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_pm.h"
  11 #include "gt/intel_ring.h"
  12
  13 #include "i915_drv.h"
  14 #include "i915_active.h"
  15 #include "i915_globals.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 i915_global_active {
  25         struct i915_global base;
  26         struct kmem_cache *slab_cache;
  27 } global;
  28
  29 struct active_node {
  30         struct i915_active_fence base;
  31         struct i915_active *ref;
  32         struct rb_node node;
  33         u64 timeline;
  34 };
  35
  36 static inline struct active_node *
  37 node_from_active(struct i915_active_fence *active)
  38 {
  39         return container_of(active, struct active_node, base);
  40 }
  41
  42 #define take_preallocated_barriers(x) llist_del_all(&(x)->preallocated_barriers)
  43
  44 static inline bool is_barrier(const struct i915_active_fence *active)
  45 {
  46         return IS_ERR(rcu_access_pointer(active->fence));
  47 }
  48
  49 static inline struct llist_node *barrier_to_ll(struct active_node *node)
  50 {
  51         GEM_BUG_ON(!is_barrier(&node->base));
  52         return (struct llist_node *)&node->base.cb.node;
  53 }
  54
  55 static inline struct intel_engine_cs *
  56 __barrier_to_engine(struct active_node *node)
  57 {
  58         return (struct intel_engine_cs *)READ_ONCE(node->base.cb.node.prev);
  59 }
  60
  61 static inline struct intel_engine_cs *
  62 barrier_to_engine(struct active_node *node)
  63 {
  64         GEM_BUG_ON(!is_barrier(&node->base));
  65         return __barrier_to_engine(node);
  66 }
  67
  68 static inline struct active_node *barrier_from_ll(struct llist_node *x)
  69 {
  70         return container_of((struct list_head *)x,
  71                             struct active_node, base.cb.node);
  72 }
  73
  74 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) && IS_ENABLED(CONFIG_DEBUG_OBJECTS)
  75
  76 static void *active_debug_hint(void *addr)
  77 {
  78         struct i915_active *ref = addr;
  79
  80         return (void *)ref->active ?: (void *)ref->retire ?: (void *)ref;
  81 }
  82
  83 static struct debug_obj_descr active_debug_desc = {
  84         .name = "i915_active",
  85         .debug_hint = active_debug_hint,
  86 };
  87
  88 static void debug_active_init(struct i915_active *ref)
  89 {
  90         debug_object_init(ref, &active_debug_desc);
  91 }
  92
  93 static void debug_active_activate(struct i915_active *ref)
  94 {
  95         lockdep_assert_held(&ref->tree_lock);
  96         if (!atomic_read(&ref->count)) /* before the first inc */
  97                 debug_object_activate(ref, &active_debug_desc);
  98 }
  99
 100 static void debug_active_deactivate(struct i915_active *ref)
 101 {
 102         lockdep_assert_held(&ref->tree_lock);
 103         if (!atomic_read(&ref->count)) /* after the last dec */
 104                 debug_object_deactivate(ref, &active_debug_desc);
 105 }
 106
 107 static void debug_active_fini(struct i915_active *ref)
 108 {
 109         debug_object_free(ref, &active_debug_desc);
 110 }
 111
 112 static void debug_active_assert(struct i915_active *ref)
 113 {
 114         debug_object_assert_init(ref, &active_debug_desc);
 115 }
 116
 117 #else
 118
 119 static inline void debug_active_init(struct i915_active *ref) { }
 120 static inline void debug_active_activate(struct i915_active *ref) { }
 121 static inline void debug_active_deactivate(struct i915_active *ref) { }
 122 static inline void debug_active_fini(struct i915_active *ref) { }
 123 static inline void debug_active_assert(struct i915_active *ref) { }
