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 struct mutex teardown_lock;
122 bool use_inline_data;
125 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
127 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
130 static LIST_HEAD(device_list);
131 static DEFINE_MUTEX(device_list_mutex);
133 static LIST_HEAD(nvme_rdma_ctrl_list);
134 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
137 * Disabling this option makes small I/O goes faster, but is fundamentally
138 * unsafe. With it turned off we will have to register a global rkey that
139 * allows read and write access to all physical memory.
141 static bool register_always = true;
142 module_param(register_always, bool, 0444);
143 MODULE_PARM_DESC(register_always,
144 "Use memory registration even for contiguous memory regions");
146 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
147 struct rdma_cm_event *event);
148 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
150 static const struct blk_mq_ops nvme_rdma_mq_ops;
151 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
153 /* XXX: really should move to a generic header sooner or later.. */
154 static inline void put_unaligned_le24(u32 val, u8 *p)
161 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
163 return queue - queue->ctrl->queues;
166 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
168 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
171 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
172 size_t capsule_size, enum dma_data_direction dir)
174 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
178 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
179 size_t capsule_size, enum dma_data_direction dir)
181 qe->data = kzalloc(capsule_size, GFP_KERNEL);
185 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
186 if (ib_dma_mapping_error(ibdev, qe->dma)) {
195 static void nvme_rdma_free_ring(struct ib_device *ibdev,
196 struct nvme_rdma_qe *ring, size_t ib_queue_size,
197 size_t capsule_size, enum dma_data_direction dir)
201 for (i = 0; i < ib_queue_size; i++)
202 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
206 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
207 size_t ib_queue_size, size_t capsule_size,
208 enum dma_data_direction dir)
210 struct nvme_rdma_qe *ring;
213 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
217 for (i = 0; i < ib_queue_size; i++) {
218 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
225 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
229 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
231 pr_debug("QP event %s (%d)\n",
232 ib_event_msg(event->event), event->event);
236 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
238 wait_for_completion_interruptible_timeout(&queue->cm_done,
239 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
240 return queue->cm_error;
243 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
245 struct nvme_rdma_device *dev = queue->device;
246 struct ib_qp_init_attr init_attr;
249 memset(&init_attr, 0, sizeof(init_attr));
250 init_attr.event_handler = nvme_rdma_qp_event;
252 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
254 init_attr.cap.max_recv_wr = queue->queue_size + 1;
255 init_attr.cap.max_recv_sge = 1;
256 init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
257 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
258 init_attr.qp_type = IB_QPT_RC;
259 init_attr.send_cq = queue->ib_cq;
260 init_attr.recv_cq = queue->ib_cq;
262 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
264 queue->qp = queue->cm_id->qp;
268 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
269 struct request *rq, unsigned int hctx_idx)
271 struct nvme_rdma_ctrl *ctrl = set->driver_data;
272 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
273 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
274 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
275 struct nvme_rdma_device *dev = queue->device;
277 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
281 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
282 struct request *rq, unsigned int hctx_idx,
283 unsigned int numa_node)
285 struct nvme_rdma_ctrl *ctrl = set->driver_data;
286 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
287 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
288 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
289 struct nvme_rdma_device *dev = queue->device;
290 struct ib_device *ibdev = dev->dev;
293 nvme_req(rq)->ctrl = &ctrl->ctrl;
294 ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
304 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
305 unsigned int hctx_idx)
307 struct nvme_rdma_ctrl *ctrl = data;
308 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
310 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
312 hctx->driver_data = queue;
316 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
317 unsigned int hctx_idx)
319 struct nvme_rdma_ctrl *ctrl = data;
320 struct nvme_rdma_queue *queue = &ctrl->queues[0];
322 BUG_ON(hctx_idx != 0);
324 hctx->driver_data = queue;
328 static void nvme_rdma_free_dev(struct kref *ref)
330 struct nvme_rdma_device *ndev =
331 container_of(ref, struct nvme_rdma_device, ref);
333 mutex_lock(&device_list_mutex);
334 list_del(&ndev->entry);
335 mutex_unlock(&device_list_mutex);
337 ib_dealloc_pd(ndev->pd);
341 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
343 kref_put(&dev->ref, nvme_rdma_free_dev);
346 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
348 return kref_get_unless_zero(&dev->ref);
351 static struct nvme_rdma_device *
352 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
354 struct nvme_rdma_device *ndev;
356 mutex_lock(&device_list_mutex);
357 list_for_each_entry(ndev, &device_list, entry) {
358 if (ndev->dev->node_guid == cm_id->device->node_guid &&
359 nvme_rdma_dev_get(ndev))
363 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
367 ndev->dev = cm_id->device;
368 kref_init(&ndev->ref);
370 ndev->pd = ib_alloc_pd(ndev->dev,
371 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
372 if (IS_ERR(ndev->pd))
375 if (!(ndev->dev->attrs.device_cap_flags &
376 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
377 dev_err(&ndev->dev->dev,
378 "Memory registrations not supported.\n");
382 ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
383 ndev->dev->attrs.max_send_sge - 1);
384 list_add(&ndev->entry, &device_list);
386 mutex_unlock(&device_list_mutex);
390 ib_dealloc_pd(ndev->pd);
394 mutex_unlock(&device_list_mutex);
398 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
400 struct nvme_rdma_device *dev;
401 struct ib_device *ibdev;
403 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
409 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
412 * The cm_id object might have been destroyed during RDMA connection
413 * establishment error flow to avoid getting other cma events, thus
414 * the destruction of the QP shouldn't use rdma_cm API.
