Resolve deadlock between SIP registration and channel based functions

In r373424, several reentrancy problems in chan_sip were addressed. As a
result, the SIP channel driver is now properly locking the channel driver
private information in certain operations that it wasn't previously. This
exposed two latent problems either in register_verify or by functions called
by register_verify. This includes:
 * Holding the private lock while calling sip_send_mwi_to_peer. This can create
   a new sip_pvt via sip_alloc, which will obtain the channel container lock.
   This is a locking inversion, as any channel related lock must be obtained
   prior to obtaining the SIP channel technology private lock.
 * Holding the privat elock while calling sip_poke_peer. In the same vein as
   sip_send_mwi_to_peer, sip_poke_peer can create a new SIP private, causing
   the same locking inversion.

Note that this locking inversion typically occured when CLI commands were run
while a SIP REGISTER request was being processed, as many CLI commands (such
as 'sip show channels', 'core show channels', etc.) have to obtain the channel
container lock.

(issue ASTERISK-21068)
Reported by: Nicolas Bouliane

(issue ASTERISK-20550)
Reported by: David Brillert

(issue ASTERISK-21314)
Reported by: Badalian Vyacheslav

(issue ASTERISK-21296)
Reported by: Gabriel Birke



git-svn-id: https://origsvn.digium.com/svn/asterisk/branches/1.8@383863 65c4cc65-6c06-0410-ace0-fbb531ad65f3
This commit is contained in:
Matthew Jordan
2013-03-26 02:23:01 +00:00
parent 21171e1684
commit 7079459be4

View File

@@ -14667,7 +14667,9 @@ static enum parse_register_result parse_register_contact(struct sip_pvt *pvt, st
ast_verbose(VERBOSE_PREFIX_3 "Registered SIP '%s' at %s\n", peer->name,
ast_sockaddr_stringify(&peer->addr));
}
sip_pvt_unlock(pvt);
sip_poke_peer(peer, 0);
sip_pvt_lock(pvt);
register_peer_exten(peer, 1);
/* Save User agent */
@@ -15455,7 +15457,9 @@ static enum check_auth_result register_verify(struct sip_pvt *p, struct ast_sock
}
if (!res) {
if (send_mwi) {
sip_pvt_unlock(p);
sip_send_mwi_to_peer(peer, 0);
sip_pvt_lock(p);
} else {
update_peer_lastmsgssent(peer, -1, 0);
}
@@ -26965,6 +26969,9 @@ static int sip_poke_noanswer(const void *data)
\note This is done with 60 seconds between each ping,
unless forced by cli or manager. If peer is unreachable,
we check every 10th second by default.
\note Do *not* hold a pvt lock while calling this function.
This function calls sip_alloc, which can cause a deadlock
if another sip_pvt is held.
*/
static int sip_poke_peer(struct sip_peer *peer, int force)
{