drivers/gpu/drm/i915/i915_active.c

   1 /*
   2  * SPDX-License-Identifier: MIT
   3  *
   4  * Copyright © 2019 Intel Corporation
   5  */
   6
   7 #include <linux/debugobjects.h>
   8
   9 #include "gt/intel_context.h"
  10 #include "gt/intel_engine_heartbeat.h"
  11 #include "gt/intel_engine_pm.h"
  12 #include "gt/intel_ring.h"
  13
  14 #include "i915_drv.h"
  15 #include "i915_active.h"
  16
  17 /*
  18  * Active refs memory management
  19  *
  20  * To be more economical with memory, we reap all the i915_active trees as
  21  * they idle (when we know the active requests are inactive) and allocate the
  22  * nodes from a local slab cache to hopefully reduce the fragmentation.
  23  */
  24 static struct kmem_cache *slab_cache;
  25
  26 struct active_node {
  27         struct rb_node node;
  28         struct i915_active_fence base;
  29         struct i915_active *ref;
  30         u64 timeline;
  31 };
  32
  33 #define fetch_node(x) rb_entry(READ_ONCE(x), typeof(struct active_node), node)
  34
  35 static inline struct active_node *
  36 node_from_active(struct i915_active_fence *active)
  37 {
  38         return container_of(active, struct active_node, base);
  39 }
  40
  41 #define take_preallocated_barriers(x) llist_del_all(&(x)->preallocated_barriers)
  42
  43 static inline bool is_barrier(const struct i915_active_fence *active)
  44 {
  45         return IS_ERR(rcu_access_pointer(active->fence));
  46 }
  47
  48 static inline struct llist_node *barrier_to_ll(struct active_node *node)
  49 {
  50         GEM_BUG_ON(!is_barrier(&node->base));
  51         return (struct llist_node *)&node->base.cb.node;
  52 }
  53
  54 static inline struct intel_engine_cs *
  55 __barrier_to_engine(struct active_node *node)
  56 {
  57         return (struct intel_engine_cs *)READ_ONCE(node->base.cb.node.prev);
  58 }
  59
  60 static inline struct intel_engine_cs *
  61 barrier_to_engine(struct active_node *node)
  62 {
  63         GEM_BUG_ON(!is_barrier(&node->base));
  64         return __barrier_to_engine(node);
  65 }
  66
  67 static inline struct active_node *barrier_from_ll(struct llist_node *x)
  68 {
  69         return container_of((struct list_head *)x,
  70                             struct active_node, base.cb.node);
  71 }
  72
  73 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) && IS_ENABLED(CONFIG_DEBUG_OBJECTS)
  74
  75 static void *active_debug_hint(void *addr)
  76 {
  77         struct i915_active *ref = addr;
  78
  79         return (void *)ref->active ?: (void *)ref->retire ?: (void *)ref;
  80 }
  81
  82 static const struct debug_obj_descr active_debug_desc = {
  83         .name = "i915_active",
  84         .debug_hint = active_debug_hint,
  85 };
  86
  87 static void debug_active_init(struct i915_active *ref)
  88 {
  89         debug_object_init(ref, &active_debug_desc);
  90 }
  91
  92 static void debug_active_activate(struct i915_active *ref)
  93 {
  94         lockdep_assert_held(&ref->tree_lock);
  95         debug_object_activate(ref, &active_debug_desc);
  96 }
  97
  98 static void debug_active_deactivate(struct i915_active *ref)
  99 {
 100         lockdep_assert_held(&ref->tree_lock);
 101         if (!atomic_read(&ref->count)) /* after the last dec */
 102                 debug_object_deactivate(ref, &active_debug_desc);
 103 }
 104
 105 static void debug_active_fini(struct i915_active *ref)
 106 {
 107         debug_object_free(ref, &active_debug_desc);
 108 }
 109
 110 static void debug_active_assert(struct i915_active *ref)
 111 {
 112         debug_object_assert_init(ref, &active_debug_desc);
 113 }
 114
 115 #else
 116
 117 static inline void debug_active_init(struct i915_active *ref) { }
 118 static inline void debug_active_activate(struct i915_active *ref) { }
 119 static inline void debug_active_deactivate(struct i915_active *ref) { }
 120 static inline void debug_active_fini(struct i915_active *ref) { }
 121 static inline void debug_active_assert(struct i915_active *ref) { }
 122
 123 #endif
 124
 125 static void
 126 __active_retire(struct i915_active *ref)
 127 {
 128         struct rb_root root = RB_ROOT;
 129         struct active_node *it, *n;
 130         unsigned long flags;
 131
 132         GEM_BUG_ON(i915_active_is_idle(ref));
 133
 134         /* return the unused nodes to our slabcache -- flushing the allocator */
 135         if (!atomic_dec_and_lock_irqsave(&ref->count, &ref->tree_lock, flags))
 136                 return;
 137
