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 /**
  35  * RC6 is a special power stage which allows the GPU to enter an very
  36  * low-voltage mode when idle, using down to 0V while at this stage.  This
  37  * stage is entered automatically when the GPU is idle when RC6 support is
  38  * enabled, and as soon as new workload arises GPU wakes up automatically as well.
  39  *
  40  * There are different RC6 modes available in Intel GPU, which differentiate
  41  * among each other with the latency required to enter and leave RC6 and
  42  * voltage consumed by the GPU in different states.
  43  *
  44  * The combination of the following flags define which states GPU is allowed
  45  * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
  46  * RC6pp is deepest RC6. Their support by hardware varies according to the
  47  * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
  48  * which brings the most power savings; deeper states save more power, but
  49  * require higher latency to switch to and wake up.
  50  */
  51 #define INTEL_RC6_ENABLE                        (1<<0)
  52 #define INTEL_RC6p_ENABLE                       (1<<1)
  53 #define INTEL_RC6pp_ENABLE                      (1<<2)
  54
  55 static void gen9_init_clock_gating(struct drm_device *dev)
  56 {
  57         struct drm_i915_private *dev_priv = dev->dev_private;
  58
  59         /* WaEnableLbsSlaRetryTimerDecrement:skl */
  60         I915_WRITE(BDW_SCRATCH1, I915_READ(BDW_SCRATCH1) |
  61                    GEN9_LBS_SLA_RETRY_TIMER_DECREMENT_ENABLE);
  62 }
  63
  64 static void skl_init_clock_gating(struct drm_device *dev)
  65 {
  66         struct drm_i915_private *dev_priv = dev->dev_private;
  67
  68         gen9_init_clock_gating(dev);
  69
  70         if (INTEL_REVID(dev) == SKL_REVID_A0) {
  71                 /*
  72                  * WaDisableSDEUnitClockGating:skl
  73                  * WaSetGAPSunitClckGateDisable:skl
  74                  */
  75                 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
  76                            GEN8_GAPSUNIT_CLOCK_GATE_DISABLE |
  77                            GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
  78         }
  79
  80         if (INTEL_REVID(dev) <= SKL_REVID_D0) {
  81                 /* WaDisableHDCInvalidation:skl */
  82                 I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) |
  83                            BDW_DISABLE_HDC_INVALIDATION);
  84
  85                 /* WaDisableChickenBitTSGBarrierAckForFFSliceCS:skl */
  86                 I915_WRITE(FF_SLICE_CS_CHICKEN2,
  87                            I915_READ(FF_SLICE_CS_CHICKEN2) |
  88                            GEN9_TSG_BARRIER_ACK_DISABLE);
  89         }
  90
  91         if (INTEL_REVID(dev) <= SKL_REVID_E0)
  92                 /* WaDisableLSQCROPERFforOCL:skl */
  93                 I915_WRITE(GEN8_L3SQCREG4, I915_READ(GEN8_L3SQCREG4) |
  94                            GEN8_LQSC_RO_PERF_DIS);
  95 }
  96
  97 static void i915_pineview_get_mem_freq(struct drm_device *dev)
  98 {
  99         struct drm_i915_private *dev_priv = dev->dev_private;
 100         u32 tmp;
 101
 102         tmp = I915_READ(CLKCFG);
 103
 104         switch (tmp & CLKCFG_FSB_MASK) {
 105         case CLKCFG_FSB_533:
 106                 dev_priv->fsb_freq = 533; /* 133*4 */
 107                 break;
 108         case CLKCFG_FSB_800:
 109                 dev_priv->fsb_freq = 800; /* 200*4 */
 110                 break;
 111         case CLKCFG_FSB_667:
 112                 dev_priv->fsb_freq =  667; /* 167*4 */
 113                 break;
 114         case CLKCFG_FSB_400:
 115                 dev_priv->fsb_freq = 400; /* 100*4 */
 116                 break;
 117         }
 118
 119         switch (tmp & CLKCFG_MEM_MASK) {
 120         case CLKCFG_MEM_533:
 121                 dev_priv->mem_freq = 533;
 122                 break;
 123         case CLKCFG_MEM_667:
 124                 dev_priv->mem_freq = 667;
 125                 break;
 126         case CLKCFG_MEM_800:
 127                 dev_priv->mem_freq = 800;
 128                 break;
 129         }
 130
 131         /* detect pineview DDR3 setting */
 132         tmp = I915_READ(CSHRDDR3CTL);
 133         dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
 134 }
 135
 136 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
 137 {
 138         struct drm_i915_private *dev_priv = dev->dev_private;
 139         u16 ddrpll, csipll;
 140
 141         ddrpll = I915_READ16(DDRMPLL1);
 142         csipll = I915_READ16(CSIPLL0);
 143
 144         switch (ddrpll & 0xff) {
 145         case 0xc:
 146                 dev_priv->mem_freq = 800;
 147                 break;
 148         case 0x10:
 149                 dev_priv->mem_freq = 1066;
 150                 break;
 151         case 0x14:
 152                 dev_priv->mem_freq = 1333;
 153                 break;
 154         case 0x18:
 155                 dev_priv->mem_freq = 1600;
 156                 break;
 157         default:
 158                 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
 159                                  ddrpll & 0xff);
 160                 dev_priv->mem_freq = 0;
 161                 break;
 162         }
 163
 164         dev_priv->ips.r_t = dev_priv->mem_freq;
 165
 166         switch (csipll & 0x3ff) {
 167         case 0x00c:
 168                 dev_priv->fsb_freq = 3200;
 169                 break;
 170         case 0x00e:
 171                 dev_priv->fsb_freq = 3733;
 172                 break;
 173         case 0x010:
 174                 dev_priv->fsb_freq = 4266;
 175                 break;
 176         case 0x012:
 177                 dev_priv->fsb_freq = 4800;
 178                 break;
 179         case 0x014:
 180                 dev_priv->fsb_freq = 5333;
 181                 break;
 182         case 0x016:
 183                 dev_priv->fsb_freq = 5866;
 184                 break;
 185         case 0x018:
 186                 dev_priv->fsb_freq = 6400;
 187                 break;
 188         default:
 189                 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
 190                                  csipll & 0x3ff);
 191                 dev_priv->fsb_freq = 0;
 192                 break;
 193         }
 194
 195         if (dev_priv->fsb_freq == 3200) {
 196                 dev_priv->ips.c_m = 0;
 197         } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
 198                 dev_priv->ips.c_m = 1;
 199         } else {
 200                 dev_priv->ips.c_m = 2;
 201         }
 202 }
 203
 204 static const struct cxsr_latency cxsr_latency_table[] = {
 205         {1, 0, 800, 400, 3382, 33382, 3983, 33983},    /* DDR2-400 SC */
 206         {1, 0, 800, 667, 3354, 33354, 3807, 33807},    /* DDR2-667 SC */
 207         {1, 0, 800, 800, 3347, 33347, 3763, 33763},    /* DDR2-800 SC */
 208         {1, 1, 800, 667, 6420, 36420, 6873, 36873},    /* DDR3-667 SC */
 209         {1, 1, 800, 800, 5902, 35902, 6318, 36318},    /* DDR3-800 SC */
 210
 211         {1, 0, 667, 400, 3400, 33400, 4021, 34021},    /* DDR2-400 SC */
 212         {1, 0, 667, 667, 3372, 33372, 3845, 33845},    /* DDR2-667 SC */
 213         {1, 0, 667, 800, 3386, 33386, 3822, 33822},    /* DDR2-800 SC */
 214         {1, 1, 667, 667, 6438, 36438, 6911, 36911},    /* DDR3-667 SC */
 215         {1, 1, 667, 800, 5941, 35941, 6377, 36377},    /* DDR3-800 SC */
 216
 217         {1, 0, 400, 400, 3472, 33472, 4173, 34173},    /* DDR2-400 SC */
 218         {1, 0, 400, 667, 3443, 33443, 3996, 33996},    /* DDR2-667 SC */
 219         {1, 0, 400, 800, 3430, 33430, 3946, 33946},    /* DDR2-800 SC */
 220         {1, 1, 400, 667, 6509, 36509, 7062, 37062},    /* DDR3-667 SC */
 221         {1, 1, 400, 800, 5985, 35985, 6501, 36501},    /* DDR3-800 SC */
 222
 223         {0, 0, 800, 400, 3438, 33438, 4065, 34065},    /* DDR2-400 SC */
 224         {0, 0, 800, 667, 3410, 33410, 3889, 33889},    /* DDR2-667 SC */
 225         {0, 0, 800, 800, 3403, 33403, 3845, 33845},    /* DDR2-800 SC */
 226         {0, 1, 800, 667, 6476, 36476, 6955, 36955},    /* DDR3-667 SC */
 227         {0, 1, 800, 800, 5958, 35958, 6400, 36400},    /* DDR3-800 SC */
 228
 229         {0, 0, 667, 400, 3456, 33456, 4103, 34106},    /* DDR2-400 SC */
 230         {0, 0, 667, 667, 3428, 33428, 3927, 33927},    /* DDR2-667 SC */
 231         {0, 0, 667, 800, 3443, 33443, 3905, 33905},    /* DDR2-800 SC */
 232         {0, 1, 667, 667, 6494, 36494, 6993, 36993},    /* DDR3-667 SC */
 233         {0, 1, 667, 800, 5998, 35998, 6460, 36460},    /* DDR3-800 SC */
 234
 235         {0, 0, 400, 400, 3528, 33528, 4255, 34255},    /* DDR2-400 SC */
 236         {0, 0, 400, 667, 3500, 33500, 4079, 34079},    /* DDR2-667 SC */
 237         {0, 0, 400, 800, 3487, 33487, 4029, 34029},    /* DDR2-800 SC */
 238         {0, 1, 400, 667, 6566, 36566, 7145, 37145},    /* DDR3-667 SC */
 239         {0, 1, 400, 800, 6042, 36042, 6584, 36584},    /* DDR3-800 SC */
 240 };
 241
 242 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
 243                                                          int is_ddr3,
 244                                                          int fsb,
 245                                                          int mem)
 246 {
 247         const struct cxsr_latency *latency;
 248         int i;
 249
 250         if (fsb == 0 || mem == 0)
 251                 return NULL;
 252
 253         for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
 254                 latency = &cxsr_latency_table[i];
 255                 if (is_desktop == latency->is_desktop &&
 256                     is_ddr3 == latency->is_ddr3 &&
 257                     fsb == latency->fsb_freq && mem == latency->mem_freq)
 258                         return latency;
 259         }
 260
 261         DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 262
 263         return NULL;
 264 }
 265
 266 void intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
 267 {
 268         struct drm_device *dev = dev_priv->dev;
 269         u32 val;
 270
 271         if (IS_VALLEYVIEW(dev)) {
 272                 I915_WRITE(FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
 273         } else if (IS_G4X(dev) || IS_CRESTLINE(dev)) {
 274                 I915_WRITE(FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
 275         } else if (IS_PINEVIEW(dev)) {
 276                 val = I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN;
 277                 val |= enable ? PINEVIEW_SELF_REFRESH_EN : 0;
 278                 I915_WRITE(DSPFW3, val);
 279         } else if (IS_I945G(dev) || IS_I945GM(dev)) {
 280                 val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
 281                                _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
 282                 I915_WRITE(FW_BLC_SELF, val);
 283         } else if (IS_I915GM(dev)) {
 284                 val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
 285                                _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
 286                 I915_WRITE(INSTPM, val);
 287         } else {
 288                 return;
 289         }
 290
 291         DRM_DEBUG_KMS("memory self-refresh is %s\n",
 292                       enable ? "enabled" : "disabled");
 293 }
 294
 295 /*
 296  * Latency for FIFO fetches is dependent on several factors:
 297  *   - memory configuration (speed, channels)
 298  *   - chipset
 299  *   - current MCH state
 300  * It can be fairly high in some situations, so here we assume a fairly
 301  * pessimal value.  It's a tradeoff between extra memory fetches (if we
 302  * set this value too high, the FIFO will fetch frequently to stay full)
 303  * and power consumption (set it too low to save power and we might see
 304  * FIFO underruns and display "flicker").
 305  *
 306  * A value of 5us seems to be a good balance; safe for very low end
 307  * platforms but not overly aggressive on lower latency configs.
 308  */
 309 static const int pessimal_latency_ns = 5000;
 310
 311 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
 312 {
 313         struct drm_i915_private *dev_priv = dev->dev_private;
 314         uint32_t dsparb = I915_READ(DSPARB);
 315         int size;
 316
 317         size = dsparb & 0x7f;
 318         if (plane)
 319                 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
 320
 321         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 322                       plane ? "B" : "A", size);
 323
 324         return size;
 325 }
 326
 327 static int i830_get_fifo_size(struct drm_device *dev, int plane)
 328 {
 329         struct drm_i915_private *dev_priv = dev->dev_private;
 330         uint32_t dsparb = I915_READ(DSPARB);
 331         int size;
 332
 333         size = dsparb & 0x1ff;
 334         if (plane)
 335                 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
 336         size >>= 1; /* Convert to cachelines */
 337
 338         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 339                       plane ? "B" : "A", size);
 340
 341         return size;
 342 }
 343
 344 static int i845_get_fifo_size(struct drm_device *dev, int plane)
 345 {
 346         struct drm_i915_private *dev_priv = dev->dev_private;
 347         uint32_t dsparb = I915_READ(DSPARB);
 348         int size;
 349
 350         size = dsparb & 0x7f;
 351         size >>= 2; /* Convert to cachelines */
 352
 353         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 354                       plane ? "B" : "A",
 355                       size);
 356
 357         return size;
 358 }
 359
 360 /* Pineview has different values for various configs */
 361 static const struct intel_watermark_params pineview_display_wm = {
 362         .fifo_size = PINEVIEW_DISPLAY_FIFO,
 363         .max_wm = PINEVIEW_MAX_WM,
 364         .default_wm = PINEVIEW_DFT_WM,
 365         .guard_size = PINEVIEW_GUARD_WM,
 366         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 367 };
 368 static const struct intel_watermark_params pineview_display_hplloff_wm = {
 369         .fifo_size = PINEVIEW_DISPLAY_FIFO,
 370         .max_wm = PINEVIEW_MAX_WM,
 371         .default_wm = PINEVIEW_DFT_HPLLOFF_WM,
 372         .guard_size = PINEVIEW_GUARD_WM,
 373         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 374 };
 375 static const struct intel_watermark_params pineview_cursor_wm = {
 376         .fifo_size = PINEVIEW_CURSOR_FIFO,
 377         .max_wm = PINEVIEW_CURSOR_MAX_WM,
 378         .default_wm = PINEVIEW_CURSOR_DFT_WM,
 379         .guard_size = PINEVIEW_CURSOR_GUARD_WM,
 380         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 381 };
 382 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
 383         .fifo_size = PINEVIEW_CURSOR_FIFO,
 384         .max_wm = PINEVIEW_CURSOR_MAX_WM,
 385         .default_wm = PINEVIEW_CURSOR_DFT_WM,
 386         .guard_size = PINEVIEW_CURSOR_GUARD_WM,
 387         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 388 };
 389 static const struct intel_watermark_params g4x_wm_info = {
 390         .fifo_size = G4X_FIFO_SIZE,
 391         .max_wm = G4X_MAX_WM,
 392         .default_wm = G4X_MAX_WM,
 393         .guard_size = 2,
 394         .cacheline_size = G4X_FIFO_LINE_SIZE,
 395 };
 396 static const struct intel_watermark_params g4x_cursor_wm_info = {
 397         .fifo_size = I965_CURSOR_FIFO,
 398         .max_wm = I965_CURSOR_MAX_WM,
 399         .default_wm = I965_CURSOR_DFT_WM,
 400         .guard_size = 2,
 401         .cacheline_size = G4X_FIFO_LINE_SIZE,
 402 };
 403 static const struct intel_watermark_params valleyview_wm_info = {
 404         .fifo_size = VALLEYVIEW_FIFO_SIZE,
 405         .max_wm = VALLEYVIEW_MAX_WM,
 406         .default_wm = VALLEYVIEW_MAX_WM,
 407         .guard_size = 2,
 408         .cacheline_size = G4X_FIFO_LINE_SIZE,
 409 };
 410 static const struct intel_watermark_params valleyview_cursor_wm_info = {
 411         .fifo_size = I965_CURSOR_FIFO,
 412         .max_wm = VALLEYVIEW_CURSOR_MAX_WM,
 413         .default_wm = I965_CURSOR_DFT_WM,
 414         .guard_size = 2,
 415         .cacheline_size = G4X_FIFO_LINE_SIZE,
 416 };
 417 static const struct intel_watermark_params i965_cursor_wm_info = {
 418         .fifo_size = I965_CURSOR_FIFO,
 419         .max_wm = I965_CURSOR_MAX_WM,
 420         .default_wm = I965_CURSOR_DFT_WM,
 421         .guard_size = 2,
 422         .cacheline_size = I915_FIFO_LINE_SIZE,
 423 };
 424 static const struct intel_watermark_params i945_wm_info = {
 425         .fifo_size = I945_FIFO_SIZE,
 426         .max_wm = I915_MAX_WM,
 427         .default_wm = 1,
 428         .guard_size = 2,
 429         .cacheline_size = I915_FIFO_LINE_SIZE,
 430 };
 431 static const struct intel_watermark_params i915_wm_info = {
 432         .fifo_size = I915_FIFO_SIZE,
 433         .max_wm = I915_MAX_WM,
 434         .default_wm = 1,
 435         .guard_size = 2,
 436         .cacheline_size = I915_FIFO_LINE_SIZE,
 437 };
 438 static const struct intel_watermark_params i830_a_wm_info = {
 439         .fifo_size = I855GM_FIFO_SIZE,
 440         .max_wm = I915_MAX_WM,
 441         .default_wm = 1,
 442         .guard_size = 2,
 443         .cacheline_size = I830_FIFO_LINE_SIZE,
 444 };
 445 static const struct intel_watermark_params i830_bc_wm_info = {
 446         .fifo_size = I855GM_FIFO_SIZE,
 447         .max_wm = I915_MAX_WM/2,
 448         .default_wm = 1,
 449         .guard_size = 2,
 450         .cacheline_size = I830_FIFO_LINE_SIZE,
 451 };
 452 static const struct intel_watermark_params i845_wm_info = {
 453         .fifo_size = I830_FIFO_SIZE,
 454         .max_wm = I915_MAX_WM,
 455         .default_wm = 1,
 456         .guard_size = 2,
 457         .cacheline_size = I830_FIFO_LINE_SIZE,
 458 };
 459
 460 /**
 461  * intel_calculate_wm - calculate watermark level
 462  * @clock_in_khz: pixel clock
 463  * @wm: chip FIFO params
 464  * @pixel_size: display pixel size
 465  * @latency_ns: memory latency for the platform
 466  *
 467  * Calculate the watermark level (the level at which the display plane will
 468  * start fetching from memory again).  Each chip has a different display
 469  * FIFO size and allocation, so the caller needs to figure that out and pass
 470  * in the correct intel_watermark_params structure.
 471  *
 472  * As the pixel clock runs, the FIFO will be drained at a rate that depends
 473  * on the pixel size.  When it reaches the watermark level, it'll start
 474  * fetching FIFO line sized based chunks from memory until the FIFO fills
 475  * past the watermark point.  If the FIFO drains completely, a FIFO underrun
 476  * will occur, and a display engine hang could result.
 477  */
 478 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
 479                                         const struct intel_watermark_params *wm,
 480                                         int fifo_size,
 481                                         int pixel_size,
 482                                         unsigned long latency_ns)
 483 {
 484         long entries_required, wm_size;
 485
 486         /*
 487          * Note: we need to make sure we don't overflow for various clock &
 488          * latency values.
 489          * clocks go from a few thousand to several hundred thousand.
 490          * latency is usually a few thousand
 491          */
 492         entries_required = ((clock_in_khz / 1000) * pixel_size * latency_ns) /
 493                 1000;
 494         entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
 495
 496         DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
 497
 498         wm_size = fifo_size - (entries_required + wm->guard_size);
 499
 500         DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
 501
 502         /* Don't promote wm_size to unsigned... */
 503         if (wm_size > (long)wm->max_wm)
 504                 wm_size = wm->max_wm;
 505         if (wm_size <= 0)
 506                 wm_size = wm->default_wm;
 507
 508         /*
 509          * Bspec seems to indicate that the value shouldn't be lower than
 510          * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
 511          * Lets go for 8 which is the burst size since certain platforms
 512          * already use a hardcoded 8 (which is what the spec says should be
 513          * done).
 514          */
 515         if (wm_size <= 8)
 516                 wm_size = 8;
 517
 518         return wm_size;
 519 }
 520
 521 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
 522 {
 523         struct drm_crtc *crtc, *enabled = NULL;
 524
 525         for_each_crtc(dev, crtc) {
 526                 if (intel_crtc_active(crtc)) {
 527                         if (enabled)
 528                                 return NULL;
 529                         enabled = crtc;
 530                 }
 531         }
 532
 533         return enabled;
 534 }
 535
 536 static void pineview_update_wm(struct drm_crtc *unused_crtc)
 537 {
 538         struct drm_device *dev = unused_crtc->dev;
 539         struct drm_i915_private *dev_priv = dev->dev_private;
 540         struct drm_crtc *crtc;
 541         const struct cxsr_latency *latency;
 542         u32 reg;
 543         unsigned long wm;
 544
 545         latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
 546                                          dev_priv->fsb_freq, dev_priv->mem_freq);
 547         if (!latency) {
 548                 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 549                 intel_set_memory_cxsr(dev_priv, false);
 550                 return;
 551         }
 552
 553         crtc = single_enabled_crtc(dev);
 554         if (crtc) {
 555                 const struct drm_display_mode *adjusted_mode;
 556                 int pixel_size = crtc->primary->fb->bits_per_pixel / 8;
 557                 int clock;
 558
 559                 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
 560                 clock = adjusted_mode->crtc_clock;
 561
 562                 /* Display SR */
 563                 wm = intel_calculate_wm(clock, &pineview_display_wm,
 564                                         pineview_display_wm.fifo_size,
 565                                         pixel_size, latency->display_sr);
 566                 reg = I915_READ(DSPFW1);
 567                 reg &= ~DSPFW_SR_MASK;
 568                 reg |= wm << DSPFW_SR_SHIFT;
 569                 I915_WRITE(DSPFW1, reg);
 570                 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
 571
 572                 /* cursor SR */
 573                 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
 574                                         pineview_display_wm.fifo_size,
 575                                         pixel_size, latency->cursor_sr);
 576                 reg = I915_READ(DSPFW3);
 577                 reg &= ~DSPFW_CURSOR_SR_MASK;
 578                 reg |= (wm & 0x3f) << DSPFW_CURSOR_SR_SHIFT;
 579                 I915_WRITE(DSPFW3, reg);
 580
 581                 /* Display HPLL off SR */
 582                 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
 583                                         pineview_display_hplloff_wm.fifo_size,
 584                                         pixel_size, latency->display_hpll_disable);
 585                 reg = I915_READ(DSPFW3);
 586                 reg &= ~DSPFW_HPLL_SR_MASK;
 587                 reg |= wm & DSPFW_HPLL_SR_MASK;
 588                 I915_WRITE(DSPFW3, reg);
 589
 590                 /* cursor HPLL off SR */
 591                 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
 592                                         pineview_display_hplloff_wm.fifo_size,
 593                                         pixel_size, latency->cursor_hpll_disable);
 594                 reg = I915_READ(DSPFW3);
 595                 reg &= ~DSPFW_HPLL_CURSOR_MASK;
 596                 reg |= (wm & 0x3f) << DSPFW_HPLL_CURSOR_SHIFT;
 597                 I915_WRITE(DSPFW3, reg);
 598                 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
 599
 600                 intel_set_memory_cxsr(dev_priv, true);
 601         } else {
 602                 intel_set_memory_cxsr(dev_priv, false);
 603         }
 604 }
 605
 606 static bool g4x_compute_wm0(struct drm_device *dev,
 607                             int plane,
 608                             const struct intel_watermark_params *display,
 609                             int display_latency_ns,
 610                             const struct intel_watermark_params *cursor,
 611                             int cursor_latency_ns,
 612                             int *plane_wm,
 613                             int *cursor_wm)
 614 {
 615         struct drm_crtc *crtc;
 616         const struct drm_display_mode *adjusted_mode;
 617         int htotal, hdisplay, clock, pixel_size;
 618         int line_time_us, line_count;
 619         int entries, tlb_miss;
 620
 621         crtc = intel_get_crtc_for_plane(dev, plane);
 622         if (!intel_crtc_active(crtc)) {
 623                 *cursor_wm = cursor->guard_size;
 624                 *plane_wm = display->guard_size;
 625                 return false;
 626         }
 627
 628         adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
 629         clock = adjusted_mode->crtc_clock;
 630         htotal = adjusted_mode->crtc_htotal;
 631         hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
 632         pixel_size = crtc->primary->fb->bits_per_pixel / 8;
 633
 634         /* Use the small buffer method to calculate plane watermark */
 635         entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
 636         tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
 637         if (tlb_miss > 0)
 638                 entries += tlb_miss;
 639         entries = DIV_ROUND_UP(entries, display->cacheline_size);
 640         *plane_wm = entries + display->guard_size;
 641         if (*plane_wm > (int)display->max_wm)
 642                 *plane_wm = display->max_wm;
 643
 644         /* Use the large buffer method to calculate cursor watermark */
 645         line_time_us = max(htotal * 1000 / clock, 1);
 646         line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
 647         entries = line_count * to_intel_crtc(crtc)->cursor_width * pixel_size;
 648         tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
 649         if (tlb_miss > 0)
 650                 entries += tlb_miss;
 651         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
 652         *cursor_wm = entries + cursor->guard_size;
 653         if (*cursor_wm > (int)cursor->max_wm)
 654                 *cursor_wm = (int)cursor->max_wm;
 655
 656         return true;
 657 }
 658
 659 /*
 660  * Check the wm result.
 661  *
 662  * If any calculated watermark values is larger than the maximum value that
 663  * can be programmed into the associated watermark register, that watermark
 664  * must be disabled.
 665  */
 666 static bool g4x_check_srwm(struct drm_device *dev,
 667                            int display_wm, int cursor_wm,
 668                            const struct intel_watermark_params *display,
 669                            const struct intel_watermark_params *cursor)
 670 {
 671         DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
 672                       display_wm, cursor_wm);
 673
 674         if (display_wm > display->max_wm) {
 675                 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
 676                               display_wm, display->max_wm);
 677                 return false;
 678         }
 679
 680         if (cursor_wm > cursor->max_wm) {
 681                 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
 682                               cursor_wm, cursor->max_wm);
 683                 return false;
 684         }
 685
 686         if (!(display_wm || cursor_wm)) {
 687                 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
 688                 return false;
 689         }
 690
 691         return true;
 692 }
 693
 694 static bool g4x_compute_srwm(struct drm_device *dev,
 695                              int plane,
 696                              int latency_ns,
 697                              const struct intel_watermark_params *display,
 698                              const struct intel_watermark_params *cursor,
 699                              int *display_wm, int *cursor_wm)
 700 {
 701         struct drm_crtc *crtc;
 702         const struct drm_display_mode *adjusted_mode;
 703         int hdisplay, htotal, pixel_size, clock;
 704         unsigned long line_time_us;
 705         int line_count, line_size;
 706         int small, large;
 707         int entries;
 708
 709         if (!latency_ns) {
 710                 *display_wm = *cursor_wm = 0;
 711                 return false;
 712         }
 713
 714         crtc = intel_get_crtc_for_plane(dev, plane);
 715         adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
 716         clock = adjusted_mode->crtc_clock;
 717         htotal = adjusted_mode->crtc_htotal;
 718         hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
 719         pixel_size = crtc->primary->fb->bits_per_pixel / 8;
 720
 721         line_time_us = max(htotal * 1000 / clock, 1);
 722         line_count = (latency_ns / line_time_us + 1000) / 1000;
 723         line_size = hdisplay * pixel_size;
 724
 725         /* Use the minimum of the small and large buffer method for primary */
 726         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
 727         large = line_count * line_size;
 728
 729         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
 730         *display_wm = entries + display->guard_size;
 731
 732         /* calculate the self-refresh watermark for display cursor */
 733         entries = line_count * pixel_size * to_intel_crtc(crtc)->cursor_width;
 734         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
 735         *cursor_wm = entries + cursor->guard_size;
 736
 737         return g4x_check_srwm(dev,
 738                               *display_wm, *cursor_wm,
 739                               display, cursor);
 740 }
 741
 742 static bool vlv_compute_drain_latency(struct drm_crtc *crtc,
 743                                       int pixel_size,
 744                                       int *prec_mult,
 745                                       int *drain_latency)
 746 {
 747         struct drm_device *dev = crtc->dev;
 748         int entries;
 749         int clock = to_intel_crtc(crtc)->config->base.adjusted_mode.crtc_clock;
 750
 751         if (WARN(clock == 0, "Pixel clock is zero!\n"))
 752                 return false;
 753
 754         if (WARN(pixel_size == 0, "Pixel size is zero!\n"))
 755                 return false;
 756
 757         entries = DIV_ROUND_UP(clock, 1000) * pixel_size;
 758         if (IS_CHERRYVIEW(dev))
 759                 *prec_mult = (entries > 128) ? DRAIN_LATENCY_PRECISION_32 :
 760                                                DRAIN_LATENCY_PRECISION_16;
 761         else
 762                 *prec_mult = (entries > 128) ? DRAIN_LATENCY_PRECISION_64 :
 763                                                DRAIN_LATENCY_PRECISION_32;
 764         *drain_latency = (64 * (*prec_mult) * 4) / entries;
 765
 766         if (*drain_latency > DRAIN_LATENCY_MASK)
 767                 *drain_latency = DRAIN_LATENCY_MASK;
 768
 769         return true;
 770 }
 771
 772 /*
 773  * Update drain latency registers of memory arbiter
 774  *
 775  * Valleyview SoC has a new memory arbiter and needs drain latency registers
 776  * to be programmed. Each plane has a drain latency multiplier and a drain
 777  * latency value.
 778  */
 779
 780 static void vlv_update_drain_latency(struct drm_crtc *crtc)
 781 {
 782         struct drm_device *dev = crtc->dev;
 783         struct drm_i915_private *dev_priv = dev->dev_private;
 784         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
 785         int pixel_size;
 786         int drain_latency;
 787         enum pipe pipe = intel_crtc->pipe;
 788         int plane_prec, prec_mult, plane_dl;
 789         const int high_precision = IS_CHERRYVIEW(dev) ?
