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
  32 #include <drm/i915_drm.h>
  33
  34 #include "i915_drv.h"
  35 #include "i915_gem_render_state.h"
  36 #include "i915_reset.h"
  37 #include "i915_trace.h"
  38 #include "intel_drv.h"
  39 #include "intel_workarounds.h"
  40
  41 /* Rough estimate of the typical request size, performing a flush,
  42  * set-context and then emitting the batch.
  43  */
  44 #define LEGACY_REQUEST_SIZE 200
  45
  46 unsigned int intel_ring_update_space(struct intel_ring *ring)
  47 {
  48         unsigned int space;
  49
  50         space = __intel_ring_space(ring->head, ring->emit, ring->size);
  51
  52         ring->space = space;
  53         return space;
  54 }
  55
  56 static int
  57 gen2_render_ring_flush(struct i915_request *rq, u32 mode)
  58 {
  59         unsigned int num_store_dw;
  60         u32 cmd, *cs;
  61
  62         cmd = MI_FLUSH;
  63         num_store_dw = 0;
  64         if (mode & EMIT_INVALIDATE)
  65                 cmd |= MI_READ_FLUSH;
  66         if (mode & EMIT_FLUSH)
  67                 num_store_dw = 4;
  68
  69         cs = intel_ring_begin(rq, 2 + 3 * num_store_dw);
  70         if (IS_ERR(cs))
  71                 return PTR_ERR(cs);
  72
  73         *cs++ = cmd;
  74         while (num_store_dw--) {
  75                 *cs++ = MI_STORE_DWORD_IMM | MI_MEM_VIRTUAL;
  76                 *cs++ = i915_scratch_offset(rq->i915);
  77                 *cs++ = 0;
  78         }
  79         *cs++ = MI_FLUSH | MI_NO_WRITE_FLUSH;
  80
  81         intel_ring_advance(rq, cs);
  82
  83         return 0;
  84 }
  85
  86 static int
  87 gen4_render_ring_flush(struct i915_request *rq, u32 mode)
  88 {
  89         u32 cmd, *cs;
  90         int i;
  91
  92         /*
  93          * read/write caches:
  94          *
  95          * I915_GEM_DOMAIN_RENDER is always invalidated, but is
  96          * only flushed if MI_NO_WRITE_FLUSH is unset.  On 965, it is
  97          * also flushed at 2d versus 3d pipeline switches.
  98          *
  99          * read-only caches:
 100          *
 101          * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
 102          * MI_READ_FLUSH is set, and is always flushed on 965.
 103          *
 104          * I915_GEM_DOMAIN_COMMAND may not exist?
 105          *
 106          * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
 107          * invalidated when MI_EXE_FLUSH is set.
 108          *
 109          * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
 110          * invalidated with every MI_FLUSH.
 111          *
 112          * TLBs:
 113          *
 114          * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
 115          * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
 116          * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
 117          * are flushed at any MI_FLUSH.
 118          */
 119
 120         cmd = MI_FLUSH;
 121         if (mode & EMIT_INVALIDATE) {
 122                 cmd |= MI_EXE_FLUSH;
 123                 if (IS_G4X(rq->i915) || IS_GEN(rq->i915, 5))
 124                         cmd |= MI_INVALIDATE_ISP;
 125         }
 126
 127         i = 2;
 128         if (mode & EMIT_INVALIDATE)
 129                 i += 20;
 130
 131         cs = intel_ring_begin(rq, i);
 132         if (IS_ERR(cs))
 133                 return PTR_ERR(cs);
 134
 135         *cs++ = cmd;
 136
 137         /*
 138          * A random delay to let the CS invalidate take effect? Without this
 139          * delay, the GPU relocation path fails as the CS does not see
 140          * the updated contents. Just as important, if we apply the flushes
 141          * to the EMIT_FLUSH branch (i.e. immediately after the relocation
 142          * write and before the invalidate on the next batch), the relocations
 143          * still fail. This implies that is a delay following invalidation
 144          * that is required to reset the caches as opposed to a delay to
 145          * ensure the memory is written.
 146          */
 147         if (mode & EMIT_INVALIDATE) {
 148                 *cs++ = GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE;
 149                 *cs++ = i915_scratch_offset(rq->i915) | PIPE_CONTROL_GLOBAL_GTT;
 150                 *cs++ = 0;
 151                 *cs++ = 0;
 152
 153                 for (i = 0; i < 12; i++)
 154                         *cs++ = MI_FLUSH;
 155
 156                 *cs++ = GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE;
 157                 *cs++ = i915_scratch_offset(rq->i915) | PIPE_CONTROL_GLOBAL_GTT;
 158                 *cs++ = 0;
 159                 *cs++ = 0;
 160         }
 161
 162         *cs++ = cmd;
 163
 164         intel_ring_advance(rq, cs);
 165
 166         return 0;
 167 }
 168
 169 /*
 170  * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
 171  * implementing two workarounds on gen6.  From section 1.4.7.1
 172  * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
 173  *
 174  * [DevSNB-C+{W/A}] Before any depth stall flush (including those
 175  * produced by non-pipelined state commands), software needs to first
 176  * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
 177  * 0.
 178  *
 179  * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
 180  * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
 181  *
 182  * And the workaround for these two requires this workaround first:
 183  *
 184  * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
 185  * BEFORE the pipe-control with a post-sync op and no write-cache
 186  * flushes.
 187  *
 188  * And this last workaround is tricky because of the requirements on
 189  * that bit.  From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
 190  * volume 2 part 1:
 191  *
 192  *     "1 of the following must also be set:
 193  *      - Render Target Cache Flush Enable ([12] of DW1)
 194  *      - Depth Cache Flush Enable ([0] of DW1)
 195  *      - Stall at Pixel Scoreboard ([1] of DW1)
 196  *      - Depth Stall ([13] of DW1)
 197  *      - Post-Sync Operation ([13] of DW1)
 198  *      - Notify Enable ([8] of DW1)"
 199  *
 200  * The cache flushes require the workaround flush that triggered this
 201  * one, so we can't use it.  Depth stall would trigger the same.
 202  * Post-sync nonzero is what triggered this second workaround, so we
 203  * can't use that one either.  Notify enable is IRQs, which aren't
 204  * really our business.  That leaves only stall at scoreboard.
 205  */
 206 static int
 207 gen6_emit_post_sync_nonzero_flush(struct i915_request *rq)
 208 {
 209         u32 scratch_addr = i915_scratch_offset(rq->i915) + 2 * CACHELINE_BYTES;
 210         u32 *cs;
 211
 212         cs = intel_ring_begin(rq, 6);
 213         if (IS_ERR(cs))
 214                 return PTR_ERR(cs);
 215
 216         *cs++ = GFX_OP_PIPE_CONTROL(5);
 217         *cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
 218         *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
 219         *cs++ = 0; /* low dword */
 220         *cs++ = 0; /* high dword */
 221         *cs++ = MI_NOOP;
 222         intel_ring_advance(rq, cs);
 223
 224         cs = intel_ring_begin(rq, 6);
 225         if (IS_ERR(cs))
 226                 return PTR_ERR(cs);
 227
 228         *cs++ = GFX_OP_PIPE_CONTROL(5);
 229         *cs++ = PIPE_CONTROL_QW_WRITE;
 230         *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
 231         *cs++ = 0;
 232         *cs++ = 0;
 233         *cs++ = MI_NOOP;
 234         intel_ring_advance(rq, cs);
 235
 236         return 0;
 237 }
 238
 239 static int
 240 gen6_render_ring_flush(struct i915_request *rq, u32 mode)
 241 {
 242         u32 scratch_addr = i915_scratch_offset(rq->i915) + 2 * CACHELINE_BYTES;
 243         u32 *cs, flags = 0;
 244         int ret;
 245
 246         /* Force SNB workarounds for PIPE_CONTROL flushes */
 247         ret = gen6_emit_post_sync_nonzero_flush(rq);
 248         if (ret)
 249                 return ret;
 250
 251         /* Just flush everything.  Experiments have shown that reducing the
 252          * number of bits based on the write domains has little performance
 253          * impact.
 254          */
 255         if (mode & EMIT_FLUSH) {
 256                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
 257                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
 258                 /*
 259                  * Ensure that any following seqno writes only happen
 260                  * when the render cache is indeed flushed.
 261                  */
 262                 flags |= PIPE_CONTROL_CS_STALL;
 263         }
 264         if (mode & EMIT_INVALIDATE) {
 265                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
 266                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 267                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 268                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
 269                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
 270                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
 271                 /*
 272                  * TLB invalidate requires a post-sync write.
 273                  */
 274                 flags |= PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_CS_STALL;
 275         }
 276
 277         cs = intel_ring_begin(rq, 4);
 278         if (IS_ERR(cs))
 279                 return PTR_ERR(cs);
 280
 281         *cs++ = GFX_OP_PIPE_CONTROL(4);
 282         *cs++ = flags;
 283         *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
 284         *cs++ = 0;
 285         intel_ring_advance(rq, cs);
 286
 287         return 0;
 288 }
 289
 290 static u32 *gen6_rcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
 291 {
 292         /* First we do the gen6_emit_post_sync_nonzero_flush w/a */
 293         *cs++ = GFX_OP_PIPE_CONTROL(4);
 294         *cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
 295         *cs++ = 0;
 296         *cs++ = 0;
 297
 298         *cs++ = GFX_OP_PIPE_CONTROL(4);
 299         *cs++ = PIPE_CONTROL_QW_WRITE;
 300         *cs++ = i915_scratch_offset(rq->i915) | PIPE_CONTROL_GLOBAL_GTT;
 301         *cs++ = 0;
 302
 303         /* Finally we can flush and with it emit the breadcrumb */
 304         *cs++ = GFX_OP_PIPE_CONTROL(4);
 305         *cs++ = (PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
 306                  PIPE_CONTROL_DEPTH_CACHE_FLUSH |
 307                  PIPE_CONTROL_DC_FLUSH_ENABLE |
 308                  PIPE_CONTROL_QW_WRITE |
 309                  PIPE_CONTROL_CS_STALL);
 310         *cs++ = rq->timeline->hwsp_offset | PIPE_CONTROL_GLOBAL_GTT;
 311         *cs++ = rq->fence.seqno;
 312
 313         *cs++ = GFX_OP_PIPE_CONTROL(4);
 314         *cs++ = PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_STORE_DATA_INDEX;
 315         *cs++ = I915_GEM_HWS_HANGCHECK_ADDR | PIPE_CONTROL_GLOBAL_GTT;
 316         *cs++ = intel_engine_next_hangcheck_seqno(rq->engine);
 317
 318         *cs++ = MI_USER_INTERRUPT;
 319         *cs++ = MI_NOOP;
 320
 321         rq->tail = intel_ring_offset(rq, cs);
 322         assert_ring_tail_valid(rq->ring, rq->tail);
 323
 324         return cs;
 325 }
 326
 327 static int
 328 gen7_render_ring_cs_stall_wa(struct i915_request *rq)
 329 {
 330         u32 *cs;
 331
 332         cs = intel_ring_begin(rq, 4);
 333         if (IS_ERR(cs))
 334                 return PTR_ERR(cs);
 335
 336         *cs++ = GFX_OP_PIPE_CONTROL(4);
 337         *cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
 338         *cs++ = 0;
 339         *cs++ = 0;
 340         intel_ring_advance(rq, cs);
 341
 342         return 0;
 343 }
 344
 345 static int
 346 gen7_render_ring_flush(struct i915_request *rq, u32 mode)
 347 {
 348         u32 scratch_addr = i915_scratch_offset(rq->i915) + 2 * CACHELINE_BYTES;
 349         u32 *cs, flags = 0;
 350
 351         /*
 352          * Ensure that any following seqno writes only happen when the render
 353          * cache is indeed flushed.
