2 * Copyright © 2008-2015 Intel Corporation
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:
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
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
25 #ifndef I915_GEM_REQUEST_H
26 #define I915_GEM_REQUEST_H
28 #include <linux/dma-fence.h>
31 #include "i915_sw_fence.h"
34 struct drm_i915_gem_object;
35 struct drm_i915_gem_request;
39 struct task_struct *tsk;
40 struct drm_i915_gem_request *request;
44 struct intel_signal_node {
46 struct intel_wait wait;
49 struct i915_dependency {
50 struct i915_priotree *signaler;
51 struct list_head signal_link;
52 struct list_head wait_link;
53 struct list_head dfs_link;
55 #define I915_DEPENDENCY_ALLOC BIT(0)
58 /* Requests exist in a complex web of interdependencies. Each request
59 * has to wait for some other request to complete before it is ready to be run
60 * (e.g. we have to wait until the pixels have been rendering into a texture
61 * before we can copy from it). We track the readiness of a request in terms
62 * of fences, but we also need to keep the dependency tree for the lifetime
63 * of the request (beyond the life of an individual fence). We use the tree
64 * at various points to reorder the requests whilst keeping the requests
65 * in order with respect to their various dependencies.
67 struct i915_priotree {
68 struct list_head signalers_list; /* those before us, we depend upon */
69 struct list_head waiters_list; /* those after us, they depend upon us */
72 #define I915_PRIORITY_MAX 1024
73 #define I915_PRIORITY_NORMAL 0
74 #define I915_PRIORITY_MIN (-I915_PRIORITY_MAX)
77 struct i915_gem_capture_list {
78 struct i915_gem_capture_list *next;
83 * Request queue structure.
85 * The request queue allows us to note sequence numbers that have been emitted
86 * and may be associated with active buffers to be retired.
88 * By keeping this list, we can avoid having to do questionable sequence
89 * number comparisons on buffer last_read|write_seqno. It also allows an
90 * emission time to be associated with the request for tracking how far ahead
91 * of the GPU the submission is.
93 * When modifying this structure be very aware that we perform a lockless
94 * RCU lookup of it that may race against reallocation of the struct
95 * from the slab freelist. We intentionally do not zero the structure on
96 * allocation so that the lookup can use the dangling pointers (and is
97 * cogniscent that those pointers may be wrong). Instead, everything that
98 * needs to be initialised must be done so explicitly.
100 * The requests are reference counted.
102 struct drm_i915_gem_request {
103 struct dma_fence fence;
106 /** On Which ring this request was generated */
107 struct drm_i915_private *i915;
110 * Context and ring buffer related to this request
111 * Contexts are refcounted, so when this request is associated with a
112 * context, we must increment the context's refcount, to guarantee that
113 * it persists while any request is linked to it. Requests themselves
114 * are also refcounted, so the request will only be freed when the last
115 * reference to it is dismissed, and the code in
116 * i915_gem_request_free() will then decrement the refcount on the
119 struct i915_gem_context *ctx;
120 struct intel_engine_cs *engine;
121 struct intel_ring *ring;
122 struct intel_timeline *timeline;
123 struct intel_signal_node signaling;
125 /* Fences for the various phases in the request's lifetime.
127 * The submit fence is used to await upon all of the request's
128 * dependencies. When it is signaled, the request is ready to run.
129 * It is used by the driver to then queue the request for execution.
131 struct i915_sw_fence submit;
132 wait_queue_t submitq;
133 wait_queue_head_t execute;
135 /* A list of everyone we wait upon, and everyone who waits upon us.
136 * Even though we will not be submitted to the hardware before the
137 * submit fence is signaled (it waits for all external events as well
138 * as our own requests), the scheduler still needs to know the
139 * dependency tree for the lifetime of the request (from execbuf
140 * to retirement), i.e. bidirectional dependency information for the
141 * request not tied to individual fences.
