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;
124 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
126 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
129 static LIST_HEAD(device_list);
130 static DEFINE_MUTEX(device_list_mutex);
132 static LIST_HEAD(nvme_rdma_ctrl_list);
133 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
136 * Disabling this option makes small I/O goes faster, but is fundamentally
137 * unsafe. With it turned off we will have to register a global rkey that
138 * allows read and write access to all physical memory.
140 static bool register_always = true;
141 module_param(register_always, bool, 0444);
142 MODULE_PARM_DESC(register_always,
143 "Use memory registration even for contiguous memory regions");
145 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
146 struct rdma_cm_event *event);
147 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
149 static const struct blk_mq_ops nvme_rdma_mq_ops;
150 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
152 /* XXX: really should move to a generic header sooner or later.. */
153 static inline void put_unaligned_le24(u32 val, u8 *p)
160 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
162 return queue - queue->ctrl->queues;
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)) {
193 static void nvme_rdma_free_ring(struct ib_device *ibdev,
194 struct nvme_rdma_qe *ring, size_t ib_queue_size,
195 size_t capsule_size, enum dma_data_direction dir)
199 for (i = 0; i < ib_queue_size; i++)
200 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
204 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
205 size_t ib_queue_size, size_t capsule_size,
206 enum dma_data_direction dir)
208 struct nvme_rdma_qe *ring;
211 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
215 for (i = 0; i < ib_queue_size; i++) {
216 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
223 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
227 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
229 pr_debug("QP event %s (%d)\n",
230 ib_event_msg(event->event), event->event);
234 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
236 wait_for_completion_interruptible_timeout(&queue->cm_done,
237 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
238 return queue->cm_error;
241 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
243 struct nvme_rdma_device *dev = queue->device;
244 struct ib_qp_init_attr init_attr;
247 memset(&init_attr, 0, sizeof(init_attr));
248 init_attr.event_handler = nvme_rdma_qp_event;
250 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
252 init_attr.cap.max_recv_wr = queue->queue_size + 1;
253 init_attr.cap.max_recv_sge = 1;
254 init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
255 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
256 init_attr.qp_type = IB_QPT_RC;
257 init_attr.send_cq = queue->ib_cq;
258 init_attr.recv_cq = queue->ib_cq;
260 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
262 queue->qp = queue->cm_id->qp;
266 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
267 struct request *rq, unsigned int hctx_idx)
269 struct nvme_rdma_ctrl *ctrl = set->driver_data;
270 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
271 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
272 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
273 struct nvme_rdma_device *dev = queue->device;
275 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
279 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
280 struct request *rq, unsigned int hctx_idx,
281 unsigned int numa_node)
283 struct nvme_rdma_ctrl *ctrl = set->driver_data;
284 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
285 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
286 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
287 struct nvme_rdma_device *dev = queue->device;
288 struct ib_device *ibdev = dev->dev;
291 nvme_req(rq)->ctrl = &ctrl->ctrl;
292 ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
302 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
303 unsigned int hctx_idx)
305 struct nvme_rdma_ctrl *ctrl = data;
306 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
308 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
310 hctx->driver_data = queue;
314 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
315 unsigned int hctx_idx)
317 struct nvme_rdma_ctrl *ctrl = data;
318 struct nvme_rdma_queue *queue = &ctrl->queues[0];
320 BUG_ON(hctx_idx != 0);
322 hctx->driver_data = queue;
326 static void nvme_rdma_free_dev(struct kref *ref)
328 struct nvme_rdma_device *ndev =
329 container_of(ref, struct nvme_rdma_device, ref);
331 mutex_lock(&device_list_mutex);
332 list_del(&ndev->entry);
333 mutex_unlock(&device_list_mutex);
335 ib_dealloc_pd(ndev->pd);
339 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
341 kref_put(&dev->ref, nvme_rdma_free_dev);
344 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
346 return kref_get_unless_zero(&dev->ref);
349 static struct nvme_rdma_device *
350 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
352 struct nvme_rdma_device *ndev;
354 mutex_lock(&device_list_mutex);
355 list_for_each_entry(ndev, &device_list, entry) {
356 if (ndev->dev->node_guid == cm_id->device->node_guid &&
357 nvme_rdma_dev_get(ndev))
361 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
365 ndev->dev = cm_id->device;
366 kref_init(&ndev->ref);
368 ndev->pd = ib_alloc_pd(ndev->dev,
369 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
370 if (IS_ERR(ndev->pd))
373 if (!(ndev->dev->attrs.device_cap_flags &
374 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
375 dev_err(&ndev->dev->dev,
376 "Memory registrations not supported.\n");
380 ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
381 ndev->dev->attrs.max_sge - 1);
382 list_add(&ndev->entry, &device_list);
384 mutex_unlock(&device_list_mutex);
388 ib_dealloc_pd(ndev->pd);
392 mutex_unlock(&device_list_mutex);
396 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
398 struct nvme_rdma_device *dev;
399 struct ib_device *ibdev;
401 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
407 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
410 * The cm_id object might have been destroyed during RDMA connection
411 * establishment error flow to avoid getting other cma events, thus
412 * the destruction of the QP shouldn't use rdma_cm API.
