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/blk-integrity.h>
17 #include <linux/types.h>
18 #include <linux/list.h>
19 #include <linux/mutex.h>
20 #include <linux/scatterlist.h>
21 #include <linux/nvme.h>
22 #include <asm/unaligned.h>
24 #include <rdma/ib_verbs.h>
25 #include <rdma/rdma_cm.h>
26 #include <linux/nvme-rdma.h>
32 #define NVME_RDMA_CM_TIMEOUT_MS 3000 /* 3 second */
34 #define NVME_RDMA_MAX_SEGMENTS 256
36 #define NVME_RDMA_MAX_INLINE_SEGMENTS 4
38 #define NVME_RDMA_DATA_SGL_SIZE \
39 (sizeof(struct scatterlist) * NVME_INLINE_SG_CNT)
40 #define NVME_RDMA_METADATA_SGL_SIZE \
41 (sizeof(struct scatterlist) * NVME_INLINE_METADATA_SG_CNT)
43 struct nvme_rdma_device {
44 struct ib_device *dev;
47 struct list_head entry;
48 unsigned int num_inline_segments;
57 struct nvme_rdma_sgl {
59 struct sg_table sg_table;
62 struct nvme_rdma_queue;
63 struct nvme_rdma_request {
64 struct nvme_request req;
66 struct nvme_rdma_qe sqe;
67 union nvme_result result;
70 struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
72 struct ib_reg_wr reg_wr;
73 struct ib_cqe reg_cqe;
74 struct nvme_rdma_queue *queue;
75 struct nvme_rdma_sgl data_sgl;
76 struct nvme_rdma_sgl *metadata_sgl;
80 enum nvme_rdma_queue_flags {
81 NVME_RDMA_Q_ALLOCATED = 0,
83 NVME_RDMA_Q_TR_READY = 2,
86 struct nvme_rdma_queue {
87 struct nvme_rdma_qe *rsp_ring;
89 size_t cmnd_capsule_len;
90 struct nvme_rdma_ctrl *ctrl;
91 struct nvme_rdma_device *device;
96 struct rdma_cm_id *cm_id;
98 struct completion cm_done;
101 struct mutex queue_lock;
104 struct nvme_rdma_ctrl {
105 /* read only in the hot path */
106 struct nvme_rdma_queue *queues;
108 /* other member variables */
109 struct blk_mq_tag_set tag_set;
110 struct work_struct err_work;
112 struct nvme_rdma_qe async_event_sqe;
114 struct delayed_work reconnect_work;
116 struct list_head list;
118 struct blk_mq_tag_set admin_tag_set;
119 struct nvme_rdma_device *device;
123 struct sockaddr_storage addr;
124 struct sockaddr_storage src_addr;
126 struct nvme_ctrl ctrl;
127 bool use_inline_data;
128 u32 io_queues[HCTX_MAX_TYPES];
131 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
133 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
136 static LIST_HEAD(device_list);
137 static DEFINE_MUTEX(device_list_mutex);
139 static LIST_HEAD(nvme_rdma_ctrl_list);
140 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
143 * Disabling this option makes small I/O goes faster, but is fundamentally
144 * unsafe. With it turned off we will have to register a global rkey that
145 * allows read and write access to all physical memory.
147 static bool register_always = true;
148 module_param(register_always, bool, 0444);
149 MODULE_PARM_DESC(register_always,
150 "Use memory registration even for contiguous memory regions");
152 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
153 struct rdma_cm_event *event);
154 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
155 static void nvme_rdma_complete_rq(struct request *rq);
157 static const struct blk_mq_ops nvme_rdma_mq_ops;
158 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
160 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
162 return queue - queue->ctrl->queues;
165 static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue)
167 return nvme_rdma_queue_idx(queue) >
168 queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] +
169 queue->ctrl->io_queues[HCTX_TYPE_READ];
172 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
174 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
177 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
178 size_t capsule_size, enum dma_data_direction dir)
180 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
184 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
185 size_t capsule_size, enum dma_data_direction dir)
187 qe->data = kzalloc(capsule_size, GFP_KERNEL);
191 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
192 if (ib_dma_mapping_error(ibdev, qe->dma)) {
201 static void nvme_rdma_free_ring(struct ib_device *ibdev,
202 struct nvme_rdma_qe *ring, size_t ib_queue_size,
203 size_t capsule_size, enum dma_data_direction dir)
207 for (i = 0; i < ib_queue_size; i++)
208 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
212 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
213 size_t ib_queue_size, size_t capsule_size,
214 enum dma_data_direction dir)
216 struct nvme_rdma_qe *ring;
219 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
224 * Bind the CQEs (post recv buffers) DMA mapping to the RDMA queue
225 * lifetime. It's safe, since any chage in the underlying RDMA device
226 * will issue error recovery and queue re-creation.
228 for (i = 0; i < ib_queue_size; i++) {
229 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
236 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
240 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
242 pr_debug("QP event %s (%d)\n",
243 ib_event_msg(event->event), event->event);
247 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
251 ret = wait_for_completion_interruptible(&queue->cm_done);
254 WARN_ON_ONCE(queue->cm_error > 0);
255 return queue->cm_error;
258 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
260 struct nvme_rdma_device *dev = queue->device;
261 struct ib_qp_init_attr init_attr;
264 memset(&init_attr, 0, sizeof(init_attr));
265 init_attr.event_handler = nvme_rdma_qp_event;
267 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
269 init_attr.cap.max_recv_wr = queue->queue_size + 1;
270 init_attr.cap.max_recv_sge = 1;
271 init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
272 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
273 init_attr.qp_type = IB_QPT_RC;
274 init_attr.send_cq = queue->ib_cq;
275 init_attr.recv_cq = queue->ib_cq;
276 if (queue->pi_support)
277 init_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
278 init_attr.qp_context = queue;
280 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
282 queue->qp = queue->cm_id->qp;
286 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
287 struct request *rq, unsigned int hctx_idx)
289 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
291 kfree(req->sqe.data);
294 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
295 struct request *rq, unsigned int hctx_idx,
296 unsigned int numa_node)
298 struct nvme_rdma_ctrl *ctrl = set->driver_data;
299 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
300 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
301 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
303 nvme_req(rq)->ctrl = &ctrl->ctrl;
304 req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL);
308 /* metadata nvme_rdma_sgl struct is located after command's data SGL */
309 if (queue->pi_support)
310 req->metadata_sgl = (void *)nvme_req(rq) +
311 sizeof(struct nvme_rdma_request) +
312 NVME_RDMA_DATA_SGL_SIZE;
315 nvme_req(rq)->cmd = req->sqe.data;
320 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
321 unsigned int hctx_idx)
323 struct nvme_rdma_ctrl *ctrl = data;
324 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
326 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
328 hctx->driver_data = queue;
332 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
333 unsigned int hctx_idx)
335 struct nvme_rdma_ctrl *ctrl = data;
336 struct nvme_rdma_queue *queue = &ctrl->queues[0];
338 BUG_ON(hctx_idx != 0);
340 hctx->driver_data = queue;
344 static void nvme_rdma_free_dev(struct kref *ref)
346 struct nvme_rdma_device *ndev =
347 container_of(ref, struct nvme_rdma_device, ref);
349 mutex_lock(&device_list_mutex);
350 list_del(&ndev->entry);
351 mutex_unlock(&device_list_mutex);
353 ib_dealloc_pd(ndev->pd);
357 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
359 kref_put(&dev->ref, nvme_rdma_free_dev);
362 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
364 return kref_get_unless_zero(&dev->ref);
367 static struct nvme_rdma_device *
368 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
370 struct nvme_rdma_device *ndev;
372 mutex_lock(&device_list_mutex);
373 list_for_each_entry(ndev, &device_list, entry) {
374 if (ndev->dev->node_guid == cm_id->device->node_guid &&
375 nvme_rdma_dev_get(ndev))
379 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
383 ndev->dev = cm_id->device;
384 kref_init(&ndev->ref);
386 ndev->pd = ib_alloc_pd(ndev->dev,
387 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
388 if (IS_ERR(ndev->pd))
391 if (!(ndev->dev->attrs.device_cap_flags &
392 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
393 dev_err(&ndev->dev->dev,
394 "Memory registrations not supported.\n");
398 ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
399 ndev->dev->attrs.max_send_sge - 1);
400 list_add(&ndev->entry, &device_list);
402 mutex_unlock(&device_list_mutex);
406 ib_dealloc_pd(ndev->pd);
410 mutex_unlock(&device_list_mutex);
414 static void nvme_rdma_free_cq(struct nvme_rdma_queue *queue)
416 if (nvme_rdma_poll_queue(queue))
417 ib_free_cq(queue->ib_cq);
419 ib_cq_pool_put(queue->ib_cq, queue->cq_size);
422 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
424 struct nvme_rdma_device *dev;
425 struct ib_device *ibdev;
427 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
433 if (queue->pi_support)
434 ib_mr_pool_destroy(queue->qp, &queue->qp->sig_mrs);
435 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
438 * The cm_id object might have been destroyed during RDMA connection
439 * establishment error flow to avoid getting other cma events, thus
440 * the destruction of the QP shouldn't use rdma_cm API.
