Re: [PATCH] crypto: scompress - eliminate percpu scratch buffers

On 21 July 2017 at 13:42, Herbert Xu [off-list ref] wrote:

On Thu, Jul 20, 2017 at 12:40:00PM +0100, Ard Biesheuvel wrote:

quoted

The scompress code unconditionally allocates 2 per-CPU scratch buffers
of 128 KB each, in order to avoid allocation overhead in the async
wrapper that encapsulates the synchronous compression algorithm, since
it may execute in atomic context.

The whole point of pre-allocation is that we cannot allocate 128K
(or 64K as it was before scomp) at run-time, and in particular,
for IPsec which runs in softirq path.  Am I missing something?

Right. And is req->dst guaranteed to be assigned in that case? Because
crypto_scomp_sg_alloc() happily allocates pages and kmalloc()s the
scatterlist if req->dst == NULL.

Is there any way we could make these scratch buffers part of the
request structure instead? Or at least defer allocating them until the
first call to crypto_scomp_init_tfm()? And on top of that, we should
probably only use the per-CPU scratch buffers if
CRYPTO_TFM_REQ_MAY_SLEEP is cleared, because in that case, we are not
pre-emptible to begin with, and the concern does not apply.