 138         GEM_BUG_ON(rcu_access_pointer(ref->excl.fence));
 139         debug_active_deactivate(ref);
 140
 141         /* Even if we have not used the cache, we may still have a barrier */
 142         if (!ref->cache)
 143                 ref->cache = fetch_node(ref->tree.rb_node);
 144
 145         /* Keep the MRU cached node for reuse */
 146         if (ref->cache) {
 147                 /* Discard all other nodes in the tree */
 148                 rb_erase(&ref->cache->node, &ref->tree);
 149                 root = ref->tree;
 150
 151                 /* Rebuild the tree with only the cached node */
 152                 rb_link_node(&ref->cache->node, NULL, &ref->tree.rb_node);
 153                 rb_insert_color(&ref->cache->node, &ref->tree);
 154                 GEM_BUG_ON(ref->tree.rb_node != &ref->cache->node);
 155
 156                 /* Make the cached node available for reuse with any timeline */
 157                 ref->cache->timeline = 0; /* needs cmpxchg(u64) */
 158         }
 159
 160         spin_unlock_irqrestore(&ref->tree_lock, flags);
 161
 162         /* After the final retire, the entire struct may be freed */
 163         if (ref->retire)
 164                 ref->retire(ref);
 165
 166         /* ... except if you wait on it, you must manage your own references! */
 167         wake_up_var(ref);
 168
 169         /* Finally free the discarded timeline tree  */
 170         rbtree_postorder_for_each_entry_safe(it, n, &root, node) {
 171                 GEM_BUG_ON(i915_active_fence_isset(&it->base));
 172                 kmem_cache_free(slab_cache, it);
 173         }
 174 }
 175
 176 static void
 177 active_work(struct work_struct *wrk)
 178 {
 179         struct i915_active *ref = container_of(wrk, typeof(*ref), work);
 180
 181         GEM_BUG_ON(!atomic_read(&ref->count));
 182         if (atomic_add_unless(&ref->count, -1, 1))
 183                 return;
 184
 185         __active_retire(ref);
 186 }
 187
 188 static void
 189 active_retire(struct i915_active *ref)
 190 {
 191         GEM_BUG_ON(!atomic_read(&ref->count));
 192         if (atomic_add_unless(&ref->count, -1, 1))
 193                 return;
 194
 195         if (ref->flags & I915_ACTIVE_RETIRE_SLEEPS) {
 196                 queue_work(system_unbound_wq, &ref->work);
 197                 return;
 198         }
 199
 200         __active_retire(ref);
 201 }
 202
 203 static inline struct dma_fence **
 204 __active_fence_slot(struct i915_active_fence *active)
 205 {
 206         return (struct dma_fence ** __force)&active->fence;
 207 }
 208
 209 static inline bool
 210 active_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
 211 {
 212         struct i915_active_fence *active =
 213                 container_of(cb, typeof(*active), cb);
 214
 215         return cmpxchg(__active_fence_slot(active), fence, NULL) == fence;
 216 }
 217
 218 static void
 219 node_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
 220 {
 221         if (active_fence_cb(fence, cb))
 222                 active_retire(container_of(cb, struct active_node, base.cb)->ref);
 223 }
 224
 225 static void
 226 excl_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
 227 {
 228         if (active_fence_cb(fence, cb))
 229                 active_retire(container_of(cb, struct i915_active, excl.cb));
 230 }
 231
 232 static struct active_node *__active_lookup(struct i915_active *ref, u64 idx)
 233 {
 234         struct active_node *it;
 235
 236         GEM_BUG_ON(idx == 0); /* 0 is the unordered timeline, rsvd for cache */
 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.
 244          */
 245         it = READ_ONCE(ref->cache);
 246         if (it) {
 247                 u64 cached = READ_ONCE(it->timeline);
 248
 249                 /* Once claimed, this slot will only belong to this idx */
 250                 if (cached == idx)
 251                         return it;
 252
 253                 /*
 254                  * An unclaimed cache [.timeline=0] can only be claimed once.
 255                  *
 256                  * If the value is already non-zero, some other thread has
 257                  * claimed the cache and we know that is does not match our
 258                  * idx. If, and only if, the timeline is currently zero is it
 259                  * worth competing to claim it atomically for ourselves (for
 260                  * only the winner of that race will cmpxchg return the old
 261                  * value of 0).
