1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2008-2021 Intel Corporation
4  */
5 
6 #include "gen2_engine_cs.h"
7 #include "gen6_engine_cs.h"
8 #include "gen6_ppgtt.h"
9 #include "gen7_renderclear.h"
10 #include "i915_drv.h"
11 #include "i915_mitigations.h"
12 #include "intel_breadcrumbs.h"
13 #include "intel_context.h"
14 #include "intel_gt.h"
15 #include "intel_gt_irq.h"
16 #include "intel_reset.h"
17 #include "intel_ring.h"
18 #include "shmem_utils.h"
19 #include "intel_engine_heartbeat.h"
20 #include "intel_engine_pm.h"
21 
22 /* Rough estimate of the typical request size, performing a flush,
23  * set-context and then emitting the batch.
24  */
25 #define LEGACY_REQUEST_SIZE 200
26 
27 static void set_hwstam(struct intel_engine_cs *engine, u32 mask)
28 {
29 	/*
30 	 * Keep the render interrupt unmasked as this papers over
31 	 * lost interrupts following a reset.
32 	 */
33 	if (engine->class == RENDER_CLASS) {
34 		if (GRAPHICS_VER(engine->i915) >= 6)
35 			mask &= ~BIT(0);
36 		else
37 			mask &= ~I915_USER_INTERRUPT;
38 	}
39 
40 	intel_engine_set_hwsp_writemask(engine, mask);
41 }
42 
43 static void set_hws_pga(struct intel_engine_cs *engine, phys_addr_t phys)
44 {
45 	u32 addr;
46 
47 	addr = lower_32_bits(phys);
48 	if (GRAPHICS_VER(engine->i915) >= 4)
49 		addr |= (phys >> 28) & 0xf0;
50 
51 	intel_uncore_write(engine->uncore, HWS_PGA, addr);
52 }
53 
54 static struct page *status_page(struct intel_engine_cs *engine)
55 {
56 	struct drm_i915_gem_object *obj = engine->status_page.vma->obj;
57 
58 	GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
59 	return sg_page(obj->mm.pages->sgl);
60 }
61 
62 static void ring_setup_phys_status_page(struct intel_engine_cs *engine)
63 {
64 	set_hws_pga(engine, PFN_PHYS(page_to_pfn(status_page(engine))));
65 	set_hwstam(engine, ~0u);
66 }
67 
68 static void set_hwsp(struct intel_engine_cs *engine, u32 offset)
69 {
70 	i915_reg_t hwsp;
71 
72 	/*
73 	 * The ring status page addresses are no longer next to the rest of
74 	 * the ring registers as of gen7.
75 	 */
76 	if (GRAPHICS_VER(engine->i915) == 7) {
77 		switch (engine->id) {
78 		/*
79 		 * No more rings exist on Gen7. Default case is only to shut up
80 		 * gcc switch check warning.
81 		 */
82 		default:
83 			GEM_BUG_ON(engine->id);
84 			fallthrough;
85 		case RCS0:
86 			hwsp = RENDER_HWS_PGA_GEN7;
87 			break;
88 		case BCS0:
89 			hwsp = BLT_HWS_PGA_GEN7;
90 			break;
91 		case VCS0:
92 			hwsp = BSD_HWS_PGA_GEN7;
93 			break;
94 		case VECS0:
95 			hwsp = VEBOX_HWS_PGA_GEN7;
96 			break;
97 		}
98 	} else if (GRAPHICS_VER(engine->i915) == 6) {
99 		hwsp = RING_HWS_PGA_GEN6(engine->mmio_base);
100 	} else {
101 		hwsp = RING_HWS_PGA(engine->mmio_base);
102 	}
103 
104 	intel_uncore_write_fw(engine->uncore, hwsp, offset);
105 	intel_uncore_posting_read_fw(engine->uncore, hwsp);
106 }
107 
108 static void flush_cs_tlb(struct intel_engine_cs *engine)
109 {
110 	if (!IS_GRAPHICS_VER(engine->i915, 6, 7))
111 		return;
112 
113 	/* ring should be idle before issuing a sync flush*/
114 	GEM_DEBUG_WARN_ON((ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
115 
116 	ENGINE_WRITE_FW(engine, RING_INSTPM,
117 			_MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE |
118 					   INSTPM_SYNC_FLUSH));
119 	if (__intel_wait_for_register_fw(engine->uncore,
120 					 RING_INSTPM(engine->mmio_base),
121 					 INSTPM_SYNC_FLUSH, 0,
122 					 2000, 0, NULL))
123 		ENGINE_TRACE(engine,
124 			     "wait for SyncFlush to complete for TLB invalidation timed out\n");
125 }
126 
127 static void ring_setup_status_page(struct intel_engine_cs *engine)
128 {
129 	set_hwsp(engine, i915_ggtt_offset(engine->status_page.vma));
130 	set_hwstam(engine, ~0u);
131 
132 	flush_cs_tlb(engine);
133 }
134 
135 static struct i915_address_space *vm_alias(struct i915_address_space *vm)
136 {
137 	if (i915_is_ggtt(vm))
138 		vm = &i915_vm_to_ggtt(vm)->alias->vm;
139 
140 	return vm;
141 }
142 
143 static u32 pp_dir(struct i915_address_space *vm)
144 {
145 	return to_gen6_ppgtt(i915_vm_to_ppgtt(vm))->pp_dir;
146 }
147 
148 static void set_pp_dir(struct intel_engine_cs *engine)
149 {
150 	struct i915_address_space *vm = vm_alias(engine->gt->vm);
151 
152 	if (!vm)
153 		return;
154 
155 	ENGINE_WRITE_FW(engine, RING_PP_DIR_DCLV, PP_DIR_DCLV_2G);
156 	ENGINE_WRITE_FW(engine, RING_PP_DIR_BASE, pp_dir(vm));
157 
158 	if (GRAPHICS_VER(engine->i915) >= 7) {
159 		ENGINE_WRITE_FW(engine,
160 				RING_MODE_GEN7,
161 				_MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
162 	}
163 }
164 
165 static bool stop_ring(struct intel_engine_cs *engine)
166 {
167 	/* Empty the ring by skipping to the end */
168 	ENGINE_WRITE_FW(engine, RING_HEAD, ENGINE_READ_FW(engine, RING_TAIL));
169 	ENGINE_POSTING_READ(engine, RING_HEAD);
170 
171 	/* The ring must be empty before it is disabled */
172 	ENGINE_WRITE_FW(engine, RING_CTL, 0);
173 	ENGINE_POSTING_READ(engine, RING_CTL);
174 
175 	/* Then reset the disabled ring */
176 	ENGINE_WRITE_FW(engine, RING_HEAD, 0);
177 	ENGINE_WRITE_FW(engine, RING_TAIL, 0);
178 
179 	return (ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR) == 0;
180 }
181 
182 static int xcs_resume(struct intel_engine_cs *engine)
183 {
184 	struct intel_ring *ring = engine->legacy.ring;
185 
186 	ENGINE_TRACE(engine, "ring:{HEAD:%04x, TAIL:%04x}\n",
187 		     ring->head, ring->tail);
188 
189 	/*
190 	 * Double check the ring is empty & disabled before we resume. Called
191 	 * from atomic context during PCI probe, so _hardirq().
