1 // SPDX-License-Identifier: GPL-2.0
3 * NVMe over Fabrics RDMA host code.
4 * Copyright (c) 2015-2016 HGST, a Western Digital Company.
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/module.h>
8 #include <linux/init.h>
9 #include <linux/slab.h>
10 #include <rdma/mr_pool.h>
11 #include <linux/err.h>
12 #include <linux/string.h>
13 #include <linux/atomic.h>
14 #include <linux/blk-mq.h>
15 #include <linux/blk-integrity.h>
16 #include <linux/types.h>
17 #include <linux/list.h>
18 #include <linux/mutex.h>
19 #include <linux/scatterlist.h>
20 #include <linux/nvme.h>
21 #include <asm/unaligned.h>
23 #include <rdma/ib_verbs.h>
24 #include <rdma/rdma_cm.h>
25 #include <linux/nvme-rdma.h>
31 #define NVME_RDMA_CM_TIMEOUT_MS 3000 /* 3 second */
33 #define NVME_RDMA_MAX_SEGMENTS 256
35 #define NVME_RDMA_MAX_INLINE_SEGMENTS 4
37 #define NVME_RDMA_DATA_SGL_SIZE \
38 (sizeof(struct scatterlist) * NVME_INLINE_SG_CNT)
39 #define NVME_RDMA_METADATA_SGL_SIZE \
40 (sizeof(struct scatterlist) * NVME_INLINE_METADATA_SG_CNT)
42 struct nvme_rdma_device {
43 struct ib_device *dev;
46 struct list_head entry;
47 unsigned int num_inline_segments;
56 struct nvme_rdma_sgl {
58 struct sg_table sg_table;
61 struct nvme_rdma_queue;
62 struct nvme_rdma_request {
63 struct nvme_request req;
65 struct nvme_rdma_qe sqe;
66 union nvme_result result;
69 struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
71 struct ib_reg_wr reg_wr;
72 struct ib_cqe reg_cqe;
73 struct nvme_rdma_queue *queue;
74 struct nvme_rdma_sgl data_sgl;
75 struct nvme_rdma_sgl *metadata_sgl;
79 enum nvme_rdma_queue_flags {
80 NVME_RDMA_Q_ALLOCATED = 0,
82 NVME_RDMA_Q_TR_READY = 2,
85 struct nvme_rdma_queue {
86 struct nvme_rdma_qe *rsp_ring;
88 size_t cmnd_capsule_len;
89 struct nvme_rdma_ctrl *ctrl;
90 struct nvme_rdma_device *device;
95 struct rdma_cm_id *cm_id;
97 struct completion cm_done;
100 struct mutex queue_lock;
103 struct nvme_rdma_ctrl {
104 /* read only in the hot path */
105 struct nvme_rdma_queue *queues;
107 /* other member variables */
108 struct blk_mq_tag_set tag_set;
109 struct work_struct err_work;
111 struct nvme_rdma_qe async_event_sqe;
113 struct delayed_work reconnect_work;
115 struct list_head list;
117 struct blk_mq_tag_set admin_tag_set;
118 struct nvme_rdma_device *device;
122 struct sockaddr_storage addr;
123 struct sockaddr_storage src_addr;
125 struct nvme_ctrl ctrl;
126 bool use_inline_data;
127 u32 io_queues[HCTX_MAX_TYPES];
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);
154 static void nvme_rdma_complete_rq(struct request *rq);
156 static const struct blk_mq_ops nvme_rdma_mq_ops;
157 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
159 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
161 return queue - queue->ctrl->queues;
164 static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue)
166 return nvme_rdma_queue_idx(queue) >
167 queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] +
168 queue->ctrl->io_queues[HCTX_TYPE_READ];
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)) {
200 static void nvme_rdma_free_ring(struct ib_device *ibdev,
201 struct nvme_rdma_qe *ring, size_t ib_queue_size,
202 size_t capsule_size, enum dma_data_direction dir)
206 for (i = 0; i < ib_queue_size; i++)
207 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
211 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
212 size_t ib_queue_size, size_t capsule_size,
213 enum dma_data_direction dir)
215 struct nvme_rdma_qe *ring;
218 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
223 * Bind the CQEs (post recv buffers) DMA mapping to the RDMA queue
224 * lifetime. It's safe, since any chage in the underlying RDMA device
225 * will issue error recovery and queue re-creation.
227 for (i = 0; i < ib_queue_size; i++) {
228 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
235 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
239 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
241 pr_debug("QP event %s (%d)\n",
242 ib_event_msg(event->event), event->event);
246 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
250 ret = wait_for_completion_interruptible(&queue->cm_done);
253 WARN_ON_ONCE(queue->cm_error > 0);
254 return queue->cm_error;
257 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
259 struct nvme_rdma_device *dev = queue->device;
260 struct ib_qp_init_attr init_attr;
263 memset(&init_attr, 0, sizeof(init_attr));
264 init_attr.event_handler = nvme_rdma_qp_event;
266 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
268 init_attr.cap.max_recv_wr = queue->queue_size + 1;
269 init_attr.cap.max_recv_sge = 1;
270 init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
271 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
272 init_attr.qp_type = IB_QPT_RC;
273 init_attr.send_cq = queue->ib_cq;
274 init_attr.recv_cq = queue->ib_cq;
275 if (queue->pi_support)
276 init_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
277 init_attr.qp_context = queue;
279 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
281 queue->qp = queue->cm_id->qp;
285 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
286 struct request *rq, unsigned int hctx_idx)
288 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
290 kfree(req->sqe.data);
293 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
294 struct request *rq, unsigned int hctx_idx,
295 unsigned int numa_node)
297 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(set->driver_data);
298 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
299 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
300 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
302 nvme_req(rq)->ctrl = &ctrl->ctrl;
303 req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL);
307 /* metadata nvme_rdma_sgl struct is located after command's data SGL */
308 if (queue->pi_support)
309 req->metadata_sgl = (void *)nvme_req(rq) +
310 sizeof(struct nvme_rdma_request) +
311 NVME_RDMA_DATA_SGL_SIZE;
314 nvme_req(rq)->cmd = req->sqe.data;
319 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
320 unsigned int hctx_idx)
322 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(data);
323 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
325 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
327 hctx->driver_data = queue;
331 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
332 unsigned int hctx_idx)
334 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(data);
335 struct nvme_rdma_queue *queue = &ctrl->queues[0];
337 BUG_ON(hctx_idx != 0);
339 hctx->driver_data = queue;
343 static void nvme_rdma_free_dev(struct kref *ref)
345 struct nvme_rdma_device *ndev =
346 container_of(ref, struct nvme_rdma_device, ref);
348 mutex_lock(&device_list_mutex);
349 list_del(&ndev->entry);
350 mutex_unlock(&device_list_mutex);
352 ib_dealloc_pd(ndev->pd);
356 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
358 kref_put(&dev->ref, nvme_rdma_free_dev);
361 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
363 return kref_get_unless_zero(&dev->ref);
366 static struct nvme_rdma_device *
367 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
369 struct nvme_rdma_device *ndev;
371 mutex_lock(&device_list_mutex);
372 list_for_each_entry(ndev, &device_list, entry) {
373 if (ndev->dev->node_guid == cm_id->device->node_guid &&
374 nvme_rdma_dev_get(ndev))
378 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
382 ndev->dev = cm_id->device;
383 kref_init(&ndev->ref);
385 ndev->pd = ib_alloc_pd(ndev->dev,
386 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
387 if (IS_ERR(ndev->pd))
390 if (!(ndev->dev->attrs.device_cap_flags &
391 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
392 dev_err(&ndev->dev->dev,
393 "Memory registrations not supported.\n");
397 ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
398 ndev->dev->attrs.max_send_sge - 1);
399 list_add(&ndev->entry, &device_list);
401 mutex_unlock(&device_list_mutex);
405 ib_dealloc_pd(ndev->pd);
409 mutex_unlock(&device_list_mutex);
413 static void nvme_rdma_free_cq(struct nvme_rdma_queue *queue)
415 if (nvme_rdma_poll_queue(queue))
416 ib_free_cq(queue->ib_cq);
418 ib_cq_pool_put(queue->ib_cq, queue->cq_size);
421 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
423 struct nvme_rdma_device *dev;
424 struct ib_device *ibdev;
426 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
432 if (queue->pi_support)
433 ib_mr_pool_destroy(queue->qp, &queue->qp->sig_mrs);
434 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
437 * The cm_id object might have been destroyed during RDMA connection
438 * establishment error flow to avoid getting other cma events, thus
439 * the destruction of the QP shouldn't use rdma_cm API.
