On Mon, Oct 22, 2007 at 02:56:36PM -0500, David Young wrote:

> I have an 8M tmpfs filesystem whose available blocks are exhausted over
> time.  There seems to be a memory leak, because I cannot account for more
> than 3M of the use with fstat(1) (for deleted files & dirs) and du(1),
..
> Filesystem        Size       Used      Avail %Cap    iUsed   iAvail %iCap Mounted on
> /dev/wd0e         230M        18M       200M   8%     1088   116670   0% /
> tmpfs             496K       196K       300K  39%     1211       53  95% /dev
> tmpfs             8.0M       8.0M         0B 100%     1856        0 100% /mfs

IIRC the unused pages won't be released unless the pagedaemon asks for them
back; there should be some soft of high water mark. On HEAD, tmpfs doesn't
free tmpfs_nodes until unmount, they are recycled through a per-mount list.

> Does no one else see this?  My application may be a bit unusual, both
> in that I use null mounts, and in that I have no swap activated.
> 
> Could the cause of the leak be an interaction between my null mounts
> and tmpfs?  Also, I am dimly aware of some reference-counting bug in
> tmpfs; it was mentioned in one of ad@'s commits to the vmlocking branch.
> (I do not run the vmlocking branch.)

On the vmlocking branch, tmpfs_nodes are freed when the linkcount drops to
zero. I'm seeing a bug where tmpfs_rmdir is called, and the directory nodes
very occasionally have an extra link, so they become orphaned. I haven't
found the cause yet.

Thanks,
Andrew