drivers/gpu/drm/i915/intel_pm.c

   1 /*
   2  * Copyright © 2012 Intel Corporation
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  * Authors:
  24  *    Eugeni Dodonov <eugeni.dodonov@intel.com>
  25  *
  26  */
  27
  28 #include <linux/cpufreq.h>
  29 #include "i915_drv.h"
  30 #include "intel_drv.h"
  31 #include "../../../platform/x86/intel_ips.h"
  32 #include <linux/module.h>
  33
  34 /**
  35  * RC6 is a special power stage which allows the GPU to enter an very
  36  * low-voltage mode when idle, using down to 0V while at this stage.  This
  37  * stage is entered automatically when the GPU is idle when RC6 support is
  38  * enabled, and as soon as new workload arises GPU wakes up automatically as well.
  39  *
  40  * There are different RC6 modes available in Intel GPU, which differentiate
  41  * among each other with the latency required to enter and leave RC6 and
  42  * voltage consumed by the GPU in different states.
  43  *
  44  * The combination of the following flags define which states GPU is allowed
  45  * to enter, while RC6 is the normal RC6 state, RC6p is the deep RC6, and
  46  * RC6pp is deepest RC6. Their support by hardware varies according to the
  47  * GPU, BIOS, chipset and platform. RC6 is usually the safest one and the one
  48  * which brings the most power savings; deeper states save more power, but
  49  * require higher latency to switch to and wake up.
  50  */
  51 #define INTEL_RC6_ENABLE                        (1<<0)
  52 #define INTEL_RC6p_ENABLE                       (1<<1)
  53 #define INTEL_RC6pp_ENABLE                      (1<<2)
  54
  55 static void bxt_init_clock_gating(struct drm_device *dev)
  56 {
  57         struct drm_i915_private *dev_priv = dev->dev_private;
  58
  59         /* WaDisableSDEUnitClockGating:bxt */
  60         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
  61                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
  62
  63         /*
  64          * FIXME:
  65          * GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ applies on 3x6 GT SKUs only.
  66          */
  67         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
  68                    GEN8_HDCUNIT_CLOCK_GATE_DISABLE_HDCREQ);
  69
  70         /*
  71          * Wa: Backlight PWM may stop in the asserted state, causing backlight
  72          * to stay fully on.
  73          */
  74         if (IS_BXT_REVID(dev_priv, BXT_REVID_B0, REVID_FOREVER))
  75                 I915_WRITE(GEN9_CLKGATE_DIS_0, I915_READ(GEN9_CLKGATE_DIS_0) |
  76                            PWM1_GATING_DIS | PWM2_GATING_DIS);
  77 }
  78
  79 static void i915_pineview_get_mem_freq(struct drm_device *dev)
  80 {
  81         struct drm_i915_private *dev_priv = dev->dev_private;
  82         u32 tmp;
  83
  84         tmp = I915_READ(CLKCFG);
  85
  86         switch (tmp & CLKCFG_FSB_MASK) {
  87         case CLKCFG_FSB_533:
  88                 dev_priv->fsb_freq = 533; /* 133*4 */
  89                 break;
  90         case CLKCFG_FSB_800:
  91                 dev_priv->fsb_freq = 800; /* 200*4 */
  92                 break;
  93         case CLKCFG_FSB_667:
  94                 dev_priv->fsb_freq =  667; /* 167*4 */
  95                 break;
  96         case CLKCFG_FSB_400:
  97                 dev_priv->fsb_freq = 400; /* 100*4 */
  98                 break;
  99         }
 100
 101         switch (tmp & CLKCFG_MEM_MASK) {
 102         case CLKCFG_MEM_533:
 103                 dev_priv->mem_freq = 533;
 104                 break;
 105         case CLKCFG_MEM_667:
 106                 dev_priv->mem_freq = 667;
 107                 break;
 108         case CLKCFG_MEM_800:
 109                 dev_priv->mem_freq = 800;
 110                 break;
 111         }
 112
 113         /* detect pineview DDR3 setting */
 114         tmp = I915_READ(CSHRDDR3CTL);
 115         dev_priv->is_ddr3 = (tmp & CSHRDDR3CTL_DDR3) ? 1 : 0;
 116 }
 117
 118 static void i915_ironlake_get_mem_freq(struct drm_device *dev)
 119 {
 120         struct drm_i915_private *dev_priv = dev->dev_private;
 121         u16 ddrpll, csipll;
 122
 123         ddrpll = I915_READ16(DDRMPLL1);
 124         csipll = I915_READ16(CSIPLL0);
 125
 126         switch (ddrpll & 0xff) {
 127         case 0xc:
 128                 dev_priv->mem_freq = 800;
 129                 break;
 130         case 0x10:
 131                 dev_priv->mem_freq = 1066;
 132                 break;
 133         case 0x14:
 134                 dev_priv->mem_freq = 1333;
 135                 break;
 136         case 0x18:
 137                 dev_priv->mem_freq = 1600;
 138                 break;
 139         default:
 140                 DRM_DEBUG_DRIVER("unknown memory frequency 0x%02x\n",
 141                                  ddrpll & 0xff);
 142                 dev_priv->mem_freq = 0;
 143                 break;
 144         }
 145
 146         dev_priv->ips.r_t = dev_priv->mem_freq;
 147
 148         switch (csipll & 0x3ff) {
 149         case 0x00c:
 150                 dev_priv->fsb_freq = 3200;
 151                 break;
 152         case 0x00e:
 153                 dev_priv->fsb_freq = 3733;
 154                 break;
 155         case 0x010:
 156                 dev_priv->fsb_freq = 4266;
 157                 break;
 158         case 0x012:
 159                 dev_priv->fsb_freq = 4800;
 160                 break;
 161         case 0x014:
 162                 dev_priv->fsb_freq = 5333;
 163                 break;
 164         case 0x016:
 165                 dev_priv->fsb_freq = 5866;
 166                 break;
 167         case 0x018:
 168                 dev_priv->fsb_freq = 6400;
 169                 break;
 170         default:
 171                 DRM_DEBUG_DRIVER("unknown fsb frequency 0x%04x\n",
 172                                  csipll & 0x3ff);
 173                 dev_priv->fsb_freq = 0;
 174                 break;
 175         }
 176
 177         if (dev_priv->fsb_freq == 3200) {
 178                 dev_priv->ips.c_m = 0;
 179         } else if (dev_priv->fsb_freq > 3200 && dev_priv->fsb_freq <= 4800) {
 180                 dev_priv->ips.c_m = 1;
 181         } else {
 182                 dev_priv->ips.c_m = 2;
 183         }
 184 }
 185
 186 static const struct cxsr_latency cxsr_latency_table[] = {
 187         {1, 0, 800, 400, 3382, 33382, 3983, 33983},    /* DDR2-400 SC */
 188         {1, 0, 800, 667, 3354, 33354, 3807, 33807},    /* DDR2-667 SC */
 189         {1, 0, 800, 800, 3347, 33347, 3763, 33763},    /* DDR2-800 SC */
 190         {1, 1, 800, 667, 6420, 36420, 6873, 36873},    /* DDR3-667 SC */
 191         {1, 1, 800, 800, 5902, 35902, 6318, 36318},    /* DDR3-800 SC */
 192
 193         {1, 0, 667, 400, 3400, 33400, 4021, 34021},    /* DDR2-400 SC */
 194         {1, 0, 667, 667, 3372, 33372, 3845, 33845},    /* DDR2-667 SC */
 195         {1, 0, 667, 800, 3386, 33386, 3822, 33822},    /* DDR2-800 SC */
 196         {1, 1, 667, 667, 6438, 36438, 6911, 36911},    /* DDR3-667 SC */
 197         {1, 1, 667, 800, 5941, 35941, 6377, 36377},    /* DDR3-800 SC */
 198
 199         {1, 0, 400, 400, 3472, 33472, 4173, 34173},    /* DDR2-400 SC */
 200         {1, 0, 400, 667, 3443, 33443, 3996, 33996},    /* DDR2-667 SC */
 201         {1, 0, 400, 800, 3430, 33430, 3946, 33946},    /* DDR2-800 SC */
 202         {1, 1, 400, 667, 6509, 36509, 7062, 37062},    /* DDR3-667 SC */
 203         {1, 1, 400, 800, 5985, 35985, 6501, 36501},    /* DDR3-800 SC */
 204
 205         {0, 0, 800, 400, 3438, 33438, 4065, 34065},    /* DDR2-400 SC */
 206         {0, 0, 800, 667, 3410, 33410, 3889, 33889},    /* DDR2-667 SC */
 207         {0, 0, 800, 800, 3403, 33403, 3845, 33845},    /* DDR2-800 SC */
 208         {0, 1, 800, 667, 6476, 36476, 6955, 36955},    /* DDR3-667 SC */
 209         {0, 1, 800, 800, 5958, 35958, 6400, 36400},    /* DDR3-800 SC */
 210
 211         {0, 0, 667, 400, 3456, 33456, 4103, 34106},    /* DDR2-400 SC */
 212         {0, 0, 667, 667, 3428, 33428, 3927, 33927},    /* DDR2-667 SC */
 213         {0, 0, 667, 800, 3443, 33443, 3905, 33905},    /* DDR2-800 SC */
 214         {0, 1, 667, 667, 6494, 36494, 6993, 36993},    /* DDR3-667 SC */
 215         {0, 1, 667, 800, 5998, 35998, 6460, 36460},    /* DDR3-800 SC */
 216
 217         {0, 0, 400, 400, 3528, 33528, 4255, 34255},    /* DDR2-400 SC */
 218         {0, 0, 400, 667, 3500, 33500, 4079, 34079},    /* DDR2-667 SC */
 219         {0, 0, 400, 800, 3487, 33487, 4029, 34029},    /* DDR2-800 SC */
 220         {0, 1, 400, 667, 6566, 36566, 7145, 37145},    /* DDR3-667 SC */
 221         {0, 1, 400, 800, 6042, 36042, 6584, 36584},    /* DDR3-800 SC */
 222 };
 223
 224 static const struct cxsr_latency *intel_get_cxsr_latency(int is_desktop,
 225                                                          int is_ddr3,
 226                                                          int fsb,
 227                                                          int mem)
 228 {
 229         const struct cxsr_latency *latency;
 230         int i;
 231
 232         if (fsb == 0 || mem == 0)
 233                 return NULL;
 234
 235         for (i = 0; i < ARRAY_SIZE(cxsr_latency_table); i++) {
 236                 latency = &cxsr_latency_table[i];
 237                 if (is_desktop == latency->is_desktop &&
 238                     is_ddr3 == latency->is_ddr3 &&
 239                     fsb == latency->fsb_freq && mem == latency->mem_freq)
 240                         return latency;
 241         }
 242
 243         DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 244
 245         return NULL;
 246 }
 247
 248 static void chv_set_memory_dvfs(struct drm_i915_private *dev_priv, bool enable)
 249 {
 250         u32 val;
 251
 252         mutex_lock(&dev_priv->rps.hw_lock);
 253
 254         val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
 255         if (enable)
 256                 val &= ~FORCE_DDR_HIGH_FREQ;
 257         else
 258                 val |= FORCE_DDR_HIGH_FREQ;
 259         val &= ~FORCE_DDR_LOW_FREQ;
 260         val |= FORCE_DDR_FREQ_REQ_ACK;
 261         vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
 262
 263         if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
 264                       FORCE_DDR_FREQ_REQ_ACK) == 0, 3))
 265                 DRM_ERROR("timed out waiting for Punit DDR DVFS request\n");
 266
 267         mutex_unlock(&dev_priv->rps.hw_lock);
 268 }
 269
 270 static void chv_set_memory_pm5(struct drm_i915_private *dev_priv, bool enable)
 271 {
 272         u32 val;
 273
 274         mutex_lock(&dev_priv->rps.hw_lock);
 275
 276         val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
 277         if (enable)
 278                 val |= DSP_MAXFIFO_PM5_ENABLE;
 279         else
 280                 val &= ~DSP_MAXFIFO_PM5_ENABLE;
 281         vlv_punit_write(dev_priv, PUNIT_REG_DSPFREQ, val);
 282
 283         mutex_unlock(&dev_priv->rps.hw_lock);
 284 }
 285
 286 #define FW_WM(value, plane) \
 287         (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK)
 288
 289 void intel_set_memory_cxsr(struct drm_i915_private *dev_priv, bool enable)
 290 {
 291         struct drm_device *dev = dev_priv->dev;
 292         u32 val;
 293
 294         if (IS_VALLEYVIEW(dev) || IS_CHERRYVIEW(dev)) {
 295                 I915_WRITE(FW_BLC_SELF_VLV, enable ? FW_CSPWRDWNEN : 0);
 296                 POSTING_READ(FW_BLC_SELF_VLV);
 297                 dev_priv->wm.vlv.cxsr = enable;
 298         } else if (IS_G4X(dev) || IS_CRESTLINE(dev)) {
 299                 I915_WRITE(FW_BLC_SELF, enable ? FW_BLC_SELF_EN : 0);
 300                 POSTING_READ(FW_BLC_SELF);
 301         } else if (IS_PINEVIEW(dev)) {
 302                 val = I915_READ(DSPFW3) & ~PINEVIEW_SELF_REFRESH_EN;
 303                 val |= enable ? PINEVIEW_SELF_REFRESH_EN : 0;
 304                 I915_WRITE(DSPFW3, val);
 305                 POSTING_READ(DSPFW3);
 306         } else if (IS_I945G(dev) || IS_I945GM(dev)) {
 307                 val = enable ? _MASKED_BIT_ENABLE(FW_BLC_SELF_EN) :
 308                                _MASKED_BIT_DISABLE(FW_BLC_SELF_EN);
 309                 I915_WRITE(FW_BLC_SELF, val);
 310                 POSTING_READ(FW_BLC_SELF);
 311         } else if (IS_I915GM(dev)) {
 312                 val = enable ? _MASKED_BIT_ENABLE(INSTPM_SELF_EN) :
 313                                _MASKED_BIT_DISABLE(INSTPM_SELF_EN);
 314                 I915_WRITE(INSTPM, val);
 315                 POSTING_READ(INSTPM);
 316         } else {
 317                 return;
 318         }
 319
 320         DRM_DEBUG_KMS("memory self-refresh is %s\n",
 321                       enable ? "enabled" : "disabled");
 322 }
 323
 324
 325 /*
 326  * Latency for FIFO fetches is dependent on several factors:
 327  *   - memory configuration (speed, channels)
 328  *   - chipset
 329  *   - current MCH state
 330  * It can be fairly high in some situations, so here we assume a fairly
 331  * pessimal value.  It's a tradeoff between extra memory fetches (if we
 332  * set this value too high, the FIFO will fetch frequently to stay full)
 333  * and power consumption (set it too low to save power and we might see
 334  * FIFO underruns and display "flicker").
 335  *
 336  * A value of 5us seems to be a good balance; safe for very low end
 337  * platforms but not overly aggressive on lower latency configs.
 338  */
 339 static const int pessimal_latency_ns = 5000;
 340
 341 #define VLV_FIFO_START(dsparb, dsparb2, lo_shift, hi_shift) \
 342         ((((dsparb) >> (lo_shift)) & 0xff) | ((((dsparb2) >> (hi_shift)) & 0x1) << 8))
 343
 344 static int vlv_get_fifo_size(struct drm_device *dev,
 345                               enum pipe pipe, int plane)
 346 {
 347         struct drm_i915_private *dev_priv = dev->dev_private;
 348         int sprite0_start, sprite1_start, size;
 349
 350         switch (pipe) {
 351                 uint32_t dsparb, dsparb2, dsparb3;
 352         case PIPE_A:
 353                 dsparb = I915_READ(DSPARB);
 354                 dsparb2 = I915_READ(DSPARB2);
 355                 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 0, 0);
 356                 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 8, 4);
 357                 break;
 358         case PIPE_B:
 359                 dsparb = I915_READ(DSPARB);
 360                 dsparb2 = I915_READ(DSPARB2);
 361                 sprite0_start = VLV_FIFO_START(dsparb, dsparb2, 16, 8);
 362                 sprite1_start = VLV_FIFO_START(dsparb, dsparb2, 24, 12);
 363                 break;
 364         case PIPE_C:
 365                 dsparb2 = I915_READ(DSPARB2);
 366                 dsparb3 = I915_READ(DSPARB3);
 367                 sprite0_start = VLV_FIFO_START(dsparb3, dsparb2, 0, 16);
 368                 sprite1_start = VLV_FIFO_START(dsparb3, dsparb2, 8, 20);
 369                 break;
 370         default:
 371                 return 0;
 372         }
 373
 374         switch (plane) {
 375         case 0:
 376                 size = sprite0_start;
 377                 break;
 378         case 1:
 379                 size = sprite1_start - sprite0_start;
 380                 break;
 381         case 2:
 382                 size = 512 - 1 - sprite1_start;
 383                 break;
 384         default:
 385                 return 0;
 386         }
 387
 388         DRM_DEBUG_KMS("Pipe %c %s %c FIFO size: %d\n",
 389                       pipe_name(pipe), plane == 0 ? "primary" : "sprite",
 390                       plane == 0 ? plane_name(pipe) : sprite_name(pipe, plane - 1),
 391                       size);
 392
 393         return size;
 394 }
 395
 396 static int i9xx_get_fifo_size(struct drm_device *dev, int plane)
 397 {
 398         struct drm_i915_private *dev_priv = dev->dev_private;
 399         uint32_t dsparb = I915_READ(DSPARB);
 400         int size;
 401
 402         size = dsparb & 0x7f;
 403         if (plane)
 404                 size = ((dsparb >> DSPARB_CSTART_SHIFT) & 0x7f) - size;
 405
 406         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 407                       plane ? "B" : "A", size);
 408
 409         return size;
 410 }
 411
 412 static int i830_get_fifo_size(struct drm_device *dev, int plane)
 413 {
 414         struct drm_i915_private *dev_priv = dev->dev_private;
 415         uint32_t dsparb = I915_READ(DSPARB);
 416         int size;
 417
 418         size = dsparb & 0x1ff;
 419         if (plane)
 420                 size = ((dsparb >> DSPARB_BEND_SHIFT) & 0x1ff) - size;
 421         size >>= 1; /* Convert to cachelines */
 422
 423         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 424                       plane ? "B" : "A", size);
 425
 426         return size;
 427 }
 428
 429 static int i845_get_fifo_size(struct drm_device *dev, int plane)
 430 {
 431         struct drm_i915_private *dev_priv = dev->dev_private;
 432         uint32_t dsparb = I915_READ(DSPARB);
 433         int size;
 434
 435         size = dsparb & 0x7f;
 436         size >>= 2; /* Convert to cachelines */
 437
 438         DRM_DEBUG_KMS("FIFO size - (0x%08x) %s: %d\n", dsparb,
 439                       plane ? "B" : "A",
 440                       size);
 441
 442         return size;
 443 }
 444
 445 /* Pineview has different values for various configs */
 446 static const struct intel_watermark_params pineview_display_wm = {
 447         .fifo_size = PINEVIEW_DISPLAY_FIFO,
 448         .max_wm = PINEVIEW_MAX_WM,
 449         .default_wm = PINEVIEW_DFT_WM,
 450         .guard_size = PINEVIEW_GUARD_WM,
 451         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 452 };
 453 static const struct intel_watermark_params pineview_display_hplloff_wm = {
 454         .fifo_size = PINEVIEW_DISPLAY_FIFO,
 455         .max_wm = PINEVIEW_MAX_WM,
 456         .default_wm = PINEVIEW_DFT_HPLLOFF_WM,
 457         .guard_size = PINEVIEW_GUARD_WM,
 458         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 459 };
 460 static const struct intel_watermark_params pineview_cursor_wm = {
 461         .fifo_size = PINEVIEW_CURSOR_FIFO,
 462         .max_wm = PINEVIEW_CURSOR_MAX_WM,
 463         .default_wm = PINEVIEW_CURSOR_DFT_WM,
 464         .guard_size = PINEVIEW_CURSOR_GUARD_WM,
 465         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 466 };
 467 static const struct intel_watermark_params pineview_cursor_hplloff_wm = {
 468         .fifo_size = PINEVIEW_CURSOR_FIFO,
 469         .max_wm = PINEVIEW_CURSOR_MAX_WM,
 470         .default_wm = PINEVIEW_CURSOR_DFT_WM,
 471         .guard_size = PINEVIEW_CURSOR_GUARD_WM,
 472         .cacheline_size = PINEVIEW_FIFO_LINE_SIZE,
 473 };
 474 static const struct intel_watermark_params g4x_wm_info = {
 475         .fifo_size = G4X_FIFO_SIZE,
 476         .max_wm = G4X_MAX_WM,
 477         .default_wm = G4X_MAX_WM,
 478         .guard_size = 2,
 479         .cacheline_size = G4X_FIFO_LINE_SIZE,
 480 };
 481 static const struct intel_watermark_params g4x_cursor_wm_info = {
 482         .fifo_size = I965_CURSOR_FIFO,
 483         .max_wm = I965_CURSOR_MAX_WM,
 484         .default_wm = I965_CURSOR_DFT_WM,
 485         .guard_size = 2,
 486         .cacheline_size = G4X_FIFO_LINE_SIZE,
 487 };
 488 static const struct intel_watermark_params valleyview_wm_info = {
 489         .fifo_size = VALLEYVIEW_FIFO_SIZE,
 490         .max_wm = VALLEYVIEW_MAX_WM,
 491         .default_wm = VALLEYVIEW_MAX_WM,
 492         .guard_size = 2,
 493         .cacheline_size = G4X_FIFO_LINE_SIZE,
 494 };
 495 static const struct intel_watermark_params valleyview_cursor_wm_info = {
 496         .fifo_size = I965_CURSOR_FIFO,
 497         .max_wm = VALLEYVIEW_CURSOR_MAX_WM,
 498         .default_wm = I965_CURSOR_DFT_WM,
 499         .guard_size = 2,
 500         .cacheline_size = G4X_FIFO_LINE_SIZE,
 501 };
 502 static const struct intel_watermark_params i965_cursor_wm_info = {
 503         .fifo_size = I965_CURSOR_FIFO,
 504         .max_wm = I965_CURSOR_MAX_WM,
 505         .default_wm = I965_CURSOR_DFT_WM,
 506         .guard_size = 2,
 507         .cacheline_size = I915_FIFO_LINE_SIZE,
 508 };
 509 static const struct intel_watermark_params i945_wm_info = {
 510         .fifo_size = I945_FIFO_SIZE,
 511         .max_wm = I915_MAX_WM,
 512         .default_wm = 1,
 513         .guard_size = 2,
 514         .cacheline_size = I915_FIFO_LINE_SIZE,
 515 };
 516 static const struct intel_watermark_params i915_wm_info = {
 517         .fifo_size = I915_FIFO_SIZE,
 518         .max_wm = I915_MAX_WM,
 519         .default_wm = 1,
 520         .guard_size = 2,
 521         .cacheline_size = I915_FIFO_LINE_SIZE,
 522 };
 523 static const struct intel_watermark_params i830_a_wm_info = {
 524         .fifo_size = I855GM_FIFO_SIZE,
 525         .max_wm = I915_MAX_WM,
 526         .default_wm = 1,
 527         .guard_size = 2,
 528         .cacheline_size = I830_FIFO_LINE_SIZE,
 529 };
 530 static const struct intel_watermark_params i830_bc_wm_info = {
 531         .fifo_size = I855GM_FIFO_SIZE,
 532         .max_wm = I915_MAX_WM/2,
 533         .default_wm = 1,
 534         .guard_size = 2,
 535         .cacheline_size = I830_FIFO_LINE_SIZE,
 536 };
 537 static const struct intel_watermark_params i845_wm_info = {
 538         .fifo_size = I830_FIFO_SIZE,
 539         .max_wm = I915_MAX_WM,
 540         .default_wm = 1,
 541         .guard_size = 2,
 542         .cacheline_size = I830_FIFO_LINE_SIZE,
 543 };
 544
 545 /**
 546  * intel_calculate_wm - calculate watermark level
 547  * @clock_in_khz: pixel clock
 548  * @wm: chip FIFO params
 549  * @pixel_size: display pixel size
 550  * @latency_ns: memory latency for the platform
 551  *
 552  * Calculate the watermark level (the level at which the display plane will
 553  * start fetching from memory again).  Each chip has a different display
 554  * FIFO size and allocation, so the caller needs to figure that out and pass
 555  * in the correct intel_watermark_params structure.
 556  *
 557  * As the pixel clock runs, the FIFO will be drained at a rate that depends
 558  * on the pixel size.  When it reaches the watermark level, it'll start
 559  * fetching FIFO line sized based chunks from memory until the FIFO fills
 560  * past the watermark point.  If the FIFO drains completely, a FIFO underrun
 561  * will occur, and a display engine hang could result.
 562  */
 563 static unsigned long intel_calculate_wm(unsigned long clock_in_khz,
 564                                         const struct intel_watermark_params *wm,
 565                                         int fifo_size,
 566                                         int pixel_size,
 567                                         unsigned long latency_ns)
 568 {
 569         long entries_required, wm_size;
 570
 571         /*
 572          * Note: we need to make sure we don't overflow for various clock &
 573          * latency values.
 574          * clocks go from a few thousand to several hundred thousand.
 575          * latency is usually a few thousand
 576          */
 577         entries_required = ((clock_in_khz / 1000) * pixel_size * latency_ns) /
 578                 1000;
 579         entries_required = DIV_ROUND_UP(entries_required, wm->cacheline_size);
 580
 581         DRM_DEBUG_KMS("FIFO entries required for mode: %ld\n", entries_required);
 582
 583         wm_size = fifo_size - (entries_required + wm->guard_size);
 584
 585         DRM_DEBUG_KMS("FIFO watermark level: %ld\n", wm_size);
 586
 587         /* Don't promote wm_size to unsigned... */
 588         if (wm_size > (long)wm->max_wm)
 589                 wm_size = wm->max_wm;
 590         if (wm_size <= 0)
 591                 wm_size = wm->default_wm;
 592
 593         /*
 594          * Bspec seems to indicate that the value shouldn't be lower than
 595          * 'burst size + 1'. Certainly 830 is quite unhappy with low values.
 596          * Lets go for 8 which is the burst size since certain platforms
 597          * already use a hardcoded 8 (which is what the spec says should be
 598          * done).
 599          */
 600         if (wm_size <= 8)
 601                 wm_size = 8;
 602
 603         return wm_size;
 604 }
 605
 606 static struct drm_crtc *single_enabled_crtc(struct drm_device *dev)
 607 {
 608         struct drm_crtc *crtc, *enabled = NULL;
 609
 610         for_each_crtc(dev, crtc) {
 611                 if (intel_crtc_active(crtc)) {
 612                         if (enabled)
 613                                 return NULL;
 614                         enabled = crtc;
 615                 }
 616         }
 617
 618         return enabled;
 619 }
 620
 621 static void pineview_update_wm(struct drm_crtc *unused_crtc)
 622 {
 623         struct drm_device *dev = unused_crtc->dev;
 624         struct drm_i915_private *dev_priv = dev->dev_private;
 625         struct drm_crtc *crtc;
 626         const struct cxsr_latency *latency;
 627         u32 reg;
 628         unsigned long wm;
 629
 630         latency = intel_get_cxsr_latency(IS_PINEVIEW_G(dev), dev_priv->is_ddr3,
 631                                          dev_priv->fsb_freq, dev_priv->mem_freq);
 632         if (!latency) {
 633                 DRM_DEBUG_KMS("Unknown FSB/MEM found, disable CxSR\n");
 634                 intel_set_memory_cxsr(dev_priv, false);
 635                 return;
 636         }
 637
 638         crtc = single_enabled_crtc(dev);
 639         if (crtc) {
 640                 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
 641                 int pixel_size = crtc->primary->state->fb->bits_per_pixel / 8;
 642                 int clock = adjusted_mode->crtc_clock;
 643
 644                 /* Display SR */
 645                 wm = intel_calculate_wm(clock, &pineview_display_wm,
 646                                         pineview_display_wm.fifo_size,
 647                                         pixel_size, latency->display_sr);
 648                 reg = I915_READ(DSPFW1);
 649                 reg &= ~DSPFW_SR_MASK;
 650                 reg |= FW_WM(wm, SR);
 651                 I915_WRITE(DSPFW1, reg);
 652                 DRM_DEBUG_KMS("DSPFW1 register is %x\n", reg);
 653
 654                 /* cursor SR */
 655                 wm = intel_calculate_wm(clock, &pineview_cursor_wm,
 656                                         pineview_display_wm.fifo_size,
 657                                         pixel_size, latency->cursor_sr);
 658                 reg = I915_READ(DSPFW3);
 659                 reg &= ~DSPFW_CURSOR_SR_MASK;
 660                 reg |= FW_WM(wm, CURSOR_SR);
 661                 I915_WRITE(DSPFW3, reg);
 662
 663                 /* Display HPLL off SR */
 664                 wm = intel_calculate_wm(clock, &pineview_display_hplloff_wm,
 665                                         pineview_display_hplloff_wm.fifo_size,
 666                                         pixel_size, latency->display_hpll_disable);
 667                 reg = I915_READ(DSPFW3);
 668                 reg &= ~DSPFW_HPLL_SR_MASK;
 669                 reg |= FW_WM(wm, HPLL_SR);
 670                 I915_WRITE(DSPFW3, reg);
 671
 672                 /* cursor HPLL off SR */
 673                 wm = intel_calculate_wm(clock, &pineview_cursor_hplloff_wm,
 674                                         pineview_display_hplloff_wm.fifo_size,
 675                                         pixel_size, latency->cursor_hpll_disable);
 676                 reg = I915_READ(DSPFW3);
 677                 reg &= ~DSPFW_HPLL_CURSOR_MASK;
 678                 reg |= FW_WM(wm, HPLL_CURSOR);
 679                 I915_WRITE(DSPFW3, reg);
 680                 DRM_DEBUG_KMS("DSPFW3 register is %x\n", reg);
 681
 682                 intel_set_memory_cxsr(dev_priv, true);
 683         } else {
 684                 intel_set_memory_cxsr(dev_priv, false);
 685         }
 686 }
 687
 688 static bool g4x_compute_wm0(struct drm_device *dev,
 689                             int plane,
 690                             const struct intel_watermark_params *display,
 691                             int display_latency_ns,
 692                             const struct intel_watermark_params *cursor,
 693                             int cursor_latency_ns,
 694                             int *plane_wm,
 695                             int *cursor_wm)
 696 {
 697         struct drm_crtc *crtc;
 698         const struct drm_display_mode *adjusted_mode;
 699         int htotal, hdisplay, clock, pixel_size;
 700         int line_time_us, line_count;
 701         int entries, tlb_miss;
 702
 703         crtc = intel_get_crtc_for_plane(dev, plane);
 704         if (!intel_crtc_active(crtc)) {
 705                 *cursor_wm = cursor->guard_size;
 706                 *plane_wm = display->guard_size;
 707                 return false;
 708         }
 709
 710         adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
 711         clock = adjusted_mode->crtc_clock;
 712         htotal = adjusted_mode->crtc_htotal;
 713         hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
 714         pixel_size = crtc->primary->state->fb->bits_per_pixel / 8;
 715
 716         /* Use the small buffer method to calculate plane watermark */
 717         entries = ((clock * pixel_size / 1000) * display_latency_ns) / 1000;
 718         tlb_miss = display->fifo_size*display->cacheline_size - hdisplay * 8;
 719         if (tlb_miss > 0)
 720                 entries += tlb_miss;
 721         entries = DIV_ROUND_UP(entries, display->cacheline_size);
 722         *plane_wm = entries + display->guard_size;
 723         if (*plane_wm > (int)display->max_wm)
 724                 *plane_wm = display->max_wm;
 725
 726         /* Use the large buffer method to calculate cursor watermark */
 727         line_time_us = max(htotal * 1000 / clock, 1);
 728         line_count = (cursor_latency_ns / line_time_us + 1000) / 1000;
 729         entries = line_count * crtc->cursor->state->crtc_w * pixel_size;
 730         tlb_miss = cursor->fifo_size*cursor->cacheline_size - hdisplay * 8;
 731         if (tlb_miss > 0)
 732                 entries += tlb_miss;
 733         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
 734         *cursor_wm = entries + cursor->guard_size;
 735         if (*cursor_wm > (int)cursor->max_wm)
 736                 *cursor_wm = (int)cursor->max_wm;
 737
 738         return true;
 739 }
 740
 741 /*
 742  * Check the wm result.
 743  *
 744  * If any calculated watermark values is larger than the maximum value that
 745  * can be programmed into the associated watermark register, that watermark
 746  * must be disabled.
 747  */
 748 static bool g4x_check_srwm(struct drm_device *dev,
 749                            int display_wm, int cursor_wm,
 750                            const struct intel_watermark_params *display,
 751                            const struct intel_watermark_params *cursor)
 752 {
 753         DRM_DEBUG_KMS("SR watermark: display plane %d, cursor %d\n",
 754                       display_wm, cursor_wm);
 755
 756         if (display_wm > display->max_wm) {
 757                 DRM_DEBUG_KMS("display watermark is too large(%d/%ld), disabling\n",
 758                               display_wm, display->max_wm);
 759                 return false;
 760         }
 761
 762         if (cursor_wm > cursor->max_wm) {
 763                 DRM_DEBUG_KMS("cursor watermark is too large(%d/%ld), disabling\n",
 764                               cursor_wm, cursor->max_wm);
 765                 return false;
 766         }
 767
 768         if (!(display_wm || cursor_wm)) {
 769                 DRM_DEBUG_KMS("SR latency is 0, disabling\n");
 770                 return false;
 771         }
 772
 773         return true;
 774 }
 775
 776 static bool g4x_compute_srwm(struct drm_device *dev,
 777                              int plane,
 778                              int latency_ns,
 779                              const struct intel_watermark_params *display,
 780                              const struct intel_watermark_params *cursor,
 781                              int *display_wm, int *cursor_wm)
 782 {
 783         struct drm_crtc *crtc;
 784         const struct drm_display_mode *adjusted_mode;
 785         int hdisplay, htotal, pixel_size, clock;
 786         unsigned long line_time_us;
 787         int line_count, line_size;
 788         int small, large;
 789         int entries;
 790
 791         if (!latency_ns) {
 792                 *display_wm = *cursor_wm = 0;
 793                 return false;
 794         }
 795
 796         crtc = intel_get_crtc_for_plane(dev, plane);
 797         adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
 798         clock = adjusted_mode->crtc_clock;
 799         htotal = adjusted_mode->crtc_htotal;
 800         hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
 801         pixel_size = crtc->primary->state->fb->bits_per_pixel / 8;
 802
 803         line_time_us = max(htotal * 1000 / clock, 1);
 804         line_count = (latency_ns / line_time_us + 1000) / 1000;
 805         line_size = hdisplay * pixel_size;
 806
 807         /* Use the minimum of the small and large buffer method for primary */
 808         small = ((clock * pixel_size / 1000) * latency_ns) / 1000;
 809         large = line_count * line_size;
 810
 811         entries = DIV_ROUND_UP(min(small, large), display->cacheline_size);
 812         *display_wm = entries + display->guard_size;
 813
 814         /* calculate the self-refresh watermark for display cursor */
 815         entries = line_count * pixel_size * crtc->cursor->state->crtc_w;
 816         entries = DIV_ROUND_UP(entries, cursor->cacheline_size);
 817         *cursor_wm = entries + cursor->guard_size;
 818
 819         return g4x_check_srwm(dev,
 820                               *display_wm, *cursor_wm,
 821                               display, cursor);
 822 }
 823
 824 #define FW_WM_VLV(value, plane) \
 825         (((value) << DSPFW_ ## plane ## _SHIFT) & DSPFW_ ## plane ## _MASK_VLV)
 826
 827 static void vlv_write_wm_values(struct intel_crtc *crtc,
 828                                 const struct vlv_wm_values *wm)
 829 {
 830         struct drm_i915_private *dev_priv = to_i915(crtc->base.dev);
 831         enum pipe pipe = crtc->pipe;
 832
 833         I915_WRITE(VLV_DDL(pipe),
 834                    (wm->ddl[pipe].cursor << DDL_CURSOR_SHIFT) |
 835                    (wm->ddl[pipe].sprite[1] << DDL_SPRITE_SHIFT(1)) |
 836                    (wm->ddl[pipe].sprite[0] << DDL_SPRITE_SHIFT(0)) |
 837                    (wm->ddl[pipe].primary << DDL_PLANE_SHIFT));
 838
 839         I915_WRITE(DSPFW1,
 840                    FW_WM(wm->sr.plane, SR) |
 841                    FW_WM(wm->pipe[PIPE_B].cursor, CURSORB) |
 842                    FW_WM_VLV(wm->pipe[PIPE_B].primary, PLANEB) |
 843                    FW_WM_VLV(wm->pipe[PIPE_A].primary, PLANEA));
 844         I915_WRITE(DSPFW2,
 845                    FW_WM_VLV(wm->pipe[PIPE_A].sprite[1], SPRITEB) |
 846                    FW_WM(wm->pipe[PIPE_A].cursor, CURSORA) |
 847                    FW_WM_VLV(wm->pipe[PIPE_A].sprite[0], SPRITEA));
 848         I915_WRITE(DSPFW3,
 849                    FW_WM(wm->sr.cursor, CURSOR_SR));
 850
 851         if (IS_CHERRYVIEW(dev_priv)) {
 852                 I915_WRITE(DSPFW7_CHV,
 853                            FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
 854                            FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
 855                 I915_WRITE(DSPFW8_CHV,
 856                            FW_WM_VLV(wm->pipe[PIPE_C].sprite[1], SPRITEF) |
 857                            FW_WM_VLV(wm->pipe[PIPE_C].sprite[0], SPRITEE));
 858                 I915_WRITE(DSPFW9_CHV,
 859                            FW_WM_VLV(wm->pipe[PIPE_C].primary, PLANEC) |
 860                            FW_WM(wm->pipe[PIPE_C].cursor, CURSORC));
 861                 I915_WRITE(DSPHOWM,
 862                            FW_WM(wm->sr.plane >> 9, SR_HI) |
 863                            FW_WM(wm->pipe[PIPE_C].sprite[1] >> 8, SPRITEF_HI) |
 864                            FW_WM(wm->pipe[PIPE_C].sprite[0] >> 8, SPRITEE_HI) |
 865                            FW_WM(wm->pipe[PIPE_C].primary >> 8, PLANEC_HI) |
 866                            FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
 867                            FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
 868                            FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
 869                            FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
 870                            FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
 871                            FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
 872         } else {
 873                 I915_WRITE(DSPFW7,
 874                            FW_WM_VLV(wm->pipe[PIPE_B].sprite[1], SPRITED) |
 875                            FW_WM_VLV(wm->pipe[PIPE_B].sprite[0], SPRITEC));
 876                 I915_WRITE(DSPHOWM,
 877                            FW_WM(wm->sr.plane >> 9, SR_HI) |
 878                            FW_WM(wm->pipe[PIPE_B].sprite[1] >> 8, SPRITED_HI) |
 879                            FW_WM(wm->pipe[PIPE_B].sprite[0] >> 8, SPRITEC_HI) |
 880                            FW_WM(wm->pipe[PIPE_B].primary >> 8, PLANEB_HI) |
 881                            FW_WM(wm->pipe[PIPE_A].sprite[1] >> 8, SPRITEB_HI) |
 882                            FW_WM(wm->pipe[PIPE_A].sprite[0] >> 8, SPRITEA_HI) |
 883                            FW_WM(wm->pipe[PIPE_A].primary >> 8, PLANEA_HI));
 884         }
 885
 886         /* zero (unused) WM1 watermarks */
 887         I915_WRITE(DSPFW4, 0);
 888         I915_WRITE(DSPFW5, 0);
 889         I915_WRITE(DSPFW6, 0);
 890         I915_WRITE(DSPHOWM1, 0);
 891
 892         POSTING_READ(DSPFW1);
 893 }
 894
 895 #undef FW_WM_VLV
 896
 897 enum vlv_wm_level {
 898         VLV_WM_LEVEL_PM2,
 899         VLV_WM_LEVEL_PM5,
 900         VLV_WM_LEVEL_DDR_DVFS,
 901 };
 902
 903 /* latency must be in 0.1us units. */
 904 static unsigned int vlv_wm_method2(unsigned int pixel_rate,
 905                                    unsigned int pipe_htotal,
 906                                    unsigned int horiz_pixels,
 907                                    unsigned int bytes_per_pixel,
 908                                    unsigned int latency)
 909 {
 910         unsigned int ret;
 911
 912         ret = (latency * pixel_rate) / (pipe_htotal * 10000);
 913         ret = (ret + 1) * horiz_pixels * bytes_per_pixel;
 914         ret = DIV_ROUND_UP(ret, 64);
 915
 916         return ret;
 917 }
 918
 919 static void vlv_setup_wm_latency(struct drm_device *dev)
 920 {
 921         struct drm_i915_private *dev_priv = dev->dev_private;
 922
 923         /* all latencies in usec */
 924         dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM2] = 3;
 925
 926         dev_priv->wm.max_level = VLV_WM_LEVEL_PM2;
 927
 928         if (IS_CHERRYVIEW(dev_priv)) {
 929                 dev_priv->wm.pri_latency[VLV_WM_LEVEL_PM5] = 12;
 930                 dev_priv->wm.pri_latency[VLV_WM_LEVEL_DDR_DVFS] = 33;
 931
 932                 dev_priv->wm.max_level = VLV_WM_LEVEL_DDR_DVFS;
 933         }
 934 }
 935
 936 static uint16_t vlv_compute_wm_level(struct intel_plane *plane,
 937                                      struct intel_crtc *crtc,
 938                                      const struct intel_plane_state *state,
 939                                      int level)
 940 {
 941         struct drm_i915_private *dev_priv = to_i915(plane->base.dev);
 942         int clock, htotal, pixel_size, width, wm;
 943
 944         if (dev_priv->wm.pri_latency[level] == 0)
 945                 return USHRT_MAX;
 946
 947         if (!state->visible)
 948                 return 0;
 949
 950         pixel_size = drm_format_plane_cpp(state->base.fb->pixel_format, 0);
 951         clock = crtc->config->base.adjusted_mode.crtc_clock;
 952         htotal = crtc->config->base.adjusted_mode.crtc_htotal;
 953         width = crtc->config->pipe_src_w;
 954         if (WARN_ON(htotal == 0))
 955                 htotal = 1;
 956
 957         if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
 958                 /*
 959                  * FIXME the formula gives values that are
 960                  * too big for the cursor FIFO, and hence we
 961                  * would never be able to use cursors. For
 962                  * now just hardcode the watermark.
