1 /*
2  * Copyright © 2008-2015 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  */
24 
25 #include <linux/dma-fence-array.h>
26 #include <linux/dma-fence-chain.h>
27 #include <linux/irq_work.h>
28 #include <linux/prefetch.h>
29 #include <linux/sched.h>
30 #include <linux/sched/clock.h>
31 #include <linux/sched/signal.h>
32 #include <linux/sched/mm.h>
33 
34 #include "gem/i915_gem_context.h"
35 #include "gt/intel_breadcrumbs.h"
36 #include "gt/intel_context.h"
37 #include "gt/intel_engine.h"
38 #include "gt/intel_engine_heartbeat.h"
39 #include "gt/intel_engine_regs.h"
40 #include "gt/intel_gpu_commands.h"
41 #include "gt/intel_reset.h"
42 #include "gt/intel_ring.h"
43 #include "gt/intel_rps.h"
44 
45 #include "i915_active.h"
46 #include "i915_deps.h"
47 #include "i915_driver.h"
48 #include "i915_drv.h"
49 #include "i915_trace.h"
50 #include "intel_pm.h"
51 
52 struct execute_cb {
53 	struct irq_work work;
54 	struct i915_sw_fence *fence;
55 	struct i915_request *signal;
56 };
57 
58 static struct kmem_cache *slab_requests;
59 static struct kmem_cache *slab_execute_cbs;
60 
i915_fence_get_driver_name(struct dma_fence * fence)61 static const char *i915_fence_get_driver_name(struct dma_fence *fence)
62 {
63 	return dev_name(to_request(fence)->i915->drm.dev);
64 }
65 
i915_fence_get_timeline_name(struct dma_fence * fence)66 static const char *i915_fence_get_timeline_name(struct dma_fence *fence)
67 {
68 	const struct i915_gem_context *ctx;
69 
70 	/*
71 	 * The timeline struct (as part of the ppgtt underneath a context)
72 	 * may be freed when the request is no longer in use by the GPU.
73 	 * We could extend the life of a context to beyond that of all
74 	 * fences, possibly keeping the hw resource around indefinitely,
75 	 * or we just give them a false name. Since
76 	 * dma_fence_ops.get_timeline_name is a debug feature, the occasional
77 	 * lie seems justifiable.
78 	 */
79 	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
80 		return "signaled";
81 
82 	ctx = i915_request_gem_context(to_request(fence));
83 	if (!ctx)
84 		return "[" DRIVER_NAME "]";
85 
86 	return ctx->name;
87 }
88 
i915_fence_signaled(struct dma_fence * fence)89 static bool i915_fence_signaled(struct dma_fence *fence)
90 {
91 	return i915_request_completed(to_request(fence));
92 }
93 
i915_fence_enable_signaling(struct dma_fence * fence)94 static bool i915_fence_enable_signaling(struct dma_fence *fence)
95 {
96 	return i915_request_enable_breadcrumb(to_request(fence));
97 }
98 
i915_fence_wait(struct dma_fence * fence,bool interruptible,signed long timeout)99 static signed long i915_fence_wait(struct dma_fence *fence,
100 				   bool interruptible,
101 				   signed long timeout)
102 {
103 	return i915_request_wait_timeout(to_request(fence),
104 					 interruptible | I915_WAIT_PRIORITY,
105 					 timeout);
106 }
107 
i915_request_slab_cache(void)108 struct kmem_cache *i915_request_slab_cache(void)
109 {
110 	return slab_requests;
111 }
112 
i915_fence_release(struct dma_fence * fence)113 static void i915_fence_release(struct dma_fence *fence)
114 {
115 	struct i915_request *rq = to_request(fence);
116 
117 	GEM_BUG_ON(rq->guc_prio != GUC_PRIO_INIT &&
118 		   rq->guc_prio != GUC_PRIO_FINI);
119 
120 	i915_request_free_capture_list(fetch_and_zero(&rq->capture_list));
121 	if (rq->batch_res) {
122 		i915_vma_resource_put(rq->batch_res);
123 		rq->batch_res = NULL;
124 	}
125 
126 	/*
127 	 * The request is put onto a RCU freelist (i.e. the address
128 	 * is immediately reused), mark the fences as being freed now.
129 	 * Otherwise the debugobjects for the fences are only marked as
130 	 * freed when the slab cache itself is freed, and so we would get
131 	 * caught trying to reuse dead objects.
132 	 */
133 	i915_sw_fence_fini(&rq->submit);
134 	i915_sw_fence_fini(&rq->semaphore);
135 
136 	/*
137 	 * Keep one request on each engine for reserved use under mempressure
138 	 * do not use with virtual engines as this really is only needed for
139 	 * kernel contexts.
140 	 *
141 	 * We do not hold a reference to the engine here and so have to be
142 	 * very careful in what rq->engine we poke. The virtual engine is
143 	 * referenced via the rq->context and we released that ref during
144 	 * i915_request_retire(), ergo we must not dereference a virtual
145 	 * engine here. Not that we would want to, as the only consumer of
146 	 * the reserved engine->request_pool is the power management parking,
147 	 * which must-not-fail, and that is only run on the physical engines.
148 	 *
149 	 * Since the request must have been executed to be have completed,
150 	 * we know that it will have been processed by the HW and will
151 	 * not be unsubmitted again, so rq->engine and rq->execution_mask
152 	 * at this point is stable. rq->execution_mask will be a single
153 	 * bit if the last and _only_ engine it could execution on was a
154 	 * physical engine, if it's multiple bits then it started on and
155 	 * could still be on a virtual engine. Thus if the mask is not a
156 	 * power-of-two we assume that rq->engine may still be a virtual
157 	 * engine and so a dangling invalid pointer that we cannot dereference
158 	 *
159 	 * For example, consider the flow of a bonded request through a virtual
160 	 * engine. The request is created with a wide engine mask (all engines
161 	 * that we might execute on). On processing the bond, the request mask
162 	 * is reduced to one or more engines. If the request is subsequently
163 	 * bound to a single engine, it will then be constrained to only
164 	 * execute on that engine and never returned to the virtual engine
165 	 * after timeslicing away, see __unwind_incomplete_requests(). Thus we
166 	 * know that if the rq->execution_mask is a single bit, rq->engine
167 	 * can be a physical engine with the exact corresponding mask.
168 	 */
169 	if (!intel_engine_is_virtual(rq->engine) &&
170 	    is_power_of_2(rq->execution_mask) &&
171 	    !cmpxchg(&rq->engine->request_pool, NULL, rq))
172 		return;
173 
174 	kmem_cache_free(slab_requests, rq);
175 }
176 
177 const struct dma_fence_ops i915_fence_ops = {
178 	.get_driver_name = i915_fence_get_driver_name,
179 	.get_timeline_name = i915_fence_get_timeline_name,
180 	.enable_signaling = i915_fence_enable_signaling,
181 	.signaled = i915_fence_signaled,
182 	.wait = i915_fence_wait,
183 	.release = i915_fence_release,
184 };
185 
irq_execute_cb(struct irq_work * wrk)186 static void irq_execute_cb(struct irq_work *wrk)
187 {
188 	struct execute_cb *cb = container_of(wrk, typeof(*cb), work);
189 
190 	i915_sw_fence_complete(cb->fence);
191 	kmem_cache_free(slab_execute_cbs, cb);
192 }
193 
194 static __always_inline void
__notify_execute_cb(struct i915_request * rq,bool (* fn)(struct irq_work * wrk))195 __notify_execute_cb(struct i915_request *rq, bool (*fn)(struct irq_work *wrk))
196 {
197 	struct execute_cb *cb, *cn;
198 
199 	if (llist_empty(&rq->execute_cb))
200 		return;
201 
202 	llist_for_each_entry_safe(cb, cn,
203 				  llist_del_all(&rq->execute_cb),
204 				  work.node.llist)
205 		fn(&cb->work);
206 }
207 
__notify_execute_cb_irq(struct i915_request * rq)208 static void __notify_execute_cb_irq(struct i915_request *rq)
209 {
210 	__notify_execute_cb(rq, irq_work_queue);
211 }
212 
irq_work_imm(struct irq_work * wrk)213 static bool irq_work_imm(struct irq_work *wrk)
214 {
215 	wrk->func(wrk);
216 	return false;
217 }
218 
i915_request_notify_execute_cb_imm(struct i915_request * rq)219 void i915_request_notify_execute_cb_imm(struct i915_request *rq)
220 {
221 	__notify_execute_cb(rq, irq_work_imm);
222 }
223 
__i915_request_fill(struct i915_request * rq,u8 val)224 static void __i915_request_fill(struct i915_request *rq, u8 val)
225 {
226 	void *vaddr = rq->ring->vaddr;
227 	u32 head;
228 
229 	head = rq->infix;
230 	if (rq->postfix < head) {
231 		memset(vaddr + head, val, rq->ring->size - head);
232 		head = 0;
233 	}
234 	memset(vaddr + head, val, rq->postfix - head);
235 }
236 
237 /**
238  * i915_request_active_engine
239  * @rq: request to inspect
240  * @active: pointer in which to return the active engine
241  *
242  * Fills the currently active engine to the @active pointer if the request
243  * is active and still not completed.
244  *
245  * Returns true if request was active or false otherwise.
246  */
247 bool
i915_request_active_engine(struct i915_request * rq,struct intel_engine_cs ** active)248 i915_request_active_engine(struct i915_request *rq,
249 			   struct intel_engine_cs **active)
250 {
251 	struct intel_engine_cs *engine, *locked;
252 	bool ret = false;
253 
254 	/*
255 	 * Serialise with __i915_request_submit() so that it sees
256 	 * is-banned?, or we know the request is already inflight.
257 	 *
258 	 * Note that rq->engine is unstable, and so we double
259 	 * check that we have acquired the lock on the final engine.
