1 // SPDX-License-Identifier: GPL-2.0
3 * NVMe over Fabrics RDMA host code.
4 * Copyright (c) 2015-2016 HGST, a Western Digital Company.
6 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
7 #include <linux/module.h>
8 #include <linux/init.h>
9 #include <linux/slab.h>
10 #include <rdma/mr_pool.h>
11 #include <linux/err.h>
12 #include <linux/string.h>
13 #include <linux/atomic.h>
14 #include <linux/blk-mq.h>
15 #include <linux/blk-mq-rdma.h>
16 #include <linux/types.h>
17 #include <linux/list.h>
18 #include <linux/mutex.h>
19 #include <linux/scatterlist.h>
20 #include <linux/nvme.h>
21 #include <asm/unaligned.h>
23 #include <rdma/ib_verbs.h>
24 #include <rdma/rdma_cm.h>
25 #include <linux/nvme-rdma.h>
31 #define NVME_RDMA_CONNECT_TIMEOUT_MS 3000 /* 3 second */
33 #define NVME_RDMA_MAX_SEGMENTS 256
35 #define NVME_RDMA_MAX_INLINE_SEGMENTS 4
37 #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_timeout(&queue->cm_done,
251 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
256 WARN_ON_ONCE(queue->cm_error > 0);
257 return queue->cm_error;
260 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
262 struct nvme_rdma_device *dev = queue->device;
263 struct ib_qp_init_attr init_attr;
266 memset(&init_attr, 0, sizeof(init_attr));
267 init_attr.event_handler = nvme_rdma_qp_event;
269 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
271 init_attr.cap.max_recv_wr = queue->queue_size + 1;
272 init_attr.cap.max_recv_sge = 1;
273 init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
274 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
275 init_attr.qp_type = IB_QPT_RC;
276 init_attr.send_cq = queue->ib_cq;
277 init_attr.recv_cq = queue->ib_cq;
278 if (queue->pi_support)
279 init_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
280 init_attr.qp_context = queue;
282 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
284 queue->qp = queue->cm_id->qp;
288 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
289 struct request *rq, unsigned int hctx_idx)
291 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
293 kfree(req->sqe.data);
296 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
297 struct request *rq, unsigned int hctx_idx,
298 unsigned int numa_node)
300 struct nvme_rdma_ctrl *ctrl = set->driver_data;
301 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
302 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
303 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
305 nvme_req(rq)->ctrl = &ctrl->ctrl;
306 req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL);
310 /* metadata nvme_rdma_sgl struct is located after command's data SGL */
311 if (queue->pi_support)
312 req->metadata_sgl = (void *)nvme_req(rq) +
313 sizeof(struct nvme_rdma_request) +
314 NVME_RDMA_DATA_SGL_SIZE;
317 nvme_req(rq)->cmd = req->sqe.data;
322 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
323 unsigned int hctx_idx)
325 struct nvme_rdma_ctrl *ctrl = data;
326 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
328 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
330 hctx->driver_data = queue;
334 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
335 unsigned int hctx_idx)
337 struct nvme_rdma_ctrl *ctrl = data;
338 struct nvme_rdma_queue *queue = &ctrl->queues[0];
340 BUG_ON(hctx_idx != 0);
342 hctx->driver_data = queue;
346 static void nvme_rdma_free_dev(struct kref *ref)
348 struct nvme_rdma_device *ndev =
349 container_of(ref, struct nvme_rdma_device, ref);
351 mutex_lock(&device_list_mutex);
352 list_del(&ndev->entry);
353 mutex_unlock(&device_list_mutex);
355 ib_dealloc_pd(ndev->pd);
359 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
361 kref_put(&dev->ref, nvme_rdma_free_dev);
364 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
366 return kref_get_unless_zero(&dev->ref);
369 static struct nvme_rdma_device *
370 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
372 struct nvme_rdma_device *ndev;
374 mutex_lock(&device_list_mutex);
375 list_for_each_entry(ndev, &device_list, entry) {
376 if (ndev->dev->node_guid == cm_id->device->node_guid &&
377 nvme_rdma_dev_get(ndev))
381 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
385 ndev->dev = cm_id->device;
386 kref_init(&ndev->ref);
388 ndev->pd = ib_alloc_pd(ndev->dev,
389 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
390 if (IS_ERR(ndev->pd))
393 if (!(ndev->dev->attrs.device_cap_flags &
394 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
395 dev_err(&ndev->dev->dev,
396 "Memory registrations not supported.\n");
400 ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
401 ndev->dev->attrs.max_send_sge - 1);
402 list_add(&ndev->entry, &device_list);
404 mutex_unlock(&device_list_mutex);
408 ib_dealloc_pd(ndev->pd);
412 mutex_unlock(&device_list_mutex);
416 static void nvme_rdma_free_cq(struct nvme_rdma_queue *queue)
418 if (nvme_rdma_poll_queue(queue))
419 ib_free_cq(queue->ib_cq);
421 ib_cq_pool_put(queue->ib_cq, queue->cq_size);
424 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
426 struct nvme_rdma_device *dev;
427 struct ib_device *ibdev;
429 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
435 if (queue->pi_support)
436 ib_mr_pool_destroy(queue->qp, &queue->qp->sig_mrs);
437 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
440 * The cm_id object might have been destroyed during RDMA connection
441 * establishment error flow to avoid getting other cma events, thus
442 * the destruction of the QP shouldn't use rdma_cm API.
444 ib_destroy_qp(queue->qp);
445 nvme_rdma_free_cq(queue);
447 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
448 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
450 nvme_rdma_dev_put(dev);
453 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev, bool pi_support)
455 u32 max_page_list_len;
458 max_page_list_len = ibdev->attrs.max_pi_fast_reg_page_list_len;
460 max_page_list_len = ibdev->attrs.max_fast_reg_page_list_len;
462 return min_t(u32, NVME_RDMA_MAX_SEGMENTS, max_page_list_len - 1);
465 static int nvme_rdma_create_cq(struct ib_device *ibdev,
466 struct nvme_rdma_queue *queue)
468 int ret, comp_vector, idx = nvme_rdma_queue_idx(queue);
469 enum ib_poll_context poll_ctx;
472 * Spread I/O queues completion vectors according their queue index.
473 * Admin queues can always go on completion vector 0.
475 comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors;
477 /* Polling queues need direct cq polling context */
478 if (nvme_rdma_poll_queue(queue)) {
479 poll_ctx = IB_POLL_DIRECT;
480 queue->ib_cq = ib_alloc_cq(ibdev, queue, queue->cq_size,
481 comp_vector, poll_ctx);
483 poll_ctx = IB_POLL_SOFTIRQ;
484 queue->ib_cq = ib_cq_pool_get(ibdev, queue->cq_size,
485 comp_vector, poll_ctx);
488 if (IS_ERR(queue->ib_cq)) {
489 ret = PTR_ERR(queue->ib_cq);
496 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
498 struct ib_device *ibdev;
499 const int send_wr_factor = 3; /* MR, SEND, INV */
500 const int cq_factor = send_wr_factor + 1; /* + RECV */
501 int ret, pages_per_mr;
503 queue->device = nvme_rdma_find_get_device(queue->cm_id);
504 if (!queue->device) {
505 dev_err(queue->cm_id->device->dev.parent,
506 "no client data found!\n");
507 return -ECONNREFUSED;
509 ibdev = queue->device->dev;
511 /* +1 for ib_stop_cq */
512 queue->cq_size = cq_factor * queue->queue_size + 1;
514 ret = nvme_rdma_create_cq(ibdev, queue);
518 ret = nvme_rdma_create_qp(queue, send_wr_factor);
520 goto out_destroy_ib_cq;
522 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
523 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
524 if (!queue->rsp_ring) {
530 * Currently we don't use SG_GAPS MR's so if the first entry is
531 * misaligned we'll end up using two entries for a single data page,
532 * so one additional entry is required.
