[PATCH] epoll_wait: fix EINTR leak

[PATCH] epoll_wait: fix EINTR leak

[Denys Vlasenko]

Usage of EINTR is wrong. For one, it causes epoll_wait
to spuriously return with EINTR on ptrace attach.
One of ERESTARTfoo's should be used.

By analogy with poll syscall family, epoll_wait should be interrupted
by signals regardless of SA_RESTART, thus ERESTARTNOHAND is used,
not ERESTARTSYS.

Signed-off-by: Denys Vlasenko <dvlasenk@redhat.com>
CC: Oleg Nesterov <oleg@redhat.com>
---
 fs/eventpoll.c | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/fs/eventpoll.c b/fs/eventpoll.c
index 0cff4434..08f8076 100644
--- a/fs/eventpoll.c
+++ b/fs/eventpoll.c
@@ -1598,7 +1598,7 @@ fetch_events:
 			if (ep_events_available(ep) || timed_out)
 				break;
 			if (signal_pending(current)) {
-				res = -EINTR;
+				res = -ERESTARTNOHAND;
 				break;
 			}
 
@@ -1990,7 +1990,7 @@ SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
 	 * the way back to userspace, before the signal mask is restored.
 	 */
 	if (sigmask) {
-		if (error == -EINTR) {
+		if (error == -ERESTARTNOHAND) {
 			memcpy(&current->saved_sigmask, &sigsaved,
 			       sizeof(sigsaved));
 			set_restore_sigmask();
@@ -2035,7 +2035,7 @@ COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
 	 * the way back to userspace, before the signal mask is restored.
 	 */
 	if (sigmask) {
-		if (err == -EINTR) {
+		if (err == -ERESTARTNOHAND) {
 			memcpy(&current->saved_sigmask, &sigsaved,
 			       sizeof(sigsaved));
 			set_restore_sigmask();
-- 
1.8.1.4

Re: [PATCH] epoll_wait: fix EINTR leak

[Oleg Nesterov]

On 06/24, Denys Vlasenko wrote:
>
> Usage of EINTR is wrong.

I agree, this is not nice. However "fix EINTR leak" doesn't look
accurate, -EINTR is fine as an error code. Just the syscall should
restart if possible.

> --- a/fs/eventpoll.c
> +++ b/fs/eventpoll.c
> @@ -1598,7 +1598,7 @@ fetch_events:
>  			if (ep_events_available(ep) || timed_out)
>  				break;
>  			if (signal_pending(current)) {
> -				res = -EINTR;
> +				res = -ERESTARTNOHAND;

This and other similar changes do look right.

Say, sys_epoll_wait(). With this patch it can sleep, then return
ERESTARTNOHAND.

And we restart it with the same timeout again. If you want to
make it restartable, you need ERESTART_RESTARTBLOCK and
do_restart_epoll_wait() which we do not have.

See for example sys_poll() which implements this logic.

Oleg.

Re: [PATCH] epoll_wait: fix EINTR leak

[Oleg Nesterov]

On 06/24, Oleg Nesterov wrote:
>
> On 06/24, Denys Vlasenko wrote:
> >
> > Usage of EINTR is wrong.
>
> I agree, this is not nice. However "fix EINTR leak" doesn't look
> accurate, -EINTR is fine as an error code. Just the syscall should
> restart if possible.
>
> > --- a/fs/eventpoll.c
> > +++ b/fs/eventpoll.c
> > @@ -1598,7 +1598,7 @@ fetch_events:
> >  			if (ep_events_available(ep) || timed_out)
> >  				break;
> >  			if (signal_pending(current)) {
> > -				res = -EINTR;
> > +				res = -ERESTARTNOHAND;
>
> This and other similar changes do look right.
>
> Say, sys_epoll_wait(). With this patch it can sleep, then return
> ERESTARTNOHAND.
>
> And we restart it with the same timeout again. If you want to
> make it restartable, you need ERESTART_RESTARTBLOCK and
> do_restart_epoll_wait() which we do not have.
>
> See for example sys_poll() which implements this logic.

But, to avoid the confusion, please note that this change won't
make it restartable wrt SA_RESTART. But it will help PTRACE_ATTACH
or PTRACE_INTERRUPT or other "spurious" signal.

We simply can't do this because there is no way to update
"timeout" later if the task actually returns to user-mode.

Oleg.