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 try_cmpxchg(__active_fence_slot(active), &fence, NULL);
 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         fence = __i915_active_fence_set(active, fence);
 453         if (!fence)
 454                 __i915_active_acquire(ref);
 455         else
 456                 dma_fence_put(fence);
 457
 458 out:
 459         i915_active_release(ref);
 460         return err;
 461 }
 462
 463 static struct dma_fence *
 464 __i915_active_set_fence(struct i915_active *ref,
 465                         struct i915_active_fence *active,
 466                         struct dma_fence *fence)
 467 {
 468         struct dma_fence *prev;
 469
 470         if (replace_barrier(ref, active)) {
 471                 RCU_INIT_POINTER(active->fence, fence);
 472                 return NULL;
 473         }
 474
 475         prev = __i915_active_fence_set(active, fence);
 476         if (!prev)
 477                 __i915_active_acquire(ref);
 478
 479         return prev;
 480 }
 481
 482 struct dma_fence *
 483 i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
 484 {
 485         /* We expect the caller to manage the exclusive timeline ordering */
 486         return __i915_active_set_fence(ref, &ref->excl, f);
 487 }
 488
 489 bool i915_active_acquire_if_busy(struct i915_active *ref)
 490 {
 491         debug_active_assert(ref);
 492         return atomic_add_unless(&ref->count, 1, 0);
 493 }
 494
 495 static void __i915_active_activate(struct i915_active *ref)
 496 {
 497         spin_lock_irq(&ref->tree_lock); /* __active_retire() */
 498         if (!atomic_fetch_inc(&ref->count))
 499                 debug_active_activate(ref);
 500         spin_unlock_irq(&ref->tree_lock);
 501 }
 502
 503 int i915_active_acquire(struct i915_active *ref)
 504 {
 505         int err;
 506
 507         if (i915_active_acquire_if_busy(ref))
 508                 return 0;
 509
 510         if (!ref->active) {
 511                 __i915_active_activate(ref);
 512                 return 0;
 513         }
 514
 515         err = mutex_lock_interruptible(&ref->mutex);
 516         if (err)
 517                 return err;
 518
 519         if (likely(!i915_active_acquire_if_busy(ref))) {
 520                 err = ref->active(ref);
 521                 if (!err)
 522                         __i915_active_activate(ref);
 523         }
 524
 525         mutex_unlock(&ref->mutex);
 526
 527         return err;
 528 }
 529
 530 void i915_active_release(struct i915_active *ref)
 531 {
 532         debug_active_assert(ref);
 533         active_retire(ref);
 534 }
 535
 536 static void enable_signaling(struct i915_active_fence *active)
 537 {
 538         struct dma_fence *fence;
 539
 540         if (unlikely(is_barrier(active)))
 541                 return;
 542
 543         fence = i915_active_fence_get(active);
 544         if (!fence)
 545                 return;
 546
 547         dma_fence_enable_sw_signaling(fence);
 548         dma_fence_put(fence);
 549 }
 550
 551 static int flush_barrier(struct active_node *it)
 552 {
 553         struct intel_engine_cs *engine;
 554
 555         if (likely(!is_barrier(&it->base)))
 556                 return 0;
 557
 558         engine = __barrier_to_engine(it);
 559         smp_rmb(); /* serialise with add_active_barriers */
 560         if (!is_barrier(&it->base))
 561                 return 0;
 562
 563         return intel_engine_flush_barriers(engine);
 564 }
 565
 566 static int flush_lazy_signals(struct i915_active *ref)
 567 {
 568         struct active_node *it, *n;
 569         int err = 0;
 570
 571         enable_signaling(&ref->excl);
 572         rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
 573                 err = flush_barrier(it); /* unconnected idle barrier? */
 574                 if (err)
 575                         break;
 576
 577                 enable_signaling(&it->base);
 578         }
 579
 580         return err;
 581 }
 582
 583 int __i915_active_wait(struct i915_active *ref, int state)
 584 {
 585         might_sleep();
 586
 587         /* Any fence added after the wait begins will not be auto-signaled */
 588         if (i915_active_acquire_if_busy(ref)) {
 589                 int err;
 590
 591                 err = flush_lazy_signals(ref);
 592                 i915_active_release(ref);
 593                 if (err)
 594                         return err;
 595
 596                 if (___wait_var_event(ref, i915_active_is_idle(ref),
 597                                       state, 0, 0, schedule()))
 598                         return -EINTR;
 599         }
 600
 601         /*
 602          * After the wait is complete, the caller may free the active.
