Re: [PATCH 3/3] NVMe: Convert to blk-mq

[Keith Busch]

On Tue, 22 Oct 2013, Matias Bjorling wrote:
> Den 22-10-2013 18:55, Keith Busch skrev:
> > On Fri, 18 Oct 2013, Matias Bjørling wrote:
> > > On 10/18/2013 05:13 PM, Keith Busch wrote:
> > > > On Fri, 18 Oct 2013, Matias Bjorling wrote:
> > > > > The nvme driver implements itself as a bio-based driver. This primarily
> > > > > because of high lock congestion for high-performance nvm devices. To
> > > > > remove the congestion within the traditional block layer, a multi-queue
> > > > > block layer is being implemented.
> >
> > > > > -    result = nvme_map_bio(nvmeq, iod, bio, dma_dir, psegs);
> > > > > -    if (result <= 0)
> > > > > +    if (nvme_map_rq(nvmeq, iod, rq, dma_dir))
> > > > >         goto free_cmdid;
> > > > > -    length = result;
> > > > >
> > > > > -    cmnd->rw.command_id = cmdid;
> > > > > +    length = blk_rq_bytes(rq);
> > > > > +
> > > > > +    cmnd->rw.command_id = rq->tag;
> > > >
> > > > The command ids have to be unique on a submission queue. Since each
> > > > namespace's blk-mq has its own 'tags' used as command ids here but share
> > > > submission queues, what's stopping the tags for commands sent to 
> > > > namespace
> > > > 1 from clashing with tags for namespace 2?
> > > >
> > > > I think this would work better if one blk-mq was created per device
> > > > rather than namespace. It would fix the tag problem above and save a
> > > > lot of memory potentially wasted on millions of requests allocated that
> > > > can't be used.
> > >
> > > You're right. I didn't see the connection. In v3 I'll push struct 
> > > request_queue to nvme_dev and map the queues appropriately. It will also 
> > > fix the command id issues.
> >
> > Just anticipating a possible issue with the suggestion. Will this separate
> > the logical block size from the request_queue? Each namespace can have
> > a different format, so the block size and request_queue can't be tied
> > together like it currently is for this to work.
>
> If only a couple of different logical sizes are to be expected (1-4), we can 
> keep a list of already initialized request queues, and use the one that match 
> an already initialized?

The spec allows a namespace to have up to 16 different block formats and
they need not be the same 16 as another namespace on the same device.

> From a practical standpoint, I don't think devices will support more
than a few formats, but even if you kept it to that many request queues,
you just get back to conflicting command id tags and some wasted memory.

> Axboe, do you know of a better solution?