2 * Copyright © 2008-2010 Intel Corporation
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:
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
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
24 * Eric Anholt <eric@anholt.net>
25 * Zou Nan hai <nanhai.zou@intel.com>
26 * Xiang Hai hao<haihao.xiang@intel.com>
30 #include <linux/log2.h>
32 #include <drm/i915_drm.h>
35 #include "i915_gem_render_state.h"
36 #include "i915_trace.h"
37 #include "intel_reset.h"
38 #include "intel_workarounds.h"
40 /* Rough estimate of the typical request size, performing a flush,
41 * set-context and then emitting the batch.
43 #define LEGACY_REQUEST_SIZE 200
45 unsigned int intel_ring_update_space(struct intel_ring *ring)
49 space = __intel_ring_space(ring->head, ring->emit, ring->size);
56 gen2_render_ring_flush(struct i915_request *rq, u32 mode)
58 unsigned int num_store_dw;
63 if (mode & EMIT_INVALIDATE)
65 if (mode & EMIT_FLUSH)
68 cs = intel_ring_begin(rq, 2 + 3 * num_store_dw);
73 while (num_store_dw--) {
74 *cs++ = MI_STORE_DWORD_IMM | MI_MEM_VIRTUAL;
75 *cs++ = i915_scratch_offset(rq->i915);
78 *cs++ = MI_FLUSH | MI_NO_WRITE_FLUSH;
80 intel_ring_advance(rq, cs);
86 gen4_render_ring_flush(struct i915_request *rq, u32 mode)
94 * I915_GEM_DOMAIN_RENDER is always invalidated, but is
95 * only flushed if MI_NO_WRITE_FLUSH is unset. On 965, it is
96 * also flushed at 2d versus 3d pipeline switches.
100 * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
101 * MI_READ_FLUSH is set, and is always flushed on 965.
103 * I915_GEM_DOMAIN_COMMAND may not exist?
105 * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
106 * invalidated when MI_EXE_FLUSH is set.
108 * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
109 * invalidated with every MI_FLUSH.
113 * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
114 * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
115 * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
116 * are flushed at any MI_FLUSH.
120 if (mode & EMIT_INVALIDATE) {
122 if (IS_G4X(rq->i915) || IS_GEN(rq->i915, 5))
123 cmd |= MI_INVALIDATE_ISP;
127 if (mode & EMIT_INVALIDATE)
130 cs = intel_ring_begin(rq, i);
137 * A random delay to let the CS invalidate take effect? Without this
138 * delay, the GPU relocation path fails as the CS does not see
139 * the updated contents. Just as important, if we apply the flushes
140 * to the EMIT_FLUSH branch (i.e. immediately after the relocation
141 * write and before the invalidate on the next batch), the relocations
142 * still fail. This implies that is a delay following invalidation
143 * that is required to reset the caches as opposed to a delay to
144 * ensure the memory is written.
146 if (mode & EMIT_INVALIDATE) {
147 *cs++ = GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE;
148 *cs++ = i915_scratch_offset(rq->i915) | PIPE_CONTROL_GLOBAL_GTT;
152 for (i = 0; i < 12; i++)
155 *cs++ = GFX_OP_PIPE_CONTROL(4) | PIPE_CONTROL_QW_WRITE;
156 *cs++ = i915_scratch_offset(rq->i915) | PIPE_CONTROL_GLOBAL_GTT;
163 intel_ring_advance(rq, cs);
169 * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
170 * implementing two workarounds on gen6. From section 1.4.7.1
171 * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
173 * [DevSNB-C+{W/A}] Before any depth stall flush (including those
174 * produced by non-pipelined state commands), software needs to first
175 * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
178 * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
179 * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
181 * And the workaround for these two requires this workaround first:
183 * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
184 * BEFORE the pipe-control with a post-sync op and no write-cache
187 * And this last workaround is tricky because of the requirements on
188 * that bit. From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
191 * "1 of the following must also be set:
192 * - Render Target Cache Flush Enable ([12] of DW1)
193 * - Depth Cache Flush Enable ([0] of DW1)
194 * - Stall at Pixel Scoreboard ([1] of DW1)
195 * - Depth Stall ([13] of DW1)
196 * - Post-Sync Operation ([13] of DW1)
197 * - Notify Enable ([8] of DW1)"
199 * The cache flushes require the workaround flush that triggered this
200 * one, so we can't use it. Depth stall would trigger the same.
201 * Post-sync nonzero is what triggered this second workaround, so we
202 * can't use that one either. Notify enable is IRQs, which aren't
203 * really our business. That leaves only stall at scoreboard.
206 gen6_emit_post_sync_nonzero_flush(struct i915_request *rq)
208 u32 scratch_addr = i915_scratch_offset(rq->i915) + 2 * CACHELINE_BYTES;
211 cs = intel_ring_begin(rq, 6);
215 *cs++ = GFX_OP_PIPE_CONTROL(5);
216 *cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
217 *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
218 *cs++ = 0; /* low dword */
219 *cs++ = 0; /* high dword */
221 intel_ring_advance(rq, cs);
223 cs = intel_ring_begin(rq, 6);
227 *cs++ = GFX_OP_PIPE_CONTROL(5);
228 *cs++ = PIPE_CONTROL_QW_WRITE;
229 *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
233 intel_ring_advance(rq, cs);
239 gen6_render_ring_flush(struct i915_request *rq, u32 mode)
241 u32 scratch_addr = i915_scratch_offset(rq->i915) + 2 * CACHELINE_BYTES;
245 /* Force SNB workarounds for PIPE_CONTROL flushes */
246 ret = gen6_emit_post_sync_nonzero_flush(rq);
250 /* Just flush everything. Experiments have shown that reducing the
251 * number of bits based on the write domains has little performance
254 if (mode & EMIT_FLUSH) {
255 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
256 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
258 * Ensure that any following seqno writes only happen
259 * when the render cache is indeed flushed.
261 flags |= PIPE_CONTROL_CS_STALL;
263 if (mode & EMIT_INVALIDATE) {
264 flags |= PIPE_CONTROL_TLB_INVALIDATE;
265 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
266 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
267 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
268 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
269 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
271 * TLB invalidate requires a post-sync write.
273 flags |= PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_CS_STALL;
276 cs = intel_ring_begin(rq, 4);
280 *cs++ = GFX_OP_PIPE_CONTROL(4);
282 *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
284 intel_ring_advance(rq, cs);
289 static u32 *gen6_rcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
291 /* First we do the gen6_emit_post_sync_nonzero_flush w/a */
292 *cs++ = GFX_OP_PIPE_CONTROL(4);
293 *cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
297 *cs++ = GFX_OP_PIPE_CONTROL(4);
298 *cs++ = PIPE_CONTROL_QW_WRITE;
299 *cs++ = i915_scratch_offset(rq->i915) | PIPE_CONTROL_GLOBAL_GTT;
302 /* Finally we can flush and with it emit the breadcrumb */
303 *cs++ = GFX_OP_PIPE_CONTROL(4);
304 *cs++ = (PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
305 PIPE_CONTROL_DEPTH_CACHE_FLUSH |
306 PIPE_CONTROL_DC_FLUSH_ENABLE |
307 PIPE_CONTROL_QW_WRITE |
308 PIPE_CONTROL_CS_STALL);
309 *cs++ = rq->timeline->hwsp_offset | PIPE_CONTROL_GLOBAL_GTT;
310 *cs++ = rq->fence.seqno;
312 *cs++ = MI_USER_INTERRUPT;
315 rq->tail = intel_ring_offset(rq, cs);
316 assert_ring_tail_valid(rq->ring, rq->tail);
322 gen7_render_ring_cs_stall_wa(struct i915_request *rq)
326 cs = intel_ring_begin(rq, 4);
330 *cs++ = GFX_OP_PIPE_CONTROL(4);
331 *cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
334 intel_ring_advance(rq, cs);
340 gen7_render_ring_flush(struct i915_request *rq, u32 mode)
342 u32 scratch_addr = i915_scratch_offset(rq->i915) + 2 * CACHELINE_BYTES;
346 * Ensure that any following seqno writes only happen when the render
347 * cache is indeed flushed.
349 * Workaround: 4th PIPE_CONTROL command (except the ones with only
350 * read-cache invalidate bits set) must have the CS_STALL bit set. We
351 * don't try to be clever and just set it unconditionally.