 963                  */
 964                 wm = 63;
 965         } else {
 966                 wm = vlv_wm_method2(clock, htotal, width, pixel_size,
 967                                     dev_priv->wm.pri_latency[level] * 10);
 968         }
 969
 970         return min_t(int, wm, USHRT_MAX);
 971 }
 972
 973 static void vlv_compute_fifo(struct intel_crtc *crtc)
 974 {
 975         struct drm_device *dev = crtc->base.dev;
 976         struct vlv_wm_state *wm_state = &crtc->wm_state;
 977         struct intel_plane *plane;
 978         unsigned int total_rate = 0;
 979         const int fifo_size = 512 - 1;
 980         int fifo_extra, fifo_left = fifo_size;
 981
 982         for_each_intel_plane_on_crtc(dev, crtc, plane) {
 983                 struct intel_plane_state *state =
 984                         to_intel_plane_state(plane->base.state);
 985
 986                 if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
 987                         continue;
 988
 989                 if (state->visible) {
 990                         wm_state->num_active_planes++;
 991                         total_rate += drm_format_plane_cpp(state->base.fb->pixel_format, 0);
 992                 }
 993         }
 994
 995         for_each_intel_plane_on_crtc(dev, crtc, plane) {
 996                 struct intel_plane_state *state =
 997                         to_intel_plane_state(plane->base.state);
 998                 unsigned int rate;
 999
1000                 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
1001                         plane->wm.fifo_size = 63;
1002                         continue;
1003                 }
1004
1005                 if (!state->visible) {
1006                         plane->wm.fifo_size = 0;
1007                         continue;
1008                 }
1009
1010                 rate = drm_format_plane_cpp(state->base.fb->pixel_format, 0);
1011                 plane->wm.fifo_size = fifo_size * rate / total_rate;
1012                 fifo_left -= plane->wm.fifo_size;
1013         }
1014
1015         fifo_extra = DIV_ROUND_UP(fifo_left, wm_state->num_active_planes ?: 1);
1016
1017         /* spread the remainder evenly */
1018         for_each_intel_plane_on_crtc(dev, crtc, plane) {
1019                 int plane_extra;
1020
1021                 if (fifo_left == 0)
1022                         break;
1023
1024                 if (plane->base.type == DRM_PLANE_TYPE_CURSOR)
1025                         continue;
1026
1027                 /* give it all to the first plane if none are active */
1028                 if (plane->wm.fifo_size == 0 &&
1029                     wm_state->num_active_planes)
1030                         continue;
1031
1032                 plane_extra = min(fifo_extra, fifo_left);
1033                 plane->wm.fifo_size += plane_extra;
1034                 fifo_left -= plane_extra;
1035         }
1036
1037         WARN_ON(fifo_left != 0);
1038 }
1039
1040 static void vlv_invert_wms(struct intel_crtc *crtc)
1041 {
1042         struct vlv_wm_state *wm_state = &crtc->wm_state;
1043         int level;
1044
1045         for (level = 0; level < wm_state->num_levels; level++) {
1046                 struct drm_device *dev = crtc->base.dev;
1047                 const int sr_fifo_size = INTEL_INFO(dev)->num_pipes * 512 - 1;
1048                 struct intel_plane *plane;
1049
1050                 wm_state->sr[level].plane = sr_fifo_size - wm_state->sr[level].plane;
1051                 wm_state->sr[level].cursor = 63 - wm_state->sr[level].cursor;
1052
1053                 for_each_intel_plane_on_crtc(dev, crtc, plane) {
1054                         switch (plane->base.type) {
1055                                 int sprite;
1056                         case DRM_PLANE_TYPE_CURSOR:
1057                                 wm_state->wm[level].cursor = plane->wm.fifo_size -
1058                                         wm_state->wm[level].cursor;
1059                                 break;
1060                         case DRM_PLANE_TYPE_PRIMARY:
1061                                 wm_state->wm[level].primary = plane->wm.fifo_size -
1062                                         wm_state->wm[level].primary;
1063                                 break;
1064                         case DRM_PLANE_TYPE_OVERLAY:
1065                                 sprite = plane->plane;
1066                                 wm_state->wm[level].sprite[sprite] = plane->wm.fifo_size -
1067                                         wm_state->wm[level].sprite[sprite];
1068                                 break;
1069                         }
1070                 }
1071         }
1072 }
1073
1074 static void vlv_compute_wm(struct intel_crtc *crtc)
1075 {
1076         struct drm_device *dev = crtc->base.dev;
1077         struct vlv_wm_state *wm_state = &crtc->wm_state;
1078         struct intel_plane *plane;
1079         int sr_fifo_size = INTEL_INFO(dev)->num_pipes * 512 - 1;
1080         int level;
1081
1082         memset(wm_state, 0, sizeof(*wm_state));
1083
1084         wm_state->cxsr = crtc->pipe != PIPE_C && crtc->wm.cxsr_allowed;
1085         wm_state->num_levels = to_i915(dev)->wm.max_level + 1;
1086
1087         wm_state->num_active_planes = 0;
1088
1089         vlv_compute_fifo(crtc);
1090
1091         if (wm_state->num_active_planes != 1)
1092                 wm_state->cxsr = false;
1093
1094         if (wm_state->cxsr) {
1095                 for (level = 0; level < wm_state->num_levels; level++) {
1096                         wm_state->sr[level].plane = sr_fifo_size;
1097                         wm_state->sr[level].cursor = 63;
1098                 }
1099         }
1100
1101         for_each_intel_plane_on_crtc(dev, crtc, plane) {
1102                 struct intel_plane_state *state =
1103                         to_intel_plane_state(plane->base.state);
1104
1105                 if (!state->visible)
1106                         continue;
1107
1108                 /* normal watermarks */
1109                 for (level = 0; level < wm_state->num_levels; level++) {
1110                         int wm = vlv_compute_wm_level(plane, crtc, state, level);
1111                         int max_wm = plane->base.type == DRM_PLANE_TYPE_CURSOR ? 63 : 511;
1112
1113                         /* hack */
1114                         if (WARN_ON(level == 0 && wm > max_wm))
1115                                 wm = max_wm;
1116
1117                         if (wm > plane->wm.fifo_size)
1118                                 break;
1119
1120                         switch (plane->base.type) {
1121                                 int sprite;
1122                         case DRM_PLANE_TYPE_CURSOR:
1123                                 wm_state->wm[level].cursor = wm;
1124                                 break;
1125                         case DRM_PLANE_TYPE_PRIMARY:
1126                                 wm_state->wm[level].primary = wm;
1127                                 break;
1128                         case DRM_PLANE_TYPE_OVERLAY:
1129                                 sprite = plane->plane;
1130                                 wm_state->wm[level].sprite[sprite] = wm;
1131                                 break;
1132                         }
1133                 }
1134
1135                 wm_state->num_levels = level;
1136
1137                 if (!wm_state->cxsr)
1138                         continue;
1139
1140                 /* maxfifo watermarks */
1141                 switch (plane->base.type) {
1142                         int sprite, level;
1143                 case DRM_PLANE_TYPE_CURSOR:
1144                         for (level = 0; level < wm_state->num_levels; level++)
1145                                 wm_state->sr[level].cursor =
1146                                         wm_state->wm[level].cursor;
1147                         break;
1148                 case DRM_PLANE_TYPE_PRIMARY:
1149                         for (level = 0; level < wm_state->num_levels; level++)
1150                                 wm_state->sr[level].plane =
1151                                         min(wm_state->sr[level].plane,
1152                                             wm_state->wm[level].primary);
1153                         break;
1154                 case DRM_PLANE_TYPE_OVERLAY:
1155                         sprite = plane->plane;
1156                         for (level = 0; level < wm_state->num_levels; level++)
1157                                 wm_state->sr[level].plane =
1158                                         min(wm_state->sr[level].plane,
1159                                             wm_state->wm[level].sprite[sprite]);
1160                         break;
1161                 }
1162         }
1163
1164         /* clear any (partially) filled invalid levels */
1165         for (level = wm_state->num_levels; level < to_i915(dev)->wm.max_level + 1; level++) {
1166                 memset(&wm_state->wm[level], 0, sizeof(wm_state->wm[level]));
1167                 memset(&wm_state->sr[level], 0, sizeof(wm_state->sr[level]));
1168         }
1169
1170         vlv_invert_wms(crtc);
1171 }
1172
1173 #define VLV_FIFO(plane, value) \
1174         (((value) << DSPARB_ ## plane ## _SHIFT_VLV) & DSPARB_ ## plane ## _MASK_VLV)
1175
1176 static void vlv_pipe_set_fifo_size(struct intel_crtc *crtc)
1177 {
1178         struct drm_device *dev = crtc->base.dev;
1179         struct drm_i915_private *dev_priv = to_i915(dev);
1180         struct intel_plane *plane;
1181         int sprite0_start = 0, sprite1_start = 0, fifo_size = 0;
1182
1183         for_each_intel_plane_on_crtc(dev, crtc, plane) {
1184                 if (plane->base.type == DRM_PLANE_TYPE_CURSOR) {
1185                         WARN_ON(plane->wm.fifo_size != 63);
1186                         continue;
1187                 }
1188
1189                 if (plane->base.type == DRM_PLANE_TYPE_PRIMARY)
1190                         sprite0_start = plane->wm.fifo_size;
1191                 else if (plane->plane == 0)
1192                         sprite1_start = sprite0_start + plane->wm.fifo_size;
1193                 else
1194                         fifo_size = sprite1_start + plane->wm.fifo_size;
1195         }
1196
1197         WARN_ON(fifo_size != 512 - 1);
1198
1199         DRM_DEBUG_KMS("Pipe %c FIFO split %d / %d / %d\n",
1200                       pipe_name(crtc->pipe), sprite0_start,
1201                       sprite1_start, fifo_size);
1202
1203         switch (crtc->pipe) {
1204                 uint32_t dsparb, dsparb2, dsparb3;
1205         case PIPE_A:
1206                 dsparb = I915_READ(DSPARB);
1207                 dsparb2 = I915_READ(DSPARB2);
1208
1209                 dsparb &= ~(VLV_FIFO(SPRITEA, 0xff) |
1210                             VLV_FIFO(SPRITEB, 0xff));
1211                 dsparb |= (VLV_FIFO(SPRITEA, sprite0_start) |
1212                            VLV_FIFO(SPRITEB, sprite1_start));
1213
1214                 dsparb2 &= ~(VLV_FIFO(SPRITEA_HI, 0x1) |
1215                              VLV_FIFO(SPRITEB_HI, 0x1));
1216                 dsparb2 |= (VLV_FIFO(SPRITEA_HI, sprite0_start >> 8) |
1217                            VLV_FIFO(SPRITEB_HI, sprite1_start >> 8));
1218
1219                 I915_WRITE(DSPARB, dsparb);
1220                 I915_WRITE(DSPARB2, dsparb2);
1221                 break;
1222         case PIPE_B:
1223                 dsparb = I915_READ(DSPARB);
1224                 dsparb2 = I915_READ(DSPARB2);
1225
1226                 dsparb &= ~(VLV_FIFO(SPRITEC, 0xff) |
1227                             VLV_FIFO(SPRITED, 0xff));
1228                 dsparb |= (VLV_FIFO(SPRITEC, sprite0_start) |
1229                            VLV_FIFO(SPRITED, sprite1_start));
1230
1231                 dsparb2 &= ~(VLV_FIFO(SPRITEC_HI, 0xff) |
1232                              VLV_FIFO(SPRITED_HI, 0xff));
1233                 dsparb2 |= (VLV_FIFO(SPRITEC_HI, sprite0_start >> 8) |
1234                            VLV_FIFO(SPRITED_HI, sprite1_start >> 8));
1235
1236                 I915_WRITE(DSPARB, dsparb);
1237                 I915_WRITE(DSPARB2, dsparb2);
1238                 break;
1239         case PIPE_C:
1240                 dsparb3 = I915_READ(DSPARB3);
1241                 dsparb2 = I915_READ(DSPARB2);
1242
1243                 dsparb3 &= ~(VLV_FIFO(SPRITEE, 0xff) |
1244                              VLV_FIFO(SPRITEF, 0xff));
1245                 dsparb3 |= (VLV_FIFO(SPRITEE, sprite0_start) |
1246                             VLV_FIFO(SPRITEF, sprite1_start));
1247
1248                 dsparb2 &= ~(VLV_FIFO(SPRITEE_HI, 0xff) |
1249                              VLV_FIFO(SPRITEF_HI, 0xff));
1250                 dsparb2 |= (VLV_FIFO(SPRITEE_HI, sprite0_start >> 8) |
1251                            VLV_FIFO(SPRITEF_HI, sprite1_start >> 8));
1252
1253                 I915_WRITE(DSPARB3, dsparb3);
1254                 I915_WRITE(DSPARB2, dsparb2);
1255                 break;
1256         default:
1257                 break;
1258         }
1259 }
1260
1261 #undef VLV_FIFO
1262
1263 static void vlv_merge_wm(struct drm_device *dev,
1264                          struct vlv_wm_values *wm)
1265 {
1266         struct intel_crtc *crtc;
1267         int num_active_crtcs = 0;
1268
1269         wm->level = to_i915(dev)->wm.max_level;
1270         wm->cxsr = true;
1271
1272         for_each_intel_crtc(dev, crtc) {
1273                 const struct vlv_wm_state *wm_state = &crtc->wm_state;
1274
1275                 if (!crtc->active)
1276                         continue;
1277
1278                 if (!wm_state->cxsr)
1279                         wm->cxsr = false;
1280
1281                 num_active_crtcs++;
1282                 wm->level = min_t(int, wm->level, wm_state->num_levels - 1);
1283         }
1284
1285         if (num_active_crtcs != 1)
1286                 wm->cxsr = false;
1287
1288         if (num_active_crtcs > 1)
1289                 wm->level = VLV_WM_LEVEL_PM2;
1290
1291         for_each_intel_crtc(dev, crtc) {
1292                 struct vlv_wm_state *wm_state = &crtc->wm_state;
1293                 enum pipe pipe = crtc->pipe;
1294
1295                 if (!crtc->active)
1296                         continue;
1297
1298                 wm->pipe[pipe] = wm_state->wm[wm->level];
1299                 if (wm->cxsr)
1300                         wm->sr = wm_state->sr[wm->level];
1301
1302                 wm->ddl[pipe].primary = DDL_PRECISION_HIGH | 2;
1303                 wm->ddl[pipe].sprite[0] = DDL_PRECISION_HIGH | 2;
1304                 wm->ddl[pipe].sprite[1] = DDL_PRECISION_HIGH | 2;
1305                 wm->ddl[pipe].cursor = DDL_PRECISION_HIGH | 2;
1306         }
1307 }
1308
1309 static void vlv_update_wm(struct drm_crtc *crtc)
1310 {
1311         struct drm_device *dev = crtc->dev;
1312         struct drm_i915_private *dev_priv = dev->dev_private;
1313         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
1314         enum pipe pipe = intel_crtc->pipe;
1315         struct vlv_wm_values wm = {};
1316
1317         vlv_compute_wm(intel_crtc);
1318         vlv_merge_wm(dev, &wm);
1319
1320         if (memcmp(&dev_priv->wm.vlv, &wm, sizeof(wm)) == 0) {
1321                 /* FIXME should be part of crtc atomic commit */
1322                 vlv_pipe_set_fifo_size(intel_crtc);
1323                 return;
1324         }
1325
1326         if (wm.level < VLV_WM_LEVEL_DDR_DVFS &&
1327             dev_priv->wm.vlv.level >= VLV_WM_LEVEL_DDR_DVFS)
1328                 chv_set_memory_dvfs(dev_priv, false);
1329
1330         if (wm.level < VLV_WM_LEVEL_PM5 &&
1331             dev_priv->wm.vlv.level >= VLV_WM_LEVEL_PM5)
1332                 chv_set_memory_pm5(dev_priv, false);
1333
1334         if (!wm.cxsr && dev_priv->wm.vlv.cxsr)
1335                 intel_set_memory_cxsr(dev_priv, false);
1336
1337         /* FIXME should be part of crtc atomic commit */
1338         vlv_pipe_set_fifo_size(intel_crtc);
1339
1340         vlv_write_wm_values(intel_crtc, &wm);
1341
1342         DRM_DEBUG_KMS("Setting FIFO watermarks - %c: plane=%d, cursor=%d, "
1343                       "sprite0=%d, sprite1=%d, SR: plane=%d, cursor=%d level=%d cxsr=%d\n",
1344                       pipe_name(pipe), wm.pipe[pipe].primary, wm.pipe[pipe].cursor,
1345                       wm.pipe[pipe].sprite[0], wm.pipe[pipe].sprite[1],
1346                       wm.sr.plane, wm.sr.cursor, wm.level, wm.cxsr);
1347
1348         if (wm.cxsr && !dev_priv->wm.vlv.cxsr)
1349                 intel_set_memory_cxsr(dev_priv, true);
1350
1351         if (wm.level >= VLV_WM_LEVEL_PM5 &&
1352             dev_priv->wm.vlv.level < VLV_WM_LEVEL_PM5)
1353                 chv_set_memory_pm5(dev_priv, true);
1354
1355         if (wm.level >= VLV_WM_LEVEL_DDR_DVFS &&
1356             dev_priv->wm.vlv.level < VLV_WM_LEVEL_DDR_DVFS)
1357                 chv_set_memory_dvfs(dev_priv, true);
1358
1359         dev_priv->wm.vlv = wm;
1360 }
1361
1362 #define single_plane_enabled(mask) is_power_of_2(mask)
1363
1364 static void g4x_update_wm(struct drm_crtc *crtc)
1365 {
1366         struct drm_device *dev = crtc->dev;
1367         static const int sr_latency_ns = 12000;
1368         struct drm_i915_private *dev_priv = dev->dev_private;
1369         int planea_wm, planeb_wm, cursora_wm, cursorb_wm;
1370         int plane_sr, cursor_sr;
1371         unsigned int enabled = 0;
1372         bool cxsr_enabled;
1373
1374         if (g4x_compute_wm0(dev, PIPE_A,
1375                             &g4x_wm_info, pessimal_latency_ns,
1376                             &g4x_cursor_wm_info, pessimal_latency_ns,
1377                             &planea_wm, &cursora_wm))
1378                 enabled |= 1 << PIPE_A;
1379
1380         if (g4x_compute_wm0(dev, PIPE_B,
1381                             &g4x_wm_info, pessimal_latency_ns,
1382                             &g4x_cursor_wm_info, pessimal_latency_ns,
1383                             &planeb_wm, &cursorb_wm))
1384                 enabled |= 1 << PIPE_B;
1385
1386         if (single_plane_enabled(enabled) &&
1387             g4x_compute_srwm(dev, ffs(enabled) - 1,
1388                              sr_latency_ns,
1389                              &g4x_wm_info,
1390                              &g4x_cursor_wm_info,
1391                              &plane_sr, &cursor_sr)) {
1392                 cxsr_enabled = true;
1393         } else {
1394                 cxsr_enabled = false;
1395                 intel_set_memory_cxsr(dev_priv, false);
1396                 plane_sr = cursor_sr = 0;
1397         }
1398
1399         DRM_DEBUG_KMS("Setting FIFO watermarks - A: plane=%d, cursor=%d, "
1400                       "B: plane=%d, cursor=%d, SR: plane=%d, cursor=%d\n",
1401                       planea_wm, cursora_wm,
1402                       planeb_wm, cursorb_wm,
1403                       plane_sr, cursor_sr);
1404
1405         I915_WRITE(DSPFW1,
1406                    FW_WM(plane_sr, SR) |
1407                    FW_WM(cursorb_wm, CURSORB) |
1408                    FW_WM(planeb_wm, PLANEB) |
1409                    FW_WM(planea_wm, PLANEA));
1410         I915_WRITE(DSPFW2,
1411                    (I915_READ(DSPFW2) & ~DSPFW_CURSORA_MASK) |
1412                    FW_WM(cursora_wm, CURSORA));
1413         /* HPLL off in SR has some issues on G4x... disable it */
1414         I915_WRITE(DSPFW3,
1415                    (I915_READ(DSPFW3) & ~(DSPFW_HPLL_SR_EN | DSPFW_CURSOR_SR_MASK)) |
1416                    FW_WM(cursor_sr, CURSOR_SR));
1417
1418         if (cxsr_enabled)
1419                 intel_set_memory_cxsr(dev_priv, true);
1420 }
1421
1422 static void i965_update_wm(struct drm_crtc *unused_crtc)
1423 {
1424         struct drm_device *dev = unused_crtc->dev;
1425         struct drm_i915_private *dev_priv = dev->dev_private;
1426         struct drm_crtc *crtc;
1427         int srwm = 1;
1428         int cursor_sr = 16;
1429         bool cxsr_enabled;
1430
1431         /* Calc sr entries for one plane configs */
1432         crtc = single_enabled_crtc(dev);
1433         if (crtc) {
1434                 /* self-refresh has much higher latency */
1435                 static const int sr_latency_ns = 12000;
1436                 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1437                 int clock = adjusted_mode->crtc_clock;
1438                 int htotal = adjusted_mode->crtc_htotal;
1439                 int hdisplay = to_intel_crtc(crtc)->config->pipe_src_w;
1440                 int pixel_size = crtc->primary->state->fb->bits_per_pixel / 8;
1441                 unsigned long line_time_us;
1442                 int entries;
1443
1444                 line_time_us = max(htotal * 1000 / clock, 1);
1445
1446                 /* Use ns/us then divide to preserve precision */
1447                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1448                         pixel_size * hdisplay;
1449                 entries = DIV_ROUND_UP(entries, I915_FIFO_LINE_SIZE);
1450                 srwm = I965_FIFO_SIZE - entries;
1451                 if (srwm < 0)
1452                         srwm = 1;
1453                 srwm &= 0x1ff;
1454                 DRM_DEBUG_KMS("self-refresh entries: %d, wm: %d\n",
1455                               entries, srwm);
1456
1457                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1458                         pixel_size * crtc->cursor->state->crtc_w;
1459                 entries = DIV_ROUND_UP(entries,
1460                                           i965_cursor_wm_info.cacheline_size);
1461                 cursor_sr = i965_cursor_wm_info.fifo_size -
1462                         (entries + i965_cursor_wm_info.guard_size);
1463
1464                 if (cursor_sr > i965_cursor_wm_info.max_wm)
1465                         cursor_sr = i965_cursor_wm_info.max_wm;
1466
1467                 DRM_DEBUG_KMS("self-refresh watermark: display plane %d "
1468                               "cursor %d\n", srwm, cursor_sr);
1469
1470                 cxsr_enabled = true;
1471         } else {
1472                 cxsr_enabled = false;
1473                 /* Turn off self refresh if both pipes are enabled */
1474                 intel_set_memory_cxsr(dev_priv, false);
1475         }
1476
1477         DRM_DEBUG_KMS("Setting FIFO watermarks - A: 8, B: 8, C: 8, SR %d\n",
1478                       srwm);
1479
1480         /* 965 has limitations... */
1481         I915_WRITE(DSPFW1, FW_WM(srwm, SR) |
1482                    FW_WM(8, CURSORB) |
1483                    FW_WM(8, PLANEB) |
1484                    FW_WM(8, PLANEA));
1485         I915_WRITE(DSPFW2, FW_WM(8, CURSORA) |
1486                    FW_WM(8, PLANEC_OLD));
1487         /* update cursor SR watermark */
1488         I915_WRITE(DSPFW3, FW_WM(cursor_sr, CURSOR_SR));
1489
1490         if (cxsr_enabled)
1491                 intel_set_memory_cxsr(dev_priv, true);
1492 }
1493
1494 #undef FW_WM
1495
1496 static void i9xx_update_wm(struct drm_crtc *unused_crtc)
1497 {
1498         struct drm_device *dev = unused_crtc->dev;
1499         struct drm_i915_private *dev_priv = dev->dev_private;
1500         const struct intel_watermark_params *wm_info;
1501         uint32_t fwater_lo;
1502         uint32_t fwater_hi;
1503         int cwm, srwm = 1;
1504         int fifo_size;
1505         int planea_wm, planeb_wm;
1506         struct drm_crtc *crtc, *enabled = NULL;
1507
1508         if (IS_I945GM(dev))
1509                 wm_info = &i945_wm_info;
1510         else if (!IS_GEN2(dev))
1511                 wm_info = &i915_wm_info;
1512         else
1513                 wm_info = &i830_a_wm_info;
1514
1515         fifo_size = dev_priv->display.get_fifo_size(dev, 0);
1516         crtc = intel_get_crtc_for_plane(dev, 0);
1517         if (intel_crtc_active(crtc)) {
1518                 const struct drm_display_mode *adjusted_mode;
1519                 int cpp = crtc->primary->state->fb->bits_per_pixel / 8;
1520                 if (IS_GEN2(dev))
1521                         cpp = 4;
1522
1523                 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1524                 planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1525                                                wm_info, fifo_size, cpp,
1526                                                pessimal_latency_ns);
1527                 enabled = crtc;
1528         } else {
1529                 planea_wm = fifo_size - wm_info->guard_size;
1530                 if (planea_wm > (long)wm_info->max_wm)
1531                         planea_wm = wm_info->max_wm;
1532         }
1533
1534         if (IS_GEN2(dev))
1535                 wm_info = &i830_bc_wm_info;
1536
1537         fifo_size = dev_priv->display.get_fifo_size(dev, 1);
1538         crtc = intel_get_crtc_for_plane(dev, 1);
1539         if (intel_crtc_active(crtc)) {
1540                 const struct drm_display_mode *adjusted_mode;
1541                 int cpp = crtc->primary->state->fb->bits_per_pixel / 8;
1542                 if (IS_GEN2(dev))
1543                         cpp = 4;
1544
1545                 adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1546                 planeb_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1547                                                wm_info, fifo_size, cpp,
1548                                                pessimal_latency_ns);
1549                 if (enabled == NULL)
1550                         enabled = crtc;
1551                 else
1552                         enabled = NULL;
1553         } else {
1554                 planeb_wm = fifo_size - wm_info->guard_size;
1555                 if (planeb_wm > (long)wm_info->max_wm)
1556                         planeb_wm = wm_info->max_wm;
1557         }
1558
1559         DRM_DEBUG_KMS("FIFO watermarks - A: %d, B: %d\n", planea_wm, planeb_wm);
1560
1561         if (IS_I915GM(dev) && enabled) {
1562                 struct drm_i915_gem_object *obj;
1563
1564                 obj = intel_fb_obj(enabled->primary->state->fb);
1565
1566                 /* self-refresh seems busted with untiled */
1567                 if (obj->tiling_mode == I915_TILING_NONE)
1568                         enabled = NULL;
1569         }
1570
1571         /*
1572          * Overlay gets an aggressive default since video jitter is bad.
1573          */
1574         cwm = 2;
1575
1576         /* Play safe and disable self-refresh before adjusting watermarks. */
1577         intel_set_memory_cxsr(dev_priv, false);
1578
1579         /* Calc sr entries for one plane configs */
1580         if (HAS_FW_BLC(dev) && enabled) {
1581                 /* self-refresh has much higher latency */
1582                 static const int sr_latency_ns = 6000;
1583                 const struct drm_display_mode *adjusted_mode = &to_intel_crtc(enabled)->config->base.adjusted_mode;
1584                 int clock = adjusted_mode->crtc_clock;
1585                 int htotal = adjusted_mode->crtc_htotal;
1586                 int hdisplay = to_intel_crtc(enabled)->config->pipe_src_w;
1587                 int pixel_size = enabled->primary->state->fb->bits_per_pixel / 8;
1588                 unsigned long line_time_us;
1589                 int entries;
1590
1591                 line_time_us = max(htotal * 1000 / clock, 1);
1592
1593                 /* Use ns/us then divide to preserve precision */
1594                 entries = (((sr_latency_ns / line_time_us) + 1000) / 1000) *
1595                         pixel_size * hdisplay;
1596                 entries = DIV_ROUND_UP(entries, wm_info->cacheline_size);
1597                 DRM_DEBUG_KMS("self-refresh entries: %d\n", entries);
1598                 srwm = wm_info->fifo_size - entries;
1599                 if (srwm < 0)
1600                         srwm = 1;
1601
1602                 if (IS_I945G(dev) || IS_I945GM(dev))
1603                         I915_WRITE(FW_BLC_SELF,
1604                                    FW_BLC_SELF_FIFO_MASK | (srwm & 0xff));
1605                 else if (IS_I915GM(dev))
1606                         I915_WRITE(FW_BLC_SELF, srwm & 0x3f);
1607         }
1608
1609         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d, B: %d, C: %d, SR %d\n",
1610                       planea_wm, planeb_wm, cwm, srwm);
1611
1612         fwater_lo = ((planeb_wm & 0x3f) << 16) | (planea_wm & 0x3f);
1613         fwater_hi = (cwm & 0x1f);
1614
1615         /* Set request length to 8 cachelines per fetch */
1616         fwater_lo = fwater_lo | (1 << 24) | (1 << 8);
1617         fwater_hi = fwater_hi | (1 << 8);
1618
1619         I915_WRITE(FW_BLC, fwater_lo);
1620         I915_WRITE(FW_BLC2, fwater_hi);
1621
1622         if (enabled)
1623                 intel_set_memory_cxsr(dev_priv, true);
1624 }
1625
1626 static void i845_update_wm(struct drm_crtc *unused_crtc)
1627 {
1628         struct drm_device *dev = unused_crtc->dev;
1629         struct drm_i915_private *dev_priv = dev->dev_private;
1630         struct drm_crtc *crtc;
1631         const struct drm_display_mode *adjusted_mode;
1632         uint32_t fwater_lo;
1633         int planea_wm;
1634
1635         crtc = single_enabled_crtc(dev);
1636         if (crtc == NULL)
1637                 return;
1638
1639         adjusted_mode = &to_intel_crtc(crtc)->config->base.adjusted_mode;
1640         planea_wm = intel_calculate_wm(adjusted_mode->crtc_clock,
1641                                        &i845_wm_info,
1642                                        dev_priv->display.get_fifo_size(dev, 0),
1643                                        4, pessimal_latency_ns);
1644         fwater_lo = I915_READ(FW_BLC) & ~0xfff;
1645         fwater_lo |= (3<<8) | planea_wm;
1646
1647         DRM_DEBUG_KMS("Setting FIFO watermarks - A: %d\n", planea_wm);
1648
1649         I915_WRITE(FW_BLC, fwater_lo);
1650 }
1651
1652 uint32_t ilk_pipe_pixel_rate(const struct intel_crtc_state *pipe_config)
1653 {
1654         uint32_t pixel_rate;
1655
1656         pixel_rate = pipe_config->base.adjusted_mode.crtc_clock;
1657
1658         /* We only use IF-ID interlacing. If we ever use PF-ID we'll need to
1659          * adjust the pixel_rate here. */
1660
1661         if (pipe_config->pch_pfit.enabled) {
1662                 uint64_t pipe_w, pipe_h, pfit_w, pfit_h;
1663                 uint32_t pfit_size = pipe_config->pch_pfit.size;
1664
1665                 pipe_w = pipe_config->pipe_src_w;
1666                 pipe_h = pipe_config->pipe_src_h;
1667
1668                 pfit_w = (pfit_size >> 16) & 0xFFFF;
1669                 pfit_h = pfit_size & 0xFFFF;
1670                 if (pipe_w < pfit_w)
1671                         pipe_w = pfit_w;
1672                 if (pipe_h < pfit_h)
1673                         pipe_h = pfit_h;
1674
1675                 pixel_rate = div_u64((uint64_t) pixel_rate * pipe_w * pipe_h,
1676                                      pfit_w * pfit_h);
1677         }
1678
1679         return pixel_rate;
1680 }
1681
1682 /* latency must be in 0.1us units. */
1683 static uint32_t ilk_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
1684                                uint32_t latency)
1685 {
1686         uint64_t ret;
1687
1688         if (WARN(latency == 0, "Latency value missing\n"))
1689                 return UINT_MAX;
1690
1691         ret = (uint64_t) pixel_rate * bytes_per_pixel * latency;
1692         ret = DIV_ROUND_UP_ULL(ret, 64 * 10000) + 2;
1693
1694         return ret;
1695 }
1696
1697 /* latency must be in 0.1us units. */
1698 static uint32_t ilk_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
1699                                uint32_t horiz_pixels, uint8_t bytes_per_pixel,
1700                                uint32_t latency)
1701 {
1702         uint32_t ret;
1703
1704         if (WARN(latency == 0, "Latency value missing\n"))
1705                 return UINT_MAX;
1706
1707         ret = (latency * pixel_rate) / (pipe_htotal * 10000);
1708         ret = (ret + 1) * horiz_pixels * bytes_per_pixel;
1709         ret = DIV_ROUND_UP(ret, 64) + 2;
1710         return ret;
1711 }
1712
1713 static uint32_t ilk_wm_fbc(uint32_t pri_val, uint32_t horiz_pixels,
1714                            uint8_t bytes_per_pixel)
1715 {
1716         return DIV_ROUND_UP(pri_val * 64, horiz_pixels * bytes_per_pixel) + 2;
1717 }
1718
1719 struct ilk_wm_maximums {
1720         uint16_t pri;
1721         uint16_t spr;
1722         uint16_t cur;
1723         uint16_t fbc;
1724 };
1725
1726 /*
1727  * For both WM_PIPE and WM_LP.
1728  * mem_value must be in 0.1us units.
1729  */
1730 static uint32_t ilk_compute_pri_wm(const struct intel_crtc_state *cstate,
1731                                    const struct intel_plane_state *pstate,
1732                                    uint32_t mem_value,
1733                                    bool is_lp)
1734 {
1735         int bpp = pstate->base.fb ? pstate->base.fb->bits_per_pixel / 8 : 0;
1736         uint32_t method1, method2;
1737
1738         if (!cstate->base.active || !pstate->visible)
1739                 return 0;
1740
1741         method1 = ilk_wm_method1(ilk_pipe_pixel_rate(cstate), bpp, mem_value);
1742
1743         if (!is_lp)
1744                 return method1;
1745
1746         method2 = ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1747                                  cstate->base.adjusted_mode.crtc_htotal,
1748                                  drm_rect_width(&pstate->dst),
1749                                  bpp,
1750                                  mem_value);
1751
1752         return min(method1, method2);
1753 }
1754
1755 /*
1756  * For both WM_PIPE and WM_LP.
1757  * mem_value must be in 0.1us units.
1758  */
1759 static uint32_t ilk_compute_spr_wm(const struct intel_crtc_state *cstate,
1760                                    const struct intel_plane_state *pstate,
1761                                    uint32_t mem_value)
1762 {
1763         int bpp = pstate->base.fb ? pstate->base.fb->bits_per_pixel / 8 : 0;
1764         uint32_t method1, method2;
1765
1766         if (!cstate->base.active || !pstate->visible)
1767                 return 0;
1768
1769         method1 = ilk_wm_method1(ilk_pipe_pixel_rate(cstate), bpp, mem_value);
1770         method2 = ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1771                                  cstate->base.adjusted_mode.crtc_htotal,
1772                                  drm_rect_width(&pstate->dst),
1773                                  bpp,
1774                                  mem_value);
1775         return min(method1, method2);
1776 }
1777
1778 /*
1779  * For both WM_PIPE and WM_LP.
1780  * mem_value must be in 0.1us units.
1781  */
1782 static uint32_t ilk_compute_cur_wm(const struct intel_crtc_state *cstate,
1783                                    const struct intel_plane_state *pstate,
1784                                    uint32_t mem_value)
1785 {
1786         /*
1787          * We treat the cursor plane as always-on for the purposes of watermark
1788          * calculation.  Until we have two-stage watermark programming merged,
1789          * this is necessary to avoid flickering.