260 	 */
261 	locked = READ_ONCE(rq->engine);
262 	spin_lock_irq(&locked->sched_engine->lock);
263 	while (unlikely(locked != (engine = READ_ONCE(rq->engine)))) {
264 		spin_unlock(&locked->sched_engine->lock);
265 		locked = engine;
266 		spin_lock(&locked->sched_engine->lock);
267 	}
268 
269 	if (i915_request_is_active(rq)) {
270 		if (!__i915_request_is_complete(rq))
271 			*active = locked;
272 		ret = true;
273 	}
274 
275 	spin_unlock_irq(&locked->sched_engine->lock);
276 
277 	return ret;
278 }
279 
__rq_init_watchdog(struct i915_request * rq)280 static void __rq_init_watchdog(struct i915_request *rq)
281 {
282 	rq->watchdog.timer.function = NULL;
283 }
284 
__rq_watchdog_expired(struct hrtimer * hrtimer)285 static enum hrtimer_restart __rq_watchdog_expired(struct hrtimer *hrtimer)
286 {
287 	struct i915_request *rq =
288 		container_of(hrtimer, struct i915_request, watchdog.timer);
289 	struct intel_gt *gt = rq->engine->gt;
290 
291 	if (!i915_request_completed(rq)) {
292 		if (llist_add(&rq->watchdog.link, &gt->watchdog.list))
293 			schedule_work(&gt->watchdog.work);
294 	} else {
295 		i915_request_put(rq);
296 	}
297 
298 	return HRTIMER_NORESTART;
299 }
300 
__rq_arm_watchdog(struct i915_request * rq)301 static void __rq_arm_watchdog(struct i915_request *rq)
302 {
303 	struct i915_request_watchdog *wdg = &rq->watchdog;
304 	struct intel_context *ce = rq->context;
305 
306 	if (!ce->watchdog.timeout_us)
307 		return;
308 
309 	i915_request_get(rq);
310 
311 	hrtimer_init(&wdg->timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
312 	wdg->timer.function = __rq_watchdog_expired;
313 	hrtimer_start_range_ns(&wdg->timer,
314 			       ns_to_ktime(ce->watchdog.timeout_us *
315 					   NSEC_PER_USEC),
316 			       NSEC_PER_MSEC,
317 			       HRTIMER_MODE_REL);
318 }
319 
__rq_cancel_watchdog(struct i915_request * rq)320 static void __rq_cancel_watchdog(struct i915_request *rq)
321 {
322 	struct i915_request_watchdog *wdg = &rq->watchdog;
323 
324 	if (wdg->timer.function && hrtimer_try_to_cancel(&wdg->timer) > 0)
325 		i915_request_put(rq);
326 }
327 
328 #if IS_ENABLED(CONFIG_DRM_I915_CAPTURE_ERROR)
329 
330 /**
331  * i915_request_free_capture_list - Free a capture list
332  * @capture: Pointer to the first list item or NULL
333  *
334  */
i915_request_free_capture_list(struct i915_capture_list * capture)335 void i915_request_free_capture_list(struct i915_capture_list *capture)
336 {
337 	while (capture) {
338 		struct i915_capture_list *next = capture->next;
339 
340 		i915_vma_resource_put(capture->vma_res);
341 		kfree(capture);
342 		capture = next;
343 	}
344 }
345 
346 #define assert_capture_list_is_null(_rq) GEM_BUG_ON((_rq)->capture_list)
347 
348 #define clear_capture_list(_rq) ((_rq)->capture_list = NULL)
349 
350 #else
351 
352 #define i915_request_free_capture_list(_a) do {} while (0)
353 
354 #define assert_capture_list_is_null(_a) do {} while (0)
355 
356 #define clear_capture_list(_rq) do {} while (0)
357 
358 #endif
359 
i915_request_retire(struct i915_request * rq)360 bool i915_request_retire(struct i915_request *rq)
361 {
362 	if (!__i915_request_is_complete(rq))
363 		return false;
364 
365 	RQ_TRACE(rq, "\n");
366 
367 	GEM_BUG_ON(!i915_sw_fence_signaled(&rq->submit));
368 	trace_i915_request_retire(rq);
369 	i915_request_mark_complete(rq);
370 
371 	__rq_cancel_watchdog(rq);
372 
373 	/*
374 	 * We know the GPU must have read the request to have
375 	 * sent us the seqno + interrupt, so use the position
376 	 * of tail of the request to update the last known position
377 	 * of the GPU head.
378 	 *
379 	 * Note this requires that we are always called in request
380 	 * completion order.
381 	 */
382 	GEM_BUG_ON(!list_is_first(&rq->link,
383 				  &i915_request_timeline(rq)->requests));
384 	if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
385 		/* Poison before we release our space in the ring */
386 		__i915_request_fill(rq, POISON_FREE);
387 	rq->ring->head = rq->postfix;
388 
389 	if (!i915_request_signaled(rq)) {
390 		spin_lock_irq(&rq->lock);
391 		dma_fence_signal_locked(&rq->fence);
392 		spin_unlock_irq(&rq->lock);
393 	}
394 
395 	if (test_and_set_bit(I915_FENCE_FLAG_BOOST, &rq->fence.flags))
396 		intel_rps_dec_waiters(&rq->engine->gt->rps);
397 
398 	/*
399 	 * We only loosely track inflight requests across preemption,
400 	 * and so we may find ourselves attempting to retire a _completed_
401 	 * request that we have removed from the HW and put back on a run
402 	 * queue.
403 	 *
404 	 * As we set I915_FENCE_FLAG_ACTIVE on the request, this should be
405 	 * after removing the breadcrumb and signaling it, so that we do not
406 	 * inadvertently attach the breadcrumb to a completed request.
407 	 */
408 	rq->engine->remove_active_request(rq);
409 	GEM_BUG_ON(!llist_empty(&rq->execute_cb));
410 
411 	__list_del_entry(&rq->link); /* poison neither prev/next (RCU walks) */
412 
413 	intel_context_exit(rq->context);
414 	intel_context_unpin(rq->context);
415 
416 	i915_sched_node_fini(&rq->sched);
417 	i915_request_put(rq);
418 
419 	return true;
420 }
421 
i915_request_retire_upto(struct i915_request * rq)422 void i915_request_retire_upto(struct i915_request *rq)
423 {
424 	struct intel_timeline * const tl = i915_request_timeline(rq);
425 	struct i915_request *tmp;
426 
427 	RQ_TRACE(rq, "\n");
428 	GEM_BUG_ON(!__i915_request_is_complete(rq));
429 
430 	do {
431 		tmp = list_first_entry(&tl->requests, typeof(*tmp), link);
432 		GEM_BUG_ON(!i915_request_completed(tmp));
433 	} while (i915_request_retire(tmp) && tmp != rq);
434 }
435 
436 static struct i915_request * const *
__engine_active(struct intel_engine_cs * engine)437 __engine_active(struct intel_engine_cs *engine)
438 {
439 	return READ_ONCE(engine->execlists.active);
440 }
441 
__request_in_flight(const struct i915_request * signal)442 static bool __request_in_flight(const struct i915_request *signal)
443 {
444 	struct i915_request * const *port, *rq;
445 	bool inflight = false;
446 
447 	if (!i915_request_is_ready(signal))
448 		return false;
449 
450 	/*
451 	 * Even if we have unwound the request, it may still be on
452 	 * the GPU (preempt-to-busy). If that request is inside an
453 	 * unpreemptible critical section, it will not be removed. Some
454 	 * GPU functions may even be stuck waiting for the paired request
455 	 * (__await_execution) to be submitted and cannot be preempted
456 	 * until the bond is executing.
457 	 *
458 	 * As we know that there are always preemption points between
459 	 * requests, we know that only the currently executing request
460 	 * may be still active even though we have cleared the flag.
461 	 * However, we can't rely on our tracking of ELSP[0] to know
462 	 * which request is currently active and so maybe stuck, as
463 	 * the tracking maybe an event behind. Instead assume that
464 	 * if the context is still inflight, then it is still active
465 	 * even if the active flag has been cleared.
466 	 *
467 	 * To further complicate matters, if there a pending promotion, the HW
468 	 * may either perform a context switch to the second inflight execlists,
469 	 * or it may switch to the pending set of execlists. In the case of the
470 	 * latter, it may send the ACK and we process the event copying the
471 	 * pending[] over top of inflight[], _overwriting_ our *active. Since
472 	 * this implies the HW is arbitrating and not struck in *active, we do
473 	 * not worry about complete accuracy, but we do require no read/write
474 	 * tearing of the pointer [the read of the pointer must be valid, even
475 	 * as the array is being overwritten, for which we require the writes
476 	 * to avoid tearing.]
477 	 *
478 	 * Note that the read of *execlists->active may race with the promotion
479 	 * of execlists->pending[] to execlists->inflight[], overwritting
480 	 * the value at *execlists->active. This is fine. The promotion implies
481 	 * that we received an ACK from the HW, and so the context is not
482 	 * stuck -- if we do not see ourselves in *active, the inflight status
483 	 * is valid. If instead we see ourselves being copied into *active,
484 	 * we are inflight and may signal the callback.
485 	 */
486 	if (!intel_context_inflight(signal->context))
487 		return false;
488 
489 	rcu_read_lock();
490 	for (port = __engine_active(signal->engine);
491 	     (rq = READ_ONCE(*port)); /* may race with promotion of pending[] */
492 	     port++) {
493 		if (rq->context == signal->context) {
494 			inflight = i915_seqno_passed(rq->fence.seqno,
495 						     signal->fence.seqno);
496 			break;
497 		}
498 	}
499 	rcu_read_unlock();
500 
501 	return inflight;
502 }
503 
504 static int
__await_execution(struct i915_request * rq,struct i915_request * signal,gfp_t gfp)505 __await_execution(struct i915_request *rq,
506 		  struct i915_request *signal,
507 		  gfp_t gfp)
508 {
509 	struct execute_cb *cb;
510 
511 	if (i915_request_is_active(signal))
512 		return 0;
513 
514 	cb = kmem_cache_alloc(slab_execute_cbs, gfp);
515 	if (!cb)
516 		return -ENOMEM;
517 
518 	cb->fence = &rq->submit;
519 	i915_sw_fence_await(cb->fence);
520 	init_irq_work(&cb->work, irq_execute_cb);
521 
522 	/*
523 	 * Register the callback first, then see if the signaler is already
524 	 * active. This ensures that if we race with the
525 	 * __notify_execute_cb from i915_request_submit() and we are not
526 	 * included in that list, we get a second bite of the cherry and
527 	 * execute it ourselves. After this point, a future
528 	 * i915_request_submit() will notify us.
529 	 *
530 	 * In i915_request_retire() we set the ACTIVE bit on a completed
531 	 * request (then flush the execute_cb). So by registering the
532 	 * callback first, then checking the ACTIVE bit, we serialise with
533 	 * the completed/retired request.
534 	 */
535 	if (llist_add(&cb->work.node.llist, &signal->execute_cb)) {
536 		if (i915_request_is_active(signal) ||
537 		    __request_in_flight(signal))
538 			i915_request_notify_execute_cb_imm(signal);
539 	}
540 
541 	return 0;
542 }
543 
fatal_error(int error)544 static bool fatal_error(int error)
545 {
546 	switch (error) {
547 	case 0: /* not an error! */
548 	case -EAGAIN: /* innocent victim of a GT reset (__i915_request_reset) */
549 	case -ETIMEDOUT: /* waiting for Godot (timer_i915_sw_fence_wake) */
550 		return false;
551 	default:
552 		return true;
553 	}
554 }
555 
__i915_request_skip(struct i915_request * rq)556 void __i915_request_skip(struct i915_request *rq)
557 {
558 	GEM_BUG_ON(!fatal_error(rq->fence.error));
559 
560 	if (rq->infix == rq->postfix)
561 		return;
562 
563 	RQ_TRACE(rq, "error: %d\n", rq->fence.error);
564 
565 	/*
566 	 * As this request likely depends on state from the lost
567 	 * context, clear out all the user operations leaving the
568 	 * breadcrumb at the end (so we get the fence notifications).