 124
 125 #endif
 126
 127 static void
 128 __active_retire(struct i915_active *ref)
 129 {
 130         struct active_node *it, *n;
 131         struct rb_root root;
 132         unsigned long flags;
 133
 134         GEM_BUG_ON(i915_active_is_idle(ref));
 135
 136         /* return the unused nodes to our slabcache -- flushing the allocator */
 137         if (!atomic_dec_and_lock_irqsave(&ref->count, &ref->tree_lock, flags))
 138                 return;
 139
 140         GEM_BUG_ON(rcu_access_pointer(ref->excl.fence));
 141         debug_active_deactivate(ref);
 142
 143         root = ref->tree;
 144         ref->tree = RB_ROOT;
 145         ref->cache = NULL;
 146
 147         spin_unlock_irqrestore(&ref->tree_lock, flags);
 148
 149         /* After the final retire, the entire struct may be freed */
 150         if (ref->retire)
 151                 ref->retire(ref);
 152
 153         /* ... except if you wait on it, you must manage your own references! */
 154         wake_up_var(ref);
 155
 156         rbtree_postorder_for_each_entry_safe(it, n, &root, node) {
 157                 GEM_BUG_ON(i915_active_fence_isset(&it->base));
 158                 kmem_cache_free(global.slab_cache, it);
 159         }
 160 }
 161
 162 static void
 163 active_work(struct work_struct *wrk)
 164 {
 165         struct i915_active *ref = container_of(wrk, typeof(*ref), work);
 166
 167         GEM_BUG_ON(!atomic_read(&ref->count));
 168         if (atomic_add_unless(&ref->count, -1, 1))
 169                 return;
 170
 171         __active_retire(ref);
 172 }
 173
 174 static void
 175 active_retire(struct i915_active *ref)
 176 {
 177         GEM_BUG_ON(!atomic_read(&ref->count));
 178         if (atomic_add_unless(&ref->count, -1, 1))
 179                 return;
 180
 181         if (ref->flags & I915_ACTIVE_RETIRE_SLEEPS) {
 182                 queue_work(system_unbound_wq, &ref->work);
 183                 return;
 184         }
 185
 186         __active_retire(ref);
 187 }
 188
 189 static inline struct dma_fence **
 190 __active_fence_slot(struct i915_active_fence *active)
 191 {
 192         return (struct dma_fence ** __force)&active->fence;
 193 }
 194
 195 static inline bool
 196 active_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
 197 {
 198         struct i915_active_fence *active =
 199                 container_of(cb, typeof(*active), cb);
 200
 201         return cmpxchg(__active_fence_slot(active), fence, NULL) == fence;
 202 }
 203
 204 static void
 205 node_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
 206 {
 207         if (active_fence_cb(fence, cb))
 208                 active_retire(container_of(cb, struct active_node, base.cb)->ref);
 209 }
 210
 211 static void
 212 excl_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
 213 {
 214         if (active_fence_cb(fence, cb))
 215                 active_retire(container_of(cb, struct i915_active, excl.cb));
 216 }
 217
 218 static struct i915_active_fence *
 219 active_instance(struct i915_active *ref, struct intel_timeline *tl)
 220 {
 221         struct active_node *node, *prealloc;
 222         struct rb_node **p, *parent;
 223         u64 idx = tl->fence_context;
 224
 225         /*
 226          * We track the most recently used timeline to skip a rbtree search
 227          * for the common case, under typical loads we never need the rbtree
 228          * at all. We can reuse the last slot if it is empty, that is
 229          * after the previous activity has been retired, or if it matches the
 230          * current timeline.
