drivers/gpu/drm/i915/intel_ringbuffer.c

   1 /*
   2  * Copyright © 2008-2010 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  *    Eric Anholt <eric@anholt.net>
  25  *    Zou Nan hai <nanhai.zou@intel.com>
  26  *    Xiang Hai hao<haihao.xiang@intel.com>
  27  *
  28  */
  29
  30 #include <drm/drmP.h>
  31 #include "i915_drv.h"
  32 #include <drm/i915_drm.h>
  33 #include "i915_trace.h"
  34 #include "intel_drv.h"
  35
  36 /* Early gen2 devices have a cacheline of just 32 bytes, using 64 is overkill,
  37  * but keeps the logic simple. Indeed, the whole purpose of this macro is just
  38  * to give some inclination as to some of the magic values used in the various
  39  * workarounds!
  40  */
  41 #define CACHELINE_BYTES 64
  42
  43 static inline int __ring_space(int head, int tail, int size)
  44 {
  45         int space = head - (tail + I915_RING_FREE_SPACE);
  46         if (space < 0)
  47                 space += size;
  48         return space;
  49 }
  50
  51 static inline int ring_space(struct intel_ringbuffer *ringbuf)
  52 {
  53         return __ring_space(ringbuf->head & HEAD_ADDR, ringbuf->tail, ringbuf->size);
  54 }
  55
  56 static bool intel_ring_stopped(struct intel_engine_cs *ring)
  57 {
  58         struct drm_i915_private *dev_priv = ring->dev->dev_private;
  59         return dev_priv->gpu_error.stop_rings & intel_ring_flag(ring);
  60 }
  61
  62 void __intel_ring_advance(struct intel_engine_cs *ring)
  63 {
  64         struct intel_ringbuffer *ringbuf = ring->buffer;
  65         ringbuf->tail &= ringbuf->size - 1;
  66         if (intel_ring_stopped(ring))
  67                 return;
  68         ring->write_tail(ring, ringbuf->tail);
  69 }
  70
  71 static int
  72 gen2_render_ring_flush(struct intel_engine_cs *ring,
  73                        u32      invalidate_domains,
  74                        u32      flush_domains)
  75 {
  76         u32 cmd;
  77         int ret;
  78
  79         cmd = MI_FLUSH;
  80         if (((invalidate_domains|flush_domains) & I915_GEM_DOMAIN_RENDER) == 0)
  81                 cmd |= MI_NO_WRITE_FLUSH;
  82
  83         if (invalidate_domains & I915_GEM_DOMAIN_SAMPLER)
  84                 cmd |= MI_READ_FLUSH;
  85
  86         ret = intel_ring_begin(ring, 2);
  87         if (ret)
  88                 return ret;
  89
  90         intel_ring_emit(ring, cmd);
  91         intel_ring_emit(ring, MI_NOOP);
  92         intel_ring_advance(ring);
  93
  94         return 0;
  95 }
  96
  97 static int
  98 gen4_render_ring_flush(struct intel_engine_cs *ring,
  99                        u32      invalidate_domains,
 100                        u32      flush_domains)
 101 {
 102         struct drm_device *dev = ring->dev;
 103         u32 cmd;
 104         int ret;
 105
 106         /*
 107          * read/write caches:
 108          *
 109          * I915_GEM_DOMAIN_RENDER is always invalidated, but is
 110          * only flushed if MI_NO_WRITE_FLUSH is unset.  On 965, it is
 111          * also flushed at 2d versus 3d pipeline switches.
 112          *
 113          * read-only caches:
 114          *
 115          * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
 116          * MI_READ_FLUSH is set, and is always flushed on 965.
 117          *
 118          * I915_GEM_DOMAIN_COMMAND may not exist?
 119          *
 120          * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
 121          * invalidated when MI_EXE_FLUSH is set.
 122          *
 123          * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
 124          * invalidated with every MI_FLUSH.
 125          *
 126          * TLBs:
 127          *
 128          * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
 129          * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
 130          * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
 131          * are flushed at any MI_FLUSH.
 132          */
 133
 134         cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
 135         if ((invalidate_domains|flush_domains) & I915_GEM_DOMAIN_RENDER)
 136                 cmd &= ~MI_NO_WRITE_FLUSH;
 137         if (invalidate_domains & I915_GEM_DOMAIN_INSTRUCTION)
 138                 cmd |= MI_EXE_FLUSH;
 139
 140         if (invalidate_domains & I915_GEM_DOMAIN_COMMAND &&
 141             (IS_G4X(dev) || IS_GEN5(dev)))
 142                 cmd |= MI_INVALIDATE_ISP;
 143
 144         ret = intel_ring_begin(ring, 2);
 145         if (ret)
 146                 return ret;
 147
 148         intel_ring_emit(ring, cmd);
 149         intel_ring_emit(ring, MI_NOOP);
 150         intel_ring_advance(ring);
 151
 152         return 0;
 153 }
 154
 155 /**
 156  * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
 157  * implementing two workarounds on gen6.  From section 1.4.7.1
 158  * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
 159  *
 160  * [DevSNB-C+{W/A}] Before any depth stall flush (including those
 161  * produced by non-pipelined state commands), software needs to first
 162  * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
 163  * 0.
 164  *
 165  * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
 166  * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
 167  *
 168  * And the workaround for these two requires this workaround first:
 169  *
 170  * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
 171  * BEFORE the pipe-control with a post-sync op and no write-cache
 172  * flushes.
 173  *
 174  * And this last workaround is tricky because of the requirements on
 175  * that bit.  From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
 176  * volume 2 part 1:
 177  *
 178  *     "1 of the following must also be set:
 179  *      - Render Target Cache Flush Enable ([12] of DW1)
 180  *      - Depth Cache Flush Enable ([0] of DW1)
 181  *      - Stall at Pixel Scoreboard ([1] of DW1)
 182  *      - Depth Stall ([13] of DW1)
 183  *      - Post-Sync Operation ([13] of DW1)
 184  *      - Notify Enable ([8] of DW1)"
 185  *
 186  * The cache flushes require the workaround flush that triggered this
 187  * one, so we can't use it.  Depth stall would trigger the same.
 188  * Post-sync nonzero is what triggered this second workaround, so we
 189  * can't use that one either.  Notify enable is IRQs, which aren't
 190  * really our business.  That leaves only stall at scoreboard.
 191  */
 192 static int
 193 intel_emit_post_sync_nonzero_flush(struct intel_engine_cs *ring)
 194 {
 195         u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES;
 196         int ret;
 197
 198
 199         ret = intel_ring_begin(ring, 6);
 200         if (ret)
 201                 return ret;
 202
 203         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(5));
 204         intel_ring_emit(ring, PIPE_CONTROL_CS_STALL |
 205                         PIPE_CONTROL_STALL_AT_SCOREBOARD);
 206         intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT); /* address */
 207         intel_ring_emit(ring, 0); /* low dword */
 208         intel_ring_emit(ring, 0); /* high dword */
 209         intel_ring_emit(ring, MI_NOOP);
 210         intel_ring_advance(ring);
 211
 212         ret = intel_ring_begin(ring, 6);
 213         if (ret)
 214                 return ret;
 215
 216         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(5));
 217         intel_ring_emit(ring, PIPE_CONTROL_QW_WRITE);
 218         intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT); /* address */
 219         intel_ring_emit(ring, 0);
 220         intel_ring_emit(ring, 0);
 221         intel_ring_emit(ring, MI_NOOP);
 222         intel_ring_advance(ring);
 223
 224         return 0;
 225 }
 226
 227 static int
 228 gen6_render_ring_flush(struct intel_engine_cs *ring,
 229                          u32 invalidate_domains, u32 flush_domains)
 230 {
 231         u32 flags = 0;
 232         u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES;
 233         int ret;
 234
 235         /* Force SNB workarounds for PIPE_CONTROL flushes */
 236         ret = intel_emit_post_sync_nonzero_flush(ring);
 237         if (ret)
 238                 return ret;
 239
 240         /* Just flush everything.  Experiments have shown that reducing the
 241          * number of bits based on the write domains has little performance
 242          * impact.
 243          */
 244         if (flush_domains) {
 245                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
 246                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
 247                 /*
 248                  * Ensure that any following seqno writes only happen
 249                  * when the render cache is indeed flushed.
 250                  */
 251                 flags |= PIPE_CONTROL_CS_STALL;
 252         }
 253         if (invalidate_domains) {
 254                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
 255                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 256                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 257                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
 258                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
 259                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
 260                 /*
 261                  * TLB invalidate requires a post-sync write.
 262                  */
 263                 flags |= PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_CS_STALL;
 264         }
 265
 266         ret = intel_ring_begin(ring, 4);
 267         if (ret)
 268                 return ret;
 269
 270         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4));
 271         intel_ring_emit(ring, flags);
 272         intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT);
 273         intel_ring_emit(ring, 0);
 274         intel_ring_advance(ring);
 275
 276         return 0;
 277 }
 278
 279 static int
 280 gen7_render_ring_cs_stall_wa(struct intel_engine_cs *ring)
 281 {
 282         int ret;
 283
 284         ret = intel_ring_begin(ring, 4);
 285         if (ret)
 286                 return ret;
 287
 288         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4));
 289         intel_ring_emit(ring, PIPE_CONTROL_CS_STALL |
 290                               PIPE_CONTROL_STALL_AT_SCOREBOARD);
 291         intel_ring_emit(ring, 0);
 292         intel_ring_emit(ring, 0);
 293         intel_ring_advance(ring);
 294
 295         return 0;
 296 }
 297
 298 static int gen7_ring_fbc_flush(struct intel_engine_cs *ring, u32 value)
 299 {
 300         int ret;
 301
 302         if (!ring->fbc_dirty)
 303                 return 0;
 304
 305         ret = intel_ring_begin(ring, 6);
 306         if (ret)
 307                 return ret;
 308         /* WaFbcNukeOn3DBlt:ivb/hsw */
 309         intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
 310         intel_ring_emit(ring, MSG_FBC_REND_STATE);
 311         intel_ring_emit(ring, value);
 312         intel_ring_emit(ring, MI_STORE_REGISTER_MEM(1) | MI_SRM_LRM_GLOBAL_GTT);
 313         intel_ring_emit(ring, MSG_FBC_REND_STATE);
 314         intel_ring_emit(ring, ring->scratch.gtt_offset + 256);
 315         intel_ring_advance(ring);
 316
 317         ring->fbc_dirty = false;
 318         return 0;
 319 }
 320
 321 static int
 322 gen7_render_ring_flush(struct intel_engine_cs *ring,
 323                        u32 invalidate_domains, u32 flush_domains)
 324 {
 325         u32 flags = 0;
 326         u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES;
 327         int ret;
 328
 329         /*
 330          * Ensure that any following seqno writes only happen when the render
 331          * cache is indeed flushed.
 332          *
 333          * Workaround: 4th PIPE_CONTROL command (except the ones with only
 334          * read-cache invalidate bits set) must have the CS_STALL bit set. We
 335          * don't try to be clever and just set it unconditionally.
 336          */
 337         flags |= PIPE_CONTROL_CS_STALL;
 338
 339         /* Just flush everything.  Experiments have shown that reducing the
 340          * number of bits based on the write domains has little performance
 341          * impact.
 342          */
 343         if (flush_domains) {
 344                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
 345                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
 346         }
 347         if (invalidate_domains) {
 348                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
 349                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 350                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 351                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
 352                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
 353                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
 354                 /*
 355                  * TLB invalidate requires a post-sync write.
