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 i915_active_fence base;
  32         struct i915_active *ref;
  33         struct rb_node node;
  34         u64 timeline;
  35 };
  36
  37 static inline struct active_node *
  38 node_from_active(struct i915_active_fence *active)
  39 {
  40         return container_of(active, struct active_node, base);
  41 }
  42
  43 #define take_preallocated_barriers(x) llist_del_all(&(x)->preallocated_barriers)
  44
  45 static inline bool is_barrier(const struct i915_active_fence *active)
  46 {
  47         return IS_ERR(rcu_access_pointer(active->fence));
  48 }
  49
  50 static inline struct llist_node *barrier_to_ll(struct active_node *node)
  51 {
  52         GEM_BUG_ON(!is_barrier(&node->base));
  53         return (struct llist_node *)&node->base.cb.node;
  54 }
  55
  56 static inline struct intel_engine_cs *
  57 __barrier_to_engine(struct active_node *node)
  58 {
  59         return (struct intel_engine_cs *)READ_ONCE(node->base.cb.node.prev);
  60 }
  61
  62 static inline struct intel_engine_cs *
  63 barrier_to_engine(struct active_node *node)
  64 {
  65         GEM_BUG_ON(!is_barrier(&node->base));
  66         return __barrier_to_engine(node);
  67 }
  68
  69 static inline struct active_node *barrier_from_ll(struct llist_node *x)
  70 {
  71         return container_of((struct list_head *)x,
  72                             struct active_node, base.cb.node);
  73 }
  74
  75 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) && IS_ENABLED(CONFIG_DEBUG_OBJECTS)
  76
  77 static void *active_debug_hint(void *addr)
  78 {
  79         struct i915_active *ref = addr;
  80
  81         return (void *)ref->active ?: (void *)ref->retire ?: (void *)ref;
  82 }
  83
  84 static struct debug_obj_descr active_debug_desc = {
  85         .name = "i915_active",
  86         .debug_hint = active_debug_hint,
  87 };
  88
  89 static void debug_active_init(struct i915_active *ref)
  90 {
  91         debug_object_init(ref, &active_debug_desc);
  92 }
  93
  94 static void debug_active_activate(struct i915_active *ref)
  95 {
  96         lockdep_assert_held(&ref->tree_lock);
  97         if (!atomic_read(&ref->count)) /* before the first inc */
  98                 debug_object_activate(ref, &active_debug_desc);
  99 }
 100
 101 static void debug_active_deactivate(struct i915_active *ref)
 102 {
 103         lockdep_assert_held(&ref->tree_lock);
 104         if (!atomic_read(&ref->count)) /* after the last dec */
 105                 debug_object_deactivate(ref, &active_debug_desc);
 106 }
 107
 108 static void debug_active_fini(struct i915_active *ref)
 109 {
 110         debug_object_free(ref, &active_debug_desc);
 111 }
 112
 113 static void debug_active_assert(struct i915_active *ref)
 114 {
 115         debug_object_assert_init(ref, &active_debug_desc);
 116 }
 117
 118 #else
 119
 120 static inline void debug_active_init(struct i915_active *ref) { }
 121 static inline void debug_active_activate(struct i915_active *ref) { }
 122 static inline void debug_active_deactivate(struct i915_active *ref) { }
 123 static inline void debug_active_fini(struct i915_active *ref) { }
 124 static inline void debug_active_assert(struct i915_active *ref) { }
 125
 126 #endif
 127
 128 static void
 129 __active_retire(struct i915_active *ref)
 130 {
 131         struct active_node *it, *n;
 132         struct rb_root root;
