tmpfs: change final i_blocks BUG to WARNING
authorHugh Dickins <hughd@google.com>
Fri, 16 Nov 2012 22:15:04 +0000 (14:15 -0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Mon, 26 Nov 2012 20:14:19 +0000 (12:14 -0800)
commit347bd4501eec64db816bcd2438e38784f0107041
tree8ce6c93bf699a4675ccb99cfa42d91ce0d52c327
parente7c1a252a7236c6a318c703a1bf29e3261a44e9c
tmpfs: change final i_blocks BUG to WARNING

commit 0f3c42f522dc1ad7e27affc0a4aa8c790bce0a66 upstream.

Under a particular load on one machine, I have hit shmem_evict_inode()'s
BUG_ON(inode->i_blocks), enough times to narrow it down to a particular
race between swapout and eviction.

It comes from the "if (freed > 0)" asymmetry in shmem_recalc_inode(),
and the lack of coherent locking between mapping's nrpages and shmem's
swapped count.  There's a window in shmem_writepage(), between lowering
nrpages in shmem_delete_from_page_cache() and then raising swapped
count, when the freed count appears to be +1 when it should be 0, and
then the asymmetry stops it from being corrected with -1 before hitting
the BUG.

One answer is coherent locking: using tree_lock throughout, without
info->lock; reasonable, but the raw_spin_lock in percpu_counter_add() on
used_blocks makes that messier than expected.  Another answer may be a
further effort to eliminate the weird shmem_recalc_inode() altogether,
but previous attempts at that failed.

So far undecided, but for now change the BUG_ON to WARN_ON: in usual
circumstances it remains a useful consistency check.

Signed-off-by: Hugh Dickins <hughd@google.com>
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>
mm/shmem.c