On Sat, Oct 24, 2020 at 1:23 AM Borislav Petkov <b...@alien8.de> wrote: > > On Fri, Oct 23, 2020 at 05:12:49PM -0700, Andy Lutomirski wrote: > > I disagree. A real CPU does exactly what I'm describing. If I stick > > A real modern CPU fetches up to 32 bytes insn window which it tries > to decode etc. I don't know, though, what it does when that fetch > encounters a fault - I will have to ask people. I'm not sure it would > even try to feed a shorter stream of bytes to the decoder but lemme > ask... >
I can pretty much guarantee that a real modern CPU is able to decode a <15 byte instruction that is followed by unmapped or non-executable pages. I don't know specifically how the CPU implements it, but it works. > > 0xcc at the end of a page and a make the next page not-present, I get > > #BP, not #PF. But if I stick 0x0F at the end of a page and mark the > > next page not-present, I get #PF. If we're trying to decode an > > instruction in user memory, we can kludge it by trying to fetch 15 > > bytes and handling -EFAULT by fetching fewer bytes, but that's gross > > and doesn't really have the right semantics. What we actually want is > > to fetch up to the page boundary and try to decode it. If it's a > > valid instruction or if it's definitely invalid, we're done. > > Otherwise we fetch across the page boundary. > > We can do that but why would you put all that logic in the insn decoder? > Is that use case sooo important? It's not sooo important, but I think it would be nice to at least try to be fully correct. > > I mean, it would work that way anyway *even* *now* - the insn decoder > will tell you that the insn it decoded wasn't valid and you, as a > caller, know that you didn't fetch the whole 15 bytes so that means > that you still need to fetch some more. You've got all the relevant > information. How so? If I have a page that ends in 0x0F followed by an unmapped page, then the correct response to an attempt to decode is SIGSEGV or -EFAULT. If there's a page there that contains garbage, then the correct response is SIGILL or -EINVAL or similar. These are different scenarios, and I don't think the current decoder API can be used to distinguish them. > > > Eventually we should wrap this whole mess up in an insn_decode_user() > > helper that does the right thing. > > Oh sure, you can do that easily. Just put the logic which determines > that it copied a shorter buffer and that it attempts to decode the > shorter buffer first in it. Yah, that can work. > > > And we can then make that helper > > extra fancy by getting PKRU and EPT-hacker-execute-only right, whereas > > we currently get these cases wrong. > > > > Does this make sense? > > Sure, but you could point me to those cases so that I can get a better > idea what they do exactly. Take a look at fixup_umip_exception(). It currently has two bugs: 1. If it tries to decode a short instruction followed by something like a userfaultfd page, it will incorrectly trigger the userfaultfd. This is because it tries to fetch MAX_INSN_SIZE even if the instruction is shorter than that. 2. It will fail on execute-only memory, and it will succeed on NX memory. copy_from_user() is the wrong API to use here. We don't have the right API, and we should add it. (Hi Dave - what's the best way to do this? New get_user_pages() mode? Try to fault it in, hold an appropriate lock, walk the page tables to check permissions, and then access the user address directly?) I don't know how much anyone really cares about this for UMIP, but with SEV-ES and such, I can see this becoming more important.
