2 * NVMe over Fabrics RDMA host code.
3 * Copyright (c) 2015-2016 HGST, a Western Digital Company.
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 #include <linux/module.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <rdma/mr_pool.h>
19 #include <linux/err.h>
20 #include <linux/string.h>
21 #include <linux/atomic.h>
22 #include <linux/blk-mq.h>
23 #include <linux/blk-mq-rdma.h>
24 #include <linux/types.h>
25 #include <linux/list.h>
26 #include <linux/mutex.h>
27 #include <linux/scatterlist.h>
28 #include <linux/nvme.h>
29 #include <asm/unaligned.h>
31 #include <rdma/ib_verbs.h>
32 #include <rdma/rdma_cm.h>
33 #include <linux/nvme-rdma.h>
39 #define NVME_RDMA_CONNECT_TIMEOUT_MS 3000 /* 3 second */
41 #define NVME_RDMA_MAX_SEGMENTS 256
43 #define NVME_RDMA_MAX_INLINE_SEGMENTS 4
45 struct nvme_rdma_device {
46 struct ib_device *dev;
49 struct list_head entry;
50 unsigned int num_inline_segments;
59 struct nvme_rdma_queue;
60 struct nvme_rdma_request {
61 struct nvme_request req;
63 struct nvme_rdma_qe sqe;
64 union nvme_result result;
67 struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
70 struct ib_reg_wr reg_wr;
71 struct ib_cqe reg_cqe;
72 struct nvme_rdma_queue *queue;
73 struct sg_table sg_table;
74 struct scatterlist first_sgl[];
77 enum nvme_rdma_queue_flags {
78 NVME_RDMA_Q_ALLOCATED = 0,
80 NVME_RDMA_Q_TR_READY = 2,
83 struct nvme_rdma_queue {
84 struct nvme_rdma_qe *rsp_ring;
86 size_t cmnd_capsule_len;
87 struct nvme_rdma_ctrl *ctrl;
88 struct nvme_rdma_device *device;
93 struct rdma_cm_id *cm_id;
95 struct completion cm_done;
98 struct nvme_rdma_ctrl {
99 /* read only in the hot path */
100 struct nvme_rdma_queue *queues;
102 /* other member variables */
103 struct blk_mq_tag_set tag_set;
104 struct work_struct err_work;
106 struct nvme_rdma_qe async_event_sqe;
108 struct delayed_work reconnect_work;
110 struct list_head list;
112 struct blk_mq_tag_set admin_tag_set;
113 struct nvme_rdma_device *device;
117 struct sockaddr_storage addr;
118 struct sockaddr_storage src_addr;
120 struct nvme_ctrl ctrl;
121 bool use_inline_data;
122 u32 io_queues[HCTX_MAX_TYPES];
125 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
127 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
130 static LIST_HEAD(device_list);
131 static DEFINE_MUTEX(device_list_mutex);
133 static LIST_HEAD(nvme_rdma_ctrl_list);
134 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
137 * Disabling this option makes small I/O goes faster, but is fundamentally
138 * unsafe. With it turned off we will have to register a global rkey that
139 * allows read and write access to all physical memory.
141 static bool register_always = true;
142 module_param(register_always, bool, 0444);
143 MODULE_PARM_DESC(register_always,
144 "Use memory registration even for contiguous memory regions");
146 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
147 struct rdma_cm_event *event);
148 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
150 static const struct blk_mq_ops nvme_rdma_mq_ops;
151 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
153 /* XXX: really should move to a generic header sooner or later.. */
154 static inline void put_unaligned_le24(u32 val, u8 *p)
161 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
163 return queue - queue->ctrl->queues;
166 static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue)
168 return nvme_rdma_queue_idx(queue) >
169 queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] +
170 queue->ctrl->io_queues[HCTX_TYPE_READ];
173 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
175 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
178 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
179 size_t capsule_size, enum dma_data_direction dir)
181 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
185 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
186 size_t capsule_size, enum dma_data_direction dir)
188 qe->data = kzalloc(capsule_size, GFP_KERNEL);
192 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
193 if (ib_dma_mapping_error(ibdev, qe->dma)) {
202 static void nvme_rdma_free_ring(struct ib_device *ibdev,
203 struct nvme_rdma_qe *ring, size_t ib_queue_size,
204 size_t capsule_size, enum dma_data_direction dir)
208 for (i = 0; i < ib_queue_size; i++)
209 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
213 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
214 size_t ib_queue_size, size_t capsule_size,
215 enum dma_data_direction dir)
217 struct nvme_rdma_qe *ring;
220 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
224 for (i = 0; i < ib_queue_size; i++) {
225 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
232 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
236 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
238 pr_debug("QP event %s (%d)\n",
239 ib_event_msg(event->event), event->event);
243 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
247 ret = wait_for_completion_interruptible_timeout(&queue->cm_done,
248 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
253 WARN_ON_ONCE(queue->cm_error > 0);
254 return queue->cm_error;
257 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
259 struct nvme_rdma_device *dev = queue->device;
260 struct ib_qp_init_attr init_attr;
263 memset(&init_attr, 0, sizeof(init_attr));
264 init_attr.event_handler = nvme_rdma_qp_event;
266 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
268 init_attr.cap.max_recv_wr = queue->queue_size + 1;
269 init_attr.cap.max_recv_sge = 1;
270 init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
271 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
272 init_attr.qp_type = IB_QPT_RC;
273 init_attr.send_cq = queue->ib_cq;
274 init_attr.recv_cq = queue->ib_cq;
276 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
278 queue->qp = queue->cm_id->qp;
282 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
283 struct request *rq, unsigned int hctx_idx)
285 struct nvme_rdma_ctrl *ctrl = set->driver_data;
286 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
287 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
288 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
289 struct nvme_rdma_device *dev = queue->device;
291 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
295 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
296 struct request *rq, unsigned int hctx_idx,
297 unsigned int numa_node)
299 struct nvme_rdma_ctrl *ctrl = set->driver_data;
300 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
301 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
302 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
303 struct nvme_rdma_device *dev = queue->device;
304 struct ib_device *ibdev = dev->dev;
307 nvme_req(rq)->ctrl = &ctrl->ctrl;
308 ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
318 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
319 unsigned int hctx_idx)
321 struct nvme_rdma_ctrl *ctrl = data;
322 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
324 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
326 hctx->driver_data = queue;
330 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
331 unsigned int hctx_idx)
333 struct nvme_rdma_ctrl *ctrl = data;
