1 /*
2  * Copyright © 2008-2010 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  * Authors:
24  *    Eric Anholt <eric@anholt.net>
25  *    Zou Nan hai <nanhai.zou@intel.com>
26  *    Xiang Hai hao<haihao.xiang@intel.com>
27  *
28  */
29 
30 #include "gen2_engine_cs.h"
31 #include "gen6_engine_cs.h"
32 #include "gen6_ppgtt.h"
33 #include "gen7_renderclear.h"
34 #include "i915_drv.h"
35 #include "i915_mitigations.h"
36 #include "intel_breadcrumbs.h"
37 #include "intel_context.h"
38 #include "intel_gt.h"
39 #include "intel_reset.h"
40 #include "intel_ring.h"
41 #include "shmem_utils.h"
42 
43 /* Rough estimate of the typical request size, performing a flush,
44  * set-context and then emitting the batch.
45  */
46 #define LEGACY_REQUEST_SIZE 200
47 
48 static void set_hwstam(struct intel_engine_cs *engine, u32 mask)
49 {
50 	/*
51 	 * Keep the render interrupt unmasked as this papers over
52 	 * lost interrupts following a reset.
53 	 */
54 	if (engine->class == RENDER_CLASS) {
55 		if (INTEL_GEN(engine->i915) >= 6)
56 			mask &= ~BIT(0);
57 		else
58 			mask &= ~I915_USER_INTERRUPT;
59 	}
60 
61 	intel_engine_set_hwsp_writemask(engine, mask);
62 }
63 
64 static void set_hws_pga(struct intel_engine_cs *engine, phys_addr_t phys)
65 {
66 	u32 addr;
67 
68 	addr = lower_32_bits(phys);
69 	if (INTEL_GEN(engine->i915) >= 4)
70 		addr |= (phys >> 28) & 0xf0;
71 
72 	intel_uncore_write(engine->uncore, HWS_PGA, addr);
73 }
74 
75 static struct page *status_page(struct intel_engine_cs *engine)
76 {
77 	struct drm_i915_gem_object *obj = engine->status_page.vma->obj;
78 
79 	GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
80 	return sg_page(obj->mm.pages->sgl);
81 }
82 
83 static void ring_setup_phys_status_page(struct intel_engine_cs *engine)
84 {
85 	set_hws_pga(engine, PFN_PHYS(page_to_pfn(status_page(engine))));
86 	set_hwstam(engine, ~0u);
87 }
88 
89 static void set_hwsp(struct intel_engine_cs *engine, u32 offset)
90 {
91 	i915_reg_t hwsp;
92 
93 	/*
94 	 * The ring status page addresses are no longer next to the rest of
95 	 * the ring registers as of gen7.
96 	 */
97 	if (IS_GEN(engine->i915, 7)) {
98 		switch (engine->id) {
99 		/*
100 		 * No more rings exist on Gen7. Default case is only to shut up
101 		 * gcc switch check warning.
102 		 */
103 		default:
104 			GEM_BUG_ON(engine->id);
105 			fallthrough;
106 		case RCS0:
107 			hwsp = RENDER_HWS_PGA_GEN7;
108 			break;
109 		case BCS0:
110 			hwsp = BLT_HWS_PGA_GEN7;
111 			break;
112 		case VCS0:
113 			hwsp = BSD_HWS_PGA_GEN7;
114 			break;
115 		case VECS0:
116 			hwsp = VEBOX_HWS_PGA_GEN7;
117 			break;
118 		}
119 	} else if (IS_GEN(engine->i915, 6)) {
120 		hwsp = RING_HWS_PGA_GEN6(engine->mmio_base);
121 	} else {
122 		hwsp = RING_HWS_PGA(engine->mmio_base);
123 	}
124 
125 	intel_uncore_write(engine->uncore, hwsp, offset);
126 	intel_uncore_posting_read(engine->uncore, hwsp);
127 }
128 
129 static void flush_cs_tlb(struct intel_engine_cs *engine)
130 {
131 	struct drm_i915_private *dev_priv = engine->i915;
132 
133 	if (!IS_GEN_RANGE(dev_priv, 6, 7))
134 		return;
135 
136 	/* ring should be idle before issuing a sync flush*/
137 	drm_WARN_ON(&dev_priv->drm,
138 		    (ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
139 
140 	ENGINE_WRITE(engine, RING_INSTPM,
141 		     _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE |
142 					INSTPM_SYNC_FLUSH));
143 	if (intel_wait_for_register(engine->uncore,
144 				    RING_INSTPM(engine->mmio_base),
145 				    INSTPM_SYNC_FLUSH, 0,
146 				    1000))
147 		drm_err(&dev_priv->drm,
148 			"%s: wait for SyncFlush to complete for TLB invalidation timed out\n",
149 			engine->name);
150 }
151 
152 static void ring_setup_status_page(struct intel_engine_cs *engine)
153 {
154 	set_hwsp(engine, i915_ggtt_offset(engine->status_page.vma));
155 	set_hwstam(engine, ~0u);
156 
157 	flush_cs_tlb(engine);
158 }
159 
160 static bool stop_ring(struct intel_engine_cs *engine)
161 {
162 	intel_engine_stop_cs(engine);
163 
164 	ENGINE_WRITE(engine, RING_HEAD, ENGINE_READ(engine, RING_TAIL));
165 
166 	ENGINE_WRITE(engine, RING_HEAD, 0);
167 	ENGINE_WRITE(engine, RING_TAIL, 0);
168 
169 	/* The ring must be empty before it is disabled */
170 	ENGINE_WRITE(engine, RING_CTL, 0);
171 
172 	return (ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) == 0;
173 }
174 
175 static struct i915_address_space *vm_alias(struct i915_address_space *vm)
176 {
177 	if (i915_is_ggtt(vm))
178 		vm = &i915_vm_to_ggtt(vm)->alias->vm;
179 
180 	return vm;
181 }
182 
183 static u32 pp_dir(struct i915_address_space *vm)
184 {
185 	return to_gen6_ppgtt(i915_vm_to_ppgtt(vm))->pp_dir;
186 }
187 
188 static void set_pp_dir(struct intel_engine_cs *engine)
189 {
190 	struct i915_address_space *vm = vm_alias(engine->gt->vm);
191 
192 	if (vm) {
