Hi,
We'd like to be able to pass algorithm-specific parameters to
comp backends. As of this moment, crypto usees hard-coded default
values and does not permit any run-time algorithm configuration,
which in some cases simply disables the most interesting functionality.
E.g. zstd can be configured to use a pre-trained (in the user-space)
compression dictionary, which significantly changes algorithms
characteristics. Another, obvious and trivial example, is algorithms
compression level.
The problem is that we need to pass params to cra_init() function,
because for some algorithms that's the only place where configuration
can take place (e.g. zstd). Changing cra_init() to accept additional
`struct crypto_comp_params` looks to be a little intrusive so before
I write any patches I'd like to hear your thoughts.
On Tue, Apr 30, 2024 at 12:04:47PM +0900, Sergey Senozhatsky wrote:
> Hi,
>
> We'd like to be able to pass algorithm-specific parameters to
> comp backends. As of this moment, crypto usees hard-coded default
> values and does not permit any run-time algorithm configuration,
> which in some cases simply disables the most interesting functionality.
> E.g. zstd can be configured to use a pre-trained (in the user-space)
> compression dictionary, which significantly changes algorithms
> characteristics. Another, obvious and trivial example, is algorithms
> compression level.
>
> The problem is that we need to pass params to cra_init() function,
> because for some algorithms that's the only place where configuration
> can take place (e.g. zstd). Changing cra_init() to accept additional
> `struct crypto_comp_params` looks to be a little intrusive so before
> I write any patches I'd like to hear your thoughts.
>
Have you considered just using the zstd library (<linux/zstd.h>) directly?
Which kernel subsystem are you working on that uses compression?
- Eric
On (24/04/29 20:18), Eric Biggers wrote:
> On Tue, Apr 30, 2024 at 12:04:47PM +0900, Sergey Senozhatsky wrote:
> > Hi,
> >
> > We'd like to be able to pass algorithm-specific parameters to
> > comp backends. As of this moment, crypto usees hard-coded default
> > values and does not permit any run-time algorithm configuration,
> > which in some cases simply disables the most interesting functionality.
> > E.g. zstd can be configured to use a pre-trained (in the user-space)
> > compression dictionary, which significantly changes algorithms
> > characteristics. Another, obvious and trivial example, is algorithms
> > compression level.
> >
> > The problem is that we need to pass params to cra_init() function,
> > because for some algorithms that's the only place where configuration
> > can take place (e.g. zstd). Changing cra_init() to accept additional
> > `struct crypto_comp_params` looks to be a little intrusive so before
> > I write any patches I'd like to hear your thoughts.
> >
>
> Have you considered just using the zstd library (<linux/zstd.h>) directly?
Yes, sort of.
> Which kernel subsystem are you working on that uses compression?
Forgot to mention - this is for zram primarily (and potentially for zswap).
Both of which take advantage of crypto API and can be configured to use
various compression algorithms (per-device).
So, on the zram side, in theory, I can stop using crypto API and start
calling lzo/zstd/lz4/etc. compression/decompression functions directly
(in fact we used to do that many many years ago, before we switched to
cypto API).