[PATCH v2 0/2] nvme-tcp: parallelize I/O queue connect

[PATCH v2 0/2] nvme-tcp: parallelize I/O queue connect

[Surabhi Gogte]

Similar to commit 2a8513091d2f ("nvme-rdma: parallelize I/O queue
allocation and startup"), I/O queue allocation and start steps
(including authentication) can be parallelized for nvme-tcp as well.

Patch 1 is a preparatory refactor: it splits the queue-count
negotiation and TLS PSK checks into helpers so that the individual
steps can be called directly from nvme_tcp_configure_io_queues(). No
functional change.

Patch 2 has the async implementation: it fans the per-queue alloc and
start out over an async domain, serializes the per-cpu queue-count in
nvme_tcp_set_queue_io_cpu() that can now run concurrently, and
switches sock_create_kern() to init_net.

Testing on a 64-core host with 64 IO-queues shows nvme-tcp connection
time reduced from 61ms to 11ms.

---
v2:
  - Resending the entire series to fix missing numbering in the
    original patch 2 subject line. No functional code changes.
---

Surabhi Gogte (2):
  nvme-tcp: refactor I/O queue setup path
  nvme-tcp: parallelize I/O queue allocation and startup

 drivers/nvme/host/tcp.c | 139 +++++++++++++++++++++++++++-------------
 1 file changed, 93 insertions(+), 46 deletions(-)

-- 
2.55.0

[PATCH v2 1/2] nvme-tcp: refactor I/O queue setup path

[Surabhi Gogte]

Split the I/O queue setup helpers apart so that the individual steps can
be called directly from nvme_tcp_configure_io_queues():

- Queue count negotiation moves out of nvme_tcp_alloc_io_queues() into a
  new nvme_tcp_io_queue_count(), leaving the allocator with just the
  per-queue allocation loop.
- TLS PSK validation moves out of __nvme_tcp_alloc_io_queues() into a
  new nvme_tcp_tls_check_psk().
- nvme_tcp_configure_io_queues() now calls the three steps directly in
  the same order as before.

Signed-off-by: Surabhi Gogte <sgogte@purestorage.com>
---
 drivers/nvme/host/tcp.c | 25 +++++++++++++++++++------
 1 file changed, 19 insertions(+), 6 deletions(-)

diff --git a/drivers/nvme/host/tcp.c b/drivers/nvme/host/tcp.c
index 5fda9661bdb7..354668ad29ac 100644
--- a/drivers/nvme/host/tcp.c
+++ b/drivers/nvme/host/tcp.c
@@ -2171,10 +2171,8 @@ static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
 	return ret;
 }
 
-static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
+static int nvme_tcp_tls_check_psk(struct nvme_ctrl *ctrl)
 {
-	int i, ret;
-
 	if (nvme_tcp_tls_configured(ctrl)) {
 		if (ctrl->opts->concat) {
 			/*
@@ -2196,6 +2194,13 @@ static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
 		}
 	}
 
+	return 0;
+}
+
+static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
+{
+	int i, ret;
+
 	for (i = 1; i < ctrl->queue_count; i++) {
 		ret = nvme_tcp_alloc_queue(ctrl, i,
 				ctrl->tls_pskid);
@@ -2212,7 +2217,7 @@ static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
 	return ret;
 }
 
-static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
+static int nvme_tcp_io_queue_count(struct nvme_ctrl *ctrl)
 {
 	unsigned int nr_io_queues;
 	int ret;
@@ -2234,14 +2239,22 @@ static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
 
 	nvmf_set_io_queues(ctrl->opts, nr_io_queues,
 			   to_tcp_ctrl(ctrl)->io_queues);
-	return __nvme_tcp_alloc_io_queues(ctrl);
+	return 0;
 }
 
 static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
 {
 	int ret, nr_queues;
 
-	ret = nvme_tcp_alloc_io_queues(ctrl);
+	ret = nvme_tcp_io_queue_count(ctrl);
+	if (ret)
+		return ret;
+
+	ret = nvme_tcp_tls_check_psk(ctrl);
+	if (ret)
+		return ret;
+
+	ret = __nvme_tcp_alloc_io_queues(ctrl);
 	if (ret)
 		return ret;
 
-- 
2.55.0

[PATCH v2 2/2] nvme-tcp: parallelize I/O queue allocation and startup

[Surabhi Gogte]

Similar to commit 2a8513091d2f ("nvme-rdma: parallelize I/O queue
allocation and startup"), refactor nvme tcp I/O queue setup to use async
API, combining allocation and startup into a single parallel operation
per queue. This reduces connection and reconnection setup time when
there are delays in establishing connections, which is especially
important for high-core-count hosts.

