Re: [PATCH 01/11] mac80211: add TX fastpath

[Johannes Berg <johannes@sipsolutions.net>]

On Tue, 2015-04-21 at 09:54 +0300, Eliad Peller wrote:

> > Hey, somebody is reviewing my patches :-)
> >
> i didn't delve into them too much, but generally they look good :)

:)

> > To fix that, I think I can hold the lock longer, so that the lifetime of
> > the key and the fast_tx pointer are more closely correlated. If I
> > acquire the spinlock before checking for the key, then the CPU that
> > invalidates the key pointer cannot race in this way with another caller,
> > since the key pointer would (for this purpose) be protected by the lock.
> > Then either the CPU that deleted the key will have to wait (while the
> > key is still pretty much valid) and then will overwrite the fast_tx w/o
> > the key, or the other CPU will have to wait and will find the key
> > pointer changed/NULL already.
> >
> > Right? what do you think?
> 
> sounds correct.
> i guess taking rcu_lock is a valid option as well (for about the same
> reasons). so either one of them should be good.

I don't think that would be correct - that just prevents the key from
being freed while we hold it, whereas here we actually need to prevent
the key from being accessed.

Anyway - I've done a bit more testing on this and will likely merge it
(the fixed version of course) in the next couple of days.

johannes