441 ib_destroy_qp(queue->qp);
442 nvme_rdma_free_cq(queue);
444 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
445 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
447 nvme_rdma_dev_put(dev);
450 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev, bool pi_support)
452 u32 max_page_list_len;
455 max_page_list_len = ibdev->attrs.max_pi_fast_reg_page_list_len;
457 max_page_list_len = ibdev->attrs.max_fast_reg_page_list_len;
459 return min_t(u32, NVME_RDMA_MAX_SEGMENTS, max_page_list_len - 1);
462 static int nvme_rdma_create_cq(struct ib_device *ibdev,
463 struct nvme_rdma_queue *queue)
465 int ret, comp_vector, idx = nvme_rdma_queue_idx(queue);
468 * Spread I/O queues completion vectors according their queue index.
469 * Admin queues can always go on completion vector 0.
471 comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors;
473 /* Polling queues need direct cq polling context */
474 if (nvme_rdma_poll_queue(queue))
475 queue->ib_cq = ib_alloc_cq(ibdev, queue, queue->cq_size,
476 comp_vector, IB_POLL_DIRECT);
478 queue->ib_cq = ib_cq_pool_get(ibdev, queue->cq_size,
479 comp_vector, IB_POLL_SOFTIRQ);
481 if (IS_ERR(queue->ib_cq)) {
482 ret = PTR_ERR(queue->ib_cq);
489 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
491 struct ib_device *ibdev;
492 const int send_wr_factor = 3; /* MR, SEND, INV */
493 const int cq_factor = send_wr_factor + 1; /* + RECV */
494 int ret, pages_per_mr;
496 queue->device = nvme_rdma_find_get_device(queue->cm_id);
497 if (!queue->device) {
498 dev_err(queue->cm_id->device->dev.parent,
499 "no client data found!\n");
500 return -ECONNREFUSED;
502 ibdev = queue->device->dev;
504 /* +1 for ib_drain_qp */
505 queue->cq_size = cq_factor * queue->queue_size + 1;
507 ret = nvme_rdma_create_cq(ibdev, queue);
511 ret = nvme_rdma_create_qp(queue, send_wr_factor);
513 goto out_destroy_ib_cq;
515 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
516 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
517 if (!queue->rsp_ring) {
523 * Currently we don't use SG_GAPS MR's so if the first entry is
524 * misaligned we'll end up using two entries for a single data page,
525 * so one additional entry is required.
527 pages_per_mr = nvme_rdma_get_max_fr_pages(ibdev, queue->pi_support) + 1;
528 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
533 dev_err(queue->ctrl->ctrl.device,
534 "failed to initialize MR pool sized %d for QID %d\n",
535 queue->queue_size, nvme_rdma_queue_idx(queue));
536 goto out_destroy_ring;
539 if (queue->pi_support) {
540 ret = ib_mr_pool_init(queue->qp, &queue->qp->sig_mrs,
541 queue->queue_size, IB_MR_TYPE_INTEGRITY,
542 pages_per_mr, pages_per_mr);
544 dev_err(queue->ctrl->ctrl.device,
545 "failed to initialize PI MR pool sized %d for QID %d\n",
546 queue->queue_size, nvme_rdma_queue_idx(queue));
547 goto out_destroy_mr_pool;
551 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
556 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
558 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
559 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
561 rdma_destroy_qp(queue->cm_id);
563 nvme_rdma_free_cq(queue);
565 nvme_rdma_dev_put(queue->device);
569 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
570 int idx, size_t queue_size)
572 struct nvme_rdma_queue *queue;
573 struct sockaddr *src_addr = NULL;
576 queue = &ctrl->queues[idx];
577 mutex_init(&queue->queue_lock);
579 if (idx && ctrl->ctrl.max_integrity_segments)
580 queue->pi_support = true;
582 queue->pi_support = false;
583 init_completion(&queue->cm_done);
586 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
588 queue->cmnd_capsule_len = sizeof(struct nvme_command);
590 queue->queue_size = queue_size;
592 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
593 RDMA_PS_TCP, IB_QPT_RC);
594 if (IS_ERR(queue->cm_id)) {
595 dev_info(ctrl->ctrl.device,
596 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
597 ret = PTR_ERR(queue->cm_id);
598 goto out_destroy_mutex;
601 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
602 src_addr = (struct sockaddr *)&ctrl->src_addr;
604 queue->cm_error = -ETIMEDOUT;
605 ret = rdma_resolve_addr(queue->cm_id, src_addr,
606 (struct sockaddr *)&ctrl->addr,
607 NVME_RDMA_CM_TIMEOUT_MS);
609 dev_info(ctrl->ctrl.device,
610 "rdma_resolve_addr failed (%d).\n", ret);
611 goto out_destroy_cm_id;
614 ret = nvme_rdma_wait_for_cm(queue);
616 dev_info(ctrl->ctrl.device,
617 "rdma connection establishment failed (%d)\n", ret);
618 goto out_destroy_cm_id;
621 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
626 rdma_destroy_id(queue->cm_id);
627 nvme_rdma_destroy_queue_ib(queue);
629 mutex_destroy(&queue->queue_lock);
633 static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
635 rdma_disconnect(queue->cm_id);
636 ib_drain_qp(queue->qp);
639 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
641 if (!test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
644 mutex_lock(&queue->queue_lock);
645 if (test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
646 __nvme_rdma_stop_queue(queue);
647 mutex_unlock(&queue->queue_lock);
650 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
652 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
655 rdma_destroy_id(queue->cm_id);
656 nvme_rdma_destroy_queue_ib(queue);
657 mutex_destroy(&queue->queue_lock);
