On Thu, May 02, 2019 at 03:12:55PM +0000, Kalyani Akula wrote: > Hi Corentin, > > Please find my response inline. > > > -----Original Message----- > > From: Corentin Labbe <clabbe.montjoie@xxxxxxxxx> > > Sent: Thursday, May 2, 2019 5:30 PM > > To: Kalyani Akula <kalyania@xxxxxxxxxx> > > Cc: herbert@xxxxxxxxxxxxxxxxxxx; kstewart@xxxxxxxxxxxxxxxxxxx; > > gregkh@xxxxxxxxxxxxxxxxxxx; tglx@xxxxxxxxxxxxx; pombredanne@xxxxxxxx; > > linux-crypto@xxxxxxxxxxxxxxx; linux-kernel@xxxxxxxxxxxxxxx; > > netdev@xxxxxxxxxxxxxxx; Sarat Chand Savitala <saratcha@xxxxxxxxxx>; Kalyani > > Akula <kalyania@xxxxxxxxxx> > > Subject: Re: [RFC PATCH V3 0/4] Add Xilinx's ZynqMP SHA3 driver support > > > > On Thu, May 02, 2019 at 04:04:38PM +0530, Kalyani Akula wrote: > > > This patch set adds support for > > > - dt-binding docs for Xilinx ZynqMP SHA3 driver > > > - Adds communication layer support for sha_hash in zynqmp.c > > > - Adds Xilinx ZynqMP driver for SHA3 Algorithm > > > - Adds device tree node for ZynqMP SHA3 driver > > > > > > V3 Changes : > > > - Removed zynqmp_sha_import and export APIs.The reason as follows The > > > user space code does an accept on an already accepted FD when we > > > create AF_ALG socket and call accept on it, it calls af_alg_accept and > > > not hash_accept. > > > import and export APIs are called from hash_accept. > > > The flow is as below > > > accept--> af_alg_accept-->hash_accept_parent-->hash_accept_parent_noke > > > accept--> y > > > for hash salg_type. > > > - Resolved comments from > > > https://patchwork.kernel.org/patch/10753719/ > > > > > > > > > Your driver still doesnt handle the case where two hash are done in parallel. > > > > Our Firmware uses IPI protocol to send this SHA3 requests to SHA3 HW engine, which doesn't support parallel processing of 2 hash requests. > The flow is > SHA3 request from App -> SHA3 driver-> ZynqMp driver-> Firmware (which doesn't support parallel processing of 2 requests) -> SHA3 HW Engine > > So your driver will just send bad result in that case. You need to export and store the intermediate result in a request context. > > Furthermore, you miss the export/import functions. > > > > When user space code does an accept on an already accepted FD as below > sockfd = socket(AF_ALG, SOCK_SEQPACKET, 0); > bind(sockfd, (struct sockaddr *)&sa, sizeof(sa)); > fd = accept(sockfd, NULL, 0); > > where my sockaddr is > struct sockaddr_alg sa = { > .salg_family = AF_ALG, > .salg_type = "hash", > .salg_name = "xilinx-sha3-384" > }; > > Upon calling accept the flow in the kernel is as mentioned > accept--> af_alg_accept-->hash_accept_parent-->hash_accept_parent_nokey > for hash salg_type. > > And where import and export functions are called from hash_accept. hence, these functions never be called from the application. > So, I removed those from the driver. > > Regards > Kalyani. > Handling your own worflow is not enough. You need to support two client doing multiple update in parallel. It seems that your driver is bugged in that case. Furthermore, i am pretty sure that export and import are mandatory, and without them self-test should fail. Do you have self test enabled and tryed to load the tcrypt module ? Regards
