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 bool
  37 intel_ring_initialized(struct intel_engine_cs *ring)
  38 {
  39         struct drm_device *dev = ring->dev;
  40
  41         if (!dev)
  42                 return false;
  43
  44         if (i915.enable_execlists) {
  45                 struct intel_context *dctx = ring->default_context;
  46                 struct intel_ringbuffer *ringbuf = dctx->engine[ring->id].ringbuf;
  47
  48                 return ringbuf->obj;
  49         } else
  50                 return ring->buffer && ring->buffer->obj;
  51 }
  52
  53 int __intel_ring_space(int head, int tail, int size)
  54 {
  55         int space = head - tail;
  56         if (space <= 0)
  57                 space += size;
  58         return space - I915_RING_FREE_SPACE;
  59 }
  60
  61 void intel_ring_update_space(struct intel_ringbuffer *ringbuf)
  62 {
  63         if (ringbuf->last_retired_head != -1) {
  64                 ringbuf->head = ringbuf->last_retired_head;
  65                 ringbuf->last_retired_head = -1;
  66         }
  67
  68         ringbuf->space = __intel_ring_space(ringbuf->head & HEAD_ADDR,
  69                                             ringbuf->tail, ringbuf->size);
  70 }
  71
  72 int intel_ring_space(struct intel_ringbuffer *ringbuf)
  73 {
  74         intel_ring_update_space(ringbuf);
  75         return ringbuf->space;
  76 }
  77
  78 bool intel_ring_stopped(struct intel_engine_cs *ring)
  79 {
  80         struct drm_i915_private *dev_priv = ring->dev->dev_private;
  81         return dev_priv->gpu_error.stop_rings & intel_ring_flag(ring);
  82 }
  83
  84 void __intel_ring_advance(struct intel_engine_cs *ring)
  85 {
  86         struct intel_ringbuffer *ringbuf = ring->buffer;
  87         ringbuf->tail &= ringbuf->size - 1;
  88         if (intel_ring_stopped(ring))
  89                 return;
  90         ring->write_tail(ring, ringbuf->tail);
  91 }
  92
  93 static int
  94 gen2_render_ring_flush(struct intel_engine_cs *ring,
  95                        u32      invalidate_domains,
  96                        u32      flush_domains)
  97 {
  98         u32 cmd;
  99         int ret;
 100
 101         cmd = MI_FLUSH;
 102         if (((invalidate_domains|flush_domains) & I915_GEM_DOMAIN_RENDER) == 0)
 103                 cmd |= MI_NO_WRITE_FLUSH;
 104
 105         if (invalidate_domains & I915_GEM_DOMAIN_SAMPLER)
 106                 cmd |= MI_READ_FLUSH;
 107
 108         ret = intel_ring_begin(ring, 2);
 109         if (ret)
 110                 return ret;
 111
 112         intel_ring_emit(ring, cmd);
 113         intel_ring_emit(ring, MI_NOOP);
 114         intel_ring_advance(ring);
 115
 116         return 0;
 117 }
 118
 119 static int
 120 gen4_render_ring_flush(struct intel_engine_cs *ring,
 121                        u32      invalidate_domains,
 122                        u32      flush_domains)
 123 {
 124         struct drm_device *dev = ring->dev;
 125         u32 cmd;
 126         int ret;
 127
 128         /*
 129          * read/write caches:
 130          *
 131          * I915_GEM_DOMAIN_RENDER is always invalidated, but is
 132          * only flushed if MI_NO_WRITE_FLUSH is unset.  On 965, it is
 133          * also flushed at 2d versus 3d pipeline switches.
 134          *
 135          * read-only caches:
 136          *
 137          * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
 138          * MI_READ_FLUSH is set, and is always flushed on 965.
 139          *
 140          * I915_GEM_DOMAIN_COMMAND may not exist?
 141          *
 142          * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
 143          * invalidated when MI_EXE_FLUSH is set.
 144          *
 145          * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
 146          * invalidated with every MI_FLUSH.
 147          *
 148          * TLBs:
 149          *
 150          * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
 151          * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
 152          * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
 153          * are flushed at any MI_FLUSH.
 154          */
 155
 156         cmd = MI_FLUSH | MI_NO_WRITE_FLUSH;
 157         if ((invalidate_domains|flush_domains) & I915_GEM_DOMAIN_RENDER)
 158                 cmd &= ~MI_NO_WRITE_FLUSH;
 159         if (invalidate_domains & I915_GEM_DOMAIN_INSTRUCTION)
 160                 cmd |= MI_EXE_FLUSH;
 161
 162         if (invalidate_domains & I915_GEM_DOMAIN_COMMAND &&
 163             (IS_G4X(dev) || IS_GEN5(dev)))
 164                 cmd |= MI_INVALIDATE_ISP;
 165
 166         ret = intel_ring_begin(ring, 2);
 167         if (ret)
 168                 return ret;
 169
 170         intel_ring_emit(ring, cmd);
 171         intel_ring_emit(ring, MI_NOOP);
 172         intel_ring_advance(ring);
 173
 174         return 0;
 175 }
 176
 177 /**
 178  * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
 179  * implementing two workarounds on gen6.  From section 1.4.7.1
 180  * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
 181  *
 182  * [DevSNB-C+{W/A}] Before any depth stall flush (including those
 183  * produced by non-pipelined state commands), software needs to first
 184  * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
 185  * 0.
 186  *
 187  * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
 188  * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
 189  *
 190  * And the workaround for these two requires this workaround first:
 191  *
 192  * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
 193  * BEFORE the pipe-control with a post-sync op and no write-cache
 194  * flushes.
 195  *
 196  * And this last workaround is tricky because of the requirements on
 197  * that bit.  From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
 198  * volume 2 part 1:
 199  *
 200  *     "1 of the following must also be set:
 201  *      - Render Target Cache Flush Enable ([12] of DW1)
 202  *      - Depth Cache Flush Enable ([0] of DW1)
 203  *      - Stall at Pixel Scoreboard ([1] of DW1)
 204  *      - Depth Stall ([13] of DW1)
 205  *      - Post-Sync Operation ([13] of DW1)
 206  *      - Notify Enable ([8] of DW1)"
 207  *
 208  * The cache flushes require the workaround flush that triggered this
 209  * one, so we can't use it.  Depth stall would trigger the same.
 210  * Post-sync nonzero is what triggered this second workaround, so we
 211  * can't use that one either.  Notify enable is IRQs, which aren't
 212  * really our business.  That leaves only stall at scoreboard.
 213  */
 214 static int
 215 intel_emit_post_sync_nonzero_flush(struct intel_engine_cs *ring)
 216 {
 217         u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES;
 218         int ret;
 219
 220
 221         ret = intel_ring_begin(ring, 6);
 222         if (ret)
 223                 return ret;
 224
 225         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(5));
 226         intel_ring_emit(ring, PIPE_CONTROL_CS_STALL |
 227                         PIPE_CONTROL_STALL_AT_SCOREBOARD);
 228         intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT); /* address */
 229         intel_ring_emit(ring, 0); /* low dword */
 230         intel_ring_emit(ring, 0); /* high dword */
 231         intel_ring_emit(ring, MI_NOOP);
 232         intel_ring_advance(ring);
 233
 234         ret = intel_ring_begin(ring, 6);
 235         if (ret)
 236                 return ret;
 237
 238         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(5));
 239         intel_ring_emit(ring, PIPE_CONTROL_QW_WRITE);
 240         intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT); /* address */
 241         intel_ring_emit(ring, 0);
 242         intel_ring_emit(ring, 0);
 243         intel_ring_emit(ring, MI_NOOP);
 244         intel_ring_advance(ring);
 245
 246         return 0;
 247 }
 248
 249 static int
 250 gen6_render_ring_flush(struct intel_engine_cs *ring,
 251                          u32 invalidate_domains, u32 flush_domains)
 252 {
 253         u32 flags = 0;
 254         u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES;
 255         int ret;
 256
 257         /* Force SNB workarounds for PIPE_CONTROL flushes */
 258         ret = intel_emit_post_sync_nonzero_flush(ring);
 259         if (ret)
 260                 return ret;
 261
 262         /* Just flush everything.  Experiments have shown that reducing the
 263          * number of bits based on the write domains has little performance
 264          * impact.
 265          */
 266         if (flush_domains) {
 267                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
 268                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
 269                 /*
 270                  * Ensure that any following seqno writes only happen
 271                  * when the render cache is indeed flushed.
 272                  */
 273                 flags |= PIPE_CONTROL_CS_STALL;
 274         }
 275         if (invalidate_domains) {
 276                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
 277                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 278                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 279                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
 280                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
 281                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
 282                 /*
 283                  * TLB invalidate requires a post-sync write.
 284                  */
 285                 flags |= PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_CS_STALL;
 286         }
 287
 288         ret = intel_ring_begin(ring, 4);
 289         if (ret)
 290                 return ret;
 291
 292         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4));
 293         intel_ring_emit(ring, flags);
 294         intel_ring_emit(ring, scratch_addr | PIPE_CONTROL_GLOBAL_GTT);
 295         intel_ring_emit(ring, 0);
 296         intel_ring_advance(ring);
 297
 298         return 0;
 299 }
 300
 301 static int
 302 gen7_render_ring_cs_stall_wa(struct intel_engine_cs *ring)
 303 {
 304         int ret;
 305
 306         ret = intel_ring_begin(ring, 4);
 307         if (ret)
 308                 return ret;
 309
 310         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4));
 311         intel_ring_emit(ring, PIPE_CONTROL_CS_STALL |
 312                               PIPE_CONTROL_STALL_AT_SCOREBOARD);
 313         intel_ring_emit(ring, 0);
 314         intel_ring_emit(ring, 0);
 315         intel_ring_advance(ring);
 316
 317         return 0;
 318 }
 319
 320 static int gen7_ring_fbc_flush(struct intel_engine_cs *ring, u32 value)
 321 {
 322         int ret;
 323
 324         if (!ring->fbc_dirty)
 325                 return 0;
 326
 327         ret = intel_ring_begin(ring, 6);
 328         if (ret)
 329                 return ret;
 330         /* WaFbcNukeOn3DBlt:ivb/hsw */
 331         intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
 332         intel_ring_emit(ring, MSG_FBC_REND_STATE);
 333         intel_ring_emit(ring, value);
 334         intel_ring_emit(ring, MI_STORE_REGISTER_MEM(1) | MI_SRM_LRM_GLOBAL_GTT);
 335         intel_ring_emit(ring, MSG_FBC_REND_STATE);
 336         intel_ring_emit(ring, ring->scratch.gtt_offset + 256);
 337         intel_ring_advance(ring);
 338
 339         ring->fbc_dirty = false;
 340         return 0;
 341 }
 342
 343 static int
 344 gen7_render_ring_flush(struct intel_engine_cs *ring,
 345                        u32 invalidate_domains, u32 flush_domains)
 346 {
 347         u32 flags = 0;
 348         u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES;
 349         int ret;
 350
 351         /*
 352          * Ensure that any following seqno writes only happen when the render
 353          * cache is indeed flushed.
 354          *
 355          * Workaround: 4th PIPE_CONTROL command (except the ones with only
 356          * read-cache invalidate bits set) must have the CS_STALL bit set. We
 357          * don't try to be clever and just set it unconditionally.
 358          */
 359         flags |= PIPE_CONTROL_CS_STALL;
 360
 361         /* Just flush everything.  Experiments have shown that reducing the
 362          * number of bits based on the write domains has little performance
 363          * impact.
 364          */
 365         if (flush_domains) {
 366                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
 367                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
 368         }
 369         if (invalidate_domains) {
 370                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
 371                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 372                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 373                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
 374                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
 375                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
 376                 /*
 377                  * TLB invalidate requires a post-sync write.