353 flags |= PIPE_CONTROL_CS_STALL;
355 /* Just flush everything. Experiments have shown that reducing the
356 * number of bits based on the write domains has little performance
359 if (mode & EMIT_FLUSH) {
360 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
361 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
362 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
363 flags |= PIPE_CONTROL_FLUSH_ENABLE;
365 if (mode & EMIT_INVALIDATE) {
366 flags |= PIPE_CONTROL_TLB_INVALIDATE;
367 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
368 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
369 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
370 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
371 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
372 flags |= PIPE_CONTROL_MEDIA_STATE_CLEAR;
374 * TLB invalidate requires a post-sync write.
376 flags |= PIPE_CONTROL_QW_WRITE;
377 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
379 flags |= PIPE_CONTROL_STALL_AT_SCOREBOARD;
381 /* Workaround: we must issue a pipe_control with CS-stall bit
382 * set before a pipe_control command that has the state cache
383 * invalidate bit set. */
384 gen7_render_ring_cs_stall_wa(rq);
387 cs = intel_ring_begin(rq, 4);
391 *cs++ = GFX_OP_PIPE_CONTROL(4);
393 *cs++ = scratch_addr;
395 intel_ring_advance(rq, cs);
400 static u32 *gen7_rcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
402 *cs++ = GFX_OP_PIPE_CONTROL(4);
403 *cs++ = (PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH |
404 PIPE_CONTROL_DEPTH_CACHE_FLUSH |
405 PIPE_CONTROL_DC_FLUSH_ENABLE |
406 PIPE_CONTROL_FLUSH_ENABLE |
407 PIPE_CONTROL_QW_WRITE |
408 PIPE_CONTROL_GLOBAL_GTT_IVB |
409 PIPE_CONTROL_CS_STALL);
410 *cs++ = rq->timeline->hwsp_offset;
411 *cs++ = rq->fence.seqno;
413 *cs++ = MI_USER_INTERRUPT;
416 rq->tail = intel_ring_offset(rq, cs);
417 assert_ring_tail_valid(rq->ring, rq->tail);
422 static u32 *gen6_xcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
424 GEM_BUG_ON(rq->timeline->hwsp_ggtt != rq->engine->status_page.vma);
425 GEM_BUG_ON(offset_in_page(rq->timeline->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
427 *cs++ = MI_FLUSH_DW | MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_DW_STORE_INDEX;
428 *cs++ = I915_GEM_HWS_SEQNO_ADDR | MI_FLUSH_DW_USE_GTT;
429 *cs++ = rq->fence.seqno;
431 *cs++ = MI_USER_INTERRUPT;
433 rq->tail = intel_ring_offset(rq, cs);
434 assert_ring_tail_valid(rq->ring, rq->tail);
439 #define GEN7_XCS_WA 32
440 static u32 *gen7_xcs_emit_breadcrumb(struct i915_request *rq, u32 *cs)
444 GEM_BUG_ON(rq->timeline->hwsp_ggtt != rq->engine->status_page.vma);
445 GEM_BUG_ON(offset_in_page(rq->timeline->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
447 *cs++ = MI_FLUSH_DW | MI_FLUSH_DW_OP_STOREDW | MI_FLUSH_DW_STORE_INDEX;
448 *cs++ = I915_GEM_HWS_SEQNO_ADDR | MI_FLUSH_DW_USE_GTT;
449 *cs++ = rq->fence.seqno;
451 for (i = 0; i < GEN7_XCS_WA; i++) {
452 *cs++ = MI_STORE_DWORD_INDEX;
453 *cs++ = I915_GEM_HWS_SEQNO_ADDR;
454 *cs++ = rq->fence.seqno;
461 *cs++ = MI_USER_INTERRUPT;
464 rq->tail = intel_ring_offset(rq, cs);
465 assert_ring_tail_valid(rq->ring, rq->tail);
471 static void set_hwstam(struct intel_engine_cs *engine, u32 mask)
474 * Keep the render interrupt unmasked as this papers over
475 * lost interrupts following a reset.
477 if (engine->class == RENDER_CLASS) {
478 if (INTEL_GEN(engine->i915) >= 6)
481 mask &= ~I915_USER_INTERRUPT;
484 intel_engine_set_hwsp_writemask(engine, mask);
487 static void set_hws_pga(struct intel_engine_cs *engine, phys_addr_t phys)
489 struct drm_i915_private *dev_priv = engine->i915;
492 addr = lower_32_bits(phys);
493 if (INTEL_GEN(dev_priv) >= 4)
494 addr |= (phys >> 28) & 0xf0;
496 I915_WRITE(HWS_PGA, addr);
499 static struct page *status_page(struct intel_engine_cs *engine)
501 struct drm_i915_gem_object *obj = engine->status_page.vma->obj;
503 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
504 return sg_page(obj->mm.pages->sgl);
507 static void ring_setup_phys_status_page(struct intel_engine_cs *engine)
509 set_hws_pga(engine, PFN_PHYS(page_to_pfn(status_page(engine))));
510 set_hwstam(engine, ~0u);
513 static void set_hwsp(struct intel_engine_cs *engine, u32 offset)
515 struct drm_i915_private *dev_priv = engine->i915;
519 * The ring status page addresses are no longer next to the rest of
520 * the ring registers as of gen7.
522 if (IS_GEN(dev_priv, 7)) {
523 switch (engine->id) {
525 * No more rings exist on Gen7. Default case is only to shut up
526 * gcc switch check warning.
529 GEM_BUG_ON(engine->id);
532 hwsp = RENDER_HWS_PGA_GEN7;
535 hwsp = BLT_HWS_PGA_GEN7;
538 hwsp = BSD_HWS_PGA_GEN7;
541 hwsp = VEBOX_HWS_PGA_GEN7;
544 } else if (IS_GEN(dev_priv, 6)) {
545 hwsp = RING_HWS_PGA_GEN6(engine->mmio_base);
547 hwsp = RING_HWS_PGA(engine->mmio_base);
550 I915_WRITE(hwsp, offset);
554 static void flush_cs_tlb(struct intel_engine_cs *engine)
556 struct drm_i915_private *dev_priv = engine->i915;
558 if (!IS_GEN_RANGE(dev_priv, 6, 7))
561 /* ring should be idle before issuing a sync flush*/
562 WARN_ON((ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
564 ENGINE_WRITE(engine, RING_INSTPM,
565 _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE |
567 if (intel_wait_for_register(engine->uncore,
568 RING_INSTPM(engine->mmio_base),
569 INSTPM_SYNC_FLUSH, 0,
571 DRM_ERROR("%s: wait for SyncFlush to complete for TLB invalidation timed out\n",
575 static void ring_setup_status_page(struct intel_engine_cs *engine)
577 set_hwsp(engine, i915_ggtt_offset(engine->status_page.vma));
578 set_hwstam(engine, ~0u);
580 flush_cs_tlb(engine);
583 static bool stop_ring(struct intel_engine_cs *engine)
585 struct drm_i915_private *dev_priv = engine->i915;
587 if (INTEL_GEN(dev_priv) > 2) {
589 RING_MI_MODE, _MASKED_BIT_ENABLE(STOP_RING));
590 if (intel_wait_for_register(engine->uncore,
591 RING_MI_MODE(engine->mmio_base),
595 DRM_ERROR("%s : timed out trying to stop ring\n",
599 * Sometimes we observe that the idle flag is not
600 * set even though the ring is empty. So double
601 * check before giving up.
