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/fence.h>
31 #include "i915_sw_fence.h"
35 struct task_struct *tsk;
39 struct intel_signal_node {
41 struct intel_wait wait;
45 * Request queue structure.
47 * The request queue allows us to note sequence numbers that have been emitted
48 * and may be associated with active buffers to be retired.
50 * By keeping this list, we can avoid having to do questionable sequence
51 * number comparisons on buffer last_read|write_seqno. It also allows an
52 * emission time to be associated with the request for tracking how far ahead
53 * of the GPU the submission is.
55 * When modifying this structure be very aware that we perform a lockless
56 * RCU lookup of it that may race against reallocation of the struct
57 * from the slab freelist. We intentionally do not zero the structure on
58 * allocation so that the lookup can use the dangling pointers (and is
59 * cogniscent that those pointers may be wrong). Instead, everything that
60 * needs to be initialised must be done so explicitly.
62 * The requests are reference counted.
64 struct drm_i915_gem_request {
68 /** On Which ring this request was generated */
69 struct drm_i915_private *i915;
72 * Context and ring buffer related to this request
73 * Contexts are refcounted, so when this request is associated with a
74 * context, we must increment the context's refcount, to guarantee that
75 * it persists while any request is linked to it. Requests themselves
76 * are also refcounted, so the request will only be freed when the last
77 * reference to it is dismissed, and the code in
78 * i915_gem_request_free() will then decrement the refcount on the
81 struct i915_gem_context *ctx;
82 struct intel_engine_cs *engine;
83 struct intel_ring *ring;
84 struct intel_signal_node signaling;
86 struct i915_sw_fence submit;
88 /** GEM sequence number associated with the previous request,
89 * when the HWS breadcrumb is equal to this the GPU is processing
94 /** Position in the ring of the start of the request */
98 * Position in the ring of the start of the postfix.
99 * This is required to calculate the maximum available ring space
100 * without overwriting the postfix.
104 /** Position in the ring of the end of the whole request */
107 /** Position in the ring of the end of any workarounds after the tail */
110 /** Preallocate space in the ring for the emitting the request */
114 * Context related to the previous request.
115 * As the contexts are accessed by the hardware until the switch is
116 * completed to a new context, the hardware may still be writing
117 * to the context object after the breadcrumb is visible. We must
118 * not unpin/unbind/prune that object whilst still active and so
119 * we keep the previous context pinned until the following (this)
120 * request is retired.
122 struct i915_gem_context *previous_context;
124 /** Batch buffer related to this request if any (used for
125 * error state dump only).
127 struct i915_vma *batch;
128 struct list_head active_list;
130 /** Time at which this request was emitted, in jiffies. */
131 unsigned long emitted_jiffies;
133 /** engine->request_list entry for this request */
134 struct list_head link;
136 /** ring->request_list entry for this request */
137 struct list_head ring_link;
139 struct drm_i915_file_private *file_priv;
140 /** file_priv list entry for this request */
141 struct list_head client_list;
143 /** Link in the execlist submission queue, guarded by execlist_lock. */
144 struct list_head execlist_link;
147 extern const struct fence_ops i915_fence_ops;
149 static inline bool fence_is_i915(struct fence *fence)
151 return fence->ops == &i915_fence_ops;
154 struct drm_i915_gem_request * __must_check
155 i915_gem_request_alloc(struct intel_engine_cs *engine,
156 struct i915_gem_context *ctx);
157 int i915_gem_request_add_to_client(struct drm_i915_gem_request *req,
158 struct drm_file *file);
159 void i915_gem_request_retire_upto(struct drm_i915_gem_request *req);
162 i915_gem_request_get_seqno(struct drm_i915_gem_request *req)
164 return req ? req->fence.seqno : 0;
167 static inline struct intel_engine_cs *