 790                 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_64;
 791
 792         plane_dl = I915_READ(VLV_DDL(pipe)) & ~(DDL_PLANE_PRECISION_HIGH |
 793                    DRAIN_LATENCY_MASK | DDL_CURSOR_PRECISION_HIGH |
 794                    (DRAIN_LATENCY_MASK << DDL_CURSOR_SHIFT));
 795
 796         if (!intel_crtc_active(crtc)) {
 797                 I915_WRITE(VLV_DDL(pipe), plane_dl);
 798                 return;
 799         }
 800
 801         /* Primary plane Drain Latency */
 802         pixel_size = crtc->primary->fb->bits_per_pixel / 8;     /* BPP */
 803         if (vlv_compute_drain_latency(crtc, pixel_size, &prec_mult, &drain_latency)) {
 804                 plane_prec = (prec_mult == high_precision) ?
 805                                            DDL_PLANE_PRECISION_HIGH :
 806                                            DDL_PLANE_PRECISION_LOW;
 807                 plane_dl |= plane_prec | drain_latency;
 808         }
 809
 810         /* Cursor Drain Latency
 811          * BPP is always 4 for cursor
 812          */
 813         pixel_size = 4;
 814
 815         /* Program cursor DL only if it is enabled */
 816         if (intel_crtc->cursor_base &&
 817             vlv_compute_drain_latency(crtc, pixel_size, &prec_mult, &drain_latency)) {
 818                 plane_prec = (prec_mult == high_precision) ?
 819                                            DDL_CURSOR_PRECISION_HIGH :
 820                                            DDL_CURSOR_PRECISION_LOW;
 821                 plane_dl |= plane_prec | (drain_latency << DDL_CURSOR_SHIFT);
 822         }
 823
 824         I915_WRITE(VLV_DDL(pipe), plane_dl);
 825 }
 826
 827 #define single_plane_enabled(mask) is_power_of_2(mask)
 828
 829 static void valleyview_update_wm(struct drm_crtc *crtc)
 830 {
 831         struct drm_device *dev = crtc->dev;
 832         static const int sr_latency_ns = 12000;
 833         struct drm_i915_private *dev_priv = dev->dev_private;
 834         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
 835         int plane_sr, cursor_sr;
 836         int ignore_plane_sr, ignore_cursor_sr;
 837         unsigned int enabled = 0;
 838         bool cxsr_enabled;
 839
 840         vlv_update_drain_latency(crtc);
 841
 842         if (g4x_compute_wm0(dev, PIPE_A,
 843                             &valleyview_wm_info, pessimal_latency_ns,
 844                             &valleyview_cursor_wm_info, pessimal_latency_ns,
 845                             &planea_wm, &cursora_wm))
 846                 enabled |= 1 << PIPE_A;
 847
 848         if (g4x_compute_wm0(dev, PIPE_B,
 849                             &valleyview_wm_info, pessimal_latency_ns,
 850                             &valleyview_cursor_wm_info, pessimal_latency_ns,
 851                             &planeb_wm, &cursorb_wm))
 852                 enabled |= 1 << PIPE_B;
 853
 854         if (single_plane_enabled(enabled) &&
 855             g4x_compute_srwm(dev, ffs(enabled) - 1,
 856                              sr_latency_ns,
 857                              &valleyview_wm_info,
 858                              &valleyview_cursor_wm_info,
 859                              &plane_sr, &ignore_cursor_sr) &&
 860             g4x_compute_srwm(dev, ffs(enabled) - 1,
 861                              2*sr_latency_ns,
 862                              &valleyview_wm_info,
 863                              &valleyview_cursor_wm_info,
 864                              &ignore_plane_sr, &cursor_sr)) {
 865                 cxsr_enabled = true;
 866         } else {
 867                 cxsr_enabled = false;
 868                 intel_set_memory_cxsr(dev_priv, false);
 869                 plane_sr = cursor_sr = 0;
 870         }
 871
 872         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, "
 873                       "B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
 874                       planea_wm, cursora_wm,
 875                       planeb_wm, cursorb_wm,
 876                       plane_sr, cursor_sr);
 877
 878         I915_WRITE(DSPFW1,
 879                    (plane_sr << DSPFW_SR_SHIFT) |
 880                    (cursorb_wm << DSPFW_CURSORB_SHIFT) |
 881                    (planeb_wm << DSPFW_PLANEB_SHIFT) |
 882                    (planea_wm << DSPFW_PLANEA_SHIFT));
 883         I915_WRITE(DSPFW2,
 884                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
 885                    (cursora_wm << DSPFW_CURSORA_SHIFT));
 886         I915_WRITE(DSPFW3,
 887                    (I915_READ(DSPFW3) & ~DSPFW_CURSOR_SR_MASK) |
 888                    (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
 889
 890         if (cxsr_enabled)
 891                 intel_set_memory_cxsr(dev_priv, true);
 892 }
 893
 894 static void cherryview_update_wm(struct drm_crtc *crtc)
 895 {
 896         struct drm_device *dev = crtc->dev;
 897         static const int sr_latency_ns = 12000;
 898         struct drm_i915_private *dev_priv = dev->dev_private;
 899         int planea_wm, planeb_wm, planec_wm;
 900         int cursora_wm, cursorb_wm, cursorc_wm;
 901         int plane_sr, cursor_sr;
 902         int ignore_plane_sr, ignore_cursor_sr;
 903         unsigned int enabled = 0;
 904         bool cxsr_enabled;
 905
 906         vlv_update_drain_latency(crtc);
 907
 908         if (g4x_compute_wm0(dev, PIPE_A,
 909                             &valleyview_wm_info, pessimal_latency_ns,
 910                             &valleyview_cursor_wm_info, pessimal_latency_ns,
 911                             &planea_wm, &cursora_wm))
 912                 enabled |= 1 << PIPE_A;
 913
 914         if (g4x_compute_wm0(dev, PIPE_B,
 915                             &valleyview_wm_info, pessimal_latency_ns,
 916                             &valleyview_cursor_wm_info, pessimal_latency_ns,
 917                             &planeb_wm, &cursorb_wm))
 918                 enabled |= 1 << PIPE_B;
 919
 920         if (g4x_compute_wm0(dev, PIPE_C,
 921                             &valleyview_wm_info, pessimal_latency_ns,
 922                             &valleyview_cursor_wm_info, pessimal_latency_ns,
 923                             &planec_wm, &cursorc_wm))
 924                 enabled |= 1 << PIPE_C;
 925
 926         if (single_plane_enabled(enabled) &&
 927             g4x_compute_srwm(dev, ffs(enabled) - 1,
 928                              sr_latency_ns,
 929                              &valleyview_wm_info,
 930                              &valleyview_cursor_wm_info,
 931                              &plane_sr, &ignore_cursor_sr) &&
 932             g4x_compute_srwm(dev, ffs(enabled) - 1,
 933                              2*sr_latency_ns,
 934                              &valleyview_wm_info,
 935                              &valleyview_cursor_wm_info,
 936                              &ignore_plane_sr, &cursor_sr)) {
 937                 cxsr_enabled = true;
 938         } else {
 939                 cxsr_enabled = false;
 940                 intel_set_memory_cxsr(dev_priv, false);
 941                 plane_sr = cursor_sr = 0;
 942         }
 943
 944         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, "
 945                       "B: plane=%d, cursor=%d, C: plane=%d, cursor=%d, "
 946                       "SR: plane=%d, cursor=%d\n",
 947                       planea_wm, cursora_wm,
 948                       planeb_wm, cursorb_wm,
 949                       planec_wm, cursorc_wm,
 950                       plane_sr, cursor_sr);
 951
 952         I915_WRITE(DSPFW1,
 953                    (plane_sr << DSPFW_SR_SHIFT) |
 954                    (cursorb_wm << DSPFW_CURSORB_SHIFT) |
 955                    (planeb_wm << DSPFW_PLANEB_SHIFT) |
 956                    (planea_wm << DSPFW_PLANEA_SHIFT));
 957         I915_WRITE(DSPFW2,
 958                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
 959                    (cursora_wm << DSPFW_CURSORA_SHIFT));
 960         I915_WRITE(DSPFW3,
 961                    (I915_READ(DSPFW3) & ~DSPFW_CURSOR_SR_MASK) |
 962                    (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
 963         I915_WRITE(DSPFW9_CHV,
 964                    (I915_READ(DSPFW9_CHV) & ~(DSPFW_PLANEC_MASK |
 965                                               DSPFW_CURSORC_MASK)) |
 966                    (planec_wm << DSPFW_PLANEC_SHIFT) |
 967                    (cursorc_wm << DSPFW_CURSORC_SHIFT));
 968
 969         if (cxsr_enabled)
 970                 intel_set_memory_cxsr(dev_priv, true);
 971 }
 972
 973 static void valleyview_update_sprite_wm(struct drm_plane *plane,
 974                                         struct drm_crtc *crtc,
 975                                         uint32_t sprite_width,
 976                                         uint32_t sprite_height,
 977                                         int pixel_size,
 978                                         bool enabled, bool scaled)
 979 {
 980         struct drm_device *dev = crtc->dev;
 981         struct drm_i915_private *dev_priv = dev->dev_private;
 982         int pipe = to_intel_plane(plane)->pipe;
 983         int sprite = to_intel_plane(plane)->plane;
 984         int drain_latency;
 985         int plane_prec;
 986         int sprite_dl;
 987         int prec_mult;
 988         const int high_precision = IS_CHERRYVIEW(dev) ?
 989                 DRAIN_LATENCY_PRECISION_32 : DRAIN_LATENCY_PRECISION_64;
 990
 991         sprite_dl = I915_READ(VLV_DDL(pipe)) & ~(DDL_SPRITE_PRECISION_HIGH(sprite) |
 992                     (DRAIN_LATENCY_MASK << DDL_SPRITE_SHIFT(sprite)));
 993
 994         if (enabled && vlv_compute_drain_latency(crtc, pixel_size, &prec_mult,
 995                                                  &drain_latency)) {
 996                 plane_prec = (prec_mult == high_precision) ?
 997                                            DDL_SPRITE_PRECISION_HIGH(sprite) :
 998                                            DDL_SPRITE_PRECISION_LOW(sprite);
 999                 sprite_dl |= plane_prec |
1000                              (drain_latency << DDL_SPRITE_SHIFT(sprite));
1001         }
1002
1003         I915_WRITE(VLV_DDL(pipe), sprite_dl);
1004 }
1005
1006 static void g4x_update_wm(struct drm_crtc *crtc)
1007 {
1008         struct drm_device *dev = crtc->dev;
1009         static const int sr_latency_ns = 12000;
1010         struct drm_i915_private *dev_priv = dev->dev_private;
1011         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1012         int plane_sr, cursor_sr;
1013         unsigned int enabled = 0;
1014         bool cxsr_enabled;
1015
1016         if (g4x_compute_wm0(dev, PIPE_A,
1017                             &g4x_wm_info, pessimal_latency_ns,
1018                             &g4x_cursor_wm_info, pessimal_latency_ns,
1019                             &planea_wm, &cursora_wm))
1020                 enabled |= 1 << PIPE_A;
1021
1022         if (g4x_compute_wm0(dev, PIPE_B,
1023                             &g4x_wm_info, pessimal_latency_ns,
1024                             &g4x_cursor_wm_info, pessimal_latency_ns,
1025                             &planeb_wm, &cursorb_wm))
1026                 enabled |= 1 << PIPE_B;
1027
1028         if (single_plane_enabled(enabled) &&
1029             g4x_compute_srwm(dev, ffs(enabled) - 1,
1030                              sr_latency_ns,
1031                              &g4x_wm_info,
1032                              &g4x_cursor_wm_info,
1033                              &plane_sr, &cursor_sr)) {
1034                 cxsr_enabled = true;
1035         } else {
1036                 cxsr_enabled = false;
1037                 intel_set_memory_cxsr(dev_priv, false);
1038                 plane_sr = cursor_sr = 0;
1039         }
1040
1041         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, "
1042                       "B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1043                       planea_wm, cursora_wm,
1044                       planeb_wm, cursorb_wm,
1045                       plane_sr, cursor_sr);
1046
1047         I915_WRITE(DSPFW1,
1048                    (plane_sr << DSPFW_SR_SHIFT) |
1049                    (cursorb_wm << DSPFW_CURSORB_SHIFT) |
1050                    (planeb_wm << DSPFW_PLANEB_SHIFT) |
1051                    (planea_wm << DSPFW_PLANEA_SHIFT));
1052         I915_WRITE(DSPFW2,
1053                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1054                    (cursora_wm << DSPFW_CURSORA_SHIFT));
1055         /* HPLL off in SR has some issues on G4x... disable it */
1056         I915_WRITE(DSPFW3,
1057                    (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1058                    (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1059
1060         if (cxsr_enabled)
1061                 intel_set_memory_cxsr(dev_priv, true);
1062 }
1063
1064 static void i965_update_wm(struct drm_crtc *unused_crtc)
1065 {
1066         struct drm_device *dev = unused_crtc->dev;
1067         struct drm_i915_private *dev_priv = dev->dev_private;
1068         struct drm_crtc *crtc;
1069         int srwm = 1;
1070         int cursor_sr = 16;
1071         bool cxsr_enabled;
1072
1073         /* Calc sr entries for one plane configs */
1074         crtc = single_enabled_crtc(dev);
1075         if (crtc) {
1076                 /* self-refresh has much higher latency */
1077                 static const int sr_latency_ns = 12000;
1078                 const struct drm_display_mode *adjusted_mode =
1079                         &to_intel_crtc(crtc)->config->base.adjusted_mode;
1080                 int clock = adjusted_mode->crtc_clock;
1081                 int htotal = adjusted_mode->crtc_htotal;
1082                 int hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
1083                 int pixel_size = crtc->primary->fb->bits_per_pixel / 8;
1084                 unsigned long line_time_us;
1085                 int entries;
1086
1087                 line_time_us = max(htotal * 1000 / clock, 1);
1088
1089                 /* Use ns/us then divide to preserve precision */
1090                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1091                         pixel_size * hdisplay;
1092                 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1093                 srwm = I965_FIFO_SIZE - entries;
1094                 if (srwm < 0)
1095                         srwm = 1;
1096                 srwm &= 0x1ff;
1097                 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1098                               entries, srwm);
1099
1100                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1101                         pixel_size * to_intel_crtc(crtc)->cursor_width;
1102                 entries = DIV_ROUND_UP(entries,
1103                                           i965_cursor_wm_info.cacheline_size);
1104                 cursor_sr = i965_cursor_wm_info.fifo_size -
1105                         (entries + i965_cursor_wm_info.guard_size);
1106
1107                 if (cursor_sr > i965_cursor_wm_info.max_wm)
1108                         cursor_sr = i965_cursor_wm_info.max_wm;
1109
1110                 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1111                               "cursor %d\n", srwm, cursor_sr);
1112
1113                 cxsr_enabled = true;
1114         } else {
1115                 cxsr_enabled = false;
1116                 /* Turn off self refresh if both pipes are enabled */
1117                 intel_set_memory_cxsr(dev_priv, false);
1118         }
1119
1120         DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1121                       srwm);
1122
1123         /* 965 has limitations... */
1124         I915_WRITE(DSPFW1, (srwm << DSPFW_SR_SHIFT) |
1125                    (8 << DSPFW_CURSORB_SHIFT) |
1126                    (8 << DSPFW_PLANEB_SHIFT) |
1127                    (8 << DSPFW_PLANEA_SHIFT));
1128         I915_WRITE(DSPFW2, (8 << DSPFW_CURSORA_SHIFT) |
1129                    (8 << DSPFW_PLANEC_SHIFT_OLD));
1130         /* update cursor SR watermark */
1131         I915_WRITE(DSPFW3, (cursor_sr << DSPFW_CURSOR_SR_SHIFT));
1132
1133         if (cxsr_enabled)
1134                 intel_set_memory_cxsr(dev_priv, true);
1135 }
1136
1137 static void i9xx_update_wm(struct drm_crtc *unused_crtc)
1138 {
1139         struct drm_device *dev = unused_crtc->dev;
1140         struct drm_i915_private *dev_priv = dev->dev_private;
1141         const struct intel_watermark_params *wm_info;
1142         uint32_t fwater_lo;
1143         uint32_t fwater_hi;
1144         int cwm, srwm = 1;
1145         int fifo_size;
1146         int planea_wm, planeb_wm;
1147         struct drm_crtc *crtc, *enabled = NULL;
1148
1149         if (IS_I945GM(dev))
1150                 wm_info = &i945_wm_info;
1151         else if (!IS_GEN2(dev))
1152                 wm_info = &i915_wm_info;
1153         else
1154                 wm_info = &i830_a_wm_info;
1155
1156         fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1157         crtc = intel_get_crtc_for_plane(dev, 0);
1158         if (intel_crtc_active(crtc)) {
1159                 const struct drm_display_mode *adjusted_mode;
1160                 int cpp = crtc->primary->fb->bits_per_pixel / 8;
1161                 if (IS_GEN2(dev))
1162                         cpp = 4;
1163
1164                 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1165                 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1166                                                wm_info, fifo_size, cpp,
1167                                                pessimal_latency_ns);
1168                 enabled = crtc;
1169         } else {
1170                 planea_wm = fifo_size - wm_info->guard_size;
1171                 if (planea_wm > (long)wm_info->max_wm)
1172                         planea_wm = wm_info->max_wm;
1173         }
1174
1175         if (IS_GEN2(dev))
1176                 wm_info = &i830_bc_wm_info;
1177
1178         fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1179         crtc = intel_get_crtc_for_plane(dev, 1);
1180         if (intel_crtc_active(crtc)) {
1181                 const struct drm_display_mode *adjusted_mode;
1182                 int cpp = crtc->primary->fb->bits_per_pixel / 8;
1183                 if (IS_GEN2(dev))
1184                         cpp = 4;
1185
1186                 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1187                 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1188                                                wm_info, fifo_size, cpp,
1189                                                pessimal_latency_ns);
1190                 if (enabled == NULL)
1191                         enabled = crtc;
1192                 else
1193                         enabled = NULL;
1194         } else {
1195                 planeb_wm = fifo_size - wm_info->guard_size;
1196                 if (planeb_wm > (long)wm_info->max_wm)
1197                         planeb_wm = wm_info->max_wm;
1198         }
1199
1200         DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1201
1202         if (IS_I915GM(dev) && enabled) {
1203                 struct drm_i915_gem_object *obj;
1204
1205                 obj = intel_fb_obj(enabled->primary->fb);
1206
1207                 /* self-refresh seems busted with untiled */
1208                 if (obj->tiling_mode == I915_TILING_NONE)
1209                         enabled = NULL;
1210         }
1211
1212         /*
1213          * Overlay gets an aggressive default since video jitter is bad.
1214          */
1215         cwm = 2;
1216
1217         /* Play safe and disable self-refresh before adjusting watermarks. */
1218         intel_set_memory_cxsr(dev_priv, false);
1219
1220         /* Calc sr entries for one plane configs */
1221         if (HAS_FW_BLC(dev) && enabled) {
1222                 /* self-refresh has much higher latency */
1223                 static const int sr_latency_ns = 6000;
1224                 const struct drm_display_mode *adjusted_mode =
1225                         &to_intel_crtc(enabled)->config->base.adjusted_mode;
1226                 int clock = adjusted_mode->crtc_clock;
1227                 int htotal = adjusted_mode->crtc_htotal;
1228                 int hdisplay = to_intel_crtc(enabled)->config->pipe_src_w;
1229                 int pixel_size = enabled->primary->fb->bits_per_pixel / 8;
1230                 unsigned long line_time_us;
1231                 int entries;
1232
1233                 line_time_us = max(htotal * 1000 / clock, 1);
1234
1235                 /* Use ns/us then divide to preserve precision */
1236                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1237                         pixel_size * hdisplay;
1238                 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1239                 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1240                 srwm = wm_info->fifo_size - entries;
1241                 if (srwm < 0)
1242                         srwm = 1;
1243
1244                 if (IS_I945G(dev) || IS_I945GM(dev))
1245                         I915_WRITE(FW_BLC_SELF,
1246                                    FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1247                 else if (IS_I915GM(dev))
1248                         I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1249         }
1250
1251         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1252                       planea_wm, planeb_wm, cwm, srwm);
1253
1254         fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1255         fwater_hi = (cwm & 0x1f);
1256
1257         /* Set request length to 8 cachelines per fetch */
1258         fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1259         fwater_hi = fwater_hi | (1 << 8);
1260
1261         I915_WRITE(FW_BLC, fwater_lo);
1262         I915_WRITE(FW_BLC2, fwater_hi);
1263
1264         if (enabled)
1265                 intel_set_memory_cxsr(dev_priv, true);
1266 }
1267
1268 static void i845_update_wm(struct drm_crtc *unused_crtc)
1269 {
1270         struct drm_device *dev = unused_crtc->dev;
1271         struct drm_i915_private *dev_priv = dev->dev_private;
1272         struct drm_crtc *crtc;
1273         const struct drm_display_mode *adjusted_mode;
1274         uint32_t fwater_lo;
1275         int planea_wm;
1276
1277         crtc = single_enabled_crtc(dev);
1278         if (crtc == NULL)
1279                 return;
1280
1281         adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1282         planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1283                                        &i845_wm_info,
1284                                        dev_priv->display.get_fifo_size(dev, 0),
1285                                        4, pessimal_latency_ns);
1286         fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1287         fwater_lo |= (3<<8) | planea_wm;
1288
1289         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1290
1291         I915_WRITE(FW_BLC, fwater_lo);
1292 }
1293
1294 static uint32_t ilk_pipe_pixel_rate(struct drm_device *dev,
1295                                     struct drm_crtc *crtc)
1296 {
1297         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1298         uint32_t pixel_rate;
1299
1300         pixel_rate = intel_crtc->config->base.adjusted_mode.crtc_clock;
1301
1302         /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
1303          * adjust the pixel_rate here. */
1304
1305         if (intel_crtc->config->pch_pfit.enabled) {
1306                 uint64_t pipe_w, pipe_h, pfit_w, pfit_h;
1307                 uint32_t pfit_size = intel_crtc->config->pch_pfit.size;
1308
1309                 pipe_w = intel_crtc->config->pipe_src_w;
1310                 pipe_h = intel_crtc->config->pipe_src_h;
1311                 pfit_w = (pfit_size >> 16) & 0xFFFF;
1312                 pfit_h = pfit_size & 0xFFFF;
1313                 if (pipe_w < pfit_w)
1314                         pipe_w = pfit_w;
1315                 if (pipe_h < pfit_h)
1316                         pipe_h = pfit_h;
1317
1318                 pixel_rate = div_u64((uint64_t) pixel_rate * pipe_w * pipe_h,
1319                                      pfit_w * pfit_h);
1320         }
1321
1322         return pixel_rate;
1323 }
1324
1325 /* latency must be in 0.1us units. */
1326 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
1327                                uint32_t latency)
1328 {
1329         uint64_t ret;
1330
1331         if (WARN(latency == 0, "Latency value missing\n"))
1332                 return UINT_MAX;
1333
1334         ret = (uint64_t) pixel_rate * bytes_per_pixel * latency;
1335         ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
1336
1337         return ret;
1338 }
1339
1340 /* latency must be in 0.1us units. */
1341 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
1342                                uint32_t horiz_pixels, uint8_t bytes_per_pixel,
1343                                uint32_t latency)
1344 {
1345         uint32_t ret;
1346
1347         if (WARN(latency == 0, "Latency value missing\n"))
1348                 return UINT_MAX;
1349
1350         ret = (latency * pixel_rate) / (pipe_htotal * 10000);
1351         ret = (ret + 1) * horiz_pixels * bytes_per_pixel;
1352         ret = DIV_ROUND_UP(ret, 64) + 2;
1353         return ret;
1354 }
1355
1356 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
1357                            uint8_t bytes_per_pixel)
1358 {
1359         return DIV_ROUND_UP(pri_val * 64, horiz_pixels * bytes_per_pixel) + 2;
1360 }
1361
1362 struct skl_pipe_wm_parameters {
1363         bool active;
1364         uint32_t pipe_htotal;
1365         uint32_t pixel_rate; /* in KHz */
1366         struct intel_plane_wm_parameters plane[I915_MAX_PLANES];
1367         struct intel_plane_wm_parameters cursor;
1368 };
1369
1370 struct ilk_pipe_wm_parameters {
1371         bool active;
1372         uint32_t pipe_htotal;
1373         uint32_t pixel_rate;
1374         struct intel_plane_wm_parameters pri;
1375         struct intel_plane_wm_parameters spr;
1376         struct intel_plane_wm_parameters cur;
1377 };
1378
1379 struct ilk_wm_maximums {
1380         uint16_t pri;
1381         uint16_t spr;
1382         uint16_t cur;
1383         uint16_t fbc;
1384 };
1385
1386 /* used in computing the new watermarks state */
1387 struct intel_wm_config {
1388         unsigned int num_pipes_active;
1389         bool sprites_enabled;
1390         bool sprites_scaled;
1391 };
1392
1393 /*
1394  * For both WM_PIPE and WM_LP.
1395  * mem_value must be in 0.1us units.
1396  */
1397 static uint32_t ilk_compute_pri_wm(const struct ilk_pipe_wm_parameters *params,
1398                                    uint32_t mem_value,
1399                                    bool is_lp)
1400 {
1401         uint32_t method1, method2;
1402
1403         if (!params->active || !params->pri.enabled)
1404                 return 0;
1405
1406         method1 = ilk_wm_method1(params->pixel_rate,
1407                                  params->pri.bytes_per_pixel,
1408                                  mem_value);
1409
1410         if (!is_lp)
1411                 return method1;
1412
1413         method2 = ilk_wm_method2(params->pixel_rate,
1414                                  params->pipe_htotal,
1415                                  params->pri.horiz_pixels,
1416                                  params->pri.bytes_per_pixel,
1417                                  mem_value);
1418
1419         return min(method1, method2);
1420 }
1421
1422 /*
1423  * For both WM_PIPE and WM_LP.
1424  * mem_value must be in 0.1us units.
1425  */
1426 static uint32_t ilk_compute_spr_wm(const struct ilk_pipe_wm_parameters *params,
1427                                    uint32_t mem_value)
1428 {
1429         uint32_t method1, method2;
1430
1431         if (!params->active || !params->spr.enabled)
1432                 return 0;
1433
1434         method1 = ilk_wm_method1(params->pixel_rate,
1435                                  params->spr.bytes_per_pixel,
1436                                  mem_value);
1437         method2 = ilk_wm_method2(params->pixel_rate,
1438                                  params->pipe_htotal,
1439                                  params->spr.horiz_pixels,
1440                                  params->spr.bytes_per_pixel,
1441                                  mem_value);
1442         return min(method1, method2);
1443 }
1444
1445 /*
1446  * For both WM_PIPE and WM_LP.
1447  * mem_value must be in 0.1us units.
1448  */
1449 static uint32_t ilk_compute_cur_wm(const struct ilk_pipe_wm_parameters *params,
1450                                    uint32_t mem_value)
1451 {
1452         if (!params->active || !params->cur.enabled)
1453                 return 0;
1454
1455         return ilk_wm_method2(params->pixel_rate,
1456                               params->pipe_htotal,
1457                               params->cur.horiz_pixels,
1458                               params->cur.bytes_per_pixel,
1459                               mem_value);
1460 }
1461
1462 /* Only for WM_LP. */
1463 static uint32_t ilk_compute_fbc_wm(const struct ilk_pipe_wm_parameters *params,
1464                                    uint32_t pri_val)
1465 {
1466         if (!params->active || !params->pri.enabled)
1467                 return 0;
1468
1469         return ilk_wm_fbc(pri_val,
1470                           params->pri.horiz_pixels,
1471                           params->pri.bytes_per_pixel);
1472 }
1473
1474 static unsigned int ilk_display_fifo_size(const struct drm_device *dev)
1475 {
1476         if (INTEL_INFO(dev)->gen >= 8)
1477                 return 3072;
1478         else if (INTEL_INFO(dev)->gen >= 7)
1479                 return 768;
1480         else
1481                 return 512;
1482 }
1483
1484 static unsigned int ilk_plane_wm_reg_max(const struct drm_device *dev,
1485                                          int level, bool is_sprite)
1486 {
1487         if (INTEL_INFO(dev)->gen >= 8)
1488                 /* BDW primary/sprite plane watermarks */
1489                 return level == 0 ? 255 : 2047;
1490         else if (INTEL_INFO(dev)->gen >= 7)
1491                 /* IVB/HSW primary/sprite plane watermarks */
1492                 return level == 0 ? 127 : 1023;
1493         else if (!is_sprite)
1494                 /* ILK/SNB primary plane watermarks */
1495                 return level == 0 ? 127 : 511;
1496         else
1497                 /* ILK/SNB sprite plane watermarks */
1498                 return level == 0 ? 63 : 255;
1499 }
1500
1501 static unsigned int ilk_cursor_wm_reg_max(const struct drm_device *dev,
1502                                           int level)
1503 {
1504         if (INTEL_INFO(dev)->gen >= 7)
1505                 return level == 0 ? 63 : 255;
1506         else
1507                 return level == 0 ? 31 : 63;
1508 }
1509
1510 static unsigned int ilk_fbc_wm_reg_max(const struct drm_device *dev)
1511 {
1512         if (INTEL_INFO(dev)->gen >= 8)
1513                 return 31;
1514         else
1515                 return 15;
1516 }
1517
1518 /* Calculate the maximum primary/sprite plane watermark */
1519 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
1520                                      int level,
1521                                      const struct intel_wm_config *config,
1522                                      enum intel_ddb_partitioning ddb_partitioning,
1523                                      bool is_sprite)
1524 {
1525         unsigned int fifo_size = ilk_display_fifo_size(dev);
1526
1527         /* if sprites aren't enabled, sprites get nothing */
1528         if (is_sprite && !config->sprites_enabled)
1529                 return 0;
1530
1531         /* HSW allows LP1+ watermarks even with multiple pipes */
1532         if (level == 0 || config->num_pipes_active > 1) {
1533                 fifo_size /= INTEL_INFO(dev)->num_pipes;
1534
1535                 /*
1536                  * For some reason the non self refresh
1537                  * FIFO size is only half of the self
1538                  * refresh FIFO size on ILK/SNB.
