net: unix: fix inflight counting bug in garbage collector
authorMiklos Szeredi <mszeredi@suse.cz>
Sun, 9 Nov 2008 14:23:57 +0000 (15:23 +0100)
committerLinus Torvalds <torvalds@linux-foundation.org>
Sun, 9 Nov 2008 19:17:33 +0000 (11:17 -0800)
Previously I assumed that the receive queues of candidates don't
change during the GC.  This is only half true, nothing can be received
from the queues (see comment in unix_gc()), but buffers could be added
through the other half of the socket pair, which may still have file
descriptors referring to it.

This can result in inc_inflight_move_tail() erronously increasing the
"inflight" counter for a unix socket for which dec_inflight() wasn't
previously called.  This in turn can trigger the "BUG_ON(total_refs <
inflight_refs)" in a later garbage collection run.

Fix this by only manipulating the "inflight" counter for sockets which
are candidates themselves.  Duplicating the file references in
unix_attach_fds() is also needed to prevent a socket becoming a
candidate for GC while the skb that contains it is not yet queued.

Reported-by: Andrea Bittau <a.bittau@cs.ucl.ac.uk>
Signed-off-by: Miklos Szeredi <mszeredi@suse.cz>
CC: stable@kernel.org
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
include/net/af_unix.h
net/unix/af_unix.c
net/unix/garbage.c

index 7dd29b7e461dead402b25d6b0fb27451d6896441..c29ff1da8a18ac2dbf905f6bd83a27066e2283e4 100644 (file)
@@ -54,6 +54,7 @@ struct unix_sock {
         atomic_long_t           inflight;
         spinlock_t             lock;
        unsigned int            gc_candidate : 1;
+       unsigned int            gc_maybe_cycle : 1;
         wait_queue_head_t       peer_wait;
 };
 #define unix_sk(__sk) ((struct unix_sock *)__sk)
index 4d3c6071b9a47102212bcc65afa55c640e7bb5b8..eb90f77bb0e294b6358db3361740deaa0bec969e 100644 (file)
@@ -1302,14 +1302,23 @@ static void unix_destruct_fds(struct sk_buff *skb)
        sock_wfree(skb);
 }
 
-static void unix_attach_fds(struct scm_cookie *scm, struct sk_buff *skb)
+static int unix_attach_fds(struct scm_cookie *scm, struct sk_buff *skb)
 {
        int i;
+
+       /*
+        * Need to duplicate file references for the sake of garbage
+        * collection.  Otherwise a socket in the fps might become a
+        * candidate for GC while the skb is not yet queued.
+        */
+       UNIXCB(skb).fp = scm_fp_dup(scm->fp);
+       if (!UNIXCB(skb).fp)
+               return -ENOMEM;
+
        for (i=scm->fp->count-1; i>=0; i--)
                unix_inflight(scm->fp->fp[i]);
-       UNIXCB(skb).fp = scm->fp;
        skb->destructor = unix_destruct_fds;
-       scm->fp = NULL;
+       return 0;
 }
 
 /*
@@ -1368,8 +1377,11 @@ static int unix_dgram_sendmsg(struct kiocb *kiocb, struct socket *sock,
                goto out;
 
        memcpy(UNIXCREDS(skb), &siocb->scm->creds, sizeof(struct ucred));
-       if (siocb->scm->fp)
-               unix_attach_fds(siocb->scm, skb);
+       if (siocb->scm->fp) {
+               err = unix_attach_fds(siocb->scm, skb);
+               if (err)
+                       goto out_free;
+       }
        unix_get_secdata(siocb->scm, skb);
 
        skb_reset_transport_header(skb);
@@ -1538,8 +1550,13 @@ static int unix_stream_sendmsg(struct kiocb *kiocb, struct socket *sock,
                size = min_t(int, size, skb_tailroom(skb));
 
                memcpy(UNIXCREDS(skb), &siocb->scm->creds, sizeof(struct ucred));
-               if (siocb->scm->fp)
-                       unix_attach_fds(siocb->scm, skb);
+               if (siocb->scm->fp) {
+                       err = unix_attach_fds(siocb->scm, skb);
+                       if (err) {
+                               kfree_skb(skb);
+                               goto out_err;
+                       }
+               }
 
