1 // SPDX-License-Identifier: MIT
3 * Copyright © 2008-2021 Intel Corporation
6 #include "gen2_engine_cs.h"
7 #include "gen6_engine_cs.h"
8 #include "gen6_ppgtt.h"
9 #include "gen7_renderclear.h"
11 #include "i915_mitigations.h"
12 #include "intel_breadcrumbs.h"
13 #include "intel_context.h"
15 #include "intel_gt_irq.h"
16 #include "intel_reset.h"
17 #include "intel_ring.h"
18 #include "shmem_utils.h"
20 /* Rough estimate of the typical request size, performing a flush,
21 * set-context and then emitting the batch.
23 #define LEGACY_REQUEST_SIZE 200
25 static void set_hwstam(struct intel_engine_cs *engine, u32 mask)
28 * Keep the render interrupt unmasked as this papers over
29 * lost interrupts following a reset.
31 if (engine->class == RENDER_CLASS) {
32 if (GRAPHICS_VER(engine->i915) >= 6)
35 mask &= ~I915_USER_INTERRUPT;
38 intel_engine_set_hwsp_writemask(engine, mask);
41 static void set_hws_pga(struct intel_engine_cs *engine, phys_addr_t phys)
45 addr = lower_32_bits(phys);
46 if (GRAPHICS_VER(engine->i915) >= 4)
47 addr |= (phys >> 28) & 0xf0;
49 intel_uncore_write(engine->uncore, HWS_PGA, addr);
52 static struct page *status_page(struct intel_engine_cs *engine)
54 struct drm_i915_gem_object *obj = engine->status_page.vma->obj;
56 GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
57 return sg_page(obj->mm.pages->sgl);
60 static void ring_setup_phys_status_page(struct intel_engine_cs *engine)
62 set_hws_pga(engine, PFN_PHYS(page_to_pfn(status_page(engine))));
63 set_hwstam(engine, ~0u);
66 static void set_hwsp(struct intel_engine_cs *engine, u32 offset)
71 * The ring status page addresses are no longer next to the rest of
72 * the ring registers as of gen7.
74 if (GRAPHICS_VER(engine->i915) == 7) {
77 * No more rings exist on Gen7. Default case is only to shut up
78 * gcc switch check warning.
81 GEM_BUG_ON(engine->id);
84 hwsp = RENDER_HWS_PGA_GEN7;
87 hwsp = BLT_HWS_PGA_GEN7;
90 hwsp = BSD_HWS_PGA_GEN7;
93 hwsp = VEBOX_HWS_PGA_GEN7;
96 } else if (GRAPHICS_VER(engine->i915) == 6) {
97 hwsp = RING_HWS_PGA_GEN6(engine->mmio_base);
99 hwsp = RING_HWS_PGA(engine->mmio_base);
102 intel_uncore_write_fw(engine->uncore, hwsp, offset);
103 intel_uncore_posting_read_fw(engine->uncore, hwsp);
106 static void flush_cs_tlb(struct intel_engine_cs *engine)
108 if (!IS_GRAPHICS_VER(engine->i915, 6, 7))
111 /* ring should be idle before issuing a sync flush*/
112 GEM_DEBUG_WARN_ON((ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
114 ENGINE_WRITE_FW(engine, RING_INSTPM,
115 _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE |
117 if (__intel_wait_for_register_fw(engine->uncore,
118 RING_INSTPM(engine->mmio_base),
119 INSTPM_SYNC_FLUSH, 0,
122 "wait for SyncFlush to complete for TLB invalidation timed out\n");
125 static void ring_setup_status_page(struct intel_engine_cs *engine)
127 set_hwsp(engine, i915_ggtt_offset(engine->status_page.vma));
128 set_hwstam(engine, ~0u);
130 flush_cs_tlb(engine);
133 static struct i915_address_space *vm_alias(struct i915_address_space *vm)
135 if (i915_is_ggtt(vm))
136 vm = &i915_vm_to_ggtt(vm)->alias->vm;
141 static u32 pp_dir(struct i915_address_space *vm)
143 return to_gen6_ppgtt(i915_vm_to_ppgtt(vm))->pp_dir;
146 static void set_pp_dir(struct intel_engine_cs *engine)
148 struct i915_address_space *vm = vm_alias(engine->gt->vm);
153 ENGINE_WRITE_FW(engine, RING_PP_DIR_DCLV, PP_DIR_DCLV_2G);
154 ENGINE_WRITE_FW(engine, RING_PP_DIR_BASE, pp_dir(vm));
156 if (GRAPHICS_VER(engine->i915) >= 7) {
157 ENGINE_WRITE_FW(engine,
159 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
163 static bool stop_ring(struct intel_engine_cs *engine)
165 /* Empty the ring by skipping to the end */
166 ENGINE_WRITE_FW(engine, RING_HEAD, ENGINE_READ_FW(engine, RING_TAIL));
167 ENGINE_POSTING_READ(engine, RING_HEAD);
169 /* The ring must be empty before it is disabled */
170 ENGINE_WRITE_FW(engine, RING_CTL, 0);
171 ENGINE_POSTING_READ(engine, RING_CTL);
173 /* Then reset the disabled ring */
174 ENGINE_WRITE_FW(engine, RING_HEAD, 0);
175 ENGINE_WRITE_FW(engine, RING_TAIL, 0);
177 return (ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR) == 0;
180 static int xcs_resume(struct intel_engine_cs *engine)
182 struct intel_ring *ring = engine->legacy.ring;
184 ENGINE_TRACE(engine, "ring:{HEAD:%04x, TAIL:%04x}\n",
185 ring->head, ring->tail);
188 * Double check the ring is empty & disabled before we resume. Called
189 * from atomic context during PCI probe, so _hardirq().
