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 #include <drm/drm_atomic_helper.h>
  35
  36 /**
  37  * DOC: RC6
  38  *
  39  * RC6 is a special power stage which allows the GPU to enter an very
  40  * low-voltage mode when idle, using down to 0V while at this stage.  This
  41  * stage is entered automatically when the GPU is idle when RC6 support is
  42  * enabled, and as soon as new workload arises GPU wakes up automatically as well.
  43  *
  44  * There are different RC6 modes available in Intel GPU, which differentiate
  45  * among each other with the latency required to enter and leave RC6 and
  46  * voltage consumed by the GPU in different states.
  47  *
  48  * The combination of the following flags define which states GPU is allowed
  49  * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
  50  * RC6pp is deepest RC6. Their support by hardware varies according to the
  51  * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
  52  * which brings the most power savings; deeper states save more power, but
  53  * require higher latency to switch to and wake up.
  54  */
  55 #define INTEL_RC6_ENABLE                        (1<<0)
  56 #define INTEL_RC6p_ENABLE                       (1<<1)
  57 #define INTEL_RC6pp_ENABLE                      (1<<2)
  58
  59 static void gen9_init_clock_gating(struct drm_i915_private *dev_priv)
  60 {
  61         /* See Bspec note for PSR2_CTL bit 31, Wa#828:skl,bxt,kbl */
  62         I915_WRITE(CHICKEN_PAR1_1,
  63                    I915_READ(CHICKEN_PAR1_1) | SKL_EDP_PSR_FIX_RDWRAP);
  64
  65         I915_WRITE(GEN8_CONFIG0,
  66                    I915_READ(GEN8_CONFIG0) | GEN9_DEFAULT_FIXES);
  67
  68         /* WaEnableChickenDCPR:skl,bxt,kbl */
  69         I915_WRITE(GEN8_CHICKEN_DCPR_1,
  70                    I915_READ(GEN8_CHICKEN_DCPR_1) | MASK_WAKEMEM);
  71
  72         /* WaFbcTurnOffFbcWatermark:skl,bxt,kbl */
  73         /* WaFbcWakeMemOn:skl,bxt,kbl */
  74         I915_WRITE(DISP_ARB_CTL, I915_READ(DISP_ARB_CTL) |
  75                    DISP_FBC_WM_DIS |
  76                    DISP_FBC_MEMORY_WAKE);
  77
  78         /* WaFbcHighMemBwCorruptionAvoidance:skl,bxt,kbl */
  79         I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
  80                    ILK_DPFC_DISABLE_DUMMY0);
  81 }
  82
  83 static void bxt_init_clock_gating(struct drm_i915_private *dev_priv)
  84 {
  85         gen9_init_clock_gating(dev_priv);
  86
  87         /* WaDisableSDEUnitClockGating:bxt */
  88         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
  89                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
  90
  91         /*
  92          * FIXME:
  93          * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
  94          */
  95         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
  96                    GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ);
  97
  98         /*
  99          * Wa: Backlight PWM may stop in the asserted state, causing backlight
 100          * to stay fully on.
 101          */
 102         if (IS_BXT_REVID(dev_priv, BXT_REVID_B0, REVID_FOREVER))
 103                 I915_WRITE(GEN9_CLKGATE_DIS_0, I915_READ(GEN9_CLKGATE_DIS_0) |
 104                            PWM1_GATING_DIS | PWM2_GATING_DIS);
 105 }
 106
 107 static void i915_pineview_get_mem_freq(struct drm_i915_private *dev_priv)
 108 {
 109         u32 tmp;
 110
 111         tmp = I915_READ(CLKCFG);
 112
 113         switch (tmp & CLKCFG_FSB_MASK) {
 114         case CLKCFG_FSB_533:
 115                 dev_priv->fsb_freq = 533; /* 133*4 */
 116                 break;
 117         case CLKCFG_FSB_800:
 118                 dev_priv->fsb_freq = 800; /* 200*4 */
 119                 break;
 120         case CLKCFG_FSB_667:
 121                 dev_priv->fsb_freq =  667; /* 167*4 */
 122                 break;
 123         case CLKCFG_FSB_400:
 124                 dev_priv->fsb_freq = 400; /* 100*4 */
 125                 break;
 126         }
 127
 128         switch (tmp & CLKCFG_MEM_MASK) {
 129         case CLKCFG_MEM_533:
 130                 dev_priv->mem_freq = 533;
 131                 break;
 132         case CLKCFG_MEM_667:
 133                 dev_priv->mem_freq = 667;
 134                 break;
 135         case CLKCFG_MEM_800:
 136                 dev_priv->mem_freq = 800;
 137                 break;
 138         }
 139
 140         /* detect pineview DDR3 setting */
 141         tmp = I915_READ(CSHRDDR3CTL);
 142         dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
 143 }
 144
 145 static void i915_ironlake_get_mem_freq(struct drm_i915_private *dev_priv)
 146 {
 147         u16 ddrpll, csipll;
 148
 149         ddrpll = I915_READ16(DDRMPLL1);
 150         csipll = I915_READ16(CSIPLL0);
 151
 152         switch (ddrpll & 0xff) {
 153         case 0xc:
 154                 dev_priv->mem_freq = 800;
 155                 break;
 156         case 0x10:
 157                 dev_priv->mem_freq = 1066;
 158                 break;
 159         case 0x14:
 160                 dev_priv->mem_freq = 1333;
 161                 break;
 162         case 0x18:
 163                 dev_priv->mem_freq = 1600;
 164                 break;
 165         default:
 166                 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
 167                                  ddrpll & 0xff);
 168                 dev_priv->mem_freq = 0;
 169                 break;
 170         }
 171
 172         dev_priv->ips.r_t = dev_priv->mem_freq;
 173
 174         switch (csipll & 0x3ff) {
 175         case 0x00c:
 176                 dev_priv->fsb_freq = 3200;
 177                 break;
 178         case 0x00e:
 179                 dev_priv->fsb_freq = 3733;
 180                 break;
 181         case 0x010:
 182                 dev_priv->fsb_freq = 4266;
 183                 break;
 184         case 0x012:
 185                 dev_priv->fsb_freq = 4800;
 186                 break;
 187         case 0x014:
 188                 dev_priv->fsb_freq = 5333;
 189                 break;
 190         case 0x016:
 191                 dev_priv->fsb_freq = 5866;
 192                 break;
 193         case 0x018:
 194                 dev_priv->fsb_freq = 6400;
 195                 break;
 196         default:
 197                 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
 198                                  csipll & 0x3ff);
 199                 dev_priv->fsb_freq = 0;
 200                 break;
 201         }
 202
 203         if (dev_priv->fsb_freq == 3200) {
 204                 dev_priv->ips.c_m = 0;
 205         } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
 206                 dev_priv->ips.c_m = 1;
 207         } else {
 208                 dev_priv->ips.c_m = 2;
 209         }
 210 }
 211
 212 static const struct cxsr_latency cxsr_latency_table[] = {
 213         {1, 0, 800, 400, 3382, 33382, 3983, 33983},    /* DDR2-400 SC */
 214         {1, 0, 800, 667, 3354, 33354, 3807, 33807},    /* DDR2-667 SC */
 215         {1, 0, 800, 800, 3347, 33347, 3763, 33763},    /* DDR2-800 SC */
 216         {1, 1, 800, 667, 6420, 36420, 6873, 36873},    /* DDR3-667 SC */
 217         {1, 1, 800, 800, 5902, 35902, 6318, 36318},    /* DDR3-800 SC */
 218
 219         {1, 0, 667, 400, 3400, 33400, 4021, 34021},    /* DDR2-400 SC */
 220         {1, 0, 667, 667, 3372, 33372, 3845, 33845},    /* DDR2-667 SC */
 221         {1, 0, 667, 800, 3386, 33386, 3822, 33822},    /* DDR2-800 SC */
 222         {1, 1, 667, 667, 6438, 36438, 6911, 36911},    /* DDR3-667 SC */
 223         {1, 1, 667, 800, 5941, 35941, 6377, 36377},    /* DDR3-800 SC */
 224
 225         {1, 0, 400, 400, 3472, 33472, 4173, 34173},    /* DDR2-400 SC */
 226         {1, 0, 400, 667, 3443, 33443, 3996, 33996},    /* DDR2-667 SC */
 227         {1, 0, 400, 800, 3430, 33430, 3946, 33946},    /* DDR2-800 SC */
 228         {1, 1, 400, 667, 6509, 36509, 7062, 37062},    /* DDR3-667 SC */
 229         {1, 1, 400, 800, 5985, 35985, 6501, 36501},    /* DDR3-800 SC */
 230
 231         {0, 0, 800, 400, 3438, 33438, 4065, 34065},    /* DDR2-400 SC */
 232         {0, 0, 800, 667, 3410, 33410, 3889, 33889},    /* DDR2-667 SC */
 233         {0, 0, 800, 800, 3403, 33403, 3845, 33845},    /* DDR2-800 SC */
 234         {0, 1, 800, 667, 6476, 36476, 6955, 36955},    /* DDR3-667 SC */
 235         {0, 1, 800, 800, 5958, 35958, 6400, 36400},    /* DDR3-800 SC */
 236
 237         {0, 0, 667, 400, 3456, 33456, 4103, 34106},    /* DDR2-400 SC */
 238         {0, 0, 667, 667, 3428, 33428, 3927, 33927},    /* DDR2-667 SC */
 239         {0, 0, 667, 800, 3443, 33443, 3905, 33905},    /* DDR2-800 SC */
 240         {0, 1, 667, 667, 6494, 36494, 6993, 36993},    /* DDR3-667 SC */
 241         {0, 1, 667, 800, 5998, 35998, 6460, 36460},    /* DDR3-800 SC */
 242
 243         {0, 0, 400, 400, 3528, 33528, 4255, 34255},    /* DDR2-400 SC */
 244         {0, 0, 400, 667, 3500, 33500, 4079, 34079},    /* DDR2-667 SC */
 245         {0, 0, 400, 800, 3487, 33487, 4029, 34029},    /* DDR2-800 SC */
 246         {0, 1, 400, 667, 6566, 36566, 7145, 37145},    /* DDR3-667 SC */
 247         {0, 1, 400, 800, 6042, 36042, 6584, 36584},    /* DDR3-800 SC */
 248 };
 249
 250 static const struct cxsr_latency *intel_get_cxsr_latency(bool is_desktop,
 251                                                          bool is_ddr3,
 252                                                          int fsb,
 253                                                          int mem)
 254 {
 255         const struct cxsr_latency *latency;
 256         int i;
 257
 258         if (fsb == 0 || mem == 0)
 259                 return NULL;
 260
 261         for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
 262                 latency = &cxsr_latency_table[i];
 263                 if (is_desktop == latency->is_desktop &&
 264                     is_ddr3 == latency->is_ddr3 &&
 265                     fsb == latency->fsb_freq && mem == latency->mem_freq)
 266                         return latency;
 267         }
 268
 269         DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 270
 271         return NULL;
 272 }
 273
 274 static void chv_set_memory_dvfs(struct drm_i915_private *dev_priv, bool enable)
 275 {
 276         u32 val;
 277
 278         mutex_lock(&dev_priv->rps.hw_lock);
 279
 280         val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
 281         if (enable)
 282                 val &= ~FORCE_DDR_HIGH_FREQ;
 283         else
 284                 val |= FORCE_DDR_HIGH_FREQ;
 285         val &= ~FORCE_DDR_LOW_FREQ;
 286         val |= FORCE_DDR_FREQ_REQ_ACK;
 287         vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
 288
 289         if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
 290                       FORCE_DDR_FREQ_REQ_ACK) == 0, 3))
 291                 DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
 292
 293         mutex_unlock(&dev_priv->rps.hw_lock);
 294 }
 295
 296 static void chv_set_memory_pm5(struct drm_i915_private *dev_priv, bool enable)
 297 {
 298         u32 val;
 299
 300         mutex_lock(&dev_priv->rps.hw_lock);
 301
 302         val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
 303         if (enable)
 304                 val |= DSP_MAXFIFO_PM5_ENABLE;
 305         else
 306                 val &= ~DSP_MAXFIFO_PM5_ENABLE;
 307         vlv_punit_write(dev_priv, PUNIT_REG_DSPFREQ, val);
 308
 309         mutex_unlock(&dev_priv->rps.hw_lock);
 310 }
 311
 312 #define FW_WM(value, plane) \
 313         (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
 314
 315 void intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
 316 {
 317         u32 val;
 318
 319         if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
 320                 I915_WRITE(FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
 321                 POSTING_READ(FW_BLC_SELF_VLV);
 322                 dev_priv->wm.vlv.cxsr = enable;
 323         } else if (IS_G4X(dev_priv) || IS_CRESTLINE(dev_priv)) {
 324                 I915_WRITE(FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
 325                 POSTING_READ(FW_BLC_SELF);
 326         } else if (IS_PINEVIEW(dev_priv)) {
 327                 val = I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN;
 328                 val |= enable ? PINEVIEW_SELF_REFRESH_EN : 0;
 329                 I915_WRITE(DSPFW3, val);
 330                 POSTING_READ(DSPFW3);
 331         } else if (IS_I945G(dev_priv) || IS_I945GM(dev_priv)) {
 332                 val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
 333                                _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
 334                 I915_WRITE(FW_BLC_SELF, val);
 335                 POSTING_READ(FW_BLC_SELF);
 336         } else if (IS_I915GM(dev_priv)) {
 337                 /*
 338                  * FIXME can't find a bit like this for 915G, and
 339                  * and yet it does have the related watermark in
 340                  * FW_BLC_SELF. What's going on?
 341                  */
 342                 val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
 343                                _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
 344                 I915_WRITE(INSTPM, val);
 345                 POSTING_READ(INSTPM);
 346         } else {
 347                 return;
 348         }
 349
 350         DRM_DEBUG_KMS("memory self-refresh is %s\n", enableddisabled(enable));
 351 }
 352
 353
 354 /*
 355  * Latency for FIFO fetches is dependent on several factors:
 356  *   - memory configuration (speed, channels)
 357  *   - chipset
 358  *   - current MCH state
 359  * It can be fairly high in some situations, so here we assume a fairly
 360  * pessimal value.  It's a tradeoff between extra memory fetches (if we
 361  * set this value too high, the FIFO will fetch frequently to stay full)
 362  * and power consumption (set it too low to save power and we might see
 363  * FIFO underruns and display "flicker").
 364  *
 365  * A value of 5us seems to be a good balance; safe for very low end
 366  * platforms but not overly aggressive on lower latency configs.
 367  */
 368 static const int pessimal_latency_ns = 5000;
 369
 370 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
 371         ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
 372
 373 static int vlv_get_fifo_size(struct drm_i915_private *dev_priv,
 374                               enum pipe pipe, int plane)
 375 {
 376         int sprite0_start, sprite1_start, size;
 377
 378         switch (pipe) {
 379                 uint32_t dsparb, dsparb2, dsparb3;
 380         case PIPE_A:
 381                 dsparb = I915_READ(DSPARB);
 382                 dsparb2 = I915_READ(DSPARB2);
 383                 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 0, 0);
 384                 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 8, 4);
 385                 break;
 386         case PIPE_B:
 387                 dsparb = I915_READ(DSPARB);
 388                 dsparb2 = I915_READ(DSPARB2);
 389                 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 16, 8);
 390                 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 24, 12);
 391                 break;
 392         case PIPE_C:
 393                 dsparb2 = I915_READ(DSPARB2);
 394                 dsparb3 = I915_READ(DSPARB3);
 395                 sprite0_start = VLV_FIFO_START(dsparb3, dsparb2, 0, 16);
 396                 sprite1_start = VLV_FIFO_START(dsparb3, dsparb2, 8, 20);
 397                 break;
 398         default:
 399                 return 0;
 400         }
 401
 402         switch (plane) {
 403         case 0:
 404                 size = sprite0_start;
 405                 break;
 406         case 1:
 407                 size = sprite1_start - sprite0_start;
 408                 break;
 409         case 2:
 410                 size = 512 - 1 - sprite1_start;
 411                 break;
 412         default:
 413                 return 0;
 414         }
 415
 416         DRM_DEBUG_KMS("Pipe %c %s %c FIFO size: %d\n",
 417                       pipe_name(pipe), plane == 0 ? "primary" : "sprite",
 418                       plane == 0 ? plane_name(pipe) : sprite_name(pipe, plane - 1),
 419                       size);
 420
 421         return size;
 422 }
 423
 424 static int i9xx_get_fifo_size(struct drm_i915_private *dev_priv, int plane)
 425 {
 426         uint32_t dsparb = I915_READ(DSPARB);
 427         int size;
 428
 429         size = dsparb & 0x7f;
 430         if (plane)
 431                 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
 432
 433         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 434                       plane ? "B" : "A", size);
 435
 436         return size;
 437 }
 438
 439 static int i830_get_fifo_size(struct drm_i915_private *dev_priv, int plane)
 440 {
 441         uint32_t dsparb = I915_READ(DSPARB);
 442         int size;
 443
 444         size = dsparb & 0x1ff;
 445         if (plane)
 446                 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
 447         size >>= 1; /* Convert to cachelines */
 448
 449         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 450                       plane ? "B" : "A", size);
 451
 452         return size;
 453 }
 454
 455 static int i845_get_fifo_size(struct drm_i915_private *dev_priv, int plane)
 456 {
 457         uint32_t dsparb = I915_READ(DSPARB);
 458         int size;
 459
 460         size = dsparb & 0x7f;
 461         size >>= 2; /* Convert to cachelines */
 462
 463         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 464                       plane ? "B" : "A",
 465                       size);
 466
 467         return size;
 468 }
 469
 470 /* Pineview has different values for various configs */
 471 static const struct intel_watermark_params pineview_display_wm = {
 472         .fifo_size = PINEVIEW_DISPLAY_FIFO,
 473         .max_wm = PINEVIEW_MAX_WM,
 474         .default_wm = PINEVIEW_DFT_WM,
 475         .guard_size = PINEVIEW_GUARD_WM,
 476         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 477 };
 478 static const struct intel_watermark_params pineview_display_hplloff_wm = {
 479         .fifo_size = PINEVIEW_DISPLAY_FIFO,
 480         .max_wm = PINEVIEW_MAX_WM,
 481         .default_wm = PINEVIEW_DFT_HPLLOFF_WM,
 482         .guard_size = PINEVIEW_GUARD_WM,
 483         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 484 };
 485 static const struct intel_watermark_params pineview_cursor_wm = {
 486         .fifo_size = PINEVIEW_CURSOR_FIFO,
 487         .max_wm = PINEVIEW_CURSOR_MAX_WM,
 488         .default_wm = PINEVIEW_CURSOR_DFT_WM,
 489         .guard_size = PINEVIEW_CURSOR_GUARD_WM,
 490         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 491 };
 492 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
 493         .fifo_size = PINEVIEW_CURSOR_FIFO,
 494         .max_wm = PINEVIEW_CURSOR_MAX_WM,
 495         .default_wm = PINEVIEW_CURSOR_DFT_WM,
 496         .guard_size = PINEVIEW_CURSOR_GUARD_WM,
 497         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 498 };
 499 static const struct intel_watermark_params g4x_wm_info = {
 500         .fifo_size = G4X_FIFO_SIZE,
 501         .max_wm = G4X_MAX_WM,
 502         .default_wm = G4X_MAX_WM,
 503         .guard_size = 2,
 504         .cacheline_size = G4X_FIFO_LINE_SIZE,
 505 };
 506 static const struct intel_watermark_params g4x_cursor_wm_info = {
 507         .fifo_size = I965_CURSOR_FIFO,
 508         .max_wm = I965_CURSOR_MAX_WM,
 509         .default_wm = I965_CURSOR_DFT_WM,
 510         .guard_size = 2,
 511         .cacheline_size = G4X_FIFO_LINE_SIZE,
 512 };
 513 static const struct intel_watermark_params i965_cursor_wm_info = {
 514         .fifo_size = I965_CURSOR_FIFO,
 515         .max_wm = I965_CURSOR_MAX_WM,
 516         .default_wm = I965_CURSOR_DFT_WM,
 517         .guard_size = 2,
 518         .cacheline_size = I915_FIFO_LINE_SIZE,
 519 };
 520 static const struct intel_watermark_params i945_wm_info = {
 521         .fifo_size = I945_FIFO_SIZE,
 522         .max_wm = I915_MAX_WM,
 523         .default_wm = 1,
 524         .guard_size = 2,
 525         .cacheline_size = I915_FIFO_LINE_SIZE,
 526 };
 527 static const struct intel_watermark_params i915_wm_info = {
 528         .fifo_size = I915_FIFO_SIZE,
 529         .max_wm = I915_MAX_WM,
 530         .default_wm = 1,
 531         .guard_size = 2,
 532         .cacheline_size = I915_FIFO_LINE_SIZE,
 533 };
 534 static const struct intel_watermark_params i830_a_wm_info = {
 535         .fifo_size = I855GM_FIFO_SIZE,
 536         .max_wm = I915_MAX_WM,
 537         .default_wm = 1,
 538         .guard_size = 2,
 539         .cacheline_size = I830_FIFO_LINE_SIZE,
 540 };
 541 static const struct intel_watermark_params i830_bc_wm_info = {
 542         .fifo_size = I855GM_FIFO_SIZE,
 543         .max_wm = I915_MAX_WM/2,
 544         .default_wm = 1,
 545         .guard_size = 2,
 546         .cacheline_size = I830_FIFO_LINE_SIZE,
 547 };
 548 static const struct intel_watermark_params i845_wm_info = {
 549         .fifo_size = I830_FIFO_SIZE,
 550         .max_wm = I915_MAX_WM,
 551         .default_wm = 1,
 552         .guard_size = 2,
 553         .cacheline_size = I830_FIFO_LINE_SIZE,
 554 };
 555
 556 /**
 557  * intel_calculate_wm - calculate watermark level
 558  * @clock_in_khz: pixel clock
 559  * @wm: chip FIFO params
 560  * @cpp: bytes per pixel
 561  * @latency_ns: memory latency for the platform
 562  *
 563  * Calculate the watermark level (the level at which the display plane will
 564  * start fetching from memory again).  Each chip has a different display
 565  * FIFO size and allocation, so the caller needs to figure that out and pass
 566  * in the correct intel_watermark_params structure.
 567  *
 568  * As the pixel clock runs, the FIFO will be drained at a rate that depends
 569  * on the pixel size.  When it reaches the watermark level, it'll start
 570  * fetching FIFO line sized based chunks from memory until the FIFO fills
 571  * past the watermark point.  If the FIFO drains completely, a FIFO underrun
 572  * will occur, and a display engine hang could result.
 573  */
 574 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
 575                                         const struct intel_watermark_params *wm,
 576                                         int fifo_size, int cpp,
 577                                         unsigned long latency_ns)
 578 {
 579         long entries_required, wm_size;
 580
 581         /*
 582          * Note: we need to make sure we don't overflow for various clock &
 583          * latency values.
 584          * clocks go from a few thousand to several hundred thousand.
 585          * latency is usually a few thousand
 586          */
 587         entries_required = ((clock_in_khz / 1000) * cpp * latency_ns) /
 588                 1000;
 589         entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
 590
 591         DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
 592
 593         wm_size = fifo_size - (entries_required + wm->guard_size);
 594
 595         DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
 596
 597         /* Don't promote wm_size to unsigned... */
 598         if (wm_size > (long)wm->max_wm)
 599                 wm_size = wm->max_wm;
 600         if (wm_size <= 0)
 601                 wm_size = wm->default_wm;
 602
 603         /*
 604          * Bspec seems to indicate that the value shouldn't be lower than
 605          * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
 606          * Lets go for 8 which is the burst size since certain platforms
 607          * already use a hardcoded 8 (which is what the spec says should be
 608          * done).
 609          */
 610         if (wm_size <= 8)
 611                 wm_size = 8;
 612
 613         return wm_size;
 614 }
 615
 616 static struct intel_crtc *single_enabled_crtc(struct drm_i915_private *dev_priv)
 617 {
 618         struct intel_crtc *crtc, *enabled = NULL;
 619
 620         for_each_intel_crtc(&dev_priv->drm, crtc) {
 621                 if (intel_crtc_active(crtc)) {
 622                         if (enabled)
 623                                 return NULL;
 624                         enabled = crtc;
 625                 }
 626         }
 627
 628         return enabled;
 629 }
 630
 631 static void pineview_update_wm(struct intel_crtc *unused_crtc)
 632 {
 633         struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
 634         struct intel_crtc *crtc;
 635         const struct cxsr_latency *latency;
 636         u32 reg;
 637         unsigned long wm;
 638
 639         latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv),
 640                                          dev_priv->is_ddr3,
 641                                          dev_priv->fsb_freq,
 642                                          dev_priv->mem_freq);
 643         if (!latency) {
 644                 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 645                 intel_set_memory_cxsr(dev_priv, false);
 646                 return;
 647         }
 648
 649         crtc = single_enabled_crtc(dev_priv);
 650         if (crtc) {
 651                 const struct drm_display_mode *adjusted_mode =
 652                         &crtc->config->base.adjusted_mode;
 653                 const struct drm_framebuffer *fb =
 654                         crtc->base.primary->state->fb;
 655                 int cpp = drm_format_plane_cpp(fb->pixel_format, 0);
 656                 int clock = adjusted_mode->crtc_clock;
 657
 658                 /* Display SR */
 659                 wm = intel_calculate_wm(clock, &pineview_display_wm,
 660                                         pineview_display_wm.fifo_size,
 661                                         cpp, latency->display_sr);
 662                 reg = I915_READ(DSPFW1);
 663                 reg &= ~DSPFW_SR_MASK;
 664                 reg |= FW_WM(wm, SR);
 665                 I915_WRITE(DSPFW1, reg);
 666                 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
 667
 668                 /* cursor SR */
 669                 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
 670                                         pineview_display_wm.fifo_size,
 671                                         cpp, latency->cursor_sr);
 672                 reg = I915_READ(DSPFW3);
 673                 reg &= ~DSPFW_CURSOR_SR_MASK;
 674                 reg |= FW_WM(wm, CURSOR_SR);
 675                 I915_WRITE(DSPFW3, reg);
 676
 677                 /* Display HPLL off SR */
 678                 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
 679                                         pineview_display_hplloff_wm.fifo_size,
 680                                         cpp, latency->display_hpll_disable);
 681                 reg = I915_READ(DSPFW3);
 682                 reg &= ~DSPFW_HPLL_SR_MASK;
 683                 reg |= FW_WM(wm, HPLL_SR);
 684                 I915_WRITE(DSPFW3, reg);
 685
 686                 /* cursor HPLL off SR */
 687                 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
 688                                         pineview_display_hplloff_wm.fifo_size,
 689                                         cpp, latency->cursor_hpll_disable);
 690                 reg = I915_READ(DSPFW3);
 691                 reg &= ~DSPFW_HPLL_CURSOR_MASK;
 692                 reg |= FW_WM(wm, HPLL_CURSOR);
 693                 I915_WRITE(DSPFW3, reg);
 694                 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
 695
 696                 intel_set_memory_cxsr(dev_priv, true);
 697         } else {
 698                 intel_set_memory_cxsr(dev_priv, false);
 699         }
 700 }
 701
 702 static bool g4x_compute_wm0(struct drm_i915_private *dev_priv,
 703                             int plane,
 704                             const struct intel_watermark_params *display,
 705                             int display_latency_ns,
 706                             const struct intel_watermark_params *cursor,
 707                             int cursor_latency_ns,
 708                             int *plane_wm,
 709                             int *cursor_wm)
 710 {
 711         struct intel_crtc *crtc;
 712         const struct drm_display_mode *adjusted_mode;
 713         const struct drm_framebuffer *fb;
 714         int htotal, hdisplay, clock, cpp;
 715         int line_time_us, line_count;
 716         int entries, tlb_miss;
 717
 718         crtc = intel_get_crtc_for_plane(dev_priv, plane);
 719         if (!intel_crtc_active(crtc)) {
 720                 *cursor_wm = cursor->guard_size;
 721                 *plane_wm = display->guard_size;
 722                 return false;
 723         }
 724
 725         adjusted_mode = &crtc->config->base.adjusted_mode;
 726         fb = crtc->base.primary->state->fb;
 727         clock = adjusted_mode->crtc_clock;
 728         htotal = adjusted_mode->crtc_htotal;
 729         hdisplay = crtc->config->pipe_src_w;
 730         cpp = drm_format_plane_cpp(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->base.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_i915_private *dev_priv,
 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/%u), 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/%u), 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_i915_private *dev_priv,
 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 intel_crtc *crtc;
 800         const struct drm_display_mode *adjusted_mode;
 801         const struct drm_framebuffer *fb;
 802         int hdisplay, htotal, cpp, clock;
 803         unsigned long line_time_us;
 804         int line_count, line_size;
 805         int small, large;
 806         int entries;
 807
 808         if (!latency_ns) {
 809                 *display_wm = *cursor_wm = 0;
 810                 return false;
 811         }
 812
 813         crtc = intel_get_crtc_for_plane(dev_priv, plane);
 814         adjusted_mode = &crtc->config->base.adjusted_mode;
 815         fb = crtc->base.primary->state->fb;
 816         clock = adjusted_mode->crtc_clock;
 817         htotal = adjusted_mode->crtc_htotal;
 818         hdisplay = crtc->config->pipe_src_w;
 819         cpp = drm_format_plane_cpp(fb->pixel_format, 0);
 820
 821         line_time_us = max(htotal * 1000 / clock, 1);
 822         line_count = (latency_ns / line_time_us + 1000) / 1000;
 823         line_size = hdisplay * cpp;
 824
 825         /* Use the minimum of the small and large buffer method for primary */
 826         small = ((clock * cpp / 1000) * latency_ns) / 1000;
 827         large = line_count * line_size;
 828
 829         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
 830         *display_wm = entries + display->guard_size;
 831
 832         /* calculate the self-refresh watermark for display cursor */
 833         entries = line_count * cpp * crtc->base.cursor->state->crtc_w;
 834         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
 835         *cursor_wm = entries + cursor->guard_size;
 836
 837         return g4x_check_srwm(dev_priv,
 838                               *display_wm, *cursor_wm,
 839                               display, cursor);
 840 }
 841
 842 #define FW_WM_VLV(value, plane) \
 843         (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
 844
 845 static void vlv_write_wm_values(struct intel_crtc *crtc,
 846                                 const struct vlv_wm_values *wm)
 847 {
 848         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
 849         enum pipe pipe = crtc->pipe;
 850
 851         I915_WRITE(VLV_DDL(pipe),
 852                    (wm->ddl[pipe].cursor << DDL_CURSOR_SHIFT) |
 853                    (wm->ddl[pipe].sprite[1] << DDL_SPRITE_SHIFT(1)) |
 854                    (wm->ddl[pipe].sprite[0] << DDL_SPRITE_SHIFT(0)) |
 855                    (wm->ddl[pipe].primary << DDL_PLANE_SHIFT));
 856
 857         I915_WRITE(DSPFW1,
 858                    FW_WM(wm->sr.plane, SR) |
 859                    FW_WM(wm->pipe[PIPE_B].cursor, CURSORB) |
 860                    FW_WM_VLV(wm->pipe[PIPE_B].primary, PLANEB) |
 861                    FW_WM_VLV(wm->pipe[PIPE_A].primary, PLANEA));
 862         I915_WRITE(DSPFW2,
 863                    FW_WM_VLV(wm->pipe[PIPE_A].sprite[1], SPRITEB) |
 864                    FW_WM(wm->pipe[PIPE_A].cursor, CURSORA) |
 865                    FW_WM_VLV(wm->pipe[PIPE_A].sprite[0], SPRITEA));
 866         I915_WRITE(DSPFW3,
 867                    FW_WM(wm->sr.cursor, CURSOR_SR));
 868
 869         if (IS_CHERRYVIEW(dev_priv)) {
 870                 I915_WRITE(DSPFW7_CHV,
 871                            FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
 872                            FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
 873                 I915_WRITE(DSPFW8_CHV,
 874                            FW_WM_VLV(wm->pipe[PIPE_C].sprite[1], SPRITEF) |
 875                            FW_WM_VLV(wm->pipe[PIPE_C].sprite[0], SPRITEE));
 876                 I915_WRITE(DSPFW9_CHV,
 877                            FW_WM_VLV(wm->pipe[PIPE_C].primary, PLANEC) |
 878                            FW_WM(wm->pipe[PIPE_C].cursor, CURSORC));
 879                 I915_WRITE(DSPHOWM,
 880                            FW_WM(wm->sr.plane >> 9, SR_HI) |
 881                            FW_WM(wm->pipe[PIPE_C].sprite[1] >> 8, SPRITEF_HI) |
 882                            FW_WM(wm->pipe[PIPE_C].sprite[0] >> 8, SPRITEE_HI) |
 883                            FW_WM(wm->pipe[PIPE_C].primary >> 8, PLANEC_HI) |
 884                            FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
 885                            FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
 886                            FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
 887                            FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
 888                            FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
 889                            FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
 890         } else {
 891                 I915_WRITE(DSPFW7,
 892                            FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
 893                            FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
 894                 I915_WRITE(DSPHOWM,
 895                            FW_WM(wm->sr.plane >> 9, SR_HI) |
 896                            FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
 897                            FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
 898                            FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
 899                            FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
 900                            FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
 901                            FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
 902         }
 903
 904         /* zero (unused) WM1 watermarks */
 905         I915_WRITE(DSPFW4, 0);
 906         I915_WRITE(DSPFW5, 0);
 907         I915_WRITE(DSPFW6, 0);
 908         I915_WRITE(DSPHOWM1, 0);
 909
 910         POSTING_READ(DSPFW1);
 911 }
 912
 913 #undef FW_WM_VLV
 914
 915 enum vlv_wm_level {
 916         VLV_WM_LEVEL_PM2,
 917         VLV_WM_LEVEL_PM5,
 918         VLV_WM_LEVEL_DDR_DVFS,
 919 };
 920
 921 /* latency must be in 0.1us units. */
 922 static unsigned int vlv_wm_method2(unsigned int pixel_rate,
 923                                    unsigned int pipe_htotal,
 924                                    unsigned int horiz_pixels,
 925                                    unsigned int cpp,
 926                                    unsigned int latency)
 927 {
 928         unsigned int ret;
 929
 930         ret = (latency * pixel_rate) / (pipe_htotal * 10000);
 931         ret = (ret + 1) * horiz_pixels * cpp;
 932         ret = DIV_ROUND_UP(ret, 64);
 933
 934         return ret;
 935 }
 936
 937 static void vlv_setup_wm_latency(struct drm_i915_private *dev_priv)
 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->base.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->base.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->base.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 =
1064                         INTEL_INFO(to_i915(dev))->num_pipes * 512 - 1;
1065                 struct intel_plane *plane;
1066
1067                 wm_state->sr[level].plane = sr_fifo_size - wm_state->sr[level].plane;
1068                 wm_state->sr[level].cursor = 63 - wm_state->sr[level].cursor;
1069
1070                 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1071                         switch (plane->base.type) {
1072                                 int sprite;
1073                         case DRM_PLANE_TYPE_CURSOR:
1074                                 wm_state->wm[level].cursor = plane->wm.fifo_size -
1075                                         wm_state->wm[level].cursor;
1076                                 break;
1077                         case DRM_PLANE_TYPE_PRIMARY:
1078                                 wm_state->wm[level].primary = plane->wm.fifo_size -
1079                                         wm_state->wm[level].primary;
1080                                 break;
1081                         case DRM_PLANE_TYPE_OVERLAY:
1082                                 sprite = plane->plane;
1083                                 wm_state->wm[level].sprite[sprite] = plane->wm.fifo_size -
1084                                         wm_state->wm[level].sprite[sprite];
1085                                 break;
1086                         }
1087                 }
1088         }
1089 }
1090
1091 static void vlv_compute_wm(struct intel_crtc *crtc)
1092 {
1093         struct drm_device *dev = crtc->base.dev;
1094         struct drm_i915_private *dev_priv = to_i915(dev);
1095         struct vlv_wm_state *wm_state = &crtc->wm_state;
1096         struct intel_plane *plane;
1097         int sr_fifo_size = INTEL_INFO(dev_priv)->num_pipes * 512 - 1;
1098         int level;
1099
1100         memset(wm_state, 0, sizeof(*wm_state));
1101
1102         wm_state->cxsr = crtc->pipe != PIPE_C && crtc->wm.cxsr_allowed;
1103         wm_state->num_levels = dev_priv->wm.max_level + 1;
1104
1105         wm_state->num_active_planes = 0;
1106
1107         vlv_compute_fifo(crtc);
1108
1109         if (wm_state->num_active_planes != 1)
1110                 wm_state->cxsr = false;
1111
1112         if (wm_state->cxsr) {
1113                 for (level = 0; level < wm_state->num_levels; level++) {
1114                         wm_state->sr[level].plane = sr_fifo_size;
1115                         wm_state->sr[level].cursor = 63;
1116                 }
1117         }
1118
1119         for_each_intel_plane_on_crtc(dev, crtc, plane) {
1120                 struct intel_plane_state *state =
1121                         to_intel_plane_state(plane->base.state);
1122
1123                 if (!state->base.visible)
1124                         continue;
1125
1126                 /* normal watermarks */
1127                 for (level = 0; level < wm_state->num_levels; level++) {
1128                         int wm = vlv_compute_wm_level(plane, crtc, state, level);
1129                         int max_wm = plane->base.type == DRM_PLANE_TYPE_CURSOR ? 63 : 511;
1130
1131                         /* hack */
1132                         if (WARN_ON(level == 0 && wm > max_wm))
1133                                 wm = max_wm;
1134
1135                         if (wm > plane->wm.fifo_size)
1136                                 break;
1137
1138                         switch (plane->base.type) {
1139                                 int sprite;
1140                         case DRM_PLANE_TYPE_CURSOR:
1141                                 wm_state->wm[level].cursor = wm;
1142                                 break;
1143                         case DRM_PLANE_TYPE_PRIMARY:
1144                                 wm_state->wm[level].primary = wm;
1145                                 break;
1146                         case DRM_PLANE_TYPE_OVERLAY:
