[dm-devel] Fwd: Re: [PATCH] dm-mpath: push back requests instead of queueing

[Hannes Reinecke]

On 11/11/2013 12:25 PM, Steffen Maier wrote:
> 
> Regarding the general approach: Yes, please!
> 
> I was pondering on the very problematic memory pressure issues during times when there is no
> working path in any of the prio groups and I came to exactly the
same conclusion: dm-mpath
> should not pretend to be an arbitrarily fast block device and
consume requests from its block
> queue despite knowing that it cannot serve those requests any time
soon.
> 
Precisely.

> Requeuing sounds good and I hope this helps us getting the back pressure / congestion such
> that blocking on the request queue mempool [1,2] will kick in to
prevent memory pressure on
> queue_if_no_path (for which "queue" has a slightly different
meaning with the change but I don't mind).
> 
Well, first of all we won't be accepting any more requests if the
first one is being requeued; blktrace shows a nice requeue game
going on (even with dd and O_DIRECT set) with just a single request
being requeued over and over again, until the 'queue_if_no_path'
scenario is gone.

> One thing I'm still wondering: Would there be any benefit of actually stopping the request
> queue until at least one path becomes available again?
[blk_{start|stop}_queue()] I.e. stop
> in map_io() after we're sure there is no path in any prio group,
and restart in reinstate_path().
> 
Hehe. Thought about that, too.

Problem with that approach is the way multipath currently works.
Currently multipath does _not_ have a flag for 'all paths down and
queue_if_no_path is active, please requeue'.
It rather evaluates all possible paths during map_io, and only
if it determines that no paths are present and queue_if_no_path
is set it'll requeue the I/O.

So if we were to use start/stop queue here the block layer would
never trigger 'map_io', and multipath would never check the path
states and no I/O will be sent, ever.

blk_start/stop_queue works best if you have _alternative_ means
of setting those, besides the normal I/O path.
(It's original idea was to be used from LLDDs, after all).
When one of these functions is being used in the normal I/O path
things are becoming iffy really fast.

That said, it _would_ make sense to use blk_start/stop_queue for
pg_init; we cannot send I/O during pg_init anyway, so there's
no point in retrying I/O here.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		      zSeries & Storage
hare at suse.de			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: J. Hawn, J. Guild, F. Imendörffer, HRB 16746 (AG Nürnberg)