Skip to contentSkip to conversation

nvme/tcp: Add wq_unbound modparam for nvme_tcp_wq

PATCH In review Li Feng 2 participants 2 messages Mar 13, 2024 NVMe

lore
Share & tools
lore

Li Feng

author
#0 lore raw

PATCH 1/1 nvme/tcp: Add wq_unbound modparam for nvme_tcp_wq

The default nvme_tcp_wq will use all CPUs to process tasks. Sometimes it is necessary to set CPU affinity to improve performance.

A new module parameter wq_unbound is added here. If set to true, users can configure cpu affinity through /sys/devices/virtual/workqueue/nvme_tcp_wq/cpumask.

---
 drivers/nvme/host/tcp.c | 21 ++++++++++++++++++---
 1 file changed, 18 insertions(+), 3 deletions(-)
drivers/nvme/host/tcp.c +18 -3
index a6d596e05602..5eaa275f436f 100644
@@ -36,6 +36,14 @@ static int so_priority;
 module_param(so_priority, int, 0644);
 MODULE_PARM_DESC(so_priority, "nvme tcp socket optimize priority");
 
+/*
+ * Use the unbound workqueue for nvme_tcp_wq, then we can set the cpu affinity
+ * from sysfs.
+ */
+static bool wq_unbound;
+module_param(wq_unbound, bool, 0644);
+MODULE_PARM_DESC(wq_unbound, "set unbound flag for nvme tcp work queue");
+
 /*
  * TLS handshake timeout
  */
@@ -1551,7 +1559,10 @@ static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
 	else if (nvme_tcp_poll_queue(queue))
 		n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
 				ctrl->io_queues[HCTX_TYPE_READ] - 1;
-	queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
+	if (wq_unbound)
+		queue->io_cpu = WORK_CPU_UNBOUND;
+	else
+		queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
 }
 
 static void nvme_tcp_tls_done(void *data, int status, key_serial_t pskid)
@@ -2790,6 +2801,8 @@ static struct nvmf_transport_ops nvme_tcp_transport = {
 
 static int __init nvme_tcp_init_module(void)
 {
+	unsigned int wq_flags = WQ_MEM_RECLAIM | WQ_HIGHPRI;
+
 	BUILD_BUG_ON(sizeof(struct nvme_tcp_hdr) != 8);
 	BUILD_BUG_ON(sizeof(struct nvme_tcp_cmd_pdu) != 72);
 	BUILD_BUG_ON(sizeof(struct nvme_tcp_data_pdu) != 24);
@@ -2799,8 +2812,10 @@ static int __init nvme_tcp_init_module(void)
 	BUILD_BUG_ON(sizeof(struct nvme_tcp_icresp_pdu) != 128);
 	BUILD_BUG_ON(sizeof(struct nvme_tcp_term_pdu) != 24);
 
-	nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
-			WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
+	if (wq_unbound)
+		wq_flags |= WQ_UNBOUND | WQ_SYSFS;
+
+	nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq", wq_flags, 0);
 	if (!nvme_tcp_wq)
 		return -ENOMEM;
 
Signature · 1 line
2.44.0
maintainer
#1 lore raw
Quoted 17 lines · Li Feng

The default nvme_tcp_wq will use all CPUs to process tasks. Sometimes it is necessary to set CPU affinity to improve performance.

A new module parameter wq_unbound is added here. If set to true, users can configure cpu affinity through /sys/devices/virtual/workqueue/nvme_tcp_wq/cpumask.

---
  drivers/nvme/host/tcp.c | 21 ++++++++++++++++++---
  1 file changed, 18 insertions(+), 3 deletions(-)
drivers/nvme/host/tcp.c +0 -0
index a6d596e05602..5eaa275f436f 100644

drivers/nvme/host/tcp.c static int so_priority; · line 45

Quoted 10 lines
drivers/nvme/host/tcp.c +7 -0
@@ -36,6 +36,14 @@ static int so_priority;
  module_param(so_priority, int, 0644);
  MODULE_PARM_DESC(so_priority, "nvme tcp socket optimize priority");
 
+/*
+ * Use the unbound workqueue for nvme_tcp_wq, then we can set the cpu affinity
+ * from sysfs.
+ */
+static bool wq_unbound;
+module_param(wq_unbound, bool, 0644);
+MODULE_PARM_DESC(wq_unbound, "set unbound flag for nvme tcp work queue");

"Use unbound workqueue for nvme-tcp IO context (default false)"

Quoted 26 lines
drivers/nvme/host/tcp.c +7 -1
+
  /*
   * TLS handshake timeout
   */
@@ -1551,7 +1559,10 @@ static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
  	else if (nvme_tcp_poll_queue(queue))
  		n = qid - ctrl->io_queues[HCTX_TYPE_DEFAULT] -
  				ctrl->io_queues[HCTX_TYPE_READ] - 1;
-	queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
+	if (wq_unbound)
+		queue->io_cpu = WORK_CPU_UNBOUND;
+	else
+		queue->io_cpu = cpumask_next_wrap(n - 1, cpu_online_mask, -1, false);
  }
 
  static void nvme_tcp_tls_done(void *data, int status, key_serial_t pskid)
@@ -2790,6 +2801,8 @@ static struct nvmf_transport_ops nvme_tcp_transport = {
 
  static int __init nvme_tcp_init_module(void)
  {
+	unsigned int wq_flags = WQ_MEM_RECLAIM | WQ_HIGHPRI;
+
  	BUILD_BUG_ON(sizeof(struct nvme_tcp_hdr) != 8);
  	BUILD_BUG_ON(sizeof(struct nvme_tcp_cmd_pdu) != 72);
  	BUILD_BUG_ON(sizeof(struct nvme_tcp_data_pdu) != 24);

drivers/nvme/host/tcp.c static int __init nvme_tcp_init_module(void) · line 2816

Quoted 7 lines
drivers/nvme/host/tcp.c +2 -2
@@ -2799,8 +2812,10 @@ static int __init nvme_tcp_init_module(void)
  	BUILD_BUG_ON(sizeof(struct nvme_tcp_icresp_pdu) != 128);
  	BUILD_BUG_ON(sizeof(struct nvme_tcp_term_pdu) != 24);
 
-	nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq",
-			WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
+	if (wq_unbound)
+		wq_flags |= WQ_UNBOUND | WQ_SYSFS;

I think we should have WQ_SYSFS exposed always. Add it in a seperate patch that comes before this.

Quoted 5 lines
drivers/nvme/host/tcp.c +2 -0
+
+	nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq", wq_flags, 0);
  	if (!nvme_tcp_wq)
  		return -ENOMEM;