drivers/gpu/drm/i915/i915_active.c

   1 /*
   2  * SPDX-License-Identifier: MIT
   3  *
   4  * Copyright © 2019 Intel Corporation
   5  */
   6
   7 #include <linux/debugobjects.h>
   8
   9 #include "gt/intel_context.h"
  10 #include "gt/intel_engine_heartbeat.h"
  11 #include "gt/intel_engine_pm.h"
  12 #include "gt/intel_ring.h"
  13
  14 #include "i915_drv.h"
  15 #include "i915_active.h"
  16 #include "i915_globals.h"
  17
  18 /*
  19  * Active refs memory management
  20  *
  21  * To be more economical with memory, we reap all the i915_active trees as
  22  * they idle (when we know the active requests are inactive) and allocate the
  23  * nodes from a local slab cache to hopefully reduce the fragmentation.
  24  */
  25 static struct i915_global_active {
  26         struct i915_global base;
  27         struct kmem_cache *slab_cache;
  28 } global;
  29
  30 struct active_node {
  31         struct rb_node node;
  32         struct i915_active_fence base;
  33         struct i915_active *ref;
  34         u64 timeline;
  35 };
  36
  37 #define fetch_node(x) rb_entry(READ_ONCE(x), typeof(struct active_node), node)
  38
  39 static inline struct active_node *
  40 node_from_active(struct i915_active_fence *active)
  41 {
  42         return container_of(active, struct active_node, base);
  43 }
  44
  45 #define take_preallocated_barriers(x) llist_del_all(&(x)->preallocated_barriers)
  46
  47 static inline bool is_barrier(const struct i915_active_fence *active)
  48 {
  49         return IS_ERR(rcu_access_pointer(active->fence));
  50 }
  51
  52 static inline struct llist_node *barrier_to_ll(struct active_node *node)
  53 {
  54         GEM_BUG_ON(!is_barrier(&node->base));
  55         return (struct llist_node *)&node->base.cb.node;
  56 }
  57
  58 static inline struct intel_engine_cs *
  59 __barrier_to_engine(struct active_node *node)
  60 {
  61         return (struct intel_engine_cs *)READ_ONCE(node->base.cb.node.prev);
  62 }
  63
  64 static inline struct intel_engine_cs *
  65 barrier_to_engine(struct active_node *node)
  66 {
  67         GEM_BUG_ON(!is_barrier(&node->base));
  68         return __barrier_to_engine(node);
  69 }
  70
  71 static inline struct active_node *barrier_from_ll(struct llist_node *x)
  72 {
  73         return container_of((struct list_head *)x,
  74                             struct active_node, base.cb.node);
  75 }
  76
  77 #if IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM) && IS_ENABLED(CONFIG_DEBUG_OBJECTS)
  78
  79 static void *active_debug_hint(void *addr)
  80 {
  81         struct i915_active *ref = addr;
  82
  83         return (void *)ref->active ?: (void *)ref->retire ?: (void *)ref;
  84 }
  85
  86 static const struct debug_obj_descr active_debug_desc = {
  87         .name = "i915_active",
  88         .debug_hint = active_debug_hint,
  89 };
  90
  91 static void debug_active_init(struct i915_active *ref)
  92 {
  93         debug_object_init(ref, &active_debug_desc);
  94 }
  95
  96 static void debug_active_activate(struct i915_active *ref)
  97 {
  98         lockdep_assert_held(&ref->tree_lock);
  99         if (!atomic_read(&ref->count)) /* before the first inc */
 100                 debug_object_activate(ref, &active_debug_desc);
 101 }
 102
 103 static void debug_active_deactivate(struct i915_active *ref)
 104 {
 105         lockdep_assert_held(&ref->tree_lock);
 106         if (!atomic_read(&ref->count)) /* after the last dec */
 107                 debug_object_deactivate(ref, &active_debug_desc);
 108 }
 109
 110 static void debug_active_fini(struct i915_active *ref)
 111 {
 112         debug_object_free(ref, &active_debug_desc);
 113 }
 114
 115 static void debug_active_assert(struct i915_active *ref)
 116 {
 117         debug_object_assert_init(ref, &active_debug_desc);
 118 }
 119
 120 #else
 121
 122 static inline void debug_active_init(struct i915_active *ref) { }
 123 static inline void debug_active_activate(struct i915_active *ref) { }
 124 static inline void debug_active_deactivate(struct i915_active *ref) { }
 125 static inline void debug_active_fini(struct i915_active *ref) { }
 126 static inline void debug_active_assert(struct i915_active *ref) { }
 127
 128 #endif
 129
 130 static void
 131 __active_retire(struct i915_active *ref)
 132 {
 133         struct rb_root root = RB_ROOT;
 134         struct active_node *it, *n;
 135         unsigned long flags;
 136
 137         GEM_BUG_ON(i915_active_is_idle(ref));
 138
 139         /* return the unused nodes to our slabcache -- flushing the allocator */
 140         if (!atomic_dec_and_lock_irqsave(&ref->count, &ref->tree_lock, flags))
 141                 return;
 142
 143         GEM_BUG_ON(rcu_access_pointer(ref->excl.fence));
 144         debug_active_deactivate(ref);
 145