416 ib_destroy_qp(queue->qp);
417 ib_free_cq(queue->ib_cq);
419 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
420 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
422 nvme_rdma_dev_put(dev);
425 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev)
427 return min_t(u32, NVME_RDMA_MAX_SEGMENTS,
428 ibdev->attrs.max_fast_reg_page_list_len);
431 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
433 struct ib_device *ibdev;
434 const int send_wr_factor = 3; /* MR, SEND, INV */
435 const int cq_factor = send_wr_factor + 1; /* + RECV */
436 int comp_vector, idx = nvme_rdma_queue_idx(queue);
439 queue->device = nvme_rdma_find_get_device(queue->cm_id);
440 if (!queue->device) {
441 dev_err(queue->cm_id->device->dev.parent,
442 "no client data found!\n");
443 return -ECONNREFUSED;
445 ibdev = queue->device->dev;
448 * Spread I/O queues completion vectors according their queue index.
449 * Admin queues can always go on completion vector 0.
451 comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors;
453 /* +1 for ib_stop_cq */
454 queue->ib_cq = ib_alloc_cq(ibdev, queue,
455 cq_factor * queue->queue_size + 1,
456 comp_vector, IB_POLL_SOFTIRQ);
457 if (IS_ERR(queue->ib_cq)) {
458 ret = PTR_ERR(queue->ib_cq);
462 ret = nvme_rdma_create_qp(queue, send_wr_factor);
464 goto out_destroy_ib_cq;
466 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
467 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
468 if (!queue->rsp_ring) {
473 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
476 nvme_rdma_get_max_fr_pages(ibdev));
478 dev_err(queue->ctrl->ctrl.device,
479 "failed to initialize MR pool sized %d for QID %d\n",
480 queue->queue_size, idx);
481 goto out_destroy_ring;
484 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
489 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
490 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
492 rdma_destroy_qp(queue->cm_id);
494 ib_free_cq(queue->ib_cq);
496 nvme_rdma_dev_put(queue->device);
500 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
501 int idx, size_t queue_size)
503 struct nvme_rdma_queue *queue;
504 struct sockaddr *src_addr = NULL;
507 queue = &ctrl->queues[idx];
509 init_completion(&queue->cm_done);
512 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
514 queue->cmnd_capsule_len = sizeof(struct nvme_command);
516 queue->queue_size = queue_size;
518 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
519 RDMA_PS_TCP, IB_QPT_RC);
520 if (IS_ERR(queue->cm_id)) {
521 dev_info(ctrl->ctrl.device,
522 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
523 return PTR_ERR(queue->cm_id);
526 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
527 src_addr = (struct sockaddr *)&ctrl->src_addr;
529 queue->cm_error = -ETIMEDOUT;
530 ret = rdma_resolve_addr(queue->cm_id, src_addr,
531 (struct sockaddr *)&ctrl->addr,
532 NVME_RDMA_CONNECT_TIMEOUT_MS);
534 dev_info(ctrl->ctrl.device,
535 "rdma_resolve_addr failed (%d).\n", ret);
536 goto out_destroy_cm_id;
539 ret = nvme_rdma_wait_for_cm(queue);
541 dev_info(ctrl->ctrl.device,
542 "rdma connection establishment failed (%d)\n", ret);
543 goto out_destroy_cm_id;
546 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
551 rdma_destroy_id(queue->cm_id);
552 nvme_rdma_destroy_queue_ib(queue);
556 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
558 if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
561 rdma_disconnect(queue->cm_id);
562 ib_drain_qp(queue->qp);
565 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
567 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
570 nvme_rdma_destroy_queue_ib(queue);
571 rdma_destroy_id(queue->cm_id);
574 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
578 for (i = 1; i < ctrl->ctrl.queue_count; i++)
579 nvme_rdma_free_queue(&ctrl->queues[i]);
582 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
586 for (i = 1; i < ctrl->ctrl.queue_count; i++)
587 nvme_rdma_stop_queue(&ctrl->queues[i]);
590 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
595 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
597 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
600 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags);
602 dev_info(ctrl->ctrl.device,
603 "failed to connect queue: %d ret=%d\n", idx, ret);
607 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
611 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
612 ret = nvme_rdma_start_queue(ctrl, i);
614 goto out_stop_queues;
620 for (i--; i >= 1; i--)
621 nvme_rdma_stop_queue(&ctrl->queues[i]);
625 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
627 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
628 struct ib_device *ibdev = ctrl->device->dev;
629 unsigned int nr_io_queues;
632 nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
635 * we map queues according to the device irq vectors for
636 * optimal locality so we don't need more queues than
637 * completion vectors.