192 	 */
193 	intel_synchronize_hardirq(engine->i915);
194 	if (!stop_ring(engine))
195 		goto err;
196 
197 	if (HWS_NEEDS_PHYSICAL(engine->i915))
198 		ring_setup_phys_status_page(engine);
199 	else
200 		ring_setup_status_page(engine);
201 
202 	intel_breadcrumbs_reset(engine->breadcrumbs);
203 
204 	/* Enforce ordering by reading HEAD register back */
205 	ENGINE_POSTING_READ(engine, RING_HEAD);
206 
207 	/*
208 	 * Initialize the ring. This must happen _after_ we've cleared the ring
209 	 * registers with the above sequence (the readback of the HEAD registers
210 	 * also enforces ordering), otherwise the hw might lose the new ring
211 	 * register values.
212 	 */
213 	ENGINE_WRITE_FW(engine, RING_START, i915_ggtt_offset(ring->vma));
214 
215 	/* Check that the ring offsets point within the ring! */
216 	GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->head));
217 	GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->tail));
218 	intel_ring_update_space(ring);
219 
220 	set_pp_dir(engine);
221 
222 	/* First wake the ring up to an empty/idle ring */
223 	ENGINE_WRITE_FW(engine, RING_HEAD, ring->head);
224 	ENGINE_WRITE_FW(engine, RING_TAIL, ring->head);
225 	ENGINE_POSTING_READ(engine, RING_TAIL);
226 
227 	ENGINE_WRITE_FW(engine, RING_CTL,
228 			RING_CTL_SIZE(ring->size) | RING_VALID);
229 
230 	/* If the head is still not zero, the ring is dead */
231 	if (__intel_wait_for_register_fw(engine->uncore,
232 					 RING_CTL(engine->mmio_base),
233 					 RING_VALID, RING_VALID,
234 					 5000, 0, NULL))
235 		goto err;
236 
237 	if (GRAPHICS_VER(engine->i915) > 2)
238 		ENGINE_WRITE_FW(engine,
239 				RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
240 
241 	/* Now awake, let it get started */
242 	if (ring->tail != ring->head) {
243 		ENGINE_WRITE_FW(engine, RING_TAIL, ring->tail);
244 		ENGINE_POSTING_READ(engine, RING_TAIL);
245 	}
246 
247 	/* Papering over lost _interrupts_ immediately following the restart */
248 	intel_engine_signal_breadcrumbs(engine);
249 	return 0;
250 
251 err:
252 	drm_err(&engine->i915->drm,
253 		"%s initialization failed; "
254 		"ctl %08x (valid? %d) head %08x [%08x] tail %08x [%08x] start %08x [expected %08x]\n",
255 		engine->name,
256 		ENGINE_READ(engine, RING_CTL),
257 		ENGINE_READ(engine, RING_CTL) & RING_VALID,
258 		ENGINE_READ(engine, RING_HEAD), ring->head,
259 		ENGINE_READ(engine, RING_TAIL), ring->tail,
260 		ENGINE_READ(engine, RING_START),
261 		i915_ggtt_offset(ring->vma));
262 	return -EIO;
263 }
264 
265 static void sanitize_hwsp(struct intel_engine_cs *engine)
266 {
267 	struct intel_timeline *tl;
268 
269 	list_for_each_entry(tl, &engine->status_page.timelines, engine_link)
270 		intel_timeline_reset_seqno(tl);
271 }
272 
273 static void xcs_sanitize(struct intel_engine_cs *engine)
274 {
275 	/*
276 	 * Poison residual state on resume, in case the suspend didn't!
277 	 *
278 	 * We have to assume that across suspend/resume (or other loss
279 	 * of control) that the contents of our pinned buffers has been
280 	 * lost, replaced by garbage. Since this doesn't always happen,
281 	 * let's poison such state so that we more quickly spot when
282 	 * we falsely assume it has been preserved.
283 	 */
284 	if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
285 		memset(engine->status_page.addr, POISON_INUSE, PAGE_SIZE);
286 
287 	/*
288 	 * The kernel_context HWSP is stored in the status_page. As above,
289 	 * that may be lost on resume/initialisation, and so we need to
290 	 * reset the value in the HWSP.
291 	 */
292 	sanitize_hwsp(engine);
293 
294 	/* And scrub the dirty cachelines for the HWSP */
295 	drm_clflush_virt_range(engine->status_page.addr, PAGE_SIZE);
296 
297 	intel_engine_reset_pinned_contexts(engine);
298 }
299 
300 static void reset_prepare(struct intel_engine_cs *engine)
301 {
302 	/*
303 	 * We stop engines, otherwise we might get failed reset and a
304 	 * dead gpu (on elk). Also as modern gpu as kbl can suffer
305 	 * from system hang if batchbuffer is progressing when
306 	 * the reset is issued, regardless of READY_TO_RESET ack.
307 	 * Thus assume it is best to stop engines on all gens
308 	 * where we have a gpu reset.
309 	 *
310 	 * WaKBLVECSSemaphoreWaitPoll:kbl (on ALL_ENGINES)
311 	 *
312 	 * WaMediaResetMainRingCleanup:ctg,elk (presumably)
313 	 * WaClearRingBufHeadRegAtInit:ctg,elk
314 	 *
315 	 * FIXME: Wa for more modern gens needs to be validated
316 	 */
317 	ENGINE_TRACE(engine, "\n");
318 	intel_engine_stop_cs(engine);
319 
320 	if (!stop_ring(engine)) {
321 		/* G45 ring initialization often fails to reset head to zero */
322 		ENGINE_TRACE(engine,
323 			     "HEAD not reset to zero, "
324 			     "{ CTL:%08x, HEAD:%08x, TAIL:%08x, START:%08x }\n",
325 			     ENGINE_READ_FW(engine, RING_CTL),
326 			     ENGINE_READ_FW(engine, RING_HEAD),
327 			     ENGINE_READ_FW(engine, RING_TAIL),
328 			     ENGINE_READ_FW(engine, RING_START));
329 		if (!stop_ring(engine)) {
330 			drm_err(&engine->i915->drm,
331 				"failed to set %s head to zero "
332 				"ctl %08x head %08x tail %08x start %08x\n",
333 				engine->name,
334 				ENGINE_READ_FW(engine, RING_CTL),
335 				ENGINE_READ_FW(engine, RING_HEAD),
336 				ENGINE_READ_FW(engine, RING_TAIL),
337 				ENGINE_READ_FW(engine, RING_START));
338 		}
339 	}
340 }
341 
342 static void reset_rewind(struct intel_engine_cs *engine, bool stalled)
343 {
344 	struct i915_request *pos, *rq;
345 	unsigned long flags;
346 	u32 head;
347 
348 	rq = NULL;
349 	spin_lock_irqsave(&engine->sched_engine->lock, flags);
350 	rcu_read_lock();
351 	list_for_each_entry(pos, &engine->sched_engine->requests, sched.link) {
352 		if (!__i915_request_is_complete(pos)) {
353 			rq = pos;
354 			break;
355 		}
356 	}
357 	rcu_read_unlock();
358 
359 	/*
360 	 * The guilty request will get skipped on a hung engine.
