Re: [PATCH v6 1/1] nvme-multipath: implement "queue-depth" iopolicy

[Chaitanya Kulkarni <chaitanyak@nvidia.com>]

On 6/11/24 17:20, John Meneghini wrote:
> From: Thomas Song <tsong@purestorage.com>
>
> The round-robin path selector is inefficient in cases where there is a
> difference in latency between paths.  In the presence of one or more
> high latency paths the round-robin selector continues to use the high
> latency path equally. This results in a bias towards the highest latency
> path and can cause a significant decrease in overall performance as IOs
> pile on the highest latency path. This problem is acute with NVMe-oF
> controllers.
>
> The queue-depth policy instead sends I/O requests down the path with the
> least amount of requests in its request queue. Paths with lower latency
> will clear requests more quickly and have less requests in their queues
> compared to higher latency paths. The goal of this path selector is to
> make more use of lower latency paths which will bring down overall IO
> latency and increase throughput and performance.
>
> Signed-off-by: Thomas Song <tsong@purestorage.com>
> [emilne: patch developed by Thomas Song @ Pure Storage, fixed whitespace
>        and compilation warnings, updated MODULE_PARM description, and
>        fixed potential issue with ->current_path[] being used]
> Co-developed-by: Ewan D. Milne <emilne@redhat.com>
> Signed-off-by: Ewan D. Milne <emilne@redhat.com>
> [jmeneghi: vairious changes and improvements, addressed review comments]
> Co-developed-by: John Meneghini <jmeneghi@redhat.com>
> Signed-off-by: John Meneghini <jmeneghi@redhat.com>
> Link: https://lore.kernel.org/linux-nvme/20240509202929.831680-1-jmeneghi@redhat.com/
> Tested-by: Marco Patalano <mpatalan@redhat.com>
> Reviewed-by: Randy Jennings <randyj@purestorage.com>
> Tested-by: Jyoti Rani <jrani@purestorage.com>
> ---
>   drivers/nvme/host/core.c      |   2 +-
>   drivers/nvme/host/multipath.c | 108 +++++++++++++++++++++++++++++++---
>   drivers/nvme/host/nvme.h      |   5 ++
>   3 files changed, 106 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/nvme/host/core.c b/drivers/nvme/host/core.c
> index 7c9f91314d36..c10ff8815d82 100644
> --- a/drivers/nvme/host/core.c
> +++ b/drivers/nvme/host/core.c
> @@ -110,7 +110,7 @@ struct workqueue_struct *nvme_delete_wq;
>   EXPORT_SYMBOL_GPL(nvme_delete_wq);
>   
>   static LIST_HEAD(nvme_subsystems);
> -static DEFINE_MUTEX(nvme_subsystems_lock);
> +DEFINE_MUTEX(nvme_subsystems_lock);
>   
>   static DEFINE_IDA(nvme_instance_ida);
>   static dev_t nvme_ctrl_base_chr_devt;
> diff --git a/drivers/nvme/host/multipath.c b/drivers/nvme/host/multipath.c
> index 03a6868f4dbc..fe10e0cebcf0 100644
> --- a/drivers/nvme/host/multipath.c
> +++ b/drivers/nvme/host/multipath.c
> @@ -17,6 +17,7 @@ MODULE_PARM_DESC(multipath,
>   static const char *nvme_iopolicy_names[] = {
>   	[NVME_IOPOLICY_NUMA]	= "numa",
>   	[NVME_IOPOLICY_RR]	= "round-robin",
> +	[NVME_IOPOLICY_QD]      = "queue-depth",
>   };
>   
>   static int iopolicy = NVME_IOPOLICY_NUMA;
> @@ -29,6 +30,8 @@ static int nvme_set_iopolicy(const char *val, const struct kernel_param *kp)
>   		iopolicy = NVME_IOPOLICY_NUMA;
>   	else if (!strncmp(val, "round-robin", 11))
>   		iopolicy = NVME_IOPOLICY_RR;
> +	else if (!strncmp(val, "queue-depth", 11))
> +		iopolicy = NVME_IOPOLICY_QD;
>   	else
>   		return -EINVAL;
>   
> @@ -43,7 +46,7 @@ static int nvme_get_iopolicy(char *buf, const struct kernel_param *kp)
>   module_param_call(iopolicy, nvme_set_iopolicy, nvme_get_iopolicy,
>   	&iopolicy, 0644);
>   MODULE_PARM_DESC(iopolicy,
> -	"Default multipath I/O policy; 'numa' (default) or 'round-robin'");
> +	"Default multipath I/O policy; 'numa' (default), 'round-robin' or 'queue-depth'");
>   
>   void nvme_mpath_default_iopolicy(struct nvme_subsystem *subsys)
>   {
> @@ -128,6 +131,11 @@ void nvme_mpath_start_request(struct request *rq)
>   	struct nvme_ns *ns = rq->q->queuedata;
>   	struct gendisk *disk = ns->head->disk;
>   
> +	if (READ_ONCE(ns->head->subsys->iopolicy) == NVME_IOPOLICY_QD) {
> +		atomic_inc(&ns->ctrl->nr_active);
> +		nvme_req(rq)->flags |= NVME_MPATH_CNT_ACTIVE;
> +	}
> +
>   	if (!blk_queue_io_stat(disk->queue) || blk_rq_is_passthrough(rq))
>   		return;
>   
> @@ -140,6 +148,12 @@ EXPORT_SYMBOL_GPL(nvme_mpath_start_request);
>   void nvme_mpath_end_request(struct request *rq)
>   {
>   	struct nvme_ns *ns = rq->q->queuedata;
> +	int result;
> +
> +	if ((nvme_req(rq)->flags & NVME_MPATH_CNT_ACTIVE)) {
> +		result = atomic_dec_if_positive(&ns->ctrl->nr_active);
> +		WARN_ON_ONCE(result < 0);
> +	}
>   
>   	if (!(nvme_req(rq)->flags & NVME_MPATH_IO_STATS))
>   		return;

