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 #include "i915_globals.h"
  17
  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 {
  26         struct i915_global base;
  27         struct kmem_cache *slab_cache;
  28 } global;
  29
  30 struct active_node {
  31         struct rb_node node;
  32         struct i915_active_fence base;
  33         struct i915_active *ref;
  34         u64 timeline;
  35 };
  36
  37 #define fetch_node(x) rb_entry(READ_ONCE(x), typeof(struct active_node), node)
  38
  39 static inline struct active_node *
  40 node_from_active(struct i915_active_fence *active)
  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
  47 static inline bool is_barrier(const struct i915_active_fence *active)
  48 {
  49         return IS_ERR(rcu_access_pointer(active->fence));
  50 }
  51
  52 static inline struct llist_node *barrier_to_ll(struct active_node *node)
  53 {
  54         GEM_BUG_ON(!is_barrier(&node->base));
  55         return (struct llist_node *)&node->base.cb.node;
  56 }
  57
  58 static inline struct intel_engine_cs *
  59 __barrier_to_engine(struct active_node *node)
  60 {
  61         return (struct intel_engine_cs *)READ_ONCE(node->base.cb.node.prev);
  62 }
  63
  64 static inline struct intel_engine_cs *
  65 barrier_to_engine(struct active_node *node)
  66 {
  67         GEM_BUG_ON(!is_barrier(&node->base));
  68         return __barrier_to_engine(node);
  69 }
  70
  71 static inline struct active_node *barrier_from_ll(struct llist_node *x)
  72 {
  73         return container_of((struct list_head *)x,
  74                             struct active_node, base.cb.node);
  75 }
  76
  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
  83         return (void *)ref->active ?: (void *)ref->retire ?: (void *)ref;
  84 }
  85
  86 static const 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 {
  98         lockdep_assert_held(&ref->tree_lock);
  99         if (!atomic_read(&ref->count)) /* before the first inc */
 100                 debug_object_activate(ref, &active_debug_desc);
 101 }
 102
 103 static void debug_active_deactivate(struct i915_active *ref)
 104 {
 105         lockdep_assert_held(&ref->tree_lock);
 106         if (!atomic_read(&ref->count)) /* after the last dec */
 107                 debug_object_deactivate(ref, &active_debug_desc);
 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
 130 static void
 131 __active_retire(struct i915_active *ref)
 132 {
 133         struct rb_root root = RB_ROOT;
 134         struct active_node *it, *n;
 135         unsigned long flags;
 136
 137         GEM_BUG_ON(i915_active_is_idle(ref));
 138
 139         /* return the unused nodes to our slabcache -- flushing the allocator */
 140         if (!atomic_dec_and_lock_irqsave(&ref->count, &ref->tree_lock, flags))
 141                 return;
 142
 143         GEM_BUG_ON(rcu_access_pointer(ref->excl.fence));
 144         debug_active_deactivate(ref);
 145
 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
 161                 /* Make the cached node available for reuse with any timeline */
 162                 ref->cache->timeline = 0; /* needs cmpxchg(u64) */
 163         }
 164
 165         spin_unlock_irqrestore(&ref->tree_lock, flags);
 166
 167         /* After the final retire, the entire struct may be freed */
 168         if (ref->retire)
 169                 ref->retire(ref);
 170
 171         /* ... except if you wait on it, you must manage your own references! */
 172         wake_up_var(ref);
 173
 174         /* Finally free the discarded timeline tree  */
 175         rbtree_postorder_for_each_entry_safe(it, n, &root, node) {
 176                 GEM_BUG_ON(i915_active_fence_isset(&it->base));
 177                 kmem_cache_free(global.slab_cache, it);
 178         }
 179 }
 180
 181 static void
 182 active_work(struct work_struct *wrk)
 183 {
 184         struct i915_active *ref = container_of(wrk, typeof(*ref), work);
 185
 186         GEM_BUG_ON(!atomic_read(&ref->count));
 187         if (atomic_add_unless(&ref->count, -1, 1))
 188                 return;
 189
 190         __active_retire(ref);
 191 }
 192
 193 static void
 194 active_retire(struct i915_active *ref)
 195 {
 196         GEM_BUG_ON(!atomic_read(&ref->count));
 197         if (atomic_add_unless(&ref->count, -1, 1))
 198                 return;
 199
 200         if (ref->flags & I915_ACTIVE_RETIRE_SLEEPS) {
 201                 queue_work(system_unbound_wq, &ref->work);
 202                 return;
 203         }
 204
 205         __active_retire(ref);
 206 }
 207
 208 static inline struct dma_fence **
 209 __active_fence_slot(struct i915_active_fence *active)
 210 {
 211         return (struct dma_fence ** __force)&active->fence;
 212 }
 213
 214 static inline bool
 215 active_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
 216 {
 217         struct i915_active_fence *active =
 218                 container_of(cb, typeof(*active), cb);
 219
 220         return cmpxchg(__active_fence_slot(active), fence, NULL) == fence;
 221 }
 222
 223 static void
 224 node_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
 225 {
 226         if (active_fence_cb(fence, cb))
 227                 active_retire(container_of(cb, struct active_node, base.cb)->ref);
 228 }
 229
 230 static void
 231 excl_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
 232 {
 233         if (active_fence_cb(fence, cb))
 234                 active_retire(container_of(cb, struct i915_active, excl.cb));
 235 }
 236
 237 static struct active_node *__active_lookup(struct i915_active *ref, u64 idx)
 238 {
 239         struct active_node *it;
 240
 241         GEM_BUG_ON(idx == 0); /* 0 is the unordered timeline, rsvd for cache */
 242
 243         /*
 244          * We track the most recently used timeline to skip a rbtree search
 245          * for the common case, under typical loads we never need the rbtree
 246          * at all. We can reuse the last slot if it is empty, that is
 247          * after the previous activity has been retired, or if it matches the
 248          * current timeline.
