Re: Strange issues with epoll since 5.0

[Eric Wong]

Eric Wong <e@xxxxxxxxx> wrote:
> Omar Kilani <omar.kilani@xxxxxxxxx> wrote:
> > Hi there,
> > 
> > Iâm still trying to piece together a reproducible test that triggers
> > this, but I wanted to post in case someone goes âhmmm... change X
> > might have done thisâ.
> 
> Maybe Davidlohr knows, since he's responsible for most of the
> epoll changes in 5.0.

Well, I am not sure if I am hitting the same problem Omar is
hitting.  But I did find an epoll_pwait regression in 5.0:

epoll_pwait seems unresponsive to SIGURG in my
heavily-parallelized use case[1] on 5.0.9.  I bisected it to
commit 854a6ed56839a40f6b5d02a2962f48841482eec4
("signal: Add restore_user_sigmask()")

Just reverting the fs/eventpoll.c change in 854a6ed56 seems
enough to fix the non-responsive epoll_pwait for me.  I have not
looked deeply into this, but perhaps the signal_pending check in
restore_user_sigmask is racy w.r.t. epoll.  It is been a while
since I have looked at kernel stuff, myself.

Anyways, this revert works; but I'm not 100% sure why...

diff --git a/fs/eventpoll.c b/fs/eventpoll.c
index a5d219d920e7..151739d76801 100644
--- a/fs/eventpoll.c
+++ b/fs/eventpoll.c
@@ -2247,7 +2247,20 @@ SYSCALL_DEFINE6(epoll_pwait, int, epfd, struct epoll_event __user *, events,
 
 	error = do_epoll_wait(epfd, events, maxevents, timeout);
 
-	restore_user_sigmask(sigmask, &sigsaved);
+	/*
+	 * If we changed the signal mask, we need to restore the original one.
+	 * In case we've got a signal while waiting, we do not restore the
+	 * signal mask yet, and we allow do_signal() to deliver the signal on
+	 * the way back to userspace, before the signal mask is restored.
+	 */
+	if (sigmask) {
+		if (error == -EINTR) {
+			memcpy(&current->saved_sigmask, &sigsaved,
+			       sizeof(sigsaved));
+			set_restore_sigmask();
+		} else
+			set_current_blocked(&sigsaved);
+	}
 
 	return error;
 }
@@ -2272,7 +2285,20 @@ COMPAT_SYSCALL_DEFINE6(epoll_pwait, int, epfd,
 
 	err = do_epoll_wait(epfd, events, maxevents, timeout);
 
-	restore_user_sigmask(sigmask, &sigsaved);
+	/*
+	 * If we changed the signal mask, we need to restore the original one.
+	 * In case we've got a signal while waiting, we do not restore the
+	 * signal mask yet, and we allow do_signal() to deliver the signal on
+	 * the way back to userspace, before the signal mask is restored.
+	 */
+	if (sigmask) {
+		if (err == -EINTR) {
+			memcpy(&current->saved_sigmask, &sigsaved,
+			       sizeof(sigsaved));
+			set_restore_sigmask();
+		} else
+			set_current_blocked(&sigsaved);
+	}
 
 	return err;
 }

Comments and/or a proper fix would be greatly appreciated.

[1] my test case is running the cmogstored 1.7.0 test suite
    in amd64 Debian stable environment.
    test/mgmt_auto_adjust would get stuck and time-out after 60s
    on vanilla v5.0.9

    tgz: https://bogomips.org/cmogstored/files/cmogstored-1.7.0.tar.gz
    # Standard autotools install,  N=32 or some high-ish number
    ./configure
    make -j$N
    make check -j$N

    # OR git clone https://bogomips.org/cmogstored.git

So, requoting the rest of Omar's original report, here; since
I am not sure if his use case involves epoll_pwait like mine does:

> Omar Kilani <omar.kilani@xxxxxxxxx> wrote:
> > Basically, somethingâs broken (or at least, has changed enough to
> > cause problems in user space) in epoll since 5.0. Itâs still broken in
> > 5.1-rc5.
> > 
> > It doesnât happen 100% of the time. Itâs sort of hard to pin down but
> > Iâve observed the following:
> > 
> > * nginx not accepting connections under load
> > * A java app which uses netty / NIO having strange writability
> > semantics on channels, which confuses netty / java enough to not
> > properly flush written data on the socket.
> > 
> > I went and tested these Linux kernels:
> > 
> > 4.20.17
> > 4.19.32
> > 4.14.111
> > 
> > And the issue(s) do not show up there.
> > 
> > Iâm still actively chasing this up, and will report back â I havenât
> > touched kernel code in 15 years so Iâm a little rusty. :)
> > 
> > Regards,
> > Omar