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 <linux/err.h>
19 #include <linux/string.h>
20 #include <linux/atomic.h>
21 #include <linux/blk-mq.h>
22 #include <linux/blk-mq-rdma.h>
23 #include <linux/types.h>
24 #include <linux/list.h>
25 #include <linux/mutex.h>
26 #include <linux/scatterlist.h>
27 #include <linux/nvme.h>
28 #include <asm/unaligned.h>
30 #include <rdma/ib_verbs.h>
31 #include <rdma/rdma_cm.h>
32 #include <linux/nvme-rdma.h>
38 #define NVME_RDMA_CONNECT_TIMEOUT_MS 3000 /* 3 second */
40 #define NVME_RDMA_MAX_SEGMENTS 256
42 #define NVME_RDMA_MAX_INLINE_SEGMENTS 1
45 * We handle AEN commands ourselves and don't even let the
46 * block layer know about them.
48 #define NVME_RDMA_NR_AEN_COMMANDS 1
49 #define NVME_RDMA_AQ_BLKMQ_DEPTH \
50 (NVME_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS)
52 struct nvme_rdma_device {
53 struct ib_device *dev;
56 struct list_head entry;
65 struct nvme_rdma_queue;
66 struct nvme_rdma_request {
67 struct nvme_request req;
69 struct nvme_rdma_qe sqe;
70 union nvme_result result;
73 struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
77 struct ib_reg_wr reg_wr;
78 struct ib_cqe reg_cqe;
79 struct nvme_rdma_queue *queue;
80 struct sg_table sg_table;
81 struct scatterlist first_sgl[];
84 enum nvme_rdma_queue_flags {
86 NVME_RDMA_Q_DELETING = 1,
89 struct nvme_rdma_queue {
90 struct nvme_rdma_qe *rsp_ring;
92 size_t cmnd_capsule_len;
93 struct nvme_rdma_ctrl *ctrl;
94 struct nvme_rdma_device *device;
99 struct rdma_cm_id *cm_id;
101 struct completion cm_done;
104 struct nvme_rdma_ctrl {
105 /* read only in the hot path */
106 struct nvme_rdma_queue *queues;
108 /* other member variables */
109 struct blk_mq_tag_set tag_set;
110 struct work_struct delete_work;
111 struct work_struct err_work;
113 struct nvme_rdma_qe async_event_sqe;
115 struct delayed_work reconnect_work;
117 struct list_head list;
119 struct blk_mq_tag_set admin_tag_set;
120 struct nvme_rdma_device *device;
124 struct sockaddr_storage addr;
125 struct sockaddr_storage src_addr;
127 struct nvme_ctrl ctrl;
130 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
132 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
135 static LIST_HEAD(device_list);
136 static DEFINE_MUTEX(device_list_mutex);
138 static LIST_HEAD(nvme_rdma_ctrl_list);
139 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
142 * Disabling this option makes small I/O goes faster, but is fundamentally
143 * unsafe. With it turned off we will have to register a global rkey that
144 * allows read and write access to all physical memory.
146 static bool register_always = true;
147 module_param(register_always, bool, 0444);
148 MODULE_PARM_DESC(register_always,
149 "Use memory registration even for contiguous memory regions");
151 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
152 struct rdma_cm_event *event);
153 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
155 static const struct blk_mq_ops nvme_rdma_mq_ops;
156 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
158 /* XXX: really should move to a generic header sooner or later.. */
159 static inline void put_unaligned_le24(u32 val, u8 *p)
166 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
168 return queue - queue->ctrl->queues;
171 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
173 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
176 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
177 size_t capsule_size, enum dma_data_direction dir)
179 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
183 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
184 size_t capsule_size, enum dma_data_direction dir)
186 qe->data = kzalloc(capsule_size, GFP_KERNEL);
190 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
191 if (ib_dma_mapping_error(ibdev, qe->dma)) {
199 static void nvme_rdma_free_ring(struct ib_device *ibdev,
200 struct nvme_rdma_qe *ring, size_t ib_queue_size,
201 size_t capsule_size, enum dma_data_direction dir)
205 for (i = 0; i < ib_queue_size; i++)
206 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
210 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
211 size_t ib_queue_size, size_t capsule_size,
212 enum dma_data_direction dir)
214 struct nvme_rdma_qe *ring;
217 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
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)
242 wait_for_completion_interruptible_timeout(&queue->cm_done,
243 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
244 return queue->cm_error;
247 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
249 struct nvme_rdma_device *dev = queue->device;
250 struct ib_qp_init_attr init_attr;
253 memset(&init_attr, 0, sizeof(init_attr));
254 init_attr.event_handler = nvme_rdma_qp_event;
256 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
258 init_attr.cap.max_recv_wr = queue->queue_size + 1;
259 init_attr.cap.max_recv_sge = 1;
260 init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
261 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
262 init_attr.qp_type = IB_QPT_RC;
263 init_attr.send_cq = queue->ib_cq;
264 init_attr.recv_cq = queue->ib_cq;
266 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
268 queue->qp = queue->cm_id->qp;
272 static int nvme_rdma_reinit_request(void *data, struct request *rq)
274 struct nvme_rdma_ctrl *ctrl = data;
275 struct nvme_rdma_device *dev = ctrl->device;
276 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
279 ib_dereg_mr(req->mr);
281 req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
283 if (IS_ERR(req->mr)) {
284 ret = PTR_ERR(req->mr);
289 req->mr->need_inval = false;
295 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
296 struct request *rq, unsigned int hctx_idx)
298 struct nvme_rdma_ctrl *ctrl = set->driver_data;
299 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
300 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
301 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
302 struct nvme_rdma_device *dev = queue->device;
305 ib_dereg_mr(req->mr);
307 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
311 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
312 struct request *rq, unsigned int hctx_idx,
313 unsigned int numa_node)
315 struct nvme_rdma_ctrl *ctrl = set->driver_data;
316 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
317 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
318 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
319 struct nvme_rdma_device *dev = queue->device;
320 struct ib_device *ibdev = dev->dev;
323 ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
