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 void i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
 394 {
 395         /* We expect the caller to manage the exclusive timeline ordering */
 396         GEM_BUG_ON(i915_active_is_idle(ref));
 397
 398         if (!__i915_active_fence_set(&ref->excl, f))
 399                 atomic_inc(&ref->count);
 400 }
 401
 402 bool i915_active_acquire_if_busy(struct i915_active *ref)
 403 {
 404         debug_active_assert(ref);
 405         return atomic_add_unless(&ref->count, 1, 0);
 406 }
 407
 408 int i915_active_acquire(struct i915_active *ref)
 409 {
 410         int err;
 411
 412         if (i915_active_acquire_if_busy(ref))
 413                 return 0;
 414
 415         err = mutex_lock_interruptible(&ref->mutex);
 416         if (err)
 417                 return err;
 418
 419         if (likely(!i915_active_acquire_if_busy(ref))) {
 420                 if (ref->active)
 421                         err = ref->active(ref);
 422                 if (!err) {
 423                         spin_lock_irq(&ref->tree_lock); /* __active_retire() */
 424                         debug_active_activate(ref);
 425                         atomic_inc(&ref->count);
 426                         spin_unlock_irq(&ref->tree_lock);
 427                 }
 428         }
 429
 430         mutex_unlock(&ref->mutex);
 431
 432         return err;
 433 }
 434
 435 void i915_active_release(struct i915_active *ref)
 436 {
 437         debug_active_assert(ref);
 438         active_retire(ref);
 439 }
 440
 441 static void enable_signaling(struct i915_active_fence *active)
 442 {
 443         struct dma_fence *fence;
 444
 445         fence = i915_active_fence_get(active);
 446         if (!fence)
 447                 return;
 448
 449         dma_fence_enable_sw_signaling(fence);
 450         dma_fence_put(fence);
 451 }
 452
 453 int i915_active_wait(struct i915_active *ref)
 454 {
 455         struct active_node *it, *n;
 456         int err = 0;
 457
 458         might_sleep();
 459
 460         if (!i915_active_acquire_if_busy(ref))
 461                 return 0;
 462
 463         /* Flush lazy signals */
 464         enable_signaling(&ref->excl);
 465         rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
 466                 if (is_barrier(&it->base)) /* unconnected idle barrier */
 467                         continue;
 468
 469                 enable_signaling(&it->base);
 470         }
 471         /* Any fence added after the wait begins will not be auto-signaled */
 472
 473         i915_active_release(ref);
 474         if (err)
 475                 return err;
 476
 477         if (wait_var_event_interruptible(ref, i915_active_is_idle(ref)))
 478                 return -EINTR;
 479
 480         flush_work(&ref->work);
 481         return 0;
 482 }
 483
 484 int i915_request_await_active(struct i915_request *rq, struct i915_active *ref)
 485 {
 486         int err = 0;
 487
 488         if (rcu_access_pointer(ref->excl.fence)) {
 489                 struct dma_fence *fence;
 490
 491                 rcu_read_lock();
 492                 fence = dma_fence_get_rcu_safe(&ref->excl.fence);
 493                 rcu_read_unlock();
 494                 if (fence) {
 495                         err = i915_request_await_dma_fence(rq, fence);
 496                         dma_fence_put(fence);
 497                 }
 498         }
 499
 500         /* In the future we may choose to await on all fences */
 501
 502         return err;
 503 }
 504
 505 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
 506 void i915_active_fini(struct i915_active *ref)
 507 {
 508         debug_active_fini(ref);
 509         GEM_BUG_ON(atomic_read(&ref->count));
 510         GEM_BUG_ON(work_pending(&ref->work));
 511         GEM_BUG_ON(!RB_EMPTY_ROOT(&ref->tree));
 512         mutex_destroy(&ref->mutex);
 513 }
 514 #endif
 515
 516 static inline bool is_idle_barrier(struct active_node *node, u64 idx)
 517 {
 518         return node->timeline == idx && !i915_active_fence_isset(&node->base);
 519 }
 520
 521 static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx)
 522 {
 523         struct rb_node *prev, *p;
 524
 525         if (RB_EMPTY_ROOT(&ref->tree))
 526                 return NULL;
 527
 528         spin_lock_irq(&ref->tree_lock);
 529         GEM_BUG_ON(i915_active_is_idle(ref));
 530
 531         /*
 532          * Try to reuse any existing barrier nodes already allocated for this
 533          * i915_active, due to overlapping active phases there is likely a
 534          * node kept alive (as we reuse before parking). We prefer to reuse
 535          * completely idle barriers (less hassle in manipulating the llists),
 536          * but otherwise any will do.
