drivers/gpu/drm/i915/gt/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
  32 #include <drm/i915_drm.h>
  33
  34 #include "gem/i915_gem_context.h"
  35
  36 #include "i915_drv.h"
  37 #include "i915_gem_render_state.h"
  38 #include "i915_trace.h"
  39 #include "intel_context.h"
  40 #include "intel_reset.h"
  41 #include "intel_workarounds.h"
  42
  43 /* Rough estimate of the typical request size, performing a flush,
  44  * set-context and then emitting the batch.
  45  */
  46 #define LEGACY_REQUEST_SIZE 200
  47
  48 unsigned int intel_ring_update_space(struct intel_ring *ring)
  49 {
  50         unsigned int space;
  51
  52         space = __intel_ring_space(ring->head, ring->emit, ring->size);
  53
  54         ring->space = space;
  55         return space;
  56 }
  57
  58 static int
  59 gen2_render_ring_flush(struct i915_request *rq, u32 mode)
  60 {
  61         unsigned int num_store_dw;
  62         u32 cmd, *cs;
  63
  64         cmd = MI_FLUSH;
  65         num_store_dw = 0;
  66         if (mode & EMIT_INVALIDATE)
  67                 cmd |= MI_READ_FLUSH;
  68         if (mode & EMIT_FLUSH)
  69                 num_store_dw = 4;
  70
  71         cs = intel_ring_begin(rq, 2 + 3 * num_store_dw);
  72         if (IS_ERR(cs))
  73                 return PTR_ERR(cs);
  74
  75         *cs++ = cmd;
  76         while (num_store_dw--) {
  77                 *cs++ = MI_STORE_DWORD_IMM | MI_MEM_VIRTUAL;
  78                 *cs++ = i915_scratch_offset(rq->i915);
  79                 *cs++ = 0;
  80         }
  81         *cs++ = MI_FLUSH | MI_NO_WRITE_FLUSH;
  82
  83         intel_ring_advance(rq, cs);
  84
  85         return 0;
  86 }
  87
  88 static int
  89 gen4_render_ring_flush(struct i915_request *rq, u32 mode)
  90 {
  91         u32 cmd, *cs;
  92         int i;
  93
  94         /*
  95          * read/write caches:
  96          *
  97          * I915_GEM_DOMAIN_RENDER is always invalidated, but is
  98          * only flushed if MI_NO_WRITE_FLUSH is unset.  On 965, it is
  99          * also flushed at 2d versus 3d pipeline switches.
 100          *
 101          * read-only caches:
 102          *
 103          * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
 104          * MI_READ_FLUSH is set, and is always flushed on 965.
 105          *
 106          * I915_GEM_DOMAIN_COMMAND may not exist?
 107          *
 108          * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
 109          * invalidated when MI_EXE_FLUSH is set.
 110          *
 111          * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
 112          * invalidated with every MI_FLUSH.
 113          *
 114          * TLBs:
 115          *
 116          * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
 117          * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
 118          * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
 119          * are flushed at any MI_FLUSH.
 120          */
 121
 122         cmd = MI_FLUSH;
 123         if (mode & EMIT_INVALIDATE) {
 124                 cmd |= MI_EXE_FLUSH;
 125                 if (IS_G4X(rq->i915) || IS_GEN(rq->i915, 5))
 126                         cmd |= MI_INVALIDATE_ISP;
 127         }
 128
 129         i = 2;
 130         if (mode & EMIT_INVALIDATE)
 131                 i += 20;
 132
 133         cs = intel_ring_begin(rq, i);
 134         if (IS_ERR(cs))
 135                 return PTR_ERR(cs);
 136
 137         *cs++ = cmd;
 138
 139         /*
 140          * A random delay to let the CS invalidate take effect? Without this
 141          * delay, the GPU relocation path fails as the CS does not see
 142          * the updated contents. Just as important, if we apply the flushes
 143          * to the EMIT_FLUSH branch (i.e. immediately after the relocation
 144          * write and before the invalidate on the next batch), the relocations
 145          * still fail. This implies that is a delay following invalidation
 146          * that is required to reset the caches as opposed to a delay to
 147          * ensure the memory is written.
 148          */
 149         if (mode & EMIT_INVALIDATE) {
 150                 *cs++ = GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE;
 151                 *cs++ = i915_scratch_offset(rq->i915) | PIPE_CONTROL_GLOBAL_GTT;
 152                 *cs++ = 0;
 153                 *cs++ = 0;
 154
 155                 for (i = 0; i < 12; i++)
 156                         *cs++ = MI_FLUSH;
 157
 158                 *cs++ = GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE;
 159                 *cs++ = i915_scratch_offset(rq->i915) | PIPE_CONTROL_GLOBAL_GTT;
 160                 *cs++ = 0;
 161                 *cs++ = 0;
 162         }
 163
 164         *cs++ = cmd;
 165
 166         intel_ring_advance(rq, cs);
 167
 168         return 0;
 169 }
 170
 171 /*
 172  * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
 173  * implementing two workarounds on gen6.  From section 1.4.7.1
 174  * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
 175  *
 176  * [DevSNB-C+{W/A}] Before any depth stall flush (including those
 177  * produced by non-pipelined state commands), software needs to first
 178  * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
 179  * 0.
 180  *
 181  * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
 182  * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
 183  *
 184  * And the workaround for these two requires this workaround first:
 185  *
 186  * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
 187  * BEFORE the pipe-control with a post-sync op and no write-cache
 188  * flushes.
 189  *
 190  * And this last workaround is tricky because of the requirements on
 191  * that bit.  From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
 192  * volume 2 part 1:
 193  *
 194  *     "1 of the following must also be set:
 195  *      - Render Target Cache Flush Enable ([12] of DW1)
 196  *      - Depth Cache Flush Enable ([0] of DW1)
 197  *      - Stall at Pixel Scoreboard ([1] of DW1)
 198  *      - Depth Stall ([13] of DW1)
 199  *      - Post-Sync Operation ([13] of DW1)
 200  *      - Notify Enable ([8] of DW1)"
 201  *
 202  * The cache flushes require the workaround flush that triggered this
 203  * one, so we can't use it.  Depth stall would trigger the same.
 204  * Post-sync nonzero is what triggered this second workaround, so we
 205  * can't use that one either.  Notify enable is IRQs, which aren't
 206  * really our business.  That leaves only stall at scoreboard.
 207  */
 208 static int
 209 gen6_emit_post_sync_nonzero_flush(struct i915_request *rq)
 210 {
 211         u32 scratch_addr = i915_scratch_offset(rq->i915) + 2 * CACHELINE_BYTES;
 212         u32 *cs;
 213
 214         cs = intel_ring_begin(rq, 6);
 215         if (IS_ERR(cs))
 216                 return PTR_ERR(cs);
 217
 218         *cs++ = GFX_OP_PIPE_CONTROL(5);
 219         *cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
 220         *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
 221         *cs++ = 0; /* low dword */
 222         *cs++ = 0; /* high dword */
 223         *cs++ = MI_NOOP;
 224         intel_ring_advance(rq, cs);
 225
 226         cs = intel_ring_begin(rq, 6);
 227         if (IS_ERR(cs))
 228                 return PTR_ERR(cs);
 229
 230         *cs++ = GFX_OP_PIPE_CONTROL(5);
 231         *cs++ = PIPE_CONTROL_QW_WRITE;
 232         *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
 233         *cs++ = 0;
 234         *cs++ = 0;
 235         *cs++ = MI_NOOP;
 236         intel_ring_advance(rq, cs);
 237
 238         return 0;
 239 }
 240
 241 static int
 242 gen6_render_ring_flush(struct i915_request *rq, u32 mode)
 243 {
 244         u32 scratch_addr = i915_scratch_offset(rq->i915) + 2 * CACHELINE_BYTES;
 245         u32 *cs, flags = 0;
 246         int ret;
 247
 248         /* Force SNB workarounds for PIPE_CONTROL flushes */
 249         ret = gen6_emit_post_sync_nonzero_flush(rq);
 250         if (ret)
 251                 return ret;
 252
 253         /* Just flush everything.  Experiments have shown that reducing the
 254          * number of bits based on the write domains has little performance
 255          * impact.
 256          */
 257         if (mode & EMIT_FLUSH) {
 258                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
 259                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
 260                 /*
 261                  * Ensure that any following seqno writes only happen
 262                  * when the render cache is indeed flushed.
 263                  */
 264                 flags |= PIPE_CONTROL_CS_STALL;
 265         }
 266         if (mode & EMIT_INVALIDATE) {
 267                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
 268                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 269                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 270                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
 271                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
 272                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
 273                 /*
 274                  * TLB invalidate requires a post-sync write.
 275                  */
 276                 flags |= PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_CS_STALL;
 277         }
 278
 279         cs = intel_ring_begin(rq, 4);
 280         if (IS_ERR(cs))
 281                 return PTR_ERR(cs);
 282
 283         *cs++ = GFX_OP_PIPE_CONTROL(4);
 284         *cs++ = flags;
 285         *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
 286         *cs++ = 0;
 287         intel_ring_advance(rq, cs);
 288
 289         return 0;
 290 }
 291
 292 static u32 *gen6_rcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
 293 {
 294         /* First we do the gen6_emit_post_sync_nonzero_flush w/a */
 295         *cs++ = GFX_OP_PIPE_CONTROL(4);
 296         *cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
 297         *cs++ = 0;
 298         *cs++ = 0;
 299
 300         *cs++ = GFX_OP_PIPE_CONTROL(4);
 301         *cs++ = PIPE_CONTROL_QW_WRITE;
 302         *cs++ = i915_scratch_offset(rq->i915) | PIPE_CONTROL_GLOBAL_GTT;
 303         *cs++ = 0;
 304
 305         /* Finally we can flush and with it emit the breadcrumb */
 306         *cs++ = GFX_OP_PIPE_CONTROL(4);
 307         *cs++ = (PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
 308                  PIPE_CONTROL_DEPTH_CACHE_FLUSH |
 309                  PIPE_CONTROL_DC_FLUSH_ENABLE |
 310                  PIPE_CONTROL_QW_WRITE |
 311                  PIPE_CONTROL_CS_STALL);
 312         *cs++ = rq->timeline->hwsp_offset | PIPE_CONTROL_GLOBAL_GTT;
 313         *cs++ = rq->fence.seqno;
 314
 315         *cs++ = MI_USER_INTERRUPT;
 316         *cs++ = MI_NOOP;
 317
 318         rq->tail = intel_ring_offset(rq, cs);
 319         assert_ring_tail_valid(rq->ring, rq->tail);
 320
 321         return cs;
 322 }
 323
 324 static int
 325 gen7_render_ring_cs_stall_wa(struct i915_request *rq)
 326 {
 327         u32 *cs;
 328
 329         cs = intel_ring_begin(rq, 4);
 330         if (IS_ERR(cs))
 331                 return PTR_ERR(cs);
 332
 333         *cs++ = GFX_OP_PIPE_CONTROL(4);
 334         *cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
 335         *cs++ = 0;
 336         *cs++ = 0;
 337         intel_ring_advance(rq, cs);
 338
 339         return 0;
 340 }
 341
 342 static int
 343 gen7_render_ring_flush(struct i915_request *rq, u32 mode)
 344 {
 345         u32 scratch_addr = i915_scratch_offset(rq->i915) + 2 * CACHELINE_BYTES;
 346         u32 *cs, flags = 0;
 347
 348         /*
 349          * Ensure that any following seqno writes only happen when the render
 350          * cache is indeed flushed.
