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"
33 #include <uapi/drm/i915_drm.h>
36 struct drm_i915_gem_object
;
37 struct drm_i915_gem_request
;
41 struct task_struct
*tsk
;
42 struct drm_i915_gem_request
*request
;
46 struct intel_signal_node
{
48 struct intel_wait wait
;
51 struct i915_dependency
{
52 struct i915_priotree
*signaler
;
53 struct list_head signal_link
;
54 struct list_head wait_link
;
55 struct list_head dfs_link
;
57 #define I915_DEPENDENCY_ALLOC BIT(0)
60 /* Requests exist in a complex web of interdependencies. Each request
61 * has to wait for some other request to complete before it is ready to be run
62 * (e.g. we have to wait until the pixels have been rendering into a texture
63 * before we can copy from it). We track the readiness of a request in terms
64 * of fences, but we also need to keep the dependency tree for the lifetime
65 * of the request (beyond the life of an individual fence). We use the tree
66 * at various points to reorder the requests whilst keeping the requests
67 * in order with respect to their various dependencies.
69 struct i915_priotree
{
70 struct list_head signalers_list
; /* those before us, we depend upon */
71 struct list_head waiters_list
; /* those after us, they depend upon us */
72 struct list_head link
;
77 I915_PRIORITY_MIN
= I915_CONTEXT_MIN_USER_PRIORITY
- 1,
78 I915_PRIORITY_NORMAL
= I915_CONTEXT_DEFAULT_PRIORITY
,
79 I915_PRIORITY_MAX
= I915_CONTEXT_MAX_USER_PRIORITY
+ 1,
81 I915_PRIORITY_INVALID
= INT_MIN
84 struct i915_gem_capture_list
{
85 struct i915_gem_capture_list
*next
;
90 * Request queue structure.
92 * The request queue allows us to note sequence numbers that have been emitted
93 * and may be associated with active buffers to be retired.
95 * By keeping this list, we can avoid having to do questionable sequence
96 * number comparisons on buffer last_read|write_seqno. It also allows an
97 * emission time to be associated with the request for tracking how far ahead
98 * of the GPU the submission is.
100 * When modifying this structure be very aware that we perform a lockless
101 * RCU lookup of it that may race against reallocation of the struct
102 * from the slab freelist. We intentionally do not zero the structure on
103 * allocation so that the lookup can use the dangling pointers (and is
104 * cogniscent that those pointers may be wrong). Instead, everything that
105 * needs to be initialised must be done so explicitly.
107 * The requests are reference counted.
109 struct drm_i915_gem_request
{
110 struct dma_fence fence
;
113 /** On Which ring this request was generated */
114 struct drm_i915_private
*i915
;
117 * Context and ring buffer related to this request
118 * Contexts are refcounted, so when this request is associated with a
119 * context, we must increment the context's refcount, to guarantee that
120 * it persists while any request is linked to it. Requests themselves
121 * are also refcounted, so the request will only be freed when the last
122 * reference to it is dismissed, and the code in
123 * i915_gem_request_free() will then decrement the refcount on the
126 struct i915_gem_context
*ctx
;
127 struct intel_engine_cs
*engine
;
128 struct intel_ring
*ring
;
129 struct intel_timeline
*timeline
;
130 struct intel_signal_node signaling
;
132 /* Fences for the various phases in the request's lifetime.
134 * The submit fence is used to await upon all of the request's
135 * dependencies. When it is signaled, the request is ready to run.
136 * It is used by the driver to then queue the request for execution.
138 struct i915_sw_fence submit
;
139 wait_queue_entry_t submitq
;
140 wait_queue_head_t execute
;
142 /* A list of everyone we wait upon, and everyone who waits upon us.
143 * Even though we will not be submitted to the hardware before the
144 * submit fence is signaled (it waits for all external events as well
145 * as our own requests), the scheduler still needs to know the
146 * dependency tree for the lifetime of the request (from execbuf
147 * to retirement), i.e. bidirectional dependency information for the
148 * request not tied to individual fences.