 537          */
 538         if (ref->cache && is_idle_barrier(ref->cache, idx)) {
 539                 p = &ref->cache->node;
 540                 goto match;
 541         }
 542
 543         prev = NULL;
 544         p = ref->tree.rb_node;
 545         while (p) {
 546                 struct active_node *node =
 547                         rb_entry(p, struct active_node, node);
 548
 549                 if (is_idle_barrier(node, idx))
 550                         goto match;
 551
 552                 prev = p;
 553                 if (node->timeline < idx)
 554                         p = p->rb_right;
 555                 else
 556                         p = p->rb_left;
 557         }
 558
 559         /*
 560          * No quick match, but we did find the leftmost rb_node for the
 561          * kernel_context. Walk the rb_tree in-order to see if there were
 562          * any idle-barriers on this timeline that we missed, or just use
 563          * the first pending barrier.
 564          */
 565         for (p = prev; p; p = rb_next(p)) {
 566                 struct active_node *node =
 567                         rb_entry(p, struct active_node, node);
 568                 struct intel_engine_cs *engine;
 569
 570                 if (node->timeline > idx)
 571                         break;
 572
 573                 if (node->timeline < idx)
 574                         continue;
 575
 576                 if (is_idle_barrier(node, idx))
 577                         goto match;
 578
 579                 /*
 580                  * The list of pending barriers is protected by the
 581                  * kernel_context timeline, which notably we do not hold
 582                  * here. i915_request_add_active_barriers() may consume
 583                  * the barrier before we claim it, so we have to check
 584                  * for success.
 585                  */
 586                 engine = __barrier_to_engine(node);
 587                 smp_rmb(); /* serialise with add_active_barriers */
 588                 if (is_barrier(&node->base) &&
 589                     ____active_del_barrier(ref, node, engine))
 590                         goto match;
 591         }
 592
 593         spin_unlock_irq(&ref->tree_lock);
 594
 595         return NULL;
 596
 597 match:
 598         rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */
 599         if (p == &ref->cache->node)
 600                 ref->cache = NULL;
 601         spin_unlock_irq(&ref->tree_lock);
 602
 603         return rb_entry(p, struct active_node, node);
 604 }
 605
 606 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
 607                                             struct intel_engine_cs *engine)
 608 {
 609         intel_engine_mask_t tmp, mask = engine->mask;
 610         struct llist_node *first = NULL, *last = NULL;
 611         struct intel_gt *gt = engine->gt;
 612         int err;
 613
 614         GEM_BUG_ON(i915_active_is_idle(ref));
 615
 616         /* Wait until the previous preallocation is completed */
 617         while (!llist_empty(&ref->preallocated_barriers))
 618                 cond_resched();
 619
 620         /*
 621          * Preallocate a node for each physical engine supporting the target
 622          * engine (remember virtual engines have more than one sibling).
 623          * We can then use the preallocated nodes in
 624          * i915_active_acquire_barrier()
 625          */
 626         for_each_engine_masked(engine, gt, mask, tmp) {
 627                 u64 idx = engine->kernel_context->timeline->fence_context;
 628                 struct llist_node *prev = first;
 629                 struct active_node *node;
 630
 631                 node = reuse_idle_barrier(ref, idx);
 632                 if (!node) {
 633                         node = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
 634                         if (!node) {
 635                                 err = ENOMEM;
 636                                 goto unwind;
 637                         }
 638
 639                         RCU_INIT_POINTER(node->base.fence, NULL);
 640                         node->base.cb.func = node_retire;
 641                         node->timeline = idx;
 642                         node->ref = ref;
 643                 }
 644
 645                 if (!i915_active_fence_isset(&node->base)) {
 646                         /*
 647                          * Mark this as being *our* unconnected proto-node.