 603          * We have to flush any concurrent retirement before returning.
 604          */
 605         flush_work(&ref->work);
 606         return 0;
 607 }
 608
 609 static int __await_active(struct i915_active_fence *active,
 610                           int (*fn)(void *arg, struct dma_fence *fence),
 611                           void *arg)
 612 {
 613         struct dma_fence *fence;
 614
 615         if (is_barrier(active)) /* XXX flush the barrier? */
 616                 return 0;
 617
 618         fence = i915_active_fence_get(active);
 619         if (fence) {
 620                 int err;
 621
 622                 err = fn(arg, fence);
 623                 dma_fence_put(fence);
 624                 if (err < 0)
 625                         return err;
 626         }
 627
 628         return 0;
 629 }
 630
 631 struct wait_barrier {
 632         struct wait_queue_entry base;
 633         struct i915_active *ref;
 634 };
 635
 636 static int
 637 barrier_wake(wait_queue_entry_t *wq, unsigned int mode, int flags, void *key)
 638 {
 639         struct wait_barrier *wb = container_of(wq, typeof(*wb), base);
 640
 641         if (i915_active_is_idle(wb->ref)) {
 642                 list_del(&wq->entry);
 643                 i915_sw_fence_complete(wq->private);
 644                 kfree(wq);
 645         }
 646
 647         return 0;
 648 }
 649
 650 static int __await_barrier(struct i915_active *ref, struct i915_sw_fence *fence)
 651 {
 652         struct wait_barrier *wb;
 653
 654         wb = kmalloc(sizeof(*wb), GFP_KERNEL);
 655         if (unlikely(!wb))
 656                 return -ENOMEM;
 657
 658         GEM_BUG_ON(i915_active_is_idle(ref));
 659         if (!i915_sw_fence_await(fence)) {
 660                 kfree(wb);
 661                 return -EINVAL;
 662         }
 663
 664         wb->base.flags = 0;
 665         wb->base.func = barrier_wake;
 666         wb->base.private = fence;
 667         wb->ref = ref;
 668
 669         add_wait_queue(__var_waitqueue(ref), &wb->base);
 670         return 0;
 671 }
 672
 673 static int await_active(struct i915_active *ref,
 674                         unsigned int flags,
 675                         int (*fn)(void *arg, struct dma_fence *fence),
 676                         void *arg, struct i915_sw_fence *barrier)
 677 {
 678         int err = 0;
 679
 680         if (!i915_active_acquire_if_busy(ref))
 681                 return 0;
 682
 683         if (flags & I915_ACTIVE_AWAIT_EXCL &&
 684             rcu_access_pointer(ref->excl.fence)) {
 685                 err = __await_active(&ref->excl, fn, arg);
 686                 if (err)
 687                         goto out;
 688         }
 689
 690         if (flags & I915_ACTIVE_AWAIT_ACTIVE) {
 691                 struct active_node *it, *n;
 692
 693                 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
 694                         err = __await_active(&it->base, fn, arg);
 695                         if (err)
 696                                 goto out;
 697                 }
 698         }
 699
 700         if (flags & I915_ACTIVE_AWAIT_BARRIER) {
 701                 err = flush_lazy_signals(ref);
 702                 if (err)
 703                         goto out;
 704
 705                 err = __await_barrier(ref, barrier);
 706                 if (err)
 707                         goto out;
 708         }
 709
 710 out:
 711         i915_active_release(ref);
 712         return err;
 713 }
 714
 715 static int rq_await_fence(void *arg, struct dma_fence *fence)
 716 {
 717         return i915_request_await_dma_fence(arg, fence);
 718 }
 719
 720 int i915_request_await_active(struct i915_request *rq,
 721                               struct i915_active *ref,
 722                               unsigned int flags)
 723 {
 724         return await_active(ref, flags, rq_await_fence, rq, &rq->submit);
 725 }
 726
 727 static int sw_await_fence(void *arg, struct dma_fence *fence)
 728 {
 729         return i915_sw_fence_await_dma_fence(arg, fence, 0,
 730                                              GFP_NOWAIT | __GFP_NOWARN);
 731 }
 732
 733 int i915_sw_fence_await_active(struct i915_sw_fence *fence,
 734                                struct i915_active *ref,
 735                                unsigned int flags)
 736 {
 737         return await_active(ref, flags, sw_await_fence, fence, fence);
 738 }
 739
 740 void i915_active_fini(struct i915_active *ref)
