]>
Commit | Line | Data |
---|---|---|
05235c53 CW |
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 | ||
f54d1867 | 28 | #include <linux/dma-fence.h> |
04769652 CW |
29 | |
30 | #include "i915_gem.h" | |
5590af3e | 31 | #include "i915_sw_fence.h" |
04769652 | 32 | |
b42fe9ca JL |
33 | struct drm_file; |
34 | struct drm_i915_gem_object; | |
56299fb7 | 35 | struct drm_i915_gem_request; |
b42fe9ca | 36 | |
dcff85c8 CW |
37 | struct intel_wait { |
38 | struct rb_node node; | |
39 | struct task_struct *tsk; | |
56299fb7 | 40 | struct drm_i915_gem_request *request; |
dcff85c8 CW |
41 | u32 seqno; |
42 | }; | |
43 | ||
44 | struct intel_signal_node { | |
45 | struct rb_node node; | |
46 | struct intel_wait wait; | |
47 | }; | |
48 | ||
52e54209 CW |
49 | struct i915_dependency { |
50 | struct i915_priotree *signaler; | |
51 | struct list_head signal_link; | |
52 | struct list_head wait_link; | |
20311bd3 | 53 | struct list_head dfs_link; |
52e54209 CW |
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 */ | |
20311bd3 CW |
70 | struct rb_node node; |
71 | int priority; | |
72 | #define I915_PRIORITY_MAX 1024 | |
73 | #define I915_PRIORITY_MIN (-I915_PRIORITY_MAX) | |
52e54209 CW |
74 | }; |
75 | ||
05235c53 CW |
76 | /** |
77 | * Request queue structure. | |
78 | * | |
79 | * The request queue allows us to note sequence numbers that have been emitted | |
80 | * and may be associated with active buffers to be retired. | |
81 | * | |
82 | * By keeping this list, we can avoid having to do questionable sequence | |
83 | * number comparisons on buffer last_read|write_seqno. It also allows an | |
84 | * emission time to be associated with the request for tracking how far ahead | |
85 | * of the GPU the submission is. | |
86 | * | |
5a198b8c CW |
87 | * When modifying this structure be very aware that we perform a lockless |
88 | * RCU lookup of it that may race against reallocation of the struct | |
89 | * from the slab freelist. We intentionally do not zero the structure on | |
90 | * allocation so that the lookup can use the dangling pointers (and is | |
91 | * cogniscent that those pointers may be wrong). Instead, everything that | |
92 | * needs to be initialised must be done so explicitly. | |
93 | * | |
04769652 | 94 | * The requests are reference counted. |
05235c53 CW |
95 | */ |
96 | struct drm_i915_gem_request { | |
f54d1867 | 97 | struct dma_fence fence; |
04769652 | 98 | spinlock_t lock; |
05235c53 CW |
99 | |
100 | /** On Which ring this request was generated */ | |
101 | struct drm_i915_private *i915; | |
102 | ||
103 | /** | |
104 | * Context and ring buffer related to this request | |
105 | * Contexts are refcounted, so when this request is associated with a | |
106 | * context, we must increment the context's refcount, to guarantee that | |
107 | * it persists while any request is linked to it. Requests themselves | |
108 | * are also refcounted, so the request will only be freed when the last | |
109 | * reference to it is dismissed, and the code in | |
110 | * i915_gem_request_free() will then decrement the refcount on the | |
111 | * context. | |
112 | */ | |
113 | struct i915_gem_context *ctx; | |
114 | struct intel_engine_cs *engine; | |
7e37f889 | 115 | struct intel_ring *ring; |
73cb9701 | 116 | struct intel_timeline *timeline; |
05235c53 CW |
117 | struct intel_signal_node signaling; |
118 | ||
23902e49 CW |
119 | /* Fences for the various phases in the request's lifetime. |
120 | * | |
121 | * The submit fence is used to await upon all of the request's | |
122 | * dependencies. When it is signaled, the request is ready to run. | |
123 | * It is used by the driver to then queue the request for execution. | |
23902e49 | 124 | */ |
5590af3e | 125 | struct i915_sw_fence submit; |
ac6424b9 | 126 | wait_queue_entry_t submitq; |
fe49789f | 127 | wait_queue_head_t execute; |
5590af3e | 128 | |
52e54209 CW |
129 | /* A list of everyone we wait upon, and everyone who waits upon us. |
