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 "intel_breadcrumbs.h"
36 #include "intel_context.h"
37 #include "intel_gt.h"
38 #include "intel_reset.h"
39 #include "intel_ring.h"
40 #include "shmem_utils.h"
41 
42 /* Rough estimate of the typical request size, performing a flush,
43  * set-context and then emitting the batch.
44  */
45 #define LEGACY_REQUEST_SIZE 200
46 
47 static void set_hwstam(struct intel_engine_cs *engine, u32 mask)
48 {
49 	/*
50 	 * Keep the render interrupt unmasked as this papers over
51 	 * lost interrupts following a reset.
52 	 */
53 	if (engine->class == RENDER_CLASS) {
54 		if (INTEL_GEN(engine->i915) >= 6)
55 			mask &= ~BIT(0);
56 		else
57 			mask &= ~I915_USER_INTERRUPT;
58 	}
59 
60 	intel_engine_set_hwsp_writemask(engine, mask);
61 }
62 
63 static void set_hws_pga(struct intel_engine_cs *engine, phys_addr_t phys)
64 {
65 	u32 addr;
66 
67 	addr = lower_32_bits(phys);
68 	if (INTEL_GEN(engine->i915) >= 4)
69 		addr |= (phys >> 28) & 0xf0;
70 
71 	intel_uncore_write(engine->uncore, HWS_PGA, addr);
72 }
73 
74 static struct page *status_page(struct intel_engine_cs *engine)
75 {
76 	struct drm_i915_gem_object *obj = engine->status_page.vma->obj;
77 
78 	GEM_BUG_ON(!i915_gem_object_has_pinned_pages(obj));
79 	return sg_page(obj->mm.pages->sgl);
80 }
81 
82 static void ring_setup_phys_status_page(struct intel_engine_cs *engine)
83 {
84 	set_hws_pga(engine, PFN_PHYS(page_to_pfn(status_page(engine))));
85 	set_hwstam(engine, ~0u);
86 }
87 
88 static void set_hwsp(struct intel_engine_cs *engine, u32 offset)
89 {
90 	i915_reg_t hwsp;
91 
92 	/*
93 	 * The ring status page addresses are no longer next to the rest of
94 	 * the ring registers as of gen7.
95 	 */
96 	if (IS_GEN(engine->i915, 7)) {
97 		switch (engine->id) {
98 		/*
99 		 * No more rings exist on Gen7. Default case is only to shut up
100 		 * gcc switch check warning.
101 		 */
102 		default:
103 			GEM_BUG_ON(engine->id);
104 			fallthrough;
105 		case RCS0:
106 			hwsp = RENDER_HWS_PGA_GEN7;
107 			break;
108 		case BCS0:
109 			hwsp = BLT_HWS_PGA_GEN7;
110 			break;
111 		case VCS0:
112 			hwsp = BSD_HWS_PGA_GEN7;
113 			break;
114 		case VECS0:
115 			hwsp = VEBOX_HWS_PGA_GEN7;
116 			break;
117 		}
118 	} else if (IS_GEN(engine->i915, 6)) {
119 		hwsp = RING_HWS_PGA_GEN6(engine->mmio_base);
120 	} else {
121 		hwsp = RING_HWS_PGA(engine->mmio_base);
122 	}
123 
124 	intel_uncore_write(engine->uncore, hwsp, offset);
125 	intel_uncore_posting_read(engine->uncore, hwsp);
126 }
127 
128 static void flush_cs_tlb(struct intel_engine_cs *engine)
129 {
130 	struct drm_i915_private *dev_priv = engine->i915;
131 
132 	if (!IS_GEN_RANGE(dev_priv, 6, 7))
133 		return;
134 
135 	/* ring should be idle before issuing a sync flush*/
136 	drm_WARN_ON(&dev_priv->drm,
137 		    (ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
138 
139 	ENGINE_WRITE(engine, RING_INSTPM,
140 		     _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE |
141 					INSTPM_SYNC_FLUSH));
142 	if (intel_wait_for_register(engine->uncore,
143 				    RING_INSTPM(engine->mmio_base),
144 				    INSTPM_SYNC_FLUSH, 0,
145 				    1000))
146 		drm_err(&dev_priv->drm,
147 			"%s: wait for SyncFlush to complete for TLB invalidation timed out\n",
148 			engine->name);
149 }
150 
151 static void ring_setup_status_page(struct intel_engine_cs *engine)
152 {
153 	set_hwsp(engine, i915_ggtt_offset(engine->status_page.vma));
154 	set_hwstam(engine, ~0u);
155 
156 	flush_cs_tlb(engine);
157 }
158 
159 static bool stop_ring(struct intel_engine_cs *engine)
160 {
161 	struct drm_i915_private *dev_priv = engine->i915;
162 
163 	if (INTEL_GEN(dev_priv) > 2) {
164 		ENGINE_WRITE(engine,
165 			     RING_MI_MODE, _MASKED_BIT_ENABLE(STOP_RING));
166 		if (intel_wait_for_register(engine->uncore,
167 					    RING_MI_MODE(engine->mmio_base),
168 					    MODE_IDLE,
169 					    MODE_IDLE,
170 					    1000)) {
171 			drm_err(&dev_priv->drm,
172 				"%s : timed out trying to stop ring\n",
173 				engine->name);
174 
175 			/*
176 			 * Sometimes we observe that the idle flag is not
177 			 * set even though the ring is empty. So double
178 			 * check before giving up.
