On 5/20/22 3:03 PM, Martin Wilck wrote:
> On Fri, 2022-05-20 at 14:08 -0500, Mike Christie wrote:
>> On 5/20/22 9:03 AM, Martin Wilck wrote:
>>> On Fri, 2022-05-20 at 14:06 +0200, Hannes Reinecke wrote:
>>>> On 5/20/22 12:57, Martin Wilck wrote:
>>>>> Brian, Martin,
>>>>>
>>>>> sorry, I've overlooked this patch previously. I have to say I
>>>>> think
>>>>> it's wrong and shouldn't have been applied. At least I need
>>>>> more
>>>>> in-
>>>>> depth explanation.
>>>>>
>>>>> On Mon, 2022-05-02 at 20:50 -0400, Martin K. Petersen wrote:
>>>>>> On Mon, 2 May 2022 08:09:17 -0700, Brian Bunker wrote:
>>>>>>
>>>>>>> The handling of the ALUA transitioning state is currently
>>>>>>> broken.
>>>>>>> When
>>>>>>> a target goes into this state, it is expected that the
>>>>>>> target
>>>>>>> is
>>>>>>> allowed to stay in this state for the implicit transition
>>>>>>> timeout
>>>>>>> without a path failure.
>>>>>
>>>>> Can you please show me a quote from the specs on which this
>>>>> expectation
>>>>> ("without a path failure") is based? AFAIK the SCSI specs don't
>>>>> say
>>>>> anything about device-mapper multipath semantics.
>>>>>
>>>>>>> The handler has this logic, but it gets
>>>>>>> skipped currently.
>>>>>>>
>>>>>>> When the target transitions, there is in-flight I/O from
>>>>>>> the
>>>>>>> initiator. The first of these responses from the target
>>>>>>> will be
>>>>>>> a
>>>>>>> unit
>>>>>>> attention letting the initiator know that the ALUA state
>>>>>>> has
>>>>>>> changed.
>>>>>>> The remaining in-flight I/Os, before the initiator finds
>>>>>>> out
>>>>>>> that
>>>>>>> the
>>>>>>> portal state has changed, will return not ready, ALUA state
>>>>>>> is
>>>>>>> transitioning. The portal state will change to
>>>>>>> SCSI_ACCESS_STATE_TRANSITIONING. This will lead to all new
>>>>>>> I/O
>>>>>>> immediately failing the path unexpectedly. The path failure
>>>>>>> happens
>>>>>>> in
>>>>>>> less than a second instead of the expected successes until
>>>>>>> the
>>>>>>> transition timer is exceeded.
>>>>>
>>>>> dm multipath has no concept of "transitioning" state. Path
>>>>> state
>>>>> can be
>>>>> either active or inactive. As Brian wrote, commands sent to the
>>>>> transitioning device will return NOT READY, TRANSITIONING, and
>>>>> require
>>>>> retries on the SCSI layer. If we know this in advance, why
>>>>> should
>>>>> we
>>>>> continue sending I/O down this semi-broken path? If other,
>>>>> healthy
>>>>> paths are available, why it would it not be the right thing to
>>>>> switch
>>>>> I/O to them ASAP?
>>>>>
>>>> But we do, don't we?
>>>> Commands are being returned with the appropriate status, and
>>>> dm-multipath should make the corresponding decisions here.
>>>> This patch just modifies the check when _sending_ commands; ie
>>>> multipath
>>>> had decided that the path is still usable.
>>>> Question rather would be why multipath did that;
>>>
>>> If alua_prep_fn() got called, the path was considered usable at the
>>> given point in time by dm-multipath. Most probably the reason was
>>> simply that no error condition had occured on this path before ALUA
>>> state switched to transitioning. I suppose this can happen
>>> if storage
>>> switches a PG consisting of multiple paths to TRANSITIONING. We get
>>> an
>>> error on one path (sda, say), issue an RTPG, and receive the new
>>> ALUA
>>> state for all paths of the PG. For all paths except sda, we'd just
>>> see
>>> a switch to TRANSITIONING without a previous SCSI error.
>>>
>>> With this patch, we'll dispatch I/O (usually an entire bunch) to
>>> these
>>> paths despite seeing them in TRANSITIONING state. Eventually, when
>>> the
>>> SCSI responses are received, this leads to path failures. If I/O
>>> latencies are small, this happens after a few ms. In that case, the
>>> goal of Brian's patch is not reached, because the time until path
>>> failure would still be on the order of milliseconds. OTOH, if
>>> latencies
>>> are high, it takes substantially longer for the kernel to realize
>>> that
>>> the path is non-functional, while other, good paths may be idle. I
>>> fail
>>> to see the benefit.
>>>
>>
>> I'm not sure everyone agrees with you on the meaning of
>> transitioning.
>>
>> If we go from non-optimized to optimized or standby to non-
>> opt/optimized
>> we don't want to try other paths because it can cause thrashing.
>
> But only with explicit ALUA, or am I missing something? I agree that
That section of the spec mentions both implicit and explicit. For
implicit, the target can want to rebalance resources for things like
a resource is down permanently, or you add more ports, or we bring
up/down resources dynamically based on usage or maintenance.
> the host shouldn't initiate a PG switch if it encounters transitioning
> state. I also agree that for transitioning towards a "better" state,
> e.g. standby to (non)-optimized, failing the path would be
> questionable. Unfortunately we don't know in which "direction" the path
> is transitioning - it could be for 'better' or 'worse'. I suppose that
For implicit, the target knows. It's initiating the transition based
on whatever metrics or resources it has. We want the initiator to let us
complete what we are doing.
For explicit, then again the target knows what it wants to do when
it gets the STPG and we only it to set the paths to optimized. So
if it goes that route where it completes the STPG before the transition
completes, then goes into transitioning, then we can just let the
device do it's transitions.
> in the case of a PG switch, it can happen that we dispatch I/O to a
> device that used to be in Standby and is now transitioning. Would it
> make sense to remember the previous state and "guess" what we're going
> to transition to? I.e. if the previous state was "Standby", it's
> probably going to be (non)optimized after the transition, and vice-
> versa?
You are referring to the issue Hannes mentions where multipath can pick
up the transitioning state and it might get confused, right? I'm not
sure what to do.
>
>> We just
>> need to transition resources before we can fully use the path. It
>> could
>> be a local cache operation or for distributed targets it could be a
>> really
>> expensive operation.
>>
>> For both though, it can take longer than the retries we get from
>> scsi-ml.
>
> So if we want to do "the right thing", we'd continue dispatching to the
> device until either the state changes or the device-reported transition
> timeout has expired?
Sort of.
Ideally I think it would be nice if we blocked the device/queue for
normal IO, then just sent a RTPG every N secs or msecs until we changed
state or until the timer expired. We then unblock and either fail upwards
or dispatch. I think this is a lot of work though.
The problem with constant dispatching is that on low latency systems we
retry too quickly. I had to add a little sleep on the target side for this
or we hammer the target/initiator too hard and we got warnings (I can't
remember the exact warn/err).
