Commit | Line | Data |
---|---|---|
64d6c500 CW |
1 | /* |
2 | * SPDX-License-Identifier: MIT | |
3 | * | |
4 | * Copyright © 2019 Intel Corporation | |
5 | */ | |
6 | ||
5361db1a CW |
7 | #include <linux/debugobjects.h> |
8 | ||
e6ba7648 | 9 | #include "gt/intel_context.h" |
d13a3177 | 10 | #include "gt/intel_engine_heartbeat.h" |
7009db14 | 11 | #include "gt/intel_engine_pm.h" |
2871ea85 | 12 | #include "gt/intel_ring.h" |
7009db14 | 13 | |
64d6c500 CW |
14 | #include "i915_drv.h" |
15 | #include "i915_active.h" | |
103b76ee | 16 | #include "i915_globals.h" |
64d6c500 | 17 | |
5f5c139d CW |
18 | /* |
19 | * Active refs memory management | |
20 | * | |
21 | * To be more economical with memory, we reap all the i915_active trees as | |
22 | * they idle (when we know the active requests are inactive) and allocate the | |
23 | * nodes from a local slab cache to hopefully reduce the fragmentation. | |
24 | */ | |
25 | static struct i915_global_active { | |
103b76ee | 26 | struct i915_global base; |
5f5c139d CW |
27 | struct kmem_cache *slab_cache; |
28 | } global; | |
29 | ||
64d6c500 | 30 | struct active_node { |
5d934137 | 31 | struct rb_node node; |
b1e3177b | 32 | struct i915_active_fence base; |
64d6c500 | 33 | struct i915_active *ref; |
64d6c500 CW |
34 | u64 timeline; |
35 | }; | |
36 | ||
5d934137 CW |
37 | #define fetch_node(x) rb_entry(READ_ONCE(x), typeof(struct active_node), node) |
38 | ||
d8af05ff | 39 | static inline struct active_node * |
b1e3177b | 40 | node_from_active(struct i915_active_fence *active) |
d8af05ff CW |
41 | { |
42 | return container_of(active, struct active_node, base); | |
43 | } | |
44 | ||
45 | #define take_preallocated_barriers(x) llist_del_all(&(x)->preallocated_barriers) | |
46 | ||
b1e3177b | 47 | static inline bool is_barrier(const struct i915_active_fence *active) |
d8af05ff | 48 | { |
b1e3177b | 49 | return IS_ERR(rcu_access_pointer(active->fence)); |
d8af05ff CW |
50 | } |
51 | ||
52 | static inline struct llist_node *barrier_to_ll(struct active_node *node) | |
53 | { | |
54 | GEM_BUG_ON(!is_barrier(&node->base)); | |
b1e3177b | 55 | return (struct llist_node *)&node->base.cb.node; |
d8af05ff CW |
56 | } |
57 | ||
f130b712 CW |
58 | static inline struct intel_engine_cs * |
59 | __barrier_to_engine(struct active_node *node) | |
60 | { | |
b1e3177b | 61 | return (struct intel_engine_cs *)READ_ONCE(node->base.cb.node.prev); |
f130b712 CW |
62 | } |
63 | ||
d8af05ff CW |
64 | static inline struct intel_engine_cs * |
65 | barrier_to_engine(struct active_node *node) | |
66 | { | |
67 | GEM_BUG_ON(!is_barrier(&node->base)); | |
f130b712 | 68 | return __barrier_to_engine(node); |
d8af05ff CW |
69 | } |
70 | ||
71 | static inline struct active_node *barrier_from_ll(struct llist_node *x) | |
72 | { | |
73 | return container_of((struct list_head *)x, | |
b1e3177b | 74 | struct active_node, base.cb.node); |
d8af05ff CW |
75 | } |
76 | ||
5361db1a CW |
77 | #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) && IS_ENABLED(CONFIG_DEBUG_OBJECTS) |
78 | ||
79 | static void *active_debug_hint(void *addr) | |
80 | { | |
81 | struct i915_active *ref = addr; | |
82 | ||
12c255b5 | 83 | return (void *)ref->active ?: (void *)ref->retire ?: (void *)ref; |
5361db1a CW |
84 | } |
85 | ||
86 | static struct debug_obj_descr active_debug_desc = { | |
87 | .name = "i915_active", | |
88 | .debug_hint = active_debug_hint, | |
89 | }; | |
90 | ||
91 | static void debug_active_init(struct i915_active *ref) | |
92 | { | |
93 | debug_object_init(ref, &active_debug_desc); | |
94 | } | |
95 | ||
96 | static void debug_active_activate(struct i915_active *ref) | |
97 | { | |
bbca083d | 98 | lockdep_assert_held(&ref->tree_lock); |
f52c6d0d CW |
99 | if (!atomic_read(&ref->count)) /* before the first inc */ |
100 | debug_object_activate(ref, &active_debug_desc); | |
5361db1a CW |
101 | } |
102 | ||
103 | static void debug_active_deactivate(struct i915_active *ref) | |
104 | { | |
c9ad602f | 105 | lockdep_assert_held(&ref->tree_lock); |
f52c6d0d CW |
106 | if (!atomic_read(&ref->count)) /* after the last dec */ |
107 | debug_object_deactivate(ref, &active_debug_desc); | |
5361db1a CW |
108 | } |
109 | ||
110 | static void debug_active_fini(struct i915_active *ref) | |
111 | { | |
112 | debug_object_free(ref, &active_debug_desc); | |
113 | } | |
114 | ||
115 | static void debug_active_assert(struct i915_active *ref) | |
116 | { | |
117 | debug_object_assert_init(ref, &active_debug_desc); | |
118 | } | |
119 | ||
120 | #else | |
121 | ||
122 | static inline void debug_active_init(struct i915_active *ref) { } | |
123 | static inline void debug_active_activate(struct i915_active *ref) { } | |
124 | static inline void debug_active_deactivate(struct i915_active *ref) { } | |
125 | static inline void debug_active_fini(struct i915_active *ref) { } | |
126 | static inline void debug_active_assert(struct i915_active *ref) { } | |
127 | ||
128 | #endif | |
129 | ||
a42375af | 130 | static void |
12c255b5 | 131 | __active_retire(struct i915_active *ref) |
a42375af | 132 | { |
99a7f4da | 133 | struct rb_root root = RB_ROOT; |
a42375af | 134 | struct active_node *it, *n; |
c9ad602f | 135 | unsigned long flags; |
12c255b5 | 136 | |
274cbf20 | 137 | GEM_BUG_ON(i915_active_is_idle(ref)); |
12c255b5 CW |
138 | |
139 | /* return the unused nodes to our slabcache -- flushing the allocator */ | |
c9ad602f | 140 | if (!atomic_dec_and_lock_irqsave(&ref->count, &ref->tree_lock, flags)) |
