Date: Thu, 17 Nov 2016 15:00:26 -0500
To: David Howells <dhowells@redhat.com>
Cc: One Thousand Gnomes <gnomes@lxorguk.ukuu.org.uk>,
        linux-fsdevel@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [RFC][PATCH 0/4] Enhanced file stat system call
Message-ID: <20161117200026.GE20937@fieldses.org>
References: <20161117143904.4ae757b4@lxorguk.ukuu.org.uk>
 <147938969703.13574.10295364502230379833.stgit@warthog.procyon.org.uk>
 <15516.1479401145@warthog.procyon.org.uk>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <15516.1479401145@warthog.procyon.org.uk>
User-Agent: Mutt/1.5.21 (2010-09-15)
From: bfields@fieldses.org (J. Bruce Fields)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 3267
Lines: 76

On Thu, Nov 17, 2016 at 04:45:45PM +0000, David Howells wrote:
> One Thousand Gnomes <gnomes@lxorguk.ukuu.org.uk> wrote:
> 
> > >  (2) Lightweight stat (AT_STATX_DONT_SYNC): Ask for just those details of
> > >      interest, and allow a network fs to approximate anything not of
> > >      interest, without going to the server.
> > > 
> > >  (3) Heavyweight stat (AT_STATX_FORCE_SYNC): Force a network fs to flush
> > >      buffers and go to the server, even if it thinks its cached attributes
> > >      are up to date.
> > 
> > That seems an odd way to do it. Wouldn't it be cleaner and more flexible
> > to give a timestamp of the oldest time you consider acceptable (and
> > obviously passing 0 indicates whatever you have)
> 
> Perhaps, though adding 6-argument syscalls is apparently frowned upon.
> 
> > > Note that no lstat() equivalent is required as that can be implemented
> > > through statx() with atflag == 0.  There is also no fstat() equivalent as
> > > that can be implemented through statx() with filename == NULL and the
> > > relevant fd passed as dfd.
> > 
> > and dfd + a name gives you fstatat() ?
> 
> Yes.
> 
> > The cover note could be clearer on this.
> 
> Fixed.
> 
> > Should the fields really be split the way they are for times rather than
> > a struct for each one so you can write code generically to handle one of
> > those rather than having to have a 4 way switch statement all the time.
> 
> It depends.  Doing so leaves 16 bytes of hole in the structure.  I could
> ameliorate the wastage by using a union to overlay useful fields in the gaps,
> but that's pretty icky and might be compiler dependent.
> 
> > Another attribute that would be nice (but migt need some trivial device
> > layer tweaking) would be STATX_ATTR_VOLATILE for filesystems that will
> > probably evaporate on a reboot. That's useful information for tools like
> > installers and also for sanity checking things like backup paths.
> 
> There's a FILE_ATTRIBUTE_TEMPORARY that I could map for windows filesystems
> that could be used with this.
> 
> > Remote needs to have clear semantics: is ext4fs over nbd 'remote' for
> > example ?
> 
> Hmmm... Interesting question.  Probably should.  But you could be insane and
> RAID an nbd and a local disk.  Further, does NFS over a loopback device to
> nfsd on the same machine qualify as root?  What if that's exposing a local fs
> on NBD?  Perhaps I should drop 'REMOTE' for now.  It sounds like something
> that a GUI filemanager might find interesting, though.

Sorry, I haven't been paying attention, just popping up for this, but:
"shared" might be a more useful term than "remote".

A filesystem that may be mounted from more than one system is "shared".
Caching performance and semantics of such a filesystem are more
complicated since the filesystem may change out from under us.  This is
what makes e.g. the lightweight/heavyweight stat difference more
interesting in the shared case.

The filesystem should be able to make that shared/unshared distinction
without knowledge of the storage it's sitting on top of.

Answering your questions by that criterion:

	- ext4/nbd: not shared
	- nfs/lo: shared

But, it's fine with me to drop any features for now as long as we can
always add them later.

--b.