334 struct nvme_rdma_queue *queue = &ctrl->queues[0];
336 BUG_ON(hctx_idx != 0);
338 hctx->driver_data = queue;
342 static void nvme_rdma_free_dev(struct kref *ref)
344 struct nvme_rdma_device *ndev =
345 container_of(ref, struct nvme_rdma_device, ref);
347 mutex_lock(&device_list_mutex);
348 list_del(&ndev->entry);
349 mutex_unlock(&device_list_mutex);
351 ib_dealloc_pd(ndev->pd);
355 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
357 kref_put(&dev->ref, nvme_rdma_free_dev);
360 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
362 return kref_get_unless_zero(&dev->ref);
365 static struct nvme_rdma_device *
366 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
368 struct nvme_rdma_device *ndev;
370 mutex_lock(&device_list_mutex);
371 list_for_each_entry(ndev, &device_list, entry) {
372 if (ndev->dev->node_guid == cm_id->device->node_guid &&
373 nvme_rdma_dev_get(ndev))
377 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
381 ndev->dev = cm_id->device;
382 kref_init(&ndev->ref);
384 ndev->pd = ib_alloc_pd(ndev->dev,
385 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
386 if (IS_ERR(ndev->pd))
389 if (!(ndev->dev->attrs.device_cap_flags &
390 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
391 dev_err(&ndev->dev->dev,
392 "Memory registrations not supported.\n");
396 ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
397 ndev->dev->attrs.max_send_sge - 1);
398 list_add(&ndev->entry, &device_list);
400 mutex_unlock(&device_list_mutex);
404 ib_dealloc_pd(ndev->pd);
408 mutex_unlock(&device_list_mutex);
412 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
414 struct nvme_rdma_device *dev;
415 struct ib_device *ibdev;
417 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
423 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
426 * The cm_id object might have been destroyed during RDMA connection
427 * establishment error flow to avoid getting other cma events, thus
428 * the destruction of the QP shouldn't use rdma_cm API.
430 ib_destroy_qp(queue->qp);
431 ib_free_cq(queue->ib_cq);
433 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
434 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
436 nvme_rdma_dev_put(dev);
439 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev)
441 return min_t(u32, NVME_RDMA_MAX_SEGMENTS,
442 ibdev->attrs.max_fast_reg_page_list_len);
445 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
447 struct ib_device *ibdev;
448 const int send_wr_factor = 3; /* MR, SEND, INV */
449 const int cq_factor = send_wr_factor + 1; /* + RECV */
450 int comp_vector, idx = nvme_rdma_queue_idx(queue);
451 enum ib_poll_context poll_ctx;
454 queue->device = nvme_rdma_find_get_device(queue->cm_id);
455 if (!queue->device) {
456 dev_err(queue->cm_id->device->dev.parent,
457 "no client data found!\n");
458 return -ECONNREFUSED;
460 ibdev = queue->device->dev;
463 * Spread I/O queues completion vectors according their queue index.
464 * Admin queues can always go on completion vector 0.
466 comp_vector = idx == 0 ? idx : idx - 1;
468 /* Polling queues need direct cq polling context */
469 if (nvme_rdma_poll_queue(queue))
470 poll_ctx = IB_POLL_DIRECT;
472 poll_ctx = IB_POLL_SOFTIRQ;
474 /* +1 for ib_stop_cq */
475 queue->ib_cq = ib_alloc_cq(ibdev, queue,
476 cq_factor * queue->queue_size + 1,
477 comp_vector, poll_ctx);
478 if (IS_ERR(queue->ib_cq)) {
479 ret = PTR_ERR(queue->ib_cq);
483 ret = nvme_rdma_create_qp(queue, send_wr_factor);
485 goto out_destroy_ib_cq;
487 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
488 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
489 if (!queue->rsp_ring) {
494 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
497 nvme_rdma_get_max_fr_pages(ibdev));
499 dev_err(queue->ctrl->ctrl.device,
500 "failed to initialize MR pool sized %d for QID %d\n",
501 queue->queue_size, idx);
502 goto out_destroy_ring;
505 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
510 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
511 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
513 rdma_destroy_qp(queue->cm_id);
515 ib_free_cq(queue->ib_cq);
517 nvme_rdma_dev_put(queue->device);
521 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
522 int idx, size_t queue_size)
524 struct nvme_rdma_queue *queue;
525 struct sockaddr *src_addr = NULL;
528 queue = &ctrl->queues[idx];
530 init_completion(&queue->cm_done);
533 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
535 queue->cmnd_capsule_len = sizeof(struct nvme_command);
537 queue->queue_size = queue_size;
539 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
540 RDMA_PS_TCP, IB_QPT_RC);
541 if (IS_ERR(queue->cm_id)) {
542 dev_info(ctrl->ctrl.device,
543 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
544 return PTR_ERR(queue->cm_id);
547 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
548 src_addr = (struct sockaddr *)&ctrl->src_addr;
550 queue->cm_error = -ETIMEDOUT;
551 ret = rdma_resolve_addr(queue->cm_id, src_addr,
552 (struct sockaddr *)&ctrl->addr,
553 NVME_RDMA_CONNECT_TIMEOUT_MS);
555 dev_info(ctrl->ctrl.device,
556 "rdma_resolve_addr failed (%d).\n", ret);
557 goto out_destroy_cm_id;
560 ret = nvme_rdma_wait_for_cm(queue);
562 dev_info(ctrl->ctrl.device,
563 "rdma connection establishment failed (%d)\n", ret);
564 goto out_destroy_cm_id;
567 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
572 rdma_destroy_id(queue->cm_id);
573 nvme_rdma_destroy_queue_ib(queue);
577 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
579 if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
582 rdma_disconnect(queue->cm_id);
583 ib_drain_qp(queue->qp);
586 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
588 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
591 nvme_rdma_destroy_queue_ib(queue);
592 rdma_destroy_id(queue->cm_id);
595 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
599 for (i = 1; i < ctrl->ctrl.queue_count; i++)
600 nvme_rdma_free_queue(&ctrl->queues[i]);
603 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
607 for (i = 1; i < ctrl->ctrl.queue_count; i++)
608 nvme_rdma_stop_queue(&ctrl->queues[i]);
611 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
613 struct nvme_rdma_queue *queue = &ctrl->queues[idx];
614 bool poll = nvme_rdma_poll_queue(queue);
618 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx, poll);
620 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
623 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
625 dev_info(ctrl->ctrl.device,
626 "failed to connect queue: %d ret=%d\n", idx, ret);
630 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
634 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
635 ret = nvme_rdma_start_queue(ctrl, i);
637 goto out_stop_queues;
643 for (i--; i >= 1; i--)
644 nvme_rdma_stop_queue(&ctrl->queues[i]);
648 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
650 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
651 struct ib_device *ibdev = ctrl->device->dev;
652 unsigned int nr_io_queues;
655 nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
658 * we map queues according to the device irq vectors for
659 * optimal locality so we don't need more queues than
660 * completion vectors.