 354          *
 355          * Workaround: 4th PIPE_CONTROL command (except the ones with only
 356          * read-cache invalidate bits set) must have the CS_STALL bit set. We
 357          * don't try to be clever and just set it unconditionally.
 358          */
 359         flags |= PIPE_CONTROL_CS_STALL;
 360
 361         /* Just flush everything.  Experiments have shown that reducing the
 362          * number of bits based on the write domains has little performance
 363          * impact.
 364          */
 365         if (mode & EMIT_FLUSH) {
 366                 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
 367                 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
 368                 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
 369                 flags |= PIPE_CONTROL_FLUSH_ENABLE;
 370         }
 371         if (mode & EMIT_INVALIDATE) {
 372                 flags |= PIPE_CONTROL_TLB_INVALIDATE;
 373                 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
 374                 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
 375                 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
 376                 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
 377                 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
 378                 flags |= PIPE_CONTROL_MEDIA_STATE_CLEAR;
 379                 /*
 380                  * TLB invalidate requires a post-sync write.
 381                  */
 382                 flags |= PIPE_CONTROL_QW_WRITE;
 383                 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
 384
 385                 flags |= PIPE_CONTROL_STALL_AT_SCOREBOARD;
 386
 387                 /* Workaround: we must issue a pipe_control with CS-stall bit
 388                  * set before a pipe_control command that has the state cache
 389                  * invalidate bit set. */
 390                 gen7_render_ring_cs_stall_wa(rq);
 391         }
 392
 393         cs = intel_ring_begin(rq, 4);
 394         if (IS_ERR(cs))
 395                 return PTR_ERR(cs);
 396
 397         *cs++ = GFX_OP_PIPE_CONTROL(4);
 398         *cs++ = flags;
 399         *cs++ = scratch_addr;
 400         *cs++ = 0;
 401         intel_ring_advance(rq, cs);
 402
 403         return 0;
 404 }
 405
 406 static u32 *gen7_rcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
 407 {
 408         *cs++ = GFX_OP_PIPE_CONTROL(4);
 409         *cs++ = (PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
 410                  PIPE_CONTROL_DEPTH_CACHE_FLUSH |
 411                  PIPE_CONTROL_DC_FLUSH_ENABLE |
 412                  PIPE_CONTROL_FLUSH_ENABLE |
 413                  PIPE_CONTROL_QW_WRITE |
 414                  PIPE_CONTROL_GLOBAL_GTT_IVB |
 415                  PIPE_CONTROL_CS_STALL);
 416         *cs++ = rq->timeline->hwsp_offset;
 417         *cs++ = rq->fence.seqno;
 418
 419         *cs++ = GFX_OP_PIPE_CONTROL(4);
 420         *cs++ = (PIPE_CONTROL_QW_WRITE |
 421                  PIPE_CONTROL_STORE_DATA_INDEX |
 422                  PIPE_CONTROL_GLOBAL_GTT_IVB);
 423         *cs++ = I915_GEM_HWS_HANGCHECK_ADDR;
 424         *cs++ = intel_engine_next_hangcheck_seqno(rq->engine);
 425
 426         *cs++ = MI_USER_INTERRUPT;
 427         *cs++ = MI_NOOP;
 428
 429         rq->tail = intel_ring_offset(rq, cs);
 430         assert_ring_tail_valid(rq->ring, rq->tail);
 431
 432         return cs;
 433 }
 434
 435 static u32 *gen6_xcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
 436 {
 437         GEM_BUG_ON(rq->timeline->hwsp_ggtt != rq->engine->status_page.vma);
 438         GEM_BUG_ON(offset_in_page(rq->timeline->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
 439
 440         *cs++ = MI_FLUSH_DW | MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_DW_STORE_INDEX;
 441         *cs++ = I915_GEM_HWS_SEQNO_ADDR | MI_FLUSH_DW_USE_GTT;
 442         *cs++ = rq->fence.seqno;
 443
 444         *cs++ = MI_FLUSH_DW | MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_DW_STORE_INDEX;
 445         *cs++ = I915_GEM_HWS_HANGCHECK_ADDR | MI_FLUSH_DW_USE_GTT;
 446         *cs++ = intel_engine_next_hangcheck_seqno(rq->engine);
 447
 448         *cs++ = MI_USER_INTERRUPT;
 449         *cs++ = MI_NOOP;
 450
 451         rq->tail = intel_ring_offset(rq, cs);
 452         assert_ring_tail_valid(rq->ring, rq->tail);
 453
 454         return cs;
 455 }
 456
 457 #define GEN7_XCS_WA 32
 458 static u32 *gen7_xcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
 459 {
 460         int i;
 461
 462         GEM_BUG_ON(rq->timeline->hwsp_ggtt != rq->engine->status_page.vma);
 463         GEM_BUG_ON(offset_in_page(rq->timeline->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
 464
 465         *cs++ = MI_FLUSH_DW | MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_DW_STORE_INDEX;
 466         *cs++ = I915_GEM_HWS_SEQNO_ADDR | MI_FLUSH_DW_USE_GTT;
 467         *cs++ = rq->fence.seqno;
 468
 469         *cs++ = MI_FLUSH_DW | MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_DW_STORE_INDEX;
 470         *cs++ = I915_GEM_HWS_HANGCHECK_ADDR | MI_FLUSH_DW_USE_GTT;
 471         *cs++ = intel_engine_next_hangcheck_seqno(rq->engine);
 472
 473         for (i = 0; i < GEN7_XCS_WA; i++) {
 474                 *cs++ = MI_STORE_DWORD_INDEX;
 475                 *cs++ = I915_GEM_HWS_SEQNO_ADDR;
 476                 *cs++ = rq->fence.seqno;
 477         }
 478
 479         *cs++ = MI_FLUSH_DW;
 480         *cs++ = 0;
 481         *cs++ = 0;
 482
 483         *cs++ = MI_USER_INTERRUPT;
 484
 485         rq->tail = intel_ring_offset(rq, cs);
 486         assert_ring_tail_valid(rq->ring, rq->tail);
 487
 488         return cs;
 489 }
 490 #undef GEN7_XCS_WA
 491
 492 static void set_hwstam(struct intel_engine_cs *engine, u32 mask)
 493 {
 494         /*
 495          * Keep the render interrupt unmasked as this papers over
 496          * lost interrupts following a reset.
 497          */
 498         if (engine->class == RENDER_CLASS) {
 499                 if (INTEL_GEN(engine->i915) >= 6)
 500                         mask &= ~BIT(0);
 501                 else
 502                         mask &= ~I915_USER_INTERRUPT;
 503         }
 504
 505         intel_engine_set_hwsp_writemask(engine, mask);
 506 }
 507
 508 static void set_hws_pga(struct intel_engine_cs *engine, phys_addr_t phys)
 509 {
 510         struct drm_i915_private *dev_priv = engine->i915;
 511         u32 addr;
 512
 513         addr = lower_32_bits(phys);
 514         if (INTEL_GEN(dev_priv) >= 4)
 515                 addr |= (phys >> 28) & 0xf0;
 516
 517         I915_WRITE(HWS_PGA, addr);
 518 }
 519
 520 static struct page *status_page(struct intel_engine_cs *engine)
 521 {
 522         struct drm_i915_gem_object *obj = engine->status_page.vma->obj;
 523
 524         GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
 525         return sg_page(obj->mm.pages->sgl);
 526 }
 527
 528 static void ring_setup_phys_status_page(struct intel_engine_cs *engine)
 529 {
 530         set_hws_pga(engine, PFN_PHYS(page_to_pfn(status_page(engine))));
 531         set_hwstam(engine, ~0u);
 532 }
 533
 534 static void set_hwsp(struct intel_engine_cs *engine, u32 offset)
 535 {
 536         struct drm_i915_private *dev_priv = engine->i915;
 537         i915_reg_t hwsp;
 538
 539         /*
 540          * The ring status page addresses are no longer next to the rest of
 541          * the ring registers as of gen7.
 542          */
 543         if (IS_GEN(dev_priv, 7)) {
 544                 switch (engine->id) {
 545                 /*
 546                  * No more rings exist on Gen7. Default case is only to shut up
 547                  * gcc switch check warning.
 548                  */
 549                 default:
 550                         GEM_BUG_ON(engine->id);
 551                         /* fallthrough */
 552                 case RCS0:
 553                         hwsp = RENDER_HWS_PGA_GEN7;
 554                         break;
 555                 case BCS0:
 556                         hwsp = BLT_HWS_PGA_GEN7;
 557                         break;
 558                 case VCS0:
 559                         hwsp = BSD_HWS_PGA_GEN7;
 560                         break;
 561                 case VECS0:
 562                         hwsp = VEBOX_HWS_PGA_GEN7;
 563                         break;
 564                 }
 565         } else if (IS_GEN(dev_priv, 6)) {
 566                 hwsp = RING_HWS_PGA_GEN6(engine->mmio_base);
 567         } else {
 568                 hwsp = RING_HWS_PGA(engine->mmio_base);
 569         }
 570
 571         I915_WRITE(hwsp, offset);
 572         POSTING_READ(hwsp);
 573 }
 574
 575 static void flush_cs_tlb(struct intel_engine_cs *engine)
 576 {
 577         struct drm_i915_private *dev_priv = engine->i915;
 578
 579         if (!IS_GEN_RANGE(dev_priv, 6, 7))
 580                 return;
 581
 582         /* ring should be idle before issuing a sync flush*/
 583         WARN_ON((ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
 584
 585         ENGINE_WRITE(engine, RING_INSTPM,
 586                      _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE |
 587                                         INSTPM_SYNC_FLUSH));
 588         if (intel_wait_for_register(engine->uncore,
 589                                     RING_INSTPM(engine->mmio_base),
 590                                     INSTPM_SYNC_FLUSH, 0,
 591                                     1000))
 592                 DRM_ERROR("%s: wait for SyncFlush to complete for TLB invalidation timed out\n",
 593                           engine->name);
 594 }
 595
 596 static void ring_setup_status_page(struct intel_engine_cs *engine)
 597 {
 598         set_hwsp(engine, i915_ggtt_offset(engine->status_page.vma));
 599         set_hwstam(engine, ~0u);
 600
 601         flush_cs_tlb(engine);
 602 }
 603
 604 static bool stop_ring(struct intel_engine_cs *engine)
 605 {
 606         struct drm_i915_private *dev_priv = engine->i915;
 607
 608         if (INTEL_GEN(dev_priv) > 2) {
 609                 ENGINE_WRITE(engine,
 610                              RING_MI_MODE, _MASKED_BIT_ENABLE(STOP_RING));
 611                 if (intel_wait_for_register(engine->uncore,
 612                                             RING_MI_MODE(engine->mmio_base),
 613                                             MODE_IDLE,
 614                                             MODE_IDLE,
 615                                             1000)) {
 616                         DRM_ERROR("%s : timed out trying to stop ring\n",
 617                                   engine->name);
 618
 619                         /*
 620                          * Sometimes we observe that the idle flag is not
 621                          * set even though the ring is empty. So double
 622                          * check before giving up.
