locking/static_key: Fix concurrent static_key_slow_inc()
authorPaolo Bonzini <pbonzini@redhat.com>
Tue, 21 Jun 2016 16:52:17 +0000 (18:52 +0200)
committerIngo Molnar <mingo@kernel.org>
Fri, 24 Jun 2016 06:23:16 +0000 (08:23 +0200)
The following scenario is possible:

    CPU 1                                   CPU 2
    static_key_slow_inc()
     atomic_inc_not_zero()
      -> key.enabled == 0, no increment
     jump_label_lock()
     atomic_inc_return()
      -> key.enabled == 1 now
                                            static_key_slow_inc()
                                             atomic_inc_not_zero()
                                              -> key.enabled == 1, inc to 2
                                             return
                                            ** static key is wrong!
     jump_label_update()
     jump_label_unlock()

Testing the static key at the point marked by (**) will follow the
wrong path for jumps that have not been patched yet.  This can
actually happen when creating many KVM virtual machines with userspace
LAPIC emulation; just run several copies of the following program:

    #include <fcntl.h>
    #include <unistd.h>
    #include <sys/ioctl.h>
    #include <linux/kvm.h>

    int main(void)
    {
        for (;;) {
            int kvmfd = open("/dev/kvm", O_RDONLY);
            int vmfd = ioctl(kvmfd, KVM_CREATE_VM, 0);
            close(ioctl(vmfd, KVM_CREATE_VCPU, 1));
            close(vmfd);
            close(kvmfd);
        }
        return 0;
    }

Every KVM_CREATE_VCPU ioctl will attempt a static_key_slow_inc() call.
The static key's purpose is to skip NULL pointer checks and indeed one
of the processes eventually dereferences NULL.

As explained in the commit that introduced the bug:

  706249c222f6 ("locking/static_keys: Rework update logic")

jump_label_update() needs key.enabled to be true.  The solution adopted
here is to temporarily make key.enabled == -1, and use go down the
slow path when key.enabled <= 0.

Reported-by: Dmitry Vyukov <dvyukov@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: <stable@vger.kernel.org> # v4.3+
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Fixes: 706249c222f6 ("locking/static_keys: Rework update logic")
Link: http://lkml.kernel.org/r/1466527937-69798-1-git-send-email-pbonzini@redhat.com
[ Small stylistic edits to the changelog and the code. ]
Signed-off-by: Ingo Molnar <mingo@kernel.org>
include/linux/jump_label.h
kernel/jump_label.c

index 0536524..6890446 100644 (file)
@@ -117,13 +117,18 @@ struct module;
 
 #include <linux/atomic.h>
 
+#ifdef HAVE_JUMP_LABEL
+
 static inline int static_key_count(struct static_key *key)
 {
-       return atomic_read(&key->enabled);
+       /*
+        * -1 means the first static_key_slow_inc() is in progress.
+        *  static_key_enabled() must return true, so return 1 here.
+        */
+       int n = atomic_read(&key->enabled);
+       return n >= 0 ? n : 1;
 }
 
-#ifdef HAVE_JUMP_LABEL
-
 #define JUMP_TYPE_FALSE        0UL
 #define JUMP_TYPE_TRUE 1UL
 #define JUMP_TYPE_MASK 1UL
@@ -162,6 +167,11 @@ extern void jump_label_apply_nops(struct module *mod);
 
 #else  /* !HAVE_JUMP_LABEL */
 
+static inline int static_key_count(struct static_key *key)
+{
+       return atomic_read(&key->enabled);
+}
+
 static __always_inline void jump_label_init(void)
 {
        static_key_initialized = true;
index 05254ee..4b353e0 100644 (file)
@@ -58,13 +58,36 @@ static void jump_label_update(struct static_key *key);
 
 void static_key_slow_inc(struct static_key *key)
 {
+       int v, v1;
+
        STATIC_KEY_CHECK_USE();
-       if (atomic_inc_not_zero(&key->enabled))
-               return;
+
+       /*
+        * Careful if we get concurrent static_key_slow_inc() calls;
+        * later calls must wait for the first one to _finish_ the
+        * jump_label_update() process.  At the same time, however,
+        * the jump_label_update() call below wants to see
+        * static_key_enabled(&key) for jumps to be updated properly.
+        *
+        * So give a special meaning to negative key->enabled: it sends
+        * static_key_slow_inc() down the slow path, and it is non-zero
+        * so it counts as "enabled" in jump_label_update().  Note that
+        * atomic_inc_unless_negative() checks >= 0, so roll our own.
+        */
+       for (v = atomic_read(&key->enabled); v > 0; v = v1) {
+               v1 = atomic_cmpxchg(&key->enabled, v, v + 1);
+               if (likely(v1 == v))
+                       return;
+       }
 
        jump_label_lock();
-       if (atomic_inc_return(&key->enabled) == 1)
+       if (atomic_read(&key->enabled) == 0) {
+               atomic_set(&key->enabled, -1);
                jump_label_update(key);
+               atomic_set(&key->enabled, 1);
+       } else {
+               atomic_inc(&key->enabled);
+       }
        jump_label_unlock();
 }
 EXPORT_SYMBOL_GPL(static_key_slow_inc);
@@ -72,6 +95,13 @@ EXPORT_SYMBOL_GPL(static_key_slow_inc);
 static void __static_key_slow_dec(struct static_key *key,
                unsigned long rate_limit, struct delayed_work *work)
 {
+       /*
+        * The negative count check is valid even when a negative
+        * key->enabled is in use by static_key_slow_inc(); a
+        * __static_key_slow_dec() before the first static_key_slow_inc()
+        * returns is unbalanced, because all other static_key_slow_inc()
+        * instances block while the update is in progress.
+        */
        if (!atomic_dec_and_mutex_lock(&key->enabled, &jump_label_mutex)) {
                WARN(atomic_read(&key->enabled) < 0,
                     "jump label: negative count!\n");