Re: [PATCH v21 17/19] mm/lru: replace pgdat lru_lock with lruvec lock

[Alex Shi]

在 2020/11/12 下午8:19, Vlastimil Babka 写道:
> On 11/5/20 9:55 AM, Alex Shi wrote:
>> This patch moves per node lru_lock into lruvec, thus bring a lru_lock for
>> each of memcg per node. So on a large machine, each of memcg don't
>> have to suffer from per node pgdat->lru_lock competition. They could go
>> fast with their self lru_lock.
>>
>> After move memcg charge before lru inserting, page isolation could
>> serialize page's memcg, then per memcg lruvec lock is stable and could
>> replace per node lru lock.
>>
>> In func isolate_migratepages_block, compact_unlock_should_abort and
>> lock_page_lruvec_irqsave are open coded to work with compact_control.
>> Also add a debug func in locking which may give some clues if there are
>> sth out of hands.
>>
>> Daniel Jordan's testing show 62% improvement on modified readtwice case
>> on his 2P * 10 core * 2 HT broadwell box.
>> https://lore.kernel.org/lkml/20200915165807.kpp7uhiw7l3loofu@xxxxxxxxxxxxxxxxxxxxxxxxxx/
>>
>> On a large machine with memcg enabled but not used, the page's lruvec
>> seeking pass a few pointers, that may lead to lru_lock holding time
>> increase and a bit regression.
>>
>> Hugh Dickins helped on the patch polish, thanks!
>>
>> Signed-off-by: Alex Shi <alex.shi@xxxxxxxxxxxxxxxxx>
>> Acked-by: Hugh Dickins <hughd@xxxxxxxxxx>
>> Cc: Rong Chen <rong.a.chen@xxxxxxxxx>
>> Cc: Hugh Dickins <hughd@xxxxxxxxxx>
>> Cc: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx>
>> Cc: Johannes Weiner <hannes@xxxxxxxxxxx>
>> Cc: Michal Hocko <mhocko@xxxxxxxxxx>
>> Cc: Vladimir Davydov <vdavydov.dev@xxxxxxxxx>
>> Cc: Yang Shi <yang.shi@xxxxxxxxxxxxxxxxx>
>> Cc: Matthew Wilcox <willy@xxxxxxxxxxxxx>
>> Cc: Konstantin Khlebnikov <khlebnikov@xxxxxxxxxxxxxx>
>> Cc: Tejun Heo <tj@xxxxxxxxxx>
>> Cc: linux-kernel@xxxxxxxxxxxxxxx
>> Cc: linux-mm@xxxxxxxxx
>> Cc: cgroups@xxxxxxxxxxxxxxx
> 
> I think I need some explanation about the rcu_read_lock() usage in lock_page_lruvec*() (and places effectively opencoding it).
> Preferably in form of some code comment, but that can be also added as a additional patch later, I don't want to block the series.
> 

Hi Vlastimil, 

Thanks for comments!

Oh, we did talk about the rcu_read_lock which is used to block memcg destroy during locking.
and the spin_lock actually includes a rcu_read_lock(). Yes, we could add this comments later.

> mem_cgroup_page_lruvec() comment says
> 
>  * This function relies on page->mem_cgroup being stable - see the
>  * access rules in commit_charge().
> 
> commit_charge() comment:
> 
>          * Any of the following ensures page->mem_cgroup stability:
>          *
>          * - the page lock
>          * - LRU isolation
>          * - lock_page_memcg()
>          * - exclusive reference
> 
> "LRU isolation" used to be quite clear, but now is it after TestClearPageLRU(page) or after deleting from the lru list as well?
> Also it doesn't mention rcu_read_lock(), should it?

The lru isolation still is same as old conception, a set actions that take a page from a lru list, and commit_charge do
need a isoltion for the page.

but the condition of page_memcg could be change since we don't rely on lru isolation for it. The comments
could be changed later.

> 
> So what exactly are we protecting by rcu_read_lock() in e.g. lock_page_lruvec()?
> 
>         rcu_read_lock();
>         lruvec = mem_cgroup_page_lruvec(page, pgdat);
>         spin_lock(&lruvec->lru_lock);
>         rcu_read_unlock();
> 
> Looks like we are protecting the lruvec from going away and it can't go away anymore after we take the lru_lock?
> 
> But then e.g. in __munlock_pagevec() we are doing this without an rcu_read_lock():
> 
>     new_lruvec = mem_cgroup_page_lruvec(page, page_pgdat(page));

TestClearPageLRU could block the page from memcg migration/destory.

Thanks
Alex

> 
> where new_lruvec is potentionally not the one that we have locked
> 
> And the last thing mem_cgroup_page_lruvec() is doing is:
> 
>         if (unlikely(lruvec->pgdat != pgdat))
>                 lruvec->pgdat = pgdat;
>         return lruvec;
> 
> So without the rcu_read_lock() is this potentionally accessing the pgdat field of lruvec that might have just gone away?
> 
> Thanks,
> Vlastimil