Key changes:
- Use async API to facilitate parallel calls for io queue setup.
- Add nvme_tcp_setup_ctx for propagating errors from async workers.
- Remove nvme_tcp_start_io_queues() and __nvme_tcp_alloc_io_queues();
  their logic is folded into nvme_tcp_setup_io_queues() and
  nvme_tcp_configure_io_queues().
- Allocate the io tag set before the queues so that the queue range is
  known, and only set up the reconnect grow case if the queue count
  actually increased.
- Serialize the cpu scan and claim in nvme_tcp_set_queue_io_cpu() with a
  spinlock, as concurrent callers would otherwise select the same cpu.
  The per-cpu counters no longer need to be atomics.
- Use init_net in nvme_tcp_alloc_queue() instead of the namespace of
  current, which is no longer the connecting task once the allocation
  runs from a worker. A controller is not guaranteed to be tied to a
  namespace, as the reconnect and error recovery paths already run from
  a workqueue in init_net.

Testing on a 64-core host with 64 IO-queues shows nvme-tcp connection
time reduced from 61ms to 11ms.

Signed-off-by: Surabhi Gogte <sgogte@purestorage.com>
---
 drivers/nvme/host/tcp.c | 126 +++++++++++++++++++++++++---------------
 1 file changed, 80 insertions(+), 46 deletions(-)

diff --git a/drivers/nvme/host/tcp.c b/drivers/nvme/host/tcp.c
index 354668ad29ac..30fe4c5abe0b 100644
--- a/drivers/nvme/host/tcp.c
+++ b/drivers/nvme/host/tcp.c
@@ -7,6 +7,7 @@
 #include <linux/module.h>
 #include <linux/init.h>
 #include <linux/slab.h>
+#include <linux/async.h>
 #include <linux/err.h>
 #include <linux/crc32.h>
 #include <linux/nvme-tcp.h>
@@ -54,7 +55,8 @@ MODULE_PARM_DESC(tls_handshake_timeout,
 		 "nvme TLS handshake timeout in seconds (default 10)");
 #endif
 
-static atomic_t nvme_tcp_cpu_queues[NR_CPUS];
+static int nvme_tcp_cpu_queues[NR_CPUS];
+static DEFINE_SPINLOCK(nvme_tcp_cpu_queues_lock);
 
 enum nvme_tcp_send_state {
 	NVME_TCP_SEND_CMD_PDU = 0,
@@ -154,6 +156,12 @@ struct nvme_tcp_queue {
 static DEFINE_MUTEX(nvme_tcp_ctrl_mutex);
 static LIST_HEAD_GUARDED(nvme_tcp_ctrl_list, nvme_tcp_ctrl_mutex);
 
+struct nvme_tcp_setup_ctx {
+	struct nvme_ctrl	*ctrl;
+	int			qid;
+	int			*err;
+};
+
 struct nvme_tcp_ctrl {
 	/* read only in the hot path */
 	struct nvme_tcp_queue	*queues;
@@ -1718,9 +1726,10 @@ static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
 		goto out;
 
 	/* Search for the least used cpu from the mq_map */
+	spin_lock(&nvme_tcp_cpu_queues_lock);
 	io_cpu = WORK_CPU_UNBOUND;
 	for_each_online_cpu(cpu) {
-		int num_queues = atomic_read(&nvme_tcp_cpu_queues[cpu]);
+		int num_queues = nvme_tcp_cpu_queues[cpu];
 
 		if (mq_map[cpu] != qid)
 			continue;
@@ -1731,9 +1740,10 @@ static void nvme_tcp_set_queue_io_cpu(struct nvme_tcp_queue *queue)
 	}
 	if (io_cpu != WORK_CPU_UNBOUND) {
 		queue->io_cpu = io_cpu;
-		atomic_inc(&nvme_tcp_cpu_queues[io_cpu]);
+		nvme_tcp_cpu_queues[io_cpu]++;
 		set_bit(NVME_TCP_Q_IO_CPU_SET, &queue->flags);
 	}
+	spin_unlock(&nvme_tcp_cpu_queues_lock);
 out:
 	dev_dbg(ctrl->ctrl.device, "queue %d: using cpu %d\n",
 		qid, queue->io_cpu);
@@ -1846,7 +1856,7 @@ static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl, int qid,
 		queue->cmnd_capsule_len = sizeof(struct nvme_command) +
 						NVME_TCP_ADMIN_CCSZ;
 
-	ret = sock_create_kern(current->nsproxy->net_ns,
+	ret = sock_create_kern(&init_net,
 			ctrl->addr.ss_family, SOCK_STREAM,
 			IPPROTO_TCP, &queue->sock);
 	if (ret) {
@@ -2010,8 +2020,11 @@ static void nvme_tcp_stop_queue_nowait(struct nvme_ctrl *nctrl, int qid)
 	if (!test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
 		return;
 
