[PATCH V4 2/2] nvme-core: use xarray for ctrl ns tracking

[Christoph Hellwig]

On Sun, Jul 19, 2020 at 08:32:03PM -0700, Chaitanya Kulkarni wrote:
> This patch replaces the ctrl->namespaces tracking from linked list to
> xarray and improves the performance. For host even though
> nvme_find_get_ns() doesn't fall into the fast path yet it does for
> NVMeOF passthru patch-series which is currently under review. This
> prepares us to improve performance for future NVMeOF passthru backend
> since nvme_find_get_ns() uses same data structure as it does in target
> to find namespace in I/O patch from nsid specified in the nvme_rw_cmd.

І'm still not really sold on that rationale..

> +	if (xa_empty(&ctrl->namespaces)) {
>  		ret = -ENOTTY;
> -		goto out_unlock;
> +		goto out;
>  	}
> -
> -	ns = list_first_entry(&ctrl->namespaces, struct nvme_ns, list);
> -	if (ns != list_last_entry(&ctrl->namespaces, struct nvme_ns, list)) {
> -		dev_warn(ctrl->device,
> -			"NVME_IOCTL_IO_CMD not supported when multiple namespaces present!\n");
> -		ret = -EINVAL;
> -		goto out_unlock;
> +	xa_for_each(&ctrl->namespaces, idx, ns) {
> +		if (count > 0)
> +			goto err;
> +		count++;

How about factoring the check into a little
nvme_has_multiple_namespaces helper and avoid the backwards jumping
goto later on?

> +	rcu_read_lock();
> +	ns = xa_load(&ctrl->namespaces, nsid);
> +	ns = ns && kref_get_unless_zero(&ns->kref) ? ns : NULL;
> +	rcu_read_unlock();

The canonical way to write this is:

	rcu_read_lock();
	ns = xa_load(&ctrl->namespaces, nsid);
	if (ns && !kref_get_unless_zero(&ns->kref))
		ns = NULL;
	rcu_read_unlock();

> + out_unregister_nvm:

Maybe call this out_unregister_lightnvm to be a little more descriptive.

> +
> +	xa_lock(&ctrl->namespaces);
> +	xa_for_each(&ctrl->namespaces, idx, ns) {
> +		tnsid = ns->head->ns_id;
> +		if (tnsid > nsid || test_bit(NVME_NS_DEAD, &ns->flags)) {
> +			__xa_erase(&ctrl->namespaces, tnsid);
> +			xa_unlock(&ctrl->namespaces);
> +			/* Even if insert fails keep going */
> +			ret = xa_insert(&rm_array, nsid, ns, GFP_KERNEL);
> +			if (ret)
> +				pr_err("xa_insert %d\n", ret);
> +			xa_lock(&ctrl->namespaces);
> +		}

And this is where I'm still worried that we now need an insert
as part of the namespace deletion.  Either keep the list here as
I suggested before, or we just need rework the deletion here first.


> @@ -4119,11 +4114,18 @@ void nvme_remove_namespaces(struct nvme_ctrl *ctrl)
>  	if (ctrl->state == NVME_CTRL_DEAD)
>  		nvme_kill_queues(ctrl);
>  
> -	down_write(&ctrl->namespaces_rwsem);
> -	list_splice_init(&ctrl->namespaces, &ns_list);
> -	up_write(&ctrl->namespaces_rwsem);
> +	xa_lock(&ctrl->namespaces);
> +	xa_for_each(&ctrl->namespaces, idx, ns) {
> +		__xa_erase(&ctrl->namespaces, ns->head->ns_id);
> +		xa_unlock(&ctrl->namespaces);
> +		ret = xa_insert(&rm_array, ns->head->ns_id, ns, GFP_KERNEL);
> +		if (ret)
> +			pr_err("xa_insert %d\n", ret);
> +		xa_lock(&ctrl->namespaces);
> +	}
> +	xa_unlock(&ctrl->namespaces);

Same here.