 146         /* Even if we have not used the cache, we may still have a barrier */
 147         if (!ref->cache)
 148                 ref->cache = fetch_node(ref->tree.rb_node);
 149
 150         /* Keep the MRU cached node for reuse */
 151         if (ref->cache) {
 152                 /* Discard all other nodes in the tree */
 153                 rb_erase(&ref->cache->node, &ref->tree);
 154                 root = ref->tree;
 155
 156                 /* Rebuild the tree with only the cached node */
 157                 rb_link_node(&ref->cache->node, NULL, &ref->tree.rb_node);
 158                 rb_insert_color(&ref->cache->node, &ref->tree);
 159                 GEM_BUG_ON(ref->tree.rb_node != &ref->cache->node);
 160
 161                 /* Make the cached node available for reuse with any timeline */
 162                 if (IS_ENABLED(CONFIG_64BIT))
 163                         ref->cache->timeline = 0; /* needs cmpxchg(u64) */
 164         }
 165
 166         spin_unlock_irqrestore(&ref->tree_lock, flags);
 167
 168         /* After the final retire, the entire struct may be freed */
 169         if (ref->retire)
 170                 ref->retire(ref);
 171
 172         /* ... except if you wait on it, you must manage your own references! */
 173         wake_up_var(ref);
 174
 175         /* Finally free the discarded timeline tree  */
 176         rbtree_postorder_for_each_entry_safe(it, n, &root, node) {
 177                 GEM_BUG_ON(i915_active_fence_isset(&it->base));
 178                 kmem_cache_free(global.slab_cache, it);
 179         }
 180 }
 181
 182 static void
 183 active_work(struct work_struct *wrk)
 184 {
 185         struct i915_active *ref = container_of(wrk, typeof(*ref), work);
 186
 187         GEM_BUG_ON(!atomic_read(&ref->count));
 188         if (atomic_add_unless(&ref->count, -1, 1))
 189                 return;
 190
 191         __active_retire(ref);
 192 }
 193
 194 static void
 195 active_retire(struct i915_active *ref)
 196 {
 197         GEM_BUG_ON(!atomic_read(&ref->count));
 198         if (atomic_add_unless(&ref->count, -1, 1))
 199                 return;
 200
 201         if (ref->flags & I915_ACTIVE_RETIRE_SLEEPS) {
 202                 queue_work(system_unbound_wq, &ref->work);
 203                 return;
 204         }
 205
 206         __active_retire(ref);
 207 }
 208
 209 static inline struct dma_fence **
 210 __active_fence_slot(struct i915_active_fence *active)
 211 {
 212         return (struct dma_fence ** __force)&active->fence;
 213 }
 214
 215 static inline bool
 216 active_fence_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
 217 {
 218         struct i915_active_fence *active =
 219                 container_of(cb, typeof(*active), cb);
 220
 221         return cmpxchg(__active_fence_slot(active), fence, NULL) == fence;
 222 }
 223
 224 static void
 225 node_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
 226 {
 227         if (active_fence_cb(fence, cb))
 228                 active_retire(container_of(cb, struct active_node, base.cb)->ref);
 229 }
 230
 231 static void
 232 excl_retire(struct dma_fence *fence, struct dma_fence_cb *cb)
 233 {
 234         if (active_fence_cb(fence, cb))
 235                 active_retire(container_of(cb, struct i915_active, excl.cb));
 236 }
 237
 238 static struct active_node *__active_lookup(struct i915_active *ref, u64 idx)
 239 {
 240         struct active_node *it;
 241
 242         GEM_BUG_ON(idx == 0); /* 0 is the unordered timeline, rsvd for cache */
 243
 244         /*
 245          * We track the most recently used timeline to skip a rbtree search
 246          * for the common case, under typical loads we never need the rbtree
 247          * at all. We can reuse the last slot if it is empty, that is
 248          * after the previous activity has been retired, or if it matches the
 249          * current timeline.
 250          */
 251         it = READ_ONCE(ref->cache);
 252         if (it) {
 253                 u64 cached = READ_ONCE(it->timeline);
 254
 255                 /* Once claimed, this slot will only belong to this idx */
 256                 if (cached == idx)
 257                         return it;
 258
 259 #ifdef CONFIG_64BIT /* for cmpxchg(u64) */
 260                 /*
 261                  * An unclaimed cache [.timeline=0] can only be claimed once.
 262                  *
 263                  * If the value is already non-zero, some other thread has
 264                  * claimed the cache and we know that is does not match our
 265                  * idx. If, and only if, the timeline is currently zero is it
 266                  * worth competing to claim it atomically for ourselves (for
 267                  * only the winner of that race will cmpxchg return the old
 268                  * value of 0).