193 		ENGINE_WRITE(engine, RING_PP_DIR_DCLV, PP_DIR_DCLV_2G);
194 		ENGINE_WRITE(engine, RING_PP_DIR_BASE, pp_dir(vm));
195 	}
196 }
197 
198 static int xcs_resume(struct intel_engine_cs *engine)
199 {
200 	struct drm_i915_private *dev_priv = engine->i915;
201 	struct intel_ring *ring = engine->legacy.ring;
202 	int ret = 0;
203 
204 	ENGINE_TRACE(engine, "ring:{HEAD:%04x, TAIL:%04x}\n",
205 		     ring->head, ring->tail);
206 
207 	intel_uncore_forcewake_get(engine->uncore, FORCEWAKE_ALL);
208 
209 	/* WaClearRingBufHeadRegAtInit:ctg,elk */
210 	if (!stop_ring(engine)) {
211 		/* G45 ring initialization often fails to reset head to zero */
212 		drm_dbg(&dev_priv->drm, "%s head not reset to zero "
213 			"ctl %08x head %08x tail %08x start %08x\n",
214 			engine->name,
215 			ENGINE_READ(engine, RING_CTL),
216 			ENGINE_READ(engine, RING_HEAD),
217 			ENGINE_READ(engine, RING_TAIL),
218 			ENGINE_READ(engine, RING_START));
219 
220 		if (!stop_ring(engine)) {
221 			drm_err(&dev_priv->drm,
222 				"failed to set %s head to zero "
223 				"ctl %08x head %08x tail %08x start %08x\n",
224 				engine->name,
225 				ENGINE_READ(engine, RING_CTL),
226 				ENGINE_READ(engine, RING_HEAD),
227 				ENGINE_READ(engine, RING_TAIL),
228 				ENGINE_READ(engine, RING_START));
229 			ret = -EIO;
230 			goto out;
231 		}
232 	}
233 
234 	if (HWS_NEEDS_PHYSICAL(dev_priv))
235 		ring_setup_phys_status_page(engine);
236 	else
237 		ring_setup_status_page(engine);
238 
239 	intel_breadcrumbs_reset(engine->breadcrumbs);
240 
241 	/* Enforce ordering by reading HEAD register back */
242 	ENGINE_POSTING_READ(engine, RING_HEAD);
243 
244 	/*
245 	 * Initialize the ring. This must happen _after_ we've cleared the ring
246 	 * registers with the above sequence (the readback of the HEAD registers
247 	 * also enforces ordering), otherwise the hw might lose the new ring
248 	 * register values.
249 	 */
250 	ENGINE_WRITE(engine, RING_START, i915_ggtt_offset(ring->vma));
251 
252 	/* Check that the ring offsets point within the ring! */
253 	GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->head));
254 	GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->tail));
255 	intel_ring_update_space(ring);
256 
257 	set_pp_dir(engine);
258 
259 	/* First wake the ring up to an empty/idle ring */
260 	ENGINE_WRITE(engine, RING_HEAD, ring->head);
261 	ENGINE_WRITE(engine, RING_TAIL, ring->head);
262 	ENGINE_POSTING_READ(engine, RING_TAIL);
263 
264 	ENGINE_WRITE(engine, RING_CTL, RING_CTL_SIZE(ring->size) | RING_VALID);
265 
266 	/* If the head is still not zero, the ring is dead */
267 	if (intel_wait_for_register(engine->uncore,
268 				    RING_CTL(engine->mmio_base),
269 				    RING_VALID, RING_VALID,
270 				    50)) {
271 		drm_err(&dev_priv->drm, "%s initialization failed "
272 			  "ctl %08x (valid? %d) head %08x [%08x] tail %08x [%08x] start %08x [expected %08x]\n",
273 			  engine->name,
274 			  ENGINE_READ(engine, RING_CTL),
275 			  ENGINE_READ(engine, RING_CTL) & RING_VALID,
276 			  ENGINE_READ(engine, RING_HEAD), ring->head,
277 			  ENGINE_READ(engine, RING_TAIL), ring->tail,
278 			  ENGINE_READ(engine, RING_START),
279 			  i915_ggtt_offset(ring->vma));
280 		ret = -EIO;
281 		goto out;
282 	}
283 
284 	if (INTEL_GEN(dev_priv) > 2)
285 		ENGINE_WRITE(engine,
286 			     RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
287 
288 	/* Now awake, let it get started */
289 	if (ring->tail != ring->head) {
290 		ENGINE_WRITE(engine, RING_TAIL, ring->tail);
291 		ENGINE_POSTING_READ(engine, RING_TAIL);
292 	}
293 
294 	/* Papering over lost _interrupts_ immediately following the restart */
295 	intel_engine_signal_breadcrumbs(engine);
296 out:
297 	intel_uncore_forcewake_put(engine->uncore, FORCEWAKE_ALL);
298 
299 	return ret;
300 }
301 
302 static void sanitize_hwsp(struct intel_engine_cs *engine)
303 {
304 	struct intel_timeline *tl;
305 
306 	list_for_each_entry(tl, &engine->status_page.timelines, engine_link)
307 		intel_timeline_reset_seqno(tl);
308 }
309 
310 static void xcs_sanitize(struct intel_engine_cs *engine)
311 {
312 	/*
313 	 * Poison residual state on resume, in case the suspend didn't!
314 	 *
315 	 * We have to assume that across suspend/resume (or other loss
316 	 * of control) that the contents of our pinned buffers has been
317 	 * lost, replaced by garbage. Since this doesn't always happen,
318 	 * let's poison such state so that we more quickly spot when
319 	 * we falsely assume it has been preserved.
320 	 */
321 	if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
322 		memset(engine->status_page.addr, POISON_INUSE, PAGE_SIZE);
323 
324 	/*
325 	 * The kernel_context HWSP is stored in the status_page. As above,
326 	 * that may be lost on resume/initialisation, and so we need to
327 	 * reset the value in the HWSP.
