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_breadcrumbs.h"
32 #include "intel_context.h"
33 #include "intel_engine_pm.h"
34 #include "intel_gt_pm.h"
35 #include "intel_gt_requests.h"
36 
37 static bool irq_enable(struct intel_engine_cs *engine)
38 {
39 	if (!engine->irq_enable)
40 		return false;
41 
42 	/* Caller disables interrupts */
43 	spin_lock(&engine->gt->irq_lock);
44 	engine->irq_enable(engine);
45 	spin_unlock(&engine->gt->irq_lock);
46 
47 	return true;
48 }
49 
50 static void irq_disable(struct intel_engine_cs *engine)
51 {
52 	if (!engine->irq_disable)
53 		return;
54 
55 	/* Caller disables interrupts */
56 	spin_lock(&engine->gt->irq_lock);
57 	engine->irq_disable(engine);
58 	spin_unlock(&engine->gt->irq_lock);
59 }
60 
61 static void __intel_breadcrumbs_arm_irq(struct intel_breadcrumbs *b)
62 {
63 	/*
64 	 * Since we are waiting on a request, the GPU should be busy
65 	 * and should have its own rpm reference.
66 	 */
67 	if (GEM_WARN_ON(!intel_gt_pm_get_if_awake(b->irq_engine->gt)))
68 		return;
69 
70 	/*
71 	 * The breadcrumb irq will be disarmed on the interrupt after the
72 	 * waiters are signaled. This gives us a single interrupt window in
73 	 * which we can add a new waiter and avoid the cost of re-enabling
74 	 * the irq.
75 	 */
76 	WRITE_ONCE(b->irq_armed, true);
77 
78 	/* Requests may have completed before we could enable the interrupt. */
79 	if (!b->irq_enabled++ && irq_enable(b->irq_engine))
80 		irq_work_queue(&b->irq_work);
81 }
82 
83 static void intel_breadcrumbs_arm_irq(struct intel_breadcrumbs *b)
84 {
85 	if (!b->irq_engine)
86 		return;
87 
88 	spin_lock(&b->irq_lock);
89 	if (!b->irq_armed)
90 		__intel_breadcrumbs_arm_irq(b);
91 	spin_unlock(&b->irq_lock);
92 }
93 
94 static void __intel_breadcrumbs_disarm_irq(struct intel_breadcrumbs *b)
95 {
96 	GEM_BUG_ON(!b->irq_enabled);
97 	if (!--b->irq_enabled)
98 		irq_disable(b->irq_engine);
99 
100 	WRITE_ONCE(b->irq_armed, false);
101 	intel_gt_pm_put_async(b->irq_engine->gt);
102 }
103 
104 static void intel_breadcrumbs_disarm_irq(struct intel_breadcrumbs *b)
105 {
106 	spin_lock(&b->irq_lock);
107 	if (b->irq_armed)
108 		__intel_breadcrumbs_disarm_irq(b);
109 	spin_unlock(&b->irq_lock);
110 }
111 
112 static void add_signaling_context(struct intel_breadcrumbs *b,
113 				  struct intel_context *ce)
114 {
115 	lockdep_assert_held(&ce->signal_lock);
116 
117 	spin_lock(&b->signalers_lock);
118 	list_add_rcu(&ce->signal_link, &b->signalers);
119 	spin_unlock(&b->signalers_lock);
120 }
121 
122 static bool remove_signaling_context(struct intel_breadcrumbs *b,
123 				     struct intel_context *ce)
124 {
125 	lockdep_assert_held(&ce->signal_lock);
126 
127 	if (!list_empty(&ce->signals))
128 		return false;
129 
130 	spin_lock(&b->signalers_lock);
131 	list_del_rcu(&ce->signal_link);
132 	spin_unlock(&b->signalers_lock);
133 
134 	return true;
135 }
136 
137 static inline bool __request_completed(const struct i915_request *rq)
138 {
139 	return i915_seqno_passed(__hwsp_seqno(rq), rq->fence.seqno);
140 }
141 
142 __maybe_unused static bool
143 check_signal_order(struct intel_context *ce, struct i915_request *rq)
144 {
145 	if (rq->context != ce)
146 		return false;
147 
148 	if (!list_is_last(&rq->signal_link, &ce->signals) &&
149 	    i915_seqno_passed(rq->fence.seqno,
150 			      list_next_entry(rq, signal_link)->fence.seqno))
151 		return false;
152 
153 	if (!list_is_first(&rq->signal_link, &ce->signals) &&
154 	    i915_seqno_passed(list_prev_entry(rq, signal_link)->fence.seqno,
155 			      rq->fence.seqno))
156 		return false;
157 
158 	return true;
159 }
160 
161 static bool
162 __dma_fence_signal(struct dma_fence *fence)
163 {
