On Oct 16, 2014, at 11:09 PM, Alex Wang <al...@nicira.com> wrote:
> Thx for the revision,
>
>
> +/* Removes 'rule' from 'cls', also destructing the 'rule'. */
> +static void
> +dpcls_remove(struct dpcls *cls, struct dpcls_rule *rule)
> +{
> + struct dpcls_subtable *subtable;
> +
> + ovs_assert(rule->mask);
> +
> + INIT_CONTAINER(subtable, rule->mask, mask);
> +
> + if (cmap_remove(&subtable->rules, &rule->cmap_node, rule->flow.hash)
> + == 0) {
> + dpcls_destroy_subtable(cls, subtable);
> + }
> +
> + rule->mask = NULL;
> +}
>
>
> The assignment here caused a segfault in test, since there could be pmd
> threads referring to the flow when dpcls_remove() is called. We should not
> reset the pointer.
>
Thank you for testing and debugging!
>
> 32 UDP flows Spirent test showed no observable improvement but let's
> wait for the per-pmd-thread classifier/flow-table~
>
This was expected, any improvement should be visible only when there is high
churn on the megaflow cache.
> Otherwise, all tested and looked good~
> Acked-by: Alex Wang <al...@nicira.com>
>
Pushed to master with the fix,
Jarno
_______________________________________________
dev mailing list
dev@openvswitch.org
http://openvswitch.org/mailman/listinfo/dev