 648                          *
 649                          * Since this node is not in any list, and we have
 650                          * decoupled it from the rbtree, we can reuse the
 651                          * request to indicate this is an idle-barrier node
 652                          * and then we can use the rb_node and list pointers
 653                          * for our tracking of the pending barrier.
 654                          */
 655                         RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
 656                         node->base.cb.node.prev = (void *)engine;
 657                         atomic_inc(&ref->count);
 658                 }
 659                 GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
 660
 661                 GEM_BUG_ON(barrier_to_engine(node) != engine);
 662                 first = barrier_to_ll(node);
 663                 first->next = prev;
 664                 if (!last)
 665                         last = first;
 666                 intel_engine_pm_get(engine);
 667         }
 668
 669         GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
 670         llist_add_batch(first, last, &ref->preallocated_barriers);
 671
 672         return 0;
 673
 674 unwind:
 675         while (first) {
 676                 struct active_node *node = barrier_from_ll(first);
 677
 678                 first = first->next;
 679
 680                 atomic_dec(&ref->count);
 681                 intel_engine_pm_put(barrier_to_engine(node));
 682
 683                 kmem_cache_free(global.slab_cache, node);
 684         }
 685         return err;
 686 }
 687
 688 void i915_active_acquire_barrier(struct i915_active *ref)
 689 {
 690         struct llist_node *pos, *next;
 691         unsigned long flags;
 692
 693         GEM_BUG_ON(i915_active_is_idle(ref));
 694
 695         /*
 696          * Transfer the list of preallocated barriers into the
 697          * i915_active rbtree, but only as proto-nodes. They will be
 698          * populated by i915_request_add_active_barriers() to point to the
 699          * request that will eventually release them.
 700          */
 701         llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
 702                 struct active_node *node = barrier_from_ll(pos);
 703                 struct intel_engine_cs *engine = barrier_to_engine(node);
 704                 struct rb_node **p, *parent;
 705
 706                 spin_lock_irqsave_nested(&ref->tree_lock, flags,
 707                                          SINGLE_DEPTH_NESTING);
 708                 parent = NULL;
 709                 p = &ref->tree.rb_node;
 710                 while (*p) {
 711                         struct active_node *it;
 712
 713                         parent = *p;
 714
 715                         it = rb_entry(parent, struct active_node, node);
 716                         if (it->timeline < node->timeline)
 717                                 p = &parent->rb_right;
 718                         else
 719                                 p = &parent->rb_left;
 720                 }
 721                 rb_link_node(&node->node, parent, p);
 722                 rb_insert_color(&node->node, &ref->tree);
 723                 spin_unlock_irqrestore(&ref->tree_lock, flags);
 724
 725                 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
 726                 llist_add(barrier_to_ll(node), &engine->barrier_tasks);
 727                 intel_engine_pm_put(engine);
 728         }
 729 }
 730
 731 static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
 732 {
 733         return __active_fence_slot(&barrier_from_ll(node)->base);
 734 }
 735
 736 void i915_request_add_active_barriers(struct i915_request *rq)
 737 {
 738         struct intel_engine_cs *engine = rq->engine;
 739         struct llist_node *node, *next;
 740         unsigned long flags;
 741
 742         GEM_BUG_ON(!intel_context_is_barrier(rq->context));
 743         GEM_BUG_ON(intel_engine_is_virtual(engine));
 744         GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline);
 745
 746         node = llist_del_all(&engine->barrier_tasks);
 747         if (!node)
 748                 return;
 749         /*
 750          * Attach the list of proto-fences to the in-flight request such
 751          * that the parent i915_active will be released when this request
 752          * is retired.
