Re: [PATCH RFC] virtio-blk: support per-device queue depth

Re: [PATCH RFC] virtio-blk: support per-device queue depth

[Jason Wang]

On 2021/1/18 上午11:58, Joseph Qi wrote:
> module parameter 'virtblk_queue_depth' was firstly introduced for
> testing/benchmarking purposes described in commit fc4324b4597c
> ("virtio-blk: base queue-depth on virtqueue ringsize or module param").
> Since we have different virtio-blk devices which have different
> capabilities, it requires that we support per-device queue depth instead
> of per-module. So defaultly use vq free elements if module parameter
> 'virtblk_queue_depth' is not set.


I wonder if it's better to use sysfs instead (or whether it has already 
had something like this in the blocker layer).

Thanks


>
> Signed-off-by: Joseph Qi <joseph.qi@linux.alibaba.com>
> ---
>   drivers/block/virtio_blk.c | 12 +++++++-----
>   1 file changed, 7 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
> index 145606d..f83a417 100644
> --- a/drivers/block/virtio_blk.c
> +++ b/drivers/block/virtio_blk.c
> @@ -705,6 +705,7 @@ static int virtblk_probe(struct virtio_device *vdev)
>   	u32 v, blk_size, max_size, sg_elems, opt_io_size;
>   	u16 min_io_size;
>   	u8 physical_block_exp, alignment_offset;
> +	unsigned int queue_depth;
>   
>   	if (!vdev->config->get) {
>   		dev_err(&vdev->dev, "%s failure: config access disabled\n",
> @@ -755,17 +756,18 @@ static int virtblk_probe(struct virtio_device *vdev)
>   		goto out_free_vq;
>   	}
>   
> -	/* Default queue sizing is to fill the ring. */
> -	if (!virtblk_queue_depth) {
> -		virtblk_queue_depth = vblk->vqs[0].vq->num_free;
> +	if (likely(!virtblk_queue_depth)) {
> +		queue_depth = vblk->vqs[0].vq->num_free;
>   		/* ... but without indirect descs, we use 2 descs per req */
>   		if (!virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC))
> -			virtblk_queue_depth /= 2;
> +			queue_depth /= 2;
> +	} else {
> +		queue_depth = virtblk_queue_depth;
>   	}
>   
>   	memset(&vblk->tag_set, 0, sizeof(vblk->tag_set));
>   	vblk->tag_set.ops = &virtio_mq_ops;
> -	vblk->tag_set.queue_depth = virtblk_queue_depth;
> +	vblk->tag_set.queue_depth = queue_depth;
>   	vblk->tag_set.numa_node = NUMA_NO_NODE;
>   	vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
>   	vblk->tag_set.cmd_size =

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH RFC] virtio-blk: support per-device queue depth

[JeffleXu]

On 1/18/21 1:25 PM, Jason Wang wrote:
> 
> On 2021/1/18 上午11:58, Joseph Qi wrote:
>> module parameter 'virtblk_queue_depth' was firstly introduced for
>> testing/benchmarking purposes described in commit fc4324b4597c
>> ("virtio-blk: base queue-depth on virtqueue ringsize or module param").
>> Since we have different virtio-blk devices which have different
>> capabilities, it requires that we support per-device queue depth instead
>> of per-module. So defaultly use vq free elements if module parameter
>> 'virtblk_queue_depth' is not set.
> 
> 
> I wonder if it's better to use sysfs instead (or whether it has already
> had something like this in the blocker layer).
> 

"/sys/block/<dev>/queue/nr_requests" indeed works, but isn't better to
set queue_depth according to the hardware capability at the very first?
AFAIK, nvme just set per-device queue_depth at initializing phase.