12c255b5 CW |
141 | return; |
142 | ||
b1e3177b | 143 | GEM_BUG_ON(rcu_access_pointer(ref->excl.fence)); |
c9ad602f CW |
144 | debug_active_deactivate(ref); |
145 | ||
99a7f4da CW |
146 | /* Even if we have not used the cache, we may still have a barrier */ |
147 | if (!ref->cache) | |
148 | ref->cache = fetch_node(ref->tree.rb_node); | |
149 | ||
150 | /* Keep the MRU cached node for reuse */ | |
151 | if (ref->cache) { | |
152 | /* Discard all other nodes in the tree */ | |
153 | rb_erase(&ref->cache->node, &ref->tree); | |
154 | root = ref->tree; | |
155 | ||
156 | /* Rebuild the tree with only the cached node */ | |
157 | rb_link_node(&ref->cache->node, NULL, &ref->tree.rb_node); | |
158 | rb_insert_color(&ref->cache->node, &ref->tree); | |
159 | GEM_BUG_ON(ref->tree.rb_node != &ref->cache->node); | |
160 | } | |
c9ad602f CW |
161 | |
162 | spin_unlock_irqrestore(&ref->tree_lock, flags); | |
e1d7b66b CW |
163 | |
164 | /* After the final retire, the entire struct may be freed */ | |
165 | if (ref->retire) | |
166 | ref->retire(ref); | |
b1e3177b CW |
167 | |
168 | /* ... except if you wait on it, you must manage your own references! */ | |
169 | wake_up_var(ref); | |
c9ad602f | 170 | |
99a7f4da | 171 | /* Finally free the discarded timeline tree */ |
c9ad602f CW |
172 | rbtree_postorder_for_each_entry_safe(it, n, &root, node) { |
173 | GEM_BUG_ON(i915_active_fence_isset(&it->base)); | |
174 | kmem_cache_free(global.slab_cache, it); | |
175 | } | |
a42375af CW |
176 | } |
177 | ||
274cbf20 CW |
178 | static void |
179 | active_work(struct work_struct *wrk) | |
180 | { | |
181 | struct i915_active *ref = container_of(wrk, typeof(*ref), work); | |
182 | ||
183 | GEM_BUG_ON(!atomic_read(&ref->count)); | |
184 | if (atomic_add_unless(&ref->count, -1, 1)) | |
185 | return; | |
186 | ||
274cbf20 CW |
187 | __active_retire(ref); |
188 | } | |
189 | ||
64d6c500 | 190 | static void |
12c255b5 | 191 | active_retire(struct i915_active *ref) |
64d6c500 | 192 | { |
12c255b5 CW |
193 | GEM_BUG_ON(!atomic_read(&ref->count)); |
194 | if (atomic_add_unless(&ref->count, -1, 1)) | |
a42375af CW |
195 | return; |
196 | ||
c9ad602f | 197 | if (ref->flags & I915_ACTIVE_RETIRE_SLEEPS) { |
274cbf20 CW |
198 | queue_work(system_unbound_wq, &ref->work); |
199 | return; | |
200 | } | |
201 | ||
12c255b5 | 202 | __active_retire(ref); |
64d6c500 CW |
203 | } |
204 | ||
df9f85d8 CW |
205 | static inline struct dma_fence ** |
206 | __active_fence_slot(struct i915_active_fence *active) | |
207 | { | |
208 | return (struct dma_fence ** __force)&active->fence; | |
209 | } | |
210 | ||
211 | static inline bool | |
212 | active_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb) | |
213 | { | |
214 | struct i915_active_fence *active = | |
215 | container_of(cb, typeof(*active), cb); | |
216 | ||
217 | return cmpxchg(__active_fence_slot(active), fence, NULL) == fence; | |
218 | } | |
219 | ||
64d6c500 | 220 | static void |
b1e3177b | 221 | node_retire(struct dma_fence *fence, struct dma_fence_cb *cb) |
64d6c500 | 222 | { |
df9f85d8 CW |
223 | if (active_fence_cb(fence, cb)) |
224 | active_retire(container_of(cb, struct active_node, base.cb)->ref); | |
64d6c500 CW |
225 | } |
226 | ||
b1e3177b CW |
227 | static void |
228 | excl_retire(struct dma_fence *fence, struct dma_fence_cb *cb) | |
229 | { | |
df9f85d8 CW |
230 | if (active_fence_cb(fence, cb)) |
231 | active_retire(container_of(cb, struct i915_active, excl.cb)); | |
b1e3177b CW |
232 | } |
233 | ||
5d934137 | 234 | static struct active_node *__active_lookup(struct i915_active *ref, u64 idx) |
64d6c500 | 235 | { |
5d934137 | 236 | struct active_node *it; |
64d6c500 CW |
237 | |
238 | /* | |
239 | * We track the most recently used timeline to skip a rbtree search | |
240 | * for the common case, under typical loads we never need the rbtree | |
241 | * at all. We can reuse the last slot if it is empty, that is | |
242 | * after the previous activity has been retired, or if it matches the | |
243 | * current timeline. | |
64d6c500 | 244 | */ |
5d934137 CW |
245 | it = READ_ONCE(ref->cache); |
246 | if (it && it->timeline == idx) | |
247 | return it; | |
248 | ||
249 | BUILD_BUG_ON(offsetof(typeof(*it), node)); | |
250 | ||
251 | /* While active, the tree can only be built; not destroyed */ | |
252 | GEM_BUG_ON(i915_active_is_idle(ref)); | |
253 | ||
254 | it = fetch_node(ref->tree.rb_node); | |
255 | while (it) { | |
256 | if (it->timeline < idx) { | |
257 | it = fetch_node(it->node.rb_right); | |
258 | } else if (it->timeline > idx) { | |
259 | it = fetch_node(it->node.rb_left); | |
260 | } else { | |
261 | WRITE_ONCE(ref->cache, it); | |
262 | break; | |
263 | } | |
264 | } | |
265 | ||
266 | /* NB: If the tree rotated beneath us, we may miss our target. */ | |
267 | return it; | |
268 | } | |
269 | ||
270 | static struct i915_active_fence * | |
271 | active_instance(struct i915_active *ref, u64 idx) | |
272 | { | |
273 | struct active_node *node, *prealloc; | |
274 | struct rb_node **p, *parent; | |
275 | ||
276 | node = __active_lookup(ref, idx); | |
277 | if (likely(node)) | |
12c255b5 CW |
278 | return &node->base; |
279 | ||
280 | /* Preallocate a replacement, just in case */ | |
281 | prealloc = kmem_cache_alloc(global.slab_cache, GFP_KERNEL); | |
282 | if (!prealloc) | |
283 | return NULL; | |
64d6c500 | 284 | |
c9ad602f | 285 | spin_lock_irq(&ref->tree_lock); |