 133         unsigned long flags;
 134
 135         GEM_BUG_ON(i915_active_is_idle(ref));
 136
 137         /* return the unused nodes to our slabcache -- flushing the allocator */
 138         if (!atomic_dec_and_lock_irqsave(&ref->count, &ref->tree_lock, flags))
 139                 return;
 140
 141         GEM_BUG_ON(rcu_access_pointer(ref->excl.fence));
 142         debug_active_deactivate(ref);
 143
 144         root = ref->tree;
 145         ref->tree = RB_ROOT;
 146         ref->cache = NULL;
 147
 148         spin_unlock_irqrestore(&ref->tree_lock, flags);
 149
 150         /* After the final retire, the entire struct may be freed */
 151         if (ref->retire)
 152                 ref->retire(ref);
 153
 154         /* ... except if you wait on it, you must manage your own references! */
 155         wake_up_var(ref);
 156
 157         rbtree_postorder_for_each_entry_safe(it, n, &root, node) {
 158                 GEM_BUG_ON(i915_active_fence_isset(&it->base));
 159                 kmem_cache_free(global.slab_cache, it);
 160         }
 161 }
 162
 163 static void
 164 active_work(struct work_struct *wrk)
 165 {
 166         struct i915_active *ref = container_of(wrk, typeof(*ref), work);
 167
 168         GEM_BUG_ON(!atomic_read(&ref->count));
 169         if (atomic_add_unless(&ref->count, -1, 1))
 170                 return;
 171
 172         __active_retire(ref);
 173 }
 174
 175 static void
 176 active_retire(struct i915_active *ref)
 177 {
 178         GEM_BUG_ON(!atomic_read(&ref->count));
 179         if (atomic_add_unless(&ref->count, -1, 1))
 180                 return;
 181
 182         if (ref->flags & I915_ACTIVE_RETIRE_SLEEPS) {
 183                 queue_work(system_unbound_wq, &ref->work);
 184                 return;
 185         }
 186
 187         __active_retire(ref);
 188 }
 189
 190 static inline struct dma_fence **
 191 __active_fence_slot(struct i915_active_fence *active)
 192 {
 193         return (struct dma_fence ** __force)&active->fence;
 194 }
 195
 196 static inline bool
 197 active_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
 198 {
 199         struct i915_active_fence *active =
 200                 container_of(cb, typeof(*active), cb);
 201
 202         return cmpxchg(__active_fence_slot(active), fence, NULL) == fence;
 203 }
 204
 205 static void
 206 node_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
 207 {
 208         if (active_fence_cb(fence, cb))
 209                 active_retire(container_of(cb, struct active_node, base.cb)->ref);
 210 }
 211
 212 static void
 213 excl_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
 214 {
 215         if (active_fence_cb(fence, cb))
 216                 active_retire(container_of(cb, struct i915_active, excl.cb));
 217 }
 218
 219 static struct i915_active_fence *
 220 active_instance(struct i915_active *ref, struct intel_timeline *tl)
 221 {
 222         struct active_node *node, *prealloc;
 223         struct rb_node **p, *parent;
 224         u64 idx = tl->fence_context;
 225
 226         /*
 227          * We track the most recently used timeline to skip a rbtree search
 228          * for the common case, under typical loads we never need the rbtree
 229          * at all. We can reuse the last slot if it is empty, that is
 230          * after the previous activity has been retired, or if it matches the
 231          * current timeline.