603 if (ENGINE_READ(engine, RING_HEAD) !=
604 ENGINE_READ(engine, RING_TAIL))
609 ENGINE_WRITE(engine, RING_HEAD, ENGINE_READ(engine, RING_TAIL));
611 ENGINE_WRITE(engine, RING_HEAD, 0);
612 ENGINE_WRITE(engine, RING_TAIL, 0);
614 /* The ring must be empty before it is disabled */
615 ENGINE_WRITE(engine, RING_CTL, 0);
617 return (ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) == 0;
620 static int xcs_resume(struct intel_engine_cs *engine)
622 struct drm_i915_private *dev_priv = engine->i915;
623 struct intel_ring *ring = engine->buffer;
626 GEM_TRACE("%s: ring:{HEAD:%04x, TAIL:%04x}\n",
627 engine->name, ring->head, ring->tail);
629 intel_uncore_forcewake_get(engine->uncore, FORCEWAKE_ALL);
631 if (!stop_ring(engine)) {
632 /* G45 ring initialization often fails to reset head to zero */
633 DRM_DEBUG_DRIVER("%s head not reset to zero "
634 "ctl %08x head %08x tail %08x start %08x\n",
636 ENGINE_READ(engine, RING_CTL),
637 ENGINE_READ(engine, RING_HEAD),
638 ENGINE_READ(engine, RING_TAIL),
639 ENGINE_READ(engine, RING_START));
641 if (!stop_ring(engine)) {
642 DRM_ERROR("failed to set %s head to zero "
643 "ctl %08x head %08x tail %08x start %08x\n",
645 ENGINE_READ(engine, RING_CTL),
646 ENGINE_READ(engine, RING_HEAD),
647 ENGINE_READ(engine, RING_TAIL),
648 ENGINE_READ(engine, RING_START));
654 if (HWS_NEEDS_PHYSICAL(dev_priv))
655 ring_setup_phys_status_page(engine);
657 ring_setup_status_page(engine);
659 intel_engine_reset_breadcrumbs(engine);
661 /* Enforce ordering by reading HEAD register back */
662 ENGINE_READ(engine, RING_HEAD);
664 /* Initialize the ring. This must happen _after_ we've cleared the ring
665 * registers with the above sequence (the readback of the HEAD registers
666 * also enforces ordering), otherwise the hw might lose the new ring
667 * register values. */
668 ENGINE_WRITE(engine, RING_START, i915_ggtt_offset(ring->vma));
670 /* WaClearRingBufHeadRegAtInit:ctg,elk */
671 if (ENGINE_READ(engine, RING_HEAD))
672 DRM_DEBUG_DRIVER("%s initialization failed [head=%08x], fudging\n",
673 engine->name, ENGINE_READ(engine, RING_HEAD));
675 /* Check that the ring offsets point within the ring! */
676 GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->head));
677 GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->tail));
678 intel_ring_update_space(ring);
680 /* First wake the ring up to an empty/idle ring */
681 ENGINE_WRITE(engine, RING_HEAD, ring->head);
682 ENGINE_WRITE(engine, RING_TAIL, ring->head);
683 ENGINE_POSTING_READ(engine, RING_TAIL);
685 ENGINE_WRITE(engine, RING_CTL, RING_CTL_SIZE(ring->size) | RING_VALID);
687 /* If the head is still not zero, the ring is dead */
688 if (intel_wait_for_register(engine->uncore,
689 RING_CTL(engine->mmio_base),
690 RING_VALID, RING_VALID,
692 DRM_ERROR("%s initialization failed "
693 "ctl %08x (valid? %d) head %08x [%08x] tail %08x [%08x] start %08x [expected %08x]\n",
695 ENGINE_READ(engine, RING_CTL),
696 ENGINE_READ(engine, RING_CTL) & RING_VALID,
697 ENGINE_READ(engine, RING_HEAD), ring->head,
698 ENGINE_READ(engine, RING_TAIL), ring->tail,
699 ENGINE_READ(engine, RING_START),
700 i915_ggtt_offset(ring->vma));
705 if (INTEL_GEN(dev_priv) > 2)
707 RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
709 /* Now awake, let it get started */
710 if (ring->tail != ring->head) {
711 ENGINE_WRITE(engine, RING_TAIL, ring->tail);
712 ENGINE_POSTING_READ(engine, RING_TAIL);
715 /* Papering over lost _interrupts_ immediately following the restart */
716 intel_engine_queue_breadcrumbs(engine);
718 intel_uncore_forcewake_put(engine->uncore, FORCEWAKE_ALL);
723 static void reset_prepare(struct intel_engine_cs *engine)
725 intel_engine_stop_cs(engine);
728 static void reset_ring(struct intel_engine_cs *engine, bool stalled)
730 struct i915_timeline *tl = &engine->timeline;
731 struct i915_request *pos, *rq;
736 spin_lock_irqsave(&tl->lock, flags);
737 list_for_each_entry(pos, &tl->requests, link) {
738 if (!i915_request_completed(pos)) {
745 * The guilty request will get skipped on a hung engine.
747 * Users of client default contexts do not rely on logical
748 * state preserved between batches so it is safe to execute
749 * queued requests following the hang. Non default contexts
750 * rely on preserved state, so skipping a batch loses the
751 * evolution of the state and it needs to be considered corrupted.
752 * Executing more queued batches on top of corrupted state is
753 * risky. But we take the risk by trying to advance through
754 * the queued requests in order to make the client behaviour
755 * more predictable around resets, by not throwing away random
756 * amount of batches it has prepared for execution. Sophisticated
757 * clients can use gem_reset_stats_ioctl and dma fence status
758 * (exported via sync_file info ioctl on explicit fences) to observe
759 * when it loses the context state and should rebuild accordingly.
761 * The context ban, and ultimately the client ban, mechanism are safety
762 * valves if client submission ends up resulting in nothing more than
768 * Try to restore the logical GPU state to match the
769 * continuation of the request queue. If we skip the
770 * context/PD restore, then the next request may try to execute
771 * assuming that its context is valid and loaded on the GPU and
772 * so may try to access invalid memory, prompting repeated GPU
775 * If the request was guilty, we still restore the logical
776 * state in case the next request requires it (e.g. the
777 * aliasing ppgtt), but skip over the hung batch.
779 * If the request was innocent, we try to replay the request
780 * with the restored context.
782 i915_reset_request(rq, stalled);
784 GEM_BUG_ON(rq->ring != engine->buffer);
787 head = engine->buffer->tail;
789 engine->buffer->head = intel_ring_wrap(engine->buffer, head);
791 spin_unlock_irqrestore(&tl->lock, flags);
794 static void reset_finish(struct intel_engine_cs *engine)
798 static int intel_rcs_ctx_init(struct i915_request *rq)
802 ret = intel_engine_emit_ctx_wa(rq);
806 ret = i915_gem_render_state_emit(rq);
813 static int rcs_resume(struct intel_engine_cs *engine)
815 struct drm_i915_private *dev_priv = engine->i915;
818 * Disable CONSTANT_BUFFER before it is loaded from the context
819 * image. For as it is loaded, it is executed and the stored
820 * address may no longer be valid, leading to a GPU hang.
822 * This imposes the requirement that userspace reload their
823 * CONSTANT_BUFFER on every batch, fortunately a requirement
824 * they are already accustomed to from before contexts were
827 if (IS_GEN(dev_priv, 4))
829 _MASKED_BIT_ENABLE(ECO_CONSTANT_BUFFER_SR_DISABLE));
831 /* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
832 if (IS_GEN_RANGE(dev_priv, 4, 6))
833 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH));
835 /* We need to disable the AsyncFlip performance optimisations in order
836 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
837 * programmed to '1' on all products.
839 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv
841 if (IS_GEN_RANGE(dev_priv, 6, 7))
842 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
844 /* Required for the hardware to program scanline values for waiting */
845 /* WaEnableFlushTlbInvalidationMode:snb */
846 if (IS_GEN(dev_priv, 6))
848 _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT));
850 /* WaBCSVCSTlbInvalidationMode:ivb,vlv,hsw */
851 if (IS_GEN(dev_priv, 7))
852 I915_WRITE(GFX_MODE_GEN7,
853 _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT) |
854 _MASKED_BIT_ENABLE(GFX_REPLAY_MODE));
856 if (IS_GEN(dev_priv, 6)) {
857 /* From the Sandybridge PRM, volume 1 part 3, page 24:
858 * "If this bit is set, STCunit will have LRA as replacement
859 * policy. [...] This bit must be reset. LRA replacement
860 * policy is not supported."
