Re: [RFC PATCH] multipathd: use 0 for initial pg if nr_priority_groups=0

[Hannes Reinecke <hare@suse.de>]

On 02/10/2011 11:00 PM, Mike Snitzer wrote:
> On Tue, Feb 01 2011 at 10:38am -0500,
> Hannes Reinecke <hare@suse.de> wrote:
> 
>> On 01/31/2011 07:34 PM, Mike Snitzer wrote:
>>> On Mon, Jan 31 2011 at  1:09pm -0500,
>>> Moger, Babu <Babu.Moger@lsi.com> wrote:
>>>
>>>> Looks good to me. 
>>>>
>>>>> -----Original Message-----
>>>>> From: Mike Snitzer [mailto:snitzer@redhat.com]
>>>>> Sent: Monday, January 31, 2011 10:14 AM
>>>>> To: dm-devel@redhat.com
>>>>> Cc: Mike Snitzer; Moger, Babu
>>>>> Subject: [PATCH 2/2] dm mpath: allow table load with 0 priority groups
>>>>>
>>>>> If an mpath device is held open when all paths in the last priority
>>>>> group have failed userspace multipathd will attempt to reload the
>>>>> associated DM table to reflect that the device no longer has any
>>>>> priority groups.  But the reload attempt always failed because the
>>>>> multipath target did not allow 0 priority groups.
>>>>>
>>>>> Adjust multipath target to allow a table with both 0 priority groups
>>>>> and 0 for the initial priority group number.
>>>>>
>>>>> All multipath target messages related to priority group (enable_group,
>>>>> disable_group, switch_group) will properly handle a priority group of
>>>>> 0 (will cause error).
>>>>>
>>>>> When reloading a multipath table with 0 priority groups, userspace
>>>>> multipathd must be updated to specify an initial priority group number
>>>>> of 0 (rather than 1).
>>>>
>>>>  Looks like we still have some action from multipath tool. CCing Ben..
>>>
>>> Right, I looked in to the multipath-tools change.  Here is an RFC patch.
>>>
>>> Inlined "FIXME"s pose the relevant questions.  I held off on having
>>> select_path_group() return 0 when !mpp->pg because I wasn't confident
>>> that I wouldn't break multipathd in some unintuitive way.
>>>
>>> I did audit the various callers (indirect through setup_map call):
>>>
>>> update_path_groups
>>> ev_add_path
>>> ev_remove_path
>>>
>>> (all above first call setup_map)
>>>
>>> setup_map()
>>>   - mpp->bestpg = select_path_group()
>>>   - assemble_map() -- establishes mp->params
>>>
>>> do_map()
>>>
>>> Even though I verified that much I wasn't completely confident in
>>> changing select_path_group()'s defualt return of 1 -- so I localized the
>>> change to where I _knew_ it was safe (assemble_map):
>>>
>>> diff --git a/libmultipath/dmparser.c b/libmultipath/dmparser.c
>>> index 1ef3aad..97b6420 100644
>>> --- a/libmultipath/dmparser.c
>>> +++ b/libmultipath/dmparser.c
>>> @@ -52,6 +52,7 @@ assemble_map (struct multipath * mp)
>>>  	int i, j;
>>>  	int shift, freechar;
>>>  	int minio;
>>> +	int nr_priority_groups, initial_pg_nr;
>>>  	char * p;
>>>  	struct pathgroup * pgp;
>>>  	struct path * pp;
>>> @@ -60,9 +61,19 @@ assemble_map (struct multipath * mp)
>>>  	p = mp->params;
>>>  	freechar = sizeof(mp->params);
>>>  
>>> +	nr_priority_groups = VECTOR_SIZE(mp->pg);
>>> +	/*
>>> +	 * FIXME: make this conditional on multipath target version?
>>> +	 * - but no existing code is conditional on multipath target version,
>>> +	 *   nor does dm_drvprereq() store the version for code to do so.
>>> +	 * - thing is older multipath never allowed nr_priority_groups=0 so
>>> +	 *   it doesn't _really_ matter if initial_pg_nr=0 here...
>>> +	 */
>>> +	initial_pg_nr = (nr_priority_groups ? mp->bestpg : 0);
>>> +
>>>  	shift = snprintf(p, freechar, "%s %s %i %i",
>>>  			 mp->features, mp->hwhandler,
>>> -			 VECTOR_SIZE(mp->pg), mp->bestpg);
>>> +			 nr_priority_groups, initial_pg_nr);
>>>  
>>>  	if (shift >= freechar) {
>>>  		fprintf(stderr, "mp->params too small\n");
>>> diff --git a/libmultipath/switchgroup.c b/libmultipath/switchgroup.c
>>> index 025a95d..82c30ad 100644
>>> --- a/libmultipath/switchgroup.c
>>> +++ b/libmultipath/switchgroup.c
>>> @@ -41,7 +41,7 @@ select_path_group (struct multipath * mpp)
>>>  	struct pathgroup * pgp;
>>>  
>>>  	if (!mpp->pg)
>>> -		return 1;
>>> +		return 1; /* FIXME: return 0 here? */
>>>  
>>>  	vector_foreach_slot (mpp->pg, pgp, i) {
>>>  		if (!pgp->paths)
>>>
>> I have a similar patchset in my multipath-tools tree, to allow maps
>> with zero paths, too.
>> However, there are quite a few places which required checking, as
>> we're basically _never_ check if a pointer is NULL before accessing
>> it. With a bit of luck I'll find some time to send out the patchset.
>>
>> Otherwise you could have a look at
>>
>> git://git.kernel.org/pub/scm/linux/kernel/git/hare/multipath-tools.git
>> branch sles11-sp1 will have the most recent stuff.
> 
> I found the following commit but it doesn't have any meaningful changes:
> 1e53326 Allow zero paths for device-mapper strings
> 
> Could you be more explicit on which commits you're thinking of?

Ah, I just rechecked the code. There used to be quite a few missing
NULL pointer checks, especially if mpp->pg or mpp->paths are NULL.
But they seem to have been fixed now.

So ignore my objection here.

Cheers,

Hannes
-- 
Dr. Hannes Reinecke		      zSeries & Storage
hare@suse.de			      +49 911 74053 688
SUSE LINUX Products GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: Markus Rex, HRB 16746 (AG Nürnberg)