drivers/dma-buf/dma-fence.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Fence mechanism for dma-buf and to allow for asynchronous dma access
   4  *
   5  * Copyright (C) 2012 Canonical Ltd
   6  * Copyright (C) 2012 Texas Instruments
   7  *
   8  * Authors:
   9  * Rob Clark <robdclark@gmail.com>
  10  * Maarten Lankhorst <maarten.lankhorst@canonical.com>
  11  */
  12
  13 #include <linux/slab.h>
  14 #include <linux/export.h>
  15 #include <linux/atomic.h>
  16 #include <linux/dma-fence.h>
  17 #include <linux/sched/signal.h>
  18
  19 #define CREATE_TRACE_POINTS
  20 #include <trace/events/dma_fence.h>
  21
  22 EXPORT_TRACEPOINT_SYMBOL(dma_fence_emit);
  23 EXPORT_TRACEPOINT_SYMBOL(dma_fence_enable_signal);
  24 EXPORT_TRACEPOINT_SYMBOL(dma_fence_signaled);
  25
  26 static DEFINE_SPINLOCK(dma_fence_stub_lock);
  27 static struct dma_fence dma_fence_stub;
  28
  29 /*
  30  * fence context counter: each execution context should have its own
  31  * fence context, this allows checking if fences belong to the same
  32  * context or not. One device can have multiple separate contexts,
  33  * and they're used if some engine can run independently of another.
  34  */
  35 static atomic64_t dma_fence_context_counter = ATOMIC64_INIT(1);
  36
  37 /**
  38  * DOC: DMA fences overview
  39  *
  40  * DMA fences, represented by &struct dma_fence, are the kernel internal
  41  * synchronization primitive for DMA operations like GPU rendering, video
  42  * encoding/decoding, or displaying buffers on a screen.
  43  *
  44  * A fence is initialized using dma_fence_init() and completed using
  45  * dma_fence_signal(). Fences are associated with a context, allocated through
  46  * dma_fence_context_alloc(), and all fences on the same context are
  47  * fully ordered.
  48  *
  49  * Since the purposes of fences is to facilitate cross-device and
  50  * cross-application synchronization, there's multiple ways to use one:
  51  *
  52  * - Individual fences can be exposed as a &sync_file, accessed as a file
  53  *   descriptor from userspace, created by calling sync_file_create(). This is
  54  *   called explicit fencing, since userspace passes around explicit
  55  *   synchronization points.
  56  *
  57  * - Some subsystems also have their own explicit fencing primitives, like
  58  *   &drm_syncobj. Compared to &sync_file, a &drm_syncobj allows the underlying
  59  *   fence to be updated.
  60  *
  61  * - Then there's also implicit fencing, where the synchronization points are
  62  *   implicitly passed around as part of shared &dma_buf instances. Such
  63  *   implicit fences are stored in &struct dma_resv through the
  64  *   &dma_buf.resv pointer.
  65  */
  66
  67 static const char *dma_fence_stub_get_name(struct dma_fence *fence)
  68 {
  69         return "stub";
  70 }
  71
  72 static const struct dma_fence_ops dma_fence_stub_ops = {
  73         .get_driver_name = dma_fence_stub_get_name,
  74         .get_timeline_name = dma_fence_stub_get_name,
  75 };
  76
  77 /**
  78  * dma_fence_get_stub - return a signaled fence
  79  *
  80  * Return a stub fence which is already signaled.
  81  */
  82 struct dma_fence *dma_fence_get_stub(void)
  83 {
  84         spin_lock(&dma_fence_stub_lock);
  85         if (!dma_fence_stub.ops) {
  86                 dma_fence_init(&dma_fence_stub,
  87                                &dma_fence_stub_ops,
  88                                &dma_fence_stub_lock,
  89                                0, 0);
  90                 dma_fence_signal_locked(&dma_fence_stub);
  91         }
  92         spin_unlock(&dma_fence_stub_lock);
  93
  94         return dma_fence_get(&dma_fence_stub);
  95 }
  96 EXPORT_SYMBOL(dma_fence_get_stub);
  97
  98 /**
  99  * dma_fence_context_alloc - allocate an array of fence contexts
 100  * @num: amount of contexts to allocate
 101  *
 102  * This function will return the first index of the number of fence contexts
 103  * allocated.  The fence context is used for setting &dma_fence.context to a
 104  * unique number by passing the context to dma_fence_init().
