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 <drm/drm_plane_helper.h>
  30 #include "i915_drv.h"
  31 #include "intel_drv.h"
  32 #include "../../../platform/x86/intel_ips.h"
  33 #include <linux/module.h>
  34
  35 /**
  36  * DOC: RC6
  37  *
  38  * RC6 is a special power stage which allows the GPU to enter an very
  39  * low-voltage mode when idle, using down to 0V while at this stage.  This
  40  * stage is entered automatically when the GPU is idle when RC6 support is
  41  * enabled, and as soon as new workload arises GPU wakes up automatically as well.
  42  *
  43  * There are different RC6 modes available in Intel GPU, which differentiate
  44  * among each other with the latency required to enter and leave RC6 and
  45  * voltage consumed by the GPU in different states.
  46  *
  47  * The combination of the following flags define which states GPU is allowed
  48  * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
  49  * RC6pp is deepest RC6. Their support by hardware varies according to the
  50  * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
  51  * which brings the most power savings; deeper states save more power, but
  52  * require higher latency to switch to and wake up.
  53  */
  54 #define INTEL_RC6_ENABLE                        (1<<0)
  55 #define INTEL_RC6p_ENABLE                       (1<<1)
  56 #define INTEL_RC6pp_ENABLE                      (1<<2)
  57
  58 static void gen9_init_clock_gating(struct drm_device *dev)
  59 {
  60         struct drm_i915_private *dev_priv = dev->dev_private;
  61
  62         /* See Bspec note for PSR2_CTL bit 31, Wa#828:skl,bxt,kbl */
  63         I915_WRITE(CHICKEN_PAR1_1,
  64                    I915_READ(CHICKEN_PAR1_1) | SKL_EDP_PSR_FIX_RDWRAP);
  65
  66         I915_WRITE(GEN8_CONFIG0,
  67                    I915_READ(GEN8_CONFIG0) | GEN9_DEFAULT_FIXES);
  68
  69         /* WaEnableChickenDCPR:skl,bxt,kbl */
  70         I915_WRITE(GEN8_CHICKEN_DCPR_1,
  71                    I915_READ(GEN8_CHICKEN_DCPR_1) | MASK_WAKEMEM);
  72
  73         /* WaFbcTurnOffFbcWatermark:skl,bxt,kbl */
  74         /* WaFbcWakeMemOn:skl,bxt,kbl */
  75         I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
  76                    DISP_FBC_WM_DIS |
  77                    DISP_FBC_MEMORY_WAKE);
  78
  79         /* WaFbcHighMemBwCorruptionAvoidance:skl,bxt,kbl */
  80         I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
  81                    ILK_DPFC_DISABLE_DUMMY0);
  82 }
  83
  84 static void bxt_init_clock_gating(struct drm_device *dev)
  85 {
  86         struct drm_i915_private *dev_priv = to_i915(dev);
  87
  88         gen9_init_clock_gating(dev);
  89
  90         /* WaDisableSDEUnitClockGating:bxt */
  91         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
  92                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
  93
  94         /*
  95          * FIXME:
  96          * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
  97          */
  98         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
  99                    GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ);
 100
 101         /*
 102          * Wa: Backlight PWM may stop in the asserted state, causing backlight
 103          * to stay fully on.
 104          */
 105         if (IS_BXT_REVID(dev_priv, BXT_REVID_B0, REVID_FOREVER))
 106                 I915_WRITE(GEN9_CLKGATE_DIS_0, I915_READ(GEN9_CLKGATE_DIS_0) |
 107                            PWM1_GATING_DIS | PWM2_GATING_DIS);
 108 }
 109
 110 static void i915_pineview_get_mem_freq(struct drm_device *dev)
 111 {
 112         struct drm_i915_private *dev_priv = to_i915(dev);
 113         u32 tmp;
 114
 115         tmp = I915_READ(CLKCFG);
 116
 117         switch (tmp & CLKCFG_FSB_MASK) {
 118         case CLKCFG_FSB_533:
 119                 dev_priv->fsb_freq = 533; /* 133*4 */
 120                 break;
 121         case CLKCFG_FSB_800:
 122                 dev_priv->fsb_freq = 800; /* 200*4 */
 123                 break;
 124         case CLKCFG_FSB_667:
 125                 dev_priv->fsb_freq =  667; /* 167*4 */
 126                 break;
 127         case CLKCFG_FSB_400:
 128                 dev_priv->fsb_freq = 400; /* 100*4 */
 129                 break;
 130         }
 131
 132         switch (tmp & CLKCFG_MEM_MASK) {
 133         case CLKCFG_MEM_533:
 134                 dev_priv->mem_freq = 533;
 135                 break;
 136         case CLKCFG_MEM_667:
 137                 dev_priv->mem_freq = 667;
 138                 break;
 139         case CLKCFG_MEM_800:
 140                 dev_priv->mem_freq = 800;
 141                 break;
 142         }
 143
 144         /* detect pineview DDR3 setting */
 145         tmp = I915_READ(CSHRDDR3CTL);
 146         dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
 147 }
 148
 149 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
 150 {
 151         struct drm_i915_private *dev_priv = to_i915(dev);
 152         u16 ddrpll, csipll;
 153
 154         ddrpll = I915_READ16(DDRMPLL1);
 155         csipll = I915_READ16(CSIPLL0);
 156
 157         switch (ddrpll & 0xff) {
 158         case 0xc:
 159                 dev_priv->mem_freq = 800;
 160                 break;
 161         case 0x10:
 162                 dev_priv->mem_freq = 1066;
 163                 break;
 164         case 0x14:
 165                 dev_priv->mem_freq = 1333;
 166                 break;
 167         case 0x18:
 168                 dev_priv->mem_freq = 1600;
 169                 break;
 170         default:
 171                 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
 172                                  ddrpll & 0xff);
 173                 dev_priv->mem_freq = 0;
 174                 break;
 175         }
 176
 177         dev_priv->ips.r_t = dev_priv->mem_freq;
 178
 179         switch (csipll & 0x3ff) {
 180         case 0x00c:
 181                 dev_priv->fsb_freq = 3200;
 182                 break;
 183         case 0x00e:
 184                 dev_priv->fsb_freq = 3733;
 185                 break;
 186         case 0x010:
 187                 dev_priv->fsb_freq = 4266;
 188                 break;
 189         case 0x012:
 190                 dev_priv->fsb_freq = 4800;
 191                 break;
 192         case 0x014:
 193                 dev_priv->fsb_freq = 5333;
 194                 break;
 195         case 0x016:
 196                 dev_priv->fsb_freq = 5866;
 197                 break;
 198         case 0x018:
 199                 dev_priv->fsb_freq = 6400;
 200                 break;
 201         default:
 202                 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
 203                                  csipll & 0x3ff);
 204                 dev_priv->fsb_freq = 0;
 205                 break;
 206         }
 207
 208         if (dev_priv->fsb_freq == 3200) {
 209                 dev_priv->ips.c_m = 0;
 210         } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
 211                 dev_priv->ips.c_m = 1;
 212         } else {
 213                 dev_priv->ips.c_m = 2;
 214         }
 215 }
 216
 217 static const struct cxsr_latency cxsr_latency_table[] = {
 218         {1, 0, 800, 400, 3382, 33382, 3983, 33983},    /* DDR2-400 SC */
 219         {1, 0, 800, 667, 3354, 33354, 3807, 33807},    /* DDR2-667 SC */
 220         {1, 0, 800, 800, 3347, 33347, 3763, 33763},    /* DDR2-800 SC */
 221         {1, 1, 800, 667, 6420, 36420, 6873, 36873},    /* DDR3-667 SC */
 222         {1, 1, 800, 800, 5902, 35902, 6318, 36318},    /* DDR3-800 SC */
 223
 224         {1, 0, 667, 400, 3400, 33400, 4021, 34021},    /* DDR2-400 SC */
 225         {1, 0, 667, 667, 3372, 33372, 3845, 33845},    /* DDR2-667 SC */
 226         {1, 0, 667, 800, 3386, 33386, 3822, 33822},    /* DDR2-800 SC */
 227         {1, 1, 667, 667, 6438, 36438, 6911, 36911},    /* DDR3-667 SC */
 228         {1, 1, 667, 800, 5941, 35941, 6377, 36377},    /* DDR3-800 SC */
 229
 230         {1, 0, 400, 400, 3472, 33472, 4173, 34173},    /* DDR2-400 SC */
 231         {1, 0, 400, 667, 3443, 33443, 3996, 33996},    /* DDR2-667 SC */
 232         {1, 0, 400, 800, 3430, 33430, 3946, 33946},    /* DDR2-800 SC */
 233         {1, 1, 400, 667, 6509, 36509, 7062, 37062},    /* DDR3-667 SC */
 234         {1, 1, 400, 800, 5985, 35985, 6501, 36501},    /* DDR3-800 SC */
 235
 236         {0, 0, 800, 400, 3438, 33438, 4065, 34065},    /* DDR2-400 SC */
 237         {0, 0, 800, 667, 3410, 33410, 3889, 33889},    /* DDR2-667 SC */
 238         {0, 0, 800, 800, 3403, 33403, 3845, 33845},    /* DDR2-800 SC */
 239         {0, 1, 800, 667, 6476, 36476, 6955, 36955},    /* DDR3-667 SC */
 240         {0, 1, 800, 800, 5958, 35958, 6400, 36400},    /* DDR3-800 SC */
 241
 242         {0, 0, 667, 400, 3456, 33456, 4103, 34106},    /* DDR2-400 SC */
 243         {0, 0, 667, 667, 3428, 33428, 3927, 33927},    /* DDR2-667 SC */
 244         {0, 0, 667, 800, 3443, 33443, 3905, 33905},    /* DDR2-800 SC */
 245         {0, 1, 667, 667, 6494, 36494, 6993, 36993},    /* DDR3-667 SC */
 246         {0, 1, 667, 800, 5998, 35998, 6460, 36460},    /* DDR3-800 SC */
 247
 248         {0, 0, 400, 400, 3528, 33528, 4255, 34255},    /* DDR2-400 SC */
 249         {0, 0, 400, 667, 3500, 33500, 4079, 34079},    /* DDR2-667 SC */
 250         {0, 0, 400, 800, 3487, 33487, 4029, 34029},    /* DDR2-800 SC */
 251         {0, 1, 400, 667, 6566, 36566, 7145, 37145},    /* DDR3-667 SC */
 252         {0, 1, 400, 800, 6042, 36042, 6584, 36584},    /* DDR3-800 SC */
 253 };
 254
 255 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
 256                                                          int is_ddr3,
 257                                                          int fsb,
 258                                                          int mem)
 259 {
 260         const struct cxsr_latency *latency;
 261         int i;
 262
 263         if (fsb == 0 || mem == 0)
 264                 return NULL;
 265
 266         for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
 267                 latency = &cxsr_latency_table[i];
 268                 if (is_desktop == latency->is_desktop &&
 269                     is_ddr3 == latency->is_ddr3 &&
 270                     fsb == latency->fsb_freq && mem == latency->mem_freq)
 271                         return latency;
 272         }
 273
 274         DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 275
 276         return NULL;
 277 }
 278
 279 static void chv_set_memory_dvfs(struct drm_i915_private *dev_priv, bool enable)
 280 {
 281         u32 val;
 282
 283         mutex_lock(&dev_priv->rps.hw_lock);
 284
 285         val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
 286         if (enable)
 287                 val &= ~FORCE_DDR_HIGH_FREQ;
 288         else
 289                 val |= FORCE_DDR_HIGH_FREQ;
 290         val &= ~FORCE_DDR_LOW_FREQ;
 291         val |= FORCE_DDR_FREQ_REQ_ACK;
 292         vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
 293
 294         if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
 295                       FORCE_DDR_FREQ_REQ_ACK) == 0, 3))
 296                 DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
 297
 298         mutex_unlock(&dev_priv->rps.hw_lock);
 299 }
 300
 301 static void chv_set_memory_pm5(struct drm_i915_private *dev_priv, bool enable)
 302 {
 303         u32 val;
 304
 305         mutex_lock(&dev_priv->rps.hw_lock);
 306
 307         val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
 308         if (enable)
 309                 val |= DSP_MAXFIFO_PM5_ENABLE;
 310         else
 311                 val &= ~DSP_MAXFIFO_PM5_ENABLE;
 312         vlv_punit_write(dev_priv, PUNIT_REG_DSPFREQ, val);
 313
 314         mutex_unlock(&dev_priv->rps.hw_lock);
 315 }
 316
 317 #define FW_WM(value, plane) \
 318         (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
 319
 320 void intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
 321 {
 322         struct drm_device *dev = &dev_priv->drm;
 323         u32 val;
 324
 325         if (IS_VALLEYVIEW(dev) || IS_CHERRYVIEW(dev)) {
 326                 I915_WRITE(FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
 327                 POSTING_READ(FW_BLC_SELF_VLV);
 328                 dev_priv->wm.vlv.cxsr = enable;
 329         } else if (IS_G4X(dev) || IS_CRESTLINE(dev)) {
 330                 I915_WRITE(FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
 331                 POSTING_READ(FW_BLC_SELF);
 332         } else if (IS_PINEVIEW(dev)) {
 333                 val = I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN;
 334                 val |= enable ? PINEVIEW_SELF_REFRESH_EN : 0;
 335                 I915_WRITE(DSPFW3, val);
 336                 POSTING_READ(DSPFW3);
 337         } else if (IS_I945G(dev) || IS_I945GM(dev)) {
 338                 val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
 339                                _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
 340                 I915_WRITE(FW_BLC_SELF, val);
 341                 POSTING_READ(FW_BLC_SELF);
 342         } else if (IS_I915GM(dev)) {
 343                 val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
 344                                _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
 345                 I915_WRITE(INSTPM, val);
 346                 POSTING_READ(INSTPM);
 347         } else {
 348                 return;
 349         }
 350
 351         DRM_DEBUG_KMS("memory self-refresh is %s\n",
 352                       enable ? "enabled" : "disabled");
 353 }
 354
 355
 356 /*
 357  * Latency for FIFO fetches is dependent on several factors:
 358  *   - memory configuration (speed, channels)
 359  *   - chipset
 360  *   - current MCH state
 361  * It can be fairly high in some situations, so here we assume a fairly
 362  * pessimal value.  It's a tradeoff between extra memory fetches (if we
 363  * set this value too high, the FIFO will fetch frequently to stay full)
 364  * and power consumption (set it too low to save power and we might see
 365  * FIFO underruns and display "flicker").
 366  *
 367  * A value of 5us seems to be a good balance; safe for very low end
 368  * platforms but not overly aggressive on lower latency configs.
 369  */
 370 static const int pessimal_latency_ns = 5000;
 371
 372 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
 373         ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
 374
 375 static int vlv_get_fifo_size(struct drm_device *dev,
 376                               enum pipe pipe, int plane)
 377 {
 378         struct drm_i915_private *dev_priv = to_i915(dev);
 379         int sprite0_start, sprite1_start, size;
 380
 381         switch (pipe) {
 382                 uint32_t dsparb, dsparb2, dsparb3;
 383         case PIPE_A:
 384                 dsparb = I915_READ(DSPARB);
 385                 dsparb2 = I915_READ(DSPARB2);
 386                 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 0, 0);
 387                 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 8, 4);
 388                 break;
 389         case PIPE_B:
 390                 dsparb = I915_READ(DSPARB);
 391                 dsparb2 = I915_READ(DSPARB2);
 392                 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 16, 8);
 393                 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 24, 12);
 394                 break;
 395         case PIPE_C:
 396                 dsparb2 = I915_READ(DSPARB2);
 397                 dsparb3 = I915_READ(DSPARB3);
 398                 sprite0_start = VLV_FIFO_START(dsparb3, dsparb2, 0, 16);
 399                 sprite1_start = VLV_FIFO_START(dsparb3, dsparb2, 8, 20);
 400                 break;
 401         default:
 402                 return 0;
 403         }
 404
 405         switch (plane) {
 406         case 0:
 407                 size = sprite0_start;
 408                 break;
 409         case 1:
 410                 size = sprite1_start - sprite0_start;
 411                 break;
 412         case 2:
 413                 size = 512 - 1 - sprite1_start;
 414                 break;
 415         default:
 416                 return 0;
 417         }
 418
 419         DRM_DEBUG_KMS("Pipe %c %s %c FIFO size: %d\n",
 420                       pipe_name(pipe), plane == 0 ? "primary" : "sprite",
 421                       plane == 0 ? plane_name(pipe) : sprite_name(pipe, plane - 1),
 422                       size);
 423
 424         return size;
 425 }
 426
 427 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
 428 {
 429         struct drm_i915_private *dev_priv = to_i915(dev);
 430         uint32_t dsparb = I915_READ(DSPARB);
 431         int size;
 432
 433         size = dsparb & 0x7f;
 434         if (plane)
 435                 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
 436
 437         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 438                       plane ? "B" : "A", size);
 439
 440         return size;
 441 }
 442
 443 static int i830_get_fifo_size(struct drm_device *dev, int plane)
 444 {
 445         struct drm_i915_private *dev_priv = to_i915(dev);
 446         uint32_t dsparb = I915_READ(DSPARB);
 447         int size;
 448
 449         size = dsparb & 0x1ff;
 450         if (plane)
 451                 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
 452         size >>= 1; /* Convert to cachelines */
 453
 454         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 455                       plane ? "B" : "A", size);
 456
 457         return size;
 458 }
 459
 460 static int i845_get_fifo_size(struct drm_device *dev, int plane)
 461 {
 462         struct drm_i915_private *dev_priv = to_i915(dev);
 463         uint32_t dsparb = I915_READ(DSPARB);
 464         int size;
 465
 466         size = dsparb & 0x7f;
 467         size >>= 2; /* Convert to cachelines */
 468
 469         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 470                       plane ? "B" : "A",
 471                       size);
 472
 473         return size;
 474 }
 475
 476 /* Pineview has different values for various configs */
 477 static const struct intel_watermark_params pineview_display_wm = {
 478         .fifo_size = PINEVIEW_DISPLAY_FIFO,
 479         .max_wm = PINEVIEW_MAX_WM,
 480         .default_wm = PINEVIEW_DFT_WM,
 481         .guard_size = PINEVIEW_GUARD_WM,
 482         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 483 };
 484 static const struct intel_watermark_params pineview_display_hplloff_wm = {
 485         .fifo_size = PINEVIEW_DISPLAY_FIFO,
 486         .max_wm = PINEVIEW_MAX_WM,
 487         .default_wm = PINEVIEW_DFT_HPLLOFF_WM,
 488         .guard_size = PINEVIEW_GUARD_WM,
 489         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 490 };
 491 static const struct intel_watermark_params pineview_cursor_wm = {
 492         .fifo_size = PINEVIEW_CURSOR_FIFO,
 493         .max_wm = PINEVIEW_CURSOR_MAX_WM,
 494         .default_wm = PINEVIEW_CURSOR_DFT_WM,
 495         .guard_size = PINEVIEW_CURSOR_GUARD_WM,
 496         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 497 };
 498 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
 499         .fifo_size = PINEVIEW_CURSOR_FIFO,
 500         .max_wm = PINEVIEW_CURSOR_MAX_WM,
 501         .default_wm = PINEVIEW_CURSOR_DFT_WM,
 502         .guard_size = PINEVIEW_CURSOR_GUARD_WM,
 503         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 504 };
 505 static const struct intel_watermark_params g4x_wm_info = {
 506         .fifo_size = G4X_FIFO_SIZE,
 507         .max_wm = G4X_MAX_WM,
 508         .default_wm = G4X_MAX_WM,
 509         .guard_size = 2,
 510         .cacheline_size = G4X_FIFO_LINE_SIZE,
 511 };
 512 static const struct intel_watermark_params g4x_cursor_wm_info = {
 513         .fifo_size = I965_CURSOR_FIFO,
 514         .max_wm = I965_CURSOR_MAX_WM,
 515         .default_wm = I965_CURSOR_DFT_WM,
 516         .guard_size = 2,
 517         .cacheline_size = G4X_FIFO_LINE_SIZE,
 518 };
 519 static const struct intel_watermark_params i965_cursor_wm_info = {
 520         .fifo_size = I965_CURSOR_FIFO,
 521         .max_wm = I965_CURSOR_MAX_WM,
 522         .default_wm = I965_CURSOR_DFT_WM,
 523         .guard_size = 2,
 524         .cacheline_size = I915_FIFO_LINE_SIZE,
 525 };
 526 static const struct intel_watermark_params i945_wm_info = {
 527         .fifo_size = I945_FIFO_SIZE,
 528         .max_wm = I915_MAX_WM,
 529         .default_wm = 1,
 530         .guard_size = 2,
 531         .cacheline_size = I915_FIFO_LINE_SIZE,
 532 };
 533 static const struct intel_watermark_params i915_wm_info = {
 534         .fifo_size = I915_FIFO_SIZE,
 535         .max_wm = I915_MAX_WM,
 536         .default_wm = 1,
 537         .guard_size = 2,
 538         .cacheline_size = I915_FIFO_LINE_SIZE,
 539 };
 540 static const struct intel_watermark_params i830_a_wm_info = {
 541         .fifo_size = I855GM_FIFO_SIZE,
 542         .max_wm = I915_MAX_WM,
 543         .default_wm = 1,
 544         .guard_size = 2,
 545         .cacheline_size = I830_FIFO_LINE_SIZE,
 546 };
 547 static const struct intel_watermark_params i830_bc_wm_info = {
 548         .fifo_size = I855GM_FIFO_SIZE,
 549         .max_wm = I915_MAX_WM/2,
 550         .default_wm = 1,
 551         .guard_size = 2,
 552         .cacheline_size = I830_FIFO_LINE_SIZE,
 553 };
 554 static const struct intel_watermark_params i845_wm_info = {
 555         .fifo_size = I830_FIFO_SIZE,
 556         .max_wm = I915_MAX_WM,
 557         .default_wm = 1,
 558         .guard_size = 2,
 559         .cacheline_size = I830_FIFO_LINE_SIZE,
 560 };
 561
 562 /**
 563  * intel_calculate_wm - calculate watermark level
 564  * @clock_in_khz: pixel clock
 565  * @wm: chip FIFO params
 566  * @cpp: bytes per pixel
 567  * @latency_ns: memory latency for the platform
 568  *
 569  * Calculate the watermark level (the level at which the display plane will
 570  * start fetching from memory again).  Each chip has a different display
 571  * FIFO size and allocation, so the caller needs to figure that out and pass
 572  * in the correct intel_watermark_params structure.
 573  *
 574  * As the pixel clock runs, the FIFO will be drained at a rate that depends
 575  * on the pixel size.  When it reaches the watermark level, it'll start
 576  * fetching FIFO line sized based chunks from memory until the FIFO fills
 577  * past the watermark point.  If the FIFO drains completely, a FIFO underrun
 578  * will occur, and a display engine hang could result.
 579  */
 580 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
 581                                         const struct intel_watermark_params *wm,
 582                                         int fifo_size, int cpp,
 583                                         unsigned long latency_ns)
 584 {
 585         long entries_required, wm_size;
 586
 587         /*
 588          * Note: we need to make sure we don't overflow for various clock &
 589          * latency values.
 590          * clocks go from a few thousand to several hundred thousand.
 591          * latency is usually a few thousand
 592          */
 593         entries_required = ((clock_in_khz / 1000) * cpp * latency_ns) /
 594                 1000;
 595         entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
 596
 597         DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
 598
 599         wm_size = fifo_size - (entries_required + wm->guard_size);
 600
 601         DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
 602
 603         /* Don't promote wm_size to unsigned... */
 604         if (wm_size > (long)wm->max_wm)
 605                 wm_size = wm->max_wm;
 606         if (wm_size <= 0)
 607                 wm_size = wm->default_wm;
 608
 609         /*
 610          * Bspec seems to indicate that the value shouldn't be lower than
 611          * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
 612          * Lets go for 8 which is the burst size since certain platforms
 613          * already use a hardcoded 8 (which is what the spec says should be
 614          * done).
 615          */
 616         if (wm_size <= 8)
 617                 wm_size = 8;
 618
 619         return wm_size;
 620 }
 621
 622 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
 623 {
 624         struct drm_crtc *crtc, *enabled = NULL;
 625
 626         for_each_crtc(dev, crtc) {
 627                 if (intel_crtc_active(crtc)) {
 628                         if (enabled)
 629                                 return NULL;
 630                         enabled = crtc;
 631                 }
 632         }
 633
 634         return enabled;
 635 }
 636
 637 static void pineview_update_wm(struct drm_crtc *unused_crtc)
 638 {
 639         struct drm_device *dev = unused_crtc->dev;
 640         struct drm_i915_private *dev_priv = to_i915(dev);
 641         struct drm_crtc *crtc;
 642         const struct cxsr_latency *latency;
 643         u32 reg;
 644         unsigned long wm;
 645
 646         latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
 647                                          dev_priv->fsb_freq, dev_priv->mem_freq);
 648         if (!latency) {
 649                 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 650                 intel_set_memory_cxsr(dev_priv, false);
 651                 return;
 652         }
 653
 654         crtc = single_enabled_crtc(dev);
 655         if (crtc) {
 656                 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
 657                 int cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
 658                 int clock = adjusted_mode->crtc_clock;
 659
 660                 /* Display SR */
 661                 wm = intel_calculate_wm(clock, &pineview_display_wm,
 662                                         pineview_display_wm.fifo_size,
 663                                         cpp, latency->display_sr);
 664                 reg = I915_READ(DSPFW1);
 665                 reg &= ~DSPFW_SR_MASK;
 666                 reg |= FW_WM(wm, SR);
 667                 I915_WRITE(DSPFW1, reg);
 668                 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
 669
 670                 /* cursor SR */
 671                 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
 672                                         pineview_display_wm.fifo_size,
 673                                         cpp, latency->cursor_sr);
 674                 reg = I915_READ(DSPFW3);
 675                 reg &= ~DSPFW_CURSOR_SR_MASK;
 676                 reg |= FW_WM(wm, CURSOR_SR);
 677                 I915_WRITE(DSPFW3, reg);
 678
 679                 /* Display HPLL off SR */
 680                 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
 681                                         pineview_display_hplloff_wm.fifo_size,
 682                                         cpp, latency->display_hpll_disable);
 683                 reg = I915_READ(DSPFW3);
 684                 reg &= ~DSPFW_HPLL_SR_MASK;
 685                 reg |= FW_WM(wm, HPLL_SR);
 686                 I915_WRITE(DSPFW3, reg);
 687
 688                 /* cursor HPLL off SR */
 689                 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
 690                                         pineview_display_hplloff_wm.fifo_size,
 691                                         cpp, latency->cursor_hpll_disable);
 692                 reg = I915_READ(DSPFW3);
 693                 reg &= ~DSPFW_HPLL_CURSOR_MASK;
 694                 reg |= FW_WM(wm, HPLL_CURSOR);
 695                 I915_WRITE(DSPFW3, reg);
 696                 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
 697
 698                 intel_set_memory_cxsr(dev_priv, true);
 699         } else {
 700                 intel_set_memory_cxsr(dev_priv, false);
 701         }
 702 }
 703
 704 static bool g4x_compute_wm0(struct drm_device *dev,
 705                             int plane,
 706                             const struct intel_watermark_params *display,
 707                             int display_latency_ns,
 708                             const struct intel_watermark_params *cursor,
 709                             int cursor_latency_ns,
 710                             int *plane_wm,
 711                             int *cursor_wm)
 712 {
 713         struct drm_crtc *crtc;
 714         const struct drm_display_mode *adjusted_mode;
 715         int htotal, hdisplay, clock, cpp;
 716         int line_time_us, line_count;
 717         int entries, tlb_miss;
 718
 719         crtc = intel_get_crtc_for_plane(dev, plane);
 720         if (!intel_crtc_active(crtc)) {
 721                 *cursor_wm = cursor->guard_size;
 722                 *plane_wm = display->guard_size;
 723                 return false;
 724         }
 725
 726         adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
 727         clock = adjusted_mode->crtc_clock;
 728         htotal = adjusted_mode->crtc_htotal;
 729         hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
 730         cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
 731
 732         /* Use the small buffer method to calculate plane watermark */
 733         entries = ((clock * cpp / 1000) * display_latency_ns) / 1000;
 734         tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
 735         if (tlb_miss > 0)
 736                 entries += tlb_miss;
 737         entries = DIV_ROUND_UP(entries, display->cacheline_size);
 738         *plane_wm = entries + display->guard_size;
 739         if (*plane_wm > (int)display->max_wm)
 740                 *plane_wm = display->max_wm;
 741
 742         /* Use the large buffer method to calculate cursor watermark */
 743         line_time_us = max(htotal * 1000 / clock, 1);
 744         line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
 745         entries = line_count * crtc->cursor->state->crtc_w * cpp;
 746         tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
 747         if (tlb_miss > 0)
 748                 entries += tlb_miss;
 749         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
 750         *cursor_wm = entries + cursor->guard_size;
 751         if (*cursor_wm > (int)cursor->max_wm)
 752                 *cursor_wm = (int)cursor->max_wm;
 753
 754         return true;
 755 }
 756
 757 /*
 758  * Check the wm result.
 759  *
 760  * If any calculated watermark values is larger than the maximum value that
 761  * can be programmed into the associated watermark register, that watermark
 762  * must be disabled.
 763  */
 764 static bool g4x_check_srwm(struct drm_device *dev,
 765                            int display_wm, int cursor_wm,
 766                            const struct intel_watermark_params *display,
 767                            const struct intel_watermark_params *cursor)
 768 {
 769         DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
 770                       display_wm, cursor_wm);
 771
 772         if (display_wm > display->max_wm) {
 773                 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
 774                               display_wm, display->max_wm);
 775                 return false;
 776         }
 777
 778         if (cursor_wm > cursor->max_wm) {
 779                 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
 780                               cursor_wm, cursor->max_wm);
 781                 return false;
 782         }
 783
 784         if (!(display_wm || cursor_wm)) {
 785                 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
 786                 return false;
 787         }
 788
 789         return true;
 790 }
 791
 792 static bool g4x_compute_srwm(struct drm_device *dev,
 793                              int plane,
 794                              int latency_ns,
 795                              const struct intel_watermark_params *display,
 796                              const struct intel_watermark_params *cursor,
 797                              int *display_wm, int *cursor_wm)
 798 {
 799         struct drm_crtc *crtc;
 800         const struct drm_display_mode *adjusted_mode;
 801         int hdisplay, htotal, cpp, clock;
 802         unsigned long line_time_us;
 803         int line_count, line_size;
 804         int small, large;
 805         int entries;
 806
 807         if (!latency_ns) {
 808                 *display_wm = *cursor_wm = 0;
 809                 return false;
 810         }
 811
 812         crtc = intel_get_crtc_for_plane(dev, plane);
 813         adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
 814         clock = adjusted_mode->crtc_clock;
 815         htotal = adjusted_mode->crtc_htotal;
 816         hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
 817         cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
 818
 819         line_time_us = max(htotal * 1000 / clock, 1);
 820         line_count = (latency_ns / line_time_us + 1000) / 1000;
 821         line_size = hdisplay * cpp;
 822
 823         /* Use the minimum of the small and large buffer method for primary */
 824         small = ((clock * cpp / 1000) * latency_ns) / 1000;
 825         large = line_count * line_size;
 826
 827         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
 828         *display_wm = entries + display->guard_size;
 829
 830         /* calculate the self-refresh watermark for display cursor */
 831         entries = line_count * cpp * crtc->cursor->state->crtc_w;
 832         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
 833         *cursor_wm = entries + cursor->guard_size;
 834
 835         return g4x_check_srwm(dev,
 836                               *display_wm, *cursor_wm,
 837                               display, cursor);
 838 }
 839
 840 #define FW_WM_VLV(value, plane) \
 841         (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
 842
 843 static void vlv_write_wm_values(struct intel_crtc *crtc,
 844                                 const struct vlv_wm_values *wm)
 845 {
 846         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
 847         enum pipe pipe = crtc->pipe;
 848
 849         I915_WRITE(VLV_DDL(pipe),
 850                    (wm->ddl[pipe].cursor << DDL_CURSOR_SHIFT) |
 851                    (wm->ddl[pipe].sprite[1] << DDL_SPRITE_SHIFT(1)) |
 852                    (wm->ddl[pipe].sprite[0] << DDL_SPRITE_SHIFT(0)) |
 853                    (wm->ddl[pipe].primary << DDL_PLANE_SHIFT));
 854
 855         I915_WRITE(DSPFW1,
 856                    FW_WM(wm->sr.plane, SR) |
 857                    FW_WM(wm->pipe[PIPE_B].cursor, CURSORB) |
 858                    FW_WM_VLV(wm->pipe[PIPE_B].primary, PLANEB) |
 859                    FW_WM_VLV(wm->pipe[PIPE_A].primary, PLANEA));
 860         I915_WRITE(DSPFW2,
 861                    FW_WM_VLV(wm->pipe[PIPE_A].sprite[1], SPRITEB) |
 862                    FW_WM(wm->pipe[PIPE_A].cursor, CURSORA) |
 863                    FW_WM_VLV(wm->pipe[PIPE_A].sprite[0], SPRITEA));
 864         I915_WRITE(DSPFW3,
 865                    FW_WM(wm->sr.cursor, CURSOR_SR));
 866
 867         if (IS_CHERRYVIEW(dev_priv)) {
 868                 I915_WRITE(DSPFW7_CHV,
 869                            FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
 870                            FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
 871                 I915_WRITE(DSPFW8_CHV,
 872                            FW_WM_VLV(wm->pipe[PIPE_C].sprite[1], SPRITEF) |
 873                            FW_WM_VLV(wm->pipe[PIPE_C].sprite[0], SPRITEE));
 874                 I915_WRITE(DSPFW9_CHV,
 875                            FW_WM_VLV(wm->pipe[PIPE_C].primary, PLANEC) |
 876                            FW_WM(wm->pipe[PIPE_C].cursor, CURSORC));
 877                 I915_WRITE(DSPHOWM,
 878                            FW_WM(wm->sr.plane >> 9, SR_HI) |
 879                            FW_WM(wm->pipe[PIPE_C].sprite[1] >> 8, SPRITEF_HI) |
 880                            FW_WM(wm->pipe[PIPE_C].sprite[0] >> 8, SPRITEE_HI) |
 881                            FW_WM(wm->pipe[PIPE_C].primary >> 8, PLANEC_HI) |
 882                            FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
 883                            FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
 884                            FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
 885                            FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
 886                            FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
 887                            FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
 888         } else {
 889                 I915_WRITE(DSPFW7,
 890                            FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
 891                            FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
 892                 I915_WRITE(DSPHOWM,
 893                            FW_WM(wm->sr.plane >> 9, SR_HI) |
 894                            FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
 895                            FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
 896                            FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
 897                            FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
 898                            FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
 899                            FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
 900         }
 901
 902         /* zero (unused) WM1 watermarks */
 903         I915_WRITE(DSPFW4, 0);
 904         I915_WRITE(DSPFW5, 0);
 905         I915_WRITE(DSPFW6, 0);
 906         I915_WRITE(DSPHOWM1, 0);
 907
 908         POSTING_READ(DSPFW1);
 909 }
 910
 911 #undef FW_WM_VLV
 912
 913 enum vlv_wm_level {
 914         VLV_WM_LEVEL_PM2,
 915         VLV_WM_LEVEL_PM5,
 916         VLV_WM_LEVEL_DDR_DVFS,
 917 };
 918
 919 /* latency must be in 0.1us units. */
 920 static unsigned int vlv_wm_method2(unsigned int pixel_rate,
 921                                    unsigned int pipe_htotal,
 922                                    unsigned int horiz_pixels,
 923                                    unsigned int cpp,
 924                                    unsigned int latency)
 925 {
 926         unsigned int ret;
 927
 928         ret = (latency * pixel_rate) / (pipe_htotal * 10000);
 929         ret = (ret + 1) * horiz_pixels * cpp;
 930         ret = DIV_ROUND_UP(ret, 64);
 931
 932         return ret;
 933 }
 934
 935 static void vlv_setup_wm_latency(struct drm_device *dev)
 936 {
 937         struct drm_i915_private *dev_priv = to_i915(dev);
 938
 939         /* all latencies in usec */
 940         dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM2] = 3;
 941
 942         dev_priv->wm.max_level = VLV_WM_LEVEL_PM2;
 943
 944         if (IS_CHERRYVIEW(dev_priv)) {
 945                 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM5] = 12;
 946                 dev_priv->wm.pri_latency[VLV_WM_LEVEL_DDR_DVFS] = 33;
 947
 948                 dev_priv->wm.max_level = VLV_WM_LEVEL_DDR_DVFS;
 949         }
 950 }
 951
 952 static uint16_t vlv_compute_wm_level(struct intel_plane *plane,
 953                                      struct intel_crtc *crtc,
 954                                      const struct intel_plane_state *state,
 955                                      int level)
 956 {
 957         struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
 958         int clock, htotal, cpp, width, wm;
 959
 960         if (dev_priv->wm.pri_latency[level] == 0)
 961                 return USHRT_MAX;
 962
 963         if (!state->visible)
 964                 return 0;
 965
 966         cpp = drm_format_plane_cpp(state->base.fb->pixel_format, 0);
 967         clock = crtc->config->base.adjusted_mode.crtc_clock;
 968         htotal = crtc->config->base.adjusted_mode.crtc_htotal;
 969         width = crtc->config->pipe_src_w;
 970         if (WARN_ON(htotal == 0))
 971                 htotal = 1;
 972
 973         if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
 974                 /*
 975                  * FIXME the formula gives values that are
 976                  * too big for the cursor FIFO, and hence we
 977                  * would never be able to use cursors. For
 978                  * now just hardcode the watermark.