                if ((err = memcpy_fromiovec(skb_put(skb,size), msg->msg_iov, size)) != 0) {
                        kfree_skb(skb);
index 2a27b84f740bc836a8c4d3debb5031209bf8441d..6d4a9a8de5ef1d8974a4f8fe17956dfd31b63fe6 100644 (file)
@@ -186,8 +186,17 @@ static void scan_inflight(struct sock *x, void (*func)(struct unix_sock *),
                                 */
                                struct sock *sk = unix_get_socket(*fp++);
                                if (sk) {
-                                       hit = true;
-                                       func(unix_sk(sk));
+                                       struct unix_sock *u = unix_sk(sk);
+
+                                       /*
+                                        * Ignore non-candidates, they could
+                                        * have been added to the queues after
+                                        * starting the garbage collection
+                                        */
+                                       if (u->gc_candidate) {
+                                               hit = true;
+                                               func(u);
+                                       }
                                }
                        }
                        if (hit && hitlist != NULL) {
@@ -249,11 +258,11 @@ static void inc_inflight_move_tail(struct unix_sock *u)
 {
        atomic_long_inc(&u->inflight);
        /*
-        * If this is still a candidate, move it to the end of the
-        * list, so that it's checked even if it was already passed
-        * over
+        * If this still might be part of a cycle, move it to the end
+        * of the list, so that it's checked even if it was already
+        * passed over
         */
-       if (u->gc_candidate)
+       if (u->gc_maybe_cycle)
                list_move_tail(&u->link, &gc_candidates);
 }
 
@@ -267,6 +276,7 @@ void unix_gc(void)
        struct unix_sock *next;
        struct sk_buff_head hitlist;
        struct list_head cursor;
+       LIST_HEAD(not_cycle_list);
 
        spin_lock(&unix_gc_lock);
 
@@ -282,10 +292,14 @@ void unix_gc(void)
         *
         * Holding unix_gc_lock will protect these candidates from
         * being detached, and hence from gaining an external
-        * reference.  This also means, that since there are no
-        * possible receivers, the receive queues of these sockets are
-        * static during the GC, even though the dequeue is done
-        * before the detach without atomicity guarantees.
+        * reference.  Since there are no possible receivers, all
+        * buffers currently on the candidates' queues stay there
+        * during the garbage collection.
+        *
+        * We also know that no new candidate can be added onto the
+        * receive queues.  Other, non candidate sockets _can_ be
+        * added to queue, so we must make sure only to touch
+        * candidates.
         */
        list_for_each_entry_safe(u, next, &gc_inflight_list, link) {
                long total_refs;
@@ -299,6 +313,7 @@ void unix_gc(void)
                if (total_refs == inflight_refs) {
                        list_move_tail(&u->link, &gc_candidates);
                        u->gc_candidate = 1;
+                       u->gc_maybe_cycle = 1;
                }
        }
 
@@ -325,13 +340,23 @@ void unix_gc(void)
                list_move(&cursor, &u->link);
 
                if (atomic_long_read(&u->inflight) > 0) {
-                       list_move_tail(&u->link, &gc_inflight_list);
-                       u->gc_candidate = 0;
+                       list_move_tail(&u->link, &not_cycle_list);
+                       u->gc_maybe_cycle = 0;
                        scan_children(&u->sk, inc_inflight_move_tail, NULL);
                }
        }
        list_del(&cursor);
 
+       /*
+        * not_cycle_list contains those sockets which do not make up a
+        * cycle.  Restore these to the inflight list.
+        */
+       while (!list_empty(&not_cycle_list)) {
+               u = list_entry(not_cycle_list.next, struct unix_sock, link);
+               u->gc_candidate = 0;
+               list_move_tail(&u->link, &gc_inflight_list);
+       }
+
        /*
         * Now gc_candidates contains only garbage.  Restore original
         * inflight counters for these as well, and remove the skbuffs