Skip to contentSkip to conversation

nvme: multipath: Implemented new iopolicy "queue-depth"

PATCH Superseded John Meneghini 5 participants 17 messages May 22, 2024 – May 23, 2024 NVMe

v5 of 8 Superseded Latest is v8 →

lore
Share & tools
lore
#0 lore raw

PATCH v5 1/1 nvme: multipath: Implemented new iopolicy "queue-depth"

From: "Ewan D. Milne" <[email protected]>

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.

[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]

---
 drivers/nvme/host/core.c      |  2 +-
 drivers/nvme/host/multipath.c | 94 +++++++++++++++++++++++++++++++++--
 drivers/nvme/host/nvme.h      |  5 ++
 3 files changed, 96 insertions(+), 5 deletions(-)
drivers/nvme/host/core.c +1 -1
index 7706df237349..66dc676677d4 100644
@@ -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;
drivers/nvme/host/multipath.c +90 -4

161 lines not on this page · raw message

drivers/nvme/host/nvme.h +5 -0
index fc31bd340a63..fa3833d88a85 100644
@@ -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),
 };
 
 static inline struct nvme_request *nvme_req(struct request *req)
@@ -359,6 +362,7 @@ 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
@@ -407,6 +411,7 @@ 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 {
Signature · 1 line
2.39.3
maintainer
#1 lore raw
Quoted 10 lines · John Meneghini

From: "Ewan D. Milne" <[email protected]>

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.

Quoted 6 lines

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.

I'm okay with this as-is, though I don't think you need either atomic_set() calls.

Christoph, Sagi, 6.10 merge window is still open and this has been iterating long before that. Any objection?

#2 lore raw
Quoted 2 lines · Keith Busch

Christoph, Sagi, 6.10 merge window is still open and this has been iterating long before that. Any objection?

No, it's not. The window for development closes with the release for 6.9. New features must be in before that.

#3 lore raw
Quoted 5 lines · Keith Busch
Quoted 2 lines

Christoph, Sagi, 6.10 merge window is still open and this has been iterating long before that. Any objection?

No, it's not. The window for development closes with the release for 6.9. New features must be in before that.

So what's the next window for new features? 6.11?

/John

#4 lore raw
Quoted 8 lines · John Meneghini
Quoted 5 lines · Keith Busch
Quoted 2 lines

Christoph, Sagi, 6.10 merge window is still open and this has been iterating long before that. Any objection?

No, it's not. The window for development closes with the release for 6.9. New features must be in before that.

So what's the next window for new features? 6.11?

What else?

#5 lore raw
Quoted 11 lines · Christoph Hellwig
Quoted 8 lines · John Meneghini
Quoted 5 lines · Keith Busch
Quoted 2 lines

Christoph, Sagi, 6.10 merge window is still open and this has been iterating long before that. Any objection?

No, it's not. The window for development closes with the release for 6.9. New features must be in before that.

So what's the next window for new features? 6.11?

What else?

Then merge it there. Duh.

maintainer
#6 lore raw
In reply to John Meneghini
Quoted 8 lines · John Meneghini
Quoted 5 lines · Keith Busch
Quoted 2 lines

Christoph, Sagi, 6.10 merge window is still open and this has been iterating long before that. Any objection?

No, it's not. The window for development closes with the release for 6.9. New features must be in before that.

So what's the next window for new features? 6.11?

The nvme 6.11 branch will be created after block creates a block-6.11 branch, which usually happens after a few -rc's are released in the current cycle. So at least a few more weeks.

author Tested-by on v5 1/1
#7 lore raw
In reply to John Meneghini

OK, I've retested this patch and verified the queue-depth policy is working as expected.

Running the same workload using queue-depth policy instead of round-robin provides a more even distribution of IO across all controllers.

https://github.com/johnmeneghini/iopolicy?tab=readme-ov-file#compare-outstanding-ios-per-controller

And it increases the overall throughput per-namespace:

https://github.com/johnmeneghini/iopolicy?tab=readme-ov-file#compare-outstanding-ios-per-namespace

Quoted 259 lines · John Meneghini

From: "Ewan D. Milne" <[email protected]>

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.

