On Tue, Jan 14, 2020 at 10:40:53AM +0000, Iuliana Prodan wrote: > On 1/14/2020 2:14 AM, Herbert Xu wrote: > > On Mon, Jan 13, 2020 at 09:48:11AM +0000, Iuliana Prodan wrote: > >> > >> Regarding the transfer request to crypto-engine: if sending all requests > >> to crypto-engine, multibuffer tests, for non-backlogging requests fail > >> after only 10 requests, since crypto-engine queue has 10 entries. > > > > That isn't right. The crypto engine should never refuse to accept > > a request > Crypto-engine accepts all request that have the backlog flag, the > non-backlog are accepted till the configured limit (of 10). > > > unless the hardware queue is really full. > Crypto-engine doesn't check the status of hardware queue. > The non-backlog requests are dropped after 10 entries. > > > Perhaps the > > crypto engine code needs to be fixed? > To me, crypto-engine seems to be made for backlogged request, that's why > I'm sending the non-backlog directly to CAAM. The implicit serialization > of request in crypto-engine is the bottleneck. > > But, as I said before, I want to update crypto-engine to set queue > length when initialize crypto-engine, and remove serialization of > requests in crypto-engine by adding knowledge about the underlying hw > accelerator (number of request that can be processed in parallel). > I'll send a RFC with my proposal for crypto-engine enhancements. > > But, until then, I would like to have a backlogging solution in CAAM driver. > Hello I have already something for queue length and processing in parallel. I will send it soon. Regards