534 pages_per_mr = nvme_rdma_get_max_fr_pages(ibdev, queue->pi_support) + 1;
535 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
540 dev_err(queue->ctrl->ctrl.device,
541 "failed to initialize MR pool sized %d for QID %d\n",
542 queue->queue_size, nvme_rdma_queue_idx(queue));
543 goto out_destroy_ring;
546 if (queue->pi_support) {
547 ret = ib_mr_pool_init(queue->qp, &queue->qp->sig_mrs,
548 queue->queue_size, IB_MR_TYPE_INTEGRITY,
549 pages_per_mr, pages_per_mr);
551 dev_err(queue->ctrl->ctrl.device,
552 "failed to initialize PI MR pool sized %d for QID %d\n",
553 queue->queue_size, nvme_rdma_queue_idx(queue));
554 goto out_destroy_mr_pool;
558 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
563 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
565 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
566 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
568 rdma_destroy_qp(queue->cm_id);
570 nvme_rdma_free_cq(queue);
572 nvme_rdma_dev_put(queue->device);
576 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
577 int idx, size_t queue_size)
579 struct nvme_rdma_queue *queue;
580 struct sockaddr *src_addr = NULL;
583 queue = &ctrl->queues[idx];
584 mutex_init(&queue->queue_lock);
586 if (idx && ctrl->ctrl.max_integrity_segments)
587 queue->pi_support = true;
589 queue->pi_support = false;
590 init_completion(&queue->cm_done);
593 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
595 queue->cmnd_capsule_len = sizeof(struct nvme_command);
597 queue->queue_size = queue_size;
599 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
600 RDMA_PS_TCP, IB_QPT_RC);
601 if (IS_ERR(queue->cm_id)) {
602 dev_info(ctrl->ctrl.device,
603 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
604 ret = PTR_ERR(queue->cm_id);
605 goto out_destroy_mutex;
608 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
609 src_addr = (struct sockaddr *)&ctrl->src_addr;
611 queue->cm_error = -ETIMEDOUT;
612 ret = rdma_resolve_addr(queue->cm_id, src_addr,
613 (struct sockaddr *)&ctrl->addr,
614 NVME_RDMA_CONNECT_TIMEOUT_MS);
616 dev_info(ctrl->ctrl.device,
617 "rdma_resolve_addr failed (%d).\n", ret);
618 goto out_destroy_cm_id;
621 ret = nvme_rdma_wait_for_cm(queue);
623 dev_info(ctrl->ctrl.device,
624 "rdma connection establishment failed (%d)\n", ret);
625 goto out_destroy_cm_id;
628 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
633 rdma_destroy_id(queue->cm_id);
634 nvme_rdma_destroy_queue_ib(queue);
636 mutex_destroy(&queue->queue_lock);
640 static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
642 rdma_disconnect(queue->cm_id);
643 ib_drain_qp(queue->qp);
646 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
648 mutex_lock(&queue->queue_lock);
649 if (test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
650 __nvme_rdma_stop_queue(queue);
651 mutex_unlock(&queue->queue_lock);
654 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
656 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
659 nvme_rdma_destroy_queue_ib(queue);
660 rdma_destroy_id(queue->cm_id);
661 mutex_destroy(&queue->queue_lock);
664 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
668 for (i = 1; i < ctrl->ctrl.queue_count; i++)
669 nvme_rdma_free_queue(&ctrl->queues[i]);
672 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
676 for (i = 1; i < ctrl->ctrl.queue_count; i++)
677 nvme_rdma_stop_queue(&ctrl->queues[i]);
680 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
682 struct nvme_rdma_queue *queue = &ctrl->queues[idx];
686 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
688 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
691 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
693 if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
694 __nvme_rdma_stop_queue(queue);
695 dev_info(ctrl->ctrl.device,
696 "failed to connect queue: %d ret=%d\n", idx, ret);
701 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
705 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
706 ret = nvme_rdma_start_queue(ctrl, i);
708 goto out_stop_queues;
714 for (i--; i >= 1; i--)
715 nvme_rdma_stop_queue(&ctrl->queues[i]);
719 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
721 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
722 struct ib_device *ibdev = ctrl->device->dev;
723 unsigned int nr_io_queues, nr_default_queues;
724 unsigned int nr_read_queues, nr_poll_queues;
727 nr_read_queues = min_t(unsigned int, ibdev->num_comp_vectors,
728 min(opts->nr_io_queues, num_online_cpus()));
729 nr_default_queues = min_t(unsigned int, ibdev->num_comp_vectors,
730 min(opts->nr_write_queues, num_online_cpus()));
731 nr_poll_queues = min(opts->nr_poll_queues, num_online_cpus());
732 nr_io_queues = nr_read_queues + nr_default_queues + nr_poll_queues;
734 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
738 ctrl->ctrl.queue_count = nr_io_queues + 1;
739 if (ctrl->ctrl.queue_count < 2) {
740 dev_err(ctrl->ctrl.device,
741 "unable to set any I/O queues\n");
745 dev_info(ctrl->ctrl.device,
746 "creating %d I/O queues.\n", nr_io_queues);
748 if (opts->nr_write_queues && nr_read_queues < nr_io_queues) {
750 * separate read/write queues
751 * hand out dedicated default queues only after we have
752 * sufficient read queues.
754 ctrl->io_queues[HCTX_TYPE_READ] = nr_read_queues;
755 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
756 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
757 min(nr_default_queues, nr_io_queues);
758 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
761 * shared read/write queues
762 * either no write queues were requested, or we don't have
763 * sufficient queue count to have dedicated default queues.
