On Thu, 23 Jul 2009 15:38:43 +0900 (JST)
Ryo Tsuruta <ryov@xxxxxxxxxxxxx> wrote:
> KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> > On Wed, 22 Jul 2009 18:40:55 +0900 (JST)
> > Ryo Tsuruta <ryov@xxxxxxxxxxxxx> wrote:
> > > KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> > > > On Tue, 21 Jul 2009 23:23:16 +0900 (JST)
> > > > Ryo Tsuruta <ryov@xxxxxxxxxxxxx> wrote:
> > > >
> > > > > This patch contains several hooks that let the blkio-cgroup framework
> > > > > to know
> > > > > which blkio-cgroup is the owner of a page before starting I/O against
> > > > > the page.
> > > >
> > > > > @@ -464,6 +465,7 @@ int add_to_page_cache_locked(struct page
> > > > > gfp_mask & GFP_RECLAIM_MASK);
> > > > > if (error)
> > > > > goto out;
> > > > > + blkio_cgroup_set_owner(page, current->mm);
> > > > >
> > > >
> > > > This part is doubtful...Is this necessary ?
> > > > I recommend you that the caller should attach owner by itself.
> > >
> > > I think that it is reasonable to add the hook right here rather than
> > > to add many hooks to a variety of places.
> > >
> > Why ? at writing, it's will be overwriten soon, IIUC. Then, this information
> > is misleading. plz add a hook like this when it means something. In this
> > case,
> > read/write callers.
> > IMO, you just increase patch's readbility but decrease easiness of
> > maintaince.
> Even though the owner is overwritten soon at writing, I'm not sure why
> inserting the hook here causes the misleading. I think that it is easy
> to understand when and where the owner is set by blkio-cgroup, and it
> does not decrease maintainability, rather than put many hooks to each
Are there _many_ callers ? I don't think so.
But okay, I don't say strong objections more if other ones say ok.
BTW, a sad information for you.
you can't call lock_page_cgroup() under radix_tree->lock.
> > Consider following situation.
> > - A process "A" has big memory. When several threads requests memory, all
> > of them are caught by a blockio cgroup of "A".
> > - A process "B" has read big file caches. When several threads requests
> > memory,
> > all of them are caught by a blockio cgroup of "B".
> > If "A" and "B" 's threshold is small, you'll see big slow down.
> > But it's not _planned_ behavior in many cases.
> > If you charges agaisnt memory owner, the admin has to set _big_ priority of
> > I/O
> > controller to "A" and "B" if it uses much memory. I think the admin can't
> > design
> > his system. It's nonsense to say "plz set I/O limit propotional to memory
> > usage of
> > your apps even if it never do I/O in usual."
> > If this blockio cgroup is introduced, people will see *unexpected* very
> > terrible slow down and the user will see heartbeat warnings/failover by
> > cluster
> > management software. Please do I/O at the priority of memory reclaiming
> > requester.
> dm-ioband gives high priority to I/O for swap-out by checking whether
> PG_swapcache flag is set on the I/O page, regardless of the assigned
> I/O bandwidth, and the bandwidth consumed for swap-out is charged to
> the owner of the pages as a debt.
> How about this approach?
I don't think it's reasonable. Why I/O device, scheduler should know about
such mm-related information ? I think layering is wrong.
And your approatch cannot be a workaround.
In follwing _typical_ case,
- A process does small logging to /var/log/mylog, once in a sec.
but it uses some amount of cold memory or shmem.
This process's logging will be delayed _unexpectedly_ by some buggy process
which does memory leak.
Xen-devel mailing list