 105  */
 106 u64 dma_fence_context_alloc(unsigned num)
 107 {
 108         WARN_ON(!num);
 109         return atomic64_add_return(num, &dma_fence_context_counter) - num;
 110 }
 111 EXPORT_SYMBOL(dma_fence_context_alloc);
 112
 113 /**
 114  * dma_fence_signal_locked - signal completion of a fence
 115  * @fence: the fence to signal
 116  *
 117  * Signal completion for software callbacks on a fence, this will unblock
 118  * dma_fence_wait() calls and run all the callbacks added with
 119  * dma_fence_add_callback(). Can be called multiple times, but since a fence
 120  * can only go from the unsignaled to the signaled state and not back, it will
 121  * only be effective the first time.
 122  *
 123  * Unlike dma_fence_signal(), this function must be called with &dma_fence.lock
 124  * held.
 125  *
 126  * Returns 0 on success and a negative error value when @fence has been
 127  * signalled already.
 128  */
 129 int dma_fence_signal_locked(struct dma_fence *fence)
 130 {
 131         struct dma_fence_cb *cur, *tmp;
 132         struct list_head cb_list;
 133
 134         lockdep_assert_held(fence->lock);
 135
 136         if (unlikely(test_and_set_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
 137                                       &fence->flags)))
 138                 return -EINVAL;
 139
 140         /* Stash the cb_list before replacing it with the timestamp */
 141         list_replace(&fence->cb_list, &cb_list);
 142
 143         fence->timestamp = ktime_get();
 144         set_bit(DMA_FENCE_FLAG_TIMESTAMP_BIT, &fence->flags);
 145         trace_dma_fence_signaled(fence);
 146
 147         list_for_each_entry_safe(cur, tmp, &cb_list, node) {
 148                 INIT_LIST_HEAD(&cur->node);
 149                 cur->func(fence, cur);
 150         }
 151
 152         return 0;
 153 }
 154 EXPORT_SYMBOL(dma_fence_signal_locked);
 155
 156 /**
 157  * dma_fence_signal - signal completion of a fence
 158  * @fence: the fence to signal
 159  *
 160  * Signal completion for software callbacks on a fence, this will unblock
 161  * dma_fence_wait() calls and run all the callbacks added with
 162  * dma_fence_add_callback(). Can be called multiple times, but since a fence
 163  * can only go from the unsignaled to the signaled state and not back, it will
 164  * only be effective the first time.
 165  *
 166  * Returns 0 on success and a negative error value when @fence has been
 167  * signalled already.
 168  */
 169 int dma_fence_signal(struct dma_fence *fence)
 170 {
 171         unsigned long flags;
 172         int ret;
 173
 174         if (!fence)
 175                 return -EINVAL;
 176
 177         spin_lock_irqsave(fence->lock, flags);
 178         ret = dma_fence_signal_locked(fence);
 179         spin_unlock_irqrestore(fence->lock, flags);
 180
 181         return ret;
 182 }
 183 EXPORT_SYMBOL(dma_fence_signal);
 184
 185 /**
 186  * dma_fence_wait_timeout - sleep until the fence gets signaled
 187  * or until timeout elapses
 188  * @fence: the fence to wait on
 189  * @intr: if true, do an interruptible wait
 190  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
 191  *
 192  * Returns -ERESTARTSYS if interrupted, 0 if the wait timed out, or the
 193  * remaining timeout in jiffies on success. Other error values may be
 194  * returned on custom implementations.
 195  *
 196  * Performs a synchronous wait on this fence. It is assumed the caller
 197  * directly or indirectly (buf-mgr between reservation and committing)
 198  * holds a reference to the fence, otherwise the fence might be
 199  * freed before return, resulting in undefined behavior.
 200  *
 201  * See also dma_fence_wait() and dma_fence_wait_any_timeout().