1539                  */
1540                 if (INTEL_INFO(dev)->gen <= 6)
1541                         fifo_size /= 2;
1542         }
1543
1544         if (config->sprites_enabled) {
1545                 /* level 0 is always calculated with 1:1 split */
1546                 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
1547                         if (is_sprite)
1548                                 fifo_size *= 5;
1549                         fifo_size /= 6;
1550                 } else {
1551                         fifo_size /= 2;
1552                 }
1553         }
1554
1555         /* clamp to max that the registers can hold */
1556         return min(fifo_size, ilk_plane_wm_reg_max(dev, level, is_sprite));
1557 }
1558
1559 /* Calculate the maximum cursor plane watermark */
1560 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
1561                                       int level,
1562                                       const struct intel_wm_config *config)
1563 {
1564         /* HSW LP1+ watermarks w/ multiple pipes */
1565         if (level > 0 && config->num_pipes_active > 1)
1566                 return 64;
1567
1568         /* otherwise just report max that registers can hold */
1569         return ilk_cursor_wm_reg_max(dev, level);
1570 }
1571
1572 static void ilk_compute_wm_maximums(const struct drm_device *dev,
1573                                     int level,
1574                                     const struct intel_wm_config *config,
1575                                     enum intel_ddb_partitioning ddb_partitioning,
1576                                     struct ilk_wm_maximums *max)
1577 {
1578         max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
1579         max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
1580         max->cur = ilk_cursor_wm_max(dev, level, config);
1581         max->fbc = ilk_fbc_wm_reg_max(dev);
1582 }
1583
1584 static void ilk_compute_wm_reg_maximums(struct drm_device *dev,
1585                                         int level,
1586                                         struct ilk_wm_maximums *max)
1587 {
1588         max->pri = ilk_plane_wm_reg_max(dev, level, false);
1589         max->spr = ilk_plane_wm_reg_max(dev, level, true);
1590         max->cur = ilk_cursor_wm_reg_max(dev, level);
1591         max->fbc = ilk_fbc_wm_reg_max(dev);
1592 }
1593
1594 static bool ilk_validate_wm_level(int level,
1595                                   const struct ilk_wm_maximums *max,
1596                                   struct intel_wm_level *result)
1597 {
1598         bool ret;
1599
1600         /* already determined to be invalid? */
1601         if (!result->enable)
1602                 return false;
1603
1604         result->enable = result->pri_val <= max->pri &&
1605                          result->spr_val <= max->spr &&
1606                          result->cur_val <= max->cur;
1607
1608         ret = result->enable;
1609
1610         /*
1611          * HACK until we can pre-compute everything,
1612          * and thus fail gracefully if LP0 watermarks
1613          * are exceeded...
1614          */
1615         if (level == 0 && !result->enable) {
1616                 if (result->pri_val > max->pri)
1617                         DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
1618                                       level, result->pri_val, max->pri);
1619                 if (result->spr_val > max->spr)
1620                         DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
1621                                       level, result->spr_val, max->spr);
1622                 if (result->cur_val > max->cur)
1623                         DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
1624                                       level, result->cur_val, max->cur);
1625
1626                 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
1627                 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
1628                 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
1629                 result->enable = true;
1630         }
1631
1632         return ret;
1633 }
1634
1635 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
1636                                  int level,
1637                                  const struct ilk_pipe_wm_parameters *p,
1638                                  struct intel_wm_level *result)
1639 {
1640         uint16_t pri_latency = dev_priv->wm.pri_latency[level];
1641         uint16_t spr_latency = dev_priv->wm.spr_latency[level];
1642         uint16_t cur_latency = dev_priv->wm.cur_latency[level];
1643
1644         /* WM1+ latency values stored in 0.5us units */
1645         if (level > 0) {
1646                 pri_latency *= 5;
1647                 spr_latency *= 5;
1648                 cur_latency *= 5;
1649         }
1650
1651         result->pri_val = ilk_compute_pri_wm(p, pri_latency, level);
1652         result->spr_val = ilk_compute_spr_wm(p, spr_latency);
1653         result->cur_val = ilk_compute_cur_wm(p, cur_latency);
1654         result->fbc_val = ilk_compute_fbc_wm(p, result->pri_val);
1655         result->enable = true;
1656 }
1657
1658 static uint32_t
1659 hsw_compute_linetime_wm(struct drm_device *dev, struct drm_crtc *crtc)
1660 {
1661         struct drm_i915_private *dev_priv = dev->dev_private;
1662         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1663         struct drm_display_mode *mode = &intel_crtc->config->base.adjusted_mode;
1664         u32 linetime, ips_linetime;
1665
1666         if (!intel_crtc_active(crtc))
1667                 return 0;
1668
1669         /* The WM are computed with base on how long it takes to fill a single
1670          * row at the given clock rate, multiplied by 8.
1671          * */
1672         linetime = DIV_ROUND_CLOSEST(mode->crtc_htotal * 1000 * 8,
1673                                      mode->crtc_clock);
1674         ips_linetime = DIV_ROUND_CLOSEST(mode->crtc_htotal * 1000 * 8,
1675                                          intel_ddi_get_cdclk_freq(dev_priv));
1676
1677         return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
1678                PIPE_WM_LINETIME_TIME(linetime);
1679 }
1680
1681 static void intel_read_wm_latency(struct drm_device *dev, uint16_t wm[8])
1682 {
1683         struct drm_i915_private *dev_priv = dev->dev_private;
1684
1685         if (IS_GEN9(dev)) {
1686                 uint32_t val;
1687                 int ret, i;
1688                 int level, max_level = ilk_wm_max_level(dev);
1689
1690                 /* read the first set of memory latencies[0:3] */
1691                 val = 0; /* data0 to be programmed to 0 for first set */
1692                 mutex_lock(&dev_priv->rps.hw_lock);
1693                 ret = sandybridge_pcode_read(dev_priv,
1694                                              GEN9_PCODE_READ_MEM_LATENCY,
1695                                              &val);
1696                 mutex_unlock(&dev_priv->rps.hw_lock);
1697
1698                 if (ret) {
1699                         DRM_ERROR("SKL Mailbox read error = %d\n", ret);
1700                         return;
1701                 }
1702
1703                 wm[0] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
1704                 wm[1] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
1705                                 GEN9_MEM_LATENCY_LEVEL_MASK;
1706                 wm[2] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
1707                                 GEN9_MEM_LATENCY_LEVEL_MASK;
1708                 wm[3] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
1709                                 GEN9_MEM_LATENCY_LEVEL_MASK;
1710
1711                 /* read the second set of memory latencies[4:7] */
1712                 val = 1; /* data0 to be programmed to 1 for second set */
1713                 mutex_lock(&dev_priv->rps.hw_lock);
1714                 ret = sandybridge_pcode_read(dev_priv,
1715                                              GEN9_PCODE_READ_MEM_LATENCY,
1716                                              &val);
1717                 mutex_unlock(&dev_priv->rps.hw_lock);
1718                 if (ret) {
1719                         DRM_ERROR("SKL Mailbox read error = %d\n", ret);
1720                         return;
1721                 }
1722
1723                 wm[4] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
1724                 wm[5] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
1725                                 GEN9_MEM_LATENCY_LEVEL_MASK;
1726                 wm[6] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
1727                                 GEN9_MEM_LATENCY_LEVEL_MASK;
1728                 wm[7] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
1729                                 GEN9_MEM_LATENCY_LEVEL_MASK;
1730
1731                 /*
1732                  * WaWmMemoryReadLatency:skl
1733                  *
1734                  * punit doesn't take into account the read latency so we need
1735                  * to add 2us to the various latency levels we retrieve from
1736                  * the punit.
1737                  *   - W0 is a bit special in that it's the only level that
1738                  *   can't be disabled if we want to have display working, so
1739                  *   we always add 2us there.
1740                  *   - For levels >=1, punit returns 0us latency when they are
1741                  *   disabled, so we respect that and don't add 2us then
1742                  *
1743                  * Additionally, if a level n (n > 1) has a 0us latency, all
1744                  * levels m (m >= n) need to be disabled. We make sure to
1745                  * sanitize the values out of the punit to satisfy this
1746                  * requirement.
1747                  */
1748                 wm[0] += 2;
1749                 for (level = 1; level <= max_level; level++)
1750                         if (wm[level] != 0)
1751                                 wm[level] += 2;
1752                         else {
1753                                 for (i = level + 1; i <= max_level; i++)
1754                                         wm[i] = 0;
1755
1756                                 break;
1757                         }
1758         } else if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
1759                 uint64_t sskpd = I915_READ64(MCH_SSKPD);
1760
1761                 wm[0] = (sskpd >> 56) & 0xFF;
1762                 if (wm[0] == 0)
1763                         wm[0] = sskpd & 0xF;
1764                 wm[1] = (sskpd >> 4) & 0xFF;
1765                 wm[2] = (sskpd >> 12) & 0xFF;
1766                 wm[3] = (sskpd >> 20) & 0x1FF;
1767                 wm[4] = (sskpd >> 32) & 0x1FF;
1768         } else if (INTEL_INFO(dev)->gen >= 6) {
1769                 uint32_t sskpd = I915_READ(MCH_SSKPD);
1770
1771                 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
1772                 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
1773                 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
1774                 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
1775         } else if (INTEL_INFO(dev)->gen >= 5) {
1776                 uint32_t mltr = I915_READ(MLTR_ILK);
1777
1778                 /* ILK primary LP0 latency is 700 ns */
1779                 wm[0] = 7;
1780                 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
1781                 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
1782         }
1783 }
1784
1785 static void intel_fixup_spr_wm_latency(struct drm_device *dev, uint16_t wm[5])
1786 {
1787         /* ILK sprite LP0 latency is 1300 ns */
1788         if (INTEL_INFO(dev)->gen == 5)
1789                 wm[0] = 13;
1790 }
1791
1792 static void intel_fixup_cur_wm_latency(struct drm_device *dev, uint16_t wm[5])
1793 {
1794         /* ILK cursor LP0 latency is 1300 ns */
1795         if (INTEL_INFO(dev)->gen == 5)
1796                 wm[0] = 13;
1797
1798         /* WaDoubleCursorLP3Latency:ivb */
1799         if (IS_IVYBRIDGE(dev))
1800                 wm[3] *= 2;
1801 }
1802
1803 int ilk_wm_max_level(const struct drm_device *dev)
1804 {
1805         /* how many WM levels are we expecting */
1806         if (IS_GEN9(dev))
1807                 return 7;
1808         else if (IS_HASWELL(dev) || IS_BROADWELL(dev))
1809                 return 4;
1810         else if (INTEL_INFO(dev)->gen >= 6)
1811                 return 3;
1812         else
1813                 return 2;
1814 }
1815
1816 static void intel_print_wm_latency(struct drm_device *dev,
1817                                    const char *name,
1818                                    const uint16_t wm[8])
1819 {
1820         int level, max_level = ilk_wm_max_level(dev);
1821
1822         for (level = 0; level <= max_level; level++) {
1823                 unsigned int latency = wm[level];
1824
1825                 if (latency == 0) {
1826                         DRM_ERROR("%s WM%d latency not provided\n",
1827                                   name, level);
1828                         continue;
1829                 }
1830
1831                 /*
1832                  * - latencies are in us on gen9.
1833                  * - before then, WM1+ latency values are in 0.5us units
1834                  */
1835                 if (IS_GEN9(dev))
1836                         latency *= 10;
1837                 else if (level > 0)
1838                         latency *= 5;
1839
1840                 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
1841                               name, level, wm[level],
1842                               latency / 10, latency % 10);
1843         }
1844 }
1845
1846 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
1847                                     uint16_t wm[5], uint16_t min)
1848 {
1849         int level, max_level = ilk_wm_max_level(dev_priv->dev);
1850
1851         if (wm[0] >= min)
1852                 return false;
1853
1854         wm[0] = max(wm[0], min);
1855         for (level = 1; level <= max_level; level++)
1856                 wm[level] = max_t(uint16_t, wm[level], DIV_ROUND_UP(min, 5));
1857
1858         return true;
1859 }
1860
1861 static void snb_wm_latency_quirk(struct drm_device *dev)
1862 {
1863         struct drm_i915_private *dev_priv = dev->dev_private;
1864         bool changed;
1865
1866         /*
1867          * The BIOS provided WM memory latency values are often
1868          * inadequate for high resolution displays. Adjust them.
1869          */
1870         changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
1871                 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
1872                 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
1873
1874         if (!changed)
1875                 return;
1876
1877         DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
1878         intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
1879         intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
1880         intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
1881 }
1882
1883 static void ilk_setup_wm_latency(struct drm_device *dev)
1884 {
1885         struct drm_i915_private *dev_priv = dev->dev_private;
1886
1887         intel_read_wm_latency(dev, dev_priv->wm.pri_latency);
1888
1889         memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
1890                sizeof(dev_priv->wm.pri_latency));
1891         memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
1892                sizeof(dev_priv->wm.pri_latency));
1893
1894         intel_fixup_spr_wm_latency(dev, dev_priv->wm.spr_latency);
1895         intel_fixup_cur_wm_latency(dev, dev_priv->wm.cur_latency);
1896
1897         intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
1898         intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
1899         intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
1900
1901         if (IS_GEN6(dev))
1902                 snb_wm_latency_quirk(dev);
1903 }
1904
1905 static void skl_setup_wm_latency(struct drm_device *dev)
1906 {
1907         struct drm_i915_private *dev_priv = dev->dev_private;
1908
1909         intel_read_wm_latency(dev, dev_priv->wm.skl_latency);
1910         intel_print_wm_latency(dev, "Gen9 Plane", dev_priv->wm.skl_latency);
1911 }
1912
1913 static void ilk_compute_wm_parameters(struct drm_crtc *crtc,
1914                                       struct ilk_pipe_wm_parameters *p)
1915 {
1916         struct drm_device *dev = crtc->dev;
1917         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1918         enum pipe pipe = intel_crtc->pipe;
1919         struct drm_plane *plane;
1920
1921         if (!intel_crtc_active(crtc))
1922                 return;
1923
1924         p->active = true;
1925         p->pipe_htotal = intel_crtc->config->base.adjusted_mode.crtc_htotal;
1926         p->pixel_rate = ilk_pipe_pixel_rate(dev, crtc);
1927         p->pri.bytes_per_pixel = crtc->primary->fb->bits_per_pixel / 8;
1928         p->cur.bytes_per_pixel = 4;
1929         p->pri.horiz_pixels = intel_crtc->config->pipe_src_w;
1930         p->cur.horiz_pixels = intel_crtc->cursor_width;
1931         /* TODO: for now, assume primary and cursor planes are always enabled. */
1932         p->pri.enabled = true;
1933         p->cur.enabled = true;
1934
1935         drm_for_each_legacy_plane(plane, &dev->mode_config.plane_list) {
1936                 struct intel_plane *intel_plane = to_intel_plane(plane);
1937
1938                 if (intel_plane->pipe == pipe) {
1939                         p->spr = intel_plane->wm;
1940                         break;
1941                 }
1942         }
1943 }
1944
1945 static void ilk_compute_wm_config(struct drm_device *dev,
1946                                   struct intel_wm_config *config)
1947 {
1948         struct intel_crtc *intel_crtc;
1949
1950         /* Compute the currently _active_ config */
1951         for_each_intel_crtc(dev, intel_crtc) {
1952                 const struct intel_pipe_wm *wm = &intel_crtc->wm.active;
1953
1954                 if (!wm->pipe_enabled)
1955                         continue;
1956
1957                 config->sprites_enabled |= wm->sprites_enabled;
1958                 config->sprites_scaled |= wm->sprites_scaled;
1959                 config->num_pipes_active++;
1960         }
1961 }
1962
1963 /* Compute new watermarks for the pipe */
1964 static bool intel_compute_pipe_wm(struct drm_crtc *crtc,
1965                                   const struct ilk_pipe_wm_parameters *params,
1966                                   struct intel_pipe_wm *pipe_wm)
1967 {
1968         struct drm_device *dev = crtc->dev;
1969         const struct drm_i915_private *dev_priv = dev->dev_private;
1970         int level, max_level = ilk_wm_max_level(dev);
1971         /* LP0 watermark maximums depend on this pipe alone */
1972         struct intel_wm_config config = {
1973                 .num_pipes_active = 1,
1974                 .sprites_enabled = params->spr.enabled,
1975                 .sprites_scaled = params->spr.scaled,
1976         };
1977         struct ilk_wm_maximums max;
1978
1979         pipe_wm->pipe_enabled = params->active;
1980         pipe_wm->sprites_enabled = params->spr.enabled;
1981         pipe_wm->sprites_scaled = params->spr.scaled;
1982
1983         /* ILK/SNB: LP2+ watermarks only w/o sprites */
1984         if (INTEL_INFO(dev)->gen <= 6 && params->spr.enabled)
1985                 max_level = 1;
1986
1987         /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
1988         if (params->spr.scaled)
1989                 max_level = 0;
1990
1991         ilk_compute_wm_level(dev_priv, 0, params, &pipe_wm->wm[0]);
1992
1993         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
1994                 pipe_wm->linetime = hsw_compute_linetime_wm(dev, crtc);
1995
1996         /* LP0 watermarks always use 1/2 DDB partitioning */
1997         ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
1998
1999         /* At least LP0 must be valid */
2000         if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0]))
2001                 return false;
2002
2003         ilk_compute_wm_reg_maximums(dev, 1, &max);
2004
2005         for (level = 1; level <= max_level; level++) {
2006                 struct intel_wm_level wm = {};
2007
2008                 ilk_compute_wm_level(dev_priv, level, params, &wm);
2009
2010                 /*
2011                  * Disable any watermark level that exceeds the
2012                  * register maximums since such watermarks are
2013                  * always invalid.
2014                  */
2015                 if (!ilk_validate_wm_level(level, &max, &wm))
2016                         break;
2017
2018                 pipe_wm->wm[level] = wm;
2019         }
2020
2021         return true;
2022 }
2023
2024 /*
2025  * Merge the watermarks from all active pipes for a specific level.
2026  */
2027 static void ilk_merge_wm_level(struct drm_device *dev,
2028                                int level,
2029                                struct intel_wm_level *ret_wm)
2030 {
2031         const struct intel_crtc *intel_crtc;
2032
2033         ret_wm->enable = true;
2034
2035         for_each_intel_crtc(dev, intel_crtc) {
2036                 const struct intel_pipe_wm *active = &intel_crtc->wm.active;
2037                 const struct intel_wm_level *wm = &active->wm[level];
2038
2039                 if (!active->pipe_enabled)
2040                         continue;
2041
2042                 /*
2043                  * The watermark values may have been used in the past,
2044                  * so we must maintain them in the registers for some
2045                  * time even if the level is now disabled.
2046                  */
2047                 if (!wm->enable)
2048                         ret_wm->enable = false;
2049
2050                 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
2051                 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
2052                 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
2053                 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
2054         }
2055 }
2056
2057 /*
2058  * Merge all low power watermarks for all active pipes.
2059  */
2060 static void ilk_wm_merge(struct drm_device *dev,
2061                          const struct intel_wm_config *config,
2062                          const struct ilk_wm_maximums *max,
2063                          struct intel_pipe_wm *merged)
2064 {
2065         int level, max_level = ilk_wm_max_level(dev);
2066         int last_enabled_level = max_level;
2067
2068         /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2069         if ((INTEL_INFO(dev)->gen <= 6 || IS_IVYBRIDGE(dev)) &&
2070             config->num_pipes_active > 1)
2071                 return;
2072
2073         /* ILK: FBC WM must be disabled always */
2074         merged->fbc_wm_enabled = INTEL_INFO(dev)->gen >= 6;
2075
2076         /* merge each WM1+ level */
2077         for (level = 1; level <= max_level; level++) {
2078                 struct intel_wm_level *wm = &merged->wm[level];
2079
2080                 ilk_merge_wm_level(dev, level, wm);
2081
2082                 if (level > last_enabled_level)
2083                         wm->enable = false;
2084                 else if (!ilk_validate_wm_level(level, max, wm))
2085                         /* make sure all following levels get disabled */
2086                         last_enabled_level = level - 1;
2087
2088                 /*
2089                  * The spec says it is preferred to disable
2090                  * FBC WMs instead of disabling a WM level.
2091                  */
2092                 if (wm->fbc_val > max->fbc) {
2093                         if (wm->enable)
2094                                 merged->fbc_wm_enabled = false;
2095                         wm->fbc_val = 0;
2096                 }
2097         }
2098
2099         /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2100         /*
2101          * FIXME this is racy. FBC might get enabled later.
2102          * What we should check here is whether FBC can be
2103          * enabled sometime later.
2104          */
2105         if (IS_GEN5(dev) && !merged->fbc_wm_enabled && intel_fbc_enabled(dev)) {
2106                 for (level = 2; level <= max_level; level++) {
2107                         struct intel_wm_level *wm = &merged->wm[level];
2108
2109                         wm->enable = false;
2110                 }
2111         }
2112 }
2113
2114 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
2115 {
2116         /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2117         return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
2118 }
2119
2120 /* The value we need to program into the WM_LPx latency field */
2121 static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
2122 {
2123         struct drm_i915_private *dev_priv = dev->dev_private;
2124
2125         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2126                 return 2 * level;
2127         else
2128                 return dev_priv->wm.pri_latency[level];
2129 }
2130
2131 static void ilk_compute_wm_results(struct drm_device *dev,
2132                                    const struct intel_pipe_wm *merged,
2133                                    enum intel_ddb_partitioning partitioning,
2134                                    struct ilk_wm_values *results)
2135 {
2136         struct intel_crtc *intel_crtc;
2137         int level, wm_lp;
2138
2139         results->enable_fbc_wm = merged->fbc_wm_enabled;
2140         results->partitioning = partitioning;
2141
2142         /* LP1+ register values */
2143         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2144                 const struct intel_wm_level *r;
2145
2146                 level = ilk_wm_lp_to_level(wm_lp, merged);
2147
2148                 r = &merged->wm[level];
2149
2150                 /*
2151                  * Maintain the watermark values even if the level is
2152                  * disabled. Doing otherwise could cause underruns.
2153                  */
2154                 results->wm_lp[wm_lp - 1] =
2155                         (ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
2156                         (r->pri_val << WM1_LP_SR_SHIFT) |
2157                         r->cur_val;
2158
2159                 if (r->enable)
2160                         results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
2161
2162                 if (INTEL_INFO(dev)->gen >= 8)
2163                         results->wm_lp[wm_lp - 1] |=
2164                                 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
2165                 else
2166                         results->wm_lp[wm_lp - 1] |=
2167                                 r->fbc_val << WM1_LP_FBC_SHIFT;
2168
2169                 /*
2170                  * Always set WM1S_LP_EN when spr_val != 0, even if the
2171                  * level is disabled. Doing otherwise could cause underruns.
2172                  */
2173                 if (INTEL_INFO(dev)->gen <= 6 && r->spr_val) {
2174                         WARN_ON(wm_lp != 1);
2175                         results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
2176                 } else
2177                         results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2178         }
2179
2180         /* LP0 register values */
2181         for_each_intel_crtc(dev, intel_crtc) {
2182                 enum pipe pipe = intel_crtc->pipe;
2183                 const struct intel_wm_level *r =
2184                         &intel_crtc->wm.active.wm[0];
2185
2186                 if (WARN_ON(!r->enable))
2187                         continue;
2188
2189                 results->wm_linetime[pipe] = intel_crtc->wm.active.linetime;
2190
2191                 results->wm_pipe[pipe] =
2192                         (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
2193                         (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
2194                         r->cur_val;
2195         }
2196 }
2197
2198 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2199  * case both are at the same level. Prefer r1 in case they're the same. */
2200 static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
2201                                                   struct intel_pipe_wm *r1,
2202                                                   struct intel_pipe_wm *r2)
2203 {
2204         int level, max_level = ilk_wm_max_level(dev);
2205         int level1 = 0, level2 = 0;
2206
2207         for (level = 1; level <= max_level; level++) {
2208                 if (r1->wm[level].enable)
2209                         level1 = level;
2210                 if (r2->wm[level].enable)
2211                         level2 = level;
2212         }
2213
2214         if (level1 == level2) {
2215                 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
2216                         return r2;
2217                 else
2218                         return r1;
2219         } else if (level1 > level2) {
2220                 return r1;
2221         } else {
2222                 return r2;
2223         }
2224 }
2225
2226 /* dirty bits used to track which watermarks need changes */
2227 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2228 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2229 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2230 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2231 #define WM_DIRTY_FBC (1 << 24)
2232 #define WM_DIRTY_DDB (1 << 25)
2233
2234 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private *dev_priv,
2235                                          const struct ilk_wm_values *old,
2236                                          const struct ilk_wm_values *new)
2237 {
2238         unsigned int dirty = 0;
2239         enum pipe pipe;
2240         int wm_lp;
2241
2242         for_each_pipe(dev_priv, pipe) {
2243                 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
2244                         dirty |= WM_DIRTY_LINETIME(pipe);
2245                         /* Must disable LP1+ watermarks too */
2246                         dirty |= WM_DIRTY_LP_ALL;
2247                 }
2248
2249                 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
2250                         dirty |= WM_DIRTY_PIPE(pipe);
2251                         /* Must disable LP1+ watermarks too */
2252                         dirty |= WM_DIRTY_LP_ALL;
2253                 }
2254         }
2255
2256         if (old->enable_fbc_wm != new->enable_fbc_wm) {
2257                 dirty |= WM_DIRTY_FBC;
2258                 /* Must disable LP1+ watermarks too */
2259                 dirty |= WM_DIRTY_LP_ALL;
2260         }
2261
2262         if (old->partitioning != new->partitioning) {
2263                 dirty |= WM_DIRTY_DDB;
2264                 /* Must disable LP1+ watermarks too */
2265                 dirty |= WM_DIRTY_LP_ALL;
2266         }
2267
2268         /* LP1+ watermarks already deemed dirty, no need to continue */
2269         if (dirty & WM_DIRTY_LP_ALL)
2270                 return dirty;
2271
2272         /* Find the lowest numbered LP1+ watermark in need of an update... */
2273         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2274                 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
2275                     old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
2276                         break;
2277         }
2278
2279         /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2280         for (; wm_lp <= 3; wm_lp++)
2281                 dirty |= WM_DIRTY_LP(wm_lp);
2282
2283         return dirty;
2284 }
2285
2286 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
2287                                unsigned int dirty)
2288 {
2289         struct ilk_wm_values *previous = &dev_priv->wm.hw;
2290         bool changed = false;
2291
2292         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
2293                 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
2294                 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
2295                 changed = true;
2296         }
2297         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
2298                 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
2299                 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
2300                 changed = true;
2301         }
2302         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
2303                 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
2304                 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
2305                 changed = true;
2306         }
2307
2308         /*
2309          * Don't touch WM1S_LP_EN here.
2310          * Doing so could cause underruns.
2311          */
2312
2313         return changed;
2314 }
2315
2316 /*
2317  * The spec says we shouldn't write when we don't need, because every write
2318  * causes WMs to be re-evaluated, expending some power.
2319  */
2320 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
2321                                 struct ilk_wm_values *results)
2322 {
2323         struct drm_device *dev = dev_priv->dev;
2324         struct ilk_wm_values *previous = &dev_priv->wm.hw;
2325         unsigned int dirty;
2326         uint32_t val;
2327
2328         dirty = ilk_compute_wm_dirty(dev_priv, previous, results);
2329         if (!dirty)
2330                 return;
2331
2332         _ilk_disable_lp_wm(dev_priv, dirty);
2333
2334         if (dirty & WM_DIRTY_PIPE(PIPE_A))
2335                 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2336         if (dirty & WM_DIRTY_PIPE(PIPE_B))
2337                 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
2338         if (dirty & WM_DIRTY_PIPE(PIPE_C))
2339                 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
2340
2341         if (dirty & WM_DIRTY_LINETIME(PIPE_A))
2342                 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
2343         if (dirty & WM_DIRTY_LINETIME(PIPE_B))
2344                 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
2345         if (dirty & WM_DIRTY_LINETIME(PIPE_C))
2346                 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
2347
2348         if (dirty & WM_DIRTY_DDB) {
2349                 if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2350                         val = I915_READ(WM_MISC);
2351                         if (results->partitioning == INTEL_DDB_PART_1_2)
2352                                 val &= ~WM_MISC_DATA_PARTITION_5_6;
2353                         else
2354                                 val |= WM_MISC_DATA_PARTITION_5_6;
2355                         I915_WRITE(WM_MISC, val);
2356                 } else {
2357                         val = I915_READ(DISP_ARB_CTL2);
2358                         if (results->partitioning == INTEL_DDB_PART_1_2)
2359                                 val &= ~DISP_DATA_PARTITION_5_6;
2360                         else
2361                                 val |= DISP_DATA_PARTITION_5_6;
2362                         I915_WRITE(DISP_ARB_CTL2, val);
2363                 }
2364         }
2365
2366         if (dirty & WM_DIRTY_FBC) {
2367                 val = I915_READ(DISP_ARB_CTL);
2368                 if (results->enable_fbc_wm)
2369                         val &= ~DISP_FBC_WM_DIS;
2370                 else
2371                         val |= DISP_FBC_WM_DIS;
2372                 I915_WRITE(DISP_ARB_CTL, val);
2373         }
2374
2375         if (dirty & WM_DIRTY_LP(1) &&
2376             previous->wm_lp_spr[0] != results->wm_lp_spr[0])
2377                 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
2378
2379         if (INTEL_INFO(dev)->gen >= 7) {
2380                 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
2381                         I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
2382                 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
2383                         I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
2384         }
2385
2386         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
2387                 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
2388         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
2389                 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
2390         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
2391                 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
2392
2393         dev_priv->wm.hw = *results;
2394 }
2395
2396 static bool ilk_disable_lp_wm(struct drm_device *dev)
2397 {
2398         struct drm_i915_private *dev_priv = dev->dev_private;
2399
2400         return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
2401 }
2402
2403 /*
2404  * On gen9, we need to allocate Display Data Buffer (DDB) portions to the
2405  * different active planes.