662 nr_io_queues = min_t(unsigned int, nr_io_queues,
663 ibdev->num_comp_vectors);
665 if (opts->nr_write_queues) {
666 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
667 min(opts->nr_write_queues, nr_io_queues);
668 nr_io_queues += ctrl->io_queues[HCTX_TYPE_DEFAULT];
670 ctrl->io_queues[HCTX_TYPE_DEFAULT] = nr_io_queues;
673 ctrl->io_queues[HCTX_TYPE_READ] = nr_io_queues;
675 if (opts->nr_poll_queues) {
676 ctrl->io_queues[HCTX_TYPE_POLL] =
677 min(opts->nr_poll_queues, num_online_cpus());
678 nr_io_queues += ctrl->io_queues[HCTX_TYPE_POLL];
681 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
685 ctrl->ctrl.queue_count = nr_io_queues + 1;
686 if (ctrl->ctrl.queue_count < 2)
689 dev_info(ctrl->ctrl.device,
690 "creating %d I/O queues.\n", nr_io_queues);
692 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
693 ret = nvme_rdma_alloc_queue(ctrl, i,
694 ctrl->ctrl.sqsize + 1);
696 goto out_free_queues;
702 for (i--; i >= 1; i--)
703 nvme_rdma_free_queue(&ctrl->queues[i]);
708 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl,
709 struct blk_mq_tag_set *set)
711 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
713 blk_mq_free_tag_set(set);
714 nvme_rdma_dev_put(ctrl->device);
717 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
720 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
721 struct blk_mq_tag_set *set;
725 set = &ctrl->admin_tag_set;
726 memset(set, 0, sizeof(*set));
727 set->ops = &nvme_rdma_admin_mq_ops;
728 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
729 set->reserved_tags = 2; /* connect + keep-alive */
730 set->numa_node = nctrl->numa_node;
731 set->cmd_size = sizeof(struct nvme_rdma_request) +
732 SG_CHUNK_SIZE * sizeof(struct scatterlist);
733 set->driver_data = ctrl;
734 set->nr_hw_queues = 1;
735 set->timeout = ADMIN_TIMEOUT;
736 set->flags = BLK_MQ_F_NO_SCHED;
738 set = &ctrl->tag_set;
739 memset(set, 0, sizeof(*set));
740 set->ops = &nvme_rdma_mq_ops;
741 set->queue_depth = nctrl->sqsize + 1;
742 set->reserved_tags = 1; /* fabric connect */
743 set->numa_node = nctrl->numa_node;
744 set->flags = BLK_MQ_F_SHOULD_MERGE;
745 set->cmd_size = sizeof(struct nvme_rdma_request) +
746 SG_CHUNK_SIZE * sizeof(struct scatterlist);
747 set->driver_data = ctrl;
748 set->nr_hw_queues = nctrl->queue_count - 1;
749 set->timeout = NVME_IO_TIMEOUT;
750 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
753 ret = blk_mq_alloc_tag_set(set);
758 * We need a reference on the device as long as the tag_set is alive,
759 * as the MRs in the request structures need a valid ib_device.
