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>
33 #include <drm/i915_drm.h>
36 #include "i915_gem_render_state.h"
37 #include "i915_trace.h"
38 #include "intel_drv.h"
39 #include "intel_workarounds.h"
41 /* Rough estimate of the typical request size, performing a flush,
42 * set-context and then emitting the batch.
44 #define LEGACY_REQUEST_SIZE 200
46 static unsigned int __intel_ring_space(unsigned int head,
51 * "If the Ring Buffer Head Pointer and the Tail Pointer are on the
52 * same cacheline, the Head Pointer must not be greater than the Tail
55 GEM_BUG_ON(!is_power_of_2(size));
56 return (head - tail - CACHELINE_BYTES) & (size - 1);
59 unsigned int intel_ring_update_space(struct intel_ring *ring)
63 space = __intel_ring_space(ring->head, ring->emit, ring->size);
70 gen2_render_ring_flush(struct i915_request *rq, u32 mode)
76 if (mode & EMIT_INVALIDATE)
79 cs = intel_ring_begin(rq, 2);
85 intel_ring_advance(rq, cs);
91 gen4_render_ring_flush(struct i915_request *rq, u32 mode)
98 * I915_GEM_DOMAIN_RENDER is always invalidated, but is
99 * only flushed if MI_NO_WRITE_FLUSH is unset. On 965, it is
100 * also flushed at 2d versus 3d pipeline switches.
104 * I915_GEM_DOMAIN_SAMPLER is flushed on pre-965 if
105 * MI_READ_FLUSH is set, and is always flushed on 965.
107 * I915_GEM_DOMAIN_COMMAND may not exist?
109 * I915_GEM_DOMAIN_INSTRUCTION, which exists on 965, is
110 * invalidated when MI_EXE_FLUSH is set.
112 * I915_GEM_DOMAIN_VERTEX, which exists on 965, is
113 * invalidated with every MI_FLUSH.
117 * On 965, TLBs associated with I915_GEM_DOMAIN_COMMAND
118 * and I915_GEM_DOMAIN_CPU in are invalidated at PTE write and
119 * I915_GEM_DOMAIN_RENDER and I915_GEM_DOMAIN_SAMPLER
120 * are flushed at any MI_FLUSH.
124 if (mode & EMIT_INVALIDATE) {
126 if (IS_G4X(rq->i915) || IS_GEN5(rq->i915))
127 cmd |= MI_INVALIDATE_ISP;
130 cs = intel_ring_begin(rq, 2);
136 intel_ring_advance(rq, cs);
142 * Emits a PIPE_CONTROL with a non-zero post-sync operation, for
143 * implementing two workarounds on gen6. From section 1.4.7.1
144 * "PIPE_CONTROL" of the Sandy Bridge PRM volume 2 part 1:
146 * [DevSNB-C+{W/A}] Before any depth stall flush (including those
147 * produced by non-pipelined state commands), software needs to first
148 * send a PIPE_CONTROL with no bits set except Post-Sync Operation !=
151 * [Dev-SNB{W/A}]: Before a PIPE_CONTROL with Write Cache Flush Enable
152 * =1, a PIPE_CONTROL with any non-zero post-sync-op is required.
154 * And the workaround for these two requires this workaround first:
156 * [Dev-SNB{W/A}]: Pipe-control with CS-stall bit set must be sent
157 * BEFORE the pipe-control with a post-sync op and no write-cache
160 * And this last workaround is tricky because of the requirements on
161 * that bit. From section 1.4.7.2.3 "Stall" of the Sandy Bridge PRM
164 * "1 of the following must also be set:
165 * - Render Target Cache Flush Enable ([12] of DW1)
166 * - Depth Cache Flush Enable ([0] of DW1)
167 * - Stall at Pixel Scoreboard ([1] of DW1)
168 * - Depth Stall ([13] of DW1)
169 * - Post-Sync Operation ([13] of DW1)
170 * - Notify Enable ([8] of DW1)"
172 * The cache flushes require the workaround flush that triggered this
173 * one, so we can't use it. Depth stall would trigger the same.
174 * Post-sync nonzero is what triggered this second workaround, so we
175 * can't use that one either. Notify enable is IRQs, which aren't
176 * really our business. That leaves only stall at scoreboard.
179 intel_emit_post_sync_nonzero_flush(struct i915_request *rq)
182 i915_ggtt_offset(rq->engine->scratch) + 2 * CACHELINE_BYTES;
185 cs = intel_ring_begin(rq, 6);
189 *cs++ = GFX_OP_PIPE_CONTROL(5);
190 *cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
191 *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
192 *cs++ = 0; /* low dword */
193 *cs++ = 0; /* high dword */
195 intel_ring_advance(rq, cs);
197 cs = intel_ring_begin(rq, 6);
201 *cs++ = GFX_OP_PIPE_CONTROL(5);
202 *cs++ = PIPE_CONTROL_QW_WRITE;
203 *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
207 intel_ring_advance(rq, cs);
213 gen6_render_ring_flush(struct i915_request *rq, u32 mode)
216 i915_ggtt_offset(rq->engine->scratch) + 2 * CACHELINE_BYTES;
220 /* Force SNB workarounds for PIPE_CONTROL flushes */
221 ret = intel_emit_post_sync_nonzero_flush(rq);
225 /* Just flush everything. Experiments have shown that reducing the
226 * number of bits based on the write domains has little performance
229 if (mode & EMIT_FLUSH) {
230 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
231 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
233 * Ensure that any following seqno writes only happen
234 * when the render cache is indeed flushed.
236 flags |= PIPE_CONTROL_CS_STALL;
238 if (mode & EMIT_INVALIDATE) {
239 flags |= PIPE_CONTROL_TLB_INVALIDATE;
240 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
241 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
242 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
243 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
244 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
246 * TLB invalidate requires a post-sync write.
248 flags |= PIPE_CONTROL_QW_WRITE | PIPE_CONTROL_CS_STALL;
251 cs = intel_ring_begin(rq, 4);
255 *cs++ = GFX_OP_PIPE_CONTROL(4);
257 *cs++ = scratch_addr | PIPE_CONTROL_GLOBAL_GTT;
259 intel_ring_advance(rq, cs);
265 gen7_render_ring_cs_stall_wa(struct i915_request *rq)
269 cs = intel_ring_begin(rq, 4);
273 *cs++ = GFX_OP_PIPE_CONTROL(4);
274 *cs++ = PIPE_CONTROL_CS_STALL | PIPE_CONTROL_STALL_AT_SCOREBOARD;
277 intel_ring_advance(rq, cs);
283 gen7_render_ring_flush(struct i915_request *rq, u32 mode)
286 i915_ggtt_offset(rq->engine->scratch) + 2 * CACHELINE_BYTES;
290 * Ensure that any following seqno writes only happen when the render
291 * cache is indeed flushed.
293 * Workaround: 4th PIPE_CONTROL command (except the ones with only
294 * read-cache invalidate bits set) must have the CS_STALL bit set. We
295 * don't try to be clever and just set it unconditionally.
297 flags |= PIPE_CONTROL_CS_STALL;
299 /* Just flush everything. Experiments have shown that reducing the
300 * number of bits based on the write domains has little performance
303 if (mode & EMIT_FLUSH) {
304 flags |= PIPE_CONTROL_RENDER_TARGET_CACHE_FLUSH;
305 flags |= PIPE_CONTROL_DEPTH_CACHE_FLUSH;
306 flags |= PIPE_CONTROL_DC_FLUSH_ENABLE;
307 flags |= PIPE_CONTROL_FLUSH_ENABLE;
309 if (mode & EMIT_INVALIDATE) {
310 flags |= PIPE_CONTROL_TLB_INVALIDATE;
311 flags |= PIPE_CONTROL_INSTRUCTION_CACHE_INVALIDATE;
312 flags |= PIPE_CONTROL_TEXTURE_CACHE_INVALIDATE;
313 flags |= PIPE_CONTROL_VF_CACHE_INVALIDATE;
314 flags |= PIPE_CONTROL_CONST_CACHE_INVALIDATE;
315 flags |= PIPE_CONTROL_STATE_CACHE_INVALIDATE;
316 flags |= PIPE_CONTROL_MEDIA_STATE_CLEAR;
318 * TLB invalidate requires a post-sync write.