 356                  */
 357                 flags |= PIPE_CONTROL_QW_WRITE;
 358                 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
 359
 360                 /* Workaround: we must issue a pipe_control with CS-stall bit
 361                  * set before a pipe_control command that has the state cache
 362                  * invalidate bit set. */
 363                 gen7_render_ring_cs_stall_wa(ring);
 364         }
 365
 366         ret = intel_ring_begin(ring, 4);
 367         if (ret)
 368                 return ret;
 369
 370         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4));
 371         intel_ring_emit(ring, flags);
 372         intel_ring_emit(ring, scratch_addr);
 373         intel_ring_emit(ring, 0);
 374         intel_ring_advance(ring);
 375
 376         if (!invalidate_domains && flush_domains)
 377                 return gen7_ring_fbc_flush(ring, FBC_REND_NUKE);
 378
 379         return 0;
 380 }
 381
 382 static int
 383 gen8_emit_pipe_control(struct intel_engine_cs *ring,
 384                        u32 flags, u32 scratch_addr)
 385 {
 386         int ret;
 387
 388         ret = intel_ring_begin(ring, 6);
 389         if (ret)
 390                 return ret;
 391
 392         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
 393         intel_ring_emit(ring, flags);
 394         intel_ring_emit(ring, scratch_addr);
 395         intel_ring_emit(ring, 0);
 396         intel_ring_emit(ring, 0);
 397         intel_ring_emit(ring, 0);
 398         intel_ring_advance(ring);
 399
 400         return 0;
 401 }
 402
 403 static int
 404 gen8_render_ring_flush(struct intel_engine_cs *ring,
 405                        u32 invalidate_domains, u32 flush_domains)
 406 {
 407         u32 flags = 0;
 408         u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES;
 409         int ret;
 410
 411         flags |= PIPE_CONTROL_CS_STALL;
 412
 413         if (flush_domains) {
 414                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
 415                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
 416         }
 417         if (invalidate_domains) {
 418                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
 419                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 420                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 421                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
 422                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
 423                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
 424                 flags |= PIPE_CONTROL_QW_WRITE;
 425                 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
 426
 427                 /* WaCsStallBeforeStateCacheInvalidate:bdw,chv */
 428                 ret = gen8_emit_pipe_control(ring,
 429                                              PIPE_CONTROL_CS_STALL |
 430                                              PIPE_CONTROL_STALL_AT_SCOREBOARD,
 431                                              0);
 432                 if (ret)
 433                         return ret;
 434         }
 435
 436         return gen8_emit_pipe_control(ring, flags, scratch_addr);
 437 }
 438
 439 static void ring_write_tail(struct intel_engine_cs *ring,
 440                             u32 value)
 441 {
 442         struct drm_i915_private *dev_priv = ring->dev->dev_private;
 443         I915_WRITE_TAIL(ring, value);
 444 }
 445
 446 u64 intel_ring_get_active_head(struct intel_engine_cs *ring)
 447 {
 448         struct drm_i915_private *dev_priv = ring->dev->dev_private;
 449         u64 acthd;
 450
 451         if (INTEL_INFO(ring->dev)->gen >= 8)
 452                 acthd = I915_READ64_2x32(RING_ACTHD(ring->mmio_base),
 453                                          RING_ACTHD_UDW(ring->mmio_base));
 454         else if (INTEL_INFO(ring->dev)->gen >= 4)
 455                 acthd = I915_READ(RING_ACTHD(ring->mmio_base));
 456         else
 457                 acthd = I915_READ(ACTHD);
 458
 459         return acthd;
 460 }
 461
 462 static void ring_setup_phys_status_page(struct intel_engine_cs *ring)
 463 {
 464         struct drm_i915_private *dev_priv = ring->dev->dev_private;
 465         u32 addr;
 466
 467         addr = dev_priv->status_page_dmah->busaddr;
 468         if (INTEL_INFO(ring->dev)->gen >= 4)
 469                 addr |= (dev_priv->status_page_dmah->busaddr >> 28) & 0xf0;
 470         I915_WRITE(HWS_PGA, addr);
 471 }
 472
 473 static bool stop_ring(struct intel_engine_cs *ring)
 474 {
 475         struct drm_i915_private *dev_priv = to_i915(ring->dev);
 476
 477         if (!IS_GEN2(ring->dev)) {
 478                 I915_WRITE_MODE(ring, _MASKED_BIT_ENABLE(STOP_RING));
 479                 if (wait_for_atomic((I915_READ_MODE(ring) & MODE_IDLE) != 0, 1000)) {
 480                         DRM_ERROR("%s :timed out trying to stop ring\n", ring->name);
 481                         return false;
 482                 }
 483         }
 484
 485         I915_WRITE_CTL(ring, 0);
 486         I915_WRITE_HEAD(ring, 0);
 487         ring->write_tail(ring, 0);
 488
 489         if (!IS_GEN2(ring->dev)) {
 490                 (void)I915_READ_CTL(ring);
 491                 I915_WRITE_MODE(ring, _MASKED_BIT_DISABLE(STOP_RING));
 492         }
 493
 494         return (I915_READ_HEAD(ring) & HEAD_ADDR) == 0;
 495 }
 496
 497 static int init_ring_common(struct intel_engine_cs *ring)
 498 {
 499         struct drm_device *dev = ring->dev;
 500         struct drm_i915_private *dev_priv = dev->dev_private;
 501         struct intel_ringbuffer *ringbuf = ring->buffer;
 502         struct drm_i915_gem_object *obj = ringbuf->obj;
 503         int ret = 0;
 504
 505         gen6_gt_force_wake_get(dev_priv, FORCEWAKE_ALL);
 506
 507         if (!stop_ring(ring)) {
 508                 /* G45 ring initialization often fails to reset head to zero */
 509                 DRM_DEBUG_KMS("%s head not reset to zero "
 510                               "ctl %08x head %08x tail %08x start %08x\n",
 511                               ring->name,
 512                               I915_READ_CTL(ring),
 513                               I915_READ_HEAD(ring),
 514                               I915_READ_TAIL(ring),
 515                               I915_READ_START(ring));
 516
 517                 if (!stop_ring(ring)) {
 518                         DRM_ERROR("failed to set %s head to zero "
 519                                   "ctl %08x head %08x tail %08x start %08x\n",
 520                                   ring->name,
 521                                   I915_READ_CTL(ring),
 522                                   I915_READ_HEAD(ring),
 523                                   I915_READ_TAIL(ring),
 524                                   I915_READ_START(ring));
 525                         ret = -EIO;
 526                         goto out;
 527                 }
 528         }
 529
 530         if (I915_NEED_GFX_HWS(dev))
 531                 intel_ring_setup_status_page(ring);
 532         else
 533                 ring_setup_phys_status_page(ring);
 534
 535         /* Enforce ordering by reading HEAD register back */
 536         I915_READ_HEAD(ring);
 537
 538         /* Initialize the ring. This must happen _after_ we've cleared the ring
 539          * registers with the above sequence (the readback of the HEAD registers
 540          * also enforces ordering), otherwise the hw might lose the new ring
 541          * register values. */
 542         I915_WRITE_START(ring, i915_gem_obj_ggtt_offset(obj));
 543         I915_WRITE_CTL(ring,
 544                         ((ringbuf->size - PAGE_SIZE) & RING_NR_PAGES)
 545                         | RING_VALID);
 546
 547         /* If the head is still not zero, the ring is dead */
 548         if (wait_for((I915_READ_CTL(ring) & RING_VALID) != 0 &&
 549                      I915_READ_START(ring) == i915_gem_obj_ggtt_offset(obj) &&
 550                      (I915_READ_HEAD(ring) & HEAD_ADDR) == 0, 50)) {
 551                 DRM_ERROR("%s initialization failed "
 552                           "ctl %08x (valid? %d) head %08x tail %08x start %08x [expected %08lx]\n",
 553                           ring->name,
 554                           I915_READ_CTL(ring), I915_READ_CTL(ring) & RING_VALID,
 555                           I915_READ_HEAD(ring), I915_READ_TAIL(ring),
 556                           I915_READ_START(ring), (unsigned long)i915_gem_obj_ggtt_offset(obj));
 557                 ret = -EIO;
 558                 goto out;
 559         }
 560
 561         if (!drm_core_check_feature(ring->dev, DRIVER_MODESET))
 562                 i915_kernel_lost_context(ring->dev);
 563         else {
 564                 ringbuf->head = I915_READ_HEAD(ring);
 565                 ringbuf->tail = I915_READ_TAIL(ring) & TAIL_ADDR;
 566                 ringbuf->space = ring_space(ringbuf);
 567                 ringbuf->last_retired_head = -1;
 568         }
 569
 570         memset(&ring->hangcheck, 0, sizeof(ring->hangcheck));
 571
 572 out:
 573         gen6_gt_force_wake_put(dev_priv, FORCEWAKE_ALL);
 574
 575         return ret;
 576 }
 577
 578 static int
 579 init_pipe_control(struct intel_engine_cs *ring)
 580 {
 581         int ret;
 582
 583         if (ring->scratch.obj)
 584                 return 0;
 585
 586         ring->scratch.obj = i915_gem_alloc_object(ring->dev, 4096);
 587         if (ring->scratch.obj == NULL) {
 588                 DRM_ERROR("Failed to allocate seqno page\n");
 589                 ret = -ENOMEM;
 590                 goto err;
 591         }
 592
 593         ret = i915_gem_object_set_cache_level(ring->scratch.obj, I915_CACHE_LLC);
 594         if (ret)
 595                 goto err_unref;
 596
 597         ret = i915_gem_obj_ggtt_pin(ring->scratch.obj, 4096, 0);
 598         if (ret)
 599                 goto err_unref;
 600
 601         ring->scratch.gtt_offset = i915_gem_obj_ggtt_offset(ring->scratch.obj);
 602         ring->scratch.cpu_page = kmap(sg_page(ring->scratch.obj->pages->sgl));
 603         if (ring->scratch.cpu_page == NULL) {
 604                 ret = -ENOMEM;
 605                 goto err_unpin;
 606         }
 607
 608         DRM_DEBUG_DRIVER("%s pipe control offset: 0x%08x\n",
 609                          ring->name, ring->scratch.gtt_offset);
 610         return 0;
 611
 612 err_unpin:
 613         i915_gem_object_ggtt_unpin(ring->scratch.obj);
 614 err_unref:
 615         drm_gem_object_unreference(&ring->scratch.obj->base);
 616 err:
 617         return ret;
 618 }
 619
 620 static int init_render_ring(struct intel_engine_cs *ring)
 621 {
 622         struct drm_device *dev = ring->dev;
 623         struct drm_i915_private *dev_priv = dev->dev_private;
 624         int ret = init_ring_common(ring);
 625         if (ret)
 626                 return ret;
 627
 628         /* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
 629         if (INTEL_INFO(dev)->gen >= 4 && INTEL_INFO(dev)->gen < 7)
 630                 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH));
 631
 632         /* We need to disable the AsyncFlip performance optimisations in order
 633          * to use MI_WAIT_FOR_EVENT within the CS. It should already be
 634          * programmed to '1' on all products.
 635          *
 636          * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
 637          */
 638         if (INTEL_INFO(dev)->gen >= 6)
 639                 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
 640
 641         /* Required for the hardware to program scanline values for waiting */
 642         /* WaEnableFlushTlbInvalidationMode:snb */
 643         if (INTEL_INFO(dev)->gen == 6)
 644                 I915_WRITE(GFX_MODE,
 645                            _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT));
 646
 647         /* WaBCSVCSTlbInvalidationMode:ivb,vlv,hsw */
 648         if (IS_GEN7(dev))
 649                 I915_WRITE(GFX_MODE_GEN7,
 650                            _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT) |
 651                            _MASKED_BIT_ENABLE(GFX_REPLAY_MODE));
 652
 653         if (INTEL_INFO(dev)->gen >= 5) {
 654                 ret = init_pipe_control(ring);
 655                 if (ret)
 656                         return ret;
 657         }
 658
 659         if (IS_GEN6(dev)) {
 660                 /* From the Sandybridge PRM, volume 1 part 3, page 24:
 661                  * "If this bit is set, STCunit will have LRA as replacement
 662                  *  policy. [...] This bit must be reset.  LRA replacement
 663                  *  policy is not supported."