862 I915_WRITE(CACHE_MODE_0,
863 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
866 if (IS_GEN_RANGE(dev_priv, 6, 7))
867 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
869 return xcs_resume(engine);
872 static void cancel_requests(struct intel_engine_cs *engine)
874 struct i915_request *request;
877 spin_lock_irqsave(&engine->timeline.lock, flags);
879 /* Mark all submitted requests as skipped. */
880 list_for_each_entry(request, &engine->timeline.requests, link) {
881 if (!i915_request_signaled(request))
882 dma_fence_set_error(&request->fence, -EIO);
884 i915_request_mark_complete(request);
887 /* Remaining _unready_ requests will be nop'ed when submitted */
889 spin_unlock_irqrestore(&engine->timeline.lock, flags);
892 static void i9xx_submit_request(struct i915_request *request)
894 i915_request_submit(request);
896 ENGINE_WRITE(request->engine, RING_TAIL,
897 intel_ring_set_tail(request->ring, request->tail));
900 static u32 *i9xx_emit_breadcrumb(struct i915_request *rq, u32 *cs)
902 GEM_BUG_ON(rq->timeline->hwsp_ggtt != rq->engine->status_page.vma);
903 GEM_BUG_ON(offset_in_page(rq->timeline->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
907 *cs++ = MI_STORE_DWORD_INDEX;
908 *cs++ = I915_GEM_HWS_SEQNO_ADDR;
909 *cs++ = rq->fence.seqno;
911 *cs++ = MI_USER_INTERRUPT;
914 rq->tail = intel_ring_offset(rq, cs);
915 assert_ring_tail_valid(rq->ring, rq->tail);
920 #define GEN5_WA_STORES 8 /* must be at least 1! */
921 static u32 *gen5_emit_breadcrumb(struct i915_request *rq, u32 *cs)
925 GEM_BUG_ON(rq->timeline->hwsp_ggtt != rq->engine->status_page.vma);
926 GEM_BUG_ON(offset_in_page(rq->timeline->hwsp_offset) != I915_GEM_HWS_SEQNO_ADDR);
930 BUILD_BUG_ON(GEN5_WA_STORES < 1);
931 for (i = 0; i < GEN5_WA_STORES; i++) {
932 *cs++ = MI_STORE_DWORD_INDEX;
933 *cs++ = I915_GEM_HWS_SEQNO_ADDR;
934 *cs++ = rq->fence.seqno;
937 *cs++ = MI_USER_INTERRUPT;
939 rq->tail = intel_ring_offset(rq, cs);
940 assert_ring_tail_valid(rq->ring, rq->tail);
944 #undef GEN5_WA_STORES
947 gen5_irq_enable(struct intel_engine_cs *engine)
949 gen5_enable_gt_irq(engine->i915, engine->irq_enable_mask);
953 gen5_irq_disable(struct intel_engine_cs *engine)
955 gen5_disable_gt_irq(engine->i915, engine->irq_enable_mask);
959 i9xx_irq_enable(struct intel_engine_cs *engine)
961 engine->i915->irq_mask &= ~engine->irq_enable_mask;
962 intel_uncore_write(engine->uncore, GEN2_IMR, engine->i915->irq_mask);
963 intel_uncore_posting_read_fw(engine->uncore, GEN2_IMR);
967 i9xx_irq_disable(struct intel_engine_cs *engine)
969 engine->i915->irq_mask |= engine->irq_enable_mask;
970 intel_uncore_write(engine->uncore, GEN2_IMR, engine->i915->irq_mask);
974 i8xx_irq_enable(struct intel_engine_cs *engine)
976 struct drm_i915_private *dev_priv = engine->i915;
978 dev_priv->irq_mask &= ~engine->irq_enable_mask;
979 I915_WRITE16(GEN2_IMR, dev_priv->irq_mask);
980 POSTING_READ16(RING_IMR(engine->mmio_base));
984 i8xx_irq_disable(struct intel_engine_cs *engine)
986 struct drm_i915_private *dev_priv = engine->i915;
988 dev_priv->irq_mask |= engine->irq_enable_mask;
989 I915_WRITE16(GEN2_IMR, dev_priv->irq_mask);
993 bsd_ring_flush(struct i915_request *rq, u32 mode)
997 cs = intel_ring_begin(rq, 2);
1003 intel_ring_advance(rq, cs);
1008 gen6_irq_enable(struct intel_engine_cs *engine)
1010 ENGINE_WRITE(engine, RING_IMR,
1011 ~(engine->irq_enable_mask | engine->irq_keep_mask));
1013 /* Flush/delay to ensure the RING_IMR is active before the GT IMR */
1014 ENGINE_POSTING_READ(engine, RING_IMR);
1016 gen5_enable_gt_irq(engine->i915, engine->irq_enable_mask);
1020 gen6_irq_disable(struct intel_engine_cs *engine)
1022 ENGINE_WRITE(engine, RING_IMR, ~engine->irq_keep_mask);
1023 gen5_disable_gt_irq(engine->i915, engine->irq_enable_mask);
1027 hsw_vebox_irq_enable(struct intel_engine_cs *engine)
1029 ENGINE_WRITE(engine, RING_IMR, ~engine->irq_enable_mask);
1031 /* Flush/delay to ensure the RING_IMR is active before the GT IMR */
1032 ENGINE_POSTING_READ(engine, RING_IMR);
1034 gen6_unmask_pm_irq(engine->i915, engine->irq_enable_mask);
1038 hsw_vebox_irq_disable(struct intel_engine_cs *engine)
1040 ENGINE_WRITE(engine, RING_IMR, ~0);
1041 gen6_mask_pm_irq(engine->i915, engine->irq_enable_mask);
1045 i965_emit_bb_start(struct i915_request *rq,
1046 u64 offset, u32 length,
1047 unsigned int dispatch_flags)
1051 cs = intel_ring_begin(rq, 2);
1055 *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT | (dispatch_flags &
1056 I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE_I965);
1058 intel_ring_advance(rq, cs);
1063 /* Just userspace ABI convention to limit the wa batch bo to a resonable size */
1064 #define I830_BATCH_LIMIT SZ_256K
1065 #define I830_TLB_ENTRIES (2)
1066 #define I830_WA_SIZE max(I830_TLB_ENTRIES*4096, I830_BATCH_LIMIT)
1068 i830_emit_bb_start(struct i915_request *rq,
1069 u64 offset, u32 len,
1070 unsigned int dispatch_flags)
1072 u32 *cs, cs_offset = i915_scratch_offset(rq->i915);
1074 GEM_BUG_ON(rq->i915->gt.scratch->size < I830_WA_SIZE);
1076 cs = intel_ring_begin(rq, 6);
1080 /* Evict the invalid PTE TLBs */
1081 *cs++ = COLOR_BLT_CMD | BLT_WRITE_RGBA;
1082 *cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | 4096;
1083 *cs++ = I830_TLB_ENTRIES << 16 | 4; /* load each page */
1087 intel_ring_advance(rq, cs);
1089 if ((dispatch_flags & I915_DISPATCH_PINNED) == 0) {
1090 if (len > I830_BATCH_LIMIT)
1093 cs = intel_ring_begin(rq, 6 + 2);
1097 /* Blit the batch (which has now all relocs applied) to the
1098 * stable batch scratch bo area (so that the CS never
1099 * stumbles over its tlb invalidation bug) ...
