| Message ID | 20210112132911.4031-1-chrubis@suse.cz |
|---|---|
| State | Accepted |
| Headers | show |
| Series | syscalls: epoll_pwait01: Work around a race | expand |
Hi Cyril, > There was a race that would manifest on slower machines. > The call to epoll_pwait() could time out before the child has chance to > run, and that would cause the signal to be sent to the parent when it > was already sleeping in wait(). > Ideally the whole test should be rewritten into new library and fixed > properly, however as we are just before a release this is an attempt for > a minimal fix. > The logic in the test is changed so that: > - epoll_wait() sleeps indefinitely > - the child: > - waits for the parent to get asleep > - sends the signal > - sleeps > - writes to the pipe > This causes the child to actually run, while the parent is blocked in > the epoll_wait(), which greatly increases the changes of the signal > arriving at the right time. Reviewed-by: Petr Vorel <pvorel@suse.cz> LGTM. > Fixes: #765 Kind regards, Petr
On Tue, Jan 12, 2021 at 9:28 PM Cyril Hrubis <chrubis@suse.cz> wrote: > There was a race that would manifest on slower machines. > > The call to epoll_pwait() could time out before the child has chance to > run, and that would cause the signal to be sent to the parent when it > was already sleeping in wait(). > > Ideally the whole test should be rewritten into new library and fixed > properly, however as we are just before a release this is an attempt for > a minimal fix. > > The logic in the test is changed so that: > > - epoll_wait() sleeps indefinitely > - the child: > - waits for the parent to get asleep > - sends the signal > - sleeps > - writes to the pipe > > This causes the child to actually run, while the parent is blocked in > the epoll_wait(), which greatly increases the changes of the signal > arriving at the right time. > > Fixes: #765 > > Signed-off-by: Cyril Hrubis <chrubis@suse.cz> > Reviewed-by: Li Wang <liwang@redhat.com>
diff --git a/testcases/kernel/syscalls/epoll_pwait/epoll_pwait01.c b/testcases/kernel/syscalls/epoll_pwait/epoll_pwait01.c index ed05da8fe..1f08112ad 100644 --- a/testcases/kernel/syscalls/epoll_pwait/epoll_pwait01.c +++ b/testcases/kernel/syscalls/epoll_pwait/epoll_pwait01.c @@ -121,8 +121,8 @@ static void verify_sigmask(void) return; } - if (TEST_RETURN != 0) { - tst_resm(TFAIL, "epoll_pwait() returned %li, expected 0", + if (TEST_RETURN != 1) { + tst_resm(TFAIL, "epoll_pwait() returned %li, expected 1", TEST_RETURN); return; } @@ -153,6 +153,7 @@ static void sighandler(int sig LTP_ATTRIBUTE_UNUSED) static void do_test(sigset_t *sigmask) { pid_t cpid; + char b; cpid = tst_fork(); if (cpid < 0) @@ -161,13 +162,15 @@ static void do_test(sigset_t *sigmask) if (cpid == 0) do_child(); - TEST(epoll_pwait(epfd, &epevs, 1, 100, sigmask)); + TEST(epoll_pwait(epfd, &epevs, 1, -1, sigmask)); if (sigmask != NULL) verify_sigmask(); else verify_nonsigmask(); + SAFE_READ(cleanup, 1, fds[0], &b, 1); + tst_record_childstatus(cleanup, cpid); } @@ -179,6 +182,8 @@ static void do_child(void) } SAFE_KILL(cleanup, getppid(), SIGUSR1); + usleep(10000); + SAFE_WRITE(cleanup, 1, fds[1], "w", 1); cleanup(); tst_exit();
There was a race that would manifest on slower machines. The call to epoll_pwait() could time out before the child has chance to run, and that would cause the signal to be sent to the parent when it was already sleeping in wait(). Ideally the whole test should be rewritten into new library and fixed properly, however as we are just before a release this is an attempt for a minimal fix. The logic in the test is changed so that: - epoll_wait() sleeps indefinitely - the child: - waits for the parent to get asleep - sends the signal - sleeps - writes to the pipe This causes the child to actually run, while the parent is blocked in the epoll_wait(), which greatly increases the changes of the signal arriving at the right time. Fixes: #765 Signed-off-by: Cyril Hrubis <chrubis@suse.cz> --- testcases/kernel/syscalls/epoll_pwait/epoll_pwait01.c | 11 ++++++++--- 1 file changed, 8 insertions(+), 3 deletions(-)