Re: [ckrm-tech] [PATCH 0/5] Allow more than PAGESIZE data read in configfs

[Matt Helsley <matthltc@us.ibm.com>]

On Fri, 2006-10-13 at 17:09 -0700, Joel Becker wrote:
> On Fri, Oct 13, 2006 at 04:37:38PM -0700, Matt Helsley wrote:
> > > 	Sure it works.  You have one per resource group.  In
> > > resource_group_make_object(), you sysfs_mkdir() the sysfs file.  There
> > 
> > 	That's the easy part. Next we need to make the pid attribute whenever a
> > new task is created. And delete it when the task dies. And move it
> > around whenever it changes groups. Is there rename() support in /sys? If
> > not, would changes to allow rename() be acceptable (I'm worried it would
> > impact alot of assumptions made in the existing code)?
> 
> 	No, you don't create a pid attribute per task.  The sysfs file
> is literally your large attribute.  So, instead of echoing a new pid to
> "/sys/kernel/config/ckrm/group1/pids", you echo to
> "/sys/ckrm/group1/pids".  To display them all, you just cat
> "/sys/ckrm/group1/pids".  It's exactly like the file you want in
> configfs, just located in a place where it is allowed.

Oh, sorry. I was still operating on the one-value-per-attribute
assumption. This indeed looks like it would work.

> > 	Consider that having two very similar (but not symlinked!) trees in
> > both /sys/ ... /res_group and /sys/kernel/config/res_group could be
> > rather confusing to userspace programmers and users alike.
> 
> 	Not really.  It's not identical (tons of attributes live in the
> configfs part but not the sysfs part), and it has a clear deliniation of
> what each does.

	Clear delineation to who? I'm not convinced this is any less confusing
to a userspace programmer than parsing a single newline between multiple
values in a configfs attribute.

> > 	It would be strange because when you rmdir a group
> > in /sys/kernel/config/res_group... a directory in /sys would also
> > disappear. Yet you can't mkdir or rmdir the /sys dirs. And to edit the
> 
> 	This is no different than tons of sysfs and procfs functionality
> today.

Yup.

> > 	There are two parts to the complexity: code complexity and the number
> > of userspace pieces to deal with. I think that in both of these
> > categories the OVPA approach is more complex. Here's how I see it:
> 
> 	By your definition, sysfs, configfs, and other fs-style control
> mechanisms are too complex.  We should all just be using ioctl() so that
> coders and users have only one namespace :-)

That's an absurd conclusion to draw from my argument that one
filesystem-based approach is less complex than another.

> > > 	You're effectively suggesting that a specific attribute type of
> > > "repeated value of type X".  No mixed types, no exploded structures,
> > > just a "list of this attr" sort of thing.  This does fit my personal
> > > requirement of avoiding a generic, abusable system.
> > 
> > Exactly.
> 
> 	How do you implement it?  Full on seq_file with restrictions
> (ops->start,stop,next,show)?

That was the plan.

Cheers,
	-Matt Helsley