179 			 */
180 			if (ENGINE_READ(engine, RING_HEAD) !=
181 			    ENGINE_READ(engine, RING_TAIL))
182 				return false;
183 		}
184 	}
185 
186 	ENGINE_WRITE(engine, RING_HEAD, ENGINE_READ(engine, RING_TAIL));
187 
188 	ENGINE_WRITE(engine, RING_HEAD, 0);
189 	ENGINE_WRITE(engine, RING_TAIL, 0);
190 
191 	/* The ring must be empty before it is disabled */
192 	ENGINE_WRITE(engine, RING_CTL, 0);
193 
194 	return (ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) == 0;
195 }
196 
197 static struct i915_address_space *vm_alias(struct i915_address_space *vm)
198 {
199 	if (i915_is_ggtt(vm))
200 		vm = &i915_vm_to_ggtt(vm)->alias->vm;
201 
202 	return vm;
203 }
204 
205 static u32 pp_dir(struct i915_address_space *vm)
206 {
207 	return to_gen6_ppgtt(i915_vm_to_ppgtt(vm))->pp_dir;
208 }
209 
210 static void set_pp_dir(struct intel_engine_cs *engine)
211 {
212 	struct i915_address_space *vm = vm_alias(engine->gt->vm);
213 
214 	if (vm) {
215 		ENGINE_WRITE(engine, RING_PP_DIR_DCLV, PP_DIR_DCLV_2G);
216 		ENGINE_WRITE(engine, RING_PP_DIR_BASE, pp_dir(vm));
217 	}
218 }
219 
220 static int xcs_resume(struct intel_engine_cs *engine)
221 {
222 	struct drm_i915_private *dev_priv = engine->i915;
223 	struct intel_ring *ring = engine->legacy.ring;
224 	int ret = 0;
225 
226 	ENGINE_TRACE(engine, "ring:{HEAD:%04x, TAIL:%04x}\n",
227 		     ring->head, ring->tail);
228 
229 	intel_uncore_forcewake_get(engine->uncore, FORCEWAKE_ALL);
230 
231 	/* WaClearRingBufHeadRegAtInit:ctg,elk */
232 	if (!stop_ring(engine)) {
233 		/* G45 ring initialization often fails to reset head to zero */
234 		drm_dbg(&dev_priv->drm, "%s head not reset to zero "
235 			"ctl %08x head %08x tail %08x start %08x\n",
236 			engine->name,
237 			ENGINE_READ(engine, RING_CTL),
238 			ENGINE_READ(engine, RING_HEAD),
239 			ENGINE_READ(engine, RING_TAIL),
240 			ENGINE_READ(engine, RING_START));
241 
242 		if (!stop_ring(engine)) {
243 			drm_err(&dev_priv->drm,
244 				"failed to set %s head to zero "
245 				"ctl %08x head %08x tail %08x start %08x\n",
246 				engine->name,
247 				ENGINE_READ(engine, RING_CTL),
248 				ENGINE_READ(engine, RING_HEAD),
249 				ENGINE_READ(engine, RING_TAIL),
250 				ENGINE_READ(engine, RING_START));
251 			ret = -EIO;
252 			goto out;
253 		}
254 	}
255 
256 	if (HWS_NEEDS_PHYSICAL(dev_priv))
257 		ring_setup_phys_status_page(engine);
258 	else
259 		ring_setup_status_page(engine);
260 
261 	intel_breadcrumbs_reset(engine->breadcrumbs);
262 
263 	/* Enforce ordering by reading HEAD register back */
264 	ENGINE_POSTING_READ(engine, RING_HEAD);
265 
266 	/*
267 	 * Initialize the ring. This must happen _after_ we've cleared the ring
268 	 * registers with the above sequence (the readback of the HEAD registers
269 	 * also enforces ordering), otherwise the hw might lose the new ring
270 	 * register values.
271 	 */
272 	ENGINE_WRITE(engine, RING_START, i915_ggtt_offset(ring->vma));
273 
274 	/* Check that the ring offsets point within the ring! */
275 	GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->head));
276 	GEM_BUG_ON(!intel_ring_offset_valid(ring, ring->tail));
277 	intel_ring_update_space(ring);
278 
279 	set_pp_dir(engine);
280 
281 	/* First wake the ring up to an empty/idle ring */
282 	ENGINE_WRITE(engine, RING_HEAD, ring->head);
283 	ENGINE_WRITE(engine, RING_TAIL, ring->head);
284 	ENGINE_POSTING_READ(engine, RING_TAIL);
285 
286 	ENGINE_WRITE(engine, RING_CTL, RING_CTL_SIZE(ring->size) | RING_VALID);
287 
288 	/* If the head is still not zero, the ring is dead */
289 	if (intel_wait_for_register(engine->uncore,
290 				    RING_CTL(engine->mmio_base),
291 				    RING_VALID, RING_VALID,
292 				    50)) {
293 		drm_err(&dev_priv->drm, "%s initialization failed "
294 			  "ctl %08x (valid? %d) head %08x [%08x] tail %08x [%08x] start %08x [expected %08x]\n",
295 			  engine->name,
296 			  ENGINE_READ(engine, RING_CTL),
297 			  ENGINE_READ(engine, RING_CTL) & RING_VALID,
298 			  ENGINE_READ(engine, RING_HEAD), ring->head,
299 			  ENGINE_READ(engine, RING_TAIL), ring->tail,
300 			  ENGINE_READ(engine, RING_START),
301 			  i915_ggtt_offset(ring->vma));
302 		ret = -EIO;
303 		goto out;
304 	}
305 
306 	if (INTEL_GEN(dev_priv) > 2)
307 		ENGINE_WRITE(engine,
308 			     RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
309 
310 	/* Now awake, let it get started */
311 	if (ring->tail != ring->head) {
312 		ENGINE_WRITE(engine, RING_TAIL, ring->tail);
313 		ENGINE_POSTING_READ(engine, RING_TAIL);
314 	}
315 
316 	/* Papering over lost _interrupts_ immediately following the restart */
317 	intel_engine_signal_breadcrumbs(engine);
318 out:
319 	intel_uncore_forcewake_put(engine->uncore, FORCEWAKE_ALL);
320 
321 	return ret;
322 }
323 
324 static void reset_prepare(struct intel_engine_cs *engine)
325 {
326 	struct intel_uncore *uncore = engine->uncore;
327 	const u32 base = engine->mmio_base;
328 
329 	/*
330 	 * We stop engines, otherwise we might get failed reset and a
331 	 * dead gpu (on elk). Also as modern gpu as kbl can suffer
332 	 * from system hang if batchbuffer is progressing when
333 	 * the reset is issued, regardless of READY_TO_RESET ack.
334 	 * Thus assume it is best to stop engines on all gens
335 	 * where we have a gpu reset.