1790          */
1791         int cpp = 4;
1792         int width = pstate->visible ? pstate->base.crtc_w : 64;
1793
1794         if (!cstate->base.active)
1795                 return 0;
1796
1797         return ilk_wm_method2(ilk_pipe_pixel_rate(cstate),
1798                               cstate->base.adjusted_mode.crtc_htotal,
1799                               width, cpp, mem_value);
1800 }
1801
1802 /* Only for WM_LP. */
1803 static uint32_t ilk_compute_fbc_wm(const struct intel_crtc_state *cstate,
1804                                    const struct intel_plane_state *pstate,
1805                                    uint32_t pri_val)
1806 {
1807         int bpp = pstate->base.fb ? pstate->base.fb->bits_per_pixel / 8 : 0;
1808
1809         if (!cstate->base.active || !pstate->visible)
1810                 return 0;
1811
1812         return ilk_wm_fbc(pri_val, drm_rect_width(&pstate->dst), bpp);
1813 }
1814
1815 static unsigned int ilk_display_fifo_size(const struct drm_device *dev)
1816 {
1817         if (INTEL_INFO(dev)->gen >= 8)
1818                 return 3072;
1819         else if (INTEL_INFO(dev)->gen >= 7)
1820                 return 768;
1821         else
1822                 return 512;
1823 }
1824
1825 static unsigned int ilk_plane_wm_reg_max(const struct drm_device *dev,
1826                                          int level, bool is_sprite)
1827 {
1828         if (INTEL_INFO(dev)->gen >= 8)
1829                 /* BDW primary/sprite plane watermarks */
1830                 return level == 0 ? 255 : 2047;
1831         else if (INTEL_INFO(dev)->gen >= 7)
1832                 /* IVB/HSW primary/sprite plane watermarks */
1833                 return level == 0 ? 127 : 1023;
1834         else if (!is_sprite)
1835                 /* ILK/SNB primary plane watermarks */
1836                 return level == 0 ? 127 : 511;
1837         else
1838                 /* ILK/SNB sprite plane watermarks */
1839                 return level == 0 ? 63 : 255;
1840 }
1841
1842 static unsigned int ilk_cursor_wm_reg_max(const struct drm_device *dev,
1843                                           int level)
1844 {
1845         if (INTEL_INFO(dev)->gen >= 7)
1846                 return level == 0 ? 63 : 255;
1847         else
1848                 return level == 0 ? 31 : 63;
1849 }
1850
1851 static unsigned int ilk_fbc_wm_reg_max(const struct drm_device *dev)
1852 {
1853         if (INTEL_INFO(dev)->gen >= 8)
1854                 return 31;
1855         else
1856                 return 15;
1857 }
1858
1859 /* Calculate the maximum primary/sprite plane watermark */
1860 static unsigned int ilk_plane_wm_max(const struct drm_device *dev,
1861                                      int level,
1862                                      const struct intel_wm_config *config,
1863                                      enum intel_ddb_partitioning ddb_partitioning,
1864                                      bool is_sprite)
1865 {
1866         unsigned int fifo_size = ilk_display_fifo_size(dev);
1867
1868         /* if sprites aren't enabled, sprites get nothing */
1869         if (is_sprite && !config->sprites_enabled)
1870                 return 0;
1871
1872         /* HSW allows LP1+ watermarks even with multiple pipes */
1873         if (level == 0 || config->num_pipes_active > 1) {
1874                 fifo_size /= INTEL_INFO(dev)->num_pipes;
1875
1876                 /*
1877                  * For some reason the non self refresh
1878                  * FIFO size is only half of the self
1879                  * refresh FIFO size on ILK/SNB.
1880                  */
1881                 if (INTEL_INFO(dev)->gen <= 6)
1882                         fifo_size /= 2;
1883         }
1884
1885         if (config->sprites_enabled) {
1886                 /* level 0 is always calculated with 1:1 split */
1887                 if (level > 0 && ddb_partitioning == INTEL_DDB_PART_5_6) {
1888                         if (is_sprite)
1889                                 fifo_size *= 5;
1890                         fifo_size /= 6;
1891                 } else {
1892                         fifo_size /= 2;
1893                 }
1894         }
1895
1896         /* clamp to max that the registers can hold */
1897         return min(fifo_size, ilk_plane_wm_reg_max(dev, level, is_sprite));
1898 }
1899
1900 /* Calculate the maximum cursor plane watermark */
1901 static unsigned int ilk_cursor_wm_max(const struct drm_device *dev,
1902                                       int level,
1903                                       const struct intel_wm_config *config)
1904 {
1905         /* HSW LP1+ watermarks w/ multiple pipes */
1906         if (level > 0 && config->num_pipes_active > 1)
1907                 return 64;
1908
1909         /* otherwise just report max that registers can hold */
1910         return ilk_cursor_wm_reg_max(dev, level);
1911 }
1912
1913 static void ilk_compute_wm_maximums(const struct drm_device *dev,
1914                                     int level,
1915                                     const struct intel_wm_config *config,
1916                                     enum intel_ddb_partitioning ddb_partitioning,
1917                                     struct ilk_wm_maximums *max)
1918 {
1919         max->pri = ilk_plane_wm_max(dev, level, config, ddb_partitioning, false);
1920         max->spr = ilk_plane_wm_max(dev, level, config, ddb_partitioning, true);
1921         max->cur = ilk_cursor_wm_max(dev, level, config);
1922         max->fbc = ilk_fbc_wm_reg_max(dev);
1923 }
1924
1925 static void ilk_compute_wm_reg_maximums(struct drm_device *dev,
1926                                         int level,
1927                                         struct ilk_wm_maximums *max)
1928 {
1929         max->pri = ilk_plane_wm_reg_max(dev, level, false);
1930         max->spr = ilk_plane_wm_reg_max(dev, level, true);
1931         max->cur = ilk_cursor_wm_reg_max(dev, level);
1932         max->fbc = ilk_fbc_wm_reg_max(dev);
1933 }
1934
1935 static bool ilk_validate_wm_level(int level,
1936                                   const struct ilk_wm_maximums *max,
1937                                   struct intel_wm_level *result)
1938 {
1939         bool ret;
1940
1941         /* already determined to be invalid? */
1942         if (!result->enable)
1943                 return false;
1944
1945         result->enable = result->pri_val <= max->pri &&
1946                          result->spr_val <= max->spr &&
1947                          result->cur_val <= max->cur;
1948
1949         ret = result->enable;
1950
1951         /*
1952          * HACK until we can pre-compute everything,
1953          * and thus fail gracefully if LP0 watermarks
1954          * are exceeded...
1955          */
1956         if (level == 0 && !result->enable) {
1957                 if (result->pri_val > max->pri)
1958                         DRM_DEBUG_KMS("Primary WM%d too large %u (max %u)\n",
1959                                       level, result->pri_val, max->pri);
1960                 if (result->spr_val > max->spr)
1961                         DRM_DEBUG_KMS("Sprite WM%d too large %u (max %u)\n",
1962                                       level, result->spr_val, max->spr);
1963                 if (result->cur_val > max->cur)
1964                         DRM_DEBUG_KMS("Cursor WM%d too large %u (max %u)\n",
1965                                       level, result->cur_val, max->cur);
1966
1967                 result->pri_val = min_t(uint32_t, result->pri_val, max->pri);
1968                 result->spr_val = min_t(uint32_t, result->spr_val, max->spr);
1969                 result->cur_val = min_t(uint32_t, result->cur_val, max->cur);
1970                 result->enable = true;
1971         }
1972
1973         return ret;
1974 }
1975
1976 static void ilk_compute_wm_level(const struct drm_i915_private *dev_priv,
1977                                  const struct intel_crtc *intel_crtc,
1978                                  int level,
1979                                  struct intel_crtc_state *cstate,
1980                                  struct intel_plane_state *pristate,
1981                                  struct intel_plane_state *sprstate,
1982                                  struct intel_plane_state *curstate,
1983                                  struct intel_wm_level *result)
1984 {
1985         uint16_t pri_latency = dev_priv->wm.pri_latency[level];
1986         uint16_t spr_latency = dev_priv->wm.spr_latency[level];
1987         uint16_t cur_latency = dev_priv->wm.cur_latency[level];
1988
1989         /* WM1+ latency values stored in 0.5us units */
1990         if (level > 0) {
1991                 pri_latency *= 5;
1992                 spr_latency *= 5;
1993                 cur_latency *= 5;
1994         }
1995
1996         result->pri_val = ilk_compute_pri_wm(cstate, pristate,
1997                                              pri_latency, level);
1998         result->spr_val = ilk_compute_spr_wm(cstate, sprstate, spr_latency);
1999         result->cur_val = ilk_compute_cur_wm(cstate, curstate, cur_latency);
2000         result->fbc_val = ilk_compute_fbc_wm(cstate, pristate, result->pri_val);
2001         result->enable = true;
2002 }
2003
2004 static uint32_t
2005 hsw_compute_linetime_wm(struct drm_device *dev, struct drm_crtc *crtc)
2006 {
2007         struct drm_i915_private *dev_priv = dev->dev_private;
2008         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2009         const struct drm_display_mode *adjusted_mode = &intel_crtc->config->base.adjusted_mode;
2010         u32 linetime, ips_linetime;
2011
2012         if (!intel_crtc->active)
2013                 return 0;
2014
2015         /* The WM are computed with base on how long it takes to fill a single
2016          * row at the given clock rate, multiplied by 8.
2017          * */
2018         linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2019                                      adjusted_mode->crtc_clock);
2020         ips_linetime = DIV_ROUND_CLOSEST(adjusted_mode->crtc_htotal * 1000 * 8,
2021                                          dev_priv->cdclk_freq);
2022
2023         return PIPE_WM_LINETIME_IPS_LINETIME(ips_linetime) |
2024                PIPE_WM_LINETIME_TIME(linetime);
2025 }
2026
2027 static void intel_read_wm_latency(struct drm_device *dev, uint16_t wm[8])
2028 {
2029         struct drm_i915_private *dev_priv = dev->dev_private;
2030
2031         if (IS_GEN9(dev)) {
2032                 uint32_t val;
2033                 int ret, i;
2034                 int level, max_level = ilk_wm_max_level(dev);
2035
2036                 /* read the first set of memory latencies[0:3] */
2037                 val = 0; /* data0 to be programmed to 0 for first set */
2038                 mutex_lock(&dev_priv->rps.hw_lock);
2039                 ret = sandybridge_pcode_read(dev_priv,
2040                                              GEN9_PCODE_READ_MEM_LATENCY,
2041                                              &val);
2042                 mutex_unlock(&dev_priv->rps.hw_lock);
2043
2044                 if (ret) {
2045                         DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2046                         return;
2047                 }
2048
2049                 wm[0] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2050                 wm[1] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2051                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2052                 wm[2] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2053                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2054                 wm[3] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2055                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2056
2057                 /* read the second set of memory latencies[4:7] */
2058                 val = 1; /* data0 to be programmed to 1 for second set */
2059                 mutex_lock(&dev_priv->rps.hw_lock);
2060                 ret = sandybridge_pcode_read(dev_priv,
2061                                              GEN9_PCODE_READ_MEM_LATENCY,
2062                                              &val);
2063                 mutex_unlock(&dev_priv->rps.hw_lock);
2064                 if (ret) {
2065                         DRM_ERROR("SKL Mailbox read error = %d\n", ret);
2066                         return;
2067                 }
2068
2069                 wm[4] = val & GEN9_MEM_LATENCY_LEVEL_MASK;
2070                 wm[5] = (val >> GEN9_MEM_LATENCY_LEVEL_1_5_SHIFT) &
2071                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2072                 wm[6] = (val >> GEN9_MEM_LATENCY_LEVEL_2_6_SHIFT) &
2073                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2074                 wm[7] = (val >> GEN9_MEM_LATENCY_LEVEL_3_7_SHIFT) &
2075                                 GEN9_MEM_LATENCY_LEVEL_MASK;
2076
2077                 /*
2078                  * WaWmMemoryReadLatency:skl
2079                  *
2080                  * punit doesn't take into account the read latency so we need
2081                  * to add 2us to the various latency levels we retrieve from
2082                  * the punit.
2083                  *   - W0 is a bit special in that it's the only level that
2084                  *   can't be disabled if we want to have display working, so
2085                  *   we always add 2us there.
2086                  *   - For levels >=1, punit returns 0us latency when they are
2087                  *   disabled, so we respect that and don't add 2us then
2088                  *
2089                  * Additionally, if a level n (n > 1) has a 0us latency, all
2090                  * levels m (m >= n) need to be disabled. We make sure to
2091                  * sanitize the values out of the punit to satisfy this
2092                  * requirement.
2093                  */
2094                 wm[0] += 2;
2095                 for (level = 1; level <= max_level; level++)
2096                         if (wm[level] != 0)
2097                                 wm[level] += 2;
2098                         else {
2099                                 for (i = level + 1; i <= max_level; i++)
2100                                         wm[i] = 0;
2101
2102                                 break;
2103                         }
2104         } else if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2105                 uint64_t sskpd = I915_READ64(MCH_SSKPD);
2106
2107                 wm[0] = (sskpd >> 56) & 0xFF;
2108                 if (wm[0] == 0)
2109                         wm[0] = sskpd & 0xF;
2110                 wm[1] = (sskpd >> 4) & 0xFF;
2111                 wm[2] = (sskpd >> 12) & 0xFF;
2112                 wm[3] = (sskpd >> 20) & 0x1FF;
2113                 wm[4] = (sskpd >> 32) & 0x1FF;
2114         } else if (INTEL_INFO(dev)->gen >= 6) {
2115                 uint32_t sskpd = I915_READ(MCH_SSKPD);
2116
2117                 wm[0] = (sskpd >> SSKPD_WM0_SHIFT) & SSKPD_WM_MASK;
2118                 wm[1] = (sskpd >> SSKPD_WM1_SHIFT) & SSKPD_WM_MASK;
2119                 wm[2] = (sskpd >> SSKPD_WM2_SHIFT) & SSKPD_WM_MASK;
2120                 wm[3] = (sskpd >> SSKPD_WM3_SHIFT) & SSKPD_WM_MASK;
2121         } else if (INTEL_INFO(dev)->gen >= 5) {
2122                 uint32_t mltr = I915_READ(MLTR_ILK);
2123
2124                 /* ILK primary LP0 latency is 700 ns */
2125                 wm[0] = 7;
2126                 wm[1] = (mltr >> MLTR_WM1_SHIFT) & ILK_SRLT_MASK;
2127                 wm[2] = (mltr >> MLTR_WM2_SHIFT) & ILK_SRLT_MASK;
2128         }
2129 }
2130
2131 static void intel_fixup_spr_wm_latency(struct drm_device *dev, uint16_t wm[5])
2132 {
2133         /* ILK sprite LP0 latency is 1300 ns */
2134         if (INTEL_INFO(dev)->gen == 5)
2135                 wm[0] = 13;
2136 }
2137
2138 static void intel_fixup_cur_wm_latency(struct drm_device *dev, uint16_t wm[5])
2139 {
2140         /* ILK cursor LP0 latency is 1300 ns */
2141         if (INTEL_INFO(dev)->gen == 5)
2142                 wm[0] = 13;
2143
2144         /* WaDoubleCursorLP3Latency:ivb */
2145         if (IS_IVYBRIDGE(dev))
2146                 wm[3] *= 2;
2147 }
2148
2149 int ilk_wm_max_level(const struct drm_device *dev)
2150 {
2151         /* how many WM levels are we expecting */
2152         if (INTEL_INFO(dev)->gen >= 9)
2153                 return 7;
2154         else if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2155                 return 4;
2156         else if (INTEL_INFO(dev)->gen >= 6)
2157                 return 3;
2158         else
2159                 return 2;
2160 }
2161
2162 static void intel_print_wm_latency(struct drm_device *dev,
2163                                    const char *name,
2164                                    const uint16_t wm[8])
2165 {
2166         int level, max_level = ilk_wm_max_level(dev);
2167
2168         for (level = 0; level <= max_level; level++) {
2169                 unsigned int latency = wm[level];
2170
2171                 if (latency == 0) {
2172                         DRM_ERROR("%s WM%d latency not provided\n",
2173                                   name, level);
2174                         continue;
2175                 }
2176
2177                 /*
2178                  * - latencies are in us on gen9.
2179                  * - before then, WM1+ latency values are in 0.5us units
2180                  */
2181                 if (IS_GEN9(dev))
2182                         latency *= 10;
2183                 else if (level > 0)
2184                         latency *= 5;
2185
2186                 DRM_DEBUG_KMS("%s WM%d latency %u (%u.%u usec)\n",
2187                               name, level, wm[level],
2188                               latency / 10, latency % 10);
2189         }
2190 }
2191
2192 static bool ilk_increase_wm_latency(struct drm_i915_private *dev_priv,
2193                                     uint16_t wm[5], uint16_t min)
2194 {
2195         int level, max_level = ilk_wm_max_level(dev_priv->dev);
2196
2197         if (wm[0] >= min)
2198                 return false;
2199
2200         wm[0] = max(wm[0], min);
2201         for (level = 1; level <= max_level; level++)
2202                 wm[level] = max_t(uint16_t, wm[level], DIV_ROUND_UP(min, 5));
2203
2204         return true;
2205 }
2206
2207 static void snb_wm_latency_quirk(struct drm_device *dev)
2208 {
2209         struct drm_i915_private *dev_priv = dev->dev_private;
2210         bool changed;
2211
2212         /*
2213          * The BIOS provided WM memory latency values are often
2214          * inadequate for high resolution displays. Adjust them.
2215          */
2216         changed = ilk_increase_wm_latency(dev_priv, dev_priv->wm.pri_latency, 12) |
2217                 ilk_increase_wm_latency(dev_priv, dev_priv->wm.spr_latency, 12) |
2218                 ilk_increase_wm_latency(dev_priv, dev_priv->wm.cur_latency, 12);
2219
2220         if (!changed)
2221                 return;
2222
2223         DRM_DEBUG_KMS("WM latency values increased to avoid potential underruns\n");
2224         intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2225         intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2226         intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2227 }
2228
2229 static void ilk_setup_wm_latency(struct drm_device *dev)
2230 {
2231         struct drm_i915_private *dev_priv = dev->dev_private;
2232
2233         intel_read_wm_latency(dev, dev_priv->wm.pri_latency);
2234
2235         memcpy(dev_priv->wm.spr_latency, dev_priv->wm.pri_latency,
2236                sizeof(dev_priv->wm.pri_latency));
2237         memcpy(dev_priv->wm.cur_latency, dev_priv->wm.pri_latency,
2238                sizeof(dev_priv->wm.pri_latency));
2239
2240         intel_fixup_spr_wm_latency(dev, dev_priv->wm.spr_latency);
2241         intel_fixup_cur_wm_latency(dev, dev_priv->wm.cur_latency);
2242
2243         intel_print_wm_latency(dev, "Primary", dev_priv->wm.pri_latency);
2244         intel_print_wm_latency(dev, "Sprite", dev_priv->wm.spr_latency);
2245         intel_print_wm_latency(dev, "Cursor", dev_priv->wm.cur_latency);
2246
2247         if (IS_GEN6(dev))
2248                 snb_wm_latency_quirk(dev);
2249 }
2250
2251 static void skl_setup_wm_latency(struct drm_device *dev)
2252 {
2253         struct drm_i915_private *dev_priv = dev->dev_private;
2254
2255         intel_read_wm_latency(dev, dev_priv->wm.skl_latency);
2256         intel_print_wm_latency(dev, "Gen9 Plane", dev_priv->wm.skl_latency);
2257 }
2258
2259 /* Compute new watermarks for the pipe */
2260 static int ilk_compute_pipe_wm(struct intel_crtc *intel_crtc,
2261                                struct drm_atomic_state *state)
2262 {
2263         struct intel_pipe_wm *pipe_wm;
2264         struct drm_device *dev = intel_crtc->base.dev;
2265         const struct drm_i915_private *dev_priv = dev->dev_private;
2266         struct intel_crtc_state *cstate = NULL;
2267         struct intel_plane *intel_plane;
2268         struct drm_plane_state *ps;
2269         struct intel_plane_state *pristate = NULL;
2270         struct intel_plane_state *sprstate = NULL;
2271         struct intel_plane_state *curstate = NULL;
2272         int level, max_level = ilk_wm_max_level(dev);
2273         /* LP0 watermark maximums depend on this pipe alone */
2274         struct intel_wm_config config = {
2275                 .num_pipes_active = 1,
2276         };
2277         struct ilk_wm_maximums max;
2278
2279         cstate = intel_atomic_get_crtc_state(state, intel_crtc);
2280         if (IS_ERR(cstate))
2281                 return PTR_ERR(cstate);
2282
2283         pipe_wm = &cstate->wm.optimal.ilk;
2284
2285         for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
2286                 ps = drm_atomic_get_plane_state(state,
2287                                                 &intel_plane->base);
2288                 if (IS_ERR(ps))
2289                         return PTR_ERR(ps);
2290
2291                 if (intel_plane->base.type == DRM_PLANE_TYPE_PRIMARY)
2292                         pristate = to_intel_plane_state(ps);
2293                 else if (intel_plane->base.type == DRM_PLANE_TYPE_OVERLAY)
2294                         sprstate = to_intel_plane_state(ps);
2295                 else if (intel_plane->base.type == DRM_PLANE_TYPE_CURSOR)
2296                         curstate = to_intel_plane_state(ps);
2297         }
2298
2299         config.sprites_enabled = sprstate->visible;
2300         config.sprites_scaled = sprstate->visible &&
2301                 (drm_rect_width(&sprstate->dst) != drm_rect_width(&sprstate->src) >> 16 ||
2302                 drm_rect_height(&sprstate->dst) != drm_rect_height(&sprstate->src) >> 16);
2303
2304         pipe_wm->pipe_enabled = cstate->base.active;
2305         pipe_wm->sprites_enabled = config.sprites_enabled;
2306         pipe_wm->sprites_scaled = config.sprites_scaled;
2307
2308         /* ILK/SNB: LP2+ watermarks only w/o sprites */
2309         if (INTEL_INFO(dev)->gen <= 6 && sprstate->visible)
2310                 max_level = 1;
2311
2312         /* ILK/SNB/IVB: LP1+ watermarks only w/o scaling */
2313         if (config.sprites_scaled)
2314                 max_level = 0;
2315
2316         ilk_compute_wm_level(dev_priv, intel_crtc, 0, cstate,
2317                              pristate, sprstate, curstate, &pipe_wm->wm[0]);
2318
2319         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2320                 pipe_wm->linetime = hsw_compute_linetime_wm(dev,
2321                                                             &intel_crtc->base);
2322
2323         /* LP0 watermarks always use 1/2 DDB partitioning */
2324         ilk_compute_wm_maximums(dev, 0, &config, INTEL_DDB_PART_1_2, &max);
2325
2326         /* At least LP0 must be valid */
2327         if (!ilk_validate_wm_level(0, &max, &pipe_wm->wm[0]))
2328                 return -EINVAL;
2329
2330         ilk_compute_wm_reg_maximums(dev, 1, &max);
2331
2332         for (level = 1; level <= max_level; level++) {
2333                 struct intel_wm_level wm = {};
2334
2335                 ilk_compute_wm_level(dev_priv, intel_crtc, level, cstate,
2336                                      pristate, sprstate, curstate, &wm);
2337
2338                 /*
2339                  * Disable any watermark level that exceeds the
2340                  * register maximums since such watermarks are
2341                  * always invalid.
2342                  */
2343                 if (!ilk_validate_wm_level(level, &max, &wm))
2344                         break;
2345
2346                 pipe_wm->wm[level] = wm;
2347         }
2348
2349         return 0;
2350 }
2351
2352 /*
2353  * Merge the watermarks from all active pipes for a specific level.
2354  */
2355 static void ilk_merge_wm_level(struct drm_device *dev,
2356                                int level,
2357                                struct intel_wm_level *ret_wm)
2358 {
2359         const struct intel_crtc *intel_crtc;
2360
2361         ret_wm->enable = true;
2362
2363         for_each_intel_crtc(dev, intel_crtc) {
2364                 const struct intel_crtc_state *cstate =
2365                         to_intel_crtc_state(intel_crtc->base.state);
2366                 const struct intel_pipe_wm *active = &cstate->wm.optimal.ilk;
2367                 const struct intel_wm_level *wm = &active->wm[level];
2368
2369                 if (!active->pipe_enabled)
2370                         continue;
2371
2372                 /*
2373                  * The watermark values may have been used in the past,
2374                  * so we must maintain them in the registers for some
2375                  * time even if the level is now disabled.
2376                  */
2377                 if (!wm->enable)
2378                         ret_wm->enable = false;
2379
2380                 ret_wm->pri_val = max(ret_wm->pri_val, wm->pri_val);
2381                 ret_wm->spr_val = max(ret_wm->spr_val, wm->spr_val);
2382                 ret_wm->cur_val = max(ret_wm->cur_val, wm->cur_val);
2383                 ret_wm->fbc_val = max(ret_wm->fbc_val, wm->fbc_val);
2384         }
2385 }
2386
2387 /*
2388  * Merge all low power watermarks for all active pipes.
2389  */
2390 static void ilk_wm_merge(struct drm_device *dev,
2391                          const struct intel_wm_config *config,
2392                          const struct ilk_wm_maximums *max,
2393                          struct intel_pipe_wm *merged)
2394 {
2395         struct drm_i915_private *dev_priv = dev->dev_private;
2396         int level, max_level = ilk_wm_max_level(dev);
2397         int last_enabled_level = max_level;
2398
2399         /* ILK/SNB/IVB: LP1+ watermarks only w/ single pipe */
2400         if ((INTEL_INFO(dev)->gen <= 6 || IS_IVYBRIDGE(dev)) &&
2401             config->num_pipes_active > 1)
2402                 return;
2403
2404         /* ILK: FBC WM must be disabled always */
2405         merged->fbc_wm_enabled = INTEL_INFO(dev)->gen >= 6;
2406
2407         /* merge each WM1+ level */
2408         for (level = 1; level <= max_level; level++) {
2409                 struct intel_wm_level *wm = &merged->wm[level];
2410
2411                 ilk_merge_wm_level(dev, level, wm);
2412
2413                 if (level > last_enabled_level)
2414                         wm->enable = false;
2415                 else if (!ilk_validate_wm_level(level, max, wm))
2416                         /* make sure all following levels get disabled */
2417                         last_enabled_level = level - 1;
2418
2419                 /*
2420                  * The spec says it is preferred to disable
2421                  * FBC WMs instead of disabling a WM level.
2422                  */
2423                 if (wm->fbc_val > max->fbc) {
2424                         if (wm->enable)
2425                                 merged->fbc_wm_enabled = false;
2426                         wm->fbc_val = 0;
2427                 }
2428         }
2429
2430         /* ILK: LP2+ must be disabled when FBC WM is disabled but FBC enabled */
2431         /*
2432          * FIXME this is racy. FBC might get enabled later.
2433          * What we should check here is whether FBC can be
2434          * enabled sometime later.
2435          */
2436         if (IS_GEN5(dev) && !merged->fbc_wm_enabled &&
2437             intel_fbc_is_active(dev_priv)) {
2438                 for (level = 2; level <= max_level; level++) {
2439                         struct intel_wm_level *wm = &merged->wm[level];
2440
2441                         wm->enable = false;
2442                 }
2443         }
2444 }
2445
2446 static int ilk_wm_lp_to_level(int wm_lp, const struct intel_pipe_wm *pipe_wm)
2447 {
2448         /* LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4 */
2449         return wm_lp + (wm_lp >= 2 && pipe_wm->wm[4].enable);
2450 }
2451
2452 /* The value we need to program into the WM_LPx latency field */
2453 static unsigned int ilk_wm_lp_latency(struct drm_device *dev, int level)
2454 {
2455         struct drm_i915_private *dev_priv = dev->dev_private;
2456
2457         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
2458                 return 2 * level;
2459         else
2460                 return dev_priv->wm.pri_latency[level];
2461 }
2462
2463 static void ilk_compute_wm_results(struct drm_device *dev,
2464                                    const struct intel_pipe_wm *merged,
2465                                    enum intel_ddb_partitioning partitioning,
2466                                    struct ilk_wm_values *results)
2467 {
2468         struct intel_crtc *intel_crtc;
2469         int level, wm_lp;
2470
2471         results->enable_fbc_wm = merged->fbc_wm_enabled;
2472         results->partitioning = partitioning;
2473
2474         /* LP1+ register values */
2475         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2476                 const struct intel_wm_level *r;
2477
2478                 level = ilk_wm_lp_to_level(wm_lp, merged);
2479
2480                 r = &merged->wm[level];
2481
2482                 /*
2483                  * Maintain the watermark values even if the level is
2484                  * disabled. Doing otherwise could cause underruns.
2485                  */
2486                 results->wm_lp[wm_lp - 1] =
2487                         (ilk_wm_lp_latency(dev, level) << WM1_LP_LATENCY_SHIFT) |
2488                         (r->pri_val << WM1_LP_SR_SHIFT) |
2489                         r->cur_val;
2490
2491                 if (r->enable)
2492                         results->wm_lp[wm_lp - 1] |= WM1_LP_SR_EN;
2493
2494                 if (INTEL_INFO(dev)->gen >= 8)
2495                         results->wm_lp[wm_lp - 1] |=
2496                                 r->fbc_val << WM1_LP_FBC_SHIFT_BDW;
2497                 else
2498                         results->wm_lp[wm_lp - 1] |=
2499                                 r->fbc_val << WM1_LP_FBC_SHIFT;
2500
2501                 /*
2502                  * Always set WM1S_LP_EN when spr_val != 0, even if the
2503                  * level is disabled. Doing otherwise could cause underruns.
2504                  */
2505                 if (INTEL_INFO(dev)->gen <= 6 && r->spr_val) {
2506                         WARN_ON(wm_lp != 1);
2507                         results->wm_lp_spr[wm_lp - 1] = WM1S_LP_EN | r->spr_val;
2508                 } else
2509                         results->wm_lp_spr[wm_lp - 1] = r->spr_val;
2510         }
2511
2512         /* LP0 register values */
2513         for_each_intel_crtc(dev, intel_crtc) {
2514                 const struct intel_crtc_state *cstate =
2515                         to_intel_crtc_state(intel_crtc->base.state);
2516                 enum pipe pipe = intel_crtc->pipe;
2517                 const struct intel_wm_level *r = &cstate->wm.optimal.ilk.wm[0];
2518
2519                 if (WARN_ON(!r->enable))
2520                         continue;
2521
2522                 results->wm_linetime[pipe] = cstate->wm.optimal.ilk.linetime;
2523
2524                 results->wm_pipe[pipe] =
2525                         (r->pri_val << WM0_PIPE_PLANE_SHIFT) |
2526                         (r->spr_val << WM0_PIPE_SPRITE_SHIFT) |
2527                         r->cur_val;
2528         }
2529 }
2530
2531 /* Find the result with the highest level enabled. Check for enable_fbc_wm in
2532  * case both are at the same level. Prefer r1 in case they're the same. */
2533 static struct intel_pipe_wm *ilk_find_best_result(struct drm_device *dev,
2534                                                   struct intel_pipe_wm *r1,
2535                                                   struct intel_pipe_wm *r2)
2536 {
2537         int level, max_level = ilk_wm_max_level(dev);
2538         int level1 = 0, level2 = 0;
2539
2540         for (level = 1; level <= max_level; level++) {
2541                 if (r1->wm[level].enable)
2542                         level1 = level;
2543                 if (r2->wm[level].enable)
2544                         level2 = level;
2545         }
2546
2547         if (level1 == level2) {
2548                 if (r2->fbc_wm_enabled && !r1->fbc_wm_enabled)
2549                         return r2;
2550                 else
2551                         return r1;
2552         } else if (level1 > level2) {
2553                 return r1;
2554         } else {
2555                 return r2;
2556         }
2557 }
2558
2559 /* dirty bits used to track which watermarks need changes */
2560 #define WM_DIRTY_PIPE(pipe) (1 << (pipe))
2561 #define WM_DIRTY_LINETIME(pipe) (1 << (8 + (pipe)))
2562 #define WM_DIRTY_LP(wm_lp) (1 << (15 + (wm_lp)))
2563 #define WM_DIRTY_LP_ALL (WM_DIRTY_LP(1) | WM_DIRTY_LP(2) | WM_DIRTY_LP(3))
2564 #define WM_DIRTY_FBC (1 << 24)
2565 #define WM_DIRTY_DDB (1 << 25)
2566
2567 static unsigned int ilk_compute_wm_dirty(struct drm_i915_private *dev_priv,
2568                                          const struct ilk_wm_values *old,
2569                                          const struct ilk_wm_values *new)
2570 {
2571         unsigned int dirty = 0;
2572         enum pipe pipe;
2573         int wm_lp;
2574
2575         for_each_pipe(dev_priv, pipe) {
2576                 if (old->wm_linetime[pipe] != new->wm_linetime[pipe]) {
2577                         dirty |= WM_DIRTY_LINETIME(pipe);
2578                         /* Must disable LP1+ watermarks too */
2579                         dirty |= WM_DIRTY_LP_ALL;
2580                 }
2581
2582                 if (old->wm_pipe[pipe] != new->wm_pipe[pipe]) {
2583                         dirty |= WM_DIRTY_PIPE(pipe);
2584                         /* Must disable LP1+ watermarks too */
2585                         dirty |= WM_DIRTY_LP_ALL;
2586                 }
2587         }
2588
2589         if (old->enable_fbc_wm != new->enable_fbc_wm) {
2590                 dirty |= WM_DIRTY_FBC;
2591                 /* Must disable LP1+ watermarks too */
2592                 dirty |= WM_DIRTY_LP_ALL;
2593         }
2594
2595         if (old->partitioning != new->partitioning) {
2596                 dirty |= WM_DIRTY_DDB;
2597                 /* Must disable LP1+ watermarks too */
2598                 dirty |= WM_DIRTY_LP_ALL;
2599         }
2600
2601         /* LP1+ watermarks already deemed dirty, no need to continue */
2602         if (dirty & WM_DIRTY_LP_ALL)
2603                 return dirty;
2604
2605         /* Find the lowest numbered LP1+ watermark in need of an update... */
2606         for (wm_lp = 1; wm_lp <= 3; wm_lp++) {
2607                 if (old->wm_lp[wm_lp - 1] != new->wm_lp[wm_lp - 1] ||
2608                     old->wm_lp_spr[wm_lp - 1] != new->wm_lp_spr[wm_lp - 1])
2609                         break;
2610         }
2611
2612         /* ...and mark it and all higher numbered LP1+ watermarks as dirty */
2613         for (; wm_lp <= 3; wm_lp++)
2614                 dirty |= WM_DIRTY_LP(wm_lp);
2615
2616         return dirty;
2617 }
2618
2619 static bool _ilk_disable_lp_wm(struct drm_i915_private *dev_priv,
2620                                unsigned int dirty)
2621 {
2622         struct ilk_wm_values *previous = &dev_priv->wm.hw;
2623         bool changed = false;
2624
2625         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] & WM1_LP_SR_EN) {
2626                 previous->wm_lp[2] &= ~WM1_LP_SR_EN;
2627                 I915_WRITE(WM3_LP_ILK, previous->wm_lp[2]);
2628                 changed = true;
2629         }
2630         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] & WM1_LP_SR_EN) {
2631                 previous->wm_lp[1] &= ~WM1_LP_SR_EN;
2632                 I915_WRITE(WM2_LP_ILK, previous->wm_lp[1]);
2633                 changed = true;
2634         }
2635         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] & WM1_LP_SR_EN) {
2636                 previous->wm_lp[0] &= ~WM1_LP_SR_EN;
2637                 I915_WRITE(WM1_LP_ILK, previous->wm_lp[0]);
2638                 changed = true;
2639         }
2640
2641         /*
2642          * Don't touch WM1S_LP_EN here.
2643          * Doing so could cause underruns.
2644          */
2645
2646         return changed;
2647 }
2648
2649 /*
2650  * The spec says we shouldn't write when we don't need, because every write
2651  * causes WMs to be re-evaluated, expending some power.
2652  */
2653 static void ilk_write_wm_values(struct drm_i915_private *dev_priv,
2654                                 struct ilk_wm_values *results)
2655 {
2656         struct drm_device *dev = dev_priv->dev;
2657         struct ilk_wm_values *previous = &dev_priv->wm.hw;
2658         unsigned int dirty;
2659         uint32_t val;
2660
2661         dirty = ilk_compute_wm_dirty(dev_priv, previous, results);
2662         if (!dirty)
2663                 return;
2664
2665         _ilk_disable_lp_wm(dev_priv, dirty);
2666
2667         if (dirty & WM_DIRTY_PIPE(PIPE_A))
2668                 I915_WRITE(WM0_PIPEA_ILK, results->wm_pipe[0]);
2669         if (dirty & WM_DIRTY_PIPE(PIPE_B))
2670                 I915_WRITE(WM0_PIPEB_ILK, results->wm_pipe[1]);
2671         if (dirty & WM_DIRTY_PIPE(PIPE_C))
2672                 I915_WRITE(WM0_PIPEC_IVB, results->wm_pipe[2]);
2673
2674         if (dirty & WM_DIRTY_LINETIME(PIPE_A))
2675                 I915_WRITE(PIPE_WM_LINETIME(PIPE_A), results->wm_linetime[0]);
2676         if (dirty & WM_DIRTY_LINETIME(PIPE_B))
2677                 I915_WRITE(PIPE_WM_LINETIME(PIPE_B), results->wm_linetime[1]);
2678         if (dirty & WM_DIRTY_LINETIME(PIPE_C))
2679                 I915_WRITE(PIPE_WM_LINETIME(PIPE_C), results->wm_linetime[2]);
2680
2681         if (dirty & WM_DIRTY_DDB) {
2682                 if (IS_HASWELL(dev) || IS_BROADWELL(dev)) {
2683                         val = I915_READ(WM_MISC);
2684                         if (results->partitioning == INTEL_DDB_PART_1_2)
2685                                 val &= ~WM_MISC_DATA_PARTITION_5_6;
2686                         else
2687                                 val |= WM_MISC_DATA_PARTITION_5_6;
2688                         I915_WRITE(WM_MISC, val);
2689                 } else {
2690                         val = I915_READ(DISP_ARB_CTL2);
2691                         if (results->partitioning == INTEL_DDB_PART_1_2)
2692                                 val &= ~DISP_DATA_PARTITION_5_6;
2693                         else
2694                                 val |= DISP_DATA_PARTITION_5_6;
2695                         I915_WRITE(DISP_ARB_CTL2, val);
2696                 }
2697         }
2698
2699         if (dirty & WM_DIRTY_FBC) {
2700                 val = I915_READ(DISP_ARB_CTL);
2701                 if (results->enable_fbc_wm)
2702                         val &= ~DISP_FBC_WM_DIS;
2703                 else
2704                         val |= DISP_FBC_WM_DIS;
2705                 I915_WRITE(DISP_ARB_CTL, val);
2706         }
2707
2708         if (dirty & WM_DIRTY_LP(1) &&
2709             previous->wm_lp_spr[0] != results->wm_lp_spr[0])
2710                 I915_WRITE(WM1S_LP_ILK, results->wm_lp_spr[0]);
2711
2712         if (INTEL_INFO(dev)->gen >= 7) {
2713                 if (dirty & WM_DIRTY_LP(2) && previous->wm_lp_spr[1] != results->wm_lp_spr[1])
2714                         I915_WRITE(WM2S_LP_IVB, results->wm_lp_spr[1]);
2715                 if (dirty & WM_DIRTY_LP(3) && previous->wm_lp_spr[2] != results->wm_lp_spr[2])
2716                         I915_WRITE(WM3S_LP_IVB, results->wm_lp_spr[2]);
2717         }
2718
2719         if (dirty & WM_DIRTY_LP(1) && previous->wm_lp[0] != results->wm_lp[0])
2720                 I915_WRITE(WM1_LP_ILK, results->wm_lp[0]);
2721         if (dirty & WM_DIRTY_LP(2) && previous->wm_lp[1] != results->wm_lp[1])
2722                 I915_WRITE(WM2_LP_ILK, results->wm_lp[1]);
2723         if (dirty & WM_DIRTY_LP(3) && previous->wm_lp[2] != results->wm_lp[2])
2724                 I915_WRITE(WM3_LP_ILK, results->wm_lp[2]);
2725
2726         dev_priv->wm.hw = *results;
2727 }
2728
2729 static bool ilk_disable_lp_wm(struct drm_device *dev)
2730 {
2731         struct drm_i915_private *dev_priv = dev->dev_private;
2732
2733         return _ilk_disable_lp_wm(dev_priv, WM_DIRTY_LP_ALL);
2734 }
2735
2736 /*
2737  * On gen9, we need to allocate Display Data Buffer (DDB) portions to the
2738  * different active planes.
