRITE(DSPFW2,
1378                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1379                    (cursora_wm << DSPFW_CURSORA_SHIFT));
1380         /* HPLL off in SR has some issues on G4x... disable it */
1381         I915_WRITE(DSPFW3,
1382                    (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1383                    (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1384 }
1385
1386 static void i965_update_wm(struct drm_device *dev)
1387 {
1388         struct drm_i915_private *dev_priv = dev->dev_private;
1389         struct drm_crtc *crtc;
1390         int srwm = 1;
1391         int cursor_sr = 16;
1392
1393         /* Calc sr entries for one plane configs */
1394         crtc = single_enabled_crtc(dev);
1395         if (crtc) {
1396                 /* self-refresh has much higher latency */
1397                 static const int sr_latency_ns = 12000;
1398                 int clock = crtc->mode.clock;
1399                 int htotal = crtc->mode.htotal;
1400                 int hdisplay = crtc->mode.hdisplay;
1401                 int pixel_size = crtc->fb->bits_per_pixel / 8;
1402                 unsigned long line_time_us;
1403                 int entries;
1404
1405                 line_time_us = ((htotal * 1000) / clock);
1406
1407                 /* Use ns/us then divide to preserve precision */
1408                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1409                         pixel_size * hdisplay;
1410                 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1411                 srwm = I965_FIFO_SIZE - entries;
1412                 if (srwm < 0)
1413                         srwm = 1;
1414                 srwm &= 0x1ff;
1415                 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1416                               entries, srwm);
1417
1418                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1419                         pixel_size * 64;
1420                 entries = DIV_ROUND_UP(entries,
1421                                           i965_cursor_wm_info.cacheline_size);
1422                 cursor_sr = i965_cursor_wm_info.fifo_size -
1423                         (entries + i965_cursor_wm_info.guard_size);
1424
1425                 if (cursor_sr > i965_cursor_wm_info.max_wm)
1426                         cursor_sr = i965_cursor_wm_info.max_wm;
1427
1428                 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1429                               "cursor %d\n", srwm, cursor_sr);
1430
1431                 if (IS_CRESTLINE(dev))
1432                         I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN);
1433         } else {
1434                 /* Turn off self refresh if both pipes are enabled */
1435                 if (IS_CRESTLINE(dev))
1436                         I915_WRITE(FW_BLC_SELF, I915_READ(FW_BLC_SELF)
1437                                    & ~FW_BLC_SELF_EN);
1438         }
1439
1440         DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1441                       srwm);
1442
1443         /* 965 has limitations... */
1444         I915_WRITE(DSPFW1, (srwm << DSPFW_SR_SHIFT) |
1445                    (8 << 16) | (8 << 8) | (8 << 0));
1446         I915_WRITE(DSPFW2, (8 << 8) | (8 << 0));
1447         /* update cursor SR watermark */
1448         I915_WRITE(DSPFW3, (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1449 }
1450
1451 static void i9xx_update_wm(struct drm_device *dev)
1452 {
1453         struct drm_i915_private *dev_priv = dev->dev_private;
1454         const struct intel_watermark_params *wm_info;
1455         uint32_t fwater_lo;
1456         uint32_t fwater_hi;
1457         int cwm, srwm = 1;
1458         int fifo_size;
1459         int planea_wm, planeb_wm;
1460         struct drm_crtc *crtc, *enabled = NULL;
1461
1462         if (IS_I945GM(dev))
1463                 wm_info = &i945_wm_info;
1464         else if (!IS_GEN2(dev))
1465                 wm_info = &i915_wm_info;
1466         else
1467                 wm_info = &i855_wm_info;
1468
1469         fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1470         crtc = intel_get_crtc_for_plane(dev, 0);
1471         if (crtc->enabled && crtc->fb) {
1472                 int cpp = crtc->fb->bits_per_pixel / 8;
1473                 if (IS_GEN2(dev))
1474                         cpp = 4;
1475
1476                 planea_wm = intel_calculate_wm(crtc->mode.clock,
1477                                                wm_info, fifo_size, cpp,
1478                                                latency_ns);
1479                 enabled = crtc;
1480         } else
1481                 planea_wm = fifo_size - wm_info->guard_size;
1482
1483         fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1484         crtc = intel_get_crtc_for_plane(dev, 1);
1485         if (crtc->enabled && crtc->fb) {
1486                 int cpp = crtc->fb->bits_per_pixel / 8;
1487                 if (IS_GEN2(dev))
1488                         cpp = 4;
1489
1490                 planeb_wm = intel_calculate_wm(crtc->mode.clock,
1491                                                wm_info, fifo_size, cpp,
1492                                                latency_ns);
1493                 if (enabled == NULL)
1494                         enabled = crtc;
1495                 else
1496                         enabled = NULL;
1497         } else
1498                 planeb_wm = fifo_size - wm_info->guard_size;
1499
1500         DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1501
1502         /*
1503          * Overlay gets an aggressive default since video jitter is bad.
1504          */
1505         cwm = 2;
1506
1507         /* Play safe and disable self-refresh before adjusting watermarks. */
1508         if (IS_I945G(dev) || IS_I945GM(dev))
1509                 I915_WRITE(FW_BLC_SELF, FW_BLC_SELF_EN_MASK | 0);
1510         else if (IS_I915GM(dev))
1511                 I915_WRITE(INSTPM, I915_READ(INSTPM) & ~INSTPM_SELF_EN);
1512
1513         /* Calc sr entries for one plane configs */
1514         if (HAS_FW_BLC(dev) && enabled) {
1515                 /* self-refresh has much higher latency */
1516                 static const int sr_latency_ns = 6000;
1517                 int clock = enabled->mode.clock;
1518                 int htotal = enabled->mode.htotal;
1519                 int hdisplay = enabled->mode.hdisplay;
1520                 int pixel_size = enabled->fb->bits_per_pixel / 8;
1521                 unsigned long line_time_us;
1522                 int entries;
1523
1524                 line_time_us = (htotal * 1000) / clock;
1525
1526                 /* Use ns/us then divide to preserve precision */
1527                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1528                         pixel_size * hdisplay;
1529                 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1530                 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1531                 srwm = wm_info->fifo_size - entries;
1532                 if (srwm < 0)
1533                         srwm = 1;
1534
1535                 if (IS_I945G(dev) || IS_I945GM(dev))
1536                         I915_WRITE(FW_BLC_SELF,
1537                                    FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1538                 else if (IS_I915GM(dev))
1539                         I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1540         }
1541
1542         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1543                       planea_wm, planeb_wm, cwm, srwm);
1544
1545         fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1546         fwater_hi = (cwm & 0x1f);
1547
1548         /* Set request length to 8 cachelines per fetch */
1549         fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1550         fwater_hi = fwater_hi | (1 << 8);
1551
1552         I915_WRITE(FW_BLC, fwater_lo);
1553         I915_WRITE(FW_BLC2, fwater_hi);
1554
1555         if (HAS_FW_BLC(dev)) {
1556                 if (enabled) {
1557                         if (IS_I945G(dev) || IS_I945GM(dev))
1558                                 I915_WRITE(FW_BLC_SELF,
1559                                            FW_BLC_SELF_EN_MASK | FW_BLC_SELF_EN);
1560                         else if (IS_I915GM(dev))
1561                                 I915_WRITE(INSTPM, I915_READ(INSTPM) | INSTPM_SELF_EN);
1562                         DRM_DEBUG_KMS("memory self refresh enabled\n");
1563                 } else
1564                         DRM_DEBUG_KMS("memory self refresh disabled\n");
1565         }
1566 }
1567
1568 static void i830_update_wm(struct drm_device *dev)
1569 {
1570         struct drm_i915_private *dev_priv = dev->dev_private;
1571         struct drm_crtc *crtc;
1572         uint32_t fwater_lo;
1573         int planea_wm;
1574
1575         crtc = single_enabled_crtc(dev);
1576         if (crtc == NULL)
1577                 return;
1578
1579         planea_wm = intel_calculate_wm(crtc->mode.clock, &i830_wm_info,
1580                                        dev_priv->display.get_fifo_size(dev, 0),
1581                                        4, latency_ns);
1582         fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1583         fwater_lo |= (3<<8) | planea_wm;
1584
1585         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1586
1587         I915_WRITE(FW_BLC, fwater_lo);
1588 }
1589
1590 #define ILK_LP0_PLANE_LATENCY           700
1591 #define ILK_LP0_CURSOR_LATENCY          1300
1592
1593 /*
1594  * Check the wm result.
1595  *
1596  * If any calculated watermark values is larger than the maximum value that
1597  * can be programmed into the associated watermark register, that watermark
1598  * must be disabled.
1599  */
1600 static bool ironlake_check_srwm(struct drm_device *dev, int level,
1601                                 int fbc_wm, int display_wm, int cursor_wm,
1602                                 const struct intel_watermark_params *display,
1603                                 const struct intel_watermark_params *cursor)
1604 {
1605         struct drm_i915_private *dev_priv = dev->dev_private;
1606
1607         DRM_DEBUG_KMS("watermark %d: display plane %d, fbc lines %d,"
1608                       " cursor %d\n", level, display_wm, fbc_wm, cursor_wm);
1609
1610         if (fbc_wm > SNB_FBC_MAX_SRWM) {
1611                 DRM_DEBUG_KMS("fbc watermark(%d) is too large(%d), disabling wm%d+\n",
1612                               fbc_wm, SNB_FBC_MAX_SRWM, level);
1613
1614                 /* fbc has it's own way to disable FBC WM */
1615                 I915_WRITE(DISP_ARB_CTL,
1616                            I915_READ(DISP_ARB_CTL) | DISP_FBC_WM_DIS);
1617                 return false;
1618         }
1619
1620         if (display_wm > display->max_wm) {
1621                 DRM_DEBUG_KMS("display watermark(%d) is too large(%d), disabling wm%d+\n",
1622                               display_wm, SNB_DISPLAY_MAX_SRWM, level);
1623                 return false;
1624         }
1625
1626         if (cursor_wm > cursor->max_wm) {
1627                 DRM_DEBUG_KMS("cursor watermark(%d) is too large(%d), disabling wm%d+\n",
1628                               cursor_wm, SNB_CURSOR_MAX_SRWM, level);
1629                 return false;
1630         }
1631
1632         if (!(fbc_wm || display_wm || cursor_wm)) {
1633                 DRM_DEBUG_KMS("latency %d is 0, disabling wm%d+\n", level, level);
1634                 return false;
1635         }
1636
1637         return true;
1638 }
1639
1640 /*
1641  * Compute watermark values of WM[1-3],
1642  */
1643 static bool ironlake_compute_srwm(struct drm_device *dev, int level, int plane,
1644                                   int latency_ns,
1645                                   const struct intel_watermark_params *display,
1646                                   const struct intel_watermark_params *cursor,
1647                                   int *fbc_wm, int *display_wm, int *cursor_wm)
1648 {
1649         struct drm_crtc *crtc;
1650         unsigned long line_time_us;
1651         int hdisplay, htotal, pixel_size, clock;
1652         int line_count, line_size;
1653         int small, large;
1654         int entries;
1655
1656         if (!latency_ns) {
1657                 *fbc_wm = *display_wm = *cursor_wm = 0;
1658                 return false;
1659         }
1660
1661         crtc = intel_get_crtc_for_plane(dev, plane);
1662         hdisplay = crtc->mode.hdisplay;
1663         htotal = crtc->mode.htotal;
1664         clock = crtc->mode.clock;
1665         pixel_size = crtc->fb->bits_per_pixel / 8;
1666
1667         line_time_us = (htotal * 1000) / clock;
1668         line_count = (latency_ns / line_time_us + 1000) / 1000;
1669         line_size = hdisplay * pixel_size;
1670
1671         /* Use the minimum of the small and large buffer method for primary */
1672         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1673         large = line_count * line_size;
1674
1675         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1676         *display_wm = entries + display->guard_size;
1677
1678         /*
1679          * Spec says:
1680          * FBC WM = ((Final Primary WM * 64) / number of bytes per line) + 2
1681          */
1682         *fbc_wm = DIV_ROUND_UP(*display_wm * 64, line_size) + 2;
1683
1684         /* calculate the self-refresh watermark for display cursor */
1685         entries = line_count * pixel_size * 64;
1686         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
1687         *cursor_wm = entries + cursor->guard_size;
1688
1689         return ironlake_check_srwm(dev, level,
1690                                    *fbc_wm, *display_wm, *cursor_wm,
1691                                    display, cursor);
1692 }
1693
1694 static void ironlake_update_wm(struct drm_device *dev)
1695 {
1696         struct drm_i915_private *dev_priv = dev->dev_private;
1697         int fbc_wm, plane_wm, cursor_wm;
1698         unsigned int enabled;
1699
1700         enabled = 0;
1701         if (g4x_compute_wm0(dev, 0,
1702                             &ironlake_display_wm_info,
1703                             ILK_LP0_PLANE_LATENCY,
1704                             &ironlake_cursor_wm_info,
1705                             ILK_LP0_CURSOR_LATENCY,
1706                             &plane_wm, &cursor_wm)) {
1707                 I915_WRITE(WM0_PIPEA_ILK,
1708                            (plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm);
1709                 DRM_DEBUG_KMS("FIFO watermarks For pipe A -"
1710                               " plane %d, " "cursor: %d\n",
1711                               plane_wm, cursor_wm);
1712                 enabled |= 1;
1713         }
1714
1715         if (g4x_compute_wm0(dev, 1,
1716                             &ironlake_display_wm_info,
1717                             ILK_LP0_PLANE_LATENCY,
1718                             &ironlake_cursor_wm_info,
1719                             ILK_LP0_CURSOR_LATENCY,
1720                             &plane_wm, &cursor_wm)) {
1721                 I915_WRITE(WM0_PIPEB_ILK,
1722                            (plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm);
1723                 DRM_DEBUG_KMS("FIFO watermarks For pipe B -"
1724                               " plane %d, cursor: %d\n",
1725                               plane_wm, cursor_wm);
1726                 enabled |= 2;
1727         }
1728
1729         /*
1730          * Calculate and update the self-refresh watermark only when one
1731          * display plane is used.