765 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
766 min(nr_read_queues, nr_io_queues);
767 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
770 if (opts->nr_poll_queues && nr_io_queues) {
771 /* map dedicated poll queues only if we have queues left */
772 ctrl->io_queues[HCTX_TYPE_POLL] =
773 min(nr_poll_queues, nr_io_queues);
776 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
777 ret = nvme_rdma_alloc_queue(ctrl, i,
778 ctrl->ctrl.sqsize + 1);
780 goto out_free_queues;
786 for (i--; i >= 1; i--)
787 nvme_rdma_free_queue(&ctrl->queues[i]);
792 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
795 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
796 struct blk_mq_tag_set *set;
800 set = &ctrl->admin_tag_set;
801 memset(set, 0, sizeof(*set));
802 set->ops = &nvme_rdma_admin_mq_ops;
803 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
804 set->reserved_tags = NVMF_RESERVED_TAGS;
805 set->numa_node = nctrl->numa_node;
806 set->cmd_size = sizeof(struct nvme_rdma_request) +
807 NVME_RDMA_DATA_SGL_SIZE;
808 set->driver_data = ctrl;
809 set->nr_hw_queues = 1;
810 set->timeout = NVME_ADMIN_TIMEOUT;
811 set->flags = BLK_MQ_F_NO_SCHED;
813 set = &ctrl->tag_set;
814 memset(set, 0, sizeof(*set));
815 set->ops = &nvme_rdma_mq_ops;
816 set->queue_depth = nctrl->sqsize + 1;
817 set->reserved_tags = NVMF_RESERVED_TAGS;
818 set->numa_node = nctrl->numa_node;
819 set->flags = BLK_MQ_F_SHOULD_MERGE;
820 set->cmd_size = sizeof(struct nvme_rdma_request) +
821 NVME_RDMA_DATA_SGL_SIZE;
822 if (nctrl->max_integrity_segments)
823 set->cmd_size += sizeof(struct nvme_rdma_sgl) +
824 NVME_RDMA_METADATA_SGL_SIZE;
825 set->driver_data = ctrl;
826 set->nr_hw_queues = nctrl->queue_count - 1;
827 set->timeout = NVME_IO_TIMEOUT;
828 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
831 ret = blk_mq_alloc_tag_set(set);
838 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
842 blk_cleanup_queue(ctrl->ctrl.admin_q);
843 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
844 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
846 if (ctrl->async_event_sqe.data) {
847 cancel_work_sync(&ctrl->ctrl.async_event_work);
848 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
849 sizeof(struct nvme_command), DMA_TO_DEVICE);
850 ctrl->async_event_sqe.data = NULL;
852 nvme_rdma_free_queue(&ctrl->queues[0]);
855 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
858 bool pi_capable = false;
861 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
865 ctrl->device = ctrl->queues[0].device;
866 ctrl->ctrl.numa_node = ibdev_to_node(ctrl->device->dev);
869 if (ctrl->device->dev->attrs.device_cap_flags &
870 IB_DEVICE_INTEGRITY_HANDOVER)
873 ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev,
877 * Bind the async event SQE DMA mapping to the admin queue lifetime.
878 * It's safe, since any chage in the underlying RDMA device will issue
879 * error recovery and queue re-creation.
881 error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
882 sizeof(struct nvme_command), DMA_TO_DEVICE);
887 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
888 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
889 error = PTR_ERR(ctrl->ctrl.admin_tagset);
890 goto out_free_async_qe;
893 ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
894 if (IS_ERR(ctrl->ctrl.fabrics_q)) {
895 error = PTR_ERR(ctrl->ctrl.fabrics_q);
896 goto out_free_tagset;
899 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
900 if (IS_ERR(ctrl->ctrl.admin_q)) {
901 error = PTR_ERR(ctrl->ctrl.admin_q);
902 goto out_cleanup_fabrics_q;
906 error = nvme_rdma_start_queue(ctrl, 0);
908 goto out_cleanup_queue;
910 error = nvme_enable_ctrl(&ctrl->ctrl);
914 ctrl->ctrl.max_segments = ctrl->max_fr_pages;
915 ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9);
917 ctrl->ctrl.max_integrity_segments = ctrl->max_fr_pages;
919 ctrl->ctrl.max_integrity_segments = 0;
921 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
923 error = nvme_init_ctrl_finish(&ctrl->ctrl);
925 goto out_quiesce_queue;
930 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
931 blk_sync_queue(ctrl->ctrl.admin_q);
933 nvme_rdma_stop_queue(&ctrl->queues[0]);
934 nvme_cancel_admin_tagset(&ctrl->ctrl);
937 blk_cleanup_queue(ctrl->ctrl.admin_q);
938 out_cleanup_fabrics_q:
940 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
943 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
945 if (ctrl->async_event_sqe.data) {
946 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
947 sizeof(struct nvme_command), DMA_TO_DEVICE);
948 ctrl->async_event_sqe.data = NULL;
951 nvme_rdma_free_queue(&ctrl->queues[0]);
955 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
959 blk_cleanup_queue(ctrl->ctrl.connect_q);
960 blk_mq_free_tag_set(ctrl->ctrl.tagset);
962 nvme_rdma_free_io_queues(ctrl);
965 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
969 ret = nvme_rdma_alloc_io_queues(ctrl);
974 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
975 if (IS_ERR(ctrl->ctrl.tagset)) {
976 ret = PTR_ERR(ctrl->ctrl.tagset);
977 goto out_free_io_queues;
980 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
981 if (IS_ERR(ctrl->ctrl.connect_q)) {
982 ret = PTR_ERR(ctrl->ctrl.connect_q);
983 goto out_free_tag_set;
987 ret = nvme_rdma_start_io_queues(ctrl);
989 goto out_cleanup_connect_q;
992 nvme_start_queues(&ctrl->ctrl);
993 if (!nvme_wait_freeze_timeout(&ctrl->ctrl, NVME_IO_TIMEOUT)) {
995 * If we timed out waiting for freeze we are likely to
996 * be stuck. Fail the controller initialization just
1000 goto out_wait_freeze_timed_out;
1002 blk_mq_update_nr_hw_queues(ctrl->ctrl.tagset,
1003 ctrl->ctrl.queue_count - 1);
1004 nvme_unfreeze(&ctrl->ctrl);
1009 out_wait_freeze_timed_out:
1010 nvme_stop_queues(&ctrl->ctrl);
1011 nvme_sync_io_queues(&ctrl->ctrl);
1012 nvme_rdma_stop_io_queues(ctrl);
1013 out_cleanup_connect_q:
1014 nvme_cancel_tagset(&ctrl->ctrl);
1016 blk_cleanup_queue(ctrl->ctrl.connect_q);
1019 blk_mq_free_tag_set(ctrl->ctrl.tagset);
1021 nvme_rdma_free_io_queues(ctrl);
1025 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
1028 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1029 blk_sync_queue(ctrl->ctrl.admin_q);
1030 nvme_rdma_stop_queue(&ctrl->queues[0]);
1031 nvme_cancel_admin_tagset(&ctrl->ctrl);
1033 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1034 nvme_rdma_destroy_admin_queue(ctrl, remove);
1037 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
1040 if (ctrl->ctrl.queue_count > 1) {
1041 nvme_start_freeze(&ctrl->ctrl);
1042 nvme_stop_queues(&ctrl->ctrl);
1043 nvme_sync_io_queues(&ctrl->ctrl);
1044 nvme_rdma_stop_io_queues(ctrl);
1045 nvme_cancel_tagset(&ctrl->ctrl);
1047 nvme_start_queues(&ctrl->ctrl);
1048 nvme_rdma_destroy_io_queues(ctrl, remove);
1052 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
1054 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1056 if (list_empty(&ctrl->list))
1059 mutex_lock(&nvme_rdma_ctrl_mutex);
1060 list_del(&ctrl->list);
1061 mutex_unlock(&nvme_rdma_ctrl_mutex);
1063 nvmf_free_options(nctrl->opts);
1065 kfree(ctrl->queues);
1069 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
1071 /* If we are resetting/deleting then do nothing */
1072 if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
1073 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
1074 ctrl->ctrl.state == NVME_CTRL_LIVE);
1078 if (nvmf_should_reconnect(&ctrl->ctrl)) {
1079 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
1080 ctrl->ctrl.opts->reconnect_delay);
1081 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
1082 ctrl->ctrl.opts->reconnect_delay * HZ);
1084 nvme_delete_ctrl(&ctrl->ctrl);
1088 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1093 ret = nvme_rdma_configure_admin_queue(ctrl, new);
1097 if (ctrl->ctrl.icdoff) {
1098 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1102 if (!(ctrl->ctrl.sgls & (1 << 2))) {
1103 dev_err(ctrl->ctrl.device,
1104 "Mandatory keyed sgls are not supported!\n");
1108 if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1109 dev_warn(ctrl->ctrl.device,
1110 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1111 ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1114 if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1115 dev_warn(ctrl->ctrl.device,
1116 "sqsize %u > ctrl maxcmd %u, clamping down\n",
1117 ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1118 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1121 if (ctrl->ctrl.sgls & (1 << 20))
1122 ctrl->use_inline_data = true;
1124 if (ctrl->ctrl.queue_count > 1) {
1125 ret = nvme_rdma_configure_io_queues(ctrl, new);
1130 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1133 * state change failure is ok if we started ctrl delete,
1134 * unless we're during creation of a new controller to
1135 * avoid races with teardown flow.