130 | * Even though we will not be submitted to the hardware before the | |
131 | * submit fence is signaled (it waits for all external events as well | |
132 | * as our own requests), the scheduler still needs to know the | |
133 | * dependency tree for the lifetime of the request (from execbuf | |
134 | * to retirement), i.e. bidirectional dependency information for the | |
135 | * request not tied to individual fences. | |
136 | */ | |
137 | struct i915_priotree priotree; | |
138 | struct i915_dependency dep; | |
139 | ||
754c9fd5 CW |
140 | /** GEM sequence number associated with this request on the |
141 | * global execution timeline. It is zero when the request is not | |
142 | * on the HW queue (i.e. not on the engine timeline list). | |
143 | * Its value is guarded by the timeline spinlock. | |
144 | */ | |
65e4760e CW |
145 | u32 global_seqno; |
146 | ||
a52abd2f | 147 | /** Position in the ring of the start of the request */ |
05235c53 CW |
148 | u32 head; |
149 | ||
150 | /** | |
a52abd2f CW |
151 | * Position in the ring of the start of the postfix. |
152 | * This is required to calculate the maximum available ring space | |
153 | * without overwriting the postfix. | |
05235c53 CW |
154 | */ |
155 | u32 postfix; | |
156 | ||
a52abd2f | 157 | /** Position in the ring of the end of the whole request */ |
05235c53 CW |
158 | u32 tail; |
159 | ||
a52abd2f CW |
160 | /** Position in the ring of the end of any workarounds after the tail */ |
161 | u32 wa_tail; | |
162 | ||
163 | /** Preallocate space in the ring for the emitting the request */ | |
05235c53 CW |
164 | u32 reserved_space; |
165 | ||
05235c53 CW |
166 | /** Batch buffer related to this request if any (used for |
167 | * error state dump only). | |
168 | */ | |
058d88c4 | 169 | struct i915_vma *batch; |
fa545cbf | 170 | struct list_head active_list; |
05235c53 CW |
171 | |
172 | /** Time at which this request was emitted, in jiffies. */ | |
173 | unsigned long emitted_jiffies; | |
174 | ||
efdf7c06 CW |
175 | /** engine->request_list entry for this request */ |
176 | struct list_head link; | |
05235c53 | 177 | |
675d9ad7 CW |
178 | /** ring->request_list entry for this request */ |
179 | struct list_head ring_link; | |
180 | ||
05235c53 CW |
181 | struct drm_i915_file_private *file_priv; |
182 | /** file_priv list entry for this request */ | |
c8659efa | 183 | struct list_head client_link; |
05235c53 CW |
184 | }; |
185 | ||
f54d1867 | 186 | extern const struct dma_fence_ops i915_fence_ops; |
04769652 | 187 | |
b52992c0 | 188 | static inline bool dma_fence_is_i915(const struct dma_fence *fence) |
04769652 CW |
189 | { |
190 | return fence->ops == &i915_fence_ops; | |
191 | } | |
192 | ||
05235c53 CW |
193 | struct drm_i915_gem_request * __must_check |
194 | i915_gem_request_alloc(struct intel_engine_cs *engine, | |
195 | struct i915_gem_context *ctx); | |
05235c53 CW |
196 | void i915_gem_request_retire_upto(struct drm_i915_gem_request *req); |
197 | ||
04769652 | 198 | static inline struct drm_i915_gem_request * |
f54d1867 | 199 | to_request(struct dma_fence *fence) |
04769652 CW |
200 | { |
201 | /* We assume that NULL fence/request are interoperable */ | |
202 | BUILD_BUG_ON(offsetof(struct drm_i915_gem_request, fence) != 0); | |
b52992c0 | 203 | GEM_BUG_ON(fence && !dma_fence_is_i915(fence)); |
04769652 CW |
204 | return container_of(fence, struct drm_i915_gem_request, fence); |
205 | } | |
206 | ||
05235c53 | 207 | static inline struct drm_i915_gem_request * |
e8a261ea | 208 | i915_gem_request_get(struct drm_i915_gem_request *req) |
05235c53 | 209 | { |
f54d1867 | 210 | return to_request(dma_fence_get(&req->fence)); |
05235c53 CW |
211 | } |
212 | ||
0eafec6d CW |
213 | static inline struct drm_i915_gem_request * |
214 | i915_gem_request_get_rcu(struct drm_i915_gem_request *req) | |
215 | { | |
f54d1867 | 216 | return to_request(dma_fence_get_rcu(&req->fence)); |
0eafec6d CW |
217 | } |
218 | ||
05235c53 | 219 | static inline void |