 232          */
 233         node = READ_ONCE(ref->cache);
 234         if (node && node->timeline == idx)
 235                 return &node->base;
 236
 237         /* Preallocate a replacement, just in case */
 238         prealloc = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
 239         if (!prealloc)
 240                 return NULL;
 241
 242         spin_lock_irq(&ref->tree_lock);
 243         GEM_BUG_ON(i915_active_is_idle(ref));
 244
 245         parent = NULL;
 246         p = &ref->tree.rb_node;
 247         while (*p) {
 248                 parent = *p;
 249
 250                 node = rb_entry(parent, struct active_node, node);
 251                 if (node->timeline == idx) {
 252                         kmem_cache_free(global.slab_cache, prealloc);
 253                         goto out;
 254                 }
 255
 256                 if (node->timeline < idx)
 257                         p = &parent->rb_right;
 258                 else
 259                         p = &parent->rb_left;
 260         }
 261
 262         node = prealloc;
 263         __i915_active_fence_init(&node->base, NULL, node_retire);
 264         node->ref = ref;
 265         node->timeline = idx;
 266
 267         rb_link_node(&node->node, parent, p);
 268         rb_insert_color(&node->node, &ref->tree);
 269
 270 out:
 271         ref->cache = node;
 272         spin_unlock_irq(&ref->tree_lock);
 273
 274         BUILD_BUG_ON(offsetof(typeof(*node), base));
 275         return &node->base;
 276 }
 277
 278 void __i915_active_init(struct i915_active *ref,
 279                         int (*active)(struct i915_active *ref),
 280                         void (*retire)(struct i915_active *ref),
 281                         struct lock_class_key *mkey,
 282                         struct lock_class_key *wkey)
 283 {
 284         unsigned long bits;
 285
 286         debug_active_init(ref);
 287
 288         ref->flags = 0;
 289         ref->active = active;
 290         ref->retire = ptr_unpack_bits(retire, &bits, 2);
 291         if (bits & I915_ACTIVE_MAY_SLEEP)
 292                 ref->flags |= I915_ACTIVE_RETIRE_SLEEPS;
 293
 294         spin_lock_init(&ref->tree_lock);
 295         ref->tree = RB_ROOT;
 296         ref->cache = NULL;
 297
 298         init_llist_head(&ref->preallocated_barriers);
 299         atomic_set(&ref->count, 0);
 300         __mutex_init(&ref->mutex, "i915_active", mkey);
 301         __i915_active_fence_init(&ref->excl, NULL, excl_retire);
 302         INIT_WORK(&ref->work, active_work);
 303 #if IS_ENABLED(CONFIG_LOCKDEP)
 304         lockdep_init_map(&ref->work.lockdep_map, "i915_active.work", wkey, 0);
 305 #endif
 306 }
 307
 308 static bool ____active_del_barrier(struct i915_active *ref,
 309                                    struct active_node *node,
 310                                    struct intel_engine_cs *engine)
 311
 312 {
 313         struct llist_node *head = NULL, *tail = NULL;
 314         struct llist_node *pos, *next;
 315
 316         GEM_BUG_ON(node->timeline != engine->kernel_context->timeline->fence_context);
 317
 318         /*
 319          * Rebuild the llist excluding our node. We may perform this
 320          * outside of the kernel_context timeline mutex and so someone
 321          * else may be manipulating the engine->barrier_tasks, in
 322          * which case either we or they will be upset :)
 323          *
 324          * A second __active_del_barrier() will report failure to claim
 325          * the active_node and the caller will just shrug and know not to
 326          * claim ownership of its node.
 327          *
 328          * A concurrent i915_request_add_active_barriers() will miss adding
 329          * any of the tasks, but we will try again on the next -- and since
 330          * we are actively using the barrier, we know that there will be
 331          * at least another opportunity when we idle.
 332          */
 333         llist_for_each_safe(pos, next, llist_del_all(&engine->barrier_tasks)) {
 334                 if (node == barrier_from_ll(pos)) {
 335                         node = NULL;
 336                         continue;
 337                 }
 338
 339                 pos->next = head;
 340                 head = pos;
 341                 if (!tail)
 342                         tail = pos;
 343         }
 344         if (head)
 345                 llist_add_batch(head, tail, &engine->barrier_tasks);
 346
 347         return !node;
 348 }
 349
 350 static bool
 351 __active_del_barrier(struct i915_active *ref, struct active_node *node)
 352 {
 353         return ____active_del_barrier(ref, node, barrier_to_engine(node));
 354 }
 355
 356 int i915_active_ref(struct i915_active *ref,
 357                     struct intel_timeline *tl,
 358                     struct dma_fence *fence)
 359 {
 360         struct i915_active_fence *active;
 361         int err;
 362
 363         lockdep_assert_held(&tl->mutex);
 364
 365         /* Prevent reaping in case we malloc/wait while building the tree */
 366         err = i915_active_acquire(ref);
 367         if (err)
 368                 return err;
 369
 370         active = active_instance(ref, tl);
 371         if (!active) {
 372                 err = -ENOMEM;
 373                 goto out;
 374         }
 375
 376         if (is_barrier(active)) { /* proto-node used by our idle barrier */
 377                 /*
 378                  * This request is on the kernel_context timeline, and so
 379                  * we can use it to substitute for the pending idle-barrer
 380                  * request that we want to emit on the kernel_context.