1732          */
1733         I915_WRITE(WM3_LP_ILK, 0);
1734         I915_WRITE(WM2_LP_ILK, 0);
1735         I915_WRITE(WM1_LP_ILK, 0);
1736
1737         if (!single_plane_enabled(enabled))
1738                 return;
1739         enabled = ffs(enabled) - 1;
1740
1741         /* WM1 */
1742         if (!ironlake_compute_srwm(dev, 1, enabled,
1743                                    ILK_READ_WM1_LATENCY() * 500,
1744                                    &ironlake_display_srwm_info,
1745                                    &ironlake_cursor_srwm_info,
1746                                    &fbc_wm, &plane_wm, &cursor_wm))
1747                 return;
1748
1749         I915_WRITE(WM1_LP_ILK,
1750                    WM1_LP_SR_EN |
1751                    (ILK_READ_WM1_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1752                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1753                    (plane_wm << WM1_LP_SR_SHIFT) |
1754                    cursor_wm);
1755
1756         /* WM2 */
1757         if (!ironlake_compute_srwm(dev, 2, enabled,
1758                                    ILK_READ_WM2_LATENCY() * 500,
1759                                    &ironlake_display_srwm_info,
1760                                    &ironlake_cursor_srwm_info,
1761                                    &fbc_wm, &plane_wm, &cursor_wm))
1762                 return;
1763
1764         I915_WRITE(WM2_LP_ILK,
1765                    WM2_LP_EN |
1766                    (ILK_READ_WM2_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1767                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1768                    (plane_wm << WM1_LP_SR_SHIFT) |
1769                    cursor_wm);
1770
1771         /*
1772          * WM3 is unsupported on ILK, probably because we don't have latency
1773          * data for that power state
1774          */
1775 }
1776
1777 static void sandybridge_update_wm(struct drm_device *dev)
1778 {
1779         struct drm_i915_private *dev_priv = dev->dev_private;
1780         int latency = SNB_READ_WM0_LATENCY() * 100;     /* In unit 0.1us */
1781         u32 val;
1782         int fbc_wm, plane_wm, cursor_wm;
1783         unsigned int enabled;
1784
1785         enabled = 0;
1786         if (g4x_compute_wm0(dev, 0,
1787                             &sandybridge_display_wm_info, latency,
1788                             &sandybridge_cursor_wm_info, latency,
1789                             &plane_wm, &cursor_wm)) {
1790                 val = I915_READ(WM0_PIPEA_ILK);
1791                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
1792                 I915_WRITE(WM0_PIPEA_ILK, val |
1793                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
1794                 DRM_DEBUG_KMS("FIFO watermarks For pipe A -"
1795                               " plane %d, " "cursor: %d\n",
1796                               plane_wm, cursor_wm);
1797                 enabled |= 1;
1798         }
1799
1800         if (g4x_compute_wm0(dev, 1,
1801                             &sandybridge_display_wm_info, latency,
1802                             &sandybridge_cursor_wm_info, latency,
1803                             &plane_wm, &cursor_wm)) {
1804                 val = I915_READ(WM0_PIPEB_ILK);
1805                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
1806                 I915_WRITE(WM0_PIPEB_ILK, val |
1807                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
1808                 DRM_DEBUG_KMS("FIFO watermarks For pipe B -"
1809                               " plane %d, cursor: %d\n",
1810                               plane_wm, cursor_wm);
1811                 enabled |= 2;
1812         }
1813
1814         if ((dev_priv->num_pipe == 3) &&
1815             g4x_compute_wm0(dev, 2,
1816                             &sandybridge_display_wm_info, latency,
1817                             &sandybridge_cursor_wm_info, latency,
1818                             &plane_wm, &cursor_wm)) {
1819                 val = I915_READ(WM0_PIPEC_IVB);
1820                 val &= ~(WM0_PIPE_PLANE_MASK | WM0_PIPE_CURSOR_MASK);
1821                 I915_WRITE(WM0_PIPEC_IVB, val |
1822                            ((plane_wm << WM0_PIPE_PLANE_SHIFT) | cursor_wm));
1823                 DRM_DEBUG_KMS("FIFO watermarks For pipe C -"
1824                               " plane %d, cursor: %d\n",
1825                               plane_wm, cursor_wm);
1826                 enabled |= 3;
1827         }
1828
1829         /*
1830          * Calculate and update the self-refresh watermark only when one
1831          * display plane is used.
1832          *
1833          * SNB support 3 levels of watermark.
1834          *
1835          * WM1/WM2/WM2 watermarks have to be enabled in the ascending order,
1836          * and disabled in the descending order
1837          *
1838          */
1839         I915_WRITE(WM3_LP_ILK, 0);
1840         I915_WRITE(WM2_LP_ILK, 0);
1841         I915_WRITE(WM1_LP_ILK, 0);
1842
1843         if (!single_plane_enabled(enabled) ||
1844             dev_priv->sprite_scaling_enabled)
1845                 return;
1846         enabled = ffs(enabled) - 1;
1847
1848         /* WM1 */
1849         if (!ironlake_compute_srwm(dev, 1, enabled,
1850                                    SNB_READ_WM1_LATENCY() * 500,
1851                                    &sandybridge_display_srwm_info,
1852                                    &sandybridge_cursor_srwm_info,
1853                                    &fbc_wm, &plane_wm, &cursor_wm))
1854                 return;
1855
1856         I915_WRITE(WM1_LP_ILK,
1857                    WM1_LP_SR_EN |
1858                    (SNB_READ_WM1_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1859                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1860                    (plane_wm << WM1_LP_SR_SHIFT) |
1861                    cursor_wm);
1862
1863         /* WM2 */
1864         if (!ironlake_compute_srwm(dev, 2, enabled,
1865                                    SNB_READ_WM2_LATENCY() * 500,
1866                                    &sandybridge_display_srwm_info,
1867                                    &sandybridge_cursor_srwm_info,
1868                                    &fbc_wm, &plane_wm, &cursor_wm))
1869                 return;
1870
1871         I915_WRITE(WM2_LP_ILK,
1872                    WM2_LP_EN |
1873                    (SNB_READ_WM2_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1874                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1875                    (plane_wm << WM1_LP_SR_SHIFT) |
1876                    cursor_wm);
1877
1878         /* WM3 */
1879         if (!ironlake_compute_srwm(dev, 3, enabled,
1880                                    SNB_READ_WM3_LATENCY() * 500,
1881                                    &sandybridge_display_srwm_info,
1882                                    &sandybridge_cursor_srwm_info,
1883                                    &fbc_wm, &plane_wm, &cursor_wm))
1884                 return;
1885
1886         I915_WRITE(WM3_LP_ILK,
1887                    WM3_LP_EN |
1888                    (SNB_READ_WM3_LATENCY() << WM1_LP_LATENCY_SHIFT) |
1889                    (fbc_wm << WM1_LP_FBC_SHIFT) |
1890                    (plane_wm << WM1_LP_SR_SHIFT) |
1891                    cursor_wm);
1892 }
1893
1894 static void
1895 haswell_update_linetime_wm(struct drm_device *dev, int pipe,
1896                                  struct drm_display_mode *mode)
1897 {
1898         struct drm_i915_private *dev_priv = dev->dev_private;
1899         u32 temp;
1900
1901         temp = I915_READ(PIPE_WM_LINETIME(pipe));
1902         temp &= ~PIPE_WM_LINETIME_MASK;
1903
1904         /* The WM are computed with base on how long it takes to fill a single
1905          * row at the given clock rate, multiplied by 8.
1906          * */
1907         temp |= PIPE_WM_LINETIME_TIME(
1908                 ((mode->crtc_hdisplay * 1000) / mode->clock) * 8);
1909
1910         /* IPS watermarks are only used by pipe A, and are ignored by
1911          * pipes B and C.  They are calculated similarly to the common
1912          * linetime values, except that we are using CD clock frequency
1913          * in MHz instead of pixel rate for the division.
1914          *
1915          * This is a placeholder for the IPS watermark calculation code.
1916          */
1917
1918         I915_WRITE(PIPE_WM_LINETIME(pipe), temp);
1919 }
1920
1921 static bool
1922 sandybridge_compute_sprite_wm(struct drm_device *dev, int plane,
1923                               uint32_t sprite_width, int pixel_size,
1924                               const struct intel_watermark_params *display,
1925                               int display_latency_ns, int *sprite_wm)
1926 {
1927         struct drm_crtc *crtc;
1928         int clock;
1929         int entries, tlb_miss;
1930
1931         crtc = intel_get_crtc_for_plane(dev, plane);
1932         if (crtc->fb == NULL || !crtc->enabled) {
1933                 *sprite_wm = display->guard_size;
1934                 return false;
1935         }
1936
1937         clock = crtc->mode.clock;
1938
1939         /* Use the small buffer method to calculate the sprite watermark */
1940         entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
1941         tlb_miss = display->fifo_size*display->cacheline_size -
1942                 sprite_width * 8;
1943         if (tlb_miss > 0)
1944                 entries += tlb_miss;
1945         entries = DIV_ROUND_UP(entries, display->cacheline_size);
1946         *sprite_wm = entries + display->guard_size;
1947         if (*sprite_wm > (int)display->max_wm)
1948                 *sprite_wm = display->max_wm;
1949
1950         return true;
1951 }
1952
1953 static bool
1954 sandybridge_compute_sprite_srwm(struct drm_device *dev, int plane,
1955                                 uint32_t sprite_width, int pixel_size,
1956                                 const struct intel_watermark_params *display,
1957                                 int latency_ns, int *sprite_wm)
1958 {
1959         struct drm_crtc *crtc;
1960         unsigned long line_time_us;
1961         int clock;
1962         int line_count, line_size;
1963         int small, large;
1964         int entries;
1965
1966         if (!latency_ns) {
1967                 *sprite_wm = 0;
1968                 return false;
1969         }
1970
1971         crtc = intel_get_crtc_for_plane(dev, plane);
1972         clock = crtc->mode.clock;
1973         if (!clock) {
1974                 *sprite_wm = 0;
1975                 return false;
1976         }
1977
1978         line_time_us = (sprite_width * 1000) / clock;
1979         if (!line_time_us) {
1980                 *sprite_wm = 0;
1981                 return false;
1982         }
1983
1984         line_count = (latency_ns / line_time_us + 1000) / 1000;
1985         line_size = sprite_width * pixel_size;
1986
1987         /* Use the minimum of the small and large buffer method for primary */
1988         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
1989         large = line_count * line_size;
1990
1991         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
1992         *sprite_wm = entries + display->guard_size;
1993
1994         return *sprite_wm > 0x3ff ? false : true;
1995 }
1996
1997 static void sandybridge_update_sprite_wm(struct drm_device *dev, int pipe,
1998                                          uint32_t sprite_width, int pixel_size)
1999 {
2000         struct drm_i915_private *dev_priv = dev->dev_private;
2001         int latency = SNB_READ_WM0_LATENCY() * 100;     /* In unit 0.1us */
2002         u32 val;
2003         int sprite_wm, reg;
2004         int ret;
2005
2006         switch (pipe) {
2007         case 0:
2008                 reg = WM0_PIPEA_ILK;
2009                 break;
2010         case 1:
2011                 reg = WM0_PIPEB_ILK;
2012                 break;
2013         case 2:
2014                 reg = WM0_PIPEC_IVB;
2015                 break;
2016         default:
2017                 return; /* bad pipe */
2018         }
2019
2020         ret = sandybridge_compute_sprite_wm(dev, pipe, sprite_width, pixel_size,
2021                                             &sandybridge_display_wm_info,
2022                                             latency, &sprite_wm);
2023         if (!ret) {
2024                 DRM_DEBUG_KMS("failed to compute sprite wm for pipe %d\n",
2025                               pipe);
2026                 return;
2027         }
2028
2029         val = I915_READ(reg);
2030         val &= ~WM0_PIPE_SPRITE_MASK;
2031         I915_WRITE(reg, val | (sprite_wm << WM0_PIPE_SPRITE_SHIFT));
2032         DRM_DEBUG_KMS("sprite watermarks For pipe %d - %d\n", pipe, sprite_wm);
2033
2034
2035         ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
2036                                               pixel_size,
2037                                               &sandybridge_display_srwm_info,
2038                                               SNB_READ_WM1_LATENCY() * 500,
2039                                               &sprite_wm);
2040         if (!ret) {
2041                 DRM_DEBUG_KMS("failed to compute sprite lp1 wm on pipe %d\n",
2042                               pipe);
2043                 return;
2044         }
2045         I915_WRITE(WM1S_LP_ILK, sprite_wm);
2046
2047         /* Only IVB has two more LP watermarks for sprite */
2048         if (!IS_IVYBRIDGE(dev))
2049                 return;
2050
2051         ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
2052                                               pixel_size,
2053                                               &sandybridge_display_srwm_info,
2054                                               SNB_READ_WM2_LATENCY() * 500,
2055                                               &sprite_wm);
2056         if (!ret) {
2057                 DRM_DEBUG_KMS("failed to compute sprite lp2 wm on pipe %d\n",
2058                               pipe);
2059                 return;
2060         }
2061         I915_WRITE(WM2S_LP_IVB, sprite_wm);
2062
2063         ret = sandybridge_compute_sprite_srwm(dev, pipe, sprite_width,
2064                                               pixel_size,
2065                                               &sandybridge_display_srwm_info,
2066                                               SNB_READ_WM3_LATENCY() * 500,
2067                                               &sprite_wm);
2068         if (!ret) {
2069                 DRM_DEBUG_KMS("failed to compute sprite lp3 wm on pipe %d\n",
2070                               pipe);
2071                 return;
2072         }
2073         I915_WRITE(WM3S_LP_IVB, sprite_wm);
2074 }
2075
2076 /**
2077  * intel_update_watermarks - update FIFO watermark values based on current modes
2078  *
2079  * Calculate watermark values for the various WM regs based on current mode
2080  * and plane configuration.
2081  *
2082  * There are several cases to deal with here:
2083  *   - normal (i.e. non-self-refresh)
2084  *   - self-refresh (SR) mode
2085  *   - lines are large relative to FIFO size (buffer can hold up to 2)
2086  *   - lines are small relative to FIFO size (buffer can hold more than 2
2087  *     lines), so need to account for TLB latency
2088  *
2089  *   The normal calculation is:
2090  *     watermark = dotclock * bytes per pixel * latency
2091  *   where latency is platform & configuration dependent (we assume pessimal
2092  *   values here).
2093  *
2094  *   The SR calculation is:
2095  *     watermark = (trunc(latency/line time)+1) * surface width *
2096  *       bytes per pixel
2097  *   where
2098  *     line time = htotal / dotclock
2099  *     surface width = hdisplay for normal plane and 64 for cursor
2100  *   and latency is assumed to be high, as above.
2101  *
2102  * The final value programmed to the register should always be rounded up,
2103  * and include an extra 2 entries to account for clock crossings.
