af_unix: limit unix_tot_inflight
authorEric Dumazet <eric.dumazet@gmail.com>
Wed, 24 Nov 2010 17:15:27 +0000 (09:15 -0800)
committerDavid S. Miller <davem@davemloft.net>
Wed, 24 Nov 2010 17:15:27 +0000 (09:15 -0800)
Vegard Nossum found a unix socket OOM was possible, posting an exploit
program.

My analysis is we can eat all LOWMEM memory before unix_gc() being
called from unix_release_sock(). Moreover, the thread blocked in
unix_gc() can consume huge amount of time to perform cleanup because of
huge working set.

One way to handle this is to have a sensible limit on unix_tot_inflight,
tested from wait_for_unix_gc() and to force a call to unix_gc() if this
limit is hit.

This solves the OOM and also reduce overall latencies, and should not
slowdown normal workloads.

Reported-by: Vegard Nossum <vegard.nossum@gmail.com>
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
net/unix/garbage.c

index c8df6fda0b1fcf124b65812f600710a3b3f17069..40df93d1cf35534a1722180b2ed1d0e783bce25a 100644 (file)
@@ -259,9 +259,16 @@ static void inc_inflight_move_tail(struct unix_sock *u)
 }
 
 static bool gc_in_progress = false;
+#define UNIX_INFLIGHT_TRIGGER_GC 16000
 
 void wait_for_unix_gc(void)
 {
+       /*
+        * If number of inflight sockets is insane,
+        * force a garbage collect right now.
+        */
+       if (unix_tot_inflight > UNIX_INFLIGHT_TRIGGER_GC && !gc_in_progress)
+               unix_gc();
        wait_event(unix_gc_wait, gc_in_progress == false);
 }