 269                  */
 270                 if (!cached && !cmpxchg(&it->timeline, 0, idx))
 271                         return it;
 272 #endif
 273         }
 274
 275         BUILD_BUG_ON(offsetof(typeof(*it), node));
 276
 277         /* While active, the tree can only be built; not destroyed */
 278         GEM_BUG_ON(i915_active_is_idle(ref));
 279
 280         it = fetch_node(ref->tree.rb_node);
 281         while (it) {
 282                 if (it->timeline < idx) {
 283                         it = fetch_node(it->node.rb_right);
 284                 } else if (it->timeline > idx) {
 285                         it = fetch_node(it->node.rb_left);
 286                 } else {
 287                         WRITE_ONCE(ref->cache, it);
 288                         break;
 289                 }
 290         }
 291
 292         /* NB: If the tree rotated beneath us, we may miss our target. */
 293         return it;
 294 }
 295
 296 static struct i915_active_fence *
 297 active_instance(struct i915_active *ref, u64 idx)
 298 {
 299         struct active_node *node, *prealloc;
 300         struct rb_node **p, *parent;
 301
 302         node = __active_lookup(ref, idx);
 303         if (likely(node))
 304                 return &node->base;
 305
 306         /* Preallocate a replacement, just in case */
 307         prealloc = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
 308         if (!prealloc)
 309                 return NULL;
 310
 311         spin_lock_irq(&ref->tree_lock);
 312         GEM_BUG_ON(i915_active_is_idle(ref));
 313
 314         parent = NULL;
 315         p = &ref->tree.rb_node;
 316         while (*p) {
 317                 parent = *p;
 318
 319                 node = rb_entry(parent, struct active_node, node);
 320                 if (node->timeline == idx) {
 321                         kmem_cache_free(global.slab_cache, prealloc);
 322                         goto out;
 323                 }
 324
 325                 if (node->timeline < idx)
 326                         p = &parent->rb_right;
 327                 else
 328                         p = &parent->rb_left;
 329         }
 330
 331         node = prealloc;
 332         __i915_active_fence_init(&node->base, NULL, node_retire);
 333         node->ref = ref;
 334         node->timeline = idx;
 335
 336         rb_link_node(&node->node, parent, p);
 337         rb_insert_color(&node->node, &ref->tree);
 338
 339 out:
 340         WRITE_ONCE(ref->cache, node);
 341         spin_unlock_irq(&ref->tree_lock);
 342
 343         return &node->base;
 344 }
 345
 346 void __i915_active_init(struct i915_active *ref,
 347                         int (*active)(struct i915_active *ref),
 348                         void (*retire)(struct i915_active *ref),
 349                         struct lock_class_key *mkey,
 350                         struct lock_class_key *wkey)
 351 {
 352         unsigned long bits;
 353
 354         debug_active_init(ref);
 355
 356         ref->flags = 0;
 357         ref->active = active;
 358         ref->retire = ptr_unpack_bits(retire, &bits, 2);
 359         if (bits & I915_ACTIVE_MAY_SLEEP)
 360                 ref->flags |= I915_ACTIVE_RETIRE_SLEEPS;
 361
 362         spin_lock_init(&ref->tree_lock);
 363         ref->tree = RB_ROOT;
 364         ref->cache = NULL;
 365
 366         init_llist_head(&ref->preallocated_barriers);
 367         atomic_set(&ref->count, 0);
 368         __mutex_init(&ref->mutex, "i915_active", mkey);
 369         __i915_active_fence_init(&ref->excl, NULL, excl_retire);
 370         INIT_WORK(&ref->work, active_work);
 371 #if IS_ENABLED(CONFIG_LOCKDEP)
 372         lockdep_init_map(&ref->work.lockdep_map, "i915_active.work", wkey, 0);
 373 #endif
 374 }
 375
 376 static bool ____active_del_barrier(struct i915_active *ref,
 377                                    struct active_node *node,
 378                                    struct intel_engine_cs *engine)
 379
 380 {
 381         struct llist_node *head = NULL, *tail = NULL;
 382         struct llist_node *pos, *next;
 383
 384         GEM_BUG_ON(node->timeline != engine->kernel_context->timeline->fence_context);
 385
 386         /*
 387          * Rebuild the llist excluding our node. We may perform this
 388          * outside of the kernel_context timeline mutex and so someone
 389          * else may be manipulating the engine->barrier_tasks, in
 390          * which case either we or they will be upset :)
 391          *
 392          * A second __active_del_barrier() will report failure to claim
 393          * the active_node and the caller will just shrug and know not to
 394          * claim ownership of its node.
 395          *
 396          * A concurrent i915_request_add_active_barriers() will miss adding
 397          * any of the tasks, but we will try again on the next -- and since
 398          * we are actively using the barrier, we know that there will be
 399          * at least another opportunity when we idle.
 400          */
 401         llist_for_each_safe(pos, next, llist_del_all(&engine->barrier_tasks)) {
 402                 if (node == barrier_from_ll(pos)) {
 403                         node = NULL;
 404                         continue;
 405                 }
 406
 407                 pos->next = head;
 408                 head = pos;
 409                 if (!tail)
 410                         tail = pos;
 411         }
 412         if (head)
 413                 llist_add_batch(head, tail, &engine->barrier_tasks);
 414
 415         return !node;
 416 }
 417
 418 static bool
 419 __active_del_barrier(struct i915_active *ref, struct active_node *node)
 420 {
 421         return ____active_del_barrier(ref, node, barrier_to_engine(node));
 422 }
 423
 424 static bool
 425 replace_barrier(struct i915_active *ref, struct i915_active_fence *active)
 426 {
 427         if (!is_barrier(active)) /* proto-node used by our idle barrier? */
 428                 return false;
 429
 430         /*
 431          * This request is on the kernel_context timeline, and so
 432          * we can use it to substitute for the pending idle-barrer
 433          * request that we want to emit on the kernel_context.
