drivers/dma-buf/dma-fence.c

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