File: | dev/pci/drm/i915/i915_request.c |
Warning: | line 809, column 2 Value stored to 'rq' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
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 | #include <linux/dma-fence-array.h> |
26 | #include <linux/dma-fence-chain.h> |
27 | #include <linux/irq_work.h> |
28 | #include <linux/prefetch.h> |
29 | #include <linux/sched.h> |
30 | #include <linux/sched/clock.h> |
31 | #include <linux/sched/signal.h> |
32 | |
33 | #include "gem/i915_gem_context.h" |
34 | #include "gt/intel_breadcrumbs.h" |
35 | #include "gt/intel_context.h" |
36 | #include "gt/intel_ring.h" |
37 | #include "gt/intel_rps.h" |
38 | |
39 | #include "i915_active.h" |
40 | #include "i915_drv.h" |
41 | #include "i915_globals.h" |
42 | #include "i915_trace.h" |
43 | #include "intel_pm.h" |
44 | |
45 | struct execute_cb { |
46 | struct irq_work work; |
47 | struct i915_sw_fence *fence; |
48 | void (*hook)(struct i915_request *rq, struct dma_fence *signal); |
49 | struct i915_request *signal; |
50 | }; |
51 | |
52 | static struct i915_global_request { |
53 | struct i915_global base; |
54 | #ifdef __linux__ |
55 | struct kmem_cache *slab_requests; |
56 | struct kmem_cache *slab_execute_cbs; |
57 | #else |
58 | struct pool slab_requests; |
59 | struct pool slab_execute_cbs; |
60 | #endif |
61 | } global; |
62 | |
63 | static const char *i915_fence_get_driver_name(struct dma_fence *fence) |
64 | { |
65 | return dev_name(to_request(fence)->engine->i915->drm.dev)""; |
66 | } |
67 | |
68 | static const char *i915_fence_get_timeline_name(struct dma_fence *fence) |
69 | { |
70 | const struct i915_gem_context *ctx; |
71 | |
72 | /* |
73 | * The timeline struct (as part of the ppgtt underneath a context) |
74 | * may be freed when the request is no longer in use by the GPU. |
75 | * We could extend the life of a context to beyond that of all |
76 | * fences, possibly keeping the hw resource around indefinitely, |
77 | * or we just give them a false name. Since |
78 | * dma_fence_ops.get_timeline_name is a debug feature, the occasional |
79 | * lie seems justifiable. |
80 | */ |
81 | if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) |
82 | return "signaled"; |
83 | |
84 | ctx = i915_request_gem_context(to_request(fence)); |
85 | if (!ctx) |
86 | return "[" DRIVER_NAME"i915" "]"; |
87 | |
88 | return ctx->name; |
89 | } |
90 | |
91 | static bool_Bool i915_fence_signaled(struct dma_fence *fence) |
92 | { |
93 | return i915_request_completed(to_request(fence)); |
94 | } |
95 | |
96 | static bool_Bool i915_fence_enable_signaling(struct dma_fence *fence) |
97 | { |
98 | return i915_request_enable_breadcrumb(to_request(fence)); |
99 | } |
100 | |
101 | static signed long i915_fence_wait(struct dma_fence *fence, |
102 | bool_Bool interruptible, |
103 | signed long timeout) |
104 | { |
105 | return i915_request_wait(to_request(fence), |
106 | interruptible | I915_WAIT_PRIORITY(1UL << (1)), |
107 | timeout); |
108 | } |
109 | |
110 | #ifdef __linux__ |
111 | struct kmem_cache *i915_request_slab_cache(void) |
112 | { |
113 | return global.slab_requests; |
114 | } |
115 | #else |
116 | struct pool *i915_request_slab_cache(void) |
117 | { |
118 | return &global.slab_requests; |
119 | } |
120 | #endif |
121 | |
122 | static void i915_fence_release(struct dma_fence *fence) |
123 | { |
124 | struct i915_request *rq = to_request(fence); |
125 | |
126 | /* |
127 | * The request is put onto a RCU freelist (i.e. the address |
128 | * is immediately reused), mark the fences as being freed now. |
129 | * Otherwise the debugobjects for the fences are only marked as |
130 | * freed when the slab cache itself is freed, and so we would get |
131 | * caught trying to reuse dead objects. |
132 | */ |
133 | i915_sw_fence_fini(&rq->submit); |
134 | i915_sw_fence_fini(&rq->semaphore); |
135 | |
136 | /* |
137 | * Keep one request on each engine for reserved use under mempressure |
138 | * |
139 | * We do not hold a reference to the engine here and so have to be |
140 | * very careful in what rq->engine we poke. The virtual engine is |
141 | * referenced via the rq->context and we released that ref during |
142 | * i915_request_retire(), ergo we must not dereference a virtual |
143 | * engine here. Not that we would want to, as the only consumer of |
144 | * the reserved engine->request_pool is the power management parking, |
145 | * which must-not-fail, and that is only run on the physical engines. |
146 | * |
147 | * Since the request must have been executed to be have completed, |
148 | * we know that it will have been processed by the HW and will |
149 | * not be unsubmitted again, so rq->engine and rq->execution_mask |
150 | * at this point is stable. rq->execution_mask will be a single |
151 | * bit if the last and _only_ engine it could execution on was a |
152 | * physical engine, if it's multiple bits then it started on and |
153 | * could still be on a virtual engine. Thus if the mask is not a |
154 | * power-of-two we assume that rq->engine may still be a virtual |
155 | * engine and so a dangling invalid pointer that we cannot dereference |
156 | * |
157 | * For example, consider the flow of a bonded request through a virtual |
158 | * engine. The request is created with a wide engine mask (all engines |
159 | * that we might execute on). On processing the bond, the request mask |
160 | * is reduced to one or more engines. If the request is subsequently |
161 | * bound to a single engine, it will then be constrained to only |
162 | * execute on that engine and never returned to the virtual engine |
163 | * after timeslicing away, see __unwind_incomplete_requests(). Thus we |
164 | * know that if the rq->execution_mask is a single bit, rq->engine |
165 | * can be a physical engine with the exact corresponding mask. |
166 | */ |
167 | if (is_power_of_2(rq->execution_mask)(((rq->execution_mask) != 0) && (((rq->execution_mask ) - 1) & (rq->execution_mask)) == 0) && |
168 | !cmpxchg(&rq->engine->request_pool, NULL, rq)__sync_val_compare_and_swap(&rq->engine->request_pool , ((void *)0), rq)) |
169 | return; |
170 | |
171 | #ifdef __linux__ |
172 | kmem_cache_free(global.slab_requests, rq); |
173 | #else |
174 | pool_put(&global.slab_requests, rq); |
175 | #endif |
176 | } |
177 | |
178 | const struct dma_fence_ops i915_fence_ops = { |
179 | .get_driver_name = i915_fence_get_driver_name, |
180 | .get_timeline_name = i915_fence_get_timeline_name, |
181 | .enable_signaling = i915_fence_enable_signaling, |
182 | .signaled = i915_fence_signaled, |
183 | .wait = i915_fence_wait, |
184 | .release = i915_fence_release, |
185 | }; |
186 | |
187 | static void irq_execute_cb(struct irq_work *wrk) |
188 | { |
189 | struct execute_cb *cb = container_of(wrk, typeof(*cb), work)({ const __typeof( ((typeof(*cb) *)0)->work ) *__mptr = (wrk ); (typeof(*cb) *)( (char *)__mptr - __builtin_offsetof(typeof (*cb), work) );}); |
190 | |
191 | i915_sw_fence_complete(cb->fence); |
192 | #ifdef __linux__ |
193 | kmem_cache_free(global.slab_execute_cbs, cb); |
194 | #else |
195 | pool_put(&global.slab_execute_cbs, cb); |
196 | #endif |
197 | } |
198 | |
199 | static void irq_execute_cb_hook(struct irq_work *wrk) |
200 | { |
201 | struct execute_cb *cb = container_of(wrk, typeof(*cb), work)({ const __typeof( ((typeof(*cb) *)0)->work ) *__mptr = (wrk ); (typeof(*cb) *)( (char *)__mptr - __builtin_offsetof(typeof (*cb), work) );}); |
202 | |
203 | cb->hook(container_of(cb->fence, struct i915_request, submit)({ const __typeof( ((struct i915_request *)0)->submit ) *__mptr = (cb->fence); (struct i915_request *)( (char *)__mptr - __builtin_offsetof (struct i915_request, submit) );}), |
204 | &cb->signal->fence); |
205 | i915_request_put(cb->signal); |
206 | |
207 | irq_execute_cb(wrk); |
208 | } |
209 | |
210 | static __always_inline__attribute__((__always_inline__)) void |
211 | __notify_execute_cb(struct i915_request *rq, bool_Bool (*fn)(struct irq_work *wrk)) |
212 | { |
213 | struct execute_cb *cb, *cn; |
214 | |
215 | if (llist_empty(&rq->execute_cb)) |
216 | return; |
217 | |
218 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
219 | #ifdef notyet |
220 | llist_for_each_entry_safe(cb, cn,for (cb = (((llist_del_all(&rq->execute_cb))) ? ({ const __typeof( ((__typeof(*cb) *)0)->work.llnode ) *__mptr = ( (llist_del_all(&rq->execute_cb))); (__typeof(*cb) *)( ( char *)__mptr - __builtin_offsetof(__typeof(*cb), work.llnode ) );}) : ((void *)0)); cb != ((void *)0) && (cn = ((cb ->work.llnode.next) ? ({ const __typeof( ((__typeof(*cb) * )0)->work.llnode ) *__mptr = (cb->work.llnode.next); (__typeof (*cb) *)( (char *)__mptr - __builtin_offsetof(__typeof(*cb), work .llnode) );}) : ((void *)0)), cb); cb = cn) |
221 | llist_del_all(&rq->execute_cb),for (cb = (((llist_del_all(&rq->execute_cb))) ? ({ const __typeof( ((__typeof(*cb) *)0)->work.llnode ) *__mptr = ( (llist_del_all(&rq->execute_cb))); (__typeof(*cb) *)( ( char *)__mptr - __builtin_offsetof(__typeof(*cb), work.llnode ) );}) : ((void *)0)); cb != ((void *)0) && (cn = ((cb ->work.llnode.next) ? ({ const __typeof( ((__typeof(*cb) * )0)->work.llnode ) *__mptr = (cb->work.llnode.next); (__typeof (*cb) *)( (char *)__mptr - __builtin_offsetof(__typeof(*cb), work .llnode) );}) : ((void *)0)), cb); cb = cn) |
222 | work.llnode)for (cb = (((llist_del_all(&rq->execute_cb))) ? ({ const __typeof( ((__typeof(*cb) *)0)->work.llnode ) *__mptr = ( (llist_del_all(&rq->execute_cb))); (__typeof(*cb) *)( ( char *)__mptr - __builtin_offsetof(__typeof(*cb), work.llnode ) );}) : ((void *)0)); cb != ((void *)0) && (cn = ((cb ->work.llnode.next) ? ({ const __typeof( ((__typeof(*cb) * )0)->work.llnode ) *__mptr = (cb->work.llnode.next); (__typeof (*cb) *)( (char *)__mptr - __builtin_offsetof(__typeof(*cb), work .llnode) );}) : ((void *)0)), cb); cb = cn) |
223 | fn(&cb->work); |
224 | #endif |
225 | } |
226 | |
227 | static void __notify_execute_cb_irq(struct i915_request *rq) |
228 | { |
229 | __notify_execute_cb(rq, irq_work_queue); |
230 | } |
231 | |
232 | static bool_Bool irq_work_imm(struct irq_work *wrk) |
233 | { |
234 | #ifdef notyet |
235 | wrk->func(wrk); |
236 | #else |
237 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
238 | #endif |
239 | return false0; |
240 | } |
241 | |
242 | static void __notify_execute_cb_imm(struct i915_request *rq) |
243 | { |
244 | __notify_execute_cb(rq, irq_work_imm); |
245 | } |
246 | |
247 | static void free_capture_list(struct i915_request *request) |
248 | { |
249 | struct i915_capture_list *capture; |
250 | |