150 struct i915_priotree priotree
;
151 struct i915_dependency dep
;
153 /** GEM sequence number associated with this request on the
154 * global execution timeline. It is zero when the request is not
155 * on the HW queue (i.e. not on the engine timeline list).
156 * Its value is guarded by the timeline spinlock.
160 /** Position in the ring of the start of the request */
164 * Position in the ring of the start of the postfix.
165 * This is required to calculate the maximum available ring space
166 * without overwriting the postfix.
170 /** Position in the ring of the end of the whole request */
173 /** Position in the ring of the end of any workarounds after the tail */
176 /** Preallocate space in the ring for the emitting the request */
179 /** Batch buffer related to this request if any (used for
180 * error state dump only).
182 struct i915_vma
*batch
;
183 /** Additional buffers requested by userspace to be captured upon
184 * a GPU hang. The vma/obj on this list are protected by their
185 * active reference - all objects on this list must also be
186 * on the active_list (of their final request).
188 struct i915_gem_capture_list
*capture_list
;
189 struct list_head active_list
;
191 /** Time at which this request was emitted, in jiffies. */
192 unsigned long emitted_jiffies
;
196 /** engine->request_list entry for this request */
197 struct list_head link
;
199 /** ring->request_list entry for this request */
200 struct list_head ring_link
;
202 struct drm_i915_file_private
*file_priv
;
203 /** file_priv list entry for this request */
204 struct list_head client_link
;
207 extern const struct dma_fence_ops i915_fence_ops
;
209 static inline bool dma_fence_is_i915(const struct dma_fence
*fence
)
211 return fence
->ops
== &i915_fence_ops
;
214 struct drm_i915_gem_request
* __must_check
215 i915_gem_request_alloc(struct intel_engine_cs
*engine
,
216 struct i915_gem_context
*ctx
);
217 void i915_gem_request_retire_upto(struct drm_i915_gem_request
*req
);
219 static inline struct drm_i915_gem_request
*
220 to_request(struct dma_fence
*fence
)
222 /* We assume that NULL fence/request are interoperable */
223 BUILD_BUG_ON(offsetof(struct drm_i915_gem_request
, fence
) != 0);
224 GEM_BUG_ON(fence
&& !dma_fence_is_i915(fence
));
225 return container_of(fence
, struct drm_i915_gem_request
, fence
);
228 static inline struct drm_i915_gem_request
*
229 i915_gem_request_get(struct drm_i915_gem_request
*req
)
231 return to_request(dma_fence_get(&req
->fence
));
234 static inline struct drm_i915_gem_request
*
235 i915_gem_request_get_rcu(struct drm_i915_gem_request
*req
)
237 return to_request(dma_fence_get_rcu(&req
->fence
));
241 i915_gem_request_put(struct drm_i915_gem_request
*req
)
243 dma_fence_put(&req
->fence
);
246 static inline void i915_gem_request_assign(struct drm_i915_gem_request
**pdst
,
247 struct drm_i915_gem_request
*src
)
250 i915_gem_request_get(src
);
253 i915_gem_request_put(*pdst
);
259 * i915_gem_request_global_seqno - report the current global seqno
260 * @request - the request
262 * A request is assigned a global seqno only when it is on the hardware
263 * execution queue. The global seqno can be used to maintain a list of
264 * requests on the same engine in retirement order, for example for
265 * constructing a priority queue for waiting. Prior to its execution, or
266 * if it is subsequently removed in the event of preemption, its global
267 * seqno is zero. As both insertion and removal from the execution queue
268 * may operate in IRQ context, it is not guarded by the usual struct_mutex
269 * BKL. Instead those relying on the global seqno must be prepared for its
270 * value to change between reads. Only when the request is complete can
271 * the global seqno be stable (due to the memory barriers on submitting
272 * the commands to the hardware to write the breadcrumb, if the HWS shows
273 * that it has passed the global seqno and the global seqno is unchanged
274 * after the read, it is indeed complete).