 231          */
 232         node = READ_ONCE(ref->cache);
 233         if (node && node->timeline == idx)
 234                 return &node->base;
 235
 236         /* Preallocate a replacement, just in case */
 237         prealloc = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
 238         if (!prealloc)
 239                 return NULL;
 240
 241         spin_lock_irq(&ref->tree_lock);
 242         GEM_BUG_ON(i915_active_is_idle(ref));
 243
 244         parent = NULL;
 245         p = &ref->tree.rb_node;
 246         while (*p) {
 247                 parent = *p;
 248
 249                 node = rb_entry(parent, struct active_node, node);
 250                 if (node->timeline == idx) {
 251                         kmem_cache_free(global.slab_cache, prealloc);
 252                         goto out;
 253                 }
 254
 255                 if (node->timeline < idx)
 256                         p = &parent->rb_right;
 257                 else
 258                         p = &parent->rb_left;
 259         }
 260
 261         node = prealloc;
 262         __i915_active_fence_init(&node->base, NULL, node_retire);
 263         node->ref = ref;
 264         node->timeline = idx;
 265
 266         rb_link_node(&node->node, parent, p);
 267         rb_insert_color(&node->node, &ref->tree);
 268
 269 out:
 270         ref->cache = node;
 271         spin_unlock_irq(&ref->tree_lock);
 272
 273         BUILD_BUG_ON(offsetof(typeof(*node), base));
 274         return &node->base;
 275 }
 276
 277 void __i915_active_init(struct i915_active *ref,
 278                         int (*active)(struct i915_active *ref),
 279                         void (*retire)(struct i915_active *ref),
 280                         struct lock_class_key *mkey,
 281                         struct lock_class_key *wkey)
 282 {
 283         unsigned long bits;
 284
 285         debug_active_init(ref);
 286
 287         ref->flags = 0;
 288         ref->active = active;
 289         ref->retire = ptr_unpack_bits(retire, &bits, 2);
 290         if (bits & I915_ACTIVE_MAY_SLEEP)
 291                 ref->flags |= I915_ACTIVE_RETIRE_SLEEPS;
 292
 293         spin_lock_init(&ref->tree_lock);
 294         ref->tree = RB_ROOT;
 295         ref->cache = NULL;
 296
 297         init_llist_head(&ref->preallocated_barriers);
 298         atomic_set(&ref->count, 0);
 299         __mutex_init(&ref->mutex, "i915_active", mkey);
 300         __i915_active_fence_init(&ref->excl, NULL, excl_retire);
 301         INIT_WORK(&ref->work, active_work);
 302 #if IS_ENABLED(CONFIG_LOCKDEP)
 303         lockdep_init_map(&ref->work.lockdep_map, "i915_active.work", wkey, 0);
 304 #endif
 305 }
 306
 307 static bool ____active_del_barrier(struct i915_active *ref,
 308                                    struct active_node *node,
 309                                    struct intel_engine_cs *engine)
 310
 311 {
 312         struct llist_node *head = NULL, *tail = NULL;
 313         struct llist_node *pos, *next;
 314
 315         GEM_BUG_ON(node->timeline != engine->kernel_context->timeline->fence_context);
 316
 317         /*
 318          * Rebuild the llist excluding our node. We may perform this
 319          * outside of the kernel_context timeline mutex and so someone
 320          * else may be manipulating the engine->barrier_tasks, in
 321          * which case either we or they will be upset :)
 322          *
 323          * A second __active_del_barrier() will report failure to claim
 324          * the active_node and the caller will just shrug and know not to
 325          * claim ownership of its node.
 326          *
 327          * A concurrent i915_request_add_active_barriers() will miss adding
 328          * any of the tasks, but we will try again on the next -- and since
 329          * we are actively using the barrier, we know that there will be
 330          * at least another opportunity when we idle.
 331          */
 332         llist_for_each_safe(pos, next, llist_del_all(&engine->barrier_tasks)) {
 333                 if (node == barrier_from_ll(pos)) {
 334                         node = NULL;
 335                         continue;
 336                 }
 337
 338                 pos->next = head;
 339                 head = pos;
 340                 if (!tail)
 341                         tail = pos;
 342         }
 343         if (head)
 344                 llist_add_batch(head, tail, &engine->barrier_tasks);
 345
 346         return !node;
 347 }
 348
 349 static bool
 350 __active_del_barrier(struct i915_active *ref, struct active_node *node)
 351 {
 352         return ____active_del_barrier(ref, node, barrier_to_engine(node));
 353 }
 354
 355 int i915_active_ref(struct i915_active *ref,
 356                     struct intel_timeline *tl,
 357                     struct dma_fence *fence)
 358 {
 359         struct i915_active_fence *active;
 360         int err;
 361
 362         lockdep_assert_held(&tl->mutex);
 363
 364         /* Prevent reaping in case we malloc/wait while building the tree */
 365         err = i915_active_acquire(ref);
 366         if (err)
 367                 return err;
 368
 369         active = active_instance(ref, tl);
 370         if (!active) {
 371                 err = -ENOMEM;
 372                 goto out;
 373         }
 374
 375         if (is_barrier(active)) { /* proto-node used by our idle barrier */
 376                 /*
 377                  * This request is on the kernel_context timeline, and so
 378                  * we can use it to substitute for the pending idle-barrer
 379                  * request that we want to emit on the kernel_context.