 262                  */
 263                 if (!cached && !cmpxchg64(&it->timeline, 0, idx))
 264                         return it;
 265         }
 266
 267         BUILD_BUG_ON(offsetof(typeof(*it), node));
 268
 269         /* While active, the tree can only be built; not destroyed */
 270         GEM_BUG_ON(i915_active_is_idle(ref));
 271
 272         it = fetch_node(ref->tree.rb_node);
 273         while (it) {
 274                 if (it->timeline < idx) {
 275                         it = fetch_node(it->node.rb_right);
 276                 } else if (it->timeline > idx) {
 277                         it = fetch_node(it->node.rb_left);
 278                 } else {
 279                         WRITE_ONCE(ref->cache, it);
 280                         break;
 281                 }
 282         }
 283
 284         /* NB: If the tree rotated beneath us, we may miss our target. */
 285         return it;
 286 }
 287
 288 static struct i915_active_fence *
 289 active_instance(struct i915_active *ref, u64 idx)
 290 {
 291         struct active_node *node;
 292         struct rb_node **p, *parent;
 293
 294         node = __active_lookup(ref, idx);
 295         if (likely(node))
 296                 return &node->base;
 297
 298         spin_lock_irq(&ref->tree_lock);
 299         GEM_BUG_ON(i915_active_is_idle(ref));
 300
 301         parent = NULL;
 302         p = &ref->tree.rb_node;
 303         while (*p) {
 304                 parent = *p;
 305
 306                 node = rb_entry(parent, struct active_node, node);
 307                 if (node->timeline == idx)
 308                         goto out;
 309
 310                 if (node->timeline < idx)
 311                         p = &parent->rb_right;
 312                 else
 313                         p = &parent->rb_left;
 314         }
 315
 316         /*
 317          * XXX: We should preallocate this before i915_active_ref() is ever
 318          *  called, but we cannot call into fs_reclaim() anyway, so use GFP_ATOMIC.
 319          */
 320         node = kmem_cache_alloc(slab_cache, GFP_ATOMIC);
 321         if (!node)
 322                 goto out;
 323
 324         __i915_active_fence_init(&node->base, NULL, node_retire);
 325         node->ref = ref;
 326         node->timeline = idx;
 327
 328         rb_link_node(&node->node, parent, p);
 329         rb_insert_color(&node->node, &ref->tree);
 330
 331 out:
 332         WRITE_ONCE(ref->cache, node);
 333         spin_unlock_irq(&ref->tree_lock);
 334
 335         return &node->base;
 336 }
 337
 338 void __i915_active_init(struct i915_active *ref,
 339                         int (*active)(struct i915_active *ref),
 340                         void (*retire)(struct i915_active *ref),
 341                         unsigned long flags,
 342                         struct lock_class_key *mkey,
 343                         struct lock_class_key *wkey)
 344 {
 345         debug_active_init(ref);
 346
 347         ref->flags = flags;
 348         ref->active = active;
 349         ref->retire = retire;
 350
 351         spin_lock_init(&ref->tree_lock);
 352         ref->tree = RB_ROOT;
 353         ref->cache = NULL;
 354
 355         init_llist_head(&ref->preallocated_barriers);
 356         atomic_set(&ref->count, 0);
 357         __mutex_init(&ref->mutex, "i915_active", mkey);
 358         __i915_active_fence_init(&ref->excl, NULL, excl_retire);
 359         INIT_WORK(&ref->work, active_work);
 360 #if IS_ENABLED(CONFIG_LOCKDEP)
 361         lockdep_init_map(&ref->work.lockdep_map, "i915_active.work", wkey, 0);
 362 #endif
 363 }
 364
 365 static bool ____active_del_barrier(struct i915_active *ref,
 366                                    struct active_node *node,
 367                                    struct intel_engine_cs *engine)
 368
 369 {
 370         struct llist_node *head = NULL, *tail = NULL;
 371         struct llist_node *pos, *next;
 372
 373         GEM_BUG_ON(node->timeline != engine->kernel_context->timeline->fence_context);
 374
 375         /*
 376          * Rebuild the llist excluding our node. We may perform this
 377          * outside of the kernel_context timeline mutex and so someone
 378          * else may be manipulating the engine->barrier_tasks, in
 379          * which case either we or they will be upset :)
 380          *
 381          * A second __active_del_barrier() will report failure to claim
 382          * the active_node and the caller will just shrug and know not to
 383          * claim ownership of its node.
 384          *
 385          * A concurrent i915_request_add_active_barriers() will miss adding
 386          * any of the tasks, but we will try again on the next -- and since
 387          * we are actively using the barrier, we know that there will be
 388          * at least another opportunity when we idle.
 389          */
 390         llist_for_each_safe(pos, next, llist_del_all(&engine->barrier_tasks)) {
 391                 if (node == barrier_from_ll(pos)) {
 392                         node = NULL;
 393                         continue;
 394                 }
 395
 396                 pos->next = head;
 397                 head = pos;
 398                 if (!tail)
 399                         tail = pos;
 400         }
 401         if (head)
 402                 llist_add_batch(head, tail, &engine->barrier_tasks);
 403
 404         return !node;
 405 }
 406
 407 static bool
 408 __active_del_barrier(struct i915_active *ref, struct active_node *node)
 409 {
 410         return ____active_del_barrier(ref, node, barrier_to_engine(node));
 411 }
 412
 413 static bool
 414 replace_barrier(struct i915_active *ref, struct i915_active_fence *active)
 415 {
 416         if (!is_barrier(active)) /* proto-node used by our idle barrier? */
 417                 return false;
 418
 419         /*
 420          * This request is on the kernel_context timeline, and so
 421          * we can use it to substitute for the pending idle-barrer
 422          * request that we want to emit on the kernel_context.
