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;
321 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
322 unsigned int hctx_idx)
324 struct nvme_rdma_ctrl *ctrl = data;
325 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
327 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
329 hctx->driver_data = queue;
333 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
334 unsigned int hctx_idx)
336 struct nvme_rdma_ctrl *ctrl = data;
337 struct nvme_rdma_queue *queue = &ctrl->queues[0];
339 BUG_ON(hctx_idx != 0);
341 hctx->driver_data = queue;
345 static void nvme_rdma_free_dev(struct kref *ref)
347 struct nvme_rdma_device *ndev =
348 container_of(ref, struct nvme_rdma_device, ref);
350 mutex_lock(&device_list_mutex);
351 list_del(&ndev->entry);
352 mutex_unlock(&device_list_mutex);
354 ib_dealloc_pd(ndev->pd);
358 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
360 kref_put(&dev->ref, nvme_rdma_free_dev);
363 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
365 return kref_get_unless_zero(&dev->ref);
368 static struct nvme_rdma_device *
369 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
371 struct nvme_rdma_device *ndev;
373 mutex_lock(&device_list_mutex);
374 list_for_each_entry(ndev, &device_list, entry) {
375 if (ndev->dev->node_guid == cm_id->device->node_guid &&
376 nvme_rdma_dev_get(ndev))
380 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
384 ndev->dev = cm_id->device;
385 kref_init(&ndev->ref);
387 ndev->pd = ib_alloc_pd(ndev->dev,
388 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
389 if (IS_ERR(ndev->pd))
392 if (!(ndev->dev->attrs.device_cap_flags &
393 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
394 dev_err(&ndev->dev->dev,
395 "Memory registrations not supported.\n");
399 ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
400 ndev->dev->attrs.max_send_sge - 1);
401 list_add(&ndev->entry, &device_list);
403 mutex_unlock(&device_list_mutex);
407 ib_dealloc_pd(ndev->pd);
411 mutex_unlock(&device_list_mutex);
415 static void nvme_rdma_free_cq(struct nvme_rdma_queue *queue)
417 if (nvme_rdma_poll_queue(queue))
418 ib_free_cq(queue->ib_cq);
420 ib_cq_pool_put(queue->ib_cq, queue->cq_size);
423 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
425 struct nvme_rdma_device *dev;
426 struct ib_device *ibdev;
428 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
434 if (queue->pi_support)
435 ib_mr_pool_destroy(queue->qp, &queue->qp->sig_mrs);
436 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
439 * The cm_id object might have been destroyed during RDMA connection
440 * establishment error flow to avoid getting other cma events, thus
441 * the destruction of the QP shouldn't use rdma_cm API.
443 ib_destroy_qp(queue->qp);
444 nvme_rdma_free_cq(queue);
446 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
447 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
449 nvme_rdma_dev_put(dev);
452 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev, bool pi_support)
454 u32 max_page_list_len;
457 max_page_list_len = ibdev->attrs.max_pi_fast_reg_page_list_len;
459 max_page_list_len = ibdev->attrs.max_fast_reg_page_list_len;
461 return min_t(u32, NVME_RDMA_MAX_SEGMENTS, max_page_list_len - 1);
464 static int nvme_rdma_create_cq(struct ib_device *ibdev,
465 struct nvme_rdma_queue *queue)
467 int ret, comp_vector, idx = nvme_rdma_queue_idx(queue);
468 enum ib_poll_context poll_ctx;
471 * Spread I/O queues completion vectors according their queue index.
472 * Admin queues can always go on completion vector 0.
