Re: [PATCH net] net_sched: act_mirred: full rcu conversion

On 16-09-11 11:12 PM, Cong Wang wrote:

On Fri, Sep 9, 2016 at 8:52 AM, John Fastabend [off-list ref] wrote:

quoted

On 16-09-08 10:26 PM, Cong Wang wrote:

quoted

On Thu, Sep 8, 2016 at 8:51 AM, Eric Dumazet [off-list ref] wrote:

quoted

On Thu, 2016-09-08 at 08:47 -0700, John Fastabend wrote:

quoted

Works for me. FWIW I find this plenty straightforward and don't really
see the need to make the hash table itself rcu friendly.

Acked-by: John Fastabend <redacted>

Yes, it seems this hash table is used in control path, with RTNL held
anyway.

Seriously? You never read hashtable in fast path?? I think you need
to wake up.

But the actions use refcnt'ing and should never be decremented to zero
as long as they can still be referenced by an active filter. If each
action handles its parameters like mirred/gact then I don't see why its
necessary.

This is correct, by "read" I meant "dereference", the tc actions
are now permanently stored in hashtable directly, so "reading"
a tc action is reading from hashtable.

Sorry if this wasn't clear.

OK, but with the current code there is no need to protect the hash table
with an RCU semantics. The ref counting ensures the hash table entries
are always available and any 'replace' commands on actions are handled
internally in the action itself with rcu_assign_pointer replacing old
params with new params. The fast path readers will always read a
consistent set of parameters in this scheme.

So no need for an rcu hash table. The reference count though should
likely be atomic because not all increments/decrements are protected by
RTNL lock.

.John