 664                  */
 665                 I915_WRITE(CACHE_MODE_0,
 666                            _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
 667         }
 668
 669         if (INTEL_INFO(dev)->gen >= 6)
 670                 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
 671
 672         if (HAS_L3_DPF(dev))
 673                 I915_WRITE_IMR(ring, ~GT_PARITY_ERROR(dev));
 674
 675         return ret;
 676 }
 677
 678 static void render_ring_cleanup(struct intel_engine_cs *ring)
 679 {
 680         struct drm_device *dev = ring->dev;
 681         struct drm_i915_private *dev_priv = dev->dev_private;
 682
 683         if (dev_priv->semaphore_obj) {
 684                 i915_gem_object_ggtt_unpin(dev_priv->semaphore_obj);
 685                 drm_gem_object_unreference(&dev_priv->semaphore_obj->base);
 686                 dev_priv->semaphore_obj = NULL;
 687         }
 688
 689         if (ring->scratch.obj == NULL)
 690                 return;
 691
 692         if (INTEL_INFO(dev)->gen >= 5) {
 693                 kunmap(sg_page(ring->scratch.obj->pages->sgl));
 694                 i915_gem_object_ggtt_unpin(ring->scratch.obj);
 695         }
 696
 697         drm_gem_object_unreference(&ring->scratch.obj->base);
 698         ring->scratch.obj = NULL;
 699 }
 700
 701 static int gen8_rcs_signal(struct intel_engine_cs *signaller,
 702                            unsigned int num_dwords)
 703 {
 704 #define MBOX_UPDATE_DWORDS 8
 705         struct drm_device *dev = signaller->dev;
 706         struct drm_i915_private *dev_priv = dev->dev_private;
 707         struct intel_engine_cs *waiter;
 708         int i, ret, num_rings;
 709
 710         num_rings = hweight32(INTEL_INFO(dev)->ring_mask);
 711         num_dwords += (num_rings-1) * MBOX_UPDATE_DWORDS;
 712 #undef MBOX_UPDATE_DWORDS
 713
 714         ret = intel_ring_begin(signaller, num_dwords);
 715         if (ret)
 716                 return ret;
 717
 718         for_each_ring(waiter, dev_priv, i) {
 719                 u64 gtt_offset = signaller->semaphore.signal_ggtt[i];
 720                 if (gtt_offset == MI_SEMAPHORE_SYNC_INVALID)
 721                         continue;
 722
 723                 intel_ring_emit(signaller, GFX_OP_PIPE_CONTROL(6));
 724                 intel_ring_emit(signaller, PIPE_CONTROL_GLOBAL_GTT_IVB |
 725                                            PIPE_CONTROL_QW_WRITE |
 726                                            PIPE_CONTROL_FLUSH_ENABLE);
 727                 intel_ring_emit(signaller, lower_32_bits(gtt_offset));
 728                 intel_ring_emit(signaller, upper_32_bits(gtt_offset));
 729                 intel_ring_emit(signaller, signaller->outstanding_lazy_seqno);
 730                 intel_ring_emit(signaller, 0);
 731                 intel_ring_emit(signaller, MI_SEMAPHORE_SIGNAL |
 732                                            MI_SEMAPHORE_TARGET(waiter->id));
 733                 intel_ring_emit(signaller, 0);
 734         }
 735
 736         return 0;
 737 }
 738
 739 static int gen8_xcs_signal(struct intel_engine_cs *signaller,
 740                            unsigned int num_dwords)
 741 {
 742 #define MBOX_UPDATE_DWORDS 6
 743         struct drm_device *dev = signaller->dev;
 744         struct drm_i915_private *dev_priv = dev->dev_private;
 745         struct intel_engine_cs *waiter;
 746         int i, ret, num_rings;
 747
 748         num_rings = hweight32(INTEL_INFO(dev)->ring_mask);
 749         num_dwords += (num_rings-1) * MBOX_UPDATE_DWORDS;
 750 #undef MBOX_UPDATE_DWORDS
 751
 752         ret = intel_ring_begin(signaller, num_dwords);
 753         if (ret)
 754                 return ret;
 755
 756         for_each_ring(waiter, dev_priv, i) {
 757                 u64 gtt_offset = signaller->semaphore.signal_ggtt[i];
 758                 if (gtt_offset == MI_SEMAPHORE_SYNC_INVALID)
 759                         continue;
 760
 761                 intel_ring_emit(signaller, (MI_FLUSH_DW + 1) |
 762                                            MI_FLUSH_DW_OP_STOREDW);
 763                 intel_ring_emit(signaller, lower_32_bits(gtt_offset) |
 764                                            MI_FLUSH_DW_USE_GTT);
 765                 intel_ring_emit(signaller, upper_32_bits(gtt_offset));
 766                 intel_ring_emit(signaller, signaller->outstanding_lazy_seqno);
 767                 intel_ring_emit(signaller, MI_SEMAPHORE_SIGNAL |
 768                                            MI_SEMAPHORE_TARGET(waiter->id));
 769                 intel_ring_emit(signaller, 0);
 770         }
 771
 772         return 0;
 773 }
 774
 775 static int gen6_signal(struct intel_engine_cs *signaller,
 776                        unsigned int num_dwords)
 777 {
 778         struct drm_device *dev = signaller->dev;
 779         struct drm_i915_private *dev_priv = dev->dev_private;
 780         struct intel_engine_cs *useless;
 781         int i, ret, num_rings;
 782
 783 #define MBOX_UPDATE_DWORDS 3
 784         num_rings = hweight32(INTEL_INFO(dev)->ring_mask);
 785         num_dwords += round_up((num_rings-1) * MBOX_UPDATE_DWORDS, 2);
 786 #undef MBOX_UPDATE_DWORDS
 787
 788         ret = intel_ring_begin(signaller, num_dwords);
 789         if (ret)
 790                 return ret;
 791
 792         for_each_ring(useless, dev_priv, i) {
 793                 u32 mbox_reg = signaller->semaphore.mbox.signal[i];
 794                 if (mbox_reg != GEN6_NOSYNC) {
 795                         intel_ring_emit(signaller, MI_LOAD_REGISTER_IMM(1));
 796                         intel_ring_emit(signaller, mbox_reg);
 797                         intel_ring_emit(signaller, signaller->outstanding_lazy_seqno);
 798                 }
 799         }
 800
 801         /* If num_dwords was rounded, make sure the tail pointer is correct */
 802         if (num_rings % 2 == 0)
 803                 intel_ring_emit(signaller, MI_NOOP);
 804
 805         return 0;
 806 }
 807
 808 /**
 809  * gen6_add_request - Update the semaphore mailbox registers
 810  *
 811  * @ring - ring that is adding a request
 812  * @seqno - return seqno stuck into the ring
 813  *
 814  * Update the mailbox registers in the *other* rings with the current seqno.
 815  * This acts like a signal in the canonical semaphore.
 816  */
 817 static int
 818 gen6_add_request(struct intel_engine_cs *ring)
 819 {
 820         int ret;
 821
 822         if (ring->semaphore.signal)
 823                 ret = ring->semaphore.signal(ring, 4);
 824         else
 825                 ret = intel_ring_begin(ring, 4);
 826
 827         if (ret)
 828                 return ret;
 829
 830         intel_ring_emit(ring, MI_STORE_DWORD_INDEX);
 831         intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
 832         intel_ring_emit(ring, ring->outstanding_lazy_seqno);
 833         intel_ring_emit(ring, MI_USER_INTERRUPT);
 834         __intel_ring_advance(ring);
 835
 836         return 0;
 837 }
 838
 839 static inline bool i915_gem_has_seqno_wrapped(struct drm_device *dev,
 840                                               u32 seqno)
 841 {
 842         struct drm_i915_private *dev_priv = dev->dev_private;
 843         return dev_priv->last_seqno < seqno;
 844 }
 845
 846 /**
 847  * intel_ring_sync - sync the waiter to the signaller on seqno
 848  *
 849  * @waiter - ring that is waiting
 850  * @signaller - ring which has, or will signal
 851  * @seqno - seqno which the waiter will block on
 852  */
 853
 854 static int
 855 gen8_ring_sync(struct intel_engine_cs *waiter,
 856                struct intel_engine_cs *signaller,
 857                u32 seqno)
 858 {
 859         struct drm_i915_private *dev_priv = waiter->dev->dev_private;
 860         int ret;
 861
 862         ret = intel_ring_begin(waiter, 4);
 863         if (ret)
 864                 return ret;
 865
 866         intel_ring_emit(waiter, MI_SEMAPHORE_WAIT |
 867                                 MI_SEMAPHORE_GLOBAL_GTT |
 868                                 MI_SEMAPHORE_POLL |
 869                                 MI_SEMAPHORE_SAD_GTE_SDD);
 870         intel_ring_emit(waiter, seqno);
 871         intel_ring_emit(waiter,
 872                         lower_32_bits(GEN8_WAIT_OFFSET(waiter, signaller->id)));
 873         intel_ring_emit(waiter,
 874                         upper_32_bits(GEN8_WAIT_OFFSET(waiter, signaller->id)));
 875         intel_ring_advance(waiter);
 876         return 0;
 877 }
 878
 879 static int
 880 gen6_ring_sync(struct intel_engine_cs *waiter,
 881                struct intel_engine_cs *signaller,
 882                u32 seqno)
 883 {
 884         u32 dw1 = MI_SEMAPHORE_MBOX |
 885                   MI_SEMAPHORE_COMPARE |
 886                   MI_SEMAPHORE_REGISTER;
 887         u32 wait_mbox = signaller->semaphore.mbox.wait[waiter->id];
 888         int ret;
 889
 890         /* Throughout all of the GEM code, seqno passed implies our current
 891          * seqno is >= the last seqno executed. However for hardware the
 892          * comparison is strictly greater than.
 893          */
 894         seqno -= 1;
 895
 896         WARN_ON(wait_mbox == MI_SEMAPHORE_SYNC_INVALID);
 897
 898         ret = intel_ring_begin(waiter, 4);
 899         if (ret)
 900                 return ret;
 901
 902         /* If seqno wrap happened, omit the wait with no-ops */
 903         if (likely(!i915_gem_has_seqno_wrapped(waiter->dev, seqno))) {
 904                 intel_ring_emit(waiter, dw1 | wait_mbox);
 905                 intel_ring_emit(waiter, seqno);
 906                 intel_ring_emit(waiter, 0);
 907                 intel_ring_emit(waiter, MI_NOOP);
 908         } else {
 909                 intel_ring_emit(waiter, MI_NOOP);
 910                 intel_ring_emit(waiter, MI_NOOP);
 911                 intel_ring_emit(waiter, MI_NOOP);
 912                 intel_ring_emit(waiter, MI_NOOP);
 913         }
 914         intel_ring_advance(waiter);
 915
 916         return 0;
 917 }
 918
 919 #define PIPE_CONTROL_FLUSH(ring__, addr__)                                      \
 920 do {                                                                    \
 921         intel_ring_emit(ring__, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE |                \
 922                  PIPE_CONTROL_DEPTH_STALL);                             \
 923         intel_ring_emit(ring__, (addr__) | PIPE_CONTROL_GLOBAL_GTT);                    \
 924         intel_ring_emit(ring__, 0);                                                     \
 925         intel_ring_emit(ring__, 0);                                                     \
 926 } while (0)
 927
 928 static int
 929 pc_render_add_request(struct intel_engine_cs *ring)
 930 {
 931         u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES;
 932         int ret;
 933
 934         /* For Ironlake, MI_USER_INTERRUPT was deprecated and apparently
 935          * incoherent with writes to memory, i.e. completely fubar,
 936          * so we need to use PIPE_NOTIFY instead.
 937          *
 938          * However, we also need to workaround the qword write
 939          * incoherence by flushing the 6 PIPE_NOTIFY buffers out to
 940          * memory before requesting an interrupt.