12c255b5 | 286 | GEM_BUG_ON(i915_active_is_idle(ref)); |
64d6c500 CW |
287 | |
288 | parent = NULL; | |
289 | p = &ref->tree.rb_node; | |
290 | while (*p) { | |
291 | parent = *p; | |
292 | ||
293 | node = rb_entry(parent, struct active_node, node); | |
12c255b5 CW |
294 | if (node->timeline == idx) { |
295 | kmem_cache_free(global.slab_cache, prealloc); | |
296 | goto out; | |
297 | } | |
64d6c500 CW |
298 | |
299 | if (node->timeline < idx) | |
300 | p = &parent->rb_right; | |
301 | else | |
302 | p = &parent->rb_left; | |
303 | } | |
304 | ||
12c255b5 | 305 | node = prealloc; |
df9f85d8 | 306 | __i915_active_fence_init(&node->base, NULL, node_retire); |
64d6c500 CW |
307 | node->ref = ref; |
308 | node->timeline = idx; | |
309 | ||
310 | rb_link_node(&node->node, parent, p); | |
311 | rb_insert_color(&node->node, &ref->tree); | |
312 | ||
64d6c500 | 313 | out: |
5d934137 | 314 | WRITE_ONCE(ref->cache, node); |
c9ad602f | 315 | spin_unlock_irq(&ref->tree_lock); |
12c255b5 CW |
316 | |
317 | return &node->base; | |
64d6c500 CW |
318 | } |
319 | ||
b1e3177b | 320 | void __i915_active_init(struct i915_active *ref, |
12c255b5 CW |
321 | int (*active)(struct i915_active *ref), |
322 | void (*retire)(struct i915_active *ref), | |
ae303004 CW |
323 | struct lock_class_key *mkey, |
324 | struct lock_class_key *wkey) | |
64d6c500 | 325 | { |
274cbf20 CW |
326 | unsigned long bits; |
327 | ||
5361db1a CW |
328 | debug_active_init(ref); |
329 | ||
79c7a28e | 330 | ref->flags = 0; |
12c255b5 | 331 | ref->active = active; |
274cbf20 CW |
332 | ref->retire = ptr_unpack_bits(retire, &bits, 2); |
333 | if (bits & I915_ACTIVE_MAY_SLEEP) | |
334 | ref->flags |= I915_ACTIVE_RETIRE_SLEEPS; | |
2850748e | 335 | |
c9ad602f | 336 | spin_lock_init(&ref->tree_lock); |
64d6c500 | 337 | ref->tree = RB_ROOT; |
12c255b5 | 338 | ref->cache = NULL; |
c9ad602f | 339 | |
d8af05ff | 340 | init_llist_head(&ref->preallocated_barriers); |
12c255b5 | 341 | atomic_set(&ref->count, 0); |
ae303004 | 342 | __mutex_init(&ref->mutex, "i915_active", mkey); |
df9f85d8 | 343 | __i915_active_fence_init(&ref->excl, NULL, excl_retire); |
274cbf20 | 344 | INIT_WORK(&ref->work, active_work); |
ae303004 CW |
345 | #if IS_ENABLED(CONFIG_LOCKDEP) |
346 | lockdep_init_map(&ref->work.lockdep_map, "i915_active.work", wkey, 0); | |
347 | #endif | |
64d6c500 CW |
348 | } |
349 | ||
f130b712 CW |
350 | static bool ____active_del_barrier(struct i915_active *ref, |
351 | struct active_node *node, | |
352 | struct intel_engine_cs *engine) | |
353 | ||
d8af05ff | 354 | { |
d8af05ff CW |
355 | struct llist_node *head = NULL, *tail = NULL; |
356 | struct llist_node *pos, *next; | |
357 | ||
75d0a7f3 | 358 | GEM_BUG_ON(node->timeline != engine->kernel_context->timeline->fence_context); |
d8af05ff CW |
359 | |
360 | /* | |
361 | * Rebuild the llist excluding our node. We may perform this | |
362 | * outside of the kernel_context timeline mutex and so someone | |
363 | * else may be manipulating the engine->barrier_tasks, in | |
364 | * which case either we or they will be upset :) | |
365 | * | |
366 | * A second __active_del_barrier() will report failure to claim | |
367 | * the active_node and the caller will just shrug and know not to | |
368 | * claim ownership of its node. | |
369 | * | |
370 | * A concurrent i915_request_add_active_barriers() will miss adding | |
371 | * any of the tasks, but we will try again on the next -- and since | |
372 | * we are actively using the barrier, we know that there will be | |
373 | * at least another opportunity when we idle. | |
374 | */ | |
375 | llist_for_each_safe(pos, next, llist_del_all(&engine->barrier_tasks)) { | |
376 | if (node == barrier_from_ll(pos)) { | |
377 | node = NULL; | |
378 | continue; | |
379 | } | |
380 | ||
381 | pos->next = head; | |
382 | head = pos; | |
383 | if (!tail) | |
384 | tail = pos; | |
385 | } | |
386 | if (head) | |
387 | llist_add_batch(head, tail, &engine->barrier_tasks); | |
388 | ||
389 | return !node; | |
390 | } | |
391 | ||
f130b712 CW |
392 | static bool |
393 | __active_del_barrier(struct i915_active *ref, struct active_node *node) | |
394 | { | |
395 | return ____active_del_barrier(ref, node, barrier_to_engine(node)); | |
396 | } | |
397 | ||
5d934137 CW |
398 | static bool |
399 | replace_barrier(struct i915_active *ref, struct i915_active_fence *active) | |
400 | { | |
401 | if (!is_barrier(active)) /* proto-node used by our idle barrier? */ | |
402 | return false; | |
403 | ||
404 | /* | |
405 | * This request is on the kernel_context timeline, and so | |
406 | * we can use it to substitute for the pending idle-barrer | |
407 | * request that we want to emit on the kernel_context. | |
408 | */ | |
409 | __active_del_barrier(ref, node_from_active(active)); | |
410 | return true; | |
411 | } | |
412 | ||
413 | int i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence) | |
64d6c500 | 414 | { |
b1e3177b | 415 | struct i915_active_fence *active; |
12c255b5 | 416 | int err; |
312c4ba1 CW |
417 | |
418 | /* Prevent reaping in case we malloc/wait while building the tree */ | |
12c255b5 CW |
419 | err = i915_active_acquire(ref); |
420 | if (err) | |
421 | return err; | |
64d6c500 | 422 | |
5d934137 | 423 | active = active_instance(ref, idx); |
12c255b5 CW |
424 | if (!active) { |
425 | err = -ENOMEM; | |
312c4ba1 CW |
426 | goto out; |
427 | } | |
64d6c500 | 428 | |
5d934137 | 429 | if (replace_barrier(ref, active)) { |