168 i915_gem_request_get_engine(struct drm_i915_gem_request *req)
170 return req ? req->engine : NULL;
173 static inline struct drm_i915_gem_request *
174 to_request(struct fence *fence)
176 /* We assume that NULL fence/request are interoperable */
177 BUILD_BUG_ON(offsetof(struct drm_i915_gem_request, fence) != 0);
178 GEM_BUG_ON(fence && !fence_is_i915(fence));
179 return container_of(fence, struct drm_i915_gem_request, fence);
182 static inline struct drm_i915_gem_request *
183 i915_gem_request_get(struct drm_i915_gem_request *req)
185 return to_request(fence_get(&req->fence));
188 static inline struct drm_i915_gem_request *
189 i915_gem_request_get_rcu(struct drm_i915_gem_request *req)
191 return to_request(fence_get_rcu(&req->fence));
195 i915_gem_request_put(struct drm_i915_gem_request *req)
197 fence_put(&req->fence);
200 static inline void i915_gem_request_assign(struct drm_i915_gem_request **pdst,
201 struct drm_i915_gem_request *src)
204 i915_gem_request_get(src);
207 i915_gem_request_put(*pdst);
212 void __i915_add_request(struct drm_i915_gem_request *req, bool flush_caches);
213 #define i915_add_request(req) \
214 __i915_add_request(req, true)
215 #define i915_add_request_no_flush(req) \
216 __i915_add_request(req, false)
218 struct intel_rps_client;
219 #define NO_WAITBOOST ERR_PTR(-1)
220 #define IS_RPS_CLIENT(p) (!IS_ERR(p))
221 #define IS_RPS_USER(p) (!IS_ERR_OR_NULL(p))
223 int i915_wait_request(struct drm_i915_gem_request *req,
226 struct intel_rps_client *rps)
227 __attribute__((nonnull(1)));
228 #define I915_WAIT_INTERRUPTIBLE BIT(0)
229 #define I915_WAIT_LOCKED BIT(1) /* struct_mutex held, handle GPU reset */
231 static inline u32 intel_engine_get_seqno(struct intel_engine_cs *engine);
234 * Returns true if seq1 is later than seq2.
236 static inline bool i915_seqno_passed(u32 seq1, u32 seq2)
238 return (s32)(seq1 - seq2) >= 0;
242 i915_gem_request_started(const struct drm_i915_gem_request *req)
244 return i915_seqno_passed(intel_engine_get_seqno(req->engine),
245 req->previous_seqno);
249 i915_gem_request_completed(const struct drm_i915_gem_request *req)
251 return i915_seqno_passed(intel_engine_get_seqno(req->engine),
255 bool __i915_spin_request(const struct drm_i915_gem_request *request,
256 int state, unsigned long timeout_us);
257 static inline bool i915_spin_request(const struct drm_i915_gem_request *request,
258 int state, unsigned long timeout_us)
260 return (i915_gem_request_started(request) &&
261 __i915_spin_request(request, state, timeout_us));
264 /* We treat requests as fences. This is not be to confused with our
265 * "fence registers" but pipeline synchronisation objects ala GL_ARB_sync.
266 * We use the fences to synchronize access from the CPU with activity on the
267 * GPU, for example, we should not rewrite an object's PTE whilst the GPU
268 * is reading them. We also track fences at a higher level to provide
269 * implicit synchronisation around GEM objects, e.g. set-domain will wait
270 * for outstanding GPU rendering before marking the object ready for CPU
271 * access, or a pageflip will wait until the GPU is complete before showing
272 * the frame on the scanout.
274 * In order to use a fence, the object must track the fence it needs to
275 * serialise with. For example, GEM objects want to track both read and
276 * write access so that we can perform concurrent read operations between
277 * the CPU and GPU engines, as well as waiting for all rendering to
278 * complete, or waiting for the last GPU user of a "fence register". The
279 * object then embeds a #i915_gem_active to track the most recent (in
280 * retirement order) request relevant for the desired mode of access.
281 * The #i915_gem_active is updated with i915_gem_active_set() to track the
282 * most recent fence request, typically this is done as part of
283 * i915_vma_move_to_active().
285 * When the #i915_gem_active completes (is retired), it will
286 * signal its completion to the owner through a callback as well as mark
287 * itself as idle (i915_gem_active.request == NULL). The owner
288 * can then perform any action, such as delayed freeing of an active
289 * resource including itself.
