Whamcloud - gitweb
LU-9971 lnet: use after free in lnet_discover_peer_locked() 91/38891/6
authorOlaf Weber <olaf.weber@hpe.com>
Tue, 12 Sep 2017 12:07:50 +0000 (14:07 +0200)
committerOleg Drokin <green@whamcloud.com>
Tue, 1 Sep 2020 03:47:19 +0000 (03:47 +0000)
When the lnet_net_lock is unlocked, the peer attached to an
lnet_peer_ni (found via lnet_peer_ni::lpni_peer_net->lpn_peer)
can change, and the old peer deallocated. If we are really
unlucky, then all the churn could give us a new, different,
peer at the same address in memory.

Change the reference counting on the lnet_peer lp so that it
is guaranteed to be alive when we relock the lnet_net_lock for
the cpt. When the reference count is dropped lp may go away if
it was unlinked, but the new peer is guaranteed to have a
different address, so we can still correctly determine whether
the peer changed and discovery should be redone.

LU-9971 lnet: fix peer ref counting

Exit from the loop after peer ref count has been incremented
to avoid wrong ref count.

The code makes sure that a peer is queued for discovery at most
once if discovery is disabled. This is done to use discovery
as a standard ping for gateways which do not have discovery feature
or discovery is disabled.

Signed-off-by: Olaf Weber <olaf.weber@hpe.com>
Change-Id: Ia44dce20074b27ec0e77d7c1908c6a44ec73d326
Reviewed-on: https://review.whamcloud.com/28944
Reviewed-by: Amir Shehata <ashehata@whamcloud.com>
Tested-by: Jenkins
Tested-by: Maloo <maloo@whamcloud.com>
Reviewed-by: James Simmons <uja.ornl@yahoo.com>
Reviewed-by: Oleg Drokin <green@whamcloud.com>
Reviewed-on: https://review.whamcloud.com/38891
Tested-by: jenkins <devops@whamcloud.com>
Reviewed-by: Chris Horn <chris.horn@hpe.com>
Reviewed-by: James Simmons <jsimmons@infradead.org>
lnet/lnet/peer.c

index 2c1b37c..135fb07 100644 (file)
@@ -2047,6 +2047,7 @@ lnet_discover_peer_locked(struct lnet_peer_ni *lpni, int cpt, bool block)
        DEFINE_WAIT(wait);
        struct lnet_peer *lp;
        int rc = 0;
+       int count = 0;
 
 again:
        lnet_net_unlock(cpt);
@@ -2059,27 +2060,38 @@ again:
         * zombie if we race with DLC, so we must check for that.
         */
        for (;;) {
+               /* Keep lp alive when the lnet_net_lock is unlocked */
+               lnet_peer_addref_locked(lp);
                prepare_to_wait(&lp->lp_dc_waitq, &wait, TASK_INTERRUPTIBLE);
                if (signal_pending(current))
                        break;
                if (the_lnet.ln_dc_state != LNET_DC_STATE_RUNNING)
                        break;
+               /*
+                * Don't repeat discovery if discovery is disabled. This is
+                * done to ensure we can use discovery as a standard ping as
+                * well for backwards compatibility with routers which do not
+                * have discovery or have discovery disabled
+                */
+               if (lnet_is_discovery_disabled(lp) && count > 0)
+                       break;
                if (lp->lp_dc_error)
                        break;
                if (lnet_peer_is_uptodate(lp))
                        break;
                lnet_peer_queue_for_discovery(lp);
+               count++;
+               CDEBUG(D_NET, "Discovery attempt # %d\n", count);
 
                /*
-                * if caller requested a non-blocking operation then
-                * return immediately. Once discovery is complete then the
-                * peer ref will be decremented and any pending messages
-                * that were stopped due to discovery will be transmitted.
+                * If caller requested a non-blocking operation then
+                * return immediately. Once discovery is complete any
+                * pending messages that were stopped due to discovery
+                * will be transmitted.
                 */
                if (!block)
                        break;
 
-               lnet_peer_addref_locked(lp);
                lnet_net_unlock(LNET_LOCK_EX);
                schedule();
                finish_wait(&lp->lp_dc_waitq, &wait);
@@ -2087,26 +2099,18 @@ again:
                lnet_peer_decref_locked(lp);
                /* Peer may have changed */
                lp = lpni->lpni_peer_net->lpn_peer;
-
-               /*
-                * Wait for discovery to complete, but don't repeat if
-                * discovery is disabled. This is done to ensure we can
-                * use discovery as a standard ping as well for backwards
-                * compatibility with routers which do not have discovery
-                * or have discovery disabled
-                */
-               if (lnet_is_discovery_disabled(lp))
-                       break;
        }
        finish_wait(&lp->lp_dc_waitq, &wait);
 
        lnet_net_unlock(LNET_LOCK_EX);
        lnet_net_lock(cpt);
-
+       lnet_peer_decref_locked(lp);
        /*
-        * If the peer has changed after we've discovered the older peer,
-        * then we need to discovery the new peer to make sure the
-        * interface information is up to date
+        * The peer may have changed, so re-check and rediscover if that turns
+        * out to have been the case. The reference count on lp ensured that
+        * even if it was unlinked from lpni the memory could not be recycled.
+        * Thus the check below is sufficient to determine whether the peer
+        * changed. If the peer changed, then lp must not be dereferenced.
         */
        if (lp != lpni->lpni_peer_net->lpn_peer)
                goto again;