 423          */
 424         return __active_del_barrier(ref, node_from_active(active));
 425 }
 426
 427 int i915_active_add_request(struct i915_active *ref, struct i915_request *rq)
 428 {
 429         u64 idx = i915_request_timeline(rq)->fence_context;
 430         struct dma_fence *fence = &rq->fence;
 431         struct i915_active_fence *active;
 432         int err;
 433
 434         /* Prevent reaping in case we malloc/wait while building the tree */
 435         err = i915_active_acquire(ref);
 436         if (err)
 437                 return err;
 438
 439         do {
 440                 active = active_instance(ref, idx);
 441                 if (!active) {
 442                         err = -ENOMEM;
 443                         goto out;
 444                 }
 445
 446                 if (replace_barrier(ref, active)) {
 447                         RCU_INIT_POINTER(active->fence, NULL);
 448                         atomic_dec(&ref->count);
 449                 }
 450         } while (unlikely(is_barrier(active)));
 451
 452         if (!__i915_active_fence_set(active, fence))
 453                 __i915_active_acquire(ref);
 454
 455 out:
 456         i915_active_release(ref);
 457         return err;
 458 }
 459
 460 static struct dma_fence *
 461 __i915_active_set_fence(struct i915_active *ref,
 462                         struct i915_active_fence *active,
 463                         struct dma_fence *fence)
 464 {
 465         struct dma_fence *prev;
 466
 467         if (replace_barrier(ref, active)) {
 468                 RCU_INIT_POINTER(active->fence, fence);
 469                 return NULL;
 470         }
 471
 472         rcu_read_lock();
 473         prev = __i915_active_fence_set(active, fence);
 474         if (prev)
 475                 prev = dma_fence_get_rcu(prev);
 476         else
 477                 __i915_active_acquire(ref);
 478         rcu_read_unlock();
 479
 480         return prev;
 481 }
 482
 483 struct dma_fence *
 484 i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
 485 {
 486         /* We expect the caller to manage the exclusive timeline ordering */
 487         return __i915_active_set_fence(ref, &ref->excl, f);
 488 }
 489
 490 bool i915_active_acquire_if_busy(struct i915_active *ref)
 491 {
 492         debug_active_assert(ref);
 493         return atomic_add_unless(&ref->count, 1, 0);
 494 }
 495
 496 static void __i915_active_activate(struct i915_active *ref)
 497 {
 498         spin_lock_irq(&ref->tree_lock); /* __active_retire() */
 499         if (!atomic_fetch_inc(&ref->count))
 500                 debug_active_activate(ref);
 501         spin_unlock_irq(&ref->tree_lock);
 502 }
 503
 504 int i915_active_acquire(struct i915_active *ref)
 505 {
 506         int err;
 507
 508         if (i915_active_acquire_if_busy(ref))
 509                 return 0;
 510
 511         if (!ref->active) {
 512                 __i915_active_activate(ref);
 513                 return 0;
 514         }
 515
 516         err = mutex_lock_interruptible(&ref->mutex);
 517         if (err)
 518                 return err;
 519
 520         if (likely(!i915_active_acquire_if_busy(ref))) {
 521                 err = ref->active(ref);
 522                 if (!err)
 523                         __i915_active_activate(ref);
 524         }
 525
 526         mutex_unlock(&ref->mutex);
 527
 528         return err;
 529 }
 530
 531 int i915_active_acquire_for_context(struct i915_active *ref, u64 idx)
 532 {
 533         struct i915_active_fence *active;
 534         int err;
 535
 536         err = i915_active_acquire(ref);
 537         if (err)
 538                 return err;
 539
 540         active = active_instance(ref, idx);
 541         if (!active) {
 542                 i915_active_release(ref);
 543                 return -ENOMEM;
 544         }
 545
 546         return 0; /* return with active ref */
 547 }
 548
 549 void i915_active_release(struct i915_active *ref)
 550 {
 551         debug_active_assert(ref);
 552         active_retire(ref);
 553 }
 554
 555 static void enable_signaling(struct i915_active_fence *active)
 556 {
 557         struct dma_fence *fence;
 558
 559         if (unlikely(is_barrier(active)))
 560                 return;
 561
 562         fence = i915_active_fence_get(active);
 563         if (!fence)
 564                 return;
 565
 566         dma_fence_enable_sw_signaling(fence);
 567         dma_fence_put(fence);
 568 }
 569
 570 static int flush_barrier(struct active_node *it)
 571 {
 572         struct intel_engine_cs *engine;
 573
 574         if (likely(!is_barrier(&it->base)))
 575                 return 0;
 576
 577         engine = __barrier_to_engine(it);
 578         smp_rmb(); /* serialise with add_active_barriers */
 579         if (!is_barrier(&it->base))
 580                 return 0;
 581
 582         return intel_engine_flush_barriers(engine);
 583 }
 584
 585 static int flush_lazy_signals(struct i915_active *ref)
 586 {
 587         struct active_node *it, *n;
 588         int err = 0;
 589
 590         enable_signaling(&ref->excl);
 591         rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
 592                 err = flush_barrier(it); /* unconnected idle barrier? */
 593                 if (err)
 594                         break;
 595
 596                 enable_signaling(&it->base);
 597         }
 598
 599         return err;
 600 }
 601
 602 int __i915_active_wait(struct i915_active *ref, int state)
 603 {
 604         might_sleep();
 605
 606         /* Any fence added after the wait begins will not be auto-signaled */
 607         if (i915_active_acquire_if_busy(ref)) {
 608                 int err;
 609
 610                 err = flush_lazy_signals(ref);
 611                 i915_active_release(ref);
 612                 if (err)
 613                         return err;
 614
 615                 if (___wait_var_event(ref, i915_active_is_idle(ref),
 616                                       state, 0, 0, schedule()))
 617                         return -EINTR;
 618         }
 619
 620         /*
 621          * After the wait is complete, the caller may free the active.