442 ib_destroy_qp(queue->qp);
443 nvme_rdma_free_cq(queue);
445 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
446 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
448 nvme_rdma_dev_put(dev);
451 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev, bool pi_support)
453 u32 max_page_list_len;
456 max_page_list_len = ibdev->attrs.max_pi_fast_reg_page_list_len;
458 max_page_list_len = ibdev->attrs.max_fast_reg_page_list_len;
460 return min_t(u32, NVME_RDMA_MAX_SEGMENTS, max_page_list_len - 1);
463 static int nvme_rdma_create_cq(struct ib_device *ibdev,
464 struct nvme_rdma_queue *queue)
466 int ret, comp_vector, idx = nvme_rdma_queue_idx(queue);
467 enum ib_poll_context poll_ctx;
470 * Spread I/O queues completion vectors according their queue index.
471 * Admin queues can always go on completion vector 0.
473 comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors;
475 /* Polling queues need direct cq polling context */
476 if (nvme_rdma_poll_queue(queue)) {
477 poll_ctx = IB_POLL_DIRECT;
478 queue->ib_cq = ib_alloc_cq(ibdev, queue, queue->cq_size,
479 comp_vector, poll_ctx);
481 poll_ctx = IB_POLL_SOFTIRQ;
482 queue->ib_cq = ib_cq_pool_get(ibdev, queue->cq_size,
483 comp_vector, poll_ctx);
486 if (IS_ERR(queue->ib_cq)) {
487 ret = PTR_ERR(queue->ib_cq);
494 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
496 struct ib_device *ibdev;
497 const int send_wr_factor = 3; /* MR, SEND, INV */
498 const int cq_factor = send_wr_factor + 1; /* + RECV */
499 int ret, pages_per_mr;
501 queue->device = nvme_rdma_find_get_device(queue->cm_id);
502 if (!queue->device) {
503 dev_err(queue->cm_id->device->dev.parent,
504 "no client data found!\n");
505 return -ECONNREFUSED;
507 ibdev = queue->device->dev;
509 /* +1 for ib_stop_cq */
510 queue->cq_size = cq_factor * queue->queue_size + 1;
512 ret = nvme_rdma_create_cq(ibdev, queue);
516 ret = nvme_rdma_create_qp(queue, send_wr_factor);
518 goto out_destroy_ib_cq;
520 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
521 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
522 if (!queue->rsp_ring) {
528 * Currently we don't use SG_GAPS MR's so if the first entry is
529 * misaligned we'll end up using two entries for a single data page,
530 * so one additional entry is required.
532 pages_per_mr = nvme_rdma_get_max_fr_pages(ibdev, queue->pi_support) + 1;
533 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
538 dev_err(queue->ctrl->ctrl.device,
539 "failed to initialize MR pool sized %d for QID %d\n",
540 queue->queue_size, nvme_rdma_queue_idx(queue));
541 goto out_destroy_ring;
544 if (queue->pi_support) {
545 ret = ib_mr_pool_init(queue->qp, &queue->qp->sig_mrs,
546 queue->queue_size, IB_MR_TYPE_INTEGRITY,
547 pages_per_mr, pages_per_mr);
549 dev_err(queue->ctrl->ctrl.device,
550 "failed to initialize PI MR pool sized %d for QID %d\n",
551 queue->queue_size, nvme_rdma_queue_idx(queue));
552 goto out_destroy_mr_pool;
556 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
561 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
563 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
564 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
566 rdma_destroy_qp(queue->cm_id);
568 nvme_rdma_free_cq(queue);
570 nvme_rdma_dev_put(queue->device);
574 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
575 int idx, size_t queue_size)
577 struct nvme_rdma_queue *queue;
578 struct sockaddr *src_addr = NULL;
581 queue = &ctrl->queues[idx];
582 mutex_init(&queue->queue_lock);
584 if (idx && ctrl->ctrl.max_integrity_segments)
585 queue->pi_support = true;
587 queue->pi_support = false;
588 init_completion(&queue->cm_done);
591 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
593 queue->cmnd_capsule_len = sizeof(struct nvme_command);
595 queue->queue_size = queue_size;
597 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
598 RDMA_PS_TCP, IB_QPT_RC);
599 if (IS_ERR(queue->cm_id)) {
600 dev_info(ctrl->ctrl.device,
601 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
602 ret = PTR_ERR(queue->cm_id);
603 goto out_destroy_mutex;
606 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
607 src_addr = (struct sockaddr *)&ctrl->src_addr;
609 queue->cm_error = -ETIMEDOUT;
610 ret = rdma_resolve_addr(queue->cm_id, src_addr,
611 (struct sockaddr *)&ctrl->addr,
612 NVME_RDMA_CM_TIMEOUT_MS);
614 dev_info(ctrl->ctrl.device,
615 "rdma_resolve_addr failed (%d).\n", ret);
616 goto out_destroy_cm_id;
619 ret = nvme_rdma_wait_for_cm(queue);
621 dev_info(ctrl->ctrl.device,
622 "rdma connection establishment failed (%d)\n", ret);
623 goto out_destroy_cm_id;
626 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
631 rdma_destroy_id(queue->cm_id);
632 nvme_rdma_destroy_queue_ib(queue);
634 mutex_destroy(&queue->queue_lock);
638 static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
640 rdma_disconnect(queue->cm_id);
641 ib_drain_qp(queue->qp);
644 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
646 mutex_lock(&queue->queue_lock);
647 if (test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
648 __nvme_rdma_stop_queue(queue);
649 mutex_unlock(&queue->queue_lock);
652 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
654 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
657 rdma_destroy_id(queue->cm_id);
658 nvme_rdma_destroy_queue_ib(queue);
659 mutex_destroy(&queue->queue_lock);
662 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
666 for (i = 1; i < ctrl->ctrl.queue_count; i++)
667 nvme_rdma_free_queue(&ctrl->queues[i]);
670 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
674 for (i = 1; i < ctrl->ctrl.queue_count; i++)
675 nvme_rdma_stop_queue(&ctrl->queues[i]);
678 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
680 struct nvme_rdma_queue *queue = &ctrl->queues[idx];
684 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
686 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
689 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
691 if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
692 __nvme_rdma_stop_queue(queue);
693 dev_info(ctrl->ctrl.device,
694 "failed to connect queue: %d ret=%d\n", idx, ret);
699 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
703 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
704 ret = nvme_rdma_start_queue(ctrl, i);
706 goto out_stop_queues;
712 for (i--; i >= 1; i--)
713 nvme_rdma_stop_queue(&ctrl->queues[i]);
717 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
719 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
720 struct ib_device *ibdev = ctrl->device->dev;
721 unsigned int nr_io_queues, nr_default_queues;
722 unsigned int nr_read_queues, nr_poll_queues;
725 nr_read_queues = min_t(unsigned int, ibdev->num_comp_vectors,
726 min(opts->nr_io_queues, num_online_cpus()));
727 nr_default_queues = min_t(unsigned int, ibdev->num_comp_vectors,
728 min(opts->nr_write_queues, num_online_cpus()));
729 nr_poll_queues = min(opts->nr_poll_queues, num_online_cpus());
730 nr_io_queues = nr_read_queues + nr_default_queues + nr_poll_queues;
732 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
736 if (nr_io_queues == 0) {
737 dev_err(ctrl->ctrl.device,
738 "unable to set any I/O queues\n");
742 ctrl->ctrl.queue_count = nr_io_queues + 1;
743 dev_info(ctrl->ctrl.device,
744 "creating %d I/O queues.\n", nr_io_queues);
746 if (opts->nr_write_queues && nr_read_queues < nr_io_queues) {
748 * separate read/write queues
749 * hand out dedicated default queues only after we have
750 * sufficient read queues.
