Re: [PATCH 1/2] ipv6: do not hold route table lock when send ndisc probe

On Thu, 2012-08-23 at 18:10 +0200, Eric Dumazet wrote:

On Tue, 2012-08-21 at 11:44 +0800, Cong Wang wrote:

quoted

Hi, Debabrata,

Could you help to test the attached patch below?

Thanks!

Hard to comment on your patch since its not inlined.

+               nw = kmalloc(sizeof(*nw), GFP_ATOMIC);
+               if (nw) {
+                       memcpy(&nw->target, &neigh->primary_key, sizeof(struct in6_addr));
+                       addrconf_addr_solict_mult(&nw->target, &nw->mcaddr);
+                       nw->dev = rt->dst.dev;
+                       INIT_WORK(&nw->work, queue_ndisc);
+                       schedule_work(&nw->work);
+               }

You cant do that without taking extra reference on dev,
and release it in queue_ndisc()

This also will add interesting side effects at device dismantle.

Right... If we call dev_hold() in the work, it is possible that the work
is still not scheduled to running when we unregister the device. What's
more, we can't flush work here as we are holding a read lock.

Hmm.