 622          * We have to flush any concurrent retirement before returning.
 623          */
 624         flush_work(&ref->work);
 625         return 0;
 626 }
 627
 628 static int __await_active(struct i915_active_fence *active,
 629                           int (*fn)(void *arg, struct dma_fence *fence),
 630                           void *arg)
 631 {
 632         struct dma_fence *fence;
 633
 634         if (is_barrier(active)) /* XXX flush the barrier? */
 635                 return 0;
 636
 637         fence = i915_active_fence_get(active);
 638         if (fence) {
 639                 int err;
 640
 641                 err = fn(arg, fence);
 642                 dma_fence_put(fence);
 643                 if (err < 0)
 644                         return err;
 645         }
 646
 647         return 0;
 648 }
 649
 650 struct wait_barrier {
 651         struct wait_queue_entry base;
 652         struct i915_active *ref;
 653 };
 654
 655 static int
 656 barrier_wake(wait_queue_entry_t *wq, unsigned int mode, int flags, void *key)
 657 {
 658         struct wait_barrier *wb = container_of(wq, typeof(*wb), base);
 659
 660         if (i915_active_is_idle(wb->ref)) {
 661                 list_del(&wq->entry);
 662                 i915_sw_fence_complete(wq->private);
 663                 kfree(wq);
 664         }
 665
 666         return 0;
 667 }
 668
 669 static int __await_barrier(struct i915_active *ref, struct i915_sw_fence *fence)
 670 {
 671         struct wait_barrier *wb;
 672
 673         wb = kmalloc(sizeof(*wb), GFP_KERNEL);
 674         if (unlikely(!wb))
 675                 return -ENOMEM;
 676
 677         GEM_BUG_ON(i915_active_is_idle(ref));
 678         if (!i915_sw_fence_await(fence)) {
 679                 kfree(wb);
 680                 return -EINVAL;
 681         }
 682
 683         wb->base.flags = 0;
 684         wb->base.func = barrier_wake;
 685         wb->base.private = fence;
 686         wb->ref = ref;
 687
 688         add_wait_queue(__var_waitqueue(ref), &wb->base);
 689         return 0;
 690 }
 691
 692 static int await_active(struct i915_active *ref,
 693                         unsigned int flags,
 694                         int (*fn)(void *arg, struct dma_fence *fence),
 695                         void *arg, struct i915_sw_fence *barrier)
 696 {
 697         int err = 0;
 698
 699         if (!i915_active_acquire_if_busy(ref))
 700                 return 0;
 701
 702         if (flags & I915_ACTIVE_AWAIT_EXCL &&
 703             rcu_access_pointer(ref->excl.fence)) {
 704                 err = __await_active(&ref->excl, fn, arg);
 705                 if (err)
 706                         goto out;
 707         }
 708
 709         if (flags & I915_ACTIVE_AWAIT_ACTIVE) {
 710                 struct active_node *it, *n;
 711
 712                 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
 713                         err = __await_active(&it->base, fn, arg);
 714                         if (err)
 715                                 goto out;
 716                 }
 717         }
 718
 719         if (flags & I915_ACTIVE_AWAIT_BARRIER) {
 720                 err = flush_lazy_signals(ref);
 721                 if (err)
 722                         goto out;
 723
 724                 err = __await_barrier(ref, barrier);
 725                 if (err)
 726                         goto out;
 727         }
 728
 729 out:
 730         i915_active_release(ref);
 731         return err;
 732 }
 733
 734 static int rq_await_fence(void *arg, struct dma_fence *fence)
 735 {
 736         return i915_request_await_dma_fence(arg, fence);
 737 }
 738
 739 int i915_request_await_active(struct i915_request *rq,
 740                               struct i915_active *ref,
 741                               unsigned int flags)
 742 {
 743         return await_active(ref, flags, rq_await_fence, rq, &rq->submit);
 744 }
 745
 746 static int sw_await_fence(void *arg, struct dma_fence *fence)
 747 {
 748         return i915_sw_fence_await_dma_fence(arg, fence, 0,
 749                                              GFP_NOWAIT | __GFP_NOWARN);
 750 }
 751
 752 int i915_sw_fence_await_active(struct i915_sw_fence *fence,
 753                                struct i915_active *ref,
 754                                unsigned int flags)
 755 {
 756         return await_active(ref, flags, sw_await_fence, fence, fence);
 757 }
 758
 759 void i915_active_fini(struct i915_active *ref)
