KVM: x86: Remove VMX support for virtualizing guest MTRR memtypes

Remove KVM's support for virtualizing guest MTRR memtypes, as full MTRR
adds no value, negatively impacts guest performance, and is a maintenance
burden due to it's complexity and oddities.

KVM's approach to virtualizating MTRRs make no sense, at all.  KVM *only*
honors guest MTRR memtypes if EPT is enabled *and* the guest has a device
that may perform non-coherent DMA access.  From a hardware virtualization
perspective of guest MTRRs, there is _nothing_ special about EPT.  Legacy
shadowing paging doesn't magically account for guest MTRRs, nor does NPT.

Unwinding and deciphering KVM's murky history, the MTRR virtualization
code appears to be the result of misdiagnosed issues when EPT + VT-d with
passthrough devices was enabled years and years ago.  And importantly, the
underlying bugs that were fudged around by honoring guest MTRR memtypes
have since been fixed (though rather poorly in some cases).

The zapping GFNs logic in the MTRR virtualization code came from:

  commit efdfe536d8c643391e19d5726b072f82964bfbdb
  Author: Xiao Guangrong <guangrong.xiao@linux.intel.com>
  Date:   Wed May 13 14:42:27 2015 +0800

    KVM: MMU: fix MTRR update

    Currently, whenever guest MTRR registers are changed
    kvm_mmu_reset_context is called to switch to the new root shadow page
    table, however, it's useless since:
    1) the cache type is not cached into shadow page's attribute so that
       the original root shadow page will be reused

    2) the cache type is set on the last spte, that means we should sync
       the last sptes when MTRR is changed

    This patch fixs this issue by drop all the spte in the gfn range which
    is being updated by MTRR

which was a fix for:

  commit 0bed3b568b68e5835ef5da888a372b9beabf7544
  Author:     Sheng Yang <sheng@linux.intel.com>
  AuthorDate: Thu Oct 9 16:01:54 2008 +0800
  Commit:     Avi Kivity <avi@redhat.com>
  CommitDate: Wed Dec 31 16:51:44 2008 +0200

      KVM: Improve MTRR structure

      As well as reset mmu context when set MTRR.

which was part of a "MTRR/PAT support for EPT" series that also added:

+       if (mt_mask) {
+               mt_mask = get_memory_type(vcpu, gfn) <<
+                         kvm_x86_ops->get_mt_mask_shift();
+               spte |= mt_mask;
+       }

where get_memory_type() was a truly gnarly helper to retrieve the guest
MTRR memtype for a given memtype.  And *very* subtly, at the time of that
change, KVM *always* set VMX_EPT_IGMT_BIT,

        kvm_mmu_set_base_ptes(VMX_EPT_READABLE_MASK |
                VMX_EPT_WRITABLE_MASK |
                VMX_EPT_DEFAULT_MT << VMX_EPT_MT_EPTE_SHIFT |
                VMX_EPT_IGMT_BIT);

which came in via:

  commit 928d4bf747e9c290b690ff515d8f81e8ee226d97
  Author:     Sheng Yang <sheng@linux.intel.com>
  AuthorDate: Thu Nov 6 14:55:45 2008 +0800
  Commit:     Avi Kivity <avi@redhat.com>
  CommitDate: Tue Nov 11 21:00:37 2008 +0200

      KVM: VMX: Set IGMT bit in EPT entry

      There is a potential issue that, when guest using pagetable without vmexit when
      EPT enabled, guest would use PAT/PCD/PWT bits to index PAT msr for it's memory,
      which would be inconsistent with host side and would cause host MCE due to
      inconsistent cache attribute.

      The patch set IGMT bit in EPT entry to ignore guest PAT and use WB as default
      memory type to protect host (notice that all memory mapped by KVM should be WB).

Note the CommitDates!  The AuthorDates strongly suggests Sheng Yang added
the whole "ignoreIGMT things as a bug fix for issues that were detected
during EPT + VT-d + passthrough enabling, but it was applied earlier
because it was a generic fix.

Jumping back to 0bed3b568b68 ("KVM: Improve MTRR structure"), the other
relevant code, or rather lack thereof, is the handling of *host* MMIO.
That fix came in a bit later, but given the author and timing, it's safe
to say it was all part of the same EPT+VT-d enabling mess.

  commit 2aaf69dcee864f4fb6402638dd2f263324ac839f
  Author:     Sheng Yang <sheng@linux.intel.com>
  AuthorDate: Wed Jan 21 16:52:16 2009 +0800
  Commit:     Avi Kivity <avi@redhat.com>
  CommitDate: Sun Feb 15 02:47:37 2009 +0200

    KVM: MMU: Map device MMIO as UC in EPT

    Software are not allow to access device MMIO using cacheable memory type, the
    patch limit MMIO region with UC and WC(guest can select WC using PAT and
    PCD/PWT).