2406  */
2407
2408 #define SKL_DDB_SIZE            896     /* in blocks */
2409
2410 static void
2411 skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
2412                                    struct drm_crtc *for_crtc,
2413                                    const struct intel_wm_config *config,
2414                                    const struct skl_pipe_wm_parameters *params,
2415                                    struct skl_ddb_entry *alloc /* out */)
2416 {
2417         struct drm_crtc *crtc;
2418         unsigned int pipe_size, ddb_size;
2419         int nth_active_pipe;
2420
2421         if (!params->active) {
2422                 alloc->start = 0;
2423                 alloc->end = 0;
2424                 return;
2425         }
2426
2427         ddb_size = SKL_DDB_SIZE;
2428
2429         ddb_size -= 4; /* 4 blocks for bypass path allocation */
2430
2431         nth_active_pipe = 0;
2432         for_each_crtc(dev, crtc) {
2433                 if (!intel_crtc_active(crtc))
2434                         continue;
2435
2436                 if (crtc == for_crtc)
2437                         break;
2438
2439                 nth_active_pipe++;
2440         }
2441
2442         pipe_size = ddb_size / config->num_pipes_active;
2443         alloc->start = nth_active_pipe * ddb_size / config->num_pipes_active;
2444         alloc->end = alloc->start + pipe_size;
2445 }
2446
2447 static unsigned int skl_cursor_allocation(const struct intel_wm_config *config)
2448 {
2449         if (config->num_pipes_active == 1)
2450                 return 32;
2451
2452         return 8;
2453 }
2454
2455 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry *entry, u32 reg)
2456 {
2457         entry->start = reg & 0x3ff;
2458         entry->end = (reg >> 16) & 0x3ff;
2459         if (entry->end)
2460                 entry->end += 1;
2461 }
2462
2463 void skl_ddb_get_hw_state(struct drm_i915_private *dev_priv,
2464                           struct skl_ddb_allocation *ddb /* out */)
2465 {
2466         struct drm_device *dev = dev_priv->dev;
2467         enum pipe pipe;
2468         int plane;
2469         u32 val;
2470
2471         for_each_pipe(dev_priv, pipe) {
2472                 for_each_plane(pipe, plane) {
2473                         val = I915_READ(PLANE_BUF_CFG(pipe, plane));
2474                         skl_ddb_entry_init_from_hw(&ddb->plane[pipe][plane],
2475                                                    val);
2476                 }
2477
2478                 val = I915_READ(CUR_BUF_CFG(pipe));
2479                 skl_ddb_entry_init_from_hw(&ddb->cursor[pipe], val);
2480         }
2481 }
2482
2483 static unsigned int
2484 skl_plane_relative_data_rate(const struct intel_plane_wm_parameters *p)
2485 {
2486         return p->horiz_pixels * p->vert_pixels * p->bytes_per_pixel;
2487 }
2488
2489 /*
2490  * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
2491  * a 8192x4096@32bpp framebuffer:
2492  *   3 * 4096 * 8192  * 4 < 2^32
2493  */
2494 static unsigned int
2495 skl_get_total_relative_data_rate(struct intel_crtc *intel_crtc,
2496                                  const struct skl_pipe_wm_parameters *params)
2497 {
2498         unsigned int total_data_rate = 0;
2499         int plane;
2500
2501         for (plane = 0; plane < intel_num_planes(intel_crtc); plane++) {
2502                 const struct intel_plane_wm_parameters *p;
2503
2504                 p = &params->plane[plane];
2505                 if (!p->enabled)
2506                         continue;
2507
2508                 total_data_rate += skl_plane_relative_data_rate(p);
2509         }
2510
2511         return total_data_rate;
2512 }
2513
2514 static void
2515 skl_allocate_pipe_ddb(struct drm_crtc *crtc,
2516                       const struct intel_wm_config *config,
2517                       const struct skl_pipe_wm_parameters *params,
2518                       struct skl_ddb_allocation *ddb /* out */)
2519 {
2520         struct drm_device *dev = crtc->dev;
2521         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2522         enum pipe pipe = intel_crtc->pipe;
2523         struct skl_ddb_entry *alloc = &ddb->pipe[pipe];
2524         uint16_t alloc_size, start, cursor_blocks;
2525         unsigned int total_data_rate;
2526         int plane;
2527
2528         skl_ddb_get_pipe_allocation_limits(dev, crtc, config, params, alloc);
2529         alloc_size = skl_ddb_entry_size(alloc);
2530         if (alloc_size == 0) {
2531                 memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
2532                 memset(&ddb->cursor[pipe], 0, sizeof(ddb->cursor[pipe]));
2533                 return;
2534         }
2535
2536         cursor_blocks = skl_cursor_allocation(config);
2537         ddb->cursor[pipe].start = alloc->end - cursor_blocks;
2538         ddb->cursor[pipe].end = alloc->end;
2539
2540         alloc_size -= cursor_blocks;
2541         alloc->end -= cursor_blocks;
2542
2543         /*
2544          * Each active plane get a portion of the remaining space, in
2545          * proportion to the amount of data they need to fetch from memory.
2546          *
2547          * FIXME: we may not allocate every single block here.
2548          */
2549         total_data_rate = skl_get_total_relative_data_rate(intel_crtc, params);
2550
2551         start = alloc->start;
2552         for (plane = 0; plane < intel_num_planes(intel_crtc); plane++) {
2553                 const struct intel_plane_wm_parameters *p;
2554                 unsigned int data_rate;
2555                 uint16_t plane_blocks;
2556
2557                 p = &params->plane[plane];
2558                 if (!p->enabled)
2559                         continue;
2560
2561                 data_rate = skl_plane_relative_data_rate(p);
2562
2563                 /*
2564                  * promote the expression to 64 bits to avoid overflowing, the
2565                  * result is < available as data_rate / total_data_rate < 1
2566                  */
2567                 plane_blocks = div_u64((uint64_t)alloc_size * data_rate,
2568                                        total_data_rate);
2569
2570                 ddb->plane[pipe][plane].start = start;
2571                 ddb->plane[pipe][plane].end = start + plane_blocks;
2572
2573                 start += plane_blocks;
2574         }
2575
2576 }
2577
2578 static uint32_t skl_pipe_pixel_rate(const struct intel_crtc_state *config)
2579 {
2580         /* TODO: Take into account the scalers once we support them */
2581         return config->base.adjusted_mode.crtc_clock;
2582 }
2583
2584 /*
2585  * The max latency should be 257 (max the punit can code is 255 and we add 2us
2586  * for the read latency) and bytes_per_pixel should always be <= 8, so that
2587  * should allow pixel_rate up to ~2 GHz which seems sufficient since max
2588  * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
2589 */
2590 static uint32_t skl_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
2591                                uint32_t latency)
2592 {
2593         uint32_t wm_intermediate_val, ret;
2594
2595         if (latency == 0)
2596                 return UINT_MAX;
2597
2598         wm_intermediate_val = latency * pixel_rate * bytes_per_pixel;
2599         ret = DIV_ROUND_UP(wm_intermediate_val, 1000);
2600
2601         return ret;
2602 }
2603
2604 static uint32_t skl_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
2605                                uint32_t horiz_pixels, uint8_t bytes_per_pixel,
2606                                uint32_t latency)
2607 {
2608         uint32_t ret, plane_bytes_per_line, wm_intermediate_val;
2609
2610         if (latency == 0)
2611                 return UINT_MAX;
2612
2613         plane_bytes_per_line = horiz_pixels * bytes_per_pixel;
2614         wm_intermediate_val = latency * pixel_rate;
2615         ret = DIV_ROUND_UP(wm_intermediate_val, pipe_htotal * 1000) *
2616                                 plane_bytes_per_line;
2617
2618         return ret;
2619 }
2620
2621 static bool skl_ddb_allocation_changed(const struct skl_ddb_allocation *new_ddb,
2622                                        const struct intel_crtc *intel_crtc)
2623 {
2624         struct drm_device *dev = intel_crtc->base.dev;
2625         struct drm_i915_private *dev_priv = dev->dev_private;
2626         const struct skl_ddb_allocation *cur_ddb = &dev_priv->wm.skl_hw.ddb;
2627         enum pipe pipe = intel_crtc->pipe;
2628
2629         if (memcmp(new_ddb->plane[pipe], cur_ddb->plane[pipe],
2630                    sizeof(new_ddb->plane[pipe])))
2631                 return true;
2632
2633         if (memcmp(&new_ddb->cursor[pipe], &cur_ddb->cursor[pipe],
2634                     sizeof(new_ddb->cursor[pipe])))
2635                 return true;
2636
2637         return false;
2638 }
2639
2640 static void skl_compute_wm_global_parameters(struct drm_device *dev,
2641                                              struct intel_wm_config *config)
2642 {
2643         struct drm_crtc *crtc;
2644         struct drm_plane *plane;
2645
2646         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
2647                 config->num_pipes_active += intel_crtc_active(crtc);
2648
2649         /* FIXME: I don't think we need those two global parameters on SKL */
2650         list_for_each_entry(plane, &dev->mode_config.plane_list, head) {
2651                 struct intel_plane *intel_plane = to_intel_plane(plane);
2652
2653                 config->sprites_enabled |= intel_plane->wm.enabled;
2654                 config->sprites_scaled |= intel_plane->wm.scaled;
2655         }
2656 }
2657
2658 static void skl_compute_wm_pipe_parameters(struct drm_crtc *crtc,
2659                                            struct skl_pipe_wm_parameters *p)
2660 {
2661         struct drm_device *dev = crtc->dev;
2662         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2663         enum pipe pipe = intel_crtc->pipe;
2664         struct drm_plane *plane;
2665         int i = 1; /* Index for sprite planes start */
2666
2667         p->active = intel_crtc_active(crtc);
2668         if (p->active) {
2669                 p->pipe_htotal = intel_crtc->config->base.adjusted_mode.crtc_htotal;
2670                 p->pixel_rate = skl_pipe_pixel_rate(intel_crtc->config);
2671
2672                 /*
2673                  * For now, assume primary and cursor planes are always enabled.
2674                  */
2675                 p->plane[0].enabled = true;
2676                 p->plane[0].bytes_per_pixel =
2677                         crtc->primary->fb->bits_per_pixel / 8;
2678                 p->plane[0].horiz_pixels = intel_crtc->config->pipe_src_w;
2679                 p->plane[0].vert_pixels = intel_crtc->config->pipe_src_h;
2680
2681                 p->cursor.enabled = true;
2682                 p->cursor.bytes_per_pixel = 4;
2683                 p->cursor.horiz_pixels = intel_crtc->cursor_width ?
2684                                          intel_crtc->cursor_width : 64;
2685         }
2686
2687         list_for_each_entry(plane, &dev->mode_config.plane_list, head) {
2688                 struct intel_plane *intel_plane = to_intel_plane(plane);
2689
2690                 if (intel_plane->pipe == pipe &&
2691                         plane->type == DRM_PLANE_TYPE_OVERLAY)
2692                         p->plane[i++] = intel_plane->wm;
2693         }
2694 }
2695
2696 static bool skl_compute_plane_wm(struct skl_pipe_wm_parameters *p,
2697                                  struct intel_plane_wm_parameters *p_params,
2698                                  uint16_t ddb_allocation,
2699                                  uint32_t mem_value,
2700                                  uint16_t *out_blocks, /* out */
2701                                  uint8_t *out_lines /* out */)
2702 {
2703         uint32_t method1, method2, plane_bytes_per_line, res_blocks, res_lines;
2704         uint32_t result_bytes;
2705
2706         if (mem_value == 0 || !p->active || !p_params->enabled)
2707                 return false;
2708
2709         method1 = skl_wm_method1(p->pixel_rate,
2710                                  p_params->bytes_per_pixel,
2711                                  mem_value);
2712         method2 = skl_wm_method2(p->pixel_rate,
2713                                  p->pipe_htotal,
2714                                  p_params->horiz_pixels,
2715                                  p_params->bytes_per_pixel,
2716                                  mem_value);
2717
2718         plane_bytes_per_line = p_params->horiz_pixels *
2719                                         p_params->bytes_per_pixel;
2720
2721         /* For now xtile and linear */
2722         if (((ddb_allocation * 512) / plane_bytes_per_line) >= 1)
2723                 result_bytes = min(method1, method2);
2724         else
2725                 result_bytes = method1;
2726
2727         res_blocks = DIV_ROUND_UP(result_bytes, 512) + 1;
2728         res_lines = DIV_ROUND_UP(result_bytes, plane_bytes_per_line);
2729
2730         if (res_blocks > ddb_allocation || res_lines > 31)
2731                 return false;
2732
2733         *out_blocks = res_blocks;
2734         *out_lines = res_lines;
2735
2736         return true;
2737 }
2738
2739 static void skl_compute_wm_level(const struct drm_i915_private *dev_priv,
2740                                  struct skl_ddb_allocation *ddb,
2741                                  struct skl_pipe_wm_parameters *p,
2742                                  enum pipe pipe,
2743                                  int level,
2744                                  int num_planes,
2745                                  struct skl_wm_level *result)
2746 {
2747         uint16_t latency = dev_priv->wm.skl_latency[level];
2748         uint16_t ddb_blocks;
2749         int i;
2750
2751         for (i = 0; i < num_planes; i++) {
2752                 ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][i]);
2753
2754                 result->plane_en[i] = skl_compute_plane_wm(p, &p->plane[i],
2755                                                 ddb_blocks,
2756                                                 latency,
2757                                                 &result->plane_res_b[i],
2758                                                 &result->plane_res_l[i]);
2759         }
2760
2761         ddb_blocks = skl_ddb_entry_size(&ddb->cursor[pipe]);
2762         result->cursor_en = skl_compute_plane_wm(p, &p->cursor, ddb_blocks,
2763                                                  latency, &result->cursor_res_b,
2764                                                  &result->cursor_res_l);
2765 }
2766
2767 static uint32_t
2768 skl_compute_linetime_wm(struct drm_crtc *crtc, struct skl_pipe_wm_parameters *p)
2769 {
2770         if (!intel_crtc_active(crtc))
2771                 return 0;
2772
2773         return DIV_ROUND_UP(8 * p->pipe_htotal * 1000, p->pixel_rate);
2774
2775 }
2776
2777 static void skl_compute_transition_wm(struct drm_crtc *crtc,
2778                                       struct skl_pipe_wm_parameters *params,
2779                                       struct skl_wm_level *trans_wm /* out */)
2780 {
2781         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2782         int i;
2783
2784         if (!params->active)
2785                 return;
2786
2787         /* Until we know more, just disable transition WMs */
2788         for (i = 0; i < intel_num_planes(intel_crtc); i++)
2789                 trans_wm->plane_en[i] = false;
2790         trans_wm->cursor_en = false;
2791 }
2792
2793 static void skl_compute_pipe_wm(struct drm_crtc *crtc,
2794                                 struct skl_ddb_allocation *ddb,
2795                                 struct skl_pipe_wm_parameters *params,
2796                                 struct skl_pipe_wm *pipe_wm)
2797 {
2798         struct drm_device *dev = crtc->dev;
2799         const struct drm_i915_private *dev_priv = dev->dev_private;
2800         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2801         int level, max_level = ilk_wm_max_level(dev);
2802
2803         for (level = 0; level <= max_level; level++) {
2804                 skl_compute_wm_level(dev_priv, ddb, params, intel_crtc->pipe,
2805                                      level, intel_num_planes(intel_crtc),
2806                                      &pipe_wm->wm[level]);
2807         }
2808         pipe_wm->linetime = skl_compute_linetime_wm(crtc, params);
2809
2810         skl_compute_transition_wm(crtc, params, &pipe_wm->trans_wm);
2811 }
2812
2813 static void skl_compute_wm_results(struct drm_device *dev,
2814                                    struct skl_pipe_wm_parameters *p,
2815                                    struct skl_pipe_wm *p_wm,
2816                                    struct skl_wm_values *r,
2817                                    struct intel_crtc *intel_crtc)
2818 {
2819         int level, max_level = ilk_wm_max_level(dev);
2820         enum pipe pipe = intel_crtc->pipe;
2821         uint32_t temp;
2822         int i;
2823
2824         for (level = 0; level <= max_level; level++) {
2825                 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
2826                         temp = 0;
2827
2828                         temp |= p_wm->wm[level].plane_res_l[i] <<
2829                                         PLANE_WM_LINES_SHIFT;
2830                         temp |= p_wm->wm[level].plane_res_b[i];
2831                         if (p_wm->wm[level].plane_en[i])
2832                                 temp |= PLANE_WM_EN;
2833
2834                         r->plane[pipe][i][level] = temp;
2835                 }
2836
2837                 temp = 0;
2838
2839                 temp |= p_wm->wm[level].cursor_res_l << PLANE_WM_LINES_SHIFT;
2840                 temp |= p_wm->wm[level].cursor_res_b;
2841
2842                 if (p_wm->wm[level].cursor_en)
2843                         temp |= PLANE_WM_EN;
2844
2845                 r->cursor[pipe][level] = temp;
2846
2847         }
2848
2849         /* transition WMs */
2850         for (i = 0; i < intel_num_planes(intel_crtc); i++) {
2851                 temp = 0;
2852                 temp |= p_wm->trans_wm.plane_res_l[i] << PLANE_WM_LINES_SHIFT;
2853                 temp |= p_wm->trans_wm.plane_res_b[i];
2854                 if (p_wm->trans_wm.plane_en[i])
2855                         temp |= PLANE_WM_EN;
2856
2857                 r->plane_trans[pipe][i] = temp;
2858         }
2859
2860         temp = 0;
2861         temp |= p_wm->trans_wm.cursor_res_l << PLANE_WM_LINES_SHIFT;
2862         temp |= p_wm->trans_wm.cursor_res_b;
2863         if (p_wm->trans_wm.cursor_en)
2864                 temp |= PLANE_WM_EN;
2865
2866         r->cursor_trans[pipe] = temp;
2867
2868         r->wm_linetime[pipe] = p_wm->linetime;
2869 }
2870
2871 static void skl_ddb_entry_write(struct drm_i915_private *dev_priv, uint32_t reg,
2872                                 const struct skl_ddb_entry *entry)
2873 {
2874         if (entry->end)
2875                 I915_WRITE(reg, (entry->end - 1) << 16 | entry->start);
2876         else
2877                 I915_WRITE(reg, 0);
2878 }
2879
2880 static void skl_write_wm_values(struct drm_i915_private *dev_priv,
2881                                 const struct skl_wm_values *new)
2882 {
2883         struct drm_device *dev = dev_priv->dev;
2884         struct intel_crtc *crtc;
2885
2886         list_for_each_entry(crtc, &dev->mode_config.crtc_list, base.head) {
2887                 int i, level, max_level = ilk_wm_max_level(dev);
2888                 enum pipe pipe = crtc->pipe;
2889
2890                 if (!new->dirty[pipe])
2891                         continue;
2892
2893                 I915_WRITE(PIPE_WM_LINETIME(pipe), new->wm_linetime[pipe]);
2894
2895                 for (level = 0; level <= max_level; level++) {
2896                         for (i = 0; i < intel_num_planes(crtc); i++)
2897                                 I915_WRITE(PLANE_WM(pipe, i, level),
2898                                            new->plane[pipe][i][level]);
2899                         I915_WRITE(CUR_WM(pipe, level),
2900                                    new->cursor[pipe][level]);
2901                 }
2902                 for (i = 0; i < intel_num_planes(crtc); i++)
2903                         I915_WRITE(PLANE_WM_TRANS(pipe, i),
2904                                    new->plane_trans[pipe][i]);
2905                 I915_WRITE(CUR_WM_TRANS(pipe), new->cursor_trans[pipe]);
2906
2907                 for (i = 0; i < intel_num_planes(crtc); i++)
2908                         skl_ddb_entry_write(dev_priv,
2909                                             PLANE_BUF_CFG(pipe, i),
2910                                             &new->ddb.plane[pipe][i]);
2911
2912                 skl_ddb_entry_write(dev_priv, CUR_BUF_CFG(pipe),
2913                                     &new->ddb.cursor[pipe]);
2914         }
2915 }
2916
2917 /*
2918  * When setting up a new DDB allocation arrangement, we need to correctly
2919  * sequence the times at which the new allocations for the pipes are taken into
2920  * account or we'll have pipes fetching from space previously allocated to
2921  * another pipe.
2922  *
2923  * Roughly the sequence looks like:
2924  *  1. re-allocate the pipe(s) with the allocation being reduced and not
2925  *     overlapping with a previous light-up pipe (another way to put it is:
2926  *     pipes with their new allocation strickly included into their old ones).
2927  *  2. re-allocate the other pipes that get their allocation reduced
2928  *  3. allocate the pipes having their allocation increased
2929  *
2930  * Steps 1. and 2. are here to take care of the following case:
2931  * - Initially DDB looks like this:
2932  *     |   B    |   C    |
2933  * - enable pipe A.
2934  * - pipe B has a reduced DDB allocation that overlaps with the old pipe C
2935  *   allocation
2936  *     |  A  |  B  |  C  |
2937  *
2938  * We need to sequence the re-allocation: C, B, A (and not B, C, A).
2939  */
2940
2941 static void
2942 skl_wm_flush_pipe(struct drm_i915_private *dev_priv, enum pipe pipe, int pass)
2943 {
2944         struct drm_device *dev = dev_priv->dev;
2945         int plane;
2946
2947         DRM_DEBUG_KMS("flush pipe %c (pass %d)\n", pipe_name(pipe), pass);
2948
2949         for_each_plane(pipe, plane) {
2950                 I915_WRITE(PLANE_SURF(pipe, plane),
2951                            I915_READ(PLANE_SURF(pipe, plane)));
2952         }
2953         I915_WRITE(CURBASE(pipe), I915_READ(CURBASE(pipe)));
2954 }
2955
2956 static bool
2957 skl_ddb_allocation_included(const struct skl_ddb_allocation *old,
2958                             const struct skl_ddb_allocation *new,
2959                             enum pipe pipe)
2960 {
2961         uint16_t old_size, new_size;
2962
2963         old_size = skl_ddb_entry_size(&old->pipe[pipe]);
2964         new_size = skl_ddb_entry_size(&new->pipe[pipe]);
2965
2966         return old_size != new_size &&
2967                new->pipe[pipe].start >= old->pipe[pipe].start &&
2968                new->pipe[pipe].end <= old->pipe[pipe].end;
2969 }
2970
2971 static void skl_flush_wm_values(struct drm_i915_private *dev_priv,
2972                                 struct skl_wm_values *new_values)
2973 {
2974         struct drm_device *dev = dev_priv->dev;
2975         struct skl_ddb_allocation *cur_ddb, *new_ddb;
2976         bool reallocated[I915_MAX_PIPES] = {false, false, false};
2977         struct intel_crtc *crtc;
2978         enum pipe pipe;
2979
2980         new_ddb = &new_values->ddb;
2981         cur_ddb = &dev_priv->wm.skl_hw.ddb;
2982
2983         /*
2984          * First pass: flush the pipes with the new allocation contained into
2985          * the old space.
2986          *
2987          * We'll wait for the vblank on those pipes to ensure we can safely
2988          * re-allocate the freed space without this pipe fetching from it.
2989          */
2990         for_each_intel_crtc(dev, crtc) {
2991                 if (!crtc->active)
2992                         continue;
2993
2994                 pipe = crtc->pipe;
2995
2996                 if (!skl_ddb_allocation_included(cur_ddb, new_ddb, pipe))
2997                         continue;
2998
2999                 skl_wm_flush_pipe(dev_priv, pipe, 1);
3000                 intel_wait_for_vblank(dev, pipe);
3001
3002                 reallocated[pipe] = true;
3003         }
3004
3005
3006         /*
3007          * Second pass: flush the pipes that are having their allocation
3008          * reduced, but overlapping with a previous allocation.
3009          *
3010          * Here as well we need to wait for the vblank to make sure the freed
3011          * space is not used anymore.
3012          */
3013         for_each_intel_crtc(dev, crtc) {
3014                 if (!crtc->active)
3015                         continue;
3016
3017                 pipe = crtc->pipe;
3018
3019                 if (reallocated[pipe])
3020                         continue;
3021
3022                 if (skl_ddb_entry_size(&new_ddb->pipe[pipe]) <
3023                     skl_ddb_entry_size(&cur_ddb->pipe[pipe])) {
3024                         skl_wm_flush_pipe(dev_priv, pipe, 2);
3025                         intel_wait_for_vblank(dev, pipe);
3026                         reallocated[pipe] = true;
3027                 }
3028         }
3029
3030         /*
3031          * Third pass: flush the pipes that got more space allocated.
3032          *
3033          * We don't need to actively wait for the update here, next vblank
3034          * will just get more DDB space with the correct WM values.
3035          */
3036         for_each_intel_crtc(dev, crtc) {
3037                 if (!crtc->active)
3038                         continue;
3039
3040                 pipe = crtc->pipe;
3041
3042                 /*
3043                  * At this point, only the pipes more space than before are
3044                  * left to re-allocate.
3045                  */
3046                 if (reallocated[pipe])
3047                         continue;
3048
3049                 skl_wm_flush_pipe(dev_priv, pipe, 3);
3050         }
3051 }
3052
3053 static bool skl_update_pipe_wm(struct drm_crtc *crtc,
3054                                struct skl_pipe_wm_parameters *params,
3055                                struct intel_wm_config *config,
3056                                struct skl_ddb_allocation *ddb, /* out */
3057                                struct skl_pipe_wm *pipe_wm /* out */)
3058 {
3059         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3060
3061         skl_compute_wm_pipe_parameters(crtc, params);
3062         skl_allocate_pipe_ddb(crtc, config, params, ddb);
3063         skl_compute_pipe_wm(crtc, ddb, params, pipe_wm);
3064
3065         if (!memcmp(&intel_crtc->wm.skl_active, pipe_wm, sizeof(*pipe_wm)))
3066                 return false;
3067
3068         intel_crtc->wm.skl_active = *pipe_wm;
3069         return true;
3070 }
3071
3072 static void skl_update_other_pipe_wm(struct drm_device *dev,
3073                                      struct drm_crtc *crtc,
3074                                      struct intel_wm_config *config,
3075                                      struct skl_wm_values *r)
3076 {
3077         struct intel_crtc *intel_crtc;
3078         struct intel_crtc *this_crtc = to_intel_crtc(crtc);
3079
3080         /*
3081          * If the WM update hasn't changed the allocation for this_crtc (the
3082          * crtc we are currently computing the new WM values for), other
3083          * enabled crtcs will keep the same allocation and we don't need to
3084          * recompute anything for them.
3085          */
3086         if (!skl_ddb_allocation_changed(&r->ddb, this_crtc))
3087                 return;
3088
3089         /*
3090          * Otherwise, because of this_crtc being freshly enabled/disabled, the
3091          * other active pipes need new DDB allocation and WM values.
3092          */
3093         list_for_each_entry(intel_crtc, &dev->mode_config.crtc_list,
3094                                 base.head) {
3095                 struct skl_pipe_wm_parameters params = {};
3096                 struct skl_pipe_wm pipe_wm = {};
3097                 bool wm_changed;
3098
3099                 if (this_crtc->pipe == intel_crtc->pipe)
3100                         continue;
3101
3102                 if (!intel_crtc->active)
3103                         continue;
3104
3105                 wm_changed = skl_update_pipe_wm(&intel_crtc->base,
3106                                                 &params, config,
3107                                                 &r->ddb, &pipe_wm);
3108
3109                 /*
3110                  * If we end up re-computing the other pipe WM values, it's
3111                  * because it was really needed, so we expect the WM values to
3112                  * be different.
3113                  */
3114                 WARN_ON(!wm_changed);
3115
3116                 skl_compute_wm_results(dev, &params, &pipe_wm, r, intel_crtc);
3117                 r->dirty[intel_crtc->pipe] = true;
3118         }
3119 }
3120
3121 static void skl_update_wm(struct drm_crtc *crtc)
3122 {
3123         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3124         struct drm_device *dev = crtc->dev;
3125         struct drm_i915_private *dev_priv = dev->dev_private;
3126         struct skl_pipe_wm_parameters params = {};
3127         struct skl_wm_values *results = &dev_priv->wm.skl_results;
3128         struct skl_pipe_wm pipe_wm = {};
3129         struct intel_wm_config config = {};
3130
3131         memset(results, 0, sizeof(*results));
3132
3133         skl_compute_wm_global_parameters(dev, &config);
3134
3135         if (!skl_update_pipe_wm(crtc, &params, &config,
3136                                 &results->ddb, &pipe_wm))
3137                 return;
3138
3139         skl_compute_wm_results(dev, &params, &pipe_wm, results, intel_crtc);
3140         results->dirty[intel_crtc->pipe] = true;
3141
3142         skl_update_other_pipe_wm(dev, crtc, &config, results);
3143         skl_write_wm_values(dev_priv, results);
3144         skl_flush_wm_values(dev_priv, results);
3145
3146         /* store the new configuration */
3147         dev_priv->wm.skl_hw = *results;
3148 }
3149
3150 static void
3151 skl_update_sprite_wm(struct drm_plane *plane, struct drm_crtc *crtc,
3152                      uint32_t sprite_width, uint32_t sprite_height,
3153                      int pixel_size, bool enabled, bool scaled)
3154 {
3155         struct intel_plane *intel_plane = to_intel_plane(plane);
3156
3157         intel_plane->wm.enabled = enabled;
3158         intel_plane->wm.scaled = scaled;
3159         intel_plane->wm.horiz_pixels = sprite_width;
3160         intel_plane->wm.vert_pixels = sprite_height;
3161         intel_plane->wm.bytes_per_pixel = pixel_size;
3162
3163         skl_update_wm(crtc);
3164 }
3165
3166 static void ilk_update_wm(struct drm_crtc *crtc)
3167 {
3168         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3169         struct drm_device *dev = crtc->dev;
3170         struct drm_i915_private *dev_priv = dev->dev_private;
3171         struct ilk_wm_maximums max;
3172         struct ilk_pipe_wm_parameters params = {};
3173         struct ilk_wm_values results = {};
3174         enum intel_ddb_partitioning partitioning;
3175         struct intel_pipe_wm pipe_wm = {};
3176         struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
3177         struct intel_wm_config config = {};
3178
3179         ilk_compute_wm_parameters(crtc, &params);
3180
3181         intel_compute_pipe_wm(crtc, &params, &pipe_wm);
3182
3183         if (!memcmp(&intel_crtc->wm.active, &pipe_wm, sizeof(pipe_wm)))
3184                 return;
3185
3186         intel_crtc->wm.active = pipe_wm;
3187
3188         ilk_compute_wm_config(dev, &config);
3189
3190         ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
3191         ilk_wm_merge(dev, &config, &max, &lp_wm_1_2);
3192
3193         /* 5/6 split only in single pipe config on IVB+ */
3194         if (INTEL_INFO(dev)->gen >= 7 &&
3195             config.num_pipes_active == 1 && config.sprites_enabled) {
3196                 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
3197                 ilk_wm_merge(dev, &config, &max, &lp_wm_5_6);
3198
3199                 best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
3200         } else {
3201                 best_lp_wm = &lp_wm_1_2;
3202         }
3203
3204         partitioning = (best_lp_wm == &lp_wm_1_2) ?
3205                        INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
3206
3207         ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
3208
3209         ilk_write_wm_values(dev_priv, &results);
3210 }
3211
3212 static void
3213 ilk_update_sprite_wm(struct drm_plane *plane,
3214                      struct drm_crtc *crtc,
3215                      uint32_t sprite_width, uint32_t sprite_height,
3216                      int pixel_size, bool enabled, bool scaled)
3217 {
3218         struct drm_device *dev = plane->dev;
3219         struct intel_plane *intel_plane = to_intel_plane(plane);
3220
3221         intel_plane->wm.enabled = enabled;
3222         intel_plane->wm.scaled = scaled;
3223         intel_plane->wm.horiz_pixels = sprite_width;
3224         intel_plane->wm.vert_pixels = sprite_width;
3225         intel_plane->wm.bytes_per_pixel = pixel_size;
3226
3227         /*
3228          * IVB workaround: must disable low power watermarks for at least
3229          * one frame before enabling scaling.  LP watermarks can be re-enabled
3230          * when scaling is disabled.