can we remove result variable ? that is only used once,
how about something like this unless there is something wrong with
totally untested :-

@@ -141,6 +149,9 @@ void nvme_mpath_end_request(struct request *rq)
  {
         struct nvme_ns *ns = rq->q->queuedata;

+       if ((nvme_req(rq)->flags & NVME_MPATH_CNT_ACTIVE))
+ WARN_ON_ONCE(atomic_dec_if_positive(&ns->ctrl->nr_active) < 0);
+
         if (!(nvme_req(rq)->flags & NVME_MPATH_IO_STATS))
                 return;
         bdev_end_io_acct(ns->head->disk->part0, req_op(rq),


> @@ -291,10 +305,15 @@ static struct nvme_ns *nvme_next_ns(struct nvme_ns_head *head,
>   	return list_first_or_null_rcu(&head->list, struct nvme_ns, siblings);
>   }
>   
> -static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
> -		int node, struct nvme_ns *old)
> +static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head)
>   {
> -	struct nvme_ns *ns, *found = NULL;
> +	struct nvme_ns *ns, *old, *found = NULL;
> +	int node = numa_node_id();
> +
> +	old = srcu_dereference(head->current_path[node], &head->srcu);
> +

nit:- no need for white-line above ?

> +	if (unlikely(!old))
> +		return __nvme_find_path(head, node);
>   
>   	if (list_is_singular(&head->list)) {
>   		if (nvme_path_is_disabled(old))
> @@ -334,13 +353,49 @@ static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head,
>   	return found;
>   }
>   
> +static struct nvme_ns *nvme_queue_depth_path(struct nvme_ns_head *head)
> +{
> +	struct nvme_ns *best_opt = NULL, *best_nonopt = NULL, *ns;
> +	unsigned int min_depth_opt = UINT_MAX, min_depth_nonopt = UINT_MAX;
> +	unsigned int depth;
> +
> +	list_for_each_entry_rcu(ns, &head->list, siblings) {
> +		if (nvme_path_is_disabled(ns))
> +			continue;
> +
> +		depth = atomic_read(&ns->ctrl->nr_active);
> +
> +		switch (ns->ana_state) {
> +		case NVME_ANA_OPTIMIZED:
> +			if (depth < min_depth_opt) {
> +				min_depth_opt = depth;
> +				best_opt = ns;
> +			}
> +			break;
> +		case NVME_ANA_NONOPTIMIZED:
> +			if (depth < min_depth_nonopt) {
> +				min_depth_nonopt = depth;
> +				best_nonopt = ns;
> +			}
> +			break;
> +		default:
> +			break;
> +		}
> +
> +		if (min_depth_opt == 0)
> +			return best_opt;
> +	}
> +
> +	return best_opt ? best_opt : best_nonopt;
> +}
> +
>   static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
>   {
>   	return nvme_ctrl_state(ns->ctrl) == NVME_CTRL_LIVE &&
>   		ns->ana_state == NVME_ANA_OPTIMIZED;
>   }
>   
> -inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
> +static struct nvme_ns *nvme_numa_path(struct nvme_ns_head *head)
>   {
>   	int node = numa_node_id();
>   	struct nvme_ns *ns;
> @@ -349,13 +404,25 @@ inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
>   	if (unlikely(!ns))
>   		return __nvme_find_path(head, node);
>   
> -	if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
> -		return nvme_round_robin_path(head, node, ns);
>   	if (unlikely(!nvme_path_is_optimized(ns)))
>   		return __nvme_find_path(head, node);
> +

nit:- do we really need above white line in this code ?

