[dm-devel] [LSF/MM TOPIC][LSF/MM ATTEND] multipath redesign

[Mike Snitzer]

On Fri, Jan 13 2017 at 10:56am -0500,
Hannes Reinecke <hare at suse.de> wrote:

> On 01/12/2017 06:29 PM, Benjamin Marzinski wrote:
> > On Thu, Jan 12, 2017 at 09:27:40AM +0100, Hannes Reinecke wrote:
> >> On 01/11/2017 11:23 PM, Mike Snitzer wrote:
> >>> On Wed, Jan 11 2017 at  4:44am -0500,
> >>> Hannes Reinecke <hare at suse.de> wrote:
> >>>
> >>>> Hi all,
> >>>>
> >>>> I'd like to attend LSF/MM this year, and would like to discuss a
> >>>> redesign of the multipath handling.
> >>>>
> >>>> With recent kernels we've got quite some functionality required for
> >>>> multipathing already implemented, making some design decisions of the
> >>>> original multipath-tools implementation quite pointless.
> >>>>
> >>>> I'm working on a proof-of-concept implementation which just uses a
> >>>> simple configfs interface and doesn't require a daemon altogether.
> >>>>
> >>>> At LSF/MM I'd like to discuss how to move forward here, and whether we'd
> >>>> like to stay with the current device-mapper integration or move away
> >>> >from that towards a stand-alone implementation.
> >>>
> >>> I'd really like open exchange of the problems you're having with the
> >>> current multipath-tools and DM multipath _before LSF_.  Last LSF only
> >>> scratched the surface on people having disdain for the complexity that is
> >>> the multipath-tools userspace.  But considering how much of the
> >>> multipath-tools you've written I find it fairly comical that you're the
> >>> person advocating switching away from it.
> >>>
> >> Yeah, I know.
> >>
> >> But I've stared long and hard at the code, and found some issues really hard
> >> to overcome. Even more so as most things it does are really pointless.
> >>
> >> multipathd _insists_ on redoing the _entire_ device layout for basically any
> >> operation (except for path checking).
> >> As the data structures allow only for a single setup it uses a lock per
> >> multipath device to protect against concurrent changes.
> >> When lots of uevents are to be processed this lock is heavily contended,
> >> leading to a slow-down of uevent processing.
> >> (cf the patchseries from Tang Junhui and my earlier pathset for
> >> lock pushdown)
> >>
> >> I've tried to move that lock down even further with distinct locks for
> >> device paths and multipath devices, but ultimately failed as it would amount
> >> to essentially a rewrite of the core engine.
> > 
> > The multipath user-space tools locking IS horrible and touches
> > everything.  I could never see a way around it that didn't involve
> > a ground-up redesign.
> >  
> :-)
> 
> >>> But if less userspace involvement is needed then fix userspace.  Fail to
> >>> see how configfs is any different than the established DM ioctl interface.
> >>>
> >>> As I just said in another email DM multipath could benefit from
> >>> factoring out the SCSI-specific bits so that they are nicely optimized
> >>> away if using new transports (e.g. NVMEoF).
> >>>
> >>> Could be lessons can be learned from your approach but I'd prefer we
> >>> provably exhaust the utility of the current DM multipath kernel
> >>> implementation.  DM multipath is one of the most actively maintained and
> >>> updated DM targets (aside from thinp and cache).  As you know DM
> >>> multipath has grown blk-mq support which yielded serious performance
> >>> improvement.  You also noted (in an earlier email) that I reintroduced
> >>> bio-based DM multipath.  On a data path level we have all possible block
> >>> core interfaces plumbed.  And yes, they all involve cloning due to the
> >>> underlying Device Mapper core.  Open to any ideas on optimization.  If
> >>> DM is imposing some inherent performance limitation then please report
> >>> it accordingly.
> >>>
> >> Ah. And I thought you disliked request-based multipathing ...
> >>
> >> It's not _actually_ the DM interface which I'm objecting to, it's more the
> >> user-space implementation.
> >> The daemon is build around some design decisions which are simply not
> >> applicable anymore:
> >> - we now _do_ have reliable device identifications, so the the 'path_id'
> >> functionality is pointless.
> > 
> > This could be largely fixed in the existing code. The route that the
> > latest patch from Tang Junhui are going still grabs the wwid if we got
> > it from the uevent, but it isn't necesary, as long was we're careful.
> > Currently rbd devices don't get their wwid from the uevent but all other
> > devices do. It would probably be possible to write an rbd device udev
> > rule to set a variable so that they can work through udev environment
> > variables too.
> > 
> But this is still only working around the problem.
> We only should need to touch the device-mapper tables when setting up
> devices or during reconfiguration.
> 
> >> - The 'alua' device handler also provides you with reliable priority
> >> information, so it should be possible to do away with the 'prio' setting,
> >> too.
> > 
> > But this isn't true for all devices. Also, Like I mentioned last year
> > when this got brought up, no matter how we group the paths, there end up
> > being users that have good reasons why they want them grouped
> > differently in their case.  The path priority/grouping seems like one
> > place where evidence has shown that we should give users the tools to
> > make policy decisions, instead of making them ourselves.
> > 
> >> - And for (most) SCSI devices the 'state' setting provides a reliable
> >> indicator if the device is useable.
> > 
> > This is also not true for all devices.
> > 
> So? The 'state' attribute reflects the internal SCSI device state.
> If _that_ doesn't work reliably you end up with I/O errors.
> Which eventually will end up with the 'state' attribute being
> synchronized with the actual device state (or being set to 'offline').
> 
> > So, are you planning on creating a multipath implementation that only
> > handles some devices? Obviously, the current userspace tools are still
> > around to handle setups that this wouldn't.
> > 
> No, certainly not.
> ATM my implementation is merely a testbed, as new
> features/functionalities can be more easily implemented there.
> I don't see any issues with porting this to device-mapper as such.
> 
> > While I've daydreamed of rewriting the multipath tools multiple times,
> > and having nothing aginst you doing it in concept, I would be happier
> > knowing that it won't simply mean that there are two sets of tools, that
> > both need to be supported to deal with all customer configurations.
> > 
> Sure. I feel the pain of supporting multipath-tools all too strongly.
> Having two tools for the same thing is always a pain, and I would like
> to avoid this if at all possible.

I welcome your work.  Should help us focus on what fat needs to be
trimmed from both multipath-tools and kernel.

Might be a good time to branch multipath-tools and get very aggressive
with trimming stuff that is outdated.

Things like the event stuff, using select interface, that Andy Grover is
working on (and Mikulas is taking a stab at finishing/optimizing) is
something that might help... but your approach described in this thread
may prove better.

Point is, everything should be on the table for revitalizing multipath
userspace (and kernel) to meet new requirements (e.g. NVMEoF, etc).

And yes, I'd prefer to ultimately see these advances land in terms of DM
multipath advances but we'll take it as it comes.