752 ctrl->io_queues[HCTX_TYPE_READ] = nr_read_queues;
753 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
754 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
755 min(nr_default_queues, nr_io_queues);
756 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
759 * shared read/write queues
760 * either no write queues were requested, or we don't have
761 * sufficient queue count to have dedicated default queues.
763 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
764 min(nr_read_queues, nr_io_queues);
765 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
768 if (opts->nr_poll_queues && nr_io_queues) {
769 /* map dedicated poll queues only if we have queues left */
770 ctrl->io_queues[HCTX_TYPE_POLL] =
771 min(nr_poll_queues, nr_io_queues);
774 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
775 ret = nvme_rdma_alloc_queue(ctrl, i,
776 ctrl->ctrl.sqsize + 1);
778 goto out_free_queues;
784 for (i--; i >= 1; i--)
785 nvme_rdma_free_queue(&ctrl->queues[i]);
790 static int nvme_rdma_alloc_admin_tag_set(struct nvme_ctrl *nctrl)
792 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
793 struct blk_mq_tag_set *set = &ctrl->admin_tag_set;
796 memset(set, 0, sizeof(*set));
797 set->ops = &nvme_rdma_admin_mq_ops;
798 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
799 set->reserved_tags = NVMF_RESERVED_TAGS;
800 set->numa_node = nctrl->numa_node;
801 set->cmd_size = sizeof(struct nvme_rdma_request) +
802 NVME_RDMA_DATA_SGL_SIZE;
803 set->driver_data = ctrl;
804 set->nr_hw_queues = 1;
805 set->timeout = NVME_ADMIN_TIMEOUT;
806 set->flags = BLK_MQ_F_NO_SCHED;
807 ret = blk_mq_alloc_tag_set(set);
809 ctrl->ctrl.admin_tagset = set;
813 static int nvme_rdma_alloc_tag_set(struct nvme_ctrl *nctrl)
815 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
816 struct blk_mq_tag_set *set = &ctrl->tag_set;
819 memset(set, 0, sizeof(*set));
820 set->ops = &nvme_rdma_mq_ops;
821 set->queue_depth = nctrl->sqsize + 1;
822 set->reserved_tags = NVMF_RESERVED_TAGS;
823 set->numa_node = nctrl->numa_node;
824 set->flags = BLK_MQ_F_SHOULD_MERGE;
825 set->cmd_size = sizeof(struct nvme_rdma_request) +
826 NVME_RDMA_DATA_SGL_SIZE;
827 if (nctrl->max_integrity_segments)
828 set->cmd_size += sizeof(struct nvme_rdma_sgl) +
829 NVME_RDMA_METADATA_SGL_SIZE;
830 set->driver_data = ctrl;
831 set->nr_hw_queues = nctrl->queue_count - 1;
832 set->timeout = NVME_IO_TIMEOUT;
833 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
834 ret = blk_mq_alloc_tag_set(set);
836 ctrl->ctrl.tagset = set;
840 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
844 blk_mq_destroy_queue(ctrl->ctrl.admin_q);
845 blk_mq_destroy_queue(ctrl->ctrl.fabrics_q);
846 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
848 if (ctrl->async_event_sqe.data) {
849 cancel_work_sync(&ctrl->ctrl.async_event_work);
850 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
851 sizeof(struct nvme_command), DMA_TO_DEVICE);
852 ctrl->async_event_sqe.data = NULL;
854 nvme_rdma_free_queue(&ctrl->queues[0]);
857 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
860 bool pi_capable = false;
863 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
867 ctrl->device = ctrl->queues[0].device;
868 ctrl->ctrl.numa_node = ibdev_to_node(ctrl->device->dev);
871 if (ctrl->device->dev->attrs.kernel_cap_flags &
872 IBK_INTEGRITY_HANDOVER)
875 ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev,
879 * Bind the async event SQE DMA mapping to the admin queue lifetime.
880 * It's safe, since any chage in the underlying RDMA device will issue
881 * error recovery and queue re-creation.