 380                  */
 381                 __active_del_barrier(ref, node_from_active(active));
 382                 RCU_INIT_POINTER(active->fence, NULL);
 383                 atomic_dec(&ref->count);
 384         }
 385         if (!__i915_active_fence_set(active, fence))
 386                 atomic_inc(&ref->count);
 387
 388 out:
 389         i915_active_release(ref);
 390         return err;
 391 }
 392
 393 struct dma_fence *
 394 i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
 395 {
 396         struct dma_fence *prev;
 397
 398         /* We expect the caller to manage the exclusive timeline ordering */
 399         GEM_BUG_ON(i915_active_is_idle(ref));
 400
 401         rcu_read_lock();
 402         prev = __i915_active_fence_set(&ref->excl, f);
 403         if (prev)
 404                 prev = dma_fence_get_rcu(prev);
 405         else
 406                 atomic_inc(&ref->count);
 407         rcu_read_unlock();
 408
 409         return prev;
 410 }
 411
 412 bool i915_active_acquire_if_busy(struct i915_active *ref)
 413 {
 414         debug_active_assert(ref);
 415         return atomic_add_unless(&ref->count, 1, 0);
 416 }
 417
 418 int i915_active_acquire(struct i915_active *ref)
 419 {
 420         int err;
 421
 422         if (i915_active_acquire_if_busy(ref))
 423                 return 0;
 424
 425         err = mutex_lock_interruptible(&ref->mutex);
 426         if (err)
 427                 return err;
 428
 429         if (likely(!i915_active_acquire_if_busy(ref))) {
 430                 if (ref->active)
 431                         err = ref->active(ref);
 432                 if (!err) {
 433                         spin_lock_irq(&ref->tree_lock); /* __active_retire() */
 434                         debug_active_activate(ref);
 435                         atomic_inc(&ref->count);
 436                         spin_unlock_irq(&ref->tree_lock);
 437                 }
 438         }
 439
 440         mutex_unlock(&ref->mutex);
 441
 442         return err;
 443 }
 444
 445 void i915_active_release(struct i915_active *ref)
 446 {
 447         debug_active_assert(ref);
 448         active_retire(ref);
 449 }
 450
 451 static void enable_signaling(struct i915_active_fence *active)
 452 {
 453         struct dma_fence *fence;
 454
 455         fence = i915_active_fence_get(active);
 456         if (!fence)
 457                 return;
 458
 459         dma_fence_enable_sw_signaling(fence);
 460         dma_fence_put(fence);
 461 }
 462
 463 int i915_active_wait(struct i915_active *ref)
 464 {
 465         struct active_node *it, *n;
 466         int err = 0;
 467
 468         might_sleep();
 469
 470         if (!i915_active_acquire_if_busy(ref))
 471                 return 0;
 472
 473         /* Flush lazy signals */
 474         enable_signaling(&ref->excl);
 475         rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
 476                 if (is_barrier(&it->base)) /* unconnected idle barrier */
 477                         continue;
 478
 479                 enable_signaling(&it->base);
 480         }
 481         /* Any fence added after the wait begins will not be auto-signaled */
 482
 483         i915_active_release(ref);
 484         if (err)
 485                 return err;
 486
 487         if (wait_var_event_interruptible(ref, i915_active_is_idle(ref)))
 488                 return -EINTR;
 489
 490         flush_work(&ref->work);
 491         return 0;
 492 }
 493
 494 int i915_request_await_active(struct i915_request *rq, struct i915_active *ref)
 495 {
 496         int err = 0;
 497
 498         if (rcu_access_pointer(ref->excl.fence)) {
 499                 struct dma_fence *fence;
 500
 501                 rcu_read_lock();
 502                 fence = dma_fence_get_rcu_safe(&ref->excl.fence);
 503                 rcu_read_unlock();
 504                 if (fence) {
 505                         err = i915_request_await_dma_fence(rq, fence);
 506                         dma_fence_put(fence);
 507                 }
 508         }
 509
 510         /* In the future we may choose to await on all fences */
 511
 512         return err;
 513 }
 514
 515 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
 516 void i915_active_fini(struct i915_active *ref)
 517 {
 518         debug_active_fini(ref);
 519         GEM_BUG_ON(atomic_read(&ref->count));
 520         GEM_BUG_ON(work_pending(&ref->work));
 521         GEM_BUG_ON(!RB_EMPTY_ROOT(&ref->tree));
 522         mutex_destroy(&ref->mutex);
 523 }
 524 #endif
 525
 526 static inline bool is_idle_barrier(struct active_node *node, u64 idx)
 527 {
 528         return node->timeline == idx && !i915_active_fence_isset(&node->base);
 529 }
 530
 531 static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx)
 532 {
 533         struct rb_node *prev, *p;
 534
 535         if (RB_EMPTY_ROOT(&ref->tree))
 536                 return NULL;
 537
 538         spin_lock_irq(&ref->tree_lock);
 539         GEM_BUG_ON(i915_active_is_idle(ref));
 540
 541         /*
 542          * Try to reuse any existing barrier nodes already allocated for this
 543          * i915_active, due to overlapping active phases there is likely a
 544          * node kept alive (as we reuse before parking). We prefer to reuse
 545          * completely idle barriers (less hassle in manipulating the llists),
 546          * but otherwise any will do.
