sctp: fix autoclose timer race
Quoted 7 lines · Andrei Pelinescu-Onciul
For very small autoclose timeouts (e.g. 0 or close to 0 due to overflows) it's possible that the timer is executed immediately after is set, before there is a chance to increase the association reference counter. To avoid this race, always reference the association before calling
mod_timer()and dereference it back if the timer was already active.
Actually, I don't think this can happen. At worst, you'd trigger a soft lock-up, since the code modifying the timer is already holding a socket lock. But that would only happen if that code is pre-empted, which would mean that a user version of the lock is held, then the autoclose would get rescheduled anyway.
-vlad
Quoted 21 lines
Signed-off-by Andrei Pelinescu-Onciul <[email protected]>--- net/sctp/output.c | 5 +++-- 1 files changed, 3 insertions(+), 2 deletions(-)net/sctp/output.c +3
-2@@ -586,8 +586,9 @@ int sctp_packet_transmit(struct sctp_packet *packet) timer = &asoc->timers[SCTP_EVENT_TIMEOUT_AUTOCLOSE ]; timeout = asoc->timeouts[SCTP_EVENT_TIMEOUT_AUTOCLOSE ]; -if (!mod_timer (timer, jiffies + timeout)) -sctp_association_hold (asoc); +sctp_association_hold (asoc); +if (mod_timer (timer, jiffies + timeout)) +sctp_association_put (asoc);/* already active, deref */ } }
Quoted 10 lines · Vlad Yasevich
Quoted 7 lines · Andrei Pelinescu-Onciul
For very small autoclose timeouts (e.g. 0 or close to 0 due to overflows) it's possible that the timer is executed immediately after is set, before there is a chance to increase the association reference counter. To avoid this race, always reference the association before calling
mod_timer()and dereference it back if the timer was already active.
Quoted 6 lines
Actually, I don't think this can happen. At worst, you'd trigger a soft lock-up, since the code modifying the timer is already holding a socket lock. But that would only happen if that code is pre-empted, which would mean that a user version of the lock is held, then the autoclose would get rescheduled anyway.
Even if the timer ends up running on a different CPU, the socket lock guarantees that only 1 instance of the state machine is running per socket. The timer handler will spin on the lock until it's released, so there doesn't appear to be a race.
-vlad