414 ib_destroy_qp(queue->qp);
415 ib_free_cq(queue->ib_cq);
417 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
418 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
420 nvme_rdma_dev_put(dev);
423 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev)
425 return min_t(u32, NVME_RDMA_MAX_SEGMENTS,
426 ibdev->attrs.max_fast_reg_page_list_len);
429 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
431 struct ib_device *ibdev;
432 const int send_wr_factor = 3; /* MR, SEND, INV */
433 const int cq_factor = send_wr_factor + 1; /* + RECV */
434 int comp_vector, idx = nvme_rdma_queue_idx(queue);
437 queue->device = nvme_rdma_find_get_device(queue->cm_id);
438 if (!queue->device) {
439 dev_err(queue->cm_id->device->dev.parent,
440 "no client data found!\n");
441 return -ECONNREFUSED;
443 ibdev = queue->device->dev;
446 * Spread I/O queues completion vectors according their queue index.
447 * Admin queues can always go on completion vector 0.
449 comp_vector = idx == 0 ? idx : idx - 1;
451 /* +1 for ib_stop_cq */
452 queue->ib_cq = ib_alloc_cq(ibdev, queue,
453 cq_factor * queue->queue_size + 1,
454 comp_vector, IB_POLL_SOFTIRQ);
455 if (IS_ERR(queue->ib_cq)) {
456 ret = PTR_ERR(queue->ib_cq);
460 ret = nvme_rdma_create_qp(queue, send_wr_factor);
462 goto out_destroy_ib_cq;
464 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
465 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
466 if (!queue->rsp_ring) {
471 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
474 nvme_rdma_get_max_fr_pages(ibdev));
476 dev_err(queue->ctrl->ctrl.device,
477 "failed to initialize MR pool sized %d for QID %d\n",
478 queue->queue_size, idx);
479 goto out_destroy_ring;
482 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
487 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
488 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
490 rdma_destroy_qp(queue->cm_id);
492 ib_free_cq(queue->ib_cq);
494 nvme_rdma_dev_put(queue->device);
498 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
499 int idx, size_t queue_size)
501 struct nvme_rdma_queue *queue;
502 struct sockaddr *src_addr = NULL;
505 queue = &ctrl->queues[idx];
507 init_completion(&queue->cm_done);
510 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
512 queue->cmnd_capsule_len = sizeof(struct nvme_command);
514 queue->queue_size = queue_size;
516 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
517 RDMA_PS_TCP, IB_QPT_RC);
518 if (IS_ERR(queue->cm_id)) {
519 dev_info(ctrl->ctrl.device,
520 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
521 return PTR_ERR(queue->cm_id);
524 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
525 src_addr = (struct sockaddr *)&ctrl->src_addr;
527 queue->cm_error = -ETIMEDOUT;
528 ret = rdma_resolve_addr(queue->cm_id, src_addr,
529 (struct sockaddr *)&ctrl->addr,
530 NVME_RDMA_CONNECT_TIMEOUT_MS);
532 dev_info(ctrl->ctrl.device,
533 "rdma_resolve_addr failed (%d).\n", ret);
534 goto out_destroy_cm_id;
537 ret = nvme_rdma_wait_for_cm(queue);
539 dev_info(ctrl->ctrl.device,
540 "rdma connection establishment failed (%d)\n", ret);
541 goto out_destroy_cm_id;
544 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
549 rdma_destroy_id(queue->cm_id);
550 nvme_rdma_destroy_queue_ib(queue);
554 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
556 if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
559 rdma_disconnect(queue->cm_id);
560 ib_drain_qp(queue->qp);
563 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
565 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
568 nvme_rdma_destroy_queue_ib(queue);
569 rdma_destroy_id(queue->cm_id);
572 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
576 for (i = 1; i < ctrl->ctrl.queue_count; i++)
577 nvme_rdma_free_queue(&ctrl->queues[i]);
580 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
584 for (i = 1; i < ctrl->ctrl.queue_count; i++)
585 nvme_rdma_stop_queue(&ctrl->queues[i]);
588 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
593 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
595 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
598 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags);
600 dev_info(ctrl->ctrl.device,
601 "failed to connect queue: %d ret=%d\n", idx, ret);
605 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
609 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
610 ret = nvme_rdma_start_queue(ctrl, i);
612 goto out_stop_queues;
618 for (i--; i >= 1; i--)
619 nvme_rdma_stop_queue(&ctrl->queues[i]);
623 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
625 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
626 struct ib_device *ibdev = ctrl->device->dev;
627 unsigned int nr_io_queues;
630 nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
633 * we map queues according to the device irq vectors for
634 * optimal locality so we don't need more queues than
635 * completion vectors.
