• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright © 2008-2015 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  */
24 
25 #ifndef I915_GEM_REQUEST_H
26 #define I915_GEM_REQUEST_H
27 
28 #include <linux/dma-fence.h>
29 
30 #include "i915_gem.h"
31 #include "i915_sw_fence.h"
32 
33 struct drm_file;
34 struct drm_i915_gem_object;
35 struct drm_i915_gem_request;
36 
37 struct intel_wait {
38 	struct rb_node node;
39 	struct task_struct *tsk;
40 	struct drm_i915_gem_request *request;
41 	u32 seqno;
42 };
43 
44 struct intel_signal_node {
45 	struct rb_node node;
46 	struct intel_wait wait;
47 };
48 
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;
54 	unsigned long flags;
55 #define I915_DEPENDENCY_ALLOC BIT(0)
56 };
57 
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.
66  */
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 */
70 	struct list_head link;
71 	int priority;
72 #define I915_PRIORITY_MAX 1024
73 #define I915_PRIORITY_NORMAL 0
74 #define I915_PRIORITY_MIN (-I915_PRIORITY_MAX)
75 };
76 
77 struct i915_gem_capture_list {
78 	struct i915_gem_capture_list *next;
79 	struct i915_vma *vma;
80 };
81 
82 /**
83  * Request queue structure.
84  *
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.
87  *
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.
92  *
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.
99  *
100  * The requests are reference counted.
101  */
102 struct drm_i915_gem_request {
103 	struct dma_fence fence;
104 	spinlock_t lock;
105 
106 	/** On Which ring this request was generated */
107 	struct drm_i915_private *i915;
108 
109 	/**
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
117 	 * context.
118 	 */
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;
124 
125 	/* Fences for the various phases in the request's lifetime.
126 	 *
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.
130 	 */
131 	struct i915_sw_fence submit;
132 	wait_queue_entry_t submitq;
133 	wait_queue_head_t execute;
134 
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.
142 	 */
143 	struct i915_priotree priotree;
144 	struct i915_dependency dep;
145 
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.
150 	 */
151 	u32 global_seqno;
152 
153 	/** Position in the ring of the start of the request */
154 	u32 head;
155 
156 	/**
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.
160 	 */
161 	u32 postfix;
162 
163 	/** Position in the ring of the end of the whole request */
164 	u32 tail;
165 
166 	/** Position in the ring of the end of any workarounds after the tail */
167 	u32 wa_tail;
168 
169 	/** Preallocate space in the ring for the emitting the request */
170 	u32 reserved_space;
171 
172 	/** Batch buffer related to this request if any (used for
173 	 * error state dump only).
174 	 */
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).
180 	 */
181 	struct i915_gem_capture_list *capture_list;
182 	struct list_head active_list;
183 
184 	/** Time at which this request was emitted, in jiffies. */
185 	unsigned long emitted_jiffies;
186 
187 	bool waitboost;
188 
189 	/** engine->request_list entry for this request */
190 	struct list_head link;
191 
192 	/** ring->request_list entry for this request */
193 	struct list_head ring_link;
194 
195 	struct drm_i915_file_private *file_priv;
196 	/** file_priv list entry for this request */
197 	struct list_head client_link;
198 };
199 
200 extern const struct dma_fence_ops i915_fence_ops;
201 
dma_fence_is_i915(const struct dma_fence * fence)202 static inline bool dma_fence_is_i915(const struct dma_fence *fence)
203 {
204 	return fence->ops == &i915_fence_ops;
205 }
206 
207 struct drm_i915_gem_request * __must_check
208 i915_gem_request_alloc(struct intel_engine_cs *engine,
209 		       struct i915_gem_context *ctx);
210 void i915_gem_request_retire_upto(struct drm_i915_gem_request *req);
211 
212 static inline struct drm_i915_gem_request *
to_request(struct dma_fence * fence)213 to_request(struct dma_fence *fence)