191 intel_synchronize_hardirq(engine->i915);
192 if (!stop_ring(engine))
195 if (HWS_NEEDS_PHYSICAL(engine->i915))
196 ring_setup_phys_status_page(engine);
198 ring_setup_status_page(engine);
200 intel_breadcrumbs_reset(engine->breadcrumbs);
202 /* Enforce ordering by reading HEAD register back */
203 ENGINE_POSTING_READ(engine, RING_HEAD);
206 * Initialize the ring. This must happen _after_ we've cleared the ring
207 * registers with the above sequence (the readback of the HEAD registers
208 * also enforces ordering), otherwise the hw might lose the new ring
211 ENGINE_WRITE_FW(engine, RING_START, i915_ggtt_offset(ring->vma));
213 /* Check that the ring offsets point within the ring! */
214 GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->head));
215 GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->tail));
216 intel_ring_update_space(ring);
220 /* First wake the ring up to an empty/idle ring */
221 ENGINE_WRITE_FW(engine, RING_HEAD, ring->head);
222 ENGINE_WRITE_FW(engine, RING_TAIL, ring->head);
223 ENGINE_POSTING_READ(engine, RING_TAIL);
225 ENGINE_WRITE_FW(engine, RING_CTL,
226 RING_CTL_SIZE(ring->size) | RING_VALID);
228 /* If the head is still not zero, the ring is dead */
229 if (__intel_wait_for_register_fw(engine->uncore,
230 RING_CTL(engine->mmio_base),
231 RING_VALID, RING_VALID,
235 if (GRAPHICS_VER(engine->i915) > 2)
236 ENGINE_WRITE_FW(engine,
237 RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
239 /* Now awake, let it get started */
240 if (ring->tail != ring->head) {
241 ENGINE_WRITE_FW(engine, RING_TAIL, ring->tail);
242 ENGINE_POSTING_READ(engine, RING_TAIL);
245 /* Papering over lost _interrupts_ immediately following the restart */
246 intel_engine_signal_breadcrumbs(engine);
250 drm_err(&engine->i915->drm,
251 "%s initialization failed; "
252 "ctl %08x (valid? %d) head %08x [%08x] tail %08x [%08x] start %08x [expected %08x]\n",
254 ENGINE_READ(engine, RING_CTL),
255 ENGINE_READ(engine, RING_CTL) & RING_VALID,
256 ENGINE_READ(engine, RING_HEAD), ring->head,
257 ENGINE_READ(engine, RING_TAIL), ring->tail,
258 ENGINE_READ(engine, RING_START),
259 i915_ggtt_offset(ring->vma));
263 static void sanitize_hwsp(struct intel_engine_cs *engine)
265 struct intel_timeline *tl;
267 list_for_each_entry(tl, &engine->status_page.timelines, engine_link)
268 intel_timeline_reset_seqno(tl);
271 static void xcs_sanitize(struct intel_engine_cs *engine)
274 * Poison residual state on resume, in case the suspend didn't!
276 * We have to assume that across suspend/resume (or other loss
277 * of control) that the contents of our pinned buffers has been
278 * lost, replaced by garbage. Since this doesn't always happen,
279 * let's poison such state so that we more quickly spot when
280 * we falsely assume it has been preserved.
282 if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
283 memset(engine->status_page.addr, POISON_INUSE, PAGE_SIZE);
286 * The kernel_context HWSP is stored in the status_page. As above,
287 * that may be lost on resume/initialisation, and so we need to
288 * reset the value in the HWSP.
290 sanitize_hwsp(engine);
292 /* And scrub the dirty cachelines for the HWSP */
293 clflush_cache_range(engine->status_page.addr, PAGE_SIZE);
296 static void reset_prepare(struct intel_engine_cs *engine)
299 * We stop engines, otherwise we might get failed reset and a
300 * dead gpu (on elk). Also as modern gpu as kbl can suffer
301 * from system hang if batchbuffer is progressing when
302 * the reset is issued, regardless of READY_TO_RESET ack.
303 * Thus assume it is best to stop engines on all gens
304 * where we have a gpu reset.