 381                  */
 382                 __active_del_barrier(ref, node_from_active(active));
 383                 RCU_INIT_POINTER(active->fence, NULL);
 384                 atomic_dec(&ref->count);
 385         }
 386         if (!__i915_active_fence_set(active, fence))
 387                 atomic_inc(&ref->count);
 388
 389 out:
 390         i915_active_release(ref);
 391         return err;
 392 }
 393
 394 struct dma_fence *
 395 i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
 396 {
 397         struct dma_fence *prev;
 398
 399         /* We expect the caller to manage the exclusive timeline ordering */
 400         GEM_BUG_ON(i915_active_is_idle(ref));
 401
 402         rcu_read_lock();
 403         prev = __i915_active_fence_set(&ref->excl, f);
 404         if (prev)
 405                 prev = dma_fence_get_rcu(prev);
 406         else
 407                 atomic_inc(&ref->count);
 408         rcu_read_unlock();
 409
 410         return prev;
 411 }
 412
 413 bool i915_active_acquire_if_busy(struct i915_active *ref)
 414 {
 415         debug_active_assert(ref);
 416         return atomic_add_unless(&ref->count, 1, 0);
 417 }
 418
 419 int i915_active_acquire(struct i915_active *ref)
 420 {
 421         int err;
 422
 423         if (i915_active_acquire_if_busy(ref))
 424                 return 0;
 425
 426         err = mutex_lock_interruptible(&ref->mutex);
 427         if (err)
 428                 return err;
 429
 430         if (likely(!i915_active_acquire_if_busy(ref))) {
 431                 if (ref->active)
 432                         err = ref->active(ref);
 433                 if (!err) {
 434                         spin_lock_irq(&ref->tree_lock); /* __active_retire() */
 435                         debug_active_activate(ref);
 436                         atomic_inc(&ref->count);
 437                         spin_unlock_irq(&ref->tree_lock);
 438                 }
 439         }
 440
 441         mutex_unlock(&ref->mutex);
 442
 443         return err;
 444 }
 445
 446 void i915_active_release(struct i915_active *ref)
 447 {
 448         debug_active_assert(ref);
 449         active_retire(ref);
 450 }
 451
 452 static void enable_signaling(struct i915_active_fence *active)
 453 {
 454         struct dma_fence *fence;
 455
 456         if (unlikely(is_barrier(active)))
 457                 return;
 458
 459         fence = i915_active_fence_get(active);
 460         if (!fence)
 461                 return;
 462
 463         dma_fence_enable_sw_signaling(fence);
 464         dma_fence_put(fence);
 465 }
 466
 467 static int flush_barrier(struct active_node *it)
 468 {
 469         struct intel_engine_cs *engine;
 470
 471         if (likely(!is_barrier(&it->base)))
 472                 return 0;
 473
 474         engine = __barrier_to_engine(it);
 475         smp_rmb(); /* serialise with add_active_barriers */
 476         if (!is_barrier(&it->base))
 477                 return 0;
 478
 479         return intel_engine_flush_barriers(engine);
 480 }
 481
 482 static int flush_lazy_signals(struct i915_active *ref)
 483 {
 484         struct active_node *it, *n;
 485         int err = 0;
 486
 487         enable_signaling(&ref->excl);
 488         rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
 489                 err = flush_barrier(it); /* unconnected idle barrier? */
 490                 if (err)
 491                         break;
 492
 493                 enable_signaling(&it->base);
 494         }
 495
 496         return err;
 497 }
 498
 499 int i915_active_wait(struct i915_active *ref)
 500 {
 501         int err;
 502
 503         might_sleep();
 504
 505         if (!i915_active_acquire_if_busy(ref))
 506                 return 0;
 507
 508         /* Any fence added after the wait begins will not be auto-signaled */
 509         err = flush_lazy_signals(ref);
 510         i915_active_release(ref);
 511         if (err)
 512                 return err;
 513
 514         if (wait_var_event_interruptible(ref, i915_active_is_idle(ref)))
