Message-ID: <c9a6f929b57e0c21c8845c211d1e3eab09d09633.camel@themaw.net>
Date:   Thu, 20 Feb 2020 12:42:15 +0800
From:   Ian Kent <raven@...maw.net>
To:     Christian Brauner <christian.brauner@...ntu.com>,
        David Howells <dhowells@...hat.com>
Cc:     viro@...iv.linux.org.uk, mszeredi@...hat.com, christian@...uner.io,
        linux-api@...r.kernel.org, linux-fsdevel@...r.kernel.org,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH 00/19] VFS: Filesystem information and notifications
 [ver #16]

On Wed, 2020-02-19 at 15:46 +0100, Christian Brauner wrote:
> On Tue, Feb 18, 2020 at 05:04:55PM +0000, David Howells wrote:
> > Here are a set of patches that adds system calls, that (a) allow
> > information about the VFS, mount topology, superblock and files to
> > be
> > retrieved and (b) allow for notifications of mount topology
> > rearrangement
> > events, mount and superblock attribute changes and other superblock
> > events,
> > such as errors.
> > 
> > ============================
> > FILESYSTEM INFORMATION QUERY
> > ============================
> > 
> > The first system call, fsinfo(), allows information about the
> > filesystem at
> > a particular path point to be queried as a set of attributes, some
> > of which
> > may have more than one value.
> > 
> > Attribute values are of four basic types:
> > 
> >  (1) Version dependent-length structure (size defined by type).
> > 
> >  (2) Variable-length string (up to 4096, including NUL).
> > 
> >  (3) List of structures (up to INT_MAX size).
> > 
> >  (4) Opaque blob (up to INT_MAX size).
> 
> I mainly have an organizational question. :) This is a huge patchset
> with lots and lots of (good) features. Wouldn't it make sense to make
> the fsinfo() syscall a completely separate patchset from the
> watch_mount() and watch_sb() syscalls? It seems that they don't need
> to
> depend on each other at all. This would make reviewing this so much
> nicer and likely would mean that fsinfo() could proceed a little
> faster.

The remainder of the fsinfo() series would need to remain useful
if this was done.

For context I want work on improving handling of large mount
tables.

Ultimately I expect to solve a very long standing autofs problem
of using large direct mount maps without prohibitive performance
overhead (and there a lot of rather challenging autofs changes to
do for this too) and I believe the fsinfo() system call, and
related bits, is the way to do this.

But improving the handling of large mount tables for autofs
will have the side effect of improvements for other mount table
users, even in the early stages of this work.

For example I want to use this for mount table handling improvements
in libmount. Clearly that ultimately needs mount change notification
in the end but ...

There's a bunch of things that need to be done alone the way
to even get started.

One thing that's needed is the ability to call fsinfo() to get
information on a mount to avoid constant reading of the proc based
mount table, which happens a lot (since the mount info. needs
to be up to date) so systemd (and others) would see an improvement
with the fsinfo() system call alone able to be used in libmount.

But for the fsinfo() system call to be used for this the file
system specific mount options need to also be obtained when
using fsinfo(). That means the super block operation fsinfo uses
to provide this must be implemented for at least most file systems.

So separating out the notifications part, leaving whatever is needed
to still be able to do this, should be fine and the system call
would be immediately useful once the super operation is implemented
for the needed file systems.

Whether the implementation of the super operation should be done
as part of this series is another question but would certainly
be a challenge and make the series more complicated. But is needed
for the change to be useful in my case.

Ian

Powered by blists - more mailing lists