306 * WaKBLVECSSemaphoreWaitPoll:kbl (on ALL_ENGINES)
308 * WaMediaResetMainRingCleanup:ctg,elk (presumably)
309 * WaClearRingBufHeadRegAtInit:ctg,elk
311 * FIXME: Wa for more modern gens needs to be validated
313 ENGINE_TRACE(engine, "\n");
314 intel_engine_stop_cs(engine);
316 if (!stop_ring(engine)) {
317 /* G45 ring initialization often fails to reset head to zero */
319 "HEAD not reset to zero, "
320 "{ CTL:%08x, HEAD:%08x, TAIL:%08x, START:%08x }\n",
321 ENGINE_READ_FW(engine, RING_CTL),
322 ENGINE_READ_FW(engine, RING_HEAD),
323 ENGINE_READ_FW(engine, RING_TAIL),
324 ENGINE_READ_FW(engine, RING_START));
325 if (!stop_ring(engine)) {
326 drm_err(&engine->i915->drm,
327 "failed to set %s head to zero "
328 "ctl %08x head %08x tail %08x start %08x\n",
330 ENGINE_READ_FW(engine, RING_CTL),
331 ENGINE_READ_FW(engine, RING_HEAD),
332 ENGINE_READ_FW(engine, RING_TAIL),
333 ENGINE_READ_FW(engine, RING_START));
338 static void reset_rewind(struct intel_engine_cs *engine, bool stalled)
340 struct i915_request *pos, *rq;
345 spin_lock_irqsave(&engine->sched_engine->lock, flags);
347 list_for_each_entry(pos, &engine->sched_engine->requests, sched.link) {
348 if (!__i915_request_is_complete(pos)) {
356 * The guilty request will get skipped on a hung engine.
358 * Users of client default contexts do not rely on logical
359 * state preserved between batches so it is safe to execute
360 * queued requests following the hang. Non default contexts
361 * rely on preserved state, so skipping a batch loses the
362 * evolution of the state and it needs to be considered corrupted.
363 * Executing more queued batches on top of corrupted state is
364 * risky. But we take the risk by trying to advance through
365 * the queued requests in order to make the client behaviour
366 * more predictable around resets, by not throwing away random
367 * amount of batches it has prepared for execution. Sophisticated
368 * clients can use gem_reset_stats_ioctl and dma fence status
369 * (exported via sync_file info ioctl on explicit fences) to observe
370 * when it loses the context state and should rebuild accordingly.
372 * The context ban, and ultimately the client ban, mechanism are safety
373 * valves if client submission ends up resulting in nothing more than
379 * Try to restore the logical GPU state to match the
380 * continuation of the request queue. If we skip the
381 * context/PD restore, then the next request may try to execute
382 * assuming that its context is valid and loaded on the GPU and
383 * so may try to access invalid memory, prompting repeated GPU
386 * If the request was guilty, we still restore the logical
387 * state in case the next request requires it (e.g. the
388 * aliasing ppgtt), but skip over the hung batch.
390 * If the request was innocent, we try to replay the request
391 * with the restored context.
393 __i915_request_reset(rq, stalled);
395 GEM_BUG_ON(rq->ring != engine->legacy.ring);
398 head = engine->legacy.ring->tail;
400 engine->legacy.ring->head = intel_ring_wrap(engine->legacy.ring, head);
402 spin_unlock_irqrestore(&engine->sched_engine->lock, flags);
405 static void reset_finish(struct intel_engine_cs *engine)
409 static void reset_cancel(struct intel_engine_cs *engine)
411 struct i915_request *request;
414 spin_lock_irqsave(&engine->sched_engine->lock, flags);
416 /* Mark all submitted requests as skipped. */
417 list_for_each_entry(request, &engine->sched_engine->requests, sched.link)
418 i915_request_put(i915_request_mark_eio(request));
419 intel_engine_signal_breadcrumbs(engine);
421 /* Remaining _unready_ requests will be nop'ed when submitted */
423 spin_unlock_irqrestore(&engine->sched_engine->lock, flags);
426 static void i9xx_submit_request(struct i915_request *request)
428 i915_request_submit(request);
429 wmb(); /* paranoid flush writes out of the WCB before mmio */
431 ENGINE_WRITE(request->engine, RING_TAIL,
432 intel_ring_set_tail(request->ring, request->tail));
435 static void __ring_context_fini(struct intel_context *ce)
437 i915_vma_put(ce->state);
440 static void ring_context_destroy(struct kref *ref)
442 struct intel_context *ce = container_of(ref, typeof(*ce), ref);
444 GEM_BUG_ON(intel_context_is_pinned(ce));
447 __ring_context_fini(ce);
449 intel_context_fini(ce);
450 intel_context_free(ce);
453 static int ring_context_init_default_state(struct intel_context *ce,
454 struct i915_gem_ww_ctx *ww)
456 struct drm_i915_gem_object *obj = ce->state->obj;
459 vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
461 return PTR_ERR(vaddr);
463 shmem_read(ce->engine->default_state, 0,
464 vaddr, ce->engine->context_size);
466 i915_gem_object_flush_map(obj);
467 __i915_gem_object_release_map(obj);
469 __set_bit(CONTEXT_VALID_BIT, &ce->flags);
473 static int ring_context_pre_pin(struct intel_context *ce,
474 struct i915_gem_ww_ctx *ww,
477 struct i915_address_space *vm;
480 if (ce->engine->default_state &&
481 !test_bit(CONTEXT_VALID_BIT, &ce->flags)) {
482 err = ring_context_init_default_state(ce, ww);
487 vm = vm_alias(ce->vm);
489 err = gen6_ppgtt_pin(i915_vm_to_ppgtt((vm)), ww);
494 static void __context_unpin_ppgtt(struct intel_context *ce)
496 struct i915_address_space *vm;
498 vm = vm_alias(ce->vm);
500 gen6_ppgtt_unpin(i915_vm_to_ppgtt(vm));
503 static void ring_context_unpin(struct intel_context *ce)
507 static void ring_context_post_unpin(struct intel_context *ce)
509 __context_unpin_ppgtt(ce);
512 static struct i915_vma *
513 alloc_context_vma(struct intel_engine_cs *engine)
515 struct drm_i915_private *i915 = engine->i915;
516 struct drm_i915_gem_object *obj;
517 struct i915_vma *vma;
520 obj = i915_gem_object_create_shmem(i915, engine->context_size);
522 return ERR_CAST(obj);
525 * Try to make the context utilize L3 as well as LLC.