[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]

---
  drivers/nvme/host/core.c      |  2 +-
  drivers/nvme/host/multipath.c | 94 +++++++++++++++++++++++++++++++++--
  drivers/nvme/host/nvme.h      |  5 ++
  3 files changed, 96 insertions(+), 5 deletions(-)
drivers/nvme/host/core.c +1 -1
index 7706df237349..66dc676677d4 100644
@@ -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;
drivers/nvme/host/multipath.c +90 -4
index 1bee176fd850..96264c5c979f 100644
@@ -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;
 
@@ -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))
+		atomic_dec_if_positive(&ns->ctrl->nr_active);
+
  	if (!(nvme_req(rq)->flags & NVME_MPATH_IO_STATS))
  		return;
  	bdev_end_io_acct(ns->head->disk->part0, req_op(rq),
@@ -332,6 +343,43 @@ 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 &&
@@ -340,15 +388,27 @@ static inline bool nvme_path_is_optimized(struct nvme_ns *ns)
 
  inline struct nvme_ns *nvme_find_path(struct nvme_ns_head *head)
  {
-	int node = numa_node_id();
+	int iopolicy = READ_ONCE(head->subsys->iopolicy);
+	int node;
  	struct nvme_ns *ns;
 
+	/*
+	 * queue-depth iopolicy does not need to reference ->current_path
+	 * but round-robin needs the last path used to advance to the
+	 * next one, and numa will continue to use the last path unless
+	 * it is or has become not optimized
+	 */
+	if (iopolicy == NVME_IOPOLICY_QD)
+		return nvme_queue_depth_path(head);
+
+	node = numa_node_id();
  	ns = srcu_dereference(head->current_path[node], &head->srcu);
  	if (unlikely(!ns))
  		return __nvme_find_path(head, node);
 
-	if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
+	if (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);
  	return ns;
@@ -800,6 +860,29 @@ 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 reset the counters and clear the mpath by design */
+	mutex_lock(&nvme_subsystems_lock);
+	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
+		atomic_set(&ctrl->nr_active, 0);
+		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)
  {
@@ -809,7 +892,7 @@ 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;
  		}
  	}
@@ -920,6 +1003,9 @@ 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 */
+	atomic_set(&ctrl->nr_active, 0);
+
  	if (!ctrl->max_namespaces ||
  	    ctrl->max_namespaces > le32_to_cpu(id->nn)) {
  		dev_err(ctrl->device,
drivers/nvme/host/nvme.h +5 -0
index fc31bd340a63..fa3833d88a85 100644
@@ -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),
  };
 
  static inline struct nvme_request *nvme_req(struct request *req)
@@ -359,6 +362,7 @@ 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
@@ -407,6 +411,7 @@ 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 {
Reviewed-by on v5 1/1
#8 lore raw
In reply to John Meneghini
Quoted 29 lines · John Meneghini

From: "Ewan D. Milne" <[email protected]>

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.

[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]

---

[...]

Quoted 10 lines
drivers/nvme/host/multipath.c +10 -0
+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;
+

drivers/nvme/host/multipath.c static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head, · line 365

Quoted 10 lines
drivers/nvme/host/multipath.c +10 -0
+		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;
+

nit:- no need to add white line needed after break above ?

Quoted 18 lines
drivers/nvme/host/multipath.c +17 -0
+		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;
+}
+

[...]

drivers/nvme/host/multipath.c static ssize_t nvme_subsys_iopolicy_show(struct device *dev, · line 863

Quoted 5 lines
drivers/nvme/host/multipath.c +1 -0
@@ -800,6 +860,29 @@ 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)

nit:- overly long line, as rest of the file is < 80 char par line ?

Quoted 9 lines
drivers/nvme/host/multipath.c +9 -0
+{
+	struct nvme_ctrl *ctrl;
+	int old_iopolicy = READ_ONCE(subsys->iopolicy);
+
+	if (old_iopolicy == iopolicy)
+		return;
+
+	WRITE_ONCE(subsys->iopolicy, iopolicy);
+

drivers/nvme/host/multipath.c static ssize_t nvme_subsys_iopolicy_show(struct device *dev, · line 882

Quoted 10 lines
drivers/nvme/host/multipath.c +10 -0
+	/* iopolicy changes reset the counters and clear the mpath by design */
+	mutex_lock(&nvme_subsystems_lock);
+	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
+		atomic_set(&ctrl->nr_active, 0);
+		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],

nit: overly long line above as rest of the file is < 80 char ...