1101 *cs++ = SRC_COPY_BLT_CMD | BLT_WRITE_RGBA;
1102 *cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | 4096;
1103 *cs++ = DIV_ROUND_UP(len, 4096) << 16 | 4096;
1110 intel_ring_advance(rq, cs);
1112 /* ... and execute it. */
1116 cs = intel_ring_begin(rq, 2);
1120 *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
1121 *cs++ = offset | (dispatch_flags & I915_DISPATCH_SECURE ? 0 :
1122 MI_BATCH_NON_SECURE);
1123 intel_ring_advance(rq, cs);
1129 i915_emit_bb_start(struct i915_request *rq,
1130 u64 offset, u32 len,
1131 unsigned int dispatch_flags)
1135 cs = intel_ring_begin(rq, 2);
1139 *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
1140 *cs++ = offset | (dispatch_flags & I915_DISPATCH_SECURE ? 0 :
1141 MI_BATCH_NON_SECURE);
1142 intel_ring_advance(rq, cs);
1147 int intel_ring_pin(struct intel_ring *ring)
1149 struct i915_vma *vma = ring->vma;
1150 enum i915_map_type map = i915_coherent_map_type(vma->vm->i915);
1155 GEM_BUG_ON(ring->vaddr);
1157 ret = i915_timeline_pin(ring->timeline);
1163 /* Ring wraparound at offset 0 sometimes hangs. No idea why. */
1164 flags |= PIN_OFFSET_BIAS | i915_ggtt_pin_bias(vma);
1166 if (vma->obj->stolen)
1167 flags |= PIN_MAPPABLE;
1171 ret = i915_vma_pin(vma, 0, 0, flags);
1173 goto unpin_timeline;
1175 if (i915_vma_is_map_and_fenceable(vma))
1176 addr = (void __force *)i915_vma_pin_iomap(vma);
1178 addr = i915_gem_object_pin_map(vma->obj, map);
1180 ret = PTR_ERR(addr);
1184 vma->obj->pin_global++;
1190 i915_vma_unpin(vma);
1192 i915_timeline_unpin(ring->timeline);
1196 void intel_ring_reset(struct intel_ring *ring, u32 tail)
1198 GEM_BUG_ON(!intel_ring_offset_valid(ring, tail));
1203 intel_ring_update_space(ring);
1206 void intel_ring_unpin(struct intel_ring *ring)
1208 GEM_BUG_ON(!ring->vma);
1209 GEM_BUG_ON(!ring->vaddr);
1211 /* Discard any unused bytes beyond that submitted to hw. */
1212 intel_ring_reset(ring, ring->tail);
1214 if (i915_vma_is_map_and_fenceable(ring->vma))
1215 i915_vma_unpin_iomap(ring->vma);
1217 i915_gem_object_unpin_map(ring->vma->obj);
1220 ring->vma->obj->pin_global--;
1221 i915_vma_unpin(ring->vma);
1223 i915_timeline_unpin(ring->timeline);
1226 static struct i915_vma *
1227 intel_ring_create_vma(struct drm_i915_private *dev_priv, int size)
1229 struct i915_address_space *vm = &dev_priv->ggtt.vm;
1230 struct drm_i915_gem_object *obj;
1231 struct i915_vma *vma;
1233 obj = i915_gem_object_create_stolen(dev_priv, size);
1235 obj = i915_gem_object_create_internal(dev_priv, size);
1237 return ERR_CAST(obj);
1240 * Mark ring buffers as read-only from GPU side (so no stray overwrites)
1241 * if supported by the platform's GGTT.
1243 if (vm->has_read_only)
1244 i915_gem_object_set_readonly(obj);
1246 vma = i915_vma_instance(obj, vm, NULL);
1253 i915_gem_object_put(obj);
1258 intel_engine_create_ring(struct intel_engine_cs *engine,
1259 struct i915_timeline *timeline,
1262 struct intel_ring *ring;
1263 struct i915_vma *vma;
1265 GEM_BUG_ON(!is_power_of_2(size));
1266 GEM_BUG_ON(RING_CTL_SIZE(size) & ~RING_NR_PAGES);
1267 GEM_BUG_ON(timeline == &engine->timeline);
1268 lockdep_assert_held(&engine->i915->drm.struct_mutex);
1270 ring = kzalloc(sizeof(*ring), GFP_KERNEL);
1272 return ERR_PTR(-ENOMEM);
1274 kref_init(&ring->ref);
1275 INIT_LIST_HEAD(&ring->request_list);
1276 ring->timeline = i915_timeline_get(timeline);
1279 /* Workaround an erratum on the i830 which causes a hang if
1280 * the TAIL pointer points to within the last 2 cachelines
1283 ring->effective_size = size;
1284 if (IS_I830(engine->i915) || IS_I845G(engine->i915))
1285 ring->effective_size -= 2 * CACHELINE_BYTES;
1287 intel_ring_update_space(ring);
1289 vma = intel_ring_create_vma(engine->i915, size);
1292 return ERR_CAST(vma);
1299 void intel_ring_free(struct kref *ref)
1301 struct intel_ring *ring = container_of(ref, typeof(*ring), ref);
1302 struct drm_i915_gem_object *obj = ring->vma->obj;
1304 i915_vma_close(ring->vma);
1305 __i915_gem_object_release_unless_active(obj);
1307 i915_timeline_put(ring->timeline);
1311 static void __ring_context_fini(struct intel_context *ce)
1313 GEM_BUG_ON(i915_gem_object_is_active(ce->state->obj));
1314 i915_gem_object_put(ce->state->obj);
1317 static void ring_context_destroy(struct kref *ref)
1319 struct intel_context *ce = container_of(ref, typeof(*ce), ref);
1321 GEM_BUG_ON(intel_context_is_pinned(ce));
1324 __ring_context_fini(ce);
1326 intel_context_free(ce);
1329 static int __context_pin_ppgtt(struct i915_gem_context *ctx)
1331 struct i915_hw_ppgtt *ppgtt;
1334 ppgtt = ctx->ppgtt ?: ctx->i915->mm.aliasing_ppgtt;
1336 err = gen6_ppgtt_pin(ppgtt);
1341 static void __context_unpin_ppgtt(struct i915_gem_context *ctx)
1343 struct i915_hw_ppgtt *ppgtt;
1345 ppgtt = ctx->ppgtt ?: ctx->i915->mm.aliasing_ppgtt;
1347 gen6_ppgtt_unpin(ppgtt);
1350 static int __context_pin(struct intel_context *ce)
1352 struct i915_vma *vma;
1359 err = i915_vma_pin(vma, 0, 0, PIN_GLOBAL | PIN_HIGH);
1364 * And mark is as a globally pinned object to let the shrinker know
1365 * it cannot reclaim the object until we release it.
1367 vma->obj->pin_global++;
1368 vma->obj->mm.dirty = true;
1373 static void __context_unpin(struct intel_context *ce)
1375 struct i915_vma *vma;
1381 vma->obj->pin_global--;
1382 i915_vma_unpin(vma);
1385 static void ring_context_unpin(struct intel_context *ce)
1387 __context_unpin_ppgtt(ce->gem_context);
1388 __context_unpin(ce);
1391 static struct i915_vma *
1392 alloc_context_vma(struct intel_engine_cs *engine)
1394 struct drm_i915_private *i915 = engine->i915;
1395 struct drm_i915_gem_object *obj;
1396 struct i915_vma *vma;
1399 obj = i915_gem_object_create_shmem(i915, engine->context_size);
1401 return ERR_CAST(obj);
1404 * Try to make the context utilize L3 as well as LLC.
1406 * On VLV we don't have L3 controls in the PTEs so we
1407 * shouldn't touch the cache level, especially as that
1408 * would make the object snooped which might have a
1409 * negative performance impact.
1411 * Snooping is required on non-llc platforms in execlist
1412 * mode, but since all GGTT accesses use PAT entry 0 we
1413 * get snooping anyway regardless of cache_level.
1415 * This is only applicable for Ivy Bridge devices since
1416 * later platforms don't have L3 control bits in the PTE.
