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 <linux/log2.h>
  31 #include <drm/drmP.h>
  32 #include "i915_drv.h"
  33 #include <drm/i915_drm.h>
  34 #include "i915_trace.h"
  35 #include "intel_drv.h"
  36
  37 /* Rough estimate of the typical request size, performing a flush,
  38  * set-context and then emitting the batch.
  39  */
  40 #define LEGACY_REQUEST_SIZE 200
  41
  42 static int __intel_ring_space(int head, int tail, int size)
  43 {
  44         int space = head - tail;
  45         if (space <= 0)
  46                 space += size;
  47         return space - I915_RING_FREE_SPACE;
  48 }
  49
  50 void intel_ring_update_space(struct intel_ring *ring)
  51 {
  52         ring->space = __intel_ring_space(ring->head, ring->emit, ring->size);
  53 }
  54
  55 static int
  56 gen2_render_ring_flush(struct drm_i915_gem_request *req, u32 mode)
  57 {
  58         u32 cmd, *cs;
  59
  60         cmd = MI_FLUSH;
  61
  62         if (mode & EMIT_INVALIDATE)
  63                 cmd |= MI_READ_FLUSH;
  64
  65         cs = intel_ring_begin(req, 2);
  66         if (IS_ERR(cs))
  67                 return PTR_ERR(cs);
  68
  69         *cs++ = cmd;
  70         *cs++ = MI_NOOP;
  71         intel_ring_advance(req, cs);
  72
  73         return 0;
  74 }
  75
  76 static int
  77 gen4_render_ring_flush(struct drm_i915_gem_request *req, u32 mode)
  78 {
  79         u32 cmd, *cs;
  80
  81         /*
  82          * read/write caches:
  83          *
  84          * I915_GEM_DOMAIN_RENDER is always invalidated, but is
  85          * only flushed if MI_NO_WRITE_FLUSH is unset.  On 965, it is
  86          * also flushed at 2d versus 3d pipeline switches.
  87          *
  88          * read-only caches:
  89          *
  90          * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
  91          * MI_READ_FLUSH is set, and is always flushed on 965.
  92          *
  93          * I915_GEM_DOMAIN_COMMAND may not exist?
  94          *
  95          * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
  96          * invalidated when MI_EXE_FLUSH is set.
  97          *
  98          * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
  99          * invalidated with every MI_FLUSH.
 100          *
 101          * TLBs:
 102          *
 103          * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
 104          * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
 105          * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
 106          * are flushed at any MI_FLUSH.
 107          */
 108
 109         cmd = MI_FLUSH;
 110         if (mode & EMIT_INVALIDATE) {
 111                 cmd |= MI_EXE_FLUSH;
 112                 if (IS_G4X(req->i915) || IS_GEN5(req->i915))
 113                         cmd |= MI_INVALIDATE_ISP;
 114         }
 115
 116         cs = intel_ring_begin(req, 2);
 117         if (IS_ERR(cs))
 118                 return PTR_ERR(cs);
 119
 120         *cs++ = cmd;
 121         *cs++ = MI_NOOP;
 122         intel_ring_advance(req, cs);
 123
 124         return 0;
 125 }
 126
 127 /**
 128  * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
 129  * implementing two workarounds on gen6.  From section 1.4.7.1
 130  * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
 131  *
 132  * [DevSNB-C+{W/A}] Before any depth stall flush (including those
 133  * produced by non-pipelined state commands), software needs to first
 134  * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
 135  * 0.
 136  *
 137  * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
 138  * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
 139  *
 140  * And the workaround for these two requires this workaround first:
 141  *
 142  * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
 143  * BEFORE the pipe-control with a post-sync op and no write-cache
 144  * flushes.
 145  *
 146  * And this last workaround is tricky because of the requirements on
 147  * that bit.  From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
 148  * volume 2 part 1:
 149  *
 150  *     "1 of the following must also be set:
 151  *      - Render Target Cache Flush Enable ([12] of DW1)
 152  *      - Depth Cache Flush Enable ([0] of DW1)
 153  *      - Stall at Pixel Scoreboard ([1] of DW1)
 154  *      - Depth Stall ([13] of DW1)
 155  *      - Post-Sync Operation ([13] of DW1)
 156  *      - Notify Enable ([8] of DW1)"
 157  *
 158  * The cache flushes require the workaround flush that triggered this
 159  * one, so we can't use it.  Depth stall would trigger the same.
 160  * Post-sync nonzero is what triggered this second workaround, so we
 161  * can't use that one either.  Notify enable is IRQs, which aren't
 162  * really our business.  That leaves only stall at scoreboard.
 163  */
 164 static int
 165 intel_emit_post_sync_nonzero_flush(struct drm_i915_gem_request *req)
 166 {
 167         u32 scratch_addr =
 168                 i915_ggtt_offset(req->engine->scratch) + 2 * CACHELINE_BYTES;
 169         u32 *cs;
 170
 171         cs = intel_ring_begin(req, 6);
 172         if (IS_ERR(cs))
 173                 return PTR_ERR(cs);
 174
 175         *cs++ = GFX_OP_PIPE_CONTROL(5);
 176         *cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
 177         *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
 178         *cs++ = 0; /* low dword */
 179         *cs++ = 0; /* high dword */
 180         *cs++ = MI_NOOP;
 181         intel_ring_advance(req, cs);
 182
 183         cs = intel_ring_begin(req, 6);
 184         if (IS_ERR(cs))
 185                 return PTR_ERR(cs);
 186
 187         *cs++ = GFX_OP_PIPE_CONTROL(5);
 188         *cs++ = PIPE_CONTROL_QW_WRITE;
 189         *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
 190         *cs++ = 0;
 191         *cs++ = 0;
 192         *cs++ = MI_NOOP;
 193         intel_ring_advance(req, cs);
 194
 195         return 0;
 196 }
 197
 198 static int
 199 gen6_render_ring_flush(struct drm_i915_gem_request *req, u32 mode)
 200 {
 201         u32 scratch_addr =
 202                 i915_ggtt_offset(req->engine->scratch) + 2 * CACHELINE_BYTES;
 203         u32 *cs, flags = 0;
 204         int ret;
 205
 206         /* Force SNB workarounds for PIPE_CONTROL flushes */
 207         ret = intel_emit_post_sync_nonzero_flush(req);
 208         if (ret)
 209                 return ret;
 210
 211         /* Just flush everything.  Experiments have shown that reducing the
 212          * number of bits based on the write domains has little performance
 213          * impact.
 214          */
 215         if (mode & EMIT_FLUSH) {
 216                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
 217                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
 218                 /*
 219                  * Ensure that any following seqno writes only happen
 220                  * when the render cache is indeed flushed.
 221                  */
 222                 flags |= PIPE_CONTROL_CS_STALL;
 223         }
 224         if (mode & EMIT_INVALIDATE) {
 225                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
 226                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 227                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 228                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
 229                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
 230                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
 231                 /*
 232                  * TLB invalidate requires a post-sync write.
 233                  */
 234                 flags |= PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_CS_STALL;
 235         }
 236
 237         cs = intel_ring_begin(req, 4);
 238         if (IS_ERR(cs))
 239                 return PTR_ERR(cs);
 240
 241         *cs++ = GFX_OP_PIPE_CONTROL(4);
 242         *cs++ = flags;
 243         *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
 244         *cs++ = 0;
 245         intel_ring_advance(req, cs);
 246
 247         return 0;
 248 }
 249
 250 static int
 251 gen7_render_ring_cs_stall_wa(struct drm_i915_gem_request *req)
 252 {
 253         u32 *cs;
 254
 255         cs = intel_ring_begin(req, 4);
 256         if (IS_ERR(cs))
 257                 return PTR_ERR(cs);
 258
 259         *cs++ = GFX_OP_PIPE_CONTROL(4);
 260         *cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
 261         *cs++ = 0;
 262         *cs++ = 0;
 263         intel_ring_advance(req, cs);
 264
 265         return 0;
 266 }
 267
 268 static int
 269 gen7_render_ring_flush(struct drm_i915_gem_request *req, u32 mode)
 270 {
 271         u32 scratch_addr =
 272                 i915_ggtt_offset(req->engine->scratch) + 2 * CACHELINE_BYTES;
 273         u32 *cs, flags = 0;
 274
 275         /*
 276          * Ensure that any following seqno writes only happen when the render
 277          * cache is indeed flushed.
 278          *
 279          * Workaround: 4th PIPE_CONTROL command (except the ones with only
 280          * read-cache invalidate bits set) must have the CS_STALL bit set. We
 281          * don't try to be clever and just set it unconditionally.
 282          */
 283         flags |= PIPE_CONTROL_CS_STALL;
 284
 285         /* Just flush everything.  Experiments have shown that reducing the
 286          * number of bits based on the write domains has little performance
 287          * impact.
 288          */
 289         if (mode & EMIT_FLUSH) {
 290                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
 291                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
 292                 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
 293                 flags |= PIPE_CONTROL_FLUSH_ENABLE;
 294         }
 295         if (mode & EMIT_INVALIDATE) {
 296                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
 297                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 298                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 299                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
 300                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
 301                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
 302                 flags |= PIPE_CONTROL_MEDIA_STATE_CLEAR;
 303                 /*
 304                  * TLB invalidate requires a post-sync write.