 979                  */
 980                 wm = 63;
 981         } else {
 982                 wm = vlv_wm_method2(clock, htotal, width, cpp,
 983                                     dev_priv->wm.pri_latency[level] * 10);
 984         }
 985
 986         return min_t(int, wm, USHRT_MAX);
 987 }
 988
 989 static void vlv_compute_fifo(struct intel_crtc *crtc)
 990 {
 991         struct drm_device *dev = crtc->base.dev;
 992         struct vlv_wm_state *wm_state = &crtc->wm_state;
 993         struct intel_plane *plane;
 994         unsigned int total_rate = 0;
 995         const int fifo_size = 512 - 1;
 996         int fifo_extra, fifo_left = fifo_size;
 997
 998         for_each_intel_plane_on_crtc(dev, crtc, plane) {
 999                 struct intel_plane_state *state =
1000                         to_intel_plane_state(plane->base.state);
1001
1002                 if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
1003                         continue;
1004
1005                 if (state->visible) {
1006                         wm_state->num_active_planes++;
1007                         total_rate += drm_format_plane_cpp(state->base.fb->pixel_format, 0);
1008                 }
1009         }
1010
1011         for_each_intel_plane_on_crtc(dev, crtc, plane) {
1012                 struct intel_plane_state *state =
1013                         to_intel_plane_state(plane->base.state);
1014                 unsigned int rate;
1015
1016                 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
1017                         plane->wm.fifo_size = 63;
1018                         continue;
1019                 }
1020
1021                 if (!state->visible) {
1022                         plane->wm.fifo_size = 0;
1023                         continue;
1024                 }
1025
1026                 rate = drm_format_plane_cpp(state->base.fb->pixel_format, 0);
1027                 plane->wm.fifo_size = fifo_size * rate / total_rate;
1028                 fifo_left -= plane->wm.fifo_size;
1029         }
1030
1031         fifo_extra = DIV_ROUND_UP(fifo_left, wm_state->num_active_planes ?: 1);
1032
1033         /* spread the remainder evenly */
1034         for_each_intel_plane_on_crtc(dev, crtc, plane) {
1035                 int plane_extra;
1036
1037                 if (fifo_left == 0)
1038                         break;
1039
1040                 if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
1041                         continue;
1042
1043                 /* give it all to the first plane if none are active */
1044                 if (plane->wm.fifo_size == 0 &&
1045                     wm_state->num_active_planes)
1046                         continue;
1047
1048                 plane_extra = min(fifo_extra, fifo_left);
1049                 plane->wm.fifo_size += plane_extra;
1050                 fifo_left -= plane_extra;
1051         }
1052
1053         WARN_ON(fifo_left != 0);
1054 }
1055
1056 static void vlv_invert_wms(struct intel_crtc *crtc)
1057 {
1058         struct vlv_wm_state *wm_state = &crtc->wm_state;
1059         int level;
1060
1061         for (level = 0; level < wm_state->num_levels; level++) {
1062                 struct drm_device *dev = crtc->base.dev;
1063                 const int sr_fifo_size = INTEL_INFO(dev)->num_pipes * 512 - 1;
1064                 struct intel_plane *plane;
1065
1066                 wm_state->sr[level].plane = sr_fifo_size - wm_state->sr[level].plane;
1067                 wm_state->sr[level].cursor = 63 - wm_state->sr[level].cursor;
1068
1069                 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1070                         switch (plane->base.type) {
1071                                 int sprite;
1072                         case DRM_PLANE_TYPE_CURSOR:
1073                                 wm_state->wm[level].cursor = plane->wm.fifo_size -
1074                                         wm_state->wm[level].cursor;
1075                                 break;
1076                         case DRM_PLANE_TYPE_PRIMARY:
1077                                 wm_state->wm[level].primary = plane->wm.fifo_size -
1078                                         wm_state->wm[level].primary;
1079                                 break;
1080                         case DRM_PLANE_TYPE_OVERLAY:
1081                                 sprite = plane->plane;
1082                                 wm_state->wm[level].sprite[sprite] = plane->wm.fifo_size -
1083                                         wm_state->wm[level].sprite[sprite];
1084                                 break;
1085                         }
1086                 }
1087         }
1088 }
1089
1090 static void vlv_compute_wm(struct intel_crtc *crtc)
1091 {
1092         struct drm_device *dev = crtc->base.dev;
1093         struct vlv_wm_state *wm_state = &crtc->wm_state;
1094         struct intel_plane *plane;
1095         int sr_fifo_size = INTEL_INFO(dev)->num_pipes * 512 - 1;
1096         int level;
1097
1098         memset(wm_state, 0, sizeof(*wm_state));
1099
1100         wm_state->cxsr = crtc->pipe != PIPE_C && crtc->wm.cxsr_allowed;
1101         wm_state->num_levels = to_i915(dev)->wm.max_level + 1;
1102
1103         wm_state->num_active_planes = 0;
1104
1105         vlv_compute_fifo(crtc);
1106
1107         if (wm_state->num_active_planes != 1)
1108                 wm_state->cxsr = false;
1109
1110         if (wm_state->cxsr) {
1111                 for (level = 0; level < wm_state->num_levels; level++) {
1112                         wm_state->sr[level].plane = sr_fifo_size;
1113                         wm_state->sr[level].cursor = 63;
1114                 }
1115         }
1116
1117         for_each_intel_plane_on_crtc(dev, crtc, plane) {
1118                 struct intel_plane_state *state =
1119                         to_intel_plane_state(plane->base.state);
1120
1121                 if (!state->visible)
1122                         continue;
1123
1124                 /* normal watermarks */
1125                 for (level = 0; level < wm_state->num_levels; level++) {
1126                         int wm = vlv_compute_wm_level(plane, crtc, state, level);
1127                         int max_wm = plane->base.type == DRM_PLANE_TYPE_CURSOR ? 63 : 511;
1128
1129                         /* hack */
1130                         if (WARN_ON(level == 0 && wm > max_wm))
1131                                 wm = max_wm;
1132
1133                         if (wm > plane->wm.fifo_size)
1134                                 break;
1135
1136                         switch (plane->base.type) {
1137                                 int sprite;
1138                         case DRM_PLANE_TYPE_CURSOR:
1139                                 wm_state->wm[level].cursor = wm;
1140                                 break;
1141                         case DRM_PLANE_TYPE_PRIMARY:
1142                                 wm_state->wm[level].primary = wm;
1143                                 break;
1144                         case DRM_PLANE_TYPE_OVERLAY:
1145                                 sprite = plane->plane;
1146                                 wm_state->wm[level].sprite[sprite] = wm;
1147                                 break;
1148                         }
1149                 }
1150
1151                 wm_state->num_levels = level;
1152
1153                 if (!wm_state->cxsr)
1154                         continue;
1155
1156                 /* maxfifo watermarks */
1157                 switch (plane->base.type) {
1158                         int sprite, level;
1159                 case DRM_PLANE_TYPE_CURSOR:
1160                         for (level = 0; level < wm_state->num_levels; level++)
1161                                 wm_state->sr[level].cursor =
1162                                         wm_state->wm[level].cursor;
1163                         break;
1164                 case DRM_PLANE_TYPE_PRIMARY:
1165                         for (level = 0; level < wm_state->num_levels; level++)
1166                                 wm_state->sr[level].plane =
1167                                         min(wm_state->sr[level].plane,
1168                                             wm_state->wm[level].primary);
1169                         break;
1170                 case DRM_PLANE_TYPE_OVERLAY:
1171                         sprite = plane->plane;
1172                         for (level = 0; level < wm_state->num_levels; level++)
1173                                 wm_state->sr[level].plane =
1174                                         min(wm_state->sr[level].plane,
1175                                             wm_state->wm[level].sprite[sprite]);
1176                         break;
1177                 }
1178         }
1179
1180         /* clear any (partially) filled invalid levels */
1181         for (level = wm_state->num_levels; level < to_i915(dev)->wm.max_level + 1; level++) {
1182                 memset(&wm_state->wm[level], 0, sizeof(wm_state->wm[level]));
1183                 memset(&wm_state->sr[level], 0, sizeof(wm_state->sr[level]));
1184         }
1185
1186         vlv_invert_wms(crtc);
1187 }
1188
1189 #define VLV_FIFO(plane, value) \
1190         (((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1191
1192 static void vlv_pipe_set_fifo_size(struct intel_crtc *crtc)
1193 {
1194         struct drm_device *dev = crtc->base.dev;
1195         struct drm_i915_private *dev_priv = to_i915(dev);
1196         struct intel_plane *plane;
1197         int sprite0_start = 0, sprite1_start = 0, fifo_size = 0;
1198
1199         for_each_intel_plane_on_crtc(dev, crtc, plane) {
1200                 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
1201                         WARN_ON(plane->wm.fifo_size != 63);
1202                         continue;
1203                 }
1204
1205                 if (plane->base.type == DRM_PLANE_TYPE_PRIMARY)
1206                         sprite0_start = plane->wm.fifo_size;
1207                 else if (plane->plane == 0)
1208                         sprite1_start = sprite0_start + plane->wm.fifo_size;
1209                 else
1210                         fifo_size = sprite1_start + plane->wm.fifo_size;
1211         }
1212
1213         WARN_ON(fifo_size != 512 - 1);
1214
1215         DRM_DEBUG_KMS("Pipe %c FIFO split %d / %d / %d\n",
1216                       pipe_name(crtc->pipe), sprite0_start,
1217                       sprite1_start, fifo_size);
1218
1219         switch (crtc->pipe) {
1220                 uint32_t dsparb, dsparb2, dsparb3;
1221         case PIPE_A:
1222                 dsparb = I915_READ(DSPARB);
1223                 dsparb2 = I915_READ(DSPARB2);
1224
1225                 dsparb &= ~(VLV_FIFO(SPRITEA, 0xff) |
1226                             VLV_FIFO(SPRITEB, 0xff));
1227                 dsparb |= (VLV_FIFO(SPRITEA, sprite0_start) |
1228                            VLV_FIFO(SPRITEB, sprite1_start));
1229
1230                 dsparb2 &= ~(VLV_FIFO(SPRITEA_HI, 0x1) |
1231                              VLV_FIFO(SPRITEB_HI, 0x1));
1232                 dsparb2 |= (VLV_FIFO(SPRITEA_HI, sprite0_start >> 8) |
1233                            VLV_FIFO(SPRITEB_HI, sprite1_start >> 8));
1234
1235                 I915_WRITE(DSPARB, dsparb);
1236                 I915_WRITE(DSPARB2, dsparb2);
1237                 break;
1238         case PIPE_B:
1239                 dsparb = I915_READ(DSPARB);
1240                 dsparb2 = I915_READ(DSPARB2);
1241
1242                 dsparb &= ~(VLV_FIFO(SPRITEC, 0xff) |
1243                             VLV_FIFO(SPRITED, 0xff));
1244                 dsparb |= (VLV_FIFO(SPRITEC, sprite0_start) |
1245                            VLV_FIFO(SPRITED, sprite1_start));
1246
1247                 dsparb2 &= ~(VLV_FIFO(SPRITEC_HI, 0xff) |
1248                              VLV_FIFO(SPRITED_HI, 0xff));
1249                 dsparb2 |= (VLV_FIFO(SPRITEC_HI, sprite0_start >> 8) |
1250                            VLV_FIFO(SPRITED_HI, sprite1_start >> 8));
1251
1252                 I915_WRITE(DSPARB, dsparb);
1253                 I915_WRITE(DSPARB2, dsparb2);
1254                 break;
1255         case PIPE_C:
1256                 dsparb3 = I915_READ(DSPARB3);
1257                 dsparb2 = I915_READ(DSPARB2);
1258
1259                 dsparb3 &= ~(VLV_FIFO(SPRITEE, 0xff) |
1260                              VLV_FIFO(SPRITEF, 0xff));
1261                 dsparb3 |= (VLV_FIFO(SPRITEE, sprite0_start) |
1262                             VLV_FIFO(SPRITEF, sprite1_start));
1263
1264                 dsparb2 &= ~(VLV_FIFO(SPRITEE_HI, 0xff) |
1265                              VLV_FIFO(SPRITEF_HI, 0xff));
1266                 dsparb2 |= (VLV_FIFO(SPRITEE_HI, sprite0_start >> 8) |
1267                            VLV_FIFO(SPRITEF_HI, sprite1_start >> 8));
1268
1269                 I915_WRITE(DSPARB3, dsparb3);
1270                 I915_WRITE(DSPARB2, dsparb2);
1271                 break;
1272         default:
1273                 break;
1274         }
1275 }
1276
1277 #undef VLV_FIFO
1278
1279 static void vlv_merge_wm(struct drm_device *dev,
1280                          struct vlv_wm_values *wm)
1281 {
1282         struct intel_crtc *crtc;
1283         int num_active_crtcs = 0;
1284
1285         wm->level = to_i915(dev)->wm.max_level;
1286         wm->cxsr = true;
1287
1288         for_each_intel_crtc(dev, crtc) {
1289                 const struct vlv_wm_state *wm_state = &crtc->wm_state;
1290
1291                 if (!crtc->active)
1292                         continue;
1293
1294                 if (!wm_state->cxsr)
1295                         wm->cxsr = false;
1296
1297                 num_active_crtcs++;
1298                 wm->level = min_t(int, wm->level, wm_state->num_levels - 1);
1299         }
1300
1301         if (num_active_crtcs != 1)
1302                 wm->cxsr = false;
1303
1304         if (num_active_crtcs > 1)
1305                 wm->level = VLV_WM_LEVEL_PM2;
1306
1307         for_each_intel_crtc(dev, crtc) {
1308                 struct vlv_wm_state *wm_state = &crtc->wm_state;
1309                 enum pipe pipe = crtc->pipe;
1310
1311                 if (!crtc->active)
1312                         continue;
1313
1314                 wm->pipe[pipe] = wm_state->wm[wm->level];
1315                 if (wm->cxsr)
1316                         wm->sr = wm_state->sr[wm->level];
1317
1318                 wm->ddl[pipe].primary = DDL_PRECISION_HIGH | 2;
1319                 wm->ddl[pipe].sprite[0] = DDL_PRECISION_HIGH | 2;
1320                 wm->ddl[pipe].sprite[1] = DDL_PRECISION_HIGH | 2;
1321                 wm->ddl[pipe].cursor = DDL_PRECISION_HIGH | 2;
1322         }
1323 }
1324
1325 static void vlv_update_wm(struct drm_crtc *crtc)
1326 {
1327         struct drm_device *dev = crtc->dev;
1328         struct drm_i915_private *dev_priv = to_i915(dev);
1329         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1330         enum pipe pipe = intel_crtc->pipe;
1331         struct vlv_wm_values wm = {};
1332
1333         vlv_compute_wm(intel_crtc);
1334         vlv_merge_wm(dev, &wm);
1335
1336         if (memcmp(&dev_priv->wm.vlv, &wm, sizeof(wm)) == 0) {
1337                 /* FIXME should be part of crtc atomic commit */
1338                 vlv_pipe_set_fifo_size(intel_crtc);
1339                 return;
1340         }
1341
1342         if (wm.level < VLV_WM_LEVEL_DDR_DVFS &&
1343             dev_priv->wm.vlv.level >= VLV_WM_LEVEL_DDR_DVFS)
1344                 chv_set_memory_dvfs(dev_priv, false);
1345
1346         if (wm.level < VLV_WM_LEVEL_PM5 &&
1347             dev_priv->wm.vlv.level >= VLV_WM_LEVEL_PM5)
1348                 chv_set_memory_pm5(dev_priv, false);
1349
1350         if (!wm.cxsr && dev_priv->wm.vlv.cxsr)
1351                 intel_set_memory_cxsr(dev_priv, false);
1352
1353         /* FIXME should be part of crtc atomic commit */
1354         vlv_pipe_set_fifo_size(intel_crtc);
1355
1356         vlv_write_wm_values(intel_crtc, &wm);
1357
1358         DRM_DEBUG_KMS("Setting FIFO watermarks - %c: plane=%d, cursor=%d, "
1359                       "sprite0=%d, sprite1=%d, SR: plane=%d, cursor=%d level=%d cxsr=%d\n",
1360                       pipe_name(pipe), wm.pipe[pipe].primary, wm.pipe[pipe].cursor,
1361                       wm.pipe[pipe].sprite[0], wm.pipe[pipe].sprite[1],
1362                       wm.sr.plane, wm.sr.cursor, wm.level, wm.cxsr);
1363
1364         if (wm.cxsr && !dev_priv->wm.vlv.cxsr)
1365                 intel_set_memory_cxsr(dev_priv, true);
1366
1367         if (wm.level >= VLV_WM_LEVEL_PM5 &&
1368             dev_priv->wm.vlv.level < VLV_WM_LEVEL_PM5)
1369                 chv_set_memory_pm5(dev_priv, true);
1370
1371         if (wm.level >= VLV_WM_LEVEL_DDR_DVFS &&
1372             dev_priv->wm.vlv.level < VLV_WM_LEVEL_DDR_DVFS)
1373                 chv_set_memory_dvfs(dev_priv, true);
1374
1375         dev_priv->wm.vlv = wm;
1376 }
1377
1378 #define single_plane_enabled(mask) is_power_of_2(mask)
1379
1380 static void g4x_update_wm(struct drm_crtc *crtc)
1381 {
1382         struct drm_device *dev = crtc->dev;
1383         static const int sr_latency_ns = 12000;
1384         struct drm_i915_private *dev_priv = to_i915(dev);
1385         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1386         int plane_sr, cursor_sr;
1387         unsigned int enabled = 0;
1388         bool cxsr_enabled;
1389
1390         if (g4x_compute_wm0(dev, PIPE_A,
1391                             &g4x_wm_info, pessimal_latency_ns,
1392                             &g4x_cursor_wm_info, pessimal_latency_ns,
1393                             &planea_wm, &cursora_wm))
1394                 enabled |= 1 << PIPE_A;
1395
1396         if (g4x_compute_wm0(dev, PIPE_B,
1397                             &g4x_wm_info, pessimal_latency_ns,
1398                             &g4x_cursor_wm_info, pessimal_latency_ns,
1399                             &planeb_wm, &cursorb_wm))
1400                 enabled |= 1 << PIPE_B;
1401
1402         if (single_plane_enabled(enabled) &&
1403             g4x_compute_srwm(dev, ffs(enabled) - 1,
1404                              sr_latency_ns,
1405                              &g4x_wm_info,
1406                              &g4x_cursor_wm_info,
1407                              &plane_sr, &cursor_sr)) {
1408                 cxsr_enabled = true;
1409         } else {
1410                 cxsr_enabled = false;
1411                 intel_set_memory_cxsr(dev_priv, false);
1412                 plane_sr = cursor_sr = 0;
1413         }
1414
1415         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, "
1416                       "B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1417                       planea_wm, cursora_wm,
1418                       planeb_wm, cursorb_wm,
1419                       plane_sr, cursor_sr);
1420
1421         I915_WRITE(DSPFW1,
1422                    FW_WM(plane_sr, SR) |
1423                    FW_WM(cursorb_wm, CURSORB) |
1424                    FW_WM(planeb_wm, PLANEB) |
1425                    FW_WM(planea_wm, PLANEA));
1426         I915_WRITE(DSPFW2,
1427                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1428                    FW_WM(cursora_wm, CURSORA));
1429         /* HPLL off in SR has some issues on G4x... disable it */
1430         I915_WRITE(DSPFW3,
1431                    (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1432                    FW_WM(cursor_sr, CURSOR_SR));
1433
1434         if (cxsr_enabled)
1435                 intel_set_memory_cxsr(dev_priv, true);
1436 }
1437
1438 static void i965_update_wm(struct drm_crtc *unused_crtc)
1439 {
1440         struct drm_device *dev = unused_crtc->dev;
1441         struct drm_i915_private *dev_priv = to_i915(dev);
1442         struct drm_crtc *crtc;
1443         int srwm = 1;
1444         int cursor_sr = 16;
1445         bool cxsr_enabled;
1446
1447         /* Calc sr entries for one plane configs */
1448         crtc = single_enabled_crtc(dev);
1449         if (crtc) {
1450                 /* self-refresh has much higher latency */
1451                 static const int sr_latency_ns = 12000;
1452                 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1453                 int clock = adjusted_mode->crtc_clock;
1454                 int htotal = adjusted_mode->crtc_htotal;
1455                 int hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
1456                 int cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
1457                 unsigned long line_time_us;
1458                 int entries;
1459
1460                 line_time_us = max(htotal * 1000 / clock, 1);
1461
1462                 /* Use ns/us then divide to preserve precision */
1463                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1464                         cpp * hdisplay;
1465                 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1466                 srwm = I965_FIFO_SIZE - entries;
1467                 if (srwm < 0)
1468                         srwm = 1;
1469                 srwm &= 0x1ff;
1470                 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1471                               entries, srwm);
1472
1473                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1474                         cpp * crtc->cursor->state->crtc_w;
1475                 entries = DIV_ROUND_UP(entries,
1476                                           i965_cursor_wm_info.cacheline_size);
1477                 cursor_sr = i965_cursor_wm_info.fifo_size -
1478                         (entries + i965_cursor_wm_info.guard_size);
1479
1480                 if (cursor_sr > i965_cursor_wm_info.max_wm)
1481                         cursor_sr = i965_cursor_wm_info.max_wm;
1482
1483                 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1484                               "cursor %d\n", srwm, cursor_sr);
1485
1486                 cxsr_enabled = true;
1487         } else {
1488                 cxsr_enabled = false;
1489                 /* Turn off self refresh if both pipes are enabled */
1490                 intel_set_memory_cxsr(dev_priv, false);
1491         }
1492
1493         DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1494                       srwm);
1495
1496         /* 965 has limitations... */
1497         I915_WRITE(DSPFW1, FW_WM(srwm, SR) |
1498                    FW_WM(8, CURSORB) |
1499                    FW_WM(8, PLANEB) |
1500                    FW_WM(8, PLANEA));
1501         I915_WRITE(DSPFW2, FW_WM(8, CURSORA) |
1502                    FW_WM(8, PLANEC_OLD));
1503         /* update cursor SR watermark */
1504         I915_WRITE(DSPFW3, FW_WM(cursor_sr, CURSOR_SR));
1505
1506         if (cxsr_enabled)
1507                 intel_set_memory_cxsr(dev_priv, true);
1508 }
1509
1510 #undef FW_WM
1511
1512 static void i9xx_update_wm(struct drm_crtc *unused_crtc)
1513 {
1514         struct drm_device *dev = unused_crtc->dev;
1515         struct drm_i915_private *dev_priv = to_i915(dev);
1516         const struct intel_watermark_params *wm_info;
1517         uint32_t fwater_lo;
1518         uint32_t fwater_hi;
1519         int cwm, srwm = 1;
1520         int fifo_size;
1521         int planea_wm, planeb_wm;
1522         struct drm_crtc *crtc, *enabled = NULL;
1523
1524         if (IS_I945GM(dev))
1525                 wm_info = &i945_wm_info;
1526         else if (!IS_GEN2(dev))
1527                 wm_info = &i915_wm_info;
1528         else
1529                 wm_info = &i830_a_wm_info;
1530
1531         fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1532         crtc = intel_get_crtc_for_plane(dev, 0);
1533         if (intel_crtc_active(crtc)) {
1534                 const struct drm_display_mode *adjusted_mode;
1535                 int cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
1536                 if (IS_GEN2(dev))
1537                         cpp = 4;
1538
1539                 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1540                 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1541                                                wm_info, fifo_size, cpp,
1542                                                pessimal_latency_ns);
1543                 enabled = crtc;
1544         } else {
1545                 planea_wm = fifo_size - wm_info->guard_size;
1546                 if (planea_wm > (long)wm_info->max_wm)
1547                         planea_wm = wm_info->max_wm;
1548         }
1549
1550         if (IS_GEN2(dev))
1551                 wm_info = &i830_bc_wm_info;
1552
1553         fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1554         crtc = intel_get_crtc_for_plane(dev, 1);
1555         if (intel_crtc_active(crtc)) {
1556                 const struct drm_display_mode *adjusted_mode;
1557                 int cpp = drm_format_plane_cpp(crtc->primary->state->fb->pixel_format, 0);
1558                 if (IS_GEN2(dev))
1559                         cpp = 4;
1560
1561                 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1562                 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1563                                                wm_info, fifo_size, cpp,
1564                                                pessimal_latency_ns);
1565                 if (enabled == NULL)
1566                         enabled = crtc;
1567                 else
1568                         enabled = NULL;
1569         } else {
1570                 planeb_wm = fifo_size - wm_info->guard_size;
1571                 if (planeb_wm > (long)wm_info->max_wm)
1572                         planeb_wm = wm_info->max_wm;
1573         }
1574
1575         DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1576
1577         if (IS_I915GM(dev) && enabled) {
1578                 struct drm_i915_gem_object *obj;
1579
1580                 obj = intel_fb_obj(enabled->primary->state->fb);
1581
1582                 /* self-refresh seems busted with untiled */
1583                 if (obj->tiling_mode == I915_TILING_NONE)
1584                         enabled = NULL;
1585         }
1586
1587         /*
1588          * Overlay gets an aggressive default since video jitter is bad.
1589          */
1590         cwm = 2;
1591
1592         /* Play safe and disable self-refresh before adjusting watermarks. */
1593         intel_set_memory_cxsr(dev_priv, false);
1594
1595         /* Calc sr entries for one plane configs */
1596         if (HAS_FW_BLC(dev) && enabled) {
1597                 /* self-refresh has much higher latency */
1598                 static const int sr_latency_ns = 6000;
1599                 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(enabled)->config->base.adjusted_mode;
1600                 int clock = adjusted_mode->crtc_clock;
1601                 int htotal = adjusted_mode->crtc_htotal;
1602                 int hdisplay = to_intel_crtc(enabled)->config->pipe_src_w;
1603                 int cpp = drm_format_plane_cpp(enabled->primary->state->fb->pixel_format, 0);
1604                 unsigned long line_time_us;
1605                 int entries;
1606
1607                 line_time_us = max(htotal * 1000 / clock, 1);
1608
1609                 /* Use ns/us then divide to preserve precision */
1610                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1611                         cpp * hdisplay;
1612                 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1613                 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1614                 srwm = wm_info->fifo_size - entries;
1615                 if (srwm < 0)
1616                         srwm = 1;
1617
1618                 if (IS_I945G(dev) || IS_I945GM(dev))
1619                         I915_WRITE(FW_BLC_SELF,
1620                                    FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1621                 else if (IS_I915GM(dev))
1622                         I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1623         }
1624
1625         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1626                       planea_wm, planeb_wm, cwm, srwm);
1627
1628         fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1629         fwater_hi = (cwm & 0x1f);
1630
1631         /* Set request length to 8 cachelines per fetch */
1632         fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1633         fwater_hi = fwater_hi | (1 << 8);
1634
1635         I915_WRITE(FW_BLC, fwater_lo);
1636         I915_WRITE(FW_BLC2, fwater_hi);
1637
1638         if (enabled)
1639                 intel_set_memory_cxsr(dev_priv, true);
1640 }
1641
1642 static void i845_update_wm(struct drm_crtc *unused_crtc)
1643 {
1644         struct drm_device *dev = unused_crtc->dev;
1645         struct drm_i915_private *dev_priv = to_i915(dev);
1646         struct drm_crtc *crtc;
1647         const struct drm_display_mode *adjusted_mode;
1648         uint32_t fwater_lo;
1649         int planea_wm;
1650
1651         crtc = single_enabled_crtc(dev);
1652         if (crtc == NULL)
1653                 return;
1654
1655         adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1656         planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1657                                        &i845_wm_info,
1658                                        dev_priv->display.get_fifo_size(dev, 0),
1659                                        4, pessimal_latency_ns);
1660         fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1661         fwater_lo |= (3<<8) | planea_wm;
1662
1663         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1664
1665         I915_WRITE(FW_BLC, fwater_lo);
1666 }
1667
1668 uint32_t ilk_pipe_pixel_rate(const struct intel_crtc_state *pipe_config)
1669 {
1670         uint32_t pixel_rate;
1671
1672         pixel_rate = pipe_config->base.adjusted_mode.crtc_clock;
1673
1674         /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
1675          * adjust the pixel_rate here. */
1676
1677         if (pipe_config->pch_pfit.enabled) {
1678                 uint64_t pipe_w, pipe_h, pfit_w, pfit_h;
1679                 uint32_t pfit_size = pipe_config->pch_pfit.size;
1680
1681                 pipe_w = pipe_config->pipe_src_w;
1682                 pipe_h = pipe_config->pipe_src_h;
1683
1684                 pfit_w = (pfit_size >> 16) & 0xFFFF;
1685                 pfit_h = pfit_size & 0xFFFF;
1686                 if (pipe_w < pfit_w)
1687                         pipe_w = pfit_w;
1688                 if (pipe_h < pfit_h)
1689                         pipe_h = pfit_h;
1690
1691                 if (WARN_ON(!pfit_w || !pfit_h))
1692                         return pixel_rate;
1693
1694                 pixel_rate = div_u64((uint64_t) pixel_rate * pipe_w * pipe_h,
1695                                      pfit_w * pfit_h);
1696         }
1697
1698         return pixel_rate;
1699 }
1700
1701 /* latency must be in 0.1us units. */
1702 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t cpp, uint32_t latency)
1703 {
1704         uint64_t ret;
1705
1706         if (WARN(latency == 0, "Latency value missing\n"))
1707                 return UINT_MAX;
1708
1709         ret = (uint64_t) pixel_rate * cpp * latency;
1710         ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
1711
1712         return ret;
1713 }
1714
1715 /* latency must be in 0.1us units. */
1716 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
1717                                uint32_t horiz_pixels, uint8_t cpp,
1718                                uint32_t latency)
1719 {
1720         uint32_t ret;
1721
1722         if (WARN(latency == 0, "Latency value missing\n"))
1723                 return UINT_MAX;
1724         if (WARN_ON(!pipe_htotal))
1725                 return UINT_MAX;
1726
1727         ret = (latency * pixel_rate) / (pipe_htotal * 10000);
1728         ret = (ret + 1) * horiz_pixels * cpp;
1729         ret = DIV_ROUND_UP(ret, 64) + 2;
1730         return ret;
1731 }
1732
1733 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
1734                            uint8_t cpp)
1735 {
1736         /*
1737          * Neither of these should be possible since this function shouldn't be
1738          * called if the CRTC is off or the plane is invisible.  But let's be
1739          * extra paranoid to avoid a potential divide-by-zero if we screw up
1740          * elsewhere in the driver.
1741          */
1742         if (WARN_ON(!cpp))
1743                 return 0;
1744         if (WARN_ON(!horiz_pixels))
1745                 return 0;
1746
1747         return DIV_ROUND_UP(pri_val * 64, horiz_pixels * cpp) + 2;
1748 }
1749
1750 struct ilk_wm_maximums {
1751         uint16_t pri;
1752         uint16_t spr;
1753         uint16_t cur;
1754         uint16_t fbc;
1755 };
1756
1757 /*
1758  * For both WM_PIPE and WM_LP.
1759  * mem_value must be in 0.1us units.
1760  */
1761 static uint32_t ilk_compute_pri_wm(const struct intel_crtc_state *cstate,
1762                                    const struct intel_plane_state *pstate,
1763                                    uint32_t mem_value,
1764                                    bool is_lp)
1765 {
1766         int cpp = pstate->base.fb ?
1767                 drm_format_plane_cpp(pstate->base.fb->pixel_format, 0) : 0;
1768         uint32_t method1, method2;
1769
1770         if (!cstate->base.active || !pstate->visible)
1771                 return 0;
1772
1773         method1 = ilk_wm_method1(ilk_pipe_pixel_rate(cstate), cpp, mem_value);
1774
1775         if (!is_lp)
1776                 return method1;
1777
1778         method2 = ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1779                                  cstate->base.adjusted_mode.crtc_htotal,
1780                                  drm_rect_width(&pstate->dst),
1781                                  cpp, mem_value);
1782
1783         return min(method1, method2);
1784 }
1785
1786 /*
1787  * For both WM_PIPE and WM_LP.
1788  * mem_value must be in 0.1us units.
1789  */
1790 static uint32_t ilk_compute_spr_wm(const struct intel_crtc_state *cstate,
1791                                    const struct intel_plane_state *pstate,
1792                                    uint32_t mem_value)
1793 {
1794         int cpp = pstate->base.fb ?
1795                 drm_format_plane_cpp(pstate->base.fb->pixel_format, 0) : 0;
1796         uint32_t method1, method2;
1797
1798         if (!cstate->base.active || !pstate->visible)
1799                 return 0;
1800
1801         method1 = ilk_wm_method1(ilk_pipe_pixel_rate(cstate), cpp, mem_value);
1802         method2 = ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1803                                  cstate->base.adjusted_mode.crtc_htotal,
1804                                  drm_rect_width(&pstate->dst),
1805                                  cpp, mem_value);
1806         return min(method1, method2);
1807 }
1808
1809 /*
1810  * For both WM_PIPE and WM_LP.
1811  * mem_value must be in 0.1us units.
1812  */
1813 static uint32_t ilk_compute_cur_wm(const struct intel_crtc_state *cstate,
1814                                    const struct intel_plane_state *pstate,
1815                                    uint32_t mem_value)
1816 {
1817         /*
1818          * We treat the cursor plane as always-on for the purposes of watermark
1819          * calculation.  Until we have two-stage watermark programming merged,
1820          * this is necessary to avoid flickering.
1821          */
1822         int cpp = 4;
1823         int width = pstate->visible ? pstate->base.crtc_w : 64;
1824
1825         if (!cstate->base.active)
1826                 return 0;
1827
1828         return ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1829                               cstate->base.adjusted_mode.crtc_htotal,
1830                               width, cpp, mem_value);
1831 }
1832
1833 /* Only for WM_LP. */
1834 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state *cstate,
1835                                    const struct intel_plane_state *pstate,
1836                                    uint32_t pri_val)
1837 {
1838         int cpp = pstate->base.fb ?
1839                 drm_format_plane_cpp(pstate->base.fb->pixel_format, 0) : 0;
1840
1841         if (!cstate->base.active || !pstate->visible)
1842                 return 0;
1843
1844         return ilk_wm_fbc(pri_val, drm_rect_width(&pstate->dst), cpp);
1845 }
1846
1847 static unsigned int ilk_display_fifo_size(const struct drm_device *dev)
1848 {
1849         if (INTEL_INFO(dev)->gen >= 8)
1850                 return 3072;
1851         else if (INTEL_INFO(dev)->gen >= 7)
1852                 return 768;
1853         else
1854                 return 512;
1855 }
1856
1857 static unsigned int ilk_plane_wm_reg_max(const struct drm_device *dev,
1858                                          int level, bool is_sprite)
1859 {
1860         if (INTEL_INFO(dev)->gen >= 8)
1861                 /* BDW primary/sprite plane watermarks */
1862                 return level == 0 ? 255 : 2047;
1863         else if (INTEL_INFO(dev)->gen >= 7)
1864                 /* IVB/HSW primary/sprite plane watermarks */
1865                 return level == 0 ? 127 : 1023;
1866         else if (!is_sprite)
1867                 /* ILK/SNB primary plane watermarks */
1868                 return level == 0 ? 127 : 511;
1869         else
1870                 /* ILK/SNB sprite plane watermarks */
1871                 return level == 0 ? 63 : 255;
1872 }
1873
1874 static unsigned int ilk_cursor_wm_reg_max(const struct drm_device *dev,
1875                                           int level)
1876 {
1877         if (INTEL_INFO(dev)->gen >= 7)
1878                 return level == 0 ? 63 : 255;
1879         else
1880                 return level == 0 ? 31 : 63;
1881 }
1882
1883 static unsigned int ilk_fbc_wm_reg_max(const struct drm_device *dev)
1884 {
1885         if (INTEL_INFO(dev)->gen >= 8)
1886                 return 31;
1887         else
1888                 return 15;
1889 }
1890
1891 /* Calculate the maximum primary/sprite plane watermark */
1892 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
1893                                      int level,
1894                                      const struct intel_wm_config *config,
1895                                      enum intel_ddb_partitioning ddb_partitioning,
1896                                      bool is_sprite)
1897 {
1898         unsigned int fifo_size = ilk_display_fifo_size(dev);
1899
1900         /* if sprites aren't enabled, sprites get nothing */
1901         if (is_sprite && !config->sprites_enabled)
1902                 return 0;
1903
1904         /* HSW allows LP1+ watermarks even with multiple pipes */
1905         if (level == 0 || config->num_pipes_active > 1) {
1906                 fifo_size /= INTEL_INFO(dev)->num_pipes;
1907
1908                 /*
1909                  * For some reason the non self refresh
1910                  * FIFO size is only half of the self
1911                  * refresh FIFO size on ILK/SNB.
