| Message ID | 20220324140450.33148-2-suwan.kim027@gmail.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | virtio-blk: support polling I/O and mq_ops->queue_rqs() | expand |
On Thu, Mar 24, 2022 at 11:04:49PM +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 = 380K, avg latency = 167.87us > -- numjobs=2 : IOPS = 409K, avg latency = 312.6us > -- numjobs=4 : IOPS = 413K, avg latency = 619.72us > > Reported-by: kernel test robot <lkp@intel.com> > Signed-off-by: Suwan Kim <suwan.kim027@gmail.com> > --- > drivers/block/virtio_blk.c | 101 +++++++++++++++++++++++++++++++++++-- > 1 file changed, 97 insertions(+), 4 deletions(-) > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > index 8c415be86732..3d16f8b753e7 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 num_poll_queues; > +module_param(num_poll_queues, uint, 0644); > +MODULE_PARM_DESC(num_poll_queues, "The number of dedicated virtqueues for polling I/O"); > + > static int major; > static DEFINE_IDA(vd_index_ida); > Is there some way to make it work reasonably without need to set module parameters? I don't see any other devices with a num_poll_queues parameter - how do they handle this? > @@ -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 +572,13 @@ 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, num_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; > + > vblk->vqs = kmalloc_array(num_vqs, sizeof(*vblk->vqs), GFP_KERNEL); > if (!vblk->vqs) > return -ENOMEM; > @@ -578,8 +592,13 @@ static int init_vq(struct virtio_blk *vblk) > } > > for (i = 0; i < num_vqs; i++) { > - callbacks[i] = virtblk_done; > - snprintf(vblk->vqs[i].name, VQ_NAME_LEN, "req.%d", 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; > } > > @@ -728,16 +747,87 @@ static const struct attribute_group *virtblk_attr_groups[] = { > static int virtblk_map_queues(struct blk_mq_tag_set *set) > { > struct virtio_blk *vblk = set->driver_data; > + int i, qoff; > + > + for (i = 0, qoff = 0; i < set->nr_maps; i++) { > + struct blk_mq_queue_map *map = &set->map[i]; > + > + map->nr_queues = vblk->io_queues[i]; > + map->queue_offset = qoff; > + qoff += map->nr_queues; > + > + 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) > + blk_mq_virtio_map_queues(&set->map[i], vblk->vdev, 0); > + else > + blk_mq_map_queues(&set->map[i]); > + } > + > + return 0; > +} > + > +static void virtblk_complete_batch(struct io_comp_batch *iob) > +{ > + struct request *req; > + struct virtblk_req *vbr; > > - return blk_mq_virtio_map_queues(&set->map[HCTX_TYPE_DEFAULT], > - vblk->vdev, 0); > + rq_list_for_each(&iob->req_list, req) { > + vbr = blk_mq_rq_to_pdu(req); > + virtblk_unmap_data(req, vbr); > + virtblk_cleanup_cmd(req); > + } > + blk_mq_end_request_batch(iob); > +} > + > +static int virtblk_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob) > +{ > + struct virtio_blk_vq *vq = hctx->driver_data; > + struct virtblk_req *vbr; > + unsigned long flags; > + unsigned int len; > + int found = 0; > + > + spin_lock_irqsave(&vq->lock, flags); > + > + while ((vbr = virtqueue_get_buf(vq->vq, &len)) != NULL) { > + struct request *req = blk_mq_rq_from_pdu(vbr); > + > + found++; > + if (!blk_mq_add_to_batch(req, iob, vbr->status, > + virtblk_complete_batch)) > + blk_mq_complete_request(req); > + } > + > + spin_unlock_irqrestore(&vq->lock, flags); > + > + return found; > +} > + > +static int virtblk_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, > + unsigned int hctx_idx) > +{ > + struct virtio_blk *vblk = data; > + struct virtio_blk_vq *vq = &vblk->vqs[hctx_idx]; > + > + WARN_ON(vblk->tag_set.tags[hctx_idx] != hctx->tags); > + hctx->driver_data = vq; > + return 0; > } > > static const struct blk_mq_ops virtio_mq_ops = { > .queue_rq = virtio_queue_rq, > .commit_rqs = virtio_commit_rqs, > + .init_hctx = virtblk_init_hctx, > .complete = virtblk_request_done, > .map_queues = virtblk_map_queues, > + .poll = virtblk_poll, > }; > > static unsigned int virtblk_queue_depth; > @@ -816,6 +906,9 @@ static int virtblk_probe(struct virtio_device *vdev) > sizeof(struct scatterlist) * VIRTIO_BLK_INLINE_SG_CNT; > vblk->tag_set.driver_data = vblk; > vblk->tag_set.nr_hw_queues = vblk->num_vqs; > + vblk->tag_set.nr_maps = 1; > + if (vblk->io_queues[HCTX_TYPE_POLL]) > + vblk->tag_set.nr_maps = 3; > > err = blk_mq_alloc_tag_set(&vblk->tag_set); > if (err) > -- > 2.26.3