639 nr_io_queues = min_t(unsigned int, nr_io_queues,
640 ibdev->num_comp_vectors);
642 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
646 if (nr_io_queues == 0) {
647 dev_err(ctrl->ctrl.device,
648 "unable to set any I/O queues\n");
652 ctrl->ctrl.queue_count = nr_io_queues + 1;
653 dev_info(ctrl->ctrl.device,
654 "creating %d I/O queues.\n", nr_io_queues);
656 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
657 ret = nvme_rdma_alloc_queue(ctrl, i,
658 ctrl->ctrl.sqsize + 1);
660 goto out_free_queues;
666 for (i--; i >= 1; i--)
667 nvme_rdma_free_queue(&ctrl->queues[i]);
672 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl,
673 struct blk_mq_tag_set *set)
675 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
677 blk_mq_free_tag_set(set);
678 nvme_rdma_dev_put(ctrl->device);
681 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
684 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
685 struct blk_mq_tag_set *set;
689 set = &ctrl->admin_tag_set;
690 memset(set, 0, sizeof(*set));
691 set->ops = &nvme_rdma_admin_mq_ops;
692 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
693 set->reserved_tags = 2; /* connect + keep-alive */
694 set->numa_node = NUMA_NO_NODE;
695 set->cmd_size = sizeof(struct nvme_rdma_request) +
696 SG_CHUNK_SIZE * sizeof(struct scatterlist);
697 set->driver_data = ctrl;
698 set->nr_hw_queues = 1;
699 set->timeout = ADMIN_TIMEOUT;
700 set->flags = BLK_MQ_F_NO_SCHED;
702 set = &ctrl->tag_set;
703 memset(set, 0, sizeof(*set));
704 set->ops = &nvme_rdma_mq_ops;
705 set->queue_depth = nctrl->sqsize + 1;
706 set->reserved_tags = 1; /* fabric connect */
707 set->numa_node = NUMA_NO_NODE;
708 set->flags = BLK_MQ_F_SHOULD_MERGE;
709 set->cmd_size = sizeof(struct nvme_rdma_request) +
710 SG_CHUNK_SIZE * sizeof(struct scatterlist);
711 set->driver_data = ctrl;
712 set->nr_hw_queues = nctrl->queue_count - 1;
713 set->timeout = NVME_IO_TIMEOUT;
716 ret = blk_mq_alloc_tag_set(set);
721 * We need a reference on the device as long as the tag_set is alive,
722 * as the MRs in the request structures need a valid ib_device.
724 ret = nvme_rdma_dev_get(ctrl->device);
727 goto out_free_tagset;
733 blk_mq_free_tag_set(set);
738 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
742 blk_cleanup_queue(ctrl->ctrl.admin_q);
743 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
745 if (ctrl->async_event_sqe.data) {
746 cancel_work_sync(&ctrl->ctrl.async_event_work);
747 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
748 sizeof(struct nvme_command), DMA_TO_DEVICE);
749 ctrl->async_event_sqe.data = NULL;
751 nvme_rdma_free_queue(&ctrl->queues[0]);
754 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
759 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
763 ctrl->device = ctrl->queues[0].device;
765 ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
767 error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
768 sizeof(struct nvme_command), DMA_TO_DEVICE);
773 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
774 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
775 error = PTR_ERR(ctrl->ctrl.admin_tagset);
776 goto out_free_async_qe;
779 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
780 if (IS_ERR(ctrl->ctrl.admin_q)) {
781 error = PTR_ERR(ctrl->ctrl.admin_q);
782 goto out_free_tagset;
786 error = nvme_rdma_start_queue(ctrl, 0);
788 goto out_cleanup_queue;
790 error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
793 dev_err(ctrl->ctrl.device,
794 "prop_get NVME_REG_CAP failed\n");
799 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
801 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
805 ctrl->ctrl.max_hw_sectors =
806 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
808 error = nvme_init_identify(&ctrl->ctrl);
815 nvme_rdma_stop_queue(&ctrl->queues[0]);
818 blk_cleanup_queue(ctrl->ctrl.admin_q);
821 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
823 if (ctrl->async_event_sqe.data) {
824 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
825 sizeof(struct nvme_command), DMA_TO_DEVICE);
826 ctrl->async_event_sqe.data = NULL;
829 nvme_rdma_free_queue(&ctrl->queues[0]);
833 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
837 blk_cleanup_queue(ctrl->ctrl.connect_q);
838 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
840 nvme_rdma_free_io_queues(ctrl);
843 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
847 ret = nvme_rdma_alloc_io_queues(ctrl);
852 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
853 if (IS_ERR(ctrl->ctrl.tagset)) {
854 ret = PTR_ERR(ctrl->ctrl.tagset);
855 goto out_free_io_queues;
858 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
859 if (IS_ERR(ctrl->ctrl.connect_q)) {
860 ret = PTR_ERR(ctrl->ctrl.connect_q);
861 goto out_free_tag_set;
864 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
865 ctrl->ctrl.queue_count - 1);
868 ret = nvme_rdma_start_io_queues(ctrl);
870 goto out_cleanup_connect_q;
874 out_cleanup_connect_q:
876 blk_cleanup_queue(ctrl->ctrl.connect_q);
879 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
881 nvme_rdma_free_io_queues(ctrl);
885 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
888 mutex_lock(&ctrl->teardown_lock);
889 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
890 nvme_rdma_stop_queue(&ctrl->queues[0]);
891 if (ctrl->ctrl.admin_tagset)
892 blk_mq_tagset_busy_iter(ctrl->ctrl.admin_tagset,
893 nvme_cancel_request, &ctrl->ctrl);
894 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
895 nvme_rdma_destroy_admin_queue(ctrl, remove);
896 mutex_unlock(&ctrl->teardown_lock);
899 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
902 mutex_lock(&ctrl->teardown_lock);
903 if (ctrl->ctrl.queue_count > 1) {
904 nvme_stop_queues(&ctrl->ctrl);
905 nvme_rdma_stop_io_queues(ctrl);
906 if (ctrl->ctrl.tagset)
907 blk_mq_tagset_busy_iter(ctrl->ctrl.tagset,
908 nvme_cancel_request, &ctrl->ctrl);
910 nvme_start_queues(&ctrl->ctrl);
911 nvme_rdma_destroy_io_queues(ctrl, remove);
913 mutex_unlock(&ctrl->teardown_lock);
916 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
918 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
920 cancel_work_sync(&ctrl->err_work);
921 cancel_delayed_work_sync(&ctrl->reconnect_work);
924 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
926 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
928 if (list_empty(&ctrl->list))
931 mutex_lock(&nvme_rdma_ctrl_mutex);
932 list_del(&ctrl->list);
933 mutex_unlock(&nvme_rdma_ctrl_mutex);
935 nvmf_free_options(nctrl->opts);
941 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
