On Wed, 2015-11-25 at 22:17 -0500, Aaron Conole wrote:

Probably that would call for a different more primitive version of this
API (sk_gfp_or_memalloc() as you suggest below). Then this could be
written in terms of that

static inline sk_gfp_or_memalloc(const struct sock *sk, gfp_t gfp_mask)
{
	return gfp_mask | (sk->sk_allocation & __GFP_MEMALLOC);
}

static inline sk_gfp_atomic(const struct sock *sk, gfp_t gfp_mask)
{
	return sk_gfp_or_memalloc(sk, gfp_mask | GFP_ATOMIC);
}

Not sure if it's "too much API".

Well, this looks like it, not sure how this is going to make code
clearer.

The only thing we bring from sk is the __GFP_MEMALLOC thing, so a single
function seems enough ?

I honestly do not care that much about function names, I mostly look at
actual implementation. And current implementation ignores the gfp_t
gfp_mask argument, for no real good reason.