Re: [PATCH v7 1/1] nvme-multipath: implement "queue-depth" iopolicy

[Christoph Hellwig <hch@lst.de>]

On Thu, Jun 20, 2024 at 01:54:29PM -0400, John Meneghini wrote:
>>> +static void nvme_subsys_iopolicy_update(struct nvme_subsystem *subsys,
>>> +		int iopolicy)
>>> +{
>>> +	struct nvme_ctrl *ctrl;
>>> +	int old_iopolicy = READ_ONCE(subsys->iopolicy);
>>> +
>>> +	if (old_iopolicy == iopolicy)
>>> +		return;
>>> +
>>> +	WRITE_ONCE(subsys->iopolicy, iopolicy);
>>
>> What is the atomicy model here?  There doesn't seem to be any
>> global lock protecting it?  Maybe move it into the
>> nvme_subsystems_lock critical section?
>
> Good question.  I didn't write this code. Yes, I agree this looks racy. 
> Updates to the subsys->iopolicy variable are not atomic. They don't need to 
> be. The process of changing the iopolicy doesn't need to be synchronized 
> and each CPU's cache will be updated lazily. This was done to avoid the 
> expense of adding (another) atomic read the io path.

Looks like all sysfs ->store calls for the same attribute are protected
by of->mutex in kernfs_fop_write_iter and we should actually be fine
here.  Sorry for the noise.

>> 	pr_notice("%s: changing iopolicy from %s to %s\n",
>> 		subsys->subnqn,
>> 		nvme_iopolicy_names[old_iopolicy],
>> 		nvme_iopolicy_names[iopolicy]);
>
> How about:
>
> pr_notice("Changed iopolicy from %s to %s for subsysnqn %s\n",
>                 nvme_iopolicy_names[old_iopolicy],
>                 nvme_iopolicy_names[iopolicy],
>                 subsys->subnqn);

Having the identification as the prefixe seems easier to parse
and grep for.