Re: [PATCH v4 1/2] virtio-blk: support polling I/O

[Suwan Kim <suwan.kim027@gmail.com>]

On Tue, Apr 05, 2022 at 01:51:40AM -0700, Christoph Hellwig wrote:
> On Tue, Apr 05, 2022 at 02:31:21PM +0900, Suwan Kim wrote:
> > This patch supports polling I/O via virtio-blk driver. Polling
> > feature is enabled by module parameter "num_poll_queues" and it
> > sets dedicated polling queues for virtio-blk. This patch improves
> > the polling I/O throughput and latency.
> > 
> > The virtio-blk driver doesn't not have a poll function and a poll
> > queue and it has been operating in interrupt driven method even if
> > the polling function is called in the upper layer.
> > 
> > virtio-blk polling is implemented upon 'batched completion' of block
> > layer. virtblk_poll() queues completed request to io_comp_batch->req_list
> > and later, virtblk_complete_batch() calls unmap function and ends
> > the requests in batch.
> > 
> > virtio-blk reads the number of poll queues from module parameter
> > "num_poll_queues". If VM sets queue parameter as below,
> > ("num-queues=N" [QEMU property], "num_poll_queues=M" [module parameter])
> > It allocates N virtqueues to virtio_blk->vqs[N] and it uses [0..(N-M-1)]
> > as default queues and [(N-M)..(N-1)] as poll queues. Unlike the default
> > queues, the poll queues have no callback function.
> > 
> > Regarding HW-SW queue mapping, the default queue mapping uses the
> > existing method that condsiders MSI irq vector. But the poll queue
> > doesn't have an irq, so it uses the regular blk-mq cpu mapping.
> > 
> > For verifying the improvement, I did Fio polling I/O performance test
> > with io_uring engine with the options below.
> > (io_uring, hipri, randread, direct=1, bs=512, iodepth=64 numjobs=N)
> > I set 4 vcpu and 4 virtio-blk queues - 2 default queues and 2 poll
> > queues for VM.
> > 
> > As a result, IOPS and average latency improved about 10%.
> > 
> > Test result:
> > 
> > - Fio io_uring poll without virtio-blk poll support
> > 	-- numjobs=1 : IOPS = 339K, avg latency = 188.33us
> > 	-- numjobs=2 : IOPS = 367K, avg latency = 347.33us
> > 	-- numjobs=4 : IOPS = 383K, avg latency = 682.06us
> > 
> > - Fio io_uring poll with virtio-blk poll support
> > 	-- numjobs=1 : IOPS = 385K, avg latency = 165.94us
> > 	-- numjobs=2 : IOPS = 408K, avg latency = 313.28us
> > 	-- numjobs=4 : IOPS = 424K, avg latency = 613.05us
> > 
> > Signed-off-by: Suwan Kim <suwan.kim027@gmail.com>
> > ---
> >  drivers/block/virtio_blk.c | 112 +++++++++++++++++++++++++++++++++++--
> >  1 file changed, 108 insertions(+), 4 deletions(-)
> > 
> > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
> > index 8c415be86732..712579dcd3cc 100644
> > --- a/drivers/block/virtio_blk.c
> > +++ b/drivers/block/virtio_blk.c
> > @@ -37,6 +37,10 @@ MODULE_PARM_DESC(num_request_queues,
> >  		 "0 for no limit. "
> >  		 "Values > nr_cpu_ids truncated to nr_cpu_ids.");
> >  
> > +static unsigned int poll_queues;
> > +module_param(poll_queues, uint, 0644);
> > +MODULE_PARM_DESC(poll_queues, "The number of dedicated virtqueues for polling I/O");
> > +
> >  static int major;
> >  static DEFINE_IDA(vd_index_ida);
> >  
> > @@ -81,6 +85,7 @@ struct virtio_blk {
> >  
> >  	/* num of vqs */
> >  	int num_vqs;
> > +	int io_queues[HCTX_MAX_TYPES];
> >  	struct virtio_blk_vq *vqs;
> >  };
> >  
> > @@ -548,6 +553,7 @@ static int init_vq(struct virtio_blk *vblk)
> >  	const char **names;
> >  	struct virtqueue **vqs;
> >  	unsigned short num_vqs;
