On 13/03/2024 10:55, Li Feng wrote:
> 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.
>
> Signed-off-by: Li Feng <[email protected]>
> ---
> drivers/nvme/host/tcp.c | 21 ++++++++++++++++++---
> 1 file changed, 18 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/nvme/host/tcp.c b/drivers/nvme/host/tcp.c
> index a6d596e05602..5eaa275f436f 100644
> --- a/drivers/nvme/host/tcp.c
> +++ b/drivers/nvme/host/tcp.c
> @@ -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)"
> +
> /*
> * 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;
I think we should have WQ_SYSFS exposed always. Add it in a seperate patch
that comes before this.
> +
> + nvme_tcp_wq = alloc_workqueue("nvme_tcp_wq", wq_flags, 0);
> if (!nvme_tcp_wq)
> return -ENOMEM;
>