 434          */
 435         __active_del_barrier(ref, node_from_active(active));
 436         return true;
 437 }
 438
 439 int i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence)
 440 {
 441         struct i915_active_fence *active;
 442         int err;
 443
 444         /* Prevent reaping in case we malloc/wait while building the tree */
 445         err = i915_active_acquire(ref);
 446         if (err)
 447                 return err;
 448
 449         active = active_instance(ref, idx);
 450         if (!active) {
 451                 err = -ENOMEM;
 452                 goto out;
 453         }
 454
 455         if (replace_barrier(ref, active)) {
 456                 RCU_INIT_POINTER(active->fence, NULL);
 457                 atomic_dec(&ref->count);
 458         }
 459         if (!__i915_active_fence_set(active, fence))
 460                 __i915_active_acquire(ref);
 461
 462 out:
 463         i915_active_release(ref);
 464         return err;
 465 }
 466
 467 static struct dma_fence *
 468 __i915_active_set_fence(struct i915_active *ref,
 469                         struct i915_active_fence *active,
 470                         struct dma_fence *fence)
 471 {
 472         struct dma_fence *prev;
 473
 474         if (replace_barrier(ref, active)) {
 475                 RCU_INIT_POINTER(active->fence, fence);
 476                 return NULL;
 477         }
 478
 479         rcu_read_lock();
 480         prev = __i915_active_fence_set(active, fence);
 481         if (prev)
 482                 prev = dma_fence_get_rcu(prev);
 483         else
 484                 __i915_active_acquire(ref);
 485         rcu_read_unlock();
 486
 487         return prev;
 488 }
 489
 490 static struct i915_active_fence *
 491 __active_fence(struct i915_active *ref, u64 idx)
 492 {
 493         struct active_node *it;
 494
 495         it = __active_lookup(ref, idx);
 496         if (unlikely(!it)) { /* Contention with parallel tree builders! */
 497                 spin_lock_irq(&ref->tree_lock);
 498                 it = __active_lookup(ref, idx);
 499                 spin_unlock_irq(&ref->tree_lock);
 500         }
 501         GEM_BUG_ON(!it); /* slot must be preallocated */
 502
 503         return &it->base;
 504 }
 505
 506 struct dma_fence *
 507 __i915_active_ref(struct i915_active *ref, u64 idx, struct dma_fence *fence)
 508 {
 509         /* Only valid while active, see i915_active_acquire_for_context() */
 510         return __i915_active_set_fence(ref, __active_fence(ref, idx), fence);
 511 }
 512
 513 struct dma_fence *
 514 i915_active_set_exclusive(struct i915_active *ref, struct dma_fence *f)
 515 {
 516         /* We expect the caller to manage the exclusive timeline ordering */
 517         return __i915_active_set_fence(ref, &ref->excl, f);
 518 }
 519
 520 bool i915_active_acquire_if_busy(struct i915_active *ref)
 521 {
 522         debug_active_assert(ref);
 523         return atomic_add_unless(&ref->count, 1, 0);
 524 }
 525
 526 static void __i915_active_activate(struct i915_active *ref)
 527 {
 528         spin_lock_irq(&ref->tree_lock); /* __active_retire() */
 529         if (!atomic_fetch_inc(&ref->count))
 530                 debug_active_activate(ref);
 531         spin_unlock_irq(&ref->tree_lock);
 532 }
 533
 534 int i915_active_acquire(struct i915_active *ref)
 535 {
 536         int err;
 537
 538         if (i915_active_acquire_if_busy(ref))
 539                 return 0;
 540
 541         if (!ref->active) {
 542                 __i915_active_activate(ref);
 543                 return 0;
 544         }
 545
 546         err = mutex_lock_interruptible(&ref->mutex);
 547         if (err)
 548                 return err;
 549
 550         if (likely(!i915_active_acquire_if_busy(ref))) {
 551                 err = ref->active(ref);
 552                 if (!err)
 553                         __i915_active_activate(ref);
 554         }
 555
 556         mutex_unlock(&ref->mutex);
 557
 558         return err;
 559 }
 560
 561 int i915_active_acquire_for_context(struct i915_active *ref, u64 idx)
 562 {
 563         struct i915_active_fence *active;
 564         int err;
 565
 566         err = i915_active_acquire(ref);
 567         if (err)
 568                 return err;
 569
 570         active = active_instance(ref, idx);
 571         if (!active) {
 572                 i915_active_release(ref);
 573                 return -ENOMEM;
 574         }
 575
 576         return 0; /* return with active ref */
 577 }
 578
 579 void i915_active_release(struct i915_active *ref)
 580 {
 581         debug_active_assert(ref);
 582         active_retire(ref);
 583 }
 584
 585 static void enable_signaling(struct i915_active_fence *active)
 586 {
 587         struct dma_fence *fence;
 588
 589         if (unlikely(is_barrier(active)))
 590                 return;
 591
 592         fence = i915_active_fence_get(active);
 593         if (!fence)
 594                 return;
 595
 596         dma_fence_enable_sw_signaling(fence);
 597         dma_fence_put(fence);
 598 }
 599
 600 static int flush_barrier(struct active_node *it)
 601 {
 602         struct intel_engine_cs *engine;
 603