474 comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors;
476 /* Polling queues need direct cq polling context */
477 if (nvme_rdma_poll_queue(queue)) {
478 poll_ctx = IB_POLL_DIRECT;
479 queue->ib_cq = ib_alloc_cq(ibdev, queue, queue->cq_size,
480 comp_vector, poll_ctx);
482 poll_ctx = IB_POLL_SOFTIRQ;
483 queue->ib_cq = ib_cq_pool_get(ibdev, queue->cq_size,
484 comp_vector, poll_ctx);
487 if (IS_ERR(queue->ib_cq)) {
488 ret = PTR_ERR(queue->ib_cq);
495 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
497 struct ib_device *ibdev;
498 const int send_wr_factor = 3; /* MR, SEND, INV */
499 const int cq_factor = send_wr_factor + 1; /* + RECV */
500 int ret, pages_per_mr;
502 queue->device = nvme_rdma_find_get_device(queue->cm_id);
503 if (!queue->device) {
504 dev_err(queue->cm_id->device->dev.parent,
505 "no client data found!\n");
506 return -ECONNREFUSED;
508 ibdev = queue->device->dev;
510 /* +1 for ib_stop_cq */
511 queue->cq_size = cq_factor * queue->queue_size + 1;
513 ret = nvme_rdma_create_cq(ibdev, queue);
517 ret = nvme_rdma_create_qp(queue, send_wr_factor);
519 goto out_destroy_ib_cq;
521 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
522 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
523 if (!queue->rsp_ring) {
529 * Currently we don't use SG_GAPS MR's so if the first entry is
530 * misaligned we'll end up using two entries for a single data page,
531 * so one additional entry is required.
533 pages_per_mr = nvme_rdma_get_max_fr_pages(ibdev, queue->pi_support) + 1;
534 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
539 dev_err(queue->ctrl->ctrl.device,
540 "failed to initialize MR pool sized %d for QID %d\n",
541 queue->queue_size, nvme_rdma_queue_idx(queue));
542 goto out_destroy_ring;
545 if (queue->pi_support) {
546 ret = ib_mr_pool_init(queue->qp, &queue->qp->sig_mrs,
547 queue->queue_size, IB_MR_TYPE_INTEGRITY,
548 pages_per_mr, pages_per_mr);
550 dev_err(queue->ctrl->ctrl.device,
551 "failed to initialize PI MR pool sized %d for QID %d\n",
552 queue->queue_size, nvme_rdma_queue_idx(queue));
553 goto out_destroy_mr_pool;
557 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
562 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
564 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
565 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
567 rdma_destroy_qp(queue->cm_id);
569 nvme_rdma_free_cq(queue);
571 nvme_rdma_dev_put(queue->device);
575 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
576 int idx, size_t queue_size)
578 struct nvme_rdma_queue *queue;
579 struct sockaddr *src_addr = NULL;
582 queue = &ctrl->queues[idx];
583 mutex_init(&queue->queue_lock);
585 if (idx && ctrl->ctrl.max_integrity_segments)
586 queue->pi_support = true;
588 queue->pi_support = false;
589 init_completion(&queue->cm_done);
592 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
594 queue->cmnd_capsule_len = sizeof(struct nvme_command);
596 queue->queue_size = queue_size;
598 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
599 RDMA_PS_TCP, IB_QPT_RC);
600 if (IS_ERR(queue->cm_id)) {
601 dev_info(ctrl->ctrl.device,
602 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
603 ret = PTR_ERR(queue->cm_id);
604 goto out_destroy_mutex;
607 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
608 src_addr = (struct sockaddr *)&ctrl->src_addr;
610 queue->cm_error = -ETIMEDOUT;
611 ret = rdma_resolve_addr(queue->cm_id, src_addr,
612 (struct sockaddr *)&ctrl->addr,
613 NVME_RDMA_CONNECT_TIMEOUT_MS);
615 dev_info(ctrl->ctrl.device,
616 "rdma_resolve_addr failed (%d).\n", ret);
617 goto out_destroy_cm_id;
620 ret = nvme_rdma_wait_for_cm(queue);
622 dev_info(ctrl->ctrl.device,
623 "rdma connection establishment failed (%d)\n", ret);
624 goto out_destroy_cm_id;
627 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
632 rdma_destroy_id(queue->cm_id);
633 nvme_rdma_destroy_queue_ib(queue);
635 mutex_destroy(&queue->queue_lock);
639 static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
641 rdma_disconnect(queue->cm_id);
642 ib_drain_qp(queue->qp);
645 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
647 mutex_lock(&queue->queue_lock);
648 if (test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
649 __nvme_rdma_stop_queue(queue);
650 mutex_unlock(&queue->queue_lock);
653 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
