OSDN Git Service

xen: fix race in xen_qlock_wait()
authorJuergen Gross <jgross@suse.com>
Mon, 1 Oct 2018 05:57:42 +0000 (07:57 +0200)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Wed, 21 Nov 2018 08:27:33 +0000 (09:27 +0100)
commit 2ac2a7d4d9ff4e01e36f9c3d116582f6f655ab47 upstream.

In the following situation a vcpu waiting for a lock might not be
woken up from xen_poll_irq():

CPU 1:                CPU 2:                      CPU 3:
takes a spinlock
                      tries to get lock
                      -> xen_qlock_wait()
frees the lock
-> xen_qlock_kick(cpu2)
                        -> xen_clear_irq_pending()

takes lock again
                                                  tries to get lock
                                                  -> *lock = _Q_SLOW_VAL
                        -> *lock == _Q_SLOW_VAL ?
                        -> xen_poll_irq()
frees the lock
-> xen_qlock_kick(cpu3)

And cpu 2 will sleep forever.

This can be avoided easily by modifying xen_qlock_wait() to call
xen_poll_irq() only if the related irq was not pending and to call
xen_clear_irq_pending() only if it was pending.

Cc: stable@vger.kernel.org
Cc: Waiman.Long@hp.com
Cc: peterz@infradead.org
Signed-off-by: Juergen Gross <jgross@suse.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Signed-off-by: Juergen Gross <jgross@suse.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
arch/x86/xen/spinlock.c

index f42e78d..55ba251 100644 (file)
@@ -47,17 +47,12 @@ static void xen_qlock_wait(u8 *byte, u8 val)
        if (irq == -1)
                return;
 
-       /* clear pending */
-       xen_clear_irq_pending(irq);
-       barrier();
+       /* If irq pending already clear it and return. */
+       if (xen_test_irq_pending(irq)) {
+               xen_clear_irq_pending(irq);
+               return;
+       }
 
-       /*
-        * We check the byte value after clearing pending IRQ to make sure
-        * that we won't miss a wakeup event because of the clearing.
-        *
-        * The sync_clear_bit() call in xen_clear_irq_pending() is atomic.
-        * So it is effectively a memory barrier for x86.
-        */
        if (READ_ONCE(*byte) != val)
                return;