[PATCH 0/1] nvmet: allow user to set req alloc flag

[Keith Busch]

On Tue, Oct 20, 2020 at 09:39:28AM -0600, Logan Gunthorpe wrote:
> 
> 
> On 2020-10-20 2:22 a.m., Sagi Grimberg wrote:
> > 
> >> Hi,
> >>
> >> When using NVMeOF target in the passthru mode we allocate the request
> >> with BLK_MQ_REQ_NOWAIT flag. This allocates the request in the following
> >> manner :-
> >>
> >> nvme_alloc_request()
> >>   blk_mq_alloc_rquest()
> >>    blk_mq_queue_enter()
> >>     if (flag & BLK_MQ_REQ_NOWAIT)
> >>    	return -EBUSY; <-- return if busy.
> >>
> >> On the NVMe controller which I've the fio random write workload running
> >> parallel on 32 namespaces with higher queue depth results in I/O error,
> >> where blk_mq_queue_enter() returning -EBUSY as shown above. This problem
> >> is not easy to reproduce but occurs once in a while with following error
> >> (See 1 for detailed log) :-
> >>
> >> test1: (groupid=0, jobs=32): err= 5
> >> (file:io_u.c:1744, func=io_u error, error=Input/output error):
> >>
> >> When the flag BLK_MQ_REQ_NOWAIT is removed from the allocation the
> >> workload doen't result in the error.
> >>
> >> This patch fixes the problem with the request allocation by adding
> >> a new configfs attribute so that user can optionally decide whether
> >> to use BLK_MQ_REQ_NOWAIT or not. We retain the default behavior by
> >> using BLK_MQ_REQ_NOWAIT when creating the nvmet passthru subsystem.
> > 
> > Why should we ever set REQ_NOWAIT at all? Nothing prevents the
> > host(s) queue depth from exceeding the controller queue depth...
> 
> I agree... I certainly found adding a configfs attribute for this rather
> off-putting. Why would the user want an option that turns on random errors?

Right, you really only want to use NOWAIT when waiting can lead to bad
things like a deadlock. I don't think the target passthrough has any
such issue, so you should be able to safely leave that flag out.