Re: Proposed enhancements to MD

[Scott Long]

Jeff Garzik wrote:
> Scott Long wrote:
>
> > I'm going to push these changes out in phases in order to keep the
>
> risk
>
> > and churn to a minimum.  The attached patch is for the partition
> > support.  It was originally from Ingo Molnar, but has changed quite a
> > bit due to the radical changes in the disk/block layer in 2.6.  The
>
> 2.4
>
> > version works quite well, while the 2.6 version is fairly fresh.  One
> > problem that I have with it is that the created partitions show up in
> > /proc/partitions after running fdisk, but not after a reboot.
>
>
> You sorta hit a bad time for 2.4 development.  Even though my employer 
> (Red Hat), Adaptec, and many others must continue to support new 
> products on 2.4.x kernels, kernel development has shifted to 2.6.x (and 
> soon 2.7.x).
>
> In general, you want a strategy of "develop on latest, then backport if 
> needed."  Once a solution is merged into the latest kernel, it 
> automatically appears in many companies' products (and perhaps more 
> importantly) product roadmaps.  Otherwise you will design various things
>
> into your software that have already been handled different in the 
> future, thus creating an automatically-obsolete solution and support 
> nightmare.

Oh, I understand completely.  This work has actually been going on for a
number of years in an on-and-off fashion.  I'm just the latest person to
pick it up, and I happened to pick it up right when the big transition
to 2.6 happened.

> Now, addressing your specific issues...
>
>
> > hile MD is fairly functional and clean, there are a number of
>
> enhancements to it that we have been working on for a while and would
>
> > like to push out to the community for review and integration.  These
> > include:
> >
> > - partition support for md devices:  MD does not support the concept
>
> of
>
> >  fdisk partitions; the only way to approximate this right now is by
> >  creating multiple arrays on the same media.  Fixing this is required
> >  for not only feature-completeness, but to allow our BIOS to
>
> recognise
>
> >  the partitions on an array and properly boot them as it would boot a
> >  normal disk.
>
>
> Neil Brown already done a significant amount of research into this 
> topic.  Given this, and his general status as md maintainer, you should 
> definitely make sure he's kept in the loop.
>
> Partitioning for md was discussed in this thread:
> http://lkml.org/lkml/2003/11/13/182
>
> In particular note Al Viro's response to Neil, in addition to Neil's own
>
> post.
>
> And I could have _sworn_ that Neil already posted a patch to do 
> partitions in md, but maybe my memory is playing tricks on me.

I thought that I had attached a patch to the end of my last mail, but I
could have messed it up.  The work to do partitioning in 2.6 looks to
be incredibly less significant than in 2.4, thankfully =-)

> > - generic device arrival notification mechanism:  This is needed to
> >  support device hot-plug, and allow arrays to be automatically
> >  configured regardless of when the md module is loaded or
>
> initialized.
>
> >  RedHat EL3 has a scaled down version of this already, but it is
> >  specific to MD and only works if MD is statically compiled into the
> >  kernel.  A general mechanism will benefit MD as well as any other
> >  storage system that wants hot-arrival notices.
>
>
> This would be via /sbin/hotplug, in the Linux world.  SCSI already does 
> this, I think, so I suppose something similar would happen for md.

A problem that we've encountered, though, is the following sequence:

1) md is inialized during boot
2) drives X Y and Z are probed during boot
3) root fs exists on array [X Y Z], but md didn't see them show up,
   so it didn't auto-configure the array

I'm not sure how this can be addressed by a userland daemon.  Remember
that we are focused on providing RAID during boot; configuring a
secondary array after boot is a much easier problem.

RHEL3 already has a mechanism to address this via the
md_autodetect_dev() hook.  This gets called by the partition code when
partition entites are discovered.  However, it is a static method, so
it only works when md is compiled into the kernel.  Our proposal to
to turn this into a generic registration mechanism, where md can
register as a listener.  When it does that, it gets a list of
previously announced devices, along with future devices as they are
discovered.

The code to do this is pretty small and simple.  The biggest question
is whether to implement it by enhancing add_partition(), or create a
new call (i.e. device_register_partition() ), like is done in RHEL3.

> > - RAID-0 fixes:  The MD RAID-0 personality is unable to perform I/O
> >  that spans a chunk boundary.  Modifications are needed so that it
>
> can
>
> >  take a request and break it up into 1 or more per-disk requests.
>
>
> I thought that raid0 was one of the few that actually did bio splitting 
> correctly?  Hum, maybe this is a 2.4-only issue.  Interesting, and 
> agreed, if so...

This is definitely still a problem in 2.6.1

> > - Metadata abstraction:  We intend to support multiple on-disk
>
> metadata
>
> >  formats, along with the 'native MD' format.  To do this, specific
> >  knowledge of MD on-disk structures must be abstracted out of the
>
> core
>
> >  and personalities modules.
>
>
> > - DDF Metadata support: Future products will use the 'DDF' on-disk
> >  metadata scheme.  These products will be bootable by the BIOS, but
> >  must have DDF support in the OS.  This will plug into the
>
> abstraction
>
> >  mentioned above.
>
>
> Neil already did the work to make 'md' support multiple types of 
> superblocks, but I'm not sure if we want to hack 'md' to support the 
> various vendor RAIDs out there.  DDF support we _definitely_ want, of 
> course.  DDF follows a very nice philosophy:  open[1] standard with no 
> vendor lock-in.
>
> IMO, your post/effort all boils down to an open design question:  device
>
> mapper or md, for doing stuff like vendor-raid1 or vendor-raid5?  And it
>
> is even possible to share (for example) raid5 engine among all the 
> various vendor RAID5's?

The stripe and parity format is not the problem here; md can be enhanced
to support different stripe and parity rotation sequences without much
trouble.

Also, think beyond just DDF.  Having plugable metadata personalities 
means that a module can be written for the existing Adaptec RAID 
products too (like the HostRAID functionality on our U320 adapters).
It also means that you can write personality modules for other vendors,
and even hardware RAID solutions.  Imagine having a PCI RAID card fail,
then plugging the drives directly into your computer and having the
array 'Just Work'.

As for the question of DM vs. MD, I think that you have to consider that
DM right now has no concept of storing configuration data on the disk
(at least that I can find, please correct me if I'm wrong).  I think
that DM will make a good LVM-like layer on top of MD, but I don't see it
replacing MD right now.


Scott

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/