 378                  */
 379                 flags |= PIPE_CONTROL_QW_WRITE;
 380                 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
 381
 382                 /* Workaround: we must issue a pipe_control with CS-stall bit
 383                  * set before a pipe_control command that has the state cache
 384                  * invalidate bit set. */
 385                 gen7_render_ring_cs_stall_wa(ring);
 386         }
 387
 388         ret = intel_ring_begin(ring, 4);
 389         if (ret)
 390                 return ret;
 391
 392         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4));
 393         intel_ring_emit(ring, flags);
 394         intel_ring_emit(ring, scratch_addr);
 395         intel_ring_emit(ring, 0);
 396         intel_ring_advance(ring);
 397
 398         if (!invalidate_domains && flush_domains)
 399                 return gen7_ring_fbc_flush(ring, FBC_REND_NUKE);
 400
 401         return 0;
 402 }
 403
 404 static int
 405 gen8_emit_pipe_control(struct intel_engine_cs *ring,
 406                        u32 flags, u32 scratch_addr)
 407 {
 408         int ret;
 409
 410         ret = intel_ring_begin(ring, 6);
 411         if (ret)
 412                 return ret;
 413
 414         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(6));
 415         intel_ring_emit(ring, flags);
 416         intel_ring_emit(ring, scratch_addr);
 417         intel_ring_emit(ring, 0);
 418         intel_ring_emit(ring, 0);
 419         intel_ring_emit(ring, 0);
 420         intel_ring_advance(ring);
 421
 422         return 0;
 423 }
 424
 425 static int
 426 gen8_render_ring_flush(struct intel_engine_cs *ring,
 427                        u32 invalidate_domains, u32 flush_domains)
 428 {
 429         u32 flags = 0;
 430         u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES;
 431         int ret;
 432
 433         flags |= PIPE_CONTROL_CS_STALL;
 434
 435         if (flush_domains) {
 436                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
 437                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
 438         }
 439         if (invalidate_domains) {
 440                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
 441                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 442                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 443                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
 444                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
 445                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
 446                 flags |= PIPE_CONTROL_QW_WRITE;
 447                 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
 448
 449                 /* WaCsStallBeforeStateCacheInvalidate:bdw,chv */
 450                 ret = gen8_emit_pipe_control(ring,
 451                                              PIPE_CONTROL_CS_STALL |
 452                                              PIPE_CONTROL_STALL_AT_SCOREBOARD,
 453                                              0);
 454                 if (ret)
 455                         return ret;
 456         }
 457
 458         ret = gen8_emit_pipe_control(ring, flags, scratch_addr);
 459         if (ret)
 460                 return ret;
 461
 462         if (!invalidate_domains && flush_domains)
 463                 return gen7_ring_fbc_flush(ring, FBC_REND_NUKE);
 464
 465         return 0;
 466 }
 467
 468 static void ring_write_tail(struct intel_engine_cs *ring,
 469                             u32 value)
 470 {
 471         struct drm_i915_private *dev_priv = ring->dev->dev_private;
 472         I915_WRITE_TAIL(ring, value);
 473 }
 474
 475 u64 intel_ring_get_active_head(struct intel_engine_cs *ring)
 476 {
 477         struct drm_i915_private *dev_priv = ring->dev->dev_private;
 478         u64 acthd;
 479
 480         if (INTEL_INFO(ring->dev)->gen >= 8)
 481                 acthd = I915_READ64_2x32(RING_ACTHD(ring->mmio_base),
 482                                          RING_ACTHD_UDW(ring->mmio_base));
 483         else if (INTEL_INFO(ring->dev)->gen >= 4)
 484                 acthd = I915_READ(RING_ACTHD(ring->mmio_base));
 485         else
 486                 acthd = I915_READ(ACTHD);
 487
 488         return acthd;
 489 }
 490
 491 static void ring_setup_phys_status_page(struct intel_engine_cs *ring)
 492 {
 493         struct drm_i915_private *dev_priv = ring->dev->dev_private;
 494         u32 addr;
 495
 496         addr = dev_priv->status_page_dmah->busaddr;
 497         if (INTEL_INFO(ring->dev)->gen >= 4)
 498                 addr |= (dev_priv->status_page_dmah->busaddr >> 28) & 0xf0;
 499         I915_WRITE(HWS_PGA, addr);
 500 }
 501
 502 static bool stop_ring(struct intel_engine_cs *ring)
 503 {
 504         struct drm_i915_private *dev_priv = to_i915(ring->dev);
 505
 506         if (!IS_GEN2(ring->dev)) {
 507                 I915_WRITE_MODE(ring, _MASKED_BIT_ENABLE(STOP_RING));
 508                 if (wait_for((I915_READ_MODE(ring) & MODE_IDLE) != 0, 1000)) {
 509                         DRM_ERROR("%s : timed out trying to stop ring\n", ring->name);
 510                         /* Sometimes we observe that the idle flag is not
 511                          * set even though the ring is empty. So double
 512                          * check before giving up.
 513                          */
 514                         if (I915_READ_HEAD(ring) != I915_READ_TAIL(ring))
 515                                 return false;
 516                 }
 517         }
 518
 519         I915_WRITE_CTL(ring, 0);
 520         I915_WRITE_HEAD(ring, 0);
 521         ring->write_tail(ring, 0);
 522
 523         if (!IS_GEN2(ring->dev)) {
 524                 (void)I915_READ_CTL(ring);
 525                 I915_WRITE_MODE(ring, _MASKED_BIT_DISABLE(STOP_RING));
 526         }
 527
 528         return (I915_READ_HEAD(ring) & HEAD_ADDR) == 0;
 529 }
 530
 531 static int init_ring_common(struct intel_engine_cs *ring)
 532 {
 533         struct drm_device *dev = ring->dev;
 534         struct drm_i915_private *dev_priv = dev->dev_private;
 535         struct intel_ringbuffer *ringbuf = ring->buffer;
 536         struct drm_i915_gem_object *obj = ringbuf->obj;
 537         int ret = 0;
 538
 539         gen6_gt_force_wake_get(dev_priv, FORCEWAKE_ALL);
 540
 541         if (!stop_ring(ring)) {
 542                 /* G45 ring initialization often fails to reset head to zero */
 543                 DRM_DEBUG_KMS("%s head not reset to zero "
 544                               "ctl %08x head %08x tail %08x start %08x\n",
 545                               ring->name,
 546                               I915_READ_CTL(ring),
 547                               I915_READ_HEAD(ring),
 548                               I915_READ_TAIL(ring),
 549                               I915_READ_START(ring));
 550
 551                 if (!stop_ring(ring)) {
 552                         DRM_ERROR("failed to set %s head to zero "
 553                                   "ctl %08x head %08x tail %08x start %08x\n",
 554                                   ring->name,
 555                                   I915_READ_CTL(ring),
 556                                   I915_READ_HEAD(ring),
 557                                   I915_READ_TAIL(ring),
 558                                   I915_READ_START(ring));
 559                         ret = -EIO;
 560                         goto out;
 561                 }
 562         }
 563
 564         if (I915_NEED_GFX_HWS(dev))
 565                 intel_ring_setup_status_page(ring);
 566         else
 567                 ring_setup_phys_status_page(ring);
 568
 569         /* Enforce ordering by reading HEAD register back */
 570         I915_READ_HEAD(ring);
 571
 572         /* Initialize the ring. This must happen _after_ we've cleared the ring
 573          * registers with the above sequence (the readback of the HEAD registers
 574          * also enforces ordering), otherwise the hw might lose the new ring
 575          * register values. */
 576         I915_WRITE_START(ring, i915_gem_obj_ggtt_offset(obj));
 577
 578         /* WaClearRingBufHeadRegAtInit:ctg,elk */
 579         if (I915_READ_HEAD(ring))
 580                 DRM_DEBUG("%s initialization failed [head=%08x], fudging\n",
 581                           ring->name, I915_READ_HEAD(ring));
 582         I915_WRITE_HEAD(ring, 0);
 583         (void)I915_READ_HEAD(ring);
 584
 585         I915_WRITE_CTL(ring,
 586                         ((ringbuf->size - PAGE_SIZE) & RING_NR_PAGES)
 587                         | RING_VALID);
 588
 589         /* If the head is still not zero, the ring is dead */
 590         if (wait_for((I915_READ_CTL(ring) & RING_VALID) != 0 &&
 591                      I915_READ_START(ring) == i915_gem_obj_ggtt_offset(obj) &&
 592                      (I915_READ_HEAD(ring) & HEAD_ADDR) == 0, 50)) {
 593                 DRM_ERROR("%s initialization failed "
 594                           "ctl %08x (valid? %d) head %08x tail %08x start %08x [expected %08lx]\n",
 595                           ring->name,
 596                           I915_READ_CTL(ring), I915_READ_CTL(ring) & RING_VALID,
 597                           I915_READ_HEAD(ring), I915_READ_TAIL(ring),
 598                           I915_READ_START(ring), (unsigned long)i915_gem_obj_ggtt_offset(obj));
 599                 ret = -EIO;
 600                 goto out;
 601         }
 602
 603         ringbuf->last_retired_head = -1;
 604         ringbuf->head = I915_READ_HEAD(ring);
 605         ringbuf->tail = I915_READ_TAIL(ring) & TAIL_ADDR;
 606         intel_ring_update_space(ringbuf);
 607
 608         memset(&ring->hangcheck, 0, sizeof(ring->hangcheck));
 609
 610 out:
 611         gen6_gt_force_wake_put(dev_priv, FORCEWAKE_ALL);
 612
 613         return ret;
 614 }
 615
 616 void
 617 intel_fini_pipe_control(struct intel_engine_cs *ring)
 618 {
 619         struct drm_device *dev = ring->dev;
 620
 621         if (ring->scratch.obj == NULL)
 622                 return;
 623
 624         if (INTEL_INFO(dev)->gen >= 5) {
 625                 kunmap(sg_page(ring->scratch.obj->pages->sgl));
 626                 i915_gem_object_ggtt_unpin(ring->scratch.obj);
 627         }
 628
 629         drm_gem_object_unreference(&ring->scratch.obj->base);
 630         ring->scratch.obj = NULL;
 631 }
 632
 633 int
 634 intel_init_pipe_control(struct intel_engine_cs *ring)
 635 {
 636         int ret;
 637
 638         WARN_ON(ring->scratch.obj);
 639
 640         ring->scratch.obj = i915_gem_alloc_object(ring->dev, 4096);
 641         if (ring->scratch.obj == NULL) {
 642                 DRM_ERROR("Failed to allocate seqno page\n");
 643                 ret = -ENOMEM;
 644                 goto err;
 645         }
 646
 647         ret = i915_gem_object_set_cache_level(ring->scratch.obj, I915_CACHE_LLC);
 648         if (ret)
 649                 goto err_unref;
 650
 651         ret = i915_gem_obj_ggtt_pin(ring->scratch.obj, 4096, 0);
 652         if (ret)
 653                 goto err_unref;
 654
 655         ring->scratch.gtt_offset = i915_gem_obj_ggtt_offset(ring->scratch.obj);
 656         ring->scratch.cpu_page = kmap(sg_page(ring->scratch.obj->pages->sgl));
 657         if (ring->scratch.cpu_page == NULL) {
 658                 ret = -ENOMEM;
 659                 goto err_unpin;
 660         }
 661
 662         DRM_DEBUG_DRIVER("%s pipe control offset: 0x%08x\n",
 663                          ring->name, ring->scratch.gtt_offset);
 664         return 0;
 665
 666 err_unpin:
 667         i915_gem_object_ggtt_unpin(ring->scratch.obj);
 668 err_unref:
 669         drm_gem_object_unreference(&ring->scratch.obj->base);
 670 err:
 671         return ret;
 672 }
 673
 674 static int intel_ring_workarounds_emit(struct intel_engine_cs *ring,
 675                                        struct intel_context *ctx)
 676 {
 677         int ret, i;
 678         struct drm_device *dev = ring->dev;
 679         struct drm_i915_private *dev_priv = dev->dev_private;
 680         struct i915_workarounds *w = &dev_priv->workarounds;
 681
 682         if (WARN_ON_ONCE(w->count == 0))
 683                 return 0;
 684
 685         ring->gpu_caches_dirty = true;
 686         ret = intel_ring_flush_all_caches(ring);
 687         if (ret)
 688                 return ret;
 689
 690         ret = intel_ring_begin(ring, (w->count * 2 + 2));
 691         if (ret)
 692                 return ret;
 693
 694         intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(w->count));
 695         for (i = 0; i < w->count; i++) {
 696                 intel_ring_emit(ring, w->reg[i].addr);
 697                 intel_ring_emit(ring, w->reg[i].value);
 698         }
 699         intel_ring_emit(ring, MI_NOOP);
 700
 701         intel_ring_advance(ring);
 702
 703         ring->gpu_caches_dirty = true;
 704         ret = intel_ring_flush_all_caches(ring);
 705         if (ret)
 706                 return ret;
 707
 708         DRM_DEBUG_DRIVER("Number of Workarounds emitted: %d\n", w->count);
 709
 710         return 0;
 711 }
 712
 713 static int intel_rcs_ctx_init(struct intel_engine_cs *ring,
 714                               struct intel_context *ctx)
 715 {
 716         int ret;
 717
 718         ret = intel_ring_workarounds_emit(ring, ctx);
 719         if (ret != 0)
 720                 return ret;
 721
 722         ret = i915_gem_render_state_init(ring);
 723         if (ret)
 724                 DRM_ERROR("init render state: %d\n", ret);
 725
 726         return ret;
 727 }
 728
 729 static int wa_add(struct drm_i915_private *dev_priv,
 730                   const u32 addr, const u32 val, const u32 mask)
 731 {
 732         const u32 idx = dev_priv->workarounds.count;
 733
 734         if (WARN_ON(idx >= I915_MAX_WA_REGS))
 735                 return -ENOSPC;
 736
 737         dev_priv->workarounds.reg[idx].addr = addr;
 738         dev_priv->workarounds.reg[idx].value = val;
 739         dev_priv->workarounds.reg[idx].mask = mask;
 740
 741         dev_priv->workarounds.count++;
 742
 743         return 0;
 744 }
 745
 746 #define WA_REG(addr, val, mask) { \
 747                 const int r = wa_add(dev_priv, (addr), (val), (mask)); \
 748                 if (r) \
 749                         return r; \
 750         }
 751
 752 #define WA_SET_BIT_MASKED(addr, mask) \
 753         WA_REG(addr, _MASKED_BIT_ENABLE(mask), (mask) & 0xffff)
 754
 755 #define WA_CLR_BIT_MASKED(addr, mask) \
 756         WA_REG(addr, _MASKED_BIT_DISABLE(mask), (mask) & 0xffff)
 757
 758 #define WA_SET_BIT(addr, mask) WA_REG(addr, I915_READ(addr) | (mask), mask)
 759 #define WA_CLR_BIT(addr, mask) WA_REG(addr, I915_READ(addr) & ~(mask), mask)
 760
 761 #define WA_WRITE(addr, val) WA_REG(addr, val, 0xffffffff)
 762
 763 static int bdw_init_workarounds(struct intel_engine_cs *ring)
 764 {
 765         struct drm_device *dev = ring->dev;
 766         struct drm_i915_private *dev_priv = dev->dev_private;
 767
 768         /* WaDisablePartialInstShootdown:bdw */
 769         /* WaDisableThreadStallDopClockGating:bdw (pre-production) */
 770         WA_SET_BIT_MASKED(GEN8_ROW_CHICKEN,
 771                           PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE |
 772                           STALL_DOP_GATING_DISABLE);
 773
 774         /* WaDisableDopClockGating:bdw */
 775         WA_SET_BIT_MASKED(GEN7_ROW_CHICKEN2,
 776                           DOP_CLOCK_GATING_DISABLE);
 777
 778         WA_SET_BIT_MASKED(HALF_SLICE_CHICKEN3,
 779                           GEN8_SAMPLER_POWER_BYPASS_DIS);
 780
 781         /* Use Force Non-Coherent whenever executing a 3D context. This is a
 782          * workaround for for a possible hang in the unlikely event a TLB
 783          * invalidation occurs during a PSD flush.