 604         if (likely(!is_barrier(&it->base)))
 605                 return 0;
 606
 607         engine = __barrier_to_engine(it);
 608         smp_rmb(); /* serialise with add_active_barriers */
 609         if (!is_barrier(&it->base))
 610                 return 0;
 611
 612         return intel_engine_flush_barriers(engine);
 613 }
 614
 615 static int flush_lazy_signals(struct i915_active *ref)
 616 {
 617         struct active_node *it, *n;
 618         int err = 0;
 619
 620         enable_signaling(&ref->excl);
 621         rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
 622                 err = flush_barrier(it); /* unconnected idle barrier? */
 623                 if (err)
 624                         break;
 625
 626                 enable_signaling(&it->base);
 627         }
 628
 629         return err;
 630 }
 631
 632 int __i915_active_wait(struct i915_active *ref, int state)
 633 {
 634         int err;
 635
 636         might_sleep();
 637
 638         if (!i915_active_acquire_if_busy(ref))
 639                 return 0;
 640
 641         /* Any fence added after the wait begins will not be auto-signaled */
 642         err = flush_lazy_signals(ref);
 643         i915_active_release(ref);
 644         if (err)
 645                 return err;
 646
 647         if (!i915_active_is_idle(ref) &&
 648             ___wait_var_event(ref, i915_active_is_idle(ref),
 649                               state, 0, 0, schedule()))
 650                 return -EINTR;
 651
 652         flush_work(&ref->work);
 653         return 0;
 654 }
 655
 656 static int __await_active(struct i915_active_fence *active,
 657                           int (*fn)(void *arg, struct dma_fence *fence),
 658                           void *arg)
 659 {
 660         struct dma_fence *fence;
 661
 662         if (is_barrier(active)) /* XXX flush the barrier? */
 663                 return 0;
 664
 665         fence = i915_active_fence_get(active);
 666         if (fence) {
 667                 int err;
 668
 669                 err = fn(arg, fence);
 670                 dma_fence_put(fence);
 671                 if (err < 0)
 672                         return err;
 673         }
 674
 675         return 0;
 676 }
 677
 678 struct wait_barrier {
 679         struct wait_queue_entry base;
 680         struct i915_active *ref;
 681 };
 682
 683 static int
 684 barrier_wake(wait_queue_entry_t *wq, unsigned int mode, int flags, void *key)
 685 {
 686         struct wait_barrier *wb = container_of(wq, typeof(*wb), base);
 687
 688         if (i915_active_is_idle(wb->ref)) {
 689                 list_del(&wq->entry);
 690                 i915_sw_fence_complete(wq->private);
 691                 kfree(wq);
 692         }
 693
 694         return 0;
 695 }
 696
 697 static int __await_barrier(struct i915_active *ref, struct i915_sw_fence *fence)
 698 {
 699         struct wait_barrier *wb;
 700
 701         wb = kmalloc(sizeof(*wb), GFP_KERNEL);
 702         if (unlikely(!wb))
 703                 return -ENOMEM;
 704
 705         GEM_BUG_ON(i915_active_is_idle(ref));
 706         if (!i915_sw_fence_await(fence)) {
 707                 kfree(wb);
 708                 return -EINVAL;
 709         }
 710
 711         wb->base.flags = 0;
 712         wb->base.func = barrier_wake;
 713         wb->base.private = fence;
 714         wb->ref = ref;
 715
 716         add_wait_queue(__var_waitqueue(ref), &wb->base);
 717         return 0;
 718 }
 719
 720 static int await_active(struct i915_active *ref,
 721                         unsigned int flags,
 722                         int (*fn)(void *arg, struct dma_fence *fence),
 723                         void *arg, struct i915_sw_fence *barrier)
 724 {
 725         int err = 0;
 726
 727         if (!i915_active_acquire_if_busy(ref))
 728                 return 0;
 729
 730         if (flags & I915_ACTIVE_AWAIT_EXCL &&
 731             rcu_access_pointer(ref->excl.fence)) {
 732                 err = __await_active(&ref->excl, fn, arg);
 733                 if (err)
 734                         goto out;
 735         }
 736
 737         if (flags & I915_ACTIVE_AWAIT_ACTIVE) {
 738                 struct active_node *it, *n;
 739
 740                 rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
 741                         err = __await_active(&it->base, fn, arg);
 742                         if (err)
 743                                 goto out;
 744                 }
 745         }
 746
 747         if (flags & I915_ACTIVE_AWAIT_BARRIER) {
 748                 err = flush_lazy_signals(ref);
 749                 if (err)
 750                         goto out;
 751
 752                 err = __await_barrier(ref, barrier);
 753                 if (err)
 754                         goto out;
 755         }
 756
 757 out:
 758         i915_active_release(ref);
 759         return err;
 760 }
 761
 762 static int rq_await_fence(void *arg, struct dma_fence *fence)
 763 {
 764         return i915_request_await_dma_fence(arg, fence);
 765 }
 766
 767 int i915_request_await_active(struct i915_request *rq,
 768                               struct i915_active *ref,