164 	return !test_and_set_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags);
165 }
166 
167 static void
168 __dma_fence_signal__timestamp(struct dma_fence *fence, ktime_t timestamp)
169 {
170 	fence->timestamp = timestamp;
171 	set_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &fence->flags);
172 	trace_dma_fence_signaled(fence);
173 }
174 
175 static void
176 __dma_fence_signal__notify(struct dma_fence *fence,
177 			   const struct list_head *list)
178 {
179 	struct dma_fence_cb *cur, *tmp;
180 
181 	lockdep_assert_held(fence->lock);
182 
183 	list_for_each_entry_safe(cur, tmp, list, node) {
184 		INIT_LIST_HEAD(&cur->node);
185 		cur->func(fence, cur);
186 	}
187 }
188 
189 static void add_retire(struct intel_breadcrumbs *b, struct intel_timeline *tl)
190 {
191 	if (b->irq_engine)
192 		intel_engine_add_retire(b->irq_engine, tl);
193 }
194 
195 static bool __signal_request(struct i915_request *rq)
196 {
197 	GEM_BUG_ON(test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags));
198 
199 	if (!__dma_fence_signal(&rq->fence)) {
200 		i915_request_put(rq);
201 		return false;
202 	}
203 
204 	return true;
205 }
206 
207 static struct llist_node *
208 slist_add(struct llist_node *node, struct llist_node *head)
209 {
210 	node->next = head;
211 	return node;
212 }
213 
214 static void signal_irq_work(struct irq_work *work)
215 {
216 	struct intel_breadcrumbs *b = container_of(work, typeof(*b), irq_work);
217 	const ktime_t timestamp = ktime_get();
218 	struct llist_node *signal, *sn;
219 	struct intel_context *ce;
220 
221 	signal = NULL;
222 	if (unlikely(!llist_empty(&b->signaled_requests)))
223 		signal = llist_del_all(&b->signaled_requests);
224 
225 	/*
226 	 * Keep the irq armed until the interrupt after all listeners are gone.
227 	 *
228 	 * Enabling/disabling the interrupt is rather costly, roughly a couple
229 	 * of hundred microseconds. If we are proactive and enable/disable
230 	 * the interrupt around every request that wants a breadcrumb, we
231 	 * quickly drown in the extra orders of magnitude of latency imposed
232 	 * on request submission.
233 	 *
234 	 * So we try to be lazy, and keep the interrupts enabled until no
235 	 * more listeners appear within a breadcrumb interrupt interval (that
236 	 * is until a request completes that no one cares about). The
237 	 * observation is that listeners come in batches, and will often
238 	 * listen to a bunch of requests in succession. Though note on icl+,
239 	 * interrupts are always enabled due to concerns with rc6 being
240 	 * dysfunctional with per-engine interrupt masking.
241 	 *
242 	 * We also try to avoid raising too many interrupts, as they may
243 	 * be generated by userspace batches and it is unfortunately rather
244 	 * too easy to drown the CPU under a flood of GPU interrupts. Thus
245 	 * whenever no one appears to be listening, we turn off the interrupts.
246 	 * Fewer interrupts should conserve power -- at the very least, fewer
247 	 * interrupt draw less ire from other users of the system and tools
248 	 * like powertop.
249 	 */
250 	if (!signal && READ_ONCE(b->irq_armed) && list_empty(&b->signalers))
251 		intel_breadcrumbs_disarm_irq(b);
252 
253 	rcu_read_lock();
254 	list_for_each_entry_rcu(ce, &b->signalers, signal_link) {
255 		struct i915_request *rq;
256 
257 		list_for_each_entry_rcu(rq, &ce->signals, signal_link) {
258 			bool release;
259 
260 			if (!__request_completed(rq))
261 				break;
262 
263 			if (!test_and_clear_bit(I915_FENCE_FLAG_SIGNAL,
264 						&rq->fence.flags))
265 				break;
266 
267 			/*
268 			 * Queue for execution after dropping the signaling
269 			 * spinlock as the callback chain may end up adding
270 			 * more signalers to the same context or engine.
