MIME-Version: 1.0
From:   Eric Dumazet <edumazet@google.com>
Date:   Mon, 20 Mar 2023 14:17:50 -0700
Message-ID: <CANn89iLy+-rQDSKCg6g=xkNLL5SmgkB7gyMSgxCEPp13-+eLbw@mail.gmail.com>
Subject: syzbot + epoll
To:     Paolo Abeni <pabeni@redhat.com>,
        LKML <linux-kernel@vger.kernel.org>, Xiumei Mu <xmu@redhat.com>,
        Jacob Keller <jacob.e.keller@intel.com>,
        Soheil Hassas Yeganeh <soheil@google.com>,
        Andrew Morton <akpm@linux-foundation.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

This is about this recent syzbot report (with a C repro)

https://lore.kernel.org/lkml/000000000000c6dc0305f75b4d74@google.com/T/#u

I think this is caused by:

commit fc02a95bb6d8bf58c6efd7e362814558eea2ef28
Author: Paolo Abeni <pabeni@redhat.com>
Date:   Tue Mar 7 19:46:37 2023 +0100

    epoll: use refcount to reduce ep_mutex contention

Problem is that __ep_remove() might return early, without removing epi
from the rbtree (ep->rbr)

This happens when epi->dying has been set to true here :

https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/tree/fs/eventpoll.c?id=6f72958a49f68553f2b6ff713e8c8e51a34c1e1e#n954

So we loop, while holding the ep->mtx held, meaning that the other
thread is blocked here

https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/tree/fs/eventpoll.c?id=6f72958a49f68553f2b6ff713e8c8e51a34c1e1e#n962

So this dead locks.

Maybe fix this with:

diff --git a/fs/eventpoll.c b/fs/eventpoll.c
index 25a59640748a0fd22a84a5aecb90815fbbca9cef..1db56c6175aab5af7bc637a452b68ed8bc11fd7f
100644
--- a/fs/eventpoll.c
+++ b/fs/eventpoll.c
@@ -784,7 +784,7 @@ static void ep_remove_safe(struct eventpoll *ep,
struct epitem *epi)

 static void ep_clear_and_put(struct eventpoll *ep)
 {
-       struct rb_node *rbp;
+       struct rb_node *rbp, *next;
        struct epitem *epi;
        bool dispose;

@@ -810,7 +810,8 @@ static void ep_clear_and_put(struct eventpoll *ep)
         * Since we still own a reference to the eventpoll struct, the
loop can't
         * dispose it.
         */
-       while ((rbp = rb_first_cached(&ep->rbr)) != NULL) {
+       for (rbp = rb_first_cached(&ep->rbr); rbp; rbp = next) {
+               next = rb_next(rbp);
                epi = rb_entry(rbp, struct epitem, rbn);
                ep_remove_safe(ep, epi);
                cond_resched();