655 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
658 rdma_destroy_id(queue->cm_id);
659 nvme_rdma_destroy_queue_ib(queue);
660 mutex_destroy(&queue->queue_lock);
663 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
667 for (i = 1; i < ctrl->ctrl.queue_count; i++)
668 nvme_rdma_free_queue(&ctrl->queues[i]);
671 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
675 for (i = 1; i < ctrl->ctrl.queue_count; i++)
676 nvme_rdma_stop_queue(&ctrl->queues[i]);
679 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
681 struct nvme_rdma_queue *queue = &ctrl->queues[idx];
682 bool poll = nvme_rdma_poll_queue(queue);
686 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx, poll);
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 if (nr_io_queues == 0) {
739 dev_err(ctrl->ctrl.device,
740 "unable to set any I/O queues\n");
744 ctrl->ctrl.queue_count = nr_io_queues + 1;
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 = 2; /* connect + keep-alive */
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 = 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 = 1; /* fabric connect */
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_identify(&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 if (ctrl->ctrl.admin_tagset) {
1032 blk_mq_tagset_busy_iter(ctrl->ctrl.admin_tagset,
1033 nvme_cancel_request, &ctrl->ctrl);
1034 blk_mq_tagset_wait_completed_request(ctrl->ctrl.admin_tagset);
1037 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1038 nvme_rdma_destroy_admin_queue(ctrl, remove);
1041 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
1044 if (ctrl->ctrl.queue_count > 1) {
1045 nvme_start_freeze(&ctrl->ctrl);
1046 nvme_stop_queues(&ctrl->ctrl);
1047 nvme_sync_io_queues(&ctrl->ctrl);
1048 nvme_rdma_stop_io_queues(ctrl);
1049 if (ctrl->ctrl.tagset) {
1050 blk_mq_tagset_busy_iter(ctrl->ctrl.tagset,
1051 nvme_cancel_request, &ctrl->ctrl);
1052 blk_mq_tagset_wait_completed_request(ctrl->ctrl.tagset);
1055 nvme_start_queues(&ctrl->ctrl);
1056 nvme_rdma_destroy_io_queues(ctrl, remove);
1060 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
1062 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1064 cancel_work_sync(&ctrl->err_work);
1065 cancel_delayed_work_sync(&ctrl->reconnect_work);
1068 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
1070 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1072 if (list_empty(&ctrl->list))
1075 mutex_lock(&nvme_rdma_ctrl_mutex);
1076 list_del(&ctrl->list);
1077 mutex_unlock(&nvme_rdma_ctrl_mutex);
1079 nvmf_free_options(nctrl->opts);
1081 kfree(ctrl->queues);
1085 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
1087 /* If we are resetting/deleting then do nothing */
1088 if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
1089 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
1090 ctrl->ctrl.state == NVME_CTRL_LIVE);
1094 if (nvmf_should_reconnect(&ctrl->ctrl)) {
1095 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
1096 ctrl->ctrl.opts->reconnect_delay);
1097 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
1098 ctrl->ctrl.opts->reconnect_delay * HZ);
1100 nvme_delete_ctrl(&ctrl->ctrl);
1104 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1109 ret = nvme_rdma_configure_admin_queue(ctrl, new);
1113 if (ctrl->ctrl.icdoff) {
1115 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1119 if (!(ctrl->ctrl.sgls & (1 << 2))) {
1121 dev_err(ctrl->ctrl.device,
1122 "Mandatory keyed sgls are not supported!\n");
1126 if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1127 dev_warn(ctrl->ctrl.device,
1128 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1129 ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1132 if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1133 dev_warn(ctrl->ctrl.device,
1134 "sqsize %u > ctrl maxcmd %u, clamping down\n",
1135 ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1136 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1139 if (ctrl->ctrl.sgls & (1 << 20))
1140 ctrl->use_inline_data = true;
1142 if (ctrl->ctrl.queue_count > 1) {
1143 ret = nvme_rdma_configure_io_queues(ctrl, new);
1148 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1151 * state change failure is ok if we started ctrl delete,
1152 * unless we're during creation of a new controller to
1153 * avoid races with teardown flow.
