1 // SPDX-License-Identifier: GPL-2.0
3 * NVMe over Fabrics RDMA host code.
4 * Copyright (c) 2015-2016 HGST, a Western Digital Company.
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/module.h>
8 #include <linux/init.h>
9 #include <linux/slab.h>
10 #include <rdma/mr_pool.h>
11 #include <linux/err.h>
12 #include <linux/string.h>
13 #include <linux/atomic.h>
14 #include <linux/blk-mq.h>
15 #include <linux/blk-mq-rdma.h>
16 #include <linux/types.h>
17 #include <linux/list.h>
18 #include <linux/mutex.h>
19 #include <linux/scatterlist.h>
20 #include <linux/nvme.h>
21 #include <asm/unaligned.h>
23 #include <rdma/ib_verbs.h>
24 #include <rdma/rdma_cm.h>
25 #include <linux/nvme-rdma.h>
31 #define NVME_RDMA_CONNECT_TIMEOUT_MS 3000 /* 3 second */
33 #define NVME_RDMA_MAX_SEGMENTS 256
35 #define NVME_RDMA_MAX_INLINE_SEGMENTS 4
37 struct nvme_rdma_device {
38 struct ib_device *dev;
41 struct list_head entry;
42 unsigned int num_inline_segments;
51 struct nvme_rdma_queue;
52 struct nvme_rdma_request {
53 struct nvme_request req;
55 struct nvme_rdma_qe sqe;
56 union nvme_result result;
59 struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
62 struct ib_reg_wr reg_wr;
63 struct ib_cqe reg_cqe;
64 struct nvme_rdma_queue *queue;
65 struct sg_table sg_table;
66 struct scatterlist first_sgl[];
69 enum nvme_rdma_queue_flags {
70 NVME_RDMA_Q_ALLOCATED = 0,
72 NVME_RDMA_Q_TR_READY = 2,
75 struct nvme_rdma_queue {
76 struct nvme_rdma_qe *rsp_ring;
78 size_t cmnd_capsule_len;
79 struct nvme_rdma_ctrl *ctrl;
80 struct nvme_rdma_device *device;
85 struct rdma_cm_id *cm_id;
87 struct completion cm_done;
90 struct nvme_rdma_ctrl {
91 /* read only in the hot path */
92 struct nvme_rdma_queue *queues;
94 /* other member variables */
95 struct blk_mq_tag_set tag_set;
96 struct work_struct err_work;
98 struct nvme_rdma_qe async_event_sqe;
100 struct delayed_work reconnect_work;
102 struct list_head list;
104 struct blk_mq_tag_set admin_tag_set;
105 struct nvme_rdma_device *device;
109 struct sockaddr_storage addr;
110 struct sockaddr_storage src_addr;
112 struct nvme_ctrl ctrl;
113 bool use_inline_data;
114 u32 io_queues[HCTX_MAX_TYPES];
117 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
119 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
122 static LIST_HEAD(device_list);
123 static DEFINE_MUTEX(device_list_mutex);
125 static LIST_HEAD(nvme_rdma_ctrl_list);
126 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
129 * Disabling this option makes small I/O goes faster, but is fundamentally
130 * unsafe. With it turned off we will have to register a global rkey that
131 * allows read and write access to all physical memory.
133 static bool register_always = true;
134 module_param(register_always, bool, 0444);
135 MODULE_PARM_DESC(register_always,
136 "Use memory registration even for contiguous memory regions");
138 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
139 struct rdma_cm_event *event);
140 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
142 static const struct blk_mq_ops nvme_rdma_mq_ops;
143 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
145 /* XXX: really should move to a generic header sooner or later.. */
146 static inline void put_unaligned_le24(u32 val, u8 *p)
153 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
155 return queue - queue->ctrl->queues;
158 static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue)
160 return nvme_rdma_queue_idx(queue) >
161 queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] +
162 queue->ctrl->io_queues[HCTX_TYPE_READ];
165 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
167 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
170 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
171 size_t capsule_size, enum dma_data_direction dir)
173 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
177 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
178 size_t capsule_size, enum dma_data_direction dir)
180 qe->data = kzalloc(capsule_size, GFP_KERNEL);
184 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
185 if (ib_dma_mapping_error(ibdev, qe->dma)) {
194 static void nvme_rdma_free_ring(struct ib_device *ibdev,
195 struct nvme_rdma_qe *ring, size_t ib_queue_size,
196 size_t capsule_size, enum dma_data_direction dir)
200 for (i = 0; i < ib_queue_size; i++)
201 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
205 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
206 size_t ib_queue_size, size_t capsule_size,
207 enum dma_data_direction dir)
209 struct nvme_rdma_qe *ring;
212 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
217 * Bind the CQEs (post recv buffers) DMA mapping to the RDMA queue
218 * lifetime. It's safe, since any chage in the underlying RDMA device
219 * will issue error recovery and queue re-creation.