361 	 *
362 	 * Users of client default contexts do not rely on logical
363 	 * state preserved between batches so it is safe to execute
364 	 * queued requests following the hang. Non default contexts
365 	 * rely on preserved state, so skipping a batch loses the
366 	 * evolution of the state and it needs to be considered corrupted.
367 	 * Executing more queued batches on top of corrupted state is
368 	 * risky. But we take the risk by trying to advance through
369 	 * the queued requests in order to make the client behaviour
370 	 * more predictable around resets, by not throwing away random
371 	 * amount of batches it has prepared for execution. Sophisticated
372 	 * clients can use gem_reset_stats_ioctl and dma fence status
373 	 * (exported via sync_file info ioctl on explicit fences) to observe
374 	 * when it loses the context state and should rebuild accordingly.
375 	 *
376 	 * The context ban, and ultimately the client ban, mechanism are safety
377 	 * valves if client submission ends up resulting in nothing more than
378 	 * subsequent hangs.
379 	 */
380 
381 	if (rq) {
382 		/*
383 		 * Try to restore the logical GPU state to match the
384 		 * continuation of the request queue. If we skip the
385 		 * context/PD restore, then the next request may try to execute
386 		 * assuming that its context is valid and loaded on the GPU and
387 		 * so may try to access invalid memory, prompting repeated GPU
388 		 * hangs.
389 		 *
390 		 * If the request was guilty, we still restore the logical
391 		 * state in case the next request requires it (e.g. the
392 		 * aliasing ppgtt), but skip over the hung batch.
393 		 *
394 		 * If the request was innocent, we try to replay the request
395 		 * with the restored context.
396 		 */
397 		__i915_request_reset(rq, stalled);
398 
399 		GEM_BUG_ON(rq->ring != engine->legacy.ring);
400 		head = rq->head;
401 	} else {
402 		head = engine->legacy.ring->tail;
403 	}
404 	engine->legacy.ring->head = intel_ring_wrap(engine->legacy.ring, head);
405 
406 	spin_unlock_irqrestore(&engine->sched_engine->lock, flags);
407 }
408 
409 static void reset_finish(struct intel_engine_cs *engine)
410 {
411 }
412 
413 static void reset_cancel(struct intel_engine_cs *engine)
414 {
415 	struct i915_request *request;
416 	unsigned long flags;
417 
418 	spin_lock_irqsave(&engine->sched_engine->lock, flags);
419 
420 	/* Mark all submitted requests as skipped. */
421 	list_for_each_entry(request, &engine->sched_engine->requests, sched.link)
422 		i915_request_put(i915_request_mark_eio(request));
423 	intel_engine_signal_breadcrumbs(engine);
424 
425 	/* Remaining _unready_ requests will be nop'ed when submitted */
426 
427 	spin_unlock_irqrestore(&engine->sched_engine->lock, flags);
428 }
429 
430 static void i9xx_submit_request(struct i915_request *request)
431 {
432 	i915_request_submit(request);
433 	wmb(); /* paranoid flush writes out of the WCB before mmio */
434 
435 	ENGINE_WRITE(request->engine, RING_TAIL,
436 		     intel_ring_set_tail(request->ring, request->tail));
437 }
438 
439 static void __ring_context_fini(struct intel_context *ce)
440 {
441 	i915_vma_put(ce->state);
442 }
443 
444 static void ring_context_destroy(struct kref *ref)
445 {
446 	struct intel_context *ce = container_of(ref, typeof(*ce), ref);
447 
448 	GEM_BUG_ON(intel_context_is_pinned(ce));
449 
450 	if (ce->state)
451 		__ring_context_fini(ce);
452 
453 	intel_context_fini(ce);
454 	intel_context_free(ce);
455 }
456 
457 static int ring_context_init_default_state(struct intel_context *ce,
458 					   struct i915_gem_ww_ctx *ww)
459 {
460 	struct drm_i915_gem_object *obj = ce->state->obj;
461 	void *vaddr;
462 
463 	vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
464 	if (IS_ERR(vaddr))
465 		return PTR_ERR(vaddr);
466 
467 	shmem_read(ce->engine->default_state, 0,
468 		   vaddr, ce->engine->context_size);
469 
470 	i915_gem_object_flush_map(obj);
471 	__i915_gem_object_release_map(obj);
472 
473 	__set_bit(CONTEXT_VALID_BIT, &ce->flags);
474 	return 0;
475 }
476 
477 static int ring_context_pre_pin(struct intel_context *ce,
478 				struct i915_gem_ww_ctx *ww,
479 				void **unused)
480 {
481 	struct i915_address_space *vm;
482 	int err = 0;
483 
484 	if (ce->engine->default_state &&
485 	    !test_bit(CONTEXT_VALID_BIT, &ce->flags)) {
486 		err = ring_context_init_default_state(ce, ww);
487 		if (err)
488 			return err;
489 	}
490 
491 	vm = vm_alias(ce->vm);
492 	if (vm)
493 		err = gen6_ppgtt_pin(i915_vm_to_ppgtt((vm)), ww);
494 
495 	return err;
496 }
497 
498 static void __context_unpin_ppgtt(struct intel_context *ce)
499 {
500 	struct i915_address_space *vm;
501 
502 	vm = vm_alias(ce->vm);
503 	if (vm)
504 		gen6_ppgtt_unpin(i915_vm_to_ppgtt(vm));
505 }
506 
507 static void ring_context_unpin(struct intel_context *ce)
508 {
509 }
510 
511 static void ring_context_post_unpin(struct intel_context *ce)
512 {
513 	__context_unpin_ppgtt(ce);
514 }
515 
516 static struct i915_vma *
517 alloc_context_vma(struct intel_engine_cs *engine)
518 {
519 	struct drm_i915_private *i915 = engine->i915;
520 	struct drm_i915_gem_object *obj;
521 	struct i915_vma *vma;
522 	int err;
523 
524 	obj = i915_gem_object_create_shmem(i915, engine->context_size);
525 	if (IS_ERR(obj))
526 		return ERR_CAST(obj);
527 
528 	/*
529 	 * Try to make the context utilize L3 as well as LLC.