On Thu, Mar 24, 2022 at 10:32:02AM -0400, Michael S. Tsirkin wrote: > On Thu, Mar 24, 2022 at 11:04:49PM +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 = 380K, avg latency = 167.87us > > -- numjobs=2 : IOPS = 409K, avg latency = 312.6us > > -- numjobs=4 : IOPS = 413K, avg latency = 619.72us > > > > Reported-by: kernel test robot <lkp@intel.com> > > Signed-off-by: Suwan Kim <suwan.kim027@gmail.com> > > --- > > drivers/block/virtio_blk.c | 101 +++++++++++++++++++++++++++++++++++-- > > 1 file changed, 97 insertions(+), 4 deletions(-) > > > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > > index 8c415be86732..3d16f8b753e7 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 num_poll_queues; > > +module_param(num_poll_queues, uint, 0644); > > +MODULE_PARM_DESC(num_poll_queues, "The number of dedicated virtqueues for polling I/O"); > > + > > static int major; > > static DEFINE_IDA(vd_index_ida); > > > > Is there some way to make it work reasonably without need to set > module parameters? I don't see any other devices with a num_poll_queues > parameter - how do they handle this? Hi Michael, NVMe driver uses module parameter. Please refer to this. ----- drivers/nvme/host/pci.c static unsigned int poll_queues; module_param_cb(poll_queues, &io_queue_count_ops, &poll_queues, 0644); MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO."); ----- Regards, Suwan Kim
Hi Suwan, The NVMe prints something like below by nvme_setup_io_queues() to confirm if the configuration takes effect. "[ 0.620458] nvme nvme0: 4/0/0 default/read/poll queues". How about to print in virtio-blk as well? Thank you very much! Dongli Zhang On 3/24/22 7:04 AM, 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 = 380K, avg latency = 167.87us > -- numjobs=2 : IOPS = 409K, avg latency = 312.6us > -- numjobs=4 : IOPS = 413K, avg latency = 619.72us > > Reported-by: kernel test robot <lkp@intel.com> > Signed-off-by: Suwan Kim <suwan.kim027@gmail.com> > --- > drivers/block/virtio_blk.c | 101 +++++++++++++++++++++++++++++++++++-- > 1 file changed, 97 insertions(+), 4 deletions(-) > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > index 8c415be86732..3d16f8b753e7 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 num_poll_queues; > +module_param(num_poll_queues, uint, 0644); > +MODULE_PARM_DESC(num_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 +572,13 @@ 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, num_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; > + > vblk->vqs = kmalloc_array(num_vqs, sizeof(*vblk->vqs), GFP_KERNEL); > if (!vblk->vqs) > return -ENOMEM; > @@ -578,8 +592,13 @@ static int init_vq(struct virtio_blk *vblk) > } > > for (i = 0; i < num_vqs; i++) { > - callbacks[i] = virtblk_done; > - snprintf(vblk->vqs[i].name, VQ_NAME_LEN, "req.