 623                          */
 624                         if (ENGINE_READ(engine, RING_HEAD) !=
 625                             ENGINE_READ(engine, RING_TAIL))
 626                                 return false;
 627                 }
 628         }
 629
 630         ENGINE_WRITE(engine, RING_HEAD, ENGINE_READ(engine, RING_TAIL));
 631
 632         ENGINE_WRITE(engine, RING_HEAD, 0);
 633         ENGINE_WRITE(engine, RING_TAIL, 0);
 634
 635         /* The ring must be empty before it is disabled */
 636         ENGINE_WRITE(engine, RING_CTL, 0);
 637
 638         return (ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) == 0;
 639 }
 640
 641 static int init_ring_common(struct intel_engine_cs *engine)
 642 {
 643         struct drm_i915_private *dev_priv = engine->i915;
 644         struct intel_ring *ring = engine->buffer;
 645         int ret = 0;
 646
 647         intel_uncore_forcewake_get(engine->uncore, FORCEWAKE_ALL);
 648
 649         if (!stop_ring(engine)) {
 650                 /* G45 ring initialization often fails to reset head to zero */
 651                 DRM_DEBUG_DRIVER("%s head not reset to zero "
 652                                 "ctl %08x head %08x tail %08x start %08x\n",
 653                                 engine->name,
 654                                 ENGINE_READ(engine, RING_CTL),
 655                                 ENGINE_READ(engine, RING_HEAD),
 656                                 ENGINE_READ(engine, RING_TAIL),
 657                                 ENGINE_READ(engine, RING_START));
 658
 659                 if (!stop_ring(engine)) {
 660                         DRM_ERROR("failed to set %s head to zero "
 661                                   "ctl %08x head %08x tail %08x start %08x\n",
 662                                   engine->name,
 663                                   ENGINE_READ(engine, RING_CTL),
 664                                   ENGINE_READ(engine, RING_HEAD),
 665                                   ENGINE_READ(engine, RING_TAIL),
 666                                   ENGINE_READ(engine, RING_START));
 667                         ret = -EIO;
 668                         goto out;
 669                 }
 670         }
 671
 672         if (HWS_NEEDS_PHYSICAL(dev_priv))
 673                 ring_setup_phys_status_page(engine);
 674         else
 675                 ring_setup_status_page(engine);
 676
 677         intel_engine_reset_breadcrumbs(engine);
 678
 679         /* Enforce ordering by reading HEAD register back */
 680         ENGINE_READ(engine, RING_HEAD);
 681
 682         /* Initialize the ring. This must happen _after_ we've cleared the ring
 683          * registers with the above sequence (the readback of the HEAD registers
 684          * also enforces ordering), otherwise the hw might lose the new ring
 685          * register values. */
 686         ENGINE_WRITE(engine, RING_START, i915_ggtt_offset(ring->vma));
 687
 688         /* WaClearRingBufHeadRegAtInit:ctg,elk */
 689         if (ENGINE_READ(engine, RING_HEAD))
 690                 DRM_DEBUG_DRIVER("%s initialization failed [head=%08x], fudging\n",
 691                                  engine->name, ENGINE_READ(engine, RING_HEAD));
 692
 693         /* Check that the ring offsets point within the ring! */
 694         GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->head));
 695         GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->tail));
 696         intel_ring_update_space(ring);
 697
 698         /* First wake the ring up to an empty/idle ring */
 699         ENGINE_WRITE(engine, RING_HEAD, ring->head);
 700         ENGINE_WRITE(engine, RING_TAIL, ring->head);
 701         ENGINE_POSTING_READ(engine, RING_TAIL);
 702
 703         ENGINE_WRITE(engine, RING_CTL, RING_CTL_SIZE(ring->size) | RING_VALID);
 704
 705         /* If the head is still not zero, the ring is dead */
 706         if (intel_wait_for_register(engine->uncore,
 707                                     RING_CTL(engine->mmio_base),
 708                                     RING_VALID, RING_VALID,
 709                                     50)) {
 710                 DRM_ERROR("%s initialization failed "
 711                           "ctl %08x (valid? %d) head %08x [%08x] tail %08x [%08x] start %08x [expected %08x]\n",
 712                           engine->name,
 713                           ENGINE_READ(engine, RING_CTL),
 714                           ENGINE_READ(engine, RING_CTL) & RING_VALID,
 715                           ENGINE_READ(engine, RING_HEAD), ring->head,
 716                           ENGINE_READ(engine, RING_TAIL), ring->tail,
 717                           ENGINE_READ(engine, RING_START),
 718                           i915_ggtt_offset(ring->vma));
 719                 ret = -EIO;
 720                 goto out;
 721         }
 722
 723         if (INTEL_GEN(dev_priv) > 2)
 724                 ENGINE_WRITE(engine,
 725                              RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
 726
 727         /* Now awake, let it get started */
 728         if (ring->tail != ring->head) {
 729                 ENGINE_WRITE(engine, RING_TAIL, ring->tail);
 730                 ENGINE_POSTING_READ(engine, RING_TAIL);
 731         }
 732
 733         /* Papering over lost _interrupts_ immediately following the restart */
 734         intel_engine_queue_breadcrumbs(engine);
 735 out:
 736         intel_uncore_forcewake_put(engine->uncore, FORCEWAKE_ALL);
 737
 738         return ret;
 739 }
 740
 741 static void reset_prepare(struct intel_engine_cs *engine)
 742 {
 743         intel_engine_stop_cs(engine);
 744 }
 745
 746 static void reset_ring(struct intel_engine_cs *engine, bool stalled)
 747 {
 748         struct i915_timeline *tl = &engine->timeline;
 749         struct i915_request *pos, *rq;
 750         unsigned long flags;
 751         u32 head;
 752
 753         rq = NULL;
 754         spin_lock_irqsave(&tl->lock, flags);
 755         list_for_each_entry(pos, &tl->requests, link) {
 756                 if (!i915_request_completed(pos)) {
 757                         rq = pos;
 758                         break;
 759                 }
 760         }
 761
 762         /*
 763          * The guilty request will get skipped on a hung engine.
 764          *
 765          * Users of client default contexts do not rely on logical
 766          * state preserved between batches so it is safe to execute
 767          * queued requests following the hang. Non default contexts
 768          * rely on preserved state, so skipping a batch loses the
 769          * evolution of the state and it needs to be considered corrupted.
 770          * Executing more queued batches on top of corrupted state is
 771          * risky. But we take the risk by trying to advance through
 772          * the queued requests in order to make the client behaviour
 773          * more predictable around resets, by not throwing away random
 774          * amount of batches it has prepared for execution. Sophisticated
 775          * clients can use gem_reset_stats_ioctl and dma fence status
 776          * (exported via sync_file info ioctl on explicit fences) to observe
 777          * when it loses the context state and should rebuild accordingly.
 778          *
 779          * The context ban, and ultimately the client ban, mechanism are safety
 780          * valves if client submission ends up resulting in nothing more than
 781          * subsequent hangs.
 782          */
 783
 784         if (rq) {
 785                 /*
 786                  * Try to restore the logical GPU state to match the
 787                  * continuation of the request queue. If we skip the
 788                  * context/PD restore, then the next request may try to execute
 789                  * assuming that its context is valid and loaded on the GPU and
 790                  * so may try to access invalid memory, prompting repeated GPU
 791                  * hangs.
 792                  *
 793                  * If the request was guilty, we still restore the logical
 794                  * state in case the next request requires it (e.g. the
 795                  * aliasing ppgtt), but skip over the hung batch.
 796                  *
 797                  * If the request was innocent, we try to replay the request
 798                  * with the restored context.
 799                  */
 800                 i915_reset_request(rq, stalled);
 801
 802                 GEM_BUG_ON(rq->ring != engine->buffer);
 803                 head = rq->head;
 804         } else {
 805                 head = engine->buffer->tail;
 806         }
 807         engine->buffer->head = intel_ring_wrap(engine->buffer, head);
 808
 809         spin_unlock_irqrestore(&tl->lock, flags);
 810 }
 811
 812 static void reset_finish(struct intel_engine_cs *engine)
 813 {
 814 }
 815
 816 static int intel_rcs_ctx_init(struct i915_request *rq)
 817 {
 818         int ret;
 819
 820         ret = intel_engine_emit_ctx_wa(rq);
 821         if (ret != 0)
 822                 return ret;
 823
 824         ret = i915_gem_render_state_emit(rq);
 825         if (ret)
 826                 return ret;
 827
 828         return 0;
 829 }
 830
 831 static int init_render_ring(struct intel_engine_cs *engine)
 832 {
 833         struct drm_i915_private *dev_priv = engine->i915;
 834         int ret = init_ring_common(engine);
 835         if (ret)
 836                 return ret;
 837
 838         /* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
 839         if (IS_GEN_RANGE(dev_priv, 4, 6))
 840                 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH));
 841
 842         /* We need to disable the AsyncFlip performance optimisations in order
 843          * to use MI_WAIT_FOR_EVENT within the CS. It should already be
 844          * programmed to '1' on all products.
 845          *
 846          * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv
 847          */
 848         if (IS_GEN_RANGE(dev_priv, 6, 7))
 849                 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
 850
 851         /* Required for the hardware to program scanline values for waiting */
 852         /* WaEnableFlushTlbInvalidationMode:snb */
 853         if (IS_GEN(dev_priv, 6))
 854                 I915_WRITE(GFX_MODE,
 855                            _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT));
 856
 857         /* WaBCSVCSTlbInvalidationMode:ivb,vlv,hsw */
 858         if (IS_GEN(dev_priv, 7))
 859                 I915_WRITE(GFX_MODE_GEN7,
 860                            _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT) |
 861                            _MASKED_BIT_ENABLE(GFX_REPLAY_MODE));
 862
 863         if (IS_GEN(dev_priv, 6)) {
 864                 /* From the Sandybridge PRM, volume 1 part 3, page 24:
 865                  * "If this bit is set, STCunit will have LRA as replacement
 866                  *  policy. [...] This bit must be reset.  LRA replacement
 867                  *  policy is not supported."