-	if (test_and_clear_bit(NVME_TCP_Q_IO_CPU_SET, &queue->flags))
-		atomic_dec(&nvme_tcp_cpu_queues[queue->io_cpu]);
+	if (test_and_clear_bit(NVME_TCP_Q_IO_CPU_SET, &queue->flags)) {
+		spin_lock(&nvme_tcp_cpu_queues_lock);
+		nvme_tcp_cpu_queues[queue->io_cpu]--;
+		spin_unlock(&nvme_tcp_cpu_queues_lock);
+	}
 
 	mutex_lock(&queue->queue_lock);
 	if (test_and_clear_bit(NVME_TCP_Q_LIVE, &queue->flags))
@@ -2118,25 +2131,6 @@ static void nvme_tcp_stop_io_queues(struct nvme_ctrl *ctrl)
 		nvme_tcp_wait_queue(ctrl, i);
 }
 
-static int nvme_tcp_start_io_queues(struct nvme_ctrl *ctrl,
-				    int first, int last)
-{
-	int i, ret;
-
-	for (i = first; i < last; i++) {
-		ret = nvme_tcp_start_queue(ctrl, i);
-		if (ret)
-			goto out_stop_queues;
-	}
-
-	return 0;
-
-out_stop_queues:
-	for (i--; i >= first; i--)
-		nvme_tcp_stop_queue(ctrl, i);
-	return ret;
-}
-
 static int nvme_tcp_alloc_admin_queue(struct nvme_ctrl *ctrl)
 {
 	int ret;
@@ -2197,22 +2191,64 @@ static int nvme_tcp_tls_check_psk(struct nvme_ctrl *ctrl)
 	return 0;
 }
 
-static int __nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
+static void nvme_tcp_setup_queue_async(void *data, async_cookie_t cookie)
 {
-	int i, ret;
+	struct nvme_tcp_setup_ctx *ctx = data;
+	struct nvme_ctrl *ctrl = ctx->ctrl;
+	int ret;
 
-	for (i = 1; i < ctrl->queue_count; i++) {
-		ret = nvme_tcp_alloc_queue(ctrl, i,
-				ctrl->tls_pskid);
-		if (ret)
-			goto out_free_queues;
+	ret = nvme_tcp_alloc_queue(ctrl, ctx->qid, ctrl->tls_pskid);
+	if (ret)
+		goto out_err;
+
+	ret = nvme_tcp_start_queue(ctrl, ctx->qid);
+	if (ret)
+		goto out_err;
+
+	return;
+
+out_err:
+	WRITE_ONCE(*ctx->err, ret);
+}
+
+static int nvme_tcp_setup_io_queues(struct nvme_ctrl *ctrl, unsigned int first,
+		unsigned int last)
+{
+	ASYNC_DOMAIN_EXCLUSIVE(queue_domain);
+	struct nvme_tcp_setup_ctx *ctxs;
+	int nr_queues = last - first;
+	int err = 0, i, ret;
+
+	ctxs = kmalloc_objs(*ctxs, nr_queues);
+	if (!ctxs)
+		return -ENOMEM;
+
+	for (i = 0; i < nr_queues; i++) {
+		ctxs[i].ctrl = ctrl;
+		ctxs[i].qid = first + i;
+		ctxs[i].err = &err;
+		async_schedule_domain(nvme_tcp_setup_queue_async, &ctxs[i],
+				&queue_domain);
 	}
 
+	async_synchronize_full_domain(&queue_domain);
+	kfree(ctxs);
+
+	ret = READ_ONCE(err);
+	if (ret)
+		goto out_free_queues;
+
 	return 0;
 
 out_free_queues:
-	for (i--; i >= 1; i--)
-		nvme_tcp_free_queue(ctrl, i);
+	for (i = first; i < last; i++) {
+		struct nvme_tcp_queue *queue = &to_tcp_ctrl(ctrl)->queues[i];
+
+		if (test_bit(NVME_TCP_Q_LIVE, &queue->flags))
+			nvme_tcp_stop_queue(ctrl, i);
+		if (test_bit(NVME_TCP_Q_ALLOCATED, &queue->flags))
+			nvme_tcp_free_queue(ctrl, i);
+	}
 
 	return ret;
 }
@@ -2254,10 +2290,6 @@ static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
 	if (ret)
 		return ret;
 
-	ret = __nvme_tcp_alloc_io_queues(ctrl);
-	if (ret)
-		return ret;
-
 	if (new) {
 		ret = nvme_alloc_io_tag_set(ctrl, &to_tcp_ctrl(ctrl)->tag_set,
 				&nvme_tcp_mq_ops,
@@ -2268,12 +2300,12 @@ static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
 	}
 