2739  */
2740
2741 #define SKL_DDB_SIZE            896     /* in blocks */
2742 #define BXT_DDB_SIZE            512
2743
2744 /*
2745  * Return the index of a plane in the SKL DDB and wm result arrays.  Primary
2746  * plane is always in slot 0, cursor is always in slot I915_MAX_PLANES-1, and
2747  * other universal planes are in indices 1..n.  Note that this may leave unused
2748  * indices between the top "sprite" plane and the cursor.
2749  */
2750 static int
2751 skl_wm_plane_id(const struct intel_plane *plane)
2752 {
2753         switch (plane->base.type) {
2754         case DRM_PLANE_TYPE_PRIMARY:
2755                 return 0;
2756         case DRM_PLANE_TYPE_CURSOR:
2757                 return PLANE_CURSOR;
2758         case DRM_PLANE_TYPE_OVERLAY:
2759                 return plane->plane + 1;
2760         default:
2761                 MISSING_CASE(plane->base.type);
2762                 return plane->plane;
2763         }
2764 }
2765
2766 static void
2767 skl_ddb_get_pipe_allocation_limits(struct drm_device *dev,
2768                                    const struct intel_crtc_state *cstate,
2769                                    const struct intel_wm_config *config,
2770                                    struct skl_ddb_entry *alloc /* out */)
2771 {
2772         struct drm_crtc *for_crtc = cstate->base.crtc;
2773         struct drm_crtc *crtc;
2774         unsigned int pipe_size, ddb_size;
2775         int nth_active_pipe;
2776
2777         if (!cstate->base.active) {
2778                 alloc->start = 0;
2779                 alloc->end = 0;
2780                 return;
2781         }
2782
2783         if (IS_BROXTON(dev))
2784                 ddb_size = BXT_DDB_SIZE;
2785         else
2786                 ddb_size = SKL_DDB_SIZE;
2787
2788         ddb_size -= 4; /* 4 blocks for bypass path allocation */
2789
2790         nth_active_pipe = 0;
2791         for_each_crtc(dev, crtc) {
2792                 if (!to_intel_crtc(crtc)->active)
2793                         continue;
2794
2795                 if (crtc == for_crtc)
2796                         break;
2797
2798                 nth_active_pipe++;
2799         }
2800
2801         pipe_size = ddb_size / config->num_pipes_active;
2802         alloc->start = nth_active_pipe * ddb_size / config->num_pipes_active;
2803         alloc->end = alloc->start + pipe_size;
2804 }
2805
2806 static unsigned int skl_cursor_allocation(const struct intel_wm_config *config)
2807 {
2808         if (config->num_pipes_active == 1)
2809                 return 32;
2810
2811         return 8;
2812 }
2813
2814 static void skl_ddb_entry_init_from_hw(struct skl_ddb_entry *entry, u32 reg)
2815 {
2816         entry->start = reg & 0x3ff;
2817         entry->end = (reg >> 16) & 0x3ff;
2818         if (entry->end)
2819                 entry->end += 1;
2820 }
2821
2822 void skl_ddb_get_hw_state(struct drm_i915_private *dev_priv,
2823                           struct skl_ddb_allocation *ddb /* out */)
2824 {
2825         enum pipe pipe;
2826         int plane;
2827         u32 val;
2828
2829         memset(ddb, 0, sizeof(*ddb));
2830
2831         for_each_pipe(dev_priv, pipe) {
2832                 if (!intel_display_power_is_enabled(dev_priv, POWER_DOMAIN_PIPE(pipe)))
2833                         continue;
2834
2835                 for_each_plane(dev_priv, pipe, plane) {
2836                         val = I915_READ(PLANE_BUF_CFG(pipe, plane));
2837                         skl_ddb_entry_init_from_hw(&ddb->plane[pipe][plane],
2838                                                    val);
2839                 }
2840
2841                 val = I915_READ(CUR_BUF_CFG(pipe));
2842                 skl_ddb_entry_init_from_hw(&ddb->plane[pipe][PLANE_CURSOR],
2843                                            val);
2844         }
2845 }
2846
2847 static unsigned int
2848 skl_plane_relative_data_rate(const struct intel_crtc_state *cstate,
2849                              const struct drm_plane_state *pstate,
2850                              int y)
2851 {
2852         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
2853         struct drm_framebuffer *fb = pstate->fb;
2854
2855         /* for planar format */
2856         if (fb->pixel_format == DRM_FORMAT_NV12) {
2857                 if (y)  /* y-plane data rate */
2858                         return intel_crtc->config->pipe_src_w *
2859                                 intel_crtc->config->pipe_src_h *
2860                                 drm_format_plane_cpp(fb->pixel_format, 0);
2861                 else    /* uv-plane data rate */
2862                         return (intel_crtc->config->pipe_src_w/2) *
2863                                 (intel_crtc->config->pipe_src_h/2) *
2864                                 drm_format_plane_cpp(fb->pixel_format, 1);
2865         }
2866
2867         /* for packed formats */
2868         return intel_crtc->config->pipe_src_w *
2869                 intel_crtc->config->pipe_src_h *
2870                 drm_format_plane_cpp(fb->pixel_format, 0);
2871 }
2872
2873 /*
2874  * We don't overflow 32 bits. Worst case is 3 planes enabled, each fetching
2875  * a 8192x4096@32bpp framebuffer:
2876  *   3 * 4096 * 8192  * 4 < 2^32
2877  */
2878 static unsigned int
2879 skl_get_total_relative_data_rate(const struct intel_crtc_state *cstate)
2880 {
2881         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
2882         struct drm_device *dev = intel_crtc->base.dev;
2883         const struct intel_plane *intel_plane;
2884         unsigned int total_data_rate = 0;
2885
2886         for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
2887                 const struct drm_plane_state *pstate = intel_plane->base.state;
2888
2889                 if (pstate->fb == NULL)
2890                         continue;
2891
2892                 if (intel_plane->base.type == DRM_PLANE_TYPE_CURSOR)
2893                         continue;
2894
2895                 /* packed/uv */
2896                 total_data_rate += skl_plane_relative_data_rate(cstate,
2897                                                                 pstate,
2898                                                                 0);
2899
2900                 if (pstate->fb->pixel_format == DRM_FORMAT_NV12)
2901                         /* y-plane */
2902                         total_data_rate += skl_plane_relative_data_rate(cstate,
2903                                                                         pstate,
2904                                                                         1);
2905         }
2906
2907         return total_data_rate;
2908 }
2909
2910 static void
2911 skl_allocate_pipe_ddb(struct intel_crtc_state *cstate,
2912                       struct skl_ddb_allocation *ddb /* out */)
2913 {
2914         struct drm_crtc *crtc = cstate->base.crtc;
2915         struct drm_device *dev = crtc->dev;
2916         struct drm_i915_private *dev_priv = to_i915(dev);
2917         struct intel_wm_config *config = &dev_priv->wm.config;
2918         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
2919         struct intel_plane *intel_plane;
2920         enum pipe pipe = intel_crtc->pipe;
2921         struct skl_ddb_entry *alloc = &ddb->pipe[pipe];
2922         uint16_t alloc_size, start, cursor_blocks;
2923         uint16_t minimum[I915_MAX_PLANES];
2924         uint16_t y_minimum[I915_MAX_PLANES];
2925         unsigned int total_data_rate;
2926
2927         skl_ddb_get_pipe_allocation_limits(dev, cstate, config, alloc);
2928         alloc_size = skl_ddb_entry_size(alloc);
2929         if (alloc_size == 0) {
2930                 memset(ddb->plane[pipe], 0, sizeof(ddb->plane[pipe]));
2931                 memset(&ddb->plane[pipe][PLANE_CURSOR], 0,
2932                        sizeof(ddb->plane[pipe][PLANE_CURSOR]));
2933                 return;
2934         }
2935
2936         cursor_blocks = skl_cursor_allocation(config);
2937         ddb->plane[pipe][PLANE_CURSOR].start = alloc->end - cursor_blocks;
2938         ddb->plane[pipe][PLANE_CURSOR].end = alloc->end;
2939
2940         alloc_size -= cursor_blocks;
2941         alloc->end -= cursor_blocks;
2942
2943         /* 1. Allocate the mininum required blocks for each active plane */
2944         for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
2945                 struct drm_plane *plane = &intel_plane->base;
2946                 struct drm_framebuffer *fb = plane->state->fb;
2947                 int id = skl_wm_plane_id(intel_plane);
2948
2949                 if (fb == NULL)
2950                         continue;
2951                 if (plane->type == DRM_PLANE_TYPE_CURSOR)
2952                         continue;
2953
2954                 minimum[id] = 8;
2955                 alloc_size -= minimum[id];
2956                 y_minimum[id] = (fb->pixel_format == DRM_FORMAT_NV12) ? 8 : 0;
2957                 alloc_size -= y_minimum[id];
2958         }
2959
2960         /*
2961          * 2. Distribute the remaining space in proportion to the amount of
2962          * data each plane needs to fetch from memory.
2963          *
2964          * FIXME: we may not allocate every single block here.
2965          */
2966         total_data_rate = skl_get_total_relative_data_rate(cstate);
2967
2968         start = alloc->start;
2969         for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
2970                 struct drm_plane *plane = &intel_plane->base;
2971                 struct drm_plane_state *pstate = intel_plane->base.state;
2972                 unsigned int data_rate, y_data_rate;
2973                 uint16_t plane_blocks, y_plane_blocks = 0;
2974                 int id = skl_wm_plane_id(intel_plane);
2975
2976                 if (pstate->fb == NULL)
2977                         continue;
2978                 if (plane->type == DRM_PLANE_TYPE_CURSOR)
2979                         continue;
2980
2981                 data_rate = skl_plane_relative_data_rate(cstate, pstate, 0);
2982
2983                 /*
2984                  * allocation for (packed formats) or (uv-plane part of planar format):
2985                  * promote the expression to 64 bits to avoid overflowing, the
2986                  * result is < available as data_rate / total_data_rate < 1
2987                  */
2988                 plane_blocks = minimum[id];
2989                 plane_blocks += div_u64((uint64_t)alloc_size * data_rate,
2990                                         total_data_rate);
2991
2992                 ddb->plane[pipe][id].start = start;
2993                 ddb->plane[pipe][id].end = start + plane_blocks;
2994
2995                 start += plane_blocks;
2996
2997                 /*
2998                  * allocation for y_plane part of planar format:
2999                  */
3000                 if (pstate->fb->pixel_format == DRM_FORMAT_NV12) {
3001                         y_data_rate = skl_plane_relative_data_rate(cstate,
3002                                                                    pstate,
3003                                                                    1);
3004                         y_plane_blocks = y_minimum[id];
3005                         y_plane_blocks += div_u64((uint64_t)alloc_size * y_data_rate,
3006                                                 total_data_rate);
3007
3008                         ddb->y_plane[pipe][id].start = start;
3009                         ddb->y_plane[pipe][id].end = start + y_plane_blocks;
3010
3011                         start += y_plane_blocks;
3012                 }
3013
3014         }
3015
3016 }
3017
3018 static uint32_t skl_pipe_pixel_rate(const struct intel_crtc_state *config)
3019 {
3020         /* TODO: Take into account the scalers once we support them */
3021         return config->base.adjusted_mode.crtc_clock;
3022 }
3023
3024 /*
3025  * The max latency should be 257 (max the punit can code is 255 and we add 2us
3026  * for the read latency) and bytes_per_pixel should always be <= 8, so that
3027  * should allow pixel_rate up to ~2 GHz which seems sufficient since max
3028  * 2xcdclk is 1350 MHz and the pixel rate should never exceed that.
3029 */
3030 static uint32_t skl_wm_method1(uint32_t pixel_rate, uint8_t bytes_per_pixel,
3031                                uint32_t latency)
3032 {
3033         uint32_t wm_intermediate_val, ret;
3034
3035         if (latency == 0)
3036                 return UINT_MAX;
3037
3038         wm_intermediate_val = latency * pixel_rate * bytes_per_pixel / 512;
3039         ret = DIV_ROUND_UP(wm_intermediate_val, 1000);
3040
3041         return ret;
3042 }
3043
3044 static uint32_t skl_wm_method2(uint32_t pixel_rate, uint32_t pipe_htotal,
3045                                uint32_t horiz_pixels, uint8_t bytes_per_pixel,
3046                                uint64_t tiling, uint32_t latency)
3047 {
3048         uint32_t ret;
3049         uint32_t plane_bytes_per_line, plane_blocks_per_line;
3050         uint32_t wm_intermediate_val;
3051
3052         if (latency == 0)
3053                 return UINT_MAX;
3054
3055         plane_bytes_per_line = horiz_pixels * bytes_per_pixel;
3056
3057         if (tiling == I915_FORMAT_MOD_Y_TILED ||
3058             tiling == I915_FORMAT_MOD_Yf_TILED) {
3059                 plane_bytes_per_line *= 4;
3060                 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
3061                 plane_blocks_per_line /= 4;
3062         } else {
3063                 plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
3064         }
3065
3066         wm_intermediate_val = latency * pixel_rate;
3067         ret = DIV_ROUND_UP(wm_intermediate_val, pipe_htotal * 1000) *
3068                                 plane_blocks_per_line;
3069
3070         return ret;
3071 }
3072
3073 static bool skl_ddb_allocation_changed(const struct skl_ddb_allocation *new_ddb,
3074                                        const struct intel_crtc *intel_crtc)
3075 {
3076         struct drm_device *dev = intel_crtc->base.dev;
3077         struct drm_i915_private *dev_priv = dev->dev_private;
3078         const struct skl_ddb_allocation *cur_ddb = &dev_priv->wm.skl_hw.ddb;
3079
3080         /*
3081          * If ddb allocation of pipes changed, it may require recalculation of
3082          * watermarks
3083          */
3084         if (memcmp(new_ddb->pipe, cur_ddb->pipe, sizeof(new_ddb->pipe)))
3085                 return true;
3086
3087         return false;
3088 }
3089
3090 static bool skl_compute_plane_wm(const struct drm_i915_private *dev_priv,
3091                                  struct intel_crtc_state *cstate,
3092                                  struct intel_plane *intel_plane,
3093                                  uint16_t ddb_allocation,
3094                                  int level,
3095                                  uint16_t *out_blocks, /* out */
3096                                  uint8_t *out_lines /* out */)
3097 {
3098         struct drm_plane *plane = &intel_plane->base;
3099         struct drm_framebuffer *fb = plane->state->fb;
3100         uint32_t latency = dev_priv->wm.skl_latency[level];
3101         uint32_t method1, method2;
3102         uint32_t plane_bytes_per_line, plane_blocks_per_line;
3103         uint32_t res_blocks, res_lines;
3104         uint32_t selected_result;
3105         uint8_t bytes_per_pixel;
3106
3107         if (latency == 0 || !cstate->base.active || !fb)
3108                 return false;
3109
3110         bytes_per_pixel = drm_format_plane_cpp(fb->pixel_format, 0);
3111         method1 = skl_wm_method1(skl_pipe_pixel_rate(cstate),
3112                                  bytes_per_pixel,
3113                                  latency);
3114         method2 = skl_wm_method2(skl_pipe_pixel_rate(cstate),
3115                                  cstate->base.adjusted_mode.crtc_htotal,
3116                                  cstate->pipe_src_w,
3117                                  bytes_per_pixel,
3118                                  fb->modifier[0],
3119                                  latency);
3120
3121         plane_bytes_per_line = cstate->pipe_src_w * bytes_per_pixel;
3122         plane_blocks_per_line = DIV_ROUND_UP(plane_bytes_per_line, 512);
3123
3124         if (fb->modifier[0] == I915_FORMAT_MOD_Y_TILED ||
3125             fb->modifier[0] == I915_FORMAT_MOD_Yf_TILED) {
3126                 uint32_t min_scanlines = 4;
3127                 uint32_t y_tile_minimum;
3128                 if (intel_rotation_90_or_270(plane->state->rotation)) {
3129                         int bpp = (fb->pixel_format == DRM_FORMAT_NV12) ?
3130                                 drm_format_plane_cpp(fb->pixel_format, 1) :
3131                                 drm_format_plane_cpp(fb->pixel_format, 0);
3132
3133                         switch (bpp) {
3134                         case 1:
3135                                 min_scanlines = 16;
3136                                 break;
3137                         case 2:
3138                                 min_scanlines = 8;
3139                                 break;
3140                         case 8:
3141                                 WARN(1, "Unsupported pixel depth for rotation");
3142                         }
3143                 }
3144                 y_tile_minimum = plane_blocks_per_line * min_scanlines;
3145                 selected_result = max(method2, y_tile_minimum);
3146         } else {
3147                 if ((ddb_allocation / plane_blocks_per_line) >= 1)
3148                         selected_result = min(method1, method2);
3149                 else
3150                         selected_result = method1;
3151         }
3152
3153         res_blocks = selected_result + 1;
3154         res_lines = DIV_ROUND_UP(selected_result, plane_blocks_per_line);
3155
3156         if (level >= 1 && level <= 7) {
3157                 if (fb->modifier[0] == I915_FORMAT_MOD_Y_TILED ||
3158                     fb->modifier[0] == I915_FORMAT_MOD_Yf_TILED)
3159                         res_lines += 4;
3160                 else
3161                         res_blocks++;
3162         }
3163
3164         if (res_blocks >= ddb_allocation || res_lines > 31)
3165                 return false;
3166
3167         *out_blocks = res_blocks;
3168         *out_lines = res_lines;
3169
3170         return true;
3171 }
3172
3173 static void skl_compute_wm_level(const struct drm_i915_private *dev_priv,
3174                                  struct skl_ddb_allocation *ddb,
3175                                  struct intel_crtc_state *cstate,
3176                                  int level,
3177                                  struct skl_wm_level *result)
3178 {
3179         struct drm_device *dev = dev_priv->dev;
3180         struct intel_crtc *intel_crtc = to_intel_crtc(cstate->base.crtc);
3181         struct intel_plane *intel_plane;
3182         uint16_t ddb_blocks;
3183         enum pipe pipe = intel_crtc->pipe;
3184
3185         for_each_intel_plane_on_crtc(dev, intel_crtc, intel_plane) {
3186                 int i = skl_wm_plane_id(intel_plane);
3187
3188                 ddb_blocks = skl_ddb_entry_size(&ddb->plane[pipe][i]);
3189
3190                 result->plane_en[i] = skl_compute_plane_wm(dev_priv,
3191                                                 cstate,
3192                                                 intel_plane,
3193                                                 ddb_blocks,
3194                                                 level,
3195                                                 &result->plane_res_b[i],
3196                                                 &result->plane_res_l[i]);
3197         }
3198 }
3199
3200 static uint32_t
3201 skl_compute_linetime_wm(struct intel_crtc_state *cstate)
3202 {
3203         if (!cstate->base.active)
3204                 return 0;
3205
3206         if (WARN_ON(skl_pipe_pixel_rate(cstate) == 0))
3207                 return 0;
3208
3209         return DIV_ROUND_UP(8 * cstate->base.adjusted_mode.crtc_htotal * 1000,
3210                             skl_pipe_pixel_rate(cstate));
3211 }
3212
3213 static void skl_compute_transition_wm(struct intel_crtc_state *cstate,
3214                                       struct skl_wm_level *trans_wm /* out */)
3215 {
3216         struct drm_crtc *crtc = cstate->base.crtc;
3217         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3218         struct intel_plane *intel_plane;
3219
3220         if (!cstate->base.active)
3221                 return;
3222
3223         /* Until we know more, just disable transition WMs */
3224         for_each_intel_plane_on_crtc(crtc->dev, intel_crtc, intel_plane) {
3225                 int i = skl_wm_plane_id(intel_plane);
3226
3227                 trans_wm->plane_en[i] = false;
3228         }
3229 }
3230
3231 static void skl_compute_pipe_wm(struct intel_crtc_state *cstate,
3232                                 struct skl_ddb_allocation *ddb,
3233                                 struct skl_pipe_wm *pipe_wm)
3234 {
3235         struct drm_device *dev = cstate->base.crtc->dev;
3236         const struct drm_i915_private *dev_priv = dev->dev_private;
3237         int level, max_level = ilk_wm_max_level(dev);
3238
3239         for (level = 0; level <= max_level; level++) {
3240                 skl_compute_wm_level(dev_priv, ddb, cstate,
3241                                      level, &pipe_wm->wm[level]);
3242         }
3243         pipe_wm->linetime = skl_compute_linetime_wm(cstate);
3244
3245         skl_compute_transition_wm(cstate, &pipe_wm->trans_wm);
3246 }
3247
3248 static void skl_compute_wm_results(struct drm_device *dev,
3249                                    struct skl_pipe_wm *p_wm,
3250                                    struct skl_wm_values *r,
3251                                    struct intel_crtc *intel_crtc)
3252 {
3253         int level, max_level = ilk_wm_max_level(dev);
3254         enum pipe pipe = intel_crtc->pipe;
3255         uint32_t temp;
3256         int i;
3257
3258         for (level = 0; level <= max_level; level++) {
3259                 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3260                         temp = 0;
3261
3262                         temp |= p_wm->wm[level].plane_res_l[i] <<
3263                                         PLANE_WM_LINES_SHIFT;
3264                         temp |= p_wm->wm[level].plane_res_b[i];
3265                         if (p_wm->wm[level].plane_en[i])
3266                                 temp |= PLANE_WM_EN;
3267
3268                         r->plane[pipe][i][level] = temp;
3269                 }
3270
3271                 temp = 0;
3272
3273                 temp |= p_wm->wm[level].plane_res_l[PLANE_CURSOR] << PLANE_WM_LINES_SHIFT;
3274                 temp |= p_wm->wm[level].plane_res_b[PLANE_CURSOR];
3275
3276                 if (p_wm->wm[level].plane_en[PLANE_CURSOR])
3277                         temp |= PLANE_WM_EN;
3278
3279                 r->plane[pipe][PLANE_CURSOR][level] = temp;
3280
3281         }
3282
3283         /* transition WMs */
3284         for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3285                 temp = 0;
3286                 temp |= p_wm->trans_wm.plane_res_l[i] << PLANE_WM_LINES_SHIFT;
3287                 temp |= p_wm->trans_wm.plane_res_b[i];
3288                 if (p_wm->trans_wm.plane_en[i])
3289                         temp |= PLANE_WM_EN;
3290
3291                 r->plane_trans[pipe][i] = temp;
3292         }
3293
3294         temp = 0;
3295         temp |= p_wm->trans_wm.plane_res_l[PLANE_CURSOR] << PLANE_WM_LINES_SHIFT;
3296         temp |= p_wm->trans_wm.plane_res_b[PLANE_CURSOR];
3297         if (p_wm->trans_wm.plane_en[PLANE_CURSOR])
3298                 temp |= PLANE_WM_EN;
3299
3300         r->plane_trans[pipe][PLANE_CURSOR] = temp;
3301
3302         r->wm_linetime[pipe] = p_wm->linetime;
3303 }
3304
3305 static void skl_ddb_entry_write(struct drm_i915_private *dev_priv,
3306                                 i915_reg_t reg,
3307                                 const struct skl_ddb_entry *entry)
3308 {
3309         if (entry->end)
3310                 I915_WRITE(reg, (entry->end - 1) << 16 | entry->start);
3311         else
3312                 I915_WRITE(reg, 0);
3313 }
3314
3315 static void skl_write_wm_values(struct drm_i915_private *dev_priv,
3316                                 const struct skl_wm_values *new)
3317 {
3318         struct drm_device *dev = dev_priv->dev;
3319         struct intel_crtc *crtc;
3320
3321         for_each_intel_crtc(dev, crtc) {
3322                 int i, level, max_level = ilk_wm_max_level(dev);
3323                 enum pipe pipe = crtc->pipe;
3324
3325                 if (!new->dirty[pipe])
3326                         continue;
3327
3328                 I915_WRITE(PIPE_WM_LINETIME(pipe), new->wm_linetime[pipe]);
3329
3330                 for (level = 0; level <= max_level; level++) {
3331                         for (i = 0; i < intel_num_planes(crtc); i++)
3332                                 I915_WRITE(PLANE_WM(pipe, i, level),
3333                                            new->plane[pipe][i][level]);
3334                         I915_WRITE(CUR_WM(pipe, level),
3335                                    new->plane[pipe][PLANE_CURSOR][level]);
3336                 }
3337                 for (i = 0; i < intel_num_planes(crtc); i++)
3338                         I915_WRITE(PLANE_WM_TRANS(pipe, i),
3339                                    new->plane_trans[pipe][i]);
3340                 I915_WRITE(CUR_WM_TRANS(pipe),
3341                            new->plane_trans[pipe][PLANE_CURSOR]);
3342
3343                 for (i = 0; i < intel_num_planes(crtc); i++) {
3344                         skl_ddb_entry_write(dev_priv,
3345                                             PLANE_BUF_CFG(pipe, i),
3346                                             &new->ddb.plane[pipe][i]);
3347                         skl_ddb_entry_write(dev_priv,
3348                                             PLANE_NV12_BUF_CFG(pipe, i),
3349                                             &new->ddb.y_plane[pipe][i]);
3350                 }
3351
3352                 skl_ddb_entry_write(dev_priv, CUR_BUF_CFG(pipe),
3353                                     &new->ddb.plane[pipe][PLANE_CURSOR]);
3354         }
3355 }
3356
3357 /*
3358  * When setting up a new DDB allocation arrangement, we need to correctly
3359  * sequence the times at which the new allocations for the pipes are taken into
3360  * account or we'll have pipes fetching from space previously allocated to
3361  * another pipe.
3362  *
3363  * Roughly the sequence looks like:
3364  *  1. re-allocate the pipe(s) with the allocation being reduced and not
3365  *     overlapping with a previous light-up pipe (another way to put it is:
3366  *     pipes with their new allocation strickly included into their old ones).
3367  *  2. re-allocate the other pipes that get their allocation reduced
3368  *  3. allocate the pipes having their allocation increased
3369  *
3370  * Steps 1. and 2. are here to take care of the following case:
3371  * - Initially DDB looks like this:
3372  *     |   B    |   C    |
3373  * - enable pipe A.
3374  * - pipe B has a reduced DDB allocation that overlaps with the old pipe C
3375  *   allocation
3376  *     |  A  |  B  |  C  |
3377  *
3378  * We need to sequence the re-allocation: C, B, A (and not B, C, A).
3379  */
3380
3381 static void
3382 skl_wm_flush_pipe(struct drm_i915_private *dev_priv, enum pipe pipe, int pass)
3383 {
3384         int plane;
3385
3386         DRM_DEBUG_KMS("flush pipe %c (pass %d)\n", pipe_name(pipe), pass);
3387
3388         for_each_plane(dev_priv, pipe, plane) {
3389                 I915_WRITE(PLANE_SURF(pipe, plane),
3390                            I915_READ(PLANE_SURF(pipe, plane)));
3391         }
3392         I915_WRITE(CURBASE(pipe), I915_READ(CURBASE(pipe)));
3393 }
3394
3395 static bool
3396 skl_ddb_allocation_included(const struct skl_ddb_allocation *old,
3397                             const struct skl_ddb_allocation *new,
3398                             enum pipe pipe)
3399 {
3400         uint16_t old_size, new_size;
3401
3402         old_size = skl_ddb_entry_size(&old->pipe[pipe]);
3403         new_size = skl_ddb_entry_size(&new->pipe[pipe]);
3404
3405         return old_size != new_size &&
3406                new->pipe[pipe].start >= old->pipe[pipe].start &&
3407                new->pipe[pipe].end <= old->pipe[pipe].end;
3408 }
3409
3410 static void skl_flush_wm_values(struct drm_i915_private *dev_priv,
3411                                 struct skl_wm_values *new_values)
3412 {
3413         struct drm_device *dev = dev_priv->dev;
3414         struct skl_ddb_allocation *cur_ddb, *new_ddb;
3415         bool reallocated[I915_MAX_PIPES] = {};
3416         struct intel_crtc *crtc;
3417         enum pipe pipe;
3418
3419         new_ddb = &new_values->ddb;
3420         cur_ddb = &dev_priv->wm.skl_hw.ddb;
3421
3422         /*
3423          * First pass: flush the pipes with the new allocation contained into
3424          * the old space.
3425          *
3426          * We'll wait for the vblank on those pipes to ensure we can safely
3427          * re-allocate the freed space without this pipe fetching from it.
3428          */
3429         for_each_intel_crtc(dev, crtc) {
3430                 if (!crtc->active)
3431                         continue;
3432
3433                 pipe = crtc->pipe;
3434
3435                 if (!skl_ddb_allocation_included(cur_ddb, new_ddb, pipe))
3436                         continue;
3437
3438                 skl_wm_flush_pipe(dev_priv, pipe, 1);
3439                 intel_wait_for_vblank(dev, pipe);
3440
3441                 reallocated[pipe] = true;
3442         }
3443
3444
3445         /*
3446          * Second pass: flush the pipes that are having their allocation
3447          * reduced, but overlapping with a previous allocation.
3448          *
3449          * Here as well we need to wait for the vblank to make sure the freed
3450          * space is not used anymore.
3451          */
3452         for_each_intel_crtc(dev, crtc) {
3453                 if (!crtc->active)
3454                         continue;
3455
3456                 pipe = crtc->pipe;
3457
3458                 if (reallocated[pipe])
3459                         continue;
3460
3461                 if (skl_ddb_entry_size(&new_ddb->pipe[pipe]) <
3462                     skl_ddb_entry_size(&cur_ddb->pipe[pipe])) {
3463                         skl_wm_flush_pipe(dev_priv, pipe, 2);
3464                         intel_wait_for_vblank(dev, pipe);
3465                         reallocated[pipe] = true;
3466                 }
3467         }
3468
3469         /*
3470          * Third pass: flush the pipes that got more space allocated.
3471          *
3472          * We don't need to actively wait for the update here, next vblank
3473          * will just get more DDB space with the correct WM values.
3474          */
3475         for_each_intel_crtc(dev, crtc) {
3476                 if (!crtc->active)
3477                         continue;
3478
3479                 pipe = crtc->pipe;
3480
3481                 /*
3482                  * At this point, only the pipes more space than before are
3483                  * left to re-allocate.
3484                  */
3485                 if (reallocated[pipe])
3486                         continue;
3487
3488                 skl_wm_flush_pipe(dev_priv, pipe, 3);
3489         }
3490 }
3491
3492 static bool skl_update_pipe_wm(struct drm_crtc *crtc,
3493                                struct skl_ddb_allocation *ddb, /* out */
3494                                struct skl_pipe_wm *pipe_wm /* out */)
3495 {
3496         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3497         struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
3498
3499         skl_allocate_pipe_ddb(cstate, ddb);
3500         skl_compute_pipe_wm(cstate, ddb, pipe_wm);
3501
3502         if (!memcmp(&intel_crtc->wm.active.skl, pipe_wm, sizeof(*pipe_wm)))
3503                 return false;
3504
3505         intel_crtc->wm.active.skl = *pipe_wm;
3506
3507         return true;
3508 }
3509
3510 static void skl_update_other_pipe_wm(struct drm_device *dev,
3511                                      struct drm_crtc *crtc,
3512                                      struct skl_wm_values *r)
3513 {
3514         struct intel_crtc *intel_crtc;
3515         struct intel_crtc *this_crtc = to_intel_crtc(crtc);
3516
3517         /*
3518          * If the WM update hasn't changed the allocation for this_crtc (the
3519          * crtc we are currently computing the new WM values for), other
3520          * enabled crtcs will keep the same allocation and we don't need to
3521          * recompute anything for them.
3522          */
3523         if (!skl_ddb_allocation_changed(&r->ddb, this_crtc))
3524                 return;
3525
3526         /*
3527          * Otherwise, because of this_crtc being freshly enabled/disabled, the
3528          * other active pipes need new DDB allocation and WM values.
3529          */
3530         for_each_intel_crtc(dev, intel_crtc) {
3531                 struct skl_pipe_wm pipe_wm = {};
3532                 bool wm_changed;
3533
3534                 if (this_crtc->pipe == intel_crtc->pipe)
3535                         continue;
3536
3537                 if (!intel_crtc->active)
3538                         continue;
3539
3540                 wm_changed = skl_update_pipe_wm(&intel_crtc->base,
3541                                                 &r->ddb, &pipe_wm);
3542
3543                 /*
3544                  * If we end up re-computing the other pipe WM values, it's
3545                  * because it was really needed, so we expect the WM values to
3546                  * be different.
3547                  */
3548                 WARN_ON(!wm_changed);
3549
3550                 skl_compute_wm_results(dev, &pipe_wm, r, intel_crtc);
3551                 r->dirty[intel_crtc->pipe] = true;
3552         }
3553 }
3554
3555 static void skl_clear_wm(struct skl_wm_values *watermarks, enum pipe pipe)
3556 {
3557         watermarks->wm_linetime[pipe] = 0;
3558         memset(watermarks->plane[pipe], 0,
3559                sizeof(uint32_t) * 8 * I915_MAX_PLANES);
3560         memset(watermarks->plane_trans[pipe],
3561                0, sizeof(uint32_t) * I915_MAX_PLANES);
3562         watermarks->plane_trans[pipe][PLANE_CURSOR] = 0;
3563
3564         /* Clear ddb entries for pipe */
3565         memset(&watermarks->ddb.pipe[pipe], 0, sizeof(struct skl_ddb_entry));
3566         memset(&watermarks->ddb.plane[pipe], 0,
3567                sizeof(struct skl_ddb_entry) * I915_MAX_PLANES);
3568         memset(&watermarks->ddb.y_plane[pipe], 0,
3569                sizeof(struct skl_ddb_entry) * I915_MAX_PLANES);
3570         memset(&watermarks->ddb.plane[pipe][PLANE_CURSOR], 0,
3571                sizeof(struct skl_ddb_entry));
3572
3573 }
3574
3575 static void skl_update_wm(struct drm_crtc *crtc)
3576 {
3577         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3578         struct drm_device *dev = crtc->dev;
3579         struct drm_i915_private *dev_priv = dev->dev_private;
3580         struct skl_wm_values *results = &dev_priv->wm.skl_results;
3581         struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
3582         struct skl_pipe_wm *pipe_wm = &cstate->wm.optimal.skl;
3583
3584
3585         /* Clear all dirty flags */
3586         memset(results->dirty, 0, sizeof(bool) * I915_MAX_PIPES);
3587
3588         skl_clear_wm(results, intel_crtc->pipe);
3589
3590         if (!skl_update_pipe_wm(crtc, &results->ddb, pipe_wm))
3591                 return;
3592
3593         skl_compute_wm_results(dev, pipe_wm, results, intel_crtc);
3594         results->dirty[intel_crtc->pipe] = true;
3595
3596         skl_update_other_pipe_wm(dev, crtc, results);
3597         skl_write_wm_values(dev_priv, results);
3598         skl_flush_wm_values(dev_priv, results);
3599
3600         /* store the new configuration */
3601         dev_priv->wm.skl_hw = *results;
3602 }
3603
3604 static void ilk_program_watermarks(struct drm_i915_private *dev_priv)
3605 {
3606         struct drm_device *dev = dev_priv->dev;
3607         struct intel_pipe_wm lp_wm_1_2 = {}, lp_wm_5_6 = {}, *best_lp_wm;
3608         struct ilk_wm_maximums max;
3609         struct intel_wm_config *config = &dev_priv->wm.config;
3610         struct ilk_wm_values results = {};
3611         enum intel_ddb_partitioning partitioning;
3612
3613         ilk_compute_wm_maximums(dev, 1, config, INTEL_DDB_PART_1_2, &max);
3614         ilk_wm_merge(dev, config, &max, &lp_wm_1_2);
3615
3616         /* 5/6 split only in single pipe config on IVB+ */
3617         if (INTEL_INFO(dev)->gen >= 7 &&
3618             config->num_pipes_active == 1 && config->sprites_enabled) {
3619                 ilk_compute_wm_maximums(dev, 1, config, INTEL_DDB_PART_5_6, &max);
3620                 ilk_wm_merge(dev, config, &max, &lp_wm_5_6);
3621
3622                 best_lp_wm = ilk_find_best_result(dev, &lp_wm_1_2, &lp_wm_5_6);
3623         } else {
3624                 best_lp_wm = &lp_wm_1_2;
3625         }
3626
3627         partitioning = (best_lp_wm == &lp_wm_1_2) ?
3628                        INTEL_DDB_PART_1_2 : INTEL_DDB_PART_5_6;
3629
3630         ilk_compute_wm_results(dev, best_lp_wm, partitioning, &results);
3631
3632         ilk_write_wm_values(dev_priv, &results);
3633 }
3634
3635 static void ilk_update_wm(struct drm_crtc *crtc)
3636 {
3637         struct drm_i915_private *dev_priv = to_i915(crtc->dev);
3638         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3639         struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
3640
3641         WARN_ON(cstate->base.active != intel_crtc->active);
3642
3643         /*
3644          * IVB workaround: must disable low power watermarks for at least
3645          * one frame before enabling scaling.  LP watermarks can be re-enabled
3646          * when scaling is disabled.