 351          *
 352          * Workaround: 4th PIPE_CONTROL command (except the ones with only
 353          * read-cache invalidate bits set) must have the CS_STALL bit set. We
 354          * don't try to be clever and just set it unconditionally.
 355          */
 356         flags |= PIPE_CONTROL_CS_STALL;
 357
 358         /* Just flush everything.  Experiments have shown that reducing the
 359          * number of bits based on the write domains has little performance
 360          * impact.
 361          */
 362         if (mode & EMIT_FLUSH) {
 363                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
 364                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
 365                 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
 366                 flags |= PIPE_CONTROL_FLUSH_ENABLE;
 367         }
 368         if (mode & EMIT_INVALIDATE) {
 369                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
 370                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 371                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 372                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
 373                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
 374                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
 375                 flags |= PIPE_CONTROL_MEDIA_STATE_CLEAR;
 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                 flags |= PIPE_CONTROL_STALL_AT_SCOREBOARD;
 383
 384                 /* Workaround: we must issue a pipe_control with CS-stall bit
 385                  * set before a pipe_control command that has the state cache
 386                  * invalidate bit set. */
 387                 gen7_render_ring_cs_stall_wa(rq);
 388         }
 389
 390         cs = intel_ring_begin(rq, 4);
 391         if (IS_ERR(cs))
 392                 return PTR_ERR(cs);
 393
 394         *cs++ = GFX_OP_PIPE_CONTROL(4);
 395         *cs++ = flags;
 396         *cs++ = scratch_addr;
 397         *cs++ = 0;
 398         intel_ring_advance(rq, cs);
 399
 400         return 0;
 401 }
 402
 403 static u32 *gen7_rcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
 404 {
 405         *cs++ = GFX_OP_PIPE_CONTROL(4);
 406         *cs++ = (PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
 407                  PIPE_CONTROL_DEPTH_CACHE_FLUSH |
 408                  PIPE_CONTROL_DC_FLUSH_ENABLE |
 409                  PIPE_CONTROL_FLUSH_ENABLE |
 410                  PIPE_CONTROL_QW_WRITE |
 411                  PIPE_CONTROL_GLOBAL_GTT_IVB |
 412                  PIPE_CONTROL_CS_STALL);
 413         *cs++ = rq->timeline->hwsp_offset;
 414         *cs++ = rq->fence.seqno;
 415
 416         *cs++ = MI_USER_INTERRUPT;
 417         *cs++ = MI_NOOP;
 418
 419         rq->tail = intel_ring_offset(rq, cs);
 420         assert_ring_tail_valid(rq->ring, rq->tail);
 421
 422         return cs;
 423 }
 424
 425 static u32 *gen6_xcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
 426 {
 427         GEM_BUG_ON(rq->timeline->hwsp_ggtt != rq->engine->status_page.vma);
 428         GEM_BUG_ON(offset_in_page(rq->timeline->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
 429
 430         *cs++ = MI_FLUSH_DW | MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_DW_STORE_INDEX;
 431         *cs++ = I915_GEM_HWS_SEQNO_ADDR | MI_FLUSH_DW_USE_GTT;
 432         *cs++ = rq->fence.seqno;
 433
 434         *cs++ = MI_USER_INTERRUPT;
 435
 436         rq->tail = intel_ring_offset(rq, cs);
 437         assert_ring_tail_valid(rq->ring, rq->tail);
 438
 439         return cs;
 440 }
 441
 442 #define GEN7_XCS_WA 32
 443 static u32 *gen7_xcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
 444 {
 445         int i;
 446
 447         GEM_BUG_ON(rq->timeline->hwsp_ggtt != rq->engine->status_page.vma);
 448         GEM_BUG_ON(offset_in_page(rq->timeline->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
 449
 450         *cs++ = MI_FLUSH_DW | MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_DW_STORE_INDEX;
 451         *cs++ = I915_GEM_HWS_SEQNO_ADDR | MI_FLUSH_DW_USE_GTT;
 452         *cs++ = rq->fence.seqno;
 453
 454         for (i = 0; i < GEN7_XCS_WA; i++) {
 455                 *cs++ = MI_STORE_DWORD_INDEX;
 456                 *cs++ = I915_GEM_HWS_SEQNO_ADDR;
 457                 *cs++ = rq->fence.seqno;
 458         }
 459
 460         *cs++ = MI_FLUSH_DW;
 461         *cs++ = 0;
 462         *cs++ = 0;
 463
 464         *cs++ = MI_USER_INTERRUPT;
 465         *cs++ = MI_NOOP;
 466
 467         rq->tail = intel_ring_offset(rq, cs);
 468         assert_ring_tail_valid(rq->ring, rq->tail);
 469
 470         return cs;
 471 }
 472 #undef GEN7_XCS_WA
 473
 474 static void set_hwstam(struct intel_engine_cs *engine, u32 mask)
 475 {
 476         /*
 477          * Keep the render interrupt unmasked as this papers over
 478          * lost interrupts following a reset.
 479          */
 480         if (engine->class == RENDER_CLASS) {
 481                 if (INTEL_GEN(engine->i915) >= 6)
 482                         mask &= ~BIT(0);
 483                 else
 484                         mask &= ~I915_USER_INTERRUPT;
 485         }
 486
 487         intel_engine_set_hwsp_writemask(engine, mask);
 488 }
 489
 490 static void set_hws_pga(struct intel_engine_cs *engine, phys_addr_t phys)
 491 {
 492         struct drm_i915_private *dev_priv = engine->i915;
 493         u32 addr;
 494
 495         addr = lower_32_bits(phys);
 496         if (INTEL_GEN(dev_priv) >= 4)
 497                 addr |= (phys >> 28) & 0xf0;
 498
 499         I915_WRITE(HWS_PGA, addr);
 500 }
 501
 502 static struct page *status_page(struct intel_engine_cs *engine)
 503 {
 504         struct drm_i915_gem_object *obj = engine->status_page.vma->obj;
 505
 506         GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
 507         return sg_page(obj->mm.pages->sgl);
 508 }
 509
 510 static void ring_setup_phys_status_page(struct intel_engine_cs *engine)
 511 {
 512         set_hws_pga(engine, PFN_PHYS(page_to_pfn(status_page(engine))));
 513         set_hwstam(engine, ~0u);
 514 }
 515
 516 static void set_hwsp(struct intel_engine_cs *engine, u32 offset)
 517 {
 518         struct drm_i915_private *dev_priv = engine->i915;
 519         i915_reg_t hwsp;
 520
 521         /*
 522          * The ring status page addresses are no longer next to the rest of
 523          * the ring registers as of gen7.
 524          */
 525         if (IS_GEN(dev_priv, 7)) {
 526                 switch (engine->id) {
 527                 /*
 528                  * No more rings exist on Gen7. Default case is only to shut up
 529                  * gcc switch check warning.
 530                  */
 531                 default:
 532                         GEM_BUG_ON(engine->id);
 533                         /* fallthrough */
 534                 case RCS0:
 535                         hwsp = RENDER_HWS_PGA_GEN7;
 536                         break;
 537                 case BCS0:
 538                         hwsp = BLT_HWS_PGA_GEN7;
 539                         break;
 540                 case VCS0:
 541                         hwsp = BSD_HWS_PGA_GEN7;
 542                         break;
 543                 case VECS0:
 544                         hwsp = VEBOX_HWS_PGA_GEN7;
 545                         break;
 546                 }
 547         } else if (IS_GEN(dev_priv, 6)) {
 548                 hwsp = RING_HWS_PGA_GEN6(engine->mmio_base);
 549         } else {
 550                 hwsp = RING_HWS_PGA(engine->mmio_base);
 551         }
 552
 553         I915_WRITE(hwsp, offset);
 554         POSTING_READ(hwsp);
 555 }
 556
 557 static void flush_cs_tlb(struct intel_engine_cs *engine)
 558 {
 559         struct drm_i915_private *dev_priv = engine->i915;
 560
 561         if (!IS_GEN_RANGE(dev_priv, 6, 7))
 562                 return;
 563
 564         /* ring should be idle before issuing a sync flush*/
 565         WARN_ON((ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
 566
 567         ENGINE_WRITE(engine, RING_INSTPM,
 568                      _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE |
 569                                         INSTPM_SYNC_FLUSH));
 570         if (intel_wait_for_register(engine->uncore,
 571                                     RING_INSTPM(engine->mmio_base),
 572                                     INSTPM_SYNC_FLUSH, 0,
 573                                     1000))
 574                 DRM_ERROR("%s: wait for SyncFlush to complete for TLB invalidation timed out\n",
 575                           engine->name);
 576 }
 577
 578 static void ring_setup_status_page(struct intel_engine_cs *engine)
 579 {
 580         set_hwsp(engine, i915_ggtt_offset(engine->status_page.vma));
 581         set_hwstam(engine, ~0u);
 582
 583         flush_cs_tlb(engine);
 584 }
 585
 586 static bool stop_ring(struct intel_engine_cs *engine)
 587 {
 588         struct drm_i915_private *dev_priv = engine->i915;
 589
 590         if (INTEL_GEN(dev_priv) > 2) {
 591                 ENGINE_WRITE(engine,
 592                              RING_MI_MODE, _MASKED_BIT_ENABLE(STOP_RING));
 593                 if (intel_wait_for_register(engine->uncore,
 594                                             RING_MI_MODE(engine->mmio_base),
 595                                             MODE_IDLE,
 596                                             MODE_IDLE,
 597                                             1000)) {
 598                         DRM_ERROR("%s : timed out trying to stop ring\n",
 599                                   engine->name);
 600
 601                         /*
 602                          * Sometimes we observe that the idle flag is not
 603                          * set even though the ring is empty. So double
 604                          * check before giving up.
