RE: [PATCH 2/2] Add /proc/PID/{smaps, numa_maps} support for DAX

-----Original Message-----
From: linux-kernel-owner@vger.kernel.org
[mailto:linux-kernel-owner@vger.kernel.org] On Behalf Of Dave Hansen
Sent: Thursday, October 26, 2017 10:51 PM
To: Michal Hocko <mhocko@kernel.org>
Cc: Du, Fan <redacted>; akpm@linux-foundation.org; hch@lst.de;
Williams, Dan J [off-list ref]; linux-kernel@vger.kernel.org;
linux-api@vger.kernel.org
Subject: Re: [PATCH 2/2] Add /proc/PID/{smaps, numa_maps} support for DAX

On 10/26/2017 07:31 AM, Michal Hocko wrote:

quoted

On Thu 26-10-17 07:24:14, Dave Hansen wrote:

quoted

Actually, I don't remember whether it was tooling or just confused
humans.  I *think* Dan was trying to write test cases for huge page DAX
support and couldn't figure out whether or not it was using large pages.

That sounds like a very weak justification to adding new stuff to smaps
to be honest.

Yep, agreed.  It can't go in _just_ for DAX, and Fan and the other DAX
folks need to elaborate on their needs here.

If user creates device DAX /dev/dax with some capacity like 512G, mmap it and
Use it will, or touched 128G. To my best knowledge at this part, there is no
statistics reported how much memory behind DAX actually used.

This is the problem our customer is facing right now.

I agree with smaps should make no special for DAX, whether or not we update
smaps for DAX, cat /proc/PID/smaps always report its part for /dev/dax.
The problem is should we report DAX usage by RSS, or introduce new fields
Like Pte@{4K,2M} for a another different purpose.

The v4 version tried to merged device DAX usage with RSS,
https://lkml.org/lkml/2017/10/26/24

Do you have any better ideas?  If we did this, we could probably make an
argument that the AnonHuge fields could go away some day.  They haven't
always been there.

The only other alternative I can think of are truly DAX-specific
interfaces, which also seem like a really bad idea.

quoted

Not only that. There have been reports that reading smaps is too
expensive. Curiously enough the overhead doesn't come up from
the data collection but rather copying to the userspace. So we should be
careful to not print data that is not of general use.

Yikes!  I just did a quick:

while true; do cat /proc/*/smaps | wc ; done

and the copying out to userspace is ~1/15th the overhead of
smaps_account().  Something sounds screwy if you're seeing the overhead
at copying to userspace.

What else can we do than continue to bloat smaps?  Could we do a file
per VMA?

/proc/$pid/smap/0x123000-0x456000