569 	 */
570 	__i915_request_fill(rq, 0);
571 	rq->infix = rq->postfix;
572 }
573 
i915_request_set_error_once(struct i915_request * rq,int error)574 bool i915_request_set_error_once(struct i915_request *rq, int error)
575 {
576 	int old;
577 
578 	GEM_BUG_ON(!IS_ERR_VALUE((long)error));
579 
580 	if (i915_request_signaled(rq))
581 		return false;
582 
583 	old = READ_ONCE(rq->fence.error);
584 	do {
585 		if (fatal_error(old))
586 			return false;
587 	} while (!try_cmpxchg(&rq->fence.error, &old, error));
588 
589 	return true;
590 }
591 
i915_request_mark_eio(struct i915_request * rq)592 struct i915_request *i915_request_mark_eio(struct i915_request *rq)
593 {
594 	if (__i915_request_is_complete(rq))
595 		return NULL;
596 
597 	GEM_BUG_ON(i915_request_signaled(rq));
598 
599 	/* As soon as the request is completed, it may be retired */
600 	rq = i915_request_get(rq);
601 
602 	i915_request_set_error_once(rq, -EIO);
603 	i915_request_mark_complete(rq);
604 
605 	return rq;
606 }
607 
__i915_request_submit(struct i915_request * request)608 bool __i915_request_submit(struct i915_request *request)
609 {
610 	struct intel_engine_cs *engine = request->engine;
611 	bool result = false;
612 
613 	RQ_TRACE(request, "\n");
614 
615 	GEM_BUG_ON(!irqs_disabled());
616 	lockdep_assert_held(&engine->sched_engine->lock);
617 
618 	/*
619 	 * With the advent of preempt-to-busy, we frequently encounter
620 	 * requests that we have unsubmitted from HW, but left running
621 	 * until the next ack and so have completed in the meantime. On
622 	 * resubmission of that completed request, we can skip
623 	 * updating the payload, and execlists can even skip submitting
624 	 * the request.
625 	 *
626 	 * We must remove the request from the caller's priority queue,
627 	 * and the caller must only call us when the request is in their
628 	 * priority queue, under the sched_engine->lock. This ensures that the
629 	 * request has *not* yet been retired and we can safely move
630 	 * the request into the engine->active.list where it will be
631 	 * dropped upon retiring. (Otherwise if resubmit a *retired*
632 	 * request, this would be a horrible use-after-free.)
633 	 */
634 	if (__i915_request_is_complete(request)) {
635 		list_del_init(&request->sched.link);
636 		goto active;
637 	}
638 
639 	if (unlikely(!intel_context_is_schedulable(request->context)))
640 		i915_request_set_error_once(request, -EIO);
641 
642 	if (unlikely(fatal_error(request->fence.error)))
643 		__i915_request_skip(request);
644 
645 	/*
646 	 * Are we using semaphores when the gpu is already saturated?
647 	 *
648 	 * Using semaphores incurs a cost in having the GPU poll a
649 	 * memory location, busywaiting for it to change. The continual
650 	 * memory reads can have a noticeable impact on the rest of the
651 	 * system with the extra bus traffic, stalling the cpu as it too
652 	 * tries to access memory across the bus (perf stat -e bus-cycles).
653 	 *
654 	 * If we installed a semaphore on this request and we only submit
655 	 * the request after the signaler completed, that indicates the
656 	 * system is overloaded and using semaphores at this time only
657 	 * increases the amount of work we are doing. If so, we disable
658 	 * further use of semaphores until we are idle again, whence we
659 	 * optimistically try again.
660 	 */
661 	if (request->sched.semaphores &&
662 	    i915_sw_fence_signaled(&request->semaphore))
663 		engine->saturated |= request->sched.semaphores;
664 
665 	engine->emit_fini_breadcrumb(request,
666 				     request->ring->vaddr + request->postfix);
667 
668 	trace_i915_request_execute(request);
669 	if (engine->bump_serial)
670 		engine->bump_serial(engine);
671 	else
672 		engine->serial++;
673 
674 	result = true;
675 
676 	GEM_BUG_ON(test_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags));
677 	engine->add_active_request(request);
678 active:
679 	clear_bit(I915_FENCE_FLAG_PQUEUE, &request->fence.flags);
680 	set_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags);
681 
682 	/*
683 	 * XXX Rollback bonded-execution on __i915_request_unsubmit()?
684 	 *
685 	 * In the future, perhaps when we have an active time-slicing scheduler,
686 	 * it will be interesting to unsubmit parallel execution and remove
687 	 * busywaits from the GPU until their master is restarted. This is
688 	 * quite hairy, we have to carefully rollback the fence and do a
689 	 * preempt-to-idle cycle on the target engine, all the while the
690 	 * master execute_cb may refire.
691 	 */
692 	__notify_execute_cb_irq(request);
693 
694 	/* We may be recursing from the signal callback of another i915 fence */
695 	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags))
696 		i915_request_enable_breadcrumb(request);
697 
698 	return result;
699 }
700 
i915_request_submit(struct i915_request * request)701 void i915_request_submit(struct i915_request *request)
702 {
703 	struct intel_engine_cs *engine = request->engine;
704 	unsigned long flags;
705 
706 	/* Will be called from irq-context when using foreign fences. */
707 	spin_lock_irqsave(&engine->sched_engine->lock, flags);
708 
709 	__i915_request_submit(request);
710 
711 	spin_unlock_irqrestore(&engine->sched_engine->lock, flags);
712 }
713 
__i915_request_unsubmit(struct i915_request * request)714 void __i915_request_unsubmit(struct i915_request *request)
715 {
716 	struct intel_engine_cs *engine = request->engine;
717 
718 	/*
719 	 * Only unwind in reverse order, required so that the per-context list
720 	 * is kept in seqno/ring order.
721 	 */
722 	RQ_TRACE(request, "\n");
723 
724 	GEM_BUG_ON(!irqs_disabled());
725 	lockdep_assert_held(&engine->sched_engine->lock);
726 
727 	/*
728 	 * Before we remove this breadcrumb from the signal list, we have
729 	 * to ensure that a concurrent dma_fence_enable_signaling() does not
730 	 * attach itself. We first mark the request as no longer active and
731 	 * make sure that is visible to other cores, and then remove the
732 	 * breadcrumb if attached.
733 	 */
734 	GEM_BUG_ON(!test_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags));
735 	clear_bit_unlock(I915_FENCE_FLAG_ACTIVE, &request->fence.flags);
736 	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags))
737 		i915_request_cancel_breadcrumb(request);
738 
739 	/* We've already spun, don't charge on resubmitting. */
740 	if (request->sched.semaphores && __i915_request_has_started(request))
741 		request->sched.semaphores = 0;
742 
743 	/*
744 	 * We don't need to wake_up any waiters on request->execute, they
745 	 * will get woken by any other event or us re-adding this request
746 	 * to the engine timeline (__i915_request_submit()). The waiters
747 	 * should be quite adapt at finding that the request now has a new
748 	 * global_seqno to the one they went to sleep on.
749 	 */
750 }
751 
i915_request_unsubmit(struct i915_request * request)752 void i915_request_unsubmit(struct i915_request *request)
753 {
754 	struct intel_engine_cs *engine = request->engine;
755 	unsigned long flags;
756 
757 	/* Will be called from irq-context when using foreign fences. */
758 	spin_lock_irqsave(&engine->sched_engine->lock, flags);
759 
760 	__i915_request_unsubmit(request);
761 
762 	spin_unlock_irqrestore(&engine->sched_engine->lock, flags);
763 }
764 
i915_request_cancel(struct i915_request * rq,int error)765 void i915_request_cancel(struct i915_request *rq, int error)
766 {
767 	if (!i915_request_set_error_once(rq, error))
768 		return;
769 
770 	set_bit(I915_FENCE_FLAG_SENTINEL, &rq->fence.flags);
771 
772 	intel_context_cancel_request(rq->context, rq);
773 }
774 
775 static int
submit_notify(struct i915_sw_fence * fence,enum i915_sw_fence_notify state)776 submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
777 {
778 	struct i915_request *request =
779 		container_of(fence, typeof(*request), submit);
780 
781 	switch (state) {
782 	case FENCE_COMPLETE:
783 		trace_i915_request_submit(request);
784 
785 		if (unlikely(fence->error))
786 			i915_request_set_error_once(request, fence->error);
787 		else
788 			__rq_arm_watchdog(request);
789 
790 		/*
791 		 * We need to serialize use of the submit_request() callback
792 		 * with its hotplugging performed during an emergency
793 		 * i915_gem_set_wedged().  We use the RCU mechanism to mark the
794 		 * critical section in order to force i915_gem_set_wedged() to
795 		 * wait until the submit_request() is completed before
796 		 * proceeding.
797 		 */
798 		rcu_read_lock();
799 		request->engine->submit_request(request);
800 		rcu_read_unlock();
801 		break;
802 
803 	case FENCE_FREE:
804 		i915_request_put(request);
805 		break;
806 	}
807 
808 	return NOTIFY_DONE;
809 }
810 
811 static int
semaphore_notify(struct i915_sw_fence * fence,enum i915_sw_fence_notify state)812 semaphore_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state)
813 {
814 	struct i915_request *rq = container_of(fence, typeof(*rq), semaphore);
815 
816 	switch (state) {
817 	case FENCE_COMPLETE:
818 		break;
819 
820 	case FENCE_FREE:
821 		i915_request_put(rq);
822 		break;
823 	}
824 
825 	return NOTIFY_DONE;
826 }
827 
retire_requests(struct intel_timeline * tl)828 static void retire_requests(struct intel_timeline *tl)
829 {
830 	struct i915_request *rq, *rn;
831 
832 	list_for_each_entry_safe(rq, rn, &tl->requests, link)
833 		if (!i915_request_retire(rq))
834 			break;
835 }
836 
837 static noinline struct i915_request *
request_alloc_slow(struct intel_timeline * tl,struct i915_request ** rsvd,gfp_t gfp)838 request_alloc_slow(struct intel_timeline *tl,
839 		   struct i915_request **rsvd,
840 		   gfp_t gfp)
841 {
842 	struct i915_request *rq;
843 
844 	/* If we cannot wait, dip into our reserves */
845 	if (!gfpflags_allow_blocking(gfp)) {
846 		rq = xchg(rsvd, NULL);
847 		if (!rq) /* Use the normal failure path for one final WARN */
848 			goto out;
849 
850 		return rq;
851 	}
852 
853 	if (list_empty(&tl->requests))
854 		goto out;
855 
856 	/* Move our oldest request to the slab-cache (if not in use!) */
857 	rq = list_first_entry(&tl->requests, typeof(*rq), link);
858 	i915_request_retire(rq);
859 
860 	rq = kmem_cache_alloc(slab_requests,
861 			      gfp | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
862 	if (rq)
863 		return rq;
864 
865 	/* Ratelimit ourselves to prevent oom from malicious clients */
866 	rq = list_last_entry(&tl->requests, typeof(*rq), link);
867 	cond_synchronize_rcu(rq->rcustate);
868 
869 	/* Retire our old requests in the hope that we free some */
870 	retire_requests(tl);
871 
872 out:
873 	return kmem_cache_alloc(slab_requests, gfp);
874 }
875 
__i915_request_ctor(void * arg)876 static void __i915_request_ctor(void *arg)
877 {
878 	struct i915_request *rq = arg;
879 
880 	spin_lock_init(&rq->lock);
881 	i915_sched_node_init(&rq->sched);
882 	i915_sw_fence_init(&rq->submit, submit_notify);
883 	i915_sw_fence_init(&rq->semaphore, semaphore_notify);
884 
885 	clear_capture_list(rq);
886 	rq->batch_res = NULL;
887 
888 	init_llist_head(&rq->execute_cb);
889 }
890 
891 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
892 #define clear_batch_ptr(_rq) ((_rq)->batch = NULL)
893 #else
894 #define clear_batch_ptr(_a) do {} while (0)
895 #endif
896 
897 struct i915_request *
__i915_request_create(struct intel_context * ce,gfp_t gfp)898 __i915_request_create(struct intel_context *ce, gfp_t gfp)
899 {
900 	struct intel_timeline *tl = ce->timeline;
901 	struct i915_request *rq;
902 	u32 seqno;
903 	int ret;
904 
905 	might_alloc(gfp);
906 
907 	/* Check that the caller provided an already pinned context */
908 	__intel_context_pin(ce);
909 
910 	/*
911 	 * Beware: Dragons be flying overhead.