336 	 *
337 	 * WaKBLVECSSemaphoreWaitPoll:kbl (on ALL_ENGINES)
338 	 *
339 	 * WaMediaResetMainRingCleanup:ctg,elk (presumably)
340 	 *
341 	 * FIXME: Wa for more modern gens needs to be validated
342 	 */
343 	ENGINE_TRACE(engine, "\n");
344 
345 	if (intel_engine_stop_cs(engine))
346 		ENGINE_TRACE(engine, "timed out on STOP_RING\n");
347 
348 	intel_uncore_write_fw(uncore,
349 			      RING_HEAD(base),
350 			      intel_uncore_read_fw(uncore, RING_TAIL(base)));
351 	intel_uncore_posting_read_fw(uncore, RING_HEAD(base)); /* paranoia */
352 
353 	intel_uncore_write_fw(uncore, RING_HEAD(base), 0);
354 	intel_uncore_write_fw(uncore, RING_TAIL(base), 0);
355 	intel_uncore_posting_read_fw(uncore, RING_TAIL(base));
356 
357 	/* The ring must be empty before it is disabled */
358 	intel_uncore_write_fw(uncore, RING_CTL(base), 0);
359 
360 	/* Check acts as a post */
361 	if (intel_uncore_read_fw(uncore, RING_HEAD(base)))
362 		ENGINE_TRACE(engine, "ring head [%x] not parked\n",
363 			     intel_uncore_read_fw(uncore, RING_HEAD(base)));
364 }
365 
366 static void reset_rewind(struct intel_engine_cs *engine, bool stalled)
367 {
368 	struct i915_request *pos, *rq;
369 	unsigned long flags;
370 	u32 head;
371 
372 	rq = NULL;
373 	spin_lock_irqsave(&engine->active.lock, flags);
374 	list_for_each_entry(pos, &engine->active.requests, sched.link) {
375 		if (!i915_request_completed(pos)) {
376 			rq = pos;
377 			break;
378 		}
379 	}
380 
381 	/*
382 	 * The guilty request will get skipped on a hung engine.
383 	 *
384 	 * Users of client default contexts do not rely on logical
385 	 * state preserved between batches so it is safe to execute
386 	 * queued requests following the hang. Non default contexts
387 	 * rely on preserved state, so skipping a batch loses the
388 	 * evolution of the state and it needs to be considered corrupted.
389 	 * Executing more queued batches on top of corrupted state is
390 	 * risky. But we take the risk by trying to advance through
391 	 * the queued requests in order to make the client behaviour
392 	 * more predictable around resets, by not throwing away random
393 	 * amount of batches it has prepared for execution. Sophisticated
394 	 * clients can use gem_reset_stats_ioctl and dma fence status
395 	 * (exported via sync_file info ioctl on explicit fences) to observe
396 	 * when it loses the context state and should rebuild accordingly.
397 	 *
398 	 * The context ban, and ultimately the client ban, mechanism are safety
399 	 * valves if client submission ends up resulting in nothing more than
400 	 * subsequent hangs.
401 	 */
402 
403 	if (rq) {
404 		/*
405 		 * Try to restore the logical GPU state to match the
406 		 * continuation of the request queue. If we skip the
407 		 * context/PD restore, then the next request may try to execute
408 		 * assuming that its context is valid and loaded on the GPU and
409 		 * so may try to access invalid memory, prompting repeated GPU
410 		 * hangs.
411 		 *
412 		 * If the request was guilty, we still restore the logical
413 		 * state in case the next request requires it (e.g. the
414 		 * aliasing ppgtt), but skip over the hung batch.
415 		 *
416 		 * If the request was innocent, we try to replay the request
417 		 * with the restored context.
418 		 */
419 		__i915_request_reset(rq, stalled);
420 
421 		GEM_BUG_ON(rq->ring != engine->legacy.ring);
422 		head = rq->head;
423 	} else {
424 		head = engine->legacy.ring->tail;
425 	}
426 	engine->legacy.ring->head = intel_ring_wrap(engine->legacy.ring, head);
427 
428 	spin_unlock_irqrestore(&engine->active.lock, flags);
429 }
430 
431 static void reset_finish(struct intel_engine_cs *engine)
432 {
433 }
434 
435 static void reset_cancel(struct intel_engine_cs *engine)
436 {
437 	struct i915_request *request;
438 	unsigned long flags;
439 
440 	spin_lock_irqsave(&engine->active.lock, flags);
441 
442 	/* Mark all submitted requests as skipped. */
443 	list_for_each_entry(request, &engine->active.requests, sched.link) {
444 		i915_request_set_error_once(request, -EIO);
445 		i915_request_mark_complete(request);
446 	}
447 	intel_engine_signal_breadcrumbs(engine);
448 
449 	/* Remaining _unready_ requests will be nop'ed when submitted */
450 
451 	spin_unlock_irqrestore(&engine->active.lock, flags);
452 }
453 
454 static void i9xx_submit_request(struct i915_request *request)
455 {
456 	i915_request_submit(request);
457 	wmb(); /* paranoid flush writes out of the WCB before mmio */
458 
459 	ENGINE_WRITE(request->engine, RING_TAIL,
460 		     intel_ring_set_tail(request->ring, request->tail));
461 }
462 
463 static void __ring_context_fini(struct intel_context *ce)
464 {
465 	i915_vma_put(ce->state);
466 }
467 
468 static void ring_context_destroy(struct kref *ref)
469 {
470 	struct intel_context *ce = container_of(ref, typeof(*ce), ref);
471 
472 	GEM_BUG_ON(intel_context_is_pinned(ce));
473 
474 	if (ce->state)
475 		__ring_context_fini(ce);
476 
477 	intel_context_fini(ce);
478 	intel_context_free(ce);
479 }
480 
481 static int ring_context_pre_pin(struct intel_context *ce,
482 				struct i915_gem_ww_ctx *ww,
483 				void **unused)
484 {
485 	struct i915_address_space *vm;
486 	int err = 0;
487 
488 	vm = vm_alias(ce->vm);
489 	if (vm)
490 		err = gen6_ppgtt_pin(i915_vm_to_ppgtt((vm)), ww);
491 
492 	return err;
493 }
494 
495 static void __context_unpin_ppgtt(struct intel_context *ce)
496 {
497 	struct i915_address_space *vm;
498 
499 	vm = vm_alias(ce->vm);
500 	if (vm)
501 		gen6_ppgtt_unpin(i915_vm_to_ppgtt(vm));
502 }
503 
504 static void ring_context_unpin(struct intel_context *ce)
505 {
506 }
507 
508 static void ring_context_post_unpin(struct intel_context *ce)
509 {
510 	__context_unpin_ppgtt(ce);
511 }
512 
513 static struct i915_vma *
514 alloc_context_vma(struct intel_engine_cs *engine)
515 {
516 	struct drm_i915_private *i915 = engine->i915;
517 	struct drm_i915_gem_object *obj;
518 	struct i915_vma *vma;
519 	int err;
520 
521 	obj = i915_gem_object_create_shmem(i915, engine->context_size);
522 	if (IS_ERR(obj))
523 		return ERR_CAST(obj);
524 
525 	/*
526 	 * Try to make the context utilize L3 as well as LLC.