943 /* If we are resetting/deleting then do nothing */
944 if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
945 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
946 ctrl->ctrl.state == NVME_CTRL_LIVE);
950 if (nvmf_should_reconnect(&ctrl->ctrl)) {
951 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
952 ctrl->ctrl.opts->reconnect_delay);
953 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
954 ctrl->ctrl.opts->reconnect_delay * HZ);
956 nvme_delete_ctrl(&ctrl->ctrl);
960 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
965 ret = nvme_rdma_configure_admin_queue(ctrl, new);
969 if (ctrl->ctrl.icdoff) {
971 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
975 if (!(ctrl->ctrl.sgls & (1 << 2))) {
977 dev_err(ctrl->ctrl.device,
978 "Mandatory keyed sgls are not supported!\n");
982 if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
983 dev_warn(ctrl->ctrl.device,
984 "queue_size %zu > ctrl sqsize %u, clamping down\n",
985 ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
988 if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
989 dev_warn(ctrl->ctrl.device,
990 "sqsize %u > ctrl maxcmd %u, clamping down\n",
991 ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
992 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
995 if (ctrl->ctrl.sgls & (1 << 20))
996 ctrl->use_inline_data = true;
998 if (ctrl->ctrl.queue_count > 1) {
999 ret = nvme_rdma_configure_io_queues(ctrl, new);
1004 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1006 /* state change failure is ok if we're in DELETING state */
1007 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1012 nvme_start_ctrl(&ctrl->ctrl);
1016 if (ctrl->ctrl.queue_count > 1)
1017 nvme_rdma_destroy_io_queues(ctrl, new);
1019 nvme_rdma_stop_queue(&ctrl->queues[0]);
1020 nvme_rdma_destroy_admin_queue(ctrl, new);
1024 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1026 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1027 struct nvme_rdma_ctrl, reconnect_work);
1029 ++ctrl->ctrl.nr_reconnects;
1031 if (nvme_rdma_setup_ctrl(ctrl, false))
1034 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1035 ctrl->ctrl.nr_reconnects);
1037 ctrl->ctrl.nr_reconnects = 0;
1042 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1043 ctrl->ctrl.nr_reconnects);
1044 nvme_rdma_reconnect_or_remove(ctrl);
1047 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1049 struct nvme_rdma_ctrl *ctrl = container_of(work,
1050 struct nvme_rdma_ctrl, err_work);
1052 nvme_stop_keep_alive(&ctrl->ctrl);
1053 flush_work(&ctrl->ctrl.async_event_work);
1054 nvme_rdma_teardown_io_queues(ctrl, false);
1055 nvme_start_queues(&ctrl->ctrl);
1056 nvme_rdma_teardown_admin_queue(ctrl, false);
1058 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1059 /* state change failure is ok if we're in DELETING state */
1060 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1064 nvme_rdma_reconnect_or_remove(ctrl);
1067 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1069 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1072 queue_work(nvme_wq, &ctrl->err_work);
1075 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1078 struct nvme_rdma_queue *queue = cq->cq_context;
1079 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1081 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1082 dev_info(ctrl->ctrl.device,
1083 "%s for CQE 0x%p failed with status %s (%d)\n",
1085 ib_wc_status_msg(wc->status), wc->status);
1086 nvme_rdma_error_recovery(ctrl);
1089 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1091 if (unlikely(wc->status != IB_WC_SUCCESS))
1092 nvme_rdma_wr_error(cq, wc, "MEMREG");
1095 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1097 struct nvme_rdma_request *req =
1098 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1099 struct request *rq = blk_mq_rq_from_pdu(req);
1101 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1102 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1106 if (refcount_dec_and_test(&req->ref))
1107 nvme_end_request(rq, req->status, req->result);
1111 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1112 struct nvme_rdma_request *req)
1114 struct ib_send_wr wr = {
1115 .opcode = IB_WR_LOCAL_INV,
1118 .send_flags = IB_SEND_SIGNALED,
1119 .ex.invalidate_rkey = req->mr->rkey,
1122 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1123 wr.wr_cqe = &req->reg_cqe;
1125 return ib_post_send(queue->qp, &wr, NULL);
1128 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1131 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1132 struct nvme_rdma_device *dev = queue->device;
1133 struct ib_device *ibdev = dev->dev;
1135 if (!blk_rq_payload_bytes(rq))
1139 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1143 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1144 req->nents, rq_data_dir(rq) ==
1145 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1147 nvme_cleanup_cmd(rq);
1148 sg_free_table_chained(&req->sg_table, true);
1151 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1153 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1156 put_unaligned_le24(0, sg->length);
1157 put_unaligned_le32(0, sg->key);
1158 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1162 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1163 struct nvme_rdma_request *req, struct nvme_command *c,
1166 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1167 struct scatterlist *sgl = req->sg_table.sgl;
1168 struct ib_sge *sge = &req->sge[1];
1172 for (i = 0; i < count; i++, sgl++, sge++) {
1173 sge->addr = sg_dma_address(sgl);
1174 sge->length = sg_dma_len(sgl);
1175 sge->lkey = queue->device->pd->local_dma_lkey;
1179 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1180 sg->length = cpu_to_le32(len);
1181 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1183 req->num_sge += count;
1187 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1188 struct nvme_rdma_request *req, struct nvme_command *c)
1190 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1192 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1193 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1194 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1195 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1199 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1200 struct nvme_rdma_request *req, struct nvme_command *c,
1203 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1206 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1207 if (WARN_ON_ONCE(!req->mr))
1211 * Align the MR to a 4K page size to match the ctrl page size and
1212 * the block virtual boundary.