 249          */
 250         it = READ_ONCE(ref->cache);
 251         if (it) {
 252                 u64 cached = READ_ONCE(it->timeline);
 253
 254                 /* Once claimed, this slot will only belong to this idx */
 255                 if (cached == idx)
 256                         return it;
 257
 258                 /*
 259                  * An unclaimed cache [.timeline=0] can only be claimed once.
 260                  *
 261                  * If the value is already non-zero, some other thread has
 262                  * claimed the cache and we know that is does not match our
 263                  * idx. If, and only if, the timeline is currently zero is it
 264                  * worth competing to claim it atomically for ourselves (for
 265                  * only the winner of that race will cmpxchg return the old
 266                  * value of 0).
 267                  */
 268                 if (!cached && !cmpxchg64(&it->timeline, 0, idx))
 269                         return it;
 270         }
 271
 272         BUILD_BUG_ON(offsetof(typeof(*it), node));
 273
 274         /* While active, the tree can only be built; not destroyed */
 275         GEM_BUG_ON(i915_active_is_idle(ref));
 276
 277         it = fetch_node(ref->tree.rb_node);
 278         while (it) {
 279                 if (it->timeline < idx) {
 280                         it = fetch_node(it->node.rb_right);
 281                 } else if (it->timeline > idx) {
 282                         it = fetch_node(it->node.rb_left);
 283                 } else {
 284                         WRITE_ONCE(ref->cache, it);
 285                         break;
 286                 }
 287         }
 288
 289         /* NB: If the tree rotated beneath us, we may miss our target. */
 290         return it;
 291 }
 292
 293 static struct i915_active_fence *
 294 active_instance(struct i915_active *ref, u64 idx)
 295 {
 296         struct active_node *node;
 297         struct rb_node **p, *parent;
 298
 299         node = __active_lookup(ref, idx);
 300         if (likely(node))
 301                 return &node->base;
 302
 303         spin_lock_irq(&ref->tree_lock);
 304         GEM_BUG_ON(i915_active_is_idle(ref));
 305
 306         parent = NULL;
 307         p = &ref->tree.rb_node;
 308         while (*p) {
 309                 parent = *p;
 310
 311                 node = rb_entry(parent, struct active_node, node);
 312                 if (node->timeline == idx)
 313                         goto out;
 314
 315                 if (node->timeline < idx)
 316                         p = &parent->rb_right;
 317                 else
 318                         p = &parent->rb_left;
 319         }
 320
 321         /*
 322          * XXX: We should preallocate this before i915_active_ref() is ever
 323          *  called, but we cannot call into fs_reclaim() anyway, so use GFP_ATOMIC.