328 	 */
329 	sanitize_hwsp(engine);
330 
331 	/* And scrub the dirty cachelines for the HWSP */
332 	clflush_cache_range(engine->status_page.addr, PAGE_SIZE);
333 }
334 
335 static void reset_prepare(struct intel_engine_cs *engine)
336 {
337 	struct intel_uncore *uncore = engine->uncore;
338 	const u32 base = engine->mmio_base;
339 
340 	/*
341 	 * We stop engines, otherwise we might get failed reset and a
342 	 * dead gpu (on elk). Also as modern gpu as kbl can suffer
343 	 * from system hang if batchbuffer is progressing when
344 	 * the reset is issued, regardless of READY_TO_RESET ack.
345 	 * Thus assume it is best to stop engines on all gens
346 	 * where we have a gpu reset.
347 	 *
348 	 * WaKBLVECSSemaphoreWaitPoll:kbl (on ALL_ENGINES)
349 	 *
350 	 * WaMediaResetMainRingCleanup:ctg,elk (presumably)
351 	 *
352 	 * FIXME: Wa for more modern gens needs to be validated
353 	 */
354 	ENGINE_TRACE(engine, "\n");
355 
356 	if (intel_engine_stop_cs(engine))
357 		ENGINE_TRACE(engine, "timed out on STOP_RING\n");
358 
359 	intel_uncore_write_fw(uncore,
360 			      RING_HEAD(base),
361 			      intel_uncore_read_fw(uncore, RING_TAIL(base)));
362 	intel_uncore_posting_read_fw(uncore, RING_HEAD(base)); /* paranoia */
363 
364 	intel_uncore_write_fw(uncore, RING_HEAD(base), 0);
365 	intel_uncore_write_fw(uncore, RING_TAIL(base), 0);
366 	intel_uncore_posting_read_fw(uncore, RING_TAIL(base));
367 
368 	/* The ring must be empty before it is disabled */
369 	intel_uncore_write_fw(uncore, RING_CTL(base), 0);
370 
371 	/* Check acts as a post */
372 	if (intel_uncore_read_fw(uncore, RING_HEAD(base)))
373 		ENGINE_TRACE(engine, "ring head [%x] not parked\n",
374 			     intel_uncore_read_fw(uncore, RING_HEAD(base)));
375 }
376 
377 static void reset_rewind(struct intel_engine_cs *engine, bool stalled)
378 {
379 	struct i915_request *pos, *rq;
380 	unsigned long flags;
381 	u32 head;
382 
383 	rq = NULL;
384 	spin_lock_irqsave(&engine->active.lock, flags);
385 	list_for_each_entry(pos, &engine->active.requests, sched.link) {
386 		if (!i915_request_completed(pos)) {
387 			rq = pos;
388 			break;
389 		}
390 	}
391 
392 	/*
393 	 * The guilty request will get skipped on a hung engine.
394 	 *
395 	 * Users of client default contexts do not rely on logical
396 	 * state preserved between batches so it is safe to execute
397 	 * queued requests following the hang. Non default contexts
398 	 * rely on preserved state, so skipping a batch loses the
399 	 * evolution of the state and it needs to be considered corrupted.
400 	 * Executing more queued batches on top of corrupted state is
401 	 * risky. But we take the risk by trying to advance through
402 	 * the queued requests in order to make the client behaviour
403 	 * more predictable around resets, by not throwing away random
404 	 * amount of batches it has prepared for execution. Sophisticated
405 	 * clients can use gem_reset_stats_ioctl and dma fence status
406 	 * (exported via sync_file info ioctl on explicit fences) to observe
407 	 * when it loses the context state and should rebuild accordingly.
408 	 *
409 	 * The context ban, and ultimately the client ban, mechanism are safety
410 	 * valves if client submission ends up resulting in nothing more than
411 	 * subsequent hangs.
412 	 */
413 
414 	if (rq) {
415 		/*
416 		 * Try to restore the logical GPU state to match the
417 		 * continuation of the request queue. If we skip the
418 		 * context/PD restore, then the next request may try to execute
419 		 * assuming that its context is valid and loaded on the GPU and
420 		 * so may try to access invalid memory, prompting repeated GPU
421 		 * hangs.
422 		 *
423 		 * If the request was guilty, we still restore the logical
424 		 * state in case the next request requires it (e.g. the
425 		 * aliasing ppgtt), but skip over the hung batch.
426 		 *
427 		 * If the request was innocent, we try to replay the request
428 		 * with the restored context.
429 		 */
430 		__i915_request_reset(rq, stalled);
431 
432 		GEM_BUG_ON(rq->ring != engine->legacy.ring);
433 		head = rq->head;
434 	} else {
435 		head = engine->legacy.ring->tail;
436 	}
437 	engine->legacy.ring->head = intel_ring_wrap(engine->legacy.ring, head);
438 
439 	spin_unlock_irqrestore(&engine->active.lock, flags);
440 }
441 
442 static void reset_finish(struct intel_engine_cs *engine)
443 {
444 }
445 
446 static void reset_cancel(struct intel_engine_cs *engine)
447 {
448 	struct i915_request *request;
449 	unsigned long flags;
450 
451 	spin_lock_irqsave(&engine->active.lock, flags);
452 
453 	/* Mark all submitted requests as skipped. */
454 	list_for_each_entry(request, &engine->active.requests, sched.link)
455 		i915_request_mark_eio(request);
456 	intel_engine_signal_breadcrumbs(engine);
457 
458 	/* Remaining _unready_ requests will be nop'ed when submitted */
459 
460 	spin_unlock_irqrestore(&engine->active.lock, flags);
461 }
462 
463 static void i9xx_submit_request(struct i915_request *request)
464 {
465 	i915_request_submit(request);
466 	wmb(); /* paranoid flush writes out of the WCB before mmio */
467 
468 	ENGINE_WRITE(request->engine, RING_TAIL,
469 		     intel_ring_set_tail(request->ring, request->tail));
470 }
471 
472 static void __ring_context_fini(struct intel_context *ce)
473 {
474 	i915_vma_put(ce->state);
475 }
476 
477 static void ring_context_destroy(struct kref *ref)
478 {
479 	struct intel_context *ce = container_of(ref, typeof(*ce), ref);
480 
481 	GEM_BUG_ON(intel_context_is_pinned(ce));
482 
483 	if (ce->state)
484 		__ring_context_fini(ce);
485 
486 	intel_context_fini(ce);
487 	intel_context_free(ce);
488 }
489 
490 static int ring_context_pre_pin(struct intel_context *ce,
491 				struct i915_gem_ww_ctx *ww,
492 				void **unused)
493 {
494 	struct i915_address_space *vm;
495 	int err = 0;
496 
497 	vm = vm_alias(ce->vm);
498 	if (vm)
499 		err = gen6_ppgtt_pin(i915_vm_to_ppgtt((vm)), ww);
500 
501 	return err;
502 }
503 
504 static void __context_unpin_ppgtt(struct intel_context *ce)
505 {
506 	struct i915_address_space *vm;
507 
508 	vm = vm_alias(ce->vm);
509 	if (vm)
510 		gen6_ppgtt_unpin(i915_vm_to_ppgtt(vm));
511 }
512 
513 static void ring_context_unpin(struct intel_context *ce)
514 {
515 }
516 
517 static void ring_context_post_unpin(struct intel_context *ce)
518 {
519 	__context_unpin_ppgtt(ce);
520 }
521 
522 static struct i915_vma *
523 alloc_context_vma(struct intel_engine_cs *engine)
524 {
525 	struct drm_i915_private *i915 = engine->i915;
526 	struct drm_i915_gem_object *obj;
527 	struct i915_vma *vma;
528 	int err;
529 
530 	obj = i915_gem_object_create_shmem(i915, engine->context_size);
531 	if (IS_ERR(obj))
532 		return ERR_CAST(obj);
533 
534 	/*
535 	 * Try to make the context utilize L3 as well as LLC.