As far as I remember, most of the nvme code uses pr_info(), but if decision has been made to use pr_notice() for a specific reason then please ignore this comment.

Quoted 3 lines
drivers/nvme/host/multipath.c +3 -0
+			subsys->subnqn);
+}
+

[...]

Quoted 14 lines
drivers/nvme/host/nvme.h +2 -0
index fc31bd340a63..fa3833d88a85 100644
@@ -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.

drivers/nvme/host/nvme.h enum { · line 200

Quoted 4 lines
drivers/nvme/host/nvme.h +1 -0
@@ -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:- I think we need to align above line to rest of the members in
      this enum ...

Quoted 12 lines
drivers/nvme/host/nvme.h +1 -0
  };
 
  static inline struct nvme_request *nvme_req(struct request *req)
@@ -359,6 +362,7 @@ 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

drivers/nvme/host/nvme.h static inline enum nvme_ctrl_state nvme_ctrl_state(struct nvme_ctrl *ctrl) · line 417

Quoted 7 lines
drivers/nvme/host/nvme.h +1 -0
@@ -407,6 +411,7 @@ 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 {

apart from the few nits patch does looks good to me.

not a blocker to merge this patch, but we need a blktests for this code in nvme
category ...

-ck

#9 lore raw
Quoted 161 lines · Chaitanya Kulkarni
Quoted 29 lines · John Meneghini

From: "Ewan D. Milne" <[email protected]>

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.

[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]

---

[...]

Quoted 10 lines
drivers/nvme/host/multipath.c +10 -0
+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;
+

drivers/nvme/host/multipath.c static struct nvme_ns *nvme_round_robin_path(struct nvme_ns_head *head, · line 365

Quoted 10 lines
drivers/nvme/host/multipath.c +10 -0
+		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;
+

nit:- no need to add white line needed after break above ?

Quoted 18 lines
drivers/nvme/host/multipath.c +17 -0
+		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;
+}
+

[...]

drivers/nvme/host/multipath.c static ssize_t nvme_subsys_iopolicy_show(struct device *dev, · line 863

Quoted 5 lines
drivers/nvme/host/multipath.c +1 -0
@@ -800,6 +860,29 @@ 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)

nit:- overly long line, as rest of the file is < 80 char par line ?

Quoted 9 lines
drivers/nvme/host/multipath.c +9 -0
+{
+	struct nvme_ctrl *ctrl;
+	int old_iopolicy = READ_ONCE(subsys->iopolicy);
+
+	if (old_iopolicy == iopolicy)
+		return;
+
+	WRITE_ONCE(subsys->iopolicy, iopolicy);
+

drivers/nvme/host/multipath.c static ssize_t nvme_subsys_iopolicy_show(struct device *dev, · line 882

Quoted 10 lines
drivers/nvme/host/multipath.c +10 -0
+	/* iopolicy changes reset the counters and clear the mpath by design */
+	mutex_lock(&nvme_subsystems_lock);
+	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
+		atomic_set(&ctrl->nr_active, 0);
+		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],

nit: overly long line above as rest of the file is < 80 char ...

As far as I remember, most of the nvme code uses pr_info(), but if decision has been made to use pr_notice() for a specific reason then please ignore this comment.

Quoted 3 lines
drivers/nvme/host/multipath.c +3 -0
+			subsys->subnqn);
+}
+

[...]

Quoted 14 lines
drivers/nvme/host/nvme.h +2 -0
index fc31bd340a63..fa3833d88a85 100644
@@ -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.

drivers/nvme/host/nvme.h enum { · line 200

Quoted 4 lines
drivers/nvme/host/nvme.h +1 -0
@@ -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:- I think we need to align above line to rest of the members in
      this enum ...

Quoted 12 lines
drivers/nvme/host/nvme.h +1 -0
   };
 
   static inline struct nvme_request *nvme_req(struct request *req)
@@ -359,6 +362,7 @@ 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

drivers/nvme/host/nvme.h static inline enum nvme_ctrl_state nvme_ctrl_state(struct nvme_ctrl *ctrl) · line 417

Quoted 7 lines
drivers/nvme/host/nvme.h +1 -0
@@ -407,6 +411,7 @@ 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 {

apart from the few nits patch does looks good to me.