 324          */
 325         node = kmem_cache_alloc(global.slab_cache, GFP_ATOMIC);
 326         if (!node)
 327                 goto out;
 328
 329         __i915_active_fence_init(&node->base, NULL, node_retire);
 330         node->ref = ref;
 331         node->timeline = idx;
 332
 333         rb_link_node(&node->node, parent, p);
 334         rb_insert_color(&node->node, &ref->tree);
 335
 336 out:
 337         WRITE_ONCE(ref->cache, node);
 338         spin_unlock_irq(&ref->tree_lock);
 339
 340         return &node->base;
 341 }
 342
 343 void __i915_active_init(struct i915_active *ref,
 344                         int (*active)(struct i915_active *ref),
 345                         void (*retire)(struct i915_active *ref),
 346                         struct lock_class_key *mkey,
 347                         struct lock_class_key *wkey)
 348 {
 349         unsigned long bits;
 350
 351         debug_active_init(ref);
 352
 353         ref->flags = 0;
 354         ref->active = active;
 355         ref->retire = ptr_unpack_bits(retire, &bits, 2);
 356         if (bits & I915_ACTIVE_MAY_SLEEP)
 357                 ref->flags |= I915_ACTIVE_RETIRE_SLEEPS;
 358
 359         spin_lock_init(&ref->tree_lock);
 360         ref->tree = RB_ROOT;
 361         ref->cache = NULL;
 362
 363         init_llist_head(&ref->preallocated_barriers);
 364         atomic_set(&ref->count, 0);
 365         __mutex_init(&ref->mutex, "i915_active", mkey);
 366         __i915_active_fence_init(&ref->excl, NULL, excl_retire);
 367         INIT_WORK(&ref->work, active_work);
 368 #if IS_ENABLED(CONFIG_LOCKDEP)
 369         lockdep_init_map(&ref->work.lockdep_map, "i915_active.work", wkey, 0);
 370 #endif
 371 }
 372
 373 static bool ____active_del_barrier(struct i915_active *ref,
 374                                    struct active_node *node,
 375                                    struct intel_engine_cs *engine)
 376
 377 {
 378         struct llist_node *head = NULL, *tail = NULL;
 379         struct llist_node *pos, *next;
 380
 381         GEM_BUG_ON(node->timeline != engine->kernel_context->timeline->fence_context);
 382
 383         /*
 384          * Rebuild the llist excluding our node. We may perform this
 385          * outside of the kernel_context timeline mutex and so someone
 386          * else may be manipulating the engine->barrier_tasks, in
 387          * which case either we or they will be upset :)
 388          *
 389          * A second __active_del_barrier() will report failure to claim
 390          * the active_node and the caller will just shrug and know not to
 391          * claim ownership of its node.
 392          *
 393          * A concurrent i915_request_add_active_barriers() will miss adding
 394          * any of the tasks, but we will try again on the next -- and since
 395          * we are actively using the barrier, we know that there will be
 396          * at least another opportunity when we idle.
 397          */
 398         llist_for_each_safe(pos, next, llist_del_all(&engine->barrier_tasks)) {
 399                 if (node == barrier_from_ll(pos)) {
 400                         node = NULL;
 401                         continue;
 402                 }
 403
 404                 pos->next = head;
 405                 head = pos;
 406                 if (!tail)
 407                         tail = pos;
 408         }
 409         if (head)
 410                 llist_add_batch(head, tail, &engine->barrier_tasks);
 411
 412         return !node;
 413 }
 414
 415 static bool
 416 __active_del_barrier(struct i915_active *ref, struct active_node *node)
 417 {
 418         return ____active_del_barrier(ref, node, barrier_to_engine(node));
 419 }
 420
 421 static bool
 422 replace_barrier(struct i915_active *ref, struct i915_active_fence *active)
 423 {
 424         if (!is_barrier(active)) /* proto-node used by our idle barrier? */
 425                 return false;
 426
 427         /*
 428          * This request is on the kernel_context timeline, and so
 429          * we can use it to substitute for the pending idle-barrer
 430          * request that we want to emit on the kernel_context.
