1 /*
2  * Copyright © 2008-2018 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 #ifndef I915_REQUEST_H
26 #define I915_REQUEST_H
27 
28 #include <linux/dma-fence.h>
29 #include <linux/irq_work.h>
30 #include <linux/lockdep.h>
31 
32 #include "gem/i915_gem_context_types.h"
33 #include "gt/intel_context_types.h"
34 #include "gt/intel_engine_types.h"
35 #include "gt/intel_timeline_types.h"
36 
37 #include "i915_gem.h"
38 #include "i915_scheduler.h"
39 #include "i915_selftest.h"
40 #include "i915_sw_fence.h"
41 
42 #include <uapi/drm/i915_drm.h>
43 
44 struct drm_file;
45 struct drm_i915_gem_object;
46 struct i915_request;
47 
48 struct i915_capture_list {
49 	struct i915_capture_list *next;
50 	struct i915_vma *vma;
51 };
52 
53 #define RQ_TRACE(rq, fmt, ...) do {					\
54 	const struct i915_request *rq__ = (rq);				\
55 	ENGINE_TRACE(rq__->engine, "fence %llx:%lld, current %d " fmt,	\
56 		     rq__->fence.context, rq__->fence.seqno,		\
57 		     hwsp_seqno(rq__), ##__VA_ARGS__);			\
58 } while (0)
59 
60 enum {
61 	/*
62 	 * I915_FENCE_FLAG_ACTIVE - this request is currently submitted to HW.
63 	 *
64 	 * Set by __i915_request_submit() on handing over to HW, and cleared
65 	 * by __i915_request_unsubmit() if we preempt this request.
66 	 *
67 	 * Finally cleared for consistency on retiring the request, when
68 	 * we know the HW is no longer running this request.
69 	 *
70 	 * See i915_request_is_active()
71 	 */
72 	I915_FENCE_FLAG_ACTIVE = DMA_FENCE_FLAG_USER_BITS,
73 
74 	/*
75 	 * I915_FENCE_FLAG_PQUEUE - this request is ready for execution
76 	 *
77 	 * Using the scheduler, when a request is ready for execution it is put
78 	 * into the priority queue, and removed from that queue when transferred
79 	 * to the HW runlists. We want to track its membership within the
80 	 * priority queue so that we can easily check before rescheduling.
81 	 *
82 	 * See i915_request_in_priority_queue()
83 	 */
84 	I915_FENCE_FLAG_PQUEUE,
85 
86 	/*
87 	 * I915_FENCE_FLAG_HOLD - this request is currently on hold
88 	 *
89 	 * This request has been suspended, pending an ongoing investigation.
90 	 */
91 	I915_FENCE_FLAG_HOLD,
92 
93 	/*
94 	 * I915_FENCE_FLAG_INITIAL_BREADCRUMB - this request has the initial
95 	 * breadcrumb that marks the end of semaphore waits and start of the
96 	 * user payload.
97 	 */
98 	I915_FENCE_FLAG_INITIAL_BREADCRUMB,
99 
100 	/*
101 	 * I915_FENCE_FLAG_SIGNAL - this request is currently on signal_list
102 	 *
103 	 * Internal bookkeeping used by the breadcrumb code to track when
104 	 * a request is on the various signal_list.
105 	 */
106 	I915_FENCE_FLAG_SIGNAL,
107 
108 	/*
109 	 * I915_FENCE_FLAG_NOPREEMPT - this request should not be preempted
110 	 *
111 	 * The execution of some requests should not be interrupted. This is
112 	 * a sensitive operation as it makes the request super important,
113 	 * blocking other higher priority work. Abuse of this flag will
114 	 * lead to quality of service issues.
115 	 */
116 	I915_FENCE_FLAG_NOPREEMPT,
117 
118 	/*
119 	 * I915_FENCE_FLAG_SENTINEL - this request should be last in the queue
120 	 *
121 	 * A high priority sentinel request may be submitted to clear the
122 	 * submission queue. As it will be the only request in-flight, upon
123 	 * execution all other active requests will have been preempted and
124 	 * unsubmitted. This preemptive pulse is used to re-evaluate the
125 	 * in-flight requests, particularly in cases where an active context
126 	 * is banned and those active requests need to be cancelled.