 605                          */
 606                         if (ENGINE_READ(engine, RING_HEAD) !=
 607                             ENGINE_READ(engine, RING_TAIL))
 608                                 return false;
 609                 }
 610         }
 611
 612         ENGINE_WRITE(engine, RING_HEAD, ENGINE_READ(engine, RING_TAIL));
 613
 614         ENGINE_WRITE(engine, RING_HEAD, 0);
 615         ENGINE_WRITE(engine, RING_TAIL, 0);
 616
 617         /* The ring must be empty before it is disabled */
 618         ENGINE_WRITE(engine, RING_CTL, 0);
 619
 620         return (ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) == 0;
 621 }
 622
 623 static int xcs_resume(struct intel_engine_cs *engine)
 624 {
 625         struct drm_i915_private *dev_priv = engine->i915;
 626         struct intel_ring *ring = engine->buffer;
 627         int ret = 0;
 628
 629         GEM_TRACE("%s: ring:{HEAD:%04x, TAIL:%04x}\n",
 630                   engine->name, ring->head, ring->tail);
 631
 632         intel_uncore_forcewake_get(engine->uncore, FORCEWAKE_ALL);
 633
 634         if (!stop_ring(engine)) {
 635                 /* G45 ring initialization often fails to reset head to zero */
 636                 DRM_DEBUG_DRIVER("%s head not reset to zero "
 637                                 "ctl %08x head %08x tail %08x start %08x\n",
 638                                 engine->name,
 639                                 ENGINE_READ(engine, RING_CTL),
 640                                 ENGINE_READ(engine, RING_HEAD),
 641                                 ENGINE_READ(engine, RING_TAIL),
 642                                 ENGINE_READ(engine, RING_START));
 643
 644                 if (!stop_ring(engine)) {
 645                         DRM_ERROR("failed to set %s head to zero "
 646                                   "ctl %08x head %08x tail %08x start %08x\n",
 647                                   engine->name,
 648                                   ENGINE_READ(engine, RING_CTL),
 649                                   ENGINE_READ(engine, RING_HEAD),
 650                                   ENGINE_READ(engine, RING_TAIL),
 651                                   ENGINE_READ(engine, RING_START));
 652                         ret = -EIO;
 653                         goto out;
 654                 }
 655         }
 656
 657         if (HWS_NEEDS_PHYSICAL(dev_priv))
 658                 ring_setup_phys_status_page(engine);
 659         else
 660                 ring_setup_status_page(engine);
 661
 662         intel_engine_reset_breadcrumbs(engine);
 663
 664         /* Enforce ordering by reading HEAD register back */
 665         ENGINE_READ(engine, RING_HEAD);
 666
 667         /* Initialize the ring. This must happen _after_ we've cleared the ring
 668          * registers with the above sequence (the readback of the HEAD registers
 669          * also enforces ordering), otherwise the hw might lose the new ring
 670          * register values. */
 671         ENGINE_WRITE(engine, RING_START, i915_ggtt_offset(ring->vma));
 672
 673         /* WaClearRingBufHeadRegAtInit:ctg,elk */
 674         if (ENGINE_READ(engine, RING_HEAD))
 675                 DRM_DEBUG_DRIVER("%s initialization failed [head=%08x], fudging\n",
 676                                  engine->name, ENGINE_READ(engine, RING_HEAD));
 677
 678         /* Check that the ring offsets point within the ring! */
 679         GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->head));
 680         GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->tail));
 681         intel_ring_update_space(ring);
 682
 683         /* First wake the ring up to an empty/idle ring */
 684         ENGINE_WRITE(engine, RING_HEAD, ring->head);
 685         ENGINE_WRITE(engine, RING_TAIL, ring->head);
 686         ENGINE_POSTING_READ(engine, RING_TAIL);
 687
 688         ENGINE_WRITE(engine, RING_CTL, RING_CTL_SIZE(ring->size) | RING_VALID);
 689
 690         /* If the head is still not zero, the ring is dead */
 691         if (intel_wait_for_register(engine->uncore,
 692                                     RING_CTL(engine->mmio_base),
 693                                     RING_VALID, RING_VALID,
 694                                     50)) {
 695                 DRM_ERROR("%s initialization failed "
 696                           "ctl %08x (valid? %d) head %08x [%08x] tail %08x [%08x] start %08x [expected %08x]\n",
 697                           engine->name,
 698                           ENGINE_READ(engine, RING_CTL),
 699                           ENGINE_READ(engine, RING_CTL) & RING_VALID,
 700                           ENGINE_READ(engine, RING_HEAD), ring->head,
 701                           ENGINE_READ(engine, RING_TAIL), ring->tail,
 702                           ENGINE_READ(engine, RING_START),
 703                           i915_ggtt_offset(ring->vma));
 704                 ret = -EIO;
 705                 goto out;
 706         }
 707
 708         if (INTEL_GEN(dev_priv) > 2)
 709                 ENGINE_WRITE(engine,
 710                              RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
 711
 712         /* Now awake, let it get started */
 713         if (ring->tail != ring->head) {
 714                 ENGINE_WRITE(engine, RING_TAIL, ring->tail);
 715                 ENGINE_POSTING_READ(engine, RING_TAIL);
 716         }
 717
 718         /* Papering over lost _interrupts_ immediately following the restart */
 719         intel_engine_queue_breadcrumbs(engine);
 720 out:
 721         intel_uncore_forcewake_put(engine->uncore, FORCEWAKE_ALL);
 722
 723         return ret;
 724 }
 725
 726 static void reset_prepare(struct intel_engine_cs *engine)
 727 {
 728         intel_engine_stop_cs(engine);
 729 }
 730
 731 static void reset_ring(struct intel_engine_cs *engine, bool stalled)
 732 {
 733         struct i915_request *pos, *rq;
 734         unsigned long flags;
 735         u32 head;
 736
 737         rq = NULL;
 738         spin_lock_irqsave(&engine->active.lock, flags);
 739         list_for_each_entry(pos, &engine->active.requests, sched.link) {
 740                 if (!i915_request_completed(pos)) {
 741                         rq = pos;
 742                         break;
 743                 }
 744         }
 745
 746         /*
 747          * The guilty request will get skipped on a hung engine.
 748          *
 749          * Users of client default contexts do not rely on logical
 750          * state preserved between batches so it is safe to execute
 751          * queued requests following the hang. Non default contexts
 752          * rely on preserved state, so skipping a batch loses the
 753          * evolution of the state and it needs to be considered corrupted.
 754          * Executing more queued batches on top of corrupted state is
 755          * risky. But we take the risk by trying to advance through
 756          * the queued requests in order to make the client behaviour
 757          * more predictable around resets, by not throwing away random
 758          * amount of batches it has prepared for execution. Sophisticated
 759          * clients can use gem_reset_stats_ioctl and dma fence status
 760          * (exported via sync_file info ioctl on explicit fences) to observe
 761          * when it loses the context state and should rebuild accordingly.
 762          *
 763          * The context ban, and ultimately the client ban, mechanism are safety
 764          * valves if client submission ends up resulting in nothing more than
 765          * subsequent hangs.
 766          */
 767
 768         if (rq) {
 769                 /*
 770                  * Try to restore the logical GPU state to match the
 771                  * continuation of the request queue. If we skip the
 772                  * context/PD restore, then the next request may try to execute
 773                  * assuming that its context is valid and loaded on the GPU and
 774                  * so may try to access invalid memory, prompting repeated GPU
 775                  * hangs.
 776                  *
 777                  * If the request was guilty, we still restore the logical
 778                  * state in case the next request requires it (e.g. the
 779                  * aliasing ppgtt), but skip over the hung batch.
 780                  *
 781                  * If the request was innocent, we try to replay the request
 782                  * with the restored context.
 783                  */
 784                 i915_reset_request(rq, stalled);
 785
 786                 GEM_BUG_ON(rq->ring != engine->buffer);
 787                 head = rq->head;
 788         } else {
 789                 head = engine->buffer->tail;
 790         }
 791         engine->buffer->head = intel_ring_wrap(engine->buffer, head);
 792
 793         spin_unlock_irqrestore(&engine->active.lock, flags);
 794 }
 795
 796 static void reset_finish(struct intel_engine_cs *engine)
 797 {
 798 }
 799
 800 static int intel_rcs_ctx_init(struct i915_request *rq)
 801 {
 802         int ret;
 803
 804         ret = intel_engine_emit_ctx_wa(rq);
 805         if (ret != 0)
 806                 return ret;
 807
 808         ret = i915_gem_render_state_emit(rq);
 809         if (ret)
 810                 return ret;
 811
 812         return 0;
 813 }
 814
 815 static int rcs_resume(struct intel_engine_cs *engine)
 816 {
 817         struct drm_i915_private *dev_priv = engine->i915;
 818
 819         /*
 820          * Disable CONSTANT_BUFFER before it is loaded from the context
 821          * image. For as it is loaded, it is executed and the stored
 822          * address may no longer be valid, leading to a GPU hang.
 823          *
 824          * This imposes the requirement that userspace reload their
 825          * CONSTANT_BUFFER on every batch, fortunately a requirement
 826          * they are already accustomed to from before contexts were
 827          * enabled.
 828          */
 829         if (IS_GEN(dev_priv, 4))
 830                 I915_WRITE(ECOSKPD,
 831                            _MASKED_BIT_ENABLE(ECO_CONSTANT_BUFFER_SR_DISABLE));
 832
 833         /* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
 834         if (IS_GEN_RANGE(dev_priv, 4, 6))
 835                 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH));
 836
 837         /* We need to disable the AsyncFlip performance optimisations in order
 838          * to use MI_WAIT_FOR_EVENT within the CS. It should already be
 839          * programmed to '1' on all products.
 840          *
 841          * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv
 842          */
 843         if (IS_GEN_RANGE(dev_priv, 6, 7))
 844                 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
 845
 846         /* Required for the hardware to program scanline values for waiting */
 847         /* WaEnableFlushTlbInvalidationMode:snb */
 848         if (IS_GEN(dev_priv, 6))
 849                 I915_WRITE(GFX_MODE,
 850                            _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT));
 851
 852         /* WaBCSVCSTlbInvalidationMode:ivb,vlv,hsw */
 853         if (IS_GEN(dev_priv, 7))
 854                 I915_WRITE(GFX_MODE_GEN7,
 855                            _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT) |
 856                            _MASKED_BIT_ENABLE(GFX_REPLAY_MODE));
 857
 858         if (IS_GEN(dev_priv, 6)) {
 859                 /* From the Sandybridge PRM, volume 1 part 3, page 24:
 860                  * "If this bit is set, STCunit will have LRA as replacement
 861                  *  policy. [...] This bit must be reset.  LRA replacement
 862                  *  policy is not supported."
