[PATCH 0/2] Fix bugs in tcp_splice_sock_handler()
Patch 1/2 here fixes what I think is the bug causing the latest issue in: https://github.com/containers/podman/issues/23686 While fixing that I spotted the bug fixed in 2/2. Definitely a bug, although I suspect we'll mostly get away with it if the --trace option is not enabled. David Gibson (2): tcp_splice: Don't double could bytes read on EINTR tcp_splice: Don't clobber errno before checking for EAGAIN tcp_splice.c | 43 ++++++++++++++++++++++--------------------- 1 file changed, 22 insertions(+), 21 deletions(-) -- 2.49.0
In tcp_splice_sock_handler(), if we get an EINTR on our second splice()
(pipe to output socket) we - as we should - go back and retry it. However,
we do so *after* we've already updated our byte counters. That does no
harm for the conn->written[] counter - since the second splice() returned
an error it will be advanced by 0. However we also advance the
conn->read[] counter, and then do so again when the splice() succeeds.
This results in the counters being out of sync, and us thinking we have
remaining data in the pipe when we don't, which can leave us in an
infinite loop once the stream finishes.
Fix this by moving the EINTR handling to directly next to the splice()
call (which is what we usually do for EINTR). As a bonus this removes one
mildly confusing goto.
For symmetry, also rework the EINTR handling on the first splice() the same
way, although that doesn't (as far as I can tell) have buggy side effects.
Link: https://github.com/containers/podman/issues/23686
Signed-off-by: David Gibson
Like many places, tcp_splice_sock_handler() needs to handle EAGAIN
specially, in this case for both of its splice() calls. Unfortunately it
tests for EAGAIN some time after those calls. In between there has been
at least a flow_trace() which could have clobbered errno. Move the test on
errno closer to the relevant system calls to avoid this problem.
Signed-off-by: David Gibson
On Wed, 9 Apr 2025 16:35:39 +1000
David Gibson
Patch 1/2 here fixes what I think is the bug causing the latest issue in: https://github.com/containers/podman/issues/23686
While fixing that I spotted the bug fixed in 2/2. Definitely a bug, although I suspect we'll mostly get away with it if the --trace option is not enabled.
David Gibson (2): tcp_splice: Don't double could bytes read on EINTR tcp_splice: Don't clobber errno before checking for EAGAIN
Applied. -- Stefano
participants (2)
-
David Gibson
-
Stefano Brivio