 305                  */
 306                 flags |= PIPE_CONTROL_QW_WRITE;
 307                 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
 308
 309                 flags |= PIPE_CONTROL_STALL_AT_SCOREBOARD;
 310
 311                 /* Workaround: we must issue a pipe_control with CS-stall bit
 312                  * set before a pipe_control command that has the state cache
 313                  * invalidate bit set. */
 314                 gen7_render_ring_cs_stall_wa(req);
 315         }
 316
 317         cs = intel_ring_begin(req, 4);
 318         if (IS_ERR(cs))
 319                 return PTR_ERR(cs);
 320
 321         *cs++ = GFX_OP_PIPE_CONTROL(4);
 322         *cs++ = flags;
 323         *cs++ = scratch_addr;
 324         *cs++ = 0;
 325         intel_ring_advance(req, cs);
 326
 327         return 0;
 328 }
 329
 330 static int
 331 gen8_render_ring_flush(struct drm_i915_gem_request *req, u32 mode)
 332 {
 333         u32 flags;
 334         u32 *cs;
 335
 336         cs = intel_ring_begin(req, mode & EMIT_INVALIDATE ? 12 : 6);
 337         if (IS_ERR(cs))
 338                 return PTR_ERR(cs);
 339
 340         flags = PIPE_CONTROL_CS_STALL;
 341
 342         if (mode & EMIT_FLUSH) {
 343                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
 344                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
 345                 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
 346                 flags |= PIPE_CONTROL_FLUSH_ENABLE;
 347         }
 348         if (mode & EMIT_INVALIDATE) {
 349                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
 350                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 351                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 352                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
 353                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
 354                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
 355                 flags |= PIPE_CONTROL_QW_WRITE;
 356                 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
 357
 358                 /* WaCsStallBeforeStateCacheInvalidate:bdw,chv */
 359                 cs = gen8_emit_pipe_control(cs,
 360                                             PIPE_CONTROL_CS_STALL |
 361                                             PIPE_CONTROL_STALL_AT_SCOREBOARD,
 362                                             0);
 363         }
 364
 365         cs = gen8_emit_pipe_control(cs, flags,
 366                                     i915_ggtt_offset(req->engine->scratch) +
 367                                     2 * CACHELINE_BYTES);
 368
 369         intel_ring_advance(req, cs);
 370
 371         return 0;
 372 }
 373
 374 static void ring_setup_phys_status_page(struct intel_engine_cs *engine)
 375 {
 376         struct drm_i915_private *dev_priv = engine->i915;
 377         u32 addr;
 378
 379         addr = dev_priv->status_page_dmah->busaddr;
 380         if (INTEL_GEN(dev_priv) >= 4)
 381                 addr |= (dev_priv->status_page_dmah->busaddr >> 28) & 0xf0;
 382         I915_WRITE(HWS_PGA, addr);
 383 }
 384
 385 static void intel_ring_setup_status_page(struct intel_engine_cs *engine)
 386 {
 387         struct drm_i915_private *dev_priv = engine->i915;
 388         i915_reg_t mmio;
 389
 390         /* The ring status page addresses are no longer next to the rest of
 391          * the ring registers as of gen7.
 392          */
 393         if (IS_GEN7(dev_priv)) {
 394                 switch (engine->id) {
 395                 case RCS:
 396                         mmio = RENDER_HWS_PGA_GEN7;
 397                         break;
 398                 case BCS:
 399                         mmio = BLT_HWS_PGA_GEN7;
 400                         break;
 401                 /*
 402                  * VCS2 actually doesn't exist on Gen7. Only shut up
 403                  * gcc switch check warning
 404                  */
 405                 case VCS2:
 406                 case VCS:
 407                         mmio = BSD_HWS_PGA_GEN7;
 408                         break;
 409                 case VECS:
 410                         mmio = VEBOX_HWS_PGA_GEN7;
 411                         break;
 412                 }
 413         } else if (IS_GEN6(dev_priv)) {
 414                 mmio = RING_HWS_PGA_GEN6(engine->mmio_base);
 415         } else {
 416                 /* XXX: gen8 returns to sanity */
 417                 mmio = RING_HWS_PGA(engine->mmio_base);
 418         }
 419
 420         I915_WRITE(mmio, engine->status_page.ggtt_offset);
 421         POSTING_READ(mmio);
 422
 423         /*
 424          * Flush the TLB for this page
 425          *
 426          * FIXME: These two bits have disappeared on gen8, so a question
 427          * arises: do we still need this and if so how should we go about
 428          * invalidating the TLB?
 429          */
 430         if (IS_GEN(dev_priv, 6, 7)) {
 431                 i915_reg_t reg = RING_INSTPM(engine->mmio_base);
 432
 433                 /* ring should be idle before issuing a sync flush*/
 434                 WARN_ON((I915_READ_MODE(engine) & MODE_IDLE) == 0);
 435
 436                 I915_WRITE(reg,
 437                            _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE |
 438                                               INSTPM_SYNC_FLUSH));
 439                 if (intel_wait_for_register(dev_priv,
 440                                             reg, INSTPM_SYNC_FLUSH, 0,
 441                                             1000))
 442                         DRM_ERROR("%s: wait for SyncFlush to complete for TLB invalidation timed out\n",
 443                                   engine->name);
 444         }
 445 }
 446
 447 static bool stop_ring(struct intel_engine_cs *engine)
 448 {
 449         struct drm_i915_private *dev_priv = engine->i915;
 450
 451         if (INTEL_GEN(dev_priv) > 2) {
 452                 I915_WRITE_MODE(engine, _MASKED_BIT_ENABLE(STOP_RING));
 453                 if (intel_wait_for_register(dev_priv,
 454                                             RING_MI_MODE(engine->mmio_base),
 455                                             MODE_IDLE,
 456                                             MODE_IDLE,
 457                                             1000)) {
 458                         DRM_ERROR("%s : timed out trying to stop ring\n",
 459                                   engine->name);
 460                         /* Sometimes we observe that the idle flag is not
 461                          * set even though the ring is empty. So double
 462                          * check before giving up.
 463                          */
 464                         if (I915_READ_HEAD(engine) != I915_READ_TAIL(engine))
 465                                 return false;
 466                 }
 467         }
 468
 469         I915_WRITE_CTL(engine, 0);
 470         I915_WRITE_HEAD(engine, 0);
 471         I915_WRITE_TAIL(engine, 0);
 472
 473         if (INTEL_GEN(dev_priv) > 2) {
 474                 (void)I915_READ_CTL(engine);
 475                 I915_WRITE_MODE(engine, _MASKED_BIT_DISABLE(STOP_RING));
 476         }
 477
 478         return (I915_READ_HEAD(engine) & HEAD_ADDR) == 0;
 479 }
 480
 481 static int init_ring_common(struct intel_engine_cs *engine)
 482 {
 483         struct drm_i915_private *dev_priv = engine->i915;
 484         struct intel_ring *ring = engine->buffer;
 485         int ret = 0;
 486
 487         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
 488
 489         if (!stop_ring(engine)) {
 490                 /* G45 ring initialization often fails to reset head to zero */
 491                 DRM_DEBUG_KMS("%s head not reset to zero "
 492                               "ctl %08x head %08x tail %08x start %08x\n",
 493                               engine->name,
 494                               I915_READ_CTL(engine),
 495                               I915_READ_HEAD(engine),
 496                               I915_READ_TAIL(engine),
 497                               I915_READ_START(engine));
 498
 499                 if (!stop_ring(engine)) {
 500                         DRM_ERROR("failed to set %s head to zero "
 501                                   "ctl %08x head %08x tail %08x start %08x\n",
 502                                   engine->name,
 503                                   I915_READ_CTL(engine),
 504                                   I915_READ_HEAD(engine),
 505                                   I915_READ_TAIL(engine),
 506                                   I915_READ_START(engine));
 507                         ret = -EIO;
 508                         goto out;
 509                 }
 510         }
 511
 512         if (HWS_NEEDS_PHYSICAL(dev_priv))
 513                 ring_setup_phys_status_page(engine);
 514         else
 515                 intel_ring_setup_status_page(engine);
 516
 517         intel_engine_reset_breadcrumbs(engine);
 518
 519         /* Enforce ordering by reading HEAD register back */
 520         I915_READ_HEAD(engine);
 521
 522         /* Initialize the ring. This must happen _after_ we've cleared the ring
 523          * registers with the above sequence (the readback of the HEAD registers
 524          * also enforces ordering), otherwise the hw might lose the new ring
 525          * register values. */
 526         I915_WRITE_START(engine, i915_ggtt_offset(ring->vma));
 527
 528         /* WaClearRingBufHeadRegAtInit:ctg,elk */
 529         if (I915_READ_HEAD(engine))
 530                 DRM_DEBUG("%s initialization failed [head=%08x], fudging\n",
 531                           engine->name, I915_READ_HEAD(engine));
 532
 533         intel_ring_update_space(ring);
 534         I915_WRITE_HEAD(engine, ring->head);
 535         I915_WRITE_TAIL(engine, ring->tail);
 536         (void)I915_READ_TAIL(engine);
 537
 538         I915_WRITE_CTL(engine, RING_CTL_SIZE(ring->size) | RING_VALID);
 539
 540         /* If the head is still not zero, the ring is dead */
 541         if (intel_wait_for_register_fw(dev_priv, RING_CTL(engine->mmio_base),
 542                                        RING_VALID, RING_VALID,
 543                                        50)) {
 544                 DRM_ERROR("%s initialization failed "
 545                           "ctl %08x (valid? %d) head %08x [%08x] tail %08x [%08x] start %08x [expected %08x]\n",
 546                           engine->name,
 547                           I915_READ_CTL(engine),
 548                           I915_READ_CTL(engine) & RING_VALID,
 549                           I915_READ_HEAD(engine), ring->head,
 550                           I915_READ_TAIL(engine), ring->tail,
 551                           I915_READ_START(engine),
 552                           i915_ggtt_offset(ring->vma));
 553                 ret = -EIO;
 554                 goto out;
 555         }
 556
 557         intel_engine_init_hangcheck(engine);
 558
 559 out:
 560         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
 561
 562         return ret;
 563 }
 564
 565 static void reset_ring_common(struct intel_engine_cs *engine,
 566                               struct drm_i915_gem_request *request)
 567 {
 568         /* Try to restore the logical GPU state to match the continuation
 569          * of the request queue. If we skip the context/PD restore, then
 570          * the next request may try to execute assuming that its context
 571          * is valid and loaded on the GPU and so may try to access invalid
 572          * memory, prompting repeated GPU hangs.
 573          *
 574          * If the request was guilty, we still restore the logical state
 575          * in case the next request requires it (e.g. the aliasing ppgtt),
 576          * but skip over the hung batch.
 577          *
 578          * If the request was innocent, we try to replay the request with
 579          * the restored context.
 580          */
 581         if (request) {
 582                 struct drm_i915_private *dev_priv = request->i915;
 583                 struct intel_context *ce = &request->ctx->engine[engine->id];
 584                 struct i915_hw_ppgtt *ppgtt;
 585
 586                 /* FIXME consider gen8 reset */
 587
 588                 if (ce->state) {
 589                         I915_WRITE(CCID,
 590                                    i915_ggtt_offset(ce->state) |
 591                                    BIT(8) /* must be set! */ |
 592                                    CCID_EXTENDED_STATE_SAVE |
 593                                    CCID_EXTENDED_STATE_RESTORE |
 594                                    CCID_EN);
 595                 }
 596
 597                 ppgtt = request->ctx->ppgtt ?: engine->i915->mm.aliasing_ppgtt;
 598                 if (ppgtt) {
 599                         u32 pd_offset = ppgtt->pd.base.ggtt_offset << 10;
 600
 601                         I915_WRITE(RING_PP_DIR_DCLV(engine), PP_DIR_DCLV_2G);
 602                         I915_WRITE(RING_PP_DIR_BASE(engine), pd_offset);
 603
 604                         /* Wait for the PD reload to complete */
 605                         if (intel_wait_for_register(dev_priv,
 606                                                     RING_PP_DIR_BASE(engine),
 607                                                     BIT(0), 0,
 608                                                     10))
 609                                 DRM_ERROR("Wait for reload of ppgtt page-directory timed out\n");
 610
 611                         ppgtt->pd_dirty_rings &= ~intel_engine_flag(engine);
 612                 }
 613
 614                 /* If the rq hung, jump to its breadcrumb and skip the batch */
 615                 if (request->fence.error == -EIO)
 616                         request->ring->head = request->postfix;
 617         } else {
 618                 engine->legacy_active_context = NULL;
 619         }
 620 }
 621
 622 static int intel_rcs_ctx_init(struct drm_i915_gem_request *req)
 623 {
 624         int ret;
 625
 626         ret = intel_ring_workarounds_emit(req);
 627         if (ret != 0)
 628                 return ret;
 629
 630         ret = i915_gem_render_state_emit(req);
 631         if (ret)
 632                 return ret;
 633
 634         return 0;
 635 }
 636
 637 static int init_render_ring(struct intel_engine_cs *engine)
 638 {
 639         struct drm_i915_private *dev_priv = engine->i915;
 640         int ret = init_ring_common(engine);
 641         if (ret)
 642                 return ret;
 643
 644         /* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
 645         if (IS_GEN(dev_priv, 4, 6))
 646                 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH));
 647
 648         /* We need to disable the AsyncFlip performance optimisations in order
 649          * to use MI_WAIT_FOR_EVENT within the CS. It should already be
 650          * programmed to '1' on all products.