637 nr_io_queues = min_t(unsigned int, nr_io_queues,
638 ibdev->num_comp_vectors);
640 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
644 ctrl->ctrl.queue_count = nr_io_queues + 1;
645 if (ctrl->ctrl.queue_count < 2)
648 dev_info(ctrl->ctrl.device,
649 "creating %d I/O queues.\n", nr_io_queues);
651 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
652 ret = nvme_rdma_alloc_queue(ctrl, i,
653 ctrl->ctrl.sqsize + 1);
655 goto out_free_queues;
661 for (i--; i >= 1; i--)
662 nvme_rdma_free_queue(&ctrl->queues[i]);
667 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl,
668 struct blk_mq_tag_set *set)
670 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
672 blk_mq_free_tag_set(set);
673 nvme_rdma_dev_put(ctrl->device);
676 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
679 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
680 struct blk_mq_tag_set *set;
684 set = &ctrl->admin_tag_set;
685 memset(set, 0, sizeof(*set));
686 set->ops = &nvme_rdma_admin_mq_ops;
687 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
688 set->reserved_tags = 2; /* connect + keep-alive */
689 set->numa_node = NUMA_NO_NODE;
690 set->cmd_size = sizeof(struct nvme_rdma_request) +
691 SG_CHUNK_SIZE * sizeof(struct scatterlist);
692 set->driver_data = ctrl;
693 set->nr_hw_queues = 1;
694 set->timeout = ADMIN_TIMEOUT;
695 set->flags = BLK_MQ_F_NO_SCHED;
697 set = &ctrl->tag_set;
698 memset(set, 0, sizeof(*set));
699 set->ops = &nvme_rdma_mq_ops;
700 set->queue_depth = nctrl->sqsize + 1;
701 set->reserved_tags = 1; /* fabric connect */
702 set->numa_node = NUMA_NO_NODE;
703 set->flags = BLK_MQ_F_SHOULD_MERGE;
704 set->cmd_size = sizeof(struct nvme_rdma_request) +
705 SG_CHUNK_SIZE * sizeof(struct scatterlist);
706 set->driver_data = ctrl;
707 set->nr_hw_queues = nctrl->queue_count - 1;
708 set->timeout = NVME_IO_TIMEOUT;
711 ret = blk_mq_alloc_tag_set(set);
716 * We need a reference on the device as long as the tag_set is alive,
717 * as the MRs in the request structures need a valid ib_device.
719 ret = nvme_rdma_dev_get(ctrl->device);
722 goto out_free_tagset;
728 blk_mq_free_tag_set(set);
733 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
737 blk_cleanup_queue(ctrl->ctrl.admin_q);
738 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
740 if (ctrl->async_event_sqe.data) {
741 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
742 sizeof(struct nvme_command), DMA_TO_DEVICE);
743 ctrl->async_event_sqe.data = NULL;
745 nvme_rdma_free_queue(&ctrl->queues[0]);
748 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
753 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
757 ctrl->device = ctrl->queues[0].device;
759 ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
761 error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
762 sizeof(struct nvme_command), DMA_TO_DEVICE);
767 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
768 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
769 error = PTR_ERR(ctrl->ctrl.admin_tagset);
770 goto out_free_async_qe;
773 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
774 if (IS_ERR(ctrl->ctrl.admin_q)) {
775 error = PTR_ERR(ctrl->ctrl.admin_q);
776 goto out_free_tagset;
780 error = nvme_rdma_start_queue(ctrl, 0);
782 goto out_cleanup_queue;
784 error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
787 dev_err(ctrl->ctrl.device,
788 "prop_get NVME_REG_CAP failed\n");
793 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
795 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
799 ctrl->ctrl.max_hw_sectors =
800 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
802 error = nvme_init_identify(&ctrl->ctrl);
809 nvme_rdma_stop_queue(&ctrl->queues[0]);
812 blk_cleanup_queue(ctrl->ctrl.admin_q);
815 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
817 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
818 sizeof(struct nvme_command), DMA_TO_DEVICE);
820 nvme_rdma_free_queue(&ctrl->queues[0]);
824 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
828 blk_cleanup_queue(ctrl->ctrl.connect_q);
829 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
831 nvme_rdma_free_io_queues(ctrl);
834 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
838 ret = nvme_rdma_alloc_io_queues(ctrl);
843 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
844 if (IS_ERR(ctrl->ctrl.tagset)) {
845 ret = PTR_ERR(ctrl->ctrl.tagset);
846 goto out_free_io_queues;
849 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
850 if (IS_ERR(ctrl->ctrl.connect_q)) {
851 ret = PTR_ERR(ctrl->ctrl.connect_q);
852 goto out_free_tag_set;
855 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
856 ctrl->ctrl.queue_count - 1);
859 ret = nvme_rdma_start_io_queues(ctrl);
861 goto out_cleanup_connect_q;
865 out_cleanup_connect_q:
867 blk_cleanup_queue(ctrl->ctrl.connect_q);
870 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
872 nvme_rdma_free_io_queues(ctrl);
876 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
878 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
880 cancel_work_sync(&ctrl->err_work);
881 cancel_delayed_work_sync(&ctrl->reconnect_work);
884 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
886 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
888 if (list_empty(&ctrl->list))
891 mutex_lock(&nvme_rdma_ctrl_mutex);
892 list_del(&ctrl->list);
893 mutex_unlock(&nvme_rdma_ctrl_mutex);
895 nvmf_free_options(nctrl->opts);
901 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
903 /* If we are resetting/deleting then do nothing */
904 if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
905 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
906 ctrl->ctrl.state == NVME_CTRL_LIVE);
910 if (nvmf_should_reconnect(&ctrl->ctrl)) {
911 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
912 ctrl->ctrl.opts->reconnect_delay);
913 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
914 ctrl->ctrl.opts->reconnect_delay * HZ);
916 nvme_delete_ctrl(&ctrl->ctrl);
920 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
925 ret = nvme_rdma_configure_admin_queue(ctrl, new);
929 if (ctrl->ctrl.icdoff) {
930 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
934 if (!(ctrl->ctrl.sgls & (1 << 2))) {
935 dev_err(ctrl->ctrl.device,
936 "Mandatory keyed sgls are not supported!\n");
