2016484 - High-bitrate AV1 freezes: DAV1DDecoder::InvokeDecode busy-waits on dav1d EAGAIN instead of draining frames

Status: NEW

(Core :: Audio/Video: Playback, defect, P2)

Description

[bran.perch-07]

Summary

High-bitrate AV1 video (30+ Mbps, 4K) freezes after a few seconds of playback when using the dav1d software decoder. The freeze timing varies with bitrate but is consistently reproducible.

Root Cause

DAV1DDecoder::InvokeDecode has a do/while loop that calls dav1d_send_data followed by a single dav1d_get_picture. When dav1d_send_data returns DAV1D_ERR(EAGAIN) (internal buffer full, no data consumed) and dav1d_get_picture also returns DAV1D_ERR(EAGAIN) (no completed frame yet), the continue statement loops back to retry dav1d_send_data immediately -- but nothing has changed. This creates a tight busy-wait spin loop with no yield that competes with dav1d's own frame decode threads for CPU time.

On high-bitrate content where software AV1 decode already needs every CPU cycle, this busy-wait starves the decode threads, preventing them from completing the frames that would unblock the pipeline. The result is a vicious cycle where decode falls further behind until playback freezes.

Two issues in the current code:

1. Single frame drain per iteration: Only one GetPicture() call per loop iteration, even when dav1d_send_data returned EAGAIN (buffer full). When the buffer is full there may be multiple completed frames waiting -- retrieving only one means multiple outer loop iterations with failed dav1d_send_data calls before the buffer has room.

2. No yield on double-EAGAIN: When both dav1d_send_data and dav1d_get_picture return EAGAIN, the loop spins with no yield. Frames are in-flight on dav1d's decode threads but the busy-wait contends for CPU time those threads need to finish.

Fix

Replace the single GetPicture() call with a draining loop (matching the pattern already used in DAV1DDecoder::Drain()), and add std::this_thread::yield() when dav1d_send_data returned EAGAIN and no frames were available to drain.

Steps to Reproduce

1. Play a high-bitrate AV1 video (30+ Mbps, 4K) with hardware AV1 decode disabled (to force dav1d software path)
2. Video freezes within a few seconds
3. Lower bitrate content may not trigger the issue because dav1d can keep up without EAGAIN contention

Comment 1

[bran.perch-07]

Attachment: Drain all buffered frames from dav1d on EAGAIN to prevent busy-wait stall

Comment 2

[bran.perch-07]

Updated patch attached (v2). The original fix was correct in concept (drain loop for EAGAIN), but review identified several hardening issues:

1. The while(true) drain loop was unbounded. Replaced with a bounded for loop (max 256 iterations) as defense-in-depth. Realistic values are under 20 (bounded by dav1d max_frame_delay).

2. std::this_thread::yield() was the wrong primitive for the EAGAIN-with-no-pictures case. Per the dav1d API contract, when dav1d_send_data returns EAGAIN, at least one picture MUST be available. The unreachable state is better served by MOZ_DIAGNOSTIC_ASSERT (crashes Nightly/early-Beta for detection, no-op in Release).

3. The Drain() method at end-of-stream had the same unbounded while(true) pattern. Now also bounded to 256 iterations for consistency. Note: the Drain() path does not get the MOZ_DIAGNOSTIC_ASSERT or hard error on limit since at end-of-stream the number of remaining frames is inherently bounded by max_frame_delay and a partial result is preferable to a hard error.

4. The diagnostic assert now checks per-iteration results count rather than cumulative total, to correctly validate the dav1d API contract on each outer loop iteration.

5. Removed #include <thread> and the drained flag from the original patch (no longer needed without std::this_thread::yield()).

Testing:
Tested with existing media playback tests. The EAGAIN state is difficult to reproduce in unit tests because it requires specific frame delay configurations and high-bitrate streams that fill dav1d internal buffer. Manual testing with 30+ Mbps 4K AV1 content confirmed the drain loop works correctly. Will include try push results before requesting review.

Comment 3

[bran.perch-07]

Attachment: bug-2016484-drain-dav1d-eagain-v2.patch bounded drain loop + MOZ_DIAGNOSTIC_ASSERT

v2 patch  see previous comment for changelog.

Comment 5

[C.M.Chang[:chunmin]]

I'll investigate this. Keep the NI for now.