 784          */
 785         /* WaForceEnableNonCoherent:bdw */
 786         /* WaHdcDisableFetchWhenMasked:bdw */
 787         /* WaDisableFenceDestinationToSLM:bdw (GT3 pre-production) */
 788         WA_SET_BIT_MASKED(HDC_CHICKEN0,
 789                           HDC_FORCE_NON_COHERENT |
 790                           HDC_DONOT_FETCH_MEM_WHEN_MASKED |
 791                           (IS_BDW_GT3(dev) ? HDC_FENCE_DEST_SLM_DISABLE : 0));
 792
 793         /* Wa4x4STCOptimizationDisable:bdw */
 794         WA_SET_BIT_MASKED(CACHE_MODE_1,
 795                           GEN8_4x4_STC_OPTIMIZATION_DISABLE);
 796
 797         /*
 798          * BSpec recommends 8x4 when MSAA is used,
 799          * however in practice 16x4 seems fastest.
 800          *
 801          * Note that PS/WM thread counts depend on the WIZ hashing
 802          * disable bit, which we don't touch here, but it's good
 803          * to keep in mind (see 3DSTATE_PS and 3DSTATE_WM).
 804          */
 805         WA_SET_BIT_MASKED(GEN7_GT_MODE,
 806                           GEN6_WIZ_HASHING_MASK | GEN6_WIZ_HASHING_16x4);
 807
 808         return 0;
 809 }
 810
 811 static int chv_init_workarounds(struct intel_engine_cs *ring)
 812 {
 813         struct drm_device *dev = ring->dev;
 814         struct drm_i915_private *dev_priv = dev->dev_private;
 815
 816         /* WaDisablePartialInstShootdown:chv */
 817         /* WaDisableThreadStallDopClockGating:chv */
 818         WA_SET_BIT_MASKED(GEN8_ROW_CHICKEN,
 819                           PARTIAL_INSTRUCTION_SHOOTDOWN_DISABLE |
 820                           STALL_DOP_GATING_DISABLE);
 821
 822         /* Use Force Non-Coherent whenever executing a 3D context. This is a
 823          * workaround for a possible hang in the unlikely event a TLB
 824          * invalidation occurs during a PSD flush.
 825          */
 826         /* WaForceEnableNonCoherent:chv */
 827         /* WaHdcDisableFetchWhenMasked:chv */
 828         WA_SET_BIT_MASKED(HDC_CHICKEN0,
 829                           HDC_FORCE_NON_COHERENT |
 830                           HDC_DONOT_FETCH_MEM_WHEN_MASKED);
 831
 832         return 0;
 833 }
 834
 835 int init_workarounds_ring(struct intel_engine_cs *ring)
 836 {
 837         struct drm_device *dev = ring->dev;
 838         struct drm_i915_private *dev_priv = dev->dev_private;
 839
 840         WARN_ON(ring->id != RCS);
 841
 842         dev_priv->workarounds.count = 0;
 843
 844         if (IS_BROADWELL(dev))
 845                 return bdw_init_workarounds(ring);
 846
 847         if (IS_CHERRYVIEW(dev))
 848                 return chv_init_workarounds(ring);
 849
 850         return 0;
 851 }
 852
 853 static int init_render_ring(struct intel_engine_cs *ring)
 854 {
 855         struct drm_device *dev = ring->dev;
 856         struct drm_i915_private *dev_priv = dev->dev_private;
 857         int ret = init_ring_common(ring);
 858         if (ret)
 859                 return ret;
 860
 861         /* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
 862         if (INTEL_INFO(dev)->gen >= 4 && INTEL_INFO(dev)->gen < 7)
 863                 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH));
 864
 865         /* We need to disable the AsyncFlip performance optimisations in order
 866          * to use MI_WAIT_FOR_EVENT within the CS. It should already be
 867          * programmed to '1' on all products.
 868          *
 869          * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv,bdw,chv
 870          */
 871         if (INTEL_INFO(dev)->gen >= 6 && INTEL_INFO(dev)->gen < 9)
 872                 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
 873
 874         /* Required for the hardware to program scanline values for waiting */
 875         /* WaEnableFlushTlbInvalidationMode:snb */
 876         if (INTEL_INFO(dev)->gen == 6)
 877                 I915_WRITE(GFX_MODE,
 878                            _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT));
 879
 880         /* WaBCSVCSTlbInvalidationMode:ivb,vlv,hsw */
 881         if (IS_GEN7(dev))
 882                 I915_WRITE(GFX_MODE_GEN7,
 883                            _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT) |
 884                            _MASKED_BIT_ENABLE(GFX_REPLAY_MODE));
 885
 886         if (IS_GEN6(dev)) {
 887                 /* From the Sandybridge PRM, volume 1 part 3, page 24:
 888                  * "If this bit is set, STCunit will have LRA as replacement
 889                  *  policy. [...] This bit must be reset.  LRA replacement
 890                  *  policy is not supported."
 891                  */
 892                 I915_WRITE(CACHE_MODE_0,
 893                            _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
 894         }
 895
 896         if (INTEL_INFO(dev)->gen >= 6)
 897                 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
 898
 899         if (HAS_L3_DPF(dev))
 900                 I915_WRITE_IMR(ring, ~GT_PARITY_ERROR(dev));
 901
 902         return init_workarounds_ring(ring);
 903 }
 904
 905 static void render_ring_cleanup(struct intel_engine_cs *ring)
 906 {
 907         struct drm_device *dev = ring->dev;
 908         struct drm_i915_private *dev_priv = dev->dev_private;
 909
 910         if (dev_priv->semaphore_obj) {
 911                 i915_gem_object_ggtt_unpin(dev_priv->semaphore_obj);
 912                 drm_gem_object_unreference(&dev_priv->semaphore_obj->base);
 913                 dev_priv->semaphore_obj = NULL;
 914         }
 915
 916         intel_fini_pipe_control(ring);
 917 }
 918
 919 static int gen8_rcs_signal(struct intel_engine_cs *signaller,
 920                            unsigned int num_dwords)
 921 {
 922 #define MBOX_UPDATE_DWORDS 8
 923         struct drm_device *dev = signaller->dev;
 924         struct drm_i915_private *dev_priv = dev->dev_private;
 925         struct intel_engine_cs *waiter;
 926         int i, ret, num_rings;
 927
 928         num_rings = hweight32(INTEL_INFO(dev)->ring_mask);
 929         num_dwords += (num_rings-1) * MBOX_UPDATE_DWORDS;
 930 #undef MBOX_UPDATE_DWORDS
 931
 932         ret = intel_ring_begin(signaller, num_dwords);
 933         if (ret)
 934                 return ret;
 935
 936         for_each_ring(waiter, dev_priv, i) {
 937                 u32 seqno;
 938                 u64 gtt_offset = signaller->semaphore.signal_ggtt[i];
 939                 if (gtt_offset == MI_SEMAPHORE_SYNC_INVALID)
 940                         continue;
 941
 942                 seqno = i915_gem_request_get_seqno(
 943                                            signaller->outstanding_lazy_request);
 944                 intel_ring_emit(signaller, GFX_OP_PIPE_CONTROL(6));
 945                 intel_ring_emit(signaller, PIPE_CONTROL_GLOBAL_GTT_IVB |
 946                                            PIPE_CONTROL_QW_WRITE |
 947                                            PIPE_CONTROL_FLUSH_ENABLE);
 948                 intel_ring_emit(signaller, lower_32_bits(gtt_offset));
 949                 intel_ring_emit(signaller, upper_32_bits(gtt_offset));
 950                 intel_ring_emit(signaller, seqno);
 951                 intel_ring_emit(signaller, 0);
 952                 intel_ring_emit(signaller, MI_SEMAPHORE_SIGNAL |
 953                                            MI_SEMAPHORE_TARGET(waiter->id));
 954                 intel_ring_emit(signaller, 0);
 955         }
 956
 957         return 0;
 958 }
 959
 960 static int gen8_xcs_signal(struct intel_engine_cs *signaller,
 961                            unsigned int num_dwords)
 962 {
 963 #define MBOX_UPDATE_DWORDS 6
 964         struct drm_device *dev = signaller->dev;
 965         struct drm_i915_private *dev_priv = dev->dev_private;
 966         struct intel_engine_cs *waiter;
 967         int i, ret, num_rings;
 968
 969         num_rings = hweight32(INTEL_INFO(dev)->ring_mask);
 970         num_dwords += (num_rings-1) * MBOX_UPDATE_DWORDS;
 971 #undef MBOX_UPDATE_DWORDS
 972
 973         ret = intel_ring_begin(signaller, num_dwords);
 974         if (ret)
 975                 return ret;
 976
 977         for_each_ring(waiter, dev_priv, i) {
 978                 u32 seqno;
 979                 u64 gtt_offset = signaller->semaphore.signal_ggtt[i];
 980                 if (gtt_offset == MI_SEMAPHORE_SYNC_INVALID)
 981                         continue;
 982
 983                 seqno = i915_gem_request_get_seqno(
 984                                            signaller->outstanding_lazy_request);
 985                 intel_ring_emit(signaller, (MI_FLUSH_DW + 1) |
 986                                            MI_FLUSH_DW_OP_STOREDW);
 987                 intel_ring_emit(signaller, lower_32_bits(gtt_offset) |
 988                                            MI_FLUSH_DW_USE_GTT);
 989                 intel_ring_emit(signaller, upper_32_bits(gtt_offset));
 990                 intel_ring_emit(signaller, seqno);
 991                 intel_ring_emit(signaller, MI_SEMAPHORE_SIGNAL |
 992                                            MI_SEMAPHORE_TARGET(waiter->id));
 993                 intel_ring_emit(signaller, 0);
 994         }
 995
 996         return 0;
 997 }
 998
 999 static int gen6_signal(struct intel_engine_cs *signaller,
1000                        unsigned int num_dwords)
1001 {
1002         struct drm_device *dev = signaller->dev;
1003         struct drm_i915_private *dev_priv = dev->dev_private;
1004         struct intel_engine_cs *useless;
1005         int i, ret, num_rings;
1006
1007 #define MBOX_UPDATE_DWORDS 3
1008         num_rings = hweight32(INTEL_INFO(dev)->ring_mask);
1009         num_dwords += round_up((num_rings-1) * MBOX_UPDATE_DWORDS, 2);
1010 #undef MBOX_UPDATE_DWORDS
1011
1012         ret = intel_ring_begin(signaller, num_dwords);
1013         if (ret)
1014                 return ret;
1015
1016         for_each_ring(useless, dev_priv, i) {
1017                 u32 mbox_reg = signaller->semaphore.mbox.signal[i];
1018                 if (mbox_reg != GEN6_NOSYNC) {
1019                         u32 seqno = i915_gem_request_get_seqno(
1020                                            signaller->outstanding_lazy_request);
1021                         intel_ring_emit(signaller, MI_LOAD_REGISTER_IMM(1));
1022                         intel_ring_emit(signaller, mbox_reg);
1023                         intel_ring_emit(signaller, seqno);
1024                 }
1025         }
1026
1027         /* If num_dwords was rounded, make sure the tail pointer is correct */
1028         if (num_rings % 2 == 0)
1029                 intel_ring_emit(signaller, MI_NOOP);
1030
1031         return 0;
1032 }
1033
1034 /**
1035  * gen6_add_request - Update the semaphore mailbox registers
1036  *
1037  * @ring - ring that is adding a request
1038  * @seqno - return seqno stuck into the ring
1039  *
1040  * Update the mailbox registers in the *other* rings with the current seqno.
1041  * This acts like a signal in the canonical semaphore.
1042  */
1043 static int
1044 gen6_add_request(struct intel_engine_cs *ring)
1045 {
1046         int ret;
1047
1048         if (ring->semaphore.signal)
1049                 ret = ring->semaphore.signal(ring, 4);
1050         else
1051                 ret = intel_ring_begin(ring, 4);
1052
1053         if (ret)
1054                 return ret;
1055
1056         intel_ring_emit(ring, MI_STORE_DWORD_INDEX);
1057         intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
1058         intel_ring_emit(ring,
1059                     i915_gem_request_get_seqno(ring->outstanding_lazy_request));
1060         intel_ring_emit(ring, MI_USER_INTERRUPT);
1061         __intel_ring_advance(ring);
1062
1063         return 0;
1064 }
1065
1066 static inline bool i915_gem_has_seqno_wrapped(struct drm_device *dev,
1067                                               u32 seqno)
1068 {
1069         struct drm_i915_private *dev_priv = dev->dev_private;
1070         return dev_priv->last_seqno < seqno;
1071 }
1072
1073 /**
1074  * intel_ring_sync - sync the waiter to the signaller on seqno
1075  *
1076  * @waiter - ring that is waiting
1077  * @signaller - ring which has, or will signal
1078  * @seqno - seqno which the waiter will block on
1079  */
1080
1081 static int
1082 gen8_ring_sync(struct intel_engine_cs *waiter,
1083                struct intel_engine_cs *signaller,
1084                u32 seqno)
1085 {
1086         struct drm_i915_private *dev_priv = waiter->dev->dev_private;
1087         int ret;
1088
1089         ret = intel_ring_begin(waiter, 4);
1090         if (ret)
1091                 return ret;
1092
1093         intel_ring_emit(waiter, MI_SEMAPHORE_WAIT |
1094                                 MI_SEMAPHORE_GLOBAL_GTT |
1095                                 MI_SEMAPHORE_POLL |
1096                                 MI_SEMAPHORE_SAD_GTE_SDD);
1097         intel_ring_emit(waiter, seqno);
1098         intel_ring_emit(waiter,
1099                         lower_32_bits(GEN8_WAIT_OFFSET(waiter, signaller->id)));
1100         intel_ring_emit(waiter,
1101                         upper_32_bits(GEN8_WAIT_OFFSET(waiter, signaller->id)));
1102         intel_ring_advance(waiter);
1103         return 0;
1104 }
1105
1106 static int
1107 gen6_ring_sync(struct intel_engine_cs *waiter,
1108                struct intel_engine_cs *signaller,
1109                u32 seqno)
1110 {
1111         u32 dw1 = MI_SEMAPHORE_MBOX |
1112                   MI_SEMAPHORE_COMPARE |
1113                   MI_SEMAPHORE_REGISTER;
1114         u32 wait_mbox = signaller->semaphore.mbox.wait[waiter->id];
1115         int ret;
1116
1117         /* Throughout all of the GEM code, seqno passed implies our current
1118          * seqno is >= the last seqno executed. However for hardware the
1119          * comparison is strictly greater than.
