swap: fix shmem swapping when more than 8 areas

commit 9b15b817f3d62409290fd56fe3cbb076a931bb0a upstream.

Minchan Kim reports that when a system has many swap areas, and tmpfs
swaps out to the ninth or more, shmem_getpage_gfp()'s attempts to read
back the page cannot locate it, and the read fails with -ENOMEM.

Whoops.  Yes, I blindly followed read_swap_header()'s pte_to_swp_entry(
swp_entry_to_pte()) technique for determining maximum usable swap
offset, without stopping to realize that that actually depends upon the
pte swap encoding shifting swap offset to the higher bits and truncating
it there.  Whereas our radix_tree swap encoding leaves offset in the
lower bits: it's swap "type" (that is, index of swap area) that was
truncated.

Fix it by reducing the SWP_TYPE_SHIFT() in swapops.h, and removing the
broken radix_to_swp_entry(swp_to_radix_entry()) from read_swap_header().

This does not reduce the usable size of a swap area any further, it
leaves it as claimed when making the original commit: no change from 3.0
on x86_64, nor on i386 without PAE; but 3.0's 512GB is reduced to 128GB
per swapfile on i386 with PAE.  It's not a change I would have risked
five years ago, but with x86_64 supported for ten years, I believe it's
appropriate now.

Hmm, and what if some architecture implements its swap pte with offset
encoded below type? That would equally break the maximum usable swap
offset check.  Happily, they all follow the same tradition of encoding
offset above type, but I'll prepare a check on that for next.

Reported-and-Reviewed-and-Tested-by: Minchan Kim <minchan@kernel.org>
Signed-off-by: Hugh Dickins <hughd@google.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

1 files changed, 5 insertions, 3 deletions

diff --git a/include/linux/swapops.h b/include/linux/swapops.h
index 792d16d9cbc..47ead515c81 100644
--- a/include/linux/swapops.h
+++ b/include/linux/swapops.h
@@ -9,13 +9,15 @@
  * get good packing density in that tree, so the index should be dense in
  * the low-order bits.
  *
- * We arrange the `type' and `offset' fields so that `type' is at the five
+ * We arrange the `type' and `offset' fields so that `type' is at the seven
  * high-order bits of the swp_entry_t and `offset' is right-aligned in the
- * remaining bits.
+ * remaining bits.  Although `type' itself needs only five bits, we allow for
+ * shmem/tmpfs to shift it all up a further two bits: see swp_to_radix_entry().
  *
  * swp_entry_t's are *never* stored anywhere in their arch-dependent format.
  */
-#define SWP_TYPE_SHIFT(e)	(sizeof(e.val) * 8 - MAX_SWAPFILES_SHIFT)
+#define SWP_TYPE_SHIFT(e)	((sizeof(e.val) * 8) - \
+			(MAX_SWAPFILES_SHIFT + RADIX_TREE_EXCEPTIONAL_SHIFT))
 #define SWP_OFFSET_MASK(e)	((1UL << SWP_TYPE_SHIFT(e)) - 1)
 
 /*