761 ret = nvme_rdma_dev_get(ctrl->device);
764 goto out_free_tagset;
770 blk_mq_free_tag_set(set);
775 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
779 blk_cleanup_queue(ctrl->ctrl.admin_q);
780 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
782 if (ctrl->async_event_sqe.data) {
783 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
784 sizeof(struct nvme_command), DMA_TO_DEVICE);
785 ctrl->async_event_sqe.data = NULL;
787 nvme_rdma_free_queue(&ctrl->queues[0]);
790 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
795 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
799 ctrl->device = ctrl->queues[0].device;
800 ctrl->ctrl.numa_node = dev_to_node(ctrl->device->dev->dma_device);
802 ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
804 error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
805 sizeof(struct nvme_command), DMA_TO_DEVICE);
810 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
811 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
812 error = PTR_ERR(ctrl->ctrl.admin_tagset);
813 goto out_free_async_qe;
816 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
817 if (IS_ERR(ctrl->ctrl.admin_q)) {
818 error = PTR_ERR(ctrl->ctrl.admin_q);
819 goto out_free_tagset;
823 error = nvme_rdma_start_queue(ctrl, 0);
825 goto out_cleanup_queue;
827 error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
830 dev_err(ctrl->ctrl.device,
831 "prop_get NVME_REG_CAP failed\n");
836 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
838 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
842 ctrl->ctrl.max_hw_sectors =
843 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
845 error = nvme_init_identify(&ctrl->ctrl);
852 nvme_rdma_stop_queue(&ctrl->queues[0]);
855 blk_cleanup_queue(ctrl->ctrl.admin_q);
858 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
860 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
861 sizeof(struct nvme_command), DMA_TO_DEVICE);
862 ctrl->async_event_sqe.data = NULL;
864 nvme_rdma_free_queue(&ctrl->queues[0]);
868 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
872 blk_cleanup_queue(ctrl->ctrl.connect_q);
873 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
875 nvme_rdma_free_io_queues(ctrl);
878 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
882 ret = nvme_rdma_alloc_io_queues(ctrl);
887 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
888 if (IS_ERR(ctrl->ctrl.tagset)) {
889 ret = PTR_ERR(ctrl->ctrl.tagset);
890 goto out_free_io_queues;
893 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
894 if (IS_ERR(ctrl->ctrl.connect_q)) {
895 ret = PTR_ERR(ctrl->ctrl.connect_q);
896 goto out_free_tag_set;
899 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
900 ctrl->ctrl.queue_count - 1);
903 ret = nvme_rdma_start_io_queues(ctrl);
905 goto out_cleanup_connect_q;
909 out_cleanup_connect_q:
911 blk_cleanup_queue(ctrl->ctrl.connect_q);
914 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
916 nvme_rdma_free_io_queues(ctrl);
920 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
923 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
924 nvme_rdma_stop_queue(&ctrl->queues[0]);
925 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set, nvme_cancel_request,
927 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
928 nvme_rdma_destroy_admin_queue(ctrl, remove);
931 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
934 if (ctrl->ctrl.queue_count > 1) {
935 nvme_stop_queues(&ctrl->ctrl);
936 nvme_rdma_stop_io_queues(ctrl);
937 blk_mq_tagset_busy_iter(&ctrl->tag_set, nvme_cancel_request,
940 nvme_start_queues(&ctrl->ctrl);
941 nvme_rdma_destroy_io_queues(ctrl, remove);
945 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
947 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
949 cancel_work_sync(&ctrl->err_work);
950 cancel_delayed_work_sync(&ctrl->reconnect_work);
953 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
955 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
957 if (list_empty(&ctrl->list))
960 mutex_lock(&nvme_rdma_ctrl_mutex);
961 list_del(&ctrl->list);
962 mutex_unlock(&nvme_rdma_ctrl_mutex);
964 nvmf_free_options(nctrl->opts);
970 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
972 /* If we are resetting/deleting then do nothing */
973 if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
974 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
975 ctrl->ctrl.state == NVME_CTRL_LIVE);
979 if (nvmf_should_reconnect(&ctrl->ctrl)) {
980 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
981 ctrl->ctrl.opts->reconnect_delay);
982 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
983 ctrl->ctrl.opts->reconnect_delay * HZ);
985 nvme_delete_ctrl(&ctrl->ctrl);
989 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
994 ret = nvme_rdma_configure_admin_queue(ctrl, new);
998 if (ctrl->ctrl.icdoff) {
999 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1003 if (!(ctrl->ctrl.sgls & (1 << 2))) {
1004 dev_err(ctrl->ctrl.device,
1005 "Mandatory keyed sgls are not supported!\n");
1009 if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1010 dev_warn(ctrl->ctrl.device,
1011 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1012 ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1015 if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1016 dev_warn(ctrl->ctrl.device,
1017 "sqsize %u > ctrl maxcmd %u, clamping down\n",
1018 ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1019 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1022 if (ctrl->ctrl.sgls & (1 << 20))
1023 ctrl->use_inline_data = true;
1025 if (ctrl->ctrl.queue_count > 1) {
1026 ret = nvme_rdma_configure_io_queues(ctrl, new);
1031 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1033 /* state change failure is ok if we're in DELETING state */
1034 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1039 nvme_start_ctrl(&ctrl->ctrl);
1043 if (ctrl->ctrl.queue_count > 1)
1044 nvme_rdma_destroy_io_queues(ctrl, new);
1046 nvme_rdma_stop_queue(&ctrl->queues[0]);
1047 nvme_rdma_destroy_admin_queue(ctrl, new);
1051 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1053 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1054 struct nvme_rdma_ctrl, reconnect_work);
1056 ++ctrl->ctrl.nr_reconnects;
1058 if (nvme_rdma_setup_ctrl(ctrl, false))
1061 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1062 ctrl->ctrl.nr_reconnects);
1064 ctrl->ctrl.nr_reconnects = 0;
1069 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1070 ctrl->ctrl.nr_reconnects);
1071 nvme_rdma_reconnect_or_remove(ctrl);
1074 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1076 struct nvme_rdma_ctrl *ctrl = container_of(work,
1077 struct nvme_rdma_ctrl, err_work);
1079 nvme_stop_keep_alive(&ctrl->ctrl);