e8a261ea | 220 | i915_gem_request_put(struct drm_i915_gem_request *req) |
05235c53 | 221 | { |
f54d1867 | 222 | dma_fence_put(&req->fence); |
05235c53 CW |
223 | } |
224 | ||
225 | static inline void i915_gem_request_assign(struct drm_i915_gem_request **pdst, | |
226 | struct drm_i915_gem_request *src) | |
227 | { | |
228 | if (src) | |
e8a261ea | 229 | i915_gem_request_get(src); |
05235c53 CW |
230 | |
231 | if (*pdst) | |
e8a261ea | 232 | i915_gem_request_put(*pdst); |
05235c53 CW |
233 | |
234 | *pdst = src; | |
235 | } | |
236 | ||
754c9fd5 CW |
237 | /** |
238 | * i915_gem_request_global_seqno - report the current global seqno | |
239 | * @request - the request | |
240 | * | |
241 | * A request is assigned a global seqno only when it is on the hardware | |
242 | * execution queue. The global seqno can be used to maintain a list of | |
243 | * requests on the same engine in retirement order, for example for | |
244 | * constructing a priority queue for waiting. Prior to its execution, or | |
245 | * if it is subsequently removed in the event of preemption, its global | |
246 | * seqno is zero. As both insertion and removal from the execution queue | |
247 | * may operate in IRQ context, it is not guarded by the usual struct_mutex | |
248 | * BKL. Instead those relying on the global seqno must be prepared for its | |
249 | * value to change between reads. Only when the request is complete can | |
250 | * the global seqno be stable (due to the memory barriers on submitting | |
251 | * the commands to the hardware to write the breadcrumb, if the HWS shows | |
252 | * that it has passed the global seqno and the global seqno is unchanged | |
253 | * after the read, it is indeed complete). | |
254 | */ | |
255 | static u32 | |
256 | i915_gem_request_global_seqno(const struct drm_i915_gem_request *request) | |
257 | { | |
258 | return READ_ONCE(request->global_seqno); | |
259 | } | |
260 | ||
a2bc4695 CW |
261 | int |
262 | i915_gem_request_await_object(struct drm_i915_gem_request *to, | |
263 | struct drm_i915_gem_object *obj, | |
264 | bool write); | |
b52992c0 CW |
265 | int i915_gem_request_await_dma_fence(struct drm_i915_gem_request *req, |
266 | struct dma_fence *fence); | |
a2bc4695 | 267 | |
17f298cf | 268 | void __i915_add_request(struct drm_i915_gem_request *req, bool flush_caches); |
05235c53 | 269 | #define i915_add_request(req) \ |
17f298cf | 270 | __i915_add_request(req, false) |
05235c53 | 271 | |
d55ac5bf CW |
272 | void __i915_gem_request_submit(struct drm_i915_gem_request *request); |
273 | void i915_gem_request_submit(struct drm_i915_gem_request *request); | |
274 | ||
d6a2289d CW |
275 | void __i915_gem_request_unsubmit(struct drm_i915_gem_request *request); |
276 | void i915_gem_request_unsubmit(struct drm_i915_gem_request *request); | |
277 | ||
05235c53 | 278 | struct intel_rps_client; |
42df2714 CW |
279 | #define NO_WAITBOOST ERR_PTR(-1) |
280 | #define IS_RPS_CLIENT(p) (!IS_ERR(p)) | |
281 | #define IS_RPS_USER(p) (!IS_ERR_OR_NULL(p)) | |
05235c53 | 282 | |
e95433c7 CW |
283 | long i915_wait_request(struct drm_i915_gem_request *req, |
284 | unsigned int flags, | |
285 | long timeout) | |
fa545cbf | 286 | __attribute__((nonnull(1))); |
22dd3bb9 CW |
287 | #define I915_WAIT_INTERRUPTIBLE BIT(0) |
288 | #define I915_WAIT_LOCKED BIT(1) /* struct_mutex held, handle GPU reset */ | |
e95433c7 | 289 | #define I915_WAIT_ALL BIT(2) /* used by i915_gem_object_wait() */ |
fa545cbf | 290 | |
05235c53 CW |
291 | static inline u32 intel_engine_get_seqno(struct intel_engine_cs *engine); |
292 | ||
293 | /** | |
294 | * Returns true if seq1 is later than seq2. | |
295 | */ | |
296 | static inline bool i915_seqno_passed(u32 seq1, u32 seq2) | |
297 | { | |
298 | return (s32)(seq1 - seq2) >= 0; | |
299 | } | |
300 | ||
301 | static inline bool | |
754c9fd5 | 302 | __i915_gem_request_started(const struct drm_i915_gem_request *req, u32 seqno) |
05235c53 | 303 | { |
754c9fd5 | 304 | GEM_BUG_ON(!seqno); |