912 	 *
913 	 * We use RCU to look up requests in flight. The lookups may
914 	 * race with the request being allocated from the slab freelist.
915 	 * That is the request we are writing to here, may be in the process
916 	 * of being read by __i915_active_request_get_rcu(). As such,
917 	 * we have to be very careful when overwriting the contents. During
918 	 * the RCU lookup, we change chase the request->engine pointer,
919 	 * read the request->global_seqno and increment the reference count.
920 	 *
921 	 * The reference count is incremented atomically. If it is zero,
922 	 * the lookup knows the request is unallocated and complete. Otherwise,
923 	 * it is either still in use, or has been reallocated and reset
924 	 * with dma_fence_init(). This increment is safe for release as we
925 	 * check that the request we have a reference to and matches the active
926 	 * request.
927 	 *
928 	 * Before we increment the refcount, we chase the request->engine
929 	 * pointer. We must not call kmem_cache_zalloc() or else we set
930 	 * that pointer to NULL and cause a crash during the lookup. If
931 	 * we see the request is completed (based on the value of the
932 	 * old engine and seqno), the lookup is complete and reports NULL.
933 	 * If we decide the request is not completed (new engine or seqno),
934 	 * then we grab a reference and double check that it is still the
935 	 * active request - which it won't be and restart the lookup.
936 	 *
937 	 * Do not use kmem_cache_zalloc() here!
938 	 */
939 	rq = kmem_cache_alloc(slab_requests,
940 			      gfp | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
941 	if (unlikely(!rq)) {
942 		rq = request_alloc_slow(tl, &ce->engine->request_pool, gfp);
943 		if (!rq) {
944 			ret = -ENOMEM;
945 			goto err_unreserve;
946 		}
947 	}
948 
949 	rq->context = ce;
950 	rq->engine = ce->engine;
951 	rq->ring = ce->ring;
952 	rq->execution_mask = ce->engine->mask;
953 	rq->i915 = ce->engine->i915;
954 
955 	ret = intel_timeline_get_seqno(tl, rq, &seqno);
956 	if (ret)
957 		goto err_free;
958 
959 	dma_fence_init(&rq->fence, &i915_fence_ops, &rq->lock,
960 		       tl->fence_context, seqno);
961 
962 	RCU_INIT_POINTER(rq->timeline, tl);
963 	rq->hwsp_seqno = tl->hwsp_seqno;
964 	GEM_BUG_ON(__i915_request_is_complete(rq));
965 
966 	rq->rcustate = get_state_synchronize_rcu(); /* acts as smp_mb() */
967 
968 	rq->guc_prio = GUC_PRIO_INIT;
969 
970 	/* We bump the ref for the fence chain */
971 	i915_sw_fence_reinit(&i915_request_get(rq)->submit);
972 	i915_sw_fence_reinit(&i915_request_get(rq)->semaphore);
973 
974 	i915_sched_node_reinit(&rq->sched);
975 
976 	/* No zalloc, everything must be cleared after use */
977 	clear_batch_ptr(rq);
978 	__rq_init_watchdog(rq);
979 	assert_capture_list_is_null(rq);
980 	GEM_BUG_ON(!llist_empty(&rq->execute_cb));
981 	GEM_BUG_ON(rq->batch_res);
982 
983 	/*
984 	 * Reserve space in the ring buffer for all the commands required to
985 	 * eventually emit this request. This is to guarantee that the
986 	 * i915_request_add() call can't fail. Note that the reserve may need
987 	 * to be redone if the request is not actually submitted straight
988 	 * away, e.g. because a GPU scheduler has deferred it.
989 	 *
990 	 * Note that due to how we add reserved_space to intel_ring_begin()
991 	 * we need to double our request to ensure that if we need to wrap
992 	 * around inside i915_request_add() there is sufficient space at
993 	 * the beginning of the ring as well.
994 	 */
995 	rq->reserved_space =
996 		2 * rq->engine->emit_fini_breadcrumb_dw * sizeof(u32);
997 
998 	/*
999 	 * Record the position of the start of the request so that
1000 	 * should we detect the updated seqno part-way through the
1001 	 * GPU processing the request, we never over-estimate the
1002 	 * position of the head.
1003 	 */
1004 	rq->head = rq->ring->emit;
1005 
1006 	ret = rq->engine->request_alloc(rq);
1007 	if (ret)
1008 		goto err_unwind;
1009 
1010 	rq->infix = rq->ring->emit; /* end of header; start of user payload */
1011 
1012 	intel_context_mark_active(ce);
1013 	list_add_tail_rcu(&rq->link, &tl->requests);
1014 
1015 	return rq;
1016 
1017 err_unwind:
1018 	ce->ring->emit = rq->head;
1019 
1020 	/* Make sure we didn't add ourselves to external state before freeing */
1021 	GEM_BUG_ON(!list_empty(&rq->sched.signalers_list));
1022 	GEM_BUG_ON(!list_empty(&rq->sched.waiters_list));
1023 
1024 err_free:
1025 	kmem_cache_free(slab_requests, rq);
1026 err_unreserve:
1027 	intel_context_unpin(ce);
1028 	return ERR_PTR(ret);
1029 }
1030 
1031 struct i915_request *
i915_request_create(struct intel_context * ce)1032 i915_request_create(struct intel_context *ce)
1033 {
1034 	struct i915_request *rq;
1035 	struct intel_timeline *tl;
1036 
1037 	tl = intel_context_timeline_lock(ce);
1038 	if (IS_ERR(tl))
1039 		return ERR_CAST(tl);
1040 
1041 	/* Move our oldest request to the slab-cache (if not in use!) */
1042 	rq = list_first_entry(&tl->requests, typeof(*rq), link);
1043 	if (!list_is_last(&rq->link, &tl->requests))
1044 		i915_request_retire(rq);
1045 
1046 	intel_context_enter(ce);
1047 	rq = __i915_request_create(ce, GFP_KERNEL);
1048 	intel_context_exit(ce); /* active reference transferred to request */
1049 	if (IS_ERR(rq))
1050 		goto err_unlock;
1051 
1052 	/* Check that we do not interrupt ourselves with a new request */
1053 	rq->cookie = lockdep_pin_lock(&tl->mutex);
1054 
1055 	return rq;
1056 
1057 err_unlock:
1058 	intel_context_timeline_unlock(tl);
1059 	return rq;
1060 }
1061 
1062 static int
i915_request_await_start(struct i915_request * rq,struct i915_request * signal)1063 i915_request_await_start(struct i915_request *rq, struct i915_request *signal)
1064 {
1065 	struct dma_fence *fence;
1066 	int err;
1067 
1068 	if (i915_request_timeline(rq) == rcu_access_pointer(signal->timeline))
1069 		return 0;
1070 
1071 	if (i915_request_started(signal))
1072 		return 0;
1073 
1074 	/*
1075 	 * The caller holds a reference on @signal, but we do not serialise
1076 	 * against it being retired and removed from the lists.
1077 	 *
1078 	 * We do not hold a reference to the request before @signal, and
1079 	 * so must be very careful to ensure that it is not _recycled_ as
1080 	 * we follow the link backwards.
1081 	 */
1082 	fence = NULL;
1083 	rcu_read_lock();
1084 	do {
1085 		struct list_head *pos = READ_ONCE(signal->link.prev);
1086 		struct i915_request *prev;
1087 
1088 		/* Confirm signal has not been retired, the link is valid */
1089 		if (unlikely(__i915_request_has_started(signal)))
1090 			break;
1091 
1092 		/* Is signal the earliest request on its timeline? */
1093 		if (pos == &rcu_dereference(signal->timeline)->requests)
1094 			break;
1095 
1096 		/*
1097 		 * Peek at the request before us in the timeline. That
1098 		 * request will only be valid before it is retired, so
1099 		 * after acquiring a reference to it, confirm that it is
1100 		 * still part of the signaler's timeline.
1101 		 */
1102 		prev = list_entry(pos, typeof(*prev), link);
1103 		if (!i915_request_get_rcu(prev))
1104 			break;
1105 
1106 		/* After the strong barrier, confirm prev is still attached */
1107 		if (unlikely(READ_ONCE(prev->link.next) != &signal->link)) {
1108 			i915_request_put(prev);
1109 			break;
1110 		}
1111 
1112 		fence = &prev->fence;
1113 	} while (0);
1114 	rcu_read_unlock();
1115 	if (!fence)
1116 		return 0;
1117 
1118 	err = 0;
1119 	if (!intel_timeline_sync_is_later(i915_request_timeline(rq), fence))
1120 		err = i915_sw_fence_await_dma_fence(&rq->submit,
1121 						    fence, 0,
1122 						    I915_FENCE_GFP);
1123 	dma_fence_put(fence);
1124 
1125 	return err;
1126 }
1127 
1128 static intel_engine_mask_t
already_busywaiting(struct i915_request * rq)1129 already_busywaiting(struct i915_request *rq)
1130 {
1131 	/*
1132 	 * Polling a semaphore causes bus traffic, delaying other users of
1133 	 * both the GPU and CPU. We want to limit the impact on others,
1134 	 * while taking advantage of early submission to reduce GPU
1135 	 * latency. Therefore we restrict ourselves to not using more
1136 	 * than one semaphore from each source, and not using a semaphore
1137 	 * if we have detected the engine is saturated (i.e. would not be
1138 	 * submitted early and cause bus traffic reading an already passed
1139 	 * semaphore).
1140 	 *
1141 	 * See the are-we-too-late? check in __i915_request_submit().
1142 	 */
1143 	return rq->sched.semaphores | READ_ONCE(rq->engine->saturated);
1144 }
1145 
1146 static int
__emit_semaphore_wait(struct i915_request * to,struct i915_request * from,u32 seqno)1147 __emit_semaphore_wait(struct i915_request *to,
1148 		      struct i915_request *from,
1149 		      u32 seqno)
1150 {
1151 	const int has_token = GRAPHICS_VER(to->engine->i915) >= 12;
1152 	u32 hwsp_offset;
1153 	int len, err;
1154 	u32 *cs;
1155 
1156 	GEM_BUG_ON(GRAPHICS_VER(to->engine->i915) < 8);
1157 	GEM_BUG_ON(i915_request_has_initial_breadcrumb(to));
1158 
1159 	/* We need to pin the signaler's HWSP until we are finished reading. */
1160 	err = intel_timeline_read_hwsp(from, to, &hwsp_offset);
1161 	if (err)
1162 		return err;
1163 
1164 	len = 4;
1165 	if (has_token)
1166 		len += 2;
1167 
1168 	cs = intel_ring_begin(to, len);
1169 	if (IS_ERR(cs))
1170 		return PTR_ERR(cs);
1171 
1172 	/*
1173 	 * Using greater-than-or-equal here means we have to worry
1174 	 * about seqno wraparound. To side step that issue, we swap
1175 	 * the timeline HWSP upon wrapping, so that everyone listening
1176 	 * for the old (pre-wrap) values do not see the much smaller
1177 	 * (post-wrap) values than they were expecting (and so wait
1178 	 * forever).