 769                               unsigned int flags)
 770 {
 771         return await_active(ref, flags, rq_await_fence, rq, &rq->submit);
 772 }
 773
 774 static int sw_await_fence(void *arg, struct dma_fence *fence)
 775 {
 776         return i915_sw_fence_await_dma_fence(arg, fence, 0,
 777                                              GFP_NOWAIT | __GFP_NOWARN);
 778 }
 779
 780 int i915_sw_fence_await_active(struct i915_sw_fence *fence,
 781                                struct i915_active *ref,
 782                                unsigned int flags)
 783 {
 784         return await_active(ref, flags, sw_await_fence, fence, fence);
 785 }
 786
 787 void i915_active_fini(struct i915_active *ref)
 788 {
 789         debug_active_fini(ref);
 790         GEM_BUG_ON(atomic_read(&ref->count));
 791         GEM_BUG_ON(work_pending(&ref->work));
 792         mutex_destroy(&ref->mutex);
 793
 794         if (ref->cache)
 795                 kmem_cache_free(global.slab_cache, ref->cache);
 796 }
 797
 798 static inline bool is_idle_barrier(struct active_node *node, u64 idx)
 799 {
 800         return node->timeline == idx && !i915_active_fence_isset(&node->base);
 801 }
 802
 803 static struct active_node *reuse_idle_barrier(struct i915_active *ref, u64 idx)
 804 {
 805         struct rb_node *prev, *p;
 806
 807         if (RB_EMPTY_ROOT(&ref->tree))
 808                 return NULL;
 809
 810         GEM_BUG_ON(i915_active_is_idle(ref));
 811
 812         /*
 813          * Try to reuse any existing barrier nodes already allocated for this
 814          * i915_active, due to overlapping active phases there is likely a
 815          * node kept alive (as we reuse before parking). We prefer to reuse
 816          * completely idle barriers (less hassle in manipulating the llists),
 817          * but otherwise any will do.
 818          */
 819         if (ref->cache && is_idle_barrier(ref->cache, idx)) {
 820                 p = &ref->cache->node;
 821                 goto match;
 822         }
 823
 824         prev = NULL;
 825         p = ref->tree.rb_node;
 826         while (p) {
 827                 struct active_node *node =
 828                         rb_entry(p, struct active_node, node);
 829
 830                 if (is_idle_barrier(node, idx))
 831                         goto match;
 832
 833                 prev = p;
 834                 if (node->timeline < idx)
 835                         p = READ_ONCE(p->rb_right);
 836                 else
 837                         p = READ_ONCE(p->rb_left);
 838         }
 839
 840         /*
 841          * No quick match, but we did find the leftmost rb_node for the
 842          * kernel_context. Walk the rb_tree in-order to see if there were
 843          * any idle-barriers on this timeline that we missed, or just use
 844          * the first pending barrier.
 845          */
 846         for (p = prev; p; p = rb_next(p)) {
 847                 struct active_node *node =
 848                         rb_entry(p, struct active_node, node);
 849                 struct intel_engine_cs *engine;
 850
 851                 if (node->timeline > idx)
 852                         break;
 853
 854                 if (node->timeline < idx)
 855                         continue;
 856
 857                 if (is_idle_barrier(node, idx))
 858                         goto match;
 859
 860                 /*
 861                  * The list of pending barriers is protected by the
 862                  * kernel_context timeline, which notably we do not hold
 863                  * here. i915_request_add_active_barriers() may consume
 864                  * the barrier before we claim it, so we have to check
 865                  * for success.
 866                  */
 867                 engine = __barrier_to_engine(node);
 868                 smp_rmb(); /* serialise with add_active_barriers */
 869                 if (is_barrier(&node->base) &&
 870                     ____active_del_barrier(ref, node, engine))
 871                         goto match;
 872         }
 873
 874         return NULL;
 875
 876 match:
 877         spin_lock_irq(&ref->tree_lock);
 878         rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */
 879         if (p == &ref->cache->node)
 880                 WRITE_ONCE(ref->cache, NULL);
 881         spin_unlock_irq(&ref->tree_lock);
 882
 883         return rb_entry(p, struct active_node, node);
 884 }
 885
 886 int i915_active_acquire_preallocate_barrier(struct i915_active *ref,
 887                                             struct intel_engine_cs *engine)
 888 {
 889         intel_engine_mask_t tmp, mask = engine->mask;
 890         struct llist_node *first = NULL, *last = NULL;
 891         struct intel_gt *gt = engine->gt;
 892
 893         GEM_BUG_ON(i915_active_is_idle(ref));
 894
 895         /* Wait until the previous preallocation is completed */
 896         while (!llist_empty(&ref->preallocated_barriers))
 897                 cond_resched();
 898
 899         /*
 900          * Preallocate a node for each physical engine supporting the target
 901          * engine (remember virtual engines have more than one sibling).