 868                  */
 869                 I915_WRITE(CACHE_MODE_0,
 870                            _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
 871         }
 872
 873         if (IS_GEN_RANGE(dev_priv, 6, 7))
 874                 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
 875
 876         if (INTEL_GEN(dev_priv) >= 6)
 877                 ENGINE_WRITE(engine, RING_IMR, ~engine->irq_keep_mask);
 878
 879         return 0;
 880 }
 881
 882 static void cancel_requests(struct intel_engine_cs *engine)
 883 {
 884         struct i915_request *request;
 885         unsigned long flags;
 886
 887         spin_lock_irqsave(&engine->timeline.lock, flags);
 888
 889         /* Mark all submitted requests as skipped. */
 890         list_for_each_entry(request, &engine->timeline.requests, link) {
 891                 if (!i915_request_signaled(request))
 892                         dma_fence_set_error(&request->fence, -EIO);
 893
 894                 i915_request_mark_complete(request);
 895         }
 896
 897         /* Remaining _unready_ requests will be nop'ed when submitted */
 898
 899         spin_unlock_irqrestore(&engine->timeline.lock, flags);
 900 }
 901
 902 static void i9xx_submit_request(struct i915_request *request)
 903 {
 904         i915_request_submit(request);
 905
 906         ENGINE_WRITE(request->engine, RING_TAIL,
 907                      intel_ring_set_tail(request->ring, request->tail));
 908 }
 909
 910 static u32 *i9xx_emit_breadcrumb(struct i915_request *rq, u32 *cs)
 911 {
 912         GEM_BUG_ON(rq->timeline->hwsp_ggtt != rq->engine->status_page.vma);
 913         GEM_BUG_ON(offset_in_page(rq->timeline->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
 914
 915         *cs++ = MI_FLUSH;
 916
 917         *cs++ = MI_STORE_DWORD_INDEX;
 918         *cs++ = I915_GEM_HWS_SEQNO_ADDR;
 919         *cs++ = rq->fence.seqno;
 920
 921         *cs++ = MI_STORE_DWORD_INDEX;
 922         *cs++ = I915_GEM_HWS_HANGCHECK_ADDR;
 923         *cs++ = intel_engine_next_hangcheck_seqno(rq->engine);
 924
 925         *cs++ = MI_USER_INTERRUPT;
 926
 927         rq->tail = intel_ring_offset(rq, cs);
 928         assert_ring_tail_valid(rq->ring, rq->tail);
 929
 930         return cs;
 931 }
 932
 933 #define GEN5_WA_STORES 8 /* must be at least 1! */
 934 static u32 *gen5_emit_breadcrumb(struct i915_request *rq, u32 *cs)
 935 {
 936         int i;
 937
 938         GEM_BUG_ON(rq->timeline->hwsp_ggtt != rq->engine->status_page.vma);
 939         GEM_BUG_ON(offset_in_page(rq->timeline->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
 940
 941         *cs++ = MI_FLUSH;
 942
 943         *cs++ = MI_STORE_DWORD_INDEX;
 944         *cs++ = I915_GEM_HWS_HANGCHECK_ADDR;
 945         *cs++ = intel_engine_next_hangcheck_seqno(rq->engine);
 946
 947         BUILD_BUG_ON(GEN5_WA_STORES < 1);
 948         for (i = 0; i < GEN5_WA_STORES; i++) {
 949                 *cs++ = MI_STORE_DWORD_INDEX;
 950                 *cs++ = I915_GEM_HWS_SEQNO_ADDR;
 951                 *cs++ = rq->fence.seqno;
 952         }
 953
 954         *cs++ = MI_USER_INTERRUPT;
 955         *cs++ = MI_NOOP;
 956
 957         rq->tail = intel_ring_offset(rq, cs);
 958         assert_ring_tail_valid(rq->ring, rq->tail);
 959
 960         return cs;
 961 }
 962 #undef GEN5_WA_STORES
 963
 964 static void
 965 gen5_irq_enable(struct intel_engine_cs *engine)
 966 {
 967         gen5_enable_gt_irq(engine->i915, engine->irq_enable_mask);
 968 }
 969
 970 static void
 971 gen5_irq_disable(struct intel_engine_cs *engine)
 972 {
 973         gen5_disable_gt_irq(engine->i915, engine->irq_enable_mask);
 974 }
 975
 976 static void
 977 i9xx_irq_enable(struct intel_engine_cs *engine)
 978 {
 979         GEM_BUG_ON(engine->id != RCS0);
 980
 981         engine->i915->irq_mask &= ~engine->irq_enable_mask;
 982         ENGINE_WRITE(engine, RING_IMR, engine->i915->irq_mask);
 983         ENGINE_POSTING_READ(engine, RING_IMR);
 984 }
 985
 986 static void
 987 i9xx_irq_disable(struct intel_engine_cs *engine)
 988 {
 989         GEM_BUG_ON(engine->id != RCS0);
 990
 991         engine->i915->irq_mask |= engine->irq_enable_mask;
 992         ENGINE_WRITE(engine, RING_IMR, engine->i915->irq_mask);
 993 }
 994
 995 static void
 996 i8xx_irq_enable(struct intel_engine_cs *engine)
 997 {
 998         struct drm_i915_private *dev_priv = engine->i915;
 999
1000         dev_priv->irq_mask &= ~engine->irq_enable_mask;
1001         I915_WRITE16(IMR, dev_priv->irq_mask);
1002         POSTING_READ16(RING_IMR(engine->mmio_base));
1003 }
1004
1005 static void
1006 i8xx_irq_disable(struct intel_engine_cs *engine)
1007 {
1008         struct drm_i915_private *dev_priv = engine->i915;
1009
1010         dev_priv->irq_mask |= engine->irq_enable_mask;
1011         I915_WRITE16(IMR, dev_priv->irq_mask);
1012 }
1013
1014 static int
1015 bsd_ring_flush(struct i915_request *rq, u32 mode)
1016 {
1017         u32 *cs;
1018
1019         cs = intel_ring_begin(rq, 2);
1020         if (IS_ERR(cs))
1021                 return PTR_ERR(cs);
1022
1023         *cs++ = MI_FLUSH;
1024         *cs++ = MI_NOOP;
1025         intel_ring_advance(rq, cs);
1026         return 0;
1027 }
1028
1029 static void
1030 gen6_irq_enable(struct intel_engine_cs *engine)
1031 {
1032         ENGINE_WRITE(engine, RING_IMR,
1033                      ~(engine->irq_enable_mask | engine->irq_keep_mask));
1034
1035         /* Flush/delay to ensure the RING_IMR is active before the GT IMR */
1036         ENGINE_POSTING_READ(engine, RING_IMR);
1037
1038         gen5_enable_gt_irq(engine->i915, engine->irq_enable_mask);
1039 }
1040
1041 static void
1042 gen6_irq_disable(struct intel_engine_cs *engine)
1043 {
1044         ENGINE_WRITE(engine, RING_IMR, ~engine->irq_keep_mask);
1045         gen5_disable_gt_irq(engine->i915, engine->irq_enable_mask);
1046 }
1047
1048 static void
1049 hsw_vebox_irq_enable(struct intel_engine_cs *engine)
1050 {
1051         ENGINE_WRITE(engine, RING_IMR, ~engine->irq_enable_mask);
1052
1053         /* Flush/delay to ensure the RING_IMR is active before the GT IMR */
1054         ENGINE_POSTING_READ(engine, RING_IMR);
1055
1056         gen6_unmask_pm_irq(engine->i915, engine->irq_enable_mask);
1057 }
1058
1059 static void
1060 hsw_vebox_irq_disable(struct intel_engine_cs *engine)
1061 {
1062         ENGINE_WRITE(engine, RING_IMR, ~0);
1063         gen6_mask_pm_irq(engine->i915, engine->irq_enable_mask);
1064 }
1065
1066 static int
1067 i965_emit_bb_start(struct i915_request *rq,
1068                    u64 offset, u32 length,
1069                    unsigned int dispatch_flags)
1070 {
1071         u32 *cs;
1072
1073         cs = intel_ring_begin(rq, 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(rq, 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 SZ_256K
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 i915_request *rq,
1091                    u64 offset, u32 len,
1092                    unsigned int dispatch_flags)
1093 {
1094         u32 *cs, cs_offset = i915_scratch_offset(rq->i915);
1095
1096         GEM_BUG_ON(rq->i915->gt.scratch->size < I830_WA_SIZE);
1097
1098         cs = intel_ring_begin(rq, 6);
1099         if (IS_ERR(cs))
1100                 return PTR_ERR(cs);
1101
1102         /* Evict the invalid PTE TLBs */
1103         *cs++ = COLOR_BLT_CMD | BLT_WRITE_RGBA;
1104         *cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | 4096;
1105         *cs++ = I830_TLB_ENTRIES << 16 | 4; /* load each page */
1106         *cs++ = cs_offset;
1107         *cs++ = 0xdeadbeef;
1108         *cs++ = MI_NOOP;
1109         intel_ring_advance(rq, cs);
1110
1111         if ((dispatch_flags & I915_DISPATCH_PINNED) == 0) {
1112                 if (len > I830_BATCH_LIMIT)
1113                         return -ENOSPC;
1114
1115                 cs = intel_ring_begin(rq, 6 + 2);
1116                 if (IS_ERR(cs))
1117                         return PTR_ERR(cs);
1118
1119                 /* Blit the batch (which has now all relocs applied) to the
1120                  * stable batch scratch bo area (so that the CS never
1121                  * stumbles over its tlb invalidation bug) ...