328 req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
330 if (IS_ERR(req->mr)) {
331 ret = PTR_ERR(req->mr);
340 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
345 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
346 unsigned int hctx_idx)
348 struct nvme_rdma_ctrl *ctrl = data;
349 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
351 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
353 hctx->driver_data = queue;
357 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
358 unsigned int hctx_idx)
360 struct nvme_rdma_ctrl *ctrl = data;
361 struct nvme_rdma_queue *queue = &ctrl->queues[0];
363 BUG_ON(hctx_idx != 0);
365 hctx->driver_data = queue;
369 static void nvme_rdma_free_dev(struct kref *ref)
371 struct nvme_rdma_device *ndev =
372 container_of(ref, struct nvme_rdma_device, ref);
374 mutex_lock(&device_list_mutex);
375 list_del(&ndev->entry);
376 mutex_unlock(&device_list_mutex);
378 ib_dealloc_pd(ndev->pd);
382 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
384 kref_put(&dev->ref, nvme_rdma_free_dev);
387 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
389 return kref_get_unless_zero(&dev->ref);
392 static struct nvme_rdma_device *
393 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
395 struct nvme_rdma_device *ndev;
397 mutex_lock(&device_list_mutex);
398 list_for_each_entry(ndev, &device_list, entry) {
399 if (ndev->dev->node_guid == cm_id->device->node_guid &&
400 nvme_rdma_dev_get(ndev))
404 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
408 ndev->dev = cm_id->device;
409 kref_init(&ndev->ref);
411 ndev->pd = ib_alloc_pd(ndev->dev,
412 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
413 if (IS_ERR(ndev->pd))
416 if (!(ndev->dev->attrs.device_cap_flags &
417 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
418 dev_err(&ndev->dev->dev,
419 "Memory registrations not supported.\n");
423 list_add(&ndev->entry, &device_list);
425 mutex_unlock(&device_list_mutex);
429 ib_dealloc_pd(ndev->pd);
433 mutex_unlock(&device_list_mutex);
437 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
439 struct nvme_rdma_device *dev;
440 struct ib_device *ibdev;
444 rdma_destroy_qp(queue->cm_id);
445 ib_free_cq(queue->ib_cq);
447 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
448 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
450 nvme_rdma_dev_put(dev);
453 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
455 struct ib_device *ibdev;
456 const int send_wr_factor = 3; /* MR, SEND, INV */
457 const int cq_factor = send_wr_factor + 1; /* + RECV */
458 int comp_vector, idx = nvme_rdma_queue_idx(queue);
461 queue->device = nvme_rdma_find_get_device(queue->cm_id);
462 if (!queue->device) {
463 dev_err(queue->cm_id->device->dev.parent,
464 "no client data found!\n");
465 return -ECONNREFUSED;
467 ibdev = queue->device->dev;
470 * Spread I/O queues completion vectors according their queue index.
471 * Admin queues can always go on completion vector 0.
473 comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors;
475 /* +1 for ib_stop_cq */
476 queue->ib_cq = ib_alloc_cq(ibdev, queue,
477 cq_factor * queue->queue_size + 1,
478 comp_vector, IB_POLL_SOFTIRQ);
479 if (IS_ERR(queue->ib_cq)) {
480 ret = PTR_ERR(queue->ib_cq);
484 ret = nvme_rdma_create_qp(queue, send_wr_factor);
486 goto out_destroy_ib_cq;
488 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
489 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
490 if (!queue->rsp_ring) {
498 ib_destroy_qp(queue->qp);
500 ib_free_cq(queue->ib_cq);
502 nvme_rdma_dev_put(queue->device);
506 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
507 int idx, size_t queue_size)
509 struct nvme_rdma_queue *queue;
510 struct sockaddr *src_addr = NULL;
513 queue = &ctrl->queues[idx];
515 init_completion(&queue->cm_done);
518 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
520 queue->cmnd_capsule_len = sizeof(struct nvme_command);
522 queue->queue_size = queue_size;
524 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
525 RDMA_PS_TCP, IB_QPT_RC);
526 if (IS_ERR(queue->cm_id)) {
527 dev_info(ctrl->ctrl.device,
528 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
529 return PTR_ERR(queue->cm_id);
532 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
533 src_addr = (struct sockaddr *)&ctrl->src_addr;
535 queue->cm_error = -ETIMEDOUT;
536 ret = rdma_resolve_addr(queue->cm_id, src_addr,
537 (struct sockaddr *)&ctrl->addr,
538 NVME_RDMA_CONNECT_TIMEOUT_MS);
540 dev_info(ctrl->ctrl.device,
541 "rdma_resolve_addr failed (%d).\n", ret);
542 goto out_destroy_cm_id;
545 ret = nvme_rdma_wait_for_cm(queue);
547 dev_info(ctrl->ctrl.device,
548 "rdma_resolve_addr wait failed (%d).\n", ret);
549 goto out_destroy_cm_id;
552 clear_bit(NVME_RDMA_Q_DELETING, &queue->flags);
557 rdma_destroy_id(queue->cm_id);
561 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
563 if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
566 rdma_disconnect(queue->cm_id);
567 ib_drain_qp(queue->qp);
570 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
572 if (test_and_set_bit(NVME_RDMA_Q_DELETING, &queue->flags))
575 if (nvme_rdma_queue_idx(queue) == 0) {
576 nvme_rdma_free_qe(queue->device->dev,
577 &queue->ctrl->async_event_sqe,
578 sizeof(struct nvme_command), DMA_TO_DEVICE);
581 nvme_rdma_destroy_queue_ib(queue);
582 rdma_destroy_id(queue->cm_id);
585 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
589 for (i = 1; i < ctrl->ctrl.queue_count; i++)
590 nvme_rdma_free_queue(&ctrl->queues[i]);
593 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
597 for (i = 1; i < ctrl->ctrl.queue_count; i++)
598 nvme_rdma_stop_queue(&ctrl->queues[i]);
601 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
606 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
608 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
611 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags);
613 dev_info(ctrl->ctrl.device,
614 "failed to connect queue: %d ret=%d\n", idx, ret);
618 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
622 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
623 ret = nvme_rdma_start_queue(ctrl, i);
625 goto out_stop_queues;
631 for (i--; i >= 1; i--)
632 nvme_rdma_stop_queue(&ctrl->queues[i]);
636 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
638 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
639 struct ib_device *ibdev = ctrl->device->dev;
640 unsigned int nr_io_queues;
643 nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
646 * we map queues according to the device irq vectors for
647 * optimal locality so we don't need more queues than
648 * completion vectors.