251 | capture = fetch_and_zero(&request->capture_list)({ typeof(*&request->capture_list) __T = *(&request ->capture_list); *(&request->capture_list) = (typeof (*&request->capture_list))0; __T; }); |
252 | while (capture) { |
253 | struct i915_capture_list *next = capture->next; |
254 | |
255 | kfree(capture); |
256 | capture = next; |
257 | } |
258 | } |
259 | |
260 | static void __i915_request_fill(struct i915_request *rq, u8 val) |
261 | { |
262 | void *vaddr = rq->ring->vaddr; |
263 | u32 head; |
264 | |
265 | head = rq->infix; |
266 | if (rq->postfix < head) { |
267 | memset(vaddr + head, val, rq->ring->size - head)__builtin_memset((vaddr + head), (val), (rq->ring->size - head)); |
268 | head = 0; |
269 | } |
270 | memset(vaddr + head, val, rq->postfix - head)__builtin_memset((vaddr + head), (val), (rq->postfix - head )); |
271 | } |
272 | |
273 | static void remove_from_engine(struct i915_request *rq) |
274 | { |
275 | struct intel_engine_cs *engine, *locked; |
276 | |
277 | /* |
278 | * Virtual engines complicate acquiring the engine timeline lock, |
279 | * as their rq->engine pointer is not stable until under that |
280 | * engine lock. The simple ploy we use is to take the lock then |
281 | * check that the rq still belongs to the newly locked engine. |
282 | */ |
283 | locked = READ_ONCE(rq->engine)({ typeof(rq->engine) __tmp = *(volatile typeof(rq->engine ) *)&(rq->engine); membar_datadep_consumer(); __tmp; } ); |
284 | spin_lock_irq(&locked->active.lock)mtx_enter(&locked->active.lock); |
285 | while (unlikely(locked != (engine = READ_ONCE(rq->engine)))__builtin_expect(!!(locked != (engine = ({ typeof(rq->engine ) __tmp = *(volatile typeof(rq->engine) *)&(rq->engine ); membar_datadep_consumer(); __tmp; }))), 0)) { |
286 | spin_unlock(&locked->active.lock)mtx_leave(&locked->active.lock); |
287 | spin_lock(&engine->active.lock)mtx_enter(&engine->active.lock); |
288 | locked = engine; |
289 | } |
290 | list_del_init(&rq->sched.link); |
291 | |
292 | clear_bit(I915_FENCE_FLAG_PQUEUE, &rq->fence.flags); |
293 | clear_bit(I915_FENCE_FLAG_HOLD, &rq->fence.flags); |
294 | |
295 | /* Prevent further __await_execution() registering a cb, then flush */ |
296 | set_bit(I915_FENCE_FLAG_ACTIVE, &rq->fence.flags); |
297 | |
298 | spin_unlock_irq(&locked->active.lock)mtx_leave(&locked->active.lock); |
299 | |
300 | __notify_execute_cb_imm(rq); |
301 | } |
302 | |
303 | bool_Bool i915_request_retire(struct i915_request *rq) |
304 | { |
305 | if (!i915_request_completed(rq)) |
306 | return false0; |
307 | |
308 | RQ_TRACE(rq, "\n")do { const struct i915_request *rq__ = (rq); do { const struct intel_engine_cs *e__ __attribute__((__unused__)) = (rq__-> engine); do { } while (0); } while (0); } while (0); |
309 | |
310 | GEM_BUG_ON(!i915_sw_fence_signaled(&rq->submit))((void)0); |
311 | trace_i915_request_retire(rq); |
312 | i915_request_mark_complete(rq); |
313 | |
314 | /* |
315 | * We know the GPU must have read the request to have |
316 | * sent us the seqno + interrupt, so use the position |
317 | * of tail of the request to update the last known position |
318 | * of the GPU head. |
319 | * |
320 | * Note this requires that we are always called in request |
321 | * completion order. |
322 | */ |
323 | GEM_BUG_ON(!list_is_first(&rq->link,((void)0) |
324 | &i915_request_timeline(rq)->requests))((void)0); |
325 | if (IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)0) |
326 | /* Poison before we release our space in the ring */ |
327 | __i915_request_fill(rq, POISON_FREE0xdf); |
328 | rq->ring->head = rq->postfix; |
329 | |
330 | if (!i915_request_signaled(rq)) { |
331 | spin_lock_irq(&rq->lock)mtx_enter(&rq->lock); |
332 | dma_fence_signal_locked(&rq->fence); |
333 | spin_unlock_irq(&rq->lock)mtx_leave(&rq->lock); |
334 | } |
335 | |
336 | if (i915_request_has_waitboost(rq)) { |
337 | GEM_BUG_ON(!atomic_read(&rq->engine->gt->rps.num_waiters))((void)0); |
338 | atomic_dec(&rq->engine->gt->rps.num_waiters)__sync_fetch_and_sub(&rq->engine->gt->rps.num_waiters , 1); |
339 | } |
340 | |
341 | /* |
342 | * We only loosely track inflight requests across preemption, |
343 | * and so we may find ourselves attempting to retire a _completed_ |
344 | * request that we have removed from the HW and put back on a run |
345 | * queue. |
346 | * |
347 | * As we set I915_FENCE_FLAG_ACTIVE on the request, this should be |
348 | * after removing the breadcrumb and signaling it, so that we do not |
349 | * inadvertently attach the breadcrumb to a completed request. |
350 | */ |
351 | remove_from_engine(rq); |
352 | GEM_BUG_ON(!llist_empty(&rq->execute_cb))((void)0); |
353 | |
354 | __list_del_entry(&rq->link)list_del(&rq->link); /* poison neither prev/next (RCU walks) */ |
355 | |
356 | intel_context_exit(rq->context); |
357 | intel_context_unpin(rq->context); |
358 | |
359 | free_capture_list(rq); |
360 | i915_sched_node_fini(&rq->sched); |
361 | i915_request_put(rq); |
362 | |
363 | return true1; |
364 | } |
365 | |
366 | void i915_request_retire_upto(struct i915_request *rq) |
367 | { |
368 | struct intel_timeline * const tl = i915_request_timeline(rq); |
369 | struct i915_request *tmp; |
370 | |
371 | RQ_TRACE(rq, "\n")do { const struct i915_request *rq__ = (rq); do { const struct intel_engine_cs *e__ __attribute__((__unused__)) = (rq__-> engine); do { } while (0); } while (0); } while (0); |
372 | |
373 | GEM_BUG_ON(!i915_request_completed(rq))((void)0); |
374 | |
375 | do { |
376 | tmp = list_first_entry(&tl->requests, typeof(*tmp), link)({ const __typeof( ((typeof(*tmp) *)0)->link ) *__mptr = ( (&tl->requests)->next); (typeof(*tmp) *)( (char *)__mptr - __builtin_offsetof(typeof(*tmp), link) );}); |
377 | } while (i915_request_retire(tmp) && tmp != rq); |
378 | } |
379 | |
380 | static struct i915_request * const * |
381 | __engine_active(struct intel_engine_cs *engine) |
382 | { |
383 | return READ_ONCE(engine->execlists.active)({ typeof(engine->execlists.active) __tmp = *(volatile typeof (engine->execlists.active) *)&(engine->execlists.active ); membar_datadep_consumer(); __tmp; }); |
384 | } |
385 | |
386 | static bool_Bool __request_in_flight(const struct i915_request *signal) |
387 | { |
388 | struct i915_request * const *port, *rq; |
389 | bool_Bool inflight = false0; |
390 | |
391 | if (!i915_request_is_ready(signal)) |
392 | return false0; |
393 | |
394 | /* |
395 | * Even if we have unwound the request, it may still be on |
396 | * the GPU (preempt-to-busy). If that request is inside an |
397 | * unpreemptible critical section, it will not be removed. Some |
398 | * GPU functions may even be stuck waiting for the paired request |
399 | * (__await_execution) to be submitted and cannot be preempted |
400 | * until the bond is executing. |
401 | * |
402 | * As we know that there are always preemption points between |
403 | * requests, we know that only the currently executing request |
404 | * may be still active even though we have cleared the flag. |
405 | * However, we can't rely on our tracking of ELSP[0] to know |
406 | * which request is currently active and so maybe stuck, as |
407 | * the tracking maybe an event behind. Instead assume that |
408 | * if the context is still inflight, then it is still active |
409 | * even if the active flag has been cleared. |
410 | * |
411 | * To further complicate matters, if there a pending promotion, the HW |
412 | * may either perform a context switch to the second inflight execlists, |
413 | * or it may switch to the pending set of execlists. In the case of the |
414 | * latter, it may send the ACK and we process the event copying the |
415 | * pending[] over top of inflight[], _overwriting_ our *active. Since |
416 | * this implies the HW is arbitrating and not struck in *active, we do |
417 | * not worry about complete accuracy, but we do require no read/write |
418 | * tearing of the pointer [the read of the pointer must be valid, even |
419 | * as the array is being overwritten, for which we require the writes |
420 | * to avoid tearing.] |
421 | * |
422 | * Note that the read of *execlists->active may race with the promotion |
423 | * of execlists->pending[] to execlists->inflight[], overwritting |
424 | * the value at *execlists->active. This is fine. The promotion implies |
425 | * that we received an ACK from the HW, and so the context is not |
426 | * stuck -- if we do not see ourselves in *active, the inflight status |
427 | * is valid. If instead we see ourselves being copied into *active, |
428 | * we are inflight and may signal the callback. |
429 | */ |
430 | if (!intel_context_inflight(signal->context)({ unsigned long __v = (unsigned long)(({ typeof((signal-> context)->inflight) __tmp = *(volatile typeof((signal-> context)->inflight) *)&((signal->context)->inflight ); membar_datadep_consumer(); __tmp; })); (typeof(({ typeof(( signal->context)->inflight) __tmp = *(volatile typeof(( signal->context)->inflight) *)&((signal->context )->inflight); membar_datadep_consumer(); __tmp; })))(__v & -(1UL << (2))); })) |
431 | return false0; |
432 | |
433 | rcu_read_lock(); |
434 | for (port = __engine_active(signal->engine); |
435 | (rq = READ_ONCE(*port)({ typeof(*port) __tmp = *(volatile typeof(*port) *)&(*port ); membar_datadep_consumer(); __tmp; })); /* may race with promotion of pending[] */ |
436 | port++) { |
437 | if (rq->context == signal->context) { |
438 | inflight = i915_seqno_passed(rq->fence.seqno, |
439 | signal->fence.seqno); |
440 | break; |
441 | } |
442 | } |
443 | rcu_read_unlock(); |
444 | |
445 | return inflight; |
446 | } |
447 | |
448 | static int |
449 | __await_execution(struct i915_request *rq, |
450 | struct i915_request *signal, |
451 | void (*hook)(struct i915_request *rq, |
452 | struct dma_fence *signal), |
453 | gfp_t gfp) |
454 | { |
455 | STUB()do { printf("%s: stub\n", __func__); } while(0); |
456 | return -ENOSYS78; |
457 | #ifdef notyet |
458 | struct execute_cb *cb; |
459 | |
460 | if (i915_request_is_active(signal)) { |
461 | if (hook) |
462 | hook(rq, &signal->fence); |
463 | return 0; |
464 | } |
465 | |
466 | #ifdef __linux__ |
467 | cb = kmem_cache_alloc(global.slab_execute_cbs, gfp); |
468 | #else |
469 | cb = pool_get(&global.slab_execute_cbs, |
470 | (gfp & GFP_NOWAIT0x0002) ? PR_NOWAIT0x0002 : PR_WAITOK0x0001); |
471 | #endif |
472 | if (!cb) |
473 | return -ENOMEM12; |
474 | |
475 | cb->fence = &rq->submit; |
476 | i915_sw_fence_await(cb->fence); |
477 | init_irq_work(&cb->work, irq_execute_cb); |
478 | |
479 | if (hook) { |
480 | cb->hook = hook; |
481 | cb->signal = i915_request_get(signal); |