 431          */
 432         __active_del_barrier(ref, node_from_active(active));
 433         return true;
 434 }
 435
 436 int i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence)
 437 {
 438         struct i915_active_fence *active;
 439         int err;
 440
 441         /* Prevent reaping in case we malloc/wait while building the tree */
 442         err = i915_active_acquire(ref);
 443         if (err)
 444                 return err;
 445
 446         active = active_instance(ref, idx);
 447         if (!active) {
 448                 err = -ENOMEM;
 449                 goto out;
 450         }
 451
 452         if (replace_barrier(ref, active)) {
 453                 RCU_INIT_POINTER(active->fence, NULL);
 454                 atomic_dec(&ref->count);
 455         }
 456         if (!__i915_active_fence_set(active, fence))
 457                 __i915_active_acquire(ref);
 458
 459 out:
 460         i915_active_release(ref);
 461         return err;
 462 }
 463
 464 static struct dma_fence *
 465 __i915_active_set_fence(struct i915_active *ref,
 466                         struct i915_active_fence *active,
 467                         struct dma_fence *fence)
 468 {
 469         struct dma_fence *prev;
 470
 471         if (replace_barrier(ref, active)) {
 472                 RCU_INIT_POINTER(active->fence, fence);
 473                 return NULL;
 474         }
 475
 476         rcu_read_lock();
 477         prev = __i915_active_fence_set(active, fence);
 478         if (prev)
 479                 prev = dma_fence_get_rcu(prev);
 480         else
 481                 __i915_active_acquire(ref);
 482         rcu_read_unlock();
 483
 484         return prev;
 485 }
 486
 487 static struct i915_active_fence *
 488 __active_fence(struct i915_active *ref, u64 idx)
 489 {
 490         struct active_node *it;
 491
 492         it = __active_lookup(ref, idx);
 493         if (unlikely(!it)) { /* Contention with parallel tree builders! */
 494                 spin_lock_irq(&ref->tree_lock);
 495                 it = __active_lookup(ref, idx);
 496                 spin_unlock_irq(&ref->tree_lock);
 497         }
 498         GEM_BUG_ON(!it); /* slot must be preallocated */
 499
 500         return &it->base;
 501 }
 502
 503 struct dma_fence *
 504 __i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence)
 505 {
 506         /* Only valid while active, see i915_active_acquire_for_context() */
 507         return __i915_active_set_fence(ref, __active_fence(ref, idx), fence);
 508 }
 509
 510 struct dma_fence *
 511 i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
 512 {
 513         /* We expect the caller to manage the exclusive timeline ordering */
 514         return __i915_active_set_fence(ref, &ref->excl, f);
 515 }
 516
 517 bool i915_active_acquire_if_busy(struct i915_active *ref)
 518 {
 519         debug_active_assert(ref);
 520         return atomic_add_unless(&ref->count, 1, 0);
 521 }
 522
 523 static void __i915_active_activate(struct i915_active *ref)
 524 {
 525         spin_lock_irq(&ref->tree_lock); /* __active_retire() */
 526         if (!atomic_fetch_inc(&ref->count))
 527                 debug_active_activate(ref);
 528         spin_unlock_irq(&ref->tree_lock);
 529 }
 530
 531 int i915_active_acquire(struct i915_active *ref)
 532 {
 533         int err;
 534
 535         if (i915_active_acquire_if_busy(ref))
 536                 return 0;
 537
 538         if (!ref->active) {
 539                 __i915_active_activate(ref);
 540                 return 0;
 541         }
 542
 543         err = mutex_lock_interruptible(&ref->mutex);
 544         if (err)
 545                 return err;
 546
 547         if (likely(!i915_active_acquire_if_busy(ref))) {
 548                 err = ref->active(ref);
 549                 if (!err)
 550                         __i915_active_activate(ref);
 551         }
 552
 553         mutex_unlock(&ref->mutex);
 554
 555         return err;
 556 }
 557
 558 int i915_active_acquire_for_context(struct i915_active *ref, u64 idx)
 559 {
 560         struct i915_active_fence *active;
 561         int err;
 562
 563         err = i915_active_acquire(ref);
 564         if (err)
 565                 return err;
 566
 567         active = active_instance(ref, idx);
 568         if (!active) {
 569                 i915_active_release(ref);
 570                 return -ENOMEM;
 571         }
 572
 573         return 0; /* return with active ref */
 574 }
 575
 576 void i915_active_release(struct i915_active *ref)
 577 {
 578         debug_active_assert(ref);
 579         active_retire(ref);
 580 }
 581
 582 static void enable_signaling(struct i915_active_fence *active)
 583 {
 584         struct dma_fence *fence;
 585
 586         if (unlikely(is_barrier(active)))
 587                 return;
 588
 589         fence = i915_active_fence_get(active);
 590         if (!fence)
 591                 return;
 592
 593         dma_fence_enable_sw_signaling(fence);
 594         dma_fence_put(fence);
 595 }
 596
 597 static int flush_barrier(struct active_node *it)
 598 {
 599         struct intel_engine_cs *engine;
 600
 601         if (likely(!is_barrier(&it->base)))
 602                 return 0;
 603
 604         engine = __barrier_to_engine(it);
 605         smp_rmb(); /* serialise with add_active_barriers */
 606         if (!is_barrier(&it->base))
 607                 return 0;
 608
 609         return intel_engine_flush_barriers(engine);
 610 }
 611
 612 static int flush_lazy_signals(struct i915_active *ref)
 613 {
 614         struct active_node *it, *n;
 615         int err = 0;
 616
 617         enable_signaling(&ref->excl);
 618         rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
 619                 err = flush_barrier(it); /* unconnected idle barrier? */
 620                 if (err)
 621                         break;
 622
 623                 enable_signaling(&it->base);
 624         }
 625
 626         return err;
 627 }
 628
 629 int __i915_active_wait(struct i915_active *ref, int state)
 630 {
 631         might_sleep();
 632
 633         /* Any fence added after the wait begins will not be auto-signaled */
 634         if (i915_active_acquire_if_busy(ref)) {
 635                 int err;
 636
 637                 err = flush_lazy_signals(ref);
 638                 i915_active_release(ref);
 639                 if (err)
 640                         return err;
 641
 642                 if (___wait_var_event(ref, i915_active_is_idle(ref),
 643                                       state, 0, 0, schedule()))
 644                         return -EINTR;
 645         }
 646
 647         /*
 648          * After the wait is complete, the caller may free the active.
