[lustre-devel] [PATCH 04/49] lnet: lnet_notify sets route aliveness incorrectly
James Simmons
jsimmons at infradead.org
Wed Apr 14 21:01:56 PDT 2021
From: Chris Horn <chris.horn at hpe.com>
lnet_notify() modifies route aliveness in two ways:
1. By setting lp_alive field of the lnet_peer struct.
2. By setting lr_alive field of the lnet_route struct (via call to
lnet_set_route_aliveness())
In both cases, the aliveness value assigned is determined by a call
to lnet_is_peer_ni_alive(), but that value only reflects the aliveness
of a particular peer NI. A gateway may have multiple peer NIs, so the
aliveness of a gateway peer (lp_alive) is not necessarily equivalent
to the aliveness of one of its NIs. Furthermore, the lr_alive field
is only used to determine route aliveness for path selection if
discovery is disabled locally or on the gateway (see
lnet_find_route_locked() and lnet_is_route_alive()).
In general, we should not set lp_alive based on an lnet_notify()
call, and we should only set lr_alive if discovery is disabled. For
lr_alive specifically, we should only set it for those routes that
have the peer NI as a next-hop.
An exception to the above exists when the reset argument to
lnet_notify() is set. The gnilnd uses this flag in its calls to
lnet_notify() because gnilnd receives out-of-band notifications of
node up and down events. Thus, when gnilnd calls lnet_notify() we
actually know whether the gateway peer is up or down and we can set
lp_alive appropriately.
net lock/EX is held by other callers of lnet_set_route_aliveness, so
we do the same in lnet_notify().
Fixes: 938a22eaf2 ("lnet: discovery off route state update")
Fixes: 1cadb960f1 ("lnet: Add peer level aliveness information")
HPE-bug-id: LUS-9034
WC-bug-id: https://jira.whamcloud.com/browse/LU-13708
Lustre-commit: e24471a722a6f23f ("LU-13708 lnet: lnet_notify sets route aliveness incorrectly")
Signed-off-by: Chris Horn <chris.horn at hpe.com>
Reviewed-on: https://review.whamcloud.com/39160
Reviewed-by: Serguei Smirnov <ssmirnov at whamcloud.com>
Reviewed-by: James Simmons <jsimmons at infradead.org>
Reviewed-by: Oleg Drokin <green at whamcloud.com>
Signed-off-by: James Simmons <jsimmons at infradead.org>
---
net/lnet/lnet/router.c | 39 +++++++++++++++++++++++++++++++++------
1 file changed, 33 insertions(+), 6 deletions(-)
diff --git a/net/lnet/lnet/router.c b/net/lnet/lnet/router.c
index e030b16..ee3c15f 100644
--- a/net/lnet/lnet/router.c
+++ b/net/lnet/lnet/router.c
@@ -391,6 +391,7 @@ bool lnet_is_route_alive(struct lnet_route *route)
lnet_check_route_inconsistency(route);
}
+/* Must hold net_lock/EX */
static inline void
lnet_set_route_aliveness(struct lnet_route *route, bool alive)
{
@@ -405,6 +406,7 @@ bool lnet_is_route_alive(struct lnet_route *route)
}
}
+/* Must hold net_lock/EX */
void
lnet_router_discovery_ping_reply(struct lnet_peer *lp)
{
@@ -1745,6 +1747,37 @@ bool lnet_router_checker_active(void)
/* recalculate aliveness */
alive = lnet_is_peer_ni_alive(lpni);
+
+ lp = lpni->lpni_peer_net->lpn_peer;
+ /* If this is an LNet router then update route aliveness */
+ if (lp->lp_rtr_refcount) {
+ if (reset)
+ /* reset flag indicates gateway peer went up or down */
+ lp->lp_alive = alive;
+
+ /* If discovery is disabled, locally or on the gateway, then
+ * any routes using lpni as next-hop need to be updated
+ *
+ * NB: We can get many notifications while a route is down, so
+ * we try and avoid the expensive net_lock/EX here for the
+ * common case of receiving duplicate lnet_notify() calls (i.e.
+ * only grab EX lock when we actually need to update the route
+ * aliveness).
+ */
+ if (lnet_is_discovery_disabled(lp)) {
+ list_for_each_entry(route, &lp->lp_routes, lr_gwlist) {
+ if (route->lr_nid == lpni->lpni_nid &&
+ route->lr_alive != alive) {
+ lnet_net_unlock(0);
+ lnet_net_lock(LNET_LOCK_EX);
+ lnet_set_route_aliveness(route, alive);
+ lnet_net_unlock(LNET_LOCK_EX);
+ lnet_net_lock(0);
+ }
+ }
+ }
+ }
+
lnet_net_unlock(0);
if (ni && !alive)
@@ -1753,12 +1786,6 @@ bool lnet_router_checker_active(void)
cpt = lpni->lpni_cpt;
lnet_net_lock(cpt);
lnet_peer_ni_decref_locked(lpni);
- if (lpni && lpni->lpni_peer_net && lpni->lpni_peer_net->lpn_peer) {
- lp = lpni->lpni_peer_net->lpn_peer;
- lp->lp_alive = alive;
- list_for_each_entry(route, &lp->lp_routes, lr_gwlist)
- lnet_set_route_aliveness(route, alive);
- }
lnet_net_unlock(cpt);
return 0;
--
1.8.3.1
More information about the lustre-devel
mailing list