1 /*
2 * SPDX-License-Identifier: MIT
3 *
4 * Copyright © 2019 Intel Corporation
5 */
6
7 #ifndef _I915_ACTIVE_H_
8 #define _I915_ACTIVE_H_
9
10 #include <linux/lockdep.h>
11
12 #include "i915_active_types.h"
13 #include "i915_request.h"
14
15 /*
16 * We treat requests as fences. This is not be to confused with our
17 * "fence registers" but pipeline synchronisation objects ala GL_ARB_sync.
18 * We use the fences to synchronize access from the CPU with activity on the
19 * GPU, for example, we should not rewrite an object's PTE whilst the GPU
20 * is reading them. We also track fences at a higher level to provide
21 * implicit synchronisation around GEM objects, e.g. set-domain will wait
22 * for outstanding GPU rendering before marking the object ready for CPU
23 * access, or a pageflip will wait until the GPU is complete before showing
24 * the frame on the scanout.
25 *
26 * In order to use a fence, the object must track the fence it needs to
27 * serialise with. For example, GEM objects want to track both read and
28 * write access so that we can perform concurrent read operations between
29 * the CPU and GPU engines, as well as waiting for all rendering to
30 * complete, or waiting for the last GPU user of a "fence register". The
31 * object then embeds a #i915_active_request to track the most recent (in
32 * retirement order) request relevant for the desired mode of access.
33 * The #i915_active_request is updated with i915_active_request_set() to
34 * track the most recent fence request, typically this is done as part of
35 * i915_vma_move_to_active().
36 *
37 * When the #i915_active_request completes (is retired), it will
38 * signal its completion to the owner through a callback as well as mark
39 * itself as idle (i915_active_request.request == NULL). The owner
40 * can then perform any action, such as delayed freeing of an active
41 * resource including itself.
42 */
43
44 void i915_active_retire_noop(struct i915_active_request *active,
45 struct i915_request *request);
46
47 /**
48 * i915_active_request_init - prepares the activity tracker for use
49 * @active - the active tracker
50 * @rq - initial request to track, can be NULL
51 * @func - a callback when then the tracker is retired (becomes idle),
52 * can be NULL
53 *
54 * i915_active_request_init() prepares the embedded @active struct for use as
55 * an activity tracker, that is for tracking the last known active request
56 * associated with it. When the last request becomes idle, when it is retired
57 * after completion, the optional callback @func is invoked.
58 */
59 static inline void
60 i915_active_request_init(struct i915_active_request *active,
61 struct mutex *lock,
62 struct i915_request *rq,
63 i915_active_retire_fn retire)
64 {
65 RCU_INIT_POINTER(active->request, rq);
66 INIT_LIST_HEAD(&active->link);
67 active->retire = retire ?: i915_active_retire_noop;
68 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
69 active->lock = lock;
70 #endif
71 }
72
73 #define INIT_ACTIVE_REQUEST(name, lock) \
74 i915_active_request_init((name), (lock), NULL, NULL)
75
76 /**
77 * i915_active_request_set - updates the tracker to watch the current request
78 * @active - the active tracker
79 * @request - the request to watch
80 *
81 * __i915_active_request_set() watches the given @request for completion. Whilst
82 * that @request is busy, the @active reports busy. When that @request is
83 * retired, the @active tracker is updated to report idle.
84 */
85 static inline void
86 __i915_active_request_set(struct i915_active_request *active,
87 struct i915_request *request)
88 {
89 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
90 lockdep_assert_held(active->lock);
91 #endif
92 list_move(&active->link, &request->active_list);
93 rcu_assign_pointer(active->request, request);
94 }
95
96 int __must_check
97 i915_active_request_set(struct i915_active_request *active,
98 struct i915_request *rq);
99
100 /**
101 * i915_active_request_raw - return the active request
102 * @active - the active tracker
103 *
104 * i915_active_request_raw() returns the current request being tracked, or NULL.
105 * It does not obtain a reference on the request for the caller, so the caller
106 * must hold struct_mutex.
107 */
108 static inline struct i915_request *
109 i915_active_request_raw(const struct i915_active_request *active,
110 struct mutex *mutex)
111 {
112 return rcu_dereference_protected(active->request,
113 lockdep_is_held(mutex));
114 }
115
116 /**
117 * i915_active_request_peek - report the active request being monitored
118 * @active - the active tracker
119 *
120 * i915_active_request_peek() returns the current request being tracked if
121 * still active, or NULL. It does not obtain a reference on the request
122 * for the caller, so the caller must hold struct_mutex.
123 */
124 static inline struct i915_request *
125 i915_active_request_peek(const struct i915_active_request *active,
126 struct mutex *mutex)
127 {
128 struct i915_request *request;
129
130 request = i915_active_request_raw(active, mutex);
131 if (!request || i915_request_completed(request))
132 return NULL;
133
134 return request;
135 }
136
137 /**
138 * i915_active_request_get - return a reference to the active request
139 * @active - the active tracker
140 *
141 * i915_active_request_get() returns a reference to the active request, or NULL
142 * if the active tracker is idle. The caller must hold struct_mutex.
