Re: file metadata via fs API

From: Karel Zak
Date: Thu Aug 13 2020 - 06:37:00 EST


On Wed, Aug 12, 2020 at 12:50:28PM -0700, Linus Torvalds wrote:
> Convince me otherwise. AGAIN. This is the exact same issue I had with
> the notification queues that I really wanted actual use-cases for, and
> feedback from actual outside users.

I thought (in last 10 years) we all agree that /proc/self/mountinfo is
the expensive, ineffective and fragile way how to deliver information to
userspace.

We have systems with thousands of mountpoints and compose mountinfo in
kernel and again parse it in userspace takes time and it's strange if
you need info about just one mountpoint.

Unfortunately, the same systems modify the huge mount table extremely
often, because it starts/stops large number of containers and every
container means a mount operation(s).

In this crazy environment, we have userspace tools like systemd or udisk
which react to VFS changes and there is no elegant way how to get
details about a modified mount node from kernel.

And of course we already have negative feedback from users who
maintain large systems -- mountinfo returns inconsistent data if you
read it by more read() calls (hopefully fixed by recent Miklos'
mountinfo cursors); system is pretty busy to compose+parse mountinfo,
etc.

> I really think this is engineering for its own sake, rather than
> responding to actual user concerns.

We're too old and too lazy for "engineering for its own sake" :-)
there is pressure from users ...

Maybe David's fsinfo() sucks, but it does not mean that
/proc/self/mountinfo is something cool. Right?

We have to dig deep grave for /proc/self/mountinfo ...

Karel

--
Karel Zak <kzak@xxxxxxxxxx>
http://karelzak.blogspot.com