05235c53 | 305 | return i915_seqno_passed(intel_engine_get_seqno(req->engine), |
754c9fd5 | 306 | seqno - 1); |
05235c53 CW |
307 | } |
308 | ||
309 | static inline bool | |
65e4760e | 310 | i915_gem_request_started(const struct drm_i915_gem_request *req) |
05235c53 | 311 | { |
754c9fd5 CW |
312 | u32 seqno; |
313 | ||
314 | seqno = i915_gem_request_global_seqno(req); | |
315 | if (!seqno) | |
65e4760e CW |
316 | return false; |
317 | ||
754c9fd5 | 318 | return __i915_gem_request_started(req, seqno); |
65e4760e CW |
319 | } |
320 | ||
321 | static inline bool | |
754c9fd5 | 322 | __i915_gem_request_completed(const struct drm_i915_gem_request *req, u32 seqno) |
65e4760e | 323 | { |
754c9fd5 CW |
324 | GEM_BUG_ON(!seqno); |
325 | return i915_seqno_passed(intel_engine_get_seqno(req->engine), seqno) && | |
326 | seqno == i915_gem_request_global_seqno(req); | |
65e4760e CW |
327 | } |
328 | ||
329 | static inline bool | |
330 | i915_gem_request_completed(const struct drm_i915_gem_request *req) | |
331 | { | |
754c9fd5 CW |
332 | u32 seqno; |
333 | ||
334 | seqno = i915_gem_request_global_seqno(req); | |
335 | if (!seqno) | |
65e4760e CW |
336 | return false; |
337 | ||
754c9fd5 | 338 | return __i915_gem_request_completed(req, seqno); |
05235c53 CW |
339 | } |
340 | ||
341 | bool __i915_spin_request(const struct drm_i915_gem_request *request, | |
754c9fd5 | 342 | u32 seqno, int state, unsigned long timeout_us); |
05235c53 CW |
343 | static inline bool i915_spin_request(const struct drm_i915_gem_request *request, |
344 | int state, unsigned long timeout_us) | |
345 | { | |
754c9fd5 CW |
346 | u32 seqno; |
347 | ||
348 | seqno = i915_gem_request_global_seqno(request); | |
5b5554c5 CW |
349 | if (!seqno) |
350 | return 0; | |
351 | ||
754c9fd5 CW |
352 | return (__i915_gem_request_started(request, seqno) && |
353 | __i915_spin_request(request, seqno, state, timeout_us)); | |
05235c53 CW |
354 | } |
355 | ||
381f371b CW |
356 | /* We treat requests as fences. This is not be to confused with our |
357 | * "fence registers" but pipeline synchronisation objects ala GL_ARB_sync. | |
358 | * We use the fences to synchronize access from the CPU with activity on the | |
359 | * GPU, for example, we should not rewrite an object's PTE whilst the GPU | |
360 | * is reading them. We also track fences at a higher level to provide | |
361 | * implicit synchronisation around GEM objects, e.g. set-domain will wait | |
362 | * for outstanding GPU rendering before marking the object ready for CPU | |
363 | * access, or a pageflip will wait until the GPU is complete before showing | |
364 | * the frame on the scanout. | |
365 | * | |
366 | * In order to use a fence, the object must track the fence it needs to | |
367 | * serialise with. For example, GEM objects want to track both read and | |
368 | * write access so that we can perform concurrent read operations between | |
369 | * the CPU and GPU engines, as well as waiting for all rendering to | |
370 | * complete, or waiting for the last GPU user of a "fence register". The | |
371 | * object then embeds a #i915_gem_active to track the most recent (in | |
372 | * retirement order) request relevant for the desired mode of access. | |
373 | * The #i915_gem_active is updated with i915_gem_active_set() to track the | |
374 | * most recent fence request, typically this is done as part of | |
375 | * i915_vma_move_to_active(). | |
376 | * | |
377 | * When the #i915_gem_active completes (is retired), it will | |
378 | * signal its completion to the owner through a callback as well as mark | |
379 | * itself as idle (i915_gem_active.request == NULL). The owner | |
380 | * can then perform any action, such as delayed freeing of an active | |
381 | * resource including itself. | |
382 | */ | |
fa545cbf CW |
383 | struct i915_gem_active; |
384 | ||
385 | typedef void (*i915_gem_retire_fn)(struct i915_gem_active *, | |
386 | struct drm_i915_gem_request *); | |
387 | ||
381f371b | 388 | struct i915_gem_active { |
