Re: [PATCH v3 2/2] crypto: hisilicon/sec2 - fix for aead invalid authsize

["linwenkai (C)" <linwenkai6@xxxxxxxxxxxxx>]

在 2024/11/10 11:36, Herbert Xu 写道:
> On Sat, Nov 02, 2024 at 10:55:59AM +0800, Chenghai Huang wrote:
> > @@ -2226,15 +2236,15 @@ static int sec_aead_spec_check(struct sec_ctx *ctx, struct sec_req *sreq)
> >   	struct device *dev = ctx->dev;
> >   	int ret;
> >   
> > -	if (unlikely(req->cryptlen + req->assoclen > MAX_INPUT_DATA_LEN ||
> > -	    req->assoclen > SEC_MAX_AAD_LEN)) {
> > -		dev_err(dev, "aead input spec error!\n");
> > +	/* Hardware does not handle cases where authsize is less than 4 bytes */
> > +	if (unlikely(sz < MIN_MAC_LEN)) {
> > +		ctx->a_ctx.fallback = true;
> This is broken.  sec_aead_spec_check is a per-request function,
> called without any locking.  Therefore it must not modify any
> field in the tfm context (at least not without additional locking),
> because multiple requests can be issued on the same tfm at any time.
>
> I suppose for this field in particular you could move it to
> set_authsize and there it would be safe to change the tfm context.
>
> Cheers,

Hi,

I have found another setup for fallback in the sec_aead_param_check 
function, so I need to fix it right too.

The orignal code:

static int sec_aead_param_check(struct sec_ctx *ctx, struct sec_req *sreq)
{

        if (ctx->sec->qm.ver == QM_HW_V2) {
                if (unlikely(!req->cryptlen || (!sreq->c_req.encrypt &&
                             req->cryptlen <= authsize))) {
                        ctx->a_ctx.fallback = true;
                        return -EINVAL;
                }
        }
}

After the modification, I used a temporary variable fallback to save the 
state:

static int sec_aead_param_check(struct sec_ctx *ctx, struct sec_req 
*sreq, bool *fallback)
{

        if (ctx->sec->qm.ver == QM_HW_V2) {
                if (unlikely(!req->cryptlen || (!sreq->c_req.encrypt &&
                             req->cryptlen <= authsize))) {
                        *fallback = true;
                        return -EINVAL;
                }
        }
}

Same with  the sec_aead_spec_check function.

static int sec_aead_spec_check(struct sec_ctx *ctx, struct sec_req 
*sreq, bool *fallback) {

        /* Hardware does not handle cases where authsize is less than 4 
bytes */
        if (unlikely(sz < MIN_MAC_LEN)) {
                *fallback = true;
                return -EINVAL;
        }

}

Do you think it is a better way?

Thanks,