1155 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1156 ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1162 nvme_start_ctrl(&ctrl->ctrl);
1166 if (ctrl->ctrl.queue_count > 1) {
1167 nvme_stop_queues(&ctrl->ctrl);
1168 nvme_sync_io_queues(&ctrl->ctrl);
1169 nvme_rdma_stop_io_queues(ctrl);
1170 nvme_cancel_tagset(&ctrl->ctrl);
1171 nvme_rdma_destroy_io_queues(ctrl, new);
1174 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1175 blk_sync_queue(ctrl->ctrl.admin_q);
1176 nvme_rdma_stop_queue(&ctrl->queues[0]);
1177 nvme_cancel_admin_tagset(&ctrl->ctrl);
1178 nvme_rdma_destroy_admin_queue(ctrl, new);
1182 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1184 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1185 struct nvme_rdma_ctrl, reconnect_work);
1187 ++ctrl->ctrl.nr_reconnects;
1189 if (nvme_rdma_setup_ctrl(ctrl, false))
1192 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1193 ctrl->ctrl.nr_reconnects);
1195 ctrl->ctrl.nr_reconnects = 0;
1200 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1201 ctrl->ctrl.nr_reconnects);
1202 nvme_rdma_reconnect_or_remove(ctrl);
1205 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1207 struct nvme_rdma_ctrl *ctrl = container_of(work,
1208 struct nvme_rdma_ctrl, err_work);
1210 nvme_stop_keep_alive(&ctrl->ctrl);
1211 flush_work(&ctrl->ctrl.async_event_work);
1212 nvme_rdma_teardown_io_queues(ctrl, false);
1213 nvme_start_queues(&ctrl->ctrl);
1214 nvme_rdma_teardown_admin_queue(ctrl, false);
1215 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1217 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1218 /* state change failure is ok if we started ctrl delete */
1219 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1220 ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1224 nvme_rdma_reconnect_or_remove(ctrl);
1227 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1229 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1232 dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1233 queue_work(nvme_reset_wq, &ctrl->err_work);
1236 static void nvme_rdma_end_request(struct nvme_rdma_request *req)
1238 struct request *rq = blk_mq_rq_from_pdu(req);
1240 if (!refcount_dec_and_test(&req->ref))
1242 if (!nvme_try_complete_req(rq, req->status, req->result))
1243 nvme_rdma_complete_rq(rq);
1246 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1249 struct nvme_rdma_queue *queue = wc->qp->qp_context;
1250 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1252 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1253 dev_info(ctrl->ctrl.device,
1254 "%s for CQE 0x%p failed with status %s (%d)\n",
1256 ib_wc_status_msg(wc->status), wc->status);
1257 nvme_rdma_error_recovery(ctrl);
1260 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1262 if (unlikely(wc->status != IB_WC_SUCCESS))
1263 nvme_rdma_wr_error(cq, wc, "MEMREG");
1266 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1268 struct nvme_rdma_request *req =
1269 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1271 if (unlikely(wc->status != IB_WC_SUCCESS))
1272 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1274 nvme_rdma_end_request(req);
1277 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1278 struct nvme_rdma_request *req)
1280 struct ib_send_wr wr = {
1281 .opcode = IB_WR_LOCAL_INV,
1284 .send_flags = IB_SEND_SIGNALED,
1285 .ex.invalidate_rkey = req->mr->rkey,
1288 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1289 wr.wr_cqe = &req->reg_cqe;
1291 return ib_post_send(queue->qp, &wr, NULL);
1294 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1297 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1298 struct nvme_rdma_device *dev = queue->device;
1299 struct ib_device *ibdev = dev->dev;
1300 struct list_head *pool = &queue->qp->rdma_mrs;
1302 if (!blk_rq_nr_phys_segments(rq))
1305 if (blk_integrity_rq(rq)) {
1306 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1307 req->metadata_sgl->nents, rq_dma_dir(rq));
1308 sg_free_table_chained(&req->metadata_sgl->sg_table,
1309 NVME_INLINE_METADATA_SG_CNT);
1312 if (req->use_sig_mr)
1313 pool = &queue->qp->sig_mrs;
1316 ib_mr_pool_put(queue->qp, pool, req->mr);
1320 ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1322 sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1325 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1327 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1330 put_unaligned_le24(0, sg->length);
1331 put_unaligned_le32(0, sg->key);
1332 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1336 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1337 struct nvme_rdma_request *req, struct nvme_command *c,
1340 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1341 struct ib_sge *sge = &req->sge[1];
1342 struct scatterlist *sgl;
1346 for_each_sg(req->data_sgl.sg_table.sgl, sgl, count, i) {
1347 sge->addr = sg_dma_address(sgl);
1348 sge->length = sg_dma_len(sgl);
1349 sge->lkey = queue->device->pd->local_dma_lkey;
1354 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1355 sg->length = cpu_to_le32(len);
1356 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1358 req->num_sge += count;
1362 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1363 struct nvme_rdma_request *req, struct nvme_command *c)
1365 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1367 sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1368 put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length);
1369 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1370 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1374 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1375 struct nvme_rdma_request *req, struct nvme_command *c,
1378 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1381 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1382 if (WARN_ON_ONCE(!req->mr))
1386 * Align the MR to a 4K page size to match the ctrl page size and
1387 * the block virtual boundary.
