[QUESTION] balloon page isolation needs LRU lock?

[QUESTION] balloon page isolation needs LRU lock?

[Joonsoo Kim]

Hello, Rafael.

I looked at some compaction code and found that some oddity about
balloon compaction. In isolate_migratepages_range(), if we meet
!PageLRU(), we check whether this page is for balloon compaction.
In this case, code needs locked. Is the lock really needed? I can't find
any relationship between balloon compaction and LRU lock.

Second question is that in above case if we don't hold a lock, we
skip this page. I guess that if we meet balloon page repeatedly, there
is no change to run isolation. Am I missing?

Please let me know what I am missing.

Thanks in advance.

Re: [QUESTION] balloon page isolation needs LRU lock?

[Rafael Aquini]

On Fri, Dec 06, 2013 at 05:53:31PM +0900, Joonsoo Kim wrote:
> Hello, Rafael.
> 
> I looked at some compaction code and found that some oddity about
> balloon compaction. In isolate_migratepages_range(), if we meet
> !PageLRU(), we check whether this page is for balloon compaction.
> In this case, code needs locked. Is the lock really needed? I can't find
> any relationship between balloon compaction and LRU lock.
> 
> Second question is that in above case if we don't hold a lock, we
> skip this page. I guess that if we meet balloon page repeatedly, there
> is no change to run isolation. Am I missing?
> 
> Please let me know what I am missing.
> 
> Thanks in advance.

Howdy Joonsoo, thanks for your question.

The major reason I left the 'locked' case in place when isolating balloon pages
was to keep consistency with the other isolation cases. Among all page types we
isolate for compaction balloon pages are an exception as, you noticed, they're
not on LRU lists. So, we (specially) fake balloon pages as LRU to isolate/compact them, 
withouth having to sort to drastic surgeries into kernel code to implement
exception cases for isolating/compacting balloon pages.

As others pages we isolate for compaction are isolated while holding the
zone->lru_lock, I left the same condition placed for balloon pages as a
safeguard for consistency. If we hit a balloon page while scanning page blocks
and we do not have the lru lock held, then the balloon page will be treated 
by the scanning mechanism just as what it is: a !PageLRU() case, and life will
go on as described by the algorithm.

OTOH, there's no direct relationship between the balloon page and the LRU lock,
other than this consistency one I aforementioned. I've never seen any major
trouble on letting the lock requirement in place during my tests on workloads
that mix balloon pages and compaction. However, if you're seeing any trouble and
that lru lock requirement is acting as an overkill or playing a bad role on your
tests, you can get rid of it easily, IMHO.

Regards,
-- Rafael

Re: [QUESTION] balloon page isolation needs LRU lock?

[Joonsoo Kim]

On Fri, Dec 06, 2013 at 10:21:43AM -0200, Rafael Aquini wrote:
> On Fri, Dec 06, 2013 at 05:53:31PM +0900, Joonsoo Kim wrote:
> > Hello, Rafael.
> > 
> > I looked at some compaction code and found that some oddity about
> > balloon compaction. In isolate_migratepages_range(), if we meet
> > !PageLRU(), we check whether this page is for balloon compaction.
> > In this case, code needs locked. Is the lock really needed? I can't find
> > any relationship between balloon compaction and LRU lock.
> > 
> > Second question is that in above case if we don't hold a lock, we
> > skip this page. I guess that if we meet balloon page repeatedly, there
> > is no change to run isolation. Am I missing?
> > 
> > Please let me know what I am missing.
> > 
> > Thanks in advance.
> 
> Howdy Joonsoo, thanks for your question.
> 
> The major reason I left the 'locked' case in place when isolating balloon pages
> was to keep consistency with the other isolation cases. Among all page types we
> isolate for compaction balloon pages are an exception as, you noticed, they're
> not on LRU lists. So, we (specially) fake balloon pages as LRU to isolate/compact them, 
> withouth having to sort to drastic surgeries into kernel code to implement
> exception cases for isolating/compacting balloon pages.
> 
> As others pages we isolate for compaction are isolated while holding the
> zone->lru_lock, I left the same condition placed for balloon pages as a
> safeguard for consistency. If we hit a balloon page while scanning page blocks
> and we do not have the lru lock held, then the balloon page will be treated 
> by the scanning mechanism just as what it is: a !PageLRU() case, and life will
> go on as described by the algorithm.
> 
> OTOH, there's no direct relationship between the balloon page and the LRU lock,
> other than this consistency one I aforementioned. I've never seen any major
> trouble on letting the lock requirement in place during my tests on workloads
> that mix balloon pages and compaction. However, if you're seeing any trouble and
> that lru lock requirement is acting as an overkill or playing a bad role on your
> tests, you can get rid of it easily, IMHO.

Hello, Rafael.

Thanks for nice explanation.
Now I totally understand what it means and why it does.

Thanks!