1// SPDX-License-Identifier: MIT
  2/*
  3 * Copyright © 2019 Intel Corporation
  4 */
  5
  6#include <linux/workqueue.h>
  7
  8#include "i915_drv.h" /* for_each_engine() */
  9#include "i915_request.h"
 10#include "intel_engine_heartbeat.h"
 11#include "intel_execlists_submission.h"
 12#include "intel_gt.h"
 13#include "intel_gt_pm.h"
 14#include "intel_gt_requests.h"
 15#include "intel_timeline.h"
 16
 17static bool retire_requests(struct intel_timeline *tl)
 18{
 19	struct i915_request *rq, *rn;
 20
 21	list_for_each_entry_safe(rq, rn, &tl->requests, link)
 22		if (!i915_request_retire(rq))
 23			return false;
 24
 25	/* And check nothing new was submitted */
 26	return !i915_active_fence_isset(&tl->last_request);
 27}
 28
 29static bool engine_active(const struct intel_engine_cs *engine)
 30{
 31	return !list_empty(&engine->kernel_context->timeline->requests);
 32}
 33
 34static bool flush_submission(struct intel_gt *gt, long timeout)
 35{
 36	struct intel_engine_cs *engine;
 37	enum intel_engine_id id;
 38	bool active = false;
 39
 40	if (!timeout)
 41		return false;
 42
 43	if (!intel_gt_pm_is_awake(gt))
 44		return false;
 45
 46	for_each_engine(engine, gt, id) {
 47		intel_engine_flush_submission(engine);
 48
 49		/* Flush the background retirement and idle barriers */
 50		flush_work(&engine->retire_work);
 51		flush_delayed_work(&engine->wakeref.work);
 52
 53		/* Is the idle barrier still outstanding? */
 54		active |= engine_active(engine);
 55	}
 56
 57	return active;
 58}
 59
 60static void engine_retire(struct work_struct *work)
 61{
 62	struct intel_engine_cs *engine =
 63		container_of(work, typeof(*engine), retire_work);
 64	struct intel_timeline *tl = xchg(&engine->retire, NULL);
 65
 66	do {
 67		struct intel_timeline *next = xchg(&tl->retire, NULL);
 68
 69		/*
 70		 * Our goal here is to retire _idle_ timelines as soon as
 71		 * possible (as they are idle, we do not expect userspace
 72		 * to be cleaning up anytime soon).
 73		 *
 74		 * If the timeline is currently locked, either it is being
 75		 * retired elsewhere or about to be!
 76		 */
 77		if (mutex_trylock(&tl->mutex)) {
 78			retire_requests(tl);
 79			mutex_unlock(&tl->mutex);
 80		}
 81		intel_timeline_put(tl);
 82
 83		GEM_BUG_ON(!next);
 84		tl = ptr_mask_bits(next, 1);
 85	} while (tl);
 86}
 87
 88static bool add_retire(struct intel_engine_cs *engine,
 89		       struct intel_timeline *tl)
 90{
 91#define STUB ((struct intel_timeline *)1)
 92	struct intel_timeline *first;
 93
 94	/*
 95	 * We open-code a llist here to include the additional tag [BIT(0)]
 96	 * so that we know when the timeline is already on a
 97	 * retirement queue: either this engine or another.
