Re: [PATCH] virtio-blk: add DISCARD support to virtio-blk driver

From: Stefan Hajnoczi
Date: Tue Mar 28 2017 - 04:38:17 EST


On Tue, Mar 28, 2017 at 3:15 AM, Liu, Changpeng <changpeng.liu@xxxxxxxxx> wrote:
>> -----Original Message-----
>> From: Stefan Hajnoczi [mailto:stefanha@xxxxxxxxx]
>> Sent: Tuesday, March 28, 2017 4:20 AM
>> To: Liu, Changpeng <changpeng.liu@xxxxxxxxx>
>> Cc: virtio-dev@xxxxxxxxxxxxxxxxxxxx; virtualization@xxxxxxxxxxxxxxxxxxxxxxxxxx; linux-
>> kernel@xxxxxxxxxxxxxxx; hch@xxxxxx; qemu-devel@xxxxxxxxxx
>> Subject: Re: [PATCH] virtio-blk: add DISCARD support to virtio-blk driver
>>
>> On Tue, Mar 28, 2017 at 04:39:25PM +0800, Changpeng Liu wrote:
>> > Currently virtio-blk driver does not provide discard feature flag, so the
>> > filesystems which built on top of the block device will not send discard
>> > command. This is okay for HDD backend, but it will impact the performance
>> > for SSD backend.
>> >
>> > Add a feature flag VIRTIO_BLK_F_DISCARD and command
>> VIRTIO_BLK_T_DISCARD
>> > to extend exist virtio-blk protocol. virtio-blk protocol uses a single
>> > 8 bytes descriptor containing type,reserved and sector, currently Linux
>> > uses the reserved field as IO priority, here we also re-use the reserved
>> > field as number of discard sectors.
>> >
>> > Signed-off-by: Changpeng Liu <changpeng.liu@xxxxxxxxx>
>> > ---
>> > drivers/block/virtio_blk.c | 38 +++++++++++++++++++++++++++++---------
>> > include/uapi/linux/virtio_blk.h | 12 ++++++++++--
>> > 2 files changed, 39 insertions(+), 11 deletions(-)
>> >
>> > diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
>> > index 1d4c9f8..550cfe7 100644
>> > --- a/drivers/block/virtio_blk.c
>> > +++ b/drivers/block/virtio_blk.c
>> > @@ -241,6 +241,9 @@ static int virtio_queue_rq(struct blk_mq_hw_ctx *hctx,
>> > case REQ_OP_FLUSH:
>> > type = VIRTIO_BLK_T_FLUSH;
>> > break;
>> > + case REQ_OP_DISCARD:
>> > + type = VIRTIO_BLK_T_DISCARD;
>> > + break;
>> > case REQ_OP_SCSI_IN:
>> > case REQ_OP_SCSI_OUT:
>> > type = VIRTIO_BLK_T_SCSI_CMD;
>> > @@ -256,16 +259,24 @@ static int virtio_queue_rq(struct blk_mq_hw_ctx *hctx,
>> > vbr->out_hdr.type = cpu_to_virtio32(vblk->vdev, type);
>> > vbr->out_hdr.sector = type ?
>> > 0 : cpu_to_virtio64(vblk->vdev, blk_rq_pos(req));
>> > - vbr->out_hdr.ioprio = cpu_to_virtio32(vblk->vdev, req_get_ioprio(req));
>> > + vbr->out_hdr.u.ioprio = cpu_to_virtio32(vblk->vdev, req_get_ioprio(req));
>> >
>> > blk_mq_start_request(req);
>> >
>> > - num = blk_rq_map_sg(hctx->queue, req, vbr->sg);
>> > - if (num) {
>> > - if (rq_data_dir(req) == WRITE)
>> > - vbr->out_hdr.type |= cpu_to_virtio32(vblk->vdev,
>> VIRTIO_BLK_T_OUT);
>> > - else
>> > - vbr->out_hdr.type |= cpu_to_virtio32(vblk->vdev,
>> VIRTIO_BLK_T_IN);
>> > + if (type == VIRTIO_BLK_T_DISCARD) {
>> > + vbr->out_hdr.u.discard_nr_sectors = cpu_to_virtio32(vblk->vdev,
>> > +
>> blk_rq_sectors(req));
>> > + num = 0;
>> > + } else {
>> > + num = blk_rq_map_sg(hctx->queue, req, vbr->sg);
>> > + if (num) {
>> > + if (rq_data_dir(req) == WRITE)
>> > + vbr->out_hdr.type |= cpu_to_virtio32(vblk->vdev,
>> > +
>> VIRTIO_BLK_T_OUT);
>> > + else
>> > + vbr->out_hdr.type |= cpu_to_virtio32(vblk->vdev,
>> > +
>> VIRTIO_BLK_T_IN);
>> > + }
>> > }
>> >
