Hi Alexander,
aahringo@xxxxxxxxxx wrote on Wed, 11 May 2022 09:09:40 -0400:
> Hi,
>
> On Tue, May 10, 2022 at 10:52 AM Miquel Raynal
> <miquel.raynal@xxxxxxxxxxx> wrote:
> >
> > Hi Alex,
> >
> > > > --- a/net/mac802154/tx.c
> > > > +++ b/net/mac802154/tx.c
> > > > @@ -106,6 +106,21 @@ ieee802154_tx(struct ieee802154_local *local, struct sk_buff *skb)
> > > > return NETDEV_TX_OK;
> > > > }
> > > >
> > > > +void ieee802154_hold_queue(struct ieee802154_local *local)
> > > > +{
> > > > + atomic_inc(&local->phy->hold_txs);
> > > > +}
> > > > +
> > > > +void ieee802154_release_queue(struct ieee802154_local *local)
> > > > +{
> > > > + atomic_dec(&local->phy->hold_txs);
> > > > +}
> > > > +
> > > > +bool ieee802154_queue_is_held(struct ieee802154_local *local)
> > > > +{
> > > > + return atomic_read(&local->phy->hold_txs);
> > > > +}
> > >
> > > I am not getting this, should the release_queue() function not do
> > > something like:
> > >
> > > if (atomic_dec_and_test(hold_txs))
> > > ieee802154_wake_queue(local);
> > >
> > > I think we don't need the test of "ieee802154_queue_is_held()" here,
> > > then we need to replace all stop_queue/wake_queue with hold and
> > > release?
> >
> > That's actually a good idea. I've implemented it and it looks nice too.
> > I'll clean this up and share a new version with:
> > - The wake call checked everytime hold_txs gets decremented
> > - The removal of the _queue_is_held() helper
> > - _wake/stop_queue() turned static
> > - _hold/release_queue() used everywhere
> >
>
> I think there is also a lock necessary for atomic inc/dec hitting zero
> and the stop/wake call afterwards...
Mmmh that is true, it can race. I've introduced a mutex (I think it's
safe but it can be turned into a spinlock if proven necessary) to
secure these increment/decrement+wakeup operations.
> ,there are also a lot of
> optimization techniques to only hold the lock for hitting zero cases
> in such areas. However we will see...
I am not aware of technical solutions to avoid the locking in these
cases, what do you have in mind? Otherwise I propose just to come up
with a working and hopefully solid solution and then we'll see how to
optimize.
Thanks,
Miquèl