277 i915_gem_request_global_seqno(const struct drm_i915_gem_request
*request
)
279 return READ_ONCE(request
->global_seqno
);
283 i915_gem_request_await_object(struct drm_i915_gem_request
*to
,
284 struct drm_i915_gem_object
*obj
,
286 int i915_gem_request_await_dma_fence(struct drm_i915_gem_request
*req
,
287 struct dma_fence
*fence
);
289 void __i915_add_request(struct drm_i915_gem_request
*req
, bool flush_caches
);
290 #define i915_add_request(req) \
291 __i915_add_request(req, false)
293 void __i915_gem_request_submit(struct drm_i915_gem_request
*request
);
294 void i915_gem_request_submit(struct drm_i915_gem_request
*request
);
296 void __i915_gem_request_unsubmit(struct drm_i915_gem_request
*request
);
297 void i915_gem_request_unsubmit(struct drm_i915_gem_request
*request
);
299 struct intel_rps_client
;
300 #define NO_WAITBOOST ERR_PTR(-1)
301 #define IS_RPS_CLIENT(p) (!IS_ERR(p))
302 #define IS_RPS_USER(p) (!IS_ERR_OR_NULL(p))
304 long i915_wait_request(struct drm_i915_gem_request
*req
,
307 __attribute__((nonnull(1)));
308 #define I915_WAIT_INTERRUPTIBLE BIT(0)
309 #define I915_WAIT_LOCKED BIT(1) /* struct_mutex held, handle GPU reset */
310 #define I915_WAIT_ALL BIT(2) /* used by i915_gem_object_wait() */
312 static inline u32
intel_engine_get_seqno(struct intel_engine_cs
*engine
);
315 * Returns true if seq1 is later than seq2.
317 static inline bool i915_seqno_passed(u32 seq1
, u32 seq2
)
319 return (s32
)(seq1
- seq2
) >= 0;
323 __i915_gem_request_completed(const struct drm_i915_gem_request
*req
, u32 seqno
)
326 return i915_seqno_passed(intel_engine_get_seqno(req
->engine
), seqno
) &&
327 seqno
== i915_gem_request_global_seqno(req
);
331 i915_gem_request_completed(const struct drm_i915_gem_request
*req
)
335 seqno
= i915_gem_request_global_seqno(req
);
339 return __i915_gem_request_completed(req
, seqno
);
342 /* We treat requests as fences. This is not be to confused with our
343 * "fence registers" but pipeline synchronisation objects ala GL_ARB_sync.
344 * We use the fences to synchronize access from the CPU with activity on the
345 * GPU, for example, we should not rewrite an object's PTE whilst the GPU
346 * is reading them. We also track fences at a higher level to provide
347 * implicit synchronisation around GEM objects, e.g. set-domain will wait
348 * for outstanding GPU rendering before marking the object ready for CPU
349 * access, or a pageflip will wait until the GPU is complete before showing
350 * the frame on the scanout.
352 * In order to use a fence, the object must track the fence it needs to
353 * serialise with. For example, GEM objects want to track both read and
354 * write access so that we can perform concurrent read operations between
355 * the CPU and GPU engines, as well as waiting for all rendering to
356 * complete, or waiting for the last GPU user of a "fence register". The
357 * object then embeds a #i915_gem_active to track the most recent (in
358 * retirement order) request relevant for the desired mode of access.
359 * The #i915_gem_active is updated with i915_gem_active_set() to track the
360 * most recent fence request, typically this is done as part of
361 * i915_vma_move_to_active().
363 * When the #i915_gem_active completes (is retired), it will
364 * signal its completion to the owner through a callback as well as mark
365 * itself as idle (i915_gem_active.request == NULL). The owner
366 * can then perform any action, such as delayed freeing of an active
367 * resource including itself.
