KAMEZAWA Hiroyuki wrote:

(2012/03/20 4:59), Konstantin Khlebnikov wrote:

quoted

KAMEZAWA Hiroyuki wrote:

quoted

This is just an RFC...test is not enough yet.

I know it's merge window..this post is just for sharing idea.

This patch merges pc->flags and pc->mem_cgroup into a word. Then,
memcg's overhead will be 8bytes per page(4096bytes?).

Because this patch will affect all memory cgroup developers, I'd like to
show patches before MM Summit. I think we can agree the direction to
reduce size of page_cgroup..and finally integrate into 'struct page'
(and remove cgroup_disable= boot option...)

Patch 1/3 - introduce pc_to_mem_cgroup and hide pc->mem_cgroup
Patch 2/3 - remove pc->mem_cgroup
Patch 3/3 - remove memory barriers.

I'm now wondering when this change should be merged....

This is cool, but maybe we should skip this temporary step and merge all this stuff into page->flags.

Why we should skip and delay reduction of size of page_cgroup
which is considered as very big problem ?

I think it would be better to solve problem completely and kill page_cgroup in one step.

quoted

I think we can replace zone-id and node-id in page->flags with cumulative dynamically allocated lruvec-id,
so there will be enough space for hundred cgroups even on 32-bit systems.

Where section-id  is ?
IIUC, now, page->section->zone/node is calculated if CONFIG_SPARSEMEM.

Yeah, sections are biggest problem there. I hope we can unravel this knot.
In the worst case we can extend page->flags upto 64-bits.

BTW, I doubt that we can modify page->flags dynamically with multi-bit operations...using
cmpxchg per each page when it's charged/uncharged/other ?

we can do atomic_xor(&page->flags, new-lruvec-id ^ old-lruvec-id) or
atomic_add(&page->flags, new-lruvec-id - old-lruvec-id) they should work faster than cmpxchg

quoted

After lru_lock splitting page to lruvec translation will be much frequently used than page to zone,
so page->zone and page->node translations can be implemented as page->lruvec->zone and page->lruvec->node.

And need to take rcu_read_lock() around page_zone() ?

Hmm, it depends. For kernel-pages there will be pointer to root-lruvec, so no protection required.
If we hold lru_lock we also don't need this rcu_read_lock.

Thanks,
-Kame