Thanks,
Jeffle

> 
> 
>>
>> Signed-off-by: Joseph Qi <joseph.qi@linux.alibaba.com>
>> ---
>>   drivers/block/virtio_blk.c | 12 +++++++-----
>>   1 file changed, 7 insertions(+), 5 deletions(-)
>>
>> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
>> index 145606d..f83a417 100644
>> --- a/drivers/block/virtio_blk.c
>> +++ b/drivers/block/virtio_blk.c
>> @@ -705,6 +705,7 @@ static int virtblk_probe(struct virtio_device *vdev)
>>       u32 v, blk_size, max_size, sg_elems, opt_io_size;
>>       u16 min_io_size;
>>       u8 physical_block_exp, alignment_offset;
>> +    unsigned int queue_depth;
>>         if (!vdev->config->get) {
>>           dev_err(&vdev->dev, "%s failure: config access disabled\n",
>> @@ -755,17 +756,18 @@ static int virtblk_probe(struct virtio_device
>> *vdev)
>>           goto out_free_vq;
>>       }
>>   -    /* Default queue sizing is to fill the ring. */
>> -    if (!virtblk_queue_depth) {
>> -        virtblk_queue_depth = vblk->vqs[0].vq->num_free;
>> +    if (likely(!virtblk_queue_depth)) {
>> +        queue_depth = vblk->vqs[0].vq->num_free;
>>           /* ... but without indirect descs, we use 2 descs per req */
>>           if (!virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC))
>> -            virtblk_queue_depth /= 2;
>> +            queue_depth /= 2;
>> +    } else {
>> +        queue_depth = virtblk_queue_depth;
>>       }
>>         memset(&vblk->tag_set, 0, sizeof(vblk->tag_set));
>>       vblk->tag_set.ops = &virtio_mq_ops;
>> -    vblk->tag_set.queue_depth = virtblk_queue_depth;
>> +    vblk->tag_set.queue_depth = queue_depth;
>>       vblk->tag_set.numa_node = NUMA_NO_NODE;
>>       vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
>>       vblk->tag_set.cmd_size =

-- 
Thanks,
Jeffle
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH RFC] virtio-blk: support per-device queue depth

[Jason Wang]

On 2021/1/19 上午9:33, JeffleXu wrote:
>
> On 1/18/21 1:25 PM, Jason Wang wrote:
>> On 2021/1/18 上午11:58, Joseph Qi wrote:
>>> module parameter 'virtblk_queue_depth' was firstly introduced for
>>> testing/benchmarking purposes described in commit fc4324b4597c
>>> ("virtio-blk: base queue-depth on virtqueue ringsize or module param").
>>> Since we have different virtio-blk devices which have different
>>> capabilities, it requires that we support per-device queue depth instead
>>> of per-module. So defaultly use vq free elements if module parameter
>>> 'virtblk_queue_depth' is not set.
>>
>> I wonder if it's better to use sysfs instead (or whether it has already
>> had something like this in the blocker layer).
>>
> "/sys/block/<dev>/queue/nr_requests" indeed works, but isn't better to
> set queue_depth according to the hardware capability at the very first?
> AFAIK, nvme just set per-device queue_depth at initializing phase.


I agree, the problem is that the current code may modify module parameter.

Thanks


>
> Thanks,
> Jeffle
>
>>
>>> Signed-off-by: Joseph Qi <joseph.qi@linux.alibaba.com>
>>> ---
>>>    drivers/block/virtio_blk.c | 12 +++++++-----
>>>    1 file changed, 7 insertions(+), 5 deletions(-)
>>>
>>> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
>>> index 145606d..f83a417 100644
>>> --- a/drivers/block/virtio_blk.c
>>> +++ b/drivers/block/virtio_blk.c
>>> @@ -705,6 +705,7 @@ static int virtblk_probe(struct virtio_device *vdev)
>>>        u32 v, blk_size, max_size, sg_elems, opt_io_size;
>>>        u16 min_io_size;
>>>        u8 physical_block_exp, alignment_offset;
>>> +    unsigned int queue_depth;
>>>          if (!vdev->config->get) {
>>>            dev_err(&vdev->dev, "%s failure: config access disabled\n",
>>> @@ -755,17 +756,18 @@ static int virtblk_probe(struct virtio_device
>>> *vdev)
>>>            goto out_free_vq;
>>>        }
>>>    -    /* Default queue sizing is to fill the ring. */
>>> -    if (!virtblk_queue_depth) {
>>> -        virtblk_queue_depth = vblk->vqs[0].vq->num_free;
>>> +    if (likely(!virtblk_queue_depth)) {
>>> +        queue_depth = vblk->vqs[0].vq->num_free;
>>>            /* ... but without indirect descs, we use 2 descs per req */
>>>            if (!virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC))
>>> -            virtblk_queue_depth /= 2;
>>> +            queue_depth /= 2;
>>> +    } else {
>>> +        queue_depth = virtblk_queue_depth;
>>>        }
>>>          memset(&vblk->tag_set, 0, sizeof(vblk->tag_set));
>>>        vblk->tag_set.ops = &virtio_mq_ops;
>>> -    vblk->tag_set.queue_depth = virtblk_queue_depth;
>>> +    vblk->tag_set.queue_depth = queue_depth;
>>>        vblk->tag_set.numa_node = NUMA_NO_NODE;
>>>        vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
>>>        vblk->tag_set.cmd_size =