2104  *
2105  * We don't use the sprite, so we can ignore that.  And on Crestline we have
2106  * to set the non-SR watermarks to 8.
2107  */
2108 void intel_update_watermarks(struct drm_device *dev)
2109 {
2110         struct drm_i915_private *dev_priv = dev->dev_private;
2111
2112         if (dev_priv->display.update_wm)
2113                 dev_priv->display.update_wm(dev);
2114 }
2115
2116 void intel_update_linetime_watermarks(struct drm_device *dev,
2117                 int pipe, struct drm_display_mode *mode)
2118 {
2119         struct drm_i915_private *dev_priv = dev->dev_private;
2120
2121         if (dev_priv->display.update_linetime_wm)
2122                 dev_priv->display.update_linetime_wm(dev, pipe, mode);
2123 }
2124
2125 void intel_update_sprite_watermarks(struct drm_device *dev, int pipe,
2126                                     uint32_t sprite_width, int pixel_size)
2127 {
2128         struct drm_i915_private *dev_priv = dev->dev_private;
2129
2130         if (dev_priv->display.update_sprite_wm)
2131                 dev_priv->display.update_sprite_wm(dev, pipe, sprite_width,
2132                                                    pixel_size);
2133 }
2134
2135 static struct drm_i915_gem_object *
2136 intel_alloc_context_page(struct drm_device *dev)
2137 {
2138         struct drm_i915_gem_object *ctx;
2139         int ret;
2140
2141         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2142
2143         ctx = i915_gem_alloc_object(dev, 4096);
2144         if (!ctx) {
2145                 DRM_DEBUG("failed to alloc power context, RC6 disabled\n");
2146                 return NULL;
2147         }
2148
2149         ret = i915_gem_object_pin(ctx, 4096, true, false);
2150         if (ret) {
2151                 DRM_ERROR("failed to pin power context: %d\n", ret);
2152                 goto err_unref;
2153         }
2154
2155         ret = i915_gem_object_set_to_gtt_domain(ctx, 1);
2156         if (ret) {
2157                 DRM_ERROR("failed to set-domain on power context: %d\n", ret);
2158                 goto err_unpin;
2159         }
2160
2161         return ctx;
2162
2163 err_unpin:
2164         i915_gem_object_unpin(ctx);
2165 err_unref:
2166         drm_gem_object_unreference(&ctx->base);
2167         mutex_unlock(&dev->struct_mutex);
2168         return NULL;
2169 }
2170
2171 /**
2172  * Lock protecting IPS related data structures
2173  */
2174 DEFINE_SPINLOCK(mchdev_lock);
2175
2176 /* Global for IPS driver to get at the current i915 device. Protected by
2177  * mchdev_lock. */
2178 static struct drm_i915_private *i915_mch_dev;
2179
2180 bool ironlake_set_drps(struct drm_device *dev, u8 val)
2181 {
2182         struct drm_i915_private *dev_priv = dev->dev_private;
2183         u16 rgvswctl;
2184
2185         assert_spin_locked(&mchdev_lock);
2186
2187         rgvswctl = I915_READ16(MEMSWCTL);
2188         if (rgvswctl & MEMCTL_CMD_STS) {
2189                 DRM_DEBUG("gpu busy, RCS change rejected\n");
2190                 return false; /* still busy with another command */
2191         }
2192
2193         rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
2194                 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
2195         I915_WRITE16(MEMSWCTL, rgvswctl);
2196         POSTING_READ16(MEMSWCTL);
2197
2198         rgvswctl |= MEMCTL_CMD_STS;
2199         I915_WRITE16(MEMSWCTL, rgvswctl);
2200
2201         return true;
2202 }
2203
2204 static void ironlake_enable_drps(struct drm_device *dev)
2205 {
2206         struct drm_i915_private *dev_priv = dev->dev_private;
2207         u32 rgvmodectl = I915_READ(MEMMODECTL);
2208         u8 fmax, fmin, fstart, vstart;
2209
2210         spin_lock_irq(&mchdev_lock);
2211
2212         /* Enable temp reporting */
2213         I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
2214         I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
2215
2216         /* 100ms RC evaluation intervals */
2217         I915_WRITE(RCUPEI, 100000);
2218         I915_WRITE(RCDNEI, 100000);
2219
2220         /* Set max/min thresholds to 90ms and 80ms respectively */
2221         I915_WRITE(RCBMAXAVG, 90000);
2222         I915_WRITE(RCBMINAVG, 80000);
2223
2224         I915_WRITE(MEMIHYST, 1);
2225
2226         /* Set up min, max, and cur for interrupt handling */
2227         fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
2228         fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
2229         fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
2230                 MEMMODE_FSTART_SHIFT;
2231
2232         vstart = (I915_READ(PXVFREQ_BASE + (fstart * 4)) & PXVFREQ_PX_MASK) >>
2233                 PXVFREQ_PX_SHIFT;
2234
2235         dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
2236         dev_priv->ips.fstart = fstart;
2237
2238         dev_priv->ips.max_delay = fstart;
2239         dev_priv->ips.min_delay = fmin;
2240         dev_priv->ips.cur_delay = fstart;
2241
2242         DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
2243                          fmax, fmin, fstart);
2244
2245         I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
2246
2247         /*
2248          * Interrupts will be enabled in ironlake_irq_postinstall
2249          */
2250
2251         I915_WRITE(VIDSTART, vstart);
2252         POSTING_READ(VIDSTART);
2253
2254         rgvmodectl |= MEMMODE_SWMODE_EN;
2255         I915_WRITE(MEMMODECTL, rgvmodectl);
2256
2257         if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
2258                 DRM_ERROR("stuck trying to change perf mode\n");
2259         mdelay(1);
2260
2261         ironlake_set_drps(dev, fstart);
2262
2263         dev_priv->ips.last_count1 = I915_READ(0x112e4) + I915_READ(0x112e8) +
2264                 I915_READ(0x112e0);
2265         dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
2266         dev_priv->ips.last_count2 = I915_READ(0x112f4);
2267         getrawmonotonic(&dev_priv->ips.last_time2);
2268
2269         spin_unlock_irq(&mchdev_lock);
2270 }
2271
2272 static void ironlake_disable_drps(struct drm_device *dev)
2273 {
2274         struct drm_i915_private *dev_priv = dev->dev_private;
2275         u16 rgvswctl;
2276
2277         spin_lock_irq(&mchdev_lock);
2278
2279         rgvswctl = I915_READ16(MEMSWCTL);
2280
2281         /* Ack interrupts, disable EFC interrupt */
2282         I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
2283         I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
2284         I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
2285         I915_WRITE(DEIIR, DE_PCU_EVENT);
2286         I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
2287
2288         /* Go back to the starting frequency */
2289         ironlake_set_drps(dev, dev_priv->ips.fstart);
2290         mdelay(1);
2291         rgvswctl |= MEMCTL_CMD_STS;
2292         I915_WRITE(MEMSWCTL, rgvswctl);
2293         mdelay(1);
2294
2295         spin_unlock_irq(&mchdev_lock);
2296 }
2297
2298 /* There's a funny hw issue where the hw returns all 0 when reading from
2299  * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
2300  * ourselves, instead of doing a rmw cycle (which might result in us clearing
2301  * all limits and the gpu stuck at whatever frequency it is at atm).
2302  */
2303 static u32 gen6_rps_limits(struct drm_i915_private *dev_priv, u8 *val)
2304 {
2305         u32 limits;
2306
2307         limits = 0;
2308
2309         if (*val >= dev_priv->rps.max_delay)
2310                 *val = dev_priv->rps.max_delay;
2311         limits |= dev_priv->rps.max_delay << 24;
2312
2313         /* Only set the down limit when we've reached the lowest level to avoid
2314          * getting more interrupts, otherwise leave this clear. This prevents a
2315          * race in the hw when coming out of rc6: There's a tiny window where
2316          * the hw runs at the minimal clock before selecting the desired
2317          * frequency, if the down threshold expires in that window we will not
2318          * receive a down interrupt. */
2319         if (*val <= dev_priv->rps.min_delay) {
2320                 *val = dev_priv->rps.min_delay;
2321                 limits |= dev_priv->rps.min_delay << 16;
2322         }
2323
2324         return limits;
2325 }
2326
2327 void gen6_set_rps(struct drm_device *dev, u8 val)
2328 {
2329         struct drm_i915_private *dev_priv = dev->dev_private;
2330         u32 limits = gen6_rps_limits(dev_priv, &val);
2331
2332         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
2333         WARN_ON(val > dev_priv->rps.max_delay);
2334         WARN_ON(val < dev_priv->rps.min_delay);
2335
2336         if (val == dev_priv->rps.cur_delay)
2337                 return;
2338
2339         I915_WRITE(GEN6_RPNSWREQ,
2340                    GEN6_FREQUENCY(val) |
2341                    GEN6_OFFSET(0) |
2342                    GEN6_AGGRESSIVE_TURBO);
2343
2344         /* Make sure we continue to get interrupts
2345          * until we hit the minimum or maximum frequencies.
2346          */
2347         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, limits);
2348
2349         POSTING_READ(GEN6_RPNSWREQ);
2350
2351         dev_priv->rps.cur_delay = val;
2352
2353         trace_intel_gpu_freq_change(val * 50);
2354 }
2355
2356 static void gen6_disable_rps(struct drm_device *dev)
2357 {
2358         struct drm_i915_private *dev_priv = dev->dev_private;
2359
2360         I915_WRITE(GEN6_RC_CONTROL, 0);
2361         I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
2362         I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
2363         I915_WRITE(GEN6_PMIER, 0);
2364         /* Complete PM interrupt masking here doesn't race with the rps work
2365          * item again unmasking PM interrupts because that is using a different
2366          * register (PMIMR) to mask PM interrupts. The only risk is in leaving
2367          * stale bits in PMIIR and PMIMR which gen6_enable_rps will clean up. */
2368
2369         spin_lock_irq(&dev_priv->rps.lock);
2370         dev_priv->rps.pm_iir = 0;
2371         spin_unlock_irq(&dev_priv->rps.lock);
2372
2373         I915_WRITE(GEN6_PMIIR, I915_READ(GEN6_PMIIR));
2374 }
2375
2376 int intel_enable_rc6(const struct drm_device *dev)
2377 {
2378         /* Respect the kernel parameter if it is set */
2379         if (i915_enable_rc6 >= 0)
2380                 return i915_enable_rc6;
2381
2382         /* Disable RC6 on Ironlake */
2383         if (INTEL_INFO(dev)->gen == 5)
2384                 return 0;
2385
2386         if (IS_HASWELL(dev)) {
2387                 DRM_DEBUG_DRIVER("Haswell: only RC6 available\n");
2388                 return INTEL_RC6_ENABLE;
2389         }
2390
2391         /* snb/ivb have more than one rc6 state. */
2392         if (INTEL_INFO(dev)->gen == 6) {
2393                 DRM_DEBUG_DRIVER("Sandybridge: deep RC6 disabled\n");
2394                 return INTEL_RC6_ENABLE;
2395         }
2396
2397         DRM_DEBUG_DRIVER("RC6 and deep RC6 enabled\n");
2398         return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
2399 }
2400
2401 static void gen6_enable_rps(struct drm_device *dev)
2402 {
2403         struct drm_i915_private *dev_priv = dev->dev_private;
2404         struct intel_ring_buffer *ring;
2405         u32 rp_state_cap;
2406         u32 gt_perf_status;
2407         u32 rc6vids, pcu_mbox, rc6_mask = 0;
2408         u32 gtfifodbg;
2409         int rc6_mode;
2410         int i, ret;
2411
2412         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
2413
2414         /* Here begins a magic sequence of register writes to enable
2415          * auto-downclocking.
2416          *
2417          * Perhaps there might be some value in exposing these to
2418          * userspace...