%d", 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; > } > > @@ -728,16 +747,87 @@ static const struct attribute_group *virtblk_attr_groups[] = { > static int virtblk_map_queues(struct blk_mq_tag_set *set) > { > struct virtio_blk *vblk = set->driver_data; > + int i, qoff; > + > + for (i = 0, qoff = 0; i < set->nr_maps; i++) { > + struct blk_mq_queue_map *map = &set->map[i]; > + > + map->nr_queues = vblk->io_queues[i]; > + map->queue_offset = qoff; > + qoff += map->nr_queues; > + > + 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) > + blk_mq_virtio_map_queues(&set->map[i], vblk->vdev, 0); > + else > + blk_mq_map_queues(&set->map[i]); > + } > + > + return 0; > +} > + > +static void virtblk_complete_batch(struct io_comp_batch *iob) > +{ > + struct request *req; > + struct virtblk_req *vbr; > > - return blk_mq_virtio_map_queues(&set->map[HCTX_TYPE_DEFAULT], > - vblk->vdev, 0); > + rq_list_for_each(&iob->req_list, req) { > + vbr = blk_mq_rq_to_pdu(req); > + virtblk_unmap_data(req, vbr); > + virtblk_cleanup_cmd(req); > + } > + blk_mq_end_request_batch(iob); > +} > + > +static int virtblk_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob) > +{ > + struct virtio_blk_vq *vq = hctx->driver_data; > + struct virtblk_req *vbr; > + unsigned long flags; > + unsigned int len; > + int found = 0; > + > + spin_lock_irqsave(&vq->lock, flags); > + > + while ((vbr = virtqueue_get_buf(vq->vq, &len)) != NULL) { > + struct request *req = blk_mq_rq_from_pdu(vbr); > + > + found++; > + if (!blk_mq_add_to_batch(req, iob, vbr->status, > + virtblk_complete_batch)) > + blk_mq_complete_request(req); > + } > + > + spin_unlock_irqrestore(&vq->lock, flags); > + > + return found; > +} > + > +static int virtblk_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, > + unsigned int hctx_idx) > +{ > + struct virtio_blk *vblk = data; > + struct virtio_blk_vq *vq = &vblk->vqs[hctx_idx]; > + > + WARN_ON(vblk->tag_set.tags[hctx_idx] != hctx->tags); > + hctx->driver_data = vq; > + return 0; > } > > static const struct blk_mq_ops virtio_mq_ops = { > .queue_rq = virtio_queue_rq, > .commit_rqs = virtio_commit_rqs, > + .init_hctx = virtblk_init_hctx, > .complete = virtblk_request_done, > .map_queues = virtblk_map_queues, > + .poll = virtblk_poll, > }; > > static unsigned int virtblk_queue_depth; > @@ -816,6 +906,9 @@ static int virtblk_probe(struct virtio_device *vdev) > sizeof(struct scatterlist) * VIRTIO_BLK_INLINE_SG_CNT; > vblk->tag_set.driver_data = vblk; > vblk->tag_set.nr_hw_queues = vblk->num_vqs; > + vblk->tag_set.nr_maps = 1; > + if (vblk->io_queues[HCTX_TYPE_POLL]) > + vblk->tag_set.nr_maps = 3; > > err = blk_mq_alloc_tag_set(&vblk->tag_set); > if (err)
On Thu, Mar 24, 2022 at 11:46:02PM +0900, Suwan Kim wrote: > On Thu, Mar 24, 2022 at 10:32:02AM -0400, Michael S. Tsirkin wrote: > > On Thu, Mar 24, 2022 at 11:04:49PM +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 = 380K, avg latency = 167.87us > > > -- numjobs=2 : IOPS = 409K, avg latency = 312.6us > > > -- numjobs=4 : IOPS = 413K, avg latency = 619.72us > > > > > > Reported-by: kernel test robot <lkp@intel.com> > > > Signed-off-by: Suwan Kim <suwan.kim027@gmail.com> > > > --- > > > drivers/block/virtio_blk.c | 101 +++++++++++++++++++++++++++++++++++-- > > > 1 file changed, 97 insertions(+), 4 deletions(-) > > > > > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > > > index 8c415be86732..3d16f8b753e7 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 num_poll_queues; > > > +module_param(num_poll_queues, uint, 0644); > > > +MODULE_PARM_DESC(num_poll_queues, "The number of dedicated virtqueues for polling I/O"); > > > + > > > static int major; > > > static DEFINE_IDA(vd_index_ida); > > > > > > > Is there some way to make it work reasonably without need to set > > module parameters? I don't see any other devices with a num_poll_queues > > parameter - how do they handle this? > > Hi Michael, > > NVMe driver uses module parameter. > > Please refer to this. > ----- > drivers/nvme/host/pci.c > > static unsigned int poll_queues; > module_param_cb(poll_queues, &io_queue_count_ops, &poll_queues, 0644); > MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO."); > ----- > > Regards, > Suwan Kim OK then. Let's maybe be consistent wrt parameter naming?