 760 {
 761         debug_active_fini(ref);
 762         GEM_BUG_ON(atomic_read(&ref->count));
 763         GEM_BUG_ON(work_pending(&ref->work));
 764         mutex_destroy(&ref->mutex);
 765
 766         if (ref->cache)
 767                 kmem_cache_free(slab_cache, ref->cache);
 768 }
 769
 770 static inline bool is_idle_barrier(struct active_node *node, u64 idx)
 771 {
 772         return node->timeline == idx && !i915_active_fence_isset(&node->base);
 773 }
 774
 775 static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx)
 776 {
 777         struct rb_node *prev, *p;
 778
 779         if (RB_EMPTY_ROOT(&ref->tree))
 780                 return NULL;
 781
 782         GEM_BUG_ON(i915_active_is_idle(ref));
 783
 784         /*
 785          * Try to reuse any existing barrier nodes already allocated for this
 786          * i915_active, due to overlapping active phases there is likely a
 787          * node kept alive (as we reuse before parking). We prefer to reuse
 788          * completely idle barriers (less hassle in manipulating the llists),
 789          * but otherwise any will do.
 790          */
 791         if (ref->cache && is_idle_barrier(ref->cache, idx)) {
 792                 p = &ref->cache->node;
 793                 goto match;
 794         }
 795
 796         prev = NULL;
 797         p = ref->tree.rb_node;
 798         while (p) {
 799                 struct active_node *node =
 800                         rb_entry(p, struct active_node, node);
 801
 802                 if (is_idle_barrier(node, idx))
 803                         goto match;
 804
 805                 prev = p;
 806                 if (node->timeline < idx)
 807                         p = READ_ONCE(p->rb_right);
 808                 else
 809                         p = READ_ONCE(p->rb_left);
 810         }
 811
 812         /*
 813          * No quick match, but we did find the leftmost rb_node for the
 814          * kernel_context. Walk the rb_tree in-order to see if there were
 815          * any idle-barriers on this timeline that we missed, or just use
 816          * the first pending barrier.
 817          */
 818         for (p = prev; p; p = rb_next(p)) {
 819                 struct active_node *node =
 820                         rb_entry(p, struct active_node, node);
 821                 struct intel_engine_cs *engine;
 822
 823                 if (node->timeline > idx)
 824                         break;
 825
 826                 if (node->timeline < idx)
 827                         continue;
 828
 829                 if (is_idle_barrier(node, idx))
 830                         goto match;
 831
 832                 /*
 833                  * The list of pending barriers is protected by the
 834                  * kernel_context timeline, which notably we do not hold
 835                  * here. i915_request_add_active_barriers() may consume
 836                  * the barrier before we claim it, so we have to check
 837                  * for success.
 838                  */
 839                 engine = __barrier_to_engine(node);
 840                 smp_rmb(); /* serialise with add_active_barriers */
 841                 if (is_barrier(&node->base) &&
 842                     ____active_del_barrier(ref, node, engine))
 843                         goto match;
 844         }
 845
 846         return NULL;
 847
 848 match:
 849         spin_lock_irq(&ref->tree_lock);
 850         rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */
 851         if (p == &ref->cache->node)
 852                 WRITE_ONCE(ref->cache, NULL);
 853         spin_unlock_irq(&ref->tree_lock);
 854
 855         return rb_entry(p, struct active_node, node);
 856 }
 857
 858 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
 859                                             struct intel_engine_cs *engine)
 860 {
 861         intel_engine_mask_t tmp, mask = engine->mask;
 862         struct llist_node *first = NULL, *last = NULL;
 863         struct intel_gt *gt = engine->gt;
 864
 865         GEM_BUG_ON(i915_active_is_idle(ref));
 866
 867         /* Wait until the previous preallocation is completed */
 868         while (!llist_empty(&ref->preallocated_barriers))
 869                 cond_resched();
 870
 871         /*
 872          * Preallocate a node for each physical engine supporting the target
 873          * engine (remember virtual engines have more than one sibling).