1912                  */
1913                 if (INTEL_INFO(dev)->gen <= 6)
1914                         fifo_size /= 2;
1915         }
1916
1917         if (config->sprites_enabled) {
1918                 /* level 0 is always calculated with 1:1 split */
1919                 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
1920                         if (is_sprite)
1921                                 fifo_size *= 5;
1922                         fifo_size /= 6;
1923                 } else {
1924                         fifo_size /= 2;
1925                 }
1926         }
1927
1928         /* clamp to max that the registers can hold */
1929         return min(fifo_size, ilk_plane_wm_reg_max(dev, level, is_sprite));
1930 }
1931
1932 /* Calculate the maximum cursor plane watermark */
1933 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
1934                                       int level,
1935                                       const struct intel_wm_config *config)
1936 {
1937         /* HSW LP1+ watermarks w/ multiple pipes */
1938         if (level > 0 && config->num_pipes_active > 1)
1939                 return 64;
1940
1941         /* otherwise just report max that registers can hold */
1942         return ilk_cursor_wm_reg_max(dev, level);
1943 }
1944
1945 static void ilk_compute_wm_maximums(const struct drm_device *dev,
1946                                     int level,
1947                                     const struct intel_wm_config *config,
1948                                     enum intel_ddb_partitioning ddb_partitioning,
1949                                     struct ilk_wm_maximums *max)
1950 {
1951         max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
1952         max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
1953         max->cur = ilk_cursor_wm_max(dev, level, config);
1954         max->fbc = ilk_fbc_wm_reg_max(dev);
1955 }
1956
1957 static void ilk_compute_wm_reg_maximums(struct drm_device *dev,
1958                                         int level,
1959                                         struct ilk_wm_maximums *max)
1960 {
1961         max->pri = ilk_plane_wm_reg_max(dev, level, false);
1962         max->spr = ilk_plane_wm_reg_max(dev, level, true);
1963         max->cur = ilk_cursor_wm_reg_max(dev, level);
1964         max->fbc = ilk_fbc_wm_reg_max(dev);
1965 }
1966
1967 static bool ilk_validate_wm_level(int level,
1968                                   const struct ilk_wm_maximums *max,
1969                                   struct intel_wm_level *result)
1970 {
1971         bool ret;
1972
1973         /* already determined to be invalid? */
1974         if (!result->enable)
1975                 return false;
1976
1977         result->enable = result->pri_val <= max->pri &&
1978                          result->spr_val <= max->spr &&
1979                          result->cur_val <= max->cur;
1980
1981         ret = result->enable;
1982
1983         /*
1984          * HACK until we can pre-compute everything,
1985          * and thus fail gracefully if LP0 watermarks
1986          * are exceeded...
1987          */
1988         if (level == 0 && !result->enable) {
1989                 if (result->pri_val > max->pri)
1990                         DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
1991                                       level, result->pri_val, max->pri);
1992                 if (result->spr_val > max->spr)
1993                         DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
1994                                       level, result->spr_val, max->spr);
1995                 if (result->cur_val > max->cur)
1996                         DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
1997                                       level, result->cur_val, max->cur);
1998
1999                 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
2000                 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
2001                 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
2002                 result->enable = true;
2003         }
2004
2005         return ret;
2006 }
2007
2008 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
2009                                  const struct intel_crtc *intel_crtc,
2010                                  int level,
2011                                  struct intel_crtc_state *cstate,
2012                                  struct intel_plane_state *pristate,
2013                                  struct intel_plane_state *sprstate,
2014                                  struct intel_plane_state *curstate,
2015                                  struct intel_wm_level *result)
2016 {
2017         uint16_t pri_latency = dev_priv->wm.pri_latency[level];
2018         uint16_t spr_latency = dev_priv->wm.spr_latency[level];
2019         uint16_t cur_latency = dev_priv->wm.cur_latency[level];
2020
2021         /* WM1+ latency values stored in 0.5us units */
2022         if (level > 0) {
2023                 pri_latency *= 5;
2024                 spr_latency *= 5;
2025                 cur_latency *= 5;
2026         }
2027
2028         if (pristate) {
2029                 result->pri_val = ilk_compute_pri_wm(cstate, pristate,
2030                                                      pri_latency, level);
2031                 result->fbc_val = ilk_compute_fbc_wm(cstate, pristate, result->pri_val);
2032         }
2033
2034         if (sprstate)
2035                 result->spr_val = ilk_compute_spr_wm(cstate, sprstate, spr_latency);
2036
2037         if (curstate)
2038                 result->cur_val = ilk_compute_cur_wm(cstate, curstate, cur_latency);
2039
2040         result->enable = true;
2041 }
2042
2043 static uint32_t
2044 hsw_compute_linetime_wm(const struct intel_crtc_state *cstate)
2045 {
2046         const struct intel_atomic_state *intel_state =
2047                 to_intel_atomic_state(cstate->base.state);
2048         const struct drm_display_mode *adjusted_mode =
2049                 &cstate->base.adjusted_mode;
2050         u32 linetime, ips_linetime;
2051
2052         if (!cstate->base.active)
2053                 return 0;
2054         if (WARN_ON(adjusted_mode->crtc_clock == 0))
2055                 return 0;
2056         if (WARN_ON(intel_state->cdclk == 0))
2057                 return 0;
2058
2059         /* The WM are computed with base on how long it takes to fill a single
2060          * row at the given clock rate, multiplied by 8.
2061          * */
2062         linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2063                                      adjusted_mode->crtc_clock);
2064         ips_linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2065                                          intel_state->cdclk);
2066
2067         return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2068                PIPE_WM_LINETIME_TIME(linetime);
2069 }
2070
2071 static void intel_read_wm_latency(struct drm_device *dev, uint16_t wm[8])
2072 {
2073         struct drm_i915_private *dev_priv = to_i915(dev);
2074
2075         if (IS_GEN9(dev)) {
2076                 uint32_t val;
2077                 int ret, i;
2078                 int level, max_level = ilk_wm_max_level(dev);
2079
2080                 /* read the first set of memory latencies[0:3] */
2081                 val = 0; /* data0 to be programmed to 0 for first set */
2082                 mutex_lock(&dev_priv->rps.hw_lock);
2083                 ret = sandybridge_pcode_read(dev_priv,
2084                                              GEN9_PCODE_READ_MEM_LATENCY,
2085                                              &val);
2086                 mutex_unlock(&dev_priv->rps.hw_lock);
2087
2088                 if (ret) {
2089                         DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2090                         return;
2091                 }
2092
2093                 wm[0] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2094                 wm[1] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2095                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2096                 wm[2] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2097                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2098                 wm[3] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2099                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2100
2101                 /* read the second set of memory latencies[4:7] */
2102                 val = 1; /* data0 to be programmed to 1 for second set */
2103                 mutex_lock(&dev_priv->rps.hw_lock);
2104                 ret = sandybridge_pcode_read(dev_priv,
2105                                              GEN9_PCODE_READ_MEM_LATENCY,
2106                                              &val);
2107                 mutex_unlock(&dev_priv->rps.hw_lock);
2108                 if (ret) {
2109                         DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2110                         return;
2111                 }
2112
2113                 wm[4] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2114                 wm[5] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2115                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2116                 wm[6] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2117                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2118                 wm[7] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2119                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2120
2121                 /*
2122                  * WaWmMemoryReadLatency:skl
2123                  *
2124                  * punit doesn't take into account the read latency so we need
2125                  * to add 2us to the various latency levels we retrieve from
2126                  * the punit.
2127                  *   - W0 is a bit special in that it's the only level that
2128                  *   can't be disabled if we want to have display working, so
2129                  *   we always add 2us there.
2130                  *   - For levels >=1, punit returns 0us latency when they are
2131                  *   disabled, so we respect that and don't add 2us then
2132                  *
2133                  * Additionally, if a level n (n > 1) has a 0us latency, all
2134                  * levels m (m >= n) need to be disabled. We make sure to
2135                  * sanitize the values out of the punit to satisfy this
2136                  * requirement.
2137                  */
2138                 wm[0] += 2;
2139                 for (level = 1; level <= max_level; level++)
2140                         if (wm[level] != 0)
2141                                 wm[level] += 2;
2142                         else {
2143                                 for (i = level + 1; i <= max_level; i++)
2144                                         wm[i] = 0;
2145
2146                                 break;
2147                         }
2148         } else if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2149                 uint64_t sskpd = I915_READ64(MCH_SSKPD);
2150
2151                 wm[0] = (sskpd >> 56) & 0xFF;
2152                 if (wm[0] == 0)
2153                         wm[0] = sskpd & 0xF;
2154                 wm[1] = (sskpd >> 4) & 0xFF;
2155                 wm[2] = (sskpd >> 12) & 0xFF;
2156                 wm[3] = (sskpd >> 20) & 0x1FF;
2157                 wm[4] = (sskpd >> 32) & 0x1FF;
2158         } else if (INTEL_INFO(dev)->gen >= 6) {
2159                 uint32_t sskpd = I915_READ(MCH_SSKPD);
2160
2161                 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2162                 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2163                 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2164                 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2165         } else if (INTEL_INFO(dev)->gen >= 5) {
2166                 uint32_t mltr = I915_READ(MLTR_ILK);
2167
2168                 /* ILK primary LP0 latency is 700 ns */
2169                 wm[0] = 7;
2170                 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2171                 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2172         }
2173 }
2174
2175 static void intel_fixup_spr_wm_latency(struct drm_device *dev, uint16_t wm[5])
2176 {
2177         /* ILK sprite LP0 latency is 1300 ns */
2178         if (IS_GEN5(dev))
2179                 wm[0] = 13;
2180 }
2181
2182 static void intel_fixup_cur_wm_latency(struct drm_device *dev, uint16_t wm[5])
2183 {
2184         /* ILK cursor LP0 latency is 1300 ns */
2185         if (IS_GEN5(dev))
2186                 wm[0] = 13;
2187
2188         /* WaDoubleCursorLP3Latency:ivb */
2189         if (IS_IVYBRIDGE(dev))
2190                 wm[3] *= 2;
2191 }
2192
2193 int ilk_wm_max_level(const struct drm_device *dev)
2194 {
2195         /* how many WM levels are we expecting */
2196         if (INTEL_INFO(dev)->gen >= 9)
2197                 return 7;
2198         else if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2199                 return 4;
2200         else if (INTEL_INFO(dev)->gen >= 6)
2201                 return 3;
2202         else
2203                 return 2;
2204 }
2205
2206 static void intel_print_wm_latency(struct drm_device *dev,
2207                                    const char *name,
2208                                    const uint16_t wm[8])
2209 {
2210         int level, max_level = ilk_wm_max_level(dev);
2211
2212         for (level = 0; level <= max_level; level++) {
2213                 unsigned int latency = wm[level];
2214
2215                 if (latency == 0) {
2216                         DRM_ERROR("%s WM%d latency not provided\n",
2217                                   name, level);
2218                         continue;
2219                 }
2220
2221                 /*
2222                  * - latencies are in us on gen9.
2223                  * - before then, WM1+ latency values are in 0.5us units
2224                  */
2225                 if (IS_GEN9(dev))
2226                         latency *= 10;
2227                 else if (level > 0)
2228                         latency *= 5;
2229
2230                 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2231                               name, level, wm[level],
2232                               latency / 10, latency % 10);
2233         }
2234 }
2235
2236 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
2237                                     uint16_t wm[5], uint16_t min)
2238 {
2239         int level, max_level = ilk_wm_max_level(&dev_priv->drm);
2240
2241         if (wm[0] >= min)
2242                 return false;
2243
2244         wm[0] = max(wm[0], min);
2245         for (level = 1; level <= max_level; level++)
2246                 wm[level] = max_t(uint16_t, wm[level], DIV_ROUND_UP(min, 5));
2247
2248         return true;
2249 }
2250
2251 static void snb_wm_latency_quirk(struct drm_device *dev)
2252 {
2253         struct drm_i915_private *dev_priv = to_i915(dev);
2254         bool changed;
2255
2256         /*
2257          * The BIOS provided WM memory latency values are often
2258          * inadequate for high resolution displays. Adjust them.
2259          */
2260         changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
2261                 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
2262                 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
2263
2264         if (!changed)
2265                 return;
2266
2267         DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2268         intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2269         intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2270         intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2271 }
2272
2273 static void ilk_setup_wm_latency(struct drm_device *dev)
2274 {
2275         struct drm_i915_private *dev_priv = to_i915(dev);
2276
2277         intel_read_wm_latency(dev, dev_priv->wm.pri_latency);
2278
2279         memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2280                sizeof(dev_priv->wm.pri_latency));
2281         memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2282                sizeof(dev_priv->wm.pri_latency));
2283
2284         intel_fixup_spr_wm_latency(dev, dev_priv->wm.spr_latency);
2285         intel_fixup_cur_wm_latency(dev, dev_priv->wm.cur_latency);
2286
2287         intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2288         intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2289         intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2290
2291         if (IS_GEN6(dev))
2292                 snb_wm_latency_quirk(dev);
2293 }
2294
2295 static void skl_setup_wm_latency(struct drm_device *dev)
2296 {
2297         struct drm_i915_private *dev_priv = to_i915(dev);
2298
2299         intel_read_wm_latency(dev, dev_priv->wm.skl_latency);
2300         intel_print_wm_latency(dev, "Gen9 Plane", dev_priv->wm.skl_latency);
2301 }
2302
2303 static bool ilk_validate_pipe_wm(struct drm_device *dev,
2304                                  struct intel_pipe_wm *pipe_wm)
2305 {
2306         /* LP0 watermark maximums depend on this pipe alone */
2307         const struct intel_wm_config config = {
2308                 .num_pipes_active = 1,
2309                 .sprites_enabled = pipe_wm->sprites_enabled,
2310                 .sprites_scaled = pipe_wm->sprites_scaled,
2311         };
2312         struct ilk_wm_maximums max;
2313
2314         /* LP0 watermarks always use 1/2 DDB partitioning */
2315         ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
2316
2317         /* At least LP0 must be valid */
2318         if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0])) {
2319                 DRM_DEBUG_KMS("LP0 watermark invalid\n");
2320                 return false;
2321         }
2322
2323         return true;
2324 }
2325
2326 /* Compute new watermarks for the pipe */
2327 static int ilk_compute_pipe_wm(struct intel_crtc_state *cstate)
2328 {
2329         struct drm_atomic_state *state = cstate->base.state;
2330         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
2331         struct intel_pipe_wm *pipe_wm;
2332         struct drm_device *dev = state->dev;
2333         const struct drm_i915_private *dev_priv = to_i915(dev);
2334         struct intel_plane *intel_plane;
2335         struct intel_plane_state *pristate = NULL;
2336         struct intel_plane_state *sprstate = NULL;
2337         struct intel_plane_state *curstate = NULL;
2338         int level, max_level = ilk_wm_max_level(dev), usable_level;
2339         struct ilk_wm_maximums max;
2340
2341         pipe_wm = &cstate->wm.ilk.optimal;
2342
2343         for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
2344                 struct intel_plane_state *ps;
2345
2346                 ps = intel_atomic_get_existing_plane_state(state,
2347                                                            intel_plane);
2348                 if (!ps)
2349                         continue;
2350
2351                 if (intel_plane->base.type == DRM_PLANE_TYPE_PRIMARY)
2352                         pristate = ps;
2353                 else if (intel_plane->base.type == DRM_PLANE_TYPE_OVERLAY)
2354                         sprstate = ps;
2355                 else if (intel_plane->base.type == DRM_PLANE_TYPE_CURSOR)
2356                         curstate = ps;
2357         }
2358
2359         pipe_wm->pipe_enabled = cstate->base.active;
2360         if (sprstate) {
2361                 pipe_wm->sprites_enabled = sprstate->visible;
2362                 pipe_wm->sprites_scaled = sprstate->visible &&
2363                         (drm_rect_width(&sprstate->dst) != drm_rect_width(&sprstate->src) >> 16 ||
2364                          drm_rect_height(&sprstate->dst) != drm_rect_height(&sprstate->src) >> 16);
2365         }
2366
2367         usable_level = max_level;
2368
2369         /* ILK/SNB: LP2+ watermarks only w/o sprites */
2370         if (INTEL_INFO(dev)->gen <= 6 && pipe_wm->sprites_enabled)
2371                 usable_level = 1;
2372
2373         /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2374         if (pipe_wm->sprites_scaled)
2375                 usable_level = 0;
2376
2377         ilk_compute_wm_level(dev_priv, intel_crtc, 0, cstate,
2378                              pristate, sprstate, curstate, &pipe_wm->raw_wm[0]);
2379
2380         memset(&pipe_wm->wm, 0, sizeof(pipe_wm->wm));
2381         pipe_wm->wm[0] = pipe_wm->raw_wm[0];
2382
2383         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2384                 pipe_wm->linetime = hsw_compute_linetime_wm(cstate);
2385
2386         if (!ilk_validate_pipe_wm(dev, pipe_wm))
2387                 return -EINVAL;
2388
2389         ilk_compute_wm_reg_maximums(dev, 1, &max);
2390
2391         for (level = 1; level <= max_level; level++) {
2392                 struct intel_wm_level *wm = &pipe_wm->raw_wm[level];
2393
2394                 ilk_compute_wm_level(dev_priv, intel_crtc, level, cstate,
2395                                      pristate, sprstate, curstate, wm);
2396
2397                 /*
2398                  * Disable any watermark level that exceeds the
2399                  * register maximums since such watermarks are
2400                  * always invalid.
2401                  */
2402                 if (level > usable_level)
2403                         continue;
2404
2405                 if (ilk_validate_wm_level(level, &max, wm))
2406                         pipe_wm->wm[level] = *wm;
2407                 else
2408                         usable_level = level;
2409         }
2410
2411         return 0;
2412 }
2413
2414 /*
2415  * Build a set of 'intermediate' watermark values that satisfy both the old
2416  * state and the new state.  These can be programmed to the hardware
2417  * immediately.
2418  */
2419 static int ilk_compute_intermediate_wm(struct drm_device *dev,
2420                                        struct intel_crtc *intel_crtc,
2421                                        struct intel_crtc_state *newstate)
2422 {
2423         struct intel_pipe_wm *a = &newstate->wm.ilk.intermediate;
2424         struct intel_pipe_wm *b = &intel_crtc->wm.active.ilk;
2425         int level, max_level = ilk_wm_max_level(dev);
2426
2427         /*
2428          * Start with the final, target watermarks, then combine with the
2429          * currently active watermarks to get values that are safe both before
2430          * and after the vblank.
2431          */
2432         *a = newstate->wm.ilk.optimal;
2433         a->pipe_enabled |= b->pipe_enabled;
2434         a->sprites_enabled |= b->sprites_enabled;
2435         a->sprites_scaled |= b->sprites_scaled;
2436
2437         for (level = 0; level <= max_level; level++) {
2438                 struct intel_wm_level *a_wm = &a->wm[level];
2439                 const struct intel_wm_level *b_wm = &b->wm[level];
2440
2441                 a_wm->enable &= b_wm->enable;
2442                 a_wm->pri_val = max(a_wm->pri_val, b_wm->pri_val);
2443                 a_wm->spr_val = max(a_wm->spr_val, b_wm->spr_val);
2444                 a_wm->cur_val = max(a_wm->cur_val, b_wm->cur_val);
2445                 a_wm->fbc_val = max(a_wm->fbc_val, b_wm->fbc_val);
2446         }
2447
2448         /*
2449          * We need to make sure that these merged watermark values are
2450          * actually a valid configuration themselves.  If they're not,
2451          * there's no safe way to transition from the old state to
2452          * the new state, so we need to fail the atomic transaction.
2453          */
2454         if (!ilk_validate_pipe_wm(dev, a))
2455                 return -EINVAL;
2456
2457         /*
2458          * If our intermediate WM are identical to the final WM, then we can
2459          * omit the post-vblank programming; only update if it's different.
2460          */
2461         if (memcmp(a, &newstate->wm.ilk.optimal, sizeof(*a)) == 0)
2462                 newstate->wm.need_postvbl_update = false;
2463
2464         return 0;
2465 }
2466
2467 /*
2468  * Merge the watermarks from all active pipes for a specific level.
2469  */
2470 static void ilk_merge_wm_level(struct drm_device *dev,
2471                                int level,
2472                                struct intel_wm_level *ret_wm)
2473 {
2474         const struct intel_crtc *intel_crtc;
2475
2476         ret_wm->enable = true;
2477
2478         for_each_intel_crtc(dev, intel_crtc) {
2479                 const struct intel_pipe_wm *active = &intel_crtc->wm.active.ilk;
2480                 const struct intel_wm_level *wm = &active->wm[level];
2481
2482                 if (!active->pipe_enabled)
2483                         continue;
2484
2485                 /*
2486                  * The watermark values may have been used in the past,
2487                  * so we must maintain them in the registers for some
2488                  * time even if the level is now disabled.
2489                  */
2490                 if (!wm->enable)
2491                         ret_wm->enable = false;
2492
2493                 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
2494                 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
2495                 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
2496                 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
2497         }
2498 }
2499
2500 /*
2501  * Merge all low power watermarks for all active pipes.
2502  */
2503 static void ilk_wm_merge(struct drm_device *dev,
2504                          const struct intel_wm_config *config,
2505                          const struct ilk_wm_maximums *max,
2506                          struct intel_pipe_wm *merged)
2507 {
2508         struct drm_i915_private *dev_priv = to_i915(dev);
2509         int level, max_level = ilk_wm_max_level(dev);
2510         int last_enabled_level = max_level;
2511
2512         /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2513         if ((INTEL_INFO(dev)->gen <= 6 || IS_IVYBRIDGE(dev)) &&
2514             config->num_pipes_active > 1)
2515                 last_enabled_level = 0;
2516
2517         /* ILK: FBC WM must be disabled always */
2518         merged->fbc_wm_enabled = INTEL_INFO(dev)->gen >= 6;
2519
2520         /* merge each WM1+ level */
2521         for (level = 1; level <= max_level; level++) {
2522                 struct intel_wm_level *wm = &merged->wm[level];
2523
2524                 ilk_merge_wm_level(dev, level, wm);
2525
2526                 if (level > last_enabled_level)
2527                         wm->enable = false;
2528                 else if (!ilk_validate_wm_level(level, max, wm))
2529                         /* make sure all following levels get disabled */
2530                         last_enabled_level = level - 1;
2531
2532                 /*
2533                  * The spec says it is preferred to disable
2534                  * FBC WMs instead of disabling a WM level.
2535                  */
2536                 if (wm->fbc_val > max->fbc) {
2537                         if (wm->enable)
2538                                 merged->fbc_wm_enabled = false;
2539                         wm->fbc_val = 0;
2540                 }
2541         }
2542
2543         /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2544         /*
2545          * FIXME this is racy. FBC might get enabled later.
2546          * What we should check here is whether FBC can be
2547          * enabled sometime later.
2548          */
2549         if (IS_GEN5(dev) && !merged->fbc_wm_enabled &&
2550             intel_fbc_is_active(dev_priv)) {
2551                 for (level = 2; level <= max_level; level++) {
2552                         struct intel_wm_level *wm = &merged->wm[level];
2553
2554                         wm->enable = false;
2555                 }
2556         }
2557 }
2558
2559 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
2560 {
2561         /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2562         return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
2563 }
2564
2565 /* The value we need to program into the WM_LPx latency field */
2566 static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
2567 {
2568         struct drm_i915_private *dev_priv = to_i915(dev);
2569
2570         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2571                 return 2 * level;
2572         else
2573                 return dev_priv->wm.pri_latency[level];
2574 }
2575
2576 static void ilk_compute_wm_results(struct drm_device *dev,
2577                                    const struct intel_pipe_wm *merged,
2578                                    enum intel_ddb_partitioning partitioning,
2579                                    struct ilk_wm_values *results)
2580 {
2581         struct intel_crtc *intel_crtc;
2582         int level, wm_lp;
2583
2584         results->enable_fbc_wm = merged->fbc_wm_enabled;
2585         results->partitioning = partitioning;
2586
2587         /* LP1+ register values */
2588         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2589                 const struct intel_wm_level *r;
2590
2591                 level = ilk_wm_lp_to_level(wm_lp, merged);
2592
2593                 r = &merged->wm[level];
2594
2595                 /*
2596                  * Maintain the watermark values even if the level is
2597                  * disabled. Doing otherwise could cause underruns.
2598                  */
2599                 results->wm_lp[wm_lp - 1] =
2600                         (ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
2601                         (r->pri_val << WM1_LP_SR_SHIFT) |
2602                         r->cur_val;
2603
2604                 if (r->enable)
2605                         results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
2606
2607                 if (INTEL_INFO(dev)->gen >= 8)
2608                         results->wm_lp[wm_lp - 1] |=
2609                                 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
2610                 else
2611                         results->wm_lp[wm_lp - 1] |=
2612                                 r->fbc_val << WM1_LP_FBC_SHIFT;
2613
2614                 /*
2615                  * Always set WM1S_LP_EN when spr_val != 0, even if the
2616                  * level is disabled. Doing otherwise could cause underruns.
2617                  */
2618                 if (INTEL_INFO(dev)->gen <= 6 && r->spr_val) {
2619                         WARN_ON(wm_lp != 1);
2620                         results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
2621                 } else
2622                         results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2623         }
2624
2625         /* LP0 register values */
2626         for_each_intel_crtc(dev, intel_crtc) {
2627                 enum pipe pipe = intel_crtc->pipe;
2628                 const struct intel_wm_level *r =
2629                         &intel_crtc->wm.active.ilk.wm[0];
2630
2631                 if (WARN_ON(!r->enable))
2632                         continue;
2633
2634                 results->wm_linetime[pipe] = intel_crtc->wm.active.ilk.linetime;
2635
2636                 results->wm_pipe[pipe] =
2637                         (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
2638                         (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
2639                         r->cur_val;
2640         }
2641 }
2642
2643 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2644  * case both are at the same level. Prefer r1 in case they're the same. */
2645 static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
2646                                                   struct intel_pipe_wm *r1,
2647                                                   struct intel_pipe_wm *r2)
2648 {
2649         int level, max_level = ilk_wm_max_level(dev);
2650         int level1 = 0, level2 = 0;
2651
2652         for (level = 1; level <= max_level; level++) {
2653                 if (r1->wm[level].enable)
2654                         level1 = level;
2655                 if (r2->wm[level].enable)
2656                         level2 = level;
2657         }
2658
2659         if (level1 == level2) {
2660                 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
2661                         return r2;
2662                 else
2663                         return r1;
2664         } else if (level1 > level2) {
2665                 return r1;
2666         } else {
2667                 return r2;
2668         }
2669 }
2670
2671 /* dirty bits used to track which watermarks need changes */
2672 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2673 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2674 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2675 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2676 #define WM_DIRTY_FBC (1 << 24)
2677 #define WM_DIRTY_DDB (1 << 25)
2678
2679 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private *dev_priv,
2680                                          const struct ilk_wm_values *old,
2681                                          const struct ilk_wm_values *new)
2682 {
2683         unsigned int dirty = 0;
2684         enum pipe pipe;
2685         int wm_lp;
2686
2687         for_each_pipe(dev_priv, pipe) {
2688                 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
2689                         dirty |= WM_DIRTY_LINETIME(pipe);
2690                         /* Must disable LP1+ watermarks too */
2691                         dirty |= WM_DIRTY_LP_ALL;
2692                 }
2693
2694                 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
2695                         dirty |= WM_DIRTY_PIPE(pipe);
2696                         /* Must disable LP1+ watermarks too */
2697                         dirty |= WM_DIRTY_LP_ALL;
2698                 }
2699         }
2700
2701         if (old->enable_fbc_wm != new->enable_fbc_wm) {
2702                 dirty |= WM_DIRTY_FBC;
2703                 /* Must disable LP1+ watermarks too */
2704                 dirty |= WM_DIRTY_LP_ALL;
2705         }
2706
2707         if (old->partitioning != new->partitioning) {
2708                 dirty |= WM_DIRTY_DDB;
2709                 /* Must disable LP1+ watermarks too */
2710                 dirty |= WM_DIRTY_LP_ALL;
2711         }
2712
2713         /* LP1+ watermarks already deemed dirty, no need to continue */
2714         if (dirty & WM_DIRTY_LP_ALL)
2715                 return dirty;
2716
2717         /* Find the lowest numbered LP1+ watermark in need of an update... */
2718         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2719                 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
2720                     old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
2721                         break;
2722         }
2723
2724         /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2725         for (; wm_lp <= 3; wm_lp++)
2726                 dirty |= WM_DIRTY_LP(wm_lp);
2727
2728         return dirty;
2729 }
2730
2731 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
2732                                unsigned int dirty)
2733 {
2734         struct ilk_wm_values *previous = &dev_priv->wm.hw;
2735         bool changed = false;
2736
2737         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
2738                 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
2739                 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
2740                 changed = true;
2741         }
2742         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
2743                 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
2744                 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
2745                 changed = true;
2746         }
2747         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
2748                 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
2749                 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
2750                 changed = true;
2751         }
2752
2753         /*
2754          * Don't touch WM1S_LP_EN here.
2755          * Doing so could cause underruns.
2756          */
2757
2758         return changed;
2759 }
2760
2761 /*
2762  * The spec says we shouldn't write when we don't need, because every write
2763  * causes WMs to be re-evaluated, expending some power.
2764  */
2765 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
2766                                 struct ilk_wm_values *results)
2767 {
2768         struct drm_device *dev = &dev_priv->drm;
2769         struct ilk_wm_values *previous = &dev_priv->wm.hw;
2770         unsigned int dirty;
2771         uint32_t val;
2772
2773         dirty = ilk_compute_wm_dirty(dev_priv, previous, results);
2774         if (!dirty)
2775                 return;
2776
2777         _ilk_disable_lp_wm(dev_priv, dirty);
2778
2779         if (dirty & WM_DIRTY_PIPE(PIPE_A))
2780                 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2781         if (dirty & WM_DIRTY_PIPE(PIPE_B))
2782                 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
2783         if (dirty & WM_DIRTY_PIPE(PIPE_C))
2784                 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
2785
2786         if (dirty & WM_DIRTY_LINETIME(PIPE_A))
2787                 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
2788         if (dirty & WM_DIRTY_LINETIME(PIPE_B))
2789                 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
2790         if (dirty & WM_DIRTY_LINETIME(PIPE_C))
2791                 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
2792
2793         if (dirty & WM_DIRTY_DDB) {
2794                 if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2795                         val = I915_READ(WM_MISC);
2796                         if (results->partitioning == INTEL_DDB_PART_1_2)
2797                                 val &= ~WM_MISC_DATA_PARTITION_5_6;
2798                         else
2799                                 val |= WM_MISC_DATA_PARTITION_5_6;
2800                         I915_WRITE(WM_MISC, val);
2801                 } else {
2802                         val = I915_READ(DISP_ARB_CTL2);
2803                         if (results->partitioning == INTEL_DDB_PART_1_2)
2804                                 val &= ~DISP_DATA_PARTITION_5_6;
2805                         else
2806                                 val |= DISP_DATA_PARTITION_5_6;
2807                         I915_WRITE(DISP_ARB_CTL2, val);
2808                 }
2809         }
2810
2811         if (dirty & WM_DIRTY_FBC) {
2812                 val = I915_READ(DISP_ARB_CTL);
2813                 if (results->enable_fbc_wm)
2814                         val &= ~DISP_FBC_WM_DIS;
2815                 else
2816                         val |= DISP_FBC_WM_DIS;
2817                 I915_WRITE(DISP_ARB_CTL, val);
2818         }
2819
2820         if (dirty & WM_DIRTY_LP(1) &&
2821             previous->wm_lp_spr[0] != results->wm_lp_spr[0])
2822                 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
2823
2824         if (INTEL_INFO(dev)->gen >= 7) {
2825                 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
2826                         I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
2827                 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
2828                         I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
2829         }
2830
2831         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
2832                 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
2833         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
2834                 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
2835         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
2836                 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
2837
2838         dev_priv->wm.hw = *results;
2839 }
2840
2841 bool ilk_disable_lp_wm(struct drm_device *dev)
2842 {
2843         struct drm_i915_private *dev_priv = to_i915(dev);
2844
2845         return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
2846 }
2847
2848 /*
2849  * On gen9, we need to allocate Display Data Buffer (DDB) portions to the
2850  * different active planes.
2851  */
2852
2853 #define SKL_DDB_SIZE            896     /* in blocks */
2854 #define BXT_DDB_SIZE            512
2855
2856 /*
2857  * Return the index of a plane in the SKL DDB and wm result arrays.  Primary
2858  * plane is always in slot 0, cursor is always in slot I915_MAX_PLANES-1, and
2859  * other universal planes are in indices 1..n.  Note that this may leave unused
2860  * indices between the top "sprite" plane and the cursor.
2861  */
2862 static int
2863 skl_wm_plane_id(const struct intel_plane *plane)
2864 {
2865         switch (plane->base.type) {
2866         case DRM_PLANE_TYPE_PRIMARY:
2867                 return 0;
2868         case DRM_PLANE_TYPE_CURSOR:
2869                 return PLANE_CURSOR;
2870         case DRM_PLANE_TYPE_OVERLAY:
2871                 return plane->plane + 1;
2872         default:
2873                 MISSING_CASE(plane->base.type);
2874                 return plane->plane;
2875         }
2876 }
2877
2878 static void
2879 skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
2880                                    const struct intel_crtc_state *cstate,
2881                                    struct skl_ddb_entry *alloc, /* out */
2882                                    int *num_active /* out */)
2883 {
2884         struct drm_atomic_state *state = cstate->base.state;
2885         struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
2886         struct drm_i915_private *dev_priv = to_i915(dev);
2887         struct drm_crtc *for_crtc = cstate->base.crtc;
2888         unsigned int pipe_size, ddb_size;
2889         int nth_active_pipe;
2890         int pipe = to_intel_crtc(for_crtc)->pipe;
2891
2892         if (WARN_ON(!state) || !cstate->base.active) {
2893                 alloc->start = 0;
2894                 alloc->end = 0;
2895                 *num_active = hweight32(dev_priv->active_crtcs);
2896                 return;
2897         }
2898
2899         if (intel_state->active_pipe_changes)
2900                 *num_active = hweight32(intel_state->active_crtcs);
2901         else
2902                 *num_active = hweight32(dev_priv->active_crtcs);
2903
2904         if (IS_BROXTON(dev))
2905                 ddb_size = BXT_DDB_SIZE;
2906         else
2907                 ddb_size = SKL_DDB_SIZE;
2908
2909         ddb_size -= 4; /* 4 blocks for bypass path allocation */
2910
2911         /*
2912          * If the state doesn't change the active CRTC's, then there's
2913          * no need to recalculate; the existing pipe allocation limits
2914          * should remain unchanged.  Note that we're safe from racing
2915          * commits since any racing commit that changes the active CRTC
2916          * list would need to grab _all_ crtc locks, including the one
2917          * we currently hold.