527 	 *
528 	 * On VLV we don't have L3 controls in the PTEs so we
529 	 * shouldn't touch the cache level, especially as that
530 	 * would make the object snooped which might have a
531 	 * negative performance impact.
532 	 *
533 	 * Snooping is required on non-llc platforms in execlist
534 	 * mode, but since all GGTT accesses use PAT entry 0 we
535 	 * get snooping anyway regardless of cache_level.
536 	 *
537 	 * This is only applicable for Ivy Bridge devices since
538 	 * later platforms don't have L3 control bits in the PTE.
539 	 */
540 	if (IS_IVYBRIDGE(i915))
541 		i915_gem_object_set_cache_coherency(obj, I915_CACHE_L3_LLC);
542 
543 	if (engine->default_state) {
544 		void *vaddr;
545 
546 		vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
547 		if (IS_ERR(vaddr)) {
548 			err = PTR_ERR(vaddr);
549 			goto err_obj;
550 		}
551 
552 		shmem_read(engine->default_state, 0,
553 			   vaddr, engine->context_size);
554 
555 		i915_gem_object_flush_map(obj);
556 		__i915_gem_object_release_map(obj);
557 	}
558 
559 	vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
560 	if (IS_ERR(vma)) {
561 		err = PTR_ERR(vma);
562 		goto err_obj;
563 	}
564 
565 	return vma;
566 
567 err_obj:
568 	i915_gem_object_put(obj);
569 	return ERR_PTR(err);
570 }
571 
572 static int ring_context_alloc(struct intel_context *ce)
573 {
574 	struct intel_engine_cs *engine = ce->engine;
575 
576 	/* One ringbuffer to rule them all */
577 	GEM_BUG_ON(!engine->legacy.ring);
578 	ce->ring = engine->legacy.ring;
579 	ce->timeline = intel_timeline_get(engine->legacy.timeline);
580 
581 	GEM_BUG_ON(ce->state);
582 	if (engine->context_size) {
583 		struct i915_vma *vma;
584 
585 		vma = alloc_context_vma(engine);
586 		if (IS_ERR(vma))
587 			return PTR_ERR(vma);
588 
589 		ce->state = vma;
590 		if (engine->default_state)
591 			__set_bit(CONTEXT_VALID_BIT, &ce->flags);
592 	}
593 
594 	return 0;
595 }
596 
597 static int ring_context_pin(struct intel_context *ce, void *unused)
598 {
599 	return 0;
600 }
601 
602 static void ring_context_reset(struct intel_context *ce)
603 {
604 	intel_ring_reset(ce->ring, ce->ring->emit);
605 }
606 
607 static const struct intel_context_ops ring_context_ops = {
608 	.alloc = ring_context_alloc,
609 
610 	.pre_pin = ring_context_pre_pin,
611 	.pin = ring_context_pin,
612 	.unpin = ring_context_unpin,
613 	.post_unpin = ring_context_post_unpin,
614 
615 	.enter = intel_context_enter_engine,
616 	.exit = intel_context_exit_engine,
617 
618 	.reset = ring_context_reset,
619 	.destroy = ring_context_destroy,
620 };
621 
622 static int load_pd_dir(struct i915_request *rq,
623 		       struct i915_address_space *vm,
624 		       u32 valid)
625 {
626 	const struct intel_engine_cs * const engine = rq->engine;
627 	u32 *cs;
628 
629 	cs = intel_ring_begin(rq, 12);
630 	if (IS_ERR(cs))
631 		return PTR_ERR(cs);
632 
633 	*cs++ = MI_LOAD_REGISTER_IMM(1);
634 	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_DCLV(engine->mmio_base));
635 	*cs++ = valid;
636 
637 	*cs++ = MI_LOAD_REGISTER_IMM(1);
638 	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
639 	*cs++ = pp_dir(vm);
640 
641 	/* Stall until the page table load is complete? */
642 	*cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
643 	*cs++ = i915_mmio_reg_offset(RING_PP_DIR_BASE(engine->mmio_base));
644 	*cs++ = intel_gt_scratch_offset(engine->gt,
645 					INTEL_GT_SCRATCH_FIELD_DEFAULT);
646 
647 	*cs++ = MI_LOAD_REGISTER_IMM(1);
648 	*cs++ = i915_mmio_reg_offset(RING_INSTPM(engine->mmio_base));
649 	*cs++ = _MASKED_BIT_ENABLE(INSTPM_TLB_INVALIDATE);
650 
651 	intel_ring_advance(rq, cs);
652 
653 	return rq->engine->emit_flush(rq, EMIT_FLUSH);
654 }
655 
656 static inline int mi_set_context(struct i915_request *rq,
657 				 struct intel_context *ce,
658 				 u32 flags)
659 {
660 	struct intel_engine_cs *engine = rq->engine;
661 	struct drm_i915_private *i915 = engine->i915;
662 	enum intel_engine_id id;
663 	const int num_engines =
664 		IS_HASWELL(i915) ? engine->gt->info.num_engines - 1 : 0;
665 	bool force_restore = false;
666 	int len;
667 	u32 *cs;
668 
669 	len = 4;
670 	if (IS_GEN(i915, 7))
671 		len += 2 + (num_engines ? 4 * num_engines + 6 : 0);
672 	else if (IS_GEN(i915, 5))
673 		len += 2;
674 	if (flags & MI_FORCE_RESTORE) {
675 		GEM_BUG_ON(flags & MI_RESTORE_INHIBIT);
676 		flags &= ~MI_FORCE_RESTORE;
677 		force_restore = true;
678 		len += 2;
679 	}
680 
681 	cs = intel_ring_begin(rq, len);
682 	if (IS_ERR(cs))
683 		return PTR_ERR(cs);
684 
685 	/* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
686 	if (IS_GEN(i915, 7)) {
687 		*cs++ = MI_ARB_ON_OFF | MI_ARB_DISABLE;
688 		if (num_engines) {
689 			struct intel_engine_cs *signaller;
690 
691 			*cs++ = MI_LOAD_REGISTER_IMM(num_engines);
692 			for_each_engine(signaller, engine->gt, id) {
693 				if (signaller == engine)
694 					continue;
695 
696 				*cs++ = i915_mmio_reg_offset(
697 					   RING_PSMI_CTL(signaller->mmio_base));
698 				*cs++ = _MASKED_BIT_ENABLE(
699 						GEN6_PSMI_SLEEP_MSG_DISABLE);
700 			}
701 		}
702 	} else if (IS_GEN(i915, 5)) {
703 		/*
704 		 * This w/a is only listed for pre-production ilk a/b steppings,
705 		 * but is also mentioned for programming the powerctx. To be
706 		 * safe, just apply the workaround; we do not use SyncFlush so
707 		 * this should never take effect and so be a no-op!