1418 if (IS_IVYBRIDGE(i915))
1419 i915_gem_object_set_cache_coherency(obj, I915_CACHE_L3_LLC);
1421 if (engine->default_state) {
1422 void *defaults, *vaddr;
1424 vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
1425 if (IS_ERR(vaddr)) {
1426 err = PTR_ERR(vaddr);
1430 defaults = i915_gem_object_pin_map(engine->default_state,
1432 if (IS_ERR(defaults)) {
1433 err = PTR_ERR(defaults);
1437 memcpy(vaddr, defaults, engine->context_size);
1438 i915_gem_object_unpin_map(engine->default_state);
1440 i915_gem_object_flush_map(obj);
1441 i915_gem_object_unpin_map(obj);
1444 vma = i915_vma_instance(obj, &i915->ggtt.vm, NULL);
1453 i915_gem_object_unpin_map(obj);
1455 i915_gem_object_put(obj);
1456 return ERR_PTR(err);
1459 static int ring_context_pin(struct intel_context *ce)
1461 struct intel_engine_cs *engine = ce->engine;
1464 /* One ringbuffer to rule them all */
1465 GEM_BUG_ON(!engine->buffer);
1466 ce->ring = engine->buffer;
1468 if (!ce->state && engine->context_size) {
1469 struct i915_vma *vma;
1471 vma = alloc_context_vma(engine);
1473 return PTR_ERR(vma);
1478 err = __context_pin(ce);
1482 err = __context_pin_ppgtt(ce->gem_context);
1489 __context_unpin(ce);
1493 static void ring_context_reset(struct intel_context *ce)
1495 intel_ring_reset(ce->ring, 0);
1498 static const struct intel_context_ops ring_context_ops = {
1499 .pin = ring_context_pin,
1500 .unpin = ring_context_unpin,
1502 .enter = intel_context_enter_engine,
1503 .exit = intel_context_exit_engine,
1505 .reset = ring_context_reset,
1506 .destroy = ring_context_destroy,
1509 static int load_pd_dir(struct i915_request *rq,
1510 const struct i915_hw_ppgtt *ppgtt)
1512 const struct intel_engine_cs * const engine = rq->engine;
1515 cs = intel_ring_begin(rq, 6);
1519 *cs++ = MI_LOAD_REGISTER_IMM(1);
1520 *cs++ = i915_mmio_reg_offset(RING_PP_DIR_DCLV(engine->mmio_base));
1521 *cs++ = PP_DIR_DCLV_2G;
1523 *cs++ = MI_LOAD_REGISTER_IMM(1);
1524 *cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
1525 *cs++ = ppgtt->pd.base.ggtt_offset << 10;
1527 intel_ring_advance(rq, cs);
1532 static int flush_pd_dir(struct i915_request *rq)
1534 const struct intel_engine_cs * const engine = rq->engine;
1537 cs = intel_ring_begin(rq, 4);
1541 /* Stall until the page table load is complete */
1542 *cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
1543 *cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
1544 *cs++ = i915_scratch_offset(rq->i915);
1547 intel_ring_advance(rq, cs);
1551 static inline int mi_set_context(struct i915_request *rq, u32 flags)
1553 struct drm_i915_private *i915 = rq->i915;
1554 struct intel_engine_cs *engine = rq->engine;
1555 enum intel_engine_id id;
1556 const int num_engines =
1557 IS_HSW_GT1(i915) ? RUNTIME_INFO(i915)->num_engines - 1 : 0;
1558 bool force_restore = false;
1562 flags |= MI_MM_SPACE_GTT;
1563 if (IS_HASWELL(i915))
1564 /* These flags are for resource streamer on HSW+ */
1565 flags |= HSW_MI_RS_SAVE_STATE_EN | HSW_MI_RS_RESTORE_STATE_EN;
1567 /* We need to save the extended state for powersaving modes */
1568 flags |= MI_SAVE_EXT_STATE_EN | MI_RESTORE_EXT_STATE_EN;
1571 if (IS_GEN(i915, 7))
1572 len += 2 + (num_engines ? 4 * num_engines + 6 : 0);
1573 else if (IS_GEN(i915, 5))
1575 if (flags & MI_FORCE_RESTORE) {
1576 GEM_BUG_ON(flags & MI_RESTORE_INHIBIT);
1577 flags &= ~MI_FORCE_RESTORE;
1578 force_restore = true;
1582 cs = intel_ring_begin(rq, len);
1586 /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
1587 if (IS_GEN(i915, 7)) {
1588 *cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
1590 struct intel_engine_cs *signaller;
1592 *cs++ = MI_LOAD_REGISTER_IMM(num_engines);
1593 for_each_engine(signaller, i915, id) {
1594 if (signaller == engine)
1597 *cs++ = i915_mmio_reg_offset(
1598 RING_PSMI_CTL(signaller->mmio_base));
1599 *cs++ = _MASKED_BIT_ENABLE(
1600 GEN6_PSMI_SLEEP_MSG_DISABLE);
1603 } else if (IS_GEN(i915, 5)) {
1605 * This w/a is only listed for pre-production ilk a/b steppings,
1606 * but is also mentioned for programming the powerctx. To be
1607 * safe, just apply the workaround; we do not use SyncFlush so
1608 * this should never take effect and so be a no-op!
1610 *cs++ = MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN;
1613 if (force_restore) {
1615 * The HW doesn't handle being told to restore the current
1616 * context very well. Quite often it likes goes to go off and
1617 * sulk, especially when it is meant to be reloading PP_DIR.
1618 * A very simple fix to force the reload is to simply switch
1619 * away from the current context and back again.
1621 * Note that the kernel_context will contain random state
1622 * following the INHIBIT_RESTORE. We accept this since we
1623 * never use the kernel_context state; it is merely a
1624 * placeholder we use to flush other contexts.
1626 *cs++ = MI_SET_CONTEXT;
1627 *cs++ = i915_ggtt_offset(engine->kernel_context->state) |
1633 *cs++ = MI_SET_CONTEXT;
1634 *cs++ = i915_ggtt_offset(rq->hw_context->state) | flags;
1636 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
1637 * WaMiSetContext_Hang:snb,ivb,vlv
1641 if (IS_GEN(i915, 7)) {
1643 struct intel_engine_cs *signaller;
1644 i915_reg_t last_reg = {}; /* keep gcc quiet */
1646 *cs++ = MI_LOAD_REGISTER_IMM(num_engines);
1647 for_each_engine(signaller, i915, id) {
1648 if (signaller == engine)
1651 last_reg = RING_PSMI_CTL(signaller->mmio_base);
1652 *cs++ = i915_mmio_reg_offset(last_reg);
1653 *cs++ = _MASKED_BIT_DISABLE(
1654 GEN6_PSMI_SLEEP_MSG_DISABLE);
1657 /* Insert a delay before the next switch! */
1658 *cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
1659 *cs++ = i915_mmio_reg_offset(last_reg);
1660 *cs++ = i915_scratch_offset(rq->i915);
1663 *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1664 } else if (IS_GEN(i915, 5)) {
1665 *cs++ = MI_SUSPEND_FLUSH;
1668 intel_ring_advance(rq, cs);
1673 static int remap_l3(struct i915_request *rq, int slice)
1675 u32 *cs, *remap_info = rq->i915->l3_parity.remap_info[slice];
1681 cs = intel_ring_begin(rq, GEN7_L3LOG_SIZE/4 * 2 + 2);
1686 * Note: We do not worry about the concurrent register cacheline hang
1687 * here because no other code should access these registers other than
1688 * at initialization time.
1690 *cs++ = MI_LOAD_REGISTER_IMM(GEN7_L3LOG_SIZE/4);
1691 for (i = 0; i < GEN7_L3LOG_SIZE/4; i++) {
1692 *cs++ = i915_mmio_reg_offset(GEN7_L3LOG(slice, i));
1693 *cs++ = remap_info[i];
1696 intel_ring_advance(rq, cs);
1701 static int switch_context(struct i915_request *rq)
1703 struct intel_engine_cs *engine = rq->engine;
1704 struct i915_gem_context *ctx = rq->gem_context;
1705 struct i915_hw_ppgtt *ppgtt = ctx->ppgtt ?: rq->i915->mm.aliasing_ppgtt;
1706 unsigned int unwind_mm = 0;
1710 GEM_BUG_ON(HAS_EXECLISTS(rq->i915));
1716 * Baytail takes a little more convincing that it really needs
1717 * to reload the PD between contexts. It is not just a little
1718 * longer, as adding more stalls after the load_pd_dir (i.e.
1719 * adding a long loop around flush_pd_dir) is not as effective
1720 * as reloading the PD umpteen times. 32 is derived from
1721 * experimentation (gem_exec_parallel/fds) and has no good
1725 if (engine->id == BCS0 && IS_VALLEYVIEW(engine->i915))
1729 ret = load_pd_dir(rq, ppgtt);
1734 if (ppgtt->pd_dirty_engines & engine->mask) {
1735 unwind_mm = engine->mask;
1736 ppgtt->pd_dirty_engines &= ~unwind_mm;
1737 hw_flags = MI_FORCE_RESTORE;
1741 if (rq->hw_context->state) {
1742 GEM_BUG_ON(engine->id != RCS0);
1745 * The kernel context(s) is treated as pure scratch and is not
1746 * expected to retain any state (as we sacrifice it during
1747 * suspend and on resume it may be corrupted). This is ok,
1748 * as nothing actually executes using the kernel context; it
1749 * is purely used for flushing user contexts.