221 for (i = 0; i < ib_queue_size; i++) {
222 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
229 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
233 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
235 pr_debug("QP event %s (%d)\n",
236 ib_event_msg(event->event), event->event);
240 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
244 ret = wait_for_completion_interruptible_timeout(&queue->cm_done,
245 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
250 WARN_ON_ONCE(queue->cm_error > 0);
251 return queue->cm_error;
254 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
256 struct nvme_rdma_device *dev = queue->device;
257 struct ib_qp_init_attr init_attr;
260 memset(&init_attr, 0, sizeof(init_attr));
261 init_attr.event_handler = nvme_rdma_qp_event;
263 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
265 init_attr.cap.max_recv_wr = queue->queue_size + 1;
266 init_attr.cap.max_recv_sge = 1;
267 init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
268 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
269 init_attr.qp_type = IB_QPT_RC;
270 init_attr.send_cq = queue->ib_cq;
271 init_attr.recv_cq = queue->ib_cq;
273 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
275 queue->qp = queue->cm_id->qp;
279 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
280 struct request *rq, unsigned int hctx_idx)
282 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
284 kfree(req->sqe.data);
287 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
288 struct request *rq, unsigned int hctx_idx,
289 unsigned int numa_node)
291 struct nvme_rdma_ctrl *ctrl = set->driver_data;
292 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
293 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
294 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
296 nvme_req(rq)->ctrl = &ctrl->ctrl;
297 req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL);
306 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
307 unsigned int hctx_idx)
309 struct nvme_rdma_ctrl *ctrl = data;
310 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
312 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
314 hctx->driver_data = queue;
318 static int nvme_rdma_init_admin_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[0];
324 BUG_ON(hctx_idx != 0);
326 hctx->driver_data = queue;
330 static void nvme_rdma_free_dev(struct kref *ref)
332 struct nvme_rdma_device *ndev =
333 container_of(ref, struct nvme_rdma_device, ref);
335 mutex_lock(&device_list_mutex);
336 list_del(&ndev->entry);
337 mutex_unlock(&device_list_mutex);
339 ib_dealloc_pd(ndev->pd);
343 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
345 kref_put(&dev->ref, nvme_rdma_free_dev);
348 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
350 return kref_get_unless_zero(&dev->ref);
353 static struct nvme_rdma_device *
354 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
356 struct nvme_rdma_device *ndev;
358 mutex_lock(&device_list_mutex);
359 list_for_each_entry(ndev, &device_list, entry) {
360 if (ndev->dev->node_guid == cm_id->device->node_guid &&
361 nvme_rdma_dev_get(ndev))
365 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
369 ndev->dev = cm_id->device;
370 kref_init(&ndev->ref);
372 ndev->pd = ib_alloc_pd(ndev->dev,
373 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
374 if (IS_ERR(ndev->pd))
377 if (!(ndev->dev->attrs.device_cap_flags &
378 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
379 dev_err(&ndev->dev->dev,
380 "Memory registrations not supported.\n");
384 ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
385 ndev->dev->attrs.max_send_sge - 1);
386 list_add(&ndev->entry, &device_list);
388 mutex_unlock(&device_list_mutex);
392 ib_dealloc_pd(ndev->pd);
396 mutex_unlock(&device_list_mutex);
400 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
402 struct nvme_rdma_device *dev;
403 struct ib_device *ibdev;
405 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
411 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
414 * The cm_id object might have been destroyed during RDMA connection
415 * establishment error flow to avoid getting other cma events, thus
416 * the destruction of the QP shouldn't use rdma_cm API.
418 ib_destroy_qp(queue->qp);
419 ib_free_cq(queue->ib_cq);
421 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
422 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
424 nvme_rdma_dev_put(dev);
427 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev)
429 return min_t(u32, NVME_RDMA_MAX_SEGMENTS,
430 ibdev->attrs.max_fast_reg_page_list_len - 1);
433 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
435 struct ib_device *ibdev;
436 const int send_wr_factor = 3; /* MR, SEND, INV */
437 const int cq_factor = send_wr_factor + 1; /* + RECV */
438 int comp_vector, idx = nvme_rdma_queue_idx(queue);
439 enum ib_poll_context poll_ctx;
440 int ret, pages_per_mr;
442 queue->device = nvme_rdma_find_get_device(queue->cm_id);
443 if (!queue->device) {
444 dev_err(queue->cm_id->device->dev.parent,
445 "no client data found!\n");
446 return -ECONNREFUSED;
448 ibdev = queue->device->dev;
451 * Spread I/O queues completion vectors according their queue index.
452 * Admin queues can always go on completion vector 0.
454 comp_vector = idx == 0 ? idx : idx - 1;
456 /* Polling queues need direct cq polling context */
457 if (nvme_rdma_poll_queue(queue))
458 poll_ctx = IB_POLL_DIRECT;
460 poll_ctx = IB_POLL_SOFTIRQ;
462 /* +1 for ib_stop_cq */
463 queue->ib_cq = ib_alloc_cq(ibdev, queue,
464 cq_factor * queue->queue_size + 1,
465 comp_vector, poll_ctx);
466 if (IS_ERR(queue->ib_cq)) {
467 ret = PTR_ERR(queue->ib_cq);
471 ret = nvme_rdma_create_qp(queue, send_wr_factor);
473 goto out_destroy_ib_cq;
475 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
476 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
477 if (!queue->rsp_ring) {
483 * Currently we don't use SG_GAPS MR's so if the first entry is
484 * misaligned we'll end up using two entries for a single data page,
485 * so one additional entry is required.
487 pages_per_mr = nvme_rdma_get_max_fr_pages(ibdev) + 1;
488 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
493 dev_err(queue->ctrl->ctrl.device,
494 "failed to initialize MR pool sized %d for QID %d\n",
495 queue->queue_size, idx);
496 goto out_destroy_ring;
499 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
504 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
505 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
507 rdma_destroy_qp(queue->cm_id);
509 ib_free_cq(queue->ib_cq);
511 nvme_rdma_dev_put(queue->device);
515 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
516 int idx, size_t queue_size)
518 struct nvme_rdma_queue *queue;
519 struct sockaddr *src_addr = NULL;
522 queue = &ctrl->queues[idx];