708 		 */
709 		*cs++ = MI_SUSPEND_FLUSH | MI_SUSPEND_FLUSH_EN;
710 	}
711 
712 	if (force_restore) {
713 		/*
714 		 * The HW doesn't handle being told to restore the current
715 		 * context very well. Quite often it likes goes to go off and
716 		 * sulk, especially when it is meant to be reloading PP_DIR.
717 		 * A very simple fix to force the reload is to simply switch
718 		 * away from the current context and back again.
719 		 *
720 		 * Note that the kernel_context will contain random state
721 		 * following the INHIBIT_RESTORE. We accept this since we
722 		 * never use the kernel_context state; it is merely a
723 		 * placeholder we use to flush other contexts.
724 		 */
725 		*cs++ = MI_SET_CONTEXT;
726 		*cs++ = i915_ggtt_offset(engine->kernel_context->state) |
727 			MI_MM_SPACE_GTT |
728 			MI_RESTORE_INHIBIT;
729 	}
730 
731 	*cs++ = MI_NOOP;
732 	*cs++ = MI_SET_CONTEXT;
733 	*cs++ = i915_ggtt_offset(ce->state) | flags;
734 	/*
735 	 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
736 	 * WaMiSetContext_Hang:snb,ivb,vlv
737 	 */
738 	*cs++ = MI_NOOP;
739 
740 	if (IS_GEN(i915, 7)) {
741 		if (num_engines) {
742 			struct intel_engine_cs *signaller;
743 			i915_reg_t last_reg = {}; /* keep gcc quiet */
744 
745 			*cs++ = MI_LOAD_REGISTER_IMM(num_engines);
746 			for_each_engine(signaller, engine->gt, id) {
747 				if (signaller == engine)
748 					continue;
749 
750 				last_reg = RING_PSMI_CTL(signaller->mmio_base);
751 				*cs++ = i915_mmio_reg_offset(last_reg);
752 				*cs++ = _MASKED_BIT_DISABLE(
753 						GEN6_PSMI_SLEEP_MSG_DISABLE);
754 			}
755 
756 			/* Insert a delay before the next switch! */
757 			*cs++ = MI_STORE_REGISTER_MEM | MI_SRM_LRM_GLOBAL_GTT;
758 			*cs++ = i915_mmio_reg_offset(last_reg);
759 			*cs++ = intel_gt_scratch_offset(engine->gt,
760 							INTEL_GT_SCRATCH_FIELD_DEFAULT);
761 			*cs++ = MI_NOOP;
762 		}
763 		*cs++ = MI_ARB_ON_OFF | MI_ARB_ENABLE;
764 	} else if (IS_GEN(i915, 5)) {
765 		*cs++ = MI_SUSPEND_FLUSH;
766 	}
767 
768 	intel_ring_advance(rq, cs);
769 
770 	return 0;
771 }
772 
773 static int remap_l3_slice(struct i915_request *rq, int slice)
774 {
775 	u32 *cs, *remap_info = rq->engine->i915->l3_parity.remap_info[slice];
776 	int i;
777 
778 	if (!remap_info)
779 		return 0;
780 
781 	cs = intel_ring_begin(rq, GEN7_L3LOG_SIZE/4 * 2 + 2);
782 	if (IS_ERR(cs))
783 		return PTR_ERR(cs);
784 
785 	/*
786 	 * Note: We do not worry about the concurrent register cacheline hang
787 	 * here because no other code should access these registers other than
788 	 * at initialization time.
789 	 */
790 	*cs++ = MI_LOAD_REGISTER_IMM(GEN7_L3LOG_SIZE/4);
791 	for (i = 0; i < GEN7_L3LOG_SIZE/4; i++) {
792 		*cs++ = i915_mmio_reg_offset(GEN7_L3LOG(slice, i));
793 		*cs++ = remap_info[i];
794 	}
795 	*cs++ = MI_NOOP;
796 	intel_ring_advance(rq, cs);
797 
798 	return 0;
799 }
800 
801 static int remap_l3(struct i915_request *rq)
802 {
803 	struct i915_gem_context *ctx = i915_request_gem_context(rq);
804 	int i, err;
805 
806 	if (!ctx || !ctx->remap_slice)
807 		return 0;
808 
809 	for (i = 0; i < MAX_L3_SLICES; i++) {
810 		if (!(ctx->remap_slice & BIT(i)))
811 			continue;
812 
813 		err = remap_l3_slice(rq, i);
814 		if (err)
815 			return err;
816 	}
817 
818 	ctx->remap_slice = 0;
819 	return 0;
820 }
821 
822 static int switch_mm(struct i915_request *rq, struct i915_address_space *vm)
823 {
824 	int ret;
825 
826 	if (!vm)
827 		return 0;
828 
829 	ret = rq->engine->emit_flush(rq, EMIT_FLUSH);
830 	if (ret)
831 		return ret;
832 
833 	/*
834 	 * Not only do we need a full barrier (post-sync write) after
835 	 * invalidating the TLBs, but we need to wait a little bit
836 	 * longer. Whether this is merely delaying us, or the
837 	 * subsequent flush is a key part of serialising with the
838 	 * post-sync op, this extra pass appears vital before a
839 	 * mm switch!