1122                  */
1123                 *cs++ = SRC_COPY_BLT_CMD | BLT_WRITE_RGBA;
1124                 *cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | 4096;
1125                 *cs++ = DIV_ROUND_UP(len, 4096) << 16 | 4096;
1126                 *cs++ = cs_offset;
1127                 *cs++ = 4096;
1128                 *cs++ = offset;
1129
1130                 *cs++ = MI_FLUSH;
1131                 *cs++ = MI_NOOP;
1132                 intel_ring_advance(rq, cs);
1133
1134                 /* ... and execute it. */
1135                 offset = cs_offset;
1136         }
1137
1138         cs = intel_ring_begin(rq, 2);
1139         if (IS_ERR(cs))
1140                 return PTR_ERR(cs);
1141
1142         *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
1143         *cs++ = offset | (dispatch_flags & I915_DISPATCH_SECURE ? 0 :
1144                 MI_BATCH_NON_SECURE);
1145         intel_ring_advance(rq, cs);
1146
1147         return 0;
1148 }
1149
1150 static int
1151 i915_emit_bb_start(struct i915_request *rq,
1152                    u64 offset, u32 len,
1153                    unsigned int dispatch_flags)
1154 {
1155         u32 *cs;
1156
1157         cs = intel_ring_begin(rq, 2);
1158         if (IS_ERR(cs))
1159                 return PTR_ERR(cs);
1160
1161         *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
1162         *cs++ = offset | (dispatch_flags & I915_DISPATCH_SECURE ? 0 :
1163                 MI_BATCH_NON_SECURE);
1164         intel_ring_advance(rq, cs);
1165
1166         return 0;
1167 }
1168
1169 int intel_ring_pin(struct intel_ring *ring)
1170 {
1171         struct i915_vma *vma = ring->vma;
1172         enum i915_map_type map = i915_coherent_map_type(vma->vm->i915);
1173         unsigned int flags;
1174         void *addr;
1175         int ret;
1176
1177         GEM_BUG_ON(ring->vaddr);
1178
1179         ret = i915_timeline_pin(ring->timeline);
1180         if (ret)
1181                 return ret;
1182
1183         flags = PIN_GLOBAL;
1184
1185         /* Ring wraparound at offset 0 sometimes hangs. No idea why. */
1186         flags |= PIN_OFFSET_BIAS | i915_ggtt_pin_bias(vma);
1187
1188         if (vma->obj->stolen)
1189                 flags |= PIN_MAPPABLE;
1190         else
1191                 flags |= PIN_HIGH;
1192
1193         ret = i915_vma_pin(vma, 0, 0, flags);
1194         if (unlikely(ret))
1195                 goto unpin_timeline;
1196
1197         if (i915_vma_is_map_and_fenceable(vma))
1198                 addr = (void __force *)i915_vma_pin_iomap(vma);
1199         else
1200                 addr = i915_gem_object_pin_map(vma->obj, map);
1201         if (IS_ERR(addr)) {
1202                 ret = PTR_ERR(addr);
1203                 goto unpin_ring;
1204         }
1205
1206         vma->obj->pin_global++;
1207
1208         ring->vaddr = addr;
1209         return 0;
1210
1211 unpin_ring:
1212         i915_vma_unpin(vma);
1213 unpin_timeline:
1214         i915_timeline_unpin(ring->timeline);
1215         return ret;
1216 }
1217
1218 void intel_ring_reset(struct intel_ring *ring, u32 tail)
1219 {
1220         GEM_BUG_ON(!intel_ring_offset_valid(ring, tail));
1221
1222         ring->tail = tail;
1223         ring->head = tail;
1224         ring->emit = tail;
1225         intel_ring_update_space(ring);
1226 }
1227
1228 void intel_ring_unpin(struct intel_ring *ring)
1229 {
1230         GEM_BUG_ON(!ring->vma);
1231         GEM_BUG_ON(!ring->vaddr);
1232
1233         /* Discard any unused bytes beyond that submitted to hw. */
1234         intel_ring_reset(ring, ring->tail);
1235
1236         if (i915_vma_is_map_and_fenceable(ring->vma))
1237                 i915_vma_unpin_iomap(ring->vma);
1238         else
1239                 i915_gem_object_unpin_map(ring->vma->obj);
1240         ring->vaddr = NULL;
1241
1242         ring->vma->obj->pin_global--;
1243         i915_vma_unpin(ring->vma);
1244
1245         i915_timeline_unpin(ring->timeline);
1246 }
1247
1248 static struct i915_vma *
1249 intel_ring_create_vma(struct drm_i915_private *dev_priv, int size)
1250 {
1251         struct i915_address_space *vm = &dev_priv->ggtt.vm;
1252         struct drm_i915_gem_object *obj;
1253         struct i915_vma *vma;
1254
1255         obj = i915_gem_object_create_stolen(dev_priv, size);
1256         if (!obj)
1257                 obj = i915_gem_object_create_internal(dev_priv, size);
1258         if (IS_ERR(obj))
1259                 return ERR_CAST(obj);
1260
1261         /*
1262          * Mark ring buffers as read-only from GPU side (so no stray overwrites)
1263          * if supported by the platform's GGTT.
1264          */
1265         if (vm->has_read_only)
1266                 i915_gem_object_set_readonly(obj);
1267
1268         vma = i915_vma_instance(obj, vm, NULL);
1269         if (IS_ERR(vma))
1270                 goto err;
1271
1272         return vma;
1273
1274 err:
1275         i915_gem_object_put(obj);
1276         return vma;
1277 }
1278
1279 struct intel_ring *
1280 intel_engine_create_ring(struct intel_engine_cs *engine,
1281                          struct i915_timeline *timeline,
1282                          int size)
1283 {
1284         struct intel_ring *ring;
1285         struct i915_vma *vma;
1286
1287         GEM_BUG_ON(!is_power_of_2(size));
1288         GEM_BUG_ON(RING_CTL_SIZE(size) & ~RING_NR_PAGES);
1289         GEM_BUG_ON(timeline == &engine->timeline);
1290         lockdep_assert_held(&engine->i915->drm.struct_mutex);
1291
1292         ring = kzalloc(sizeof(*ring), GFP_KERNEL);
1293         if (!ring)
1294                 return ERR_PTR(-ENOMEM);
1295
1296         kref_init(&ring->ref);
1297         INIT_LIST_HEAD(&ring->request_list);
1298         ring->timeline = i915_timeline_get(timeline);
1299
1300         ring->size = size;
1301         /* Workaround an erratum on the i830 which causes a hang if
1302          * the TAIL pointer points to within the last 2 cachelines
1303          * of the buffer.
1304          */
1305         ring->effective_size = size;
1306         if (IS_I830(engine->i915) || IS_I845G(engine->i915))
1307                 ring->effective_size -= 2 * CACHELINE_BYTES;
1308
1309         intel_ring_update_space(ring);
1310
1311         vma = intel_ring_create_vma(engine->i915, size);
1312         if (IS_ERR(vma)) {
1313                 kfree(ring);
1314                 return ERR_CAST(vma);
1315         }
1316         ring->vma = vma;
1317
1318         return ring;
1319 }
1320
1321 void intel_ring_free(struct kref *ref)
1322 {
1323         struct intel_ring *ring = container_of(ref, typeof(*ring), ref);
1324         struct drm_i915_gem_object *obj = ring->vma->obj;
1325
1326         i915_vma_close(ring->vma);
1327         __i915_gem_object_release_unless_active(obj);
1328
1329         i915_timeline_put(ring->timeline);
1330         kfree(ring);
1331 }
1332
1333 static void __ring_context_fini(struct intel_context *ce)
1334 {
1335         GEM_BUG_ON(i915_gem_object_is_active(ce->state->obj));
1336         i915_gem_object_put(ce->state->obj);
1337 }
1338
1339 static void ring_context_destroy(struct kref *ref)
1340 {
1341         struct intel_context *ce = container_of(ref, typeof(*ce), ref);
1342
1343         GEM_BUG_ON(intel_context_is_pinned(ce));
1344
1345         if (ce->state)
1346                 __ring_context_fini(ce);
1347
1348         intel_context_free(ce);
1349 }
1350
1351 static int __context_pin_ppgtt(struct i915_gem_context *ctx)
1352 {
1353         struct i915_hw_ppgtt *ppgtt;
1354         int err = 0;
1355
1356         ppgtt = ctx->ppgtt ?: ctx->i915->mm.aliasing_ppgtt;
1357         if (ppgtt)
1358                 err = gen6_ppgtt_pin(ppgtt);
1359
1360         return err;
1361 }
1362
1363 static void __context_unpin_ppgtt(struct i915_gem_context *ctx)
1364 {
1365         struct i915_hw_ppgtt *ppgtt;
1366
1367         ppgtt = ctx->ppgtt ?: ctx->i915->mm.aliasing_ppgtt;
1368         if (ppgtt)
1369                 gen6_ppgtt_unpin(ppgtt);
1370 }
1371
1372 static int __context_pin(struct intel_context *ce)
1373 {
1374         struct i915_vma *vma;
1375         int err;
1376
1377         vma = ce->state;
1378         if (!vma)
1379                 return 0;
1380
1381         err = i915_vma_pin(vma, 0, 0, PIN_GLOBAL | PIN_HIGH);
1382         if (err)
1383                 return err;
1384
1385         /*
1386          * And mark is as a globally pinned object to let the shrinker know
1387          * it cannot reclaim the object until we release it.
1388          */
1389         vma->obj->pin_global++;
1390         vma->obj->mm.dirty = true;
1391
1392         return 0;
1393 }
1394
1395 static void __context_unpin(struct intel_context *ce)
1396 {
1397         struct i915_vma *vma;
1398
1399         vma = ce->state;
1400         if (!vma)
1401                 return;
1402
1403         vma->obj->pin_global--;
1404         i915_vma_unpin(vma);
1405 }
1406
1407 static void ring_context_unpin(struct intel_context *ce)
1408 {
1409         __context_unpin_ppgtt(ce->gem_context);
1410         __context_unpin(ce);
1411 }
1412
1413 static struct i915_vma *
1414 alloc_context_vma(struct intel_engine_cs *engine)
1415 {
1416         struct drm_i915_private *i915 = engine->i915;
1417         struct drm_i915_gem_object *obj;
1418         struct i915_vma *vma;
1419         int err;
1420
1421         obj = i915_gem_object_create(i915, engine->context_size);
1422         if (IS_ERR(obj))
1423                 return ERR_CAST(obj);
1424
1425         /*
1426          * Try to make the context utilize L3 as well as LLC.
1427          *
1428          * On VLV we don't have L3 controls in the PTEs so we
1429          * shouldn't touch the cache level, especially as that
1430          * would make the object snooped which might have a
1431          * negative performance impact.
1432          *
1433          * Snooping is required on non-llc platforms in execlist
1434          * mode, but since all GGTT accesses use PAT entry 0 we
1435          * get snooping anyway regardless of cache_level.
1436          *
1437          * This is only applicable for Ivy Bridge devices since
1438          * later platforms don't have L3 control bits in the PTE.