883 error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
884 sizeof(struct nvme_command), DMA_TO_DEVICE);
889 error = nvme_rdma_alloc_admin_tag_set(&ctrl->ctrl);
891 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 nvme_start_admin_queue(&ctrl->ctrl);
923 error = nvme_init_ctrl_finish(&ctrl->ctrl);
925 goto out_quiesce_queue;
930 nvme_stop_admin_queue(&ctrl->ctrl);
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_mq_destroy_queue(ctrl->ctrl.admin_q);
938 out_cleanup_fabrics_q:
940 blk_mq_destroy_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_mq_destroy_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 ret = nvme_rdma_alloc_tag_set(&ctrl->ctrl);
976 goto out_free_io_queues;
978 ret = nvme_ctrl_init_connect_q(&(ctrl->ctrl));
980 goto out_free_tag_set;
983 ret = nvme_rdma_start_io_queues(ctrl);
985 goto out_cleanup_connect_q;
988 nvme_start_queues(&ctrl->ctrl);
989 if (!nvme_wait_freeze_timeout(&ctrl->ctrl, NVME_IO_TIMEOUT)) {
991 * If we timed out waiting for freeze we are likely to
992 * be stuck. Fail the controller initialization just
996 goto out_wait_freeze_timed_out;
998 blk_mq_update_nr_hw_queues(ctrl->ctrl.tagset,
999 ctrl->ctrl.queue_count - 1);
1000 nvme_unfreeze(&ctrl->ctrl);
1005 out_wait_freeze_timed_out:
1006 nvme_stop_queues(&ctrl->ctrl);
1007 nvme_sync_io_queues(&ctrl->ctrl);
1008 nvme_rdma_stop_io_queues(ctrl);
1009 out_cleanup_connect_q:
1010 nvme_cancel_tagset(&ctrl->ctrl);
1012 blk_mq_destroy_queue(ctrl->ctrl.connect_q);
1015 blk_mq_free_tag_set(ctrl->ctrl.tagset);
1017 nvme_rdma_free_io_queues(ctrl);
1021 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
1024 nvme_stop_admin_queue(&ctrl->ctrl);
1025 blk_sync_queue(ctrl->ctrl.admin_q);
1026 nvme_rdma_stop_queue(&ctrl->queues[0]);
1027 nvme_cancel_admin_tagset(&ctrl->ctrl);
1029 nvme_start_admin_queue(&ctrl->ctrl);
1030 nvme_rdma_destroy_admin_queue(ctrl, remove);
1033 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
1036 if (ctrl->ctrl.queue_count > 1) {
1037 nvme_start_freeze(&ctrl->ctrl);
1038 nvme_stop_queues(&ctrl->ctrl);
1039 nvme_sync_io_queues(&ctrl->ctrl);
1040 nvme_rdma_stop_io_queues(ctrl);
1041 nvme_cancel_tagset(&ctrl->ctrl);
1043 nvme_start_queues(&ctrl->ctrl);
1044 nvme_rdma_destroy_io_queues(ctrl, remove);
1048 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
1050 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1052 cancel_work_sync(&ctrl->err_work);
1053 cancel_delayed_work_sync(&ctrl->reconnect_work);
1056 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
1058 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1060 if (list_empty(&ctrl->list))
1063 mutex_lock(&nvme_rdma_ctrl_mutex);
1064 list_del(&ctrl->list);
1065 mutex_unlock(&nvme_rdma_ctrl_mutex);
1067 nvmf_free_options(nctrl->opts);
1069 kfree(ctrl->queues);
1073 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
1075 /* If we are resetting/deleting then do nothing */
1076 if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
1077 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
1078 ctrl->ctrl.state == NVME_CTRL_LIVE);
1082 if (nvmf_should_reconnect(&ctrl->ctrl)) {
1083 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
1084 ctrl->ctrl.opts->reconnect_delay);
1085 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
1086 ctrl->ctrl.opts->reconnect_delay * HZ);
1088 nvme_delete_ctrl(&ctrl->ctrl);
1092 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1097 ret = nvme_rdma_configure_admin_queue(ctrl, new);
1101 if (ctrl->ctrl.icdoff) {
1103 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1107 if (!(ctrl->ctrl.sgls & (1 << 2))) {
1109 dev_err(ctrl->ctrl.device,
1110 "Mandatory keyed sgls are not supported!\n");
1114 if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1115 dev_warn(ctrl->ctrl.device,
1116 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1117 ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1120 if (ctrl->ctrl.sqsize + 1 > NVME_RDMA_MAX_QUEUE_SIZE) {
1121 dev_warn(ctrl->ctrl.device,
1122 "ctrl sqsize %u > max queue size %u, clamping down\n",
1123 ctrl->ctrl.sqsize + 1, NVME_RDMA_MAX_QUEUE_SIZE);
1124 ctrl->ctrl.sqsize = NVME_RDMA_MAX_QUEUE_SIZE - 1;
1127 if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1128 dev_warn(ctrl->ctrl.device,
1129 "sqsize %u > ctrl maxcmd %u, clamping down\n",
1130 ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1131 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1134 if (ctrl->ctrl.sgls & (1 << 20))
1135 ctrl->use_inline_data = true;
1137 if (ctrl->ctrl.queue_count > 1) {
1138 ret = nvme_rdma_configure_io_queues(ctrl, new);
1143 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1146 * state change failure is ok if we started ctrl delete,
1147 * unless we're during creation of a new controller to
1148 * avoid races with teardown flow.
1150 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1151 ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1157 nvme_start_ctrl(&ctrl->ctrl);
1161 if (ctrl->ctrl.queue_count > 1) {
1162 nvme_stop_queues(&ctrl->ctrl);
1163 nvme_sync_io_queues(&ctrl->ctrl);
1164 nvme_rdma_stop_io_queues(ctrl);
1165 nvme_cancel_tagset(&ctrl->ctrl);
1166 nvme_rdma_destroy_io_queues(ctrl, new);
1169 nvme_stop_admin_queue(&ctrl->ctrl);
1170 blk_sync_queue(ctrl->ctrl.admin_q);
1171 nvme_rdma_stop_queue(&ctrl->queues[0]);
1172 nvme_cancel_admin_tagset(&ctrl->ctrl);
1173 nvme_rdma_destroy_admin_queue(ctrl, new);
1177 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1179 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1180 struct nvme_rdma_ctrl, reconnect_work);
1182 ++ctrl->ctrl.nr_reconnects;
1184 if (nvme_rdma_setup_ctrl(ctrl, false))
1187 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1188 ctrl->ctrl.nr_reconnects);
1190 ctrl->ctrl.nr_reconnects = 0;
1195 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1196 ctrl->ctrl.nr_reconnects);
1197 nvme_rdma_reconnect_or_remove(ctrl);
1200 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1202 struct nvme_rdma_ctrl *ctrl = container_of(work,
1203 struct nvme_rdma_ctrl, err_work);
1205 nvme_auth_stop(&ctrl->ctrl);
1206 nvme_stop_keep_alive(&ctrl->ctrl);
1207 flush_work(&ctrl->ctrl.async_event_work);
1208 nvme_rdma_teardown_io_queues(ctrl, false);
1209 nvme_start_queues(&ctrl->ctrl);
1210 nvme_rdma_teardown_admin_queue(ctrl, false);
1211 nvme_start_admin_queue(&ctrl->ctrl);
1213 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1214 /* state change failure is ok if we started ctrl delete */
1215 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1216 ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1220 nvme_rdma_reconnect_or_remove(ctrl);
1223 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1225 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1228 dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1229 queue_work(nvme_reset_wq, &ctrl->err_work);
1232 static void nvme_rdma_end_request(struct nvme_rdma_request *req)
1234 struct request *rq = blk_mq_rq_from_pdu(req);
1236 if (!refcount_dec_and_test(&req->ref))
1238 if (!nvme_try_complete_req(rq, req->status, req->result))
1239 nvme_rdma_complete_rq(rq);
1242 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1245 struct nvme_rdma_queue *queue = wc->qp->qp_context;
1246 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1248 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1249 dev_info(ctrl->ctrl.device,
1250 "%s for CQE 0x%p failed with status %s (%d)\n",
1252 ib_wc_status_msg(wc->status), wc->status);
1253 nvme_rdma_error_recovery(ctrl);
1256 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1258 if (unlikely(wc->status != IB_WC_SUCCESS))
1259 nvme_rdma_wr_error(cq, wc, "MEMREG");
1262 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1264 struct nvme_rdma_request *req =
1265 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1267 if (unlikely(wc->status != IB_WC_SUCCESS))
1268 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1270 nvme_rdma_end_request(req);
1273 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1274 struct nvme_rdma_request *req)
1276 struct ib_send_wr wr = {
1277 .opcode = IB_WR_LOCAL_INV,
1280 .send_flags = IB_SEND_SIGNALED,
1281 .ex.invalidate_rkey = req->mr->rkey,
1284 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1285 wr.wr_cqe = &req->reg_cqe;
1287 return ib_post_send(queue->qp, &wr, NULL);
1290 static void nvme_rdma_dma_unmap_req(struct ib_device *ibdev, struct request *rq)
1292 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1294 if (blk_integrity_rq(rq)) {
1295 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1296 req->metadata_sgl->nents, rq_dma_dir(rq));
1297 sg_free_table_chained(&req->metadata_sgl->sg_table,
1298 NVME_INLINE_METADATA_SG_CNT);
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 void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1309 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1310 struct nvme_rdma_device *dev = queue->device;
1311 struct ib_device *ibdev = dev->dev;
1312 struct list_head *pool = &queue->qp->rdma_mrs;
1314 if (!blk_rq_nr_phys_segments(rq))
1317 if (req->use_sig_mr)
1318 pool = &queue->qp->sig_mrs;
1321 ib_mr_pool_put(queue->qp, pool, req->mr);
1325 nvme_rdma_dma_unmap_req(ibdev, rq);
1328 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1330 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1333 put_unaligned_le24(0, sg->length);
1334 put_unaligned_le32(0, sg->key);
1335 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1339 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1340 struct nvme_rdma_request *req, struct nvme_command *c,
1343 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1344 struct ib_sge *sge = &req->sge[1];
1345 struct scatterlist *sgl;
1349 for_each_sg(req->data_sgl.sg_table.sgl, sgl, count, i) {
1350 sge->addr = sg_dma_address(sgl);
1351 sge->length = sg_dma_len(sgl);
1352 sge->lkey = queue->device->pd->local_dma_lkey;
1357 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1358 sg->length = cpu_to_le32(len);
1359 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1361 req->num_sge += count;
1365 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1366 struct nvme_rdma_request *req, struct nvme_command *c)
1368 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1370 sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1371 put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length);
1372 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1373 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1377 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1378 struct nvme_rdma_request *req, struct nvme_command *c,
1381 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1384 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1385 if (WARN_ON_ONCE(!req->mr))
1389 * Align the MR to a 4K page size to match the ctrl page size and
1390 * the block virtual boundary.