291 struct i915_gem_active;
293 typedef void (*i915_gem_retire_fn)(struct i915_gem_active *,
294 struct drm_i915_gem_request *);
296 struct i915_gem_active {
297 struct drm_i915_gem_request __rcu *request;
298 struct list_head link;
299 i915_gem_retire_fn retire;
302 void i915_gem_retire_noop(struct i915_gem_active *,
303 struct drm_i915_gem_request *request);
306 * init_request_active - prepares the activity tracker for use
307 * @active - the active tracker
308 * @func - a callback when then the tracker is retired (becomes idle),
311 * init_request_active() prepares the embedded @active struct for use as
312 * an activity tracker, that is for tracking the last known active request
313 * associated with it. When the last request becomes idle, when it is retired
314 * after completion, the optional callback @func is invoked.
317 init_request_active(struct i915_gem_active *active,
318 i915_gem_retire_fn retire)
320 INIT_LIST_HEAD(&active->link);
321 active->retire = retire ?: i915_gem_retire_noop;
325 * i915_gem_active_set - updates the tracker to watch the current request
326 * @active - the active tracker
327 * @request - the request to watch
329 * i915_gem_active_set() watches the given @request for completion. Whilst
330 * that @request is busy, the @active reports busy. When that @request is
331 * retired, the @active tracker is updated to report idle.
334 i915_gem_active_set(struct i915_gem_active *active,
335 struct drm_i915_gem_request *request)
337 list_move(&active->link, &request->active_list);
338 rcu_assign_pointer(active->request, request);
341 static inline struct drm_i915_gem_request *
342 __i915_gem_active_peek(const struct i915_gem_active *active)
344 /* Inside the error capture (running with the driver in an unknown
345 * state), we want to bend the rules slightly (a lot).
347 * Work is in progress to make it safer, in the meantime this keeps
348 * the known issue from spamming the logs.
350 return rcu_dereference_protected(active->request, 1);
354 * i915_gem_active_raw - return the active request
355 * @active - the active tracker
357 * i915_gem_active_raw() returns the current request being tracked, or NULL.
358 * It does not obtain a reference on the request for the caller, so the caller
359 * must hold struct_mutex.
361 static inline struct drm_i915_gem_request *
362 i915_gem_active_raw(const struct i915_gem_active *active, struct mutex *mutex)
364 return rcu_dereference_protected(active->request,
365 lockdep_is_held(mutex));
369 * i915_gem_active_peek - report the active request being monitored
370 * @active - the active tracker
372 * i915_gem_active_peek() returns the current request being tracked if
373 * still active, or NULL. It does not obtain a reference on the request
374 * for the caller, so the caller must hold struct_mutex.
376 static inline struct drm_i915_gem_request *
377 i915_gem_active_peek(const struct i915_gem_active *active, struct mutex *mutex)
379 struct drm_i915_gem_request *request;
381 request = i915_gem_active_raw(active, mutex);
382 if (!request || i915_gem_request_completed(request))
389 * i915_gem_active_get - return a reference to the active request
390 * @active - the active tracker
392 * i915_gem_active_get() returns a reference to the active request, or NULL
393 * if the active tracker is idle. The caller must hold struct_mutex.
395 static inline struct drm_i915_gem_request *
396 i915_gem_active_get(const struct i915_gem_active *active, struct mutex *mutex)
398 return i915_gem_request_get(i915_gem_active_peek(active, mutex));
402 * __i915_gem_active_get_rcu - return a reference to the active request
403 * @active - the active tracker
405 * __i915_gem_active_get() returns a reference to the active request, or NULL
406 * if the active tracker is idle. The caller must hold the RCU read lock, but
407 * the returned pointer is safe to use outside of RCU.
