Re: [PATCH v2 1/2] crypto: engine - support for parallel requests

[Corentin Labbe]

On Tue, Feb 04, 2020 at 02:34:19PM +0200, Iuliana Prodan wrote:
> Added support for executing multiple requests, in parallel,
> for crypto engine.
> A new callback is added, can_enqueue_more, which asks the
> driver if the hardware has free space, to enqueue a new request.
> The new crypto_engine_alloc_init_and_set function, initialize
> crypto-engine, sets the maximum size for crypto-engine software
> queue (not hardcoded anymore) and the can_enqueue_more callback.
> On crypto_pump_requests, if can_enqueue_more callback returns true,
> a new request is send to hardware, until there is no space and the
> callback returns false.
> 
> Signed-off-by: Iuliana Prodan <iuliana.prodan@xxxxxxx>
> ---
>  crypto/crypto_engine.c  | 106 ++++++++++++++++++++++++++++++------------------
>  include/crypto/engine.h |  10 +++--
>  2 files changed, 72 insertions(+), 44 deletions(-)
> 
> diff --git a/crypto/crypto_engine.c b/crypto/crypto_engine.c
> index eb029ff..aba934f 100644
> --- a/crypto/crypto_engine.c
> +++ b/crypto/crypto_engine.c
> @@ -22,32 +22,18 @@
>   * @err: error number
>   */
>  static void crypto_finalize_request(struct crypto_engine *engine,
> -			     struct crypto_async_request *req, int err)
> +				    struct crypto_async_request *req, int err)
>  {
> -	unsigned long flags;
> -	bool finalize_cur_req = false;
>  	int ret;
>  	struct crypto_engine_ctx *enginectx;
>  
> -	spin_lock_irqsave(&engine->queue_lock, flags);
> -	if (engine->cur_req == req)
> -		finalize_cur_req = true;
> -	spin_unlock_irqrestore(&engine->queue_lock, flags);
> -
> -	if (finalize_cur_req) {
> -		enginectx = crypto_tfm_ctx(req->tfm);
> -		if (engine->cur_req_prepared &&
> -		    enginectx->op.unprepare_request) {
> -			ret = enginectx->op.unprepare_request(engine, req);
> -			if (ret)
> -				dev_err(engine->dev, "failed to unprepare request\n");
> -		}
> -		spin_lock_irqsave(&engine->queue_lock, flags);
> -		engine->cur_req = NULL;
> -		engine->cur_req_prepared = false;
> -		spin_unlock_irqrestore(&engine->queue_lock, flags);
> +	enginectx = crypto_tfm_ctx(req->tfm);
> +	if (enginectx->op.prepare_request &&
> +	    enginectx->op.unprepare_request) {
> +		ret = enginectx->op.unprepare_request(engine, req);
> +		if (ret)
> +			dev_err(engine->dev, "failed to unprepare request\n");
>  	}
> -
>  	req->complete(req, err);
>  
>  	kthread_queue_work(engine->kworker, &engine->pump_requests);
> @@ -73,10 +59,6 @@ static void crypto_pump_requests(struct crypto_engine *engine,
>  
>  	spin_lock_irqsave(&engine->queue_lock, flags);
>  
> -	/* Make sure we are not already running a request */
> -	if (engine->cur_req)
> -		goto out;
> -

Hello

Your patch has the same problem than mine reported by Horia.
If a queue has more than one request, a first crypto_pump_requests() will send a request and for drivers which do not block on do_one_request() crypto_pump_requests() will end.
Then another crypto_pump_requests() will fire sending a second request while the driver does not support that.

So we need to replace engine->cur_req by another locking mechanism.
Perhaps the cleaner is to add a "request count" (increased when do_one_request, decreased in crypto_finalize_request)
I know that the early version have that and it was removed, but I do not see any better way.

Regards