In addition to the host MMIO and IGMT issues, KVM's MTRR virtualization
was obviously never tested on NPT until much later, which lends further
credence to the theory/argument that this was all the result of
misdiagnosed issues.

Discussion from the EPT+MTRR enabling thread[*] more or less confirms that
Sheng Yang was trying to resolve issues with passthrough MMIO.

 * Sheng Yang
  : Do you mean host(qemu) would access this memory and if we set it to guest
  : MTRR, host access would be broken? We would cover this in our shadow MTRR
  : patch, for we encountered this in video ram when doing some experiment with
  : VGA assignment.

And in the same thread, there's also what appears to be confirmation of
Intel running into issues with Windows XP related to a guest device driver
mapping DMA with WC in the PAT.

 * Avi Kavity
  : Sheng Yang wrote:
  : > Yes... But it's easy to do with assigned devices' mmio, but what if guest
  : > specific some non-mmio memory's memory type? E.g. we have met one issue in
  : > Xen, that a assigned-device's XP driver specific one memory region as buffer,
  : > and modify the memory type then do DMA.
  : >
  : > Only map MMIO space can be first step, but I guess we can modify assigned
  : > memory region memory type follow guest's?
  : >
  :
  : With ept/npt, we can't, since the memory type is in the guest's
  : pagetable entries, and these are not accessible.

[*] https://lore.kernel.org/all/1223539317-32379-1-git-send-email-sheng@linux.intel.com

So, for the most part, what likely happened is that 15 years ago, a few
engineers (a) fixed a #MC problem by ignoring guest PAT and (b) initially
"fixed" passthrough device MMIO by emulating *guest* MTRRs.  Except for
the below case, everything since then has been a result of those two
intertwined changes.

The one exception, which is actually yet more confirmation of all of the
above, is the revert of Paolo's attempt at "full" virtualization of guest
MTRRs:

  commit 606decd67049217684e3cb5a54104d51ddd4ef35
  Author: Paolo Bonzini <pbonzini@redhat.com>
  Date:   Thu Oct 1 13:12:47 2015 +0200

    Revert "KVM: x86: apply guest MTRR virtualization on host reserved pages"

    This reverts commit fd717f11015f673487ffc826e59b2bad69d20fe5.
    It was reported to cause Machine Check Exceptions (bug 104091).

...

  commit fd717f11015f673487ffc826e59b2bad69d20fe5
  Author: Paolo Bonzini <pbonzini@redhat.com>
  Date:   Tue Jul 7 14:38:13 2015 +0200

    KVM: x86: apply guest MTRR virtualization on host reserved pages

    Currently guest MTRR is avoided if kvm_is_reserved_pfn returns true.
    However, the guest could prefer a different page type than UC for
    such pages. A good example is that pass-throughed VGA frame buffer is
    not always UC as host expected.

    This patch enables full use of virtual guest MTRRs.

I.e. Paolo tried to add back KVM's behavior before "Map device MMIO as UC
in EPT" and got the same result: machine checks, likely due to the guest
MTRRs not being trustworthy/sane at all times.

Note, Paolo also tried to enable MTRR virtualization on SVM+NPT, but that
too got reverted.  Unfortunately, it doesn't appear that anyone ever found
a smoking gun, i.e. exactly why emulating guest MTRRs via NPT PAT caused
extremely slow boot times doesn't appear to have a definitive root cause.

  commit fc07e76ac7ffa3afd621a1c3858a503386a14281
  Author: Paolo Bonzini <pbonzini@redhat.com>
  Date:   Thu Oct 1 13:20:22 2015 +0200

    Revert "KVM: SVM: use NPT page attributes"

    This reverts commit 3c2e7f7de3240216042b61073803b61b9b3cfb22.
    Initializing the mapping from MTRR to PAT values was reported to
    fail nondeterministically, and it also caused extremely slow boot
    (due to caching getting disabled---bug 103321) with assigned devices.