b1e3177b CW |
430 | RCU_INIT_POINTER(active->fence, NULL); |
431 | atomic_dec(&ref->count); | |
d8af05ff | 432 | } |
b1e3177b | 433 | if (!__i915_active_fence_set(active, fence)) |
5d934137 | 434 | __i915_active_acquire(ref); |
64d6c500 | 435 | |
312c4ba1 CW |
436 | out: |
437 | i915_active_release(ref); | |
438 | return err; | |
64d6c500 CW |
439 | } |
440 | ||
5d934137 CW |
441 | static struct dma_fence * |
442 | __i915_active_set_fence(struct i915_active *ref, | |
443 | struct i915_active_fence *active, | |
444 | struct dma_fence *fence) | |
2850748e | 445 | { |
e3793468 CW |
446 | struct dma_fence *prev; |
447 | ||
5d934137 CW |
448 | if (replace_barrier(ref, active)) { |
449 | RCU_INIT_POINTER(active->fence, fence); | |
450 | return NULL; | |
451 | } | |
2850748e | 452 | |
30ca04e1 | 453 | rcu_read_lock(); |
5d934137 | 454 | prev = __i915_active_fence_set(active, fence); |
30ca04e1 CW |
455 | if (prev) |
456 | prev = dma_fence_get_rcu(prev); | |
457 | else | |
5d934137 | 458 | __i915_active_acquire(ref); |
30ca04e1 | 459 | rcu_read_unlock(); |
e3793468 CW |
460 | |
461 | return prev; | |
b1e3177b | 462 | } |
2850748e | 463 | |
5d934137 CW |
464 | static struct i915_active_fence * |
465 | __active_fence(struct i915_active *ref, u64 idx) | |
466 | { | |
467 | struct active_node *it; | |
468 | ||
469 | it = __active_lookup(ref, idx); | |
470 | if (unlikely(!it)) { /* Contention with parallel tree builders! */ | |
471 | spin_lock_irq(&ref->tree_lock); | |
472 | it = __active_lookup(ref, idx); | |
473 | spin_unlock_irq(&ref->tree_lock); | |
474 | } | |
475 | GEM_BUG_ON(!it); /* slot must be preallocated */ | |
476 | ||
477 | return &it->base; | |
478 | } | |
479 | ||
480 | struct dma_fence * | |
481 | __i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence) | |
482 | { | |
483 | /* Only valid while active, see i915_active_acquire_for_context() */ | |
484 | return __i915_active_set_fence(ref, __active_fence(ref, idx), fence); | |
485 | } | |
486 | ||
487 | struct dma_fence * | |
488 | i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f) | |
489 | { | |
490 | /* We expect the caller to manage the exclusive timeline ordering */ | |
491 | return __i915_active_set_fence(ref, &ref->excl, f); | |
492 | } | |
493 | ||
b1e3177b CW |
494 | bool i915_active_acquire_if_busy(struct i915_active *ref) |
495 | { | |
496 | debug_active_assert(ref); | |
497 | return atomic_add_unless(&ref->count, 1, 0); | |
2850748e CW |
498 | } |
499 | ||
04240e30 CW |
500 | static void __i915_active_activate(struct i915_active *ref) |
501 | { | |
502 | spin_lock_irq(&ref->tree_lock); /* __active_retire() */ | |
503 | if (!atomic_fetch_inc(&ref->count)) | |
504 | debug_active_activate(ref); | |
505 | spin_unlock_irq(&ref->tree_lock); | |
506 | } | |
507 | ||
12c255b5 | 508 | int i915_active_acquire(struct i915_active *ref) |
64d6c500 | 509 | { |
12c255b5 CW |
510 | int err; |
511 | ||
b1e3177b | 512 | if (i915_active_acquire_if_busy(ref)) |
12c255b5 | 513 | return 0; |
5361db1a | 514 | |
04240e30 CW |
515 | if (!ref->active) { |
516 | __i915_active_activate(ref); | |
517 | return 0; | |
518 | } | |
519 | ||
12c255b5 CW |
520 | err = mutex_lock_interruptible(&ref->mutex); |
521 | if (err) | |
522 | return err; | |
5361db1a | 523 | |
ac0e331a | 524 | if (likely(!i915_active_acquire_if_busy(ref))) { |
04240e30 CW |
525 | err = ref->active(ref); |
526 | if (!err) | |
527 | __i915_active_activate(ref); | |
12c255b5 CW |
528 | } |
529 | ||
530 | mutex_unlock(&ref->mutex); | |
531 | ||
532 | return err; | |
64d6c500 CW |
533 | } |
534 | ||
5d934137 CW |
535 | int i915_active_acquire_for_context(struct i915_active *ref, u64 idx) |
536 | { | |
537 | struct i915_active_fence *active; | |
538 | int err; | |
539 | ||
540 | err = i915_active_acquire(ref); | |
541 | if (err) | |
542 | return err; | |
543 | ||
544 | active = active_instance(ref, idx); | |
545 | if (!active) { | |
546 | i915_active_release(ref); | |
547 | return -ENOMEM; | |
548 | } | |
549 | ||
550 | return 0; /* return with active ref */ | |
551 | } | |
552 | ||
64d6c500 CW |
553 | void i915_active_release(struct i915_active *ref) |
554 | { | |
5361db1a | 555 | debug_active_assert(ref); |
12c255b5 | 556 | active_retire(ref); |
64d6c500 CW |
557 | } |
558 | ||
b1e3177b | 559 | static void enable_signaling(struct i915_active_fence *active) |
79c7a28e | 560 | { |
b1e3177b | 561 | struct dma_fence *fence; |
79c7a28e | 562 | |
c0e31018 CW |
563 | if (unlikely(is_barrier(active))) |
564 | return; | |
565 | ||
b1e3177b CW |
566 | fence = i915_active_fence_get(active); |
567 | if (!fence) | |
568 | return; | |
2850748e | 569 | |
b1e3177b CW |
570 | dma_fence_enable_sw_signaling(fence); |
571 | dma_fence_put(fence); | |
2850748e CW |
572 | } |
573 | ||
d13a3177 | 574 | static int flush_barrier(struct active_node *it) |
64d6c500 | 575 | { |
d13a3177 | 576 | struct intel_engine_cs *engine; |
64d6c500 | 577 | |
d13a3177 CW |
578 | if (likely(!is_barrier(&it->base))) |
579 | return 0; | |
12c255b5 | 580 | |
d13a3177 CW |
581 | engine = __barrier_to_engine(it); |
582 | smp_rmb(); /* serialise with add_active_barriers */ | |
583 | if (!is_barrier(&it->base)) | |
12c255b5 | 584 | return 0; |
2850748e | 585 | |
d13a3177 CW |
586 | return intel_engine_flush_barriers(engine); |
587 | } | |
588 | ||
589 | static int flush_lazy_signals(struct i915_active *ref) | |
590 | { | |
591 | struct active_node *it, *n; | |
592 | int err = 0; | |