 547          */
 548         if (ref->cache && is_idle_barrier(ref->cache, idx)) {
 549                 p = &ref->cache->node;
 550                 goto match;
 551         }
 552
 553         prev = NULL;
 554         p = ref->tree.rb_node;
 555         while (p) {
 556                 struct active_node *node =
 557                         rb_entry(p, struct active_node, node);
 558
 559                 if (is_idle_barrier(node, idx))
 560                         goto match;
 561
 562                 prev = p;
 563                 if (node->timeline < idx)
 564                         p = p->rb_right;
 565                 else
 566                         p = p->rb_left;
 567         }
 568
 569         /*
 570          * No quick match, but we did find the leftmost rb_node for the
 571          * kernel_context. Walk the rb_tree in-order to see if there were
 572          * any idle-barriers on this timeline that we missed, or just use
 573          * the first pending barrier.
 574          */
 575         for (p = prev; p; p = rb_next(p)) {
 576                 struct active_node *node =
 577                         rb_entry(p, struct active_node, node);
 578                 struct intel_engine_cs *engine;
 579
 580                 if (node->timeline > idx)
 581                         break;
 582
 583                 if (node->timeline < idx)
 584                         continue;
 585
 586                 if (is_idle_barrier(node, idx))
 587                         goto match;
 588
 589                 /*
 590                  * The list of pending barriers is protected by the
 591                  * kernel_context timeline, which notably we do not hold
 592                  * here. i915_request_add_active_barriers() may consume
 593                  * the barrier before we claim it, so we have to check
 594                  * for success.
 595                  */
 596                 engine = __barrier_to_engine(node);
 597                 smp_rmb(); /* serialise with add_active_barriers */
 598                 if (is_barrier(&node->base) &&
 599                     ____active_del_barrier(ref, node, engine))
 600                         goto match;
 601         }
 602
 603         spin_unlock_irq(&ref->tree_lock);
 604
 605         return NULL;
 606
 607 match:
 608         rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */
 609         if (p == &ref->cache->node)
 610                 ref->cache = NULL;
 611         spin_unlock_irq(&ref->tree_lock);
 612
 613         return rb_entry(p, struct active_node, node);
 614 }
 615
 616 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
 617                                             struct intel_engine_cs *engine)
 618 {
 619         intel_engine_mask_t tmp, mask = engine->mask;
 620         struct llist_node *first = NULL, *last = NULL;
 621         struct intel_gt *gt = engine->gt;
 622         int err;
 623
 624         GEM_BUG_ON(i915_active_is_idle(ref));
 625
 626         /* Wait until the previous preallocation is completed */
 627         while (!llist_empty(&ref->preallocated_barriers))
 628                 cond_resched();
 629
 630         /*
 631          * Preallocate a node for each physical engine supporting the target
 632          * engine (remember virtual engines have more than one sibling).