 202  */
 203 signed long
 204 dma_fence_wait_timeout(struct dma_fence *fence, bool intr, signed long timeout)
 205 {
 206         signed long ret;
 207
 208         if (WARN_ON(timeout < 0))
 209                 return -EINVAL;
 210
 211         trace_dma_fence_wait_start(fence);
 212         if (fence->ops->wait)
 213                 ret = fence->ops->wait(fence, intr, timeout);
 214         else
 215                 ret = dma_fence_default_wait(fence, intr, timeout);
 216         trace_dma_fence_wait_end(fence);
 217         return ret;
 218 }
 219 EXPORT_SYMBOL(dma_fence_wait_timeout);
 220
 221 /**
 222  * dma_fence_release - default relese function for fences
 223  * @kref: &dma_fence.recfount
 224  *
 225  * This is the default release functions for &dma_fence. Drivers shouldn't call
 226  * this directly, but instead call dma_fence_put().
 227  */
 228 void dma_fence_release(struct kref *kref)
 229 {
 230         struct dma_fence *fence =
 231                 container_of(kref, struct dma_fence, refcount);
 232
 233         trace_dma_fence_destroy(fence);
 234
 235         if (WARN(!list_empty(&fence->cb_list) &&
 236                  !test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags),
 237                  "Fence %s:%s:%llx:%llx released with pending signals!\n",
 238                  fence->ops->get_driver_name(fence),
 239                  fence->ops->get_timeline_name(fence),
 240                  fence->context, fence->seqno)) {
 241                 unsigned long flags;
 242
 243                 /*
 244                  * Failed to signal before release, likely a refcounting issue.
 245                  *
 246                  * This should never happen, but if it does make sure that we
 247                  * don't leave chains dangling. We set the error flag first
 248                  * so that the callbacks know this signal is due to an error.
 249                  */
 250                 spin_lock_irqsave(fence->lock, flags);
 251                 fence->error = -EDEADLK;
 252                 dma_fence_signal_locked(fence);
 253                 spin_unlock_irqrestore(fence->lock, flags);
 254         }
 255
 256         if (fence->ops->release)
 257                 fence->ops->release(fence);
 258         else
 259                 dma_fence_free(fence);
 260 }
 261 EXPORT_SYMBOL(dma_fence_release);
 262
 263 /**
 264  * dma_fence_free - default release function for &dma_fence.
 265  * @fence: fence to release
 266  *
 267  * This is the default implementation for &dma_fence_ops.release. It calls
 268  * kfree_rcu() on @fence.
 269  */
 270 void dma_fence_free(struct dma_fence *fence)
 271 {
 272         kfree_rcu(fence, rcu);
 273 }
 274 EXPORT_SYMBOL(dma_fence_free);
 275
 276 static bool __dma_fence_enable_signaling(struct dma_fence *fence)
 277 {
 278         bool was_set;
 279
 280         lockdep_assert_held(fence->lock);
 281
 282         was_set = test_and_set_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
 283                                    &fence->flags);
 284
 285         if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
 286                 return false;
 287
 288         if (!was_set && fence->ops->enable_signaling) {
 289                 trace_dma_fence_enable_signal(fence);
 290
 291                 if (!fence->ops->enable_signaling(fence)) {
 292                         dma_fence_signal_locked(fence);
 293                         return false;
 294                 }
 295         }
 296
 297         return true;
 298 }
 299
 300 /**
 301  * dma_fence_enable_sw_signaling - enable signaling on fence
 302  * @fence: the fence to enable
 303  *
 304  * This will request for sw signaling to be enabled, to make the fence
 305  * complete as soon as possible. This calls &dma_fence_ops.enable_signaling
 306  * internally.
 307  */
 308 void dma_fence_enable_sw_signaling(struct dma_fence *fence)
 309 {
 310         unsigned long flags;
 311
 312         if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
 313                 return;
 314
 315         spin_lock_irqsave(fence->lock, flags);
 316         __dma_fence_enable_signaling(fence);
 317         spin_unlock_irqrestore(fence->lock, flags);
 318 }
 319 EXPORT_SYMBOL(dma_fence_enable_sw_signaling);
 320
 321 /**
 322  * dma_fence_add_callback - add a callback to be called when the fence
 323  * is signaled
 324  * @fence: the fence to wait on
 325  * @cb: the callback to register
 326  * @func: the function to call
 327  *
 328  * @cb will be initialized by dma_fence_add_callback(), no initialization
 329  * by the caller is required. Any number of callbacks can be registered
 330  * to a fence, but a callback can only be registered to one fence at a time.
