Re: [PATCH] mm: preallocate page before lock_page at filemap COW.(WasRe: [PATCH V2] mm: Do not keep page locked during page fault whilecharging it for memcg

[Hiroyuki Kamezawa]

2011/6/23 Michal Hocko <mhocko@xxxxxxx>:
> On Thu 23-06-11 19:01:57, KAMEZAWA Hiroyuki wrote:
>> On Thu, 23 Jun 2011 11:02:04 +0200
>> Michal Hocko <mhocko@xxxxxxx> wrote:
>>
>> > On Thu 23-06-11 17:08:11, KAMEZAWA Hiroyuki wrote:
>> > > On Thu, 23 Jun 2011 09:41:33 +0200
>> > > Michal Hocko <mhocko@xxxxxxx> wrote:
>> > [...]
>> > > > Other than that:
>> > > > Reviewed-by: Michal Hocko <mhocko@xxxxxxx>
>> > > >
>> > >
>> > > I found the page is added to LRU before charging. (In this case,
>> > > memcg's LRU is ignored.) I'll post a new version with a fix.
>> >
>> > Yes, you are right. I have missed that.
>> > This means that we might race with reclaim which could evict the COWed
>> > page wich in turn would uncharge that page even though we haven't
>> > charged it yet.
>> >
>> > Can we postpone page_add_new_anon_rmap to the charging path or it would
>> > just race somewhere else?
>> >
>>
>> I got a different idea. How about this ?
>> I think this will have benefit for non-memcg users under OOM, too.
>
> Could you be more specific? I do not see how preallocation which might
> turn out to be pointless could help under OOM.
>

We'll have no page allocation under lock_page() held in this path.
I think it is good.

>>
>> A concerns is VM_FAULT_RETRY case but wait-for-lock will be much heavier
>> than preallocation + free-for-retry cost.
>
> Preallocation is rather costly when fault handler fails (e.g. SIGBUS
> which is the easiest one to trigger).
>
I think pcp cache of free page allocater does enough good job and I guess
we'll see no problem even if there is a storm of SIGBUS.

> I am not saying this approach is bad but I think that preallocation can
> be much more costly than unlock, charge and lock&recheck approach.

memcg_is_disabled() cannot help ROOT cgroup. And additional
lock/unlock method may kill FAULT_RETRY at lock contention optimization
which was added recently.


Thanks,
-Kame
--
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/