Re: [PATCH v4] sctp: fix ASCONF list handling
From: David Miller <davem@davemloft.net>
Date: 2015-06-11 23:31:23
Also in:
linux-sctp
From: mleitner@redhat.com Date: Thu, 11 Jun 2015 11:30:46 -0300
Attempts to circumvent this lock invertion with RCU and/or list splicing
were unsuccessful, as they led to more and more code to handle it
properly.
Back when Hannes started reviewing the patches, he had asked if I
couldn't take the lock earlier during the socket destruction. I had said
no because sctp_destroy_sock() is called with socket lock already held
on sctp_close_sock() and such would not be possible to handle on error
handling situations like when sctp_init_sock() fails and
sctp_destroy_sock() is called right after that.
But if we take care that nothing fails after initializing asconf on
sctp_init_sock(), this is possible, and less complicated than my RCU and
list splicing attempts.This is definitely a cleaner/simpler fix, but:
quoted hunk ↗ jump to hunk
@@ -1528,7 +1528,10 @@ static void sctp_close(struct sock *sk, long timeout) /* Supposedly, no process has access to the socket, but * the net layers still may. + * Also, sctp_destroy_sock() needs to be called with addr_wq_lock + * held and that should be grabbed before socket lock. */ + spin_lock_bh(&net->sctp.addr_wq_lock); local_bh_disable(); bh_lock_sock(sk);@@ -1540,6 +1543,7 @@ static void sctp_close(struct sock *sk, long timeout) bh_unlock_sock(sk); local_bh_enable(); + spin_unlock_bh(&net->sctp.addr_wq_lock); sock_put(sk);
The local_bh_{enable,disable}() now appear to be superfluous and thus
can be removed.