143 struct i915_priotree priotree;
144 struct i915_dependency dep;
146 /** GEM sequence number associated with this request on the
147 * global execution timeline. It is zero when the request is not
148 * on the HW queue (i.e. not on the engine timeline list).
149 * Its value is guarded by the timeline spinlock.
153 /** Position in the ring of the start of the request */
157 * Position in the ring of the start of the postfix.
158 * This is required to calculate the maximum available ring space
159 * without overwriting the postfix.
163 /** Position in the ring of the end of the whole request */
166 /** Position in the ring of the end of any workarounds after the tail */
169 /** Preallocate space in the ring for the emitting the request */
172 /** Batch buffer related to this request if any (used for
173 * error state dump only).
175 struct i915_vma *batch;
176 /** Additional buffers requested by userspace to be captured upon
177 * a GPU hang. The vma/obj on this list are protected by their
178 * active reference - all objects on this list must also be
179 * on the active_list (of their final request).
181 struct i915_gem_capture_list *capture_list;
182 struct list_head active_list;
184 /** Time at which this request was emitted, in jiffies. */
185 unsigned long emitted_jiffies;
187 /** engine->request_list entry for this request */
188 struct list_head link;
190 /** ring->request_list entry for this request */
191 struct list_head ring_link;
193 struct drm_i915_file_private *file_priv;
194 /** file_priv list entry for this request */
195 struct list_head client_link;
198 extern const struct dma_fence_ops i915_fence_ops;
200 static inline bool dma_fence_is_i915(const struct dma_fence *fence)
202 return fence->ops == &i915_fence_ops;
205 struct drm_i915_gem_request * __must_check
206 i915_gem_request_alloc(struct intel_engine_cs *engine,
207 struct i915_gem_context *ctx);
208 void i915_gem_request_retire_upto(struct drm_i915_gem_request *req);
210 static inline struct drm_i915_gem_request *
211 to_request(struct dma_fence *fence)
213 /* We assume that NULL fence/request are interoperable */
214 BUILD_BUG_ON(offsetof(struct drm_i915_gem_request, fence) != 0);
215 GEM_BUG_ON(fence && !dma_fence_is_i915(fence));
216 return container_of(fence, struct drm_i915_gem_request, fence);
219 static inline struct drm_i915_gem_request *
220 i915_gem_request_get(struct drm_i915_gem_request *req)
222 return to_request(dma_fence_get(&req->fence));
225 static inline struct drm_i915_gem_request *
226 i915_gem_request_get_rcu(struct drm_i915_gem_request *req)
228 return to_request(dma_fence_get_rcu(&req->fence));
232 i915_gem_request_put(struct drm_i915_gem_request *req)
234 dma_fence_put(&req->fence);
237 static inline void i915_gem_request_assign(struct drm_i915_gem_request **pdst,
238 struct drm_i915_gem_request *src)
241 i915_gem_request_get(src);
244 i915_gem_request_put(*pdst);
250 * i915_gem_request_global_seqno - report the current global seqno
251 * @request - the request
253 * A request is assigned a global seqno only when it is on the hardware
254 * execution queue. The global seqno can be used to maintain a list of
255 * requests on the same engine in retirement order, for example for
256 * constructing a priority queue for waiting. Prior to its execution, or
257 * if it is subsequently removed in the event of preemption, its global
258 * seqno is zero. As both insertion and removal from the execution queue
259 * may operate in IRQ context, it is not guarded by the usual struct_mutex
260 * BKL. Instead those relying on the global seqno must be prepared for its
261 * value to change between reads. Only when the request is complete can
262 * the global seqno be stable (due to the memory barriers on submitting
263 * the commands to the hardware to write the breadcrumb, if the HWS shows
264 * that it has passed the global seqno and the global seqno is unchanged
265 * after the read, it is indeed complete).