1080 nvme_rdma_teardown_io_queues(ctrl, false);
1081 nvme_start_queues(&ctrl->ctrl);
1082 nvme_rdma_teardown_admin_queue(ctrl, false);
1084 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1085 /* state change failure is ok if we're in DELETING state */
1086 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1090 nvme_rdma_reconnect_or_remove(ctrl);
1093 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1095 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1098 queue_work(nvme_wq, &ctrl->err_work);
1101 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1104 struct nvme_rdma_queue *queue = cq->cq_context;
1105 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1107 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1108 dev_info(ctrl->ctrl.device,
1109 "%s for CQE 0x%p failed with status %s (%d)\n",
1111 ib_wc_status_msg(wc->status), wc->status);
1112 nvme_rdma_error_recovery(ctrl);
1115 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1117 if (unlikely(wc->status != IB_WC_SUCCESS))
1118 nvme_rdma_wr_error(cq, wc, "MEMREG");
1121 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1123 struct nvme_rdma_request *req =
1124 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1125 struct request *rq = blk_mq_rq_from_pdu(req);
1127 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1128 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1132 if (refcount_dec_and_test(&req->ref))
1133 nvme_end_request(rq, req->status, req->result);
1137 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1138 struct nvme_rdma_request *req)
1140 struct ib_send_wr wr = {
1141 .opcode = IB_WR_LOCAL_INV,
1144 .send_flags = IB_SEND_SIGNALED,
1145 .ex.invalidate_rkey = req->mr->rkey,
1148 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1149 wr.wr_cqe = &req->reg_cqe;
1151 return ib_post_send(queue->qp, &wr, NULL);
1154 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1157 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1158 struct nvme_rdma_device *dev = queue->device;
1159 struct ib_device *ibdev = dev->dev;
1161 if (!blk_rq_payload_bytes(rq))
1165 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1169 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1170 req->nents, rq_data_dir(rq) ==
1171 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1173 nvme_cleanup_cmd(rq);
1174 sg_free_table_chained(&req->sg_table, true);
1177 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1179 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1182 put_unaligned_le24(0, sg->length);
1183 put_unaligned_le32(0, sg->key);
1184 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1188 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1189 struct nvme_rdma_request *req, struct nvme_command *c,
1192 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1193 struct scatterlist *sgl = req->sg_table.sgl;
1194 struct ib_sge *sge = &req->sge[1];
1198 for (i = 0; i < count; i++, sgl++, sge++) {
1199 sge->addr = sg_dma_address(sgl);
1200 sge->length = sg_dma_len(sgl);
1201 sge->lkey = queue->device->pd->local_dma_lkey;
1205 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1206 sg->length = cpu_to_le32(len);
1207 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1209 req->num_sge += count;
1213 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1214 struct nvme_rdma_request *req, struct nvme_command *c)
1216 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1218 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1219 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1220 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1221 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1225 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1226 struct nvme_rdma_request *req, struct nvme_command *c,
1229 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1232 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1233 if (WARN_ON_ONCE(!req->mr))
1237 * Align the MR to a 4K page size to match the ctrl page size and
1238 * the block virtual boundary.
1240 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1241 if (unlikely(nr < count)) {
1242 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1249 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1251 req->reg_cqe.done = nvme_rdma_memreg_done;
1252 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1253 req->reg_wr.wr.opcode = IB_WR_REG_MR;
1254 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1255 req->reg_wr.wr.num_sge = 0;
1256 req->reg_wr.mr = req->mr;
1257 req->reg_wr.key = req->mr->rkey;
1258 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1259 IB_ACCESS_REMOTE_READ |
1260 IB_ACCESS_REMOTE_WRITE;
1262 sg->addr = cpu_to_le64(req->mr->iova);
1263 put_unaligned_le24(req->mr->length, sg->length);
1264 put_unaligned_le32(req->mr->rkey, sg->key);
1265 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1266 NVME_SGL_FMT_INVALIDATE;
1271 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1272 struct request *rq, struct nvme_command *c)
1274 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1275 struct nvme_rdma_device *dev = queue->device;
1276 struct ib_device *ibdev = dev->dev;
1280 refcount_set(&req->ref, 2); /* send and recv completions */
1282 c->common.flags |= NVME_CMD_SGL_METABUF;
1284 if (!blk_rq_payload_bytes(rq))
1285 return nvme_rdma_set_sg_null(c);
1287 req->sg_table.sgl = req->first_sgl;
1288 ret = sg_alloc_table_chained(&req->sg_table,
1289 blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1293 req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1295 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1296 rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1297 if (unlikely(count <= 0)) {
1299 goto out_free_table;
1302 if (count <= dev->num_inline_segments) {
1303 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1304 queue->ctrl->use_inline_data &&
1305 blk_rq_payload_bytes(rq) <=
1306 nvme_rdma_inline_data_size(queue)) {
1307 ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1311 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1312 ret = nvme_rdma_map_sg_single(queue, req, c);
1317 ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1325 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1326 req->nents, rq_data_dir(rq) ==
1327 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1329 sg_free_table_chained(&req->sg_table, true);
1333 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1335 struct nvme_rdma_qe *qe =
1336 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1337 struct nvme_rdma_request *req =
1338 container_of(qe, struct nvme_rdma_request, sqe);
1339 struct request *rq = blk_mq_rq_from_pdu(req);
1341 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1342 nvme_rdma_wr_error(cq, wc, "SEND");
1346 if (refcount_dec_and_test(&req->ref))