On Thu, Mar 24, 2022 at 11:04:49PM +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 = 380K, avg latency = 167.87us > -- numjobs=2 : IOPS = 409K, avg latency = 312.6us > -- numjobs=4 : IOPS = 413K, avg latency = 619.72us > > Reported-by: kernel test robot <lkp@intel.com> > Signed-off-by: Suwan Kim <suwan.kim027@gmail.com> > --- > drivers/block/virtio_blk.c | 101 +++++++++++++++++++++++++++++++++++-- > 1 file changed, 97 insertions(+), 4 deletions(-) > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > index 8c415be86732..3d16f8b753e7 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 num_poll_queues; > +module_param(num_poll_queues, uint, 0644); > +MODULE_PARM_DESC(num_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 +572,13 @@ 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, num_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; > + > vblk->vqs = kmalloc_array(num_vqs, sizeof(*vblk->vqs), GFP_KERNEL); > if (!vblk->vqs) > return -ENOMEM; > @@ -578,8 +592,13 @@ static int init_vq(struct virtio_blk *vblk) > } > > for (i = 0; i < num_vqs; i++) { > - callbacks[i] = virtblk_done; > - snprintf(vblk->vqs[i].name, VQ_NAME_LEN, "req.%d", 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; > } > > @@ -728,16 +747,87 @@ static const struct attribute_group *virtblk_attr_groups[] = { > static int virtblk_map_queues(struct blk_mq_tag_set *set) > { > struct virtio_blk *vblk = set->driver_data; > + int i, qoff; > + > + for (i = 0, qoff = 0; i < set->nr_maps; i++) { > + struct blk_mq_queue_map *map = &set->map[i]; > + > + map->nr_queues = vblk->io_queues[i]; > + map->queue_offset = qoff; > + qoff += map->nr_queues; > + > + 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) > + blk_mq_virtio_map_queues(&set->map[i], vblk->vdev, 0); > + else > + blk_mq_map_queues(&set->map[i]); > + } > + > + return 0; > +} > + > +static void virtblk_complete_batch(struct io_comp_batch *iob) > +{ > + struct request *req; > + struct virtblk_req *vbr; > > - return blk_mq_virtio_map_queues(&set->map[HCTX_TYPE_DEFAULT], > - vblk->vdev, 0); > + rq_list_for_each(&iob->req_list, req) { > + vbr = blk_mq_rq_to_pdu(req); > + virtblk_unmap_data(req, vbr); > + virtblk_cleanup_cmd(req); > + } > + blk_mq_end_request_batch(iob); > +} > + > +static int virtblk_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob) > +{ > + struct virtio_blk_vq *vq = hctx->driver_data; > + struct virtblk_req *vbr; > + unsigned long flags; > + unsigned int len; > + int found = 0; > + > + spin_lock_irqsave(&vq->lock, flags); > + > + while ((vbr = virtqueue_get_buf(vq->vq, &len)) != NULL) { > + struct request *req = blk_mq_rq_from_pdu(vbr); > + > + found++; > + if (!blk_mq_add_to_batch(req, iob, vbr->status, > + virtblk_complete_batch)) > + blk_mq_complete_request(req); > + } > + > + spin_unlock_irqrestore(&vq->lock, flags); > + > + return found; > +} > + > +static int virtblk_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, > + unsigned int hctx_idx) > +{ > + struct virtio_blk *vblk = data; > + struct virtio_blk_vq *vq = &vblk->vqs[hctx_idx]; > + > + WARN_ON(vblk->tag_set.tags[hctx_idx] != hctx->tags); > + hctx->driver_data = vq; > + return 0; > } > > static const struct blk_mq_ops virtio_mq_ops = { > .queue_rq = virtio_queue_rq, > .commit_rqs = virtio_commit_rqs, > + .init_hctx = virtblk_init_hctx, > .complete = virtblk_request_done, > .map_queues = virtblk_map_queues, > + .poll = virtblk_poll, > }; > > static unsigned int virtblk_queue_depth; > @@ -816,6 +906,9 @@ static int virtblk_probe(struct virtio_device *vdev) > sizeof(struct scatterlist) * VIRTIO_BLK_INLINE_SG_CNT; > vblk->tag_set.driver_data = vblk; > vblk->tag_set.nr_hw_queues = vblk->num_vqs; > + vblk->tag_set.nr_maps = 1; > + if (vblk->io_queues[HCTX_TYPE_POLL]) > + vblk->tag_set.nr_maps = 3; > > err = blk_mq_alloc_tag_set(&vblk->tag_set); > if (err) So wrt cleanup, does something poll for all buffers to be used when device is removed? > -- > 2.26.3
On Thu, Mar 24, 2022 at 10:34:04AM -0700, Dongli Zhang wrote: > Hi Suwan, > > The NVMe prints something like below by nvme_setup_io_queues() to confirm > if the configuration takes effect. > > "[ 0.620458] nvme nvme0: 4/0/0 default/read/poll queues". > > How about to print in virtio-blk as well? Hi Dongli, Thansk for your feedback. It is good idea. I will add it in next version. Regards, Suwan Kim
On Thu, Mar 24, 2022 at 01:56:18PM -0400, Michael S. Tsirkin wrote: > On Thu, Mar 24, 2022 at 11:46:02PM +0900, Suwan Kim wrote: > > On Thu, Mar 24, 2022 at 10:32:02AM -0400, Michael S. Tsirkin wrote: > > > On Thu, Mar 24, 2022 at 11:04:49PM +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 = 380K, avg latency = 167.87us > > > > -- numjobs=2 : IOPS = 409K, avg latency = 312.6us > > > > -- numjobs=4 : IOPS = 413K, avg latency = 619.72us > > > > > > > > Reported-by: kernel test robot <lkp@intel.com> > > > > Signed-off-by: Suwan Kim <suwan.kim027@gmail.com> > > > > --- > > > > drivers/block/virtio_blk.c | 101 +++++++++++++++++++++++++++++++++++-- > > > > 1 file changed, 97 insertions(+), 4 deletions(-) > > > > > > > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > > > > index 8c415be86732..3d16f8b753e7 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 num_poll_queues; > > > > +module_param(num_poll_queues, uint, 0644); > > > > +MODULE_PARM_DESC(num_poll_queues, "The number of dedicated virtqueues for polling I/O"); > > > > + > > > > static int major; > > > > static DEFINE_IDA(vd_index_ida); > > > > > > > > > > Is there some way to make it work reasonably without need to set > > > module parameters? I don't see any other devices with a num_poll_queues > > > parameter - how do they handle this? > > > > Hi Michael, > > > > NVMe driver uses module parameter. > > > > Please refer to this. > > ----- > > drivers/nvme/host/pci.c > > > > static unsigned int poll_queues; > > module_param_cb(poll_queues, &io_queue_count_ops, &poll_queues, 0644); > > MODULE_PARM_DESC(poll_queues, "Number of queues to use for polled IO."); > > ----- > > > > Regards, > > Suwan Kim > > OK then. Let's maybe be consistent wrt parameter naming? Ok. Consistent naming scheme seems to be better for code readability. I will rename it to 'poll_queues' in next version. Regards, Suwan Kim