143 */
144 static inline struct i915_request *
145 i915_active_request_get(const struct i915_active_request *active,
146 struct mutex *mutex)
147 {
148 return i915_request_get(i915_active_request_peek(active, mutex));
149 }
150
151 /**
152 * __i915_active_request_get_rcu - return a reference to the active request
153 * @active - the active tracker
154 *
155 * __i915_active_request_get() returns a reference to the active request,
156 * or NULL if the active tracker is idle. The caller must hold the RCU read
157 * lock, but the returned pointer is safe to use outside of RCU.
158 */
159 static inline struct i915_request *
160 __i915_active_request_get_rcu(const struct i915_active_request *active)
161 {
162 /*
163 * Performing a lockless retrieval of the active request is super
164 * tricky. SLAB_TYPESAFE_BY_RCU merely guarantees that the backing
165 * slab of request objects will not be freed whilst we hold the
166 * RCU read lock. It does not guarantee that the request itself
167 * will not be freed and then *reused*. Viz,
168 *
169 * Thread A Thread B
170 *
171 * rq = active.request
172 * retire(rq) -> free(rq);
173 * (rq is now first on the slab freelist)
174 * active.request = NULL
175 *
176 * rq = new submission on a new object
177 * ref(rq)
178 *
179 * To prevent the request from being reused whilst the caller
180 * uses it, we take a reference like normal. Whilst acquiring
181 * the reference we check that it is not in a destroyed state
182 * (refcnt == 0). That prevents the request being reallocated
183 * whilst the caller holds on to it. To check that the request
184 * was not reallocated as we acquired the reference we have to
185 * check that our request remains the active request across
186 * the lookup, in the same manner as a seqlock. The visibility
187 * of the pointer versus the reference counting is controlled
188 * by using RCU barriers (rcu_dereference and rcu_assign_pointer).
189 *
190 * In the middle of all that, we inspect whether the request is
191 * complete. Retiring is lazy so the request may be completed long
192 * before the active tracker is updated. Querying whether the
193 * request is complete is far cheaper (as it involves no locked
194 * instructions setting cachelines to exclusive) than acquiring
195 * the reference, so we do it first. The RCU read lock ensures the
196 * pointer dereference is valid, but does not ensure that the
197 * seqno nor HWS is the right one! However, if the request was
198 * reallocated, that means the active tracker's request was complete.
199 * If the new request is also complete, then both are and we can
200 * just report the active tracker is idle. If the new request is
201 * incomplete, then we acquire a reference on it and check that
202 * it remained the active request.
203 *
204 * It is then imperative that we do not zero the request on
205 * reallocation, so that we can chase the dangling pointers!
206 * See i915_request_alloc().
207 */
208 do {
209 struct i915_request *request;
210
211 request = rcu_dereference(active->request);
212 if (!request || i915_request_completed(request))
213 return NULL;
214
215 /*
216 * An especially silly compiler could decide to recompute the
217 * result of i915_request_completed, more specifically
218 * re-emit the load for request->fence.seqno. A race would catch
219 * a later seqno value, which could flip the result from true to
220 * false. Which means part of the instructions below might not
221 * be executed, while later on instructions are executed. Due to
222 * barriers within the refcounting the inconsistency can't reach
223 * past the call to i915_request_get_rcu, but not executing
224 * that while still executing i915_request_put() creates
225 * havoc enough. Prevent this with a compiler barrier.
226 */
227 barrier();
228
229 request = i915_request_get_rcu(request);
230
231 /*
232 * What stops the following rcu_access_pointer() from occurring
233 * before the above i915_request_get_rcu()? If we were
234 * to read the value before pausing to get the reference to
235 * the request, we may not notice a change in the active
236 * tracker.
237 *
238 * The rcu_access_pointer() is a mere compiler barrier, which
239 * means both the CPU and compiler are free to perform the
240 * memory read without constraint. The compiler only has to
241 * ensure that any operations after the rcu_access_pointer()
242 * occur afterwards in program order. This means the read may
243 * be performed earlier by an out-of-order CPU, or adventurous
244 * compiler.
245 *
246 * The atomic operation at the heart of
247 * i915_request_get_rcu(), see dma_fence_get_rcu(), is
248 * atomic_inc_not_zero() which is only a full memory barrier
249 * when successful. That is, if i915_request_get_rcu()
250 * returns the request (and so with the reference counted
251 * incremented) then the following read for rcu_access_pointer()
252 * must occur after the atomic operation and so confirm
253 * that this request is the one currently being tracked.
254 *
255 * The corresponding write barrier is part of
256 * rcu_assign_pointer().
