On Mon, 8 Mar 2010 17:31:00 +0900, KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> On Mon, 8 Mar 2010 17:07:11 +0900
> Daisuke Nishimura <nishimura@xxxxxxxxxxxxxxxxx> wrote:
>
> > On Mon, 8 Mar 2010 11:37:11 +0900, KAMEZAWA Hiroyuki <kamezawa.hiroyu@xxxxxxxxxxxxxx> wrote:
> > > On Mon, 8 Mar 2010 11:17:24 +0900
> > > Daisuke Nishimura <nishimura@xxxxxxxxxxxxxxxxx> wrote:
> > >
> > > > > But IIRC, clear_writeback is done under treelock.... No ?
> > > > >
> > > > The place where NR_WRITEBACK is updated is out of tree_lock.
> > > >
> > > > 1311 int test_clear_page_writeback(struct page *page)
> > > > 1312 {
> > > > 1313 struct address_space *mapping = page_mapping(page);
> > > > 1314 int ret;
> > > > 1315
> > > > 1316 if (mapping) {
> > > > 1317 struct backing_dev_info *bdi = mapping->backing_dev_info;
> > > > 1318 unsigned long flags;
> > > > 1319
> > > > 1320 spin_lock_irqsave(&mapping->tree_lock, flags);
> > > > 1321 ret = TestClearPageWriteback(page);
> > > > 1322 if (ret) {
> > > > 1323 radix_tree_tag_clear(&mapping->page_tree,
> > > > 1324 page_index(page),
> > > > 1325 PAGECACHE_TAG_WRITEBACK);
> > > > 1326 if (bdi_cap_account_writeback(bdi)) {
> > > > 1327 __dec_bdi_stat(bdi, BDI_WRITEBACK);
> > > > 1328 __bdi_writeout_inc(bdi);
> > > > 1329 }
> > > > 1330 }
> > > > 1331 spin_unlock_irqrestore(&mapping->tree_lock, flags);
> > > > 1332 } else {
> > > > 1333 ret = TestClearPageWriteback(page);
> > > > 1334 }
> > > > 1335 if (ret)
> > > > 1336 dec_zone_page_state(page, NR_WRITEBACK);
> > > > 1337 return ret;
> > > > 1338 }
> > >
> > > We can move this up to under tree_lock. Considering memcg, all our target has "mapping".
> > >
> > > If we newly account bounce-buffers (for NILFS, FUSE, etc..), which has no ->mapping,
> > > we need much more complex new charge/uncharge theory.
> > >
> > > But yes, adding new lock scheme seems complicated. (Sorry Andrea.)
> > > My concerns is performance. We may need somehing new re-implementation of
> > > locks/migrate/charge/uncharge.
> > >
> > I agree. Performance is my concern too.
> >
> > I made a patch below and measured the time(average of 10 times) of kernel build
> > on tmpfs(make -j8 on 8 CPU machine with 2.6.33 defconfig).
> >
> > <before>
> > - root cgroup: 190.47 sec
> > - child cgroup: 192.81 sec
> >
> > <after>
> > - root cgroup: 191.06 sec
> > - child cgroup: 193.06 sec
> >
> > Hmm... about 0.3% slower for root, 0.1% slower for child.
> >
>
> Hmm...accepatable ? (sounds it's in error-range)
>
> BTW, why local_irq_disable() ?
> local_irq_save()/restore() isn't better ?
>
I don't have any strong reason. All of lock_page_cgroup() is *now* called w/o irq disabled,
so I used just disable()/enable() instead of save()/restore().
I think disable()/enable() is better in those cases because we need not to save/restore
eflags register by pushf/popf, but, I don't have any numbers though, there wouldn't be a big
difference in performance.
Thanks,
Daisuke Nishimura.
--
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>