530 	 *
531 	 * On VLV we don't have L3 controls in the PTEs so we
532 	 * shouldn't touch the cache level, especially as that
533 	 * would make the object snooped which might have a
534 	 * negative performance impact.
535 	 *
536 	 * Snooping is required on non-llc platforms in execlist
537 	 * mode, but since all GGTT accesses use PAT entry 0 we
538 	 * get snooping anyway regardless of cache_level.
539 	 *
540 	 * This is only applicable for Ivy Bridge devices since
541 	 * later platforms don't have L3 control bits in the PTE.
542 	 */
543 	if (IS_IVYBRIDGE(i915))
544 		i915_gem_object_set_cache_coherency(obj, I915_CACHE_L3_LLC);
545 
546 	vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
547 	if (IS_ERR(vma)) {
548 		err = PTR_ERR(vma);
549 		goto err_obj;
550 	}
551 
552 	return vma;
553 
554 err_obj:
555 	i915_gem_object_put(obj);
556 	return ERR_PTR(err);
557 }
558 
559 static int ring_context_alloc(struct intel_context *ce)
560 {
561 	struct intel_engine_cs *engine = ce->engine;
562 
563 	/* One ringbuffer to rule them all */
564 	GEM_BUG_ON(!engine->legacy.ring);
565 	ce->ring = engine->legacy.ring;
566 	ce->timeline = intel_timeline_get(engine->legacy.timeline);
567 
568 	GEM_BUG_ON(ce->state);
569 	if (engine->context_size) {
570 		struct i915_vma *vma;
571 
572 		vma = alloc_context_vma(engine);
573 		if (IS_ERR(vma))
574 			return PTR_ERR(vma);
575 
576 		ce->state = vma;
577 	}
578 
579 	return 0;
580 }
581 
582 static int ring_context_pin(struct intel_context *ce, void *unused)
583 {
584 	return 0;
585 }
586 
587 static void ring_context_reset(struct intel_context *ce)
588 {
589 	intel_ring_reset(ce->ring, ce->ring->emit);
590 	clear_bit(CONTEXT_VALID_BIT, &ce->flags);
591 }
592 
593 static void ring_context_ban(struct intel_context *ce,
594 			     struct i915_request *rq)
595 {
596 	struct intel_engine_cs *engine;
597 
598 	if (!rq || !i915_request_is_active(rq))
599 		return;
600 
601 	engine = rq->engine;
602 	lockdep_assert_held(&engine->sched_engine->lock);
603 	list_for_each_entry_continue(rq, &engine->sched_engine->requests,
604 				     sched.link)
605 		if (rq->context == ce) {
606 			i915_request_set_error_once(rq, -EIO);
607 			__i915_request_skip(rq);
608 		}
609 }
610 
611 static void ring_context_cancel_request(struct intel_context *ce,
612 					struct i915_request *rq)
613 {
614 	struct intel_engine_cs *engine = NULL;
615 
616 	i915_request_active_engine(rq, &engine);
617 
618 	if (engine && intel_engine_pulse(engine))
619 		intel_gt_handle_error(engine->gt, engine->mask, 0,
620 				      "request cancellation by %s",
621 				      current->comm);
622 }
623 
624 static const struct intel_context_ops ring_context_ops = {
625 	.alloc = ring_context_alloc,
626 
627 	.cancel_request = ring_context_cancel_request,
628 
629 	.ban = ring_context_ban,
630 
631 	.pre_pin = ring_context_pre_pin,
632 	.pin = ring_context_pin,
633 	.unpin = ring_context_unpin,
634 	.post_unpin = ring_context_post_unpin,
635 
636 	.enter = intel_context_enter_engine,
637 	.exit = intel_context_exit_engine,
638 
639 	.reset = ring_context_reset,
640 	.destroy = ring_context_destroy,
641 };
642 
643 static int load_pd_dir(struct i915_request *rq,
644 		       struct i915_address_space *vm,
645 		       u32 valid)
646 {
647 	const struct intel_engine_cs * const engine = rq->engine;
648 	u32 *cs;
649 
650 	cs = intel_ring_begin(rq, 12);
651 	if (IS_ERR(cs))
652 		return PTR_ERR(cs);
653 
654 	*cs++ = MI_LOAD_REGISTER_IMM(1);
655 	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_DCLV(engine->mmio_base));
656 	*cs++ = valid;
657 
658 	*cs++ = MI_LOAD_REGISTER_IMM(1);
659 	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
660 	*cs++ = pp_dir(vm);
661 
662 	/* Stall until the page table load is complete? */
663 	*cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
664 	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
665 	*cs++ = intel_gt_scratch_offset(engine->gt,
666 					INTEL_GT_SCRATCH_FIELD_DEFAULT);
667 
668 	*cs++ = MI_LOAD_REGISTER_IMM(1);
669 	*cs++ = i915_mmio_reg_offset(RING_INSTPM(engine->mmio_base));
670 	*cs++ = _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE);
671 
672 	intel_ring_advance(rq, cs);
673 
674 	return rq->engine->emit_flush(rq, EMIT_FLUSH);
675 }
676 
677 static int mi_set_context(struct i915_request *rq,
678 			  struct intel_context *ce,
679 			  u32 flags)
680 {
681 	struct intel_engine_cs *engine = rq->engine;
682 	struct drm_i915_private *i915 = engine->i915;
683 	enum intel_engine_id id;
684 	const int num_engines =
685 		IS_HASWELL(i915) ? engine->gt->info.num_engines - 1 : 0;
686 	bool force_restore = false;
687 	int len;
688 	u32 *cs;
689 
690 	len = 4;
691 	if (GRAPHICS_VER(i915) == 7)
692 		len += 2 + (num_engines ? 4 * num_engines + 6 : 0);
693 	else if (GRAPHICS_VER(i915) == 5)
694 		len += 2;
695 	if (flags & MI_FORCE_RESTORE) {
696 		GEM_BUG_ON(flags & MI_RESTORE_INHIBIT);
697 		flags &= ~MI_FORCE_RESTORE;
698 		force_restore = true;
699 		len += 2;
700 	}
701 
702 	cs = intel_ring_begin(rq, len);
703 	if (IS_ERR(cs))
704 		return PTR_ERR(cs);
705 
706 	/* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
707 	if (GRAPHICS_VER(i915) == 7) {
708 		*cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
709 		if (num_engines) {
710 			struct intel_engine_cs *signaller;
711 
712 			*cs++ = MI_LOAD_REGISTER_IMM(num_engines);
713 			for_each_engine(signaller, engine->gt, id) {
714 				if (signaller == engine)
715 					continue;
716 
717 				*cs++ = i915_mmio_reg_offset(
718 					   RING_PSMI_CTL(signaller->mmio_base));
719 				*cs++ = _MASKED_BIT_ENABLE(
720 						GEN6_PSMI_SLEEP_MSG_DISABLE);
721 			}
722 		}
723 	} else if (GRAPHICS_VER(i915) == 5) {
724 		/*
725 		 * This w/a is only listed for pre-production ilk a/b steppings,
726 		 * but is also mentioned for programming the powerctx. To be
727 		 * safe, just apply the workaround; we do not use SyncFlush so
728 		 * this should never take effect and so be a no-op!