214 {
215 	/* We assume that NULL fence/request are interoperable */
216 	BUILD_BUG_ON(offsetof(struct drm_i915_gem_request, fence) != 0);
217 	GEM_BUG_ON(fence && !dma_fence_is_i915(fence));
218 	return container_of(fence, struct drm_i915_gem_request, fence);
219 }
220 
221 static inline struct drm_i915_gem_request *
i915_gem_request_get(struct drm_i915_gem_request * req)222 i915_gem_request_get(struct drm_i915_gem_request *req)
223 {
224 	return to_request(dma_fence_get(&req->fence));
225 }
226 
227 static inline struct drm_i915_gem_request *
i915_gem_request_get_rcu(struct drm_i915_gem_request * req)228 i915_gem_request_get_rcu(struct drm_i915_gem_request *req)
229 {
230 	return to_request(dma_fence_get_rcu(&req->fence));
231 }
232 
233 static inline void
i915_gem_request_put(struct drm_i915_gem_request * req)234 i915_gem_request_put(struct drm_i915_gem_request *req)
235 {
236 	dma_fence_put(&req->fence);
237 }
238 
i915_gem_request_assign(struct drm_i915_gem_request ** pdst,struct drm_i915_gem_request * src)239 static inline void i915_gem_request_assign(struct drm_i915_gem_request **pdst,
240 					   struct drm_i915_gem_request *src)
241 {
242 	if (src)
243 		i915_gem_request_get(src);
244 
245 	if (*pdst)
246 		i915_gem_request_put(*pdst);
247 
248 	*pdst = src;
249 }
250 
251 /**
252  * i915_gem_request_global_seqno - report the current global seqno
253  * @request - the request
254  *
255  * A request is assigned a global seqno only when it is on the hardware
256  * execution queue. The global seqno can be used to maintain a list of
257  * requests on the same engine in retirement order, for example for
258  * constructing a priority queue for waiting. Prior to its execution, or
259  * if it is subsequently removed in the event of preemption, its global
260  * seqno is zero. As both insertion and removal from the execution queue
261  * may operate in IRQ context, it is not guarded by the usual struct_mutex
262  * BKL. Instead those relying on the global seqno must be prepared for its
263  * value to change between reads. Only when the request is complete can
264  * the global seqno be stable (due to the memory barriers on submitting
265  * the commands to the hardware to write the breadcrumb, if the HWS shows
266  * that it has passed the global seqno and the global seqno is unchanged
267  * after the read, it is indeed complete).
268  */
269 static u32
i915_gem_request_global_seqno(const struct drm_i915_gem_request * request)270 i915_gem_request_global_seqno(const struct drm_i915_gem_request *request)
271 {
272 	return READ_ONCE(request->global_seqno);
273 }
274 
275 int
276 i915_gem_request_await_object(struct drm_i915_gem_request *to,
277 			      struct drm_i915_gem_object *obj,
278 			      bool write);
279 int i915_gem_request_await_dma_fence(struct drm_i915_gem_request *req,
280 				     struct dma_fence *fence);
281 
282 void __i915_add_request(struct drm_i915_gem_request *req, bool flush_caches);
283 #define i915_add_request(req) \
284 	__i915_add_request(req, false)
285 
286 void __i915_gem_request_submit(struct drm_i915_gem_request *request);
287 void i915_gem_request_submit(struct drm_i915_gem_request *request);
288 
289 void __i915_gem_request_unsubmit(struct drm_i915_gem_request *request);
290 void i915_gem_request_unsubmit(struct drm_i915_gem_request *request);
291 
292 struct intel_rps_client;
293 #define NO_WAITBOOST ERR_PTR(-1)
294 #define IS_RPS_CLIENT(p) (!IS_ERR(p))
295 #define IS_RPS_USER(p) (!IS_ERR_OR_NULL(p))
296 
297 long i915_wait_request(struct drm_i915_gem_request *req,
298 		       unsigned int flags,
299 		       long timeout)
300 	__attribute__((nonnull(1)));
301 #define I915_WAIT_INTERRUPTIBLE	BIT(0)
302 #define I915_WAIT_LOCKED	BIT(1) /* struct_mutex held, handle GPU reset */
303 #define I915_WAIT_ALL		BIT(2) /* used by i915_gem_object_wait() */
304 
305 static inline u32 intel_engine_get_seqno(struct intel_engine_cs *engine);
306 
307 /**
308  * Returns true if seq1 is later than seq2.