593 | ||
b1e3177b | 594 | enable_signaling(&ref->excl); |
64d6c500 | 595 | rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) { |
d13a3177 CW |
596 | err = flush_barrier(it); /* unconnected idle barrier? */ |
597 | if (err) | |
598 | break; | |
d8af05ff | 599 | |
b1e3177b | 600 | enable_signaling(&it->base); |
64d6c500 CW |
601 | } |
602 | ||
d13a3177 CW |
603 | return err; |
604 | } | |
605 | ||
d75a92a8 | 606 | int __i915_active_wait(struct i915_active *ref, int state) |
d13a3177 CW |
607 | { |
608 | int err; | |
609 | ||
610 | might_sleep(); | |
611 | ||
612 | if (!i915_active_acquire_if_busy(ref)) | |
613 | return 0; | |
614 | ||
615 | /* Any fence added after the wait begins will not be auto-signaled */ | |
616 | err = flush_lazy_signals(ref); | |
b1e3177b | 617 | i915_active_release(ref); |
afd1bcd4 CW |
618 | if (err) |
619 | return err; | |
620 | ||
d75a92a8 CW |
621 | if (!i915_active_is_idle(ref) && |
622 | ___wait_var_event(ref, i915_active_is_idle(ref), | |
623 | state, 0, 0, schedule())) | |
79c7a28e CW |
624 | return -EINTR; |
625 | ||
e1cda6a5 | 626 | flush_work(&ref->work); |
afd1bcd4 | 627 | return 0; |
64d6c500 CW |
628 | } |
629 | ||
29e6ecf3 CW |
630 | static int __await_active(struct i915_active_fence *active, |
631 | int (*fn)(void *arg, struct dma_fence *fence), | |
632 | void *arg) | |
633 | { | |
634 | struct dma_fence *fence; | |
635 | ||
636 | if (is_barrier(active)) /* XXX flush the barrier? */ | |
637 | return 0; | |
638 | ||
639 | fence = i915_active_fence_get(active); | |
640 | if (fence) { | |
641 | int err; | |
642 | ||
643 | err = fn(arg, fence); | |
644 | dma_fence_put(fence); | |
645 | if (err < 0) | |
646 | return err; | |
647 | } | |
648 | ||
649 | return 0; | |
650 | } | |
651 | ||
3b0a0579 CW |
652 | struct wait_barrier { |
653 | struct wait_queue_entry base; | |
654 | struct i915_active *ref; | |
655 | }; | |
656 | ||
657 | static int | |
658 | barrier_wake(wait_queue_entry_t *wq, unsigned int mode, int flags, void *key) | |
659 | { | |
660 | struct wait_barrier *wb = container_of(wq, typeof(*wb), base); | |
661 | ||
662 | if (i915_active_is_idle(wb->ref)) { | |
663 | list_del(&wq->entry); | |
664 | i915_sw_fence_complete(wq->private); | |
665 | kfree(wq); | |
666 | } | |
667 | ||
668 | return 0; | |
669 | } | |
670 | ||
671 | static int __await_barrier(struct i915_active *ref, struct i915_sw_fence *fence) | |
672 | { | |
673 | struct wait_barrier *wb; | |
674 | ||
675 | wb = kmalloc(sizeof(*wb), GFP_KERNEL); | |
676 | if (unlikely(!wb)) | |
677 | return -ENOMEM; | |
678 | ||
679 | GEM_BUG_ON(i915_active_is_idle(ref)); | |
680 | if (!i915_sw_fence_await(fence)) { | |
681 | kfree(wb); | |
682 | return -EINVAL; | |
683 | } | |
684 | ||
685 | wb->base.flags = 0; | |
686 | wb->base.func = barrier_wake; | |
687 | wb->base.private = fence; | |
688 | wb->ref = ref; | |
689 | ||
690 | add_wait_queue(__var_waitqueue(ref), &wb->base); | |
691 | return 0; | |
692 | } | |
693 | ||
29e6ecf3 CW |
694 | static int await_active(struct i915_active *ref, |
695 | unsigned int flags, | |
696 | int (*fn)(void *arg, struct dma_fence *fence), | |
3b0a0579 | 697 | void *arg, struct i915_sw_fence *barrier) |
64d6c500 | 698 | { |
2850748e | 699 | int err = 0; |
64d6c500 | 700 | |
3b0a0579 CW |
701 | if (!i915_active_acquire_if_busy(ref)) |
702 | return 0; | |
703 | ||
442dbc5c CW |
704 | if (flags & I915_ACTIVE_AWAIT_EXCL && |
705 | rcu_access_pointer(ref->excl.fence)) { | |
29e6ecf3 CW |
706 | err = __await_active(&ref->excl, fn, arg); |
707 | if (err) | |
3b0a0579 | 708 | goto out; |
29e6ecf3 CW |
709 | } |
710 | ||
3b0a0579 | 711 | if (flags & I915_ACTIVE_AWAIT_ACTIVE) { |
29e6ecf3 CW |
712 | struct active_node *it, *n; |
713 | ||
714 | rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) { | |
715 | err = __await_active(&it->base, fn, arg); | |
716 | if (err) | |
3b0a0579 | 717 | goto out; |
2850748e | 718 | } |
3b0a0579 CW |
719 | } |
720 | ||
721 | if (flags & I915_ACTIVE_AWAIT_BARRIER) { | |
722 | err = flush_lazy_signals(ref); | |
29e6ecf3 | 723 | if (err) |
3b0a0579 CW |
724 | goto out; |
725 | ||
726 | err = __await_barrier(ref, barrier); | |
727 | if (err) | |
728 | goto out; | |
64d6c500 CW |
729 | } |
730 | ||
3b0a0579 CW |
731 | out: |
732 | i915_active_release(ref); | |
733 | return err; | |
29e6ecf3 CW |
734 | } |
735 | ||
736 | static int rq_await_fence(void *arg, struct dma_fence *fence) | |
737 | { | |
738 | return i915_request_await_dma_fence(arg, fence); | |
739 | } | |
2850748e | 740 | |
29e6ecf3 CW |
741 | int i915_request_await_active(struct i915_request *rq, |
742 | struct i915_active *ref, | |
743 | unsigned int flags) | |
744 | { | |
3b0a0579 | 745 | return await_active(ref, flags, rq_await_fence, rq, &rq->submit); |
29e6ecf3 CW |
746 | } |
747 | ||
748 | static int sw_await_fence(void *arg, struct dma_fence *fence) | |
749 | { | |
750 | return i915_sw_fence_await_dma_fence(arg, fence, 0, | |
751 | GFP_NOWAIT | __GFP_NOWARN); | |
752 | } | |
753 | ||
754 | int i915_sw_fence_await_active(struct i915_sw_fence *fence, | |
755 | struct i915_active *ref, | |
756 | unsigned int flags) | |
757 | { | |
3b0a0579 | 758 | return await_active(ref, flags, sw_await_fence, fence, fence); |
64d6c500 CW |
759 | } |
760 | ||
761 | void i915_active_fini(struct i915_active *ref) | |
762 | { | |
5361db1a | 763 | debug_active_fini(ref); |
12c255b5 | 764 | GEM_BUG_ON(atomic_read(&ref->count)); |