2419          */
2420         I915_WRITE(GEN6_RC_STATE, 0);
2421
2422         /* Clear the DBG now so we don't confuse earlier errors */
2423         if ((gtfifodbg = I915_READ(GTFIFODBG))) {
2424                 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
2425                 I915_WRITE(GTFIFODBG, gtfifodbg);
2426         }
2427
2428         gen6_gt_force_wake_get(dev_priv);
2429
2430         rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
2431         gt_perf_status = I915_READ(GEN6_GT_PERF_STATUS);
2432
2433         /* In units of 100MHz */
2434         dev_priv->rps.max_delay = rp_state_cap & 0xff;
2435         dev_priv->rps.min_delay = (rp_state_cap & 0xff0000) >> 16;
2436         dev_priv->rps.cur_delay = 0;
2437
2438         /* disable the counters and set deterministic thresholds */
2439         I915_WRITE(GEN6_RC_CONTROL, 0);
2440
2441         I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
2442         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
2443         I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
2444         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
2445         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
2446
2447         for_each_ring(ring, dev_priv, i)
2448                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
2449
2450         I915_WRITE(GEN6_RC_SLEEP, 0);
2451         I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
2452         I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
2453         I915_WRITE(GEN6_RC6p_THRESHOLD, 100000);
2454         I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
2455
2456         /* Check if we are enabling RC6 */
2457         rc6_mode = intel_enable_rc6(dev_priv->dev);
2458         if (rc6_mode & INTEL_RC6_ENABLE)
2459                 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
2460
2461         /* We don't use those on Haswell */
2462         if (!IS_HASWELL(dev)) {
2463                 if (rc6_mode & INTEL_RC6p_ENABLE)
2464                         rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
2465
2466                 if (rc6_mode & INTEL_RC6pp_ENABLE)
2467                         rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
2468         }
2469
2470         DRM_INFO("Enabling RC6 states: RC6 %s, RC6p %s, RC6pp %s\n",
2471                         (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off",
2472                         (rc6_mask & GEN6_RC_CTL_RC6p_ENABLE) ? "on" : "off",
2473                         (rc6_mask & GEN6_RC_CTL_RC6pp_ENABLE) ? "on" : "off");
2474
2475         I915_WRITE(GEN6_RC_CONTROL,
2476                    rc6_mask |
2477                    GEN6_RC_CTL_EI_MODE(1) |
2478                    GEN6_RC_CTL_HW_ENABLE);
2479
2480         I915_WRITE(GEN6_RPNSWREQ,
2481                    GEN6_FREQUENCY(10) |
2482                    GEN6_OFFSET(0) |
2483                    GEN6_AGGRESSIVE_TURBO);
2484         I915_WRITE(GEN6_RC_VIDEO_FREQ,
2485                    GEN6_FREQUENCY(12));
2486
2487         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
2488         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
2489                    dev_priv->rps.max_delay << 24 |
2490                    dev_priv->rps.min_delay << 16);
2491
2492         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
2493         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
2494         I915_WRITE(GEN6_RP_UP_EI, 66000);
2495         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
2496
2497         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
2498         I915_WRITE(GEN6_RP_CONTROL,
2499                    GEN6_RP_MEDIA_TURBO |
2500                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
2501                    GEN6_RP_MEDIA_IS_GFX |
2502                    GEN6_RP_ENABLE |
2503                    GEN6_RP_UP_BUSY_AVG |
2504                    (IS_HASWELL(dev) ? GEN7_RP_DOWN_IDLE_AVG : GEN6_RP_DOWN_IDLE_CONT));
2505
2506         ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_MIN_FREQ_TABLE, 0);
2507         if (!ret) {
2508                 pcu_mbox = 0;
2509                 ret = sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &pcu_mbox);
2510                 if (ret && pcu_mbox & (1<<31)) { /* OC supported */
2511                         dev_priv->rps.max_delay = pcu_mbox & 0xff;
2512                         DRM_DEBUG_DRIVER("overclocking supported, adjusting frequency max to %dMHz\n", pcu_mbox * 50);
2513                 }
2514         } else {
2515                 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
2516         }
2517
2518         gen6_set_rps(dev_priv->dev, (gt_perf_status & 0xff00) >> 8);
2519
2520         /* requires MSI enabled */
2521         I915_WRITE(GEN6_PMIER, GEN6_PM_DEFERRED_EVENTS);
2522         spin_lock_irq(&dev_priv->rps.lock);
2523         WARN_ON(dev_priv->rps.pm_iir != 0);
2524         I915_WRITE(GEN6_PMIMR, 0);
2525         spin_unlock_irq(&dev_priv->rps.lock);
2526         /* enable all PM interrupts */
2527         I915_WRITE(GEN6_PMINTRMSK, 0);
2528
2529         rc6vids = 0;
2530         ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
2531         if (IS_GEN6(dev) && ret) {
2532                 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
2533         } else if (IS_GEN6(dev) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
2534                 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
2535                           GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
2536                 rc6vids &= 0xffff00;
2537                 rc6vids |= GEN6_ENCODE_RC6_VID(450);
2538                 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
2539                 if (ret)
2540                         DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
2541         }
2542
2543         gen6_gt_force_wake_put(dev_priv);
2544 }
2545
2546 static void gen6_update_ring_freq(struct drm_device *dev)
2547 {
2548         struct drm_i915_private *dev_priv = dev->dev_private;
2549         int min_freq = 15;
2550         int gpu_freq;
2551         unsigned int ia_freq, max_ia_freq;
2552         int scaling_factor = 180;
2553
2554         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
2555
2556         max_ia_freq = cpufreq_quick_get_max(0);
2557         /*
2558          * Default to measured freq if none found, PCU will ensure we don't go
2559          * over
2560          */
2561         if (!max_ia_freq)
2562                 max_ia_freq = tsc_khz;
2563
2564         /* Convert from kHz to MHz */
2565         max_ia_freq /= 1000;
2566
2567         /*
2568          * For each potential GPU frequency, load a ring frequency we'd like
2569          * to use for memory access.  We do this by specifying the IA frequency
2570          * the PCU should use as a reference to determine the ring frequency.
2571          */
2572         for (gpu_freq = dev_priv->rps.max_delay; gpu_freq >= dev_priv->rps.min_delay;
2573              gpu_freq--) {
2574                 int diff = dev_priv->rps.max_delay - gpu_freq;
2575
2576                 /*
2577                  * For GPU frequencies less than 750MHz, just use the lowest
2578                  * ring freq.
2579                  */
2580                 if (gpu_freq < min_freq)
2581                         ia_freq = 800;
2582                 else
2583                         ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
2584                 ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
2585                 ia_freq <<= GEN6_PCODE_FREQ_IA_RATIO_SHIFT;
2586
2587                 sandybridge_pcode_write(dev_priv,
2588                                         GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
2589                                         ia_freq | gpu_freq);
2590         }
2591 }
2592
2593 void ironlake_teardown_rc6(struct drm_device *dev)
2594 {
2595         struct drm_i915_private *dev_priv = dev->dev_private;
2596
2597         if (dev_priv->ips.renderctx) {
2598                 i915_gem_object_unpin(dev_priv->ips.renderctx);
2599                 drm_gem_object_unreference(&dev_priv->ips.renderctx->base);
2600                 dev_priv->ips.renderctx = NULL;
2601         }
2602
2603         if (dev_priv->ips.pwrctx) {
2604                 i915_gem_object_unpin(dev_priv->ips.pwrctx);
2605                 drm_gem_object_unreference(&dev_priv->ips.pwrctx->base);
2606                 dev_priv->ips.pwrctx = NULL;
2607         }
2608 }
2609
2610 static void ironlake_disable_rc6(struct drm_device *dev)
2611 {
2612         struct drm_i915_private *dev_priv = dev->dev_private;
2613
2614         if (I915_READ(PWRCTXA)) {
2615                 /* Wake the GPU, prevent RC6, then restore RSTDBYCTL */
2616                 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) | RCX_SW_EXIT);
2617                 wait_for(((I915_READ(RSTDBYCTL) & RSX_STATUS_MASK) == RSX_STATUS_ON),
2618                          50);
2619
2620                 I915_WRITE(PWRCTXA, 0);
2621                 POSTING_READ(PWRCTXA);
2622
2623                 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
2624                 POSTING_READ(RSTDBYCTL);
2625         }
2626 }
2627
2628 static int ironlake_setup_rc6(struct drm_device *dev)
2629 {
2630         struct drm_i915_private *dev_priv = dev->dev_private;
2631
2632         if (dev_priv->ips.renderctx == NULL)
2633                 dev_priv->ips.renderctx = intel_alloc_context_page(dev);
2634         if (!dev_priv->ips.renderctx)
2635                 return -ENOMEM;
2636
2637         if (dev_priv->ips.pwrctx == NULL)
2638                 dev_priv->ips.pwrctx = intel_alloc_context_page(dev);
2639         if (!dev_priv->ips.pwrctx) {
2640                 ironlake_teardown_rc6(dev);
2641                 return -ENOMEM;
2642         }
2643
2644         return 0;
2645 }
2646
2647 static void ironlake_enable_rc6(struct drm_device *dev)
2648 {
2649         struct drm_i915_private *dev_priv = dev->dev_private;
2650         struct intel_ring_buffer *ring = &dev_priv->ring[RCS];
2651         bool was_interruptible;
2652         int ret;
2653
2654         /* rc6 disabled by default due to repeated reports of hanging during
2655          * boot and resume.
2656          */
2657         if (!intel_enable_rc6(dev))
2658                 return;
2659
2660         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
2661
2662         ret = ironlake_setup_rc6(dev);
2663         if (ret)
2664                 return;
2665
2666         was_interruptible = dev_priv->mm.interruptible;
2667         dev_priv->mm.interruptible = false;
2668
2669         /*
2670          * GPU can automatically power down the render unit if given a page
2671          * to save state.
2672          */
2673         ret = intel_ring_begin(ring, 6);
2674         if (ret) {
2675                 ironlake_teardown_rc6(dev);
2676                 dev_priv->mm.interruptible = was_interruptible;
2677                 return;
2678         }
2679
2680         intel_ring_emit(ring, MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN);
2681         intel_ring_emit(ring, MI_SET_CONTEXT);
2682         intel_ring_emit(ring, dev_priv->ips.renderctx->gtt_offset |
2683                         MI_MM_SPACE_GTT |
2684                         MI_SAVE_EXT_STATE_EN |
2685                         MI_RESTORE_EXT_STATE_EN |
2686                         MI_RESTORE_INHIBIT);
2687         intel_ring_emit(ring, MI_SUSPEND_FLUSH);
2688         intel_ring_emit(ring, MI_NOOP);
2689         intel_ring_emit(ring, MI_FLUSH);
2690         intel_ring_advance(ring);
2691
2692         /*
2693          * Wait for the command parser to advance past MI_SET_CONTEXT. The HW
2694          * does an implicit flush, combined with MI_FLUSH above, it should be
2695          * safe to assume that renderctx is valid
2696          */
2697         ret = intel_ring_idle(ring);
2698         dev_priv->mm.interruptible = was_interruptible;
2699         if (ret) {
2700                 DRM_ERROR("failed to enable ironlake power power savings\n");
2701                 ironlake_teardown_rc6(dev);
2702                 return;
2703         }
2704
2705         I915_WRITE(PWRCTXA, dev_priv->ips.pwrctx->gtt_offset | PWRCTX_EN);
2706         I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
2707 }
2708
2709 static unsigned long intel_pxfreq(u32 vidfreq)
2710 {
2711         unsigned long freq;
2712         int div = (vidfreq & 0x3f0000) >> 16;
2713         int post = (vidfreq & 0x3000) >> 12;
2714         int pre = (vidfreq & 0x7);
2715
2716         if (!pre)
2717                 return 0;
2718
2719         freq = ((div * 133333) / ((1<<post) * pre));
2720
2721         return freq;
2722 }
2723
2724 static const struct cparams {
2725         u16 i;
2726         u16 t;
2727         u16 m;
2728         u16 c;
2729 } cparams[] = {
2730         { 1, 1333, 301, 28664 },
2731         { 1, 1066, 294, 24460 },
2732         { 1, 800, 294, 25192 },
2733         { 0, 1333, 276, 27605 },
2734         { 0, 1066, 276, 27605 },
2735         { 0, 800, 231, 23784 },
2736 };
2737
2738 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
2739 {
2740         u64 total_count, diff, ret;
2741         u32 count1, count2, count3, m = 0, c = 0;
2742         unsigned long now = jiffies_to_msecs(jiffies), diff1;
2743         int i;
2744
2745         assert_spin_locked(&mchdev_lock);
2746
2747         diff1 = now - dev_priv->ips.last_time1;
2748
2749         /* Prevent division-by-zero if we are asking too fast.
2750          * Also, we don't get interesting results if we are polling
2751          * faster than once in 10ms, so just return the saved value
2752          * in such cases.