1179 	 */
1180 	*cs++ = (MI_SEMAPHORE_WAIT |
1181 		 MI_SEMAPHORE_GLOBAL_GTT |
1182 		 MI_SEMAPHORE_POLL |
1183 		 MI_SEMAPHORE_SAD_GTE_SDD) +
1184 		has_token;
1185 	*cs++ = seqno;
1186 	*cs++ = hwsp_offset;
1187 	*cs++ = 0;
1188 	if (has_token) {
1189 		*cs++ = 0;
1190 		*cs++ = MI_NOOP;
1191 	}
1192 
1193 	intel_ring_advance(to, cs);
1194 	return 0;
1195 }
1196 
1197 static bool
can_use_semaphore_wait(struct i915_request * to,struct i915_request * from)1198 can_use_semaphore_wait(struct i915_request *to, struct i915_request *from)
1199 {
1200 	return to->engine->gt->ggtt == from->engine->gt->ggtt;
1201 }
1202 
1203 static int
emit_semaphore_wait(struct i915_request * to,struct i915_request * from,gfp_t gfp)1204 emit_semaphore_wait(struct i915_request *to,
1205 		    struct i915_request *from,
1206 		    gfp_t gfp)
1207 {
1208 	const intel_engine_mask_t mask = READ_ONCE(from->engine)->mask;
1209 	struct i915_sw_fence *wait = &to->submit;
1210 
1211 	if (!can_use_semaphore_wait(to, from))
1212 		goto await_fence;
1213 
1214 	if (!intel_context_use_semaphores(to->context))
1215 		goto await_fence;
1216 
1217 	if (i915_request_has_initial_breadcrumb(to))
1218 		goto await_fence;
1219 
1220 	/*
1221 	 * If this or its dependents are waiting on an external fence
1222 	 * that may fail catastrophically, then we want to avoid using
1223 	 * sempahores as they bypass the fence signaling metadata, and we
1224 	 * lose the fence->error propagation.
1225 	 */
1226 	if (from->sched.flags & I915_SCHED_HAS_EXTERNAL_CHAIN)
1227 		goto await_fence;
1228 
1229 	/* Just emit the first semaphore we see as request space is limited. */
1230 	if (already_busywaiting(to) & mask)
1231 		goto await_fence;
1232 
1233 	if (i915_request_await_start(to, from) < 0)
1234 		goto await_fence;
1235 
1236 	/* Only submit our spinner after the signaler is running! */
1237 	if (__await_execution(to, from, gfp))
1238 		goto await_fence;
1239 
1240 	if (__emit_semaphore_wait(to, from, from->fence.seqno))
1241 		goto await_fence;
1242 
1243 	to->sched.semaphores |= mask;
1244 	wait = &to->semaphore;
1245 
1246 await_fence:
1247 	return i915_sw_fence_await_dma_fence(wait,
1248 					     &from->fence, 0,
1249 					     I915_FENCE_GFP);
1250 }
1251 
intel_timeline_sync_has_start(struct intel_timeline * tl,struct dma_fence * fence)1252 static bool intel_timeline_sync_has_start(struct intel_timeline *tl,
1253 					  struct dma_fence *fence)
1254 {
1255 	return __intel_timeline_sync_is_later(tl,
1256 					      fence->context,
1257 					      fence->seqno - 1);
1258 }
1259 
intel_timeline_sync_set_start(struct intel_timeline * tl,const struct dma_fence * fence)1260 static int intel_timeline_sync_set_start(struct intel_timeline *tl,
1261 					 const struct dma_fence *fence)
1262 {
1263 	return __intel_timeline_sync_set(tl, fence->context, fence->seqno - 1);
1264 }
1265 
1266 static int
__i915_request_await_execution(struct i915_request * to,struct i915_request * from)1267 __i915_request_await_execution(struct i915_request *to,
1268 			       struct i915_request *from)
1269 {
1270 	int err;
1271 
1272 	GEM_BUG_ON(intel_context_is_barrier(from->context));
1273 
1274 	/* Submit both requests at the same time */
1275 	err = __await_execution(to, from, I915_FENCE_GFP);
1276 	if (err)
1277 		return err;
1278 
1279 	/* Squash repeated depenendices to the same timelines */
1280 	if (intel_timeline_sync_has_start(i915_request_timeline(to),
1281 					  &from->fence))
1282 		return 0;
1283 
1284 	/*
1285 	 * Wait until the start of this request.
1286 	 *
1287 	 * The execution cb fires when we submit the request to HW. But in
1288 	 * many cases this may be long before the request itself is ready to
1289 	 * run (consider that we submit 2 requests for the same context, where
1290 	 * the request of interest is behind an indefinite spinner). So we hook
1291 	 * up to both to reduce our queues and keep the execution lag minimised
1292 	 * in the worst case, though we hope that the await_start is elided.
1293 	 */
1294 	err = i915_request_await_start(to, from);
1295 	if (err < 0)
1296 		return err;
1297 
1298 	/*
1299 	 * Ensure both start together [after all semaphores in signal]
1300 	 *
1301 	 * Now that we are queued to the HW at roughly the same time (thanks
1302 	 * to the execute cb) and are ready to run at roughly the same time
1303 	 * (thanks to the await start), our signaler may still be indefinitely
1304 	 * delayed by waiting on a semaphore from a remote engine. If our
1305 	 * signaler depends on a semaphore, so indirectly do we, and we do not
1306 	 * want to start our payload until our signaler also starts theirs.
1307 	 * So we wait.
1308 	 *
1309 	 * However, there is also a second condition for which we need to wait
1310 	 * for the precise start of the signaler. Consider that the signaler
1311 	 * was submitted in a chain of requests following another context
1312 	 * (with just an ordinary intra-engine fence dependency between the
1313 	 * two). In this case the signaler is queued to HW, but not for
1314 	 * immediate execution, and so we must wait until it reaches the
1315 	 * active slot.
1316 	 */
1317 	if (can_use_semaphore_wait(to, from) &&
1318 	    intel_engine_has_semaphores(to->engine) &&
1319 	    !i915_request_has_initial_breadcrumb(to)) {
1320 		err = __emit_semaphore_wait(to, from, from->fence.seqno - 1);
1321 		if (err < 0)
1322 			return err;
1323 	}
1324 
1325 	/* Couple the dependency tree for PI on this exposed to->fence */
1326 	if (to->engine->sched_engine->schedule) {
1327 		err = i915_sched_node_add_dependency(&to->sched,
1328 						     &from->sched,
1329 						     I915_DEPENDENCY_WEAK);
1330 		if (err < 0)
1331 			return err;
1332 	}
1333 
1334 	return intel_timeline_sync_set_start(i915_request_timeline(to),
1335 					     &from->fence);
1336 }
1337 
mark_external(struct i915_request * rq)1338 static void mark_external(struct i915_request *rq)
1339 {
1340 	/*
1341 	 * The downside of using semaphores is that we lose metadata passing
1342 	 * along the signaling chain. This is particularly nasty when we
1343 	 * need to pass along a fatal error such as EFAULT or EDEADLK. For
1344 	 * fatal errors we want to scrub the request before it is executed,
1345 	 * which means that we cannot preload the request onto HW and have
1346 	 * it wait upon a semaphore.
1347 	 */
1348 	rq->sched.flags |= I915_SCHED_HAS_EXTERNAL_CHAIN;
1349 }
1350 
1351 static int
__i915_request_await_external(struct i915_request * rq,struct dma_fence * fence)1352 __i915_request_await_external(struct i915_request *rq, struct dma_fence *fence)
1353 {
1354 	mark_external(rq);
1355 	return i915_sw_fence_await_dma_fence(&rq->submit, fence,
1356 					     i915_fence_context_timeout(rq->engine->i915,
1357 									fence->context),
1358 					     I915_FENCE_GFP);
1359 }
1360 
1361 static int
i915_request_await_external(struct i915_request * rq,struct dma_fence * fence)1362 i915_request_await_external(struct i915_request *rq, struct dma_fence *fence)
1363 {
1364 	struct dma_fence *iter;
1365 	int err = 0;
1366 
1367 	if (!to_dma_fence_chain(fence))
1368 		return __i915_request_await_external(rq, fence);
1369 
1370 	dma_fence_chain_for_each(iter, fence) {
1371 		struct dma_fence_chain *chain = to_dma_fence_chain(iter);
1372 
1373 		if (!dma_fence_is_i915(chain->fence)) {
1374 			err = __i915_request_await_external(rq, iter);
1375 			break;
1376 		}
1377 
1378 		err = i915_request_await_dma_fence(rq, chain->fence);
1379 		if (err < 0)
1380 			break;
1381 	}
1382 
1383 	dma_fence_put(iter);
1384 	return err;
1385 }
1386 
is_parallel_rq(struct i915_request * rq)1387 static inline bool is_parallel_rq(struct i915_request *rq)
1388 {
1389 	return intel_context_is_parallel(rq->context);
1390 }
1391 
request_to_parent(struct i915_request * rq)1392 static inline struct intel_context *request_to_parent(struct i915_request *rq)
1393 {
1394 	return intel_context_to_parent(rq->context);
1395 }
1396 
is_same_parallel_context(struct i915_request * to,struct i915_request * from)1397 static bool is_same_parallel_context(struct i915_request *to,
1398 				     struct i915_request *from)
1399 {
1400 	if (is_parallel_rq(to))
1401 		return request_to_parent(to) == request_to_parent(from);
1402 
1403 	return false;
1404 }
1405 
1406 int
i915_request_await_execution(struct i915_request * rq,struct dma_fence * fence)1407 i915_request_await_execution(struct i915_request *rq,
1408 			     struct dma_fence *fence)
1409 {
1410 	struct dma_fence **child = &fence;
1411 	unsigned int nchild = 1;
1412 	int ret;
1413 
1414 	if (dma_fence_is_array(fence)) {
1415 		struct dma_fence_array *array = to_dma_fence_array(fence);
1416 
1417 		/* XXX Error for signal-on-any fence arrays */
1418 
1419 		child = array->fences;
1420 		nchild = array->num_fences;
1421 		GEM_BUG_ON(!nchild);
1422 	}
1423 
1424 	do {
1425 		fence = *child++;
1426 		if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
1427 			continue;
1428 
1429 		if (fence->context == rq->fence.context)
1430 			continue;
1431 
1432 		/*
1433 		 * We don't squash repeated fence dependencies here as we
1434 		 * want to run our callback in all cases.