320 flags |= PIPE_CONTROL_QW_WRITE;
321 flags |= PIPE_CONTROL_GLOBAL_GTT_IVB;
323 flags |= PIPE_CONTROL_STALL_AT_SCOREBOARD;
325 /* Workaround: we must issue a pipe_control with CS-stall bit
326 * set before a pipe_control command that has the state cache
327 * invalidate bit set. */
328 gen7_render_ring_cs_stall_wa(rq);
331 cs = intel_ring_begin(rq, 4);
335 *cs++ = GFX_OP_PIPE_CONTROL(4);
337 *cs++ = scratch_addr;
339 intel_ring_advance(rq, cs);
344 static void ring_setup_phys_status_page(struct intel_engine_cs *engine)
346 struct drm_i915_private *dev_priv = engine->i915;
349 addr = dev_priv->status_page_dmah->busaddr;
350 if (INTEL_GEN(dev_priv) >= 4)
351 addr |= (dev_priv->status_page_dmah->busaddr >> 28) & 0xf0;
352 I915_WRITE(HWS_PGA, addr);
355 static void intel_ring_setup_status_page(struct intel_engine_cs *engine)
357 struct drm_i915_private *dev_priv = engine->i915;
360 /* The ring status page addresses are no longer next to the rest of
361 * the ring registers as of gen7.
363 if (IS_GEN7(dev_priv)) {
364 switch (engine->id) {
366 * No more rings exist on Gen7. Default case is only to shut up
367 * gcc switch check warning.
370 GEM_BUG_ON(engine->id);
372 mmio = RENDER_HWS_PGA_GEN7;
375 mmio = BLT_HWS_PGA_GEN7;
378 mmio = BSD_HWS_PGA_GEN7;
381 mmio = VEBOX_HWS_PGA_GEN7;
384 } else if (IS_GEN6(dev_priv)) {
385 mmio = RING_HWS_PGA_GEN6(engine->mmio_base);
387 mmio = RING_HWS_PGA(engine->mmio_base);
390 if (INTEL_GEN(dev_priv) >= 6)
391 I915_WRITE(RING_HWSTAM(engine->mmio_base), 0xffffffff);
393 I915_WRITE(mmio, engine->status_page.ggtt_offset);
396 /* Flush the TLB for this page */
397 if (IS_GEN(dev_priv, 6, 7)) {
398 i915_reg_t reg = RING_INSTPM(engine->mmio_base);
400 /* ring should be idle before issuing a sync flush*/
401 WARN_ON((I915_READ_MODE(engine) & MODE_IDLE) == 0);
404 _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE |
406 if (intel_wait_for_register(dev_priv,
407 reg, INSTPM_SYNC_FLUSH, 0,
409 DRM_ERROR("%s: wait for SyncFlush to complete for TLB invalidation timed out\n",
414 static bool stop_ring(struct intel_engine_cs *engine)
416 struct drm_i915_private *dev_priv = engine->i915;
418 if (INTEL_GEN(dev_priv) > 2) {
419 I915_WRITE_MODE(engine, _MASKED_BIT_ENABLE(STOP_RING));
420 if (intel_wait_for_register(dev_priv,
421 RING_MI_MODE(engine->mmio_base),
425 DRM_ERROR("%s : timed out trying to stop ring\n",
427 /* Sometimes we observe that the idle flag is not
428 * set even though the ring is empty. So double
429 * check before giving up.
431 if (I915_READ_HEAD(engine) != I915_READ_TAIL(engine))
436 I915_WRITE_HEAD(engine, I915_READ_TAIL(engine));
438 I915_WRITE_HEAD(engine, 0);
439 I915_WRITE_TAIL(engine, 0);
441 /* The ring must be empty before it is disabled */
442 I915_WRITE_CTL(engine, 0);
444 return (I915_READ_HEAD(engine) & HEAD_ADDR) == 0;
447 static int init_ring_common(struct intel_engine_cs *engine)
449 struct drm_i915_private *dev_priv = engine->i915;
450 struct intel_ring *ring = engine->buffer;
453 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
455 if (!stop_ring(engine)) {
456 /* G45 ring initialization often fails to reset head to zero */
457 DRM_DEBUG_DRIVER("%s head not reset to zero "
458 "ctl %08x head %08x tail %08x start %08x\n",
460 I915_READ_CTL(engine),
461 I915_READ_HEAD(engine),
462 I915_READ_TAIL(engine),
463 I915_READ_START(engine));
465 if (!stop_ring(engine)) {
466 DRM_ERROR("failed to set %s head to zero "
467 "ctl %08x head %08x tail %08x start %08x\n",
469 I915_READ_CTL(engine),
470 I915_READ_HEAD(engine),
471 I915_READ_TAIL(engine),
472 I915_READ_START(engine));
478 if (HWS_NEEDS_PHYSICAL(dev_priv))
479 ring_setup_phys_status_page(engine);
481 intel_ring_setup_status_page(engine);
483 intel_engine_reset_breadcrumbs(engine);
485 /* Enforce ordering by reading HEAD register back */
486 I915_READ_HEAD(engine);
488 /* Initialize the ring. This must happen _after_ we've cleared the ring
489 * registers with the above sequence (the readback of the HEAD registers
490 * also enforces ordering), otherwise the hw might lose the new ring
491 * register values. */
492 I915_WRITE_START(engine, i915_ggtt_offset(ring->vma));
494 /* WaClearRingBufHeadRegAtInit:ctg,elk */
495 if (I915_READ_HEAD(engine))
496 DRM_DEBUG_DRIVER("%s initialization failed [head=%08x], fudging\n",
497 engine->name, I915_READ_HEAD(engine));
499 intel_ring_update_space(ring);
500 I915_WRITE_HEAD(engine, ring->head);
501 I915_WRITE_TAIL(engine, ring->tail);
502 (void)I915_READ_TAIL(engine);
504 I915_WRITE_CTL(engine, RING_CTL_SIZE(ring->size) | RING_VALID);
506 /* If the head is still not zero, the ring is dead */
507 if (intel_wait_for_register(dev_priv, RING_CTL(engine->mmio_base),
508 RING_VALID, RING_VALID,
510 DRM_ERROR("%s initialization failed "
511 "ctl %08x (valid? %d) head %08x [%08x] tail %08x [%08x] start %08x [expected %08x]\n",
513 I915_READ_CTL(engine),
514 I915_READ_CTL(engine) & RING_VALID,
515 I915_READ_HEAD(engine), ring->head,
516 I915_READ_TAIL(engine), ring->tail,
517 I915_READ_START(engine),
518 i915_ggtt_offset(ring->vma));
523 intel_engine_init_hangcheck(engine);
525 if (INTEL_GEN(dev_priv) > 2)
526 I915_WRITE_MODE(engine, _MASKED_BIT_DISABLE(STOP_RING));
529 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
534 static struct i915_request *reset_prepare(struct intel_engine_cs *engine)
536 intel_engine_stop_cs(engine);
538 if (engine->irq_seqno_barrier)
539 engine->irq_seqno_barrier(engine);
541 return i915_gem_find_active_request(engine);
544 static void skip_request(struct i915_request *rq)
546 void *vaddr = rq->ring->vaddr;
550 if (rq->postfix < head) {
551 memset32(vaddr + head, MI_NOOP,
552 (rq->ring->size - head) / sizeof(u32));
555 memset32(vaddr + head, MI_NOOP, (rq->postfix - head) / sizeof(u32));
558 static void reset_ring(struct intel_engine_cs *engine, struct i915_request *rq)
560 GEM_TRACE("%s seqno=%x\n", engine->name, rq ? rq->global_seqno : 0);
563 * RC6 must be prevented until the reset is complete and the engine
564 * reinitialised. If it occurs in the middle of this sequence, the
565 * state written to/loaded from the power context is ill-defined (e.g.
566 * the PP_BASE_DIR may be lost).
568 assert_forcewakes_active(engine->i915, FORCEWAKE_ALL);
571 * Try to restore the logical GPU state to match the continuation
572 * of the request queue. If we skip the context/PD restore, then
573 * the next request may try to execute assuming that its context
574 * is valid and loaded on the GPU and so may try to access invalid
575 * memory, prompting repeated GPU hangs.
