On 6 September 2018 at 09:21, Ard Biesheuvel [off-list ref] wrote:

On 6 September 2018 at 06:53, Gilad Ben-Yossef [off-list ref] wrote:

quoted

On Thu, Sep 6, 2018 at 1:49 AM, Ard Biesheuvel
[off-list ref] wrote:

quoted

On 5 September 2018 at 23:05, Kees Cook [off-list ref] wrote:

quoted

On Wed, Sep 5, 2018 at 2:18 AM, Ard Biesheuvel
[off-list ref] wrote:

quoted

On 4 September 2018 at 20:16, Kees Cook [off-list ref] wrote:

quoted

In the quest to remove all stack VLA usage from the kernel[1], this
caps the skcipher request size similar to other limits and adds a sanity
check at registration. Looking at instrumented tcrypt output, the largest
is for lrw:

        crypt: testing lrw(aes)
        crypto_skcipher_set_reqsize: 8
        crypto_skcipher_set_reqsize: 88
        crypto_skcipher_set_reqsize: 472

Are you sure this is a representative sampling? I haven't double
checked myself, but we have plenty of drivers for peripherals in
drivers/crypto that implement block ciphers, and they would not turn
up in tcrypt unless you are running on a platform that provides the
hardware in question.

Hrm, excellent point. Looking at this again:

The core part of the VLA is using this in the ON_STACK macro:

static inline unsigned int crypto_skcipher_reqsize(struct crypto_skcipher *tfm)
{
        return tfm->reqsize;
}

I don't find any struct crypto_skcipher .reqsize static initializers,
and the initial reqsize is here:

static int crypto_init_skcipher_ops_ablkcipher(struct crypto_tfm *tfm)
{
...
        skcipher->reqsize = crypto_ablkcipher_reqsize(ablkcipher) +
                            sizeof(struct ablkcipher_request);

with updates via crypto_skcipher_set_reqsize().

So I have to examine ablkcipher reqsize too:

static inline unsigned int crypto_ablkcipher_reqsize(
        struct crypto_ablkcipher *tfm)
{
        return crypto_ablkcipher_crt(tfm)->reqsize;
}

And of the crt_ablkcipher.reqsize assignments/initializers, I found:

ablkcipher reqsize:
1       struct dcp_aes_req_ctx
8       struct atmel_tdes_reqctx
8       struct cryptd_blkcipher_request_ctx
8       struct mtk_aes_reqctx
8       struct omap_des_reqctx
8       struct s5p_aes_reqctx
8       struct sahara_aes_reqctx
8       struct stm32_cryp_reqctx
8       struct stm32_cryp_reqctx
16      struct ablk_ctx
24      struct atmel_aes_reqctx
48      struct omap_aes_reqctx
48      struct omap_aes_reqctx
48      struct qat_crypto_request
56      struct artpec6_crypto_request_context
64      struct chcr_blkcipher_req_ctx
80      struct spacc_req
80      struct virtio_crypto_sym_request
136     struct qce_cipher_reqctx
168     struct n2_request_context
328     struct ccp_des3_req_ctx
400     struct ccp_aes_req_ctx
536     struct hifn_request_context
992     struct cvm_req_ctx
2456    struct iproc_reqctx_s

The base ablkcipher wrapper is:
80      struct ablkcipher_request

And in my earlier skcipher wrapper analysis, lrw was the largest
skcipher wrapper:
384     struct rctx

iproc_reqctx_s is an extreme outlier, with cvm_req_ctx at a bit less than half.

Making this a 2920 byte fixed array doesn't seem sensible at all
(though that's what's already possible to use with existing
SKCIPHER_REQUEST_ON_STACK users).

What's the right path forward here?

The skcipher implementations based on crypto IP blocks are typically
asynchronous, and I wouldn't be surprised if a fair number of
SKCIPHER_REQUEST_ON_STACK() users are limited to synchronous
skciphers.

According to Herbert, SKCIPHER_REQUEST_ON_STACK() may only be used
for invoking synchronous ciphers.

In fact, due to the way the crypto API is built, if you try using it
with any transformation that uses DMA
you would most probably end up trying to DMA to/from the stack which
as we all know is not a great idea.

Ah yes, I found [0] which contains that quote.

So that means that Kees can disregard the occurrences that are async
only, but it still implies that we cannot limit the reqsize like he
proposes unless we take the sync/async nature into account.
It also means we should probably BUG() or WARN() in
SKCIPHER_REQUEST_ON_STACK() when used with an async algo.

Something like this should do the trick:

diff --git a/include/crypto/skcipher.h b/include/crypto/skcipher.h
index 2f327f090c3e..70584e0f26bc 100644
--- a/include/crypto/skcipher.h
+++ b/include/crypto/skcipher.h

@@ -142,7 +142,9 @@ struct skcipher_alg {
 #define SKCIPHER_REQUEST_ON_STACK(name, tfm) \
        char __##name##_desc[sizeof(struct skcipher_request) + \
                crypto_skcipher_reqsize(tfm)] CRYPTO_MINALIGN_ATTR; \
-       struct skcipher_request *name = (void *)__##name##_desc
+       struct skcipher_request *name = WARN_ON( \
+               crypto_skcipher_alg(tfm)->base.cra_flags & CRYPTO_ALG_ASYNC) \
+               ? NULL : (void *)__##name##_desc

 /**
  * DOC: Symmetric Key Cipher API

That way, we will almost certainly oops on a NULL pointer dereference
right after, but we at least the stack corruption.

quoted

quoted

So we could formalize this and limit SKCIPHER_REQUEST_ON_STACK() to
synchronous skciphers, which implies that the reqsize limit only has
to apply synchronous skciphers as well. But before we can do this, we
have to identify the remaining occurrences that allow asynchronous
skciphers to be used, and replace them with heap allocations.

Any such occurrences are almost for sure broken already due to the DMA
issue I've mentioned.

I am not convinced of this. The skcipher request struct does not
contain any payload buffers, and whether the algo specific ctx struct
is used for DMA is completely up to the driver. So I am quite sure
there are plenty of async algos that work fine with
SKCIPHER_REQUEST_ON_STACK() and vmapped stacks.

quoted

Gilad

--
Gilad Ben-Yossef
Chief Coffee Drinker

values of ? will give rise to dom!

[0] https://www.redhat.com/archives/dm-devel/2018-January/msg00087.html