 633          * We can then use the preallocated nodes in
 634          * i915_active_acquire_barrier()
 635          */
 636         GEM_BUG_ON(!mask);
 637         for_each_engine_masked(engine, gt, mask, tmp) {
 638                 u64 idx = engine->kernel_context->timeline->fence_context;
 639                 struct llist_node *prev = first;
 640                 struct active_node *node;
 641
 642                 node = reuse_idle_barrier(ref, idx);
 643                 if (!node) {
 644                         node = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
 645                         if (!node) {
 646                                 err = ENOMEM;
 647                                 goto unwind;
 648                         }
 649
 650                         RCU_INIT_POINTER(node->base.fence, NULL);
 651                         node->base.cb.func = node_retire;
 652                         node->timeline = idx;
 653                         node->ref = ref;
 654                 }
 655
 656                 if (!i915_active_fence_isset(&node->base)) {
 657                         /*
 658                          * Mark this as being *our* unconnected proto-node.
 659                          *
 660                          * Since this node is not in any list, and we have
 661                          * decoupled it from the rbtree, we can reuse the
 662                          * request to indicate this is an idle-barrier node
 663                          * and then we can use the rb_node and list pointers
 664                          * for our tracking of the pending barrier.
 665                          */
 666                         RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
 667                         node->base.cb.node.prev = (void *)engine;
 668                         atomic_inc(&ref->count);
 669                 }
 670                 GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
 671
 672                 GEM_BUG_ON(barrier_to_engine(node) != engine);
 673                 first = barrier_to_ll(node);
 674                 first->next = prev;
 675                 if (!last)
 676                         last = first;
 677                 intel_engine_pm_get(engine);
 678         }
 679
 680         GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
 681         llist_add_batch(first, last, &ref->preallocated_barriers);
 682
 683         return 0;
 684
 685 unwind:
 686         while (first) {
 687                 struct active_node *node = barrier_from_ll(first);
 688
 689                 first = first->next;
 690
 691                 atomic_dec(&ref->count);
 692                 intel_engine_pm_put(barrier_to_engine(node));
 693
 694                 kmem_cache_free(global.slab_cache, node);
 695         }
 696         return err;
 697 }
 698
 699 void i915_active_acquire_barrier(struct i915_active *ref)
 700 {
 701         struct llist_node *pos, *next;
 702         unsigned long flags;
 703
 704         GEM_BUG_ON(i915_active_is_idle(ref));
 705
 706         /*
 707          * Transfer the list of preallocated barriers into the
 708          * i915_active rbtree, but only as proto-nodes. They will be
 709          * populated by i915_request_add_active_barriers() to point to the
 710          * request that will eventually release them.
 711          */
 712         llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
 713                 struct active_node *node = barrier_from_ll(pos);
 714                 struct intel_engine_cs *engine = barrier_to_engine(node);
 715                 struct rb_node **p, *parent;
 716
 717                 spin_lock_irqsave_nested(&ref->tree_lock, flags,
 718                                          SINGLE_DEPTH_NESTING);
 719                 parent = NULL;
 720                 p = &ref->tree.rb_node;
 721                 while (*p) {
 722                         struct active_node *it;
 723
 724                         parent = *p;
 725
 726                         it = rb_entry(parent, struct active_node, node);
 727                         if (it->timeline < node->timeline)
 728                                 p = &parent->rb_right;
 729                         else
 730                                 p = &parent->rb_left;
 731                 }
 732                 rb_link_node(&node->node, parent, p);
 733                 rb_insert_color(&node->node, &ref->tree);
 734                 spin_unlock_irqrestore(&ref->tree_lock, flags);
 735
 736                 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
 737                 llist_add(barrier_to_ll(node), &engine->barrier_tasks);
 738                 intel_engine_pm_put(engine);
 739         }
 740 }
 741
 742 static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
 743 {
 744         return __active_fence_slot(&barrier_from_ll(node)->base);
 745 }
 746
 747 void i915_request_add_active_barriers(struct i915_request *rq)
 748 {
 749         struct intel_engine_cs *engine = rq->engine;
 750         struct llist_node *node, *next;
 751         unsigned long flags;
 752
 753         GEM_BUG_ON(!intel_context_is_barrier(rq->context));
 754         GEM_BUG_ON(intel_engine_is_virtual(engine));
 755         GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline);
 756
 757         node = llist_del_all(&engine->barrier_tasks);
 758         if (!node)
 759                 return;
 760         /*
 761          * Attach the list of proto-fences to the in-flight request such
 762          * that the parent i915_active will be released when this request
 763          * is retired.