1347 nvme_end_request(rq, req->status, req->result);
1350 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1351 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1352 struct ib_send_wr *first)
1354 struct ib_send_wr wr;
1357 sge->addr = qe->dma;
1358 sge->length = sizeof(struct nvme_command),
1359 sge->lkey = queue->device->pd->local_dma_lkey;
1362 wr.wr_cqe = &qe->cqe;
1364 wr.num_sge = num_sge;
1365 wr.opcode = IB_WR_SEND;
1366 wr.send_flags = IB_SEND_SIGNALED;
1373 ret = ib_post_send(queue->qp, first, NULL);
1374 if (unlikely(ret)) {
1375 dev_err(queue->ctrl->ctrl.device,
1376 "%s failed with error code %d\n", __func__, ret);
1381 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1382 struct nvme_rdma_qe *qe)
1384 struct ib_recv_wr wr;
1388 list.addr = qe->dma;
1389 list.length = sizeof(struct nvme_completion);
1390 list.lkey = queue->device->pd->local_dma_lkey;
1392 qe->cqe.done = nvme_rdma_recv_done;
1395 wr.wr_cqe = &qe->cqe;
1399 ret = ib_post_recv(queue->qp, &wr, NULL);
1400 if (unlikely(ret)) {
1401 dev_err(queue->ctrl->ctrl.device,
1402 "%s failed with error code %d\n", __func__, ret);
1407 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1409 u32 queue_idx = nvme_rdma_queue_idx(queue);
1412 return queue->ctrl->admin_tag_set.tags[queue_idx];
1413 return queue->ctrl->tag_set.tags[queue_idx - 1];
1416 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1418 if (unlikely(wc->status != IB_WC_SUCCESS))
1419 nvme_rdma_wr_error(cq, wc, "ASYNC");
1422 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1424 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1425 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1426 struct ib_device *dev = queue->device->dev;
1427 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1428 struct nvme_command *cmd = sqe->data;
1432 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1434 memset(cmd, 0, sizeof(*cmd));
1435 cmd->common.opcode = nvme_admin_async_event;
1436 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1437 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1438 nvme_rdma_set_sg_null(cmd);
1440 sqe->cqe.done = nvme_rdma_async_done;
1442 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1445 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1449 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1450 struct nvme_completion *cqe, struct ib_wc *wc)
1453 struct nvme_rdma_request *req;
1455 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1457 dev_err(queue->ctrl->ctrl.device,
1458 "tag 0x%x on QP %#x not found\n",
1459 cqe->command_id, queue->qp->qp_num);
1460 nvme_rdma_error_recovery(queue->ctrl);
1463 req = blk_mq_rq_to_pdu(rq);
1465 req->status = cqe->status;
1466 req->result = cqe->result;
1468 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1469 if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1470 dev_err(queue->ctrl->ctrl.device,
1471 "Bogus remote invalidation for rkey %#x\n",
1473 nvme_rdma_error_recovery(queue->ctrl);
1475 } else if (req->mr) {
1478 ret = nvme_rdma_inv_rkey(queue, req);
1479 if (unlikely(ret < 0)) {
1480 dev_err(queue->ctrl->ctrl.device,
1481 "Queueing INV WR for rkey %#x failed (%d)\n",
1482 req->mr->rkey, ret);
1483 nvme_rdma_error_recovery(queue->ctrl);
1485 /* the local invalidation completion will end the request */
1489 if (refcount_dec_and_test(&req->ref))
1490 nvme_end_request(rq, req->status, req->result);
1493 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1495 struct nvme_rdma_qe *qe =
1496 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1497 struct nvme_rdma_queue *queue = cq->cq_context;
1498 struct ib_device *ibdev = queue->device->dev;
1499 struct nvme_completion *cqe = qe->data;
1500 const size_t len = sizeof(struct nvme_completion);
1502 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1503 nvme_rdma_wr_error(cq, wc, "RECV");
1507 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1509 * AEN requests are special as they don't time out and can
1510 * survive any kind of queue freeze and often don't respond to
1511 * aborts. We don't even bother to allocate a struct request
1512 * for them but rather special case them here.
1514 if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1515 cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1516 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1519 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1520 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1522 nvme_rdma_post_recv(queue, qe);
1525 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1529 for (i = 0; i < queue->queue_size; i++) {
1530 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1532 goto out_destroy_queue_ib;
1537 out_destroy_queue_ib:
1538 nvme_rdma_destroy_queue_ib(queue);
1542 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1543 struct rdma_cm_event *ev)
1545 struct rdma_cm_id *cm_id = queue->cm_id;
1546 int status = ev->status;
1547 const char *rej_msg;
1548 const struct nvme_rdma_cm_rej *rej_data;
1551 rej_msg = rdma_reject_msg(cm_id, status);
1552 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1554 if (rej_data && rej_data_len >= sizeof(u16)) {
1555 u16 sts = le16_to_cpu(rej_data->sts);
1557 dev_err(queue->ctrl->ctrl.device,
1558 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1559 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1561 dev_err(queue->ctrl->ctrl.device,
1562 "Connect rejected: status %d (%s).\n", status, rej_msg);
1568 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1572 ret = nvme_rdma_create_queue_ib(queue);
1576 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1578 dev_err(queue->ctrl->ctrl.device,
1579 "rdma_resolve_route failed (%d).\n",
1581 goto out_destroy_queue;
1587 nvme_rdma_destroy_queue_ib(queue);
1591 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1593 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1594 struct rdma_conn_param param = { };
1595 struct nvme_rdma_cm_req priv = { };
1598 param.qp_num = queue->qp->qp_num;
1599 param.flow_control = 1;
1601 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1602 /* maximum retry count */
1603 param.retry_count = 7;
1604 param.rnr_retry_count = 7;
1605 param.private_data = &priv;
1606 param.private_data_len = sizeof(priv);
1608 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1609 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1611 * set the admin queue depth to the minimum size
1612 * specified by the Fabrics standard.
1614 if (priv.qid == 0) {
1615 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1616 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1619 * current interpretation of the fabrics spec
1620 * is at minimum you make hrqsize sqsize+1, or a
1621 * 1's based representation of sqsize.