482 | #ifdef __linux__ |
483 | cb->work.func = irq_execute_cb_hook; |
484 | #else |
485 | init_irq_work(&cb->work, irq_execute_cb_hook); |
486 | #endif |
487 | } |
488 | |
489 | /* |
490 | * Register the callback first, then see if the signaler is already |
491 | * active. This ensures that if we race with the |
492 | * __notify_execute_cb from i915_request_submit() and we are not |
493 | * included in that list, we get a second bite of the cherry and |
494 | * execute it ourselves. After this point, a future |
495 | * i915_request_submit() will notify us. |
496 | * |
497 | * In i915_request_retire() we set the ACTIVE bit on a completed |
498 | * request (then flush the execute_cb). So by registering the |
499 | * callback first, then checking the ACTIVE bit, we serialise with |
500 | * the completed/retired request. |
501 | */ |
502 | if (llist_add(&cb->work.llnode, &signal->execute_cb)) { |
503 | if (i915_request_is_active(signal) || |
504 | __request_in_flight(signal)) |
505 | __notify_execute_cb_imm(signal); |
506 | } |
507 | |
508 | return 0; |
509 | #endif |
510 | } |
511 | |
512 | static bool_Bool fatal_error(int error) |
513 | { |
514 | switch (error) { |
515 | case 0: /* not an error! */ |
516 | case -EAGAIN35: /* innocent victim of a GT reset (__i915_request_reset) */ |
517 | case -ETIMEDOUT60: /* waiting for Godot (timer_i915_sw_fence_wake) */ |
518 | return false0; |
519 | default: |
520 | return true1; |
521 | } |
522 | } |
523 | |
524 | void __i915_request_skip(struct i915_request *rq) |
525 | { |
526 | GEM_BUG_ON(!fatal_error(rq->fence.error))((void)0); |
527 | |
528 | if (rq->infix == rq->postfix) |
529 | return; |
530 | |
531 | /* |
532 | * As this request likely depends on state from the lost |
533 | * context, clear out all the user operations leaving the |
534 | * breadcrumb at the end (so we get the fence notifications). |
535 | */ |
536 | __i915_request_fill(rq, 0); |
537 | rq->infix = rq->postfix; |
538 | } |
539 | |
540 | void i915_request_set_error_once(struct i915_request *rq, int error) |
541 | { |
542 | int old; |
543 | |
544 | GEM_BUG_ON(!IS_ERR_VALUE((long)error))((void)0); |
545 | |
546 | if (i915_request_signaled(rq)) |
547 | return; |
548 | |
549 | old = READ_ONCE(rq->fence.error)({ typeof(rq->fence.error) __tmp = *(volatile typeof(rq-> fence.error) *)&(rq->fence.error); membar_datadep_consumer (); __tmp; }); |
550 | do { |
551 | if (fatal_error(old)) |
552 | return; |
553 | } while (!try_cmpxchg(&rq->fence.error, &old, error)({ __typeof(&rq->fence.error) __op = (__typeof((&rq ->fence.error)))(&old); __typeof(*(&rq->fence.error )) __o = *__op; __typeof(*(&rq->fence.error)) __p = __sync_val_compare_and_swap ((&rq->fence.error), (__o), (error)); if (__p != __o) * __op = __p; (__p == __o); })); |
554 | } |
555 | |
556 | bool_Bool __i915_request_submit(struct i915_request *request) |
557 | { |
558 | struct intel_engine_cs *engine = request->engine; |
559 | bool_Bool result = false0; |
560 | |
561 | RQ_TRACE(request, "\n")do { const struct i915_request *rq__ = (request); do { const struct intel_engine_cs *e__ __attribute__((__unused__)) = (rq__-> engine); do { } while (0); } while (0); } while (0); |
562 | |
563 | GEM_BUG_ON(!irqs_disabled())((void)0); |
564 | lockdep_assert_held(&engine->active.lock)do { (void)(&engine->active.lock); } while(0); |
565 | |
566 | /* |
567 | * With the advent of preempt-to-busy, we frequently encounter |
568 | * requests that we have unsubmitted from HW, but left running |
569 | * until the next ack and so have completed in the meantime. On |
570 | * resubmission of that completed request, we can skip |
571 | * updating the payload, and execlists can even skip submitting |
572 | * the request. |
573 | * |
574 | * We must remove the request from the caller's priority queue, |
575 | * and the caller must only call us when the request is in their |
576 | * priority queue, under the active.lock. This ensures that the |
577 | * request has *not* yet been retired and we can safely move |
578 | * the request into the engine->active.list where it will be |
579 | * dropped upon retiring. (Otherwise if resubmit a *retired* |
580 | * request, this would be a horrible use-after-free.) |
581 | */ |
582 | if (i915_request_completed(request)) |
583 | goto xfer; |
584 | |
585 | if (unlikely(intel_context_is_closed(request->context) &&__builtin_expect(!!(intel_context_is_closed(request->context ) && !intel_engine_has_heartbeat(engine)), 0) |
586 | !intel_engine_has_heartbeat(engine))__builtin_expect(!!(intel_context_is_closed(request->context ) && !intel_engine_has_heartbeat(engine)), 0)) |
587 | intel_context_set_banned(request->context); |
588 | |
589 | if (unlikely(intel_context_is_banned(request->context))__builtin_expect(!!(intel_context_is_banned(request->context )), 0)) |
590 | i915_request_set_error_once(request, -EIO5); |
591 | |
592 | if (unlikely(fatal_error(request->fence.error))__builtin_expect(!!(fatal_error(request->fence.error)), 0)) |
593 | __i915_request_skip(request); |
594 | |
595 | /* |
596 | * Are we using semaphores when the gpu is already saturated? |
597 | * |
598 | * Using semaphores incurs a cost in having the GPU poll a |
599 | * memory location, busywaiting for it to change. The continual |
600 | * memory reads can have a noticeable impact on the rest of the |
601 | * system with the extra bus traffic, stalling the cpu as it too |
602 | * tries to access memory across the bus (perf stat -e bus-cycles). |
603 | * |
604 | * If we installed a semaphore on this request and we only submit |
605 | * the request after the signaler completed, that indicates the |
606 | * system is overloaded and using semaphores at this time only |
607 | * increases the amount of work we are doing. If so, we disable |
608 | * further use of semaphores until we are idle again, whence we |
609 | * optimistically try again. |
610 | */ |
611 | if (request->sched.semaphores && |
612 | i915_sw_fence_signaled(&request->semaphore)) |
613 | engine->saturated |= request->sched.semaphores; |
614 | |
615 | engine->emit_fini_breadcrumb(request, |
616 | request->ring->vaddr + request->postfix); |
617 | |
618 | trace_i915_request_execute(request); |
619 | engine->serial++; |
620 | result = true1; |
621 | |
622 | xfer: |
623 | if (!test_and_set_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags)) { |
624 | list_move_tail(&request->sched.link, &engine->active.requests); |
625 | clear_bit(I915_FENCE_FLAG_PQUEUE, &request->fence.flags); |
626 | } |
627 | |
628 | /* |
629 | * XXX Rollback bonded-execution on __i915_request_unsubmit()? |
630 | * |
631 | * In the future, perhaps when we have an active time-slicing scheduler, |
632 | * it will be interesting to unsubmit parallel execution and remove |
633 | * busywaits from the GPU until their master is restarted. This is |
634 | * quite hairy, we have to carefully rollback the fence and do a |
635 | * preempt-to-idle cycle on the target engine, all the while the |
636 | * master execute_cb may refire. |
637 | */ |
638 | __notify_execute_cb_irq(request); |
639 | |
640 | /* We may be recursing from the signal callback of another i915 fence */ |
641 | if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags)) |
642 | i915_request_enable_breadcrumb(request); |
643 | |
644 | return result; |
645 | } |
646 | |
647 | void i915_request_submit(struct i915_request *request) |
648 | { |
649 | struct intel_engine_cs *engine = request->engine; |
650 | unsigned long flags; |
651 | |
652 | /* Will be called from irq-context when using foreign fences. */ |
653 | spin_lock_irqsave(&engine->active.lock, flags)do { flags = 0; mtx_enter(&engine->active.lock); } while (0); |
654 | |
655 | __i915_request_submit(request); |
656 | |
657 | spin_unlock_irqrestore(&engine->active.lock, flags)do { (void)(flags); mtx_leave(&engine->active.lock); } while (0); |
658 | } |
659 | |
660 | void __i915_request_unsubmit(struct i915_request *request) |
661 | { |
662 | struct intel_engine_cs *engine = request->engine; |
663 | |
664 | /* |
665 | * Only unwind in reverse order, required so that the per-context list |
666 | * is kept in seqno/ring order. |
667 | */ |
668 | RQ_TRACE(request, "\n")do { const struct i915_request *rq__ = (request); do { const struct intel_engine_cs *e__ __attribute__((__unused__)) = (rq__-> engine); do { } while (0); } while (0); } while (0); |
669 | |
670 | GEM_BUG_ON(!irqs_disabled())((void)0); |
671 | lockdep_assert_held(&engine->active.lock)do { (void)(&engine->active.lock); } while(0); |
672 | |
673 | /* |
674 | * Before we remove this breadcrumb from the signal list, we have |
675 | * to ensure that a concurrent dma_fence_enable_signaling() does not |
676 | * attach itself. We first mark the request as no longer active and |
677 | * make sure that is visible to other cores, and then remove the |
678 | * breadcrumb if attached. |
679 | */ |
680 | GEM_BUG_ON(!test_bit(I915_FENCE_FLAG_ACTIVE, &request->fence.flags))((void)0); |
681 | clear_bit_unlock(I915_FENCE_FLAG_ACTIVE, &request->fence.flags); |
682 | if (test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT, &request->fence.flags)) |
683 | i915_request_cancel_breadcrumb(request); |
684 | |
685 | /* We've already spun, don't charge on resubmitting. */ |
686 | if (request->sched.semaphores && i915_request_started(request)) |
687 | request->sched.semaphores = 0; |
688 | |
689 | /* |
690 | * We don't need to wake_up any waiters on request->execute, they |
691 | * will get woken by any other event or us re-adding this request |
692 | * to the engine timeline (__i915_request_submit()). The waiters |
693 | * should be quite adapt at finding that the request now has a new |
694 | * global_seqno to the one they went to sleep on. |
695 | */ |
696 | } |
697 | |
698 | void i915_request_unsubmit(struct i915_request *request) |
699 | { |
700 | struct intel_engine_cs *engine = request->engine; |
701 | unsigned long flags; |
702 | |
703 | /* Will be called from irq-context when using foreign fences. */ |
704 | spin_lock_irqsave(&engine->active.lock, flags)do { flags = 0; mtx_enter(&engine->active.lock); } while (0); |
705 | |
706 | __i915_request_unsubmit(request); |
707 | |
708 | spin_unlock_irqrestore(&engine->active.lock, flags)do { (void)(flags); mtx_leave(&engine->active.lock); } while (0); |
709 | } |
710 | |
711 | static int __i915_sw_fence_call__attribute__((__aligned__(4))) |
712 | submit_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) |
713 | { |
714 | struct i915_request *request = |