0eafec6d | 389 | struct drm_i915_gem_request __rcu *request; |
fa545cbf CW |
390 | struct list_head link; |
391 | i915_gem_retire_fn retire; | |
381f371b CW |
392 | }; |
393 | ||
fa545cbf CW |
394 | void i915_gem_retire_noop(struct i915_gem_active *, |
395 | struct drm_i915_gem_request *request); | |
396 | ||
397 | /** | |
398 | * init_request_active - prepares the activity tracker for use | |
399 | * @active - the active tracker | |
400 | * @func - a callback when then the tracker is retired (becomes idle), | |
401 | * can be NULL | |
402 | * | |
403 | * init_request_active() prepares the embedded @active struct for use as | |
404 | * an activity tracker, that is for tracking the last known active request | |
405 | * associated with it. When the last request becomes idle, when it is retired | |
406 | * after completion, the optional callback @func is invoked. | |
407 | */ | |
408 | static inline void | |
409 | init_request_active(struct i915_gem_active *active, | |
410 | i915_gem_retire_fn retire) | |
411 | { | |
412 | INIT_LIST_HEAD(&active->link); | |
413 | active->retire = retire ?: i915_gem_retire_noop; | |
414 | } | |
415 | ||
27c01aae CW |
416 | /** |
417 | * i915_gem_active_set - updates the tracker to watch the current request | |
418 | * @active - the active tracker | |
419 | * @request - the request to watch | |
420 | * | |
421 | * i915_gem_active_set() watches the given @request for completion. Whilst | |
422 | * that @request is busy, the @active reports busy. When that @request is | |
423 | * retired, the @active tracker is updated to report idle. | |
424 | */ | |
381f371b CW |
425 | static inline void |
426 | i915_gem_active_set(struct i915_gem_active *active, | |
427 | struct drm_i915_gem_request *request) | |
428 | { | |
fa545cbf | 429 | list_move(&active->link, &request->active_list); |
0eafec6d | 430 | rcu_assign_pointer(active->request, request); |
381f371b CW |
431 | } |
432 | ||
ecd9caa0 VS |
433 | /** |
434 | * i915_gem_active_set_retire_fn - updates the retirement callback | |
435 | * @active - the active tracker | |
436 | * @fn - the routine called when the request is retired | |
437 | * @mutex - struct_mutex used to guard retirements | |
438 | * | |
439 | * i915_gem_active_set_retire_fn() updates the function pointer that | |
440 | * is called when the final request associated with the @active tracker | |
441 | * is retired. | |
442 | */ | |
443 | static inline void | |
444 | i915_gem_active_set_retire_fn(struct i915_gem_active *active, | |
445 | i915_gem_retire_fn fn, | |
446 | struct mutex *mutex) | |
447 | { | |
448 | lockdep_assert_held(mutex); | |
449 | active->retire = fn ?: i915_gem_retire_noop; | |
450 | } | |
451 | ||
d72d908b CW |
452 | static inline struct drm_i915_gem_request * |
453 | __i915_gem_active_peek(const struct i915_gem_active *active) | |
454 | { | |
0eafec6d CW |
455 | /* Inside the error capture (running with the driver in an unknown |
456 | * state), we want to bend the rules slightly (a lot). | |
457 | * | |
458 | * Work is in progress to make it safer, in the meantime this keeps | |
459 | * the known issue from spamming the logs. | |
460 | */ | |
461 | return rcu_dereference_protected(active->request, 1); | |
d72d908b CW |
462 | } |
463 | ||
385384a8 CW |
464 | /** |
465 | * i915_gem_active_raw - return the active request | |
466 | * @active - the active tracker | |
467 | * | |
468 | * i915_gem_active_raw() returns the current request being tracked, or NULL. | |
469 | * It does not obtain a reference on the request for the caller, so the caller | |
470 | * must hold struct_mutex. | |
471 | */ | |
472 | static inline struct drm_i915_gem_request * | |
473 | i915_gem_active_raw(const struct i915_gem_active *active, struct mutex *mutex) | |
474 | { | |
475 | return rcu_dereference_protected(active->request, | |
476 | lockdep_is_held(mutex)); | |
477 | } | |
478 | ||
27c01aae | 479 | /** |
fa545cbf | 480 | * i915_gem_active_peek - report the active request being monitored |