309  */
i915_seqno_passed(u32 seq1,u32 seq2)310 static inline bool i915_seqno_passed(u32 seq1, u32 seq2)
311 {
312 	return (s32)(seq1 - seq2) >= 0;
313 }
314 
315 static inline bool
__i915_gem_request_started(const struct drm_i915_gem_request * req,u32 seqno)316 __i915_gem_request_started(const struct drm_i915_gem_request *req, u32 seqno)
317 {
318 	GEM_BUG_ON(!seqno);
319 	return i915_seqno_passed(intel_engine_get_seqno(req->engine),
320 				 seqno - 1);
321 }
322 
323 static inline bool
i915_gem_request_started(const struct drm_i915_gem_request * req)324 i915_gem_request_started(const struct drm_i915_gem_request *req)
325 {
326 	u32 seqno;
327 
328 	seqno = i915_gem_request_global_seqno(req);
329 	if (!seqno)
330 		return false;
331 
332 	return __i915_gem_request_started(req, seqno);
333 }
334 
335 static inline bool
__i915_gem_request_completed(const struct drm_i915_gem_request * req,u32 seqno)336 __i915_gem_request_completed(const struct drm_i915_gem_request *req, u32 seqno)
337 {
338 	GEM_BUG_ON(!seqno);
339 	return i915_seqno_passed(intel_engine_get_seqno(req->engine), seqno) &&
340 		seqno == i915_gem_request_global_seqno(req);
341 }
342 
343 static inline bool
i915_gem_request_completed(const struct drm_i915_gem_request * req)344 i915_gem_request_completed(const struct drm_i915_gem_request *req)
345 {
346 	u32 seqno;
347 
348 	seqno = i915_gem_request_global_seqno(req);
349 	if (!seqno)
350 		return false;
351 
352 	return __i915_gem_request_completed(req, seqno);
353 }
354 
355 bool __i915_spin_request(const struct drm_i915_gem_request *request,
356 			 u32 seqno, int state, unsigned long timeout_us);
i915_spin_request(const struct drm_i915_gem_request * request,int state,unsigned long timeout_us)357 static inline bool i915_spin_request(const struct drm_i915_gem_request *request,
358 				     int state, unsigned long timeout_us)
359 {
360 	u32 seqno;
361 
362 	seqno = i915_gem_request_global_seqno(request);
363 	if (!seqno)
364 		return 0;
365 
366 	return (__i915_gem_request_started(request, seqno) &&
367 		__i915_spin_request(request, seqno, state, timeout_us));
368 }
369 
370 /* We treat requests as fences. This is not be to confused with our
371  * "fence registers" but pipeline synchronisation objects ala GL_ARB_sync.
372  * We use the fences to synchronize access from the CPU with activity on the
373  * GPU, for example, we should not rewrite an object's PTE whilst the GPU
374  * is reading them. We also track fences at a higher level to provide
375  * implicit synchronisation around GEM objects, e.g. set-domain will wait
376  * for outstanding GPU rendering before marking the object ready for CPU
377  * access, or a pageflip will wait until the GPU is complete before showing
378  * the frame on the scanout.
379  *
380  * In order to use a fence, the object must track the fence it needs to
381  * serialise with. For example, GEM objects want to track both read and
382  * write access so that we can perform concurrent read operations between
383  * the CPU and GPU engines, as well as waiting for all rendering to
384  * complete, or waiting for the last GPU user of a "fence register". The
385  * object then embeds a #i915_gem_active to track the most recent (in
386  * retirement order) request relevant for the desired mode of access.
387  * The #i915_gem_active is updated with i915_gem_active_set() to track the
388  * most recent fence request, typically this is done as part of
389  * i915_vma_move_to_active().
390  *
391  * When the #i915_gem_active completes (is retired), it will
392  * signal its completion to the owner through a callback as well as mark
393  * itself as idle (i915_gem_active.request == NULL). The owner
394  * can then perform any action, such as delayed freeing of an active
395  * resource including itself.
396  */
397 struct i915_gem_active;
398 
399 typedef void (*i915_gem_retire_fn)(struct i915_gem_active *,
400 				   struct drm_i915_gem_request *);
401 
402 struct i915_gem_active {
403 	struct drm_i915_gem_request __rcu *request;
404 	struct list_head link;
405 	i915_gem_retire_fn retire;
406 };
407 
408 void i915_gem_retire_noop(struct i915_gem_active *,
409 			  struct drm_i915_gem_request *request);
410 
411 /**
412  * init_request_active - prepares the activity tracker for use
413  * @active - the active tracker
414  * @func - a callback when then the tracker is retired (becomes idle),
415  *         can be NULL
416  *
417  * init_request_active() prepares the embedded @active struct for use as
418  * an activity tracker, that is for tracking the last known active request
419  * associated with it. When the last request becomes idle, when it is retired
420  * after completion, the optional callback @func is invoked.