 902          * We can then use the preallocated nodes in
 903          * i915_active_acquire_barrier()
 904          */
 905         GEM_BUG_ON(!mask);
 906         for_each_engine_masked(engine, gt, mask, tmp) {
 907                 u64 idx = engine->kernel_context->timeline->fence_context;
 908                 struct llist_node *prev = first;
 909                 struct active_node *node;
 910
 911                 rcu_read_lock();
 912                 node = reuse_idle_barrier(ref, idx);
 913                 rcu_read_unlock();
 914                 if (!node) {
 915                         node = kmem_cache_alloc(global.slab_cache, GFP_KERNEL);
 916                         if (!node)
 917                                 goto unwind;
 918
 919                         RCU_INIT_POINTER(node->base.fence, NULL);
 920                         node->base.cb.func = node_retire;
 921                         node->timeline = idx;
 922                         node->ref = ref;
 923                 }
 924
 925                 if (!i915_active_fence_isset(&node->base)) {
 926                         /*
 927                          * Mark this as being *our* unconnected proto-node.
 928                          *
 929                          * Since this node is not in any list, and we have
 930                          * decoupled it from the rbtree, we can reuse the
 931                          * request to indicate this is an idle-barrier node
 932                          * and then we can use the rb_node and list pointers
 933                          * for our tracking of the pending barrier.
 934                          */
 935                         RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
 936                         node->base.cb.node.prev = (void *)engine;
 937                         __i915_active_acquire(ref);
 938                 }
 939                 GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
 940
 941                 GEM_BUG_ON(barrier_to_engine(node) != engine);
 942                 first = barrier_to_ll(node);
 943                 first->next = prev;
 944                 if (!last)
 945                         last = first;
 946                 intel_engine_pm_get(engine);
 947         }
 948
 949         GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
 950         llist_add_batch(first, last, &ref->preallocated_barriers);
 951
 952         return 0;
 953
 954 unwind:
 955         while (first) {
 956                 struct active_node *node = barrier_from_ll(first);
 957
 958                 first = first->next;
 959
 960                 atomic_dec(&ref->count);
 961                 intel_engine_pm_put(barrier_to_engine(node));
 962
 963                 kmem_cache_free(global.slab_cache, node);
 964         }
 965         return -ENOMEM;
 966 }
 967
 968 void i915_active_acquire_barrier(struct i915_active *ref)
 969 {
 970         struct llist_node *pos, *next;
 971         unsigned long flags;
 972
 973         GEM_BUG_ON(i915_active_is_idle(ref));
 974
 975         /*
 976          * Transfer the list of preallocated barriers into the
 977          * i915_active rbtree, but only as proto-nodes. They will be
 978          * populated by i915_request_add_active_barriers() to point to the
 979          * request that will eventually release them.
 980          */
 981         llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) {
 982                 struct active_node *node = barrier_from_ll(pos);
 983                 struct intel_engine_cs *engine = barrier_to_engine(node);
 984                 struct rb_node **p, *parent;
 985
 986                 spin_lock_irqsave_nested(&ref->tree_lock, flags,
 987                                          SINGLE_DEPTH_NESTING);
 988                 parent = NULL;
 989                 p = &ref->tree.rb_node;
 990                 while (*p) {
 991                         struct active_node *it;
 992
 993                         parent = *p;
 994
 995                         it = rb_entry(parent, struct active_node, node);
 996                         if (it->timeline < node->timeline)
 997                                 p = &parent->rb_right;
 998                         else
 999                                 p = &parent->rb_left;
1000                 }
1001                 rb_link_node(&node->node, parent, p);
1002                 rb_insert_color(&node->node, &ref->tree);
1003                 spin_unlock_irqrestore(&ref->tree_lock, flags);
1004
1005                 GEM_BUG_ON(!intel_engine_pm_is_awake(engine));
1006                 llist_add(barrier_to_ll(node), &engine->barrier_tasks);
1007                 intel_engine_pm_put_delay(engine, 1);
1008         }
1009 }
1010
1011 static struct dma_fence **ll_to_fence_slot(struct llist_node *node)
1012 {
1013         return __active_fence_slot(&barrier_from_ll(node)->base);
1014 }
1015
1016 void i915_request_add_active_barriers(struct i915_request *rq)
1017 {
1018         struct intel_engine_cs *engine = rq->engine;
1019         struct llist_node *node, *next;
1020         unsigned long flags;
1021
1022         GEM_BUG_ON(!intel_context_is_barrier(rq->context));
1023         GEM_BUG_ON(intel_engine_is_virtual(engine));
1024         GEM_BUG_ON(i915_request_timeline(rq) != engine->kernel_context->timeline);
1025
1026         node = llist_del_all(&engine->barrier_tasks);
1027         if (!node)
1028                 return;
1029         /*
1030          * Attach the list of proto-fences to the in-flight request such
1031          * that the parent i915_active will be released when this request
1032          * is retired.
1033          */
1034         spin_lock_irqsave(&rq->lock, flags);
1035         llist_for_each_safe(node, next, node) {
1036                 /* serialise with reuse_idle_barrier */
1037                 smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
1038                 list_add_tail((struct list_head *)node, &rq->fence.cb_list);
1039         }
1040         spin_unlock_irqrestore(&rq->lock, flags);
1041 }
1042
1043 /*
1044  * __i915_active_fence_set: Update the last active fence along its timeline
1045  * @active: the active tracker
1046  * @fence: the new fence (under construction)
1047  *
1048  * Records the new @fence as the last active fence along its timeline in
1049  * this active tracker, moving the tracking callbacks from the previous
1050  * fence onto this one. Returns the previous fence (if not already completed),
1051  * which the caller must ensure is executed before the new fence. To ensure
1052  * that the order of fences within the timeline of the i915_active_fence is
1053  * understood, it should be locked by the caller.
