Le 24/07/2014 08:00, Herbert Xu a écrit :
> On Sat, Jul 12, 2014 at 02:59:13PM +0200, LABBE Corentin wrote:
>>
>> +/* sunxi_hash_init: initialize request context
>> + * Activate the SS, and configure it for MD5 or SHA1
>> + */
>> +int sunxi_hash_init(struct ahash_request *areq)
>> +{
>> + const char *hash_type;
>> + struct crypto_ahash *tfm = crypto_ahash_reqtfm(areq);
>> + struct sunxi_req_ctx *op = crypto_ahash_ctx(tfm);
>> +
>> + mutex_lock(&ss->lock);
>> +
>> + hash_type = crypto_tfm_alg_name(areq->base.tfm);
>> +
>> + op->byte_count = 0;
>> + op->nbwait = 0;
>> + op->waitbuf = 0;
>> +
>> + /* Enable and configure SS for MD5 or SHA1 */
>> + if (strcmp(hash_type, "sha1") == 0)
>> + op->mode = SS_OP_SHA1;
>> + else
>> + op->mode = SS_OP_MD5;
>> +
>> + writel(op->mode | SS_ENABLED, ss->base + SS_CTL);
>> + return 0;
>
> The hash driver is completely broken. You are modifying tfm
> ctx data which is shared by all users of a single tfm. So
> if two users conduct hashes in parallel they will step all
> over each other.
So where can I store data for each request ?
>
> Worse, the unpaired mutex_lock will quickly lead to dead locks.
>
> You cannot assume that final will be called.
An user reported an equivalent problem when using openssl speed test with cryptodev.
Does cryptoqueue is a good answer to that problem since the device could handle only one transformation at a time ?
And perhaps with cryptoqueue, my first question is useless.
--
To unsubscribe from this list: send the line "unsubscribe linux-crypto" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