2753          */
2754         if (diff1 <= 10)
2755                 return dev_priv->ips.chipset_power;
2756
2757         count1 = I915_READ(DMIEC);
2758         count2 = I915_READ(DDREC);
2759         count3 = I915_READ(CSIEC);
2760
2761         total_count = count1 + count2 + count3;
2762
2763         /* FIXME: handle per-counter overflow */
2764         if (total_count < dev_priv->ips.last_count1) {
2765                 diff = ~0UL - dev_priv->ips.last_count1;
2766                 diff += total_count;
2767         } else {
2768                 diff = total_count - dev_priv->ips.last_count1;
2769         }
2770
2771         for (i = 0; i < ARRAY_SIZE(cparams); i++) {
2772                 if (cparams[i].i == dev_priv->ips.c_m &&
2773                     cparams[i].t == dev_priv->ips.r_t) {
2774                         m = cparams[i].m;
2775                         c = cparams[i].c;
2776                         break;
2777                 }
2778         }
2779
2780         diff = div_u64(diff, diff1);
2781         ret = ((m * diff) + c);
2782         ret = div_u64(ret, 10);
2783
2784         dev_priv->ips.last_count1 = total_count;
2785         dev_priv->ips.last_time1 = now;
2786
2787         dev_priv->ips.chipset_power = ret;
2788
2789         return ret;
2790 }
2791
2792 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
2793 {
2794         unsigned long val;
2795
2796         if (dev_priv->info->gen != 5)
2797                 return 0;
2798
2799         spin_lock_irq(&mchdev_lock);
2800
2801         val = __i915_chipset_val(dev_priv);
2802
2803         spin_unlock_irq(&mchdev_lock);
2804
2805         return val;
2806 }
2807
2808 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
2809 {
2810         unsigned long m, x, b;
2811         u32 tsfs;
2812
2813         tsfs = I915_READ(TSFS);
2814
2815         m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
2816         x = I915_READ8(TR1);
2817
2818         b = tsfs & TSFS_INTR_MASK;
2819
2820         return ((m * x) / 127) - b;
2821 }
2822
2823 static u16 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
2824 {
2825         static const struct v_table {
2826                 u16 vd; /* in .1 mil */
2827                 u16 vm; /* in .1 mil */
2828         } v_table[] = {
2829                 { 0, 0, },
2830                 { 375, 0, },
2831                 { 500, 0, },
2832                 { 625, 0, },
2833                 { 750, 0, },
2834                 { 875, 0, },
2835                 { 1000, 0, },
2836                 { 1125, 0, },
2837                 { 4125, 3000, },
2838                 { 4125, 3000, },
2839                 { 4125, 3000, },
2840                 { 4125, 3000, },
2841                 { 4125, 3000, },
2842                 { 4125, 3000, },
2843                 { 4125, 3000, },
2844                 { 4125, 3000, },
2845                 { 4125, 3000, },
2846                 { 4125, 3000, },
2847                 { 4125, 3000, },
2848                 { 4125, 3000, },
2849                 { 4125, 3000, },
2850                 { 4125, 3000, },
2851                 { 4125, 3000, },
2852                 { 4125, 3000, },
2853                 { 4125, 3000, },
2854                 { 4125, 3000, },
2855                 { 4125, 3000, },
2856                 { 4125, 3000, },
2857                 { 4125, 3000, },
2858                 { 4125, 3000, },
2859                 { 4125, 3000, },
2860                 { 4125, 3000, },
2861                 { 4250, 3125, },
2862                 { 4375, 3250, },
2863                 { 4500, 3375, },
2864                 { 4625, 3500, },
2865                 { 4750, 3625, },
2866                 { 4875, 3750, },
2867                 { 5000, 3875, },
2868                 { 5125, 4000, },
2869                 { 5250, 4125, },
2870                 { 5375, 4250, },
2871                 { 5500, 4375, },
2872                 { 5625, 4500, },
2873                 { 5750, 4625, },
2874                 { 5875, 4750, },
2875                 { 6000, 4875, },
2876                 { 6125, 5000, },
2877                 { 6250, 5125, },
2878                 { 6375, 5250, },
2879                 { 6500, 5375, },
2880                 { 6625, 5500, },
2881                 { 6750, 5625, },
2882                 { 6875, 5750, },
2883                 { 7000, 5875, },
2884                 { 7125, 6000, },
2885                 { 7250, 6125, },
2886                 { 7375, 6250, },
2887                 { 7500, 6375, },
2888                 { 7625, 6500, },
2889                 { 7750, 6625, },
2890                 { 7875, 6750, },
2891                 { 8000, 6875, },
2892                 { 8125, 7000, },
2893                 { 8250, 7125, },
2894                 { 8375, 7250, },
2895                 { 8500, 7375, },
2896                 { 8625, 7500, },
2897                 { 8750, 7625, },
2898                 { 8875, 7750, },
2899                 { 9000, 7875, },
2900                 { 9125, 8000, },
2901                 { 9250, 8125, },
2902                 { 9375, 8250, },
2903                 { 9500, 8375, },
2904                 { 9625, 8500, },
2905                 { 9750, 8625, },
2906                 { 9875, 8750, },
2907                 { 10000, 8875, },
2908                 { 10125, 9000, },
2909                 { 10250, 9125, },
2910                 { 10375, 9250, },
2911                 { 10500, 9375, },
2912                 { 10625, 9500, },
2913                 { 10750, 9625, },
2914                 { 10875, 9750, },
2915                 { 11000, 9875, },
2916                 { 11125, 10000, },
2917                 { 11250, 10125, },
2918                 { 11375, 10250, },
2919                 { 11500, 10375, },
2920                 { 11625, 10500, },
2921                 { 11750, 10625, },
2922                 { 11875, 10750, },
2923                 { 12000, 10875, },
2924                 { 12125, 11000, },
2925                 { 12250, 11125, },
2926                 { 12375, 11250, },
2927                 { 12500, 11375, },
2928                 { 12625, 11500, },
2929                 { 12750, 11625, },
2930                 { 12875, 11750, },
2931                 { 13000, 11875, },
2932                 { 13125, 12000, },
2933                 { 13250, 12125, },
2934                 { 13375, 12250, },
2935                 { 13500, 12375, },
2936                 { 13625, 12500, },
2937                 { 13750, 12625, },
2938                 { 13875, 12750, },
2939                 { 14000, 12875, },
2940                 { 14125, 13000, },
2941                 { 14250, 13125, },
2942                 { 14375, 13250, },
2943                 { 14500, 13375, },
2944                 { 14625, 13500, },
2945                 { 14750, 13625, },
2946                 { 14875, 13750, },
2947                 { 15000, 13875, },
2948                 { 15125, 14000, },
2949                 { 15250, 14125, },
2950                 { 15375, 14250, },
2951                 { 15500, 14375, },
2952                 { 15625, 14500, },
2953                 { 15750, 14625, },
2954                 { 15875, 14750, },
2955                 { 16000, 14875, },
2956                 { 16125, 15000, },
2957         };
2958         if (dev_priv->info->is_mobile)
2959                 return v_table[pxvid].vm;
2960         else
2961                 return v_table[pxvid].vd;
2962 }
2963
2964 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
2965 {
2966         struct timespec now, diff1;
2967         u64 diff;
2968         unsigned long diffms;
2969         u32 count;
2970
2971         assert_spin_locked(&mchdev_lock);
2972
2973         getrawmonotonic(&now);
2974         diff1 = timespec_sub(now, dev_priv->ips.last_time2);
2975
2976         /* Don't divide by 0 */
2977         diffms = diff1.tv_sec * 1000 + diff1.tv_nsec / 1000000;
2978         if (!diffms)
2979                 return;
2980
2981         count = I915_READ(GFXEC);
2982
2983         if (count < dev_priv->ips.last_count2) {
2984                 diff = ~0UL - dev_priv->ips.last_count2;
2985                 diff += count;
2986         } else {
2987                 diff = count - dev_priv->ips.last_count2;
2988         }
2989
2990         dev_priv->ips.last_count2 = count;
2991         dev_priv->ips.last_time2 = now;
2992
2993         /* More magic constants... */
2994         diff = diff * 1181;
2995         diff = div_u64(diff, diffms * 10);
2996         dev_priv->ips.gfx_power = diff;
2997 }
2998
2999 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
3000 {
3001         if (dev_priv->info->gen != 5)
3002                 return;
3003
3004         spin_lock_irq(&mchdev_lock);
3005
3006         __i915_update_gfx_val(dev_priv);
3007
3008         spin_unlock_irq(&mchdev_lock);
3009 }
3010
3011 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
3012 {
3013         unsigned long t, corr, state1, corr2, state2;
3014         u32 pxvid, ext_v;
3015
3016         assert_spin_locked(&mchdev_lock);
3017
3018         pxvid = I915_READ(PXVFREQ_BASE + (dev_priv->rps.cur_delay * 4));
3019         pxvid = (pxvid >> 24) & 0x7f;
3020         ext_v = pvid_to_extvid(dev_priv, pxvid);
3021
3022         state1 = ext_v;
3023
3024         t = i915_mch_val(dev_priv);
3025
3026         /* Revel in the empirically derived constants */
3027
3028         /* Correction factor in 1/100000 units */
3029         if (t > 80)
3030                 corr = ((t * 2349) + 135940);
3031         else if (t >= 50)
3032                 corr = ((t * 964) + 29317);
3033         else /* < 50 */
3034                 corr = ((t * 301) + 1004);
3035
3036         corr = corr * ((150142 * state1) / 10000 - 78642);
3037         corr /= 100000;
3038         corr2 = (corr * dev_priv->ips.corr);
3039
3040         state2 = (corr2 * state1) / 10000;
3041         state2 /= 100; /* convert to mW */
3042
3043         __i915_update_gfx_val(dev_priv);
3044
3045         return dev_priv->ips.gfx_power + state2;
3046 }
3047
3048 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
3049 {
3050         unsigned long val;
3051
3052         if (dev_priv->info->gen != 5)
3053                 return 0;
3054
3055         spin_lock_irq(&mchdev_lock);
3056
3057         val = __i915_gfx_val(dev_priv);
3058
3059         spin_unlock_irq(&mchdev_lock);
3060
3061         return val;
3062 }
3063
3064 /**
3065  * i915_read_mch_val - return value for IPS use
3066  *
3067  * Calculate and return a value for the IPS driver to use when deciding whether
3068  * we have thermal and power headroom to increase CPU or GPU power budget.
3069  */
3070 unsigned long i915_read_mch_val(void)
3071 {
3072         struct drm_i915_private *dev_priv;
3073         unsigned long chipset_val, graphics_val, ret = 0;
3074
3075         spin_lock_irq(&mchdev_lock);
3076         if (!i915_mch_dev)
3077                 goto out_unlock;
3078         dev_priv = i915_mch_dev;
3079
3080         chipset_val = __i915_chipset_val(dev_priv);
3081         graphics_val = __i915_gfx_val(dev_priv);
3082
3083         ret = chipset_val + graphics_val;
3084
3085 out_unlock:
3086         spin_unlock_irq(&mchdev_lock);
3087
3088         return ret;
3089 }
3090 EXPORT_SYMBOL_GPL(i915_read_mch_val);
3091
3092 /**
3093  * i915_gpu_raise - raise GPU frequency limit
3094  *
3095  * Raise the limit; IPS indicates we have thermal headroom.
3096  */
3097 bool i915_gpu_raise(void)
3098 {
3099         struct drm_i915_private *dev_priv;
3100         bool ret = true;
3101
3102         spin_lock_irq(&mchdev_lock);
3103         if (!i915_mch_dev) {
3104                 ret = false;
3105                 goto out_unlock;
3106         }
3107         dev_priv = i915_mch_dev;
3108
3109         if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
3110                 dev_priv->ips.max_delay--;
3111
3112 out_unlock:
3113         spin_unlock_irq(&mchdev_lock);
3114
3115         return ret;
3116 }
3117 EXPORT_SYMBOL_GPL(i915_gpu_raise);
3118
3119 /**
3120  * i915_gpu_lower - lower GPU frequency limit
3121  *
3122  * IPS indicates we're close to a thermal limit, so throttle back the GPU
3123  * frequency maximum.
3124  */
3125 bool i915_gpu_lower(void)
3126 {
3127         struct drm_i915_private *dev_priv;
3128         bool ret = true;
3129
3130         spin_lock_irq(&mchdev_lock);
3131         if (!i915_mch_dev) {
3132                 ret = false;
3133                 goto out_unlock;
3134         }
3135         dev_priv = i915_mch_dev;
3136
3137         if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
3138                 dev_priv->ips.max_delay++;
3139
3140 out_unlock:
3141         spin_unlock_irq(&mchdev_lock);
3142
3143         return ret;
3144 }
3145 EXPORT_SYMBOL_GPL(i915_gpu_lower);
3146
3147 /**
3148  * i915_gpu_busy - indicate GPU business to IPS
3149  *
3150  * Tell the IPS driver whether or not the GPU is busy.
3151  */
3152 bool i915_gpu_busy(void)
3153 {
3154         struct drm_i915_private *dev_priv;
3155         struct intel_ring_buffer *ring;
3156         bool ret = false;
3157         int i;
3158
3159         spin_lock_irq(&mchdev_lock);
3160         if (!i915_mch_dev)
3161                 goto out_unlock;
3162         dev_priv = i915_mch_dev;
3163
3164         for_each_ring(ring, dev_priv, i)
3165                 ret |= !list_empty(&ring->request_list);
3166
3167 out_unlock:
3168         spin_unlock_irq(&mchdev_lock);
3169
3170         return ret;
3171 }
3172 EXPORT_SYMBOL_GPL(i915_gpu_busy);
3173
3174 /**
3175  * i915_gpu_turbo_disable - disable graphics turbo
3176  *
3177  * Disable graphics turbo by resetting the max frequency and setting the
3178  * current frequency to the default.
3179  */
3180 bool i915_gpu_turbo_disable(void)
3181 {
3182         struct drm_i915_private *dev_priv;
3183         bool ret = true;
3184
3185         spin_lock_irq(&mchdev_lock);
3186         if (!i915_mch_dev) {
3187                 ret = false;
3188                 goto out_unlock;
3189         }
3190         dev_priv = i915_mch_dev;
3191
3192         dev_priv->ips.max_delay = dev_priv->ips.fstart;
3193
3194         if (!ironlake_set_drps(dev_priv->dev, dev_priv->ips.fstart))
3195                 ret = false;
3196
3197 out_unlock:
3198         spin_unlock_irq(&mchdev_lock);
3199
3200         return ret;
3201 }
3202 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
3203
3204 /**
3205  * Tells the intel_ips driver that the i915 driver is now loaded, if
3206  * IPS got loaded first.
3207  *
3208  * This awkward dance is so that neither module has to depend on the
3209  * other in order for IPS to do the appropriate communication of
3210  * GPU turbo limits to i915.
3211  */
3212 static void
3213 ips_ping_for_i915_load(void)
3214 {
3215         void (*link)(void);
3216
3217         link = symbol_get(ips_link_to_i915_driver);
3218         if (link) {
3219                 link();
3220                 symbol_put(ips_link_to_i915_driver);
3221         }
3222 }
3223
3224 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
3225 {
3226         /* We only register the i915 ips part with intel-ips once everything is
3227          * set up, to avoid intel-ips sneaking in and reading bogus values. */
3228         spin_lock_irq(&mchdev_lock);
3229         i915_mch_dev = dev_priv;
3230         spin_unlock_irq(&mchdev_lock);
3231
3232         ips_ping_for_i915_load();
3233 }
3234
3235 void intel_gpu_ips_teardown(void)
3236 {
3237         spin_lock_irq(&mchdev_lock);
3238         i915_mch_dev = NULL;
3239         spin_unlock_irq(&mchdev_lock);
3240 }
3241 static void intel_init_emon(struct drm_device *dev)
3242 {
3243         struct drm_i915_private *dev_priv = dev->dev_private;
3244         u32 lcfuse;
3245         u8 pxw[16];
3246         int i;
3247
3248         /* Disable to program */
3249         I915_WRITE(ECR, 0);
3250         POSTING_READ(ECR);
3251
3252         /* Program energy weights for various events */
3253         I915_WRITE(SDEW, 0x15040d00);
3254         I915_WRITE(CSIEW0, 0x007f0000);
3255         I915_WRITE(CSIEW1, 0x1e220004);
3256         I915_WRITE(CSIEW2, 0x04000004);
3257
3258         for (i = 0; i < 5; i++)
3259                 I915_WRITE(PEW + (i * 4), 0);
3260         for (i = 0; i < 3; i++)
3261                 I915_WRITE(DEW + (i * 4), 0);
3262
3263         /* Program P-state weights to account for frequency power adjustment */
3264         for (i = 0; i < 16; i++) {
3265                 u32 pxvidfreq = I915_READ(PXVFREQ_BASE + (i * 4));
3266                 unsigned long freq = intel_pxfreq(pxvidfreq);
3267                 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
3268                         PXVFREQ_PX_SHIFT;
3269                 unsigned long val;
3270
3271                 val = vid * vid;
3272                 val *= (freq / 1000);
3273                 val *= 255;
3274                 val /= (127*127*900);
3275                 if (val > 0xff)
3276                         DRM_ERROR("bad pxval: %ld\n", val);
3277                 pxw[i] = val;
3278         }
3279         /* Render standby states get 0 weight */
3280         pxw[14] = 0;
3281         pxw[15] = 0;
3282
3283         for (i = 0; i < 4; i++) {
3284                 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
3285                         (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
3286                 I915_WRITE(PXW + (i * 4), val);
3287         }
3288
3289         /* Adjust magic regs to magic values (more experimental results) */
3290         I915_WRITE(OGW0, 0);
3291         I915_WRITE(OGW1, 0);
3292         I915_WRITE(EG0, 0x00007f00);
3293         I915_WRITE(EG1, 0x0000000e);
3294         I915_WRITE(EG2, 0x000e0000);
3295         I915_WRITE(EG3, 0x68000300);
3296         I915_WRITE(EG4, 0x42000000);
3297         I915_WRITE(EG5, 0x00140031);
3298         I915_WRITE(EG6, 0);
3299         I915_WRITE(EG7, 0);
3300
3301         for (i = 0; i < 8; i++)
3302                 I915_WRITE(PXWL + (i * 4), 0);
3303
3304         /* Enable PMON + select events */
3305         I915_WRITE(ECR, 0x80000019);
3306
3307         lcfuse = I915_READ(LCFUSE02);
3308
3309         dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
3310 }
3311
3312 void intel_disable_gt_powersave(struct drm_device *dev)
3313 {
3314         struct drm_i915_private *dev_priv = dev->dev_private;
3315
3316         if (IS_IRONLAKE_M(dev)) {
3317                 ironlake_disable_drps(dev);
3318                 ironlake_disable_rc6(dev);
3319         } else if (INTEL_INFO(dev)->gen >= 6 && !IS_VALLEYVIEW(dev)) {
3320                 cancel_delayed_work_sync(&dev_priv->rps.delayed_resume_work);
3321                 mutex_lock(&dev_priv->rps.hw_lock);
3322                 gen6_disable_rps(dev);
3323                 mutex_unlock(&dev_priv->rps.hw_lock);
3324         }
3325 }
3326
3327 static void intel_gen6_powersave_work(struct work_struct *work)
3328 {
3329         struct drm_i915_private *dev_priv =
3330                 container_of(work, struct drm_i915_private,
3331                              rps.delayed_resume_work.work);
3332         struct drm_device *dev = dev_priv->dev;
3333
3334         mutex_lock(&dev_priv->rps.hw_lock);
3335         gen6_enable_rps(dev);
3336         gen6_update_ring_freq(dev);
3337         mutex_unlock(&dev_priv->rps.hw_lock);
3338 }
3339
3340 void intel_enable_gt_powersave(struct drm_device *dev)
3341 {
3342         struct drm_i915_private *dev_priv = dev->dev_private;
3343
3344         if (IS_IRONLAKE_M(dev)) {
3345                 ironlake_enable_drps(dev);
3346                 ironlake_enable_rc6(dev);
3347                 intel_init_emon(dev);
3348         } else if ((IS_GEN6(dev) || IS_GEN7(dev)) && !IS_VALLEYVIEW(dev)) {
3349                 /*
3350                  * PCU communication is slow and this doesn't need to be
3351                  * done at any specific time, so do this out of our fast path
3352                  * to make resume and init faster.
