In case of removing ib_ipoib module ipoib_ib_dev_stop() function will be
called and all address handles (ah) in dead_ahs list will be reaped.
But some ah will be added to the dead list after ipoib_ib_dev_stop done
by ipoib_mcast_free. These ahs will not be freed.
The solution here is to wait till multicast_list will be empty. So, all
ahs will be added to dead_ahs list.
Signed-off-by: Vladimir Sokolovsky <[EMAIL PROTECTED]>
---
Roland,
There may be some extremely slight window for leaking address handles
still, since the multicast list is emptied in ipoib_mcast_dev_flush() before
it calls ipoib_mcast_free() (which adds address handles to the dead list).
However, this seems to be the best compromise that I can see without
a lot of nasty (and possibly buggy) changes.
drivers/infiniband/ulp/ipoib/ipoib_ib.c | 2 +-
1 files changed, 1 insertions(+), 1 deletions(-)
diff --git a/drivers/infiniband/ulp/ipoib/ipoib_ib.c
b/drivers/infiniband/ulp/ipoib/ipoib_ib.c
index 66cafa2..6cc0c59 100644
--- a/drivers/infiniband/ulp/ipoib/ipoib_ib.c
+++ b/drivers/infiniband/ulp/ipoib/ipoib_ib.c
@@ -863,7 +863,7 @@ timeout:
begin = jiffies;
- while (!list_empty(&priv->dead_ahs)) {
+ while (!list_empty(&priv->dead_ahs) ||
!list_empty(&priv->multicast_list)) {
__ipoib_reap_ah(dev);
if (time_after(jiffies, begin + HZ)) {
--
1.5.6.3
_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