715 | container_of(fence, typeof(*request), submit)({ const __typeof( ((typeof(*request) *)0)->submit ) *__mptr = (fence); (typeof(*request) *)( (char *)__mptr - __builtin_offsetof (typeof(*request), submit) );}); |
716 | |
717 | switch (state) { |
718 | case FENCE_COMPLETE: |
719 | trace_i915_request_submit(request); |
720 | |
721 | if (unlikely(fence->error)__builtin_expect(!!(fence->error), 0)) |
722 | i915_request_set_error_once(request, fence->error); |
723 | |
724 | /* |
725 | * We need to serialize use of the submit_request() callback |
726 | * with its hotplugging performed during an emergency |
727 | * i915_gem_set_wedged(). We use the RCU mechanism to mark the |
728 | * critical section in order to force i915_gem_set_wedged() to |
729 | * wait until the submit_request() is completed before |
730 | * proceeding. |
731 | */ |
732 | rcu_read_lock(); |
733 | request->engine->submit_request(request); |
734 | rcu_read_unlock(); |
735 | break; |
736 | |
737 | case FENCE_FREE: |
738 | i915_request_put(request); |
739 | break; |
740 | } |
741 | |
742 | return NOTIFY_DONE0; |
743 | } |
744 | |
745 | static int __i915_sw_fence_call__attribute__((__aligned__(4))) |
746 | semaphore_notify(struct i915_sw_fence *fence, enum i915_sw_fence_notify state) |
747 | { |
748 | struct i915_request *rq = container_of(fence, typeof(*rq), semaphore)({ const __typeof( ((typeof(*rq) *)0)->semaphore ) *__mptr = (fence); (typeof(*rq) *)( (char *)__mptr - __builtin_offsetof (typeof(*rq), semaphore) );}); |
749 | |
750 | switch (state) { |
751 | case FENCE_COMPLETE: |
752 | break; |
753 | |
754 | case FENCE_FREE: |
755 | i915_request_put(rq); |
756 | break; |
757 | } |
758 | |
759 | return NOTIFY_DONE0; |
760 | } |
761 | |
762 | static void retire_requests(struct intel_timeline *tl) |
763 | { |
764 | struct i915_request *rq, *rn; |
765 | |
766 | list_for_each_entry_safe(rq, rn, &tl->requests, link)for (rq = ({ const __typeof( ((__typeof(*rq) *)0)->link ) * __mptr = ((&tl->requests)->next); (__typeof(*rq) *) ( (char *)__mptr - __builtin_offsetof(__typeof(*rq), link) ); }), rn = ({ const __typeof( ((__typeof(*rq) *)0)->link ) * __mptr = (rq->link.next); (__typeof(*rq) *)( (char *)__mptr - __builtin_offsetof(__typeof(*rq), link) );}); &rq-> link != (&tl->requests); rq = rn, rn = ({ const __typeof ( ((__typeof(*rn) *)0)->link ) *__mptr = (rn->link.next ); (__typeof(*rn) *)( (char *)__mptr - __builtin_offsetof(__typeof (*rn), link) );})) |
767 | if (!i915_request_retire(rq)) |
768 | break; |
769 | } |
770 | |
771 | static void __i915_request_ctor(void *); |
772 | |
773 | static noinline__attribute__((__noinline__)) struct i915_request * |
774 | request_alloc_slow(struct intel_timeline *tl, |
775 | struct i915_request **rsvd, |
776 | gfp_t gfp) |
777 | { |
778 | struct i915_request *rq; |
779 | |
780 | /* If we cannot wait, dip into our reserves */ |
781 | if (!gfpflags_allow_blocking(gfp)) { |
782 | rq = xchg(rsvd, NULL)__sync_lock_test_and_set(rsvd, ((void *)0)); |
783 | if (!rq) /* Use the normal failure path for one final WARN */ |
784 | goto out; |
785 | |
786 | return rq; |
787 | } |
788 | |
789 | if (list_empty(&tl->requests)) |
790 | goto out; |
791 | |
792 | /* Move our oldest request to the slab-cache (if not in use!) */ |
793 | rq = list_first_entry(&tl->requests, typeof(*rq), link)({ const __typeof( ((typeof(*rq) *)0)->link ) *__mptr = (( &tl->requests)->next); (typeof(*rq) *)( (char *)__mptr - __builtin_offsetof(typeof(*rq), link) );}); |
794 | i915_request_retire(rq); |
795 | |
796 | #ifdef __linux__ |
797 | rq = kmem_cache_alloc(global.slab_requests, |
798 | gfp | __GFP_RETRY_MAYFAIL0 | __GFP_NOWARN0); |
799 | #else |
800 | rq = pool_get(&global.slab_requests, |
801 | (gfp & GFP_NOWAIT0x0002) ? PR_NOWAIT0x0002 : PR_WAITOK0x0001); |
802 | if (rq) |
803 | __i915_request_ctor(rq); |
804 | #endif |
805 | if (rq) |
806 | return rq; |
807 | |
808 | /* Ratelimit ourselves to prevent oom from malicious clients */ |
809 | rq = list_last_entry(&tl->requests, typeof(*rq), link)({ const __typeof( ((typeof(*rq) *)0)->link ) *__mptr = (( &tl->requests)->prev); (typeof(*rq) *)( (char *)__mptr - __builtin_offsetof(typeof(*rq), link) );}); |
Value stored to 'rq' is never read | |
810 | cond_synchronize_rcu(rq->rcustate); |
811 | |
812 | /* Retire our old requests in the hope that we free some */ |
813 | retire_requests(tl); |
814 | |
815 | out: |
816 | #ifdef __linux__ |
817 | return kmem_cache_alloc(global.slab_requests, gfp); |
818 | #else |
819 | rq = pool_get(&global.slab_requests, |
820 | (gfp & GFP_NOWAIT0x0002) ? PR_NOWAIT0x0002 : PR_WAITOK0x0001); |
821 | if (rq) |
822 | __i915_request_ctor(rq); |
823 | return rq; |
824 | #endif |
825 | } |
826 | |
827 | static void __i915_request_ctor(void *arg) |
828 | { |
829 | struct i915_request *rq = arg; |
830 | |
831 | /* |
832 | * witness does not understand spin_lock_nested() |
833 | * order reversal in i915 with this lock |
834 | */ |
835 | mtx_init_flags(&rq->lock, IPL_TTY, NULL, MTX_NOWITNESS)do { (void)(((void *)0)); (void)(0x01); __mtx_init((&rq-> lock), ((((0x9)) > 0x0 && ((0x9)) < 0x9) ? 0x9 : ((0x9)))); } while (0); |
836 | i915_sched_node_init(&rq->sched); |
837 | i915_sw_fence_init(&rq->submit, submit_notify)__i915_sw_fence_init((&rq->submit), (submit_notify), ( (void *)0), ((void *)0)); |
838 | i915_sw_fence_init(&rq->semaphore, semaphore_notify)__i915_sw_fence_init((&rq->semaphore), (semaphore_notify ), ((void *)0), ((void *)0)); |
839 | |
840 | rq->capture_list = NULL((void *)0); |
841 | |
842 | init_llist_head(&rq->execute_cb); |
843 | } |
844 | |
845 | struct i915_request * |
846 | __i915_request_create(struct intel_context *ce, gfp_t gfp) |
847 | { |
848 | struct intel_timeline *tl = ce->timeline; |
849 | struct i915_request *rq; |
850 | u32 seqno; |
851 | int ret; |
852 | |
853 | might_sleep_if(gfpflags_allow_blocking(gfp))do { if (gfpflags_allow_blocking(gfp)) assertwaitok(); } while (0); |
854 | |
855 | /* Check that the caller provided an already pinned context */ |
856 | __intel_context_pin(ce); |
857 | |
858 | /* |
859 | * Beware: Dragons be flying overhead. |
860 | * |
861 | * We use RCU to look up requests in flight. The lookups may |
862 | * race with the request being allocated from the slab freelist. |
863 | * That is the request we are writing to here, may be in the process |
864 | * of being read by __i915_active_request_get_rcu(). As such, |
865 | * we have to be very careful when overwriting the contents. During |
866 | * the RCU lookup, we change chase the request->engine pointer, |
867 | * read the request->global_seqno and increment the reference count. |
868 | * |
869 | * The reference count is incremented atomically. If it is zero, |
870 | * the lookup knows the request is unallocated and complete. Otherwise, |
871 | * it is either still in use, or has been reallocated and reset |
872 | * with dma_fence_init(). This increment is safe for release as we |
873 | * check that the request we have a reference to and matches the active |
874 | * request. |
875 | * |
876 | * Before we increment the refcount, we chase the request->engine |
877 | * pointer. We must not call kmem_cache_zalloc() or else we set |
878 | * that pointer to NULL and cause a crash during the lookup. If |
879 | * we see the request is completed (based on the value of the |
880 | * old engine and seqno), the lookup is complete and reports NULL. |
881 | * If we decide the request is not completed (new engine or seqno), |
882 | * then we grab a reference and double check that it is still the |
883 | * active request - which it won't be and restart the lookup. |
884 | * |
885 | * Do not use kmem_cache_zalloc() here! |
886 | */ |
887 | #ifdef __linux__ |
888 | rq = kmem_cache_alloc(global.slab_requests, |
889 | gfp | __GFP_RETRY_MAYFAIL0 | __GFP_NOWARN0); |
890 | #else |
891 | rq = pool_get(&global.slab_requests, |
892 | (gfp & GFP_NOWAIT0x0002) ? PR_NOWAIT0x0002 : PR_WAITOK0x0001); |
893 | if (rq) |
894 | __i915_request_ctor(rq); |
895 | #endif |
896 | if (unlikely(!rq)__builtin_expect(!!(!rq), 0)) { |
897 | rq = request_alloc_slow(tl, &ce->engine->request_pool, gfp); |
898 | if (!rq) { |
899 | ret = -ENOMEM12; |
900 | goto err_unreserve; |
901 | } |
902 | } |
903 | |
904 | rq->context = ce; |
905 | rq->engine = ce->engine; |
906 | rq->ring = ce->ring; |
907 | rq->execution_mask = ce->engine->mask; |
908 | |
909 | ret = intel_timeline_get_seqno(tl, rq, &seqno); |
910 | if (ret) |
911 | goto err_free; |
912 | |
913 | dma_fence_init(&rq->fence, &i915_fence_ops, &rq->lock, |
914 | tl->fence_context, seqno); |
915 | |
916 | RCU_INIT_POINTER(rq->timeline, tl)do { (rq->timeline) = (tl); } while(0); |
917 | RCU_INIT_POINTER(rq->hwsp_cacheline, tl->hwsp_cacheline)do { (rq->hwsp_cacheline) = (tl->hwsp_cacheline); } while (0); |
918 | rq->hwsp_seqno = tl->hwsp_seqno; |
919 | GEM_BUG_ON(i915_request_completed(rq))((void)0); |
920 | |
921 | rq->rcustate = get_state_synchronize_rcu()0; /* acts as smp_mb() */ |
922 | |
923 | /* We bump the ref for the fence chain */ |
924 | i915_sw_fence_reinit(&i915_request_get(rq)->submit); |
925 | i915_sw_fence_reinit(&i915_request_get(rq)->semaphore); |
926 | |
927 | i915_sched_node_reinit(&rq->sched); |
928 | |
929 | /* No zalloc, everything must be cleared after use */ |
930 | rq->batch = NULL((void *)0); |
931 | GEM_BUG_ON(rq->capture_list)((void)0); |
932 | GEM_BUG_ON(!llist_empty(&rq->execute_cb))((void)0); |
933 | |
934 | /* |
935 | * Reserve space in the ring buffer for all the commands required to |
936 | * eventually emit this request. This is to guarantee that the |
937 | * i915_request_add() call can't fail. Note that the reserve may need |
938 | * to be redone if the request is not actually submitted straight |
939 | * away, e.g. because a GPU scheduler has deferred it. |
940 | * |
941 | * Note that due to how we add reserved_space to intel_ring_begin() |
942 | * we need to double our request to ensure that if we need to wrap |
943 | * around inside i915_request_add() there is sufficient space at |
944 | * the beginning of the ring as well. |
945 | */ |
946 | rq->reserved_space = |
947 | 2 * rq->engine->emit_fini_breadcrumb_dw * sizeof(u32); |
948 | |
949 | /* |
950 | * Record the position of the start of the request so that |
951 | * should we detect the updated seqno part-way through the |
952 | * GPU processing the request, we never over-estimate the |
953 | * position of the head. |
954 | */ |
955 | rq->head = rq->ring->emit; |
956 | |
957 | ret = rq->engine->request_alloc(rq); |
958 | if (ret) |
959 | goto err_unwind; |
