Commit 1215e51edad1 "ipv4: fix a deadlock in ip_ra_control"
made rtnl_lock() be used in raw_close(). This function is called
on every RAW socket destruction, so that rtnl_mutex is taken
every time. This scales very sadly. I observe cleanup_net()
spending a lot of time in rtnl_lock() and raw_close() is one
of the biggest rtnl user (since we have percpu net->ipv4.icmp_sk).
Another problem is that commit does not explain actual call path
mrtsock_destruct() takes sk lock and the way to deadlock.
But there is no sk lock taking is visible in mrtsock_destruct().
So, it is a question does we need it at all.
This patchset reworks the locking: reverts the problem commit
and its descendant, and introduces rtnl-independent locking.
This may have a continuation, and someone may work on killing
rtnl_lock() in mrtsock_destruct() in the future.
Thanks,
Kirill
---
Kirill Tkhai (5):
net: Revert "ipv4: get rid of ip_ra_lock"
net: Revert "ipv4: fix a deadlock in ip_ra_control"
net: Move IP_ROUTER_ALERT out of lock_sock(sk)
net: Make ip_ra_chain per struct net
net: Replace ip_ra_lock with per-net mutex
include/net/ip.h | 13 +++++++++++--
include/net/netns/ipv4.h | 2 ++
net/core/net_namespace.c | 1 +
net/ipv4/ip_input.c | 5 ++---
net/ipv4/ip_sockglue.c | 34 +++++++++++++---------------------
net/ipv4/ipmr.c | 11 +++++++++--
net/ipv4/raw.c | 2 --
7 files changed, 38 insertions(+), 30 deletions(-)
--
Signed-off-by: Kirill Tkhai <ktk...@virtuozzo.com>