274cbf20 | 765 | GEM_BUG_ON(work_pending(&ref->work)); |
12c255b5 | 766 | mutex_destroy(&ref->mutex); |
99a7f4da CW |
767 | |
768 | if (ref->cache) | |
769 | kmem_cache_free(global.slab_cache, ref->cache); | |
64d6c500 CW |
770 | } |
771 | ||
d8af05ff CW |
772 | static inline bool is_idle_barrier(struct active_node *node, u64 idx) |
773 | { | |
b1e3177b | 774 | return node->timeline == idx && !i915_active_fence_isset(&node->base); |
d8af05ff CW |
775 | } |
776 | ||
777 | static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx) | |
778 | { | |
779 | struct rb_node *prev, *p; | |
780 | ||
781 | if (RB_EMPTY_ROOT(&ref->tree)) | |
782 | return NULL; | |
783 | ||
c9ad602f | 784 | spin_lock_irq(&ref->tree_lock); |
d8af05ff CW |
785 | GEM_BUG_ON(i915_active_is_idle(ref)); |
786 | ||
787 | /* | |
788 | * Try to reuse any existing barrier nodes already allocated for this | |
789 | * i915_active, due to overlapping active phases there is likely a | |
790 | * node kept alive (as we reuse before parking). We prefer to reuse | |
791 | * completely idle barriers (less hassle in manipulating the llists), | |
792 | * but otherwise any will do. | |
793 | */ | |
794 | if (ref->cache && is_idle_barrier(ref->cache, idx)) { | |
795 | p = &ref->cache->node; | |
796 | goto match; | |
797 | } | |
798 | ||
799 | prev = NULL; | |
800 | p = ref->tree.rb_node; | |
801 | while (p) { | |
802 | struct active_node *node = | |
803 | rb_entry(p, struct active_node, node); | |
804 | ||
805 | if (is_idle_barrier(node, idx)) | |
806 | goto match; | |
807 | ||
808 | prev = p; | |
809 | if (node->timeline < idx) | |
810 | p = p->rb_right; | |
811 | else | |
812 | p = p->rb_left; | |
813 | } | |
814 | ||
815 | /* | |
816 | * No quick match, but we did find the leftmost rb_node for the | |
817 | * kernel_context. Walk the rb_tree in-order to see if there were | |
818 | * any idle-barriers on this timeline that we missed, or just use | |
819 | * the first pending barrier. | |
820 | */ | |
821 | for (p = prev; p; p = rb_next(p)) { | |
822 | struct active_node *node = | |
823 | rb_entry(p, struct active_node, node); | |
f130b712 | 824 | struct intel_engine_cs *engine; |
d8af05ff CW |
825 | |
826 | if (node->timeline > idx) | |
827 | break; | |
828 | ||
829 | if (node->timeline < idx) | |
830 | continue; | |
831 | ||
832 | if (is_idle_barrier(node, idx)) | |
833 | goto match; | |
834 | ||
835 | /* | |
836 | * The list of pending barriers is protected by the | |
837 | * kernel_context timeline, which notably we do not hold | |
838 | * here. i915_request_add_active_barriers() may consume | |
839 | * the barrier before we claim it, so we have to check | |
840 | * for success. | |
841 | */ | |
f130b712 CW |
842 | engine = __barrier_to_engine(node); |
843 | smp_rmb(); /* serialise with add_active_barriers */ | |
844 | if (is_barrier(&node->base) && | |
845 | ____active_del_barrier(ref, node, engine)) | |
d8af05ff CW |
846 | goto match; |
847 | } | |
848 | ||
c9ad602f | 849 | spin_unlock_irq(&ref->tree_lock); |
d8af05ff CW |
850 | |
851 | return NULL; | |
852 | ||
853 | match: | |
854 | rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */ | |
855 | if (p == &ref->cache->node) | |
5d934137 | 856 | WRITE_ONCE(ref->cache, NULL); |
c9ad602f | 857 | spin_unlock_irq(&ref->tree_lock); |
d8af05ff CW |
858 | |
859 | return rb_entry(p, struct active_node, node); | |
860 | } | |
861 | ||
ce476c80 CW |
862 | int i915_active_acquire_preallocate_barrier(struct i915_active *ref, |
863 | struct intel_engine_cs *engine) | |
864 | { | |
3f99a614 | 865 | intel_engine_mask_t tmp, mask = engine->mask; |
d4c3c0b8 | 866 | struct llist_node *first = NULL, *last = NULL; |
a50134b1 | 867 | struct intel_gt *gt = engine->gt; |
ce476c80 | 868 | |
b5e8e954 | 869 | GEM_BUG_ON(i915_active_is_idle(ref)); |
84135022 CW |
870 | |
871 | /* Wait until the previous preallocation is completed */ | |
872 | while (!llist_empty(&ref->preallocated_barriers)) | |
873 | cond_resched(); | |
d8af05ff CW |
874 | |
875 | /* | |
876 | * Preallocate a node for each physical engine supporting the target | |
877 | * engine (remember virtual engines have more than one sibling). | |
878 | * We can then use the preallocated nodes in | |
879 | * i915_active_acquire_barrier() | |
880 | */ | |
416d3838 | 881 | GEM_BUG_ON(!mask); |
a50134b1 | 882 | for_each_engine_masked(engine, gt, mask, tmp) { |
75d0a7f3 | 883 | u64 idx = engine->kernel_context->timeline->fence_context; |
d4c3c0b8 | 884 | struct llist_node *prev = first; |
ce476c80 CW |
885 | struct active_node *node; |
886 | ||
d8af05ff CW |
887 | node = reuse_idle_barrier(ref, idx); |
888 | if (!node) { | |
889 | node = kmem_cache_alloc(global.slab_cache, GFP_KERNEL); | |
e714977e | 890 | if (!node) |
d8af05ff | 891 | goto unwind; |
d8af05ff | 892 | |
b1e3177b CW |
893 | RCU_INIT_POINTER(node->base.fence, NULL); |
894 | node->base.cb.func = node_retire; | |
d8af05ff CW |
895 | node->timeline = idx; |
896 | node->ref = ref; | |
ce476c80 CW |
897 | } |
898 | ||
b1e3177b | 899 | if (!i915_active_fence_isset(&node->base)) { |
d8af05ff CW |
900 | /* |
901 | * Mark this as being *our* unconnected proto-node. | |
902 | * | |
903 | * Since this node is not in any list, and we have | |
904 | * decoupled it from the rbtree, we can reuse the | |
905 | * request to indicate this is an idle-barrier node | |
906 | * and then we can use the rb_node and list pointers | |
907 | * for our tracking of the pending barrier. | |