Quoted 6 lines

not a blocker to merge this patch, but we need a blktests for this code in nvme
category ...

As presented at LSF by Daniel; ALUA support (and, with that, multipath support) is one of the topics to be implemented for blktests. And without that we can't have a meaningful QD test.

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                  Kernel Storage Architect
[email protected]                                +49 911 74053 688
SUSE Software Solutions GmbH, Frankenstr. 146, 90461 Nürnberg
HRB 36809 (AG Nürnberg), GF: I. Totev, A. McDonald, W. Knoblich
#10 lore raw
Quoted 2 lines · Hannes Reinecke

As presented at LSF by Daniel; ALUA support (and, with that, multipath support) is one of the topics to be implemented for blktests. And without that we can't have a meaningful QD test.

So as a part of this patch you want a blktest that verifies queue-depth multipathing?

Or are your just tying acceptance of this patch to a blktest that tests multipath testing over all?

Seems to me you have a patch designed to do that...

https://github.com/osandov/blktests/pull/114

/John

#11 lore raw
Quoted 6 lines · John Meneghini
Quoted 3 lines · Hannes Reinecke

As presented at LSF by Daniel; ALUA support (and, with that, multipath support) is one of the topics to be implemented for blktests. And without that we can't have a meaningful QD test.

Quoted 6 lines

So as a part of this patch you want a blktest that verifies queue-depth multipathing?

Or are your just tying acceptance of this patch to a blktest that tests multipath testing over all?

as I mentioned this earlier having a blktest is not a blocker for acceptance of this patch at all

once this patch is merged it'd be nice to have it, but you have to wait for Daniel's ALUA support
as mentioned by Hannes in previous email ...

Quoted 5 lines

Seems to me you have a patch designed to do that...

https://github.com/osandov/blktests/pull/114

/John

-ck

#12 lore raw
In reply to Hannes Reinecke
Quoted 9 lines
Quoted 9 lines

apart from the few nits patch does looks good to me.

not a blocker to merge this patch, but we need a blktests for this code in nvme
category ...

Quoted 7 lines

As presented at LSF by Daniel; ALUA support (and, with that, multipath support) is one of the topics to be implemented for blktests. And without that we can't have a meaningful QD test.

Cheers,

Hannes

we can wait for ALUA, as I said it is not a blocker but thanks pointing that out ...

-ck

#13 lore raw
In reply to John Meneghini

drivers/nvme/host/multipath.c static inline bool nvme_path_is_optimized(struct nvme_ns *ns) · line 400

Quoted 6 lines
drivers/nvme/host/multipath.c +6 -0
+	/*
+	 * queue-depth iopolicy does not need to reference ->current_path
+	 * but round-robin needs the last path used to advance to the
+	 * next one, and numa will continue to use the last path unless
+	 * it is or has become not optimized
+	 */

Can we please turn this into a full sentence? I.e.:

/*
* The queue-depth iopolicy does not need to reference ->current_path,
* but the round-robin iopolicy needs the last path used to advance to
* the next one, and numa will continue to use the last path unless
* it is or has become non-optimized.
*/

?

Quoted 5 lines
drivers/nvme/host/multipath.c +4 -0
+	if (iopolicy == NVME_IOPOLICY_QD)
+		return nvme_queue_depth_path(head);
+
+	node = numa_node_id();
 	ns = srcu_dereference(head->current_path[node], &head->srcu);

drivers/nvme/host/multipath.c static inline bool nvme_path_is_optimized(struct nvme_ns *ns) · line 414

Quoted 10 lines
drivers/nvme/host/multipath.c +2 -1
 	if (unlikely(!ns))
 		return __nvme_find_path(head, node);
 
-	if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
+	if (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);
 	return ns;

Also this is growing into the kind of spaghetti code that is on the fast path to become unmaintainable. I'd much rather see the srcu_dereference + __nvme_find_path duplicated and have a switch over the iopolicies with a separate helper for each of them here than the various ifs at different levels.

Quoted 1 line

+static void nvme_subsys_iopolicy_update(struct nvme_subsystem *subsys, int iopolicy)

Overly long line here.