1623 priv.hrqsize = cpu_to_le16(queue->queue_size);
1624 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1627 ret = rdma_connect(queue->cm_id, ¶m);
1629 dev_err(ctrl->ctrl.device,
1630 "rdma_connect failed (%d).\n", ret);
1631 goto out_destroy_queue_ib;
1636 out_destroy_queue_ib:
1637 nvme_rdma_destroy_queue_ib(queue);
1641 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1642 struct rdma_cm_event *ev)
1644 struct nvme_rdma_queue *queue = cm_id->context;
1647 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1648 rdma_event_msg(ev->event), ev->event,
1651 switch (ev->event) {
1652 case RDMA_CM_EVENT_ADDR_RESOLVED:
1653 cm_error = nvme_rdma_addr_resolved(queue);
1655 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1656 cm_error = nvme_rdma_route_resolved(queue);
1658 case RDMA_CM_EVENT_ESTABLISHED:
1659 queue->cm_error = nvme_rdma_conn_established(queue);
1660 /* complete cm_done regardless of success/failure */
1661 complete(&queue->cm_done);
1663 case RDMA_CM_EVENT_REJECTED:
1664 nvme_rdma_destroy_queue_ib(queue);
1665 cm_error = nvme_rdma_conn_rejected(queue, ev);
1667 case RDMA_CM_EVENT_ROUTE_ERROR:
1668 case RDMA_CM_EVENT_CONNECT_ERROR:
1669 case RDMA_CM_EVENT_UNREACHABLE:
1670 nvme_rdma_destroy_queue_ib(queue);
1672 case RDMA_CM_EVENT_ADDR_ERROR:
1673 dev_dbg(queue->ctrl->ctrl.device,
1674 "CM error event %d\n", ev->event);
1675 cm_error = -ECONNRESET;
1677 case RDMA_CM_EVENT_DISCONNECTED:
1678 case RDMA_CM_EVENT_ADDR_CHANGE:
1679 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1680 dev_dbg(queue->ctrl->ctrl.device,
1681 "disconnect received - connection closed\n");
1682 nvme_rdma_error_recovery(queue->ctrl);
1684 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1685 /* device removal is handled via the ib_client API */
1688 dev_err(queue->ctrl->ctrl.device,
1689 "Unexpected RDMA CM event (%d)\n", ev->event);
1690 nvme_rdma_error_recovery(queue->ctrl);
1695 queue->cm_error = cm_error;
1696 complete(&queue->cm_done);
1702 static enum blk_eh_timer_return
1703 nvme_rdma_timeout(struct request *rq, bool reserved)
1705 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1706 struct nvme_rdma_queue *queue = req->queue;
1707 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1709 dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
1710 rq->tag, nvme_rdma_queue_idx(queue));
1712 if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
1714 * Teardown immediately if controller times out while starting
1715 * or we are already started error recovery. all outstanding
1716 * requests are completed on shutdown, so we return BLK_EH_DONE.
1718 flush_work(&ctrl->err_work);
1719 nvme_rdma_teardown_io_queues(ctrl, false);
1720 nvme_rdma_teardown_admin_queue(ctrl, false);
1724 dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1725 nvme_rdma_error_recovery(ctrl);
1727 return BLK_EH_RESET_TIMER;
1730 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1731 const struct blk_mq_queue_data *bd)
1733 struct nvme_ns *ns = hctx->queue->queuedata;
1734 struct nvme_rdma_queue *queue = hctx->driver_data;
1735 struct request *rq = bd->rq;
1736 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1737 struct nvme_rdma_qe *sqe = &req->sqe;
1738 struct nvme_command *c = sqe->data;
1739 struct ib_device *dev;
1740 bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1744 WARN_ON_ONCE(rq->tag < 0);
1746 if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1747 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
1749 dev = queue->device->dev;
1750 ib_dma_sync_single_for_cpu(dev, sqe->dma,
1751 sizeof(struct nvme_command), DMA_TO_DEVICE);
1753 ret = nvme_setup_cmd(ns, rq, c);
1757 blk_mq_start_request(rq);
1759 err = nvme_rdma_map_data(queue, rq, c);
1760 if (unlikely(err < 0)) {
1761 dev_err(queue->ctrl->ctrl.device,
1762 "Failed to map data (%d)\n", err);
1763 nvme_cleanup_cmd(rq);
1767 sqe->cqe.done = nvme_rdma_send_done;
1769 ib_dma_sync_single_for_device(dev, sqe->dma,
1770 sizeof(struct nvme_command), DMA_TO_DEVICE);
1772 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1773 req->mr ? &req->reg_wr.wr : NULL);
1774 if (unlikely(err)) {
1775 nvme_rdma_unmap_data(queue, rq);
1781 if (err == -ENOMEM || err == -EAGAIN)
1782 return BLK_STS_RESOURCE;
1783 return BLK_STS_IOERR;
1786 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
1788 struct nvme_rdma_queue *queue = hctx->driver_data;
1790 return ib_process_cq_direct(queue->ib_cq, -1);
1793 static void nvme_rdma_complete_rq(struct request *rq)
1795 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1797 nvme_rdma_unmap_data(req->queue, rq);
1798 nvme_complete_rq(rq);
1801 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1803 struct nvme_rdma_ctrl *ctrl = set->driver_data;
1805 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
1806 set->map[HCTX_TYPE_DEFAULT].nr_queues =
1807 ctrl->io_queues[HCTX_TYPE_DEFAULT];
1808 set->map[HCTX_TYPE_READ].nr_queues = ctrl->io_queues[HCTX_TYPE_READ];
1809 if (ctrl->ctrl.opts->nr_write_queues) {
1810 /* separate read/write queues */
1811 set->map[HCTX_TYPE_READ].queue_offset =
1812 ctrl->io_queues[HCTX_TYPE_DEFAULT];
1814 /* mixed read/write queues */
1815 set->map[HCTX_TYPE_READ].queue_offset = 0;
1817 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
1818 ctrl->device->dev, 0);
1819 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
1820 ctrl->device->dev, 0);
1822 if (ctrl->ctrl.opts->nr_poll_queues) {
1823 set->map[HCTX_TYPE_POLL].nr_queues =
1824 ctrl->io_queues[HCTX_TYPE_POLL];
1825 set->map[HCTX_TYPE_POLL].queue_offset =
1826 ctrl->io_queues[HCTX_TYPE_DEFAULT];
1827 if (ctrl->ctrl.opts->nr_write_queues)
1828 set->map[HCTX_TYPE_POLL].queue_offset +=
1829 ctrl->io_queues[HCTX_TYPE_READ];
1830 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
1835 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1836 .queue_rq = nvme_rdma_queue_rq,
1837 .complete = nvme_rdma_complete_rq,
1838 .init_request = nvme_rdma_init_request,
1839 .exit_request = nvme_rdma_exit_request,
1840 .init_hctx = nvme_rdma_init_hctx,
1841 .timeout = nvme_rdma_timeout,
