On 09/24/2012 05:43 PM, David Rientjes wrote:

On Fri, 21 Sep 2012, Nico Schottelius wrote:

quoted

 Active / Total Objects (% used)    : 1165130 / 1198087 (97.2%)
 Active / Total Slabs (% used)      : 81027 / 81027 (100.0%)
 Active / Total Caches (% used)     : 69 / 101 (68.3%)
 Active / Total Size (% used)       : 1237249.81K / 1246521.94K (99.3%)
 Minimum / Average / Maximum Object : 0.01K / 1.04K / 15.23K

  OBJS ACTIVE  USE OBJ SIZE  SLABS OBJ/SLAB CACHE SIZE NAME                   
993607 993607 100%    1.21K  75358       26   2411456K jfs_ip

Well that doesn't look good.  100% of the inode cache for jfs are being 
used which either means

 - there's a memory leak, or

maybe a missing iput() somewhere?

Nico, does unmounting the usb drive after killing the backup clean up
the jfs inode cache?

 - there's some sort of throttling issue in jfs.

And those objects are consuming ~2.3GB of slab on your 4GB machine and 
seems to only have occurred between v3.4.2 to v3.5.3.

Almost nothing in jfs has changed between these releases. Only this:

vfs: Rename end_writeback() to clear_inode()

It would be interesting to see what kmemleak would tell us.

Adding Dave Kleikamp to the cc.