1214 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1215 if (unlikely(nr < count)) {
1216 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1223 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1225 req->reg_cqe.done = nvme_rdma_memreg_done;
1226 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1227 req->reg_wr.wr.opcode = IB_WR_REG_MR;
1228 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1229 req->reg_wr.wr.num_sge = 0;
1230 req->reg_wr.mr = req->mr;
1231 req->reg_wr.key = req->mr->rkey;
1232 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1233 IB_ACCESS_REMOTE_READ |
1234 IB_ACCESS_REMOTE_WRITE;
1236 sg->addr = cpu_to_le64(req->mr->iova);
1237 put_unaligned_le24(req->mr->length, sg->length);
1238 put_unaligned_le32(req->mr->rkey, sg->key);
1239 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1240 NVME_SGL_FMT_INVALIDATE;
1245 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1246 struct request *rq, struct nvme_command *c)
1248 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1249 struct nvme_rdma_device *dev = queue->device;
1250 struct ib_device *ibdev = dev->dev;
1254 refcount_set(&req->ref, 2); /* send and recv completions */
1256 c->common.flags |= NVME_CMD_SGL_METABUF;
1258 if (!blk_rq_payload_bytes(rq))
1259 return nvme_rdma_set_sg_null(c);
1261 req->sg_table.sgl = req->first_sgl;
1262 ret = sg_alloc_table_chained(&req->sg_table,
1263 blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1267 req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1269 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1270 rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1271 if (unlikely(count <= 0)) {
1273 goto out_free_table;
1276 if (count <= dev->num_inline_segments) {
1277 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1278 queue->ctrl->use_inline_data &&
1279 blk_rq_payload_bytes(rq) <=
1280 nvme_rdma_inline_data_size(queue)) {
1281 ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1285 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1286 ret = nvme_rdma_map_sg_single(queue, req, c);
1291 ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1299 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1300 req->nents, rq_data_dir(rq) ==
1301 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1303 sg_free_table_chained(&req->sg_table, true);
1307 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1309 struct nvme_rdma_qe *qe =
1310 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1311 struct nvme_rdma_request *req =
1312 container_of(qe, struct nvme_rdma_request, sqe);
1313 struct request *rq = blk_mq_rq_from_pdu(req);
1315 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1316 nvme_rdma_wr_error(cq, wc, "SEND");
1320 if (refcount_dec_and_test(&req->ref))
1321 nvme_end_request(rq, req->status, req->result);
1324 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1325 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1326 struct ib_send_wr *first)
1328 struct ib_send_wr wr;
1331 sge->addr = qe->dma;
1332 sge->length = sizeof(struct nvme_command),
1333 sge->lkey = queue->device->pd->local_dma_lkey;
1336 wr.wr_cqe = &qe->cqe;
1338 wr.num_sge = num_sge;
1339 wr.opcode = IB_WR_SEND;
1340 wr.send_flags = IB_SEND_SIGNALED;
1347 ret = ib_post_send(queue->qp, first, NULL);
1348 if (unlikely(ret)) {
1349 dev_err(queue->ctrl->ctrl.device,
1350 "%s failed with error code %d\n", __func__, ret);
1355 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1356 struct nvme_rdma_qe *qe)
1358 struct ib_recv_wr wr;
1362 list.addr = qe->dma;
1363 list.length = sizeof(struct nvme_completion);
1364 list.lkey = queue->device->pd->local_dma_lkey;
1366 qe->cqe.done = nvme_rdma_recv_done;
1369 wr.wr_cqe = &qe->cqe;
1373 ret = ib_post_recv(queue->qp, &wr, NULL);
1374 if (unlikely(ret)) {
1375 dev_err(queue->ctrl->ctrl.device,
1376 "%s failed with error code %d\n", __func__, ret);
1381 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1383 u32 queue_idx = nvme_rdma_queue_idx(queue);
1386 return queue->ctrl->admin_tag_set.tags[queue_idx];
1387 return queue->ctrl->tag_set.tags[queue_idx - 1];
1390 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1392 if (unlikely(wc->status != IB_WC_SUCCESS))
1393 nvme_rdma_wr_error(cq, wc, "ASYNC");
1396 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1398 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1399 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1400 struct ib_device *dev = queue->device->dev;
1401 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1402 struct nvme_command *cmd = sqe->data;
1406 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1408 memset(cmd, 0, sizeof(*cmd));
1409 cmd->common.opcode = nvme_admin_async_event;
1410 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1411 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1412 nvme_rdma_set_sg_null(cmd);
1414 sqe->cqe.done = nvme_rdma_async_done;
1416 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1419 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1423 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1424 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1427 struct nvme_rdma_request *req;
1430 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1432 dev_err(queue->ctrl->ctrl.device,
1433 "tag 0x%x on QP %#x not found\n",
1434 cqe->command_id, queue->qp->qp_num);
1435 nvme_rdma_error_recovery(queue->ctrl);
1438 req = blk_mq_rq_to_pdu(rq);
1440 req->status = cqe->status;
1441 req->result = cqe->result;
1443 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1444 if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1445 dev_err(queue->ctrl->ctrl.device,
1446 "Bogus remote invalidation for rkey %#x\n",
1448 nvme_rdma_error_recovery(queue->ctrl);
1450 } else if (req->mr) {
1451 ret = nvme_rdma_inv_rkey(queue, req);
1452 if (unlikely(ret < 0)) {
1453 dev_err(queue->ctrl->ctrl.device,
1454 "Queueing INV WR for rkey %#x failed (%d)\n",
1455 req->mr->rkey, ret);
1456 nvme_rdma_error_recovery(queue->ctrl);
1458 /* the local invalidation completion will end the request */
1462 if (refcount_dec_and_test(&req->ref)) {
1465 nvme_end_request(rq, req->status, req->result);
1471 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1473 struct nvme_rdma_qe *qe =
1474 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1475 struct nvme_rdma_queue *queue = cq->cq_context;
1476 struct ib_device *ibdev = queue->device->dev;
1477 struct nvme_completion *cqe = qe->data;
1478 const size_t len = sizeof(struct nvme_completion);
1481 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1482 nvme_rdma_wr_error(cq, wc, "RECV");
1486 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1488 * AEN requests are special as they don't time out and can
1489 * survive any kind of queue freeze and often don't respond to
1490 * aborts. We don't even bother to allocate a struct request
1491 * for them but rather special case them here.