2918          */
2919         if (!intel_state->active_pipe_changes) {
2920                 *alloc = dev_priv->wm.skl_hw.ddb.pipe[pipe];
2921                 return;
2922         }
2923
2924         nth_active_pipe = hweight32(intel_state->active_crtcs &
2925                                     (drm_crtc_mask(for_crtc) - 1));
2926         pipe_size = ddb_size / hweight32(intel_state->active_crtcs);
2927         alloc->start = nth_active_pipe * ddb_size / *num_active;
2928         alloc->end = alloc->start + pipe_size;
2929 }
2930
2931 static unsigned int skl_cursor_allocation(int num_active)
2932 {
2933         if (num_active == 1)
2934                 return 32;
2935
2936         return 8;
2937 }
2938
2939 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry *entry, u32 reg)
2940 {
2941         entry->start = reg & 0x3ff;
2942         entry->end = (reg >> 16) & 0x3ff;
2943         if (entry->end)
2944                 entry->end += 1;
2945 }
2946
2947 void skl_ddb_get_hw_state(struct drm_i915_private *dev_priv,
2948                           struct skl_ddb_allocation *ddb /* out */)
2949 {
2950         enum pipe pipe;
2951         int plane;
2952         u32 val;
2953
2954         memset(ddb, 0, sizeof(*ddb));
2955
2956         for_each_pipe(dev_priv, pipe) {
2957                 enum intel_display_power_domain power_domain;
2958
2959                 power_domain = POWER_DOMAIN_PIPE(pipe);
2960                 if (!intel_display_power_get_if_enabled(dev_priv, power_domain))
2961                         continue;
2962
2963                 for_each_plane(dev_priv, pipe, plane) {
2964                         val = I915_READ(PLANE_BUF_CFG(pipe, plane));
2965                         skl_ddb_entry_init_from_hw(&ddb->plane[pipe][plane],
2966                                                    val);
2967                 }
2968
2969                 val = I915_READ(CUR_BUF_CFG(pipe));
2970                 skl_ddb_entry_init_from_hw(&ddb->plane[pipe][PLANE_CURSOR],
2971                                            val);
2972
2973                 intel_display_power_put(dev_priv, power_domain);
2974         }
2975 }
2976
2977 /*
2978  * Determines the downscale amount of a plane for the purposes of watermark calculations.
2979  * The bspec defines downscale amount as:
2980  *
2981  * """
2982  * Horizontal down scale amount = maximum[1, Horizontal source size /
2983  *                                           Horizontal destination size]
2984  * Vertical down scale amount = maximum[1, Vertical source size /
2985  *                                         Vertical destination size]
2986  * Total down scale amount = Horizontal down scale amount *
2987  *                           Vertical down scale amount
2988  * """
2989  *
2990  * Return value is provided in 16.16 fixed point form to retain fractional part.
2991  * Caller should take care of dividing & rounding off the value.
2992  */
2993 static uint32_t
2994 skl_plane_downscale_amount(const struct intel_plane_state *pstate)
2995 {
2996         uint32_t downscale_h, downscale_w;
2997         uint32_t src_w, src_h, dst_w, dst_h;
2998
2999         if (WARN_ON(!pstate->visible))
3000                 return DRM_PLANE_HELPER_NO_SCALING;
3001
3002         /* n.b., src is 16.16 fixed point, dst is whole integer */
3003         src_w = drm_rect_width(&pstate->src);
3004         src_h = drm_rect_height(&pstate->src);
3005         dst_w = drm_rect_width(&pstate->dst);
3006         dst_h = drm_rect_height(&pstate->dst);
3007         if (intel_rotation_90_or_270(pstate->base.rotation))
3008                 swap(dst_w, dst_h);
3009
3010         downscale_h = max(src_h / dst_h, (uint32_t)DRM_PLANE_HELPER_NO_SCALING);
3011         downscale_w = max(src_w / dst_w, (uint32_t)DRM_PLANE_HELPER_NO_SCALING);
3012
3013         /* Provide result in 16.16 fixed point */
3014         return (uint64_t)downscale_w * downscale_h >> 16;
3015 }
3016
3017 static unsigned int
3018 skl_plane_relative_data_rate(const struct intel_crtc_state *cstate,
3019                              const struct drm_plane_state *pstate,
3020                              int y)
3021 {
3022         struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
3023         struct drm_framebuffer *fb = pstate->fb;
3024         uint32_t down_scale_amount, data_rate;
3025         uint32_t width = 0, height = 0;
3026         unsigned format = fb ? fb->pixel_format : DRM_FORMAT_XRGB8888;
3027
3028         if (!intel_pstate->visible)
3029                 return 0;
3030         if (pstate->plane->type == DRM_PLANE_TYPE_CURSOR)
3031                 return 0;
3032         if (y && format != DRM_FORMAT_NV12)
3033                 return 0;
3034
3035         width = drm_rect_width(&intel_pstate->src) >> 16;
3036         height = drm_rect_height(&intel_pstate->src) >> 16;
3037
3038         if (intel_rotation_90_or_270(pstate->rotation))
3039                 swap(width, height);
3040
3041         /* for planar format */
3042         if (format == DRM_FORMAT_NV12) {
3043                 if (y)  /* y-plane data rate */
3044                         data_rate = width * height *
3045                                 drm_format_plane_cpp(format, 0);
3046                 else    /* uv-plane data rate */
3047                         data_rate = (width / 2) * (height / 2) *
3048                                 drm_format_plane_cpp(format, 1);
3049         } else {
3050                 /* for packed formats */
3051                 data_rate = width * height * drm_format_plane_cpp(format, 0);
3052         }
3053
3054         down_scale_amount = skl_plane_downscale_amount(intel_pstate);
3055
3056         return (uint64_t)data_rate * down_scale_amount >> 16;
3057 }
3058
3059 /*
3060  * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
3061  * a 8192x4096@32bpp framebuffer:
3062  *   3 * 4096 * 8192  * 4 < 2^32
3063  */
3064 static unsigned int
3065 skl_get_total_relative_data_rate(struct intel_crtc_state *intel_cstate)
3066 {
3067         struct drm_crtc_state *cstate = &intel_cstate->base;
3068         struct drm_atomic_state *state = cstate->state;
3069         struct drm_crtc *crtc = cstate->crtc;
3070         struct drm_device *dev = crtc->dev;
3071         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3072         const struct drm_plane *plane;
3073         const struct intel_plane *intel_plane;
3074         struct drm_plane_state *pstate;
3075         unsigned int rate, total_data_rate = 0;
3076         int id;
3077         int i;
3078
3079         if (WARN_ON(!state))
3080                 return 0;
3081
3082         /* Calculate and cache data rate for each plane */
3083         for_each_plane_in_state(state, plane, pstate, i) {
3084                 id = skl_wm_plane_id(to_intel_plane(plane));
3085                 intel_plane = to_intel_plane(plane);
3086
3087                 if (intel_plane->pipe != intel_crtc->pipe)
3088                         continue;
3089
3090                 /* packed/uv */
3091                 rate = skl_plane_relative_data_rate(intel_cstate,
3092                                                     pstate, 0);
3093                 intel_cstate->wm.skl.plane_data_rate[id] = rate;
3094
3095                 /* y-plane */
3096                 rate = skl_plane_relative_data_rate(intel_cstate,
3097                                                     pstate, 1);
3098                 intel_cstate->wm.skl.plane_y_data_rate[id] = rate;
3099         }
3100
3101         /* Calculate CRTC's total data rate from cached values */
3102         for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
3103                 int id = skl_wm_plane_id(intel_plane);
3104
3105                 /* packed/uv */
3106                 total_data_rate += intel_cstate->wm.skl.plane_data_rate[id];
3107                 total_data_rate += intel_cstate->wm.skl.plane_y_data_rate[id];
3108         }
3109
3110         WARN_ON(cstate->plane_mask && total_data_rate == 0);
3111
3112         return total_data_rate;
3113 }
3114
3115 static uint16_t
3116 skl_ddb_min_alloc(const struct drm_plane_state *pstate,
3117                   const int y)
3118 {
3119         struct drm_framebuffer *fb = pstate->fb;
3120         struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
3121         uint32_t src_w, src_h;
3122         uint32_t min_scanlines = 8;
3123         uint8_t plane_bpp;
3124
3125         if (WARN_ON(!fb))
3126                 return 0;
3127
3128         /* For packed formats, no y-plane, return 0 */
3129         if (y && fb->pixel_format != DRM_FORMAT_NV12)
3130                 return 0;
3131
3132         /* For Non Y-tile return 8-blocks */
3133         if (fb->modifier[0] != I915_FORMAT_MOD_Y_TILED &&
3134             fb->modifier[0] != I915_FORMAT_MOD_Yf_TILED)
3135                 return 8;
3136
3137         src_w = drm_rect_width(&intel_pstate->src) >> 16;
3138         src_h = drm_rect_height(&intel_pstate->src) >> 16;
3139
3140         if (intel_rotation_90_or_270(pstate->rotation))
3141                 swap(src_w, src_h);
3142
3143         /* Halve UV plane width and height for NV12 */
3144         if (fb->pixel_format == DRM_FORMAT_NV12 && !y) {
3145                 src_w /= 2;
3146                 src_h /= 2;
3147         }
3148
3149         if (fb->pixel_format == DRM_FORMAT_NV12 && !y)
3150                 plane_bpp = drm_format_plane_cpp(fb->pixel_format, 1);
3151         else
3152                 plane_bpp = drm_format_plane_cpp(fb->pixel_format, 0);
3153
3154         if (intel_rotation_90_or_270(pstate->rotation)) {
3155                 switch (plane_bpp) {
3156                 case 1:
3157                         min_scanlines = 32;
3158                         break;
3159                 case 2:
3160                         min_scanlines = 16;
3161                         break;
3162                 case 4:
3163                         min_scanlines = 8;
3164                         break;
3165                 case 8:
3166                         min_scanlines = 4;
3167                         break;
3168                 default:
3169                         WARN(1, "Unsupported pixel depth %u for rotation",
3170                              plane_bpp);
3171                         min_scanlines = 32;
3172                 }
3173         }
3174
3175         return DIV_ROUND_UP((4 * src_w * plane_bpp), 512) * min_scanlines/4 + 3;
3176 }
3177
3178 static int
3179 skl_allocate_pipe_ddb(struct intel_crtc_state *cstate,
3180                       struct skl_ddb_allocation *ddb /* out */)
3181 {
3182         struct drm_atomic_state *state = cstate->base.state;
3183         struct drm_crtc *crtc = cstate->base.crtc;
3184         struct drm_device *dev = crtc->dev;
3185         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3186         struct intel_plane *intel_plane;
3187         struct drm_plane *plane;
3188         struct drm_plane_state *pstate;
3189         enum pipe pipe = intel_crtc->pipe;
3190         struct skl_ddb_entry *alloc = &ddb->pipe[pipe];
3191         uint16_t alloc_size, start, cursor_blocks;
3192         uint16_t *minimum = cstate->wm.skl.minimum_blocks;
3193         uint16_t *y_minimum = cstate->wm.skl.minimum_y_blocks;
3194         unsigned int total_data_rate;
3195         int num_active;
3196         int id, i;
3197
3198         if (WARN_ON(!state))
3199                 return 0;
3200
3201         if (!cstate->base.active) {
3202                 ddb->pipe[pipe].start = ddb->pipe[pipe].end = 0;
3203                 memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
3204                 memset(ddb->y_plane[pipe], 0, sizeof(ddb->y_plane[pipe]));
3205                 return 0;
3206         }
3207
3208         skl_ddb_get_pipe_allocation_limits(dev, cstate, alloc, &num_active);
3209         alloc_size = skl_ddb_entry_size(alloc);
3210         if (alloc_size == 0) {
3211                 memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
3212                 return 0;
3213         }
3214
3215         cursor_blocks = skl_cursor_allocation(num_active);
3216         ddb->plane[pipe][PLANE_CURSOR].start = alloc->end - cursor_blocks;
3217         ddb->plane[pipe][PLANE_CURSOR].end = alloc->end;
3218
3219         alloc_size -= cursor_blocks;
3220
3221         /* 1. Allocate the mininum required blocks for each active plane */
3222         for_each_plane_in_state(state, plane, pstate, i) {
3223                 intel_plane = to_intel_plane(plane);
3224                 id = skl_wm_plane_id(intel_plane);
3225
3226                 if (intel_plane->pipe != pipe)
3227                         continue;
3228
3229                 if (!to_intel_plane_state(pstate)->visible) {
3230                         minimum[id] = 0;
3231                         y_minimum[id] = 0;
3232                         continue;
3233                 }
3234                 if (plane->type == DRM_PLANE_TYPE_CURSOR) {
3235                         minimum[id] = 0;
3236                         y_minimum[id] = 0;
3237                         continue;
3238                 }
3239
3240                 minimum[id] = skl_ddb_min_alloc(pstate, 0);
3241                 y_minimum[id] = skl_ddb_min_alloc(pstate, 1);
3242         }
3243
3244         for (i = 0; i < PLANE_CURSOR; i++) {
3245                 alloc_size -= minimum[i];
3246                 alloc_size -= y_minimum[i];
3247         }
3248
3249         /*
3250          * 2. Distribute the remaining space in proportion to the amount of
3251          * data each plane needs to fetch from memory.
3252          *
3253          * FIXME: we may not allocate every single block here.
3254          */
3255         total_data_rate = skl_get_total_relative_data_rate(cstate);
3256         if (total_data_rate == 0)
3257                 return 0;
3258
3259         start = alloc->start;
3260         for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
3261                 unsigned int data_rate, y_data_rate;
3262                 uint16_t plane_blocks, y_plane_blocks = 0;
3263                 int id = skl_wm_plane_id(intel_plane);
3264
3265                 data_rate = cstate->wm.skl.plane_data_rate[id];
3266
3267                 /*
3268                  * allocation for (packed formats) or (uv-plane part of planar format):
3269                  * promote the expression to 64 bits to avoid overflowing, the
3270                  * result is < available as data_rate / total_data_rate < 1
3271                  */
3272                 plane_blocks = minimum[id];
3273                 plane_blocks += div_u64((uint64_t)alloc_size * data_rate,
3274                                         total_data_rate);
3275
3276                 /* Leave disabled planes at (0,0) */
3277                 if (data_rate) {
3278                         ddb->plane[pipe][id].start = start;
3279                         ddb->plane[pipe][id].end = start + plane_blocks;
3280                 }
3281
3282                 start += plane_blocks;
3283
3284                 /*
3285                  * allocation for y_plane part of planar format:
3286                  */
3287                 y_data_rate = cstate->wm.skl.plane_y_data_rate[id];
3288
3289                 y_plane_blocks = y_minimum[id];
3290                 y_plane_blocks += div_u64((uint64_t)alloc_size * y_data_rate,
3291                                         total_data_rate);
3292
3293                 if (y_data_rate) {
3294                         ddb->y_plane[pipe][id].start = start;
3295                         ddb->y_plane[pipe][id].end = start + y_plane_blocks;
3296                 }
3297
3298                 start += y_plane_blocks;
3299         }
3300
3301         return 0;
3302 }
3303
3304 static uint32_t skl_pipe_pixel_rate(const struct intel_crtc_state *config)
3305 {
3306         /* TODO: Take into account the scalers once we support them */
3307         return config->base.adjusted_mode.crtc_clock;
3308 }
3309
3310 /*
3311  * The max latency should be 257 (max the punit can code is 255 and we add 2us
3312  * for the read latency) and cpp should always be <= 8, so that
3313  * should allow pixel_rate up to ~2 GHz which seems sufficient since max
3314  * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
3315 */
3316 static uint32_t skl_wm_method1(uint32_t pixel_rate, uint8_t cpp, uint32_t latency)
3317 {
3318         uint32_t wm_intermediate_val, ret;
3319
3320         if (latency == 0)
3321                 return UINT_MAX;
3322
3323         wm_intermediate_val = latency * pixel_rate * cpp / 512;
3324         ret = DIV_ROUND_UP(wm_intermediate_val, 1000);
3325
3326         return ret;
3327 }
3328
3329 static uint32_t skl_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
3330                                uint32_t horiz_pixels, uint8_t cpp,
3331                                uint64_t tiling, uint32_t latency)
3332 {
3333         uint32_t ret;
3334         uint32_t plane_bytes_per_line, plane_blocks_per_line;
3335         uint32_t wm_intermediate_val;
3336
3337         if (latency == 0)
3338                 return UINT_MAX;
3339
3340         plane_bytes_per_line = horiz_pixels * cpp;
3341
3342         if (tiling == I915_FORMAT_MOD_Y_TILED ||
3343             tiling == I915_FORMAT_MOD_Yf_TILED) {
3344                 plane_bytes_per_line *= 4;
3345                 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
3346                 plane_blocks_per_line /= 4;
3347         } else {
3348                 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
3349         }
3350
3351         wm_intermediate_val = latency * pixel_rate;
3352         ret = DIV_ROUND_UP(wm_intermediate_val, pipe_htotal * 1000) *
3353                                 plane_blocks_per_line;
3354
3355         return ret;
3356 }
3357
3358 static uint32_t skl_adjusted_plane_pixel_rate(const struct intel_crtc_state *cstate,
3359                                               struct intel_plane_state *pstate)
3360 {
3361         uint64_t adjusted_pixel_rate;
3362         uint64_t downscale_amount;
3363         uint64_t pixel_rate;
3364
3365         /* Shouldn't reach here on disabled planes... */
3366         if (WARN_ON(!pstate->visible))
3367                 return 0;
3368
3369         /*
3370          * Adjusted plane pixel rate is just the pipe's adjusted pixel rate
3371          * with additional adjustments for plane-specific scaling.
3372          */
3373         adjusted_pixel_rate = skl_pipe_pixel_rate(cstate);
3374         downscale_amount = skl_plane_downscale_amount(pstate);
3375
3376         pixel_rate = adjusted_pixel_rate * downscale_amount >> 16;
3377         WARN_ON(pixel_rate != clamp_t(uint32_t, pixel_rate, 0, ~0));
3378
3379         return pixel_rate;
3380 }
3381
3382 static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv,
3383                                 struct intel_crtc_state *cstate,
3384                                 struct intel_plane_state *intel_pstate,
3385                                 uint16_t ddb_allocation,
3386                                 int level,
3387                                 uint16_t *out_blocks, /* out */
3388                                 uint8_t *out_lines, /* out */
3389                                 bool *enabled /* out */)
3390 {
3391         struct drm_plane_state *pstate = &intel_pstate->base;
3392         struct drm_framebuffer *fb = pstate->fb;
3393         uint32_t latency = dev_priv->wm.skl_latency[level];
3394         uint32_t method1, method2;
3395         uint32_t plane_bytes_per_line, plane_blocks_per_line;
3396         uint32_t res_blocks, res_lines;
3397         uint32_t selected_result;
3398         uint8_t cpp;
3399         uint32_t width = 0, height = 0;
3400         uint32_t plane_pixel_rate;
3401
3402         if (latency == 0 || !cstate->base.active || !intel_pstate->visible) {
3403                 *enabled = false;
3404                 return 0;
3405         }
3406
3407         width = drm_rect_width(&intel_pstate->src) >> 16;
3408         height = drm_rect_height(&intel_pstate->src) >> 16;
3409
3410         if (intel_rotation_90_or_270(pstate->rotation))
3411                 swap(width, height);
3412
3413         cpp = drm_format_plane_cpp(fb->pixel_format, 0);
3414         plane_pixel_rate = skl_adjusted_plane_pixel_rate(cstate, intel_pstate);
3415
3416         method1 = skl_wm_method1(plane_pixel_rate, cpp, latency);
3417         method2 = skl_wm_method2(plane_pixel_rate,
3418                                  cstate->base.adjusted_mode.crtc_htotal,
3419                                  width,
3420                                  cpp,
3421                                  fb->modifier[0],
3422                                  latency);
3423
3424         plane_bytes_per_line = width * cpp;
3425         plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
3426
3427         if (fb->modifier[0] == I915_FORMAT_MOD_Y_TILED ||
3428             fb->modifier[0] == I915_FORMAT_MOD_Yf_TILED) {
3429                 uint32_t min_scanlines = 4;
3430                 uint32_t y_tile_minimum;
3431                 if (intel_rotation_90_or_270(pstate->rotation)) {
3432                         int cpp = (fb->pixel_format == DRM_FORMAT_NV12) ?
3433                                 drm_format_plane_cpp(fb->pixel_format, 1) :
3434                                 drm_format_plane_cpp(fb->pixel_format, 0);
3435
3436                         switch (cpp) {
3437                         case 1:
3438                                 min_scanlines = 16;
3439                                 break;
3440                         case 2:
3441                                 min_scanlines = 8;
3442                                 break;
3443                         case 8:
3444                                 WARN(1, "Unsupported pixel depth for rotation");
3445                         }
3446                 }
3447                 y_tile_minimum = plane_blocks_per_line * min_scanlines;
3448                 selected_result = max(method2, y_tile_minimum);
3449         } else {
3450                 if ((ddb_allocation / plane_blocks_per_line) >= 1)
3451                         selected_result = min(method1, method2);
3452                 else
3453                         selected_result = method1;
3454         }
3455
3456         res_blocks = selected_result + 1;
3457         res_lines = DIV_ROUND_UP(selected_result, plane_blocks_per_line);
3458
3459         if (level >= 1 && level <= 7) {
3460                 if (fb->modifier[0] == I915_FORMAT_MOD_Y_TILED ||
3461                     fb->modifier[0] == I915_FORMAT_MOD_Yf_TILED)
3462                         res_lines += 4;
3463                 else
3464                         res_blocks++;
3465         }
3466
3467         if (res_blocks >= ddb_allocation || res_lines > 31) {
3468                 *enabled = false;
3469
3470                 /*
3471                  * If there are no valid level 0 watermarks, then we can't
3472                  * support this display configuration.
3473                  */
3474                 if (level) {
3475                         return 0;
3476                 } else {
3477                         DRM_DEBUG_KMS("Requested display configuration exceeds system watermark limitations\n");
3478                         DRM_DEBUG_KMS("Plane %d.%d: blocks required = %u/%u, lines required = %u/31\n",
3479                                       to_intel_crtc(cstate->base.crtc)->pipe,
3480                                       skl_wm_plane_id(to_intel_plane(pstate->plane)),
3481                                       res_blocks, ddb_allocation, res_lines);
3482
3483                         return -EINVAL;
3484                 }
3485         }
3486
3487         *out_blocks = res_blocks;
3488         *out_lines = res_lines;
3489         *enabled = true;
3490
3491         return 0;
3492 }
3493
3494 static int
3495 skl_compute_wm_level(const struct drm_i915_private *dev_priv,
3496                      struct skl_ddb_allocation *ddb,
3497                      struct intel_crtc_state *cstate,
3498                      int level,
3499                      struct skl_wm_level *result)
3500 {
3501         struct drm_atomic_state *state = cstate->base.state;
3502         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
3503         struct drm_plane *plane;
3504         struct intel_plane *intel_plane;
3505         struct intel_plane_state *intel_pstate;
3506         uint16_t ddb_blocks;
3507         enum pipe pipe = intel_crtc->pipe;
3508         int ret;
3509
3510         /*
3511          * We'll only calculate watermarks for planes that are actually
3512          * enabled, so make sure all other planes are set as disabled.
3513          */
3514         memset(result, 0, sizeof(*result));
3515
3516         for_each_intel_plane_mask(&dev_priv->drm,
3517                                   intel_plane,
3518                                   cstate->base.plane_mask) {
3519                 int i = skl_wm_plane_id(intel_plane);
3520
3521                 plane = &intel_plane->base;
3522                 intel_pstate = NULL;
3523                 if (state)
3524                         intel_pstate =
3525                                 intel_atomic_get_existing_plane_state(state,
3526                                                                       intel_plane);
3527
3528                 /*
3529                  * Note: If we start supporting multiple pending atomic commits
3530                  * against the same planes/CRTC's in the future, plane->state
3531                  * will no longer be the correct pre-state to use for the
3532                  * calculations here and we'll need to change where we get the
3533                  * 'unchanged' plane data from.
3534                  *
3535                  * For now this is fine because we only allow one queued commit
3536                  * against a CRTC.  Even if the plane isn't modified by this
3537                  * transaction and we don't have a plane lock, we still have
3538                  * the CRTC's lock, so we know that no other transactions are
3539                  * racing with us to update it.
3540                  */
3541                 if (!intel_pstate)
3542                         intel_pstate = to_intel_plane_state(plane->state);
3543
3544                 WARN_ON(!intel_pstate->base.fb);
3545
3546                 ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][i]);
3547
3548                 ret = skl_compute_plane_wm(dev_priv,
3549                                            cstate,
3550                                            intel_pstate,
3551                                            ddb_blocks,
3552                                            level,
3553                                            &result->plane_res_b[i],
3554                                            &result->plane_res_l[i],
3555                                            &result->plane_en[i]);
3556                 if (ret)
3557                         return ret;
3558         }
3559
3560         return 0;
3561 }
3562
3563 static uint32_t
3564 skl_compute_linetime_wm(struct intel_crtc_state *cstate)
3565 {
3566         if (!cstate->base.active)
3567                 return 0;
3568
3569         if (WARN_ON(skl_pipe_pixel_rate(cstate) == 0))
3570                 return 0;
3571
3572         return DIV_ROUND_UP(8 * cstate->base.adjusted_mode.crtc_htotal * 1000,
3573                             skl_pipe_pixel_rate(cstate));
3574 }
3575
3576 static void skl_compute_transition_wm(struct intel_crtc_state *cstate,
3577                                       struct skl_wm_level *trans_wm /* out */)
3578 {
3579         struct drm_crtc *crtc = cstate->base.crtc;
3580         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3581         struct intel_plane *intel_plane;
3582
3583         if (!cstate->base.active)
3584                 return;
3585
3586         /* Until we know more, just disable transition WMs */
3587         for_each_intel_plane_on_crtc(crtc->dev, intel_crtc, intel_plane) {
3588                 int i = skl_wm_plane_id(intel_plane);
3589
3590                 trans_wm->plane_en[i] = false;
3591         }
3592 }
3593
3594 static int skl_build_pipe_wm(struct intel_crtc_state *cstate,
3595                              struct skl_ddb_allocation *ddb,
3596                              struct skl_pipe_wm *pipe_wm)
3597 {
3598         struct drm_device *dev = cstate->base.crtc->dev;
3599         const struct drm_i915_private *dev_priv = to_i915(dev);
3600         int level, max_level = ilk_wm_max_level(dev);
3601         int ret;
3602
3603         for (level = 0; level <= max_level; level++) {
3604                 ret = skl_compute_wm_level(dev_priv, ddb, cstate,
3605                                            level, &pipe_wm->wm[level]);
3606                 if (ret)
3607                         return ret;
3608         }
3609         pipe_wm->linetime = skl_compute_linetime_wm(cstate);
3610
3611         skl_compute_transition_wm(cstate, &pipe_wm->trans_wm);
3612
3613         return 0;
3614 }
3615
3616 static void skl_compute_wm_results(struct drm_device *dev,
3617                                    struct skl_pipe_wm *p_wm,
3618                                    struct skl_wm_values *r,
3619                                    struct intel_crtc *intel_crtc)
3620 {
3621         int level, max_level = ilk_wm_max_level(dev);
3622         enum pipe pipe = intel_crtc->pipe;
3623         uint32_t temp;
3624         int i;
3625
3626         for (level = 0; level <= max_level; level++) {
3627                 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3628                         temp = 0;
3629
3630                         temp |= p_wm->wm[level].plane_res_l[i] <<
3631                                         PLANE_WM_LINES_SHIFT;
3632                         temp |= p_wm->wm[level].plane_res_b[i];
3633                         if (p_wm->wm[level].plane_en[i])
3634                                 temp |= PLANE_WM_EN;
3635
3636                         r->plane[pipe][i][level] = temp;
3637                 }
3638
3639                 temp = 0;
3640
3641                 temp |= p_wm->wm[level].plane_res_l[PLANE_CURSOR] << PLANE_WM_LINES_SHIFT;
3642                 temp |= p_wm->wm[level].plane_res_b[PLANE_CURSOR];
3643
3644                 if (p_wm->wm[level].plane_en[PLANE_CURSOR])
3645                         temp |= PLANE_WM_EN;
3646
3647                 r->plane[pipe][PLANE_CURSOR][level] = temp;
3648
3649         }
3650
3651         /* transition WMs */
3652         for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3653                 temp = 0;
3654                 temp |= p_wm->trans_wm.plane_res_l[i] << PLANE_WM_LINES_SHIFT;
3655                 temp |= p_wm->trans_wm.plane_res_b[i];
3656                 if (p_wm->trans_wm.plane_en[i])
3657                         temp |= PLANE_WM_EN;
3658
3659                 r->plane_trans[pipe][i] = temp;
3660         }
3661
3662         temp = 0;
3663         temp |= p_wm->trans_wm.plane_res_l[PLANE_CURSOR] << PLANE_WM_LINES_SHIFT;
3664         temp |= p_wm->trans_wm.plane_res_b[PLANE_CURSOR];
3665         if (p_wm->trans_wm.plane_en[PLANE_CURSOR])
3666                 temp |= PLANE_WM_EN;
3667
3668         r->plane_trans[pipe][PLANE_CURSOR] = temp;
3669
3670         r->wm_linetime[pipe] = p_wm->linetime;
3671 }
3672
3673 static void skl_ddb_entry_write(struct drm_i915_private *dev_priv,
3674                                 i915_reg_t reg,
3675                                 const struct skl_ddb_entry *entry)
3676 {
3677         if (entry->end)
3678                 I915_WRITE(reg, (entry->end - 1) << 16 | entry->start);
3679         else
3680                 I915_WRITE(reg, 0);
3681 }
3682
3683 static void skl_write_wm_values(struct drm_i915_private *dev_priv,
3684                                 const struct skl_wm_values *new)
3685 {
3686         struct drm_device *dev = &dev_priv->drm;
3687         struct intel_crtc *crtc;
3688
3689         for_each_intel_crtc(dev, crtc) {
3690                 int i, level, max_level = ilk_wm_max_level(dev);
3691                 enum pipe pipe = crtc->pipe;
3692
3693                 if ((new->dirty_pipes & drm_crtc_mask(&crtc->base)) == 0)
3694                         continue;
3695                 if (!crtc->active)
3696                         continue;
3697
3698                 I915_WRITE(PIPE_WM_LINETIME(pipe), new->wm_linetime[pipe]);
3699
3700                 for (level = 0; level <= max_level; level++) {
3701                         for (i = 0; i < intel_num_planes(crtc); i++)
3702                                 I915_WRITE(PLANE_WM(pipe, i, level),
3703                                            new->plane[pipe][i][level]);
3704                         I915_WRITE(CUR_WM(pipe, level),
3705                                    new->plane[pipe][PLANE_CURSOR][level]);
3706                 }
3707                 for (i = 0; i < intel_num_planes(crtc); i++)
3708                         I915_WRITE(PLANE_WM_TRANS(pipe, i),
3709                                    new->plane_trans[pipe][i]);
3710                 I915_WRITE(CUR_WM_TRANS(pipe),
3711                            new->plane_trans[pipe][PLANE_CURSOR]);
3712
3713                 for (i = 0; i < intel_num_planes(crtc); i++) {
3714                         skl_ddb_entry_write(dev_priv,
3715                                             PLANE_BUF_CFG(pipe, i),
3716                                             &new->ddb.plane[pipe][i]);
3717                         skl_ddb_entry_write(dev_priv,
3718                                             PLANE_NV12_BUF_CFG(pipe, i),
3719                                             &new->ddb.y_plane[pipe][i]);
3720                 }
3721
3722                 skl_ddb_entry_write(dev_priv, CUR_BUF_CFG(pipe),
3723                                     &new->ddb.plane[pipe][PLANE_CURSOR]);
3724         }
3725 }
3726
3727 /*
3728  * When setting up a new DDB allocation arrangement, we need to correctly
3729  * sequence the times at which the new allocations for the pipes are taken into
3730  * account or we'll have pipes fetching from space previously allocated to
3731  * another pipe.
3732  *
3733  * Roughly the sequence looks like:
3734  *  1. re-allocate the pipe(s) with the allocation being reduced and not
3735  *     overlapping with a previous light-up pipe (another way to put it is:
3736  *     pipes with their new allocation strickly included into their old ones).
3737  *  2. re-allocate the other pipes that get their allocation reduced
3738  *  3. allocate the pipes having their allocation increased
3739  *
3740  * Steps 1. and 2. are here to take care of the following case:
3741  * - Initially DDB looks like this:
3742  *     |   B    |   C    |
3743  * - enable pipe A.
3744  * - pipe B has a reduced DDB allocation that overlaps with the old pipe C
3745  *   allocation
3746  *     |  A  |  B  |  C  |
3747  *
3748  * We need to sequence the re-allocation: C, B, A (and not B, C, A).
3749  */
3750
3751 static void
3752 skl_wm_flush_pipe(struct drm_i915_private *dev_priv, enum pipe pipe, int pass)
3753 {
3754         int plane;
3755
3756         DRM_DEBUG_KMS("flush pipe %c (pass %d)\n", pipe_name(pipe), pass);
3757
3758         for_each_plane(dev_priv, pipe, plane) {
3759                 I915_WRITE(PLANE_SURF(pipe, plane),
3760                            I915_READ(PLANE_SURF(pipe, plane)));
3761         }
3762         I915_WRITE(CURBASE(pipe), I915_READ(CURBASE(pipe)));
3763 }
3764
3765 static bool
3766 skl_ddb_allocation_included(const struct skl_ddb_allocation *old,
3767                             const struct skl_ddb_allocation *new,
3768                             enum pipe pipe)
3769 {
3770         uint16_t old_size, new_size;
3771
3772         old_size = skl_ddb_entry_size(&old->pipe[pipe]);
3773         new_size = skl_ddb_entry_size(&new->pipe[pipe]);
3774
3775         return old_size != new_size &&
3776                new->pipe[pipe].start >= old->pipe[pipe].start &&
3777                new->pipe[pipe].end <= old->pipe[pipe].end;
3778 }
3779
3780 static void skl_flush_wm_values(struct drm_i915_private *dev_priv,
3781                                 struct skl_wm_values *new_values)
3782 {
3783         struct drm_device *dev = &dev_priv->drm;
3784         struct skl_ddb_allocation *cur_ddb, *new_ddb;
3785         bool reallocated[I915_MAX_PIPES] = {};
3786         struct intel_crtc *crtc;
3787         enum pipe pipe;
3788
3789         new_ddb = &new_values->ddb;
3790         cur_ddb = &dev_priv->wm.skl_hw.ddb;
3791
3792         /*
3793          * First pass: flush the pipes with the new allocation contained into
3794          * the old space.
3795          *
3796          * We'll wait for the vblank on those pipes to ensure we can safely
3797          * re-allocate the freed space without this pipe fetching from it.
3798          */
3799         for_each_intel_crtc(dev, crtc) {
3800                 if (!crtc->active)
3801                         continue;
3802
3803                 pipe = crtc->pipe;
3804
3805                 if (!skl_ddb_allocation_included(cur_ddb, new_ddb, pipe))
3806                         continue;
3807
3808                 skl_wm_flush_pipe(dev_priv, pipe, 1);
3809                 intel_wait_for_vblank(dev, pipe);
3810
3811                 reallocated[pipe] = true;
3812         }
3813
3814
3815         /*
3816          * Second pass: flush the pipes that are having their allocation
3817          * reduced, but overlapping with a previous allocation.
3818          *
3819          * Here as well we need to wait for the vblank to make sure the freed
3820          * space is not used anymore.
3821          */
3822         for_each_intel_crtc(dev, crtc) {
3823                 if (!crtc->active)
3824                         continue;
3825
3826                 pipe = crtc->pipe;
3827
3828                 if (reallocated[pipe])
3829                         continue;
3830
3831                 if (skl_ddb_entry_size(&new_ddb->pipe[pipe]) <
3832                     skl_ddb_entry_size(&cur_ddb->pipe[pipe])) {
3833                         skl_wm_flush_pipe(dev_priv, pipe, 2);
3834                         intel_wait_for_vblank(dev, pipe);
3835                         reallocated[pipe] = true;
3836                 }
3837         }
3838
3839         /*
3840          * Third pass: flush the pipes that got more space allocated.
3841          *
3842          * We don't need to actively wait for the update here, next vblank
3843          * will just get more DDB space with the correct WM values.
3844          */
3845         for_each_intel_crtc(dev, crtc) {
3846                 if (!crtc->active)
3847                         continue;
3848
3849                 pipe = crtc->pipe;
3850
3851                 /*
3852                  * At this point, only the pipes more space than before are
3853                  * left to re-allocate.