3353                  */
3354                 schedule_delayed_work(&dev_priv->rps.delayed_resume_work,
3355                                       round_jiffies_up_relative(HZ));
3356         }
3357 }
3358
3359 static void ibx_init_clock_gating(struct drm_device *dev)
3360 {
3361         struct drm_i915_private *dev_priv = dev->dev_private;
3362
3363         /*
3364          * On Ibex Peak and Cougar Point, we need to disable clock
3365          * gating for the panel power sequencer or it will fail to
3366          * start up when no ports are active.
3367          */
3368         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
3369 }
3370
3371 static void ironlake_init_clock_gating(struct drm_device *dev)
3372 {
3373         struct drm_i915_private *dev_priv = dev->dev_private;
3374         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
3375
3376         /* Required for FBC */
3377         dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
3378                    ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
3379                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
3380
3381         I915_WRITE(PCH_3DCGDIS0,
3382                    MARIUNIT_CLOCK_GATE_DISABLE |
3383                    SVSMUNIT_CLOCK_GATE_DISABLE);
3384         I915_WRITE(PCH_3DCGDIS1,
3385                    VFMUNIT_CLOCK_GATE_DISABLE);
3386
3387         /*
3388          * According to the spec the following bits should be set in
3389          * order to enable memory self-refresh
3390          * The bit 22/21 of 0x42004
3391          * The bit 5 of 0x42020
3392          * The bit 15 of 0x45000
3393          */
3394         I915_WRITE(ILK_DISPLAY_CHICKEN2,
3395                    (I915_READ(ILK_DISPLAY_CHICKEN2) |
3396                     ILK_DPARB_GATE | ILK_VSDPFD_FULL));
3397         dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
3398         I915_WRITE(DISP_ARB_CTL,
3399                    (I915_READ(DISP_ARB_CTL) |
3400                     DISP_FBC_WM_DIS));
3401         I915_WRITE(WM3_LP_ILK, 0);
3402         I915_WRITE(WM2_LP_ILK, 0);
3403         I915_WRITE(WM1_LP_ILK, 0);
3404
3405         /*
3406          * Based on the document from hardware guys the following bits
3407          * should be set unconditionally in order to enable FBC.
3408          * The bit 22 of 0x42000
3409          * The bit 22 of 0x42004
3410          * The bit 7,8,9 of 0x42020.
3411          */
3412         if (IS_IRONLAKE_M(dev)) {
3413                 I915_WRITE(ILK_DISPLAY_CHICKEN1,
3414                            I915_READ(ILK_DISPLAY_CHICKEN1) |
3415                            ILK_FBCQ_DIS);
3416                 I915_WRITE(ILK_DISPLAY_CHICKEN2,
3417                            I915_READ(ILK_DISPLAY_CHICKEN2) |
3418                            ILK_DPARB_GATE);
3419         }
3420
3421         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
3422
3423         I915_WRITE(ILK_DISPLAY_CHICKEN2,
3424                    I915_READ(ILK_DISPLAY_CHICKEN2) |
3425                    ILK_ELPIN_409_SELECT);
3426         I915_WRITE(_3D_CHICKEN2,
3427                    _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
3428                    _3D_CHICKEN2_WM_READ_PIPELINED);
3429
3430         /* WaDisableRenderCachePipelinedFlush */
3431         I915_WRITE(CACHE_MODE_0,
3432                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
3433
3434         ibx_init_clock_gating(dev);
3435 }
3436
3437 static void cpt_init_clock_gating(struct drm_device *dev)
3438 {
3439         struct drm_i915_private *dev_priv = dev->dev_private;
3440         int pipe;
3441
3442         /*
3443          * On Ibex Peak and Cougar Point, we need to disable clock
3444          * gating for the panel power sequencer or it will fail to
3445          * start up when no ports are active.
3446          */
3447         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
3448         I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
3449                    DPLS_EDP_PPS_FIX_DIS);
3450         /* The below fixes the weird display corruption, a few pixels shifted
3451          * downward, on (only) LVDS of some HP laptops with IVY.
3452          */
3453         for_each_pipe(pipe)
3454                 I915_WRITE(TRANS_CHICKEN2(pipe), TRANS_CHICKEN2_TIMING_OVERRIDE);
3455         /* WADP0ClockGatingDisable */
3456         for_each_pipe(pipe) {
3457                 I915_WRITE(TRANS_CHICKEN1(pipe),
3458                            TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
3459         }
3460 }
3461
3462 static void gen6_init_clock_gating(struct drm_device *dev)
3463 {
3464         struct drm_i915_private *dev_priv = dev->dev_private;
3465         int pipe;
3466         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
3467
3468         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
3469
3470         I915_WRITE(ILK_DISPLAY_CHICKEN2,
3471                    I915_READ(ILK_DISPLAY_CHICKEN2) |
3472                    ILK_ELPIN_409_SELECT);
3473
3474         I915_WRITE(WM3_LP_ILK, 0);
3475         I915_WRITE(WM2_LP_ILK, 0);
3476         I915_WRITE(WM1_LP_ILK, 0);
3477
3478         I915_WRITE(CACHE_MODE_0,
3479                    _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
3480
3481         I915_WRITE(GEN6_UCGCTL1,
3482                    I915_READ(GEN6_UCGCTL1) |
3483                    GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
3484                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
3485
3486         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
3487          * gating disable must be set.  Failure to set it results in
3488          * flickering pixels due to Z write ordering failures after
3489          * some amount of runtime in the Mesa "fire" demo, and Unigine
3490          * Sanctuary and Tropics, and apparently anything else with
3491          * alpha test or pixel discard.
3492          *
3493          * According to the spec, bit 11 (RCCUNIT) must also be set,
3494          * but we didn't debug actual testcases to find it out.
3495          *
3496          * Also apply WaDisableVDSUnitClockGating and
3497          * WaDisableRCPBUnitClockGating.
3498          */
3499         I915_WRITE(GEN6_UCGCTL2,
3500                    GEN7_VDSUNIT_CLOCK_GATE_DISABLE |
3501                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
3502                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
3503
3504         /* Bspec says we need to always set all mask bits. */
3505         I915_WRITE(_3D_CHICKEN3, (0xFFFF << 16) |
3506                    _3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL);
3507
3508         /*
3509          * According to the spec the following bits should be
3510          * set in order to enable memory self-refresh and fbc:
3511          * The bit21 and bit22 of 0x42000
3512          * The bit21 and bit22 of 0x42004
3513          * The bit5 and bit7 of 0x42020
3514          * The bit14 of 0x70180
3515          * The bit14 of 0x71180
3516          */
3517         I915_WRITE(ILK_DISPLAY_CHICKEN1,
3518                    I915_READ(ILK_DISPLAY_CHICKEN1) |
3519                    ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
3520         I915_WRITE(ILK_DISPLAY_CHICKEN2,
3521                    I915_READ(ILK_DISPLAY_CHICKEN2) |
3522                    ILK_DPARB_GATE | ILK_VSDPFD_FULL);
3523         I915_WRITE(ILK_DSPCLK_GATE_D,
3524                    I915_READ(ILK_DSPCLK_GATE_D) |
3525                    ILK_DPARBUNIT_CLOCK_GATE_ENABLE  |
3526                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
3527
3528         /* WaMbcDriverBootEnable */
3529         I915_WRITE(GEN6_MBCTL, I915_READ(GEN6_MBCTL) |
3530                    GEN6_MBCTL_ENABLE_BOOT_FETCH);
3531
3532         for_each_pipe(pipe) {
3533                 I915_WRITE(DSPCNTR(pipe),
3534                            I915_READ(DSPCNTR(pipe)) |
3535                            DISPPLANE_TRICKLE_FEED_DISABLE);
3536                 intel_flush_display_plane(dev_priv, pipe);
3537         }
3538
3539         /* The default value should be 0x200 according to docs, but the two
3540          * platforms I checked have a 0 for this. (Maybe BIOS overrides?) */
3541         I915_WRITE(GEN6_GT_MODE, _MASKED_BIT_DISABLE(0xffff));
3542         I915_WRITE(GEN6_GT_MODE, _MASKED_BIT_ENABLE(GEN6_GT_MODE_HI));
3543
3544         cpt_init_clock_gating(dev);
3545 }
3546
3547 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
3548 {
3549         uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
3550
3551         reg &= ~GEN7_FF_SCHED_MASK;
3552         reg |= GEN7_FF_TS_SCHED_HW;
3553         reg |= GEN7_FF_VS_SCHED_HW;
3554         reg |= GEN7_FF_DS_SCHED_HW;
3555
3556         I915_WRITE(GEN7_FF_THREAD_MODE, reg);
3557 }
3558
3559 static void lpt_init_clock_gating(struct drm_device *dev)
3560 {
3561         struct drm_i915_private *dev_priv = dev->dev_private;
3562
3563         /*
3564          * TODO: this bit should only be enabled when really needed, then
3565          * disabled when not needed anymore in order to save power.
3566          */
3567         if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE)
3568                 I915_WRITE(SOUTH_DSPCLK_GATE_D,
3569                            I915_READ(SOUTH_DSPCLK_GATE_D) |
3570                            PCH_LP_PARTITION_LEVEL_DISABLE);
3571 }
3572
3573 static void haswell_init_clock_gating(struct drm_device *dev)
3574 {
3575         struct drm_i915_private *dev_priv = dev->dev_private;
3576         int pipe;
3577
3578         I915_WRITE(WM3_LP_ILK, 0);
3579         I915_WRITE(WM2_LP_ILK, 0);
3580         I915_WRITE(WM1_LP_ILK, 0);
3581
3582         /* According to the spec, bit 13 (RCZUNIT) must be set on IVB.
3583          * This implements the WaDisableRCZUnitClockGating workaround.
3584          */
3585         I915_WRITE(GEN6_UCGCTL2, GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
3586
3587         /* Apply the WaDisableRHWOOptimizationForRenderHang workaround. */
3588         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
3589                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
3590
3591         /* WaApplyL3ControlAndL3ChickenMode requires those two on Ivy Bridge */
3592         I915_WRITE(GEN7_L3CNTLREG1,
3593                         GEN7_WA_FOR_GEN7_L3_CONTROL);
3594         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
3595                         GEN7_WA_L3_CHICKEN_MODE);
3596
3597         /* This is required by WaCatErrorRejectionIssue */
3598         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
3599                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
3600                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
3601
3602         for_each_pipe(pipe) {
3603                 I915_WRITE(DSPCNTR(pipe),
3604                            I915_READ(DSPCNTR(pipe)) |
3605                            DISPPLANE_TRICKLE_FEED_DISABLE);
3606                 intel_flush_display_plane(dev_priv, pipe);
3607         }
3608
3609         gen7_setup_fixed_func_scheduler(dev_priv);
3610
3611         /* WaDisable4x2SubspanOptimization */
3612         I915_WRITE(CACHE_MODE_1,
3613                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
3614
3615         /* WaMbcDriverBootEnable */
3616         I915_WRITE(GEN6_MBCTL, I915_READ(GEN6_MBCTL) |
3617                    GEN6_MBCTL_ENABLE_BOOT_FETCH);
3618
3619         /* XXX: This is a workaround for early silicon revisions and should be
3620          * removed later.
3621          */
3622         I915_WRITE(WM_DBG,
3623                         I915_READ(WM_DBG) |
3624                         WM_DBG_DISALLOW_MULTIPLE_LP |
3625                         WM_DBG_DISALLOW_SPRITE |
3626                         WM_DBG_DISALLOW_MAXFIFO);
3627
3628         lpt_init_clock_gating(dev);
3629 }
3630
3631 static void ivybridge_init_clock_gating(struct drm_device *dev)
3632 {
3633         struct drm_i915_private *dev_priv = dev->dev_private;
3634         int pipe;
3635         uint32_t snpcr;
3636
3637         I915_WRITE(WM3_LP_ILK, 0);
3638         I915_WRITE(WM2_LP_ILK, 0);
3639         I915_WRITE(WM1_LP_ILK, 0);
3640
3641         I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
3642
3643         /* WaDisableEarlyCull */
3644         I915_WRITE(_3D_CHICKEN3,
3645                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
3646
3647         /* WaDisableBackToBackFlipFix */
3648         I915_WRITE(IVB_CHICKEN3,
3649                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
3650                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
3651
3652         /* WaDisablePSDDualDispatchEnable */
3653         if (IS_IVB_GT1(dev))
3654                 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
3655                            _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
3656         else
3657                 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1_GT2,
3658                            _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
3659
3660         /* Apply the WaDisableRHWOOptimizationForRenderHang workaround. */
3661         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
3662                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
3663
3664         /* WaApplyL3ControlAndL3ChickenMode requires those two on Ivy Bridge */
3665         I915_WRITE(GEN7_L3CNTLREG1,
3666                         GEN7_WA_FOR_GEN7_L3_CONTROL);
3667         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
3668                    GEN7_WA_L3_CHICKEN_MODE);
3669         if (IS_IVB_GT1(dev))
3670                 I915_WRITE(GEN7_ROW_CHICKEN2,
3671                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
3672         else
3673                 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
3674                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
3675
3676
3677         /* WaForceL3Serialization */
3678         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
3679                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
3680
3681         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
3682          * gating disable must be set.  Failure to set it results in
3683          * flickering pixels due to Z write ordering failures after
3684          * some amount of runtime in the Mesa "fire" demo, and Unigine
3685          * Sanctuary and Tropics, and apparently anything else with
3686          * alpha test or pixel discard.
