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 	if (unlikely(intel_engine_is_virtual(engine)))
140 		engine = intel_virtual_engine_get_sibling(engine, 0);
141 
142 	intel_engine_add_retire(engine, tl);
143 }
144 
145 static void signal_irq_work(struct irq_work *work)
146 {
147 	struct intel_breadcrumbs *b = container_of(work, typeof(*b), irq_work);
148 	const ktime_t timestamp = ktime_get();
149 	struct intel_context *ce, *cn;
150 	struct list_head *pos, *next;
151 	LIST_HEAD(signal);
152 
153 	spin_lock(&b->irq_lock);
154 
155 	if (b->irq_armed && list_empty(&b->signalers))
156 		__intel_breadcrumbs_disarm_irq(b);
157 
158 	list_for_each_entry_safe(ce, cn, &b->signalers, signal_link) {
159 		GEM_BUG_ON(list_empty(&ce->signals));
160 
161 		list_for_each_safe(pos, next, &ce->signals) {
162 			struct i915_request *rq =
163 				list_entry(pos, typeof(*rq), signal_link);
164 
165 			GEM_BUG_ON(!check_signal_order(ce, rq));
166 
167 			if (!__request_completed(rq))
168 				break;
169 
170 			GEM_BUG_ON(!test_bit(I915_FENCE_FLAG_SIGNAL,
171 					     &rq->fence.flags));
172 			clear_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
173 
174 			if (!__dma_fence_signal(&rq->fence))
175 				continue;
176 
177 			/*
178 			 * Queue for execution after dropping the signaling
179 			 * spinlock as the callback chain may end up adding
180 			 * more signalers to the same context or engine.
181 			 */
182 			i915_request_get(rq);
183 			list_add_tail(&rq->signal_link, &signal);
184 		}
185 
186 		/*
187 		 * We process the list deletion in bulk, only using a list_add
188 		 * (not list_move) above but keeping the status of
189 		 * rq->signal_link known with the I915_FENCE_FLAG_SIGNAL bit.
190 		 */
191 		if (!list_is_first(pos, &ce->signals)) {
192 			/* Advance the list to the first incomplete request */
193 			__list_del_many(&ce->signals, pos);
194 			if (&ce->signals == pos) { /* now empty */
195 				list_del_init(&ce->signal_link);
196 				add_retire(b, ce->timeline);
197 			}
198 		}
199 	}
200 
201 	spin_unlock(&b->irq_lock);
202 
203 	list_for_each_safe(pos, next, &signal) {
204 		struct i915_request *rq =
205 			list_entry(pos, typeof(*rq), signal_link);
206 		struct list_head cb_list;
207 
208 		spin_lock(&rq->lock);
209 		list_replace(&rq->fence.cb_list, &cb_list);
210 		__dma_fence_signal__timestamp(&rq->fence, timestamp);
211 		__dma_fence_signal__notify(&rq->fence, &cb_list);
212 		spin_unlock(&rq->lock);
213 
214 		i915_request_put(rq);
215 	}
216 }
217 
218 static bool __intel_breadcrumbs_arm_irq(struct intel_breadcrumbs *b)
219 {
220 	struct intel_engine_cs *engine =
221 		container_of(b, struct intel_engine_cs, breadcrumbs);
222 
223 	lockdep_assert_held(&b->irq_lock);
224 	if (b->irq_armed)
225 		return true;
226 
227 	if (!intel_gt_pm_get_if_awake(engine->gt))
228 		return false;
229 
230 	/*
231 	 * The breadcrumb irq will be disarmed on the interrupt after the
232 	 * waiters are signaled. This gives us a single interrupt window in
233 	 * which we can add a new waiter and avoid the cost of re-enabling
234 	 * the irq.
235 	 */
236 	b->irq_armed = true;
237 
238 	/*
239 	 * Since we are waiting on a request, the GPU should be busy
240 	 * and should have its own rpm reference. This is tracked
241 	 * by i915->gt.awake, we can forgo holding our own wakref
242 	 * for the interrupt as before i915->gt.awake is released (when
243 	 * the driver is idle) we disarm the breadcrumbs.
