On Wed 17 Apr 2019 at 19:34, Jakub Kicinski <jakub.kicin...@netronome.com> 
wrote:
> On Wed, 17 Apr 2019 07:29:36 +0000, Vlad Buslov wrote:
>> On Wed 17 Apr 2019 at 00:49, Jakub Kicinski <jakub.kicin...@netronome.com> 
>> wrote:
>> > On Tue, 16 Apr 2019 17:20:47 +0300, Vlad Buslov wrote:
>> >> @@ -1551,6 +1558,10 @@ static int fl_change(struct net *net, struct 
>> >> sk_buff *in_skb,
>> >>           goto errout_mask;
>> >>
>> >>   if (!tc_skip_hw(fnew->flags)) {
>> >> +         spin_lock(&tp->lock);
>> >> +         list_add(&fnew->hw_list, &head->hw_filters);
>> >> +         spin_unlock(&tp->lock);
>> >> +
>> >>           err = fl_hw_replace_filter(tp, fnew, rtnl_held, extack);
>> >>           if (err)
>> >>                   goto errout_ht;
>> >
>> > Duplicated deletes should be fine, but I'm not sure same is true for
>> > adds.  Won't seeing an add with the same cookie twice confuse drivers?
>> >
>> > There's also the minor issue of offloaded count being off in that
>> > case :)
>>
>> Hmmm, okay. Rejecting duplicate cookies should be a trivial change to
>> drivers though. Do you see any faults with this approach in general?
>
> Trivial or not it adds up, the stack should make driver authors' job as
> easy as possible.  The simplest thing to do would be to add a mutex
> around the HW calls.  But that obviously doesn't work for you, cause
> you want multiple outstanding requests to the FW for a single tp, right?
>
> How about a RW lock, that would take R on normal add/replace/del paths
> and W on replays?  That should scale, no?

I've been thinking some more about possible ways to mitigate the
problem. First of all I tried to implement POC of rwlock in flower and
it isn't straightforward because of lock ordering. Observe that
fl_reoffload() is always called with rtnl lock taken (I didn't do any work to
unlock bind/unbind), but fl_change() can be called without rtnl lock and
needs to obtain it before offloading rules. This means that we have
deadlock here, if fl_change() obtains locks in order rwlock --->
rtnl_lock and fl_reoffload() obtains locks in order rtnl_lock --->
rwlock.

Considering this, I tried to improve my solution to remove possibility
of multiple adds of same filter and it seems to me that it would be
enough to move hw_filters list management in flower offloads functions:
add filter to list while holding rtnl lock in fl_hw_replace_filter() and
remove it from list while holding rtnl lock in fl_hw_destroy_filter().
What do you think?

Reply via email to