Re: [PATCH v2] [RFC] mm: fs: Invalidate BH LRU during page migration

[Chris Goldsworthy]

On 2021-02-11 06:09, Matthew Wilcox wrote:
> On Wed, Feb 10, 2021 at 09:35:40PM -0800, Chris Goldsworthy wrote:
> > +/* These are used to control the BH LRU invalidation during page 
> > migration */
> > +static struct cpumask lru_needs_invalidation;
> > +static bool bh_lru_disabled = false;
>
> As I asked before, what protects this on an SMP system?

Sorry Matthew, I misconstrued your earlier question in V1, and thought 
you had been referring to compile-time protection (so as to prevent 
build breakages).  It is not protected, so I'll need to change this into 
an atomic counter that is incremented and decremented by bh_lru_enable() 
and bh_lru_disable() respectively (such that if the counter is greater 
than zero, we bail).

> > @@ -1292,7 +1296,9 @@ static inline void check_irqs_on(void)
> >  /*
> >   * Install a buffer_head into this cpu's LRU.  If not already in the 
> > LRU, it is
> >   * inserted at the front, and the buffer_head at the back if any is 
> > evicted.
> > - * Or, if already in the LRU it is moved to the front.
> > + * Or, if already in the LRU it is moved to the front. Note that if 
> > LRU is
> > + * disabled because of an ongoing page migration, we won't insert bh 
> > into the
> > + * LRU.
>
> And also, why do we need to do this?  The page LRU has no equivalent
> mechanism to prevent new pages being added to the per-CPU LRU lists.
> If a BH has just been used, isn't that a strong hint that this page is
> a bad candidate for migration?

I had assumed that up until now, that pages in the page cache aren't an 
issue, such that they're dropped during migration as needed. Looking at 
try_to_free_buffers[1], I don't see any handling for the page cache.  I 
will need to do due diligence and follow up on this.

As for the question on necessity, if there is a case in which preventing 
buffer_heads from being added to the BH LRU ensures that the containing 
page can be migrated, then I would say that the change is justified, 
since adds another scenario in which migration is guaranteed (I will 
follow up on this as well).

Regards,

Chris.

[1] https://elixir.bootlin.com/linux/latest/source/fs/buffer.c#L3225

--
The Qualcomm Innovation Center, Inc.
The Qualcomm Innovation Center, Inc. is a member of the Code Aurora 
Forum,
a Linux Foundation Collaborative Project