 515                 return -EINTR;
 516
 517         flush_work(&ref->work);
 518         return 0;
 519 }
 520
 521 static int __await_active(struct i915_active_fence *active,
 522                           int (*fn)(void *arg, struct dma_fence *fence),
 523                           void *arg)
 524 {
 525         struct dma_fence *fence;
 526
 527         if (is_barrier(active)) /* XXX flush the barrier? */
 528                 return 0;
 529
 530         fence = i915_active_fence_get(active);
 531         if (fence) {
 532                 int err;
 533
 534                 err = fn(arg, fence);
 535                 dma_fence_put(fence);
 536                 if (err < 0)
 537                         return err;
 538         }
 539
 540         return 0;
 541 }
 542
 543 static int await_active(struct i915_active *ref,
 544                         unsigned int flags,
 545                         int (*fn)(void *arg, struct dma_fence *fence),
 546                         void *arg)
 547 {
 548         int err = 0;
 549
 550         /* We must always wait for the exclusive fence! */
 551         if (rcu_access_pointer(ref->excl.fence)) {
 552                 err = __await_active(&ref->excl, fn, arg);
 553                 if (err)
 554                         return err;
 555         }
 556
 557         if (flags & I915_ACTIVE_AWAIT_ALL && i915_active_acquire_if_busy(ref)) {
 558                 struct active_node *it, *n;
 559
 560                 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
 561                         err = __await_active(&it->base, fn, arg);
 562                         if (err)
 563                                 break;
 564                 }
 565                 i915_active_release(ref);
 566                 if (err)
 567                         return err;
 568         }
 569
 570         return 0;
 571 }
 572
 573 static int rq_await_fence(void *arg, struct dma_fence *fence)
 574 {
 575         return i915_request_await_dma_fence(arg, fence);
 576 }
 577
 578 int i915_request_await_active(struct i915_request *rq,
 579                               struct i915_active *ref,
 580                               unsigned int flags)
 581 {
 582         return await_active(ref, flags, rq_await_fence, rq);
 583 }
 584
 585 static int sw_await_fence(void *arg, struct dma_fence *fence)
 586 {
 587         return i915_sw_fence_await_dma_fence(arg, fence, 0,
 588                                              GFP_NOWAIT | __GFP_NOWARN);
 589 }
 590
 591 int i915_sw_fence_await_active(struct i915_sw_fence *fence,
 592                                struct i915_active *ref,
 593                                unsigned int flags)
 594 {
 595         return await_active(ref, flags, sw_await_fence, fence);
 596 }
 597
 598 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM)
 599 void i915_active_fini(struct i915_active *ref)
 600 {
 601         debug_active_fini(ref);
 602         GEM_BUG_ON(atomic_read(&ref->count));
 603         GEM_BUG_ON(work_pending(&ref->work));
 604         GEM_BUG_ON(!RB_EMPTY_ROOT(&ref->tree));
 605         mutex_destroy(&ref->mutex);
 606 }
 607 #endif
 608
 609 static inline bool is_idle_barrier(struct active_node *node, u64 idx)
 610 {
 611         return node->timeline == idx && !i915_active_fence_isset(&node->base);
 612 }
 613
 614 static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx)
 615 {
 616         struct rb_node *prev, *p;
 617
 618         if (RB_EMPTY_ROOT(&ref->tree))
 619                 return NULL;
 620
 621         spin_lock_irq(&ref->tree_lock);
 622         GEM_BUG_ON(i915_active_is_idle(ref));
 623
 624         /*
 625          * Try to reuse any existing barrier nodes already allocated for this
 626          * i915_active, due to overlapping active phases there is likely a
 627          * node kept alive (as we reuse before parking). We prefer to reuse
 628          * completely idle barriers (less hassle in manipulating the llists),
 629          * but otherwise any will do.