1493 if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1494 cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1495 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1498 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1499 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1501 nvme_rdma_post_recv(queue, qe);
1505 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1507 __nvme_rdma_recv_done(cq, wc, -1);
1510 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1514 for (i = 0; i < queue->queue_size; i++) {
1515 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1517 goto out_destroy_queue_ib;
1522 out_destroy_queue_ib:
1523 nvme_rdma_destroy_queue_ib(queue);
1527 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1528 struct rdma_cm_event *ev)
1530 struct rdma_cm_id *cm_id = queue->cm_id;
1531 int status = ev->status;
1532 const char *rej_msg;
1533 const struct nvme_rdma_cm_rej *rej_data;
1536 rej_msg = rdma_reject_msg(cm_id, status);
1537 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1539 if (rej_data && rej_data_len >= sizeof(u16)) {
1540 u16 sts = le16_to_cpu(rej_data->sts);
1542 dev_err(queue->ctrl->ctrl.device,
1543 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1544 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1546 dev_err(queue->ctrl->ctrl.device,
1547 "Connect rejected: status %d (%s).\n", status, rej_msg);
1553 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1557 ret = nvme_rdma_create_queue_ib(queue);
1561 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1563 dev_err(queue->ctrl->ctrl.device,
1564 "rdma_resolve_route failed (%d).\n",
1566 goto out_destroy_queue;
1572 nvme_rdma_destroy_queue_ib(queue);
1576 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1578 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1579 struct rdma_conn_param param = { };
1580 struct nvme_rdma_cm_req priv = { };
1583 param.qp_num = queue->qp->qp_num;
1584 param.flow_control = 1;
1586 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1587 /* maximum retry count */
1588 param.retry_count = 7;
1589 param.rnr_retry_count = 7;
1590 param.private_data = &priv;
1591 param.private_data_len = sizeof(priv);
1593 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1594 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1596 * set the admin queue depth to the minimum size
1597 * specified by the Fabrics standard.
1599 if (priv.qid == 0) {
1600 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1601 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1604 * current interpretation of the fabrics spec
1605 * is at minimum you make hrqsize sqsize+1, or a
1606 * 1's based representation of sqsize.
1608 priv.hrqsize = cpu_to_le16(queue->queue_size);
1609 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1612 ret = rdma_connect(queue->cm_id, ¶m);
1614 dev_err(ctrl->ctrl.device,
1615 "rdma_connect failed (%d).\n", ret);
1616 goto out_destroy_queue_ib;
1621 out_destroy_queue_ib:
1622 nvme_rdma_destroy_queue_ib(queue);
1626 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1627 struct rdma_cm_event *ev)
1629 struct nvme_rdma_queue *queue = cm_id->context;
1632 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1633 rdma_event_msg(ev->event), ev->event,
1636 switch (ev->event) {
1637 case RDMA_CM_EVENT_ADDR_RESOLVED:
1638 cm_error = nvme_rdma_addr_resolved(queue);
1640 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1641 cm_error = nvme_rdma_route_resolved(queue);
1643 case RDMA_CM_EVENT_ESTABLISHED:
1644 queue->cm_error = nvme_rdma_conn_established(queue);
1645 /* complete cm_done regardless of success/failure */
1646 complete(&queue->cm_done);
1648 case RDMA_CM_EVENT_REJECTED:
1649 cm_error = nvme_rdma_conn_rejected(queue, ev);
1651 case RDMA_CM_EVENT_ROUTE_ERROR:
1652 case RDMA_CM_EVENT_CONNECT_ERROR:
1653 case RDMA_CM_EVENT_UNREACHABLE:
1654 nvme_rdma_destroy_queue_ib(queue);
1656 case RDMA_CM_EVENT_ADDR_ERROR:
1657 dev_dbg(queue->ctrl->ctrl.device,
1658 "CM error event %d\n", ev->event);
1659 cm_error = -ECONNRESET;
1661 case RDMA_CM_EVENT_DISCONNECTED:
1662 case RDMA_CM_EVENT_ADDR_CHANGE:
1663 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1664 dev_dbg(queue->ctrl->ctrl.device,
1665 "disconnect received - connection closed\n");
1666 nvme_rdma_error_recovery(queue->ctrl);
1668 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1669 /* device removal is handled via the ib_client API */
1672 dev_err(queue->ctrl->ctrl.device,
1673 "Unexpected RDMA CM event (%d)\n", ev->event);
1674 nvme_rdma_error_recovery(queue->ctrl);
1679 queue->cm_error = cm_error;
1680 complete(&queue->cm_done);
1686 static enum blk_eh_timer_return
1687 nvme_rdma_timeout(struct request *rq, bool reserved)
1689 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1690 struct nvme_rdma_queue *queue = req->queue;
1691 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1693 dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
1694 rq->tag, nvme_rdma_queue_idx(queue));
1696 if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
1698 * Teardown immediately if controller times out while starting
1699 * or we are already started error recovery. all outstanding
1700 * requests are completed on shutdown, so we return BLK_EH_DONE.