524 init_completion(&queue->cm_done);
527 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
529 queue->cmnd_capsule_len = sizeof(struct nvme_command);
531 queue->queue_size = queue_size;
533 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
534 RDMA_PS_TCP, IB_QPT_RC);
535 if (IS_ERR(queue->cm_id)) {
536 dev_info(ctrl->ctrl.device,
537 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
538 return PTR_ERR(queue->cm_id);
541 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
542 src_addr = (struct sockaddr *)&ctrl->src_addr;
544 queue->cm_error = -ETIMEDOUT;
545 ret = rdma_resolve_addr(queue->cm_id, src_addr,
546 (struct sockaddr *)&ctrl->addr,
547 NVME_RDMA_CONNECT_TIMEOUT_MS);
549 dev_info(ctrl->ctrl.device,
550 "rdma_resolve_addr failed (%d).\n", ret);
551 goto out_destroy_cm_id;
554 ret = nvme_rdma_wait_for_cm(queue);
556 dev_info(ctrl->ctrl.device,
557 "rdma connection establishment failed (%d)\n", ret);
558 goto out_destroy_cm_id;
561 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
566 rdma_destroy_id(queue->cm_id);
567 nvme_rdma_destroy_queue_ib(queue);
571 static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
573 rdma_disconnect(queue->cm_id);
574 ib_drain_qp(queue->qp);
577 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
579 if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
581 __nvme_rdma_stop_queue(queue);
584 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
586 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
589 nvme_rdma_destroy_queue_ib(queue);
590 rdma_destroy_id(queue->cm_id);
593 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
597 for (i = 1; i < ctrl->ctrl.queue_count; i++)
598 nvme_rdma_free_queue(&ctrl->queues[i]);
601 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
605 for (i = 1; i < ctrl->ctrl.queue_count; i++)
606 nvme_rdma_stop_queue(&ctrl->queues[i]);
609 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
611 struct nvme_rdma_queue *queue = &ctrl->queues[idx];
612 bool poll = nvme_rdma_poll_queue(queue);
616 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx, poll);
618 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
621 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
623 if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
624 __nvme_rdma_stop_queue(queue);
625 dev_info(ctrl->ctrl.device,
626 "failed to connect queue: %d ret=%d\n", idx, ret);
631 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
635 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
636 ret = nvme_rdma_start_queue(ctrl, i);
638 goto out_stop_queues;
644 for (i--; i >= 1; i--)
645 nvme_rdma_stop_queue(&ctrl->queues[i]);
649 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
651 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
652 struct ib_device *ibdev = ctrl->device->dev;
653 unsigned int nr_io_queues, nr_default_queues;
654 unsigned int nr_read_queues, nr_poll_queues;
657 nr_read_queues = min_t(unsigned int, ibdev->num_comp_vectors,
658 min(opts->nr_io_queues, num_online_cpus()));
659 nr_default_queues = min_t(unsigned int, ibdev->num_comp_vectors,
660 min(opts->nr_write_queues, num_online_cpus()));
661 nr_poll_queues = min(opts->nr_poll_queues, num_online_cpus());
662 nr_io_queues = nr_read_queues + nr_default_queues + nr_poll_queues;
664 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
668 ctrl->ctrl.queue_count = nr_io_queues + 1;
669 if (ctrl->ctrl.queue_count < 2)
672 dev_info(ctrl->ctrl.device,
673 "creating %d I/O queues.\n", nr_io_queues);
675 if (opts->nr_write_queues && nr_read_queues < nr_io_queues) {
677 * separate read/write queues
678 * hand out dedicated default queues only after we have
679 * sufficient read queues.
681 ctrl->io_queues[HCTX_TYPE_READ] = nr_read_queues;
682 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
683 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
684 min(nr_default_queues, nr_io_queues);
685 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
688 * shared read/write queues
689 * either no write queues were requested, or we don't have
690 * sufficient queue count to have dedicated default queues.
692 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
693 min(nr_read_queues, nr_io_queues);
694 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
697 if (opts->nr_poll_queues && nr_io_queues) {
698 /* map dedicated poll queues only if we have queues left */
699 ctrl->io_queues[HCTX_TYPE_POLL] =
700 min(nr_poll_queues, nr_io_queues);
703 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
704 ret = nvme_rdma_alloc_queue(ctrl, i,
705 ctrl->ctrl.sqsize + 1);
707 goto out_free_queues;
713 for (i--; i >= 1; i--)
714 nvme_rdma_free_queue(&ctrl->queues[i]);
719 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
722 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
723 struct blk_mq_tag_set *set;
727 set = &ctrl->admin_tag_set;
728 memset(set, 0, sizeof(*set));
729 set->ops = &nvme_rdma_admin_mq_ops;
730 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
731 set->reserved_tags = 2; /* connect + keep-alive */
732 set->numa_node = nctrl->numa_node;
733 set->cmd_size = sizeof(struct nvme_rdma_request) +
734 NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
735 set->driver_data = ctrl;
736 set->nr_hw_queues = 1;
737 set->timeout = ADMIN_TIMEOUT;
738 set->flags = BLK_MQ_F_NO_SCHED;
740 set = &ctrl->tag_set;
741 memset(set, 0, sizeof(*set));
742 set->ops = &nvme_rdma_mq_ops;
743 set->queue_depth = nctrl->sqsize + 1;
744 set->reserved_tags = 1; /* fabric connect */
745 set->numa_node = nctrl->numa_node;
746 set->flags = BLK_MQ_F_SHOULD_MERGE;
747 set->cmd_size = sizeof(struct nvme_rdma_request) +
748 NVME_INLINE_SG_CNT * sizeof(struct scatterlist);
749 set->driver_data = ctrl;
750 set->nr_hw_queues = nctrl->queue_count - 1;
751 set->timeout = NVME_IO_TIMEOUT;
752 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
755 ret = blk_mq_alloc_tag_set(set);
762 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
766 blk_cleanup_queue(ctrl->ctrl.admin_q);
767 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
768 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
770 if (ctrl->async_event_sqe.data) {
771 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
772 sizeof(struct nvme_command), DMA_TO_DEVICE);
773 ctrl->async_event_sqe.data = NULL;
775 nvme_rdma_free_queue(&ctrl->queues[0]);
778 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
783 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
787 ctrl->device = ctrl->queues[0].device;
788 ctrl->ctrl.numa_node = dev_to_node(ctrl->device->dev->dma_device);
790 ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
793 * Bind the async event SQE DMA mapping to the admin queue lifetime.
794 * It's safe, since any chage in the underlying RDMA device will issue
795 * error recovery and queue re-creation.