 863                  */
 864                 I915_WRITE(CACHE_MODE_0,
 865                            _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
 866         }
 867
 868         if (IS_GEN_RANGE(dev_priv, 6, 7))
 869                 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
 870
 871         return xcs_resume(engine);
 872 }
 873
 874 static void cancel_requests(struct intel_engine_cs *engine)
 875 {
 876         struct i915_request *request;
 877         unsigned long flags;
 878
 879         spin_lock_irqsave(&engine->active.lock, flags);
 880
 881         /* Mark all submitted requests as skipped. */
 882         list_for_each_entry(request, &engine->active.requests, sched.link) {
 883                 if (!i915_request_signaled(request))
 884                         dma_fence_set_error(&request->fence, -EIO);
 885
 886                 i915_request_mark_complete(request);
 887         }
 888
 889         /* Remaining _unready_ requests will be nop'ed when submitted */
 890
 891         spin_unlock_irqrestore(&engine->active.lock, flags);
 892 }
 893
 894 static void i9xx_submit_request(struct i915_request *request)
 895 {
 896         i915_request_submit(request);
 897
 898         ENGINE_WRITE(request->engine, RING_TAIL,
 899                      intel_ring_set_tail(request->ring, request->tail));
 900 }
 901
 902 static u32 *i9xx_emit_breadcrumb(struct i915_request *rq, u32 *cs)
 903 {
 904         GEM_BUG_ON(rq->timeline->hwsp_ggtt != rq->engine->status_page.vma);
 905         GEM_BUG_ON(offset_in_page(rq->timeline->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
 906
 907         *cs++ = MI_FLUSH;
 908
 909         *cs++ = MI_STORE_DWORD_INDEX;
 910         *cs++ = I915_GEM_HWS_SEQNO_ADDR;
 911         *cs++ = rq->fence.seqno;
 912
 913         *cs++ = MI_USER_INTERRUPT;
 914         *cs++ = MI_NOOP;
 915
 916         rq->tail = intel_ring_offset(rq, cs);
 917         assert_ring_tail_valid(rq->ring, rq->tail);
 918
 919         return cs;
 920 }
 921
 922 #define GEN5_WA_STORES 8 /* must be at least 1! */
 923 static u32 *gen5_emit_breadcrumb(struct i915_request *rq, u32 *cs)
 924 {
 925         int i;
 926
 927         GEM_BUG_ON(rq->timeline->hwsp_ggtt != rq->engine->status_page.vma);
 928         GEM_BUG_ON(offset_in_page(rq->timeline->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
 929
 930         *cs++ = MI_FLUSH;
 931
 932         BUILD_BUG_ON(GEN5_WA_STORES < 1);
 933         for (i = 0; i < GEN5_WA_STORES; i++) {
 934                 *cs++ = MI_STORE_DWORD_INDEX;
 935                 *cs++ = I915_GEM_HWS_SEQNO_ADDR;
 936                 *cs++ = rq->fence.seqno;
 937         }
 938
 939         *cs++ = MI_USER_INTERRUPT;
 940
 941         rq->tail = intel_ring_offset(rq, cs);
 942         assert_ring_tail_valid(rq->ring, rq->tail);
 943
 944         return cs;
 945 }
 946 #undef GEN5_WA_STORES
 947
 948 static void
 949 gen5_irq_enable(struct intel_engine_cs *engine)
 950 {
 951         gen5_enable_gt_irq(engine->i915, engine->irq_enable_mask);
 952 }
 953
 954 static void
 955 gen5_irq_disable(struct intel_engine_cs *engine)
 956 {
 957         gen5_disable_gt_irq(engine->i915, engine->irq_enable_mask);
 958 }
 959
 960 static void
 961 i9xx_irq_enable(struct intel_engine_cs *engine)
 962 {
 963         engine->i915->irq_mask &= ~engine->irq_enable_mask;
 964         intel_uncore_write(engine->uncore, GEN2_IMR, engine->i915->irq_mask);
 965         intel_uncore_posting_read_fw(engine->uncore, GEN2_IMR);
 966 }
 967
 968 static void
 969 i9xx_irq_disable(struct intel_engine_cs *engine)
 970 {
 971         engine->i915->irq_mask |= engine->irq_enable_mask;
 972         intel_uncore_write(engine->uncore, GEN2_IMR, engine->i915->irq_mask);
 973 }
 974
 975 static void
 976 i8xx_irq_enable(struct intel_engine_cs *engine)
 977 {
 978         struct drm_i915_private *i915 = engine->i915;
 979
 980         i915->irq_mask &= ~engine->irq_enable_mask;
 981         intel_uncore_write16(&i915->uncore, GEN2_IMR, i915->irq_mask);
 982         ENGINE_POSTING_READ16(engine, RING_IMR);
 983 }
 984
 985 static void
 986 i8xx_irq_disable(struct intel_engine_cs *engine)
 987 {
 988         struct drm_i915_private *i915 = engine->i915;
 989
 990         i915->irq_mask |= engine->irq_enable_mask;
 991         intel_uncore_write16(&i915->uncore, GEN2_IMR, i915->irq_mask);
 992 }
 993
 994 static int
 995 bsd_ring_flush(struct i915_request *rq, u32 mode)
 996 {
 997         u32 *cs;
 998
 999         cs = intel_ring_begin(rq, 2);
1000         if (IS_ERR(cs))
1001                 return PTR_ERR(cs);
1002
1003         *cs++ = MI_FLUSH;
1004         *cs++ = MI_NOOP;
1005         intel_ring_advance(rq, cs);
1006         return 0;
1007 }
1008
1009 static void
1010 gen6_irq_enable(struct intel_engine_cs *engine)
1011 {
1012         ENGINE_WRITE(engine, RING_IMR,
1013                      ~(engine->irq_enable_mask | engine->irq_keep_mask));
1014
1015         /* Flush/delay to ensure the RING_IMR is active before the GT IMR */
1016         ENGINE_POSTING_READ(engine, RING_IMR);
1017
1018         gen5_enable_gt_irq(engine->i915, engine->irq_enable_mask);
1019 }
1020
1021 static void
1022 gen6_irq_disable(struct intel_engine_cs *engine)
1023 {
1024         ENGINE_WRITE(engine, RING_IMR, ~engine->irq_keep_mask);
1025         gen5_disable_gt_irq(engine->i915, engine->irq_enable_mask);
1026 }
1027
1028 static void
1029 hsw_vebox_irq_enable(struct intel_engine_cs *engine)
1030 {
1031         ENGINE_WRITE(engine, RING_IMR, ~engine->irq_enable_mask);
1032
1033         /* Flush/delay to ensure the RING_IMR is active before the GT IMR */
1034         ENGINE_POSTING_READ(engine, RING_IMR);
1035
1036         gen6_unmask_pm_irq(engine->i915, engine->irq_enable_mask);
1037 }
1038
1039 static void
1040 hsw_vebox_irq_disable(struct intel_engine_cs *engine)
1041 {
1042         ENGINE_WRITE(engine, RING_IMR, ~0);
1043         gen6_mask_pm_irq(engine->i915, engine->irq_enable_mask);
1044 }
1045
1046 static int
1047 i965_emit_bb_start(struct i915_request *rq,
1048                    u64 offset, u32 length,
1049                    unsigned int dispatch_flags)
1050 {
1051         u32 *cs;
1052
1053         cs = intel_ring_begin(rq, 2);
1054         if (IS_ERR(cs))
1055                 return PTR_ERR(cs);
1056
1057         *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT | (dispatch_flags &
1058                 I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE_I965);
1059         *cs++ = offset;
1060         intel_ring_advance(rq, cs);
1061
1062         return 0;
1063 }
1064
1065 /* Just userspace ABI convention to limit the wa batch bo to a resonable size */
1066 #define I830_BATCH_LIMIT SZ_256K
1067 #define I830_TLB_ENTRIES (2)
1068 #define I830_WA_SIZE max(I830_TLB_ENTRIES*4096, I830_BATCH_LIMIT)
1069 static int
1070 i830_emit_bb_start(struct i915_request *rq,
1071                    u64 offset, u32 len,
1072                    unsigned int dispatch_flags)
1073 {
1074         u32 *cs, cs_offset = i915_scratch_offset(rq->i915);
1075
1076         GEM_BUG_ON(rq->i915->gt.scratch->size < I830_WA_SIZE);
1077
1078         cs = intel_ring_begin(rq, 6);
1079         if (IS_ERR(cs))
1080                 return PTR_ERR(cs);
1081
1082         /* Evict the invalid PTE TLBs */
1083         *cs++ = COLOR_BLT_CMD | BLT_WRITE_RGBA;
1084         *cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | 4096;
1085         *cs++ = I830_TLB_ENTRIES << 16 | 4; /* load each page */
1086         *cs++ = cs_offset;
1087         *cs++ = 0xdeadbeef;
1088         *cs++ = MI_NOOP;
1089         intel_ring_advance(rq, cs);
1090
1091         if ((dispatch_flags & I915_DISPATCH_PINNED) == 0) {
1092                 if (len > I830_BATCH_LIMIT)
1093                         return -ENOSPC;
1094
1095                 cs = intel_ring_begin(rq, 6 + 2);
1096                 if (IS_ERR(cs))
1097                         return PTR_ERR(cs);
1098
1099                 /* Blit the batch (which has now all relocs applied) to the
1100                  * stable batch scratch bo area (so that the CS never
1101                  * stumbles over its tlb invalidation bug) ...
