Re: [net-next v5 3/3] net/tls: Remove redundant array allocation.

[Dave Watson <davejwatson@fb.com>]

On 07/27/18 09:34 AM, Vakul Garg wrote:
> 
> 
> > -----Original Message-----
> > From: Dave Watson [mailto:davejwatson@fb.com]
> > Sent: Thursday, July 26, 2018 2:31 AM
> > To: Vakul Garg <vakul.garg@nxp.com>
> > Cc: David Miller <davem@davemloft.net>; netdev@vger.kernel.org;
> > borisp@mellanox.com; aviadye@mellanox.com; Doron Roberts-Kedes
> > <doronrk@fb.com>
> > Subject: Re: [net-next v5 3/3] net/tls: Remove redundant array allocation.
> > 
> > On 07/24/18 08:22 AM, Vakul Garg wrote:
> > > Will it be a bad idea to get rid of array 'sgin' on stack and simply
> > > kmalloc 'sgin' for whatever the number the number of pages returned by
> > iov_iter_npages()?
> > > We can allocate for sgout too with the same kmalloc().
> > >
> > > (Using a local array based 'sgin' is coming in the way to achieve
> > > sending multiple async record decryption requests to the accelerator
> > > without waiting for previous one to complete.)
> > 
> > Yes we could do this, and yes we would need to heap allocate if you want to
> > support multiple outstanding decryption requests.  I think async crypto
> > prevents any sort of zerocopy-fastpath, however.
> 
> We already do a aead_request_alloc (which internally does kmalloc).
> To mitigate the cost of kmalloc/kfree for sg lists and aad, I am allocating a 
> combined memory chunk for all of these and then segmenting it into
> aead_req, aad, sgin, sgout. This way there should be no extra cost for
> memory allocations in non-async.

Makes sense, sounds good to me.