1751 if (i915_gem_context_is_kernel(ctx))
1752 hw_flags = MI_RESTORE_INHIBIT;
1754 ret = mi_set_context(rq, hw_flags);
1760 ret = engine->emit_flush(rq, EMIT_INVALIDATE);
1764 ret = flush_pd_dir(rq);
1769 * Not only do we need a full barrier (post-sync write) after
1770 * invalidating the TLBs, but we need to wait a little bit
1771 * longer. Whether this is merely delaying us, or the
1772 * subsequent flush is a key part of serialising with the
1773 * post-sync op, this extra pass appears vital before a
1776 ret = engine->emit_flush(rq, EMIT_INVALIDATE);
1780 ret = engine->emit_flush(rq, EMIT_FLUSH);
1785 if (ctx->remap_slice) {
1786 for (i = 0; i < MAX_L3_SLICES; i++) {
1787 if (!(ctx->remap_slice & BIT(i)))
1790 ret = remap_l3(rq, i);
1795 ctx->remap_slice = 0;
1802 ppgtt->pd_dirty_engines |= unwind_mm;
1807 static int ring_request_alloc(struct i915_request *request)
1811 GEM_BUG_ON(!intel_context_is_pinned(request->hw_context));
1812 GEM_BUG_ON(request->timeline->has_initial_breadcrumb);
1815 * Flush enough space to reduce the likelihood of waiting after
1816 * we start building the request - in which case we will just
1817 * have to repeat work.
1819 request->reserved_space += LEGACY_REQUEST_SIZE;
1821 /* Unconditionally invalidate GPU caches and TLBs. */
1822 ret = request->engine->emit_flush(request, EMIT_INVALIDATE);
1826 ret = switch_context(request);
1830 request->reserved_space -= LEGACY_REQUEST_SIZE;
1834 static noinline int wait_for_space(struct intel_ring *ring, unsigned int bytes)
1836 struct i915_request *target;
1839 if (intel_ring_update_space(ring) >= bytes)
1842 GEM_BUG_ON(list_empty(&ring->request_list));
1843 list_for_each_entry(target, &ring->request_list, ring_link) {
1844 /* Would completion of this request free enough space? */
1845 if (bytes <= __intel_ring_space(target->postfix,
1846 ring->emit, ring->size))
1850 if (WARN_ON(&target->ring_link == &ring->request_list))
1853 timeout = i915_request_wait(target,
1854 I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
1855 MAX_SCHEDULE_TIMEOUT);
1859 i915_request_retire_upto(target);
1861 intel_ring_update_space(ring);
1862 GEM_BUG_ON(ring->space < bytes);
1866 u32 *intel_ring_begin(struct i915_request *rq, unsigned int num_dwords)
1868 struct intel_ring *ring = rq->ring;
1869 const unsigned int remain_usable = ring->effective_size - ring->emit;
1870 const unsigned int bytes = num_dwords * sizeof(u32);
1871 unsigned int need_wrap = 0;
1872 unsigned int total_bytes;
1875 /* Packets must be qword aligned. */
1876 GEM_BUG_ON(num_dwords & 1);
1878 total_bytes = bytes + rq->reserved_space;
1879 GEM_BUG_ON(total_bytes > ring->effective_size);
1881 if (unlikely(total_bytes > remain_usable)) {
1882 const int remain_actual = ring->size - ring->emit;
1884 if (bytes > remain_usable) {
1886 * Not enough space for the basic request. So need to
1887 * flush out the remainder and then wait for
1890 total_bytes += remain_actual;
1891 need_wrap = remain_actual | 1;
1894 * The base request will fit but the reserved space
1895 * falls off the end. So we don't need an immediate
1896 * wrap and only need to effectively wait for the
1897 * reserved size from the start of ringbuffer.
1899 total_bytes = rq->reserved_space + remain_actual;
1903 if (unlikely(total_bytes > ring->space)) {
1907 * Space is reserved in the ringbuffer for finalising the
1908 * request, as that cannot be allowed to fail. During request
1909 * finalisation, reserved_space is set to 0 to stop the
1910 * overallocation and the assumption is that then we never need
1911 * to wait (which has the risk of failing with EINTR).
1913 * See also i915_request_alloc() and i915_request_add().
1915 GEM_BUG_ON(!rq->reserved_space);
1917 ret = wait_for_space(ring, total_bytes);
1919 return ERR_PTR(ret);
1922 if (unlikely(need_wrap)) {
1924 GEM_BUG_ON(need_wrap > ring->space);
1925 GEM_BUG_ON(ring->emit + need_wrap > ring->size);
1926 GEM_BUG_ON(!IS_ALIGNED(need_wrap, sizeof(u64)));
1928 /* Fill the tail with MI_NOOP */
1929 memset64(ring->vaddr + ring->emit, 0, need_wrap / sizeof(u64));
1930 ring->space -= need_wrap;
1934 GEM_BUG_ON(ring->emit > ring->size - bytes);
1935 GEM_BUG_ON(ring->space < bytes);
1936 cs = ring->vaddr + ring->emit;
1937 GEM_DEBUG_EXEC(memset32(cs, POISON_INUSE, bytes / sizeof(*cs)));
1938 ring->emit += bytes;
1939 ring->space -= bytes;
1944 /* Align the ring tail to a cacheline boundary */
1945 int intel_ring_cacheline_align(struct i915_request *rq)
1950 num_dwords = (rq->ring->emit & (CACHELINE_BYTES - 1)) / sizeof(u32);
1951 if (num_dwords == 0)
1954 num_dwords = CACHELINE_DWORDS - num_dwords;
1955 GEM_BUG_ON(num_dwords & 1);
1957 cs = intel_ring_begin(rq, num_dwords);
1961 memset64(cs, (u64)MI_NOOP << 32 | MI_NOOP, num_dwords / 2);
1962 intel_ring_advance(rq, cs);
1964 GEM_BUG_ON(rq->ring->emit & (CACHELINE_BYTES - 1));
1968 static void gen6_bsd_submit_request(struct i915_request *request)
1970 struct intel_uncore *uncore = request->engine->uncore;
1972 intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
1974 /* Every tail move must follow the sequence below */
1976 /* Disable notification that the ring is IDLE. The GT
1977 * will then assume that it is busy and bring it out of rc6.
1979 intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
1980 _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1982 /* Clear the context id. Here be magic! */
1983 intel_uncore_write64_fw(uncore, GEN6_BSD_RNCID, 0x0);
1985 /* Wait for the ring not to be idle, i.e. for it to wake up. */
1986 if (__intel_wait_for_register_fw(uncore,
1987 GEN6_BSD_SLEEP_PSMI_CONTROL,
1988 GEN6_BSD_SLEEP_INDICATOR,
1991 DRM_ERROR("timed out waiting for the BSD ring to wake up\n");
1993 /* Now that the ring is fully powered up, update the tail */
1994 i9xx_submit_request(request);
1996 /* Let the ring send IDLE messages to the GT again,
1997 * and so let it sleep to conserve power when idle.
1999 intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
2000 _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
2002 intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
2005 static int mi_flush_dw(struct i915_request *rq, u32 flags)
2009 cs = intel_ring_begin(rq, 4);
2016 * We always require a command barrier so that subsequent
2017 * commands, such as breadcrumb interrupts, are strictly ordered
2018 * wrt the contents of the write cache being flushed to memory
2019 * (and thus being coherent from the CPU).
2021 cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
2024 * Bspec vol 1c.3 - blitter engine command streamer:
2025 * "If ENABLED, all TLBs will be invalidated once the flush
2026 * operation is complete. This bit is only valid when the
2027 * Post-Sync Operation field is a value of 1h or 3h."
2032 *cs++ = I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT;
2036 intel_ring_advance(rq, cs);
2041 static int gen6_flush_dw(struct i915_request *rq, u32 mode, u32 invflags)
2043 return mi_flush_dw(rq, mode & EMIT_INVALIDATE ? invflags : 0);
2046 static int gen6_bsd_ring_flush(struct i915_request *rq, u32 mode)
2048 return gen6_flush_dw(rq, mode, MI_INVALIDATE_TLB | MI_INVALIDATE_BSD);
2052 hsw_emit_bb_start(struct i915_request *rq,
2053 u64 offset, u32 len,
2054 unsigned int dispatch_flags)
2058 cs = intel_ring_begin(rq, 2);
2062 *cs++ = MI_BATCH_BUFFER_START | (dispatch_flags & I915_DISPATCH_SECURE ?
2063 0 : MI_BATCH_PPGTT_HSW | MI_BATCH_NON_SECURE_HSW);
2064 /* bit0-7 is the length on GEN6+ */
2066 intel_ring_advance(rq, cs);
2072 gen6_emit_bb_start(struct i915_request *rq,
2073 u64 offset, u32 len,
2074 unsigned int dispatch_flags)
2078 cs = intel_ring_begin(rq, 2);
2082 *cs++ = MI_BATCH_BUFFER_START | (dispatch_flags & I915_DISPATCH_SECURE ?