527 * On VLV we don't have L3 controls in the PTEs so we
528 * shouldn't touch the cache level, especially as that
529 * would make the object snooped which might have a
530 * negative performance impact.
532 * Snooping is required on non-llc platforms in execlist
533 * mode, but since all GGTT accesses use PAT entry 0 we
534 * get snooping anyway regardless of cache_level.
536 * This is only applicable for Ivy Bridge devices since
537 * later platforms don't have L3 control bits in the PTE.
539 if (IS_IVYBRIDGE(i915))
540 i915_gem_object_set_cache_coherency(obj, I915_CACHE_L3_LLC);
542 vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
551 i915_gem_object_put(obj);
555 static int ring_context_alloc(struct intel_context *ce)
557 struct intel_engine_cs *engine = ce->engine;
559 /* One ringbuffer to rule them all */
560 GEM_BUG_ON(!engine->legacy.ring);
561 ce->ring = engine->legacy.ring;
562 ce->timeline = intel_timeline_get(engine->legacy.timeline);
564 GEM_BUG_ON(ce->state);
565 if (engine->context_size) {
566 struct i915_vma *vma;
568 vma = alloc_context_vma(engine);
578 static int ring_context_pin(struct intel_context *ce, void *unused)
583 static void ring_context_reset(struct intel_context *ce)
585 intel_ring_reset(ce->ring, ce->ring->emit);
586 clear_bit(CONTEXT_VALID_BIT, &ce->flags);
589 static const struct intel_context_ops ring_context_ops = {
590 .alloc = ring_context_alloc,
592 .pre_pin = ring_context_pre_pin,
593 .pin = ring_context_pin,
594 .unpin = ring_context_unpin,
595 .post_unpin = ring_context_post_unpin,
597 .enter = intel_context_enter_engine,
598 .exit = intel_context_exit_engine,
600 .reset = ring_context_reset,
601 .destroy = ring_context_destroy,
604 static int load_pd_dir(struct i915_request *rq,
605 struct i915_address_space *vm,
608 const struct intel_engine_cs * const engine = rq->engine;
611 cs = intel_ring_begin(rq, 12);
615 *cs++ = MI_LOAD_REGISTER_IMM(1);
616 *cs++ = i915_mmio_reg_offset(RING_PP_DIR_DCLV(engine->mmio_base));
619 *cs++ = MI_LOAD_REGISTER_IMM(1);
620 *cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
623 /* Stall until the page table load is complete? */
624 *cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
625 *cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
626 *cs++ = intel_gt_scratch_offset(engine->gt,
627 INTEL_GT_SCRATCH_FIELD_DEFAULT);
629 *cs++ = MI_LOAD_REGISTER_IMM(1);
630 *cs++ = i915_mmio_reg_offset(RING_INSTPM(engine->mmio_base));
631 *cs++ = _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE);
633 intel_ring_advance(rq, cs);
635 return rq->engine->emit_flush(rq, EMIT_FLUSH);
638 static int mi_set_context(struct i915_request *rq,
639 struct intel_context *ce,
642 struct intel_engine_cs *engine = rq->engine;
643 struct drm_i915_private *i915 = engine->i915;
644 enum intel_engine_id id;
645 const int num_engines =
646 IS_HASWELL(i915) ? engine->gt->info.num_engines - 1 : 0;
647 bool force_restore = false;
652 if (GRAPHICS_VER(i915) == 7)
653 len += 2 + (num_engines ? 4 * num_engines + 6 : 0);
654 else if (GRAPHICS_VER(i915) == 5)
656 if (flags & MI_FORCE_RESTORE) {
657 GEM_BUG_ON(flags & MI_RESTORE_INHIBIT);
658 flags &= ~MI_FORCE_RESTORE;
659 force_restore = true;
663 cs = intel_ring_begin(rq, len);
667 /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
668 if (GRAPHICS_VER(i915) == 7) {
669 *cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
671 struct intel_engine_cs *signaller;
673 *cs++ = MI_LOAD_REGISTER_IMM(num_engines);
674 for_each_engine(signaller, engine->gt, id) {
675 if (signaller == engine)
678 *cs++ = i915_mmio_reg_offset(
679 RING_PSMI_CTL(signaller->mmio_base));
680 *cs++ = _MASKED_BIT_ENABLE(
681 GEN6_PSMI_SLEEP_MSG_DISABLE);
684 } else if (GRAPHICS_VER(i915) == 5) {
686 * This w/a is only listed for pre-production ilk a/b steppings,
687 * but is also mentioned for programming the powerctx. To be
688 * safe, just apply the workaround; we do not use SyncFlush so
689 * this should never take effect and so be a no-op!