 649          * We have to flush any concurrent retirement before returning.
 650          */
 651         flush_work(&ref->work);
 652         return 0;
 653 }
 654
 655 static int __await_active(struct i915_active_fence *active,
 656                           int (*fn)(void *arg, struct dma_fence *fence),
 657                           void *arg)
 658 {
 659         struct dma_fence *fence;
 660
 661         if (is_barrier(active)) /* XXX flush the barrier? */
 662                 return 0;
 663
 664         fence = i915_active_fence_get(active);
 665         if (fence) {
 666                 int err;
 667
 668                 err = fn(arg, fence);
 669                 dma_fence_put(fence);
 670                 if (err < 0)
 671                         return err;
 672         }
 673
 674         return 0;
 675 }
 676
 677 struct wait_barrier {
 678         struct wait_queue_entry base;
 679         struct i915_active *ref;
 680 };
 681
 682 static int
 683 barrier_wake(wait_queue_entry_t *wq, unsigned int mode, int flags, void *key)
 684 {
 685         struct wait_barrier *wb = container_of(wq, typeof(*wb), base);
 686
 687         if (i915_active_is_idle(wb->ref)) {
 688                 list_del(&wq->entry);
 689                 i915_sw_fence_complete(wq->private);
 690                 kfree(wq);
 691         }
 692
 693         return 0;
 694 }
 695
 696 static int __await_barrier(struct i915_active *ref, struct i915_sw_fence *fence)
 697 {
 698         struct wait_barrier *wb;
 699
 700         wb = kmalloc(sizeof(*wb), GFP_KERNEL);
 701         if (unlikely(!wb))
 702                 return -ENOMEM;
 703
 704         GEM_BUG_ON(i915_active_is_idle(ref));
 705         if (!i915_sw_fence_await(fence)) {
 706                 kfree(wb);
 707                 return -EINVAL;
 708         }
 709
 710         wb->base.flags = 0;
 711         wb->base.func = barrier_wake;
 712         wb->base.private = fence;
 713         wb->ref = ref;
 714
 715         add_wait_queue(__var_waitqueue(ref), &wb->base);
 716         return 0;
 717 }
 718
 719 static int await_active(struct i915_active *ref,
 720                         unsigned int flags,
 721                         int (*fn)(void *arg, struct dma_fence *fence),
 722                         void *arg, struct i915_sw_fence *barrier)
 723 {
 724         int err = 0;
 725
 726         if (!i915_active_acquire_if_busy(ref))
 727                 return 0;
 728
 729         if (flags & I915_ACTIVE_AWAIT_EXCL &&
 730             rcu_access_pointer(ref->excl.fence)) {
 731                 err = __await_active(&ref->excl, fn, arg);
 732                 if (err)
 733                         goto out;
 734         }
 735
 736         if (flags & I915_ACTIVE_AWAIT_ACTIVE) {
 737                 struct active_node *it, *n;
 738
 739                 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
 740                         err = __await_active(&it->base, fn, arg);
 741                         if (err)
 742                                 goto out;
 743                 }
 744         }
 745
 746         if (flags & I915_ACTIVE_AWAIT_BARRIER) {
 747                 err = flush_lazy_signals(ref);
 748                 if (err)
 749                         goto out;
 750
 751                 err = __await_barrier(ref, barrier);
 752                 if (err)
 753                         goto out;
 754         }
 755
 756 out:
 757         i915_active_release(ref);
 758         return err;
 759 }
 760
 761 static int rq_await_fence(void *arg, struct dma_fence *fence)
 762 {
 763         return i915_request_await_dma_fence(arg, fence);
 764 }
 765
 766 int i915_request_await_active(struct i915_request *rq,
 767                               struct i915_active *ref,
 768                               unsigned int flags)
 769 {
 770         return await_active(ref, flags, rq_await_fence, rq, &rq->submit);
 771 }
 772
 773 static int sw_await_fence(void *arg, struct dma_fence *fence)
 774 {
 775         return i915_sw_fence_await_dma_fence(arg, fence, 0,
 776                                              GFP_NOWAIT | __GFP_NOWARN);
 777 }
 778
 779 int i915_sw_fence_await_active(struct i915_sw_fence *fence,
 780                                struct i915_active *ref,
 781                                unsigned int flags)
 782 {
 783         return await_active(ref, flags, sw_await_fence, fence, fence);
 784 }
 785
 786 void i915_active_fini(struct i915_active *ref)
 787 {
 788         debug_active_fini(ref);
 789         GEM_BUG_ON(atomic_read(&ref->count));
 790         GEM_BUG_ON(work_pending(&ref->work));
 791         mutex_destroy(&ref->mutex);
 792
 793         if (ref->cache)
 794                 kmem_cache_free(global.slab_cache, ref->cache);
 795 }
 796
 797 static inline bool is_idle_barrier(struct active_node *node, u64 idx)
 798 {
 799         return node->timeline == idx && !i915_active_fence_isset(&node->base);
 800 }
 801
 802 static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx)
 803 {
 804         struct rb_node *prev, *p;
 805
 806         if (RB_EMPTY_ROOT(&ref->tree))
 807                 return NULL;
 808
 809         GEM_BUG_ON(i915_active_is_idle(ref));
 810
 811         /*
 812          * Try to reuse any existing barrier nodes already allocated for this
 813          * i915_active, due to overlapping active phases there is likely a
 814          * node kept alive (as we reuse before parking). We prefer to reuse
 815          * completely idle barriers (less hassle in manipulating the llists),
 816          * but otherwise any will do.
