On Wed, Nov 20, 2019 at 07:41:44PM +0800, Alex Shi wrote:
> 在 2019/11/20 上午12:04, Johannes Weiner 写道:
> >> @@ -1246,6 +1245,46 @@ struct lruvec *mem_cgroup_page_lruvec(struct page *page, struct pglist_data *pgd
> >> return lruvec;
> >> }
> >>
> >> +struct lruvec *lock_page_lruvec_irq(struct page *page,
> >> + struct pglist_data *pgdat)
> >> +{
> >> + struct lruvec *lruvec;
> >> +
> >> +again:
> >> + rcu_read_lock();
> >> + lruvec = mem_cgroup_page_lruvec(page, pgdat);
> >> + spin_lock_irq(&lruvec->lru_lock);
> >> + rcu_read_unlock();
> > The spinlock doesn't prevent the lruvec from being freed
> >
> > You deleted the rules from the mem_cgroup_page_lruvec() documentation,
> > but they still apply: if the page is already !PageLRU() by the time
> > you get here, it could get reclaimed or migrated to another cgroup,
> > and that can free the memcg/lruvec. Merely having the lru_lock held
> > does not prevent this.
>
>
> Forgive my idiot, I still don't know the details of unsafe lruvec here.
> From my shortsight, the spin_lock_irq(embedded a preempt_disable) could block all rcu syncing thus, keep all memcg alive until the preempt_enabled in unspinlock, is this right?
> If so even the page->mem_cgroup is migrated to others cgroups, the new and old cgroup should still be alive here.
You are right about the freeing part, I missed this. And I should have
read this email here before sending out my "fix" to the current code;
thankfully Hugh re-iterated my mistake on that thread. My apologies.
But I still don't understand how the moving part is safe. You look up
the lruvec optimistically, lock it, then verify the lookup. What keeps
page->mem_cgroup from changing after you verified it?
lock_page_lruvec(): mem_cgroup_move_account():
again:
rcu_read_lock()
lruvec = page->mem_cgroup->lruvec
isolate_lru_page()
spin_lock_irq(&lruvec->lru_lock)
rcu_read_unlock()
if page->mem_cgroup->lruvec != lruvec:
spin_unlock_irq(&lruvec->lru_lock)
goto again;
page->mem_cgroup = new cgroup
putback_lru_page() // new lruvec
SetPageLRU()
return lruvec; // old lruvec
The caller assumes page belongs to the returned lruvec and will then
change the page's lru state with a mismatched page and lruvec.
If we could restrict lock_page_lruvec() to working only on PageLRU
pages, we could fix the problem with memory barriers. But this won't
work for split_huge_page(), which is AFAICT the only user that needs
to freeze the lru state of a page that could be isolated elsewhere.
So AFAICS the only option is to lock out mem_cgroup_move_account()
entirely when the lru_lock is held. Which I guess should be fine.