244 	 */
245 
246 	if (!b->irq_enabled++)
247 		irq_enable(engine);
248 
249 	return true;
250 }
251 
252 void intel_engine_init_breadcrumbs(struct intel_engine_cs *engine)
253 {
254 	struct intel_breadcrumbs *b = &engine->breadcrumbs;
255 
256 	spin_lock_init(&b->irq_lock);
257 	INIT_LIST_HEAD(&b->signalers);
258 
259 	init_irq_work(&b->irq_work, signal_irq_work);
260 }
261 
262 void intel_engine_reset_breadcrumbs(struct intel_engine_cs *engine)
263 {
264 	struct intel_breadcrumbs *b = &engine->breadcrumbs;
265 	unsigned long flags;
266 
267 	spin_lock_irqsave(&b->irq_lock, flags);
268 
269 	if (b->irq_enabled)
270 		irq_enable(engine);
271 	else
272 		irq_disable(engine);
273 
274 	spin_unlock_irqrestore(&b->irq_lock, flags);
275 }
276 
277 void intel_engine_fini_breadcrumbs(struct intel_engine_cs *engine)
278 {
279 }
280 
281 bool i915_request_enable_breadcrumb(struct i915_request *rq)
282 {
283 	lockdep_assert_held(&rq->lock);
284 
285 	if (test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags)) {
286 		struct intel_breadcrumbs *b = &rq->engine->breadcrumbs;
287 		struct intel_context *ce = rq->context;
288 		struct list_head *pos;
289 
290 		spin_lock(&b->irq_lock);
291 		GEM_BUG_ON(test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags));
292 
293 		if (!__intel_breadcrumbs_arm_irq(b))
294 			goto unlock;
295 
296 		/*
297 		 * We keep the seqno in retirement order, so we can break
298 		 * inside intel_engine_signal_breadcrumbs as soon as we've
299 		 * passed the last completed request (or seen a request that
300 		 * hasn't event started). We could walk the timeline->requests,
301 		 * but keeping a separate signalers_list has the advantage of
302 		 * hopefully being much smaller than the full list and so
303 		 * provides faster iteration and detection when there are no
304 		 * more interrupts required for this context.
305 		 *
306 		 * We typically expect to add new signalers in order, so we
307 		 * start looking for our insertion point from the tail of
308 		 * the list.
309 		 */
310 		list_for_each_prev(pos, &ce->signals) {
311 			struct i915_request *it =
312 				list_entry(pos, typeof(*it), signal_link);
313 
314 			if (i915_seqno_passed(rq->fence.seqno, it->fence.seqno))
315 				break;
316 		}
317 		list_add(&rq->signal_link, pos);
318 		if (pos == &ce->signals) /* catch transitions from empty list */
319 			list_move_tail(&ce->signal_link, &b->signalers);
320 		GEM_BUG_ON(!check_signal_order(ce, rq));
321 
322 		set_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
323 unlock:
324 		spin_unlock(&b->irq_lock);
325 	}
326 
327 	return !__request_completed(rq);
328 }
329 
330 void i915_request_cancel_breadcrumb(struct i915_request *rq)
331 {
332 	struct intel_breadcrumbs *b = &rq->engine->breadcrumbs;
333 
334 	lockdep_assert_held(&rq->lock);
335 
336 	/*
337 	 * We must wait for b->irq_lock so that we know the interrupt handler
338 	 * has released its reference to the intel_context and has completed
339 	 * the DMA_FENCE_FLAG_SIGNALED_BIT/I915_FENCE_FLAG_SIGNAL dance (if
340 	 * required).
341 	 */
342 	spin_lock(&b->irq_lock);
343 	if (test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags)) {
344 		struct intel_context *ce = rq->context;
345 
346 		list_del(&rq->signal_link);
347 		if (list_empty(&ce->signals))
348 			list_del_init(&ce->signal_link);
349 
350 		clear_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
351 	}
352 	spin_unlock(&b->irq_lock);
353 }
354 
355 void intel_engine_print_breadcrumbs(struct intel_engine_cs *engine,
356 				    struct drm_printer *p)
357 {
358 	struct intel_breadcrumbs *b = &engine->breadcrumbs;
359 	struct intel_context *ce;
360 	struct i915_request *rq;
361 
362 	if (list_empty(&b->signalers))
363 		return;
364 
365 	drm_printf(p, "Signals:\n");
366 
367 	spin_lock_irq(&b->irq_lock);
368 	list_for_each_entry(ce, &b->signalers, signal_link) {
369 		list_for_each_entry(rq, &ce->signals, signal_link) {
370 			drm_printf(p, "\t[%llx:%llx%s] @ %dms\n",
371 				   rq->fence.context, rq->fence.seqno,
372 				   i915_request_completed(rq) ? "!" :
373 				   i915_request_started(rq) ? "*" :
374 				   "",
375 				   jiffies_to_msecs(jiffies - rq->emitted_jiffies));
376 		}
377 	}
378 	spin_unlock_irq(&b->irq_lock);
379 }
380