 941          */
 942         ret = intel_ring_begin(ring, 32);
 943         if (ret)
 944                 return ret;
 945
 946         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE |
 947                         PIPE_CONTROL_WRITE_FLUSH |
 948                         PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE);
 949         intel_ring_emit(ring, ring->scratch.gtt_offset | PIPE_CONTROL_GLOBAL_GTT);
 950         intel_ring_emit(ring, ring->outstanding_lazy_seqno);
 951         intel_ring_emit(ring, 0);
 952         PIPE_CONTROL_FLUSH(ring, scratch_addr);
 953         scratch_addr += 2 * CACHELINE_BYTES; /* write to separate cachelines */
 954         PIPE_CONTROL_FLUSH(ring, scratch_addr);
 955         scratch_addr += 2 * CACHELINE_BYTES;
 956         PIPE_CONTROL_FLUSH(ring, scratch_addr);
 957         scratch_addr += 2 * CACHELINE_BYTES;
 958         PIPE_CONTROL_FLUSH(ring, scratch_addr);
 959         scratch_addr += 2 * CACHELINE_BYTES;
 960         PIPE_CONTROL_FLUSH(ring, scratch_addr);
 961         scratch_addr += 2 * CACHELINE_BYTES;
 962         PIPE_CONTROL_FLUSH(ring, scratch_addr);
 963
 964         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE |
 965                         PIPE_CONTROL_WRITE_FLUSH |
 966                         PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
 967                         PIPE_CONTROL_NOTIFY);
 968         intel_ring_emit(ring, ring->scratch.gtt_offset | PIPE_CONTROL_GLOBAL_GTT);
 969         intel_ring_emit(ring, ring->outstanding_lazy_seqno);
 970         intel_ring_emit(ring, 0);
 971         __intel_ring_advance(ring);
 972
 973         return 0;
 974 }
 975
 976 static u32
 977 gen6_ring_get_seqno(struct intel_engine_cs *ring, bool lazy_coherency)
 978 {
 979         /* Workaround to force correct ordering between irq and seqno writes on
 980          * ivb (and maybe also on snb) by reading from a CS register (like
 981          * ACTHD) before reading the status page. */
 982         if (!lazy_coherency) {
 983                 struct drm_i915_private *dev_priv = ring->dev->dev_private;
 984                 POSTING_READ(RING_ACTHD(ring->mmio_base));
 985         }
 986
 987         return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
 988 }
 989
 990 static u32
 991 ring_get_seqno(struct intel_engine_cs *ring, bool lazy_coherency)
 992 {
 993         return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
 994 }
 995
 996 static void
 997 ring_set_seqno(struct intel_engine_cs *ring, u32 seqno)
 998 {
 999         intel_write_status_page(ring, I915_GEM_HWS_INDEX, seqno);
1000 }
1001
1002 static u32
1003 pc_render_get_seqno(struct intel_engine_cs *ring, bool lazy_coherency)
1004 {
1005         return ring->scratch.cpu_page[0];
1006 }
1007
1008 static void
1009 pc_render_set_seqno(struct intel_engine_cs *ring, u32 seqno)
1010 {
1011         ring->scratch.cpu_page[0] = seqno;
1012 }
1013
1014 static bool
1015 gen5_ring_get_irq(struct intel_engine_cs *ring)
1016 {
1017         struct drm_device *dev = ring->dev;
1018         struct drm_i915_private *dev_priv = dev->dev_private;
1019         unsigned long flags;
1020
1021         if (!dev->irq_enabled)
1022                 return false;
1023
1024         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1025         if (ring->irq_refcount++ == 0)
1026                 gen5_enable_gt_irq(dev_priv, ring->irq_enable_mask);
1027         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1028
1029         return true;
1030 }
1031
1032 static void
1033 gen5_ring_put_irq(struct intel_engine_cs *ring)
1034 {
1035         struct drm_device *dev = ring->dev;
1036         struct drm_i915_private *dev_priv = dev->dev_private;
1037         unsigned long flags;
1038
1039         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1040         if (--ring->irq_refcount == 0)
1041                 gen5_disable_gt_irq(dev_priv, ring->irq_enable_mask);
1042         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1043 }
1044
1045 static bool
1046 i9xx_ring_get_irq(struct intel_engine_cs *ring)
1047 {
1048         struct drm_device *dev = ring->dev;
1049         struct drm_i915_private *dev_priv = dev->dev_private;
1050         unsigned long flags;
1051
1052         if (!dev->irq_enabled)
1053                 return false;
1054
1055         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1056         if (ring->irq_refcount++ == 0) {
1057                 dev_priv->irq_mask &= ~ring->irq_enable_mask;
1058                 I915_WRITE(IMR, dev_priv->irq_mask);
1059                 POSTING_READ(IMR);
1060         }
1061         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1062
1063         return true;
1064 }
1065
1066 static void
1067 i9xx_ring_put_irq(struct intel_engine_cs *ring)
1068 {
1069         struct drm_device *dev = ring->dev;
1070         struct drm_i915_private *dev_priv = dev->dev_private;
1071         unsigned long flags;
1072
1073         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1074         if (--ring->irq_refcount == 0) {
1075                 dev_priv->irq_mask |= ring->irq_enable_mask;
1076                 I915_WRITE(IMR, dev_priv->irq_mask);
1077                 POSTING_READ(IMR);
1078         }
1079         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1080 }
1081
1082 static bool
1083 i8xx_ring_get_irq(struct intel_engine_cs *ring)
1084 {
1085         struct drm_device *dev = ring->dev;
1086         struct drm_i915_private *dev_priv = dev->dev_private;
1087         unsigned long flags;
1088
1089         if (!dev->irq_enabled)
1090                 return false;
1091
1092         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1093         if (ring->irq_refcount++ == 0) {
1094                 dev_priv->irq_mask &= ~ring->irq_enable_mask;
1095                 I915_WRITE16(IMR, dev_priv->irq_mask);
1096                 POSTING_READ16(IMR);
1097         }
1098         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1099
1100         return true;
1101 }
1102
1103 static void
1104 i8xx_ring_put_irq(struct intel_engine_cs *ring)
1105 {
1106         struct drm_device *dev = ring->dev;
1107         struct drm_i915_private *dev_priv = dev->dev_private;
1108         unsigned long flags;
1109
1110         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1111         if (--ring->irq_refcount == 0) {
1112                 dev_priv->irq_mask |= ring->irq_enable_mask;
1113                 I915_WRITE16(IMR, dev_priv->irq_mask);
1114                 POSTING_READ16(IMR);
1115         }
1116         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1117 }
1118
1119 void intel_ring_setup_status_page(struct intel_engine_cs *ring)
1120 {
1121         struct drm_device *dev = ring->dev;
1122         struct drm_i915_private *dev_priv = ring->dev->dev_private;
1123         u32 mmio = 0;
1124
1125         /* The ring status page addresses are no longer next to the rest of
1126          * the ring registers as of gen7.
1127          */
1128         if (IS_GEN7(dev)) {
1129                 switch (ring->id) {
1130                 case RCS:
1131                         mmio = RENDER_HWS_PGA_GEN7;
1132                         break;
1133                 case BCS:
1134                         mmio = BLT_HWS_PGA_GEN7;
1135                         break;
1136                 /*
1137                  * VCS2 actually doesn't exist on Gen7. Only shut up
1138                  * gcc switch check warning
1139                  */
1140                 case VCS2:
1141                 case VCS:
1142                         mmio = BSD_HWS_PGA_GEN7;
1143                         break;
1144                 case VECS:
1145                         mmio = VEBOX_HWS_PGA_GEN7;
1146                         break;
1147                 }
1148         } else if (IS_GEN6(ring->dev)) {
1149                 mmio = RING_HWS_PGA_GEN6(ring->mmio_base);
1150         } else {
1151                 /* XXX: gen8 returns to sanity */
1152                 mmio = RING_HWS_PGA(ring->mmio_base);
1153         }
1154
1155         I915_WRITE(mmio, (u32)ring->status_page.gfx_addr);
1156         POSTING_READ(mmio);
1157
1158         /*
1159          * Flush the TLB for this page
1160          *
1161          * FIXME: These two bits have disappeared on gen8, so a question
1162          * arises: do we still need this and if so how should we go about
1163          * invalidating the TLB?
1164          */
1165         if (INTEL_INFO(dev)->gen >= 6 && INTEL_INFO(dev)->gen < 8) {
1166                 u32 reg = RING_INSTPM(ring->mmio_base);
1167
1168                 /* ring should be idle before issuing a sync flush*/
1169                 WARN_ON((I915_READ_MODE(ring) & MODE_IDLE) == 0);
1170
1171                 I915_WRITE(reg,
1172                            _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE |
1173                                               INSTPM_SYNC_FLUSH));
1174                 if (wait_for((I915_READ(reg) & INSTPM_SYNC_FLUSH) == 0,
1175                              1000))
1176                         DRM_ERROR("%s: wait for SyncFlush to complete for TLB invalidation timed out\n",
1177                                   ring->name);
1178         }
1179 }
1180
1181 static int
1182 bsd_ring_flush(struct intel_engine_cs *ring,
1183                u32     invalidate_domains,
1184                u32     flush_domains)
1185 {
1186         int ret;
1187
1188         ret = intel_ring_begin(ring, 2);
1189         if (ret)
1190                 return ret;
1191
1192         intel_ring_emit(ring, MI_FLUSH);
1193         intel_ring_emit(ring, MI_NOOP);
1194         intel_ring_advance(ring);
1195         return 0;
1196 }
1197
1198 static int
1199 i9xx_add_request(struct intel_engine_cs *ring)
1200 {
1201         int ret;
1202
1203         ret = intel_ring_begin(ring, 4);
1204         if (ret)
1205                 return ret;
1206
1207         intel_ring_emit(ring, MI_STORE_DWORD_INDEX);
1208         intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
1209         intel_ring_emit(ring, ring->outstanding_lazy_seqno);
1210         intel_ring_emit(ring, MI_USER_INTERRUPT);
1211         __intel_ring_advance(ring);
1212
1213         return 0;
1214 }
1215
1216 static bool
1217 gen6_ring_get_irq(struct intel_engine_cs *ring)
1218 {
1219         struct drm_device *dev = ring->dev;
1220         struct drm_i915_private *dev_priv = dev->dev_private;
1221         unsigned long flags;
1222
1223         if (!dev->irq_enabled)
1224                return false;
1225
1226         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1227         if (ring->irq_refcount++ == 0) {
1228                 if (HAS_L3_DPF(dev) && ring->id == RCS)
1229                         I915_WRITE_IMR(ring,
1230                                        ~(ring->irq_enable_mask |
1231                                          GT_PARITY_ERROR(dev)));
1232                 else
1233                         I915_WRITE_IMR(ring, ~ring->irq_enable_mask);
1234                 gen5_enable_gt_irq(dev_priv, ring->irq_enable_mask);
1235         }
1236         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1237
1238         return true;
1239 }
1240
1241 static void
1242 gen6_ring_put_irq(struct intel_engine_cs *ring)
1243 {
1244         struct drm_device *dev = ring->dev;
1245         struct drm_i915_private *dev_priv = dev->dev_private;
1246         unsigned long flags;
1247
1248         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1249         if (--ring->irq_refcount == 0) {
1250                 if (HAS_L3_DPF(dev) && ring->id == RCS)
1251                         I915_WRITE_IMR(ring, ~GT_PARITY_ERROR(dev));
1252                 else
1253                         I915_WRITE_IMR(ring, ~0);
1254                 gen5_disable_gt_irq(dev_priv, ring->irq_enable_mask);
1255         }
1256         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1257 }
1258
1259 static bool
1260 hsw_vebox_get_irq(struct intel_engine_cs *ring)
1261 {
1262         struct drm_device *dev = ring->dev;
1263         struct drm_i915_private *dev_priv = dev->dev_private;
1264         unsigned long flags;
1265
1266         if (!dev->irq_enabled)
1267                 return false;
1268
1269         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1270         if (ring->irq_refcount++ == 0) {
1271                 I915_WRITE_IMR(ring, ~ring->irq_enable_mask);
1272                 gen6_enable_pm_irq(dev_priv, ring->irq_enable_mask);
1273         }
1274         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1275
1276         return true;
1277 }
1278
1279 static void
1280 hsw_vebox_put_irq(struct intel_engine_cs *ring)
1281 {
1282         struct drm_device *dev = ring->dev;
1283         struct drm_i915_private *dev_priv = dev->dev_private;
1284         unsigned long flags;
1285
1286         if (!dev->irq_enabled)
1287                 return;
1288
1289         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1290         if (--ring->irq_refcount == 0) {
1291                 I915_WRITE_IMR(ring, ~0);
1292                 gen6_disable_pm_irq(dev_priv, ring->irq_enable_mask);
1293         }
1294         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1295 }
1296
1297 static bool
1298 gen8_ring_get_irq(struct intel_engine_cs *ring)
1299 {
1300         struct drm_device *dev = ring->dev;
1301         struct drm_i915_private *dev_priv = dev->dev_private;
1302         unsigned long flags;
1303
1304         if (!dev->irq_enabled)
1305                 return false;
1306
1307         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1308         if (ring->irq_refcount++ == 0) {
1309                 if (HAS_L3_DPF(dev) && ring->id == RCS) {
1310                         I915_WRITE_IMR(ring,
1311                                        ~(ring->irq_enable_mask |
1312                                          GT_RENDER_L3_PARITY_ERROR_INTERRUPT));
1313                 } else {
1314                         I915_WRITE_IMR(ring, ~ring->irq_enable_mask);
1315                 }
1316                 POSTING_READ(RING_IMR(ring->mmio_base));
1317         }
1318         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1319
1320         return true;
1321 }
1322
1323 static void
1324 gen8_ring_put_irq(struct intel_engine_cs *ring)
1325 {
1326         struct drm_device *dev = ring->dev;
1327         struct drm_i915_private *dev_priv = dev->dev_private;
1328         unsigned long flags;
1329
1330         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1331         if (--ring->irq_refcount == 0) {
1332                 if (HAS_L3_DPF(dev) && ring->id == RCS) {
1333                         I915_WRITE_IMR(ring,
1334                                        ~GT_RENDER_L3_PARITY_ERROR_INTERRUPT);
1335                 } else {
1336                         I915_WRITE_IMR(ring, ~0);
1337                 }
1338                 POSTING_READ(RING_IMR(ring->mmio_base));
1339         }
1340         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1341 }
1342
1343 static int
1344 i965_dispatch_execbuffer(struct intel_engine_cs *ring,
1345                          u64 offset, u32 length,
1346                          unsigned flags)
1347 {
1348         int ret;
1349
1350         ret = intel_ring_begin(ring, 2);
1351         if (ret)
1352                 return ret;
1353
1354         intel_ring_emit(ring,
1355                         MI_BATCH_BUFFER_START |
1356                         MI_BATCH_GTT |
1357                         (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE_I965));
1358         intel_ring_emit(ring, offset);
1359         intel_ring_advance(ring);
1360
1361         return 0;
1362 }
1363
1364 /* Just userspace ABI convention to limit the wa batch bo to a resonable size */
1365 #define I830_BATCH_LIMIT (256*1024)
1366 #define I830_TLB_ENTRIES (2)
1367 #define I830_WA_SIZE max(I830_TLB_ENTRIES*4096, I830_BATCH_LIMIT)
1368 static int
1369 i830_dispatch_execbuffer(struct intel_engine_cs *ring,
1370                                 u64 offset, u32 len,
1371                                 unsigned flags)
1372 {
1373         u32 cs_offset = ring->scratch.gtt_offset;
1374         int ret;
1375
1376         ret = intel_ring_begin(ring, 6);
1377         if (ret)
1378                 return ret;
1379
1380         /* Evict the invalid PTE TLBs */
1381         intel_ring_emit(ring, COLOR_BLT_CMD | BLT_WRITE_RGBA);
1382         intel_ring_emit(ring, BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | 4096);
1383         intel_ring_emit(ring, I830_TLB_ENTRIES << 16 | 4); /* load each page */
1384         intel_ring_emit(ring, cs_offset);
1385         intel_ring_emit(ring, 0xdeadbeef);
1386         intel_ring_emit(ring, MI_NOOP);
1387         intel_ring_advance(ring);
1388
1389         if ((flags & I915_DISPATCH_PINNED) == 0) {
1390                 if (len > I830_BATCH_LIMIT)
1391                         return -ENOSPC;
1392
1393                 ret = intel_ring_begin(ring, 6 + 2);
1394                 if (ret)
1395                         return ret;
1396
1397                 /* Blit the batch (which has now all relocs applied) to the
1398                  * stable batch scratch bo area (so that the CS never
1399                  * stumbles over its tlb invalidation bug) ...
