On 2/20/20 10:13 AM, Sean Christopherson wrote: > On Tue, Feb 18, 2020 at 07:26:31PM -0800, Jordan Hand wrote: >> During mprotect (in mm/mprotect.c line 525) the following checks if >> READ_IMPLIES_EXECUTE and a PROT_READ is being requested. If so and >> VM_MAYEXEC is set, it also adds PROT_EXEC to the request. >> >> if (rier && (vma->vm_flags & VM_MAYEXEC)) >> prot |= PROT_EXEC; >> >> But if we look at sgx_encl_page_alloc(), we see vm_max_prot_bits is set >> without taking VM_MAYEXEC into account: >> >> encl_page->vm_max_prot_bits = calc_vm_prot_bits(prot, 0); >> >> sgx_encl_may_map() checks that the requested protection can be added with: >> >> if (!page || (~page->vm_max_prot_bits & vm_prot_bits)) >> return -EACCESS >> >> This means that for any process where READ_IMPLIES_EXECUTE is set and >> page where (vma->vm_flags & VM_MAYEXEC) == true, mmap/mprotect calls to >> that request PROT_READ on a page that was not added with PROT_EXEC will >> fail. > > I could've sworn this was discussed on the SGX list at one point, but > apparently we only discussed it internally. Anyways... > > More than likely, the READ_IMPLIES_EXECUTE (RIE) crud rears its head > because part of the enclave loader is written in assembly. Unless > explicitly told otherwise, the linker assumes that any program with > assembly code may need an executable stack, which leads to the RIE > personality being set for the process. Here's a fantastic write up for > more details: https://www.airs.com/blog/archives/518 > > There are essentially two paths we can take: > > 1) Exempt EPC pages from RIE during mmap()/mprotect(), i.e. don't add > PROT_EXEC for enclaves. > > 2) Punt the issue to userspace. > > Option (1) is desirable in some ways: > > - Enclaves will get an executable stack if and only if the loader/creator > intentionally configures it to have an executable stack. > > - Separates enclaves from the personality of the loader. > > - Userspace doesn't have to do anything for the common case of not > wanting an executable stack for its enclaves. > > The big down side to (1) is that it'd require an ugly hook in architecture > agnostic code. And arguably, it reduces the overall security of the > platform (more below). > > For (2), userspace has a few options: > > a) Tell the linker the enclave loader doesn't need RIE, either via a .note > in assembly files or via the global "-z noexecstack" flag. > > b) Spawn a separate process to run/map the enclave if the enclave loader > needs RIE. > > c) Require enclaves to allow PROT_EXEC on all pages. Note, this is an > absolutely terrible idea and only included for completeness. > > As shown by the lack of a mmap()/mprotect() hook in this series to squash > RIE, we chose option (2). Given that enclave loaders are not legacy code > and hopefully following decent coding practices, option (2a) should suffice > for all loaders. The security benefit mentioned above is that forcing > enclave loaders to squash RIE eliminates an exectuable stack as an attack > vector on the loader. I see your point and I do agree that there are security benefits to (2a) and I think we could do that for our loader. That said, it does concern me that this breaks perfectly valid userspace behavior. If a userspace process decides to use RIE, I don't know that the SGX driver should disobey that decision. So option (3) would be to just honor RIE for enclave pages and when page permissions are set to PROT_READ in sgx_encl_page_alloc and RIE is set, also add PROT_EXEC. I understand your concerns that this using RIE is bad security practice and I'm not convinced that (3) is the way to go, but from a philosophy perspective I don't know that the kernel should be in the business of stopping userspace from doing valid things. If option (3) can't/shouldn't be done for some reason, option (1) at least keeps from breaking expected userspace behavior. But I do agree that (1) is ugly to implement. -Jordan