 817          */
 818         if (ref->cache && is_idle_barrier(ref->cache, idx)) {
 819                 p = &ref->cache->node;
 820                 goto match;
 821         }
 822
 823         prev = NULL;
 824         p = ref->tree.rb_node;
 825         while (p) {
 826                 struct active_node *node =
 827                         rb_entry(p, struct active_node, node);
 828
 829                 if (is_idle_barrier(node, idx))
 830                         goto match;
 831
 832                 prev = p;
 833                 if (node->timeline < idx)
 834                         p = READ_ONCE(p->rb_right);
 835                 else
 836                         p = READ_ONCE(p->rb_left);
 837         }
 838
 839         /*
 840          * No quick match, but we did find the leftmost rb_node for the
 841          * kernel_context. Walk the rb_tree in-order to see if there were
 842          * any idle-barriers on this timeline that we missed, or just use
 843          * the first pending barrier.
 844          */
 845         for (p = prev; p; p = rb_next(p)) {
 846                 struct active_node *node =
 847                         rb_entry(p, struct active_node, node);
 848                 struct intel_engine_cs *engine;
 849
 850                 if (node->timeline > idx)
 851                         break;
 852
 853                 if (node->timeline < idx)
 854                         continue;
 855
 856                 if (is_idle_barrier(node, idx))
 857                         goto match;
 858
 859                 /*
 860                  * The list of pending barriers is protected by the
 861                  * kernel_context timeline, which notably we do not hold
 862                  * here. i915_request_add_active_barriers() may consume
 863                  * the barrier before we claim it, so we have to check
 864                  * for success.
 865                  */
 866                 engine = __barrier_to_engine(node);
 867                 smp_rmb(); /* serialise with add_active_barriers */
 868                 if (is_barrier(&node->base) &&
 869                     ____active_del_barrier(ref, node, engine))
 870                         goto match;
 871         }
 872
 873         return NULL;
 874
 875 match:
 876         spin_lock_irq(&ref->tree_lock);
 877         rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */
 878         if (p == &ref->cache->node)
 879                 WRITE_ONCE(ref->cache, NULL);
 880         spin_unlock_irq(&ref->tree_lock);
 881
 882         return rb_entry(p, struct active_node, node);
 883 }
 884
 885 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
 886                                             struct intel_engine_cs *engine)
 887 {
 888         intel_engine_mask_t tmp, mask = engine->mask;
 889         struct llist_node *first = NULL, *last = NULL;
 890         struct intel_gt *gt = engine->gt;
 891
 892         GEM_BUG_ON(i915_active_is_idle(ref));
 893
 894         /* Wait until the previous preallocation is completed */
 895         while (!llist_empty(&ref->preallocated_barriers))
 896                 cond_resched();
 897
 898         /*
 899          * Preallocate a node for each physical engine supporting the target
 900          * engine (remember virtual engines have more than one sibling).
 901          * We can then use the preallocated nodes in
 902          * i915_active_acquire_barrier()
 903          */
 904         GEM_BUG_ON(!mask);
 905         for_each_engine_masked(engine, gt, mask, tmp) {
 906                 u64 idx = engine->kernel_context->timeline->fence_context;
 907                 struct llist_node *prev = first;
 908                 struct active_node *node;
 909
 910                 rcu_read_lock();
 911                 node = reuse_idle_barrier(ref, idx);
 912                 rcu_read_unlock();
 913                 if (!node) {
 914                         node = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
 915                         if (!node)
 916                                 goto unwind;
 917
 918                         RCU_INIT_POINTER(node->base.fence, NULL);
 919                         node->base.cb.func = node_retire;
 920                         node->timeline = idx;
 921                         node->ref = ref;
 922                 }
 923
 924                 if (!i915_active_fence_isset(&node->base)) {
 925                         /*
 926                          * Mark this as being *our* unconnected proto-node.