1400                  */
1401                 intel_ring_emit(ring, SRC_COPY_BLT_CMD | BLT_WRITE_RGBA);
1402                 intel_ring_emit(ring, BLT_DEPTH_32 | BLT_ROP_SRC_COPY | 4096);
1403                 intel_ring_emit(ring, DIV_ROUND_UP(len, 4096) << 16 | 4096);
1404                 intel_ring_emit(ring, cs_offset);
1405                 intel_ring_emit(ring, 4096);
1406                 intel_ring_emit(ring, offset);
1407
1408                 intel_ring_emit(ring, MI_FLUSH);
1409                 intel_ring_emit(ring, MI_NOOP);
1410                 intel_ring_advance(ring);
1411
1412                 /* ... and execute it. */
1413                 offset = cs_offset;
1414         }
1415
1416         ret = intel_ring_begin(ring, 4);
1417         if (ret)
1418                 return ret;
1419
1420         intel_ring_emit(ring, MI_BATCH_BUFFER);
1421         intel_ring_emit(ring, offset | (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE));
1422         intel_ring_emit(ring, offset + len - 8);
1423         intel_ring_emit(ring, MI_NOOP);
1424         intel_ring_advance(ring);
1425
1426         return 0;
1427 }
1428
1429 static int
1430 i915_dispatch_execbuffer(struct intel_engine_cs *ring,
1431                          u64 offset, u32 len,
1432                          unsigned flags)
1433 {
1434         int ret;
1435
1436         ret = intel_ring_begin(ring, 2);
1437         if (ret)
1438                 return ret;
1439
1440         intel_ring_emit(ring, MI_BATCH_BUFFER_START | MI_BATCH_GTT);
1441         intel_ring_emit(ring, offset | (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE));
1442         intel_ring_advance(ring);
1443
1444         return 0;
1445 }
1446
1447 static void cleanup_status_page(struct intel_engine_cs *ring)
1448 {
1449         struct drm_i915_gem_object *obj;
1450
1451         obj = ring->status_page.obj;
1452         if (obj == NULL)
1453                 return;
1454
1455         kunmap(sg_page(obj->pages->sgl));
1456         i915_gem_object_ggtt_unpin(obj);
1457         drm_gem_object_unreference(&obj->base);
1458         ring->status_page.obj = NULL;
1459 }
1460
1461 static int init_status_page(struct intel_engine_cs *ring)
1462 {
1463         struct drm_i915_gem_object *obj;
1464
1465         if ((obj = ring->status_page.obj) == NULL) {
1466                 unsigned flags;
1467                 int ret;
1468
1469                 obj = i915_gem_alloc_object(ring->dev, 4096);
1470                 if (obj == NULL) {
1471                         DRM_ERROR("Failed to allocate status page\n");
1472                         return -ENOMEM;
1473                 }
1474
1475                 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_LLC);
1476                 if (ret)
1477                         goto err_unref;
1478
1479                 flags = 0;
1480                 if (!HAS_LLC(ring->dev))
1481                         /* On g33, we cannot place HWS above 256MiB, so
1482                          * restrict its pinning to the low mappable arena.
1483                          * Though this restriction is not documented for
1484                          * gen4, gen5, or byt, they also behave similarly
1485                          * and hang if the HWS is placed at the top of the
1486                          * GTT. To generalise, it appears that all !llc
1487                          * platforms have issues with us placing the HWS
1488                          * above the mappable region (even though we never
1489                          * actualy map it).
1490                          */
1491                         flags |= PIN_MAPPABLE;
1492                 ret = i915_gem_obj_ggtt_pin(obj, 4096, flags);
1493                 if (ret) {
1494 err_unref:
1495                         drm_gem_object_unreference(&obj->base);
1496                         return ret;
1497                 }
1498
1499                 ring->status_page.obj = obj;
1500         }
1501
1502         ring->status_page.gfx_addr = i915_gem_obj_ggtt_offset(obj);
1503         ring->status_page.page_addr = kmap(sg_page(obj->pages->sgl));
1504         memset(ring->status_page.page_addr, 0, PAGE_SIZE);
1505
1506         DRM_DEBUG_DRIVER("%s hws offset: 0x%08x\n",
1507                         ring->name, ring->status_page.gfx_addr);
1508
1509         return 0;
1510 }
1511
1512 static int init_phys_status_page(struct intel_engine_cs *ring)
1513 {
1514         struct drm_i915_private *dev_priv = ring->dev->dev_private;
1515
1516         if (!dev_priv->status_page_dmah) {
1517                 dev_priv->status_page_dmah =
1518                         drm_pci_alloc(ring->dev, PAGE_SIZE, PAGE_SIZE);
1519                 if (!dev_priv->status_page_dmah)
1520                         return -ENOMEM;
1521         }
1522
1523         ring->status_page.page_addr = dev_priv->status_page_dmah->vaddr;
1524         memset(ring->status_page.page_addr, 0, PAGE_SIZE);
1525
1526         return 0;
1527 }
1528
1529 static void intel_destroy_ringbuffer_obj(struct intel_ringbuffer *ringbuf)
1530 {
1531         if (!ringbuf->obj)
1532                 return;
1533
1534         iounmap(ringbuf->virtual_start);
1535         i915_gem_object_ggtt_unpin(ringbuf->obj);
1536         drm_gem_object_unreference(&ringbuf->obj->base);
1537         ringbuf->obj = NULL;
1538 }
1539
1540 static int intel_alloc_ringbuffer_obj(struct drm_device *dev,
1541                                       struct intel_ringbuffer *ringbuf)
1542 {
1543         struct drm_i915_private *dev_priv = to_i915(dev);
1544         struct drm_i915_gem_object *obj;
1545         int ret;
1546
1547         if (ringbuf->obj)
1548                 return 0;
1549
1550         obj = NULL;
1551         if (!HAS_LLC(dev))
1552                 obj = i915_gem_object_create_stolen(dev, ringbuf->size);
1553         if (obj == NULL)
1554                 obj = i915_gem_alloc_object(dev, ringbuf->size);
1555         if (obj == NULL)
1556                 return -ENOMEM;
1557
1558         /* mark ring buffers as read-only from GPU side by default */
1559         obj->gt_ro = 1;
1560
1561         ret = i915_gem_obj_ggtt_pin(obj, PAGE_SIZE, PIN_MAPPABLE);
1562         if (ret)
1563                 goto err_unref;
1564
1565         ret = i915_gem_object_set_to_gtt_domain(obj, true);
1566         if (ret)
1567                 goto err_unpin;
1568
1569         ringbuf->virtual_start =
1570                 ioremap_wc(dev_priv->gtt.mappable_base + i915_gem_obj_ggtt_offset(obj),
1571                                 ringbuf->size);
1572         if (ringbuf->virtual_start == NULL) {
1573                 ret = -EINVAL;
1574                 goto err_unpin;
1575         }
1576
1577         ringbuf->obj = obj;
1578         return 0;
1579
1580 err_unpin:
1581         i915_gem_object_ggtt_unpin(obj);
1582 err_unref:
1583         drm_gem_object_unreference(&obj->base);
1584         return ret;
1585 }
1586
1587 static int intel_init_ring_buffer(struct drm_device *dev,
1588                                   struct intel_engine_cs *ring)
1589 {
1590         struct intel_ringbuffer *ringbuf = ring->buffer;
1591         int ret;
1592
1593         if (ringbuf == NULL) {
1594                 ringbuf = kzalloc(sizeof(*ringbuf), GFP_KERNEL);
1595                 if (!ringbuf)
1596                         return -ENOMEM;
1597                 ring->buffer = ringbuf;
1598         }
1599
1600         ring->dev = dev;
1601         INIT_LIST_HEAD(&ring->active_list);
1602         INIT_LIST_HEAD(&ring->request_list);
1603         ringbuf->size = 32 * PAGE_SIZE;
1604         memset(ring->semaphore.sync_seqno, 0, sizeof(ring->semaphore.sync_seqno));
1605
1606         init_waitqueue_head(&ring->irq_queue);
1607
1608         if (I915_NEED_GFX_HWS(dev)) {
1609                 ret = init_status_page(ring);
1610                 if (ret)
1611                         goto error;
1612         } else {
1613                 BUG_ON(ring->id != RCS);
1614                 ret = init_phys_status_page(ring);
1615                 if (ret)
1616                         goto error;
1617         }
1618
1619         ret = intel_alloc_ringbuffer_obj(dev, ringbuf);
1620         if (ret) {
1621                 DRM_ERROR("Failed to allocate ringbuffer %s: %d\n", ring->name, ret);
1622                 goto error;
1623         }
1624
1625         /* Workaround an erratum on the i830 which causes a hang if
1626          * the TAIL pointer points to within the last 2 cachelines
1627          * of the buffer.