409 static inline struct drm_i915_gem_request *
410 __i915_gem_active_get_rcu(const struct i915_gem_active *active)
412 /* Performing a lockless retrieval of the active request is super
413 * tricky. SLAB_DESTROY_BY_RCU merely guarantees that the backing
414 * slab of request objects will not be freed whilst we hold the
415 * RCU read lock. It does not guarantee that the request itself
416 * will not be freed and then *reused*. Viz,
420 * req = active.request
421 * retire(req) -> free(req);
422 * (req is now first on the slab freelist)
423 * active.request = NULL
425 * req = new submission on a new object
428 * To prevent the request from being reused whilst the caller
429 * uses it, we take a reference like normal. Whilst acquiring
430 * the reference we check that it is not in a destroyed state
431 * (refcnt == 0). That prevents the request being reallocated
432 * whilst the caller holds on to it. To check that the request
433 * was not reallocated as we acquired the reference we have to
434 * check that our request remains the active request across
435 * the lookup, in the same manner as a seqlock. The visibility
436 * of the pointer versus the reference counting is controlled
437 * by using RCU barriers (rcu_dereference and rcu_assign_pointer).
439 * In the middle of all that, we inspect whether the request is
440 * complete. Retiring is lazy so the request may be completed long
441 * before the active tracker is updated. Querying whether the
442 * request is complete is far cheaper (as it involves no locked
443 * instructions setting cachelines to exclusive) than acquiring
444 * the reference, so we do it first. The RCU read lock ensures the
445 * pointer dereference is valid, but does not ensure that the
446 * seqno nor HWS is the right one! However, if the request was
447 * reallocated, that means the active tracker's request was complete.
448 * If the new request is also complete, then both are and we can
449 * just report the active tracker is idle. If the new request is
450 * incomplete, then we acquire a reference on it and check that
451 * it remained the active request.
453 * It is then imperative that we do not zero the request on
454 * reallocation, so that we can chase the dangling pointers!
455 * See i915_gem_request_alloc().
458 struct drm_i915_gem_request *request;
460 request = rcu_dereference(active->request);
461 if (!request || i915_gem_request_completed(request))
464 /* An especially silly compiler could decide to recompute the
465 * result of i915_gem_request_completed, more specifically
466 * re-emit the load for request->fence.seqno. A race would catch
467 * a later seqno value, which could flip the result from true to
468 * false. Which means part of the instructions below might not
469 * be executed, while later on instructions are executed. Due to
470 * barriers within the refcounting the inconsistency can't reach
471 * past the call to i915_gem_request_get_rcu, but not executing
472 * that while still executing i915_gem_request_put() creates
473 * havoc enough. Prevent this with a compiler barrier.
477 request = i915_gem_request_get_rcu(request);
479 /* What stops the following rcu_access_pointer() from occurring
480 * before the above i915_gem_request_get_rcu()? If we were
481 * to read the value before pausing to get the reference to
482 * the request, we may not notice a change in the active
485 * The rcu_access_pointer() is a mere compiler barrier, which
486 * means both the CPU and compiler are free to perform the
487 * memory read without constraint. The compiler only has to
488 * ensure that any operations after the rcu_access_pointer()
489 * occur afterwards in program order. This means the read may
490 * be performed earlier by an out-of-order CPU, or adventurous
493 * The atomic operation at the heart of
494 * i915_gem_request_get_rcu(), see fence_get_rcu(), is
495 * atomic_inc_not_zero() which is only a full memory barrier
496 * when successful. That is, if i915_gem_request_get_rcu()
497 * returns the request (and so with the reference counted
498 * incremented) then the following read for rcu_access_pointer()
499 * must occur after the atomic operation and so confirm
500 * that this request is the one currently being tracked.
502 * The corresponding write barrier is part of
503 * rcu_assign_pointer().
505 if (!request || request == rcu_access_pointer(active->request))
506 return rcu_pointer_handoff(request);
508 i915_gem_request_put(request);
513 * i915_gem_active_get_unlocked - return a reference to the active request
514 * @active - the active tracker
516 * i915_gem_active_get_unlocked() returns a reference to the active request,
517 * or NULL if the active tracker is idle. The reference is obtained under RCU,
518 * so no locking is required by the caller.
520 * The reference should be freed with i915_gem_request_put().