127 	 */
128 	I915_FENCE_FLAG_SENTINEL,
129 
130 	/*
131 	 * I915_FENCE_FLAG_BOOST - upclock the gpu for this request
132 	 *
133 	 * Some requests are more important than others! In particular, a
134 	 * request that the user is waiting on is typically required for
135 	 * interactive latency, for which we want to minimise by upclocking
136 	 * the GPU. Here we track such boost requests on a per-request basis.
137 	 */
138 	I915_FENCE_FLAG_BOOST,
139 };
140 
141 /**
142  * Request queue structure.
143  *
144  * The request queue allows us to note sequence numbers that have been emitted
145  * and may be associated with active buffers to be retired.
146  *
147  * By keeping this list, we can avoid having to do questionable sequence
148  * number comparisons on buffer last_read|write_seqno. It also allows an
149  * emission time to be associated with the request for tracking how far ahead
150  * of the GPU the submission is.
151  *
152  * When modifying this structure be very aware that we perform a lockless
153  * RCU lookup of it that may race against reallocation of the struct
154  * from the slab freelist. We intentionally do not zero the structure on
155  * allocation so that the lookup can use the dangling pointers (and is
156  * cogniscent that those pointers may be wrong). Instead, everything that
157  * needs to be initialised must be done so explicitly.
158  *
159  * The requests are reference counted.
160  */
161 struct i915_request {
162 	struct dma_fence fence;
163 	spinlock_t lock;
164 
165 	/**
166 	 * Context and ring buffer related to this request
167 	 * Contexts are refcounted, so when this request is associated with a
168 	 * context, we must increment the context's refcount, to guarantee that
169 	 * it persists while any request is linked to it. Requests themselves
170 	 * are also refcounted, so the request will only be freed when the last
171 	 * reference to it is dismissed, and the code in
172 	 * i915_request_free() will then decrement the refcount on the
173 	 * context.
174 	 */
175 	struct intel_engine_cs *engine;
176 	struct intel_context *context;
177 	struct intel_ring *ring;
178 	struct intel_timeline __rcu *timeline;
179 	struct list_head signal_link;
180 
181 	/*
182 	 * The rcu epoch of when this request was allocated. Used to judiciously
183 	 * apply backpressure on future allocations to ensure that under
184 	 * mempressure there is sufficient RCU ticks for us to reclaim our
185 	 * RCU protected slabs.
186 	 */
187 	unsigned long rcustate;
188 
189 	/*
190 	 * We pin the timeline->mutex while constructing the request to
191 	 * ensure that no caller accidentally drops it during construction.
192 	 * The timeline->mutex must be held to ensure that only this caller
193 	 * can use the ring and manipulate the associated timeline during
194 	 * construction.
195 	 */
196 	struct pin_cookie cookie;
197 
198 	/*
199 	 * Fences for the various phases in the request's lifetime.
200 	 *
201 	 * The submit fence is used to await upon all of the request's
202 	 * dependencies. When it is signaled, the request is ready to run.
203 	 * It is used by the driver to then queue the request for execution.
204 	 */
205 	struct i915_sw_fence submit;
206 	union {
207 		wait_queue_entry_t submitq;
208 		struct i915_sw_dma_fence_cb dmaq;
209 		struct i915_request_duration_cb {
210 			struct dma_fence_cb cb;
211 			ktime_t emitted;
212 		} duration;
213 	};
214 	struct llist_head execute_cb;
215 	struct i915_sw_fence semaphore;
216 
217 	/*
218 	 * A list of everyone we wait upon, and everyone who waits upon us.
219 	 * Even though we will not be submitted to the hardware before the
220 	 * submit fence is signaled (it waits for all external events as well
221 	 * as our own requests), the scheduler still needs to know the
222 	 * dependency tree for the lifetime of the request (from execbuf
223 	 * to retirement), i.e. bidirectional dependency information for the
224 	 * request not tied to individual fences.
225 	 */
226 	struct i915_sched_node sched;
227 	struct i915_dependency dep;
228 	intel_engine_mask_t execution_mask;
229 
230 	/*
231 	 * A convenience pointer to the current breadcrumb value stored in
232 	 * the HW status page (or our timeline's local equivalent). The full
233 	 * path would be rq->hw_context->ring->timeline->hwsp_seqno.
