1 /*
2  * SPDX-License-Identifier: MIT
3  *
4  * Copyright © 2019 Intel Corporation
5  */
6 
7 #include <linux/workqueue.h>
8 
9 #include "i915_drv.h" /* for_each_engine() */
10 #include "i915_request.h"
11 #include "intel_engine_heartbeat.h"
12 #include "intel_gt.h"
13 #include "intel_gt_pm.h"
14 #include "intel_gt_requests.h"
15 #include "intel_timeline.h"
16 
17 static 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 
29 static bool flush_submission(struct intel_gt *gt)
30 {
31 	struct intel_engine_cs *engine;
32 	enum intel_engine_id id;
33 	bool active = false;
34 
35 	if (!intel_gt_pm_is_awake(gt))
36 		return false;
37 
38 	for_each_engine(engine, gt, id) {
39 		intel_engine_flush_submission(engine);
40 		active |= flush_work(&engine->retire_work);
41 		active |= flush_work(&engine->wakeref.work);
42 	}
43 
44 	return active;
45 }
46 
47 static void engine_retire(struct work_struct *work)
48 {
49 	struct intel_engine_cs *engine =
50 		container_of(work, typeof(*engine), retire_work);
51 	struct intel_timeline *tl = xchg(&engine->retire, NULL);
52 
53 	do {
54 		struct intel_timeline *next = xchg(&tl->retire, NULL);
55 
56 		/*
57 		 * Our goal here is to retire _idle_ timelines as soon as
58 		 * possible (as they are idle, we do not expect userspace
59 		 * to be cleaning up anytime soon).
60 		 *
61 		 * If the timeline is currently locked, either it is being
62 		 * retired elsewhere or about to be!
63 		 */
64 		if (mutex_trylock(&tl->mutex)) {
65 			retire_requests(tl);
66 			mutex_unlock(&tl->mutex);
67 		}
68 		intel_timeline_put(tl);
69 
70 		GEM_BUG_ON(!next);
71 		tl = ptr_mask_bits(next, 1);
72 	} while (tl);
73 }
74 
75 static bool add_retire(struct intel_engine_cs *engine,
76 		       struct intel_timeline *tl)
77 {
78 #define STUB ((struct intel_timeline *)1)
79 	struct intel_timeline *first;
80 
81 	/*
82 	 * We open-code a llist here to include the additional tag [BIT(0)]
83 	 * so that we know when the timeline is already on a
84 	 * retirement queue: either this engine or another.
85 	 */
86 
87 	if (cmpxchg(&tl->retire, NULL, STUB)) /* already queued */
88 		return false;
89 
90 	intel_timeline_get(tl);
91 	first = READ_ONCE(engine->retire);
92 	do
93 		tl->retire = ptr_pack_bits(first, 1, 1);
94 	while (!try_cmpxchg(&engine->retire, &first, tl));
95 
96 	return !first;
97 }
98 
99 void intel_engine_add_retire(struct intel_engine_cs *engine,
100 			     struct intel_timeline *tl)
101 {
102 	if (add_retire(engine, tl))
103 		schedule_work(&engine->retire_work);
104 }
105 
106 void intel_engine_init_retire(struct intel_engine_cs *engine)
107 {
108 	INIT_WORK(&engine->retire_work, engine_retire);
109 }
110 
111 void intel_engine_fini_retire(struct intel_engine_cs *engine)
112 {
113 	flush_work(&engine->retire_work);
114 	GEM_BUG_ON(engine->retire);
115 }
116 
117 long intel_gt_retire_requests_timeout(struct intel_gt *gt, long timeout)
118 {
119 	struct intel_gt_timelines *timelines = &gt->timelines;
120 	struct intel_timeline *tl, *tn;
121 	unsigned long active_count = 0;
122 	bool interruptible;
123 	LIST_HEAD(free);
124 
125 	interruptible = true;
126 	if (unlikely(timeout < 0))
127 		timeout = -timeout, interruptible = false;
128 
129 	flush_submission(gt); /* kick the ksoftirqd tasklets */
130 	spin_lock(&timelines->lock);
131 	list_for_each_entry_safe(tl, tn, &timelines->active_list, link) {
132 		if (!mutex_trylock(&tl->mutex)) {
133 			active_count++; /* report busy to caller, try again? */
134 			continue;
135 		}
136 
137 		intel_timeline_get(tl);
138 		GEM_BUG_ON(!atomic_read(&tl->active_count));
139 		atomic_inc(&tl->active_count); /* pin the list element */
140 		spin_unlock(&timelines->lock);
141 
142 		if (timeout > 0) {
143 			struct dma_fence *fence;
144 
145 			fence = i915_active_fence_get(&tl->last_request);
146 			if (fence) {
147 				timeout = dma_fence_wait_timeout(fence,
148 								 interruptible,
149 								 timeout);
150 				dma_fence_put(fence);
151 			}
152 		}
153 
154 		if (!retire_requests(tl) || flush_submission(gt))
155 			active_count++;
156 
157 		spin_lock(&timelines->lock);
158 
159 		/* Resume iteration after dropping lock */
160 		list_safe_reset_next(tl, tn, link);
161 		if (atomic_dec_and_test(&tl->active_count))
162 			list_del(&tl->link);
163 
164 		mutex_unlock(&tl->mutex);
165 
166 		/* Defer the final release to after the spinlock */
167 		if (refcount_dec_and_test(&tl->kref.refcount)) {
168 			GEM_BUG_ON(atomic_read(&tl->active_count));
169 			list_add(&tl->link, &free);
170 		}
171 	}
172 	spin_unlock(&timelines->lock);
173 
174 	list_for_each_entry_safe(tl, tn, &free, link)
175 		__intel_timeline_free(&tl->kref);
176 
177 	return active_count ? timeout : 0;
178 }
179 
180 int intel_gt_wait_for_idle(struct intel_gt *gt, long timeout)
181 {
182 	/* If the device is asleep, we have no requests outstanding */
183 	if (!intel_gt_pm_is_awake(gt))
184 		return 0;
185 
186 	while ((timeout = intel_gt_retire_requests_timeout(gt, timeout)) > 0) {
187 		cond_resched();
188 		if (signal_pending(current))
189 			return -EINTR;
190 	}
191 
192 	return timeout;
193 }
194 
195 static void retire_work_handler(struct work_struct *work)
196 {
197 	struct intel_gt *gt =
198 		container_of(work, typeof(*gt), requests.retire_work.work);
199 
200 	schedule_delayed_work(&gt->requests.retire_work,
201 			      round_jiffies_up_relative(HZ));
202 	intel_gt_retire_requests(gt);
203 }
204 
205 void intel_gt_init_requests(struct intel_gt *gt)
206 {
207 	INIT_DELAYED_WORK(&gt->requests.retire_work, retire_work_handler);
208 }
209 
210 void intel_gt_park_requests(struct intel_gt *gt)
211 {
212 	cancel_delayed_work(&gt->requests.retire_work);
213 }
214 
215 void intel_gt_unpark_requests(struct intel_gt *gt)
216 {
217 	schedule_delayed_work(&gt->requests.retire_work,
218 			      round_jiffies_up_relative(HZ));
219 }
220 
221 void intel_gt_fini_requests(struct intel_gt *gt)
222 {
223 	/* Wait until the work is marked as finished before unloading! */
224 	cancel_delayed_work_sync(&gt->requests.retire_work);
225 }
226