1147                                 sprite = plane->plane;
1148                                 wm_state->wm[level].sprite[sprite] = wm;
1149                                 break;
1150                         }
1151                 }
1152
1153                 wm_state->num_levels = level;
1154
1155                 if (!wm_state->cxsr)
1156                         continue;
1157
1158                 /* maxfifo watermarks */
1159                 switch (plane->base.type) {
1160                         int sprite, level;
1161                 case DRM_PLANE_TYPE_CURSOR:
1162                         for (level = 0; level < wm_state->num_levels; level++)
1163                                 wm_state->sr[level].cursor =
1164                                         wm_state->wm[level].cursor;
1165                         break;
1166                 case DRM_PLANE_TYPE_PRIMARY:
1167                         for (level = 0; level < wm_state->num_levels; level++)
1168                                 wm_state->sr[level].plane =
1169                                         min(wm_state->sr[level].plane,
1170                                             wm_state->wm[level].primary);
1171                         break;
1172                 case DRM_PLANE_TYPE_OVERLAY:
1173                         sprite = plane->plane;
1174                         for (level = 0; level < wm_state->num_levels; level++)
1175                                 wm_state->sr[level].plane =
1176                                         min(wm_state->sr[level].plane,
1177                                             wm_state->wm[level].sprite[sprite]);
1178                         break;
1179                 }
1180         }
1181
1182         /* clear any (partially) filled invalid levels */
1183         for (level = wm_state->num_levels; level < dev_priv->wm.max_level + 1; level++) {
1184                 memset(&wm_state->wm[level], 0, sizeof(wm_state->wm[level]));
1185                 memset(&wm_state->sr[level], 0, sizeof(wm_state->sr[level]));
1186         }
1187
1188         vlv_invert_wms(crtc);
1189 }
1190
1191 #define VLV_FIFO(plane, value) \
1192         (((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1193
1194 static void vlv_pipe_set_fifo_size(struct intel_crtc *crtc)
1195 {
1196         struct drm_device *dev = crtc->base.dev;
1197         struct drm_i915_private *dev_priv = to_i915(dev);
1198         struct intel_plane *plane;
1199         int sprite0_start = 0, sprite1_start = 0, fifo_size = 0;
1200
1201         for_each_intel_plane_on_crtc(dev, crtc, plane) {
1202                 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
1203                         WARN_ON(plane->wm.fifo_size != 63);
1204                         continue;
1205                 }
1206
1207                 if (plane->base.type == DRM_PLANE_TYPE_PRIMARY)
1208                         sprite0_start = plane->wm.fifo_size;
1209                 else if (plane->plane == 0)
1210                         sprite1_start = sprite0_start + plane->wm.fifo_size;
1211                 else
1212                         fifo_size = sprite1_start + plane->wm.fifo_size;
1213         }
1214
1215         WARN_ON(fifo_size != 512 - 1);
1216
1217         DRM_DEBUG_KMS("Pipe %c FIFO split %d / %d / %d\n",
1218                       pipe_name(crtc->pipe), sprite0_start,
1219                       sprite1_start, fifo_size);
1220
1221         switch (crtc->pipe) {
1222                 uint32_t dsparb, dsparb2, dsparb3;
1223         case PIPE_A:
1224                 dsparb = I915_READ(DSPARB);
1225                 dsparb2 = I915_READ(DSPARB2);
1226
1227                 dsparb &= ~(VLV_FIFO(SPRITEA, 0xff) |
1228                             VLV_FIFO(SPRITEB, 0xff));
1229                 dsparb |= (VLV_FIFO(SPRITEA, sprite0_start) |
1230                            VLV_FIFO(SPRITEB, sprite1_start));
1231
1232                 dsparb2 &= ~(VLV_FIFO(SPRITEA_HI, 0x1) |
1233                              VLV_FIFO(SPRITEB_HI, 0x1));
1234                 dsparb2 |= (VLV_FIFO(SPRITEA_HI, sprite0_start >> 8) |
1235                            VLV_FIFO(SPRITEB_HI, sprite1_start >> 8));
1236
1237                 I915_WRITE(DSPARB, dsparb);
1238                 I915_WRITE(DSPARB2, dsparb2);
1239                 break;
1240         case PIPE_B:
1241                 dsparb = I915_READ(DSPARB);
1242                 dsparb2 = I915_READ(DSPARB2);
1243
1244                 dsparb &= ~(VLV_FIFO(SPRITEC, 0xff) |
1245                             VLV_FIFO(SPRITED, 0xff));
1246                 dsparb |= (VLV_FIFO(SPRITEC, sprite0_start) |
1247                            VLV_FIFO(SPRITED, sprite1_start));
1248
1249                 dsparb2 &= ~(VLV_FIFO(SPRITEC_HI, 0xff) |
1250                              VLV_FIFO(SPRITED_HI, 0xff));
1251                 dsparb2 |= (VLV_FIFO(SPRITEC_HI, sprite0_start >> 8) |
1252                            VLV_FIFO(SPRITED_HI, sprite1_start >> 8));
1253
1254                 I915_WRITE(DSPARB, dsparb);
1255                 I915_WRITE(DSPARB2, dsparb2);
1256                 break;
1257         case PIPE_C:
1258                 dsparb3 = I915_READ(DSPARB3);
1259                 dsparb2 = I915_READ(DSPARB2);
1260
1261                 dsparb3 &= ~(VLV_FIFO(SPRITEE, 0xff) |
1262                              VLV_FIFO(SPRITEF, 0xff));
1263                 dsparb3 |= (VLV_FIFO(SPRITEE, sprite0_start) |
1264                             VLV_FIFO(SPRITEF, sprite1_start));
1265
1266                 dsparb2 &= ~(VLV_FIFO(SPRITEE_HI, 0xff) |
1267                              VLV_FIFO(SPRITEF_HI, 0xff));
1268                 dsparb2 |= (VLV_FIFO(SPRITEE_HI, sprite0_start >> 8) |
1269                            VLV_FIFO(SPRITEF_HI, sprite1_start >> 8));
1270
1271                 I915_WRITE(DSPARB3, dsparb3);
1272                 I915_WRITE(DSPARB2, dsparb2);
1273                 break;
1274         default:
1275                 break;
1276         }
1277 }
1278
1279 #undef VLV_FIFO
1280
1281 static void vlv_merge_wm(struct drm_device *dev,
1282                          struct vlv_wm_values *wm)
1283 {
1284         struct intel_crtc *crtc;
1285         int num_active_crtcs = 0;
1286
1287         wm->level = to_i915(dev)->wm.max_level;
1288         wm->cxsr = true;
1289
1290         for_each_intel_crtc(dev, crtc) {
1291                 const struct vlv_wm_state *wm_state = &crtc->wm_state;
1292
1293                 if (!crtc->active)
1294                         continue;
1295
1296                 if (!wm_state->cxsr)
1297                         wm->cxsr = false;
1298
1299                 num_active_crtcs++;
1300                 wm->level = min_t(int, wm->level, wm_state->num_levels - 1);
1301         }
1302
1303         if (num_active_crtcs != 1)
1304                 wm->cxsr = false;
1305
1306         if (num_active_crtcs > 1)
1307                 wm->level = VLV_WM_LEVEL_PM2;
1308
1309         for_each_intel_crtc(dev, crtc) {
1310                 struct vlv_wm_state *wm_state = &crtc->wm_state;
1311                 enum pipe pipe = crtc->pipe;
1312
1313                 if (!crtc->active)
1314                         continue;
1315
1316                 wm->pipe[pipe] = wm_state->wm[wm->level];
1317                 if (wm->cxsr)
1318                         wm->sr = wm_state->sr[wm->level];
1319
1320                 wm->ddl[pipe].primary = DDL_PRECISION_HIGH | 2;
1321                 wm->ddl[pipe].sprite[0] = DDL_PRECISION_HIGH | 2;
1322                 wm->ddl[pipe].sprite[1] = DDL_PRECISION_HIGH | 2;
1323                 wm->ddl[pipe].cursor = DDL_PRECISION_HIGH | 2;
1324         }
1325 }
1326
1327 static void vlv_update_wm(struct intel_crtc *crtc)
1328 {
1329         struct drm_device *dev = crtc->base.dev;
1330         struct drm_i915_private *dev_priv = to_i915(dev);
1331         enum pipe pipe = crtc->pipe;
1332         struct vlv_wm_values wm = {};
1333
1334         vlv_compute_wm(crtc);
1335         vlv_merge_wm(dev, &wm);
1336
1337         if (memcmp(&dev_priv->wm.vlv, &wm, sizeof(wm)) == 0) {
1338                 /* FIXME should be part of crtc atomic commit */
1339                 vlv_pipe_set_fifo_size(crtc);
1340                 return;
1341         }
1342
1343         if (wm.level < VLV_WM_LEVEL_DDR_DVFS &&
1344             dev_priv->wm.vlv.level >= VLV_WM_LEVEL_DDR_DVFS)
1345                 chv_set_memory_dvfs(dev_priv, false);
1346
1347         if (wm.level < VLV_WM_LEVEL_PM5 &&
1348             dev_priv->wm.vlv.level >= VLV_WM_LEVEL_PM5)
1349                 chv_set_memory_pm5(dev_priv, false);
1350
1351         if (!wm.cxsr && dev_priv->wm.vlv.cxsr)
1352                 intel_set_memory_cxsr(dev_priv, false);
1353
1354         /* FIXME should be part of crtc atomic commit */
1355         vlv_pipe_set_fifo_size(crtc);
1356
1357         vlv_write_wm_values(crtc, &wm);
1358
1359         DRM_DEBUG_KMS("Setting FIFO watermarks - %c: plane=%d, cursor=%d, "
1360                       "sprite0=%d, sprite1=%d, SR: plane=%d, cursor=%d level=%d cxsr=%d\n",
1361                       pipe_name(pipe), wm.pipe[pipe].primary, wm.pipe[pipe].cursor,
1362                       wm.pipe[pipe].sprite[0], wm.pipe[pipe].sprite[1],
1363                       wm.sr.plane, wm.sr.cursor, wm.level, wm.cxsr);
1364
1365         if (wm.cxsr && !dev_priv->wm.vlv.cxsr)
1366                 intel_set_memory_cxsr(dev_priv, true);
1367
1368         if (wm.level >= VLV_WM_LEVEL_PM5 &&
1369             dev_priv->wm.vlv.level < VLV_WM_LEVEL_PM5)
1370                 chv_set_memory_pm5(dev_priv, true);
1371
1372         if (wm.level >= VLV_WM_LEVEL_DDR_DVFS &&
1373             dev_priv->wm.vlv.level < VLV_WM_LEVEL_DDR_DVFS)
1374                 chv_set_memory_dvfs(dev_priv, true);
1375
1376         dev_priv->wm.vlv = wm;
1377 }
1378
1379 #define single_plane_enabled(mask) is_power_of_2(mask)
1380
1381 static void g4x_update_wm(struct intel_crtc *crtc)
1382 {
1383         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
1384         static const int sr_latency_ns = 12000;
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_priv, 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_priv, 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_priv, 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 intel_crtc *unused_crtc)
1439 {
1440         struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
1441         struct intel_crtc *crtc;
1442         int srwm = 1;
1443         int cursor_sr = 16;
1444         bool cxsr_enabled;
1445
1446         /* Calc sr entries for one plane configs */
1447         crtc = single_enabled_crtc(dev_priv);
1448         if (crtc) {
1449                 /* self-refresh has much higher latency */
1450                 static const int sr_latency_ns = 12000;
1451                 const struct drm_display_mode *adjusted_mode =
1452                         &crtc->config->base.adjusted_mode;
1453                 const struct drm_framebuffer *fb =
1454                         crtc->base.primary->state->fb;
1455                 int clock = adjusted_mode->crtc_clock;
1456                 int htotal = adjusted_mode->crtc_htotal;
1457                 int hdisplay = crtc->config->pipe_src_w;
1458                 int cpp = drm_format_plane_cpp(fb->pixel_format, 0);
1459                 unsigned long line_time_us;
1460                 int entries;
1461
1462                 line_time_us = max(htotal * 1000 / clock, 1);
1463
1464                 /* Use ns/us then divide to preserve precision */
1465                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1466                         cpp * hdisplay;
1467                 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1468                 srwm = I965_FIFO_SIZE - entries;
1469                 if (srwm < 0)
1470                         srwm = 1;
1471                 srwm &= 0x1ff;
1472                 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1473                               entries, srwm);
1474
1475                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1476                         cpp * crtc->base.cursor->state->crtc_w;
1477                 entries = DIV_ROUND_UP(entries,
1478                                           i965_cursor_wm_info.cacheline_size);
1479                 cursor_sr = i965_cursor_wm_info.fifo_size -
1480                         (entries + i965_cursor_wm_info.guard_size);
1481
1482                 if (cursor_sr > i965_cursor_wm_info.max_wm)
1483                         cursor_sr = i965_cursor_wm_info.max_wm;
1484
1485                 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1486                               "cursor %d\n", srwm, cursor_sr);
1487
1488                 cxsr_enabled = true;
1489         } else {
1490                 cxsr_enabled = false;
1491                 /* Turn off self refresh if both pipes are enabled */
1492                 intel_set_memory_cxsr(dev_priv, false);
1493         }
1494
1495         DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1496                       srwm);
1497
1498         /* 965 has limitations... */
1499         I915_WRITE(DSPFW1, FW_WM(srwm, SR) |
1500                    FW_WM(8, CURSORB) |
1501                    FW_WM(8, PLANEB) |
1502                    FW_WM(8, PLANEA));
1503         I915_WRITE(DSPFW2, FW_WM(8, CURSORA) |
1504                    FW_WM(8, PLANEC_OLD));
1505         /* update cursor SR watermark */
1506         I915_WRITE(DSPFW3, FW_WM(cursor_sr, CURSOR_SR));
1507
1508         if (cxsr_enabled)
1509                 intel_set_memory_cxsr(dev_priv, true);
1510 }
1511
1512 #undef FW_WM
1513
1514 static void i9xx_update_wm(struct intel_crtc *unused_crtc)
1515 {
1516         struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
1517         const struct intel_watermark_params *wm_info;
1518         uint32_t fwater_lo;
1519         uint32_t fwater_hi;
1520         int cwm, srwm = 1;
1521         int fifo_size;
1522         int planea_wm, planeb_wm;
1523         struct intel_crtc *crtc, *enabled = NULL;
1524
1525         if (IS_I945GM(dev_priv))
1526                 wm_info = &i945_wm_info;
1527         else if (!IS_GEN2(dev_priv))
1528                 wm_info = &i915_wm_info;
1529         else
1530                 wm_info = &i830_a_wm_info;
1531
1532         fifo_size = dev_priv->display.get_fifo_size(dev_priv, 0);
1533         crtc = intel_get_crtc_for_plane(dev_priv, 0);
1534         if (intel_crtc_active(crtc)) {
1535                 const struct drm_display_mode *adjusted_mode =
1536                         &crtc->config->base.adjusted_mode;
1537                 const struct drm_framebuffer *fb =
1538                         crtc->base.primary->state->fb;
1539                 int cpp;
1540
1541                 if (IS_GEN2(dev_priv))
1542                         cpp = 4;
1543                 else
1544                         cpp = drm_format_plane_cpp(fb->pixel_format, 0);
1545
1546                 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1547                                                wm_info, fifo_size, cpp,
1548                                                pessimal_latency_ns);
1549                 enabled = crtc;
1550         } else {
1551                 planea_wm = fifo_size - wm_info->guard_size;
1552                 if (planea_wm > (long)wm_info->max_wm)
1553                         planea_wm = wm_info->max_wm;
1554         }
1555
1556         if (IS_GEN2(dev_priv))
1557                 wm_info = &i830_bc_wm_info;
1558
1559         fifo_size = dev_priv->display.get_fifo_size(dev_priv, 1);
1560         crtc = intel_get_crtc_for_plane(dev_priv, 1);
1561         if (intel_crtc_active(crtc)) {
1562                 const struct drm_display_mode *adjusted_mode =
1563                         &crtc->config->base.adjusted_mode;
1564                 const struct drm_framebuffer *fb =
1565                         crtc->base.primary->state->fb;
1566                 int cpp;
1567
1568                 if (IS_GEN2(dev_priv))
1569                         cpp = 4;
1570                 else
1571                         cpp = drm_format_plane_cpp(fb->pixel_format, 0);
1572
1573                 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1574                                                wm_info, fifo_size, cpp,
1575                                                pessimal_latency_ns);
1576                 if (enabled == NULL)
1577                         enabled = crtc;
1578                 else
1579                         enabled = NULL;
1580         } else {
1581                 planeb_wm = fifo_size - wm_info->guard_size;
1582                 if (planeb_wm > (long)wm_info->max_wm)
1583                         planeb_wm = wm_info->max_wm;
1584         }
1585
1586         DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1587
1588         if (IS_I915GM(dev_priv) && enabled) {
1589                 struct drm_i915_gem_object *obj;
1590
1591                 obj = intel_fb_obj(enabled->base.primary->state->fb);
1592
1593                 /* self-refresh seems busted with untiled */
1594                 if (!i915_gem_object_is_tiled(obj))
1595                         enabled = NULL;
1596         }
1597
1598         /*
1599          * Overlay gets an aggressive default since video jitter is bad.
1600          */
1601         cwm = 2;
1602
1603         /* Play safe and disable self-refresh before adjusting watermarks. */
1604         intel_set_memory_cxsr(dev_priv, false);
1605
1606         /* Calc sr entries for one plane configs */
1607         if (HAS_FW_BLC(dev_priv) && enabled) {
1608                 /* self-refresh has much higher latency */
1609                 static const int sr_latency_ns = 6000;
1610                 const struct drm_display_mode *adjusted_mode =
1611                         &enabled->config->base.adjusted_mode;
1612                 const struct drm_framebuffer *fb =
1613                         enabled->base.primary->state->fb;
1614                 int clock = adjusted_mode->crtc_clock;
1615                 int htotal = adjusted_mode->crtc_htotal;
1616                 int hdisplay = enabled->config->pipe_src_w;
1617                 int cpp;
1618                 unsigned long line_time_us;
1619                 int entries;
1620
1621                 if (IS_I915GM(dev_priv) || IS_I945GM(dev_priv))
1622                         cpp = 4;
1623                 else
1624                         cpp = drm_format_plane_cpp(fb->pixel_format, 0);
1625
1626                 line_time_us = max(htotal * 1000 / clock, 1);
1627
1628                 /* Use ns/us then divide to preserve precision */
1629                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1630                         cpp * hdisplay;
1631                 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1632                 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1633                 srwm = wm_info->fifo_size - entries;
1634                 if (srwm < 0)
1635                         srwm = 1;
1636
1637                 if (IS_I945G(dev_priv) || IS_I945GM(dev_priv))
1638                         I915_WRITE(FW_BLC_SELF,
1639                                    FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1640                 else
1641                         I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1642         }
1643
1644         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1645                       planea_wm, planeb_wm, cwm, srwm);
1646
1647         fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1648         fwater_hi = (cwm & 0x1f);
1649
1650         /* Set request length to 8 cachelines per fetch */
1651         fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1652         fwater_hi = fwater_hi | (1 << 8);
1653
1654         I915_WRITE(FW_BLC, fwater_lo);
1655         I915_WRITE(FW_BLC2, fwater_hi);
1656
1657         if (enabled)
1658                 intel_set_memory_cxsr(dev_priv, true);
1659 }
1660
1661 static void i845_update_wm(struct intel_crtc *unused_crtc)
1662 {
1663         struct drm_i915_private *dev_priv = to_i915(unused_crtc->base.dev);
1664         struct intel_crtc *crtc;
1665         const struct drm_display_mode *adjusted_mode;
1666         uint32_t fwater_lo;
1667         int planea_wm;
1668
1669         crtc = single_enabled_crtc(dev_priv);
1670         if (crtc == NULL)
1671                 return;
1672
1673         adjusted_mode = &crtc->config->base.adjusted_mode;
1674         planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1675                                        &i845_wm_info,
1676                                        dev_priv->display.get_fifo_size(dev_priv, 0),
1677                                        4, pessimal_latency_ns);
1678         fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1679         fwater_lo |= (3<<8) | planea_wm;
1680
1681         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1682
1683         I915_WRITE(FW_BLC, fwater_lo);
1684 }
1685
1686 uint32_t ilk_pipe_pixel_rate(const struct intel_crtc_state *pipe_config)
1687 {
1688         uint32_t pixel_rate;
1689
1690         pixel_rate = pipe_config->base.adjusted_mode.crtc_clock;
1691
1692         /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
1693          * adjust the pixel_rate here. */
1694
1695         if (pipe_config->pch_pfit.enabled) {
1696                 uint64_t pipe_w, pipe_h, pfit_w, pfit_h;
1697                 uint32_t pfit_size = pipe_config->pch_pfit.size;
1698
1699                 pipe_w = pipe_config->pipe_src_w;
1700                 pipe_h = pipe_config->pipe_src_h;
1701
1702                 pfit_w = (pfit_size >> 16) & 0xFFFF;
1703                 pfit_h = pfit_size & 0xFFFF;
1704                 if (pipe_w < pfit_w)
1705                         pipe_w = pfit_w;
1706                 if (pipe_h < pfit_h)
1707                         pipe_h = pfit_h;
1708
1709                 if (WARN_ON(!pfit_w || !pfit_h))
1710                         return pixel_rate;
1711
1712                 pixel_rate = div_u64((uint64_t) pixel_rate * pipe_w * pipe_h,
1713                                      pfit_w * pfit_h);
1714         }
1715
1716         return pixel_rate;
1717 }
1718
1719 /* latency must be in 0.1us units. */
1720 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t cpp, uint32_t latency)
1721 {
1722         uint64_t ret;
1723
1724         if (WARN(latency == 0, "Latency value missing\n"))
1725                 return UINT_MAX;
1726
1727         ret = (uint64_t) pixel_rate * cpp * latency;
1728         ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
1729
1730         return ret;
1731 }
1732
1733 /* latency must be in 0.1us units. */
1734 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
1735                                uint32_t horiz_pixels, uint8_t cpp,
1736                                uint32_t latency)
1737 {
1738         uint32_t ret;
1739
1740         if (WARN(latency == 0, "Latency value missing\n"))
1741                 return UINT_MAX;
1742         if (WARN_ON(!pipe_htotal))
1743                 return UINT_MAX;
1744
1745         ret = (latency * pixel_rate) / (pipe_htotal * 10000);
1746         ret = (ret + 1) * horiz_pixels * cpp;
1747         ret = DIV_ROUND_UP(ret, 64) + 2;
1748         return ret;
1749 }
1750
1751 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
1752                            uint8_t cpp)
1753 {
1754         /*
1755          * Neither of these should be possible since this function shouldn't be
1756          * called if the CRTC is off or the plane is invisible.  But let's be
1757          * extra paranoid to avoid a potential divide-by-zero if we screw up
1758          * elsewhere in the driver.
1759          */
1760         if (WARN_ON(!cpp))
1761                 return 0;
1762         if (WARN_ON(!horiz_pixels))
1763                 return 0;
1764
1765         return DIV_ROUND_UP(pri_val * 64, horiz_pixels * cpp) + 2;
1766 }
1767
1768 struct ilk_wm_maximums {
1769         uint16_t pri;
1770         uint16_t spr;
1771         uint16_t cur;
1772         uint16_t fbc;
1773 };
1774
1775 /*
1776  * For both WM_PIPE and WM_LP.
1777  * mem_value must be in 0.1us units.
1778  */
1779 static uint32_t ilk_compute_pri_wm(const struct intel_crtc_state *cstate,
1780                                    const struct intel_plane_state *pstate,
1781                                    uint32_t mem_value,
1782                                    bool is_lp)
1783 {
1784         int cpp = pstate->base.fb ?
1785                 drm_format_plane_cpp(pstate->base.fb->pixel_format, 0) : 0;
1786         uint32_t method1, method2;
1787
1788         if (!cstate->base.active || !pstate->base.visible)
1789                 return 0;
1790
1791         method1 = ilk_wm_method1(ilk_pipe_pixel_rate(cstate), cpp, mem_value);
1792
1793         if (!is_lp)
1794                 return method1;
1795
1796         method2 = ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1797                                  cstate->base.adjusted_mode.crtc_htotal,
1798                                  drm_rect_width(&pstate->base.dst),
1799                                  cpp, mem_value);
1800
1801         return min(method1, method2);
1802 }
1803
1804 /*
1805  * For both WM_PIPE and WM_LP.
1806  * mem_value must be in 0.1us units.
1807  */
1808 static uint32_t ilk_compute_spr_wm(const struct intel_crtc_state *cstate,
1809                                    const struct intel_plane_state *pstate,
1810                                    uint32_t mem_value)
1811 {
1812         int cpp = pstate->base.fb ?
1813                 drm_format_plane_cpp(pstate->base.fb->pixel_format, 0) : 0;
1814         uint32_t method1, method2;
1815
1816         if (!cstate->base.active || !pstate->base.visible)
1817                 return 0;
1818
1819         method1 = ilk_wm_method1(ilk_pipe_pixel_rate(cstate), cpp, mem_value);
1820         method2 = ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1821                                  cstate->base.adjusted_mode.crtc_htotal,
1822                                  drm_rect_width(&pstate->base.dst),
1823                                  cpp, mem_value);
1824         return min(method1, method2);
1825 }
1826
1827 /*
1828  * For both WM_PIPE and WM_LP.
1829  * mem_value must be in 0.1us units.
1830  */
1831 static uint32_t ilk_compute_cur_wm(const struct intel_crtc_state *cstate,
1832                                    const struct intel_plane_state *pstate,
1833                                    uint32_t mem_value)
1834 {
1835         /*
1836          * We treat the cursor plane as always-on for the purposes of watermark
1837          * calculation.  Until we have two-stage watermark programming merged,
1838          * this is necessary to avoid flickering.
1839          */
1840         int cpp = 4;
1841         int width = pstate->base.visible ? pstate->base.crtc_w : 64;
1842
1843         if (!cstate->base.active)
1844                 return 0;
1845
1846         return ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1847                               cstate->base.adjusted_mode.crtc_htotal,
1848                               width, cpp, mem_value);
1849 }
1850
1851 /* Only for WM_LP. */
1852 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state *cstate,
1853                                    const struct intel_plane_state *pstate,
1854                                    uint32_t pri_val)
1855 {
1856         int cpp = pstate->base.fb ?
1857                 drm_format_plane_cpp(pstate->base.fb->pixel_format, 0) : 0;
1858
1859         if (!cstate->base.active || !pstate->base.visible)
1860                 return 0;
1861
1862         return ilk_wm_fbc(pri_val, drm_rect_width(&pstate->base.dst), cpp);
1863 }
1864
1865 static unsigned int
1866 ilk_display_fifo_size(const struct drm_i915_private *dev_priv)
1867 {
1868         if (INTEL_GEN(dev_priv) >= 8)
1869                 return 3072;
1870         else if (INTEL_GEN(dev_priv) >= 7)
1871                 return 768;
1872         else
1873                 return 512;
1874 }
1875
1876 static unsigned int
1877 ilk_plane_wm_reg_max(const struct drm_i915_private *dev_priv,
1878                      int level, bool is_sprite)
1879 {
1880         if (INTEL_GEN(dev_priv) >= 8)
1881                 /* BDW primary/sprite plane watermarks */
1882                 return level == 0 ? 255 : 2047;
1883         else if (INTEL_GEN(dev_priv) >= 7)
1884                 /* IVB/HSW primary/sprite plane watermarks */
1885                 return level == 0 ? 127 : 1023;
1886         else if (!is_sprite)
1887                 /* ILK/SNB primary plane watermarks */
1888                 return level == 0 ? 127 : 511;
1889         else
1890                 /* ILK/SNB sprite plane watermarks */
1891                 return level == 0 ? 63 : 255;
1892 }
1893
1894 static unsigned int
1895 ilk_cursor_wm_reg_max(const struct drm_i915_private *dev_priv, int level)
1896 {
1897         if (INTEL_GEN(dev_priv) >= 7)
1898                 return level == 0 ? 63 : 255;
1899         else
1900                 return level == 0 ? 31 : 63;
1901 }
1902
1903 static unsigned int ilk_fbc_wm_reg_max(const struct drm_i915_private *dev_priv)
1904 {
1905         if (INTEL_GEN(dev_priv) >= 8)
1906                 return 31;
1907         else
1908                 return 15;
1909 }
1910
1911 /* Calculate the maximum primary/sprite plane watermark */
1912 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
1913                                      int level,
1914                                      const struct intel_wm_config *config,
1915                                      enum intel_ddb_partitioning ddb_partitioning,
1916                                      bool is_sprite)
1917 {
1918         struct drm_i915_private *dev_priv = to_i915(dev);
1919         unsigned int fifo_size = ilk_display_fifo_size(dev_priv);
1920
1921         /* if sprites aren't enabled, sprites get nothing */
1922         if (is_sprite && !config->sprites_enabled)
1923                 return 0;
1924
1925         /* HSW allows LP1+ watermarks even with multiple pipes */
1926         if (level == 0 || config->num_pipes_active > 1) {
1927                 fifo_size /= INTEL_INFO(dev_priv)->num_pipes;
1928
1929                 /*
1930                  * For some reason the non self refresh
1931                  * FIFO size is only half of the self
1932                  * refresh FIFO size on ILK/SNB.
1933                  */
1934                 if (INTEL_GEN(dev_priv) <= 6)
1935                         fifo_size /= 2;
1936         }
1937
1938         if (config->sprites_enabled) {
1939                 /* level 0 is always calculated with 1:1 split */
1940                 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
1941                         if (is_sprite)
1942                                 fifo_size *= 5;
1943                         fifo_size /= 6;
1944                 } else {
1945                         fifo_size /= 2;
1946                 }
1947         }
1948
1949         /* clamp to max that the registers can hold */
1950         return min(fifo_size, ilk_plane_wm_reg_max(dev_priv, level, is_sprite));
1951 }
1952
1953 /* Calculate the maximum cursor plane watermark */
1954 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
1955                                       int level,
1956                                       const struct intel_wm_config *config)
1957 {
1958         /* HSW LP1+ watermarks w/ multiple pipes */
1959         if (level > 0 && config->num_pipes_active > 1)
1960                 return 64;
1961
1962         /* otherwise just report max that registers can hold */
1963         return ilk_cursor_wm_reg_max(to_i915(dev), level);
1964 }
1965
1966 static void ilk_compute_wm_maximums(const struct drm_device *dev,
1967                                     int level,
1968                                     const struct intel_wm_config *config,
1969                                     enum intel_ddb_partitioning ddb_partitioning,
1970                                     struct ilk_wm_maximums *max)
1971 {
1972         max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
1973         max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
1974         max->cur = ilk_cursor_wm_max(dev, level, config);
1975         max->fbc = ilk_fbc_wm_reg_max(to_i915(dev));
1976 }
1977
1978 static void ilk_compute_wm_reg_maximums(const struct drm_i915_private *dev_priv,
1979                                         int level,
1980                                         struct ilk_wm_maximums *max)
1981 {
1982         max->pri = ilk_plane_wm_reg_max(dev_priv, level, false);
1983         max->spr = ilk_plane_wm_reg_max(dev_priv, level, true);
1984         max->cur = ilk_cursor_wm_reg_max(dev_priv, level);
1985         max->fbc = ilk_fbc_wm_reg_max(dev_priv);
1986 }
1987
1988 static bool ilk_validate_wm_level(int level,
1989                                   const struct ilk_wm_maximums *max,
1990                                   struct intel_wm_level *result)
1991 {
1992         bool ret;
1993
1994         /* already determined to be invalid? */
1995         if (!result->enable)
1996                 return false;
1997
1998         result->enable = result->pri_val <= max->pri &&
1999                          result->spr_val <= max->spr &&
2000                          result->cur_val <= max->cur;
2001
2002         ret = result->enable;
2003
2004         /*
2005          * HACK until we can pre-compute everything,
2006          * and thus fail gracefully if LP0 watermarks
2007          * are exceeded...
2008          */
2009         if (level == 0 && !result->enable) {
2010                 if (result->pri_val > max->pri)
2011                         DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
2012                                       level, result->pri_val, max->pri);
2013                 if (result->spr_val > max->spr)
2014                         DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
2015                                       level, result->spr_val, max->spr);
2016                 if (result->cur_val > max->cur)
2017                         DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
2018                                       level, result->cur_val, max->cur);
2019
2020                 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
2021                 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
2022                 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
2023                 result->enable = true;
2024         }
2025
2026         return ret;
2027 }
2028
2029 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
2030                                  const struct intel_crtc *intel_crtc,
2031                                  int level,
2032                                  struct intel_crtc_state *cstate,
2033                                  struct intel_plane_state *pristate,
2034                                  struct intel_plane_state *sprstate,
2035                                  struct intel_plane_state *curstate,
2036                                  struct intel_wm_level *result)
2037 {
2038         uint16_t pri_latency = dev_priv->wm.pri_latency[level];
2039         uint16_t spr_latency = dev_priv->wm.spr_latency[level];
2040         uint16_t cur_latency = dev_priv->wm.cur_latency[level];
2041
2042         /* WM1+ latency values stored in 0.5us units */
2043         if (level > 0) {
2044                 pri_latency *= 5;
2045                 spr_latency *= 5;
2046                 cur_latency *= 5;
2047         }
2048
2049         if (pristate) {
2050                 result->pri_val = ilk_compute_pri_wm(cstate, pristate,
2051                                                      pri_latency, level);
2052                 result->fbc_val = ilk_compute_fbc_wm(cstate, pristate, result->pri_val);
2053         }
2054
2055         if (sprstate)
2056                 result->spr_val = ilk_compute_spr_wm(cstate, sprstate, spr_latency);
2057
2058         if (curstate)
2059                 result->cur_val = ilk_compute_cur_wm(cstate, curstate, cur_latency);
2060
2061         result->enable = true;
2062 }
2063
2064 static uint32_t
2065 hsw_compute_linetime_wm(const struct intel_crtc_state *cstate)
2066 {
2067         const struct intel_atomic_state *intel_state =
2068                 to_intel_atomic_state(cstate->base.state);
2069         const struct drm_display_mode *adjusted_mode =
2070                 &cstate->base.adjusted_mode;
2071         u32 linetime, ips_linetime;
2072
2073         if (!cstate->base.active)
2074                 return 0;
2075         if (WARN_ON(adjusted_mode->crtc_clock == 0))
2076                 return 0;
2077         if (WARN_ON(intel_state->cdclk == 0))
2078                 return 0;
2079
2080         /* The WM are computed with base on how long it takes to fill a single
2081          * row at the given clock rate, multiplied by 8.
2082          * */
2083         linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2084                                      adjusted_mode->crtc_clock);
2085         ips_linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2086                                          intel_state->cdclk);
2087
2088         return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2089                PIPE_WM_LINETIME_TIME(linetime);
2090 }
2091
2092 static void intel_read_wm_latency(struct drm_i915_private *dev_priv,
2093                                   uint16_t wm[8])
2094 {
2095         if (IS_GEN9(dev_priv)) {
2096                 uint32_t val;
2097                 int ret, i;
2098                 int level, max_level = ilk_wm_max_level(dev_priv);
2099
2100                 /* read the first set of memory latencies[0:3] */
2101                 val = 0; /* data0 to be programmed to 0 for first set */
2102                 mutex_lock(&dev_priv->rps.hw_lock);
2103                 ret = sandybridge_pcode_read(dev_priv,
2104                                              GEN9_PCODE_READ_MEM_LATENCY,
2105                                              &val);
2106                 mutex_unlock(&dev_priv->rps.hw_lock);
2107
2108                 if (ret) {
2109                         DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2110                         return;
2111                 }
2112
2113                 wm[0] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2114                 wm[1] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2115                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2116                 wm[2] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2117                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2118                 wm[3] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2119                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2120
2121                 /* read the second set of memory latencies[4:7] */
2122                 val = 1; /* data0 to be programmed to 1 for second set */
2123                 mutex_lock(&dev_priv->rps.hw_lock);
2124                 ret = sandybridge_pcode_read(dev_priv,
2125                                              GEN9_PCODE_READ_MEM_LATENCY,
2126                                              &val);
2127                 mutex_unlock(&dev_priv->rps.hw_lock);
2128                 if (ret) {
2129                         DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2130                         return;
2131                 }
2132
2133                 wm[4] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2134                 wm[5] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2135                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2136                 wm[6] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2137                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2138                 wm[7] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2139                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2140
2141                 /*
2142                  * If a level n (n > 1) has a 0us latency, all levels m (m >= n)
2143                  * need to be disabled. We make sure to sanitize the values out
2144                  * of the punit to satisfy this requirement.