536 	 *
537 	 * On VLV we don't have L3 controls in the PTEs so we
538 	 * shouldn't touch the cache level, especially as that
539 	 * would make the object snooped which might have a
540 	 * negative performance impact.
541 	 *
542 	 * Snooping is required on non-llc platforms in execlist
543 	 * mode, but since all GGTT accesses use PAT entry 0 we
544 	 * get snooping anyway regardless of cache_level.
545 	 *
546 	 * This is only applicable for Ivy Bridge devices since
547 	 * later platforms don't have L3 control bits in the PTE.
548 	 */
549 	if (IS_IVYBRIDGE(i915))
550 		i915_gem_object_set_cache_coherency(obj, I915_CACHE_L3_LLC);
551 
552 	if (engine->default_state) {
553 		void *vaddr;
554 
555 		vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
556 		if (IS_ERR(vaddr)) {
557 			err = PTR_ERR(vaddr);
558 			goto err_obj;
559 		}
560 
561 		shmem_read(engine->default_state, 0,
562 			   vaddr, engine->context_size);
563 
564 		i915_gem_object_flush_map(obj);
565 		__i915_gem_object_release_map(obj);
566 	}
567 
568 	vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
569 	if (IS_ERR(vma)) {
570 		err = PTR_ERR(vma);
571 		goto err_obj;
572 	}
573 
574 	return vma;
575 
576 err_obj:
577 	i915_gem_object_put(obj);
578 	return ERR_PTR(err);
579 }
580 
581 static int ring_context_alloc(struct intel_context *ce)
582 {
583 	struct intel_engine_cs *engine = ce->engine;
584 
585 	/* One ringbuffer to rule them all */
586 	GEM_BUG_ON(!engine->legacy.ring);
587 	ce->ring = engine->legacy.ring;
588 	ce->timeline = intel_timeline_get(engine->legacy.timeline);
589 
590 	GEM_BUG_ON(ce->state);
591 	if (engine->context_size) {
592 		struct i915_vma *vma;
593 
594 		vma = alloc_context_vma(engine);
595 		if (IS_ERR(vma))
596 			return PTR_ERR(vma);
597 
598 		ce->state = vma;
599 		if (engine->default_state)
600 			__set_bit(CONTEXT_VALID_BIT, &ce->flags);
601 	}
602 
603 	return 0;
604 }
605 
606 static int ring_context_pin(struct intel_context *ce, void *unused)
607 {
608 	return 0;
609 }
610 
611 static void ring_context_reset(struct intel_context *ce)
612 {
613 	intel_ring_reset(ce->ring, ce->ring->emit);
614 	clear_bit(CONTEXT_VALID_BIT, &ce->flags);
615 }
616 
617 static const struct intel_context_ops ring_context_ops = {
618 	.alloc = ring_context_alloc,
619 
620 	.pre_pin = ring_context_pre_pin,
621 	.pin = ring_context_pin,
622 	.unpin = ring_context_unpin,
623 	.post_unpin = ring_context_post_unpin,
624 
625 	.enter = intel_context_enter_engine,
626 	.exit = intel_context_exit_engine,
627 
628 	.reset = ring_context_reset,
629 	.destroy = ring_context_destroy,
630 };
631 
632 static int load_pd_dir(struct i915_request *rq,
633 		       struct i915_address_space *vm,
634 		       u32 valid)
635 {
636 	const struct intel_engine_cs * const engine = rq->engine;
637 	u32 *cs;
638 
639 	cs = intel_ring_begin(rq, 12);
640 	if (IS_ERR(cs))
641 		return PTR_ERR(cs);
642 
643 	*cs++ = MI_LOAD_REGISTER_IMM(1);
644 	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_DCLV(engine->mmio_base));
645 	*cs++ = valid;
646 
647 	*cs++ = MI_LOAD_REGISTER_IMM(1);
648 	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
649 	*cs++ = pp_dir(vm);
650 
651 	/* Stall until the page table load is complete? */
652 	*cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
653 	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
654 	*cs++ = intel_gt_scratch_offset(engine->gt,
655 					INTEL_GT_SCRATCH_FIELD_DEFAULT);
656 
657 	*cs++ = MI_LOAD_REGISTER_IMM(1);
658 	*cs++ = i915_mmio_reg_offset(RING_INSTPM(engine->mmio_base));
659 	*cs++ = _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE);
660 
661 	intel_ring_advance(rq, cs);
662 
663 	return rq->engine->emit_flush(rq, EMIT_FLUSH);
664 }
665 
666 static inline int mi_set_context(struct i915_request *rq,
667 				 struct intel_context *ce,
668 				 u32 flags)
669 {
670 	struct intel_engine_cs *engine = rq->engine;
671 	struct drm_i915_private *i915 = engine->i915;
672 	enum intel_engine_id id;
673 	const int num_engines =
674 		IS_HASWELL(i915) ? engine->gt->info.num_engines - 1 : 0;
675 	bool force_restore = false;
676 	int len;
677 	u32 *cs;
678 
679 	len = 4;
680 	if (IS_GEN(i915, 7))
681 		len += 2 + (num_engines ? 4 * num_engines + 6 : 0);
682 	else if (IS_GEN(i915, 5))
683 		len += 2;
684 	if (flags & MI_FORCE_RESTORE) {
685 		GEM_BUG_ON(flags & MI_RESTORE_INHIBIT);
686 		flags &= ~MI_FORCE_RESTORE;
687 		force_restore = true;
688 		len += 2;
689 	}
690 
691 	cs = intel_ring_begin(rq, len);
692 	if (IS_ERR(cs))
693 		return PTR_ERR(cs);
694 
695 	/* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
696 	if (IS_GEN(i915, 7)) {
697 		*cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
698 		if (num_engines) {
699 			struct intel_engine_cs *signaller;
700 
701 			*cs++ = MI_LOAD_REGISTER_IMM(num_engines);
702 			for_each_engine(signaller, engine->gt, id) {
703 				if (signaller == engine)
704 					continue;
705 
706 				*cs++ = i915_mmio_reg_offset(
707 					   RING_PSMI_CTL(signaller->mmio_base));
708 				*cs++ = _MASKED_BIT_ENABLE(
709 						GEN6_PSMI_SLEEP_MSG_DISABLE);
710 			}
711 		}
712 	} else if (IS_GEN(i915, 5)) {
713 		/*
714 		 * This w/a is only listed for pre-production ilk a/b steppings,
715 		 * but is also mentioned for programming the powerctx. To be
716 		 * safe, just apply the workaround; we do not use SyncFlush so
717 		 * this should never take effect and so be a no-op!
