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