 630          */
 631         if (ref->cache && is_idle_barrier(ref->cache, idx)) {
 632                 p = &ref->cache->node;
 633                 goto match;
 634         }
 635
 636         prev = NULL;
 637         p = ref->tree.rb_node;
 638         while (p) {
 639                 struct active_node *node =
 640                         rb_entry(p, struct active_node, node);
 641
 642                 if (is_idle_barrier(node, idx))
 643                         goto match;
 644
 645                 prev = p;
 646                 if (node->timeline < idx)
 647                         p = p->rb_right;
 648                 else
 649                         p = p->rb_left;
 650         }
 651
 652         /*
 653          * No quick match, but we did find the leftmost rb_node for the
 654          * kernel_context. Walk the rb_tree in-order to see if there were
 655          * any idle-barriers on this timeline that we missed, or just use
 656          * the first pending barrier.
 657          */
 658         for (p = prev; p; p = rb_next(p)) {
 659                 struct active_node *node =
 660                         rb_entry(p, struct active_node, node);
 661                 struct intel_engine_cs *engine;
 662
 663                 if (node->timeline > idx)
 664                         break;
 665
 666                 if (node->timeline < idx)
 667                         continue;
 668
 669                 if (is_idle_barrier(node, idx))
 670                         goto match;
 671
 672                 /*
 673                  * The list of pending barriers is protected by the
 674                  * kernel_context timeline, which notably we do not hold
 675                  * here. i915_request_add_active_barriers() may consume
 676                  * the barrier before we claim it, so we have to check
 677                  * for success.
 678                  */
 679                 engine = __barrier_to_engine(node);
 680                 smp_rmb(); /* serialise with add_active_barriers */
 681                 if (is_barrier(&node->base) &&
 682                     ____active_del_barrier(ref, node, engine))
 683                         goto match;
 684         }
 685
 686         spin_unlock_irq(&ref->tree_lock);
 687
 688         return NULL;
 689
 690 match:
 691         rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */
 692         if (p == &ref->cache->node)
 693                 ref->cache = NULL;
 694         spin_unlock_irq(&ref->tree_lock);
 695
 696         return rb_entry(p, struct active_node, node);
 697 }
 698
 699 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
 700                                             struct intel_engine_cs *engine)
 701 {
 702         intel_engine_mask_t tmp, mask = engine->mask;
 703         struct llist_node *first = NULL, *last = NULL;
 704         struct intel_gt *gt = engine->gt;
 705         int err;
 706
 707         GEM_BUG_ON(i915_active_is_idle(ref));
 708
 709         /* Wait until the previous preallocation is completed */
 710         while (!llist_empty(&ref->preallocated_barriers))
 711                 cond_resched();
 712
 713         /*
 714          * Preallocate a node for each physical engine supporting the target
 715          * engine (remember virtual engines have more than one sibling).
 716          * We can then use the preallocated nodes in
 717          * i915_active_acquire_barrier()
 718          */
 719         GEM_BUG_ON(!mask);
 720         for_each_engine_masked(engine, gt, mask, tmp) {
 721                 u64 idx = engine->kernel_context->timeline->fence_context;
 722                 struct llist_node *prev = first;
 723                 struct active_node *node;
 724
 725                 node = reuse_idle_barrier(ref, idx);
 726                 if (!node) {
 727                         node = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
 728                         if (!node) {
 729                                 err = ENOMEM;
 730                                 goto unwind;
 731                         }
 732
 733                         RCU_INIT_POINTER(node->base.fence, NULL);
 734                         node->base.cb.func = node_retire;
 735                         node->timeline = idx;
 736                         node->ref = ref;
 737                 }
 738
 739                 if (!i915_active_fence_isset(&node->base)) {
 740                         /*
 741                          * Mark this as being *our* unconnected proto-node.