 651          *
 652          * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv
 653          */
 654         if (IS_GEN(dev_priv, 6, 7))
 655                 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
 656
 657         /* Required for the hardware to program scanline values for waiting */
 658         /* WaEnableFlushTlbInvalidationMode:snb */
 659         if (IS_GEN6(dev_priv))
 660                 I915_WRITE(GFX_MODE,
 661                            _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT));
 662
 663         /* WaBCSVCSTlbInvalidationMode:ivb,vlv,hsw */
 664         if (IS_GEN7(dev_priv))
 665                 I915_WRITE(GFX_MODE_GEN7,
 666                            _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT) |
 667                            _MASKED_BIT_ENABLE(GFX_REPLAY_MODE));
 668
 669         if (IS_GEN6(dev_priv)) {
 670                 /* From the Sandybridge PRM, volume 1 part 3, page 24:
 671                  * "If this bit is set, STCunit will have LRA as replacement
 672                  *  policy. [...] This bit must be reset.  LRA replacement
 673                  *  policy is not supported."
 674                  */
 675                 I915_WRITE(CACHE_MODE_0,
 676                            _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
 677         }
 678
 679         if (IS_GEN(dev_priv, 6, 7))
 680                 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
 681
 682         if (INTEL_INFO(dev_priv)->gen >= 6)
 683                 I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
 684
 685         return init_workarounds_ring(engine);
 686 }
 687
 688 static void render_ring_cleanup(struct intel_engine_cs *engine)
 689 {
 690         struct drm_i915_private *dev_priv = engine->i915;
 691
 692         i915_vma_unpin_and_release(&dev_priv->semaphore);
 693 }
 694
 695 static u32 *gen8_rcs_signal(struct drm_i915_gem_request *req, u32 *cs)
 696 {
 697         struct drm_i915_private *dev_priv = req->i915;
 698         struct intel_engine_cs *waiter;
 699         enum intel_engine_id id;
 700
 701         for_each_engine(waiter, dev_priv, id) {
 702                 u64 gtt_offset = req->engine->semaphore.signal_ggtt[id];
 703                 if (gtt_offset == MI_SEMAPHORE_SYNC_INVALID)
 704                         continue;
 705
 706                 *cs++ = GFX_OP_PIPE_CONTROL(6);
 707                 *cs++ = PIPE_CONTROL_GLOBAL_GTT_IVB | PIPE_CONTROL_QW_WRITE |
 708                         PIPE_CONTROL_CS_STALL;
 709                 *cs++ = lower_32_bits(gtt_offset);
 710                 *cs++ = upper_32_bits(gtt_offset);
 711                 *cs++ = req->global_seqno;
 712                 *cs++ = 0;
 713                 *cs++ = MI_SEMAPHORE_SIGNAL |
 714                         MI_SEMAPHORE_TARGET(waiter->hw_id);
 715                 *cs++ = 0;
 716         }
 717
 718         return cs;
 719 }
 720
 721 static u32 *gen8_xcs_signal(struct drm_i915_gem_request *req, u32 *cs)
 722 {
 723         struct drm_i915_private *dev_priv = req->i915;
 724         struct intel_engine_cs *waiter;
 725         enum intel_engine_id id;
 726
 727         for_each_engine(waiter, dev_priv, id) {
 728                 u64 gtt_offset = req->engine->semaphore.signal_ggtt[id];
 729                 if (gtt_offset == MI_SEMAPHORE_SYNC_INVALID)
 730                         continue;
 731
 732                 *cs++ = (MI_FLUSH_DW + 1) | MI_FLUSH_DW_OP_STOREDW;
 733                 *cs++ = lower_32_bits(gtt_offset) | MI_FLUSH_DW_USE_GTT;
 734                 *cs++ = upper_32_bits(gtt_offset);
 735                 *cs++ = req->global_seqno;
 736                 *cs++ = MI_SEMAPHORE_SIGNAL |
 737                         MI_SEMAPHORE_TARGET(waiter->hw_id);
 738                 *cs++ = 0;
 739         }
 740
 741         return cs;
 742 }
 743
 744 static u32 *gen6_signal(struct drm_i915_gem_request *req, u32 *cs)
 745 {
 746         struct drm_i915_private *dev_priv = req->i915;
 747         struct intel_engine_cs *engine;
 748         enum intel_engine_id id;
 749         int num_rings = 0;
 750
 751         for_each_engine(engine, dev_priv, id) {
 752                 i915_reg_t mbox_reg;
 753
 754                 if (!(BIT(engine->hw_id) & GEN6_SEMAPHORES_MASK))
 755                         continue;
 756
 757                 mbox_reg = req->engine->semaphore.mbox.signal[engine->hw_id];
 758                 if (i915_mmio_reg_valid(mbox_reg)) {
 759                         *cs++ = MI_LOAD_REGISTER_IMM(1);
 760                         *cs++ = i915_mmio_reg_offset(mbox_reg);
 761                         *cs++ = req->global_seqno;
 762                         num_rings++;
 763                 }
 764         }
 765         if (num_rings & 1)
 766                 *cs++ = MI_NOOP;
 767
 768         return cs;
 769 }
 770
 771 static void i9xx_submit_request(struct drm_i915_gem_request *request)
 772 {
 773         struct drm_i915_private *dev_priv = request->i915;
 774
 775         i915_gem_request_submit(request);
 776
 777         I915_WRITE_TAIL(request->engine,
 778                         intel_ring_set_tail(request->ring, request->tail));
 779 }
 780
 781 static void i9xx_emit_breadcrumb(struct drm_i915_gem_request *req, u32 *cs)
 782 {
 783         *cs++ = MI_STORE_DWORD_INDEX;
 784         *cs++ = I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT;
 785         *cs++ = req->global_seqno;
 786         *cs++ = MI_USER_INTERRUPT;
 787
 788         req->tail = intel_ring_offset(req, cs);
 789         assert_ring_tail_valid(req->ring, req->tail);
 790 }
 791
 792 static const int i9xx_emit_breadcrumb_sz = 4;
 793
 794 /**
 795  * gen6_sema_emit_breadcrumb - Update the semaphore mailbox registers
 796  *
 797  * @request - request to write to the ring
 798  *
 799  * Update the mailbox registers in the *other* rings with the current seqno.
 800  * This acts like a signal in the canonical semaphore.
 801  */
 802 static void gen6_sema_emit_breadcrumb(struct drm_i915_gem_request *req, u32 *cs)
 803 {
 804         return i9xx_emit_breadcrumb(req,
 805                                     req->engine->semaphore.signal(req, cs));
 806 }
 807
 808 static void gen8_render_emit_breadcrumb(struct drm_i915_gem_request *req,
 809                                         u32 *cs)
 810 {
 811         struct intel_engine_cs *engine = req->engine;
 812
 813         if (engine->semaphore.signal)
 814                 cs = engine->semaphore.signal(req, cs);
 815
 816         *cs++ = GFX_OP_PIPE_CONTROL(6);
 817         *cs++ = PIPE_CONTROL_GLOBAL_GTT_IVB | PIPE_CONTROL_CS_STALL |
 818                 PIPE_CONTROL_QW_WRITE;
 819         *cs++ = intel_hws_seqno_address(engine);
 820         *cs++ = 0;
 821         *cs++ = req->global_seqno;
 822         /* We're thrashing one dword of HWS. */
 823         *cs++ = 0;
 824         *cs++ = MI_USER_INTERRUPT;
 825         *cs++ = MI_NOOP;
 826
 827         req->tail = intel_ring_offset(req, cs);
 828         assert_ring_tail_valid(req->ring, req->tail);
 829 }
 830
 831 static const int gen8_render_emit_breadcrumb_sz = 8;
 832
 833 /**
 834  * intel_ring_sync - sync the waiter to the signaller on seqno
 835  *
 836  * @waiter - ring that is waiting
 837  * @signaller - ring which has, or will signal
 838  * @seqno - seqno which the waiter will block on
 839  */
 840
 841 static int
 842 gen8_ring_sync_to(struct drm_i915_gem_request *req,
 843                   struct drm_i915_gem_request *signal)
 844 {
 845         struct drm_i915_private *dev_priv = req->i915;
 846         u64 offset = GEN8_WAIT_OFFSET(req->engine, signal->engine->id);
 847         struct i915_hw_ppgtt *ppgtt;
 848         u32 *cs;
 849
 850         cs = intel_ring_begin(req, 4);
 851         if (IS_ERR(cs))
 852                 return PTR_ERR(cs);
 853
 854         *cs++ = MI_SEMAPHORE_WAIT | MI_SEMAPHORE_GLOBAL_GTT |
 855                 MI_SEMAPHORE_SAD_GTE_SDD;
 856         *cs++ = signal->global_seqno;
 857         *cs++ = lower_32_bits(offset);
 858         *cs++ = upper_32_bits(offset);
 859         intel_ring_advance(req, cs);
 860
 861         /* When the !RCS engines idle waiting upon a semaphore, they lose their
 862          * pagetables and we must reload them before executing the batch.
 863          * We do this on the i915_switch_context() following the wait and
 864          * before the dispatch.