660 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
664 for (i = 1; i < ctrl->ctrl.queue_count; i++)
665 nvme_rdma_free_queue(&ctrl->queues[i]);
668 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
672 for (i = 1; i < ctrl->ctrl.queue_count; i++)
673 nvme_rdma_stop_queue(&ctrl->queues[i]);
676 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
678 struct nvme_rdma_queue *queue = &ctrl->queues[idx];
682 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
684 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
687 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
689 if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
690 __nvme_rdma_stop_queue(queue);
691 dev_info(ctrl->ctrl.device,
692 "failed to connect queue: %d ret=%d\n", idx, ret);
697 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl,
702 for (i = first; i < last; i++) {
703 ret = nvme_rdma_start_queue(ctrl, i);
705 goto out_stop_queues;
711 for (i--; i >= first; i--)
712 nvme_rdma_stop_queue(&ctrl->queues[i]);
716 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
718 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
719 unsigned int nr_io_queues;
722 nr_io_queues = nvmf_nr_io_queues(opts);
723 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
727 if (nr_io_queues == 0) {
728 dev_err(ctrl->ctrl.device,
729 "unable to set any I/O queues\n");
733 ctrl->ctrl.queue_count = nr_io_queues + 1;
734 dev_info(ctrl->ctrl.device,
735 "creating %d I/O queues.\n", nr_io_queues);
737 nvmf_set_io_queues(opts, nr_io_queues, ctrl->io_queues);
738 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
739 ret = nvme_rdma_alloc_queue(ctrl, i,
740 ctrl->ctrl.sqsize + 1);
742 goto out_free_queues;
748 for (i--; i >= 1; i--)
749 nvme_rdma_free_queue(&ctrl->queues[i]);
754 static int nvme_rdma_alloc_tag_set(struct nvme_ctrl *ctrl)
756 unsigned int cmd_size = sizeof(struct nvme_rdma_request) +
757 NVME_RDMA_DATA_SGL_SIZE;
759 if (ctrl->max_integrity_segments)
760 cmd_size += sizeof(struct nvme_rdma_sgl) +
761 NVME_RDMA_METADATA_SGL_SIZE;
763 return nvme_alloc_io_tag_set(ctrl, &to_rdma_ctrl(ctrl)->tag_set,
765 ctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2,
769 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
771 if (ctrl->async_event_sqe.data) {
772 cancel_work_sync(&ctrl->ctrl.async_event_work);
773 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
774 sizeof(struct nvme_command), DMA_TO_DEVICE);
775 ctrl->async_event_sqe.data = NULL;
777 nvme_rdma_free_queue(&ctrl->queues[0]);
780 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
783 bool pi_capable = false;
786 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
790 ctrl->device = ctrl->queues[0].device;
791 ctrl->ctrl.numa_node = ibdev_to_node(ctrl->device->dev);
794 if (ctrl->device->dev->attrs.kernel_cap_flags &
795 IBK_INTEGRITY_HANDOVER)
798 ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev,
802 * Bind the async event SQE DMA mapping to the admin queue lifetime.
803 * It's safe, since any chage in the underlying RDMA device will issue
804 * error recovery and queue re-creation.
806 error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
807 sizeof(struct nvme_command), DMA_TO_DEVICE);
812 error = nvme_alloc_admin_tag_set(&ctrl->ctrl,
813 &ctrl->admin_tag_set, &nvme_rdma_admin_mq_ops,
814 sizeof(struct nvme_rdma_request) +
815 NVME_RDMA_DATA_SGL_SIZE);
817 goto out_free_async_qe;
821 error = nvme_rdma_start_queue(ctrl, 0);
823 goto out_remove_admin_tag_set;
825 error = nvme_enable_ctrl(&ctrl->ctrl);
829 ctrl->ctrl.max_segments = ctrl->max_fr_pages;
830 ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9);
832 ctrl->ctrl.max_integrity_segments = ctrl->max_fr_pages;
834 ctrl->ctrl.max_integrity_segments = 0;
836 nvme_unquiesce_admin_queue(&ctrl->ctrl);
838 error = nvme_init_ctrl_finish(&ctrl->ctrl, false);
840 goto out_quiesce_queue;
845 nvme_quiesce_admin_queue(&ctrl->ctrl);
846 blk_sync_queue(ctrl->ctrl.admin_q);
848 nvme_rdma_stop_queue(&ctrl->queues[0]);
849 nvme_cancel_admin_tagset(&ctrl->ctrl);
850 out_remove_admin_tag_set:
852 nvme_remove_admin_tag_set(&ctrl->ctrl);
854 if (ctrl->async_event_sqe.data) {
855 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
856 sizeof(struct nvme_command), DMA_TO_DEVICE);
857 ctrl->async_event_sqe.data = NULL;
860 nvme_rdma_free_queue(&ctrl->queues[0]);
864 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
868 ret = nvme_rdma_alloc_io_queues(ctrl);
873 ret = nvme_rdma_alloc_tag_set(&ctrl->ctrl);
875 goto out_free_io_queues;
879 * Only start IO queues for which we have allocated the tagset
880 * and limitted it to the available queues. On reconnects, the
881 * queue number might have changed.
883 nr_queues = min(ctrl->tag_set.nr_hw_queues + 1, ctrl->ctrl.queue_count);
884 ret = nvme_rdma_start_io_queues(ctrl, 1, nr_queues);
886 goto out_cleanup_tagset;
889 nvme_start_freeze(&ctrl->ctrl);
890 nvme_unquiesce_io_queues(&ctrl->ctrl);
891 if (!nvme_wait_freeze_timeout(&ctrl->ctrl, NVME_IO_TIMEOUT)) {
893 * If we timed out waiting for freeze we are likely to
894 * be stuck. Fail the controller initialization just
898 nvme_unfreeze(&ctrl->ctrl);
899 goto out_wait_freeze_timed_out;
901 blk_mq_update_nr_hw_queues(ctrl->ctrl.tagset,
902 ctrl->ctrl.queue_count - 1);
903 nvme_unfreeze(&ctrl->ctrl);
907 * If the number of queues has increased (reconnect case)
908 * start all new queues now.
