huge tmpfs: fix Committed_AS leak

Under swapping load on huge tmpfs, /proc/meminfo's Committed_AS grows
bigger and bigger: just a cosmetic issue for most users, but disabling
for those who run without overcommit (/proc/sys/vm/overcommit_memory 2).

shmem_uncharge() was forgetting to unaccount __vm_enough_memory's
charge, and shmem_charge() was forgetting it on the filesystem-full
error path.

Fixes: 800d8c6 ("shmem: add huge pages support")
Signed-off-by: Hugh Dickins <hughd@google.com>
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>

mm/shmem.c

@@ -270,7 +270,7 @@ bool shmem_charge(struct inode *inode, long pages)
 		info->alloced -= pages;
 		shmem_recalc_inode(inode);
 		spin_unlock_irqrestore(&info->lock, flags);
-
+		shmem_unacct_blocks(info->flags, pages);
 		return false;
 	}
 	percpu_counter_add(&sbinfo->used_blocks, pages);
@@ -291,6 +291,7 @@ void shmem_uncharge(struct inode *inode, long pages)
 
 	if (sbinfo->max_blocks)
 		percpu_counter_sub(&sbinfo->used_blocks, pages);
+	shmem_unacct_blocks(info->flags, pages);
 }
 
 /*