940 if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
941 dev_warn(ctrl->ctrl.device,
942 "queue_size %zu > ctrl sqsize %u, clamping down\n",
943 ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
946 if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
947 dev_warn(ctrl->ctrl.device,
948 "sqsize %u > ctrl maxcmd %u, clamping down\n",
949 ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
950 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
953 if (ctrl->ctrl.sgls & (1 << 20))
954 ctrl->use_inline_data = true;
956 if (ctrl->ctrl.queue_count > 1) {
957 ret = nvme_rdma_configure_io_queues(ctrl, new);
962 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
964 /* state change failure is ok if we're in DELETING state */
965 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
970 nvme_start_ctrl(&ctrl->ctrl);
974 if (ctrl->ctrl.queue_count > 1)
975 nvme_rdma_destroy_io_queues(ctrl, new);
977 nvme_rdma_stop_queue(&ctrl->queues[0]);
978 nvme_rdma_destroy_admin_queue(ctrl, new);
982 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
984 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
985 struct nvme_rdma_ctrl, reconnect_work);
987 ++ctrl->ctrl.nr_reconnects;
989 if (nvme_rdma_setup_ctrl(ctrl, false))
992 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
993 ctrl->ctrl.nr_reconnects);
995 ctrl->ctrl.nr_reconnects = 0;
1000 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1001 ctrl->ctrl.nr_reconnects);
1002 nvme_rdma_reconnect_or_remove(ctrl);
1005 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1007 struct nvme_rdma_ctrl *ctrl = container_of(work,
1008 struct nvme_rdma_ctrl, err_work);
1010 nvme_stop_keep_alive(&ctrl->ctrl);
1012 if (ctrl->ctrl.queue_count > 1) {
1013 nvme_stop_queues(&ctrl->ctrl);
1014 nvme_rdma_stop_io_queues(ctrl);
1015 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1016 nvme_cancel_request, &ctrl->ctrl);
1017 nvme_rdma_destroy_io_queues(ctrl, false);
1020 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1021 nvme_rdma_stop_queue(&ctrl->queues[0]);
1022 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1023 nvme_cancel_request, &ctrl->ctrl);
1024 nvme_rdma_destroy_admin_queue(ctrl, false);
1027 * queues are not a live anymore, so restart the queues to fail fast
1030 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1031 nvme_start_queues(&ctrl->ctrl);
1033 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1034 /* state change failure is ok if we're in DELETING state */
1035 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1039 nvme_rdma_reconnect_or_remove(ctrl);
1042 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1044 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1047 queue_work(nvme_wq, &ctrl->err_work);
1050 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1053 struct nvme_rdma_queue *queue = cq->cq_context;
1054 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1056 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1057 dev_info(ctrl->ctrl.device,
1058 "%s for CQE 0x%p failed with status %s (%d)\n",
1060 ib_wc_status_msg(wc->status), wc->status);
1061 nvme_rdma_error_recovery(ctrl);
1064 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1066 if (unlikely(wc->status != IB_WC_SUCCESS))
1067 nvme_rdma_wr_error(cq, wc, "MEMREG");
1070 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1072 struct nvme_rdma_request *req =
1073 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1074 struct request *rq = blk_mq_rq_from_pdu(req);
1076 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1077 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1081 if (refcount_dec_and_test(&req->ref))
1082 nvme_end_request(rq, req->status, req->result);
1086 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1087 struct nvme_rdma_request *req)
1089 struct ib_send_wr *bad_wr;
1090 struct ib_send_wr wr = {
1091 .opcode = IB_WR_LOCAL_INV,
1094 .send_flags = IB_SEND_SIGNALED,
1095 .ex.invalidate_rkey = req->mr->rkey,
1098 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1099 wr.wr_cqe = &req->reg_cqe;
1101 return ib_post_send(queue->qp, &wr, &bad_wr);
1104 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1107 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1108 struct nvme_rdma_device *dev = queue->device;
1109 struct ib_device *ibdev = dev->dev;
1111 if (!blk_rq_payload_bytes(rq))
1115 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1119 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1120 req->nents, rq_data_dir(rq) ==
1121 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1123 nvme_cleanup_cmd(rq);
1124 sg_free_table_chained(&req->sg_table, true);
1127 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1129 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1132 put_unaligned_le24(0, sg->length);
1133 put_unaligned_le32(0, sg->key);
1134 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1138 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1139 struct nvme_rdma_request *req, struct nvme_command *c,
1142 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1143 struct scatterlist *sgl = req->sg_table.sgl;
1144 struct ib_sge *sge = &req->sge[1];
1148 for (i = 0; i < count; i++, sgl++, sge++) {
1149 sge->addr = sg_dma_address(sgl);
1150 sge->length = sg_dma_len(sgl);
1151 sge->lkey = queue->device->pd->local_dma_lkey;
1155 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1156 sg->length = cpu_to_le32(len);
1157 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1159 req->num_sge += count;
1163 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1164 struct nvme_rdma_request *req, struct nvme_command *c)
1166 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1168 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1169 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1170 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1171 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1175 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1176 struct nvme_rdma_request *req, struct nvme_command *c,
1179 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1182 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1183 if (WARN_ON_ONCE(!req->mr))
1187 * Align the MR to a 4K page size to match the ctrl page size and
1188 * the block virtual boundary.
