On Tue, Sep 04, 2018 at 11:17:35AM -0700, Sean Christopherson wrote: > On Tue, Sep 04, 2018 at 09:01:15PM +0300, Andy Shevchenko wrote: > > On Tue, Sep 4, 2018 a> +/** > > > > > > > + va = ioremap_cache(addr, size); > > > > > + if (!va) > > > > > + return -ENOMEM; > > > > > > > > I'm not sure this is a right API. Do we operate with memory? Does it > > > > have I/O side effects? > > > > If no, memremap() would be better to use. > > > > > > Preserving __iomem is desirable. There aren't side effects per se, > > > but direct non-enclave accesses to the EPC get abort page semantics so > > > the kernel shouldn't be directly dereferencing a pointer to the EPC. > > > Though by that argument, sgx_epc_bank.va, sgx_epc_addr's return and > > > all ENCLS helpers should be tagged __iomem. > > > > Why? > > Does it related to *any* I/O? > > No, hence my other comment that __private or a new tag altogether may > be more appropriate. The noderef attribute is what we truly care > about. My proposal is that we go with memremap() and use #define __sgx_epc __attribute__((noderef)) It makes sense to check that direct EPC pointers are not passed to functions when they are not supposed to. /Jarkko