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH RFC] virtio-blk: support per-device queue depth

[JeffleXu]

On 1/19/21 12:06 PM, Jason Wang wrote:
> 
> On 2021/1/19 上午9:33, JeffleXu wrote:
>>
>> On 1/18/21 1:25 PM, Jason Wang wrote:
>>> On 2021/1/18 上午11:58, Joseph Qi wrote:
>>>> module parameter 'virtblk_queue_depth' was firstly introduced for
>>>> testing/benchmarking purposes described in commit fc4324b4597c
>>>> ("virtio-blk: base queue-depth on virtqueue ringsize or module param").
>>>> Since we have different virtio-blk devices which have different
>>>> capabilities, it requires that we support per-device queue depth
>>>> instead
>>>> of per-module. So defaultly use vq free elements if module parameter
>>>> 'virtblk_queue_depth' is not set.
>>>
>>> I wonder if it's better to use sysfs instead (or whether it has already
>>> had something like this in the blocker layer).
>>>
>> "/sys/block/<dev>/queue/nr_requests" indeed works, but isn't better to
>> set queue_depth according to the hardware capability at the very first?
>> AFAIK, nvme just set per-device queue_depth at initializing phase.
> 
> 
> I agree, the problem is that the current code may modify module parameter.

The module parameter 'virtblk_queue_depth' is actually remained untainted.

Actually it is the original code before this patch that changes the
module parameter. When the module parameter is not set by boot cmdline
(i.e., default to 0), it will be initialized to the queue_depth of the
vring of the first probed virtio-blk device, and will be revealed to
user space through '/sys/module/virtio_blk/parameters/queue_depth'. I'm
not sure if this behavior is reasonable or not.

The only side effect of this patch is that, now
'/sys/module/virtio_blk/parameters/queue_depth' will be kept as '0' when
the module parameter is not set manually.


Thanks,
Jeffle


>>
>>>
>>>> Signed-off-by: Joseph Qi <joseph.qi@linux.alibaba.com>
>>>> ---
>>>>    drivers/block/virtio_blk.c | 12 +++++++-----
>>>>    1 file changed, 7 insertions(+), 5 deletions(-)
>>>>
>>>> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
>>>> index 145606d..f83a417 100644
>>>> --- a/drivers/block/virtio_blk.c
>>>> +++ b/drivers/block/virtio_blk.c
>>>> @@ -705,6 +705,7 @@ static int virtblk_probe(struct virtio_device
>>>> *vdev)
>>>>        u32 v, blk_size, max_size, sg_elems, opt_io_size;
>>>>        u16 min_io_size;
>>>>        u8 physical_block_exp, alignment_offset;
>>>> +    unsigned int queue_depth;
>>>>          if (!vdev->config->get) {
>>>>            dev_err(&vdev->dev, "%s failure: config access disabled\n",
>>>> @@ -755,17 +756,18 @@ static int virtblk_probe(struct virtio_device
>>>> *vdev)
>>>>            goto out_free_vq;
>>>>        }
>>>>    -    /* Default queue sizing is to fill the ring. */
>>>> -    if (!virtblk_queue_depth) {
>>>> -        virtblk_queue_depth = vblk->vqs[0].vq->num_free;
>>>> +    if (likely(!virtblk_queue_depth)) {
>>>> +        queue_depth = vblk->vqs[0].vq->num_free;
>>>>            /* ... but without indirect descs, we use 2 descs per req */
>>>>            if (!virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC))
>>>> -            virtblk_queue_depth /= 2;
>>>> +            queue_depth /= 2;
>>>> +    } else {
>>>> +        queue_depth = virtblk_queue_depth;
>>>>        }
>>>>          memset(&vblk->tag_set, 0, sizeof(vblk->tag_set));
>>>>        vblk->tag_set.ops = &virtio_mq_ops;
>>>> -    vblk->tag_set.queue_depth = virtblk_queue_depth;
>>>> +    vblk->tag_set.queue_depth = queue_depth;
>>>>        vblk->tag_set.numa_node = NUMA_NO_NODE;
>>>>        vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
>>>>        vblk->tag_set.cmd_size =

-- 
Thanks,
Jeffle
_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH RFC] virtio-blk: support per-device queue depth

[Jason Wang]