> > +	unsigned int num_poll_vqs;
> >  	struct virtio_device *vdev = vblk->vdev;
> >  	struct irq_affinity desc = { 0, };
> >  
> > @@ -556,6 +562,7 @@ static int init_vq(struct virtio_blk *vblk)
> >  				   &num_vqs);
> >  	if (err)
> >  		num_vqs = 1;
> > +
> >  	if (!err && !num_vqs) {
> >  		dev_err(&vdev->dev, "MQ advertised but zero queues reported\n");
> >  		return -EINVAL;
> > @@ -565,6 +572,18 @@ static int init_vq(struct virtio_blk *vblk)
> >  			min_not_zero(num_request_queues, nr_cpu_ids),
> >  			num_vqs);
> >  
> > +	num_poll_vqs = min_t(unsigned int, poll_queues, num_vqs - 1);
> > +
> > +	memset(vblk->io_queues, 0, sizeof(int) * HCTX_MAX_TYPES);
> > +	vblk->io_queues[HCTX_TYPE_DEFAULT] = num_vqs - num_poll_vqs;
> > +	vblk->io_queues[HCTX_TYPE_READ] = 0;
> > +	vblk->io_queues[HCTX_TYPE_POLL] = num_poll_vqs;
> > +
> > +	dev_info(&vdev->dev, "%d/%d/%d default/read/poll queues\n",
> > +				vblk->io_queues[HCTX_TYPE_DEFAULT],
> > +				vblk->io_queues[HCTX_TYPE_READ],
> > +				vblk->io_queues[HCTX_TYPE_POLL]);
> > +
> >  	vblk->vqs = kmalloc_array(num_vqs, sizeof(*vblk->vqs), GFP_KERNEL);
> >  	if (!vblk->vqs)
> >  		return -ENOMEM;
> > @@ -578,8 +597,13 @@ static int init_vq(struct virtio_blk *vblk)
> >  	}
> >  
> >  	for (i = 0; i < num_vqs; i++) {
> > +		if (i < num_vqs - num_poll_vqs) {
> > +			callbacks[i] = virtblk_done;
> > +			snprintf(vblk->vqs[i].name, VQ_NAME_LEN, "req.%d", i);
> > +		} else {
> > +			callbacks[i] = NULL;
> > +			snprintf(vblk->vqs[i].name, VQ_NAME_LEN, "req_poll.%d", i);
> > +		}
> >  		names[i] = vblk->vqs[i].name;
> 
> This would look a little cleaner with two loops:
> 
>  	for (i = 0; i < num_vqs - num_poll_vqs; i++) {
> 		callbacks[i] = virtblk_done;
> 		snprintf(vblk->vqs[i].name, VQ_NAME_LEN, "req.%d", i);
>  		names[i] = vblk->vqs[i].name;
> 	}
>  	for (; i < num_vqs; i++) {
> 		callbacks[i] = NULL;
> 		snprintf(vblk->vqs[i].name, VQ_NAME_LEN, "req_poll.%d", i);
> 		names[i] = vblk->vqs[i].name;
> 	}
> 
> > +
> > +		if (map->nr_queues == 0)
> > +			continue;
> > +
> > +		/*
> > +		 * Regular queues have interrupts and hence CPU affinity is
> > +		 * defined by the core virtio code, but polling queues have
> > +		 * no interrupts so we let the block layer assign CPU affinity.
> > +		 */
> > +		if (i == HCTX_TYPE_DEFAULT)
> 
> I'd check for
> 		i != HCTX_TYPE_POLL
> 
> here instead to make the check a little more explicit and future proof
> for the potential addition of read queues (which would be a Linux only
> change without hypervisor or spec changes).  In fact you might as well
> add that support now as doing it is completely trivial once a driver
> supports multiple map types.
> 
> > +static void virtblk_complete_batch(struct io_comp_batch *iob)
> > +{
> > +	struct request *req;
> > +	struct virtblk_req *vbr;
> > +
> > +	rq_list_for_each(&iob->req_list, req) {
> > +		vbr = blk_mq_rq_to_pdu(req);
> > +		virtblk_unmap_data(req, vbr);
> > +		virtblk_cleanup_cmd(req);
> 
> vbr is only used ones, so why not just:
> 
> 		virtblk_unmap_data(req, blk_mq_rq_to_pdu);
> ?
> 
> Or even better add a cleanup patch to just remove the vbr argument to
> virtblk_unmap_data as it is not needed at all.

I replied that I will add cleanup path but after thinking about it
again, it seems better not to remove the argument because
blk_mq_rq_to_pdu() needs type casting anyway.

So I will fix it to "virtblk_unmap_data(req, blk_mq_rq_to_pdu);"
as you mentioned.

Regards,
Suwan Kim