 764          */
 765         spin_lock_irqsave(&rq->lock, flags);
 766         llist_for_each_safe(node, next, node) {
 767                 /* serialise with reuse_idle_barrier */
 768                 smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
 769                 list_add_tail((struct list_head *)node, &rq->fence.cb_list);
 770         }
 771         spin_unlock_irqrestore(&rq->lock, flags);
 772 }
 773
 774 /*
 775  * __i915_active_fence_set: Update the last active fence along its timeline
 776  * @active: the active tracker
 777  * @fence: the new fence (under construction)
 778  *
 779  * Records the new @fence as the last active fence along its timeline in
 780  * this active tracker, moving the tracking callbacks from the previous
 781  * fence onto this one. Returns the previous fence (if not already completed),
 782  * which the caller must ensure is executed before the new fence. To ensure
 783  * that the order of fences within the timeline of the i915_active_fence is
 784  * understood, it should be locked by the caller.
 785  */
 786 struct dma_fence *
 787 __i915_active_fence_set(struct i915_active_fence *active,
 788                         struct dma_fence *fence)
 789 {
 790         struct dma_fence *prev;
 791         unsigned long flags;
 792
 793         if (fence == rcu_access_pointer(active->fence))
 794                 return fence;
 795
 796         GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
 797
 798         /*
 799          * Consider that we have two threads arriving (A and B), with
 800          * C already resident as the active->fence.
 801          *
 802          * A does the xchg first, and so it sees C or NULL depending
 803          * on the timing of the interrupt handler. If it is NULL, the
 804          * previous fence must have been signaled and we know that
 805          * we are first on the timeline. If it is still present,
 806          * we acquire the lock on that fence and serialise with the interrupt
 807          * handler, in the process removing it from any future interrupt
 808          * callback. A will then wait on C before executing (if present).
 809          *
 810          * As B is second, it sees A as the previous fence and so waits for
 811          * it to complete its transition and takes over the occupancy for
 812          * itself -- remembering that it needs to wait on A before executing.
 813          *
 814          * Note the strong ordering of the timeline also provides consistent
 815          * nesting rules for the fence->lock; the inner lock is always the
 816          * older lock.
 817          */
 818         spin_lock_irqsave(fence->lock, flags);
 819         prev = xchg(__active_fence_slot(active), fence);
 820         if (prev) {
 821                 GEM_BUG_ON(prev == fence);
 822                 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
 823                 __list_del_entry(&active->cb.node);
 824                 spin_unlock(prev->lock); /* serialise with prev->cb_list */
 825         }
 826         list_add_tail(&active->cb.node, &fence->cb_list);
 827         spin_unlock_irqrestore(fence->lock, flags);
 828
 829         return prev;
 830 }
 831
 832 int i915_active_fence_set(struct i915_active_fence *active,
 833                           struct i915_request *rq)
 834 {
 835         struct dma_fence *fence;
 836         int err = 0;
 837
 838         /* Must maintain timeline ordering wrt previous active requests */
 839         rcu_read_lock();
 840         fence = __i915_active_fence_set(active, &rq->fence);
 841         if (fence) /* but the previous fence may not belong to that timeline! */
 842                 fence = dma_fence_get_rcu(fence);
 843         rcu_read_unlock();
 844         if (fence) {
 845                 err = i915_request_await_dma_fence(rq, fence);
 846                 dma_fence_put(fence);
 847         }
 848
 849         return err;
 850 }
 851
 852 void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb)
 853 {
 854         active_fence_cb(fence, cb);
 855 }
 856
 857 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
 858 #include "selftests/i915_active.c"
 859 #endif
 860
 861 static void i915_global_active_shrink(void)
 862 {
 863         kmem_cache_shrink(global.slab_cache);
 864 }
 865
 866 static void i915_global_active_exit(void)
 867 {
 868         kmem_cache_destroy(global.slab_cache);
 869 }
 870
 871 static struct i915_global_active global = { {
 872         .shrink = i915_global_active_shrink,
 873         .exit = i915_global_active_exit,
 874 } };
 875
 876 int __init i915_global_active_init(void)
 877 {
 878         global.slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN);
 879         if (!global.slab_cache)
 880                 return -ENOMEM;
 881
 882         i915_global_register(&global.base);
 883         return 0;
 884 }