3854                  */
3855                 if (reallocated[pipe])
3856                         continue;
3857
3858                 skl_wm_flush_pipe(dev_priv, pipe, 3);
3859         }
3860 }
3861
3862 static int skl_update_pipe_wm(struct drm_crtc_state *cstate,
3863                               struct skl_ddb_allocation *ddb, /* out */
3864                               struct skl_pipe_wm *pipe_wm, /* out */
3865                               bool *changed /* out */)
3866 {
3867         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->crtc);
3868         struct intel_crtc_state *intel_cstate = to_intel_crtc_state(cstate);
3869         int ret;
3870
3871         ret = skl_build_pipe_wm(intel_cstate, ddb, pipe_wm);
3872         if (ret)
3873                 return ret;
3874
3875         if (!memcmp(&intel_crtc->wm.active.skl, pipe_wm, sizeof(*pipe_wm)))
3876                 *changed = false;
3877         else
3878                 *changed = true;
3879
3880         return 0;
3881 }
3882
3883 static uint32_t
3884 pipes_modified(struct drm_atomic_state *state)
3885 {
3886         struct drm_crtc *crtc;
3887         struct drm_crtc_state *cstate;
3888         uint32_t i, ret = 0;
3889
3890         for_each_crtc_in_state(state, crtc, cstate, i)
3891                 ret |= drm_crtc_mask(crtc);
3892
3893         return ret;
3894 }
3895
3896 static int
3897 skl_compute_ddb(struct drm_atomic_state *state)
3898 {
3899         struct drm_device *dev = state->dev;
3900         struct drm_i915_private *dev_priv = to_i915(dev);
3901         struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3902         struct intel_crtc *intel_crtc;
3903         struct skl_ddb_allocation *ddb = &intel_state->wm_results.ddb;
3904         uint32_t realloc_pipes = pipes_modified(state);
3905         int ret;
3906
3907         /*
3908          * If this is our first atomic update following hardware readout,
3909          * we can't trust the DDB that the BIOS programmed for us.  Let's
3910          * pretend that all pipes switched active status so that we'll
3911          * ensure a full DDB recompute.
3912          */
3913         if (dev_priv->wm.distrust_bios_wm)
3914                 intel_state->active_pipe_changes = ~0;
3915
3916         /*
3917          * If the modeset changes which CRTC's are active, we need to
3918          * recompute the DDB allocation for *all* active pipes, even
3919          * those that weren't otherwise being modified in any way by this
3920          * atomic commit.  Due to the shrinking of the per-pipe allocations
3921          * when new active CRTC's are added, it's possible for a pipe that
3922          * we were already using and aren't changing at all here to suddenly
3923          * become invalid if its DDB needs exceeds its new allocation.
3924          *
3925          * Note that if we wind up doing a full DDB recompute, we can't let
3926          * any other display updates race with this transaction, so we need
3927          * to grab the lock on *all* CRTC's.
3928          */
3929         if (intel_state->active_pipe_changes) {
3930                 realloc_pipes = ~0;
3931                 intel_state->wm_results.dirty_pipes = ~0;
3932         }
3933
3934         for_each_intel_crtc_mask(dev, intel_crtc, realloc_pipes) {
3935                 struct intel_crtc_state *cstate;
3936
3937                 cstate = intel_atomic_get_crtc_state(state, intel_crtc);
3938                 if (IS_ERR(cstate))
3939                         return PTR_ERR(cstate);
3940
3941                 ret = skl_allocate_pipe_ddb(cstate, ddb);
3942                 if (ret)
3943                         return ret;
3944         }
3945
3946         return 0;
3947 }
3948
3949 static int
3950 skl_compute_wm(struct drm_atomic_state *state)
3951 {
3952         struct drm_crtc *crtc;
3953         struct drm_crtc_state *cstate;
3954         struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3955         struct skl_wm_values *results = &intel_state->wm_results;
3956         struct skl_pipe_wm *pipe_wm;
3957         bool changed = false;
3958         int ret, i;
3959
3960         /*
3961          * If this transaction isn't actually touching any CRTC's, don't
3962          * bother with watermark calculation.  Note that if we pass this
3963          * test, we're guaranteed to hold at least one CRTC state mutex,
3964          * which means we can safely use values like dev_priv->active_crtcs
3965          * since any racing commits that want to update them would need to
3966          * hold _all_ CRTC state mutexes.
3967          */
3968         for_each_crtc_in_state(state, crtc, cstate, i)
3969                 changed = true;
3970         if (!changed)
3971                 return 0;
3972
3973         /* Clear all dirty flags */
3974         results->dirty_pipes = 0;
3975
3976         ret = skl_compute_ddb(state);
3977         if (ret)
3978                 return ret;
3979
3980         /*
3981          * Calculate WM's for all pipes that are part of this transaction.
3982          * Note that the DDB allocation above may have added more CRTC's that
3983          * weren't otherwise being modified (and set bits in dirty_pipes) if
3984          * pipe allocations had to change.
3985          *
3986          * FIXME:  Now that we're doing this in the atomic check phase, we
3987          * should allow skl_update_pipe_wm() to return failure in cases where
3988          * no suitable watermark values can be found.
3989          */
3990         for_each_crtc_in_state(state, crtc, cstate, i) {
3991                 struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3992                 struct intel_crtc_state *intel_cstate =
3993                         to_intel_crtc_state(cstate);
3994
3995                 pipe_wm = &intel_cstate->wm.skl.optimal;
3996                 ret = skl_update_pipe_wm(cstate, &results->ddb, pipe_wm,
3997                                          &changed);
3998                 if (ret)
3999                         return ret;
4000
4001                 if (changed)
4002                         results->dirty_pipes |= drm_crtc_mask(crtc);
4003
4004                 if ((results->dirty_pipes & drm_crtc_mask(crtc)) == 0)
4005                         /* This pipe's WM's did not change */
4006                         continue;
4007
4008                 intel_cstate->update_wm_pre = true;
4009                 skl_compute_wm_results(crtc->dev, pipe_wm, results, intel_crtc);
4010         }
4011
4012         return 0;
4013 }
4014
4015 static void skl_update_wm(struct drm_crtc *crtc)
4016 {
4017         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4018         struct drm_device *dev = crtc->dev;
4019         struct drm_i915_private *dev_priv = to_i915(dev);
4020         struct skl_wm_values *results = &dev_priv->wm.skl_results;
4021         struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
4022         struct skl_pipe_wm *pipe_wm = &cstate->wm.skl.optimal;
4023
4024         if ((results->dirty_pipes & drm_crtc_mask(crtc)) == 0)
4025                 return;
4026
4027         intel_crtc->wm.active.skl = *pipe_wm;
4028
4029         mutex_lock(&dev_priv->wm.wm_mutex);
4030
4031         skl_write_wm_values(dev_priv, results);
4032         skl_flush_wm_values(dev_priv, results);
4033
4034         /* store the new configuration */
4035         dev_priv->wm.skl_hw = *results;
4036
4037         mutex_unlock(&dev_priv->wm.wm_mutex);
4038 }
4039
4040 static void ilk_compute_wm_config(struct drm_device *dev,
4041                                   struct intel_wm_config *config)
4042 {
4043         struct intel_crtc *crtc;
4044
4045         /* Compute the currently _active_ config */
4046         for_each_intel_crtc(dev, crtc) {
4047                 const struct intel_pipe_wm *wm = &crtc->wm.active.ilk;
4048
4049                 if (!wm->pipe_enabled)
4050                         continue;
4051
4052                 config->sprites_enabled |= wm->sprites_enabled;
4053                 config->sprites_scaled |= wm->sprites_scaled;
4054                 config->num_pipes_active++;
4055         }
4056 }
4057
4058 static void ilk_program_watermarks(struct drm_i915_private *dev_priv)
4059 {
4060         struct drm_device *dev = &dev_priv->drm;
4061         struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
4062         struct ilk_wm_maximums max;
4063         struct intel_wm_config config = {};
4064         struct ilk_wm_values results = {};
4065         enum intel_ddb_partitioning partitioning;
4066
4067         ilk_compute_wm_config(dev, &config);
4068
4069         ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
4070         ilk_wm_merge(dev, &config, &max, &lp_wm_1_2);
4071
4072         /* 5/6 split only in single pipe config on IVB+ */
4073         if (INTEL_INFO(dev)->gen >= 7 &&
4074             config.num_pipes_active == 1 && config.sprites_enabled) {
4075                 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
4076                 ilk_wm_merge(dev, &config, &max, &lp_wm_5_6);
4077
4078                 best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
4079         } else {
4080                 best_lp_wm = &lp_wm_1_2;
4081         }
4082
4083         partitioning = (best_lp_wm == &lp_wm_1_2) ?
4084                        INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
4085
4086         ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
4087
4088         ilk_write_wm_values(dev_priv, &results);
4089 }
4090
4091 static void ilk_initial_watermarks(struct intel_crtc_state *cstate)
4092 {
4093         struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
4094         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4095
4096         mutex_lock(&dev_priv->wm.wm_mutex);
4097         intel_crtc->wm.active.ilk = cstate->wm.ilk.intermediate;
4098         ilk_program_watermarks(dev_priv);
4099         mutex_unlock(&dev_priv->wm.wm_mutex);
4100 }
4101
4102 static void ilk_optimize_watermarks(struct intel_crtc_state *cstate)
4103 {
4104         struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
4105         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4106
4107         mutex_lock(&dev_priv->wm.wm_mutex);
4108         if (cstate->wm.need_postvbl_update) {
4109                 intel_crtc->wm.active.ilk = cstate->wm.ilk.optimal;
4110                 ilk_program_watermarks(dev_priv);
4111         }
4112         mutex_unlock(&dev_priv->wm.wm_mutex);
4113 }
4114
4115 static void skl_pipe_wm_active_state(uint32_t val,
4116                                      struct skl_pipe_wm *active,
4117                                      bool is_transwm,
4118                                      bool is_cursor,
4119                                      int i,
4120                                      int level)
4121 {
4122         bool is_enabled = (val & PLANE_WM_EN) != 0;
4123
4124         if (!is_transwm) {
4125                 if (!is_cursor) {
4126                         active->wm[level].plane_en[i] = is_enabled;
4127                         active->wm[level].plane_res_b[i] =
4128                                         val & PLANE_WM_BLOCKS_MASK;
4129                         active->wm[level].plane_res_l[i] =
4130                                         (val >> PLANE_WM_LINES_SHIFT) &
4131                                                 PLANE_WM_LINES_MASK;
4132                 } else {
4133                         active->wm[level].plane_en[PLANE_CURSOR] = is_enabled;
4134                         active->wm[level].plane_res_b[PLANE_CURSOR] =
4135                                         val & PLANE_WM_BLOCKS_MASK;
4136                         active->wm[level].plane_res_l[PLANE_CURSOR] =
4137                                         (val >> PLANE_WM_LINES_SHIFT) &
4138                                                 PLANE_WM_LINES_MASK;
4139                 }
4140         } else {
4141                 if (!is_cursor) {
4142                         active->trans_wm.plane_en[i] = is_enabled;
4143                         active->trans_wm.plane_res_b[i] =
4144                                         val & PLANE_WM_BLOCKS_MASK;
4145                         active->trans_wm.plane_res_l[i] =
4146                                         (val >> PLANE_WM_LINES_SHIFT) &
4147                                                 PLANE_WM_LINES_MASK;
4148                 } else {
4149                         active->trans_wm.plane_en[PLANE_CURSOR] = is_enabled;
4150                         active->trans_wm.plane_res_b[PLANE_CURSOR] =
4151                                         val & PLANE_WM_BLOCKS_MASK;
4152                         active->trans_wm.plane_res_l[PLANE_CURSOR] =
4153                                         (val >> PLANE_WM_LINES_SHIFT) &
4154                                                 PLANE_WM_LINES_MASK;
4155                 }
4156         }
4157 }
4158
4159 static void skl_pipe_wm_get_hw_state(struct drm_crtc *crtc)
4160 {
4161         struct drm_device *dev = crtc->dev;
4162         struct drm_i915_private *dev_priv = to_i915(dev);
4163         struct skl_wm_values *hw = &dev_priv->wm.skl_hw;
4164         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4165         struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
4166         struct skl_pipe_wm *active = &cstate->wm.skl.optimal;
4167         enum pipe pipe = intel_crtc->pipe;
4168         int level, i, max_level;
4169         uint32_t temp;
4170
4171         max_level = ilk_wm_max_level(dev);
4172
4173         hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
4174
4175         for (level = 0; level <= max_level; level++) {
4176                 for (i = 0; i < intel_num_planes(intel_crtc); i++)
4177                         hw->plane[pipe][i][level] =
4178                                         I915_READ(PLANE_WM(pipe, i, level));
4179                 hw->plane[pipe][PLANE_CURSOR][level] = I915_READ(CUR_WM(pipe, level));
4180         }
4181
4182         for (i = 0; i < intel_num_planes(intel_crtc); i++)
4183                 hw->plane_trans[pipe][i] = I915_READ(PLANE_WM_TRANS(pipe, i));
4184         hw->plane_trans[pipe][PLANE_CURSOR] = I915_READ(CUR_WM_TRANS(pipe));
4185
4186         if (!intel_crtc->active)
4187                 return;
4188
4189         hw->dirty_pipes |= drm_crtc_mask(crtc);
4190
4191         active->linetime = hw->wm_linetime[pipe];
4192
4193         for (level = 0; level <= max_level; level++) {
4194                 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
4195                         temp = hw->plane[pipe][i][level];
4196                         skl_pipe_wm_active_state(temp, active, false,
4197                                                 false, i, level);
4198                 }
4199                 temp = hw->plane[pipe][PLANE_CURSOR][level];
4200                 skl_pipe_wm_active_state(temp, active, false, true, i, level);
4201         }
4202
4203         for (i = 0; i < intel_num_planes(intel_crtc); i++) {
4204                 temp = hw->plane_trans[pipe][i];
4205                 skl_pipe_wm_active_state(temp, active, true, false, i, 0);
4206         }
4207
4208         temp = hw->plane_trans[pipe][PLANE_CURSOR];
4209         skl_pipe_wm_active_state(temp, active, true, true, i, 0);
4210
4211         intel_crtc->wm.active.skl = *active;
4212 }
4213
4214 void skl_wm_get_hw_state(struct drm_device *dev)
4215 {
4216         struct drm_i915_private *dev_priv = to_i915(dev);
4217         struct skl_ddb_allocation *ddb = &dev_priv->wm.skl_hw.ddb;
4218         struct drm_crtc *crtc;
4219
4220         skl_ddb_get_hw_state(dev_priv, ddb);
4221         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
4222                 skl_pipe_wm_get_hw_state(crtc);
4223
4224         if (dev_priv->active_crtcs) {
4225                 /* Fully recompute DDB on first atomic commit */
4226                 dev_priv->wm.distrust_bios_wm = true;
4227         } else {
4228                 /* Easy/common case; just sanitize DDB now if everything off */
4229                 memset(ddb, 0, sizeof(*ddb));
4230         }
4231 }
4232
4233 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
4234 {
4235         struct drm_device *dev = crtc->dev;
4236         struct drm_i915_private *dev_priv = to_i915(dev);
4237         struct ilk_wm_values *hw = &dev_priv->wm.hw;
4238         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4239         struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
4240         struct intel_pipe_wm *active = &cstate->wm.ilk.optimal;
4241         enum pipe pipe = intel_crtc->pipe;
4242         static const i915_reg_t wm0_pipe_reg[] = {
4243                 [PIPE_A] = WM0_PIPEA_ILK,
4244                 [PIPE_B] = WM0_PIPEB_ILK,
4245                 [PIPE_C] = WM0_PIPEC_IVB,
4246         };
4247
4248         hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
4249         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
4250                 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
4251
4252         memset(active, 0, sizeof(*active));
4253
4254         active->pipe_enabled = intel_crtc->active;
4255
4256         if (active->pipe_enabled) {
4257                 u32 tmp = hw->wm_pipe[pipe];
4258
4259                 /*
4260                  * For active pipes LP0 watermark is marked as
4261                  * enabled, and LP1+ watermaks as disabled since
4262                  * we can't really reverse compute them in case
4263                  * multiple pipes are active.
4264                  */
4265                 active->wm[0].enable = true;
4266                 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
4267                 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
4268                 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
4269                 active->linetime = hw->wm_linetime[pipe];
4270         } else {
4271                 int level, max_level = ilk_wm_max_level(dev);
4272
4273                 /*
4274                  * For inactive pipes, all watermark levels
4275                  * should be marked as enabled but zeroed,
4276                  * which is what we'd compute them to.
4277                  */
4278                 for (level = 0; level <= max_level; level++)
4279                         active->wm[level].enable = true;
4280         }
4281
4282         intel_crtc->wm.active.ilk = *active;
4283 }
4284
4285 #define _FW_WM(value, plane) \
4286         (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
4287 #define _FW_WM_VLV(value, plane) \
4288         (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
4289
4290 static void vlv_read_wm_values(struct drm_i915_private *dev_priv,
4291                                struct vlv_wm_values *wm)
4292 {
4293         enum pipe pipe;
4294         uint32_t tmp;
4295
4296         for_each_pipe(dev_priv, pipe) {
4297                 tmp = I915_READ(VLV_DDL(pipe));
4298
4299                 wm->ddl[pipe].primary =
4300                         (tmp >> DDL_PLANE_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4301                 wm->ddl[pipe].cursor =
4302                         (tmp >> DDL_CURSOR_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4303                 wm->ddl[pipe].sprite[0] =
4304                         (tmp >> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4305                 wm->ddl[pipe].sprite[1] =
4306                         (tmp >> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4307         }
4308
4309         tmp = I915_READ(DSPFW1);
4310         wm->sr.plane = _FW_WM(tmp, SR);
4311         wm->pipe[PIPE_B].cursor = _FW_WM(tmp, CURSORB);
4312         wm->pipe[PIPE_B].primary = _FW_WM_VLV(tmp, PLANEB);
4313         wm->pipe[PIPE_A].primary = _FW_WM_VLV(tmp, PLANEA);
4314
4315         tmp = I915_READ(DSPFW2);
4316         wm->pipe[PIPE_A].sprite[1] = _FW_WM_VLV(tmp, SPRITEB);
4317         wm->pipe[PIPE_A].cursor = _FW_WM(tmp, CURSORA);
4318         wm->pipe[PIPE_A].sprite[0] = _FW_WM_VLV(tmp, SPRITEA);
4319
4320         tmp = I915_READ(DSPFW3);
4321         wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
4322
4323         if (IS_CHERRYVIEW(dev_priv)) {
4324                 tmp = I915_READ(DSPFW7_CHV);
4325                 wm->pipe[PIPE_B].sprite[1] = _FW_WM_VLV(tmp, SPRITED);
4326                 wm->pipe[PIPE_B].sprite[0] = _FW_WM_VLV(tmp, SPRITEC);
4327
4328                 tmp = I915_READ(DSPFW8_CHV);
4329                 wm->pipe[PIPE_C].sprite[1] = _FW_WM_VLV(tmp, SPRITEF);
4330                 wm->pipe[PIPE_C].sprite[0] = _FW_WM_VLV(tmp, SPRITEE);
4331
4332                 tmp = I915_READ(DSPFW9_CHV);
4333                 wm->pipe[PIPE_C].primary = _FW_WM_VLV(tmp, PLANEC);
4334                 wm->pipe[PIPE_C].cursor = _FW_WM(tmp, CURSORC);
4335
4336                 tmp = I915_READ(DSPHOWM);
4337                 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
4338                 wm->pipe[PIPE_C].sprite[1] |= _FW_WM(tmp, SPRITEF_HI) << 8;
4339                 wm->pipe[PIPE_C].sprite[0] |= _FW_WM(tmp, SPRITEE_HI) << 8;
4340                 wm->pipe[PIPE_C].primary |= _FW_WM(tmp, PLANEC_HI) << 8;
4341                 wm->pipe[PIPE_B].sprite[1] |= _FW_WM(tmp, SPRITED_HI) << 8;
4342                 wm->pipe[PIPE_B].sprite[0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
4343                 wm->pipe[PIPE_B].primary |= _FW_WM(tmp, PLANEB_HI) << 8;
4344                 wm->pipe[PIPE_A].sprite[1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
4345                 wm->pipe[PIPE_A].sprite[0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
4346                 wm->pipe[PIPE_A].primary |= _FW_WM(tmp, PLANEA_HI) << 8;
4347         } else {
4348                 tmp = I915_READ(DSPFW7);
4349                 wm->pipe[PIPE_B].sprite[1] = _FW_WM_VLV(tmp, SPRITED);
4350                 wm->pipe[PIPE_B].sprite[0] = _FW_WM_VLV(tmp, SPRITEC);
4351
4352                 tmp = I915_READ(DSPHOWM);
4353                 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
4354                 wm->pipe[PIPE_B].sprite[1] |= _FW_WM(tmp, SPRITED_HI) << 8;
4355                 wm->pipe[PIPE_B].sprite[0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
4356                 wm->pipe[PIPE_B].primary |= _FW_WM(tmp, PLANEB_HI) << 8;
4357                 wm->pipe[PIPE_A].sprite[1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
4358                 wm->pipe[PIPE_A].sprite[0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
4359                 wm->pipe[PIPE_A].primary |= _FW_WM(tmp, PLANEA_HI) << 8;
4360         }
4361 }
4362
4363 #undef _FW_WM
4364 #undef _FW_WM_VLV
4365
4366 void vlv_wm_get_hw_state(struct drm_device *dev)
4367 {
4368         struct drm_i915_private *dev_priv = to_i915(dev);
4369         struct vlv_wm_values *wm = &dev_priv->wm.vlv;
4370         struct intel_plane *plane;
4371         enum pipe pipe;
4372         u32 val;
4373
4374         vlv_read_wm_values(dev_priv, wm);
4375
4376         for_each_intel_plane(dev, plane) {
4377                 switch (plane->base.type) {
4378                         int sprite;
4379                 case DRM_PLANE_TYPE_CURSOR:
4380                         plane->wm.fifo_size = 63;
4381                         break;
4382                 case DRM_PLANE_TYPE_PRIMARY:
4383                         plane->wm.fifo_size = vlv_get_fifo_size(dev, plane->pipe, 0);
4384                         break;
4385                 case DRM_PLANE_TYPE_OVERLAY:
4386                         sprite = plane->plane;
4387                         plane->wm.fifo_size = vlv_get_fifo_size(dev, plane->pipe, sprite + 1);
4388                         break;
4389                 }
4390         }
4391
4392         wm->cxsr = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
4393         wm->level = VLV_WM_LEVEL_PM2;
4394
4395         if (IS_CHERRYVIEW(dev_priv)) {
4396                 mutex_lock(&dev_priv->rps.hw_lock);
4397
4398                 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
4399                 if (val & DSP_MAXFIFO_PM5_ENABLE)
4400                         wm->level = VLV_WM_LEVEL_PM5;
4401
4402                 /*
4403                  * If DDR DVFS is disabled in the BIOS, Punit
4404                  * will never ack the request. So if that happens
4405                  * assume we don't have to enable/disable DDR DVFS
4406                  * dynamically. To test that just set the REQ_ACK
4407                  * bit to poke the Punit, but don't change the
4408                  * HIGH/LOW bits so that we don't actually change
4409                  * the current state.
4410                  */
4411                 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
4412                 val |= FORCE_DDR_FREQ_REQ_ACK;
4413                 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
4414
4415                 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
4416                               FORCE_DDR_FREQ_REQ_ACK) == 0, 3)) {
4417                         DRM_DEBUG_KMS("Punit not acking DDR DVFS request, "
4418                                       "assuming DDR DVFS is disabled\n");
4419                         dev_priv->wm.max_level = VLV_WM_LEVEL_PM5;
4420                 } else {
4421                         val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
4422                         if ((val & FORCE_DDR_HIGH_FREQ) == 0)
4423                                 wm->level = VLV_WM_LEVEL_DDR_DVFS;
4424                 }
4425
4426                 mutex_unlock(&dev_priv->rps.hw_lock);
4427         }
4428
4429         for_each_pipe(dev_priv, pipe)
4430                 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
4431                               pipe_name(pipe), wm->pipe[pipe].primary, wm->pipe[pipe].cursor,
4432                               wm->pipe[pipe].sprite[0], wm->pipe[pipe].sprite[1]);
4433
4434         DRM_DEBUG_KMS("Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
4435                       wm->sr.plane, wm->sr.cursor, wm->level, wm->cxsr);
4436 }
4437
4438 void ilk_wm_get_hw_state(struct drm_device *dev)
4439 {
4440         struct drm_i915_private *dev_priv = to_i915(dev);
4441         struct ilk_wm_values *hw = &dev_priv->wm.hw;
4442         struct drm_crtc *crtc;
4443
4444         for_each_crtc(dev, crtc)
4445                 ilk_pipe_wm_get_hw_state(crtc);
4446
4447         hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
4448         hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
4449         hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
4450
4451         hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
4452         if (INTEL_INFO(dev)->gen >= 7) {
4453                 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
4454                 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
4455         }
4456
4457         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
4458                 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
4459                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4460         else if (IS_IVYBRIDGE(dev))
4461                 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
4462                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4463
4464         hw->enable_fbc_wm =
4465                 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
4466 }
4467
4468 /**
4469  * intel_update_watermarks - update FIFO watermark values based on current modes
4470  *
4471  * Calculate watermark values for the various WM regs based on current mode
4472  * and plane configuration.
4473  *
4474  * There are several cases to deal with here:
4475  *   - normal (i.e. non-self-refresh)
4476  *   - self-refresh (SR) mode
4477  *   - lines are large relative to FIFO size (buffer can hold up to 2)
4478  *   - lines are small relative to FIFO size (buffer can hold more than 2
4479  *     lines), so need to account for TLB latency
4480  *
4481  *   The normal calculation is:
4482  *     watermark = dotclock * bytes per pixel * latency
4483  *   where latency is platform & configuration dependent (we assume pessimal
4484  *   values here).
4485  *
4486  *   The SR calculation is:
4487  *     watermark = (trunc(latency/line time)+1) * surface width *
4488  *       bytes per pixel
4489  *   where
4490  *     line time = htotal / dotclock
4491  *     surface width = hdisplay for normal plane and 64 for cursor
4492  *   and latency is assumed to be high, as above.
4493  *
4494  * The final value programmed to the register should always be rounded up,
4495  * and include an extra 2 entries to account for clock crossings.
4496  *
4497  * We don't use the sprite, so we can ignore that.  And on Crestline we have
4498  * to set the non-SR watermarks to 8.
4499  */
4500 void intel_update_watermarks(struct drm_crtc *crtc)
4501 {
4502         struct drm_i915_private *dev_priv = to_i915(crtc->dev);
4503
4504         if (dev_priv->display.update_wm)
4505                 dev_priv->display.update_wm(crtc);
4506 }
4507
4508 /*
4509  * Lock protecting IPS related data structures
4510  */
4511 DEFINE_SPINLOCK(mchdev_lock);
4512
4513 /* Global for IPS driver to get at the current i915 device. Protected by
4514  * mchdev_lock. */
4515 static struct drm_i915_private *i915_mch_dev;
4516
4517 bool ironlake_set_drps(struct drm_i915_private *dev_priv, u8 val)
4518 {
4519         u16 rgvswctl;
4520
4521         assert_spin_locked(&mchdev_lock);
4522
4523         rgvswctl = I915_READ16(MEMSWCTL);
4524         if (rgvswctl & MEMCTL_CMD_STS) {
4525                 DRM_DEBUG("gpu busy, RCS change rejected\n");
4526                 return false; /* still busy with another command */
4527         }
4528
4529         rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
4530                 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
4531         I915_WRITE16(MEMSWCTL, rgvswctl);
4532         POSTING_READ16(MEMSWCTL);
4533
4534         rgvswctl |= MEMCTL_CMD_STS;
4535         I915_WRITE16(MEMSWCTL, rgvswctl);
4536
4537         return true;
4538 }
4539
4540 static void ironlake_enable_drps(struct drm_i915_private *dev_priv)
4541 {
4542         u32 rgvmodectl;
4543         u8 fmax, fmin, fstart, vstart;
4544
4545         spin_lock_irq(&mchdev_lock);
4546
4547         rgvmodectl = I915_READ(MEMMODECTL);
4548
4549         /* Enable temp reporting */
4550         I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
4551         I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
4552
4553         /* 100ms RC evaluation intervals */
4554         I915_WRITE(RCUPEI, 100000);
4555         I915_WRITE(RCDNEI, 100000);
4556
4557         /* Set max/min thresholds to 90ms and 80ms respectively */
4558         I915_WRITE(RCBMAXAVG, 90000);
4559         I915_WRITE(RCBMINAVG, 80000);
4560
4561         I915_WRITE(MEMIHYST, 1);
4562
4563         /* Set up min, max, and cur for interrupt handling */
4564         fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
4565         fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
4566         fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
4567                 MEMMODE_FSTART_SHIFT;
4568
4569         vstart = (I915_READ(PXVFREQ(fstart)) & PXVFREQ_PX_MASK) >>
4570                 PXVFREQ_PX_SHIFT;
4571
4572         dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
4573         dev_priv->ips.fstart = fstart;
4574
4575         dev_priv->ips.max_delay = fstart;
4576         dev_priv->ips.min_delay = fmin;
4577         dev_priv->ips.cur_delay = fstart;
4578
4579         DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
4580                          fmax, fmin, fstart);
4581
4582         I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
4583
4584         /*
4585          * Interrupts will be enabled in ironlake_irq_postinstall
4586          */
4587
4588         I915_WRITE(VIDSTART, vstart);
4589         POSTING_READ(VIDSTART);
4590
4591         rgvmodectl |= MEMMODE_SWMODE_EN;
4592         I915_WRITE(MEMMODECTL, rgvmodectl);
4593
4594         if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
4595                 DRM_ERROR("stuck trying to change perf mode\n");
4596         mdelay(1);
4597
4598         ironlake_set_drps(dev_priv, fstart);
4599
4600         dev_priv->ips.last_count1 = I915_READ(DMIEC) +
4601                 I915_READ(DDREC) + I915_READ(CSIEC);
4602         dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
4603         dev_priv->ips.last_count2 = I915_READ(GFXEC);
4604         dev_priv->ips.last_time2 = ktime_get_raw_ns();
4605
4606         spin_unlock_irq(&mchdev_lock);
4607 }
4608
4609 static void ironlake_disable_drps(struct drm_i915_private *dev_priv)
4610 {
4611         u16 rgvswctl;
4612
4613         spin_lock_irq(&mchdev_lock);
4614
4615         rgvswctl = I915_READ16(MEMSWCTL);
4616
4617         /* Ack interrupts, disable EFC interrupt */
4618         I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
4619         I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
4620         I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
4621         I915_WRITE(DEIIR, DE_PCU_EVENT);
4622         I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
4623
4624         /* Go back to the starting frequency */
4625         ironlake_set_drps(dev_priv, dev_priv->ips.fstart);
4626         mdelay(1);
4627         rgvswctl |= MEMCTL_CMD_STS;
4628         I915_WRITE(MEMSWCTL, rgvswctl);
4629         mdelay(1);
4630
4631         spin_unlock_irq(&mchdev_lock);
4632 }
4633
4634 /* There's a funny hw issue where the hw returns all 0 when reading from
4635  * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
4636  * ourselves, instead of doing a rmw cycle (which might result in us clearing
4637  * all limits and the gpu stuck at whatever frequency it is at atm).
4638  */
4639 static u32 intel_rps_limits(struct drm_i915_private *dev_priv, u8 val)
4640 {
4641         u32 limits;
4642
4643         /* Only set the down limit when we've reached the lowest level to avoid
4644          * getting more interrupts, otherwise leave this clear. This prevents a
4645          * race in the hw when coming out of rc6: There's a tiny window where
4646          * the hw runs at the minimal clock before selecting the desired
4647          * frequency, if the down threshold expires in that window we will not
4648          * receive a down interrupt. */
4649         if (IS_GEN9(dev_priv)) {
4650                 limits = (dev_priv->rps.max_freq_softlimit) << 23;
4651                 if (val <= dev_priv->rps.min_freq_softlimit)
4652                         limits |= (dev_priv->rps.min_freq_softlimit) << 14;
4653         } else {
4654                 limits = dev_priv->rps.max_freq_softlimit << 24;
4655                 if (val <= dev_priv->rps.min_freq_softlimit)
4656                         limits |= dev_priv->rps.min_freq_softlimit << 16;
4657         }
4658
4659         return limits;
4660 }
4661
4662 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
4663 {
4664         int new_power;
4665         u32 threshold_up = 0, threshold_down = 0; /* in % */
4666         u32 ei_up = 0, ei_down = 0;
4667
4668         new_power = dev_priv->rps.power;
4669         switch (dev_priv->rps.power) {
4670         case LOW_POWER:
4671                 if (val > dev_priv->rps.efficient_freq + 1 &&
4672                     val > dev_priv->rps.cur_freq)
4673                         new_power = BETWEEN;
4674                 break;
4675
4676         case BETWEEN:
4677                 if (val <= dev_priv->rps.efficient_freq &&
4678                     val < dev_priv->rps.cur_freq)
4679                         new_power = LOW_POWER;
4680                 else if (val >= dev_priv->rps.rp0_freq &&
4681                          val > dev_priv->rps.cur_freq)
4682                         new_power = HIGH_POWER;
4683                 break;
4684
4685         case HIGH_POWER:
4686                 if (val < (dev_priv->rps.rp1_freq + dev_priv->rps.rp0_freq) >> 1 &&
4687                     val < dev_priv->rps.cur_freq)
4688                         new_power = BETWEEN;
4689                 break;
4690         }
4691         /* Max/min bins are special */
4692         if (val <= dev_priv->rps.min_freq_softlimit)
4693                 new_power = LOW_POWER;
4694         if (val >= dev_priv->rps.max_freq_softlimit)
4695                 new_power = HIGH_POWER;
4696         if (new_power == dev_priv->rps.power)
4697                 return;
4698
4699         /* Note the units here are not exactly 1us, but 1280ns. */
4700         switch (new_power) {
4701         case LOW_POWER:
4702                 /* Upclock if more than 95% busy over 16ms */
4703                 ei_up = 16000;
4704                 threshold_up = 95;
4705
4706                 /* Downclock if less than 85% busy over 32ms */
4707                 ei_down = 32000;
4708                 threshold_down = 85;
4709                 break;
4710
4711         case BETWEEN:
4712                 /* Upclock if more than 90% busy over 13ms */
4713                 ei_up = 13000;
4714                 threshold_up = 90;
4715
4716                 /* Downclock if less than 75% busy over 32ms */
4717                 ei_down = 32000;
4718                 threshold_down = 75;
4719                 break;
4720
4721         case HIGH_POWER:
4722                 /* Upclock if more than 85% busy over 10ms */
4723                 ei_up = 10000;
4724                 threshold_up = 85;
4725
4726                 /* Downclock if less than 60% busy over 32ms */
4727                 ei_down = 32000;
4728                 threshold_down = 60;
4729                 break;
4730         }
4731
4732         I915_WRITE(GEN6_RP_UP_EI,
4733                    GT_INTERVAL_FROM_US(dev_priv, ei_up));
4734         I915_WRITE(GEN6_RP_UP_THRESHOLD,
4735                    GT_INTERVAL_FROM_US(dev_priv,
4736                                        ei_up * threshold_up / 100));
4737
4738         I915_WRITE(GEN6_RP_DOWN_EI,
4739                    GT_INTERVAL_FROM_US(dev_priv, ei_down));
4740         I915_WRITE(GEN6_RP_DOWN_THRESHOLD,
4741                    GT_INTERVAL_FROM_US(dev_priv,
4742                                        ei_down * threshold_down / 100));
4743
4744         I915_WRITE(GEN6_RP_CONTROL,
4745                    GEN6_RP_MEDIA_TURBO |
4746                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
4747                    GEN6_RP_MEDIA_IS_GFX |
4748                    GEN6_RP_ENABLE |
4749                    GEN6_RP_UP_BUSY_AVG |
4750                    GEN6_RP_DOWN_IDLE_AVG);
4751
4752         dev_priv->rps.power = new_power;
4753         dev_priv->rps.up_threshold = threshold_up;
4754         dev_priv->rps.down_threshold = threshold_down;
4755         dev_priv->rps.last_adj = 0;
4756 }
4757
4758 static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
4759 {
4760         u32 mask = 0;
4761
4762         if (val > dev_priv->rps.min_freq_softlimit)
4763                 mask |= GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
4764         if (val < dev_priv->rps.max_freq_softlimit)
4765                 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD;
4766
4767         mask &= dev_priv->pm_rps_events;
4768
4769         return gen6_sanitize_rps_pm_mask(dev_priv, ~mask);
4770 }
4771
4772 /* gen6_set_rps is called to update the frequency request, but should also be
4773  * called when the range (min_delay and max_delay) is modified so that we can
4774  * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
4775 static void gen6_set_rps(struct drm_i915_private *dev_priv, u8 val)
4776 {
4777         /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
4778         if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1))
4779                 return;
4780
4781         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4782         WARN_ON(val > dev_priv->rps.max_freq);
4783         WARN_ON(val < dev_priv->rps.min_freq);
4784
4785         /* min/max delay may still have been modified so be sure to
4786          * write the limits value.
