Re: ZAC target (Was: Re: dm-multipath: Accept failed paths for multipath maps)
From: Hannes Reinecke
Date: Tue Jul 22 2014 - 01:47:07 EST
On 07/21/2014 09:28 PM, Kent Overstreet wrote:
On Mon, Jul 21, 2014 at 04:23:41PM +0200, Hannes Reinecke wrote:
On 07/18/2014 07:04 PM, John Utz wrote:
On 07/18/2014 05:31 AM, John Utz wrote:
Thankyou very much for the exhaustive answer! I forwarded on to my
project peers because i don't think any of us where aware of the
existing infrastructure.
Of course, said infrastructure would have to be taught about ZAC,
but it seems like it would be a nice place to start testing from....
ZAC is a different beast altogether; I've posted an initial set of
patches a while back on linux-scsi.
But I don't think multipath needs to be changed for that.
Other areas of device-mapper most certainly do.
Pretty sure John is working on a new ZAC-oriented DM target.
YUP.
Per Ted T'so's suggestion several months ago, the goal is to create
a new DM target that implements the ZAC/ZBC command set and the SMR
write pointer architecture so that FSfolksen can try their hand at
porting their stuff to it.
It's in the very early stages so there is nothing to show yet, but
development is ongoing. There are a few unknowns about how to surface
some specific behaviors (new verbs and errors, particularly errors
with sense codes that return a write pointer) but i have not gotten
far enuf along in development to be able to construct succint and
specific questions on the topic so that will have to wait for a bit.
I was pondering the 'best' ZAC implementation, too, and found the
'report zones' command _very_ cumbersome to use.
Especially the fact that in theory each zone could have a different size
_and_ plenty of zones could be present will be making zone lookup hellish.
However: it seems to me that we might benefit from a generic
'block boundaries' implementation.
Reasoning here is that several subsystems (RAID, ZAC/ZBC, and things like
referrals) impose I/O scheduling boundaries which must not be crossed when
assembling requests.
Wasn't Ted working on such a thing?
Seeing that we already have some block limitations I was wondering if we
couldn't have some set of 'I/O scheduling boundaries' as part
of the request_queue structure.
I'd prefer not to dump yet more crap in request_queue, but that's a fairly minor
quibble :)
I also tend to think having different size zones is crazy and I would avoid
making any effort to support that in practice, but OTOH there's good reason for
wanting one or two "normal" zones and the rest append only so the interface is
going to have to accomadate some differences between zones.
Also, depending on the approach supporting different size zones might not
actually be problematic. If you're starting with something that's pure COW and
you're just plugging in this "ZAC allocation" stuff (which I think is what I'm
going to do in bcache) then it might not actually be an issue.
No, what I was suggesting is to introduce 'I/O scheduling barriers'.
Some devices like RAID or indeed ZAC have internal boundaries which
cannot be crossed by any I/O. So either the I/O has to be split up
or the I/O scheduler have to be made aware of these boundaries.
I have had this issue several times now (once with implementing
Referrals, now with ZAC) so I was wondering whether we can have some
sort of generic implementation in the block layer.
And as we're already having request queue limits this might fall
quite naturally into it. Or so I thought.
Hmm. Guess I should start coding here.
Cheers,
Hannes
--
Dr. Hannes Reinecke zSeries & Storage
hare@xxxxxxx +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)
--
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/