27c01aae CW |
481 | * @active - the active tracker |
482 | * | |
fa545cbf CW |
483 | * i915_gem_active_peek() returns the current request being tracked if |
484 | * still active, or NULL. It does not obtain a reference on the request | |
485 | * for the caller, so the caller must hold struct_mutex. | |
27c01aae CW |
486 | */ |
487 | static inline struct drm_i915_gem_request * | |
d72d908b | 488 | i915_gem_active_peek(const struct i915_gem_active *active, struct mutex *mutex) |
27c01aae | 489 | { |
fa545cbf CW |
490 | struct drm_i915_gem_request *request; |
491 | ||
385384a8 | 492 | request = i915_gem_active_raw(active, mutex); |
0eafec6d CW |
493 | if (!request || i915_gem_request_completed(request)) |
494 | return NULL; | |
495 | ||
496 | return request; | |
497 | } | |
498 | ||
27c01aae CW |
499 | /** |
500 | * i915_gem_active_get - return a reference to the active request | |
501 | * @active - the active tracker | |
502 | * | |
503 | * i915_gem_active_get() returns a reference to the active request, or NULL | |
504 | * if the active tracker is idle. The caller must hold struct_mutex. | |
505 | */ | |
506 | static inline struct drm_i915_gem_request * | |
d72d908b | 507 | i915_gem_active_get(const struct i915_gem_active *active, struct mutex *mutex) |
27c01aae | 508 | { |
fa545cbf | 509 | return i915_gem_request_get(i915_gem_active_peek(active, mutex)); |
27c01aae CW |
510 | } |
511 | ||
0eafec6d CW |
512 | /** |
513 | * __i915_gem_active_get_rcu - return a reference to the active request | |
514 | * @active - the active tracker | |
515 | * | |
516 | * __i915_gem_active_get() returns a reference to the active request, or NULL | |
517 | * if the active tracker is idle. The caller must hold the RCU read lock, but | |
518 | * the returned pointer is safe to use outside of RCU. | |
519 | */ | |
520 | static inline struct drm_i915_gem_request * | |
521 | __i915_gem_active_get_rcu(const struct i915_gem_active *active) | |
522 | { | |
523 | /* Performing a lockless retrieval of the active request is super | |
5f0d5a3a | 524 | * tricky. SLAB_TYPESAFE_BY_RCU merely guarantees that the backing |
0eafec6d CW |
525 | * slab of request objects will not be freed whilst we hold the |
526 | * RCU read lock. It does not guarantee that the request itself | |
527 | * will not be freed and then *reused*. Viz, | |
528 | * | |
529 | * Thread A Thread B | |
530 | * | |
531 | * req = active.request | |
532 | * retire(req) -> free(req); | |
533 | * (req is now first on the slab freelist) | |
534 | * active.request = NULL | |
535 | * | |
536 | * req = new submission on a new object | |
537 | * ref(req) | |
538 | * | |
539 | * To prevent the request from being reused whilst the caller | |
540 | * uses it, we take a reference like normal. Whilst acquiring | |
541 | * the reference we check that it is not in a destroyed state | |
542 | * (refcnt == 0). That prevents the request being reallocated | |
543 | * whilst the caller holds on to it. To check that the request | |
544 | * was not reallocated as we acquired the reference we have to | |
545 | * check that our request remains the active request across | |
546 | * the lookup, in the same manner as a seqlock. The visibility | |
547 | * of the pointer versus the reference counting is controlled | |
548 | * by using RCU barriers (rcu_dereference and rcu_assign_pointer). | |
549 | * | |
550 | * In the middle of all that, we inspect whether the request is | |
551 | * complete. Retiring is lazy so the request may be completed long | |
552 | * before the active tracker is updated. Querying whether the | |
553 | * request is complete is far cheaper (as it involves no locked | |
554 | * instructions setting cachelines to exclusive) than acquiring | |
555 | * the reference, so we do it first. The RCU read lock ensures the | |
556 | * pointer dereference is valid, but does not ensure that the | |
557 | * seqno nor HWS is the right one! However, if the request was | |