650 nr_io_queues = min_t(unsigned int, nr_io_queues,
651 ibdev->num_comp_vectors);
653 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
657 if (nr_io_queues == 0) {
658 dev_err(ctrl->ctrl.device,
659 "unable to set any I/O queues\n");
663 ctrl->ctrl.queue_count = nr_io_queues + 1;
664 dev_info(ctrl->ctrl.device,
665 "creating %d I/O queues.\n", nr_io_queues);
667 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
668 ret = nvme_rdma_alloc_queue(ctrl, i,
669 ctrl->ctrl.sqsize + 1);
671 goto out_free_queues;
677 for (i--; i >= 1; i--)
678 nvme_rdma_free_queue(&ctrl->queues[i]);
683 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl, bool admin)
685 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
686 struct blk_mq_tag_set *set = admin ?
687 &ctrl->admin_tag_set : &ctrl->tag_set;
689 blk_mq_free_tag_set(set);
690 nvme_rdma_dev_put(ctrl->device);
693 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
696 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
697 struct blk_mq_tag_set *set;
701 set = &ctrl->admin_tag_set;
702 memset(set, 0, sizeof(*set));
703 set->ops = &nvme_rdma_admin_mq_ops;
704 set->queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
705 set->reserved_tags = 2; /* connect + keep-alive */
706 set->numa_node = NUMA_NO_NODE;
707 set->cmd_size = sizeof(struct nvme_rdma_request) +
708 SG_CHUNK_SIZE * sizeof(struct scatterlist);
709 set->driver_data = ctrl;
710 set->nr_hw_queues = 1;
711 set->timeout = ADMIN_TIMEOUT;
713 set = &ctrl->tag_set;
714 memset(set, 0, sizeof(*set));
715 set->ops = &nvme_rdma_mq_ops;
716 set->queue_depth = nctrl->opts->queue_size;
717 set->reserved_tags = 1; /* fabric connect */
718 set->numa_node = NUMA_NO_NODE;
719 set->flags = BLK_MQ_F_SHOULD_MERGE;
720 set->cmd_size = sizeof(struct nvme_rdma_request) +
721 SG_CHUNK_SIZE * sizeof(struct scatterlist);
722 set->driver_data = ctrl;
723 set->nr_hw_queues = nctrl->queue_count - 1;
724 set->timeout = NVME_IO_TIMEOUT;
727 ret = blk_mq_alloc_tag_set(set);
732 * We need a reference on the device as long as the tag_set is alive,
733 * as the MRs in the request structures need a valid ib_device.
735 ret = nvme_rdma_dev_get(ctrl->device);
738 goto out_free_tagset;
744 blk_mq_free_tag_set(set);
749 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
752 nvme_rdma_stop_queue(&ctrl->queues[0]);
754 blk_cleanup_queue(ctrl->ctrl.admin_q);
755 nvme_rdma_free_tagset(&ctrl->ctrl, true);
757 nvme_rdma_free_queue(&ctrl->queues[0]);
760 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
765 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
769 ctrl->device = ctrl->queues[0].device;
771 ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
772 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
775 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
776 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
777 error = PTR_ERR(ctrl->ctrl.admin_tagset);
781 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
782 if (IS_ERR(ctrl->ctrl.admin_q)) {
783 error = PTR_ERR(ctrl->ctrl.admin_q);
784 goto out_free_tagset;
787 error = blk_mq_reinit_tagset(&ctrl->admin_tag_set,
788 nvme_rdma_reinit_request);
793 error = nvme_rdma_start_queue(ctrl, 0);
795 goto out_cleanup_queue;
797 error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
800 dev_err(ctrl->ctrl.device,
801 "prop_get NVME_REG_CAP failed\n");
806 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
808 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
812 ctrl->ctrl.max_hw_sectors =
813 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
815 error = nvme_init_identify(&ctrl->ctrl);
819 error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
820 &ctrl->async_event_sqe, sizeof(struct nvme_command),
828 nvme_rdma_stop_queue(&ctrl->queues[0]);
831 blk_cleanup_queue(ctrl->ctrl.admin_q);
834 nvme_rdma_free_tagset(&ctrl->ctrl, true);
836 nvme_rdma_free_queue(&ctrl->queues[0]);
840 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
843 nvme_rdma_stop_io_queues(ctrl);
845 blk_cleanup_queue(ctrl->ctrl.connect_q);
846 nvme_rdma_free_tagset(&ctrl->ctrl, false);
848 nvme_rdma_free_io_queues(ctrl);
851 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
855 ret = nvme_rdma_alloc_io_queues(ctrl);
860 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
861 if (IS_ERR(ctrl->ctrl.tagset)) {
862 ret = PTR_ERR(ctrl->ctrl.tagset);
863 goto out_free_io_queues;
866 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
867 if (IS_ERR(ctrl->ctrl.connect_q)) {
868 ret = PTR_ERR(ctrl->ctrl.connect_q);
869 goto out_free_tag_set;
872 ret = blk_mq_reinit_tagset(&ctrl->tag_set,
873 nvme_rdma_reinit_request);
875 goto out_free_io_queues;
877 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
878 ctrl->ctrl.queue_count - 1);
881 ret = nvme_rdma_start_io_queues(ctrl);
883 goto out_cleanup_connect_q;
887 out_cleanup_connect_q:
889 blk_cleanup_queue(ctrl->ctrl.connect_q);
892 nvme_rdma_free_tagset(&ctrl->ctrl, false);
894 nvme_rdma_free_io_queues(ctrl);
898 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
900 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
902 if (list_empty(&ctrl->list))
905 mutex_lock(&nvme_rdma_ctrl_mutex);
906 list_del(&ctrl->list);
907 mutex_unlock(&nvme_rdma_ctrl_mutex);
910 nvmf_free_options(nctrl->opts);
915 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
917 /* If we are resetting/deleting then do nothing */
918 if (ctrl->ctrl.state != NVME_CTRL_RECONNECTING) {
919 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
920 ctrl->ctrl.state == NVME_CTRL_LIVE);
924 if (nvmf_should_reconnect(&ctrl->ctrl)) {
