net: unix: properly re-increment inflight counter of GC discarded candidates
authorAndrey Ulanov <andreyu@google.com>
Wed, 15 Mar 2017 03:16:42 +0000 (20:16 -0700)
committerDavid S. Miller <davem@davemloft.net>
Tue, 21 Mar 2017 22:25:10 +0000 (15:25 -0700)
Dmitry has reported that a BUG_ON() condition in unix_notinflight()
may be triggered by a simple code that forwards unix socket in an
SCM_RIGHTS message.
That is caused by incorrect unix socket GC implementation in unix_gc().

The GC first collects list of candidates, then (a) decrements their
"children's" inflight counter, (b) checks which inflight counters are
now 0, and then (c) increments all inflight counters back.
(a) and (c) are done by calling scan_children() with inc_inflight or
dec_inflight as the second argument.

Commit 6209344f5a37 ("net: unix: fix inflight counting bug in garbage
collector") changed scan_children() such that it no longer considers
sockets that do not have UNIX_GC_CANDIDATE flag. It also added a block
of code that that unsets this flag _before_ invoking
scan_children(, dec_iflight, ). This may lead to incorrect inflight
counters for some sockets.

This change fixes this bug by changing order of operations:
UNIX_GC_CANDIDATE is now unset only after all inflight counters are
restored to the original state.

  kernel BUG at net/unix/garbage.c:149!
  RIP: 0010:[<ffffffff8717ebf4>]  [<ffffffff8717ebf4>]
  unix_notinflight+0x3b4/0x490 net/unix/garbage.c:149
  Call Trace:
   [<ffffffff8716cfbf>] unix_detach_fds.isra.19+0xff/0x170 net/unix/af_unix.c:1487
   [<ffffffff8716f6a9>] unix_destruct_scm+0xf9/0x210 net/unix/af_unix.c:1496
   [<ffffffff86a90a01>] skb_release_head_state+0x101/0x200 net/core/skbuff.c:655
   [<ffffffff86a9808a>] skb_release_all+0x1a/0x60 net/core/skbuff.c:668
   [<ffffffff86a980ea>] __kfree_skb+0x1a/0x30 net/core/skbuff.c:684
   [<ffffffff86a98284>] kfree_skb+0x184/0x570 net/core/skbuff.c:705
   [<ffffffff871789d5>] unix_release_sock+0x5b5/0xbd0 net/unix/af_unix.c:559
   [<ffffffff87179039>] unix_release+0x49/0x90 net/unix/af_unix.c:836
   [<ffffffff86a694b2>] sock_release+0x92/0x1f0 net/socket.c:570
   [<ffffffff86a6962b>] sock_close+0x1b/0x20 net/socket.c:1017
   [<ffffffff81a76b8e>] __fput+0x34e/0x910 fs/file_table.c:208
   [<ffffffff81a771da>] ____fput+0x1a/0x20 fs/file_table.c:244
   [<ffffffff81483ab0>] task_work_run+0x1a0/0x280 kernel/task_work.c:116
   [<     inline     >] exit_task_work include/linux/task_work.h:21
   [<ffffffff8141287a>] do_exit+0x183a/0x2640 kernel/exit.c:828
   [<ffffffff8141383e>] do_group_exit+0x14e/0x420 kernel/exit.c:931
   [<ffffffff814429d3>] get_signal+0x663/0x1880 kernel/signal.c:2307
   [<ffffffff81239b45>] do_signal+0xc5/0x2190 arch/x86/kernel/signal.c:807
   [<ffffffff8100666a>] exit_to_usermode_loop+0x1ea/0x2d0
  arch/x86/entry/common.c:156
   [<     inline     >] prepare_exit_to_usermode arch/x86/entry/common.c:190
   [<ffffffff81009693>] syscall_return_slowpath+0x4d3/0x570
  arch/x86/entry/common.c:259
   [<ffffffff881478e6>] entry_SYSCALL_64_fastpath+0xc4/0xc6

Link: https://lkml.org/lkml/2017/3/6/252
Signed-off-by: Andrey Ulanov <andreyu@google.com>
Reported-by: Dmitry Vyukov <dvyukov@google.com>
Fixes: 6209344 ("net: unix: fix inflight counting bug in garbage collector")
Signed-off-by: David S. Miller <davem@davemloft.net>
net/unix/garbage.c

index 6a0d48525fcf9a71f54bb43495b200b300f5341e..c36757e728442bb936c17f8a975b420f5f8a5972 100644 (file)
@@ -146,6 +146,7 @@ void unix_notinflight(struct user_struct *user, struct file *fp)
        if (s) {
                struct unix_sock *u = unix_sk(s);
 
+               BUG_ON(!atomic_long_read(&u->inflight));
                BUG_ON(list_empty(&u->link));
 
                if (atomic_long_dec_and_test(&u->inflight))
@@ -341,6 +342,14 @@ void unix_gc(void)
        }
        list_del(&cursor);
 
+       /* Now gc_candidates contains only garbage.  Restore original
+        * inflight counters for these as well, and remove the skbuffs
+        * which are creating the cycle(s).
+        */
+       skb_queue_head_init(&hitlist);
+       list_for_each_entry(u, &gc_candidates, link)
+               scan_children(&u->sk, inc_inflight, &hitlist);
+
        /* not_cycle_list contains those sockets which do not make up a
         * cycle.  Restore these to the inflight list.
         */
@@ -350,14 +359,6 @@ void unix_gc(void)
                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
-        * which are creating the cycle(s).
-        */
-       skb_queue_head_init(&hitlist);
-       list_for_each_entry(u, &gc_candidates, link)
-       scan_children(&u->sk, inc_inflight, &hitlist);
-
        spin_unlock(&unix_gc_lock);
 
        /* Here we are. Hitlist is filled. Die. */