1392 nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL,
1394 if (unlikely(nr < count)) {
1395 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1402 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1404 req->reg_cqe.done = nvme_rdma_memreg_done;
1405 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1406 req->reg_wr.wr.opcode = IB_WR_REG_MR;
1407 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1408 req->reg_wr.wr.num_sge = 0;
1409 req->reg_wr.mr = req->mr;
1410 req->reg_wr.key = req->mr->rkey;
1411 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1412 IB_ACCESS_REMOTE_READ |
1413 IB_ACCESS_REMOTE_WRITE;
1415 sg->addr = cpu_to_le64(req->mr->iova);
1416 put_unaligned_le24(req->mr->length, sg->length);
1417 put_unaligned_le32(req->mr->rkey, sg->key);
1418 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1419 NVME_SGL_FMT_INVALIDATE;
1424 static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1425 struct nvme_command *cmd, struct ib_sig_domain *domain,
1426 u16 control, u8 pi_type)
1428 domain->sig_type = IB_SIG_TYPE_T10_DIF;
1429 domain->sig.dif.bg_type = IB_T10DIF_CRC;
1430 domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1431 domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1432 if (control & NVME_RW_PRINFO_PRCHK_REF)
1433 domain->sig.dif.ref_remap = true;
1435 domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
1436 domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
1437 domain->sig.dif.app_escape = true;
1438 if (pi_type == NVME_NS_DPS_PI_TYPE3)
1439 domain->sig.dif.ref_escape = true;
1442 static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1443 struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1446 u16 control = le16_to_cpu(cmd->rw.control);
1448 memset(sig_attrs, 0, sizeof(*sig_attrs));
1449 if (control & NVME_RW_PRINFO_PRACT) {
1450 /* for WRITE_INSERT/READ_STRIP no memory domain */
1451 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1452 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1454 /* Clear the PRACT bit since HCA will generate/verify the PI */
1455 control &= ~NVME_RW_PRINFO_PRACT;
1456 cmd->rw.control = cpu_to_le16(control);
1458 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1459 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1461 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
1466 static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1469 if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1470 *mask |= IB_SIG_CHECK_REFTAG;
1471 if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1472 *mask |= IB_SIG_CHECK_GUARD;
1475 static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1477 if (unlikely(wc->status != IB_WC_SUCCESS))
1478 nvme_rdma_wr_error(cq, wc, "SIG");
1481 static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1482 struct nvme_rdma_request *req, struct nvme_command *c,
1483 int count, int pi_count)
1485 struct nvme_rdma_sgl *sgl = &req->data_sgl;
1486 struct ib_reg_wr *wr = &req->reg_wr;
1487 struct request *rq = blk_mq_rq_from_pdu(req);
1488 struct nvme_ns *ns = rq->q->queuedata;
1489 struct bio *bio = rq->bio;
1490 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1493 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs);
1494 if (WARN_ON_ONCE(!req->mr))
1497 nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL,
1498 req->metadata_sgl->sg_table.sgl, pi_count, NULL,
1503 nvme_rdma_set_sig_attrs(blk_get_integrity(bio->bi_bdev->bd_disk), c,
1504 req->mr->sig_attrs, ns->pi_type);
1505 nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask);
1507 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1509 req->reg_cqe.done = nvme_rdma_sig_done;
1510 memset(wr, 0, sizeof(*wr));
1511 wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1512 wr->wr.wr_cqe = &req->reg_cqe;
1514 wr->wr.send_flags = 0;
1516 wr->key = req->mr->rkey;
1517 wr->access = IB_ACCESS_LOCAL_WRITE |
1518 IB_ACCESS_REMOTE_READ |
1519 IB_ACCESS_REMOTE_WRITE;
1521 sg->addr = cpu_to_le64(req->mr->iova);
1522 put_unaligned_le24(req->mr->length, sg->length);
1523 put_unaligned_le32(req->mr->rkey, sg->key);
1524 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1529 ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr);
1536 static int nvme_rdma_dma_map_req(struct ib_device *ibdev, struct request *rq,
1537 int *count, int *pi_count)
1539 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1542 req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1543 ret = sg_alloc_table_chained(&req->data_sgl.sg_table,
1544 blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl,
1545 NVME_INLINE_SG_CNT);
1549 req->data_sgl.nents = blk_rq_map_sg(rq->q, rq,
1550 req->data_sgl.sg_table.sgl);
1552 *count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl,
1553 req->data_sgl.nents, rq_dma_dir(rq));
1554 if (unlikely(*count <= 0)) {
1556 goto out_free_table;
1559 if (blk_integrity_rq(rq)) {
1560 req->metadata_sgl->sg_table.sgl =
1561 (struct scatterlist *)(req->metadata_sgl + 1);
1562 ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table,
1563 blk_rq_count_integrity_sg(rq->q, rq->bio),
1564 req->metadata_sgl->sg_table.sgl,
1565 NVME_INLINE_METADATA_SG_CNT);
1566 if (unlikely(ret)) {
1571 req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q,
1572 rq->bio, req->metadata_sgl->sg_table.sgl);
1573 *pi_count = ib_dma_map_sg(ibdev,
1574 req->metadata_sgl->sg_table.sgl,
1575 req->metadata_sgl->nents,
1577 if (unlikely(*pi_count <= 0)) {
1579 goto out_free_pi_table;
1586 sg_free_table_chained(&req->metadata_sgl->sg_table,
1587 NVME_INLINE_METADATA_SG_CNT);
1589 ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1592 sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1596 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1597 struct request *rq, struct nvme_command *c)
1599 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1600 struct nvme_rdma_device *dev = queue->device;
1601 struct ib_device *ibdev = dev->dev;
1606 refcount_set(&req->ref, 2); /* send and recv completions */
1608 c->common.flags |= NVME_CMD_SGL_METABUF;
1610 if (!blk_rq_nr_phys_segments(rq))
1611 return nvme_rdma_set_sg_null(c);
1613 ret = nvme_rdma_dma_map_req(ibdev, rq, &count, &pi_count);
1617 if (req->use_sig_mr) {
1618 ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1622 if (count <= dev->num_inline_segments) {
1623 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1624 queue->ctrl->use_inline_data &&
1625 blk_rq_payload_bytes(rq) <=
1626 nvme_rdma_inline_data_size(queue)) {
1627 ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1631 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1632 ret = nvme_rdma_map_sg_single(queue, req, c);
1637 ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1640 goto out_dma_unmap_req;
1645 nvme_rdma_dma_unmap_req(ibdev, rq);
1649 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1651 struct nvme_rdma_qe *qe =
1652 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1653 struct nvme_rdma_request *req =
1654 container_of(qe, struct nvme_rdma_request, sqe);