718 		 */
719 		*cs++ = MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN;
720 	}
721 
722 	if (force_restore) {
723 		/*
724 		 * The HW doesn't handle being told to restore the current
725 		 * context very well. Quite often it likes goes to go off and
726 		 * sulk, especially when it is meant to be reloading PP_DIR.
727 		 * A very simple fix to force the reload is to simply switch
728 		 * away from the current context and back again.
729 		 *
730 		 * Note that the kernel_context will contain random state
731 		 * following the INHIBIT_RESTORE. We accept this since we
732 		 * never use the kernel_context state; it is merely a
733 		 * placeholder we use to flush other contexts.
734 		 */
735 		*cs++ = MI_SET_CONTEXT;
736 		*cs++ = i915_ggtt_offset(engine->kernel_context->state) |
737 			MI_MM_SPACE_GTT |
738 			MI_RESTORE_INHIBIT;
739 	}
740 
741 	*cs++ = MI_NOOP;
742 	*cs++ = MI_SET_CONTEXT;
743 	*cs++ = i915_ggtt_offset(ce->state) | flags;
744 	/*
745 	 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
746 	 * WaMiSetContext_Hang:snb,ivb,vlv
747 	 */
748 	*cs++ = MI_NOOP;
749 
750 	if (IS_GEN(i915, 7)) {
751 		if (num_engines) {
752 			struct intel_engine_cs *signaller;
753 			i915_reg_t last_reg = {}; /* keep gcc quiet */
754 
755 			*cs++ = MI_LOAD_REGISTER_IMM(num_engines);
756 			for_each_engine(signaller, engine->gt, id) {
757 				if (signaller == engine)
758 					continue;
759 
760 				last_reg = RING_PSMI_CTL(signaller->mmio_base);
761 				*cs++ = i915_mmio_reg_offset(last_reg);
762 				*cs++ = _MASKED_BIT_DISABLE(
763 						GEN6_PSMI_SLEEP_MSG_DISABLE);
764 			}
765 
766 			/* Insert a delay before the next switch! */
767 			*cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
768 			*cs++ = i915_mmio_reg_offset(last_reg);
769 			*cs++ = intel_gt_scratch_offset(engine->gt,
770 							INTEL_GT_SCRATCH_FIELD_DEFAULT);
771 			*cs++ = MI_NOOP;
772 		}
773 		*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
774 	} else if (IS_GEN(i915, 5)) {
775 		*cs++ = MI_SUSPEND_FLUSH;
776 	}
777 
778 	intel_ring_advance(rq, cs);
779 
780 	return 0;
781 }
782 
783 static int remap_l3_slice(struct i915_request *rq, int slice)
784 {
785 	u32 *cs, *remap_info = rq->engine->i915->l3_parity.remap_info[slice];
786 	int i;
787 
788 	if (!remap_info)
789 		return 0;
790 
791 	cs = intel_ring_begin(rq, GEN7_L3LOG_SIZE/4 * 2 + 2);
792 	if (IS_ERR(cs))
793 		return PTR_ERR(cs);
794 
795 	/*
796 	 * Note: We do not worry about the concurrent register cacheline hang
797 	 * here because no other code should access these registers other than
798 	 * at initialization time.
799 	 */
800 	*cs++ = MI_LOAD_REGISTER_IMM(GEN7_L3LOG_SIZE/4);
801 	for (i = 0; i < GEN7_L3LOG_SIZE/4; i++) {
802 		*cs++ = i915_mmio_reg_offset(GEN7_L3LOG(slice, i));
803 		*cs++ = remap_info[i];
804 	}
805 	*cs++ = MI_NOOP;
806 	intel_ring_advance(rq, cs);
807 
808 	return 0;
809 }
810 
811 static int remap_l3(struct i915_request *rq)
812 {
813 	struct i915_gem_context *ctx = i915_request_gem_context(rq);
814 	int i, err;
815 
816 	if (!ctx || !ctx->remap_slice)
817 		return 0;
818 
819 	for (i = 0; i < MAX_L3_SLICES; i++) {
820 		if (!(ctx->remap_slice & BIT(i)))
821 			continue;
822 
823 		err = remap_l3_slice(rq, i);
824 		if (err)
825 			return err;
826 	}
827 
828 	ctx->remap_slice = 0;
829 	return 0;
830 }
831 
832 static int switch_mm(struct i915_request *rq, struct i915_address_space *vm)
833 {
834 	int ret;
835 
836 	if (!vm)
837 		return 0;
838 
839 	ret = rq->engine->emit_flush(rq, EMIT_FLUSH);
840 	if (ret)
841 		return ret;
842 
843 	/*
844 	 * Not only do we need a full barrier (post-sync write) after
845 	 * invalidating the TLBs, but we need to wait a little bit
846 	 * longer. Whether this is merely delaying us, or the
847 	 * subsequent flush is a key part of serialising with the
848 	 * post-sync op, this extra pass appears vital before a
849 	 * mm switch!