1702 flush_work(&ctrl->err_work);
1703 nvme_rdma_teardown_io_queues(ctrl, false);
1704 nvme_rdma_teardown_admin_queue(ctrl, false);
1708 dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1709 nvme_rdma_error_recovery(ctrl);
1711 return BLK_EH_RESET_TIMER;
1714 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1715 const struct blk_mq_queue_data *bd)
1717 struct nvme_ns *ns = hctx->queue->queuedata;
1718 struct nvme_rdma_queue *queue = hctx->driver_data;
1719 struct request *rq = bd->rq;
1720 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1721 struct nvme_rdma_qe *sqe = &req->sqe;
1722 struct nvme_command *c = sqe->data;
1723 struct ib_device *dev;
1724 bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1728 WARN_ON_ONCE(rq->tag < 0);
1730 if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1731 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
1733 dev = queue->device->dev;
1734 ib_dma_sync_single_for_cpu(dev, sqe->dma,
1735 sizeof(struct nvme_command), DMA_TO_DEVICE);
1737 ret = nvme_setup_cmd(ns, rq, c);
1741 blk_mq_start_request(rq);
1743 err = nvme_rdma_map_data(queue, rq, c);
1744 if (unlikely(err < 0)) {
1745 dev_err(queue->ctrl->ctrl.device,
1746 "Failed to map data (%d)\n", err);
1747 nvme_cleanup_cmd(rq);
1751 sqe->cqe.done = nvme_rdma_send_done;
1753 ib_dma_sync_single_for_device(dev, sqe->dma,
1754 sizeof(struct nvme_command), DMA_TO_DEVICE);
1756 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1757 req->mr ? &req->reg_wr.wr : NULL);
1758 if (unlikely(err)) {
1759 nvme_rdma_unmap_data(queue, rq);
1765 if (err == -ENOMEM || err == -EAGAIN)
1766 return BLK_STS_RESOURCE;
1767 return BLK_STS_IOERR;
1770 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1772 struct nvme_rdma_queue *queue = hctx->driver_data;
1773 struct ib_cq *cq = queue->ib_cq;
1777 while (ib_poll_cq(cq, 1, &wc) > 0) {
1778 struct ib_cqe *cqe = wc.wr_cqe;
1781 if (cqe->done == nvme_rdma_recv_done)
1782 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1791 static void nvme_rdma_complete_rq(struct request *rq)
1793 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1795 nvme_rdma_unmap_data(req->queue, rq);
1796 nvme_complete_rq(rq);
1799 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1801 struct nvme_rdma_ctrl *ctrl = set->driver_data;
1803 return blk_mq_rdma_map_queues(set, ctrl->device->dev, 0);
1806 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1807 .queue_rq = nvme_rdma_queue_rq,
1808 .complete = nvme_rdma_complete_rq,
1809 .init_request = nvme_rdma_init_request,
1810 .exit_request = nvme_rdma_exit_request,
1811 .init_hctx = nvme_rdma_init_hctx,
1812 .poll = nvme_rdma_poll,
1813 .timeout = nvme_rdma_timeout,
1814 .map_queues = nvme_rdma_map_queues,
1817 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1818 .queue_rq = nvme_rdma_queue_rq,
1819 .complete = nvme_rdma_complete_rq,
1820 .init_request = nvme_rdma_init_request,
1821 .exit_request = nvme_rdma_exit_request,
1822 .init_hctx = nvme_rdma_init_admin_hctx,
1823 .timeout = nvme_rdma_timeout,
1826 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1828 nvme_rdma_teardown_io_queues(ctrl, shutdown);
1830 nvme_shutdown_ctrl(&ctrl->ctrl);
1832 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1833 nvme_rdma_teardown_admin_queue(ctrl, shutdown);
1836 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1838 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1841 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1843 struct nvme_rdma_ctrl *ctrl =
1844 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1846 nvme_stop_ctrl(&ctrl->ctrl);
1847 nvme_rdma_shutdown_ctrl(ctrl, false);
1849 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1850 /* state change failure should never happen */
1855 if (nvme_rdma_setup_ctrl(ctrl, false))
1861 ++ctrl->ctrl.nr_reconnects;
1862 nvme_rdma_reconnect_or_remove(ctrl);
1865 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1867 .module = THIS_MODULE,
1868 .flags = NVME_F_FABRICS,
1869 .reg_read32 = nvmf_reg_read32,
1870 .reg_read64 = nvmf_reg_read64,
1871 .reg_write32 = nvmf_reg_write32,
1872 .free_ctrl = nvme_rdma_free_ctrl,
1873 .submit_async_event = nvme_rdma_submit_async_event,
1874 .delete_ctrl = nvme_rdma_delete_ctrl,
1875 .get_address = nvmf_get_address,
1876 .stop_ctrl = nvme_rdma_stop_ctrl,
1880 __nvme_rdma_options_match(struct nvme_rdma_ctrl *ctrl,
1881 struct nvmf_ctrl_options *opts)
1883 char *stdport = __stringify(NVME_RDMA_IP_PORT);
1886 if (!nvmf_ctlr_matches_baseopts(&ctrl->ctrl, opts) ||
1887 strcmp(opts->traddr, ctrl->ctrl.opts->traddr))
1890 if (opts->mask & NVMF_OPT_TRSVCID &&
1891 ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1892 if (strcmp(opts->trsvcid, ctrl->ctrl.opts->trsvcid))
1894 } else if (opts->mask & NVMF_OPT_TRSVCID) {
1895 if (strcmp(opts->trsvcid, stdport))
1897 } else if (ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1898 if (strcmp(stdport, ctrl->ctrl.opts->trsvcid))
1901 /* else, it's a match as both have stdport. Fall to next checks */
1904 * checking the local address is rough. In most cases, one
1905 * is not specified and the host port is selected by the stack.
