Re: [RFC PATCH 0/5] kmalloc-reclaimable caches

[Vlastimil Babka]

On 05/24/2018 05:32 PM, Johannes Weiner wrote:
> On Thu, May 24, 2018 at 01:00:06PM +0200, Vlastimil Babka wrote:
>> - the vmstat/meminfo counter name is rather general and might suggest it also
>>   includes reclaimable page caches, which it doesn't
>>
>> Suggestions welcome for all three points. For the last one, we might also keep
>> the counter separate from nr_slab_reclaimable, not superset. I did a superset
>> as IIRC somebody suggested that in the older threads or at LSF.
> 
> Yeah, the "reclaimable" name is too generic. How about KReclaimable?
> 
> The counter being a superset sounds good to me. We use this info for
> both load balancing and manual debugging. For load balancing code it's
> nice not having to worry about finding all the counters that hold
> reclaimable memory depending on kernel version; it's always simply
> user cache + user anon + kernel reclaimable. And for debugging, we can
> always add more specific subset counters later on if we need them.

Hm, Christoph in his reply to patch 4/5 expressed a different opinion.
It's true that updating two counters has extra overhead, especially if
there are two separate critical sections:

mod_lruvec_page_state(page, NR_SLAB_RECLAIMABLE, nr_pages);
mod_node_page_state(page_pgdat(page), NR_RECLAIMABLE, nr_pages);

The first disables irq for CONFIG_MEMCG or defers to
mod_node_page_state() otherwise.
mod_node_page_state() is different depending on CONFIG_SMP and
CONFIG_HAVE_CMPXCHG_LOCAL.

I don't see an easy way to make this optimal? Different counter would be
indeed simpler. /proc/vmstat would then print separate counters, but we
could have both separate and summary counter in /proc/meminfo. Would
that be enough?