729 		 */
730 		*cs++ = MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN;
731 	}
732 
733 	if (force_restore) {
734 		/*
735 		 * The HW doesn't handle being told to restore the current
736 		 * context very well. Quite often it likes goes to go off and
737 		 * sulk, especially when it is meant to be reloading PP_DIR.
738 		 * A very simple fix to force the reload is to simply switch
739 		 * away from the current context and back again.
740 		 *
741 		 * Note that the kernel_context will contain random state
742 		 * following the INHIBIT_RESTORE. We accept this since we
743 		 * never use the kernel_context state; it is merely a
744 		 * placeholder we use to flush other contexts.
745 		 */
746 		*cs++ = MI_SET_CONTEXT;
747 		*cs++ = i915_ggtt_offset(engine->kernel_context->state) |
748 			MI_MM_SPACE_GTT |
749 			MI_RESTORE_INHIBIT;
750 	}
751 
752 	*cs++ = MI_NOOP;
753 	*cs++ = MI_SET_CONTEXT;
754 	*cs++ = i915_ggtt_offset(ce->state) | flags;
755 	/*
756 	 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
757 	 * WaMiSetContext_Hang:snb,ivb,vlv
758 	 */
759 	*cs++ = MI_NOOP;
760 
761 	if (GRAPHICS_VER(i915) == 7) {
762 		if (num_engines) {
763 			struct intel_engine_cs *signaller;
764 			i915_reg_t last_reg = {}; /* keep gcc quiet */
765 
766 			*cs++ = MI_LOAD_REGISTER_IMM(num_engines);
767 			for_each_engine(signaller, engine->gt, id) {
768 				if (signaller == engine)
769 					continue;
770 
771 				last_reg = RING_PSMI_CTL(signaller->mmio_base);
772 				*cs++ = i915_mmio_reg_offset(last_reg);
773 				*cs++ = _MASKED_BIT_DISABLE(
774 						GEN6_PSMI_SLEEP_MSG_DISABLE);
775 			}
776 
777 			/* Insert a delay before the next switch! */
778 			*cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
779 			*cs++ = i915_mmio_reg_offset(last_reg);
780 			*cs++ = intel_gt_scratch_offset(engine->gt,
781 							INTEL_GT_SCRATCH_FIELD_DEFAULT);
782 			*cs++ = MI_NOOP;
783 		}
784 		*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
785 	} else if (GRAPHICS_VER(i915) == 5) {
786 		*cs++ = MI_SUSPEND_FLUSH;
787 	}
788 
789 	intel_ring_advance(rq, cs);
790 
791 	return 0;
792 }
793 
794 static int remap_l3_slice(struct i915_request *rq, int slice)
795 {
796 #define L3LOG_DW (GEN7_L3LOG_SIZE / sizeof(u32))
797 	u32 *cs, *remap_info = rq->engine->i915->l3_parity.remap_info[slice];
798 	int i;
799 
800 	if (!remap_info)
801 		return 0;
802 
803 	cs = intel_ring_begin(rq, L3LOG_DW * 2 + 2);
804 	if (IS_ERR(cs))
805 		return PTR_ERR(cs);
806 
807 	/*
808 	 * Note: We do not worry about the concurrent register cacheline hang
809 	 * here because no other code should access these registers other than
810 	 * at initialization time.
811 	 */
812 	*cs++ = MI_LOAD_REGISTER_IMM(L3LOG_DW);
813 	for (i = 0; i < L3LOG_DW; i++) {
814 		*cs++ = i915_mmio_reg_offset(GEN7_L3LOG(slice, i));
815 		*cs++ = remap_info[i];
816 	}
817 	*cs++ = MI_NOOP;
818 	intel_ring_advance(rq, cs);
819 
820 	return 0;
821 #undef L3LOG_DW
822 }
823 
824 static int remap_l3(struct i915_request *rq)
825 {
826 	struct i915_gem_context *ctx = i915_request_gem_context(rq);
827 	int i, err;
828 
829 	if (!ctx || !ctx->remap_slice)
830 		return 0;
831 
832 	for (i = 0; i < MAX_L3_SLICES; i++) {
833 		if (!(ctx->remap_slice & BIT(i)))
834 			continue;
835 
836 		err = remap_l3_slice(rq, i);
837 		if (err)
838 			return err;
839 	}
840 
841 	ctx->remap_slice = 0;
842 	return 0;
843 }
844 
845 static int switch_mm(struct i915_request *rq, struct i915_address_space *vm)
846 {
847 	int ret;
848 
849 	if (!vm)
850 		return 0;
851 
852 	ret = rq->engine->emit_flush(rq, EMIT_FLUSH);
853 	if (ret)
854 		return ret;
855 
856 	/*
857 	 * Not only do we need a full barrier (post-sync write) after
858 	 * invalidating the TLBs, but we need to wait a little bit
859 	 * longer. Whether this is merely delaying us, or the
860 	 * subsequent flush is a key part of serialising with the
861 	 * post-sync op, this extra pass appears vital before a
862 	 * mm switch!