1054  */
1055 struct dma_fence *
1056 __i915_active_fence_set(struct i915_active_fence *active,
1057                         struct dma_fence *fence)
1058 {
1059         struct dma_fence *prev;
1060         unsigned long flags;
1061
1062         if (fence == rcu_access_pointer(active->fence))
1063                 return fence;
1064
1065         GEM_BUG_ON(test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags));
1066
1067         /*
1068          * Consider that we have two threads arriving (A and B), with
1069          * C already resident as the active->fence.
1070          *
1071          * A does the xchg first, and so it sees C or NULL depending
1072          * on the timing of the interrupt handler. If it is NULL, the
1073          * previous fence must have been signaled and we know that
1074          * we are first on the timeline. If it is still present,
1075          * we acquire the lock on that fence and serialise with the interrupt
1076          * handler, in the process removing it from any future interrupt
1077          * callback. A will then wait on C before executing (if present).
1078          *
1079          * As B is second, it sees A as the previous fence and so waits for
1080          * it to complete its transition and takes over the occupancy for
1081          * itself -- remembering that it needs to wait on A before executing.
1082          *
1083          * Note the strong ordering of the timeline also provides consistent
1084          * nesting rules for the fence->lock; the inner lock is always the
1085          * older lock.
1086          */
1087         spin_lock_irqsave(fence->lock, flags);
1088         prev = xchg(__active_fence_slot(active), fence);
1089         if (prev) {
1090                 GEM_BUG_ON(prev == fence);
1091                 spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
1092                 __list_del_entry(&active->cb.node);
1093                 spin_unlock(prev->lock); /* serialise with prev->cb_list */
1094         }
1095         list_add_tail(&active->cb.node, &fence->cb_list);
1096         spin_unlock_irqrestore(fence->lock, flags);
1097
1098         return prev;
1099 }
1100
1101 int i915_active_fence_set(struct i915_active_fence *active,
1102                           struct i915_request *rq)
1103 {
1104         struct dma_fence *fence;
1105         int err = 0;
1106
1107         /* Must maintain timeline ordering wrt previous active requests */
1108         rcu_read_lock();
1109         fence = __i915_active_fence_set(active, &rq->fence);
1110         if (fence) /* but the previous fence may not belong to that timeline! */
1111                 fence = dma_fence_get_rcu(fence);
1112         rcu_read_unlock();
1113         if (fence) {
1114                 err = i915_request_await_dma_fence(rq, fence);
1115                 dma_fence_put(fence);
1116         }
1117
1118         return err;
1119 }
1120
1121 void i915_active_noop(struct dma_fence *fence, struct dma_fence_cb *cb)
1122 {
1123         active_fence_cb(fence, cb);
1124 }
1125
1126 struct auto_active {
1127         struct i915_active base;
1128         struct kref ref;
1129 };
1130
1131 struct i915_active *i915_active_get(struct i915_active *ref)
1132 {
1133         struct auto_active *aa = container_of(ref, typeof(*aa), base);
1134
1135         kref_get(&aa->ref);
1136         return &aa->base;
1137 }
1138
1139 static void auto_release(struct kref *ref)
1140 {
1141         struct auto_active *aa = container_of(ref, typeof(*aa), ref);
1142
1143         i915_active_fini(&aa->base);
1144         kfree(aa);
1145 }
1146
1147 void i915_active_put(struct i915_active *ref)
1148 {
1149         struct auto_active *aa = container_of(ref, typeof(*aa), base);
1150
1151         kref_put(&aa->ref, auto_release);
1152 }
1153
1154 static int auto_active(struct i915_active *ref)
1155 {
1156         i915_active_get(ref);
1157         return 0;
1158 }
1159
1160 static void auto_retire(struct i915_active *ref)
1161 {
1162         i915_active_put(ref);
1163 }
1164
1165 struct i915_active *i915_active_create(void)
1166 {
1167         struct auto_active *aa;
1168
1169         aa = kmalloc(sizeof(*aa), GFP_KERNEL);
1170         if (!aa)
1171                 return NULL;
1172
1173         kref_init(&aa->ref);
1174         i915_active_init(&aa->base, auto_active, auto_retire);
1175
1176         return &aa->base;
1177 }
1178
1179 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1180 #include "selftests/i915_active.c"
1181 #endif
1182
1183 static void i915_global_active_shrink(void)
1184 {
1185         kmem_cache_shrink(global.slab_cache);
1186 }
1187
1188 static void i915_global_active_exit(void)
1189 {
1190         kmem_cache_destroy(global.slab_cache);
1191 }
1192
1193 static struct i915_global_active global = { {
1194         .shrink = i915_global_active_shrink,
1195         .exit = i915_global_active_exit,
1196 } };
1197
1198 int __init i915_global_active_init(void)
1199 {
1200         global.slab_cache = KMEM_CACHE(active_node, SLAB_HWCACHE_ALIGN);
1201         if (!global.slab_cache)
1202                 return -ENOMEM;
1203
1204         i915_global_register(&global.base);
1205         return 0;
1206 }