1137 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1138 ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1144 nvme_start_ctrl(&ctrl->ctrl);
1148 if (ctrl->ctrl.queue_count > 1) {
1149 nvme_stop_queues(&ctrl->ctrl);
1150 nvme_sync_io_queues(&ctrl->ctrl);
1151 nvme_rdma_stop_io_queues(ctrl);
1152 nvme_cancel_tagset(&ctrl->ctrl);
1153 nvme_rdma_destroy_io_queues(ctrl, new);
1156 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1157 blk_sync_queue(ctrl->ctrl.admin_q);
1158 nvme_rdma_stop_queue(&ctrl->queues[0]);
1159 nvme_cancel_admin_tagset(&ctrl->ctrl);
1160 nvme_rdma_destroy_admin_queue(ctrl, new);
1164 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1166 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1167 struct nvme_rdma_ctrl, reconnect_work);
1169 ++ctrl->ctrl.nr_reconnects;
1171 if (nvme_rdma_setup_ctrl(ctrl, false))
1174 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1175 ctrl->ctrl.nr_reconnects);
1177 ctrl->ctrl.nr_reconnects = 0;
1182 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1183 ctrl->ctrl.nr_reconnects);
1184 nvme_rdma_reconnect_or_remove(ctrl);
1187 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1189 struct nvme_rdma_ctrl *ctrl = container_of(work,
1190 struct nvme_rdma_ctrl, err_work);
1192 nvme_stop_keep_alive(&ctrl->ctrl);
1193 nvme_rdma_teardown_io_queues(ctrl, false);
1194 nvme_start_queues(&ctrl->ctrl);
1195 nvme_rdma_teardown_admin_queue(ctrl, false);
1196 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1198 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1199 /* state change failure is ok if we started ctrl delete */
1200 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1201 ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1205 nvme_rdma_reconnect_or_remove(ctrl);
1208 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1210 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1213 dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1214 queue_work(nvme_reset_wq, &ctrl->err_work);
1217 static void nvme_rdma_end_request(struct nvme_rdma_request *req)
1219 struct request *rq = blk_mq_rq_from_pdu(req);
1221 if (!refcount_dec_and_test(&req->ref))
1223 if (!nvme_try_complete_req(rq, req->status, req->result))
1224 nvme_rdma_complete_rq(rq);
1227 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1230 struct nvme_rdma_queue *queue = wc->qp->qp_context;
1231 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1233 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1234 dev_info(ctrl->ctrl.device,
1235 "%s for CQE 0x%p failed with status %s (%d)\n",
1237 ib_wc_status_msg(wc->status), wc->status);
1238 nvme_rdma_error_recovery(ctrl);
1241 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1243 if (unlikely(wc->status != IB_WC_SUCCESS))
1244 nvme_rdma_wr_error(cq, wc, "MEMREG");
1247 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1249 struct nvme_rdma_request *req =
1250 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1252 if (unlikely(wc->status != IB_WC_SUCCESS))
1253 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1255 nvme_rdma_end_request(req);
1258 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1259 struct nvme_rdma_request *req)
1261 struct ib_send_wr wr = {
1262 .opcode = IB_WR_LOCAL_INV,
1265 .send_flags = IB_SEND_SIGNALED,
1266 .ex.invalidate_rkey = req->mr->rkey,
1269 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1270 wr.wr_cqe = &req->reg_cqe;
1272 return ib_post_send(queue->qp, &wr, NULL);
1275 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1278 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1279 struct nvme_rdma_device *dev = queue->device;
1280 struct ib_device *ibdev = dev->dev;
1281 struct list_head *pool = &queue->qp->rdma_mrs;
1283 if (!blk_rq_nr_phys_segments(rq))
1286 if (blk_integrity_rq(rq)) {
1287 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1288 req->metadata_sgl->nents, rq_dma_dir(rq));
1289 sg_free_table_chained(&req->metadata_sgl->sg_table,
1290 NVME_INLINE_METADATA_SG_CNT);
1293 if (req->use_sig_mr)
1294 pool = &queue->qp->sig_mrs;
1297 ib_mr_pool_put(queue->qp, pool, req->mr);
1301 ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1303 sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1306 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1308 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1311 put_unaligned_le24(0, sg->length);
1312 put_unaligned_le32(0, sg->key);
1313 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1317 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1318 struct nvme_rdma_request *req, struct nvme_command *c,
1321 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1322 struct ib_sge *sge = &req->sge[1];
1323 struct scatterlist *sgl;
1327 for_each_sg(req->data_sgl.sg_table.sgl, sgl, count, i) {
1328 sge->addr = sg_dma_address(sgl);
1329 sge->length = sg_dma_len(sgl);
1330 sge->lkey = queue->device->pd->local_dma_lkey;
1335 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1336 sg->length = cpu_to_le32(len);
1337 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1339 req->num_sge += count;
1343 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1344 struct nvme_rdma_request *req, struct nvme_command *c)
1346 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1348 sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1349 put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length);
1350 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1351 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1355 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1356 struct nvme_rdma_request *req, struct nvme_command *c,
1359 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1362 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1363 if (WARN_ON_ONCE(!req->mr))
1367 * Align the MR to a 4K page size to match the ctrl page size and
1368 * the block virtual boundary.