 753          */
 754         spin_lock_irqsave(&rq->lock, flags);
 755         llist_for_each_safe(node, next, node) {
 756                 /* serialise with reuse_idle_barrier */
 757                 smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
 758                 list_add_tail((struct list_head *)node, &rq->fence.cb_list);
 759         }
 760         spin_unlock_irqrestore(&rq->lock, flags);
 761 }
 762
 763 /*
 764  * __i915_active_fence_set: Update the last active fence along its timeline
 765  * @active: the active tracker
 766  * @fence: the new fence (under construction)
 767  *
 768  * Records the new @fence as the last active fence along its timeline in
 769  * this active tracker, moving the tracking callbacks from the previous
 770  * fence onto this one. Returns the previous fence (if not already completed),
 771  * which the caller must ensure is executed before the new fence. To ensure
 772  * that the order of fences within the timeline of the i915_active_fence is
 773  * understood, it should be locked by the caller.
 774  */
 775 struct dma_fence *
 776 __i915_active_fence_set(struct i915_active_fence *active,
 777                         struct dma_fence *fence)
 778 {
 779         struct dma_fence *prev;
 780         unsigned long flags;
 781
 782         if (fence == rcu_access_pointer(active->fence))
 783                 return fence;
 784
 785         GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
 786
 787         /*
 788          * Consider that we have two threads arriving (A and B), with
 789          * C already resident as the active->fence.
 790          *
 791          * A does the xchg first, and so it sees C or NULL depending
 792          * on the timing of the interrupt handler. If it is NULL, the
 793          * previous fence must have been signaled and we know that
 794          * we are first on the timeline. If it is still present,
 795          * we acquire the lock on that fence and serialise with the interrupt
 796          * handler, in the process removing it from any future interrupt
 797          * callback. A will then wait on C before executing (if present).
 798          *
 799          * As B is second, it sees A as the previous fence and so waits for
 800          * it to complete its transition and takes over the occupancy for
 801          * itself -- remembering that it needs to wait on A before executing.
 802          *
 803          * Note the strong ordering of the timeline also provides consistent
 804          * nesting rules for the fence->lock; the inner lock is always the
 805          * older lock.
 806          */
 807         spin_lock_irqsave(fence->lock, flags);
 808         prev = xchg(__active_fence_slot(active), fence);
 809         if (prev) {
 810                 GEM_BUG_ON(prev == fence);
 811                 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
 812                 __list_del_entry(&active->cb.node);
 813                 spin_unlock(prev->lock); /* serialise with prev->cb_list */
 814         }
 815         GEM_BUG_ON(rcu_access_pointer(active->fence) != fence);
 816         list_add_tail(&active->cb.node, &fence->cb_list);
 817         spin_unlock_irqrestore(fence->lock, flags);
 818
 819         return prev;
 820 }
 821
 822 int i915_active_fence_set(struct i915_active_fence *active,
 823                           struct i915_request *rq)
 824 {
 825         struct dma_fence *fence;
 826         int err = 0;
 827
 828         /* Must maintain timeline ordering wrt previous active requests */
 829         rcu_read_lock();
 830         fence = __i915_active_fence_set(active, &rq->fence);
 831         if (fence) /* but the previous fence may not belong to that timeline! */
 832                 fence = dma_fence_get_rcu(fence);
 833         rcu_read_unlock();
 834         if (fence) {
 835                 err = i915_request_await_dma_fence(rq, fence);
 836                 dma_fence_put(fence);
 837         }
 838
 839         return err;
 840 }
 841
 842 void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb)
 843 {
 844         active_fence_cb(fence, cb);
 845 }
 846
 847 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
 848 #include "selftests/i915_active.c"
 849 #endif
 850
 851 static void i915_global_active_shrink(void)
 852 {
 853         kmem_cache_shrink(global.slab_cache);
 854 }
 855
 856 static void i915_global_active_exit(void)
 857 {
 858         kmem_cache_destroy(global.slab_cache);
 859 }
 860
 861 static struct i915_global_active global = { {
 862         .shrink = i915_global_active_shrink,
 863         .exit = i915_global_active_exit,
 864 } };
 865
 866 int __init i915_global_active_init(void)
 867 {
 868         global.slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN);
 869         if (!global.slab_cache)
 870                 return -ENOMEM;
 871
 872         i915_global_register(&global.base);
 873         return 0;
 874 }