369 struct i915_gem_active
;
371 typedef void (*i915_gem_retire_fn
)(struct i915_gem_active
*,
372 struct drm_i915_gem_request
*);
374 struct i915_gem_active
{
375 struct drm_i915_gem_request __rcu
*request
;
376 struct list_head link
;
377 i915_gem_retire_fn retire
;
380 void i915_gem_retire_noop(struct i915_gem_active
*,
381 struct drm_i915_gem_request
*request
);
384 * init_request_active - prepares the activity tracker for use
385 * @active - the active tracker
386 * @func - a callback when then the tracker is retired (becomes idle),
389 * init_request_active() prepares the embedded @active struct for use as
390 * an activity tracker, that is for tracking the last known active request
391 * associated with it. When the last request becomes idle, when it is retired
392 * after completion, the optional callback @func is invoked.
395 init_request_active(struct i915_gem_active
*active
,
396 i915_gem_retire_fn retire
)
398 INIT_LIST_HEAD(&active
->link
);
399 active
->retire
= retire
?: i915_gem_retire_noop
;
403 * i915_gem_active_set - updates the tracker to watch the current request
404 * @active - the active tracker
405 * @request - the request to watch
407 * i915_gem_active_set() watches the given @request for completion. Whilst
408 * that @request is busy, the @active reports busy. When that @request is
409 * retired, the @active tracker is updated to report idle.
412 i915_gem_active_set(struct i915_gem_active
*active
,
413 struct drm_i915_gem_request
*request
)
415 list_move(&active
->link
, &request
->active_list
);
416 rcu_assign_pointer(active
->request
, request
);
420 * i915_gem_active_set_retire_fn - updates the retirement callback
421 * @active - the active tracker
422 * @fn - the routine called when the request is retired
423 * @mutex - struct_mutex used to guard retirements
425 * i915_gem_active_set_retire_fn() updates the function pointer that
426 * is called when the final request associated with the @active tracker
430 i915_gem_active_set_retire_fn(struct i915_gem_active
*active
,
431 i915_gem_retire_fn fn
,
434 lockdep_assert_held(mutex
);
435 active
->retire
= fn
?: i915_gem_retire_noop
;
438 static inline struct drm_i915_gem_request
*
439 __i915_gem_active_peek(const struct i915_gem_active
*active
)
441 /* Inside the error capture (running with the driver in an unknown
442 * state), we want to bend the rules slightly (a lot).
444 * Work is in progress to make it safer, in the meantime this keeps
445 * the known issue from spamming the logs.
447 return rcu_dereference_protected(active
->request
, 1);
451 * i915_gem_active_raw - return the active request
452 * @active - the active tracker
454 * i915_gem_active_raw() returns the current request being tracked, or NULL.
455 * It does not obtain a reference on the request for the caller, so the caller
456 * must hold struct_mutex.
458 static inline struct drm_i915_gem_request
*
459 i915_gem_active_raw(const struct i915_gem_active
*active
, struct lock
*mutex
)
461 return rcu_dereference_protected(active
->request
,
462 lockdep_is_held(mutex
));
466 * i915_gem_active_peek - report the active request being monitored
467 * @active - the active tracker
469 * i915_gem_active_peek() returns the current request being tracked if
470 * still active, or NULL. It does not obtain a reference on the request
471 * for the caller, so the caller must hold struct_mutex.
473 static inline struct drm_i915_gem_request
*
474 i915_gem_active_peek(const struct i915_gem_active
*active
, struct lock
*mutex
)
476 struct drm_i915_gem_request
*request
;
478 request
= i915_gem_active_raw(active
, mutex
);
479 if (!request
|| i915_gem_request_completed(request
))
486 * i915_gem_active_get - return a reference to the active request
487 * @active - the active tracker
489 * i915_gem_active_get() returns a reference to the active request, or NULL
490 * if the active tracker is idle. The caller must hold struct_mutex.