1370 nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL,
1372 if (unlikely(nr < count)) {
1373 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1380 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1382 req->reg_cqe.done = nvme_rdma_memreg_done;
1383 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1384 req->reg_wr.wr.opcode = IB_WR_REG_MR;
1385 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1386 req->reg_wr.wr.num_sge = 0;
1387 req->reg_wr.mr = req->mr;
1388 req->reg_wr.key = req->mr->rkey;
1389 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1390 IB_ACCESS_REMOTE_READ |
1391 IB_ACCESS_REMOTE_WRITE;
1393 sg->addr = cpu_to_le64(req->mr->iova);
1394 put_unaligned_le24(req->mr->length, sg->length);
1395 put_unaligned_le32(req->mr->rkey, sg->key);
1396 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1397 NVME_SGL_FMT_INVALIDATE;
1402 static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1403 struct nvme_command *cmd, struct ib_sig_domain *domain,
1404 u16 control, u8 pi_type)
1406 domain->sig_type = IB_SIG_TYPE_T10_DIF;
1407 domain->sig.dif.bg_type = IB_T10DIF_CRC;
1408 domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1409 domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1410 if (control & NVME_RW_PRINFO_PRCHK_REF)
1411 domain->sig.dif.ref_remap = true;
1413 domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
1414 domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
1415 domain->sig.dif.app_escape = true;
1416 if (pi_type == NVME_NS_DPS_PI_TYPE3)
1417 domain->sig.dif.ref_escape = true;
1420 static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1421 struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1424 u16 control = le16_to_cpu(cmd->rw.control);
1426 memset(sig_attrs, 0, sizeof(*sig_attrs));
1427 if (control & NVME_RW_PRINFO_PRACT) {
1428 /* for WRITE_INSERT/READ_STRIP no memory domain */
1429 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1430 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1432 /* Clear the PRACT bit since HCA will generate/verify the PI */
1433 control &= ~NVME_RW_PRINFO_PRACT;
1434 cmd->rw.control = cpu_to_le16(control);
1436 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1437 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1439 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
1444 static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1447 if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1448 *mask |= IB_SIG_CHECK_REFTAG;
1449 if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1450 *mask |= IB_SIG_CHECK_GUARD;
1453 static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1455 if (unlikely(wc->status != IB_WC_SUCCESS))
1456 nvme_rdma_wr_error(cq, wc, "SIG");
1459 static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1460 struct nvme_rdma_request *req, struct nvme_command *c,
1461 int count, int pi_count)
1463 struct nvme_rdma_sgl *sgl = &req->data_sgl;
1464 struct ib_reg_wr *wr = &req->reg_wr;
1465 struct request *rq = blk_mq_rq_from_pdu(req);
1466 struct nvme_ns *ns = rq->q->queuedata;
1467 struct bio *bio = rq->bio;
1468 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1471 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs);
1472 if (WARN_ON_ONCE(!req->mr))
1475 nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL,
1476 req->metadata_sgl->sg_table.sgl, pi_count, NULL,
1481 nvme_rdma_set_sig_attrs(blk_get_integrity(bio->bi_bdev->bd_disk), c,
1482 req->mr->sig_attrs, ns->pi_type);
1483 nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask);
1485 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1487 req->reg_cqe.done = nvme_rdma_sig_done;
1488 memset(wr, 0, sizeof(*wr));
1489 wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1490 wr->wr.wr_cqe = &req->reg_cqe;
1492 wr->wr.send_flags = 0;
1494 wr->key = req->mr->rkey;
1495 wr->access = IB_ACCESS_LOCAL_WRITE |
1496 IB_ACCESS_REMOTE_READ |
1497 IB_ACCESS_REMOTE_WRITE;
1499 sg->addr = cpu_to_le64(req->mr->iova);
1500 put_unaligned_le24(req->mr->length, sg->length);
1501 put_unaligned_le32(req->mr->rkey, sg->key);
1502 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1507 ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr);
1514 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1515 struct request *rq, struct nvme_command *c)
1517 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1518 struct nvme_rdma_device *dev = queue->device;
1519 struct ib_device *ibdev = dev->dev;
1524 refcount_set(&req->ref, 2); /* send and recv completions */
1526 c->common.flags |= NVME_CMD_SGL_METABUF;
1528 if (!blk_rq_nr_phys_segments(rq))
1529 return nvme_rdma_set_sg_null(c);
1531 req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1532 ret = sg_alloc_table_chained(&req->data_sgl.sg_table,
1533 blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl,
1534 NVME_INLINE_SG_CNT);
1538 req->data_sgl.nents = blk_rq_map_sg(rq->q, rq,
1539 req->data_sgl.sg_table.sgl);
1541 count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl,
1542 req->data_sgl.nents, rq_dma_dir(rq));
1543 if (unlikely(count <= 0)) {
1545 goto out_free_table;
1548 if (blk_integrity_rq(rq)) {
1549 req->metadata_sgl->sg_table.sgl =
1550 (struct scatterlist *)(req->metadata_sgl + 1);
1551 ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table,
1552 blk_rq_count_integrity_sg(rq->q, rq->bio),
1553 req->metadata_sgl->sg_table.sgl,
1554 NVME_INLINE_METADATA_SG_CNT);
1555 if (unlikely(ret)) {
1560 req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q,
1561 rq->bio, req->metadata_sgl->sg_table.sgl);
1562 pi_count = ib_dma_map_sg(ibdev,
1563 req->metadata_sgl->sg_table.sgl,
1564 req->metadata_sgl->nents,
1566 if (unlikely(pi_count <= 0)) {
1568 goto out_free_pi_table;
1572 if (req->use_sig_mr) {
1573 ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1577 if (count <= dev->num_inline_segments) {
1578 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1579 queue->ctrl->use_inline_data &&
1580 blk_rq_payload_bytes(rq) <=
1581 nvme_rdma_inline_data_size(queue)) {
1582 ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1586 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1587 ret = nvme_rdma_map_sg_single(queue, req, c);
1592 ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1595 goto out_unmap_pi_sg;
1600 if (blk_integrity_rq(rq))
1601 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1602 req->metadata_sgl->nents, rq_dma_dir(rq));
1604 if (blk_integrity_rq(rq))
1605 sg_free_table_chained(&req->metadata_sgl->sg_table,
1606 NVME_INLINE_METADATA_SG_CNT);
1608 ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1611 sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1615 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1617 struct nvme_rdma_qe *qe =
1618 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1619 struct nvme_rdma_request *req =
1620 container_of(qe, struct nvme_rdma_request, sqe);
1622 if (unlikely(wc->status != IB_WC_SUCCESS))
1623 nvme_rdma_wr_error(cq, wc, "SEND");