1628          */
1629         ringbuf->effective_size = ringbuf->size;
1630         if (IS_I830(dev) || IS_845G(dev))
1631                 ringbuf->effective_size -= 2 * CACHELINE_BYTES;
1632
1633         ret = i915_cmd_parser_init_ring(ring);
1634         if (ret)
1635                 goto error;
1636
1637         ret = ring->init(ring);
1638         if (ret)
1639                 goto error;
1640
1641         return 0;
1642
1643 error:
1644         kfree(ringbuf);
1645         ring->buffer = NULL;
1646         return ret;
1647 }
1648
1649 void intel_cleanup_ring_buffer(struct intel_engine_cs *ring)
1650 {
1651         struct drm_i915_private *dev_priv = to_i915(ring->dev);
1652         struct intel_ringbuffer *ringbuf = ring->buffer;
1653
1654         if (!intel_ring_initialized(ring))
1655                 return;
1656
1657         intel_stop_ring_buffer(ring);
1658         WARN_ON(!IS_GEN2(ring->dev) && (I915_READ_MODE(ring) & MODE_IDLE) == 0);
1659
1660         intel_destroy_ringbuffer_obj(ringbuf);
1661         ring->preallocated_lazy_request = NULL;
1662         ring->outstanding_lazy_seqno = 0;
1663
1664         if (ring->cleanup)
1665                 ring->cleanup(ring);
1666
1667         cleanup_status_page(ring);
1668
1669         i915_cmd_parser_fini_ring(ring);
1670
1671         kfree(ringbuf);
1672         ring->buffer = NULL;
1673 }
1674
1675 static int intel_ring_wait_request(struct intel_engine_cs *ring, int n)
1676 {
1677         struct intel_ringbuffer *ringbuf = ring->buffer;
1678         struct drm_i915_gem_request *request;
1679         u32 seqno = 0;
1680         int ret;
1681
1682         if (ringbuf->last_retired_head != -1) {
1683                 ringbuf->head = ringbuf->last_retired_head;
1684                 ringbuf->last_retired_head = -1;
1685
1686                 ringbuf->space = ring_space(ringbuf);
1687                 if (ringbuf->space >= n)
1688                         return 0;
1689         }
1690
1691         list_for_each_entry(request, &ring->request_list, list) {
1692                 if (__ring_space(request->tail, ringbuf->tail, ringbuf->size) >= n) {
1693                         seqno = request->seqno;
1694                         break;
1695                 }
1696         }
1697
1698         if (seqno == 0)
1699                 return -ENOSPC;
1700
1701         ret = i915_wait_seqno(ring, seqno);
1702         if (ret)
1703                 return ret;
1704
1705         i915_gem_retire_requests_ring(ring);
1706         ringbuf->head = ringbuf->last_retired_head;
1707         ringbuf->last_retired_head = -1;
1708
1709         ringbuf->space = ring_space(ringbuf);
1710         return 0;
1711 }
1712
1713 static int ring_wait_for_space(struct intel_engine_cs *ring, int n)
1714 {
1715         struct drm_device *dev = ring->dev;
1716         struct drm_i915_private *dev_priv = dev->dev_private;
1717         struct intel_ringbuffer *ringbuf = ring->buffer;
1718         unsigned long end;
1719         int ret;
1720
1721         ret = intel_ring_wait_request(ring, n);
1722         if (ret != -ENOSPC)
1723                 return ret;
1724
1725         /* force the tail write in case we have been skipping them */
1726         __intel_ring_advance(ring);
1727
1728         /* With GEM the hangcheck timer should kick us out of the loop,
1729          * leaving it early runs the risk of corrupting GEM state (due
1730          * to running on almost untested codepaths). But on resume
1731          * timers don't work yet, so prevent a complete hang in that
1732          * case by choosing an insanely large timeout. */
1733         end = jiffies + 60 * HZ;
1734
1735         trace_i915_ring_wait_begin(ring);
1736         do {
1737                 ringbuf->head = I915_READ_HEAD(ring);
1738                 ringbuf->space = ring_space(ringbuf);
1739                 if (ringbuf->space >= n) {
1740                         ret = 0;
1741                         break;
1742                 }
1743
1744                 if (!drm_core_check_feature(dev, DRIVER_MODESET) &&
1745                     dev->primary->master) {
1746                         struct drm_i915_master_private *master_priv = dev->primary->master->driver_priv;
1747                         if (master_priv->sarea_priv)
1748                                 master_priv->sarea_priv->perf_boxes |= I915_BOX_WAIT;
1749                 }
1750
1751                 msleep(1);
1752
1753                 if (dev_priv->mm.interruptible && signal_pending(current)) {
1754                         ret = -ERESTARTSYS;
1755                         break;
1756                 }
1757
1758                 ret = i915_gem_check_wedge(&dev_priv->gpu_error,
1759                                            dev_priv->mm.interruptible);
1760                 if (ret)
1761                         break;
1762
1763                 if (time_after(jiffies, end)) {
1764                         ret = -EBUSY;
1765                         break;
1766                 }
1767         } while (1);
1768         trace_i915_ring_wait_end(ring);
1769         return ret;
1770 }
1771
1772 static int intel_wrap_ring_buffer(struct intel_engine_cs *ring)
1773 {
1774         uint32_t __iomem *virt;
1775         struct intel_ringbuffer *ringbuf = ring->buffer;
1776         int rem = ringbuf->size - ringbuf->tail;
1777
1778         if (ringbuf->space < rem) {
1779                 int ret = ring_wait_for_space(ring, rem);
1780                 if (ret)
1781                         return ret;
1782         }
1783
1784         virt = ringbuf->virtual_start + ringbuf->tail;
1785         rem /= 4;
1786         while (rem--)
1787                 iowrite32(MI_NOOP, virt++);
1788
1789         ringbuf->tail = 0;
1790         ringbuf->space = ring_space(ringbuf);
1791
1792         return 0;
1793 }
1794
1795 int intel_ring_idle(struct intel_engine_cs *ring)
1796 {
1797         u32 seqno;
1798         int ret;
1799
1800         /* We need to add any requests required to flush the objects and ring */
1801         if (ring->outstanding_lazy_seqno) {
1802                 ret = i915_add_request(ring, NULL);
1803                 if (ret)
1804                         return ret;
1805         }
1806
1807         /* Wait upon the last request to be completed */
1808         if (list_empty(&ring->request_list))
1809                 return 0;
1810
1811         seqno = list_entry(ring->request_list.prev,
1812                            struct drm_i915_gem_request,
1813                            list)->seqno;
1814
1815         return i915_wait_seqno(ring, seqno);
1816 }
1817
1818 static int
1819 intel_ring_alloc_seqno(struct intel_engine_cs *ring)
1820 {
1821         if (ring->outstanding_lazy_seqno)
1822                 return 0;
1823
1824         if (ring->preallocated_lazy_request == NULL) {
1825                 struct drm_i915_gem_request *request;
1826
1827                 request = kmalloc(sizeof(*request), GFP_KERNEL);
1828                 if (request == NULL)
1829                         return -ENOMEM;
1830
1831                 ring->preallocated_lazy_request = request;
1832         }
1833
1834         return i915_gem_get_seqno(ring->dev, &ring->outstanding_lazy_seqno);
1835 }
1836
1837 static int __intel_ring_prepare(struct intel_engine_cs *ring,
1838                                 int bytes)
1839 {
1840         struct intel_ringbuffer *ringbuf = ring->buffer;
1841         int ret;
1842
1843         if (unlikely(ringbuf->tail + bytes > ringbuf->effective_size)) {
1844                 ret = intel_wrap_ring_buffer(ring);
1845                 if (unlikely(ret))
1846                         return ret;
1847         }
1848
1849         if (unlikely(ringbuf->space < bytes)) {
1850                 ret = ring_wait_for_space(ring, bytes);
1851                 if (unlikely(ret))
1852                         return ret;
1853         }
1854
1855         return 0;
1856 }
1857
1858 int intel_ring_begin(struct intel_engine_cs *ring,
1859                      int num_dwords)
1860 {
1861         struct drm_i915_private *dev_priv = ring->dev->dev_private;
1862         int ret;
1863
1864         ret = i915_gem_check_wedge(&dev_priv->gpu_error,
1865                                    dev_priv->mm.interruptible);
1866         if (ret)
1867                 return ret;
1868
1869         ret = __intel_ring_prepare(ring, num_dwords * sizeof(uint32_t));
1870         if (ret)
1871                 return ret;
1872
1873         /* Preallocate the olr before touching the ring */
1874         ret = intel_ring_alloc_seqno(ring);
1875         if (ret)
1876                 return ret;
1877
1878         ring->buffer->space -= num_dwords * sizeof(uint32_t);
1879         return 0;
1880 }
1881
1882 /* Align the ring tail to a cacheline boundary */
1883 int intel_ring_cacheline_align(struct intel_engine_cs *ring)
1884 {
1885         int num_dwords = (ring->buffer->tail & (CACHELINE_BYTES - 1)) / sizeof(uint32_t);
1886         int ret;
1887
1888         if (num_dwords == 0)
1889                 return 0;
1890
1891         num_dwords = CACHELINE_BYTES / sizeof(uint32_t) - num_dwords;
1892         ret = intel_ring_begin(ring, num_dwords);
1893         if (ret)
1894                 return ret;
1895
1896         while (num_dwords--)
1897                 intel_ring_emit(ring, MI_NOOP);
1898
1899         intel_ring_advance(ring);
1900
1901         return 0;
1902 }
1903
1904 void intel_ring_init_seqno(struct intel_engine_cs *ring, u32 seqno)
1905 {
1906         struct drm_device *dev = ring->dev;
1907         struct drm_i915_private *dev_priv = dev->dev_private;
1908
1909         BUG_ON(ring->outstanding_lazy_seqno);
1910
1911         if (INTEL_INFO(dev)->gen == 6 || INTEL_INFO(dev)->gen == 7) {
1912                 I915_WRITE(RING_SYNC_0(ring->mmio_base), 0);
1913                 I915_WRITE(RING_SYNC_1(ring->mmio_base), 0);
1914                 if (HAS_VEBOX(dev))
1915                         I915_WRITE(RING_SYNC_2(ring->mmio_base), 0);
1916         }
1917
1918         ring->set_seqno(ring, seqno);
1919         ring->hangcheck.seqno = seqno;
1920 }
1921
1922 static void gen6_bsd_ring_write_tail(struct intel_engine_cs *ring,
1923                                      u32 value)
1924 {
1925         struct drm_i915_private *dev_priv = ring->dev->dev_private;
1926
1927        /* Every tail move must follow the sequence below */
1928
1929         /* Disable notification that the ring is IDLE. The GT
1930          * will then assume that it is busy and bring it out of rc6.
1931          */
1932         I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL,
1933                    _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1934
1935         /* Clear the context id. Here be magic! */
1936         I915_WRITE64(GEN6_BSD_RNCID, 0x0);
1937
1938         /* Wait for the ring not to be idle, i.e. for it to wake up. */
1939         if (wait_for((I915_READ(GEN6_BSD_SLEEP_PSMI_CONTROL) &
1940                       GEN6_BSD_SLEEP_INDICATOR) == 0,
1941                      50))
1942                 DRM_ERROR("timed out waiting for the BSD ring to wake up\n");
1943
1944         /* Now that the ring is fully powered up, update the tail */
1945         I915_WRITE_TAIL(ring, value);
1946         POSTING_READ(RING_TAIL(ring->mmio_base));
1947
1948         /* Let the ring send IDLE messages to the GT again,
1949          * and so let it sleep to conserve power when idle.
1950          */
1951         I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL,
1952                    _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1953 }
1954
1955 static int gen6_bsd_ring_flush(struct intel_engine_cs *ring,
1956                                u32 invalidate, u32 flush)
1957 {
1958         uint32_t cmd;
1959         int ret;
1960
1961         ret = intel_ring_begin(ring, 4);
1962         if (ret)
1963                 return ret;
1964
1965         cmd = MI_FLUSH_DW;
1966         if (INTEL_INFO(ring->dev)->gen >= 8)
1967                 cmd += 1;
1968         /*
1969          * Bspec vol 1c.5 - video engine command streamer:
1970          * "If ENABLED, all TLBs will be invalidated once the flush
1971          * operation is complete. This bit is only valid when the
1972          * Post-Sync Operation field is a value of 1h or 3h."
1973          */
1974         if (invalidate & I915_GEM_GPU_DOMAINS)
1975                 cmd |= MI_INVALIDATE_TLB | MI_INVALIDATE_BSD |
1976                         MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1977         intel_ring_emit(ring, cmd);
1978         intel_ring_emit(ring, I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT);
1979         if (INTEL_INFO(ring->dev)->gen >= 8) {
1980                 intel_ring_emit(ring, 0); /* upper addr */
1981                 intel_ring_emit(ring, 0); /* value */
1982         } else  {
1983                 intel_ring_emit(ring, 0);
1984                 intel_ring_emit(ring, MI_NOOP);
1985         }
1986         intel_ring_advance(ring);
1987         return 0;
1988 }
1989
1990 static int
1991 gen8_ring_dispatch_execbuffer(struct intel_engine_cs *ring,
1992                               u64 offset, u32 len,
1993                               unsigned flags)
1994 {
1995         struct drm_i915_private *dev_priv = ring->dev->dev_private;
1996         bool ppgtt = dev_priv->mm.aliasing_ppgtt != NULL &&
1997                 !(flags & I915_DISPATCH_SECURE);
1998         int ret;
1999
2000         ret = intel_ring_begin(ring, 4);
2001         if (ret)
2002                 return ret;
2003
2004         /* FIXME(BDW): Address space and security selectors. */
2005         intel_ring_emit(ring, MI_BATCH_BUFFER_START_GEN8 | (ppgtt<<8));
2006         intel_ring_emit(ring, lower_32_bits(offset));
2007         intel_ring_emit(ring, upper_32_bits(offset));
2008         intel_ring_emit(ring, MI_NOOP);
2009         intel_ring_advance(ring);
2010
2011         return 0;
2012 }
2013
2014 static int
2015 hsw_ring_dispatch_execbuffer(struct intel_engine_cs *ring,
2016                               u64 offset, u32 len,
2017                               unsigned flags)
2018 {
2019         int ret;
2020
2021         ret = intel_ring_begin(ring, 2);
2022         if (ret)
2023                 return ret;
2024
2025         intel_ring_emit(ring,
2026                         MI_BATCH_BUFFER_START | MI_BATCH_PPGTT_HSW |
2027                         (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE_HSW));
2028         /* bit0-7 is the length on GEN6+ */
2029         intel_ring_emit(ring, offset);
2030         intel_ring_advance(ring);
2031
2032         return 0;
2033 }
2034
2035 static int
2036 gen6_ring_dispatch_execbuffer(struct intel_engine_cs *ring,
2037                               u64 offset, u32 len,
2038                               unsigned flags)
2039 {
2040         int ret;
2041
2042         ret = intel_ring_begin(ring, 2);
2043         if (ret)
2044                 return ret;
2045
2046         intel_ring_emit(ring,
2047                         MI_BATCH_BUFFER_START |
2048                         (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE_I965));
2049         /* bit0-7 is the length on GEN6+ */
2050         intel_ring_emit(ring, offset);
2051         intel_ring_advance(ring);
2052
2053         return 0;
2054 }
2055
2056 /* Blitter support (SandyBridge+) */
2057
2058 static int gen6_ring_flush(struct intel_engine_cs *ring,
2059                            u32 invalidate, u32 flush)
2060 {
2061         struct drm_device *dev = ring->dev;
2062         uint32_t cmd;
2063         int ret;
2064
2065         ret = intel_ring_begin(ring, 4);
2066         if (ret)
2067                 return ret;
2068
2069         cmd = MI_FLUSH_DW;
2070         if (INTEL_INFO(ring->dev)->gen >= 8)
2071                 cmd += 1;
2072         /*
2073          * Bspec vol 1c.3 - blitter engine command streamer:
2074          * "If ENABLED, all TLBs will be invalidated once the flush
2075          * operation is complete. This bit is only valid when the
2076          * Post-Sync Operation field is a value of 1h or 3h."