691 *cs++ = MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN;
696 * The HW doesn't handle being told to restore the current
697 * context very well. Quite often it likes goes to go off and
698 * sulk, especially when it is meant to be reloading PP_DIR.
699 * A very simple fix to force the reload is to simply switch
700 * away from the current context and back again.
702 * Note that the kernel_context will contain random state
703 * following the INHIBIT_RESTORE. We accept this since we
704 * never use the kernel_context state; it is merely a
705 * placeholder we use to flush other contexts.
707 *cs++ = MI_SET_CONTEXT;
708 *cs++ = i915_ggtt_offset(engine->kernel_context->state) |
714 *cs++ = MI_SET_CONTEXT;
715 *cs++ = i915_ggtt_offset(ce->state) | flags;
717 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
718 * WaMiSetContext_Hang:snb,ivb,vlv
722 if (GRAPHICS_VER(i915) == 7) {
724 struct intel_engine_cs *signaller;
725 i915_reg_t last_reg = {}; /* keep gcc quiet */
727 *cs++ = MI_LOAD_REGISTER_IMM(num_engines);
728 for_each_engine(signaller, engine->gt, id) {
729 if (signaller == engine)
732 last_reg = RING_PSMI_CTL(signaller->mmio_base);
733 *cs++ = i915_mmio_reg_offset(last_reg);
734 *cs++ = _MASKED_BIT_DISABLE(
735 GEN6_PSMI_SLEEP_MSG_DISABLE);
738 /* Insert a delay before the next switch! */
739 *cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
740 *cs++ = i915_mmio_reg_offset(last_reg);
741 *cs++ = intel_gt_scratch_offset(engine->gt,
742 INTEL_GT_SCRATCH_FIELD_DEFAULT);
745 *cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
746 } else if (GRAPHICS_VER(i915) == 5) {
747 *cs++ = MI_SUSPEND_FLUSH;
750 intel_ring_advance(rq, cs);
755 static int remap_l3_slice(struct i915_request *rq, int slice)
757 #define L3LOG_DW (GEN7_L3LOG_SIZE / sizeof(u32))
758 u32 *cs, *remap_info = rq->engine->i915->l3_parity.remap_info[slice];
764 cs = intel_ring_begin(rq, L3LOG_DW * 2 + 2);
769 * Note: We do not worry about the concurrent register cacheline hang
770 * here because no other code should access these registers other than
771 * at initialization time.
773 *cs++ = MI_LOAD_REGISTER_IMM(L3LOG_DW);
774 for (i = 0; i < L3LOG_DW; i++) {
775 *cs++ = i915_mmio_reg_offset(GEN7_L3LOG(slice, i));
776 *cs++ = remap_info[i];
779 intel_ring_advance(rq, cs);
785 static int remap_l3(struct i915_request *rq)
787 struct i915_gem_context *ctx = i915_request_gem_context(rq);
790 if (!ctx || !ctx->remap_slice)
793 for (i = 0; i < MAX_L3_SLICES; i++) {
794 if (!(ctx->remap_slice & BIT(i)))
797 err = remap_l3_slice(rq, i);
802 ctx->remap_slice = 0;
806 static int switch_mm(struct i915_request *rq, struct i915_address_space *vm)
813 ret = rq->engine->emit_flush(rq, EMIT_FLUSH);
818 * Not only do we need a full barrier (post-sync write) after
819 * invalidating the TLBs, but we need to wait a little bit
820 * longer. Whether this is merely delaying us, or the
821 * subsequent flush is a key part of serialising with the
822 * post-sync op, this extra pass appears vital before a
825 ret = load_pd_dir(rq, vm, PP_DIR_DCLV_2G);
829 return rq->engine->emit_flush(rq, EMIT_INVALIDATE);
832 static int clear_residuals(struct i915_request *rq)
834 struct intel_engine_cs *engine = rq->engine;
837 ret = switch_mm(rq, vm_alias(engine->kernel_context->vm));
841 if (engine->kernel_context->state) {
842 ret = mi_set_context(rq,
