nvme-tcp: support simple polling
authorSagi Grimberg <sagi@grimberg.me>
Wed, 3 Jul 2019 21:08:04 +0000 (14:08 -0700)
committerSagi Grimberg <sagi@grimberg.me>
Thu, 29 Aug 2019 19:55:01 +0000 (12:55 -0700)
Simple polling support via socket busy_poll interface.
Although we do not shutdown interrupts but simply hammer
the socket poll, we can sometimes find completions faster
than the normal interrupt driven RX path.

We add per queue nr_cqe counter that resets every time
RX path is invoked such that .poll callback can return it
to stay consistent with the semantics.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
drivers/nvme/host/tcp.c

index 0317721..1e2e5ab 100644 (file)
@@ -13,6 +13,7 @@
 #include <net/tcp.h>
 #include <linux/blk-mq.h>
 #include <crypto/hash.h>
+#include <net/busy_poll.h>
 
 #include "nvme.h"
 #include "fabrics.h"
@@ -72,6 +73,7 @@ struct nvme_tcp_queue {
        int                     pdu_offset;
        size_t                  data_remaining;
        size_t                  ddgst_remaining;
+       unsigned int            nr_cqe;
 
        /* send state */
        struct nvme_tcp_request *request;
@@ -438,6 +440,7 @@ static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
        }
 
        nvme_end_request(rq, cqe->status, cqe->result);
+       queue->nr_cqe++;
 
        return 0;
 }
@@ -696,8 +699,10 @@ static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
                        nvme_tcp_ddgst_final(queue->rcv_hash, &queue->exp_ddgst);
                        queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
                } else {
-                       if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS)
+                       if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
                                nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
+                               queue->nr_cqe++;
+                       }
                        nvme_tcp_init_recv_ctx(queue);
                }
        }
@@ -737,6 +742,7 @@ static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
                                                pdu->command_id);
 
                nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
+               queue->nr_cqe++;
        }
 
        nvme_tcp_init_recv_ctx(queue);
@@ -1026,6 +1032,7 @@ static int nvme_tcp_try_recv(struct nvme_tcp_queue *queue)
        rd_desc.arg.data = queue;
        rd_desc.count = 1;
        lock_sock(sk);
+       queue->nr_cqe = 0;
        consumed = sock->ops->read_sock(sk, &rd_desc, nvme_tcp_recv_skb);
        release_sock(sk);
        return consumed;
@@ -1367,6 +1374,7 @@ static int nvme_tcp_alloc_queue(struct nvme_ctrl *nctrl,
        queue->sock->sk->sk_data_ready = nvme_tcp_data_ready;
        queue->sock->sk->sk_state_change = nvme_tcp_state_change;
        queue->sock->sk->sk_write_space = nvme_tcp_write_space;
+       queue->sock->sk->sk_ll_usec = 1;
        write_unlock_bh(&queue->sock->sk->sk_callback_lock);
 
        return 0;
@@ -1465,7 +1473,7 @@ static struct blk_mq_tag_set *nvme_tcp_alloc_tagset(struct nvme_ctrl *nctrl,
                set->driver_data = ctrl;
                set->nr_hw_queues = nctrl->queue_count - 1;
                set->timeout = NVME_IO_TIMEOUT;
-               set->nr_maps = 2 /* default + read */;
+               set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
        }
 
        ret = blk_mq_alloc_tag_set(set);
@@ -1564,6 +1572,7 @@ static unsigned int nvme_tcp_nr_io_queues(struct nvme_ctrl *ctrl)
 
        nr_io_queues = min(ctrl->opts->nr_io_queues, num_online_cpus());
        nr_io_queues += min(ctrl->opts->nr_write_queues, num_online_cpus());
+       nr_io_queues += min(ctrl->opts->nr_poll_queues, num_online_cpus());
 
        return nr_io_queues;
 }
@@ -1595,6 +1604,12 @@ static void nvme_tcp_set_io_queues(struct nvme_ctrl *nctrl,
                        min(opts->nr_io_queues, nr_io_queues);
                nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
        }
+
+       if (opts->nr_poll_queues && nr_io_queues) {
+               /* map dedicated poll queues only if we have queues left */
+               ctrl->io_queues[HCTX_TYPE_POLL] =
+                       min(opts->nr_poll_queues, nr_io_queues);
+       }
 }
 
 static int nvme_tcp_alloc_io_queues(struct nvme_ctrl *ctrl)
@@ -2142,14 +2157,36 @@ static int nvme_tcp_map_queues(struct blk_mq_tag_set *set)
        blk_mq_map_queues(&set->map[HCTX_TYPE_DEFAULT]);
        blk_mq_map_queues(&set->map[HCTX_TYPE_READ]);
 
+       if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
+               /* map dedicated poll queues only if we have queues left */
+               set->map[HCTX_TYPE_POLL].nr_queues =
+                               ctrl->io_queues[HCTX_TYPE_POLL];
+               set->map[HCTX_TYPE_POLL].queue_offset =
+                       ctrl->io_queues[HCTX_TYPE_DEFAULT] +
+                       ctrl->io_queues[HCTX_TYPE_READ];
+               blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
+       }
+
        dev_info(ctrl->ctrl.device,
-               "mapped %d/%d default/read queues.\n",
+               "mapped %d/%d/%d default/read/poll queues.\n",
                ctrl->io_queues[HCTX_TYPE_DEFAULT],
-               ctrl->io_queues[HCTX_TYPE_READ]);
+               ctrl->io_queues[HCTX_TYPE_READ],
+               ctrl->io_queues[HCTX_TYPE_POLL]);
 
        return 0;
 }
 
+static int nvme_tcp_poll(struct blk_mq_hw_ctx *hctx)
+{
+       struct nvme_tcp_queue *queue = hctx->driver_data;
+       struct sock *sk = queue->sock->sk;
+
+       if (sk_can_busy_loop(sk) && skb_queue_empty(&sk->sk_receive_queue))
+               sk_busy_loop(sk, true);
+       nvme_tcp_try_recv(queue);
+       return queue->nr_cqe;
+}
+
 static struct blk_mq_ops nvme_tcp_mq_ops = {
        .queue_rq       = nvme_tcp_queue_rq,
        .complete       = nvme_complete_rq,
@@ -2158,6 +2195,7 @@ static struct blk_mq_ops nvme_tcp_mq_ops = {
        .init_hctx      = nvme_tcp_init_hctx,
        .timeout        = nvme_tcp_timeout,
        .map_queues     = nvme_tcp_map_queues,
+       .poll           = nvme_tcp_poll,
 };
 
 static struct blk_mq_ops nvme_tcp_admin_mq_ops = {
@@ -2211,7 +2249,8 @@ static struct nvme_ctrl *nvme_tcp_create_ctrl(struct device *dev,
 
        INIT_LIST_HEAD(&ctrl->list);
        ctrl->ctrl.opts = opts;
-       ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues + 1;
+       ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
+                               opts->nr_poll_queues + 1;
        ctrl->ctrl.sqsize = opts->queue_size - 1;
        ctrl->ctrl.kato = opts->kato;
 
@@ -2305,7 +2344,7 @@ static struct nvmf_transport_ops nvme_tcp_transport = {
        .allowed_opts   = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
                          NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
                          NVMF_OPT_HDR_DIGEST | NVMF_OPT_DATA_DIGEST |
-                         NVMF_OPT_NR_WRITE_QUEUES,
+                         NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES,
        .create_ctrl    = nvme_tcp_create_ctrl,
 };