2145                  */
2146                 for (level = 1; level <= max_level; level++) {
2147                         if (wm[level] == 0) {
2148                                 for (i = level + 1; i <= max_level; i++)
2149                                         wm[i] = 0;
2150                                 break;
2151                         }
2152                 }
2153
2154                 /*
2155                  * WaWmMemoryReadLatency:skl
2156                  *
2157                  * punit doesn't take into account the read latency so we need
2158                  * to add 2us to the various latency levels we retrieve from the
2159                  * punit when level 0 response data us 0us.
2160                  */
2161                 if (wm[0] == 0) {
2162                         wm[0] += 2;
2163                         for (level = 1; level <= max_level; level++) {
2164                                 if (wm[level] == 0)
2165                                         break;
2166                                 wm[level] += 2;
2167                         }
2168                 }
2169
2170         } else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
2171                 uint64_t sskpd = I915_READ64(MCH_SSKPD);
2172
2173                 wm[0] = (sskpd >> 56) & 0xFF;
2174                 if (wm[0] == 0)
2175                         wm[0] = sskpd & 0xF;
2176                 wm[1] = (sskpd >> 4) & 0xFF;
2177                 wm[2] = (sskpd >> 12) & 0xFF;
2178                 wm[3] = (sskpd >> 20) & 0x1FF;
2179                 wm[4] = (sskpd >> 32) & 0x1FF;
2180         } else if (INTEL_GEN(dev_priv) >= 6) {
2181                 uint32_t sskpd = I915_READ(MCH_SSKPD);
2182
2183                 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2184                 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2185                 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2186                 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2187         } else if (INTEL_GEN(dev_priv) >= 5) {
2188                 uint32_t mltr = I915_READ(MLTR_ILK);
2189
2190                 /* ILK primary LP0 latency is 700 ns */
2191                 wm[0] = 7;
2192                 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2193                 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2194         }
2195 }
2196
2197 static void intel_fixup_spr_wm_latency(struct drm_i915_private *dev_priv,
2198                                        uint16_t wm[5])
2199 {
2200         /* ILK sprite LP0 latency is 1300 ns */
2201         if (IS_GEN5(dev_priv))
2202                 wm[0] = 13;
2203 }
2204
2205 static void intel_fixup_cur_wm_latency(struct drm_i915_private *dev_priv,
2206                                        uint16_t wm[5])
2207 {
2208         /* ILK cursor LP0 latency is 1300 ns */
2209         if (IS_GEN5(dev_priv))
2210                 wm[0] = 13;
2211
2212         /* WaDoubleCursorLP3Latency:ivb */
2213         if (IS_IVYBRIDGE(dev_priv))
2214                 wm[3] *= 2;
2215 }
2216
2217 int ilk_wm_max_level(const struct drm_i915_private *dev_priv)
2218 {
2219         /* how many WM levels are we expecting */
2220         if (INTEL_GEN(dev_priv) >= 9)
2221                 return 7;
2222         else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
2223                 return 4;
2224         else if (INTEL_GEN(dev_priv) >= 6)
2225                 return 3;
2226         else
2227                 return 2;
2228 }
2229
2230 static void intel_print_wm_latency(struct drm_i915_private *dev_priv,
2231                                    const char *name,
2232                                    const uint16_t wm[8])
2233 {
2234         int level, max_level = ilk_wm_max_level(dev_priv);
2235
2236         for (level = 0; level <= max_level; level++) {
2237                 unsigned int latency = wm[level];
2238
2239                 if (latency == 0) {
2240                         DRM_ERROR("%s WM%d latency not provided\n",
2241                                   name, level);
2242                         continue;
2243                 }
2244
2245                 /*
2246                  * - latencies are in us on gen9.
2247                  * - before then, WM1+ latency values are in 0.5us units
2248                  */
2249                 if (IS_GEN9(dev_priv))
2250                         latency *= 10;
2251                 else if (level > 0)
2252                         latency *= 5;
2253
2254                 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2255                               name, level, wm[level],
2256                               latency / 10, latency % 10);
2257         }
2258 }
2259
2260 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
2261                                     uint16_t wm[5], uint16_t min)
2262 {
2263         int level, max_level = ilk_wm_max_level(dev_priv);
2264
2265         if (wm[0] >= min)
2266                 return false;
2267
2268         wm[0] = max(wm[0], min);
2269         for (level = 1; level <= max_level; level++)
2270                 wm[level] = max_t(uint16_t, wm[level], DIV_ROUND_UP(min, 5));
2271
2272         return true;
2273 }
2274
2275 static void snb_wm_latency_quirk(struct drm_i915_private *dev_priv)
2276 {
2277         bool changed;
2278
2279         /*
2280          * The BIOS provided WM memory latency values are often
2281          * inadequate for high resolution displays. Adjust them.
2282          */
2283         changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
2284                 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
2285                 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
2286
2287         if (!changed)
2288                 return;
2289
2290         DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2291         intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
2292         intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
2293         intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
2294 }
2295
2296 static void ilk_setup_wm_latency(struct drm_i915_private *dev_priv)
2297 {
2298         intel_read_wm_latency(dev_priv, dev_priv->wm.pri_latency);
2299
2300         memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2301                sizeof(dev_priv->wm.pri_latency));
2302         memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2303                sizeof(dev_priv->wm.pri_latency));
2304
2305         intel_fixup_spr_wm_latency(dev_priv, dev_priv->wm.spr_latency);
2306         intel_fixup_cur_wm_latency(dev_priv, dev_priv->wm.cur_latency);
2307
2308         intel_print_wm_latency(dev_priv, "Primary", dev_priv->wm.pri_latency);
2309         intel_print_wm_latency(dev_priv, "Sprite", dev_priv->wm.spr_latency);
2310         intel_print_wm_latency(dev_priv, "Cursor", dev_priv->wm.cur_latency);
2311
2312         if (IS_GEN6(dev_priv))
2313                 snb_wm_latency_quirk(dev_priv);
2314 }
2315
2316 static void skl_setup_wm_latency(struct drm_i915_private *dev_priv)
2317 {
2318         intel_read_wm_latency(dev_priv, dev_priv->wm.skl_latency);
2319         intel_print_wm_latency(dev_priv, "Gen9 Plane", dev_priv->wm.skl_latency);
2320 }
2321
2322 static bool ilk_validate_pipe_wm(struct drm_device *dev,
2323                                  struct intel_pipe_wm *pipe_wm)
2324 {
2325         /* LP0 watermark maximums depend on this pipe alone */
2326         const struct intel_wm_config config = {
2327                 .num_pipes_active = 1,
2328                 .sprites_enabled = pipe_wm->sprites_enabled,
2329                 .sprites_scaled = pipe_wm->sprites_scaled,
2330         };
2331         struct ilk_wm_maximums max;
2332
2333         /* LP0 watermarks always use 1/2 DDB partitioning */
2334         ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
2335
2336         /* At least LP0 must be valid */
2337         if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0])) {
2338                 DRM_DEBUG_KMS("LP0 watermark invalid\n");
2339                 return false;
2340         }
2341
2342         return true;
2343 }
2344
2345 /* Compute new watermarks for the pipe */
2346 static int ilk_compute_pipe_wm(struct intel_crtc_state *cstate)
2347 {
2348         struct drm_atomic_state *state = cstate->base.state;
2349         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
2350         struct intel_pipe_wm *pipe_wm;
2351         struct drm_device *dev = state->dev;
2352         const struct drm_i915_private *dev_priv = to_i915(dev);
2353         struct intel_plane *intel_plane;
2354         struct intel_plane_state *pristate = NULL;
2355         struct intel_plane_state *sprstate = NULL;
2356         struct intel_plane_state *curstate = NULL;
2357         int level, max_level = ilk_wm_max_level(dev_priv), usable_level;
2358         struct ilk_wm_maximums max;
2359
2360         pipe_wm = &cstate->wm.ilk.optimal;
2361
2362         for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
2363                 struct intel_plane_state *ps;
2364
2365                 ps = intel_atomic_get_existing_plane_state(state,
2366                                                            intel_plane);
2367                 if (!ps)
2368                         continue;
2369
2370                 if (intel_plane->base.type == DRM_PLANE_TYPE_PRIMARY)
2371                         pristate = ps;
2372                 else if (intel_plane->base.type == DRM_PLANE_TYPE_OVERLAY)
2373                         sprstate = ps;
2374                 else if (intel_plane->base.type == DRM_PLANE_TYPE_CURSOR)
2375                         curstate = ps;
2376         }
2377
2378         pipe_wm->pipe_enabled = cstate->base.active;
2379         if (sprstate) {
2380                 pipe_wm->sprites_enabled = sprstate->base.visible;
2381                 pipe_wm->sprites_scaled = sprstate->base.visible &&
2382                         (drm_rect_width(&sprstate->base.dst) != drm_rect_width(&sprstate->base.src) >> 16 ||
2383                          drm_rect_height(&sprstate->base.dst) != drm_rect_height(&sprstate->base.src) >> 16);
2384         }
2385
2386         usable_level = max_level;
2387
2388         /* ILK/SNB: LP2+ watermarks only w/o sprites */
2389         if (INTEL_GEN(dev_priv) <= 6 && pipe_wm->sprites_enabled)
2390                 usable_level = 1;
2391
2392         /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2393         if (pipe_wm->sprites_scaled)
2394                 usable_level = 0;
2395
2396         ilk_compute_wm_level(dev_priv, intel_crtc, 0, cstate,
2397                              pristate, sprstate, curstate, &pipe_wm->raw_wm[0]);
2398
2399         memset(&pipe_wm->wm, 0, sizeof(pipe_wm->wm));
2400         pipe_wm->wm[0] = pipe_wm->raw_wm[0];
2401
2402         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
2403                 pipe_wm->linetime = hsw_compute_linetime_wm(cstate);
2404
2405         if (!ilk_validate_pipe_wm(dev, pipe_wm))
2406                 return -EINVAL;
2407
2408         ilk_compute_wm_reg_maximums(dev_priv, 1, &max);
2409
2410         for (level = 1; level <= max_level; level++) {
2411                 struct intel_wm_level *wm = &pipe_wm->raw_wm[level];
2412
2413                 ilk_compute_wm_level(dev_priv, intel_crtc, level, cstate,
2414                                      pristate, sprstate, curstate, wm);
2415
2416                 /*
2417                  * Disable any watermark level that exceeds the
2418                  * register maximums since such watermarks are
2419                  * always invalid.
2420                  */
2421                 if (level > usable_level)
2422                         continue;
2423
2424                 if (ilk_validate_wm_level(level, &max, wm))
2425                         pipe_wm->wm[level] = *wm;
2426                 else
2427                         usable_level = level;
2428         }
2429
2430         return 0;
2431 }
2432
2433 /*
2434  * Build a set of 'intermediate' watermark values that satisfy both the old
2435  * state and the new state.  These can be programmed to the hardware
2436  * immediately.
2437  */
2438 static int ilk_compute_intermediate_wm(struct drm_device *dev,
2439                                        struct intel_crtc *intel_crtc,
2440                                        struct intel_crtc_state *newstate)
2441 {
2442         struct intel_pipe_wm *a = &newstate->wm.ilk.intermediate;
2443         struct intel_pipe_wm *b = &intel_crtc->wm.active.ilk;
2444         int level, max_level = ilk_wm_max_level(to_i915(dev));
2445
2446         /*
2447          * Start with the final, target watermarks, then combine with the
2448          * currently active watermarks to get values that are safe both before
2449          * and after the vblank.
2450          */
2451         *a = newstate->wm.ilk.optimal;
2452         a->pipe_enabled |= b->pipe_enabled;
2453         a->sprites_enabled |= b->sprites_enabled;
2454         a->sprites_scaled |= b->sprites_scaled;
2455
2456         for (level = 0; level <= max_level; level++) {
2457                 struct intel_wm_level *a_wm = &a->wm[level];
2458                 const struct intel_wm_level *b_wm = &b->wm[level];
2459
2460                 a_wm->enable &= b_wm->enable;
2461                 a_wm->pri_val = max(a_wm->pri_val, b_wm->pri_val);
2462                 a_wm->spr_val = max(a_wm->spr_val, b_wm->spr_val);
2463                 a_wm->cur_val = max(a_wm->cur_val, b_wm->cur_val);
2464                 a_wm->fbc_val = max(a_wm->fbc_val, b_wm->fbc_val);
2465         }
2466
2467         /*
2468          * We need to make sure that these merged watermark values are
2469          * actually a valid configuration themselves.  If they're not,
2470          * there's no safe way to transition from the old state to
2471          * the new state, so we need to fail the atomic transaction.
2472          */
2473         if (!ilk_validate_pipe_wm(dev, a))
2474                 return -EINVAL;
2475
2476         /*
2477          * If our intermediate WM are identical to the final WM, then we can
2478          * omit the post-vblank programming; only update if it's different.
2479          */
2480         if (memcmp(a, &newstate->wm.ilk.optimal, sizeof(*a)) == 0)
2481                 newstate->wm.need_postvbl_update = false;
2482
2483         return 0;
2484 }
2485
2486 /*
2487  * Merge the watermarks from all active pipes for a specific level.
2488  */
2489 static void ilk_merge_wm_level(struct drm_device *dev,
2490                                int level,
2491                                struct intel_wm_level *ret_wm)
2492 {
2493         const struct intel_crtc *intel_crtc;
2494
2495         ret_wm->enable = true;
2496
2497         for_each_intel_crtc(dev, intel_crtc) {
2498                 const struct intel_pipe_wm *active = &intel_crtc->wm.active.ilk;
2499                 const struct intel_wm_level *wm = &active->wm[level];
2500
2501                 if (!active->pipe_enabled)
2502                         continue;
2503
2504                 /*
2505                  * The watermark values may have been used in the past,
2506                  * so we must maintain them in the registers for some
2507                  * time even if the level is now disabled.
2508                  */
2509                 if (!wm->enable)
2510                         ret_wm->enable = false;
2511
2512                 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
2513                 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
2514                 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
2515                 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
2516         }
2517 }
2518
2519 /*
2520  * Merge all low power watermarks for all active pipes.
2521  */
2522 static void ilk_wm_merge(struct drm_device *dev,
2523                          const struct intel_wm_config *config,
2524                          const struct ilk_wm_maximums *max,
2525                          struct intel_pipe_wm *merged)
2526 {
2527         struct drm_i915_private *dev_priv = to_i915(dev);
2528         int level, max_level = ilk_wm_max_level(dev_priv);
2529         int last_enabled_level = max_level;
2530
2531         /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2532         if ((INTEL_GEN(dev_priv) <= 6 || IS_IVYBRIDGE(dev_priv)) &&
2533             config->num_pipes_active > 1)
2534                 last_enabled_level = 0;
2535
2536         /* ILK: FBC WM must be disabled always */
2537         merged->fbc_wm_enabled = INTEL_GEN(dev_priv) >= 6;
2538
2539         /* merge each WM1+ level */
2540         for (level = 1; level <= max_level; level++) {
2541                 struct intel_wm_level *wm = &merged->wm[level];
2542
2543                 ilk_merge_wm_level(dev, level, wm);
2544
2545                 if (level > last_enabled_level)
2546                         wm->enable = false;
2547                 else if (!ilk_validate_wm_level(level, max, wm))
2548                         /* make sure all following levels get disabled */
2549                         last_enabled_level = level - 1;
2550
2551                 /*
2552                  * The spec says it is preferred to disable
2553                  * FBC WMs instead of disabling a WM level.
2554                  */
2555                 if (wm->fbc_val > max->fbc) {
2556                         if (wm->enable)
2557                                 merged->fbc_wm_enabled = false;
2558                         wm->fbc_val = 0;
2559                 }
2560         }
2561
2562         /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2563         /*
2564          * FIXME this is racy. FBC might get enabled later.
2565          * What we should check here is whether FBC can be
2566          * enabled sometime later.
2567          */
2568         if (IS_GEN5(dev_priv) && !merged->fbc_wm_enabled &&
2569             intel_fbc_is_active(dev_priv)) {
2570                 for (level = 2; level <= max_level; level++) {
2571                         struct intel_wm_level *wm = &merged->wm[level];
2572
2573                         wm->enable = false;
2574                 }
2575         }
2576 }
2577
2578 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
2579 {
2580         /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2581         return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
2582 }
2583
2584 /* The value we need to program into the WM_LPx latency field */
2585 static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
2586 {
2587         struct drm_i915_private *dev_priv = to_i915(dev);
2588
2589         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
2590                 return 2 * level;
2591         else
2592                 return dev_priv->wm.pri_latency[level];
2593 }
2594
2595 static void ilk_compute_wm_results(struct drm_device *dev,
2596                                    const struct intel_pipe_wm *merged,
2597                                    enum intel_ddb_partitioning partitioning,
2598                                    struct ilk_wm_values *results)
2599 {
2600         struct drm_i915_private *dev_priv = to_i915(dev);
2601         struct intel_crtc *intel_crtc;
2602         int level, wm_lp;
2603
2604         results->enable_fbc_wm = merged->fbc_wm_enabled;
2605         results->partitioning = partitioning;
2606
2607         /* LP1+ register values */
2608         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2609                 const struct intel_wm_level *r;
2610
2611                 level = ilk_wm_lp_to_level(wm_lp, merged);
2612
2613                 r = &merged->wm[level];
2614
2615                 /*
2616                  * Maintain the watermark values even if the level is
2617                  * disabled. Doing otherwise could cause underruns.
2618                  */
2619                 results->wm_lp[wm_lp - 1] =
2620                         (ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
2621                         (r->pri_val << WM1_LP_SR_SHIFT) |
2622                         r->cur_val;
2623
2624                 if (r->enable)
2625                         results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
2626
2627                 if (INTEL_GEN(dev_priv) >= 8)
2628                         results->wm_lp[wm_lp - 1] |=
2629                                 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
2630                 else
2631                         results->wm_lp[wm_lp - 1] |=
2632                                 r->fbc_val << WM1_LP_FBC_SHIFT;
2633
2634                 /*
2635                  * Always set WM1S_LP_EN when spr_val != 0, even if the
2636                  * level is disabled. Doing otherwise could cause underruns.
2637                  */
2638                 if (INTEL_GEN(dev_priv) <= 6 && r->spr_val) {
2639                         WARN_ON(wm_lp != 1);
2640                         results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
2641                 } else
2642                         results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2643         }
2644
2645         /* LP0 register values */
2646         for_each_intel_crtc(dev, intel_crtc) {
2647                 enum pipe pipe = intel_crtc->pipe;
2648                 const struct intel_wm_level *r =
2649                         &intel_crtc->wm.active.ilk.wm[0];
2650
2651                 if (WARN_ON(!r->enable))
2652                         continue;
2653
2654                 results->wm_linetime[pipe] = intel_crtc->wm.active.ilk.linetime;
2655
2656                 results->wm_pipe[pipe] =
2657                         (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
2658                         (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
2659                         r->cur_val;
2660         }
2661 }
2662
2663 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2664  * case both are at the same level. Prefer r1 in case they're the same. */
2665 static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
2666                                                   struct intel_pipe_wm *r1,
2667                                                   struct intel_pipe_wm *r2)
2668 {
2669         int level, max_level = ilk_wm_max_level(to_i915(dev));
2670         int level1 = 0, level2 = 0;
2671
2672         for (level = 1; level <= max_level; level++) {
2673                 if (r1->wm[level].enable)
2674                         level1 = level;
2675                 if (r2->wm[level].enable)
2676                         level2 = level;
2677         }
2678
2679         if (level1 == level2) {
2680                 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
2681                         return r2;
2682                 else
2683                         return r1;
2684         } else if (level1 > level2) {
2685                 return r1;
2686         } else {
2687                 return r2;
2688         }
2689 }
2690
2691 /* dirty bits used to track which watermarks need changes */
2692 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2693 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2694 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2695 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2696 #define WM_DIRTY_FBC (1 << 24)
2697 #define WM_DIRTY_DDB (1 << 25)
2698
2699 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private *dev_priv,
2700                                          const struct ilk_wm_values *old,
2701                                          const struct ilk_wm_values *new)
2702 {
2703         unsigned int dirty = 0;
2704         enum pipe pipe;
2705         int wm_lp;
2706
2707         for_each_pipe(dev_priv, pipe) {
2708                 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
2709                         dirty |= WM_DIRTY_LINETIME(pipe);
2710                         /* Must disable LP1+ watermarks too */
2711                         dirty |= WM_DIRTY_LP_ALL;
2712                 }
2713
2714                 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
2715                         dirty |= WM_DIRTY_PIPE(pipe);
2716                         /* Must disable LP1+ watermarks too */
2717                         dirty |= WM_DIRTY_LP_ALL;
2718                 }
2719         }
2720
2721         if (old->enable_fbc_wm != new->enable_fbc_wm) {
2722                 dirty |= WM_DIRTY_FBC;
2723                 /* Must disable LP1+ watermarks too */
2724                 dirty |= WM_DIRTY_LP_ALL;
2725         }
2726
2727         if (old->partitioning != new->partitioning) {
2728                 dirty |= WM_DIRTY_DDB;
2729                 /* Must disable LP1+ watermarks too */
2730                 dirty |= WM_DIRTY_LP_ALL;
2731         }
2732
2733         /* LP1+ watermarks already deemed dirty, no need to continue */
2734         if (dirty & WM_DIRTY_LP_ALL)
2735                 return dirty;
2736
2737         /* Find the lowest numbered LP1+ watermark in need of an update... */
2738         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2739                 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
2740                     old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
2741                         break;
2742         }
2743
2744         /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2745         for (; wm_lp <= 3; wm_lp++)
2746                 dirty |= WM_DIRTY_LP(wm_lp);
2747
2748         return dirty;
2749 }
2750
2751 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
2752                                unsigned int dirty)
2753 {
2754         struct ilk_wm_values *previous = &dev_priv->wm.hw;
2755         bool changed = false;
2756
2757         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
2758                 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
2759                 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
2760                 changed = true;
2761         }
2762         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
2763                 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
2764                 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
2765                 changed = true;
2766         }
2767         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
2768                 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
2769                 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
2770                 changed = true;
2771         }
2772
2773         /*
2774          * Don't touch WM1S_LP_EN here.
2775          * Doing so could cause underruns.
2776          */
2777
2778         return changed;
2779 }
2780
2781 /*
2782  * The spec says we shouldn't write when we don't need, because every write
2783  * causes WMs to be re-evaluated, expending some power.
2784  */
2785 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
2786                                 struct ilk_wm_values *results)
2787 {
2788         struct ilk_wm_values *previous = &dev_priv->wm.hw;
2789         unsigned int dirty;
2790         uint32_t val;
2791
2792         dirty = ilk_compute_wm_dirty(dev_priv, previous, results);
2793         if (!dirty)
2794                 return;
2795
2796         _ilk_disable_lp_wm(dev_priv, dirty);
2797
2798         if (dirty & WM_DIRTY_PIPE(PIPE_A))
2799                 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2800         if (dirty & WM_DIRTY_PIPE(PIPE_B))
2801                 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
2802         if (dirty & WM_DIRTY_PIPE(PIPE_C))
2803                 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
2804
2805         if (dirty & WM_DIRTY_LINETIME(PIPE_A))
2806                 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
2807         if (dirty & WM_DIRTY_LINETIME(PIPE_B))
2808                 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
2809         if (dirty & WM_DIRTY_LINETIME(PIPE_C))
2810                 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
2811
2812         if (dirty & WM_DIRTY_DDB) {
2813                 if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv)) {
2814                         val = I915_READ(WM_MISC);
2815                         if (results->partitioning == INTEL_DDB_PART_1_2)
2816                                 val &= ~WM_MISC_DATA_PARTITION_5_6;
2817                         else
2818                                 val |= WM_MISC_DATA_PARTITION_5_6;
2819                         I915_WRITE(WM_MISC, val);
2820                 } else {
2821                         val = I915_READ(DISP_ARB_CTL2);
2822                         if (results->partitioning == INTEL_DDB_PART_1_2)
2823                                 val &= ~DISP_DATA_PARTITION_5_6;
2824                         else
2825                                 val |= DISP_DATA_PARTITION_5_6;
2826                         I915_WRITE(DISP_ARB_CTL2, val);
2827                 }
2828         }
2829
2830         if (dirty & WM_DIRTY_FBC) {
2831                 val = I915_READ(DISP_ARB_CTL);
2832                 if (results->enable_fbc_wm)
2833                         val &= ~DISP_FBC_WM_DIS;
2834                 else
2835                         val |= DISP_FBC_WM_DIS;
2836                 I915_WRITE(DISP_ARB_CTL, val);
2837         }
2838
2839         if (dirty & WM_DIRTY_LP(1) &&
2840             previous->wm_lp_spr[0] != results->wm_lp_spr[0])
2841                 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
2842
2843         if (INTEL_GEN(dev_priv) >= 7) {
2844                 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
2845                         I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
2846                 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
2847                         I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
2848         }
2849
2850         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
2851                 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
2852         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
2853                 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
2854         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
2855                 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
2856
2857         dev_priv->wm.hw = *results;
2858 }
2859
2860 bool ilk_disable_lp_wm(struct drm_device *dev)
2861 {
2862         struct drm_i915_private *dev_priv = to_i915(dev);
2863
2864         return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
2865 }
2866
2867 #define SKL_SAGV_BLOCK_TIME     30 /* µs */
2868
2869 /*
2870  * Return the index of a plane in the SKL DDB and wm result arrays.  Primary
2871  * plane is always in slot 0, cursor is always in slot I915_MAX_PLANES-1, and
2872  * other universal planes are in indices 1..n.  Note that this may leave unused
2873  * indices between the top "sprite" plane and the cursor.
2874  */
2875 static int
2876 skl_wm_plane_id(const struct intel_plane *plane)
2877 {
2878         switch (plane->base.type) {
2879         case DRM_PLANE_TYPE_PRIMARY:
2880                 return 0;
2881         case DRM_PLANE_TYPE_CURSOR:
2882                 return PLANE_CURSOR;
2883         case DRM_PLANE_TYPE_OVERLAY:
2884                 return plane->plane + 1;
2885         default:
2886                 MISSING_CASE(plane->base.type);
2887                 return plane->plane;
2888         }
2889 }
2890
2891 /*
2892  * FIXME: We still don't have the proper code detect if we need to apply the WA,
2893  * so assume we'll always need it in order to avoid underruns.
2894  */
2895 static bool skl_needs_memory_bw_wa(struct intel_atomic_state *state)
2896 {
2897         struct drm_i915_private *dev_priv = to_i915(state->base.dev);
2898
2899         if (IS_SKYLAKE(dev_priv) || IS_BROXTON(dev_priv) ||
2900             IS_KABYLAKE(dev_priv))
2901                 return true;
2902
2903         return false;
2904 }
2905
2906 static bool
2907 intel_has_sagv(struct drm_i915_private *dev_priv)
2908 {
2909         if (IS_KABYLAKE(dev_priv))
2910                 return true;
2911
2912         if (IS_SKYLAKE(dev_priv) &&
2913             dev_priv->sagv_status != I915_SAGV_NOT_CONTROLLED)
2914                 return true;
2915
2916         return false;
2917 }
2918
2919 /*
2920  * SAGV dynamically adjusts the system agent voltage and clock frequencies
2921  * depending on power and performance requirements. The display engine access
2922  * to system memory is blocked during the adjustment time. Because of the
2923  * blocking time, having this enabled can cause full system hangs and/or pipe
2924  * underruns if we don't meet all of the following requirements:
2925  *
2926  *  - <= 1 pipe enabled
2927  *  - All planes can enable watermarks for latencies >= SAGV engine block time
2928  *  - We're not using an interlaced display configuration
2929  */
2930 int
2931 intel_enable_sagv(struct drm_i915_private *dev_priv)
2932 {
2933         int ret;
2934
2935         if (!intel_has_sagv(dev_priv))
2936                 return 0;
2937
2938         if (dev_priv->sagv_status == I915_SAGV_ENABLED)
2939                 return 0;
2940
2941         DRM_DEBUG_KMS("Enabling the SAGV\n");
2942         mutex_lock(&dev_priv->rps.hw_lock);
2943
2944         ret = sandybridge_pcode_write(dev_priv, GEN9_PCODE_SAGV_CONTROL,
2945                                       GEN9_SAGV_ENABLE);
2946
2947         /* We don't need to wait for the SAGV when enabling */
2948         mutex_unlock(&dev_priv->rps.hw_lock);
2949
2950         /*
2951          * Some skl systems, pre-release machines in particular,
2952          * don't actually have an SAGV.
2953          */
2954         if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
2955                 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
2956                 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
2957                 return 0;
2958         } else if (ret < 0) {
2959                 DRM_ERROR("Failed to enable the SAGV\n");
2960                 return ret;
2961         }
2962
2963         dev_priv->sagv_status = I915_SAGV_ENABLED;
2964         return 0;
2965 }
2966
2967 int
2968 intel_disable_sagv(struct drm_i915_private *dev_priv)
2969 {
2970         int ret;
2971
2972         if (!intel_has_sagv(dev_priv))
2973                 return 0;
2974
2975         if (dev_priv->sagv_status == I915_SAGV_DISABLED)
2976                 return 0;
2977
2978         DRM_DEBUG_KMS("Disabling the SAGV\n");
2979         mutex_lock(&dev_priv->rps.hw_lock);
2980
2981         /* bspec says to keep retrying for at least 1 ms */
2982         ret = skl_pcode_request(dev_priv, GEN9_PCODE_SAGV_CONTROL,
2983                                 GEN9_SAGV_DISABLE,
2984                                 GEN9_SAGV_IS_DISABLED, GEN9_SAGV_IS_DISABLED,
2985                                 1);
2986         mutex_unlock(&dev_priv->rps.hw_lock);
2987
2988         /*
2989          * Some skl systems, pre-release machines in particular,
2990          * don't actually have an SAGV.
2991          */
2992         if (IS_SKYLAKE(dev_priv) && ret == -ENXIO) {
2993                 DRM_DEBUG_DRIVER("No SAGV found on system, ignoring\n");
2994                 dev_priv->sagv_status = I915_SAGV_NOT_CONTROLLED;
2995                 return 0;
2996         } else if (ret < 0) {
2997                 DRM_ERROR("Failed to disable the SAGV (%d)\n", ret);
2998                 return ret;
2999         }
3000
3001         dev_priv->sagv_status = I915_SAGV_DISABLED;
3002         return 0;
3003 }
3004
3005 bool intel_can_enable_sagv(struct drm_atomic_state *state)
3006 {
3007         struct drm_device *dev = state->dev;
3008         struct drm_i915_private *dev_priv = to_i915(dev);
3009         struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3010         struct intel_crtc *crtc;
3011         struct intel_plane *plane;
3012         struct intel_crtc_state *cstate;
3013         struct skl_plane_wm *wm;
3014         enum pipe pipe;
3015         int level, latency;
3016
3017         if (!intel_has_sagv(dev_priv))
3018                 return false;
3019
3020         /*
3021          * SKL workaround: bspec recommends we disable the SAGV when we have
3022          * more then one pipe enabled
3023          *
3024          * If there are no active CRTCs, no additional checks need be performed
3025          */
3026         if (hweight32(intel_state->active_crtcs) == 0)
3027                 return true;
3028         else if (hweight32(intel_state->active_crtcs) > 1)
3029                 return false;
3030
3031         /* Since we're now guaranteed to only have one active CRTC... */
3032         pipe = ffs(intel_state->active_crtcs) - 1;
3033         crtc = intel_get_crtc_for_pipe(dev_priv, pipe);
3034         cstate = to_intel_crtc_state(crtc->base.state);
3035
3036         if (crtc->base.state->adjusted_mode.flags & DRM_MODE_FLAG_INTERLACE)
3037                 return false;
3038
3039         for_each_intel_plane_on_crtc(dev, crtc, plane) {
3040                 wm = &cstate->wm.skl.optimal.planes[skl_wm_plane_id(plane)];
3041
3042                 /* Skip this plane if it's not enabled */
3043                 if (!wm->wm[0].plane_en)
3044                         continue;
3045
3046                 /* Find the highest enabled wm level for this plane */
3047                 for (level = ilk_wm_max_level(dev_priv);
3048                      !wm->wm[level].plane_en; --level)
3049                      { }
3050
3051                 latency = dev_priv->wm.skl_latency[level];
3052
3053                 if (skl_needs_memory_bw_wa(intel_state) &&
3054                     plane->base.state->fb->modifier ==
3055                     I915_FORMAT_MOD_X_TILED)
3056                         latency += 15;
3057
3058                 /*
3059                  * If any of the planes on this pipe don't enable wm levels
3060                  * that incur memory latencies higher then 30µs we can't enable
3061                  * the SAGV
3062                  */
3063                 if (latency < SKL_SAGV_BLOCK_TIME)
3064                         return false;
3065         }
3066
3067         return true;
3068 }
3069
3070 static void
3071 skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
3072                                    const struct intel_crtc_state *cstate,
3073                                    struct skl_ddb_entry *alloc, /* out */
3074                                    int *num_active /* out */)
3075 {
3076         struct drm_atomic_state *state = cstate->base.state;
3077         struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3078         struct drm_i915_private *dev_priv = to_i915(dev);
3079         struct drm_crtc *for_crtc = cstate->base.crtc;
3080         unsigned int pipe_size, ddb_size;
3081         int nth_active_pipe;
3082
3083         if (WARN_ON(!state) || !cstate->base.active) {
3084                 alloc->start = 0;
3085                 alloc->end = 0;
3086                 *num_active = hweight32(dev_priv->active_crtcs);
3087                 return;
3088         }
3089
3090         if (intel_state->active_pipe_changes)
3091                 *num_active = hweight32(intel_state->active_crtcs);
3092         else
3093                 *num_active = hweight32(dev_priv->active_crtcs);
3094
3095         ddb_size = INTEL_INFO(dev_priv)->ddb_size;
3096         WARN_ON(ddb_size == 0);
3097
3098         ddb_size -= 4; /* 4 blocks for bypass path allocation */
3099
3100         /*
3101          * If the state doesn't change the active CRTC's, then there's
3102          * no need to recalculate; the existing pipe allocation limits
3103          * should remain unchanged.  Note that we're safe from racing
3104          * commits since any racing commit that changes the active CRTC
3105          * list would need to grab _all_ crtc locks, including the one
3106          * we currently hold.
3107          */
3108         if (!intel_state->active_pipe_changes) {
3109                 /*
3110                  * alloc may be cleared by clear_intel_crtc_state,
3111                  * copy from old state to be sure
3112                  */
3113                 *alloc = to_intel_crtc_state(for_crtc->state)->wm.skl.ddb;
3114                 return;
3115         }
3116
3117         nth_active_pipe = hweight32(intel_state->active_crtcs &
3118                                     (drm_crtc_mask(for_crtc) - 1));
3119         pipe_size = ddb_size / hweight32(intel_state->active_crtcs);
3120         alloc->start = nth_active_pipe * ddb_size / *num_active;
3121         alloc->end = alloc->start + pipe_size;
3122 }
3123
3124 static unsigned int skl_cursor_allocation(int num_active)
3125 {
3126         if (num_active == 1)
3127                 return 32;
3128
3129         return 8;
3130 }
3131
3132 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry *entry, u32 reg)
3133 {
3134         entry->start = reg & 0x3ff;
3135         entry->end = (reg >> 16) & 0x3ff;
3136         if (entry->end)
3137                 entry->end += 1;
3138 }
3139
3140 void skl_ddb_get_hw_state(struct drm_i915_private *dev_priv,
3141                           struct skl_ddb_allocation *ddb /* out */)
3142 {
3143         enum pipe pipe;
3144         int plane;
3145         u32 val;
3146
3147         memset(ddb, 0, sizeof(*ddb));
3148
3149         for_each_pipe(dev_priv, pipe) {
3150                 enum intel_display_power_domain power_domain;
3151
3152                 power_domain = POWER_DOMAIN_PIPE(pipe);
3153                 if (!intel_display_power_get_if_enabled(dev_priv, power_domain))
3154                         continue;
3155
3156                 for_each_universal_plane(dev_priv, pipe, plane) {
3157                         val = I915_READ(PLANE_BUF_CFG(pipe, plane));
3158                         skl_ddb_entry_init_from_hw(&ddb->plane[pipe][plane],
3159                                                    val);
3160                 }
3161
3162                 val = I915_READ(CUR_BUF_CFG(pipe));
3163                 skl_ddb_entry_init_from_hw(&ddb->plane[pipe][PLANE_CURSOR],
3164                                            val);
3165
3166                 intel_display_power_put(dev_priv, power_domain);
3167         }
3168 }
3169
3170 /*
3171  * Determines the downscale amount of a plane for the purposes of watermark calculations.