1435 		 */
1436 
1437 		if (dma_fence_is_i915(fence)) {
1438 			if (is_same_parallel_context(rq, to_request(fence)))
1439 				continue;
1440 			ret = __i915_request_await_execution(rq,
1441 							     to_request(fence));
1442 		} else {
1443 			ret = i915_request_await_external(rq, fence);
1444 		}
1445 		if (ret < 0)
1446 			return ret;
1447 	} while (--nchild);
1448 
1449 	return 0;
1450 }
1451 
1452 static int
await_request_submit(struct i915_request * to,struct i915_request * from)1453 await_request_submit(struct i915_request *to, struct i915_request *from)
1454 {
1455 	/*
1456 	 * If we are waiting on a virtual engine, then it may be
1457 	 * constrained to execute on a single engine *prior* to submission.
1458 	 * When it is submitted, it will be first submitted to the virtual
1459 	 * engine and then passed to the physical engine. We cannot allow
1460 	 * the waiter to be submitted immediately to the physical engine
1461 	 * as it may then bypass the virtual request.
1462 	 */
1463 	if (to->engine == READ_ONCE(from->engine))
1464 		return i915_sw_fence_await_sw_fence_gfp(&to->submit,
1465 							&from->submit,
1466 							I915_FENCE_GFP);
1467 	else
1468 		return __i915_request_await_execution(to, from);
1469 }
1470 
1471 static int
i915_request_await_request(struct i915_request * to,struct i915_request * from)1472 i915_request_await_request(struct i915_request *to, struct i915_request *from)
1473 {
1474 	int ret;
1475 
1476 	GEM_BUG_ON(to == from);
1477 	GEM_BUG_ON(to->timeline == from->timeline);
1478 
1479 	if (i915_request_completed(from)) {
1480 		i915_sw_fence_set_error_once(&to->submit, from->fence.error);
1481 		return 0;
1482 	}
1483 
1484 	if (to->engine->sched_engine->schedule) {
1485 		ret = i915_sched_node_add_dependency(&to->sched,
1486 						     &from->sched,
1487 						     I915_DEPENDENCY_EXTERNAL);
1488 		if (ret < 0)
1489 			return ret;
1490 	}
1491 
1492 	if (!intel_engine_uses_guc(to->engine) &&
1493 	    is_power_of_2(to->execution_mask | READ_ONCE(from->execution_mask)))
1494 		ret = await_request_submit(to, from);
1495 	else
1496 		ret = emit_semaphore_wait(to, from, I915_FENCE_GFP);
1497 	if (ret < 0)
1498 		return ret;
1499 
1500 	return 0;
1501 }
1502 
1503 int
i915_request_await_dma_fence(struct i915_request * rq,struct dma_fence * fence)1504 i915_request_await_dma_fence(struct i915_request *rq, struct dma_fence *fence)
1505 {
1506 	struct dma_fence **child = &fence;
1507 	unsigned int nchild = 1;
1508 	int ret;
1509 
1510 	/*
1511 	 * Note that if the fence-array was created in signal-on-any mode,
1512 	 * we should *not* decompose it into its individual fences. However,
1513 	 * we don't currently store which mode the fence-array is operating
1514 	 * in. Fortunately, the only user of signal-on-any is private to
1515 	 * amdgpu and we should not see any incoming fence-array from
1516 	 * sync-file being in signal-on-any mode.
1517 	 */
1518 	if (dma_fence_is_array(fence)) {
1519 		struct dma_fence_array *array = to_dma_fence_array(fence);
1520 
1521 		child = array->fences;
1522 		nchild = array->num_fences;
1523 		GEM_BUG_ON(!nchild);
1524 	}
1525 
1526 	do {
1527 		fence = *child++;
1528 		if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
1529 			continue;
1530 
1531 		/*
1532 		 * Requests on the same timeline are explicitly ordered, along
1533 		 * with their dependencies, by i915_request_add() which ensures
1534 		 * that requests are submitted in-order through each ring.
1535 		 */
1536 		if (fence->context == rq->fence.context)
1537 			continue;
1538 
1539 		/* Squash repeated waits to the same timelines */
1540 		if (fence->context &&
1541 		    intel_timeline_sync_is_later(i915_request_timeline(rq),
1542 						 fence))
1543 			continue;
1544 
1545 		if (dma_fence_is_i915(fence)) {
1546 			if (is_same_parallel_context(rq, to_request(fence)))
1547 				continue;
1548 			ret = i915_request_await_request(rq, to_request(fence));
1549 		} else {
1550 			ret = i915_request_await_external(rq, fence);
1551 		}
1552 		if (ret < 0)
1553 			return ret;
1554 
1555 		/* Record the latest fence used against each timeline */
1556 		if (fence->context)
1557 			intel_timeline_sync_set(i915_request_timeline(rq),
1558 						fence);
1559 	} while (--nchild);
1560 
1561 	return 0;
1562 }
1563 
1564 /**
1565  * i915_request_await_deps - set this request to (async) wait upon a struct
1566  * i915_deps dma_fence collection
1567  * @rq: request we are wishing to use
1568  * @deps: The struct i915_deps containing the dependencies.
1569  *
1570  * Returns 0 if successful, negative error code on error.
1571  */
i915_request_await_deps(struct i915_request * rq,const struct i915_deps * deps)1572 int i915_request_await_deps(struct i915_request *rq, const struct i915_deps *deps)
1573 {
1574 	int i, err;
1575 
1576 	for (i = 0; i < deps->num_deps; ++i) {
1577 		err = i915_request_await_dma_fence(rq, deps->fences[i]);
1578 		if (err)
1579 			return err;
1580 	}
1581 
1582 	return 0;
1583 }
1584 
1585 /**
1586  * i915_request_await_object - set this request to (async) wait upon a bo
1587  * @to: request we are wishing to use
1588  * @obj: object which may be in use on another ring.
1589  * @write: whether the wait is on behalf of a writer
1590  *
1591  * This code is meant to abstract object synchronization with the GPU.
1592  * Conceptually we serialise writes between engines inside the GPU.
1593  * We only allow one engine to write into a buffer at any time, but
1594  * multiple readers. To ensure each has a coherent view of memory, we must:
1595  *
1596  * - If there is an outstanding write request to the object, the new
1597  *   request must wait for it to complete (either CPU or in hw, requests
1598  *   on the same ring will be naturally ordered).
1599  *
1600  * - If we are a write request (pending_write_domain is set), the new
1601  *   request must wait for outstanding read requests to complete.
1602  *
1603  * Returns 0 if successful, else propagates up the lower layer error.
1604  */
1605 int
i915_request_await_object(struct i915_request * to,struct drm_i915_gem_object * obj,bool write)1606 i915_request_await_object(struct i915_request *to,
1607 			  struct drm_i915_gem_object *obj,
1608 			  bool write)
1609 {
1610 	struct dma_resv_iter cursor;
1611 	struct dma_fence *fence;
1612 	int ret = 0;
1613 
1614 	dma_resv_for_each_fence(&cursor, obj->base.resv,
1615 				dma_resv_usage_rw(write), fence) {
1616 		ret = i915_request_await_dma_fence(to, fence);
1617 		if (ret)
1618 			break;
1619 	}
1620 
1621 	return ret;
1622 }
1623 
1624 static struct i915_request *
__i915_request_ensure_parallel_ordering(struct i915_request * rq,struct intel_timeline * timeline)1625 __i915_request_ensure_parallel_ordering(struct i915_request *rq,
1626 					struct intel_timeline *timeline)
1627 {
1628 	struct i915_request *prev;
1629 
1630 	GEM_BUG_ON(!is_parallel_rq(rq));
1631 
1632 	prev = request_to_parent(rq)->parallel.last_rq;
1633 	if (prev) {
1634 		if (!__i915_request_is_complete(prev)) {
1635 			i915_sw_fence_await_sw_fence(&rq->submit,
1636 						     &prev->submit,
1637 						     &rq->submitq);
1638 
1639 			if (rq->engine->sched_engine->schedule)
1640 				__i915_sched_node_add_dependency(&rq->sched,
1641 								 &prev->sched,
1642 								 &rq->dep,
1643 								 0);
1644 		}
1645 		i915_request_put(prev);
1646 	}
1647 
1648 	request_to_parent(rq)->parallel.last_rq = i915_request_get(rq);
1649 
1650 	return to_request(__i915_active_fence_set(&timeline->last_request,
1651 						  &rq->fence));
1652 }
1653 
1654 static struct i915_request *
__i915_request_ensure_ordering(struct i915_request * rq,struct intel_timeline * timeline)1655 __i915_request_ensure_ordering(struct i915_request *rq,
1656 			       struct intel_timeline *timeline)
1657 {
1658 	struct i915_request *prev;
1659 
1660 	GEM_BUG_ON(is_parallel_rq(rq));
1661 
1662 	prev = to_request(__i915_active_fence_set(&timeline->last_request,
1663 						  &rq->fence));
1664 
1665 	if (prev && !__i915_request_is_complete(prev)) {
1666 		bool uses_guc = intel_engine_uses_guc(rq->engine);
1667 		bool pow2 = is_power_of_2(READ_ONCE(prev->engine)->mask |
1668 					  rq->engine->mask);
1669 		bool same_context = prev->context == rq->context;
1670 
1671 		/*
1672 		 * The requests are supposed to be kept in order. However,
1673 		 * we need to be wary in case the timeline->last_request
1674 		 * is used as a barrier for external modification to this
1675 		 * context.
1676 		 */
1677 		GEM_BUG_ON(same_context &&
1678 			   i915_seqno_passed(prev->fence.seqno,
1679 					     rq->fence.seqno));
1680 
1681 		if ((same_context && uses_guc) || (!uses_guc && pow2))
1682 			i915_sw_fence_await_sw_fence(&rq->submit,
1683 						     &prev->submit,
1684 						     &rq->submitq);
1685 		else
1686 			__i915_sw_fence_await_dma_fence(&rq->submit,
1687 							&prev->fence,
1688 							&rq->dmaq);
1689 		if (rq->engine->sched_engine->schedule)
1690 			__i915_sched_node_add_dependency(&rq->sched,
1691 							 &prev->sched,
1692 							 &rq->dep,
1693 							 0);
1694 	}
1695 
1696 	return prev;
1697 }
1698 
1699 static struct i915_request *
__i915_request_add_to_timeline(struct i915_request * rq)1700 __i915_request_add_to_timeline(struct i915_request *rq)
1701 {
1702 	struct intel_timeline *timeline = i915_request_timeline(rq);
1703 	struct i915_request *prev;
1704 
1705 	/*
1706 	 * Dependency tracking and request ordering along the timeline
1707 	 * is special cased so that we can eliminate redundant ordering
1708 	 * operations while building the request (we know that the timeline
1709 	 * itself is ordered, and here we guarantee it).
1710 	 *
1711 	 * As we know we will need to emit tracking along the timeline,
1712 	 * we embed the hooks into our request struct -- at the cost of
1713 	 * having to have specialised no-allocation interfaces (which will
1714 	 * be beneficial elsewhere).
1715 	 *
1716 	 * A second benefit to open-coding i915_request_await_request is
1717 	 * that we can apply a slight variant of the rules specialised
1718 	 * for timelines that jump between engines (such as virtual engines).
1719 	 * If we consider the case of virtual engine, we must emit a dma-fence
1720 	 * to prevent scheduling of the second request until the first is
1721 	 * complete (to maximise our greedy late load balancing) and this
1722 	 * precludes optimising to use semaphores serialisation of a single
1723 	 * timeline across engines.