840 	 */
841 	ret = load_pd_dir(rq, vm, PP_DIR_DCLV_2G);
842 	if (ret)
843 		return ret;
844 
845 	return rq->engine->emit_flush(rq, EMIT_INVALIDATE);
846 }
847 
848 static int clear_residuals(struct i915_request *rq)
849 {
850 	struct intel_engine_cs *engine = rq->engine;
851 	int ret;
852 
853 	ret = switch_mm(rq, vm_alias(engine->kernel_context->vm));
854 	if (ret)
855 		return ret;
856 
857 	if (engine->kernel_context->state) {
858 		ret = mi_set_context(rq,
859 				     engine->kernel_context,
860 				     MI_MM_SPACE_GTT | MI_RESTORE_INHIBIT);
861 		if (ret)
862 			return ret;
863 	}
864 
865 	ret = engine->emit_bb_start(rq,
866 				    engine->wa_ctx.vma->node.start, 0,
867 				    0);
868 	if (ret)
869 		return ret;
870 
871 	ret = engine->emit_flush(rq, EMIT_FLUSH);
872 	if (ret)
873 		return ret;
874 
875 	/* Always invalidate before the next switch_mm() */
876 	return engine->emit_flush(rq, EMIT_INVALIDATE);
877 }
878 
879 static int switch_context(struct i915_request *rq)
880 {
881 	struct intel_engine_cs *engine = rq->engine;
882 	struct intel_context *ce = rq->context;
883 	void **residuals = NULL;
884 	int ret;
885 
886 	GEM_BUG_ON(HAS_EXECLISTS(engine->i915));
887 
888 	if (engine->wa_ctx.vma && ce != engine->kernel_context) {
889 		if (engine->wa_ctx.vma->private != ce) {
890 			ret = clear_residuals(rq);
891 			if (ret)
892 				return ret;
893 
894 			residuals = &engine->wa_ctx.vma->private;
895 		}
896 	}
897 
898 	ret = switch_mm(rq, vm_alias(ce->vm));
899 	if (ret)
900 		return ret;
901 
902 	if (ce->state) {
903 		u32 flags;
904 
905 		GEM_BUG_ON(engine->id != RCS0);
906 
907 		/* For resource streamer on HSW+ and power context elsewhere */
908 		BUILD_BUG_ON(HSW_MI_RS_SAVE_STATE_EN != MI_SAVE_EXT_STATE_EN);
909 		BUILD_BUG_ON(HSW_MI_RS_RESTORE_STATE_EN != MI_RESTORE_EXT_STATE_EN);
910 
911 		flags = MI_SAVE_EXT_STATE_EN | MI_MM_SPACE_GTT;
912 		if (test_bit(CONTEXT_VALID_BIT, &ce->flags))
913 			flags |= MI_RESTORE_EXT_STATE_EN;
914 		else
915 			flags |= MI_RESTORE_INHIBIT;
916 
917 		ret = mi_set_context(rq, ce, flags);
918 		if (ret)
919 			return ret;
920 	}
921 
922 	ret = remap_l3(rq);
923 	if (ret)
924 		return ret;
925 
926 	/*
927 	 * Now past the point of no return, this request _will_ be emitted.
928 	 *
929 	 * Or at least this preamble will be emitted, the request may be
930 	 * interrupted prior to submitting the user payload. If so, we
931 	 * still submit the "empty" request in order to preserve global
932 	 * state tracking such as this, our tracking of the current
933 	 * dirty context.
934 	 */
935 	if (residuals) {
936 		intel_context_put(*residuals);
937 		*residuals = intel_context_get(ce);
938 	}
939 
940 	return 0;
941 }
942 
943 static int ring_request_alloc(struct i915_request *request)
944 {
945 	int ret;
946 
947 	GEM_BUG_ON(!intel_context_is_pinned(request->context));
948 	GEM_BUG_ON(i915_request_timeline(request)->has_initial_breadcrumb);
949 
950 	/*
951 	 * Flush enough space to reduce the likelihood of waiting after
952 	 * we start building the request - in which case we will just
953 	 * have to repeat work.
954 	 */
955 	request->reserved_space += LEGACY_REQUEST_SIZE;
956 
957 	/* Unconditionally invalidate GPU caches and TLBs. */
958 	ret = request->engine->emit_flush(request, EMIT_INVALIDATE);
959 	if (ret)
960 		return ret;
961 
962 	ret = switch_context(request);
963 	if (ret)
964 		return ret;
965 
966 	request->reserved_space -= LEGACY_REQUEST_SIZE;
967 	return 0;
968 }
969 
970 static void gen6_bsd_submit_request(struct i915_request *request)
971 {
972 	struct intel_uncore *uncore = request->engine->uncore;
973 
974 	intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
975 
976        /* Every tail move must follow the sequence below */
977 
978 	/* Disable notification that the ring is IDLE. The GT
979 	 * will then assume that it is busy and bring it out of rc6.
980 	 */
981 	intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
982 			      _MASKED_BIT_ENABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
983 
984 	/* Clear the context id. Here be magic! */
985 	intel_uncore_write64_fw(uncore, GEN6_BSD_RNCID, 0x0);
986 
987 	/* Wait for the ring not to be idle, i.e. for it to wake up. */
988 	if (__intel_wait_for_register_fw(uncore,
989 					 GEN6_BSD_SLEEP_PSMI_CONTROL,
990 					 GEN6_BSD_SLEEP_INDICATOR,
991 					 0,
992 					 1000, 0, NULL))
993 		drm_err(&uncore->i915->drm,
994 			"timed out waiting for the BSD ring to wake up\n");
995 
996 	/* Now that the ring is fully powered up, update the tail */
997 	i9xx_submit_request(request);
998 
999 	/* Let the ring send IDLE messages to the GT again,
1000 	 * and so let it sleep to conserve power when idle.
