Re: quicklists confuse meminfo

[Thomas Gleixner <tglx@linutronix.de>]

On Sun, 9 Mar 2008, Ingo Molnar wrote:

> 
> * Thomas Gleixner <tglx@linutronix.de> wrote:
> 
> > Bart reported http://bugzilla.kernel.org/show_bug.cgi?id=9991. He 
> > assumed a memory leak in 32bit kernels when he analyzed the output of 
> > /proc/meminfo.
> > 
> > The leak is not a leak, it's an accounting bug. quicklists keep a 
> > large amount of pages which are accounted as used memory.
> [...]
> > Another strange observation about quicklists is the imbalance of the 
> > quicklists across CPUs. Running the above loop on a 2way machine I can 
> > observe that the quicklist pages are acuumulating on one CPU. Stopping 
> > and restarting the loop a couple of times can shift the accumulation 
> > from one to the other CPU.
> 
> hm. I think we should not let this much RAM hang around in a 
> special-purpose allocator like quicklists. Shouldnt the quicklists be 
> temporary in nature, and be trimmed much more agressively?
> 
> in fact, we have a check_pgt_cache() call in cpu_idle(), which does:
> 
>         quicklist_trim(0, pgd_dtor, 25, 16);
> 
> but it appears we dont do quicklist trimming anywhere else! So if a 
> system has no idle time, the quicklist can grow unbounded, and that's a 
> real memory leak IMO.

Right, also the quicklist_trim() in idle() is freeing at max 16 pages
in one go. According to the quicklist_trim() code we keep up to
(node_free_pages / 16) in the quicklist unconditionally, which seems
rather odd as well.

Thanks,
	tglx