mlock: operate on any regions with protection != PROT_NONE
authorMichel Lespinasse <walken@google.com>
Tue, 1 Feb 2011 01:03:41 +0000 (17:03 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 1 Feb 2011 23:20:50 +0000 (10:20 +1100)
As Tao Ma noticed, change 5ecfda0 breaks blktrace. This is because
blktrace mmaps a file with PROT_WRITE permissions but without PROT_READ,
so my attempt to not unnecessarity break COW during mlock ended up
causing mlock to fail with a permission problem.

I am proposing to let mlock ignore vma protection in all cases except
PROT_NONE. In particular, mlock should not fail for PROT_WRITE regions
(as in the blktrace case, which broke at 5ecfda0) or for PROT_EXEC
regions (which seem to me like they were always broken).

Signed-off-by: Michel Lespinasse <walken@google.com>
Acked-by: Rik van Riel <riel@redhat.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/mlock.c

index 13e81ee..c3924c7 100644 (file)
@@ -178,6 +178,13 @@ static long __mlock_vma_pages_range(struct vm_area_struct *vma,
        if ((vma->vm_flags & (VM_WRITE | VM_SHARED)) == VM_WRITE)
                gup_flags |= FOLL_WRITE;
 
+       /*
+        * We want mlock to succeed for regions that have any permissions
+        * other than PROT_NONE.
+        */
+       if (vma->vm_flags & (VM_READ | VM_WRITE | VM_EXEC))
+               gup_flags |= FOLL_FORCE;
+
        if (vma->vm_flags & VM_LOCKED)
                gup_flags |= FOLL_MLOCK;