1190 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1191 if (unlikely(nr < count)) {
1192 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1199 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1201 req->reg_cqe.done = nvme_rdma_memreg_done;
1202 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1203 req->reg_wr.wr.opcode = IB_WR_REG_MR;
1204 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1205 req->reg_wr.wr.num_sge = 0;
1206 req->reg_wr.mr = req->mr;
1207 req->reg_wr.key = req->mr->rkey;
1208 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1209 IB_ACCESS_REMOTE_READ |
1210 IB_ACCESS_REMOTE_WRITE;
1212 sg->addr = cpu_to_le64(req->mr->iova);
1213 put_unaligned_le24(req->mr->length, sg->length);
1214 put_unaligned_le32(req->mr->rkey, sg->key);
1215 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1216 NVME_SGL_FMT_INVALIDATE;
1221 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1222 struct request *rq, struct nvme_command *c)
1224 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1225 struct nvme_rdma_device *dev = queue->device;
1226 struct ib_device *ibdev = dev->dev;
1230 refcount_set(&req->ref, 2); /* send and recv completions */
1232 c->common.flags |= NVME_CMD_SGL_METABUF;
1234 if (!blk_rq_payload_bytes(rq))
1235 return nvme_rdma_set_sg_null(c);
1237 req->sg_table.sgl = req->first_sgl;
1238 ret = sg_alloc_table_chained(&req->sg_table,
1239 blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1243 req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1245 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1246 rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1247 if (unlikely(count <= 0)) {
1249 goto out_free_table;
1252 if (count <= dev->num_inline_segments) {
1253 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1254 queue->ctrl->use_inline_data &&
1255 blk_rq_payload_bytes(rq) <=
1256 nvme_rdma_inline_data_size(queue)) {
1257 ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1261 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1262 ret = nvme_rdma_map_sg_single(queue, req, c);
1267 ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1275 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1276 req->nents, rq_data_dir(rq) ==
1277 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1279 sg_free_table_chained(&req->sg_table, true);
1283 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1285 struct nvme_rdma_qe *qe =
1286 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1287 struct nvme_rdma_request *req =
1288 container_of(qe, struct nvme_rdma_request, sqe);
1289 struct request *rq = blk_mq_rq_from_pdu(req);
1291 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1292 nvme_rdma_wr_error(cq, wc, "SEND");
1296 if (refcount_dec_and_test(&req->ref))
1297 nvme_end_request(rq, req->status, req->result);
1300 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1301 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1302 struct ib_send_wr *first)
1304 struct ib_send_wr wr, *bad_wr;
1307 sge->addr = qe->dma;
1308 sge->length = sizeof(struct nvme_command),
1309 sge->lkey = queue->device->pd->local_dma_lkey;
1312 wr.wr_cqe = &qe->cqe;
1314 wr.num_sge = num_sge;
1315 wr.opcode = IB_WR_SEND;
1316 wr.send_flags = IB_SEND_SIGNALED;
1323 ret = ib_post_send(queue->qp, first, &bad_wr);
1324 if (unlikely(ret)) {
1325 dev_err(queue->ctrl->ctrl.device,
1326 "%s failed with error code %d\n", __func__, ret);
1331 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1332 struct nvme_rdma_qe *qe)
1334 struct ib_recv_wr wr, *bad_wr;
1338 list.addr = qe->dma;
1339 list.length = sizeof(struct nvme_completion);
1340 list.lkey = queue->device->pd->local_dma_lkey;
1342 qe->cqe.done = nvme_rdma_recv_done;
1345 wr.wr_cqe = &qe->cqe;
1349 ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1350 if (unlikely(ret)) {
1351 dev_err(queue->ctrl->ctrl.device,
1352 "%s failed with error code %d\n", __func__, ret);
1357 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1359 u32 queue_idx = nvme_rdma_queue_idx(queue);
1362 return queue->ctrl->admin_tag_set.tags[queue_idx];
1363 return queue->ctrl->tag_set.tags[queue_idx - 1];
1366 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1368 if (unlikely(wc->status != IB_WC_SUCCESS))
1369 nvme_rdma_wr_error(cq, wc, "ASYNC");
1372 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1374 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1375 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1376 struct ib_device *dev = queue->device->dev;
1377 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1378 struct nvme_command *cmd = sqe->data;
1382 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1384 memset(cmd, 0, sizeof(*cmd));
1385 cmd->common.opcode = nvme_admin_async_event;
1386 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1387 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1388 nvme_rdma_set_sg_null(cmd);
1390 sqe->cqe.done = nvme_rdma_async_done;
1392 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1395 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1399 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1400 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1403 struct nvme_rdma_request *req;
1406 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1408 dev_err(queue->ctrl->ctrl.device,
1409 "tag 0x%x on QP %#x not found\n",
1410 cqe->command_id, queue->qp->qp_num);
1411 nvme_rdma_error_recovery(queue->ctrl);
1414 req = blk_mq_rq_to_pdu(rq);
1416 req->status = cqe->status;
1417 req->result = cqe->result;
1419 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1420 if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1421 dev_err(queue->ctrl->ctrl.device,
1422 "Bogus remote invalidation for rkey %#x\n",
1424 nvme_rdma_error_recovery(queue->ctrl);
1426 } else if (req->mr) {
1427 ret = nvme_rdma_inv_rkey(queue, req);
1428 if (unlikely(ret < 0)) {
1429 dev_err(queue->ctrl->ctrl.device,
1430 "Queueing INV WR for rkey %#x failed (%d)\n",
1431 req->mr->rkey, ret);
1432 nvme_rdma_error_recovery(queue->ctrl);
1434 /* the local invalidation completion will end the request */
1438 if (refcount_dec_and_test(&req->ref)) {
1441 nvme_end_request(rq, req->status, req->result);
1447 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1449 struct nvme_rdma_qe *qe =
1450 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1451 struct nvme_rdma_queue *queue = cq->cq_context;
1452 struct ib_device *ibdev = queue->device->dev;
1453 struct nvme_completion *cqe = qe->data;
1454 const size_t len = sizeof(struct nvme_completion);
1457 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1458 nvme_rdma_wr_error(cq, wc, "RECV");
1462 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1464 * AEN requests are special as they don't time out and can
1465 * survive any kind of queue freeze and often don't respond to
1466 * aborts. We don't even bother to allocate a struct request
1467 * for them but rather special case them here.