1001 	 */
1002 	intel_uncore_write_fw(uncore, GEN6_BSD_SLEEP_PSMI_CONTROL,
1003 			      _MASKED_BIT_DISABLE(GEN6_BSD_SLEEP_MSG_DISABLE));
1004 
1005 	intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
1006 }
1007 
1008 static void i9xx_set_default_submission(struct intel_engine_cs *engine)
1009 {
1010 	engine->submit_request = i9xx_submit_request;
1011 
1012 	engine->park = NULL;
1013 	engine->unpark = NULL;
1014 }
1015 
1016 static void gen6_bsd_set_default_submission(struct intel_engine_cs *engine)
1017 {
1018 	i9xx_set_default_submission(engine);
1019 	engine->submit_request = gen6_bsd_submit_request;
1020 }
1021 
1022 static void ring_release(struct intel_engine_cs *engine)
1023 {
1024 	struct drm_i915_private *dev_priv = engine->i915;
1025 
1026 	drm_WARN_ON(&dev_priv->drm, INTEL_GEN(dev_priv) > 2 &&
1027 		    (ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE) == 0);
1028 
1029 	intel_engine_cleanup_common(engine);
1030 
1031 	if (engine->wa_ctx.vma) {
1032 		intel_context_put(engine->wa_ctx.vma->private);
1033 		i915_vma_unpin_and_release(&engine->wa_ctx.vma, 0);
1034 	}
1035 
1036 	intel_ring_unpin(engine->legacy.ring);
1037 	intel_ring_put(engine->legacy.ring);
1038 
1039 	intel_timeline_unpin(engine->legacy.timeline);
1040 	intel_timeline_put(engine->legacy.timeline);
1041 }
1042 
1043 static void setup_irq(struct intel_engine_cs *engine)
1044 {
1045 	struct drm_i915_private *i915 = engine->i915;
1046 
1047 	if (INTEL_GEN(i915) >= 6) {
1048 		engine->irq_enable = gen6_irq_enable;
1049 		engine->irq_disable = gen6_irq_disable;
1050 	} else if (INTEL_GEN(i915) >= 5) {
1051 		engine->irq_enable = gen5_irq_enable;
1052 		engine->irq_disable = gen5_irq_disable;
1053 	} else if (INTEL_GEN(i915) >= 3) {
1054 		engine->irq_enable = gen3_irq_enable;
1055 		engine->irq_disable = gen3_irq_disable;
1056 	} else {
1057 		engine->irq_enable = gen2_irq_enable;
1058 		engine->irq_disable = gen2_irq_disable;
1059 	}
1060 }
1061 
1062 static void setup_common(struct intel_engine_cs *engine)
1063 {
1064 	struct drm_i915_private *i915 = engine->i915;
1065 
1066 	/* gen8+ are only supported with execlists */
1067 	GEM_BUG_ON(INTEL_GEN(i915) >= 8);
1068 
1069 	setup_irq(engine);
1070 
1071 	engine->resume = xcs_resume;
1072 	engine->reset.prepare = reset_prepare;
1073 	engine->reset.rewind = reset_rewind;
1074 	engine->reset.cancel = reset_cancel;
1075 	engine->reset.finish = reset_finish;
1076 
1077 	engine->cops = &ring_context_ops;
1078 	engine->request_alloc = ring_request_alloc;
1079 
1080 	/*
1081 	 * Using a global execution timeline; the previous final breadcrumb is
1082 	 * equivalent to our next initial bread so we can elide
1083 	 * engine->emit_init_breadcrumb().
1084 	 */
1085 	engine->emit_fini_breadcrumb = gen3_emit_breadcrumb;
1086 	if (IS_GEN(i915, 5))
1087 		engine->emit_fini_breadcrumb = gen5_emit_breadcrumb;
1088 
1089 	engine->set_default_submission = i9xx_set_default_submission;
1090 
1091 	if (INTEL_GEN(i915) >= 6)
1092 		engine->emit_bb_start = gen6_emit_bb_start;
1093 	else if (INTEL_GEN(i915) >= 4)
1094 		engine->emit_bb_start = gen4_emit_bb_start;
1095 	else if (IS_I830(i915) || IS_I845G(i915))
1096 		engine->emit_bb_start = i830_emit_bb_start;
1097 	else
1098 		engine->emit_bb_start = gen3_emit_bb_start;
1099 }
1100 
1101 static void setup_rcs(struct intel_engine_cs *engine)
1102 {
1103 	struct drm_i915_private *i915 = engine->i915;
1104 
1105 	if (HAS_L3_DPF(i915))
1106 		engine->irq_keep_mask = GT_RENDER_L3_PARITY_ERROR_INTERRUPT;
1107 
1108 	engine->irq_enable_mask = GT_RENDER_USER_INTERRUPT;
1109 
1110 	if (INTEL_GEN(i915) >= 7) {
1111 		engine->emit_flush = gen7_emit_flush_rcs;
1112 		engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_rcs;
1113 	} else if (IS_GEN(i915, 6)) {
1114 		engine->emit_flush = gen6_emit_flush_rcs;
1115 		engine->emit_fini_breadcrumb = gen6_emit_breadcrumb_rcs;
1116 	} else if (IS_GEN(i915, 5)) {
1117 		engine->emit_flush = gen4_emit_flush_rcs;
1118 	} else {
1119 		if (INTEL_GEN(i915) < 4)
1120 			engine->emit_flush = gen2_emit_flush;
1121 		else
1122 			engine->emit_flush = gen4_emit_flush_rcs;
1123 		engine->irq_enable_mask = I915_USER_INTERRUPT;
1124 	}
1125 
1126 	if (IS_HASWELL(i915))
1127 		engine->emit_bb_start = hsw_emit_bb_start;
1128 }
1129 
1130 static void setup_vcs(struct intel_engine_cs *engine)
1131 {
1132 	struct drm_i915_private *i915 = engine->i915;
1133 
1134 	if (INTEL_GEN(i915) >= 6) {
1135 		/* gen6 bsd needs a special wa for tail updates */
1136 		if (IS_GEN(i915, 6))
1137 			engine->set_default_submission = gen6_bsd_set_default_submission;
1138 		engine->emit_flush = gen6_emit_flush_vcs;
1139 		engine->irq_enable_mask = GT_BSD_USER_INTERRUPT;
1140 
1141 		if (IS_GEN(i915, 6))
1142 			engine->emit_fini_breadcrumb = gen6_emit_breadcrumb_xcs;