 331  *
 332  * Note that the callback can be called from an atomic context.  If
 333  * fence is already signaled, this function will return -ENOENT (and
 334  * *not* call the callback).
 335  *
 336  * Add a software callback to the fence. Same restrictions apply to
 337  * refcount as it does to dma_fence_wait(), however the caller doesn't need to
 338  * keep a refcount to fence afterward dma_fence_add_callback() has returned:
 339  * when software access is enabled, the creator of the fence is required to keep
 340  * the fence alive until after it signals with dma_fence_signal(). The callback
 341  * itself can be called from irq context.
 342  *
 343  * Returns 0 in case of success, -ENOENT if the fence is already signaled
 344  * and -EINVAL in case of error.
 345  */
 346 int dma_fence_add_callback(struct dma_fence *fence, struct dma_fence_cb *cb,
 347                            dma_fence_func_t func)
 348 {
 349         unsigned long flags;
 350         int ret = 0;
 351
 352         if (WARN_ON(!fence || !func))
 353                 return -EINVAL;
 354
 355         if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
 356                 INIT_LIST_HEAD(&cb->node);
 357                 return -ENOENT;
 358         }
 359
 360         spin_lock_irqsave(fence->lock, flags);
 361
 362         if (__dma_fence_enable_signaling(fence)) {
 363                 cb->func = func;
 364                 list_add_tail(&cb->node, &fence->cb_list);
 365         } else {
 366                 INIT_LIST_HEAD(&cb->node);
 367                 ret = -ENOENT;
 368         }
 369
 370         spin_unlock_irqrestore(fence->lock, flags);
 371
 372         return ret;
 373 }
 374 EXPORT_SYMBOL(dma_fence_add_callback);
 375
 376 /**
 377  * dma_fence_get_status - returns the status upon completion
 378  * @fence: the dma_fence to query
 379  *
 380  * This wraps dma_fence_get_status_locked() to return the error status
 381  * condition on a signaled fence. See dma_fence_get_status_locked() for more
 382  * details.
 383  *
 384  * Returns 0 if the fence has not yet been signaled, 1 if the fence has
 385  * been signaled without an error condition, or a negative error code
 386  * if the fence has been completed in err.
 387  */
 388 int dma_fence_get_status(struct dma_fence *fence)
 389 {
 390         unsigned long flags;
 391         int status;
 392
 393         spin_lock_irqsave(fence->lock, flags);
 394         status = dma_fence_get_status_locked(fence);
 395         spin_unlock_irqrestore(fence->lock, flags);
 396
 397         return status;
 398 }
 399 EXPORT_SYMBOL(dma_fence_get_status);
 400
 401 /**
 402  * dma_fence_remove_callback - remove a callback from the signaling list
 403  * @fence: the fence to wait on
 404  * @cb: the callback to remove
 405  *
 406  * Remove a previously queued callback from the fence. This function returns
 407  * true if the callback is successfully removed, or false if the fence has
 408  * already been signaled.
 409  *
 410  * *WARNING*:
 411  * Cancelling a callback should only be done if you really know what you're
 412  * doing, since deadlocks and race conditions could occur all too easily. For
 413  * this reason, it should only ever be done on hardware lockup recovery,
 414  * with a reference held to the fence.
 415  *
 416  * Behaviour is undefined if @cb has not been added to @fence using
 417  * dma_fence_add_callback() beforehand.