1656 if (unlikely(wc->status != IB_WC_SUCCESS))
1657 nvme_rdma_wr_error(cq, wc, "SEND");
1659 nvme_rdma_end_request(req);
1662 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1663 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1664 struct ib_send_wr *first)
1666 struct ib_send_wr wr;
1669 sge->addr = qe->dma;
1670 sge->length = sizeof(struct nvme_command);
1671 sge->lkey = queue->device->pd->local_dma_lkey;
1674 wr.wr_cqe = &qe->cqe;
1676 wr.num_sge = num_sge;
1677 wr.opcode = IB_WR_SEND;
1678 wr.send_flags = IB_SEND_SIGNALED;
1685 ret = ib_post_send(queue->qp, first, NULL);
1686 if (unlikely(ret)) {
1687 dev_err(queue->ctrl->ctrl.device,
1688 "%s failed with error code %d\n", __func__, ret);
1693 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1694 struct nvme_rdma_qe *qe)
1696 struct ib_recv_wr wr;
1700 list.addr = qe->dma;
1701 list.length = sizeof(struct nvme_completion);
1702 list.lkey = queue->device->pd->local_dma_lkey;
1704 qe->cqe.done = nvme_rdma_recv_done;
1707 wr.wr_cqe = &qe->cqe;
1711 ret = ib_post_recv(queue->qp, &wr, NULL);
1712 if (unlikely(ret)) {
1713 dev_err(queue->ctrl->ctrl.device,
1714 "%s failed with error code %d\n", __func__, ret);
1719 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1721 u32 queue_idx = nvme_rdma_queue_idx(queue);
1724 return queue->ctrl->admin_tag_set.tags[queue_idx];
1725 return queue->ctrl->tag_set.tags[queue_idx - 1];
1728 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1730 if (unlikely(wc->status != IB_WC_SUCCESS))
1731 nvme_rdma_wr_error(cq, wc, "ASYNC");
1734 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1736 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1737 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1738 struct ib_device *dev = queue->device->dev;
1739 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1740 struct nvme_command *cmd = sqe->data;
1744 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1746 memset(cmd, 0, sizeof(*cmd));
1747 cmd->common.opcode = nvme_admin_async_event;
1748 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1749 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1750 nvme_rdma_set_sg_null(cmd);
1752 sqe->cqe.done = nvme_rdma_async_done;
1754 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1757 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1761 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1762 struct nvme_completion *cqe, struct ib_wc *wc)
1765 struct nvme_rdma_request *req;
1767 rq = nvme_find_rq(nvme_rdma_tagset(queue), cqe->command_id);
1769 dev_err(queue->ctrl->ctrl.device,
1770 "got bad command_id %#x on QP %#x\n",
1771 cqe->command_id, queue->qp->qp_num);
1772 nvme_rdma_error_recovery(queue->ctrl);
1775 req = blk_mq_rq_to_pdu(rq);
1777 req->status = cqe->status;
1778 req->result = cqe->result;
1780 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1781 if (unlikely(!req->mr ||
1782 wc->ex.invalidate_rkey != req->mr->rkey)) {
1783 dev_err(queue->ctrl->ctrl.device,
1784 "Bogus remote invalidation for rkey %#x\n",
1785 req->mr ? req->mr->rkey : 0);
1786 nvme_rdma_error_recovery(queue->ctrl);
1788 } else if (req->mr) {
1791 ret = nvme_rdma_inv_rkey(queue, req);
1792 if (unlikely(ret < 0)) {
1793 dev_err(queue->ctrl->ctrl.device,
1794 "Queueing INV WR for rkey %#x failed (%d)\n",
1795 req->mr->rkey, ret);
1796 nvme_rdma_error_recovery(queue->ctrl);
1798 /* the local invalidation completion will end the request */
1802 nvme_rdma_end_request(req);
1805 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1807 struct nvme_rdma_qe *qe =
1808 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1809 struct nvme_rdma_queue *queue = wc->qp->qp_context;
1810 struct ib_device *ibdev = queue->device->dev;
1811 struct nvme_completion *cqe = qe->data;
1812 const size_t len = sizeof(struct nvme_completion);
1814 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1815 nvme_rdma_wr_error(cq, wc, "RECV");
1819 /* sanity checking for received data length */
1820 if (unlikely(wc->byte_len < len)) {
1821 dev_err(queue->ctrl->ctrl.device,
1822 "Unexpected nvme completion length(%d)\n", wc->byte_len);
1823 nvme_rdma_error_recovery(queue->ctrl);
1827 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1829 * AEN requests are special as they don't time out and can
1830 * survive any kind of queue freeze and often don't respond to
1831 * aborts. We don't even bother to allocate a struct request
1832 * for them but rather special case them here.
1834 if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1836 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1839 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1840 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1842 nvme_rdma_post_recv(queue, qe);
1845 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1849 for (i = 0; i < queue->queue_size; i++) {
1850 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1858 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1859 struct rdma_cm_event *ev)
1861 struct rdma_cm_id *cm_id = queue->cm_id;
1862 int status = ev->status;
1863 const char *rej_msg;
1864 const struct nvme_rdma_cm_rej *rej_data;
1867 rej_msg = rdma_reject_msg(cm_id, status);
1868 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1870 if (rej_data && rej_data_len >= sizeof(u16)) {
1871 u16 sts = le16_to_cpu(rej_data->sts);
1873 dev_err(queue->ctrl->ctrl.device,
1874 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1875 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1877 dev_err(queue->ctrl->ctrl.device,
1878 "Connect rejected: status %d (%s).\n", status, rej_msg);
1884 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1886 struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1889 ret = nvme_rdma_create_queue_ib(queue);
1893 if (ctrl->opts->tos >= 0)
1894 rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1895 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CM_TIMEOUT_MS);
1897 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1899 goto out_destroy_queue;
1905 nvme_rdma_destroy_queue_ib(queue);
1909 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1911 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1912 struct rdma_conn_param param = { };
1913 struct nvme_rdma_cm_req priv = { };
1916 param.qp_num = queue->qp->qp_num;
1917 param.flow_control = 1;
1919 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1920 /* maximum retry count */
1921 param.retry_count = 7;
1922 param.rnr_retry_count = 7;
1923 param.private_data = &priv;
1924 param.private_data_len = sizeof(priv);
1926 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1927 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1929 * set the admin queue depth to the minimum size
1930 * specified by the Fabrics standard.
1932 if (priv.qid == 0) {
1933 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1934 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1937 * current interpretation of the fabrics spec
1938 * is at minimum you make hrqsize sqsize+1, or a
1939 * 1's based representation of sqsize.
