Re: [PATCH 1/1]: scsi scsi_dh_alua: don't fail I/O until transition time expires

[Brian Bunker <brian@purestorage.com>]

> On Jun 9, 2021, at 12:03 AM, Hannes Reinecke <hare@suse.de> wrote:
> 
> On 6/8/21 2:03 AM, Brian Bunker wrote:
>>> Do not return an error to multipath which will result in a failed path until the \
>>> transition time expires.
>>> The current patch which returns BLK_STS_AGAIN for ALUA transitioning breaks the \
>>> assumptions in our target regarding ALUA states. With that change an error is very \
>>> quickly returned to multipath which in turn immediately fails the path. The \
>>> assumption in that patch seems to be that another path will be available for \
>>> multipath to use. That assumption I don't believe is fair to make since while one \
>>> path is in a transitioning state it is reasonable to assume that other paths may \
>>> also be in non active states.
>>> I beg to disagree. Path groups are nominally independent, and might
>>> change states independent on the other path groups.
>>> While for some arrays a 'transitioning' state is indeed system-wide,
>>> other arrays might be able to serve I/O on other paths whilst one is in
>>> transitioning.
>>> So I'd rather not presume anything here.
>> I agree. No problem there. Our array could and does return transitioning on
>> some portal groups while others might be active/online or unavailable.
>>> As outlined above, we cannot assume that all paths will be set to
>>> 'transitioning' once we hit the 'transitioning' condition on one path.
>>> As such, we need to retry the I/O on other paths, to ensure failover
>>> does work in these cases. Hence it's perfectly okay to set this path to
>> ‘> failed' as we cannot currently send I/O to that path.
>>> If, however, we are hitting a 'transitioning' status on _all_ paths (ie
>>> all paths are set to 'failed') we need to ensure that we do _not_ fail
>>> the I/O (as technically the paths are still alive), but retry with
>>> TUR/RTPG until one path reaches a final state.
>>> Then we should reinstate that path and continue with I/O.
>> I am not saying that all paths should be changed to transitioning, but
>> I/Os sent to the path that is in transitioning should not immediately
>> fail if there is not an online path like what does happen without
>> this patch or one like it.
>> The other paths which are in other states should succeed or fail
>> I/O as they would based on their state. I am only concerned about
>> the portal group in the transitioning state and making sure it doesn’t
>> immediately bubble errors back to the multipath layer which fails the
>> path which is what we see and don’t want to see.
>>> So what is the error you are seeing?
>> Right now this is what fails and used to work before the patch
>> This worked in previous Linux versions and continues to work
>> in Windows, ESXi, Solaris, AIX, and HP-UX. I have tested those.
>> It might work on others as well, but that list is good enough for me.
>> We have an array with two controllers and when all is good
>> each controller reports active/optimized for all of it ports. There
>> Is a TPG per controller.
>> CT0 - Primary - AO - TPG 0
>> CT1 - Secondary - AO - TPG 1
>> In any upgrade there is a point where we have to have the
>> secondary promote to primary. In our world we call this a
>> giveback. This is done by returning unavailable for I/O
>> that is sent to the previous primary CT0 and transitioning
>> for CT1, the promoting secondary:
>> CT0 - was primary - unavailable - TPG 0
>> CT1 - promoting not yet primary - transitioning - TPG 1
>> This is where we hit the issue. The paths to CT0 fail
>> since its ALUA state is unavailable as expected. The paths
>> to CT1 also quickly fail in the same second after some
>> retries. There are no paths which can serve I/O for a
>> short time as the secondary promotes to primary. We
>> expect ALUA state transitioning to protect this path
>> against an I/O error returning to multipath which it
>> no longer does.
>> If it worked we would expect:
>> CT0 - becoming secondary - still unavailable - TPG 0
>> CT1 - Primary - AO - TPG 1
>> And a short time later:
>> CT0 - secondary - AO - TPG 0
>> CT1 - primary - AO - TPG 1
>> Hopefully that helps with the context and why we
>> are proposing what we are.
>> Ah-ha.
> 'Unavailable' state. Right.
> 
> Hmm. Seems that we need to distinguish (at the device-mapper multipath layer) between temporarily failed paths (like transitioning), which could become available at a later time, and permanently failed paths (like unavailable or standby), for which a retry would not yield different results. I thought we did that, but apparently there's an issue somewhere.
> 
> Lemme see ...
> 
> Cheers,
> 
> Hannes
> -- 
> Dr. Hannes Reinecke                Kernel Storage Architect
> hare@suse.de                              +49 911 74053 688
> SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
> HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer

Thanks. For us, the current behavior after the fix I mentioned which broke things leaves us in a bad state. The only thing we can do is use no_path_retry in multipath.conf, but that brings with it just a new set of problems. There are times we do want an I/O error to return to the caller as quickly as possible (We actual don’t have any paths that can serve I/O). The state where one controller is unavailable and the other is transitioning is not a condition where we would want an I/O error returned. We expect that the transitioning path would protect us until it transitions to an online state in our case. These transition times in our case are very short, just a few seconds, but we would prefer to not hold on to I/O on the target if we can’t serve it. The transitioning state had been working great for this up until Centos 8 stream picked up the change.

Thanks,
Brian