1724 	 *
1725 	 * We do not order parallel submission requests on the timeline as each
1726 	 * parallel submission context has its own timeline and the ordering
1727 	 * rules for parallel requests are that they must be submitted in the
1728 	 * order received from the execbuf IOCTL. So rather than using the
1729 	 * timeline we store a pointer to last request submitted in the
1730 	 * relationship in the gem context and insert a submission fence
1731 	 * between that request and request passed into this function or
1732 	 * alternatively we use completion fence if gem context has a single
1733 	 * timeline and this is the first submission of an execbuf IOCTL.
1734 	 */
1735 	if (likely(!is_parallel_rq(rq)))
1736 		prev = __i915_request_ensure_ordering(rq, timeline);
1737 	else
1738 		prev = __i915_request_ensure_parallel_ordering(rq, timeline);
1739 
1740 	/*
1741 	 * Make sure that no request gazumped us - if it was allocated after
1742 	 * our i915_request_alloc() and called __i915_request_add() before
1743 	 * us, the timeline will hold its seqno which is later than ours.
1744 	 */
1745 	GEM_BUG_ON(timeline->seqno != rq->fence.seqno);
1746 
1747 	return prev;
1748 }
1749 
1750 /*
1751  * NB: This function is not allowed to fail. Doing so would mean the the
1752  * request is not being tracked for completion but the work itself is
1753  * going to happen on the hardware. This would be a Bad Thing(tm).
1754  */
__i915_request_commit(struct i915_request * rq)1755 struct i915_request *__i915_request_commit(struct i915_request *rq)
1756 {
1757 	struct intel_engine_cs *engine = rq->engine;
1758 	struct intel_ring *ring = rq->ring;
1759 	u32 *cs;
1760 
1761 	RQ_TRACE(rq, "\n");
1762 
1763 	/*
1764 	 * To ensure that this call will not fail, space for its emissions
1765 	 * should already have been reserved in the ring buffer. Let the ring
1766 	 * know that it is time to use that space up.
1767 	 */
1768 	GEM_BUG_ON(rq->reserved_space > ring->space);
1769 	rq->reserved_space = 0;
1770 	rq->emitted_jiffies = jiffies;
1771 
1772 	/*
1773 	 * Record the position of the start of the breadcrumb so that
1774 	 * should we detect the updated seqno part-way through the
1775 	 * GPU processing the request, we never over-estimate the
1776 	 * position of the ring's HEAD.
1777 	 */
1778 	cs = intel_ring_begin(rq, engine->emit_fini_breadcrumb_dw);
1779 	GEM_BUG_ON(IS_ERR(cs));
1780 	rq->postfix = intel_ring_offset(rq, cs);
1781 
1782 	return __i915_request_add_to_timeline(rq);
1783 }
1784 
__i915_request_queue_bh(struct i915_request * rq)1785 void __i915_request_queue_bh(struct i915_request *rq)
1786 {
1787 	i915_sw_fence_commit(&rq->semaphore);
1788 	i915_sw_fence_commit(&rq->submit);
1789 }
1790 
__i915_request_queue(struct i915_request * rq,const struct i915_sched_attr * attr)1791 void __i915_request_queue(struct i915_request *rq,
1792 			  const struct i915_sched_attr *attr)
1793 {
1794 	/*
1795 	 * Let the backend know a new request has arrived that may need
1796 	 * to adjust the existing execution schedule due to a high priority
1797 	 * request - i.e. we may want to preempt the current request in order
1798 	 * to run a high priority dependency chain *before* we can execute this
1799 	 * request.
1800 	 *
1801 	 * This is called before the request is ready to run so that we can
1802 	 * decide whether to preempt the entire chain so that it is ready to
1803 	 * run at the earliest possible convenience.
1804 	 */
1805 	if (attr && rq->engine->sched_engine->schedule)
1806 		rq->engine->sched_engine->schedule(rq, attr);
1807 
1808 	local_bh_disable();
1809 	__i915_request_queue_bh(rq);
1810 	local_bh_enable(); /* kick tasklets */
1811 }
1812 
i915_request_add(struct i915_request * rq)1813 void i915_request_add(struct i915_request *rq)
1814 {
1815 	struct intel_timeline * const tl = i915_request_timeline(rq);
1816 	struct i915_sched_attr attr = {};
1817 	struct i915_gem_context *ctx;
1818 
1819 	lockdep_assert_held(&tl->mutex);
1820 	lockdep_unpin_lock(&tl->mutex, rq->cookie);
1821 
1822 	trace_i915_request_add(rq);
1823 	__i915_request_commit(rq);
1824 
1825 	/* XXX placeholder for selftests */
1826 	rcu_read_lock();
1827 	ctx = rcu_dereference(rq->context->gem_context);
1828 	if (ctx)
1829 		attr = ctx->sched;
1830 	rcu_read_unlock();
1831 
1832 	__i915_request_queue(rq, &attr);
1833 
1834 	mutex_unlock(&tl->mutex);
1835 }
1836 
local_clock_ns(unsigned int * cpu)1837 static unsigned long local_clock_ns(unsigned int *cpu)
1838 {
1839 	unsigned long t;
1840 
1841 	/*
1842 	 * Cheaply and approximately convert from nanoseconds to microseconds.
1843 	 * The result and subsequent calculations are also defined in the same
1844 	 * approximate microseconds units. The principal source of timing
1845 	 * error here is from the simple truncation.
1846 	 *
1847 	 * Note that local_clock() is only defined wrt to the current CPU;
1848 	 * the comparisons are no longer valid if we switch CPUs. Instead of
1849 	 * blocking preemption for the entire busywait, we can detect the CPU
1850 	 * switch and use that as indicator of system load and a reason to
1851 	 * stop busywaiting, see busywait_stop().
1852 	 */
1853 	*cpu = get_cpu();
1854 	t = local_clock();
1855 	put_cpu();
1856 
1857 	return t;
1858 }
1859 
busywait_stop(unsigned long timeout,unsigned int cpu)1860 static bool busywait_stop(unsigned long timeout, unsigned int cpu)
1861 {
1862 	unsigned int this_cpu;
1863 
1864 	if (time_after(local_clock_ns(&this_cpu), timeout))
1865 		return true;
1866 
1867 	return this_cpu != cpu;
1868 }
1869 
__i915_spin_request(struct i915_request * const rq,int state)1870 static bool __i915_spin_request(struct i915_request * const rq, int state)
1871 {
1872 	unsigned long timeout_ns;
1873 	unsigned int cpu;
1874 
1875 	/*
1876 	 * Only wait for the request if we know it is likely to complete.
1877 	 *
1878 	 * We don't track the timestamps around requests, nor the average
1879 	 * request length, so we do not have a good indicator that this
1880 	 * request will complete within the timeout. What we do know is the
1881 	 * order in which requests are executed by the context and so we can
1882 	 * tell if the request has been started. If the request is not even
1883 	 * running yet, it is a fair assumption that it will not complete
1884 	 * within our relatively short timeout.
1885 	 */
1886 	if (!i915_request_is_running(rq))
1887 		return false;
1888 
1889 	/*
1890 	 * When waiting for high frequency requests, e.g. during synchronous
1891 	 * rendering split between the CPU and GPU, the finite amount of time
1892 	 * required to set up the irq and wait upon it limits the response
1893 	 * rate. By busywaiting on the request completion for a short while we
1894 	 * can service the high frequency waits as quick as possible. However,
1895 	 * if it is a slow request, we want to sleep as quickly as possible.
1896 	 * The tradeoff between waiting and sleeping is roughly the time it
1897 	 * takes to sleep on a request, on the order of a microsecond.
1898 	 */
1899 
1900 	timeout_ns = READ_ONCE(rq->engine->props.max_busywait_duration_ns);
1901 	timeout_ns += local_clock_ns(&cpu);
1902 	do {
1903 		if (dma_fence_is_signaled(&rq->fence))
1904 			return true;
1905 
1906 		if (signal_pending_state(state, current))
1907 			break;
1908 
1909 		if (busywait_stop(timeout_ns, cpu))
1910 			break;
1911 
1912 		cpu_relax();
1913 	} while (!need_resched());
1914 
1915 	return false;
1916 }
1917 
1918 struct request_wait {
1919 	struct dma_fence_cb cb;
1920 	struct task_struct *tsk;
1921 };
1922 
request_wait_wake(struct dma_fence * fence,struct dma_fence_cb * cb)1923 static void request_wait_wake(struct dma_fence *fence, struct dma_fence_cb *cb)
1924 {
1925 	struct request_wait *wait = container_of(cb, typeof(*wait), cb);
1926 
1927 	wake_up_process(fetch_and_zero(&wait->tsk));
1928 }
1929 
1930 /**
1931  * i915_request_wait_timeout - wait until execution of request has finished
1932  * @rq: the request to wait upon
1933  * @flags: how to wait
1934  * @timeout: how long to wait in jiffies
1935  *
1936  * i915_request_wait_timeout() waits for the request to be completed, for a
1937  * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an
1938  * unbounded wait).
1939  *
1940  * Returns the remaining time (in jiffies) if the request completed, which may
1941  * be zero if the request is unfinished after the timeout expires.
1942  * If the timeout is 0, it will return 1 if the fence is signaled.
1943  *
1944  * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is
1945  * pending before the request completes.
1946  *
1947  * NOTE: This function has the same wait semantics as dma-fence.
1948  */
i915_request_wait_timeout(struct i915_request * rq,unsigned int flags,long timeout)1949 long i915_request_wait_timeout(struct i915_request *rq,
1950 			       unsigned int flags,
1951 			       long timeout)
1952 {
1953 	const int state = flags & I915_WAIT_INTERRUPTIBLE ?
1954 		TASK_INTERRUPTIBLE : TASK_UNINTERRUPTIBLE;
1955 	struct request_wait wait;
1956 
1957 	might_sleep();
1958 	GEM_BUG_ON(timeout < 0);
1959 
1960 	if (dma_fence_is_signaled(&rq->fence))
1961 		return timeout ?: 1;
1962 
1963 	if (!timeout)
1964 		return -ETIME;
1965 
1966 	trace_i915_request_wait_begin(rq, flags);
1967 
1968 	/*
1969 	 * We must never wait on the GPU while holding a lock as we
1970 	 * may need to perform a GPU reset. So while we don't need to
1971 	 * serialise wait/reset with an explicit lock, we do want
1972 	 * lockdep to detect potential dependency cycles.
1973 	 */
1974 	mutex_acquire(&rq->engine->gt->reset.mutex.dep_map, 0, 0, _THIS_IP_);
1975 
1976 	/*
1977 	 * Optimistic spin before touching IRQs.
1978 	 *
1979 	 * We may use a rather large value here to offset the penalty of
1980 	 * switching away from the active task. Frequently, the client will
1981 	 * wait upon an old swapbuffer to throttle itself to remain within a
1982 	 * frame of the gpu. If the client is running in lockstep with the gpu,
1983 	 * then it should not be waiting long at all, and a sleep now will incur
1984 	 * extra scheduler latency in producing the next frame. To try to
1985 	 * avoid adding the cost of enabling/disabling the interrupt to the
1986 	 * short wait, we first spin to see if the request would have completed
1987 	 * in the time taken to setup the interrupt.