234 	 */
235 	const u32 *hwsp_seqno;
236 
237 	/*
238 	 * If we need to access the timeline's seqno for this request in
239 	 * another request, we need to keep a read reference to this associated
240 	 * cacheline, so that we do not free and recycle it before the foreign
241 	 * observers have completed. Hence, we keep a pointer to the cacheline
242 	 * inside the timeline's HWSP vma, but it is only valid while this
243 	 * request has not completed and guarded by the timeline mutex.
244 	 */
245 	struct intel_timeline_cacheline __rcu *hwsp_cacheline;
246 
247 	/** Position in the ring of the start of the request */
248 	u32 head;
249 
250 	/** Position in the ring of the start of the user packets */
251 	u32 infix;
252 
253 	/**
254 	 * Position in the ring of the start of the postfix.
255 	 * This is required to calculate the maximum available ring space
256 	 * without overwriting the postfix.
257 	 */
258 	u32 postfix;
259 
260 	/** Position in the ring of the end of the whole request */
261 	u32 tail;
262 
263 	/** Position in the ring of the end of any workarounds after the tail */
264 	u32 wa_tail;
265 
266 	/** Preallocate space in the ring for the emitting the request */
267 	u32 reserved_space;
268 
269 	/** Batch buffer related to this request if any (used for
270 	 * error state dump only).
271 	 */
272 	struct i915_vma *batch;
273 	/**
274 	 * Additional buffers requested by userspace to be captured upon
275 	 * a GPU hang. The vma/obj on this list are protected by their
276 	 * active reference - all objects on this list must also be
277 	 * on the active_list (of their final request).
278 	 */
279 	struct i915_capture_list *capture_list;
280 
281 	/** Time at which this request was emitted, in jiffies. */
282 	unsigned long emitted_jiffies;
283 
284 	/** timeline->request entry for this request */
285 	struct list_head link;
286 
287 	I915_SELFTEST_DECLARE(struct {
288 		struct list_head link;
289 		unsigned long delay;
290 	} mock;)
291 };
292 
293 #define I915_FENCE_GFP (GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN)
294 
295 extern const struct dma_fence_ops i915_fence_ops;
296 
297 static inline bool dma_fence_is_i915(const struct dma_fence *fence)
298 {
299 	return fence->ops == &i915_fence_ops;
300 }
301 
302 struct kmem_cache *i915_request_slab_cache(void);
303 
304 struct i915_request * __must_check
305 __i915_request_create(struct intel_context *ce, gfp_t gfp);
306 struct i915_request * __must_check
307 i915_request_create(struct intel_context *ce);
308 
309 void i915_request_set_error_once(struct i915_request *rq, int error);
310 void __i915_request_skip(struct i915_request *rq);
311 
312 struct i915_request *__i915_request_commit(struct i915_request *request);
313 void __i915_request_queue(struct i915_request *rq,
314 			  const struct i915_sched_attr *attr);
315 
316 bool i915_request_retire(struct i915_request *rq);
317 void i915_request_retire_upto(struct i915_request *rq);
318 
319 static inline struct i915_request *
320 to_request(struct dma_fence *fence)
321 {
322 	/* We assume that NULL fence/request are interoperable */
323 	BUILD_BUG_ON(offsetof(struct i915_request, fence) != 0);
324 	GEM_BUG_ON(fence && !dma_fence_is_i915(fence));
325 	return container_of(fence, struct i915_request, fence);
326 }
327 
328 static inline struct i915_request *
329 i915_request_get(struct i915_request *rq)
330 {
331 	return to_request(dma_fence_get(&rq->fence));
332 }
333 
334 static inline struct i915_request *
335 i915_request_get_rcu(struct i915_request *rq)
336 {
337 	return to_request(dma_fence_get_rcu(&rq->fence));
338 }
339 
340 static inline void
341 i915_request_put(struct i915_request *rq)
342 {
343 	dma_fence_put(&rq->fence);
344 }
345 
346 int i915_request_await_object(struct i915_request *to,
347 			      struct drm_i915_gem_object *obj,
348 			      bool write);
349 int i915_request_await_dma_fence(struct i915_request *rq,
350 				 struct dma_fence *fence);
351 int i915_request_await_execution(struct i915_request *rq,