797 error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
798 sizeof(struct nvme_command), DMA_TO_DEVICE);
803 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
804 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
805 error = PTR_ERR(ctrl->ctrl.admin_tagset);
806 goto out_free_async_qe;
809 ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
810 if (IS_ERR(ctrl->ctrl.fabrics_q)) {
811 error = PTR_ERR(ctrl->ctrl.fabrics_q);
812 goto out_free_tagset;
815 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
816 if (IS_ERR(ctrl->ctrl.admin_q)) {
817 error = PTR_ERR(ctrl->ctrl.admin_q);
818 goto out_cleanup_fabrics_q;
822 error = nvme_rdma_start_queue(ctrl, 0);
824 goto out_cleanup_queue;
826 error = nvme_enable_ctrl(&ctrl->ctrl);
830 ctrl->ctrl.max_segments = ctrl->max_fr_pages;
831 ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9);
833 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
835 error = nvme_init_identify(&ctrl->ctrl);
842 nvme_rdma_stop_queue(&ctrl->queues[0]);
845 blk_cleanup_queue(ctrl->ctrl.admin_q);
846 out_cleanup_fabrics_q:
848 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
851 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
853 if (ctrl->async_event_sqe.data) {
854 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
855 sizeof(struct nvme_command), DMA_TO_DEVICE);
856 ctrl->async_event_sqe.data = NULL;
859 nvme_rdma_free_queue(&ctrl->queues[0]);
863 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
867 blk_cleanup_queue(ctrl->ctrl.connect_q);
868 blk_mq_free_tag_set(ctrl->ctrl.tagset);
870 nvme_rdma_free_io_queues(ctrl);
873 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
877 ret = nvme_rdma_alloc_io_queues(ctrl);
882 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
883 if (IS_ERR(ctrl->ctrl.tagset)) {
884 ret = PTR_ERR(ctrl->ctrl.tagset);
885 goto out_free_io_queues;
888 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
889 if (IS_ERR(ctrl->ctrl.connect_q)) {
890 ret = PTR_ERR(ctrl->ctrl.connect_q);
891 goto out_free_tag_set;
894 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
895 ctrl->ctrl.queue_count - 1);
898 ret = nvme_rdma_start_io_queues(ctrl);
900 goto out_cleanup_connect_q;
904 out_cleanup_connect_q:
906 blk_cleanup_queue(ctrl->ctrl.connect_q);
909 blk_mq_free_tag_set(ctrl->ctrl.tagset);
911 nvme_rdma_free_io_queues(ctrl);
915 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
918 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
919 nvme_rdma_stop_queue(&ctrl->queues[0]);
920 if (ctrl->ctrl.admin_tagset) {
921 blk_mq_tagset_busy_iter(ctrl->ctrl.admin_tagset,
922 nvme_cancel_request, &ctrl->ctrl);
923 blk_mq_tagset_wait_completed_request(ctrl->ctrl.admin_tagset);
926 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
927 nvme_rdma_destroy_admin_queue(ctrl, remove);
930 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
933 if (ctrl->ctrl.queue_count > 1) {
934 nvme_stop_queues(&ctrl->ctrl);
935 nvme_rdma_stop_io_queues(ctrl);
936 if (ctrl->ctrl.tagset) {
937 blk_mq_tagset_busy_iter(ctrl->ctrl.tagset,
938 nvme_cancel_request, &ctrl->ctrl);
939 blk_mq_tagset_wait_completed_request(ctrl->ctrl.tagset);
942 nvme_start_queues(&ctrl->ctrl);
943 nvme_rdma_destroy_io_queues(ctrl, remove);
947 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
949 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
951 if (list_empty(&ctrl->list))
954 mutex_lock(&nvme_rdma_ctrl_mutex);
955 list_del(&ctrl->list);
956 mutex_unlock(&nvme_rdma_ctrl_mutex);
958 nvmf_free_options(nctrl->opts);
964 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
966 /* If we are resetting/deleting then do nothing */
967 if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
968 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
969 ctrl->ctrl.state == NVME_CTRL_LIVE);
973 if (nvmf_should_reconnect(&ctrl->ctrl)) {
974 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
975 ctrl->ctrl.opts->reconnect_delay);
976 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
977 ctrl->ctrl.opts->reconnect_delay * HZ);
979 nvme_delete_ctrl(&ctrl->ctrl);
983 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
988 ret = nvme_rdma_configure_admin_queue(ctrl, new);
992 if (ctrl->ctrl.icdoff) {
993 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
997 if (!(ctrl->ctrl.sgls & (1 << 2))) {
998 dev_err(ctrl->ctrl.device,
999 "Mandatory keyed sgls are not supported!\n");
1003 if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1004 dev_warn(ctrl->ctrl.device,
1005 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1006 ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1009 if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1010 dev_warn(ctrl->ctrl.device,
1011 "sqsize %u > ctrl maxcmd %u, clamping down\n",
1012 ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1013 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1016 if (ctrl->ctrl.sgls & (1 << 20))
1017 ctrl->use_inline_data = true;
1019 if (ctrl->ctrl.queue_count > 1) {
1020 ret = nvme_rdma_configure_io_queues(ctrl, new);
1025 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1027 /* state change failure is ok if we're in DELETING state */
1028 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1033 nvme_start_ctrl(&ctrl->ctrl);
1037 if (ctrl->ctrl.queue_count > 1)
1038 nvme_rdma_destroy_io_queues(ctrl, new);
1040 nvme_rdma_stop_queue(&ctrl->queues[0]);
1041 nvme_rdma_destroy_admin_queue(ctrl, new);
1045 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1047 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1048 struct nvme_rdma_ctrl, reconnect_work);
1050 ++ctrl->ctrl.nr_reconnects;
1052 if (nvme_rdma_setup_ctrl(ctrl, false))
1055 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1056 ctrl->ctrl.nr_reconnects);
1058 ctrl->ctrl.nr_reconnects = 0;
1063 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1064 ctrl->ctrl.nr_reconnects);
1065 nvme_rdma_reconnect_or_remove(ctrl);
1068 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1070 struct nvme_rdma_ctrl *ctrl = container_of(work,
1071 struct nvme_rdma_ctrl, err_work);
1073 nvme_stop_keep_alive(&ctrl->ctrl);
1074 nvme_rdma_teardown_io_queues(ctrl, false);
1075 nvme_start_queues(&ctrl->ctrl);
1076 nvme_rdma_teardown_admin_queue(ctrl, false);
1077 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1079 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1080 /* state change failure is ok if we're in DELETING state */
1081 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1085 nvme_rdma_reconnect_or_remove(ctrl);