1469 if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1470 cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1471 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1474 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1475 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1477 nvme_rdma_post_recv(queue, qe);
1481 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1483 __nvme_rdma_recv_done(cq, wc, -1);
1486 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1490 for (i = 0; i < queue->queue_size; i++) {
1491 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1493 goto out_destroy_queue_ib;
1498 out_destroy_queue_ib:
1499 nvme_rdma_destroy_queue_ib(queue);
1503 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1504 struct rdma_cm_event *ev)
1506 struct rdma_cm_id *cm_id = queue->cm_id;
1507 int status = ev->status;
1508 const char *rej_msg;
1509 const struct nvme_rdma_cm_rej *rej_data;
1512 rej_msg = rdma_reject_msg(cm_id, status);
1513 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1515 if (rej_data && rej_data_len >= sizeof(u16)) {
1516 u16 sts = le16_to_cpu(rej_data->sts);
1518 dev_err(queue->ctrl->ctrl.device,
1519 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1520 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1522 dev_err(queue->ctrl->ctrl.device,
1523 "Connect rejected: status %d (%s).\n", status, rej_msg);
1529 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1533 ret = nvme_rdma_create_queue_ib(queue);
1537 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1539 dev_err(queue->ctrl->ctrl.device,
1540 "rdma_resolve_route failed (%d).\n",
1542 goto out_destroy_queue;
1548 nvme_rdma_destroy_queue_ib(queue);
1552 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1554 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1555 struct rdma_conn_param param = { };
1556 struct nvme_rdma_cm_req priv = { };
1559 param.qp_num = queue->qp->qp_num;
1560 param.flow_control = 1;
1562 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1563 /* maximum retry count */
1564 param.retry_count = 7;
1565 param.rnr_retry_count = 7;
1566 param.private_data = &priv;
1567 param.private_data_len = sizeof(priv);
1569 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1570 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1572 * set the admin queue depth to the minimum size
1573 * specified by the Fabrics standard.
1575 if (priv.qid == 0) {
1576 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1577 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1580 * current interpretation of the fabrics spec
1581 * is at minimum you make hrqsize sqsize+1, or a
1582 * 1's based representation of sqsize.
1584 priv.hrqsize = cpu_to_le16(queue->queue_size);
1585 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1588 ret = rdma_connect(queue->cm_id, ¶m);
1590 dev_err(ctrl->ctrl.device,
1591 "rdma_connect failed (%d).\n", ret);
1592 goto out_destroy_queue_ib;
1597 out_destroy_queue_ib:
1598 nvme_rdma_destroy_queue_ib(queue);
1602 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1603 struct rdma_cm_event *ev)
1605 struct nvme_rdma_queue *queue = cm_id->context;
1608 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1609 rdma_event_msg(ev->event), ev->event,
1612 switch (ev->event) {
1613 case RDMA_CM_EVENT_ADDR_RESOLVED:
1614 cm_error = nvme_rdma_addr_resolved(queue);
1616 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1617 cm_error = nvme_rdma_route_resolved(queue);
1619 case RDMA_CM_EVENT_ESTABLISHED:
1620 queue->cm_error = nvme_rdma_conn_established(queue);
1621 /* complete cm_done regardless of success/failure */
1622 complete(&queue->cm_done);
1624 case RDMA_CM_EVENT_REJECTED:
1625 nvme_rdma_destroy_queue_ib(queue);
1626 cm_error = nvme_rdma_conn_rejected(queue, ev);
1628 case RDMA_CM_EVENT_ROUTE_ERROR:
1629 case RDMA_CM_EVENT_CONNECT_ERROR:
1630 case RDMA_CM_EVENT_UNREACHABLE:
1631 nvme_rdma_destroy_queue_ib(queue);
1633 case RDMA_CM_EVENT_ADDR_ERROR:
1634 dev_dbg(queue->ctrl->ctrl.device,
1635 "CM error event %d\n", ev->event);
1636 cm_error = -ECONNRESET;
1638 case RDMA_CM_EVENT_DISCONNECTED:
1639 case RDMA_CM_EVENT_ADDR_CHANGE:
1640 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1641 dev_dbg(queue->ctrl->ctrl.device,
1642 "disconnect received - connection closed\n");
1643 nvme_rdma_error_recovery(queue->ctrl);
1645 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1646 /* device removal is handled via the ib_client API */
1649 dev_err(queue->ctrl->ctrl.device,
1650 "Unexpected RDMA CM event (%d)\n", ev->event);
1651 nvme_rdma_error_recovery(queue->ctrl);
1656 queue->cm_error = cm_error;
1657 complete(&queue->cm_done);
1663 static enum blk_eh_timer_return
1664 nvme_rdma_timeout(struct request *rq, bool reserved)
1666 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1668 dev_warn(req->queue->ctrl->ctrl.device,
1669 "I/O %d QID %d timeout, reset controller\n",
1670 rq->tag, nvme_rdma_queue_idx(req->queue));
1672 /* queue error recovery */
1673 nvme_rdma_error_recovery(req->queue->ctrl);
1675 /* fail with DNR on cmd timeout */
1676 nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1681 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1682 const struct blk_mq_queue_data *bd)