352 				 struct dma_fence *fence,
353 				 void (*hook)(struct i915_request *rq,
354 					      struct dma_fence *signal));
355 
356 void i915_request_add(struct i915_request *rq);
357 
358 bool __i915_request_submit(struct i915_request *request);
359 void i915_request_submit(struct i915_request *request);
360 
361 void __i915_request_unsubmit(struct i915_request *request);
362 void i915_request_unsubmit(struct i915_request *request);
363 
364 long i915_request_wait(struct i915_request *rq,
365 		       unsigned int flags,
366 		       long timeout)
367 	__attribute__((nonnull(1)));
368 #define I915_WAIT_INTERRUPTIBLE	BIT(0)
369 #define I915_WAIT_PRIORITY	BIT(1) /* small priority bump for the request */
370 #define I915_WAIT_ALL		BIT(2) /* used by i915_gem_object_wait() */
371 
372 static inline bool i915_request_signaled(const struct i915_request *rq)
373 {
374 	/* The request may live longer than its HWSP, so check flags first! */
375 	return test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &rq->fence.flags);
376 }
377 
378 static inline bool i915_request_is_active(const struct i915_request *rq)
379 {
380 	return test_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags);
381 }
382 
383 static inline bool i915_request_in_priority_queue(const struct i915_request *rq)
384 {
385 	return test_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags);
386 }
387 
388 static inline bool
389 i915_request_has_initial_breadcrumb(const struct i915_request *rq)
390 {
391 	return test_bit(I915_FENCE_FLAG_INITIAL_BREADCRUMB, &rq->fence.flags);
392 }
393 
394 /**
395  * Returns true if seq1 is later than seq2.
396  */
397 static inline bool i915_seqno_passed(u32 seq1, u32 seq2)
398 {
399 	return (s32)(seq1 - seq2) >= 0;
400 }
401 
402 static inline u32 __hwsp_seqno(const struct i915_request *rq)
403 {
404 	const u32 *hwsp = READ_ONCE(rq->hwsp_seqno);
405 
406 	return READ_ONCE(*hwsp);
407 }
408 
409 /**
410  * hwsp_seqno - the current breadcrumb value in the HW status page
411  * @rq: the request, to chase the relevant HW status page
412  *
413  * The emphasis in naming here is that hwsp_seqno() is not a property of the
414  * request, but an indication of the current HW state (associated with this
415  * request). Its value will change as the GPU executes more requests.
416  *
417  * Returns the current breadcrumb value in the associated HW status page (or
418  * the local timeline's equivalent) for this request. The request itself
419  * has the associated breadcrumb value of rq->fence.seqno, when the HW
420  * status page has that breadcrumb or later, this request is complete.
421  */
422 static inline u32 hwsp_seqno(const struct i915_request *rq)
423 {
424 	u32 seqno;
425 
426 	rcu_read_lock(); /* the HWSP may be freed at runtime */
427 	seqno = __hwsp_seqno(rq);
428 	rcu_read_unlock();
429 
430 	return seqno;
431 }
432 
433 static inline bool __i915_request_has_started(const struct i915_request *rq)
434 {
435 	return i915_seqno_passed(hwsp_seqno(rq), rq->fence.seqno - 1);
436 }
437 
438 /**
439  * i915_request_started - check if the request has begun being executed
440  * @rq: the request
441  *
442  * If the timeline is not using initial breadcrumbs, a request is
443  * considered started if the previous request on its timeline (i.e.
444  * context) has been signaled.
445  *
446  * If the timeline is using semaphores, it will also be emitting an
447  * "initial breadcrumb" after the semaphores are complete and just before
448  * it began executing the user payload. A request can therefore be active
449  * on the HW and not yet started as it is still busywaiting on its
450  * dependencies (via HW semaphores).
451  *
452  * If the request has started, its dependencies will have been signaled
453  * (either by fences or by semaphores) and it will have begun processing
454  * the user payload.
455  *
456  * However, even if a request has started, it may have been preempted and
457  * so no longer active, or it may have already completed.
458  *
459  * See also i915_request_is_active().