925 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
926 ctrl->ctrl.opts->reconnect_delay);
927 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
928 ctrl->ctrl.opts->reconnect_delay * HZ);
930 dev_info(ctrl->ctrl.device, "Removing controller...\n");
931 queue_work(nvme_wq, &ctrl->delete_work);
935 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
937 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
938 struct nvme_rdma_ctrl, reconnect_work);
942 ++ctrl->ctrl.nr_reconnects;
944 if (ctrl->ctrl.queue_count > 1)
945 nvme_rdma_destroy_io_queues(ctrl, false);
947 nvme_rdma_destroy_admin_queue(ctrl, false);
948 ret = nvme_rdma_configure_admin_queue(ctrl, false);
952 if (ctrl->ctrl.queue_count > 1) {
953 ret = nvme_rdma_configure_io_queues(ctrl, false);
958 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
960 /* state change failure is ok if we're in DELETING state */
961 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
965 ctrl->ctrl.nr_reconnects = 0;
967 nvme_start_ctrl(&ctrl->ctrl);
969 dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
974 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
975 ctrl->ctrl.nr_reconnects);
976 nvme_rdma_reconnect_or_remove(ctrl);
979 static void nvme_rdma_error_recovery_work(struct work_struct *work)
981 struct nvme_rdma_ctrl *ctrl = container_of(work,
982 struct nvme_rdma_ctrl, err_work);
984 nvme_stop_keep_alive(&ctrl->ctrl);
986 if (ctrl->ctrl.queue_count > 1) {
987 nvme_stop_queues(&ctrl->ctrl);
988 nvme_rdma_stop_io_queues(ctrl);
990 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
991 nvme_rdma_stop_queue(&ctrl->queues[0]);
993 /* We must take care of fastfail/requeue all our inflight requests */
994 if (ctrl->ctrl.queue_count > 1)
995 blk_mq_tagset_busy_iter(&ctrl->tag_set,
996 nvme_cancel_request, &ctrl->ctrl);
997 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
998 nvme_cancel_request, &ctrl->ctrl);
1001 * queues are not a live anymore, so restart the queues to fail fast
1004 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1005 nvme_start_queues(&ctrl->ctrl);
1007 nvme_rdma_reconnect_or_remove(ctrl);
1010 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1012 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
1015 queue_work(nvme_wq, &ctrl->err_work);
1018 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1021 struct nvme_rdma_queue *queue = cq->cq_context;
1022 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1024 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1025 dev_info(ctrl->ctrl.device,
1026 "%s for CQE 0x%p failed with status %s (%d)\n",
1028 ib_wc_status_msg(wc->status), wc->status);
1029 nvme_rdma_error_recovery(ctrl);
1032 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1034 if (unlikely(wc->status != IB_WC_SUCCESS))
1035 nvme_rdma_wr_error(cq, wc, "MEMREG");
1038 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1040 if (unlikely(wc->status != IB_WC_SUCCESS))
1041 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1044 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1045 struct nvme_rdma_request *req)
1047 struct ib_send_wr *bad_wr;
1048 struct ib_send_wr wr = {
1049 .opcode = IB_WR_LOCAL_INV,
1053 .ex.invalidate_rkey = req->mr->rkey,
1056 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1057 wr.wr_cqe = &req->reg_cqe;
1059 return ib_post_send(queue->qp, &wr, &bad_wr);
1062 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1065 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1066 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1067 struct nvme_rdma_device *dev = queue->device;
1068 struct ib_device *ibdev = dev->dev;
1071 if (!blk_rq_bytes(rq))
1074 if (req->mr->need_inval) {
1075 res = nvme_rdma_inv_rkey(queue, req);
1076 if (unlikely(res < 0)) {
1077 dev_err(ctrl->ctrl.device,
1078 "Queueing INV WR for rkey %#x failed (%d)\n",
1079 req->mr->rkey, res);
1080 nvme_rdma_error_recovery(queue->ctrl);
1084 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1085 req->nents, rq_data_dir(rq) ==
1086 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1088 nvme_cleanup_cmd(rq);
1089 sg_free_table_chained(&req->sg_table, true);
1092 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1094 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1097 put_unaligned_le24(0, sg->length);
1098 put_unaligned_le32(0, sg->key);
1099 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1103 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1104 struct nvme_rdma_request *req, struct nvme_command *c)
1106 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1108 req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
1109 req->sge[1].length = sg_dma_len(req->sg_table.sgl);
1110 req->sge[1].lkey = queue->device->pd->local_dma_lkey;
1112 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1113 sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
1114 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1116 req->inline_data = true;
1121 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1122 struct nvme_rdma_request *req, struct nvme_command *c)
1124 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1126 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1127 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1128 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1129 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1133 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1134 struct nvme_rdma_request *req, struct nvme_command *c,
1137 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1141 * Align the MR to a 4K page size to match the ctrl page size and
1142 * the block virtual boundary.