421  */
422 static inline void
init_request_active(struct i915_gem_active * active,i915_gem_retire_fn retire)423 init_request_active(struct i915_gem_active *active,
424 		    i915_gem_retire_fn retire)
425 {
426 	INIT_LIST_HEAD(&active->link);
427 	active->retire = retire ?: i915_gem_retire_noop;
428 }
429 
430 /**
431  * i915_gem_active_set - updates the tracker to watch the current request
432  * @active - the active tracker
433  * @request - the request to watch
434  *
435  * i915_gem_active_set() watches the given @request for completion. Whilst
436  * that @request is busy, the @active reports busy. When that @request is
437  * retired, the @active tracker is updated to report idle.
438  */
439 static inline void
i915_gem_active_set(struct i915_gem_active * active,struct drm_i915_gem_request * request)440 i915_gem_active_set(struct i915_gem_active *active,
441 		    struct drm_i915_gem_request *request)
442 {
443 	list_move(&active->link, &request->active_list);
444 	rcu_assign_pointer(active->request, request);
445 }
446 
447 /**
448  * i915_gem_active_set_retire_fn - updates the retirement callback
449  * @active - the active tracker
450  * @fn - the routine called when the request is retired
451  * @mutex - struct_mutex used to guard retirements
452  *
453  * i915_gem_active_set_retire_fn() updates the function pointer that
454  * is called when the final request associated with the @active tracker
455  * is retired.
456  */
457 static inline void
i915_gem_active_set_retire_fn(struct i915_gem_active * active,i915_gem_retire_fn fn,struct mutex * mutex)458 i915_gem_active_set_retire_fn(struct i915_gem_active *active,
459 			      i915_gem_retire_fn fn,
460 			      struct mutex *mutex)
461 {
462 	lockdep_assert_held(mutex);
463 	active->retire = fn ?: i915_gem_retire_noop;
464 }
465 
466 static inline struct drm_i915_gem_request *
__i915_gem_active_peek(const struct i915_gem_active * active)467 __i915_gem_active_peek(const struct i915_gem_active *active)
468 {
469 	/* Inside the error capture (running with the driver in an unknown
470 	 * state), we want to bend the rules slightly (a lot).
471 	 *
472 	 * Work is in progress to make it safer, in the meantime this keeps
473 	 * the known issue from spamming the logs.
474 	 */
475 	return rcu_dereference_protected(active->request, 1);
476 }
477 
478 /**
479  * i915_gem_active_raw - return the active request
480  * @active - the active tracker
481  *
482  * i915_gem_active_raw() returns the current request being tracked, or NULL.
483  * It does not obtain a reference on the request for the caller, so the caller
484  * must hold struct_mutex.
485  */
486 static inline struct drm_i915_gem_request *
i915_gem_active_raw(const struct i915_gem_active * active,struct mutex * mutex)487 i915_gem_active_raw(const struct i915_gem_active *active, struct mutex *mutex)
488 {
489 	return rcu_dereference_protected(active->request,
490 					 lockdep_is_held(mutex));
491 }
492 
493 /**
494  * i915_gem_active_peek - report the active request being monitored
495  * @active - the active tracker
496  *
497  * i915_gem_active_peek() returns the current request being tracked if
498  * still active, or NULL. It does not obtain a reference on the request
499  * for the caller, so the caller must hold struct_mutex.
500  */
501 static inline struct drm_i915_gem_request *
i915_gem_active_peek(const struct i915_gem_active * active,struct mutex * mutex)502 i915_gem_active_peek(const struct i915_gem_active *active, struct mutex *mutex)
503 {
504 	struct drm_i915_gem_request *request;
505 
506 	request = i915_gem_active_raw(active, mutex);
507 	if (!request || i915_gem_request_completed(request))
508 		return NULL;
509 
510 	return request;
511 }
512 
513 /**
514  * i915_gem_active_get - return a reference to the active request
515  * @active - the active tracker
516  *
517  * i915_gem_active_get() returns a reference to the active request, or NULL
518  * if the active tracker is idle. The caller must hold struct_mutex.
519  */
520 static inline struct drm_i915_gem_request *
i915_gem_active_get(const struct i915_gem_active * active,struct mutex * mutex)521 i915_gem_active_get(const struct i915_gem_active *active, struct mutex *mutex)
522 {
523 	return i915_gem_request_get(i915_gem_active_peek(active, mutex));
524 }
525 
526 /**
527  * __i915_gem_active_get_rcu - return a reference to the active request
528  * @active - the active tracker
529  *
530  * __i915_gem_active_get() returns a reference to the active request, or NULL
531  * if the active tracker is idle. The caller must hold the RCU read lock, but
532  * the returned pointer is safe to use outside of RCU.