271 			 */
272 			spin_lock(&ce->signal_lock);
273 			list_del_rcu(&rq->signal_link);
274 			release = remove_signaling_context(b, ce);
275 			spin_unlock(&ce->signal_lock);
276 
277 			if (__signal_request(rq))
278 				/* We own signal_node now, xfer to local list */
279 				signal = slist_add(&rq->signal_node, signal);
280 
281 			if (release) {
282 				add_retire(b, ce->timeline);
283 				intel_context_put(ce);
284 			}
285 		}
286 	}
287 	rcu_read_unlock();
288 
289 	llist_for_each_safe(signal, sn, signal) {
290 		struct i915_request *rq =
291 			llist_entry(signal, typeof(*rq), signal_node);
292 		struct list_head cb_list;
293 
294 		spin_lock(&rq->lock);
295 		list_replace(&rq->fence.cb_list, &cb_list);
296 		__dma_fence_signal__timestamp(&rq->fence, timestamp);
297 		__dma_fence_signal__notify(&rq->fence, &cb_list);
298 		spin_unlock(&rq->lock);
299 
300 		i915_request_put(rq);
301 	}
302 
303 	if (!READ_ONCE(b->irq_armed) && !list_empty(&b->signalers))
304 		intel_breadcrumbs_arm_irq(b);
305 }
306 
307 struct intel_breadcrumbs *
308 intel_breadcrumbs_create(struct intel_engine_cs *irq_engine)
309 {
310 	struct intel_breadcrumbs *b;
311 
312 	b = kzalloc(sizeof(*b), GFP_KERNEL);
313 	if (!b)
314 		return NULL;
315 
316 	b->irq_engine = irq_engine;
317 
318 	spin_lock_init(&b->signalers_lock);
319 	INIT_LIST_HEAD(&b->signalers);
320 	init_llist_head(&b->signaled_requests);
321 
322 	spin_lock_init(&b->irq_lock);
323 	init_irq_work(&b->irq_work, signal_irq_work);
324 
325 	return b;
326 }
327 
328 void intel_breadcrumbs_reset(struct intel_breadcrumbs *b)
329 {
330 	unsigned long flags;
331 
332 	if (!b->irq_engine)
333 		return;
334 
335 	spin_lock_irqsave(&b->irq_lock, flags);
336 
337 	if (b->irq_enabled)
338 		irq_enable(b->irq_engine);
339 	else
340 		irq_disable(b->irq_engine);
341 
342 	spin_unlock_irqrestore(&b->irq_lock, flags);
343 }
344 
345 void intel_breadcrumbs_park(struct intel_breadcrumbs *b)
346 {
347 	/* Kick the work once more to drain the signalers */
348 	irq_work_sync(&b->irq_work);
349 	while (unlikely(READ_ONCE(b->irq_armed))) {
350 		local_irq_disable();
351 		signal_irq_work(&b->irq_work);
352 		local_irq_enable();
353 		cond_resched();
354 	}
355 	GEM_BUG_ON(!list_empty(&b->signalers));
356 }
357 
358 void intel_breadcrumbs_free(struct intel_breadcrumbs *b)
359 {
360 	irq_work_sync(&b->irq_work);
361 	GEM_BUG_ON(!list_empty(&b->signalers));
362 	GEM_BUG_ON(b->irq_armed);
363 	kfree(b);
364 }
365 
366 static void insert_breadcrumb(struct i915_request *rq)
367 {
368 	struct intel_breadcrumbs *b = READ_ONCE(rq->engine)->breadcrumbs;
369 	struct intel_context *ce = rq->context;
370 	struct list_head *pos;
371 
372 	if (test_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags))
373 		return;
374 
375 	i915_request_get(rq);
376 
377 	/*
378 	 * If the request is already completed, we can transfer it
379 	 * straight onto a signaled list, and queue the irq worker for
380 	 * its signal completion.
