On Tue, Nov 20, 2018 at 04:11:08AM -0500, Jeff King wrote:
> On Mon, Nov 19, 2018 at 11:36:08AM -0800, Carlo Arenas wrote:
>
> > tests 3-8 seem to fail because perl is hardcoded to /urs/bin/perl in
> > t5562/invoke-with-content-length.pl, while I seem to be getting some
> > sporadic errors in 9 with the following output :
> >
> > ++ env CONTENT_TYPE=application/x-git-receive-pack-request
> > QUERY_STRING=/repo.git/git-receive-pack
> > 'PATH_TRANSLATED=/home/carenas/src/git/t/trash
> > directory.t5562-http-backend-content-length/.git/git-receive-pack'
> > GIT_HTTP_EXPORT_ALL=TRUE REQUEST_METHOD=POST
> > /home/carenas/src/git/t/t5562/invoke-with-content-length.pl push_body
> > git http-backend
> > ++ verify_http_result '200 OK'
> > ++ grep fatal: act.err
> > Binary file act.err matches
> > ++ return 1
> > error: last command exited with $?=1
> > not ok 9 - push plain
> >
> > and the following output in act.err (with a 200 in act)
> >
> > fatal: the remote end hung up unexpectedly
>
> This bit me today, too, and I can reproduce it by running under my
> stress-testing script.
I saw this a few times on Travis CI as well.
> Curiously, the act.err file also has 54 NUL bytes before the "fatal:"
> message.
I think those NUL bytes come from the file system.
The contents of 'act.err' from the previous test ('fetch gzipped
empty') is usually:
fatal: request ended in the middle of the gzip stream
fatal: the remote end hung up unexpectedly
Notice that the length of the first line is 54 bytes (including the
trailing newline). So I suspect that the following is happening:
- http-backend in the previous test writes the first line,
- that test finishes and this one starts,
- this test truncates 'act.err',
- and then the still-running http-backend from the previous test
finally writes the second line.
So at this point 'act.err' is empty, but the offset of the fd of the
redirection still open from the previous test is at 54, so the file
system fills those bytes with NULs.
> I tried adding an "strace" to see who was producing that
> output, but I can't seem to get it to fail when running under strace
> (presumably because the timing is quite different, and this likely some
> kind of pipe race).
>
> -Peff