 98	 */
 99
100	if (cmpxchg(&tl->retire, NULL, STUB)) /* already queued */
101		return false;
102
103	intel_timeline_get(tl);
104	first = READ_ONCE(engine->retire);
105	do
106		tl->retire = ptr_pack_bits(first, 1, 1);
107	while (!try_cmpxchg(&engine->retire, &first, tl));
108
109	return !first;
110}
111
112void intel_engine_add_retire(struct intel_engine_cs *engine,
113			     struct intel_timeline *tl)
114{
115	/* We don't deal well with the engine disappearing beneath us */
116	GEM_BUG_ON(intel_engine_is_virtual(engine));
117
118	if (add_retire(engine, tl))
119		schedule_work(&engine->retire_work);
120}
121
122void intel_engine_init_retire(struct intel_engine_cs *engine)
123{
124	INIT_WORK(&engine->retire_work, engine_retire);
125}
126
127void intel_engine_fini_retire(struct intel_engine_cs *engine)
128{
129	flush_work(&engine->retire_work);
130	GEM_BUG_ON(engine->retire);
131}
132
133long intel_gt_retire_requests_timeout(struct intel_gt *gt, long timeout,
134				      long *remaining_timeout)
135{
136	struct intel_gt_timelines *timelines = &gt->timelines;
137	struct intel_timeline *tl, *tn;
138	unsigned long active_count = 0;
139	LIST_HEAD(free);
140
141	flush_submission(gt, timeout); /* kick the ksoftirqd tasklets */
142	spin_lock(&timelines->lock);
143	list_for_each_entry_safe(tl, tn, &timelines->active_list, link) {
144		if (!mutex_trylock(&tl->mutex)) {
145			active_count++; /* report busy to caller, try again? */
146			continue;
147		}
148
149		intel_timeline_get(tl);
150		GEM_BUG_ON(!atomic_read(&tl->active_count));
151		atomic_inc(&tl->active_count); /* pin the list element */
152		spin_unlock(&timelines->lock);
153
154		if (timeout > 0) {
155			struct dma_fence *fence;
156
157			fence = i915_active_fence_get(&tl->last_request);
158			if (fence) {
159				mutex_unlock(&tl->mutex);
160
161				timeout = dma_fence_wait_timeout(fence,
162								 true,
163								 timeout);
164				dma_fence_put(fence);
165
166				/* Retirement is best effort */
167				if (!mutex_trylock(&tl->mutex)) {
168					active_count++;
169					goto out_active;
170				}
171			}
172		}
173
174		if (!retire_requests(tl))
175			active_count++;
176		mutex_unlock(&tl->mutex);
177
178out_active:	spin_lock(&timelines->lock);
179
180		/* Resume list iteration after reacquiring spinlock */
181		list_safe_reset_next(tl, tn, link);
182		if (atomic_dec_and_test(&tl->active_count))
183			list_del(&tl->link);
184
185		/* Defer the final release to after the spinlock */
186		if (refcount_dec_and_test(&tl->kref.refcount)) {
187			GEM_BUG_ON(atomic_read(&tl->active_count));
188			list_add(&tl->link, &free);
189		}
190	}
191	spin_unlock(&timelines->lock);
192
193	list_for_each_entry_safe(tl, tn, &free, link)
194		__intel_timeline_free(&tl->kref);
195
196	if (flush_submission(gt, timeout)) /* Wait, there's more! */
197		active_count++;
198
199	if (remaining_timeout)
200		*remaining_timeout = timeout;
201
202	return active_count ? timeout ?: -ETIME : 0;
203}
204
205static void retire_work_handler(struct work_struct *work)
206{
207	struct intel_gt *gt =
208		container_of(work, typeof(*gt), requests.retire_work.work);
209
210	schedule_delayed_work(&gt->requests.retire_work,
211			      round_jiffies_up_relative(HZ));
212	intel_gt_retire_requests(gt);
213}
214
215void intel_gt_init_requests(struct intel_gt *gt)
216{
217	INIT_DELAYED_WORK(&gt->requests.retire_work, retire_work_handler);
218}
219
220void intel_gt_park_requests(struct intel_gt *gt)
221{
222	cancel_delayed_work(&gt->requests.retire_work);
223}
224
225void intel_gt_unpark_requests(struct intel_gt *gt)
226{
227	schedule_delayed_work(&gt->requests.retire_work,
228			      round_jiffies_up_relative(HZ));
229}
230
231void intel_gt_fini_requests(struct intel_gt *gt)
232{
233	/* Wait until the work is marked as finished before unloading! */
234	cancel_delayed_work_sync(&gt->requests.retire_work);
235
236	flush_work(&gt->watchdog.work);
237}
238
239void intel_gt_watchdog_work(struct work_struct *work)
240{
241	struct intel_gt *gt =
242		container_of(work, typeof(*gt), watchdog.work);
243	struct i915_request *rq, *rn;
244	struct llist_node *first;
245
246	first = llist_del_all(&gt->watchdog.list);
247	if (!first)
248		return;
249
250	llist_for_each_entry_safe(rq, rn, first, watchdog.link) {
251		if (!i915_request_completed(rq)) {
252			struct dma_fence *f = &rq->fence;
253
254			pr_notice("Fence expiration time out i915-%s:%s:%llx!\n",
255				  f->ops->get_driver_name(f),
256				  f->ops->get_timeline_name(f),
257				  f->seqno);
258			i915_request_cancel(rq, -EINTR);
259		}
260		i915_request_put(rq);
261	}
262}