843 engine->kernel_context,
844 MI_MM_SPACE_GTT | MI_RESTORE_INHIBIT);
849 ret = engine->emit_bb_start(rq,
850 engine->wa_ctx.vma->node.start, 0,
855 ret = engine->emit_flush(rq, EMIT_FLUSH);
859 /* Always invalidate before the next switch_mm() */
860 return engine->emit_flush(rq, EMIT_INVALIDATE);
863 static int switch_context(struct i915_request *rq)
865 struct intel_engine_cs *engine = rq->engine;
866 struct intel_context *ce = rq->context;
867 void **residuals = NULL;
870 GEM_BUG_ON(HAS_EXECLISTS(engine->i915));
872 if (engine->wa_ctx.vma && ce != engine->kernel_context) {
873 if (engine->wa_ctx.vma->private != ce &&
874 i915_mitigate_clear_residuals()) {
875 ret = clear_residuals(rq);
879 residuals = &engine->wa_ctx.vma->private;
883 ret = switch_mm(rq, vm_alias(ce->vm));
890 GEM_BUG_ON(engine->id != RCS0);
892 /* For resource streamer on HSW+ and power context elsewhere */
893 BUILD_BUG_ON(HSW_MI_RS_SAVE_STATE_EN != MI_SAVE_EXT_STATE_EN);
894 BUILD_BUG_ON(HSW_MI_RS_RESTORE_STATE_EN != MI_RESTORE_EXT_STATE_EN);
896 flags = MI_SAVE_EXT_STATE_EN | MI_MM_SPACE_GTT;
897 if (test_bit(CONTEXT_VALID_BIT, &ce->flags))
898 flags |= MI_RESTORE_EXT_STATE_EN;
900 flags |= MI_RESTORE_INHIBIT;
902 ret = mi_set_context(rq, ce, flags);
912 * Now past the point of no return, this request _will_ be emitted.
914 * Or at least this preamble will be emitted, the request may be
915 * interrupted prior to submitting the user payload. If so, we
916 * still submit the "empty" request in order to preserve global
917 * state tracking such as this, our tracking of the current
921 intel_context_put(*residuals);
922 *residuals = intel_context_get(ce);
928 static int ring_request_alloc(struct i915_request *request)
932 GEM_BUG_ON(!intel_context_is_pinned(request->context));
933 GEM_BUG_ON(i915_request_timeline(request)->has_initial_breadcrumb);
936 * Flush enough space to reduce the likelihood of waiting after
937 * we start building the request - in which case we will just
938 * have to repeat work.
940 request->reserved_space += LEGACY_REQUEST_SIZE;
942 /* Unconditionally invalidate GPU caches and TLBs. */
943 ret = request->engine->emit_flush(request, EMIT_INVALIDATE);
947 ret = switch_context(request);
951 request->reserved_space -= LEGACY_REQUEST_SIZE;
955 static void gen6_bsd_submit_request(struct i915_request *request)
957 struct intel_uncore *uncore = request->engine->uncore;
959 intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
961 /* Every tail move must follow the sequence below */
963 /* Disable notification that the ring is IDLE. The GT
964 * will then assume that it is busy and bring it out of rc6.
966 intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
967 _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
969 /* Clear the context id. Here be magic! */
970 intel_uncore_write64_fw(uncore, GEN6_BSD_RNCID, 0x0);
972 /* Wait for the ring not to be idle, i.e. for it to wake up. */
973 if (__intel_wait_for_register_fw(uncore,
974 GEN6_BSD_SLEEP_PSMI_CONTROL,
975 GEN6_BSD_SLEEP_INDICATOR,
978 drm_err(&uncore->i915->drm,
979 "timed out waiting for the BSD ring to wake up\n");
981 /* Now that the ring is fully powered up, update the tail */
982 i9xx_submit_request(request);
984 /* Let the ring send IDLE messages to the GT again,
985 * and so let it sleep to conserve power when idle.