910 ret = nvme_rdma_start_io_queues(ctrl, nr_queues,
911 ctrl->tag_set.nr_hw_queues + 1);
913 goto out_wait_freeze_timed_out;
917 out_wait_freeze_timed_out:
918 nvme_quiesce_io_queues(&ctrl->ctrl);
919 nvme_sync_io_queues(&ctrl->ctrl);
920 nvme_rdma_stop_io_queues(ctrl);
922 nvme_cancel_tagset(&ctrl->ctrl);
924 nvme_remove_io_tag_set(&ctrl->ctrl);
926 nvme_rdma_free_io_queues(ctrl);
930 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
933 nvme_quiesce_admin_queue(&ctrl->ctrl);
934 blk_sync_queue(ctrl->ctrl.admin_q);
935 nvme_rdma_stop_queue(&ctrl->queues[0]);
936 nvme_cancel_admin_tagset(&ctrl->ctrl);
938 nvme_unquiesce_admin_queue(&ctrl->ctrl);
939 nvme_remove_admin_tag_set(&ctrl->ctrl);
941 nvme_rdma_destroy_admin_queue(ctrl);
944 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
947 if (ctrl->ctrl.queue_count > 1) {
948 nvme_quiesce_io_queues(&ctrl->ctrl);
949 nvme_sync_io_queues(&ctrl->ctrl);
950 nvme_rdma_stop_io_queues(ctrl);
951 nvme_cancel_tagset(&ctrl->ctrl);
953 nvme_unquiesce_io_queues(&ctrl->ctrl);
954 nvme_remove_io_tag_set(&ctrl->ctrl);
956 nvme_rdma_free_io_queues(ctrl);
960 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
962 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
964 flush_work(&ctrl->err_work);
965 cancel_delayed_work_sync(&ctrl->reconnect_work);
968 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
970 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
972 if (list_empty(&ctrl->list))
975 mutex_lock(&nvme_rdma_ctrl_mutex);
976 list_del(&ctrl->list);
977 mutex_unlock(&nvme_rdma_ctrl_mutex);
979 nvmf_free_options(nctrl->opts);
985 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
987 enum nvme_ctrl_state state = nvme_ctrl_state(&ctrl->ctrl);
989 /* If we are resetting/deleting then do nothing */
990 if (state != NVME_CTRL_CONNECTING) {
991 WARN_ON_ONCE(state == NVME_CTRL_NEW || state == NVME_CTRL_LIVE);
995 if (nvmf_should_reconnect(&ctrl->ctrl)) {
996 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
997 ctrl->ctrl.opts->reconnect_delay);
998 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
999 ctrl->ctrl.opts->reconnect_delay * HZ);
1001 nvme_delete_ctrl(&ctrl->ctrl);
1005 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1010 ret = nvme_rdma_configure_admin_queue(ctrl, new);
1014 if (ctrl->ctrl.icdoff) {
1016 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1020 if (!(ctrl->ctrl.sgls & (1 << 2))) {
1022 dev_err(ctrl->ctrl.device,
1023 "Mandatory keyed sgls are not supported!\n");
1027 if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1028 dev_warn(ctrl->ctrl.device,
1029 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1030 ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1033 if (ctrl->ctrl.sqsize + 1 > NVME_RDMA_MAX_QUEUE_SIZE) {
1034 dev_warn(ctrl->ctrl.device,
1035 "ctrl sqsize %u > max queue size %u, clamping down\n",
1036 ctrl->ctrl.sqsize + 1, NVME_RDMA_MAX_QUEUE_SIZE);
1037 ctrl->ctrl.sqsize = NVME_RDMA_MAX_QUEUE_SIZE - 1;
1040 if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1041 dev_warn(ctrl->ctrl.device,
1042 "sqsize %u > ctrl maxcmd %u, clamping down\n",
1043 ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1044 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1047 if (ctrl->ctrl.sgls & (1 << 20))
1048 ctrl->use_inline_data = true;
1050 if (ctrl->ctrl.queue_count > 1) {
1051 ret = nvme_rdma_configure_io_queues(ctrl, new);
1056 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1059 * state change failure is ok if we started ctrl delete,
1060 * unless we're during creation of a new controller to
1061 * avoid races with teardown flow.
1063 enum nvme_ctrl_state state = nvme_ctrl_state(&ctrl->ctrl);
1065 WARN_ON_ONCE(state != NVME_CTRL_DELETING &&
1066 state != NVME_CTRL_DELETING_NOIO);
1072 nvme_start_ctrl(&ctrl->ctrl);
1076 if (ctrl->ctrl.queue_count > 1) {
1077 nvme_quiesce_io_queues(&ctrl->ctrl);
1078 nvme_sync_io_queues(&ctrl->ctrl);
1079 nvme_rdma_stop_io_queues(ctrl);
1080 nvme_cancel_tagset(&ctrl->ctrl);
1082 nvme_remove_io_tag_set(&ctrl->ctrl);
1083 nvme_rdma_free_io_queues(ctrl);
1086 nvme_stop_keep_alive(&ctrl->ctrl);
1087 nvme_quiesce_admin_queue(&ctrl->ctrl);
1088 blk_sync_queue(ctrl->ctrl.admin_q);
1089 nvme_rdma_stop_queue(&ctrl->queues[0]);
1090 nvme_cancel_admin_tagset(&ctrl->ctrl);
1092 nvme_remove_admin_tag_set(&ctrl->ctrl);
1093 nvme_rdma_destroy_admin_queue(ctrl);
1097 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1099 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1100 struct nvme_rdma_ctrl, reconnect_work);
1102 ++ctrl->ctrl.nr_reconnects;
1104 if (nvme_rdma_setup_ctrl(ctrl, false))
1107 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1108 ctrl->ctrl.nr_reconnects);
1110 ctrl->ctrl.nr_reconnects = 0;
1115 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1116 ctrl->ctrl.nr_reconnects);
1117 nvme_rdma_reconnect_or_remove(ctrl);
1120 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1122 struct nvme_rdma_ctrl *ctrl = container_of(work,
1123 struct nvme_rdma_ctrl, err_work);
1125 nvme_stop_keep_alive(&ctrl->ctrl);
1126 flush_work(&ctrl->ctrl.async_event_work);
1127 nvme_rdma_teardown_io_queues(ctrl, false);
1128 nvme_unquiesce_io_queues(&ctrl->ctrl);
1129 nvme_rdma_teardown_admin_queue(ctrl, false);
1130 nvme_unquiesce_admin_queue(&ctrl->ctrl);
1131 nvme_auth_stop(&ctrl->ctrl);
1133 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1134 /* state change failure is ok if we started ctrl delete */
1135 enum nvme_ctrl_state state = nvme_ctrl_state(&ctrl->ctrl);
1137 WARN_ON_ONCE(state != NVME_CTRL_DELETING &&
1138 state != NVME_CTRL_DELETING_NOIO);
1142 nvme_rdma_reconnect_or_remove(ctrl);
1145 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1147 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1150 dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1151 queue_work(nvme_reset_wq, &ctrl->err_work);
1154 static void nvme_rdma_end_request(struct nvme_rdma_request *req)
1156 struct request *rq = blk_mq_rq_from_pdu(req);
1158 if (!refcount_dec_and_test(&req->ref))
1160 if (!nvme_try_complete_req(rq, req->status, req->result))
1161 nvme_rdma_complete_rq(rq);
1164 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1167 struct nvme_rdma_queue *queue = wc->qp->qp_context;
1168 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1170 if (nvme_ctrl_state(&ctrl->ctrl) == NVME_CTRL_LIVE)
1171 dev_info(ctrl->ctrl.device,
1172 "%s for CQE 0x%p failed with status %s (%d)\n",
1174 ib_wc_status_msg(wc->status), wc->status);
1175 nvme_rdma_error_recovery(ctrl);
1178 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1180 if (unlikely(wc->status != IB_WC_SUCCESS))
1181 nvme_rdma_wr_error(cq, wc, "MEMREG");
1184 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1186 struct nvme_rdma_request *req =
1187 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1189 if (unlikely(wc->status != IB_WC_SUCCESS))
1190 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1192 nvme_rdma_end_request(req);
1195 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1196 struct nvme_rdma_request *req)
1198 struct ib_send_wr wr = {
1199 .opcode = IB_WR_LOCAL_INV,
1202 .send_flags = IB_SEND_SIGNALED,
1203 .ex.invalidate_rkey = req->mr->rkey,
1206 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1207 wr.wr_cqe = &req->reg_cqe;
1209 return ib_post_send(queue->qp, &wr, NULL);
1212 static void nvme_rdma_dma_unmap_req(struct ib_device *ibdev, struct request *rq)
1214 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1216 if (blk_integrity_rq(rq)) {
1217 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1218 req->metadata_sgl->nents, rq_dma_dir(rq));
1219 sg_free_table_chained(&req->metadata_sgl->sg_table,
1220 NVME_INLINE_METADATA_SG_CNT);
1223 ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1225 sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1228 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1231 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1232 struct nvme_rdma_device *dev = queue->device;
1233 struct ib_device *ibdev = dev->dev;
1234 struct list_head *pool = &queue->qp->rdma_mrs;
1236 if (!blk_rq_nr_phys_segments(rq))
1239 if (req->use_sig_mr)
1240 pool = &queue->qp->sig_mrs;
1243 ib_mr_pool_put(queue->qp, pool, req->mr);
1247 nvme_rdma_dma_unmap_req(ibdev, rq);
1250 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1252 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1255 put_unaligned_le24(0, sg->length);
1256 put_unaligned_le32(0, sg->key);
1257 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1261 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1262 struct nvme_rdma_request *req, struct nvme_command *c,
1265 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1266 struct ib_sge *sge = &req->sge[1];
1267 struct scatterlist *sgl;
1271 for_each_sg(req->data_sgl.sg_table.sgl, sgl, count, i) {
1272 sge->addr = sg_dma_address(sgl);
1273 sge->length = sg_dma_len(sgl);
1274 sge->lkey = queue->device->pd->local_dma_lkey;
1279 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1280 sg->length = cpu_to_le32(len);
1281 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1283 req->num_sge += count;
1287 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1288 struct nvme_rdma_request *req, struct nvme_command *c)
1290 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1292 sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1293 put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length);
1294 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1295 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1299 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1300 struct nvme_rdma_request *req, struct nvme_command *c,
1303 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1306 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1307 if (WARN_ON_ONCE(!req->mr))
1311 * Align the MR to a 4K page size to match the ctrl page size and
1312 * the block virtual boundary.