257 */
258 if (!request || request == rcu_access_pointer(active->request))
259 return rcu_pointer_handoff(request);
260
261 i915_request_put(request);
262 } while (1);
263 }
264
265 /**
266 * i915_active_request_get_unlocked - return a reference to the active request
267 * @active - the active tracker
268 *
269 * i915_active_request_get_unlocked() returns a reference to the active request,
270 * or NULL if the active tracker is idle. The reference is obtained under RCU,
271 * so no locking is required by the caller.
272 *
273 * The reference should be freed with i915_request_put().
274 */
275 static inline struct i915_request *
276 i915_active_request_get_unlocked(const struct i915_active_request *active)
277 {
278 struct i915_request *request;
279
280 rcu_read_lock();
281 request = __i915_active_request_get_rcu(active);
282 rcu_read_unlock();
283
284 return request;
285 }
286
287 /**
288 * i915_active_request_isset - report whether the active tracker is assigned
289 * @active - the active tracker
290 *
291 * i915_active_request_isset() returns true if the active tracker is currently
292 * assigned to a request. Due to the lazy retiring, that request may be idle
293 * and this may report stale information.
294 */
295 static inline bool
296 i915_active_request_isset(const struct i915_active_request *active)
297 {
298 return rcu_access_pointer(active->request);
299 }
300
301 /**
302 * i915_active_request_retire - waits until the request is retired
303 * @active - the active request on which to wait
304 *
305 * i915_active_request_retire() waits until the request is completed,
306 * and then ensures that at least the retirement handler for this
307 * @active tracker is called before returning. If the @active
308 * tracker is idle, the function returns immediately.
309 */
310 static inline int __must_check
311 i915_active_request_retire(struct i915_active_request *active,
312 struct mutex *mutex, i915_active_retire_fn retire)
313 {
314 struct i915_request *request;
315 long ret;
316
317 request = i915_active_request_raw(active, mutex);
318 if (!request)
319 return 0;
320
321 ret = i915_request_wait(request,
322 I915_WAIT_INTERRUPTIBLE,
323 MAX_SCHEDULE_TIMEOUT);
324 if (ret < 0)
325 return ret;
326
327 list_del_init(&active->link);
328 RCU_INIT_POINTER(active->request, NULL);
329
330 retire(active, request);
331
332 return 0;
333 }
334
335 /*
336 * GPU activity tracking
337 *
338 * Each set of commands submitted to the GPU compromises a single request that
339 * signals a fence upon completion. struct i915_request combines the
340 * command submission, scheduling and fence signaling roles. If we want to see
341 * if a particular task is complete, we need to grab the fence (struct
342 * i915_request) for that task and check or wait for it to be signaled. More
343 * often though we want to track the status of a bunch of tasks, for example
344 * to wait for the GPU to finish accessing some memory across a variety of
345 * different command pipelines from different clients. We could choose to
346 * track every single request associated with the task, but knowing that
347 * each request belongs to an ordered timeline (later requests within a
348 * timeline must wait for earlier requests), we need only track the
349 * latest request in each timeline to determine the overall status of the
350 * task.
351 *
352 * struct i915_active provides this tracking across timelines. It builds a
353 * composite shared-fence, and is updated as new work is submitted to the task,
354 * forming a snapshot of the current status. It should be embedded into the
355 * different resources that need to track their associated GPU activity to
356 * provide a callback when that GPU activity has ceased, or otherwise to
357 * provide a serialisation point either for request submission or for CPU
358 * synchronisation.
359 */
360
361 void __i915_active_init(struct drm_i915_private *i915,
362 struct i915_active *ref,
363 int (*active)(struct i915_active *ref),
364 void (*retire)(struct i915_active *ref),
365 struct lock_class_key *key);
366 #define i915_active_init(i915, ref, active, retire) do { \
367 static struct lock_class_key __key; \
368 \
369 __i915_active_init(i915, ref, active, retire, &__key); \
370 } while (0)
371
372 int i915_active_ref(struct i915_active *ref,
373 struct intel_timeline *tl,
374 struct i915_request *rq);
375
376 int i915_active_wait(struct i915_active *ref);
377
378 int i915_request_await_active(struct i915_request *rq,
379 struct i915_active *ref);
380 int i915_request_await_active_request(struct i915_request *rq,
381 struct i915_active_request *active);
382
383 int i915_active_acquire(struct i915_active *ref);
384 void i915_active_release(struct i915_active *ref);
385 void __i915_active_release_nested(struct i915_active *ref, int subclass);
386
387 bool i915_active_trygrab(struct i915_active *ref);
388 void i915_active_ungrab(struct i915_active *ref);
389
390 static inline bool
391 i915_active_is_idle(const struct i915_active *ref)
392 {
393 return !atomic_read(&ref->count);
394 }
395
396 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
397 void i915_active_fini(struct i915_active *ref);
398 #else
399 static inline void i915_active_fini(struct i915_active *ref) { }
400 #endif
401
402 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
403 struct intel_engine_cs *engine);
404 void i915_active_acquire_barrier(struct i915_active *ref);
405 void i915_request_add_active_barriers(struct i915_request *rq);
406
407 #endif /* _I915_ACTIVE_H_ */