4787          */
4788         if (val != dev_priv->rps.cur_freq) {
4789                 gen6_set_rps_thresholds(dev_priv, val);
4790
4791                 if (IS_GEN9(dev_priv))
4792                         I915_WRITE(GEN6_RPNSWREQ,
4793                                    GEN9_FREQUENCY(val));
4794                 else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
4795                         I915_WRITE(GEN6_RPNSWREQ,
4796                                    HSW_FREQUENCY(val));
4797                 else
4798                         I915_WRITE(GEN6_RPNSWREQ,
4799                                    GEN6_FREQUENCY(val) |
4800                                    GEN6_OFFSET(0) |
4801                                    GEN6_AGGRESSIVE_TURBO);
4802         }
4803
4804         /* Make sure we continue to get interrupts
4805          * until we hit the minimum or maximum frequencies.
4806          */
4807         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, intel_rps_limits(dev_priv, val));
4808         I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
4809
4810         POSTING_READ(GEN6_RPNSWREQ);
4811
4812         dev_priv->rps.cur_freq = val;
4813         trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
4814 }
4815
4816 static void valleyview_set_rps(struct drm_i915_private *dev_priv, u8 val)
4817 {
4818         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4819         WARN_ON(val > dev_priv->rps.max_freq);
4820         WARN_ON(val < dev_priv->rps.min_freq);
4821
4822         if (WARN_ONCE(IS_CHERRYVIEW(dev_priv) && (val & 1),
4823                       "Odd GPU freq value\n"))
4824                 val &= ~1;
4825
4826         I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
4827
4828         if (val != dev_priv->rps.cur_freq) {
4829                 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
4830                 if (!IS_CHERRYVIEW(dev_priv))
4831                         gen6_set_rps_thresholds(dev_priv, val);
4832         }
4833
4834         dev_priv->rps.cur_freq = val;
4835         trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
4836 }
4837
4838 /* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
4839  *
4840  * * If Gfx is Idle, then
4841  * 1. Forcewake Media well.
4842  * 2. Request idle freq.
4843  * 3. Release Forcewake of Media well.
4844 */
4845 static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
4846 {
4847         u32 val = dev_priv->rps.idle_freq;
4848
4849         if (dev_priv->rps.cur_freq <= val)
4850                 return;
4851
4852         /* Wake up the media well, as that takes a lot less
4853          * power than the Render well. */
4854         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_MEDIA);
4855         valleyview_set_rps(dev_priv, val);
4856         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_MEDIA);
4857 }
4858
4859 void gen6_rps_busy(struct drm_i915_private *dev_priv)
4860 {
4861         mutex_lock(&dev_priv->rps.hw_lock);
4862         if (dev_priv->rps.enabled) {
4863                 if (dev_priv->pm_rps_events & (GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_UP_EI_EXPIRED))
4864                         gen6_rps_reset_ei(dev_priv);
4865                 I915_WRITE(GEN6_PMINTRMSK,
4866                            gen6_rps_pm_mask(dev_priv, dev_priv->rps.cur_freq));
4867
4868                 gen6_enable_rps_interrupts(dev_priv);
4869
4870                 /* Ensure we start at the user's desired frequency */
4871                 intel_set_rps(dev_priv,
4872                               clamp(dev_priv->rps.cur_freq,
4873                                     dev_priv->rps.min_freq_softlimit,
4874                                     dev_priv->rps.max_freq_softlimit));
4875         }
4876         mutex_unlock(&dev_priv->rps.hw_lock);
4877 }
4878
4879 void gen6_rps_idle(struct drm_i915_private *dev_priv)
4880 {
4881         /* Flush our bottom-half so that it does not race with us
4882          * setting the idle frequency and so that it is bounded by
4883          * our rpm wakeref. And then disable the interrupts to stop any
4884          * futher RPS reclocking whilst we are asleep.
4885          */
4886         gen6_disable_rps_interrupts(dev_priv);
4887
4888         mutex_lock(&dev_priv->rps.hw_lock);
4889         if (dev_priv->rps.enabled) {
4890                 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
4891                         vlv_set_rps_idle(dev_priv);
4892                 else
4893                         gen6_set_rps(dev_priv, dev_priv->rps.idle_freq);
4894                 dev_priv->rps.last_adj = 0;
4895                 I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
4896         }
4897         mutex_unlock(&dev_priv->rps.hw_lock);
4898
4899         spin_lock(&dev_priv->rps.client_lock);
4900         while (!list_empty(&dev_priv->rps.clients))
4901                 list_del_init(dev_priv->rps.clients.next);
4902         spin_unlock(&dev_priv->rps.client_lock);
4903 }
4904
4905 void gen6_rps_boost(struct drm_i915_private *dev_priv,
4906                     struct intel_rps_client *rps,
4907                     unsigned long submitted)
4908 {
4909         /* This is intentionally racy! We peek at the state here, then
4910          * validate inside the RPS worker.
4911          */
4912         if (!(dev_priv->gt.awake &&
4913               dev_priv->rps.enabled &&
4914               dev_priv->rps.cur_freq < dev_priv->rps.max_freq_softlimit))
4915                 return;
4916
4917         /* Force a RPS boost (and don't count it against the client) if
4918          * the GPU is severely congested.
4919          */
4920         if (rps && time_after(jiffies, submitted + DRM_I915_THROTTLE_JIFFIES))
4921                 rps = NULL;
4922
4923         spin_lock(&dev_priv->rps.client_lock);
4924         if (rps == NULL || list_empty(&rps->link)) {
4925                 spin_lock_irq(&dev_priv->irq_lock);
4926                 if (dev_priv->rps.interrupts_enabled) {
4927                         dev_priv->rps.client_boost = true;
4928                         schedule_work(&dev_priv->rps.work);
4929                 }
4930                 spin_unlock_irq(&dev_priv->irq_lock);
4931
4932                 if (rps != NULL) {
4933                         list_add(&rps->link, &dev_priv->rps.clients);
4934                         rps->boosts++;
4935                 } else
4936                         dev_priv->rps.boosts++;
4937         }
4938         spin_unlock(&dev_priv->rps.client_lock);
4939 }
4940
4941 void intel_set_rps(struct drm_i915_private *dev_priv, u8 val)
4942 {
4943         if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
4944                 valleyview_set_rps(dev_priv, val);
4945         else
4946                 gen6_set_rps(dev_priv, val);
4947 }
4948
4949 static void gen9_disable_rc6(struct drm_i915_private *dev_priv)
4950 {
4951         I915_WRITE(GEN6_RC_CONTROL, 0);
4952         I915_WRITE(GEN9_PG_ENABLE, 0);
4953 }
4954
4955 static void gen9_disable_rps(struct drm_i915_private *dev_priv)
4956 {
4957         I915_WRITE(GEN6_RP_CONTROL, 0);
4958 }
4959
4960 static void gen6_disable_rps(struct drm_i915_private *dev_priv)
4961 {
4962         I915_WRITE(GEN6_RC_CONTROL, 0);
4963         I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
4964         I915_WRITE(GEN6_RP_CONTROL, 0);
4965 }
4966
4967 static void cherryview_disable_rps(struct drm_i915_private *dev_priv)
4968 {
4969         I915_WRITE(GEN6_RC_CONTROL, 0);
4970 }
4971
4972 static void valleyview_disable_rps(struct drm_i915_private *dev_priv)
4973 {
4974         /* we're doing forcewake before Disabling RC6,
4975          * This what the BIOS expects when going into suspend */
4976         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4977
4978         I915_WRITE(GEN6_RC_CONTROL, 0);
4979
4980         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4981 }
4982
4983 static void intel_print_rc6_info(struct drm_i915_private *dev_priv, u32 mode)
4984 {
4985         if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
4986                 if (mode & (GEN7_RC_CTL_TO_MODE | GEN6_RC_CTL_EI_MODE(1)))
4987                         mode = GEN6_RC_CTL_RC6_ENABLE;
4988                 else
4989                         mode = 0;
4990         }
4991         if (HAS_RC6p(dev_priv))
4992                 DRM_DEBUG_DRIVER("Enabling RC6 states: "
4993                                  "RC6 %s RC6p %s RC6pp %s\n",
4994                                  onoff(mode & GEN6_RC_CTL_RC6_ENABLE),
4995                                  onoff(mode & GEN6_RC_CTL_RC6p_ENABLE),
4996                                  onoff(mode & GEN6_RC_CTL_RC6pp_ENABLE));
4997
4998         else
4999                 DRM_DEBUG_DRIVER("Enabling RC6 states: RC6 %s\n",
5000                                  onoff(mode & GEN6_RC_CTL_RC6_ENABLE));
5001 }
5002
5003 static bool bxt_check_bios_rc6_setup(struct drm_i915_private *dev_priv)
5004 {
5005         struct i915_ggtt *ggtt = &dev_priv->ggtt;
5006         bool enable_rc6 = true;
5007         unsigned long rc6_ctx_base;
5008         u32 rc_ctl;
5009         int rc_sw_target;
5010
5011         rc_ctl = I915_READ(GEN6_RC_CONTROL);
5012         rc_sw_target = (I915_READ(GEN6_RC_STATE) & RC_SW_TARGET_STATE_MASK) >>
5013                        RC_SW_TARGET_STATE_SHIFT;
5014         DRM_DEBUG_DRIVER("BIOS enabled RC states: "
5015                          "HW_CTRL %s HW_RC6 %s SW_TARGET_STATE %x\n",
5016                          onoff(rc_ctl & GEN6_RC_CTL_HW_ENABLE),
5017                          onoff(rc_ctl & GEN6_RC_CTL_RC6_ENABLE),
5018                          rc_sw_target);
5019
5020         if (!(I915_READ(RC6_LOCATION) & RC6_CTX_IN_DRAM)) {
5021                 DRM_DEBUG_DRIVER("RC6 Base location not set properly.\n");
5022                 enable_rc6 = false;
5023         }
5024
5025         /*
5026          * The exact context size is not known for BXT, so assume a page size
5027          * for this check.
5028          */
5029         rc6_ctx_base = I915_READ(RC6_CTX_BASE) & RC6_CTX_BASE_MASK;
5030         if (!((rc6_ctx_base >= ggtt->stolen_reserved_base) &&
5031               (rc6_ctx_base + PAGE_SIZE <= ggtt->stolen_reserved_base +
5032                                         ggtt->stolen_reserved_size))) {
5033                 DRM_DEBUG_DRIVER("RC6 Base address not as expected.\n");
5034                 enable_rc6 = false;
5035         }
5036
5037         if (!(((I915_READ(PWRCTX_MAXCNT_RCSUNIT) & IDLE_TIME_MASK) > 1) &&
5038               ((I915_READ(PWRCTX_MAXCNT_VCSUNIT0) & IDLE_TIME_MASK) > 1) &&
5039               ((I915_READ(PWRCTX_MAXCNT_BCSUNIT) & IDLE_TIME_MASK) > 1) &&
5040               ((I915_READ(PWRCTX_MAXCNT_VECSUNIT) & IDLE_TIME_MASK) > 1))) {
5041                 DRM_DEBUG_DRIVER("Engine Idle wait time not set properly.\n");
5042                 enable_rc6 = false;
5043         }
5044
5045         if (!I915_READ(GEN8_PUSHBUS_CONTROL) ||
5046             !I915_READ(GEN8_PUSHBUS_ENABLE) ||
5047             !I915_READ(GEN8_PUSHBUS_SHIFT)) {
5048                 DRM_DEBUG_DRIVER("Pushbus not setup properly.\n");
5049                 enable_rc6 = false;
5050         }
5051
5052         if (!I915_READ(GEN6_GFXPAUSE)) {
5053                 DRM_DEBUG_DRIVER("GFX pause not setup properly.\n");
5054                 enable_rc6 = false;
5055         }
5056
5057         if (!I915_READ(GEN8_MISC_CTRL0)) {
5058                 DRM_DEBUG_DRIVER("GPM control not setup properly.\n");
5059                 enable_rc6 = false;
5060         }
5061
5062         return enable_rc6;
5063 }
5064
5065 int sanitize_rc6_option(struct drm_i915_private *dev_priv, int enable_rc6)
5066 {
5067         /* No RC6 before Ironlake and code is gone for ilk. */
5068         if (INTEL_INFO(dev_priv)->gen < 6)
5069                 return 0;
5070
5071         if (!enable_rc6)
5072                 return 0;
5073
5074         if (IS_BROXTON(dev_priv) && !bxt_check_bios_rc6_setup(dev_priv)) {
5075                 DRM_INFO("RC6 disabled by BIOS\n");
5076                 return 0;
5077         }
5078
5079         /* Respect the kernel parameter if it is set */
5080         if (enable_rc6 >= 0) {
5081                 int mask;
5082
5083                 if (HAS_RC6p(dev_priv))
5084                         mask = INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE |
5085                                INTEL_RC6pp_ENABLE;
5086                 else
5087                         mask = INTEL_RC6_ENABLE;
5088
5089                 if ((enable_rc6 & mask) != enable_rc6)
5090                         DRM_DEBUG_DRIVER("Adjusting RC6 mask to %d "
5091                                          "(requested %d, valid %d)\n",
5092                                          enable_rc6 & mask, enable_rc6, mask);
5093
5094                 return enable_rc6 & mask;
5095         }
5096
5097         if (IS_IVYBRIDGE(dev_priv))
5098                 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
5099
5100         return INTEL_RC6_ENABLE;
5101 }
5102
5103 static void gen6_init_rps_frequencies(struct drm_i915_private *dev_priv)
5104 {
5105         uint32_t rp_state_cap;
5106         u32 ddcc_status = 0;
5107         int ret;
5108
5109         /* All of these values are in units of 50MHz */
5110         dev_priv->rps.cur_freq          = 0;
5111         /* static values from HW: RP0 > RP1 > RPn (min_freq) */
5112         if (IS_BROXTON(dev_priv)) {
5113                 rp_state_cap = I915_READ(BXT_RP_STATE_CAP);
5114                 dev_priv->rps.rp0_freq = (rp_state_cap >> 16) & 0xff;
5115                 dev_priv->rps.rp1_freq = (rp_state_cap >>  8) & 0xff;
5116                 dev_priv->rps.min_freq = (rp_state_cap >>  0) & 0xff;
5117         } else {
5118                 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
5119                 dev_priv->rps.rp0_freq = (rp_state_cap >>  0) & 0xff;
5120                 dev_priv->rps.rp1_freq = (rp_state_cap >>  8) & 0xff;
5121                 dev_priv->rps.min_freq = (rp_state_cap >> 16) & 0xff;
5122         }
5123
5124         /* hw_max = RP0 until we check for overclocking */
5125         dev_priv->rps.max_freq          = dev_priv->rps.rp0_freq;
5126
5127         dev_priv->rps.efficient_freq = dev_priv->rps.rp1_freq;
5128         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv) ||
5129             IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5130                 ret = sandybridge_pcode_read(dev_priv,
5131                                         HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL,
5132                                         &ddcc_status);
5133                 if (0 == ret)
5134                         dev_priv->rps.efficient_freq =
5135                                 clamp_t(u8,
5136                                         ((ddcc_status >> 8) & 0xff),
5137                                         dev_priv->rps.min_freq,
5138                                         dev_priv->rps.max_freq);
5139         }
5140
5141         if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5142                 /* Store the frequency values in 16.66 MHZ units, which is
5143                    the natural hardware unit for SKL */
5144                 dev_priv->rps.rp0_freq *= GEN9_FREQ_SCALER;
5145                 dev_priv->rps.rp1_freq *= GEN9_FREQ_SCALER;
5146                 dev_priv->rps.min_freq *= GEN9_FREQ_SCALER;
5147                 dev_priv->rps.max_freq *= GEN9_FREQ_SCALER;
5148                 dev_priv->rps.efficient_freq *= GEN9_FREQ_SCALER;
5149         }
5150
5151         dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
5152
5153         /* Preserve min/max settings in case of re-init */
5154         if (dev_priv->rps.max_freq_softlimit == 0)
5155                 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
5156
5157         if (dev_priv->rps.min_freq_softlimit == 0) {
5158                 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
5159                         dev_priv->rps.min_freq_softlimit =
5160                                 max_t(int, dev_priv->rps.efficient_freq,
5161                                       intel_freq_opcode(dev_priv, 450));
5162                 else
5163                         dev_priv->rps.min_freq_softlimit =
5164                                 dev_priv->rps.min_freq;
5165         }
5166 }
5167
5168 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
5169 static void gen9_enable_rps(struct drm_i915_private *dev_priv)
5170 {
5171         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5172
5173         gen6_init_rps_frequencies(dev_priv);
5174
5175         /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
5176         if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) {
5177                 /*
5178                  * BIOS could leave the Hw Turbo enabled, so need to explicitly
5179                  * clear out the Control register just to avoid inconsitency
5180                  * with debugfs interface, which will show  Turbo as enabled
5181                  * only and that is not expected by the User after adding the
5182                  * WaGsvDisableTurbo. Apart from this there is no problem even
5183                  * if the Turbo is left enabled in the Control register, as the
5184                  * Up/Down interrupts would remain masked.
5185                  */
5186                 gen9_disable_rps(dev_priv);
5187                 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5188                 return;
5189         }
5190
5191         /* Program defaults and thresholds for RPS*/
5192         I915_WRITE(GEN6_RC_VIDEO_FREQ,
5193                 GEN9_FREQUENCY(dev_priv->rps.rp1_freq));
5194
5195         /* 1 second timeout*/
5196         I915_WRITE(GEN6_RP_DOWN_TIMEOUT,
5197                 GT_INTERVAL_FROM_US(dev_priv, 1000000));
5198
5199         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 0xa);
5200
5201         /* Leaning on the below call to gen6_set_rps to program/setup the
5202          * Up/Down EI & threshold registers, as well as the RP_CONTROL,
5203          * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
5204         dev_priv->rps.power = HIGH_POWER; /* force a reset */
5205         gen6_set_rps(dev_priv, dev_priv->rps.idle_freq);
5206
5207         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5208 }
5209
5210 static void gen9_enable_rc6(struct drm_i915_private *dev_priv)
5211 {
5212         struct intel_engine_cs *engine;
5213         uint32_t rc6_mask = 0;
5214
5215         /* 1a: Software RC state - RC0 */
5216         I915_WRITE(GEN6_RC_STATE, 0);
5217
5218         /* 1b: Get forcewake during program sequence. Although the driver
5219          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5220         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5221
5222         /* 2a: Disable RC states. */
5223         I915_WRITE(GEN6_RC_CONTROL, 0);
5224
5225         /* 2b: Program RC6 thresholds.*/
5226
5227         /* WaRsDoubleRc6WrlWithCoarsePowerGating: Doubling WRL only when CPG is enabled */
5228         if (IS_SKYLAKE(dev_priv))
5229                 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 108 << 16);
5230         else
5231                 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16);
5232         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5233         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
5234         for_each_engine(engine, dev_priv)
5235                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5236
5237         if (HAS_GUC(dev_priv))
5238                 I915_WRITE(GUC_MAX_IDLE_COUNT, 0xA);
5239
5240         I915_WRITE(GEN6_RC_SLEEP, 0);
5241
5242         /* 2c: Program Coarse Power Gating Policies. */
5243         I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS, 25);
5244         I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS, 25);
5245
5246         /* 3a: Enable RC6 */
5247         if (intel_enable_rc6() & INTEL_RC6_ENABLE)
5248                 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
5249         DRM_INFO("RC6 %s\n", onoff(rc6_mask & GEN6_RC_CTL_RC6_ENABLE));
5250         /* WaRsUseTimeoutMode */
5251         if (IS_SKL_REVID(dev_priv, 0, SKL_REVID_D0) ||
5252             IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) {
5253                 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us */
5254                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5255                            GEN7_RC_CTL_TO_MODE |
5256                            rc6_mask);
5257         } else {
5258                 I915_WRITE(GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */
5259                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5260                            GEN6_RC_CTL_EI_MODE(1) |
5261                            rc6_mask);
5262         }
5263
5264         /*
5265          * 3b: Enable Coarse Power Gating only when RC6 is enabled.
5266          * WaRsDisableCoarsePowerGating:skl,bxt - Render/Media PG need to be disabled with RC6.
5267          */
5268         if (NEEDS_WaRsDisableCoarsePowerGating(dev_priv))
5269                 I915_WRITE(GEN9_PG_ENABLE, 0);
5270         else
5271                 I915_WRITE(GEN9_PG_ENABLE, (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ?
5272                                 (GEN9_RENDER_PG_ENABLE | GEN9_MEDIA_PG_ENABLE) : 0);
5273
5274         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5275 }
5276
5277 static void gen8_enable_rps(struct drm_i915_private *dev_priv)
5278 {
5279         struct intel_engine_cs *engine;
5280         uint32_t rc6_mask = 0;
5281
5282         /* 1a: Software RC state - RC0 */
5283         I915_WRITE(GEN6_RC_STATE, 0);
5284
5285         /* 1c & 1d: Get forcewake during program sequence. Although the driver
5286          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5287         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5288
5289         /* 2a: Disable RC states. */
5290         I915_WRITE(GEN6_RC_CONTROL, 0);
5291
5292         /* Initialize rps frequencies */
5293         gen6_init_rps_frequencies(dev_priv);
5294
5295         /* 2b: Program RC6 thresholds.*/
5296         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
5297         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5298         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
5299         for_each_engine(engine, dev_priv)
5300                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5301         I915_WRITE(GEN6_RC_SLEEP, 0);
5302         if (IS_BROADWELL(dev_priv))
5303                 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */
5304         else
5305                 I915_WRITE(GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
5306
5307         /* 3: Enable RC6 */
5308         if (intel_enable_rc6() & INTEL_RC6_ENABLE)
5309                 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
5310         intel_print_rc6_info(dev_priv, rc6_mask);
5311         if (IS_BROADWELL(dev_priv))
5312                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5313                                 GEN7_RC_CTL_TO_MODE |
5314                                 rc6_mask);
5315         else
5316                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5317                                 GEN6_RC_CTL_EI_MODE(1) |
5318                                 rc6_mask);
5319
5320         /* 4 Program defaults and thresholds for RPS*/
5321         I915_WRITE(GEN6_RPNSWREQ,
5322                    HSW_FREQUENCY(dev_priv->rps.rp1_freq));
5323         I915_WRITE(GEN6_RC_VIDEO_FREQ,
5324                    HSW_FREQUENCY(dev_priv->rps.rp1_freq));
5325         /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
5326         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
5327
5328         /* Docs recommend 900MHz, and 300 MHz respectively */
5329         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
5330                    dev_priv->rps.max_freq_softlimit << 24 |
5331                    dev_priv->rps.min_freq_softlimit << 16);
5332
5333         I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
5334         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
5335         I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
5336         I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
5337
5338         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5339
5340         /* 5: Enable RPS */
5341         I915_WRITE(GEN6_RP_CONTROL,
5342                    GEN6_RP_MEDIA_TURBO |
5343                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
5344                    GEN6_RP_MEDIA_IS_GFX |
5345                    GEN6_RP_ENABLE |
5346                    GEN6_RP_UP_BUSY_AVG |
5347                    GEN6_RP_DOWN_IDLE_AVG);
5348
5349         /* 6: Ring frequency + overclocking (our driver does this later */
5350
5351         dev_priv->rps.power = HIGH_POWER; /* force a reset */
5352         gen6_set_rps(dev_priv, dev_priv->rps.idle_freq);
5353
5354         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5355 }
5356
5357 static void gen6_enable_rps(struct drm_i915_private *dev_priv)
5358 {
5359         struct intel_engine_cs *engine;
5360         u32 rc6vids, pcu_mbox = 0, rc6_mask = 0;
5361         u32 gtfifodbg;
5362         int rc6_mode;
5363         int ret;
5364
5365         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5366
5367         /* Here begins a magic sequence of register writes to enable
5368          * auto-downclocking.
5369          *
5370          * Perhaps there might be some value in exposing these to
5371          * userspace...
5372          */
5373         I915_WRITE(GEN6_RC_STATE, 0);
5374
5375         /* Clear the DBG now so we don't confuse earlier errors */
5376         gtfifodbg = I915_READ(GTFIFODBG);
5377         if (gtfifodbg) {
5378                 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
5379                 I915_WRITE(GTFIFODBG, gtfifodbg);
5380         }
5381
5382         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5383
5384         /* Initialize rps frequencies */
5385         gen6_init_rps_frequencies(dev_priv);
5386
5387         /* disable the counters and set deterministic thresholds */
5388         I915_WRITE(GEN6_RC_CONTROL, 0);
5389
5390         I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
5391         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
5392         I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
5393         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
5394         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
5395
5396         for_each_engine(engine, dev_priv)
5397                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5398
5399         I915_WRITE(GEN6_RC_SLEEP, 0);
5400         I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
5401         if (IS_IVYBRIDGE(dev_priv))
5402                 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
5403         else
5404                 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
5405         I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
5406         I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
5407
5408         /* Check if we are enabling RC6 */
5409         rc6_mode = intel_enable_rc6();
5410         if (rc6_mode & INTEL_RC6_ENABLE)
5411                 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
5412
5413         /* We don't use those on Haswell */
5414         if (!IS_HASWELL(dev_priv)) {
5415                 if (rc6_mode & INTEL_RC6p_ENABLE)
5416                         rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
5417
5418                 if (rc6_mode & INTEL_RC6pp_ENABLE)
5419                         rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
5420         }
5421
5422         intel_print_rc6_info(dev_priv, rc6_mask);
5423
5424         I915_WRITE(GEN6_RC_CONTROL,
5425                    rc6_mask |
5426                    GEN6_RC_CTL_EI_MODE(1) |
5427                    GEN6_RC_CTL_HW_ENABLE);
5428
5429         /* Power down if completely idle for over 50ms */
5430         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
5431         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5432
5433         ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_MIN_FREQ_TABLE, 0);
5434         if (ret)
5435                 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
5436
5437         ret = sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &pcu_mbox);
5438         if (!ret && (pcu_mbox & (1<<31))) { /* OC supported */
5439                 DRM_DEBUG_DRIVER("Overclocking supported. Max: %dMHz, Overclock max: %dMHz\n",
5440                                  (dev_priv->rps.max_freq_softlimit & 0xff) * 50,
5441                                  (pcu_mbox & 0xff) * 50);
5442                 dev_priv->rps.max_freq = pcu_mbox & 0xff;
5443         }
5444
5445         dev_priv->rps.power = HIGH_POWER; /* force a reset */
5446         gen6_set_rps(dev_priv, dev_priv->rps.idle_freq);
5447
5448         rc6vids = 0;
5449         ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
5450         if (IS_GEN6(dev_priv) && ret) {
5451                 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
5452         } else if (IS_GEN6(dev_priv) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
5453                 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
5454                           GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
5455                 rc6vids &= 0xffff00;
5456                 rc6vids |= GEN6_ENCODE_RC6_VID(450);
5457                 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
5458                 if (ret)
5459                         DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
5460         }
5461
5462         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5463 }
5464
5465 static void __gen6_update_ring_freq(struct drm_i915_private *dev_priv)
5466 {
5467         int min_freq = 15;
5468         unsigned int gpu_freq;
5469         unsigned int max_ia_freq, min_ring_freq;
5470         unsigned int max_gpu_freq, min_gpu_freq;
5471         int scaling_factor = 180;
5472         struct cpufreq_policy *policy;
5473
5474         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5475
5476         policy = cpufreq_cpu_get(0);
5477         if (policy) {
5478                 max_ia_freq = policy->cpuinfo.max_freq;
5479                 cpufreq_cpu_put(policy);
5480         } else {
5481                 /*
5482                  * Default to measured freq if none found, PCU will ensure we
5483                  * don't go over
5484                  */
5485                 max_ia_freq = tsc_khz;
5486         }
5487
5488         /* Convert from kHz to MHz */
5489         max_ia_freq /= 1000;
5490
5491         min_ring_freq = I915_READ(DCLK) & 0xf;
5492         /* convert DDR frequency from units of 266.6MHz to bandwidth */
5493         min_ring_freq = mult_frac(min_ring_freq, 8, 3);
5494
5495         if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5496                 /* Convert GT frequency to 50 HZ units */
5497                 min_gpu_freq = dev_priv->rps.min_freq / GEN9_FREQ_SCALER;
5498                 max_gpu_freq = dev_priv->rps.max_freq / GEN9_FREQ_SCALER;
5499         } else {
5500                 min_gpu_freq = dev_priv->rps.min_freq;
5501                 max_gpu_freq = dev_priv->rps.max_freq;
5502         }
5503
5504         /*
5505          * For each potential GPU frequency, load a ring frequency we'd like
5506          * to use for memory access.  We do this by specifying the IA frequency
5507          * the PCU should use as a reference to determine the ring frequency.
5508          */
5509         for (gpu_freq = max_gpu_freq; gpu_freq >= min_gpu_freq; gpu_freq--) {
5510                 int diff = max_gpu_freq - gpu_freq;
5511                 unsigned int ia_freq = 0, ring_freq = 0;
5512
5513                 if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5514                         /*
5515                          * ring_freq = 2 * GT. ring_freq is in 100MHz units
5516                          * No floor required for ring frequency on SKL.
5517                          */
5518                         ring_freq = gpu_freq;
5519                 } else if (INTEL_INFO(dev_priv)->gen >= 8) {
5520                         /* max(2 * GT, DDR). NB: GT is 50MHz units */
5521                         ring_freq = max(min_ring_freq, gpu_freq);
5522                 } else if (IS_HASWELL(dev_priv)) {
5523                         ring_freq = mult_frac(gpu_freq, 5, 4);
5524                         ring_freq = max(min_ring_freq, ring_freq);
5525                         /* leave ia_freq as the default, chosen by cpufreq */
5526                 } else {
5527                         /* On older processors, there is no separate ring
5528                          * clock domain, so in order to boost the bandwidth
5529                          * of the ring, we need to upclock the CPU (ia_freq).
5530                          *
5531                          * For GPU frequencies less than 750MHz,
5532                          * just use the lowest ring freq.
5533                          */
5534                         if (gpu_freq < min_freq)
5535                                 ia_freq = 800;
5536                         else
5537                                 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
5538                         ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
5539                 }
5540
5541                 sandybridge_pcode_write(dev_priv,
5542                                         GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
5543                                         ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
5544                                         ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
5545                                         gpu_freq);
5546         }
5547 }
5548
5549 void gen6_update_ring_freq(struct drm_i915_private *dev_priv)
5550 {
5551         if (!HAS_CORE_RING_FREQ(dev_priv))
5552                 return;
5553
5554         mutex_lock(&dev_priv->rps.hw_lock);
5555         __gen6_update_ring_freq(dev_priv);
5556         mutex_unlock(&dev_priv->rps.hw_lock);
5557 }
5558
5559 static int cherryview_rps_max_freq(struct drm_i915_private *dev_priv)
5560 {
5561         u32 val, rp0;
5562
5563         val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
5564
5565         switch (INTEL_INFO(dev_priv)->eu_total) {
5566         case 8:
5567                 /* (2 * 4) config */
5568                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT);
5569                 break;
5570         case 12:
5571                 /* (2 * 6) config */
5572                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT);
5573                 break;
5574         case 16:
5575                 /* (2 * 8) config */
5576         default:
5577                 /* Setting (2 * 8) Min RP0 for any other combination */
5578                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT);
5579                 break;
5580         }
5581
5582         rp0 = (rp0 & FB_GFX_FREQ_FUSE_MASK);
5583
5584         return rp0;
5585 }
5586
5587 static int cherryview_rps_rpe_freq(struct drm_i915_private *dev_priv)
5588 {
5589         u32 val, rpe;
5590
5591         val = vlv_punit_read(dev_priv, PUNIT_GPU_DUTYCYCLE_REG);
5592         rpe = (val >> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK;
5593
5594         return rpe;
5595 }
5596
5597 static int cherryview_rps_guar_freq(struct drm_i915_private *dev_priv)
5598 {
5599         u32 val, rp1;
5600
5601         val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
5602         rp1 = (val & FB_GFX_FREQ_FUSE_MASK);
5603
5604         return rp1;
5605 }
5606
5607 static int valleyview_rps_guar_freq(struct drm_i915_private *dev_priv)
5608 {
5609         u32 val, rp1;
5610
5611         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
5612
5613         rp1 = (val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT;
5614
5615         return rp1;
5616 }
5617
5618 static int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
5619 {
5620         u32 val, rp0;
5621
5622         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
5623
5624         rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
5625         /* Clamp to max */
5626         rp0 = min_t(u32, rp0, 0xea);
5627
5628         return rp0;
5629 }
5630
5631 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
5632 {
5633         u32 val, rpe;
5634
5635         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
5636         rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
5637         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
5638         rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
5639
5640         return rpe;
5641 }
5642
5643 static int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
5644 {
5645         u32 val;
5646
5647         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
5648         /*
5649          * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
5650          * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
5651          * a BYT-M B0 the above register contains 0xbf. Moreover when setting
5652          * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
5653          * to make sure it matches what Punit accepts.
5654          */
5655         return max_t(u32, val, 0xc0);
5656 }
5657
5658 /* Check that the pctx buffer wasn't move under us. */
5659 static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
5660 {
5661         unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
5662
5663         WARN_ON(pctx_addr != dev_priv->mm.stolen_base +
5664                              dev_priv->vlv_pctx->stolen->start);
5665 }
5666
5667
5668 /* Check that the pcbr address is not empty. */
5669 static void cherryview_check_pctx(struct drm_i915_private *dev_priv)
5670 {
5671         unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
5672
5673         WARN_ON((pctx_addr >> VLV_PCBR_ADDR_SHIFT) == 0);
5674 }
5675
5676 static void cherryview_setup_pctx(struct drm_i915_private *dev_priv)
5677 {
5678         struct i915_ggtt *ggtt = &dev_priv->ggtt;
5679         unsigned long pctx_paddr, paddr;
5680         u32 pcbr;
5681         int pctx_size = 32*1024;
5682
5683         pcbr = I915_READ(VLV_PCBR);
5684         if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) {
5685                 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5686                 paddr = (dev_priv->mm.stolen_base +
5687                          (ggtt->stolen_size - pctx_size));
5688
5689                 pctx_paddr = (paddr & (~4095));
5690                 I915_WRITE(VLV_PCBR, pctx_paddr);
5691         }
5692
5693         DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
5694 }
5695
5696 static void valleyview_setup_pctx(struct drm_i915_private *dev_priv)
5697 {
5698         struct drm_i915_gem_object *pctx;
5699         unsigned long pctx_paddr;
5700         u32 pcbr;
5701         int pctx_size = 24*1024;
5702
5703         mutex_lock(&dev_priv->drm.struct_mutex);
5704
5705         pcbr = I915_READ(VLV_PCBR);
5706         if (pcbr) {
5707                 /* BIOS set it up already, grab the pre-alloc'd space */
5708                 int pcbr_offset;
5709
5710                 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
5711                 pctx = i915_gem_object_create_stolen_for_preallocated(&dev_priv->drm,
5712                                                                       pcbr_offset,
5713                                                                       I915_GTT_OFFSET_NONE,
5714                                                                       pctx_size);
5715                 goto out;
5716         }
5717
5718         DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5719
5720         /*
5721          * From the Gunit register HAS:
5722          * The Gfx driver is expected to program this register and ensure
5723          * proper allocation within Gfx stolen memory.  For example, this
5724          * register should be programmed such than the PCBR range does not
5725          * overlap with other ranges, such as the frame buffer, protected
5726          * memory, or any other relevant ranges.
