On 17 October 2016 at 14:16, Johannes Berg <johan...@sipsolutions.net> wrote: > On Mon, 2016-10-17 at 14:06 +0100, Ard Biesheuvel wrote: >> >> Actually, while I think it will be worthwhile going forward to >> implement such an 'auxiliary data' feature in a generic way, I still >> think we should address the issue at hand without too much >> complication. >> >> If we pedal back to the version of 'mac80211: move struct aead_req >> off the stack' that uses kzalloc() instead of aead_request_alloc(), >> we can simply add some space for aad and/or zero, and get rid of >> the kmem cache entirely. >> >> If you're past this point already, i won't bother but otherwise I can >> rework 'mac80211: move struct aead_req off the stack' so that the >> other patch is no longer required (and IIRC, this is actually >> something you proposed yourself a couple of iterations ago?) > > Yes, I did consider that. > > It makes some sense, and I guess the extra memcpy() would be cheaper > than the extra alloc? > > I'd happily use that instead of the combination of my two patches. The > aead_request_alloc() is just a simple inline anyway, so no real problem > not using it. >
Indeed. And it keeps the clutter inside the aes_xxx.c files, which could easily be updated in the future to use some auxdata feature if it ever materializes. I think it would help this code, but also the ESP code you pointed out, to have some kind of 'ordered synchronous' CRYPTO_xxx flag, where the crypto API could manage the kmem cache and percpu pointers to allocations. This goes well beyond what we can do as a fix, though, so we need an intermediate solution in any case. Shall I propose the patch?