 865          */
 866         ppgtt = req->ctx->ppgtt;
 867         if (ppgtt && req->engine->id != RCS)
 868                 ppgtt->pd_dirty_rings |= intel_engine_flag(req->engine);
 869         return 0;
 870 }
 871
 872 static int
 873 gen6_ring_sync_to(struct drm_i915_gem_request *req,
 874                   struct drm_i915_gem_request *signal)
 875 {
 876         u32 dw1 = MI_SEMAPHORE_MBOX |
 877                   MI_SEMAPHORE_COMPARE |
 878                   MI_SEMAPHORE_REGISTER;
 879         u32 wait_mbox = signal->engine->semaphore.mbox.wait[req->engine->hw_id];
 880         u32 *cs;
 881
 882         WARN_ON(wait_mbox == MI_SEMAPHORE_SYNC_INVALID);
 883
 884         cs = intel_ring_begin(req, 4);
 885         if (IS_ERR(cs))
 886                 return PTR_ERR(cs);
 887
 888         *cs++ = dw1 | wait_mbox;
 889         /* Throughout all of the GEM code, seqno passed implies our current
 890          * seqno is >= the last seqno executed. However for hardware the
 891          * comparison is strictly greater than.
 892          */
 893         *cs++ = signal->global_seqno - 1;
 894         *cs++ = 0;
 895         *cs++ = MI_NOOP;
 896         intel_ring_advance(req, cs);
 897
 898         return 0;
 899 }
 900
 901 static void
 902 gen5_seqno_barrier(struct intel_engine_cs *engine)
 903 {
 904         /* MI_STORE are internally buffered by the GPU and not flushed
 905          * either by MI_FLUSH or SyncFlush or any other combination of
 906          * MI commands.
 907          *
 908          * "Only the submission of the store operation is guaranteed.
 909          * The write result will be complete (coherent) some time later
 910          * (this is practically a finite period but there is no guaranteed
 911          * latency)."
 912          *
 913          * Empirically, we observe that we need a delay of at least 75us to
 914          * be sure that the seqno write is visible by the CPU.
 915          */
 916         usleep_range(125, 250);
 917 }
 918
 919 static void
 920 gen6_seqno_barrier(struct intel_engine_cs *engine)
 921 {
 922         struct drm_i915_private *dev_priv = engine->i915;
 923
 924         /* Workaround to force correct ordering between irq and seqno writes on
 925          * ivb (and maybe also on snb) by reading from a CS register (like
 926          * ACTHD) before reading the status page.
 927          *
 928          * Note that this effectively stalls the read by the time it takes to
 929          * do a memory transaction, which more or less ensures that the write
 930          * from the GPU has sufficient time to invalidate the CPU cacheline.
 931          * Alternatively we could delay the interrupt from the CS ring to give
 932          * the write time to land, but that would incur a delay after every
 933          * batch i.e. much more frequent than a delay when waiting for the
 934          * interrupt (with the same net latency).
 935          *
 936          * Also note that to prevent whole machine hangs on gen7, we have to
 937          * take the spinlock to guard against concurrent cacheline access.
 938          */
 939         spin_lock_irq(&dev_priv->uncore.lock);
 940         POSTING_READ_FW(RING_ACTHD(engine->mmio_base));
 941         spin_unlock_irq(&dev_priv->uncore.lock);
 942 }
 943
 944 static void
 945 gen5_irq_enable(struct intel_engine_cs *engine)
 946 {
 947         gen5_enable_gt_irq(engine->i915, engine->irq_enable_mask);
 948 }
 949
 950 static void
 951 gen5_irq_disable(struct intel_engine_cs *engine)
 952 {
 953         gen5_disable_gt_irq(engine->i915, engine->irq_enable_mask);
 954 }
 955
 956 static void
 957 i9xx_irq_enable(struct intel_engine_cs *engine)
 958 {
 959         struct drm_i915_private *dev_priv = engine->i915;
 960
 961         dev_priv->irq_mask &= ~engine->irq_enable_mask;
 962         I915_WRITE(IMR, dev_priv->irq_mask);
 963         POSTING_READ_FW(RING_IMR(engine->mmio_base));
 964 }
 965
 966 static void
 967 i9xx_irq_disable(struct intel_engine_cs *engine)
 968 {
 969         struct drm_i915_private *dev_priv = engine->i915;
 970
 971         dev_priv->irq_mask |= engine->irq_enable_mask;
 972         I915_WRITE(IMR, dev_priv->irq_mask);
 973 }
 974
 975 static void
 976 i8xx_irq_enable(struct intel_engine_cs *engine)
 977 {
 978         struct drm_i915_private *dev_priv = engine->i915;
 979
 980         dev_priv->irq_mask &= ~engine->irq_enable_mask;
 981         I915_WRITE16(IMR, dev_priv->irq_mask);
 982         POSTING_READ16(RING_IMR(engine->mmio_base));
 983 }
 984
 985 static void
 986 i8xx_irq_disable(struct intel_engine_cs *engine)
 987 {
 988         struct drm_i915_private *dev_priv = engine->i915;
 989
 990         dev_priv->irq_mask |= engine->irq_enable_mask;
 991         I915_WRITE16(IMR, dev_priv->irq_mask);
 992 }
 993
 994 static int
 995 bsd_ring_flush(struct drm_i915_gem_request *req, u32 mode)
 996 {
 997         u32 *cs;
 998
 999         cs = intel_ring_begin(req, 2);
1000         if (IS_ERR(cs))
1001                 return PTR_ERR(cs);
1002
1003         *cs++ = MI_FLUSH;
1004         *cs++ = MI_NOOP;
1005         intel_ring_advance(req, cs);
1006         return 0;
1007 }
1008
1009 static void
1010 gen6_irq_enable(struct intel_engine_cs *engine)
1011 {
1012         struct drm_i915_private *dev_priv = engine->i915;
1013
1014         I915_WRITE_IMR(engine,
1015                        ~(engine->irq_enable_mask |
1016                          engine->irq_keep_mask));
1017         gen5_enable_gt_irq(dev_priv, engine->irq_enable_mask);
1018 }
1019
1020 static void
1021 gen6_irq_disable(struct intel_engine_cs *engine)
1022 {
1023         struct drm_i915_private *dev_priv = engine->i915;
1024
1025         I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
1026         gen5_disable_gt_irq(dev_priv, engine->irq_enable_mask);
1027 }
1028
1029 static void
1030 hsw_vebox_irq_enable(struct intel_engine_cs *engine)
1031 {
1032         struct drm_i915_private *dev_priv = engine->i915;
1033
1034         I915_WRITE_IMR(engine, ~engine->irq_enable_mask);
1035         gen6_unmask_pm_irq(dev_priv, engine->irq_enable_mask);
1036 }
1037
1038 static void
1039 hsw_vebox_irq_disable(struct intel_engine_cs *engine)
1040 {
1041         struct drm_i915_private *dev_priv = engine->i915;
1042
1043         I915_WRITE_IMR(engine, ~0);
1044         gen6_mask_pm_irq(dev_priv, engine->irq_enable_mask);
1045 }
1046
1047 static void
1048 gen8_irq_enable(struct intel_engine_cs *engine)
1049 {
1050         struct drm_i915_private *dev_priv = engine->i915;
1051
1052         I915_WRITE_IMR(engine,
1053                        ~(engine->irq_enable_mask |
1054                          engine->irq_keep_mask));
1055         POSTING_READ_FW(RING_IMR(engine->mmio_base));
1056 }
1057
1058 static void
1059 gen8_irq_disable(struct intel_engine_cs *engine)
1060 {
1061         struct drm_i915_private *dev_priv = engine->i915;
1062
1063         I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
1064 }
1065
1066 static int
1067 i965_emit_bb_start(struct drm_i915_gem_request *req,
1068                    u64 offset, u32 length,
1069                    unsigned int dispatch_flags)
1070 {
1071         u32 *cs;
1072
1073         cs = intel_ring_begin(req, 2);
1074         if (IS_ERR(cs))
1075                 return PTR_ERR(cs);
1076
1077         *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT | (dispatch_flags &
1078                 I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE_I965);
1079         *cs++ = offset;
1080         intel_ring_advance(req, cs);
1081
1082         return 0;
1083 }
1084
1085 /* Just userspace ABI convention to limit the wa batch bo to a resonable size */
1086 #define I830_BATCH_LIMIT (256*1024)
1087 #define I830_TLB_ENTRIES (2)
1088 #define I830_WA_SIZE max(I830_TLB_ENTRIES*4096, I830_BATCH_LIMIT)
1089 static int
1090 i830_emit_bb_start(struct drm_i915_gem_request *req,
1091                    u64 offset, u32 len,
1092                    unsigned int dispatch_flags)
1093 {
1094         u32 *cs, cs_offset = i915_ggtt_offset(req->engine->scratch);
1095
1096         cs = intel_ring_begin(req, 6);
1097         if (IS_ERR(cs))
1098                 return PTR_ERR(cs);
1099
1100         /* Evict the invalid PTE TLBs */
1101         *cs++ = COLOR_BLT_CMD | BLT_WRITE_RGBA;
1102         *cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | 4096;
1103         *cs++ = I830_TLB_ENTRIES << 16 | 4; /* load each page */
1104         *cs++ = cs_offset;
1105         *cs++ = 0xdeadbeef;
1106         *cs++ = MI_NOOP;
1107         intel_ring_advance(req, cs);
1108
1109         if ((dispatch_flags & I915_DISPATCH_PINNED) == 0) {
1110                 if (len > I830_BATCH_LIMIT)
1111                         return -ENOSPC;
1112
1113                 cs = intel_ring_begin(req, 6 + 2);
1114                 if (IS_ERR(cs))
1115                         return PTR_ERR(cs);
1116
1117                 /* Blit the batch (which has now all relocs applied) to the
1118                  * stable batch scratch bo area (so that the CS never
1119                  * stumbles over its tlb invalidation bug) ...