1120          */
1121         seqno -= 1;
1122
1123         WARN_ON(wait_mbox == MI_SEMAPHORE_SYNC_INVALID);
1124
1125         ret = intel_ring_begin(waiter, 4);
1126         if (ret)
1127                 return ret;
1128
1129         /* If seqno wrap happened, omit the wait with no-ops */
1130         if (likely(!i915_gem_has_seqno_wrapped(waiter->dev, seqno))) {
1131                 intel_ring_emit(waiter, dw1 | wait_mbox);
1132                 intel_ring_emit(waiter, seqno);
1133                 intel_ring_emit(waiter, 0);
1134                 intel_ring_emit(waiter, MI_NOOP);
1135         } else {
1136                 intel_ring_emit(waiter, MI_NOOP);
1137                 intel_ring_emit(waiter, MI_NOOP);
1138                 intel_ring_emit(waiter, MI_NOOP);
1139                 intel_ring_emit(waiter, MI_NOOP);
1140         }
1141         intel_ring_advance(waiter);
1142
1143         return 0;
1144 }
1145
1146 #define PIPE_CONTROL_FLUSH(ring__, addr__)                                      \
1147 do {                                                                    \
1148         intel_ring_emit(ring__, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE |                \
1149                  PIPE_CONTROL_DEPTH_STALL);                             \
1150         intel_ring_emit(ring__, (addr__) | PIPE_CONTROL_GLOBAL_GTT);                    \
1151         intel_ring_emit(ring__, 0);                                                     \
1152         intel_ring_emit(ring__, 0);                                                     \
1153 } while (0)
1154
1155 static int
1156 pc_render_add_request(struct intel_engine_cs *ring)
1157 {
1158         u32 scratch_addr = ring->scratch.gtt_offset + 2 * CACHELINE_BYTES;
1159         int ret;
1160
1161         /* For Ironlake, MI_USER_INTERRUPT was deprecated and apparently
1162          * incoherent with writes to memory, i.e. completely fubar,
1163          * so we need to use PIPE_NOTIFY instead.
1164          *
1165          * However, we also need to workaround the qword write
1166          * incoherence by flushing the 6 PIPE_NOTIFY buffers out to
1167          * memory before requesting an interrupt.
1168          */
1169         ret = intel_ring_begin(ring, 32);
1170         if (ret)
1171                 return ret;
1172
1173         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE |
1174                         PIPE_CONTROL_WRITE_FLUSH |
1175                         PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE);
1176         intel_ring_emit(ring, ring->scratch.gtt_offset | PIPE_CONTROL_GLOBAL_GTT);
1177         intel_ring_emit(ring,
1178                     i915_gem_request_get_seqno(ring->outstanding_lazy_request));
1179         intel_ring_emit(ring, 0);
1180         PIPE_CONTROL_FLUSH(ring, scratch_addr);
1181         scratch_addr += 2 * CACHELINE_BYTES; /* write to separate cachelines */
1182         PIPE_CONTROL_FLUSH(ring, scratch_addr);
1183         scratch_addr += 2 * CACHELINE_BYTES;
1184         PIPE_CONTROL_FLUSH(ring, scratch_addr);
1185         scratch_addr += 2 * CACHELINE_BYTES;
1186         PIPE_CONTROL_FLUSH(ring, scratch_addr);
1187         scratch_addr += 2 * CACHELINE_BYTES;
1188         PIPE_CONTROL_FLUSH(ring, scratch_addr);
1189         scratch_addr += 2 * CACHELINE_BYTES;
1190         PIPE_CONTROL_FLUSH(ring, scratch_addr);
1191
1192         intel_ring_emit(ring, GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE |
1193                         PIPE_CONTROL_WRITE_FLUSH |
1194                         PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE |
1195                         PIPE_CONTROL_NOTIFY);
1196         intel_ring_emit(ring, ring->scratch.gtt_offset | PIPE_CONTROL_GLOBAL_GTT);
1197         intel_ring_emit(ring,
1198                     i915_gem_request_get_seqno(ring->outstanding_lazy_request));
1199         intel_ring_emit(ring, 0);
1200         __intel_ring_advance(ring);
1201
1202         return 0;
1203 }
1204
1205 static u32
1206 gen6_ring_get_seqno(struct intel_engine_cs *ring, bool lazy_coherency)
1207 {
1208         /* Workaround to force correct ordering between irq and seqno writes on
1209          * ivb (and maybe also on snb) by reading from a CS register (like
1210          * ACTHD) before reading the status page. */
1211         if (!lazy_coherency) {
1212                 struct drm_i915_private *dev_priv = ring->dev->dev_private;
1213                 POSTING_READ(RING_ACTHD(ring->mmio_base));
1214         }
1215
1216         return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
1217 }
1218
1219 static u32
1220 ring_get_seqno(struct intel_engine_cs *ring, bool lazy_coherency)
1221 {
1222         return intel_read_status_page(ring, I915_GEM_HWS_INDEX);
1223 }
1224
1225 static void
1226 ring_set_seqno(struct intel_engine_cs *ring, u32 seqno)
1227 {
1228         intel_write_status_page(ring, I915_GEM_HWS_INDEX, seqno);
1229 }
1230
1231 static u32
1232 pc_render_get_seqno(struct intel_engine_cs *ring, bool lazy_coherency)
1233 {
1234         return ring->scratch.cpu_page[0];
1235 }
1236
1237 static void
1238 pc_render_set_seqno(struct intel_engine_cs *ring, u32 seqno)
1239 {
1240         ring->scratch.cpu_page[0] = seqno;
1241 }
1242
1243 static bool
1244 gen5_ring_get_irq(struct intel_engine_cs *ring)
1245 {
1246         struct drm_device *dev = ring->dev;
1247         struct drm_i915_private *dev_priv = dev->dev_private;
1248         unsigned long flags;
1249
1250         if (WARN_ON(!intel_irqs_enabled(dev_priv)))
1251                 return false;
1252
1253         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1254         if (ring->irq_refcount++ == 0)
1255                 gen5_enable_gt_irq(dev_priv, ring->irq_enable_mask);
1256         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1257
1258         return true;
1259 }
1260
1261 static void
1262 gen5_ring_put_irq(struct intel_engine_cs *ring)
1263 {
1264         struct drm_device *dev = ring->dev;
1265         struct drm_i915_private *dev_priv = dev->dev_private;
1266         unsigned long flags;
1267
1268         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1269         if (--ring->irq_refcount == 0)
1270                 gen5_disable_gt_irq(dev_priv, ring->irq_enable_mask);
1271         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1272 }
1273
1274 static bool
1275 i9xx_ring_get_irq(struct intel_engine_cs *ring)
1276 {
1277         struct drm_device *dev = ring->dev;
1278         struct drm_i915_private *dev_priv = dev->dev_private;
1279         unsigned long flags;
1280
1281         if (!intel_irqs_enabled(dev_priv))
1282                 return false;
1283
1284         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1285         if (ring->irq_refcount++ == 0) {
1286                 dev_priv->irq_mask &= ~ring->irq_enable_mask;
1287                 I915_WRITE(IMR, dev_priv->irq_mask);
1288                 POSTING_READ(IMR);
1289         }
1290         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1291
1292         return true;
1293 }
1294
1295 static void
1296 i9xx_ring_put_irq(struct intel_engine_cs *ring)
1297 {
1298         struct drm_device *dev = ring->dev;
1299         struct drm_i915_private *dev_priv = dev->dev_private;
1300         unsigned long flags;
1301
1302         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1303         if (--ring->irq_refcount == 0) {
1304                 dev_priv->irq_mask |= ring->irq_enable_mask;
1305                 I915_WRITE(IMR, dev_priv->irq_mask);
1306                 POSTING_READ(IMR);
1307         }
1308         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1309 }
1310
1311 static bool
1312 i8xx_ring_get_irq(struct intel_engine_cs *ring)
1313 {
1314         struct drm_device *dev = ring->dev;
1315         struct drm_i915_private *dev_priv = dev->dev_private;
1316         unsigned long flags;
1317
1318         if (!intel_irqs_enabled(dev_priv))
1319                 return false;
1320
1321         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1322         if (ring->irq_refcount++ == 0) {
1323                 dev_priv->irq_mask &= ~ring->irq_enable_mask;
1324                 I915_WRITE16(IMR, dev_priv->irq_mask);
1325                 POSTING_READ16(IMR);
1326         }
1327         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1328
1329         return true;
1330 }
1331
1332 static void
1333 i8xx_ring_put_irq(struct intel_engine_cs *ring)
1334 {
1335         struct drm_device *dev = ring->dev;
1336         struct drm_i915_private *dev_priv = dev->dev_private;
1337         unsigned long flags;
1338
1339         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1340         if (--ring->irq_refcount == 0) {
1341                 dev_priv->irq_mask |= ring->irq_enable_mask;
1342                 I915_WRITE16(IMR, dev_priv->irq_mask);
1343                 POSTING_READ16(IMR);
1344         }
1345         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1346 }
1347
1348 void intel_ring_setup_status_page(struct intel_engine_cs *ring)
1349 {
1350         struct drm_device *dev = ring->dev;
1351         struct drm_i915_private *dev_priv = ring->dev->dev_private;
1352         u32 mmio = 0;
1353
1354         /* The ring status page addresses are no longer next to the rest of
1355          * the ring registers as of gen7.
1356          */
1357         if (IS_GEN7(dev)) {
1358                 switch (ring->id) {
1359                 case RCS:
1360                         mmio = RENDER_HWS_PGA_GEN7;
1361                         break;
1362                 case BCS:
1363                         mmio = BLT_HWS_PGA_GEN7;
1364                         break;
1365                 /*
1366                  * VCS2 actually doesn't exist on Gen7. Only shut up
1367                  * gcc switch check warning
1368                  */
1369                 case VCS2:
1370                 case VCS:
1371                         mmio = BSD_HWS_PGA_GEN7;
1372                         break;
1373                 case VECS:
1374                         mmio = VEBOX_HWS_PGA_GEN7;
1375                         break;
1376                 }
1377         } else if (IS_GEN6(ring->dev)) {
1378                 mmio = RING_HWS_PGA_GEN6(ring->mmio_base);
1379         } else {
1380                 /* XXX: gen8 returns to sanity */
1381                 mmio = RING_HWS_PGA(ring->mmio_base);
1382         }
1383
1384         I915_WRITE(mmio, (u32)ring->status_page.gfx_addr);
1385         POSTING_READ(mmio);
1386
1387         /*
1388          * Flush the TLB for this page
1389          *
1390          * FIXME: These two bits have disappeared on gen8, so a question
1391          * arises: do we still need this and if so how should we go about
1392          * invalidating the TLB?
