KVM: Disable large pages on misaligned memory slots
Avi Kivity [Mon, 8 Jun 2009 12:52:39 +0000 (15:52 +0300)]
If a slots guest physical address and host virtual address unequal (mod
large page size), then we would erronously try to back guest large pages
with host large pages.  Detect this misalignment and diable large page
support for the trouble slot.

Cc: stable@kernel.org
Signed-off-by: Avi Kivity <avi@redhat.com>

virt/kvm/kvm_main.c

index 5fed9bf..5f865ed 100644 (file)
@@ -1086,7 +1086,7 @@ int __kvm_set_memory_region(struct kvm *kvm,
 {
        int r;
        gfn_t base_gfn;
-       unsigned long npages;
+       unsigned long npages, ugfn;
        int largepages;
        unsigned long i;
        struct kvm_memory_slot *memslot;
@@ -1177,6 +1177,14 @@ int __kvm_set_memory_region(struct kvm *kvm,
                        new.lpage_info[0].write_count = 1;
                if ((base_gfn+npages) % KVM_PAGES_PER_HPAGE)
                        new.lpage_info[largepages-1].write_count = 1;
+               ugfn = new.userspace_addr >> PAGE_SHIFT;
+               /*
+                * If the gfn and userspace address are not aligned wrt each
+                * other, disable large page support for this slot
+                */
+               if ((base_gfn ^ ugfn) & (KVM_PAGES_PER_HPAGE - 1))
+                       for (i = 0; i < largepages; ++i)
+                               new.lpage_info[i].write_count = 1;
        }
 
        /* Allocate page dirty bitmap if needed */