577 * If the request was guilty, we still restore the logical state
578 * in case the next request requires it (e.g. the aliasing ppgtt),
579 * but skip over the hung batch.
581 * If the request was innocent, we try to replay the request with
582 * the restored context.
585 /* If the rq hung, jump to its breadcrumb and skip the batch */
586 rq->ring->head = intel_ring_wrap(rq->ring, rq->head);
587 if (rq->fence.error == -EIO)
592 static void reset_finish(struct intel_engine_cs *engine)
596 static int intel_rcs_ctx_init(struct i915_request *rq)
600 ret = intel_ctx_workarounds_emit(rq);
604 ret = i915_gem_render_state_emit(rq);
611 static int init_render_ring(struct intel_engine_cs *engine)
613 struct drm_i915_private *dev_priv = engine->i915;
614 int ret = init_ring_common(engine);
618 intel_whitelist_workarounds_apply(engine);
620 /* WaTimedSingleVertexDispatch:cl,bw,ctg,elk,ilk,snb */
621 if (IS_GEN(dev_priv, 4, 6))
622 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(VS_TIMER_DISPATCH));
624 /* We need to disable the AsyncFlip performance optimisations in order
625 * to use MI_WAIT_FOR_EVENT within the CS. It should already be
626 * programmed to '1' on all products.
628 * WaDisableAsyncFlipPerfMode:snb,ivb,hsw,vlv
630 if (IS_GEN(dev_priv, 6, 7))
631 I915_WRITE(MI_MODE, _MASKED_BIT_ENABLE(ASYNC_FLIP_PERF_DISABLE));
633 /* Required for the hardware to program scanline values for waiting */
634 /* WaEnableFlushTlbInvalidationMode:snb */
635 if (IS_GEN6(dev_priv))
637 _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT));
639 /* WaBCSVCSTlbInvalidationMode:ivb,vlv,hsw */
640 if (IS_GEN7(dev_priv))
641 I915_WRITE(GFX_MODE_GEN7,
642 _MASKED_BIT_ENABLE(GFX_TLB_INVALIDATE_EXPLICIT) |
643 _MASKED_BIT_ENABLE(GFX_REPLAY_MODE));
645 if (IS_GEN6(dev_priv)) {
646 /* From the Sandybridge PRM, volume 1 part 3, page 24:
647 * "If this bit is set, STCunit will have LRA as replacement
648 * policy. [...] This bit must be reset. LRA replacement
649 * policy is not supported."
651 I915_WRITE(CACHE_MODE_0,
652 _MASKED_BIT_DISABLE(CM0_STC_EVICT_DISABLE_LRA_SNB));
655 if (IS_GEN(dev_priv, 6, 7))
656 I915_WRITE(INSTPM, _MASKED_BIT_ENABLE(INSTPM_FORCE_ORDERING));
658 if (INTEL_GEN(dev_priv) >= 6)
659 I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
664 static u32 *gen6_signal(struct i915_request *rq, u32 *cs)
666 struct drm_i915_private *dev_priv = rq->i915;
667 struct intel_engine_cs *engine;
668 enum intel_engine_id id;
671 for_each_engine(engine, dev_priv, id) {
674 if (!(BIT(engine->hw_id) & GEN6_SEMAPHORES_MASK))
677 mbox_reg = rq->engine->semaphore.mbox.signal[engine->hw_id];
678 if (i915_mmio_reg_valid(mbox_reg)) {
679 *cs++ = MI_LOAD_REGISTER_IMM(1);
680 *cs++ = i915_mmio_reg_offset(mbox_reg);
681 *cs++ = rq->global_seqno;
691 static void cancel_requests(struct intel_engine_cs *engine)
693 struct i915_request *request;
696 spin_lock_irqsave(&engine->timeline.lock, flags);
698 /* Mark all submitted requests as skipped. */
699 list_for_each_entry(request, &engine->timeline.requests, link) {
700 GEM_BUG_ON(!request->global_seqno);
701 if (!i915_request_completed(request))
702 dma_fence_set_error(&request->fence, -EIO);
704 /* Remaining _unready_ requests will be nop'ed when submitted */
706 spin_unlock_irqrestore(&engine->timeline.lock, flags);
709 static void i9xx_submit_request(struct i915_request *request)
711 struct drm_i915_private *dev_priv = request->i915;
713 i915_request_submit(request);
715 I915_WRITE_TAIL(request->engine,
716 intel_ring_set_tail(request->ring, request->tail));
719 static void i9xx_emit_breadcrumb(struct i915_request *rq, u32 *cs)
721 *cs++ = MI_STORE_DWORD_INDEX;
722 *cs++ = I915_GEM_HWS_INDEX << MI_STORE_DWORD_INDEX_SHIFT;
723 *cs++ = rq->global_seqno;
724 *cs++ = MI_USER_INTERRUPT;
726 rq->tail = intel_ring_offset(rq, cs);
727 assert_ring_tail_valid(rq->ring, rq->tail);
730 static const int i9xx_emit_breadcrumb_sz = 4;
732 static void gen6_sema_emit_breadcrumb(struct i915_request *rq, u32 *cs)
734 return i9xx_emit_breadcrumb(rq, rq->engine->semaphore.signal(rq, cs));
738 gen6_ring_sync_to(struct i915_request *rq, struct i915_request *signal)
740 u32 dw1 = MI_SEMAPHORE_MBOX |
741 MI_SEMAPHORE_COMPARE |
742 MI_SEMAPHORE_REGISTER;
743 u32 wait_mbox = signal->engine->semaphore.mbox.wait[rq->engine->hw_id];
746 WARN_ON(wait_mbox == MI_SEMAPHORE_SYNC_INVALID);
748 cs = intel_ring_begin(rq, 4);
752 *cs++ = dw1 | wait_mbox;
753 /* Throughout all of the GEM code, seqno passed implies our current
754 * seqno is >= the last seqno executed. However for hardware the
755 * comparison is strictly greater than.
757 *cs++ = signal->global_seqno - 1;
760 intel_ring_advance(rq, cs);
766 gen5_seqno_barrier(struct intel_engine_cs *engine)
768 /* MI_STORE are internally buffered by the GPU and not flushed
769 * either by MI_FLUSH or SyncFlush or any other combination of
772 * "Only the submission of the store operation is guaranteed.
773 * The write result will be complete (coherent) some time later
774 * (this is practically a finite period but there is no guaranteed
777 * Empirically, we observe that we need a delay of at least 75us to
778 * be sure that the seqno write is visible by the CPU.
780 usleep_range(125, 250);
784 gen6_seqno_barrier(struct intel_engine_cs *engine)
786 struct drm_i915_private *dev_priv = engine->i915;
788 /* Workaround to force correct ordering between irq and seqno writes on
789 * ivb (and maybe also on snb) by reading from a CS register (like
790 * ACTHD) before reading the status page.
792 * Note that this effectively stalls the read by the time it takes to
793 * do a memory transaction, which more or less ensures that the write
794 * from the GPU has sufficient time to invalidate the CPU cacheline.
795 * Alternatively we could delay the interrupt from the CS ring to give
796 * the write time to land, but that would incur a delay after every
797 * batch i.e. much more frequent than a delay when waiting for the
798 * interrupt (with the same net latency).
800 * Also note that to prevent whole machine hangs on gen7, we have to
801 * take the spinlock to guard against concurrent cacheline access.