1088 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1090 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1093 queue_work(nvme_reset_wq, &ctrl->err_work);
1096 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1099 struct nvme_rdma_queue *queue = cq->cq_context;
1100 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1102 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1103 dev_info(ctrl->ctrl.device,
1104 "%s for CQE 0x%p failed with status %s (%d)\n",
1106 ib_wc_status_msg(wc->status), wc->status);
1107 nvme_rdma_error_recovery(ctrl);
1110 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1112 if (unlikely(wc->status != IB_WC_SUCCESS))
1113 nvme_rdma_wr_error(cq, wc, "MEMREG");
1116 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1118 struct nvme_rdma_request *req =
1119 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1120 struct request *rq = blk_mq_rq_from_pdu(req);
1122 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1123 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1127 if (refcount_dec_and_test(&req->ref))
1128 nvme_end_request(rq, req->status, req->result);
1132 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1133 struct nvme_rdma_request *req)
1135 struct ib_send_wr wr = {
1136 .opcode = IB_WR_LOCAL_INV,
1139 .send_flags = IB_SEND_SIGNALED,
1140 .ex.invalidate_rkey = req->mr->rkey,
1143 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1144 wr.wr_cqe = &req->reg_cqe;
1146 return ib_post_send(queue->qp, &wr, NULL);
1149 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1152 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1153 struct nvme_rdma_device *dev = queue->device;
1154 struct ib_device *ibdev = dev->dev;
1156 if (!blk_rq_nr_phys_segments(rq))
1160 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1164 ib_dma_unmap_sg(ibdev, req->sg_table.sgl, req->nents, rq_dma_dir(rq));
1165 sg_free_table_chained(&req->sg_table, NVME_INLINE_SG_CNT);
1168 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1170 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1173 put_unaligned_le24(0, sg->length);
1174 put_unaligned_le32(0, sg->key);
1175 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1179 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1180 struct nvme_rdma_request *req, struct nvme_command *c,
1183 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1184 struct scatterlist *sgl = req->sg_table.sgl;
1185 struct ib_sge *sge = &req->sge[1];
1189 for (i = 0; i < count; i++, sgl++, sge++) {
1190 sge->addr = sg_dma_address(sgl);
1191 sge->length = sg_dma_len(sgl);
1192 sge->lkey = queue->device->pd->local_dma_lkey;
1196 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1197 sg->length = cpu_to_le32(len);
1198 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1200 req->num_sge += count;
1204 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1205 struct nvme_rdma_request *req, struct nvme_command *c)
1207 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1209 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1210 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1211 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1212 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1216 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1217 struct nvme_rdma_request *req, struct nvme_command *c,
1220 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1223 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1224 if (WARN_ON_ONCE(!req->mr))
1228 * Align the MR to a 4K page size to match the ctrl page size and
1229 * the block virtual boundary.
1231 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1232 if (unlikely(nr < count)) {
1233 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1240 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1242 req->reg_cqe.done = nvme_rdma_memreg_done;
1243 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1244 req->reg_wr.wr.opcode = IB_WR_REG_MR;
1245 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1246 req->reg_wr.wr.num_sge = 0;
1247 req->reg_wr.mr = req->mr;
1248 req->reg_wr.key = req->mr->rkey;
1249 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1250 IB_ACCESS_REMOTE_READ |
1251 IB_ACCESS_REMOTE_WRITE;
1253 sg->addr = cpu_to_le64(req->mr->iova);
1254 put_unaligned_le24(req->mr->length, sg->length);
1255 put_unaligned_le32(req->mr->rkey, sg->key);
1256 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1257 NVME_SGL_FMT_INVALIDATE;
1262 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1263 struct request *rq, struct nvme_command *c)
1265 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1266 struct nvme_rdma_device *dev = queue->device;
1267 struct ib_device *ibdev = dev->dev;
1271 refcount_set(&req->ref, 2); /* send and recv completions */
1273 c->common.flags |= NVME_CMD_SGL_METABUF;
1275 if (!blk_rq_nr_phys_segments(rq))
1276 return nvme_rdma_set_sg_null(c);
1278 req->sg_table.sgl = req->first_sgl;
1279 ret = sg_alloc_table_chained(&req->sg_table,
1280 blk_rq_nr_phys_segments(rq), req->sg_table.sgl,
1281 NVME_INLINE_SG_CNT);
1285 req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1287 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1289 if (unlikely(count <= 0)) {
1291 goto out_free_table;
1294 if (count <= dev->num_inline_segments) {
1295 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1296 queue->ctrl->use_inline_data &&
1297 blk_rq_payload_bytes(rq) <=
1298 nvme_rdma_inline_data_size(queue)) {
1299 ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1303 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1304 ret = nvme_rdma_map_sg_single(queue, req, c);
1309 ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1317 ib_dma_unmap_sg(ibdev, req->sg_table.sgl, req->nents, rq_dma_dir(rq));
1319 sg_free_table_chained(&req->sg_table, NVME_INLINE_SG_CNT);
1323 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1325 struct nvme_rdma_qe *qe =
1326 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1327 struct nvme_rdma_request *req =
1328 container_of(qe, struct nvme_rdma_request, sqe);
1329 struct request *rq = blk_mq_rq_from_pdu(req);
1331 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1332 nvme_rdma_wr_error(cq, wc, "SEND");
1336 if (refcount_dec_and_test(&req->ref))
1337 nvme_end_request(rq, req->status, req->result);
1340 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1341 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1342 struct ib_send_wr *first)
1344 struct ib_send_wr wr;
1347 sge->addr = qe->dma;