 741 {
 742         debug_active_fini(ref);
 743         GEM_BUG_ON(atomic_read(&ref->count));
 744         GEM_BUG_ON(work_pending(&ref->work));
 745         mutex_destroy(&ref->mutex);
 746
 747         if (ref->cache)
 748                 kmem_cache_free(slab_cache, ref->cache);
 749 }
 750
 751 static inline bool is_idle_barrier(struct active_node *node, u64 idx)
 752 {
 753         return node->timeline == idx && !i915_active_fence_isset(&node->base);
 754 }
 755
 756 static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx)
 757 {
 758         struct rb_node *prev, *p;
 759
 760         if (RB_EMPTY_ROOT(&ref->tree))
 761                 return NULL;
 762
 763         GEM_BUG_ON(i915_active_is_idle(ref));
 764
 765         /*
 766          * Try to reuse any existing barrier nodes already allocated for this
 767          * i915_active, due to overlapping active phases there is likely a
 768          * node kept alive (as we reuse before parking). We prefer to reuse
 769          * completely idle barriers (less hassle in manipulating the llists),
 770          * but otherwise any will do.
 771          */
 772         if (ref->cache && is_idle_barrier(ref->cache, idx)) {
 773                 p = &ref->cache->node;
 774                 goto match;
 775         }
 776
 777         prev = NULL;
 778         p = ref->tree.rb_node;
 779         while (p) {
 780                 struct active_node *node =
 781                         rb_entry(p, struct active_node, node);
 782
 783                 if (is_idle_barrier(node, idx))
 784                         goto match;
 785
 786                 prev = p;
 787                 if (node->timeline < idx)
 788                         p = READ_ONCE(p->rb_right);
 789                 else
 790                         p = READ_ONCE(p->rb_left);
 791         }
 792
 793         /*
 794          * No quick match, but we did find the leftmost rb_node for the
 795          * kernel_context. Walk the rb_tree in-order to see if there were
 796          * any idle-barriers on this timeline that we missed, or just use
 797          * the first pending barrier.
 798          */
 799         for (p = prev; p; p = rb_next(p)) {
 800                 struct active_node *node =
 801                         rb_entry(p, struct active_node, node);
 802                 struct intel_engine_cs *engine;
 803
 804                 if (node->timeline > idx)
 805                         break;
 806
 807                 if (node->timeline < idx)
 808                         continue;
 809
 810                 if (is_idle_barrier(node, idx))
 811                         goto match;
 812
 813                 /*
 814                  * The list of pending barriers is protected by the
 815                  * kernel_context timeline, which notably we do not hold
 816                  * here. i915_request_add_active_barriers() may consume
 817                  * the barrier before we claim it, so we have to check
 818                  * for success.
 819                  */
 820                 engine = __barrier_to_engine(node);
 821                 smp_rmb(); /* serialise with add_active_barriers */
 822                 if (is_barrier(&node->base) &&
 823                     ____active_del_barrier(ref, node, engine))
 824                         goto match;
 825         }
 826
 827         return NULL;
 828
 829 match:
 830         spin_lock_irq(&ref->tree_lock);
 831         rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */
 832         if (p == &ref->cache->node)
 833                 WRITE_ONCE(ref->cache, NULL);
 834         spin_unlock_irq(&ref->tree_lock);
 835
 836         return rb_entry(p, struct active_node, node);
 837 }
 838
 839 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
 840                                             struct intel_engine_cs *engine)
 841 {
 842         intel_engine_mask_t tmp, mask = engine->mask;
 843         struct llist_node *first = NULL, *last = NULL;
 844         struct intel_gt *gt = engine->gt;
 845
 846         GEM_BUG_ON(i915_active_is_idle(ref));
 847
 848         /* Wait until the previous preallocation is completed */
 849         while (!llist_empty(&ref->preallocated_barriers))
 850                 cond_resched();
 851
 852         /*
 853          * Preallocate a node for each physical engine supporting the target
 854          * engine (remember virtual engines have more than one sibling).