On 2021/1/20 上午9:51, JeffleXu wrote:
>
> On 1/19/21 12:06 PM, Jason Wang wrote:
>> On 2021/1/19 上午9:33, JeffleXu wrote:
>>> On 1/18/21 1:25 PM, Jason Wang wrote:
>>>> On 2021/1/18 上午11:58, Joseph Qi wrote:
>>>>> module parameter 'virtblk_queue_depth' was firstly introduced for
>>>>> testing/benchmarking purposes described in commit fc4324b4597c
>>>>> ("virtio-blk: base queue-depth on virtqueue ringsize or module param").
>>>>> Since we have different virtio-blk devices which have different
>>>>> capabilities, it requires that we support per-device queue depth
>>>>> instead
>>>>> of per-module. So defaultly use vq free elements if module parameter
>>>>> 'virtblk_queue_depth' is not set.
>>>> I wonder if it's better to use sysfs instead (or whether it has already
>>>> had something like this in the blocker layer).
>>>>
>>> "/sys/block/<dev>/queue/nr_requests" indeed works, but isn't better to
>>> set queue_depth according to the hardware capability at the very first?
>>> AFAIK, nvme just set per-device queue_depth at initializing phase.
>>
>> I agree, the problem is that the current code may modify module parameter.
> The module parameter 'virtblk_queue_depth' is actually remained untainted.
>
> Actually it is the original code before this patch that changes the
> module parameter.


Yes, that's what I meant.


> When the module parameter is not set by boot cmdline
> (i.e., default to 0), it will be initialized to the queue_depth of the
> vring of the first probed virtio-blk device, and will be revealed to
> user space through '/sys/module/virtio_blk/parameters/queue_depth'. I'm
> not sure if this behavior is reasonable or not.


Right, it means the virtio-blk devices that is probed after the first 
one can only use the queue_depth that is set according to the capability 
of the first virtio-blk device.


>
> The only side effect of this patch is that, now
> '/sys/module/virtio_blk/parameters/queue_depth' will be kept as '0' when
> the module parameter is not set manually.


I think it's not an issue, the nr_request should be the correct way to  
get per device queue depth.

Thanks


>
>
> Thanks,
> Jeffle
>
>
>>>>> Signed-off-by: Joseph Qi <joseph.qi@linux.alibaba.com>
>>>>> ---
>>>>>     drivers/block/virtio_blk.c | 12 +++++++-----
>>>>>     1 file changed, 7 insertions(+), 5 deletions(-)
>>>>>
>>>>> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
>>>>> index 145606d..f83a417 100644
>>>>> --- a/drivers/block/virtio_blk.c
>>>>> +++ b/drivers/block/virtio_blk.c
>>>>> @@ -705,6 +705,7 @@ static int virtblk_probe(struct virtio_device
>>>>> *vdev)
>>>>>         u32 v, blk_size, max_size, sg_elems, opt_io_size;
>>>>>         u16 min_io_size;
>>>>>         u8 physical_block_exp, alignment_offset;
>>>>> +    unsigned int queue_depth;
>>>>>           if (!vdev->config->get) {
>>>>>             dev_err(&vdev->dev, "%s failure: config access disabled\n",
>>>>> @@ -755,17 +756,18 @@ static int virtblk_probe(struct virtio_device
>>>>> *vdev)
>>>>>             goto out_free_vq;
>>>>>         }
>>>>>     -    /* Default queue sizing is to fill the ring. */
>>>>> -    if (!virtblk_queue_depth) {
>>>>> -        virtblk_queue_depth = vblk->vqs[0].vq->num_free;
>>>>> +    if (likely(!virtblk_queue_depth)) {
>>>>> +        queue_depth = vblk->vqs[0].vq->num_free;
>>>>>             /* ... but without indirect descs, we use 2 descs per req */
>>>>>             if (!virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC))
>>>>> -            virtblk_queue_depth /= 2;
>>>>> +            queue_depth /= 2;
>>>>> +    } else {
>>>>> +        queue_depth = virtblk_queue_depth;
>>>>>         }
>>>>>           memset(&vblk->tag_set, 0, sizeof(vblk->tag_set));
>>>>>         vblk->tag_set.ops = &virtio_mq_ops;
>>>>> -    vblk->tag_set.queue_depth = virtblk_queue_depth;
>>>>> +    vblk->tag_set.queue_depth = queue_depth;
>>>>>         vblk->tag_set.numa_node = NUMA_NO_NODE;
>>>>>         vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
>>>>>         vblk->tag_set.cmd_size =

