On Fri, Jul 17, 2020 at 12:20:39AM -0700, ira.weiny@xxxxxxxxx wrote:
> From: Ira Weiny <ira.weiny@xxxxxxxxx>
>
> This RFC series has been reviewed by Dave Hansen.
>
> Changes from RFC:
> Clean up commit messages based on Peter Zijlstra's and Dave Hansen's
> feedback
> Fix static branch anti-pattern
> New patch:
> (memremap: Convert devmap static branch to {inc,dec})
> This was the code I used as a model for my static branch which
> I believe is wrong now.
> New Patch:
> (x86/entry: Preserve PKRS MSR through exceptions)
> This attempts to preserve the per-logical-processor MSR, and
> reference counting during exceptions. I'd really like feed
> back on this because I _think_ it should work but I'm afraid
> I'm missing something as my testing has shown a lot of spotty
> crashes which don't make sense to me.
>
> This patch set introduces a new page protection mechanism for supervisor pages,
> Protection Key Supervisor (PKS) and an initial user of them, persistent memory,
> PMEM.
>
> PKS enables protections on 'domains' of supervisor pages to limit supervisor
> mode access to those pages beyond the normal paging protections. They work in
> a similar fashion to user space pkeys. Like User page pkeys (PKU), supervisor
> pkeys are checked in addition to normal paging protections and Access or Writes
> can be disabled via a MSR update without TLB flushes when permissions change.
> A page mapping is assigned to a domain by setting a pkey in the page table
> entry.
>
> Unlike User pkeys no new instructions are added; rather WRMSR/RDMSR are used to
> update the PKRS register.
>
> XSAVE is not supported for the PKRS MSR. To reduce software complexity the
> implementation saves/restores the MSR across context switches but not during
> irqs. This is a compromise which results is a hardening of unwanted access
> without absolute restriction.
>
> For consistent behavior with current paging protections, pkey 0 is reserved and
> configured to allow full access via the pkey mechanism, thus preserving the
> default paging protections on mappings with the default pkey value of 0.
>
> Other keys, (1-15) are allocated by an allocator which prepares us for key
> contention from day one. Kernel users should be prepared for the allocator to
> fail either because of key exhaustion or due to PKS not being supported on the
> arch and/or CPU instance.
>
> Protecting against stray writes is particularly important for PMEM because,
> unlike writes to anonymous memory, writes to PMEM persists across a reboot.
> Thus data corruption could result in permanent loss of data.
>
> The following attributes of PKS makes it perfect as a mechanism to protect PMEM
> from stray access within the kernel:
>
> 1) Fast switching of permissions
> 2) Prevents access without page table manipulations
> 3) Works on a per thread basis
> 4) No TLB flushes required
Cool! This seems like it'd be very handy to make other types of kernel
data "read-only at rest" (as was long ago proposed via X86_CR0_WP[1],
which only provided to protection levels, not 15). For example, I think
at least a few other kinds of areas stand out to me that are in need
of PKS markings (i.e. only things that actually manipulate these areas
should gain temporary PK access):
- Page Tables themselves
- Identity mapping
- The "read-only at rest" stuff, though it'll need special plumbing to
make it work with the slab allocator, etc (more like the later "static
allocation" work[2]).
[1] https://lore.kernel.org/lkml/1490811363-93944-1-git-send-email-keescook@xxxxxxxxxxxx/
[2] https://lore.kernel.org/lkml/cover.1550097697.git.igor.stoppa@xxxxxxxxxx/
--
Kees Cook