>   	return ns;
>   }
>   
> +
> +inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
> +{
> +	switch (READ_ONCE(head->subsys->iopolicy)) {
> +	case NVME_IOPOLICY_QD:
> +		return nvme_queue_depth_path(head);
> +	case NVME_IOPOLICY_RR:
> +		return nvme_round_robin_path(head);
> +	default:
> +		return nvme_numa_path(head);
> +	}

should we use another case for NVME_IOPOLICY_NUMA that will call
nvme_numa_path() and report ratelimited error on the default lable
before settling on nvme_numa_path()?

something like this totally untested :-

inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
{
         switch (READ_ONCE(head->subsys->iopolicy)) {
         case NVME_IOPOLICY_QD:
                 return nvme_queue_depth_path(head);
         case NVME_IOPOLICY_RR:
                 return nvme_round_robin_path(head);
         case NVME_IOPOLICY_NUMA:
                 return nvme_numa_path(head);
         default:
                 dev_warn_ratelimited(disk_to_dev(head->disk),
                                      "invalid iopolicy %u",
head->subsys->iopolicy);
                 return nvme_numa_path(head);
         }
}

but if it has already been discussed to not have default case
then please ignore this comment ...

> +}
> +
>   static bool nvme_available_path(struct nvme_ns_head *head)
>   {
>   	struct nvme_ns *ns;
> @@ -803,6 +870,28 @@ static ssize_t nvme_subsys_iopolicy_show(struct device *dev,
>   			  nvme_iopolicy_names[READ_ONCE(subsys->iopolicy)]);
>   }
>   
> +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);
> +
> +	/* iopolicy changes clear the mpath by design */
> +	mutex_lock(&nvme_subsystems_lock);
> +	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry)
> +		nvme_mpath_clear_ctrl_paths(ctrl);
> +	mutex_unlock(&nvme_subsystems_lock);
> +
> +	pr_notice("%s: changed from %s to %s for subsysnqn %s\n", __func__,
> +			nvme_iopolicy_names[old_iopolicy], nvme_iopolicy_names[iopolicy],
> +			subsys->subnqn);
> +}
> +
>   static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
>   		struct device_attribute *attr, const char *buf, size_t count)
>   {
> @@ -812,7 +901,7 @@ static ssize_t nvme_subsys_iopolicy_store(struct device *dev,
>   
>   	for (i = 0; i < ARRAY_SIZE(nvme_iopolicy_names); i++) {
>   		if (sysfs_streq(buf, nvme_iopolicy_names[i])) {
> -			WRITE_ONCE(subsys->iopolicy, i);
> +			nvme_subsys_iopolicy_update(subsys, i);
>   			return count;
>   		}
>   	}
> @@ -923,6 +1012,9 @@ int nvme_mpath_init_identify(struct nvme_ctrl *ctrl, struct nvme_id_ctrl *id)
>   	    !(ctrl->subsys->cmic & NVME_CTRL_CMIC_ANA))
>   		return 0;
>   
> +	/* initialize this in the identify path to cover controller resets */

nit: If I'm not wrong, this function gets called from 
|nvme_init_identify()|,
so it's pretty clear. That makes above comment kind of redundant ?
However, if others want that comment here, please ignore this message.

> +	atomic_set(&ctrl->nr_active, 0);
> +
>   	if (!ctrl->max_namespaces ||
>   	    ctrl->max_namespaces > le32_to_cpu(id->nn)) {
>   		dev_err(ctrl->device,
> diff --git a/drivers/nvme/host/nvme.h b/drivers/nvme/host/nvme.h
> index 73442d3f504b..d6c1fe3e2832 100644
> --- a/drivers/nvme/host/nvme.h
> +++ b/drivers/nvme/host/nvme.h
> @@ -50,6 +50,8 @@ extern struct workqueue_struct *nvme_wq;
>   extern struct workqueue_struct *nvme_reset_wq;
>   extern struct workqueue_struct *nvme_delete_wq;
>   
> +extern struct mutex nvme_subsystems_lock;
> +
>   /*
>    * List of workarounds for devices that required behavior not specified in
>    * the standard.
> @@ -195,6 +197,7 @@ enum {
>   	NVME_REQ_CANCELLED		= (1 << 0),
>   	NVME_REQ_USERCMD		= (1 << 1),
>   	NVME_MPATH_IO_STATS		= (1 << 2),
> +	NVME_MPATH_CNT_ACTIVE	= (1 << 3),

nit:- please align above to existing code ...

>   };
>   
>   static inline struct nvme_request *nvme_req(struct request *req)
> @@ -360,6 +363,7 @@ struct nvme_ctrl {
>   	size_t ana_log_size;
>   	struct timer_list anatt_timer;
>   	struct work_struct ana_work;
> +	atomic_t nr_active;
>   #endif
>   
>   #ifdef CONFIG_NVME_HOST_AUTH
> @@ -408,6 +412,7 @@ static inline enum nvme_ctrl_state nvme_ctrl_state(struct nvme_ctrl *ctrl)
>   enum nvme_iopolicy {
>   	NVME_IOPOLICY_NUMA,
>   	NVME_IOPOLICY_RR,
> +	NVME_IOPOLICY_QD,
>   };
>   
>   struct nvme_subsystem {

-ck