1144 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1145 if (unlikely(nr < count)) {
1151 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1153 req->reg_cqe.done = nvme_rdma_memreg_done;
1154 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1155 req->reg_wr.wr.opcode = IB_WR_REG_MR;
1156 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1157 req->reg_wr.wr.num_sge = 0;
1158 req->reg_wr.mr = req->mr;
1159 req->reg_wr.key = req->mr->rkey;
1160 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1161 IB_ACCESS_REMOTE_READ |
1162 IB_ACCESS_REMOTE_WRITE;
1164 req->mr->need_inval = true;
1166 sg->addr = cpu_to_le64(req->mr->iova);
1167 put_unaligned_le24(req->mr->length, sg->length);
1168 put_unaligned_le32(req->mr->rkey, sg->key);
1169 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1170 NVME_SGL_FMT_INVALIDATE;
1175 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1176 struct request *rq, struct nvme_command *c)
1178 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1179 struct nvme_rdma_device *dev = queue->device;
1180 struct ib_device *ibdev = dev->dev;
1184 req->inline_data = false;
1185 req->mr->need_inval = false;
1186 refcount_set(&req->ref, 2); /* send and recv completions */
1188 c->common.flags |= NVME_CMD_SGL_METABUF;
1190 if (!blk_rq_bytes(rq))
1191 return nvme_rdma_set_sg_null(c);
1193 req->sg_table.sgl = req->first_sgl;
1194 ret = sg_alloc_table_chained(&req->sg_table,
1195 blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1199 req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1201 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1202 rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1203 if (unlikely(count <= 0)) {
1204 sg_free_table_chained(&req->sg_table, true);
1209 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1210 blk_rq_payload_bytes(rq) <=
1211 nvme_rdma_inline_data_size(queue))
1212 return nvme_rdma_map_sg_inline(queue, req, c);
1214 if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
1215 return nvme_rdma_map_sg_single(queue, req, c);
1218 return nvme_rdma_map_sg_fr(queue, req, c, count);
1221 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1223 struct nvme_rdma_qe *qe =
1224 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1225 struct nvme_rdma_request *req =
1226 container_of(qe, struct nvme_rdma_request, sqe);
1227 struct request *rq = blk_mq_rq_from_pdu(req);
1229 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1230 nvme_rdma_wr_error(cq, wc, "SEND");
1234 if (refcount_dec_and_test(&req->ref))
1235 nvme_end_request(rq, req->status, req->result);
1238 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1239 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1240 struct ib_send_wr *first)
1242 struct ib_send_wr wr, *bad_wr;
1245 sge->addr = qe->dma;
1246 sge->length = sizeof(struct nvme_command),
1247 sge->lkey = queue->device->pd->local_dma_lkey;
1250 wr.wr_cqe = &qe->cqe;
1252 wr.num_sge = num_sge;
1253 wr.opcode = IB_WR_SEND;
1254 wr.send_flags = IB_SEND_SIGNALED;
1261 ret = ib_post_send(queue->qp, first, &bad_wr);
1262 if (unlikely(ret)) {
1263 dev_err(queue->ctrl->ctrl.device,
1264 "%s failed with error code %d\n", __func__, ret);
1269 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1270 struct nvme_rdma_qe *qe)
1272 struct ib_recv_wr wr, *bad_wr;
1276 list.addr = qe->dma;
1277 list.length = sizeof(struct nvme_completion);
1278 list.lkey = queue->device->pd->local_dma_lkey;
1280 qe->cqe.done = nvme_rdma_recv_done;
1283 wr.wr_cqe = &qe->cqe;
1287 ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1288 if (unlikely(ret)) {
1289 dev_err(queue->ctrl->ctrl.device,
1290 "%s failed with error code %d\n", __func__, ret);
1295 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1297 u32 queue_idx = nvme_rdma_queue_idx(queue);
1300 return queue->ctrl->admin_tag_set.tags[queue_idx];
1301 return queue->ctrl->tag_set.tags[queue_idx - 1];
1304 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1306 if (unlikely(wc->status != IB_WC_SUCCESS))
1307 nvme_rdma_wr_error(cq, wc, "ASYNC");
1310 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1312 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1313 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1314 struct ib_device *dev = queue->device->dev;
1315 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1316 struct nvme_command *cmd = sqe->data;
1320 if (WARN_ON_ONCE(aer_idx != 0))
1323 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1325 memset(cmd, 0, sizeof(*cmd));
1326 cmd->common.opcode = nvme_admin_async_event;
1327 cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1328 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1329 nvme_rdma_set_sg_null(cmd);
1331 sqe->cqe.done = nvme_rdma_async_done;
1333 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1336 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1340 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1341 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1344 struct nvme_rdma_request *req;
1347 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1349 dev_err(queue->ctrl->ctrl.device,
1350 "tag 0x%x on QP %#x not found\n",
1351 cqe->command_id, queue->qp->qp_num);
1352 nvme_rdma_error_recovery(queue->ctrl);
1355 req = blk_mq_rq_to_pdu(rq);
1357 req->status = cqe->status;
1358 req->result = cqe->result;
1360 if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1361 wc->ex.invalidate_rkey == req->mr->rkey)
1362 req->mr->need_inval = false;
1364 if (refcount_dec_and_test(&req->ref)) {
1367 nvme_end_request(rq, req->status, req->result);
1373 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1375 struct nvme_rdma_qe *qe =
1376 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1377 struct nvme_rdma_queue *queue = cq->cq_context;
1378 struct ib_device *ibdev = queue->device->dev;
1379 struct nvme_completion *cqe = qe->data;
1380 const size_t len = sizeof(struct nvme_completion);
1383 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1384 nvme_rdma_wr_error(cq, wc, "RECV");
1388 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1390 * AEN requests are special as they don't time out and can
1391 * survive any kind of queue freeze and often don't respond to
1392 * aborts. We don't even bother to allocate a struct request
1393 * for them but rather special case them here.
