On Tue, 3 Aug 2010 09:15:13 +0530
Balbir Singh <balbir@xxxxxxxxxxxxxxxxxx> wrote:
> * KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> [2010-08-02 19:14:10]:
>
> > From: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
> >
> > Now, addresses of memory cgroup can be calculated by their ID without complex.
> > This patch relplaces pc->mem_cgroup from a pointer to a unsigned short.
> > On 64bit architecture, this offers us more 6bytes room per page_cgroup.
> > Use 2bytes for blkio-cgroup's page tracking. More 4bytes will be used for
> > some light-weight concurrent access.
> >
> > We may able to move this id onto flags field but ...go step by step.
> >
> > Changelog: 20100730
> > - fixed some garbage added by debug code in early stage
> >
> > Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx>
> > ---
> > include/linux/page_cgroup.h | 3 ++-
> > mm/memcontrol.c | 32 +++++++++++++++++++-------------
> > mm/page_cgroup.c | 2 +-
> > 3 files changed, 22 insertions(+), 15 deletions(-)
> >
> > Index: mmotm-0727/include/linux/page_cgroup.h
> > ===================================================================
> > --- mmotm-0727.orig/include/linux/page_cgroup.h
> > +++ mmotm-0727/include/linux/page_cgroup.h
> > @@ -12,7 +12,8 @@
> > */
> > struct page_cgroup {
> > unsigned long flags;
> > - struct mem_cgroup *mem_cgroup;
> > + unsigned short mem_cgroup; /* ID of assigned memory cgroup */
> > + unsigned short blk_cgroup; /* Not Used..but will be. */
> > struct page *page;
> > struct list_head lru; /* per cgroup LRU list */
> > };
>
> Can I recommend that on 64 bit systems, we merge the flag, mem_cgroup
> and blk_cgroup into one 8 byte value. We could use
> __attribute("packed") and do something like this
>
It's a next step.
> struct page_cgroup {
> unsigned int flags;
> unsigned short mem_cgroup;
> unsigned short blk_cgroup;
> ...
> } __attribute(("packed"));
>
> Then we need to make sure we don't use more that 32 bits for flags,
> which is very much under control at the moment.
>
set_bit() requires "long" as its argument. more some trick is required.
And, IIUC, packing implies
pc->mem_cgroup = mem_cgroup_id; or
pc->blk_cgroup = blk_cgroup_id; will have race with
set/clear_bit(BIT_XXXX, &pc->flags)
This "packing" is not very easy. we have to consider all possible combinations
of operations.
> This will save us 8 bytes in total on 64 bit systems and nothing on 32
> bit systems, but will enable blkio cgroup to co-exist.
>
yes. But I have cocnerns of race condition. to do that, we need
patch 3-5. (But patch 5 adds spinlock, then no 8bytes reduce.)
Let me go step by step. I'm _really_ afraid of race conditions.
Thanks,
-Kame
--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@xxxxxxxxxx For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>