[PATCH V2 1/2] nvme-core: use xarray for ctrl ns tracking

[Chaitanya Kulkarni]

On 7/1/20 6:12 AM, Christoph Hellwig wrote:
> [willy: a comment/request on the xa_load API below]
That will be great.
> 
> On Tue, Jun 30, 2020 at 07:25:16PM -0700, Chaitanya Kulkarni wrote:
>> This patch replaces the ctrl->namespaces tracking from linked list to
>> xarray and improves the performance.
> 
> The performance improvement needs to be clearly stated here.
> 

This is a preparation for the passthru code which uses 
nvme_find_get_ns() which falls in the fast path from passthru to host
core. Having linked list will have same performance issue with which
I've reported on for the NVMeOF gen-blk target using nvme-loop when we 
integrate passthru.

How about we get this series in a good shape and before you apply I'll 
forward port on the passthru V14 and document the performance numbers 
for both the non gen-blk and passthru NVMeOF target ?

OR you want to see the numbers now with the comments fixed in V3 ?

I'm fine either way.

>>   static int nvme_dev_user_cmd(struct nvme_ctrl *ctrl, void __user *argp)
>>   {
>> +	struct nvme_id_ctrl *id;
>>   	struct nvme_ns *ns;
>> +	int ret = 0;
>>   
>> +	if (xa_empty(&ctrl->namespaces)) {
>>   		ret = -ENOTTY;
>> +		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");
>> +	/* Let the scan work finish updating ctrl->namespaces */
>> +	flush_work(&ctrl->scan_work);
>> +	if (nvme_identify_ctrl(ctrl, &id)) {
>> +		dev_err(ctrl->device, "nvme_identify_ctrl() failed\n");
>>   		ret = -EINVAL;
>> -		goto out_unlock;
>> +		goto out;
>> +	}
>> +	if (le32_to_cpu(id->nn) > 1) {
>> +		dev_warn(ctrl->device,
>> +		"NVME_IOCTL_IO_CMD not supported when multiple namespaces present!\n");
>> +		goto out;
>>   	}
> 
> This code doesn't make any sense at all.  Why does a patch changing
> data structures add new calls that go out on the wire?
> 
Yes, this should not be here, I'll remove that and only keep the code to
check multiple namespaces and if needed this needs to be a separate
patch.

>> +	struct nvme_ns *ns;
>> +	XA_STATE(xas, &ctrl->namespaces, nsid);
>>   
>> +	rcu_read_lock();
>> +	do {
>> +		ns = xas_load(&xas);
>> +		if (xa_is_zero(ns))
>> +			ns = NULL;
>> +	} while (xas_retry(&xas, ns));
>> +	ns = ns && kref_get_unless_zero(&ns->kref) ? ns : NULL;
>> +	rcu_read_unlock();
> 
> This looks pretty weird, but I think the problem is one in the xarray
> API, as for the typical lookup pattern we'd want an xa_load with
> external RCU locking:
> 

The Kref needs to be taken under the lock so I've open coded the 
xa_load() and added kref get under rcu locking.

Matthew can shed more light above pattern this ?

> 	rcu_read_lock();
> 	ns = xa_load_rcu(&ctrl->namespaces, nsid);
> 	if (ns && !kref_get_unless_zero(&ns->kref))
> 		ns = NULL;
> 	rcu_read_unlock();
> 
> instead of duplicating this fairly arcane loop in all kinds of callers.
> 
>> -	down_write(&ctrl->namespaces_rwsem);
>> -	list_add_tail(&ns->list, &ctrl->namespaces);
>> -	up_write(&ctrl->namespaces_rwsem);
>> +	ret = xa_insert(&ctrl->namespaces, nsid, ns, GFP_KERNEL);
>> +	if (ret) {
>> +		switch (ret) {
>> +		case -ENOMEM:
>> +			dev_err(ctrl->device,
>> +				"xa insert memory allocation\n");
>> +			break;
>> +		case -EBUSY:
>> +			dev_err(ctrl->device,
>> +				"xa insert entry already present\n");
>> +			break;
>> +		}
>> +	}
> 
> No need for the switch and the detailed printks here, but we do need
> actual error handling.
> 

We need to add a wrapper nvme_xa_insert() to take care for the error 
handling with pr_debug() and switch() which can be used everywhere
to not bloat the functions calling xa_insert() are you okay with that ?

I'll have goto above based on if you are okay with the xa_insert() 
wrapper nvme_xa_insert() to undo things that alloc ns does.