3172  * The bspec defines downscale amount as:
3173  *
3174  * """
3175  * Horizontal down scale amount = maximum[1, Horizontal source size /
3176  *                                           Horizontal destination size]
3177  * Vertical down scale amount = maximum[1, Vertical source size /
3178  *                                         Vertical destination size]
3179  * Total down scale amount = Horizontal down scale amount *
3180  *                           Vertical down scale amount
3181  * """
3182  *
3183  * Return value is provided in 16.16 fixed point form to retain fractional part.
3184  * Caller should take care of dividing & rounding off the value.
3185  */
3186 static uint32_t
3187 skl_plane_downscale_amount(const struct intel_plane_state *pstate)
3188 {
3189         uint32_t downscale_h, downscale_w;
3190         uint32_t src_w, src_h, dst_w, dst_h;
3191
3192         if (WARN_ON(!pstate->base.visible))
3193                 return DRM_PLANE_HELPER_NO_SCALING;
3194
3195         /* n.b., src is 16.16 fixed point, dst is whole integer */
3196         src_w = drm_rect_width(&pstate->base.src);
3197         src_h = drm_rect_height(&pstate->base.src);
3198         dst_w = drm_rect_width(&pstate->base.dst);
3199         dst_h = drm_rect_height(&pstate->base.dst);
3200         if (drm_rotation_90_or_270(pstate->base.rotation))
3201                 swap(dst_w, dst_h);
3202
3203         downscale_h = max(src_h / dst_h, (uint32_t)DRM_PLANE_HELPER_NO_SCALING);
3204         downscale_w = max(src_w / dst_w, (uint32_t)DRM_PLANE_HELPER_NO_SCALING);
3205
3206         /* Provide result in 16.16 fixed point */
3207         return (uint64_t)downscale_w * downscale_h >> 16;
3208 }
3209
3210 static unsigned int
3211 skl_plane_relative_data_rate(const struct intel_crtc_state *cstate,
3212                              const struct drm_plane_state *pstate,
3213                              int y)
3214 {
3215         struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
3216         struct drm_framebuffer *fb = pstate->fb;
3217         uint32_t down_scale_amount, data_rate;
3218         uint32_t width = 0, height = 0;
3219         unsigned format = fb ? fb->pixel_format : DRM_FORMAT_XRGB8888;
3220
3221         if (!intel_pstate->base.visible)
3222                 return 0;
3223         if (pstate->plane->type == DRM_PLANE_TYPE_CURSOR)
3224                 return 0;
3225         if (y && format != DRM_FORMAT_NV12)
3226                 return 0;
3227
3228         width = drm_rect_width(&intel_pstate->base.src) >> 16;
3229         height = drm_rect_height(&intel_pstate->base.src) >> 16;
3230
3231         if (drm_rotation_90_or_270(pstate->rotation))
3232                 swap(width, height);
3233
3234         /* for planar format */
3235         if (format == DRM_FORMAT_NV12) {
3236                 if (y)  /* y-plane data rate */
3237                         data_rate = width * height *
3238                                 drm_format_plane_cpp(format, 0);
3239                 else    /* uv-plane data rate */
3240                         data_rate = (width / 2) * (height / 2) *
3241                                 drm_format_plane_cpp(format, 1);
3242         } else {
3243                 /* for packed formats */
3244                 data_rate = width * height * drm_format_plane_cpp(format, 0);
3245         }
3246
3247         down_scale_amount = skl_plane_downscale_amount(intel_pstate);
3248
3249         return (uint64_t)data_rate * down_scale_amount >> 16;
3250 }
3251
3252 /*
3253  * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
3254  * a 8192x4096@32bpp framebuffer:
3255  *   3 * 4096 * 8192  * 4 < 2^32
3256  */
3257 static unsigned int
3258 skl_get_total_relative_data_rate(struct intel_crtc_state *intel_cstate,
3259                                  unsigned *plane_data_rate,
3260                                  unsigned *plane_y_data_rate)
3261 {
3262         struct drm_crtc_state *cstate = &intel_cstate->base;
3263         struct drm_atomic_state *state = cstate->state;
3264         struct drm_plane *plane;
3265         const struct intel_plane *intel_plane;
3266         const struct drm_plane_state *pstate;
3267         unsigned int rate, total_data_rate = 0;
3268         int id;
3269
3270         if (WARN_ON(!state))
3271                 return 0;
3272
3273         /* Calculate and cache data rate for each plane */
3274         drm_atomic_crtc_state_for_each_plane_state(plane, pstate, cstate) {
3275                 id = skl_wm_plane_id(to_intel_plane(plane));
3276                 intel_plane = to_intel_plane(plane);
3277
3278                 /* packed/uv */
3279                 rate = skl_plane_relative_data_rate(intel_cstate,
3280                                                     pstate, 0);
3281                 plane_data_rate[id] = rate;
3282
3283                 total_data_rate += rate;
3284
3285                 /* y-plane */
3286                 rate = skl_plane_relative_data_rate(intel_cstate,
3287                                                     pstate, 1);
3288                 plane_y_data_rate[id] = rate;
3289
3290                 total_data_rate += rate;
3291         }
3292
3293         return total_data_rate;
3294 }
3295
3296 static uint16_t
3297 skl_ddb_min_alloc(const struct drm_plane_state *pstate,
3298                   const int y)
3299 {
3300         struct drm_framebuffer *fb = pstate->fb;
3301         struct intel_plane_state *intel_pstate = to_intel_plane_state(pstate);
3302         uint32_t src_w, src_h;
3303         uint32_t min_scanlines = 8;
3304         uint8_t plane_bpp;
3305
3306         if (WARN_ON(!fb))
3307                 return 0;
3308
3309         /* For packed formats, no y-plane, return 0 */
3310         if (y && fb->pixel_format != DRM_FORMAT_NV12)
3311                 return 0;
3312
3313         /* For Non Y-tile return 8-blocks */
3314         if (fb->modifier != I915_FORMAT_MOD_Y_TILED &&
3315             fb->modifier != I915_FORMAT_MOD_Yf_TILED)
3316                 return 8;
3317
3318         src_w = drm_rect_width(&intel_pstate->base.src) >> 16;
3319         src_h = drm_rect_height(&intel_pstate->base.src) >> 16;
3320
3321         if (drm_rotation_90_or_270(pstate->rotation))
3322                 swap(src_w, src_h);
3323
3324         /* Halve UV plane width and height for NV12 */
3325         if (fb->pixel_format == DRM_FORMAT_NV12 && !y) {
3326                 src_w /= 2;
3327                 src_h /= 2;
3328         }
3329
3330         if (fb->pixel_format == DRM_FORMAT_NV12 && !y)
3331                 plane_bpp = drm_format_plane_cpp(fb->pixel_format, 1);
3332         else
3333                 plane_bpp = drm_format_plane_cpp(fb->pixel_format, 0);
3334
3335         if (drm_rotation_90_or_270(pstate->rotation)) {
3336                 switch (plane_bpp) {
3337                 case 1:
3338                         min_scanlines = 32;
3339                         break;
3340                 case 2:
3341                         min_scanlines = 16;
3342                         break;
3343                 case 4:
3344                         min_scanlines = 8;
3345                         break;
3346                 case 8:
3347                         min_scanlines = 4;
3348                         break;
3349                 default:
3350                         WARN(1, "Unsupported pixel depth %u for rotation",
3351                              plane_bpp);
3352                         min_scanlines = 32;
3353                 }
3354         }
3355
3356         return DIV_ROUND_UP((4 * src_w * plane_bpp), 512) * min_scanlines/4 + 3;
3357 }
3358
3359 static void
3360 skl_ddb_calc_min(const struct intel_crtc_state *cstate, int num_active,
3361                  uint16_t *minimum, uint16_t *y_minimum)
3362 {
3363         const struct drm_plane_state *pstate;
3364         struct drm_plane *plane;
3365
3366         drm_atomic_crtc_state_for_each_plane_state(plane, pstate, &cstate->base) {
3367                 struct intel_plane *intel_plane = to_intel_plane(plane);
3368                 int id = skl_wm_plane_id(intel_plane);
3369
3370                 if (id == PLANE_CURSOR)
3371                         continue;
3372
3373                 if (!pstate->visible)
3374                         continue;
3375
3376                 minimum[id] = skl_ddb_min_alloc(pstate, 0);
3377                 y_minimum[id] = skl_ddb_min_alloc(pstate, 1);
3378         }
3379
3380         minimum[PLANE_CURSOR] = skl_cursor_allocation(num_active);
3381 }
3382
3383 static int
3384 skl_allocate_pipe_ddb(struct intel_crtc_state *cstate,
3385                       struct skl_ddb_allocation *ddb /* out */)
3386 {
3387         struct drm_atomic_state *state = cstate->base.state;
3388         struct drm_crtc *crtc = cstate->base.crtc;
3389         struct drm_device *dev = crtc->dev;
3390         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3391         enum pipe pipe = intel_crtc->pipe;
3392         struct skl_ddb_entry *alloc = &cstate->wm.skl.ddb;
3393         uint16_t alloc_size, start;
3394         uint16_t minimum[I915_MAX_PLANES] = {};
3395         uint16_t y_minimum[I915_MAX_PLANES] = {};
3396         unsigned int total_data_rate;
3397         int num_active;
3398         int id, i;
3399         unsigned plane_data_rate[I915_MAX_PLANES] = {};
3400         unsigned plane_y_data_rate[I915_MAX_PLANES] = {};
3401
3402         /* Clear the partitioning for disabled planes. */
3403         memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
3404         memset(ddb->y_plane[pipe], 0, sizeof(ddb->y_plane[pipe]));
3405
3406         if (WARN_ON(!state))
3407                 return 0;
3408
3409         if (!cstate->base.active) {
3410                 alloc->start = alloc->end = 0;
3411                 return 0;
3412         }
3413
3414         skl_ddb_get_pipe_allocation_limits(dev, cstate, alloc, &num_active);
3415         alloc_size = skl_ddb_entry_size(alloc);
3416         if (alloc_size == 0) {
3417                 memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
3418                 return 0;
3419         }
3420
3421         skl_ddb_calc_min(cstate, num_active, minimum, y_minimum);
3422
3423         /*
3424          * 1. Allocate the mininum required blocks for each active plane
3425          * and allocate the cursor, it doesn't require extra allocation
3426          * proportional to the data rate.
3427          */
3428
3429         for (i = 0; i < I915_MAX_PLANES; i++) {
3430                 alloc_size -= minimum[i];
3431                 alloc_size -= y_minimum[i];
3432         }
3433
3434         ddb->plane[pipe][PLANE_CURSOR].start = alloc->end - minimum[PLANE_CURSOR];
3435         ddb->plane[pipe][PLANE_CURSOR].end = alloc->end;
3436
3437         /*
3438          * 2. Distribute the remaining space in proportion to the amount of
3439          * data each plane needs to fetch from memory.
3440          *
3441          * FIXME: we may not allocate every single block here.
3442          */
3443         total_data_rate = skl_get_total_relative_data_rate(cstate,
3444                                                            plane_data_rate,
3445                                                            plane_y_data_rate);
3446         if (total_data_rate == 0)
3447                 return 0;
3448
3449         start = alloc->start;
3450         for (id = 0; id < I915_MAX_PLANES; id++) {
3451                 unsigned int data_rate, y_data_rate;
3452                 uint16_t plane_blocks, y_plane_blocks = 0;
3453
3454                 if (id == PLANE_CURSOR)
3455                         continue;
3456
3457                 data_rate = plane_data_rate[id];
3458
3459                 /*
3460                  * allocation for (packed formats) or (uv-plane part of planar format):
3461                  * promote the expression to 64 bits to avoid overflowing, the
3462                  * result is < available as data_rate / total_data_rate < 1
3463                  */
3464                 plane_blocks = minimum[id];
3465                 plane_blocks += div_u64((uint64_t)alloc_size * data_rate,
3466                                         total_data_rate);
3467
3468                 /* Leave disabled planes at (0,0) */
3469                 if (data_rate) {
3470                         ddb->plane[pipe][id].start = start;
3471                         ddb->plane[pipe][id].end = start + plane_blocks;
3472                 }
3473
3474                 start += plane_blocks;
3475
3476                 /*
3477                  * allocation for y_plane part of planar format:
3478                  */
3479                 y_data_rate = plane_y_data_rate[id];
3480
3481                 y_plane_blocks = y_minimum[id];
3482                 y_plane_blocks += div_u64((uint64_t)alloc_size * y_data_rate,
3483                                         total_data_rate);
3484
3485                 if (y_data_rate) {
3486                         ddb->y_plane[pipe][id].start = start;
3487                         ddb->y_plane[pipe][id].end = start + y_plane_blocks;
3488                 }
3489
3490                 start += y_plane_blocks;
3491         }
3492
3493         return 0;
3494 }
3495
3496 /*
3497  * The max latency should be 257 (max the punit can code is 255 and we add 2us
3498  * for the read latency) and cpp should always be <= 8, so that
3499  * should allow pixel_rate up to ~2 GHz which seems sufficient since max
3500  * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
3501 */
3502 static uint32_t skl_wm_method1(uint32_t pixel_rate, uint8_t cpp, uint32_t latency)
3503 {
3504         uint32_t wm_intermediate_val, ret;
3505
3506         if (latency == 0)
3507                 return UINT_MAX;
3508
3509         wm_intermediate_val = latency * pixel_rate * cpp / 512;
3510         ret = DIV_ROUND_UP(wm_intermediate_val, 1000);
3511
3512         return ret;
3513 }
3514
3515 static uint32_t skl_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
3516                                uint32_t latency, uint32_t plane_blocks_per_line)
3517 {
3518         uint32_t ret;
3519         uint32_t wm_intermediate_val;
3520
3521         if (latency == 0)
3522                 return UINT_MAX;
3523
3524         wm_intermediate_val = latency * pixel_rate;
3525         ret = DIV_ROUND_UP(wm_intermediate_val, pipe_htotal * 1000) *
3526                                 plane_blocks_per_line;
3527
3528         return ret;
3529 }
3530
3531 static uint32_t skl_adjusted_plane_pixel_rate(const struct intel_crtc_state *cstate,
3532                                               struct intel_plane_state *pstate)
3533 {
3534         uint64_t adjusted_pixel_rate;
3535         uint64_t downscale_amount;
3536         uint64_t pixel_rate;
3537
3538         /* Shouldn't reach here on disabled planes... */
3539         if (WARN_ON(!pstate->base.visible))
3540                 return 0;
3541
3542         /*
3543          * Adjusted plane pixel rate is just the pipe's adjusted pixel rate
3544          * with additional adjustments for plane-specific scaling.
3545          */
3546         adjusted_pixel_rate = ilk_pipe_pixel_rate(cstate);
3547         downscale_amount = skl_plane_downscale_amount(pstate);
3548
3549         pixel_rate = adjusted_pixel_rate * downscale_amount >> 16;
3550         WARN_ON(pixel_rate != clamp_t(uint32_t, pixel_rate, 0, ~0));
3551
3552         return pixel_rate;
3553 }
3554
3555 static int skl_compute_plane_wm(const struct drm_i915_private *dev_priv,
3556                                 struct intel_crtc_state *cstate,
3557                                 struct intel_plane_state *intel_pstate,
3558                                 uint16_t ddb_allocation,
3559                                 int level,
3560                                 uint16_t *out_blocks, /* out */
3561                                 uint8_t *out_lines, /* out */
3562                                 bool *enabled /* out */)
3563 {
3564         struct drm_plane_state *pstate = &intel_pstate->base;
3565         struct drm_framebuffer *fb = pstate->fb;
3566         uint32_t latency = dev_priv->wm.skl_latency[level];
3567         uint32_t method1, method2;
3568         uint32_t plane_bytes_per_line, plane_blocks_per_line;
3569         uint32_t res_blocks, res_lines;
3570         uint32_t selected_result;
3571         uint8_t cpp;
3572         uint32_t width = 0, height = 0;
3573         uint32_t plane_pixel_rate;
3574         uint32_t y_tile_minimum, y_min_scanlines;
3575         struct intel_atomic_state *state =
3576                 to_intel_atomic_state(cstate->base.state);
3577         bool apply_memory_bw_wa = skl_needs_memory_bw_wa(state);
3578
3579         if (latency == 0 || !cstate->base.active || !intel_pstate->base.visible) {
3580                 *enabled = false;
3581                 return 0;
3582         }
3583
3584         if (apply_memory_bw_wa && fb->modifier == I915_FORMAT_MOD_X_TILED)
3585                 latency += 15;
3586
3587         width = drm_rect_width(&intel_pstate->base.src) >> 16;
3588         height = drm_rect_height(&intel_pstate->base.src) >> 16;
3589
3590         if (drm_rotation_90_or_270(pstate->rotation))
3591                 swap(width, height);
3592
3593         cpp = drm_format_plane_cpp(fb->pixel_format, 0);
3594         plane_pixel_rate = skl_adjusted_plane_pixel_rate(cstate, intel_pstate);
3595
3596         if (drm_rotation_90_or_270(pstate->rotation)) {
3597                 int cpp = (fb->pixel_format == DRM_FORMAT_NV12) ?
3598                         drm_format_plane_cpp(fb->pixel_format, 1) :
3599                         drm_format_plane_cpp(fb->pixel_format, 0);
3600
3601                 switch (cpp) {
3602                 case 1:
3603                         y_min_scanlines = 16;
3604                         break;
3605                 case 2:
3606                         y_min_scanlines = 8;
3607                         break;
3608                 case 4:
3609                         y_min_scanlines = 4;
3610                         break;
3611                 default:
3612                         MISSING_CASE(cpp);
3613                         return -EINVAL;
3614                 }
3615         } else {
3616                 y_min_scanlines = 4;
3617         }
3618
3619         if (apply_memory_bw_wa)
3620                 y_min_scanlines *= 2;
3621
3622         plane_bytes_per_line = width * cpp;
3623         if (fb->modifier == I915_FORMAT_MOD_Y_TILED ||
3624             fb->modifier == I915_FORMAT_MOD_Yf_TILED) {
3625                 plane_blocks_per_line =
3626                       DIV_ROUND_UP(plane_bytes_per_line * y_min_scanlines, 512);
3627                 plane_blocks_per_line /= y_min_scanlines;
3628         } else if (fb->modifier == DRM_FORMAT_MOD_NONE) {
3629                 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512)
3630                                         + 1;
3631         } else {
3632                 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
3633         }
3634
3635         method1 = skl_wm_method1(plane_pixel_rate, cpp, latency);
3636         method2 = skl_wm_method2(plane_pixel_rate,
3637                                  cstate->base.adjusted_mode.crtc_htotal,
3638                                  latency,
3639                                  plane_blocks_per_line);
3640
3641         y_tile_minimum = plane_blocks_per_line * y_min_scanlines;
3642
3643         if (fb->modifier == I915_FORMAT_MOD_Y_TILED ||
3644             fb->modifier == I915_FORMAT_MOD_Yf_TILED) {
3645                 selected_result = max(method2, y_tile_minimum);
3646         } else {
3647                 if ((cpp * cstate->base.adjusted_mode.crtc_htotal / 512 < 1) &&
3648                     (plane_bytes_per_line / 512 < 1))
3649                         selected_result = method2;
3650                 else if ((ddb_allocation / plane_blocks_per_line) >= 1)
3651                         selected_result = min(method1, method2);
3652                 else
3653                         selected_result = method1;
3654         }
3655
3656         res_blocks = selected_result + 1;
3657         res_lines = DIV_ROUND_UP(selected_result, plane_blocks_per_line);
3658
3659         if (level >= 1 && level <= 7) {
3660                 if (fb->modifier == I915_FORMAT_MOD_Y_TILED ||
3661                     fb->modifier == I915_FORMAT_MOD_Yf_TILED) {
3662                         res_blocks += y_tile_minimum;
3663                         res_lines += y_min_scanlines;
3664                 } else {
3665                         res_blocks++;
3666                 }
3667         }
3668
3669         if (res_blocks >= ddb_allocation || res_lines > 31) {
3670                 *enabled = false;
3671
3672                 /*
3673                  * If there are no valid level 0 watermarks, then we can't
3674                  * support this display configuration.
3675                  */
3676                 if (level) {
3677                         return 0;
3678                 } else {
3679                         DRM_DEBUG_KMS("Requested display configuration exceeds system watermark limitations\n");
3680                         DRM_DEBUG_KMS("Plane %d.%d: blocks required = %u/%u, lines required = %u/31\n",
3681                                       to_intel_crtc(cstate->base.crtc)->pipe,
3682                                       skl_wm_plane_id(to_intel_plane(pstate->plane)),
3683                                       res_blocks, ddb_allocation, res_lines);
3684
3685                         return -EINVAL;
3686                 }
3687         }
3688
3689         *out_blocks = res_blocks;
3690         *out_lines = res_lines;
3691         *enabled = true;
3692
3693         return 0;
3694 }
3695
3696 static int
3697 skl_compute_wm_level(const struct drm_i915_private *dev_priv,
3698                      struct skl_ddb_allocation *ddb,
3699                      struct intel_crtc_state *cstate,
3700                      struct intel_plane *intel_plane,
3701                      int level,
3702                      struct skl_wm_level *result)
3703 {
3704         struct drm_atomic_state *state = cstate->base.state;
3705         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
3706         struct drm_plane *plane = &intel_plane->base;
3707         struct intel_plane_state *intel_pstate = NULL;
3708         uint16_t ddb_blocks;
3709         enum pipe pipe = intel_crtc->pipe;
3710         int ret;
3711         int i = skl_wm_plane_id(intel_plane);
3712
3713         if (state)
3714                 intel_pstate =
3715                         intel_atomic_get_existing_plane_state(state,
3716                                                               intel_plane);
3717
3718         /*
3719          * Note: If we start supporting multiple pending atomic commits against
3720          * the same planes/CRTC's in the future, plane->state will no longer be
3721          * the correct pre-state to use for the calculations here and we'll
3722          * need to change where we get the 'unchanged' plane data from.
3723          *
3724          * For now this is fine because we only allow one queued commit against
3725          * a CRTC.  Even if the plane isn't modified by this transaction and we
3726          * don't have a plane lock, we still have the CRTC's lock, so we know
3727          * that no other transactions are racing with us to update it.
3728          */
3729         if (!intel_pstate)
3730                 intel_pstate = to_intel_plane_state(plane->state);
3731
3732         WARN_ON(!intel_pstate->base.fb);
3733
3734         ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][i]);
3735
3736         ret = skl_compute_plane_wm(dev_priv,
3737                                    cstate,
3738                                    intel_pstate,
3739                                    ddb_blocks,
3740                                    level,
3741                                    &result->plane_res_b,
3742                                    &result->plane_res_l,
3743                                    &result->plane_en);
3744         if (ret)
3745                 return ret;
3746
3747         return 0;
3748 }
3749
3750 static uint32_t
3751 skl_compute_linetime_wm(struct intel_crtc_state *cstate)
3752 {
3753         uint32_t pixel_rate;
3754
3755         if (!cstate->base.active)
3756                 return 0;
3757
3758         pixel_rate = ilk_pipe_pixel_rate(cstate);
3759
3760         if (WARN_ON(pixel_rate == 0))
3761                 return 0;
3762
3763         return DIV_ROUND_UP(8 * cstate->base.adjusted_mode.crtc_htotal * 1000,
3764                             pixel_rate);
3765 }
3766
3767 static void skl_compute_transition_wm(struct intel_crtc_state *cstate,
3768                                       struct skl_wm_level *trans_wm /* out */)
3769 {
3770         if (!cstate->base.active)
3771                 return;
3772
3773         /* Until we know more, just disable transition WMs */
3774         trans_wm->plane_en = false;
3775 }
3776
3777 static int skl_build_pipe_wm(struct intel_crtc_state *cstate,
3778                              struct skl_ddb_allocation *ddb,
3779                              struct skl_pipe_wm *pipe_wm)
3780 {
3781         struct drm_device *dev = cstate->base.crtc->dev;
3782         const struct drm_i915_private *dev_priv = to_i915(dev);
3783         struct intel_plane *intel_plane;
3784         struct skl_plane_wm *wm;
3785         int level, max_level = ilk_wm_max_level(dev_priv);
3786         int ret;
3787
3788         /*
3789          * We'll only calculate watermarks for planes that are actually
3790          * enabled, so make sure all other planes are set as disabled.
3791          */
3792         memset(pipe_wm->planes, 0, sizeof(pipe_wm->planes));
3793
3794         for_each_intel_plane_mask(&dev_priv->drm,
3795                                   intel_plane,
3796                                   cstate->base.plane_mask) {
3797                 wm = &pipe_wm->planes[skl_wm_plane_id(intel_plane)];
3798
3799                 for (level = 0; level <= max_level; level++) {
3800                         ret = skl_compute_wm_level(dev_priv, ddb, cstate,
3801                                                    intel_plane, level,
3802                                                    &wm->wm[level]);
3803                         if (ret)
3804                                 return ret;
3805                 }
3806                 skl_compute_transition_wm(cstate, &wm->trans_wm);
3807         }
3808         pipe_wm->linetime = skl_compute_linetime_wm(cstate);
3809
3810         return 0;
3811 }
3812
3813 static void skl_ddb_entry_write(struct drm_i915_private *dev_priv,
3814                                 i915_reg_t reg,
3815                                 const struct skl_ddb_entry *entry)
3816 {
3817         if (entry->end)
3818                 I915_WRITE(reg, (entry->end - 1) << 16 | entry->start);
3819         else
3820                 I915_WRITE(reg, 0);
3821 }
3822
3823 static void skl_write_wm_level(struct drm_i915_private *dev_priv,
3824                                i915_reg_t reg,
3825                                const struct skl_wm_level *level)
3826 {
3827         uint32_t val = 0;
3828
3829         if (level->plane_en) {
3830                 val |= PLANE_WM_EN;
3831                 val |= level->plane_res_b;
3832                 val |= level->plane_res_l << PLANE_WM_LINES_SHIFT;
3833         }
3834
3835         I915_WRITE(reg, val);
3836 }
3837
3838 static void skl_write_plane_wm(struct intel_crtc *intel_crtc,
3839                                const struct skl_plane_wm *wm,
3840                                const struct skl_ddb_allocation *ddb,
3841                                int plane)
3842 {
3843         struct drm_crtc *crtc = &intel_crtc->base;
3844         struct drm_device *dev = crtc->dev;
3845         struct drm_i915_private *dev_priv = to_i915(dev);
3846         int level, max_level = ilk_wm_max_level(dev_priv);
3847         enum pipe pipe = intel_crtc->pipe;
3848
3849         for (level = 0; level <= max_level; level++) {
3850                 skl_write_wm_level(dev_priv, PLANE_WM(pipe, plane, level),
3851                                    &wm->wm[level]);
3852         }
3853         skl_write_wm_level(dev_priv, PLANE_WM_TRANS(pipe, plane),
3854                            &wm->trans_wm);
3855
3856         skl_ddb_entry_write(dev_priv, PLANE_BUF_CFG(pipe, plane),
3857                             &ddb->plane[pipe][plane]);
3858         skl_ddb_entry_write(dev_priv, PLANE_NV12_BUF_CFG(pipe, plane),
3859                             &ddb->y_plane[pipe][plane]);
3860 }
3861
3862 static void skl_write_cursor_wm(struct intel_crtc *intel_crtc,
3863                                 const struct skl_plane_wm *wm,
3864                                 const struct skl_ddb_allocation *ddb)
3865 {
3866         struct drm_crtc *crtc = &intel_crtc->base;
3867         struct drm_device *dev = crtc->dev;
3868         struct drm_i915_private *dev_priv = to_i915(dev);
3869         int level, max_level = ilk_wm_max_level(dev_priv);
3870         enum pipe pipe = intel_crtc->pipe;
3871
3872         for (level = 0; level <= max_level; level++) {
3873                 skl_write_wm_level(dev_priv, CUR_WM(pipe, level),
3874                                    &wm->wm[level]);
3875         }
3876         skl_write_wm_level(dev_priv, CUR_WM_TRANS(pipe), &wm->trans_wm);
3877
3878         skl_ddb_entry_write(dev_priv, CUR_BUF_CFG(pipe),
3879                             &ddb->plane[pipe][PLANE_CURSOR]);
3880 }
3881
3882 bool skl_wm_level_equals(const struct skl_wm_level *l1,
3883                          const struct skl_wm_level *l2)
3884 {
3885         if (l1->plane_en != l2->plane_en)
3886                 return false;
3887
3888         /* If both planes aren't enabled, the rest shouldn't matter */
3889         if (!l1->plane_en)
3890                 return true;
3891
3892         return (l1->plane_res_l == l2->plane_res_l &&
3893                 l1->plane_res_b == l2->plane_res_b);
3894 }
3895
3896 static inline bool skl_ddb_entries_overlap(const struct skl_ddb_entry *a,
3897                                            const struct skl_ddb_entry *b)
3898 {
3899         return a->start < b->end && b->start < a->end;
3900 }
3901
3902 bool skl_ddb_allocation_overlaps(const struct skl_ddb_entry **entries,
3903                                  const struct skl_ddb_entry *ddb,
3904                                  int ignore)
3905 {
3906         int i;
3907
3908         for (i = 0; i < I915_MAX_PIPES; i++)
3909                 if (i != ignore && entries[i] &&
3910                     skl_ddb_entries_overlap(ddb, entries[i]))
3911                         return true;
3912
3913         return false;
3914 }
3915
3916 static int skl_update_pipe_wm(struct drm_crtc_state *cstate,
3917                               const struct skl_pipe_wm *old_pipe_wm,
3918                               struct skl_pipe_wm *pipe_wm, /* out */
3919                               struct skl_ddb_allocation *ddb, /* out */
3920                               bool *changed /* out */)
3921 {
3922         struct intel_crtc_state *intel_cstate = to_intel_crtc_state(cstate);
3923         int ret;
3924
3925         ret = skl_build_pipe_wm(intel_cstate, ddb, pipe_wm);
3926         if (ret)
3927                 return ret;
3928
3929         if (!memcmp(old_pipe_wm, pipe_wm, sizeof(*pipe_wm)))
3930                 *changed = false;
3931         else
3932                 *changed = true;
3933
3934         return 0;
3935 }
3936
3937 static uint32_t
3938 pipes_modified(struct drm_atomic_state *state)
3939 {
3940         struct drm_crtc *crtc;
3941         struct drm_crtc_state *cstate;
3942         uint32_t i, ret = 0;
3943
3944         for_each_crtc_in_state(state, crtc, cstate, i)
3945                 ret |= drm_crtc_mask(crtc);
3946
3947         return ret;
3948 }
3949
3950 static int
3951 skl_ddb_add_affected_planes(struct intel_crtc_state *cstate)
3952 {
3953         struct drm_atomic_state *state = cstate->base.state;
3954         struct drm_device *dev = state->dev;
3955         struct drm_crtc *crtc = cstate->base.crtc;
3956         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3957         struct drm_i915_private *dev_priv = to_i915(dev);
3958         struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3959         struct skl_ddb_allocation *new_ddb = &intel_state->wm_results.ddb;
3960         struct skl_ddb_allocation *cur_ddb = &dev_priv->wm.skl_hw.ddb;
3961         struct drm_plane_state *plane_state;
3962         struct drm_plane *plane;
3963         enum pipe pipe = intel_crtc->pipe;
3964         int id;
3965
3966         WARN_ON(!drm_atomic_get_existing_crtc_state(state, crtc));
3967
3968         drm_for_each_plane_mask(plane, dev, cstate->base.plane_mask) {
3969                 id = skl_wm_plane_id(to_intel_plane(plane));
3970
3971                 if (skl_ddb_entry_equal(&cur_ddb->plane[pipe][id],
3972                                         &new_ddb->plane[pipe][id]) &&
3973                     skl_ddb_entry_equal(&cur_ddb->y_plane[pipe][id],
3974                                         &new_ddb->y_plane[pipe][id]))
3975                         continue;
3976
3977                 plane_state = drm_atomic_get_plane_state(state, plane);
3978                 if (IS_ERR(plane_state))
3979                         return PTR_ERR(plane_state);
3980         }
3981
3982         return 0;
3983 }
3984
3985 static int
3986 skl_compute_ddb(struct drm_atomic_state *state)
3987 {
3988         struct drm_device *dev = state->dev;
3989         struct drm_i915_private *dev_priv = to_i915(dev);
3990         struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
3991         struct intel_crtc *intel_crtc;
3992         struct skl_ddb_allocation *ddb = &intel_state->wm_results.ddb;
3993         uint32_t realloc_pipes = pipes_modified(state);
3994         int ret;
3995
3996         /*
3997          * If this is our first atomic update following hardware readout,
3998          * we can't trust the DDB that the BIOS programmed for us.  Let's
3999          * pretend that all pipes switched active status so that we'll
4000          * ensure a full DDB recompute.
4001          */
4002         if (dev_priv->wm.distrust_bios_wm) {
4003                 ret = drm_modeset_lock(&dev->mode_config.connection_mutex,
4004                                        state->acquire_ctx);
4005                 if (ret)
4006                         return ret;
4007
4008                 intel_state->active_pipe_changes = ~0;
4009
4010                 /*
4011                  * We usually only initialize intel_state->active_crtcs if we
4012                  * we're doing a modeset; make sure this field is always
4013                  * initialized during the sanitization process that happens
4014                  * on the first commit too.
4015                  */
4016                 if (!intel_state->modeset)
4017                         intel_state->active_crtcs = dev_priv->active_crtcs;
4018         }
4019
4020         /*
4021          * If the modeset changes which CRTC's are active, we need to
4022          * recompute the DDB allocation for *all* active pipes, even
4023          * those that weren't otherwise being modified in any way by this
4024          * atomic commit.  Due to the shrinking of the per-pipe allocations
4025          * when new active CRTC's are added, it's possible for a pipe that
4026          * we were already using and aren't changing at all here to suddenly
4027          * become invalid if its DDB needs exceeds its new allocation.
4028          *
4029          * Note that if we wind up doing a full DDB recompute, we can't let
4030          * any other display updates race with this transaction, so we need
4031          * to grab the lock on *all* CRTC's.
4032          */
4033         if (intel_state->active_pipe_changes) {
4034                 realloc_pipes = ~0;
4035                 intel_state->wm_results.dirty_pipes = ~0;
4036         }
4037
4038         /*
4039          * We're not recomputing for the pipes not included in the commit, so
4040          * make sure we start with the current state.
4041          */
4042         memcpy(ddb, &dev_priv->wm.skl_hw.ddb, sizeof(*ddb));
4043
4044         for_each_intel_crtc_mask(dev, intel_crtc, realloc_pipes) {
4045                 struct intel_crtc_state *cstate;
4046
4047                 cstate = intel_atomic_get_crtc_state(state, intel_crtc);
4048                 if (IS_ERR(cstate))
4049                         return PTR_ERR(cstate);
4050
4051                 ret = skl_allocate_pipe_ddb(cstate, ddb);
4052                 if (ret)
4053                         return ret;
4054
4055                 ret = skl_ddb_add_affected_planes(cstate);
4056                 if (ret)
4057                         return ret;
4058         }
4059
4060         return 0;
4061 }
4062
4063 static void
4064 skl_copy_wm_for_pipe(struct skl_wm_values *dst,
4065                      struct skl_wm_values *src,
4066                      enum pipe pipe)
4067 {
4068         memcpy(dst->ddb.y_plane[pipe], src->ddb.y_plane[pipe],
4069                sizeof(dst->ddb.y_plane[pipe]));
4070         memcpy(dst->ddb.plane[pipe], src->ddb.plane[pipe],
4071                sizeof(dst->ddb.plane[pipe]));
4072 }
4073
4074 static void
4075 skl_print_wm_changes(const struct drm_atomic_state *state)
4076 {
4077         const struct drm_device *dev = state->dev;
4078         const struct drm_i915_private *dev_priv = to_i915(dev);
4079         const struct intel_atomic_state *intel_state =
4080                 to_intel_atomic_state(state);
4081         const struct drm_crtc *crtc;
4082         const struct drm_crtc_state *cstate;
4083         const struct intel_plane *intel_plane;
4084         const struct skl_ddb_allocation *old_ddb = &dev_priv->wm.skl_hw.ddb;
4085         const struct skl_ddb_allocation *new_ddb = &intel_state->wm_results.ddb;
4086         int id;
4087         int i;
4088
4089         for_each_crtc_in_state(state, crtc, cstate, i) {
4090                 const struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4091                 enum pipe pipe = intel_crtc->pipe;
4092
4093                 for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
4094                         const struct skl_ddb_entry *old, *new;
4095
4096                         id = skl_wm_plane_id(intel_plane);
4097                         old = &old_ddb->plane[pipe][id];
4098                         new = &new_ddb->plane[pipe][id];
4099
4100                         if (skl_ddb_entry_equal(old, new))
4101                                 continue;
4102
4103                         DRM_DEBUG_ATOMIC("[PLANE:%d:%s] ddb (%d - %d) -> (%d - %d)\n",
4104                                          intel_plane->base.base.id,
4105                                          intel_plane->base.name,
4106                                          old->start, old->end,
4107                                          new->start, new->end);
4108                 }
4109         }
4110 }
4111
4112 static int
4113 skl_compute_wm(struct drm_atomic_state *state)
4114 {
4115         struct drm_crtc *crtc;
4116         struct drm_crtc_state *cstate;
4117         struct intel_atomic_state *intel_state = to_intel_atomic_state(state);
4118         struct skl_wm_values *results = &intel_state->wm_results;
4119         struct skl_pipe_wm *pipe_wm;
4120         bool changed = false;
4121         int ret, i;
4122
4123         /*
4124          * If this transaction isn't actually touching any CRTC's, don't
4125          * bother with watermark calculation.  Note that if we pass this
4126          * test, we're guaranteed to hold at least one CRTC state mutex,
4127          * which means we can safely use values like dev_priv->active_crtcs
4128          * since any racing commits that want to update them would need to
4129          * hold _all_ CRTC state mutexes.
