Re: [BUG] crypto: export() overran state buffer on test vector

[Eric Biggers]

On Thu, Feb 06, 2020 at 09:54:42AM +0100, Corentin Labbe wrote:
> Hello
> 
> When working on adding hash support on sun8i-ce, I made a simple version which always fallback.
> but booting it lead to this:
> [   52.274278] sun8i-ce 1c15000.crypto: Register sha1
> [   52.279286] sun8i-ce 1c15000.crypto: sun8i_hash_crainit statesize is 96
> [   52.285933] sun8i-ce 1c15000.crypto: Fallback for sha1-sun8i-ce is sha1-ce
> [   52.312423] shash_default_export descsize=104
> [   52.316021] alg: ahash: sha1-sun8i-ce export() overran state buffer on test vector 0, cfg=\"import/export\" statesize=96
> [   52.333189] sun8i-ce 1c15000.crypto: Register sha224
> [   52.338387] sun8i-ce 1c15000.crypto: sun8i_hash_crainit statesize is 104
> [   52.345097] sun8i-ce 1c15000.crypto: Fallback for sha224-sun8i-ce is sha224-ce
> [   52.371865] shash_default_export descsize=112
> [   52.375459] alg: ahash: sha224-sun8i-ce export() overran state buffer on test vector 0, cfg=\"import/export\" statesize=104
> [   52.393039] sun8i-ce 1c15000.crypto: Register sha256
> [   52.398219] sun8i-ce 1c15000.crypto: sun8i_hash_crainit statesize is 104
> [   52.404937] sun8i-ce 1c15000.crypto: Fallback for sha256-sun8i-ce is sha256-ce
> [   52.431476] shash_default_export descsize=112
> [   52.435073] alg: ahash: sha256-sun8i-ce export() overran state buffer on test vector 0, cfg=\"import/export\" statesize=104
> 
> For sha1, sha224 and sha256, my driver fail to pass the test.
> This is due to the fact that export() (and so shash_async_export/shash_default_export) use crypto_shash_descsize() as length but selftest expect it to be statesize.

That doesn't appear to actually be the problem.  shash_default_export() does
assume descsize == statesize, but it's only used when that's the case.
See shash_prepare_alg().  

> Just in case, this is my export code:
> int sun8i_hash_crainit(struct crypto_tfm *tfm)
> {
>         struct sun8i_hash_tfm_ctx *op = crypto_tfm_ctx(tfm);
>         struct ahash_alg *alg = __crypto_ahash_alg(tfm->__crt_alg);
>         struct sun8i_ce_alg_template *algt;
> 
>         memset(op, 0, sizeof(struct sun8i_hash_tfm_ctx));
> 
>         crypto_ahash_set_reqsize(__crypto_ahash_cast(tfm), sizeof(struct sun8i_hash_reqctx));
> 
>         op->fallback_tfm = crypto_alloc_ahash(crypto_tfm_alg_name(tfm), 0, CRYPTO_ALG_NEED_FALLBACK);
>         if (IS_ERR(op->fallback_tfm)) {
>                 dev_err(algt->ce->dev, "Fallback driver cound no be loaded\n");
>                 return PTR_ERR(op->fallback_tfm);
>         }
>         dev_info(op->ce->dev, "%s statesize is %u\n", __func__, algt->alg.hash.halg.statesize);
>         dev_info(op->ce->dev, "Fallback for %s is %s\n",
>                 crypto_tfm_alg_driver_name(tfm),
>                 crypto_tfm_alg_driver_name(&op->fallback_tfm->base));
>         return 0;
> }
> 
> int sun8i_hash_init(struct ahash_request *areq)
> {
>         struct sun8i_hash_reqctx *rctx = ahash_request_ctx(areq);
>         struct crypto_ahash *tfm = crypto_ahash_reqtfm(areq);
>         struct sun8i_hash_tfm_ctx *tfmctx = crypto_ahash_ctx(tfm);
> 
>         memset(rctx, 0, sizeof(struct sun8i_hash_reqctx));
> 
>         ahash_request_set_tfm(&rctx->fallback_req, tfmctx->fallback_tfm);
>         rctx->fallback_req.base.flags = areq->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
> 
>         return crypto_ahash_init(&rctx->fallback_req);
> }
> 
> int sun8i_hash_export(struct ahash_request *areq, void *out)
> {
>         struct sun8i_hash_reqctx *rctx = ahash_request_ctx(areq);
>         struct crypto_ahash *tfm = crypto_ahash_reqtfm(areq);
>         struct sun8i_hash_tfm_ctx *tfmctx = crypto_ahash_ctx(tfm);
> 
>         ahash_request_set_tfm(&rctx->fallback_req, tfmctx->fallback_tfm);
>         rctx->fallback_req.base.flags = areq->base.flags & CRYPTO_TFM_REQ_MAY_SLEEP;
>                                                                                 
>         return crypto_ahash_export(&rctx->fallback_req, out);                   
> }

It seems the actual problem is that you're doing the export using a fallback
algorithm, which may have a statesize different from the one you're setting.

But I'm not sure what you should do here, since the correct statesize can only
be known when a tfm is allocated, not when the algorithm is registered.

Possibly statesize needs to be made a property of the tfm (struct crypto_ahash
and crypto_shash) rather than the algorithm (struct hash_alg_common).

But are you sure you actually need a fallback algorithm for any hash algorithms
in your driver in the first place?

- Eric