5727          */
5728         pctx = i915_gem_object_create_stolen(&dev_priv->drm, pctx_size);
5729         if (!pctx) {
5730                 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
5731                 goto out;
5732         }
5733
5734         pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
5735         I915_WRITE(VLV_PCBR, pctx_paddr);
5736
5737 out:
5738         DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
5739         dev_priv->vlv_pctx = pctx;
5740         mutex_unlock(&dev_priv->drm.struct_mutex);
5741 }
5742
5743 static void valleyview_cleanup_pctx(struct drm_i915_private *dev_priv)
5744 {
5745         if (WARN_ON(!dev_priv->vlv_pctx))
5746                 return;
5747
5748         drm_gem_object_unreference_unlocked(&dev_priv->vlv_pctx->base);
5749         dev_priv->vlv_pctx = NULL;
5750 }
5751
5752 static void vlv_init_gpll_ref_freq(struct drm_i915_private *dev_priv)
5753 {
5754         dev_priv->rps.gpll_ref_freq =
5755                 vlv_get_cck_clock(dev_priv, "GPLL ref",
5756                                   CCK_GPLL_CLOCK_CONTROL,
5757                                   dev_priv->czclk_freq);
5758
5759         DRM_DEBUG_DRIVER("GPLL reference freq: %d kHz\n",
5760                          dev_priv->rps.gpll_ref_freq);
5761 }
5762
5763 static void valleyview_init_gt_powersave(struct drm_i915_private *dev_priv)
5764 {
5765         u32 val;
5766
5767         valleyview_setup_pctx(dev_priv);
5768
5769         vlv_init_gpll_ref_freq(dev_priv);
5770
5771         mutex_lock(&dev_priv->rps.hw_lock);
5772
5773         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5774         switch ((val >> 6) & 3) {
5775         case 0:
5776         case 1:
5777                 dev_priv->mem_freq = 800;
5778                 break;
5779         case 2:
5780                 dev_priv->mem_freq = 1066;
5781                 break;
5782         case 3:
5783                 dev_priv->mem_freq = 1333;
5784                 break;
5785         }
5786         DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
5787
5788         dev_priv->rps.max_freq = valleyview_rps_max_freq(dev_priv);
5789         dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
5790         DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5791                          intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
5792                          dev_priv->rps.max_freq);
5793
5794         dev_priv->rps.efficient_freq = valleyview_rps_rpe_freq(dev_priv);
5795         DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5796                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5797                          dev_priv->rps.efficient_freq);
5798
5799         dev_priv->rps.rp1_freq = valleyview_rps_guar_freq(dev_priv);
5800         DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
5801                          intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
5802                          dev_priv->rps.rp1_freq);
5803
5804         dev_priv->rps.min_freq = valleyview_rps_min_freq(dev_priv);
5805         DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5806                          intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
5807                          dev_priv->rps.min_freq);
5808
5809         dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
5810
5811         /* Preserve min/max settings in case of re-init */
5812         if (dev_priv->rps.max_freq_softlimit == 0)
5813                 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
5814
5815         if (dev_priv->rps.min_freq_softlimit == 0)
5816                 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
5817
5818         mutex_unlock(&dev_priv->rps.hw_lock);
5819 }
5820
5821 static void cherryview_init_gt_powersave(struct drm_i915_private *dev_priv)
5822 {
5823         u32 val;
5824
5825         cherryview_setup_pctx(dev_priv);
5826
5827         vlv_init_gpll_ref_freq(dev_priv);
5828
5829         mutex_lock(&dev_priv->rps.hw_lock);
5830
5831         mutex_lock(&dev_priv->sb_lock);
5832         val = vlv_cck_read(dev_priv, CCK_FUSE_REG);
5833         mutex_unlock(&dev_priv->sb_lock);
5834
5835         switch ((val >> 2) & 0x7) {
5836         case 3:
5837                 dev_priv->mem_freq = 2000;
5838                 break;
5839         default:
5840                 dev_priv->mem_freq = 1600;
5841                 break;
5842         }
5843         DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
5844
5845         dev_priv->rps.max_freq = cherryview_rps_max_freq(dev_priv);
5846         dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
5847         DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5848                          intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
5849                          dev_priv->rps.max_freq);
5850
5851         dev_priv->rps.efficient_freq = cherryview_rps_rpe_freq(dev_priv);
5852         DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5853                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5854                          dev_priv->rps.efficient_freq);
5855
5856         dev_priv->rps.rp1_freq = cherryview_rps_guar_freq(dev_priv);
5857         DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
5858                          intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
5859                          dev_priv->rps.rp1_freq);
5860
5861         /* PUnit validated range is only [RPe, RP0] */
5862         dev_priv->rps.min_freq = dev_priv->rps.efficient_freq;
5863         DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5864                          intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
5865                          dev_priv->rps.min_freq);
5866
5867         WARN_ONCE((dev_priv->rps.max_freq |
5868                    dev_priv->rps.efficient_freq |
5869                    dev_priv->rps.rp1_freq |
5870                    dev_priv->rps.min_freq) & 1,
5871                   "Odd GPU freq values\n");
5872
5873         dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
5874
5875         /* Preserve min/max settings in case of re-init */
5876         if (dev_priv->rps.max_freq_softlimit == 0)
5877                 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
5878
5879         if (dev_priv->rps.min_freq_softlimit == 0)
5880                 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
5881
5882         mutex_unlock(&dev_priv->rps.hw_lock);
5883 }
5884
5885 static void valleyview_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
5886 {
5887         valleyview_cleanup_pctx(dev_priv);
5888 }
5889
5890 static void cherryview_enable_rps(struct drm_i915_private *dev_priv)
5891 {
5892         struct intel_engine_cs *engine;
5893         u32 gtfifodbg, val, rc6_mode = 0, pcbr;
5894
5895         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5896
5897         gtfifodbg = I915_READ(GTFIFODBG) & ~(GT_FIFO_SBDEDICATE_FREE_ENTRY_CHV |
5898                                              GT_FIFO_FREE_ENTRIES_CHV);
5899         if (gtfifodbg) {
5900                 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5901                                  gtfifodbg);
5902                 I915_WRITE(GTFIFODBG, gtfifodbg);
5903         }
5904
5905         cherryview_check_pctx(dev_priv);
5906
5907         /* 1a & 1b: Get forcewake during program sequence. Although the driver
5908          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5909         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5910
5911         /*  Disable RC states. */
5912         I915_WRITE(GEN6_RC_CONTROL, 0);
5913
5914         /* 2a: Program RC6 thresholds.*/
5915         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
5916         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5917         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
5918
5919         for_each_engine(engine, dev_priv)
5920                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5921         I915_WRITE(GEN6_RC_SLEEP, 0);
5922
5923         /* TO threshold set to 500 us ( 0x186 * 1.28 us) */
5924         I915_WRITE(GEN6_RC6_THRESHOLD, 0x186);
5925
5926         /* allows RC6 residency counter to work */
5927         I915_WRITE(VLV_COUNTER_CONTROL,
5928                    _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
5929                                       VLV_MEDIA_RC6_COUNT_EN |
5930                                       VLV_RENDER_RC6_COUNT_EN));
5931
5932         /* For now we assume BIOS is allocating and populating the PCBR  */
5933         pcbr = I915_READ(VLV_PCBR);
5934
5935         /* 3: Enable RC6 */
5936         if ((intel_enable_rc6() & INTEL_RC6_ENABLE) &&
5937             (pcbr >> VLV_PCBR_ADDR_SHIFT))
5938                 rc6_mode = GEN7_RC_CTL_TO_MODE;
5939
5940         I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
5941
5942         /* 4 Program defaults and thresholds for RPS*/
5943         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
5944         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
5945         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
5946         I915_WRITE(GEN6_RP_UP_EI, 66000);
5947         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
5948
5949         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5950
5951         /* 5: Enable RPS */
5952         I915_WRITE(GEN6_RP_CONTROL,
5953                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
5954                    GEN6_RP_MEDIA_IS_GFX |
5955                    GEN6_RP_ENABLE |
5956                    GEN6_RP_UP_BUSY_AVG |
5957                    GEN6_RP_DOWN_IDLE_AVG);
5958
5959         /* Setting Fixed Bias */
5960         val = VLV_OVERRIDE_EN |
5961                   VLV_SOC_TDP_EN |
5962                   CHV_BIAS_CPU_50_SOC_50;
5963         vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
5964
5965         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5966
5967         /* RPS code assumes GPLL is used */
5968         WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
5969
5970         DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
5971         DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
5972
5973         dev_priv->rps.cur_freq = (val >> 8) & 0xff;
5974         DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
5975                          intel_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
5976                          dev_priv->rps.cur_freq);
5977
5978         DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
5979                          intel_gpu_freq(dev_priv, dev_priv->rps.idle_freq),
5980                          dev_priv->rps.idle_freq);
5981
5982         valleyview_set_rps(dev_priv, dev_priv->rps.idle_freq);
5983
5984         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5985 }
5986
5987 static void valleyview_enable_rps(struct drm_i915_private *dev_priv)
5988 {
5989         struct intel_engine_cs *engine;
5990         u32 gtfifodbg, val, rc6_mode = 0;
5991
5992         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5993
5994         valleyview_check_pctx(dev_priv);
5995
5996         gtfifodbg = I915_READ(GTFIFODBG);
5997         if (gtfifodbg) {
5998                 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5999                                  gtfifodbg);
6000                 I915_WRITE(GTFIFODBG, gtfifodbg);
6001         }
6002
6003         /* If VLV, Forcewake all wells, else re-direct to regular path */
6004         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6005
6006         /*  Disable RC states. */
6007         I915_WRITE(GEN6_RC_CONTROL, 0);
6008
6009         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
6010         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
6011         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
6012         I915_WRITE(GEN6_RP_UP_EI, 66000);
6013         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
6014
6015         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
6016
6017         I915_WRITE(GEN6_RP_CONTROL,
6018                    GEN6_RP_MEDIA_TURBO |
6019                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
6020                    GEN6_RP_MEDIA_IS_GFX |
6021                    GEN6_RP_ENABLE |
6022                    GEN6_RP_UP_BUSY_AVG |
6023                    GEN6_RP_DOWN_IDLE_CONT);
6024
6025         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
6026         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
6027         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
6028
6029         for_each_engine(engine, dev_priv)
6030                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6031
6032         I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
6033
6034         /* allows RC6 residency counter to work */
6035         I915_WRITE(VLV_COUNTER_CONTROL,
6036                    _MASKED_BIT_ENABLE(VLV_MEDIA_RC0_COUNT_EN |
6037                                       VLV_RENDER_RC0_COUNT_EN |
6038                                       VLV_MEDIA_RC6_COUNT_EN |
6039                                       VLV_RENDER_RC6_COUNT_EN));
6040
6041         if (intel_enable_rc6() & INTEL_RC6_ENABLE)
6042                 rc6_mode = GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
6043
6044         intel_print_rc6_info(dev_priv, rc6_mode);
6045
6046         I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
6047
6048         /* Setting Fixed Bias */
6049         val = VLV_OVERRIDE_EN |
6050                   VLV_SOC_TDP_EN |
6051                   VLV_BIAS_CPU_125_SOC_875;
6052         vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
6053
6054         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
6055
6056         /* RPS code assumes GPLL is used */
6057         WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
6058
6059         DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
6060         DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
6061
6062         dev_priv->rps.cur_freq = (val >> 8) & 0xff;
6063         DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
6064                          intel_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
6065                          dev_priv->rps.cur_freq);
6066
6067         DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
6068                          intel_gpu_freq(dev_priv, dev_priv->rps.idle_freq),
6069                          dev_priv->rps.idle_freq);
6070
6071         valleyview_set_rps(dev_priv, dev_priv->rps.idle_freq);
6072
6073         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6074 }
6075
6076 static unsigned long intel_pxfreq(u32 vidfreq)
6077 {
6078         unsigned long freq;
6079         int div = (vidfreq & 0x3f0000) >> 16;
6080         int post = (vidfreq & 0x3000) >> 12;
6081         int pre = (vidfreq & 0x7);
6082
6083         if (!pre)
6084                 return 0;
6085
6086         freq = ((div * 133333) / ((1<<post) * pre));
6087
6088         return freq;
6089 }
6090
6091 static const struct cparams {
6092         u16 i;
6093         u16 t;
6094         u16 m;
6095         u16 c;
6096 } cparams[] = {
6097         { 1, 1333, 301, 28664 },
6098         { 1, 1066, 294, 24460 },
6099         { 1, 800, 294, 25192 },
6100         { 0, 1333, 276, 27605 },
6101         { 0, 1066, 276, 27605 },
6102         { 0, 800, 231, 23784 },
6103 };
6104
6105 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
6106 {
6107         u64 total_count, diff, ret;
6108         u32 count1, count2, count3, m = 0, c = 0;
6109         unsigned long now = jiffies_to_msecs(jiffies), diff1;
6110         int i;
6111
6112         assert_spin_locked(&mchdev_lock);
6113
6114         diff1 = now - dev_priv->ips.last_time1;
6115
6116         /* Prevent division-by-zero if we are asking too fast.
6117          * Also, we don't get interesting results if we are polling
6118          * faster than once in 10ms, so just return the saved value
6119          * in such cases.
6120          */
6121         if (diff1 <= 10)
6122                 return dev_priv->ips.chipset_power;
6123
6124         count1 = I915_READ(DMIEC);
6125         count2 = I915_READ(DDREC);
6126         count3 = I915_READ(CSIEC);
6127
6128         total_count = count1 + count2 + count3;
6129
6130         /* FIXME: handle per-counter overflow */
6131         if (total_count < dev_priv->ips.last_count1) {
6132                 diff = ~0UL - dev_priv->ips.last_count1;
6133                 diff += total_count;
6134         } else {
6135                 diff = total_count - dev_priv->ips.last_count1;
6136         }
6137
6138         for (i = 0; i < ARRAY_SIZE(cparams); i++) {
6139                 if (cparams[i].i == dev_priv->ips.c_m &&
6140                     cparams[i].t == dev_priv->ips.r_t) {
6141                         m = cparams[i].m;
6142                         c = cparams[i].c;
6143                         break;
6144                 }
6145         }
6146
6147         diff = div_u64(diff, diff1);
6148         ret = ((m * diff) + c);
6149         ret = div_u64(ret, 10);
6150
6151         dev_priv->ips.last_count1 = total_count;
6152         dev_priv->ips.last_time1 = now;
6153
6154         dev_priv->ips.chipset_power = ret;
6155
6156         return ret;
6157 }
6158
6159 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
6160 {
6161         unsigned long val;
6162
6163         if (INTEL_INFO(dev_priv)->gen != 5)
6164                 return 0;
6165
6166         spin_lock_irq(&mchdev_lock);
6167
6168         val = __i915_chipset_val(dev_priv);
6169
6170         spin_unlock_irq(&mchdev_lock);
6171
6172         return val;
6173 }
6174
6175 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
6176 {
6177         unsigned long m, x, b;
6178         u32 tsfs;
6179
6180         tsfs = I915_READ(TSFS);
6181
6182         m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
6183         x = I915_READ8(TR1);
6184
6185         b = tsfs & TSFS_INTR_MASK;
6186
6187         return ((m * x) / 127) - b;
6188 }
6189
6190 static int _pxvid_to_vd(u8 pxvid)
6191 {
6192         if (pxvid == 0)
6193                 return 0;
6194
6195         if (pxvid >= 8 && pxvid < 31)
6196                 pxvid = 31;
6197
6198         return (pxvid + 2) * 125;
6199 }
6200
6201 static u32 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
6202 {
6203         const int vd = _pxvid_to_vd(pxvid);
6204         const int vm = vd - 1125;
6205
6206         if (INTEL_INFO(dev_priv)->is_mobile)
6207                 return vm > 0 ? vm : 0;
6208
6209         return vd;
6210 }
6211
6212 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
6213 {
6214         u64 now, diff, diffms;
6215         u32 count;
6216
6217         assert_spin_locked(&mchdev_lock);
6218
6219         now = ktime_get_raw_ns();
6220         diffms = now - dev_priv->ips.last_time2;
6221         do_div(diffms, NSEC_PER_MSEC);
6222
6223         /* Don't divide by 0 */
6224         if (!diffms)
6225                 return;
6226
6227         count = I915_READ(GFXEC);
6228
6229         if (count < dev_priv->ips.last_count2) {
6230                 diff = ~0UL - dev_priv->ips.last_count2;
6231                 diff += count;
6232         } else {
6233                 diff = count - dev_priv->ips.last_count2;
6234         }
6235
6236         dev_priv->ips.last_count2 = count;
6237         dev_priv->ips.last_time2 = now;
6238
6239         /* More magic constants... */
6240         diff = diff * 1181;
6241         diff = div_u64(diff, diffms * 10);
6242         dev_priv->ips.gfx_power = diff;
6243 }
6244
6245 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
6246 {
6247         if (INTEL_INFO(dev_priv)->gen != 5)
6248                 return;
6249
6250         spin_lock_irq(&mchdev_lock);
6251
6252         __i915_update_gfx_val(dev_priv);
6253
6254         spin_unlock_irq(&mchdev_lock);
6255 }
6256
6257 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
6258 {
6259         unsigned long t, corr, state1, corr2, state2;
6260         u32 pxvid, ext_v;
6261
6262         assert_spin_locked(&mchdev_lock);
6263
6264         pxvid = I915_READ(PXVFREQ(dev_priv->rps.cur_freq));
6265         pxvid = (pxvid >> 24) & 0x7f;
6266         ext_v = pvid_to_extvid(dev_priv, pxvid);
6267
6268         state1 = ext_v;
6269
6270         t = i915_mch_val(dev_priv);
6271
6272         /* Revel in the empirically derived constants */
6273
6274         /* Correction factor in 1/100000 units */
6275         if (t > 80)
6276                 corr = ((t * 2349) + 135940);
6277         else if (t >= 50)
6278                 corr = ((t * 964) + 29317);
6279         else /* < 50 */
6280                 corr = ((t * 301) + 1004);
6281
6282         corr = corr * ((150142 * state1) / 10000 - 78642);
6283         corr /= 100000;
6284         corr2 = (corr * dev_priv->ips.corr);
6285
6286         state2 = (corr2 * state1) / 10000;
6287         state2 /= 100; /* convert to mW */
6288
6289         __i915_update_gfx_val(dev_priv);
6290
6291         return dev_priv->ips.gfx_power + state2;
6292 }
6293
6294 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
6295 {
6296         unsigned long val;
6297
6298         if (INTEL_INFO(dev_priv)->gen != 5)
6299                 return 0;
6300
6301         spin_lock_irq(&mchdev_lock);
6302
6303         val = __i915_gfx_val(dev_priv);
6304
6305         spin_unlock_irq(&mchdev_lock);
6306
6307         return val;
6308 }
6309
6310 /**
6311  * i915_read_mch_val - return value for IPS use
6312  *
6313  * Calculate and return a value for the IPS driver to use when deciding whether
6314  * we have thermal and power headroom to increase CPU or GPU power budget.
6315  */
6316 unsigned long i915_read_mch_val(void)
6317 {
6318         struct drm_i915_private *dev_priv;
6319         unsigned long chipset_val, graphics_val, ret = 0;
6320
6321         spin_lock_irq(&mchdev_lock);
6322         if (!i915_mch_dev)
6323                 goto out_unlock;
6324         dev_priv = i915_mch_dev;
6325
6326         chipset_val = __i915_chipset_val(dev_priv);
6327         graphics_val = __i915_gfx_val(dev_priv);
6328
6329         ret = chipset_val + graphics_val;
6330
6331 out_unlock:
6332         spin_unlock_irq(&mchdev_lock);
6333
6334         return ret;
6335 }
6336 EXPORT_SYMBOL_GPL(i915_read_mch_val);
6337
6338 /**
6339  * i915_gpu_raise - raise GPU frequency limit
6340  *
6341  * Raise the limit; IPS indicates we have thermal headroom.
6342  */
6343 bool i915_gpu_raise(void)
6344 {
6345         struct drm_i915_private *dev_priv;
6346         bool ret = true;
6347
6348         spin_lock_irq(&mchdev_lock);
6349         if (!i915_mch_dev) {
6350                 ret = false;
6351                 goto out_unlock;
6352         }
6353         dev_priv = i915_mch_dev;
6354
6355         if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
6356                 dev_priv->ips.max_delay--;
6357
6358 out_unlock:
6359         spin_unlock_irq(&mchdev_lock);
6360
6361         return ret;
6362 }
6363 EXPORT_SYMBOL_GPL(i915_gpu_raise);
6364
6365 /**
6366  * i915_gpu_lower - lower GPU frequency limit
6367  *
6368  * IPS indicates we're close to a thermal limit, so throttle back the GPU
6369  * frequency maximum.
6370  */
6371 bool i915_gpu_lower(void)
6372 {
6373         struct drm_i915_private *dev_priv;
6374         bool ret = true;
6375
6376         spin_lock_irq(&mchdev_lock);
6377         if (!i915_mch_dev) {
6378                 ret = false;
6379                 goto out_unlock;
6380         }
6381         dev_priv = i915_mch_dev;
6382
6383         if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
6384                 dev_priv->ips.max_delay++;
6385
6386 out_unlock:
6387         spin_unlock_irq(&mchdev_lock);
6388
6389         return ret;
6390 }
6391 EXPORT_SYMBOL_GPL(i915_gpu_lower);
6392
6393 /**
6394  * i915_gpu_busy - indicate GPU business to IPS
6395  *
6396  * Tell the IPS driver whether or not the GPU is busy.
6397  */
6398 bool i915_gpu_busy(void)
6399 {
6400         struct drm_i915_private *dev_priv;
6401         struct intel_engine_cs *engine;
6402         bool ret = false;
6403
6404         spin_lock_irq(&mchdev_lock);
6405         if (!i915_mch_dev)
6406                 goto out_unlock;
6407         dev_priv = i915_mch_dev;
6408
6409         for_each_engine(engine, dev_priv)
6410                 ret |= !list_empty(&engine->request_list);
6411
6412 out_unlock:
6413         spin_unlock_irq(&mchdev_lock);
6414
6415         return ret;
6416 }
6417 EXPORT_SYMBOL_GPL(i915_gpu_busy);
6418
6419 /**
6420  * i915_gpu_turbo_disable - disable graphics turbo
6421  *
6422  * Disable graphics turbo by resetting the max frequency and setting the
6423  * current frequency to the default.
6424  */
6425 bool i915_gpu_turbo_disable(void)
6426 {
6427         struct drm_i915_private *dev_priv;
6428         bool ret = true;
6429
6430         spin_lock_irq(&mchdev_lock);
6431         if (!i915_mch_dev) {
6432                 ret = false;
6433                 goto out_unlock;
6434         }
6435         dev_priv = i915_mch_dev;
6436
6437         dev_priv->ips.max_delay = dev_priv->ips.fstart;
6438
6439         if (!ironlake_set_drps(dev_priv, dev_priv->ips.fstart))
6440                 ret = false;
6441
6442 out_unlock:
6443         spin_unlock_irq(&mchdev_lock);
6444
6445         return ret;
6446 }
6447 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
6448
6449 /**
6450  * Tells the intel_ips driver that the i915 driver is now loaded, if
6451  * IPS got loaded first.
6452  *
6453  * This awkward dance is so that neither module has to depend on the
6454  * other in order for IPS to do the appropriate communication of
6455  * GPU turbo limits to i915.
6456  */
6457 static void
6458 ips_ping_for_i915_load(void)
6459 {
6460         void (*link)(void);
6461
6462         link = symbol_get(ips_link_to_i915_driver);
6463         if (link) {
6464                 link();
6465                 symbol_put(ips_link_to_i915_driver);
6466         }
6467 }
6468
6469 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
6470 {
6471         /* We only register the i915 ips part with intel-ips once everything is
6472          * set up, to avoid intel-ips sneaking in and reading bogus values. */
6473         spin_lock_irq(&mchdev_lock);
6474         i915_mch_dev = dev_priv;
6475         spin_unlock_irq(&mchdev_lock);
6476
6477         ips_ping_for_i915_load();
6478 }
6479
6480 void intel_gpu_ips_teardown(void)
6481 {
6482         spin_lock_irq(&mchdev_lock);
6483         i915_mch_dev = NULL;
6484         spin_unlock_irq(&mchdev_lock);
6485 }
6486
6487 static void intel_init_emon(struct drm_i915_private *dev_priv)
6488 {
6489         u32 lcfuse;
6490         u8 pxw[16];
6491         int i;
6492
6493         /* Disable to program */
6494         I915_WRITE(ECR, 0);
6495         POSTING_READ(ECR);
6496
6497         /* Program energy weights for various events */
6498         I915_WRITE(SDEW, 0x15040d00);
6499         I915_WRITE(CSIEW0, 0x007f0000);
6500         I915_WRITE(CSIEW1, 0x1e220004);
6501         I915_WRITE(CSIEW2, 0x04000004);
6502
6503         for (i = 0; i < 5; i++)
6504                 I915_WRITE(PEW(i), 0);
6505         for (i = 0; i < 3; i++)
6506                 I915_WRITE(DEW(i), 0);
6507
6508         /* Program P-state weights to account for frequency power adjustment */
6509         for (i = 0; i < 16; i++) {
6510                 u32 pxvidfreq = I915_READ(PXVFREQ(i));
6511                 unsigned long freq = intel_pxfreq(pxvidfreq);
6512                 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
6513                         PXVFREQ_PX_SHIFT;
6514                 unsigned long val;
6515
6516                 val = vid * vid;
6517                 val *= (freq / 1000);
6518                 val *= 255;
6519                 val /= (127*127*900);
6520                 if (val > 0xff)
6521                         DRM_ERROR("bad pxval: %ld\n", val);
6522                 pxw[i] = val;
6523         }
6524         /* Render standby states get 0 weight */
6525         pxw[14] = 0;
6526         pxw[15] = 0;
6527
6528         for (i = 0; i < 4; i++) {
6529                 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
6530                         (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
6531                 I915_WRITE(PXW(i), val);
6532         }
6533
6534         /* Adjust magic regs to magic values (more experimental results) */
6535         I915_WRITE(OGW0, 0);
6536         I915_WRITE(OGW1, 0);
6537         I915_WRITE(EG0, 0x00007f00);
6538         I915_WRITE(EG1, 0x0000000e);
6539         I915_WRITE(EG2, 0x000e0000);
6540         I915_WRITE(EG3, 0x68000300);
6541         I915_WRITE(EG4, 0x42000000);
6542         I915_WRITE(EG5, 0x00140031);
6543         I915_WRITE(EG6, 0);
6544         I915_WRITE(EG7, 0);
6545
6546         for (i = 0; i < 8; i++)
6547                 I915_WRITE(PXWL(i), 0);
6548
6549         /* Enable PMON + select events */
6550         I915_WRITE(ECR, 0x80000019);
6551
6552         lcfuse = I915_READ(LCFUSE02);
6553
6554         dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
6555 }
6556
6557 void intel_init_gt_powersave(struct drm_i915_private *dev_priv)
6558 {
6559         /*
6560          * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
6561          * requirement.
6562          */
6563         if (!i915.enable_rc6) {
6564                 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
6565                 intel_runtime_pm_get(dev_priv);
6566         }
6567
6568         if (IS_CHERRYVIEW(dev_priv))
6569                 cherryview_init_gt_powersave(dev_priv);
6570         else if (IS_VALLEYVIEW(dev_priv))
6571                 valleyview_init_gt_powersave(dev_priv);
6572 }
6573
6574 void intel_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
6575 {
6576         if (IS_CHERRYVIEW(dev_priv))
6577                 return;
6578         else if (IS_VALLEYVIEW(dev_priv))
6579                 valleyview_cleanup_gt_powersave(dev_priv);
6580
6581         if (!i915.enable_rc6)
6582                 intel_runtime_pm_put(dev_priv);
6583 }
6584
6585 static void gen6_suspend_rps(struct drm_i915_private *dev_priv)
6586 {
6587         flush_delayed_work(&dev_priv->rps.delayed_resume_work);
6588
6589         gen6_disable_rps_interrupts(dev_priv);
6590 }
6591
6592 /**
6593  * intel_suspend_gt_powersave - suspend PM work and helper threads
6594  * @dev_priv: i915 device
6595  *
6596  * We don't want to disable RC6 or other features here, we just want
6597  * to make sure any work we've queued has finished and won't bother
6598  * us while we're suspended.
6599  */
6600 void intel_suspend_gt_powersave(struct drm_i915_private *dev_priv)
6601 {
6602         if (INTEL_GEN(dev_priv) < 6)
6603                 return;
6604
6605         gen6_suspend_rps(dev_priv);
6606
6607         /* Force GPU to min freq during suspend */
6608         gen6_rps_idle(dev_priv);
6609 }
6610
6611 void intel_disable_gt_powersave(struct drm_i915_private *dev_priv)
6612 {
6613         if (IS_IRONLAKE_M(dev_priv)) {
6614                 ironlake_disable_drps(dev_priv);
6615         } else if (INTEL_INFO(dev_priv)->gen >= 6) {
6616                 intel_suspend_gt_powersave(dev_priv);
6617
6618                 mutex_lock(&dev_priv->rps.hw_lock);
6619                 if (INTEL_INFO(dev_priv)->gen >= 9) {
6620                         gen9_disable_rc6(dev_priv);
6621                         gen9_disable_rps(dev_priv);
6622                 } else if (IS_CHERRYVIEW(dev_priv))
6623                         cherryview_disable_rps(dev_priv);
6624                 else if (IS_VALLEYVIEW(dev_priv))
6625                         valleyview_disable_rps(dev_priv);
6626                 else
6627                         gen6_disable_rps(dev_priv);
6628
6629                 dev_priv->rps.enabled = false;
6630                 mutex_unlock(&dev_priv->rps.hw_lock);
6631         }
6632 }
6633
6634 static void intel_gen6_powersave_work(struct work_struct *work)
6635 {
6636         struct drm_i915_private *dev_priv =
6637                 container_of(work, struct drm_i915_private,
6638                              rps.delayed_resume_work.work);
6639
6640         mutex_lock(&dev_priv->rps.hw_lock);
6641
6642         gen6_reset_rps_interrupts(dev_priv);
6643
6644         if (IS_CHERRYVIEW(dev_priv)) {
6645                 cherryview_enable_rps(dev_priv);
6646         } else if (IS_VALLEYVIEW(dev_priv)) {
6647                 valleyview_enable_rps(dev_priv);
6648         } else if (INTEL_INFO(dev_priv)->gen >= 9) {
6649                 gen9_enable_rc6(dev_priv);
6650                 gen9_enable_rps(dev_priv);
6651                 if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv))
6652                         __gen6_update_ring_freq(dev_priv);
6653         } else if (IS_BROADWELL(dev_priv)) {
6654                 gen8_enable_rps(dev_priv);
6655                 __gen6_update_ring_freq(dev_priv);
6656         } else {
6657                 gen6_enable_rps(dev_priv);
6658                 __gen6_update_ring_freq(dev_priv);
6659         }
6660
6661         WARN_ON(dev_priv->rps.max_freq < dev_priv->rps.min_freq);
6662         WARN_ON(dev_priv->rps.idle_freq > dev_priv->rps.max_freq);
6663
6664         WARN_ON(dev_priv->rps.efficient_freq < dev_priv->rps.min_freq);
6665         WARN_ON(dev_priv->rps.efficient_freq > dev_priv->rps.max_freq);
6666
6667         dev_priv->rps.enabled = true;
6668
6669         gen6_enable_rps_interrupts(dev_priv);
6670
6671         mutex_unlock(&dev_priv->rps.hw_lock);
6672
6673         intel_runtime_pm_put(dev_priv);
6674 }
6675
6676 void intel_enable_gt_powersave(struct drm_i915_private *dev_priv)
6677 {
6678         /* Powersaving is controlled by the host when inside a VM */
6679         if (intel_vgpu_active(dev_priv))
6680                 return;
6681
6682         if (IS_IRONLAKE_M(dev_priv)) {
6683                 ironlake_enable_drps(dev_priv);
6684                 mutex_lock(&dev_priv->drm.struct_mutex);
6685                 intel_init_emon(dev_priv);
6686                 mutex_unlock(&dev_priv->drm.struct_mutex);
6687         } else if (INTEL_INFO(dev_priv)->gen >= 6) {
6688                 /*
6689                  * PCU communication is slow and this doesn't need to be
6690                  * done at any specific time, so do this out of our fast path
6691                  * to make resume and init faster.
6692                  *
6693                  * We depend on the HW RC6 power context save/restore
6694                  * mechanism when entering D3 through runtime PM suspend. So
6695                  * disable RPM until RPS/RC6 is properly setup. We can only
6696                  * get here via the driver load/system resume/runtime resume
6697                  * paths, so the _noresume version is enough (and in case of
6698                  * runtime resume it's necessary).
6699                  */
6700                 if (schedule_delayed_work(&dev_priv->rps.delayed_resume_work,
6701                                            round_jiffies_up_relative(HZ)))
6702                         intel_runtime_pm_get_noresume(dev_priv);
6703         }
6704 }
6705
6706 void intel_reset_gt_powersave(struct drm_i915_private *dev_priv)
6707 {
6708         if (INTEL_INFO(dev_priv)->gen < 6)
6709                 return;
6710
6711         gen6_suspend_rps(dev_priv);
6712         dev_priv->rps.enabled = false;
6713 }
6714
6715 static void ibx_init_clock_gating(struct drm_device *dev)
6716 {
6717         struct drm_i915_private *dev_priv = to_i915(dev);
6718
6719         /*
6720          * On Ibex Peak and Cougar Point, we need to disable clock
6721          * gating for the panel power sequencer or it will fail to
6722          * start up when no ports are active.
6723          */
6724         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
6725 }
6726
6727 static void g4x_disable_trickle_feed(struct drm_device *dev)
6728 {
6729         struct drm_i915_private *dev_priv = to_i915(dev);
6730         enum pipe pipe;
6731
6732         for_each_pipe(dev_priv, pipe) {
6733                 I915_WRITE(DSPCNTR(pipe),
6734                            I915_READ(DSPCNTR(pipe)) |
6735                            DISPPLANE_TRICKLE_FEED_DISABLE);
6736
6737                 I915_WRITE(DSPSURF(pipe), I915_READ(DSPSURF(pipe)));
6738                 POSTING_READ(DSPSURF(pipe));
6739         }
6740 }
6741
6742 static void ilk_init_lp_watermarks(struct drm_device *dev)
6743 {
6744         struct drm_i915_private *dev_priv = to_i915(dev);
6745
6746         I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
6747         I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
6748         I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
6749
6750         /*
6751          * Don't touch WM1S_LP_EN here.
6752          * Doing so could cause underruns.
6753          */
6754 }
6755
6756 static void ironlake_init_clock_gating(struct drm_device *dev)
6757 {
6758         struct drm_i915_private *dev_priv = to_i915(dev);
6759         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6760
6761         /*
6762          * Required for FBC
6763          * WaFbcDisableDpfcClockGating:ilk
6764          */
6765         dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
6766                    ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
6767                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
6768
6769         I915_WRITE(PCH_3DCGDIS0,
6770                    MARIUNIT_CLOCK_GATE_DISABLE |
6771                    SVSMUNIT_CLOCK_GATE_DISABLE);
6772         I915_WRITE(PCH_3DCGDIS1,
6773                    VFMUNIT_CLOCK_GATE_DISABLE);
6774
6775         /*
6776          * According to the spec the following bits should be set in
6777          * order to enable memory self-refresh
6778          * The bit 22/21 of 0x42004
6779          * The bit 5 of 0x42020
6780          * The bit 15 of 0x45000
6781          */
6782         I915_WRITE(ILK_DISPLAY_CHICKEN2,
6783                    (I915_READ(ILK_DISPLAY_CHICKEN2) |
6784                     ILK_DPARB_GATE | ILK_VSDPFD_FULL));
6785         dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
6786         I915_WRITE(DISP_ARB_CTL,
6787                    (I915_READ(DISP_ARB_CTL) |
6788                     DISP_FBC_WM_DIS));
6789
6790         ilk_init_lp_watermarks(dev);
6791
6792         /*
6793          * Based on the document from hardware guys the following bits
6794          * should be set unconditionally in order to enable FBC.
6795          * The bit 22 of 0x42000
6796          * The bit 22 of 0x42004
6797          * The bit 7,8,9 of 0x42020.
6798          */
6799         if (IS_IRONLAKE_M(dev)) {
6800                 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
6801                 I915_WRITE(ILK_DISPLAY_CHICKEN1,
6802                            I915_READ(ILK_DISPLAY_CHICKEN1) |
6803                            ILK_FBCQ_DIS);
6804                 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6805                            I915_READ(ILK_DISPLAY_CHICKEN2) |
6806                            ILK_DPARB_GATE);
6807         }
6808
6809         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
6810
6811         I915_WRITE(ILK_DISPLAY_CHICKEN2,
6812                    I915_READ(ILK_DISPLAY_CHICKEN2) |
6813                    ILK_ELPIN_409_SELECT);
6814         I915_WRITE(_3D_CHICKEN2,
6815                    _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
6816                    _3D_CHICKEN2_WM_READ_PIPELINED);
6817
6818         /* WaDisableRenderCachePipelinedFlush:ilk */
6819         I915_WRITE(CACHE_MODE_0,
6820                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
6821
6822         /* WaDisable_RenderCache_OperationalFlush:ilk */
6823         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6824
6825         g4x_disable_trickle_feed(dev);
6826
6827         ibx_init_clock_gating(dev);
6828 }
6829
6830 static void cpt_init_clock_gating(struct drm_device *dev)
6831 {
6832         struct drm_i915_private *dev_priv = to_i915(dev);
6833         int pipe;
6834         uint32_t val;
6835
6836         /*
6837          * On Ibex Peak and Cougar Point, we need to disable clock
6838          * gating for the panel power sequencer or it will fail to
6839          * start up when no ports are active.