4130          */
4131         for_each_crtc_in_state(state, crtc, cstate, i)
4132                 changed = true;
4133         if (!changed)
4134                 return 0;
4135
4136         /* Clear all dirty flags */
4137         results->dirty_pipes = 0;
4138
4139         ret = skl_compute_ddb(state);
4140         if (ret)
4141                 return ret;
4142
4143         /*
4144          * Calculate WM's for all pipes that are part of this transaction.
4145          * Note that the DDB allocation above may have added more CRTC's that
4146          * weren't otherwise being modified (and set bits in dirty_pipes) if
4147          * pipe allocations had to change.
4148          *
4149          * FIXME:  Now that we're doing this in the atomic check phase, we
4150          * should allow skl_update_pipe_wm() to return failure in cases where
4151          * no suitable watermark values can be found.
4152          */
4153         for_each_crtc_in_state(state, crtc, cstate, i) {
4154                 struct intel_crtc_state *intel_cstate =
4155                         to_intel_crtc_state(cstate);
4156                 const struct skl_pipe_wm *old_pipe_wm =
4157                         &to_intel_crtc_state(crtc->state)->wm.skl.optimal;
4158
4159                 pipe_wm = &intel_cstate->wm.skl.optimal;
4160                 ret = skl_update_pipe_wm(cstate, old_pipe_wm, pipe_wm,
4161                                          &results->ddb, &changed);
4162                 if (ret)
4163                         return ret;
4164
4165                 if (changed)
4166                         results->dirty_pipes |= drm_crtc_mask(crtc);
4167
4168                 if ((results->dirty_pipes & drm_crtc_mask(crtc)) == 0)
4169                         /* This pipe's WM's did not change */
4170                         continue;
4171
4172                 intel_cstate->update_wm_pre = true;
4173         }
4174
4175         skl_print_wm_changes(state);
4176
4177         return 0;
4178 }
4179
4180 static void skl_atomic_update_crtc_wm(struct intel_atomic_state *state,
4181                                       struct intel_crtc_state *cstate)
4182 {
4183         struct intel_crtc *crtc = to_intel_crtc(cstate->base.crtc);
4184         struct drm_i915_private *dev_priv = to_i915(state->base.dev);
4185         struct skl_pipe_wm *pipe_wm = &cstate->wm.skl.optimal;
4186         const struct skl_ddb_allocation *ddb = &state->wm_results.ddb;
4187         enum pipe pipe = crtc->pipe;
4188         int plane;
4189
4190         if (!(state->wm_results.dirty_pipes & drm_crtc_mask(&crtc->base)))
4191                 return;
4192
4193         I915_WRITE(PIPE_WM_LINETIME(pipe), pipe_wm->linetime);
4194
4195         for_each_universal_plane(dev_priv, pipe, plane)
4196                 skl_write_plane_wm(crtc, &pipe_wm->planes[plane], ddb, plane);
4197
4198         skl_write_cursor_wm(crtc, &pipe_wm->planes[PLANE_CURSOR], ddb);
4199 }
4200
4201 static void skl_initial_wm(struct intel_atomic_state *state,
4202                            struct intel_crtc_state *cstate)
4203 {
4204         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4205         struct drm_device *dev = intel_crtc->base.dev;
4206         struct drm_i915_private *dev_priv = to_i915(dev);
4207         struct skl_wm_values *results = &state->wm_results;
4208         struct skl_wm_values *hw_vals = &dev_priv->wm.skl_hw;
4209         enum pipe pipe = intel_crtc->pipe;
4210
4211         if ((results->dirty_pipes & drm_crtc_mask(&intel_crtc->base)) == 0)
4212                 return;
4213
4214         mutex_lock(&dev_priv->wm.wm_mutex);
4215
4216         if (cstate->base.active_changed)
4217                 skl_atomic_update_crtc_wm(state, cstate);
4218
4219         skl_copy_wm_for_pipe(hw_vals, results, pipe);
4220
4221         mutex_unlock(&dev_priv->wm.wm_mutex);
4222 }
4223
4224 static void ilk_compute_wm_config(struct drm_device *dev,
4225                                   struct intel_wm_config *config)
4226 {
4227         struct intel_crtc *crtc;
4228
4229         /* Compute the currently _active_ config */
4230         for_each_intel_crtc(dev, crtc) {
4231                 const struct intel_pipe_wm *wm = &crtc->wm.active.ilk;
4232
4233                 if (!wm->pipe_enabled)
4234                         continue;
4235
4236                 config->sprites_enabled |= wm->sprites_enabled;
4237                 config->sprites_scaled |= wm->sprites_scaled;
4238                 config->num_pipes_active++;
4239         }
4240 }
4241
4242 static void ilk_program_watermarks(struct drm_i915_private *dev_priv)
4243 {
4244         struct drm_device *dev = &dev_priv->drm;
4245         struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
4246         struct ilk_wm_maximums max;
4247         struct intel_wm_config config = {};
4248         struct ilk_wm_values results = {};
4249         enum intel_ddb_partitioning partitioning;
4250
4251         ilk_compute_wm_config(dev, &config);
4252
4253         ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_1_2, &max);
4254         ilk_wm_merge(dev, &config, &max, &lp_wm_1_2);
4255
4256         /* 5/6 split only in single pipe config on IVB+ */
4257         if (INTEL_GEN(dev_priv) >= 7 &&
4258             config.num_pipes_active == 1 && config.sprites_enabled) {
4259                 ilk_compute_wm_maximums(dev, 1, &config, INTEL_DDB_PART_5_6, &max);
4260                 ilk_wm_merge(dev, &config, &max, &lp_wm_5_6);
4261
4262                 best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
4263         } else {
4264                 best_lp_wm = &lp_wm_1_2;
4265         }
4266
4267         partitioning = (best_lp_wm == &lp_wm_1_2) ?
4268                        INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
4269
4270         ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
4271
4272         ilk_write_wm_values(dev_priv, &results);
4273 }
4274
4275 static void ilk_initial_watermarks(struct intel_atomic_state *state,
4276                                    struct intel_crtc_state *cstate)
4277 {
4278         struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
4279         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4280
4281         mutex_lock(&dev_priv->wm.wm_mutex);
4282         intel_crtc->wm.active.ilk = cstate->wm.ilk.intermediate;
4283         ilk_program_watermarks(dev_priv);
4284         mutex_unlock(&dev_priv->wm.wm_mutex);
4285 }
4286
4287 static void ilk_optimize_watermarks(struct intel_atomic_state *state,
4288                                     struct intel_crtc_state *cstate)
4289 {
4290         struct drm_i915_private *dev_priv = to_i915(cstate->base.crtc->dev);
4291         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
4292
4293         mutex_lock(&dev_priv->wm.wm_mutex);
4294         if (cstate->wm.need_postvbl_update) {
4295                 intel_crtc->wm.active.ilk = cstate->wm.ilk.optimal;
4296                 ilk_program_watermarks(dev_priv);
4297         }
4298         mutex_unlock(&dev_priv->wm.wm_mutex);
4299 }
4300
4301 static inline void skl_wm_level_from_reg_val(uint32_t val,
4302                                              struct skl_wm_level *level)
4303 {
4304         level->plane_en = val & PLANE_WM_EN;
4305         level->plane_res_b = val & PLANE_WM_BLOCKS_MASK;
4306         level->plane_res_l = (val >> PLANE_WM_LINES_SHIFT) &
4307                 PLANE_WM_LINES_MASK;
4308 }
4309
4310 void skl_pipe_wm_get_hw_state(struct drm_crtc *crtc,
4311                               struct skl_pipe_wm *out)
4312 {
4313         struct drm_device *dev = crtc->dev;
4314         struct drm_i915_private *dev_priv = to_i915(dev);
4315         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4316         struct intel_plane *intel_plane;
4317         struct skl_plane_wm *wm;
4318         enum pipe pipe = intel_crtc->pipe;
4319         int level, id, max_level;
4320         uint32_t val;
4321
4322         max_level = ilk_wm_max_level(dev_priv);
4323
4324         for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
4325                 id = skl_wm_plane_id(intel_plane);
4326                 wm = &out->planes[id];
4327
4328                 for (level = 0; level <= max_level; level++) {
4329                         if (id != PLANE_CURSOR)
4330                                 val = I915_READ(PLANE_WM(pipe, id, level));
4331                         else
4332                                 val = I915_READ(CUR_WM(pipe, level));
4333
4334                         skl_wm_level_from_reg_val(val, &wm->wm[level]);
4335                 }
4336
4337                 if (id != PLANE_CURSOR)
4338                         val = I915_READ(PLANE_WM_TRANS(pipe, id));
4339                 else
4340                         val = I915_READ(CUR_WM_TRANS(pipe));
4341
4342                 skl_wm_level_from_reg_val(val, &wm->trans_wm);
4343         }
4344
4345         if (!intel_crtc->active)
4346                 return;
4347
4348         out->linetime = I915_READ(PIPE_WM_LINETIME(pipe));
4349 }
4350
4351 void skl_wm_get_hw_state(struct drm_device *dev)
4352 {
4353         struct drm_i915_private *dev_priv = to_i915(dev);
4354         struct skl_wm_values *hw = &dev_priv->wm.skl_hw;
4355         struct skl_ddb_allocation *ddb = &dev_priv->wm.skl_hw.ddb;
4356         struct drm_crtc *crtc;
4357         struct intel_crtc *intel_crtc;
4358         struct intel_crtc_state *cstate;
4359
4360         skl_ddb_get_hw_state(dev_priv, ddb);
4361         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head) {
4362                 intel_crtc = to_intel_crtc(crtc);
4363                 cstate = to_intel_crtc_state(crtc->state);
4364
4365                 skl_pipe_wm_get_hw_state(crtc, &cstate->wm.skl.optimal);
4366
4367                 if (intel_crtc->active)
4368                         hw->dirty_pipes |= drm_crtc_mask(crtc);
4369         }
4370
4371         if (dev_priv->active_crtcs) {
4372                 /* Fully recompute DDB on first atomic commit */
4373                 dev_priv->wm.distrust_bios_wm = true;
4374         } else {
4375                 /* Easy/common case; just sanitize DDB now if everything off */
4376                 memset(ddb, 0, sizeof(*ddb));
4377         }
4378 }
4379
4380 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
4381 {
4382         struct drm_device *dev = crtc->dev;
4383         struct drm_i915_private *dev_priv = to_i915(dev);
4384         struct ilk_wm_values *hw = &dev_priv->wm.hw;
4385         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
4386         struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
4387         struct intel_pipe_wm *active = &cstate->wm.ilk.optimal;
4388         enum pipe pipe = intel_crtc->pipe;
4389         static const i915_reg_t wm0_pipe_reg[] = {
4390                 [PIPE_A] = WM0_PIPEA_ILK,
4391                 [PIPE_B] = WM0_PIPEB_ILK,
4392                 [PIPE_C] = WM0_PIPEC_IVB,
4393         };
4394
4395         hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
4396         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
4397                 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
4398
4399         memset(active, 0, sizeof(*active));
4400
4401         active->pipe_enabled = intel_crtc->active;
4402
4403         if (active->pipe_enabled) {
4404                 u32 tmp = hw->wm_pipe[pipe];
4405
4406                 /*
4407                  * For active pipes LP0 watermark is marked as
4408                  * enabled, and LP1+ watermaks as disabled since
4409                  * we can't really reverse compute them in case
4410                  * multiple pipes are active.
4411                  */
4412                 active->wm[0].enable = true;
4413                 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
4414                 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
4415                 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
4416                 active->linetime = hw->wm_linetime[pipe];
4417         } else {
4418                 int level, max_level = ilk_wm_max_level(dev_priv);
4419
4420                 /*
4421                  * For inactive pipes, all watermark levels
4422                  * should be marked as enabled but zeroed,
4423                  * which is what we'd compute them to.
4424                  */
4425                 for (level = 0; level <= max_level; level++)
4426                         active->wm[level].enable = true;
4427         }
4428
4429         intel_crtc->wm.active.ilk = *active;
4430 }
4431
4432 #define _FW_WM(value, plane) \
4433         (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
4434 #define _FW_WM_VLV(value, plane) \
4435         (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
4436
4437 static void vlv_read_wm_values(struct drm_i915_private *dev_priv,
4438                                struct vlv_wm_values *wm)
4439 {
4440         enum pipe pipe;
4441         uint32_t tmp;
4442
4443         for_each_pipe(dev_priv, pipe) {
4444                 tmp = I915_READ(VLV_DDL(pipe));
4445
4446                 wm->ddl[pipe].primary =
4447                         (tmp >> DDL_PLANE_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4448                 wm->ddl[pipe].cursor =
4449                         (tmp >> DDL_CURSOR_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4450                 wm->ddl[pipe].sprite[0] =
4451                         (tmp >> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4452                 wm->ddl[pipe].sprite[1] =
4453                         (tmp >> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
4454         }
4455
4456         tmp = I915_READ(DSPFW1);
4457         wm->sr.plane = _FW_WM(tmp, SR);
4458         wm->pipe[PIPE_B].cursor = _FW_WM(tmp, CURSORB);
4459         wm->pipe[PIPE_B].primary = _FW_WM_VLV(tmp, PLANEB);
4460         wm->pipe[PIPE_A].primary = _FW_WM_VLV(tmp, PLANEA);
4461
4462         tmp = I915_READ(DSPFW2);
4463         wm->pipe[PIPE_A].sprite[1] = _FW_WM_VLV(tmp, SPRITEB);
4464         wm->pipe[PIPE_A].cursor = _FW_WM(tmp, CURSORA);
4465         wm->pipe[PIPE_A].sprite[0] = _FW_WM_VLV(tmp, SPRITEA);
4466
4467         tmp = I915_READ(DSPFW3);
4468         wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
4469
4470         if (IS_CHERRYVIEW(dev_priv)) {
4471                 tmp = I915_READ(DSPFW7_CHV);
4472                 wm->pipe[PIPE_B].sprite[1] = _FW_WM_VLV(tmp, SPRITED);
4473                 wm->pipe[PIPE_B].sprite[0] = _FW_WM_VLV(tmp, SPRITEC);
4474
4475                 tmp = I915_READ(DSPFW8_CHV);
4476                 wm->pipe[PIPE_C].sprite[1] = _FW_WM_VLV(tmp, SPRITEF);
4477                 wm->pipe[PIPE_C].sprite[0] = _FW_WM_VLV(tmp, SPRITEE);
4478
4479                 tmp = I915_READ(DSPFW9_CHV);
4480                 wm->pipe[PIPE_C].primary = _FW_WM_VLV(tmp, PLANEC);
4481                 wm->pipe[PIPE_C].cursor = _FW_WM(tmp, CURSORC);
4482
4483                 tmp = I915_READ(DSPHOWM);
4484                 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
4485                 wm->pipe[PIPE_C].sprite[1] |= _FW_WM(tmp, SPRITEF_HI) << 8;
4486                 wm->pipe[PIPE_C].sprite[0] |= _FW_WM(tmp, SPRITEE_HI) << 8;
4487                 wm->pipe[PIPE_C].primary |= _FW_WM(tmp, PLANEC_HI) << 8;
4488                 wm->pipe[PIPE_B].sprite[1] |= _FW_WM(tmp, SPRITED_HI) << 8;
4489                 wm->pipe[PIPE_B].sprite[0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
4490                 wm->pipe[PIPE_B].primary |= _FW_WM(tmp, PLANEB_HI) << 8;
4491                 wm->pipe[PIPE_A].sprite[1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
4492                 wm->pipe[PIPE_A].sprite[0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
4493                 wm->pipe[PIPE_A].primary |= _FW_WM(tmp, PLANEA_HI) << 8;
4494         } else {
4495                 tmp = I915_READ(DSPFW7);
4496                 wm->pipe[PIPE_B].sprite[1] = _FW_WM_VLV(tmp, SPRITED);
4497                 wm->pipe[PIPE_B].sprite[0] = _FW_WM_VLV(tmp, SPRITEC);
4498
4499                 tmp = I915_READ(DSPHOWM);
4500                 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
4501                 wm->pipe[PIPE_B].sprite[1] |= _FW_WM(tmp, SPRITED_HI) << 8;
4502                 wm->pipe[PIPE_B].sprite[0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
4503                 wm->pipe[PIPE_B].primary |= _FW_WM(tmp, PLANEB_HI) << 8;
4504                 wm->pipe[PIPE_A].sprite[1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
4505                 wm->pipe[PIPE_A].sprite[0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
4506                 wm->pipe[PIPE_A].primary |= _FW_WM(tmp, PLANEA_HI) << 8;
4507         }
4508 }
4509
4510 #undef _FW_WM
4511 #undef _FW_WM_VLV
4512
4513 void vlv_wm_get_hw_state(struct drm_device *dev)
4514 {
4515         struct drm_i915_private *dev_priv = to_i915(dev);
4516         struct vlv_wm_values *wm = &dev_priv->wm.vlv;
4517         struct intel_plane *plane;
4518         enum pipe pipe;
4519         u32 val;
4520
4521         vlv_read_wm_values(dev_priv, wm);
4522
4523         for_each_intel_plane(dev, plane) {
4524                 switch (plane->base.type) {
4525                         int sprite;
4526                 case DRM_PLANE_TYPE_CURSOR:
4527                         plane->wm.fifo_size = 63;
4528                         break;
4529                 case DRM_PLANE_TYPE_PRIMARY:
4530                         plane->wm.fifo_size = vlv_get_fifo_size(dev_priv, plane->pipe, 0);
4531                         break;
4532                 case DRM_PLANE_TYPE_OVERLAY:
4533                         sprite = plane->plane;
4534                         plane->wm.fifo_size = vlv_get_fifo_size(dev_priv, plane->pipe, sprite + 1);
4535                         break;
4536                 }
4537         }
4538
4539         wm->cxsr = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
4540         wm->level = VLV_WM_LEVEL_PM2;
4541
4542         if (IS_CHERRYVIEW(dev_priv)) {
4543                 mutex_lock(&dev_priv->rps.hw_lock);
4544
4545                 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
4546                 if (val & DSP_MAXFIFO_PM5_ENABLE)
4547                         wm->level = VLV_WM_LEVEL_PM5;
4548
4549                 /*
4550                  * If DDR DVFS is disabled in the BIOS, Punit
4551                  * will never ack the request. So if that happens
4552                  * assume we don't have to enable/disable DDR DVFS
4553                  * dynamically. To test that just set the REQ_ACK
4554                  * bit to poke the Punit, but don't change the
4555                  * HIGH/LOW bits so that we don't actually change
4556                  * the current state.
4557                  */
4558                 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
4559                 val |= FORCE_DDR_FREQ_REQ_ACK;
4560                 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
4561
4562                 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
4563                               FORCE_DDR_FREQ_REQ_ACK) == 0, 3)) {
4564                         DRM_DEBUG_KMS("Punit not acking DDR DVFS request, "
4565                                       "assuming DDR DVFS is disabled\n");
4566                         dev_priv->wm.max_level = VLV_WM_LEVEL_PM5;
4567                 } else {
4568                         val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
4569                         if ((val & FORCE_DDR_HIGH_FREQ) == 0)
4570                                 wm->level = VLV_WM_LEVEL_DDR_DVFS;
4571                 }
4572
4573                 mutex_unlock(&dev_priv->rps.hw_lock);
4574         }
4575
4576         for_each_pipe(dev_priv, pipe)
4577                 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
4578                               pipe_name(pipe), wm->pipe[pipe].primary, wm->pipe[pipe].cursor,
4579                               wm->pipe[pipe].sprite[0], wm->pipe[pipe].sprite[1]);
4580
4581         DRM_DEBUG_KMS("Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
4582                       wm->sr.plane, wm->sr.cursor, wm->level, wm->cxsr);
4583 }
4584
4585 void ilk_wm_get_hw_state(struct drm_device *dev)
4586 {
4587         struct drm_i915_private *dev_priv = to_i915(dev);
4588         struct ilk_wm_values *hw = &dev_priv->wm.hw;
4589         struct drm_crtc *crtc;
4590
4591         for_each_crtc(dev, crtc)
4592                 ilk_pipe_wm_get_hw_state(crtc);
4593
4594         hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
4595         hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
4596         hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
4597
4598         hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
4599         if (INTEL_GEN(dev_priv) >= 7) {
4600                 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
4601                 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
4602         }
4603
4604         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
4605                 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
4606                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4607         else if (IS_IVYBRIDGE(dev_priv))
4608                 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
4609                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
4610
4611         hw->enable_fbc_wm =
4612                 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
4613 }
4614
4615 /**
4616  * intel_update_watermarks - update FIFO watermark values based on current modes
4617  *
4618  * Calculate watermark values for the various WM regs based on current mode
4619  * and plane configuration.
4620  *
4621  * There are several cases to deal with here:
4622  *   - normal (i.e. non-self-refresh)
4623  *   - self-refresh (SR) mode
4624  *   - lines are large relative to FIFO size (buffer can hold up to 2)
4625  *   - lines are small relative to FIFO size (buffer can hold more than 2
4626  *     lines), so need to account for TLB latency
4627  *
4628  *   The normal calculation is:
4629  *     watermark = dotclock * bytes per pixel * latency
4630  *   where latency is platform & configuration dependent (we assume pessimal
4631  *   values here).
4632  *
4633  *   The SR calculation is:
4634  *     watermark = (trunc(latency/line time)+1) * surface width *
4635  *       bytes per pixel
4636  *   where
4637  *     line time = htotal / dotclock
4638  *     surface width = hdisplay for normal plane and 64 for cursor
4639  *   and latency is assumed to be high, as above.
4640  *
4641  * The final value programmed to the register should always be rounded up,
4642  * and include an extra 2 entries to account for clock crossings.
4643  *
4644  * We don't use the sprite, so we can ignore that.  And on Crestline we have
4645  * to set the non-SR watermarks to 8.
4646  */
4647 void intel_update_watermarks(struct intel_crtc *crtc)
4648 {
4649         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
4650
4651         if (dev_priv->display.update_wm)
4652                 dev_priv->display.update_wm(crtc);
4653 }
4654
4655 /*
4656  * Lock protecting IPS related data structures
4657  */
4658 DEFINE_SPINLOCK(mchdev_lock);
4659
4660 /* Global for IPS driver to get at the current i915 device. Protected by
4661  * mchdev_lock. */
4662 static struct drm_i915_private *i915_mch_dev;
4663
4664 bool ironlake_set_drps(struct drm_i915_private *dev_priv, u8 val)
4665 {
4666         u16 rgvswctl;
4667
4668         assert_spin_locked(&mchdev_lock);
4669
4670         rgvswctl = I915_READ16(MEMSWCTL);
4671         if (rgvswctl & MEMCTL_CMD_STS) {
4672                 DRM_DEBUG("gpu busy, RCS change rejected\n");
4673                 return false; /* still busy with another command */
4674         }
4675
4676         rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
4677                 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
4678         I915_WRITE16(MEMSWCTL, rgvswctl);
4679         POSTING_READ16(MEMSWCTL);
4680
4681         rgvswctl |= MEMCTL_CMD_STS;
4682         I915_WRITE16(MEMSWCTL, rgvswctl);
4683
4684         return true;
4685 }
4686
4687 static void ironlake_enable_drps(struct drm_i915_private *dev_priv)
4688 {
4689         u32 rgvmodectl;
4690         u8 fmax, fmin, fstart, vstart;
4691
4692         spin_lock_irq(&mchdev_lock);
4693
4694         rgvmodectl = I915_READ(MEMMODECTL);
4695
4696         /* Enable temp reporting */
4697         I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
4698         I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
4699
4700         /* 100ms RC evaluation intervals */
4701         I915_WRITE(RCUPEI, 100000);
4702         I915_WRITE(RCDNEI, 100000);
4703
4704         /* Set max/min thresholds to 90ms and 80ms respectively */
4705         I915_WRITE(RCBMAXAVG, 90000);
4706         I915_WRITE(RCBMINAVG, 80000);
4707
4708         I915_WRITE(MEMIHYST, 1);
4709
4710         /* Set up min, max, and cur for interrupt handling */
4711         fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
4712         fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
4713         fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
4714                 MEMMODE_FSTART_SHIFT;
4715
4716         vstart = (I915_READ(PXVFREQ(fstart)) & PXVFREQ_PX_MASK) >>
4717                 PXVFREQ_PX_SHIFT;
4718
4719         dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
4720         dev_priv->ips.fstart = fstart;
4721
4722         dev_priv->ips.max_delay = fstart;
4723         dev_priv->ips.min_delay = fmin;
4724         dev_priv->ips.cur_delay = fstart;
4725
4726         DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
4727                          fmax, fmin, fstart);
4728
4729         I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
4730
4731         /*
4732          * Interrupts will be enabled in ironlake_irq_postinstall
4733          */
4734
4735         I915_WRITE(VIDSTART, vstart);
4736         POSTING_READ(VIDSTART);
4737
4738         rgvmodectl |= MEMMODE_SWMODE_EN;
4739         I915_WRITE(MEMMODECTL, rgvmodectl);
4740
4741         if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
4742                 DRM_ERROR("stuck trying to change perf mode\n");
4743         mdelay(1);
4744
4745         ironlake_set_drps(dev_priv, fstart);
4746
4747         dev_priv->ips.last_count1 = I915_READ(DMIEC) +
4748                 I915_READ(DDREC) + I915_READ(CSIEC);
4749         dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
4750         dev_priv->ips.last_count2 = I915_READ(GFXEC);
4751         dev_priv->ips.last_time2 = ktime_get_raw_ns();
4752
4753         spin_unlock_irq(&mchdev_lock);
4754 }
4755
4756 static void ironlake_disable_drps(struct drm_i915_private *dev_priv)
4757 {
4758         u16 rgvswctl;
4759
4760         spin_lock_irq(&mchdev_lock);
4761
4762         rgvswctl = I915_READ16(MEMSWCTL);
4763
4764         /* Ack interrupts, disable EFC interrupt */
4765         I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
4766         I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
4767         I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
4768         I915_WRITE(DEIIR, DE_PCU_EVENT);
4769         I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
4770
4771         /* Go back to the starting frequency */
4772         ironlake_set_drps(dev_priv, dev_priv->ips.fstart);
4773         mdelay(1);
4774         rgvswctl |= MEMCTL_CMD_STS;
4775         I915_WRITE(MEMSWCTL, rgvswctl);
4776         mdelay(1);
4777
4778         spin_unlock_irq(&mchdev_lock);
4779 }
4780
4781 /* There's a funny hw issue where the hw returns all 0 when reading from
4782  * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
4783  * ourselves, instead of doing a rmw cycle (which might result in us clearing
4784  * all limits and the gpu stuck at whatever frequency it is at atm).
4785  */
4786 static u32 intel_rps_limits(struct drm_i915_private *dev_priv, u8 val)
4787 {
4788         u32 limits;
4789
4790         /* Only set the down limit when we've reached the lowest level to avoid
4791          * getting more interrupts, otherwise leave this clear. This prevents a
4792          * race in the hw when coming out of rc6: There's a tiny window where
4793          * the hw runs at the minimal clock before selecting the desired
4794          * frequency, if the down threshold expires in that window we will not
4795          * receive a down interrupt. */
4796         if (IS_GEN9(dev_priv)) {
4797                 limits = (dev_priv->rps.max_freq_softlimit) << 23;
4798                 if (val <= dev_priv->rps.min_freq_softlimit)
4799                         limits |= (dev_priv->rps.min_freq_softlimit) << 14;
4800         } else {
4801                 limits = dev_priv->rps.max_freq_softlimit << 24;
4802                 if (val <= dev_priv->rps.min_freq_softlimit)
4803                         limits |= dev_priv->rps.min_freq_softlimit << 16;
4804         }
4805
4806         return limits;
4807 }
4808
4809 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
4810 {
4811         int new_power;
4812         u32 threshold_up = 0, threshold_down = 0; /* in % */
4813         u32 ei_up = 0, ei_down = 0;
4814
4815         new_power = dev_priv->rps.power;
4816         switch (dev_priv->rps.power) {
4817         case LOW_POWER:
4818                 if (val > dev_priv->rps.efficient_freq + 1 &&
4819                     val > dev_priv->rps.cur_freq)
4820                         new_power = BETWEEN;
4821                 break;
4822
4823         case BETWEEN:
4824                 if (val <= dev_priv->rps.efficient_freq &&
4825                     val < dev_priv->rps.cur_freq)
4826                         new_power = LOW_POWER;
4827                 else if (val >= dev_priv->rps.rp0_freq &&
4828                          val > dev_priv->rps.cur_freq)
4829                         new_power = HIGH_POWER;
4830                 break;
4831
4832         case HIGH_POWER:
4833                 if (val < (dev_priv->rps.rp1_freq + dev_priv->rps.rp0_freq) >> 1 &&
4834                     val < dev_priv->rps.cur_freq)
4835                         new_power = BETWEEN;
4836                 break;
4837         }
4838         /* Max/min bins are special */
4839         if (val <= dev_priv->rps.min_freq_softlimit)
4840                 new_power = LOW_POWER;
4841         if (val >= dev_priv->rps.max_freq_softlimit)
4842                 new_power = HIGH_POWER;
4843         if (new_power == dev_priv->rps.power)
4844                 return;
4845
4846         /* Note the units here are not exactly 1us, but 1280ns. */
4847         switch (new_power) {
4848         case LOW_POWER:
4849                 /* Upclock if more than 95% busy over 16ms */
4850                 ei_up = 16000;
4851                 threshold_up = 95;
4852
4853                 /* Downclock if less than 85% busy over 32ms */
4854                 ei_down = 32000;
4855                 threshold_down = 85;
4856                 break;
4857
4858         case BETWEEN:
4859                 /* Upclock if more than 90% busy over 13ms */
4860                 ei_up = 13000;
4861                 threshold_up = 90;
4862
4863                 /* Downclock if less than 75% busy over 32ms */
4864                 ei_down = 32000;
4865                 threshold_down = 75;
4866                 break;
4867
4868         case HIGH_POWER:
4869                 /* Upclock if more than 85% busy over 10ms */
4870                 ei_up = 10000;
4871                 threshold_up = 85;
4872
4873                 /* Downclock if less than 60% busy over 32ms */
4874                 ei_down = 32000;
4875                 threshold_down = 60;
4876                 break;
4877         }
4878
4879         I915_WRITE(GEN6_RP_UP_EI,
4880                    GT_INTERVAL_FROM_US(dev_priv, ei_up));
4881         I915_WRITE(GEN6_RP_UP_THRESHOLD,
4882                    GT_INTERVAL_FROM_US(dev_priv,
4883                                        ei_up * threshold_up / 100));
4884
4885         I915_WRITE(GEN6_RP_DOWN_EI,
4886                    GT_INTERVAL_FROM_US(dev_priv, ei_down));
4887         I915_WRITE(GEN6_RP_DOWN_THRESHOLD,
4888                    GT_INTERVAL_FROM_US(dev_priv,
4889                                        ei_down * threshold_down / 100));
4890
4891         I915_WRITE(GEN6_RP_CONTROL,
4892                    GEN6_RP_MEDIA_TURBO |
4893                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
4894                    GEN6_RP_MEDIA_IS_GFX |
4895                    GEN6_RP_ENABLE |
4896                    GEN6_RP_UP_BUSY_AVG |
4897                    GEN6_RP_DOWN_IDLE_AVG);
4898
4899         dev_priv->rps.power = new_power;
4900         dev_priv->rps.up_threshold = threshold_up;
4901         dev_priv->rps.down_threshold = threshold_down;
4902         dev_priv->rps.last_adj = 0;
4903 }
4904
4905 static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
4906 {
4907         u32 mask = 0;
4908
4909         if (val > dev_priv->rps.min_freq_softlimit)
4910                 mask |= GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
4911         if (val < dev_priv->rps.max_freq_softlimit)
4912                 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD;
4913
4914         mask &= dev_priv->pm_rps_events;
4915
4916         return gen6_sanitize_rps_pm_mask(dev_priv, ~mask);
4917 }
4918
4919 /* gen6_set_rps is called to update the frequency request, but should also be
4920  * called when the range (min_delay and max_delay) is modified so that we can
4921  * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
4922 static void gen6_set_rps(struct drm_i915_private *dev_priv, u8 val)
4923 {
4924         /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
4925         if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1))
4926                 return;
4927
4928         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4929         WARN_ON(val > dev_priv->rps.max_freq);
4930         WARN_ON(val < dev_priv->rps.min_freq);
4931
4932         /* min/max delay may still have been modified so be sure to
4933          * write the limits value.
4934          */
4935         if (val != dev_priv->rps.cur_freq) {
4936                 gen6_set_rps_thresholds(dev_priv, val);
4937
4938                 if (IS_GEN9(dev_priv))
4939                         I915_WRITE(GEN6_RPNSWREQ,
4940                                    GEN9_FREQUENCY(val));
4941                 else if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
4942                         I915_WRITE(GEN6_RPNSWREQ,
4943                                    HSW_FREQUENCY(val));
4944                 else
4945                         I915_WRITE(GEN6_RPNSWREQ,
4946                                    GEN6_FREQUENCY(val) |
4947                                    GEN6_OFFSET(0) |
4948                                    GEN6_AGGRESSIVE_TURBO);
4949         }
4950
4951         /* Make sure we continue to get interrupts
4952          * until we hit the minimum or maximum frequencies.
4953          */
4954         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, intel_rps_limits(dev_priv, val));
4955         I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
4956
4957         POSTING_READ(GEN6_RPNSWREQ);
4958
4959         dev_priv->rps.cur_freq = val;
4960         trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
4961 }
4962
4963 static void valleyview_set_rps(struct drm_i915_private *dev_priv, u8 val)
4964 {
4965         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4966         WARN_ON(val > dev_priv->rps.max_freq);
4967         WARN_ON(val < dev_priv->rps.min_freq);
4968
4969         if (WARN_ONCE(IS_CHERRYVIEW(dev_priv) && (val & 1),
4970                       "Odd GPU freq value\n"))
4971                 val &= ~1;
4972
4973         I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
4974
4975         if (val != dev_priv->rps.cur_freq) {
4976                 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
4977                 if (!IS_CHERRYVIEW(dev_priv))
4978                         gen6_set_rps_thresholds(dev_priv, val);
4979         }
4980
4981         dev_priv->rps.cur_freq = val;
4982         trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
4983 }
4984
4985 /* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
4986  *
4987  * * If Gfx is Idle, then
4988  * 1. Forcewake Media well.
4989  * 2. Request idle freq.
4990  * 3. Release Forcewake of Media well.
4991 */
4992 static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
4993 {
4994         u32 val = dev_priv->rps.idle_freq;
4995
4996         if (dev_priv->rps.cur_freq <= val)
4997                 return;
4998
4999         /* Wake up the media well, as that takes a lot less
5000          * power than the Render well. */
5001         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_MEDIA);
5002         valleyview_set_rps(dev_priv, val);
5003         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_MEDIA);
5004 }
5005
5006 void gen6_rps_busy(struct drm_i915_private *dev_priv)
5007 {
5008         mutex_lock(&dev_priv->rps.hw_lock);
5009         if (dev_priv->rps.enabled) {
5010                 if (dev_priv->pm_rps_events & (GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_UP_EI_EXPIRED))
5011                         gen6_rps_reset_ei(dev_priv);
5012                 I915_WRITE(GEN6_PMINTRMSK,
5013                            gen6_rps_pm_mask(dev_priv, dev_priv->rps.cur_freq));
5014
5015                 gen6_enable_rps_interrupts(dev_priv);
5016
5017                 /* Ensure we start at the user's desired frequency */
5018                 intel_set_rps(dev_priv,
5019                               clamp(dev_priv->rps.cur_freq,
5020                                     dev_priv->rps.min_freq_softlimit,
5021                                     dev_priv->rps.max_freq_softlimit));
5022         }
5023         mutex_unlock(&dev_priv->rps.hw_lock);
5024 }
5025
5026 void gen6_rps_idle(struct drm_i915_private *dev_priv)
5027 {
5028         /* Flush our bottom-half so that it does not race with us
5029          * setting the idle frequency and so that it is bounded by
5030          * our rpm wakeref. And then disable the interrupts to stop any
5031          * futher RPS reclocking whilst we are asleep.
