[ThomasWaldmann]

Fixes borgbackup/borg#10566: a daemonized borg mount (2.0.0b24+, rest/ssh stdio repo) never exits after borg umount.

Cause

borg mount opens the store and then forks to daemonize. The StdioSession stderr drain thread does not survive fork(), so in the daemon:

• close() hangs: at fork time the drain thread was blocked in read() on the server's stderr, holding the BufferedReader's internal lock. process.stderr.close() waits for that lock forever (the reported hang).
• nothing drains the server's stderr: server messages get lost (no Remote: ... debug lines, no stderr tail in error messages), and once the pipe buffer (64 KiB on Linux) is full, the server blocks writing stderr and stops answering requests.

Fix

The session records the pid that started the drain thread. When request() or close() runs in a different process (a forked child), _check_fork():

1. wraps an os.dup() of the stderr fd in a new reader (fresh, unheld lock) and starts a new drain thread on it,
2. closes the old reader's unbuffered raw file (no lock involved), which also makes the old reader count as closed, so neither we nor its finalizer close it via the stuck lock.

close() then uses the normal path. Same-process behaviour is unchanged (one os.getpid() per request). This is done lazily rather than via os.register_at_fork(), because at-fork hooks also run in Popen(..., preexec_fn=...) children before exec, where starting a thread is unsafe.

Remaining limitation (as before): the server is a child of the original process, not of the daemon, so wait() in the daemon cannot see its exit code.

Tests

• New test_stdio_session_works_in_forked_child mimics daemonizing: a child opens the session and forks again, the intermediate parent exits, and the grandchild does a request where the server writes ~200 KB to stderr, then closes. It passes with this fix and hangs (killed after 10 s, test fails) both without any fix and with only a non-blocking close.
• Linux (Debian 13), borg master + borgstore 0.7.0, daemonized borg mount over ssh://: without the fix the daemon stays stuck in stderr.close() after borg umount; with it, it exits ~2 s after umount, no leftover processes, repo lock released. With borg serve --debug, the daemon's server stderr pipe holds 0 unread bytes with the fix vs. stale unread output without it.

🤖 Generated with Claude Code