1120                  */
1121                 *cs++ = SRC_COPY_BLT_CMD | BLT_WRITE_RGBA;
1122                 *cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | 4096;
1123                 *cs++ = DIV_ROUND_UP(len, 4096) << 16 | 4096;
1124                 *cs++ = cs_offset;
1125                 *cs++ = 4096;
1126                 *cs++ = offset;
1127
1128                 *cs++ = MI_FLUSH;
1129                 *cs++ = MI_NOOP;
1130                 intel_ring_advance(req, cs);
1131
1132                 /* ... and execute it. */
1133                 offset = cs_offset;
1134         }
1135
1136         cs = intel_ring_begin(req, 2);
1137         if (IS_ERR(cs))
1138                 return PTR_ERR(cs);
1139
1140         *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
1141         *cs++ = offset | (dispatch_flags & I915_DISPATCH_SECURE ? 0 :
1142                 MI_BATCH_NON_SECURE);
1143         intel_ring_advance(req, cs);
1144
1145         return 0;
1146 }
1147
1148 static int
1149 i915_emit_bb_start(struct drm_i915_gem_request *req,
1150                    u64 offset, u32 len,
1151                    unsigned int dispatch_flags)
1152 {
1153         u32 *cs;
1154
1155         cs = intel_ring_begin(req, 2);
1156         if (IS_ERR(cs))
1157                 return PTR_ERR(cs);
1158
1159         *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
1160         *cs++ = offset | (dispatch_flags & I915_DISPATCH_SECURE ? 0 :
1161                 MI_BATCH_NON_SECURE);
1162         intel_ring_advance(req, cs);
1163
1164         return 0;
1165 }
1166
1167 static void cleanup_phys_status_page(struct intel_engine_cs *engine)
1168 {
1169         struct drm_i915_private *dev_priv = engine->i915;
1170
1171         if (!dev_priv->status_page_dmah)
1172                 return;
1173
1174         drm_pci_free(&dev_priv->drm, dev_priv->status_page_dmah);
1175         engine->status_page.page_addr = NULL;
1176 }
1177
1178 static void cleanup_status_page(struct intel_engine_cs *engine)
1179 {
1180         struct i915_vma *vma;
1181         struct drm_i915_gem_object *obj;
1182
1183         vma = fetch_and_zero(&engine->status_page.vma);
1184         if (!vma)
1185                 return;
1186
1187         obj = vma->obj;
1188
1189         i915_vma_unpin(vma);
1190         i915_vma_close(vma);
1191
1192         i915_gem_object_unpin_map(obj);
1193         __i915_gem_object_release_unless_active(obj);
1194 }
1195
1196 static int init_status_page(struct intel_engine_cs *engine)
1197 {
1198         struct drm_i915_gem_object *obj;
1199         struct i915_vma *vma;
1200         unsigned int flags;
1201         void *vaddr;
1202         int ret;
1203
1204         obj = i915_gem_object_create_internal(engine->i915, PAGE_SIZE);
1205         if (IS_ERR(obj)) {
1206                 DRM_ERROR("Failed to allocate status page\n");
1207                 return PTR_ERR(obj);
1208         }
1209
1210         ret = i915_gem_object_set_cache_level(obj, I915_CACHE_LLC);
1211         if (ret)
1212                 goto err;
1213
1214         vma = i915_vma_instance(obj, &engine->i915->ggtt.base, NULL);
1215         if (IS_ERR(vma)) {
1216                 ret = PTR_ERR(vma);
1217                 goto err;
1218         }
1219
1220         flags = PIN_GLOBAL;
1221         if (!HAS_LLC(engine->i915))
1222                 /* On g33, we cannot place HWS above 256MiB, so
1223                  * restrict its pinning to the low mappable arena.
1224                  * Though this restriction is not documented for
1225                  * gen4, gen5, or byt, they also behave similarly
1226                  * and hang if the HWS is placed at the top of the
1227                  * GTT. To generalise, it appears that all !llc
1228                  * platforms have issues with us placing the HWS
1229                  * above the mappable region (even though we never
1230                  * actualy map it).
1231                  */
1232                 flags |= PIN_MAPPABLE;
1233         ret = i915_vma_pin(vma, 0, 4096, flags);
1234         if (ret)
1235                 goto err;
1236
1237         vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
1238         if (IS_ERR(vaddr)) {
1239                 ret = PTR_ERR(vaddr);
1240                 goto err_unpin;
1241         }
1242
1243         engine->status_page.vma = vma;
1244         engine->status_page.ggtt_offset = i915_ggtt_offset(vma);
1245         engine->status_page.page_addr = memset(vaddr, 0, PAGE_SIZE);
1246
1247         DRM_DEBUG_DRIVER("%s hws offset: 0x%08x\n",
1248                          engine->name, i915_ggtt_offset(vma));
1249         return 0;
1250
1251 err_unpin:
1252         i915_vma_unpin(vma);
1253 err:
1254         i915_gem_object_put(obj);
1255         return ret;
1256 }
1257
1258 static int init_phys_status_page(struct intel_engine_cs *engine)
1259 {
1260         struct drm_i915_private *dev_priv = engine->i915;
1261
1262         dev_priv->status_page_dmah =
1263                 drm_pci_alloc(&dev_priv->drm, PAGE_SIZE, PAGE_SIZE);
1264         if (!dev_priv->status_page_dmah)
1265                 return -ENOMEM;
1266
1267         engine->status_page.page_addr = dev_priv->status_page_dmah->vaddr;
1268         memset(engine->status_page.page_addr, 0, PAGE_SIZE);
1269
1270         return 0;
1271 }
1272
1273 int intel_ring_pin(struct intel_ring *ring, unsigned int offset_bias)
1274 {
1275         unsigned int flags;
1276         enum i915_map_type map;
1277         struct i915_vma *vma = ring->vma;
1278         void *addr;
1279         int ret;
1280
1281         GEM_BUG_ON(ring->vaddr);
1282
1283         map = HAS_LLC(ring->engine->i915) ? I915_MAP_WB : I915_MAP_WC;
1284
1285         flags = PIN_GLOBAL;
1286         if (offset_bias)
1287                 flags |= PIN_OFFSET_BIAS | offset_bias;
1288         if (vma->obj->stolen)
1289                 flags |= PIN_MAPPABLE;
1290
1291         if (!(vma->flags & I915_VMA_GLOBAL_BIND)) {
1292                 if (flags & PIN_MAPPABLE || map == I915_MAP_WC)
1293                         ret = i915_gem_object_set_to_gtt_domain(vma->obj, true);
1294                 else
1295                         ret = i915_gem_object_set_to_cpu_domain(vma->obj, true);
1296                 if (unlikely(ret))
1297                         return ret;
1298         }
1299
1300         ret = i915_vma_pin(vma, 0, PAGE_SIZE, flags);
1301         if (unlikely(ret))
1302                 return ret;
1303
1304         if (i915_vma_is_map_and_fenceable(vma))
1305                 addr = (void __force *)i915_vma_pin_iomap(vma);
1306         else
1307                 addr = i915_gem_object_pin_map(vma->obj, map);
1308         if (IS_ERR(addr))
1309                 goto err;
1310
1311         ring->vaddr = addr;
1312         return 0;
1313
1314 err:
1315         i915_vma_unpin(vma);
1316         return PTR_ERR(addr);
1317 }
1318
1319 void intel_ring_reset(struct intel_ring *ring, u32 tail)
1320 {
1321         GEM_BUG_ON(!list_empty(&ring->request_list));
1322         ring->tail = tail;
1323         ring->head = tail;
1324         ring->emit = tail;
1325         intel_ring_update_space(ring);
1326 }
1327
1328 void intel_ring_unpin(struct intel_ring *ring)
1329 {
1330         GEM_BUG_ON(!ring->vma);
1331         GEM_BUG_ON(!ring->vaddr);
1332
1333         /* Discard any unused bytes beyond that submitted to hw. */
1334         intel_ring_reset(ring, ring->tail);
1335
1336         if (i915_vma_is_map_and_fenceable(ring->vma))
1337                 i915_vma_unpin_iomap(ring->vma);
1338         else
1339                 i915_gem_object_unpin_map(ring->vma->obj);
1340         ring->vaddr = NULL;
1341
1342         i915_vma_unpin(ring->vma);
1343 }
1344
1345 static struct i915_vma *
1346 intel_ring_create_vma(struct drm_i915_private *dev_priv, int size)
1347 {
1348         struct drm_i915_gem_object *obj;
1349         struct i915_vma *vma;
1350
1351         obj = i915_gem_object_create_stolen(dev_priv, size);
1352         if (!obj)
1353                 obj = i915_gem_object_create(dev_priv, size);
1354         if (IS_ERR(obj))
1355                 return ERR_CAST(obj);
1356
1357         /* mark ring buffers as read-only from GPU side by default */
1358         obj->gt_ro = 1;
1359
1360         vma = i915_vma_instance(obj, &dev_priv->ggtt.base, NULL);
1361         if (IS_ERR(vma))
1362                 goto err;
1363
1364         return vma;
1365
1366 err:
1367         i915_gem_object_put(obj);
1368         return vma;
1369 }
1370
1371 struct intel_ring *
1372 intel_engine_create_ring(struct intel_engine_cs *engine, int size)
1373 {
1374         struct intel_ring *ring;
1375         struct i915_vma *vma;
1376
1377         GEM_BUG_ON(!is_power_of_2(size));
1378         GEM_BUG_ON(RING_CTL_SIZE(size) & ~RING_NR_PAGES);
1379
1380         ring = kzalloc(sizeof(*ring), GFP_KERNEL);
1381         if (!ring)
1382                 return ERR_PTR(-ENOMEM);
1383
1384         ring->engine = engine;
1385
1386         INIT_LIST_HEAD(&ring->request_list);
1387
1388         ring->size = size;
1389         /* Workaround an erratum on the i830 which causes a hang if
1390          * the TAIL pointer points to within the last 2 cachelines
1391          * of the buffer.
1392          */
1393         ring->effective_size = size;
1394         if (IS_I830(engine->i915) || IS_I845G(engine->i915))
1395                 ring->effective_size -= 2 * CACHELINE_BYTES;
1396
1397         intel_ring_update_space(ring);
1398
1399         vma = intel_ring_create_vma(engine->i915, size);
1400         if (IS_ERR(vma)) {
1401                 kfree(ring);
1402                 return ERR_CAST(vma);
1403         }
1404         ring->vma = vma;
1405
1406         return ring;
1407 }
1408
1409 void
1410 intel_ring_free(struct intel_ring *ring)
1411 {
1412         struct drm_i915_gem_object *obj = ring->vma->obj;
1413
1414         i915_vma_close(ring->vma);
1415         __i915_gem_object_release_unless_active(obj);
1416
1417         kfree(ring);
1418 }
1419
1420 static int context_pin(struct i915_gem_context *ctx)
1421 {
1422         struct i915_vma *vma = ctx->engine[RCS].state;
1423         int ret;
1424
1425         /* Clear this page out of any CPU caches for coherent swap-in/out.
1426          * We only want to do this on the first bind so that we do not stall
1427          * on an active context (which by nature is already on the GPU).
1428          */
1429         if (!(vma->flags & I915_VMA_GLOBAL_BIND)) {
1430                 ret = i915_gem_object_set_to_gtt_domain(vma->obj, false);
1431                 if (ret)
1432                         return ret;
1433         }
1434
1435         return i915_vma_pin(vma, 0, I915_GTT_MIN_ALIGNMENT,
1436                             PIN_GLOBAL | PIN_HIGH);
1437 }
1438
1439 static int intel_ring_context_pin(struct intel_engine_cs *engine,
1440                                   struct i915_gem_context *ctx)
1441 {
1442         struct intel_context *ce = &ctx->engine[engine->id];
1443         int ret;
1444
1445         lockdep_assert_held(&ctx->i915->drm.struct_mutex);
1446
1447         if (ce->pin_count++)
1448                 return 0;
1449         GEM_BUG_ON(!ce->pin_count); /* no overflow please! */
1450
1451         if (ce->state) {
1452                 ret = context_pin(ctx);
1453                 if (ret)
1454                         goto error;
1455
1456                 ce->state->obj->mm.dirty = true;
1457         }
1458
1459         /* The kernel context is only used as a placeholder for flushing the
1460          * active context. It is never used for submitting user rendering and
1461          * as such never requires the golden render context, and so we can skip
1462          * emitting it when we switch to the kernel context. This is required
1463          * as during eviction we cannot allocate and pin the renderstate in
1464          * order to initialise the context.