803 spin_lock_irq(&dev_priv->uncore.lock);
804 POSTING_READ_FW(RING_ACTHD(engine->mmio_base));
805 spin_unlock_irq(&dev_priv->uncore.lock);
809 gen5_irq_enable(struct intel_engine_cs *engine)
811 gen5_enable_gt_irq(engine->i915, engine->irq_enable_mask);
815 gen5_irq_disable(struct intel_engine_cs *engine)
817 gen5_disable_gt_irq(engine->i915, engine->irq_enable_mask);
821 i9xx_irq_enable(struct intel_engine_cs *engine)
823 struct drm_i915_private *dev_priv = engine->i915;
825 dev_priv->irq_mask &= ~engine->irq_enable_mask;
826 I915_WRITE(IMR, dev_priv->irq_mask);
827 POSTING_READ_FW(RING_IMR(engine->mmio_base));
831 i9xx_irq_disable(struct intel_engine_cs *engine)
833 struct drm_i915_private *dev_priv = engine->i915;
835 dev_priv->irq_mask |= engine->irq_enable_mask;
836 I915_WRITE(IMR, dev_priv->irq_mask);
840 i8xx_irq_enable(struct intel_engine_cs *engine)
842 struct drm_i915_private *dev_priv = engine->i915;
844 dev_priv->irq_mask &= ~engine->irq_enable_mask;
845 I915_WRITE16(IMR, dev_priv->irq_mask);
846 POSTING_READ16(RING_IMR(engine->mmio_base));
850 i8xx_irq_disable(struct intel_engine_cs *engine)
852 struct drm_i915_private *dev_priv = engine->i915;
854 dev_priv->irq_mask |= engine->irq_enable_mask;
855 I915_WRITE16(IMR, dev_priv->irq_mask);
859 bsd_ring_flush(struct i915_request *rq, u32 mode)
863 cs = intel_ring_begin(rq, 2);
869 intel_ring_advance(rq, cs);
874 gen6_irq_enable(struct intel_engine_cs *engine)
876 struct drm_i915_private *dev_priv = engine->i915;
878 I915_WRITE_IMR(engine,
879 ~(engine->irq_enable_mask |
880 engine->irq_keep_mask));
881 gen5_enable_gt_irq(dev_priv, engine->irq_enable_mask);
885 gen6_irq_disable(struct intel_engine_cs *engine)
887 struct drm_i915_private *dev_priv = engine->i915;
889 I915_WRITE_IMR(engine, ~engine->irq_keep_mask);
890 gen5_disable_gt_irq(dev_priv, engine->irq_enable_mask);
894 hsw_vebox_irq_enable(struct intel_engine_cs *engine)
896 struct drm_i915_private *dev_priv = engine->i915;
898 I915_WRITE_IMR(engine, ~engine->irq_enable_mask);
899 gen6_unmask_pm_irq(dev_priv, engine->irq_enable_mask);
903 hsw_vebox_irq_disable(struct intel_engine_cs *engine)
905 struct drm_i915_private *dev_priv = engine->i915;
907 I915_WRITE_IMR(engine, ~0);
908 gen6_mask_pm_irq(dev_priv, engine->irq_enable_mask);
912 i965_emit_bb_start(struct i915_request *rq,
913 u64 offset, u32 length,
914 unsigned int dispatch_flags)
918 cs = intel_ring_begin(rq, 2);
922 *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT | (dispatch_flags &
923 I915_DISPATCH_SECURE ? 0 : MI_BATCH_NON_SECURE_I965);
925 intel_ring_advance(rq, cs);
930 /* Just userspace ABI convention to limit the wa batch bo to a resonable size */
931 #define I830_BATCH_LIMIT (256*1024)
932 #define I830_TLB_ENTRIES (2)
933 #define I830_WA_SIZE max(I830_TLB_ENTRIES*4096, I830_BATCH_LIMIT)
935 i830_emit_bb_start(struct i915_request *rq,
937 unsigned int dispatch_flags)
939 u32 *cs, cs_offset = i915_ggtt_offset(rq->engine->scratch);
941 cs = intel_ring_begin(rq, 6);
945 /* Evict the invalid PTE TLBs */
946 *cs++ = COLOR_BLT_CMD | BLT_WRITE_RGBA;
947 *cs++ = BLT_DEPTH_32 | BLT_ROP_COLOR_COPY | 4096;
948 *cs++ = I830_TLB_ENTRIES << 16 | 4; /* load each page */
952 intel_ring_advance(rq, cs);
954 if ((dispatch_flags & I915_DISPATCH_PINNED) == 0) {
955 if (len > I830_BATCH_LIMIT)
958 cs = intel_ring_begin(rq, 6 + 2);
962 /* Blit the batch (which has now all relocs applied) to the
963 * stable batch scratch bo area (so that the CS never
964 * stumbles over its tlb invalidation bug) ...
966 *cs++ = SRC_COPY_BLT_CMD | BLT_WRITE_RGBA;
967 *cs++ = BLT_DEPTH_32 | BLT_ROP_SRC_COPY | 4096;
968 *cs++ = DIV_ROUND_UP(len, 4096) << 16 | 4096;
975 intel_ring_advance(rq, cs);
977 /* ... and execute it. */
981 cs = intel_ring_begin(rq, 2);
985 *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
986 *cs++ = offset | (dispatch_flags & I915_DISPATCH_SECURE ? 0 :
987 MI_BATCH_NON_SECURE);
988 intel_ring_advance(rq, cs);
994 i915_emit_bb_start(struct i915_request *rq,
996 unsigned int dispatch_flags)
1000 cs = intel_ring_begin(rq, 2);
1004 *cs++ = MI_BATCH_BUFFER_START | MI_BATCH_GTT;
1005 *cs++ = offset | (dispatch_flags & I915_DISPATCH_SECURE ? 0 :
1006 MI_BATCH_NON_SECURE);
1007 intel_ring_advance(rq, cs);
1014 int intel_ring_pin(struct intel_ring *ring,
1015 struct drm_i915_private *i915,
1016 unsigned int offset_bias)
1018 enum i915_map_type map = HAS_LLC(i915) ? I915_MAP_WB : I915_MAP_WC;
1019 struct i915_vma *vma = ring->vma;
1024 GEM_BUG_ON(ring->vaddr);
1029 flags |= PIN_OFFSET_BIAS | offset_bias;
1030 if (vma->obj->stolen)
1031 flags |= PIN_MAPPABLE;
1033 if (!(vma->flags & I915_VMA_GLOBAL_BIND)) {
1034 if (flags & PIN_MAPPABLE || map == I915_MAP_WC)
1035 ret = i915_gem_object_set_to_gtt_domain(vma->obj, true);
1037 ret = i915_gem_object_set_to_cpu_domain(vma->obj, true);
1042 ret = i915_vma_pin(vma, 0, PAGE_SIZE, flags);
1046 if (i915_vma_is_map_and_fenceable(vma))
1047 addr = (void __force *)i915_vma_pin_iomap(vma);
1049 addr = i915_gem_object_pin_map(vma->obj, map);
1053 vma->obj->pin_global++;
1059 i915_vma_unpin(vma);
1060 return PTR_ERR(addr);
1063 void intel_ring_reset(struct intel_ring *ring, u32 tail)
1068 intel_ring_update_space(ring);
1071 void intel_ring_unpin(struct intel_ring *ring)
1073 GEM_BUG_ON(!ring->vma);
1074 GEM_BUG_ON(!ring->vaddr);
1076 /* Discard any unused bytes beyond that submitted to hw. */
1077 intel_ring_reset(ring, ring->tail);
1079 if (i915_vma_is_map_and_fenceable(ring->vma))
1080 i915_vma_unpin_iomap(ring->vma);
1082 i915_gem_object_unpin_map(ring->vma->obj);
1085 ring->vma->obj->pin_global--;
1086 i915_vma_unpin(ring->vma);
1089 static struct i915_vma *
1090 intel_ring_create_vma(struct drm_i915_private *dev_priv, int size)
1092 struct drm_i915_gem_object *obj;
1093 struct i915_vma *vma;
1095 obj = i915_gem_object_create_stolen(dev_priv, size);
1097 obj = i915_gem_object_create_internal(dev_priv, size);
1099 return ERR_CAST(obj);
1101 /* mark ring buffers as read-only from GPU side by default */
1104 vma = i915_vma_instance(obj, &dev_priv->ggtt.vm, NULL);
1111 i915_gem_object_put(obj);
1116 intel_engine_create_ring(struct intel_engine_cs *engine,
1117 struct i915_timeline *timeline,
1120 struct intel_ring *ring;
1121 struct i915_vma *vma;
1123 GEM_BUG_ON(!is_power_of_2(size));
1124 GEM_BUG_ON(RING_CTL_SIZE(size) & ~RING_NR_PAGES);
1125 GEM_BUG_ON(timeline == &engine->timeline);
1126 lockdep_assert_held(&engine->i915->drm.struct_mutex);
1128 ring = kzalloc(sizeof(*ring), GFP_KERNEL);
1130 return ERR_PTR(-ENOMEM);
1132 INIT_LIST_HEAD(&ring->request_list);
1133 ring->timeline = i915_timeline_get(timeline);
1136 /* Workaround an erratum on the i830 which causes a hang if
1137 * the TAIL pointer points to within the last 2 cachelines
1140 ring->effective_size = size;
1141 if (IS_I830(engine->i915) || IS_I845G(engine->i915))
1142 ring->effective_size -= 2 * CACHELINE_BYTES;
1144 intel_ring_update_space(ring);
1146 vma = intel_ring_create_vma(engine->i915, size);
1149 return ERR_CAST(vma);
1157 intel_ring_free(struct intel_ring *ring)
1159 struct drm_i915_gem_object *obj = ring->vma->obj;
1161 i915_vma_close(ring->vma);
1162 __i915_gem_object_release_unless_active(obj);
1164 i915_timeline_put(ring->timeline);
1168 static void intel_ring_context_destroy(struct intel_context *ce)
1170 GEM_BUG_ON(ce->pin_count);
1173 __i915_gem_object_release_unless_active(ce->state->obj);
1176 static int __context_pin(struct intel_context *ce)
1178 struct i915_vma *vma;
1186 * Clear this page out of any CPU caches for coherent swap-in/out.
