Is there a way to do an architecture specific shake of memory?

Is there a way to do an architecture specific shake of memory?

[Robin Holt]

I am in the process of fixing the quicklist handling for the ia64 page
tables.  The problem we are running into is the quicklists are per-cpu.
When memory is running short, the only current callout to shrink the
quicklist is coming from cpu_idle().  The most recent suggestion is to
schedule_delayed_work_on() for every cpu in the system and have them
check to see if the quicklist needs to be shrunk.  This feels wrong.

What I would like to do is have wakup_kswapd(), kswapd() or
balance_pdgat() do the actual call to smp_call_function_single() as
needed to try to shrink the quicklists.  This would need to be an ia64
only change.  Is there already a method for the architecture to get
control during any part of the process?  If not, where would be the most
acceptable place?

Any suggestions are welcome.

Thanks,
Robin Holt
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"aart@kvack.org"> aart@kvack.org </a>

Re: Is there a way to do an architecture specific shake of memory?

[Christoph Lameter]

On Tue, 8 Mar 2005, Robin Holt wrote:

> Any suggestions are welcome.

Check when you free items how long the list of free items is and if its
too long free some of them.


--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"aart@kvack.org"> aart@kvack.org </a>

Re: Is there a way to do an architecture specific shake of memory?

[Robin Holt]

On Tue, Mar 08, 2005 at 06:35:07PM -0800, Christoph Lameter wrote:
> On Tue, 8 Mar 2005, Robin Holt wrote:
> 
> > Any suggestions are welcome.
> 
> Check when you free items how long the list of free items is and if its
> too long free some of them.

That is done as well as a check from cpu_idle.  The amount of free is
being changed from a boot-time very large number computed from total
memory to a per-node percentage of free memory.  A concern was raised
for the eventuality of a process constantly running on a cpu so the idle
calls never happen, a memory hog application begins to consume memory on
the node, and nothing ever shakes the memory free from the quicklists.
This led to the suggestion of a timer based shaker.  I would rather put
it into the blocked allocation path than have it timer based and hope
we pick the correct resolution of timer.  The problem is the quicklists
are per-cpu so we need to try flushing the quicklists for the cpus on
the affected node.

Thanks,
Robin Holt
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"aart@kvack.org"> aart@kvack.org </a>