On Mon, Dec 17, 2018 at 2:20 PM Sean Christopherson <sean.j.christopherson@xxxxxxxxx> wrote: > > My brain is still sorting out the details, but I generally like the idea > of allocating an anon inode when creating an enclave, and exposing the > other ioctls() via the returned fd. This is essentially the approach > used by KVM to manage multiple "layers" of ioctls across KVM itself, VMs > and vCPUS. There are even similarities to accessing physical memory via > multiple disparate domains, e.g. host kernel, host userspace and guest. > In my mind, opening /dev/sgx would give you the requisite inode. I'm not 100% sure that the chardev infrastructure allows this, but I think it does. > The only potential hiccup I can see is the build flow. Currently, > EADD+EEXTEND is done via a work queue to avoid major performance issues > (10x regression) when userspace is building multiple enclaves in parallel > using goroutines to wrap Cgo (the issue might apply to any M:N scheduler, > but I've only confirmed the Golang case). The issue is that allocating > an EPC page acts like a blocking syscall when the EPC is under pressure, > i.e. an EPC page isn't immediately available. This causes Go's scheduler > to thrash and tank performance[1]. What's the issue, and how does a workqueue help? I'm wondering if a nicer solution would be an ioctl to add lots of pages in a single call. > > Alternatively, we could change the EADD+EEXTEND flow to not insert the > added page's PFN into the owner's process space, i.e. force userspace to > fault when it runs the enclave. But that only delays the issue because > eventually we'll want to account EPC pages, i.e. add a cgroup, at which > point we'll likely need current->mm anyways. You should be able to account the backing pages to a cgroup without actually sticking them into the EPC, no? Or am I misunderstanding? I guess we'll eventually want a cgroup to limit use of the limited EPC resources.