492 static inline struct drm_i915_gem_request
*
493 i915_gem_active_get(const struct i915_gem_active
*active
, struct lock
*mutex
)
495 return i915_gem_request_get(i915_gem_active_peek(active
, mutex
));
499 * __i915_gem_active_get_rcu - return a reference to the active request
500 * @active - the active tracker
502 * __i915_gem_active_get() returns a reference to the active request, or NULL
503 * if the active tracker is idle. The caller must hold the RCU read lock, but
504 * the returned pointer is safe to use outside of RCU.
506 static inline struct drm_i915_gem_request
*
507 __i915_gem_active_get_rcu(const struct i915_gem_active
*active
)
509 /* Performing a lockless retrieval of the active request is super
510 * tricky. SLAB_TYPESAFE_BY_RCU merely guarantees that the backing
511 * slab of request objects will not be freed whilst we hold the
512 * RCU read lock. It does not guarantee that the request itself
513 * will not be freed and then *reused*. Viz,
517 * req = active.request
518 * retire(req) -> free(req);
519 * (req is now first on the slab freelist)
520 * active.request = NULL
522 * req = new submission on a new object
525 * To prevent the request from being reused whilst the caller
526 * uses it, we take a reference like normal. Whilst acquiring
527 * the reference we check that it is not in a destroyed state
528 * (refcnt == 0). That prevents the request being reallocated
529 * whilst the caller holds on to it. To check that the request
530 * was not reallocated as we acquired the reference we have to
531 * check that our request remains the active request across
532 * the lookup, in the same manner as a seqlock. The visibility
533 * of the pointer versus the reference counting is controlled
534 * by using RCU barriers (rcu_dereference and rcu_assign_pointer).
536 * In the middle of all that, we inspect whether the request is
537 * complete. Retiring is lazy so the request may be completed long
538 * before the active tracker is updated. Querying whether the
539 * request is complete is far cheaper (as it involves no locked
540 * instructions setting cachelines to exclusive) than acquiring
541 * the reference, so we do it first. The RCU read lock ensures the
542 * pointer dereference is valid, but does not ensure that the
543 * seqno nor HWS is the right one! However, if the request was
544 * reallocated, that means the active tracker's request was complete.
545 * If the new request is also complete, then both are and we can
546 * just report the active tracker is idle. If the new request is
547 * incomplete, then we acquire a reference on it and check that
548 * it remained the active request.
550 * It is then imperative that we do not zero the request on
551 * reallocation, so that we can chase the dangling pointers!
552 * See i915_gem_request_alloc().
555 struct drm_i915_gem_request
*request
;
557 request
= rcu_dereference(active
->request
);
558 if (!request
|| i915_gem_request_completed(request
))
561 /* An especially silly compiler could decide to recompute the
562 * result of i915_gem_request_completed, more specifically
563 * re-emit the load for request->fence.seqno. A race would catch
564 * a later seqno value, which could flip the result from true to
565 * false. Which means part of the instructions below might not
566 * be executed, while later on instructions are executed. Due to
567 * barriers within the refcounting the inconsistency can't reach
568 * past the call to i915_gem_request_get_rcu, but not executing
569 * that while still executing i915_gem_request_put() creates
570 * havoc enough. Prevent this with a compiler barrier.
574 request
= i915_gem_request_get_rcu(request
);
576 /* What stops the following rcu_access_pointer() from occurring
577 * before the above i915_gem_request_get_rcu()? If we were
578 * to read the value before pausing to get the reference to
579 * the request, we may not notice a change in the active
582 * The rcu_access_pointer() is a mere compiler barrier, which
583 * means both the CPU and compiler are free to perform the
584 * memory read without constraint. The compiler only has to
585 * ensure that any operations after the rcu_access_pointer()
586 * occur afterwards in program order. This means the read may
587 * be performed earlier by an out-of-order CPU, or adventurous
590 * The atomic operation at the heart of
591 * i915_gem_request_get_rcu(), see dma_fence_get_rcu(), is
592 * atomic_inc_not_zero() which is only a full memory barrier
593 * when successful. That is, if i915_gem_request_get_rcu()
594 * returns the request (and so with the reference counted
595 * incremented) then the following read for rcu_access_pointer()
596 * must occur after the atomic operation and so confirm
597 * that this request is the one currently being tracked.