1187 * We only want to do this on the first bind so that we do not stall
1188 * on an active context (which by nature is already on the GPU).
1190 if (!(vma->flags & I915_VMA_GLOBAL_BIND)) {
1191 err = i915_gem_object_set_to_gtt_domain(vma->obj, true);
1196 err = i915_vma_pin(vma, 0, I915_GTT_MIN_ALIGNMENT,
1197 PIN_GLOBAL | PIN_HIGH);
1202 * And mark is as a globally pinned object to let the shrinker know
1203 * it cannot reclaim the object until we release it.
1205 vma->obj->pin_global++;
1210 static void __context_unpin(struct intel_context *ce)
1212 struct i915_vma *vma;
1218 vma->obj->pin_global--;
1219 i915_vma_unpin(vma);
1222 static void intel_ring_context_unpin(struct intel_context *ce)
1224 __context_unpin(ce);
1226 i915_gem_context_put(ce->gem_context);
1229 static struct i915_vma *
1230 alloc_context_vma(struct intel_engine_cs *engine)
1232 struct drm_i915_private *i915 = engine->i915;
1233 struct drm_i915_gem_object *obj;
1234 struct i915_vma *vma;
1237 obj = i915_gem_object_create(i915, engine->context_size);
1239 return ERR_CAST(obj);
1241 if (engine->default_state) {
1242 void *defaults, *vaddr;
1244 vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
1245 if (IS_ERR(vaddr)) {
1246 err = PTR_ERR(vaddr);
1250 defaults = i915_gem_object_pin_map(engine->default_state,
1252 if (IS_ERR(defaults)) {
1253 err = PTR_ERR(defaults);
1257 memcpy(vaddr, defaults, engine->context_size);
1259 i915_gem_object_unpin_map(engine->default_state);
1260 i915_gem_object_unpin_map(obj);
1264 * Try to make the context utilize L3 as well as LLC.
1266 * On VLV we don't have L3 controls in the PTEs so we
1267 * shouldn't touch the cache level, especially as that
1268 * would make the object snooped which might have a
1269 * negative performance impact.
1271 * Snooping is required on non-llc platforms in execlist
1272 * mode, but since all GGTT accesses use PAT entry 0 we
1273 * get snooping anyway regardless of cache_level.
1275 * This is only applicable for Ivy Bridge devices since
1276 * later platforms don't have L3 control bits in the PTE.
1278 if (IS_IVYBRIDGE(i915)) {
1279 /* Ignore any error, regard it as a simple optimisation */
1280 i915_gem_object_set_cache_level(obj, I915_CACHE_L3_LLC);
1283 vma = i915_vma_instance(obj, &i915->ggtt.vm, NULL);
1292 i915_gem_object_unpin_map(obj);
1294 i915_gem_object_put(obj);
1295 return ERR_PTR(err);
1298 static struct intel_context *
1299 __ring_context_pin(struct intel_engine_cs *engine,
1300 struct i915_gem_context *ctx,
1301 struct intel_context *ce)
1305 if (!ce->state && engine->context_size) {
1306 struct i915_vma *vma;
1308 vma = alloc_context_vma(engine);
1317 err = __context_pin(ce);
1321 i915_gem_context_get(ctx);
1323 /* One ringbuffer to rule them all */
1324 GEM_BUG_ON(!engine->buffer);
1325 ce->ring = engine->buffer;
1331 return ERR_PTR(err);
1334 static const struct intel_context_ops ring_context_ops = {
1335 .unpin = intel_ring_context_unpin,
1336 .destroy = intel_ring_context_destroy,
1339 static struct intel_context *
1340 intel_ring_context_pin(struct intel_engine_cs *engine,
1341 struct i915_gem_context *ctx)
1343 struct intel_context *ce = to_intel_context(ctx, engine);
1345 lockdep_assert_held(&ctx->i915->drm.struct_mutex);
1347 if (likely(ce->pin_count++))
1349 GEM_BUG_ON(!ce->pin_count); /* no overflow please! */
1351 ce->ops = &ring_context_ops;
1353 return __ring_context_pin(engine, ctx, ce);
1356 static int intel_init_ring_buffer(struct intel_engine_cs *engine)
1358 struct intel_ring *ring;
1359 struct i915_timeline *timeline;
1362 intel_engine_setup_common(engine);
1364 timeline = i915_timeline_create(engine->i915, engine->name);
1365 if (IS_ERR(timeline)) {
1366 err = PTR_ERR(timeline);
1370 ring = intel_engine_create_ring(engine, timeline, 32 * PAGE_SIZE);
1371 i915_timeline_put(timeline);
1373 err = PTR_ERR(ring);
1377 /* Ring wraparound at offset 0 sometimes hangs. No idea why. */
1378 err = intel_ring_pin(ring, engine->i915, I915_GTT_PAGE_SIZE);
1382 GEM_BUG_ON(engine->buffer);
1383 engine->buffer = ring;
1385 err = intel_engine_init_common(engine);
1392 intel_ring_unpin(ring);
1394 intel_ring_free(ring);
1396 intel_engine_cleanup_common(engine);
1400 void intel_engine_cleanup(struct intel_engine_cs *engine)
1402 struct drm_i915_private *dev_priv = engine->i915;
1404 WARN_ON(INTEL_GEN(dev_priv) > 2 &&
1405 (I915_READ_MODE(engine) & MODE_IDLE) == 0);
1407 intel_ring_unpin(engine->buffer);
1408 intel_ring_free(engine->buffer);
1410 if (engine->cleanup)
1411 engine->cleanup(engine);
1413 intel_engine_cleanup_common(engine);
1415 dev_priv->engine[engine->id] = NULL;
1419 void intel_legacy_submission_resume(struct drm_i915_private *dev_priv)
1421 struct intel_engine_cs *engine;
1422 enum intel_engine_id id;
1424 /* Restart from the beginning of the rings for convenience */
1425 for_each_engine(engine, dev_priv, id)
1426 intel_ring_reset(engine->buffer, 0);
1429 static int load_pd_dir(struct i915_request *rq,
1430 const struct i915_hw_ppgtt *ppgtt)
1432 const struct intel_engine_cs * const engine = rq->engine;
1435 cs = intel_ring_begin(rq, 6);
1439 *cs++ = MI_LOAD_REGISTER_IMM(1);
1440 *cs++ = i915_mmio_reg_offset(RING_PP_DIR_DCLV(engine));
1441 *cs++ = PP_DIR_DCLV_2G;
1443 *cs++ = MI_LOAD_REGISTER_IMM(1);
1444 *cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine));
1445 *cs++ = ppgtt->pd.base.ggtt_offset << 10;
1447 intel_ring_advance(rq, cs);
1452 static inline int mi_set_context(struct i915_request *rq, u32 flags)
1454 struct drm_i915_private *i915 = rq->i915;
1455 struct intel_engine_cs *engine = rq->engine;
1456 enum intel_engine_id id;
1457 const int num_rings =
1458 /* Use an extended w/a on gen7 if signalling from other rings */
1459 (HAS_LEGACY_SEMAPHORES(i915) && IS_GEN7(i915)) ?