2077          */
2078         if (invalidate & I915_GEM_DOMAIN_RENDER)
2079                 cmd |= MI_INVALIDATE_TLB | MI_FLUSH_DW_STORE_INDEX |
2080                         MI_FLUSH_DW_OP_STOREDW;
2081         intel_ring_emit(ring, cmd);
2082         intel_ring_emit(ring, I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT);
2083         if (INTEL_INFO(ring->dev)->gen >= 8) {
2084                 intel_ring_emit(ring, 0); /* upper addr */
2085                 intel_ring_emit(ring, 0); /* value */
2086         } else  {
2087                 intel_ring_emit(ring, 0);
2088                 intel_ring_emit(ring, MI_NOOP);
2089         }
2090         intel_ring_advance(ring);
2091
2092         if (IS_GEN7(dev) && !invalidate && flush)
2093                 return gen7_ring_fbc_flush(ring, FBC_REND_CACHE_CLEAN);
2094
2095         return 0;
2096 }
2097
2098 int intel_init_render_ring_buffer(struct drm_device *dev)
2099 {
2100         struct drm_i915_private *dev_priv = dev->dev_private;
2101         struct intel_engine_cs *ring = &dev_priv->ring[RCS];
2102         struct drm_i915_gem_object *obj;
2103         int ret;
2104
2105         ring->name = "render ring";
2106         ring->id = RCS;
2107         ring->mmio_base = RENDER_RING_BASE;
2108
2109         if (INTEL_INFO(dev)->gen >= 8) {
2110                 if (i915_semaphore_is_enabled(dev)) {
2111                         obj = i915_gem_alloc_object(dev, 4096);
2112                         if (obj == NULL) {
2113                                 DRM_ERROR("Failed to allocate semaphore bo. Disabling semaphores\n");
2114                                 i915.semaphores = 0;
2115                         } else {
2116                                 i915_gem_object_set_cache_level(obj, I915_CACHE_LLC);
2117                                 ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_NONBLOCK);
2118                                 if (ret != 0) {
2119                                         drm_gem_object_unreference(&obj->base);
2120                                         DRM_ERROR("Failed to pin semaphore bo. Disabling semaphores\n");
2121                                         i915.semaphores = 0;
2122                                 } else
2123                                         dev_priv->semaphore_obj = obj;
2124                         }
2125                 }
2126                 ring->add_request = gen6_add_request;
2127                 ring->flush = gen8_render_ring_flush;
2128                 ring->irq_get = gen8_ring_get_irq;
2129                 ring->irq_put = gen8_ring_put_irq;
2130                 ring->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
2131                 ring->get_seqno = gen6_ring_get_seqno;
2132                 ring->set_seqno = ring_set_seqno;
2133                 if (i915_semaphore_is_enabled(dev)) {
2134                         WARN_ON(!dev_priv->semaphore_obj);
2135                         ring->semaphore.sync_to = gen8_ring_sync;
2136                         ring->semaphore.signal = gen8_rcs_signal;
2137                         GEN8_RING_SEMAPHORE_INIT;
2138                 }
2139         } else if (INTEL_INFO(dev)->gen >= 6) {
2140                 ring->add_request = gen6_add_request;
2141                 ring->flush = gen7_render_ring_flush;
2142                 if (INTEL_INFO(dev)->gen == 6)
2143                         ring->flush = gen6_render_ring_flush;
2144                 ring->irq_get = gen6_ring_get_irq;
2145                 ring->irq_put = gen6_ring_put_irq;
2146                 ring->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
2147                 ring->get_seqno = gen6_ring_get_seqno;
2148                 ring->set_seqno = ring_set_seqno;
2149                 if (i915_semaphore_is_enabled(dev)) {
2150                         ring->semaphore.sync_to = gen6_ring_sync;
2151                         ring->semaphore.signal = gen6_signal;
2152                         /*
2153                          * The current semaphore is only applied on pre-gen8
2154                          * platform.  And there is no VCS2 ring on the pre-gen8
2155                          * platform. So the semaphore between RCS and VCS2 is
2156                          * initialized as INVALID.  Gen8 will initialize the
2157                          * sema between VCS2 and RCS later.
2158                          */
2159                         ring->semaphore.mbox.wait[RCS] = MI_SEMAPHORE_SYNC_INVALID;
2160                         ring->semaphore.mbox.wait[VCS] = MI_SEMAPHORE_SYNC_RV;
2161                         ring->semaphore.mbox.wait[BCS] = MI_SEMAPHORE_SYNC_RB;
2162                         ring->semaphore.mbox.wait[VECS] = MI_SEMAPHORE_SYNC_RVE;
2163                         ring->semaphore.mbox.wait[VCS2] = MI_SEMAPHORE_SYNC_INVALID;
2164                         ring->semaphore.mbox.signal[RCS] = GEN6_NOSYNC;
2165                         ring->semaphore.mbox.signal[VCS] = GEN6_VRSYNC;
2166                         ring->semaphore.mbox.signal[BCS] = GEN6_BRSYNC;
2167                         ring->semaphore.mbox.signal[VECS] = GEN6_VERSYNC;
2168                         ring->semaphore.mbox.signal[VCS2] = GEN6_NOSYNC;
2169                 }
2170         } else if (IS_GEN5(dev)) {
2171                 ring->add_request = pc_render_add_request;
2172                 ring->flush = gen4_render_ring_flush;
2173                 ring->get_seqno = pc_render_get_seqno;
2174                 ring->set_seqno = pc_render_set_seqno;
2175                 ring->irq_get = gen5_ring_get_irq;
2176                 ring->irq_put = gen5_ring_put_irq;
2177                 ring->irq_enable_mask = GT_RENDER_USER_INTERRUPT |
2178                                         GT_RENDER_PIPECTL_NOTIFY_INTERRUPT;
2179         } else {
2180                 ring->add_request = i9xx_add_request;
2181                 if (INTEL_INFO(dev)->gen < 4)
2182                         ring->flush = gen2_render_ring_flush;
2183                 else
2184                         ring->flush = gen4_render_ring_flush;
2185                 ring->get_seqno = ring_get_seqno;
2186                 ring->set_seqno = ring_set_seqno;
2187                 if (IS_GEN2(dev)) {
2188                         ring->irq_get = i8xx_ring_get_irq;
2189                         ring->irq_put = i8xx_ring_put_irq;
2190                 } else {
2191                         ring->irq_get = i9xx_ring_get_irq;
2192                         ring->irq_put = i9xx_ring_put_irq;
2193                 }
2194                 ring->irq_enable_mask = I915_USER_INTERRUPT;
2195         }
2196         ring->write_tail = ring_write_tail;
2197
2198         if (IS_HASWELL(dev))
2199                 ring->dispatch_execbuffer = hsw_ring_dispatch_execbuffer;
2200         else if (IS_GEN8(dev))
2201                 ring->dispatch_execbuffer = gen8_ring_dispatch_execbuffer;
2202         else if (INTEL_INFO(dev)->gen >= 6)
2203                 ring->dispatch_execbuffer = gen6_ring_dispatch_execbuffer;
2204         else if (INTEL_INFO(dev)->gen >= 4)
2205                 ring->dispatch_execbuffer = i965_dispatch_execbuffer;
2206         else if (IS_I830(dev) || IS_845G(dev))
2207                 ring->dispatch_execbuffer = i830_dispatch_execbuffer;
2208         else
2209                 ring->dispatch_execbuffer = i915_dispatch_execbuffer;
2210         ring->init = init_render_ring;
2211         ring->cleanup = render_ring_cleanup;
2212
2213         /* Workaround batchbuffer to combat CS tlb bug. */
2214         if (HAS_BROKEN_CS_TLB(dev)) {
2215                 obj = i915_gem_alloc_object(dev, I830_WA_SIZE);
2216                 if (obj == NULL) {
2217                         DRM_ERROR("Failed to allocate batch bo\n");
2218                         return -ENOMEM;
2219                 }
2220
2221                 ret = i915_gem_obj_ggtt_pin(obj, 0, 0);
2222                 if (ret != 0) {
2223                         drm_gem_object_unreference(&obj->base);
2224                         DRM_ERROR("Failed to ping batch bo\n");
2225                         return ret;
2226                 }
2227
2228                 ring->scratch.obj = obj;
2229                 ring->scratch.gtt_offset = i915_gem_obj_ggtt_offset(obj);
2230         }
2231
2232         return intel_init_ring_buffer(dev, ring);
2233 }
2234
2235 int intel_render_ring_init_dri(struct drm_device *dev, u64 start, u32 size)
2236 {
2237         struct drm_i915_private *dev_priv = dev->dev_private;
2238         struct intel_engine_cs *ring = &dev_priv->ring[RCS];
2239         struct intel_ringbuffer *ringbuf = ring->buffer;
2240         int ret;
2241
2242         if (ringbuf == NULL) {
2243                 ringbuf = kzalloc(sizeof(*ringbuf), GFP_KERNEL);
2244                 if (!ringbuf)
2245                         return -ENOMEM;
2246                 ring->buffer = ringbuf;
2247         }
2248
2249         ring->name = "render ring";
2250         ring->id = RCS;
2251         ring->mmio_base = RENDER_RING_BASE;
2252
2253         if (INTEL_INFO(dev)->gen >= 6) {
2254                 /* non-kms not supported on gen6+ */
2255                 ret = -ENODEV;
2256                 goto err_ringbuf;
2257         }
2258
2259         /* Note: gem is not supported on gen5/ilk without kms (the corresponding
2260          * gem_init ioctl returns with -ENODEV). Hence we do not need to set up
2261          * the special gen5 functions. */
2262         ring->add_request = i9xx_add_request;
2263         if (INTEL_INFO(dev)->gen < 4)
2264                 ring->flush = gen2_render_ring_flush;
2265         else
2266                 ring->flush = gen4_render_ring_flush;
2267         ring->get_seqno = ring_get_seqno;
2268         ring->set_seqno = ring_set_seqno;
2269         if (IS_GEN2(dev)) {
2270                 ring->irq_get = i8xx_ring_get_irq;
2271                 ring->irq_put = i8xx_ring_put_irq;
2272         } else {
2273                 ring->irq_get = i9xx_ring_get_irq;
2274                 ring->irq_put = i9xx_ring_put_irq;
2275         }
2276         ring->irq_enable_mask = I915_USER_INTERRUPT;
2277         ring->write_tail = ring_write_tail;
2278         if (INTEL_INFO(dev)->gen >= 4)
2279                 ring->dispatch_execbuffer = i965_dispatch_execbuffer;
2280         else if (IS_I830(dev) || IS_845G(dev))
2281                 ring->dispatch_execbuffer = i830_dispatch_execbuffer;
2282         else
2283                 ring->dispatch_execbuffer = i915_dispatch_execbuffer;
2284         ring->init = init_render_ring;
2285         ring->cleanup = render_ring_cleanup;
2286
2287         ring->dev = dev;
2288         INIT_LIST_HEAD(&ring->active_list);
2289         INIT_LIST_HEAD(&ring->request_list);
2290
2291         ringbuf->size = size;
2292         ringbuf->effective_size = ringbuf->size;
2293         if (IS_I830(ring->dev) || IS_845G(ring->dev))
2294                 ringbuf->effective_size -= 2 * CACHELINE_BYTES;
2295
2296         ringbuf->virtual_start = ioremap_wc(start, size);
2297         if (ringbuf->virtual_start == NULL) {
2298                 DRM_ERROR("can not ioremap virtual address for"
2299                           " ring buffer\n");
2300                 ret = -ENOMEM;
2301                 goto err_ringbuf;
2302         }
2303
2304         if (!I915_NEED_GFX_HWS(dev)) {
2305                 ret = init_phys_status_page(ring);
2306                 if (ret)
2307                         goto err_vstart;
2308         }
2309
2310         return 0;
2311
2312 err_vstart:
2313         iounmap(ringbuf->virtual_start);
2314 err_ringbuf:
2315         kfree(ringbuf);
2316         ring->buffer = NULL;
2317         return ret;
2318 }
2319
2320 int intel_init_bsd_ring_buffer(struct drm_device *dev)
2321 {
2322         struct drm_i915_private *dev_priv = dev->dev_private;
2323         struct intel_engine_cs *ring = &dev_priv->ring[VCS];
2324
2325         ring->name = "bsd ring";
2326         ring->id = VCS;
2327
2328         ring->write_tail = ring_write_tail;
2329         if (INTEL_INFO(dev)->gen >= 6) {
2330                 ring->mmio_base = GEN6_BSD_RING_BASE;
2331                 /* gen6 bsd needs a special wa for tail updates */
2332                 if (IS_GEN6(dev))
2333                         ring->write_tail = gen6_bsd_ring_write_tail;
2334                 ring->flush = gen6_bsd_ring_flush;
2335                 ring->add_request = gen6_add_request;
2336                 ring->get_seqno = gen6_ring_get_seqno;
2337                 ring->set_seqno = ring_set_seqno;
2338                 if (INTEL_INFO(dev)->gen >= 8) {
2339                         ring->irq_enable_mask =
2340                                 GT_RENDER_USER_INTERRUPT << GEN8_VCS1_IRQ_SHIFT;
2341                         ring->irq_get = gen8_ring_get_irq;
2342                         ring->irq_put = gen8_ring_put_irq;
2343                         ring->dispatch_execbuffer =
2344                                 gen8_ring_dispatch_execbuffer;
2345                         if (i915_semaphore_is_enabled(dev)) {
2346                                 ring->semaphore.sync_to = gen8_ring_sync;
2347                                 ring->semaphore.signal = gen8_xcs_signal;
2348                                 GEN8_RING_SEMAPHORE_INIT;
2349                         }
2350                 } else {
2351                         ring->irq_enable_mask = GT_BSD_USER_INTERRUPT;
2352                         ring->irq_get = gen6_ring_get_irq;
2353                         ring->irq_put = gen6_ring_put_irq;
2354                         ring->dispatch_execbuffer =
2355                                 gen6_ring_dispatch_execbuffer;
2356                         if (i915_semaphore_is_enabled(dev)) {
2357                                 ring->semaphore.sync_to = gen6_ring_sync;
2358                                 ring->semaphore.signal = gen6_signal;
2359                                 ring->semaphore.mbox.wait[RCS] = MI_SEMAPHORE_SYNC_VR;
2360                                 ring->semaphore.mbox.wait[VCS] = MI_SEMAPHORE_SYNC_INVALID;
2361                                 ring->semaphore.mbox.wait[BCS] = MI_SEMAPHORE_SYNC_VB;
2362                                 ring->semaphore.mbox.wait[VECS] = MI_SEMAPHORE_SYNC_VVE;
2363                                 ring->semaphore.mbox.wait[VCS2] = MI_SEMAPHORE_SYNC_INVALID;
2364                                 ring->semaphore.mbox.signal[RCS] = GEN6_RVSYNC;
2365                                 ring->semaphore.mbox.signal[VCS] = GEN6_NOSYNC;
2366                                 ring->semaphore.mbox.signal[BCS] = GEN6_BVSYNC;
2367                                 ring->semaphore.mbox.signal[VECS] = GEN6_VEVSYNC;
2368                                 ring->semaphore.mbox.signal[VCS2] = GEN6_NOSYNC;
2369                         }
2370                 }
2371         } else {
2372                 ring->mmio_base = BSD_RING_BASE;
2373                 ring->flush = bsd_ring_flush;
2374                 ring->add_request = i9xx_add_request;
2375                 ring->get_seqno = ring_get_seqno;
2376                 ring->set_seqno = ring_set_seqno;
2377                 if (IS_GEN5(dev)) {
2378                         ring->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
2379                         ring->irq_get = gen5_ring_get_irq;
2380                         ring->irq_put = gen5_ring_put_irq;
2381                 } else {
2382                         ring->irq_enable_mask = I915_BSD_USER_INTERRUPT;
2383                         ring->irq_get = i9xx_ring_get_irq;
2384                         ring->irq_put = i9xx_ring_put_irq;
2385                 }
2386                 ring->dispatch_execbuffer = i965_dispatch_execbuffer;
2387         }
2388         ring->init = init_ring_common;
2389
2390         return intel_init_ring_buffer(dev, ring);
2391 }
2392
2393 /**
2394  * Initialize the second BSD ring for Broadwell GT3.
