Re: [PATCH v3 2/2] tmpfs: Make tmpfs scalable with percpu_counter for used blocks

[Minchan Kim <minchan.kim@gmail.com>]

On Thu, Jun 17, 2010 at 09:08:59PM -0700, Andrew Morton wrote:
> On Fri, 18 Jun 2010 10:35:51 +0900 Minchan Kim <minchan.kim@gmail.com> wrote:
> 
> > > __struct shmem_sb_info {
> > > __ __ __ __unsigned long max_blocks; __ /* How many blocks are allowed */
> > > - __ __ __ unsigned long free_blocks; __/* How many are left for allocation */
> > > + __ __ __ struct percpu_counter used_blocks; __/* How many are allocated */
> > 
> > Just a nitpick.
> > Why do you change free_blocks and used_blocks?
> > I think we can use free_blocks following as.
> > 
> > ex)
> > if (percpu_counter_compare(&sbinfo->free_blocks, 0))
> 
> See previous lengthy discussion ;)
> 
> If we count free_blocks then we need to alter the value of free_blocks
> in remount_fs, and reinitialising distributed counters on-the-fly is
> ugly.
> 
> I suppose we could have done it by doing a large add or sub in
> remount_fs, and keeping track of the exact value of free_blocks
> elsewhere in the superblock, but it's far simpler this way.
> 

Thanks, Andrew. I found your comment in previous mail thread. 
Tim, Please add below Andrew's comment in description. 

It would make reviewer happy in future.

--
Andrew's comment
" It a bit buggy - using percpu_counter_init() against an
already-initialised percpu_counter() is leaky.  I suspect that's
happening in remount_fs.

A better approach would be to remove free_blocks altogether and add a
new `percpu_counter used_blocks;' which simply counts how many blocks
are presently in use.  Such a thing would then never need to be
reinitialised."


-- 
Kind regards,
Minchan Kim