960 | |
961 | rq->infix = rq->ring->emit; /* end of header; start of user payload */ |
962 | |
963 | intel_context_mark_active(ce); |
964 | list_add_tail_rcu(&rq->link, &tl->requests)list_add_tail(&rq->link, &tl->requests); |
965 | |
966 | return rq; |
967 | |
968 | err_unwind: |
969 | ce->ring->emit = rq->head; |
970 | |
971 | /* Make sure we didn't add ourselves to external state before freeing */ |
972 | GEM_BUG_ON(!list_empty(&rq->sched.signalers_list))((void)0); |
973 | GEM_BUG_ON(!list_empty(&rq->sched.waiters_list))((void)0); |
974 | |
975 | err_free: |
976 | #ifdef __linux__ |
977 | kmem_cache_free(global.slab_requests, rq); |
978 | #else |
979 | pool_put(&global.slab_requests, rq); |
980 | #endif |
981 | err_unreserve: |
982 | intel_context_unpin(ce); |
983 | return ERR_PTR(ret); |
984 | } |
985 | |
986 | struct i915_request * |
987 | i915_request_create(struct intel_context *ce) |
988 | { |
989 | struct i915_request *rq; |
990 | struct intel_timeline *tl; |
991 | |
992 | tl = intel_context_timeline_lock(ce); |
993 | if (IS_ERR(tl)) |
994 | return ERR_CAST(tl); |
995 | |
996 | /* Move our oldest request to the slab-cache (if not in use!) */ |
997 | rq = list_first_entry(&tl->requests, typeof(*rq), link)({ const __typeof( ((typeof(*rq) *)0)->link ) *__mptr = (( &tl->requests)->next); (typeof(*rq) *)( (char *)__mptr - __builtin_offsetof(typeof(*rq), link) );}); |
998 | if (!list_is_last(&rq->link, &tl->requests)) |
999 | i915_request_retire(rq); |
1000 | |
1001 | intel_context_enter(ce); |
1002 | rq = __i915_request_create(ce, GFP_KERNEL(0x0001 | 0x0004)); |
1003 | intel_context_exit(ce); /* active reference transferred to request */ |
1004 | if (IS_ERR(rq)) |
1005 | goto err_unlock; |
1006 | |
1007 | /* Check that we do not interrupt ourselves with a new request */ |
1008 | rq->cookie = lockdep_pin_lock(&tl->mutex)({ struct pin_cookie pc = {}; pc; }); |
1009 | |
1010 | return rq; |
1011 | |
1012 | err_unlock: |
1013 | intel_context_timeline_unlock(tl); |
1014 | return rq; |
1015 | } |
1016 | |
1017 | static int |
1018 | i915_request_await_start(struct i915_request *rq, struct i915_request *signal) |
1019 | { |
1020 | struct dma_fence *fence; |
1021 | int err; |
1022 | |
1023 | if (i915_request_timeline(rq) == rcu_access_pointer(signal->timeline)(signal->timeline)) |
1024 | return 0; |
1025 | |
1026 | if (i915_request_started(signal)) |
1027 | return 0; |
1028 | |
1029 | fence = NULL((void *)0); |
1030 | rcu_read_lock(); |
1031 | spin_lock_irq(&signal->lock)mtx_enter(&signal->lock); |
1032 | do { |
1033 | struct list_head *pos = READ_ONCE(signal->link.prev)({ typeof(signal->link.prev) __tmp = *(volatile typeof(signal ->link.prev) *)&(signal->link.prev); membar_datadep_consumer (); __tmp; }); |
1034 | struct i915_request *prev; |
1035 | |
1036 | /* Confirm signal has not been retired, the link is valid */ |
1037 | if (unlikely(i915_request_started(signal))__builtin_expect(!!(i915_request_started(signal)), 0)) |
1038 | break; |
1039 | |
1040 | /* Is signal the earliest request on its timeline? */ |
1041 | if (pos == &rcu_dereference(signal->timeline)(signal->timeline)->requests) |
1042 | break; |
1043 | |
1044 | /* |
1045 | * Peek at the request before us in the timeline. That |
1046 | * request will only be valid before it is retired, so |
1047 | * after acquiring a reference to it, confirm that it is |
1048 | * still part of the signaler's timeline. |
1049 | */ |
1050 | prev = list_entry(pos, typeof(*prev), link)({ const __typeof( ((typeof(*prev) *)0)->link ) *__mptr = ( pos); (typeof(*prev) *)( (char *)__mptr - __builtin_offsetof( typeof(*prev), link) );}); |
1051 | if (!i915_request_get_rcu(prev)) |
1052 | break; |
1053 | |
1054 | /* After the strong barrier, confirm prev is still attached */ |
1055 | if (unlikely(READ_ONCE(prev->link.next) != &signal->link)__builtin_expect(!!(({ typeof(prev->link.next) __tmp = *(volatile typeof(prev->link.next) *)&(prev->link.next); membar_datadep_consumer (); __tmp; }) != &signal->link), 0)) { |
1056 | i915_request_put(prev); |
1057 | break; |
1058 | } |
1059 | |
1060 | fence = &prev->fence; |
1061 | } while (0); |
1062 | spin_unlock_irq(&signal->lock)mtx_leave(&signal->lock); |
1063 | rcu_read_unlock(); |
1064 | if (!fence) |
1065 | return 0; |
1066 | |
1067 | err = 0; |
1068 | if (!intel_timeline_sync_is_later(i915_request_timeline(rq), fence)) |
1069 | err = i915_sw_fence_await_dma_fence(&rq->submit, |
1070 | fence, 0, |
1071 | I915_FENCE_GFP((0x0001 | 0x0004) | 0 | 0)); |
1072 | dma_fence_put(fence); |
1073 | |
1074 | return err; |
1075 | } |
1076 | |
1077 | static intel_engine_mask_t |
1078 | already_busywaiting(struct i915_request *rq) |
1079 | { |
1080 | /* |
1081 | * Polling a semaphore causes bus traffic, delaying other users of |
1082 | * both the GPU and CPU. We want to limit the impact on others, |
1083 | * while taking advantage of early submission to reduce GPU |
1084 | * latency. Therefore we restrict ourselves to not using more |
1085 | * than one semaphore from each source, and not using a semaphore |
1086 | * if we have detected the engine is saturated (i.e. would not be |
1087 | * submitted early and cause bus traffic reading an already passed |
1088 | * semaphore). |
1089 | * |
1090 | * See the are-we-too-late? check in __i915_request_submit(). |
1091 | */ |
1092 | return rq->sched.semaphores | READ_ONCE(rq->engine->saturated)({ typeof(rq->engine->saturated) __tmp = *(volatile typeof (rq->engine->saturated) *)&(rq->engine->saturated ); membar_datadep_consumer(); __tmp; }); |
1093 | } |
1094 | |
1095 | static int |
1096 | __emit_semaphore_wait(struct i915_request *to, |
1097 | struct i915_request *from, |
1098 | u32 seqno) |
1099 | { |
1100 | const int has_token = INTEL_GEN(to->engine->i915)((&(to->engine->i915)->__info)->gen) >= 12; |
1101 | u32 hwsp_offset; |
1102 | int len, err; |
1103 | u32 *cs; |
1104 | |
1105 | GEM_BUG_ON(INTEL_GEN(to->engine->i915) < 8)((void)0); |
1106 | GEM_BUG_ON(i915_request_has_initial_breadcrumb(to))((void)0); |
1107 | |
1108 | /* We need to pin the signaler's HWSP until we are finished reading. */ |
1109 | err = intel_timeline_read_hwsp(from, to, &hwsp_offset); |
1110 | if (err) |
1111 | return err; |
1112 | |
1113 | len = 4; |
1114 | if (has_token) |
1115 | len += 2; |
1116 | |
1117 | cs = intel_ring_begin(to, len); |
1118 | if (IS_ERR(cs)) |
1119 | return PTR_ERR(cs); |
1120 | |
1121 | /* |
1122 | * Using greater-than-or-equal here means we have to worry |
1123 | * about seqno wraparound. To side step that issue, we swap |
1124 | * the timeline HWSP upon wrapping, so that everyone listening |
1125 | * for the old (pre-wrap) values do not see the much smaller |
1126 | * (post-wrap) values than they were expecting (and so wait |
1127 | * forever). |
1128 | */ |
1129 | *cs++ = (MI_SEMAPHORE_WAIT(((0x1c) << 23) | (2)) | |
1130 | MI_SEMAPHORE_GLOBAL_GTT(1<<22) | |
1131 | MI_SEMAPHORE_POLL(1 << 15) | |
1132 | MI_SEMAPHORE_SAD_GTE_SDD(1 << 12)) + |
1133 | has_token; |
1134 | *cs++ = seqno; |
1135 | *cs++ = hwsp_offset; |
1136 | *cs++ = 0; |
1137 | if (has_token) { |
1138 | *cs++ = 0; |
1139 | *cs++ = MI_NOOP(((0) << 23) | (0)); |
1140 | } |
1141 | |
1142 | intel_ring_advance(to, cs); |
1143 | return 0; |
1144 | } |
1145 | |
1146 | static int |
1147 | emit_semaphore_wait(struct i915_request *to, |
1148 | struct i915_request *from, |
1149 | gfp_t gfp) |
1150 | { |
1151 | const intel_engine_mask_t mask = READ_ONCE(from->engine)({ typeof(from->engine) __tmp = *(volatile typeof(from-> engine) *)&(from->engine); membar_datadep_consumer(); __tmp ; })->mask; |
1152 | struct i915_sw_fence *wait = &to->submit; |
1153 | |
1154 | if (!intel_context_use_semaphores(to->context)) |
1155 | goto await_fence; |
1156 | |
1157 | if (i915_request_has_initial_breadcrumb(to)) |
1158 | goto await_fence; |
1159 | |
1160 | if (!rcu_access_pointer(from->hwsp_cacheline)(from->hwsp_cacheline)) |
1161 | goto await_fence; |
1162 | |
1163 | /* |
1164 | * If this or its dependents are waiting on an external fence |
1165 | * that may fail catastrophically, then we want to avoid using |
1166 | * sempahores as they bypass the fence signaling metadata, and we |
1167 | * lose the fence->error propagation. |
1168 | */ |
1169 | if (from->sched.flags & I915_SCHED_HAS_EXTERNAL_CHAIN(1UL << (0))) |
1170 | goto await_fence; |
1171 | |
1172 | /* Just emit the first semaphore we see as request space is limited. */ |
1173 | if (already_busywaiting(to) & mask) |
1174 | goto await_fence; |
1175 | |
1176 | if (i915_request_await_start(to, from) < 0) |
1177 | goto await_fence; |
1178 | |
1179 | /* Only submit our spinner after the signaler is running! */ |
1180 | if (__await_execution(to, from, NULL((void *)0), gfp)) |
1181 | goto await_fence; |
1182 | |
1183 | if (__emit_semaphore_wait(to, from, from->fence.seqno)) |
1184 | goto await_fence; |
1185 | |
1186 | to->sched.semaphores |= mask; |
1187 | wait = &to->semaphore; |
1188 | |
1189 | await_fence: |
1190 | return i915_sw_fence_await_dma_fence(wait, |
1191 | &from->fence, 0, |
1192 | I915_FENCE_GFP((0x0001 | 0x0004) | 0 | 0)); |
1193 | } |
1194 | |
1195 | static bool_Bool intel_timeline_sync_has_start(struct intel_timeline *tl, |
1196 | struct dma_fence *fence) |
1197 | { |
1198 | return __intel_timeline_sync_is_later(tl, |
1199 | fence->context, |
1200 | fence->seqno - 1); |
1201 | } |
1202 | |
1203 | static int intel_timeline_sync_set_start(struct intel_timeline *tl, |
1204 | const struct dma_fence *fence) |
1205 | { |
1206 | return __intel_timeline_sync_set(tl, fence->context, fence->seqno - 1); |
1207 | } |
1208 | |
1209 | static int |
1210 | __i915_request_await_execution(struct i915_request *to, |
1211 | struct i915_request *from, |
1212 | void (*hook)(struct i915_request *rq, |
1213 | struct dma_fence *signal)) |
1214 | { |
1215 | int err; |
1216 | |
1217 | GEM_BUG_ON(intel_context_is_barrier(from->context))((void)0); |
1218 | |
1219 | /* Submit both requests at the same time */ |
1220 | err = __await_execution(to, from, hook, I915_FENCE_GFP((0x0001 | 0x0004) | 0 | 0)); |
1221 | if (err) |
1222 | return err; |
1223 | |
1224 | /* Squash repeated depenendices to the same timelines */ |
1225 | if (intel_timeline_sync_has_start(i915_request_timeline(to), |
1226 | &from->fence)) |
1227 | return 0; |
1228 | |
1229 | /* |
1230 | * Wait until the start of this request. |
1231 | * |
1232 | * The execution cb fires when we submit the request to HW. But in |
1233 | * many cases this may be long before the request itself is ready to |