 742                          *
 743                          * Since this node is not in any list, and we have
 744                          * decoupled it from the rbtree, we can reuse the
 745                          * request to indicate this is an idle-barrier node
 746                          * and then we can use the rb_node and list pointers
 747                          * for our tracking of the pending barrier.
 748                          */
 749                         RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
 750                         node->base.cb.node.prev = (void *)engine;
 751                         atomic_inc(&ref->count);
 752                 }
 753                 GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
 754
 755                 GEM_BUG_ON(barrier_to_engine(node) != engine);
 756                 first = barrier_to_ll(node);
 757                 first->next = prev;
 758                 if (!last)
 759                         last = first;
 760                 intel_engine_pm_get(engine);
 761         }
 762
 763         GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
 764         llist_add_batch(first, last, &ref->preallocated_barriers);
 765
 766         return 0;
 767
 768 unwind:
 769         while (first) {
 770                 struct active_node *node = barrier_from_ll(first);
 771
 772                 first = first->next;
 773
 774                 atomic_dec(&ref->count);
 775                 intel_engine_pm_put(barrier_to_engine(node));
 776
 777                 kmem_cache_free(global.slab_cache, node);
 778         }
 779         return err;
 780 }
 781
 782 void i915_active_acquire_barrier(struct i915_active *ref)
 783 {
 784         struct llist_node *pos, *next;
 785         unsigned long flags;
 786
 787         GEM_BUG_ON(i915_active_is_idle(ref));
 788
 789         /*
 790          * Transfer the list of preallocated barriers into the
 791          * i915_active rbtree, but only as proto-nodes. They will be
 792          * populated by i915_request_add_active_barriers() to point to the
 793          * request that will eventually release them.
 794          */
 795         llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
 796                 struct active_node *node = barrier_from_ll(pos);
 797                 struct intel_engine_cs *engine = barrier_to_engine(node);
 798                 struct rb_node **p, *parent;
 799
 800                 spin_lock_irqsave_nested(&ref->tree_lock, flags,
 801                                          SINGLE_DEPTH_NESTING);
 802                 parent = NULL;
 803                 p = &ref->tree.rb_node;
 804                 while (*p) {
 805                         struct active_node *it;
 806
 807                         parent = *p;
 808
 809                         it = rb_entry(parent, struct active_node, node);
 810                         if (it->timeline < node->timeline)
 811                                 p = &parent->rb_right;
 812                         else
 813                                 p = &parent->rb_left;
 814                 }
 815                 rb_link_node(&node->node, parent, p);
 816                 rb_insert_color(&node->node, &ref->tree);
 817                 spin_unlock_irqrestore(&ref->tree_lock, flags);
 818
 819                 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
 820                 llist_add(barrier_to_ll(node), &engine->barrier_tasks);
 821                 intel_engine_pm_put(engine);
 822         }
 823 }
 824
 825 static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
 826 {
 827         return __active_fence_slot(&barrier_from_ll(node)->base);
 828 }
 829
 830 void i915_request_add_active_barriers(struct i915_request *rq)
 831 {
 832         struct intel_engine_cs *engine = rq->engine;
 833         struct llist_node *node, *next;
 834         unsigned long flags;
 835
 836         GEM_BUG_ON(!intel_context_is_barrier(rq->context));
 837         GEM_BUG_ON(intel_engine_is_virtual(engine));
 838         GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline);
 839
 840         node = llist_del_all(&engine->barrier_tasks);
 841         if (!node)
 842                 return;
 843         /*
 844          * Attach the list of proto-fences to the in-flight request such
 845          * that the parent i915_active will be released when this request
 846          * is retired.