1460 INTEL_INFO(i915)->num_rings - 1 :
1462 bool force_restore = false;
1466 flags |= MI_MM_SPACE_GTT;
1467 if (IS_HASWELL(i915))
1468 /* These flags are for resource streamer on HSW+ */
1469 flags |= HSW_MI_RS_SAVE_STATE_EN | HSW_MI_RS_RESTORE_STATE_EN;
1471 flags |= MI_SAVE_EXT_STATE_EN | MI_RESTORE_EXT_STATE_EN;
1475 len += 2 + (num_rings ? 4*num_rings + 6 : 0);
1476 if (flags & MI_FORCE_RESTORE) {
1477 GEM_BUG_ON(flags & MI_RESTORE_INHIBIT);
1478 flags &= ~MI_FORCE_RESTORE;
1479 force_restore = true;
1483 cs = intel_ring_begin(rq, len);
1487 /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
1488 if (IS_GEN7(i915)) {
1489 *cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
1491 struct intel_engine_cs *signaller;
1493 *cs++ = MI_LOAD_REGISTER_IMM(num_rings);
1494 for_each_engine(signaller, i915, id) {
1495 if (signaller == engine)
1498 *cs++ = i915_mmio_reg_offset(
1499 RING_PSMI_CTL(signaller->mmio_base));
1500 *cs++ = _MASKED_BIT_ENABLE(
1501 GEN6_PSMI_SLEEP_MSG_DISABLE);
1506 if (force_restore) {
1508 * The HW doesn't handle being told to restore the current
1509 * context very well. Quite often it likes goes to go off and
1510 * sulk, especially when it is meant to be reloading PP_DIR.
1511 * A very simple fix to force the reload is to simply switch
1512 * away from the current context and back again.
1514 * Note that the kernel_context will contain random state
1515 * following the INHIBIT_RESTORE. We accept this since we
1516 * never use the kernel_context state; it is merely a
1517 * placeholder we use to flush other contexts.
1519 *cs++ = MI_SET_CONTEXT;
1520 *cs++ = i915_ggtt_offset(to_intel_context(i915->kernel_context,
1527 *cs++ = MI_SET_CONTEXT;
1528 *cs++ = i915_ggtt_offset(rq->hw_context->state) | flags;
1530 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
1531 * WaMiSetContext_Hang:snb,ivb,vlv
1535 if (IS_GEN7(i915)) {
1537 struct intel_engine_cs *signaller;
1538 i915_reg_t last_reg = {}; /* keep gcc quiet */
1540 *cs++ = MI_LOAD_REGISTER_IMM(num_rings);
1541 for_each_engine(signaller, i915, id) {
1542 if (signaller == engine)
1545 last_reg = RING_PSMI_CTL(signaller->mmio_base);
1546 *cs++ = i915_mmio_reg_offset(last_reg);
1547 *cs++ = _MASKED_BIT_DISABLE(
1548 GEN6_PSMI_SLEEP_MSG_DISABLE);
1551 /* Insert a delay before the next switch! */
1552 *cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
1553 *cs++ = i915_mmio_reg_offset(last_reg);
1554 *cs++ = i915_ggtt_offset(engine->scratch);
1557 *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
1560 intel_ring_advance(rq, cs);
1565 static int remap_l3(struct i915_request *rq, int slice)
1567 u32 *cs, *remap_info = rq->i915->l3_parity.remap_info[slice];
1573 cs = intel_ring_begin(rq, GEN7_L3LOG_SIZE/4 * 2 + 2);
1578 * Note: We do not worry about the concurrent register cacheline hang
1579 * here because no other code should access these registers other than
1580 * at initialization time.
1582 *cs++ = MI_LOAD_REGISTER_IMM(GEN7_L3LOG_SIZE/4);
1583 for (i = 0; i < GEN7_L3LOG_SIZE/4; i++) {
1584 *cs++ = i915_mmio_reg_offset(GEN7_L3LOG(slice, i));
1585 *cs++ = remap_info[i];
1588 intel_ring_advance(rq, cs);
1593 static int switch_context(struct i915_request *rq)
1595 struct intel_engine_cs *engine = rq->engine;
1596 struct i915_gem_context *ctx = rq->gem_context;
1597 struct i915_hw_ppgtt *ppgtt = ctx->ppgtt ?: rq->i915->mm.aliasing_ppgtt;
1598 unsigned int unwind_mm = 0;
1602 lockdep_assert_held(&rq->i915->drm.struct_mutex);
1603 GEM_BUG_ON(HAS_EXECLISTS(rq->i915));
1606 ret = load_pd_dir(rq, ppgtt);
1610 if (intel_engine_flag(engine) & ppgtt->pd_dirty_rings) {
1611 unwind_mm = intel_engine_flag(engine);
1612 ppgtt->pd_dirty_rings &= ~unwind_mm;
1613 hw_flags = MI_FORCE_RESTORE;
1617 if (rq->hw_context->state) {
1618 GEM_BUG_ON(engine->id != RCS);
1621 * The kernel context(s) is treated as pure scratch and is not
1622 * expected to retain any state (as we sacrifice it during
1623 * suspend and on resume it may be corrupted). This is ok,
1624 * as nothing actually executes using the kernel context; it
1625 * is purely used for flushing user contexts.
1627 if (i915_gem_context_is_kernel(ctx))
1628 hw_flags = MI_RESTORE_INHIBIT;
1630 ret = mi_set_context(rq, hw_flags);
1635 if (ctx->remap_slice) {
1636 for (i = 0; i < MAX_L3_SLICES; i++) {
1637 if (!(ctx->remap_slice & BIT(i)))
1640 ret = remap_l3(rq, i);
1645 ctx->remap_slice = 0;
1652 ppgtt->pd_dirty_rings |= unwind_mm;
1657 static int ring_request_alloc(struct i915_request *request)
1661 GEM_BUG_ON(!request->hw_context->pin_count);
1663 /* Flush enough space to reduce the likelihood of waiting after
1664 * we start building the request - in which case we will just
1665 * have to repeat work.
1667 request->reserved_space += LEGACY_REQUEST_SIZE;
1669 ret = intel_ring_wait_for_space(request->ring, request->reserved_space);
1673 ret = switch_context(request);
1677 request->reserved_space -= LEGACY_REQUEST_SIZE;
1681 static noinline int wait_for_space(struct intel_ring *ring, unsigned int bytes)
1683 struct i915_request *target;
1686 lockdep_assert_held(&ring->vma->vm->i915->drm.struct_mutex);
1688 if (intel_ring_update_space(ring) >= bytes)
1691 GEM_BUG_ON(list_empty(&ring->request_list));
1692 list_for_each_entry(target, &ring->request_list, ring_link) {
1693 /* Would completion of this request free enough space? */
1694 if (bytes <= __intel_ring_space(target->postfix,
1695 ring->emit, ring->size))
1699 if (WARN_ON(&target->ring_link == &ring->request_list))
1702 timeout = i915_request_wait(target,
1703 I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
1704 MAX_SCHEDULE_TIMEOUT);
1708 i915_request_retire_upto(target);
1710 intel_ring_update_space(ring);
1711 GEM_BUG_ON(ring->space < bytes);
1715 int intel_ring_wait_for_space(struct intel_ring *ring, unsigned int bytes)
1717 GEM_BUG_ON(bytes > ring->effective_size);
1718 if (unlikely(bytes > ring->effective_size - ring->emit))
1719 bytes += ring->size - ring->emit;
1721 if (unlikely(bytes > ring->space)) {
1722 int ret = wait_for_space(ring, bytes);
1727 GEM_BUG_ON(ring->space < bytes);
1731 u32 *intel_ring_begin(struct i915_request *rq, unsigned int num_dwords)
1733 struct intel_ring *ring = rq->ring;
1734 const unsigned int remain_usable = ring->effective_size - ring->emit;
1735 const unsigned int bytes = num_dwords * sizeof(u32);
1736 unsigned int need_wrap = 0;
1737 unsigned int total_bytes;
1740 /* Packets must be qword aligned. */
1741 GEM_BUG_ON(num_dwords & 1);
1743 total_bytes = bytes + rq->reserved_space;
1744 GEM_BUG_ON(total_bytes > ring->effective_size);
1746 if (unlikely(total_bytes > remain_usable)) {
1747 const int remain_actual = ring->size - ring->emit;
1749 if (bytes > remain_usable) {
1751 * Not enough space for the basic request. So need to
1752 * flush out the remainder and then wait for
1755 total_bytes += remain_actual;
1756 need_wrap = remain_actual | 1;
1759 * The base request will fit but the reserved space
1760 * falls off the end. So we don't need an immediate
1761 * wrap and only need to effectively wait for the
1762 * reserved size from the start of ringbuffer.