558 | * reallocated, that means the active tracker's request was complete. | |
559 | * If the new request is also complete, then both are and we can | |
560 | * just report the active tracker is idle. If the new request is | |
561 | * incomplete, then we acquire a reference on it and check that | |
562 | * it remained the active request. | |
5a198b8c CW |
563 | * |
564 | * It is then imperative that we do not zero the request on | |
565 | * reallocation, so that we can chase the dangling pointers! | |
566 | * See i915_gem_request_alloc(). | |
0eafec6d CW |
567 | */ |
568 | do { | |
569 | struct drm_i915_gem_request *request; | |
570 | ||
571 | request = rcu_dereference(active->request); | |
572 | if (!request || i915_gem_request_completed(request)) | |
573 | return NULL; | |
574 | ||
c75870d8 DV |
575 | /* An especially silly compiler could decide to recompute the |
576 | * result of i915_gem_request_completed, more specifically | |
577 | * re-emit the load for request->fence.seqno. A race would catch | |
578 | * a later seqno value, which could flip the result from true to | |
579 | * false. Which means part of the instructions below might not | |
580 | * be executed, while later on instructions are executed. Due to | |
581 | * barriers within the refcounting the inconsistency can't reach | |
582 | * past the call to i915_gem_request_get_rcu, but not executing | |
583 | * that while still executing i915_gem_request_put() creates | |
584 | * havoc enough. Prevent this with a compiler barrier. | |
585 | */ | |
586 | barrier(); | |
587 | ||
0eafec6d CW |
588 | request = i915_gem_request_get_rcu(request); |
589 | ||
590 | /* What stops the following rcu_access_pointer() from occurring | |
591 | * before the above i915_gem_request_get_rcu()? If we were | |
592 | * to read the value before pausing to get the reference to | |
593 | * the request, we may not notice a change in the active | |
594 | * tracker. | |
595 | * | |
596 | * The rcu_access_pointer() is a mere compiler barrier, which | |
597 | * means both the CPU and compiler are free to perform the | |
598 | * memory read without constraint. The compiler only has to | |
599 | * ensure that any operations after the rcu_access_pointer() | |
600 | * occur afterwards in program order. This means the read may | |
601 | * be performed earlier by an out-of-order CPU, or adventurous | |
602 | * compiler. | |
603 | * | |
604 | * The atomic operation at the heart of | |
f54d1867 | 605 | * i915_gem_request_get_rcu(), see dma_fence_get_rcu(), is |
0eafec6d CW |
606 | * atomic_inc_not_zero() which is only a full memory barrier |
607 | * when successful. That is, if i915_gem_request_get_rcu() | |
608 | * returns the request (and so with the reference counted | |
609 | * incremented) then the following read for rcu_access_pointer() | |
610 | * must occur after the atomic operation and so confirm | |
611 | * that this request is the one currently being tracked. | |
edf6b76f CW |
612 | * |
613 | * The corresponding write barrier is part of | |
614 | * rcu_assign_pointer(). | |
0eafec6d CW |
615 | */ |
616 | if (!request || request == rcu_access_pointer(active->request)) | |
617 | return rcu_pointer_handoff(request); | |
618 | ||
619 | i915_gem_request_put(request); | |
620 | } while (1); | |
621 | } | |
622 | ||
623 | /** | |
624 | * i915_gem_active_get_unlocked - return a reference to the active request | |
625 | * @active - the active tracker | |
626 | * | |
627 | * i915_gem_active_get_unlocked() returns a reference to the active request, | |
628 | * or NULL if the active tracker is idle. The reference is obtained under RCU, | |
629 | * so no locking is required by the caller. | |
630 | * | |
631 | * The reference should be freed with i915_gem_request_put(). | |
632 | */ | |
633 | static inline struct drm_i915_gem_request * | |
634 | i915_gem_active_get_unlocked(const struct i915_gem_active *active) | |
635 | { | |
636 | struct drm_i915_gem_request *request; | |
637 | ||
638 | rcu_read_lock(); | |