6840          */
6841         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
6842                    PCH_DPLUNIT_CLOCK_GATE_DISABLE |
6843                    PCH_CPUNIT_CLOCK_GATE_DISABLE);
6844         I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
6845                    DPLS_EDP_PPS_FIX_DIS);
6846         /* The below fixes the weird display corruption, a few pixels shifted
6847          * downward, on (only) LVDS of some HP laptops with IVY.
6848          */
6849         for_each_pipe(dev_priv, pipe) {
6850                 val = I915_READ(TRANS_CHICKEN2(pipe));
6851                 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
6852                 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6853                 if (dev_priv->vbt.fdi_rx_polarity_inverted)
6854                         val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6855                 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
6856                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
6857                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
6858                 I915_WRITE(TRANS_CHICKEN2(pipe), val);
6859         }
6860         /* WADP0ClockGatingDisable */
6861         for_each_pipe(dev_priv, pipe) {
6862                 I915_WRITE(TRANS_CHICKEN1(pipe),
6863                            TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
6864         }
6865 }
6866
6867 static void gen6_check_mch_setup(struct drm_device *dev)
6868 {
6869         struct drm_i915_private *dev_priv = to_i915(dev);
6870         uint32_t tmp;
6871
6872         tmp = I915_READ(MCH_SSKPD);
6873         if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL)
6874                 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
6875                               tmp);
6876 }
6877
6878 static void gen6_init_clock_gating(struct drm_device *dev)
6879 {
6880         struct drm_i915_private *dev_priv = to_i915(dev);
6881         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6882
6883         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
6884
6885         I915_WRITE(ILK_DISPLAY_CHICKEN2,
6886                    I915_READ(ILK_DISPLAY_CHICKEN2) |
6887                    ILK_ELPIN_409_SELECT);
6888
6889         /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
6890         I915_WRITE(_3D_CHICKEN,
6891                    _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
6892
6893         /* WaDisable_RenderCache_OperationalFlush:snb */
6894         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6895
6896         /*
6897          * BSpec recoomends 8x4 when MSAA is used,
6898          * however in practice 16x4 seems fastest.
6899          *
6900          * Note that PS/WM thread counts depend on the WIZ hashing
6901          * disable bit, which we don't touch here, but it's good
6902          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6903          */
6904         I915_WRITE(GEN6_GT_MODE,
6905                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6906
6907         ilk_init_lp_watermarks(dev);
6908
6909         I915_WRITE(CACHE_MODE_0,
6910                    _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
6911
6912         I915_WRITE(GEN6_UCGCTL1,
6913                    I915_READ(GEN6_UCGCTL1) |
6914                    GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
6915                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
6916
6917         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
6918          * gating disable must be set.  Failure to set it results in
6919          * flickering pixels due to Z write ordering failures after
6920          * some amount of runtime in the Mesa "fire" demo, and Unigine
6921          * Sanctuary and Tropics, and apparently anything else with
6922          * alpha test or pixel discard.
6923          *
6924          * According to the spec, bit 11 (RCCUNIT) must also be set,
6925          * but we didn't debug actual testcases to find it out.
6926          *
6927          * WaDisableRCCUnitClockGating:snb
6928          * WaDisableRCPBUnitClockGating:snb
6929          */
6930         I915_WRITE(GEN6_UCGCTL2,
6931                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
6932                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
6933
6934         /* WaStripsFansDisableFastClipPerformanceFix:snb */
6935         I915_WRITE(_3D_CHICKEN3,
6936                    _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
6937
6938         /*
6939          * Bspec says:
6940          * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
6941          * 3DSTATE_SF number of SF output attributes is more than 16."
6942          */
6943         I915_WRITE(_3D_CHICKEN3,
6944                    _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
6945
6946         /*
6947          * According to the spec the following bits should be
6948          * set in order to enable memory self-refresh and fbc:
6949          * The bit21 and bit22 of 0x42000
6950          * The bit21 and bit22 of 0x42004
6951          * The bit5 and bit7 of 0x42020
6952          * The bit14 of 0x70180
6953          * The bit14 of 0x71180
6954          *
6955          * WaFbcAsynchFlipDisableFbcQueue:snb
6956          */
6957         I915_WRITE(ILK_DISPLAY_CHICKEN1,
6958                    I915_READ(ILK_DISPLAY_CHICKEN1) |
6959                    ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
6960         I915_WRITE(ILK_DISPLAY_CHICKEN2,
6961                    I915_READ(ILK_DISPLAY_CHICKEN2) |
6962                    ILK_DPARB_GATE | ILK_VSDPFD_FULL);
6963         I915_WRITE(ILK_DSPCLK_GATE_D,
6964                    I915_READ(ILK_DSPCLK_GATE_D) |
6965                    ILK_DPARBUNIT_CLOCK_GATE_ENABLE  |
6966                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
6967
6968         g4x_disable_trickle_feed(dev);
6969
6970         cpt_init_clock_gating(dev);
6971
6972         gen6_check_mch_setup(dev);
6973 }
6974
6975 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
6976 {
6977         uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
6978
6979         /*
6980          * WaVSThreadDispatchOverride:ivb,vlv
6981          *
6982          * This actually overrides the dispatch
6983          * mode for all thread types.
6984          */
6985         reg &= ~GEN7_FF_SCHED_MASK;
6986         reg |= GEN7_FF_TS_SCHED_HW;
6987         reg |= GEN7_FF_VS_SCHED_HW;
6988         reg |= GEN7_FF_DS_SCHED_HW;
6989
6990         I915_WRITE(GEN7_FF_THREAD_MODE, reg);
6991 }
6992
6993 static void lpt_init_clock_gating(struct drm_device *dev)
6994 {
6995         struct drm_i915_private *dev_priv = to_i915(dev);
6996
6997         /*
6998          * TODO: this bit should only be enabled when really needed, then
6999          * disabled when not needed anymore in order to save power.
7000          */
7001         if (HAS_PCH_LPT_LP(dev))
7002                 I915_WRITE(SOUTH_DSPCLK_GATE_D,
7003                            I915_READ(SOUTH_DSPCLK_GATE_D) |
7004                            PCH_LP_PARTITION_LEVEL_DISABLE);
7005
7006         /* WADPOClockGatingDisable:hsw */
7007         I915_WRITE(TRANS_CHICKEN1(PIPE_A),
7008                    I915_READ(TRANS_CHICKEN1(PIPE_A)) |
7009                    TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
7010 }
7011
7012 static void lpt_suspend_hw(struct drm_device *dev)
7013 {
7014         struct drm_i915_private *dev_priv = to_i915(dev);
7015
7016         if (HAS_PCH_LPT_LP(dev)) {
7017                 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
7018
7019                 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
7020                 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
7021         }
7022 }
7023
7024 static void gen8_set_l3sqc_credits(struct drm_i915_private *dev_priv,
7025                                    int general_prio_credits,
7026                                    int high_prio_credits)
7027 {
7028         u32 misccpctl;
7029
7030         /* WaTempDisableDOPClkGating:bdw */
7031         misccpctl = I915_READ(GEN7_MISCCPCTL);
7032         I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
7033
7034         I915_WRITE(GEN8_L3SQCREG1,
7035                    L3_GENERAL_PRIO_CREDITS(general_prio_credits) |
7036                    L3_HIGH_PRIO_CREDITS(high_prio_credits));
7037
7038         /*
7039          * Wait at least 100 clocks before re-enabling clock gating.
7040          * See the definition of L3SQCREG1 in BSpec.
7041          */
7042         POSTING_READ(GEN8_L3SQCREG1);
7043         udelay(1);
7044         I915_WRITE(GEN7_MISCCPCTL, misccpctl);
7045 }
7046
7047 static void kabylake_init_clock_gating(struct drm_device *dev)
7048 {
7049         struct drm_i915_private *dev_priv = dev->dev_private;
7050
7051         gen9_init_clock_gating(dev);
7052
7053         /* WaDisableSDEUnitClockGating:kbl */
7054         if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
7055                 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7056                            GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7057
7058         /* WaDisableGamClockGating:kbl */
7059         if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
7060                 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
7061                            GEN6_GAMUNIT_CLOCK_GATE_DISABLE);
7062
7063         /* WaFbcNukeOnHostModify:kbl */
7064         I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
7065                    ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
7066 }
7067
7068 static void skylake_init_clock_gating(struct drm_device *dev)
7069 {
7070         struct drm_i915_private *dev_priv = dev->dev_private;
7071
7072         gen9_init_clock_gating(dev);
7073
7074         /* WAC6entrylatency:skl */
7075         I915_WRITE(FBC_LLC_READ_CTRL, I915_READ(FBC_LLC_READ_CTRL) |
7076                    FBC_LLC_FULLY_OPEN);
7077
7078         /* WaFbcNukeOnHostModify:skl */
7079         I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
7080                    ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
7081 }
7082
7083 static void broadwell_init_clock_gating(struct drm_device *dev)
7084 {
7085         struct drm_i915_private *dev_priv = to_i915(dev);
7086         enum pipe pipe;
7087
7088         ilk_init_lp_watermarks(dev);
7089
7090         /* WaSwitchSolVfFArbitrationPriority:bdw */
7091         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
7092
7093         /* WaPsrDPAMaskVBlankInSRD:bdw */
7094         I915_WRITE(CHICKEN_PAR1_1,
7095                    I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
7096
7097         /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
7098         for_each_pipe(dev_priv, pipe) {
7099                 I915_WRITE(CHICKEN_PIPESL_1(pipe),
7100                            I915_READ(CHICKEN_PIPESL_1(pipe)) |
7101                            BDW_DPRS_MASK_VBLANK_SRD);
7102         }
7103
7104         /* WaVSRefCountFullforceMissDisable:bdw */
7105         /* WaDSRefCountFullforceMissDisable:bdw */
7106         I915_WRITE(GEN7_FF_THREAD_MODE,
7107                    I915_READ(GEN7_FF_THREAD_MODE) &
7108                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
7109
7110         I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
7111                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
7112
7113         /* WaDisableSDEUnitClockGating:bdw */
7114         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7115                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7116
7117         /* WaProgramL3SqcReg1Default:bdw */
7118         gen8_set_l3sqc_credits(dev_priv, 30, 2);
7119
7120         /*
7121          * WaGttCachingOffByDefault:bdw
7122          * GTT cache may not work with big pages, so if those
7123          * are ever enabled GTT cache may need to be disabled.
7124          */
7125         I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
7126
7127         /* WaKVMNotificationOnConfigChange:bdw */
7128         I915_WRITE(CHICKEN_PAR2_1, I915_READ(CHICKEN_PAR2_1)
7129                    | KVM_CONFIG_CHANGE_NOTIFICATION_SELECT);
7130
7131         lpt_init_clock_gating(dev);
7132 }
7133
7134 static void haswell_init_clock_gating(struct drm_device *dev)
7135 {
7136         struct drm_i915_private *dev_priv = to_i915(dev);
7137
7138         ilk_init_lp_watermarks(dev);
7139
7140         /* L3 caching of data atomics doesn't work -- disable it. */
7141         I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
7142         I915_WRITE(HSW_ROW_CHICKEN3,
7143                    _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
7144
7145         /* This is required by WaCatErrorRejectionIssue:hsw */
7146         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7147                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7148                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7149
7150         /* WaVSRefCountFullforceMissDisable:hsw */
7151         I915_WRITE(GEN7_FF_THREAD_MODE,
7152                    I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
7153
7154         /* WaDisable_RenderCache_OperationalFlush:hsw */
7155         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7156
7157         /* enable HiZ Raw Stall Optimization */
7158         I915_WRITE(CACHE_MODE_0_GEN7,
7159                    _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
7160
7161         /* WaDisable4x2SubspanOptimization:hsw */
7162         I915_WRITE(CACHE_MODE_1,
7163                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7164
7165         /*
7166          * BSpec recommends 8x4 when MSAA is used,
7167          * however in practice 16x4 seems fastest.
7168          *
7169          * Note that PS/WM thread counts depend on the WIZ hashing
7170          * disable bit, which we don't touch here, but it's good
7171          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7172          */
7173         I915_WRITE(GEN7_GT_MODE,
7174                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7175
7176         /* WaSampleCChickenBitEnable:hsw */
7177         I915_WRITE(HALF_SLICE_CHICKEN3,
7178                    _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE));
7179
7180         /* WaSwitchSolVfFArbitrationPriority:hsw */
7181         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
7182
7183         /* WaRsPkgCStateDisplayPMReq:hsw */
7184         I915_WRITE(CHICKEN_PAR1_1,
7185                    I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
7186
7187         lpt_init_clock_gating(dev);
7188 }
7189
7190 static void ivybridge_init_clock_gating(struct drm_device *dev)
7191 {
7192         struct drm_i915_private *dev_priv = to_i915(dev);
7193         uint32_t snpcr;
7194
7195         ilk_init_lp_watermarks(dev);
7196
7197         I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
7198
7199         /* WaDisableEarlyCull:ivb */
7200         I915_WRITE(_3D_CHICKEN3,
7201                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
7202
7203         /* WaDisableBackToBackFlipFix:ivb */
7204         I915_WRITE(IVB_CHICKEN3,
7205                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7206                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
7207
7208         /* WaDisablePSDDualDispatchEnable:ivb */
7209         if (IS_IVB_GT1(dev))
7210                 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
7211                            _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
7212
7213         /* WaDisable_RenderCache_OperationalFlush:ivb */
7214         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7215
7216         /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
7217         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
7218                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
7219
7220         /* WaApplyL3ControlAndL3ChickenMode:ivb */
7221         I915_WRITE(GEN7_L3CNTLREG1,
7222                         GEN7_WA_FOR_GEN7_L3_CONTROL);
7223         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
7224                    GEN7_WA_L3_CHICKEN_MODE);
7225         if (IS_IVB_GT1(dev))
7226                 I915_WRITE(GEN7_ROW_CHICKEN2,
7227                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7228         else {
7229                 /* must write both registers */
7230                 I915_WRITE(GEN7_ROW_CHICKEN2,
7231                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7232                 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
7233                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7234         }
7235
7236         /* WaForceL3Serialization:ivb */
7237         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
7238                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
7239
7240         /*
7241          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7242          * This implements the WaDisableRCZUnitClockGating:ivb workaround.
7243          */
7244         I915_WRITE(GEN6_UCGCTL2,
7245                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7246
7247         /* This is required by WaCatErrorRejectionIssue:ivb */
7248         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7249                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7250                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7251
7252         g4x_disable_trickle_feed(dev);
7253
7254         gen7_setup_fixed_func_scheduler(dev_priv);
7255
7256         if (0) { /* causes HiZ corruption on ivb:gt1 */
7257                 /* enable HiZ Raw Stall Optimization */
7258                 I915_WRITE(CACHE_MODE_0_GEN7,
7259                            _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
7260         }
7261
7262         /* WaDisable4x2SubspanOptimization:ivb */
7263         I915_WRITE(CACHE_MODE_1,
7264                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7265
7266         /*
7267          * BSpec recommends 8x4 when MSAA is used,
7268          * however in practice 16x4 seems fastest.
7269          *
7270          * Note that PS/WM thread counts depend on the WIZ hashing
7271          * disable bit, which we don't touch here, but it's good
7272          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7273          */
7274         I915_WRITE(GEN7_GT_MODE,
7275                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7276
7277         snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
7278         snpcr &= ~GEN6_MBC_SNPCR_MASK;
7279         snpcr |= GEN6_MBC_SNPCR_MED;
7280         I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
7281
7282         if (!HAS_PCH_NOP(dev))
7283                 cpt_init_clock_gating(dev);
7284
7285         gen6_check_mch_setup(dev);
7286 }
7287
7288 static void valleyview_init_clock_gating(struct drm_device *dev)
7289 {
7290         struct drm_i915_private *dev_priv = to_i915(dev);
7291
7292         /* WaDisableEarlyCull:vlv */
7293         I915_WRITE(_3D_CHICKEN3,
7294                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
7295
7296         /* WaDisableBackToBackFlipFix:vlv */
7297         I915_WRITE(IVB_CHICKEN3,
7298                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7299                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
7300
7301         /* WaPsdDispatchEnable:vlv */
7302         /* WaDisablePSDDualDispatchEnable:vlv */
7303         I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
7304                    _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
7305                                       GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
7306
7307         /* WaDisable_RenderCache_OperationalFlush:vlv */
7308         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7309
7310         /* WaForceL3Serialization:vlv */
7311         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
7312                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
7313
7314         /* WaDisableDopClockGating:vlv */
7315         I915_WRITE(GEN7_ROW_CHICKEN2,
7316                    _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7317
7318         /* This is required by WaCatErrorRejectionIssue:vlv */
7319         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7320                    I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7321                    GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7322
7323         gen7_setup_fixed_func_scheduler(dev_priv);
7324
7325         /*
7326          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7327          * This implements the WaDisableRCZUnitClockGating:vlv workaround.
7328          */
7329         I915_WRITE(GEN6_UCGCTL2,
7330                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7331
7332         /* WaDisableL3Bank2xClockGate:vlv
7333          * Disabling L3 clock gating- MMIO 940c[25] = 1
7334          * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
7335         I915_WRITE(GEN7_UCGCTL4,
7336                    I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
7337
7338         /*
7339          * BSpec says this must be set, even though
7340          * WaDisable4x2SubspanOptimization isn't listed for VLV.
7341          */
7342         I915_WRITE(CACHE_MODE_1,
7343                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7344
7345         /*
7346          * BSpec recommends 8x4 when MSAA is used,
7347          * however in practice 16x4 seems fastest.
7348          *
7349          * Note that PS/WM thread counts depend on the WIZ hashing
7350          * disable bit, which we don't touch here, but it's good
7351          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7352          */
7353         I915_WRITE(GEN7_GT_MODE,
7354                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7355
7356         /*
7357          * WaIncreaseL3CreditsForVLVB0:vlv
7358          * This is the hardware default actually.
7359          */
7360         I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
7361
7362         /*
7363          * WaDisableVLVClockGating_VBIIssue:vlv
7364          * Disable clock gating on th GCFG unit to prevent a delay
7365          * in the reporting of vblank events.
7366          */
7367         I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
7368 }
7369
7370 static void cherryview_init_clock_gating(struct drm_device *dev)
7371 {
7372         struct drm_i915_private *dev_priv = to_i915(dev);
7373
7374         /* WaVSRefCountFullforceMissDisable:chv */
7375         /* WaDSRefCountFullforceMissDisable:chv */
7376         I915_WRITE(GEN7_FF_THREAD_MODE,
7377                    I915_READ(GEN7_FF_THREAD_MODE) &
7378                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
7379
7380         /* WaDisableSemaphoreAndSyncFlipWait:chv */
7381         I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
7382                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
7383
7384         /* WaDisableCSUnitClockGating:chv */
7385         I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
7386                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
7387
7388         /* WaDisableSDEUnitClockGating:chv */
7389         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7390                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7391
7392         /*
7393          * WaProgramL3SqcReg1Default:chv
7394          * See gfxspecs/Related Documents/Performance Guide/
7395          * LSQC Setting Recommendations.
7396          */
7397         gen8_set_l3sqc_credits(dev_priv, 38, 2);
7398
7399         /*
7400          * GTT cache may not work with big pages, so if those
7401          * are ever enabled GTT cache may need to be disabled.
7402          */
7403         I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
7404 }
7405
7406 static void g4x_init_clock_gating(struct drm_device *dev)
7407 {
7408         struct drm_i915_private *dev_priv = to_i915(dev);
7409         uint32_t dspclk_gate;
7410
7411         I915_WRITE(RENCLK_GATE_D1, 0);
7412         I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
7413                    GS_UNIT_CLOCK_GATE_DISABLE |
7414                    CL_UNIT_CLOCK_GATE_DISABLE);
7415         I915_WRITE(RAMCLK_GATE_D, 0);
7416         dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
7417                 OVRUNIT_CLOCK_GATE_DISABLE |
7418                 OVCUNIT_CLOCK_GATE_DISABLE;
7419         if (IS_GM45(dev))
7420                 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
7421         I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
7422
7423         /* WaDisableRenderCachePipelinedFlush */
7424         I915_WRITE(CACHE_MODE_0,
7425                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
7426
7427         /* WaDisable_RenderCache_OperationalFlush:g4x */
7428         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7429
7430         g4x_disable_trickle_feed(dev);
7431 }
7432
7433 static void crestline_init_clock_gating(struct drm_device *dev)
7434 {
7435         struct drm_i915_private *dev_priv = to_i915(dev);
7436
7437         I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
7438         I915_WRITE(RENCLK_GATE_D2, 0);
7439         I915_WRITE(DSPCLK_GATE_D, 0);
7440         I915_WRITE(RAMCLK_GATE_D, 0);
7441         I915_WRITE16(DEUC, 0);
7442         I915_WRITE(MI_ARB_STATE,
7443                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7444
7445         /* WaDisable_RenderCache_OperationalFlush:gen4 */
7446         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7447 }
7448
7449 static void broadwater_init_clock_gating(struct drm_device *dev)
7450 {
7451         struct drm_i915_private *dev_priv = to_i915(dev);
7452
7453         I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
7454                    I965_RCC_CLOCK_GATE_DISABLE |
7455                    I965_RCPB_CLOCK_GATE_DISABLE |
7456                    I965_ISC_CLOCK_GATE_DISABLE |
7457                    I965_FBC_CLOCK_GATE_DISABLE);
7458         I915_WRITE(RENCLK_GATE_D2, 0);
7459         I915_WRITE(MI_ARB_STATE,
7460                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7461
7462         /* WaDisable_RenderCache_OperationalFlush:gen4 */
7463         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7464 }
7465
7466 static void gen3_init_clock_gating(struct drm_device *dev)
7467 {
7468         struct drm_i915_private *dev_priv = to_i915(dev);
7469         u32 dstate = I915_READ(D_STATE);
7470
7471         dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
7472                 DSTATE_DOT_CLOCK_GATING;
7473         I915_WRITE(D_STATE, dstate);
7474
7475         if (IS_PINEVIEW(dev))
7476                 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
7477
7478         /* IIR "flip pending" means done if this bit is set */
7479         I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
7480
7481         /* interrupts should cause a wake up from C3 */
7482         I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
7483
7484         /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
7485         I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
7486
7487         I915_WRITE(MI_ARB_STATE,
7488                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7489 }
7490
7491 static void i85x_init_clock_gating(struct drm_device *dev)
7492 {
7493         struct drm_i915_private *dev_priv = to_i915(dev);
7494
7495         I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
7496
7497         /* interrupts should cause a wake up from C3 */
7498         I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
7499                    _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
7500
7501         I915_WRITE(MEM_MODE,
7502                    _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE));
7503 }
7504
7505 static void i830_init_clock_gating(struct drm_device *dev)
7506 {
7507         struct drm_i915_private *dev_priv = to_i915(dev);
7508
7509         I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
7510
7511         I915_WRITE(MEM_MODE,
7512                    _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE) |
7513                    _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE));
7514 }
7515
7516 void intel_init_clock_gating(struct drm_device *dev)
7517 {
7518         struct drm_i915_private *dev_priv = to_i915(dev);
7519
7520         dev_priv->display.init_clock_gating(dev);
7521 }
7522
7523 void intel_suspend_hw(struct drm_device *dev)
7524 {
7525         if (HAS_PCH_LPT(dev))
7526                 lpt_suspend_hw(dev);
7527 }
7528
7529 static void nop_init_clock_gating(struct drm_device *dev)
7530 {
7531         DRM_DEBUG_KMS("No clock gating settings or workarounds applied.\n");
7532 }
7533
7534 /**
7535  * intel_init_clock_gating_hooks - setup the clock gating hooks
7536  * @dev_priv: device private
7537  *
7538  * Setup the hooks that configure which clocks of a given platform can be
7539  * gated and also apply various GT and display specific workarounds for these
7540  * platforms. Note that some GT specific workarounds are applied separately
7541  * when GPU contexts or batchbuffers start their execution.
7542  */
7543 void intel_init_clock_gating_hooks(struct drm_i915_private *dev_priv)
7544 {
7545         if (IS_SKYLAKE(dev_priv))
7546                 dev_priv->display.init_clock_gating = skylake_init_clock_gating;
7547         else if (IS_KABYLAKE(dev_priv))
7548                 dev_priv->display.init_clock_gating = kabylake_init_clock_gating;
7549         else if (IS_BROXTON(dev_priv))
7550                 dev_priv->display.init_clock_gating = bxt_init_clock_gating;
7551         else if (IS_BROADWELL(dev_priv))
7552                 dev_priv->display.init_clock_gating = broadwell_init_clock_gating;
7553         else if (IS_CHERRYVIEW(dev_priv))
7554                 dev_priv->display.init_clock_gating = cherryview_init_clock_gating;
7555         else if (IS_HASWELL(dev_priv))
7556                 dev_priv->display.init_clock_gating = haswell_init_clock_gating;
7557         else if (IS_IVYBRIDGE(dev_priv))
7558                 dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
7559         else if (IS_VALLEYVIEW(dev_priv))
7560                 dev_priv->display.init_clock_gating = valleyview_init_clock_gating;
7561         else if (IS_GEN6(dev_priv))
7562                 dev_priv->display.init_clock_gating = gen6_init_clock_gating;
7563         else if (IS_GEN5(dev_priv))
7564                 dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
7565         else if (IS_G4X(dev_priv))
7566                 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
7567         else if (IS_CRESTLINE(dev_priv))
7568                 dev_priv->display.init_clock_gating = crestline_init_clock_gating;
7569         else if (IS_BROADWATER(dev_priv))
7570                 dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
7571         else if (IS_GEN3(dev_priv))
7572                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
7573         else if (IS_I85X(dev_priv) || IS_I865G(dev_priv))
7574                 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
7575         else if (IS_GEN2(dev_priv))
7576                 dev_priv->display.init_clock_gating = i830_init_clock_gating;
7577         else {
7578                 MISSING_CASE(INTEL_DEVID(dev_priv));
7579                 dev_priv->display.init_clock_gating = nop_init_clock_gating;
7580         }
7581 }
7582
7583 /* Set up chip specific power management-related functions */
7584 void intel_init_pm(struct drm_device *dev)
7585 {
7586         struct drm_i915_private *dev_priv = to_i915(dev);
7587
7588         intel_fbc_init(dev_priv);
7589
7590         /* For cxsr */
7591         if (IS_PINEVIEW(dev))
7592                 i915_pineview_get_mem_freq(dev);
7593         else if (IS_GEN5(dev))
7594                 i915_ironlake_get_mem_freq(dev);
7595
7596         /* For FIFO watermark updates */
7597         if (INTEL_INFO(dev)->gen >= 9) {
7598                 skl_setup_wm_latency(dev);
7599                 dev_priv->display.update_wm = skl_update_wm;
7600                 dev_priv->display.compute_global_watermarks = skl_compute_wm;
7601         } else if (HAS_PCH_SPLIT(dev)) {
7602                 ilk_setup_wm_latency(dev);
7603
7604                 if ((IS_GEN5(dev) && dev_priv->wm.pri_latency[1] &&
7605                      dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
7606                     (!IS_GEN5(dev) && dev_priv->wm.pri_latency[0] &&
7607                      dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
7608                         dev_priv->display.compute_pipe_wm = ilk_compute_pipe_wm;
7609                         dev_priv->display.compute_intermediate_wm =
7610                                 ilk_compute_intermediate_wm;
7611                         dev_priv->display.initial_watermarks =
7612                                 ilk_initial_watermarks;
7613                         dev_priv->display.optimize_watermarks =
7614                                 ilk_optimize_watermarks;
7615                 } else {
7616                         DRM_DEBUG_KMS("Failed to read display plane latency. "
7617                                       "Disable CxSR\n");
7618                 }
7619         } else if (IS_CHERRYVIEW(dev)) {
7620                 vlv_setup_wm_latency(dev);
7621                 dev_priv->display.update_wm = vlv_update_wm;
7622         } else if (IS_VALLEYVIEW(dev)) {
7623                 vlv_setup_wm_latency(dev);
7624                 dev_priv->display.update_wm = vlv_update_wm;
7625         } else if (IS_PINEVIEW(dev)) {
7626                 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
7627                                             dev_priv->is_ddr3,
7628                                             dev_priv->fsb_freq,
7629                                             dev_priv->mem_freq)) {
7630                         DRM_INFO("failed to find known CxSR latency "
7631                                  "(found ddr%s fsb freq %d, mem freq %d), "
7632                                  "disabling CxSR\n",
7633                                  (dev_priv->is_ddr3 == 1) ? "3" : "2",
7634                                  dev_priv->fsb_freq, dev_priv->mem_freq);
7635                         /* Disable CxSR and never update its watermark again */
7636                         intel_set_memory_cxsr(dev_priv, false);
7637                         dev_priv->display.update_wm = NULL;
7638                 } else
7639                         dev_priv->display.update_wm = pineview_update_wm;
7640         } else if (IS_G4X(dev)) {
7641                 dev_priv->display.update_wm = g4x_update_wm;
7642         } else if (IS_GEN4(dev)) {
7643                 dev_priv->display.update_wm = i965_update_wm;
7644         } else if (IS_GEN3(dev)) {
7645                 dev_priv->display.update_wm = i9xx_update_wm;
7646                 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
7647         } else if (IS_GEN2(dev)) {
7648                 if (INTEL_INFO(dev)->num_pipes == 1) {
7649                         dev_priv->display.update_wm = i845_update_wm;
7650                         dev_priv->display.get_fifo_size = i845_get_fifo_size;
7651                 } else {
7652                         dev_priv->display.update_wm = i9xx_update_wm;
7653                         dev_priv->display.get_fifo_size = i830_get_fifo_size;
7654                 }
7655         } else {
7656                 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
7657         }
7658 }
7659
7660 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u32 mbox, u32 *val)
7661 {
7662         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7663
7664         /* GEN6_PCODE_* are outside of the forcewake domain, we can
7665          * use te fw I915_READ variants to reduce the amount of work
7666          * required when reading/writing.
7667          */
7668
7669         if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
7670                 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
7671                 return -EAGAIN;
7672         }
7673
7674         I915_WRITE_FW(GEN6_PCODE_DATA, *val);
7675         I915_WRITE_FW(GEN6_PCODE_DATA1, 0);
7676         I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
7677
7678         if (intel_wait_for_register_fw(dev_priv,
7679                                        GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
7680                                        500)) {
7681                 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
7682                 return -ETIMEDOUT;
7683         }
7684
7685         *val = I915_READ_FW(GEN6_PCODE_DATA);
7686         I915_WRITE_FW(GEN6_PCODE_DATA, 0);
7687
7688         return 0;
7689 }
7690
7691 int sandybridge_pcode_write(struct drm_i915_private *dev_priv,
7692                                u32 mbox, u32 val)
7693 {
7694         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7695
7696         /* GEN6_PCODE_* are outside of the forcewake domain, we can
7697          * use te fw I915_READ variants to reduce the amount of work
7698          * required when reading/writing.
7699          */
7700
7701         if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
7702                 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
7703                 return -EAGAIN;
7704         }
7705
7706         I915_WRITE_FW(GEN6_PCODE_DATA, val);
7707         I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
7708
7709         if (intel_wait_for_register_fw(dev_priv,
7710                                        GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
7711                                        500)) {
7712                 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
7713                 return -ETIMEDOUT;
7714         }
7715
7716         I915_WRITE_FW(GEN6_PCODE_DATA, 0);
7717
7718         return 0;
7719 }
7720
7721 static int byt_gpu_freq(struct drm_i915_private *dev_priv, int val)
7722 {
7723         /*
7724          * N = val - 0xb7
7725          * Slow = Fast = GPLL ref * N
7726          */
7727         return DIV_ROUND_CLOSEST(dev_priv->rps.gpll_ref_freq * (val - 0xb7), 1000);
7728 }
7729
7730 static int byt_freq_opcode(struct drm_i915_private *dev_priv, int val)
7731 {
7732         return DIV_ROUND_CLOSEST(1000 * val, dev_priv->rps.gpll_ref_freq) + 0xb7;
7733 }
7734
7735 static int chv_gpu_freq(struct drm_i915_private *dev_priv, int val)
7736 {
7737         /*
7738          * N = val / 2
7739          * CU (slow) = CU2x (fast) / 2 = GPLL ref * N / 2
7740          */
7741         return DIV_ROUND_CLOSEST(dev_priv->rps.gpll_ref_freq * val, 2 * 2 * 1000);
7742 }
7743
7744 static int chv_freq_opcode(struct drm_i915_private *dev_priv, int val)
7745 {
7746         /* CHV needs even values */
7747         return DIV_ROUND_CLOSEST(2 * 1000 * val, dev_priv->rps.gpll_ref_freq) * 2;
7748 }
7749
7750 int intel_gpu_freq(struct drm_i915_private *dev_priv, int val)
7751 {
7752         if (IS_GEN9(dev_priv))
7753                 return DIV_ROUND_CLOSEST(val * GT_FREQUENCY_MULTIPLIER,
7754                                          GEN9_FREQ_SCALER);
7755         else if (IS_CHERRYVIEW(dev_priv))
7756                 return chv_gpu_freq(dev_priv, val);
7757         else if (IS_VALLEYVIEW(dev_priv))
7758                 return byt_gpu_freq(dev_priv, val);
7759         else
7760                 return val * GT_FREQUENCY_MULTIPLIER;
7761 }
7762
7763 int intel_freq_opcode(struct drm_i915_private *dev_priv, int val)
7764 {
7765         if (IS_GEN9(dev_priv))
7766                 return DIV_ROUND_CLOSEST(val * GEN9_FREQ_SCALER,
7767                                          GT_FREQUENCY_MULTIPLIER);
7768         else if (IS_CHERRYVIEW(dev_priv))
7769                 return chv_freq_opcode(dev_priv, val);
7770         else if (IS_VALLEYVIEW(dev_priv))
7771                 return byt_freq_opcode(dev_priv, val);
7772         else
7773                 return DIV_ROUND_CLOSEST(val, GT_FREQUENCY_MULTIPLIER);
7774 }
7775
7776 struct request_boost {
7777         struct work_struct work;
7778         struct drm_i915_gem_request *req;
7779 };
7780
7781 static void __intel_rps_boost_work(struct work_struct *work)
7782 {
7783         struct request_boost *boost = container_of(work, struct request_boost, work);
7784         struct drm_i915_gem_request *req = boost->req;
7785
7786         if (!i915_gem_request_completed(req))
7787                 gen6_rps_boost(req->i915, NULL, req->emitted_jiffies);
7788
7789         i915_gem_request_unreference(req);
7790         kfree(boost);
7791 }
7792
7793 void intel_queue_rps_boost_for_request(struct drm_i915_gem_request *req)
7794 {
7795         struct request_boost *boost;
7796
7797         if (req == NULL || INTEL_GEN(req->i915) < 6)
7798                 return;
7799
7800         if (i915_gem_request_completed(req))
7801                 return;
7802
7803         boost = kmalloc(sizeof(*boost), GFP_ATOMIC);
7804         if (boost == NULL)
7805                 return;
7806
7807         i915_gem_request_reference(req);
7808         boost->req = req;
7809
7810         INIT_WORK(&boost->work, __intel_rps_boost_work);
7811         queue_work(req->i915->wq, &boost->work);
7812 }
7813
7814 void intel_pm_setup(struct drm_device *dev)
7815 {
7816         struct drm_i915_private *dev_priv = to_i915(dev);
7817
7818         mutex_init(&dev_priv->rps.hw_lock);
7819         spin_lock_init(&dev_priv->rps.client_lock);
7820
7821         INIT_DELAYED_WORK(&dev_priv->rps.delayed_resume_work,
7822                           intel_gen6_powersave_work);
7823         INIT_LIST_HEAD(&dev_priv->rps.clients);
7824         INIT_LIST_HEAD(&dev_priv->rps.semaphores.link);
7825         INIT_LIST_HEAD(&dev_priv->rps.mmioflips.link);
7826
7827         dev_priv->pm.suspended = false;
7828         atomic_set(&dev_priv->pm.wakeref_count, 0);
7829         atomic_set(&dev_priv->pm.atomic_seq, 0);
7830 }