533  */
534 static inline struct drm_i915_gem_request *
__i915_gem_active_get_rcu(const struct i915_gem_active * active)535 __i915_gem_active_get_rcu(const struct i915_gem_active *active)
536 {
537 	/* Performing a lockless retrieval of the active request is super
538 	 * tricky. SLAB_TYPESAFE_BY_RCU merely guarantees that the backing
539 	 * slab of request objects will not be freed whilst we hold the
540 	 * RCU read lock. It does not guarantee that the request itself
541 	 * will not be freed and then *reused*. Viz,
542 	 *
543 	 * Thread A			Thread B
544 	 *
545 	 * req = active.request
546 	 *				retire(req) -> free(req);
547 	 *				(req is now first on the slab freelist)
548 	 *				active.request = NULL
549 	 *
550 	 *				req = new submission on a new object
551 	 * ref(req)
552 	 *
553 	 * To prevent the request from being reused whilst the caller
554 	 * uses it, we take a reference like normal. Whilst acquiring
555 	 * the reference we check that it is not in a destroyed state
556 	 * (refcnt == 0). That prevents the request being reallocated
557 	 * whilst the caller holds on to it. To check that the request
558 	 * was not reallocated as we acquired the reference we have to
559 	 * check that our request remains the active request across
560 	 * the lookup, in the same manner as a seqlock. The visibility
561 	 * of the pointer versus the reference counting is controlled
562 	 * by using RCU barriers (rcu_dereference and rcu_assign_pointer).
563 	 *
564 	 * In the middle of all that, we inspect whether the request is
565 	 * complete. Retiring is lazy so the request may be completed long
566 	 * before the active tracker is updated. Querying whether the
567 	 * request is complete is far cheaper (as it involves no locked
568 	 * instructions setting cachelines to exclusive) than acquiring
569 	 * the reference, so we do it first. The RCU read lock ensures the
570 	 * pointer dereference is valid, but does not ensure that the
571 	 * seqno nor HWS is the right one! However, if the request was
572 	 * reallocated, that means the active tracker's request was complete.
573 	 * If the new request is also complete, then both are and we can
574 	 * just report the active tracker is idle. If the new request is
575 	 * incomplete, then we acquire a reference on it and check that
576 	 * it remained the active request.
577 	 *
578 	 * It is then imperative that we do not zero the request on
579 	 * reallocation, so that we can chase the dangling pointers!
580 	 * See i915_gem_request_alloc().
581 	 */
582 	do {
583 		struct drm_i915_gem_request *request;
584 
585 		request = rcu_dereference(active->request);
586 		if (!request || i915_gem_request_completed(request))
587 			return NULL;
588 
589 		/* An especially silly compiler could decide to recompute the
590 		 * result of i915_gem_request_completed, more specifically
591 		 * re-emit the load for request->fence.seqno. A race would catch
592 		 * a later seqno value, which could flip the result from true to
593 		 * false. Which means part of the instructions below might not
594 		 * be executed, while later on instructions are executed. Due to
595 		 * barriers within the refcounting the inconsistency can't reach
596 		 * past the call to i915_gem_request_get_rcu, but not executing
597 		 * that while still executing i915_gem_request_put() creates
598 		 * havoc enough.  Prevent this with a compiler barrier.
599 		 */
600 		barrier();
601 
602 		request = i915_gem_request_get_rcu(request);
603 
604 		/* What stops the following rcu_access_pointer() from occurring
605 		 * before the above i915_gem_request_get_rcu()? If we were
606 		 * to read the value before pausing to get the reference to
607 		 * the request, we may not notice a change in the active
608 		 * tracker.
609 		 *
610 		 * The rcu_access_pointer() is a mere compiler barrier, which
611 		 * means both the CPU and compiler are free to perform the
612 		 * memory read without constraint. The compiler only has to
613 		 * ensure that any operations after the rcu_access_pointer()
614 		 * occur afterwards in program order. This means the read may
615 		 * be performed earlier by an out-of-order CPU, or adventurous
616 		 * compiler.
617 		 *
618 		 * The atomic operation at the heart of
619 		 * i915_gem_request_get_rcu(), see dma_fence_get_rcu(), is
620 		 * atomic_inc_not_zero() which is only a full memory barrier
621 		 * when successful. That is, if i915_gem_request_get_rcu()
622 		 * returns the request (and so with the reference counted
623 		 * incremented) then the following read for rcu_access_pointer()
624 		 * must occur after the atomic operation and so confirm
625 		 * that this request is the one currently being tracked.