1389 nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL,
1391 if (unlikely(nr < count)) {
1392 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1399 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1401 req->reg_cqe.done = nvme_rdma_memreg_done;
1402 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1403 req->reg_wr.wr.opcode = IB_WR_REG_MR;
1404 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1405 req->reg_wr.wr.num_sge = 0;
1406 req->reg_wr.mr = req->mr;
1407 req->reg_wr.key = req->mr->rkey;
1408 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1409 IB_ACCESS_REMOTE_READ |
1410 IB_ACCESS_REMOTE_WRITE;
1412 sg->addr = cpu_to_le64(req->mr->iova);
1413 put_unaligned_le24(req->mr->length, sg->length);
1414 put_unaligned_le32(req->mr->rkey, sg->key);
1415 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1416 NVME_SGL_FMT_INVALIDATE;
1421 static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1422 struct nvme_command *cmd, struct ib_sig_domain *domain,
1423 u16 control, u8 pi_type)
1425 domain->sig_type = IB_SIG_TYPE_T10_DIF;
1426 domain->sig.dif.bg_type = IB_T10DIF_CRC;
1427 domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1428 domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1429 if (control & NVME_RW_PRINFO_PRCHK_REF)
1430 domain->sig.dif.ref_remap = true;
1432 domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
1433 domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
1434 domain->sig.dif.app_escape = true;
1435 if (pi_type == NVME_NS_DPS_PI_TYPE3)
1436 domain->sig.dif.ref_escape = true;
1439 static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1440 struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1443 u16 control = le16_to_cpu(cmd->rw.control);
1445 memset(sig_attrs, 0, sizeof(*sig_attrs));
1446 if (control & NVME_RW_PRINFO_PRACT) {
1447 /* for WRITE_INSERT/READ_STRIP no memory domain */
1448 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1449 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1451 /* Clear the PRACT bit since HCA will generate/verify the PI */
1452 control &= ~NVME_RW_PRINFO_PRACT;
1453 cmd->rw.control = cpu_to_le16(control);
1455 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1456 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1458 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
1463 static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1466 if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1467 *mask |= IB_SIG_CHECK_REFTAG;
1468 if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1469 *mask |= IB_SIG_CHECK_GUARD;
1472 static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1474 if (unlikely(wc->status != IB_WC_SUCCESS))
1475 nvme_rdma_wr_error(cq, wc, "SIG");
1478 static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1479 struct nvme_rdma_request *req, struct nvme_command *c,
1480 int count, int pi_count)
1482 struct nvme_rdma_sgl *sgl = &req->data_sgl;
1483 struct ib_reg_wr *wr = &req->reg_wr;
1484 struct request *rq = blk_mq_rq_from_pdu(req);
1485 struct nvme_ns *ns = rq->q->queuedata;
1486 struct bio *bio = rq->bio;
1487 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1490 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs);
1491 if (WARN_ON_ONCE(!req->mr))
1494 nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL,
1495 req->metadata_sgl->sg_table.sgl, pi_count, NULL,
1500 nvme_rdma_set_sig_attrs(blk_get_integrity(bio->bi_disk), c,
1501 req->mr->sig_attrs, ns->pi_type);
1502 nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask);
1504 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1506 req->reg_cqe.done = nvme_rdma_sig_done;
1507 memset(wr, 0, sizeof(*wr));
1508 wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1509 wr->wr.wr_cqe = &req->reg_cqe;
