zsmalloc: fix migrate_zspage-zs_free race condition
authorJunil Lee <junil0814.lee@lge.com>
Wed, 20 Jan 2016 22:58:18 +0000 (14:58 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Thu, 21 Jan 2016 01:09:18 +0000 (17:09 -0800)
record_obj() in migrate_zspage() does not preserve handle's
HANDLE_PIN_BIT, set by find_aloced_obj()->trypin_tag(), and implicitly
(accidentally) un-pins the handle, while migrate_zspage() still performs
an explicit unpin_tag() on the that handle.  This additional explicit
unpin_tag() introduces a race condition with zs_free(), which can pin
that handle by this time, so the handle becomes un-pinned.

Schematically, it goes like this:

  CPU0                                        CPU1
  migrate_zspage
    find_alloced_obj
      trypin_tag
        set HANDLE_PIN_BIT                    zs_free()
                                                pin_tag()
  obj_malloc() -- new object, no tag
  record_obj() -- remove HANDLE_PIN_BIT           set HANDLE_PIN_BIT
  unpin_tag()  -- remove zs_free's HANDLE_PIN_BIT

The race condition may result in a NULL pointer dereference:

  Unable to handle kernel NULL pointer dereference at virtual address 00000000
  CPU: 0 PID: 19001 Comm: CookieMonsterCl Tainted:
  PC is at get_zspage_mapping+0x0/0x24
  LR is at obj_free.isra.22+0x64/0x128
  Call trace:
     get_zspage_mapping+0x0/0x24
     zs_free+0x88/0x114
     zram_free_page+0x64/0xcc
     zram_slot_free_notify+0x90/0x108
     swap_entry_free+0x278/0x294
     free_swap_and_cache+0x38/0x11c
     unmap_single_vma+0x480/0x5c8
     unmap_vmas+0x44/0x60
     exit_mmap+0x50/0x110
     mmput+0x58/0xe0
     do_exit+0x320/0x8dc
     do_group_exit+0x44/0xa8
     get_signal+0x538/0x580
     do_signal+0x98/0x4b8
     do_notify_resume+0x14/0x5c

This patch keeps the lock bit in migration path and update value
atomically.

Signed-off-by: Junil Lee <junil0814.lee@lge.com>
Signed-off-by: Minchan Kim <minchan@kernel.org>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Sergey Senozhatsky <sergey.senozhatsky.work@gmail.com>
Cc: <stable@vger.kernel.org> [4.1+]
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
mm/zsmalloc.c

index e7414cec220b3bce6cd5abbdd6a769f1f2c334a5..2d7c4c11fc635d3822d64e8b2d796414945305f7 100644 (file)
@@ -309,7 +309,12 @@ static void free_handle(struct zs_pool *pool, unsigned long handle)
 
 static void record_obj(unsigned long handle, unsigned long obj)
 {
-       *(unsigned long *)handle = obj;
+       /*
+        * lsb of @obj represents handle lock while other bits
+        * represent object value the handle is pointing so
+        * updating shouldn't do store tearing.
+        */
+       WRITE_ONCE(*(unsigned long *)handle, obj);
 }
 
 /* zpool driver */
@@ -1635,6 +1640,13 @@ static int migrate_zspage(struct zs_pool *pool, struct size_class *class,
                free_obj = obj_malloc(d_page, class, handle);
                zs_object_copy(free_obj, used_obj, class);
                index++;
+               /*
+                * record_obj updates handle's value to free_obj and it will
+                * invalidate lock bit(ie, HANDLE_PIN_BIT) of handle, which
+                * breaks synchronization using pin_tag(e,g, zs_free) so
+                * let's keep the lock bit.
+                */
+               free_obj |= BIT(HANDLE_PIN_BIT);
                record_obj(handle, free_obj);
                unpin_tag(handle);
                obj_free(pool, class, used_obj);