...

  commit 3c2e7f7de3240216042b61073803b61b9b3cfb22
  Author: Paolo Bonzini <pbonzini@redhat.com>
  Date:   Tue Jul 7 14:32:17 2015 +0200

    KVM: SVM: use NPT page attributes

    Right now, NPT page attributes are not used, and the final page
    attribute depends solely on gPAT (which however is not synced
    correctly), the guest MTRRs and the guest page attributes.

    However, we can do better by mimicking what is done for VMX.
    In the absence of PCI passthrough, the guest PAT can be ignored
    and the page attributes can be just WB.  If passthrough is being
    used, instead, keep respecting the guest PAT, and emulate the guest
    MTRRs through the PAT field of the nested page tables.

    The only snag is that WP memory cannot be emulated correctly,
    because Linux's default PAT setting only includes the other types.

In short, honoring guest MTRRs for VMX was initially a workaround of
sorts for KVM ignoring guest PAT *and* for KVM not forcing UC for host
MMIO.  And while there *are* known cases where honoring guest MTRRs is
desirable, e.g. passthrough VGA frame buffers, the desired behavior in
that case is to get WC instead of UC, i.e. at this point it's for
performance, not correctness.

Furthermore, the complete absence of MTRR virtualization on NPT and
shadow paging proves that, while KVM theoretically can do better, it's
by no means necessary for correctnesss.

Lastly, since kernels mostly rely on firmware to do MTRR setup, and the
host typically provides guest firmware, honoring guest MTRRs is effectively
honoring *host* userspace memtypes, which is also backwards.  I.e. it
would be far better for host userspace to communicate its desired memtype
directly to KVM (or perhaps indirectly via VMAs in the host kernel), not
through guest MTRRs.

Tested-by: Xiangfei Ma <xiangfeix.ma@intel.com>
Tested-by: Yongwei Ma <yongwei.ma@intel.com>
Link: https://lore.kernel.org/r/20240309010929.1403984-2-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>

1 files changed, 8 insertions, 25 deletions

diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
index 662f62dfb2aa..561e7c29cf9e 100644
--- a/arch/x86/kvm/mmu/mmu.c
+++ b/arch/x86/kvm/mmu/mmu.c
@@ -4613,38 +4613,21 @@ out_unlock:
 }
 #endif
 
-bool __kvm_mmu_honors_guest_mtrrs(bool vm_has_noncoherent_dma)
+bool kvm_mmu_may_ignore_guest_pat(void)
 {
 	/*
-	 * If host MTRRs are ignored (shadow_memtype_mask is non-zero), and the
-	 * VM has non-coherent DMA (DMA doesn't snoop CPU caches), KVM's ABI is
-	 * to honor the memtype from the guest's MTRRs so that guest accesses
-	 * to memory that is DMA'd aren't cached against the guest's wishes.
-	 *
-	 * Note, KVM may still ultimately ignore guest MTRRs for certain PFNs,
-	 * e.g. KVM will force UC memtype for host MMIO.
+	 * When EPT is enabled (shadow_memtype_mask is non-zero), and the VM
+	 * has non-coherent DMA (DMA doesn't snoop CPU caches), KVM's ABI is to
+	 * honor the memtype from the guest's PAT so that guest accesses to
+	 * memory that is DMA'd aren't cached against the guest's wishes.  As a
+	 * result, KVM _may_ ignore guest PAT, whereas without non-coherent DMA,
+	 * KVM _always_ ignores guest PAT (when EPT is enabled).
 	 */
-	return vm_has_noncoherent_dma && shadow_memtype_mask;
+	return shadow_memtype_mask;
 }
 
 int kvm_tdp_page_fault(struct kvm_vcpu *vcpu, struct kvm_page_fault *fault)
 {
-	/*
-	 * If the guest's MTRRs may be used to compute the "real" memtype,
-	 * restrict the mapping level to ensure KVM uses a consistent memtype
-	 * across the entire mapping.
-	 */
-	if (kvm_mmu_honors_guest_mtrrs(vcpu->kvm)) {
-		for ( ; fault->max_level > PG_LEVEL_4K; --fault->max_level) {
-			int page_num = KVM_PAGES_PER_HPAGE(fault->max_level);
-			gfn_t base = gfn_round_for_level(fault->gfn,
-							 fault->max_level);
-
-			if (kvm_mtrr_check_gfn_range_consistency(vcpu, base, page_num))
-				break;
-		}
-	}
-
 #ifdef CONFIG_X86_64
 	if (tdp_mmu_enabled)
 		return kvm_tdp_mmu_page_fault(vcpu, fault);