1102                  */
1103                 *cs++ = SRC_COPY_BLT_CMD | BLT_WRITE_RGBA;
1104                 *cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | 4096;
1105                 *cs++ = DIV_ROUND_UP(len, 4096) << 16 | 4096;
1106                 *cs++ = cs_offset;
1107                 *cs++ = 4096;
1108                 *cs++ = offset;
1109
1110                 *cs++ = MI_FLUSH;
1111                 *cs++ = MI_NOOP;
1112                 intel_ring_advance(rq, cs);
1113
1114                 /* ... and execute it. */
1115                 offset = cs_offset;
1116         }
1117
1118         cs = intel_ring_begin(rq, 2);
1119         if (IS_ERR(cs))
1120                 return PTR_ERR(cs);
1121
1122         *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
1123         *cs++ = offset | (dispatch_flags & I915_DISPATCH_SECURE ? 0 :
1124                 MI_BATCH_NON_SECURE);
1125         intel_ring_advance(rq, cs);
1126
1127         return 0;
1128 }
1129
1130 static int
1131 i915_emit_bb_start(struct i915_request *rq,
1132                    u64 offset, u32 len,
1133                    unsigned int dispatch_flags)
1134 {
1135         u32 *cs;
1136
1137         cs = intel_ring_begin(rq, 2);
1138         if (IS_ERR(cs))
1139                 return PTR_ERR(cs);
1140
1141         *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
1142         *cs++ = offset | (dispatch_flags & I915_DISPATCH_SECURE ? 0 :
1143                 MI_BATCH_NON_SECURE);
1144         intel_ring_advance(rq, cs);
1145
1146         return 0;
1147 }
1148
1149 int intel_ring_pin(struct intel_ring *ring)
1150 {
1151         struct i915_vma *vma = ring->vma;
1152         unsigned int flags;
1153         void *addr;
1154         int ret;
1155
1156         if (atomic_fetch_inc(&ring->pin_count))
1157                 return 0;
1158
1159         ret = i915_timeline_pin(ring->timeline);
1160         if (ret)
1161                 goto err_unpin;
1162
1163         flags = PIN_GLOBAL;
1164
1165         /* Ring wraparound at offset 0 sometimes hangs. No idea why. */
1166         flags |= PIN_OFFSET_BIAS | i915_ggtt_pin_bias(vma);
1167
1168         if (vma->obj->stolen)
1169                 flags |= PIN_MAPPABLE;
1170         else
1171                 flags |= PIN_HIGH;
1172
1173         ret = i915_vma_pin(vma, 0, 0, flags);
1174         if (unlikely(ret))
1175                 goto err_timeline;
1176
1177         if (i915_vma_is_map_and_fenceable(vma))
1178                 addr = (void __force *)i915_vma_pin_iomap(vma);
1179         else
1180                 addr = i915_gem_object_pin_map(vma->obj,
1181                                                i915_coherent_map_type(vma->vm->i915));
1182         if (IS_ERR(addr)) {
1183                 ret = PTR_ERR(addr);
1184                 goto err_ring;
1185         }
1186
1187         vma->obj->pin_global++;
1188
1189         GEM_BUG_ON(ring->vaddr);
1190         ring->vaddr = addr;
1191
1192         return 0;
1193
1194 err_ring:
1195         i915_vma_unpin(vma);
1196 err_timeline:
1197         i915_timeline_unpin(ring->timeline);
1198 err_unpin:
1199         atomic_dec(&ring->pin_count);
1200         return ret;
1201 }
1202
1203 void intel_ring_reset(struct intel_ring *ring, u32 tail)
1204 {
1205         GEM_BUG_ON(!intel_ring_offset_valid(ring, tail));
1206
1207         ring->tail = tail;
1208         ring->head = tail;
1209         ring->emit = tail;
1210         intel_ring_update_space(ring);
1211 }
1212
1213 void intel_ring_unpin(struct intel_ring *ring)
1214 {
1215         if (!atomic_dec_and_test(&ring->pin_count))
1216                 return;
1217
1218         /* Discard any unused bytes beyond that submitted to hw. */
1219         intel_ring_reset(ring, ring->tail);
1220
1221         GEM_BUG_ON(!ring->vma);
1222         if (i915_vma_is_map_and_fenceable(ring->vma))
1223                 i915_vma_unpin_iomap(ring->vma);
1224         else
1225                 i915_gem_object_unpin_map(ring->vma->obj);
1226
1227         GEM_BUG_ON(!ring->vaddr);
1228         ring->vaddr = NULL;
1229
1230         ring->vma->obj->pin_global--;
1231         i915_vma_unpin(ring->vma);
1232
1233         i915_timeline_unpin(ring->timeline);
1234 }
1235
1236 static struct i915_vma *
1237 intel_ring_create_vma(struct drm_i915_private *dev_priv, int size)
1238 {
1239         struct i915_address_space *vm = &dev_priv->ggtt.vm;
1240         struct drm_i915_gem_object *obj;
1241         struct i915_vma *vma;
1242
1243         obj = i915_gem_object_create_stolen(dev_priv, size);
1244         if (!obj)
1245                 obj = i915_gem_object_create_internal(dev_priv, size);
1246         if (IS_ERR(obj))
1247                 return ERR_CAST(obj);
1248
1249         /*
1250          * Mark ring buffers as read-only from GPU side (so no stray overwrites)
1251          * if supported by the platform's GGTT.
1252          */
1253         if (vm->has_read_only)
1254                 i915_gem_object_set_readonly(obj);
1255
1256         vma = i915_vma_instance(obj, vm, NULL);
1257         if (IS_ERR(vma))
1258                 goto err;
1259
1260         return vma;
1261
1262 err:
1263         i915_gem_object_put(obj);
1264         return vma;
1265 }
1266
1267 struct intel_ring *
1268 intel_engine_create_ring(struct intel_engine_cs *engine,
1269                          struct i915_timeline *timeline,
1270                          int size)
1271 {
1272         struct intel_ring *ring;
1273         struct i915_vma *vma;
1274
1275         GEM_BUG_ON(!is_power_of_2(size));
1276         GEM_BUG_ON(RING_CTL_SIZE(size) & ~RING_NR_PAGES);
1277
1278         ring = kzalloc(sizeof(*ring), GFP_KERNEL);
1279         if (!ring)
1280                 return ERR_PTR(-ENOMEM);
1281
1282         kref_init(&ring->ref);
1283         INIT_LIST_HEAD(&ring->request_list);
1284         ring->timeline = i915_timeline_get(timeline);
1285
1286         ring->size = size;
1287         /* Workaround an erratum on the i830 which causes a hang if
1288          * the TAIL pointer points to within the last 2 cachelines
1289          * of the buffer.
1290          */
1291         ring->effective_size = size;
1292         if (IS_I830(engine->i915) || IS_I845G(engine->i915))
1293                 ring->effective_size -= 2 * CACHELINE_BYTES;
1294
1295         intel_ring_update_space(ring);
1296
1297         vma = intel_ring_create_vma(engine->i915, size);
1298         if (IS_ERR(vma)) {
1299                 kfree(ring);
1300                 return ERR_CAST(vma);
1301         }
1302         ring->vma = vma;
1303
1304         return ring;
1305 }
1306
1307 void intel_ring_free(struct kref *ref)
1308 {
1309         struct intel_ring *ring = container_of(ref, typeof(*ring), ref);
1310
1311         i915_vma_close(ring->vma);
1312         i915_vma_put(ring->vma);
1313
1314         i915_timeline_put(ring->timeline);
1315         kfree(ring);
1316 }
1317
1318 static void __ring_context_fini(struct intel_context *ce)
1319 {
1320         GEM_BUG_ON(i915_gem_object_is_active(ce->state->obj));
1321         i915_gem_object_put(ce->state->obj);
1322 }
1323
1324 static void ring_context_destroy(struct kref *ref)
1325 {
1326         struct intel_context *ce = container_of(ref, typeof(*ce), ref);
1327
1328         GEM_BUG_ON(intel_context_is_pinned(ce));
1329
1330         if (ce->state)
1331                 __ring_context_fini(ce);
1332
1333         intel_context_free(ce);
1334 }
1335
1336 static int __context_pin_ppgtt(struct i915_gem_context *ctx)
1337 {
1338         struct i915_address_space *vm;
1339         int err = 0;
1340
1341         vm = ctx->vm ?: &ctx->i915->mm.aliasing_ppgtt->vm;
1342         if (vm)
1343                 err = gen6_ppgtt_pin(i915_vm_to_ppgtt((vm)));
1344
1345         return err;
1346 }
1347
1348 static void __context_unpin_ppgtt(struct i915_gem_context *ctx)
1349 {
1350         struct i915_address_space *vm;
1351
1352         vm = ctx->vm ?: &ctx->i915->mm.aliasing_ppgtt->vm;
1353         if (vm)
1354                 gen6_ppgtt_unpin(i915_vm_to_ppgtt(vm));
1355 }
1356
1357 static void ring_context_unpin(struct intel_context *ce)
1358 {
1359         __context_unpin_ppgtt(ce->gem_context);
1360 }
1361
1362 static struct i915_vma *
1363 alloc_context_vma(struct intel_engine_cs *engine)
1364 {
1365         struct drm_i915_private *i915 = engine->i915;
1366         struct drm_i915_gem_object *obj;
1367         struct i915_vma *vma;
1368         int err;
1369
1370         obj = i915_gem_object_create_shmem(i915, engine->context_size);
1371         if (IS_ERR(obj))
1372                 return ERR_CAST(obj);
1373
1374         /*
1375          * Try to make the context utilize L3 as well as LLC.
1376          *
1377          * On VLV we don't have L3 controls in the PTEs so we
1378          * shouldn't touch the cache level, especially as that
1379          * would make the object snooped which might have a
1380          * negative performance impact.
1381          *
1382          * Snooping is required on non-llc platforms in execlist
1383          * mode, but since all GGTT accesses use PAT entry 0 we
1384          * get snooping anyway regardless of cache_level.
1385          *
1386          * This is only applicable for Ivy Bridge devices since
1387          * later platforms don't have L3 control bits in the PTE.
