Re: [PATCH 10/12] crypto: qat - use memzero_explicit() for algs

[Giovanni Cabiddu <giovanni.cabiddu@intel.com>]

On Fri, May 06, 2022 at 04:38:15PM +0200, Greg KH wrote:
> On Fri, May 06, 2022 at 10:54:07AM +0100, Giovanni Cabiddu wrote:
> > On Fri, May 06, 2022 at 11:22:39AM +0200, Greg KH wrote:
> > > On Fri, May 06, 2022 at 09:23:25AM +0100, Giovanni Cabiddu wrote:
> > > > Use memzero_explicit(), instead of a memset(.., 0, ..) in the
> > > > implementation of the algorithms, for buffers containing sensitive
> > > > information to ensure they are wiped out before free.
> > > > 
> > > > Cc: stable@vger.kernel.org
> > > > Signed-off-by: Giovanni Cabiddu <giovanni.cabiddu@intel.com>
> > > > Reviewed-by: Adam Guerin <adam.guerin@intel.com>
> > > > Reviewed-by: Wojciech Ziemba <wojciech.ziemba@intel.com>
> > > > ---
> > > >  drivers/crypto/qat/qat_common/qat_algs.c      | 20 +++++++++----------
> > > >  drivers/crypto/qat/qat_common/qat_asym_algs.c | 20 +++++++++----------
> > > >  2 files changed, 20 insertions(+), 20 deletions(-)
> > > > 
> > > > diff --git a/drivers/crypto/qat/qat_common/qat_algs.c b/drivers/crypto/qat/qat_common/qat_algs.c
> > > > index 873533dc43a7..c42df18e02b2 100644
> > > > --- a/drivers/crypto/qat/qat_common/qat_algs.c
> > > > +++ b/drivers/crypto/qat/qat_common/qat_algs.c
> > > > @@ -637,12 +637,12 @@ static int qat_alg_aead_newkey(struct crypto_aead *tfm, const u8 *key,
> > > >  	return 0;
> > > >  
> > > >  out_free_all:
> > > > -	memset(ctx->dec_cd, 0, sizeof(struct qat_alg_cd));
> > > > +	memzero_explicit(ctx->dec_cd, sizeof(struct qat_alg_cd));
> > > 
> > > This is for structure fields, why does memset() not work properly here?
> > > The compiler should always call this, it doesn't know what
> > > dma_free_coherent() does.  You are referencing this pointer after the
> > > memset() call so all should be working as intended here.
> > > 
> > > Because of this, I don't see why this change is needed.  Do you have
> > > reports of compilers not calling memset() for all of this properly?
> > Apologies, I had a wrong assumption.
> > Based on a comment in the memzero_explicit() documentation I assumed it
> > should be always used to zero sensitive data.
> > 
> >      * memzero_explicit - Fill a region of memory (e.g. sensitive
> >      *			  keying data) with 0s.
> 
> Yes, that's what it is for, when the compiler thinks it is "smarter than
> you" for stack variables.
> 
> It's great for functions like this:
> 	int foo(...)
> 	{
> 		struct key secret_key;
> 
> 		do something and set secret_key...
> 
> 		/* All done, clean up and return */
> 		memset(&secret_key, 0, sizeof(secret_key));
> 		return 0;
> 	}
> 
> For that, some compilers now go "hey, they just want to set this to 0
> and then never touch it again, that is pointless, let's not call
> memset() at all!".
> 
> But when you call:
> 	memset(foo->key, 0, sizeof(key));
> 	do_something_with_foo(foo);
> 
> the compiler can NOT go and ignore the call to memset() as it does not
> know what do_something_with_foo() does.  Or at least it better not.
> 
> Try out this with a small example and look at the asm output for proof.
Thanks for the explanation. It is clear now.

> 
> You aren't the first to be confused about this, I see this happening at
> least once a month with a patch to change code like you did.  Don't know
> why it keeps coming up, is this a checkpatch() recommentation?
It is not a checkpatch recommendation.
I got that assumption looking at kfree_sensitive() which contains a call
to memzero_explicit(). This was introduced in 2020 by
8982ae527fbe ("mm/slab: use memzero_explicit() in kzfree()" when the
function was still called kzfree().
I assume now that the call to memzero_explicit() in kfree_sensitive() is
also redundant, unless I'm missing something.

Regards,

-- 
Giovanni