863 	 */
864 	ret = load_pd_dir(rq, vm, PP_DIR_DCLV_2G);
865 	if (ret)
866 		return ret;
867 
868 	return rq->engine->emit_flush(rq, EMIT_INVALIDATE);
869 }
870 
871 static int clear_residuals(struct i915_request *rq)
872 {
873 	struct intel_engine_cs *engine = rq->engine;
874 	int ret;
875 
876 	ret = switch_mm(rq, vm_alias(engine->kernel_context->vm));
877 	if (ret)
878 		return ret;
879 
880 	if (engine->kernel_context->state) {
881 		ret = mi_set_context(rq,
882 				     engine->kernel_context,
883 				     MI_MM_SPACE_GTT | MI_RESTORE_INHIBIT);
884 		if (ret)
885 			return ret;
886 	}
887 
888 	ret = engine->emit_bb_start(rq,
889 				    engine->wa_ctx.vma->node.start, 0,
890 				    0);
891 	if (ret)
892 		return ret;
893 
894 	ret = engine->emit_flush(rq, EMIT_FLUSH);
895 	if (ret)
896 		return ret;
897 
898 	/* Always invalidate before the next switch_mm() */
899 	return engine->emit_flush(rq, EMIT_INVALIDATE);
900 }
901 
902 static int switch_context(struct i915_request *rq)
903 {
904 	struct intel_engine_cs *engine = rq->engine;
905 	struct intel_context *ce = rq->context;
906 	void **residuals = NULL;
907 	int ret;
908 
909 	GEM_BUG_ON(HAS_EXECLISTS(engine->i915));
910 
911 	if (engine->wa_ctx.vma && ce != engine->kernel_context) {
912 		if (engine->wa_ctx.vma->private != ce &&
913 		    i915_mitigate_clear_residuals()) {
914 			ret = clear_residuals(rq);
915 			if (ret)
916 				return ret;
917 
918 			residuals = &engine->wa_ctx.vma->private;
919 		}
920 	}
921 
922 	ret = switch_mm(rq, vm_alias(ce->vm));
923 	if (ret)
924 		return ret;
925 
926 	if (ce->state) {
927 		u32 flags;
928 
929 		GEM_BUG_ON(engine->id != RCS0);
930 
931 		/* For resource streamer on HSW+ and power context elsewhere */
932 		BUILD_BUG_ON(HSW_MI_RS_SAVE_STATE_EN != MI_SAVE_EXT_STATE_EN);
933 		BUILD_BUG_ON(HSW_MI_RS_RESTORE_STATE_EN != MI_RESTORE_EXT_STATE_EN);
934 
935 		flags = MI_SAVE_EXT_STATE_EN | MI_MM_SPACE_GTT;
936 		if (test_bit(CONTEXT_VALID_BIT, &ce->flags))
937 			flags |= MI_RESTORE_EXT_STATE_EN;
938 		else
939 			flags |= MI_RESTORE_INHIBIT;
940 
941 		ret = mi_set_context(rq, ce, flags);
942 		if (ret)
943 			return ret;
944 	}
945 
946 	ret = remap_l3(rq);
947 	if (ret)
948 		return ret;
949 
950 	/*
951 	 * Now past the point of no return, this request _will_ be emitted.
952 	 *
953 	 * Or at least this preamble will be emitted, the request may be
954 	 * interrupted prior to submitting the user payload. If so, we
955 	 * still submit the "empty" request in order to preserve global
956 	 * state tracking such as this, our tracking of the current
957 	 * dirty context.
958 	 */
959 	if (residuals) {
960 		intel_context_put(*residuals);
961 		*residuals = intel_context_get(ce);
962 	}
963 
964 	return 0;
965 }
966 
967 static int ring_request_alloc(struct i915_request *request)
968 {
969 	int ret;
970 
971 	GEM_BUG_ON(!intel_context_is_pinned(request->context));
972 	GEM_BUG_ON(i915_request_timeline(request)->has_initial_breadcrumb);
973 
974 	/*
975 	 * Flush enough space to reduce the likelihood of waiting after
976 	 * we start building the request - in which case we will just
977 	 * have to repeat work.
978 	 */
979 	request->reserved_space += LEGACY_REQUEST_SIZE;
980 
981 	/* Unconditionally invalidate GPU caches and TLBs. */
982 	ret = request->engine->emit_flush(request, EMIT_INVALIDATE);
983 	if (ret)
984 		return ret;
985 
986 	ret = switch_context(request);
987 	if (ret)
988 		return ret;
989 
990 	request->reserved_space -= LEGACY_REQUEST_SIZE;
991 	return 0;
992 }
993 
994 static void gen6_bsd_submit_request(struct i915_request *request)
995 {
996 	struct intel_uncore *uncore = request->engine->uncore;
997 
998 	intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
999 
1000        /* Every tail move must follow the sequence below */
1001 
1002 	/* Disable notification that the ring is IDLE. The GT
1003 	 * will then assume that it is busy and bring it out of rc6.
1004 	 */
1005 	intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
1006 			      _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1007 
1008 	/* Clear the context id. Here be magic! */
1009 	intel_uncore_write64_fw(uncore, GEN6_BSD_RNCID, 0x0);
1010 
1011 	/* Wait for the ring not to be idle, i.e. for it to wake up. */
1012 	if (__intel_wait_for_register_fw(uncore,
1013 					 GEN6_BSD_SLEEP_PSMI_CONTROL,
1014 					 GEN6_BSD_SLEEP_INDICATOR,
1015 					 0,
1016 					 1000, 0, NULL))
1017 		drm_err(&uncore->i915->drm,
1018 			"timed out waiting for the BSD ring to wake up\n");
1019 
1020 	/* Now that the ring is fully powered up, update the tail */
1021 	i9xx_submit_request(request);
1022 
1023 	/* Let the ring send IDLE messages to the GT again,
1024 	 * and so let it sleep to conserve power when idle.