1234 | * run (consider that we submit 2 requests for the same context, where |
1235 | * the request of interest is behind an indefinite spinner). So we hook |
1236 | * up to both to reduce our queues and keep the execution lag minimised |
1237 | * in the worst case, though we hope that the await_start is elided. |
1238 | */ |
1239 | err = i915_request_await_start(to, from); |
1240 | if (err < 0) |
1241 | return err; |
1242 | |
1243 | /* |
1244 | * Ensure both start together [after all semaphores in signal] |
1245 | * |
1246 | * Now that we are queued to the HW at roughly the same time (thanks |
1247 | * to the execute cb) and are ready to run at roughly the same time |
1248 | * (thanks to the await start), our signaler may still be indefinitely |
1249 | * delayed by waiting on a semaphore from a remote engine. If our |
1250 | * signaler depends on a semaphore, so indirectly do we, and we do not |
1251 | * want to start our payload until our signaler also starts theirs. |
1252 | * So we wait. |
1253 | * |
1254 | * However, there is also a second condition for which we need to wait |
1255 | * for the precise start of the signaler. Consider that the signaler |
1256 | * was submitted in a chain of requests following another context |
1257 | * (with just an ordinary intra-engine fence dependency between the |
1258 | * two). In this case the signaler is queued to HW, but not for |
1259 | * immediate execution, and so we must wait until it reaches the |
1260 | * active slot. |
1261 | */ |
1262 | if (intel_engine_has_semaphores(to->engine) && |
1263 | !i915_request_has_initial_breadcrumb(to)) { |
1264 | err = __emit_semaphore_wait(to, from, from->fence.seqno - 1); |
1265 | if (err < 0) |
1266 | return err; |
1267 | } |
1268 | |
1269 | /* Couple the dependency tree for PI on this exposed to->fence */ |
1270 | if (to->engine->schedule) { |
1271 | err = i915_sched_node_add_dependency(&to->sched, |
1272 | &from->sched, |
1273 | I915_DEPENDENCY_WEAK(1UL << (2))); |
1274 | if (err < 0) |
1275 | return err; |
1276 | } |
1277 | |
1278 | return intel_timeline_sync_set_start(i915_request_timeline(to), |
1279 | &from->fence); |
1280 | } |
1281 | |
1282 | static void mark_external(struct i915_request *rq) |
1283 | { |
1284 | /* |
1285 | * The downside of using semaphores is that we lose metadata passing |
1286 | * along the signaling chain. This is particularly nasty when we |
1287 | * need to pass along a fatal error such as EFAULT or EDEADLK. For |
1288 | * fatal errors we want to scrub the request before it is executed, |
1289 | * which means that we cannot preload the request onto HW and have |
1290 | * it wait upon a semaphore. |
1291 | */ |
1292 | rq->sched.flags |= I915_SCHED_HAS_EXTERNAL_CHAIN(1UL << (0)); |
1293 | } |
1294 | |
1295 | static int |
1296 | __i915_request_await_external(struct i915_request *rq, struct dma_fence *fence) |
1297 | { |
1298 | mark_external(rq); |
1299 | return i915_sw_fence_await_dma_fence(&rq->submit, fence, |
1300 | i915_fence_context_timeout(rq->engine->i915, |
1301 | fence->context), |
1302 | I915_FENCE_GFP((0x0001 | 0x0004) | 0 | 0)); |
1303 | } |
1304 | |
1305 | static int |
1306 | i915_request_await_external(struct i915_request *rq, struct dma_fence *fence) |
1307 | { |
1308 | struct dma_fence *iter; |
1309 | int err = 0; |
1310 | |
1311 | if (!to_dma_fence_chain(fence)) |
1312 | return __i915_request_await_external(rq, fence); |
1313 | |
1314 | dma_fence_chain_for_each(iter, fence)for (iter = dma_fence_get(fence); iter != ((void *)0); iter = dma_fence_chain_next(iter)) { |
1315 | struct dma_fence_chain *chain = to_dma_fence_chain(iter); |
1316 | |
1317 | if (!dma_fence_is_i915(chain->fence)) { |
1318 | err = __i915_request_await_external(rq, iter); |
1319 | break; |
1320 | } |
1321 | |
1322 | err = i915_request_await_dma_fence(rq, chain->fence); |
1323 | if (err < 0) |
1324 | break; |
1325 | } |
1326 | |
1327 | dma_fence_put(iter); |
1328 | return err; |
1329 | } |
1330 | |
1331 | int |
1332 | i915_request_await_execution(struct i915_request *rq, |
1333 | struct dma_fence *fence, |
1334 | void (*hook)(struct i915_request *rq, |
1335 | struct dma_fence *signal)) |
1336 | { |
1337 | struct dma_fence **child = &fence; |
1338 | unsigned int nchild = 1; |
1339 | int ret; |
1340 | |
1341 | if (dma_fence_is_array(fence)) { |
1342 | struct dma_fence_array *array = to_dma_fence_array(fence); |
1343 | |
1344 | /* XXX Error for signal-on-any fence arrays */ |
1345 | |
1346 | child = array->fences; |
1347 | nchild = array->num_fences; |
1348 | GEM_BUG_ON(!nchild)((void)0); |
1349 | } |
1350 | |
1351 | do { |
1352 | fence = *child++; |
1353 | if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) |
1354 | continue; |
1355 | |
1356 | if (fence->context == rq->fence.context) |
1357 | continue; |
1358 | |
1359 | /* |
1360 | * We don't squash repeated fence dependencies here as we |
1361 | * want to run our callback in all cases. |
1362 | */ |
1363 | |
1364 | if (dma_fence_is_i915(fence)) |
1365 | ret = __i915_request_await_execution(rq, |
1366 | to_request(fence), |
1367 | hook); |
1368 | else |
1369 | ret = i915_request_await_external(rq, fence); |
1370 | if (ret < 0) |
1371 | return ret; |
1372 | } while (--nchild); |
1373 | |
1374 | return 0; |
1375 | } |
1376 | |
1377 | static int |
1378 | await_request_submit(struct i915_request *to, struct i915_request *from) |
1379 | { |
1380 | /* |
1381 | * If we are waiting on a virtual engine, then it may be |
1382 | * constrained to execute on a single engine *prior* to submission. |
1383 | * When it is submitted, it will be first submitted to the virtual |
1384 | * engine and then passed to the physical engine. We cannot allow |
1385 | * the waiter to be submitted immediately to the physical engine |
1386 | * as it may then bypass the virtual request. |
1387 | */ |
1388 | if (to->engine == READ_ONCE(from->engine)({ typeof(from->engine) __tmp = *(volatile typeof(from-> engine) *)&(from->engine); membar_datadep_consumer(); __tmp ; })) |
1389 | return i915_sw_fence_await_sw_fence_gfp(&to->submit, |
1390 | &from->submit, |
1391 | I915_FENCE_GFP((0x0001 | 0x0004) | 0 | 0)); |
1392 | else |
1393 | return __i915_request_await_execution(to, from, NULL((void *)0)); |
1394 | } |
1395 | |
1396 | static int |
1397 | i915_request_await_request(struct i915_request *to, struct i915_request *from) |
1398 | { |
1399 | int ret; |
1400 | |
1401 | GEM_BUG_ON(to == from)((void)0); |
1402 | GEM_BUG_ON(to->timeline == from->timeline)((void)0); |
1403 | |
1404 | if (i915_request_completed(from)) { |
1405 | i915_sw_fence_set_error_once(&to->submit, from->fence.error); |
1406 | return 0; |
1407 | } |
1408 | |
1409 | if (to->engine->schedule) { |
1410 | ret = i915_sched_node_add_dependency(&to->sched, |
1411 | &from->sched, |
1412 | I915_DEPENDENCY_EXTERNAL(1UL << (1))); |
1413 | if (ret < 0) |
1414 | return ret; |
1415 | } |
1416 | |
1417 | if (is_power_of_2(to->execution_mask | READ_ONCE(from->execution_mask))(((to->execution_mask | ({ typeof(from->execution_mask) __tmp = *(volatile typeof(from->execution_mask) *)&(from ->execution_mask); membar_datadep_consumer(); __tmp; })) != 0) && (((to->execution_mask | ({ typeof(from-> execution_mask) __tmp = *(volatile typeof(from->execution_mask ) *)&(from->execution_mask); membar_datadep_consumer() ; __tmp; })) - 1) & (to->execution_mask | ({ typeof(from ->execution_mask) __tmp = *(volatile typeof(from->execution_mask ) *)&(from->execution_mask); membar_datadep_consumer() ; __tmp; }))) == 0)) |
1418 | ret = await_request_submit(to, from); |
1419 | else |
1420 | ret = emit_semaphore_wait(to, from, I915_FENCE_GFP((0x0001 | 0x0004) | 0 | 0)); |
1421 | if (ret < 0) |
1422 | return ret; |
1423 | |
1424 | return 0; |
1425 | } |
1426 | |
1427 | int |
1428 | i915_request_await_dma_fence(struct i915_request *rq, struct dma_fence *fence) |
1429 | { |
1430 | struct dma_fence **child = &fence; |
1431 | unsigned int nchild = 1; |
1432 | int ret; |
1433 | |
1434 | /* |
1435 | * Note that if the fence-array was created in signal-on-any mode, |
1436 | * we should *not* decompose it into its individual fences. However, |
1437 | * we don't currently store which mode the fence-array is operating |
1438 | * in. Fortunately, the only user of signal-on-any is private to |
1439 | * amdgpu and we should not see any incoming fence-array from |
1440 | * sync-file being in signal-on-any mode. |
1441 | */ |
1442 | if (dma_fence_is_array(fence)) { |
1443 | struct dma_fence_array *array = to_dma_fence_array(fence); |
1444 | |
1445 | child = array->fences; |
1446 | nchild = array->num_fences; |
1447 | GEM_BUG_ON(!nchild)((void)0); |
1448 | } |
1449 | |
1450 | do { |
1451 | fence = *child++; |
1452 | if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) |
1453 | continue; |
1454 | |
1455 | /* |
1456 | * Requests on the same timeline are explicitly ordered, along |
1457 | * with their dependencies, by i915_request_add() which ensures |
1458 | * that requests are submitted in-order through each ring. |
1459 | */ |
1460 | if (fence->context == rq->fence.context) |
1461 | continue; |
1462 | |
1463 | /* Squash repeated waits to the same timelines */ |
1464 | if (fence->context && |
1465 | intel_timeline_sync_is_later(i915_request_timeline(rq), |
1466 | fence)) |
1467 | continue; |
1468 | |
1469 | if (dma_fence_is_i915(fence)) |
1470 | ret = i915_request_await_request(rq, to_request(fence)); |
1471 | else |
1472 | ret = i915_request_await_external(rq, fence); |
1473 | if (ret < 0) |
1474 | return ret; |
1475 | |
1476 | /* Record the latest fence used against each timeline */ |
1477 | if (fence->context) |
1478 | intel_timeline_sync_set(i915_request_timeline(rq), |
1479 | fence); |
1480 | } while (--nchild); |
1481 | |
1482 | return 0; |
1483 | } |
1484 | |
1485 | /** |
1486 | * i915_request_await_object - set this request to (async) wait upon a bo |
1487 | * @to: request we are wishing to use |
1488 | * @obj: object which may be in use on another ring. |
1489 | * @write: whether the wait is on behalf of a writer |
1490 | * |
1491 | * This code is meant to abstract object synchronization with the GPU. |
1492 | * Conceptually we serialise writes between engines inside the GPU. |
1493 | * We only allow one engine to write into a buffer at any time, but |
1494 | * multiple readers. To ensure each has a coherent view of memory, we must: |
1495 | * |
1496 | * - If there is an outstanding write request to the object, the new |
1497 | * request must wait for it to complete (either CPU or in hw, requests |
1498 | * on the same ring will be naturally ordered). |