1388          */
1389         if (IS_IVYBRIDGE(i915))
1390                 i915_gem_object_set_cache_coherency(obj, I915_CACHE_L3_LLC);
1391
1392         if (engine->default_state) {
1393                 void *defaults, *vaddr;
1394
1395                 vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
1396                 if (IS_ERR(vaddr)) {
1397                         err = PTR_ERR(vaddr);
1398                         goto err_obj;
1399                 }
1400
1401                 defaults = i915_gem_object_pin_map(engine->default_state,
1402                                                    I915_MAP_WB);
1403                 if (IS_ERR(defaults)) {
1404                         err = PTR_ERR(defaults);
1405                         goto err_map;
1406                 }
1407
1408                 memcpy(vaddr, defaults, engine->context_size);
1409                 i915_gem_object_unpin_map(engine->default_state);
1410
1411                 i915_gem_object_flush_map(obj);
1412                 i915_gem_object_unpin_map(obj);
1413         }
1414
1415         vma = i915_vma_instance(obj, &i915->ggtt.vm, NULL);
1416         if (IS_ERR(vma)) {
1417                 err = PTR_ERR(vma);
1418                 goto err_obj;
1419         }
1420
1421         return vma;
1422
1423 err_map:
1424         i915_gem_object_unpin_map(obj);
1425 err_obj:
1426         i915_gem_object_put(obj);
1427         return ERR_PTR(err);
1428 }
1429
1430 static int ring_context_pin(struct intel_context *ce)
1431 {
1432         struct intel_engine_cs *engine = ce->engine;
1433         int err;
1434
1435         /* One ringbuffer to rule them all */
1436         GEM_BUG_ON(!engine->buffer);
1437         ce->ring = engine->buffer;
1438
1439         if (!ce->state && engine->context_size) {
1440                 struct i915_vma *vma;
1441
1442                 vma = alloc_context_vma(engine);
1443                 if (IS_ERR(vma))
1444                         return PTR_ERR(vma);
1445
1446                 ce->state = vma;
1447         }
1448
1449         err = intel_context_active_acquire(ce, PIN_HIGH);
1450         if (err)
1451                 return err;
1452
1453         err = __context_pin_ppgtt(ce->gem_context);
1454         if (err)
1455                 goto err_active;
1456
1457         return 0;
1458
1459 err_active:
1460         intel_context_active_release(ce);
1461         return err;
1462 }
1463
1464 static void ring_context_reset(struct intel_context *ce)
1465 {
1466         intel_ring_reset(ce->ring, 0);
1467 }
1468
1469 static const struct intel_context_ops ring_context_ops = {
1470         .pin = ring_context_pin,
1471         .unpin = ring_context_unpin,
1472
1473         .enter = intel_context_enter_engine,
1474         .exit = intel_context_exit_engine,
1475
1476         .reset = ring_context_reset,
1477         .destroy = ring_context_destroy,
1478 };
1479
1480 static int load_pd_dir(struct i915_request *rq, const struct i915_ppgtt *ppgtt)
1481 {
1482         const struct intel_engine_cs * const engine = rq->engine;
1483         u32 *cs;
1484
1485         cs = intel_ring_begin(rq, 6);
1486         if (IS_ERR(cs))
1487                 return PTR_ERR(cs);
1488
1489         *cs++ = MI_LOAD_REGISTER_IMM(1);
1490         *cs++ = i915_mmio_reg_offset(RING_PP_DIR_DCLV(engine->mmio_base));
1491         *cs++ = PP_DIR_DCLV_2G;
1492
1493         *cs++ = MI_LOAD_REGISTER_IMM(1);
1494         *cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
1495         *cs++ = ppgtt->pd->base.ggtt_offset << 10;
1496
1497         intel_ring_advance(rq, cs);
1498
1499         return 0;
1500 }
1501
1502 static int flush_pd_dir(struct i915_request *rq)
1503 {
1504         const struct intel_engine_cs * const engine = rq->engine;
1505         u32 *cs;
1506
1507         cs = intel_ring_begin(rq, 4);
1508         if (IS_ERR(cs))
1509                 return PTR_ERR(cs);
1510
1511         /* Stall until the page table load is complete */
1512         *cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
1513         *cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
1514         *cs++ = i915_scratch_offset(rq->i915);
1515         *cs++ = MI_NOOP;
1516
1517         intel_ring_advance(rq, cs);
1518         return 0;
1519 }
1520
1521 static inline int mi_set_context(struct i915_request *rq, u32 flags)
1522 {
1523         struct drm_i915_private *i915 = rq->i915;
1524         struct intel_engine_cs *engine = rq->engine;
1525         enum intel_engine_id id;
1526         const int num_engines =
1527                 IS_HSW_GT1(i915) ? RUNTIME_INFO(i915)->num_engines - 1 : 0;
1528         bool force_restore = false;
1529         int len;
1530         u32 *cs;
1531
1532         flags |= MI_MM_SPACE_GTT;
1533         if (IS_HASWELL(i915))
1534                 /* These flags are for resource streamer on HSW+ */
1535                 flags |= HSW_MI_RS_SAVE_STATE_EN | HSW_MI_RS_RESTORE_STATE_EN;
1536         else
1537                 /* We need to save the extended state for powersaving modes */
1538                 flags |= MI_SAVE_EXT_STATE_EN | MI_RESTORE_EXT_STATE_EN;
1539
1540         len = 4;
1541         if (IS_GEN(i915, 7))
1542                 len += 2 + (num_engines ? 4 * num_engines + 6 : 0);
1543         else if (IS_GEN(i915, 5))
1544                 len += 2;
1545         if (flags & MI_FORCE_RESTORE) {
1546                 GEM_BUG_ON(flags & MI_RESTORE_INHIBIT);
1547                 flags &= ~MI_FORCE_RESTORE;
1548                 force_restore = true;
1549                 len += 2;
1550         }
1551
1552         cs = intel_ring_begin(rq, len);
1553         if (IS_ERR(cs))
1554                 return PTR_ERR(cs);
1555
1556         /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
1557         if (IS_GEN(i915, 7)) {
1558                 *cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
1559                 if (num_engines) {
1560                         struct intel_engine_cs *signaller;
1561
1562                         *cs++ = MI_LOAD_REGISTER_IMM(num_engines);
1563                         for_each_engine(signaller, i915, id) {
1564                                 if (signaller == engine)
1565                                         continue;
1566
1567                                 *cs++ = i915_mmio_reg_offset(
1568                                            RING_PSMI_CTL(signaller->mmio_base));
1569                                 *cs++ = _MASKED_BIT_ENABLE(
1570                                                 GEN6_PSMI_SLEEP_MSG_DISABLE);
1571                         }
1572                 }
1573         } else if (IS_GEN(i915, 5)) {
1574                 /*
1575                  * This w/a is only listed for pre-production ilk a/b steppings,
1576                  * but is also mentioned for programming the powerctx. To be
1577                  * safe, just apply the workaround; we do not use SyncFlush so
1578                  * this should never take effect and so be a no-op!
1579                  */
1580                 *cs++ = MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN;
1581         }
1582
1583         if (force_restore) {
1584                 /*
1585                  * The HW doesn't handle being told to restore the current
1586                  * context very well. Quite often it likes goes to go off and
1587                  * sulk, especially when it is meant to be reloading PP_DIR.
1588                  * A very simple fix to force the reload is to simply switch
1589                  * away from the current context and back again.
1590                  *
1591                  * Note that the kernel_context will contain random state
1592                  * following the INHIBIT_RESTORE. We accept this since we
1593                  * never use the kernel_context state; it is merely a
1594                  * placeholder we use to flush other contexts.
1595                  */
1596                 *cs++ = MI_SET_CONTEXT;
1597                 *cs++ = i915_ggtt_offset(engine->kernel_context->state) |
1598                         MI_MM_SPACE_GTT |
1599                         MI_RESTORE_INHIBIT;
1600         }
1601
1602         *cs++ = MI_NOOP;
1603         *cs++ = MI_SET_CONTEXT;
1604         *cs++ = i915_ggtt_offset(rq->hw_context->state) | flags;
1605         /*
1606          * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
1607          * WaMiSetContext_Hang:snb,ivb,vlv
1608          */
1609         *cs++ = MI_NOOP;
1610
1611         if (IS_GEN(i915, 7)) {
1612                 if (num_engines) {
1613                         struct intel_engine_cs *signaller;
1614                         i915_reg_t last_reg = {}; /* keep gcc quiet */
1615
1616                         *cs++ = MI_LOAD_REGISTER_IMM(num_engines);
1617                         for_each_engine(signaller, i915, id) {
1618                                 if (signaller == engine)
1619                                         continue;
1620
1621                                 last_reg = RING_PSMI_CTL(signaller->mmio_base);
1622                                 *cs++ = i915_mmio_reg_offset(last_reg);
1623                                 *cs++ = _MASKED_BIT_DISABLE(
1624                                                 GEN6_PSMI_SLEEP_MSG_DISABLE);
1625                         }
1626
1627                         /* Insert a delay before the next switch! */
1628                         *cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
1629                         *cs++ = i915_mmio_reg_offset(last_reg);
1630                         *cs++ = i915_scratch_offset(rq->i915);
1631                         *cs++ = MI_NOOP;
1632                 }
1633                 *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1634         } else if (IS_GEN(i915, 5)) {
1635                 *cs++ = MI_SUSPEND_FLUSH;
1636         }
1637
1638         intel_ring_advance(rq, cs);
1639
1640         return 0;
1641 }
1642
1643 static int remap_l3(struct i915_request *rq, int slice)
1644 {
1645         u32 *cs, *remap_info = rq->i915->l3_parity.remap_info[slice];
1646         int i;
1647
1648         if (!remap_info)
1649                 return 0;
1650
1651         cs = intel_ring_begin(rq, GEN7_L3LOG_SIZE/4 * 2 + 2);
1652         if (IS_ERR(cs))
1653                 return PTR_ERR(cs);
1654
1655         /*
1656          * Note: We do not worry about the concurrent register cacheline hang
1657          * here because no other code should access these registers other than
1658          * at initialization time.
1659          */
1660         *cs++ = MI_LOAD_REGISTER_IMM(GEN7_L3LOG_SIZE/4);
1661         for (i = 0; i < GEN7_L3LOG_SIZE/4; i++) {
1662                 *cs++ = i915_mmio_reg_offset(GEN7_L3LOG(slice, i));
1663                 *cs++ = remap_info[i];
1664         }
1665         *cs++ = MI_NOOP;
1666         intel_ring_advance(rq, cs);
1667
1668         return 0;
1669 }
1670
1671 static int switch_context(struct i915_request *rq)
1672 {
1673         struct intel_engine_cs *engine = rq->engine;
1674         struct i915_gem_context *ctx = rq->gem_context;
1675         struct i915_address_space *vm =
1676                 ctx->vm ?: &rq->i915->mm.aliasing_ppgtt->vm;
1677         unsigned int unwind_mm = 0;
1678         u32 hw_flags = 0;
1679         int ret, i;
1680
1681         GEM_BUG_ON(HAS_EXECLISTS(rq->i915));
1682
1683         if (vm) {
1684                 struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1685                 int loops;
1686
1687                 /*
1688                  * Baytail takes a little more convincing that it really needs
1689                  * to reload the PD between contexts. It is not just a little
1690                  * longer, as adding more stalls after the load_pd_dir (i.e.
1691                  * adding a long loop around flush_pd_dir) is not as effective
1692                  * as reloading the PD umpteen times. 32 is derived from
1693                  * experimentation (gem_exec_parallel/fds) and has no good
1694                  * explanation.
1695                  */
1696                 loops = 1;
1697                 if (engine->id == BCS0 && IS_VALLEYVIEW(engine->i915))
1698                         loops = 32;
1699
1700                 do {
1701                         ret = load_pd_dir(rq, ppgtt);
1702                         if (ret)
1703                                 goto err;
1704                 } while (--loops);
1705
1706                 if (ppgtt->pd_dirty_engines & engine->mask) {
1707                         unwind_mm = engine->mask;
1708                         ppgtt->pd_dirty_engines &= ~unwind_mm;
1709                         hw_flags = MI_FORCE_RESTORE;
1710                 }
1711         }
1712
1713         if (rq->hw_context->state) {
1714                 GEM_BUG_ON(engine->id != RCS0);
1715
1716                 /*
1717                  * The kernel context(s) is treated as pure scratch and is not
1718                  * expected to retain any state (as we sacrifice it during
1719                  * suspend and on resume it may be corrupted). This is ok,
1720                  * as nothing actually executes using the kernel context; it
1721                  * is purely used for flushing user contexts.
1722                  */
1723                 if (i915_gem_context_is_kernel(ctx))
1724                         hw_flags = MI_RESTORE_INHIBIT;
1725
1726                 ret = mi_set_context(rq, hw_flags);
1727                 if (ret)
1728                         goto err_mm;
1729         }
1730
1731         if (vm) {
1732                 ret = engine->emit_flush(rq, EMIT_INVALIDATE);
1733                 if (ret)
1734                         goto err_mm;
1735
1736                 ret = flush_pd_dir(rq);
1737                 if (ret)
1738                         goto err_mm;
1739
1740                 /*
1741                  * Not only do we need a full barrier (post-sync write) after
1742                  * invalidating the TLBs, but we need to wait a little bit
1743                  * longer. Whether this is merely delaying us, or the
1744                  * subsequent flush is a key part of serialising with the
1745                  * post-sync op, this extra pass appears vital before a
1746                  * mm switch!