5032          */
5033         gen6_disable_rps_interrupts(dev_priv);
5034
5035         mutex_lock(&dev_priv->rps.hw_lock);
5036         if (dev_priv->rps.enabled) {
5037                 if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
5038                         vlv_set_rps_idle(dev_priv);
5039                 else
5040                         gen6_set_rps(dev_priv, dev_priv->rps.idle_freq);
5041                 dev_priv->rps.last_adj = 0;
5042                 I915_WRITE(GEN6_PMINTRMSK,
5043                            gen6_sanitize_rps_pm_mask(dev_priv, ~0));
5044         }
5045         mutex_unlock(&dev_priv->rps.hw_lock);
5046
5047         spin_lock(&dev_priv->rps.client_lock);
5048         while (!list_empty(&dev_priv->rps.clients))
5049                 list_del_init(dev_priv->rps.clients.next);
5050         spin_unlock(&dev_priv->rps.client_lock);
5051 }
5052
5053 void gen6_rps_boost(struct drm_i915_private *dev_priv,
5054                     struct intel_rps_client *rps,
5055                     unsigned long submitted)
5056 {
5057         /* This is intentionally racy! We peek at the state here, then
5058          * validate inside the RPS worker.
5059          */
5060         if (!(dev_priv->gt.awake &&
5061               dev_priv->rps.enabled &&
5062               dev_priv->rps.cur_freq < dev_priv->rps.boost_freq))
5063                 return;
5064
5065         /* Force a RPS boost (and don't count it against the client) if
5066          * the GPU is severely congested.
5067          */
5068         if (rps && time_after(jiffies, submitted + DRM_I915_THROTTLE_JIFFIES))
5069                 rps = NULL;
5070
5071         spin_lock(&dev_priv->rps.client_lock);
5072         if (rps == NULL || list_empty(&rps->link)) {
5073                 spin_lock_irq(&dev_priv->irq_lock);
5074                 if (dev_priv->rps.interrupts_enabled) {
5075                         dev_priv->rps.client_boost = true;
5076                         schedule_work(&dev_priv->rps.work);
5077                 }
5078                 spin_unlock_irq(&dev_priv->irq_lock);
5079
5080                 if (rps != NULL) {
5081                         list_add(&rps->link, &dev_priv->rps.clients);
5082                         rps->boosts++;
5083                 } else
5084                         dev_priv->rps.boosts++;
5085         }
5086         spin_unlock(&dev_priv->rps.client_lock);
5087 }
5088
5089 void intel_set_rps(struct drm_i915_private *dev_priv, u8 val)
5090 {
5091         if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv))
5092                 valleyview_set_rps(dev_priv, val);
5093         else
5094                 gen6_set_rps(dev_priv, val);
5095 }
5096
5097 static void gen9_disable_rc6(struct drm_i915_private *dev_priv)
5098 {
5099         I915_WRITE(GEN6_RC_CONTROL, 0);
5100         I915_WRITE(GEN9_PG_ENABLE, 0);
5101 }
5102
5103 static void gen9_disable_rps(struct drm_i915_private *dev_priv)
5104 {
5105         I915_WRITE(GEN6_RP_CONTROL, 0);
5106 }
5107
5108 static void gen6_disable_rps(struct drm_i915_private *dev_priv)
5109 {
5110         I915_WRITE(GEN6_RC_CONTROL, 0);
5111         I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
5112         I915_WRITE(GEN6_RP_CONTROL, 0);
5113 }
5114
5115 static void cherryview_disable_rps(struct drm_i915_private *dev_priv)
5116 {
5117         I915_WRITE(GEN6_RC_CONTROL, 0);
5118 }
5119
5120 static void valleyview_disable_rps(struct drm_i915_private *dev_priv)
5121 {
5122         /* we're doing forcewake before Disabling RC6,
5123          * This what the BIOS expects when going into suspend */
5124         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5125
5126         I915_WRITE(GEN6_RC_CONTROL, 0);
5127
5128         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5129 }
5130
5131 static void intel_print_rc6_info(struct drm_i915_private *dev_priv, u32 mode)
5132 {
5133         if (IS_VALLEYVIEW(dev_priv) || IS_CHERRYVIEW(dev_priv)) {
5134                 if (mode & (GEN7_RC_CTL_TO_MODE | GEN6_RC_CTL_EI_MODE(1)))
5135                         mode = GEN6_RC_CTL_RC6_ENABLE;
5136                 else
5137                         mode = 0;
5138         }
5139         if (HAS_RC6p(dev_priv))
5140                 DRM_DEBUG_DRIVER("Enabling RC6 states: "
5141                                  "RC6 %s RC6p %s RC6pp %s\n",
5142                                  onoff(mode & GEN6_RC_CTL_RC6_ENABLE),
5143                                  onoff(mode & GEN6_RC_CTL_RC6p_ENABLE),
5144                                  onoff(mode & GEN6_RC_CTL_RC6pp_ENABLE));
5145
5146         else
5147                 DRM_DEBUG_DRIVER("Enabling RC6 states: RC6 %s\n",
5148                                  onoff(mode & GEN6_RC_CTL_RC6_ENABLE));
5149 }
5150
5151 static bool bxt_check_bios_rc6_setup(struct drm_i915_private *dev_priv)
5152 {
5153         struct i915_ggtt *ggtt = &dev_priv->ggtt;
5154         bool enable_rc6 = true;
5155         unsigned long rc6_ctx_base;
5156         u32 rc_ctl;
5157         int rc_sw_target;
5158
5159         rc_ctl = I915_READ(GEN6_RC_CONTROL);
5160         rc_sw_target = (I915_READ(GEN6_RC_STATE) & RC_SW_TARGET_STATE_MASK) >>
5161                        RC_SW_TARGET_STATE_SHIFT;
5162         DRM_DEBUG_DRIVER("BIOS enabled RC states: "
5163                          "HW_CTRL %s HW_RC6 %s SW_TARGET_STATE %x\n",
5164                          onoff(rc_ctl & GEN6_RC_CTL_HW_ENABLE),
5165                          onoff(rc_ctl & GEN6_RC_CTL_RC6_ENABLE),
5166                          rc_sw_target);
5167
5168         if (!(I915_READ(RC6_LOCATION) & RC6_CTX_IN_DRAM)) {
5169                 DRM_DEBUG_DRIVER("RC6 Base location not set properly.\n");
5170                 enable_rc6 = false;
5171         }
5172
5173         /*
5174          * The exact context size is not known for BXT, so assume a page size
5175          * for this check.
5176          */
5177         rc6_ctx_base = I915_READ(RC6_CTX_BASE) & RC6_CTX_BASE_MASK;
5178         if (!((rc6_ctx_base >= ggtt->stolen_reserved_base) &&
5179               (rc6_ctx_base + PAGE_SIZE <= ggtt->stolen_reserved_base +
5180                                         ggtt->stolen_reserved_size))) {
5181                 DRM_DEBUG_DRIVER("RC6 Base address not as expected.\n");
5182                 enable_rc6 = false;
5183         }
5184
5185         if (!(((I915_READ(PWRCTX_MAXCNT_RCSUNIT) & IDLE_TIME_MASK) > 1) &&
5186               ((I915_READ(PWRCTX_MAXCNT_VCSUNIT0) & IDLE_TIME_MASK) > 1) &&
5187               ((I915_READ(PWRCTX_MAXCNT_BCSUNIT) & IDLE_TIME_MASK) > 1) &&
5188               ((I915_READ(PWRCTX_MAXCNT_VECSUNIT) & IDLE_TIME_MASK) > 1))) {
5189                 DRM_DEBUG_DRIVER("Engine Idle wait time not set properly.\n");
5190                 enable_rc6 = false;
5191         }
5192
5193         if (!I915_READ(GEN8_PUSHBUS_CONTROL) ||
5194             !I915_READ(GEN8_PUSHBUS_ENABLE) ||
5195             !I915_READ(GEN8_PUSHBUS_SHIFT)) {
5196                 DRM_DEBUG_DRIVER("Pushbus not setup properly.\n");
5197                 enable_rc6 = false;
5198         }
5199
5200         if (!I915_READ(GEN6_GFXPAUSE)) {
5201                 DRM_DEBUG_DRIVER("GFX pause not setup properly.\n");
5202                 enable_rc6 = false;
5203         }
5204
5205         if (!I915_READ(GEN8_MISC_CTRL0)) {
5206                 DRM_DEBUG_DRIVER("GPM control not setup properly.\n");
5207                 enable_rc6 = false;
5208         }
5209
5210         return enable_rc6;
5211 }
5212
5213 int sanitize_rc6_option(struct drm_i915_private *dev_priv, int enable_rc6)
5214 {
5215         /* No RC6 before Ironlake and code is gone for ilk. */
5216         if (INTEL_INFO(dev_priv)->gen < 6)
5217                 return 0;
5218
5219         if (!enable_rc6)
5220                 return 0;
5221
5222         if (IS_BROXTON(dev_priv) && !bxt_check_bios_rc6_setup(dev_priv)) {
5223                 DRM_INFO("RC6 disabled by BIOS\n");
5224                 return 0;
5225         }
5226
5227         /* Respect the kernel parameter if it is set */
5228         if (enable_rc6 >= 0) {
5229                 int mask;
5230
5231                 if (HAS_RC6p(dev_priv))
5232                         mask = INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE |
5233                                INTEL_RC6pp_ENABLE;
5234                 else
5235                         mask = INTEL_RC6_ENABLE;
5236
5237                 if ((enable_rc6 & mask) != enable_rc6)
5238                         DRM_DEBUG_DRIVER("Adjusting RC6 mask to %d "
5239                                          "(requested %d, valid %d)\n",
5240                                          enable_rc6 & mask, enable_rc6, mask);
5241
5242                 return enable_rc6 & mask;
5243         }
5244
5245         if (IS_IVYBRIDGE(dev_priv))
5246                 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
5247
5248         return INTEL_RC6_ENABLE;
5249 }
5250
5251 static void gen6_init_rps_frequencies(struct drm_i915_private *dev_priv)
5252 {
5253         /* All of these values are in units of 50MHz */
5254
5255         /* static values from HW: RP0 > RP1 > RPn (min_freq) */
5256         if (IS_BROXTON(dev_priv)) {
5257                 u32 rp_state_cap = I915_READ(BXT_RP_STATE_CAP);
5258                 dev_priv->rps.rp0_freq = (rp_state_cap >> 16) & 0xff;
5259                 dev_priv->rps.rp1_freq = (rp_state_cap >>  8) & 0xff;
5260                 dev_priv->rps.min_freq = (rp_state_cap >>  0) & 0xff;
5261         } else {
5262                 u32 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
5263                 dev_priv->rps.rp0_freq = (rp_state_cap >>  0) & 0xff;
5264                 dev_priv->rps.rp1_freq = (rp_state_cap >>  8) & 0xff;
5265                 dev_priv->rps.min_freq = (rp_state_cap >> 16) & 0xff;
5266         }
5267         /* hw_max = RP0 until we check for overclocking */
5268         dev_priv->rps.max_freq = dev_priv->rps.rp0_freq;
5269
5270         dev_priv->rps.efficient_freq = dev_priv->rps.rp1_freq;
5271         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv) ||
5272             IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5273                 u32 ddcc_status = 0;
5274
5275                 if (sandybridge_pcode_read(dev_priv,
5276                                            HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL,
5277                                            &ddcc_status) == 0)
5278                         dev_priv->rps.efficient_freq =
5279                                 clamp_t(u8,
5280                                         ((ddcc_status >> 8) & 0xff),
5281                                         dev_priv->rps.min_freq,
5282                                         dev_priv->rps.max_freq);
5283         }
5284
5285         if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5286                 /* Store the frequency values in 16.66 MHZ units, which is
5287                  * the natural hardware unit for SKL
5288                  */
5289                 dev_priv->rps.rp0_freq *= GEN9_FREQ_SCALER;
5290                 dev_priv->rps.rp1_freq *= GEN9_FREQ_SCALER;
5291                 dev_priv->rps.min_freq *= GEN9_FREQ_SCALER;
5292                 dev_priv->rps.max_freq *= GEN9_FREQ_SCALER;
5293                 dev_priv->rps.efficient_freq *= GEN9_FREQ_SCALER;
5294         }
5295 }
5296
5297 static void reset_rps(struct drm_i915_private *dev_priv,
5298                       void (*set)(struct drm_i915_private *, u8))
5299 {
5300         u8 freq = dev_priv->rps.cur_freq;
5301
5302         /* force a reset */
5303         dev_priv->rps.power = -1;
5304         dev_priv->rps.cur_freq = -1;
5305
5306         set(dev_priv, freq);
5307 }
5308
5309 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
5310 static void gen9_enable_rps(struct drm_i915_private *dev_priv)
5311 {
5312         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5313
5314         /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
5315         if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) {
5316                 /*
5317                  * BIOS could leave the Hw Turbo enabled, so need to explicitly
5318                  * clear out the Control register just to avoid inconsitency
5319                  * with debugfs interface, which will show  Turbo as enabled
5320                  * only and that is not expected by the User after adding the
5321                  * WaGsvDisableTurbo. Apart from this there is no problem even
5322                  * if the Turbo is left enabled in the Control register, as the
5323                  * Up/Down interrupts would remain masked.
5324                  */
5325                 gen9_disable_rps(dev_priv);
5326                 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5327                 return;
5328         }
5329
5330         /* Program defaults and thresholds for RPS*/
5331         I915_WRITE(GEN6_RC_VIDEO_FREQ,
5332                 GEN9_FREQUENCY(dev_priv->rps.rp1_freq));
5333
5334         /* 1 second timeout*/
5335         I915_WRITE(GEN6_RP_DOWN_TIMEOUT,
5336                 GT_INTERVAL_FROM_US(dev_priv, 1000000));
5337
5338         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 0xa);
5339
5340         /* Leaning on the below call to gen6_set_rps to program/setup the
5341          * Up/Down EI & threshold registers, as well as the RP_CONTROL,
5342          * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
5343         reset_rps(dev_priv, gen6_set_rps);
5344
5345         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5346 }
5347
5348 static void gen9_enable_rc6(struct drm_i915_private *dev_priv)
5349 {
5350         struct intel_engine_cs *engine;
5351         enum intel_engine_id id;
5352         uint32_t rc6_mask = 0;
5353
5354         /* 1a: Software RC state - RC0 */
5355         I915_WRITE(GEN6_RC_STATE, 0);
5356
5357         /* 1b: Get forcewake during program sequence. Although the driver
5358          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5359         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5360
5361         /* 2a: Disable RC states. */
5362         I915_WRITE(GEN6_RC_CONTROL, 0);
5363
5364         /* 2b: Program RC6 thresholds.*/
5365
5366         /* WaRsDoubleRc6WrlWithCoarsePowerGating: Doubling WRL only when CPG is enabled */
5367         if (IS_SKYLAKE(dev_priv))
5368                 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 108 << 16);
5369         else
5370                 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16);
5371         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5372         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
5373         for_each_engine(engine, dev_priv, id)
5374                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5375
5376         if (HAS_GUC(dev_priv))
5377                 I915_WRITE(GUC_MAX_IDLE_COUNT, 0xA);
5378
5379         I915_WRITE(GEN6_RC_SLEEP, 0);
5380
5381         /* 2c: Program Coarse Power Gating Policies. */
5382         I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS, 25);
5383         I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS, 25);
5384
5385         /* 3a: Enable RC6 */
5386         if (intel_enable_rc6() & INTEL_RC6_ENABLE)
5387                 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
5388         DRM_INFO("RC6 %s\n", onoff(rc6_mask & GEN6_RC_CTL_RC6_ENABLE));
5389         /* WaRsUseTimeoutMode:bxt */
5390         if (IS_BXT_REVID(dev_priv, 0, BXT_REVID_A1)) {
5391                 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us */
5392                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5393                            GEN7_RC_CTL_TO_MODE |
5394                            rc6_mask);
5395         } else {
5396                 I915_WRITE(GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */
5397                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5398                            GEN6_RC_CTL_EI_MODE(1) |
5399                            rc6_mask);
5400         }
5401
5402         /*
5403          * 3b: Enable Coarse Power Gating only when RC6 is enabled.
5404          * WaRsDisableCoarsePowerGating:skl,bxt - Render/Media PG need to be disabled with RC6.
5405          */
5406         if (NEEDS_WaRsDisableCoarsePowerGating(dev_priv))
5407                 I915_WRITE(GEN9_PG_ENABLE, 0);
5408         else
5409                 I915_WRITE(GEN9_PG_ENABLE, (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ?
5410                                 (GEN9_RENDER_PG_ENABLE | GEN9_MEDIA_PG_ENABLE) : 0);
5411
5412         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5413 }
5414
5415 static void gen8_enable_rps(struct drm_i915_private *dev_priv)
5416 {
5417         struct intel_engine_cs *engine;
5418         enum intel_engine_id id;
5419         uint32_t rc6_mask = 0;
5420
5421         /* 1a: Software RC state - RC0 */
5422         I915_WRITE(GEN6_RC_STATE, 0);
5423
5424         /* 1c & 1d: Get forcewake during program sequence. Although the driver
5425          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5426         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5427
5428         /* 2a: Disable RC states. */
5429         I915_WRITE(GEN6_RC_CONTROL, 0);
5430
5431         /* 2b: Program RC6 thresholds.*/
5432         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
5433         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5434         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
5435         for_each_engine(engine, dev_priv, id)
5436                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5437         I915_WRITE(GEN6_RC_SLEEP, 0);
5438         if (IS_BROADWELL(dev_priv))
5439                 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */
5440         else
5441                 I915_WRITE(GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
5442
5443         /* 3: Enable RC6 */
5444         if (intel_enable_rc6() & INTEL_RC6_ENABLE)
5445                 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
5446         intel_print_rc6_info(dev_priv, rc6_mask);
5447         if (IS_BROADWELL(dev_priv))
5448                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5449                                 GEN7_RC_CTL_TO_MODE |
5450                                 rc6_mask);
5451         else
5452                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
5453                                 GEN6_RC_CTL_EI_MODE(1) |
5454                                 rc6_mask);
5455
5456         /* 4 Program defaults and thresholds for RPS*/
5457         I915_WRITE(GEN6_RPNSWREQ,
5458                    HSW_FREQUENCY(dev_priv->rps.rp1_freq));
5459         I915_WRITE(GEN6_RC_VIDEO_FREQ,
5460                    HSW_FREQUENCY(dev_priv->rps.rp1_freq));
5461         /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
5462         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
5463
5464         /* Docs recommend 900MHz, and 300 MHz respectively */
5465         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
5466                    dev_priv->rps.max_freq_softlimit << 24 |
5467                    dev_priv->rps.min_freq_softlimit << 16);
5468
5469         I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
5470         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
5471         I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
5472         I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
5473
5474         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5475
5476         /* 5: Enable RPS */
5477         I915_WRITE(GEN6_RP_CONTROL,
5478                    GEN6_RP_MEDIA_TURBO |
5479                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
5480                    GEN6_RP_MEDIA_IS_GFX |
5481                    GEN6_RP_ENABLE |
5482                    GEN6_RP_UP_BUSY_AVG |
5483                    GEN6_RP_DOWN_IDLE_AVG);
5484
5485         /* 6: Ring frequency + overclocking (our driver does this later */
5486
5487         reset_rps(dev_priv, gen6_set_rps);
5488
5489         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5490 }
5491
5492 static void gen6_enable_rps(struct drm_i915_private *dev_priv)
5493 {
5494         struct intel_engine_cs *engine;
5495         enum intel_engine_id id;
5496         u32 rc6vids, rc6_mask = 0;
5497         u32 gtfifodbg;
5498         int rc6_mode;
5499         int ret;
5500
5501         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5502
5503         /* Here begins a magic sequence of register writes to enable
5504          * auto-downclocking.
5505          *
5506          * Perhaps there might be some value in exposing these to
5507          * userspace...
5508          */
5509         I915_WRITE(GEN6_RC_STATE, 0);
5510
5511         /* Clear the DBG now so we don't confuse earlier errors */
5512         gtfifodbg = I915_READ(GTFIFODBG);
5513         if (gtfifodbg) {
5514                 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
5515                 I915_WRITE(GTFIFODBG, gtfifodbg);
5516         }
5517
5518         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5519
5520         /* disable the counters and set deterministic thresholds */
5521         I915_WRITE(GEN6_RC_CONTROL, 0);
5522
5523         I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
5524         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
5525         I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
5526         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
5527         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
5528
5529         for_each_engine(engine, dev_priv, id)
5530                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
5531
5532         I915_WRITE(GEN6_RC_SLEEP, 0);
5533         I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
5534         if (IS_IVYBRIDGE(dev_priv))
5535                 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
5536         else
5537                 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
5538         I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
5539         I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
5540
5541         /* Check if we are enabling RC6 */
5542         rc6_mode = intel_enable_rc6();
5543         if (rc6_mode & INTEL_RC6_ENABLE)
5544                 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
5545
5546         /* We don't use those on Haswell */
5547         if (!IS_HASWELL(dev_priv)) {
5548                 if (rc6_mode & INTEL_RC6p_ENABLE)
5549                         rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
5550
5551                 if (rc6_mode & INTEL_RC6pp_ENABLE)
5552                         rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
5553         }
5554
5555         intel_print_rc6_info(dev_priv, rc6_mask);
5556
5557         I915_WRITE(GEN6_RC_CONTROL,
5558                    rc6_mask |
5559                    GEN6_RC_CTL_EI_MODE(1) |
5560                    GEN6_RC_CTL_HW_ENABLE);
5561
5562         /* Power down if completely idle for over 50ms */
5563         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
5564         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5565
5566         reset_rps(dev_priv, gen6_set_rps);
5567
5568         rc6vids = 0;
5569         ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
5570         if (IS_GEN6(dev_priv) && ret) {
5571                 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
5572         } else if (IS_GEN6(dev_priv) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
5573                 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
5574                           GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
5575                 rc6vids &= 0xffff00;
5576                 rc6vids |= GEN6_ENCODE_RC6_VID(450);
5577                 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
5578                 if (ret)
5579                         DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
5580         }
5581
5582         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5583 }
5584
5585 static void gen6_update_ring_freq(struct drm_i915_private *dev_priv)
5586 {
5587         int min_freq = 15;
5588         unsigned int gpu_freq;
5589         unsigned int max_ia_freq, min_ring_freq;
5590         unsigned int max_gpu_freq, min_gpu_freq;
5591         int scaling_factor = 180;
5592         struct cpufreq_policy *policy;
5593
5594         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5595
5596         policy = cpufreq_cpu_get(0);
5597         if (policy) {
5598                 max_ia_freq = policy->cpuinfo.max_freq;
5599                 cpufreq_cpu_put(policy);
5600         } else {
5601                 /*
5602                  * Default to measured freq if none found, PCU will ensure we
5603                  * don't go over
5604                  */
5605                 max_ia_freq = tsc_khz;
5606         }
5607
5608         /* Convert from kHz to MHz */
5609         max_ia_freq /= 1000;
5610
5611         min_ring_freq = I915_READ(DCLK) & 0xf;
5612         /* convert DDR frequency from units of 266.6MHz to bandwidth */
5613         min_ring_freq = mult_frac(min_ring_freq, 8, 3);
5614
5615         if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5616                 /* Convert GT frequency to 50 HZ units */
5617                 min_gpu_freq = dev_priv->rps.min_freq / GEN9_FREQ_SCALER;
5618                 max_gpu_freq = dev_priv->rps.max_freq / GEN9_FREQ_SCALER;
5619         } else {
5620                 min_gpu_freq = dev_priv->rps.min_freq;
5621                 max_gpu_freq = dev_priv->rps.max_freq;
5622         }
5623
5624         /*
5625          * For each potential GPU frequency, load a ring frequency we'd like
5626          * to use for memory access.  We do this by specifying the IA frequency
5627          * the PCU should use as a reference to determine the ring frequency.
5628          */
5629         for (gpu_freq = max_gpu_freq; gpu_freq >= min_gpu_freq; gpu_freq--) {
5630                 int diff = max_gpu_freq - gpu_freq;
5631                 unsigned int ia_freq = 0, ring_freq = 0;
5632
5633                 if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv)) {
5634                         /*
5635                          * ring_freq = 2 * GT. ring_freq is in 100MHz units
5636                          * No floor required for ring frequency on SKL.
5637                          */
5638                         ring_freq = gpu_freq;
5639                 } else if (INTEL_INFO(dev_priv)->gen >= 8) {
5640                         /* max(2 * GT, DDR). NB: GT is 50MHz units */
5641                         ring_freq = max(min_ring_freq, gpu_freq);
5642                 } else if (IS_HASWELL(dev_priv)) {
5643                         ring_freq = mult_frac(gpu_freq, 5, 4);
5644                         ring_freq = max(min_ring_freq, ring_freq);
5645                         /* leave ia_freq as the default, chosen by cpufreq */
5646                 } else {
5647                         /* On older processors, there is no separate ring
5648                          * clock domain, so in order to boost the bandwidth
5649                          * of the ring, we need to upclock the CPU (ia_freq).
5650                          *
5651                          * For GPU frequencies less than 750MHz,
5652                          * just use the lowest ring freq.
5653                          */
5654                         if (gpu_freq < min_freq)
5655                                 ia_freq = 800;
5656                         else
5657                                 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
5658                         ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
5659                 }
5660
5661                 sandybridge_pcode_write(dev_priv,
5662                                         GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
5663                                         ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
5664                                         ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
5665                                         gpu_freq);
5666         }
5667 }
5668
5669 static int cherryview_rps_max_freq(struct drm_i915_private *dev_priv)
5670 {
5671         u32 val, rp0;
5672
5673         val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
5674
5675         switch (INTEL_INFO(dev_priv)->sseu.eu_total) {
5676         case 8:
5677                 /* (2 * 4) config */
5678                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT);
5679                 break;
5680         case 12:
5681                 /* (2 * 6) config */
5682                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT);
5683                 break;
5684         case 16:
5685                 /* (2 * 8) config */
5686         default:
5687                 /* Setting (2 * 8) Min RP0 for any other combination */
5688                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT);
5689                 break;
5690         }
5691
5692         rp0 = (rp0 & FB_GFX_FREQ_FUSE_MASK);
5693
5694         return rp0;
5695 }
5696
5697 static int cherryview_rps_rpe_freq(struct drm_i915_private *dev_priv)
5698 {
5699         u32 val, rpe;
5700
5701         val = vlv_punit_read(dev_priv, PUNIT_GPU_DUTYCYCLE_REG);
5702         rpe = (val >> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK;
5703
5704         return rpe;
5705 }
5706
5707 static int cherryview_rps_guar_freq(struct drm_i915_private *dev_priv)
5708 {
5709         u32 val, rp1;
5710
5711         val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
5712         rp1 = (val & FB_GFX_FREQ_FUSE_MASK);
5713
5714         return rp1;
5715 }
5716
5717 static int valleyview_rps_guar_freq(struct drm_i915_private *dev_priv)
5718 {
5719         u32 val, rp1;
5720
5721         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
5722
5723         rp1 = (val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT;
5724
5725         return rp1;
5726 }
5727
5728 static int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
5729 {
5730         u32 val, rp0;
5731
5732         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
5733
5734         rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
5735         /* Clamp to max */
5736         rp0 = min_t(u32, rp0, 0xea);
5737
5738         return rp0;
5739 }
5740
5741 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
5742 {
5743         u32 val, rpe;
5744
5745         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
5746         rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
5747         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
5748         rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
5749
5750         return rpe;
5751 }
5752
5753 static int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
5754 {
5755         u32 val;
5756
5757         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
5758         /*
5759          * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
5760          * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
5761          * a BYT-M B0 the above register contains 0xbf. Moreover when setting
5762          * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
5763          * to make sure it matches what Punit accepts.
5764          */
5765         return max_t(u32, val, 0xc0);
5766 }
5767
5768 /* Check that the pctx buffer wasn't move under us. */
5769 static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
5770 {
5771         unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
5772
5773         WARN_ON(pctx_addr != dev_priv->mm.stolen_base +
5774                              dev_priv->vlv_pctx->stolen->start);
5775 }
5776
5777
5778 /* Check that the pcbr address is not empty. */
5779 static void cherryview_check_pctx(struct drm_i915_private *dev_priv)
5780 {
5781         unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
5782
5783         WARN_ON((pctx_addr >> VLV_PCBR_ADDR_SHIFT) == 0);
5784 }
5785
5786 static void cherryview_setup_pctx(struct drm_i915_private *dev_priv)
5787 {
5788         struct i915_ggtt *ggtt = &dev_priv->ggtt;
5789         unsigned long pctx_paddr, paddr;
5790         u32 pcbr;
5791         int pctx_size = 32*1024;
5792
5793         pcbr = I915_READ(VLV_PCBR);
5794         if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) {
5795                 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5796                 paddr = (dev_priv->mm.stolen_base +
5797                          (ggtt->stolen_size - pctx_size));
5798
5799                 pctx_paddr = (paddr & (~4095));
5800                 I915_WRITE(VLV_PCBR, pctx_paddr);
5801         }
5802
5803         DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
5804 }
5805
5806 static void valleyview_setup_pctx(struct drm_i915_private *dev_priv)
5807 {
5808         struct drm_i915_gem_object *pctx;
5809         unsigned long pctx_paddr;
5810         u32 pcbr;
5811         int pctx_size = 24*1024;
5812
5813         pcbr = I915_READ(VLV_PCBR);
5814         if (pcbr) {
5815                 /* BIOS set it up already, grab the pre-alloc'd space */
5816                 int pcbr_offset;
5817
5818                 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
5819                 pctx = i915_gem_object_create_stolen_for_preallocated(&dev_priv->drm,
5820                                                                       pcbr_offset,
5821                                                                       I915_GTT_OFFSET_NONE,
5822                                                                       pctx_size);
5823                 goto out;
5824         }
5825
5826         DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5827
5828         /*
5829          * From the Gunit register HAS:
5830          * The Gfx driver is expected to program this register and ensure
5831          * proper allocation within Gfx stolen memory.  For example, this
5832          * register should be programmed such than the PCBR range does not
5833          * overlap with other ranges, such as the frame buffer, protected
5834          * memory, or any other relevant ranges.
5835          */
5836         pctx = i915_gem_object_create_stolen(&dev_priv->drm, pctx_size);
5837         if (!pctx) {
5838                 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
5839                 goto out;
5840         }
5841
5842         pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
5843         I915_WRITE(VLV_PCBR, pctx_paddr);
5844
5845 out:
5846         DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
5847         dev_priv->vlv_pctx = pctx;
5848 }
5849
5850 static void valleyview_cleanup_pctx(struct drm_i915_private *dev_priv)
5851 {
5852         if (WARN_ON(!dev_priv->vlv_pctx))
5853                 return;
5854
5855         i915_gem_object_put(dev_priv->vlv_pctx);
5856         dev_priv->vlv_pctx = NULL;
5857 }
5858
5859 static void vlv_init_gpll_ref_freq(struct drm_i915_private *dev_priv)
5860 {
5861         dev_priv->rps.gpll_ref_freq =
5862                 vlv_get_cck_clock(dev_priv, "GPLL ref",
5863                                   CCK_GPLL_CLOCK_CONTROL,
5864                                   dev_priv->czclk_freq);
5865
5866         DRM_DEBUG_DRIVER("GPLL reference freq: %d kHz\n",
5867                          dev_priv->rps.gpll_ref_freq);
5868 }
5869
5870 static void valleyview_init_gt_powersave(struct drm_i915_private *dev_priv)
5871 {
5872         u32 val;
5873
5874         valleyview_setup_pctx(dev_priv);
5875
5876         vlv_init_gpll_ref_freq(dev_priv);
5877
5878         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5879         switch ((val >> 6) & 3) {
5880         case 0:
5881         case 1:
5882                 dev_priv->mem_freq = 800;
5883                 break;
5884         case 2:
5885                 dev_priv->mem_freq = 1066;
5886                 break;
5887         case 3:
5888                 dev_priv->mem_freq = 1333;
5889                 break;
5890         }
5891         DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
5892
5893         dev_priv->rps.max_freq = valleyview_rps_max_freq(dev_priv);
5894         dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
5895         DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5896                          intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
5897                          dev_priv->rps.max_freq);
5898
5899         dev_priv->rps.efficient_freq = valleyview_rps_rpe_freq(dev_priv);
5900         DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5901                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5902                          dev_priv->rps.efficient_freq);
5903
5904         dev_priv->rps.rp1_freq = valleyview_rps_guar_freq(dev_priv);
5905         DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
5906                          intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
5907                          dev_priv->rps.rp1_freq);
5908
5909         dev_priv->rps.min_freq = valleyview_rps_min_freq(dev_priv);
5910         DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5911                          intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
5912                          dev_priv->rps.min_freq);
5913 }
5914
5915 static void cherryview_init_gt_powersave(struct drm_i915_private *dev_priv)
5916 {
5917         u32 val;
5918
5919         cherryview_setup_pctx(dev_priv);
5920
5921         vlv_init_gpll_ref_freq(dev_priv);
5922
5923         mutex_lock(&dev_priv->sb_lock);
5924         val = vlv_cck_read(dev_priv, CCK_FUSE_REG);
5925         mutex_unlock(&dev_priv->sb_lock);
5926
5927         switch ((val >> 2) & 0x7) {
5928         case 3:
5929                 dev_priv->mem_freq = 2000;
5930                 break;
5931         default:
5932                 dev_priv->mem_freq = 1600;
5933                 break;
5934         }
5935         DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
5936
5937         dev_priv->rps.max_freq = cherryview_rps_max_freq(dev_priv);
5938         dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
5939         DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5940                          intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
5941                          dev_priv->rps.max_freq);
5942
5943         dev_priv->rps.efficient_freq = cherryview_rps_rpe_freq(dev_priv);
5944         DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5945                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5946                          dev_priv->rps.efficient_freq);
5947
5948         dev_priv->rps.rp1_freq = cherryview_rps_guar_freq(dev_priv);
5949         DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
5950                          intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
5951                          dev_priv->rps.rp1_freq);
5952
5953         /* PUnit validated range is only [RPe, RP0] */
5954         dev_priv->rps.min_freq = dev_priv->rps.efficient_freq;
5955         DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5956                          intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
5957                          dev_priv->rps.min_freq);
5958
5959         WARN_ONCE((dev_priv->rps.max_freq |
5960                    dev_priv->rps.efficient_freq |
5961                    dev_priv->rps.rp1_freq |
5962                    dev_priv->rps.min_freq) & 1,
5963                   "Odd GPU freq values\n");
5964 }
5965
5966 static void valleyview_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
5967 {
5968         valleyview_cleanup_pctx(dev_priv);
5969 }
5970
5971 static void cherryview_enable_rps(struct drm_i915_private *dev_priv)
5972 {
5973         struct intel_engine_cs *engine;
5974         enum intel_engine_id id;
5975         u32 gtfifodbg, val, rc6_mode = 0, pcbr;
5976
5977         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5978
5979         gtfifodbg = I915_READ(GTFIFODBG) & ~(GT_FIFO_SBDEDICATE_FREE_ENTRY_CHV |
5980                                              GT_FIFO_FREE_ENTRIES_CHV);
5981         if (gtfifodbg) {
5982                 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5983                                  gtfifodbg);
5984                 I915_WRITE(GTFIFODBG, gtfifodbg);
5985         }
5986
5987         cherryview_check_pctx(dev_priv);
5988
5989         /* 1a & 1b: Get forcewake during program sequence. Although the driver
5990          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5991         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5992
5993         /*  Disable RC states. */
5994         I915_WRITE(GEN6_RC_CONTROL, 0);
5995
5996         /* 2a: Program RC6 thresholds.*/
5997         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
5998         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5999         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
6000
6001         for_each_engine(engine, dev_priv, id)
6002                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6003         I915_WRITE(GEN6_RC_SLEEP, 0);
6004
6005         /* TO threshold set to 500 us ( 0x186 * 1.28 us) */
6006         I915_WRITE(GEN6_RC6_THRESHOLD, 0x186);
6007
6008         /* allows RC6 residency counter to work */
6009         I915_WRITE(VLV_COUNTER_CONTROL,
6010                    _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
6011                                       VLV_MEDIA_RC6_COUNT_EN |
6012                                       VLV_RENDER_RC6_COUNT_EN));
6013
6014         /* For now we assume BIOS is allocating and populating the PCBR  */
6015         pcbr = I915_READ(VLV_PCBR);
6016
6017         /* 3: Enable RC6 */
6018         if ((intel_enable_rc6() & INTEL_RC6_ENABLE) &&
6019             (pcbr >> VLV_PCBR_ADDR_SHIFT))
6020                 rc6_mode = GEN7_RC_CTL_TO_MODE;
6021
6022         I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
6023
6024         /* 4 Program defaults and thresholds for RPS*/
6025         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
6026         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
6027         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
6028         I915_WRITE(GEN6_RP_UP_EI, 66000);
6029         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
6030
6031         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
6032
6033         /* 5: Enable RPS */
6034         I915_WRITE(GEN6_RP_CONTROL,
6035                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
6036                    GEN6_RP_MEDIA_IS_GFX |
6037                    GEN6_RP_ENABLE |
6038                    GEN6_RP_UP_BUSY_AVG |
6039                    GEN6_RP_DOWN_IDLE_AVG);
6040
6041         /* Setting Fixed Bias */
6042         val = VLV_OVERRIDE_EN |
6043                   VLV_SOC_TDP_EN |
6044                   CHV_BIAS_CPU_50_SOC_50;
6045         vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
6046
6047         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
6048
6049         /* RPS code assumes GPLL is used */
6050         WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
6051
6052         DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
6053         DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
6054
6055         reset_rps(dev_priv, valleyview_set_rps);
6056
6057         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6058 }
6059
6060 static void valleyview_enable_rps(struct drm_i915_private *dev_priv)
6061 {
6062         struct intel_engine_cs *engine;
6063         enum intel_engine_id id;
6064         u32 gtfifodbg, val, rc6_mode = 0;
6065
6066         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
6067
6068         valleyview_check_pctx(dev_priv);
6069
6070         gtfifodbg = I915_READ(GTFIFODBG);
6071         if (gtfifodbg) {
6072                 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
6073                                  gtfifodbg);
6074                 I915_WRITE(GTFIFODBG, gtfifodbg);
6075         }
6076
6077         /* If VLV, Forcewake all wells, else re-direct to regular path */
6078         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
6079
6080         /*  Disable RC states. */
6081         I915_WRITE(GEN6_RC_CONTROL, 0);
6082
6083         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
6084         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
6085         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
6086         I915_WRITE(GEN6_RP_UP_EI, 66000);
6087         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
6088
6089         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
6090
6091         I915_WRITE(GEN6_RP_CONTROL,
6092                    GEN6_RP_MEDIA_TURBO |
6093                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
6094                    GEN6_RP_MEDIA_IS_GFX |
6095                    GEN6_RP_ENABLE |
6096                    GEN6_RP_UP_BUSY_AVG |
6097                    GEN6_RP_DOWN_IDLE_CONT);
6098
6099         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
6100         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
6101         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
6102
6103         for_each_engine(engine, dev_priv, id)
6104                 I915_WRITE(RING_MAX_IDLE(engine->mmio_base), 10);
6105
6106         I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
6107
6108         /* allows RC6 residency counter to work */
6109         I915_WRITE(VLV_COUNTER_CONTROL,
6110                    _MASKED_BIT_ENABLE(VLV_MEDIA_RC0_COUNT_EN |
6111                                       VLV_RENDER_RC0_COUNT_EN |
6112                                       VLV_MEDIA_RC6_COUNT_EN |
6113                                       VLV_RENDER_RC6_COUNT_EN));
6114
6115         if (intel_enable_rc6() & INTEL_RC6_ENABLE)
6116                 rc6_mode = GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
6117
6118         intel_print_rc6_info(dev_priv, rc6_mode);
6119
6120         I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
6121
6122         /* Setting Fixed Bias */
6123         val = VLV_OVERRIDE_EN |
6124                   VLV_SOC_TDP_EN |
6125                   VLV_BIAS_CPU_125_SOC_875;
6126         vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
6127
6128         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
6129
6130         /* RPS code assumes GPLL is used */
6131         WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
6132
6133         DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
6134         DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
6135
6136         reset_rps(dev_priv, valleyview_set_rps);
6137
6138         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
6139 }
6140
6141 static unsigned long intel_pxfreq(u32 vidfreq)
6142 {
6143         unsigned long freq;
6144         int div = (vidfreq & 0x3f0000) >> 16;
6145         int post = (vidfreq & 0x3000) >> 12;
6146         int pre = (vidfreq & 0x7);
6147
6148         if (!pre)
6149                 return 0;
6150
6151         freq = ((div * 133333) / ((1<<post) * pre));
6152
6153         return freq;
6154 }
6155
6156 static const struct cparams {
6157         u16 i;
6158         u16 t;
6159         u16 m;
6160         u16 c;
6161 } cparams[] = {
6162         { 1, 1333, 301, 28664 },
6163         { 1, 1066, 294, 24460 },
6164         { 1, 800, 294, 25192 },
6165         { 0, 1333, 276, 27605 },
6166         { 0, 1066, 276, 27605 },
6167         { 0, 800, 231, 23784 },
6168 };
6169
6170 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
6171 {
6172         u64 total_count, diff, ret;
6173         u32 count1, count2, count3, m = 0, c = 0;
6174         unsigned long now = jiffies_to_msecs(jiffies), diff1;
6175         int i;
6176
6177         assert_spin_locked(&mchdev_lock);
6178
6179         diff1 = now - dev_priv->ips.last_time1;
6180
6181         /* Prevent division-by-zero if we are asking too fast.
