char: remove watch callback on chardev detach from frontend

If a frontend device releases the chardev (via unplug), the chr handlers
are set to NULL via qdev's exit callbacks invoking
qemu_chr_add_handlers().  If the chardev had a pending operation, a
callback will be invoked, which will try to access data in the
just-released frontend, causing a segfault.

Ensure the callbacks are disabled when frontends release chardevs.

This was seen when a virtio-serial port was unplugged when heavy
guest->host IO was in progress (causing a callback to be registered).
In the window in which the throttling was active, unplugging ports
caused a qemu segfault.

https://bugzilla.redhat.com/show_bug.cgi?id=985205

CC: <qemu-stable@nongnu.org>
Reported-by: Sibiao Luo <sluo@redhat.com>
Reviewed-by: Gerd Hoffmann <kraxel@redhat.com>
Signed-off-by: Amit Shah <amit.shah@redhat.com>
(cherry picked from commit 386a5a1e0057e220f79c48fe3689e3dfb17f1b09)

Signed-off-by: Michael Roth <mdroth@linux.vnet.ibm.com>

diff --git a/qemu-char.c b/qemu-char.c
index fa005174ddf0805720bdb46e943fe64e61135a01..fc1c23d2c3c3691b5cdfb1fe27602b122aee0c1b 100644 (file)
--- a/qemu-char.c
+++ b/qemu-char.c
@@ -193,6 +193,8 @@ void qemu_chr_fe_printf(CharDriverState *s, const char *fmt, ...)
     va_end(ap);
 }
 
+static void remove_fd_in_watch(CharDriverState *chr);
+
 void qemu_chr_add_handlers(CharDriverState *s,
                            IOCanReadHandler *fd_can_read,
                            IOReadHandler *fd_read,
@@ -203,6 +205,7 @@ void qemu_chr_add_handlers(CharDriverState *s,
 
     if (!opaque && !fd_can_read && !fd_read && !fd_event) {
         fe_open = 0;
+        remove_fd_in_watch(s);
     } else {
         fe_open = 1;
     }