Re: [PATCH v18 00/32] per memcg lru_lock: reviews

From: Hugh Dickins
Date: Wed Sep 09 2020 - 22:42:19 EST

On Wed, 9 Sep 2020, Alex Shi wrote:
> 在 2020/9/9 上午7:41, Hugh Dickins 写道:
> >
> > [PATCH v18 05/32] mm/thp: remove code path which never got into
> > This is a good simplification, but I see no sign that you understand
> > why it's valid: it relies on lru_add_page_tail() being called while
> > head refcount is frozen to 0: we would not get this far if someone
> > else holds a reference to the THP - which they must hold if they have
> > isolated the page from its lru (and that's true before or after your
> > per-memcg changes - but even truer after those changes, since PageLRU
> > can then be flipped without lru_lock at any instant): please explain
> > something of this in the commit message.
> Is the following commit log better?
> split_huge_page() will never call on a page which isn't on lru list, so
> this code never got a chance to run, and should not be run, to add tail
> pages on a lru list which head page isn't there.
> Hugh Dickins' mentioned:
> The path should never be called since lru_add_page_tail() being called
> while head refcount is frozen to 0: we would not get this far if someone
> else holds a reference to the THP - which they must hold if they have
> isolated the page from its lru.
> Although the bug was never triggered, it'better be removed for code
> correctness, and add a warn for unexpected calling.

Not much better, no. split_huge_page() can easily be called for a page
which is not on the lru list at the time, and I don't know what was the
bug which was never triggered. Stick with whatever text you end up with
for the combination of 05/32 and 18/32, and I'll rewrite it after.

> > [PATCH v18 06/32] mm/thp: narrow lru locking
> > Why? What part does this play in the series? "narrow lru locking" can
> > also be described as "widen page cache locking":
> Uh, the page cache locking isn't widen, it's still on the old place.

I'm not sure if you're joking there. Perhaps just a misunderstanding.

Yes, patch 06/32 does not touch the xa_lock(&mapping->i_pages) and
xa_lock(&swap_cache->i_pages) lines (odd how we've arrived at two of
those, but please do not get into cleaning it up now); but it removes
the spin_lock_irqsave(&pgdata->lru_lock, flags) which used to come
before them, and inserts a spin_lock(&pgdat->lru_lock) after them.

You call that narrowing the lru locking, okay, but I see it as also
pushing the page cache locking outwards: before this patch, page cache
lock was taken inside lru_lock; after this patch, page cache lock is
taken outside lru_lock. If you cannot see that, then I think you
should not have touched this code at all; but it's what we have
been testing, and I think we should go forward with it.

> > But I wish you could give some reason for it in the commit message!
> It's a head scratch task. Would you like to tell me what's detailed info
> should be there? Thanks!

So, you don't know why you did it either: then it will be hard to
justify. I guess I'll have to write something for it later. I'm
strongly tempted just to drop the patch, but expect it will become
useful later, for using lock_page_memcg() before getting lru_lock.

> > Signed-off-by: Wei Yang <richard.weiyang@xxxxxxxxx>
> > Is that correct? Or Wei Yang suggested some part of it perhaps?
> Yes, we talked a lot to confirm the locking change is safe.

Okay, but the patch was written by you, and sent by you to Andrew:
that is not a case for "Signed-off-by: Someone Else".

> > [PATCH v18 27/32] mm/swap.c: optimizing __pagevec_lru_add lru_lock
> > Could we please drop this one for the moment? And come back to it later
> > when the basic series is safely in. It's a good idea to try sorting
> > together those pages which come under the same lock (though my guess is
> > that they naturally gather themselves together quite well already); but
> > I'm not happy adding 360 bytes to the kernel stack here (and that in
> > addition to 192 bytes of horrid pseudo-vma in the shmem swapin case),
> > though that could be avoided by making it per-cpu. But I hope there's
> > a simpler way of doing it, as efficient, but also useful for the other
> > pagevec operations here: perhaps scanning the pagevec for same page->
> > mem_cgroup (and flags node bits), NULLing entries as they are done.
> > Another, easily fixed, minor defect in this patch: if I'm reading it
> > right, it reverses the order in which the pages are put on the lru?
> this patch could give about 10+% performance gain on my multiple memcg
> readtwice testing. fairness locking cost the performance much.

Good to know, should have been mentioned. s/fairness/Repeated/

But what was the gain or loss on your multiple memcg readtwice
testing without this patch, compared against node-only lru_lock?
The 80% gain mentioned before, I presume. So this further
optimization can wait until the rest is solid.

> I also tried per cpu solution but that cause much trouble of per cpu func
> things, and looks no benefit except a bit struct size of stack, so if
> stack size still fine. May we could use the solution and improve it better.
> like, functionlize, fix the reverse issue etc.

I don't know how important the stack depth consideration is nowadays:
I still care, maybe others don't, since VMAP_STACK became an option.

Yes, please fix the reversal (if I was right on that); and I expect
you could use a singly linked list instead of the double.

But I'll look for an alternative - later, once the urgent stuff
is completed - and leave the acks on this patch to others.