 847          */
 848         spin_lock_irqsave(&rq->lock, flags);
 849         llist_for_each_safe(node, next, node) {
 850                 /* serialise with reuse_idle_barrier */
 851                 smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
 852                 list_add_tail((struct list_head *)node, &rq->fence.cb_list);
 853         }
 854         spin_unlock_irqrestore(&rq->lock, flags);
 855 }
 856
 857 /*
 858  * __i915_active_fence_set: Update the last active fence along its timeline
 859  * @active: the active tracker
 860  * @fence: the new fence (under construction)
 861  *
 862  * Records the new @fence as the last active fence along its timeline in
 863  * this active tracker, moving the tracking callbacks from the previous
 864  * fence onto this one. Returns the previous fence (if not already completed),
 865  * which the caller must ensure is executed before the new fence. To ensure
 866  * that the order of fences within the timeline of the i915_active_fence is
 867  * understood, it should be locked by the caller.
 868  */
 869 struct dma_fence *
 870 __i915_active_fence_set(struct i915_active_fence *active,
 871                         struct dma_fence *fence)
 872 {
 873         struct dma_fence *prev;
 874         unsigned long flags;
 875
 876         if (fence == rcu_access_pointer(active->fence))
 877                 return fence;
 878
 879         GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
 880
 881         /*
 882          * Consider that we have two threads arriving (A and B), with
 883          * C already resident as the active->fence.
 884          *
 885          * A does the xchg first, and so it sees C or NULL depending
 886          * on the timing of the interrupt handler. If it is NULL, the
 887          * previous fence must have been signaled and we know that
 888          * we are first on the timeline. If it is still present,
 889          * we acquire the lock on that fence and serialise with the interrupt
 890          * handler, in the process removing it from any future interrupt
 891          * callback. A will then wait on C before executing (if present).
 892          *
 893          * As B is second, it sees A as the previous fence and so waits for
 894          * it to complete its transition and takes over the occupancy for
 895          * itself -- remembering that it needs to wait on A before executing.
 896          *
 897          * Note the strong ordering of the timeline also provides consistent
 898          * nesting rules for the fence->lock; the inner lock is always the
 899          * older lock.
 900          */
 901         spin_lock_irqsave(fence->lock, flags);
 902         prev = xchg(__active_fence_slot(active), fence);
 903         if (prev) {
 904                 GEM_BUG_ON(prev == fence);
 905                 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
 906                 __list_del_entry(&active->cb.node);
 907                 spin_unlock(prev->lock); /* serialise with prev->cb_list */
 908         }
 909         list_add_tail(&active->cb.node, &fence->cb_list);
 910         spin_unlock_irqrestore(fence->lock, flags);
 911
 912         return prev;
 913 }
 914
 915 int i915_active_fence_set(struct i915_active_fence *active,
 916                           struct i915_request *rq)
 917 {
 918         struct dma_fence *fence;
 919         int err = 0;
 920
 921         /* Must maintain timeline ordering wrt previous active requests */
 922         rcu_read_lock();
 923         fence = __i915_active_fence_set(active, &rq->fence);
 924         if (fence) /* but the previous fence may not belong to that timeline! */
 925                 fence = dma_fence_get_rcu(fence);
 926         rcu_read_unlock();
 927         if (fence) {
 928                 err = i915_request_await_dma_fence(rq, fence);
 929                 dma_fence_put(fence);
 930         }
 931
 932         return err;
 933 }
 934
 935 void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb)
 936 {
 937         active_fence_cb(fence, cb);
 938 }
 939
 940 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
 941 #include "selftests/i915_active.c"
 942 #endif
 943
 944 static void i915_global_active_shrink(void)
 945 {
 946         kmem_cache_shrink(global.slab_cache);
 947 }
 948
 949 static void i915_global_active_exit(void)
 950 {
 951         kmem_cache_destroy(global.slab_cache);
 952 }
 953
 954 static struct i915_global_active global = { {
 955         .shrink = i915_global_active_shrink,
 956         .exit = i915_global_active_exit,
 957 } };
 958
 959 int __init i915_global_active_init(void)
 960 {
 961         global.slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN);
 962         if (!global.slab_cache)
 963                 return -ENOMEM;
 964
 965         i915_global_register(&global.base);
 966         return 0;
 967 }