1842 .map_queues = nvme_rdma_map_queues,
1843 .poll = nvme_rdma_poll,
1846 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1847 .queue_rq = nvme_rdma_queue_rq,
1848 .complete = nvme_rdma_complete_rq,
1849 .init_request = nvme_rdma_init_request,
1850 .exit_request = nvme_rdma_exit_request,
1851 .init_hctx = nvme_rdma_init_admin_hctx,
1852 .timeout = nvme_rdma_timeout,
1855 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1857 nvme_rdma_teardown_io_queues(ctrl, shutdown);
1859 nvme_shutdown_ctrl(&ctrl->ctrl);
1861 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1862 nvme_rdma_teardown_admin_queue(ctrl, shutdown);
1865 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1867 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1870 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1872 struct nvme_rdma_ctrl *ctrl =
1873 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1875 nvme_stop_ctrl(&ctrl->ctrl);
1876 nvme_rdma_shutdown_ctrl(ctrl, false);
1878 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1879 /* state change failure should never happen */
1884 if (nvme_rdma_setup_ctrl(ctrl, false))
1890 ++ctrl->ctrl.nr_reconnects;
1891 nvme_rdma_reconnect_or_remove(ctrl);
1894 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1896 .module = THIS_MODULE,
1897 .flags = NVME_F_FABRICS,
1898 .reg_read32 = nvmf_reg_read32,
1899 .reg_read64 = nvmf_reg_read64,
1900 .reg_write32 = nvmf_reg_write32,
1901 .free_ctrl = nvme_rdma_free_ctrl,
1902 .submit_async_event = nvme_rdma_submit_async_event,
1903 .delete_ctrl = nvme_rdma_delete_ctrl,
1904 .get_address = nvmf_get_address,
1905 .stop_ctrl = nvme_rdma_stop_ctrl,
1909 * Fails a connection request if it matches an existing controller
1910 * (association) with the same tuple:
1911 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1913 * if local address is not specified in the request, it will match an
1914 * existing controller with all the other parameters the same and no
1915 * local port address specified as well.
1917 * The ports don't need to be compared as they are intrinsically
1918 * already matched by the port pointers supplied.
1921 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1923 struct nvme_rdma_ctrl *ctrl;
1926 mutex_lock(&nvme_rdma_ctrl_mutex);
1927 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1928 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
1932 mutex_unlock(&nvme_rdma_ctrl_mutex);
1937 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1938 struct nvmf_ctrl_options *opts)
1940 struct nvme_rdma_ctrl *ctrl;
1944 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1946 return ERR_PTR(-ENOMEM);
1947 ctrl->ctrl.opts = opts;
1948 INIT_LIST_HEAD(&ctrl->list);
1950 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
1952 kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
1953 if (!opts->trsvcid) {
1957 opts->mask |= NVMF_OPT_TRSVCID;
1960 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1961 opts->traddr, opts->trsvcid, &ctrl->addr);
1963 pr_err("malformed address passed: %s:%s\n",
1964 opts->traddr, opts->trsvcid);
1968 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1969 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1970 opts->host_traddr, NULL, &ctrl->src_addr);
1972 pr_err("malformed src address passed: %s\n",
1978 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1983 INIT_DELAYED_WORK(&ctrl->reconnect_work,
1984 nvme_rdma_reconnect_ctrl_work);
1985 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1986 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1988 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
1989 opts->nr_poll_queues + 1;
1990 ctrl->ctrl.sqsize = opts->queue_size - 1;
1991 ctrl->ctrl.kato = opts->kato;
1994 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1999 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2000 0 /* no quirks, we're perfect! */);
2002 goto out_kfree_queues;
2004 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2005 WARN_ON_ONCE(!changed);
2007 ret = nvme_rdma_setup_ctrl(ctrl, true);
2009 goto out_uninit_ctrl;
2011 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2012 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2014 nvme_get_ctrl(&ctrl->ctrl);
2016 mutex_lock(&nvme_rdma_ctrl_mutex);
2017 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2018 mutex_unlock(&nvme_rdma_ctrl_mutex);
2023 nvme_uninit_ctrl(&ctrl->ctrl);
2024 nvme_put_ctrl(&ctrl->ctrl);
2027 return ERR_PTR(ret);
2029 kfree(ctrl->queues);
2032 return ERR_PTR(ret);
2035 static struct nvmf_transport_ops nvme_rdma_transport = {
2037 .module = THIS_MODULE,
2038 .required_opts = NVMF_OPT_TRADDR,
2039 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2040 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2041 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES,
2042 .create_ctrl = nvme_rdma_create_ctrl,
2045 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2047 struct nvme_rdma_ctrl *ctrl;
2048 struct nvme_rdma_device *ndev;
2051 mutex_lock(&device_list_mutex);
2052 list_for_each_entry(ndev, &device_list, entry) {
2053 if (ndev->dev == ib_device) {
2058 mutex_unlock(&device_list_mutex);
2063 /* Delete all controllers using this device */
2064 mutex_lock(&nvme_rdma_ctrl_mutex);
2065 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2066 if (ctrl->device->dev != ib_device)
2068 nvme_delete_ctrl(&ctrl->ctrl);
2070 mutex_unlock(&nvme_rdma_ctrl_mutex);
2072 flush_workqueue(nvme_delete_wq);
2075 static struct ib_client nvme_rdma_ib_client = {
2076 .name = "nvme_rdma",
2077 .remove = nvme_rdma_remove_one
2080 static int __init nvme_rdma_init_module(void)
2084 ret = ib_register_client(&nvme_rdma_ib_client);
2088 ret = nvmf_register_transport(&nvme_rdma_transport);
2090 goto err_unreg_client;
2095 ib_unregister_client(&nvme_rdma_ib_client);
2099 static void __exit nvme_rdma_cleanup_module(void)
2101 nvmf_unregister_transport(&nvme_rdma_transport);
2102 ib_unregister_client(&nvme_rdma_ib_client);
2105 module_init(nvme_rdma_init_module);
2106 module_exit(nvme_rdma_cleanup_module);
2108 MODULE_LICENSE("GPL v2");