599 * The corresponding write barrier is part of
600 * rcu_assign_pointer().
602 if (!request
|| request
== rcu_access_pointer(active
->request
))
603 return rcu_pointer_handoff(request
);
605 i915_gem_request_put(request
);
610 * i915_gem_active_get_unlocked - return a reference to the active request
611 * @active - the active tracker
613 * i915_gem_active_get_unlocked() returns a reference to the active request,
614 * or NULL if the active tracker is idle. The reference is obtained under RCU,
615 * so no locking is required by the caller.
617 * The reference should be freed with i915_gem_request_put().
619 static inline struct drm_i915_gem_request
*
620 i915_gem_active_get_unlocked(const struct i915_gem_active
*active
)
622 struct drm_i915_gem_request
*request
;
625 request
= __i915_gem_active_get_rcu(active
);
632 * i915_gem_active_isset - report whether the active tracker is assigned
633 * @active - the active tracker
635 * i915_gem_active_isset() returns true if the active tracker is currently
636 * assigned to a request. Due to the lazy retiring, that request may be idle
637 * and this may report stale information.
640 i915_gem_active_isset(const struct i915_gem_active
*active
)
642 return rcu_access_pointer(active
->request
);
646 * i915_gem_active_wait - waits until the request is completed
647 * @active - the active request on which to wait
648 * @flags - how to wait
649 * @timeout - how long to wait at most
650 * @rps - userspace client to charge for a waitboost
652 * i915_gem_active_wait() waits until the request is completed before
653 * returning, without requiring any locks to be held. Note that it does not
654 * retire any requests before returning.
656 * This function relies on RCU in order to acquire the reference to the active
657 * request without holding any locks. See __i915_gem_active_get_rcu() for the
658 * glory details on how that is managed. Once the reference is acquired, we
659 * can then wait upon the request, and afterwards release our reference,
660 * free of any locking.
662 * This function wraps i915_wait_request(), see it for the full details on
665 * Returns 0 if successful, or a negative error code.
668 i915_gem_active_wait(const struct i915_gem_active
*active
, unsigned int flags
)
670 struct drm_i915_gem_request
*request
;
673 request
= i915_gem_active_get_unlocked(active
);
675 ret
= i915_wait_request(request
, flags
, MAX_SCHEDULE_TIMEOUT
);
676 i915_gem_request_put(request
);
679 return ret
< 0 ? ret
: 0;
683 * i915_gem_active_retire - waits until the request is retired
684 * @active - the active request on which to wait
686 * i915_gem_active_retire() waits until the request is completed,
687 * and then ensures that at least the retirement handler for this
688 * @active tracker is called before returning. If the @active
689 * tracker is idle, the function returns immediately.
691 static inline int __must_check
692 i915_gem_active_retire(struct i915_gem_active
*active
,
695 struct drm_i915_gem_request
*request
;
698 request
= i915_gem_active_raw(active
, mutex
);
702 ret
= i915_wait_request(request
,
703 I915_WAIT_INTERRUPTIBLE
| I915_WAIT_LOCKED
,
704 MAX_SCHEDULE_TIMEOUT
);
708 list_del_init(&active
->link
);
709 RCU_INIT_POINTER(active
->request
, NULL
);
711 active
->retire(active
, request
);
716 #define for_each_active(mask, idx) \
717 for (; mask ? idx = ffs(mask) - 1, 1 : 0; mask &= ~BIT(idx))
719 #endif /* I915_GEM_REQUEST_H */