850 	 */
851 	ret = load_pd_dir(rq, vm, PP_DIR_DCLV_2G);
852 	if (ret)
853 		return ret;
854 
855 	return rq->engine->emit_flush(rq, EMIT_INVALIDATE);
856 }
857 
858 static int clear_residuals(struct i915_request *rq)
859 {
860 	struct intel_engine_cs *engine = rq->engine;
861 	int ret;
862 
863 	ret = switch_mm(rq, vm_alias(engine->kernel_context->vm));
864 	if (ret)
865 		return ret;
866 
867 	if (engine->kernel_context->state) {
868 		ret = mi_set_context(rq,
869 				     engine->kernel_context,
870 				     MI_MM_SPACE_GTT | MI_RESTORE_INHIBIT);
871 		if (ret)
872 			return ret;
873 	}
874 
875 	ret = engine->emit_bb_start(rq,
876 				    engine->wa_ctx.vma->node.start, 0,
877 				    0);
878 	if (ret)
879 		return ret;
880 
881 	ret = engine->emit_flush(rq, EMIT_FLUSH);
882 	if (ret)
883 		return ret;
884 
885 	/* Always invalidate before the next switch_mm() */
886 	return engine->emit_flush(rq, EMIT_INVALIDATE);
887 }
888 
889 static int switch_context(struct i915_request *rq)
890 {
891 	struct intel_engine_cs *engine = rq->engine;
892 	struct intel_context *ce = rq->context;
893 	void **residuals = NULL;
894 	int ret;
895 
896 	GEM_BUG_ON(HAS_EXECLISTS(engine->i915));
897 
898 	if (engine->wa_ctx.vma && ce != engine->kernel_context) {
899 		if (engine->wa_ctx.vma->private != ce &&
900 		    i915_mitigate_clear_residuals()) {
901 			ret = clear_residuals(rq);
902 			if (ret)
903 				return ret;
904 
905 			residuals = &engine->wa_ctx.vma->private;
906 		}
907 	}
908 
909 	ret = switch_mm(rq, vm_alias(ce->vm));
910 	if (ret)
911 		return ret;
912 
913 	if (ce->state) {
914 		u32 flags;
915 
916 		GEM_BUG_ON(engine->id != RCS0);
917 
918 		/* For resource streamer on HSW+ and power context elsewhere */
919 		BUILD_BUG_ON(HSW_MI_RS_SAVE_STATE_EN != MI_SAVE_EXT_STATE_EN);
920 		BUILD_BUG_ON(HSW_MI_RS_RESTORE_STATE_EN != MI_RESTORE_EXT_STATE_EN);
921 
922 		flags = MI_SAVE_EXT_STATE_EN | MI_MM_SPACE_GTT;
923 		if (test_bit(CONTEXT_VALID_BIT, &ce->flags))
924 			flags |= MI_RESTORE_EXT_STATE_EN;
925 		else
926 			flags |= MI_RESTORE_INHIBIT;
927 
928 		ret = mi_set_context(rq, ce, flags);
929 		if (ret)
930 			return ret;
931 	}
932 
933 	ret = remap_l3(rq);
934 	if (ret)
935 		return ret;
936 
937 	/*
938 	 * Now past the point of no return, this request _will_ be emitted.
939 	 *
940 	 * Or at least this preamble will be emitted, the request may be
941 	 * interrupted prior to submitting the user payload. If so, we
942 	 * still submit the "empty" request in order to preserve global
943 	 * state tracking such as this, our tracking of the current
944 	 * dirty context.
945 	 */
946 	if (residuals) {
947 		intel_context_put(*residuals);
948 		*residuals = intel_context_get(ce);
949 	}
950 
951 	return 0;
952 }
953 
954 static int ring_request_alloc(struct i915_request *request)
955 {
956 	int ret;
957 
958 	GEM_BUG_ON(!intel_context_is_pinned(request->context));
959 	GEM_BUG_ON(i915_request_timeline(request)->has_initial_breadcrumb);
960 
961 	/*
962 	 * Flush enough space to reduce the likelihood of waiting after
963 	 * we start building the request - in which case we will just
964 	 * have to repeat work.
965 	 */
966 	request->reserved_space += LEGACY_REQUEST_SIZE;
967 
968 	/* Unconditionally invalidate GPU caches and TLBs. */
969 	ret = request->engine->emit_flush(request, EMIT_INVALIDATE);
970 	if (ret)
971 		return ret;
972 
973 	ret = switch_context(request);
974 	if (ret)
975 		return ret;
976 
977 	request->reserved_space -= LEGACY_REQUEST_SIZE;
978 	return 0;
979 }
980 
981 static void gen6_bsd_submit_request(struct i915_request *request)
982 {
983 	struct intel_uncore *uncore = request->engine->uncore;
984 
985 	intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
986 
987        /* Every tail move must follow the sequence below */
988 
989 	/* Disable notification that the ring is IDLE. The GT
990 	 * will then assume that it is busy and bring it out of rc6.
991 	 */
992 	intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
993 			      _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
994 
995 	/* Clear the context id. Here be magic! */
996 	intel_uncore_write64_fw(uncore, GEN6_BSD_RNCID, 0x0);
997 
998 	/* Wait for the ring not to be idle, i.e. for it to wake up. */
999 	if (__intel_wait_for_register_fw(uncore,
1000 					 GEN6_BSD_SLEEP_PSMI_CONTROL,
1001 					 GEN6_BSD_SLEEP_INDICATOR,
1002 					 0,
1003 					 1000, 0, NULL))
1004 		drm_err(&uncore->i915->drm,
1005 			"timed out waiting for the BSD ring to wake up\n");
1006 
1007 	/* Now that the ring is fully powered up, update the tail */
1008 	i9xx_submit_request(request);
1009 
1010 	/* Let the ring send IDLE messages to the GT again,
1011 	 * and so let it sleep to conserve power when idle.