1625 nvme_rdma_end_request(req);
1628 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1629 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1630 struct ib_send_wr *first)
1632 struct ib_send_wr wr;
1635 sge->addr = qe->dma;
1636 sge->length = sizeof(struct nvme_command);
1637 sge->lkey = queue->device->pd->local_dma_lkey;
1640 wr.wr_cqe = &qe->cqe;
1642 wr.num_sge = num_sge;
1643 wr.opcode = IB_WR_SEND;
1644 wr.send_flags = IB_SEND_SIGNALED;
1651 ret = ib_post_send(queue->qp, first, NULL);
1652 if (unlikely(ret)) {
1653 dev_err(queue->ctrl->ctrl.device,
1654 "%s failed with error code %d\n", __func__, ret);
1659 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1660 struct nvme_rdma_qe *qe)
1662 struct ib_recv_wr wr;
1666 list.addr = qe->dma;
1667 list.length = sizeof(struct nvme_completion);
1668 list.lkey = queue->device->pd->local_dma_lkey;
1670 qe->cqe.done = nvme_rdma_recv_done;
1673 wr.wr_cqe = &qe->cqe;
1677 ret = ib_post_recv(queue->qp, &wr, NULL);
1678 if (unlikely(ret)) {
1679 dev_err(queue->ctrl->ctrl.device,
1680 "%s failed with error code %d\n", __func__, ret);
1685 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1687 u32 queue_idx = nvme_rdma_queue_idx(queue);
1690 return queue->ctrl->admin_tag_set.tags[queue_idx];
1691 return queue->ctrl->tag_set.tags[queue_idx - 1];
1694 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1696 if (unlikely(wc->status != IB_WC_SUCCESS))
1697 nvme_rdma_wr_error(cq, wc, "ASYNC");
1700 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1702 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1703 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1704 struct ib_device *dev = queue->device->dev;
1705 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1706 struct nvme_command *cmd = sqe->data;
1710 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1712 memset(cmd, 0, sizeof(*cmd));
1713 cmd->common.opcode = nvme_admin_async_event;
1714 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1715 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1716 nvme_rdma_set_sg_null(cmd);
1718 sqe->cqe.done = nvme_rdma_async_done;
1720 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1723 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1727 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1728 struct nvme_completion *cqe, struct ib_wc *wc)
1731 struct nvme_rdma_request *req;
1733 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1735 dev_err(queue->ctrl->ctrl.device,
1736 "tag 0x%x on QP %#x not found\n",
1737 cqe->command_id, queue->qp->qp_num);
1738 nvme_rdma_error_recovery(queue->ctrl);
1741 req = blk_mq_rq_to_pdu(rq);
1743 req->status = cqe->status;
1744 req->result = cqe->result;
1746 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1747 if (unlikely(!req->mr ||
1748 wc->ex.invalidate_rkey != req->mr->rkey)) {
1749 dev_err(queue->ctrl->ctrl.device,
1750 "Bogus remote invalidation for rkey %#x\n",
1751 req->mr ? req->mr->rkey : 0);
1752 nvme_rdma_error_recovery(queue->ctrl);
1754 } else if (req->mr) {
1757 ret = nvme_rdma_inv_rkey(queue, req);
1758 if (unlikely(ret < 0)) {
1759 dev_err(queue->ctrl->ctrl.device,
1760 "Queueing INV WR for rkey %#x failed (%d)\n",
1761 req->mr->rkey, ret);
1762 nvme_rdma_error_recovery(queue->ctrl);
1764 /* the local invalidation completion will end the request */
1768 nvme_rdma_end_request(req);
1771 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1773 struct nvme_rdma_qe *qe =
1774 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1775 struct nvme_rdma_queue *queue = wc->qp->qp_context;
1776 struct ib_device *ibdev = queue->device->dev;
1777 struct nvme_completion *cqe = qe->data;
1778 const size_t len = sizeof(struct nvme_completion);
1780 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1781 nvme_rdma_wr_error(cq, wc, "RECV");
1785 /* sanity checking for received data length */
1786 if (unlikely(wc->byte_len < len)) {
1787 dev_err(queue->ctrl->ctrl.device,
1788 "Unexpected nvme completion length(%d)\n", wc->byte_len);
1789 nvme_rdma_error_recovery(queue->ctrl);
1793 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1795 * AEN requests are special as they don't time out and can
1796 * survive any kind of queue freeze and often don't respond to
1797 * aborts. We don't even bother to allocate a struct request
1798 * for them but rather special case them here.
1800 if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1802 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1805 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1806 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1808 nvme_rdma_post_recv(queue, qe);
1811 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1815 for (i = 0; i < queue->queue_size; i++) {
1816 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1818 goto out_destroy_queue_ib;
1823 out_destroy_queue_ib:
1824 nvme_rdma_destroy_queue_ib(queue);
1828 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1829 struct rdma_cm_event *ev)
1831 struct rdma_cm_id *cm_id = queue->cm_id;
1832 int status = ev->status;
1833 const char *rej_msg;
1834 const struct nvme_rdma_cm_rej *rej_data;
1837 rej_msg = rdma_reject_msg(cm_id, status);
1838 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1840 if (rej_data && rej_data_len >= sizeof(u16)) {
1841 u16 sts = le16_to_cpu(rej_data->sts);
1843 dev_err(queue->ctrl->ctrl.device,
1844 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1845 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1847 dev_err(queue->ctrl->ctrl.device,
1848 "Connect rejected: status %d (%s).\n", status, rej_msg);
1854 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1856 struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1859 ret = nvme_rdma_create_queue_ib(queue);
1863 if (ctrl->opts->tos >= 0)
1864 rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1865 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1867 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1869 goto out_destroy_queue;
1875 nvme_rdma_destroy_queue_ib(queue);
1879 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1881 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1882 struct rdma_conn_param param = { };
1883 struct nvme_rdma_cm_req priv = { };
1886 param.qp_num = queue->qp->qp_num;
1887 param.flow_control = 1;
1889 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1890 /* maximum retry count */
1891 param.retry_count = 7;
1892 param.rnr_retry_count = 7;
1893 param.private_data = &priv;
1894 param.private_data_len = sizeof(priv);
1896 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1897 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1899 * set the admin queue depth to the minimum size
1900 * specified by the Fabrics standard.
1902 if (priv.qid == 0) {
1903 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1904 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1907 * current interpretation of the fabrics spec
1908 * is at minimum you make hrqsize sqsize+1, or a
1909 * 1's based representation of sqsize.