1499 | * |
1500 | * - If we are a write request (pending_write_domain is set), the new |
1501 | * request must wait for outstanding read requests to complete. |
1502 | * |
1503 | * Returns 0 if successful, else propagates up the lower layer error. |
1504 | */ |
1505 | int |
1506 | i915_request_await_object(struct i915_request *to, |
1507 | struct drm_i915_gem_object *obj, |
1508 | bool_Bool write) |
1509 | { |
1510 | struct dma_fence *excl; |
1511 | int ret = 0; |
1512 | |
1513 | if (write) { |
1514 | struct dma_fence **shared; |
1515 | unsigned int count, i; |
1516 | |
1517 | ret = dma_resv_get_fences_rcu(obj->base.resv, |
1518 | &excl, &count, &shared); |
1519 | if (ret) |
1520 | return ret; |
1521 | |
1522 | for (i = 0; i < count; i++) { |
1523 | ret = i915_request_await_dma_fence(to, shared[i]); |
1524 | if (ret) |
1525 | break; |
1526 | |
1527 | dma_fence_put(shared[i]); |
1528 | } |
1529 | |
1530 | for (; i < count; i++) |
1531 | dma_fence_put(shared[i]); |
1532 | kfree(shared); |
1533 | } else { |
1534 | excl = dma_resv_get_excl_rcu(obj->base.resv); |
1535 | } |
1536 | |
1537 | if (excl) { |
1538 | if (ret == 0) |
1539 | ret = i915_request_await_dma_fence(to, excl); |
1540 | |
1541 | dma_fence_put(excl); |
1542 | } |
1543 | |
1544 | return ret; |
1545 | } |
1546 | |
1547 | static struct i915_request * |
1548 | __i915_request_add_to_timeline(struct i915_request *rq) |
1549 | { |
1550 | struct intel_timeline *timeline = i915_request_timeline(rq); |
1551 | struct i915_request *prev; |
1552 | |
1553 | /* |
1554 | * Dependency tracking and request ordering along the timeline |
1555 | * is special cased so that we can eliminate redundant ordering |
1556 | * operations while building the request (we know that the timeline |
1557 | * itself is ordered, and here we guarantee it). |
1558 | * |
1559 | * As we know we will need to emit tracking along the timeline, |
1560 | * we embed the hooks into our request struct -- at the cost of |
1561 | * having to have specialised no-allocation interfaces (which will |
1562 | * be beneficial elsewhere). |
1563 | * |
1564 | * A second benefit to open-coding i915_request_await_request is |
1565 | * that we can apply a slight variant of the rules specialised |
1566 | * for timelines that jump between engines (such as virtual engines). |
1567 | * If we consider the case of virtual engine, we must emit a dma-fence |
1568 | * to prevent scheduling of the second request until the first is |
1569 | * complete (to maximise our greedy late load balancing) and this |
1570 | * precludes optimising to use semaphores serialisation of a single |
1571 | * timeline across engines. |
1572 | */ |
1573 | prev = to_request(__i915_active_fence_set(&timeline->last_request, |
1574 | &rq->fence)); |
1575 | if (prev && !i915_request_completed(prev)) { |
1576 | /* |
1577 | * The requests are supposed to be kept in order. However, |
1578 | * we need to be wary in case the timeline->last_request |
1579 | * is used as a barrier for external modification to this |
1580 | * context. |
1581 | */ |
1582 | GEM_BUG_ON(prev->context == rq->context &&((void)0) |
1583 | i915_seqno_passed(prev->fence.seqno,((void)0) |
1584 | rq->fence.seqno))((void)0); |
1585 | |
1586 | if (is_power_of_2(READ_ONCE(prev->engine)->mask | rq->engine->mask)(((({ typeof(prev->engine) __tmp = *(volatile typeof(prev-> engine) *)&(prev->engine); membar_datadep_consumer(); __tmp ; })->mask | rq->engine->mask) != 0) && (((( { typeof(prev->engine) __tmp = *(volatile typeof(prev-> engine) *)&(prev->engine); membar_datadep_consumer(); __tmp ; })->mask | rq->engine->mask) - 1) & (({ typeof (prev->engine) __tmp = *(volatile typeof(prev->engine) * )&(prev->engine); membar_datadep_consumer(); __tmp; }) ->mask | rq->engine->mask)) == 0)) |
1587 | i915_sw_fence_await_sw_fence(&rq->submit, |
1588 | &prev->submit, |
1589 | &rq->submitq); |
1590 | else |
1591 | __i915_sw_fence_await_dma_fence(&rq->submit, |
1592 | &prev->fence, |
1593 | &rq->dmaq); |
1594 | if (rq->engine->schedule) |
1595 | __i915_sched_node_add_dependency(&rq->sched, |
1596 | &prev->sched, |
1597 | &rq->dep, |
1598 | 0); |
1599 | } |
1600 | |
1601 | /* |
1602 | * Make sure that no request gazumped us - if it was allocated after |
1603 | * our i915_request_alloc() and called __i915_request_add() before |
1604 | * us, the timeline will hold its seqno which is later than ours. |
1605 | */ |
1606 | GEM_BUG_ON(timeline->seqno != rq->fence.seqno)((void)0); |
1607 | |
1608 | return prev; |
1609 | } |
1610 | |
1611 | /* |
1612 | * NB: This function is not allowed to fail. Doing so would mean the the |
1613 | * request is not being tracked for completion but the work itself is |
1614 | * going to happen on the hardware. This would be a Bad Thing(tm). |
1615 | */ |
1616 | struct i915_request *__i915_request_commit(struct i915_request *rq) |
1617 | { |
1618 | struct intel_engine_cs *engine = rq->engine; |
1619 | struct intel_ring *ring = rq->ring; |
1620 | u32 *cs; |
1621 | |
1622 | RQ_TRACE(rq, "\n")do { const struct i915_request *rq__ = (rq); do { const struct intel_engine_cs *e__ __attribute__((__unused__)) = (rq__-> engine); do { } while (0); } while (0); } while (0); |
1623 | |
1624 | /* |
1625 | * To ensure that this call will not fail, space for its emissions |
1626 | * should already have been reserved in the ring buffer. Let the ring |
1627 | * know that it is time to use that space up. |
1628 | */ |
1629 | GEM_BUG_ON(rq->reserved_space > ring->space)((void)0); |
1630 | rq->reserved_space = 0; |
1631 | rq->emitted_jiffies = jiffies; |
1632 | |
1633 | /* |
1634 | * Record the position of the start of the breadcrumb so that |
1635 | * should we detect the updated seqno part-way through the |
1636 | * GPU processing the request, we never over-estimate the |
1637 | * position of the ring's HEAD. |
1638 | */ |
1639 | cs = intel_ring_begin(rq, engine->emit_fini_breadcrumb_dw); |
1640 | GEM_BUG_ON(IS_ERR(cs))((void)0); |
1641 | rq->postfix = intel_ring_offset(rq, cs); |
1642 | |
1643 | return __i915_request_add_to_timeline(rq); |
1644 | } |
1645 | |
1646 | void __i915_request_queue(struct i915_request *rq, |
1647 | const struct i915_sched_attr *attr) |
1648 | { |
1649 | /* |
1650 | * Let the backend know a new request has arrived that may need |
1651 | * to adjust the existing execution schedule due to a high priority |
1652 | * request - i.e. we may want to preempt the current request in order |
1653 | * to run a high priority dependency chain *before* we can execute this |
1654 | * request. |
1655 | * |
1656 | * This is called before the request is ready to run so that we can |
1657 | * decide whether to preempt the entire chain so that it is ready to |
1658 | * run at the earliest possible convenience. |
1659 | */ |
1660 | if (attr && rq->engine->schedule) |
1661 | rq->engine->schedule(rq, attr); |
1662 | i915_sw_fence_commit(&rq->semaphore); |
1663 | i915_sw_fence_commit(&rq->submit); |
1664 | } |
1665 | |
1666 | void i915_request_add(struct i915_request *rq) |
1667 | { |
1668 | struct intel_timeline * const tl = i915_request_timeline(rq); |
1669 | struct i915_sched_attr attr = {}; |
1670 | struct i915_gem_context *ctx; |
1671 | |
1672 | lockdep_assert_held(&tl->mutex)do { (void)(&tl->mutex); } while(0); |
1673 | lockdep_unpin_lock(&tl->mutex, rq->cookie); |
1674 | |
1675 | trace_i915_request_add(rq); |
1676 | __i915_request_commit(rq); |
1677 | |
1678 | /* XXX placeholder for selftests */ |
1679 | rcu_read_lock(); |
1680 | ctx = rcu_dereference(rq->context->gem_context)(rq->context->gem_context); |
1681 | if (ctx) |
1682 | attr = ctx->sched; |
1683 | rcu_read_unlock(); |
1684 | |
1685 | __i915_request_queue(rq, &attr); |
1686 | |
1687 | mutex_unlock(&tl->mutex)rw_exit_write(&tl->mutex); |
1688 | } |
1689 | |
1690 | static unsigned long local_clock_ns(unsigned int *cpu) |
1691 | { |
1692 | unsigned long t; |
1693 | |
1694 | /* |
1695 | * Cheaply and approximately convert from nanoseconds to microseconds. |
1696 | * The result and subsequent calculations are also defined in the same |
1697 | * approximate microseconds units. The principal source of timing |
1698 | * error here is from the simple truncation. |
1699 | * |
1700 | * Note that local_clock() is only defined wrt to the current CPU; |
1701 | * the comparisons are no longer valid if we switch CPUs. Instead of |
1702 | * blocking preemption for the entire busywait, we can detect the CPU |
1703 | * switch and use that as indicator of system load and a reason to |
1704 | * stop busywaiting, see busywait_stop(). |
1705 | */ |
1706 | *cpu = get_cpu()(({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_cpuid); |
1707 | t = local_clock(); |
1708 | put_cpu(); |
1709 | |
1710 | return t; |
1711 | } |
1712 | |
1713 | static bool_Bool busywait_stop(unsigned long timeout, unsigned int cpu) |
1714 | { |
1715 | unsigned int this_cpu; |
1716 | |
1717 | if (time_after(local_clock_ns(&this_cpu), timeout)((long)(timeout) - (long)(local_clock_ns(&this_cpu)) < 0)) |
1718 | return true1; |
1719 | |
1720 | return this_cpu != cpu; |
1721 | } |
1722 | |
1723 | static bool_Bool __i915_spin_request(struct i915_request * const rq, int state) |
1724 | { |
1725 | unsigned long timeout_ns; |
1726 | unsigned int cpu; |
1727 | |
1728 | /* |
1729 | * Only wait for the request if we know it is likely to complete. |
1730 | * |
1731 | * We don't track the timestamps around requests, nor the average |
1732 | * request length, so we do not have a good indicator that this |
1733 | * request will complete within the timeout. What we do know is the |
1734 | * order in which requests are executed by the context and so we can |
1735 | * tell if the request has been started. If the request is not even |
1736 | * running yet, it is a fair assumption that it will not complete |
1737 | * within our relatively short timeout. |
1738 | */ |
1739 | if (!i915_request_is_running(rq)) |
1740 | return false0; |
1741 | |
1742 | /* |
1743 | * When waiting for high frequency requests, e.g. during synchronous |
1744 | * rendering split between the CPU and GPU, the finite amount of time |
1745 | * required to set up the irq and wait upon it limits the response |
1746 | * rate. By busywaiting on the request completion for a short while we |
1747 | * can service the high frequency waits as quick as possible. However, |
1748 | * if it is a slow request, we want to sleep as quickly as possible. |
1749 | * The tradeoff between waiting and sleeping is roughly the time it |
1750 | * takes to sleep on a request, on the order of a microsecond. |