1012 	 */
1013 	intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
1014 			      _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1015 
1016 	intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
1017 }
1018 
1019 static void i9xx_set_default_submission(struct intel_engine_cs *engine)
1020 {
1021 	engine->submit_request = i9xx_submit_request;
1022 
1023 	engine->park = NULL;
1024 	engine->unpark = NULL;
1025 }
1026 
1027 static void gen6_bsd_set_default_submission(struct intel_engine_cs *engine)
1028 {
1029 	i9xx_set_default_submission(engine);
1030 	engine->submit_request = gen6_bsd_submit_request;
1031 }
1032 
1033 static void ring_release(struct intel_engine_cs *engine)
1034 {
1035 	struct drm_i915_private *dev_priv = engine->i915;
1036 
1037 	drm_WARN_ON(&dev_priv->drm, INTEL_GEN(dev_priv) > 2 &&
1038 		    (ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
1039 
1040 	intel_engine_cleanup_common(engine);
1041 
1042 	if (engine->wa_ctx.vma) {
1043 		intel_context_put(engine->wa_ctx.vma->private);
1044 		i915_vma_unpin_and_release(&engine->wa_ctx.vma, 0);
1045 	}
1046 
1047 	intel_ring_unpin(engine->legacy.ring);
1048 	intel_ring_put(engine->legacy.ring);
1049 
1050 	intel_timeline_unpin(engine->legacy.timeline);
1051 	intel_timeline_put(engine->legacy.timeline);
1052 }
1053 
1054 static void setup_irq(struct intel_engine_cs *engine)
1055 {
1056 	struct drm_i915_private *i915 = engine->i915;
1057 
1058 	if (INTEL_GEN(i915) >= 6) {
1059 		engine->irq_enable = gen6_irq_enable;
1060 		engine->irq_disable = gen6_irq_disable;
1061 	} else if (INTEL_GEN(i915) >= 5) {
1062 		engine->irq_enable = gen5_irq_enable;
1063 		engine->irq_disable = gen5_irq_disable;
1064 	} else if (INTEL_GEN(i915) >= 3) {
1065 		engine->irq_enable = gen3_irq_enable;
1066 		engine->irq_disable = gen3_irq_disable;
1067 	} else {
1068 		engine->irq_enable = gen2_irq_enable;
1069 		engine->irq_disable = gen2_irq_disable;
1070 	}
1071 }
1072 
1073 static void setup_common(struct intel_engine_cs *engine)
1074 {
1075 	struct drm_i915_private *i915 = engine->i915;
1076 
1077 	/* gen8+ are only supported with execlists */
1078 	GEM_BUG_ON(INTEL_GEN(i915) >= 8);
1079 
1080 	setup_irq(engine);
1081 
1082 	engine->resume = xcs_resume;
1083 	engine->sanitize = xcs_sanitize;
1084 
1085 	engine->reset.prepare = reset_prepare;
1086 	engine->reset.rewind = reset_rewind;
1087 	engine->reset.cancel = reset_cancel;
1088 	engine->reset.finish = reset_finish;
1089 
1090 	engine->cops = &ring_context_ops;
1091 	engine->request_alloc = ring_request_alloc;
1092 
1093 	/*
1094 	 * Using a global execution timeline; the previous final breadcrumb is
1095 	 * equivalent to our next initial bread so we can elide
1096 	 * engine->emit_init_breadcrumb().
1097 	 */
1098 	engine->emit_fini_breadcrumb = gen3_emit_breadcrumb;
1099 	if (IS_GEN(i915, 5))
1100 		engine->emit_fini_breadcrumb = gen5_emit_breadcrumb;
1101 
1102 	engine->set_default_submission = i9xx_set_default_submission;
1103 
1104 	if (INTEL_GEN(i915) >= 6)
1105 		engine->emit_bb_start = gen6_emit_bb_start;
1106 	else if (INTEL_GEN(i915) >= 4)
1107 		engine->emit_bb_start = gen4_emit_bb_start;
1108 	else if (IS_I830(i915) || IS_I845G(i915))
1109 		engine->emit_bb_start = i830_emit_bb_start;
1110 	else
1111 		engine->emit_bb_start = gen3_emit_bb_start;
1112 }
1113 
1114 static void setup_rcs(struct intel_engine_cs *engine)
1115 {
1116 	struct drm_i915_private *i915 = engine->i915;
1117 
1118 	if (HAS_L3_DPF(i915))
1119 		engine->irq_keep_mask = GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
1120 
1121 	engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
1122 
1123 	if (INTEL_GEN(i915) >= 7) {
1124 		engine->emit_flush = gen7_emit_flush_rcs;
1125 		engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_rcs;
1126 	} else if (IS_GEN(i915, 6)) {
1127 		engine->emit_flush = gen6_emit_flush_rcs;
1128 		engine->emit_fini_breadcrumb = gen6_emit_breadcrumb_rcs;
1129 	} else if (IS_GEN(i915, 5)) {
1130 		engine->emit_flush = gen4_emit_flush_rcs;
1131 	} else {
1132 		if (INTEL_GEN(i915) < 4)
1133 			engine->emit_flush = gen2_emit_flush;
1134 		else
1135 			engine->emit_flush = gen4_emit_flush_rcs;
1136 		engine->irq_enable_mask = I915_USER_INTERRUPT;
1137 	}
1138 
1139 	if (IS_HASWELL(i915))
1140 		engine->emit_bb_start = hsw_emit_bb_start;
1141 }
1142 
1143 static void setup_vcs(struct intel_engine_cs *engine)
1144 {
1145 	struct drm_i915_private *i915 = engine->i915;
1146 
1147 	if (INTEL_GEN(i915) >= 6) {
1148 		/* gen6 bsd needs a special wa for tail updates */
1149 		if (IS_GEN(i915, 6))
1150 			engine->set_default_submission = gen6_bsd_set_default_submission;
1151 		engine->emit_flush = gen6_emit_flush_vcs;
1152 		engine->irq_enable_mask = GT_BSD_USER_INTERRUPT;
1153 
1154 		if (IS_GEN(i915, 6))