 	/*
-	 * Only start IO queues for which we have allocated the tagset
+	 * Only setup IO queues for which we have allocated the tagset
 	 * and limited it to the available queues. On reconnects, the
 	 * queue number might have changed.
 	 */
 	nr_queues = min(ctrl->tagset->nr_hw_queues + 1, ctrl->queue_count);
-	ret = nvme_tcp_start_io_queues(ctrl, 1, nr_queues);
+	ret = nvme_tcp_setup_io_queues(ctrl, 1, nr_queues);
 	if (ret)
 		goto out_cleanup_connect_q;
 
@@ -2297,12 +2329,14 @@ static int nvme_tcp_configure_io_queues(struct nvme_ctrl *ctrl, bool new)
 
 	/*
 	 * If the number of queues has increased (reconnect case)
-	 * start all new queues now.
+	 * setup all new queues now.
 	 */
-	ret = nvme_tcp_start_io_queues(ctrl, nr_queues,
-				       ctrl->tagset->nr_hw_queues + 1);
-	if (ret)
-		goto out_wait_freeze_timed_out;
+	if (ctrl->tagset->nr_hw_queues + 1 > nr_queues) {
+		ret = nvme_tcp_setup_io_queues(ctrl, nr_queues,
+				ctrl->tagset->nr_hw_queues + 1);
+		if (ret)
+			goto out_wait_freeze_timed_out;
+	}
 
 	return 0;
 
@@ -3140,7 +3174,7 @@ static int __init nvme_tcp_init_module(void)
 		return -ENOMEM;
 
 	for_each_possible_cpu(cpu)
-		atomic_set(&nvme_tcp_cpu_queues[cpu], 0);
+		nvme_tcp_cpu_queues[cpu] = 0;
 
 	nvmf_register_transport(&nvme_tcp_transport);
 	return 0;
-- 
2.55.0

Re: [PATCH v2 1/2] nvme-tcp: refactor I/O queue setup path

[Randy Jennings]

On Mon, Aug 24, 2026 at 3:57 PM Surabhi Gogte <sgogte@purestorage.com> wrote:
>
> Split the I/O queue setup helpers apart so that the individual steps can
> be called directly from nvme_tcp_configure_io_queues():
>
> - Queue count negotiation moves out of nvme_tcp_alloc_io_queues() into a
>   new nvme_tcp_io_queue_count(), leaving the allocator with just the
>   per-queue allocation loop.
> - TLS PSK validation moves out of __nvme_tcp_alloc_io_queues() into a
>   new nvme_tcp_tls_check_psk().
> - nvme_tcp_configure_io_queues() now calls the three steps directly in
>   the same order as before.
>
> Signed-off-by: Surabhi Gogte <sgogte@purestorage.com>
Reviewed-by: Randy Jennings <randyj@purestorage.com>

Re: [PATCH v2 2/2] nvme-tcp: parallelize I/O queue allocation and startup

[Randy Jennings]

On Mon, Aug 24, 2026 at 3:57 PM Surabhi Gogte <sgogte@purestorage.com> wrote:
>
> Similar to commit 2a8513091d2f ("nvme-rdma: parallelize I/O queue
> allocation and startup"), refactor nvme tcp I/O queue setup to use async
> API, combining allocation and startup into a single parallel operation
> per queue. This reduces connection and reconnection setup time when
> there are delays in establishing connections, which is especially
> important for high-core-count hosts.
>
> Key changes:
> - Use async API to facilitate parallel calls for io queue setup.
> - Add nvme_tcp_setup_ctx for propagating errors from async workers.
> - Remove nvme_tcp_start_io_queues() and __nvme_tcp_alloc_io_queues();
>   their logic is folded into nvme_tcp_setup_io_queues() and
>   nvme_tcp_configure_io_queues().
> - Allocate the io tag set before the queues so that the queue range is
>   known, and only set up the reconnect grow case if the queue count
>   actually increased.
> - Serialize the cpu scan and claim in nvme_tcp_set_queue_io_cpu() with a
>   spinlock, as concurrent callers would otherwise select the same cpu.
>   The per-cpu counters no longer need to be atomics.
> - Use init_net in nvme_tcp_alloc_queue() instead of the namespace of
>   current, which is no longer the connecting task once the allocation
>   runs from a worker. A controller is not guaranteed to be tied to a
>   namespace, as the reconnect and error recovery paths already run from
>   a workqueue in init_net.
>
> Testing on a 64-core host with 64 IO-queues shows nvme-tcp connection
> time reduced from 61ms to 11ms.
>
> Signed-off-by: Surabhi Gogte <sgogte@purestorage.com>
:~$
Reviewed-by: Randy Jennings <randyj@purestorage.com>