2395  * It is noted that this only exists on Broadwell GT3.
2396  */
2397 int intel_init_bsd2_ring_buffer(struct drm_device *dev)
2398 {
2399         struct drm_i915_private *dev_priv = dev->dev_private;
2400         struct intel_engine_cs *ring = &dev_priv->ring[VCS2];
2401
2402         if ((INTEL_INFO(dev)->gen != 8)) {
2403                 DRM_ERROR("No dual-BSD ring on non-BDW machine\n");
2404                 return -EINVAL;
2405         }
2406
2407         ring->name = "bsd2 ring";
2408         ring->id = VCS2;
2409
2410         ring->write_tail = ring_write_tail;
2411         ring->mmio_base = GEN8_BSD2_RING_BASE;
2412         ring->flush = gen6_bsd_ring_flush;
2413         ring->add_request = gen6_add_request;
2414         ring->get_seqno = gen6_ring_get_seqno;
2415         ring->set_seqno = ring_set_seqno;
2416         ring->irq_enable_mask =
2417                         GT_RENDER_USER_INTERRUPT << GEN8_VCS2_IRQ_SHIFT;
2418         ring->irq_get = gen8_ring_get_irq;
2419         ring->irq_put = gen8_ring_put_irq;
2420         ring->dispatch_execbuffer =
2421                         gen8_ring_dispatch_execbuffer;
2422         if (i915_semaphore_is_enabled(dev)) {
2423                 ring->semaphore.sync_to = gen8_ring_sync;
2424                 ring->semaphore.signal = gen8_xcs_signal;
2425                 GEN8_RING_SEMAPHORE_INIT;
2426         }
2427         ring->init = init_ring_common;
2428
2429         return intel_init_ring_buffer(dev, ring);
2430 }
2431
2432 int intel_init_blt_ring_buffer(struct drm_device *dev)
2433 {
2434         struct drm_i915_private *dev_priv = dev->dev_private;
2435         struct intel_engine_cs *ring = &dev_priv->ring[BCS];
2436
2437         ring->name = "blitter ring";
2438         ring->id = BCS;
2439
2440         ring->mmio_base = BLT_RING_BASE;
2441         ring->write_tail = ring_write_tail;
2442         ring->flush = gen6_ring_flush;
2443         ring->add_request = gen6_add_request;
2444         ring->get_seqno = gen6_ring_get_seqno;
2445         ring->set_seqno = ring_set_seqno;
2446         if (INTEL_INFO(dev)->gen >= 8) {
2447                 ring->irq_enable_mask =
2448                         GT_RENDER_USER_INTERRUPT << GEN8_BCS_IRQ_SHIFT;
2449                 ring->irq_get = gen8_ring_get_irq;
2450                 ring->irq_put = gen8_ring_put_irq;
2451                 ring->dispatch_execbuffer = gen8_ring_dispatch_execbuffer;
2452                 if (i915_semaphore_is_enabled(dev)) {
2453                         ring->semaphore.sync_to = gen8_ring_sync;
2454                         ring->semaphore.signal = gen8_xcs_signal;
2455                         GEN8_RING_SEMAPHORE_INIT;
2456                 }
2457         } else {
2458                 ring->irq_enable_mask = GT_BLT_USER_INTERRUPT;
2459                 ring->irq_get = gen6_ring_get_irq;
2460                 ring->irq_put = gen6_ring_put_irq;
2461                 ring->dispatch_execbuffer = gen6_ring_dispatch_execbuffer;
2462                 if (i915_semaphore_is_enabled(dev)) {
2463                         ring->semaphore.signal = gen6_signal;
2464                         ring->semaphore.sync_to = gen6_ring_sync;
2465                         /*
2466                          * The current semaphore is only applied on pre-gen8
2467                          * platform.  And there is no VCS2 ring on the pre-gen8
2468                          * platform. So the semaphore between BCS and VCS2 is
2469                          * initialized as INVALID.  Gen8 will initialize the
2470                          * sema between BCS and VCS2 later.
2471                          */
2472                         ring->semaphore.mbox.wait[RCS] = MI_SEMAPHORE_SYNC_BR;
2473                         ring->semaphore.mbox.wait[VCS] = MI_SEMAPHORE_SYNC_BV;
2474                         ring->semaphore.mbox.wait[BCS] = MI_SEMAPHORE_SYNC_INVALID;
2475                         ring->semaphore.mbox.wait[VECS] = MI_SEMAPHORE_SYNC_BVE;
2476                         ring->semaphore.mbox.wait[VCS2] = MI_SEMAPHORE_SYNC_INVALID;
2477                         ring->semaphore.mbox.signal[RCS] = GEN6_RBSYNC;
2478                         ring->semaphore.mbox.signal[VCS] = GEN6_VBSYNC;
2479                         ring->semaphore.mbox.signal[BCS] = GEN6_NOSYNC;
2480                         ring->semaphore.mbox.signal[VECS] = GEN6_VEBSYNC;
2481                         ring->semaphore.mbox.signal[VCS2] = GEN6_NOSYNC;
2482                 }
2483         }
2484         ring->init = init_ring_common;
2485
2486         return intel_init_ring_buffer(dev, ring);
2487 }
2488
2489 int intel_init_vebox_ring_buffer(struct drm_device *dev)
2490 {
2491         struct drm_i915_private *dev_priv = dev->dev_private;
2492         struct intel_engine_cs *ring = &dev_priv->ring[VECS];
2493
2494         ring->name = "video enhancement ring";
2495         ring->id = VECS;
2496
2497         ring->mmio_base = VEBOX_RING_BASE;
2498         ring->write_tail = ring_write_tail;
2499         ring->flush = gen6_ring_flush;
2500         ring->add_request = gen6_add_request;
2501         ring->get_seqno = gen6_ring_get_seqno;
2502         ring->set_seqno = ring_set_seqno;
2503
2504         if (INTEL_INFO(dev)->gen >= 8) {
2505                 ring->irq_enable_mask =
2506                         GT_RENDER_USER_INTERRUPT << GEN8_VECS_IRQ_SHIFT;
2507                 ring->irq_get = gen8_ring_get_irq;
2508                 ring->irq_put = gen8_ring_put_irq;
2509                 ring->dispatch_execbuffer = gen8_ring_dispatch_execbuffer;
2510                 if (i915_semaphore_is_enabled(dev)) {
2511                         ring->semaphore.sync_to = gen8_ring_sync;
2512                         ring->semaphore.signal = gen8_xcs_signal;
2513                         GEN8_RING_SEMAPHORE_INIT;
2514                 }
2515         } else {
2516                 ring->irq_enable_mask = PM_VEBOX_USER_INTERRUPT;
2517                 ring->irq_get = hsw_vebox_get_irq;
2518                 ring->irq_put = hsw_vebox_put_irq;
2519                 ring->dispatch_execbuffer = gen6_ring_dispatch_execbuffer;
2520                 if (i915_semaphore_is_enabled(dev)) {
2521                         ring->semaphore.sync_to = gen6_ring_sync;
2522                         ring->semaphore.signal = gen6_signal;
2523                         ring->semaphore.mbox.wait[RCS] = MI_SEMAPHORE_SYNC_VER;
2524                         ring->semaphore.mbox.wait[VCS] = MI_SEMAPHORE_SYNC_VEV;
2525                         ring->semaphore.mbox.wait[BCS] = MI_SEMAPHORE_SYNC_VEB;
2526                         ring->semaphore.mbox.wait[VECS] = MI_SEMAPHORE_SYNC_INVALID;
2527                         ring->semaphore.mbox.wait[VCS2] = MI_SEMAPHORE_SYNC_INVALID;
2528                         ring->semaphore.mbox.signal[RCS] = GEN6_RVESYNC;
2529                         ring->semaphore.mbox.signal[VCS] = GEN6_VVESYNC;
2530                         ring->semaphore.mbox.signal[BCS] = GEN6_BVESYNC;
2531                         ring->semaphore.mbox.signal[VECS] = GEN6_NOSYNC;
2532                         ring->semaphore.mbox.signal[VCS2] = GEN6_NOSYNC;
2533                 }
2534         }
2535         ring->init = init_ring_common;
2536
2537         return intel_init_ring_buffer(dev, ring);
2538 }
2539
2540 int
2541 intel_ring_flush_all_caches(struct intel_engine_cs *ring)
2542 {
2543         int ret;
2544
2545         if (!ring->gpu_caches_dirty)
2546                 return 0;
2547
2548         ret = ring->flush(ring, 0, I915_GEM_GPU_DOMAINS);
2549         if (ret)
2550                 return ret;
2551
2552         trace_i915_gem_ring_flush(ring, 0, I915_GEM_GPU_DOMAINS);
2553
2554         ring->gpu_caches_dirty = false;
2555         return 0;
2556 }
2557
2558 int
2559 intel_ring_invalidate_all_caches(struct intel_engine_cs *ring)
2560 {
2561         uint32_t flush_domains;
2562         int ret;
2563
2564         flush_domains = 0;
2565         if (ring->gpu_caches_dirty)
2566                 flush_domains = I915_GEM_GPU_DOMAINS;
2567
2568         ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, flush_domains);
2569         if (ret)
2570                 return ret;
2571
2572         trace_i915_gem_ring_flush(ring, I915_GEM_GPU_DOMAINS, flush_domains);
2573
2574         ring->gpu_caches_dirty = false;
2575         return 0;
2576 }
2577
2578 void
2579 intel_stop_ring_buffer(struct intel_engine_cs *ring)
2580 {
2581         int ret;
2582
2583         if (!intel_ring_initialized(ring))
2584                 return;
2585
2586         ret = intel_ring_idle(ring);
2587         if (ret && !i915_reset_in_progress(&to_i915(ring->dev)->gpu_error))
2588                 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
2589                           ring->name, ret);
2590
2591         stop_ring(ring);
2592 }