>> > spin_lock_irqsave(&vblk->vqs[qid].lock, flags);
>> > @@ -775,6 +786,15 @@ static int virtblk_probe(struct virtio_device *vdev)
>> > if (!err && opt_io_size)
>> > blk_queue_io_opt(q, blk_size * opt_io_size);
>> >
>> > + if (virtio_has_feature(vdev, VIRTIO_BLK_F_DISCARD)) {
>> > + q->limits.discard_zeroes_data = 0;
>> > + q->limits.discard_alignment = blk_size;
>> > + q->limits.discard_granularity = blk_size;
>> > + blk_queue_max_discard_sectors(q, UINT_MAX);
>> > + blk_queue_max_discard_segments(q, 1);
>> > + queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
>> > + }
>>
>> Please add configuration space fields for these limits. Looking at the
>> virtio-scsi block limits code in QEMU's scsi_disk_emulate_inquiry() I
>> can see that the hypervisor has useful values that it wants to
>> communicate. They shouldn't be hardcoded to blk_size.
> Yes, move discard related parameters to configuration space make sense.
>>
>> > +
>> > virtio_device_ready(vdev);
>> >
>> > device_add_disk(&vdev->dev, vblk->disk);
>> > @@ -882,14 +902,14 @@ static int virtblk_restore(struct virtio_device *vdev)
>> > VIRTIO_BLK_F_SCSI,
>> > #endif
>> > VIRTIO_BLK_F_FLUSH, VIRTIO_BLK_F_TOPOLOGY,
>> VIRTIO_BLK_F_CONFIG_WCE,
>> > - VIRTIO_BLK_F_MQ,
>> > + VIRTIO_BLK_F_MQ, VIRTIO_BLK_F_DISCARD,
>> > }
>> > ;
>> > static unsigned int features[] = {
>> > VIRTIO_BLK_F_SEG_MAX, VIRTIO_BLK_F_SIZE_MAX,
>> VIRTIO_BLK_F_GEOMETRY,
>> > VIRTIO_BLK_F_RO, VIRTIO_BLK_F_BLK_SIZE,
>> > VIRTIO_BLK_F_FLUSH, VIRTIO_BLK_F_TOPOLOGY,
>> VIRTIO_BLK_F_CONFIG_WCE,
>> > - VIRTIO_BLK_F_MQ,
>> > + VIRTIO_BLK_F_MQ, VIRTIO_BLK_F_DISCARD,
>> > };
>> >
>> > static struct virtio_driver virtio_blk = {
>> > diff --git a/include/uapi/linux/virtio_blk.h b/include/uapi/linux/virtio_blk.h
>> > index 9ebe4d9..d608649 100644
>> > --- a/include/uapi/linux/virtio_blk.h
>> > +++ b/include/uapi/linux/virtio_blk.h
>> > @@ -38,6 +38,7 @@
>> > #define VIRTIO_BLK_F_BLK_SIZE 6 /* Block size of disk is available*/
>> > #define VIRTIO_BLK_F_TOPOLOGY 10 /* Topology information is
>> available */
>> > #define VIRTIO_BLK_F_MQ 12 /* support more than one vq */
>> > +#define VIRTIO_BLK_F_DISCARD 13 /* DISCARD command is supported
>> */
>> >
>> > /* Legacy feature bits */
>> > #ifndef VIRTIO_BLK_NO_LEGACY
>> > @@ -114,6 +115,9 @@ struct virtio_blk_config {
>> > /* Get device ID command */
>> > #define VIRTIO_BLK_T_GET_ID 8
>> >
>> > +/* Discard command */
>> > +#define VIRTIO_BLK_T_DISCARD 16
>> > +
>> > #ifndef VIRTIO_BLK_NO_LEGACY
>> > /* Barrier before this op. */
>> > #define VIRTIO_BLK_T_BARRIER 0x80000000
>> > @@ -127,8 +131,12 @@ struct virtio_blk_config {
>> > struct virtio_blk_outhdr {
>> > /* VIRTIO_BLK_T* */
>> > __virtio32 type;
>> > - /* io priority. */
>> > - __virtio32 ioprio;
>> > + union {
>> > + /* io priority. */
>> > + __virtio32 ioprio;
>> > + /* discard number of sectors */
>> > + __virtio32 discard_nr_sectors;
>> > + } u;
>>
>> DISCARD commands have no io priority? Perhaps it's better to add an
>> extended header.
> What I think now is that, keep the ioprio field, and let DISCARD command has input data buffers, 16 bytes aligned descriptor for each
> DISCARD segment(can support mult-range feature), and this is also aligned with SCSI and NVMe specification.

Sounds good, thanks.

Stefan