1911 priv.hrqsize = cpu_to_le16(queue->queue_size);
1912 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1915 ret = rdma_connect_locked(queue->cm_id, ¶m);
1917 dev_err(ctrl->ctrl.device,
1918 "rdma_connect_locked failed (%d).\n", ret);
1919 goto out_destroy_queue_ib;
1924 out_destroy_queue_ib:
1925 nvme_rdma_destroy_queue_ib(queue);
1929 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1930 struct rdma_cm_event *ev)
1932 struct nvme_rdma_queue *queue = cm_id->context;
1935 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1936 rdma_event_msg(ev->event), ev->event,
1939 switch (ev->event) {
1940 case RDMA_CM_EVENT_ADDR_RESOLVED:
1941 cm_error = nvme_rdma_addr_resolved(queue);
1943 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1944 cm_error = nvme_rdma_route_resolved(queue);
1946 case RDMA_CM_EVENT_ESTABLISHED:
1947 queue->cm_error = nvme_rdma_conn_established(queue);
1948 /* complete cm_done regardless of success/failure */
1949 complete(&queue->cm_done);
1951 case RDMA_CM_EVENT_REJECTED:
1952 cm_error = nvme_rdma_conn_rejected(queue, ev);
1954 case RDMA_CM_EVENT_ROUTE_ERROR:
1955 case RDMA_CM_EVENT_CONNECT_ERROR:
1956 case RDMA_CM_EVENT_UNREACHABLE:
1957 nvme_rdma_destroy_queue_ib(queue);
1959 case RDMA_CM_EVENT_ADDR_ERROR:
1960 dev_dbg(queue->ctrl->ctrl.device,
1961 "CM error event %d\n", ev->event);
1962 cm_error = -ECONNRESET;
1964 case RDMA_CM_EVENT_DISCONNECTED:
1965 case RDMA_CM_EVENT_ADDR_CHANGE:
1966 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1967 dev_dbg(queue->ctrl->ctrl.device,
1968 "disconnect received - connection closed\n");
1969 nvme_rdma_error_recovery(queue->ctrl);
1971 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1972 /* device removal is handled via the ib_client API */
1975 dev_err(queue->ctrl->ctrl.device,
1976 "Unexpected RDMA CM event (%d)\n", ev->event);
1977 nvme_rdma_error_recovery(queue->ctrl);
1982 queue->cm_error = cm_error;
1983 complete(&queue->cm_done);
1989 static void nvme_rdma_complete_timed_out(struct request *rq)
1991 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1992 struct nvme_rdma_queue *queue = req->queue;
1994 nvme_rdma_stop_queue(queue);
1995 if (blk_mq_request_started(rq) && !blk_mq_request_completed(rq)) {
1996 nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
1997 blk_mq_complete_request(rq);
2001 static enum blk_eh_timer_return
2002 nvme_rdma_timeout(struct request *rq, bool reserved)
2004 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2005 struct nvme_rdma_queue *queue = req->queue;
2006 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
2008 dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
2009 rq->tag, nvme_rdma_queue_idx(queue));
2011 if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
2013 * If we are resetting, connecting or deleting we should
2014 * complete immediately because we may block controller
2015 * teardown or setup sequence
2016 * - ctrl disable/shutdown fabrics requests
2017 * - connect requests
2018 * - initialization admin requests
2019 * - I/O requests that entered after unquiescing and
2020 * the controller stopped responding
2022 * All other requests should be cancelled by the error
2023 * recovery work, so it's fine that we fail it here.
2025 nvme_rdma_complete_timed_out(rq);
2030 * LIVE state should trigger the normal error recovery which will
2031 * handle completing this request.
2033 nvme_rdma_error_recovery(ctrl);
2034 return BLK_EH_RESET_TIMER;
2037 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
2038 const struct blk_mq_queue_data *bd)
2040 struct nvme_ns *ns = hctx->queue->queuedata;
2041 struct nvme_rdma_queue *queue = hctx->driver_data;
2042 struct request *rq = bd->rq;
2043 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2044 struct nvme_rdma_qe *sqe = &req->sqe;
2045 struct nvme_command *c = nvme_req(rq)->cmd;
2046 struct ib_device *dev;
2047 bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
2051 WARN_ON_ONCE(rq->tag < 0);
2053 if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2054 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2056 dev = queue->device->dev;
2058 req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
2059 sizeof(struct nvme_command),
2061 err = ib_dma_mapping_error(dev, req->sqe.dma);
2063 return BLK_STS_RESOURCE;
2065 ib_dma_sync_single_for_cpu(dev, sqe->dma,
2066 sizeof(struct nvme_command), DMA_TO_DEVICE);
2068 ret = nvme_setup_cmd(ns, rq);
2072 blk_mq_start_request(rq);
2074 if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
2075 queue->pi_support &&
2076 (c->common.opcode == nvme_cmd_write ||
2077 c->common.opcode == nvme_cmd_read) &&
2079 req->use_sig_mr = true;
2081 req->use_sig_mr = false;
2083 err = nvme_rdma_map_data(queue, rq, c);
2084 if (unlikely(err < 0)) {
2085 dev_err(queue->ctrl->ctrl.device,
2086 "Failed to map data (%d)\n", err);
2090 sqe->cqe.done = nvme_rdma_send_done;
2092 ib_dma_sync_single_for_device(dev, sqe->dma,
2093 sizeof(struct nvme_command), DMA_TO_DEVICE);
2095 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
2096 req->mr ? &req->reg_wr.wr : NULL);
2103 nvme_rdma_unmap_data(queue, rq);
2106 ret = nvme_host_path_error(rq);
2107 else if (err == -ENOMEM || err == -EAGAIN)
2108 ret = BLK_STS_RESOURCE;
2110 ret = BLK_STS_IOERR;
2111 nvme_cleanup_cmd(rq);
2113 ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
2118 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
2120 struct nvme_rdma_queue *queue = hctx->driver_data;
2122 return ib_process_cq_direct(queue->ib_cq, -1);
2125 static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req)
2127 struct request *rq = blk_mq_rq_from_pdu(req);
2128 struct ib_mr_status mr_status;
2131 ret = ib_check_mr_status(req->mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
2133 pr_err("ib_check_mr_status failed, ret %d\n", ret);
2134 nvme_req(rq)->status = NVME_SC_INVALID_PI;
2138 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
2139 switch (mr_status.sig_err.err_type) {
2140 case IB_SIG_BAD_GUARD:
2141 nvme_req(rq)->status = NVME_SC_GUARD_CHECK;
2143 case IB_SIG_BAD_REFTAG:
2144 nvme_req(rq)->status = NVME_SC_REFTAG_CHECK;
2146 case IB_SIG_BAD_APPTAG:
2147 nvme_req(rq)->status = NVME_SC_APPTAG_CHECK;
2150 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
2151 mr_status.sig_err.err_type, mr_status.sig_err.expected,
2152 mr_status.sig_err.actual);
2156 static void nvme_rdma_complete_rq(struct request *rq)
2158 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2159 struct nvme_rdma_queue *queue = req->queue;
2160 struct ib_device *ibdev = queue->device->dev;
2162 if (req->use_sig_mr)
2163 nvme_rdma_check_pi_status(req);
2165 nvme_rdma_unmap_data(queue, rq);
2166 ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
2168 nvme_complete_rq(rq);
2171 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
2173 struct nvme_rdma_ctrl *ctrl = set->driver_data;
2174 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2176 if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2177 /* separate read/write queues */
2178 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2179 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2180 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2181 set->map[HCTX_TYPE_READ].nr_queues =
2182 ctrl->io_queues[HCTX_TYPE_READ];
2183 set->map[HCTX_TYPE_READ].queue_offset =
2184 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2186 /* shared read/write queues */
2187 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2188 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2189 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2190 set->map[HCTX_TYPE_READ].nr_queues =