 874          * We can then use the preallocated nodes in
 875          * i915_active_acquire_barrier()
 876          */
 877         GEM_BUG_ON(!mask);
 878         for_each_engine_masked(engine, gt, mask, tmp) {
 879                 u64 idx = engine->kernel_context->timeline->fence_context;
 880                 struct llist_node *prev = first;
 881                 struct active_node *node;
 882
 883                 rcu_read_lock();
 884                 node = reuse_idle_barrier(ref, idx);
 885                 rcu_read_unlock();
 886                 if (!node) {
 887                         node = kmem_cache_alloc(slab_cache, GFP_KERNEL);
 888                         if (!node)
 889                                 goto unwind;
 890
 891                         RCU_INIT_POINTER(node->base.fence, NULL);
 892                         node->base.cb.func = node_retire;
 893                         node->timeline = idx;
 894                         node->ref = ref;
 895                 }
 896
 897                 if (!i915_active_fence_isset(&node->base)) {
 898                         /*
 899                          * Mark this as being *our* unconnected proto-node.
 900                          *
 901                          * Since this node is not in any list, and we have
 902                          * decoupled it from the rbtree, we can reuse the
 903                          * request to indicate this is an idle-barrier node
 904                          * and then we can use the rb_node and list pointers
 905                          * for our tracking of the pending barrier.
 906                          */
 907                         RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
 908                         node->base.cb.node.prev = (void *)engine;
 909                         __i915_active_acquire(ref);
 910                 }
 911                 GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
 912
 913                 GEM_BUG_ON(barrier_to_engine(node) != engine);
 914                 first = barrier_to_ll(node);
 915                 first->next = prev;
 916                 if (!last)
 917                         last = first;
 918                 intel_engine_pm_get(engine);
 919         }
 920
 921         GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
 922         llist_add_batch(first, last, &ref->preallocated_barriers);
 923
 924         return 0;
 925
 926 unwind:
 927         while (first) {
 928                 struct active_node *node = barrier_from_ll(first);
 929
 930                 first = first->next;
 931
 932                 atomic_dec(&ref->count);
 933                 intel_engine_pm_put(barrier_to_engine(node));
 934
 935                 kmem_cache_free(slab_cache, node);
 936         }
 937         return -ENOMEM;
 938 }
 939
 940 void i915_active_acquire_barrier(struct i915_active *ref)
 941 {
 942         struct llist_node *pos, *next;
 943         unsigned long flags;
 944
 945         GEM_BUG_ON(i915_active_is_idle(ref));
 946
 947         /*
 948          * Transfer the list of preallocated barriers into the
 949          * i915_active rbtree, but only as proto-nodes. They will be
 950          * populated by i915_request_add_active_barriers() to point to the
 951          * request that will eventually release them.
 952          */
 953         llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
 954                 struct active_node *node = barrier_from_ll(pos);
 955                 struct intel_engine_cs *engine = barrier_to_engine(node);
 956                 struct rb_node **p, *parent;
 957
 958                 spin_lock_irqsave_nested(&ref->tree_lock, flags,
 959                                          SINGLE_DEPTH_NESTING);
 960                 parent = NULL;
 961                 p = &ref->tree.rb_node;
 962                 while (*p) {
 963                         struct active_node *it;
 964
 965                         parent = *p;
 966
 967                         it = rb_entry(parent, struct active_node, node);
 968                         if (it->timeline < node->timeline)
 969                                 p = &parent->rb_right;
 970                         else
 971                                 p = &parent->rb_left;
 972                 }
 973                 rb_link_node(&node->node, parent, p);
 974                 rb_insert_color(&node->node, &ref->tree);
 975                 spin_unlock_irqrestore(&ref->tree_lock, flags);
 976
 977                 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
 978                 llist_add(barrier_to_ll(node), &engine->barrier_tasks);
 979                 intel_engine_pm_put_delay(engine, 2);
 980         }
 981 }
 982
 983 static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
 984 {
 985         return __active_fence_slot(&barrier_from_ll(node)->base);
 986 }
 987
 988 void i915_request_add_active_barriers(struct i915_request *rq)
 989 {
 990         struct intel_engine_cs *engine = rq->engine;
 991         struct llist_node *node, *next;
 992         unsigned long flags;
 993
 994         GEM_BUG_ON(!intel_context_is_barrier(rq->context));
 995         GEM_BUG_ON(intel_engine_is_virtual(engine));
 996         GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline);
 997
 998         node = llist_del_all(&engine->barrier_tasks);
 999         if (!node)
1000                 return;
1001         /*
1002          * Attach the list of proto-fences to the in-flight request such
1003          * that the parent i915_active will be released when this request
1004          * is retired.
1005          */
1006         spin_lock_irqsave(&rq->lock, flags);
1007         llist_for_each_safe(node, next, node) {
1008                 /* serialise with reuse_idle_barrier */
1009                 smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
1010                 list_add_tail((struct list_head *)node, &rq->fence.cb_list);
1011         }
1012         spin_unlock_irqrestore(&rq->lock, flags);
1013 }
1014
1015 /*
1016  * __i915_active_fence_set: Update the last active fence along its timeline
1017  * @active: the active tracker
1018  * @fence: the new fence (under construction)
1019  *
1020  * Records the new @fence as the last active fence along its timeline in
1021  * this active tracker, moving the tracking callbacks from the previous
1022  * fence onto this one. Returns the previous fence (if not already completed),
1023  * which the caller must ensure is executed before the new fence. To ensure
1024  * that the order of fences within the timeline of the i915_active_fence is
1025  * understood, it should be locked by the caller.
1026  */
1027 struct dma_fence *
1028 __i915_active_fence_set(struct i915_active_fence *active,
1029                         struct dma_fence *fence)
1030 {
1031         struct dma_fence *prev;
1032         unsigned long flags;
1033
1034         if (fence == rcu_access_pointer(active->fence))
1035                 return fence;
1036
1037         GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
1038
1039         /*
1040          * Consider that we have two threads arriving (A and B), with
1041          * C already resident as the active->fence.
1042          *
1043          * A does the xchg first, and so it sees C or NULL depending
1044          * on the timing of the interrupt handler. If it is NULL, the
1045          * previous fence must have been signaled and we know that
1046          * we are first on the timeline. If it is still present,
1047          * we acquire the lock on that fence and serialise with the interrupt
1048          * handler, in the process removing it from any future interrupt
1049          * callback. A will then wait on C before executing (if present).
1050          *
1051          * As B is second, it sees A as the previous fence and so waits for
1052          * it to complete its transition and takes over the occupancy for
1053          * itself -- remembering that it needs to wait on A before executing.
1054          *
1055          * Note the strong ordering of the timeline also provides consistent
1056          * nesting rules for the fence->lock; the inner lock is always the
1057          * older lock.
1058          */
1059         spin_lock_irqsave(fence->lock, flags);
1060         prev = xchg(__active_fence_slot(active), fence);
1061         if (prev) {
1062                 GEM_BUG_ON(prev == fence);
1063                 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
1064                 __list_del_entry(&active->cb.node);
1065                 spin_unlock(prev->lock); /* serialise with prev->cb_list */
1066         }
1067         list_add_tail(&active->cb.node, &fence->cb_list);
1068         spin_unlock_irqrestore(fence->lock, flags);
1069
1070         return prev;
1071 }
1072
1073 int i915_active_fence_set(struct i915_active_fence *active,
1074                           struct i915_request *rq)
1075 {
1076         struct dma_fence *fence;
1077         int err = 0;
1078
1079         /* Must maintain timeline ordering wrt previous active requests */
1080         rcu_read_lock();
1081         fence = __i915_active_fence_set(active, &rq->fence);
1082         if (fence) /* but the previous fence may not belong to that timeline! */
1083                 fence = dma_fence_get_rcu(fence);
1084         rcu_read_unlock();
1085         if (fence) {
1086                 err = i915_request_await_dma_fence(rq, fence);
1087                 dma_fence_put(fence);
1088         }
1089
1090         return err;
1091 }
1092
1093 void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb)
1094 {
1095         active_fence_cb(fence, cb);
1096 }
1097
1098 struct auto_active {
1099         struct i915_active base;
1100         struct kref ref;
1101 };
1102
1103 struct i915_active *i915_active_get(struct i915_active *ref)
1104 {
1105         struct auto_active *aa = container_of(ref, typeof(*aa), base);
1106
1107         kref_get(&aa->ref);
1108         return &aa->base;
1109 }
1110
1111 static void auto_release(struct kref *ref)
1112 {
1113         struct auto_active *aa = container_of(ref, typeof(*aa), ref);
1114
1115         i915_active_fini(&aa->base);
1116         kfree(aa);
1117 }
1118
1119 void i915_active_put(struct i915_active *ref)
1120 {
1121         struct auto_active *aa = container_of(ref, typeof(*aa), base);
1122
1123         kref_put(&aa->ref, auto_release);
1124 }
1125
1126 static int auto_active(struct i915_active *ref)
1127 {
1128         i915_active_get(ref);
1129         return 0;
1130 }
1131
1132 static void auto_retire(struct i915_active *ref)
1133 {
1134         i915_active_put(ref);
1135 }
1136
1137 struct i915_active *i915_active_create(void)
1138 {
1139         struct auto_active *aa;
1140
1141         aa = kmalloc(sizeof(*aa), GFP_KERNEL);
1142         if (!aa)
1143                 return NULL;
1144
1145         kref_init(&aa->ref);
1146         i915_active_init(&aa->base, auto_active, auto_retire, 0);
1147
1148         return &aa->base;
1149 }
1150
1151 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1152 #include "selftests/i915_active.c"
1153 #endif
1154
1155 void i915_active_module_exit(void)
1156 {
1157         kmem_cache_destroy(slab_cache);
1158 }
1159
1160 int __init i915_active_module_init(void)
1161 {
1162         slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN);
1163         if (!slab_cache)
1164                 return -ENOMEM;
1165
1166         return 0;
1167 }