987 intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
988 _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
990 intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
993 static void i9xx_set_default_submission(struct intel_engine_cs *engine)
995 engine->submit_request = i9xx_submit_request;
998 static void gen6_bsd_set_default_submission(struct intel_engine_cs *engine)
1000 engine->submit_request = gen6_bsd_submit_request;
1003 static void ring_release(struct intel_engine_cs *engine)
1005 struct drm_i915_private *dev_priv = engine->i915;
1007 drm_WARN_ON(&dev_priv->drm, GRAPHICS_VER(dev_priv) > 2 &&
1008 (ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
1010 intel_engine_cleanup_common(engine);
1012 if (engine->wa_ctx.vma) {
1013 intel_context_put(engine->wa_ctx.vma->private);
1014 i915_vma_unpin_and_release(&engine->wa_ctx.vma, 0);
1017 intel_ring_unpin(engine->legacy.ring);
1018 intel_ring_put(engine->legacy.ring);
1020 intel_timeline_unpin(engine->legacy.timeline);
1021 intel_timeline_put(engine->legacy.timeline);
1024 static void irq_handler(struct intel_engine_cs *engine, u16 iir)
1026 intel_engine_signal_breadcrumbs(engine);
1029 static void setup_irq(struct intel_engine_cs *engine)
1031 struct drm_i915_private *i915 = engine->i915;
1033 intel_engine_set_irq_handler(engine, irq_handler);
1035 if (GRAPHICS_VER(i915) >= 6) {
1036 engine->irq_enable = gen6_irq_enable;
1037 engine->irq_disable = gen6_irq_disable;
1038 } else if (GRAPHICS_VER(i915) >= 5) {
1039 engine->irq_enable = gen5_irq_enable;
1040 engine->irq_disable = gen5_irq_disable;
1041 } else if (GRAPHICS_VER(i915) >= 3) {
1042 engine->irq_enable = gen3_irq_enable;
1043 engine->irq_disable = gen3_irq_disable;
1045 engine->irq_enable = gen2_irq_enable;
1046 engine->irq_disable = gen2_irq_disable;
1050 static void setup_common(struct intel_engine_cs *engine)
1052 struct drm_i915_private *i915 = engine->i915;
1054 /* gen8+ are only supported with execlists */
1055 GEM_BUG_ON(GRAPHICS_VER(i915) >= 8);
1059 engine->resume = xcs_resume;
1060 engine->sanitize = xcs_sanitize;
1062 engine->reset.prepare = reset_prepare;
1063 engine->reset.rewind = reset_rewind;
1064 engine->reset.cancel = reset_cancel;
1065 engine->reset.finish = reset_finish;
1067 engine->cops = &ring_context_ops;
1068 engine->request_alloc = ring_request_alloc;
1071 * Using a global execution timeline; the previous final breadcrumb is
1072 * equivalent to our next initial bread so we can elide
1073 * engine->emit_init_breadcrumb().
1075 engine->emit_fini_breadcrumb = gen3_emit_breadcrumb;
1076 if (GRAPHICS_VER(i915) == 5)
1077 engine->emit_fini_breadcrumb = gen5_emit_breadcrumb;
1079 engine->set_default_submission = i9xx_set_default_submission;
1081 if (GRAPHICS_VER(i915) >= 6)
1082 engine->emit_bb_start = gen6_emit_bb_start;
1083 else if (GRAPHICS_VER(i915) >= 4)
1084 engine->emit_bb_start = gen4_emit_bb_start;
1085 else if (IS_I830(i915) || IS_I845G(i915))
1086 engine->emit_bb_start = i830_emit_bb_start;
1088 engine->emit_bb_start = gen3_emit_bb_start;
1091 static void setup_rcs(struct intel_engine_cs *engine)
1093 struct drm_i915_private *i915 = engine->i915;
1095 if (HAS_L3_DPF(i915))
1096 engine->irq_keep_mask = GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
1098 engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
1100 if (GRAPHICS_VER(i915) >= 7) {
1101 engine->emit_flush = gen7_emit_flush_rcs;
1102 engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_rcs;
1103 } else if (GRAPHICS_VER(i915) == 6) {
1104 engine->emit_flush = gen6_emit_flush_rcs;
1105 engine->emit_fini_breadcrumb = gen6_emit_breadcrumb_rcs;
1106 } else if (GRAPHICS_VER(i915) == 5) {
1107 engine->emit_flush = gen4_emit_flush_rcs;
1109 if (GRAPHICS_VER(i915) < 4)
1110 engine->emit_flush = gen2_emit_flush;
1112 engine->emit_flush = gen4_emit_flush_rcs;
1113 engine->irq_enable_mask = I915_USER_INTERRUPT;
1116 if (IS_HASWELL(i915))
1117 engine->emit_bb_start = hsw_emit_bb_start;
1120 static void setup_vcs(struct intel_engine_cs *engine)
1122 struct drm_i915_private *i915 = engine->i915;
1124 if (GRAPHICS_VER(i915) >= 6) {
1125 /* gen6 bsd needs a special wa for tail updates */
1126 if (GRAPHICS_VER(i915) == 6)
1127 engine->set_default_submission = gen6_bsd_set_default_submission;
1128 engine->emit_flush = gen6_emit_flush_vcs;
1129 engine->irq_enable_mask = GT_BSD_USER_INTERRUPT;
1131 if (GRAPHICS_VER(i915) == 6)
