Re: [PATCH] crypto: do not free algorithm before using

[PanBian]

On Thu, Nov 22, 2018 at 10:44:41PM +0800, Herbert Xu wrote:
> On Thu, Nov 22, 2018 at 06:00:16PM +0800, Pan Bian wrote:
> > In multiple functions, the algorithm fields are read after its reference
> > is dropped through crypto_mod_put. In this case, the algorithm memory
> > may be freed, resulting in use-after-free bugs. This patch delays the
> > put operation until the algorithm is never used.
> > 
> > Signed-off-by: Pan Bian <bianpan2016@xxxxxxx>
> 
> I don't think this patch is needed.
> 
> > ---
> >  crypto/cbc.c  | 6 ++++--
> >  crypto/cfb.c  | 6 ++++--
> >  crypto/pcbc.c | 6 ++++--
> >  3 files changed, 12 insertions(+), 6 deletions(-)
> > 
> > diff --git a/crypto/cbc.c b/crypto/cbc.c
> > index b761b1f..dd5f332 100644
> > --- a/crypto/cbc.c
> > +++ b/crypto/cbc.c
> > @@ -140,9 +140,8 @@ static int crypto_cbc_create(struct crypto_template *tmpl, struct rtattr **tb)
> >  	spawn = skcipher_instance_ctx(inst);
> >  	err = crypto_init_spawn(spawn, alg, skcipher_crypto_instance(inst),
> >  				CRYPTO_ALG_TYPE_MASK);
> > -	crypto_mod_put(alg);
> >  	if (err)
> > -		goto err_free_inst;
> > +		goto err_put_alg;
> 
> We can safely drop the reference to the algorithm because the spawn
> is now meant to hold a reference to it.  As long as the spawn is
> alive so will the algorithm.

Thanks for your explanation! But I find that the function
crypto_init_spawn just lets spawn->alg point to the algorithm without 
increasing the reference count, i.e., alg->cra_refcnt. So I am confused
about how this can protect the algorithm from being freed. Maybe I
missed some key points. Could you please explain it in more details?

Thank you!

Best regards,
Pan Bian

> 
> Cheers,
> -- 
> Email: Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx>
> Home Page: http://gondor.apana.org.au/~herbert/
> PGP Key: http://gondor.apana.org.au/~herbert/pubkey.txt