1395 if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1396 cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
1397 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1400 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1401 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1403 nvme_rdma_post_recv(queue, qe);
1407 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1409 __nvme_rdma_recv_done(cq, wc, -1);
1412 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1416 for (i = 0; i < queue->queue_size; i++) {
1417 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1419 goto out_destroy_queue_ib;
1424 out_destroy_queue_ib:
1425 nvme_rdma_destroy_queue_ib(queue);
1429 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1430 struct rdma_cm_event *ev)
1432 struct rdma_cm_id *cm_id = queue->cm_id;
1433 int status = ev->status;
1434 const char *rej_msg;
1435 const struct nvme_rdma_cm_rej *rej_data;
1438 rej_msg = rdma_reject_msg(cm_id, status);
1439 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1441 if (rej_data && rej_data_len >= sizeof(u16)) {
1442 u16 sts = le16_to_cpu(rej_data->sts);
1444 dev_err(queue->ctrl->ctrl.device,
1445 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1446 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1448 dev_err(queue->ctrl->ctrl.device,
1449 "Connect rejected: status %d (%s).\n", status, rej_msg);
1455 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1459 ret = nvme_rdma_create_queue_ib(queue);
1463 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1465 dev_err(queue->ctrl->ctrl.device,
1466 "rdma_resolve_route failed (%d).\n",
1468 goto out_destroy_queue;
1474 nvme_rdma_destroy_queue_ib(queue);
1478 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1480 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1481 struct rdma_conn_param param = { };
1482 struct nvme_rdma_cm_req priv = { };
1485 param.qp_num = queue->qp->qp_num;
1486 param.flow_control = 1;
1488 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1489 /* maximum retry count */
1490 param.retry_count = 7;
1491 param.rnr_retry_count = 7;
1492 param.private_data = &priv;
1493 param.private_data_len = sizeof(priv);
1495 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1496 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1498 * set the admin queue depth to the minimum size
1499 * specified by the Fabrics standard.
1501 if (priv.qid == 0) {
1502 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1503 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1506 * current interpretation of the fabrics spec
1507 * is at minimum you make hrqsize sqsize+1, or a
1508 * 1's based representation of sqsize.
1510 priv.hrqsize = cpu_to_le16(queue->queue_size);
1511 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1514 ret = rdma_connect(queue->cm_id, ¶m);
1516 dev_err(ctrl->ctrl.device,
1517 "rdma_connect failed (%d).\n", ret);
1518 goto out_destroy_queue_ib;
1523 out_destroy_queue_ib:
1524 nvme_rdma_destroy_queue_ib(queue);
1528 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1529 struct rdma_cm_event *ev)
1531 struct nvme_rdma_queue *queue = cm_id->context;
1534 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1535 rdma_event_msg(ev->event), ev->event,
1538 switch (ev->event) {
1539 case RDMA_CM_EVENT_ADDR_RESOLVED:
1540 cm_error = nvme_rdma_addr_resolved(queue);
1542 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1543 cm_error = nvme_rdma_route_resolved(queue);
1545 case RDMA_CM_EVENT_ESTABLISHED:
1546 queue->cm_error = nvme_rdma_conn_established(queue);
1547 /* complete cm_done regardless of success/failure */
1548 complete(&queue->cm_done);
1550 case RDMA_CM_EVENT_REJECTED:
1551 cm_error = nvme_rdma_conn_rejected(queue, ev);
1553 case RDMA_CM_EVENT_ROUTE_ERROR:
1554 case RDMA_CM_EVENT_CONNECT_ERROR:
1555 case RDMA_CM_EVENT_UNREACHABLE:
1556 nvme_rdma_destroy_queue_ib(queue);
1557 case RDMA_CM_EVENT_ADDR_ERROR:
1558 dev_dbg(queue->ctrl->ctrl.device,
1559 "CM error event %d\n", ev->event);
1560 cm_error = -ECONNRESET;
1562 case RDMA_CM_EVENT_DISCONNECTED:
1563 case RDMA_CM_EVENT_ADDR_CHANGE:
1564 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1565 dev_dbg(queue->ctrl->ctrl.device,
1566 "disconnect received - connection closed\n");
1567 nvme_rdma_error_recovery(queue->ctrl);
1569 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1570 /* device removal is handled via the ib_client API */
1573 dev_err(queue->ctrl->ctrl.device,
1574 "Unexpected RDMA CM event (%d)\n", ev->event);
1575 nvme_rdma_error_recovery(queue->ctrl);
1580 queue->cm_error = cm_error;
1581 complete(&queue->cm_done);
1587 static enum blk_eh_timer_return
1588 nvme_rdma_timeout(struct request *rq, bool reserved)
1590 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1592 /* queue error recovery */
1593 nvme_rdma_error_recovery(req->queue->ctrl);
1595 /* fail with DNR on cmd timeout */
1596 nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1598 return BLK_EH_HANDLED;
1602 * We cannot accept any other command until the Connect command has completed.