3231          *
3232          * WaCxSRDisabledForSpriteScaling:ivb
3233          */
3234         if (IS_IVYBRIDGE(dev) && scaled && ilk_disable_lp_wm(dev))
3235                 intel_wait_for_vblank(dev, intel_plane->pipe);
3236
3237         ilk_update_wm(crtc);
3238 }
3239
3240 static void skl_pipe_wm_active_state(uint32_t val,
3241                                      struct skl_pipe_wm *active,
3242                                      bool is_transwm,
3243                                      bool is_cursor,
3244                                      int i,
3245                                      int level)
3246 {
3247         bool is_enabled = (val & PLANE_WM_EN) != 0;
3248
3249         if (!is_transwm) {
3250                 if (!is_cursor) {
3251                         active->wm[level].plane_en[i] = is_enabled;
3252                         active->wm[level].plane_res_b[i] =
3253                                         val & PLANE_WM_BLOCKS_MASK;
3254                         active->wm[level].plane_res_l[i] =
3255                                         (val >> PLANE_WM_LINES_SHIFT) &
3256                                                 PLANE_WM_LINES_MASK;
3257                 } else {
3258                         active->wm[level].cursor_en = is_enabled;
3259                         active->wm[level].cursor_res_b =
3260                                         val & PLANE_WM_BLOCKS_MASK;
3261                         active->wm[level].cursor_res_l =
3262                                         (val >> PLANE_WM_LINES_SHIFT) &
3263                                                 PLANE_WM_LINES_MASK;
3264                 }
3265         } else {
3266                 if (!is_cursor) {
3267                         active->trans_wm.plane_en[i] = is_enabled;
3268                         active->trans_wm.plane_res_b[i] =
3269                                         val & PLANE_WM_BLOCKS_MASK;
3270                         active->trans_wm.plane_res_l[i] =
3271                                         (val >> PLANE_WM_LINES_SHIFT) &
3272                                                 PLANE_WM_LINES_MASK;
3273                 } else {
3274                         active->trans_wm.cursor_en = is_enabled;
3275                         active->trans_wm.cursor_res_b =
3276                                         val & PLANE_WM_BLOCKS_MASK;
3277                         active->trans_wm.cursor_res_l =
3278                                         (val >> PLANE_WM_LINES_SHIFT) &
3279                                                 PLANE_WM_LINES_MASK;
3280                 }
3281         }
3282 }
3283
3284 static void skl_pipe_wm_get_hw_state(struct drm_crtc *crtc)
3285 {
3286         struct drm_device *dev = crtc->dev;
3287         struct drm_i915_private *dev_priv = dev->dev_private;
3288         struct skl_wm_values *hw = &dev_priv->wm.skl_hw;
3289         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3290         struct skl_pipe_wm *active = &intel_crtc->wm.skl_active;
3291         enum pipe pipe = intel_crtc->pipe;
3292         int level, i, max_level;
3293         uint32_t temp;
3294
3295         max_level = ilk_wm_max_level(dev);
3296
3297         hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
3298
3299         for (level = 0; level <= max_level; level++) {
3300                 for (i = 0; i < intel_num_planes(intel_crtc); i++)
3301                         hw->plane[pipe][i][level] =
3302                                         I915_READ(PLANE_WM(pipe, i, level));
3303                 hw->cursor[pipe][level] = I915_READ(CUR_WM(pipe, level));
3304         }
3305
3306         for (i = 0; i < intel_num_planes(intel_crtc); i++)
3307                 hw->plane_trans[pipe][i] = I915_READ(PLANE_WM_TRANS(pipe, i));
3308         hw->cursor_trans[pipe] = I915_READ(CUR_WM_TRANS(pipe));
3309
3310         if (!intel_crtc_active(crtc))
3311                 return;
3312
3313         hw->dirty[pipe] = true;
3314
3315         active->linetime = hw->wm_linetime[pipe];
3316
3317         for (level = 0; level <= max_level; level++) {
3318                 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3319                         temp = hw->plane[pipe][i][level];
3320                         skl_pipe_wm_active_state(temp, active, false,
3321                                                 false, i, level);
3322                 }
3323                 temp = hw->cursor[pipe][level];
3324                 skl_pipe_wm_active_state(temp, active, false, true, i, level);
3325         }
3326
3327         for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3328                 temp = hw->plane_trans[pipe][i];
3329                 skl_pipe_wm_active_state(temp, active, true, false, i, 0);
3330         }
3331
3332         temp = hw->cursor_trans[pipe];
3333         skl_pipe_wm_active_state(temp, active, true, true, i, 0);
3334 }
3335
3336 void skl_wm_get_hw_state(struct drm_device *dev)
3337 {
3338         struct drm_i915_private *dev_priv = dev->dev_private;
3339         struct skl_ddb_allocation *ddb = &dev_priv->wm.skl_hw.ddb;
3340         struct drm_crtc *crtc;
3341
3342         skl_ddb_get_hw_state(dev_priv, ddb);
3343         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
3344                 skl_pipe_wm_get_hw_state(crtc);
3345 }
3346
3347 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
3348 {
3349         struct drm_device *dev = crtc->dev;
3350         struct drm_i915_private *dev_priv = dev->dev_private;
3351         struct ilk_wm_values *hw = &dev_priv->wm.hw;
3352         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3353         struct intel_pipe_wm *active = &intel_crtc->wm.active;
3354         enum pipe pipe = intel_crtc->pipe;
3355         static const unsigned int wm0_pipe_reg[] = {
3356                 [PIPE_A] = WM0_PIPEA_ILK,
3357                 [PIPE_B] = WM0_PIPEB_ILK,
3358                 [PIPE_C] = WM0_PIPEC_IVB,
3359         };
3360
3361         hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
3362         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
3363                 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
3364
3365         active->pipe_enabled = intel_crtc_active(crtc);
3366
3367         if (active->pipe_enabled) {
3368                 u32 tmp = hw->wm_pipe[pipe];
3369
3370                 /*
3371                  * For active pipes LP0 watermark is marked as
3372                  * enabled, and LP1+ watermaks as disabled since
3373                  * we can't really reverse compute them in case
3374                  * multiple pipes are active.
3375                  */
3376                 active->wm[0].enable = true;
3377                 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
3378                 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
3379                 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
3380                 active->linetime = hw->wm_linetime[pipe];
3381         } else {
3382                 int level, max_level = ilk_wm_max_level(dev);
3383
3384                 /*
3385                  * For inactive pipes, all watermark levels
3386                  * should be marked as enabled but zeroed,
3387                  * which is what we'd compute them to.
3388                  */
3389                 for (level = 0; level <= max_level; level++)
3390                         active->wm[level].enable = true;
3391         }
3392 }
3393
3394 void ilk_wm_get_hw_state(struct drm_device *dev)
3395 {
3396         struct drm_i915_private *dev_priv = dev->dev_private;
3397         struct ilk_wm_values *hw = &dev_priv->wm.hw;
3398         struct drm_crtc *crtc;
3399
3400         for_each_crtc(dev, crtc)
3401                 ilk_pipe_wm_get_hw_state(crtc);
3402
3403         hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
3404         hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
3405         hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
3406
3407         hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
3408         if (INTEL_INFO(dev)->gen >= 7) {
3409                 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
3410                 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
3411         }
3412
3413         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
3414                 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
3415                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
3416         else if (IS_IVYBRIDGE(dev))
3417                 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
3418                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
3419
3420         hw->enable_fbc_wm =
3421                 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
3422 }
3423
3424 /**
3425  * intel_update_watermarks - update FIFO watermark values based on current modes
3426  *
3427  * Calculate watermark values for the various WM regs based on current mode
3428  * and plane configuration.
3429  *
3430  * There are several cases to deal with here:
3431  *   - normal (i.e. non-self-refresh)
3432  *   - self-refresh (SR) mode
3433  *   - lines are large relative to FIFO size (buffer can hold up to 2)
3434  *   - lines are small relative to FIFO size (buffer can hold more than 2
3435  *     lines), so need to account for TLB latency
3436  *
3437  *   The normal calculation is:
3438  *     watermark = dotclock * bytes per pixel * latency
3439  *   where latency is platform & configuration dependent (we assume pessimal
3440  *   values here).
3441  *
3442  *   The SR calculation is:
3443  *     watermark = (trunc(latency/line time)+1) * surface width *
3444  *       bytes per pixel
3445  *   where
3446  *     line time = htotal / dotclock
3447  *     surface width = hdisplay for normal plane and 64 for cursor
3448  *   and latency is assumed to be high, as above.
3449  *
3450  * The final value programmed to the register should always be rounded up,
3451  * and include an extra 2 entries to account for clock crossings.
3452  *
3453  * We don't use the sprite, so we can ignore that.  And on Crestline we have
3454  * to set the non-SR watermarks to 8.
3455  */
3456 void intel_update_watermarks(struct drm_crtc *crtc)
3457 {
3458         struct drm_i915_private *dev_priv = crtc->dev->dev_private;
3459
3460         if (dev_priv->display.update_wm)
3461                 dev_priv->display.update_wm(crtc);
3462 }
3463
3464 void intel_update_sprite_watermarks(struct drm_plane *plane,
3465                                     struct drm_crtc *crtc,
3466                                     uint32_t sprite_width,
3467                                     uint32_t sprite_height,
3468                                     int pixel_size,
3469                                     bool enabled, bool scaled)
3470 {
3471         struct drm_i915_private *dev_priv = plane->dev->dev_private;
3472
3473         if (dev_priv->display.update_sprite_wm)
3474                 dev_priv->display.update_sprite_wm(plane, crtc,
3475                                                    sprite_width, sprite_height,
3476                                                    pixel_size, enabled, scaled);
3477 }
3478
3479 static struct drm_i915_gem_object *
3480 intel_alloc_context_page(struct drm_device *dev)
3481 {
3482         struct drm_i915_gem_object *ctx;
3483         int ret;
3484
3485         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
3486
3487         ctx = i915_gem_alloc_object(dev, 4096);
3488         if (!ctx) {
3489                 DRM_DEBUG("failed to alloc power context, RC6 disabled\n");
3490                 return NULL;
3491         }
3492
3493         ret = i915_gem_obj_ggtt_pin(ctx, 4096, 0);
3494         if (ret) {
3495                 DRM_ERROR("failed to pin power context: %d\n", ret);
3496                 goto err_unref;
3497         }
3498
3499         ret = i915_gem_object_set_to_gtt_domain(ctx, 1);
3500         if (ret) {
3501                 DRM_ERROR("failed to set-domain on power context: %d\n", ret);
3502                 goto err_unpin;
3503         }
3504
3505         return ctx;
3506
3507 err_unpin:
3508         i915_gem_object_ggtt_unpin(ctx);
3509 err_unref:
3510         drm_gem_object_unreference(&ctx->base);
3511         return NULL;
3512 }
3513
3514 /**
3515  * Lock protecting IPS related data structures
3516  */
3517 DEFINE_SPINLOCK(mchdev_lock);
3518
3519 /* Global for IPS driver to get at the current i915 device. Protected by
3520  * mchdev_lock. */
3521 static struct drm_i915_private *i915_mch_dev;
3522
3523 bool ironlake_set_drps(struct drm_device *dev, u8 val)
3524 {
3525         struct drm_i915_private *dev_priv = dev->dev_private;
3526         u16 rgvswctl;
3527
3528         assert_spin_locked(&mchdev_lock);
3529
3530         rgvswctl = I915_READ16(MEMSWCTL);
3531         if (rgvswctl & MEMCTL_CMD_STS) {
3532                 DRM_DEBUG("gpu busy, RCS change rejected\n");
3533                 return false; /* still busy with another command */
3534         }
3535
3536         rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
3537                 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
3538         I915_WRITE16(MEMSWCTL, rgvswctl);
3539         POSTING_READ16(MEMSWCTL);
3540
3541         rgvswctl |= MEMCTL_CMD_STS;
3542         I915_WRITE16(MEMSWCTL, rgvswctl);
3543
3544         return true;
3545 }
3546
3547 static void ironlake_enable_drps(struct drm_device *dev)
3548 {
3549         struct drm_i915_private *dev_priv = dev->dev_private;
3550         u32 rgvmodectl = I915_READ(MEMMODECTL);
3551         u8 fmax, fmin, fstart, vstart;
3552
3553         spin_lock_irq(&mchdev_lock);
3554
3555         /* Enable temp reporting */
3556         I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
3557         I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
3558
3559         /* 100ms RC evaluation intervals */
3560         I915_WRITE(RCUPEI, 100000);
3561         I915_WRITE(RCDNEI, 100000);
3562
3563         /* Set max/min thresholds to 90ms and 80ms respectively */
3564         I915_WRITE(RCBMAXAVG, 90000);
3565         I915_WRITE(RCBMINAVG, 80000);
3566
3567         I915_WRITE(MEMIHYST, 1);
3568
3569         /* Set up min, max, and cur for interrupt handling */
3570         fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
3571         fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
3572         fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
3573                 MEMMODE_FSTART_SHIFT;
3574
3575         vstart = (I915_READ(PXVFREQ_BASE + (fstart * 4)) & PXVFREQ_PX_MASK) >>
3576                 PXVFREQ_PX_SHIFT;
3577
3578         dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
3579         dev_priv->ips.fstart = fstart;
3580
3581         dev_priv->ips.max_delay = fstart;
3582         dev_priv->ips.min_delay = fmin;
3583         dev_priv->ips.cur_delay = fstart;
3584
3585         DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
3586                          fmax, fmin, fstart);
3587
3588         I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
3589
3590         /*
3591          * Interrupts will be enabled in ironlake_irq_postinstall
3592          */
3593
3594         I915_WRITE(VIDSTART, vstart);
3595         POSTING_READ(VIDSTART);
3596
3597         rgvmodectl |= MEMMODE_SWMODE_EN;
3598         I915_WRITE(MEMMODECTL, rgvmodectl);
3599
3600         if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
3601                 DRM_ERROR("stuck trying to change perf mode\n");
3602         mdelay(1);
3603
3604         ironlake_set_drps(dev, fstart);
3605
3606         dev_priv->ips.last_count1 = I915_READ(0x112e4) + I915_READ(0x112e8) +
3607                 I915_READ(0x112e0);
3608         dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
3609         dev_priv->ips.last_count2 = I915_READ(0x112f4);
3610         dev_priv->ips.last_time2 = ktime_get_raw_ns();
3611
3612         spin_unlock_irq(&mchdev_lock);
3613 }
3614
3615 static void ironlake_disable_drps(struct drm_device *dev)
3616 {
3617         struct drm_i915_private *dev_priv = dev->dev_private;
3618         u16 rgvswctl;
3619
3620         spin_lock_irq(&mchdev_lock);
3621
3622         rgvswctl = I915_READ16(MEMSWCTL);
3623
3624         /* Ack interrupts, disable EFC interrupt */
3625         I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
3626         I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
3627         I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
3628         I915_WRITE(DEIIR, DE_PCU_EVENT);
3629         I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
3630
3631         /* Go back to the starting frequency */
3632         ironlake_set_drps(dev, dev_priv->ips.fstart);
3633         mdelay(1);
3634         rgvswctl |= MEMCTL_CMD_STS;
3635         I915_WRITE(MEMSWCTL, rgvswctl);
3636         mdelay(1);
3637
3638         spin_unlock_irq(&mchdev_lock);
3639 }
3640
3641 /* There's a funny hw issue where the hw returns all 0 when reading from
3642  * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
3643  * ourselves, instead of doing a rmw cycle (which might result in us clearing
3644  * all limits and the gpu stuck at whatever frequency it is at atm).
3645  */
3646 static u32 gen6_rps_limits(struct drm_i915_private *dev_priv, u8 val)
3647 {
3648         u32 limits;
3649
3650         /* Only set the down limit when we've reached the lowest level to avoid
3651          * getting more interrupts, otherwise leave this clear. This prevents a
3652          * race in the hw when coming out of rc6: There's a tiny window where
3653          * the hw runs at the minimal clock before selecting the desired
3654          * frequency, if the down threshold expires in that window we will not
3655          * receive a down interrupt. */
3656         limits = dev_priv->rps.max_freq_softlimit << 24;
3657         if (val <= dev_priv->rps.min_freq_softlimit)
3658                 limits |= dev_priv->rps.min_freq_softlimit << 16;
3659
3660         return limits;
3661 }
3662
3663 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
3664 {
3665         int new_power;
3666
3667         new_power = dev_priv->rps.power;
3668         switch (dev_priv->rps.power) {
3669         case LOW_POWER:
3670                 if (val > dev_priv->rps.efficient_freq + 1 && val > dev_priv->rps.cur_freq)
3671                         new_power = BETWEEN;
3672                 break;
3673
3674         case BETWEEN:
3675                 if (val <= dev_priv->rps.efficient_freq && val < dev_priv->rps.cur_freq)
3676                         new_power = LOW_POWER;
3677                 else if (val >= dev_priv->rps.rp0_freq && val > dev_priv->rps.cur_freq)
3678                         new_power = HIGH_POWER;
3679                 break;
3680
3681         case HIGH_POWER:
3682                 if (val < (dev_priv->rps.rp1_freq + dev_priv->rps.rp0_freq) >> 1 && val < dev_priv->rps.cur_freq)
3683                         new_power = BETWEEN;
3684                 break;
3685         }
3686         /* Max/min bins are special */
3687         if (val == dev_priv->rps.min_freq_softlimit)
3688                 new_power = LOW_POWER;
3689         if (val == dev_priv->rps.max_freq_softlimit)
3690                 new_power = HIGH_POWER;
3691         if (new_power == dev_priv->rps.power)
3692                 return;
3693
3694         /* Note the units here are not exactly 1us, but 1280ns. */
3695         switch (new_power) {
3696         case LOW_POWER:
3697                 /* Upclock if more than 95% busy over 16ms */
3698                 I915_WRITE(GEN6_RP_UP_EI, 12500);
3699                 I915_WRITE(GEN6_RP_UP_THRESHOLD, 11800);
3700
3701                 /* Downclock if less than 85% busy over 32ms */
3702                 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3703                 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 21250);
3704
3705                 I915_WRITE(GEN6_RP_CONTROL,
3706                            GEN6_RP_MEDIA_TURBO |
3707                            GEN6_RP_MEDIA_HW_NORMAL_MODE |
3708                            GEN6_RP_MEDIA_IS_GFX |
3709                            GEN6_RP_ENABLE |
3710                            GEN6_RP_UP_BUSY_AVG |
3711                            GEN6_RP_DOWN_IDLE_AVG);
3712                 break;
3713
3714         case BETWEEN:
3715                 /* Upclock if more than 90% busy over 13ms */
3716                 I915_WRITE(GEN6_RP_UP_EI, 10250);
3717                 I915_WRITE(GEN6_RP_UP_THRESHOLD, 9225);
3718
3719                 /* Downclock if less than 75% busy over 32ms */
3720                 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3721                 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 18750);
3722
3723                 I915_WRITE(GEN6_RP_CONTROL,
3724                            GEN6_RP_MEDIA_TURBO |
3725                            GEN6_RP_MEDIA_HW_NORMAL_MODE |
3726                            GEN6_RP_MEDIA_IS_GFX |
3727                            GEN6_RP_ENABLE |
3728                            GEN6_RP_UP_BUSY_AVG |
3729                            GEN6_RP_DOWN_IDLE_AVG);
3730                 break;
3731
3732         case HIGH_POWER:
3733                 /* Upclock if more than 85% busy over 10ms */
3734                 I915_WRITE(GEN6_RP_UP_EI, 8000);
3735                 I915_WRITE(GEN6_RP_UP_THRESHOLD, 6800);
3736
3737                 /* Downclock if less than 60% busy over 32ms */
3738                 I915_WRITE(GEN6_RP_DOWN_EI, 25000);
3739                 I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 15000);
3740
3741                 I915_WRITE(GEN6_RP_CONTROL,
3742                            GEN6_RP_MEDIA_TURBO |
3743                            GEN6_RP_MEDIA_HW_NORMAL_MODE |
3744                            GEN6_RP_MEDIA_IS_GFX |
3745                            GEN6_RP_ENABLE |
3746                            GEN6_RP_UP_BUSY_AVG |
3747                            GEN6_RP_DOWN_IDLE_AVG);
3748                 break;
3749         }
3750
3751         dev_priv->rps.power = new_power;
3752         dev_priv->rps.last_adj = 0;
3753 }
3754
3755 static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
3756 {
3757         u32 mask = 0;
3758
3759         if (val > dev_priv->rps.min_freq_softlimit)
3760                 mask |= GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
3761         if (val < dev_priv->rps.max_freq_softlimit)
3762                 mask |= GEN6_PM_RP_UP_THRESHOLD;
3763
3764         mask |= dev_priv->pm_rps_events & (GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_UP_EI_EXPIRED);
3765         mask &= dev_priv->pm_rps_events;
3766
3767         return gen6_sanitize_rps_pm_mask(dev_priv, ~mask);
3768 }
3769
3770 /* gen6_set_rps is called to update the frequency request, but should also be
3771  * called when the range (min_delay and max_delay) is modified so that we can
3772  * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
3773 static void gen6_set_rps(struct drm_device *dev, u8 val)
3774 {
3775         struct drm_i915_private *dev_priv = dev->dev_private;
3776
3777         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3778         WARN_ON(val > dev_priv->rps.max_freq_softlimit);
3779         WARN_ON(val < dev_priv->rps.min_freq_softlimit);
3780
3781         /* min/max delay may still have been modified so be sure to
3782          * write the limits value.
3783          */
3784         if (val != dev_priv->rps.cur_freq) {
3785                 gen6_set_rps_thresholds(dev_priv, val);
3786
3787                 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
3788                         I915_WRITE(GEN6_RPNSWREQ,
3789                                    HSW_FREQUENCY(val));
3790                 else
3791                         I915_WRITE(GEN6_RPNSWREQ,
3792                                    GEN6_FREQUENCY(val) |
3793                                    GEN6_OFFSET(0) |
3794                                    GEN6_AGGRESSIVE_TURBO);
3795         }
3796
3797         /* Make sure we continue to get interrupts
3798          * until we hit the minimum or maximum frequencies.
3799          */
3800         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, gen6_rps_limits(dev_priv, val));
3801         I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
3802
3803         POSTING_READ(GEN6_RPNSWREQ);
3804
3805         dev_priv->rps.cur_freq = val;
3806         trace_intel_gpu_freq_change(val * 50);
3807 }
3808
3809 static void valleyview_set_rps(struct drm_device *dev, u8 val)
3810 {
3811         struct drm_i915_private *dev_priv = dev->dev_private;
3812
3813         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
3814         WARN_ON(val > dev_priv->rps.max_freq_softlimit);
3815         WARN_ON(val < dev_priv->rps.min_freq_softlimit);
3816
3817         if (WARN_ONCE(IS_CHERRYVIEW(dev) && (val & 1),
3818                       "Odd GPU freq value\n"))
3819                 val &= ~1;
3820
3821         if (val != dev_priv->rps.cur_freq)
3822                 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
3823
3824         I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
3825
3826         dev_priv->rps.cur_freq = val;
3827         trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
3828 }
3829
3830 /* vlv_set_rps_idle: Set the frequency to Rpn if Gfx clocks are down
3831  *
3832  * * If Gfx is Idle, then
3833  * 1. Mask Turbo interrupts
3834  * 2. Bring up Gfx clock
3835  * 3. Change the freq to Rpn and wait till P-Unit updates freq
3836  * 4. Clear the Force GFX CLK ON bit so that Gfx can down
3837  * 5. Unmask Turbo interrupts
3838 */
3839 static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
3840 {
3841         struct drm_device *dev = dev_priv->dev;
3842
3843         /* CHV and latest VLV don't need to force the gfx clock */
3844         if (IS_CHERRYVIEW(dev) || dev->pdev->revision >= 0xd) {
3845                 valleyview_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
3846                 return;
3847         }
3848
3849         /*
3850          * When we are idle.  Drop to min voltage state.
3851          */
3852
3853         if (dev_priv->rps.cur_freq <= dev_priv->rps.min_freq_softlimit)
3854                 return;
3855
3856         /* Mask turbo interrupt so that they will not come in between */
3857         I915_WRITE(GEN6_PMINTRMSK,
3858                    gen6_sanitize_rps_pm_mask(dev_priv, ~0));
3859
3860         vlv_force_gfx_clock(dev_priv, true);
3861
3862         dev_priv->rps.cur_freq = dev_priv->rps.min_freq_softlimit;
3863
3864         vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ,
3865                                         dev_priv->rps.min_freq_softlimit);
3866
3867         if (wait_for(((vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS))
3868                                 & GENFREQSTATUS) == 0, 100))
3869                 DRM_ERROR("timed out waiting for Punit\n");
3870
3871         vlv_force_gfx_clock(dev_priv, false);
3872
3873         I915_WRITE(GEN6_PMINTRMSK,
3874                    gen6_rps_pm_mask(dev_priv, dev_priv->rps.cur_freq));
3875 }
3876
3877 void gen6_rps_idle(struct drm_i915_private *dev_priv)
3878 {
3879         struct drm_device *dev = dev_priv->dev;
3880
3881         mutex_lock(&dev_priv->rps.hw_lock);
3882         if (dev_priv->rps.enabled) {
3883                 if (IS_VALLEYVIEW(dev))
3884                         vlv_set_rps_idle(dev_priv);
3885                 else
3886                         gen6_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
3887                 dev_priv->rps.last_adj = 0;
3888         }
3889         mutex_unlock(&dev_priv->rps.hw_lock);
3890 }
3891
3892 void gen6_rps_boost(struct drm_i915_private *dev_priv)
3893 {
3894         mutex_lock(&dev_priv->rps.hw_lock);
3895         if (dev_priv->rps.enabled) {
3896                 intel_set_rps(dev_priv->dev, dev_priv->rps.max_freq_softlimit);
3897                 dev_priv->rps.last_adj = 0;
3898         }
3899         mutex_unlock(&dev_priv->rps.hw_lock);
3900 }
3901
3902 void intel_set_rps(struct drm_device *dev, u8 val)
3903 {
3904         if (IS_VALLEYVIEW(dev))
3905                 valleyview_set_rps(dev, val);
3906         else
3907                 gen6_set_rps(dev, val);
3908 }
3909
3910 static void gen9_disable_rps(struct drm_device *dev)
3911 {
3912         struct drm_i915_private *dev_priv = dev->dev_private;
3913
3914         I915_WRITE(GEN6_RC_CONTROL, 0);
3915         I915_WRITE(GEN9_PG_ENABLE, 0);
3916 }
3917
3918 static void gen6_disable_rps(struct drm_device *dev)
3919 {
3920         struct drm_i915_private *dev_priv = dev->dev_private;
3921
3922         I915_WRITE(GEN6_RC_CONTROL, 0);
3923         I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
3924 }
3925
3926 static void cherryview_disable_rps(struct drm_device *dev)
3927 {
3928         struct drm_i915_private *dev_priv = dev->dev_private;
3929
3930         I915_WRITE(GEN6_RC_CONTROL, 0);
3931 }
3932
3933 static void valleyview_disable_rps(struct drm_device *dev)
3934 {
3935         struct drm_i915_private *dev_priv = dev->dev_private;
3936
3937         /* we're doing forcewake before Disabling RC6,
3938          * This what the BIOS expects when going into suspend */
3939         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
3940
3941         I915_WRITE(GEN6_RC_CONTROL, 0);
3942
3943         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
3944 }
3945
3946 static void intel_print_rc6_info(struct drm_device *dev, u32 mode)
3947 {
3948         if (IS_VALLEYVIEW(dev)) {
3949                 if (mode & (GEN7_RC_CTL_TO_MODE | GEN6_RC_CTL_EI_MODE(1)))
3950                         mode = GEN6_RC_CTL_RC6_ENABLE;
3951                 else
3952                         mode = 0;
3953         }
3954         if (HAS_RC6p(dev))
3955                 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s RC6p %s RC6pp %s\n",
3956                               (mode & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off",
3957                               (mode & GEN6_RC_CTL_RC6p_ENABLE) ? "on" : "off",
3958                               (mode & GEN6_RC_CTL_RC6pp_ENABLE) ? "on" : "off");
3959
3960         else
3961                 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s\n",
3962                               (mode & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off");
3963 }
3964
3965 static int sanitize_rc6_option(const struct drm_device *dev, int enable_rc6)
3966 {
3967         /* No RC6 before Ironlake */
3968         if (INTEL_INFO(dev)->gen < 5)
3969                 return 0;
3970
3971         /* RC6 is only on Ironlake mobile not on desktop */
3972         if (INTEL_INFO(dev)->gen == 5 && !IS_IRONLAKE_M(dev))
3973                 return 0;
3974
3975         /* Respect the kernel parameter if it is set */
3976         if (enable_rc6 >= 0) {
3977                 int mask;
3978
3979                 if (HAS_RC6p(dev))
3980                         mask = INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE |
3981                                INTEL_RC6pp_ENABLE;
3982                 else
3983                         mask = INTEL_RC6_ENABLE;
3984
3985                 if ((enable_rc6 & mask) != enable_rc6)
3986                         DRM_DEBUG_KMS("Adjusting RC6 mask to %d (requested %d, valid %d)\n",
3987                                       enable_rc6 & mask, enable_rc6, mask);
3988
3989                 return enable_rc6 & mask;
3990         }
3991
3992         /* Disable RC6 on Ironlake */
3993         if (INTEL_INFO(dev)->gen == 5)
3994                 return 0;
3995
3996         if (IS_IVYBRIDGE(dev))
3997                 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
3998
3999         return INTEL_RC6_ENABLE;
4000 }
4001
4002 int intel_enable_rc6(const struct drm_device *dev)
4003 {
4004         return i915.enable_rc6;
4005 }
4006
4007 static void gen6_init_rps_frequencies(struct drm_device *dev)
4008 {
4009         struct drm_i915_private *dev_priv = dev->dev_private;
4010         uint32_t rp_state_cap;
4011         u32 ddcc_status = 0;
4012         int ret;
4013
4014         rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
4015         /* All of these values are in units of 50MHz */
4016         dev_priv->rps.cur_freq          = 0;
4017         /* static values from HW: RP0 > RP1 > RPn (min_freq) */
4018         dev_priv->rps.rp0_freq          = (rp_state_cap >>  0) & 0xff;
4019         dev_priv->rps.rp1_freq          = (rp_state_cap >>  8) & 0xff;
4020         dev_priv->rps.min_freq          = (rp_state_cap >> 16) & 0xff;
4021         /* hw_max = RP0 until we check for overclocking */
4022         dev_priv->rps.max_freq          = dev_priv->rps.rp0_freq;
4023
4024         dev_priv->rps.efficient_freq = dev_priv->rps.rp1_freq;
4025         if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
4026                 ret = sandybridge_pcode_read(dev_priv,
4027                                         HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL,
4028                                         &ddcc_status);
4029                 if (0 == ret)
4030                         dev_priv->rps.efficient_freq =
4031                                 clamp_t(u8,
4032                                         ((ddcc_status >> 8) & 0xff),
4033                                         dev_priv->rps.min_freq,
4034                                         dev_priv->rps.max_freq);
4035         }
4036
4037         /* Preserve min/max settings in case of re-init */
4038         if (dev_priv->rps.max_freq_softlimit == 0)
4039                 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
4040
4041         if (dev_priv->rps.min_freq_softlimit == 0) {
4042                 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
4043                         dev_priv->rps.min_freq_softlimit =
4044                                 /* max(RPe, 450 MHz) */
4045                                 max(dev_priv->rps.efficient_freq, (u8) 9);
4046                 else
4047                         dev_priv->rps.min_freq_softlimit =
4048                                 dev_priv->rps.min_freq;
4049         }
4050 }
4051
4052 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
4053 static void gen9_enable_rps(struct drm_device *dev)
4054 {
4055         struct drm_i915_private *dev_priv = dev->dev_private;
4056
4057         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4058
4059         gen6_init_rps_frequencies(dev);
4060
4061         I915_WRITE(GEN6_RPNSWREQ, 0xc800000);
4062         I915_WRITE(GEN6_RC_VIDEO_FREQ, 0xc800000);
4063
4064         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 0xf4240);
4065         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, 0x12060000);
4066         I915_WRITE(GEN6_RP_UP_THRESHOLD, 0xe808);
4067         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 0x3bd08);
4068         I915_WRITE(GEN6_RP_UP_EI, 0x101d0);
4069         I915_WRITE(GEN6_RP_DOWN_EI, 0x55730);
4070         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 0xa);
4071         I915_WRITE(GEN6_PMINTRMSK, 0x6);
4072         I915_WRITE(GEN6_RP_CONTROL, GEN6_RP_MEDIA_TURBO |
4073                    GEN6_RP_MEDIA_HW_MODE | GEN6_RP_MEDIA_IS_GFX |
4074                    GEN6_RP_ENABLE | GEN6_RP_UP_BUSY_AVG |
4075                    GEN6_RP_DOWN_IDLE_AVG);
4076
4077         gen6_enable_rps_interrupts(dev);
4078
4079         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4080 }
4081
4082 static void gen9_enable_rc6(struct drm_device *dev)
4083 {
4084         struct drm_i915_private *dev_priv = dev->dev_private;
4085         struct intel_engine_cs *ring;
4086         uint32_t rc6_mask = 0;
4087         int unused;
4088
4089         /* 1a: Software RC state - RC0 */
4090         I915_WRITE(GEN6_RC_STATE, 0);
4091
4092         /* 1b: Get forcewake during program sequence. Although the driver
4093          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4094         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4095
4096         /* 2a: Disable RC states. */
4097         I915_WRITE(GEN6_RC_CONTROL, 0);
4098
4099         /* 2b: Program RC6 thresholds.*/
4100         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16);
4101         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
4102         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
4103         for_each_ring(ring, dev_priv, unused)
4104                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4105         I915_WRITE(GEN6_RC_SLEEP, 0);