 418  */
 419 bool
 420 dma_fence_remove_callback(struct dma_fence *fence, struct dma_fence_cb *cb)
 421 {
 422         unsigned long flags;
 423         bool ret;
 424
 425         spin_lock_irqsave(fence->lock, flags);
 426
 427         ret = !list_empty(&cb->node);
 428         if (ret)
 429                 list_del_init(&cb->node);
 430
 431         spin_unlock_irqrestore(fence->lock, flags);
 432
 433         return ret;
 434 }
 435 EXPORT_SYMBOL(dma_fence_remove_callback);
 436
 437 struct default_wait_cb {
 438         struct dma_fence_cb base;
 439         struct task_struct *task;
 440 };
 441
 442 static void
 443 dma_fence_default_wait_cb(struct dma_fence *fence, struct dma_fence_cb *cb)
 444 {
 445         struct default_wait_cb *wait =
 446                 container_of(cb, struct default_wait_cb, base);
 447
 448         wake_up_state(wait->task, TASK_NORMAL);
 449 }
 450
 451 /**
 452  * dma_fence_default_wait - default sleep until the fence gets signaled
 453  * or until timeout elapses
 454  * @fence: the fence to wait on
 455  * @intr: if true, do an interruptible wait
 456  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
 457  *
 458  * Returns -ERESTARTSYS if interrupted, 0 if the wait timed out, or the
 459  * remaining timeout in jiffies on success. If timeout is zero the value one is
 460  * returned if the fence is already signaled for consistency with other
 461  * functions taking a jiffies timeout.
 462  */
 463 signed long
 464 dma_fence_default_wait(struct dma_fence *fence, bool intr, signed long timeout)
 465 {
 466         struct default_wait_cb cb;
 467         unsigned long flags;
 468         signed long ret = timeout ? timeout : 1;
 469
 470         if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags))
 471                 return ret;
 472
 473         spin_lock_irqsave(fence->lock, flags);
 474
 475         if (intr && signal_pending(current)) {
 476                 ret = -ERESTARTSYS;
 477                 goto out;
 478         }
 479
 480         if (!__dma_fence_enable_signaling(fence))
 481                 goto out;
 482
 483         if (!timeout) {
 484                 ret = 0;
 485                 goto out;
 486         }
 487
 488         cb.base.func = dma_fence_default_wait_cb;
 489         cb.task = current;
 490         list_add(&cb.base.node, &fence->cb_list);
 491
 492         while (!test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags) && ret > 0) {
 493                 if (intr)
 494                         __set_current_state(TASK_INTERRUPTIBLE);
 495                 else
 496                         __set_current_state(TASK_UNINTERRUPTIBLE);
 497                 spin_unlock_irqrestore(fence->lock, flags);
 498
 499                 ret = schedule_timeout(ret);
 500
 501                 spin_lock_irqsave(fence->lock, flags);
 502                 if (ret > 0 && intr && signal_pending(current))
 503                         ret = -ERESTARTSYS;
 504         }
 505
 506         if (!list_empty(&cb.base.node))
 507                 list_del(&cb.base.node);
 508         __set_current_state(TASK_RUNNING);
 509
 510 out:
 511         spin_unlock_irqrestore(fence->lock, flags);
 512         return ret;
 513 }
 514 EXPORT_SYMBOL(dma_fence_default_wait);
 515
 516 static bool
 517 dma_fence_test_signaled_any(struct dma_fence **fences, uint32_t count,
 518                             uint32_t *idx)
 519 {
 520         int i;
 521
 522         for (i = 0; i < count; ++i) {
 523                 struct dma_fence *fence = fences[i];
 524                 if (test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags)) {
 525                         if (idx)
 526                                 *idx = i;
 527                         return true;
 528                 }
 529         }
 530         return false;
 531 }
 532
 533 /**
 534  * dma_fence_wait_any_timeout - sleep until any fence gets signaled
 535  * or until timeout elapses
 536  * @fences: array of fences to wait on
 537  * @count: number of fences to wait on
 538  * @intr: if true, do an interruptible wait
 539  * @timeout: timeout value in jiffies, or MAX_SCHEDULE_TIMEOUT
 540  * @idx: used to store the first signaled fence index, meaningful only on
 541  *      positive return
 542  *
 543  * Returns -EINVAL on custom fence wait implementation, -ERESTARTSYS if
 544  * interrupted, 0 if the wait timed out, or the remaining timeout in jiffies
 545  * on success.