1393          */
1394         if (INTEL_INFO(dev)->gen >= 6 && INTEL_INFO(dev)->gen < 8) {
1395                 u32 reg = RING_INSTPM(ring->mmio_base);
1396
1397                 /* ring should be idle before issuing a sync flush*/
1398                 WARN_ON((I915_READ_MODE(ring) & MODE_IDLE) == 0);
1399
1400                 I915_WRITE(reg,
1401                            _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE |
1402                                               INSTPM_SYNC_FLUSH));
1403                 if (wait_for((I915_READ(reg) & INSTPM_SYNC_FLUSH) == 0,
1404                              1000))
1405                         DRM_ERROR("%s: wait for SyncFlush to complete for TLB invalidation timed out\n",
1406                                   ring->name);
1407         }
1408 }
1409
1410 static int
1411 bsd_ring_flush(struct intel_engine_cs *ring,
1412                u32     invalidate_domains,
1413                u32     flush_domains)
1414 {
1415         int ret;
1416
1417         ret = intel_ring_begin(ring, 2);
1418         if (ret)
1419                 return ret;
1420
1421         intel_ring_emit(ring, MI_FLUSH);
1422         intel_ring_emit(ring, MI_NOOP);
1423         intel_ring_advance(ring);
1424         return 0;
1425 }
1426
1427 static int
1428 i9xx_add_request(struct intel_engine_cs *ring)
1429 {
1430         int ret;
1431
1432         ret = intel_ring_begin(ring, 4);
1433         if (ret)
1434                 return ret;
1435
1436         intel_ring_emit(ring, MI_STORE_DWORD_INDEX);
1437         intel_ring_emit(ring, I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT);
1438         intel_ring_emit(ring,
1439                     i915_gem_request_get_seqno(ring->outstanding_lazy_request));
1440         intel_ring_emit(ring, MI_USER_INTERRUPT);
1441         __intel_ring_advance(ring);
1442
1443         return 0;
1444 }
1445
1446 static bool
1447 gen6_ring_get_irq(struct intel_engine_cs *ring)
1448 {
1449         struct drm_device *dev = ring->dev;
1450         struct drm_i915_private *dev_priv = dev->dev_private;
1451         unsigned long flags;
1452
1453         if (WARN_ON(!intel_irqs_enabled(dev_priv)))
1454                 return false;
1455
1456         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1457         if (ring->irq_refcount++ == 0) {
1458                 if (HAS_L3_DPF(dev) && ring->id == RCS)
1459                         I915_WRITE_IMR(ring,
1460                                        ~(ring->irq_enable_mask |
1461                                          GT_PARITY_ERROR(dev)));
1462                 else
1463                         I915_WRITE_IMR(ring, ~ring->irq_enable_mask);
1464                 gen5_enable_gt_irq(dev_priv, ring->irq_enable_mask);
1465         }
1466         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1467
1468         return true;
1469 }
1470
1471 static void
1472 gen6_ring_put_irq(struct intel_engine_cs *ring)
1473 {
1474         struct drm_device *dev = ring->dev;
1475         struct drm_i915_private *dev_priv = dev->dev_private;
1476         unsigned long flags;
1477
1478         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1479         if (--ring->irq_refcount == 0) {
1480                 if (HAS_L3_DPF(dev) && ring->id == RCS)
1481                         I915_WRITE_IMR(ring, ~GT_PARITY_ERROR(dev));
1482                 else
1483                         I915_WRITE_IMR(ring, ~0);
1484                 gen5_disable_gt_irq(dev_priv, ring->irq_enable_mask);
1485         }
1486         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1487 }
1488
1489 static bool
1490 hsw_vebox_get_irq(struct intel_engine_cs *ring)
1491 {
1492         struct drm_device *dev = ring->dev;
1493         struct drm_i915_private *dev_priv = dev->dev_private;
1494         unsigned long flags;
1495
1496         if (WARN_ON(!intel_irqs_enabled(dev_priv)))
1497                 return false;
1498
1499         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1500         if (ring->irq_refcount++ == 0) {
1501                 I915_WRITE_IMR(ring, ~ring->irq_enable_mask);
1502                 gen6_enable_pm_irq(dev_priv, ring->irq_enable_mask);
1503         }
1504         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1505
1506         return true;
1507 }
1508
1509 static void
1510 hsw_vebox_put_irq(struct intel_engine_cs *ring)
1511 {
1512         struct drm_device *dev = ring->dev;
1513         struct drm_i915_private *dev_priv = dev->dev_private;
1514         unsigned long flags;
1515
1516         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1517         if (--ring->irq_refcount == 0) {
1518                 I915_WRITE_IMR(ring, ~0);
1519                 gen6_disable_pm_irq(dev_priv, ring->irq_enable_mask);
1520         }
1521         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1522 }
1523
1524 static bool
1525 gen8_ring_get_irq(struct intel_engine_cs *ring)
1526 {
1527         struct drm_device *dev = ring->dev;
1528         struct drm_i915_private *dev_priv = dev->dev_private;
1529         unsigned long flags;
1530
1531         if (WARN_ON(!intel_irqs_enabled(dev_priv)))
1532                 return false;
1533
1534         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1535         if (ring->irq_refcount++ == 0) {
1536                 if (HAS_L3_DPF(dev) && ring->id == RCS) {
1537                         I915_WRITE_IMR(ring,
1538                                        ~(ring->irq_enable_mask |
1539                                          GT_RENDER_L3_PARITY_ERROR_INTERRUPT));
1540                 } else {
1541                         I915_WRITE_IMR(ring, ~ring->irq_enable_mask);
1542                 }
1543                 POSTING_READ(RING_IMR(ring->mmio_base));
1544         }
1545         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1546
1547         return true;
1548 }
1549
1550 static void
1551 gen8_ring_put_irq(struct intel_engine_cs *ring)
1552 {
1553         struct drm_device *dev = ring->dev;
1554         struct drm_i915_private *dev_priv = dev->dev_private;
1555         unsigned long flags;
1556
1557         spin_lock_irqsave(&dev_priv->irq_lock, flags);
1558         if (--ring->irq_refcount == 0) {
1559                 if (HAS_L3_DPF(dev) && ring->id == RCS) {
1560                         I915_WRITE_IMR(ring,
1561                                        ~GT_RENDER_L3_PARITY_ERROR_INTERRUPT);
1562                 } else {
1563                         I915_WRITE_IMR(ring, ~0);
1564                 }
1565                 POSTING_READ(RING_IMR(ring->mmio_base));
1566         }
1567         spin_unlock_irqrestore(&dev_priv->irq_lock, flags);
1568 }
1569
1570 static int
1571 i965_dispatch_execbuffer(struct intel_engine_cs *ring,
1572                          u64 offset, u32 length,
1573                          unsigned flags)
1574 {
1575         int ret;
1576
1577         ret = intel_ring_begin(ring, 2);
1578         if (ret)
1579                 return ret;
1580
1581         intel_ring_emit(ring,
1582                         MI_BATCH_BUFFER_START |
1583                         MI_BATCH_GTT |
1584                         (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE_I965));
1585         intel_ring_emit(ring, offset);
1586         intel_ring_advance(ring);
1587
1588         return 0;
1589 }
1590
1591 /* Just userspace ABI convention to limit the wa batch bo to a resonable size */
1592 #define I830_BATCH_LIMIT (256*1024)
1593 #define I830_TLB_ENTRIES (2)
1594 #define I830_WA_SIZE max(I830_TLB_ENTRIES*4096, I830_BATCH_LIMIT)
1595 static int
1596 i830_dispatch_execbuffer(struct intel_engine_cs *ring,
1597                                 u64 offset, u32 len,
1598                                 unsigned flags)
1599 {
1600         u32 cs_offset = ring->scratch.gtt_offset;
1601         int ret;
1602
1603         ret = intel_ring_begin(ring, 6);
1604         if (ret)
1605                 return ret;
1606
1607         /* Evict the invalid PTE TLBs */
1608         intel_ring_emit(ring, COLOR_BLT_CMD | BLT_WRITE_RGBA);
1609         intel_ring_emit(ring, BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | 4096);
1610         intel_ring_emit(ring, I830_TLB_ENTRIES << 16 | 4); /* load each page */
1611         intel_ring_emit(ring, cs_offset);
1612         intel_ring_emit(ring, 0xdeadbeef);
1613         intel_ring_emit(ring, MI_NOOP);
1614         intel_ring_advance(ring);
1615
1616         if ((flags & I915_DISPATCH_PINNED) == 0) {
1617                 if (len > I830_BATCH_LIMIT)
1618                         return -ENOSPC;
1619
1620                 ret = intel_ring_begin(ring, 6 + 2);
1621                 if (ret)
1622                         return ret;
1623
1624                 /* Blit the batch (which has now all relocs applied) to the
1625                  * stable batch scratch bo area (so that the CS never
1626                  * stumbles over its tlb invalidation bug) ...
1627                  */
1628                 intel_ring_emit(ring, SRC_COPY_BLT_CMD | BLT_WRITE_RGBA);
1629                 intel_ring_emit(ring, BLT_DEPTH_32 | BLT_ROP_SRC_COPY | 4096);
1630                 intel_ring_emit(ring, DIV_ROUND_UP(len, 4096) << 16 | 4096);
1631                 intel_ring_emit(ring, cs_offset);
1632                 intel_ring_emit(ring, 4096);
1633                 intel_ring_emit(ring, offset);
1634
1635                 intel_ring_emit(ring, MI_FLUSH);
1636                 intel_ring_emit(ring, MI_NOOP);
1637                 intel_ring_advance(ring);
1638
1639                 /* ... and execute it. */
1640                 offset = cs_offset;
1641         }
1642
1643         ret = intel_ring_begin(ring, 4);
1644         if (ret)
1645                 return ret;
1646
1647         intel_ring_emit(ring, MI_BATCH_BUFFER);
1648         intel_ring_emit(ring, offset | (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE));
1649         intel_ring_emit(ring, offset + len - 8);
1650         intel_ring_emit(ring, MI_NOOP);
1651         intel_ring_advance(ring);
1652
1653         return 0;
1654 }
1655
1656 static int
1657 i915_dispatch_execbuffer(struct intel_engine_cs *ring,
1658                          u64 offset, u32 len,
1659                          unsigned flags)
1660 {
1661         int ret;
1662
1663         ret = intel_ring_begin(ring, 2);
1664         if (ret)
1665                 return ret;
1666
1667         intel_ring_emit(ring, MI_BATCH_BUFFER_START | MI_BATCH_GTT);
1668         intel_ring_emit(ring, offset | (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE));
1669         intel_ring_advance(ring);
1670
1671         return 0;
1672 }
1673
1674 static void cleanup_status_page(struct intel_engine_cs *ring)
1675 {
1676         struct drm_i915_gem_object *obj;
1677
1678         obj = ring->status_page.obj;
1679         if (obj == NULL)
1680                 return;
1681
1682         kunmap(sg_page(obj->pages->sgl));
1683         i915_gem_object_ggtt_unpin(obj);
1684         drm_gem_object_unreference(&obj->base);
1685         ring->status_page.obj = NULL;
1686 }
1687
1688 static int init_status_page(struct intel_engine_cs *ring)
1689 {
1690         struct drm_i915_gem_object *obj;
1691
1692         if ((obj = ring->status_page.obj) == NULL) {
1693                 unsigned flags;
1694                 int ret;
1695
1696                 obj = i915_gem_alloc_object(ring->dev, 4096);
1697                 if (obj == NULL) {
1698                         DRM_ERROR("Failed to allocate status page\n");
1699                         return -ENOMEM;
1700                 }
1701
1702                 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_LLC);
1703                 if (ret)
1704                         goto err_unref;
1705
1706                 flags = 0;
1707                 if (!HAS_LLC(ring->dev))
1708                         /* On g33, we cannot place HWS above 256MiB, so
1709                          * restrict its pinning to the low mappable arena.
1710                          * Though this restriction is not documented for
1711                          * gen4, gen5, or byt, they also behave similarly
1712                          * and hang if the HWS is placed at the top of the
1713                          * GTT. To generalise, it appears that all !llc
1714                          * platforms have issues with us placing the HWS
1715                          * above the mappable region (even though we never
1716                          * actualy map it).
1717                          */
1718                         flags |= PIN_MAPPABLE;
1719                 ret = i915_gem_obj_ggtt_pin(obj, 4096, flags);
1720                 if (ret) {
1721 err_unref:
1722                         drm_gem_object_unreference(&obj->base);
1723                         return ret;
1724                 }
1725
1726                 ring->status_page.obj = obj;
1727         }
1728
1729         ring->status_page.gfx_addr = i915_gem_obj_ggtt_offset(obj);
1730         ring->status_page.page_addr = kmap(sg_page(obj->pages->sgl));
1731         memset(ring->status_page.page_addr, 0, PAGE_SIZE);
1732
1733         DRM_DEBUG_DRIVER("%s hws offset: 0x%08x\n",
1734                         ring->name, ring->status_page.gfx_addr);
1735
1736         return 0;
1737 }
1738
1739 static int init_phys_status_page(struct intel_engine_cs *ring)
1740 {
1741         struct drm_i915_private *dev_priv = ring->dev->dev_private;
1742
1743         if (!dev_priv->status_page_dmah) {
1744                 dev_priv->status_page_dmah =
1745                         drm_pci_alloc(ring->dev, PAGE_SIZE, PAGE_SIZE);
1746                 if (!dev_priv->status_page_dmah)
1747                         return -ENOMEM;
1748         }
1749
1750         ring->status_page.page_addr = dev_priv->status_page_dmah->vaddr;
1751         memset(ring->status_page.page_addr, 0, PAGE_SIZE);
1752
1753         return 0;
1754 }
1755
1756 void intel_unpin_ringbuffer_obj(struct intel_ringbuffer *ringbuf)
1757 {
1758         iounmap(ringbuf->virtual_start);
1759         ringbuf->virtual_start = NULL;
1760         i915_gem_object_ggtt_unpin(ringbuf->obj);
1761 }
1762
1763 int intel_pin_and_map_ringbuffer_obj(struct drm_device *dev,
1764                                      struct intel_ringbuffer *ringbuf)
1765 {
1766         struct drm_i915_private *dev_priv = to_i915(dev);
1767         struct drm_i915_gem_object *obj = ringbuf->obj;
1768         int ret;
1769
1770         ret = i915_gem_obj_ggtt_pin(obj, PAGE_SIZE, PIN_MAPPABLE);
1771         if (ret)
1772                 return ret;
1773
1774         ret = i915_gem_object_set_to_gtt_domain(obj, true);
1775         if (ret) {
1776                 i915_gem_object_ggtt_unpin(obj);
1777                 return ret;
1778         }
1779
1780         ringbuf->virtual_start = ioremap_wc(dev_priv->gtt.mappable_base +
1781                         i915_gem_obj_ggtt_offset(obj), ringbuf->size);
1782         if (ringbuf->virtual_start == NULL) {
1783                 i915_gem_object_ggtt_unpin(obj);
1784                 return -EINVAL;
1785         }
1786
1787         return 0;
1788 }
1789
1790 void intel_destroy_ringbuffer_obj(struct intel_ringbuffer *ringbuf)
1791 {
1792         drm_gem_object_unreference(&ringbuf->obj->base);
1793         ringbuf->obj = NULL;
1794 }
1795
1796 int intel_alloc_ringbuffer_obj(struct drm_device *dev,
1797                                struct intel_ringbuffer *ringbuf)
1798 {
1799         struct drm_i915_gem_object *obj;
1800
1801         obj = NULL;
1802         if (!HAS_LLC(dev))
1803                 obj = i915_gem_object_create_stolen(dev, ringbuf->size);
1804         if (obj == NULL)
1805                 obj = i915_gem_alloc_object(dev, ringbuf->size);
1806         if (obj == NULL)
1807                 return -ENOMEM;
1808
1809         /* mark ring buffers as read-only from GPU side by default */
1810         obj->gt_ro = 1;
1811
1812         ringbuf->obj = obj;
1813
1814         return 0;
1815 }
1816
1817 static int intel_init_ring_buffer(struct drm_device *dev,
1818                                   struct intel_engine_cs *ring)
1819 {
1820         struct intel_ringbuffer *ringbuf;
1821         int ret;
1822
1823         WARN_ON(ring->buffer);
1824
1825         ringbuf = kzalloc(sizeof(*ringbuf), GFP_KERNEL);
1826         if (!ringbuf)
1827                 return -ENOMEM;
1828         ring->buffer = ringbuf;
1829
1830         ring->dev = dev;
1831         INIT_LIST_HEAD(&ring->active_list);
1832         INIT_LIST_HEAD(&ring->request_list);
1833         INIT_LIST_HEAD(&ring->execlist_queue);
1834         ringbuf->size = 32 * PAGE_SIZE;
1835         ringbuf->ring = ring;
1836         memset(ring->semaphore.sync_seqno, 0, sizeof(ring->semaphore.sync_seqno));
1837
1838         init_waitqueue_head(&ring->irq_queue);
1839
1840         if (I915_NEED_GFX_HWS(dev)) {
1841                 ret = init_status_page(ring);
1842                 if (ret)
1843                         goto error;
1844         } else {
1845                 BUG_ON(ring->id != RCS);
1846                 ret = init_phys_status_page(ring);
1847                 if (ret)
1848                         goto error;
1849         }
1850
1851         WARN_ON(ringbuf->obj);
1852
1853         ret = intel_alloc_ringbuffer_obj(dev, ringbuf);
1854         if (ret) {
1855                 DRM_ERROR("Failed to allocate ringbuffer %s: %d\n",
1856                                 ring->name, ret);
1857                 goto error;
1858         }
1859
1860         ret = intel_pin_and_map_ringbuffer_obj(dev, ringbuf);
1861         if (ret) {
1862                 DRM_ERROR("Failed to pin and map ringbuffer %s: %d\n",
1863                                 ring->name, ret);
1864                 intel_destroy_ringbuffer_obj(ringbuf);
1865                 goto error;
1866         }
1867
1868         /* Workaround an erratum on the i830 which causes a hang if
1869          * the TAIL pointer points to within the last 2 cachelines
1870          * of the buffer.