3647          *
3648          * WaCxSRDisabledForSpriteScaling:ivb
3649          */
3650         if (cstate->disable_lp_wm) {
3651                 ilk_disable_lp_wm(crtc->dev);
3652                 intel_wait_for_vblank(crtc->dev, intel_crtc->pipe);
3653         }
3654
3655         intel_crtc->wm.active.ilk = cstate->wm.optimal.ilk;
3656
3657         ilk_program_watermarks(dev_priv);
3658 }
3659
3660 static void skl_pipe_wm_active_state(uint32_t val,
3661                                      struct skl_pipe_wm *active,
3662                                      bool is_transwm,
3663                                      bool is_cursor,
3664                                      int i,
3665                                      int level)
3666 {
3667         bool is_enabled = (val & PLANE_WM_EN) != 0;
3668
3669         if (!is_transwm) {
3670                 if (!is_cursor) {
3671                         active->wm[level].plane_en[i] = is_enabled;
3672                         active->wm[level].plane_res_b[i] =
3673                                         val & PLANE_WM_BLOCKS_MASK;
3674                         active->wm[level].plane_res_l[i] =
3675                                         (val >> PLANE_WM_LINES_SHIFT) &
3676                                                 PLANE_WM_LINES_MASK;
3677                 } else {
3678                         active->wm[level].plane_en[PLANE_CURSOR] = is_enabled;
3679                         active->wm[level].plane_res_b[PLANE_CURSOR] =
3680                                         val & PLANE_WM_BLOCKS_MASK;
3681                         active->wm[level].plane_res_l[PLANE_CURSOR] =
3682                                         (val >> PLANE_WM_LINES_SHIFT) &
3683                                                 PLANE_WM_LINES_MASK;
3684                 }
3685         } else {
3686                 if (!is_cursor) {
3687                         active->trans_wm.plane_en[i] = is_enabled;
3688                         active->trans_wm.plane_res_b[i] =
3689                                         val & PLANE_WM_BLOCKS_MASK;
3690                         active->trans_wm.plane_res_l[i] =
3691                                         (val >> PLANE_WM_LINES_SHIFT) &
3692                                                 PLANE_WM_LINES_MASK;
3693                 } else {
3694                         active->trans_wm.plane_en[PLANE_CURSOR] = is_enabled;
3695                         active->trans_wm.plane_res_b[PLANE_CURSOR] =
3696                                         val & PLANE_WM_BLOCKS_MASK;
3697                         active->trans_wm.plane_res_l[PLANE_CURSOR] =
3698                                         (val >> PLANE_WM_LINES_SHIFT) &
3699                                                 PLANE_WM_LINES_MASK;
3700                 }
3701         }
3702 }
3703
3704 static void skl_pipe_wm_get_hw_state(struct drm_crtc *crtc)
3705 {
3706         struct drm_device *dev = crtc->dev;
3707         struct drm_i915_private *dev_priv = dev->dev_private;
3708         struct skl_wm_values *hw = &dev_priv->wm.skl_hw;
3709         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3710         struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
3711         struct skl_pipe_wm *active = &cstate->wm.optimal.skl;
3712         enum pipe pipe = intel_crtc->pipe;
3713         int level, i, max_level;
3714         uint32_t temp;
3715
3716         max_level = ilk_wm_max_level(dev);
3717
3718         hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
3719
3720         for (level = 0; level <= max_level; level++) {
3721                 for (i = 0; i < intel_num_planes(intel_crtc); i++)
3722                         hw->plane[pipe][i][level] =
3723                                         I915_READ(PLANE_WM(pipe, i, level));
3724                 hw->plane[pipe][PLANE_CURSOR][level] = I915_READ(CUR_WM(pipe, level));
3725         }
3726
3727         for (i = 0; i < intel_num_planes(intel_crtc); i++)
3728                 hw->plane_trans[pipe][i] = I915_READ(PLANE_WM_TRANS(pipe, i));
3729         hw->plane_trans[pipe][PLANE_CURSOR] = I915_READ(CUR_WM_TRANS(pipe));
3730
3731         if (!intel_crtc->active)
3732                 return;
3733
3734         hw->dirty[pipe] = true;
3735
3736         active->linetime = hw->wm_linetime[pipe];
3737
3738         for (level = 0; level <= max_level; level++) {
3739                 for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3740                         temp = hw->plane[pipe][i][level];
3741                         skl_pipe_wm_active_state(temp, active, false,
3742                                                 false, i, level);
3743                 }
3744                 temp = hw->plane[pipe][PLANE_CURSOR][level];
3745                 skl_pipe_wm_active_state(temp, active, false, true, i, level);
3746         }
3747
3748         for (i = 0; i < intel_num_planes(intel_crtc); i++) {
3749                 temp = hw->plane_trans[pipe][i];
3750                 skl_pipe_wm_active_state(temp, active, true, false, i, 0);
3751         }
3752
3753         temp = hw->plane_trans[pipe][PLANE_CURSOR];
3754         skl_pipe_wm_active_state(temp, active, true, true, i, 0);
3755
3756         intel_crtc->wm.active.skl = *active;
3757 }
3758
3759 void skl_wm_get_hw_state(struct drm_device *dev)
3760 {
3761         struct drm_i915_private *dev_priv = dev->dev_private;
3762         struct skl_ddb_allocation *ddb = &dev_priv->wm.skl_hw.ddb;
3763         struct drm_crtc *crtc;
3764
3765         skl_ddb_get_hw_state(dev_priv, ddb);
3766         list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
3767                 skl_pipe_wm_get_hw_state(crtc);
3768 }
3769
3770 static void ilk_pipe_wm_get_hw_state(struct drm_crtc *crtc)
3771 {
3772         struct drm_device *dev = crtc->dev;
3773         struct drm_i915_private *dev_priv = dev->dev_private;
3774         struct ilk_wm_values *hw = &dev_priv->wm.hw;
3775         struct intel_crtc *intel_crtc = to_intel_crtc(crtc);
3776         struct intel_crtc_state *cstate = to_intel_crtc_state(crtc->state);
3777         struct intel_pipe_wm *active = &cstate->wm.optimal.ilk;
3778         enum pipe pipe = intel_crtc->pipe;
3779         static const i915_reg_t wm0_pipe_reg[] = {
3780                 [PIPE_A] = WM0_PIPEA_ILK,
3781                 [PIPE_B] = WM0_PIPEB_ILK,
3782                 [PIPE_C] = WM0_PIPEC_IVB,
3783         };
3784
3785         hw->wm_pipe[pipe] = I915_READ(wm0_pipe_reg[pipe]);
3786         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
3787                 hw->wm_linetime[pipe] = I915_READ(PIPE_WM_LINETIME(pipe));
3788
3789         active->pipe_enabled = intel_crtc->active;
3790
3791         if (active->pipe_enabled) {
3792                 u32 tmp = hw->wm_pipe[pipe];
3793
3794                 /*
3795                  * For active pipes LP0 watermark is marked as
3796                  * enabled, and LP1+ watermaks as disabled since
3797                  * we can't really reverse compute them in case
3798                  * multiple pipes are active.
3799                  */
3800                 active->wm[0].enable = true;
3801                 active->wm[0].pri_val = (tmp & WM0_PIPE_PLANE_MASK) >> WM0_PIPE_PLANE_SHIFT;
3802                 active->wm[0].spr_val = (tmp & WM0_PIPE_SPRITE_MASK) >> WM0_PIPE_SPRITE_SHIFT;
3803                 active->wm[0].cur_val = tmp & WM0_PIPE_CURSOR_MASK;
3804                 active->linetime = hw->wm_linetime[pipe];
3805         } else {
3806                 int level, max_level = ilk_wm_max_level(dev);
3807
3808                 /*
3809                  * For inactive pipes, all watermark levels
3810                  * should be marked as enabled but zeroed,
3811                  * which is what we'd compute them to.
3812                  */
3813                 for (level = 0; level <= max_level; level++)
3814                         active->wm[level].enable = true;
3815         }
3816
3817         intel_crtc->wm.active.ilk = *active;
3818 }
3819
3820 #define _FW_WM(value, plane) \
3821         (((value) & DSPFW_ ## plane ## _MASK) >> DSPFW_ ## plane ## _SHIFT)
3822 #define _FW_WM_VLV(value, plane) \
3823         (((value) & DSPFW_ ## plane ## _MASK_VLV) >> DSPFW_ ## plane ## _SHIFT)
3824
3825 static void vlv_read_wm_values(struct drm_i915_private *dev_priv,
3826                                struct vlv_wm_values *wm)
3827 {
3828         enum pipe pipe;
3829         uint32_t tmp;
3830
3831         for_each_pipe(dev_priv, pipe) {
3832                 tmp = I915_READ(VLV_DDL(pipe));
3833
3834                 wm->ddl[pipe].primary =
3835                         (tmp >> DDL_PLANE_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
3836                 wm->ddl[pipe].cursor =
3837                         (tmp >> DDL_CURSOR_SHIFT) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
3838                 wm->ddl[pipe].sprite[0] =
3839                         (tmp >> DDL_SPRITE_SHIFT(0)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
3840                 wm->ddl[pipe].sprite[1] =
3841                         (tmp >> DDL_SPRITE_SHIFT(1)) & (DDL_PRECISION_HIGH | DRAIN_LATENCY_MASK);
3842         }
3843
3844         tmp = I915_READ(DSPFW1);
3845         wm->sr.plane = _FW_WM(tmp, SR);
3846         wm->pipe[PIPE_B].cursor = _FW_WM(tmp, CURSORB);
3847         wm->pipe[PIPE_B].primary = _FW_WM_VLV(tmp, PLANEB);
3848         wm->pipe[PIPE_A].primary = _FW_WM_VLV(tmp, PLANEA);
3849
3850         tmp = I915_READ(DSPFW2);
3851         wm->pipe[PIPE_A].sprite[1] = _FW_WM_VLV(tmp, SPRITEB);
3852         wm->pipe[PIPE_A].cursor = _FW_WM(tmp, CURSORA);
3853         wm->pipe[PIPE_A].sprite[0] = _FW_WM_VLV(tmp, SPRITEA);
3854
3855         tmp = I915_READ(DSPFW3);
3856         wm->sr.cursor = _FW_WM(tmp, CURSOR_SR);
3857
3858         if (IS_CHERRYVIEW(dev_priv)) {
3859                 tmp = I915_READ(DSPFW7_CHV);
3860                 wm->pipe[PIPE_B].sprite[1] = _FW_WM_VLV(tmp, SPRITED);
3861                 wm->pipe[PIPE_B].sprite[0] = _FW_WM_VLV(tmp, SPRITEC);
3862
3863                 tmp = I915_READ(DSPFW8_CHV);
3864                 wm->pipe[PIPE_C].sprite[1] = _FW_WM_VLV(tmp, SPRITEF);
3865                 wm->pipe[PIPE_C].sprite[0] = _FW_WM_VLV(tmp, SPRITEE);
3866
3867                 tmp = I915_READ(DSPFW9_CHV);
3868                 wm->pipe[PIPE_C].primary = _FW_WM_VLV(tmp, PLANEC);
3869                 wm->pipe[PIPE_C].cursor = _FW_WM(tmp, CURSORC);
3870
3871                 tmp = I915_READ(DSPHOWM);
3872                 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
3873                 wm->pipe[PIPE_C].sprite[1] |= _FW_WM(tmp, SPRITEF_HI) << 8;
3874                 wm->pipe[PIPE_C].sprite[0] |= _FW_WM(tmp, SPRITEE_HI) << 8;
3875                 wm->pipe[PIPE_C].primary |= _FW_WM(tmp, PLANEC_HI) << 8;
3876                 wm->pipe[PIPE_B].sprite[1] |= _FW_WM(tmp, SPRITED_HI) << 8;
3877                 wm->pipe[PIPE_B].sprite[0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
3878                 wm->pipe[PIPE_B].primary |= _FW_WM(tmp, PLANEB_HI) << 8;
3879                 wm->pipe[PIPE_A].sprite[1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
3880                 wm->pipe[PIPE_A].sprite[0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
3881                 wm->pipe[PIPE_A].primary |= _FW_WM(tmp, PLANEA_HI) << 8;
3882         } else {
3883                 tmp = I915_READ(DSPFW7);
3884                 wm->pipe[PIPE_B].sprite[1] = _FW_WM_VLV(tmp, SPRITED);
3885                 wm->pipe[PIPE_B].sprite[0] = _FW_WM_VLV(tmp, SPRITEC);
3886
3887                 tmp = I915_READ(DSPHOWM);
3888                 wm->sr.plane |= _FW_WM(tmp, SR_HI) << 9;
3889                 wm->pipe[PIPE_B].sprite[1] |= _FW_WM(tmp, SPRITED_HI) << 8;
3890                 wm->pipe[PIPE_B].sprite[0] |= _FW_WM(tmp, SPRITEC_HI) << 8;
3891                 wm->pipe[PIPE_B].primary |= _FW_WM(tmp, PLANEB_HI) << 8;
3892                 wm->pipe[PIPE_A].sprite[1] |= _FW_WM(tmp, SPRITEB_HI) << 8;
3893                 wm->pipe[PIPE_A].sprite[0] |= _FW_WM(tmp, SPRITEA_HI) << 8;
3894                 wm->pipe[PIPE_A].primary |= _FW_WM(tmp, PLANEA_HI) << 8;
3895         }
3896 }
3897
3898 #undef _FW_WM
3899 #undef _FW_WM_VLV
3900
3901 void vlv_wm_get_hw_state(struct drm_device *dev)
3902 {
3903         struct drm_i915_private *dev_priv = to_i915(dev);
3904         struct vlv_wm_values *wm = &dev_priv->wm.vlv;
3905         struct intel_plane *plane;
3906         enum pipe pipe;
3907         u32 val;
3908
3909         vlv_read_wm_values(dev_priv, wm);
3910
3911         for_each_intel_plane(dev, plane) {
3912                 switch (plane->base.type) {
3913                         int sprite;
3914                 case DRM_PLANE_TYPE_CURSOR:
3915                         plane->wm.fifo_size = 63;
3916                         break;
3917                 case DRM_PLANE_TYPE_PRIMARY:
3918                         plane->wm.fifo_size = vlv_get_fifo_size(dev, plane->pipe, 0);
3919                         break;
3920                 case DRM_PLANE_TYPE_OVERLAY:
3921                         sprite = plane->plane;
3922                         plane->wm.fifo_size = vlv_get_fifo_size(dev, plane->pipe, sprite + 1);
3923                         break;
3924                 }
3925         }
3926
3927         wm->cxsr = I915_READ(FW_BLC_SELF_VLV) & FW_CSPWRDWNEN;
3928         wm->level = VLV_WM_LEVEL_PM2;
3929
3930         if (IS_CHERRYVIEW(dev_priv)) {
3931                 mutex_lock(&dev_priv->rps.hw_lock);
3932
3933                 val = vlv_punit_read(dev_priv, PUNIT_REG_DSPFREQ);
3934                 if (val & DSP_MAXFIFO_PM5_ENABLE)
3935                         wm->level = VLV_WM_LEVEL_PM5;
3936
3937                 /*
3938                  * If DDR DVFS is disabled in the BIOS, Punit
3939                  * will never ack the request. So if that happens
3940                  * assume we don't have to enable/disable DDR DVFS
3941                  * dynamically. To test that just set the REQ_ACK
3942                  * bit to poke the Punit, but don't change the
3943                  * HIGH/LOW bits so that we don't actually change
3944                  * the current state.
3945                  */
3946                 val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
3947                 val |= FORCE_DDR_FREQ_REQ_ACK;
3948                 vlv_punit_write(dev_priv, PUNIT_REG_DDR_SETUP2, val);
3949
3950                 if (wait_for((vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2) &
3951                               FORCE_DDR_FREQ_REQ_ACK) == 0, 3)) {
3952                         DRM_DEBUG_KMS("Punit not acking DDR DVFS request, "
3953                                       "assuming DDR DVFS is disabled\n");
3954                         dev_priv->wm.max_level = VLV_WM_LEVEL_PM5;
3955                 } else {
3956                         val = vlv_punit_read(dev_priv, PUNIT_REG_DDR_SETUP2);
3957                         if ((val & FORCE_DDR_HIGH_FREQ) == 0)
3958                                 wm->level = VLV_WM_LEVEL_DDR_DVFS;
3959                 }
3960
3961                 mutex_unlock(&dev_priv->rps.hw_lock);
3962         }
3963
3964         for_each_pipe(dev_priv, pipe)
3965                 DRM_DEBUG_KMS("Initial watermarks: pipe %c, plane=%d, cursor=%d, sprite0=%d, sprite1=%d\n",
3966                               pipe_name(pipe), wm->pipe[pipe].primary, wm->pipe[pipe].cursor,
3967                               wm->pipe[pipe].sprite[0], wm->pipe[pipe].sprite[1]);
3968
3969         DRM_DEBUG_KMS("Initial watermarks: SR plane=%d, SR cursor=%d level=%d cxsr=%d\n",
3970                       wm->sr.plane, wm->sr.cursor, wm->level, wm->cxsr);
3971 }
3972
3973 void ilk_wm_get_hw_state(struct drm_device *dev)
3974 {
3975         struct drm_i915_private *dev_priv = dev->dev_private;
3976         struct ilk_wm_values *hw = &dev_priv->wm.hw;
3977         struct drm_crtc *crtc;
3978
3979         for_each_crtc(dev, crtc)
3980                 ilk_pipe_wm_get_hw_state(crtc);
3981
3982         hw->wm_lp[0] = I915_READ(WM1_LP_ILK);
3983         hw->wm_lp[1] = I915_READ(WM2_LP_ILK);
3984         hw->wm_lp[2] = I915_READ(WM3_LP_ILK);
3985
3986         hw->wm_lp_spr[0] = I915_READ(WM1S_LP_ILK);
3987         if (INTEL_INFO(dev)->gen >= 7) {
3988                 hw->wm_lp_spr[1] = I915_READ(WM2S_LP_IVB);
3989                 hw->wm_lp_spr[2] = I915_READ(WM3S_LP_IVB);
3990         }
3991
3992         if (IS_HASWELL(dev) || IS_BROADWELL(dev))
3993                 hw->partitioning = (I915_READ(WM_MISC) & WM_MISC_DATA_PARTITION_5_6) ?
3994                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
3995         else if (IS_IVYBRIDGE(dev))
3996                 hw->partitioning = (I915_READ(DISP_ARB_CTL2) & DISP_DATA_PARTITION_5_6) ?
3997                         INTEL_DDB_PART_5_6 : INTEL_DDB_PART_1_2;
3998
3999         hw->enable_fbc_wm =
4000                 !(I915_READ(DISP_ARB_CTL) & DISP_FBC_WM_DIS);
4001 }
4002
4003 /**
4004  * intel_update_watermarks - update FIFO watermark values based on current modes
4005  *
4006  * Calculate watermark values for the various WM regs based on current mode
4007  * and plane configuration.
4008  *
4009  * There are several cases to deal with here:
4010  *   - normal (i.e. non-self-refresh)
4011  *   - self-refresh (SR) mode
4012  *   - lines are large relative to FIFO size (buffer can hold up to 2)
4013  *   - lines are small relative to FIFO size (buffer can hold more than 2
4014  *     lines), so need to account for TLB latency
4015  *
4016  *   The normal calculation is:
4017  *     watermark = dotclock * bytes per pixel * latency
4018  *   where latency is platform & configuration dependent (we assume pessimal
4019  *   values here).
4020  *
4021  *   The SR calculation is:
4022  *     watermark = (trunc(latency/line time)+1) * surface width *
4023  *       bytes per pixel
4024  *   where
4025  *     line time = htotal / dotclock
4026  *     surface width = hdisplay for normal plane and 64 for cursor
4027  *   and latency is assumed to be high, as above.
4028  *
4029  * The final value programmed to the register should always be rounded up,
4030  * and include an extra 2 entries to account for clock crossings.
4031  *
4032  * We don't use the sprite, so we can ignore that.  And on Crestline we have
4033  * to set the non-SR watermarks to 8.
4034  */
4035 void intel_update_watermarks(struct drm_crtc *crtc)
4036 {
4037         struct drm_i915_private *dev_priv = crtc->dev->dev_private;
4038
4039         if (dev_priv->display.update_wm)
4040                 dev_priv->display.update_wm(crtc);
4041 }
4042
4043 /**
4044  * Lock protecting IPS related data structures
4045  */
4046 DEFINE_SPINLOCK(mchdev_lock);
4047
4048 /* Global for IPS driver to get at the current i915 device. Protected by
4049  * mchdev_lock. */
4050 static struct drm_i915_private *i915_mch_dev;
4051
4052 bool ironlake_set_drps(struct drm_device *dev, u8 val)
4053 {
4054         struct drm_i915_private *dev_priv = dev->dev_private;
4055         u16 rgvswctl;
4056
4057         assert_spin_locked(&mchdev_lock);
4058
4059         rgvswctl = I915_READ16(MEMSWCTL);
4060         if (rgvswctl & MEMCTL_CMD_STS) {
4061                 DRM_DEBUG("gpu busy, RCS change rejected\n");
4062                 return false; /* still busy with another command */
4063         }
4064
4065         rgvswctl = (MEMCTL_CMD_CHFREQ << MEMCTL_CMD_SHIFT) |
4066                 (val << MEMCTL_FREQ_SHIFT) | MEMCTL_SFCAVM;
4067         I915_WRITE16(MEMSWCTL, rgvswctl);
4068         POSTING_READ16(MEMSWCTL);
4069
4070         rgvswctl |= MEMCTL_CMD_STS;
4071         I915_WRITE16(MEMSWCTL, rgvswctl);
4072
4073         return true;
4074 }
4075
4076 static void ironlake_enable_drps(struct drm_device *dev)
4077 {
4078         struct drm_i915_private *dev_priv = dev->dev_private;
4079         u32 rgvmodectl = I915_READ(MEMMODECTL);
4080         u8 fmax, fmin, fstart, vstart;
4081
4082         spin_lock_irq(&mchdev_lock);
4083
4084         /* Enable temp reporting */
4085         I915_WRITE16(PMMISC, I915_READ(PMMISC) | MCPPCE_EN);
4086         I915_WRITE16(TSC1, I915_READ(TSC1) | TSE);
4087
4088         /* 100ms RC evaluation intervals */
4089         I915_WRITE(RCUPEI, 100000);
4090         I915_WRITE(RCDNEI, 100000);
4091
4092         /* Set max/min thresholds to 90ms and 80ms respectively */
4093         I915_WRITE(RCBMAXAVG, 90000);
4094         I915_WRITE(RCBMINAVG, 80000);
4095
4096         I915_WRITE(MEMIHYST, 1);
4097
4098         /* Set up min, max, and cur for interrupt handling */
4099         fmax = (rgvmodectl & MEMMODE_FMAX_MASK) >> MEMMODE_FMAX_SHIFT;
4100         fmin = (rgvmodectl & MEMMODE_FMIN_MASK);
4101         fstart = (rgvmodectl & MEMMODE_FSTART_MASK) >>
4102                 MEMMODE_FSTART_SHIFT;
4103
4104         vstart = (I915_READ(PXVFREQ(fstart)) & PXVFREQ_PX_MASK) >>
4105                 PXVFREQ_PX_SHIFT;
4106
4107         dev_priv->ips.fmax = fmax; /* IPS callback will increase this */
4108         dev_priv->ips.fstart = fstart;
4109
4110         dev_priv->ips.max_delay = fstart;
4111         dev_priv->ips.min_delay = fmin;
4112         dev_priv->ips.cur_delay = fstart;
4113
4114         DRM_DEBUG_DRIVER("fmax: %d, fmin: %d, fstart: %d\n",
4115                          fmax, fmin, fstart);
4116
4117         I915_WRITE(MEMINTREN, MEMINT_CX_SUPR_EN | MEMINT_EVAL_CHG_EN);
4118
4119         /*
4120          * Interrupts will be enabled in ironlake_irq_postinstall
4121          */
4122
4123         I915_WRITE(VIDSTART, vstart);
4124         POSTING_READ(VIDSTART);
4125
4126         rgvmodectl |= MEMMODE_SWMODE_EN;
4127         I915_WRITE(MEMMODECTL, rgvmodectl);
4128
4129         if (wait_for_atomic((I915_READ(MEMSWCTL) & MEMCTL_CMD_STS) == 0, 10))
4130                 DRM_ERROR("stuck trying to change perf mode\n");
4131         mdelay(1);
4132
4133         ironlake_set_drps(dev, fstart);
4134
4135         dev_priv->ips.last_count1 = I915_READ(DMIEC) +
4136                 I915_READ(DDREC) + I915_READ(CSIEC);
4137         dev_priv->ips.last_time1 = jiffies_to_msecs(jiffies);
4138         dev_priv->ips.last_count2 = I915_READ(GFXEC);
4139         dev_priv->ips.last_time2 = ktime_get_raw_ns();
4140
4141         spin_unlock_irq(&mchdev_lock);
4142 }
4143
4144 static void ironlake_disable_drps(struct drm_device *dev)
4145 {
4146         struct drm_i915_private *dev_priv = dev->dev_private;
4147         u16 rgvswctl;
4148
4149         spin_lock_irq(&mchdev_lock);
4150
4151         rgvswctl = I915_READ16(MEMSWCTL);
4152
4153         /* Ack interrupts, disable EFC interrupt */
4154         I915_WRITE(MEMINTREN, I915_READ(MEMINTREN) & ~MEMINT_EVAL_CHG_EN);
4155         I915_WRITE(MEMINTRSTS, MEMINT_EVAL_CHG);
4156         I915_WRITE(DEIER, I915_READ(DEIER) & ~DE_PCU_EVENT);
4157         I915_WRITE(DEIIR, DE_PCU_EVENT);
4158         I915_WRITE(DEIMR, I915_READ(DEIMR) | DE_PCU_EVENT);
4159
4160         /* Go back to the starting frequency */
4161         ironlake_set_drps(dev, dev_priv->ips.fstart);
4162         mdelay(1);
4163         rgvswctl |= MEMCTL_CMD_STS;
4164         I915_WRITE(MEMSWCTL, rgvswctl);
4165         mdelay(1);
4166
4167         spin_unlock_irq(&mchdev_lock);
4168 }
4169
4170 /* There's a funny hw issue where the hw returns all 0 when reading from
4171  * GEN6_RP_INTERRUPT_LIMITS. Hence we always need to compute the desired value
4172  * ourselves, instead of doing a rmw cycle (which might result in us clearing
4173  * all limits and the gpu stuck at whatever frequency it is at atm).
4174  */
4175 static u32 intel_rps_limits(struct drm_i915_private *dev_priv, u8 val)
4176 {
4177         u32 limits;
4178
4179         /* Only set the down limit when we've reached the lowest level to avoid
4180          * getting more interrupts, otherwise leave this clear. This prevents a
4181          * race in the hw when coming out of rc6: There's a tiny window where
4182          * the hw runs at the minimal clock before selecting the desired
4183          * frequency, if the down threshold expires in that window we will not
4184          * receive a down interrupt. */
4185         if (IS_GEN9(dev_priv->dev)) {
4186                 limits = (dev_priv->rps.max_freq_softlimit) << 23;
4187                 if (val <= dev_priv->rps.min_freq_softlimit)
4188                         limits |= (dev_priv->rps.min_freq_softlimit) << 14;
4189         } else {
4190                 limits = dev_priv->rps.max_freq_softlimit << 24;
4191                 if (val <= dev_priv->rps.min_freq_softlimit)
4192                         limits |= dev_priv->rps.min_freq_softlimit << 16;
4193         }
4194
4195         return limits;
4196 }
4197
4198 static void gen6_set_rps_thresholds(struct drm_i915_private *dev_priv, u8 val)
4199 {
4200         int new_power;
4201         u32 threshold_up = 0, threshold_down = 0; /* in % */
4202         u32 ei_up = 0, ei_down = 0;
4203
4204         new_power = dev_priv->rps.power;
4205         switch (dev_priv->rps.power) {
4206         case LOW_POWER:
4207                 if (val > dev_priv->rps.efficient_freq + 1 && val > dev_priv->rps.cur_freq)
4208                         new_power = BETWEEN;
4209                 break;
4210
4211         case BETWEEN:
4212                 if (val <= dev_priv->rps.efficient_freq && val < dev_priv->rps.cur_freq)
4213                         new_power = LOW_POWER;
4214                 else if (val >= dev_priv->rps.rp0_freq && val > dev_priv->rps.cur_freq)
4215                         new_power = HIGH_POWER;
4216                 break;
4217
4218         case HIGH_POWER:
4219                 if (val < (dev_priv->rps.rp1_freq + dev_priv->rps.rp0_freq) >> 1 && val < dev_priv->rps.cur_freq)
4220                         new_power = BETWEEN;
4221                 break;
4222         }
4223         /* Max/min bins are special */
4224         if (val <= dev_priv->rps.min_freq_softlimit)
4225                 new_power = LOW_POWER;
4226         if (val >= dev_priv->rps.max_freq_softlimit)
4227                 new_power = HIGH_POWER;
4228         if (new_power == dev_priv->rps.power)
4229                 return;
4230
4231         /* Note the units here are not exactly 1us, but 1280ns. */
4232         switch (new_power) {
4233         case LOW_POWER:
4234                 /* Upclock if more than 95% busy over 16ms */
4235                 ei_up = 16000;
4236                 threshold_up = 95;
4237
4238                 /* Downclock if less than 85% busy over 32ms */
4239                 ei_down = 32000;
4240                 threshold_down = 85;
4241                 break;
4242
4243         case BETWEEN:
4244                 /* Upclock if more than 90% busy over 13ms */
4245                 ei_up = 13000;
4246                 threshold_up = 90;
4247
4248                 /* Downclock if less than 75% busy over 32ms */
4249                 ei_down = 32000;
4250                 threshold_down = 75;
4251                 break;
4252
4253         case HIGH_POWER:
4254                 /* Upclock if more than 85% busy over 10ms */
4255                 ei_up = 10000;
4256                 threshold_up = 85;
4257
4258                 /* Downclock if less than 60% busy over 32ms */
4259                 ei_down = 32000;
4260                 threshold_down = 60;
4261                 break;
4262         }
4263
4264         I915_WRITE(GEN6_RP_UP_EI,
4265                 GT_INTERVAL_FROM_US(dev_priv, ei_up));
4266         I915_WRITE(GEN6_RP_UP_THRESHOLD,
4267                 GT_INTERVAL_FROM_US(dev_priv, (ei_up * threshold_up / 100)));
4268
4269         I915_WRITE(GEN6_RP_DOWN_EI,
4270                 GT_INTERVAL_FROM_US(dev_priv, ei_down));
4271         I915_WRITE(GEN6_RP_DOWN_THRESHOLD,
4272                 GT_INTERVAL_FROM_US(dev_priv, (ei_down * threshold_down / 100)));
4273
4274          I915_WRITE(GEN6_RP_CONTROL,
4275                     GEN6_RP_MEDIA_TURBO |
4276                     GEN6_RP_MEDIA_HW_NORMAL_MODE |
4277                     GEN6_RP_MEDIA_IS_GFX |
4278                     GEN6_RP_ENABLE |
4279                     GEN6_RP_UP_BUSY_AVG |
4280                     GEN6_RP_DOWN_IDLE_AVG);
4281
4282         dev_priv->rps.power = new_power;
4283         dev_priv->rps.up_threshold = threshold_up;
4284         dev_priv->rps.down_threshold = threshold_down;
4285         dev_priv->rps.last_adj = 0;
4286 }
4287
4288 static u32 gen6_rps_pm_mask(struct drm_i915_private *dev_priv, u8 val)
4289 {
4290         u32 mask = 0;
4291
4292         if (val > dev_priv->rps.min_freq_softlimit)
4293                 mask |= GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_DOWN_THRESHOLD | GEN6_PM_RP_DOWN_TIMEOUT;
4294         if (val < dev_priv->rps.max_freq_softlimit)
4295                 mask |= GEN6_PM_RP_UP_EI_EXPIRED | GEN6_PM_RP_UP_THRESHOLD;
4296
4297         mask &= dev_priv->pm_rps_events;
4298
4299         return gen6_sanitize_rps_pm_mask(dev_priv, ~mask);
4300 }
4301
4302 /* gen6_set_rps is called to update the frequency request, but should also be
4303  * called when the range (min_delay and max_delay) is modified so that we can
4304  * update the GEN6_RP_INTERRUPT_LIMITS register accordingly. */
4305 static void gen6_set_rps(struct drm_device *dev, u8 val)
4306 {
4307         struct drm_i915_private *dev_priv = dev->dev_private;
4308
4309         /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
4310         if (IS_BXT_REVID(dev, 0, BXT_REVID_A1))
4311                 return;
4312
4313         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4314         WARN_ON(val > dev_priv->rps.max_freq);
4315         WARN_ON(val < dev_priv->rps.min_freq);
4316
4317         /* min/max delay may still have been modified so be sure to
4318          * write the limits value.
4319          */
4320         if (val != dev_priv->rps.cur_freq) {
4321                 gen6_set_rps_thresholds(dev_priv, val);
4322
4323                 if (IS_GEN9(dev))
4324                         I915_WRITE(GEN6_RPNSWREQ,
4325                                    GEN9_FREQUENCY(val));
4326                 else if (IS_HASWELL(dev) || IS_BROADWELL(dev))
4327                         I915_WRITE(GEN6_RPNSWREQ,
4328                                    HSW_FREQUENCY(val));
4329                 else
4330                         I915_WRITE(GEN6_RPNSWREQ,
4331                                    GEN6_FREQUENCY(val) |
4332                                    GEN6_OFFSET(0) |
4333                                    GEN6_AGGRESSIVE_TURBO);
4334         }
4335
4336         /* Make sure we continue to get interrupts
4337          * until we hit the minimum or maximum frequencies.
4338          */
4339         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS, intel_rps_limits(dev_priv, val));
4340         I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
4341
4342         POSTING_READ(GEN6_RPNSWREQ);
4343
4344         dev_priv->rps.cur_freq = val;
4345         trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
4346 }
4347
4348 static void valleyview_set_rps(struct drm_device *dev, u8 val)
4349 {
4350         struct drm_i915_private *dev_priv = dev->dev_private;
4351
4352         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4353         WARN_ON(val > dev_priv->rps.max_freq);
4354         WARN_ON(val < dev_priv->rps.min_freq);
4355
4356         if (WARN_ONCE(IS_CHERRYVIEW(dev) && (val & 1),
4357                       "Odd GPU freq value\n"))
4358                 val &= ~1;
4359
4360         I915_WRITE(GEN6_PMINTRMSK, gen6_rps_pm_mask(dev_priv, val));
4361
4362         if (val != dev_priv->rps.cur_freq) {
4363                 vlv_punit_write(dev_priv, PUNIT_REG_GPU_FREQ_REQ, val);
4364                 if (!IS_CHERRYVIEW(dev_priv))
4365                         gen6_set_rps_thresholds(dev_priv, val);
4366         }
4367
4368         dev_priv->rps.cur_freq = val;
4369         trace_intel_gpu_freq_change(intel_gpu_freq(dev_priv, val));
4370 }
4371
4372 /* vlv_set_rps_idle: Set the frequency to idle, if Gfx clocks are down
4373  *
4374  * * If Gfx is Idle, then
4375  * 1. Forcewake Media well.
4376  * 2. Request idle freq.
4377  * 3. Release Forcewake of Media well.
4378 */
4379 static void vlv_set_rps_idle(struct drm_i915_private *dev_priv)
4380 {
4381         u32 val = dev_priv->rps.idle_freq;
4382
4383         if (dev_priv->rps.cur_freq <= val)
4384                 return;
4385
4386         /* Wake up the media well, as that takes a lot less
4387          * power than the Render well. */
4388         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_MEDIA);
4389         valleyview_set_rps(dev_priv->dev, val);
4390         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_MEDIA);
4391 }
4392
4393 void gen6_rps_busy(struct drm_i915_private *dev_priv)
4394 {
4395         mutex_lock(&dev_priv->rps.hw_lock);
4396         if (dev_priv->rps.enabled) {
4397                 if (dev_priv->pm_rps_events & (GEN6_PM_RP_DOWN_EI_EXPIRED | GEN6_PM_RP_UP_EI_EXPIRED))
4398                         gen6_rps_reset_ei(dev_priv);
4399                 I915_WRITE(GEN6_PMINTRMSK,
4400                            gen6_rps_pm_mask(dev_priv, dev_priv->rps.cur_freq));
4401         }
4402         mutex_unlock(&dev_priv->rps.hw_lock);
4403 }
4404
4405 void gen6_rps_idle(struct drm_i915_private *dev_priv)
4406 {
4407         struct drm_device *dev = dev_priv->dev;
4408
4409         mutex_lock(&dev_priv->rps.hw_lock);
4410         if (dev_priv->rps.enabled) {
4411                 if (IS_VALLEYVIEW(dev) || IS_CHERRYVIEW(dev))
4412                         vlv_set_rps_idle(dev_priv);
4413                 else
4414                         gen6_set_rps(dev_priv->dev, dev_priv->rps.idle_freq);
4415                 dev_priv->rps.last_adj = 0;
4416                 I915_WRITE(GEN6_PMINTRMSK, 0xffffffff);
4417         }
4418         mutex_unlock(&dev_priv->rps.hw_lock);
4419
4420         spin_lock(&dev_priv->rps.client_lock);
4421         while (!list_empty(&dev_priv->rps.clients))
4422                 list_del_init(dev_priv->rps.clients.next);
4423         spin_unlock(&dev_priv->rps.client_lock);
4424 }
4425
4426 void gen6_rps_boost(struct drm_i915_private *dev_priv,
4427                     struct intel_rps_client *rps,
4428                     unsigned long submitted)
4429 {
4430         /* This is intentionally racy! We peek at the state here, then
4431          * validate inside the RPS worker.
4432          */
4433         if (!(dev_priv->mm.busy &&
4434               dev_priv->rps.enabled &&
4435               dev_priv->rps.cur_freq < dev_priv->rps.max_freq_softlimit))
4436                 return;
4437
4438         /* Force a RPS boost (and don't count it against the client) if
4439          * the GPU is severely congested.