1604 static inline blk_status_t
1605 nvme_rdma_is_ready(struct nvme_rdma_queue *queue, struct request *rq)
1607 if (unlikely(!test_bit(NVME_RDMA_Q_LIVE, &queue->flags)))
1608 return nvmf_check_init_req(&queue->ctrl->ctrl, rq);
1612 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1613 const struct blk_mq_queue_data *bd)
1615 struct nvme_ns *ns = hctx->queue->queuedata;
1616 struct nvme_rdma_queue *queue = hctx->driver_data;
1617 struct request *rq = bd->rq;
1618 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1619 struct nvme_rdma_qe *sqe = &req->sqe;
1620 struct nvme_command *c = sqe->data;
1621 struct ib_device *dev;
1625 WARN_ON_ONCE(rq->tag < 0);
1627 ret = nvme_rdma_is_ready(queue, rq);
1631 dev = queue->device->dev;
1632 ib_dma_sync_single_for_cpu(dev, sqe->dma,
1633 sizeof(struct nvme_command), DMA_TO_DEVICE);
1635 ret = nvme_setup_cmd(ns, rq, c);
1639 blk_mq_start_request(rq);
1641 err = nvme_rdma_map_data(queue, rq, c);
1642 if (unlikely(err < 0)) {
1643 dev_err(queue->ctrl->ctrl.device,
1644 "Failed to map data (%d)\n", err);
1645 nvme_cleanup_cmd(rq);
1649 sqe->cqe.done = nvme_rdma_send_done;
1651 ib_dma_sync_single_for_device(dev, sqe->dma,
1652 sizeof(struct nvme_command), DMA_TO_DEVICE);
1654 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1655 req->mr->need_inval ? &req->reg_wr.wr : NULL);
1656 if (unlikely(err)) {
1657 nvme_rdma_unmap_data(queue, rq);
1663 if (err == -ENOMEM || err == -EAGAIN)
1664 return BLK_STS_RESOURCE;
1665 return BLK_STS_IOERR;
1668 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1670 struct nvme_rdma_queue *queue = hctx->driver_data;
1671 struct ib_cq *cq = queue->ib_cq;
1675 while (ib_poll_cq(cq, 1, &wc) > 0) {
1676 struct ib_cqe *cqe = wc.wr_cqe;
1679 if (cqe->done == nvme_rdma_recv_done)
1680 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1689 static void nvme_rdma_complete_rq(struct request *rq)
1691 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1693 nvme_rdma_unmap_data(req->queue, rq);
1694 nvme_complete_rq(rq);
1697 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1699 struct nvme_rdma_ctrl *ctrl = set->driver_data;
1701 return blk_mq_rdma_map_queues(set, ctrl->device->dev, 0);
1704 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1705 .queue_rq = nvme_rdma_queue_rq,
1706 .complete = nvme_rdma_complete_rq,
1707 .init_request = nvme_rdma_init_request,
1708 .exit_request = nvme_rdma_exit_request,
1709 .init_hctx = nvme_rdma_init_hctx,
1710 .poll = nvme_rdma_poll,
1711 .timeout = nvme_rdma_timeout,
1712 .map_queues = nvme_rdma_map_queues,
1715 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1716 .queue_rq = nvme_rdma_queue_rq,
1717 .complete = nvme_rdma_complete_rq,
1718 .init_request = nvme_rdma_init_request,
1719 .exit_request = nvme_rdma_exit_request,
1720 .init_hctx = nvme_rdma_init_admin_hctx,
1721 .timeout = nvme_rdma_timeout,
1724 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1726 cancel_work_sync(&ctrl->err_work);
1727 cancel_delayed_work_sync(&ctrl->reconnect_work);
1729 if (ctrl->ctrl.queue_count > 1) {
1730 nvme_stop_queues(&ctrl->ctrl);
1731 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1732 nvme_cancel_request, &ctrl->ctrl);
1734 nvme_start_queues(&ctrl->ctrl);
1735 nvme_rdma_destroy_io_queues(ctrl, shutdown);
1739 nvme_shutdown_ctrl(&ctrl->ctrl);
1741 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1743 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1744 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1745 nvme_cancel_request, &ctrl->ctrl);
1746 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1747 nvme_rdma_destroy_admin_queue(ctrl, shutdown);
1750 static void nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl)
1752 nvme_remove_namespaces(&ctrl->ctrl);
1753 nvme_rdma_shutdown_ctrl(ctrl, true);
1754 nvme_uninit_ctrl(&ctrl->ctrl);
1755 nvme_put_ctrl(&ctrl->ctrl);
1758 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1760 struct nvme_rdma_ctrl *ctrl = container_of(work,
1761 struct nvme_rdma_ctrl, delete_work);
1763 nvme_stop_ctrl(&ctrl->ctrl);
1764 nvme_rdma_remove_ctrl(ctrl);
1767 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1769 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1772 if (!queue_work(nvme_wq, &ctrl->delete_work))
1778 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1780 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1784 * Keep a reference until all work is flushed since
1785 * __nvme_rdma_del_ctrl can free the ctrl mem
1787 if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1789 ret = __nvme_rdma_del_ctrl(ctrl);
1791 flush_work(&ctrl->delete_work);
1792 nvme_put_ctrl(&ctrl->ctrl);
1796 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1798 struct nvme_rdma_ctrl *ctrl =
1799 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1803 nvme_stop_ctrl(&ctrl->ctrl);
1804 nvme_rdma_shutdown_ctrl(ctrl, false);
1806 ret = nvme_rdma_configure_admin_queue(ctrl, false);
1810 if (ctrl->ctrl.queue_count > 1) {
1811 ret = nvme_rdma_configure_io_queues(ctrl, false);