 546  *
 547  * Synchronous waits for the first fence in the array to be signaled. The
 548  * caller needs to hold a reference to all fences in the array, otherwise a
 549  * fence might be freed before return, resulting in undefined behavior.
 550  *
 551  * See also dma_fence_wait() and dma_fence_wait_timeout().
 552  */
 553 signed long
 554 dma_fence_wait_any_timeout(struct dma_fence **fences, uint32_t count,
 555                            bool intr, signed long timeout, uint32_t *idx)
 556 {
 557         struct default_wait_cb *cb;
 558         signed long ret = timeout;
 559         unsigned i;
 560
 561         if (WARN_ON(!fences || !count || timeout < 0))
 562                 return -EINVAL;
 563
 564         if (timeout == 0) {
 565                 for (i = 0; i < count; ++i)
 566                         if (dma_fence_is_signaled(fences[i])) {
 567                                 if (idx)
 568                                         *idx = i;
 569                                 return 1;
 570                         }
 571
 572                 return 0;
 573         }
 574
 575         cb = kcalloc(count, sizeof(struct default_wait_cb), GFP_KERNEL);
 576         if (cb == NULL) {
 577                 ret = -ENOMEM;
 578                 goto err_free_cb;
 579         }
 580
 581         for (i = 0; i < count; ++i) {
 582                 struct dma_fence *fence = fences[i];
 583
 584                 cb[i].task = current;
 585                 if (dma_fence_add_callback(fence, &cb[i].base,
 586                                            dma_fence_default_wait_cb)) {
 587                         /* This fence is already signaled */
 588                         if (idx)
 589                                 *idx = i;
 590                         goto fence_rm_cb;
 591                 }
 592         }
 593
 594         while (ret > 0) {
 595                 if (intr)
 596                         set_current_state(TASK_INTERRUPTIBLE);
 597                 else
 598                         set_current_state(TASK_UNINTERRUPTIBLE);
 599
 600                 if (dma_fence_test_signaled_any(fences, count, idx))
 601                         break;
 602
 603                 ret = schedule_timeout(ret);
 604
 605                 if (ret > 0 && intr && signal_pending(current))
 606                         ret = -ERESTARTSYS;
 607         }
 608
 609         __set_current_state(TASK_RUNNING);
 610
 611 fence_rm_cb:
 612         while (i-- > 0)
 613                 dma_fence_remove_callback(fences[i], &cb[i].base);
 614
 615 err_free_cb:
 616         kfree(cb);
 617
 618         return ret;
 619 }
 620 EXPORT_SYMBOL(dma_fence_wait_any_timeout);
 621
 622 /**
 623  * dma_fence_init - Initialize a custom fence.
 624  * @fence: the fence to initialize
 625  * @ops: the dma_fence_ops for operations on this fence
 626  * @lock: the irqsafe spinlock to use for locking this fence
 627  * @context: the execution context this fence is run on
 628  * @seqno: a linear increasing sequence number for this context
 629  *
 630  * Initializes an allocated fence, the caller doesn't have to keep its
 631  * refcount after committing with this fence, but it will need to hold a
 632  * refcount again if &dma_fence_ops.enable_signaling gets called.
 633  *
 634  * context and seqno are used for easy comparison between fences, allowing
 635  * to check which fence is later by simply using dma_fence_later().
 636  */
 637 void
 638 dma_fence_init(struct dma_fence *fence, const struct dma_fence_ops *ops,
 639                spinlock_t *lock, u64 context, u64 seqno)
 640 {
 641         BUG_ON(!lock);
 642         BUG_ON(!ops || !ops->get_driver_name || !ops->get_timeline_name);
 643
 644         kref_init(&fence->refcount);
 645         fence->ops = ops;
 646         INIT_LIST_HEAD(&fence->cb_list);
 647         fence->lock = lock;
 648         fence->context = context;
 649         fence->seqno = seqno;
 650         fence->flags = 0UL;
 651         fence->error = 0;
 652
 653         trace_dma_fence_init(fence);
 654 }
 655 EXPORT_SYMBOL(dma_fence_init);