2083 0 : MI_BATCH_NON_SECURE_I965);
2084 /* bit0-7 is the length on GEN6+ */
2086 intel_ring_advance(rq, cs);
2091 /* Blitter support (SandyBridge+) */
2093 static int gen6_ring_flush(struct i915_request *rq, u32 mode)
2095 return gen6_flush_dw(rq, mode, MI_INVALIDATE_TLB);
2098 static void i9xx_set_default_submission(struct intel_engine_cs *engine)
2100 engine->submit_request = i9xx_submit_request;
2101 engine->cancel_requests = cancel_requests;
2103 engine->park = NULL;
2104 engine->unpark = NULL;
2107 static void gen6_bsd_set_default_submission(struct intel_engine_cs *engine)
2109 i9xx_set_default_submission(engine);
2110 engine->submit_request = gen6_bsd_submit_request;
2113 static void ring_destroy(struct intel_engine_cs *engine)
2115 struct drm_i915_private *dev_priv = engine->i915;
2117 WARN_ON(INTEL_GEN(dev_priv) > 2 &&
2118 (ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
2120 intel_ring_unpin(engine->buffer);
2121 intel_ring_put(engine->buffer);
2123 intel_engine_cleanup_common(engine);
2127 static void setup_irq(struct intel_engine_cs *engine)
2129 struct drm_i915_private *i915 = engine->i915;
2131 if (INTEL_GEN(i915) >= 6) {
2132 engine->irq_enable = gen6_irq_enable;
2133 engine->irq_disable = gen6_irq_disable;
2134 } else if (INTEL_GEN(i915) >= 5) {
2135 engine->irq_enable = gen5_irq_enable;
2136 engine->irq_disable = gen5_irq_disable;
2137 } else if (INTEL_GEN(i915) >= 3) {
2138 engine->irq_enable = i9xx_irq_enable;
2139 engine->irq_disable = i9xx_irq_disable;
2141 engine->irq_enable = i8xx_irq_enable;
2142 engine->irq_disable = i8xx_irq_disable;
2146 static void setup_common(struct intel_engine_cs *engine)
2148 struct drm_i915_private *i915 = engine->i915;
2150 /* gen8+ are only supported with execlists */
2151 GEM_BUG_ON(INTEL_GEN(i915) >= 8);
2155 engine->destroy = ring_destroy;
2157 engine->resume = xcs_resume;
2158 engine->reset.prepare = reset_prepare;
2159 engine->reset.reset = reset_ring;
2160 engine->reset.finish = reset_finish;
2162 engine->cops = &ring_context_ops;
2163 engine->request_alloc = ring_request_alloc;
2166 * Using a global execution timeline; the previous final breadcrumb is
2167 * equivalent to our next initial bread so we can elide
2168 * engine->emit_init_breadcrumb().
2170 engine->emit_fini_breadcrumb = i9xx_emit_breadcrumb;
2171 if (IS_GEN(i915, 5))
2172 engine->emit_fini_breadcrumb = gen5_emit_breadcrumb;
2174 engine->set_default_submission = i9xx_set_default_submission;
2176 if (INTEL_GEN(i915) >= 6)
2177 engine->emit_bb_start = gen6_emit_bb_start;
2178 else if (INTEL_GEN(i915) >= 4)
2179 engine->emit_bb_start = i965_emit_bb_start;
2180 else if (IS_I830(i915) || IS_I845G(i915))
2181 engine->emit_bb_start = i830_emit_bb_start;
2183 engine->emit_bb_start = i915_emit_bb_start;
2186 static void setup_rcs(struct intel_engine_cs *engine)
2188 struct drm_i915_private *i915 = engine->i915;
2190 if (HAS_L3_DPF(i915))
2191 engine->irq_keep_mask = GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
2193 engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
2195 if (INTEL_GEN(i915) >= 7) {
2196 engine->init_context = intel_rcs_ctx_init;
2197 engine->emit_flush = gen7_render_ring_flush;
2198 engine->emit_fini_breadcrumb = gen7_rcs_emit_breadcrumb;
2199 } else if (IS_GEN(i915, 6)) {
2200 engine->init_context = intel_rcs_ctx_init;
2201 engine->emit_flush = gen6_render_ring_flush;
2202 engine->emit_fini_breadcrumb = gen6_rcs_emit_breadcrumb;
2203 } else if (IS_GEN(i915, 5)) {
2204 engine->emit_flush = gen4_render_ring_flush;
2206 if (INTEL_GEN(i915) < 4)
2207 engine->emit_flush = gen2_render_ring_flush;
2209 engine->emit_flush = gen4_render_ring_flush;
2210 engine->irq_enable_mask = I915_USER_INTERRUPT;
2213 if (IS_HASWELL(i915))
2214 engine->emit_bb_start = hsw_emit_bb_start;
2216 engine->resume = rcs_resume;
2219 static void setup_vcs(struct intel_engine_cs *engine)
2221 struct drm_i915_private *i915 = engine->i915;
2223 if (INTEL_GEN(i915) >= 6) {
2224 /* gen6 bsd needs a special wa for tail updates */
2225 if (IS_GEN(i915, 6))
2226 engine->set_default_submission = gen6_bsd_set_default_submission;
2227 engine->emit_flush = gen6_bsd_ring_flush;
2228 engine->irq_enable_mask = GT_BSD_USER_INTERRUPT;
2230 if (IS_GEN(i915, 6))
2231 engine->emit_fini_breadcrumb = gen6_xcs_emit_breadcrumb;
2233 engine->emit_fini_breadcrumb = gen7_xcs_emit_breadcrumb;
2235 engine->emit_flush = bsd_ring_flush;
2236 if (IS_GEN(i915, 5))
2237 engine->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
2239 engine->irq_enable_mask = I915_BSD_USER_INTERRUPT;
2243 static void setup_bcs(struct intel_engine_cs *engine)
2245 struct drm_i915_private *i915 = engine->i915;
2247 engine->emit_flush = gen6_ring_flush;
2248 engine->irq_enable_mask = GT_BLT_USER_INTERRUPT;
2250 if (IS_GEN(i915, 6))
2251 engine->emit_fini_breadcrumb = gen6_xcs_emit_breadcrumb;
2253 engine->emit_fini_breadcrumb = gen7_xcs_emit_breadcrumb;
2256 static void setup_vecs(struct intel_engine_cs *engine)
2258 struct drm_i915_private *i915 = engine->i915;
2260 GEM_BUG_ON(INTEL_GEN(i915) < 7);
2262 engine->emit_flush = gen6_ring_flush;
2263 engine->irq_enable_mask = PM_VEBOX_USER_INTERRUPT;
2264 engine->irq_enable = hsw_vebox_irq_enable;
2265 engine->irq_disable = hsw_vebox_irq_disable;
2267 engine->emit_fini_breadcrumb = gen7_xcs_emit_breadcrumb;
2270 int intel_ring_submission_setup(struct intel_engine_cs *engine)
2272 setup_common(engine);
2274 switch (engine->class) {
2278 case VIDEO_DECODE_CLASS:
2281 case COPY_ENGINE_CLASS:
2284 case VIDEO_ENHANCEMENT_CLASS:
2288 MISSING_CASE(engine->class);
2295 int intel_ring_submission_init(struct intel_engine_cs *engine)
2297 struct i915_timeline *timeline;
2298 struct intel_ring *ring;
2301 timeline = i915_timeline_create(engine->i915, engine->status_page.vma);
2302 if (IS_ERR(timeline)) {
2303 err = PTR_ERR(timeline);
2306 GEM_BUG_ON(timeline->has_initial_breadcrumb);
2308 ring = intel_engine_create_ring(engine, timeline, 32 * PAGE_SIZE);
2309 i915_timeline_put(timeline);
2311 err = PTR_ERR(ring);
2315 err = intel_ring_pin(ring);
2319 GEM_BUG_ON(engine->buffer);
2320 engine->buffer = ring;
2322 err = intel_engine_init_common(engine);
2326 GEM_BUG_ON(ring->timeline->hwsp_ggtt != engine->status_page.vma);
2331 intel_ring_unpin(ring);
2333 intel_ring_put(ring);
2335 intel_engine_cleanup_common(engine);