avformat/shared: stale BLOCK_PENDING never times out in non-blocking mode

[ronag]

Summary

The fix for #22 returns AVERROR(EAGAIN) from the BLOCK_PENDING wait when AVIO_FLAG_NONBLOCK is set. However, pending_since is a local in shared_read(), so it resets on every call. The cache_timeout expiry check can only fire within a single call, and a nonblocking call never gets past its first iteration: it sets pending_since and returns EAGAIN.

If the owner died while holding BLOCK_PENDING, which is persistent state in the spacemap, a nonblocking reader gets EAGAIN forever. A blocking reader recovers after cache_timeout.

Location

• pending_since local:

  FFmpeg/libavformat/shared.c

  Line 660 in fe63f8f

  int64_t pending_since = 0;

• Wait loop:

  FFmpeg/libavformat/shared.c

  Lines 764 to 786 in fe63f8f

  	case BLOCK_PENDING:
  	/* Another thread is busy fetching this block, wait for it to finish */
  	if (!s->timeout) {
  	break; /* no timeout requested, immediately race to fetch block */
  	} else if (pending_since) {
  	int64_t new = av_gettime_relative();
  	if (new - pending_since >= s->timeout)
  	break; /* timeout expired, try to fetch the block ourselves */
  	} else {
  	pending_since = av_gettime_relative();
  	}
  	if (h->flags & AVIO_FLAG_NONBLOCK)
  	return AVERROR(EAGAIN);
  	/* Make sure we try a few times before giving up */
  	av_usleep(FFMIN(s->timeout >> 4, 10000));
  	if (ff_check_interrupt(&h->interrupt_callback))
  	return AVERROR_EXIT;
  	state = atomic_load_explicit(&block->state, memory_order_acquire);
  	goto retry;
  	}

Reproduction

Cache block 0, then simulate a crashed owner by writing BLOCK_PENDING (1) into block 1's slot in the .spacemap (byte offset 128 + 4). Then seek to 32768 and read with cache_timeout=100000 (harness from #71, with AVIO_FLAG_NONBLOCK added to the open flags):

blocking:          [0 EAGAINs, 0.108s] read: 4 (ok)
AVIO_FLAG_NONBLOCK: [2375 EAGAINs, 3.000s] read: -35 (Resource temporarily unavailable)   # never recovers

Suggested fix

Keep the wait start in SharedContext, together with the block id it applies to. Reset it when the reader moves to a different block or the block leaves BLOCK_PENDING.

[haasn]

Real bug but unlikely to affect anything in production