4106         I915_WRITE(GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */
4107
4108         /* 2c: Program Coarse Power Gating Policies. */
4109         I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS, 25);
4110         I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS, 25);
4111
4112         /* 3a: Enable RC6 */
4113         if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
4114                 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
4115         DRM_INFO("RC6 %s\n", (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ?
4116                         "on" : "off");
4117         I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4118                                    GEN6_RC_CTL_EI_MODE(1) |
4119                                    rc6_mask);
4120
4121         /* 3b: Enable Coarse Power Gating only when RC6 is enabled */
4122         I915_WRITE(GEN9_PG_ENABLE, (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ? 3 : 0);
4123
4124         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4125
4126 }
4127
4128 static void gen8_enable_rps(struct drm_device *dev)
4129 {
4130         struct drm_i915_private *dev_priv = dev->dev_private;
4131         struct intel_engine_cs *ring;
4132         uint32_t rc6_mask = 0;
4133         int unused;
4134
4135         /* 1a: Software RC state - RC0 */
4136         I915_WRITE(GEN6_RC_STATE, 0);
4137
4138         /* 1c & 1d: Get forcewake during program sequence. Although the driver
4139          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4140         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4141
4142         /* 2a: Disable RC states. */
4143         I915_WRITE(GEN6_RC_CONTROL, 0);
4144
4145         /* Initialize rps frequencies */
4146         gen6_init_rps_frequencies(dev);
4147
4148         /* 2b: Program RC6 thresholds.*/
4149         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
4150         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
4151         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
4152         for_each_ring(ring, dev_priv, unused)
4153                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4154         I915_WRITE(GEN6_RC_SLEEP, 0);
4155         if (IS_BROADWELL(dev))
4156                 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */
4157         else
4158                 I915_WRITE(GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
4159
4160         /* 3: Enable RC6 */
4161         if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
4162                 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
4163         intel_print_rc6_info(dev, rc6_mask);
4164         if (IS_BROADWELL(dev))
4165                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4166                                 GEN7_RC_CTL_TO_MODE |
4167                                 rc6_mask);
4168         else
4169                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4170                                 GEN6_RC_CTL_EI_MODE(1) |
4171                                 rc6_mask);
4172
4173         /* 4 Program defaults and thresholds for RPS*/
4174         I915_WRITE(GEN6_RPNSWREQ,
4175                    HSW_FREQUENCY(dev_priv->rps.rp1_freq));
4176         I915_WRITE(GEN6_RC_VIDEO_FREQ,
4177                    HSW_FREQUENCY(dev_priv->rps.rp1_freq));
4178         /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
4179         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
4180
4181         /* Docs recommend 900MHz, and 300 MHz respectively */
4182         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
4183                    dev_priv->rps.max_freq_softlimit << 24 |
4184                    dev_priv->rps.min_freq_softlimit << 16);
4185
4186         I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
4187         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
4188         I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
4189         I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
4190
4191         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
4192
4193         /* 5: Enable RPS */
4194         I915_WRITE(GEN6_RP_CONTROL,
4195                    GEN6_RP_MEDIA_TURBO |
4196                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
4197                    GEN6_RP_MEDIA_IS_GFX |
4198                    GEN6_RP_ENABLE |
4199                    GEN6_RP_UP_BUSY_AVG |
4200                    GEN6_RP_DOWN_IDLE_AVG);
4201
4202         /* 6: Ring frequency + overclocking (our driver does this later */
4203
4204         dev_priv->rps.power = HIGH_POWER; /* force a reset */
4205         gen6_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
4206
4207         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4208 }
4209
4210 static void gen6_enable_rps(struct drm_device *dev)
4211 {
4212         struct drm_i915_private *dev_priv = dev->dev_private;
4213         struct intel_engine_cs *ring;
4214         u32 rc6vids, pcu_mbox = 0, rc6_mask = 0;
4215         u32 gtfifodbg;
4216         int rc6_mode;
4217         int i, ret;
4218
4219         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4220
4221         /* Here begins a magic sequence of register writes to enable
4222          * auto-downclocking.
4223          *
4224          * Perhaps there might be some value in exposing these to
4225          * userspace...
4226          */
4227         I915_WRITE(GEN6_RC_STATE, 0);
4228
4229         /* Clear the DBG now so we don't confuse earlier errors */
4230         if ((gtfifodbg = I915_READ(GTFIFODBG))) {
4231                 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
4232                 I915_WRITE(GTFIFODBG, gtfifodbg);
4233         }
4234
4235         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4236
4237         /* Initialize rps frequencies */
4238         gen6_init_rps_frequencies(dev);
4239
4240         /* disable the counters and set deterministic thresholds */
4241         I915_WRITE(GEN6_RC_CONTROL, 0);
4242
4243         I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
4244         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
4245         I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
4246         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
4247         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
4248
4249         for_each_ring(ring, dev_priv, i)
4250                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4251
4252         I915_WRITE(GEN6_RC_SLEEP, 0);
4253         I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
4254         if (IS_IVYBRIDGE(dev))
4255                 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
4256         else
4257                 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
4258         I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
4259         I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
4260
4261         /* Check if we are enabling RC6 */
4262         rc6_mode = intel_enable_rc6(dev_priv->dev);
4263         if (rc6_mode & INTEL_RC6_ENABLE)
4264                 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
4265
4266         /* We don't use those on Haswell */
4267         if (!IS_HASWELL(dev)) {
4268                 if (rc6_mode & INTEL_RC6p_ENABLE)
4269                         rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
4270
4271                 if (rc6_mode & INTEL_RC6pp_ENABLE)
4272                         rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
4273         }
4274
4275         intel_print_rc6_info(dev, rc6_mask);
4276
4277         I915_WRITE(GEN6_RC_CONTROL,
4278                    rc6_mask |
4279                    GEN6_RC_CTL_EI_MODE(1) |
4280                    GEN6_RC_CTL_HW_ENABLE);
4281
4282         /* Power down if completely idle for over 50ms */
4283         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
4284         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
4285
4286         ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_MIN_FREQ_TABLE, 0);
4287         if (ret)
4288                 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
4289
4290         ret = sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &pcu_mbox);
4291         if (!ret && (pcu_mbox & (1<<31))) { /* OC supported */
4292                 DRM_DEBUG_DRIVER("Overclocking supported. Max: %dMHz, Overclock max: %dMHz\n",
4293                                  (dev_priv->rps.max_freq_softlimit & 0xff) * 50,
4294                                  (pcu_mbox & 0xff) * 50);
4295                 dev_priv->rps.max_freq = pcu_mbox & 0xff;
4296         }
4297
4298         dev_priv->rps.power = HIGH_POWER; /* force a reset */
4299         gen6_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
4300
4301         rc6vids = 0;
4302         ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
4303         if (IS_GEN6(dev) && ret) {
4304                 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
4305         } else if (IS_GEN6(dev) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
4306                 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
4307                           GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
4308                 rc6vids &= 0xffff00;
4309                 rc6vids |= GEN6_ENCODE_RC6_VID(450);
4310                 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
4311                 if (ret)
4312                         DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
4313         }
4314
4315         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4316 }
4317
4318 static void __gen6_update_ring_freq(struct drm_device *dev)
4319 {
4320         struct drm_i915_private *dev_priv = dev->dev_private;
4321         int min_freq = 15;
4322         unsigned int gpu_freq;
4323         unsigned int max_ia_freq, min_ring_freq;
4324         int scaling_factor = 180;
4325         struct cpufreq_policy *policy;
4326
4327         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4328
4329         policy = cpufreq_cpu_get(0);
4330         if (policy) {
4331                 max_ia_freq = policy->cpuinfo.max_freq;
4332                 cpufreq_cpu_put(policy);
4333         } else {
4334                 /*
4335                  * Default to measured freq if none found, PCU will ensure we
4336                  * don't go over
4337                  */
4338                 max_ia_freq = tsc_khz;
4339         }
4340
4341         /* Convert from kHz to MHz */
4342         max_ia_freq /= 1000;
4343
4344         min_ring_freq = I915_READ(DCLK) & 0xf;
4345         /* convert DDR frequency from units of 266.6MHz to bandwidth */
4346         min_ring_freq = mult_frac(min_ring_freq, 8, 3);
4347
4348         /*
4349          * For each potential GPU frequency, load a ring frequency we'd like
4350          * to use for memory access.  We do this by specifying the IA frequency
4351          * the PCU should use as a reference to determine the ring frequency.
4352          */
4353         for (gpu_freq = dev_priv->rps.max_freq; gpu_freq >= dev_priv->rps.min_freq;
4354              gpu_freq--) {
4355                 int diff = dev_priv->rps.max_freq - gpu_freq;
4356                 unsigned int ia_freq = 0, ring_freq = 0;
4357
4358                 if (INTEL_INFO(dev)->gen >= 8) {
4359                         /* max(2 * GT, DDR). NB: GT is 50MHz units */
4360                         ring_freq = max(min_ring_freq, gpu_freq);
4361                 } else if (IS_HASWELL(dev)) {
4362                         ring_freq = mult_frac(gpu_freq, 5, 4);
4363                         ring_freq = max(min_ring_freq, ring_freq);
4364                         /* leave ia_freq as the default, chosen by cpufreq */
4365                 } else {
4366                         /* On older processors, there is no separate ring
4367                          * clock domain, so in order to boost the bandwidth
4368                          * of the ring, we need to upclock the CPU (ia_freq).
4369                          *
4370                          * For GPU frequencies less than 750MHz,
4371                          * just use the lowest ring freq.
4372                          */
4373                         if (gpu_freq < min_freq)
4374                                 ia_freq = 800;
4375                         else
4376                                 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
4377                         ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
4378                 }
4379
4380                 sandybridge_pcode_write(dev_priv,
4381                                         GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
4382                                         ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
4383                                         ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
4384                                         gpu_freq);
4385         }
4386 }
4387
4388 void gen6_update_ring_freq(struct drm_device *dev)
4389 {
4390         struct drm_i915_private *dev_priv = dev->dev_private;
4391
4392         if (INTEL_INFO(dev)->gen < 6 || IS_VALLEYVIEW(dev))
4393                 return;
4394
4395         mutex_lock(&dev_priv->rps.hw_lock);
4396         __gen6_update_ring_freq(dev);
4397         mutex_unlock(&dev_priv->rps.hw_lock);
4398 }
4399
4400 static int cherryview_rps_max_freq(struct drm_i915_private *dev_priv)
4401 {
4402         struct drm_device *dev = dev_priv->dev;
4403         u32 val, rp0;
4404
4405         if (dev->pdev->revision >= 0x20) {
4406                 val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
4407
4408                 switch (INTEL_INFO(dev)->eu_total) {
4409                 case 8:
4410                                 /* (2 * 4) config */
4411                                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT);
4412                                 break;
4413                 case 12:
4414                                 /* (2 * 6) config */
4415                                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT);
4416                                 break;
4417                 case 16:
4418                                 /* (2 * 8) config */
4419                 default:
4420                                 /* Setting (2 * 8) Min RP0 for any other combination */
4421                                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT);
4422                                 break;
4423                 }
4424                 rp0 = (rp0 & FB_GFX_FREQ_FUSE_MASK);
4425         } else {
4426                 /* For pre-production hardware */
4427                 val = vlv_punit_read(dev_priv, PUNIT_GPU_STATUS_REG);
4428                 rp0 = (val >> PUNIT_GPU_STATUS_MAX_FREQ_SHIFT) &
4429                        PUNIT_GPU_STATUS_MAX_FREQ_MASK;
4430         }
4431         return rp0;
4432 }
4433
4434 static int cherryview_rps_rpe_freq(struct drm_i915_private *dev_priv)
4435 {
4436         u32 val, rpe;
4437
4438         val = vlv_punit_read(dev_priv, PUNIT_GPU_DUTYCYCLE_REG);
4439         rpe = (val >> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK;
4440
4441         return rpe;
4442 }
4443
4444 static int cherryview_rps_guar_freq(struct drm_i915_private *dev_priv)
4445 {
4446         struct drm_device *dev = dev_priv->dev;
4447         u32 val, rp1;
4448
4449         if (dev->pdev->revision >= 0x20) {
4450                 val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
4451                 rp1 = (val & FB_GFX_FREQ_FUSE_MASK);
4452         } else {
4453                 /* For pre-production hardware */
4454                 val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
4455                 rp1 = ((val >> PUNIT_GPU_STATUS_MAX_FREQ_SHIFT) &
4456                        PUNIT_GPU_STATUS_MAX_FREQ_MASK);
4457         }
4458         return rp1;
4459 }
4460
4461 static int cherryview_rps_min_freq(struct drm_i915_private *dev_priv)
4462 {
4463         struct drm_device *dev = dev_priv->dev;
4464         u32 val, rpn;
4465
4466         if (dev->pdev->revision >= 0x20) {
4467                 val = vlv_punit_read(dev_priv, FB_GFX_FMIN_AT_VMIN_FUSE);
4468                 rpn = ((val >> FB_GFX_FMIN_AT_VMIN_FUSE_SHIFT) &
4469                        FB_GFX_FREQ_FUSE_MASK);
4470         } else { /* For pre-production hardware */
4471                 val = vlv_punit_read(dev_priv, PUNIT_GPU_STATUS_REG);
4472                 rpn = ((val >> PUNIT_GPU_STATIS_GFX_MIN_FREQ_SHIFT) &
4473                        PUNIT_GPU_STATUS_GFX_MIN_FREQ_MASK);
4474         }
4475
4476         return rpn;
4477 }
4478
4479 static int valleyview_rps_guar_freq(struct drm_i915_private *dev_priv)
4480 {
4481         u32 val, rp1;
4482
4483         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
4484
4485         rp1 = (val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT;
4486
4487         return rp1;
4488 }
4489
4490 static int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
4491 {
4492         u32 val, rp0;
4493
4494         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
4495
4496         rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
4497         /* Clamp to max */
4498         rp0 = min_t(u32, rp0, 0xea);
4499
4500         return rp0;
4501 }
4502
4503 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
4504 {
4505         u32 val, rpe;
4506
4507         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
4508         rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
4509         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
4510         rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
4511
4512         return rpe;
4513 }
4514
4515 static int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
4516 {
4517         return vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
4518 }
4519
4520 /* Check that the pctx buffer wasn't move under us. */
4521 static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
4522 {
4523         unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
4524
4525         WARN_ON(pctx_addr != dev_priv->mm.stolen_base +
4526                              dev_priv->vlv_pctx->stolen->start);
4527 }
4528
4529
4530 /* Check that the pcbr address is not empty. */
4531 static void cherryview_check_pctx(struct drm_i915_private *dev_priv)
4532 {
4533         unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
4534
4535         WARN_ON((pctx_addr >> VLV_PCBR_ADDR_SHIFT) == 0);
4536 }
4537
4538 static void cherryview_setup_pctx(struct drm_device *dev)
4539 {
4540         struct drm_i915_private *dev_priv = dev->dev_private;
4541         unsigned long pctx_paddr, paddr;
4542         struct i915_gtt *gtt = &dev_priv->gtt;
4543         u32 pcbr;
4544         int pctx_size = 32*1024;
4545
4546         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
4547
4548         pcbr = I915_READ(VLV_PCBR);
4549         if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) {
4550                 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
4551                 paddr = (dev_priv->mm.stolen_base +
4552                          (gtt->stolen_size - pctx_size));
4553
4554                 pctx_paddr = (paddr & (~4095));
4555                 I915_WRITE(VLV_PCBR, pctx_paddr);
4556         }
4557
4558         DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
4559 }
4560
4561 static void valleyview_setup_pctx(struct drm_device *dev)
4562 {
4563         struct drm_i915_private *dev_priv = dev->dev_private;
4564         struct drm_i915_gem_object *pctx;
4565         unsigned long pctx_paddr;
4566         u32 pcbr;
4567         int pctx_size = 24*1024;
4568
4569         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
4570
4571         pcbr = I915_READ(VLV_PCBR);
4572         if (pcbr) {
4573                 /* BIOS set it up already, grab the pre-alloc'd space */
4574                 int pcbr_offset;
4575
4576                 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
4577                 pctx = i915_gem_object_create_stolen_for_preallocated(dev_priv->dev,
4578                                                                       pcbr_offset,
4579                                                                       I915_GTT_OFFSET_NONE,
4580                                                                       pctx_size);
4581                 goto out;
4582         }
4583
4584         DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
4585
4586         /*
4587          * From the Gunit register HAS:
4588          * The Gfx driver is expected to program this register and ensure
4589          * proper allocation within Gfx stolen memory.  For example, this
4590          * register should be programmed such than the PCBR range does not
4591          * overlap with other ranges, such as the frame buffer, protected
4592          * memory, or any other relevant ranges.
4593          */
4594         pctx = i915_gem_object_create_stolen(dev, pctx_size);
4595         if (!pctx) {
4596                 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
4597                 return;
4598         }
4599
4600         pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
4601         I915_WRITE(VLV_PCBR, pctx_paddr);
4602
4603 out:
4604         DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
4605         dev_priv->vlv_pctx = pctx;
4606 }
4607
4608 static void valleyview_cleanup_pctx(struct drm_device *dev)
4609 {
4610         struct drm_i915_private *dev_priv = dev->dev_private;
4611
4612         if (WARN_ON(!dev_priv->vlv_pctx))
4613                 return;
4614
4615         drm_gem_object_unreference(&dev_priv->vlv_pctx->base);
4616         dev_priv->vlv_pctx = NULL;
4617 }
4618
4619 static void valleyview_init_gt_powersave(struct drm_device *dev)
4620 {
4621         struct drm_i915_private *dev_priv = dev->dev_private;
4622         u32 val;
4623
4624         valleyview_setup_pctx(dev);
4625
4626         mutex_lock(&dev_priv->rps.hw_lock);
4627
4628         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
4629         switch ((val >> 6) & 3) {
4630         case 0:
4631         case 1:
4632                 dev_priv->mem_freq = 800;
4633                 break;
4634         case 2:
4635                 dev_priv->mem_freq = 1066;
4636                 break;
4637         case 3:
4638                 dev_priv->mem_freq = 1333;
4639                 break;
4640         }
4641         DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
4642
4643         dev_priv->rps.max_freq = valleyview_rps_max_freq(dev_priv);
4644         dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
4645         DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
4646                          intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
4647                          dev_priv->rps.max_freq);
4648
4649         dev_priv->rps.efficient_freq = valleyview_rps_rpe_freq(dev_priv);
4650         DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
4651                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
4652                          dev_priv->rps.efficient_freq);
4653
4654         dev_priv->rps.rp1_freq = valleyview_rps_guar_freq(dev_priv);
4655         DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
4656                          intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
4657                          dev_priv->rps.rp1_freq);
4658
4659         dev_priv->rps.min_freq = valleyview_rps_min_freq(dev_priv);
4660         DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
4661                          intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
4662                          dev_priv->rps.min_freq);
4663
4664         /* Preserve min/max settings in case of re-init */
4665         if (dev_priv->rps.max_freq_softlimit == 0)
4666                 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
4667
4668         if (dev_priv->rps.min_freq_softlimit == 0)
4669                 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
4670
4671         mutex_unlock(&dev_priv->rps.hw_lock);
4672 }
4673
4674 static void cherryview_init_gt_powersave(struct drm_device *dev)
4675 {
4676         struct drm_i915_private *dev_priv = dev->dev_private;
4677         u32 val;
4678
4679         cherryview_setup_pctx(dev);
4680
4681         mutex_lock(&dev_priv->rps.hw_lock);
4682
4683         mutex_lock(&dev_priv->dpio_lock);
4684         val = vlv_cck_read(dev_priv, CCK_FUSE_REG);
4685         mutex_unlock(&dev_priv->dpio_lock);
4686
4687         switch ((val >> 2) & 0x7) {
4688         case 0:
4689         case 1:
4690                 dev_priv->rps.cz_freq = 200;
4691                 dev_priv->mem_freq = 1600;
4692                 break;
4693         case 2:
4694                 dev_priv->rps.cz_freq = 267;
4695                 dev_priv->mem_freq = 1600;
4696                 break;
4697         case 3:
4698                 dev_priv->rps.cz_freq = 333;
4699                 dev_priv->mem_freq = 2000;
4700                 break;
4701         case 4:
4702                 dev_priv->rps.cz_freq = 320;
4703                 dev_priv->mem_freq = 1600;
4704                 break;
4705         case 5:
4706                 dev_priv->rps.cz_freq = 400;
4707                 dev_priv->mem_freq = 1600;
4708                 break;
4709         }
4710         DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
4711
4712         dev_priv->rps.max_freq = cherryview_rps_max_freq(dev_priv);
4713         dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
4714         DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
4715                          intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
4716                          dev_priv->rps.max_freq);
4717
4718         dev_priv->rps.efficient_freq = cherryview_rps_rpe_freq(dev_priv);
4719         DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
4720                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
4721                          dev_priv->rps.efficient_freq);
4722
4723         dev_priv->rps.rp1_freq = cherryview_rps_guar_freq(dev_priv);
4724         DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
4725                          intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
4726                          dev_priv->rps.rp1_freq);
4727
4728         dev_priv->rps.min_freq = cherryview_rps_min_freq(dev_priv);
4729         DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
4730                          intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
4731                          dev_priv->rps.min_freq);
4732
4733         WARN_ONCE((dev_priv->rps.max_freq |
4734                    dev_priv->rps.efficient_freq |
4735                    dev_priv->rps.rp1_freq |
4736                    dev_priv->rps.min_freq) & 1,
4737                   "Odd GPU freq values\n");
4738
4739         /* Preserve min/max settings in case of re-init */
4740         if (dev_priv->rps.max_freq_softlimit == 0)
4741                 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
4742
4743         if (dev_priv->rps.min_freq_softlimit == 0)
4744                 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
4745
4746         mutex_unlock(&dev_priv->rps.hw_lock);
4747 }
4748
4749 static void valleyview_cleanup_gt_powersave(struct drm_device *dev)
4750 {
4751         valleyview_cleanup_pctx(dev);
4752 }
4753
4754 static void cherryview_enable_rps(struct drm_device *dev)
4755 {
4756         struct drm_i915_private *dev_priv = dev->dev_private;
4757         struct intel_engine_cs *ring;
4758         u32 gtfifodbg, val, rc6_mode = 0, pcbr;
4759         int i;
4760
4761         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4762
4763         gtfifodbg = I915_READ(GTFIFODBG);
4764         if (gtfifodbg) {
4765                 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
4766                                  gtfifodbg);
4767                 I915_WRITE(GTFIFODBG, gtfifodbg);
4768         }
4769
4770         cherryview_check_pctx(dev_priv);
4771
4772         /* 1a & 1b: Get forcewake during program sequence. Although the driver
4773          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4774         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4775
4776         /*  Disable RC states. */
4777         I915_WRITE(GEN6_RC_CONTROL, 0);
4778
4779         /* 2a: Program RC6 thresholds.*/
4780         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
4781         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
4782         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
4783
4784         for_each_ring(ring, dev_priv, i)
4785                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4786         I915_WRITE(GEN6_RC_SLEEP, 0);
4787
4788         /* TO threshold set to 1750 us ( 0x557 * 1.28 us) */
4789         I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
4790
4791         /* allows RC6 residency counter to work */
4792         I915_WRITE(VLV_COUNTER_CONTROL,
4793                    _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
4794                                       VLV_MEDIA_RC6_COUNT_EN |
4795                                       VLV_RENDER_RC6_COUNT_EN));
4796
4797         /* For now we assume BIOS is allocating and populating the PCBR  */
4798         pcbr = I915_READ(VLV_PCBR);
4799
4800         /* 3: Enable RC6 */
4801         if ((intel_enable_rc6(dev) & INTEL_RC6_ENABLE) &&
4802                                                 (pcbr >> VLV_PCBR_ADDR_SHIFT))
4803                 rc6_mode = GEN7_RC_CTL_TO_MODE;
4804
4805         I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
4806
4807         /* 4 Program defaults and thresholds for RPS*/
4808         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
4809         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
4810         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
4811         I915_WRITE(GEN6_RP_UP_EI, 66000);
4812         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
4813
4814         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
4815
4816         /* 5: Enable RPS */
4817         I915_WRITE(GEN6_RP_CONTROL,
4818                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
4819                    GEN6_RP_MEDIA_IS_GFX |
4820                    GEN6_RP_ENABLE |
4821                    GEN6_RP_UP_BUSY_AVG |
4822                    GEN6_RP_DOWN_IDLE_AVG);
4823
4824         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
4825
4826         /* RPS code assumes GPLL is used */
4827         WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
4828
4829         DRM_DEBUG_DRIVER("GPLL enabled? %s\n", val & GPLLENABLE ? "yes" : "no");
4830         DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
4831
4832         dev_priv->rps.cur_freq = (val >> 8) & 0xff;
4833         DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
4834                          intel_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
4835                          dev_priv->rps.cur_freq);
4836
4837         DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
4838                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
4839                          dev_priv->rps.efficient_freq);
4840
4841         valleyview_set_rps(dev_priv->dev, dev_priv->rps.efficient_freq);
4842
4843         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4844 }
4845
4846 static void valleyview_enable_rps(struct drm_device *dev)
4847 {
4848         struct drm_i915_private *dev_priv = dev->dev_private;
4849         struct intel_engine_cs *ring;
4850         u32 gtfifodbg, val, rc6_mode = 0;
4851         int i;
4852
4853         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4854
4855         valleyview_check_pctx(dev_priv);
4856
4857         if ((gtfifodbg = I915_READ(GTFIFODBG))) {
4858                 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
4859                                  gtfifodbg);
4860                 I915_WRITE(GTFIFODBG, gtfifodbg);
4861         }
4862
4863         /* If VLV, Forcewake all wells, else re-direct to regular path */
4864         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4865
4866         /*  Disable RC states. */
4867         I915_WRITE(GEN6_RC_CONTROL, 0);
4868
4869         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
4870         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
4871         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
4872         I915_WRITE(GEN6_RP_UP_EI, 66000);
4873         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
4874
4875         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
4876
4877         I915_WRITE(GEN6_RP_CONTROL,
4878                    GEN6_RP_MEDIA_TURBO |
4879                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
4880                    GEN6_RP_MEDIA_IS_GFX |
4881                    GEN6_RP_ENABLE |
4882                    GEN6_RP_UP_BUSY_AVG |
4883                    GEN6_RP_DOWN_IDLE_CONT);
4884
4885         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
4886         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
4887         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
4888
4889         for_each_ring(ring, dev_priv, i)
4890                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4891
4892         I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
4893
4894         /* allows RC6 residency counter to work */
4895         I915_WRITE(VLV_COUNTER_CONTROL,
4896                    _MASKED_BIT_ENABLE(VLV_MEDIA_RC0_COUNT_EN |
4897                                       VLV_RENDER_RC0_COUNT_EN |
4898                                       VLV_MEDIA_RC6_COUNT_EN |
4899                                       VLV_RENDER_RC6_COUNT_EN));
4900
4901         if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
4902                 rc6_mode = GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
4903
4904         intel_print_rc6_info(dev, rc6_mode);
4905
4906         I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
4907
4908         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
4909
4910         /* RPS code assumes GPLL is used */
4911         WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
4912
4913         DRM_DEBUG_DRIVER("GPLL enabled? %s\n", val & GPLLENABLE ? "yes" : "no");
4914         DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
4915
4916         dev_priv->rps.cur_freq = (val >> 8) & 0xff;
4917         DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
4918                          intel_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
4919                          dev_priv->rps.cur_freq);
4920
4921         DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
4922                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
4923                          dev_priv->rps.efficient_freq);
4924
4925         valleyview_set_rps(dev_priv->dev, dev_priv->rps.efficient_freq);
4926
4927         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4928 }
4929
4930 void ironlake_teardown_rc6(struct drm_device *dev)
4931 {
4932         struct drm_i915_private *dev_priv = dev->dev_private;
4933
4934         if (dev_priv->ips.renderctx) {
4935                 i915_gem_object_ggtt_unpin(dev_priv->ips.renderctx);
4936                 drm_gem_object_unreference(&dev_priv->ips.renderctx->base);
4937                 dev_priv->ips.renderctx = NULL;
4938         }
4939
4940         if (dev_priv->ips.pwrctx) {
4941                 i915_gem_object_ggtt_unpin(dev_priv->ips.pwrctx);
4942                 drm_gem_object_unreference(&dev_priv->ips.pwrctx->base);
4943                 dev_priv->ips.pwrctx = NULL;
4944         }
4945 }
4946
4947 static void ironlake_disable_rc6(struct drm_device *dev)
4948 {
4949         struct drm_i915_private *dev_priv = dev->dev_private;
4950
4951         if (I915_READ(PWRCTXA)) {
4952                 /* Wake the GPU, prevent RC6, then restore RSTDBYCTL */
4953                 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) | RCX_SW_EXIT);
4954                 wait_for(((I915_READ(RSTDBYCTL) & RSX_STATUS_MASK) == RSX_STATUS_ON),
4955                          50);
4956
4957                 I915_WRITE(PWRCTXA, 0);
4958                 POSTING_READ(PWRCTXA);
4959
4960                 I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
4961                 POSTING_READ(RSTDBYCTL);
4962         }
4963 }
4964
4965 static int ironlake_setup_rc6(struct drm_device *dev)
4966 {
4967         struct drm_i915_private *dev_priv = dev->dev_private;
4968
4969         if (dev_priv->ips.renderctx == NULL)
4970                 dev_priv->ips.renderctx = intel_alloc_context_page(dev);
4971         if (!dev_priv->ips.renderctx)
4972                 return -ENOMEM;
4973
4974         if (dev_priv->ips.pwrctx == NULL)
4975                 dev_priv->ips.pwrctx = intel_alloc_context_page(dev);
4976         if (!dev_priv->ips.pwrctx) {
4977                 ironlake_teardown_rc6(dev);
4978                 return -ENOMEM;
4979         }
4980
4981         return 0;
4982 }
4983
4984 static void ironlake_enable_rc6(struct drm_device *dev)
4985 {
4986         struct drm_i915_private *dev_priv = dev->dev_private;
4987         struct intel_engine_cs *ring = &dev_priv->ring[RCS];
4988         bool was_interruptible;
4989         int ret;
4990
4991         /* rc6 disabled by default due to repeated reports of hanging during
4992          * boot and resume.