1941 priv.hrqsize = cpu_to_le16(queue->queue_size);
1942 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1945 ret = rdma_connect_locked(queue->cm_id, ¶m);
1947 dev_err(ctrl->ctrl.device,
1948 "rdma_connect_locked failed (%d).\n", ret);
1955 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1956 struct rdma_cm_event *ev)
1958 struct nvme_rdma_queue *queue = cm_id->context;
1961 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1962 rdma_event_msg(ev->event), ev->event,
1965 switch (ev->event) {
1966 case RDMA_CM_EVENT_ADDR_RESOLVED:
1967 cm_error = nvme_rdma_addr_resolved(queue);
1969 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1970 cm_error = nvme_rdma_route_resolved(queue);
1972 case RDMA_CM_EVENT_ESTABLISHED:
1973 queue->cm_error = nvme_rdma_conn_established(queue);
1974 /* complete cm_done regardless of success/failure */
1975 complete(&queue->cm_done);
1977 case RDMA_CM_EVENT_REJECTED:
1978 cm_error = nvme_rdma_conn_rejected(queue, ev);
1980 case RDMA_CM_EVENT_ROUTE_ERROR:
1981 case RDMA_CM_EVENT_CONNECT_ERROR:
1982 case RDMA_CM_EVENT_UNREACHABLE:
1983 case RDMA_CM_EVENT_ADDR_ERROR:
1984 dev_dbg(queue->ctrl->ctrl.device,
1985 "CM error event %d\n", ev->event);
1986 cm_error = -ECONNRESET;
1988 case RDMA_CM_EVENT_DISCONNECTED:
1989 case RDMA_CM_EVENT_ADDR_CHANGE:
1990 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1991 dev_dbg(queue->ctrl->ctrl.device,
1992 "disconnect received - connection closed\n");
1993 nvme_rdma_error_recovery(queue->ctrl);
1995 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1996 /* device removal is handled via the ib_client API */
1999 dev_err(queue->ctrl->ctrl.device,
2000 "Unexpected RDMA CM event (%d)\n", ev->event);
2001 nvme_rdma_error_recovery(queue->ctrl);
2006 queue->cm_error = cm_error;
2007 complete(&queue->cm_done);
2013 static void nvme_rdma_complete_timed_out(struct request *rq)
2015 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2016 struct nvme_rdma_queue *queue = req->queue;
2018 nvme_rdma_stop_queue(queue);
2019 nvmf_complete_timed_out_request(rq);
2022 static enum blk_eh_timer_return nvme_rdma_timeout(struct request *rq)
2024 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2025 struct nvme_rdma_queue *queue = req->queue;
2026 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
2028 dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
2029 rq->tag, nvme_rdma_queue_idx(queue));
2031 if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
2033 * If we are resetting, connecting or deleting we should
2034 * complete immediately because we may block controller
2035 * teardown or setup sequence
2036 * - ctrl disable/shutdown fabrics requests
2037 * - connect requests
2038 * - initialization admin requests
2039 * - I/O requests that entered after unquiescing and
2040 * the controller stopped responding
2042 * All other requests should be cancelled by the error
2043 * recovery work, so it's fine that we fail it here.
2045 nvme_rdma_complete_timed_out(rq);
2050 * LIVE state should trigger the normal error recovery which will
2051 * handle completing this request.
2053 nvme_rdma_error_recovery(ctrl);
2054 return BLK_EH_RESET_TIMER;
2057 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
2058 const struct blk_mq_queue_data *bd)
2060 struct nvme_ns *ns = hctx->queue->queuedata;
2061 struct nvme_rdma_queue *queue = hctx->driver_data;
2062 struct request *rq = bd->rq;
2063 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2064 struct nvme_rdma_qe *sqe = &req->sqe;
2065 struct nvme_command *c = nvme_req(rq)->cmd;
2066 struct ib_device *dev;
2067 bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
2071 WARN_ON_ONCE(rq->tag < 0);
2073 if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2074 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2076 dev = queue->device->dev;
2078 req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
2079 sizeof(struct nvme_command),
2081 err = ib_dma_mapping_error(dev, req->sqe.dma);
2083 return BLK_STS_RESOURCE;
2085 ib_dma_sync_single_for_cpu(dev, sqe->dma,
2086 sizeof(struct nvme_command), DMA_TO_DEVICE);
2088 ret = nvme_setup_cmd(ns, rq);
2092 blk_mq_start_request(rq);
2094 if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
2095 queue->pi_support &&
2096 (c->common.opcode == nvme_cmd_write ||
2097 c->common.opcode == nvme_cmd_read) &&
2099 req->use_sig_mr = true;
2101 req->use_sig_mr = false;
2103 err = nvme_rdma_map_data(queue, rq, c);
2104 if (unlikely(err < 0)) {
2105 dev_err(queue->ctrl->ctrl.device,
2106 "Failed to map data (%d)\n", err);
2110 sqe->cqe.done = nvme_rdma_send_done;
2112 ib_dma_sync_single_for_device(dev, sqe->dma,
2113 sizeof(struct nvme_command), DMA_TO_DEVICE);
2115 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
2116 req->mr ? &req->reg_wr.wr : NULL);
2123 nvme_rdma_unmap_data(queue, rq);
2126 ret = nvme_host_path_error(rq);
2127 else if (err == -ENOMEM || err == -EAGAIN)
2128 ret = BLK_STS_RESOURCE;
2130 ret = BLK_STS_IOERR;
2131 nvme_cleanup_cmd(rq);
2133 ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
2138 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2140 struct nvme_rdma_queue *queue = hctx->driver_data;
2142 return ib_process_cq_direct(queue->ib_cq, -1);
2145 static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req)
2147 struct request *rq = blk_mq_rq_from_pdu(req);
2148 struct ib_mr_status mr_status;
2151 ret = ib_check_mr_status(req->mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
2153 pr_err("ib_check_mr_status failed, ret %d\n", ret);
2154 nvme_req(rq)->status = NVME_SC_INVALID_PI;
2158 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
2159 switch (mr_status.sig_err.err_type) {
2160 case IB_SIG_BAD_GUARD:
2161 nvme_req(rq)->status = NVME_SC_GUARD_CHECK;
2163 case IB_SIG_BAD_REFTAG:
2164 nvme_req(rq)->status = NVME_SC_REFTAG_CHECK;
2166 case IB_SIG_BAD_APPTAG:
2167 nvme_req(rq)->status = NVME_SC_APPTAG_CHECK;
2170 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
2171 mr_status.sig_err.err_type, mr_status.sig_err.expected,
2172 mr_status.sig_err.actual);
2176 static void nvme_rdma_complete_rq(struct request *rq)
2178 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2179 struct nvme_rdma_queue *queue = req->queue;
2180 struct ib_device *ibdev = queue->device->dev;
2182 if (req->use_sig_mr)
2183 nvme_rdma_check_pi_status(req);
2185 nvme_rdma_unmap_data(queue, rq);
2186 ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
2188 nvme_complete_rq(rq);
2191 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
2193 struct nvme_rdma_ctrl *ctrl = set->driver_data;
2194 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2196 if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2197 /* separate read/write queues */
2198 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2199 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2200 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2201 set->map[HCTX_TYPE_READ].nr_queues =
2202 ctrl->io_queues[HCTX_TYPE_READ];
2203 set->map[HCTX_TYPE_READ].queue_offset =
2204 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2206 /* shared read/write queues */
2207 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2208 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2209 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2210 set->map[HCTX_TYPE_READ].nr_queues =
2211 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2212 set->map[HCTX_TYPE_READ].queue_offset = 0;