1314 nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL,
1316 if (unlikely(nr < count)) {
1317 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1324 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1326 req->reg_cqe.done = nvme_rdma_memreg_done;
1327 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1328 req->reg_wr.wr.opcode = IB_WR_REG_MR;
1329 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1330 req->reg_wr.wr.num_sge = 0;
1331 req->reg_wr.mr = req->mr;
1332 req->reg_wr.key = req->mr->rkey;
1333 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1334 IB_ACCESS_REMOTE_READ |
1335 IB_ACCESS_REMOTE_WRITE;
1337 sg->addr = cpu_to_le64(req->mr->iova);
1338 put_unaligned_le24(req->mr->length, sg->length);
1339 put_unaligned_le32(req->mr->rkey, sg->key);
1340 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1341 NVME_SGL_FMT_INVALIDATE;
1346 static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1347 struct nvme_command *cmd, struct ib_sig_domain *domain,
1348 u16 control, u8 pi_type)
1350 domain->sig_type = IB_SIG_TYPE_T10_DIF;
1351 domain->sig.dif.bg_type = IB_T10DIF_CRC;
1352 domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1353 domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1354 if (control & NVME_RW_PRINFO_PRCHK_REF)
1355 domain->sig.dif.ref_remap = true;
1357 domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
1358 domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
1359 domain->sig.dif.app_escape = true;
1360 if (pi_type == NVME_NS_DPS_PI_TYPE3)
1361 domain->sig.dif.ref_escape = true;
1364 static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1365 struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1368 u16 control = le16_to_cpu(cmd->rw.control);
1370 memset(sig_attrs, 0, sizeof(*sig_attrs));
1371 if (control & NVME_RW_PRINFO_PRACT) {
1372 /* for WRITE_INSERT/READ_STRIP no memory domain */
1373 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1374 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1376 /* Clear the PRACT bit since HCA will generate/verify the PI */
1377 control &= ~NVME_RW_PRINFO_PRACT;
1378 cmd->rw.control = cpu_to_le16(control);
1380 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1381 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1383 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
1388 static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1391 if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1392 *mask |= IB_SIG_CHECK_REFTAG;
1393 if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1394 *mask |= IB_SIG_CHECK_GUARD;
1397 static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1399 if (unlikely(wc->status != IB_WC_SUCCESS))
1400 nvme_rdma_wr_error(cq, wc, "SIG");
1403 static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1404 struct nvme_rdma_request *req, struct nvme_command *c,
1405 int count, int pi_count)
1407 struct nvme_rdma_sgl *sgl = &req->data_sgl;
1408 struct ib_reg_wr *wr = &req->reg_wr;
1409 struct request *rq = blk_mq_rq_from_pdu(req);
1410 struct nvme_ns *ns = rq->q->queuedata;
1411 struct bio *bio = rq->bio;
1412 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1415 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs);
1416 if (WARN_ON_ONCE(!req->mr))
1419 nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL,
1420 req->metadata_sgl->sg_table.sgl, pi_count, NULL,
1425 nvme_rdma_set_sig_attrs(blk_get_integrity(bio->bi_bdev->bd_disk), c,
1426 req->mr->sig_attrs, ns->pi_type);
1427 nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask);
1429 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1431 req->reg_cqe.done = nvme_rdma_sig_done;
1432 memset(wr, 0, sizeof(*wr));
1433 wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1434 wr->wr.wr_cqe = &req->reg_cqe;
1436 wr->wr.send_flags = 0;
1438 wr->key = req->mr->rkey;
1439 wr->access = IB_ACCESS_LOCAL_WRITE |
1440 IB_ACCESS_REMOTE_READ |
1441 IB_ACCESS_REMOTE_WRITE;
1443 sg->addr = cpu_to_le64(req->mr->iova);
1444 put_unaligned_le24(req->mr->length, sg->length);
1445 put_unaligned_le32(req->mr->rkey, sg->key);
1446 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1451 ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr);
1458 static int nvme_rdma_dma_map_req(struct ib_device *ibdev, struct request *rq,
1459 int *count, int *pi_count)
1461 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1464 req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1465 ret = sg_alloc_table_chained(&req->data_sgl.sg_table,
1466 blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl,
1467 NVME_INLINE_SG_CNT);
1471 req->data_sgl.nents = blk_rq_map_sg(rq->q, rq,
1472 req->data_sgl.sg_table.sgl);
1474 *count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl,
1475 req->data_sgl.nents, rq_dma_dir(rq));
1476 if (unlikely(*count <= 0)) {
1478 goto out_free_table;
1481 if (blk_integrity_rq(rq)) {
1482 req->metadata_sgl->sg_table.sgl =
1483 (struct scatterlist *)(req->metadata_sgl + 1);
1484 ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table,
1485 blk_rq_count_integrity_sg(rq->q, rq->bio),
1486 req->metadata_sgl->sg_table.sgl,
1487 NVME_INLINE_METADATA_SG_CNT);
1488 if (unlikely(ret)) {
1493 req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q,
1494 rq->bio, req->metadata_sgl->sg_table.sgl);
1495 *pi_count = ib_dma_map_sg(ibdev,
1496 req->metadata_sgl->sg_table.sgl,
1497 req->metadata_sgl->nents,
1499 if (unlikely(*pi_count <= 0)) {
1501 goto out_free_pi_table;
1508 sg_free_table_chained(&req->metadata_sgl->sg_table,
1509 NVME_INLINE_METADATA_SG_CNT);
1511 ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1514 sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1518 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1519 struct request *rq, struct nvme_command *c)
1521 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1522 struct nvme_rdma_device *dev = queue->device;
1523 struct ib_device *ibdev = dev->dev;
1528 refcount_set(&req->ref, 2); /* send and recv completions */
1530 c->common.flags |= NVME_CMD_SGL_METABUF;
1532 if (!blk_rq_nr_phys_segments(rq))
1533 return nvme_rdma_set_sg_null(c);
1535 ret = nvme_rdma_dma_map_req(ibdev, rq, &count, &pi_count);
1539 if (req->use_sig_mr) {
1540 ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1544 if (count <= dev->num_inline_segments) {
1545 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1546 queue->ctrl->use_inline_data &&
1547 blk_rq_payload_bytes(rq) <=
1548 nvme_rdma_inline_data_size(queue)) {
1549 ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1553 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1554 ret = nvme_rdma_map_sg_single(queue, req, c);
1559 ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1562 goto out_dma_unmap_req;
1567 nvme_rdma_dma_unmap_req(ibdev, rq);
1571 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1573 struct nvme_rdma_qe *qe =