1871          */
1872         ringbuf->effective_size = ringbuf->size;
1873         if (IS_I830(dev) || IS_845G(dev))
1874                 ringbuf->effective_size -= 2 * CACHELINE_BYTES;
1875
1876         ret = i915_cmd_parser_init_ring(ring);
1877         if (ret)
1878                 goto error;
1879
1880         return 0;
1881
1882 error:
1883         kfree(ringbuf);
1884         ring->buffer = NULL;
1885         return ret;
1886 }
1887
1888 void intel_cleanup_ring_buffer(struct intel_engine_cs *ring)
1889 {
1890         struct drm_i915_private *dev_priv;
1891         struct intel_ringbuffer *ringbuf;
1892
1893         if (!intel_ring_initialized(ring))
1894                 return;
1895
1896         dev_priv = to_i915(ring->dev);
1897         ringbuf = ring->buffer;
1898
1899         intel_stop_ring_buffer(ring);
1900         WARN_ON(!IS_GEN2(ring->dev) && (I915_READ_MODE(ring) & MODE_IDLE) == 0);
1901
1902         intel_unpin_ringbuffer_obj(ringbuf);
1903         intel_destroy_ringbuffer_obj(ringbuf);
1904         i915_gem_request_assign(&ring->outstanding_lazy_request, NULL);
1905
1906         if (ring->cleanup)
1907                 ring->cleanup(ring);
1908
1909         cleanup_status_page(ring);
1910
1911         i915_cmd_parser_fini_ring(ring);
1912
1913         kfree(ringbuf);
1914         ring->buffer = NULL;
1915 }
1916
1917 static int intel_ring_wait_request(struct intel_engine_cs *ring, int n)
1918 {
1919         struct intel_ringbuffer *ringbuf = ring->buffer;
1920         struct drm_i915_gem_request *request;
1921         int ret;
1922
1923         if (intel_ring_space(ringbuf) >= n)
1924                 return 0;
1925
1926         list_for_each_entry(request, &ring->request_list, list) {
1927                 if (__intel_ring_space(request->tail, ringbuf->tail,
1928                                        ringbuf->size) >= n) {
1929                         break;
1930                 }
1931         }
1932
1933         if (&request->list == &ring->request_list)
1934                 return -ENOSPC;
1935
1936         ret = i915_wait_request(request);
1937         if (ret)
1938                 return ret;
1939
1940         i915_gem_retire_requests_ring(ring);
1941
1942         return 0;
1943 }
1944
1945 static int ring_wait_for_space(struct intel_engine_cs *ring, int n)
1946 {
1947         struct drm_device *dev = ring->dev;
1948         struct drm_i915_private *dev_priv = dev->dev_private;
1949         struct intel_ringbuffer *ringbuf = ring->buffer;
1950         unsigned long end;
1951         int ret;
1952
1953         ret = intel_ring_wait_request(ring, n);
1954         if (ret != -ENOSPC)
1955                 return ret;
1956
1957         /* force the tail write in case we have been skipping them */
1958         __intel_ring_advance(ring);
1959
1960         /* With GEM the hangcheck timer should kick us out of the loop,
1961          * leaving it early runs the risk of corrupting GEM state (due
1962          * to running on almost untested codepaths). But on resume
1963          * timers don't work yet, so prevent a complete hang in that
1964          * case by choosing an insanely large timeout. */
1965         end = jiffies + 60 * HZ;
1966
1967         ret = 0;
1968         trace_i915_ring_wait_begin(ring);
1969         do {
1970                 if (intel_ring_space(ringbuf) >= n)
1971                         break;
1972                 ringbuf->head = I915_READ_HEAD(ring);
1973                 if (intel_ring_space(ringbuf) >= n)
1974                         break;
1975
1976                 msleep(1);
1977
1978                 if (dev_priv->mm.interruptible && signal_pending(current)) {
1979                         ret = -ERESTARTSYS;
1980                         break;
1981                 }
1982
1983                 ret = i915_gem_check_wedge(&dev_priv->gpu_error,
1984                                            dev_priv->mm.interruptible);
1985                 if (ret)
1986                         break;
1987
1988                 if (time_after(jiffies, end)) {
1989                         ret = -EBUSY;
1990                         break;
1991                 }
1992         } while (1);
1993         trace_i915_ring_wait_end(ring);
1994         return ret;
1995 }
1996
1997 static int intel_wrap_ring_buffer(struct intel_engine_cs *ring)
1998 {
1999         uint32_t __iomem *virt;
2000         struct intel_ringbuffer *ringbuf = ring->buffer;
2001         int rem = ringbuf->size - ringbuf->tail;
2002
2003         if (ringbuf->space < rem) {
2004                 int ret = ring_wait_for_space(ring, rem);
2005                 if (ret)
2006                         return ret;
2007         }
2008
2009         virt = ringbuf->virtual_start + ringbuf->tail;
2010         rem /= 4;
2011         while (rem--)
2012                 iowrite32(MI_NOOP, virt++);
2013
2014         ringbuf->tail = 0;
2015         intel_ring_update_space(ringbuf);
2016
2017         return 0;
2018 }
2019
2020 int intel_ring_idle(struct intel_engine_cs *ring)
2021 {
2022         struct drm_i915_gem_request *req;
2023         int ret;
2024
2025         /* We need to add any requests required to flush the objects and ring */
2026         if (ring->outstanding_lazy_request) {
2027                 ret = i915_add_request(ring);
2028                 if (ret)
2029                         return ret;
2030         }
2031
2032         /* Wait upon the last request to be completed */
2033         if (list_empty(&ring->request_list))
2034                 return 0;
2035
2036         req = list_entry(ring->request_list.prev,
2037                            struct drm_i915_gem_request,
2038                            list);
2039
2040         return i915_wait_request(req);
2041 }
2042
2043 static int
2044 intel_ring_alloc_request(struct intel_engine_cs *ring)
2045 {
2046         int ret;
2047         struct drm_i915_gem_request *request;
2048         struct drm_i915_private *dev_private = ring->dev->dev_private;
2049
2050         if (ring->outstanding_lazy_request)
2051                 return 0;
2052
2053         request = kzalloc(sizeof(*request), GFP_KERNEL);
2054         if (request == NULL)
2055                 return -ENOMEM;
2056
2057         kref_init(&request->ref);
2058         request->ring = ring;
2059         request->uniq = dev_private->request_uniq++;
2060
2061         ret = i915_gem_get_seqno(ring->dev, &request->seqno);
2062         if (ret) {
2063                 kfree(request);
2064                 return ret;
2065         }
2066
2067         ring->outstanding_lazy_request = request;
2068         return 0;
2069 }
2070
2071 static int __intel_ring_prepare(struct intel_engine_cs *ring,
2072                                 int bytes)
2073 {
2074         struct intel_ringbuffer *ringbuf = ring->buffer;
2075         int ret;
2076
2077         if (unlikely(ringbuf->tail + bytes > ringbuf->effective_size)) {
2078                 ret = intel_wrap_ring_buffer(ring);
2079                 if (unlikely(ret))
2080                         return ret;
2081         }
2082
2083         if (unlikely(ringbuf->space < bytes)) {
2084                 ret = ring_wait_for_space(ring, bytes);
2085                 if (unlikely(ret))
2086                         return ret;
2087         }
2088
2089         return 0;
2090 }
2091
2092 int intel_ring_begin(struct intel_engine_cs *ring,
2093                      int num_dwords)
2094 {
2095         struct drm_i915_private *dev_priv = ring->dev->dev_private;
2096         int ret;
2097
2098         ret = i915_gem_check_wedge(&dev_priv->gpu_error,
2099                                    dev_priv->mm.interruptible);
2100         if (ret)
2101                 return ret;
2102
2103         ret = __intel_ring_prepare(ring, num_dwords * sizeof(uint32_t));
2104         if (ret)
2105                 return ret;
2106
2107         /* Preallocate the olr before touching the ring */
2108         ret = intel_ring_alloc_request(ring);
2109         if (ret)
2110                 return ret;
2111
2112         ring->buffer->space -= num_dwords * sizeof(uint32_t);
2113         return 0;
2114 }
2115
2116 /* Align the ring tail to a cacheline boundary */
2117 int intel_ring_cacheline_align(struct intel_engine_cs *ring)
2118 {
2119         int num_dwords = (ring->buffer->tail & (CACHELINE_BYTES - 1)) / sizeof(uint32_t);
2120         int ret;
2121
2122         if (num_dwords == 0)
2123                 return 0;
2124
2125         num_dwords = CACHELINE_BYTES / sizeof(uint32_t) - num_dwords;
2126         ret = intel_ring_begin(ring, num_dwords);
2127         if (ret)
2128                 return ret;
2129
2130         while (num_dwords--)
2131                 intel_ring_emit(ring, MI_NOOP);
2132
2133         intel_ring_advance(ring);
2134
2135         return 0;
2136 }
2137
2138 void intel_ring_init_seqno(struct intel_engine_cs *ring, u32 seqno)
2139 {
2140         struct drm_device *dev = ring->dev;
2141         struct drm_i915_private *dev_priv = dev->dev_private;
2142
2143         BUG_ON(ring->outstanding_lazy_request);
2144
2145         if (INTEL_INFO(dev)->gen == 6 || INTEL_INFO(dev)->gen == 7) {
2146                 I915_WRITE(RING_SYNC_0(ring->mmio_base), 0);
2147                 I915_WRITE(RING_SYNC_1(ring->mmio_base), 0);
2148                 if (HAS_VEBOX(dev))
2149                         I915_WRITE(RING_SYNC_2(ring->mmio_base), 0);
2150         }
2151
2152         ring->set_seqno(ring, seqno);
2153         ring->hangcheck.seqno = seqno;
2154 }
2155
2156 static void gen6_bsd_ring_write_tail(struct intel_engine_cs *ring,
2157                                      u32 value)
2158 {
2159         struct drm_i915_private *dev_priv = ring->dev->dev_private;
2160
2161        /* Every tail move must follow the sequence below */
2162
2163         /* Disable notification that the ring is IDLE. The GT
2164          * will then assume that it is busy and bring it out of rc6.
2165          */
2166         I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL,
2167                    _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
2168
2169         /* Clear the context id. Here be magic! */
2170         I915_WRITE64(GEN6_BSD_RNCID, 0x0);
2171
2172         /* Wait for the ring not to be idle, i.e. for it to wake up. */
2173         if (wait_for((I915_READ(GEN6_BSD_SLEEP_PSMI_CONTROL) &
2174                       GEN6_BSD_SLEEP_INDICATOR) == 0,
2175                      50))
2176                 DRM_ERROR("timed out waiting for the BSD ring to wake up\n");
2177
2178         /* Now that the ring is fully powered up, update the tail */
2179         I915_WRITE_TAIL(ring, value);
2180         POSTING_READ(RING_TAIL(ring->mmio_base));
2181
2182         /* Let the ring send IDLE messages to the GT again,
2183          * and so let it sleep to conserve power when idle.