1465          */
1466         if (i915_gem_context_is_kernel(ctx))
1467                 ce->initialised = true;
1468
1469         i915_gem_context_get(ctx);
1470         return 0;
1471
1472 error:
1473         ce->pin_count = 0;
1474         return ret;
1475 }
1476
1477 static void intel_ring_context_unpin(struct intel_engine_cs *engine,
1478                                      struct i915_gem_context *ctx)
1479 {
1480         struct intel_context *ce = &ctx->engine[engine->id];
1481
1482         lockdep_assert_held(&ctx->i915->drm.struct_mutex);
1483         GEM_BUG_ON(ce->pin_count == 0);
1484
1485         if (--ce->pin_count)
1486                 return;
1487
1488         if (ce->state)
1489                 i915_vma_unpin(ce->state);
1490
1491         i915_gem_context_put(ctx);
1492 }
1493
1494 static int intel_init_ring_buffer(struct intel_engine_cs *engine)
1495 {
1496         struct drm_i915_private *dev_priv = engine->i915;
1497         struct intel_ring *ring;
1498         int ret;
1499
1500         WARN_ON(engine->buffer);
1501
1502         intel_engine_setup_common(engine);
1503
1504         ret = intel_engine_init_common(engine);
1505         if (ret)
1506                 goto error;
1507
1508         ring = intel_engine_create_ring(engine, 32 * PAGE_SIZE);
1509         if (IS_ERR(ring)) {
1510                 ret = PTR_ERR(ring);
1511                 goto error;
1512         }
1513
1514         if (HWS_NEEDS_PHYSICAL(dev_priv)) {
1515                 WARN_ON(engine->id != RCS);
1516                 ret = init_phys_status_page(engine);
1517                 if (ret)
1518                         goto error;
1519         } else {
1520                 ret = init_status_page(engine);
1521                 if (ret)
1522                         goto error;
1523         }
1524
1525         /* Ring wraparound at offset 0 sometimes hangs. No idea why. */
1526         ret = intel_ring_pin(ring, I915_GTT_PAGE_SIZE);
1527         if (ret) {
1528                 intel_ring_free(ring);
1529                 goto error;
1530         }
1531         engine->buffer = ring;
1532
1533         return 0;
1534
1535 error:
1536         intel_engine_cleanup(engine);
1537         return ret;
1538 }
1539
1540 void intel_engine_cleanup(struct intel_engine_cs *engine)
1541 {
1542         struct drm_i915_private *dev_priv;
1543
1544         dev_priv = engine->i915;
1545
1546         if (engine->buffer) {
1547                 WARN_ON(INTEL_GEN(dev_priv) > 2 &&
1548                         (I915_READ_MODE(engine) & MODE_IDLE) == 0);
1549
1550                 intel_ring_unpin(engine->buffer);
1551                 intel_ring_free(engine->buffer);
1552                 engine->buffer = NULL;
1553         }
1554
1555         if (engine->cleanup)
1556                 engine->cleanup(engine);
1557
1558         if (HWS_NEEDS_PHYSICAL(dev_priv)) {
1559                 WARN_ON(engine->id != RCS);
1560                 cleanup_phys_status_page(engine);
1561         } else {
1562                 cleanup_status_page(engine);
1563         }
1564
1565         intel_engine_cleanup_common(engine);
1566
1567         engine->i915 = NULL;
1568         dev_priv->engine[engine->id] = NULL;
1569         kfree(engine);
1570 }
1571
1572 void intel_legacy_submission_resume(struct drm_i915_private *dev_priv)
1573 {
1574         struct intel_engine_cs *engine;
1575         enum intel_engine_id id;
1576
1577         /* Restart from the beginning of the rings for convenience */
1578         for_each_engine(engine, dev_priv, id)
1579                 intel_ring_reset(engine->buffer, 0);
1580 }
1581
1582 static int ring_request_alloc(struct drm_i915_gem_request *request)
1583 {
1584         u32 *cs;
1585
1586         GEM_BUG_ON(!request->ctx->engine[request->engine->id].pin_count);
1587
1588         /* Flush enough space to reduce the likelihood of waiting after
1589          * we start building the request - in which case we will just
1590          * have to repeat work.
1591          */
1592         request->reserved_space += LEGACY_REQUEST_SIZE;
1593
1594         GEM_BUG_ON(!request->engine->buffer);
1595         request->ring = request->engine->buffer;
1596
1597         cs = intel_ring_begin(request, 0);
1598         if (IS_ERR(cs))
1599                 return PTR_ERR(cs);
1600
1601         request->reserved_space -= LEGACY_REQUEST_SIZE;
1602         return 0;
1603 }
1604
1605 static int wait_for_space(struct drm_i915_gem_request *req, int bytes)
1606 {
1607         struct intel_ring *ring = req->ring;
1608         struct drm_i915_gem_request *target;
1609         long timeout;
1610
1611         lockdep_assert_held(&req->i915->drm.struct_mutex);
1612
1613         intel_ring_update_space(ring);
1614         if (ring->space >= bytes)
1615                 return 0;
1616
1617         /*
1618          * Space is reserved in the ringbuffer for finalising the request,
1619          * as that cannot be allowed to fail. During request finalisation,
1620          * reserved_space is set to 0 to stop the overallocation and the
1621          * assumption is that then we never need to wait (which has the
1622          * risk of failing with EINTR).
1623          *
1624          * See also i915_gem_request_alloc() and i915_add_request().
1625          */
1626         GEM_BUG_ON(!req->reserved_space);
1627
1628         list_for_each_entry(target, &ring->request_list, ring_link) {
1629                 unsigned space;
1630
1631                 /* Would completion of this request free enough space? */
1632                 space = __intel_ring_space(target->postfix, ring->emit,
1633                                            ring->size);
1634                 if (space >= bytes)
1635                         break;
1636         }
1637
1638         if (WARN_ON(&target->ring_link == &ring->request_list))
1639                 return -ENOSPC;
1640
1641         timeout = i915_wait_request(target,
1642                                     I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
1643                                     MAX_SCHEDULE_TIMEOUT);
1644         if (timeout < 0)
1645                 return timeout;
1646
1647         i915_gem_request_retire_upto(target);
1648
1649         intel_ring_update_space(ring);
1650         GEM_BUG_ON(ring->space < bytes);
1651         return 0;
1652 }
1653
1654 u32 *intel_ring_begin(struct drm_i915_gem_request *req, int num_dwords)
1655 {
1656         struct intel_ring *ring = req->ring;
1657         int remain_actual = ring->size - ring->emit;
1658         int remain_usable = ring->effective_size - ring->emit;
1659         int bytes = num_dwords * sizeof(u32);
1660         int total_bytes, wait_bytes;
1661         bool need_wrap = false;
1662         u32 *cs;
1663
1664         total_bytes = bytes + req->reserved_space;
1665
1666         if (unlikely(bytes > remain_usable)) {
1667                 /*
1668                  * Not enough space for the basic request. So need to flush
1669                  * out the remainder and then wait for base + reserved.
1670                  */
1671                 wait_bytes = remain_actual + total_bytes;
1672                 need_wrap = true;
1673         } else if (unlikely(total_bytes > remain_usable)) {
1674                 /*
1675                  * The base request will fit but the reserved space
1676                  * falls off the end. So we don't need an immediate wrap
1677                  * and only need to effectively wait for the reserved
1678                  * size space from the start of ringbuffer.
1679                  */
1680                 wait_bytes = remain_actual + req->reserved_space;
1681         } else {
1682                 /* No wrapping required, just waiting. */
1683                 wait_bytes = total_bytes;
1684         }
1685
1686         if (wait_bytes > ring->space) {
1687                 int ret = wait_for_space(req, wait_bytes);
1688                 if (unlikely(ret))
1689                         return ERR_PTR(ret);
1690         }
1691
1692         if (unlikely(need_wrap)) {
1693                 GEM_BUG_ON(remain_actual > ring->space);
1694                 GEM_BUG_ON(ring->emit + remain_actual > ring->size);
1695
1696                 /* Fill the tail with MI_NOOP */
1697                 memset(ring->vaddr + ring->emit, 0, remain_actual);
1698                 ring->emit = 0;
1699                 ring->space -= remain_actual;
1700         }
1701
1702         GEM_BUG_ON(ring->emit > ring->size - bytes);
1703         cs = ring->vaddr + ring->emit;
1704         ring->emit += bytes;
1705         ring->space -= bytes;
1706         GEM_BUG_ON(ring->space < 0);
1707
1708         return cs;
1709 }
1710
1711 /* Align the ring tail to a cacheline boundary */
1712 int intel_ring_cacheline_align(struct drm_i915_gem_request *req)
1713 {
1714         int num_dwords =
1715                 (req->ring->emit & (CACHELINE_BYTES - 1)) / sizeof(uint32_t);
1716         u32 *cs;
1717
1718         if (num_dwords == 0)
1719                 return 0;
1720
1721         num_dwords = CACHELINE_BYTES / sizeof(uint32_t) - num_dwords;
1722         cs = intel_ring_begin(req, num_dwords);
1723         if (IS_ERR(cs))
1724                 return PTR_ERR(cs);
1725
1726         while (num_dwords--)
1727                 *cs++ = MI_NOOP;
1728
1729         intel_ring_advance(req, cs);
1730
1731         return 0;
1732 }
1733
1734 static void gen6_bsd_submit_request(struct drm_i915_gem_request *request)
1735 {
1736         struct drm_i915_private *dev_priv = request->i915;
1737
1738         intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
1739
1740        /* Every tail move must follow the sequence below */
1741
1742         /* Disable notification that the ring is IDLE. The GT
1743          * will then assume that it is busy and bring it out of rc6.