1574 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1575 struct nvme_rdma_request *req =
1576 container_of(qe, struct nvme_rdma_request, sqe);
1578 if (unlikely(wc->status != IB_WC_SUCCESS))
1579 nvme_rdma_wr_error(cq, wc, "SEND");
1581 nvme_rdma_end_request(req);
1584 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1585 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1586 struct ib_send_wr *first)
1588 struct ib_send_wr wr;
1591 sge->addr = qe->dma;
1592 sge->length = sizeof(struct nvme_command);
1593 sge->lkey = queue->device->pd->local_dma_lkey;
1596 wr.wr_cqe = &qe->cqe;
1598 wr.num_sge = num_sge;
1599 wr.opcode = IB_WR_SEND;
1600 wr.send_flags = IB_SEND_SIGNALED;
1607 ret = ib_post_send(queue->qp, first, NULL);
1608 if (unlikely(ret)) {
1609 dev_err(queue->ctrl->ctrl.device,
1610 "%s failed with error code %d\n", __func__, ret);
1615 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1616 struct nvme_rdma_qe *qe)
1618 struct ib_recv_wr wr;
1622 list.addr = qe->dma;
1623 list.length = sizeof(struct nvme_completion);
1624 list.lkey = queue->device->pd->local_dma_lkey;
1626 qe->cqe.done = nvme_rdma_recv_done;
1629 wr.wr_cqe = &qe->cqe;
1633 ret = ib_post_recv(queue->qp, &wr, NULL);
1634 if (unlikely(ret)) {
1635 dev_err(queue->ctrl->ctrl.device,
1636 "%s failed with error code %d\n", __func__, ret);
1641 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1643 u32 queue_idx = nvme_rdma_queue_idx(queue);
1646 return queue->ctrl->admin_tag_set.tags[queue_idx];
1647 return queue->ctrl->tag_set.tags[queue_idx - 1];
1650 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1652 if (unlikely(wc->status != IB_WC_SUCCESS))
1653 nvme_rdma_wr_error(cq, wc, "ASYNC");
1656 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1658 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1659 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1660 struct ib_device *dev = queue->device->dev;
1661 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1662 struct nvme_command *cmd = sqe->data;
1666 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1668 memset(cmd, 0, sizeof(*cmd));
1669 cmd->common.opcode = nvme_admin_async_event;
1670 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1671 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1672 nvme_rdma_set_sg_null(cmd);
1674 sqe->cqe.done = nvme_rdma_async_done;
1676 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1679 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1683 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1684 struct nvme_completion *cqe, struct ib_wc *wc)
1687 struct nvme_rdma_request *req;
1689 rq = nvme_find_rq(nvme_rdma_tagset(queue), cqe->command_id);
1691 dev_err(queue->ctrl->ctrl.device,
1692 "got bad command_id %#x on QP %#x\n",
1693 cqe->command_id, queue->qp->qp_num);
1694 nvme_rdma_error_recovery(queue->ctrl);
1697 req = blk_mq_rq_to_pdu(rq);
1699 req->status = cqe->status;
1700 req->result = cqe->result;
1702 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1703 if (unlikely(!req->mr ||
1704 wc->ex.invalidate_rkey != req->mr->rkey)) {
1705 dev_err(queue->ctrl->ctrl.device,
1706 "Bogus remote invalidation for rkey %#x\n",
1707 req->mr ? req->mr->rkey : 0);
1708 nvme_rdma_error_recovery(queue->ctrl);
1710 } else if (req->mr) {
1713 ret = nvme_rdma_inv_rkey(queue, req);
1714 if (unlikely(ret < 0)) {
1715 dev_err(queue->ctrl->ctrl.device,
1716 "Queueing INV WR for rkey %#x failed (%d)\n",
1717 req->mr->rkey, ret);
1718 nvme_rdma_error_recovery(queue->ctrl);
1720 /* the local invalidation completion will end the request */
1724 nvme_rdma_end_request(req);
1727 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1729 struct nvme_rdma_qe *qe =
1730 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1731 struct nvme_rdma_queue *queue = wc->qp->qp_context;
1732 struct ib_device *ibdev = queue->device->dev;
1733 struct nvme_completion *cqe = qe->data;
1734 const size_t len = sizeof(struct nvme_completion);
1736 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1737 nvme_rdma_wr_error(cq, wc, "RECV");
1741 /* sanity checking for received data length */
1742 if (unlikely(wc->byte_len < len)) {
1743 dev_err(queue->ctrl->ctrl.device,
1744 "Unexpected nvme completion length(%d)\n", wc->byte_len);
1745 nvme_rdma_error_recovery(queue->ctrl);
1749 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1751 * AEN requests are special as they don't time out and can
1752 * survive any kind of queue freeze and often don't respond to
1753 * aborts. We don't even bother to allocate a struct request
1754 * for them but rather special case them here.
1756 if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1758 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1761 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1762 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1764 nvme_rdma_post_recv(queue, qe);
1767 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1771 for (i = 0; i < queue->queue_size; i++) {
1772 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1780 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1781 struct rdma_cm_event *ev)
1783 struct rdma_cm_id *cm_id = queue->cm_id;
1784 int status = ev->status;
1785 const char *rej_msg;
1786 const struct nvme_rdma_cm_rej *rej_data;
1789 rej_msg = rdma_reject_msg(cm_id, status);
1790 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1792 if (rej_data && rej_data_len >= sizeof(u16)) {
1793 u16 sts = le16_to_cpu(rej_data->sts);
1795 dev_err(queue->ctrl->ctrl.device,
1796 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1797 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1799 dev_err(queue->ctrl->ctrl.device,
1800 "Connect rejected: status %d (%s).\n", status, rej_msg);
1806 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1808 struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1811 ret = nvme_rdma_create_queue_ib(queue);
1815 if (ctrl->opts->tos >= 0)
1816 rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1817 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CM_TIMEOUT_MS);
1819 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1821 goto out_destroy_queue;
1827 nvme_rdma_destroy_queue_ib(queue);
1831 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1833 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1834 struct rdma_conn_param param = { };
1835 struct nvme_rdma_cm_req priv = { };
1838 param.qp_num = queue->qp->qp_num;
1839 param.flow_control = 1;
1841 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1842 /* maximum retry count */
1843 param.retry_count = 7;
1844 param.rnr_retry_count = 7;
1845 param.private_data = &priv;
1846 param.private_data_len = sizeof(priv);
1848 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1849 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1851 * set the admin queue depth to the minimum size
1852 * specified by the Fabrics standard.