1348 sge->length = sizeof(struct nvme_command),
1349 sge->lkey = queue->device->pd->local_dma_lkey;
1352 wr.wr_cqe = &qe->cqe;
1354 wr.num_sge = num_sge;
1355 wr.opcode = IB_WR_SEND;
1356 wr.send_flags = IB_SEND_SIGNALED;
1363 ret = ib_post_send(queue->qp, first, NULL);
1364 if (unlikely(ret)) {
1365 dev_err(queue->ctrl->ctrl.device,
1366 "%s failed with error code %d\n", __func__, ret);
1371 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1372 struct nvme_rdma_qe *qe)
1374 struct ib_recv_wr wr;
1378 list.addr = qe->dma;
1379 list.length = sizeof(struct nvme_completion);
1380 list.lkey = queue->device->pd->local_dma_lkey;
1382 qe->cqe.done = nvme_rdma_recv_done;
1385 wr.wr_cqe = &qe->cqe;
1389 ret = ib_post_recv(queue->qp, &wr, NULL);
1390 if (unlikely(ret)) {
1391 dev_err(queue->ctrl->ctrl.device,
1392 "%s failed with error code %d\n", __func__, ret);
1397 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1399 u32 queue_idx = nvme_rdma_queue_idx(queue);
1402 return queue->ctrl->admin_tag_set.tags[queue_idx];
1403 return queue->ctrl->tag_set.tags[queue_idx - 1];
1406 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1408 if (unlikely(wc->status != IB_WC_SUCCESS))
1409 nvme_rdma_wr_error(cq, wc, "ASYNC");
1412 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1414 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1415 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1416 struct ib_device *dev = queue->device->dev;
1417 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1418 struct nvme_command *cmd = sqe->data;
1422 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1424 memset(cmd, 0, sizeof(*cmd));
1425 cmd->common.opcode = nvme_admin_async_event;
1426 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1427 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1428 nvme_rdma_set_sg_null(cmd);
1430 sqe->cqe.done = nvme_rdma_async_done;
1432 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1435 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1439 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1440 struct nvme_completion *cqe, struct ib_wc *wc)
1443 struct nvme_rdma_request *req;
1445 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1447 dev_err(queue->ctrl->ctrl.device,
1448 "tag 0x%x on QP %#x not found\n",
1449 cqe->command_id, queue->qp->qp_num);
1450 nvme_rdma_error_recovery(queue->ctrl);
1453 req = blk_mq_rq_to_pdu(rq);
1455 req->status = cqe->status;
1456 req->result = cqe->result;
1458 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1459 if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1460 dev_err(queue->ctrl->ctrl.device,
1461 "Bogus remote invalidation for rkey %#x\n",
1463 nvme_rdma_error_recovery(queue->ctrl);
1465 } else if (req->mr) {
1468 ret = nvme_rdma_inv_rkey(queue, req);
1469 if (unlikely(ret < 0)) {
1470 dev_err(queue->ctrl->ctrl.device,
1471 "Queueing INV WR for rkey %#x failed (%d)\n",
1472 req->mr->rkey, ret);
1473 nvme_rdma_error_recovery(queue->ctrl);
1475 /* the local invalidation completion will end the request */
1479 if (refcount_dec_and_test(&req->ref))
1480 nvme_end_request(rq, req->status, req->result);
1483 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1485 struct nvme_rdma_qe *qe =
1486 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1487 struct nvme_rdma_queue *queue = cq->cq_context;
1488 struct ib_device *ibdev = queue->device->dev;
1489 struct nvme_completion *cqe = qe->data;
1490 const size_t len = sizeof(struct nvme_completion);
1492 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1493 nvme_rdma_wr_error(cq, wc, "RECV");
1497 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1499 * AEN requests are special as they don't time out and can
1500 * survive any kind of queue freeze and often don't respond to
1501 * aborts. We don't even bother to allocate a struct request
1502 * for them but rather special case them here.
1504 if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1506 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1509 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1510 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1512 nvme_rdma_post_recv(queue, qe);
1515 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1519 for (i = 0; i < queue->queue_size; i++) {
1520 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1522 goto out_destroy_queue_ib;
1527 out_destroy_queue_ib:
1528 nvme_rdma_destroy_queue_ib(queue);
1532 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1533 struct rdma_cm_event *ev)
1535 struct rdma_cm_id *cm_id = queue->cm_id;
1536 int status = ev->status;
1537 const char *rej_msg;
1538 const struct nvme_rdma_cm_rej *rej_data;
1541 rej_msg = rdma_reject_msg(cm_id, status);
1542 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1544 if (rej_data && rej_data_len >= sizeof(u16)) {
1545 u16 sts = le16_to_cpu(rej_data->sts);
1547 dev_err(queue->ctrl->ctrl.device,
1548 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1549 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1551 dev_err(queue->ctrl->ctrl.device,
1552 "Connect rejected: status %d (%s).\n", status, rej_msg);
1558 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1560 struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1563 ret = nvme_rdma_create_queue_ib(queue);
1567 if (ctrl->opts->tos >= 0)
1568 rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1569 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1571 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1573 goto out_destroy_queue;
1579 nvme_rdma_destroy_queue_ib(queue);
1583 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1585 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1586 struct rdma_conn_param param = { };
1587 struct nvme_rdma_cm_req priv = { };
1590 param.qp_num = queue->qp->qp_num;
1591 param.flow_control = 1;
1593 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1594 /* maximum retry count */
1595 param.retry_count = 7;
1596 param.rnr_retry_count = 7;
1597 param.private_data = &priv;
1598 param.private_data_len = sizeof(priv);
1600 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1601 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1603 * set the admin queue depth to the minimum size
1604 * specified by the Fabrics standard.
1606 if (priv.qid == 0) {
1607 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1608 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1611 * current interpretation of the fabrics spec
1612 * is at minimum you make hrqsize sqsize+1, or a
1613 * 1's based representation of sqsize.
