t5400: avoid concurrent writes into a trace file
authorJeff King <peff@peff.net>
Thu, 18 May 2017 05:02:09 +0000 (01:02 -0400)
committerJunio C Hamano <gitster@pobox.com>
Thu, 18 May 2017 05:09:23 +0000 (14:09 +0900)
One test in t5400 examines the packet exchange between git-push and
git-receive-pack. The latter inherits the GIT_TRACE_PACKET environment
variable, so that both processes dump trace data into the same file
concurrently. This should not be a problem because the trace file is
opened with O_APPEND.

On Windows, however, O_APPEND is not atomic as it should be: it is
emulated as lseek(SEEK_END) followed by write(). For this reason, the
test is unreliable: it can happen that one process overwrites a line
that was just written by the other process. As a consequence, the test
sometimes does not find one or another line that is expected (and it is
also successful occasionally).

The test case is actually only interested in the output of git-push.
To ensure that only git-push writes to the trace file, override the
receive-pack command such that it does not even open the trace file.

Reported-by: Johannes Sixt <j6t@kdbg.org>
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>

index 3331e0f..d375d71 100755 (executable)
@@ -288,7 +288,10 @@ test_expect_success 'receive-pack de-dupes .have lines' '
        $shared .have
-       GIT_TRACE_PACKET=$(pwd)/trace git push fork HEAD:foo &&
+       GIT_TRACE_PACKET=$(pwd)/trace \
+           git push \
+               --receive-pack="unset GIT_TRACE_PACKET; git-receive-pack" \
+               fork HEAD:foo &&
        extract_ref_advertisement <trace >refs &&
        test_cmp expect refs