[LSF/MM TOPIC] block, dm: restack queue_limits

[LSF/MM TOPIC] block, dm: restack queue_limits

[Mike Snitzer]

We currently don't restack the queue_limits if the lowest, or
intermediate, layer of an IO stack changes.

This is particularly unfortunate in the case of FLUSH/FUA which may
change if/when a HW controller's BBU fails; whereby requiring the device
advertise that it has a volatile write cache (WCE=1).

But in the context of DM, really it'd be best if the entire stack of
devices had their limits restacked if any underlying layer's limits
change.

In the past, Martin and I discussed that we should "just do it" but
never did.  Not sure we need a lengthy discussion but figured I'd put it
out there.

Maybe I'll find time, between now and April, to try implementing it.

Thanks,
Mike

Re: [dm-devel] [LSF/MM TOPIC] block, dm: restack queue_limits

[Hannes Reinecke]

On 01/29/2018 10:08 PM, Mike Snitzer wrote:
> We currently don't restack the queue_limits if the lowest, or
> intermediate, layer of an IO stack changes.
> 
> This is particularly unfortunate in the case of FLUSH/FUA which may
> change if/when a HW controller's BBU fails; whereby requiring the device
> advertise that it has a volatile write cache (WCE=1).
> 
Uh-oh. Device rescan.
Would be a valid topic on its own...

> But in the context of DM, really it'd be best if the entire stack of
> devices had their limits restacked if any underlying layer's limits
> change.
> 
> In the past, Martin and I discussed that we should "just do it" but
> never did.  Not sure we need a lengthy discussion but figured I'd put it
> out there.
> 
> Maybe I'll find time, between now and April, to try implementing it.
> 
For SCSI the device capabilities are pretty much set in stone after the
initial scan; there are hooks for rescanning, but they will only work
half of the time.
Plus we can't really change the device type on the fly (eg if the SCSI
device type changes; if it moves away from '0' we would need to unbind
the sd driver, and if it moves to '0' we'll need to rescan the sd
device. None of this is happening right now.)

So I'd be glad to have a discussion around this.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		   Teamlead Storage & Networking
hare@suse.de			               +49 911 74053 688
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: F. Imendörffer, J. Smithard, J. Guild, D. Upmanyu, G. Norton
HRB 21284 (AG Nürnberg)

Re: [dm-devel] [LSF/MM TOPIC] block, dm: restack queue_limits

[Ewan D. Milne]

On Tue, 2018-01-30 at 16:07 +0100, Hannes Reinecke wrote:
> On 01/29/2018 10:08 PM, Mike Snitzer wrote:
> > We currently don't restack the queue_limits if the lowest, or
> > intermediate, layer of an IO stack changes.
> > 
> > This is particularly unfortunate in the case of FLUSH/FUA which may
> > change if/when a HW controller's BBU fails; whereby requiring the device
> > advertise that it has a volatile write cache (WCE=1).
> > 
> Uh-oh. Device rescan.
> Would be a valid topic on its own...
> 
> > But in the context of DM, really it'd be best if the entire stack of
> > devices had their limits restacked if any underlying layer's limits
> > change.
> > 
> > In the past, Martin and I discussed that we should "just do it" but
> > never did.  Not sure we need a lengthy discussion but figured I'd put it
> > out there.
> > 
> > Maybe I'll find time, between now and April, to try implementing it.
> > 
> For SCSI the device capabilities are pretty much set in stone after the
> initial scan; there are hooks for rescanning, but they will only work
> half of the time.
> Plus we can't really change the device type on the fly (eg if the SCSI
> device type changes; if it moves away from '0' we would need to unbind
> the sd driver, and if it moves to '0' we'll need to rescan the sd
> device. None of this is happening right now.)
> 
> So I'd be glad to have a discussion around this.

At least array vendor has also desired the ability to change various
attributes of the device after the initial scan, such as the model name.
Not sure what would break if we did this, since who knows what userspace
software might be caching this info, but...

-Ewan

Re: [dm-devel] [LSF/MM TOPIC] block, dm: restack queue_limits

[NeilBrown]

[-- Attachment #1: Type: text/plain, Size: 1311 bytes --]

On Mon, Jan 29 2018, Mike Snitzer wrote:

> We currently don't restack the queue_limits if the lowest, or
> intermediate, layer of an IO stack changes.
>
> This is particularly unfortunate in the case of FLUSH/FUA which may
> change if/when a HW controller's BBU fails; whereby requiring the device
> advertise that it has a volatile write cache (WCE=1).
>
> But in the context of DM, really it'd be best if the entire stack of
> devices had their limits restacked if any underlying layer's limits
> change.
>
> In the past, Martin and I discussed that we should "just do it" but
> never did.  Not sure we need a lengthy discussion but figured I'd put it
> out there.

So much "yes"!!
Just a notifier chain would probably do.

I would like the notification to support changing the size of the device
too.
I see this as being two-stage.
1/ I'm going to change the device size to X - are you all OK with that?
2/ Device size is now X.

That allows md and dm to check that filesystems aren't going to get mad
when devices are made smaller, and can adjust (if they want to) when
devices get bigger.

Thanks,
NeilBrown

>
> Maybe I'll find time, between now and April, to try implementing it.
>
> Thanks,
> Mike
>
> --
> dm-devel mailing list
> dm-devel@redhat.com
> https://www.redhat.com/mailman/listinfo/dm-devel

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 832 bytes --]