1143 		else
1144 			engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_xcs;
1145 	} else {
1146 		engine->emit_flush = gen4_emit_flush_vcs;
1147 		if (IS_GEN(i915, 5))
1148 			engine->irq_enable_mask = ILK_BSD_USER_INTERRUPT;
1149 		else
1150 			engine->irq_enable_mask = I915_BSD_USER_INTERRUPT;
1151 	}
1152 }
1153 
1154 static void setup_bcs(struct intel_engine_cs *engine)
1155 {
1156 	struct drm_i915_private *i915 = engine->i915;
1157 
1158 	engine->emit_flush = gen6_emit_flush_xcs;
1159 	engine->irq_enable_mask = GT_BLT_USER_INTERRUPT;
1160 
1161 	if (IS_GEN(i915, 6))
1162 		engine->emit_fini_breadcrumb = gen6_emit_breadcrumb_xcs;
1163 	else
1164 		engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_xcs;
1165 }
1166 
1167 static void setup_vecs(struct intel_engine_cs *engine)
1168 {
1169 	struct drm_i915_private *i915 = engine->i915;
1170 
1171 	GEM_BUG_ON(INTEL_GEN(i915) < 7);
1172 
1173 	engine->emit_flush = gen6_emit_flush_xcs;
1174 	engine->irq_enable_mask = PM_VEBOX_USER_INTERRUPT;
1175 	engine->irq_enable = hsw_irq_enable_vecs;
1176 	engine->irq_disable = hsw_irq_disable_vecs;
1177 
1178 	engine->emit_fini_breadcrumb = gen7_emit_breadcrumb_xcs;
1179 }
1180 
1181 static int gen7_ctx_switch_bb_setup(struct intel_engine_cs * const engine,
1182 				    struct i915_vma * const vma)
1183 {
1184 	return gen7_setup_clear_gpr_bb(engine, vma);
1185 }
1186 
1187 static int gen7_ctx_switch_bb_init(struct intel_engine_cs *engine)
1188 {
1189 	struct drm_i915_gem_object *obj;
1190 	struct i915_vma *vma;
1191 	int size;
1192 	int err;
1193 
1194 	size = gen7_ctx_switch_bb_setup(engine, NULL /* probe size */);
1195 	if (size <= 0)
1196 		return size;
1197 
1198 	size = ALIGN(size, PAGE_SIZE);
1199 	obj = i915_gem_object_create_internal(engine->i915, size);
1200 	if (IS_ERR(obj))
1201 		return PTR_ERR(obj);
1202 
1203 	vma = i915_vma_instance(obj, engine->gt->vm, NULL);
1204 	if (IS_ERR(vma)) {
1205 		err = PTR_ERR(vma);
1206 		goto err_obj;
1207 	}
1208 
1209 	vma->private = intel_context_create(engine); /* dummy residuals */
1210 	if (IS_ERR(vma->private)) {
1211 		err = PTR_ERR(vma->private);
1212 		goto err_obj;
1213 	}
1214 
1215 	err = i915_vma_pin(vma, 0, 0, PIN_USER | PIN_HIGH);
1216 	if (err)
1217 		goto err_private;
1218 
1219 	err = i915_vma_sync(vma);
1220 	if (err)
1221 		goto err_unpin;
1222 
1223 	err = gen7_ctx_switch_bb_setup(engine, vma);
1224 	if (err)
1225 		goto err_unpin;
1226 
1227 	engine->wa_ctx.vma = vma;
1228 	return 0;
1229 
1230 err_unpin:
1231 	i915_vma_unpin(vma);
1232 err_private:
1233 	intel_context_put(vma->private);
1234 err_obj:
1235 	i915_gem_object_put(obj);
1236 	return err;
1237 }
1238 
1239 int intel_ring_submission_setup(struct intel_engine_cs *engine)
1240 {
1241 	struct intel_timeline *timeline;
1242 	struct intel_ring *ring;
1243 	int err;
1244 
1245 	setup_common(engine);
1246 
1247 	switch (engine->class) {
1248 	case RENDER_CLASS:
1249 		setup_rcs(engine);
1250 		break;
1251 	case VIDEO_DECODE_CLASS:
1252 		setup_vcs(engine);
1253 		break;
1254 	case COPY_ENGINE_CLASS:
1255 		setup_bcs(engine);
1256 		break;
1257 	case VIDEO_ENHANCEMENT_CLASS:
1258 		setup_vecs(engine);
1259 		break;
1260 	default:
1261 		MISSING_CASE(engine->class);
1262 		return -ENODEV;
1263 	}
1264 
1265 	timeline = intel_timeline_create_from_engine(engine,
1266 						     I915_GEM_HWS_SEQNO_ADDR);
1267 	if (IS_ERR(timeline)) {
1268 		err = PTR_ERR(timeline);
1269 		goto err;
1270 	}
1271 	GEM_BUG_ON(timeline->has_initial_breadcrumb);
1272 
1273 	err = intel_timeline_pin(timeline, NULL);
1274 	if (err)
1275 		goto err_timeline;
1276 
1277 	ring = intel_engine_create_ring(engine, SZ_16K);
1278 	if (IS_ERR(ring)) {
1279 		err = PTR_ERR(ring);
1280 		goto err_timeline_unpin;
1281 	}
1282 
1283 	err = intel_ring_pin(ring, NULL);
1284 	if (err)
1285 		goto err_ring;
1286 
1287 	GEM_BUG_ON(engine->legacy.ring);
1288 	engine->legacy.ring = ring;
1289 	engine->legacy.timeline = timeline;
1290 
1291 	GEM_BUG_ON(timeline->hwsp_ggtt != engine->status_page.vma);
1292 
1293 	if (IS_HASWELL(engine->i915) && engine->class == RENDER_CLASS) {
1294 		err = gen7_ctx_switch_bb_init(engine);
1295 		if (err)
1296 			goto err_ring_unpin;
1297 	}
1298 
1299 	/* Finally, take ownership and responsibility for cleanup! */
1300 	engine->release = ring_release;
1301 
1302 	return 0;
1303 
1304 err_ring_unpin:
1305 	intel_ring_unpin(ring);
1306 err_ring:
1307 	intel_ring_put(ring);
1308 err_timeline_unpin:
1309 	intel_timeline_unpin(timeline);
1310 err_timeline:
1311 	intel_timeline_put(timeline);
1312 err:
1313 	intel_engine_cleanup_common(engine);
1314 	return err;
1315 }
1316 
1317 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1318 #include "selftest_ring_submission.c"
1319 #endif
1320