2184          */
2185         I915_WRITE(GEN6_BSD_SLEEP_PSMI_CONTROL,
2186                    _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
2187 }
2188
2189 static int gen6_bsd_ring_flush(struct intel_engine_cs *ring,
2190                                u32 invalidate, u32 flush)
2191 {
2192         uint32_t cmd;
2193         int ret;
2194
2195         ret = intel_ring_begin(ring, 4);
2196         if (ret)
2197                 return ret;
2198
2199         cmd = MI_FLUSH_DW;
2200         if (INTEL_INFO(ring->dev)->gen >= 8)
2201                 cmd += 1;
2202         /*
2203          * Bspec vol 1c.5 - video engine command streamer:
2204          * "If ENABLED, all TLBs will be invalidated once the flush
2205          * operation is complete. This bit is only valid when the
2206          * Post-Sync Operation field is a value of 1h or 3h."
2207          */
2208         if (invalidate & I915_GEM_GPU_DOMAINS)
2209                 cmd |= MI_INVALIDATE_TLB | MI_INVALIDATE_BSD |
2210                         MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
2211         intel_ring_emit(ring, cmd);
2212         intel_ring_emit(ring, I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT);
2213         if (INTEL_INFO(ring->dev)->gen >= 8) {
2214                 intel_ring_emit(ring, 0); /* upper addr */
2215                 intel_ring_emit(ring, 0); /* value */
2216         } else  {
2217                 intel_ring_emit(ring, 0);
2218                 intel_ring_emit(ring, MI_NOOP);
2219         }
2220         intel_ring_advance(ring);
2221         return 0;
2222 }
2223
2224 static int
2225 gen8_ring_dispatch_execbuffer(struct intel_engine_cs *ring,
2226                               u64 offset, u32 len,
2227                               unsigned flags)
2228 {
2229         bool ppgtt = USES_PPGTT(ring->dev) && !(flags & I915_DISPATCH_SECURE);
2230         int ret;
2231
2232         ret = intel_ring_begin(ring, 4);
2233         if (ret)
2234                 return ret;
2235
2236         /* FIXME(BDW): Address space and security selectors. */
2237         intel_ring_emit(ring, MI_BATCH_BUFFER_START_GEN8 | (ppgtt<<8));
2238         intel_ring_emit(ring, lower_32_bits(offset));
2239         intel_ring_emit(ring, upper_32_bits(offset));
2240         intel_ring_emit(ring, MI_NOOP);
2241         intel_ring_advance(ring);
2242
2243         return 0;
2244 }
2245
2246 static int
2247 hsw_ring_dispatch_execbuffer(struct intel_engine_cs *ring,
2248                               u64 offset, u32 len,
2249                               unsigned flags)
2250 {
2251         int ret;
2252
2253         ret = intel_ring_begin(ring, 2);
2254         if (ret)
2255                 return ret;
2256
2257         intel_ring_emit(ring,
2258                         MI_BATCH_BUFFER_START |
2259                         (flags & I915_DISPATCH_SECURE ?
2260                          0 : MI_BATCH_PPGTT_HSW | MI_BATCH_NON_SECURE_HSW));
2261         /* bit0-7 is the length on GEN6+ */
2262         intel_ring_emit(ring, offset);
2263         intel_ring_advance(ring);
2264
2265         return 0;
2266 }
2267
2268 static int
2269 gen6_ring_dispatch_execbuffer(struct intel_engine_cs *ring,
2270                               u64 offset, u32 len,
2271                               unsigned flags)
2272 {
2273         int ret;
2274
2275         ret = intel_ring_begin(ring, 2);
2276         if (ret)
2277                 return ret;
2278
2279         intel_ring_emit(ring,
2280                         MI_BATCH_BUFFER_START |
2281                         (flags & I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE_I965));
2282         /* bit0-7 is the length on GEN6+ */
2283         intel_ring_emit(ring, offset);
2284         intel_ring_advance(ring);
2285
2286         return 0;
2287 }
2288
2289 /* Blitter support (SandyBridge+) */
2290
2291 static int gen6_ring_flush(struct intel_engine_cs *ring,
2292                            u32 invalidate, u32 flush)
2293 {
2294         struct drm_device *dev = ring->dev;
2295         struct drm_i915_private *dev_priv = dev->dev_private;
2296         uint32_t cmd;
2297         int ret;
2298
2299         ret = intel_ring_begin(ring, 4);
2300         if (ret)
2301                 return ret;
2302
2303         cmd = MI_FLUSH_DW;
2304         if (INTEL_INFO(ring->dev)->gen >= 8)
2305                 cmd += 1;
2306         /*
2307          * Bspec vol 1c.3 - blitter engine command streamer:
2308          * "If ENABLED, all TLBs will be invalidated once the flush
2309          * operation is complete. This bit is only valid when the
2310          * Post-Sync Operation field is a value of 1h or 3h."
2311          */
2312         if (invalidate & I915_GEM_DOMAIN_RENDER)
2313                 cmd |= MI_INVALIDATE_TLB | MI_FLUSH_DW_STORE_INDEX |
2314                         MI_FLUSH_DW_OP_STOREDW;
2315         intel_ring_emit(ring, cmd);
2316         intel_ring_emit(ring, I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT);
2317         if (INTEL_INFO(ring->dev)->gen >= 8) {
2318                 intel_ring_emit(ring, 0); /* upper addr */
2319                 intel_ring_emit(ring, 0); /* value */
2320         } else  {
2321                 intel_ring_emit(ring, 0);
2322                 intel_ring_emit(ring, MI_NOOP);
2323         }
2324         intel_ring_advance(ring);
2325
2326         if (!invalidate && flush) {
2327                 if (IS_GEN7(dev))
2328                         return gen7_ring_fbc_flush(ring, FBC_REND_CACHE_CLEAN);
2329                 else if (IS_BROADWELL(dev))
2330                         dev_priv->fbc.need_sw_cache_clean = true;
2331         }
2332
2333         return 0;
2334 }
2335
2336 int intel_init_render_ring_buffer(struct drm_device *dev)
2337 {
2338         struct drm_i915_private *dev_priv = dev->dev_private;
2339         struct intel_engine_cs *ring = &dev_priv->ring[RCS];
2340         struct drm_i915_gem_object *obj;
2341         int ret;
2342
2343         ring->name = "render ring";
2344         ring->id = RCS;
2345         ring->mmio_base = RENDER_RING_BASE;
2346
2347         if (INTEL_INFO(dev)->gen >= 8) {
2348                 if (i915_semaphore_is_enabled(dev)) {
2349                         obj = i915_gem_alloc_object(dev, 4096);
2350                         if (obj == NULL) {
2351                                 DRM_ERROR("Failed to allocate semaphore bo. Disabling semaphores\n");
2352                                 i915.semaphores = 0;
2353                         } else {
2354                                 i915_gem_object_set_cache_level(obj, I915_CACHE_LLC);
2355                                 ret = i915_gem_obj_ggtt_pin(obj, 0, PIN_NONBLOCK);
2356                                 if (ret != 0) {
2357                                         drm_gem_object_unreference(&obj->base);
2358                                         DRM_ERROR("Failed to pin semaphore bo. Disabling semaphores\n");
2359                                         i915.semaphores = 0;
2360                                 } else
2361                                         dev_priv->semaphore_obj = obj;
2362                         }
2363                 }
2364
2365                 ring->init_context = intel_rcs_ctx_init;
2366                 ring->add_request = gen6_add_request;
2367                 ring->flush = gen8_render_ring_flush;
2368                 ring->irq_get = gen8_ring_get_irq;
2369                 ring->irq_put = gen8_ring_put_irq;
2370                 ring->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
2371                 ring->get_seqno = gen6_ring_get_seqno;
2372                 ring->set_seqno = ring_set_seqno;
2373                 if (i915_semaphore_is_enabled(dev)) {
2374                         WARN_ON(!dev_priv->semaphore_obj);
2375                         ring->semaphore.sync_to = gen8_ring_sync;
2376                         ring->semaphore.signal = gen8_rcs_signal;
2377                         GEN8_RING_SEMAPHORE_INIT;
2378                 }
2379         } else if (INTEL_INFO(dev)->gen >= 6) {
2380                 ring->add_request = gen6_add_request;
2381                 ring->flush = gen7_render_ring_flush;
2382                 if (INTEL_INFO(dev)->gen == 6)
2383                         ring->flush = gen6_render_ring_flush;
2384                 ring->irq_get = gen6_ring_get_irq;
2385                 ring->irq_put = gen6_ring_put_irq;
2386                 ring->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
2387                 ring->get_seqno = gen6_ring_get_seqno;
2388                 ring->set_seqno = ring_set_seqno;
2389                 if (i915_semaphore_is_enabled(dev)) {
2390                         ring->semaphore.sync_to = gen6_ring_sync;
2391                         ring->semaphore.signal = gen6_signal;
2392                         /*
2393                          * The current semaphore is only applied on pre-gen8
2394                          * platform.  And there is no VCS2 ring on the pre-gen8
2395                          * platform. So the semaphore between RCS and VCS2 is
2396                          * initialized as INVALID.  Gen8 will initialize the
2397                          * sema between VCS2 and RCS later.
2398                          */
2399                         ring->semaphore.mbox.wait[RCS] = MI_SEMAPHORE_SYNC_INVALID;
2400                         ring->semaphore.mbox.wait[VCS] = MI_SEMAPHORE_SYNC_RV;
2401                         ring->semaphore.mbox.wait[BCS] = MI_SEMAPHORE_SYNC_RB;
2402                         ring->semaphore.mbox.wait[VECS] = MI_SEMAPHORE_SYNC_RVE;
2403                         ring->semaphore.mbox.wait[VCS2] = MI_SEMAPHORE_SYNC_INVALID;
2404                         ring->semaphore.mbox.signal[RCS] = GEN6_NOSYNC;
2405                         ring->semaphore.mbox.signal[VCS] = GEN6_VRSYNC;
2406                         ring->semaphore.mbox.signal[BCS] = GEN6_BRSYNC;
2407                         ring->semaphore.mbox.signal[VECS] = GEN6_VERSYNC;
2408                         ring->semaphore.mbox.signal[VCS2] = GEN6_NOSYNC;
2409                 }
2410         } else if (IS_GEN5(dev)) {
2411                 ring->add_request = pc_render_add_request;
2412                 ring->flush = gen4_render_ring_flush;
2413                 ring->get_seqno = pc_render_get_seqno;
2414                 ring->set_seqno = pc_render_set_seqno;
2415                 ring->irq_get = gen5_ring_get_irq;
2416                 ring->irq_put = gen5_ring_put_irq;
2417                 ring->irq_enable_mask = GT_RENDER_USER_INTERRUPT |
2418                                         GT_RENDER_PIPECTL_NOTIFY_INTERRUPT;
2419         } else {
2420                 ring->add_request = i9xx_add_request;
2421                 if (INTEL_INFO(dev)->gen < 4)
2422                         ring->flush = gen2_render_ring_flush;
2423                 else
2424                         ring->flush = gen4_render_ring_flush;
2425                 ring->get_seqno = ring_get_seqno;
2426                 ring->set_seqno = ring_set_seqno;
2427                 if (IS_GEN2(dev)) {
2428                         ring->irq_get = i8xx_ring_get_irq;
2429                         ring->irq_put = i8xx_ring_put_irq;
2430                 } else {
2431                         ring->irq_get = i9xx_ring_get_irq;
2432                         ring->irq_put = i9xx_ring_put_irq;
2433                 }
2434                 ring->irq_enable_mask = I915_USER_INTERRUPT;
2435         }
2436         ring->write_tail = ring_write_tail;
2437
2438         if (IS_HASWELL(dev))
2439                 ring->dispatch_execbuffer = hsw_ring_dispatch_execbuffer;
2440         else if (IS_GEN8(dev))
2441                 ring->dispatch_execbuffer = gen8_ring_dispatch_execbuffer;
2442         else if (INTEL_INFO(dev)->gen >= 6)
2443                 ring->dispatch_execbuffer = gen6_ring_dispatch_execbuffer;
2444         else if (INTEL_INFO(dev)->gen >= 4)
2445                 ring->dispatch_execbuffer = i965_dispatch_execbuffer;
2446         else if (IS_I830(dev) || IS_845G(dev))
2447                 ring->dispatch_execbuffer = i830_dispatch_execbuffer;
2448         else
2449                 ring->dispatch_execbuffer = i915_dispatch_execbuffer;
2450         ring->init_hw = init_render_ring;
2451         ring->cleanup = render_ring_cleanup;
2452
2453         /* Workaround batchbuffer to combat CS tlb bug. */
2454         if (HAS_BROKEN_CS_TLB(dev)) {
2455                 obj = i915_gem_alloc_object(dev, I830_WA_SIZE);
2456                 if (obj == NULL) {
2457                         DRM_ERROR("Failed to allocate batch bo\n");
2458                         return -ENOMEM;
2459                 }
2460
2461                 ret = i915_gem_obj_ggtt_pin(obj, 0, 0);
2462                 if (ret != 0) {
2463                         drm_gem_object_unreference(&obj->base);
2464                         DRM_ERROR("Failed to ping batch bo\n");
2465                         return ret;
2466                 }
2467
2468                 ring->scratch.obj = obj;
2469                 ring->scratch.gtt_offset = i915_gem_obj_ggtt_offset(obj);
2470         }
2471
2472         ret = intel_init_ring_buffer(dev, ring);
2473         if (ret)
2474                 return ret;
2475
2476         if (INTEL_INFO(dev)->gen >= 5) {
2477                 ret = intel_init_pipe_control(ring);
2478                 if (ret)
2479                         return ret;
2480         }
2481
2482         return 0;
2483 }
2484
2485 int intel_init_bsd_ring_buffer(struct drm_device *dev)
2486 {
2487         struct drm_i915_private *dev_priv = dev->dev_private;
2488         struct intel_engine_cs *ring = &dev_priv->ring[VCS];
2489
2490         ring->name = "bsd ring";
2491         ring->id = VCS;
2492
2493         ring->write_tail = ring_write_tail;
2494         if (INTEL_INFO(dev)->gen >= 6) {
2495                 ring->mmio_base = GEN6_BSD_RING_BASE;
2496                 /* gen6 bsd needs a special wa for tail updates */
2497                 if (IS_GEN6(dev))
2498                         ring->write_tail = gen6_bsd_ring_write_tail;
2499                 ring->flush = gen6_bsd_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                 if (INTEL_INFO(dev)->gen >= 8) {
2504                         ring->irq_enable_mask =
2505                                 GT_RENDER_USER_INTERRUPT << GEN8_VCS1_IRQ_SHIFT;
2506                         ring->irq_get = gen8_ring_get_irq;
2507                         ring->irq_put = gen8_ring_put_irq;
2508                         ring->dispatch_execbuffer =
2509                                 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 = GT_BSD_USER_INTERRUPT;
2517                         ring->irq_get = gen6_ring_get_irq;
2518                         ring->irq_put = gen6_ring_put_irq;
2519                         ring->dispatch_execbuffer =
2520                                 gen6_ring_dispatch_execbuffer;
2521                         if (i915_semaphore_is_enabled(dev)) {
2522                                 ring->semaphore.sync_to = gen6_ring_sync;
2523                                 ring->semaphore.signal = gen6_signal;
2524                                 ring->semaphore.mbox.wait[RCS] = MI_SEMAPHORE_SYNC_VR;
2525                                 ring->semaphore.mbox.wait[VCS] = MI_SEMAPHORE_SYNC_INVALID;
2526                                 ring->semaphore.mbox.wait[BCS] = MI_SEMAPHORE_SYNC_VB;
2527                                 ring->semaphore.mbox.wait[VECS] = MI_SEMAPHORE_SYNC_VVE;
2528                                 ring->semaphore.mbox.wait[VCS2] = MI_SEMAPHORE_SYNC_INVALID;
2529                                 ring->semaphore.mbox.signal[RCS] = GEN6_RVSYNC;
2530                                 ring->semaphore.mbox.signal[VCS] = GEN6_NOSYNC;
2531                                 ring->semaphore.mbox.signal[BCS] = GEN6_BVSYNC;
2532                                 ring->semaphore.mbox.signal[VECS] = GEN6_VEVSYNC;
2533                                 ring->semaphore.mbox.signal[VCS2] = GEN6_NOSYNC;
2534                         }
2535                 }
2536         } else {
2537                 ring->mmio_base = BSD_RING_BASE;
2538                 ring->flush = bsd_ring_flush;
2539                 ring->add_request = i9xx_add_request;
2540                 ring->get_seqno = ring_get_seqno;
2541                 ring->set_seqno = ring_set_seqno;
2542                 if (IS_GEN5(dev)) {
2543                         ring->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
2544                         ring->irq_get = gen5_ring_get_irq;
2545                         ring->irq_put = gen5_ring_put_irq;
2546                 } else {
2547                         ring->irq_enable_mask = I915_BSD_USER_INTERRUPT;
2548                         ring->irq_get = i9xx_ring_get_irq;
2549                         ring->irq_put = i9xx_ring_put_irq;
2550                 }
2551                 ring->dispatch_execbuffer = i965_dispatch_execbuffer;
2552         }
2553         ring->init_hw = init_ring_common;
2554
2555         return intel_init_ring_buffer(dev, ring);
2556 }
2557
2558 /**
2559  * Initialize the second BSD ring for Broadwell GT3.
