Re: [PATCH 2/3] mm/page_alloc: convert zone_pcp_update() to use on_each_cpu()instead of stop_machine()

From: Cody P Schafer
Date: Mon Apr 08 2013 - 15:49:50 EST


On 04/08/2013 12:26 PM, KOSAKI Motohiro wrote:
(4/8/13 1:32 PM), Cody P Schafer wrote:
On 04/07/2013 08:39 AM, KOSAKI Motohiro wrote:
(4/5/13 4:33 PM), Cody P Schafer wrote:
No off-cpu users of the percpu pagesets exist.

zone_pcp_update()'s goal is to adjust the ->high and ->mark members of a
percpu pageset based on a zone's ->managed_pages. We don't need to drain
the entire percpu pageset just to modify these fields. Avoid calling
setup_pageset() (and the draining required to call it) and instead just
set the fields' values.

This does change the behavior of zone_pcp_update() as the percpu
pagesets will not be drained when zone_pcp_update() is called (they will
end up being shrunk, not completely drained, later when a 0-order page
is freed in free_hot_cold_page()).

Signed-off-by: Cody P Schafer <cody@xxxxxxxxxxxxxxxxxx>

NAK.

1) zone_pcp_update() is only used from memory hotplug and it require page drain.

I'm looking at this code because I'm currently working on a patchset
which adds another interface which modifies zone sizes, so "only used
from memory hotplug" is a temporary thing (unless I discover that
zone_pcp_update() is not intended to do what I want it to do).

maybe yes, maybe no. I don't know temporary or not. However the fact is,
you must not break anywhere. You need to look all caller always.

Right, which is why I want to understand memory hotplug's actual requirements.

2) stop_machin is used for avoiding race. just removing it is insane.

What race? Is there a cross cpu access to ->high & ->batch that makes
using on_each_cpu() instead of stop_machine() inappropriate? It is
absolutely not just being removed.

OK, I missed that. however your code is still wrong.
However you can't call free_pcppages_bulk() from interrupt context and
then you can't use on_each_cpu() anyway.

Given drain_pages() implementation, I find that hard to believe (It uses on_each_cpu_mask() and eventually calls free_pcppages_bulk()).

Can you provide a reference backing up your statement?

If this turns out to be an issue, schedule_on_each_cpu() could be an alternative.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/