522 static inline struct drm_i915_gem_request *
523 i915_gem_active_get_unlocked(const struct i915_gem_active *active)
525 struct drm_i915_gem_request *request;
528 request = __i915_gem_active_get_rcu(active);
535 * i915_gem_active_isset - report whether the active tracker is assigned
536 * @active - the active tracker
538 * i915_gem_active_isset() returns true if the active tracker is currently
539 * assigned to a request. Due to the lazy retiring, that request may be idle
540 * and this may report stale information.
543 i915_gem_active_isset(const struct i915_gem_active *active)
545 return rcu_access_pointer(active->request);
549 * i915_gem_active_is_idle - report whether the active tracker is idle
550 * @active - the active tracker
552 * i915_gem_active_is_idle() returns true if the active tracker is currently
553 * unassigned or if the request is complete (but not yet retired). Requires
554 * the caller to hold struct_mutex (but that can be relaxed if desired).
557 i915_gem_active_is_idle(const struct i915_gem_active *active,
560 return !i915_gem_active_peek(active, mutex);
564 * i915_gem_active_wait - waits until the request is completed
565 * @active - the active request on which to wait
567 * i915_gem_active_wait() waits until the request is completed before
568 * returning. Note that it does not guarantee that the request is
569 * retired first, see i915_gem_active_retire().
571 * i915_gem_active_wait() returns immediately if the active
572 * request is already complete.
574 static inline int __must_check
575 i915_gem_active_wait(const struct i915_gem_active *active, struct mutex *mutex)
577 struct drm_i915_gem_request *request;
579 request = i915_gem_active_peek(active, mutex);
583 return i915_wait_request(request,
584 I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
589 * i915_gem_active_wait_unlocked - waits until the request is completed
590 * @active - the active request on which to wait
591 * @flags - how to wait
592 * @timeout - how long to wait at most
593 * @rps - userspace client to charge for a waitboost
595 * i915_gem_active_wait_unlocked() waits until the request is completed before
596 * returning, without requiring any locks to be held. Note that it does not
597 * retire any requests before returning.
599 * This function relies on RCU in order to acquire the reference to the active
600 * request without holding any locks. See __i915_gem_active_get_rcu() for the
601 * glory details on how that is managed. Once the reference is acquired, we
602 * can then wait upon the request, and afterwards release our reference,
603 * free of any locking.
605 * This function wraps i915_wait_request(), see it for the full details on
608 * Returns 0 if successful, or a negative error code.
611 i915_gem_active_wait_unlocked(const struct i915_gem_active *active,
614 struct intel_rps_client *rps)
616 struct drm_i915_gem_request *request;
619 request = i915_gem_active_get_unlocked(active);
621 ret = i915_wait_request(request, flags, timeout, rps);
622 i915_gem_request_put(request);
629 * i915_gem_active_retire - waits until the request is retired
630 * @active - the active request on which to wait
632 * i915_gem_active_retire() waits until the request is completed,
633 * and then ensures that at least the retirement handler for this
634 * @active tracker is called before returning. If the @active
635 * tracker is idle, the function returns immediately.
637 static inline int __must_check
638 i915_gem_active_retire(struct i915_gem_active *active,
641 struct drm_i915_gem_request *request;
644 request = i915_gem_active_raw(active, mutex);
648 ret = i915_wait_request(request,
649 I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
654 list_del_init(&active->link);
655 RCU_INIT_POINTER(active->request, NULL);
657 active->retire(active, request);
662 /* Convenience functions for peeking at state inside active's request whilst
663 * guarded by the struct_mutex.
666 static inline uint32_t
667 i915_gem_active_get_seqno(const struct i915_gem_active *active,
670 return i915_gem_request_get_seqno(i915_gem_active_peek(active, mutex));
673 static inline struct intel_engine_cs *
674 i915_gem_active_get_engine(const struct i915_gem_active *active,
677 return i915_gem_request_get_engine(i915_gem_active_peek(active, mutex));
680 #define for_each_active(mask, idx) \
681 for (; mask ? idx = ffs(mask) - 1, 1 : 0; mask &= ~BIT(idx))
683 #endif /* I915_GEM_REQUEST_H */