1751 | */ |
1752 | |
1753 | timeout_ns = READ_ONCE(rq->engine->props.max_busywait_duration_ns)({ typeof(rq->engine->props.max_busywait_duration_ns) __tmp = *(volatile typeof(rq->engine->props.max_busywait_duration_ns ) *)&(rq->engine->props.max_busywait_duration_ns); membar_datadep_consumer (); __tmp; }); |
1754 | timeout_ns += local_clock_ns(&cpu); |
1755 | do { |
1756 | if (dma_fence_is_signaled(&rq->fence)) |
1757 | return true1; |
1758 | |
1759 | if (signal_pending_state(state, current)((state) & 0x100 ? (((({struct cpu_info *__ci; asm volatile ("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc)->p_siglist | (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc)->p_p->ps_siglist) & ~(({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci ) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci; })->ci_curproc)->p_sigmask) : 0)) |
1760 | break; |
1761 | |
1762 | if (busywait_stop(timeout_ns, cpu)) |
1763 | break; |
1764 | |
1765 | cpu_relax(); |
1766 | } while (!drm_need_resched()(({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_schedstate.spc_schedflags & 0x0002)); |
1767 | |
1768 | return false0; |
1769 | } |
1770 | |
1771 | struct request_wait { |
1772 | struct dma_fence_cb cb; |
1773 | #ifdef __linux__ |
1774 | struct task_struct *tsk; |
1775 | #else |
1776 | struct proc *tsk; |
1777 | #endif |
1778 | }; |
1779 | |
1780 | static void request_wait_wake(struct dma_fence *fence, struct dma_fence_cb *cb) |
1781 | { |
1782 | struct request_wait *wait = container_of(cb, typeof(*wait), cb)({ const __typeof( ((typeof(*wait) *)0)->cb ) *__mptr = (cb ); (typeof(*wait) *)( (char *)__mptr - __builtin_offsetof(typeof (*wait), cb) );}); |
1783 | |
1784 | wake_up_process(fetch_and_zero(&wait->tsk)({ typeof(*&wait->tsk) __T = *(&wait->tsk); *(& wait->tsk) = (typeof(*&wait->tsk))0; __T; })); |
1785 | } |
1786 | |
1787 | /** |
1788 | * i915_request_wait - wait until execution of request has finished |
1789 | * @rq: the request to wait upon |
1790 | * @flags: how to wait |
1791 | * @timeout: how long to wait in jiffies |
1792 | * |
1793 | * i915_request_wait() waits for the request to be completed, for a |
1794 | * maximum of @timeout jiffies (with MAX_SCHEDULE_TIMEOUT implying an |
1795 | * unbounded wait). |
1796 | * |
1797 | * Returns the remaining time (in jiffies) if the request completed, which may |
1798 | * be zero or -ETIME if the request is unfinished after the timeout expires. |
1799 | * May return -EINTR is called with I915_WAIT_INTERRUPTIBLE and a signal is |
1800 | * pending before the request completes. |
1801 | */ |
1802 | long i915_request_wait(struct i915_request *rq, |
1803 | unsigned int flags, |
1804 | long timeout) |
1805 | { |
1806 | const int state = flags & I915_WAIT_INTERRUPTIBLE(1UL << (0)) ? |
1807 | TASK_INTERRUPTIBLE0x100 : TASK_UNINTERRUPTIBLE0; |
1808 | struct request_wait wait; |
1809 | |
1810 | might_sleep()assertwaitok(); |
1811 | GEM_BUG_ON(timeout < 0)((void)0); |
1812 | |
1813 | if (dma_fence_is_signaled(&rq->fence)) |
1814 | return timeout; |
1815 | |
1816 | if (!timeout) |
1817 | return -ETIME60; |
1818 | |
1819 | trace_i915_request_wait_begin(rq, flags); |
1820 | |
1821 | /* |
1822 | * We must never wait on the GPU while holding a lock as we |
1823 | * may need to perform a GPU reset. So while we don't need to |
1824 | * serialise wait/reset with an explicit lock, we do want |
1825 | * lockdep to detect potential dependency cycles. |
1826 | */ |
1827 | mutex_acquire(&rq->engine->gt->reset.mutex.dep_map, 0, 0, _THIS_IP_); |
1828 | |
1829 | /* |
1830 | * Optimistic spin before touching IRQs. |
1831 | * |
1832 | * We may use a rather large value here to offset the penalty of |
1833 | * switching away from the active task. Frequently, the client will |
1834 | * wait upon an old swapbuffer to throttle itself to remain within a |
1835 | * frame of the gpu. If the client is running in lockstep with the gpu, |
1836 | * then it should not be waiting long at all, and a sleep now will incur |
1837 | * extra scheduler latency in producing the next frame. To try to |
1838 | * avoid adding the cost of enabling/disabling the interrupt to the |
1839 | * short wait, we first spin to see if the request would have completed |
1840 | * in the time taken to setup the interrupt. |
1841 | * |
1842 | * We need upto 5us to enable the irq, and upto 20us to hide the |
1843 | * scheduler latency of a context switch, ignoring the secondary |
1844 | * impacts from a context switch such as cache eviction. |
1845 | * |
1846 | * The scheme used for low-latency IO is called "hybrid interrupt |
1847 | * polling". The suggestion there is to sleep until just before you |
1848 | * expect to be woken by the device interrupt and then poll for its |
1849 | * completion. That requires having a good predictor for the request |
1850 | * duration, which we currently lack. |
1851 | */ |
1852 | if (IS_ACTIVE(CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT)((8000) != 0) && |
1853 | __i915_spin_request(rq, state)) |
1854 | goto out; |
1855 | |
1856 | /* |
1857 | * This client is about to stall waiting for the GPU. In many cases |
1858 | * this is undesirable and limits the throughput of the system, as |
1859 | * many clients cannot continue processing user input/output whilst |
1860 | * blocked. RPS autotuning may take tens of milliseconds to respond |
1861 | * to the GPU load and thus incurs additional latency for the client. |
1862 | * We can circumvent that by promoting the GPU frequency to maximum |
1863 | * before we sleep. This makes the GPU throttle up much more quickly |
1864 | * (good for benchmarks and user experience, e.g. window animations), |
1865 | * but at a cost of spending more power processing the workload |
1866 | * (bad for battery). |
1867 | */ |
1868 | if (flags & I915_WAIT_PRIORITY(1UL << (1)) && !i915_request_started(rq)) |
1869 | intel_rps_boost(rq); |
1870 | |
1871 | #ifdef __linux__ |
1872 | wait.tsk = current; |
1873 | #else |
1874 | wait.tsk = curproc({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc; |
1875 | #endif |
1876 | if (dma_fence_add_callback(&rq->fence, &wait.cb, request_wait_wake)) |
1877 | goto out; |
1878 | |
1879 | /* |
1880 | * Flush the submission tasklet, but only if it may help this request. |
1881 | * |
1882 | * We sometimes experience some latency between the HW interrupts and |
1883 | * tasklet execution (mostly due to ksoftirqd latency, but it can also |
1884 | * be due to lazy CS events), so lets run the tasklet manually if there |
1885 | * is a chance it may submit this request. If the request is not ready |
1886 | * to run, as it is waiting for other fences to be signaled, flushing |
1887 | * the tasklet is busy work without any advantage for this client. |
1888 | * |
1889 | * If the HW is being lazy, this is the last chance before we go to |
1890 | * sleep to catch any pending events. We will check periodically in |
1891 | * the heartbeat to flush the submission tasklets as a last resort |
1892 | * for unhappy HW. |
1893 | */ |
1894 | if (i915_request_is_ready(rq)) |
1895 | intel_engine_flush_submission(rq->engine); |
1896 | |
1897 | for (;;) { |
1898 | set_current_state(state); |
1899 | |
1900 | if (dma_fence_is_signaled(&rq->fence)) |
1901 | break; |
1902 | |
1903 | if (signal_pending_state(state, current)((state) & 0x100 ? (((({struct cpu_info *__ci; asm volatile ("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc)->p_siglist | (({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_curproc)->p_p->ps_siglist) & ~(({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci ) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci; })->ci_curproc)->p_sigmask) : 0)) { |
1904 | timeout = -ERESTARTSYS4; |
1905 | break; |
1906 | } |
1907 | |
1908 | if (!timeout) { |
1909 | timeout = -ETIME60; |
1910 | break; |
1911 | } |
1912 | |
1913 | timeout = io_schedule_timeout(timeout)schedule_timeout(timeout); |
1914 | } |
1915 | __set_current_state(TASK_RUNNING-1); |
1916 | |
1917 | if (READ_ONCE(wait.tsk)({ typeof(wait.tsk) __tmp = *(volatile typeof(wait.tsk) *)& (wait.tsk); membar_datadep_consumer(); __tmp; })) |
1918 | dma_fence_remove_callback(&rq->fence, &wait.cb); |
1919 | GEM_BUG_ON(!list_empty(&wait.cb.node))((void)0); |
1920 | |
1921 | out: |
1922 | mutex_release(&rq->engine->gt->reset.mutex.dep_map, _THIS_IP_); |
1923 | trace_i915_request_wait_end(rq); |
1924 | return timeout; |
1925 | } |
1926 | |
1927 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)0 |
1928 | #include "selftests/mock_request.c" |
1929 | #include "selftests/i915_request.c" |
1930 | #endif |
1931 | |
1932 | static void i915_global_request_shrink(void) |
1933 | { |
1934 | #ifdef notyet |
1935 | kmem_cache_shrink(global.slab_execute_cbs); |
1936 | kmem_cache_shrink(global.slab_requests); |
1937 | #endif |
1938 | } |
1939 | |
1940 | static void i915_global_request_exit(void) |
1941 | { |
1942 | #ifdef __linux__ |
1943 | kmem_cache_destroy(global.slab_execute_cbs); |
1944 | kmem_cache_destroy(global.slab_requests); |
1945 | #else |
1946 | pool_destroy(&global.slab_execute_cbs); |
1947 | pool_destroy(&global.slab_requests); |
1948 | #endif |
1949 | } |
1950 | |
1951 | static struct i915_global_request global = { { |
1952 | .shrink = i915_global_request_shrink, |
1953 | .exit = i915_global_request_exit, |
1954 | } }; |
1955 | |
1956 | int __init i915_global_request_init(void) |
1957 | { |
1958 | #ifdef __linux__ |
1959 | global.slab_requests = |
1960 | kmem_cache_create("i915_request", |
1961 | sizeof(struct i915_request), |
1962 | __alignof__(struct i915_request), |
1963 | SLAB_HWCACHE_ALIGN | |
1964 | SLAB_RECLAIM_ACCOUNT | |
1965 | SLAB_TYPESAFE_BY_RCU, |
1966 | __i915_request_ctor); |
1967 | if (!global.slab_requests) |
1968 | return -ENOMEM12; |
1969 | |
1970 | global.slab_execute_cbs = KMEM_CACHE(execute_cb, |
1971 | SLAB_HWCACHE_ALIGN | |
1972 | SLAB_RECLAIM_ACCOUNT | |
1973 | SLAB_TYPESAFE_BY_RCU); |
1974 | if (!global.slab_execute_cbs) |
1975 | goto err_requests; |
1976 | #else |
1977 | pool_init(&global.slab_requests, sizeof(struct i915_request), |
1978 | CACHELINESIZE64, IPL_TTY0x9, 0, "i915_request", NULL((void *)0)); |
1979 | pool_init(&global.slab_execute_cbs, sizeof(struct execute_cb), |
1980 | CACHELINESIZE64, IPL_TTY0x9, 0, "i915_exec", NULL((void *)0)); |
1981 | #endif |
1982 | |
1983 | i915_global_register(&global.base); |
1984 | return 0; |
1985 | |
1986 | #ifdef __linux__ |
1987 | err_requests: |
1988 | kmem_cache_destroy(global.slab_requests); |
1989 | return -ENOMEM12; |
1990 | #endif |
1991 | } |