On Thu, Mar 24, 2022 at 01:58:28PM -0400, Michael S. Tsirkin wrote: > On Thu, Mar 24, 2022 at 11:04:49PM +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 = 380K, avg latency = 167.87us > > -- numjobs=2 : IOPS = 409K, avg latency = 312.6us > > -- numjobs=4 : IOPS = 413K, avg latency = 619.72us > > > > Reported-by: kernel test robot <lkp@intel.com> > > Signed-off-by: Suwan Kim <suwan.kim027@gmail.com> > > --- > > drivers/block/virtio_blk.c | 101 +++++++++++++++++++++++++++++++++++-- > > 1 file changed, 97 insertions(+), 4 deletions(-) > > > > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c > > index 8c415be86732..3d16f8b753e7 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 num_poll_queues; > > +module_param(num_poll_queues, uint, 0644); > > +MODULE_PARM_DESC(num_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 +572,13 @@ 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, num_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; > > + > > vblk->vqs = kmalloc_array(num_vqs, sizeof(*vblk->vqs), GFP_KERNEL); > > if (!vblk->vqs) > > return -ENOMEM; > > @@ -578,8 +592,13 @@ static int init_vq(struct virtio_blk *vblk) > > } > > > > for (i = 0; i < num_vqs; i++) { > > - callbacks[i] = virtblk_done; > > - snprintf(vblk->vqs[i].name, VQ_NAME_LEN, "req.%d", 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; > > } > > > > @@ -728,16 +747,87 @@ static const struct attribute_group *virtblk_attr_groups[] = { > > static int virtblk_map_queues(struct blk_mq_tag_set *set) > > { > > struct virtio_blk *vblk = set->driver_data; > > + int i, qoff; > > + > > + for (i = 0, qoff = 0; i < set->nr_maps; i++) { > > + struct blk_mq_queue_map *map = &set->map[i]; > > + > > + map->nr_queues = vblk->io_queues[i]; > > + map->queue_offset = qoff; > > + qoff += map->nr_queues; > > + > > + 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) > > + blk_mq_virtio_map_queues(&set->map[i], vblk->vdev, 0); > > + else > > + blk_mq_map_queues(&set->map[i]); > > + } > > + > > + return 0; > > +} > > + > > +static void virtblk_complete_batch(struct io_comp_batch *iob) > > +{ > > + struct request *req; > > + struct virtblk_req *vbr; > > > > - return blk_mq_virtio_map_queues(&set->map[HCTX_TYPE_DEFAULT], > > - vblk->vdev, 0); > > + rq_list_for_each(&iob->req_list, req) { > > + vbr = blk_mq_rq_to_pdu(req); > > + virtblk_unmap_data(req, vbr); > > + virtblk_cleanup_cmd(req); > > + } > > + blk_mq_end_request_batch(iob); > > +} > > + > > +static int virtblk_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob) > > +{ > > + struct virtio_blk_vq *vq = hctx->driver_data; > > + struct virtblk_req *vbr; > > + unsigned long flags; > > + unsigned int len; > > + int found = 0; > > + > > + spin_lock_irqsave(&vq->lock, flags); > > + > > + while ((vbr = virtqueue_get_buf(vq->vq, &len)) != NULL) { > > + struct request *req = blk_mq_rq_from_pdu(vbr); > > + > > + found++; > > + if (!blk_mq_add_to_batch(req, iob, vbr->status, > > + virtblk_complete_batch)) > > + blk_mq_complete_request(req); > > + } > > + > > + spin_unlock_irqrestore(&vq->lock, flags); > > + > > + return found; > > +} > > + > > +static int virtblk_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, > > + unsigned int hctx_idx) > > +{ > > + struct virtio_blk *vblk = data; > > + struct virtio_blk_vq *vq = &vblk->vqs[hctx_idx]; > > + > > + WARN_ON(vblk->tag_set.tags[hctx_idx] != hctx->tags); > > + hctx->driver_data = vq; > > + return 0; > > } > > > > static const struct blk_mq_ops virtio_mq_ops = { > > .queue_rq = virtio_queue_rq, > > .commit_rqs = virtio_commit_rqs, > > + .init_hctx = virtblk_init_hctx, > > .complete = virtblk_request_done, > > .map_queues = virtblk_map_queues, > > + .poll = virtblk_poll, > > }; > > > > static unsigned int virtblk_queue_depth; > > @@ -816,6 +906,9 @@ static int virtblk_probe(struct virtio_device *vdev) > > sizeof(struct scatterlist) * VIRTIO_BLK_INLINE_SG_CNT; > > vblk->tag_set.driver_data = vblk; > > vblk->tag_set.nr_hw_queues = vblk->num_vqs; > > + vblk->tag_set.nr_maps = 1; > > + if (vblk->io_queues[HCTX_TYPE_POLL]) > > + vblk->tag_set.nr_maps = 3; > > > > err = blk_mq_alloc_tag_set(&vblk->tag_set); > > if (err) > > > > So wrt cleanup, does something poll for all buffers to be > used when device is removed? Sorry for late reply. Maybe below function calls iterate each HW queue and flush requests before device is removed? ----- virtblk_remove() -> blk_cleanup_disk()/blk_cleanup_queue() -> blk_queue_start_drain()/blk_freeze_queue() ----- Regards, Suwan Kim
On Thu, Mar 24, 2022 at 11:04:49PM +0900, Suwan Kim wrote: > +static int virtblk_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob) > +{ > + struct virtio_blk_vq *vq = hctx->driver_data; > + struct virtblk_req *vbr; > + unsigned long flags; > + unsigned int len; > + int found = 0; > + > + spin_lock_irqsave(&vq->lock, flags); > + > + while ((vbr = virtqueue_get_buf(vq->vq, &len)) != NULL) { > + struct request *req = blk_mq_rq_from_pdu(vbr); > + > + found++; > + if (!blk_mq_add_to_batch(req, iob, vbr->status, > + virtblk_complete_batch)) > + blk_mq_complete_request(req); > + } > + > + spin_unlock_irqrestore(&vq->lock, flags); virtblk_done() does: /* In case queue is stopped waiting for more buffers. */ if (req_done) blk_mq_start_stopped_hw_queues(vblk->disk->queue, true); Is the same thing needed here in virtblk_poll() so that stopped queues are restarted when requests complete?
On Mon, Mar 28, 2022 at 01:53:46PM +0100, Stefan Hajnoczi wrote: > On Thu, Mar 24, 2022 at 11:04:49PM +0900, Suwan Kim wrote: > > +static int virtblk_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob) > > +{ > > + struct virtio_blk_vq *vq = hctx->driver_data; > > + struct virtblk_req *vbr; > > + unsigned long flags; > > + unsigned int len; > > + int found = 0; > > + > > + spin_lock_irqsave(&vq->lock, flags); > > + > > + while ((vbr = virtqueue_get_buf(vq->vq, &len)) != NULL) { > > + struct request *req = blk_mq_rq_from_pdu(vbr); > > + > > + found++; > > + if (!blk_mq_add_to_batch(req, iob, vbr->status, > > + virtblk_complete_batch)) > > + blk_mq_complete_request(req); > > + } > > + > > + spin_unlock_irqrestore(&vq->lock, flags); > > virtblk_done() does: > > /* In case queue is stopped waiting for more buffers. */ > if (req_done) > blk_mq_start_stopped_hw_queues(vblk->disk->queue, true); > > Is the same thing needed here in virtblk_poll() so that stopped queues > are restarted when requests complete? I think you are right. I missed that. I just added blk_mq_start_stopped_hw_queues() to virtblk_poll as you commented and did performance test again. It showed higher peak performance than virtblk_poll without blk_mq_start_stopped_hw_queues(). I will add it in next version. Thanks for the comment! Regards, Suwan Kim
diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c index 8c415be86732..3d16f8b753e7 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 num_poll_queues; +module_param(num_poll_queues, uint, 0644); +MODULE_PARM_DESC(num_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 +572,13 @@ 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, num_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; + vblk->vqs = kmalloc_array(num_vqs, sizeof(*vblk->vqs), GFP_KERNEL); if (!vblk->vqs) return -ENOMEM; @@ -578,8 +592,13 @@ static int init_vq(struct virtio_blk *vblk) } for (i = 0; i < num_vqs; i++) { - callbacks[i] = virtblk_done; - snprintf(vblk->vqs[i].name, VQ_NAME_LEN, "req.%d", 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; } @@ -728,16 +747,87 @@ static const struct attribute_group *virtblk_attr_groups[] = { static int virtblk_map_queues(struct blk_mq_tag_set *set) { struct virtio_blk *vblk = set->driver_data; + int i, qoff; + + for (i = 0, qoff = 0; i < set->nr_maps; i++) { + struct blk_mq_queue_map *map = &set->map[i]; + + map->nr_queues = vblk->io_queues[i]; + map->queue_offset = qoff; + qoff += map->nr_queues; + + 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) + blk_mq_virtio_map_queues(&set->map[i], vblk->vdev, 0); + else + blk_mq_map_queues(&set->map[i]); + } + + return 0; +} + +static void virtblk_complete_batch(struct io_comp_batch *iob) +{ + struct request *req; + struct virtblk_req *vbr; - return blk_mq_virtio_map_queues(&set->map[HCTX_TYPE_DEFAULT], - vblk->vdev, 0); + rq_list_for_each(&iob->req_list, req) { + vbr = blk_mq_rq_to_pdu(req); + virtblk_unmap_data(req, vbr); + virtblk_cleanup_cmd(req); + } + blk_mq_end_request_batch(iob); +} + +static int virtblk_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob) +{ + struct virtio_blk_vq *vq = hctx->driver_data; + struct virtblk_req *vbr; + unsigned long flags; + unsigned int len; + int found = 0; + + spin_lock_irqsave(&vq->lock, flags); + + while ((vbr = virtqueue_get_buf(vq->vq, &len)) != NULL) { + struct request *req = blk_mq_rq_from_pdu(vbr); + + found++; + if (!blk_mq_add_to_batch(req, iob, vbr->status, + virtblk_complete_batch)) + blk_mq_complete_request(req); + } + + spin_unlock_irqrestore(&vq->lock, flags); + + return found; +} + +static int virtblk_init_hctx(struct blk_mq_hw_ctx *hctx, void *data, + unsigned int hctx_idx) +{ + struct virtio_blk *vblk = data; + struct virtio_blk_vq *vq = &vblk->vqs[hctx_idx]; + + WARN_ON(vblk->tag_set.tags[hctx_idx] != hctx->tags); + hctx->driver_data = vq; + return 0; } static const struct blk_mq_ops virtio_mq_ops = { .queue_rq = virtio_queue_rq, .commit_rqs = virtio_commit_rqs, + .init_hctx = virtblk_init_hctx, .complete = virtblk_request_done, .map_queues = virtblk_map_queues, + .poll = virtblk_poll, }; static unsigned int virtblk_queue_depth; @@ -816,6 +906,9 @@ static int virtblk_probe(struct virtio_device *vdev) sizeof(struct scatterlist) * VIRTIO_BLK_INLINE_SG_CNT; vblk->tag_set.driver_data = vblk; vblk->tag_set.nr_hw_queues = vblk->num_vqs; + vblk->tag_set.nr_maps = 1; + if (vblk->io_queues[HCTX_TYPE_POLL]) + vblk->tag_set.nr_maps = 3; err = blk_mq_alloc_tag_set(&vblk->tag_set); if (err)
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 = 380K, avg latency = 167.87us -- numjobs=2 : IOPS = 409K, avg latency = 312.6us -- numjobs=4 : IOPS = 413K, avg latency = 619.72us Reported-by: kernel test robot <lkp@intel.com> Signed-off-by: Suwan Kim <suwan.kim027@gmail.com> --- drivers/block/virtio_blk.c | 101 +++++++++++++++++++++++++++++++++++-- 1 file changed, 97 insertions(+), 4 deletions(-)