1747                  */
1748                 ret = engine->emit_flush(rq, EMIT_INVALIDATE);
1749                 if (ret)
1750                         goto err_mm;
1751
1752                 ret = engine->emit_flush(rq, EMIT_FLUSH);
1753                 if (ret)
1754                         goto err_mm;
1755         }
1756
1757         if (ctx->remap_slice) {
1758                 for (i = 0; i < MAX_L3_SLICES; i++) {
1759                         if (!(ctx->remap_slice & BIT(i)))
1760                                 continue;
1761
1762                         ret = remap_l3(rq, i);
1763                         if (ret)
1764                                 goto err_mm;
1765                 }
1766
1767                 ctx->remap_slice = 0;
1768         }
1769
1770         return 0;
1771
1772 err_mm:
1773         if (unwind_mm)
1774                 i915_vm_to_ppgtt(vm)->pd_dirty_engines |= unwind_mm;
1775 err:
1776         return ret;
1777 }
1778
1779 static int ring_request_alloc(struct i915_request *request)
1780 {
1781         int ret;
1782
1783         GEM_BUG_ON(!intel_context_is_pinned(request->hw_context));
1784         GEM_BUG_ON(request->timeline->has_initial_breadcrumb);
1785
1786         /*
1787          * Flush enough space to reduce the likelihood of waiting after
1788          * we start building the request - in which case we will just
1789          * have to repeat work.
1790          */
1791         request->reserved_space += LEGACY_REQUEST_SIZE;
1792
1793         /* Unconditionally invalidate GPU caches and TLBs. */
1794         ret = request->engine->emit_flush(request, EMIT_INVALIDATE);
1795         if (ret)
1796                 return ret;
1797
1798         ret = switch_context(request);
1799         if (ret)
1800                 return ret;
1801
1802         request->reserved_space -= LEGACY_REQUEST_SIZE;
1803         return 0;
1804 }
1805
1806 static noinline int wait_for_space(struct intel_ring *ring, unsigned int bytes)
1807 {
1808         struct i915_request *target;
1809         long timeout;
1810
1811         if (intel_ring_update_space(ring) >= bytes)
1812                 return 0;
1813
1814         GEM_BUG_ON(list_empty(&ring->request_list));
1815         list_for_each_entry(target, &ring->request_list, ring_link) {
1816                 /* Would completion of this request free enough space? */
1817                 if (bytes <= __intel_ring_space(target->postfix,
1818                                                 ring->emit, ring->size))
1819                         break;
1820         }
1821
1822         if (WARN_ON(&target->ring_link == &ring->request_list))
1823                 return -ENOSPC;
1824
1825         timeout = i915_request_wait(target,
1826                                     I915_WAIT_INTERRUPTIBLE,
1827                                     MAX_SCHEDULE_TIMEOUT);
1828         if (timeout < 0)
1829                 return timeout;
1830
1831         i915_request_retire_upto(target);
1832
1833         intel_ring_update_space(ring);
1834         GEM_BUG_ON(ring->space < bytes);
1835         return 0;
1836 }
1837
1838 u32 *intel_ring_begin(struct i915_request *rq, unsigned int num_dwords)
1839 {
1840         struct intel_ring *ring = rq->ring;
1841         const unsigned int remain_usable = ring->effective_size - ring->emit;
1842         const unsigned int bytes = num_dwords * sizeof(u32);
1843         unsigned int need_wrap = 0;
1844         unsigned int total_bytes;
1845         u32 *cs;
1846
1847         /* Packets must be qword aligned. */
1848         GEM_BUG_ON(num_dwords & 1);
1849
1850         total_bytes = bytes + rq->reserved_space;
1851         GEM_BUG_ON(total_bytes > ring->effective_size);
1852
1853         if (unlikely(total_bytes > remain_usable)) {
1854                 const int remain_actual = ring->size - ring->emit;
1855
1856                 if (bytes > remain_usable) {
1857                         /*
1858                          * Not enough space for the basic request. So need to
1859                          * flush out the remainder and then wait for
1860                          * base + reserved.
1861                          */
1862                         total_bytes += remain_actual;
1863                         need_wrap = remain_actual | 1;
1864                 } else  {
1865                         /*
1866                          * The base request will fit but the reserved space
1867                          * falls off the end. So we don't need an immediate
1868                          * wrap and only need to effectively wait for the
1869                          * reserved size from the start of ringbuffer.
1870                          */
1871                         total_bytes = rq->reserved_space + remain_actual;
1872                 }
1873         }
1874
1875         if (unlikely(total_bytes > ring->space)) {
1876                 int ret;
1877
1878                 /*
1879                  * Space is reserved in the ringbuffer for finalising the
1880                  * request, as that cannot be allowed to fail. During request
1881                  * finalisation, reserved_space is set to 0 to stop the
1882                  * overallocation and the assumption is that then we never need
1883                  * to wait (which has the risk of failing with EINTR).
1884                  *
1885                  * See also i915_request_alloc() and i915_request_add().
1886                  */
1887                 GEM_BUG_ON(!rq->reserved_space);
1888
1889                 ret = wait_for_space(ring, total_bytes);
1890                 if (unlikely(ret))
1891                         return ERR_PTR(ret);
1892         }
1893
1894         if (unlikely(need_wrap)) {
1895                 need_wrap &= ~1;
1896                 GEM_BUG_ON(need_wrap > ring->space);
1897                 GEM_BUG_ON(ring->emit + need_wrap > ring->size);
1898                 GEM_BUG_ON(!IS_ALIGNED(need_wrap, sizeof(u64)));
1899
1900                 /* Fill the tail with MI_NOOP */
1901                 memset64(ring->vaddr + ring->emit, 0, need_wrap / sizeof(u64));
1902                 ring->space -= need_wrap;
1903                 ring->emit = 0;
1904         }
1905
1906         GEM_BUG_ON(ring->emit > ring->size - bytes);
1907         GEM_BUG_ON(ring->space < bytes);
1908         cs = ring->vaddr + ring->emit;
1909         GEM_DEBUG_EXEC(memset32(cs, POISON_INUSE, bytes / sizeof(*cs)));
1910         ring->emit += bytes;
1911         ring->space -= bytes;
1912
1913         return cs;
1914 }
1915
1916 /* Align the ring tail to a cacheline boundary */
1917 int intel_ring_cacheline_align(struct i915_request *rq)
1918 {
1919         int num_dwords;
1920         void *cs;
1921
1922         num_dwords = (rq->ring->emit & (CACHELINE_BYTES - 1)) / sizeof(u32);
1923         if (num_dwords == 0)
1924                 return 0;
1925
1926         num_dwords = CACHELINE_DWORDS - num_dwords;
1927         GEM_BUG_ON(num_dwords & 1);
1928
1929         cs = intel_ring_begin(rq, num_dwords);
1930         if (IS_ERR(cs))
1931                 return PTR_ERR(cs);
1932
1933         memset64(cs, (u64)MI_NOOP << 32 | MI_NOOP, num_dwords / 2);
1934         intel_ring_advance(rq, cs);
1935
1936         GEM_BUG_ON(rq->ring->emit & (CACHELINE_BYTES - 1));
1937         return 0;
1938 }
1939
1940 static void gen6_bsd_submit_request(struct i915_request *request)
1941 {
1942         struct intel_uncore *uncore = request->engine->uncore;
1943
1944         intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
1945
1946        /* Every tail move must follow the sequence below */
1947
1948         /* Disable notification that the ring is IDLE. The GT
1949          * will then assume that it is busy and bring it out of rc6.
1950          */
1951         intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
1952                               _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1953
1954         /* Clear the context id. Here be magic! */
1955         intel_uncore_write64_fw(uncore, GEN6_BSD_RNCID, 0x0);
1956
1957         /* Wait for the ring not to be idle, i.e. for it to wake up. */
1958         if (__intel_wait_for_register_fw(uncore,
1959                                          GEN6_BSD_SLEEP_PSMI_CONTROL,
1960                                          GEN6_BSD_SLEEP_INDICATOR,
1961                                          0,
1962                                          1000, 0, NULL))
1963                 DRM_ERROR("timed out waiting for the BSD ring to wake up\n");
1964
1965         /* Now that the ring is fully powered up, update the tail */
1966         i9xx_submit_request(request);
1967
1968         /* Let the ring send IDLE messages to the GT again,
1969          * and so let it sleep to conserve power when idle.
1970          */
1971         intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
1972                               _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1973
1974         intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
1975 }
1976
1977 static int mi_flush_dw(struct i915_request *rq, u32 flags)
1978 {
1979         u32 cmd, *cs;
1980
1981         cs = intel_ring_begin(rq, 4);
1982         if (IS_ERR(cs))
1983                 return PTR_ERR(cs);
1984
1985         cmd = MI_FLUSH_DW;
1986
1987         /*
1988          * We always require a command barrier so that subsequent
1989          * commands, such as breadcrumb interrupts, are strictly ordered
1990          * wrt the contents of the write cache being flushed to memory
1991          * (and thus being coherent from the CPU).
1992          */
1993         cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1994
1995         /*
1996          * Bspec vol 1c.3 - blitter engine command streamer:
1997          * "If ENABLED, all TLBs will be invalidated once the flush
1998          * operation is complete. This bit is only valid when the
1999          * Post-Sync Operation field is a value of 1h or 3h."
2000          */
2001         cmd |= flags;
2002
2003         *cs++ = cmd;
2004         *cs++ = I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT;
2005         *cs++ = 0;
2006         *cs++ = MI_NOOP;
2007
2008         intel_ring_advance(rq, cs);
2009
2010         return 0;
2011 }
2012
2013 static int gen6_flush_dw(struct i915_request *rq, u32 mode, u32 invflags)
2014 {
2015         return mi_flush_dw(rq, mode & EMIT_INVALIDATE ? invflags : 0);
2016 }
2017
2018 static int gen6_bsd_ring_flush(struct i915_request *rq, u32 mode)
2019 {
2020         return gen6_flush_dw(rq, mode, MI_INVALIDATE_TLB | MI_INVALIDATE_BSD);
2021 }
2022
2023 static int
2024 hsw_emit_bb_start(struct i915_request *rq,
2025                   u64 offset, u32 len,
2026                   unsigned int dispatch_flags)
2027 {
2028         u32 *cs;
2029
2030         cs = intel_ring_begin(rq, 2);
2031         if (IS_ERR(cs))
2032                 return PTR_ERR(cs);
2033
2034         *cs++ = MI_BATCH_BUFFER_START | (dispatch_flags & I915_DISPATCH_SECURE ?