2560  * It is noted that this only exists on Broadwell GT3.
2561  */
2562 int intel_init_bsd2_ring_buffer(struct drm_device *dev)
2563 {
2564         struct drm_i915_private *dev_priv = dev->dev_private;
2565         struct intel_engine_cs *ring = &dev_priv->ring[VCS2];
2566
2567         if ((INTEL_INFO(dev)->gen != 8)) {
2568                 DRM_ERROR("No dual-BSD ring on non-BDW machine\n");
2569                 return -EINVAL;
2570         }
2571
2572         ring->name = "bsd2 ring";
2573         ring->id = VCS2;
2574
2575         ring->write_tail = ring_write_tail;
2576         ring->mmio_base = GEN8_BSD2_RING_BASE;
2577         ring->flush = gen6_bsd_ring_flush;
2578         ring->add_request = gen6_add_request;
2579         ring->get_seqno = gen6_ring_get_seqno;
2580         ring->set_seqno = ring_set_seqno;
2581         ring->irq_enable_mask =
2582                         GT_RENDER_USER_INTERRUPT << GEN8_VCS2_IRQ_SHIFT;
2583         ring->irq_get = gen8_ring_get_irq;
2584         ring->irq_put = gen8_ring_put_irq;
2585         ring->dispatch_execbuffer =
2586                         gen8_ring_dispatch_execbuffer;
2587         if (i915_semaphore_is_enabled(dev)) {
2588                 ring->semaphore.sync_to = gen8_ring_sync;
2589                 ring->semaphore.signal = gen8_xcs_signal;
2590                 GEN8_RING_SEMAPHORE_INIT;
2591         }
2592         ring->init_hw = init_ring_common;
2593
2594         return intel_init_ring_buffer(dev, ring);
2595 }
2596
2597 int intel_init_blt_ring_buffer(struct drm_device *dev)
2598 {
2599         struct drm_i915_private *dev_priv = dev->dev_private;
2600         struct intel_engine_cs *ring = &dev_priv->ring[BCS];
2601
2602         ring->name = "blitter ring";
2603         ring->id = BCS;
2604
2605         ring->mmio_base = BLT_RING_BASE;
2606         ring->write_tail = ring_write_tail;
2607         ring->flush = gen6_ring_flush;
2608         ring->add_request = gen6_add_request;
2609         ring->get_seqno = gen6_ring_get_seqno;
2610         ring->set_seqno = ring_set_seqno;
2611         if (INTEL_INFO(dev)->gen >= 8) {
2612                 ring->irq_enable_mask =
2613                         GT_RENDER_USER_INTERRUPT << GEN8_BCS_IRQ_SHIFT;
2614                 ring->irq_get = gen8_ring_get_irq;
2615                 ring->irq_put = gen8_ring_put_irq;
2616                 ring->dispatch_execbuffer = gen8_ring_dispatch_execbuffer;
2617                 if (i915_semaphore_is_enabled(dev)) {
2618                         ring->semaphore.sync_to = gen8_ring_sync;
2619                         ring->semaphore.signal = gen8_xcs_signal;
2620                         GEN8_RING_SEMAPHORE_INIT;
2621                 }
2622         } else {
2623                 ring->irq_enable_mask = GT_BLT_USER_INTERRUPT;
2624                 ring->irq_get = gen6_ring_get_irq;
2625                 ring->irq_put = gen6_ring_put_irq;
2626                 ring->dispatch_execbuffer = gen6_ring_dispatch_execbuffer;
2627                 if (i915_semaphore_is_enabled(dev)) {
2628                         ring->semaphore.signal = gen6_signal;
2629                         ring->semaphore.sync_to = gen6_ring_sync;
2630                         /*
2631                          * The current semaphore is only applied on pre-gen8
2632                          * platform.  And there is no VCS2 ring on the pre-gen8
2633                          * platform. So the semaphore between BCS and VCS2 is
2634                          * initialized as INVALID.  Gen8 will initialize the
2635                          * sema between BCS and VCS2 later.
2636                          */
2637                         ring->semaphore.mbox.wait[RCS] = MI_SEMAPHORE_SYNC_BR;
2638                         ring->semaphore.mbox.wait[VCS] = MI_SEMAPHORE_SYNC_BV;
2639                         ring->semaphore.mbox.wait[BCS] = MI_SEMAPHORE_SYNC_INVALID;
2640                         ring->semaphore.mbox.wait[VECS] = MI_SEMAPHORE_SYNC_BVE;
2641                         ring->semaphore.mbox.wait[VCS2] = MI_SEMAPHORE_SYNC_INVALID;
2642                         ring->semaphore.mbox.signal[RCS] = GEN6_RBSYNC;
2643                         ring->semaphore.mbox.signal[VCS] = GEN6_VBSYNC;
2644                         ring->semaphore.mbox.signal[BCS] = GEN6_NOSYNC;
2645                         ring->semaphore.mbox.signal[VECS] = GEN6_VEBSYNC;
2646                         ring->semaphore.mbox.signal[VCS2] = GEN6_NOSYNC;
2647                 }
2648         }
2649         ring->init_hw = init_ring_common;
2650
2651         return intel_init_ring_buffer(dev, ring);
2652 }
2653
2654 int intel_init_vebox_ring_buffer(struct drm_device *dev)
2655 {
2656         struct drm_i915_private *dev_priv = dev->dev_private;
2657         struct intel_engine_cs *ring = &dev_priv->ring[VECS];
2658
2659         ring->name = "video enhancement ring";
2660         ring->id = VECS;
2661
2662         ring->mmio_base = VEBOX_RING_BASE;
2663         ring->write_tail = ring_write_tail;
2664         ring->flush = gen6_ring_flush;
2665         ring->add_request = gen6_add_request;
2666         ring->get_seqno = gen6_ring_get_seqno;
2667         ring->set_seqno = ring_set_seqno;
2668
2669         if (INTEL_INFO(dev)->gen >= 8) {
2670                 ring->irq_enable_mask =
2671                         GT_RENDER_USER_INTERRUPT << GEN8_VECS_IRQ_SHIFT;
2672                 ring->irq_get = gen8_ring_get_irq;
2673                 ring->irq_put = gen8_ring_put_irq;
2674                 ring->dispatch_execbuffer = gen8_ring_dispatch_execbuffer;
2675                 if (i915_semaphore_is_enabled(dev)) {
2676                         ring->semaphore.sync_to = gen8_ring_sync;
2677                         ring->semaphore.signal = gen8_xcs_signal;
2678                         GEN8_RING_SEMAPHORE_INIT;
2679                 }
2680         } else {
2681                 ring->irq_enable_mask = PM_VEBOX_USER_INTERRUPT;
2682                 ring->irq_get = hsw_vebox_get_irq;
2683                 ring->irq_put = hsw_vebox_put_irq;
2684                 ring->dispatch_execbuffer = gen6_ring_dispatch_execbuffer;
2685                 if (i915_semaphore_is_enabled(dev)) {
2686                         ring->semaphore.sync_to = gen6_ring_sync;
2687                         ring->semaphore.signal = gen6_signal;
2688                         ring->semaphore.mbox.wait[RCS] = MI_SEMAPHORE_SYNC_VER;
2689                         ring->semaphore.mbox.wait[VCS] = MI_SEMAPHORE_SYNC_VEV;
2690                         ring->semaphore.mbox.wait[BCS] = MI_SEMAPHORE_SYNC_VEB;
2691                         ring->semaphore.mbox.wait[VECS] = MI_SEMAPHORE_SYNC_INVALID;
2692                         ring->semaphore.mbox.wait[VCS2] = MI_SEMAPHORE_SYNC_INVALID;
2693                         ring->semaphore.mbox.signal[RCS] = GEN6_RVESYNC;
2694                         ring->semaphore.mbox.signal[VCS] = GEN6_VVESYNC;
2695                         ring->semaphore.mbox.signal[BCS] = GEN6_BVESYNC;
2696                         ring->semaphore.mbox.signal[VECS] = GEN6_NOSYNC;
2697                         ring->semaphore.mbox.signal[VCS2] = GEN6_NOSYNC;
2698                 }
2699         }
2700         ring->init_hw = init_ring_common;
2701
2702         return intel_init_ring_buffer(dev, ring);
2703 }
2704
2705 int
2706 intel_ring_flush_all_caches(struct intel_engine_cs *ring)
2707 {
2708         int ret;
2709
2710         if (!ring->gpu_caches_dirty)
2711                 return 0;
2712
2713         ret = ring->flush(ring, 0, I915_GEM_GPU_DOMAINS);
2714         if (ret)
2715                 return ret;
2716
2717         trace_i915_gem_ring_flush(ring, 0, I915_GEM_GPU_DOMAINS);
2718
2719         ring->gpu_caches_dirty = false;
2720         return 0;
2721 }
2722
2723 int
2724 intel_ring_invalidate_all_caches(struct intel_engine_cs *ring)
2725 {
2726         uint32_t flush_domains;
2727         int ret;
2728
2729         flush_domains = 0;
2730         if (ring->gpu_caches_dirty)
2731                 flush_domains = I915_GEM_GPU_DOMAINS;
2732
2733         ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, flush_domains);
2734         if (ret)
2735                 return ret;
2736
2737         trace_i915_gem_ring_flush(ring, I915_GEM_GPU_DOMAINS, flush_domains);
2738
2739         ring->gpu_caches_dirty = false;
2740         return 0;
2741 }
2742
2743 void
2744 intel_stop_ring_buffer(struct intel_engine_cs *ring)
2745 {
2746         int ret;
2747
2748         if (!intel_ring_initialized(ring))
2749                 return;
2750
2751         ret = intel_ring_idle(ring);
2752         if (ret && !i915_reset_in_progress(&to_i915(ring->dev)->gpu_error))
2753                 DRM_ERROR("failed to quiesce %s whilst cleaning up: %d\n",
2754                           ring->name, ret);
2755
2756         stop_ring(ring);
2757 }