908 | */ | |
b1e3177b CW |
909 | RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN)); |
910 | node->base.cb.node.prev = (void *)engine; | |
5d934137 | 911 | __i915_active_acquire(ref); |
d8af05ff | 912 | } |
df9f85d8 | 913 | GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN)); |
ce476c80 | 914 | |
d8af05ff | 915 | GEM_BUG_ON(barrier_to_engine(node) != engine); |
d4c3c0b8 JRS |
916 | first = barrier_to_ll(node); |
917 | first->next = prev; | |
918 | if (!last) | |
919 | last = first; | |
7009db14 | 920 | intel_engine_pm_get(engine); |
ce476c80 CW |
921 | } |
922 | ||
84135022 | 923 | GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers)); |
d4c3c0b8 | 924 | llist_add_batch(first, last, &ref->preallocated_barriers); |
84135022 | 925 | |
7009db14 CW |
926 | return 0; |
927 | ||
928 | unwind: | |
d4c3c0b8 JRS |
929 | while (first) { |
930 | struct active_node *node = barrier_from_ll(first); | |
7009db14 | 931 | |
d4c3c0b8 | 932 | first = first->next; |
84135022 | 933 | |
d8af05ff CW |
934 | atomic_dec(&ref->count); |
935 | intel_engine_pm_put(barrier_to_engine(node)); | |
7009db14 | 936 | |
7009db14 CW |
937 | kmem_cache_free(global.slab_cache, node); |
938 | } | |
e714977e | 939 | return -ENOMEM; |
ce476c80 CW |
940 | } |
941 | ||
942 | void i915_active_acquire_barrier(struct i915_active *ref) | |
943 | { | |
944 | struct llist_node *pos, *next; | |
c9ad602f | 945 | unsigned long flags; |
ce476c80 | 946 | |
12c255b5 | 947 | GEM_BUG_ON(i915_active_is_idle(ref)); |
ce476c80 | 948 | |
d8af05ff CW |
949 | /* |
950 | * Transfer the list of preallocated barriers into the | |
951 | * i915_active rbtree, but only as proto-nodes. They will be | |
952 | * populated by i915_request_add_active_barriers() to point to the | |
953 | * request that will eventually release them. | |
954 | */ | |
d8af05ff CW |
955 | llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) { |
956 | struct active_node *node = barrier_from_ll(pos); | |
957 | struct intel_engine_cs *engine = barrier_to_engine(node); | |
ce476c80 CW |
958 | struct rb_node **p, *parent; |
959 | ||
07779a76 CW |
960 | spin_lock_irqsave_nested(&ref->tree_lock, flags, |
961 | SINGLE_DEPTH_NESTING); | |
ce476c80 CW |
962 | parent = NULL; |
963 | p = &ref->tree.rb_node; | |
964 | while (*p) { | |
d8af05ff CW |
965 | struct active_node *it; |
966 | ||
ce476c80 | 967 | parent = *p; |
d8af05ff CW |
968 | |
969 | it = rb_entry(parent, struct active_node, node); | |
970 | if (it->timeline < node->timeline) | |
ce476c80 CW |
971 | p = &parent->rb_right; |
972 | else | |
973 | p = &parent->rb_left; | |
974 | } | |
975 | rb_link_node(&node->node, parent, p); | |
976 | rb_insert_color(&node->node, &ref->tree); | |
07779a76 | 977 | spin_unlock_irqrestore(&ref->tree_lock, flags); |
ce476c80 | 978 | |
b7234840 | 979 | GEM_BUG_ON(!intel_engine_pm_is_awake(engine)); |
d8af05ff | 980 | llist_add(barrier_to_ll(node), &engine->barrier_tasks); |
edee52c9 | 981 | intel_engine_pm_put_delay(engine, 1); |
ce476c80 | 982 | } |
ce476c80 CW |
983 | } |
984 | ||
df9f85d8 CW |
985 | static struct dma_fence **ll_to_fence_slot(struct llist_node *node) |
986 | { | |
987 | return __active_fence_slot(&barrier_from_ll(node)->base); | |
988 | } | |
989 | ||
d8af05ff | 990 | void i915_request_add_active_barriers(struct i915_request *rq) |
ce476c80 CW |
991 | { |
992 | struct intel_engine_cs *engine = rq->engine; | |
993 | struct llist_node *node, *next; | |
b1e3177b | 994 | unsigned long flags; |
ce476c80 | 995 | |
e6ba7648 | 996 | GEM_BUG_ON(!intel_context_is_barrier(rq->context)); |
d8af05ff | 997 | GEM_BUG_ON(intel_engine_is_virtual(engine)); |
d19d71fc | 998 | GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline); |
d8af05ff | 999 | |
b1e3177b CW |
1000 | node = llist_del_all(&engine->barrier_tasks); |
1001 | if (!node) | |
1002 | return; | |
d8af05ff CW |
1003 | /* |
1004 | * Attach the list of proto-fences to the in-flight request such | |
1005 | * that the parent i915_active will be released when this request | |
1006 | * is retired. | |
1007 | */ | |
b1e3177b CW |
1008 | spin_lock_irqsave(&rq->lock, flags); |
1009 | llist_for_each_safe(node, next, node) { | |
df9f85d8 CW |
1010 | /* serialise with reuse_idle_barrier */ |
1011 | smp_store_mb(*ll_to_fence_slot(node), &rq->fence); | |
b1e3177b CW |
1012 | list_add_tail((struct list_head *)node, &rq->fence.cb_list); |
1013 | } | |
1014 | spin_unlock_irqrestore(&rq->lock, flags); | |
1015 | } | |
1016 | ||
b1e3177b CW |
1017 | /* |
1018 | * __i915_active_fence_set: Update the last active fence along its timeline | |
1019 | * @active: the active tracker | |
1020 | * @fence: the new fence (under construction) | |
1021 | * | |
1022 | * Records the new @fence as the last active fence along its timeline in | |
1023 | * this active tracker, moving the tracking callbacks from the previous | |
1024 | * fence onto this one. Returns the previous fence (if not already completed), | |
1025 | * which the caller must ensure is executed before the new fence. To ensure | |
1026 | * that the order of fences within the timeline of the i915_active_fence is | |
df9f85d8 | 1027 | * understood, it should be locked by the caller. |
b1e3177b CW |
1028 | */ |
1029 | struct dma_fence * | |
1030 | __i915_active_fence_set(struct i915_active_fence *active, | |
1031 | struct dma_fence *fence) | |
1032 | { | |
1033 | struct dma_fence *prev; | |
1034 | unsigned long flags; | |
