1 /*
2  * Copyright © 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/kthread.h>
26 #include <trace/events/dma_fence.h>
27 #include <uapi/linux/sched/types.h>
28 
29 #include "i915_drv.h"
30 #include "i915_trace.h"
31 #include "intel_gt_pm.h"
32 #include "intel_gt_requests.h"
33 
34 static void irq_enable(struct intel_engine_cs *engine)
35 {
36 	if (!engine->irq_enable)
37 		return;
38 
39 	/* Caller disables interrupts */
40 	spin_lock(&engine->gt->irq_lock);
41 	engine->irq_enable(engine);
42 	spin_unlock(&engine->gt->irq_lock);
43 }
44 
45 static void irq_disable(struct intel_engine_cs *engine)
46 {
47 	if (!engine->irq_disable)
48 		return;
49 
50 	/* Caller disables interrupts */
51 	spin_lock(&engine->gt->irq_lock);
52 	engine->irq_disable(engine);
53 	spin_unlock(&engine->gt->irq_lock);
54 }
55 
56 static void __intel_breadcrumbs_disarm_irq(struct intel_breadcrumbs *b)
57 {
58 	struct intel_engine_cs *engine =
59 		container_of(b, struct intel_engine_cs, breadcrumbs);
60 
61 	lockdep_assert_held(&b->irq_lock);
62 
63 	GEM_BUG_ON(!b->irq_enabled);
64 	if (!--b->irq_enabled)
65 		irq_disable(engine);
66 
67 	b->irq_armed = false;
68 	intel_gt_pm_put_async(engine->gt);
69 }
70 
71 void intel_engine_disarm_breadcrumbs(struct intel_engine_cs *engine)
72 {
73 	struct intel_breadcrumbs *b = &engine->breadcrumbs;
74 	unsigned long flags;
75 
76 	if (!b->irq_armed)
77 		return;
78 
79 	spin_lock_irqsave(&b->irq_lock, flags);
80 	if (b->irq_armed)
81 		__intel_breadcrumbs_disarm_irq(b);
82 	spin_unlock_irqrestore(&b->irq_lock, flags);
83 }
84 
85 static inline bool __request_completed(const struct i915_request *rq)
86 {
87 	return i915_seqno_passed(__hwsp_seqno(rq), rq->fence.seqno);
88 }
89 
90 __maybe_unused static bool
91 check_signal_order(struct intel_context *ce, struct i915_request *rq)
92 {
93 	if (!list_is_last(&rq->signal_link, &ce->signals) &&
94 	    i915_seqno_passed(rq->fence.seqno,
95 			      list_next_entry(rq, signal_link)->fence.seqno))
96 		return false;
97 
98 	if (!list_is_first(&rq->signal_link, &ce->signals) &&
99 	    i915_seqno_passed(list_prev_entry(rq, signal_link)->fence.seqno,
100 			      rq->fence.seqno))
101 		return false;
102 
103 	return true;
104 }
105 
106 static bool
107 __dma_fence_signal(struct dma_fence *fence)
108 {
109 	return !test_and_set_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags);
110 }
111 
112 static void
113 __dma_fence_signal__timestamp(struct dma_fence *fence, ktime_t timestamp)
114 {
115 	fence->timestamp = timestamp;
116 	set_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &fence->flags);
117 	trace_dma_fence_signaled(fence);
118 }
119 
120 static void
121 __dma_fence_signal__notify(struct dma_fence *fence,
122 			   const struct list_head *list)
123 {
124 	struct dma_fence_cb *cur, *tmp;
125 
126 	lockdep_assert_held(fence->lock);
127 
128 	list_for_each_entry_safe(cur, tmp, list, node) {
129 		INIT_LIST_HEAD(&cur->node);
130 		cur->func(fence, cur);
131 	}
132 }
133 
134 static void add_retire(struct intel_breadcrumbs *b, struct intel_timeline *tl)
135 {
136 	struct intel_engine_cs *engine =
137 		container_of(b, struct intel_engine_cs, breadcrumbs);
138 
139 	intel_engine_add_retire(engine, tl);
140 }
141 
142 static void signal_irq_work(struct irq_work *work)
143 {
144 	struct intel_breadcrumbs *b = container_of(work, typeof(*b), irq_work);
145 	const ktime_t timestamp = ktime_get();
146 	struct intel_context *ce, *cn;
147 	struct list_head *pos, *next;
148 	LIST_HEAD(signal);
149 
150 	spin_lock(&b->irq_lock);
151 
152 	if (b->irq_armed && list_empty(&b->signalers))
153 		__intel_breadcrumbs_disarm_irq(b);
154 
155 	list_for_each_entry_safe(ce, cn, &b->signalers, signal_link) {
156 		GEM_BUG_ON(list_empty(&ce->signals));
157 
158 		list_for_each_safe(pos, next, &ce->signals) {
159 			struct i915_request *rq =
160 				list_entry(pos, typeof(*rq), signal_link);
161 
162 			GEM_BUG_ON(!check_signal_order(ce, rq));
163 
164 			if (!__request_completed(rq))
165 				break;
166 
167 			GEM_BUG_ON(!test_bit(I915_FENCE_FLAG_SIGNAL,
168 					     &rq->fence.flags));
169 			clear_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
170 
171 			if (!__dma_fence_signal(&rq->fence))
172 				continue;
173 
174 			/*
175 			 * Queue for execution after dropping the signaling
176 			 * spinlock as the callback chain may end up adding
177 			 * more signalers to the same context or engine.
