The t5562 script occasionally takes 60 extra seconds to complete due to
a race condition in the invoke-with-content-length.pl helper.
The way it's supposed to work is this:
- we set up a SIGCLD handler
- we kick off http-backend and write to it with a set content-length,
but _don't_ close the pipe
- we sleep for 60 seconds, assuming that SIGCLD from http-backend
finishing will interrupt us
- after the sleep finishes (whetherby 60 seconds or because it was
interrupted by the signal), we check a flag to see if our SIGCLD
handler was called. If not, then we complain.
This usually completes immediately, because the signal interrupts our
sleep. But very occasionally the child process dies _before_ we hit the
sleep, so we don't realize it. The test still completes successfully
(because our $exited flag is set), but it takes an extra 60 seconds.
There's no way to check the flag and sleep atomically. So the best we
can do with this approach is to sleep in smaller chunks (say, 1 second)
and check the flag incrementally. Then we waste a maximum of 1 second if
we lose the race. This was proposed in:
https://lore.kernel.org/git/20190218205028.32486-1-max@max630.net/
and it does work. But we can do better.
Instead of blocking on sleep and waiting for the child signal to
interrupt us, we can block on the child exiting and set an alarm signal
to trigger the timeout.
This lets us exit the script immediately when the child behaves (with no
race possible), and wait a maximum of 60 seconds when it doesn't.
Note one small subtlety: perl is very willing to restart the waitpid()
call after the alarm is delivered, even if we've thrown an exception via
die. "perldoc -f alarm" claims you can get around this with an eval/die
combo (and even has some example code), but it doesn't seem to work for
me with waitpid(); instead, we continue waiting until the child exits.
So instead, we'll instruct the child process to exit in the alarm
handler itself. In the original code this was done by calling
close($out). That would continue to work, since our child is always
http-backend, which should exit when its stdin closes. But we can be
even more robust against a hung or confused child by sending a KILL
signal, which should terminate it immediately.
Reported-by: SZEDER Gábor <szeder.dev@gmail.com>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Some systems do not have perl installed to /usr/bin. Use the variable
from the build settiings, and call perl directly than via shebang.
Signed-off-by: Max Kirillov <max@max630.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Push passes to another commands, as described in
https://public-inbox.org/git/20171129032214.GB32345@sigill.intra.peff.net/
As it gets complicated to correctly track the data length, instead transfer
the data through parent process and cut the pipe as the specified length is
reached. Do it only when CONTENT_LENGTH is set, otherwise pass the input
directly to the forked commands.
Add tests for cases:
* CONTENT_LENGTH is set, script's stdin has more data, with all combinations
of variations: fetch or push, plain or compressed body, correct or truncated
input.
* CONTENT_LENGTH is specified to a value which does not fit into ssize_t.
Helped-by: Junio C Hamano <gitster@pobox.com>
Signed-off-by: Max Kirillov <max@max630.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Jeff King
Max Kirillov