Messages in this thread | | | Date | Sat, 19 Nov 2011 08:09:41 +1100 | From | Dave Chinner <> | Subject | Re: [PATCH] mm: remove struct reclaim_state |
| |
On Fri, Nov 18, 2011 at 02:07:02PM +0400, Konstantin Khlebnikov wrote: > Dave Chinner wrote: > >On Fri, Nov 18, 2011 at 01:28:06PM +0300, Konstantin Khlebnikov wrote: > >>Memory reclaimer want to know how much pages was reclaimed during shrinking slabs. > >>Currently there is special struct reclaim_state with single counter and pointer from > >>task-struct. Let's store counter direcly on task struct and account freed pages > >>unconditionally. This will reduce stack usage and simplify code in reclaimer and slab. > >> > >>Logic in do_try_to_free_pages() is slightly changed, but this is ok. > >>Nobody calls shrink_slab() explicitly before do_try_to_free_pages(), > > > >Except for drop_slab() and shake_page().... > > Indeed, but they do not care about accounting reclaimed pages and > they do not call do_try_to_free_pages() after all.
Right, so you're effectively leaving a landmine for someone to trip over - anyone that cares about accounting during shrink_slab needs to zero the value first. The current code makes this obvious by not having a reclaim structure in the cases where callers don't care about accounting - after your change the correct usage is undocumented....
Cheers,
Dave. -- Dave Chinner david@fromorbit.com
| |