1025 	 */
1026 	intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
1027 			      _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1028 
1029 	intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
1030 }
1031 
1032 static void i9xx_set_default_submission(struct intel_engine_cs *engine)
1033 {
1034 	engine->submit_request = i9xx_submit_request;
1035 }
1036 
1037 static void gen6_bsd_set_default_submission(struct intel_engine_cs *engine)
1038 {
1039 	engine->submit_request = gen6_bsd_submit_request;
1040 }
1041 
1042 static void ring_release(struct intel_engine_cs *engine)
1043 {
1044 	struct drm_i915_private *dev_priv = engine->i915;
1045 
1046 	drm_WARN_ON(&dev_priv->drm, GRAPHICS_VER(dev_priv) > 2 &&
1047 		    (ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
1048 
1049 	intel_engine_cleanup_common(engine);
1050 
1051 	if (engine->wa_ctx.vma) {
1052 		intel_context_put(engine->wa_ctx.vma->private);
1053 		i915_vma_unpin_and_release(&engine->wa_ctx.vma, 0);
1054 	}
1055 
1056 	intel_ring_unpin(engine->legacy.ring);
1057 	intel_ring_put(engine->legacy.ring);
1058 
1059 	intel_timeline_unpin(engine->legacy.timeline);
1060 	intel_timeline_put(engine->legacy.timeline);
1061 }
1062 
1063 static void irq_handler(struct intel_engine_cs *engine, u16 iir)
1064 {
1065 	intel_engine_signal_breadcrumbs(engine);
1066 }
1067 
1068 static void setup_irq(struct intel_engine_cs *engine)
1069 {
1070 	struct drm_i915_private *i915 = engine->i915;
1071 
1072 	intel_engine_set_irq_handler(engine, irq_handler);
1073 
1074 	if (GRAPHICS_VER(i915) >= 6) {
1075 		engine->irq_enable = gen6_irq_enable;
1076 		engine->irq_disable = gen6_irq_disable;
1077 	} else if (GRAPHICS_VER(i915) >= 5) {
1078 		engine->irq_enable = gen5_irq_enable;
1079 		engine->irq_disable = gen5_irq_disable;
1080 	} else if (GRAPHICS_VER(i915) >= 3) {
1081 		engine->irq_enable = gen3_irq_enable;
1082 		engine->irq_disable = gen3_irq_disable;
1083 	} else {
1084 		engine->irq_enable = gen2_irq_enable;
1085 		engine->irq_disable = gen2_irq_disable;
1086 	}
1087 }
1088 
1089 static void add_to_engine(struct i915_request *rq)
1090 {
1091 	lockdep_assert_held(&rq->engine->sched_engine->lock);
1092 	list_move_tail(&rq->sched.link, &rq->engine->sched_engine->requests);
1093 }
1094 
1095 static void remove_from_engine(struct i915_request *rq)
1096 {
1097 	spin_lock_irq(&rq->engine->sched_engine->lock);
1098 	list_del_init(&rq->sched.link);
1099 
1100 	/* Prevent further __await_execution() registering a cb, then flush */
1101 	set_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags);
1102 
1103 	spin_unlock_irq(&rq->engine->sched_engine->lock);
1104 
1105 	i915_request_notify_execute_cb_imm(rq);
1106 }
1107 
1108 static void setup_common(struct intel_engine_cs *engine)
1109 {
1110 	struct drm_i915_private *i915 = engine->i915;
1111 
1112 	/* gen8+ are only supported with execlists */
1113 	GEM_BUG_ON(GRAPHICS_VER(i915) >= 8);
1114 
1115 	setup_irq(engine);
1116 
1117 	engine->resume = xcs_resume;
1118 	engine->sanitize = xcs_sanitize;
1119 
1120 	engine->reset.prepare = reset_prepare;
1121 	engine->reset.rewind = reset_rewind;
1122 	engine->reset.cancel = reset_cancel;
1123 	engine->reset.finish = reset_finish;
1124 
1125 	engine->add_active_request = add_to_engine;
1126 	engine->remove_active_request = remove_from_engine;
1127 
1128 	engine->cops = &ring_context_ops;
1129 	engine->request_alloc = ring_request_alloc;
1130 
1131 	/*
1132 	 * Using a global execution timeline; the previous final breadcrumb is
1133 	 * equivalent to our next initial bread so we can elide
1134 	 * engine->emit_init_breadcrumb().
1135 	 */
1136 	engine->emit_fini_breadcrumb = gen3_emit_breadcrumb;
1137 	if (GRAPHICS_VER(i915) == 5)
1138 		engine->emit_fini_breadcrumb = gen5_emit_breadcrumb;
1139 
1140 	engine->set_default_submission = i9xx_set_default_submission;
1141 
1142 	if (GRAPHICS_VER(i915) >= 6)
1143 		engine->emit_bb_start = gen6_emit_bb_start;
1144 	else if (GRAPHICS_VER(i915) >= 4)
1145 		engine->emit_bb_start = gen4_emit_bb_start;
1146 	else if (IS_I830(i915) || IS_I845G(i915))
1147 		engine->emit_bb_start = i830_emit_bb_start;
1148 	else
1149 		engine->emit_bb_start = gen3_emit_bb_start;
1150 }
1151 
1152 static void setup_rcs(struct intel_engine_cs *engine)
1153 {
1154 	struct drm_i915_private *i915 = engine->i915;
1155 
1156 	if (HAS_L3_DPF(i915))
1157 		engine->irq_keep_mask = GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
1158 
1159 	engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
1160 
1161 	if (GRAPHICS_VER(i915) >= 7) {
1162 		engine->emit_flush = gen7_emit_flush_rcs;
1163 		engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_rcs;
1164 	} else if (GRAPHICS_VER(i915) == 6) {
1165 		engine->emit_flush = gen6_emit_flush_rcs;
1166 		engine->emit_fini_breadcrumb = gen6_emit_breadcrumb_rcs;
1167 	} else if (GRAPHICS_VER(i915) == 5) {
1168 		engine->emit_flush = gen4_emit_flush_rcs;
1169 	} else {
1170 		if (GRAPHICS_VER(i915) < 4)
1171 			engine->emit_flush = gen2_emit_flush;
1172 		else
1173 			engine->emit_flush = gen4_emit_flush_rcs;
1174 		engine->irq_enable_mask = I915_USER_INTERRUPT;
1175 	}
1176 
1177 	if (IS_HASWELL(i915))
1178 		engine->emit_bb_start = hsw_emit_bb_start;
1179 }
1180 
1181 static void setup_vcs(struct intel_engine_cs *engine)
1182 {
1183 	struct drm_i915_private *i915 = engine->i915;
1184 
1185 	if (GRAPHICS_VER(i915) >= 6) {
1186 		/* gen6 bsd needs a special wa for tail updates */
1187 		if (GRAPHICS_VER(i915) == 6)
1188 			engine->set_default_submission = gen6_bsd_set_default_submission;
1189 		engine->emit_flush = gen6_emit_flush_vcs;
1190 		engine->irq_enable_mask = GT_BSD_USER_INTERRUPT;
1191 
1192 		if (GRAPHICS_VER(i915) == 6)
1193 			engine->emit_fini_breadcrumb = gen6_emit_breadcrumb_xcs;
1194 		else
1195 			engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_xcs;