1854 if (priv.qid == 0) {
1855 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1856 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1859 * current interpretation of the fabrics spec
1860 * is at minimum you make hrqsize sqsize+1, or a
1861 * 1's based representation of sqsize.
1863 priv.hrqsize = cpu_to_le16(queue->queue_size);
1864 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1867 ret = rdma_connect_locked(queue->cm_id, ¶m);
1869 dev_err(ctrl->ctrl.device,
1870 "rdma_connect_locked failed (%d).\n", ret);
1877 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1878 struct rdma_cm_event *ev)
1880 struct nvme_rdma_queue *queue = cm_id->context;
1883 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1884 rdma_event_msg(ev->event), ev->event,
1887 switch (ev->event) {
1888 case RDMA_CM_EVENT_ADDR_RESOLVED:
1889 cm_error = nvme_rdma_addr_resolved(queue);
1891 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1892 cm_error = nvme_rdma_route_resolved(queue);
1894 case RDMA_CM_EVENT_ESTABLISHED:
1895 queue->cm_error = nvme_rdma_conn_established(queue);
1896 /* complete cm_done regardless of success/failure */
1897 complete(&queue->cm_done);
1899 case RDMA_CM_EVENT_REJECTED:
1900 cm_error = nvme_rdma_conn_rejected(queue, ev);
1902 case RDMA_CM_EVENT_ROUTE_ERROR:
1903 case RDMA_CM_EVENT_CONNECT_ERROR:
1904 case RDMA_CM_EVENT_UNREACHABLE:
1905 case RDMA_CM_EVENT_ADDR_ERROR:
1906 dev_dbg(queue->ctrl->ctrl.device,
1907 "CM error event %d\n", ev->event);
1908 cm_error = -ECONNRESET;
1910 case RDMA_CM_EVENT_DISCONNECTED:
1911 case RDMA_CM_EVENT_ADDR_CHANGE:
1912 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1913 dev_dbg(queue->ctrl->ctrl.device,
1914 "disconnect received - connection closed\n");
1915 nvme_rdma_error_recovery(queue->ctrl);
1917 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1918 /* device removal is handled via the ib_client API */
1921 dev_err(queue->ctrl->ctrl.device,
1922 "Unexpected RDMA CM event (%d)\n", ev->event);
1923 nvme_rdma_error_recovery(queue->ctrl);
1928 queue->cm_error = cm_error;
1929 complete(&queue->cm_done);
1935 static void nvme_rdma_complete_timed_out(struct request *rq)
1937 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1938 struct nvme_rdma_queue *queue = req->queue;
1940 nvme_rdma_stop_queue(queue);
1941 nvmf_complete_timed_out_request(rq);
1944 static enum blk_eh_timer_return nvme_rdma_timeout(struct request *rq)
1946 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1947 struct nvme_rdma_queue *queue = req->queue;
1948 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1950 dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
1951 rq->tag, nvme_rdma_queue_idx(queue));
1953 if (nvme_ctrl_state(&ctrl->ctrl) != NVME_CTRL_LIVE) {
1955 * If we are resetting, connecting or deleting we should
1956 * complete immediately because we may block controller
1957 * teardown or setup sequence
1958 * - ctrl disable/shutdown fabrics requests
1959 * - connect requests
1960 * - initialization admin requests
1961 * - I/O requests that entered after unquiescing and
1962 * the controller stopped responding
1964 * All other requests should be cancelled by the error
1965 * recovery work, so it's fine that we fail it here.
1967 nvme_rdma_complete_timed_out(rq);
1972 * LIVE state should trigger the normal error recovery which will
1973 * handle completing this request.
1975 nvme_rdma_error_recovery(ctrl);
1976 return BLK_EH_RESET_TIMER;
1979 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1980 const struct blk_mq_queue_data *bd)
1982 struct nvme_ns *ns = hctx->queue->queuedata;
1983 struct nvme_rdma_queue *queue = hctx->driver_data;
1984 struct request *rq = bd->rq;
1985 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1986 struct nvme_rdma_qe *sqe = &req->sqe;
1987 struct nvme_command *c = nvme_req(rq)->cmd;
1988 struct ib_device *dev;
1989 bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1993 WARN_ON_ONCE(rq->tag < 0);
1995 if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1996 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
1998 dev = queue->device->dev;
2000 req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
2001 sizeof(struct nvme_command),
2003 err = ib_dma_mapping_error(dev, req->sqe.dma);
2005 return BLK_STS_RESOURCE;
2007 ib_dma_sync_single_for_cpu(dev, sqe->dma,
2008 sizeof(struct nvme_command), DMA_TO_DEVICE);
2010 ret = nvme_setup_cmd(ns, rq);
2014 nvme_start_request(rq);
2016 if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
2017 queue->pi_support &&
2018 (c->common.opcode == nvme_cmd_write ||
2019 c->common.opcode == nvme_cmd_read) &&
2021 req->use_sig_mr = true;
2023 req->use_sig_mr = false;
2025 err = nvme_rdma_map_data(queue, rq, c);
2026 if (unlikely(err < 0)) {
2027 dev_err(queue->ctrl->ctrl.device,
2028 "Failed to map data (%d)\n", err);
2032 sqe->cqe.done = nvme_rdma_send_done;
2034 ib_dma_sync_single_for_device(dev, sqe->dma,
2035 sizeof(struct nvme_command), DMA_TO_DEVICE);
2037 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
2038 req->mr ? &req->reg_wr.wr : NULL);
2045 nvme_rdma_unmap_data(queue, rq);
2048 ret = nvme_host_path_error(rq);
2049 else if (err == -ENOMEM || err == -EAGAIN)
2050 ret = BLK_STS_RESOURCE;
2052 ret = BLK_STS_IOERR;
2053 nvme_cleanup_cmd(rq);
2055 ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
2060 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2062 struct nvme_rdma_queue *queue = hctx->driver_data;
2064 return ib_process_cq_direct(queue->ib_cq, -1);
2067 static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req)
2069 struct request *rq = blk_mq_rq_from_pdu(req);
2070 struct ib_mr_status mr_status;
2073 ret = ib_check_mr_status(req->mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
2075 pr_err("ib_check_mr_status failed, ret %d\n", ret);
2076 nvme_req(rq)->status = NVME_SC_INVALID_PI;
2080 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
2081 switch (mr_status.sig_err.err_type) {
2082 case IB_SIG_BAD_GUARD:
2083 nvme_req(rq)->status = NVME_SC_GUARD_CHECK;
2085 case IB_SIG_BAD_REFTAG:
2086 nvme_req(rq)->status = NVME_SC_REFTAG_CHECK;
2088 case IB_SIG_BAD_APPTAG:
2089 nvme_req(rq)->status = NVME_SC_APPTAG_CHECK;
2092 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
2093 mr_status.sig_err.err_type, mr_status.sig_err.expected,
2094 mr_status.sig_err.actual);
2098 static void nvme_rdma_complete_rq(struct request *rq)
2100 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2101 struct nvme_rdma_queue *queue = req->queue;
2102 struct ib_device *ibdev = queue->device->dev;
2104 if (req->use_sig_mr)
2105 nvme_rdma_check_pi_status(req);
2107 nvme_rdma_unmap_data(queue, rq);
2108 ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
2110 nvme_complete_rq(rq);
2113 static void nvme_rdma_map_queues(struct blk_mq_tag_set *set)
