On Thu, 22 Jul 2010 16:54:25 +0900
Daisuke Nishimura <nishimura@xxxxxxxxxxxxxxxxx> wrote:
> On Thu, 22 Jul 2010 15:04:45 +0900
> KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
>
> > On Thu, 22 Jul 2010 14:43:56 +0900
> > Daisuke Nishimura <nishimura@xxxxxxxxxxxxxxxxx> wrote:
> >
> > > On Wed, 21 Jul 2010 19:58:31 +0900
> > > KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> > >
> > > > These are just a _toy_ level patches yet. My final purpose is to use indexed array
> > > > for mem_cgroup itself, it has IDs.
> > > >
> > > > Background:
> > > > memory cgroup uses struct page_cgroup for tracking all used pages. It's defined as
> > > > ==
> > > > struct page_cgroup {
> > > > unsigned long flags;
> > > > struct mem_cgroup *mem_cgroup;
> > > > struct page *page;
> > > > struct list_head lru; /* per cgroup LRU list */
> > > > };
> > > > ==
> > > > and this increase the cost of per-page-objects dramatically. Now, we have
> > > > troubles on this object.
> > > > 1. Recently, a blkio-tracking guy wants to add "blockio-cgroup" information
> > > > to page_cgroup. But our concern is extra 8bytes per page.
> > > > 2. At tracking dirty page status etc...we need some trick for safe access
> > > > to page_cgroup and memcgroup's information. For example, a small seqlock.
> > > >
> > > > Now, each memory cgroup has its own ID (0-65535). So, if we can replace
> > > > 8byte of pointer "pc->mem_cgroup" with an ID, which is 2 bytes, we may able
> > > > to have another room. (Moreover, I think we can reduce the number of IDs...)
> > > >
> > > > This patch is a trial for implement a virually-indexed on-demand array and
> > > > an example of usage. Any commetns are welcome.
> > > >
> >
> > Hi,
> > > So, your purpose is to:
> > >
> > > - make the size of mem_croup small(by [2/2])
> > It's just an example to test virt-array. I don't convice it can
> > save memory or make something fast. and I found a bug in free routine.)
> >
> >
> > > - manage all the mem_cgroup in virt-array indexed by its ID(it would be faster
> > > than using css_lookup)
> > yes.
> >
> > > - replace pc->mem_cgroup by its ID and make the size of page_cgroup small
> > >
> > yes.
> >
> > Final style I'm thinking is
> > struct page_cgroup {
> > unsigned long flags;
> > spinlock_t lock; # for lock_page_cgroup()
> > unsigned short memcg;
> > unsigned short blkio;
> > struct page *page;
> > struct list_head list;
> > };
> > This will be benefical in 64bit. About 32bit, I may have to merge some fields.
> > Or I may have to add some "version" field for updating memcg's statistics
> > without locks. memcg field may be able to be moved onto high-bits of "flags"
> > because it's stable value unless it's not under move_charge.
> > (IIUC, at move_charge, memcg is off-LRU and there are no race with AcctLRU bit
> > v.s. pc->mem_cgroup field. With other flags, lock_page_cgroup() works enough.)
> >
> > Anyway, race with move_charge() will be the last enemy for us to track
> > dirty pages etc...at least, this kind of "make room" job is required, I feel.
> >
> > There are many things to be considered, but I'm a bit in hurry. I'd like to do
> > some preparation before Mel at el rewrites memory-reclaim+writeback complelety.
> >
> Thank you for clarifying your thought.
>
> I have one comment for this patch.
> > +static int idx_used(const struct virt_array *v, int idx)
> > +{
> > + return test_bit(idx, v->map);
> > +}
> > +
> Who set the bit ?
> Shouldn't we set it at alloc_varray_item() ?
>
yes. That's bug. my host crashed ;(
and....clear_bit() is also lacked .
I'll prepare a patch for memory_cgroup_on_array _after_ tests ;)
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>