Re: [PATCH 01/04] Adding cipher mode context information to crypto_tfm

[Andrew Morton <akpm@osdl.org>]

Fruhwirth Clemens <clemens@endorphin.org> wrote:
>
> On Wed, 2005-02-09 at 17:19 -0800, Andrew Morton wrote:
> > Fruhwirth Clemens <clemens@endorphin.org> wrote:
> > >
> > > It must be
> > >  possible to process more than 2 mappings in softirq context.
> > 
> > Adding a few more fixmap slots wouldn't hurt anyone.  But if you want an
> > arbitrarily large number of them then no, we cannot do that.
> 
> What magnitude is "few more"? 2, 10, 100?

Not 100.  10 would seem excessive.

> > Taking more than one sleeping kmap at a time within the same process is
> > deadlocky, btw.  You can end up with N such tasks all holding one kmap and
> > waiting for someone else to release one.
> > 
> > Possibly one could arrange for the pages to not be in highmem at all.
> 
> Is there an easy way to bring pages to lowmem? The cryptoapi is called
> from the backlog of the networking stack, which is assigned in irq
> context first and processed softirq context.

Are networking frames ever allocated from highmem?  Don't think so.

> There is little opportunity
> to bringt stuff to low mem. And we can't bringt stuff to lowmem on our
> own as well, because (as I guess) this involves a page allocation, which
> would have to be GFP_ATOMIC, which can fail. So we would need fallback
> for the fallback mechanism, and that's still the tiny set of scratch
> buffers.. hairy business..

yup.

> Ok, what about the following plan:
> 
> If context == softirq, use kmap_atomic until they all used, fall-back to
> scratch buffers, and printk in some DEBUG mode:"Warning slow, redesign
> your client or raise the number of fixmaps".

Sounds painful.

> If context == user, use kmap_atomic until they all used, and fall-back
> to kmap.

Taking multiple kmaps can deadlock due to kmap exhaustion though.