kvm: optimize away THP checks in kvm_is_mmio_pfn()
authorAndrea Arcangeli <aarcange@redhat.com>
Thu, 25 Jul 2013 01:04:38 +0000 (03:04 +0200)
committerGleb Natapov <gleb@redhat.com>
Tue, 27 Aug 2013 08:01:10 +0000 (11:01 +0300)
The checks on PG_reserved in the page structure on head and tail pages
aren't necessary because split_huge_page wouldn't transfer the
PG_reserved bit from head to tail anyway.

This was a forward-thinking check done in the case PageReserved was
set by a driver-owned page mapped in userland with something like
remap_pfn_range in a VM_PFNMAP region, but using hugepmds (not
possible right now). It was meant to be very safe, but it's overkill
as it's unlikely split_huge_page could ever run without the driver
noticing and tearing down the hugepage itself.

And if a driver in the future will really want to map a reserved
hugepage in userland using an huge pmd it should simply take care of
marking all subpages reserved too to keep KVM safe. This of course
would require such a hypothetical driver to tear down the huge pmd
itself and splitting the hugepage itself, instead of relaying on
split_huge_page, but that sounds very reasonable, especially
considering split_huge_page wouldn't currently transfer the reserved
bit anyway.

Signed-off-by: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Gleb Natapov <gleb@redhat.com>
virt/kvm/kvm_main.c

index d2836788561e1f42437f71ec0bbe08c637384668..0fc25aed79a8d84222c124f52bd3d131d5403070 100644 (file)
@@ -102,28 +102,8 @@ static bool largepages_enabled = true;
 
 bool kvm_is_mmio_pfn(pfn_t pfn)
 {
-       if (pfn_valid(pfn)) {
-               int reserved;
-               struct page *tail = pfn_to_page(pfn);
-               struct page *head = compound_trans_head(tail);
-               reserved = PageReserved(head);
-               if (head != tail) {
-                       /*
-                        * "head" is not a dangling pointer
-                        * (compound_trans_head takes care of that)
-                        * but the hugepage may have been splitted
-                        * from under us (and we may not hold a
-                        * reference count on the head page so it can
-                        * be reused before we run PageReferenced), so
-                        * we've to check PageTail before returning
-                        * what we just read.
-                        */
-                       smp_rmb();
-                       if (PageTail(tail))
-                               return reserved;
-               }
-               return PageReserved(tail);
-       }
+       if (pfn_valid(pfn))
+               return PageReserved(pfn_to_page(pfn));
 
        return true;
 }