1744          */
1745         I915_WRITE_FW(GEN6_BSD_SLEEP_PSMI_CONTROL,
1746                       _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1747
1748         /* Clear the context id. Here be magic! */
1749         I915_WRITE64_FW(GEN6_BSD_RNCID, 0x0);
1750
1751         /* Wait for the ring not to be idle, i.e. for it to wake up. */
1752         if (intel_wait_for_register_fw(dev_priv,
1753                                        GEN6_BSD_SLEEP_PSMI_CONTROL,
1754                                        GEN6_BSD_SLEEP_INDICATOR,
1755                                        0,
1756                                        50))
1757                 DRM_ERROR("timed out waiting for the BSD ring to wake up\n");
1758
1759         /* Now that the ring is fully powered up, update the tail */
1760         i9xx_submit_request(request);
1761
1762         /* Let the ring send IDLE messages to the GT again,
1763          * and so let it sleep to conserve power when idle.
1764          */
1765         I915_WRITE_FW(GEN6_BSD_SLEEP_PSMI_CONTROL,
1766                       _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1767
1768         intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
1769 }
1770
1771 static int gen6_bsd_ring_flush(struct drm_i915_gem_request *req, u32 mode)
1772 {
1773         u32 cmd, *cs;
1774
1775         cs = intel_ring_begin(req, 4);
1776         if (IS_ERR(cs))
1777                 return PTR_ERR(cs);
1778
1779         cmd = MI_FLUSH_DW;
1780         if (INTEL_GEN(req->i915) >= 8)
1781                 cmd += 1;
1782
1783         /* We always require a command barrier so that subsequent
1784          * commands, such as breadcrumb interrupts, are strictly ordered
1785          * wrt the contents of the write cache being flushed to memory
1786          * (and thus being coherent from the CPU).
1787          */
1788         cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1789
1790         /*
1791          * Bspec vol 1c.5 - video engine command streamer:
1792          * "If ENABLED, all TLBs will be invalidated once the flush
1793          * operation is complete. This bit is only valid when the
1794          * Post-Sync Operation field is a value of 1h or 3h."
1795          */
1796         if (mode & EMIT_INVALIDATE)
1797                 cmd |= MI_INVALIDATE_TLB | MI_INVALIDATE_BSD;
1798
1799         *cs++ = cmd;
1800         *cs++ = I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT;
1801         if (INTEL_GEN(req->i915) >= 8) {
1802                 *cs++ = 0; /* upper addr */
1803                 *cs++ = 0; /* value */
1804         } else  {
1805                 *cs++ = 0;
1806                 *cs++ = MI_NOOP;
1807         }
1808         intel_ring_advance(req, cs);
1809         return 0;
1810 }
1811
1812 static int
1813 gen8_emit_bb_start(struct drm_i915_gem_request *req,
1814                    u64 offset, u32 len,
1815                    unsigned int dispatch_flags)
1816 {
1817         bool ppgtt = USES_PPGTT(req->i915) &&
1818                         !(dispatch_flags & I915_DISPATCH_SECURE);
1819         u32 *cs;
1820
1821         cs = intel_ring_begin(req, 4);
1822         if (IS_ERR(cs))
1823                 return PTR_ERR(cs);
1824
1825         /* FIXME(BDW): Address space and security selectors. */
1826         *cs++ = MI_BATCH_BUFFER_START_GEN8 | (ppgtt << 8) | (dispatch_flags &
1827                 I915_DISPATCH_RS ? MI_BATCH_RESOURCE_STREAMER : 0);
1828         *cs++ = lower_32_bits(offset);
1829         *cs++ = upper_32_bits(offset);
1830         *cs++ = MI_NOOP;
1831         intel_ring_advance(req, cs);
1832
1833         return 0;
1834 }
1835
1836 static int
1837 hsw_emit_bb_start(struct drm_i915_gem_request *req,
1838                   u64 offset, u32 len,
1839                   unsigned int dispatch_flags)
1840 {
1841         u32 *cs;
1842
1843         cs = intel_ring_begin(req, 2);
1844         if (IS_ERR(cs))
1845                 return PTR_ERR(cs);
1846
1847         *cs++ = MI_BATCH_BUFFER_START | (dispatch_flags & I915_DISPATCH_SECURE ?
1848                 0 : MI_BATCH_PPGTT_HSW | MI_BATCH_NON_SECURE_HSW) |
1849                 (dispatch_flags & I915_DISPATCH_RS ?
1850                 MI_BATCH_RESOURCE_STREAMER : 0);
1851         /* bit0-7 is the length on GEN6+ */
1852         *cs++ = offset;
1853         intel_ring_advance(req, cs);
1854
1855         return 0;
1856 }
1857
1858 static int
1859 gen6_emit_bb_start(struct drm_i915_gem_request *req,
1860                    u64 offset, u32 len,
1861                    unsigned int dispatch_flags)
1862 {
1863         u32 *cs;
1864
1865         cs = intel_ring_begin(req, 2);
1866         if (IS_ERR(cs))
1867                 return PTR_ERR(cs);
1868
1869         *cs++ = MI_BATCH_BUFFER_START | (dispatch_flags & I915_DISPATCH_SECURE ?
1870                 0 : MI_BATCH_NON_SECURE_I965);
1871         /* bit0-7 is the length on GEN6+ */
1872         *cs++ = offset;
1873         intel_ring_advance(req, cs);
1874
1875         return 0;
1876 }
1877
1878 /* Blitter support (SandyBridge+) */
1879
1880 static int gen6_ring_flush(struct drm_i915_gem_request *req, u32 mode)
1881 {
1882         u32 cmd, *cs;
1883
1884         cs = intel_ring_begin(req, 4);
1885         if (IS_ERR(cs))
1886                 return PTR_ERR(cs);
1887
1888         cmd = MI_FLUSH_DW;
1889         if (INTEL_GEN(req->i915) >= 8)
1890                 cmd += 1;
1891
1892         /* We always require a command barrier so that subsequent
1893          * commands, such as breadcrumb interrupts, are strictly ordered
1894          * wrt the contents of the write cache being flushed to memory
1895          * (and thus being coherent from the CPU).
1896          */
1897         cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1898
1899         /*
1900          * Bspec vol 1c.3 - blitter engine command streamer:
1901          * "If ENABLED, all TLBs will be invalidated once the flush
1902          * operation is complete. This bit is only valid when the
1903          * Post-Sync Operation field is a value of 1h or 3h."
1904          */
1905         if (mode & EMIT_INVALIDATE)
1906                 cmd |= MI_INVALIDATE_TLB;
1907         *cs++ = cmd;
1908         *cs++ = I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT;
1909         if (INTEL_GEN(req->i915) >= 8) {
1910                 *cs++ = 0; /* upper addr */
1911                 *cs++ = 0; /* value */
1912         } else  {
1913                 *cs++ = 0;
1914                 *cs++ = MI_NOOP;
1915         }
1916         intel_ring_advance(req, cs);
1917
1918         return 0;
1919 }
1920
1921 static void intel_ring_init_semaphores(struct drm_i915_private *dev_priv,
1922                                        struct intel_engine_cs *engine)
1923 {
1924         struct drm_i915_gem_object *obj;
1925         int ret, i;
1926
1927         if (!i915.semaphores)
1928                 return;
1929
1930         if (INTEL_GEN(dev_priv) >= 8 && !dev_priv->semaphore) {
1931                 struct i915_vma *vma;
1932
1933                 obj = i915_gem_object_create(dev_priv, PAGE_SIZE);
1934                 if (IS_ERR(obj))
1935                         goto err;
1936
1937                 vma = i915_vma_instance(obj, &dev_priv->ggtt.base, NULL);
1938                 if (IS_ERR(vma))
1939                         goto err_obj;
1940
1941                 ret = i915_gem_object_set_to_gtt_domain(obj, false);
1942                 if (ret)
1943                         goto err_obj;
1944
1945                 ret = i915_vma_pin(vma, 0, 0, PIN_GLOBAL | PIN_HIGH);
1946                 if (ret)
1947                         goto err_obj;
1948
1949                 dev_priv->semaphore = vma;
1950         }
1951
1952         if (INTEL_GEN(dev_priv) >= 8) {
1953                 u32 offset = i915_ggtt_offset(dev_priv->semaphore);
1954
1955                 engine->semaphore.sync_to = gen8_ring_sync_to;
1956                 engine->semaphore.signal = gen8_xcs_signal;
1957
1958                 for (i = 0; i < I915_NUM_ENGINES; i++) {
1959                         u32 ring_offset;
1960
1961                         if (i != engine->id)
1962                                 ring_offset = offset + GEN8_SEMAPHORE_OFFSET(engine->id, i);
1963                         else
1964                                 ring_offset = MI_SEMAPHORE_SYNC_INVALID;
1965
1966                         engine->semaphore.signal_ggtt[i] = ring_offset;
1967                 }
1968         } else if (INTEL_GEN(dev_priv) >= 6) {
1969                 engine->semaphore.sync_to = gen6_ring_sync_to;
1970                 engine->semaphore.signal = gen6_signal;
1971
1972                 /*
1973                  * The current semaphore is only applied on pre-gen8
1974                  * platform.  And there is no VCS2 ring on the pre-gen8
1975                  * platform. So the semaphore between RCS and VCS2 is
1976                  * initialized as INVALID.  Gen8 will initialize the
1977                  * sema between VCS2 and RCS later.