268 i915_gem_request_global_seqno(const struct drm_i915_gem_request *request)
270 return READ_ONCE(request->global_seqno);
274 i915_gem_request_await_object(struct drm_i915_gem_request *to,
275 struct drm_i915_gem_object *obj,
277 int i915_gem_request_await_dma_fence(struct drm_i915_gem_request *req,
278 struct dma_fence *fence);
280 void __i915_add_request(struct drm_i915_gem_request *req, bool flush_caches);
281 #define i915_add_request(req) \
282 __i915_add_request(req, false)
284 void __i915_gem_request_submit(struct drm_i915_gem_request *request);
285 void i915_gem_request_submit(struct drm_i915_gem_request *request);
287 void __i915_gem_request_unsubmit(struct drm_i915_gem_request *request);
288 void i915_gem_request_unsubmit(struct drm_i915_gem_request *request);
290 struct intel_rps_client;
291 #define NO_WAITBOOST ERR_PTR(-1)
292 #define IS_RPS_CLIENT(p) (!IS_ERR(p))
293 #define IS_RPS_USER(p) (!IS_ERR_OR_NULL(p))
295 long i915_wait_request(struct drm_i915_gem_request *req,
298 __attribute__((nonnull(1)));
299 #define I915_WAIT_INTERRUPTIBLE BIT(0)
300 #define I915_WAIT_LOCKED BIT(1) /* struct_mutex held, handle GPU reset */
301 #define I915_WAIT_ALL BIT(2) /* used by i915_gem_object_wait() */
303 static inline u32 intel_engine_get_seqno(struct intel_engine_cs *engine);
306 * Returns true if seq1 is later than seq2.
308 static inline bool i915_seqno_passed(u32 seq1, u32 seq2)
310 return (s32)(seq1 - seq2) >= 0;
314 __i915_gem_request_started(const struct drm_i915_gem_request *req, u32 seqno)
317 return i915_seqno_passed(intel_engine_get_seqno(req->engine),
322 i915_gem_request_started(const struct drm_i915_gem_request *req)
326 seqno = i915_gem_request_global_seqno(req);
330 return __i915_gem_request_started(req, seqno);
334 __i915_gem_request_completed(const struct drm_i915_gem_request *req, u32 seqno)
337 return i915_seqno_passed(intel_engine_get_seqno(req->engine), seqno) &&
338 seqno == i915_gem_request_global_seqno(req);
342 i915_gem_request_completed(const struct drm_i915_gem_request *req)
346 seqno = i915_gem_request_global_seqno(req);
350 return __i915_gem_request_completed(req, seqno);
353 bool __i915_spin_request(const struct drm_i915_gem_request *request,
354 u32 seqno, int state, unsigned long timeout_us);
355 static inline bool i915_spin_request(const struct drm_i915_gem_request *request,
356 int state, unsigned long timeout_us)
360 seqno = i915_gem_request_global_seqno(request);
364 return (__i915_gem_request_started(request, seqno) &&
365 __i915_spin_request(request, seqno, state, timeout_us));
368 /* We treat requests as fences. This is not be to confused with our
369 * "fence registers" but pipeline synchronisation objects ala GL_ARB_sync.
370 * We use the fences to synchronize access from the CPU with activity on the
371 * GPU, for example, we should not rewrite an object's PTE whilst the GPU
372 * is reading them. We also track fences at a higher level to provide
373 * implicit synchronisation around GEM objects, e.g. set-domain will wait
374 * for outstanding GPU rendering before marking the object ready for CPU
375 * access, or a pageflip will wait until the GPU is complete before showing
376 * the frame on the scanout.
378 * In order to use a fence, the object must track the fence it needs to
379 * serialise with. For example, GEM objects want to track both read and
380 * write access so that we can perform concurrent read operations between
381 * the CPU and GPU engines, as well as waiting for all rendering to
382 * complete, or waiting for the last GPU user of a "fence register". The
383 * object then embeds a #i915_gem_active to track the most recent (in
384 * retirement order) request relevant for the desired mode of access.