1684 struct nvme_ns *ns = hctx->queue->queuedata;
1685 struct nvme_rdma_queue *queue = hctx->driver_data;
1686 struct request *rq = bd->rq;
1687 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1688 struct nvme_rdma_qe *sqe = &req->sqe;
1689 struct nvme_command *c = sqe->data;
1690 struct ib_device *dev;
1691 bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1695 WARN_ON_ONCE(rq->tag < 0);
1697 if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1698 return nvmf_fail_nonready_command(rq);
1700 dev = queue->device->dev;
1701 ib_dma_sync_single_for_cpu(dev, sqe->dma,
1702 sizeof(struct nvme_command), DMA_TO_DEVICE);
1704 ret = nvme_setup_cmd(ns, rq, c);
1708 blk_mq_start_request(rq);
1710 err = nvme_rdma_map_data(queue, rq, c);
1711 if (unlikely(err < 0)) {
1712 dev_err(queue->ctrl->ctrl.device,
1713 "Failed to map data (%d)\n", err);
1714 nvme_cleanup_cmd(rq);
1718 sqe->cqe.done = nvme_rdma_send_done;
1720 ib_dma_sync_single_for_device(dev, sqe->dma,
1721 sizeof(struct nvme_command), DMA_TO_DEVICE);
1723 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1724 req->mr ? &req->reg_wr.wr : NULL);
1725 if (unlikely(err)) {
1726 nvme_rdma_unmap_data(queue, rq);
1732 if (err == -ENOMEM || err == -EAGAIN)
1733 return BLK_STS_RESOURCE;
1734 return BLK_STS_IOERR;
1737 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1739 struct nvme_rdma_queue *queue = hctx->driver_data;
1740 struct ib_cq *cq = queue->ib_cq;
1744 while (ib_poll_cq(cq, 1, &wc) > 0) {
1745 struct ib_cqe *cqe = wc.wr_cqe;
1748 if (cqe->done == nvme_rdma_recv_done)
1749 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1758 static void nvme_rdma_complete_rq(struct request *rq)
1760 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1762 nvme_rdma_unmap_data(req->queue, rq);
1763 nvme_complete_rq(rq);
1766 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1768 struct nvme_rdma_ctrl *ctrl = set->driver_data;
1770 return blk_mq_rdma_map_queues(set, ctrl->device->dev, 0);
1773 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1774 .queue_rq = nvme_rdma_queue_rq,
1775 .complete = nvme_rdma_complete_rq,
1776 .init_request = nvme_rdma_init_request,
1777 .exit_request = nvme_rdma_exit_request,
1778 .init_hctx = nvme_rdma_init_hctx,
1779 .poll = nvme_rdma_poll,
1780 .timeout = nvme_rdma_timeout,
1781 .map_queues = nvme_rdma_map_queues,
1784 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1785 .queue_rq = nvme_rdma_queue_rq,
1786 .complete = nvme_rdma_complete_rq,
1787 .init_request = nvme_rdma_init_request,
1788 .exit_request = nvme_rdma_exit_request,
1789 .init_hctx = nvme_rdma_init_admin_hctx,
1790 .timeout = nvme_rdma_timeout,
1793 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1795 if (ctrl->ctrl.queue_count > 1) {
1796 nvme_stop_queues(&ctrl->ctrl);
1797 nvme_rdma_stop_io_queues(ctrl);
1798 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1799 nvme_cancel_request, &ctrl->ctrl);
1801 nvme_start_queues(&ctrl->ctrl);
1802 nvme_rdma_destroy_io_queues(ctrl, shutdown);
1806 nvme_shutdown_ctrl(&ctrl->ctrl);
1808 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1810 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1811 nvme_rdma_stop_queue(&ctrl->queues[0]);
1812 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1813 nvme_cancel_request, &ctrl->ctrl);
1814 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1815 nvme_rdma_destroy_admin_queue(ctrl, shutdown);
1818 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1820 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1823 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1825 struct nvme_rdma_ctrl *ctrl =
1826 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1828 nvme_stop_ctrl(&ctrl->ctrl);
1829 nvme_rdma_shutdown_ctrl(ctrl, false);
1831 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1832 /* state change failure should never happen */
1837 if (nvme_rdma_setup_ctrl(ctrl, false))
1843 ++ctrl->ctrl.nr_reconnects;
1844 nvme_rdma_reconnect_or_remove(ctrl);
1847 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1849 .module = THIS_MODULE,
1850 .flags = NVME_F_FABRICS,
1851 .reg_read32 = nvmf_reg_read32,
1852 .reg_read64 = nvmf_reg_read64,
1853 .reg_write32 = nvmf_reg_write32,
1854 .free_ctrl = nvme_rdma_free_ctrl,
1855 .submit_async_event = nvme_rdma_submit_async_event,
1856 .delete_ctrl = nvme_rdma_delete_ctrl,
1857 .get_address = nvmf_get_address,
1858 .stop_ctrl = nvme_rdma_stop_ctrl,
1862 __nvme_rdma_options_match(struct nvme_rdma_ctrl *ctrl,
1863 struct nvmf_ctrl_options *opts)
1865 char *stdport = __stringify(NVME_RDMA_IP_PORT);
1868 if (!nvmf_ctlr_matches_baseopts(&ctrl->ctrl, opts) ||
1869 strcmp(opts->traddr, ctrl->ctrl.opts->traddr))
1872 if (opts->mask & NVMF_OPT_TRSVCID &&
1873 ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1874 if (strcmp(opts->trsvcid, ctrl->ctrl.opts->trsvcid))
1876 } else if (opts->mask & NVMF_OPT_TRSVCID) {
1877 if (strcmp(opts->trsvcid, stdport))
1879 } else if (ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1880 if (strcmp(stdport, ctrl->ctrl.opts->trsvcid))
1883 /* else, it's a match as both have stdport. Fall to next checks */
1886 * checking the local address is rough. In most cases, one
1887 * is not specified and the host port is selected by the stack.