2035                 0 : MI_BATCH_PPGTT_HSW | MI_BATCH_NON_SECURE_HSW);
2036         /* bit0-7 is the length on GEN6+ */
2037         *cs++ = offset;
2038         intel_ring_advance(rq, cs);
2039
2040         return 0;
2041 }
2042
2043 static int
2044 gen6_emit_bb_start(struct i915_request *rq,
2045                    u64 offset, u32 len,
2046                    unsigned int dispatch_flags)
2047 {
2048         u32 *cs;
2049
2050         cs = intel_ring_begin(rq, 2);
2051         if (IS_ERR(cs))
2052                 return PTR_ERR(cs);
2053
2054         *cs++ = MI_BATCH_BUFFER_START | (dispatch_flags & I915_DISPATCH_SECURE ?
2055                 0 : MI_BATCH_NON_SECURE_I965);
2056         /* bit0-7 is the length on GEN6+ */
2057         *cs++ = offset;
2058         intel_ring_advance(rq, cs);
2059
2060         return 0;
2061 }
2062
2063 /* Blitter support (SandyBridge+) */
2064
2065 static int gen6_ring_flush(struct i915_request *rq, u32 mode)
2066 {
2067         return gen6_flush_dw(rq, mode, MI_INVALIDATE_TLB);
2068 }
2069
2070 static void i9xx_set_default_submission(struct intel_engine_cs *engine)
2071 {
2072         engine->submit_request = i9xx_submit_request;
2073         engine->cancel_requests = cancel_requests;
2074
2075         engine->park = NULL;
2076         engine->unpark = NULL;
2077 }
2078
2079 static void gen6_bsd_set_default_submission(struct intel_engine_cs *engine)
2080 {
2081         i9xx_set_default_submission(engine);
2082         engine->submit_request = gen6_bsd_submit_request;
2083 }
2084
2085 static void ring_destroy(struct intel_engine_cs *engine)
2086 {
2087         struct drm_i915_private *dev_priv = engine->i915;
2088
2089         WARN_ON(INTEL_GEN(dev_priv) > 2 &&
2090                 (ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
2091
2092         intel_engine_cleanup_common(engine);
2093
2094         intel_ring_unpin(engine->buffer);
2095         intel_ring_put(engine->buffer);
2096
2097         kfree(engine);
2098 }
2099
2100 static void setup_irq(struct intel_engine_cs *engine)
2101 {
2102         struct drm_i915_private *i915 = engine->i915;
2103
2104         if (INTEL_GEN(i915) >= 6) {
2105                 engine->irq_enable = gen6_irq_enable;
2106                 engine->irq_disable = gen6_irq_disable;
2107         } else if (INTEL_GEN(i915) >= 5) {
2108                 engine->irq_enable = gen5_irq_enable;
2109                 engine->irq_disable = gen5_irq_disable;
2110         } else if (INTEL_GEN(i915) >= 3) {
2111                 engine->irq_enable = i9xx_irq_enable;
2112                 engine->irq_disable = i9xx_irq_disable;
2113         } else {
2114                 engine->irq_enable = i8xx_irq_enable;
2115                 engine->irq_disable = i8xx_irq_disable;
2116         }
2117 }
2118
2119 static void setup_common(struct intel_engine_cs *engine)
2120 {
2121         struct drm_i915_private *i915 = engine->i915;
2122
2123         /* gen8+ are only supported with execlists */
2124         GEM_BUG_ON(INTEL_GEN(i915) >= 8);
2125
2126         setup_irq(engine);
2127
2128         engine->destroy = ring_destroy;
2129
2130         engine->resume = xcs_resume;
2131         engine->reset.prepare = reset_prepare;
2132         engine->reset.reset = reset_ring;
2133         engine->reset.finish = reset_finish;
2134
2135         engine->cops = &ring_context_ops;
2136         engine->request_alloc = ring_request_alloc;
2137
2138         /*
2139          * Using a global execution timeline; the previous final breadcrumb is
2140          * equivalent to our next initial bread so we can elide
2141          * engine->emit_init_breadcrumb().
2142          */
2143         engine->emit_fini_breadcrumb = i9xx_emit_breadcrumb;
2144         if (IS_GEN(i915, 5))
2145                 engine->emit_fini_breadcrumb = gen5_emit_breadcrumb;
2146
2147         engine->set_default_submission = i9xx_set_default_submission;
2148
2149         if (INTEL_GEN(i915) >= 6)
2150                 engine->emit_bb_start = gen6_emit_bb_start;
2151         else if (INTEL_GEN(i915) >= 4)
2152                 engine->emit_bb_start = i965_emit_bb_start;
2153         else if (IS_I830(i915) || IS_I845G(i915))
2154                 engine->emit_bb_start = i830_emit_bb_start;
2155         else
2156                 engine->emit_bb_start = i915_emit_bb_start;
2157 }
2158
2159 static void setup_rcs(struct intel_engine_cs *engine)
2160 {
2161         struct drm_i915_private *i915 = engine->i915;
2162
2163         if (HAS_L3_DPF(i915))
2164                 engine->irq_keep_mask = GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
2165
2166         engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
2167
2168         if (INTEL_GEN(i915) >= 7) {
2169                 engine->init_context = intel_rcs_ctx_init;
2170                 engine->emit_flush = gen7_render_ring_flush;
2171                 engine->emit_fini_breadcrumb = gen7_rcs_emit_breadcrumb;
2172         } else if (IS_GEN(i915, 6)) {
2173                 engine->init_context = intel_rcs_ctx_init;
2174                 engine->emit_flush = gen6_render_ring_flush;
2175                 engine->emit_fini_breadcrumb = gen6_rcs_emit_breadcrumb;
2176         } else if (IS_GEN(i915, 5)) {
2177                 engine->emit_flush = gen4_render_ring_flush;
2178         } else {
2179                 if (INTEL_GEN(i915) < 4)
2180                         engine->emit_flush = gen2_render_ring_flush;
2181                 else
2182                         engine->emit_flush = gen4_render_ring_flush;
2183                 engine->irq_enable_mask = I915_USER_INTERRUPT;
2184         }
2185
2186         if (IS_HASWELL(i915))
2187                 engine->emit_bb_start = hsw_emit_bb_start;
2188
2189         engine->resume = rcs_resume;
2190 }
2191
2192 static void setup_vcs(struct intel_engine_cs *engine)
2193 {
2194         struct drm_i915_private *i915 = engine->i915;
2195
2196         if (INTEL_GEN(i915) >= 6) {
2197                 /* gen6 bsd needs a special wa for tail updates */
2198                 if (IS_GEN(i915, 6))
2199                         engine->set_default_submission = gen6_bsd_set_default_submission;
2200                 engine->emit_flush = gen6_bsd_ring_flush;
2201                 engine->irq_enable_mask = GT_BSD_USER_INTERRUPT;
2202
2203                 if (IS_GEN(i915, 6))
2204                         engine->emit_fini_breadcrumb = gen6_xcs_emit_breadcrumb;
2205                 else
2206                         engine->emit_fini_breadcrumb = gen7_xcs_emit_breadcrumb;
2207         } else {
2208                 engine->emit_flush = bsd_ring_flush;
2209                 if (IS_GEN(i915, 5))
2210                         engine->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
2211                 else
2212                         engine->irq_enable_mask = I915_BSD_USER_INTERRUPT;
2213         }
2214 }
2215
2216 static void setup_bcs(struct intel_engine_cs *engine)
2217 {
2218         struct drm_i915_private *i915 = engine->i915;
2219
2220         engine->emit_flush = gen6_ring_flush;
2221         engine->irq_enable_mask = GT_BLT_USER_INTERRUPT;
2222
2223         if (IS_GEN(i915, 6))
2224                 engine->emit_fini_breadcrumb = gen6_xcs_emit_breadcrumb;
2225         else
2226                 engine->emit_fini_breadcrumb = gen7_xcs_emit_breadcrumb;
2227 }
2228
2229 static void setup_vecs(struct intel_engine_cs *engine)
2230 {
2231         struct drm_i915_private *i915 = engine->i915;
2232
2233         GEM_BUG_ON(INTEL_GEN(i915) < 7);
2234
2235         engine->emit_flush = gen6_ring_flush;
2236         engine->irq_enable_mask = PM_VEBOX_USER_INTERRUPT;
2237         engine->irq_enable = hsw_vebox_irq_enable;
2238         engine->irq_disable = hsw_vebox_irq_disable;
2239
2240         engine->emit_fini_breadcrumb = gen7_xcs_emit_breadcrumb;
2241 }
2242
2243 int intel_ring_submission_setup(struct intel_engine_cs *engine)
2244 {
2245         setup_common(engine);
2246
2247         switch (engine->class) {
2248         case RENDER_CLASS:
2249                 setup_rcs(engine);
2250                 break;
2251         case VIDEO_DECODE_CLASS:
2252                 setup_vcs(engine);
2253                 break;
2254         case COPY_ENGINE_CLASS:
2255                 setup_bcs(engine);
2256                 break;
2257         case VIDEO_ENHANCEMENT_CLASS:
2258                 setup_vecs(engine);
2259                 break;
2260         default:
2261                 MISSING_CASE(engine->class);
2262                 return -ENODEV;
2263         }
2264
2265         return 0;
2266 }
2267
2268 int intel_ring_submission_init(struct intel_engine_cs *engine)
2269 {
2270         struct i915_timeline *timeline;
2271         struct intel_ring *ring;
2272         int err;
2273
2274         timeline = i915_timeline_create(engine->i915, engine->status_page.vma);
2275         if (IS_ERR(timeline)) {
2276                 err = PTR_ERR(timeline);
2277                 goto err;
2278         }
2279         GEM_BUG_ON(timeline->has_initial_breadcrumb);
2280
2281         ring = intel_engine_create_ring(engine, timeline, 32 * PAGE_SIZE);
2282         i915_timeline_put(timeline);
2283         if (IS_ERR(ring)) {
2284                 err = PTR_ERR(ring);
2285                 goto err;
2286         }
2287
2288         err = intel_ring_pin(ring);
2289         if (err)
2290                 goto err_ring;
2291
2292         GEM_BUG_ON(engine->buffer);
2293         engine->buffer = ring;
2294
2295         err = intel_engine_init_common(engine);
2296         if (err)
2297                 goto err_unpin;
2298
2299         GEM_BUG_ON(ring->timeline->hwsp_ggtt != engine->status_page.vma);
2300
2301         return 0;
2302
2303 err_unpin:
2304         intel_ring_unpin(ring);
2305 err_ring:
2306         intel_ring_put(ring);
2307 err:
2308         intel_engine_cleanup_common(engine);
2309         return err;
2310 }