381 	 */
382 	if (__request_completed(rq)) {
383 		if (__signal_request(rq) &&
384 		    llist_add(&rq->signal_node, &b->signaled_requests))
385 			irq_work_queue(&b->irq_work);
386 		return;
387 	}
388 
389 	if (list_empty(&ce->signals)) {
390 		intel_context_get(ce);
391 		add_signaling_context(b, ce);
392 		pos = &ce->signals;
393 	} else {
394 		/*
395 		 * We keep the seqno in retirement order, so we can break
396 		 * inside intel_engine_signal_breadcrumbs as soon as we've
397 		 * passed the last completed request (or seen a request that
398 		 * hasn't event started). We could walk the timeline->requests,
399 		 * but keeping a separate signalers_list has the advantage of
400 		 * hopefully being much smaller than the full list and so
401 		 * provides faster iteration and detection when there are no
402 		 * more interrupts required for this context.
403 		 *
404 		 * We typically expect to add new signalers in order, so we
405 		 * start looking for our insertion point from the tail of
406 		 * the list.
407 		 */
408 		list_for_each_prev(pos, &ce->signals) {
409 			struct i915_request *it =
410 				list_entry(pos, typeof(*it), signal_link);
411 
412 			if (i915_seqno_passed(rq->fence.seqno, it->fence.seqno))
413 				break;
414 		}
415 	}
416 	list_add_rcu(&rq->signal_link, pos);
417 	GEM_BUG_ON(!check_signal_order(ce, rq));
418 	GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags));
419 	set_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags);
420 
421 	/*
422 	 * Defer enabling the interrupt to after HW submission and recheck
423 	 * the request as it may have completed and raised the interrupt as
424 	 * we were attaching it into the lists.
425 	 */
426 	irq_work_queue(&b->irq_work);
427 }
428 
429 bool i915_request_enable_breadcrumb(struct i915_request *rq)
430 {
431 	struct intel_context *ce = rq->context;
432 
433 	/* Serialises with i915_request_retire() using rq->lock */
434 	if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags))
435 		return true;
436 
437 	/*
438 	 * Peek at i915_request_submit()/i915_request_unsubmit() status.
439 	 *
440 	 * If the request is not yet active (and not signaled), we will
441 	 * attach the breadcrumb later.
442 	 */
443 	if (!test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags))
444 		return true;
445 
446 	spin_lock(&ce->signal_lock);
447 	if (test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags))
448 		insert_breadcrumb(rq);
449 	spin_unlock(&ce->signal_lock);
450 
451 	return true;
452 }
453 
454 void i915_request_cancel_breadcrumb(struct i915_request *rq)
455 {
456 	struct intel_context *ce = rq->context;
457 	bool release;
458 
459 	if (!test_and_clear_bit(I915_FENCE_FLAG_SIGNAL, &rq->fence.flags))
460 		return;
461 
462 	spin_lock(&ce->signal_lock);
463 	list_del_rcu(&rq->signal_link);
464 	release = remove_signaling_context(rq->engine->breadcrumbs, ce);
465 	spin_unlock(&ce->signal_lock);
466 	if (release)
467 		intel_context_put(ce);
468 
469 	i915_request_put(rq);
470 }
471 
472 static void print_signals(struct intel_breadcrumbs *b, struct drm_printer *p)
473 {
474 	struct intel_context *ce;
475 	struct i915_request *rq;
476 
477 	drm_printf(p, "Signals:\n");
478 
479 	rcu_read_lock();
480 	list_for_each_entry_rcu(ce, &b->signalers, signal_link) {
481 		list_for_each_entry_rcu(rq, &ce->signals, signal_link)
482 			drm_printf(p, "\t[%llx:%llx%s] @ %dms\n",
483 				   rq->fence.context, rq->fence.seqno,
484 				   i915_request_completed(rq) ? "!" :
485 				   i915_request_started(rq) ? "*" :
486 				   "",
487 				   jiffies_to_msecs(jiffies - rq->emitted_jiffies));
488 	}
489 	rcu_read_unlock();
490 }
491 
492 void intel_engine_print_breadcrumbs(struct intel_engine_cs *engine,
493 				    struct drm_printer *p)
494 {
495 	struct intel_breadcrumbs *b;
496 
497 	b = engine->breadcrumbs;
498 	if (!b)
499 		return;
500 
501 	drm_printf(p, "IRQ: %s\n", enableddisabled(b->irq_armed));
502 	if (!list_empty(&b->signalers))
503 		print_signals(b, p);
504 }
505