6182          * Also, we don't get interesting results if we are polling
6183          * faster than once in 10ms, so just return the saved value
6184          * in such cases.
6185          */
6186         if (diff1 <= 10)
6187                 return dev_priv->ips.chipset_power;
6188
6189         count1 = I915_READ(DMIEC);
6190         count2 = I915_READ(DDREC);
6191         count3 = I915_READ(CSIEC);
6192
6193         total_count = count1 + count2 + count3;
6194
6195         /* FIXME: handle per-counter overflow */
6196         if (total_count < dev_priv->ips.last_count1) {
6197                 diff = ~0UL - dev_priv->ips.last_count1;
6198                 diff += total_count;
6199         } else {
6200                 diff = total_count - dev_priv->ips.last_count1;
6201         }
6202
6203         for (i = 0; i < ARRAY_SIZE(cparams); i++) {
6204                 if (cparams[i].i == dev_priv->ips.c_m &&
6205                     cparams[i].t == dev_priv->ips.r_t) {
6206                         m = cparams[i].m;
6207                         c = cparams[i].c;
6208                         break;
6209                 }
6210         }
6211
6212         diff = div_u64(diff, diff1);
6213         ret = ((m * diff) + c);
6214         ret = div_u64(ret, 10);
6215
6216         dev_priv->ips.last_count1 = total_count;
6217         dev_priv->ips.last_time1 = now;
6218
6219         dev_priv->ips.chipset_power = ret;
6220
6221         return ret;
6222 }
6223
6224 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
6225 {
6226         unsigned long val;
6227
6228         if (INTEL_INFO(dev_priv)->gen != 5)
6229                 return 0;
6230
6231         spin_lock_irq(&mchdev_lock);
6232
6233         val = __i915_chipset_val(dev_priv);
6234
6235         spin_unlock_irq(&mchdev_lock);
6236
6237         return val;
6238 }
6239
6240 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
6241 {
6242         unsigned long m, x, b;
6243         u32 tsfs;
6244
6245         tsfs = I915_READ(TSFS);
6246
6247         m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
6248         x = I915_READ8(TR1);
6249
6250         b = tsfs & TSFS_INTR_MASK;
6251
6252         return ((m * x) / 127) - b;
6253 }
6254
6255 static int _pxvid_to_vd(u8 pxvid)
6256 {
6257         if (pxvid == 0)
6258                 return 0;
6259
6260         if (pxvid >= 8 && pxvid < 31)
6261                 pxvid = 31;
6262
6263         return (pxvid + 2) * 125;
6264 }
6265
6266 static u32 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
6267 {
6268         const int vd = _pxvid_to_vd(pxvid);
6269         const int vm = vd - 1125;
6270
6271         if (INTEL_INFO(dev_priv)->is_mobile)
6272                 return vm > 0 ? vm : 0;
6273
6274         return vd;
6275 }
6276
6277 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
6278 {
6279         u64 now, diff, diffms;
6280         u32 count;
6281
6282         assert_spin_locked(&mchdev_lock);
6283
6284         now = ktime_get_raw_ns();
6285         diffms = now - dev_priv->ips.last_time2;
6286         do_div(diffms, NSEC_PER_MSEC);
6287
6288         /* Don't divide by 0 */
6289         if (!diffms)
6290                 return;
6291
6292         count = I915_READ(GFXEC);
6293
6294         if (count < dev_priv->ips.last_count2) {
6295                 diff = ~0UL - dev_priv->ips.last_count2;
6296                 diff += count;
6297         } else {
6298                 diff = count - dev_priv->ips.last_count2;
6299         }
6300
6301         dev_priv->ips.last_count2 = count;
6302         dev_priv->ips.last_time2 = now;
6303
6304         /* More magic constants... */
6305         diff = diff * 1181;
6306         diff = div_u64(diff, diffms * 10);
6307         dev_priv->ips.gfx_power = diff;
6308 }
6309
6310 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
6311 {
6312         if (INTEL_INFO(dev_priv)->gen != 5)
6313                 return;
6314
6315         spin_lock_irq(&mchdev_lock);
6316
6317         __i915_update_gfx_val(dev_priv);
6318
6319         spin_unlock_irq(&mchdev_lock);
6320 }
6321
6322 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
6323 {
6324         unsigned long t, corr, state1, corr2, state2;
6325         u32 pxvid, ext_v;
6326
6327         assert_spin_locked(&mchdev_lock);
6328
6329         pxvid = I915_READ(PXVFREQ(dev_priv->rps.cur_freq));
6330         pxvid = (pxvid >> 24) & 0x7f;
6331         ext_v = pvid_to_extvid(dev_priv, pxvid);
6332
6333         state1 = ext_v;
6334
6335         t = i915_mch_val(dev_priv);
6336
6337         /* Revel in the empirically derived constants */
6338
6339         /* Correction factor in 1/100000 units */
6340         if (t > 80)
6341                 corr = ((t * 2349) + 135940);
6342         else if (t >= 50)
6343                 corr = ((t * 964) + 29317);
6344         else /* < 50 */
6345                 corr = ((t * 301) + 1004);
6346
6347         corr = corr * ((150142 * state1) / 10000 - 78642);
6348         corr /= 100000;
6349         corr2 = (corr * dev_priv->ips.corr);
6350
6351         state2 = (corr2 * state1) / 10000;
6352         state2 /= 100; /* convert to mW */
6353
6354         __i915_update_gfx_val(dev_priv);
6355
6356         return dev_priv->ips.gfx_power + state2;
6357 }
6358
6359 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
6360 {
6361         unsigned long val;
6362
6363         if (INTEL_INFO(dev_priv)->gen != 5)
6364                 return 0;
6365
6366         spin_lock_irq(&mchdev_lock);
6367
6368         val = __i915_gfx_val(dev_priv);
6369
6370         spin_unlock_irq(&mchdev_lock);
6371
6372         return val;
6373 }
6374
6375 /**
6376  * i915_read_mch_val - return value for IPS use
6377  *
6378  * Calculate and return a value for the IPS driver to use when deciding whether
6379  * we have thermal and power headroom to increase CPU or GPU power budget.
6380  */
6381 unsigned long i915_read_mch_val(void)
6382 {
6383         struct drm_i915_private *dev_priv;
6384         unsigned long chipset_val, graphics_val, ret = 0;
6385
6386         spin_lock_irq(&mchdev_lock);
6387         if (!i915_mch_dev)
6388                 goto out_unlock;
6389         dev_priv = i915_mch_dev;
6390
6391         chipset_val = __i915_chipset_val(dev_priv);
6392         graphics_val = __i915_gfx_val(dev_priv);
6393
6394         ret = chipset_val + graphics_val;
6395
6396 out_unlock:
6397         spin_unlock_irq(&mchdev_lock);
6398
6399         return ret;
6400 }
6401 EXPORT_SYMBOL_GPL(i915_read_mch_val);
6402
6403 /**
6404  * i915_gpu_raise - raise GPU frequency limit
6405  *
6406  * Raise the limit; IPS indicates we have thermal headroom.
6407  */
6408 bool i915_gpu_raise(void)
6409 {
6410         struct drm_i915_private *dev_priv;
6411         bool ret = true;
6412
6413         spin_lock_irq(&mchdev_lock);
6414         if (!i915_mch_dev) {
6415                 ret = false;
6416                 goto out_unlock;
6417         }
6418         dev_priv = i915_mch_dev;
6419
6420         if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
6421                 dev_priv->ips.max_delay--;
6422
6423 out_unlock:
6424         spin_unlock_irq(&mchdev_lock);
6425
6426         return ret;
6427 }
6428 EXPORT_SYMBOL_GPL(i915_gpu_raise);
6429
6430 /**
6431  * i915_gpu_lower - lower GPU frequency limit
6432  *
6433  * IPS indicates we're close to a thermal limit, so throttle back the GPU
6434  * frequency maximum.
6435  */
6436 bool i915_gpu_lower(void)
6437 {
6438         struct drm_i915_private *dev_priv;
6439         bool ret = true;
6440
6441         spin_lock_irq(&mchdev_lock);
6442         if (!i915_mch_dev) {
6443                 ret = false;
6444                 goto out_unlock;
6445         }
6446         dev_priv = i915_mch_dev;
6447
6448         if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
6449                 dev_priv->ips.max_delay++;
6450
6451 out_unlock:
6452         spin_unlock_irq(&mchdev_lock);
6453
6454         return ret;
6455 }
6456 EXPORT_SYMBOL_GPL(i915_gpu_lower);
6457
6458 /**
6459  * i915_gpu_busy - indicate GPU business to IPS
6460  *
6461  * Tell the IPS driver whether or not the GPU is busy.
6462  */
6463 bool i915_gpu_busy(void)
6464 {
6465         bool ret = false;
6466
6467         spin_lock_irq(&mchdev_lock);
6468         if (i915_mch_dev)
6469                 ret = i915_mch_dev->gt.awake;
6470         spin_unlock_irq(&mchdev_lock);
6471
6472         return ret;
6473 }
6474 EXPORT_SYMBOL_GPL(i915_gpu_busy);
6475
6476 /**
6477  * i915_gpu_turbo_disable - disable graphics turbo
6478  *
6479  * Disable graphics turbo by resetting the max frequency and setting the
6480  * current frequency to the default.
6481  */
6482 bool i915_gpu_turbo_disable(void)
6483 {
6484         struct drm_i915_private *dev_priv;
6485         bool ret = true;
6486
6487         spin_lock_irq(&mchdev_lock);
6488         if (!i915_mch_dev) {
6489                 ret = false;
6490                 goto out_unlock;
6491         }
6492         dev_priv = i915_mch_dev;
6493
6494         dev_priv->ips.max_delay = dev_priv->ips.fstart;
6495
6496         if (!ironlake_set_drps(dev_priv, dev_priv->ips.fstart))
6497                 ret = false;
6498
6499 out_unlock:
6500         spin_unlock_irq(&mchdev_lock);
6501
6502         return ret;
6503 }
6504 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
6505
6506 /**
6507  * Tells the intel_ips driver that the i915 driver is now loaded, if
6508  * IPS got loaded first.
6509  *
6510  * This awkward dance is so that neither module has to depend on the
6511  * other in order for IPS to do the appropriate communication of
6512  * GPU turbo limits to i915.
6513  */
6514 static void
6515 ips_ping_for_i915_load(void)
6516 {
6517         void (*link)(void);
6518
6519         link = symbol_get(ips_link_to_i915_driver);
6520         if (link) {
6521                 link();
6522                 symbol_put(ips_link_to_i915_driver);
6523         }
6524 }
6525
6526 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
6527 {
6528         /* We only register the i915 ips part with intel-ips once everything is
6529          * set up, to avoid intel-ips sneaking in and reading bogus values. */
6530         spin_lock_irq(&mchdev_lock);
6531         i915_mch_dev = dev_priv;
6532         spin_unlock_irq(&mchdev_lock);
6533
6534         ips_ping_for_i915_load();
6535 }
6536
6537 void intel_gpu_ips_teardown(void)
6538 {
6539         spin_lock_irq(&mchdev_lock);
6540         i915_mch_dev = NULL;
6541         spin_unlock_irq(&mchdev_lock);
6542 }
6543
6544 static void intel_init_emon(struct drm_i915_private *dev_priv)
6545 {
6546         u32 lcfuse;
6547         u8 pxw[16];
6548         int i;
6549
6550         /* Disable to program */
6551         I915_WRITE(ECR, 0);
6552         POSTING_READ(ECR);
6553
6554         /* Program energy weights for various events */
6555         I915_WRITE(SDEW, 0x15040d00);
6556         I915_WRITE(CSIEW0, 0x007f0000);
6557         I915_WRITE(CSIEW1, 0x1e220004);
6558         I915_WRITE(CSIEW2, 0x04000004);
6559
6560         for (i = 0; i < 5; i++)
6561                 I915_WRITE(PEW(i), 0);
6562         for (i = 0; i < 3; i++)
6563                 I915_WRITE(DEW(i), 0);
6564
6565         /* Program P-state weights to account for frequency power adjustment */
6566         for (i = 0; i < 16; i++) {
6567                 u32 pxvidfreq = I915_READ(PXVFREQ(i));
6568                 unsigned long freq = intel_pxfreq(pxvidfreq);
6569                 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
6570                         PXVFREQ_PX_SHIFT;
6571                 unsigned long val;
6572
6573                 val = vid * vid;
6574                 val *= (freq / 1000);
6575                 val *= 255;
6576                 val /= (127*127*900);
6577                 if (val > 0xff)
6578                         DRM_ERROR("bad pxval: %ld\n", val);
6579                 pxw[i] = val;
6580         }
6581         /* Render standby states get 0 weight */
6582         pxw[14] = 0;
6583         pxw[15] = 0;
6584
6585         for (i = 0; i < 4; i++) {
6586                 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
6587                         (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
6588                 I915_WRITE(PXW(i), val);
6589         }
6590
6591         /* Adjust magic regs to magic values (more experimental results) */
6592         I915_WRITE(OGW0, 0);
6593         I915_WRITE(OGW1, 0);
6594         I915_WRITE(EG0, 0x00007f00);
6595         I915_WRITE(EG1, 0x0000000e);
6596         I915_WRITE(EG2, 0x000e0000);
6597         I915_WRITE(EG3, 0x68000300);
6598         I915_WRITE(EG4, 0x42000000);
6599         I915_WRITE(EG5, 0x00140031);
6600         I915_WRITE(EG6, 0);
6601         I915_WRITE(EG7, 0);
6602
6603         for (i = 0; i < 8; i++)
6604                 I915_WRITE(PXWL(i), 0);
6605
6606         /* Enable PMON + select events */
6607         I915_WRITE(ECR, 0x80000019);
6608
6609         lcfuse = I915_READ(LCFUSE02);
6610
6611         dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
6612 }
6613
6614 void intel_init_gt_powersave(struct drm_i915_private *dev_priv)
6615 {
6616         /*
6617          * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
6618          * requirement.
6619          */
6620         if (!i915.enable_rc6) {
6621                 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
6622                 intel_runtime_pm_get(dev_priv);
6623         }
6624
6625         mutex_lock(&dev_priv->drm.struct_mutex);
6626         mutex_lock(&dev_priv->rps.hw_lock);
6627
6628         /* Initialize RPS limits (for userspace) */
6629         if (IS_CHERRYVIEW(dev_priv))
6630                 cherryview_init_gt_powersave(dev_priv);
6631         else if (IS_VALLEYVIEW(dev_priv))
6632                 valleyview_init_gt_powersave(dev_priv);
6633         else if (INTEL_GEN(dev_priv) >= 6)
6634                 gen6_init_rps_frequencies(dev_priv);
6635
6636         /* Derive initial user preferences/limits from the hardware limits */
6637         dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
6638         dev_priv->rps.cur_freq = dev_priv->rps.idle_freq;
6639
6640         dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
6641         dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
6642
6643         if (IS_HASWELL(dev_priv) || IS_BROADWELL(dev_priv))
6644                 dev_priv->rps.min_freq_softlimit =
6645                         max_t(int,
6646                               dev_priv->rps.efficient_freq,
6647                               intel_freq_opcode(dev_priv, 450));
6648
6649         /* After setting max-softlimit, find the overclock max freq */
6650         if (IS_GEN6(dev_priv) ||
6651             IS_IVYBRIDGE(dev_priv) || IS_HASWELL(dev_priv)) {
6652                 u32 params = 0;
6653
6654                 sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &params);
6655                 if (params & BIT(31)) { /* OC supported */
6656                         DRM_DEBUG_DRIVER("Overclocking supported, max: %dMHz, overclock: %dMHz\n",
6657                                          (dev_priv->rps.max_freq & 0xff) * 50,
6658                                          (params & 0xff) * 50);
6659                         dev_priv->rps.max_freq = params & 0xff;
6660                 }
6661         }
6662
6663         /* Finally allow us to boost to max by default */
6664         dev_priv->rps.boost_freq = dev_priv->rps.max_freq;
6665
6666         mutex_unlock(&dev_priv->rps.hw_lock);
6667         mutex_unlock(&dev_priv->drm.struct_mutex);
6668
6669         intel_autoenable_gt_powersave(dev_priv);
6670 }
6671
6672 void intel_cleanup_gt_powersave(struct drm_i915_private *dev_priv)
6673 {
6674         if (IS_VALLEYVIEW(dev_priv))
6675                 valleyview_cleanup_gt_powersave(dev_priv);
6676
6677         if (!i915.enable_rc6)
6678                 intel_runtime_pm_put(dev_priv);
6679 }
6680
6681 /**
6682  * intel_suspend_gt_powersave - suspend PM work and helper threads
6683  * @dev_priv: i915 device
6684  *
6685  * We don't want to disable RC6 or other features here, we just want
6686  * to make sure any work we've queued has finished and won't bother
6687  * us while we're suspended.
6688  */
6689 void intel_suspend_gt_powersave(struct drm_i915_private *dev_priv)
6690 {
6691         if (INTEL_GEN(dev_priv) < 6)
6692                 return;
6693
6694         if (cancel_delayed_work_sync(&dev_priv->rps.autoenable_work))
6695                 intel_runtime_pm_put(dev_priv);
6696
6697         /* gen6_rps_idle() will be called later to disable interrupts */
6698 }
6699
6700 void intel_sanitize_gt_powersave(struct drm_i915_private *dev_priv)
6701 {
6702         dev_priv->rps.enabled = true; /* force disabling */
6703         intel_disable_gt_powersave(dev_priv);
6704
6705         gen6_reset_rps_interrupts(dev_priv);
6706 }
6707
6708 void intel_disable_gt_powersave(struct drm_i915_private *dev_priv)
6709 {
6710         if (!READ_ONCE(dev_priv->rps.enabled))
6711                 return;
6712
6713         mutex_lock(&dev_priv->rps.hw_lock);
6714
6715         if (INTEL_GEN(dev_priv) >= 9) {
6716                 gen9_disable_rc6(dev_priv);
6717                 gen9_disable_rps(dev_priv);
6718         } else if (IS_CHERRYVIEW(dev_priv)) {
6719                 cherryview_disable_rps(dev_priv);
6720         } else if (IS_VALLEYVIEW(dev_priv)) {
6721                 valleyview_disable_rps(dev_priv);
6722         } else if (INTEL_GEN(dev_priv) >= 6) {
6723                 gen6_disable_rps(dev_priv);
6724         }  else if (IS_IRONLAKE_M(dev_priv)) {
6725                 ironlake_disable_drps(dev_priv);
6726         }
6727
6728         dev_priv->rps.enabled = false;
6729         mutex_unlock(&dev_priv->rps.hw_lock);
6730 }
6731
6732 void intel_enable_gt_powersave(struct drm_i915_private *dev_priv)
6733 {
6734         /* We shouldn't be disabling as we submit, so this should be less
6735          * racy than it appears!
6736          */
6737         if (READ_ONCE(dev_priv->rps.enabled))
6738                 return;
6739
6740         /* Powersaving is controlled by the host when inside a VM */
6741         if (intel_vgpu_active(dev_priv))
6742                 return;
6743
6744         mutex_lock(&dev_priv->rps.hw_lock);
6745
6746         if (IS_CHERRYVIEW(dev_priv)) {
6747                 cherryview_enable_rps(dev_priv);
6748         } else if (IS_VALLEYVIEW(dev_priv)) {
6749                 valleyview_enable_rps(dev_priv);
6750         } else if (INTEL_GEN(dev_priv) >= 9) {
6751                 gen9_enable_rc6(dev_priv);
6752                 gen9_enable_rps(dev_priv);
6753                 if (IS_SKYLAKE(dev_priv) || IS_KABYLAKE(dev_priv))
6754                         gen6_update_ring_freq(dev_priv);
6755         } else if (IS_BROADWELL(dev_priv)) {
6756                 gen8_enable_rps(dev_priv);
6757                 gen6_update_ring_freq(dev_priv);
6758         } else if (INTEL_GEN(dev_priv) >= 6) {
6759                 gen6_enable_rps(dev_priv);
6760                 gen6_update_ring_freq(dev_priv);
6761         } else if (IS_IRONLAKE_M(dev_priv)) {
6762                 ironlake_enable_drps(dev_priv);
6763                 intel_init_emon(dev_priv);
6764         }
6765
6766         WARN_ON(dev_priv->rps.max_freq < dev_priv->rps.min_freq);
6767         WARN_ON(dev_priv->rps.idle_freq > dev_priv->rps.max_freq);
6768
6769         WARN_ON(dev_priv->rps.efficient_freq < dev_priv->rps.min_freq);
6770         WARN_ON(dev_priv->rps.efficient_freq > dev_priv->rps.max_freq);
6771
6772         dev_priv->rps.enabled = true;
6773         mutex_unlock(&dev_priv->rps.hw_lock);
6774 }
6775
6776 static void __intel_autoenable_gt_powersave(struct work_struct *work)
6777 {
6778         struct drm_i915_private *dev_priv =
6779                 container_of(work, typeof(*dev_priv), rps.autoenable_work.work);
6780         struct intel_engine_cs *rcs;
6781         struct drm_i915_gem_request *req;
6782
6783         if (READ_ONCE(dev_priv->rps.enabled))
6784                 goto out;
6785
6786         rcs = dev_priv->engine[RCS];
6787         if (rcs->last_context)
6788                 goto out;
6789
6790         if (!rcs->init_context)
6791                 goto out;
6792
6793         mutex_lock(&dev_priv->drm.struct_mutex);
6794
6795         req = i915_gem_request_alloc(rcs, dev_priv->kernel_context);
6796         if (IS_ERR(req))
6797                 goto unlock;
6798
6799         if (!i915.enable_execlists && i915_switch_context(req) == 0)
6800                 rcs->init_context(req);
6801
6802         /* Mark the device busy, calling intel_enable_gt_powersave() */
6803         i915_add_request_no_flush(req);
6804
6805 unlock:
6806         mutex_unlock(&dev_priv->drm.struct_mutex);
6807 out:
6808         intel_runtime_pm_put(dev_priv);
6809 }
6810
6811 void intel_autoenable_gt_powersave(struct drm_i915_private *dev_priv)
6812 {
6813         if (READ_ONCE(dev_priv->rps.enabled))
6814                 return;
6815
6816         if (IS_IRONLAKE_M(dev_priv)) {
6817                 ironlake_enable_drps(dev_priv);
6818                 intel_init_emon(dev_priv);
6819         } else if (INTEL_INFO(dev_priv)->gen >= 6) {
6820                 /*
6821                  * PCU communication is slow and this doesn't need to be
6822                  * done at any specific time, so do this out of our fast path
6823                  * to make resume and init faster.
6824                  *
6825                  * We depend on the HW RC6 power context save/restore
6826                  * mechanism when entering D3 through runtime PM suspend. So
6827                  * disable RPM until RPS/RC6 is properly setup. We can only
6828                  * get here via the driver load/system resume/runtime resume
6829                  * paths, so the _noresume version is enough (and in case of
6830                  * runtime resume it's necessary).
6831                  */
6832                 if (queue_delayed_work(dev_priv->wq,
6833                                        &dev_priv->rps.autoenable_work,
6834                                        round_jiffies_up_relative(HZ)))
6835                         intel_runtime_pm_get_noresume(dev_priv);
6836         }
6837 }
6838
6839 static void ibx_init_clock_gating(struct drm_i915_private *dev_priv)
6840 {
6841         /*
6842          * On Ibex Peak and Cougar Point, we need to disable clock
6843          * gating for the panel power sequencer or it will fail to
6844          * start up when no ports are active.
6845          */
6846         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
6847 }
6848
6849 static void g4x_disable_trickle_feed(struct drm_i915_private *dev_priv)
6850 {
6851         enum pipe pipe;
6852
6853         for_each_pipe(dev_priv, pipe) {
6854                 I915_WRITE(DSPCNTR(pipe),
6855                            I915_READ(DSPCNTR(pipe)) |
6856                            DISPPLANE_TRICKLE_FEED_DISABLE);
6857
6858                 I915_WRITE(DSPSURF(pipe), I915_READ(DSPSURF(pipe)));
6859                 POSTING_READ(DSPSURF(pipe));
6860         }
6861 }
6862
6863 static void ilk_init_lp_watermarks(struct drm_i915_private *dev_priv)
6864 {
6865         I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
6866         I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
6867         I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
6868
6869         /*
6870          * Don't touch WM1S_LP_EN here.
6871          * Doing so could cause underruns.
6872          */
6873 }
6874
6875 static void ironlake_init_clock_gating(struct drm_i915_private *dev_priv)
6876 {
6877         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6878
6879         /*
6880          * Required for FBC
6881          * WaFbcDisableDpfcClockGating:ilk
6882          */
6883         dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
6884                    ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
6885                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
6886
6887         I915_WRITE(PCH_3DCGDIS0,
6888                    MARIUNIT_CLOCK_GATE_DISABLE |
6889                    SVSMUNIT_CLOCK_GATE_DISABLE);
6890         I915_WRITE(PCH_3DCGDIS1,
6891                    VFMUNIT_CLOCK_GATE_DISABLE);
6892
6893         /*
6894          * According to the spec the following bits should be set in
6895          * order to enable memory self-refresh
6896          * The bit 22/21 of 0x42004
6897          * The bit 5 of 0x42020
6898          * The bit 15 of 0x45000
6899          */
6900         I915_WRITE(ILK_DISPLAY_CHICKEN2,
6901                    (I915_READ(ILK_DISPLAY_CHICKEN2) |
6902                     ILK_DPARB_GATE | ILK_VSDPFD_FULL));
6903         dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
6904         I915_WRITE(DISP_ARB_CTL,
6905                    (I915_READ(DISP_ARB_CTL) |
6906                     DISP_FBC_WM_DIS));
6907
6908         ilk_init_lp_watermarks(dev_priv);
6909
6910         /*
6911          * Based on the document from hardware guys the following bits
6912          * should be set unconditionally in order to enable FBC.
6913          * The bit 22 of 0x42000
6914          * The bit 22 of 0x42004
6915          * The bit 7,8,9 of 0x42020.
6916          */
6917         if (IS_IRONLAKE_M(dev_priv)) {
6918                 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
6919                 I915_WRITE(ILK_DISPLAY_CHICKEN1,
6920                            I915_READ(ILK_DISPLAY_CHICKEN1) |
6921                            ILK_FBCQ_DIS);
6922                 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6923                            I915_READ(ILK_DISPLAY_CHICKEN2) |
6924                            ILK_DPARB_GATE);
6925         }
6926
6927         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
6928
6929         I915_WRITE(ILK_DISPLAY_CHICKEN2,
6930                    I915_READ(ILK_DISPLAY_CHICKEN2) |
6931                    ILK_ELPIN_409_SELECT);
6932         I915_WRITE(_3D_CHICKEN2,
6933                    _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
6934                    _3D_CHICKEN2_WM_READ_PIPELINED);
6935
6936         /* WaDisableRenderCachePipelinedFlush:ilk */
6937         I915_WRITE(CACHE_MODE_0,
6938                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
6939
6940         /* WaDisable_RenderCache_OperationalFlush:ilk */
6941         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6942
6943         g4x_disable_trickle_feed(dev_priv);
6944
6945         ibx_init_clock_gating(dev_priv);
6946 }
6947
6948 static void cpt_init_clock_gating(struct drm_i915_private *dev_priv)
6949 {
6950         int pipe;
6951         uint32_t val;
6952
6953         /*
6954          * On Ibex Peak and Cougar Point, we need to disable clock
6955          * gating for the panel power sequencer or it will fail to
6956          * start up when no ports are active.
6957          */
6958         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
6959                    PCH_DPLUNIT_CLOCK_GATE_DISABLE |
6960                    PCH_CPUNIT_CLOCK_GATE_DISABLE);
6961         I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
6962                    DPLS_EDP_PPS_FIX_DIS);
6963         /* The below fixes the weird display corruption, a few pixels shifted
6964          * downward, on (only) LVDS of some HP laptops with IVY.
6965          */
6966         for_each_pipe(dev_priv, pipe) {
6967                 val = I915_READ(TRANS_CHICKEN2(pipe));
6968                 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
6969                 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6970                 if (dev_priv->vbt.fdi_rx_polarity_inverted)
6971                         val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6972                 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
6973                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
6974                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
6975                 I915_WRITE(TRANS_CHICKEN2(pipe), val);
6976         }
6977         /* WADP0ClockGatingDisable */
6978         for_each_pipe(dev_priv, pipe) {
6979                 I915_WRITE(TRANS_CHICKEN1(pipe),
6980                            TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
6981         }
6982 }
6983
6984 static void gen6_check_mch_setup(struct drm_i915_private *dev_priv)
6985 {
6986         uint32_t tmp;
6987
6988         tmp = I915_READ(MCH_SSKPD);
6989         if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL)
6990                 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
6991                               tmp);
6992 }
6993
6994 static void gen6_init_clock_gating(struct drm_i915_private *dev_priv)
6995 {
6996         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6997
6998         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
6999
7000         I915_WRITE(ILK_DISPLAY_CHICKEN2,
7001                    I915_READ(ILK_DISPLAY_CHICKEN2) |
7002                    ILK_ELPIN_409_SELECT);
7003
7004         /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
7005         I915_WRITE(_3D_CHICKEN,
7006                    _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
7007
7008         /* WaDisable_RenderCache_OperationalFlush:snb */
7009         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7010
7011         /*
7012          * BSpec recoomends 8x4 when MSAA is used,
7013          * however in practice 16x4 seems fastest.
7014          *
7015          * Note that PS/WM thread counts depend on the WIZ hashing
7016          * disable bit, which we don't touch here, but it's good
7017          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7018          */
7019         I915_WRITE(GEN6_GT_MODE,
7020                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7021
7022         ilk_init_lp_watermarks(dev_priv);
7023
7024         I915_WRITE(CACHE_MODE_0,
7025                    _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
7026
7027         I915_WRITE(GEN6_UCGCTL1,
7028                    I915_READ(GEN6_UCGCTL1) |
7029                    GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
7030                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
7031
7032         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
7033          * gating disable must be set.  Failure to set it results in
7034          * flickering pixels due to Z write ordering failures after
7035          * some amount of runtime in the Mesa "fire" demo, and Unigine
7036          * Sanctuary and Tropics, and apparently anything else with
7037          * alpha test or pixel discard.
7038          *
7039          * According to the spec, bit 11 (RCCUNIT) must also be set,
7040          * but we didn't debug actual testcases to find it out.
7041          *
7042          * WaDisableRCCUnitClockGating:snb
7043          * WaDisableRCPBUnitClockGating:snb
7044          */
7045         I915_WRITE(GEN6_UCGCTL2,
7046                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
7047                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
7048
7049         /* WaStripsFansDisableFastClipPerformanceFix:snb */
7050         I915_WRITE(_3D_CHICKEN3,
7051                    _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
7052
7053         /*
7054          * Bspec says:
7055          * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
7056          * 3DSTATE_SF number of SF output attributes is more than 16."