Quoted 6 lines
drivers/nvme/host/multipath.c +6 -0
+{
+	struct nvme_ctrl *ctrl;
+	int old_iopolicy = READ_ONCE(subsys->iopolicy);
+
+	if (old_iopolicy == iopolicy)
+		return;

drivers/nvme/host/multipath.c static ssize_t nvme_subsys_iopolicy_show(struct device *dev, · line 879

Quoted 10 lines
drivers/nvme/host/multipath.c +10 -0
+
+	WRITE_ONCE(subsys->iopolicy, iopolicy);
+
+	/* iopolicy changes reset the counters and clear the mpath by design */
+	mutex_lock(&nvme_subsystems_lock);
+	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
+		atomic_set(&ctrl->nr_active, 0);
+		nvme_mpath_clear_ctrl_paths(ctrl);
+	}
+	mutex_unlock(&nvme_subsystems_lock);

You probably want to take the lock over the iopolicy assignment to serialize it. And why do we need the atomic_set here?

drivers/nvme/host/multipath.c static ssize_t nvme_subsys_iopolicy_show(struct device *dev, · line 882

Quoted 3 lines
drivers/nvme/host/multipath.c +3 -0
+
+	pr_notice("%s: changed from %s to %s for subsysnqn %s\n", __func__,
+			nvme_iopolicy_names[old_iopolicy], nvme_iopolicy_names[iopolicy],

Pleae avoid the overly long line here as well.

drivers/nvme/host/nvme.h enum { · line 200

Quoted 4 lines
drivers/nvme/host/nvme.h +1 -0
 	NVME_REQ_CANCELLED		= (1 << 0),
 	NVME_REQ_USERCMD		= (1 << 1),
 	NVME_MPATH_IO_STATS		= (1 << 2),
+	NVME_MPATH_CNT_ACTIVE	= (1 << 3),

This does not match the indentation above.

#14 lore raw
Quoted 9 lines · Christoph Hellwig

drivers/nvme/host/multipath.c static inline bool nvme_path_is_optimized(struct nvme_ns *ns) · line 400

Quoted 6 lines
drivers/nvme/host/multipath.c +6 -0
+	/*
+	 * queue-depth iopolicy does not need to reference ->current_path
+	 * but round-robin needs the last path used to advance to the
+	 * next one, and numa will continue to use the last path unless
+	 * it is or has become not optimized
+	 */

Can we please turn this into a full sentence? I.e.:

Quoted 8 lines

/*
* The queue-depth iopolicy does not need to reference ->current_path,
* but the round-robin iopolicy needs the last path used to advance to
* the next one, and numa will continue to use the last path unless
* it is or has become non-optimized.
*/

?

I can do that.

Quoted 8 lines
Quoted 8 lines
drivers/nvme/host/multipath.c +4 -0
+	if (iopolicy == NVME_IOPOLICY_QD)
+		return nvme_queue_depth_path(head);
+
+	node = numa_node_id();
  	ns = srcu_dereference(head->current_path[node], &head->srcu);
  	if (unlikely(!ns))
  		return __nvme_find_path(head, node);
Quoted 13 lines

drivers/nvme/host/multipath.c static inline bool nvme_path_is_optimized(struct nvme_ns *ns) · line 414

Quoted 7 lines
drivers/nvme/host/multipath.c +2 -1
-	if (READ_ONCE(head->subsys->iopolicy) == NVME_IOPOLICY_RR)
+	if (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);
  	return ns;

Also this is growing into the kind of spaghetti code that is on the fast path to become unmaintainable. I'd much rather see the srcu_dereference + __nvme_find_path duplicated and have a switch over the iopolicies with a separate helper for each of them here than the various ifs at different levels.

OK I will turn this into a switch statement.

Quoted 3 lines
Quoted 1 line

+static void nvme_subsys_iopolicy_update(struct nvme_subsystem *subsys, int iopolicy)

Overly long line here.

I can fix this.

Quoted 11 lines
Quoted 11 lines
drivers/nvme/host/multipath.c +11 -0
+{
+	struct nvme_ctrl *ctrl;
+	int old_iopolicy = READ_ONCE(subsys->iopolicy);
+
+	if (old_iopolicy == iopolicy)
+		return;
+
+	WRITE_ONCE(subsys->iopolicy, iopolicy);
+
+	/* iopolicy changes reset the counters and clear the mpath by design */
+	mutex_lock(&nvme_subsystems_lock);
Quoted 8 lines

drivers/nvme/host/multipath.c static ssize_t nvme_subsys_iopolicy_show(struct device *dev, · line 879

Quoted 5 lines
drivers/nvme/host/multipath.c +5 -0
+	list_for_each_entry(ctrl, &subsys->ctrls, subsys_entry) {
+		atomic_set(&ctrl->nr_active, 0);
+		nvme_mpath_clear_ctrl_paths(ctrl);
+	}
+	mutex_unlock(&nvme_subsystems_lock);

You probably want to take the lock over the iopolicy assignment to serialize it. And why do we need the atomic_set here?