1615 priv.hrqsize = cpu_to_le16(queue->queue_size);
1616 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1619 ret = rdma_connect(queue->cm_id, ¶m);
1621 dev_err(ctrl->ctrl.device,
1622 "rdma_connect failed (%d).\n", ret);
1623 goto out_destroy_queue_ib;
1628 out_destroy_queue_ib:
1629 nvme_rdma_destroy_queue_ib(queue);
1633 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1634 struct rdma_cm_event *ev)
1636 struct nvme_rdma_queue *queue = cm_id->context;
1639 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1640 rdma_event_msg(ev->event), ev->event,
1643 switch (ev->event) {
1644 case RDMA_CM_EVENT_ADDR_RESOLVED:
1645 cm_error = nvme_rdma_addr_resolved(queue);
1647 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1648 cm_error = nvme_rdma_route_resolved(queue);
1650 case RDMA_CM_EVENT_ESTABLISHED:
1651 queue->cm_error = nvme_rdma_conn_established(queue);
1652 /* complete cm_done regardless of success/failure */
1653 complete(&queue->cm_done);
1655 case RDMA_CM_EVENT_REJECTED:
1656 nvme_rdma_destroy_queue_ib(queue);
1657 cm_error = nvme_rdma_conn_rejected(queue, ev);
1659 case RDMA_CM_EVENT_ROUTE_ERROR:
1660 case RDMA_CM_EVENT_CONNECT_ERROR:
1661 case RDMA_CM_EVENT_UNREACHABLE:
1662 nvme_rdma_destroy_queue_ib(queue);
1664 case RDMA_CM_EVENT_ADDR_ERROR:
1665 dev_dbg(queue->ctrl->ctrl.device,
1666 "CM error event %d\n", ev->event);
1667 cm_error = -ECONNRESET;
1669 case RDMA_CM_EVENT_DISCONNECTED:
1670 case RDMA_CM_EVENT_ADDR_CHANGE:
1671 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1672 dev_dbg(queue->ctrl->ctrl.device,
1673 "disconnect received - connection closed\n");
1674 nvme_rdma_error_recovery(queue->ctrl);
1676 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1677 /* device removal is handled via the ib_client API */
1680 dev_err(queue->ctrl->ctrl.device,
1681 "Unexpected RDMA CM event (%d)\n", ev->event);
1682 nvme_rdma_error_recovery(queue->ctrl);
1687 queue->cm_error = cm_error;
1688 complete(&queue->cm_done);
1694 static enum blk_eh_timer_return
1695 nvme_rdma_timeout(struct request *rq, bool reserved)
1697 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1698 struct nvme_rdma_queue *queue = req->queue;
1699 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1701 dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
1702 rq->tag, nvme_rdma_queue_idx(queue));
1705 * Restart the timer if a controller reset is already scheduled. Any
1706 * timed out commands would be handled before entering the connecting
1709 if (ctrl->ctrl.state == NVME_CTRL_RESETTING)
1710 return BLK_EH_RESET_TIMER;
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;
1751 req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
1752 sizeof(struct nvme_command),
1754 err = ib_dma_mapping_error(dev, req->sqe.dma);
1756 return BLK_STS_RESOURCE;
1758 ib_dma_sync_single_for_cpu(dev, sqe->dma,
1759 sizeof(struct nvme_command), DMA_TO_DEVICE);
1761 ret = nvme_setup_cmd(ns, rq, c);
1765 blk_mq_start_request(rq);
1767 err = nvme_rdma_map_data(queue, rq, c);
1768 if (unlikely(err < 0)) {
1769 dev_err(queue->ctrl->ctrl.device,
1770 "Failed to map data (%d)\n", err);
1774 sqe->cqe.done = nvme_rdma_send_done;
1776 ib_dma_sync_single_for_device(dev, sqe->dma,
1777 sizeof(struct nvme_command), DMA_TO_DEVICE);
1779 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1780 req->mr ? &req->reg_wr.wr : NULL);
1787 nvme_rdma_unmap_data(queue, rq);
1789 if (err == -ENOMEM || err == -EAGAIN)
1790 ret = BLK_STS_RESOURCE;
1792 ret = BLK_STS_IOERR;
1793 nvme_cleanup_cmd(rq);
1795 ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
1800 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
1802 struct nvme_rdma_queue *queue = hctx->driver_data;
1804 return ib_process_cq_direct(queue->ib_cq, -1);
1807 static void nvme_rdma_complete_rq(struct request *rq)
1809 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1810 struct nvme_rdma_queue *queue = req->queue;
1811 struct ib_device *ibdev = queue->device->dev;
1813 nvme_rdma_unmap_data(queue, rq);
1814 ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
1816 nvme_complete_rq(rq);
1819 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1821 struct nvme_rdma_ctrl *ctrl = set->driver_data;
1822 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
1824 if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
1825 /* separate read/write queues */
1826 set->map[HCTX_TYPE_DEFAULT].nr_queues =
1827 ctrl->io_queues[HCTX_TYPE_DEFAULT];
1828 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
1829 set->map[HCTX_TYPE_READ].nr_queues =
1830 ctrl->io_queues[HCTX_TYPE_READ];
1831 set->map[HCTX_TYPE_READ].queue_offset =
1832 ctrl->io_queues[HCTX_TYPE_DEFAULT];
1834 /* shared read/write queues */
1835 set->map[HCTX_TYPE_DEFAULT].nr_queues =
1836 ctrl->io_queues[HCTX_TYPE_DEFAULT];
1837 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
1838 set->map[HCTX_TYPE_READ].nr_queues =
1839 ctrl->io_queues[HCTX_TYPE_DEFAULT];
1840 set->map[HCTX_TYPE_READ].queue_offset = 0;
1842 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
1843 ctrl->device->dev, 0);
1844 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
1845 ctrl->device->dev, 0);
1847 if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
1848 /* map dedicated poll queues only if we have queues left */
1849 set->map[HCTX_TYPE_POLL].nr_queues =
1850 ctrl->io_queues[HCTX_TYPE_POLL];
1851 set->map[HCTX_TYPE_POLL].queue_offset =
1852 ctrl->io_queues[HCTX_TYPE_DEFAULT] +
1853 ctrl->io_queues[HCTX_TYPE_READ];
1854 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
1857 dev_info(ctrl->ctrl.device,
1858 "mapped %d/%d/%d default/read/poll queues.\n",
1859 ctrl->io_queues[HCTX_TYPE_DEFAULT],
1860 ctrl->io_queues[HCTX_TYPE_READ],
1861 ctrl->io_queues[HCTX_TYPE_POLL]);
1866 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1867 .queue_rq = nvme_rdma_queue_rq,
1868 .complete = nvme_rdma_complete_rq,
1869 .init_request = nvme_rdma_init_request,
1870 .exit_request = nvme_rdma_exit_request,
1871 .init_hctx = nvme_rdma_init_hctx,
1872 .timeout = nvme_rdma_timeout,
1873 .map_queues = nvme_rdma_map_queues,
1874 .poll = nvme_rdma_poll,
1877 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1878 .queue_rq = nvme_rdma_queue_rq,
1879 .complete = nvme_rdma_complete_rq,
1880 .init_request = nvme_rdma_init_request,
1881 .exit_request = nvme_rdma_exit_request,
1882 .init_hctx = nvme_rdma_init_admin_hctx,
1883 .timeout = nvme_rdma_timeout,
1886 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1888 cancel_work_sync(&ctrl->err_work);
1889 cancel_delayed_work_sync(&ctrl->reconnect_work);
1891 nvme_rdma_teardown_io_queues(ctrl, shutdown);
1892 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1894 nvme_shutdown_ctrl(&ctrl->ctrl);
1896 nvme_disable_ctrl(&ctrl->ctrl);
1897 nvme_rdma_teardown_admin_queue(ctrl, shutdown);
1900 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1902 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1905 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1907 struct nvme_rdma_ctrl *ctrl =
1908 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1910 nvme_stop_ctrl(&ctrl->ctrl);
1911 nvme_rdma_shutdown_ctrl(ctrl, false);
1913 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1914 /* state change failure should never happen */
1919 if (nvme_rdma_setup_ctrl(ctrl, false))
1925 ++ctrl->ctrl.nr_reconnects;
1926 nvme_rdma_reconnect_or_remove(ctrl);
1929 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1931 .module = THIS_MODULE,
1932 .flags = NVME_F_FABRICS,
1933 .reg_read32 = nvmf_reg_read32,
1934 .reg_read64 = nvmf_reg_read64,
1935 .reg_write32 = nvmf_reg_write32,
1936 .free_ctrl = nvme_rdma_free_ctrl,
1937 .submit_async_event = nvme_rdma_submit_async_event,
1938 .delete_ctrl = nvme_rdma_delete_ctrl,
1939 .get_address = nvmf_get_address,
1943 * Fails a connection request if it matches an existing controller
1944 * (association) with the same tuple:
1945 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1947 * if local address is not specified in the request, it will match an
1948 * existing controller with all the other parameters the same and no
1949 * local port address specified as well.