1439          */
1440         if (IS_IVYBRIDGE(i915))
1441                 i915_gem_object_set_cache_coherency(obj, I915_CACHE_L3_LLC);
1442
1443         if (engine->default_state) {
1444                 void *defaults, *vaddr;
1445
1446                 vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
1447                 if (IS_ERR(vaddr)) {
1448                         err = PTR_ERR(vaddr);
1449                         goto err_obj;
1450                 }
1451
1452                 defaults = i915_gem_object_pin_map(engine->default_state,
1453                                                    I915_MAP_WB);
1454                 if (IS_ERR(defaults)) {
1455                         err = PTR_ERR(defaults);
1456                         goto err_map;
1457                 }
1458
1459                 memcpy(vaddr, defaults, engine->context_size);
1460                 i915_gem_object_unpin_map(engine->default_state);
1461
1462                 i915_gem_object_flush_map(obj);
1463                 i915_gem_object_unpin_map(obj);
1464         }
1465
1466         vma = i915_vma_instance(obj, &i915->ggtt.vm, NULL);
1467         if (IS_ERR(vma)) {
1468                 err = PTR_ERR(vma);
1469                 goto err_obj;
1470         }
1471
1472         return vma;
1473
1474 err_map:
1475         i915_gem_object_unpin_map(obj);
1476 err_obj:
1477         i915_gem_object_put(obj);
1478         return ERR_PTR(err);
1479 }
1480
1481 static int ring_context_pin(struct intel_context *ce)
1482 {
1483         struct intel_engine_cs *engine = ce->engine;
1484         int err;
1485
1486         /* One ringbuffer to rule them all */
1487         GEM_BUG_ON(!engine->buffer);
1488         ce->ring = engine->buffer;
1489
1490         if (!ce->state && engine->context_size) {
1491                 struct i915_vma *vma;
1492
1493                 vma = alloc_context_vma(engine);
1494                 if (IS_ERR(vma))
1495                         return PTR_ERR(vma);
1496
1497                 ce->state = vma;
1498         }
1499
1500         err = __context_pin(ce);
1501         if (err)
1502                 return err;
1503
1504         err = __context_pin_ppgtt(ce->gem_context);
1505         if (err)
1506                 goto err_unpin;
1507
1508         return 0;
1509
1510 err_unpin:
1511         __context_unpin(ce);
1512         return err;
1513 }
1514
1515 static const struct intel_context_ops ring_context_ops = {
1516         .pin = ring_context_pin,
1517         .unpin = ring_context_unpin,
1518         .destroy = ring_context_destroy,
1519 };
1520
1521 static int intel_init_ring_buffer(struct intel_engine_cs *engine)
1522 {
1523         struct i915_timeline *timeline;
1524         struct intel_ring *ring;
1525         int err;
1526
1527         err = intel_engine_setup_common(engine);
1528         if (err)
1529                 return err;
1530
1531         timeline = i915_timeline_create(engine->i915, engine->status_page.vma);
1532         if (IS_ERR(timeline)) {
1533                 err = PTR_ERR(timeline);
1534                 goto err;
1535         }
1536         GEM_BUG_ON(timeline->has_initial_breadcrumb);
1537
1538         ring = intel_engine_create_ring(engine, timeline, 32 * PAGE_SIZE);
1539         i915_timeline_put(timeline);
1540         if (IS_ERR(ring)) {
1541                 err = PTR_ERR(ring);
1542                 goto err;
1543         }
1544
1545         err = intel_ring_pin(ring);
1546         if (err)
1547                 goto err_ring;
1548
1549         GEM_BUG_ON(engine->buffer);
1550         engine->buffer = ring;
1551
1552         err = intel_engine_init_common(engine);
1553         if (err)
1554                 goto err_unpin;
1555
1556         GEM_BUG_ON(ring->timeline->hwsp_ggtt != engine->status_page.vma);
1557
1558         return 0;
1559
1560 err_unpin:
1561         intel_ring_unpin(ring);
1562 err_ring:
1563         intel_ring_put(ring);
1564 err:
1565         intel_engine_cleanup_common(engine);
1566         return err;
1567 }
1568
1569 void intel_engine_cleanup(struct intel_engine_cs *engine)
1570 {
1571         struct drm_i915_private *dev_priv = engine->i915;
1572
1573         WARN_ON(INTEL_GEN(dev_priv) > 2 &&
1574                 (ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
1575
1576         intel_ring_unpin(engine->buffer);
1577         intel_ring_put(engine->buffer);
1578
1579         if (engine->cleanup)
1580                 engine->cleanup(engine);
1581
1582         intel_engine_cleanup_common(engine);
1583
1584         dev_priv->engine[engine->id] = NULL;
1585         kfree(engine);
1586 }
1587
1588 void intel_legacy_submission_resume(struct drm_i915_private *dev_priv)
1589 {
1590         struct intel_engine_cs *engine;
1591         enum intel_engine_id id;
1592
1593         /* Restart from the beginning of the rings for convenience */
1594         for_each_engine(engine, dev_priv, id)
1595                 intel_ring_reset(engine->buffer, 0);
1596 }
1597
1598 static int load_pd_dir(struct i915_request *rq,
1599                        const struct i915_hw_ppgtt *ppgtt)
1600 {
1601         const struct intel_engine_cs * const engine = rq->engine;
1602         u32 *cs;
1603
1604         cs = intel_ring_begin(rq, 6);
1605         if (IS_ERR(cs))
1606                 return PTR_ERR(cs);
1607
1608         *cs++ = MI_LOAD_REGISTER_IMM(1);
1609         *cs++ = i915_mmio_reg_offset(RING_PP_DIR_DCLV(engine->mmio_base));
1610         *cs++ = PP_DIR_DCLV_2G;
1611
1612         *cs++ = MI_LOAD_REGISTER_IMM(1);
1613         *cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
1614         *cs++ = ppgtt->pd.base.ggtt_offset << 10;
1615
1616         intel_ring_advance(rq, cs);
1617
1618         return 0;
1619 }
1620
1621 static int flush_pd_dir(struct i915_request *rq)
1622 {
1623         const struct intel_engine_cs * const engine = rq->engine;
1624         u32 *cs;
1625
1626         cs = intel_ring_begin(rq, 4);
1627         if (IS_ERR(cs))
1628                 return PTR_ERR(cs);
1629
1630         /* Stall until the page table load is complete */
1631         *cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
1632         *cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
1633         *cs++ = i915_scratch_offset(rq->i915);
1634         *cs++ = MI_NOOP;
1635
1636         intel_ring_advance(rq, cs);
1637         return 0;
1638 }
1639
1640 static inline int mi_set_context(struct i915_request *rq, u32 flags)
1641 {
1642         struct drm_i915_private *i915 = rq->i915;
1643         struct intel_engine_cs *engine = rq->engine;
1644         enum intel_engine_id id;
1645         const int num_engines =
1646                 IS_HSW_GT1(i915) ? RUNTIME_INFO(i915)->num_engines - 1 : 0;
1647         bool force_restore = false;
1648         int len;
1649         u32 *cs;
1650
1651         flags |= MI_MM_SPACE_GTT;
1652         if (IS_HASWELL(i915))
1653                 /* These flags are for resource streamer on HSW+ */
1654                 flags |= HSW_MI_RS_SAVE_STATE_EN | HSW_MI_RS_RESTORE_STATE_EN;
1655         else
1656                 flags |= MI_SAVE_EXT_STATE_EN | MI_RESTORE_EXT_STATE_EN;
1657
1658         len = 4;
1659         if (IS_GEN(i915, 7))
1660                 len += 2 + (num_engines ? 4 * num_engines + 6 : 0);
1661         if (flags & MI_FORCE_RESTORE) {
1662                 GEM_BUG_ON(flags & MI_RESTORE_INHIBIT);
1663                 flags &= ~MI_FORCE_RESTORE;
1664                 force_restore = true;
1665                 len += 2;
1666         }
1667
1668         cs = intel_ring_begin(rq, len);
1669         if (IS_ERR(cs))
1670                 return PTR_ERR(cs);
1671
1672         /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
1673         if (IS_GEN(i915, 7)) {
1674                 *cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
1675                 if (num_engines) {
1676                         struct intel_engine_cs *signaller;
1677
1678                         *cs++ = MI_LOAD_REGISTER_IMM(num_engines);
1679                         for_each_engine(signaller, i915, id) {
1680                                 if (signaller == engine)
1681                                         continue;
1682
1683                                 *cs++ = i915_mmio_reg_offset(
1684                                            RING_PSMI_CTL(signaller->mmio_base));
1685                                 *cs++ = _MASKED_BIT_ENABLE(
1686                                                 GEN6_PSMI_SLEEP_MSG_DISABLE);
1687                         }
1688                 }
1689         }
1690
1691         if (force_restore) {
1692                 /*
1693                  * The HW doesn't handle being told to restore the current
1694                  * context very well. Quite often it likes goes to go off and
1695                  * sulk, especially when it is meant to be reloading PP_DIR.
1696                  * A very simple fix to force the reload is to simply switch
1697                  * away from the current context and back again.
1698                  *
1699                  * Note that the kernel_context will contain random state
1700                  * following the INHIBIT_RESTORE. We accept this since we
1701                  * never use the kernel_context state; it is merely a
1702                  * placeholder we use to flush other contexts.
1703                  */
1704                 *cs++ = MI_SET_CONTEXT;
1705                 *cs++ = i915_ggtt_offset(engine->kernel_context->state) |
1706                         MI_MM_SPACE_GTT |
1707                         MI_RESTORE_INHIBIT;
1708         }
1709
1710         *cs++ = MI_NOOP;
1711         *cs++ = MI_SET_CONTEXT;
1712         *cs++ = i915_ggtt_offset(rq->hw_context->state) | flags;
1713         /*
1714          * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
1715          * WaMiSetContext_Hang:snb,ivb,vlv
1716          */
1717         *cs++ = MI_NOOP;
1718
1719         if (IS_GEN(i915, 7)) {
1720                 if (num_engines) {
1721                         struct intel_engine_cs *signaller;
1722                         i915_reg_t last_reg = {}; /* keep gcc quiet */
1723
1724                         *cs++ = MI_LOAD_REGISTER_IMM(num_engines);
1725                         for_each_engine(signaller, i915, id) {
1726                                 if (signaller == engine)
1727                                         continue;
1728
1729                                 last_reg = RING_PSMI_CTL(signaller->mmio_base);
1730                                 *cs++ = i915_mmio_reg_offset(last_reg);
1731                                 *cs++ = _MASKED_BIT_DISABLE(
1732                                                 GEN6_PSMI_SLEEP_MSG_DISABLE);
1733                         }
1734
1735                         /* Insert a delay before the next switch! */
1736                         *cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
1737                         *cs++ = i915_mmio_reg_offset(last_reg);
1738                         *cs++ = i915_scratch_offset(rq->i915);
1739                         *cs++ = MI_NOOP;
1740                 }
1741                 *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1742         }
1743
1744         intel_ring_advance(rq, cs);
1745
1746         return 0;
1747 }
1748
1749 static int remap_l3(struct i915_request *rq, int slice)
1750 {
1751         u32 *cs, *remap_info = rq->i915->l3_parity.remap_info[slice];
1752         int i;
1753
1754         if (!remap_info)
1755                 return 0;
1756
1757         cs = intel_ring_begin(rq, GEN7_L3LOG_SIZE/4 * 2 + 2);
1758         if (IS_ERR(cs))
1759                 return PTR_ERR(cs);
1760
1761         /*
1762          * Note: We do not worry about the concurrent register cacheline hang
1763          * here because no other code should access these registers other than
1764          * at initialization time.
1765          */
1766         *cs++ = MI_LOAD_REGISTER_IMM(GEN7_L3LOG_SIZE/4);
1767         for (i = 0; i < GEN7_L3LOG_SIZE/4; i++) {
1768                 *cs++ = i915_mmio_reg_offset(GEN7_L3LOG(slice, i));
1769                 *cs++ = remap_info[i];
1770         }
1771         *cs++ = MI_NOOP;
1772         intel_ring_advance(rq, cs);
1773
1774         return 0;
1775 }
1776
1777 static int switch_context(struct i915_request *rq)
1778 {
1779         struct intel_engine_cs *engine = rq->engine;
1780         struct i915_gem_context *ctx = rq->gem_context;
1781         struct i915_hw_ppgtt *ppgtt = ctx->ppgtt ?: rq->i915->mm.aliasing_ppgtt;
1782         unsigned int unwind_mm = 0;
1783         u32 hw_flags = 0;
1784         int ret, i;
1785
1786         lockdep_assert_held(&rq->i915->drm.struct_mutex);
1787         GEM_BUG_ON(HAS_EXECLISTS(rq->i915));
1788
1789         if (ppgtt) {
1790                 int loops;
1791
1792                 /*
1793                  * Baytail takes a little more convincing that it really needs
1794                  * to reload the PD between contexts. It is not just a little
1795                  * longer, as adding more stalls after the load_pd_dir (i.e.