385 * The #i915_gem_active is updated with i915_gem_active_set() to track the
386 * most recent fence request, typically this is done as part of
387 * i915_vma_move_to_active().
389 * When the #i915_gem_active completes (is retired), it will
390 * signal its completion to the owner through a callback as well as mark
391 * itself as idle (i915_gem_active.request == NULL). The owner
392 * can then perform any action, such as delayed freeing of an active
393 * resource including itself.
395 struct i915_gem_active;
397 typedef void (*i915_gem_retire_fn)(struct i915_gem_active *,
398 struct drm_i915_gem_request *);
400 struct i915_gem_active {
401 struct drm_i915_gem_request __rcu *request;
402 struct list_head link;
403 i915_gem_retire_fn retire;
406 void i915_gem_retire_noop(struct i915_gem_active *,
407 struct drm_i915_gem_request *request);
410 * init_request_active - prepares the activity tracker for use
411 * @active - the active tracker
412 * @func - a callback when then the tracker is retired (becomes idle),
415 * init_request_active() prepares the embedded @active struct for use as
416 * an activity tracker, that is for tracking the last known active request
417 * associated with it. When the last request becomes idle, when it is retired
418 * after completion, the optional callback @func is invoked.
421 init_request_active(struct i915_gem_active *active,
422 i915_gem_retire_fn retire)
424 INIT_LIST_HEAD(&active->link);
425 active->retire = retire ?: i915_gem_retire_noop;
429 * i915_gem_active_set - updates the tracker to watch the current request
430 * @active - the active tracker
431 * @request - the request to watch
433 * i915_gem_active_set() watches the given @request for completion. Whilst
434 * that @request is busy, the @active reports busy. When that @request is
435 * retired, the @active tracker is updated to report idle.
438 i915_gem_active_set(struct i915_gem_active *active,
439 struct drm_i915_gem_request *request)
441 list_move(&active->link, &request->active_list);
442 rcu_assign_pointer(active->request, request);
446 * i915_gem_active_set_retire_fn - updates the retirement callback
447 * @active - the active tracker
448 * @fn - the routine called when the request is retired
449 * @mutex - struct_mutex used to guard retirements
451 * i915_gem_active_set_retire_fn() updates the function pointer that
452 * is called when the final request associated with the @active tracker
456 i915_gem_active_set_retire_fn(struct i915_gem_active *active,
457 i915_gem_retire_fn fn,
460 lockdep_assert_held(mutex);
461 active->retire = fn ?: i915_gem_retire_noop;
464 static inline struct drm_i915_gem_request *
465 __i915_gem_active_peek(const struct i915_gem_active *active)
467 /* Inside the error capture (running with the driver in an unknown
468 * state), we want to bend the rules slightly (a lot).
470 * Work is in progress to make it safer, in the meantime this keeps
471 * the known issue from spamming the logs.
473 return rcu_dereference_protected(active->request, 1);
477 * i915_gem_active_raw - return the active request
478 * @active - the active tracker
480 * i915_gem_active_raw() returns the current request being tracked, or NULL.
481 * It does not obtain a reference on the request for the caller, so the caller
482 * must hold struct_mutex.
484 static inline struct drm_i915_gem_request *
485 i915_gem_active_raw(const struct i915_gem_active *active, struct mutex *mutex)
487 return rcu_dereference_protected(active->request,
488 lockdep_is_held(mutex));
492 * i915_gem_active_peek - report the active request being monitored
493 * @active - the active tracker
495 * i915_gem_active_peek() returns the current request being tracked if
496 * still active, or NULL. It does not obtain a reference on the request
497 * for the caller, so the caller must hold struct_mutex.
