On 07/12/19 00:57, Sean Christopherson wrote: > The original purpose of this series was to call thp_adjust() from > __direct_map() and FNAME(fetch) to eliminate a page refcounting quirk[*]. > Before doing that, I wanted to clean up the large page handling so that > the map/fetch funtions weren't being passed multiple booleans that tracked > the same basic info. While trying to decipher all the the interactions, > I stumbled across a handful of fun things: > > - 32-bit KVM w/ TDP is completely broken with respect to 64-bit GPAs due > to the page fault handlers and all related flows dropping bits 63:32 > of the GPA. As a result, KVM inserts the wrong GPA and the guest hangs > because it generates EPT/NPT faults until it's killed. > > - The TDP and non-paging page fault flows are identical except for > one-off constraints on guest page size. > > - The !VALID_PAGE(root_hpa) checks in the page fault flows are bogus. > They were added a few years ago to "fix" a nVMX bug and are no longer > needed now that nVMX is in much better shape. > > Patch 1 fixes the 32-bit KVM w/ TDP issue. More details below. > > Patches 2-12 are 99% refactoring to merge TDP and non-paging page fault > handling, and to do the thp_adjust() move. These are basically nops from > a functional perspective. There are technically functional changes in a > few patches, but they are very superficial and in theory won't be > observable in normal usage. > > Patches 13-16 add WARNs on the !VALID_PAGE(root_hpa) checks to make it > clear that root_hpa is expected to be valid when handling page faults, > e.g. for the longest time I thought KVM relied on the checks in map/fetch > to correctly handle kvm_mmu_zap_all(). > > > 32-bit KVM w/ TDP: > > I marked this patch for stable because it's obviously a bug fix, but I'm > entirely not sure we want to backport the fix. Obviously no userspace VMM > is actually exposing 64-bit GPAs to its guests, i.e. odds are this won't > actually fix any real world use cases. And, the scope of the changes are > likely going to make backporting a pain. But, on the other hand, if it's > not backported then future bug fixes in related code are likely to > conflict, and it does fix the case where a buggy guest kernel accesses a > non-existent 64-bit GPA (crashes instead of hanging indefinitely). > > I'm also not confident I found all the cases where KVM is truncating the > GPA. AFAIK, 32-bit Qemu simply doesn't support 64-bit GPAs. To confirm > the bug and verify the fix, I hacked KVM and the guest kernel to generate > 64-bit GPAs when remapping MMIO, which covers a tiny fragment of KVM. Queued, thanks! Paolo > diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c > index fa46fbed60013..49a59bcb32117 100644 > --- a/arch/x86/kvm/x86.c > +++ b/arch/x86/kvm/x86.c > @@ -5737,7 +5737,7 @@ static int emulator_read_write(struct x86_emulate_ctxt *ctxt, > vcpu->run->mmio.len = min(8u, vcpu->mmio_fragments[0].len); > vcpu->run->mmio.is_write = vcpu->mmio_is_write = ops->write; > vcpu->run->exit_reason = KVM_EXIT_MMIO; > - vcpu->run->mmio.phys_addr = gpa; > + vcpu->run->mmio.phys_addr = gpa & 0xffffffffull; > > return ops->read_write_exit_mmio(vcpu, gpa, val, bytes); > } > > > diff --git a/arch/x86/mm/ioremap.c b/arch/x86/mm/ioremap.c > index b9c78f3bcd673..e22f254987bea 100644 > --- a/arch/x86/mm/ioremap.c > +++ b/arch/x86/mm/ioremap.c > @@ -184,7 +184,10 @@ static void __iomem *__ioremap_caller(resource_size_t phys_addr, > if (kernel_map_sync_memtype(phys_addr, size, pcm)) > goto err_free_area; > > - if (ioremap_page_range(vaddr, vaddr + size, phys_addr, prot)) > + BUG_ON(!boot_cpu_data.x86_phys_bits); > + > + if (ioremap_page_range(vaddr, vaddr + size, > + phys_addr | BIT_ULL(boot_cpu_data.x86_phys_bits - 1), prot)) > goto err_free_area; > > ret_addr = (void __iomem *) (vaddr + offset); > > [*] https://lkml.kernel.org/r/20191126174603.GB22233@xxxxxxxxxxxxxxx > > Sean Christopherson (16): > KVM: x86: Use gpa_t for cr2/gpa to fix TDP support on 32-bit KVM > KVM: x86/mmu: Move definition of make_mmu_pages_available() up > KVM: x86/mmu: Fold nonpaging_map() into nonpaging_page_fault() > KVM: x86/mmu: Move nonpaging_page_fault() below try_async_pf() > KVM: x86/mmu: Refactor handling of cache consistency with TDP > KVM: x86/mmu: Refactor the per-slot level calculation in > mapping_level() > KVM: x86/mmu: Refactor handling of forced 4k pages in page faults > KVM: x86/mmu: Incorporate guest's page level into max level for shadow > MMU > KVM: x86/mmu: Persist gfn_lpage_is_disallowed() to max_level > KVM: x86/mmu: Rename lpage_disallowed to account_disallowed_nx_lpage > KVM: x86/mmu: Consolidate tdp_page_fault() and nonpaging_page_fault() > KVM: x86/mmu: Move transparent_hugepage_adjust() above __direct_map() > KVM: x86/mmu: Move calls to thp_adjust() down a level > KVM: x86/mmu: Move root_hpa validity checks to top of page fault > handler > KVM: x86/mmu: WARN on an invalid root_hpa > KVM: x86/mmu: WARN if root_hpa is invalid when handling a page fault > > arch/x86/include/asm/kvm_host.h | 8 +- > arch/x86/kvm/mmu/mmu.c | 438 ++++++++++++++------------------ > arch/x86/kvm/mmu/paging_tmpl.h | 58 +++-- > arch/x86/kvm/mmutrace.h | 12 +- > arch/x86/kvm/x86.c | 40 ++- > arch/x86/kvm/x86.h | 2 +- > include/linux/kvm_host.h | 6 +- > virt/kvm/async_pf.c | 10 +- > 8 files changed, 259 insertions(+), 315 deletions(-) >