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH RFC] virtio-blk: support per-device queue depth

[Jason Wang]

On 2021/1/18 下午2:06, Joseph Qi wrote:
> Hi Jason,
>
> On 1/18/21 1:25 PM, Jason Wang wrote:
>> On 2021/1/18 上午11:58, Joseph Qi wrote:
>>> module parameter 'virtblk_queue_depth' was firstly introduced for
>>> testing/benchmarking purposes described in commit fc4324b4597c
>>> ("virtio-blk: base queue-depth on virtqueue ringsize or module param").
>>> Since we have different virtio-blk devices which have different
>>> capabilities, it requires that we support per-device queue depth instead
>>> of per-module. So defaultly use vq free elements if module parameter
>>> 'virtblk_queue_depth' is not set.
>>
>> I wonder if it's better to use sysfs instead (or whether it has already had something like this in the blocker layer).
>>
> Thanks for quick response.
> Do you mean adjust /sys/block/vdX/queue/nr_requests?
> But current logic in virtblk_probe() is, virtblk_queue_depth is
> used as a saved value for first probed vdev, not purely module
> parameter.


Right, I see. So I think the patch is fine.

Thanks


>
> Thanks,
> Joseph
>

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH RFC] virtio-blk: support per-device queue depth

[Jason Wang]

On 2021/1/18 上午11:58, Joseph Qi wrote:
> module parameter 'virtblk_queue_depth' was firstly introduced for
> testing/benchmarking purposes described in commit fc4324b4597c
> ("virtio-blk: base queue-depth on virtqueue ringsize or module param").
> Since we have different virtio-blk devices which have different
> capabilities, it requires that we support per-device queue depth instead
> of per-module. So defaultly use vq free elements if module parameter
> 'virtblk_queue_depth' is not set.
>
> Signed-off-by: Joseph Qi <joseph.qi@linux.alibaba.com>


Acked-by: Jason Wang <jasowang@redhat.com>


> ---
>   drivers/block/virtio_blk.c | 12 +++++++-----
>   1 file changed, 7 insertions(+), 5 deletions(-)
>
> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
> index 145606d..f83a417 100644
> --- a/drivers/block/virtio_blk.c
> +++ b/drivers/block/virtio_blk.c
> @@ -705,6 +705,7 @@ static int virtblk_probe(struct virtio_device *vdev)
>   	u32 v, blk_size, max_size, sg_elems, opt_io_size;
>   	u16 min_io_size;
>   	u8 physical_block_exp, alignment_offset;
> +	unsigned int queue_depth;
>   
>   	if (!vdev->config->get) {
>   		dev_err(&vdev->dev, "%s failure: config access disabled\n",
> @@ -755,17 +756,18 @@ static int virtblk_probe(struct virtio_device *vdev)
>   		goto out_free_vq;
>   	}
>   
> -	/* Default queue sizing is to fill the ring. */
> -	if (!virtblk_queue_depth) {
> -		virtblk_queue_depth = vblk->vqs[0].vq->num_free;
> +	if (likely(!virtblk_queue_depth)) {
> +		queue_depth = vblk->vqs[0].vq->num_free;
>   		/* ... but without indirect descs, we use 2 descs per req */
>   		if (!virtio_has_feature(vdev, VIRTIO_RING_F_INDIRECT_DESC))
> -			virtblk_queue_depth /= 2;
> +			queue_depth /= 2;
> +	} else {
> +		queue_depth = virtblk_queue_depth;
>   	}
>   
>   	memset(&vblk->tag_set, 0, sizeof(vblk->tag_set));
>   	vblk->tag_set.ops = &virtio_mq_ops;
> -	vblk->tag_set.queue_depth = virtblk_queue_depth;
> +	vblk->tag_set.queue_depth = queue_depth;
>   	vblk->tag_set.numa_node = NUMA_NO_NODE;
>   	vblk->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
>   	vblk->tag_set.cmd_size =

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization

Re: [PATCH RFC] virtio-blk: support per-device queue depth

[Michael S. Tsirkin]

On Fri, Jan 22, 2021 at 09:43:27AM +0800, Joseph Qi wrote:
> Hi Michael,
> 
> Any comments on this patch?
> 
> Thanks,
> Joseph

Suggest copying all reviewers, including Paolo Bonzini
<pbonzini@redhat.com> and Stefan Hajnoczi <stefanha@redhat.com>.

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization