On 14 October 2016 at 10:25, Johannes Berg [off-list ref] wrote:

On Fri, 2016-10-14 at 10:21 +0100, Ard Biesheuvel wrote:

quoted

It is annotated with a TODO, though :-)

38320c70d282b (Herbert Xu               2008-01-28 19:35:05
-0800  41)
 * TODO: Use spare space in skb for this where possible.

I saw that, but I don't think generally there will be spare space for
it - the stuff there is likely far too big. Anyway ... same problem
that we have.

I'm not inclined to allocate ~500 bytes temporarily for every frame
either though.

Maybe we could try to manage it in mac80211, we'd "only" need 5 AEAD
structs (which are today on the stack) in parallel for each key (4 TX,
1 RX), but in a typical case of having 3 keys that's already 7.5K worth
of memory that we almost never use. Again, with more complexity, we
could know that the TX will not be used if the driver does the TX, but
the single RX one we'd need unconditionally... decisions decisions...

So why is the performance hit acceptable for ESP but not for WPA? We
could easily implement the same thing, i.e., kmalloc(GFP_ATOMIC)/kfree
the aead_req struct rather than allocate it on the stack