499 static inline struct drm_i915_gem_request *
500 i915_gem_active_peek(const struct i915_gem_active *active, struct mutex *mutex)
502 struct drm_i915_gem_request *request;
504 request = i915_gem_active_raw(active, mutex);
505 if (!request || i915_gem_request_completed(request))
512 * i915_gem_active_get - return a reference to the active request
513 * @active - the active tracker
515 * i915_gem_active_get() returns a reference to the active request, or NULL
516 * if the active tracker is idle. The caller must hold struct_mutex.
518 static inline struct drm_i915_gem_request *
519 i915_gem_active_get(const struct i915_gem_active *active, struct mutex *mutex)
521 return i915_gem_request_get(i915_gem_active_peek(active, mutex));
525 * __i915_gem_active_get_rcu - return a reference to the active request
526 * @active - the active tracker
528 * __i915_gem_active_get() returns a reference to the active request, or NULL
529 * if the active tracker is idle. The caller must hold the RCU read lock, but
530 * the returned pointer is safe to use outside of RCU.
532 static inline struct drm_i915_gem_request *
533 __i915_gem_active_get_rcu(const struct i915_gem_active *active)
535 /* Performing a lockless retrieval of the active request is super
536 * tricky. SLAB_DESTROY_BY_RCU merely guarantees that the backing
537 * slab of request objects will not be freed whilst we hold the
538 * RCU read lock. It does not guarantee that the request itself
539 * will not be freed and then *reused*. Viz,
543 * req = active.request
544 * retire(req) -> free(req);
545 * (req is now first on the slab freelist)
546 * active.request = NULL
548 * req = new submission on a new object
551 * To prevent the request from being reused whilst the caller
552 * uses it, we take a reference like normal. Whilst acquiring
553 * the reference we check that it is not in a destroyed state
554 * (refcnt == 0). That prevents the request being reallocated
555 * whilst the caller holds on to it. To check that the request
556 * was not reallocated as we acquired the reference we have to
557 * check that our request remains the active request across
558 * the lookup, in the same manner as a seqlock. The visibility
559 * of the pointer versus the reference counting is controlled
560 * by using RCU barriers (rcu_dereference and rcu_assign_pointer).
562 * In the middle of all that, we inspect whether the request is
563 * complete. Retiring is lazy so the request may be completed long
564 * before the active tracker is updated. Querying whether the
565 * request is complete is far cheaper (as it involves no locked
566 * instructions setting cachelines to exclusive) than acquiring
567 * the reference, so we do it first. The RCU read lock ensures the
568 * pointer dereference is valid, but does not ensure that the
569 * seqno nor HWS is the right one! However, if the request was
570 * reallocated, that means the active tracker's request was complete.
571 * If the new request is also complete, then both are and we can
572 * just report the active tracker is idle. If the new request is
573 * incomplete, then we acquire a reference on it and check that
574 * it remained the active request.
576 * It is then imperative that we do not zero the request on
577 * reallocation, so that we can chase the dangling pointers!
578 * See i915_gem_request_alloc().
581 struct drm_i915_gem_request *request;
583 request = rcu_dereference(active->request);
584 if (!request || i915_gem_request_completed(request))
587 /* An especially silly compiler could decide to recompute the
588 * result of i915_gem_request_completed, more specifically
589 * re-emit the load for request->fence.seqno. A race would catch
590 * a later seqno value, which could flip the result from true to
591 * false. Which means part of the instructions below might not
592 * be executed, while later on instructions are executed. Due to
593 * barriers within the refcounting the inconsistency can't reach
594 * past the call to i915_gem_request_get_rcu, but not executing
595 * that while still executing i915_gem_request_put() creates
596 * havoc enough. Prevent this with a compiler barrier.