1155 			engine->emit_fini_breadcrumb = gen6_emit_breadcrumb_xcs;
1156 		else
1157 			engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_xcs;
1158 	} else {
1159 		engine->emit_flush = gen4_emit_flush_vcs;
1160 		if (IS_GEN(i915, 5))
1161 			engine->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
1162 		else
1163 			engine->irq_enable_mask = I915_BSD_USER_INTERRUPT;
1164 	}
1165 }
1166 
1167 static void setup_bcs(struct intel_engine_cs *engine)
1168 {
1169 	struct drm_i915_private *i915 = engine->i915;
1170 
1171 	engine->emit_flush = gen6_emit_flush_xcs;
1172 	engine->irq_enable_mask = GT_BLT_USER_INTERRUPT;
1173 
1174 	if (IS_GEN(i915, 6))
1175 		engine->emit_fini_breadcrumb = gen6_emit_breadcrumb_xcs;
1176 	else
1177 		engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_xcs;
1178 }
1179 
1180 static void setup_vecs(struct intel_engine_cs *engine)
1181 {
1182 	struct drm_i915_private *i915 = engine->i915;
1183 
1184 	GEM_BUG_ON(INTEL_GEN(i915) < 7);
1185 
1186 	engine->emit_flush = gen6_emit_flush_xcs;
1187 	engine->irq_enable_mask = PM_VEBOX_USER_INTERRUPT;
1188 	engine->irq_enable = hsw_irq_enable_vecs;
1189 	engine->irq_disable = hsw_irq_disable_vecs;
1190 
1191 	engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_xcs;
1192 }
1193 
1194 static int gen7_ctx_switch_bb_setup(struct intel_engine_cs * const engine,
1195 				    struct i915_vma * const vma)
1196 {
1197 	return gen7_setup_clear_gpr_bb(engine, vma);
1198 }
1199 
1200 static int gen7_ctx_switch_bb_init(struct intel_engine_cs *engine)
1201 {
1202 	struct drm_i915_gem_object *obj;
1203 	struct i915_vma *vma;
1204 	int size;
1205 	int err;
1206 
1207 	size = gen7_ctx_switch_bb_setup(engine, NULL /* probe size */);
1208 	if (size <= 0)
1209 		return size;
1210 
1211 	size = ALIGN(size, PAGE_SIZE);
1212 	obj = i915_gem_object_create_internal(engine->i915, size);
1213 	if (IS_ERR(obj))
1214 		return PTR_ERR(obj);
1215 
1216 	vma = i915_vma_instance(obj, engine->gt->vm, NULL);
1217 	if (IS_ERR(vma)) {
1218 		err = PTR_ERR(vma);
1219 		goto err_obj;
1220 	}
1221 
1222 	vma->private = intel_context_create(engine); /* dummy residuals */
1223 	if (IS_ERR(vma->private)) {
1224 		err = PTR_ERR(vma->private);
1225 		goto err_obj;
1226 	}
1227 
1228 	err = i915_vma_pin(vma, 0, 0, PIN_USER | PIN_HIGH);
1229 	if (err)
1230 		goto err_private;
1231 
1232 	err = i915_vma_sync(vma);
1233 	if (err)
1234 		goto err_unpin;
1235 
1236 	err = gen7_ctx_switch_bb_setup(engine, vma);
1237 	if (err)
1238 		goto err_unpin;
1239 
1240 	engine->wa_ctx.vma = vma;
1241 	return 0;
1242 
1243 err_unpin:
1244 	i915_vma_unpin(vma);
1245 err_private:
1246 	intel_context_put(vma->private);
1247 err_obj:
1248 	i915_gem_object_put(obj);
1249 	return err;
1250 }
1251 
1252 int intel_ring_submission_setup(struct intel_engine_cs *engine)
1253 {
1254 	struct intel_timeline *timeline;
1255 	struct intel_ring *ring;
1256 	int err;
1257 
1258 	setup_common(engine);
1259 
1260 	switch (engine->class) {
1261 	case RENDER_CLASS:
1262 		setup_rcs(engine);
1263 		break;
1264 	case VIDEO_DECODE_CLASS:
1265 		setup_vcs(engine);
1266 		break;
1267 	case COPY_ENGINE_CLASS:
1268 		setup_bcs(engine);
1269 		break;
1270 	case VIDEO_ENHANCEMENT_CLASS:
1271 		setup_vecs(engine);
1272 		break;
1273 	default:
1274 		MISSING_CASE(engine->class);
1275 		return -ENODEV;
1276 	}
1277 
1278 	timeline = intel_timeline_create_from_engine(engine,
1279 						     I915_GEM_HWS_SEQNO_ADDR);
1280 	if (IS_ERR(timeline)) {
1281 		err = PTR_ERR(timeline);
1282 		goto err;
1283 	}
1284 	GEM_BUG_ON(timeline->has_initial_breadcrumb);
1285 
1286 	err = intel_timeline_pin(timeline, NULL);
1287 	if (err)
1288 		goto err_timeline;
1289 
1290 	ring = intel_engine_create_ring(engine, SZ_16K);
1291 	if (IS_ERR(ring)) {
1292 		err = PTR_ERR(ring);
1293 		goto err_timeline_unpin;
1294 	}
1295 
1296 	err = intel_ring_pin(ring, NULL);
1297 	if (err)
1298 		goto err_ring;
1299 
1300 	GEM_BUG_ON(engine->legacy.ring);
1301 	engine->legacy.ring = ring;
1302 	engine->legacy.timeline = timeline;
1303 
1304 	GEM_BUG_ON(timeline->hwsp_ggtt != engine->status_page.vma);
1305 
1306 	if (IS_GEN(engine->i915, 7) && engine->class == RENDER_CLASS) {
1307 		err = gen7_ctx_switch_bb_init(engine);
1308 		if (err)
1309 			goto err_ring_unpin;
1310 	}
1311 
1312 	/* Finally, take ownership and responsibility for cleanup! */
1313 	engine->release = ring_release;
1314 
1315 	return 0;
1316 
1317 err_ring_unpin:
1318 	intel_ring_unpin(ring);
1319 err_ring:
1320 	intel_ring_put(ring);
1321 err_timeline_unpin:
1322 	intel_timeline_unpin(timeline);
1323 err_timeline:
1324 	intel_timeline_put(timeline);
1325 err:
1326 	intel_engine_cleanup_common(engine);
1327 	return err;
1328 }
1329 
1330 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1331 #include "selftest_ring_submission.c"
1332 #endif
1333