Since we are targeting this to 6.11, I will work on refactoring this code.

I'll remove the atomic set here, but I may also add a WARN_ON_ONCE someplace just to be sure our assumptions about the nr_active counter state is correct. I agree with Keith that these counters need to be accurate in order for queue-depth to work.

Quoted 5 lines

drivers/nvme/host/multipath.c static ssize_t nvme_subsys_iopolicy_show(struct device *dev, · line 882

Quoted 3 lines
drivers/nvme/host/multipath.c +3 -0
+
+	pr_notice("%s: changed from %s to %s for subsysnqn %s\n", __func__,
+			nvme_iopolicy_names[old_iopolicy], nvme_iopolicy_names[iopolicy],

Pleae avoid the overly long line here as well.

Hmm... I thought I fixed this already. Will do.

Quoted 6 lines

drivers/nvme/host/nvme.h enum { · line 200

Quoted 4 lines

NVME_REQ_CANCELLED = (1 << 0),
NVME_REQ_USERCMD = (1 << 1),
NVME_MPATH_IO_STATS = (1 << 2),

drivers/nvme/host/nvme.h +1 -0
+	NVME_MPATH_CNT_ACTIVE	= (1 << 3),

This does not match the indentation above.

I must have my tab stop set incorrectly in .vimrc or something. I'll fix this.

/John

#15 lore raw
In reply to John Meneghini

Oh, and there is really something of with the patch format.

First the subject line is completely wrong, this should be:

nvme-multipath: implement "queue-depth" iopolicy.

Quoted 7 lines · John Meneghini

From: "Ewan D. Milne" <[email protected]>

[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]

Second the patch author needs to be the first signoff. I'm not entirely sure who is supposed to be the author, but either it should be Thomas, or if the patch has changed so much that it's someone else the original signoff should turn into a Co-Developed-by.

author Tested-by Marco Patalano Reviewed-by Randy Jennings Tested-by Jyoti Rani
#16 lore raw
Quoted 5 lines · Christoph Hellwig

Oh, and there is really something of with the patch format.

First the subject line is completely wrong, this should be:

nvme-multipath: implement "queue-depth" iopolicy.

Will fix.

Quoted 9 lines · John Meneghini
Quoted 7 lines

From: "Ewan D. Milne" <[email protected]>

[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]

Quoted 4 lines

Second the patch author needs to be the first signoff. I'm not entirely sure who is supposed to be the author, but either it should be Thomas, or if the patch has changed so much that it's someone else the original signoff should turn into a Co-Developed-by.

This patch was originally authored by Thomas Song, an intern who worked at Purestorage. The original patch was almost a year ago. Thomas no longer works for Purestorage, and he is no where to be found. Ewan and I have both made many changes since then, but the basic concept, which adds a controller scoped atomic counter in nvme_find_path(), remains the same. Ewan worked on the patch for several months, trying different counter implementations, all in an effort to avoid the atomic counter. However, nothing else worked and in the end this patch retains the essential features of Thomas Songs original patch.

I'd like Thomas to get credit for this patch, so I've changed the author from Ewan Milne to Thomas Song.

As for the Co-Developed-By: checkpatch.pl is insisting that that each "Co-developed-by:" be followed by a signoff.

jmeneghi:linux > scripts/checkpatch.pl --git HEAD
WARNING: Co-developed-by: must be immediately followed by Signed-off-by:
#26:

[jmeneghi: vairious changes and improvements, addressed review comments]

WARNING: Co-developed-by: must be immediately followed by Signed-off-by:
#28:

total: 0 errors, 2 warnings, 197 lines checked

So this is what I currently have for a message log.

Author: Thomas Song <[email protected]>
Date:   Tue Nov 7 16:23:29 2023 -0500

nvme-multipath: implemented new iopolicy "queue-depth"

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.

[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]