1951 * The ports don't need to be compared as they are intrinsically
1952 * already matched by the port pointers supplied.
1955 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1957 struct nvme_rdma_ctrl *ctrl;
1960 mutex_lock(&nvme_rdma_ctrl_mutex);
1961 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1962 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
1966 mutex_unlock(&nvme_rdma_ctrl_mutex);
1971 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1972 struct nvmf_ctrl_options *opts)
1974 struct nvme_rdma_ctrl *ctrl;
1978 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1980 return ERR_PTR(-ENOMEM);
1981 ctrl->ctrl.opts = opts;
1982 INIT_LIST_HEAD(&ctrl->list);
1984 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
1986 kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
1987 if (!opts->trsvcid) {
1991 opts->mask |= NVMF_OPT_TRSVCID;
1994 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1995 opts->traddr, opts->trsvcid, &ctrl->addr);
1997 pr_err("malformed address passed: %s:%s\n",
1998 opts->traddr, opts->trsvcid);
2002 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2003 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2004 opts->host_traddr, NULL, &ctrl->src_addr);
2006 pr_err("malformed src address passed: %s\n",
2012 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2017 INIT_DELAYED_WORK(&ctrl->reconnect_work,
2018 nvme_rdma_reconnect_ctrl_work);
2019 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2020 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2022 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2023 opts->nr_poll_queues + 1;
2024 ctrl->ctrl.sqsize = opts->queue_size - 1;
2025 ctrl->ctrl.kato = opts->kato;
2028 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2033 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2034 0 /* no quirks, we're perfect! */);
2036 goto out_kfree_queues;
2038 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2039 WARN_ON_ONCE(!changed);
2041 ret = nvme_rdma_setup_ctrl(ctrl, true);
2043 goto out_uninit_ctrl;
2045 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2046 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2048 nvme_get_ctrl(&ctrl->ctrl);
2050 mutex_lock(&nvme_rdma_ctrl_mutex);
2051 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2052 mutex_unlock(&nvme_rdma_ctrl_mutex);
2057 nvme_uninit_ctrl(&ctrl->ctrl);
2058 nvme_put_ctrl(&ctrl->ctrl);
2061 return ERR_PTR(ret);
2063 kfree(ctrl->queues);
2066 return ERR_PTR(ret);
2069 static struct nvmf_transport_ops nvme_rdma_transport = {
2071 .module = THIS_MODULE,
2072 .required_opts = NVMF_OPT_TRADDR,
2073 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2074 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2075 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2077 .create_ctrl = nvme_rdma_create_ctrl,
2080 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2082 struct nvme_rdma_ctrl *ctrl;
2083 struct nvme_rdma_device *ndev;
2086 mutex_lock(&device_list_mutex);
2087 list_for_each_entry(ndev, &device_list, entry) {
2088 if (ndev->dev == ib_device) {
2093 mutex_unlock(&device_list_mutex);
2098 /* Delete all controllers using this device */
2099 mutex_lock(&nvme_rdma_ctrl_mutex);
2100 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2101 if (ctrl->device->dev != ib_device)
2103 nvme_delete_ctrl(&ctrl->ctrl);
2105 mutex_unlock(&nvme_rdma_ctrl_mutex);
2107 flush_workqueue(nvme_delete_wq);
2110 static struct ib_client nvme_rdma_ib_client = {
2111 .name = "nvme_rdma",
2112 .remove = nvme_rdma_remove_one
2115 static int __init nvme_rdma_init_module(void)
2119 ret = ib_register_client(&nvme_rdma_ib_client);
2123 ret = nvmf_register_transport(&nvme_rdma_transport);
2125 goto err_unreg_client;
2130 ib_unregister_client(&nvme_rdma_ib_client);
2134 static void __exit nvme_rdma_cleanup_module(void)
2136 struct nvme_rdma_ctrl *ctrl;
2138 nvmf_unregister_transport(&nvme_rdma_transport);
2139 ib_unregister_client(&nvme_rdma_ib_client);
2141 mutex_lock(&nvme_rdma_ctrl_mutex);
2142 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2143 nvme_delete_ctrl(&ctrl->ctrl);
2144 mutex_unlock(&nvme_rdma_ctrl_mutex);
2145 flush_workqueue(nvme_delete_wq);
2148 module_init(nvme_rdma_init_module);
2149 module_exit(nvme_rdma_cleanup_module);
2151 MODULE_LICENSE("GPL v2");