1796                  * adding a long loop around flush_pd_dir) is not as effective
1797                  * as reloading the PD umpteen times. 32 is derived from
1798                  * experimentation (gem_exec_parallel/fds) and has no good
1799                  * explanation.
1800                  */
1801                 loops = 1;
1802                 if (engine->id == BCS0 && IS_VALLEYVIEW(engine->i915))
1803                         loops = 32;
1804
1805                 do {
1806                         ret = load_pd_dir(rq, ppgtt);
1807                         if (ret)
1808                                 goto err;
1809                 } while (--loops);
1810
1811                 if (ppgtt->pd_dirty_engines & engine->mask) {
1812                         unwind_mm = engine->mask;
1813                         ppgtt->pd_dirty_engines &= ~unwind_mm;
1814                         hw_flags = MI_FORCE_RESTORE;
1815                 }
1816         }
1817
1818         if (rq->hw_context->state) {
1819                 GEM_BUG_ON(engine->id != RCS0);
1820
1821                 /*
1822                  * The kernel context(s) is treated as pure scratch and is not
1823                  * expected to retain any state (as we sacrifice it during
1824                  * suspend and on resume it may be corrupted). This is ok,
1825                  * as nothing actually executes using the kernel context; it
1826                  * is purely used for flushing user contexts.
1827                  */
1828                 if (i915_gem_context_is_kernel(ctx))
1829                         hw_flags = MI_RESTORE_INHIBIT;
1830
1831                 ret = mi_set_context(rq, hw_flags);
1832                 if (ret)
1833                         goto err_mm;
1834         }
1835
1836         if (ppgtt) {
1837                 ret = engine->emit_flush(rq, EMIT_INVALIDATE);
1838                 if (ret)
1839                         goto err_mm;
1840
1841                 ret = flush_pd_dir(rq);
1842                 if (ret)
1843                         goto err_mm;
1844
1845                 /*
1846                  * Not only do we need a full barrier (post-sync write) after
1847                  * invalidating the TLBs, but we need to wait a little bit
1848                  * longer. Whether this is merely delaying us, or the
1849                  * subsequent flush is a key part of serialising with the
1850                  * post-sync op, this extra pass appears vital before a
1851                  * mm switch!
1852                  */
1853                 ret = engine->emit_flush(rq, EMIT_INVALIDATE);
1854                 if (ret)
1855                         goto err_mm;
1856
1857                 ret = engine->emit_flush(rq, EMIT_FLUSH);
1858                 if (ret)
1859                         goto err_mm;
1860         }
1861
1862         if (ctx->remap_slice) {
1863                 for (i = 0; i < MAX_L3_SLICES; i++) {
1864                         if (!(ctx->remap_slice & BIT(i)))
1865                                 continue;
1866
1867                         ret = remap_l3(rq, i);
1868                         if (ret)
1869                                 goto err_mm;
1870                 }
1871
1872                 ctx->remap_slice = 0;
1873         }
1874
1875         return 0;
1876
1877 err_mm:
1878         if (unwind_mm)
1879                 ppgtt->pd_dirty_engines |= unwind_mm;
1880 err:
1881         return ret;
1882 }
1883
1884 static int ring_request_alloc(struct i915_request *request)
1885 {
1886         int ret;
1887
1888         GEM_BUG_ON(!intel_context_is_pinned(request->hw_context));
1889         GEM_BUG_ON(request->timeline->has_initial_breadcrumb);
1890
1891         /*
1892          * Flush enough space to reduce the likelihood of waiting after
1893          * we start building the request - in which case we will just
1894          * have to repeat work.
1895          */
1896         request->reserved_space += LEGACY_REQUEST_SIZE;
1897
1898         ret = switch_context(request);
1899         if (ret)
1900                 return ret;
1901
1902         /* Unconditionally invalidate GPU caches and TLBs. */
1903         ret = request->engine->emit_flush(request, EMIT_INVALIDATE);
1904         if (ret)
1905                 return ret;
1906
1907         request->reserved_space -= LEGACY_REQUEST_SIZE;
1908         return 0;
1909 }
1910
1911 static noinline int wait_for_space(struct intel_ring *ring, unsigned int bytes)
1912 {
1913         struct i915_request *target;
1914         long timeout;
1915
1916         lockdep_assert_held(&ring->vma->vm->i915->drm.struct_mutex);
1917
1918         if (intel_ring_update_space(ring) >= bytes)
1919                 return 0;
1920
1921         GEM_BUG_ON(list_empty(&ring->request_list));
1922         list_for_each_entry(target, &ring->request_list, ring_link) {
1923                 /* Would completion of this request free enough space? */
1924                 if (bytes <= __intel_ring_space(target->postfix,
1925                                                 ring->emit, ring->size))
1926                         break;
1927         }
1928
1929         if (WARN_ON(&target->ring_link == &ring->request_list))
1930                 return -ENOSPC;
1931
1932         timeout = i915_request_wait(target,
1933                                     I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
1934                                     MAX_SCHEDULE_TIMEOUT);
1935         if (timeout < 0)
1936                 return timeout;
1937
1938         i915_request_retire_upto(target);
1939
1940         intel_ring_update_space(ring);
1941         GEM_BUG_ON(ring->space < bytes);
1942         return 0;
1943 }
1944
1945 u32 *intel_ring_begin(struct i915_request *rq, unsigned int num_dwords)
1946 {
1947         struct intel_ring *ring = rq->ring;
1948         const unsigned int remain_usable = ring->effective_size - ring->emit;
1949         const unsigned int bytes = num_dwords * sizeof(u32);
1950         unsigned int need_wrap = 0;
1951         unsigned int total_bytes;
1952         u32 *cs;
1953
1954         /* Packets must be qword aligned. */
1955         GEM_BUG_ON(num_dwords & 1);
1956
1957         total_bytes = bytes + rq->reserved_space;
1958         GEM_BUG_ON(total_bytes > ring->effective_size);
1959
1960         if (unlikely(total_bytes > remain_usable)) {
1961                 const int remain_actual = ring->size - ring->emit;
1962
1963                 if (bytes > remain_usable) {
1964                         /*
1965                          * Not enough space for the basic request. So need to
1966                          * flush out the remainder and then wait for
1967                          * base + reserved.
1968                          */
1969                         total_bytes += remain_actual;
1970                         need_wrap = remain_actual | 1;
1971                 } else  {
1972                         /*
1973                          * The base request will fit but the reserved space
1974                          * falls off the end. So we don't need an immediate
1975                          * wrap and only need to effectively wait for the
1976                          * reserved size from the start of ringbuffer.
1977                          */
1978                         total_bytes = rq->reserved_space + remain_actual;
1979                 }
1980         }
1981
1982         if (unlikely(total_bytes > ring->space)) {
1983                 int ret;
1984
1985                 /*
1986                  * Space is reserved in the ringbuffer for finalising the
1987                  * request, as that cannot be allowed to fail. During request
1988                  * finalisation, reserved_space is set to 0 to stop the
1989                  * overallocation and the assumption is that then we never need
1990                  * to wait (which has the risk of failing with EINTR).
1991                  *
1992                  * See also i915_request_alloc() and i915_request_add().
1993                  */
1994                 GEM_BUG_ON(!rq->reserved_space);
1995
1996                 ret = wait_for_space(ring, total_bytes);
1997                 if (unlikely(ret))
1998                         return ERR_PTR(ret);
1999         }
2000
2001         if (unlikely(need_wrap)) {
2002                 need_wrap &= ~1;
2003                 GEM_BUG_ON(need_wrap > ring->space);
2004                 GEM_BUG_ON(ring->emit + need_wrap > ring->size);
2005                 GEM_BUG_ON(!IS_ALIGNED(need_wrap, sizeof(u64)));
2006
2007                 /* Fill the tail with MI_NOOP */
2008                 memset64(ring->vaddr + ring->emit, 0, need_wrap / sizeof(u64));
2009                 ring->space -= need_wrap;
2010                 ring->emit = 0;
2011         }
2012
2013         GEM_BUG_ON(ring->emit > ring->size - bytes);
2014         GEM_BUG_ON(ring->space < bytes);
2015         cs = ring->vaddr + ring->emit;
2016         GEM_DEBUG_EXEC(memset32(cs, POISON_INUSE, bytes / sizeof(*cs)));
2017         ring->emit += bytes;
2018         ring->space -= bytes;
2019
2020         return cs;
2021 }
2022
2023 /* Align the ring tail to a cacheline boundary */
2024 int intel_ring_cacheline_align(struct i915_request *rq)
2025 {
2026         int num_dwords;
2027         void *cs;
2028
2029         num_dwords = (rq->ring->emit & (CACHELINE_BYTES - 1)) / sizeof(u32);
2030         if (num_dwords == 0)
2031                 return 0;
2032
2033         num_dwords = CACHELINE_DWORDS - num_dwords;
2034         GEM_BUG_ON(num_dwords & 1);
2035
2036         cs = intel_ring_begin(rq, num_dwords);
2037         if (IS_ERR(cs))
2038                 return PTR_ERR(cs);
2039
2040         memset64(cs, (u64)MI_NOOP << 32 | MI_NOOP, num_dwords / 2);
2041         intel_ring_advance(rq, cs);
2042
2043         GEM_BUG_ON(rq->ring->emit & (CACHELINE_BYTES - 1));
2044         return 0;
2045 }
2046
2047 static void gen6_bsd_submit_request(struct i915_request *request)
2048 {
2049         struct intel_uncore *uncore = request->engine->uncore;
2050
2051         intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
2052
2053        /* Every tail move must follow the sequence below */
2054
2055         /* Disable notification that the ring is IDLE. The GT
2056          * will then assume that it is busy and bring it out of rc6.
2057          */
2058         intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
2059                               _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
2060
2061         /* Clear the context id. Here be magic! */
2062         intel_uncore_write64_fw(uncore, GEN6_BSD_RNCID, 0x0);
2063
2064         /* Wait for the ring not to be idle, i.e. for it to wake up. */
2065         if (__intel_wait_for_register_fw(uncore,
2066                                          GEN6_BSD_SLEEP_PSMI_CONTROL,
2067                                          GEN6_BSD_SLEEP_INDICATOR,
2068                                          0,
2069                                          1000, 0, NULL))
2070                 DRM_ERROR("timed out waiting for the BSD ring to wake up\n");
2071
2072         /* Now that the ring is fully powered up, update the tail */
2073         i9xx_submit_request(request);
2074
2075         /* Let the ring send IDLE messages to the GT again,
2076          * and so let it sleep to conserve power when idle.
