mm: allow arch code to control the user page table ceiling
Hugh Dickins [Mon, 29 Apr 2013 22:07:44 +0000 (15:07 -0700)]
commit 6ee8630e02be6dd89926ca0fbc21af68b23dc087 upstream.

On architectures where a pgd entry may be shared between user and kernel
(e.g.  ARM+LPAE), freeing page tables needs a ceiling other than 0.
This patch introduces a generic USER_PGTABLES_CEILING that arch code can
override.  It is the responsibility of the arch code setting the ceiling
to ensure the complete freeing of the page tables (usually in
pgd_free()).

[catalin.marinas@arm.com: commit log; shift_arg_pages(), asm-generic/pgtables.h changes]
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Catalin Marinas <catalin.marinas@arm.com>
Cc: Russell King <linux@arm.linux.org.uk>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

fs/exec.c
include/asm-generic/pgtable.h
mm/mmap.c

index e3a7e36..2b7f5ff 100644 (file)
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -627,7 +627,7 @@ static int shift_arg_pages(struct vm_area_struct *vma, unsigned long shift)
                 * when the old and new regions overlap clear from new_end.
                 */
                free_pgd_range(&tlb, new_end, old_end, new_end,
-                       vma->vm_next ? vma->vm_next->vm_start : 0);
+                       vma->vm_next ? vma->vm_next->vm_start : USER_PGTABLES_CEILING);
        } else {
                /*
                 * otherwise, clean from old_start; this is done to not touch
@@ -636,7 +636,7 @@ static int shift_arg_pages(struct vm_area_struct *vma, unsigned long shift)
                 * for the others its just a little faster.
                 */
                free_pgd_range(&tlb, old_start, old_end, new_end,
-                       vma->vm_next ? vma->vm_next->vm_start : 0);
+                       vma->vm_next ? vma->vm_next->vm_start : USER_PGTABLES_CEILING);
        }
        tlb_finish_mmu(&tlb, new_end, old_end);
 
index c7ec2cd..9a6eb25 100644 (file)
@@ -7,6 +7,16 @@
 #include <linux/mm_types.h>
 #include <linux/bug.h>
 
+/*
+ * On almost all architectures and configurations, 0 can be used as the
+ * upper ceiling to free_pgtables(): on many architectures it has the same
+ * effect as using TASK_SIZE.  However, there is one configuration which
+ * must impose a more careful limit, to avoid freeing kernel pgtables.
+ */
+#ifndef USER_PGTABLES_CEILING
+#define USER_PGTABLES_CEILING  0UL
+#endif
+
 #ifndef __HAVE_ARCH_PTEP_SET_ACCESS_FLAGS
 extern int ptep_set_access_flags(struct vm_area_struct *vma,
                                 unsigned long address, pte_t *ptep,
index 2add0a1..3635d47 100644 (file)
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -1920,7 +1920,7 @@ static void unmap_region(struct mm_struct *mm,
        unmap_vmas(&tlb, vma, start, end, &nr_accounted, NULL);
        vm_unacct_memory(nr_accounted);
        free_pgtables(&tlb, vma, prev ? prev->vm_end : FIRST_USER_ADDRESS,
-                                next ? next->vm_start : 0);
+                                next ? next->vm_start : USER_PGTABLES_CEILING);
        tlb_finish_mmu(&tlb, start, end);
 }
 
@@ -2308,7 +2308,7 @@ void exit_mmap(struct mm_struct *mm)
        unmap_vmas(&tlb, vma, 0, -1, &nr_accounted, NULL);
        vm_unacct_memory(nr_accounted);
 
-       free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, 0);
+       free_pgtables(&tlb, vma, FIRST_USER_ADDRESS, USER_PGTABLES_CEILING);
        tlb_finish_mmu(&tlb, 0, -1);
 
        /*