On Mon, 23 Jan 2012 14:05:33 -0800 Ying Han <yinghan@xxxxxxxxxx> wrote: > On Wed, Jan 18, 2012 at 3:53 PM, KAMEZAWA Hiroyuki > <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote: > > On Wed, 18 Jan 2012 11:47:03 +0100 > > Michal Hocko <mhocko@xxxxxxx> wrote: > > > >> On Wed 18-01-12 09:12:26, KAMEZAWA Hiroyuki wrote: > >> > On Tue, 17 Jan 2012 17:46:05 +0100 > >> > Michal Hocko <mhocko@xxxxxxx> wrote: > >> > > >> > > On Fri 13-01-12 17:40:19, KAMEZAWA Hiroyuki wrote: > >> [...] > >> > > > This patch removes PCG_MOVE_LOCK and add hashed rwlock array > >> > > > instead of it. This works well enough. Even when we need to > >> > > > take the lock, > >> > > > >> > > Hmmm, rwlocks are not popular these days very much. > >> > > Anyway, can we rather make it (source) memcg (bit)spinlock instead. We > >> > > would reduce false sharing this way and would penalize only pages from > >> > > the moving group. > >> > > > >> > per-memcg spinlock ? > >> > >> Yes > >> > >> > The reason I used rwlock() is to avoid disabling IRQ. This routine > >> > will be called by IRQ context (for dirty ratio support). So, IRQ > >> > disable will be required if we use spinlock. > >> > >> OK, I have missed the comment about disabling IRQs. It's true that we do > >> not have to be afraid about deadlocks if the lock is held only for > >> reading from the irq context but does the spinlock makes a performance > >> bottleneck? We are talking about the slowpath. > >> I could see the reason for the read lock when doing hashed locks because > >> they are global but if we make the lock per memcg then we shouldn't > >> interfere with other updates which are not blocked by the move. > >> > > > > Hm, ok. In the next version, I'll use per-memcg spinlock (with hash if necessary) > > Just want to make sure I understand it, even we make the lock > per-memcg, there is still a false sharing of pc within one memcg. Do > we need to demonstrate the effect ? > Hmm, I'll try some. Account_move occurs when a) a task is moved to other cgroup b) a cgroup is removed. I think checking case a) will be enough because there is no task in a memcg while it is being removed. Then, I'll measure performace of file mapping while moving task repeatedly. There will be spinlock conflict. - I'll consider to make the range of spinlock small. - I'll consider have a hash of spinlock or spinlock based of page-zone and types. (It's easy to make spinlock as to be per-memcg-per-zone.) > Also, I don't get the point of why spinlock instead of rwlock in this case? > >From Documentation/spinlocks.txt > NOTE! reader-writer locks require more atomic memory operations than > simple spinlocks. Unless the reader critical section is long, you > are better off just using spinlocks. > NOTE! We are working hard to remove reader-writer spinlocks in most > cases, so please don't add a new one without consensus. (Instead, see > Documentation/RCU/rcu.txt for complete information.) > I don't have enough strong motivation to use rwlock. But if rwlock works enough well rather than spinlocks, it will be a choice. Thanks, -Kame -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx"> email@xxxxxxxxx </a>