1907 * Assume no match if:
1908 * local address is specified and address is not the same
1909 * local address is not specified but remote is, or vice versa
1910 * (admin using specific host_traddr when it matters).
1912 if (opts->mask & NVMF_OPT_HOST_TRADDR &&
1913 ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR) {
1914 if (strcmp(opts->host_traddr, ctrl->ctrl.opts->host_traddr))
1916 } else if (opts->mask & NVMF_OPT_HOST_TRADDR ||
1917 ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
1920 * if neither controller had an host port specified, assume it's
1921 * a match as everything else matched.
1928 * Fails a connection request if it matches an existing controller
1929 * (association) with the same tuple:
1930 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1932 * if local address is not specified in the request, it will match an
1933 * existing controller with all the other parameters the same and no
1934 * local port address specified as well.
1936 * The ports don't need to be compared as they are intrinsically
1937 * already matched by the port pointers supplied.
1940 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1942 struct nvme_rdma_ctrl *ctrl;
1945 mutex_lock(&nvme_rdma_ctrl_mutex);
1946 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1947 found = __nvme_rdma_options_match(ctrl, opts);
1951 mutex_unlock(&nvme_rdma_ctrl_mutex);
1956 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1957 struct nvmf_ctrl_options *opts)
1959 struct nvme_rdma_ctrl *ctrl;
1964 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1966 return ERR_PTR(-ENOMEM);
1967 ctrl->ctrl.opts = opts;
1968 INIT_LIST_HEAD(&ctrl->list);
1969 mutex_init(&ctrl->teardown_lock);
1971 if (opts->mask & NVMF_OPT_TRSVCID)
1972 port = opts->trsvcid;
1974 port = __stringify(NVME_RDMA_IP_PORT);
1976 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1977 opts->traddr, port, &ctrl->addr);
1979 pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1983 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1984 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1985 opts->host_traddr, NULL, &ctrl->src_addr);
1987 pr_err("malformed src address passed: %s\n",
1993 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1998 INIT_DELAYED_WORK(&ctrl->reconnect_work,
1999 nvme_rdma_reconnect_ctrl_work);
2000 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2001 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2003 ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
2004 ctrl->ctrl.sqsize = opts->queue_size - 1;
2005 ctrl->ctrl.kato = opts->kato;
2008 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2013 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2014 0 /* no quirks, we're perfect! */);
2016 goto out_kfree_queues;
2018 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2019 WARN_ON_ONCE(!changed);
2021 ret = nvme_rdma_setup_ctrl(ctrl, true);
2023 goto out_uninit_ctrl;
2025 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2026 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2028 nvme_get_ctrl(&ctrl->ctrl);
2030 mutex_lock(&nvme_rdma_ctrl_mutex);
2031 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2032 mutex_unlock(&nvme_rdma_ctrl_mutex);
2037 nvme_uninit_ctrl(&ctrl->ctrl);
2038 nvme_put_ctrl(&ctrl->ctrl);
2041 return ERR_PTR(ret);
2043 kfree(ctrl->queues);
2046 return ERR_PTR(ret);
2049 static struct nvmf_transport_ops nvme_rdma_transport = {
2051 .module = THIS_MODULE,
2052 .required_opts = NVMF_OPT_TRADDR,
2053 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2054 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
2055 .create_ctrl = nvme_rdma_create_ctrl,
2058 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2060 struct nvme_rdma_ctrl *ctrl;
2061 struct nvme_rdma_device *ndev;
2064 mutex_lock(&device_list_mutex);
2065 list_for_each_entry(ndev, &device_list, entry) {
2066 if (ndev->dev == ib_device) {
2071 mutex_unlock(&device_list_mutex);
2076 /* Delete all controllers using this device */
2077 mutex_lock(&nvme_rdma_ctrl_mutex);
2078 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2079 if (ctrl->device->dev != ib_device)
2081 nvme_delete_ctrl(&ctrl->ctrl);
2083 mutex_unlock(&nvme_rdma_ctrl_mutex);
2085 flush_workqueue(nvme_delete_wq);
2088 static struct ib_client nvme_rdma_ib_client = {
2089 .name = "nvme_rdma",
2090 .remove = nvme_rdma_remove_one
2093 static int __init nvme_rdma_init_module(void)
2097 ret = ib_register_client(&nvme_rdma_ib_client);
2101 ret = nvmf_register_transport(&nvme_rdma_transport);
2103 goto err_unreg_client;
2108 ib_unregister_client(&nvme_rdma_ib_client);
2112 static void __exit nvme_rdma_cleanup_module(void)
2114 nvmf_unregister_transport(&nvme_rdma_transport);
2115 ib_unregister_client(&nvme_rdma_ib_client);
2118 module_init(nvme_rdma_init_module);
2119 module_exit(nvme_rdma_cleanup_module);
2121 MODULE_LICENSE("GPL v2");