4440          */
4441         if (rps && time_after(jiffies, submitted + DRM_I915_THROTTLE_JIFFIES))
4442                 rps = NULL;
4443
4444         spin_lock(&dev_priv->rps.client_lock);
4445         if (rps == NULL || list_empty(&rps->link)) {
4446                 spin_lock_irq(&dev_priv->irq_lock);
4447                 if (dev_priv->rps.interrupts_enabled) {
4448                         dev_priv->rps.client_boost = true;
4449                         queue_work(dev_priv->wq, &dev_priv->rps.work);
4450                 }
4451                 spin_unlock_irq(&dev_priv->irq_lock);
4452
4453                 if (rps != NULL) {
4454                         list_add(&rps->link, &dev_priv->rps.clients);
4455                         rps->boosts++;
4456                 } else
4457                         dev_priv->rps.boosts++;
4458         }
4459         spin_unlock(&dev_priv->rps.client_lock);
4460 }
4461
4462 void intel_set_rps(struct drm_device *dev, u8 val)
4463 {
4464         if (IS_VALLEYVIEW(dev) || IS_CHERRYVIEW(dev))
4465                 valleyview_set_rps(dev, val);
4466         else
4467                 gen6_set_rps(dev, val);
4468 }
4469
4470 static void gen9_disable_rps(struct drm_device *dev)
4471 {
4472         struct drm_i915_private *dev_priv = dev->dev_private;
4473
4474         I915_WRITE(GEN6_RC_CONTROL, 0);
4475         I915_WRITE(GEN9_PG_ENABLE, 0);
4476 }
4477
4478 static void gen6_disable_rps(struct drm_device *dev)
4479 {
4480         struct drm_i915_private *dev_priv = dev->dev_private;
4481
4482         I915_WRITE(GEN6_RC_CONTROL, 0);
4483         I915_WRITE(GEN6_RPNSWREQ, 1 << 31);
4484 }
4485
4486 static void cherryview_disable_rps(struct drm_device *dev)
4487 {
4488         struct drm_i915_private *dev_priv = dev->dev_private;
4489
4490         I915_WRITE(GEN6_RC_CONTROL, 0);
4491 }
4492
4493 static void valleyview_disable_rps(struct drm_device *dev)
4494 {
4495         struct drm_i915_private *dev_priv = dev->dev_private;
4496
4497         /* we're doing forcewake before Disabling RC6,
4498          * This what the BIOS expects when going into suspend */
4499         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4500
4501         I915_WRITE(GEN6_RC_CONTROL, 0);
4502
4503         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4504 }
4505
4506 static void intel_print_rc6_info(struct drm_device *dev, u32 mode)
4507 {
4508         if (IS_VALLEYVIEW(dev) || IS_CHERRYVIEW(dev)) {
4509                 if (mode & (GEN7_RC_CTL_TO_MODE | GEN6_RC_CTL_EI_MODE(1)))
4510                         mode = GEN6_RC_CTL_RC6_ENABLE;
4511                 else
4512                         mode = 0;
4513         }
4514         if (HAS_RC6p(dev))
4515                 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s RC6p %s RC6pp %s\n",
4516                               (mode & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off",
4517                               (mode & GEN6_RC_CTL_RC6p_ENABLE) ? "on" : "off",
4518                               (mode & GEN6_RC_CTL_RC6pp_ENABLE) ? "on" : "off");
4519
4520         else
4521                 DRM_DEBUG_KMS("Enabling RC6 states: RC6 %s\n",
4522                               (mode & GEN6_RC_CTL_RC6_ENABLE) ? "on" : "off");
4523 }
4524
4525 static int sanitize_rc6_option(const struct drm_device *dev, int enable_rc6)
4526 {
4527         /* No RC6 before Ironlake and code is gone for ilk. */
4528         if (INTEL_INFO(dev)->gen < 6)
4529                 return 0;
4530
4531         /* Respect the kernel parameter if it is set */
4532         if (enable_rc6 >= 0) {
4533                 int mask;
4534
4535                 if (HAS_RC6p(dev))
4536                         mask = INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE |
4537                                INTEL_RC6pp_ENABLE;
4538                 else
4539                         mask = INTEL_RC6_ENABLE;
4540
4541                 if ((enable_rc6 & mask) != enable_rc6)
4542                         DRM_DEBUG_KMS("Adjusting RC6 mask to %d (requested %d, valid %d)\n",
4543                                       enable_rc6 & mask, enable_rc6, mask);
4544
4545                 return enable_rc6 & mask;
4546         }
4547
4548         if (IS_IVYBRIDGE(dev))
4549                 return (INTEL_RC6_ENABLE | INTEL_RC6p_ENABLE);
4550
4551         return INTEL_RC6_ENABLE;
4552 }
4553
4554 int intel_enable_rc6(const struct drm_device *dev)
4555 {
4556         return i915.enable_rc6;
4557 }
4558
4559 static void gen6_init_rps_frequencies(struct drm_device *dev)
4560 {
4561         struct drm_i915_private *dev_priv = dev->dev_private;
4562         uint32_t rp_state_cap;
4563         u32 ddcc_status = 0;
4564         int ret;
4565
4566         /* All of these values are in units of 50MHz */
4567         dev_priv->rps.cur_freq          = 0;
4568         /* static values from HW: RP0 > RP1 > RPn (min_freq) */
4569         if (IS_BROXTON(dev)) {
4570                 rp_state_cap = I915_READ(BXT_RP_STATE_CAP);
4571                 dev_priv->rps.rp0_freq = (rp_state_cap >> 16) & 0xff;
4572                 dev_priv->rps.rp1_freq = (rp_state_cap >>  8) & 0xff;
4573                 dev_priv->rps.min_freq = (rp_state_cap >>  0) & 0xff;
4574         } else {
4575                 rp_state_cap = I915_READ(GEN6_RP_STATE_CAP);
4576                 dev_priv->rps.rp0_freq = (rp_state_cap >>  0) & 0xff;
4577                 dev_priv->rps.rp1_freq = (rp_state_cap >>  8) & 0xff;
4578                 dev_priv->rps.min_freq = (rp_state_cap >> 16) & 0xff;
4579         }
4580
4581         /* hw_max = RP0 until we check for overclocking */
4582         dev_priv->rps.max_freq          = dev_priv->rps.rp0_freq;
4583
4584         dev_priv->rps.efficient_freq = dev_priv->rps.rp1_freq;
4585         if (IS_HASWELL(dev) || IS_BROADWELL(dev) ||
4586             IS_SKYLAKE(dev) || IS_KABYLAKE(dev)) {
4587                 ret = sandybridge_pcode_read(dev_priv,
4588                                         HSW_PCODE_DYNAMIC_DUTY_CYCLE_CONTROL,
4589                                         &ddcc_status);
4590                 if (0 == ret)
4591                         dev_priv->rps.efficient_freq =
4592                                 clamp_t(u8,
4593                                         ((ddcc_status >> 8) & 0xff),
4594                                         dev_priv->rps.min_freq,
4595                                         dev_priv->rps.max_freq);
4596         }
4597
4598         if (IS_SKYLAKE(dev) || IS_KABYLAKE(dev)) {
4599                 /* Store the frequency values in 16.66 MHZ units, which is
4600                    the natural hardware unit for SKL */
4601                 dev_priv->rps.rp0_freq *= GEN9_FREQ_SCALER;
4602                 dev_priv->rps.rp1_freq *= GEN9_FREQ_SCALER;
4603                 dev_priv->rps.min_freq *= GEN9_FREQ_SCALER;
4604                 dev_priv->rps.max_freq *= GEN9_FREQ_SCALER;
4605                 dev_priv->rps.efficient_freq *= GEN9_FREQ_SCALER;
4606         }
4607
4608         dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
4609
4610         /* Preserve min/max settings in case of re-init */
4611         if (dev_priv->rps.max_freq_softlimit == 0)
4612                 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
4613
4614         if (dev_priv->rps.min_freq_softlimit == 0) {
4615                 if (IS_HASWELL(dev) || IS_BROADWELL(dev))
4616                         dev_priv->rps.min_freq_softlimit =
4617                                 max_t(int, dev_priv->rps.efficient_freq,
4618                                       intel_freq_opcode(dev_priv, 450));
4619                 else
4620                         dev_priv->rps.min_freq_softlimit =
4621                                 dev_priv->rps.min_freq;
4622         }
4623 }
4624
4625 /* See the Gen9_GT_PM_Programming_Guide doc for the below */
4626 static void gen9_enable_rps(struct drm_device *dev)
4627 {
4628         struct drm_i915_private *dev_priv = dev->dev_private;
4629
4630         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4631
4632         gen6_init_rps_frequencies(dev);
4633
4634         /* WaGsvDisableTurbo: Workaround to disable turbo on BXT A* */
4635         if (IS_BXT_REVID(dev, 0, BXT_REVID_A1)) {
4636                 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4637                 return;
4638         }
4639
4640         /* Program defaults and thresholds for RPS*/
4641         I915_WRITE(GEN6_RC_VIDEO_FREQ,
4642                 GEN9_FREQUENCY(dev_priv->rps.rp1_freq));
4643
4644         /* 1 second timeout*/
4645         I915_WRITE(GEN6_RP_DOWN_TIMEOUT,
4646                 GT_INTERVAL_FROM_US(dev_priv, 1000000));
4647
4648         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 0xa);
4649
4650         /* Leaning on the below call to gen6_set_rps to program/setup the
4651          * Up/Down EI & threshold registers, as well as the RP_CONTROL,
4652          * RP_INTERRUPT_LIMITS & RPNSWREQ registers */
4653         dev_priv->rps.power = HIGH_POWER; /* force a reset */
4654         gen6_set_rps(dev_priv->dev, dev_priv->rps.min_freq_softlimit);
4655
4656         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4657 }
4658
4659 static void gen9_enable_rc6(struct drm_device *dev)
4660 {
4661         struct drm_i915_private *dev_priv = dev->dev_private;
4662         struct intel_engine_cs *ring;
4663         uint32_t rc6_mask = 0;
4664         int unused;
4665
4666         /* 1a: Software RC state - RC0 */
4667         I915_WRITE(GEN6_RC_STATE, 0);
4668
4669         /* 1b: Get forcewake during program sequence. Although the driver
4670          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4671         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4672
4673         /* 2a: Disable RC states. */
4674         I915_WRITE(GEN6_RC_CONTROL, 0);
4675
4676         /* 2b: Program RC6 thresholds.*/
4677
4678         /* WaRsDoubleRc6WrlWithCoarsePowerGating: Doubling WRL only when CPG is enabled */
4679         if (IS_SKYLAKE(dev))
4680                 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 108 << 16);
4681         else
4682                 I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 54 << 16);
4683         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
4684         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
4685         for_each_ring(ring, dev_priv, unused)
4686                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4687
4688         if (HAS_GUC_UCODE(dev))
4689                 I915_WRITE(GUC_MAX_IDLE_COUNT, 0xA);
4690
4691         I915_WRITE(GEN6_RC_SLEEP, 0);
4692
4693         /* 2c: Program Coarse Power Gating Policies. */
4694         I915_WRITE(GEN9_MEDIA_PG_IDLE_HYSTERESIS, 25);
4695         I915_WRITE(GEN9_RENDER_PG_IDLE_HYSTERESIS, 25);
4696
4697         /* 3a: Enable RC6 */
4698         if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
4699                 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
4700         DRM_INFO("RC6 %s\n", (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ?
4701                         "on" : "off");
4702         /* WaRsUseTimeoutMode */
4703         if (IS_SKL_REVID(dev, 0, SKL_REVID_D0) ||
4704             IS_BXT_REVID(dev, 0, BXT_REVID_A1)) {
4705                 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us */
4706                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4707                            GEN7_RC_CTL_TO_MODE |
4708                            rc6_mask);
4709         } else {
4710                 I915_WRITE(GEN6_RC6_THRESHOLD, 37500); /* 37.5/125ms per EI */
4711                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4712                            GEN6_RC_CTL_EI_MODE(1) |
4713                            rc6_mask);
4714         }
4715
4716         /*
4717          * 3b: Enable Coarse Power Gating only when RC6 is enabled.
4718          * WaRsDisableCoarsePowerGating:skl,bxt - Render/Media PG need to be disabled with RC6.
4719          */
4720         if ((IS_BROXTON(dev) && (INTEL_REVID(dev) < BXT_REVID_B0)) ||
4721             ((IS_SKL_GT3(dev) || IS_SKL_GT4(dev)) && (INTEL_REVID(dev) <= SKL_REVID_F0)))
4722                 I915_WRITE(GEN9_PG_ENABLE, 0);
4723         else
4724                 I915_WRITE(GEN9_PG_ENABLE, (rc6_mask & GEN6_RC_CTL_RC6_ENABLE) ?
4725                                 (GEN9_RENDER_PG_ENABLE | GEN9_MEDIA_PG_ENABLE) : 0);
4726
4727         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4728
4729 }
4730
4731 static void gen8_enable_rps(struct drm_device *dev)
4732 {
4733         struct drm_i915_private *dev_priv = dev->dev_private;
4734         struct intel_engine_cs *ring;
4735         uint32_t rc6_mask = 0;
4736         int unused;
4737
4738         /* 1a: Software RC state - RC0 */
4739         I915_WRITE(GEN6_RC_STATE, 0);
4740
4741         /* 1c & 1d: Get forcewake during program sequence. Although the driver
4742          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
4743         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4744
4745         /* 2a: Disable RC states. */
4746         I915_WRITE(GEN6_RC_CONTROL, 0);
4747
4748         /* Initialize rps frequencies */
4749         gen6_init_rps_frequencies(dev);
4750
4751         /* 2b: Program RC6 thresholds.*/
4752         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
4753         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
4754         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
4755         for_each_ring(ring, dev_priv, unused)
4756                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4757         I915_WRITE(GEN6_RC_SLEEP, 0);
4758         if (IS_BROADWELL(dev))
4759                 I915_WRITE(GEN6_RC6_THRESHOLD, 625); /* 800us/1.28 for TO */
4760         else
4761                 I915_WRITE(GEN6_RC6_THRESHOLD, 50000); /* 50/125ms per EI */
4762
4763         /* 3: Enable RC6 */
4764         if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
4765                 rc6_mask = GEN6_RC_CTL_RC6_ENABLE;
4766         intel_print_rc6_info(dev, rc6_mask);
4767         if (IS_BROADWELL(dev))
4768                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4769                                 GEN7_RC_CTL_TO_MODE |
4770                                 rc6_mask);
4771         else
4772                 I915_WRITE(GEN6_RC_CONTROL, GEN6_RC_CTL_HW_ENABLE |
4773                                 GEN6_RC_CTL_EI_MODE(1) |
4774                                 rc6_mask);
4775
4776         /* 4 Program defaults and thresholds for RPS*/
4777         I915_WRITE(GEN6_RPNSWREQ,
4778                    HSW_FREQUENCY(dev_priv->rps.rp1_freq));
4779         I915_WRITE(GEN6_RC_VIDEO_FREQ,
4780                    HSW_FREQUENCY(dev_priv->rps.rp1_freq));
4781         /* NB: Docs say 1s, and 1000000 - which aren't equivalent */
4782         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 100000000 / 128); /* 1 second timeout */
4783
4784         /* Docs recommend 900MHz, and 300 MHz respectively */
4785         I915_WRITE(GEN6_RP_INTERRUPT_LIMITS,
4786                    dev_priv->rps.max_freq_softlimit << 24 |
4787                    dev_priv->rps.min_freq_softlimit << 16);
4788
4789         I915_WRITE(GEN6_RP_UP_THRESHOLD, 7600000 / 128); /* 76ms busyness per EI, 90% */
4790         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 31300000 / 128); /* 313ms busyness per EI, 70%*/
4791         I915_WRITE(GEN6_RP_UP_EI, 66000); /* 84.48ms, XXX: random? */
4792         I915_WRITE(GEN6_RP_DOWN_EI, 350000); /* 448ms, XXX: random? */
4793
4794         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
4795
4796         /* 5: Enable RPS */
4797         I915_WRITE(GEN6_RP_CONTROL,
4798                    GEN6_RP_MEDIA_TURBO |
4799                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
4800                    GEN6_RP_MEDIA_IS_GFX |
4801                    GEN6_RP_ENABLE |
4802                    GEN6_RP_UP_BUSY_AVG |
4803                    GEN6_RP_DOWN_IDLE_AVG);
4804
4805         /* 6: Ring frequency + overclocking (our driver does this later */
4806
4807         dev_priv->rps.power = HIGH_POWER; /* force a reset */
4808         gen6_set_rps(dev_priv->dev, dev_priv->rps.idle_freq);
4809
4810         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4811 }
4812
4813 static void gen6_enable_rps(struct drm_device *dev)
4814 {
4815         struct drm_i915_private *dev_priv = dev->dev_private;
4816         struct intel_engine_cs *ring;
4817         u32 rc6vids, pcu_mbox = 0, rc6_mask = 0;
4818         u32 gtfifodbg;
4819         int rc6_mode;
4820         int i, ret;
4821
4822         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4823
4824         /* Here begins a magic sequence of register writes to enable
4825          * auto-downclocking.
4826          *
4827          * Perhaps there might be some value in exposing these to
4828          * userspace...
4829          */
4830         I915_WRITE(GEN6_RC_STATE, 0);
4831
4832         /* Clear the DBG now so we don't confuse earlier errors */
4833         if ((gtfifodbg = I915_READ(GTFIFODBG))) {
4834                 DRM_ERROR("GT fifo had a previous error %x\n", gtfifodbg);
4835                 I915_WRITE(GTFIFODBG, gtfifodbg);
4836         }
4837
4838         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
4839
4840         /* Initialize rps frequencies */
4841         gen6_init_rps_frequencies(dev);
4842
4843         /* disable the counters and set deterministic thresholds */
4844         I915_WRITE(GEN6_RC_CONTROL, 0);
4845
4846         I915_WRITE(GEN6_RC1_WAKE_RATE_LIMIT, 1000 << 16);
4847         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16 | 30);
4848         I915_WRITE(GEN6_RC6pp_WAKE_RATE_LIMIT, 30);
4849         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
4850         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
4851
4852         for_each_ring(ring, dev_priv, i)
4853                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
4854
4855         I915_WRITE(GEN6_RC_SLEEP, 0);
4856         I915_WRITE(GEN6_RC1e_THRESHOLD, 1000);
4857         if (IS_IVYBRIDGE(dev))
4858                 I915_WRITE(GEN6_RC6_THRESHOLD, 125000);
4859         else
4860                 I915_WRITE(GEN6_RC6_THRESHOLD, 50000);
4861         I915_WRITE(GEN6_RC6p_THRESHOLD, 150000);
4862         I915_WRITE(GEN6_RC6pp_THRESHOLD, 64000); /* unused */
4863
4864         /* Check if we are enabling RC6 */
4865         rc6_mode = intel_enable_rc6(dev_priv->dev);
4866         if (rc6_mode & INTEL_RC6_ENABLE)
4867                 rc6_mask |= GEN6_RC_CTL_RC6_ENABLE;
4868
4869         /* We don't use those on Haswell */
4870         if (!IS_HASWELL(dev)) {
4871                 if (rc6_mode & INTEL_RC6p_ENABLE)
4872                         rc6_mask |= GEN6_RC_CTL_RC6p_ENABLE;
4873
4874                 if (rc6_mode & INTEL_RC6pp_ENABLE)
4875                         rc6_mask |= GEN6_RC_CTL_RC6pp_ENABLE;
4876         }
4877
4878         intel_print_rc6_info(dev, rc6_mask);
4879
4880         I915_WRITE(GEN6_RC_CONTROL,
4881                    rc6_mask |
4882                    GEN6_RC_CTL_EI_MODE(1) |
4883                    GEN6_RC_CTL_HW_ENABLE);
4884
4885         /* Power down if completely idle for over 50ms */
4886         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 50000);
4887         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
4888
4889         ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_MIN_FREQ_TABLE, 0);
4890         if (ret)
4891                 DRM_DEBUG_DRIVER("Failed to set the min frequency\n");
4892
4893         ret = sandybridge_pcode_read(dev_priv, GEN6_READ_OC_PARAMS, &pcu_mbox);
4894         if (!ret && (pcu_mbox & (1<<31))) { /* OC supported */
4895                 DRM_DEBUG_DRIVER("Overclocking supported. Max: %dMHz, Overclock max: %dMHz\n",
4896                                  (dev_priv->rps.max_freq_softlimit & 0xff) * 50,
4897                                  (pcu_mbox & 0xff) * 50);
4898                 dev_priv->rps.max_freq = pcu_mbox & 0xff;
4899         }
4900
4901         dev_priv->rps.power = HIGH_POWER; /* force a reset */
4902         gen6_set_rps(dev_priv->dev, dev_priv->rps.idle_freq);
4903
4904         rc6vids = 0;
4905         ret = sandybridge_pcode_read(dev_priv, GEN6_PCODE_READ_RC6VIDS, &rc6vids);
4906         if (IS_GEN6(dev) && ret) {
4907                 DRM_DEBUG_DRIVER("Couldn't check for BIOS workaround\n");
4908         } else if (IS_GEN6(dev) && (GEN6_DECODE_RC6_VID(rc6vids & 0xff) < 450)) {
4909                 DRM_DEBUG_DRIVER("You should update your BIOS. Correcting minimum rc6 voltage (%dmV->%dmV)\n",
4910                           GEN6_DECODE_RC6_VID(rc6vids & 0xff), 450);
4911                 rc6vids &= 0xffff00;
4912                 rc6vids |= GEN6_ENCODE_RC6_VID(450);
4913                 ret = sandybridge_pcode_write(dev_priv, GEN6_PCODE_WRITE_RC6VIDS, rc6vids);
4914                 if (ret)
4915                         DRM_ERROR("Couldn't fix incorrect rc6 voltage\n");
4916         }
4917
4918         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
4919 }
4920
4921 static void __gen6_update_ring_freq(struct drm_device *dev)
4922 {
4923         struct drm_i915_private *dev_priv = dev->dev_private;
4924         int min_freq = 15;
4925         unsigned int gpu_freq;
4926         unsigned int max_ia_freq, min_ring_freq;
4927         unsigned int max_gpu_freq, min_gpu_freq;
4928         int scaling_factor = 180;
4929         struct cpufreq_policy *policy;
4930
4931         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
4932
4933         policy = cpufreq_cpu_get(0);
4934         if (policy) {
4935                 max_ia_freq = policy->cpuinfo.max_freq;
4936                 cpufreq_cpu_put(policy);
4937         } else {
4938                 /*
4939                  * Default to measured freq if none found, PCU will ensure we
4940                  * don't go over
4941                  */
4942                 max_ia_freq = tsc_khz;
4943         }
4944
4945         /* Convert from kHz to MHz */
4946         max_ia_freq /= 1000;
4947
4948         min_ring_freq = I915_READ(DCLK) & 0xf;
4949         /* convert DDR frequency from units of 266.6MHz to bandwidth */
4950         min_ring_freq = mult_frac(min_ring_freq, 8, 3);
4951
4952         if (IS_SKYLAKE(dev) || IS_KABYLAKE(dev)) {
4953                 /* Convert GT frequency to 50 HZ units */
4954                 min_gpu_freq = dev_priv->rps.min_freq / GEN9_FREQ_SCALER;
4955                 max_gpu_freq = dev_priv->rps.max_freq / GEN9_FREQ_SCALER;
4956         } else {
4957                 min_gpu_freq = dev_priv->rps.min_freq;
4958                 max_gpu_freq = dev_priv->rps.max_freq;
4959         }
4960
4961         /*
4962          * For each potential GPU frequency, load a ring frequency we'd like
4963          * to use for memory access.  We do this by specifying the IA frequency
4964          * the PCU should use as a reference to determine the ring frequency.
4965          */
4966         for (gpu_freq = max_gpu_freq; gpu_freq >= min_gpu_freq; gpu_freq--) {
4967                 int diff = max_gpu_freq - gpu_freq;
4968                 unsigned int ia_freq = 0, ring_freq = 0;
4969
4970                 if (IS_SKYLAKE(dev) || IS_KABYLAKE(dev)) {
4971                         /*
4972                          * ring_freq = 2 * GT. ring_freq is in 100MHz units
4973                          * No floor required for ring frequency on SKL.
4974                          */
4975                         ring_freq = gpu_freq;
4976                 } else if (INTEL_INFO(dev)->gen >= 8) {
4977                         /* max(2 * GT, DDR). NB: GT is 50MHz units */
4978                         ring_freq = max(min_ring_freq, gpu_freq);
4979                 } else if (IS_HASWELL(dev)) {
4980                         ring_freq = mult_frac(gpu_freq, 5, 4);
4981                         ring_freq = max(min_ring_freq, ring_freq);
4982                         /* leave ia_freq as the default, chosen by cpufreq */
4983                 } else {
4984                         /* On older processors, there is no separate ring
4985                          * clock domain, so in order to boost the bandwidth
4986                          * of the ring, we need to upclock the CPU (ia_freq).
4987                          *
4988                          * For GPU frequencies less than 750MHz,
4989                          * just use the lowest ring freq.
4990                          */
4991                         if (gpu_freq < min_freq)
4992                                 ia_freq = 800;
4993                         else
4994                                 ia_freq = max_ia_freq - ((diff * scaling_factor) / 2);
4995                         ia_freq = DIV_ROUND_CLOSEST(ia_freq, 100);
4996                 }
4997
4998                 sandybridge_pcode_write(dev_priv,
4999                                         GEN6_PCODE_WRITE_MIN_FREQ_TABLE,
5000                                         ia_freq << GEN6_PCODE_FREQ_IA_RATIO_SHIFT |
5001                                         ring_freq << GEN6_PCODE_FREQ_RING_RATIO_SHIFT |
5002                                         gpu_freq);
5003         }
5004 }
5005
5006 void gen6_update_ring_freq(struct drm_device *dev)
5007 {
5008         struct drm_i915_private *dev_priv = dev->dev_private;
5009
5010         if (!HAS_CORE_RING_FREQ(dev))
5011                 return;
5012
5013         mutex_lock(&dev_priv->rps.hw_lock);
5014         __gen6_update_ring_freq(dev);
5015         mutex_unlock(&dev_priv->rps.hw_lock);
5016 }
5017
5018 static int cherryview_rps_max_freq(struct drm_i915_private *dev_priv)
5019 {
5020         struct drm_device *dev = dev_priv->dev;
5021         u32 val, rp0;
5022
5023         val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
5024
5025         switch (INTEL_INFO(dev)->eu_total) {
5026         case 8:
5027                 /* (2 * 4) config */
5028                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS4EU_FUSE_SHIFT);
5029                 break;
5030         case 12:
5031                 /* (2 * 6) config */
5032                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS6EU_FUSE_SHIFT);
5033                 break;
5034         case 16:
5035                 /* (2 * 8) config */
5036         default:
5037                 /* Setting (2 * 8) Min RP0 for any other combination */
5038                 rp0 = (val >> FB_GFX_FMAX_AT_VMAX_2SS8EU_FUSE_SHIFT);
5039                 break;
5040         }
5041
5042         rp0 = (rp0 & FB_GFX_FREQ_FUSE_MASK);
5043
5044         return rp0;
5045 }
5046
5047 static int cherryview_rps_rpe_freq(struct drm_i915_private *dev_priv)
5048 {
5049         u32 val, rpe;
5050
5051         val = vlv_punit_read(dev_priv, PUNIT_GPU_DUTYCYCLE_REG);
5052         rpe = (val >> PUNIT_GPU_DUTYCYCLE_RPE_FREQ_SHIFT) & PUNIT_GPU_DUTYCYCLE_RPE_FREQ_MASK;
5053
5054         return rpe;
5055 }
5056
5057 static int cherryview_rps_guar_freq(struct drm_i915_private *dev_priv)
5058 {
5059         u32 val, rp1;
5060
5061         val = vlv_punit_read(dev_priv, FB_GFX_FMAX_AT_VMAX_FUSE);
5062         rp1 = (val & FB_GFX_FREQ_FUSE_MASK);
5063
5064         return rp1;
5065 }
5066
5067 static int valleyview_rps_guar_freq(struct drm_i915_private *dev_priv)
5068 {
5069         u32 val, rp1;
5070
5071         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
5072
5073         rp1 = (val & FB_GFX_FGUARANTEED_FREQ_FUSE_MASK) >> FB_GFX_FGUARANTEED_FREQ_FUSE_SHIFT;
5074
5075         return rp1;
5076 }
5077
5078 static int valleyview_rps_max_freq(struct drm_i915_private *dev_priv)
5079 {
5080         u32 val, rp0;
5081
5082         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FREQ_FUSE);
5083
5084         rp0 = (val & FB_GFX_MAX_FREQ_FUSE_MASK) >> FB_GFX_MAX_FREQ_FUSE_SHIFT;
5085         /* Clamp to max */
5086         rp0 = min_t(u32, rp0, 0xea);
5087
5088         return rp0;
5089 }
5090
5091 static int valleyview_rps_rpe_freq(struct drm_i915_private *dev_priv)
5092 {
5093         u32 val, rpe;
5094
5095         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_LO);
5096         rpe = (val & FB_FMAX_VMIN_FREQ_LO_MASK) >> FB_FMAX_VMIN_FREQ_LO_SHIFT;
5097         val = vlv_nc_read(dev_priv, IOSF_NC_FB_GFX_FMAX_FUSE_HI);
5098         rpe |= (val & FB_FMAX_VMIN_FREQ_HI_MASK) << 5;
5099
5100         return rpe;
5101 }
5102
5103 static int valleyview_rps_min_freq(struct drm_i915_private *dev_priv)
5104 {
5105         u32 val;
5106
5107         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_LFM) & 0xff;
5108         /*
5109          * According to the BYT Punit GPU turbo HAS 1.1.6.3 the minimum value
5110          * for the minimum frequency in GPLL mode is 0xc1. Contrary to this on
5111          * a BYT-M B0 the above register contains 0xbf. Moreover when setting
5112          * a frequency Punit will not allow values below 0xc0. Clamp it 0xc0
5113          * to make sure it matches what Punit accepts.
5114          */
5115         return max_t(u32, val, 0xc0);
5116 }
5117
5118 /* Check that the pctx buffer wasn't move under us. */
5119 static void valleyview_check_pctx(struct drm_i915_private *dev_priv)
5120 {
5121         unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
5122
5123         WARN_ON(pctx_addr != dev_priv->mm.stolen_base +
5124                              dev_priv->vlv_pctx->stolen->start);
5125 }
5126
5127
5128 /* Check that the pcbr address is not empty. */
5129 static void cherryview_check_pctx(struct drm_i915_private *dev_priv)
5130 {
5131         unsigned long pctx_addr = I915_READ(VLV_PCBR) & ~4095;
5132
5133         WARN_ON((pctx_addr >> VLV_PCBR_ADDR_SHIFT) == 0);
5134 }
5135
5136 static void cherryview_setup_pctx(struct drm_device *dev)
5137 {
5138         struct drm_i915_private *dev_priv = dev->dev_private;
5139         unsigned long pctx_paddr, paddr;
5140         struct i915_gtt *gtt = &dev_priv->gtt;
5141         u32 pcbr;
5142         int pctx_size = 32*1024;
5143
5144         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
5145
5146         pcbr = I915_READ(VLV_PCBR);
5147         if ((pcbr >> VLV_PCBR_ADDR_SHIFT) == 0) {
5148                 DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5149                 paddr = (dev_priv->mm.stolen_base +
5150                          (gtt->stolen_size - pctx_size));
5151
5152                 pctx_paddr = (paddr & (~4095));
5153                 I915_WRITE(VLV_PCBR, pctx_paddr);
5154         }
5155
5156         DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
5157 }
5158
5159 static void valleyview_setup_pctx(struct drm_device *dev)
5160 {
5161         struct drm_i915_private *dev_priv = dev->dev_private;
5162         struct drm_i915_gem_object *pctx;
5163         unsigned long pctx_paddr;
5164         u32 pcbr;
5165         int pctx_size = 24*1024;
5166
5167         WARN_ON(!mutex_is_locked(&dev->struct_mutex));
5168
5169         pcbr = I915_READ(VLV_PCBR);
5170         if (pcbr) {
5171                 /* BIOS set it up already, grab the pre-alloc'd space */
5172                 int pcbr_offset;
5173
5174                 pcbr_offset = (pcbr & (~4095)) - dev_priv->mm.stolen_base;
5175                 pctx = i915_gem_object_create_stolen_for_preallocated(dev_priv->dev,
5176                                                                       pcbr_offset,
5177                                                                       I915_GTT_OFFSET_NONE,
5178                                                                       pctx_size);
5179                 goto out;
5180         }
5181
5182         DRM_DEBUG_DRIVER("BIOS didn't set up PCBR, fixing up\n");
5183
5184         /*
5185          * From the Gunit register HAS:
5186          * The Gfx driver is expected to program this register and ensure
5187          * proper allocation within Gfx stolen memory.  For example, this
5188          * register should be programmed such than the PCBR range does not
5189          * overlap with other ranges, such as the frame buffer, protected
5190          * memory, or any other relevant ranges.
5191          */
5192         pctx = i915_gem_object_create_stolen(dev, pctx_size);
5193         if (!pctx) {
5194                 DRM_DEBUG("not enough stolen space for PCTX, disabling\n");
5195                 return;
5196         }
5197
5198         pctx_paddr = dev_priv->mm.stolen_base + pctx->stolen->start;
5199         I915_WRITE(VLV_PCBR, pctx_paddr);
5200
5201 out:
5202         DRM_DEBUG_DRIVER("PCBR: 0x%08x\n", I915_READ(VLV_PCBR));
5203         dev_priv->vlv_pctx = pctx;
5204 }
5205
5206 static void valleyview_cleanup_pctx(struct drm_device *dev)
5207 {
5208         struct drm_i915_private *dev_priv = dev->dev_private;
5209
5210         if (WARN_ON(!dev_priv->vlv_pctx))
5211                 return;
5212
5213         drm_gem_object_unreference(&dev_priv->vlv_pctx->base);
5214         dev_priv->vlv_pctx = NULL;
5215 }
5216
5217 static void valleyview_init_gt_powersave(struct drm_device *dev)
5218 {
5219         struct drm_i915_private *dev_priv = dev->dev_private;
5220         u32 val;
5221
5222         valleyview_setup_pctx(dev);
5223
5224         mutex_lock(&dev_priv->rps.hw_lock);
5225
5226         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5227         switch ((val >> 6) & 3) {
5228         case 0:
5229         case 1:
5230                 dev_priv->mem_freq = 800;
5231                 break;
5232         case 2:
5233                 dev_priv->mem_freq = 1066;
5234                 break;
5235         case 3:
5236                 dev_priv->mem_freq = 1333;
5237                 break;
5238         }
5239         DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
5240
5241         dev_priv->rps.max_freq = valleyview_rps_max_freq(dev_priv);
5242         dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
5243         DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5244                          intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
5245                          dev_priv->rps.max_freq);
5246
5247         dev_priv->rps.efficient_freq = valleyview_rps_rpe_freq(dev_priv);
5248         DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5249                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5250                          dev_priv->rps.efficient_freq);
5251
5252         dev_priv->rps.rp1_freq = valleyview_rps_guar_freq(dev_priv);
5253         DRM_DEBUG_DRIVER("RP1(Guar Freq) GPU freq: %d MHz (%u)\n",
5254                          intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
5255                          dev_priv->rps.rp1_freq);
5256
5257         dev_priv->rps.min_freq = valleyview_rps_min_freq(dev_priv);
5258         DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5259                          intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
5260                          dev_priv->rps.min_freq);
5261
5262         dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
5263
5264         /* Preserve min/max settings in case of re-init */
5265         if (dev_priv->rps.max_freq_softlimit == 0)
5266                 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
5267
5268         if (dev_priv->rps.min_freq_softlimit == 0)
5269                 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
5270
5271         mutex_unlock(&dev_priv->rps.hw_lock);
5272 }
5273
5274 static void cherryview_init_gt_powersave(struct drm_device *dev)
5275 {
5276         struct drm_i915_private *dev_priv = dev->dev_private;
5277         u32 val;
5278
5279         cherryview_setup_pctx(dev);
5280
5281         mutex_lock(&dev_priv->rps.hw_lock);
5282
5283         mutex_lock(&dev_priv->sb_lock);
5284         val = vlv_cck_read(dev_priv, CCK_FUSE_REG);
5285         mutex_unlock(&dev_priv->sb_lock);
5286
5287         switch ((val >> 2) & 0x7) {
5288         case 3:
5289                 dev_priv->mem_freq = 2000;
5290                 break;
5291         default:
5292                 dev_priv->mem_freq = 1600;
5293                 break;
5294         }
5295         DRM_DEBUG_DRIVER("DDR speed: %d MHz\n", dev_priv->mem_freq);
5296
5297         dev_priv->rps.max_freq = cherryview_rps_max_freq(dev_priv);
5298         dev_priv->rps.rp0_freq = dev_priv->rps.max_freq;
5299         DRM_DEBUG_DRIVER("max GPU freq: %d MHz (%u)\n",
5300                          intel_gpu_freq(dev_priv, dev_priv->rps.max_freq),
5301                          dev_priv->rps.max_freq);
5302
5303         dev_priv->rps.efficient_freq = cherryview_rps_rpe_freq(dev_priv);
5304         DRM_DEBUG_DRIVER("RPe GPU freq: %d MHz (%u)\n",
5305                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5306                          dev_priv->rps.efficient_freq);
5307
5308         dev_priv->rps.rp1_freq = cherryview_rps_guar_freq(dev_priv);
5309         DRM_DEBUG_DRIVER("RP1(Guar) GPU freq: %d MHz (%u)\n",
5310                          intel_gpu_freq(dev_priv, dev_priv->rps.rp1_freq),
5311                          dev_priv->rps.rp1_freq);
5312
5313         /* PUnit validated range is only [RPe, RP0] */
5314         dev_priv->rps.min_freq = dev_priv->rps.efficient_freq;
5315         DRM_DEBUG_DRIVER("min GPU freq: %d MHz (%u)\n",
5316                          intel_gpu_freq(dev_priv, dev_priv->rps.min_freq),
5317                          dev_priv->rps.min_freq);
5318
5319         WARN_ONCE((dev_priv->rps.max_freq |
5320                    dev_priv->rps.efficient_freq |
5321                    dev_priv->rps.rp1_freq |
5322                    dev_priv->rps.min_freq) & 1,
5323                   "Odd GPU freq values\n");
5324
5325         dev_priv->rps.idle_freq = dev_priv->rps.min_freq;
5326
5327         /* Preserve min/max settings in case of re-init */
5328         if (dev_priv->rps.max_freq_softlimit == 0)
5329                 dev_priv->rps.max_freq_softlimit = dev_priv->rps.max_freq;
5330
5331         if (dev_priv->rps.min_freq_softlimit == 0)
5332                 dev_priv->rps.min_freq_softlimit = dev_priv->rps.min_freq;
5333
5334         mutex_unlock(&dev_priv->rps.hw_lock);
5335 }
5336
5337 static void valleyview_cleanup_gt_powersave(struct drm_device *dev)
5338 {
5339         valleyview_cleanup_pctx(dev);
5340 }
5341
5342 static void cherryview_enable_rps(struct drm_device *dev)
5343 {
5344         struct drm_i915_private *dev_priv = dev->dev_private;
5345         struct intel_engine_cs *ring;
5346         u32 gtfifodbg, val, rc6_mode = 0, pcbr;
5347         int i;
5348
5349         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5350
5351         gtfifodbg = I915_READ(GTFIFODBG);
5352         if (gtfifodbg) {
5353                 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5354                                  gtfifodbg);
5355                 I915_WRITE(GTFIFODBG, gtfifodbg);
5356         }
5357
5358         cherryview_check_pctx(dev_priv);
5359
5360         /* 1a & 1b: Get forcewake during program sequence. Although the driver
5361          * hasn't enabled a state yet where we need forcewake, BIOS may have.*/
5362         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5363
5364         /*  Disable RC states. */
5365         I915_WRITE(GEN6_RC_CONTROL, 0);
5366
5367         /* 2a: Program RC6 thresholds.*/
5368         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 40 << 16);
5369         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000); /* 12500 * 1280ns */
5370         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25); /* 25 * 1280ns */
5371
5372         for_each_ring(ring, dev_priv, i)
5373                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
5374         I915_WRITE(GEN6_RC_SLEEP, 0);
5375
5376         /* TO threshold set to 500 us ( 0x186 * 1.28 us) */
5377         I915_WRITE(GEN6_RC6_THRESHOLD, 0x186);
5378
5379         /* allows RC6 residency counter to work */
5380         I915_WRITE(VLV_COUNTER_CONTROL,
5381                    _MASKED_BIT_ENABLE(VLV_COUNT_RANGE_HIGH |
5382                                       VLV_MEDIA_RC6_COUNT_EN |
5383                                       VLV_RENDER_RC6_COUNT_EN));
5384
5385         /* For now we assume BIOS is allocating and populating the PCBR  */
5386         pcbr = I915_READ(VLV_PCBR);
5387
5388         /* 3: Enable RC6 */
5389         if ((intel_enable_rc6(dev) & INTEL_RC6_ENABLE) &&
5390                                                 (pcbr >> VLV_PCBR_ADDR_SHIFT))
5391                 rc6_mode = GEN7_RC_CTL_TO_MODE;
5392
5393         I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
5394
5395         /* 4 Program defaults and thresholds for RPS*/
5396         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
5397         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
5398         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
5399         I915_WRITE(GEN6_RP_UP_EI, 66000);
5400         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
5401
5402         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5403
5404         /* 5: Enable RPS */
5405         I915_WRITE(GEN6_RP_CONTROL,
5406                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
5407                    GEN6_RP_MEDIA_IS_GFX |
5408                    GEN6_RP_ENABLE |
5409                    GEN6_RP_UP_BUSY_AVG |
5410                    GEN6_RP_DOWN_IDLE_AVG);
5411
5412         /* Setting Fixed Bias */
5413         val = VLV_OVERRIDE_EN |
5414                   VLV_SOC_TDP_EN |
5415                   CHV_BIAS_CPU_50_SOC_50;
5416         vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
5417
5418         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5419
5420         /* RPS code assumes GPLL is used */
5421         WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
5422
5423         DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
5424         DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
5425
5426         dev_priv->rps.cur_freq = (val >> 8) & 0xff;
5427         DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
5428                          intel_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
5429                          dev_priv->rps.cur_freq);
5430
5431         DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
5432                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5433                          dev_priv->rps.efficient_freq);
5434
5435         valleyview_set_rps(dev_priv->dev, dev_priv->rps.efficient_freq);
5436
5437         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5438 }
5439
5440 static void valleyview_enable_rps(struct drm_device *dev)
5441 {
5442         struct drm_i915_private *dev_priv = dev->dev_private;
5443         struct intel_engine_cs *ring;
5444         u32 gtfifodbg, val, rc6_mode = 0;
5445         int i;
5446
5447         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
5448
5449         valleyview_check_pctx(dev_priv);
5450
5451         if ((gtfifodbg = I915_READ(GTFIFODBG))) {
5452                 DRM_DEBUG_DRIVER("GT fifo had a previous error %x\n",
5453                                  gtfifodbg);
5454                 I915_WRITE(GTFIFODBG, gtfifodbg);
5455         }
5456
5457         /* If VLV, Forcewake all wells, else re-direct to regular path */
5458         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
5459
5460         /*  Disable RC states. */
5461         I915_WRITE(GEN6_RC_CONTROL, 0);
5462
5463         I915_WRITE(GEN6_RP_DOWN_TIMEOUT, 1000000);
5464         I915_WRITE(GEN6_RP_UP_THRESHOLD, 59400);
5465         I915_WRITE(GEN6_RP_DOWN_THRESHOLD, 245000);
5466         I915_WRITE(GEN6_RP_UP_EI, 66000);
5467         I915_WRITE(GEN6_RP_DOWN_EI, 350000);
5468
5469         I915_WRITE(GEN6_RP_IDLE_HYSTERSIS, 10);
5470
5471         I915_WRITE(GEN6_RP_CONTROL,
5472                    GEN6_RP_MEDIA_TURBO |
5473                    GEN6_RP_MEDIA_HW_NORMAL_MODE |
5474                    GEN6_RP_MEDIA_IS_GFX |
5475                    GEN6_RP_ENABLE |
5476                    GEN6_RP_UP_BUSY_AVG |
5477                    GEN6_RP_DOWN_IDLE_CONT);
5478
5479         I915_WRITE(GEN6_RC6_WAKE_RATE_LIMIT, 0x00280000);
5480         I915_WRITE(GEN6_RC_EVALUATION_INTERVAL, 125000);
5481         I915_WRITE(GEN6_RC_IDLE_HYSTERSIS, 25);
5482
5483         for_each_ring(ring, dev_priv, i)
5484                 I915_WRITE(RING_MAX_IDLE(ring->mmio_base), 10);
5485
5486         I915_WRITE(GEN6_RC6_THRESHOLD, 0x557);
5487
5488         /* allows RC6 residency counter to work */
5489         I915_WRITE(VLV_COUNTER_CONTROL,
5490                    _MASKED_BIT_ENABLE(VLV_MEDIA_RC0_COUNT_EN |
5491                                       VLV_RENDER_RC0_COUNT_EN |
5492                                       VLV_MEDIA_RC6_COUNT_EN |
5493                                       VLV_RENDER_RC6_COUNT_EN));
5494
5495         if (intel_enable_rc6(dev) & INTEL_RC6_ENABLE)
5496                 rc6_mode = GEN7_RC_CTL_TO_MODE | VLV_RC_CTL_CTX_RST_PARALLEL;
5497
5498         intel_print_rc6_info(dev, rc6_mode);
5499
5500         I915_WRITE(GEN6_RC_CONTROL, rc6_mode);
5501
5502         /* Setting Fixed Bias */
5503         val = VLV_OVERRIDE_EN |
5504                   VLV_SOC_TDP_EN |
5505                   VLV_BIAS_CPU_125_SOC_875;
5506         vlv_punit_write(dev_priv, VLV_TURBO_SOC_OVERRIDE, val);
5507
5508         val = vlv_punit_read(dev_priv, PUNIT_REG_GPU_FREQ_STS);
5509
5510         /* RPS code assumes GPLL is used */
5511         WARN_ONCE((val & GPLLENABLE) == 0, "GPLL not enabled\n");
5512
5513         DRM_DEBUG_DRIVER("GPLL enabled? %s\n", yesno(val & GPLLENABLE));
5514         DRM_DEBUG_DRIVER("GPU status: 0x%08x\n", val);
5515
5516         dev_priv->rps.cur_freq = (val >> 8) & 0xff;
5517         DRM_DEBUG_DRIVER("current GPU freq: %d MHz (%u)\n",
5518                          intel_gpu_freq(dev_priv, dev_priv->rps.cur_freq),
5519                          dev_priv->rps.cur_freq);
5520
5521         DRM_DEBUG_DRIVER("setting GPU freq to %d MHz (%u)\n",
5522                          intel_gpu_freq(dev_priv, dev_priv->rps.efficient_freq),
5523                          dev_priv->rps.efficient_freq);
5524
5525         valleyview_set_rps(dev_priv->dev, dev_priv->rps.efficient_freq);
5526
5527         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
5528 }
5529
5530 static unsigned long intel_pxfreq(u32 vidfreq)
5531 {
5532         unsigned long freq;
5533         int div = (vidfreq & 0x3f0000) >> 16;
5534         int post = (vidfreq & 0x3000) >> 12;
5535         int pre = (vidfreq & 0x7);
5536
5537         if (!pre)
5538                 return 0;
5539
5540         freq = ((div * 133333) / ((1<<post) * pre));
5541
5542         return freq;
5543 }
5544
5545 static const struct cparams {
5546         u16 i;
5547         u16 t;
5548         u16 m;
5549         u16 c;
5550 } cparams[] = {
5551         { 1, 1333, 301, 28664 },
5552         { 1, 1066, 294, 24460 },
5553         { 1, 800, 294, 25192 },
5554         { 0, 1333, 276, 27605 },
5555         { 0, 1066, 276, 27605 },
5556         { 0, 800, 231, 23784 },
5557 };
5558
5559 static unsigned long __i915_chipset_val(struct drm_i915_private *dev_priv)
5560 {
5561         u64 total_count, diff, ret;
5562         u32 count1, count2, count3, m = 0, c = 0;
5563         unsigned long now = jiffies_to_msecs(jiffies), diff1;
5564         int i;
5565
5566         assert_spin_locked(&mchdev_lock);
5567
5568         diff1 = now - dev_priv->ips.last_time1;
5569
5570         /* Prevent division-by-zero if we are asking too fast.