626 		 *
627 		 * The corresponding write barrier is part of
628 		 * rcu_assign_pointer().
629 		 */
630 		if (!request || request == rcu_access_pointer(active->request))
631 			return rcu_pointer_handoff(request);
632 
633 		i915_gem_request_put(request);
634 	} while (1);
635 }
636 
637 /**
638  * i915_gem_active_get_unlocked - return a reference to the active request
639  * @active - the active tracker
640  *
641  * i915_gem_active_get_unlocked() returns a reference to the active request,
642  * or NULL if the active tracker is idle. The reference is obtained under RCU,
643  * so no locking is required by the caller.
644  *
645  * The reference should be freed with i915_gem_request_put().
646  */
647 static inline struct drm_i915_gem_request *
i915_gem_active_get_unlocked(const struct i915_gem_active * active)648 i915_gem_active_get_unlocked(const struct i915_gem_active *active)
649 {
650 	struct drm_i915_gem_request *request;
651 
652 	rcu_read_lock();
653 	request = __i915_gem_active_get_rcu(active);
654 	rcu_read_unlock();
655 
656 	return request;
657 }
658 
659 /**
660  * i915_gem_active_isset - report whether the active tracker is assigned
661  * @active - the active tracker
662  *
663  * i915_gem_active_isset() returns true if the active tracker is currently
664  * assigned to a request. Due to the lazy retiring, that request may be idle
665  * and this may report stale information.
666  */
667 static inline bool
i915_gem_active_isset(const struct i915_gem_active * active)668 i915_gem_active_isset(const struct i915_gem_active *active)
669 {
670 	return rcu_access_pointer(active->request);
671 }
672 
673 /**
674  * i915_gem_active_wait - waits until the request is completed
675  * @active - the active request on which to wait
676  * @flags - how to wait
677  * @timeout - how long to wait at most
678  * @rps - userspace client to charge for a waitboost
679  *
680  * i915_gem_active_wait() waits until the request is completed before
681  * returning, without requiring any locks to be held. Note that it does not
682  * retire any requests before returning.
683  *
684  * This function relies on RCU in order to acquire the reference to the active
685  * request without holding any locks. See __i915_gem_active_get_rcu() for the
686  * glory details on how that is managed. Once the reference is acquired, we
687  * can then wait upon the request, and afterwards release our reference,
688  * free of any locking.
689  *
690  * This function wraps i915_wait_request(), see it for the full details on
691  * the arguments.
692  *
693  * Returns 0 if successful, or a negative error code.
694  */
695 static inline int
i915_gem_active_wait(const struct i915_gem_active * active,unsigned int flags)696 i915_gem_active_wait(const struct i915_gem_active *active, unsigned int flags)
697 {
698 	struct drm_i915_gem_request *request;
699 	long ret = 0;
700 
701 	request = i915_gem_active_get_unlocked(active);
702 	if (request) {
703 		ret = i915_wait_request(request, flags, MAX_SCHEDULE_TIMEOUT);
704 		i915_gem_request_put(request);
705 	}
706 
707 	return ret < 0 ? ret : 0;
708 }
709 
710 /**
711  * i915_gem_active_retire - waits until the request is retired
712  * @active - the active request on which to wait
713  *
714  * i915_gem_active_retire() waits until the request is completed,
715  * and then ensures that at least the retirement handler for this
716  * @active tracker is called before returning. If the @active
717  * tracker is idle, the function returns immediately.
718  */
719 static inline int __must_check
i915_gem_active_retire(struct i915_gem_active * active,struct mutex * mutex)720 i915_gem_active_retire(struct i915_gem_active *active,
721 		       struct mutex *mutex)
722 {
723 	struct drm_i915_gem_request *request;
724 	long ret;
725 
726 	request = i915_gem_active_raw(active, mutex);
727 	if (!request)
728 		return 0;
729 
730 	ret = i915_wait_request(request,
731 				I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED,
732 				MAX_SCHEDULE_TIMEOUT);
733 	if (ret < 0)
734 		return ret;
735 
736 	list_del_init(&active->link);
737 	RCU_INIT_POINTER(active->request, NULL);
738 
739 	active->retire(active, request);
740 
741 	return 0;
742 }
743 
744 #define for_each_active(mask, idx) \
745 	for (; mask ? idx = ffs(mask) - 1, 1 : 0; mask &= ~BIT(idx))
746 
747 #endif /* I915_GEM_REQUEST_H */
748