639 | request = __i915_gem_active_get_rcu(active); | |
640 | rcu_read_unlock(); | |
641 | ||
642 | return request; | |
643 | } | |
644 | ||
27c01aae CW |
645 | /** |
646 | * i915_gem_active_isset - report whether the active tracker is assigned | |
647 | * @active - the active tracker | |
648 | * | |
649 | * i915_gem_active_isset() returns true if the active tracker is currently | |
650 | * assigned to a request. Due to the lazy retiring, that request may be idle | |
651 | * and this may report stale information. | |
652 | */ | |
653 | static inline bool | |
654 | i915_gem_active_isset(const struct i915_gem_active *active) | |
655 | { | |
0eafec6d | 656 | return rcu_access_pointer(active->request); |
27c01aae CW |
657 | } |
658 | ||
659 | /** | |
d07f0e59 | 660 | * i915_gem_active_wait - waits until the request is completed |
2467658e | 661 | * @active - the active request on which to wait |
ea746f36 | 662 | * @flags - how to wait |
2467658e CW |
663 | * @timeout - how long to wait at most |
664 | * @rps - userspace client to charge for a waitboost | |
665 | * | |
2e36991a | 666 | * i915_gem_active_wait() waits until the request is completed before |
2467658e CW |
667 | * returning, without requiring any locks to be held. Note that it does not |
668 | * retire any requests before returning. | |
669 | * | |
670 | * This function relies on RCU in order to acquire the reference to the active | |
671 | * request without holding any locks. See __i915_gem_active_get_rcu() for the | |
672 | * glory details on how that is managed. Once the reference is acquired, we | |
673 | * can then wait upon the request, and afterwards release our reference, | |
674 | * free of any locking. | |
675 | * | |
676 | * This function wraps i915_wait_request(), see it for the full details on | |
677 | * the arguments. | |
678 | * | |
679 | * Returns 0 if successful, or a negative error code. | |
680 | */ | |
681 | static inline int | |
2e36991a | 682 | i915_gem_active_wait(const struct i915_gem_active *active, unsigned int flags) |
2467658e CW |
683 | { |
684 | struct drm_i915_gem_request *request; | |
e95433c7 | 685 | long ret = 0; |
2467658e CW |
686 | |
687 | request = i915_gem_active_get_unlocked(active); | |
688 | if (request) { | |
e95433c7 | 689 | ret = i915_wait_request(request, flags, MAX_SCHEDULE_TIMEOUT); |
2467658e CW |
690 | i915_gem_request_put(request); |
691 | } | |
692 | ||
e95433c7 | 693 | return ret < 0 ? ret : 0; |
2467658e CW |
694 | } |
695 | ||
27c01aae CW |
696 | /** |
697 | * i915_gem_active_retire - waits until the request is retired | |
698 | * @active - the active request on which to wait | |
699 | * | |
700 | * i915_gem_active_retire() waits until the request is completed, | |
701 | * and then ensures that at least the retirement handler for this | |
702 | * @active tracker is called before returning. If the @active | |
703 | * tracker is idle, the function returns immediately. | |
704 | */ | |
705 | static inline int __must_check | |
fa545cbf | 706 | i915_gem_active_retire(struct i915_gem_active *active, |
d72d908b | 707 | struct mutex *mutex) |
27c01aae | 708 | { |
fa545cbf | 709 | struct drm_i915_gem_request *request; |
e95433c7 | 710 | long ret; |
fa545cbf | 711 | |
385384a8 | 712 | request = i915_gem_active_raw(active, mutex); |
fa545cbf CW |
713 | if (!request) |
714 | return 0; | |
715 | ||
22dd3bb9 CW |
716 | ret = i915_wait_request(request, |
717 | I915_WAIT_INTERRUPTIBLE | I915_WAIT_LOCKED, | |
e95433c7 CW |
718 | MAX_SCHEDULE_TIMEOUT); |
719 | if (ret < 0) | |
fa545cbf CW |
720 | return ret; |
721 | ||
722 | list_del_init(&active->link); | |
0eafec6d CW |
723 | RCU_INIT_POINTER(active->request, NULL); |
724 | ||
fa545cbf CW |
725 | active->retire(active, request); |
726 | ||
727 | return 0; | |
27c01aae CW |
728 | } |
729 | ||
381f371b CW |
730 | #define for_each_active(mask, idx) \ |
731 | for (; mask ? idx = ffs(mask) - 1, 1 : 0; mask &= ~BIT(idx)) | |
732 | ||
05235c53 | 733 | #endif /* I915_GEM_REQUEST_H */ |