2214 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
2215 ctrl->device->dev, 0);
2216 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
2217 ctrl->device->dev, 0);
2219 if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2220 /* map dedicated poll queues only if we have queues left */
2221 set->map[HCTX_TYPE_POLL].nr_queues =
2222 ctrl->io_queues[HCTX_TYPE_POLL];
2223 set->map[HCTX_TYPE_POLL].queue_offset =
2224 ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2225 ctrl->io_queues[HCTX_TYPE_READ];
2226 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2229 dev_info(ctrl->ctrl.device,
2230 "mapped %d/%d/%d default/read/poll queues.\n",
2231 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2232 ctrl->io_queues[HCTX_TYPE_READ],
2233 ctrl->io_queues[HCTX_TYPE_POLL]);
2238 static const struct blk_mq_ops nvme_rdma_mq_ops = {
2239 .queue_rq = nvme_rdma_queue_rq,
2240 .complete = nvme_rdma_complete_rq,
2241 .init_request = nvme_rdma_init_request,
2242 .exit_request = nvme_rdma_exit_request,
2243 .init_hctx = nvme_rdma_init_hctx,
2244 .timeout = nvme_rdma_timeout,
2245 .map_queues = nvme_rdma_map_queues,
2246 .poll = nvme_rdma_poll,
2249 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
2250 .queue_rq = nvme_rdma_queue_rq,
2251 .complete = nvme_rdma_complete_rq,
2252 .init_request = nvme_rdma_init_request,
2253 .exit_request = nvme_rdma_exit_request,
2254 .init_hctx = nvme_rdma_init_admin_hctx,
2255 .timeout = nvme_rdma_timeout,
2258 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
2260 nvme_rdma_teardown_io_queues(ctrl, shutdown);
2261 nvme_stop_admin_queue(&ctrl->ctrl);
2263 nvme_shutdown_ctrl(&ctrl->ctrl);
2265 nvme_disable_ctrl(&ctrl->ctrl);
2266 nvme_rdma_teardown_admin_queue(ctrl, shutdown);
2269 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
2271 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
2274 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
2276 struct nvme_rdma_ctrl *ctrl =
2277 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
2279 nvme_stop_ctrl(&ctrl->ctrl);
2280 nvme_rdma_shutdown_ctrl(ctrl, false);
2282 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2283 /* state change failure should never happen */
2288 if (nvme_rdma_setup_ctrl(ctrl, false))
2294 ++ctrl->ctrl.nr_reconnects;
2295 nvme_rdma_reconnect_or_remove(ctrl);
2298 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
2300 .module = THIS_MODULE,
2301 .flags = NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED,
2302 .reg_read32 = nvmf_reg_read32,
2303 .reg_read64 = nvmf_reg_read64,
2304 .reg_write32 = nvmf_reg_write32,
2305 .free_ctrl = nvme_rdma_free_ctrl,
2306 .submit_async_event = nvme_rdma_submit_async_event,
2307 .delete_ctrl = nvme_rdma_delete_ctrl,
2308 .get_address = nvmf_get_address,
2309 .stop_ctrl = nvme_rdma_stop_ctrl,
2313 * Fails a connection request if it matches an existing controller
2314 * (association) with the same tuple:
2315 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
2317 * if local address is not specified in the request, it will match an
2318 * existing controller with all the other parameters the same and no
2319 * local port address specified as well.
2321 * The ports don't need to be compared as they are intrinsically
2322 * already matched by the port pointers supplied.
2325 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
2327 struct nvme_rdma_ctrl *ctrl;
2330 mutex_lock(&nvme_rdma_ctrl_mutex);
2331 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2332 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2336 mutex_unlock(&nvme_rdma_ctrl_mutex);
2341 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
2342 struct nvmf_ctrl_options *opts)
2344 struct nvme_rdma_ctrl *ctrl;
2348 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2350 return ERR_PTR(-ENOMEM);
2351 ctrl->ctrl.opts = opts;
2352 INIT_LIST_HEAD(&ctrl->list);
2354 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2356 kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
2357 if (!opts->trsvcid) {
2361 opts->mask |= NVMF_OPT_TRSVCID;
2364 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2365 opts->traddr, opts->trsvcid, &ctrl->addr);
2367 pr_err("malformed address passed: %s:%s\n",
2368 opts->traddr, opts->trsvcid);
2372 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2373 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2374 opts->host_traddr, NULL, &ctrl->src_addr);
2376 pr_err("malformed src address passed: %s\n",
2382 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2387 INIT_DELAYED_WORK(&ctrl->reconnect_work,
2388 nvme_rdma_reconnect_ctrl_work);
2389 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2390 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2392 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2393 opts->nr_poll_queues + 1;
2394 ctrl->ctrl.sqsize = opts->queue_size - 1;
2395 ctrl->ctrl.kato = opts->kato;
2398 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2403 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2404 0 /* no quirks, we're perfect! */);
2406 goto out_kfree_queues;
2408 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2409 WARN_ON_ONCE(!changed);
2411 ret = nvme_rdma_setup_ctrl(ctrl, true);
2413 goto out_uninit_ctrl;
2415 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2416 nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr);
2418 mutex_lock(&nvme_rdma_ctrl_mutex);
2419 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2420 mutex_unlock(&nvme_rdma_ctrl_mutex);
2425 nvme_uninit_ctrl(&ctrl->ctrl);
2426 nvme_put_ctrl(&ctrl->ctrl);
2429 return ERR_PTR(ret);
2431 kfree(ctrl->queues);
2434 return ERR_PTR(ret);
2437 static struct nvmf_transport_ops nvme_rdma_transport = {
2439 .module = THIS_MODULE,
2440 .required_opts = NVMF_OPT_TRADDR,
2441 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2442 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2443 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2445 .create_ctrl = nvme_rdma_create_ctrl,
2448 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2450 struct nvme_rdma_ctrl *ctrl;
2451 struct nvme_rdma_device *ndev;
2454 mutex_lock(&device_list_mutex);
2455 list_for_each_entry(ndev, &device_list, entry) {
2456 if (ndev->dev == ib_device) {
2461 mutex_unlock(&device_list_mutex);
2466 /* Delete all controllers using this device */
2467 mutex_lock(&nvme_rdma_ctrl_mutex);
2468 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2469 if (ctrl->device->dev != ib_device)
2471 nvme_delete_ctrl(&ctrl->ctrl);
2473 mutex_unlock(&nvme_rdma_ctrl_mutex);
2475 flush_workqueue(nvme_delete_wq);
2478 static struct ib_client nvme_rdma_ib_client = {
2479 .name = "nvme_rdma",
2480 .remove = nvme_rdma_remove_one
2483 static int __init nvme_rdma_init_module(void)
2487 ret = ib_register_client(&nvme_rdma_ib_client);
2491 ret = nvmf_register_transport(&nvme_rdma_transport);
2493 goto err_unreg_client;
2498 ib_unregister_client(&nvme_rdma_ib_client);
2502 static void __exit nvme_rdma_cleanup_module(void)
2504 struct nvme_rdma_ctrl *ctrl;
2506 nvmf_unregister_transport(&nvme_rdma_transport);
2507 ib_unregister_client(&nvme_rdma_ib_client);
2509 mutex_lock(&nvme_rdma_ctrl_mutex);
2510 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2511 nvme_delete_ctrl(&ctrl->ctrl);
2512 mutex_unlock(&nvme_rdma_ctrl_mutex);
2513 flush_workqueue(nvme_delete_wq);
2516 module_init(nvme_rdma_init_module);
2517 module_exit(nvme_rdma_cleanup_module);
2519 MODULE_LICENSE("GPL v2");