1978                  */
1979                 for (i = 0; i < GEN6_NUM_SEMAPHORES; i++) {
1980                         static const struct {
1981                                 u32 wait_mbox;
1982                                 i915_reg_t mbox_reg;
1983                         } sem_data[GEN6_NUM_SEMAPHORES][GEN6_NUM_SEMAPHORES] = {
1984                                 [RCS_HW] = {
1985                                         [VCS_HW] =  { .wait_mbox = MI_SEMAPHORE_SYNC_RV,  .mbox_reg = GEN6_VRSYNC },
1986                                         [BCS_HW] =  { .wait_mbox = MI_SEMAPHORE_SYNC_RB,  .mbox_reg = GEN6_BRSYNC },
1987                                         [VECS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_RVE, .mbox_reg = GEN6_VERSYNC },
1988                                 },
1989                                 [VCS_HW] = {
1990                                         [RCS_HW] =  { .wait_mbox = MI_SEMAPHORE_SYNC_VR,  .mbox_reg = GEN6_RVSYNC },
1991                                         [BCS_HW] =  { .wait_mbox = MI_SEMAPHORE_SYNC_VB,  .mbox_reg = GEN6_BVSYNC },
1992                                         [VECS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_VVE, .mbox_reg = GEN6_VEVSYNC },
1993                                 },
1994                                 [BCS_HW] = {
1995                                         [RCS_HW] =  { .wait_mbox = MI_SEMAPHORE_SYNC_BR,  .mbox_reg = GEN6_RBSYNC },
1996                                         [VCS_HW] =  { .wait_mbox = MI_SEMAPHORE_SYNC_BV,  .mbox_reg = GEN6_VBSYNC },
1997                                         [VECS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_BVE, .mbox_reg = GEN6_VEBSYNC },
1998                                 },
1999                                 [VECS_HW] = {
2000                                         [RCS_HW] =  { .wait_mbox = MI_SEMAPHORE_SYNC_VER, .mbox_reg = GEN6_RVESYNC },
2001                                         [VCS_HW] =  { .wait_mbox = MI_SEMAPHORE_SYNC_VEV, .mbox_reg = GEN6_VVESYNC },
2002                                         [BCS_HW] =  { .wait_mbox = MI_SEMAPHORE_SYNC_VEB, .mbox_reg = GEN6_BVESYNC },
2003                                 },
2004                         };
2005                         u32 wait_mbox;
2006                         i915_reg_t mbox_reg;
2007
2008                         if (i == engine->hw_id) {
2009                                 wait_mbox = MI_SEMAPHORE_SYNC_INVALID;
2010                                 mbox_reg = GEN6_NOSYNC;
2011                         } else {
2012                                 wait_mbox = sem_data[engine->hw_id][i].wait_mbox;
2013                                 mbox_reg = sem_data[engine->hw_id][i].mbox_reg;
2014                         }
2015
2016                         engine->semaphore.mbox.wait[i] = wait_mbox;
2017                         engine->semaphore.mbox.signal[i] = mbox_reg;
2018                 }
2019         }
2020
2021         return;
2022
2023 err_obj:
2024         i915_gem_object_put(obj);
2025 err:
2026         DRM_DEBUG_DRIVER("Failed to allocate space for semaphores, disabling\n");
2027         i915.semaphores = 0;
2028 }
2029
2030 static void intel_ring_init_irq(struct drm_i915_private *dev_priv,
2031                                 struct intel_engine_cs *engine)
2032 {
2033         engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT << engine->irq_shift;
2034
2035         if (INTEL_GEN(dev_priv) >= 8) {
2036                 engine->irq_enable = gen8_irq_enable;
2037                 engine->irq_disable = gen8_irq_disable;
2038                 engine->irq_seqno_barrier = gen6_seqno_barrier;
2039         } else if (INTEL_GEN(dev_priv) >= 6) {
2040                 engine->irq_enable = gen6_irq_enable;
2041                 engine->irq_disable = gen6_irq_disable;
2042                 engine->irq_seqno_barrier = gen6_seqno_barrier;
2043         } else if (INTEL_GEN(dev_priv) >= 5) {
2044                 engine->irq_enable = gen5_irq_enable;
2045                 engine->irq_disable = gen5_irq_disable;
2046                 engine->irq_seqno_barrier = gen5_seqno_barrier;
2047         } else if (INTEL_GEN(dev_priv) >= 3) {
2048                 engine->irq_enable = i9xx_irq_enable;
2049                 engine->irq_disable = i9xx_irq_disable;
2050         } else {
2051                 engine->irq_enable = i8xx_irq_enable;
2052                 engine->irq_disable = i8xx_irq_disable;
2053         }
2054 }
2055
2056 static void i9xx_set_default_submission(struct intel_engine_cs *engine)
2057 {
2058         engine->submit_request = i9xx_submit_request;
2059 }
2060
2061 static void gen6_bsd_set_default_submission(struct intel_engine_cs *engine)
2062 {
2063         engine->submit_request = gen6_bsd_submit_request;
2064 }
2065
2066 static void intel_ring_default_vfuncs(struct drm_i915_private *dev_priv,
2067                                       struct intel_engine_cs *engine)
2068 {
2069         intel_ring_init_irq(dev_priv, engine);
2070         intel_ring_init_semaphores(dev_priv, engine);
2071
2072         engine->init_hw = init_ring_common;
2073         engine->reset_hw = reset_ring_common;
2074
2075         engine->context_pin = intel_ring_context_pin;
2076         engine->context_unpin = intel_ring_context_unpin;
2077
2078         engine->request_alloc = ring_request_alloc;
2079
2080         engine->emit_breadcrumb = i9xx_emit_breadcrumb;
2081         engine->emit_breadcrumb_sz = i9xx_emit_breadcrumb_sz;
2082         if (i915.semaphores) {
2083                 int num_rings;
2084
2085                 engine->emit_breadcrumb = gen6_sema_emit_breadcrumb;
2086
2087                 num_rings = hweight32(INTEL_INFO(dev_priv)->ring_mask) - 1;
2088                 if (INTEL_GEN(dev_priv) >= 8) {
2089                         engine->emit_breadcrumb_sz += num_rings * 6;
2090                 } else {
2091                         engine->emit_breadcrumb_sz += num_rings * 3;
2092                         if (num_rings & 1)
2093                                 engine->emit_breadcrumb_sz++;
2094                 }
2095         }
2096
2097         engine->set_default_submission = i9xx_set_default_submission;
2098
2099         if (INTEL_GEN(dev_priv) >= 8)
2100                 engine->emit_bb_start = gen8_emit_bb_start;
2101         else if (INTEL_GEN(dev_priv) >= 6)
2102                 engine->emit_bb_start = gen6_emit_bb_start;
2103         else if (INTEL_GEN(dev_priv) >= 4)
2104                 engine->emit_bb_start = i965_emit_bb_start;
2105         else if (IS_I830(dev_priv) || IS_I845G(dev_priv))
2106                 engine->emit_bb_start = i830_emit_bb_start;
2107         else
2108                 engine->emit_bb_start = i915_emit_bb_start;
2109 }
2110
2111 int intel_init_render_ring_buffer(struct intel_engine_cs *engine)
2112 {
2113         struct drm_i915_private *dev_priv = engine->i915;
2114         int ret;
2115
2116         intel_ring_default_vfuncs(dev_priv, engine);
2117
2118         if (HAS_L3_DPF(dev_priv))
2119                 engine->irq_keep_mask = GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
2120
2121         if (INTEL_GEN(dev_priv) >= 8) {
2122                 engine->init_context = intel_rcs_ctx_init;
2123                 engine->emit_breadcrumb = gen8_render_emit_breadcrumb;
2124                 engine->emit_breadcrumb_sz = gen8_render_emit_breadcrumb_sz;
2125                 engine->emit_flush = gen8_render_ring_flush;
2126                 if (i915.semaphores) {
2127                         int num_rings;
2128
2129                         engine->semaphore.signal = gen8_rcs_signal;
2130
2131                         num_rings =
2132                                 hweight32(INTEL_INFO(dev_priv)->ring_mask) - 1;
2133                         engine->emit_breadcrumb_sz += num_rings * 8;
2134                 }
2135         } else if (INTEL_GEN(dev_priv) >= 6) {
2136                 engine->init_context = intel_rcs_ctx_init;
2137                 engine->emit_flush = gen7_render_ring_flush;
2138                 if (IS_GEN6(dev_priv))
2139                         engine->emit_flush = gen6_render_ring_flush;
2140         } else if (IS_GEN5(dev_priv)) {
2141                 engine->emit_flush = gen4_render_ring_flush;
2142         } else {
2143                 if (INTEL_GEN(dev_priv) < 4)
2144                         engine->emit_flush = gen2_render_ring_flush;
2145                 else
2146                         engine->emit_flush = gen4_render_ring_flush;
2147                 engine->irq_enable_mask = I915_USER_INTERRUPT;
2148         }
2149
2150         if (IS_HASWELL(dev_priv))
2151                 engine->emit_bb_start = hsw_emit_bb_start;
2152
2153         engine->init_hw = init_render_ring;
2154         engine->cleanup = render_ring_cleanup;
2155
2156         ret = intel_init_ring_buffer(engine);
2157         if (ret)
2158                 return ret;
2159
2160         if (INTEL_GEN(dev_priv) >= 6) {
2161                 ret = intel_engine_create_scratch(engine, PAGE_SIZE);
2162                 if (ret)
2163                         return ret;
2164         } else if (HAS_BROKEN_CS_TLB(dev_priv)) {
2165                 ret = intel_engine_create_scratch(engine, I830_WA_SIZE);
2166                 if (ret)
2167                         return ret;
2168         }
2169
2170         return 0;
2171 }
2172
2173 int intel_init_bsd_ring_buffer(struct intel_engine_cs *engine)
2174 {
2175         struct drm_i915_private *dev_priv = engine->i915;
2176
2177         intel_ring_default_vfuncs(dev_priv, engine);
2178
2179         if (INTEL_GEN(dev_priv) >= 6) {
2180                 /* gen6 bsd needs a special wa for tail updates */
2181                 if (IS_GEN6(dev_priv))
2182                         engine->set_default_submission = gen6_bsd_set_default_submission;
2183                 engine->emit_flush = gen6_bsd_ring_flush;
2184                 if (INTEL_GEN(dev_priv) < 8)
2185                         engine->irq_enable_mask = GT_BSD_USER_INTERRUPT;
2186         } else {
2187                 engine->mmio_base = BSD_RING_BASE;
2188                 engine->emit_flush = bsd_ring_flush;
2189                 if (IS_GEN5(dev_priv))
2190                         engine->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
2191                 else
2192                         engine->irq_enable_mask = I915_BSD_USER_INTERRUPT;
2193         }
2194
2195         return intel_init_ring_buffer(engine);
2196 }
2197
2198 /**
2199  * Initialize the second BSD ring (eg. Broadwell GT3, Skylake GT3)
2200  */
2201 int intel_init_bsd2_ring_buffer(struct intel_engine_cs *engine)
2202 {
2203         struct drm_i915_private *dev_priv = engine->i915;
2204
2205         intel_ring_default_vfuncs(dev_priv, engine);
2206
2207         engine->emit_flush = gen6_bsd_ring_flush;
2208
2209         return intel_init_ring_buffer(engine);
2210 }
2211
2212 int intel_init_blt_ring_buffer(struct intel_engine_cs *engine)
2213 {
2214         struct drm_i915_private *dev_priv = engine->i915;
2215
2216         intel_ring_default_vfuncs(dev_priv, engine);
2217
2218         engine->emit_flush = gen6_ring_flush;
2219         if (INTEL_GEN(dev_priv) < 8)
2220                 engine->irq_enable_mask = GT_BLT_USER_INTERRUPT;
2221
2222         return intel_init_ring_buffer(engine);
2223 }
2224
2225 int intel_init_vebox_ring_buffer(struct intel_engine_cs *engine)
2226 {
2227         struct drm_i915_private *dev_priv = engine->i915;
2228
2229         intel_ring_default_vfuncs(dev_priv, engine);
2230
2231         engine->emit_flush = gen6_ring_flush;
2232
2233         if (INTEL_GEN(dev_priv) < 8) {
2234                 engine->irq_enable_mask = PM_VEBOX_USER_INTERRUPT;
2235                 engine->irq_enable = hsw_vebox_irq_enable;
2236                 engine->irq_disable = hsw_vebox_irq_disable;
2237         }
2238
2239         return intel_init_ring_buffer(engine);
2240 }