1816 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1817 WARN_ON_ONCE(!changed);
1819 nvme_start_ctrl(&ctrl->ctrl);
1824 dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1825 nvme_rdma_remove_ctrl(ctrl);
1828 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1830 .module = THIS_MODULE,
1831 .flags = NVME_F_FABRICS,
1832 .reg_read32 = nvmf_reg_read32,
1833 .reg_read64 = nvmf_reg_read64,
1834 .reg_write32 = nvmf_reg_write32,
1835 .free_ctrl = nvme_rdma_free_ctrl,
1836 .submit_async_event = nvme_rdma_submit_async_event,
1837 .delete_ctrl = nvme_rdma_del_ctrl,
1838 .get_address = nvmf_get_address,
1841 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1842 struct nvmf_ctrl_options *opts)
1844 struct nvme_rdma_ctrl *ctrl;
1849 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1851 return ERR_PTR(-ENOMEM);
1852 ctrl->ctrl.opts = opts;
1853 INIT_LIST_HEAD(&ctrl->list);
1855 if (opts->mask & NVMF_OPT_TRSVCID)
1856 port = opts->trsvcid;
1858 port = __stringify(NVME_RDMA_IP_PORT);
1860 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1861 opts->traddr, port, &ctrl->addr);
1863 pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1867 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1868 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1869 opts->host_traddr, NULL, &ctrl->src_addr);
1871 pr_err("malformed src address passed: %s\n",
1877 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1878 0 /* no quirks, we're perfect! */);
1882 INIT_DELAYED_WORK(&ctrl->reconnect_work,
1883 nvme_rdma_reconnect_ctrl_work);
1884 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1885 INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1886 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1888 ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1889 ctrl->ctrl.sqsize = opts->queue_size - 1;
1890 ctrl->ctrl.kato = opts->kato;
1893 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1896 goto out_uninit_ctrl;
1898 ret = nvme_rdma_configure_admin_queue(ctrl, true);
1900 goto out_kfree_queues;
1902 /* sanity check icdoff */
1903 if (ctrl->ctrl.icdoff) {
1904 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1906 goto out_remove_admin_queue;
1909 /* sanity check keyed sgls */
1910 if (!(ctrl->ctrl.sgls & (1 << 20))) {
1911 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1913 goto out_remove_admin_queue;
1916 if (opts->queue_size > ctrl->ctrl.maxcmd) {
1917 /* warn if maxcmd is lower than queue_size */
1918 dev_warn(ctrl->ctrl.device,
1919 "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1920 opts->queue_size, ctrl->ctrl.maxcmd);
1921 opts->queue_size = ctrl->ctrl.maxcmd;
1924 if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
1925 /* warn if sqsize is lower than queue_size */
1926 dev_warn(ctrl->ctrl.device,
1927 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1928 opts->queue_size, ctrl->ctrl.sqsize + 1);
1929 opts->queue_size = ctrl->ctrl.sqsize + 1;
1932 if (opts->nr_io_queues) {
1933 ret = nvme_rdma_configure_io_queues(ctrl, true);
1935 goto out_remove_admin_queue;
1938 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1939 WARN_ON_ONCE(!changed);
1941 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1942 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1944 kref_get(&ctrl->ctrl.kref);
1946 mutex_lock(&nvme_rdma_ctrl_mutex);
1947 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1948 mutex_unlock(&nvme_rdma_ctrl_mutex);
1950 nvme_start_ctrl(&ctrl->ctrl);
1954 out_remove_admin_queue:
1955 nvme_rdma_destroy_admin_queue(ctrl, true);
1957 kfree(ctrl->queues);
1959 nvme_uninit_ctrl(&ctrl->ctrl);
1960 nvme_put_ctrl(&ctrl->ctrl);
1963 return ERR_PTR(ret);
1966 return ERR_PTR(ret);
1969 static struct nvmf_transport_ops nvme_rdma_transport = {
1971 .required_opts = NVMF_OPT_TRADDR,
1972 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
1973 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
1974 .create_ctrl = nvme_rdma_create_ctrl,
1977 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1979 struct nvme_rdma_ctrl *ctrl;
1981 /* Delete all controllers using this device */
1982 mutex_lock(&nvme_rdma_ctrl_mutex);
1983 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1984 if (ctrl->device->dev != ib_device)
1986 dev_info(ctrl->ctrl.device,
1987 "Removing ctrl: NQN \"%s\", addr %pISp\n",
1988 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1989 __nvme_rdma_del_ctrl(ctrl);
1991 mutex_unlock(&nvme_rdma_ctrl_mutex);
1993 flush_workqueue(nvme_wq);
1996 static struct ib_client nvme_rdma_ib_client = {
1997 .name = "nvme_rdma",
1998 .remove = nvme_rdma_remove_one
2001 static int __init nvme_rdma_init_module(void)
2005 ret = ib_register_client(&nvme_rdma_ib_client);
2009 ret = nvmf_register_transport(&nvme_rdma_transport);
2011 goto err_unreg_client;
2016 ib_unregister_client(&nvme_rdma_ib_client);
2020 static void __exit nvme_rdma_cleanup_module(void)
2022 nvmf_unregister_transport(&nvme_rdma_transport);
2023 ib_unregister_client(&nvme_rdma_ib_client);
2026 module_init(nvme_rdma_init_module);
2027 module_exit(nvme_rdma_cleanup_module);
2029 MODULE_LICENSE("GPL v2");