On Fri, Aug 10, 2018 at 3:28 PM Eric Blake <ebl...@redhat.com> wrote:
>
> In kill_qemu() we have an assert that checks that the QEMU process
> didn't dump core:
> assert(!WCOREDUMP(wstatus));
>
> Unfortunately the WCOREDUMP macro here means the resulting message
> is not very easy to comprehend on at least some systems:
>
> ahci-test: tests/libqtest.c:113: kill_qemu: Assertion `!(((__extension__
> (((union { __typeof(wstatus) __in; int __i; }) { .__in = (wstatus) }).__i)))
> & 0x80)' failed.
>
> and it doesn't identify what signal the process took. What's more,
> WCOREDUMP is not reliable - in some cases, setrlimit() coupled with
> kernel dump settings can result in the flag not being set. It's
> better to log ALL death by signal, instead of caring whether a core
> dump was attempted (although once we know a signal happened, also
> mentioning if a core dump is present can be helpful).
>
> Furthermore, we are NOT detecting EINTR (while EINTR shouldn't be
> happening if we didn't install signal handlers, it's still better
> to always be robust).
A nice thing about this patch is that "make check" will now return an
error when compiled with --enable-sanitizers and there are leaks
(instead of just reporting warnings).
With the few leak fixes pending merge, x86-64 ASAN make check is
clean: I wish we can enable it in our CI, to avoid regressions.
That I know of, there is at least patchew (make docker), travis, and
Peter has his own set of integration tools/build iirc.
Travis looks overloaded to me.
May I suggest to replace docker-test-clang@ubuntu by
docker-test-debug@ubuntu in patchew? (test-debug uses clang, with some
default sanitizers)
The 2 patches missing upstream for a clean x86-64 run are:
- megasas: fix sglist leak
- tests: fix bdrv-drain leak
thanks
>
> Finally, even non-signal death with a non-zero status is suspicious,
> since qemu's SIGINT handler is supposed to result in exit(0).
>
> Instead of using a raw assert, print the information in an
> easier to understand way:
>
> /i386/ahci/sanity: tests/libqtest.c:119: kill_qemu() detected QEMU death from
> signal 11 (Segmentation fault) (core dumped)
> Aborted (core dumped)
>
> (Of course, the really useful information would be why the QEMU
> process dumped core in the first place, but we don't have that
> by the time the test program has picked up the exit status.)
>
> Suggested-by: Peter Maydell <peter.mayd...@linaro.org>
> Signed-off-by: Eric Blake <ebl...@redhat.com>
>
> ---
> v4: don't special-case death without WCOREDUMP [Markus]
> v3: use TFR() instead of open-coding the retry loop [Thomas]
> ---
> tests/libqtest.c | 24 +++++++++++++++++++++---
> 1 file changed, 21 insertions(+), 3 deletions(-)
>
> diff --git a/tests/libqtest.c b/tests/libqtest.c
> index 098af6aec44..57cf89015e0 100644
> --- a/tests/libqtest.c
> +++ b/tests/libqtest.c
> @@ -107,10 +107,28 @@ static void kill_qemu(QTestState *s)
> pid_t pid;
>
> kill(s->qemu_pid, SIGTERM);
> - pid = waitpid(s->qemu_pid, &wstatus, 0);
> + TFR(pid = waitpid(s->qemu_pid, &wstatus, 0));
>
> - if (pid == s->qemu_pid && WIFSIGNALED(wstatus)) {
> - assert(!WCOREDUMP(wstatus));
> + assert(pid == s->qemu_pid);
> + /*
> + * We expect qemu to exit with status 0; anything else is
> + * fishy and should be logged with as much detail as possible.
> + */
> + if (wstatus) {
> + if (WIFEXITED(wstatus)) {
> + fprintf(stderr, "%s:%d: kill_qemu() tried to terminate QEMU "
> + "process but encountered exit status %d\n",
> + __FILE__, __LINE__, WEXITSTATUS(wstatus));
> + } else if (WIFSIGNALED(wstatus)) {
> + int sig = WTERMSIG(wstatus);
> + const char *signame = strsignal(sig) ?: "unknown ???";
> + const char *dump = WCOREDUMP(wstatus) ? " (dumped core)" :
> "";
> +
> + fprintf(stderr, "%s:%d: kill_qemu() detected QEMU death "
> + "from signal %d (%s)%s\n",
> + __FILE__, __LINE__, sig, signame, dump);
> + }
> + abort();
> }
> }
> }
> --
> 2.14.4
>
>
--
Marc-André Lureau