1988 	 *
1989 	 * We need upto 5us to enable the irq, and upto 20us to hide the
1990 	 * scheduler latency of a context switch, ignoring the secondary
1991 	 * impacts from a context switch such as cache eviction.
1992 	 *
1993 	 * The scheme used for low-latency IO is called "hybrid interrupt
1994 	 * polling". The suggestion there is to sleep until just before you
1995 	 * expect to be woken by the device interrupt and then poll for its
1996 	 * completion. That requires having a good predictor for the request
1997 	 * duration, which we currently lack.
1998 	 */
1999 	if (CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT &&
2000 	    __i915_spin_request(rq, state))
2001 		goto out;
2002 
2003 	/*
2004 	 * This client is about to stall waiting for the GPU. In many cases
2005 	 * this is undesirable and limits the throughput of the system, as
2006 	 * many clients cannot continue processing user input/output whilst
2007 	 * blocked. RPS autotuning may take tens of milliseconds to respond
2008 	 * to the GPU load and thus incurs additional latency for the client.
2009 	 * We can circumvent that by promoting the GPU frequency to maximum
2010 	 * before we sleep. This makes the GPU throttle up much more quickly
2011 	 * (good for benchmarks and user experience, e.g. window animations),
2012 	 * but at a cost of spending more power processing the workload
2013 	 * (bad for battery).
2014 	 */
2015 	if (flags & I915_WAIT_PRIORITY && !i915_request_started(rq))
2016 		intel_rps_boost(rq);
2017 
2018 	wait.tsk = current;
2019 	if (dma_fence_add_callback(&rq->fence, &wait.cb, request_wait_wake))
2020 		goto out;
2021 
2022 	/*
2023 	 * Flush the submission tasklet, but only if it may help this request.
2024 	 *
2025 	 * We sometimes experience some latency between the HW interrupts and
2026 	 * tasklet execution (mostly due to ksoftirqd latency, but it can also
2027 	 * be due to lazy CS events), so lets run the tasklet manually if there
2028 	 * is a chance it may submit this request. If the request is not ready
2029 	 * to run, as it is waiting for other fences to be signaled, flushing
2030 	 * the tasklet is busy work without any advantage for this client.
2031 	 *
2032 	 * If the HW is being lazy, this is the last chance before we go to
2033 	 * sleep to catch any pending events. We will check periodically in
2034 	 * the heartbeat to flush the submission tasklets as a last resort
2035 	 * for unhappy HW.
2036 	 */
2037 	if (i915_request_is_ready(rq))
2038 		__intel_engine_flush_submission(rq->engine, false);
2039 
2040 	for (;;) {
2041 		set_current_state(state);
2042 
2043 		if (dma_fence_is_signaled(&rq->fence))
2044 			break;
2045 
2046 		if (signal_pending_state(state, current)) {
2047 			timeout = -ERESTARTSYS;
2048 			break;
2049 		}
2050 
2051 		if (!timeout) {
2052 			timeout = -ETIME;
2053 			break;
2054 		}
2055 
2056 		timeout = io_schedule_timeout(timeout);
2057 	}
2058 	__set_current_state(TASK_RUNNING);
2059 
2060 	if (READ_ONCE(wait.tsk))
2061 		dma_fence_remove_callback(&rq->fence, &wait.cb);
2062 	GEM_BUG_ON(!list_empty(&wait.cb.node));
2063 
2064 out:
2065 	mutex_release(&rq->engine->gt->reset.mutex.dep_map, _THIS_IP_);
2066 	trace_i915_request_wait_end(rq);
2067 	return timeout;
2068 }
2069 
2070 /**
2071  * i915_request_wait - wait until execution of request has finished
2072  * @rq: the request to wait upon
2073  * @flags: how to wait
2074  * @timeout: how long to wait in jiffies
2075  *
2076  * i915_request_wait() waits for the request to be completed, for a
2077  * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an
2078  * unbounded wait).
2079  *
2080  * Returns the remaining time (in jiffies) if the request completed, which may
2081  * be zero or -ETIME if the request is unfinished after the timeout expires.
2082  * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is
2083  * pending before the request completes.
2084  *
2085  * NOTE: This function behaves differently from dma-fence wait semantics for
2086  * timeout = 0. It returns 0 on success, and -ETIME if not signaled.
2087  */
i915_request_wait(struct i915_request * rq,unsigned int flags,long timeout)2088 long i915_request_wait(struct i915_request *rq,
2089 		       unsigned int flags,
2090 		       long timeout)
2091 {
2092 	long ret = i915_request_wait_timeout(rq, flags, timeout);
2093 
2094 	if (!ret)
2095 		return -ETIME;
2096 
2097 	if (ret > 0 && !timeout)
2098 		return 0;
2099 
2100 	return ret;
2101 }
2102 
print_sched_attr(const struct i915_sched_attr * attr,char * buf,int x,int len)2103 static int print_sched_attr(const struct i915_sched_attr *attr,
2104 			    char *buf, int x, int len)
2105 {
2106 	if (attr->priority == I915_PRIORITY_INVALID)
2107 		return x;
2108 
2109 	x += snprintf(buf + x, len - x,
2110 		      " prio=%d", attr->priority);
2111 
2112 	return x;
2113 }
2114 
queue_status(const struct i915_request * rq)2115 static char queue_status(const struct i915_request *rq)
2116 {
2117 	if (i915_request_is_active(rq))
2118 		return 'E';
2119 
2120 	if (i915_request_is_ready(rq))
2121 		return intel_engine_is_virtual(rq->engine) ? 'V' : 'R';
2122 
2123 	return 'U';
2124 }
2125 
run_status(const struct i915_request * rq)2126 static const char *run_status(const struct i915_request *rq)
2127 {
2128 	if (__i915_request_is_complete(rq))
2129 		return "!";
2130 
2131 	if (__i915_request_has_started(rq))
2132 		return "*";
2133 
2134 	if (!i915_sw_fence_signaled(&rq->semaphore))
2135 		return "&";
2136 
2137 	return "";
2138 }
2139 
fence_status(const struct i915_request * rq)2140 static const char *fence_status(const struct i915_request *rq)
2141 {
2142 	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags))
2143 		return "+";
2144 
2145 	if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &rq->fence.flags))
2146 		return "-";
2147 
2148 	return "";
2149 }
2150 
i915_request_show(struct drm_printer * m,const struct i915_request * rq,const char * prefix,int indent)2151 void i915_request_show(struct drm_printer *m,
2152 		       const struct i915_request *rq,
2153 		       const char *prefix,
2154 		       int indent)
2155 {
2156 	const char *name = rq->fence.ops->get_timeline_name((struct dma_fence *)&rq->fence);
2157 	char buf[80] = "";
2158 	int x = 0;
2159 
2160 	/*
2161 	 * The prefix is used to show the queue status, for which we use
2162 	 * the following flags:
2163 	 *
2164 	 *  U [Unready]
2165 	 *    - initial status upon being submitted by the user
2166 	 *
2167 	 *    - the request is not ready for execution as it is waiting
2168 	 *      for external fences
2169 	 *
2170 	 *  R [Ready]
2171 	 *    - all fences the request was waiting on have been signaled,
2172 	 *      and the request is now ready for execution and will be
2173 	 *      in a backend queue
2174 	 *
2175 	 *    - a ready request may still need to wait on semaphores
2176 	 *      [internal fences]
2177 	 *
2178 	 *  V [Ready/virtual]
2179 	 *    - same as ready, but queued over multiple backends
2180 	 *
2181 	 *  E [Executing]
2182 	 *    - the request has been transferred from the backend queue and
2183 	 *      submitted for execution on HW
2184 	 *
2185 	 *    - a completed request may still be regarded as executing, its
2186 	 *      status may not be updated until it is retired and removed
2187 	 *      from the lists
2188 	 */
2189 
2190 	x = print_sched_attr(&rq->sched.attr, buf, x, sizeof(buf));
2191 
2192 	drm_printf(m, "%s%.*s%c %llx:%lld%s%s %s @ %dms: %s\n",
2193 		   prefix, indent, "                ",
2194 		   queue_status(rq),
2195 		   rq->fence.context, rq->fence.seqno,
2196 		   run_status(rq),
2197 		   fence_status(rq),
2198 		   buf,
2199 		   jiffies_to_msecs(jiffies - rq->emitted_jiffies),
2200 		   name);
2201 }
2202 
engine_match_ring(struct intel_engine_cs * engine,struct i915_request * rq)2203 static bool engine_match_ring(struct intel_engine_cs *engine, struct i915_request *rq)
2204 {
2205 	u32 ring = ENGINE_READ(engine, RING_START);
2206 
2207 	return ring == i915_ggtt_offset(rq->ring->vma);
2208 }
2209 
match_ring(struct i915_request * rq)2210 static bool match_ring(struct i915_request *rq)
2211 {
2212 	struct intel_engine_cs *engine;
2213 	bool found;
2214 	int i;
2215 
2216 	if (!intel_engine_is_virtual(rq->engine))
2217 		return engine_match_ring(rq->engine, rq);
2218 
2219 	found = false;
2220 	i = 0;
2221 	while ((engine = intel_engine_get_sibling(rq->engine, i++))) {
2222 		found = engine_match_ring(engine, rq);
2223 		if (found)
2224 			break;
2225 	}
2226 
2227 	return found;
2228 }
2229 
i915_test_request_state(struct i915_request * rq)2230 enum i915_request_state i915_test_request_state(struct i915_request *rq)
2231 {
2232 	if (i915_request_completed(rq))
2233 		return I915_REQUEST_COMPLETE;
2234 
2235 	if (!i915_request_started(rq))
2236 		return I915_REQUEST_PENDING;
2237 
2238 	if (match_ring(rq))
2239 		return I915_REQUEST_ACTIVE;
2240 
2241 	return I915_REQUEST_QUEUED;
2242 }
2243 
2244 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2245 #include "selftests/mock_request.c"
2246 #include "selftests/i915_request.c"
2247 #endif
2248 
i915_request_module_exit(void)2249 void i915_request_module_exit(void)
2250 {
2251 	kmem_cache_destroy(slab_execute_cbs);
2252 	kmem_cache_destroy(slab_requests);
2253 }
2254 
i915_request_module_init(void)2255 int __init i915_request_module_init(void)
2256 {
2257 	slab_requests =
2258 		kmem_cache_create("i915_request",
2259 				  sizeof(struct i915_request),
2260 				  __alignof__(struct i915_request),
2261 				  SLAB_HWCACHE_ALIGN |
2262 				  SLAB_RECLAIM_ACCOUNT |
2263 				  SLAB_TYPESAFE_BY_RCU,
2264 				  __i915_request_ctor);
2265 	if (!slab_requests)
2266 		return -ENOMEM;
2267 
2268 	slab_execute_cbs = KMEM_CACHE(execute_cb,
2269 					     SLAB_HWCACHE_ALIGN |
2270 					     SLAB_RECLAIM_ACCOUNT |
2271 					     SLAB_TYPESAFE_BY_RCU);
2272 	if (!slab_execute_cbs)
2273 		goto err_requests;
2274 
2275 	return 0;
2276 
2277 err_requests:
2278 	kmem_cache_destroy(slab_requests);
2279 	return -ENOMEM;
2280 }
2281