 927                          *
 928                          * Since this node is not in any list, and we have
 929                          * decoupled it from the rbtree, we can reuse the
 930                          * request to indicate this is an idle-barrier node
 931                          * and then we can use the rb_node and list pointers
 932                          * for our tracking of the pending barrier.
 933                          */
 934                         RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
 935                         node->base.cb.node.prev = (void *)engine;
 936                         __i915_active_acquire(ref);
 937                 }
 938                 GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
 939
 940                 GEM_BUG_ON(barrier_to_engine(node) != engine);
 941                 first = barrier_to_ll(node);
 942                 first->next = prev;
 943                 if (!last)
 944                         last = first;
 945                 intel_engine_pm_get(engine);
 946         }
 947
 948         GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
 949         llist_add_batch(first, last, &ref->preallocated_barriers);
 950
 951         return 0;
 952
 953 unwind:
 954         while (first) {
 955                 struct active_node *node = barrier_from_ll(first);
 956
 957                 first = first->next;
 958
 959                 atomic_dec(&ref->count);
 960                 intel_engine_pm_put(barrier_to_engine(node));
 961
 962                 kmem_cache_free(global.slab_cache, node);
 963         }
 964         return -ENOMEM;
 965 }
 966
 967 void i915_active_acquire_barrier(struct i915_active *ref)
 968 {
 969         struct llist_node *pos, *next;
 970         unsigned long flags;
 971
 972         GEM_BUG_ON(i915_active_is_idle(ref));
 973
 974         /*
 975          * Transfer the list of preallocated barriers into the
 976          * i915_active rbtree, but only as proto-nodes. They will be
 977          * populated by i915_request_add_active_barriers() to point to the
 978          * request that will eventually release them.
 979          */
 980         llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
 981                 struct active_node *node = barrier_from_ll(pos);
 982                 struct intel_engine_cs *engine = barrier_to_engine(node);
 983                 struct rb_node **p, *parent;
 984
 985                 spin_lock_irqsave_nested(&ref->tree_lock, flags,
 986                                          SINGLE_DEPTH_NESTING);
 987                 parent = NULL;
 988                 p = &ref->tree.rb_node;
 989                 while (*p) {
 990                         struct active_node *it;
 991
 992                         parent = *p;
 993
 994                         it = rb_entry(parent, struct active_node, node);
 995                         if (it->timeline < node->timeline)
 996                                 p = &parent->rb_right;
 997                         else
 998                                 p = &parent->rb_left;
 999                 }
1000                 rb_link_node(&node->node, parent, p);
1001                 rb_insert_color(&node->node, &ref->tree);
1002                 spin_unlock_irqrestore(&ref->tree_lock, flags);
1003
1004                 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
1005                 llist_add(barrier_to_ll(node), &engine->barrier_tasks);
1006                 intel_engine_pm_put_delay(engine, 1);
1007         }
1008 }
1009
1010 static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
1011 {
1012         return __active_fence_slot(&barrier_from_ll(node)->base);
1013 }
1014
1015 void i915_request_add_active_barriers(struct i915_request *rq)
1016 {
1017         struct intel_engine_cs *engine = rq->engine;
1018         struct llist_node *node, *next;
1019         unsigned long flags;
1020
1021         GEM_BUG_ON(!intel_context_is_barrier(rq->context));
1022         GEM_BUG_ON(intel_engine_is_virtual(engine));
1023         GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline);
1024
1025         node = llist_del_all(&engine->barrier_tasks);
1026         if (!node)
1027                 return;
1028         /*
1029          * Attach the list of proto-fences to the in-flight request such
1030          * that the parent i915_active will be released when this request
1031          * is retired.
1032          */
1033         spin_lock_irqsave(&rq->lock, flags);
1034         llist_for_each_safe(node, next, node) {
1035                 /* serialise with reuse_idle_barrier */
1036                 smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
1037                 list_add_tail((struct list_head *)node, &rq->fence.cb_list);
1038         }
1039         spin_unlock_irqrestore(&rq->lock, flags);
1040 }
1041
1042 /*
1043  * __i915_active_fence_set: Update the last active fence along its timeline
1044  * @active: the active tracker
1045  * @fence: the new fence (under construction)
1046  *
1047  * Records the new @fence as the last active fence along its timeline in
1048  * this active tracker, moving the tracking callbacks from the previous
1049  * fence onto this one. Returns the previous fence (if not already completed),
1050  * which the caller must ensure is executed before the new fence. To ensure
1051  * that the order of fences within the timeline of the i915_active_fence is
1052  * understood, it should be locked by the caller.