5571          * Also, we don't get interesting results if we are polling
5572          * faster than once in 10ms, so just return the saved value
5573          * in such cases.
5574          */
5575         if (diff1 <= 10)
5576                 return dev_priv->ips.chipset_power;
5577
5578         count1 = I915_READ(DMIEC);
5579         count2 = I915_READ(DDREC);
5580         count3 = I915_READ(CSIEC);
5581
5582         total_count = count1 + count2 + count3;
5583
5584         /* FIXME: handle per-counter overflow */
5585         if (total_count < dev_priv->ips.last_count1) {
5586                 diff = ~0UL - dev_priv->ips.last_count1;
5587                 diff += total_count;
5588         } else {
5589                 diff = total_count - dev_priv->ips.last_count1;
5590         }
5591
5592         for (i = 0; i < ARRAY_SIZE(cparams); i++) {
5593                 if (cparams[i].i == dev_priv->ips.c_m &&
5594                     cparams[i].t == dev_priv->ips.r_t) {
5595                         m = cparams[i].m;
5596                         c = cparams[i].c;
5597                         break;
5598                 }
5599         }
5600
5601         diff = div_u64(diff, diff1);
5602         ret = ((m * diff) + c);
5603         ret = div_u64(ret, 10);
5604
5605         dev_priv->ips.last_count1 = total_count;
5606         dev_priv->ips.last_time1 = now;
5607
5608         dev_priv->ips.chipset_power = ret;
5609
5610         return ret;
5611 }
5612
5613 unsigned long i915_chipset_val(struct drm_i915_private *dev_priv)
5614 {
5615         struct drm_device *dev = dev_priv->dev;
5616         unsigned long val;
5617
5618         if (INTEL_INFO(dev)->gen != 5)
5619                 return 0;
5620
5621         spin_lock_irq(&mchdev_lock);
5622
5623         val = __i915_chipset_val(dev_priv);
5624
5625         spin_unlock_irq(&mchdev_lock);
5626
5627         return val;
5628 }
5629
5630 unsigned long i915_mch_val(struct drm_i915_private *dev_priv)
5631 {
5632         unsigned long m, x, b;
5633         u32 tsfs;
5634
5635         tsfs = I915_READ(TSFS);
5636
5637         m = ((tsfs & TSFS_SLOPE_MASK) >> TSFS_SLOPE_SHIFT);
5638         x = I915_READ8(TR1);
5639
5640         b = tsfs & TSFS_INTR_MASK;
5641
5642         return ((m * x) / 127) - b;
5643 }
5644
5645 static int _pxvid_to_vd(u8 pxvid)
5646 {
5647         if (pxvid == 0)
5648                 return 0;
5649
5650         if (pxvid >= 8 && pxvid < 31)
5651                 pxvid = 31;
5652
5653         return (pxvid + 2) * 125;
5654 }
5655
5656 static u32 pvid_to_extvid(struct drm_i915_private *dev_priv, u8 pxvid)
5657 {
5658         struct drm_device *dev = dev_priv->dev;
5659         const int vd = _pxvid_to_vd(pxvid);
5660         const int vm = vd - 1125;
5661
5662         if (INTEL_INFO(dev)->is_mobile)
5663                 return vm > 0 ? vm : 0;
5664
5665         return vd;
5666 }
5667
5668 static void __i915_update_gfx_val(struct drm_i915_private *dev_priv)
5669 {
5670         u64 now, diff, diffms;
5671         u32 count;
5672
5673         assert_spin_locked(&mchdev_lock);
5674
5675         now = ktime_get_raw_ns();
5676         diffms = now - dev_priv->ips.last_time2;
5677         do_div(diffms, NSEC_PER_MSEC);
5678
5679         /* Don't divide by 0 */
5680         if (!diffms)
5681                 return;
5682
5683         count = I915_READ(GFXEC);
5684
5685         if (count < dev_priv->ips.last_count2) {
5686                 diff = ~0UL - dev_priv->ips.last_count2;
5687                 diff += count;
5688         } else {
5689                 diff = count - dev_priv->ips.last_count2;
5690         }
5691
5692         dev_priv->ips.last_count2 = count;
5693         dev_priv->ips.last_time2 = now;
5694
5695         /* More magic constants... */
5696         diff = diff * 1181;
5697         diff = div_u64(diff, diffms * 10);
5698         dev_priv->ips.gfx_power = diff;
5699 }
5700
5701 void i915_update_gfx_val(struct drm_i915_private *dev_priv)
5702 {
5703         struct drm_device *dev = dev_priv->dev;
5704
5705         if (INTEL_INFO(dev)->gen != 5)
5706                 return;
5707
5708         spin_lock_irq(&mchdev_lock);
5709
5710         __i915_update_gfx_val(dev_priv);
5711
5712         spin_unlock_irq(&mchdev_lock);
5713 }
5714
5715 static unsigned long __i915_gfx_val(struct drm_i915_private *dev_priv)
5716 {
5717         unsigned long t, corr, state1, corr2, state2;
5718         u32 pxvid, ext_v;
5719
5720         assert_spin_locked(&mchdev_lock);
5721
5722         pxvid = I915_READ(PXVFREQ(dev_priv->rps.cur_freq));
5723         pxvid = (pxvid >> 24) & 0x7f;
5724         ext_v = pvid_to_extvid(dev_priv, pxvid);
5725
5726         state1 = ext_v;
5727
5728         t = i915_mch_val(dev_priv);
5729
5730         /* Revel in the empirically derived constants */
5731
5732         /* Correction factor in 1/100000 units */
5733         if (t > 80)
5734                 corr = ((t * 2349) + 135940);
5735         else if (t >= 50)
5736                 corr = ((t * 964) + 29317);
5737         else /* < 50 */
5738                 corr = ((t * 301) + 1004);
5739
5740         corr = corr * ((150142 * state1) / 10000 - 78642);
5741         corr /= 100000;
5742         corr2 = (corr * dev_priv->ips.corr);
5743
5744         state2 = (corr2 * state1) / 10000;
5745         state2 /= 100; /* convert to mW */
5746
5747         __i915_update_gfx_val(dev_priv);
5748
5749         return dev_priv->ips.gfx_power + state2;
5750 }
5751
5752 unsigned long i915_gfx_val(struct drm_i915_private *dev_priv)
5753 {
5754         struct drm_device *dev = dev_priv->dev;
5755         unsigned long val;
5756
5757         if (INTEL_INFO(dev)->gen != 5)
5758                 return 0;
5759
5760         spin_lock_irq(&mchdev_lock);
5761
5762         val = __i915_gfx_val(dev_priv);
5763
5764         spin_unlock_irq(&mchdev_lock);
5765
5766         return val;
5767 }
5768
5769 /**
5770  * i915_read_mch_val - return value for IPS use
5771  *
5772  * Calculate and return a value for the IPS driver to use when deciding whether
5773  * we have thermal and power headroom to increase CPU or GPU power budget.
5774  */
5775 unsigned long i915_read_mch_val(void)
5776 {
5777         struct drm_i915_private *dev_priv;
5778         unsigned long chipset_val, graphics_val, ret = 0;
5779
5780         spin_lock_irq(&mchdev_lock);
5781         if (!i915_mch_dev)
5782                 goto out_unlock;
5783         dev_priv = i915_mch_dev;
5784
5785         chipset_val = __i915_chipset_val(dev_priv);
5786         graphics_val = __i915_gfx_val(dev_priv);
5787
5788         ret = chipset_val + graphics_val;
5789
5790 out_unlock:
5791         spin_unlock_irq(&mchdev_lock);
5792
5793         return ret;
5794 }
5795 EXPORT_SYMBOL_GPL(i915_read_mch_val);
5796
5797 /**
5798  * i915_gpu_raise - raise GPU frequency limit
5799  *
5800  * Raise the limit; IPS indicates we have thermal headroom.
5801  */
5802 bool i915_gpu_raise(void)
5803 {
5804         struct drm_i915_private *dev_priv;
5805         bool ret = true;
5806
5807         spin_lock_irq(&mchdev_lock);
5808         if (!i915_mch_dev) {
5809                 ret = false;
5810                 goto out_unlock;
5811         }
5812         dev_priv = i915_mch_dev;
5813
5814         if (dev_priv->ips.max_delay > dev_priv->ips.fmax)
5815                 dev_priv->ips.max_delay--;
5816
5817 out_unlock:
5818         spin_unlock_irq(&mchdev_lock);
5819
5820         return ret;
5821 }
5822 EXPORT_SYMBOL_GPL(i915_gpu_raise);
5823
5824 /**
5825  * i915_gpu_lower - lower GPU frequency limit
5826  *
5827  * IPS indicates we're close to a thermal limit, so throttle back the GPU
5828  * frequency maximum.
5829  */
5830 bool i915_gpu_lower(void)
5831 {
5832         struct drm_i915_private *dev_priv;
5833         bool ret = true;
5834
5835         spin_lock_irq(&mchdev_lock);
5836         if (!i915_mch_dev) {
5837                 ret = false;
5838                 goto out_unlock;
5839         }
5840         dev_priv = i915_mch_dev;
5841
5842         if (dev_priv->ips.max_delay < dev_priv->ips.min_delay)
5843                 dev_priv->ips.max_delay++;
5844
5845 out_unlock:
5846         spin_unlock_irq(&mchdev_lock);
5847
5848         return ret;
5849 }
5850 EXPORT_SYMBOL_GPL(i915_gpu_lower);
5851
5852 /**
5853  * i915_gpu_busy - indicate GPU business to IPS
5854  *
5855  * Tell the IPS driver whether or not the GPU is busy.
5856  */
5857 bool i915_gpu_busy(void)
5858 {
5859         struct drm_i915_private *dev_priv;
5860         struct intel_engine_cs *ring;
5861         bool ret = false;
5862         int i;
5863
5864         spin_lock_irq(&mchdev_lock);
5865         if (!i915_mch_dev)
5866                 goto out_unlock;
5867         dev_priv = i915_mch_dev;
5868
5869         for_each_ring(ring, dev_priv, i)
5870                 ret |= !list_empty(&ring->request_list);
5871
5872 out_unlock:
5873         spin_unlock_irq(&mchdev_lock);
5874
5875         return ret;
5876 }
5877 EXPORT_SYMBOL_GPL(i915_gpu_busy);
5878
5879 /**
5880  * i915_gpu_turbo_disable - disable graphics turbo
5881  *
5882  * Disable graphics turbo by resetting the max frequency and setting the
5883  * current frequency to the default.
5884  */
5885 bool i915_gpu_turbo_disable(void)
5886 {
5887         struct drm_i915_private *dev_priv;
5888         bool ret = true;
5889
5890         spin_lock_irq(&mchdev_lock);
5891         if (!i915_mch_dev) {
5892                 ret = false;
5893                 goto out_unlock;
5894         }
5895         dev_priv = i915_mch_dev;
5896
5897         dev_priv->ips.max_delay = dev_priv->ips.fstart;
5898
5899         if (!ironlake_set_drps(dev_priv->dev, dev_priv->ips.fstart))
5900                 ret = false;
5901
5902 out_unlock:
5903         spin_unlock_irq(&mchdev_lock);
5904
5905         return ret;
5906 }
5907 EXPORT_SYMBOL_GPL(i915_gpu_turbo_disable);
5908
5909 /**
5910  * Tells the intel_ips driver that the i915 driver is now loaded, if
5911  * IPS got loaded first.
5912  *
5913  * This awkward dance is so that neither module has to depend on the
5914  * other in order for IPS to do the appropriate communication of
5915  * GPU turbo limits to i915.
5916  */
5917 static void
5918 ips_ping_for_i915_load(void)
5919 {
5920         void (*link)(void);
5921
5922         link = symbol_get(ips_link_to_i915_driver);
5923         if (link) {
5924                 link();
5925                 symbol_put(ips_link_to_i915_driver);
5926         }
5927 }
5928
5929 void intel_gpu_ips_init(struct drm_i915_private *dev_priv)
5930 {
5931         /* We only register the i915 ips part with intel-ips once everything is
5932          * set up, to avoid intel-ips sneaking in and reading bogus values. */
5933         spin_lock_irq(&mchdev_lock);
5934         i915_mch_dev = dev_priv;
5935         spin_unlock_irq(&mchdev_lock);
5936
5937         ips_ping_for_i915_load();
5938 }
5939
5940 void intel_gpu_ips_teardown(void)
5941 {
5942         spin_lock_irq(&mchdev_lock);
5943         i915_mch_dev = NULL;
5944         spin_unlock_irq(&mchdev_lock);
5945 }
5946
5947 static void intel_init_emon(struct drm_device *dev)
5948 {
5949         struct drm_i915_private *dev_priv = dev->dev_private;
5950         u32 lcfuse;
5951         u8 pxw[16];
5952         int i;
5953
5954         /* Disable to program */
5955         I915_WRITE(ECR, 0);
5956         POSTING_READ(ECR);
5957
5958         /* Program energy weights for various events */
5959         I915_WRITE(SDEW, 0x15040d00);
5960         I915_WRITE(CSIEW0, 0x007f0000);
5961         I915_WRITE(CSIEW1, 0x1e220004);
5962         I915_WRITE(CSIEW2, 0x04000004);
5963
5964         for (i = 0; i < 5; i++)
5965                 I915_WRITE(PEW(i), 0);
5966         for (i = 0; i < 3; i++)
5967                 I915_WRITE(DEW(i), 0);
5968
5969         /* Program P-state weights to account for frequency power adjustment */
5970         for (i = 0; i < 16; i++) {
5971                 u32 pxvidfreq = I915_READ(PXVFREQ(i));
5972                 unsigned long freq = intel_pxfreq(pxvidfreq);
5973                 unsigned long vid = (pxvidfreq & PXVFREQ_PX_MASK) >>
5974                         PXVFREQ_PX_SHIFT;
5975                 unsigned long val;
5976
5977                 val = vid * vid;
5978                 val *= (freq / 1000);
5979                 val *= 255;
5980                 val /= (127*127*900);
5981                 if (val > 0xff)
5982                         DRM_ERROR("bad pxval: %ld\n", val);
5983                 pxw[i] = val;
5984         }
5985         /* Render standby states get 0 weight */
5986         pxw[14] = 0;
5987         pxw[15] = 0;
5988
5989         for (i = 0; i < 4; i++) {
5990                 u32 val = (pxw[i*4] << 24) | (pxw[(i*4)+1] << 16) |
5991                         (pxw[(i*4)+2] << 8) | (pxw[(i*4)+3]);
5992                 I915_WRITE(PXW(i), val);
5993         }
5994
5995         /* Adjust magic regs to magic values (more experimental results) */
5996         I915_WRITE(OGW0, 0);
5997         I915_WRITE(OGW1, 0);
5998         I915_WRITE(EG0, 0x00007f00);
5999         I915_WRITE(EG1, 0x0000000e);
6000         I915_WRITE(EG2, 0x000e0000);
6001         I915_WRITE(EG3, 0x68000300);
6002         I915_WRITE(EG4, 0x42000000);
6003         I915_WRITE(EG5, 0x00140031);
6004         I915_WRITE(EG6, 0);
6005         I915_WRITE(EG7, 0);
6006
6007         for (i = 0; i < 8; i++)
6008                 I915_WRITE(PXWL(i), 0);
6009
6010         /* Enable PMON + select events */
6011         I915_WRITE(ECR, 0x80000019);
6012
6013         lcfuse = I915_READ(LCFUSE02);
6014
6015         dev_priv->ips.corr = (lcfuse & LCFUSE_HIV_MASK);
6016 }
6017
6018 void intel_init_gt_powersave(struct drm_device *dev)
6019 {
6020         struct drm_i915_private *dev_priv = dev->dev_private;
6021
6022         i915.enable_rc6 = sanitize_rc6_option(dev, i915.enable_rc6);
6023         /*
6024          * RPM depends on RC6 to save restore the GT HW context, so make RC6 a
6025          * requirement.
6026          */
6027         if (!i915.enable_rc6) {
6028                 DRM_INFO("RC6 disabled, disabling runtime PM support\n");
6029                 intel_runtime_pm_get(dev_priv);
6030         }
6031
6032         if (IS_CHERRYVIEW(dev))
6033                 cherryview_init_gt_powersave(dev);
6034         else if (IS_VALLEYVIEW(dev))
6035                 valleyview_init_gt_powersave(dev);
6036 }
6037
6038 void intel_cleanup_gt_powersave(struct drm_device *dev)
6039 {
6040         struct drm_i915_private *dev_priv = dev->dev_private;
6041
6042         if (IS_CHERRYVIEW(dev))
6043                 return;
6044         else if (IS_VALLEYVIEW(dev))
6045                 valleyview_cleanup_gt_powersave(dev);
6046
6047         if (!i915.enable_rc6)
6048                 intel_runtime_pm_put(dev_priv);
6049 }
6050
6051 static void gen6_suspend_rps(struct drm_device *dev)
6052 {
6053         struct drm_i915_private *dev_priv = dev->dev_private;
6054
6055         flush_delayed_work(&dev_priv->rps.delayed_resume_work);
6056
6057         gen6_disable_rps_interrupts(dev);
6058 }
6059
6060 /**
6061  * intel_suspend_gt_powersave - suspend PM work and helper threads
6062  * @dev: drm device
6063  *
6064  * We don't want to disable RC6 or other features here, we just want
6065  * to make sure any work we've queued has finished and won't bother
6066  * us while we're suspended.
6067  */
6068 void intel_suspend_gt_powersave(struct drm_device *dev)
6069 {
6070         struct drm_i915_private *dev_priv = dev->dev_private;
6071
6072         if (INTEL_INFO(dev)->gen < 6)
6073                 return;
6074
6075         gen6_suspend_rps(dev);
6076
6077         /* Force GPU to min freq during suspend */
6078         gen6_rps_idle(dev_priv);
6079 }
6080
6081 void intel_disable_gt_powersave(struct drm_device *dev)
6082 {
6083         struct drm_i915_private *dev_priv = dev->dev_private;
6084
6085         if (IS_IRONLAKE_M(dev)) {
6086                 ironlake_disable_drps(dev);
6087         } else if (INTEL_INFO(dev)->gen >= 6) {
6088                 intel_suspend_gt_powersave(dev);
6089
6090                 mutex_lock(&dev_priv->rps.hw_lock);
6091                 if (INTEL_INFO(dev)->gen >= 9)
6092                         gen9_disable_rps(dev);
6093                 else if (IS_CHERRYVIEW(dev))
6094                         cherryview_disable_rps(dev);
6095                 else if (IS_VALLEYVIEW(dev))
6096                         valleyview_disable_rps(dev);
6097                 else
6098                         gen6_disable_rps(dev);
6099
6100                 dev_priv->rps.enabled = false;
6101                 mutex_unlock(&dev_priv->rps.hw_lock);
6102         }
6103 }
6104
6105 static void intel_gen6_powersave_work(struct work_struct *work)
6106 {
6107         struct drm_i915_private *dev_priv =
6108                 container_of(work, struct drm_i915_private,
6109                              rps.delayed_resume_work.work);
6110         struct drm_device *dev = dev_priv->dev;
6111
6112         mutex_lock(&dev_priv->rps.hw_lock);
6113
6114         gen6_reset_rps_interrupts(dev);
6115
6116         if (IS_CHERRYVIEW(dev)) {
6117                 cherryview_enable_rps(dev);
6118         } else if (IS_VALLEYVIEW(dev)) {
6119                 valleyview_enable_rps(dev);
6120         } else if (INTEL_INFO(dev)->gen >= 9) {
6121                 gen9_enable_rc6(dev);
6122                 gen9_enable_rps(dev);
6123                 if (IS_SKYLAKE(dev) || IS_KABYLAKE(dev))
6124                         __gen6_update_ring_freq(dev);
6125         } else if (IS_BROADWELL(dev)) {
6126                 gen8_enable_rps(dev);
6127                 __gen6_update_ring_freq(dev);
6128         } else {
6129                 gen6_enable_rps(dev);
6130                 __gen6_update_ring_freq(dev);
6131         }
6132
6133         WARN_ON(dev_priv->rps.max_freq < dev_priv->rps.min_freq);
6134         WARN_ON(dev_priv->rps.idle_freq > dev_priv->rps.max_freq);
6135
6136         WARN_ON(dev_priv->rps.efficient_freq < dev_priv->rps.min_freq);
6137         WARN_ON(dev_priv->rps.efficient_freq > dev_priv->rps.max_freq);
6138
6139         dev_priv->rps.enabled = true;
6140
6141         gen6_enable_rps_interrupts(dev);
6142
6143         mutex_unlock(&dev_priv->rps.hw_lock);
6144
6145         intel_runtime_pm_put(dev_priv);
6146 }
6147
6148 void intel_enable_gt_powersave(struct drm_device *dev)
6149 {
6150         struct drm_i915_private *dev_priv = dev->dev_private;
6151
6152         /* Powersaving is controlled by the host when inside a VM */
6153         if (intel_vgpu_active(dev))
6154                 return;
6155
6156         if (IS_IRONLAKE_M(dev)) {
6157                 mutex_lock(&dev->struct_mutex);
6158                 ironlake_enable_drps(dev);
6159                 intel_init_emon(dev);
6160                 mutex_unlock(&dev->struct_mutex);
6161         } else if (INTEL_INFO(dev)->gen >= 6) {
6162                 /*
6163                  * PCU communication is slow and this doesn't need to be
6164                  * done at any specific time, so do this out of our fast path
6165                  * to make resume and init faster.
6166                  *
6167                  * We depend on the HW RC6 power context save/restore
6168                  * mechanism when entering D3 through runtime PM suspend. So
6169                  * disable RPM until RPS/RC6 is properly setup. We can only
6170                  * get here via the driver load/system resume/runtime resume
6171                  * paths, so the _noresume version is enough (and in case of
6172                  * runtime resume it's necessary).
6173                  */
6174                 if (schedule_delayed_work(&dev_priv->rps.delayed_resume_work,
6175                                            round_jiffies_up_relative(HZ)))
6176                         intel_runtime_pm_get_noresume(dev_priv);
6177         }
6178 }
6179
6180 void intel_reset_gt_powersave(struct drm_device *dev)
6181 {
6182         struct drm_i915_private *dev_priv = dev->dev_private;
6183
6184         if (INTEL_INFO(dev)->gen < 6)
6185                 return;
6186
6187         gen6_suspend_rps(dev);
6188         dev_priv->rps.enabled = false;
6189 }
6190
6191 static void ibx_init_clock_gating(struct drm_device *dev)
6192 {
6193         struct drm_i915_private *dev_priv = dev->dev_private;
6194
6195         /*
6196          * On Ibex Peak and Cougar Point, we need to disable clock
6197          * gating for the panel power sequencer or it will fail to
6198          * start up when no ports are active.
6199          */
6200         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE);
6201 }
6202
6203 static void g4x_disable_trickle_feed(struct drm_device *dev)
6204 {
6205         struct drm_i915_private *dev_priv = dev->dev_private;
6206         enum pipe pipe;
6207
6208         for_each_pipe(dev_priv, pipe) {
6209                 I915_WRITE(DSPCNTR(pipe),
6210                            I915_READ(DSPCNTR(pipe)) |
6211                            DISPPLANE_TRICKLE_FEED_DISABLE);
6212
6213                 I915_WRITE(DSPSURF(pipe), I915_READ(DSPSURF(pipe)));
6214                 POSTING_READ(DSPSURF(pipe));
6215         }
6216 }
6217
6218 static void ilk_init_lp_watermarks(struct drm_device *dev)
6219 {
6220         struct drm_i915_private *dev_priv = dev->dev_private;
6221
6222         I915_WRITE(WM3_LP_ILK, I915_READ(WM3_LP_ILK) & ~WM1_LP_SR_EN);
6223         I915_WRITE(WM2_LP_ILK, I915_READ(WM2_LP_ILK) & ~WM1_LP_SR_EN);
6224         I915_WRITE(WM1_LP_ILK, I915_READ(WM1_LP_ILK) & ~WM1_LP_SR_EN);
6225
6226         /*
6227          * Don't touch WM1S_LP_EN here.
6228          * Doing so could cause underruns.
6229          */
6230 }
6231
6232 static void ironlake_init_clock_gating(struct drm_device *dev)
6233 {
6234         struct drm_i915_private *dev_priv = dev->dev_private;
6235         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6236
6237         /*
6238          * Required for FBC
6239          * WaFbcDisableDpfcClockGating:ilk
6240          */
6241         dspclk_gate |= ILK_DPFCRUNIT_CLOCK_GATE_DISABLE |
6242                    ILK_DPFCUNIT_CLOCK_GATE_DISABLE |
6243                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE;
6244
6245         I915_WRITE(PCH_3DCGDIS0,
6246                    MARIUNIT_CLOCK_GATE_DISABLE |
6247                    SVSMUNIT_CLOCK_GATE_DISABLE);
6248         I915_WRITE(PCH_3DCGDIS1,
6249                    VFMUNIT_CLOCK_GATE_DISABLE);
6250
6251         /*
6252          * According to the spec the following bits should be set in
6253          * order to enable memory self-refresh
6254          * The bit 22/21 of 0x42004
6255          * The bit 5 of 0x42020
6256          * The bit 15 of 0x45000
6257          */
6258         I915_WRITE(ILK_DISPLAY_CHICKEN2,
6259                    (I915_READ(ILK_DISPLAY_CHICKEN2) |
6260                     ILK_DPARB_GATE | ILK_VSDPFD_FULL));
6261         dspclk_gate |= ILK_DPARBUNIT_CLOCK_GATE_ENABLE;
6262         I915_WRITE(DISP_ARB_CTL,
6263                    (I915_READ(DISP_ARB_CTL) |
6264                     DISP_FBC_WM_DIS));
6265
6266         ilk_init_lp_watermarks(dev);
6267
6268         /*
6269          * Based on the document from hardware guys the following bits
6270          * should be set unconditionally in order to enable FBC.
6271          * The bit 22 of 0x42000
6272          * The bit 22 of 0x42004
6273          * The bit 7,8,9 of 0x42020.
6274          */
6275         if (IS_IRONLAKE_M(dev)) {
6276                 /* WaFbcAsynchFlipDisableFbcQueue:ilk */
6277                 I915_WRITE(ILK_DISPLAY_CHICKEN1,
6278                            I915_READ(ILK_DISPLAY_CHICKEN1) |
6279                            ILK_FBCQ_DIS);
6280                 I915_WRITE(ILK_DISPLAY_CHICKEN2,
6281                            I915_READ(ILK_DISPLAY_CHICKEN2) |
6282                            ILK_DPARB_GATE);
6283         }
6284
6285         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
6286
6287         I915_WRITE(ILK_DISPLAY_CHICKEN2,
6288                    I915_READ(ILK_DISPLAY_CHICKEN2) |
6289                    ILK_ELPIN_409_SELECT);
6290         I915_WRITE(_3D_CHICKEN2,
6291                    _3D_CHICKEN2_WM_READ_PIPELINED << 16 |
6292                    _3D_CHICKEN2_WM_READ_PIPELINED);
6293
6294         /* WaDisableRenderCachePipelinedFlush:ilk */
6295         I915_WRITE(CACHE_MODE_0,
6296                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
6297
6298         /* WaDisable_RenderCache_OperationalFlush:ilk */
6299         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6300
6301         g4x_disable_trickle_feed(dev);
6302
6303         ibx_init_clock_gating(dev);
6304 }
6305
6306 static void cpt_init_clock_gating(struct drm_device *dev)
6307 {
6308         struct drm_i915_private *dev_priv = dev->dev_private;
6309         int pipe;
6310         uint32_t val;
6311
6312         /*
6313          * On Ibex Peak and Cougar Point, we need to disable clock
6314          * gating for the panel power sequencer or it will fail to
6315          * start up when no ports are active.
6316          */
6317         I915_WRITE(SOUTH_DSPCLK_GATE_D, PCH_DPLSUNIT_CLOCK_GATE_DISABLE |
6318                    PCH_DPLUNIT_CLOCK_GATE_DISABLE |
6319                    PCH_CPUNIT_CLOCK_GATE_DISABLE);
6320         I915_WRITE(SOUTH_CHICKEN2, I915_READ(SOUTH_CHICKEN2) |
6321                    DPLS_EDP_PPS_FIX_DIS);
6322         /* The below fixes the weird display corruption, a few pixels shifted
6323          * downward, on (only) LVDS of some HP laptops with IVY.
6324          */
6325         for_each_pipe(dev_priv, pipe) {
6326                 val = I915_READ(TRANS_CHICKEN2(pipe));
6327                 val |= TRANS_CHICKEN2_TIMING_OVERRIDE;
6328                 val &= ~TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6329                 if (dev_priv->vbt.fdi_rx_polarity_inverted)
6330                         val |= TRANS_CHICKEN2_FDI_POLARITY_REVERSED;
6331                 val &= ~TRANS_CHICKEN2_FRAME_START_DELAY_MASK;
6332                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_COUNTER;
6333                 val &= ~TRANS_CHICKEN2_DISABLE_DEEP_COLOR_MODESWITCH;
6334                 I915_WRITE(TRANS_CHICKEN2(pipe), val);
6335         }
6336         /* WADP0ClockGatingDisable */
6337         for_each_pipe(dev_priv, pipe) {
6338                 I915_WRITE(TRANS_CHICKEN1(pipe),
6339                            TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
6340         }
6341 }
6342
6343 static void gen6_check_mch_setup(struct drm_device *dev)
6344 {
6345         struct drm_i915_private *dev_priv = dev->dev_private;
6346         uint32_t tmp;
6347
6348         tmp = I915_READ(MCH_SSKPD);
6349         if ((tmp & MCH_SSKPD_WM0_MASK) != MCH_SSKPD_WM0_VAL)
6350                 DRM_DEBUG_KMS("Wrong MCH_SSKPD value: 0x%08x This can cause underruns.\n",
6351                               tmp);
6352 }
6353
6354 static void gen6_init_clock_gating(struct drm_device *dev)
6355 {
6356         struct drm_i915_private *dev_priv = dev->dev_private;
6357         uint32_t dspclk_gate = ILK_VRHUNIT_CLOCK_GATE_DISABLE;
6358
6359         I915_WRITE(ILK_DSPCLK_GATE_D, dspclk_gate);
6360
6361         I915_WRITE(ILK_DISPLAY_CHICKEN2,
6362                    I915_READ(ILK_DISPLAY_CHICKEN2) |
6363                    ILK_ELPIN_409_SELECT);
6364
6365         /* WaDisableHiZPlanesWhenMSAAEnabled:snb */
6366         I915_WRITE(_3D_CHICKEN,
6367                    _MASKED_BIT_ENABLE(_3D_CHICKEN_HIZ_PLANE_DISABLE_MSAA_4X_SNB));
6368
6369         /* WaDisable_RenderCache_OperationalFlush:snb */
6370         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6371
6372         /*
6373          * BSpec recoomends 8x4 when MSAA is used,
6374          * however in practice 16x4 seems fastest.
6375          *
6376          * Note that PS/WM thread counts depend on the WIZ hashing
6377          * disable bit, which we don't touch here, but it's good
6378          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6379          */
6380         I915_WRITE(GEN6_GT_MODE,
6381                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6382
6383         ilk_init_lp_watermarks(dev);
6384
6385         I915_WRITE(CACHE_MODE_0,
6386                    _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
6387
6388         I915_WRITE(GEN6_UCGCTL1,
6389                    I915_READ(GEN6_UCGCTL1) |
6390                    GEN6_BLBUNIT_CLOCK_GATE_DISABLE |
6391                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
6392
6393         /* According to the BSpec vol1g, bit 12 (RCPBUNIT) clock
6394          * gating disable must be set.  Failure to set it results in
6395          * flickering pixels due to Z write ordering failures after
6396          * some amount of runtime in the Mesa "fire" demo, and Unigine
6397          * Sanctuary and Tropics, and apparently anything else with
6398          * alpha test or pixel discard.