3687          *
3688          * According to the spec, bit 11 (RCCUNIT) must also be set,
3689          * but we didn't debug actual testcases to find it out.
3690          *
3691          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
3692          * This implements the WaDisableRCZUnitClockGating workaround.
3693          */
3694         I915_WRITE(GEN6_UCGCTL2,
3695                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE |
3696                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
3697
3698         /* This is required by WaCatErrorRejectionIssue */
3699         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
3700                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
3701                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
3702
3703         for_each_pipe(pipe) {
3704                 I915_WRITE(DSPCNTR(pipe),
3705                            I915_READ(DSPCNTR(pipe)) |
3706                            DISPPLANE_TRICKLE_FEED_DISABLE);
3707                 intel_flush_display_plane(dev_priv, pipe);
3708         }
3709
3710         /* WaMbcDriverBootEnable */
3711         I915_WRITE(GEN6_MBCTL, I915_READ(GEN6_MBCTL) |
3712                    GEN6_MBCTL_ENABLE_BOOT_FETCH);
3713
3714         gen7_setup_fixed_func_scheduler(dev_priv);
3715
3716         /* WaDisable4x2SubspanOptimization */
3717         I915_WRITE(CACHE_MODE_1,
3718                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
3719
3720         snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
3721         snpcr &= ~GEN6_MBC_SNPCR_MASK;
3722         snpcr |= GEN6_MBC_SNPCR_MED;
3723         I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
3724
3725         cpt_init_clock_gating(dev);
3726 }
3727
3728 static void valleyview_init_clock_gating(struct drm_device *dev)
3729 {
3730         struct drm_i915_private *dev_priv = dev->dev_private;
3731         int pipe;
3732
3733         I915_WRITE(WM3_LP_ILK, 0);
3734         I915_WRITE(WM2_LP_ILK, 0);
3735         I915_WRITE(WM1_LP_ILK, 0);
3736
3737         I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
3738
3739         /* WaDisableEarlyCull */
3740         I915_WRITE(_3D_CHICKEN3,
3741                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
3742
3743         /* WaDisableBackToBackFlipFix */
3744         I915_WRITE(IVB_CHICKEN3,
3745                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
3746                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
3747
3748         I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
3749                    _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
3750
3751         /* Apply the WaDisableRHWOOptimizationForRenderHang workaround. */
3752         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
3753                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
3754
3755         /* WaApplyL3ControlAndL3ChickenMode requires those two on Ivy Bridge */
3756         I915_WRITE(GEN7_L3CNTLREG1, I915_READ(GEN7_L3CNTLREG1) | GEN7_L3AGDIS);
3757         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER, GEN7_WA_L3_CHICKEN_MODE);
3758
3759         /* WaForceL3Serialization */
3760         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
3761                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
3762
3763         /* WaDisableDopClockGating */
3764         I915_WRITE(GEN7_ROW_CHICKEN2,
3765                    _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
3766
3767         /* WaForceL3Serialization */
3768         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
3769                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
3770
3771         /* This is required by WaCatErrorRejectionIssue */
3772         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
3773                    I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
3774                    GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
3775
3776         /* WaMbcDriverBootEnable */
3777         I915_WRITE(GEN6_MBCTL, I915_READ(GEN6_MBCTL) |
3778                    GEN6_MBCTL_ENABLE_BOOT_FETCH);
3779
3780
3781         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
3782          * gating disable must be set.  Failure to set it results in
3783          * flickering pixels due to Z write ordering failures after
3784          * some amount of runtime in the Mesa "fire" demo, and Unigine
3785          * Sanctuary and Tropics, and apparently anything else with
3786          * alpha test or pixel discard.
3787          *
3788          * According to the spec, bit 11 (RCCUNIT) must also be set,
3789          * but we didn't debug actual testcases to find it out.
3790          *
3791          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
3792          * This implements the WaDisableRCZUnitClockGating workaround.
3793          *
3794          * Also apply WaDisableVDSUnitClockGating and
3795          * WaDisableRCPBUnitClockGating.
3796          */
3797         I915_WRITE(GEN6_UCGCTL2,
3798                    GEN7_VDSUNIT_CLOCK_GATE_DISABLE |
3799                    GEN7_TDLUNIT_CLOCK_GATE_DISABLE |
3800                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE |
3801                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
3802                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
3803
3804         I915_WRITE(GEN7_UCGCTL4, GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
3805
3806         for_each_pipe(pipe) {
3807                 I915_WRITE(DSPCNTR(pipe),
3808                            I915_READ(DSPCNTR(pipe)) |
3809                            DISPPLANE_TRICKLE_FEED_DISABLE);
3810                 intel_flush_display_plane(dev_priv, pipe);
3811         }
3812
3813         I915_WRITE(CACHE_MODE_1,
3814                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
3815
3816         /*
3817          * On ValleyView, the GUnit needs to signal the GT
3818          * when flip and other events complete.  So enable
3819          * all the GUnit->GT interrupts here
3820          */
3821         I915_WRITE(VLV_DPFLIPSTAT, PIPEB_LINE_COMPARE_INT_EN |
3822                    PIPEB_HLINE_INT_EN | PIPEB_VBLANK_INT_EN |
3823                    SPRITED_FLIPDONE_INT_EN | SPRITEC_FLIPDONE_INT_EN |
3824                    PLANEB_FLIPDONE_INT_EN | PIPEA_LINE_COMPARE_INT_EN |
3825                    PIPEA_HLINE_INT_EN | PIPEA_VBLANK_INT_EN |
3826                    SPRITEB_FLIPDONE_INT_EN | SPRITEA_FLIPDONE_INT_EN |
3827                    PLANEA_FLIPDONE_INT_EN);
3828
3829         /*
3830          * WaDisableVLVClockGating_VBIIssue
3831          * Disable clock gating on th GCFG unit to prevent a delay
3832          * in the reporting of vblank events.
3833          */
3834         I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
3835 }
3836
3837 static void g4x_init_clock_gating(struct drm_device *dev)
3838 {
3839         struct drm_i915_private *dev_priv = dev->dev_private;
3840         uint32_t dspclk_gate;
3841
3842         I915_WRITE(RENCLK_GATE_D1, 0);
3843         I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
3844                    GS_UNIT_CLOCK_GATE_DISABLE |
3845                    CL_UNIT_CLOCK_GATE_DISABLE);
3846         I915_WRITE(RAMCLK_GATE_D, 0);
3847         dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
3848                 OVRUNIT_CLOCK_GATE_DISABLE |
3849                 OVCUNIT_CLOCK_GATE_DISABLE;
3850         if (IS_GM45(dev))
3851                 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
3852         I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
3853
3854         /* WaDisableRenderCachePipelinedFlush */
3855         I915_WRITE(CACHE_MODE_0,
3856                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
3857 }
3858
3859 static void crestline_init_clock_gating(struct drm_device *dev)
3860 {
3861         struct drm_i915_private *dev_priv = dev->dev_private;
3862
3863         I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
3864         I915_WRITE(RENCLK_GATE_D2, 0);
3865         I915_WRITE(DSPCLK_GATE_D, 0);
3866         I915_WRITE(RAMCLK_GATE_D, 0);
3867         I915_WRITE16(DEUC, 0);
3868 }
3869
3870 static void broadwater_init_clock_gating(struct drm_device *dev)
3871 {
3872         struct drm_i915_private *dev_priv = dev->dev_private;
3873
3874         I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
3875                    I965_RCC_CLOCK_GATE_DISABLE |
3876                    I965_RCPB_CLOCK_GATE_DISABLE |
3877                    I965_ISC_CLOCK_GATE_DISABLE |
3878                    I965_FBC_CLOCK_GATE_DISABLE);
3879         I915_WRITE(RENCLK_GATE_D2, 0);
3880 }
3881
3882 static void gen3_init_clock_gating(struct drm_device *dev)
3883 {
3884         struct drm_i915_private *dev_priv = dev->dev_private;
3885         u32 dstate = I915_READ(D_STATE);
3886
3887         dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
3888                 DSTATE_DOT_CLOCK_GATING;
3889         I915_WRITE(D_STATE, dstate);
3890
3891         if (IS_PINEVIEW(dev))
3892                 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
3893
3894         /* IIR "flip pending" means done if this bit is set */
3895         I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
3896 }
3897
3898 static void i85x_init_clock_gating(struct drm_device *dev)
3899 {
3900         struct drm_i915_private *dev_priv = dev->dev_private;
3901
3902         I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
3903 }
3904
3905 static void i830_init_clock_gating(struct drm_device *dev)
3906 {
3907         struct drm_i915_private *dev_priv = dev->dev_private;
3908
3909         I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
3910 }
3911
3912 void intel_init_clock_gating(struct drm_device *dev)
3913 {
3914         struct drm_i915_private *dev_priv = dev->dev_private;
3915
3916         dev_priv->display.init_clock_gating(dev);
3917 }
3918
3919 /* Starting with Haswell, we have different power wells for
3920  * different parts of the GPU. This attempts to enable them all.
3921  */
3922 void intel_init_power_wells(struct drm_device *dev)
3923 {
3924         struct drm_i915_private *dev_priv = dev->dev_private;
3925         unsigned long power_wells[] = {
3926                 HSW_PWR_WELL_CTL1,
3927                 HSW_PWR_WELL_CTL2,
3928                 HSW_PWR_WELL_CTL4
3929         };
3930         int i;
3931
3932         if (!IS_HASWELL(dev))
3933                 return;
3934
3935         mutex_lock(&dev->struct_mutex);
3936
3937         for (i = 0; i < ARRAY_SIZE(power_wells); i++) {
3938                 int well = I915_READ(power_wells[i]);
3939
3940                 if ((well & HSW_PWR_WELL_STATE) == 0) {
3941                         I915_WRITE(power_wells[i], well & HSW_PWR_WELL_ENABLE);
3942                         if (wait_for((I915_READ(power_wells[i]) & HSW_PWR_WELL_STATE), 20))
3943                                 DRM_ERROR("Error enabling power well %lx\n", power_wells[i]);
3944                 }
3945         }
3946
3947         mutex_unlock(&dev->struct_mutex);
3948 }
3949
3950 /* Set up chip specific power management-related functions */
3951 void intel_init_pm(struct drm_device *dev)
3952 {
3953         struct drm_i915_private *dev_priv = dev->dev_private;
3954
3955         if (I915_HAS_FBC(dev)) {
3956                 if (HAS_PCH_SPLIT(dev)) {
3957                         dev_priv->display.fbc_enabled = ironlake_fbc_enabled;
3958                         dev_priv->display.enable_fbc = ironlake_enable_fbc;
3959                         dev_priv->display.disable_fbc = ironlake_disable_fbc;
3960                 } else if (IS_GM45(dev)) {
3961                         dev_priv->display.fbc_enabled = g4x_fbc_enabled;
3962                         dev_priv->display.enable_fbc = g4x_enable_fbc;
3963                         dev_priv->display.disable_fbc = g4x_disable_fbc;
3964                 } else if (IS_CRESTLINE(dev)) {
3965                         dev_priv->display.fbc_enabled = i8xx_fbc_enabled;
3966                         dev_priv->display.enable_fbc = i8xx_enable_fbc;
3967                         dev_priv->display.disable_fbc = i8xx_disable_fbc;
3968                 }
3969                 /* 855GM needs testing */
3970         }
3971
3972         /* For cxsr */
3973         if (IS_PINEVIEW(dev))
3974                 i915_pineview_get_mem_freq(dev);
3975         else if (IS_GEN5(dev))
3976                 i915_ironlake_get_mem_freq(dev);
3977
3978         /* For FIFO watermark updates */
3979         if (HAS_PCH_SPLIT(dev)) {
3980                 if (IS_GEN5(dev)) {
3981                         if (I915_READ(MLTR_ILK) & ILK_SRLT_MASK)
3982                                 dev_priv->display.update_wm = ironlake_update_wm;
3983                         else {
3984                                 DRM_DEBUG_KMS("Failed to get proper latency. "
3985                                               "Disable CxSR\n");
3986                                 dev_priv->display.update_wm = NULL;
3987                         }
3988                         dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
3989                 } else if (IS_GEN6(dev)) {
3990                         if (SNB_READ_WM0_LATENCY()) {
3991                                 dev_priv->display.update_wm = sandybridge_update_wm;
3992                                 dev_priv->display.update_sprite_wm = sandybridge_update_sprite_wm;
3993                         } else {
3994                                 DRM_DEBUG_KMS("Failed to read display plane latency. "
3995                                               "Disable CxSR\n");
3996                                 dev_priv->display.update_wm = NULL;
3997                         }
3998                         dev_priv->display.init_clock_gating = gen6_init_clock_gating;
3999                 } else if (IS_IVYBRIDGE(dev)) {
4000                         /* FIXME: detect B0+ stepping and use auto training */
4001                         if (SNB_READ_WM0_LATENCY()) {
4002                                 dev_priv->display.update_wm = sandybridge_update_wm;
4003                                 dev_priv->display.update_sprite_wm = sandybridge_update_sprite_wm;
4004                         } else {
4005                                 DRM_DEBUG_KMS("Failed to read display plane latency. "
4006                                               "Disable CxSR\n");
4007                                 dev_priv->display.update_wm = NULL;
4008                         }
4009                         dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
4010                 } else if (IS_HASWELL(dev)) {
4011                         if (SNB_READ_WM0_LATENCY()) {
4012                                 dev_priv->display.update_wm = sandybridge_update_wm;
4013                                 dev_priv->display.update_sprite_wm = sandybridge_update_sprite_wm;
4014                                 dev_priv->display.update_linetime_wm = haswell_update_linetime_wm;
4015                         } else {
4016                                 DRM_DEBUG_KMS("Failed to read display plane latency. "
4017                                               "Disable CxSR\n");
4018                                 dev_priv->display.update_wm = NULL;
4019                         }
4020                         dev_priv->display.init_clock_gating = haswell_init_clock_gating;
4021                 } else
4022                         dev_priv->display.update_wm = NULL;
4023         } else if (IS_VALLEYVIEW(dev)) {
4024                 dev_priv->display.update_wm = valleyview_update_wm;
4025                 dev_priv->display.init_clock_gating =
4026                         valleyview_init_clock_gating;
4027         } else if (IS_PINEVIEW(dev)) {
4028                 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
4029                                             dev_priv->is_ddr3,
4030                                             dev_priv->fsb_freq,
4031                                             dev_priv->mem_freq)) {
4032                         DRM_INFO("failed to find known CxSR latency "
4033                                  "(found ddr%s fsb freq %d, mem freq %d), "
4034                                  "disabling CxSR\n",
4035                                  (dev_priv->is_ddr3 == 1) ? "3" : "2",
4036                                  dev_priv->fsb_freq, dev_priv->mem_freq);
4037                         /* Disable CxSR and never update its watermark again */
4038                         pineview_disable_cxsr(dev);
4039                         dev_priv->display.update_wm = NULL;
4040                 } else
4041                         dev_priv->display.update_wm = pineview_update_wm;
4042                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
4043         } else if (IS_G4X(dev)) {
4044                 dev_priv->display.update_wm = g4x_update_wm;
4045                 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
4046         } else if (IS_GEN4(dev)) {
4047                 dev_priv->display.update_wm = i965_update_wm;
4048                 if (IS_CRESTLINE(dev))
4049                         dev_priv->display.init_clock_gating = crestline_init_clock_gating;
4050                 else if (IS_BROADWATER(dev))
4051                         dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
4052         } else if (IS_GEN3(dev)) {
4053                 dev_priv->display.update_wm = i9xx_update_wm;
4054                 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
4055                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
4056         } else if (IS_I865G(dev)) {
4057                 dev_priv->display.update_wm = i830_update_wm;
4058                 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
4059                 dev_priv->display.get_fifo_size = i830_get_fifo_size;
4060         } else if (IS_I85X(dev)) {
4061                 dev_priv->display.update_wm = i9xx_update_wm;
4062                 dev_priv->display.get_fifo_size = i85x_get_fifo_size;
4063                 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
4064         } else {
4065                 dev_priv->display.update_wm = i830_update_wm;
4066                 dev_priv->display.init_clock_gating = i830_init_clock_gating;
4067                 if (IS_845G(dev))
4068                         dev_priv->display.get_fifo_size = i845_get_fifo_size;
4069                 else
4070                         dev_priv->display.get_fifo_size = i830_get_fifo_size;
4071         }
4072 }
4073
4074 static void __gen6_gt_wait_for_thread_c0(struct drm_i915_private *dev_priv)
4075 {
4076         u32 gt_thread_status_mask;
4077
4078         if (IS_HASWELL(dev_priv->dev))
4079                 gt_thread_status_mask = GEN6_GT_THREAD_STATUS_CORE_MASK_HSW;
4080         else
4081                 gt_thread_status_mask = GEN6_GT_THREAD_STATUS_CORE_MASK;
4082
4083         /* w/a for a sporadic read returning 0 by waiting for the GT
4084          * thread to wake up.
