Repository navigation
Reliability and safety fixes for the receiver and the recovery phase. The wire protocol is unchanged (v3), so 1.2.0 interoperates with 1.1.x.
Fixed
- Recovery no longer gives up on a responsive sender. The receiver used to send its whole RESEND burst without reading replies; with many missing parts the burst outlasted
--ttland the transfer failed even though the sender had answered everything. - A receiver started mid-stream now joins the transfer. The sender repeats its ANNOUNCE about once a second for the whole stream, not just the first second.
- A complete file is no longer thrown away when every FINISH is lost. The receiver verifies as soon as the last part lands.
- A resumed receive no longer times out while the sender re-streams the part it already holds.
- One forged ANNOUNCE can no longer truncate or delete a
.part/.part.idxthe receiver did not create, including a real--resumesnapshot. - The resume index is written atomically, and a failed index write no longer deletes the received data.
- Windows: the sender no longer leaves the resend phase early when a read fails because of the datagram rather than silence.
- A flood of rejected announcements is reported at most once a second instead of one line per packet.
- A timeout always says why the receiver exited.
Behaviour changes
- A leftover
<name>.partor<name>.part.idxthat does not match the announced transfer is left untouched and the announcement is ignored with a warning. Delete the file or pass--overwrite. --overwritealso replaces such a stale.part.- An existing output file is refused before anything is received, not after the whole transfer.
- On macOS and Linux an announced name that is not valid UTF-8 is saved as
file.out.
Full Changelog: v1.1.0...v1.2.0