ipv4: fix race in concurrent ip_route_input_slow()
authorAlexei Starovoitov <ast@plumgrid.com>
Wed, 20 Nov 2013 03:12:34 +0000 (19:12 -0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sun, 8 Dec 2013 15:29:26 +0000 (07:29 -0800)
commit62713c4b6bc10c2d082ee1540e11b01a2b2162ab
tree0c6232d64c58c883c2db4289c9066826861fb574
parent8392415810ff70ffb61ed1d4dbe5672ce34de9a2
ipv4: fix race in concurrent ip_route_input_slow()

[ Upstream commit dcdfdf56b4a6c9437fc37dbc9cee94a788f9b0c4 ]

CPUs can ask for local route via ip_route_input_noref() concurrently.
if nh_rth_input is not cached yet, CPUs will proceed to allocate
equivalent DSTs on 'lo' and then will try to cache them in nh_rth_input
via rt_cache_route()
Most of the time they succeed, but on occasion the following two lines:
orig = *p;
prev = cmpxchg(p, orig, rt);
in rt_cache_route() do race and one of the cpus fails to complete cmpxchg.
But ip_route_input_slow() doesn't check the return code of rt_cache_route(),
so dst is leaking. dst_destroy() is never called and 'lo' device
refcnt doesn't go to zero, which can be seen in the logs as:
unregister_netdevice: waiting for lo to become free. Usage count = 1
Adding mdelay() between above two lines makes it easily reproducible.
Fix it similar to nh_pcpu_rth_output case.

Fixes: d2d68ba9fe8b ("ipv4: Cache input routes in fib_info nexthops.")
Signed-off-by: Alexei Starovoitov <ast@plumgrid.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
net/ipv4/route.c