4085          */
4086         if (wait_for_atomic_us((I915_READ_NOTRACE(GEN6_GT_THREAD_STATUS_REG) & gt_thread_status_mask) == 0, 500))
4087                 DRM_ERROR("GT thread status wait timed out\n");
4088 }
4089
4090 static void __gen6_gt_force_wake_reset(struct drm_i915_private *dev_priv)
4091 {
4092         I915_WRITE_NOTRACE(FORCEWAKE, 0);
4093         POSTING_READ(ECOBUS); /* something from same cacheline, but !FORCEWAKE */
4094 }
4095
4096 static void __gen6_gt_force_wake_get(struct drm_i915_private *dev_priv)
4097 {
4098         u32 forcewake_ack;
4099
4100         if (IS_HASWELL(dev_priv->dev))
4101                 forcewake_ack = FORCEWAKE_ACK_HSW;
4102         else
4103                 forcewake_ack = FORCEWAKE_ACK;
4104
4105         if (wait_for_atomic((I915_READ_NOTRACE(forcewake_ack) & 1) == 0,
4106                             FORCEWAKE_ACK_TIMEOUT_MS))
4107                 DRM_ERROR("Timed out waiting for forcewake old ack to clear.\n");
4108
4109         I915_WRITE_NOTRACE(FORCEWAKE, FORCEWAKE_KERNEL);
4110         POSTING_READ(ECOBUS); /* something from same cacheline, but !FORCEWAKE */
4111
4112         if (wait_for_atomic((I915_READ_NOTRACE(forcewake_ack) & 1),
4113                             FORCEWAKE_ACK_TIMEOUT_MS))
4114                 DRM_ERROR("Timed out waiting for forcewake to ack request.\n");
4115
4116         __gen6_gt_wait_for_thread_c0(dev_priv);
4117 }
4118
4119 static void __gen6_gt_force_wake_mt_reset(struct drm_i915_private *dev_priv)
4120 {
4121         I915_WRITE_NOTRACE(FORCEWAKE_MT, _MASKED_BIT_DISABLE(0xffff));
4122         POSTING_READ(ECOBUS); /* something from same cacheline, but !FORCEWAKE */
4123 }
4124
4125 static void __gen6_gt_force_wake_mt_get(struct drm_i915_private *dev_priv)
4126 {
4127         u32 forcewake_ack;
4128
4129         if (IS_HASWELL(dev_priv->dev))
4130                 forcewake_ack = FORCEWAKE_ACK_HSW;
4131         else
4132                 forcewake_ack = FORCEWAKE_MT_ACK;
4133
4134         if (wait_for_atomic((I915_READ_NOTRACE(forcewake_ack) & 1) == 0,
4135                             FORCEWAKE_ACK_TIMEOUT_MS))
4136                 DRM_ERROR("Timed out waiting for forcewake old ack to clear.\n");
4137
4138         I915_WRITE_NOTRACE(FORCEWAKE_MT, _MASKED_BIT_ENABLE(FORCEWAKE_KERNEL));
4139         POSTING_READ(ECOBUS); /* something from same cacheline, but !FORCEWAKE */
4140
4141         if (wait_for_atomic((I915_READ_NOTRACE(forcewake_ack) & 1),
4142                             FORCEWAKE_ACK_TIMEOUT_MS))
4143                 DRM_ERROR("Timed out waiting for forcewake to ack request.\n");
4144
4145         __gen6_gt_wait_for_thread_c0(dev_priv);
4146 }
4147
4148 /*
4149  * Generally this is called implicitly by the register read function. However,
4150  * if some sequence requires the GT to not power down then this function should
4151  * be called at the beginning of the sequence followed by a call to
4152  * gen6_gt_force_wake_put() at the end of the sequence.
4153  */
4154 void gen6_gt_force_wake_get(struct drm_i915_private *dev_priv)
4155 {
4156         unsigned long irqflags;
4157
4158         spin_lock_irqsave(&dev_priv->gt_lock, irqflags);
4159         if (dev_priv->forcewake_count++ == 0)
4160                 dev_priv->gt.force_wake_get(dev_priv);
4161         spin_unlock_irqrestore(&dev_priv->gt_lock, irqflags);
4162 }
4163
4164 void gen6_gt_check_fifodbg(struct drm_i915_private *dev_priv)
4165 {
4166         u32 gtfifodbg;
4167         gtfifodbg = I915_READ_NOTRACE(GTFIFODBG);
4168         if (WARN(gtfifodbg & GT_FIFO_CPU_ERROR_MASK,
4169              "MMIO read or write has been dropped %x\n", gtfifodbg))
4170                 I915_WRITE_NOTRACE(GTFIFODBG, GT_FIFO_CPU_ERROR_MASK);
4171 }
4172
4173 static void __gen6_gt_force_wake_put(struct drm_i915_private *dev_priv)
4174 {
4175         I915_WRITE_NOTRACE(FORCEWAKE, 0);
4176         /* gen6_gt_check_fifodbg doubles as the POSTING_READ */
4177         gen6_gt_check_fifodbg(dev_priv);
4178 }
4179
4180 static void __gen6_gt_force_wake_mt_put(struct drm_i915_private *dev_priv)
4181 {
4182         I915_WRITE_NOTRACE(FORCEWAKE_MT, _MASKED_BIT_DISABLE(FORCEWAKE_KERNEL));
4183         /* gen6_gt_check_fifodbg doubles as the POSTING_READ */
4184         gen6_gt_check_fifodbg(dev_priv);
4185 }
4186
4187 /*
4188  * see gen6_gt_force_wake_get()
4189  */
4190 void gen6_gt_force_wake_put(struct drm_i915_private *dev_priv)
4191 {
4192         unsigned long irqflags;
4193
4194         spin_lock_irqsave(&dev_priv->gt_lock, irqflags);
4195         if (--dev_priv->forcewake_count == 0)
4196                 dev_priv->gt.force_wake_put(dev_priv);
4197         spin_unlock_irqrestore(&dev_priv->gt_lock, irqflags);
4198 }
4199
4200 int __gen6_gt_wait_for_fifo(struct drm_i915_private *dev_priv)
4201 {
4202         int ret = 0;
4203
4204         if (dev_priv->gt_fifo_count < GT_FIFO_NUM_RESERVED_ENTRIES) {
4205                 int loop = 500;
4206                 u32 fifo = I915_READ_NOTRACE(GT_FIFO_FREE_ENTRIES);
4207                 while (fifo <= GT_FIFO_NUM_RESERVED_ENTRIES && loop--) {
4208                         udelay(10);
4209                         fifo = I915_READ_NOTRACE(GT_FIFO_FREE_ENTRIES);
4210                 }
4211                 if (WARN_ON(loop < 0 && fifo <= GT_FIFO_NUM_RESERVED_ENTRIES))
4212                         ++ret;
4213                 dev_priv->gt_fifo_count = fifo;
4214         }
4215         dev_priv->gt_fifo_count--;
4216
4217         return ret;
4218 }
4219
4220 static void vlv_force_wake_reset(struct drm_i915_private *dev_priv)
4221 {
4222         I915_WRITE_NOTRACE(FORCEWAKE_VLV, _MASKED_BIT_DISABLE(0xffff));
4223 }
4224
4225 static void vlv_force_wake_get(struct drm_i915_private *dev_priv)
4226 {
4227         if (wait_for_atomic((I915_READ_NOTRACE(FORCEWAKE_ACK_VLV) & 1) == 0,
4228                             FORCEWAKE_ACK_TIMEOUT_MS))
4229                 DRM_ERROR("Timed out waiting for forcewake old ack to clear.\n");
4230
4231         I915_WRITE_NOTRACE(FORCEWAKE_VLV, _MASKED_BIT_ENABLE(FORCEWAKE_KERNEL));
4232
4233         if (wait_for_atomic((I915_READ_NOTRACE(FORCEWAKE_ACK_VLV) & 1),
4234                             FORCEWAKE_ACK_TIMEOUT_MS))
4235                 DRM_ERROR("Timed out waiting for forcewake to ack request.\n");
4236
4237         __gen6_gt_wait_for_thread_c0(dev_priv);
4238 }
4239
4240 static void vlv_force_wake_put(struct drm_i915_private *dev_priv)
4241 {
4242         I915_WRITE_NOTRACE(FORCEWAKE_VLV, _MASKED_BIT_DISABLE(FORCEWAKE_KERNEL));
4243         /* The below doubles as a POSTING_READ */
4244         gen6_gt_check_fifodbg(dev_priv);
4245 }
4246
4247 void intel_gt_reset(struct drm_device *dev)
4248 {
4249         struct drm_i915_private *dev_priv = dev->dev_private;
4250
4251         if (IS_VALLEYVIEW(dev)) {
4252                 vlv_force_wake_reset(dev_priv);
4253         } else if (INTEL_INFO(dev)->gen >= 6) {
4254                 __gen6_gt_force_wake_reset(dev_priv);
4255                 if (IS_IVYBRIDGE(dev) || IS_HASWELL(dev))
4256                         __gen6_gt_force_wake_mt_reset(dev_priv);
4257         }
4258 }
4259
4260 void intel_gt_init(struct drm_device *dev)
4261 {
4262         struct drm_i915_private *dev_priv = dev->dev_private;
4263
4264         spin_lock_init(&dev_priv->gt_lock);
4265
4266         intel_gt_reset(dev);
4267
4268         if (IS_VALLEYVIEW(dev)) {
4269                 dev_priv->gt.force_wake_get = vlv_force_wake_get;
4270                 dev_priv->gt.force_wake_put = vlv_force_wake_put;
4271         } else if (IS_IVYBRIDGE(dev) || IS_HASWELL(dev)) {
4272                 dev_priv->gt.force_wake_get = __gen6_gt_force_wake_mt_get;
4273                 dev_priv->gt.force_wake_put = __gen6_gt_force_wake_mt_put;
4274         } else if (IS_GEN6(dev)) {
4275                 dev_priv->gt.force_wake_get = __gen6_gt_force_wake_get;
4276                 dev_priv->gt.force_wake_put = __gen6_gt_force_wake_put;
4277         }
4278         INIT_DELAYED_WORK(&dev_priv->rps.delayed_resume_work,
4279                           intel_gen6_powersave_work);
4280 }
4281
4282 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u8 mbox, u32 *val)
4283 {
4284         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4285
4286         if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
4287                 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
4288                 return -EAGAIN;
4289         }
4290
4291         I915_WRITE(GEN6_PCODE_DATA, *val);
4292         I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
4293
4294         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
4295                      500)) {
4296                 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
4297                 return -ETIMEDOUT;
4298         }
4299
4300         *val = I915_READ(GEN6_PCODE_DATA);
4301         I915_WRITE(GEN6_PCODE_DATA, 0);
4302
4303         return 0;
4304 }
4305
4306 int sandybridge_pcode_write(struct drm_i915_private *dev_priv, u8 mbox, u32 val)
4307 {
4308         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4309
4310         if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
4311                 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
4312                 return -EAGAIN;
4313         }
4314
4315         I915_WRITE(GEN6_PCODE_DATA, val);
4316         I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
4317
4318         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
4319                      500)) {
4320                 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
4321                 return -ETIMEDOUT;
4322         }
4323
4324         I915_WRITE(GEN6_PCODE_DATA, 0);
4325
4326         return 0;
4327 }