4993          */
4994         if (!intel_enable_rc6(dev))
4995                 return;
4996
4997         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
4998
4999         ret = ironlake_setup_rc6(dev);
5000         if (ret)
5001                 return;
5002
5003         was_interruptible = dev_priv->mm.interruptible;
5004         dev_priv->mm.interruptible = false;
5005
5006         /*
5007          * GPU can automatically power down the render unit if given a page
5008          * to save state.
5009          */
5010         ret = intel_ring_begin(ring, 6);
5011         if (ret) {
5012                 ironlake_teardown_rc6(dev);
5013                 dev_priv->mm.interruptible = was_interruptible;
5014                 return;
5015         }
5016
5017         intel_ring_emit(ring, MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN);
5018         intel_ring_emit(ring, MI_SET_CONTEXT);
5019         intel_ring_emit(ring, i915_gem_obj_ggtt_offset(dev_priv->ips.renderctx) |
5020                         MI_MM_SPACE_GTT |
5021                         MI_SAVE_EXT_STATE_EN |
5022                         MI_RESTORE_EXT_STATE_EN |
5023                         MI_RESTORE_INHIBIT);
5024         intel_ring_emit(ring, MI_SUSPEND_FLUSH);
5025         intel_ring_emit(ring, MI_NOOP);
5026         intel_ring_emit(ring, MI_FLUSH);
5027         intel_ring_advance(ring);
5028
5029         /*
5030          * Wait for the command parser to advance past MI_SET_CONTEXT. The HW
5031          * does an implicit flush, combined with MI_FLUSH above, it should be
5032          * safe to assume that renderctx is valid
5033          */
5034         ret = intel_ring_idle(ring);
5035         dev_priv->mm.interruptible = was_interruptible;
5036         if (ret) {
5037                 DRM_ERROR("failed to enable ironlake power savings\n");
5038                 ironlake_teardown_rc6(dev);
5039                 return;
5040         }
5041
5042         I915_WRITE(PWRCTXA, i915_gem_obj_ggtt_offset(dev_priv->ips.pwrctx) | PWRCTX_EN);
5043         I915_WRITE(RSTDBYCTL, I915_READ(RSTDBYCTL) & ~RCX_SW_EXIT);
5044
5045         intel_print_rc6_info(dev, GEN6_RC_CTL_RC6_ENABLE);
5046 }
5047
5048 static unsigned long intel_pxfreq(u32 vidfreq)
5049 {
5050         unsigned long freq;
5051         int div = (vidfreq & 0x3f0000) >> 16;
5052         int post = (vidfreq & 0x3000) >> 12;
5053         int pre = (vidfreq & 0x7);
5054
5055         if (!pre)
5056                 return 0;
5057
5058         freq = ((div * 133333) / ((1<<post) * pre));
5059
5060         return freq;
5061 }
5062
5063 static const struct cparams {
5064         u16 i;
5065         u16 t;
5066         u16 m;
5067         u16 c;
5068 } cparams[] = {
5069         { 1, 1333, 301, 28664 },
5070         { 1, 1066, 294, 24460 },
5071         { 1, 800, 294, 25192 },
5072         { 0, 1333, 276, 27605 },
5073         { 0, 1066, 276, 27605 },
5074         { 0, 800, 231, 23784 },
5075 };
5076
5077 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
5078 {
5079         u64 total_count, diff, ret;
5080         u32 count1, count2, count3, m = 0, c = 0;
5081         unsigned long now = jiffies_to_msecs(jiffies), diff1;
5082         int i;
5083
5084         assert_spin_locked(&mchdev_lock);
5085
5086         diff1 = now - dev_priv->ips.last_time1;
5087
5088         /* Prevent division-by-zero if we are asking too fast.
5089          * Also, we don't get interesting results if we are polling
5090          * faster than once in 10ms, so just return the saved value
5091          * in such cases.
5092          */
5093         if (diff1 <= 10)
5094                 return dev_priv->ips.chipset_power;
5095
5096         count1 = I915_READ(DMIEC);
5097         count2 = I915_READ(DDREC);
5098         count3 = I915_READ(CSIEC);
5099
5100         total_count = count1 + count2 + count3;
5101
5102         /* FIXME: handle per-counter overflow */
5103         if (total_count < dev_priv->ips.last_count1) {
5104                 diff = ~0UL - dev_priv->ips.last_count1;
5105                 diff += total_count;
5106         } else {
5107                 diff = total_count - dev_priv->ips.last_count1;
5108         }
5109
5110         for (i = 0; i < ARRAY_SIZE(cparams); i++) {
5111                 if (cparams[i].i == dev_priv->ips.c_m &&
5112                     cparams[i].t == dev_priv->ips.r_t) {
5113                         m = cparams[i].m;
5114                         c = cparams[i].c;
5115                         break;
5116                 }
5117         }
5118
5119         diff = div_u64(diff, diff1);
5120         ret = ((m * diff) + c);
5121         ret = div_u64(ret, 10);
5122
5123         dev_priv->ips.last_count1 = total_count;
5124         dev_priv->ips.last_time1 = now;
5125
5126         dev_priv->ips.chipset_power = ret;
5127
5128         return ret;
5129 }
5130
5131 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
5132 {
5133         struct drm_device *dev = dev_priv->dev;
5134         unsigned long val;
5135
5136         if (INTEL_INFO(dev)->gen != 5)
5137                 return 0;
5138
5139         spin_lock_irq(&mchdev_lock);
5140
5141         val = __i915_chipset_val(dev_priv);
5142
5143         spin_unlock_irq(&mchdev_lock);
5144
5145         return val;
5146 }
5147
5148 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
5149 {
5150         unsigned long m, x, b;
5151         u32 tsfs;
5152
5153         tsfs = I915_READ(TSFS);
5154
5155         m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
5156         x = I915_READ8(TR1);
5157
5158         b = tsfs & TSFS_INTR_MASK;
5159
5160         return ((m * x) / 127) - b;
5161 }
5162
5163 static int _pxvid_to_vd(u8 pxvid)
5164 {
5165         if (pxvid == 0)
5166                 return 0;
5167
5168         if (pxvid >= 8 && pxvid < 31)
5169                 pxvid = 31;
5170
5171         return (pxvid + 2) * 125;
5172 }
5173
5174 static u32 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
5175 {
5176         struct drm_device *dev = dev_priv->dev;
5177         const int vd = _pxvid_to_vd(pxvid);
5178         const int vm = vd - 1125;
5179
5180         if (INTEL_INFO(dev)->is_mobile)
5181                 return vm > 0 ? vm : 0;
5182
5183         return vd;
5184 }
5185
5186 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
5187 {
5188         u64 now, diff, diffms;
5189         u32 count;
5190
5191         assert_spin_locked(&mchdev_lock);
5192
5193         now = ktime_get_raw_ns();
5194         diffms = now - dev_priv->ips.last_time2;
5195         do_div(diffms, NSEC_PER_MSEC);
5196
5197         /* Don't divide by 0 */
5198         if (!diffms)
5199                 return;
5200
5201         count = I915_READ(GFXEC);
5202
5203         if (count < dev_priv->ips.last_count2) {
5204                 diff = ~0UL - dev_priv->ips.last_count2;
5205                 diff += count;
5206         } else {
5207                 diff = count - dev_priv->ips.last_count2;
5208         }
5209
5210         dev_priv->ips.last_count2 = count;
5211         dev_priv->ips.last_time2 = now;
5212
5213         /* More magic constants... */
5214         diff = diff * 1181;
5215         diff = div_u64(diff, diffms * 10);
5216         dev_priv->ips.gfx_power = diff;
5217 }
5218
5219 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
5220 {
5221         struct drm_device *dev = dev_priv->dev;
5222
5223         if (INTEL_INFO(dev)->gen != 5)
5224                 return;
5225
5226         spin_lock_irq(&mchdev_lock);
5227
5228         __i915_update_gfx_val(dev_priv);
5229
5230         spin_unlock_irq(&mchdev_lock);
5231 }
5232
5233 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
5234 {
5235         unsigned long t, corr, state1, corr2, state2;
5236         u32 pxvid, ext_v;
5237
5238         assert_spin_locked(&mchdev_lock);
5239
5240         pxvid = I915_READ(PXVFREQ_BASE + (dev_priv->rps.cur_freq * 4));
5241         pxvid = (pxvid >> 24) & 0x7f;
5242         ext_v = pvid_to_extvid(dev_priv, pxvid);
5243
5244         state1 = ext_v;
5245
5246         t = i915_mch_val(dev_priv);
5247
5248         /* Revel in the empirically derived constants */
5249
5250         /* Correction factor in 1/100000 units */
5251         if (t > 80)
5252                 corr = ((t * 2349) + 135940);
5253         else if (t >= 50)
5254                 corr = ((t * 964) + 29317);
5255         else /* < 50 */
5256                 corr = ((t * 301) + 1004);
5257
5258         corr = corr * ((150142 * state1) / 10000 - 78642);
5259         corr /= 100000;
5260         corr2 = (corr * dev_priv->ips.corr);
5261
5262         state2 = (corr2 * state1) / 10000;
5263         state2 /= 100; /* convert to mW */
5264
5265         __i915_update_gfx_val(dev_priv);
5266
5267         return dev_priv->ips.gfx_power + state2;
5268 }
5269
5270 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
5271 {
5272         struct drm_device *dev = dev_priv->dev;
5273         unsigned long val;
5274
5275         if (INTEL_INFO(dev)->gen != 5)
5276                 return 0;
5277
5278         spin_lock_irq(&mchdev_lock);
5279
5280         val = __i915_gfx_val(dev_priv);
5281
5282         spin_unlock_irq(&mchdev_lock);
5283
5284         return val;
5285 }
5286
5287 /**
5288  * i915_read_mch_val - return value for IPS use
5289  *
5290  * Calculate and return a value for the IPS driver to use when deciding whether
5291  * we have thermal and power headroom to increase CPU or GPU power budget.
5292  */
5293 unsigned long i915_read_mch_val(void)
5294 {
5295         struct drm_i915_private *dev_priv;
5296         unsigned long chipset_val, graphics_val, ret = 0;
5297
5298         spin_lock_irq(&mchdev_lock);
5299         if (!i915_mch_dev)
5300                 goto out_unlock;
5301         dev_priv = i915_mch_dev;
5302
5303         chipset_val = __i915_chipset_val(dev_priv);
5304         graphics_val = __i915_gfx_val(dev_priv);
5305
5306         ret = chipset_val + graphics_val;
5307
5308 out_unlock:
5309         spin_unlock_irq(&mchdev_lock);
5310
5311         return ret;
5312 }
5313 EXPORT_SYMBOL_GPL(i915_read_mch_val);
5314
5315 /**
5316  * i915_gpu_raise - raise GPU frequency limit
5317  *
5318  * Raise the limit; IPS indicates we have thermal headroom.
5319  */
5320 bool i915_gpu_raise(void)
5321 {
5322         struct drm_i915_private *dev_priv;
5323         bool ret = true;
5324
5325         spin_lock_irq(&mchdev_lock);
5326         if (!i915_mch_dev) {
5327                 ret = false;
5328                 goto out_unlock;
5329         }
5330         dev_priv = i915_mch_dev;
5331
5332         if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
5333                 dev_priv->ips.max_delay--;
5334
5335 out_unlock:
5336         spin_unlock_irq(&mchdev_lock);
5337
5338         return ret;
5339 }
5340 EXPORT_SYMBOL_GPL(i915_gpu_raise);
5341
5342 /**
5343  * i915_gpu_lower - lower GPU frequency limit
5344  *
5345  * IPS indicates we're close to a thermal limit, so throttle back the GPU
5346  * frequency maximum.
5347  */
5348 bool i915_gpu_lower(void)
5349 {
5350         struct drm_i915_private *dev_priv;
5351         bool ret = true;
5352
5353         spin_lock_irq(&mchdev_lock);
5354         if (!i915_mch_dev) {
5355                 ret = false;
5356                 goto out_unlock;
5357         }
5358         dev_priv = i915_mch_dev;
5359
5360         if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
5361                 dev_priv->ips.max_delay++;
5362
5363 out_unlock:
5364         spin_unlock_irq(&mchdev_lock);
5365
5366         return ret;
5367 }
5368 EXPORT_SYMBOL_GPL(i915_gpu_lower);
5369
5370 /**
5371  * i915_gpu_busy - indicate GPU business to IPS
5372  *
5373  * Tell the IPS driver whether or not the GPU is busy.
5374  */
5375 bool i915_gpu_busy(void)
5376 {
5377         struct drm_i915_private *dev_priv;
5378         struct intel_engine_cs *ring;
5379         bool ret = false;
5380         int i;
5381
5382         spin_lock_irq(&mchdev_lock);
5383         if (!i915_mch_dev)
5384                 goto out_unlock;
5385         dev_priv = i915_mch_dev;
5386
5387         for_each_ring(ring, dev_priv, i)
5388                 ret |= !list_empty(&ring->request_list);
5389
5390 out_unlock:
5391         spin_unlock_irq(&mchdev_lock);
5392
5393         return ret;
5394 }
5395 EXPORT_SYMBOL_GPL(i915_gpu_busy);
5396
5397 /**
5398  * i915_gpu_turbo_disable - disable graphics turbo
5399  *
5400  * Disable graphics turbo by resetting the max frequency and setting the
5401  * current frequency to the default.
5402  */
5403 bool i915_gpu_turbo_disable(void)
5404 {
5405         struct drm_i915_private *dev_priv;
5406         bool ret = true;
5407
5408         spin_lock_irq(&mchdev_lock);
5409         if (!i915_mch_dev) {
5410                 ret = false;
5411                 goto out_unlock;
5412         }
5413         dev_priv = i915_mch_dev;
5414
5415         dev_priv->ips.max_delay = dev_priv->ips.fstart;
5416
5417         if (!ironlake_set_drps(dev_priv->dev, dev_priv->ips.fstart))
5418                 ret = false;
5419
5420 out_unlock:
5421         spin_unlock_irq(&mchdev_lock);
5422
5423         return ret;
5424 }
5425 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
5426
5427 /**
5428  * Tells the intel_ips driver that the i915 driver is now loaded, if
5429  * IPS got loaded first.
5430  *
5431  * This awkward dance is so that neither module has to depend on the
5432  * other in order for IPS to do the appropriate communication of
5433  * GPU turbo limits to i915.
5434  */
5435 static void
5436 ips_ping_for_i915_load(void)
5437 {
5438         void (*link)(void);
5439
5440         link = symbol_get(ips_link_to_i915_driver);
5441         if (link) {
5442                 link();
5443                 symbol_put(ips_link_to_i915_driver);
5444         }
5445 }
5446
5447 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
5448 {
5449         /* We only register the i915 ips part with intel-ips once everything is
5450          * set up, to avoid intel-ips sneaking in and reading bogus values. */
5451         spin_lock_irq(&mchdev_lock);
5452         i915_mch_dev = dev_priv;
5453         spin_unlock_irq(&mchdev_lock);
5454
5455         ips_ping_for_i915_load();
5456 }
5457
5458 void intel_gpu_ips_teardown(void)
5459 {
5460         spin_lock_irq(&mchdev_lock);
5461         i915_mch_dev = NULL;
5462         spin_unlock_irq(&mchdev_lock);
5463 }
5464
5465 static void intel_init_emon(struct drm_device *dev)
5466 {
5467         struct drm_i915_private *dev_priv = dev->dev_private;
5468         u32 lcfuse;
5469         u8 pxw[16];
5470         int i;
5471
5472         /* Disable to program */
5473         I915_WRITE(ECR, 0);
5474         POSTING_READ(ECR);
5475
5476         /* Program energy weights for various events */
5477         I915_WRITE(SDEW, 0x15040d00);
5478         I915_WRITE(CSIEW0, 0x007f0000);
5479         I915_WRITE(CSIEW1, 0x1e220004);
5480         I915_WRITE(CSIEW2, 0x04000004);
5481
5482         for (i = 0; i < 5; i++)
5483                 I915_WRITE(PEW + (i * 4), 0);
5484         for (i = 0; i < 3; i++)
5485                 I915_WRITE(DEW + (i * 4), 0);
5486
5487         /* Program P-state weights to account for frequency power adjustment */
5488         for (i = 0; i < 16; i++) {
5489                 u32 pxvidfreq = I915_READ(PXVFREQ_BASE + (i * 4));
5490                 unsigned long freq = intel_pxfreq(pxvidfreq);
5491                 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
5492                         PXVFREQ_PX_SHIFT;
5493                 unsigned long val;
5494
5495                 val = vid * vid;
5496                 val *= (freq / 1000);
5497                 val *= 255;
5498                 val /= (127*127*900);
5499                 if (val > 0xff)
5500                         DRM_ERROR("bad pxval: %ld\n", val);
5501                 pxw[i] = val;
5502         }
5503         /* Render standby states get 0 weight */
5504         pxw[14] = 0;
5505         pxw[15] = 0;
5506
5507         for (i = 0; i < 4; i++) {
5508                 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
5509                         (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
5510                 I915_WRITE(PXW + (i * 4), val);
5511         }
5512
5513         /* Adjust magic regs to magic values (more experimental results) */
5514         I915_WRITE(OGW0, 0);
5515         I915_WRITE(OGW1, 0);
5516         I915_WRITE(EG0, 0x00007f00);
5517         I915_WRITE(EG1, 0x0000000e);
5518         I915_WRITE(EG2, 0x000e0000);
5519         I915_WRITE(EG3, 0x68000300);
5520         I915_WRITE(EG4, 0x42000000);
5521         I915_WRITE(EG5, 0x00140031);
5522         I915_WRITE(EG6, 0);
5523         I915_WRITE(EG7, 0);
5524
5525         for (i = 0; i < 8; i++)
5526                 I915_WRITE(PXWL + (i * 4), 0);
5527
5528         /* Enable PMON + select events */
5529         I915_WRITE(ECR, 0x80000019);
5530
5531         lcfuse = I915_READ(LCFUSE02);
5532
5533         dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
5534 }
5535
5536 void intel_init_gt_powersave(struct drm_device *dev)
5537 {
5538         i915.enable_rc6 = sanitize_rc6_option(dev, i915.enable_rc6);
5539
5540         if (IS_CHERRYVIEW(dev))
5541                 cherryview_init_gt_powersave(dev);
5542         else if (IS_VALLEYVIEW(dev))
5543                 valleyview_init_gt_powersave(dev);
5544 }
5545
5546 void intel_cleanup_gt_powersave(struct drm_device *dev)
5547 {
5548         if (IS_CHERRYVIEW(dev))
5549                 return;
5550         else if (IS_VALLEYVIEW(dev))
5551                 valleyview_cleanup_gt_powersave(dev);
5552 }
5553
5554 static void gen6_suspend_rps(struct drm_device *dev)
5555 {
5556         struct drm_i915_private *dev_priv = dev->dev_private;
5557
5558         flush_delayed_work(&dev_priv->rps.delayed_resume_work);
5559
5560         /*
5561          * TODO: disable RPS interrupts on GEN9+ too once RPS support
5562          * is added for it.
5563          */
5564         if (INTEL_INFO(dev)->gen < 9)
5565                 gen6_disable_rps_interrupts(dev);
5566 }
5567
5568 /**
5569  * intel_suspend_gt_powersave - suspend PM work and helper threads
5570  * @dev: drm device
5571  *
5572  * We don't want to disable RC6 or other features here, we just want
5573  * to make sure any work we've queued has finished and won't bother
5574  * us while we're suspended.
5575  */
5576 void intel_suspend_gt_powersave(struct drm_device *dev)
5577 {
5578         struct drm_i915_private *dev_priv = dev->dev_private;
5579
5580         if (INTEL_INFO(dev)->gen < 6)
5581                 return;
5582
5583         gen6_suspend_rps(dev);
5584
5585         /* Force GPU to min freq during suspend */
5586         gen6_rps_idle(dev_priv);
5587 }
5588
5589 void intel_disable_gt_powersave(struct drm_device *dev)
5590 {
5591         struct drm_i915_private *dev_priv = dev->dev_private;
5592
5593         if (IS_IRONLAKE_M(dev)) {
5594                 ironlake_disable_drps(dev);
5595                 ironlake_disable_rc6(dev);
5596         } else if (INTEL_INFO(dev)->gen >= 6) {
5597                 intel_suspend_gt_powersave(dev);
5598
5599                 mutex_lock(&dev_priv->rps.hw_lock);
5600                 if (INTEL_INFO(dev)->gen >= 9)
5601                         gen9_disable_rps(dev);
5602                 else if (IS_CHERRYVIEW(dev))
5603                         cherryview_disable_rps(dev);
5604                 else if (IS_VALLEYVIEW(dev))
5605                         valleyview_disable_rps(dev);
5606                 else
5607                         gen6_disable_rps(dev);
5608
5609                 dev_priv->rps.enabled = false;
5610                 mutex_unlock(&dev_priv->rps.hw_lock);
5611         }
5612 }
5613
5614 static void intel_gen6_powersave_work(struct work_struct *work)
5615 {
5616         struct drm_i915_private *dev_priv =
5617                 container_of(work, struct drm_i915_private,
5618                              rps.delayed_resume_work.work);
5619         struct drm_device *dev = dev_priv->dev;
5620
5621         mutex_lock(&dev_priv->rps.hw_lock);
5622
5623         /*
5624          * TODO: reset/enable RPS interrupts on GEN9+ too, once RPS support is
5625          * added for it.
5626          */
5627         if (INTEL_INFO(dev)->gen < 9)
5628                 gen6_reset_rps_interrupts(dev);
5629
5630         if (IS_CHERRYVIEW(dev)) {
5631                 cherryview_enable_rps(dev);
5632         } else if (IS_VALLEYVIEW(dev)) {
5633                 valleyview_enable_rps(dev);
5634         } else if (INTEL_INFO(dev)->gen >= 9) {
5635                 gen9_enable_rc6(dev);
5636                 gen9_enable_rps(dev);
5637                 __gen6_update_ring_freq(dev);
5638         } else if (IS_BROADWELL(dev)) {
5639                 gen8_enable_rps(dev);
5640                 __gen6_update_ring_freq(dev);
5641         } else {
5642                 gen6_enable_rps(dev);
5643                 __gen6_update_ring_freq(dev);
5644         }
5645         dev_priv->rps.enabled = true;
5646
5647         if (INTEL_INFO(dev)->gen < 9)
5648                 gen6_enable_rps_interrupts(dev);
5649
5650         mutex_unlock(&dev_priv->rps.hw_lock);
5651
5652         intel_runtime_pm_put(dev_priv);
5653 }
5654
5655 void intel_enable_gt_powersave(struct drm_device *dev)
5656 {
5657         struct drm_i915_private *dev_priv = dev->dev_private;
5658
5659         /* Powersaving is controlled by the host when inside a VM */
5660         if (intel_vgpu_active(dev))
5661                 return;
5662
5663         if (IS_IRONLAKE_M(dev)) {
5664                 mutex_lock(&dev->struct_mutex);
5665                 ironlake_enable_drps(dev);
5666                 ironlake_enable_rc6(dev);
5667                 intel_init_emon(dev);
5668                 mutex_unlock(&dev->struct_mutex);
5669         } else if (INTEL_INFO(dev)->gen >= 6) {
5670                 /*
5671                  * PCU communication is slow and this doesn't need to be
5672                  * done at any specific time, so do this out of our fast path
5673                  * to make resume and init faster.
5674                  *
5675                  * We depend on the HW RC6 power context save/restore
5676                  * mechanism when entering D3 through runtime PM suspend. So
5677                  * disable RPM until RPS/RC6 is properly setup. We can only
5678                  * get here via the driver load/system resume/runtime resume
5679                  * paths, so the _noresume version is enough (and in case of
5680                  * runtime resume it's necessary).
5681                  */
5682                 if (schedule_delayed_work(&dev_priv->rps.delayed_resume_work,
5683                                            round_jiffies_up_relative(HZ)))
5684                         intel_runtime_pm_get_noresume(dev_priv);
5685         }
5686 }
5687
5688 void intel_reset_gt_powersave(struct drm_device *dev)
5689 {
5690         struct drm_i915_private *dev_priv = dev->dev_private;
5691
5692         if (INTEL_INFO(dev)->gen < 6)
5693                 return;
5694
5695         gen6_suspend_rps(dev);
5696         dev_priv->rps.enabled = false;
5697 }
5698
5699 static void ibx_init_clock_gating(struct drm_device *dev)
5700 {
5701         struct drm_i915_private *dev_priv = dev->dev_private;
5702
5703         /*
5704          * On Ibex Peak and Cougar Point, we need to disable clock
5705          * gating for the panel power sequencer or it will fail to
5706          * start up when no ports are active.
5707          */
5708         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
5709 }
5710
5711 static void g4x_disable_trickle_feed(struct drm_device *dev)
5712 {
5713         struct drm_i915_private *dev_priv = dev->dev_private;
5714         int pipe;
5715
5716         for_each_pipe(dev_priv, pipe) {
5717                 I915_WRITE(DSPCNTR(pipe),
5718                            I915_READ(DSPCNTR(pipe)) |
5719                            DISPPLANE_TRICKLE_FEED_DISABLE);
5720                 intel_flush_primary_plane(dev_priv, pipe);
5721         }
5722 }
5723
5724 static void ilk_init_lp_watermarks(struct drm_device *dev)
5725 {
5726         struct drm_i915_private *dev_priv = dev->dev_private;
5727
5728         I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
5729         I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
5730         I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
5731
5732         /*
5733          * Don't touch WM1S_LP_EN here.
5734          * Doing so could cause underruns.
5735          */
5736 }
5737
5738 static void ironlake_init_clock_gating(struct drm_device *dev)
5739 {
5740         struct drm_i915_private *dev_priv = dev->dev_private;
5741         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
5742
5743         /*
5744          * Required for FBC
5745          * WaFbcDisableDpfcClockGating:ilk
5746          */
5747         dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
5748                    ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
5749                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
5750
5751         I915_WRITE(PCH_3DCGDIS0,
5752                    MARIUNIT_CLOCK_GATE_DISABLE |
5753                    SVSMUNIT_CLOCK_GATE_DISABLE);
5754         I915_WRITE(PCH_3DCGDIS1,
5755                    VFMUNIT_CLOCK_GATE_DISABLE);
5756
5757         /*
5758          * According to the spec the following bits should be set in
5759          * order to enable memory self-refresh
5760          * The bit 22/21 of 0x42004
5761          * The bit 5 of 0x42020
5762          * The bit 15 of 0x45000
5763          */
5764         I915_WRITE(ILK_DISPLAY_CHICKEN2,
5765                    (I915_READ(ILK_DISPLAY_CHICKEN2) |
5766                     ILK_DPARB_GATE | ILK_VSDPFD_FULL));
5767         dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
5768         I915_WRITE(DISP_ARB_CTL,
5769                    (I915_READ(DISP_ARB_CTL) |
5770                     DISP_FBC_WM_DIS));
5771
5772         ilk_init_lp_watermarks(dev);
5773
5774         /*
5775          * Based on the document from hardware guys the following bits
5776          * should be set unconditionally in order to enable FBC.
5777          * The bit 22 of 0x42000
5778          * The bit 22 of 0x42004
5779          * The bit 7,8,9 of 0x42020.
5780          */
5781         if (IS_IRONLAKE_M(dev)) {
5782                 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
5783                 I915_WRITE(ILK_DISPLAY_CHICKEN1,
5784                            I915_READ(ILK_DISPLAY_CHICKEN1) |
5785                            ILK_FBCQ_DIS);
5786                 I915_WRITE(ILK_DISPLAY_CHICKEN2,
5787                            I915_READ(ILK_DISPLAY_CHICKEN2) |
5788                            ILK_DPARB_GATE);
5789         }
5790
5791         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
5792
5793         I915_WRITE(ILK_DISPLAY_CHICKEN2,
5794                    I915_READ(ILK_DISPLAY_CHICKEN2) |
5795                    ILK_ELPIN_409_SELECT);
5796         I915_WRITE(_3D_CHICKEN2,
5797                    _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
5798                    _3D_CHICKEN2_WM_READ_PIPELINED);
5799
5800         /* WaDisableRenderCachePipelinedFlush:ilk */
5801         I915_WRITE(CACHE_MODE_0,
5802                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
5803
5804         /* WaDisable_RenderCache_OperationalFlush:ilk */
5805         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5806
5807         g4x_disable_trickle_feed(dev);
5808
5809         ibx_init_clock_gating(dev);
5810 }
5811
5812 static void cpt_init_clock_gating(struct drm_device *dev)
5813 {
5814         struct drm_i915_private *dev_priv = dev->dev_private;
5815         int pipe;
5816         uint32_t val;
5817
5818         /*
5819          * On Ibex Peak and Cougar Point, we need to disable clock
5820          * gating for the panel power sequencer or it will fail to
5821          * start up when no ports are active.