1053  */
1054 struct dma_fence *
1055 __i915_active_fence_set(struct i915_active_fence *active,
1056                         struct dma_fence *fence)
1057 {
1058         struct dma_fence *prev;
1059         unsigned long flags;
1060
1061         if (fence == rcu_access_pointer(active->fence))
1062                 return fence;
1063
1064         GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
1065
1066         /*
1067          * Consider that we have two threads arriving (A and B), with
1068          * C already resident as the active->fence.
1069          *
1070          * A does the xchg first, and so it sees C or NULL depending
1071          * on the timing of the interrupt handler. If it is NULL, the
1072          * previous fence must have been signaled and we know that
1073          * we are first on the timeline. If it is still present,
1074          * we acquire the lock on that fence and serialise with the interrupt
1075          * handler, in the process removing it from any future interrupt
1076          * callback. A will then wait on C before executing (if present).
1077          *
1078          * As B is second, it sees A as the previous fence and so waits for
1079          * it to complete its transition and takes over the occupancy for
1080          * itself -- remembering that it needs to wait on A before executing.
1081          *
1082          * Note the strong ordering of the timeline also provides consistent
1083          * nesting rules for the fence->lock; the inner lock is always the
1084          * older lock.
1085          */
1086         spin_lock_irqsave(fence->lock, flags);
1087         prev = xchg(__active_fence_slot(active), fence);
1088         if (prev) {
1089                 GEM_BUG_ON(prev == fence);
1090                 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
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         rcu_read_lock();
1108         fence = __i915_active_fence_set(active, &rq->fence);
1109         if (fence) /* but the previous fence may not belong to that timeline! */
1110                 fence = dma_fence_get_rcu(fence);
1111         rcu_read_unlock();
1112         if (fence) {
1113                 err = i915_request_await_dma_fence(rq, fence);
1114                 dma_fence_put(fence);
1115         }
1116
1117         return err;
1118 }
1119
1120 void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb)
1121 {
1122         active_fence_cb(fence, cb);
1123 }
1124
1125 struct auto_active {
1126         struct i915_active base;
1127         struct kref ref;
1128 };
1129
1130 struct i915_active *i915_active_get(struct i915_active *ref)
1131 {
1132         struct auto_active *aa = container_of(ref, typeof(*aa), base);
1133
1134         kref_get(&aa->ref);
1135         return &aa->base;
1136 }
1137
1138 static void auto_release(struct kref *ref)
1139 {
1140         struct auto_active *aa = container_of(ref, typeof(*aa), ref);
1141
1142         i915_active_fini(&aa->base);
1143         kfree(aa);
1144 }
1145
1146 void i915_active_put(struct i915_active *ref)
1147 {
1148         struct auto_active *aa = container_of(ref, typeof(*aa), base);
1149
1150         kref_put(&aa->ref, auto_release);
1151 }
1152
1153 static int auto_active(struct i915_active *ref)
1154 {
1155         i915_active_get(ref);
1156         return 0;
1157 }
1158
1159 __i915_active_call static void
1160 auto_retire(struct i915_active *ref)
1161 {
1162         i915_active_put(ref);
1163 }
1164
1165 struct i915_active *i915_active_create(void)
1166 {
1167         struct auto_active *aa;
1168
1169         aa = kmalloc(sizeof(*aa), GFP_KERNEL);
1170         if (!aa)
1171                 return NULL;
1172
1173         kref_init(&aa->ref);
1174         i915_active_init(&aa->base, auto_active, auto_retire);
1175
1176         return &aa->base;
1177 }
1178
1179 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1180 #include "selftests/i915_active.c"
1181 #endif
1182
1183 static void i915_global_active_shrink(void)
1184 {
1185         kmem_cache_shrink(global.slab_cache);
1186 }
1187
1188 static void i915_global_active_exit(void)
1189 {
1190         kmem_cache_destroy(global.slab_cache);
1191 }
1192
1193 static struct i915_global_active global = { {
1194         .shrink = i915_global_active_shrink,
1195         .exit = i915_global_active_exit,
1196 } };
1197
1198 int __init i915_global_active_init(void)
1199 {
1200         global.slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN);
1201         if (!global.slab_cache)
1202                 return -ENOMEM;
1203
1204         i915_global_register(&global.base);
1205         return 0;
1206 }