1132 engine->emit_fini_breadcrumb = gen6_emit_breadcrumb_xcs;
1134 engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_xcs;
1136 engine->emit_flush = gen4_emit_flush_vcs;
1137 if (GRAPHICS_VER(i915) == 5)
1138 engine->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
1140 engine->irq_enable_mask = I915_BSD_USER_INTERRUPT;
1144 static void setup_bcs(struct intel_engine_cs *engine)
1146 struct drm_i915_private *i915 = engine->i915;
1148 engine->emit_flush = gen6_emit_flush_xcs;
1149 engine->irq_enable_mask = GT_BLT_USER_INTERRUPT;
1151 if (GRAPHICS_VER(i915) == 6)
1152 engine->emit_fini_breadcrumb = gen6_emit_breadcrumb_xcs;
1154 engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_xcs;
1157 static void setup_vecs(struct intel_engine_cs *engine)
1159 struct drm_i915_private *i915 = engine->i915;
1161 GEM_BUG_ON(GRAPHICS_VER(i915) < 7);
1163 engine->emit_flush = gen6_emit_flush_xcs;
1164 engine->irq_enable_mask = PM_VEBOX_USER_INTERRUPT;
1165 engine->irq_enable = hsw_irq_enable_vecs;
1166 engine->irq_disable = hsw_irq_disable_vecs;
1168 engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_xcs;
1171 static int gen7_ctx_switch_bb_setup(struct intel_engine_cs * const engine,
1172 struct i915_vma * const vma)
1174 return gen7_setup_clear_gpr_bb(engine, vma);
1177 static int gen7_ctx_switch_bb_init(struct intel_engine_cs *engine,
1178 struct i915_gem_ww_ctx *ww,
1179 struct i915_vma *vma)
1183 err = i915_vma_pin_ww(vma, ww, 0, 0, PIN_USER | PIN_HIGH);
1187 err = i915_vma_sync(vma);
1191 err = gen7_ctx_switch_bb_setup(engine, vma);
1195 engine->wa_ctx.vma = vma;
1199 i915_vma_unpin(vma);
1203 static struct i915_vma *gen7_ctx_vma(struct intel_engine_cs *engine)
1205 struct drm_i915_gem_object *obj;
1206 struct i915_vma *vma;
1209 if (GRAPHICS_VER(engine->i915) != 7 || engine->class != RENDER_CLASS)
1212 err = gen7_ctx_switch_bb_setup(engine, NULL /* probe size */);
1214 return ERR_PTR(err);
1218 size = ALIGN(err, PAGE_SIZE);
1220 obj = i915_gem_object_create_internal(engine->i915, size);
1222 return ERR_CAST(obj);
1224 vma = i915_vma_instance(obj, engine->gt->vm, NULL);
1226 i915_gem_object_put(obj);
1227 return ERR_CAST(vma);
1230 vma->private = intel_context_create(engine); /* dummy residuals */
1231 if (IS_ERR(vma->private)) {
1232 err = PTR_ERR(vma->private);
1233 vma->private = NULL;
1234 i915_gem_object_put(obj);
1235 return ERR_PTR(err);
1241 int intel_ring_submission_setup(struct intel_engine_cs *engine)
1243 struct i915_gem_ww_ctx ww;
1244 struct intel_timeline *timeline;
1245 struct intel_ring *ring;
1246 struct i915_vma *gen7_wa_vma;
1249 setup_common(engine);
1251 switch (engine->class) {
1255 case VIDEO_DECODE_CLASS:
1258 case COPY_ENGINE_CLASS:
1261 case VIDEO_ENHANCEMENT_CLASS:
1265 MISSING_CASE(engine->class);
1269 timeline = intel_timeline_create_from_engine(engine,
1270 I915_GEM_HWS_SEQNO_ADDR);
1271 if (IS_ERR(timeline)) {
1272 err = PTR_ERR(timeline);
1275 GEM_BUG_ON(timeline->has_initial_breadcrumb);
1277 ring = intel_engine_create_ring(engine, SZ_16K);
1279 err = PTR_ERR(ring);
1283 GEM_BUG_ON(engine->legacy.ring);
1284 engine->legacy.ring = ring;
1285 engine->legacy.timeline = timeline;
1287 gen7_wa_vma = gen7_ctx_vma(engine);
1288 if (IS_ERR(gen7_wa_vma)) {
1289 err = PTR_ERR(gen7_wa_vma);
1293 i915_gem_ww_ctx_init(&ww, false);
1296 err = i915_gem_object_lock(timeline->hwsp_ggtt->obj, &ww);
1297 if (!err && gen7_wa_vma)
1298 err = i915_gem_object_lock(gen7_wa_vma->obj, &ww);
1299 if (!err && engine->legacy.ring->vma->obj)
1300 err = i915_gem_object_lock(engine->legacy.ring->vma->obj, &ww);
1302 err = intel_timeline_pin(timeline, &ww);
1304 err = intel_ring_pin(ring, &ww);
1306 intel_timeline_unpin(timeline);
1311 GEM_BUG_ON(timeline->hwsp_ggtt != engine->status_page.vma);
1314 err = gen7_ctx_switch_bb_init(engine, &ww, gen7_wa_vma);
1316 intel_ring_unpin(ring);
1317 intel_timeline_unpin(timeline);
1322 if (err == -EDEADLK) {
1323 err = i915_gem_ww_ctx_backoff(&ww);
1327 i915_gem_ww_ctx_fini(&ww);
1331 /* Finally, take ownership and responsibility for cleanup! */
1332 engine->release = ring_release;
1338 intel_context_put(gen7_wa_vma->private);
1339 i915_gem_object_put(gen7_wa_vma->obj);
1342 intel_ring_put(ring);
1344 intel_timeline_put(timeline);
1346 intel_engine_cleanup_common(engine);
1350 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1351 #include "selftest_ring_submission.c"