Re: [PATCH v2 1/1] virtio-blk: add num_io_queues module parameter

[Max Gurtovoy <mgurtovoy@nvidia.com>]

On 9/9/2021 6:40 PM, Michael S. Tsirkin wrote:
> On Thu, Sep 09, 2021 at 06:37:37PM +0300, Max Gurtovoy wrote:
>> On 9/9/2021 4:42 PM, Michael S. Tsirkin wrote:
>>> On Mon, Sep 06, 2021 at 02:59:40PM +0300, Max Gurtovoy wrote:
>>>> On 9/6/2021 2:20 PM, Michael S. Tsirkin wrote:
>>>>> On Mon, Sep 06, 2021 at 01:31:32AM +0300, Max Gurtovoy wrote:
>>>>>> On 9/5/2021 7:02 PM, Michael S. Tsirkin wrote:
>>>>>>> On Thu, Sep 02, 2021 at 02:45:52PM +0100, Stefan Hajnoczi wrote:
>>>>>>>> On Tue, Aug 31, 2021 at 04:50:35PM +0300, Max Gurtovoy wrote:
>>>>>>>>> Sometimes a user would like to control the amount of IO queues to be
>>>>>>>>> created for a block device. For example, for limiting the memory
>>>>>>>>> footprint of virtio-blk devices.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Max Gurtovoy <mgurtovoy@nvidia.com>
>>>>>>>>> ---
>>>>>>>>>
>>>>>>>>> changes from v1:
>>>>>>>>>      - use param_set_uint_minmax (from Christoph)
>>>>>>>>>      - added "Should > 0" to module description
>>>>>>>>>
>>>>>>>>> Note: This commit apply on top of Jens's branch for-5.15/drivers
>>>>>>>>> ---
>>>>>>>>>      drivers/block/virtio_blk.c | 20 +++++++++++++++++++-
>>>>>>>>>      1 file changed, 19 insertions(+), 1 deletion(-)
>>>>>>>>>
>>>>>>>>> diff --git a/drivers/block/virtio_blk.c b/drivers/block/virtio_blk.c
>>>>>>>>> index 4b49df2dfd23..9332fc4e9b31 100644
>>>>>>>>> --- a/drivers/block/virtio_blk.c
>>>>>>>>> +++ b/drivers/block/virtio_blk.c
>>>>>>>>> @@ -24,6 +24,22 @@
>>>>>>>>>      /* The maximum number of sg elements that fit into a virtqueue */
>>>>>>>>>      #define VIRTIO_BLK_MAX_SG_ELEMS 32768
>>>>>>>>> +static int virtblk_queue_count_set(const char *val,
>>>>>>>>> +		const struct kernel_param *kp)
>>>>>>>>> +{
>>>>>>>>> +	return param_set_uint_minmax(val, kp, 1, nr_cpu_ids);
>>>>>>>>> +}
>>>>> Hmm which tree is this for?
>>>> I've mentioned in the note that it apply on branch for-5.15/drivers. But now
>>>> you can apply it on linus/master as well.
>>>>
>>>>
>>>>>>>>> +
>>>>>>>>> +static const struct kernel_param_ops queue_count_ops = {
>>>>>>>>> +	.set = virtblk_queue_count_set,
>>>>>>>>> +	.get = param_get_uint,
>>>>>>>>> +};
>>>>>>>>> +
>>>>>>>>> +static unsigned int num_io_queues;
>>>>>>>>> +module_param_cb(num_io_queues, &queue_count_ops, &num_io_queues, 0644);
>>>>>>>>> +MODULE_PARM_DESC(num_io_queues,
>>>>>>>>> +		 "Number of IO virt queues to use for blk device. Should > 0");
>>>>> better:
>>>>>
>>>>> +MODULE_PARM_DESC(num_io_request_queues,
>>>>> +                "Limit number of IO request virt queues to use for each device. 0 for now limit");
>>>> You proposed it and I replied on it bellow.
>>> Can't say I understand 100% what you are saying. Want to send
>>> a description that does make sense to you but
>>> also reflects reality? 0 is the default so
>>> "should > 0" besides being ungrammatical does not seem t"
>>> reflect what it does ...
>> if you "modprobe virtio_blk" the previous behavior will happen.
>>
>> You can't "modprobe virtio_blk num_io_request_queues=0" since the minimal
>> value is 1.
>>
>> So your description is not reflecting the code.
>>
>> We can do:
>>
>> MODULE_PARM_DESC(num_io_request_queues, "Number of request virt queues to use for blk device. Minimum value is 1 queue");
> What's the default value? We should document that.

default value for static global variables is 0.

MODULE_PARM_DESC(num_io_request_queues, "Number of request virt queues 
to use for blk device. Minimum value is 1 queue. Default and Maximum 
value is equal to the total number of CPUs");


>
>>>>>>>>> +
>>>>>>>>>      static int major;
>>>>>>>>>      static DEFINE_IDA(vd_index_ida);
>>>>>>>>> @@ -501,7 +517,9 @@ static int init_vq(struct virtio_blk *vblk)
>>>>>>>>>      	if (err)
>>>>>>>>>      		num_vqs = 1;
>>>>>>>>> -	num_vqs = min_t(unsigned int, nr_cpu_ids, num_vqs);
>>>>>>>>> +	num_vqs = min_t(unsigned int,
>>>>>>>>> +			min_not_zero(num_io_queues, nr_cpu_ids),
>>>>>>>>> +			num_vqs);
>>>>>>>> If you respin, please consider calling them request queues. That's the
>>>>>>>> terminology from the VIRTIO spec and it's nice to keep it consistent.
>>>>>>>> But the purpose of num_io_queues is clear, so:
>>>>>>>>
>>>>>>>> Reviewed-by: Stefan Hajnoczi <stefanha@redhat.com>
>>>>>>> I did this:
>>>>>>> +static unsigned int num_io_request_queues;
>>>>>>> +module_param_cb(num_io_request_queues, &queue_count_ops, &num_io_request_queues, 0644);
>>>>>>> +MODULE_PARM_DESC(num_io_request_queues,
>>>>>>> +                "Limit number of IO request virt queues to use for each device. 0 for now limit");
>>>>>> The parameter is writable and can be changed and then new devices might be
>>>>>> probed with new value.
>>>>>>
>>>>>> It can't be zero in the code. we can change param_set_uint_minmax args and
>>>>>> say that 0 says nr_cpus.
>>>>>>
>>>>>> I'm ok with the renaming but I prefer to stick to the description we gave in
>>>>>> V3 of this patch (and maybe enable value of 0 as mentioned above).