 855          * We can then use the preallocated nodes in
 856          * i915_active_acquire_barrier()
 857          */
 858         GEM_BUG_ON(!mask);
 859         for_each_engine_masked(engine, gt, mask, tmp) {
 860                 u64 idx = engine->kernel_context->timeline->fence_context;
 861                 struct llist_node *prev = first;
 862                 struct active_node *node;
 863
 864                 rcu_read_lock();
 865                 node = reuse_idle_barrier(ref, idx);
 866                 rcu_read_unlock();
 867                 if (!node) {
 868                         node = kmem_cache_alloc(slab_cache, GFP_KERNEL);
 869                         if (!node)
 870                                 goto unwind;
 871
 872                         RCU_INIT_POINTER(node->base.fence, NULL);
 873                         node->base.cb.func = node_retire;
 874                         node->timeline = idx;
 875                         node->ref = ref;
 876                 }
 877
 878                 if (!i915_active_fence_isset(&node->base)) {
 879                         /*
 880                          * Mark this as being *our* unconnected proto-node.
 881                          *
 882                          * Since this node is not in any list, and we have
 883                          * decoupled it from the rbtree, we can reuse the
 884                          * request to indicate this is an idle-barrier node
 885                          * and then we can use the rb_node and list pointers
 886                          * for our tracking of the pending barrier.
 887                          */
 888                         RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
 889                         node->base.cb.node.prev = (void *)engine;
 890                         __i915_active_acquire(ref);
 891                 }
 892                 GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
 893
 894                 GEM_BUG_ON(barrier_to_engine(node) != engine);
 895                 first = barrier_to_ll(node);
 896                 first->next = prev;
 897                 if (!last)
 898                         last = first;
 899                 intel_engine_pm_get(engine);
 900         }
 901
 902         GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
 903         llist_add_batch(first, last, &ref->preallocated_barriers);
 904
 905         return 0;
 906
 907 unwind:
 908         while (first) {
 909                 struct active_node *node = barrier_from_ll(first);
 910
 911                 first = first->next;
 912
 913                 atomic_dec(&ref->count);
 914                 intel_engine_pm_put(barrier_to_engine(node));
 915
 916                 kmem_cache_free(slab_cache, node);
 917         }
 918         return -ENOMEM;
 919 }
 920
 921 void i915_active_acquire_barrier(struct i915_active *ref)
 922 {
 923         struct llist_node *pos, *next;
 924         unsigned long flags;
 925
 926         GEM_BUG_ON(i915_active_is_idle(ref));
 927
 928         /*
 929          * Transfer the list of preallocated barriers into the
 930          * i915_active rbtree, but only as proto-nodes. They will be
 931          * populated by i915_request_add_active_barriers() to point to the
 932          * request that will eventually release them.
 933          */
 934         llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
 935                 struct active_node *node = barrier_from_ll(pos);
 936                 struct intel_engine_cs *engine = barrier_to_engine(node);
 937                 struct rb_node **p, *parent;
 938
 939                 spin_lock_irqsave_nested(&ref->tree_lock, flags,
 940                                          SINGLE_DEPTH_NESTING);
 941                 parent = NULL;
 942                 p = &ref->tree.rb_node;
 943                 while (*p) {
 944                         struct active_node *it;
 945
 946                         parent = *p;
 947
 948                         it = rb_entry(parent, struct active_node, node);
 949                         if (it->timeline < node->timeline)
 950                                 p = &parent->rb_right;
 951                         else
 952                                 p = &parent->rb_left;
 953                 }
 954                 rb_link_node(&node->node, parent, p);
 955                 rb_insert_color(&node->node, &ref->tree);
 956                 spin_unlock_irqrestore(&ref->tree_lock, flags);
 957
 958                 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
 959                 llist_add(barrier_to_ll(node), &engine->barrier_tasks);
 960                 intel_engine_pm_put_delay(engine, 2);
 961         }
 962 }
 963
 964 static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
 965 {
 966         return __active_fence_slot(&barrier_from_ll(node)->base);
 967 }
 968
 969 void i915_request_add_active_barriers(struct i915_request *rq)
 970 {
 971         struct intel_engine_cs *engine = rq->engine;
 972         struct llist_node *node, *next;
 973         unsigned long flags;
 974
 975         GEM_BUG_ON(!intel_context_is_barrier(rq->context));
 976         GEM_BUG_ON(intel_engine_is_virtual(engine));
 977         GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline);
 978
 979         node = llist_del_all(&engine->barrier_tasks);
 980         if (!node)
 981                 return;
 982         /*
 983          * Attach the list of proto-fences to the in-flight request such
 984          * that the parent i915_active will be released when this request
 985          * is retired.
 986          */
 987         spin_lock_irqsave(&rq->lock, flags);
 988         llist_for_each_safe(node, next, node) {
 989                 /* serialise with reuse_idle_barrier */
 990                 smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
 991                 list_add_tail((struct list_head *)node, &rq->fence.cb_list);
 992         }
 993         spin_unlock_irqrestore(&rq->lock, flags);
 994 }
 995
 996 /*
 997  * __i915_active_fence_set: Update the last active fence along its timeline
 998  * @active: the active tracker
 999  * @fence: the new fence (under construction)
1000  *
1001  * Records the new @fence as the last active fence along its timeline in
1002  * this active tracker, moving the tracking callbacks from the previous
1003  * fence onto this one. Gets and returns a reference to the previous fence
1004  * (if not already completed), which the caller must put after making sure
1005  * that it is executed before the new fence. To ensure that the order of
1006  * fences within the timeline of the i915_active_fence is understood, it
1007  * should be locked by the caller.
1008  */
1009 struct dma_fence *
1010 __i915_active_fence_set(struct i915_active_fence *active,
1011                         struct dma_fence *fence)
1012 {
1013         struct dma_fence *prev;
1014         unsigned long flags;
1015
1016         /*
1017          * In case of fences embedded in i915_requests, their memory is
1018          * SLAB_FAILSAFE_BY_RCU, then it can be reused right after release
1019          * by new requests.  Then, there is a risk of passing back a pointer
1020          * to a new, completely unrelated fence that reuses the same memory
1021          * while tracked under a different active tracker.  Combined with i915
1022          * perf open/close operations that build await dependencies between
1023          * engine kernel context requests and user requests from different
1024          * timelines, this can lead to dependency loops and infinite waits.
1025          *
1026          * As a countermeasure, we try to get a reference to the active->fence
1027          * first, so if we succeed and pass it back to our user then it is not
1028          * released and potentially reused by an unrelated request before the
1029          * user has a chance to set up an await dependency on it.
1030          */
1031         prev = i915_active_fence_get(active);
1032         if (fence == prev)
1033                 return fence;
1034
1035         GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
1036
1037         /*
1038          * Consider that we have two threads arriving (A and B), with
1039          * C already resident as the active->fence.
1040          *
1041          * Both A and B have got a reference to C or NULL, depending on the
1042          * timing of the interrupt handler.  Let's assume that if A has got C
1043          * then it has locked C first (before B).
1044          *
1045          * Note the strong ordering of the timeline also provides consistent
1046          * nesting rules for the fence->lock; the inner lock is always the
1047          * older lock.
1048          */
1049         spin_lock_irqsave(fence->lock, flags);
1050         if (prev)
1051                 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
1052
1053         /*
1054          * A does the cmpxchg first, and so it sees C or NULL, as before, or
1055          * something else, depending on the timing of other threads and/or
1056          * interrupt handler.  If not the same as before then A unlocks C if
1057          * applicable and retries, starting from an attempt to get a new
1058          * active->fence.  Meanwhile, B follows the same path as A.
1059          * Once A succeeds with cmpxch, B fails again, retires, gets A from
1060          * active->fence, locks it as soon as A completes, and possibly
1061          * succeeds with cmpxchg.
1062          */
1063         while (cmpxchg(__active_fence_slot(active), prev, fence) != prev) {
1064                 if (prev) {
1065                         spin_unlock(prev->lock);
1066                         dma_fence_put(prev);
1067                 }
1068                 spin_unlock_irqrestore(fence->lock, flags);
1069
1070                 prev = i915_active_fence_get(active);
1071                 GEM_BUG_ON(prev == fence);
1072
1073                 spin_lock_irqsave(fence->lock, flags);
1074                 if (prev)
1075                         spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
1076         }
1077
1078         /*
1079          * If prev is NULL then the previous fence must have been signaled
1080          * and we know that we are first on the timeline.  If it is still
1081          * present then, having the lock on that fence already acquired, we
1082          * serialise with the interrupt handler, in the process of removing it
1083          * from any future interrupt callback.  A will then wait on C before
1084          * executing (if present).
1085          *
1086          * As B is second, it sees A as the previous fence and so waits for
1087          * it to complete its transition and takes over the occupancy for
1088          * itself -- remembering that it needs to wait on A before executing.
1089          */
1090         if (prev) {
1091                 __list_del_entry(&active->cb.node);
1092                 spin_unlock(prev->lock); /* serialise with prev->cb_list */
1093         }
1094         list_add_tail(&active->cb.node, &fence->cb_list);
1095         spin_unlock_irqrestore(fence->lock, flags);
1096
1097         return prev;
1098 }
1099
1100 int i915_active_fence_set(struct i915_active_fence *active,
1101                           struct i915_request *rq)
1102 {
1103         struct dma_fence *fence;
1104         int err = 0;
1105
1106         /* Must maintain timeline ordering wrt previous active requests */
1107         fence = __i915_active_fence_set(active, &rq->fence);
1108         if (fence) {
1109                 err = i915_request_await_dma_fence(rq, fence);
1110                 dma_fence_put(fence);
1111         }
1112
1113         return err;
1114 }
1115
1116 void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb)
1117 {
1118         active_fence_cb(fence, cb);
1119 }
1120
1121 struct auto_active {
1122         struct i915_active base;
1123         struct kref ref;
1124 };
1125
1126 struct i915_active *i915_active_get(struct i915_active *ref)
1127 {
1128         struct auto_active *aa = container_of(ref, typeof(*aa), base);
1129
1130         kref_get(&aa->ref);
1131         return &aa->base;
1132 }
1133
1134 static void auto_release(struct kref *ref)
1135 {
1136         struct auto_active *aa = container_of(ref, typeof(*aa), ref);
1137
1138         i915_active_fini(&aa->base);
1139         kfree(aa);
1140 }
1141
1142 void i915_active_put(struct i915_active *ref)
1143 {
1144         struct auto_active *aa = container_of(ref, typeof(*aa), base);
1145
1146         kref_put(&aa->ref, auto_release);
1147 }
1148
1149 static int auto_active(struct i915_active *ref)
1150 {
1151         i915_active_get(ref);
1152         return 0;
1153 }
1154
1155 static void auto_retire(struct i915_active *ref)
1156 {
1157         i915_active_put(ref);
1158 }
1159
1160 struct i915_active *i915_active_create(void)
1161 {
1162         struct auto_active *aa;
1163
1164         aa = kmalloc(sizeof(*aa), GFP_KERNEL);
1165         if (!aa)
1166                 return NULL;
1167
1168         kref_init(&aa->ref);
1169         i915_active_init(&aa->base, auto_active, auto_retire, 0);
1170
1171         return &aa->base;
1172 }
1173
1174 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1175 #include "selftests/i915_active.c"
1176 #endif
1177
1178 void i915_active_module_exit(void)
1179 {
1180         kmem_cache_destroy(slab_cache);
1181 }
1182
1183 int __init i915_active_module_init(void)
1184 {
1185         slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN);
1186         if (!slab_cache)
1187                 return -ENOMEM;
1188
1189         return 0;
1190 }