178 			 */
179 			i915_request_get(rq);
180 			list_add_tail(&rq->signal_link, &signal);
181 		}
182 
183 		/*
184 		 * We process the list deletion in bulk, only using a list_add
185 		 * (not list_move) above but keeping the status of
186 		 * rq->signal_link known with the I915_FENCE_FLAG_SIGNAL bit.
187 		 */
188 		if (!list_is_first(pos, &ce->signals)) {
189 			/* Advance the list to the first incomplete request */
190 			__list_del_many(&ce->signals, pos);
191 			if (&ce->signals == pos) { /* now empty */
192 				list_del_init(&ce->signal_link);
193 				add_retire(b, ce->timeline);
194 			}
195 		}
196 	}
197 
198 	spin_unlock(&b->irq_lock);
199 
200 	list_for_each_safe(pos, next, &signal) {
201 		struct i915_request *rq =
202 			list_entry(pos, typeof(*rq), signal_link);
203 		struct list_head cb_list;
204 
205 		spin_lock(&rq->lock);
206 		list_replace(&rq->fence.cb_list, &cb_list);
207 		__dma_fence_signal__timestamp(&rq->fence, timestamp);
208 		__dma_fence_signal__notify(&rq->fence, &cb_list);
209 		spin_unlock(&rq->lock);
210 
211 		i915_request_put(rq);
212 	}
213 }
214 
215 static bool __intel_breadcrumbs_arm_irq(struct intel_breadcrumbs *b)
216 {
217 	struct intel_engine_cs *engine =
218 		container_of(b, struct intel_engine_cs, breadcrumbs);
219 
220 	lockdep_assert_held(&b->irq_lock);
221 	if (b->irq_armed)
222 		return true;
223 
224 	if (!intel_gt_pm_get_if_awake(engine->gt))
225 		return false;
226 
227 	/*
228 	 * The breadcrumb irq will be disarmed on the interrupt after the
229 	 * waiters are signaled. This gives us a single interrupt window in
230 	 * which we can add a new waiter and avoid the cost of re-enabling
231 	 * the irq.
232 	 */
233 	b->irq_armed = true;
234 
235 	/*
236 	 * Since we are waiting on a request, the GPU should be busy
237 	 * and should have its own rpm reference. This is tracked
238 	 * by i915->gt.awake, we can forgo holding our own wakref
239 	 * for the interrupt as before i915->gt.awake is released (when
240 	 * the driver is idle) we disarm the breadcrumbs.
