Re: [RFC 00/10] Process-local memory allocations for hiding KVM secrets

[Andy Lutomirski]

On Wed, Jun 12, 2019 at 6:50 PM Nadav Amit <namit@xxxxxxxxxx> wrote:
>
> > On Jun 12, 2019, at 6:30 PM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
> >
> > On Wed, Jun 12, 2019 at 1:27 PM Andy Lutomirski <luto@xxxxxxxxxxxxxx> wrote:
> >>> On Jun 12, 2019, at 12:55 PM, Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
> >>>
> >>>> On 6/12/19 10:08 AM, Marius Hillenbrand wrote:
> >>>> This patch series proposes to introduce a region for what we call
> >>>> process-local memory into the kernel's virtual address space.
> >>>
> >>> It might be fun to cc some x86 folks on this series.  They might have
> >>> some relevant opinions. ;)
> >>>
> >>> A few high-level questions:
> >>>
> >>> Why go to all this trouble to hide guest state like registers if all the
> >>> guest data itself is still mapped?
> >>>
> >>> Where's the context-switching code?  Did I just miss it?
> >>>
> >>> We've discussed having per-cpu page tables where a given PGD is only in
> >>> use from one CPU at a time.  I *think* this scheme still works in such a
> >>> case, it just adds one more PGD entry that would have to context-switched.
> >>
> >> Fair warning: Linus is on record as absolutely hating this idea. He might change his mind, but itâs an uphill battle.
> >
> > I looked at the patch, and it (sensibly) has nothing to do with
> > per-cpu PGDs.  So it's in great shape!
> >
> > Seriously, though, here are some very high-level review comments:
> >
> > Please don't call it "process local", since "process" is meaningless.
> > Call it "mm local" or something like that.
> >
> > We already have a per-mm kernel mapping: the LDT.  So please nix all
> > the code that adds a new VA region, etc, except to the extent that
> > some of it consists of valid cleanups in and of itself.  Instead,
> > please refactor the LDT code (arch/x86/kernel/ldt.c, mainly) to make
> > it use a more general "mm local" address range, and then reuse the
> > same infrastructure for other fancy things.  The code that makes it
> > KASLR-able should be in its very own patch that applies *after* the
> > code that makes it all work so that, when the KASLR part causes a
> > crash, we can bisect it.
> >
> > + /*
> > + * Faults in process-local memory may be caused by process-local
> > + * addresses leaking into other contexts.
> > + * tbd: warn and handle gracefully.
> > + */
> > + if (unlikely(fault_in_process_local(address))) {
> > + pr_err("page fault in PROCLOCAL at %lx", address);
> > + force_sig_fault(SIGSEGV, SEGV_MAPERR, (void __user *)address, current);
> > + }
> > +
> >
> > Huh?  Either it's an OOPS or you shouldn't print any special
> > debugging.  As it is, you're just blatantly leaking the address of the
> > mm-local range to malicious user programs.
> >
> > Also, you should IMO consider using this mechanism for kmap_atomic().
> > Hi, Nadav!
>
> Well, some context for the âhiâ would have been helpful. (Do I have a bug
> and I still donât understand it?)

Fair enough :)

>
> Perhaps you regard some use-case for a similar mechanism that I mentioned
> before. I did implement something similar (but not the way that you wanted)
> to improve the performance of seccomp and system-calls when retpolines are
> used. I set per-mm code area that held code that used direct calls to invoke
> seccomp filters and frequently used system-calls.
>
> My mechanism, I think, is more not suitable for this use-case. I needed my
> code-page to be at the same 2GB range as the kernel text/modules, which does
> complicate things. Due to the same reason, it is also limited in the size of
> the data/code that it can hold.
>

I actually meant the opposite.  If we had a general-purpose per-mm
kernel address range, could it be used to optimize kmap_atomic() by
limiting the scope of any shootdowns?  As a rough sketch, we'd have
some kmap_atomic slots for each cpu *in the mm-local region*.  I'm not
entirely sure this is a win.

--Andy