1035 | ||
df9f85d8 CW |
1036 | if (fence == rcu_access_pointer(active->fence)) |
1037 | return fence; | |
1038 | ||
b1e3177b CW |
1039 | GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)); |
1040 | ||
df9f85d8 CW |
1041 | /* |
1042 | * Consider that we have two threads arriving (A and B), with | |
1043 | * C already resident as the active->fence. | |
1044 | * | |
1045 | * A does the xchg first, and so it sees C or NULL depending | |
1046 | * on the timing of the interrupt handler. If it is NULL, the | |
1047 | * previous fence must have been signaled and we know that | |
1048 | * we are first on the timeline. If it is still present, | |
1049 | * we acquire the lock on that fence and serialise with the interrupt | |
1050 | * handler, in the process removing it from any future interrupt | |
1051 | * callback. A will then wait on C before executing (if present). | |
1052 | * | |
1053 | * As B is second, it sees A as the previous fence and so waits for | |
1054 | * it to complete its transition and takes over the occupancy for | |
1055 | * itself -- remembering that it needs to wait on A before executing. | |
1056 | * | |
1057 | * Note the strong ordering of the timeline also provides consistent | |
1058 | * nesting rules for the fence->lock; the inner lock is always the | |
1059 | * older lock. | |
1060 | */ | |
1061 | spin_lock_irqsave(fence->lock, flags); | |
1062 | prev = xchg(__active_fence_slot(active), fence); | |
b1e3177b CW |
1063 | if (prev) { |
1064 | GEM_BUG_ON(prev == fence); | |
1065 | spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING); | |
1066 | __list_del_entry(&active->cb.node); | |
1067 | spin_unlock(prev->lock); /* serialise with prev->cb_list */ | |
d8af05ff | 1068 | } |
b1e3177b | 1069 | list_add_tail(&active->cb.node, &fence->cb_list); |
b1e3177b CW |
1070 | spin_unlock_irqrestore(fence->lock, flags); |
1071 | ||
1072 | return prev; | |
ce476c80 CW |
1073 | } |
1074 | ||
b1e3177b CW |
1075 | int i915_active_fence_set(struct i915_active_fence *active, |
1076 | struct i915_request *rq) | |
21950ee7 | 1077 | { |
b1e3177b CW |
1078 | struct dma_fence *fence; |
1079 | int err = 0; | |
21950ee7 | 1080 | |
b1e3177b CW |
1081 | /* Must maintain timeline ordering wrt previous active requests */ |
1082 | rcu_read_lock(); | |
1083 | fence = __i915_active_fence_set(active, &rq->fence); | |
1084 | if (fence) /* but the previous fence may not belong to that timeline! */ | |
1085 | fence = dma_fence_get_rcu(fence); | |
1086 | rcu_read_unlock(); | |
1087 | if (fence) { | |
1088 | err = i915_request_await_dma_fence(rq, fence); | |
1089 | dma_fence_put(fence); | |
1090 | } | |
21950ee7 | 1091 | |
b1e3177b | 1092 | return err; |
21950ee7 CW |
1093 | } |
1094 | ||
b1e3177b | 1095 | void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb) |
21950ee7 | 1096 | { |
df9f85d8 | 1097 | active_fence_cb(fence, cb); |
21950ee7 CW |
1098 | } |
1099 | ||
229007e0 CW |
1100 | struct auto_active { |
1101 | struct i915_active base; | |
1102 | struct kref ref; | |
1103 | }; | |
1104 | ||
1105 | struct i915_active *i915_active_get(struct i915_active *ref) | |
1106 | { | |
1107 | struct auto_active *aa = container_of(ref, typeof(*aa), base); | |
1108 | ||
1109 | kref_get(&aa->ref); | |
1110 | return &aa->base; | |
1111 | } | |
1112 | ||
1113 | static void auto_release(struct kref *ref) | |
1114 | { | |
1115 | struct auto_active *aa = container_of(ref, typeof(*aa), ref); | |
1116 | ||
1117 | i915_active_fini(&aa->base); | |
1118 | kfree(aa); | |
1119 | } | |
1120 | ||
1121 | void i915_active_put(struct i915_active *ref) | |
1122 | { | |
1123 | struct auto_active *aa = container_of(ref, typeof(*aa), base); | |
1124 | ||
1125 | kref_put(&aa->ref, auto_release); | |
1126 | } | |
1127 | ||
1128 | static int auto_active(struct i915_active *ref) | |
1129 | { | |
1130 | i915_active_get(ref); | |
1131 | return 0; | |
1132 | } | |
1133 | ||
1134 | static void auto_retire(struct i915_active *ref) | |
1135 | { | |
1136 | i915_active_put(ref); | |
1137 | } | |
1138 | ||
1139 | struct i915_active *i915_active_create(void) | |
1140 | { | |
1141 | struct auto_active *aa; | |
1142 | ||
1143 | aa = kmalloc(sizeof(*aa), GFP_KERNEL); | |
1144 | if (!aa) | |
1145 | return NULL; | |
1146 | ||
1147 | kref_init(&aa->ref); | |
1148 | i915_active_init(&aa->base, auto_active, auto_retire); | |
1149 | ||
1150 | return &aa->base; | |
1151 | } | |
1152 | ||
64d6c500 CW |
1153 | #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST) |
1154 | #include "selftests/i915_active.c" | |
1155 | #endif | |
5f5c139d | 1156 | |
103b76ee | 1157 | static void i915_global_active_shrink(void) |
5f5c139d | 1158 | { |
103b76ee | 1159 | kmem_cache_shrink(global.slab_cache); |
5f5c139d CW |
1160 | } |
1161 | ||
103b76ee | 1162 | static void i915_global_active_exit(void) |
32eb6bcf | 1163 | { |
103b76ee | 1164 | kmem_cache_destroy(global.slab_cache); |
32eb6bcf CW |
1165 | } |
1166 | ||
103b76ee CW |
1167 | static struct i915_global_active global = { { |
1168 | .shrink = i915_global_active_shrink, | |
1169 | .exit = i915_global_active_exit, | |
1170 | } }; | |
1171 | ||
1172 | int __init i915_global_active_init(void) | |
5f5c139d | 1173 | { |
103b76ee CW |
1174 | global.slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN); |
1175 | if (!global.slab_cache) | |
1176 | return -ENOMEM; | |
1177 | ||
1178 | i915_global_register(&global.base); | |
1179 | return 0; | |
5f5c139d | 1180 | } |