Re: [PATCH] Revert "IB/core: Fix use workqueue without WQ_MEM_RECLAIM"

[Jason Gunthorpe <jgg@ziepe.ca>]

On Tue, Jun 06, 2023 at 02:01:04AM +0300, Sagi Grimberg wrote:
> 
> > > > > workqueue: WQ_MEM_RECLAIM nvme-wq:nvme_rdma_reconnect_ctrl_work
> > > > > [nvme_rdma] is flushing !WQ_MEM_RECLAIM ib_addr:process_one_req [ib_core]
> > > > 
> > > > And why does nvme-wq need WQ_MEM_RECLAIM flag? I wonder if it is really
> > > > needed.
> > > 
> > > Adding Sagi Grimberg to cc, he probably knows and can explain it better than me.
> > 
> > We already allocate so much memory on these paths it is pretty
> > nonsense to claim they are a reclaim context. One allocation on the WQ
> > is not going to be the problem.
> > 
> > Probably this nvme stuff should not be re-using a reclaim marke dWQ
> > for memory allocating work like this, it is kind of nonsensical.
> 
> A controller reset runs on this workqueue, which should succeed to allow
> for pages to be flushed to the nvme disk. So I'd say its kind of
> essential that this sequence has a rescuer thread.

So don't run the CM stuff on the same WQ, go to another one without
the reclaim mark?

Jason