460  *
461  * Returns true if the request has begun executing the user payload, or
462  * has completed:
463  */
464 static inline bool i915_request_started(const struct i915_request *rq)
465 {
466 	if (i915_request_signaled(rq))
467 		return true;
468 
469 	/* Remember: started but may have since been preempted! */
470 	return __i915_request_has_started(rq);
471 }
472 
473 /**
474  * i915_request_is_running - check if the request may actually be executing
475  * @rq: the request
476  *
477  * Returns true if the request is currently submitted to hardware, has passed
478  * its start point (i.e. the context is setup and not busywaiting). Note that
479  * it may no longer be running by the time the function returns!
480  */
481 static inline bool i915_request_is_running(const struct i915_request *rq)
482 {
483 	if (!i915_request_is_active(rq))
484 		return false;
485 
486 	return __i915_request_has_started(rq);
487 }
488 
489 /**
490  * i915_request_is_ready - check if the request is ready for execution
491  * @rq: the request
492  *
493  * Upon construction, the request is instructed to wait upon various
494  * signals before it is ready to be executed by the HW. That is, we do
495  * not want to start execution and read data before it is written. In practice,
496  * this is controlled with a mixture of interrupts and semaphores. Once
497  * the submit fence is completed, the backend scheduler will place the
498  * request into its queue and from there submit it for execution. So we
499  * can detect when a request is eligible for execution (and is under control
500  * of the scheduler) by querying where it is in any of the scheduler's lists.
501  *
502  * Returns true if the request is ready for execution (it may be inflight),
503  * false otherwise.
504  */
505 static inline bool i915_request_is_ready(const struct i915_request *rq)
506 {
507 	return !list_empty(&rq->sched.link);
508 }
509 
510 static inline bool i915_request_completed(const struct i915_request *rq)
511 {
512 	if (i915_request_signaled(rq))
513 		return true;
514 
515 	return i915_seqno_passed(hwsp_seqno(rq), rq->fence.seqno);
516 }
517 
518 static inline void i915_request_mark_complete(struct i915_request *rq)
519 {
520 	WRITE_ONCE(rq->hwsp_seqno, /* decouple from HWSP */
521 		   (u32 *)&rq->fence.seqno);
522 }
523 
524 static inline bool i915_request_has_waitboost(const struct i915_request *rq)
525 {
526 	return test_bit(I915_FENCE_FLAG_BOOST, &rq->fence.flags);
527 }
528 
529 static inline bool i915_request_has_nopreempt(const struct i915_request *rq)
530 {
531 	/* Preemption should only be disabled very rarely */
532 	return unlikely(test_bit(I915_FENCE_FLAG_NOPREEMPT, &rq->fence.flags));
533 }
534 
535 static inline bool i915_request_has_sentinel(const struct i915_request *rq)
536 {
537 	return unlikely(test_bit(I915_FENCE_FLAG_SENTINEL, &rq->fence.flags));
538 }
539 
540 static inline bool i915_request_on_hold(const struct i915_request *rq)
541 {
542 	return unlikely(test_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags));
543 }
544 
545 static inline void i915_request_set_hold(struct i915_request *rq)
546 {
547 	set_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags);
548 }
549 
550 static inline void i915_request_clear_hold(struct i915_request *rq)
551 {
552 	clear_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags);
553 }
554 
555 static inline struct intel_timeline *
556 i915_request_timeline(const struct i915_request *rq)
557 {
558 	/* Valid only while the request is being constructed (or retired). */
559 	return rcu_dereference_protected(rq->timeline,
560 					 lockdep_is_held(&rcu_access_pointer(rq->timeline)->mutex));
561 }
562 
563 static inline struct i915_gem_context *
564 i915_request_gem_context(const struct i915_request *rq)
565 {
566 	/* Valid only while the request is being constructed (or retired). */
567 	return rcu_dereference_protected(rq->context->gem_context, true);
568 }
569 
570 static inline struct intel_timeline *
571 i915_request_active_timeline(const struct i915_request *rq)
572 {
573 	/*
574 	 * When in use during submission, we are protected by a guarantee that
575 	 * the context/timeline is pinned and must remain pinned until after
576 	 * this submission.
577 	 */
578 	return rcu_dereference_protected(rq->timeline,
579 					 lockdep_is_held(&rq->engine->active.lock));
580 }
581 
582 #endif /* I915_REQUEST_H */
583