[SRU][Bionic][Cosmic][PATCH 0/1] tcp: do not release socket ownership in tcp_close()

Ioanna Alifieraki ioanna-maria.alifieraki at canonical.com
Wed May 29 00:22:56 UTC 2019

BugLink: https://bugs.launchpad.net/bugs/1830813

SRU Justification

Impact: WARN_ON messages casued by race condition between the close of
a TCP socket and another process  inspecting the same socket.

The code of interest is the following; in tcp_close function :

While in release_sock(sk), sock_release_owner function is called which sets the sk->sk_lock.owned=0.
When WARN_ON(sock_owned_by_user(sk)) is called it expects to find that the socket is not
owned by anyone.
According to upstream commit 8873c064d1de579ea2341,
while a socket is being closed is possible that other threads find it in rtnetlink dump.
tcp_get_info() function acquires the socket lock ( and sets sk_lock.owned=1 ) for 
a short amount of time, however long enough to trigger this warning.

Fixed by upstream commit in v4.20:
Commit: 8873c064d1de579ea23412a6d3eee972593f142b
"tcp: do not release socket ownership in tcp_close()"

Commit 8873c064d1de579ea23412a6d3eee972 fixes this bug by delegating the release of ownership
(calling release_sock(sk)) to later; just before exiting tcp_close function.

Reporter has tested and verified test 4.15 test kernel for Bionic.
This bug is difficult to be reproduced locally because the race condition cannot 
be triggered in a deterministic way.
To hit this bug we need the following :
a) a process closing a socket and while the execution is between release_sock(s)
and WARN_ON(sock_owned_by_user(sk))
b) another process inspecting the same socket to get into tcp_get_info(), acquire
ownership of the socket and not release it until the first process reaches the

This scenario is difficult to be achieved in a testing environment.

Regression Potential:
As far as Bionic (4.15 kernel) is concerned the reporter of the bug has tested and
verified a test kernel with the fix.
Concerning Cosmic (4.18 kernel) the fix has not been tested.
However, given that 
a) this fix essentially removes the WARN_ON(sock_owned_by_user(sk))
and delegates the release of the ownership to later in the tcp_close function, and
b) the relevant code paths in 4.15 and 4.18 are largely the same
the regression potential should be minimal.   

Eric Dumazet (1):
  tcp: do not release socket ownership in tcp_close()

 include/net/sock.h |  1 +
 net/core/sock.c    |  2 +-
 net/ipv4/tcp.c     | 11 +++--------
 3 files changed, 5 insertions(+), 9 deletions(-)


More information about the kernel-team mailing list