2077          */
2078         intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
2079                               _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
2080
2081         intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
2082 }
2083
2084 static int mi_flush_dw(struct i915_request *rq, u32 flags)
2085 {
2086         u32 cmd, *cs;
2087
2088         cs = intel_ring_begin(rq, 4);
2089         if (IS_ERR(cs))
2090                 return PTR_ERR(cs);
2091
2092         cmd = MI_FLUSH_DW;
2093
2094         /*
2095          * We always require a command barrier so that subsequent
2096          * commands, such as breadcrumb interrupts, are strictly ordered
2097          * wrt the contents of the write cache being flushed to memory
2098          * (and thus being coherent from the CPU).
2099          */
2100         cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
2101
2102         /*
2103          * Bspec vol 1c.3 - blitter engine command streamer:
2104          * "If ENABLED, all TLBs will be invalidated once the flush
2105          * operation is complete. This bit is only valid when the
2106          * Post-Sync Operation field is a value of 1h or 3h."
2107          */
2108         cmd |= flags;
2109
2110         *cs++ = cmd;
2111         *cs++ = I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT;
2112         *cs++ = 0;
2113         *cs++ = MI_NOOP;
2114
2115         intel_ring_advance(rq, cs);
2116
2117         return 0;
2118 }
2119
2120 static int gen6_flush_dw(struct i915_request *rq, u32 mode, u32 invflags)
2121 {
2122         return mi_flush_dw(rq, mode & EMIT_INVALIDATE ? invflags : 0);
2123 }
2124
2125 static int gen6_bsd_ring_flush(struct i915_request *rq, u32 mode)
2126 {
2127         return gen6_flush_dw(rq, mode, MI_INVALIDATE_TLB | MI_INVALIDATE_BSD);
2128 }
2129
2130 static int
2131 hsw_emit_bb_start(struct i915_request *rq,
2132                   u64 offset, u32 len,
2133                   unsigned int dispatch_flags)
2134 {
2135         u32 *cs;
2136
2137         cs = intel_ring_begin(rq, 2);
2138         if (IS_ERR(cs))
2139                 return PTR_ERR(cs);
2140
2141         *cs++ = MI_BATCH_BUFFER_START | (dispatch_flags & I915_DISPATCH_SECURE ?
2142                 0 : MI_BATCH_PPGTT_HSW | MI_BATCH_NON_SECURE_HSW);
2143         /* bit0-7 is the length on GEN6+ */
2144         *cs++ = offset;
2145         intel_ring_advance(rq, cs);
2146
2147         return 0;
2148 }
2149
2150 static int
2151 gen6_emit_bb_start(struct i915_request *rq,
2152                    u64 offset, u32 len,
2153                    unsigned int dispatch_flags)
2154 {
2155         u32 *cs;
2156
2157         cs = intel_ring_begin(rq, 2);
2158         if (IS_ERR(cs))
2159                 return PTR_ERR(cs);
2160
2161         *cs++ = MI_BATCH_BUFFER_START | (dispatch_flags & I915_DISPATCH_SECURE ?
2162                 0 : MI_BATCH_NON_SECURE_I965);
2163         /* bit0-7 is the length on GEN6+ */
2164         *cs++ = offset;
2165         intel_ring_advance(rq, cs);
2166
2167         return 0;
2168 }
2169
2170 /* Blitter support (SandyBridge+) */
2171
2172 static int gen6_ring_flush(struct i915_request *rq, u32 mode)
2173 {
2174         return gen6_flush_dw(rq, mode, MI_INVALIDATE_TLB);
2175 }
2176
2177 static void intel_ring_init_irq(struct drm_i915_private *dev_priv,
2178                                 struct intel_engine_cs *engine)
2179 {
2180         if (INTEL_GEN(dev_priv) >= 6) {
2181                 engine->irq_enable = gen6_irq_enable;
2182                 engine->irq_disable = gen6_irq_disable;
2183         } else if (INTEL_GEN(dev_priv) >= 5) {
2184                 engine->irq_enable = gen5_irq_enable;
2185                 engine->irq_disable = gen5_irq_disable;
2186         } else if (INTEL_GEN(dev_priv) >= 3) {
2187                 engine->irq_enable = i9xx_irq_enable;
2188                 engine->irq_disable = i9xx_irq_disable;
2189         } else {
2190                 engine->irq_enable = i8xx_irq_enable;
2191                 engine->irq_disable = i8xx_irq_disable;
2192         }
2193 }
2194
2195 static void i9xx_set_default_submission(struct intel_engine_cs *engine)
2196 {
2197         engine->submit_request = i9xx_submit_request;
2198         engine->cancel_requests = cancel_requests;
2199
2200         engine->park = NULL;
2201         engine->unpark = NULL;
2202 }
2203
2204 static void gen6_bsd_set_default_submission(struct intel_engine_cs *engine)
2205 {
2206         i9xx_set_default_submission(engine);
2207         engine->submit_request = gen6_bsd_submit_request;
2208 }
2209
2210 static void intel_ring_default_vfuncs(struct drm_i915_private *dev_priv,
2211                                       struct intel_engine_cs *engine)
2212 {
2213         /* gen8+ are only supported with execlists */
2214         GEM_BUG_ON(INTEL_GEN(dev_priv) >= 8);
2215
2216         intel_ring_init_irq(dev_priv, engine);
2217
2218         engine->init_hw = init_ring_common;
2219         engine->reset.prepare = reset_prepare;
2220         engine->reset.reset = reset_ring;
2221         engine->reset.finish = reset_finish;
2222
2223         engine->cops = &ring_context_ops;
2224         engine->request_alloc = ring_request_alloc;
2225
2226         /*
2227          * Using a global execution timeline; the previous final breadcrumb is
2228          * equivalent to our next initial bread so we can elide
2229          * engine->emit_init_breadcrumb().
2230          */
2231         engine->emit_fini_breadcrumb = i9xx_emit_breadcrumb;
2232         if (IS_GEN(dev_priv, 5))
2233                 engine->emit_fini_breadcrumb = gen5_emit_breadcrumb;
2234
2235         engine->set_default_submission = i9xx_set_default_submission;
2236
2237         if (INTEL_GEN(dev_priv) >= 6)
2238                 engine->emit_bb_start = gen6_emit_bb_start;
2239         else if (INTEL_GEN(dev_priv) >= 4)
2240                 engine->emit_bb_start = i965_emit_bb_start;
2241         else if (IS_I830(dev_priv) || IS_I845G(dev_priv))
2242                 engine->emit_bb_start = i830_emit_bb_start;
2243         else
2244                 engine->emit_bb_start = i915_emit_bb_start;
2245 }
2246
2247 int intel_init_render_ring_buffer(struct intel_engine_cs *engine)
2248 {
2249         struct drm_i915_private *dev_priv = engine->i915;
2250         int ret;
2251
2252         intel_ring_default_vfuncs(dev_priv, engine);
2253
2254         if (HAS_L3_DPF(dev_priv))
2255                 engine->irq_keep_mask = GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
2256
2257         engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
2258
2259         if (INTEL_GEN(dev_priv) >= 7) {
2260                 engine->init_context = intel_rcs_ctx_init;
2261                 engine->emit_flush = gen7_render_ring_flush;
2262                 engine->emit_fini_breadcrumb = gen7_rcs_emit_breadcrumb;
2263         } else if (IS_GEN(dev_priv, 6)) {
2264                 engine->init_context = intel_rcs_ctx_init;
2265                 engine->emit_flush = gen6_render_ring_flush;
2266                 engine->emit_fini_breadcrumb = gen6_rcs_emit_breadcrumb;
2267         } else if (IS_GEN(dev_priv, 5)) {
2268                 engine->emit_flush = gen4_render_ring_flush;
2269         } else {
2270                 if (INTEL_GEN(dev_priv) < 4)
2271                         engine->emit_flush = gen2_render_ring_flush;
2272                 else
2273                         engine->emit_flush = gen4_render_ring_flush;
2274                 engine->irq_enable_mask = I915_USER_INTERRUPT;
2275         }
2276
2277         if (IS_HASWELL(dev_priv))
2278                 engine->emit_bb_start = hsw_emit_bb_start;
2279
2280         engine->init_hw = init_render_ring;
2281
2282         ret = intel_init_ring_buffer(engine);
2283         if (ret)
2284                 return ret;
2285
2286         return 0;
2287 }
2288
2289 int intel_init_bsd_ring_buffer(struct intel_engine_cs *engine)
2290 {
2291         struct drm_i915_private *dev_priv = engine->i915;
2292
2293         intel_ring_default_vfuncs(dev_priv, engine);
2294
2295         if (INTEL_GEN(dev_priv) >= 6) {
2296                 /* gen6 bsd needs a special wa for tail updates */
2297                 if (IS_GEN(dev_priv, 6))
2298                         engine->set_default_submission = gen6_bsd_set_default_submission;
2299                 engine->emit_flush = gen6_bsd_ring_flush;
2300                 engine->irq_enable_mask = GT_BSD_USER_INTERRUPT;
2301
2302                 if (IS_GEN(dev_priv, 6))
2303                         engine->emit_fini_breadcrumb = gen6_xcs_emit_breadcrumb;
2304                 else
2305                         engine->emit_fini_breadcrumb = gen7_xcs_emit_breadcrumb;
2306         } else {
2307                 engine->emit_flush = bsd_ring_flush;
2308                 if (IS_GEN(dev_priv, 5))
2309                         engine->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
2310                 else
2311                         engine->irq_enable_mask = I915_BSD_USER_INTERRUPT;
2312         }
2313
2314         return intel_init_ring_buffer(engine);
2315 }
2316
2317 int intel_init_blt_ring_buffer(struct intel_engine_cs *engine)
2318 {
2319         struct drm_i915_private *dev_priv = engine->i915;
2320
2321         GEM_BUG_ON(INTEL_GEN(dev_priv) < 6);
2322
2323         intel_ring_default_vfuncs(dev_priv, engine);
2324
2325         engine->emit_flush = gen6_ring_flush;
2326         engine->irq_enable_mask = GT_BLT_USER_INTERRUPT;
2327
2328         if (IS_GEN(dev_priv, 6))
2329                 engine->emit_fini_breadcrumb = gen6_xcs_emit_breadcrumb;
2330         else
2331                 engine->emit_fini_breadcrumb = gen7_xcs_emit_breadcrumb;
2332
2333         return intel_init_ring_buffer(engine);
2334 }
2335
2336 int intel_init_vebox_ring_buffer(struct intel_engine_cs *engine)
2337 {
2338         struct drm_i915_private *dev_priv = engine->i915;
2339
2340         GEM_BUG_ON(INTEL_GEN(dev_priv) < 7);
2341
2342         intel_ring_default_vfuncs(dev_priv, engine);
2343
2344         engine->emit_flush = gen6_ring_flush;
2345         engine->irq_enable_mask = PM_VEBOX_USER_INTERRUPT;
2346         engine->irq_enable = hsw_vebox_irq_enable;
2347         engine->irq_disable = hsw_vebox_irq_disable;
2348
2349         engine->emit_fini_breadcrumb = gen7_xcs_emit_breadcrumb;
2350
2351         return intel_init_ring_buffer(engine);
2352 }