closures: Fix race in closure_sync()
authorKent Overstreet <kent.overstreet@linux.dev>
Tue, 24 Oct 2023 18:46:58 +0000 (14:46 -0400)
committerKent Overstreet <kent.overstreet@linux.dev>
Tue, 31 Oct 2023 01:48:22 +0000 (21:48 -0400)
As pointed out by Linus, closure_sync() was racy; we could skip blocking
immediately after a get() and a put(), but then that would skip any
barrier corresponding to the other thread's put() barrier.

To fix this, always do the full __closure_sync() sequence whenever any
get() has happened and the closure might have been used by other
threads.

Signed-off-by: Kent Overstreet <kent.overstreet@linux.dev>
fs/bcachefs/fs-io-direct.c
include/linux/closure.h
lib/closure.c

index 6a9557e7ecabb47d1a30b7c665367decb3fecbe4..5b42a76c4796f90062bb86e2914d0301e52cf7d0 100644 (file)
@@ -113,6 +113,7 @@ static int bch2_direct_IO_read(struct kiocb *req, struct iov_iter *iter)
        } else {
                atomic_set(&dio->cl.remaining,
                           CLOSURE_REMAINING_INITIALIZER + 1);
+               dio->cl.closure_get_happened = true;
        }
 
        dio->req        = req;
index bdab17050bc889bbbe3431cbea607112deae7f50..de7bb47d8a46ace38d95a81ed6df231d91ac725b 100644 (file)
@@ -154,6 +154,7 @@ struct closure {
        struct closure          *parent;
 
        atomic_t                remaining;
+       bool                    closure_get_happened;
 
 #ifdef CONFIG_DEBUG_CLOSURES
 #define CLOSURE_MAGIC_DEAD     0xc054dead
@@ -185,7 +186,11 @@ static inline unsigned closure_nr_remaining(struct closure *cl)
  */
 static inline void closure_sync(struct closure *cl)
 {
-       if (closure_nr_remaining(cl) != 1)
+#ifdef CONFIG_DEBUG_CLOSURES
+       BUG_ON(closure_nr_remaining(cl) != 1 && !cl->closure_get_happened);
+#endif
+
+       if (cl->closure_get_happened)
                __closure_sync(cl);
 }
 
@@ -257,6 +262,8 @@ static inline void closure_queue(struct closure *cl)
  */
 static inline void closure_get(struct closure *cl)
 {
+       cl->closure_get_happened = true;
+
 #ifdef CONFIG_DEBUG_CLOSURES
        BUG_ON((atomic_inc_return(&cl->remaining) &
                CLOSURE_REMAINING_MASK) <= 1);
@@ -279,6 +286,7 @@ static inline void closure_init(struct closure *cl, struct closure *parent)
                closure_get(parent);
 
        atomic_set(&cl->remaining, CLOSURE_REMAINING_INITIALIZER);
+       cl->closure_get_happened = false;
 
        closure_debug_create(cl);
        closure_set_ip(cl);
index 501dfa277b59e94b4eb6108bcc72b72d9cdcce20..f86c9eeafb35ad9da21ebddda8a182ea27970ff8 100644 (file)
@@ -23,6 +23,8 @@ static inline void closure_put_after_sub(struct closure *cl, int flags)
        if (!r) {
                smp_acquire__after_ctrl_dep();
 
+               cl->closure_get_happened = false;
+
                if (cl->fn && !(flags & CLOSURE_DESTRUCTOR)) {
                        atomic_set(&cl->remaining,
                                   CLOSURE_REMAINING_INITIALIZER);
@@ -92,6 +94,7 @@ bool closure_wait(struct closure_waitlist *waitlist, struct closure *cl)
        if (atomic_read(&cl->remaining) & CLOSURE_WAITING)
                return false;
 
+       cl->closure_get_happened = true;
        closure_set_waiting(cl, _RET_IP_);
        atomic_add(CLOSURE_WAITING + 1, &cl->remaining);
        llist_add(&cl->list, &waitlist->list);