1196 	} else {
1197 		engine->emit_flush = gen4_emit_flush_vcs;
1198 		if (GRAPHICS_VER(i915) == 5)
1199 			engine->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
1200 		else
1201 			engine->irq_enable_mask = I915_BSD_USER_INTERRUPT;
1202 	}
1203 }
1204 
1205 static void setup_bcs(struct intel_engine_cs *engine)
1206 {
1207 	struct drm_i915_private *i915 = engine->i915;
1208 
1209 	engine->emit_flush = gen6_emit_flush_xcs;
1210 	engine->irq_enable_mask = GT_BLT_USER_INTERRUPT;
1211 
1212 	if (GRAPHICS_VER(i915) == 6)
1213 		engine->emit_fini_breadcrumb = gen6_emit_breadcrumb_xcs;
1214 	else
1215 		engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_xcs;
1216 }
1217 
1218 static void setup_vecs(struct intel_engine_cs *engine)
1219 {
1220 	struct drm_i915_private *i915 = engine->i915;
1221 
1222 	GEM_BUG_ON(GRAPHICS_VER(i915) < 7);
1223 
1224 	engine->emit_flush = gen6_emit_flush_xcs;
1225 	engine->irq_enable_mask = PM_VEBOX_USER_INTERRUPT;
1226 	engine->irq_enable = hsw_irq_enable_vecs;
1227 	engine->irq_disable = hsw_irq_disable_vecs;
1228 
1229 	engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_xcs;
1230 }
1231 
1232 static int gen7_ctx_switch_bb_setup(struct intel_engine_cs * const engine,
1233 				    struct i915_vma * const vma)
1234 {
1235 	return gen7_setup_clear_gpr_bb(engine, vma);
1236 }
1237 
1238 static int gen7_ctx_switch_bb_init(struct intel_engine_cs *engine,
1239 				   struct i915_gem_ww_ctx *ww,
1240 				   struct i915_vma *vma)
1241 {
1242 	int err;
1243 
1244 	err = i915_vma_pin_ww(vma, ww, 0, 0, PIN_USER | PIN_HIGH);
1245 	if (err)
1246 		return err;
1247 
1248 	err = i915_vma_sync(vma);
1249 	if (err)
1250 		goto err_unpin;
1251 
1252 	err = gen7_ctx_switch_bb_setup(engine, vma);
1253 	if (err)
1254 		goto err_unpin;
1255 
1256 	engine->wa_ctx.vma = vma;
1257 	return 0;
1258 
1259 err_unpin:
1260 	i915_vma_unpin(vma);
1261 	return err;
1262 }
1263 
1264 static struct i915_vma *gen7_ctx_vma(struct intel_engine_cs *engine)
1265 {
1266 	struct drm_i915_gem_object *obj;
1267 	struct i915_vma *vma;
1268 	int size, err;
1269 
1270 	if (GRAPHICS_VER(engine->i915) != 7 || engine->class != RENDER_CLASS)
1271 		return NULL;
1272 
1273 	err = gen7_ctx_switch_bb_setup(engine, NULL /* probe size */);
1274 	if (err < 0)
1275 		return ERR_PTR(err);
1276 	if (!err)
1277 		return NULL;
1278 
1279 	size = ALIGN(err, PAGE_SIZE);
1280 
1281 	obj = i915_gem_object_create_internal(engine->i915, size);
1282 	if (IS_ERR(obj))
1283 		return ERR_CAST(obj);
1284 
1285 	vma = i915_vma_instance(obj, engine->gt->vm, NULL);
1286 	if (IS_ERR(vma)) {
1287 		i915_gem_object_put(obj);
1288 		return ERR_CAST(vma);
1289 	}
1290 
1291 	vma->private = intel_context_create(engine); /* dummy residuals */
1292 	if (IS_ERR(vma->private)) {
1293 		err = PTR_ERR(vma->private);
1294 		vma->private = NULL;
1295 		i915_gem_object_put(obj);
1296 		return ERR_PTR(err);
1297 	}
1298 
1299 	return vma;
1300 }
1301 
1302 int intel_ring_submission_setup(struct intel_engine_cs *engine)
1303 {
1304 	struct i915_gem_ww_ctx ww;
1305 	struct intel_timeline *timeline;
1306 	struct intel_ring *ring;
1307 	struct i915_vma *gen7_wa_vma;
1308 	int err;
1309 
1310 	setup_common(engine);
1311 
1312 	switch (engine->class) {
1313 	case RENDER_CLASS:
1314 		setup_rcs(engine);
1315 		break;
1316 	case VIDEO_DECODE_CLASS:
1317 		setup_vcs(engine);
1318 		break;
1319 	case COPY_ENGINE_CLASS:
1320 		setup_bcs(engine);
1321 		break;
1322 	case VIDEO_ENHANCEMENT_CLASS:
1323 		setup_vecs(engine);
1324 		break;
1325 	default:
1326 		MISSING_CASE(engine->class);
1327 		return -ENODEV;
1328 	}
1329 
1330 	timeline = intel_timeline_create_from_engine(engine,
1331 						     I915_GEM_HWS_SEQNO_ADDR);
1332 	if (IS_ERR(timeline)) {
1333 		err = PTR_ERR(timeline);
1334 		goto err;
1335 	}
1336 	GEM_BUG_ON(timeline->has_initial_breadcrumb);
1337 
1338 	ring = intel_engine_create_ring(engine, SZ_16K);
1339 	if (IS_ERR(ring)) {
1340 		err = PTR_ERR(ring);
1341 		goto err_timeline;
1342 	}
1343 
1344 	GEM_BUG_ON(engine->legacy.ring);
1345 	engine->legacy.ring = ring;
1346 	engine->legacy.timeline = timeline;
1347 
1348 	gen7_wa_vma = gen7_ctx_vma(engine);
1349 	if (IS_ERR(gen7_wa_vma)) {
1350 		err = PTR_ERR(gen7_wa_vma);
1351 		goto err_ring;
1352 	}
1353 
1354 	i915_gem_ww_ctx_init(&ww, false);
1355 
1356 retry:
1357 	err = i915_gem_object_lock(timeline->hwsp_ggtt->obj, &ww);
1358 	if (!err && gen7_wa_vma)
1359 		err = i915_gem_object_lock(gen7_wa_vma->obj, &ww);
1360 	if (!err)
1361 		err = i915_gem_object_lock(engine->legacy.ring->vma->obj, &ww);
1362 	if (!err)
1363 		err = intel_timeline_pin(timeline, &ww);
1364 	if (!err) {
1365 		err = intel_ring_pin(ring, &ww);
1366 		if (err)
1367 			intel_timeline_unpin(timeline);
1368 	}
1369 	if (err)
1370 		goto out;
1371 
1372 	GEM_BUG_ON(timeline->hwsp_ggtt != engine->status_page.vma);
1373 
1374 	if (gen7_wa_vma) {
1375 		err = gen7_ctx_switch_bb_init(engine, &ww, gen7_wa_vma);
1376 		if (err) {
1377 			intel_ring_unpin(ring);
1378 			intel_timeline_unpin(timeline);
1379 		}
1380 	}
1381 
1382 out:
1383 	if (err == -EDEADLK) {
1384 		err = i915_gem_ww_ctx_backoff(&ww);
1385 		if (!err)
1386 			goto retry;
1387 	}
1388 	i915_gem_ww_ctx_fini(&ww);
1389 	if (err)
1390 		goto err_gen7_put;
1391 
1392 	/* Finally, take ownership and responsibility for cleanup! */
1393 	engine->release = ring_release;
1394 
1395 	return 0;
1396 
1397 err_gen7_put:
1398 	if (gen7_wa_vma) {
1399 		intel_context_put(gen7_wa_vma->private);
1400 		i915_gem_object_put(gen7_wa_vma->obj);
1401 	}
1402 err_ring:
1403 	intel_ring_put(ring);
1404 err_timeline:
1405 	intel_timeline_put(timeline);
1406 err:
1407 	intel_engine_cleanup_common(engine);
1408 	return err;
1409 }
1410 
1411 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1412 #include "selftest_ring_submission.c"
1413 #endif
1414