On Tue, 2022-09-27 at 07:41 +0000, Patel, Himanshu wrote: > ++linux-sgx@xxxxxxxxxxxxxxx > > Please explain me the process for my below questions. > > Thanks > Himanshu > > From: Patel, Himanshu > Sent: Tuesday, September 27, 2022 10:48 AM > To: Hansen, Dave <dave.hansen@xxxxxxxxx>; Huang, Kai <kai.huang@xxxxxxxxx> > Cc: Hossain, Md Iqbal <md.iqbal.hossain@xxxxxxxxx>; Kuvaiskii, Dmitrii > <Dmitrii.Kuvaiskii@xxxxxxxxx> > Subject: SGX Information needed > > Hi Dave/Kai, > > I have some questions related to SGX. I am putting a scenario here. > > suppose we have 16G PRM (per socket) and sgx_enclave_size=64G , so workload > will need remote socket's enclave memory as well right. > which entity (SGX driver?) is doing the job to decide remote sockets to > fulfill workload's requirement? > > If SGX driver is deciding , it must be looking somewhere(like any mapped table > or anything else) to select remote sockets, what it that? > Hi Himanshu, Jarkko should be the person to answer, but I'll try first: Yes the SGX driver decides which node(s) to allocate. Firstly the drive tries to allocate from the local node. If local node is running out of EPC page, the driver tries to allocate EPC page from the 'next-highest node' (i.e. node 0 -> 1 -> 2 ...). EPC doesn't have its own SRAT type, so it depends on SRAT table's memory resource (which covers EPC in practice, if I recall correctly) affinity to decide which EPC page belongs to which node. Hope above helps. -- Thanks, -Kai