6399          *
6400          * According to the spec, bit 11 (RCCUNIT) must also be set,
6401          * but we didn't debug actual testcases to find it out.
6402          *
6403          * WaDisableRCCUnitClockGating:snb
6404          * WaDisableRCPBUnitClockGating:snb
6405          */
6406         I915_WRITE(GEN6_UCGCTL2,
6407                    GEN6_RCPBUNIT_CLOCK_GATE_DISABLE |
6408                    GEN6_RCCUNIT_CLOCK_GATE_DISABLE);
6409
6410         /* WaStripsFansDisableFastClipPerformanceFix:snb */
6411         I915_WRITE(_3D_CHICKEN3,
6412                    _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_FASTCLIP_CULL));
6413
6414         /*
6415          * Bspec says:
6416          * "This bit must be set if 3DSTATE_CLIP clip mode is set to normal and
6417          * 3DSTATE_SF number of SF output attributes is more than 16."
6418          */
6419         I915_WRITE(_3D_CHICKEN3,
6420                    _MASKED_BIT_ENABLE(_3D_CHICKEN3_SF_DISABLE_PIPELINED_ATTR_FETCH));
6421
6422         /*
6423          * According to the spec the following bits should be
6424          * set in order to enable memory self-refresh and fbc:
6425          * The bit21 and bit22 of 0x42000
6426          * The bit21 and bit22 of 0x42004
6427          * The bit5 and bit7 of 0x42020
6428          * The bit14 of 0x70180
6429          * The bit14 of 0x71180
6430          *
6431          * WaFbcAsynchFlipDisableFbcQueue:snb
6432          */
6433         I915_WRITE(ILK_DISPLAY_CHICKEN1,
6434                    I915_READ(ILK_DISPLAY_CHICKEN1) |
6435                    ILK_FBCQ_DIS | ILK_PABSTRETCH_DIS);
6436         I915_WRITE(ILK_DISPLAY_CHICKEN2,
6437                    I915_READ(ILK_DISPLAY_CHICKEN2) |
6438                    ILK_DPARB_GATE | ILK_VSDPFD_FULL);
6439         I915_WRITE(ILK_DSPCLK_GATE_D,
6440                    I915_READ(ILK_DSPCLK_GATE_D) |
6441                    ILK_DPARBUNIT_CLOCK_GATE_ENABLE  |
6442                    ILK_DPFDUNIT_CLOCK_GATE_ENABLE);
6443
6444         g4x_disable_trickle_feed(dev);
6445
6446         cpt_init_clock_gating(dev);
6447
6448         gen6_check_mch_setup(dev);
6449 }
6450
6451 static void gen7_setup_fixed_func_scheduler(struct drm_i915_private *dev_priv)
6452 {
6453         uint32_t reg = I915_READ(GEN7_FF_THREAD_MODE);
6454
6455         /*
6456          * WaVSThreadDispatchOverride:ivb,vlv
6457          *
6458          * This actually overrides the dispatch
6459          * mode for all thread types.
6460          */
6461         reg &= ~GEN7_FF_SCHED_MASK;
6462         reg |= GEN7_FF_TS_SCHED_HW;
6463         reg |= GEN7_FF_VS_SCHED_HW;
6464         reg |= GEN7_FF_DS_SCHED_HW;
6465
6466         I915_WRITE(GEN7_FF_THREAD_MODE, reg);
6467 }
6468
6469 static void lpt_init_clock_gating(struct drm_device *dev)
6470 {
6471         struct drm_i915_private *dev_priv = dev->dev_private;
6472
6473         /*
6474          * TODO: this bit should only be enabled when really needed, then
6475          * disabled when not needed anymore in order to save power.
6476          */
6477         if (HAS_PCH_LPT_LP(dev))
6478                 I915_WRITE(SOUTH_DSPCLK_GATE_D,
6479                            I915_READ(SOUTH_DSPCLK_GATE_D) |
6480                            PCH_LP_PARTITION_LEVEL_DISABLE);
6481
6482         /* WADPOClockGatingDisable:hsw */
6483         I915_WRITE(TRANS_CHICKEN1(PIPE_A),
6484                    I915_READ(TRANS_CHICKEN1(PIPE_A)) |
6485                    TRANS_CHICKEN1_DP0UNIT_GC_DISABLE);
6486 }
6487
6488 static void lpt_suspend_hw(struct drm_device *dev)
6489 {
6490         struct drm_i915_private *dev_priv = dev->dev_private;
6491
6492         if (HAS_PCH_LPT_LP(dev)) {
6493                 uint32_t val = I915_READ(SOUTH_DSPCLK_GATE_D);
6494
6495                 val &= ~PCH_LP_PARTITION_LEVEL_DISABLE;
6496                 I915_WRITE(SOUTH_DSPCLK_GATE_D, val);
6497         }
6498 }
6499
6500 static void broadwell_init_clock_gating(struct drm_device *dev)
6501 {
6502         struct drm_i915_private *dev_priv = dev->dev_private;
6503         enum pipe pipe;
6504         uint32_t misccpctl;
6505
6506         ilk_init_lp_watermarks(dev);
6507
6508         /* WaSwitchSolVfFArbitrationPriority:bdw */
6509         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
6510
6511         /* WaPsrDPAMaskVBlankInSRD:bdw */
6512         I915_WRITE(CHICKEN_PAR1_1,
6513                    I915_READ(CHICKEN_PAR1_1) | DPA_MASK_VBLANK_SRD);
6514
6515         /* WaPsrDPRSUnmaskVBlankInSRD:bdw */
6516         for_each_pipe(dev_priv, pipe) {
6517                 I915_WRITE(CHICKEN_PIPESL_1(pipe),
6518                            I915_READ(CHICKEN_PIPESL_1(pipe)) |
6519                            BDW_DPRS_MASK_VBLANK_SRD);
6520         }
6521
6522         /* WaVSRefCountFullforceMissDisable:bdw */
6523         /* WaDSRefCountFullforceMissDisable:bdw */
6524         I915_WRITE(GEN7_FF_THREAD_MODE,
6525                    I915_READ(GEN7_FF_THREAD_MODE) &
6526                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
6527
6528         I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
6529                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
6530
6531         /* WaDisableSDEUnitClockGating:bdw */
6532         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
6533                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
6534
6535         /*
6536          * WaProgramL3SqcReg1Default:bdw
6537          * WaTempDisableDOPClkGating:bdw
6538          */
6539         misccpctl = I915_READ(GEN7_MISCCPCTL);
6540         I915_WRITE(GEN7_MISCCPCTL, misccpctl & ~GEN7_DOP_CLOCK_GATE_ENABLE);
6541         I915_WRITE(GEN8_L3SQCREG1, BDW_WA_L3SQCREG1_DEFAULT);
6542         I915_WRITE(GEN7_MISCCPCTL, misccpctl);
6543
6544         /*
6545          * WaGttCachingOffByDefault:bdw
6546          * GTT cache may not work with big pages, so if those
6547          * are ever enabled GTT cache may need to be disabled.
6548          */
6549         I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
6550
6551         lpt_init_clock_gating(dev);
6552 }
6553
6554 static void haswell_init_clock_gating(struct drm_device *dev)
6555 {
6556         struct drm_i915_private *dev_priv = dev->dev_private;
6557
6558         ilk_init_lp_watermarks(dev);
6559
6560         /* L3 caching of data atomics doesn't work -- disable it. */
6561         I915_WRITE(HSW_SCRATCH1, HSW_SCRATCH1_L3_DATA_ATOMICS_DISABLE);
6562         I915_WRITE(HSW_ROW_CHICKEN3,
6563                    _MASKED_BIT_ENABLE(HSW_ROW_CHICKEN3_L3_GLOBAL_ATOMICS_DISABLE));
6564
6565         /* This is required by WaCatErrorRejectionIssue:hsw */
6566         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
6567                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
6568                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
6569
6570         /* WaVSRefCountFullforceMissDisable:hsw */
6571         I915_WRITE(GEN7_FF_THREAD_MODE,
6572                    I915_READ(GEN7_FF_THREAD_MODE) & ~GEN7_FF_VS_REF_CNT_FFME);
6573
6574         /* WaDisable_RenderCache_OperationalFlush:hsw */
6575         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6576
6577         /* enable HiZ Raw Stall Optimization */
6578         I915_WRITE(CACHE_MODE_0_GEN7,
6579                    _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
6580
6581         /* WaDisable4x2SubspanOptimization:hsw */
6582         I915_WRITE(CACHE_MODE_1,
6583                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
6584
6585         /*
6586          * BSpec recommends 8x4 when MSAA is used,
6587          * however in practice 16x4 seems fastest.
6588          *
6589          * Note that PS/WM thread counts depend on the WIZ hashing
6590          * disable bit, which we don't touch here, but it's good
6591          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6592          */
6593         I915_WRITE(GEN7_GT_MODE,
6594                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6595
6596         /* WaSampleCChickenBitEnable:hsw */
6597         I915_WRITE(HALF_SLICE_CHICKEN3,
6598                    _MASKED_BIT_ENABLE(HSW_SAMPLE_C_PERFORMANCE));
6599
6600         /* WaSwitchSolVfFArbitrationPriority:hsw */
6601         I915_WRITE(GAM_ECOCHK, I915_READ(GAM_ECOCHK) | HSW_ECOCHK_ARB_PRIO_SOL);
6602
6603         /* WaRsPkgCStateDisplayPMReq:hsw */
6604         I915_WRITE(CHICKEN_PAR1_1,
6605                    I915_READ(CHICKEN_PAR1_1) | FORCE_ARB_IDLE_PLANES);
6606
6607         lpt_init_clock_gating(dev);
6608 }
6609
6610 static void ivybridge_init_clock_gating(struct drm_device *dev)
6611 {
6612         struct drm_i915_private *dev_priv = dev->dev_private;
6613         uint32_t snpcr;
6614
6615         ilk_init_lp_watermarks(dev);
6616
6617         I915_WRITE(ILK_DSPCLK_GATE_D, ILK_VRHUNIT_CLOCK_GATE_DISABLE);
6618
6619         /* WaDisableEarlyCull:ivb */
6620         I915_WRITE(_3D_CHICKEN3,
6621                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
6622
6623         /* WaDisableBackToBackFlipFix:ivb */
6624         I915_WRITE(IVB_CHICKEN3,
6625                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
6626                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
6627
6628         /* WaDisablePSDDualDispatchEnable:ivb */
6629         if (IS_IVB_GT1(dev))
6630                 I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
6631                            _MASKED_BIT_ENABLE(GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
6632
6633         /* WaDisable_RenderCache_OperationalFlush:ivb */
6634         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6635
6636         /* Apply the WaDisableRHWOOptimizationForRenderHang:ivb workaround. */
6637         I915_WRITE(GEN7_COMMON_SLICE_CHICKEN1,
6638                    GEN7_CSC1_RHWO_OPT_DISABLE_IN_RCC);
6639
6640         /* WaApplyL3ControlAndL3ChickenMode:ivb */
6641         I915_WRITE(GEN7_L3CNTLREG1,
6642                         GEN7_WA_FOR_GEN7_L3_CONTROL);
6643         I915_WRITE(GEN7_L3_CHICKEN_MODE_REGISTER,
6644                    GEN7_WA_L3_CHICKEN_MODE);
6645         if (IS_IVB_GT1(dev))
6646                 I915_WRITE(GEN7_ROW_CHICKEN2,
6647                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6648         else {
6649                 /* must write both registers */
6650                 I915_WRITE(GEN7_ROW_CHICKEN2,
6651                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6652                 I915_WRITE(GEN7_ROW_CHICKEN2_GT2,
6653                            _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6654         }
6655
6656         /* WaForceL3Serialization:ivb */
6657         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
6658                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
6659
6660         /*
6661          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6662          * This implements the WaDisableRCZUnitClockGating:ivb workaround.
6663          */
6664         I915_WRITE(GEN6_UCGCTL2,
6665                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
6666
6667         /* This is required by WaCatErrorRejectionIssue:ivb */
6668         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
6669                         I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
6670                         GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
6671
6672         g4x_disable_trickle_feed(dev);
6673
6674         gen7_setup_fixed_func_scheduler(dev_priv);
6675
6676         if (0) { /* causes HiZ corruption on ivb:gt1 */
6677                 /* enable HiZ Raw Stall Optimization */
6678                 I915_WRITE(CACHE_MODE_0_GEN7,
6679                            _MASKED_BIT_DISABLE(HIZ_RAW_STALL_OPT_DISABLE));
6680         }
6681
6682         /* WaDisable4x2SubspanOptimization:ivb */
6683         I915_WRITE(CACHE_MODE_1,
6684                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
6685
6686         /*
6687          * BSpec recommends 8x4 when MSAA is used,
6688          * however in practice 16x4 seems fastest.
6689          *
6690          * Note that PS/WM thread counts depend on the WIZ hashing
6691          * disable bit, which we don't touch here, but it's good
6692          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6693          */
6694         I915_WRITE(GEN7_GT_MODE,
6695                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6696
6697         snpcr = I915_READ(GEN6_MBCUNIT_SNPCR);
6698         snpcr &= ~GEN6_MBC_SNPCR_MASK;
6699         snpcr |= GEN6_MBC_SNPCR_MED;
6700         I915_WRITE(GEN6_MBCUNIT_SNPCR, snpcr);
6701
6702         if (!HAS_PCH_NOP(dev))
6703                 cpt_init_clock_gating(dev);
6704
6705         gen6_check_mch_setup(dev);
6706 }
6707
6708 static void vlv_init_display_clock_gating(struct drm_i915_private *dev_priv)
6709 {
6710         I915_WRITE(DSPCLK_GATE_D, VRHUNIT_CLOCK_GATE_DISABLE);
6711
6712         /*
6713          * Disable trickle feed and enable pnd deadline calculation
6714          */
6715         I915_WRITE(MI_ARB_VLV, MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE);
6716         I915_WRITE(CBR1_VLV, 0);
6717 }
6718
6719 static void valleyview_init_clock_gating(struct drm_device *dev)
6720 {
6721         struct drm_i915_private *dev_priv = dev->dev_private;
6722
6723         vlv_init_display_clock_gating(dev_priv);
6724
6725         /* WaDisableEarlyCull:vlv */
6726         I915_WRITE(_3D_CHICKEN3,
6727                    _MASKED_BIT_ENABLE(_3D_CHICKEN_SF_DISABLE_OBJEND_CULL));
6728
6729         /* WaDisableBackToBackFlipFix:vlv */
6730         I915_WRITE(IVB_CHICKEN3,
6731                    CHICKEN3_DGMG_REQ_OUT_FIX_DISABLE |
6732                    CHICKEN3_DGMG_DONE_FIX_DISABLE);
6733
6734         /* WaPsdDispatchEnable:vlv */
6735         /* WaDisablePSDDualDispatchEnable:vlv */
6736         I915_WRITE(GEN7_HALF_SLICE_CHICKEN1,
6737                    _MASKED_BIT_ENABLE(GEN7_MAX_PS_THREAD_DEP |
6738                                       GEN7_PSD_SINGLE_PORT_DISPATCH_ENABLE));
6739
6740         /* WaDisable_RenderCache_OperationalFlush:vlv */
6741         I915_WRITE(CACHE_MODE_0_GEN7, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6742
6743         /* WaForceL3Serialization:vlv */
6744         I915_WRITE(GEN7_L3SQCREG4, I915_READ(GEN7_L3SQCREG4) &
6745                    ~L3SQ_URB_READ_CAM_MATCH_DISABLE);
6746
6747         /* WaDisableDopClockGating:vlv */
6748         I915_WRITE(GEN7_ROW_CHICKEN2,
6749                    _MASKED_BIT_ENABLE(DOP_CLOCK_GATING_DISABLE));
6750
6751         /* This is required by WaCatErrorRejectionIssue:vlv */
6752         I915_WRITE(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG,
6753                    I915_READ(GEN7_SQ_CHICKEN_MBCUNIT_CONFIG) |
6754                    GEN7_SQ_CHICKEN_MBCUNIT_SQINTMOB);
6755
6756         gen7_setup_fixed_func_scheduler(dev_priv);
6757
6758         /*
6759          * According to the spec, bit 13 (RCZUNIT) must be set on IVB.
6760          * This implements the WaDisableRCZUnitClockGating:vlv workaround.
6761          */
6762         I915_WRITE(GEN6_UCGCTL2,
6763                    GEN6_RCZUNIT_CLOCK_GATE_DISABLE);
6764
6765         /* WaDisableL3Bank2xClockGate:vlv
6766          * Disabling L3 clock gating- MMIO 940c[25] = 1
6767          * Set bit 25, to disable L3_BANK_2x_CLK_GATING */
6768         I915_WRITE(GEN7_UCGCTL4,
6769                    I915_READ(GEN7_UCGCTL4) | GEN7_L3BANK2X_CLOCK_GATE_DISABLE);
6770
6771         /*
6772          * BSpec says this must be set, even though
6773          * WaDisable4x2SubspanOptimization isn't listed for VLV.
6774          */
6775         I915_WRITE(CACHE_MODE_1,
6776                    _MASKED_BIT_ENABLE(PIXEL_SUBSPAN_COLLECT_OPT_DISABLE));
6777
6778         /*
6779          * BSpec recommends 8x4 when MSAA is used,
6780          * however in practice 16x4 seems fastest.
6781          *
6782          * Note that PS/WM thread counts depend on the WIZ hashing
6783          * disable bit, which we don't touch here, but it's good
6784          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
6785          */
6786         I915_WRITE(GEN7_GT_MODE,
6787                    _MASKED_FIELD(GEN6_WIZ_HASHING_MASK, GEN6_WIZ_HASHING_16x4));
6788
6789         /*
6790          * WaIncreaseL3CreditsForVLVB0:vlv
6791          * This is the hardware default actually.
6792          */
6793         I915_WRITE(GEN7_L3SQCREG1, VLV_B0_WA_L3SQCREG1_VALUE);
6794
6795         /*
6796          * WaDisableVLVClockGating_VBIIssue:vlv
6797          * Disable clock gating on th GCFG unit to prevent a delay
6798          * in the reporting of vblank events.
6799          */
6800         I915_WRITE(VLV_GUNIT_CLOCK_GATE, GCFG_DIS);
6801 }
6802
6803 static void cherryview_init_clock_gating(struct drm_device *dev)
6804 {
6805         struct drm_i915_private *dev_priv = dev->dev_private;
6806
6807         vlv_init_display_clock_gating(dev_priv);
6808
6809         /* WaVSRefCountFullforceMissDisable:chv */
6810         /* WaDSRefCountFullforceMissDisable:chv */
6811         I915_WRITE(GEN7_FF_THREAD_MODE,
6812                    I915_READ(GEN7_FF_THREAD_MODE) &
6813                    ~(GEN8_FF_DS_REF_CNT_FFME | GEN7_FF_VS_REF_CNT_FFME));
6814
6815         /* WaDisableSemaphoreAndSyncFlipWait:chv */
6816         I915_WRITE(GEN6_RC_SLEEP_PSMI_CONTROL,
6817                    _MASKED_BIT_ENABLE(GEN8_RC_SEMA_IDLE_MSG_DISABLE));
6818
6819         /* WaDisableCSUnitClockGating:chv */
6820         I915_WRITE(GEN6_UCGCTL1, I915_READ(GEN6_UCGCTL1) |
6821                    GEN6_CSUNIT_CLOCK_GATE_DISABLE);
6822
6823         /* WaDisableSDEUnitClockGating:chv */
6824         I915_WRITE(GEN8_UCGCTL6, I915_READ(GEN8_UCGCTL6) |
6825                    GEN8_SDEUNIT_CLOCK_GATE_DISABLE);
6826
6827         /*
6828          * GTT cache may not work with big pages, so if those
6829          * are ever enabled GTT cache may need to be disabled.
6830          */
6831         I915_WRITE(HSW_GTT_CACHE_EN, GTT_CACHE_EN_ALL);
6832 }
6833
6834 static void g4x_init_clock_gating(struct drm_device *dev)
6835 {
6836         struct drm_i915_private *dev_priv = dev->dev_private;
6837         uint32_t dspclk_gate;
6838
6839         I915_WRITE(RENCLK_GATE_D1, 0);
6840         I915_WRITE(RENCLK_GATE_D2, VF_UNIT_CLOCK_GATE_DISABLE |
6841                    GS_UNIT_CLOCK_GATE_DISABLE |
6842                    CL_UNIT_CLOCK_GATE_DISABLE);
6843         I915_WRITE(RAMCLK_GATE_D, 0);
6844         dspclk_gate = VRHUNIT_CLOCK_GATE_DISABLE |
6845                 OVRUNIT_CLOCK_GATE_DISABLE |
6846                 OVCUNIT_CLOCK_GATE_DISABLE;
6847         if (IS_GM45(dev))
6848                 dspclk_gate |= DSSUNIT_CLOCK_GATE_DISABLE;
6849         I915_WRITE(DSPCLK_GATE_D, dspclk_gate);
6850
6851         /* WaDisableRenderCachePipelinedFlush */
6852         I915_WRITE(CACHE_MODE_0,
6853                    _MASKED_BIT_ENABLE(CM0_PIPELINED_RENDER_FLUSH_DISABLE));
6854
6855         /* WaDisable_RenderCache_OperationalFlush:g4x */
6856         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6857
6858         g4x_disable_trickle_feed(dev);
6859 }
6860
6861 static void crestline_init_clock_gating(struct drm_device *dev)
6862 {
6863         struct drm_i915_private *dev_priv = dev->dev_private;
6864
6865         I915_WRITE(RENCLK_GATE_D1, I965_RCC_CLOCK_GATE_DISABLE);
6866         I915_WRITE(RENCLK_GATE_D2, 0);
6867         I915_WRITE(DSPCLK_GATE_D, 0);
6868         I915_WRITE(RAMCLK_GATE_D, 0);
6869         I915_WRITE16(DEUC, 0);
6870         I915_WRITE(MI_ARB_STATE,
6871                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
6872
6873         /* WaDisable_RenderCache_OperationalFlush:gen4 */
6874         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6875 }
6876
6877 static void broadwater_init_clock_gating(struct drm_device *dev)
6878 {
6879         struct drm_i915_private *dev_priv = dev->dev_private;
6880
6881         I915_WRITE(RENCLK_GATE_D1, I965_RCZ_CLOCK_GATE_DISABLE |
6882                    I965_RCC_CLOCK_GATE_DISABLE |
6883                    I965_RCPB_CLOCK_GATE_DISABLE |
6884                    I965_ISC_CLOCK_GATE_DISABLE |
6885                    I965_FBC_CLOCK_GATE_DISABLE);
6886         I915_WRITE(RENCLK_GATE_D2, 0);
6887         I915_WRITE(MI_ARB_STATE,
6888                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
6889
6890         /* WaDisable_RenderCache_OperationalFlush:gen4 */
6891         I915_WRITE(CACHE_MODE_0, _MASKED_BIT_DISABLE(RC_OP_FLUSH_ENABLE));
6892 }
6893
6894 static void gen3_init_clock_gating(struct drm_device *dev)
6895 {
6896         struct drm_i915_private *dev_priv = dev->dev_private;
6897         u32 dstate = I915_READ(D_STATE);
6898
6899         dstate |= DSTATE_PLL_D3_OFF | DSTATE_GFX_CLOCK_GATING |
6900                 DSTATE_DOT_CLOCK_GATING;
6901         I915_WRITE(D_STATE, dstate);
6902
6903         if (IS_PINEVIEW(dev))
6904                 I915_WRITE(ECOSKPD, _MASKED_BIT_ENABLE(ECO_GATING_CX_ONLY));
6905
6906         /* IIR "flip pending" means done if this bit is set */
6907         I915_WRITE(ECOSKPD, _MASKED_BIT_DISABLE(ECO_FLIP_DONE));
6908
6909         /* interrupts should cause a wake up from C3 */
6910         I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_AGPBUSY_INT_EN));
6911
6912         /* On GEN3 we really need to make sure the ARB C3 LP bit is set */
6913         I915_WRITE(MI_ARB_STATE, _MASKED_BIT_ENABLE(MI_ARB_C3_LP_WRITE_ENABLE));
6914
6915         I915_WRITE(MI_ARB_STATE,
6916                    _MASKED_BIT_ENABLE(MI_ARB_DISPLAY_TRICKLE_FEED_DISABLE));
6917 }
6918
6919 static void i85x_init_clock_gating(struct drm_device *dev)
6920 {
6921         struct drm_i915_private *dev_priv = dev->dev_private;
6922
6923         I915_WRITE(RENCLK_GATE_D1, SV_CLOCK_GATE_DISABLE);
6924
6925         /* interrupts should cause a wake up from C3 */
6926         I915_WRITE(MI_STATE, _MASKED_BIT_ENABLE(MI_AGPBUSY_INT_EN) |
6927                    _MASKED_BIT_DISABLE(MI_AGPBUSY_830_MODE));
6928
6929         I915_WRITE(MEM_MODE,
6930                    _MASKED_BIT_ENABLE(MEM_DISPLAY_TRICKLE_FEED_DISABLE));
6931 }
6932
6933 static void i830_init_clock_gating(struct drm_device *dev)
6934 {
6935         struct drm_i915_private *dev_priv = dev->dev_private;
6936
6937         I915_WRITE(DSPCLK_GATE_D, OVRUNIT_CLOCK_GATE_DISABLE);
6938
6939         I915_WRITE(MEM_MODE,
6940                    _MASKED_BIT_ENABLE(MEM_DISPLAY_A_TRICKLE_FEED_DISABLE) |
6941                    _MASKED_BIT_ENABLE(MEM_DISPLAY_B_TRICKLE_FEED_DISABLE));
6942 }
6943
6944 void intel_init_clock_gating(struct drm_device *dev)
6945 {
6946         struct drm_i915_private *dev_priv = dev->dev_private;
6947
6948         if (dev_priv->display.init_clock_gating)
6949                 dev_priv->display.init_clock_gating(dev);
6950 }
6951
6952 void intel_suspend_hw(struct drm_device *dev)
6953 {
6954         if (HAS_PCH_LPT(dev))
6955                 lpt_suspend_hw(dev);
6956 }
6957
6958 /* Set up chip specific power management-related functions */
6959 void intel_init_pm(struct drm_device *dev)
6960 {
6961         struct drm_i915_private *dev_priv = dev->dev_private;
6962
6963         intel_fbc_init(dev_priv);
6964
6965         /* For cxsr */
6966         if (IS_PINEVIEW(dev))
6967                 i915_pineview_get_mem_freq(dev);
6968         else if (IS_GEN5(dev))
6969                 i915_ironlake_get_mem_freq(dev);
6970
6971         /* For FIFO watermark updates */
6972         if (INTEL_INFO(dev)->gen >= 9) {
6973                 skl_setup_wm_latency(dev);
6974
6975                 if (IS_BROXTON(dev))
6976                         dev_priv->display.init_clock_gating =
6977                                 bxt_init_clock_gating;
6978                 dev_priv->display.update_wm = skl_update_wm;
6979         } else if (HAS_PCH_SPLIT(dev)) {
6980                 ilk_setup_wm_latency(dev);
6981
6982                 if ((IS_GEN5(dev) && dev_priv->wm.pri_latency[1] &&
6983                      dev_priv->wm.spr_latency[1] && dev_priv->wm.cur_latency[1]) ||
6984                     (!IS_GEN5(dev) && dev_priv->wm.pri_latency[0] &&
6985                      dev_priv->wm.spr_latency[0] && dev_priv->wm.cur_latency[0])) {
6986                         dev_priv->display.update_wm = ilk_update_wm;
6987                         dev_priv->display.compute_pipe_wm = ilk_compute_pipe_wm;
6988                 } else {
6989                         DRM_DEBUG_KMS("Failed to read display plane latency. "
6990                                       "Disable CxSR\n");
6991                 }
6992
6993                 if (IS_GEN5(dev))
6994                         dev_priv->display.init_clock_gating = ironlake_init_clock_gating;
6995                 else if (IS_GEN6(dev))
6996                         dev_priv->display.init_clock_gating = gen6_init_clock_gating;
6997                 else if (IS_IVYBRIDGE(dev))
6998                         dev_priv->display.init_clock_gating = ivybridge_init_clock_gating;
6999                 else if (IS_HASWELL(dev))
7000                         dev_priv->display.init_clock_gating = haswell_init_clock_gating;
7001                 else if (INTEL_INFO(dev)->gen == 8)
7002                         dev_priv->display.init_clock_gating = broadwell_init_clock_gating;
7003         } else if (IS_CHERRYVIEW(dev)) {
7004                 vlv_setup_wm_latency(dev);
7005
7006                 dev_priv->display.update_wm = vlv_update_wm;
7007                 dev_priv->display.init_clock_gating =
7008                         cherryview_init_clock_gating;
7009         } else if (IS_VALLEYVIEW(dev)) {
7010                 vlv_setup_wm_latency(dev);
7011
7012                 dev_priv->display.update_wm = vlv_update_wm;
7013                 dev_priv->display.init_clock_gating =
7014                         valleyview_init_clock_gating;
7015         } else if (IS_PINEVIEW(dev)) {
7016                 if (!intel_get_cxsr_latency(IS_PINEVIEW_G(dev),
7017                                             dev_priv->is_ddr3,
7018                                             dev_priv->fsb_freq,
7019                                             dev_priv->mem_freq)) {
7020                         DRM_INFO("failed to find known CxSR latency "
7021                                  "(found ddr%s fsb freq %d, mem freq %d), "
7022                                  "disabling CxSR\n",
7023                                  (dev_priv->is_ddr3 == 1) ? "3" : "2",
7024                                  dev_priv->fsb_freq, dev_priv->mem_freq);
7025                         /* Disable CxSR and never update its watermark again */
7026                         intel_set_memory_cxsr(dev_priv, false);
7027                         dev_priv->display.update_wm = NULL;
7028                 } else
7029                         dev_priv->display.update_wm = pineview_update_wm;
7030                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
7031         } else if (IS_G4X(dev)) {
7032                 dev_priv->display.update_wm = g4x_update_wm;
7033                 dev_priv->display.init_clock_gating = g4x_init_clock_gating;
7034         } else if (IS_GEN4(dev)) {
7035                 dev_priv->display.update_wm = i965_update_wm;
7036                 if (IS_CRESTLINE(dev))
7037                         dev_priv->display.init_clock_gating = crestline_init_clock_gating;
7038                 else if (IS_BROADWATER(dev))
7039                         dev_priv->display.init_clock_gating = broadwater_init_clock_gating;
7040         } else if (IS_GEN3(dev)) {
7041                 dev_priv->display.update_wm = i9xx_update_wm;
7042                 dev_priv->display.get_fifo_size = i9xx_get_fifo_size;
7043                 dev_priv->display.init_clock_gating = gen3_init_clock_gating;
7044         } else if (IS_GEN2(dev)) {
7045                 if (INTEL_INFO(dev)->num_pipes == 1) {
7046                         dev_priv->display.update_wm = i845_update_wm;
7047                         dev_priv->display.get_fifo_size = i845_get_fifo_size;
7048                 } else {
7049                         dev_priv->display.update_wm = i9xx_update_wm;
7050                         dev_priv->display.get_fifo_size = i830_get_fifo_size;
7051                 }
7052
7053                 if (IS_I85X(dev) || IS_I865G(dev))
7054                         dev_priv->display.init_clock_gating = i85x_init_clock_gating;
7055                 else
7056                         dev_priv->display.init_clock_gating = i830_init_clock_gating;
7057         } else {
7058                 DRM_ERROR("unexpected fall-through in intel_init_pm\n");
7059         }
7060 }
7061
7062 int sandybridge_pcode_read(struct drm_i915_private *dev_priv, u32 mbox, u32 *val)
7063 {
7064         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7065
7066         if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
7067                 DRM_DEBUG_DRIVER("warning: pcode (read) mailbox access failed\n");
7068                 return -EAGAIN;
7069         }
7070
7071         I915_WRITE(GEN6_PCODE_DATA, *val);
7072         I915_WRITE(GEN6_PCODE_DATA1, 0);
7073         I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
7074
7075         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
7076                      500)) {
7077                 DRM_ERROR("timeout waiting for pcode read (%d) to finish\n", mbox);
7078                 return -ETIMEDOUT;
7079         }
7080
7081         *val = I915_READ(GEN6_PCODE_DATA);
7082         I915_WRITE(GEN6_PCODE_DATA, 0);
7083
7084         return 0;
7085 }
7086
7087 int sandybridge_pcode_write(struct drm_i915_private *dev_priv, u32 mbox, u32 val)
7088 {
7089         WARN_ON(!mutex_is_locked(&dev_priv->rps.hw_lock));
7090
7091         if (I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) {
7092                 DRM_DEBUG_DRIVER("warning: pcode (write) mailbox access failed\n");
7093                 return -EAGAIN;
7094         }
7095
7096         I915_WRITE(GEN6_PCODE_DATA, val);
7097         I915_WRITE(GEN6_PCODE_MAILBOX, GEN6_PCODE_READY | mbox);
7098
7099         if (wait_for((I915_READ(GEN6_PCODE_MAILBOX) & GEN6_PCODE_READY) == 0,
7100                      500)) {
7101                 DRM_ERROR("timeout waiting for pcode write (%d) to finish\n", mbox);
7102                 return -ETIMEDOUT;
7103         }
7104
7105         I915_WRITE(GEN6_PCODE_DATA, 0);
7106
7107         return 0;
7108 }
7109
7110 static int vlv_gpu_freq_div(unsigned int czclk_freq)
7111 {
7112         switch (czclk_freq) {
7113         case 200:
7114                 return 10;
7115         case 267:
7116                 return 12;
7117         case 320:
7118         case 333:
7119                 return 16;
7120         case 400:
7121                 return 20;
7122         default:
7123                 return -1;
7124         }
7125 }
7126
7127 static int byt_gpu_freq(struct drm_i915_private *dev_priv, int val)
7128 {
7129         int div, czclk_freq = DIV_ROUND_CLOSEST(dev_priv->czclk_freq, 1000);
7130
7131         div = vlv_gpu_freq_div(czclk_freq);
7132         if (div < 0)
7133                 return div;
7134
7135         return DIV_ROUND_CLOSEST(czclk_freq * (val + 6 - 0xbd), div);
7136 }
7137
7138 static int byt_freq_opcode(struct drm_i915_private *dev_priv, int val)
7139 {
7140         int mul, czclk_freq = DIV_ROUND_CLOSEST(dev_priv->czclk_freq, 1000);
7141
7142         mul = vlv_gpu_freq_div(czclk_freq);
7143         if (mul < 0)
7144                 return mul;
7145
7146         return DIV_ROUND_CLOSEST(mul * val, czclk_freq) + 0xbd - 6;
7147 }
7148
7149 static int chv_gpu_freq(struct drm_i915_private *dev_priv, int val)
7150 {
7151         int div, czclk_freq = DIV_ROUND_CLOSEST(dev_priv->czclk_freq, 1000);
7152
7153         div = vlv_gpu_freq_div(czclk_freq) / 2;
7154         if (div < 0)
7155                 return div;
7156
7157         return DIV_ROUND_CLOSEST(czclk_freq * val, 2 * div) / 2;
7158 }
7159
7160 static int chv_freq_opcode(struct drm_i915_private *dev_priv, int val)
7161 {
7162         int mul, czclk_freq = DIV_ROUND_CLOSEST(dev_priv->czclk_freq, 1000);
7163
7164         mul = vlv_gpu_freq_div(czclk_freq) / 2;
7165         if (mul < 0)
7166                 return mul;
7167
7168         /* CHV needs even values */
7169         return DIV_ROUND_CLOSEST(val * 2 * mul, czclk_freq) * 2;
7170 }
7171
7172 int intel_gpu_freq(struct drm_i915_private *dev_priv, int val)
7173 {
7174         if (IS_GEN9(dev_priv->dev))
7175                 return DIV_ROUND_CLOSEST(val * GT_FREQUENCY_MULTIPLIER,
7176                                          GEN9_FREQ_SCALER);
7177         else if (IS_CHERRYVIEW(dev_priv->dev))
7178                 return chv_gpu_freq(dev_priv, val);
7179         else if (IS_VALLEYVIEW(dev_priv->dev))
7180                 return byt_gpu_freq(dev_priv, val);
7181         else
7182                 return val * GT_FREQUENCY_MULTIPLIER;
7183 }
7184
7185 int intel_freq_opcode(struct drm_i915_private *dev_priv, int val)
7186 {
7187         if (IS_GEN9(dev_priv->dev))
7188                 return DIV_ROUND_CLOSEST(val * GEN9_FREQ_SCALER,
7189                                          GT_FREQUENCY_MULTIPLIER);
7190         else if (IS_CHERRYVIEW(dev_priv->dev))
7191                 return chv_freq_opcode(dev_priv, val);
7192         else if (IS_VALLEYVIEW(dev_priv->dev))
7193                 return byt_freq_opcode(dev_priv, val);
7194         else
7195                 return DIV_ROUND_CLOSEST(val, GT_FREQUENCY_MULTIPLIER);
7196 }
7197
7198 struct request_boost {
7199         struct work_struct work;
7200         struct drm_i915_gem_request *req;
7201 };
7202
7203 static void __intel_rps_boost_work(struct work_struct *work)
7204 {
7205         struct request_boost *boost = container_of(work, struct request_boost, work);
7206         struct drm_i915_gem_request *req = boost->req;
7207
7208         if (!i915_gem_request_completed(req, true))
7209                 gen6_rps_boost(to_i915(req->ring->dev), NULL,
7210                                req->emitted_jiffies);
7211
7212         i915_gem_request_unreference__unlocked(req);
7213         kfree(boost);
7214 }
7215
7216 void intel_queue_rps_boost_for_request(struct drm_device *dev,
7217                                        struct drm_i915_gem_request *req)
7218 {
7219         struct request_boost *boost;
7220
7221         if (req == NULL || INTEL_INFO(dev)->gen < 6)
7222                 return;
7223
7224         if (i915_gem_request_completed(req, true))
7225                 return;
7226
7227         boost = kmalloc(sizeof(*boost), GFP_ATOMIC);
7228         if (boost == NULL)
7229                 return;
7230
7231         i915_gem_request_reference(req);
7232         boost->req = req;
7233
7234         INIT_WORK(&boost->work, __intel_rps_boost_work);
7235         queue_work(to_i915(dev)->wq, &boost->work);
7236 }
7237
7238 void intel_pm_setup(struct drm_device *dev)
7239 {
7240         struct drm_i915_private *dev_priv = dev->dev_private;
7241
7242         mutex_init(&dev_priv->rps.hw_lock);
7243         spin_lock_init(&dev_priv->rps.client_lock);
7244
7245         INIT_DELAYED_WORK(&dev_priv->rps.delayed_resume_work,
7246                           intel_gen6_powersave_work);
7247         INIT_LIST_HEAD(&dev_priv->rps.clients);
7248         INIT_LIST_HEAD(&dev_priv->rps.semaphores.link);
7249         INIT_LIST_HEAD(&dev_priv->rps.mmioflips.link);
7250
7251         dev_priv->pm.suspended = false;
7252         atomic_set(&dev_priv->pm.wakeref_count, 0);
7253         atomic_set(&dev_priv->pm.atomic_seq, 0);
7254 }