Allow zero paths for multipath priority groups

Commit Message

Nikanth Karthikesan July 8, 2009, 8:45 a.m. UTC

From: Hannes Reinecke <[e-mail]>
Subject: Allow zero paths for multipath priority groups

For correct handling of the all-paths-down scenario we have to
allow zero paths as a valid argument for priority groups.

Signed-off-by: Hannes Reinecke <[e-mail]>
Signed-off-by: Nikanth Karthikesan <[e-mail]>

---
 drivers/md/dm-mpath.c |    4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)


--
dm-devel mailing list
[e-mail]
https://www.redhat.com/mailman/listinfo/dm-devel

Comments

Alasdair G Kergon July 20, 2009, 8:37 p.m. UTC | #1

On Wed, Jul 08, 2009 at 02:15:39PM +0530, Nikanth Karthikesan wrote:
> From: Hannes Reinecke <[e-mail]>
> Subject: Allow zero paths for multipath priority groups
> 
> For correct handling of the all-paths-down scenario we have to
> allow zero paths as a valid argument for priority groups.
 
That seems counter-intuitive.
Can you be more specific, with example scenarios?

Alasdair

--
dm-devel mailing list
[e-mail]
https://www.redhat.com/mailman/listinfo/dm-devel

Hannes Reinecke July 21, 2009, 2:46 p.m. UTC | #2

Alasdair G Kergon wrote:
> On Wed, Jul 08, 2009 at 02:15:39PM +0530, Nikanth Karthikesan wrote:
>> From: Hannes Reinecke <[e-mail]>
>> Subject: Allow zero paths for multipath priority groups
>>
>> For correct handling of the all-paths-down scenario we have to
>> allow zero paths as a valid argument for priority groups.
>  
> That seems counter-intuitive.
> Can you be more specific, with example scenarios?
> 
Currently we're not allowing the last path to be removed
from a multipath target. With earlier kernel revisions
this wasn't a problem as dev_loss_tmo didn't work correctly
and so in effect no devices were ever removed from the system.

However, with newer kernels dev_loss_tmo works correctly, so
any dead path will be removed from the system.
So when the last path goes down, we're having a stale reference
to an sdev which for all intents and purposes should have been
removed.

And even worse, when this path comes back the SCSI subsystem
isn't even able to assign it to the old device number, as this
is already taken.

Cheers,

Hannes

Alasdair G Kergon July 21, 2009, 2:55 p.m. UTC | #3

Then effectively the multipath target with no paths turns into a pure
queueing target.  I'll have a think about the options, but that's probably
quite a reasonable idea.

Alasdair

--
dm-devel mailing list
[e-mail]
https://www.redhat.com/mailman/listinfo/dm-devel

Hannes Reinecke July 22, 2009, 6:08 a.m. UTC | #4

Alasdair G Kergon wrote:
> Then effectively the multipath target with no paths turns into a pure
> queueing target.  I'll have a think about the options, but that's probably
> quite a reasonable idea.
> 
Quite so.

Cheers,

Hannes

Patch

Index: linux-2.6-dm/drivers/md/dm-mpath.c
===================================================================
--- linux-2.6-dm.orig/drivers/md/dm-mpath.c
+++ linux-2.6-dm/drivers/md/dm-mpath.c
@@ -787,8 +787,8 @@  static int multipath_ctr(struct dm_targe
 {
 	/* target parameters */
 	static struct param _params[] = {
-		{1, 1024, "invalid number of priority groups"},
-		{1, 1024, "invalid initial priority group number"},
+		{0, 1024, "invalid number of priority groups"},
+		{0, 1024, "invalid initial priority group number"},
 	};
 
 	int r;