1889 * Assume no match if:
1890 * local address is specified and address is not the same
1891 * local address is not specified but remote is, or vice versa
1892 * (admin using specific host_traddr when it matters).
1894 if (opts->mask & NVMF_OPT_HOST_TRADDR &&
1895 ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR) {
1896 if (strcmp(opts->host_traddr, ctrl->ctrl.opts->host_traddr))
1898 } else if (opts->mask & NVMF_OPT_HOST_TRADDR ||
1899 ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
1902 * if neither controller had an host port specified, assume it's
1903 * a match as everything else matched.
1910 * Fails a connection request if it matches an existing controller
1911 * (association) with the same tuple:
1912 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1914 * if local address is not specified in the request, it will match an
1915 * existing controller with all the other parameters the same and no
1916 * local port address specified as well.
1918 * The ports don't need to be compared as they are intrinsically
1919 * already matched by the port pointers supplied.
1922 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1924 struct nvme_rdma_ctrl *ctrl;
1927 mutex_lock(&nvme_rdma_ctrl_mutex);
1928 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1929 found = __nvme_rdma_options_match(ctrl, opts);
1933 mutex_unlock(&nvme_rdma_ctrl_mutex);
1938 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1939 struct nvmf_ctrl_options *opts)
1941 struct nvme_rdma_ctrl *ctrl;
1946 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1948 return ERR_PTR(-ENOMEM);
1949 ctrl->ctrl.opts = opts;
1950 INIT_LIST_HEAD(&ctrl->list);
1952 if (opts->mask & NVMF_OPT_TRSVCID)
1953 port = opts->trsvcid;
1955 port = __stringify(NVME_RDMA_IP_PORT);
1957 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1958 opts->traddr, port, &ctrl->addr);
1960 pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1964 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1965 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1966 opts->host_traddr, NULL, &ctrl->src_addr);
1968 pr_err("malformed src address passed: %s\n",
1974 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1979 INIT_DELAYED_WORK(&ctrl->reconnect_work,
1980 nvme_rdma_reconnect_ctrl_work);
1981 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1982 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1984 ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1985 ctrl->ctrl.sqsize = opts->queue_size - 1;
1986 ctrl->ctrl.kato = opts->kato;
1989 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1994 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1995 0 /* no quirks, we're perfect! */);
1997 goto out_kfree_queues;
1999 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2000 WARN_ON_ONCE(!changed);
2002 ret = nvme_rdma_setup_ctrl(ctrl, true);
2004 goto out_uninit_ctrl;
2006 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2007 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2009 nvme_get_ctrl(&ctrl->ctrl);
2011 mutex_lock(&nvme_rdma_ctrl_mutex);
2012 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2013 mutex_unlock(&nvme_rdma_ctrl_mutex);
2018 nvme_uninit_ctrl(&ctrl->ctrl);
2019 nvme_put_ctrl(&ctrl->ctrl);
2022 return ERR_PTR(ret);
2024 kfree(ctrl->queues);
2027 return ERR_PTR(ret);
2030 static struct nvmf_transport_ops nvme_rdma_transport = {
2032 .module = THIS_MODULE,
2033 .required_opts = NVMF_OPT_TRADDR,
2034 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2035 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
2036 .create_ctrl = nvme_rdma_create_ctrl,
2039 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2041 struct nvme_rdma_ctrl *ctrl;
2042 struct nvme_rdma_device *ndev;
2045 mutex_lock(&device_list_mutex);
2046 list_for_each_entry(ndev, &device_list, entry) {
2047 if (ndev->dev == ib_device) {
2052 mutex_unlock(&device_list_mutex);
2057 /* Delete all controllers using this device */
2058 mutex_lock(&nvme_rdma_ctrl_mutex);
2059 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2060 if (ctrl->device->dev != ib_device)
2062 nvme_delete_ctrl(&ctrl->ctrl);
2064 mutex_unlock(&nvme_rdma_ctrl_mutex);
2066 flush_workqueue(nvme_delete_wq);
2069 static struct ib_client nvme_rdma_ib_client = {
2070 .name = "nvme_rdma",
2071 .remove = nvme_rdma_remove_one
2074 static int __init nvme_rdma_init_module(void)
2078 ret = ib_register_client(&nvme_rdma_ib_client);
2082 ret = nvmf_register_transport(&nvme_rdma_transport);
2084 goto err_unreg_client;
2089 ib_unregister_client(&nvme_rdma_ib_client);
2093 static void __exit nvme_rdma_cleanup_module(void)
2095 nvmf_unregister_transport(&nvme_rdma_transport);
2096 ib_unregister_client(&nvme_rdma_ib_client);
2099 module_init(nvme_rdma_init_module);
2100 module_exit(nvme_rdma_cleanup_module);
2102 MODULE_LICENSE("GPL v2");