2191 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2192 set->map[HCTX_TYPE_READ].queue_offset = 0;
2194 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
2195 ctrl->device->dev, 0);
2196 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
2197 ctrl->device->dev, 0);
2199 if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2200 /* map dedicated poll queues only if we have queues left */
2201 set->map[HCTX_TYPE_POLL].nr_queues =
2202 ctrl->io_queues[HCTX_TYPE_POLL];
2203 set->map[HCTX_TYPE_POLL].queue_offset =
2204 ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2205 ctrl->io_queues[HCTX_TYPE_READ];
2206 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2209 dev_info(ctrl->ctrl.device,
2210 "mapped %d/%d/%d default/read/poll queues.\n",
2211 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2212 ctrl->io_queues[HCTX_TYPE_READ],
2213 ctrl->io_queues[HCTX_TYPE_POLL]);
2218 static const struct blk_mq_ops nvme_rdma_mq_ops = {
2219 .queue_rq = nvme_rdma_queue_rq,
2220 .complete = nvme_rdma_complete_rq,
2221 .init_request = nvme_rdma_init_request,
2222 .exit_request = nvme_rdma_exit_request,
2223 .init_hctx = nvme_rdma_init_hctx,
2224 .timeout = nvme_rdma_timeout,
2225 .map_queues = nvme_rdma_map_queues,
2226 .poll = nvme_rdma_poll,
2229 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
2230 .queue_rq = nvme_rdma_queue_rq,
2231 .complete = nvme_rdma_complete_rq,
2232 .init_request = nvme_rdma_init_request,
2233 .exit_request = nvme_rdma_exit_request,
2234 .init_hctx = nvme_rdma_init_admin_hctx,
2235 .timeout = nvme_rdma_timeout,
2238 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
2240 cancel_work_sync(&ctrl->err_work);
2241 cancel_delayed_work_sync(&ctrl->reconnect_work);
2243 nvme_rdma_teardown_io_queues(ctrl, shutdown);
2244 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
2246 nvme_shutdown_ctrl(&ctrl->ctrl);
2248 nvme_disable_ctrl(&ctrl->ctrl);
2249 nvme_rdma_teardown_admin_queue(ctrl, shutdown);
2252 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
2254 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
2257 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
2259 struct nvme_rdma_ctrl *ctrl =
2260 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
2262 nvme_stop_ctrl(&ctrl->ctrl);
2263 nvme_rdma_shutdown_ctrl(ctrl, false);
2265 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2266 /* state change failure should never happen */
2271 if (nvme_rdma_setup_ctrl(ctrl, false))
2277 ++ctrl->ctrl.nr_reconnects;
2278 nvme_rdma_reconnect_or_remove(ctrl);
2281 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
2283 .module = THIS_MODULE,
2284 .flags = NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED,
2285 .reg_read32 = nvmf_reg_read32,
2286 .reg_read64 = nvmf_reg_read64,
2287 .reg_write32 = nvmf_reg_write32,
2288 .free_ctrl = nvme_rdma_free_ctrl,
2289 .submit_async_event = nvme_rdma_submit_async_event,
2290 .delete_ctrl = nvme_rdma_delete_ctrl,
2291 .get_address = nvmf_get_address,
2295 * Fails a connection request if it matches an existing controller
2296 * (association) with the same tuple:
2297 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
2299 * if local address is not specified in the request, it will match an
2300 * existing controller with all the other parameters the same and no
2301 * local port address specified as well.
2303 * The ports don't need to be compared as they are intrinsically
2304 * already matched by the port pointers supplied.
2307 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
2309 struct nvme_rdma_ctrl *ctrl;
2312 mutex_lock(&nvme_rdma_ctrl_mutex);
2313 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2314 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2318 mutex_unlock(&nvme_rdma_ctrl_mutex);
2323 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
2324 struct nvmf_ctrl_options *opts)
2326 struct nvme_rdma_ctrl *ctrl;
2330 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2332 return ERR_PTR(-ENOMEM);
2333 ctrl->ctrl.opts = opts;
2334 INIT_LIST_HEAD(&ctrl->list);
2336 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2338 kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
2339 if (!opts->trsvcid) {
2343 opts->mask |= NVMF_OPT_TRSVCID;
2346 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2347 opts->traddr, opts->trsvcid, &ctrl->addr);
2349 pr_err("malformed address passed: %s:%s\n",
2350 opts->traddr, opts->trsvcid);
2354 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2355 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2356 opts->host_traddr, NULL, &ctrl->src_addr);
2358 pr_err("malformed src address passed: %s\n",
2364 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2369 INIT_DELAYED_WORK(&ctrl->reconnect_work,
2370 nvme_rdma_reconnect_ctrl_work);
2371 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2372 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2374 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2375 opts->nr_poll_queues + 1;
2376 ctrl->ctrl.sqsize = opts->queue_size - 1;
2377 ctrl->ctrl.kato = opts->kato;
2380 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2385 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2386 0 /* no quirks, we're perfect! */);
2388 goto out_kfree_queues;
2390 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2391 WARN_ON_ONCE(!changed);
2393 ret = nvme_rdma_setup_ctrl(ctrl, true);
2395 goto out_uninit_ctrl;
2397 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2398 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2400 mutex_lock(&nvme_rdma_ctrl_mutex);
2401 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2402 mutex_unlock(&nvme_rdma_ctrl_mutex);
2407 nvme_uninit_ctrl(&ctrl->ctrl);
2408 nvme_put_ctrl(&ctrl->ctrl);
2411 return ERR_PTR(ret);
2413 kfree(ctrl->queues);
2416 return ERR_PTR(ret);
2419 static struct nvmf_transport_ops nvme_rdma_transport = {
2421 .module = THIS_MODULE,
2422 .required_opts = NVMF_OPT_TRADDR,
2423 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2424 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2425 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2427 .create_ctrl = nvme_rdma_create_ctrl,
2430 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2432 struct nvme_rdma_ctrl *ctrl;
2433 struct nvme_rdma_device *ndev;
2436 mutex_lock(&device_list_mutex);
2437 list_for_each_entry(ndev, &device_list, entry) {
2438 if (ndev->dev == ib_device) {
2443 mutex_unlock(&device_list_mutex);
2448 /* Delete all controllers using this device */
2449 mutex_lock(&nvme_rdma_ctrl_mutex);
2450 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2451 if (ctrl->device->dev != ib_device)
2453 nvme_delete_ctrl(&ctrl->ctrl);
2455 mutex_unlock(&nvme_rdma_ctrl_mutex);
2457 flush_workqueue(nvme_delete_wq);
2460 static struct ib_client nvme_rdma_ib_client = {
2461 .name = "nvme_rdma",
2462 .remove = nvme_rdma_remove_one
2465 static int __init nvme_rdma_init_module(void)
2469 ret = ib_register_client(&nvme_rdma_ib_client);
2473 ret = nvmf_register_transport(&nvme_rdma_transport);
2475 goto err_unreg_client;
2480 ib_unregister_client(&nvme_rdma_ib_client);
2484 static void __exit nvme_rdma_cleanup_module(void)
2486 struct nvme_rdma_ctrl *ctrl;
2488 nvmf_unregister_transport(&nvme_rdma_transport);
2489 ib_unregister_client(&nvme_rdma_ib_client);
2491 mutex_lock(&nvme_rdma_ctrl_mutex);
2492 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2493 nvme_delete_ctrl(&ctrl->ctrl);
2494 mutex_unlock(&nvme_rdma_ctrl_mutex);
2495 flush_workqueue(nvme_delete_wq);
2498 module_init(nvme_rdma_init_module);
2499 module_exit(nvme_rdma_cleanup_module);
2501 MODULE_LICENSE("GPL v2");