241 	 */
242 
243 	if (!b->irq_enabled++)
244 		irq_enable(engine);
245 
246 	return true;
247 }
248 
249 void intel_engine_init_breadcrumbs(struct intel_engine_cs *engine)
250 {
251 	struct intel_breadcrumbs *b = &engine->breadcrumbs;
252 
253 	spin_lock_init(&b->irq_lock);
254 	INIT_LIST_HEAD(&b->signalers);
255 
256 	init_irq_work(&b->irq_work, signal_irq_work);
257 }
258 
259 void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine)
260 {
261 	struct intel_breadcrumbs *b = &engine->breadcrumbs;
262 	unsigned long flags;
263 
264 	spin_lock_irqsave(&b->irq_lock, flags);
265 
266 	if (b->irq_enabled)
267 		irq_enable(engine);
268 	else
269 		irq_disable(engine);
270 
271 	spin_unlock_irqrestore(&b->irq_lock, flags);
272 }
273 
274 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine)
275 {
276 }
277 
278 bool i915_request_enable_breadcrumb(struct i915_request *rq)
279 {
280 	lockdep_assert_held(&rq->lock);
281 
282 	if (test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags)) {
283 		struct intel_breadcrumbs *b = &rq->engine->breadcrumbs;
284 		struct intel_context *ce = rq->context;
285 		struct list_head *pos;
286 
287 		spin_lock(&b->irq_lock);
288 		GEM_BUG_ON(test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags));
289 
290 		if (!__intel_breadcrumbs_arm_irq(b))
291 			goto unlock;
292 
293 		/*
294 		 * We keep the seqno in retirement order, so we can break
295 		 * inside intel_engine_signal_breadcrumbs as soon as we've
296 		 * passed the last completed request (or seen a request that
297 		 * hasn't event started). We could walk the timeline->requests,
298 		 * but keeping a separate signalers_list has the advantage of
299 		 * hopefully being much smaller than the full list and so
300 		 * provides faster iteration and detection when there are no
301 		 * more interrupts required for this context.
302 		 *
303 		 * We typically expect to add new signalers in order, so we
304 		 * start looking for our insertion point from the tail of
305 		 * the list.
306 		 */
307 		list_for_each_prev(pos, &ce->signals) {
308 			struct i915_request *it =
309 				list_entry(pos, typeof(*it), signal_link);
310 
311 			if (i915_seqno_passed(rq->fence.seqno, it->fence.seqno))
312 				break;
313 		}
314 		list_add(&rq->signal_link, pos);
315 		if (pos == &ce->signals) /* catch transitions from empty list */
316 			list_move_tail(&ce->signal_link, &b->signalers);
317 		GEM_BUG_ON(!check_signal_order(ce, rq));
318 
319 		set_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
320 unlock:
321 		spin_unlock(&b->irq_lock);
322 	}
323 
324 	return !__request_completed(rq);
325 }
326 
327 void i915_request_cancel_breadcrumb(struct i915_request *rq)
328 {
329 	struct intel_breadcrumbs *b = &rq->engine->breadcrumbs;
330 
331 	lockdep_assert_held(&rq->lock);
332 
333 	/*
334 	 * We must wait for b->irq_lock so that we know the interrupt handler
335 	 * has released its reference to the intel_context and has completed
336 	 * the DMA_FENCE_FLAG_SIGNALED_BIT/I915_FENCE_FLAG_SIGNAL dance (if
337 	 * required).
338 	 */
339 	spin_lock(&b->irq_lock);
340 	if (test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags)) {
341 		struct intel_context *ce = rq->context;
342 
343 		list_del(&rq->signal_link);
344 		if (list_empty(&ce->signals))
345 			list_del_init(&ce->signal_link);
346 
347 		clear_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
348 	}
349 	spin_unlock(&b->irq_lock);
350 }
351 
352 void intel_engine_print_breadcrumbs(struct intel_engine_cs *engine,
353 				    struct drm_printer *p)
354 {
355 	struct intel_breadcrumbs *b = &engine->breadcrumbs;
356 	struct intel_context *ce;
357 	struct i915_request *rq;
358 
359 	if (list_empty(&b->signalers))
360 		return;
361 
362 	drm_printf(p, "Signals:\n");
363 
364 	spin_lock_irq(&b->irq_lock);
365 	list_for_each_entry(ce, &b->signalers, signal_link) {
366 		list_for_each_entry(rq, &ce->signals, signal_link) {
367 			drm_printf(p, "\t[%llx:%llx%s] @ %dms\n",
368 				   rq->fence.context, rq->fence.seqno,
369 				   i915_request_completed(rq) ? "!" :
370 				   i915_request_started(rq) ? "*" :
371 				   "",
372 				   jiffies_to_msecs(jiffies - rq->emitted_jiffies));
373 		}
374 	}
375 	spin_unlock_irq(&b->irq_lock);
376 }
377