600 request = i915_gem_request_get_rcu(request);
602 /* What stops the following rcu_access_pointer() from occurring
603 * before the above i915_gem_request_get_rcu()? If we were
604 * to read the value before pausing to get the reference to
605 * the request, we may not notice a change in the active
608 * The rcu_access_pointer() is a mere compiler barrier, which
609 * means both the CPU and compiler are free to perform the
610 * memory read without constraint. The compiler only has to
611 * ensure that any operations after the rcu_access_pointer()
612 * occur afterwards in program order. This means the read may
613 * be performed earlier by an out-of-order CPU, or adventurous
616 * The atomic operation at the heart of
617 * i915_gem_request_get_rcu(), see dma_fence_get_rcu(), is
618 * atomic_inc_not_zero() which is only a full memory barrier
619 * when successful. That is, if i915_gem_request_get_rcu()
620 * returns the request (and so with the reference counted
621 * incremented) then the following read for rcu_access_pointer()
622 * must occur after the atomic operation and so confirm
623 * that this request is the one currently being tracked.
625 * The corresponding write barrier is part of
626 * rcu_assign_pointer().
628 if (!request || request == rcu_access_pointer(active->request))
629 return rcu_pointer_handoff(request);
631 i915_gem_request_put(request);
636 * i915_gem_active_get_unlocked - return a reference to the active request
637 * @active - the active tracker
639 * i915_gem_active_get_unlocked() returns a reference to the active request,
640 * or NULL if the active tracker is idle. The reference is obtained under RCU,
641 * so no locking is required by the caller.
643 * The reference should be freed with i915_gem_request_put().
645 static inline struct drm_i915_gem_request *
646 i915_gem_active_get_unlocked(const struct i915_gem_active *active)
648 struct drm_i915_gem_request *request;
651 request = __i915_gem_active_get_rcu(active);
658 * i915_gem_active_isset - report whether the active tracker is assigned
659 * @active - the active tracker
661 * i915_gem_active_isset() returns true if the active tracker is currently
662 * assigned to a request. Due to the lazy retiring, that request may be idle
663 * and this may report stale information.
666 i915_gem_active_isset(const struct i915_gem_active *active)
668 return rcu_access_pointer(active->request);
672 * i915_gem_active_wait - waits until the request is completed
673 * @active - the active request on which to wait
674 * @flags - how to wait
675 * @timeout - how long to wait at most
676 * @rps - userspace client to charge for a waitboost
678 * i915_gem_active_wait() waits until the request is completed before
679 * returning, without requiring any locks to be held. Note that it does not
680 * retire any requests before returning.
682 * This function relies on RCU in order to acquire the reference to the active
683 * request without holding any locks. See __i915_gem_active_get_rcu() for the
684 * glory details on how that is managed. Once the reference is acquired, we
685 * can then wait upon the request, and afterwards release our reference,
686 * free of any locking.
688 * This function wraps i915_wait_request(), see it for the full details on
691 * Returns 0 if successful, or a negative error code.
694 i915_gem_active_wait(const struct i915_gem_active *active, unsigned int flags)
696 struct drm_i915_gem_request *request;
699 request = i915_gem_active_get_unlocked(active);
701 ret = i915_wait_request(request, flags, MAX_SCHEDULE_TIMEOUT);
702 i915_gem_request_put(request);
705 return ret < 0 ? ret : 0;
709 * i915_gem_active_retire - waits until the request is retired
710 * @active - the active request on which to wait
712 * i915_gem_active_retire() waits until the request is completed,
713 * and then ensures that at least the retirement handler for this
714 * @active tracker is called before returning. If the @active
715 * tracker is idle, the function returns immediately.
717 static inline int __must_check
718 i915_gem_active_retire(struct i915_gem_active *active,
721 struct drm_i915_gem_request *request;
724 request = i915_gem_active_raw(active, mutex);
728 ret = i915_wait_request(request,
729 I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
730 MAX_SCHEDULE_TIMEOUT);
734 list_del_init(&active->link);
735 RCU_INIT_POINTER(active->request, NULL);
737 active->retire(active, request);
742 #define for_each_active(mask, idx) \
743 for (; mask ? idx = ffs(mask) - 1, 1 : 0; mask &= ~BIT(idx))
745 #endif /* I915_GEM_REQUEST_H */