5822          */
5823         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
5824                    PCH_DPLUNIT_CLOCK_GATE_DISABLE |
5825                    PCH_CPUNIT_CLOCK_GATE_DISABLE);
5826         I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
5827                    DPLS_EDP_PPS_FIX_DIS);
5828         /* The below fixes the weird display corruption, a few pixels shifted
5829          * downward, on (only) LVDS of some HP laptops with IVY.
5830          */
5831         for_each_pipe(dev_priv, pipe) {
5832                 val = I915_READ(TRANS_CHICKEN2(pipe));
5833                 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
5834                 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
5835                 if (dev_priv->vbt.fdi_rx_polarity_inverted)
5836                         val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
5837                 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
5838                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
5839                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
5840                 I915_WRITE(TRANS_CHICKEN2(pipe), val);
5841         }
5842         /* WADP0ClockGatingDisable */
5843         for_each_pipe(dev_priv, pipe) {
5844                 I915_WRITE(TRANS_CHICKEN1(pipe),
5845                            TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
5846         }
5847 }
5848
5849 static void gen6_check_mch_setup(struct drm_device *dev)
5850 {
5851         struct drm_i915_private *dev_priv = dev->dev_private;
5852         uint32_t tmp;
5853
5854         tmp = I915_READ(MCH_SSKPD);
5855         if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL)
5856                 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
5857                               tmp);
5858 }
5859
5860 static void gen6_init_clock_gating(struct drm_device *dev)
5861 {
5862         struct drm_i915_private *dev_priv = dev->dev_private;
5863         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
5864
5865         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
5866
5867         I915_WRITE(ILK_DISPLAY_CHICKEN2,
5868                    I915_READ(ILK_DISPLAY_CHICKEN2) |
5869                    ILK_ELPIN_409_SELECT);
5870
5871         /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
5872         I915_WRITE(_3D_CHICKEN,
5873                    _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
5874
5875         /* WaDisable_RenderCache_OperationalFlush:snb */
5876         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
5877
5878         /*
5879          * BSpec recoomends 8x4 when MSAA is used,
5880          * however in practice 16x4 seems fastest.
5881          *
5882          * Note that PS/WM thread counts depend on the WIZ hashing
5883          * disable bit, which we don't touch here, but it's good
5884          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
5885          */
5886         I915_WRITE(GEN6_GT_MODE,
5887                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
5888
5889         ilk_init_lp_watermarks(dev);
5890
5891         I915_WRITE(CACHE_MODE_0,
5892                    _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
5893
5894         I915_WRITE(GEN6_UCGCTL1,
5895                    I915_READ(GEN6_UCGCTL1) |
5896                    GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
5897                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
5898
5899         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
5900          * gating disable must be set.  Failure to set it results in
5901          * flickering pixels due to Z write ordering failures after
5902          * some amount of runtime in the Mesa "fire" demo, and Unigine
5903          * Sanctuary and Tropics, and apparently anything else with
5904          * alpha test or pixel discard.
5905          *
5906          * According to the spec, bit 11 (RCCUNIT) must also be set,
5907          * but we didn't debug actual testcases to find it out.
5908          *
5909          * WaDisableRCCUnitClockGating:snb
5910          * WaDisableRCPBUnitClockGating:snb
5911          */
5912         I915_WRITE(GEN6_UCGCTL2,
5913                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
5914                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
5915
5916         /* WaStripsFansDisableFastClipPerformanceFix:snb */
5917         I915_WRITE(_3D_CHICKEN3,
5918                    _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
5919
5920         /*
5921          * Bspec says:
5922          * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
5923          * 3DSTATE_SF number of SF output attributes is more than 16."
5924          */
5925         I915_WRITE(_3D_CHICKEN3,
5926                    _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
5927
5928         /*
5929          * According to the spec the following bits should be
5930          * set in order to enable memory self-refresh and fbc:
5931          * The bit21 and bit22 of 0x42000
5932          * The bit21 and bit22 of 0x42004
5933          * The bit5 and bit7 of 0x42020
5934          * The bit14 of 0x70180
5935          * The bit14 of 0x71180
5936          *
5937          * WaFbcAsynchFlipDisableFbcQueue:snb
5938          */
5939         I915_WRITE(ILK_DISPLAY_CHICKEN1,
5940                    I915_READ(ILK_DISPLAY_CHICKEN1) |
5941                    ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
5942         I915_WRITE(ILK_DISPLAY_CHICKEN2,
5943                    I915_READ(ILK_DISPLAY_CHICKEN2) |
5944                    ILK_DPARB_GATE | ILK_VSDPFD_FULL);
5945         I915_WRITE(ILK_DSPCLK_GATE_D,
5946                    I915_READ(ILK_DSPCLK_GATE_D) |
5947                    ILK_DPARBUNIT_CLOCK_GATE_ENABLE  |
5948                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
5949
5950         g4x_disable_trickle_feed(dev);
5951
5952         cpt_init_clock_gating(dev);
5953
5954         gen6_check_mch_setup(dev);
5955 }
5956
5957 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
5958 {
5959         uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
5960
5961         /*
5962          * WaVSThreadDispatchOverride:ivb,vlv
5963          *
5964          * This actually overrides the dispatch
5965          * mode for all thread types.
5966          */
5967         reg &= ~GEN7_FF_SCHED_MASK;
5968         reg |= GEN7_FF_TS_SCHED_HW;
5969         reg |= GEN7_FF_VS_SCHED_HW;
5970         reg |= GEN7_FF_DS_SCHED_HW;
5971
5972         I915_WRITE(GEN7_FF_THREAD_MODE, reg);
5973 }
5974
5975 static void lpt_init_clock_gating(struct drm_device *dev)
5976 {
5977         struct drm_i915_private *dev_priv = dev->dev_private;
5978
5979         /*
5980          * TODO: this bit should only be enabled when really needed, then
5981          * disabled when not needed anymore in order to save power.
5982          */
5983         if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE)
5984                 I915_WRITE(SOUTH_DSPCLK_GATE_D,
5985                            I915_READ(SOUTH_DSPCLK_GATE_D) |
5986                            PCH_LP_PARTITION_LEVEL_DISABLE);
5987
5988         /* WADPOClockGatingDisable:hsw */
5989         I915_WRITE(_TRANSA_CHICKEN1,
5990                    I915_READ(_TRANSA_CHICKEN1) |
5991                    TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
5992 }
5993
5994 static void lpt_suspend_hw(struct drm_device *dev)
5995 {
5996         struct drm_i915_private *dev_priv = dev->dev_private;
5997
5998         if (dev_priv->pch_id == INTEL_PCH_LPT_LP_DEVICE_ID_TYPE) {
5999                 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
6000
6001                 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
6002                 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
6003         }
6004 }
6005
6006 static void broadwell_init_clock_gating(struct drm_device *dev)
6007 {
6008         struct drm_i915_private *dev_priv = dev->dev_private;
6009         enum pipe pipe;
6010
6011         I915_WRITE(WM3_LP_ILK, 0);
6012         I915_WRITE(WM2_LP_ILK, 0);
6013         I915_WRITE(WM1_LP_ILK, 0);
6014
6015         /* WaSwitchSolVfFArbitrationPriority:bdw */
6016         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
6017
6018         /* WaPsrDPAMaskVBlankInSRD:bdw */
6019         I915_WRITE(CHICKEN_PAR1_1,
6020                    I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
6021
6022         /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
6023         for_each_pipe(dev_priv, pipe) {
6024                 I915_WRITE(CHICKEN_PIPESL_1(pipe),
6025                            I915_READ(CHICKEN_PIPESL_1(pipe)) |
6026                            BDW_DPRS_MASK_VBLANK_SRD);
6027         }
6028
6029         /* WaVSRefCountFullforceMissDisable:bdw */
6030         /* WaDSRefCountFullforceMissDisable:bdw */
6031         I915_WRITE(GEN7_FF_THREAD_MODE,
6032                    I915_READ(GEN7_FF_THREAD_MODE) &
6033                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
6034
6035         I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
6036                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
6037
6038         /* WaDisableSDEUnitClockGating:bdw */
6039         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
6040                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
6041
6042         lpt_init_clock_gating(dev);
6043 }
6044
6045 static void haswell_init_clock_gating(struct drm_device *dev)
6046 {
6047         struct drm_i915_private *dev_priv = dev->dev_private;
6048
6049         ilk_init_lp_watermarks(dev);
6050
6051         /* L3 caching of data atomics doesn't work -- disable it. */
6052         I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
6053         I915_WRITE(HSW_ROW_CHICKEN3,
6054                    _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
6055
6056         /* This is required by WaCatErrorRejectionIssue:hsw */
6057         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
6058                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
6059                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
6060
6061         /* WaVSRefCountFullforceMissDisable:hsw */
6062         I915_WRITE(GEN7_FF_THREAD_MODE,
6063                    I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
6064
6065         /* WaDisable_RenderCache_OperationalFlush:hsw */
6066         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6067
6068         /* enable HiZ Raw Stall Optimization */
6069         I915_WRITE(CACHE_MODE_0_GEN7,
6070                    _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
6071
6072         /* WaDisable4x2SubspanOptimization:hsw */
6073         I915_WRITE(CACHE_MODE_1,
6074                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
6075
6076         /*
6077          * BSpec recommends 8x4 when MSAA is used,
6078          * however in practice 16x4 seems fastest.
6079          *
6080          * Note that PS/WM thread counts depend on the WIZ hashing
6081          * disable bit, which we don't touch here, but it's good
6082          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6083          */
6084         I915_WRITE(GEN7_GT_MODE,
6085                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6086
6087         /* WaSampleCChickenBitEnable:hsw */
6088         I915_WRITE(HALF_SLICE_CHICKEN3,
6089                    _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE));
6090
6091         /* WaSwitchSolVfFArbitrationPriority:hsw */
6092         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
6093
6094         /* WaRsPkgCStateDisplayPMReq:hsw */
6095         I915_WRITE(CHICKEN_PAR1_1,
6096                    I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
6097
6098         lpt_init_clock_gating(dev);
6099 }
6100
6101 static void ivybridge_init_clock_gating(struct drm_device *dev)
6102 {
6103         struct drm_i915_private *dev_priv = dev->dev_private;
6104         uint32_t snpcr;
6105
6106         ilk_init_lp_watermarks(dev);
6107
6108         I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
6109
6110         /* WaDisableEarlyCull:ivb */
6111         I915_WRITE(_3D_CHICKEN3,
6112                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
6113
6114         /* WaDisableBackToBackFlipFix:ivb */
6115         I915_WRITE(IVB_CHICKEN3,
6116                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
6117                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
6118
6119         /* WaDisablePSDDualDispatchEnable:ivb */
6120         if (IS_IVB_GT1(dev))
6121                 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
6122                            _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
6123
6124         /* WaDisable_RenderCache_OperationalFlush:ivb */
6125         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6126
6127         /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
6128         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
6129                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
6130
6131         /* WaApplyL3ControlAndL3ChickenMode:ivb */
6132         I915_WRITE(GEN7_L3CNTLREG1,
6133                         GEN7_WA_FOR_GEN7_L3_CONTROL);
6134         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
6135                    GEN7_WA_L3_CHICKEN_MODE);
6136         if (IS_IVB_GT1(dev))
6137                 I915_WRITE(GEN7_ROW_CHICKEN2,
6138                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6139         else {
6140                 /* must write both registers */
6141                 I915_WRITE(GEN7_ROW_CHICKEN2,
6142                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6143                 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
6144                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6145         }
6146
6147         /* WaForceL3Serialization:ivb */
6148         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
6149                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
6150
6151         /*
6152          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6153          * This implements the WaDisableRCZUnitClockGating:ivb workaround.
6154          */
6155         I915_WRITE(GEN6_UCGCTL2,
6156                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
6157
6158         /* This is required by WaCatErrorRejectionIssue:ivb */
6159         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
6160                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
6161                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
6162
6163         g4x_disable_trickle_feed(dev);
6164
6165         gen7_setup_fixed_func_scheduler(dev_priv);
6166
6167         if (0) { /* causes HiZ corruption on ivb:gt1 */
6168                 /* enable HiZ Raw Stall Optimization */
6169                 I915_WRITE(CACHE_MODE_0_GEN7,
6170                            _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
6171         }
6172
6173         /* WaDisable4x2SubspanOptimization:ivb */
6174         I915_WRITE(CACHE_MODE_1,
6175                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
6176
6177         /*
6178          * BSpec recommends 8x4 when MSAA is used,
6179          * however in practice 16x4 seems fastest.
6180          *
6181          * Note that PS/WM thread counts depend on the WIZ hashing
6182          * disable bit, which we don't touch here, but it's good
6183          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6184          */
6185         I915_WRITE(GEN7_GT_MODE,
6186                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6187
6188         snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
6189         snpcr &= ~GEN6_MBC_SNPCR_MASK;
6190         snpcr |= GEN6_MBC_SNPCR_MED;
6191         I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
6192
6193         if (!HAS_PCH_NOP(dev))
6194                 cpt_init_clock_gating(dev);
6195
6196         gen6_check_mch_setup(dev);
6197 }
6198
6199 static void valleyview_init_clock_gating(struct drm_device *dev)
6200 {
6201         struct drm_i915_private *dev_priv = dev->dev_private;
6202
6203         I915_WRITE(DSPCLK_GATE_D, VRHUNIT_CLOCK_GATE_DISABLE);
6204
6205         /* WaDisableEarlyCull:vlv */
6206         I915_WRITE(_3D_CHICKEN3,
6207                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
6208
6209         /* WaDisableBackToBackFlipFix:vlv */
6210         I915_WRITE(IVB_CHICKEN3,
6211                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
6212                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
6213
6214         /* WaPsdDispatchEnable:vlv */
6215         /* WaDisablePSDDualDispatchEnable:vlv */
6216         I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
6217                    _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
6218                                       GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
6219
6220         /* WaDisable_RenderCache_OperationalFlush:vlv */
6221         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6222
6223         /* WaForceL3Serialization:vlv */
6224         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
6225                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
6226
6227         /* WaDisableDopClockGating:vlv */
6228         I915_WRITE(GEN7_ROW_CHICKEN2,
6229                    _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6230
6231         /* This is required by WaCatErrorRejectionIssue:vlv */
6232         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
6233                    I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
6234                    GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
6235
6236         gen7_setup_fixed_func_scheduler(dev_priv);
6237
6238         /*
6239          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6240          * This implements the WaDisableRCZUnitClockGating:vlv workaround.
6241          */
6242         I915_WRITE(GEN6_UCGCTL2,
6243                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
6244
6245         /* WaDisableL3Bank2xClockGate:vlv
6246          * Disabling L3 clock gating- MMIO 940c[25] = 1
6247          * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
6248         I915_WRITE(GEN7_UCGCTL4,
6249                    I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
6250
6251         I915_WRITE(MI_ARB_VLV, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE);
6252
6253         /*
6254          * BSpec says this must be set, even though
6255          * WaDisable4x2SubspanOptimization isn't listed for VLV.
6256          */
6257         I915_WRITE(CACHE_MODE_1,
6258                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
6259
6260         /*
6261          * BSpec recommends 8x4 when MSAA is used,
6262          * however in practice 16x4 seems fastest.
6263          *
6264          * Note that PS/WM thread counts depend on the WIZ hashing
6265          * disable bit, which we don't touch here, but it's good
6266          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6267          */
6268         I915_WRITE(GEN7_GT_MODE,
6269                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6270
6271         /*
6272          * WaIncreaseL3CreditsForVLVB0:vlv
6273          * This is the hardware default actually.
6274          */
6275         I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
6276
6277         /*
6278          * WaDisableVLVClockGating_VBIIssue:vlv
6279          * Disable clock gating on th GCFG unit to prevent a delay
6280          * in the reporting of vblank events.
6281          */
6282         I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
6283 }
6284
6285 static void cherryview_init_clock_gating(struct drm_device *dev)
6286 {
6287         struct drm_i915_private *dev_priv = dev->dev_private;
6288
6289         I915_WRITE(DSPCLK_GATE_D, VRHUNIT_CLOCK_GATE_DISABLE);
6290
6291         I915_WRITE(MI_ARB_VLV, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE);
6292
6293         /* WaVSRefCountFullforceMissDisable:chv */
6294         /* WaDSRefCountFullforceMissDisable:chv */
6295         I915_WRITE(GEN7_FF_THREAD_MODE,
6296                    I915_READ(GEN7_FF_THREAD_MODE) &
6297                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
6298
6299         /* WaDisableSemaphoreAndSyncFlipWait:chv */
6300         I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
6301                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
6302
6303         /* WaDisableCSUnitClockGating:chv */
6304         I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
6305                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
6306
6307         /* WaDisableSDEUnitClockGating:chv */
6308         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
6309                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
6310 }
6311
6312 static void g4x_init_clock_gating(struct drm_device *dev)
6313 {
6314         struct drm_i915_private *dev_priv = dev->dev_private;
6315         uint32_t dspclk_gate;
6316
6317         I915_WRITE(RENCLK_GATE_D1, 0);
6318         I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
6319                    GS_UNIT_CLOCK_GATE_DISABLE |
6320                    CL_UNIT_CLOCK_GATE_DISABLE);
6321         I915_WRITE(RAMCLK_GATE_D, 0);
6322         dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
6323                 OVRUNIT_CLOCK_GATE_DISABLE |
6324                 OVCUNIT_CLOCK_GATE_DISABLE;
6325         if (IS_GM45(dev))
6326                 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
6327         I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
6328
6329         /* WaDisableRenderCachePipelinedFlush */
6330         I915_WRITE(CACHE_MODE_0,
6331                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
6332
6333         /* WaDisable_RenderCache_OperationalFlush:g4x */
6334         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6335
6336         g4x_disable_trickle_feed(dev);
6337 }
6338
6339 static void crestline_init_clock_gating(struct drm_device *dev)
6340 {
6341         struct drm_i915_private *dev_priv = dev->dev_private;
6342
6343         I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
6344         I915_WRITE(RENCLK_GATE_D2, 0);
6345         I915_WRITE(DSPCLK_GATE_D, 0);
6346         I915_WRITE(RAMCLK_GATE_D, 0);
6347         I915_WRITE16(DEUC, 0);
6348         I915_WRITE(MI_ARB_STATE,
6349                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
6350
6351         /* WaDisable_RenderCache_OperationalFlush:gen4 */
6352         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6353 }
6354
6355 static void broadwater_init_clock_gating(struct drm_device *dev)
6356 {
6357         struct drm_i915_private *dev_priv = dev->dev_private;
6358
6359         I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
6360                    I965_RCC_CLOCK_GATE_DISABLE |
6361                    I965_RCPB_CLOCK_GATE_DISABLE |
6362                    I965_ISC_CLOCK_GATE_DISABLE |
6363                    I965_FBC_CLOCK_GATE_DISABLE);
6364         I915_WRITE(RENCLK_GATE_D2, 0);
6365         I915_WRITE(MI_ARB_STATE,
6366                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
6367
6368         /* WaDisable_RenderCache_OperationalFlush:gen4 */
6369         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6370 }
6371
6372 static void gen3_init_clock_gating(struct drm_device *dev)
6373 {
6374         struct drm_i915_private *dev_priv = dev->dev_private;
6375         u32 dstate = I915_READ(D_STATE);
6376
6377         dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
6378                 DSTATE_DOT_CLOCK_GATING;
6379         I915_WRITE(D_STATE, dstate);
6380
6381         if (IS_PINEVIEW(dev))
6382                 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
6383
6384         /* IIR "flip pending" means done if this bit is set */
6385         I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
6386
6387         /* interrupts should cause a wake up from C3 */
6388         I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
6389
6390         /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
6391         I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
6392
6393         I915_WRITE(MI_ARB_STATE,
6394                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
6395 }
6396
6397 static void i85x_init_clock_gating(struct drm_device *dev)
6398 {
6399         struct drm_i915_private *dev_priv = dev->dev_private;
6400
6401         I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
6402
6403         /* interrupts should cause a wake up from C3 */
6404         I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
6405                    _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
6406
6407         I915_WRITE(MEM_MODE,
6408                    _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE));
6409 }
6410
6411 static void i830_init_clock_gating(struct drm_device *dev)
6412 {
6413         struct drm_i915_private *dev_priv = dev->dev_private;
6414
6415         I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
6416
6417         I915_WRITE(MEM_MODE,
6418                    _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE) |
6419                    _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE));
6420 }
6421
6422 void intel_init_clock_gating(struct drm_device *dev)
6423 {
6424         struct drm_i915_private *dev_priv = dev->dev_private;
6425
6426         if (dev_priv->display.init_clock_gating)
6427                 dev_priv->display.init_clock_gating(dev);
6428 }
6429
6430 void intel_suspend_hw(struct drm_device *dev)
6431 {
6432         if (HAS_PCH_LPT(dev))
6433                 lpt_suspend_hw(dev);
6434 }
6435
6436 /* Set up chip specific power management-related functions */
6437 void intel_init_pm(struct drm_device *dev)
6438 {
6439         struct drm_i915_private *dev_priv = dev->dev_private;
6440
6441         intel_fbc_init(dev_priv);
6442
6443         /* For cxsr */
6444         if (IS_PINEVIEW(dev))
6445                 i915_pineview_get_mem_freq(dev);
6446         else if (IS_GEN5(dev))
6447                 i915_ironlake_get_mem_freq(dev);
6448
6449         /* For FIFO watermark updates */
6450         if (INTEL_INFO(dev)->gen >= 9) {
6451                 skl_setup_wm_latency(dev);
6452
6453                 dev_priv->display.init_clock_gating = skl_init_clock_gating;
6454                 dev_priv->display.update_wm = skl_update_wm;
6455                 dev_priv->display.update_sprite_wm = skl_update_sprite_wm;
6456         } else if (HAS_PCH_SPLIT(dev)) {
6457                 ilk_setup_wm_latency(dev);
6458
6459                 if ((IS_GEN5(dev) && dev_priv->wm.pri_latency[1] &&
6460                      dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
6461                     (!IS_GEN5(dev) && dev_priv->wm.pri_latency[0] &&
6462                      dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
6463                         dev_priv->display.update_wm = ilk_update_wm;
6464                         dev_priv->display.update_sprite_wm = ilk_update_sprite_wm;
6465                 } else {
6466                         DRM_DEBUG_KMS("Failed to read display plane latency. "
6467                                       "Disable CxSR\n");
6468                 }
6469
6470                 if (IS_GEN5(dev))
6471                         dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
6472                 else if (IS_GEN6(dev))
6473                         dev_priv->display.init_clock_gating = gen6_init_clock_gating;
6474                 else if (IS_IVYBRIDGE(dev))
6475                         dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
6476                 else if (IS_HASWELL(dev))
6477                         dev_priv->display.init_clock_gating = haswell_init_clock_gating;
6478                 else if (INTEL_INFO(dev)->gen == 8)
6479                         dev_priv->display.init_clock_gating = broadwell_init_clock_gating;
6480         } else if (IS_CHERRYVIEW(dev)) {
6481                 dev_priv->display.update_wm = cherryview_update_wm;
6482                 dev_priv->display.update_sprite_wm = valleyview_update_sprite_wm;
6483                 dev_priv->display.init_clock_gating =
6484                         cherryview_init_clock_gating;
6485         } else if (IS_VALLEYVIEW(dev)) {
6486                 dev_priv->display.update_wm = valleyview_update_wm;
6487                 dev_priv->display.update_sprite_wm = valleyview_update_sprite_wm;
6488                 dev_priv->display.init_clock_gating =
6489                         valleyview_init_clock_gating;
6490         } else if (IS_PINEVIEW(dev)) {
6491                 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
6492                                             dev_priv->is_ddr3,
6493                                             dev_priv->fsb_freq,
6494                                             dev_priv->mem_freq)) {
6495                         DRM_INFO("failed to find known CxSR latency "
6496                                  "(found ddr%s fsb freq %d, mem freq %d), "
6497                                  "disabling CxSR\n",
6498                                  (dev_priv->is_ddr3 == 1) ? "3" : "2",
6499                                  dev_priv->fsb_freq, dev_priv->mem_freq);
6500                         /* Disable CxSR and never update its watermark again */
6501                         intel_set_memory_cxsr(dev_priv, false);
6502                         dev_priv->display.update_wm = NULL;
6503                 } else
6504                         dev_priv->display.update_wm = pineview_update_wm;
6505                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
6506         } else if (IS_G4X(dev)) {
6507                 dev_priv->display.update_wm = g4x_update_wm;
6508                 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
6509         } else if (IS_GEN4(dev)) {
6510                 dev_priv->display.update_wm = i965_update_wm;
6511                 if (IS_CRESTLINE(dev))
6512                         dev_priv->display.init_clock_gating = crestline_init_clock_gating;
6513                 else if (IS_BROADWATER(dev))
6514                         dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
6515         } else if (IS_GEN3(dev)) {
6516                 dev_priv->display.update_wm = i9xx_update_wm;
6517                 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
6518                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
6519         } else if (IS_GEN2(dev)) {
6520                 if (INTEL_INFO(dev)->num_pipes == 1) {
6521                         dev_priv->display.update_wm = i845_update_wm;
6522                         dev_priv->display.get_fifo_size = i845_get_fifo_size;
6523                 } else {
6524                         dev_priv->display.update_wm = i9xx_update_wm;
6525                         dev_priv->display.get_fifo_size = i830_get_fifo_size;
6526                 }
6527
6528                 if (IS_I85X(dev) || IS_I865G(dev))
6529                         dev_priv->display.init_clock_gating = i85x_init_clock_gating;
6530                 else
6531                         dev_priv->display.init_clock_gating = i830_init_clock_gating;
6532         } else {
6533                 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
6534         }
6535 }
6536
6537 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u32 mbox, u32 *val)
6538 {
6539         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6540
6541         if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
6542                 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
6543                 return -EAGAIN;
6544         }
6545
6546         I915_WRITE(GEN6_PCODE_DATA, *val);
6547         I915_WRITE(GEN6_PCODE_DATA1, 0);
6548         I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
6549
6550         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
6551                      500)) {
6552                 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
6553                 return -ETIMEDOUT;
6554         }
6555
6556         *val = I915_READ(GEN6_PCODE_DATA);
6557         I915_WRITE(GEN6_PCODE_DATA, 0);
6558
6559         return 0;
6560 }
6561
6562 int sandybridge_pcode_write(struct drm_i915_private *dev_priv, u32 mbox, u32 val)
6563 {
6564         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6565
6566         if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
6567                 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
6568                 return -EAGAIN;
6569         }
6570
6571         I915_WRITE(GEN6_PCODE_DATA, val);
6572         I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
6573
6574         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
6575                      500)) {
6576                 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
6577                 return -ETIMEDOUT;
6578         }
6579
6580         I915_WRITE(GEN6_PCODE_DATA, 0);
6581
6582         return 0;
6583 }
6584
6585 static int vlv_gpu_freq_div(unsigned int czclk_freq)
6586 {
6587         switch (czclk_freq) {
6588         case 200:
6589                 return 10;
6590         case 267:
6591                 return 12;
6592         case 320:
6593         case 333:
6594                 return 16;
6595         case 400:
6596                 return 20;
6597         default:
6598                 return -1;
6599         }
6600 }
6601
6602 static int byt_gpu_freq(struct drm_i915_private *dev_priv, int val)
6603 {
6604         int div, czclk_freq = DIV_ROUND_CLOSEST(dev_priv->mem_freq, 4);
6605
6606         div = vlv_gpu_freq_div(czclk_freq);
6607         if (div < 0)
6608                 return div;
6609
6610         return DIV_ROUND_CLOSEST(czclk_freq * (val + 6 - 0xbd), div);
6611 }
6612
6613 static int byt_freq_opcode(struct drm_i915_private *dev_priv, int val)
6614 {
6615         int mul, czclk_freq = DIV_ROUND_CLOSEST(dev_priv->mem_freq, 4);
6616
6617         mul = vlv_gpu_freq_div(czclk_freq);
6618         if (mul < 0)
6619                 return mul;
6620
6621         return DIV_ROUND_CLOSEST(mul * val, czclk_freq) + 0xbd - 6;
6622 }
6623
6624 static int chv_gpu_freq(struct drm_i915_private *dev_priv, int val)
6625 {
6626         int div, czclk_freq = dev_priv->rps.cz_freq;
6627
6628         div = vlv_gpu_freq_div(czclk_freq) / 2;
6629         if (div < 0)
6630                 return div;
6631
6632         return DIV_ROUND_CLOSEST(czclk_freq * val, 2 * div) / 2;
6633 }
6634
6635 static int chv_freq_opcode(struct drm_i915_private *dev_priv, int val)
6636 {
6637         int mul, czclk_freq = dev_priv->rps.cz_freq;
6638
6639         mul = vlv_gpu_freq_div(czclk_freq) / 2;
6640         if (mul < 0)
6641                 return mul;
6642
6643         /* CHV needs even values */
6644         return DIV_ROUND_CLOSEST(val * 2 * mul, czclk_freq) * 2;
6645 }
6646
6647 int intel_gpu_freq(struct drm_i915_private *dev_priv, int val)
6648 {
6649         if (IS_CHERRYVIEW(dev_priv->dev))
6650                 return chv_gpu_freq(dev_priv, val);
6651         else if (IS_VALLEYVIEW(dev_priv->dev))
6652                 return byt_gpu_freq(dev_priv, val);
6653         else
6654                 return val * GT_FREQUENCY_MULTIPLIER;
6655 }
6656
6657 int intel_freq_opcode(struct drm_i915_private *dev_priv, int val)
6658 {
6659         if (IS_CHERRYVIEW(dev_priv->dev))
6660                 return chv_freq_opcode(dev_priv, val);
6661         else if (IS_VALLEYVIEW(dev_priv->dev))
6662                 return byt_freq_opcode(dev_priv, val);
6663         else
6664                 return val / GT_FREQUENCY_MULTIPLIER;
6665 }
6666
6667 void intel_pm_setup(struct drm_device *dev)
6668 {
6669         struct drm_i915_private *dev_priv = dev->dev_private;
6670
6671         mutex_init(&dev_priv->rps.hw_lock);
6672
6673         INIT_DELAYED_WORK(&dev_priv->rps.delayed_resume_work,
6674                           intel_gen6_powersave_work);
6675
6676         dev_priv->pm.suspended = false;
6677 }