7057          */
7058         I915_WRITE(_3D_CHICKEN3,
7059                    _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
7060
7061         /*
7062          * According to the spec the following bits should be
7063          * set in order to enable memory self-refresh and fbc:
7064          * The bit21 and bit22 of 0x42000
7065          * The bit21 and bit22 of 0x42004
7066          * The bit5 and bit7 of 0x42020
7067          * The bit14 of 0x70180
7068          * The bit14 of 0x71180
7069          *
7070          * WaFbcAsynchFlipDisableFbcQueue:snb
7071          */
7072         I915_WRITE(ILK_DISPLAY_CHICKEN1,
7073                    I915_READ(ILK_DISPLAY_CHICKEN1) |
7074                    ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
7075         I915_WRITE(ILK_DISPLAY_CHICKEN2,
7076                    I915_READ(ILK_DISPLAY_CHICKEN2) |
7077                    ILK_DPARB_GATE | ILK_VSDPFD_FULL);
7078         I915_WRITE(ILK_DSPCLK_GATE_D,
7079                    I915_READ(ILK_DSPCLK_GATE_D) |
7080                    ILK_DPARBUNIT_CLOCK_GATE_ENABLE  |
7081                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
7082
7083         g4x_disable_trickle_feed(dev_priv);
7084
7085         cpt_init_clock_gating(dev_priv);
7086
7087         gen6_check_mch_setup(dev_priv);
7088 }
7089
7090 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
7091 {
7092         uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
7093
7094         /*
7095          * WaVSThreadDispatchOverride:ivb,vlv
7096          *
7097          * This actually overrides the dispatch
7098          * mode for all thread types.
7099          */
7100         reg &= ~GEN7_FF_SCHED_MASK;
7101         reg |= GEN7_FF_TS_SCHED_HW;
7102         reg |= GEN7_FF_VS_SCHED_HW;
7103         reg |= GEN7_FF_DS_SCHED_HW;
7104
7105         I915_WRITE(GEN7_FF_THREAD_MODE, reg);
7106 }
7107
7108 static void lpt_init_clock_gating(struct drm_i915_private *dev_priv)
7109 {
7110         /*
7111          * TODO: this bit should only be enabled when really needed, then
7112          * disabled when not needed anymore in order to save power.
7113          */
7114         if (HAS_PCH_LPT_LP(dev_priv))
7115                 I915_WRITE(SOUTH_DSPCLK_GATE_D,
7116                            I915_READ(SOUTH_DSPCLK_GATE_D) |
7117                            PCH_LP_PARTITION_LEVEL_DISABLE);
7118
7119         /* WADPOClockGatingDisable:hsw */
7120         I915_WRITE(TRANS_CHICKEN1(PIPE_A),
7121                    I915_READ(TRANS_CHICKEN1(PIPE_A)) |
7122                    TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
7123 }
7124
7125 static void lpt_suspend_hw(struct drm_i915_private *dev_priv)
7126 {
7127         if (HAS_PCH_LPT_LP(dev_priv)) {
7128                 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
7129
7130                 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
7131                 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
7132         }
7133 }
7134
7135 static void gen8_set_l3sqc_credits(struct drm_i915_private *dev_priv,
7136                                    int general_prio_credits,
7137                                    int high_prio_credits)
7138 {
7139         u32 misccpctl;
7140
7141         /* WaTempDisableDOPClkGating:bdw */
7142         misccpctl = I915_READ(GEN7_MISCCPCTL);
7143         I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
7144
7145         I915_WRITE(GEN8_L3SQCREG1,
7146                    L3_GENERAL_PRIO_CREDITS(general_prio_credits) |
7147                    L3_HIGH_PRIO_CREDITS(high_prio_credits));
7148
7149         /*
7150          * Wait at least 100 clocks before re-enabling clock gating.
7151          * See the definition of L3SQCREG1 in BSpec.
7152          */
7153         POSTING_READ(GEN8_L3SQCREG1);
7154         udelay(1);
7155         I915_WRITE(GEN7_MISCCPCTL, misccpctl);
7156 }
7157
7158 static void kabylake_init_clock_gating(struct drm_i915_private *dev_priv)
7159 {
7160         gen9_init_clock_gating(dev_priv);
7161
7162         /* WaDisableSDEUnitClockGating:kbl */
7163         if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
7164                 I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7165                            GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7166
7167         /* WaDisableGamClockGating:kbl */
7168         if (IS_KBL_REVID(dev_priv, 0, KBL_REVID_B0))
7169                 I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
7170                            GEN6_GAMUNIT_CLOCK_GATE_DISABLE);
7171
7172         /* WaFbcNukeOnHostModify:kbl */
7173         I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
7174                    ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
7175 }
7176
7177 static void skylake_init_clock_gating(struct drm_i915_private *dev_priv)
7178 {
7179         gen9_init_clock_gating(dev_priv);
7180
7181         /* WAC6entrylatency:skl */
7182         I915_WRITE(FBC_LLC_READ_CTRL, I915_READ(FBC_LLC_READ_CTRL) |
7183                    FBC_LLC_FULLY_OPEN);
7184
7185         /* WaFbcNukeOnHostModify:skl */
7186         I915_WRITE(ILK_DPFC_CHICKEN, I915_READ(ILK_DPFC_CHICKEN) |
7187                    ILK_DPFC_NUKE_ON_ANY_MODIFICATION);
7188 }
7189
7190 static void broadwell_init_clock_gating(struct drm_i915_private *dev_priv)
7191 {
7192         enum pipe pipe;
7193
7194         ilk_init_lp_watermarks(dev_priv);
7195
7196         /* WaSwitchSolVfFArbitrationPriority:bdw */
7197         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
7198
7199         /* WaPsrDPAMaskVBlankInSRD:bdw */
7200         I915_WRITE(CHICKEN_PAR1_1,
7201                    I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
7202
7203         /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
7204         for_each_pipe(dev_priv, pipe) {
7205                 I915_WRITE(CHICKEN_PIPESL_1(pipe),
7206                            I915_READ(CHICKEN_PIPESL_1(pipe)) |
7207                            BDW_DPRS_MASK_VBLANK_SRD);
7208         }
7209
7210         /* WaVSRefCountFullforceMissDisable:bdw */
7211         /* WaDSRefCountFullforceMissDisable:bdw */
7212         I915_WRITE(GEN7_FF_THREAD_MODE,
7213                    I915_READ(GEN7_FF_THREAD_MODE) &
7214                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
7215
7216         I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
7217                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
7218
7219         /* WaDisableSDEUnitClockGating:bdw */
7220         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7221                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7222
7223         /* WaProgramL3SqcReg1Default:bdw */
7224         gen8_set_l3sqc_credits(dev_priv, 30, 2);
7225
7226         /*
7227          * WaGttCachingOffByDefault:bdw
7228          * GTT cache may not work with big pages, so if those
7229          * are ever enabled GTT cache may need to be disabled.
7230          */
7231         I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
7232
7233         /* WaKVMNotificationOnConfigChange:bdw */
7234         I915_WRITE(CHICKEN_PAR2_1, I915_READ(CHICKEN_PAR2_1)
7235                    | KVM_CONFIG_CHANGE_NOTIFICATION_SELECT);
7236
7237         lpt_init_clock_gating(dev_priv);
7238 }
7239
7240 static void haswell_init_clock_gating(struct drm_i915_private *dev_priv)
7241 {
7242         ilk_init_lp_watermarks(dev_priv);
7243
7244         /* L3 caching of data atomics doesn't work -- disable it. */
7245         I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
7246         I915_WRITE(HSW_ROW_CHICKEN3,
7247                    _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
7248
7249         /* This is required by WaCatErrorRejectionIssue:hsw */
7250         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7251                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7252                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7253
7254         /* WaVSRefCountFullforceMissDisable:hsw */
7255         I915_WRITE(GEN7_FF_THREAD_MODE,
7256                    I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
7257
7258         /* WaDisable_RenderCache_OperationalFlush:hsw */
7259         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7260
7261         /* enable HiZ Raw Stall Optimization */
7262         I915_WRITE(CACHE_MODE_0_GEN7,
7263                    _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
7264
7265         /* WaDisable4x2SubspanOptimization:hsw */
7266         I915_WRITE(CACHE_MODE_1,
7267                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7268
7269         /*
7270          * BSpec recommends 8x4 when MSAA is used,
7271          * however in practice 16x4 seems fastest.
7272          *
7273          * Note that PS/WM thread counts depend on the WIZ hashing
7274          * disable bit, which we don't touch here, but it's good
7275          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7276          */
7277         I915_WRITE(GEN7_GT_MODE,
7278                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7279
7280         /* WaSampleCChickenBitEnable:hsw */
7281         I915_WRITE(HALF_SLICE_CHICKEN3,
7282                    _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE));
7283
7284         /* WaSwitchSolVfFArbitrationPriority:hsw */
7285         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
7286
7287         /* WaRsPkgCStateDisplayPMReq:hsw */
7288         I915_WRITE(CHICKEN_PAR1_1,
7289                    I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
7290
7291         lpt_init_clock_gating(dev_priv);
7292 }
7293
7294 static void ivybridge_init_clock_gating(struct drm_i915_private *dev_priv)
7295 {
7296         uint32_t snpcr;
7297
7298         ilk_init_lp_watermarks(dev_priv);
7299
7300         I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
7301
7302         /* WaDisableEarlyCull:ivb */
7303         I915_WRITE(_3D_CHICKEN3,
7304                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
7305
7306         /* WaDisableBackToBackFlipFix:ivb */
7307         I915_WRITE(IVB_CHICKEN3,
7308                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7309                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
7310
7311         /* WaDisablePSDDualDispatchEnable:ivb */
7312         if (IS_IVB_GT1(dev_priv))
7313                 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
7314                            _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
7315
7316         /* WaDisable_RenderCache_OperationalFlush:ivb */
7317         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7318
7319         /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
7320         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
7321                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
7322
7323         /* WaApplyL3ControlAndL3ChickenMode:ivb */
7324         I915_WRITE(GEN7_L3CNTLREG1,
7325                         GEN7_WA_FOR_GEN7_L3_CONTROL);
7326         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
7327                    GEN7_WA_L3_CHICKEN_MODE);
7328         if (IS_IVB_GT1(dev_priv))
7329                 I915_WRITE(GEN7_ROW_CHICKEN2,
7330                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7331         else {
7332                 /* must write both registers */
7333                 I915_WRITE(GEN7_ROW_CHICKEN2,
7334                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7335                 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
7336                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7337         }
7338
7339         /* WaForceL3Serialization:ivb */
7340         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
7341                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
7342
7343         /*
7344          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7345          * This implements the WaDisableRCZUnitClockGating:ivb workaround.
7346          */
7347         I915_WRITE(GEN6_UCGCTL2,
7348                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7349
7350         /* This is required by WaCatErrorRejectionIssue:ivb */
7351         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7352                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7353                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7354
7355         g4x_disable_trickle_feed(dev_priv);
7356
7357         gen7_setup_fixed_func_scheduler(dev_priv);
7358
7359         if (0) { /* causes HiZ corruption on ivb:gt1 */
7360                 /* enable HiZ Raw Stall Optimization */
7361                 I915_WRITE(CACHE_MODE_0_GEN7,
7362                            _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
7363         }
7364
7365         /* WaDisable4x2SubspanOptimization:ivb */
7366         I915_WRITE(CACHE_MODE_1,
7367                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7368
7369         /*
7370          * BSpec recommends 8x4 when MSAA is used,
7371          * however in practice 16x4 seems fastest.
7372          *
7373          * Note that PS/WM thread counts depend on the WIZ hashing
7374          * disable bit, which we don't touch here, but it's good
7375          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7376          */
7377         I915_WRITE(GEN7_GT_MODE,
7378                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7379
7380         snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
7381         snpcr &= ~GEN6_MBC_SNPCR_MASK;
7382         snpcr |= GEN6_MBC_SNPCR_MED;
7383         I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
7384
7385         if (!HAS_PCH_NOP(dev_priv))
7386                 cpt_init_clock_gating(dev_priv);
7387
7388         gen6_check_mch_setup(dev_priv);
7389 }
7390
7391 static void valleyview_init_clock_gating(struct drm_i915_private *dev_priv)
7392 {
7393         /* WaDisableEarlyCull:vlv */
7394         I915_WRITE(_3D_CHICKEN3,
7395                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
7396
7397         /* WaDisableBackToBackFlipFix:vlv */
7398         I915_WRITE(IVB_CHICKEN3,
7399                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
7400                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
7401
7402         /* WaPsdDispatchEnable:vlv */
7403         /* WaDisablePSDDualDispatchEnable:vlv */
7404         I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
7405                    _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
7406                                       GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
7407
7408         /* WaDisable_RenderCache_OperationalFlush:vlv */
7409         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7410
7411         /* WaForceL3Serialization:vlv */
7412         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
7413                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
7414
7415         /* WaDisableDopClockGating:vlv */
7416         I915_WRITE(GEN7_ROW_CHICKEN2,
7417                    _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
7418
7419         /* This is required by WaCatErrorRejectionIssue:vlv */
7420         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
7421                    I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
7422                    GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
7423
7424         gen7_setup_fixed_func_scheduler(dev_priv);
7425
7426         /*
7427          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
7428          * This implements the WaDisableRCZUnitClockGating:vlv workaround.
7429          */
7430         I915_WRITE(GEN6_UCGCTL2,
7431                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
7432
7433         /* WaDisableL3Bank2xClockGate:vlv
7434          * Disabling L3 clock gating- MMIO 940c[25] = 1
7435          * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
7436         I915_WRITE(GEN7_UCGCTL4,
7437                    I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
7438
7439         /*
7440          * BSpec says this must be set, even though
7441          * WaDisable4x2SubspanOptimization isn't listed for VLV.
7442          */
7443         I915_WRITE(CACHE_MODE_1,
7444                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
7445
7446         /*
7447          * BSpec recommends 8x4 when MSAA is used,
7448          * however in practice 16x4 seems fastest.
7449          *
7450          * Note that PS/WM thread counts depend on the WIZ hashing
7451          * disable bit, which we don't touch here, but it's good
7452          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
7453          */
7454         I915_WRITE(GEN7_GT_MODE,
7455                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
7456
7457         /*
7458          * WaIncreaseL3CreditsForVLVB0:vlv
7459          * This is the hardware default actually.
7460          */
7461         I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
7462
7463         /*
7464          * WaDisableVLVClockGating_VBIIssue:vlv
7465          * Disable clock gating on th GCFG unit to prevent a delay
7466          * in the reporting of vblank events.
7467          */
7468         I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
7469 }
7470
7471 static void cherryview_init_clock_gating(struct drm_i915_private *dev_priv)
7472 {
7473         /* WaVSRefCountFullforceMissDisable:chv */
7474         /* WaDSRefCountFullforceMissDisable:chv */
7475         I915_WRITE(GEN7_FF_THREAD_MODE,
7476                    I915_READ(GEN7_FF_THREAD_MODE) &
7477                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
7478
7479         /* WaDisableSemaphoreAndSyncFlipWait:chv */
7480         I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
7481                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
7482
7483         /* WaDisableCSUnitClockGating:chv */
7484         I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
7485                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
7486
7487         /* WaDisableSDEUnitClockGating:chv */
7488         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
7489                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
7490
7491         /*
7492          * WaProgramL3SqcReg1Default:chv
7493          * See gfxspecs/Related Documents/Performance Guide/
7494          * LSQC Setting Recommendations.
7495          */
7496         gen8_set_l3sqc_credits(dev_priv, 38, 2);
7497
7498         /*
7499          * GTT cache may not work with big pages, so if those
7500          * are ever enabled GTT cache may need to be disabled.
7501          */
7502         I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
7503 }
7504
7505 static void g4x_init_clock_gating(struct drm_i915_private *dev_priv)
7506 {
7507         uint32_t dspclk_gate;
7508
7509         I915_WRITE(RENCLK_GATE_D1, 0);
7510         I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
7511                    GS_UNIT_CLOCK_GATE_DISABLE |
7512                    CL_UNIT_CLOCK_GATE_DISABLE);
7513         I915_WRITE(RAMCLK_GATE_D, 0);
7514         dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
7515                 OVRUNIT_CLOCK_GATE_DISABLE |
7516                 OVCUNIT_CLOCK_GATE_DISABLE;
7517         if (IS_GM45(dev_priv))
7518                 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
7519         I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
7520
7521         /* WaDisableRenderCachePipelinedFlush */
7522         I915_WRITE(CACHE_MODE_0,
7523                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
7524
7525         /* WaDisable_RenderCache_OperationalFlush:g4x */
7526         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7527
7528         g4x_disable_trickle_feed(dev_priv);
7529 }
7530
7531 static void crestline_init_clock_gating(struct drm_i915_private *dev_priv)
7532 {
7533         I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
7534         I915_WRITE(RENCLK_GATE_D2, 0);
7535         I915_WRITE(DSPCLK_GATE_D, 0);
7536         I915_WRITE(RAMCLK_GATE_D, 0);
7537         I915_WRITE16(DEUC, 0);
7538         I915_WRITE(MI_ARB_STATE,
7539                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7540
7541         /* WaDisable_RenderCache_OperationalFlush:gen4 */
7542         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7543 }
7544
7545 static void broadwater_init_clock_gating(struct drm_i915_private *dev_priv)
7546 {
7547         I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
7548                    I965_RCC_CLOCK_GATE_DISABLE |
7549                    I965_RCPB_CLOCK_GATE_DISABLE |
7550                    I965_ISC_CLOCK_GATE_DISABLE |
7551                    I965_FBC_CLOCK_GATE_DISABLE);
7552         I915_WRITE(RENCLK_GATE_D2, 0);
7553         I915_WRITE(MI_ARB_STATE,
7554                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7555
7556         /* WaDisable_RenderCache_OperationalFlush:gen4 */
7557         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
7558 }
7559
7560 static void gen3_init_clock_gating(struct drm_i915_private *dev_priv)
7561 {
7562         u32 dstate = I915_READ(D_STATE);
7563
7564         dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
7565                 DSTATE_DOT_CLOCK_GATING;
7566         I915_WRITE(D_STATE, dstate);
7567
7568         if (IS_PINEVIEW(dev_priv))
7569                 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
7570
7571         /* IIR "flip pending" means done if this bit is set */
7572         I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
7573
7574         /* interrupts should cause a wake up from C3 */
7575         I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
7576
7577         /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
7578         I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
7579
7580         I915_WRITE(MI_ARB_STATE,
7581                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
7582 }
7583
7584 static void i85x_init_clock_gating(struct drm_i915_private *dev_priv)
7585 {
7586         I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
7587
7588         /* interrupts should cause a wake up from C3 */
7589         I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
7590                    _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
7591
7592         I915_WRITE(MEM_MODE,
7593                    _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE));
7594 }
7595
7596 static void i830_init_clock_gating(struct drm_i915_private *dev_priv)
7597 {
7598         I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
7599
7600         I915_WRITE(MEM_MODE,
7601                    _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE) |
7602                    _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE));
7603 }
7604
7605 void intel_init_clock_gating(struct drm_i915_private *dev_priv)
7606 {
7607         dev_priv->display.init_clock_gating(dev_priv);
7608 }
7609
7610 void intel_suspend_hw(struct drm_i915_private *dev_priv)
7611 {
7612         if (HAS_PCH_LPT(dev_priv))
7613                 lpt_suspend_hw(dev_priv);
7614 }
7615
7616 static void nop_init_clock_gating(struct drm_i915_private *dev_priv)
7617 {
7618         DRM_DEBUG_KMS("No clock gating settings or workarounds applied.\n");
7619 }
7620
7621 /**
7622  * intel_init_clock_gating_hooks - setup the clock gating hooks
7623  * @dev_priv: device private
7624  *
7625  * Setup the hooks that configure which clocks of a given platform can be
7626  * gated and also apply various GT and display specific workarounds for these
7627  * platforms. Note that some GT specific workarounds are applied separately
7628  * when GPU contexts or batchbuffers start their execution.
7629  */
7630 void intel_init_clock_gating_hooks(struct drm_i915_private *dev_priv)
7631 {
7632         if (IS_SKYLAKE(dev_priv))
7633                 dev_priv->display.init_clock_gating = skylake_init_clock_gating;
7634         else if (IS_KABYLAKE(dev_priv))
7635                 dev_priv->display.init_clock_gating = kabylake_init_clock_gating;
7636         else if (IS_BROXTON(dev_priv))
7637                 dev_priv->display.init_clock_gating = bxt_init_clock_gating;
7638         else if (IS_BROADWELL(dev_priv))
7639                 dev_priv->display.init_clock_gating = broadwell_init_clock_gating;
7640         else if (IS_CHERRYVIEW(dev_priv))
7641                 dev_priv->display.init_clock_gating = cherryview_init_clock_gating;
7642         else if (IS_HASWELL(dev_priv))
7643                 dev_priv->display.init_clock_gating = haswell_init_clock_gating;
7644         else if (IS_IVYBRIDGE(dev_priv))
7645                 dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
7646         else if (IS_VALLEYVIEW(dev_priv))
7647                 dev_priv->display.init_clock_gating = valleyview_init_clock_gating;
7648         else if (IS_GEN6(dev_priv))
7649                 dev_priv->display.init_clock_gating = gen6_init_clock_gating;
7650         else if (IS_GEN5(dev_priv))
7651                 dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
7652         else if (IS_G4X(dev_priv))
7653                 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
7654         else if (IS_CRESTLINE(dev_priv))
7655                 dev_priv->display.init_clock_gating = crestline_init_clock_gating;
7656         else if (IS_BROADWATER(dev_priv))
7657                 dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
7658         else if (IS_GEN3(dev_priv))
7659                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
7660         else if (IS_I85X(dev_priv) || IS_I865G(dev_priv))
7661                 dev_priv->display.init_clock_gating = i85x_init_clock_gating;
7662         else if (IS_GEN2(dev_priv))
7663                 dev_priv->display.init_clock_gating = i830_init_clock_gating;
7664         else {
7665                 MISSING_CASE(INTEL_DEVID(dev_priv));
7666                 dev_priv->display.init_clock_gating = nop_init_clock_gating;
7667         }
7668 }
7669
7670 /* Set up chip specific power management-related functions */
7671 void intel_init_pm(struct drm_i915_private *dev_priv)
7672 {
7673         intel_fbc_init(dev_priv);
7674
7675         /* For cxsr */
7676         if (IS_PINEVIEW(dev_priv))
7677                 i915_pineview_get_mem_freq(dev_priv);
7678         else if (IS_GEN5(dev_priv))
7679                 i915_ironlake_get_mem_freq(dev_priv);
7680
7681         /* For FIFO watermark updates */
7682         if (INTEL_GEN(dev_priv) >= 9) {
7683                 skl_setup_wm_latency(dev_priv);
7684                 dev_priv->display.initial_watermarks = skl_initial_wm;
7685                 dev_priv->display.atomic_update_watermarks = skl_atomic_update_crtc_wm;
7686                 dev_priv->display.compute_global_watermarks = skl_compute_wm;
7687         } else if (HAS_PCH_SPLIT(dev_priv)) {
7688                 ilk_setup_wm_latency(dev_priv);
7689
7690                 if ((IS_GEN5(dev_priv) && dev_priv->wm.pri_latency[1] &&
7691                      dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
7692                     (!IS_GEN5(dev_priv) && dev_priv->wm.pri_latency[0] &&
7693                      dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
7694                         dev_priv->display.compute_pipe_wm = ilk_compute_pipe_wm;
7695                         dev_priv->display.compute_intermediate_wm =
7696                                 ilk_compute_intermediate_wm;
7697                         dev_priv->display.initial_watermarks =
7698                                 ilk_initial_watermarks;
7699                         dev_priv->display.optimize_watermarks =
7700                                 ilk_optimize_watermarks;
7701                 } else {
7702                         DRM_DEBUG_KMS("Failed to read display plane latency. "
7703                                       "Disable CxSR\n");
7704                 }
7705         } else if (IS_CHERRYVIEW(dev_priv)) {
7706                 vlv_setup_wm_latency(dev_priv);
7707                 dev_priv->display.update_wm = vlv_update_wm;
7708         } else if (IS_VALLEYVIEW(dev_priv)) {
7709                 vlv_setup_wm_latency(dev_priv);
7710                 dev_priv->display.update_wm = vlv_update_wm;
7711         } else if (IS_PINEVIEW(dev_priv)) {
7712                 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev_priv),
7713                                             dev_priv->is_ddr3,
7714                                             dev_priv->fsb_freq,
7715                                             dev_priv->mem_freq)) {
7716                         DRM_INFO("failed to find known CxSR latency "
7717                                  "(found ddr%s fsb freq %d, mem freq %d), "
7718                                  "disabling CxSR\n",
7719                                  (dev_priv->is_ddr3 == 1) ? "3" : "2",
7720                                  dev_priv->fsb_freq, dev_priv->mem_freq);
7721                         /* Disable CxSR and never update its watermark again */
7722                         intel_set_memory_cxsr(dev_priv, false);
7723                         dev_priv->display.update_wm = NULL;
7724                 } else
7725                         dev_priv->display.update_wm = pineview_update_wm;
7726         } else if (IS_G4X(dev_priv)) {
7727                 dev_priv->display.update_wm = g4x_update_wm;
7728         } else if (IS_GEN4(dev_priv)) {
7729                 dev_priv->display.update_wm = i965_update_wm;
7730         } else if (IS_GEN3(dev_priv)) {
7731                 dev_priv->display.update_wm = i9xx_update_wm;
7732                 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
7733         } else if (IS_GEN2(dev_priv)) {
7734                 if (INTEL_INFO(dev_priv)->num_pipes == 1) {
7735                         dev_priv->display.update_wm = i845_update_wm;
7736                         dev_priv->display.get_fifo_size = i845_get_fifo_size;
7737                 } else {
7738                         dev_priv->display.update_wm = i9xx_update_wm;
7739                         dev_priv->display.get_fifo_size = i830_get_fifo_size;
7740                 }
7741         } else {
7742                 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
7743         }
7744 }
7745
7746 static inline int gen6_check_mailbox_status(struct drm_i915_private *dev_priv)
7747 {
7748         uint32_t flags =
7749                 I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_ERROR_MASK;
7750
7751         switch (flags) {
7752         case GEN6_PCODE_SUCCESS:
7753                 return 0;
7754         case GEN6_PCODE_UNIMPLEMENTED_CMD:
7755         case GEN6_PCODE_ILLEGAL_CMD:
7756                 return -ENXIO;
7757         case GEN6_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
7758         case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
7759                 return -EOVERFLOW;
7760         case GEN6_PCODE_TIMEOUT:
7761                 return -ETIMEDOUT;
7762         default:
7763                 MISSING_CASE(flags)
7764                 return 0;
7765         }
7766 }
7767
7768 static inline int gen7_check_mailbox_status(struct drm_i915_private *dev_priv)
7769 {
7770         uint32_t flags =
7771                 I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_ERROR_MASK;
7772
7773         switch (flags) {
7774         case GEN6_PCODE_SUCCESS:
7775                 return 0;
7776         case GEN6_PCODE_ILLEGAL_CMD:
7777                 return -ENXIO;
7778         case GEN7_PCODE_TIMEOUT:
7779                 return -ETIMEDOUT;
7780         case GEN7_PCODE_ILLEGAL_DATA:
7781                 return -EINVAL;
7782         case GEN7_PCODE_MIN_FREQ_TABLE_GT_RATIO_OUT_OF_RANGE:
7783                 return -EOVERFLOW;
7784         default:
7785                 MISSING_CASE(flags);
7786                 return 0;
7787         }
7788 }
7789
7790 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u32 mbox, u32 *val)
7791 {
7792         int status;
7793
7794         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7795
7796         /* GEN6_PCODE_* are outside of the forcewake domain, we can
7797          * use te fw I915_READ variants to reduce the amount of work
7798          * required when reading/writing.
7799          */
7800
7801         if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
7802                 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
7803                 return -EAGAIN;
7804         }
7805
7806         I915_WRITE_FW(GEN6_PCODE_DATA, *val);
7807         I915_WRITE_FW(GEN6_PCODE_DATA1, 0);
7808         I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
7809
7810         if (intel_wait_for_register_fw(dev_priv,
7811                                        GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
7812                                        500)) {
7813                 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
7814                 return -ETIMEDOUT;
7815         }
7816
7817         *val = I915_READ_FW(GEN6_PCODE_DATA);
7818         I915_WRITE_FW(GEN6_PCODE_DATA, 0);
7819
7820         if (INTEL_GEN(dev_priv) > 6)
7821                 status = gen7_check_mailbox_status(dev_priv);
7822         else
7823                 status = gen6_check_mailbox_status(dev_priv);
7824
7825         if (status) {
7826                 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed: %d\n",
7827                                  status);
7828                 return status;
7829         }
7830
7831         return 0;
7832 }
7833
7834 int sandybridge_pcode_write(struct drm_i915_private *dev_priv,
7835                             u32 mbox, u32 val)
7836 {
7837         int status;
7838
7839         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7840
7841         /* GEN6_PCODE_* are outside of the forcewake domain, we can
7842          * use te fw I915_READ variants to reduce the amount of work
7843          * required when reading/writing.
7844          */
7845
7846         if (I915_READ_FW(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
7847                 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
7848                 return -EAGAIN;
7849         }
7850
7851         I915_WRITE_FW(GEN6_PCODE_DATA, val);
7852         I915_WRITE_FW(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
7853
7854         if (intel_wait_for_register_fw(dev_priv,
7855                                        GEN6_PCODE_MAILBOX, GEN6_PCODE_READY, 0,
7856                                        500)) {
7857                 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
7858                 return -ETIMEDOUT;
7859         }
7860
7861         I915_WRITE_FW(GEN6_PCODE_DATA, 0);
7862
7863         if (INTEL_GEN(dev_priv) > 6)
7864                 status = gen7_check_mailbox_status(dev_priv);
7865         else
7866                 status = gen6_check_mailbox_status(dev_priv);
7867
7868         if (status) {
7869                 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed: %d\n",
7870                                  status);
7871                 return status;
7872         }
7873
7874         return 0;
7875 }
7876
7877 static bool skl_pcode_try_request(struct drm_i915_private *dev_priv, u32 mbox,
7878                                   u32 request, u32 reply_mask, u32 reply,
7879                                   u32 *status)
7880 {
7881         u32 val = request;
7882
7883         *status = sandybridge_pcode_read(dev_priv, mbox, &val);
7884
7885         return *status || ((val & reply_mask) == reply);
7886 }
7887
7888 /**
7889  * skl_pcode_request - send PCODE request until acknowledgment
7890  * @dev_priv: device private
7891  * @mbox: PCODE mailbox ID the request is targeted for
7892  * @request: request ID
7893  * @reply_mask: mask used to check for request acknowledgment
7894  * @reply: value used to check for request acknowledgment
7895  * @timeout_base_ms: timeout for polling with preemption enabled
7896  *
7897  * Keep resending the @request to @mbox until PCODE acknowledges it, PCODE
7898  * reports an error or an overall timeout of @timeout_base_ms+10 ms expires.
7899  * The request is acknowledged once the PCODE reply dword equals @reply after
7900  * applying @reply_mask. Polling is first attempted with preemption enabled
7901  * for @timeout_base_ms and if this times out for another 10 ms with
7902  * preemption disabled.
7903  *
7904  * Returns 0 on success, %-ETIMEDOUT in case of a timeout, <0 in case of some
7905  * other error as reported by PCODE.
7906  */
7907 int skl_pcode_request(struct drm_i915_private *dev_priv, u32 mbox, u32 request,
7908                       u32 reply_mask, u32 reply, int timeout_base_ms)
7909 {
7910         u32 status;
7911         int ret;
7912
7913         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7914
7915 #define COND skl_pcode_try_request(dev_priv, mbox, request, reply_mask, reply, \
7916                                    &status)
7917
7918         /*
7919          * Prime the PCODE by doing a request first. Normally it guarantees
7920          * that a subsequent request, at most @timeout_base_ms later, succeeds.
7921          * _wait_for() doesn't guarantee when its passed condition is evaluated
7922          * first, so send the first request explicitly.
7923          */
7924         if (COND) {
7925                 ret = 0;
7926                 goto out;
7927         }
7928         ret = _wait_for(COND, timeout_base_ms * 1000, 10);
7929         if (!ret)
7930                 goto out;
7931
7932         /*
7933          * The above can time out if the number of requests was low (2 in the
7934          * worst case) _and_ PCODE was busy for some reason even after a
7935          * (queued) request and @timeout_base_ms delay. As a workaround retry
7936          * the poll with preemption disabled to maximize the number of
7937          * requests. Increase the timeout from @timeout_base_ms to 10ms to
7938          * account for interrupts that could reduce the number of these
7939          * requests.
7940          */
7941         DRM_DEBUG_KMS("PCODE timeout, retrying with preemption disabled\n");
7942         WARN_ON_ONCE(timeout_base_ms > 3);
7943         preempt_disable();
7944         ret = wait_for_atomic(COND, 10);
7945         preempt_enable();
7946
7947 out:
7948         return ret ? ret : status;
7949 #undef COND
7950 }
7951
7952 static int byt_gpu_freq(struct drm_i915_private *dev_priv, int val)
7953 {
7954         /*
7955          * N = val - 0xb7
7956          * Slow = Fast = GPLL ref * N
7957          */
7958         return DIV_ROUND_CLOSEST(dev_priv->rps.gpll_ref_freq * (val - 0xb7), 1000);
7959 }
7960
7961 static int byt_freq_opcode(struct drm_i915_private *dev_priv, int val)
7962 {
7963         return DIV_ROUND_CLOSEST(1000 * val, dev_priv->rps.gpll_ref_freq) + 0xb7;
7964 }
7965
7966 static int chv_gpu_freq(struct drm_i915_private *dev_priv, int val)
7967 {
7968         /*
7969          * N = val / 2
7970          * CU (slow) = CU2x (fast) / 2 = GPLL ref * N / 2
7971          */
7972         return DIV_ROUND_CLOSEST(dev_priv->rps.gpll_ref_freq * val, 2 * 2 * 1000);
7973 }
7974
7975 static int chv_freq_opcode(struct drm_i915_private *dev_priv, int val)
7976 {
7977         /* CHV needs even values */
7978         return DIV_ROUND_CLOSEST(2 * 1000 * val, dev_priv->rps.gpll_ref_freq) * 2;
7979 }
7980
7981 int intel_gpu_freq(struct drm_i915_private *dev_priv, int val)
7982 {
7983         if (IS_GEN9(dev_priv))
7984                 return DIV_ROUND_CLOSEST(val * GT_FREQUENCY_MULTIPLIER,
7985                                          GEN9_FREQ_SCALER);
7986         else if (IS_CHERRYVIEW(dev_priv))
7987                 return chv_gpu_freq(dev_priv, val);
7988         else if (IS_VALLEYVIEW(dev_priv))
7989                 return byt_gpu_freq(dev_priv, val);
7990         else
7991                 return val * GT_FREQUENCY_MULTIPLIER;
7992 }
7993
7994 int intel_freq_opcode(struct drm_i915_private *dev_priv, int val)
7995 {
7996         if (IS_GEN9(dev_priv))
7997                 return DIV_ROUND_CLOSEST(val * GEN9_FREQ_SCALER,
7998                                          GT_FREQUENCY_MULTIPLIER);
7999         else if (IS_CHERRYVIEW(dev_priv))
8000                 return chv_freq_opcode(dev_priv, val);
8001         else if (IS_VALLEYVIEW(dev_priv))
8002                 return byt_freq_opcode(dev_priv, val);
8003         else
8004                 return DIV_ROUND_CLOSEST(val, GT_FREQUENCY_MULTIPLIER);
8005 }
8006
8007 struct request_boost {
8008         struct work_struct work;
8009         struct drm_i915_gem_request *req;
8010 };
8011
8012 static void __intel_rps_boost_work(struct work_struct *work)
8013 {
8014         struct request_boost *boost = container_of(work, struct request_boost, work);
8015         struct drm_i915_gem_request *req = boost->req;
8016
8017         if (!i915_gem_request_completed(req))
8018                 gen6_rps_boost(req->i915, NULL, req->emitted_jiffies);
8019
8020         i915_gem_request_put(req);
8021         kfree(boost);
8022 }
8023
8024 void intel_queue_rps_boost_for_request(struct drm_i915_gem_request *req)
8025 {
8026         struct request_boost *boost;
8027
8028         if (req == NULL || INTEL_GEN(req->i915) < 6)
8029                 return;
8030
8031         if (i915_gem_request_completed(req))
8032                 return;
8033
8034         boost = kmalloc(sizeof(*boost), GFP_ATOMIC);
8035         if (boost == NULL)
8036                 return;
8037
8038         boost->req = i915_gem_request_get(req);
8039
8040         INIT_WORK(&boost->work, __intel_rps_boost_work);
8041         queue_work(req->i915->wq, &boost->work);
8042 }
8043
8044 void intel_pm_setup(struct drm_device *dev)
8045 {
8046         struct drm_i915_private *dev_priv = to_i915(dev);
8047
8048         mutex_init(&dev_priv->rps.hw_lock);
8049         spin_lock_init(&dev_priv->rps.client_lock);
8050
8051         INIT_DELAYED_WORK(&dev_priv->rps.autoenable_work,
8052                           __intel_autoenable_gt_powersave);
8053         INIT_LIST_HEAD(&dev_priv->rps.clients);
8054
8055         dev_priv->pm.suspended = false;
8056         atomic_set(&dev_priv->pm.wakeref_count, 0);
8057 }