>>   static void nvme_ns_remove(struct nvme_ns *ns)
>>   {
>> +	struct xarray *xa = &ns->ctrl->namespaces;
>> +	bool free;
>> +
>>   	if (test_and_set_bit(NVME_NS_REMOVING, &ns->flags))
>>   		return;
>>   
>> @@ -3740,12 +3749,14 @@ static void nvme_ns_remove(struct nvme_ns *ns)
>>   			blk_integrity_unregister(ns->disk);
>>   	}
>>   
>> -	down_write(&ns->ctrl->namespaces_rwsem);
>> -	list_del_init(&ns->list);
>> -	up_write(&ns->ctrl->namespaces_rwsem);
>> +	xa_lock(xa);
>> +	__xa_erase(xa, ns->head->ns_id);
>> +	free = refcount_dec_and_test(&ns->kref.refcount) ? true : false;
>> +	xa_unlock(xa);
>>   
>>   	nvme_mpath_check_last_path(ns);
>> -	nvme_put_ns(ns);
>> +	if (free)
>> +		__nvme_free_ns(ns);
> 
> This looks very strange to me.  Shoudn't this be a normal xa_erase
> followed by a normal nvme_put_ns?  For certain the driver code has
> no business poking into the kref internals.
> 

There is a race when kref is manipulated in nvme_find_get_ns() and
nvme _ns_remove() pointed by Keith which needs ns->kref to be guarded 
with locks. Let me know I'll share a detail scenario.

Given that xarray locking only uses spinlocks we cannot call
nvme_put_ns() since it will sleep, so it separates the code.

>>   static void nvme_remove_invalid_namespaces(struct nvme_ctrl *ctrl,
>>   					unsigned nsid)
>>   {
>> +	struct xarray *namespaces = &ctrl->namespaces;
>> +	struct xarray rm_array;
>> +	unsigned long tnsid;
>> +	struct nvme_ns *ns;
>> +	unsigned long idx;
>> +	int ret;
>>   
>> +	xa_init(&rm_array);
>> +
>> +	xa_lock(namespaces);
>> +	xa_for_each(namespaces, idx, ns) {
>> +		tnsid = ns->head->ns_id;
>> +		if (tnsid > nsid || test_bit(NVME_NS_DEAD, &ns->flags)) {
>> +			xa_unlock(namespaces);
>> +			xa_erase(namespaces, tnsid);
>> +			/* Even if insert fails keep going */
>> +			ret = xa_insert(&rm_array, nsid, ns, GFP_KERNEL);
>> +			switch (ret) {
>> +			case -ENOMEM:
>> +				pr_err("xa insert memory allocation failed\n");
>> +				break;
>> +			case -EBUSY:
>> +				pr_err("xa insert entry already present\n");
>> +				break;
>> +			}
>> +			xa_lock(namespaces);
>> +		}
>>   	}
>> +	xa_unlock(namespaces);
> 
> I don't think you want an xarray for the delete list.  Just keep the
> list head for that now - once we moved to RCU read side locking some
> of this could potentially be simplified later.
Okay that make sense will keep it simple, will do.
> 
>>    */
>>   void nvme_remove_namespaces(struct nvme_ctrl *ctrl)
>>   {
>> -	struct nvme_ns *ns, *next;
>> -	LIST_HEAD(ns_list);
>> +	struct xarray rm_array;
>> +	unsigned long tnsid;
>> +	struct nvme_ns *ns;
>> +	unsigned long idx;
>> +	int ret;
>> +
>> +	xa_init(&rm_array);
>>   
>>   	/*
>>   	 * make sure to requeue I/O to all namespaces as these
>> @@ -3919,11 +3950,30 @@ 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) {
>> +		tnsid = ns->head->ns_id;
>> +		xa_unlock(&ctrl->namespaces);
>> +		xa_erase(&ctrl->namespaces, tnsid);
>> +		/* Even if insert fails keep going */
>> +		ret = xa_insert(&rm_array, tnsid, ns, GFP_KERNEL);
>> +		if (ret) {
>> +			switch (ret) {
>> +			case -ENOMEM:
>> +				dev_err(ctrl->device,
>> +					"xa insert memory allocation\n");
>> +				break;
>> +			case -EBUSY:
>> +				dev_err(ctrl->device,
>> +					"xa insert entry already present\n");
>> +				break;
>> +			}
>> +		}
>> +		xa_lock(&ctrl->namespaces);
>> +	}
>> +	xa_unlock(&ctrl->namespaces);
> 
> Same here.
Yes there is pattern here and I was wondering if we can ge rid of this 
duplicate code will see.
>