1764 total_bytes = rq->reserved_space + remain_actual;
1768 if (unlikely(total_bytes > ring->space)) {
1772 * Space is reserved in the ringbuffer for finalising the
1773 * request, as that cannot be allowed to fail. During request
1774 * finalisation, reserved_space is set to 0 to stop the
1775 * overallocation and the assumption is that then we never need
1776 * to wait (which has the risk of failing with EINTR).
1778 * See also i915_request_alloc() and i915_request_add().
1780 GEM_BUG_ON(!rq->reserved_space);
1782 ret = wait_for_space(ring, total_bytes);
1784 return ERR_PTR(ret);
1787 if (unlikely(need_wrap)) {
1789 GEM_BUG_ON(need_wrap > ring->space);
1790 GEM_BUG_ON(ring->emit + need_wrap > ring->size);
1791 GEM_BUG_ON(!IS_ALIGNED(need_wrap, sizeof(u64)));
1793 /* Fill the tail with MI_NOOP */
1794 memset64(ring->vaddr + ring->emit, 0, need_wrap / sizeof(u64));
1795 ring->space -= need_wrap;
1799 GEM_BUG_ON(ring->emit > ring->size - bytes);
1800 GEM_BUG_ON(ring->space < bytes);
1801 cs = ring->vaddr + ring->emit;
1802 GEM_DEBUG_EXEC(memset32(cs, POISON_INUSE, bytes / sizeof(*cs)));
1803 ring->emit += bytes;
1804 ring->space -= bytes;
1809 /* Align the ring tail to a cacheline boundary */
1810 int intel_ring_cacheline_align(struct i915_request *rq)
1815 num_dwords = (rq->ring->emit & (CACHELINE_BYTES - 1)) / sizeof(u32);
1816 if (num_dwords == 0)
1819 num_dwords = CACHELINE_DWORDS - num_dwords;
1820 GEM_BUG_ON(num_dwords & 1);
1822 cs = intel_ring_begin(rq, num_dwords);
1826 memset64(cs, (u64)MI_NOOP << 32 | MI_NOOP, num_dwords / 2);
1827 intel_ring_advance(rq, cs);
1829 GEM_BUG_ON(rq->ring->emit & (CACHELINE_BYTES - 1));
1833 static void gen6_bsd_submit_request(struct i915_request *request)
1835 struct drm_i915_private *dev_priv = request->i915;
1837 intel_uncore_forcewake_get(dev_priv, FORCEWAKE_ALL);
1839 /* Every tail move must follow the sequence below */
1841 /* Disable notification that the ring is IDLE. The GT
1842 * will then assume that it is busy and bring it out of rc6.
1844 I915_WRITE_FW(GEN6_BSD_SLEEP_PSMI_CONTROL,
1845 _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1847 /* Clear the context id. Here be magic! */
1848 I915_WRITE64_FW(GEN6_BSD_RNCID, 0x0);
1850 /* Wait for the ring not to be idle, i.e. for it to wake up. */
1851 if (__intel_wait_for_register_fw(dev_priv,
1852 GEN6_BSD_SLEEP_PSMI_CONTROL,
1853 GEN6_BSD_SLEEP_INDICATOR,
1856 DRM_ERROR("timed out waiting for the BSD ring to wake up\n");
1858 /* Now that the ring is fully powered up, update the tail */
1859 i9xx_submit_request(request);
1861 /* Let the ring send IDLE messages to the GT again,
1862 * and so let it sleep to conserve power when idle.
1864 I915_WRITE_FW(GEN6_BSD_SLEEP_PSMI_CONTROL,
1865 _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1867 intel_uncore_forcewake_put(dev_priv, FORCEWAKE_ALL);
1870 static int gen6_bsd_ring_flush(struct i915_request *rq, u32 mode)
1874 cs = intel_ring_begin(rq, 4);
1880 /* We always require a command barrier so that subsequent
1881 * commands, such as breadcrumb interrupts, are strictly ordered
1882 * wrt the contents of the write cache being flushed to memory
1883 * (and thus being coherent from the CPU).
1885 cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1888 * Bspec vol 1c.5 - video engine command streamer:
1889 * "If ENABLED, all TLBs will be invalidated once the flush
1890 * operation is complete. This bit is only valid when the
1891 * Post-Sync Operation field is a value of 1h or 3h."
1893 if (mode & EMIT_INVALIDATE)
1894 cmd |= MI_INVALIDATE_TLB | MI_INVALIDATE_BSD;
1897 *cs++ = I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT;
1900 intel_ring_advance(rq, cs);
1905 hsw_emit_bb_start(struct i915_request *rq,
1906 u64 offset, u32 len,
1907 unsigned int dispatch_flags)
1911 cs = intel_ring_begin(rq, 2);
1915 *cs++ = MI_BATCH_BUFFER_START | (dispatch_flags & I915_DISPATCH_SECURE ?
1916 0 : MI_BATCH_PPGTT_HSW | MI_BATCH_NON_SECURE_HSW) |
1917 (dispatch_flags & I915_DISPATCH_RS ?
1918 MI_BATCH_RESOURCE_STREAMER : 0);
1919 /* bit0-7 is the length on GEN6+ */
1921 intel_ring_advance(rq, cs);
1927 gen6_emit_bb_start(struct i915_request *rq,
1928 u64 offset, u32 len,
1929 unsigned int dispatch_flags)
1933 cs = intel_ring_begin(rq, 2);
1937 *cs++ = MI_BATCH_BUFFER_START | (dispatch_flags & I915_DISPATCH_SECURE ?
1938 0 : MI_BATCH_NON_SECURE_I965);
1939 /* bit0-7 is the length on GEN6+ */
1941 intel_ring_advance(rq, cs);
1946 /* Blitter support (SandyBridge+) */
1948 static int gen6_ring_flush(struct i915_request *rq, u32 mode)
1952 cs = intel_ring_begin(rq, 4);
1958 /* We always require a command barrier so that subsequent
1959 * commands, such as breadcrumb interrupts, are strictly ordered
1960 * wrt the contents of the write cache being flushed to memory
1961 * (and thus being coherent from the CPU).
1963 cmd |= MI_FLUSH_DW_STORE_INDEX | MI_FLUSH_DW_OP_STOREDW;
1966 * Bspec vol 1c.3 - blitter engine command streamer:
1967 * "If ENABLED, all TLBs will be invalidated once the flush
1968 * operation is complete. This bit is only valid when the
1969 * Post-Sync Operation field is a value of 1h or 3h."
1971 if (mode & EMIT_INVALIDATE)
1972 cmd |= MI_INVALIDATE_TLB;
1974 *cs++ = I915_GEM_HWS_SCRATCH_ADDR | MI_FLUSH_DW_USE_GTT;
1977 intel_ring_advance(rq, cs);
1982 static void intel_ring_init_semaphores(struct drm_i915_private *dev_priv,
1983 struct intel_engine_cs *engine)
1987 if (!HAS_LEGACY_SEMAPHORES(dev_priv))
1990 GEM_BUG_ON(INTEL_GEN(dev_priv) < 6);
1991 engine->semaphore.sync_to = gen6_ring_sync_to;
1992 engine->semaphore.signal = gen6_signal;
1995 * The current semaphore is only applied on pre-gen8
1996 * platform. And there is no VCS2 ring on the pre-gen8
1997 * platform. So the semaphore between RCS and VCS2 is
1998 * initialized as INVALID.
2000 for (i = 0; i < GEN6_NUM_SEMAPHORES; i++) {
2001 static const struct {
2003 i915_reg_t mbox_reg;
2004 } sem_data[GEN6_NUM_SEMAPHORES][GEN6_NUM_SEMAPHORES] = {
2006 [VCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_RV, .mbox_reg = GEN6_VRSYNC },
2007 [BCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_RB, .mbox_reg = GEN6_BRSYNC },
2008 [VECS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_RVE, .mbox_reg = GEN6_VERSYNC },
2011 [RCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_VR, .mbox_reg = GEN6_RVSYNC },
2012 [BCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_VB, .mbox_reg = GEN6_BVSYNC },
2013 [VECS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_VVE, .mbox_reg = GEN6_VEVSYNC },
2016 [RCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_BR, .mbox_reg = GEN6_RBSYNC },
2017 [VCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_BV, .mbox_reg = GEN6_VBSYNC },
2018 [VECS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_BVE, .mbox_reg = GEN6_VEBSYNC },
2021 [RCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_VER, .mbox_reg = GEN6_RVESYNC },
2022 [VCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_VEV, .mbox_reg = GEN6_VVESYNC },
2023 [BCS_HW] = { .wait_mbox = MI_SEMAPHORE_SYNC_VEB, .mbox_reg = GEN6_BVESYNC },
2027 i915_reg_t mbox_reg;
2029 if (i == engine->hw_id) {
2030 wait_mbox = MI_SEMAPHORE_SYNC_INVALID;
2031 mbox_reg = GEN6_NOSYNC;
2033 wait_mbox = sem_data[engine->hw_id][i].wait_mbox;
2034 mbox_reg = sem_data[engine->hw_id][i].mbox_reg;
2037 engine->semaphore.mbox.wait[i] = wait_mbox;
2038 engine->semaphore.mbox.signal[i] = mbox_reg;
2042 static void intel_ring_init_irq(struct drm_i915_private *dev_priv,
2043 struct intel_engine_cs *engine)
2045 if (INTEL_GEN(dev_priv) >= 6) {
2046 engine->irq_enable = gen6_irq_enable;
2047 engine->irq_disable = gen6_irq_disable;
2048 engine->irq_seqno_barrier = gen6_seqno_barrier;
2049 } else if (INTEL_GEN(dev_priv) >= 5) {
2050 engine->irq_enable = gen5_irq_enable;
2051 engine->irq_disable = gen5_irq_disable;
2052 engine->irq_seqno_barrier = gen5_seqno_barrier;
2053 } else if (INTEL_GEN(dev_priv) >= 3) {
2054 engine->irq_enable = i9xx_irq_enable;
2055 engine->irq_disable = i9xx_irq_disable;
2057 engine->irq_enable = i8xx_irq_enable;
2058 engine->irq_disable = i8xx_irq_disable;
2062 static void i9xx_set_default_submission(struct intel_engine_cs *engine)
2064 engine->submit_request = i9xx_submit_request;
2065 engine->cancel_requests = cancel_requests;
2067 engine->park = NULL;
2068 engine->unpark = NULL;
2071 static void gen6_bsd_set_default_submission(struct intel_engine_cs *engine)
2073 i9xx_set_default_submission(engine);
2074 engine->submit_request = gen6_bsd_submit_request;
2077 static void intel_ring_default_vfuncs(struct drm_i915_private *dev_priv,
2078 struct intel_engine_cs *engine)
2080 /* gen8+ are only supported with execlists */
2081 GEM_BUG_ON(INTEL_GEN(dev_priv) >= 8);
2083 intel_ring_init_irq(dev_priv, engine);
2084 intel_ring_init_semaphores(dev_priv, engine);
2086 engine->init_hw = init_ring_common;
2087 engine->reset.prepare = reset_prepare;
2088 engine->reset.reset = reset_ring;
2089 engine->reset.finish = reset_finish;
2091 engine->context_pin = intel_ring_context_pin;
2092 engine->request_alloc = ring_request_alloc;
2094 engine->emit_breadcrumb = i9xx_emit_breadcrumb;
2095 engine->emit_breadcrumb_sz = i9xx_emit_breadcrumb_sz;
2096 if (HAS_LEGACY_SEMAPHORES(dev_priv)) {
2099 engine->emit_breadcrumb = gen6_sema_emit_breadcrumb;
2101 num_rings = INTEL_INFO(dev_priv)->num_rings - 1;
2102 engine->emit_breadcrumb_sz += num_rings * 3;
2104 engine->emit_breadcrumb_sz++;
2107 engine->set_default_submission = i9xx_set_default_submission;
2109 if (INTEL_GEN(dev_priv) >= 6)
2110 engine->emit_bb_start = gen6_emit_bb_start;
2111 else if (INTEL_GEN(dev_priv) >= 4)
2112 engine->emit_bb_start = i965_emit_bb_start;
2113 else if (IS_I830(dev_priv) || IS_I845G(dev_priv))
2114 engine->emit_bb_start = i830_emit_bb_start;
2116 engine->emit_bb_start = i915_emit_bb_start;
2119 int intel_init_render_ring_buffer(struct intel_engine_cs *engine)
2121 struct drm_i915_private *dev_priv = engine->i915;
2124 intel_ring_default_vfuncs(dev_priv, engine);
2126 if (HAS_L3_DPF(dev_priv))
2127 engine->irq_keep_mask = GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
2129 engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
2131 if (INTEL_GEN(dev_priv) >= 6) {
2132 engine->init_context = intel_rcs_ctx_init;
2133 engine->emit_flush = gen7_render_ring_flush;
2134 if (IS_GEN6(dev_priv))
2135 engine->emit_flush = gen6_render_ring_flush;
2136 } else if (IS_GEN5(dev_priv)) {
2137 engine->emit_flush = gen4_render_ring_flush;
2139 if (INTEL_GEN(dev_priv) < 4)
2140 engine->emit_flush = gen2_render_ring_flush;
2142 engine->emit_flush = gen4_render_ring_flush;
2143 engine->irq_enable_mask = I915_USER_INTERRUPT;
2146 if (IS_HASWELL(dev_priv))
2147 engine->emit_bb_start = hsw_emit_bb_start;
2149 engine->init_hw = init_render_ring;
2151 ret = intel_init_ring_buffer(engine);
2155 if (INTEL_GEN(dev_priv) >= 6) {
2156 ret = intel_engine_create_scratch(engine, PAGE_SIZE);
2159 } else if (HAS_BROKEN_CS_TLB(dev_priv)) {
2160 ret = intel_engine_create_scratch(engine, I830_WA_SIZE);
2168 int intel_init_bsd_ring_buffer(struct intel_engine_cs *engine)
2170 struct drm_i915_private *dev_priv = engine->i915;
2172 intel_ring_default_vfuncs(dev_priv, engine);
2174 if (INTEL_GEN(dev_priv) >= 6) {
2175 /* gen6 bsd needs a special wa for tail updates */
2176 if (IS_GEN6(dev_priv))
2177 engine->set_default_submission = gen6_bsd_set_default_submission;
2178 engine->emit_flush = gen6_bsd_ring_flush;
2179 engine->irq_enable_mask = GT_BSD_USER_INTERRUPT;
2181 engine->emit_flush = bsd_ring_flush;
2182 if (IS_GEN5(dev_priv))
2183 engine->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
2185 engine->irq_enable_mask = I915_BSD_USER_INTERRUPT;
2188 return intel_init_ring_buffer(engine);
2191 int intel_init_blt_ring_buffer(struct intel_engine_cs *engine)
2193 struct drm_i915_private *dev_priv = engine->i915;
2195 intel_ring_default_vfuncs(dev_priv, engine);
2197 engine->emit_flush = gen6_ring_flush;
2198 engine->irq_enable_mask = GT_BLT_USER_INTERRUPT;
2200 return intel_init_ring_buffer(engine);
2203 int intel_init_vebox_ring_buffer(struct intel_engine_cs *engine)
2205 struct drm_i915_private *dev_priv = engine->i915;
2207 intel_ring_default_vfuncs(dev_priv, engine);
2209 engine->emit_flush = gen6_ring_flush;
2210 engine->irq_enable_mask = PM_VEBOX_USER_INTERRUPT;
2211 engine->irq_enable = hsw_vebox_irq_enable;
2212 engine->irq_disable = hsw_vebox_irq_disable;
2214 return intel_init_ring_buffer(engine);