2115 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(set->driver_data);
2117 nvmf_map_queues(set, &ctrl->ctrl, ctrl->io_queues);
2120 static const struct blk_mq_ops nvme_rdma_mq_ops = {
2121 .queue_rq = nvme_rdma_queue_rq,
2122 .complete = nvme_rdma_complete_rq,
2123 .init_request = nvme_rdma_init_request,
2124 .exit_request = nvme_rdma_exit_request,
2125 .init_hctx = nvme_rdma_init_hctx,
2126 .timeout = nvme_rdma_timeout,
2127 .map_queues = nvme_rdma_map_queues,
2128 .poll = nvme_rdma_poll,
2131 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
2132 .queue_rq = nvme_rdma_queue_rq,
2133 .complete = nvme_rdma_complete_rq,
2134 .init_request = nvme_rdma_init_request,
2135 .exit_request = nvme_rdma_exit_request,
2136 .init_hctx = nvme_rdma_init_admin_hctx,
2137 .timeout = nvme_rdma_timeout,
2140 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
2142 nvme_rdma_teardown_io_queues(ctrl, shutdown);
2143 nvme_quiesce_admin_queue(&ctrl->ctrl);
2144 nvme_disable_ctrl(&ctrl->ctrl, shutdown);
2145 nvme_rdma_teardown_admin_queue(ctrl, shutdown);
2148 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
2150 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
2153 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
2155 struct nvme_rdma_ctrl *ctrl =
2156 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
2158 nvme_stop_ctrl(&ctrl->ctrl);
2159 nvme_rdma_shutdown_ctrl(ctrl, false);
2161 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2162 /* state change failure should never happen */
2167 if (nvme_rdma_setup_ctrl(ctrl, false))
2173 ++ctrl->ctrl.nr_reconnects;
2174 nvme_rdma_reconnect_or_remove(ctrl);
2177 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
2179 .module = THIS_MODULE,
2180 .flags = NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED,
2181 .reg_read32 = nvmf_reg_read32,
2182 .reg_read64 = nvmf_reg_read64,
2183 .reg_write32 = nvmf_reg_write32,
2184 .free_ctrl = nvme_rdma_free_ctrl,
2185 .submit_async_event = nvme_rdma_submit_async_event,
2186 .delete_ctrl = nvme_rdma_delete_ctrl,
2187 .get_address = nvmf_get_address,
2188 .stop_ctrl = nvme_rdma_stop_ctrl,
2192 * Fails a connection request if it matches an existing controller
2193 * (association) with the same tuple:
2194 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
2196 * if local address is not specified in the request, it will match an
2197 * existing controller with all the other parameters the same and no
2198 * local port address specified as well.
2200 * The ports don't need to be compared as they are intrinsically
2201 * already matched by the port pointers supplied.
2204 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
2206 struct nvme_rdma_ctrl *ctrl;
2209 mutex_lock(&nvme_rdma_ctrl_mutex);
2210 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2211 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2215 mutex_unlock(&nvme_rdma_ctrl_mutex);
2220 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
2221 struct nvmf_ctrl_options *opts)
2223 struct nvme_rdma_ctrl *ctrl;
2227 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2229 return ERR_PTR(-ENOMEM);
2230 ctrl->ctrl.opts = opts;
2231 INIT_LIST_HEAD(&ctrl->list);
2233 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2235 kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
2236 if (!opts->trsvcid) {
2240 opts->mask |= NVMF_OPT_TRSVCID;
2243 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2244 opts->traddr, opts->trsvcid, &ctrl->addr);
2246 pr_err("malformed address passed: %s:%s\n",
2247 opts->traddr, opts->trsvcid);
2251 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2252 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2253 opts->host_traddr, NULL, &ctrl->src_addr);
2255 pr_err("malformed src address passed: %s\n",
2261 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2266 INIT_DELAYED_WORK(&ctrl->reconnect_work,
2267 nvme_rdma_reconnect_ctrl_work);
2268 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2269 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2271 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2272 opts->nr_poll_queues + 1;
2273 ctrl->ctrl.sqsize = opts->queue_size - 1;
2274 ctrl->ctrl.kato = opts->kato;
2277 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2282 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2283 0 /* no quirks, we're perfect! */);
2285 goto out_kfree_queues;
2287 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2288 WARN_ON_ONCE(!changed);
2290 ret = nvme_rdma_setup_ctrl(ctrl, true);
2292 goto out_uninit_ctrl;
2294 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2295 nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr);
2297 mutex_lock(&nvme_rdma_ctrl_mutex);
2298 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2299 mutex_unlock(&nvme_rdma_ctrl_mutex);
2304 nvme_uninit_ctrl(&ctrl->ctrl);
2305 nvme_put_ctrl(&ctrl->ctrl);
2308 return ERR_PTR(ret);
2310 kfree(ctrl->queues);
2313 return ERR_PTR(ret);
2316 static struct nvmf_transport_ops nvme_rdma_transport = {
2318 .module = THIS_MODULE,
2319 .required_opts = NVMF_OPT_TRADDR,
2320 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2321 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2322 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2324 .create_ctrl = nvme_rdma_create_ctrl,
2327 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2329 struct nvme_rdma_ctrl *ctrl;
2330 struct nvme_rdma_device *ndev;
2333 mutex_lock(&device_list_mutex);
2334 list_for_each_entry(ndev, &device_list, entry) {
2335 if (ndev->dev == ib_device) {
2340 mutex_unlock(&device_list_mutex);
2345 /* Delete all controllers using this device */
2346 mutex_lock(&nvme_rdma_ctrl_mutex);
2347 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2348 if (ctrl->device->dev != ib_device)
2350 nvme_delete_ctrl(&ctrl->ctrl);
2352 mutex_unlock(&nvme_rdma_ctrl_mutex);
2354 flush_workqueue(nvme_delete_wq);
2357 static struct ib_client nvme_rdma_ib_client = {
2358 .name = "nvme_rdma",
2359 .remove = nvme_rdma_remove_one
2362 static int __init nvme_rdma_init_module(void)
2366 ret = ib_register_client(&nvme_rdma_ib_client);
2370 ret = nvmf_register_transport(&nvme_rdma_transport);
2372 goto err_unreg_client;
2377 ib_unregister_client(&nvme_rdma_ib_client);
2381 static void __exit nvme_rdma_cleanup_module(void)
2383 struct nvme_rdma_ctrl *ctrl;
2385 nvmf_unregister_transport(&nvme_rdma_transport);
2386 ib_unregister_client(&nvme_rdma_ib_client);
2388 mutex_lock(&nvme_rdma_ctrl_mutex);
2389 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2390 nvme_delete_ctrl(&ctrl->ctrl);
2391 mutex_unlock(&nvme_rdma_ctrl_mutex);
2392 flush_workqueue(nvme_delete_wq);
2395 module_init(nvme_rdma_init_module);
2396 module_exit(nvme_rdma_cleanup_module);
2398 MODULE_LICENSE("GPL v2");