Re: + zram-support-compaction.patch added to -mm tree

[Minchan Kim <minchan@kernel.org>]

On Mon, Mar 09, 2015 at 10:27:28AM +0900, Sergey Senozhatsky wrote:
> On (03/09/15 10:05), Minchan Kim wrote:
> > > well, to be fair, compaction is a zsmalloc internal. zram has nothing to do with
> > > it.
> > > 
> > > but do we we even need this stat? it seems that
> > > 
> > >    mem_total_used (before compaction) - mem_total_user (after comapction)
> > > 
> > > will give user an idea on how much memory was compacted.
> > 
> > It's not enough. What I want to know is compaction efficiency per client of
> > zsmalloc(ie, zram). 
> >
> 
> so what a typical user can do with this information? isn't it an entirely
> debug info that makes some hidden sense only to developers?

Absolutely true.

> 
> if you insist on exporting this as a zram stat for everyone how obout
> starting to move away from per-stat RO sysfs attrs. it seems that we have
> uncomfortably a lot of sysfs attrs, and that doesn't make life easier in
> user space. for example, block devices have /sys/block/.../stat file:
> 
> /sys/block/sda$ cat stat
>    45931       59  2075686   289906    55768     9229  1967800   318033        0   193583   607806
> 
> and there are no num_reads, num_writes, num_failed_reads, num_failed_writes,
> etc., etc. per-stat sysfs attrs force user-space to do lots of syscalls:
> open(), read(), close() with error control on every step; for every stat.

I absoulte agree with you and I really wanted to tidy it up but was no
time. Sergey, Could you contribute? If you have no time, I will do by
myself but it would be low priority now.


> 
> so how about introducing zram<id>/malloc_stats (or any similar name) and
> provide compaction and all future allocator related stats there
> (via s*printf("%d %d %d", ....)) ?
> 
> 	-ss
> 
> > IOW, (how many of freed pages / how many of objects) per
> > zs_compact.
> > 
> > > 
> > > 	-ss
> > > 
> > > > That's why it is on debugfs. If we add the stat into zsmalloc, we should turn on debugfs
> > > > and CONFIG_ZSMALLOC_STAT to see *a* stat. Even, CONFIG_ZSMALLOC_STAT will add
> > > > unncessary overheads to account another stats fo zsmalloc internals.
> > > > 
> > > > As well, if we add auto-compacion like stuff in zsmalloc(ie, it will trigger
> > > > by itself if fragmention is over to predefined theshold), the stat will
> > > > accumulate stat while someone want to see snapshot compaction effiecieny
> > > > of the moment.
> > > > 
> > > > So, I want to keep it in zram now.
> > > > 
> > > > -- 
> > > > Kind regards,
> > > > Minchan Kim
> > > > 
> > 
> > -- 
> > Kind regards,
> > Minchan Kim
> > 

-- 
Kind regards,
Minchan Kim