**2.4 — buffer geometry correctness on the Linux CPU path.**
L5: the self-mmap de-pad mapped the fd from offset 0 and then indexed by `chunk.offset()` ALONE,
ignoring `spa_data.mapoffset` — where that spa_data's region actually begins inside the fd. For a
pooled-memfd producer (every buffer a slice of one fd) it therefore read the WRONG buffer, and the
`needed > avail` guard could not catch it: `avail` came from the whole-fd mapping, so it was ample
for any single buffer's span. PipeWire's own MAP_BUFFERS slice (the fallback) was always correct
because it already starts at `mapoffset` — only the hand-rolled "fix" path was wrong, so the two
paths now compute their base separately (checked add; both halves are producer-controlled).
L15: when `fstat` failed, the mmap length was invented as `offset + needed` — producer-controlled
geometry. That defeats the entire point of the fstat (the "buffer smaller than the frame span"
guard then compares against the number the producer just supplied) and can map, and read, past the
end of the object — a SIGBUS, not an `Err`. Without a real length we now decline to self-map and
let PipeWire's own slice serve, which is bounded by construction.
L12: `bytes_per_pixel()`'s catch-all answers 4 for NV12, so a producer-native NV12 buffer computed
`row = 4w` against a stride of ~w and ALWAYS tripped the stride guard — reporting "chunk stride <
row", i.e. blaming the producer for a host limitation. NV12 cannot be de-padded on this path at all
(plane 1 is not even in `datas[0]`'s span), so reject it with a message that says which side is at
fault and under what condition the NV12 offer is valid.
L14: `spa_meta_bitmap` was read field-by-field through an aligned `*const` at a
PRODUCER-controlled offset, with a SAFETY comment asserting an alignment the code never
established. One `read_unaligned` of the `Copy` POD instead.
**2.6 — hand-off semantics (L9, L10).** The frame channel was `sync_channel(8)` + `try_send`, i.e.
drop-NEWEST, while the trait documented drop-oldest. Neither remedy the plan offered works: a
`SyncSender` cannot pop, so "drain one and retry" is not expressible, and shrinking the bound makes
staleness WORSE, not better (with bound N the queue holds the first N frames of a stall and
everything after is discarded, so a larger N actually yields a fresher frame — the defect is the
discarding, not the depth). Replaced with a one-deep OVERWRITING mailbox plus a depth-1 wakeup
channel: publishing overwrites, so a stalled consumer loses the intermediate frames and is always
handed the freshest one. `next_frame` used to return the OLDEST of eight stale frames.
Falls out of it:
- `wait_arrival` now honours its documented "must NOT consume" contract by peeking the slot. The
`pending` field existed only to stash a frame the un-peekable channel forced it to consume;
it is gone.
- a queued `CapturedFrame` can own a dup'd dmabuf fd or a CUDA buffer, so the old 8-deep backlog
could pin eight compositor buffers. One-deep by construction now.
- L10: `set_active(false)` flushes the mailbox, so a reused capturer no longer opens the next
stream with the previous session's last frame (wrong content, and a `pts_ns` from the old
clock). Two lines, where flushing a queue + a `pending` stash was fiddlier.
- the producer-liveness signal moves to the wakeup channel's `Disconnected`, and a final frame
left in the slot as the thread exits is still served before the error.
pf-capture 13/13; workspace clippy --all-targets clean on Linux and windows-msvc.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
enricobuehlerandClaude Opus 5
759fdf767c