1511 wr->wr.send_flags = 0;
1513 wr->key = req->mr->rkey;
1514 wr->access = IB_ACCESS_LOCAL_WRITE |
1515 IB_ACCESS_REMOTE_READ |
1516 IB_ACCESS_REMOTE_WRITE;
1518 sg->addr = cpu_to_le64(req->mr->iova);
1519 put_unaligned_le24(req->mr->length, sg->length);
1520 put_unaligned_le32(req->mr->rkey, sg->key);
1521 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1526 ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr);
1533 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1534 struct request *rq, struct nvme_command *c)
1536 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1537 struct nvme_rdma_device *dev = queue->device;
1538 struct ib_device *ibdev = dev->dev;
1543 refcount_set(&req->ref, 2); /* send and recv completions */
1545 c->common.flags |= NVME_CMD_SGL_METABUF;
1547 if (!blk_rq_nr_phys_segments(rq))
1548 return nvme_rdma_set_sg_null(c);
1550 req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1551 ret = sg_alloc_table_chained(&req->data_sgl.sg_table,
1552 blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl,
1553 NVME_INLINE_SG_CNT);
1557 req->data_sgl.nents = blk_rq_map_sg(rq->q, rq,
1558 req->data_sgl.sg_table.sgl);
1560 count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl,
1561 req->data_sgl.nents, rq_dma_dir(rq));
1562 if (unlikely(count <= 0)) {
1564 goto out_free_table;
1567 if (blk_integrity_rq(rq)) {
1568 req->metadata_sgl->sg_table.sgl =
1569 (struct scatterlist *)(req->metadata_sgl + 1);
1570 ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table,
1571 blk_rq_count_integrity_sg(rq->q, rq->bio),
1572 req->metadata_sgl->sg_table.sgl,
1573 NVME_INLINE_METADATA_SG_CNT);
1574 if (unlikely(ret)) {
1579 req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q,
1580 rq->bio, req->metadata_sgl->sg_table.sgl);
1581 pi_count = ib_dma_map_sg(ibdev,
1582 req->metadata_sgl->sg_table.sgl,
1583 req->metadata_sgl->nents,
1585 if (unlikely(pi_count <= 0)) {
1587 goto out_free_pi_table;
1591 if (req->use_sig_mr) {
1592 ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1596 if (count <= dev->num_inline_segments) {
1597 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1598 queue->ctrl->use_inline_data &&
1599 blk_rq_payload_bytes(rq) <=
1600 nvme_rdma_inline_data_size(queue)) {
1601 ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1605 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1606 ret = nvme_rdma_map_sg_single(queue, req, c);
1611 ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1614 goto out_unmap_pi_sg;
1619 if (blk_integrity_rq(rq))
1620 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1621 req->metadata_sgl->nents, rq_dma_dir(rq));
1623 if (blk_integrity_rq(rq))
1624 sg_free_table_chained(&req->metadata_sgl->sg_table,
1625 NVME_INLINE_METADATA_SG_CNT);
1627 ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1630 sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1634 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1636 struct nvme_rdma_qe *qe =
1637 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1638 struct nvme_rdma_request *req =
1639 container_of(qe, struct nvme_rdma_request, sqe);
1641 if (unlikely(wc->status != IB_WC_SUCCESS))
1642 nvme_rdma_wr_error(cq, wc, "SEND");
1644 nvme_rdma_end_request(req);
1647 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1648 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1649 struct ib_send_wr *first)
1651 struct ib_send_wr wr;
1654 sge->addr = qe->dma;
1655 sge->length = sizeof(struct nvme_command);
1656 sge->lkey = queue->device->pd->local_dma_lkey;
1659 wr.wr_cqe = &qe->cqe;
1661 wr.num_sge = num_sge;
1662 wr.opcode = IB_WR_SEND;
1663 wr.send_flags = IB_SEND_SIGNALED;
1670 ret = ib_post_send(queue->qp, first, NULL);
1671 if (unlikely(ret)) {
1672 dev_err(queue->ctrl->ctrl.device,
1673 "%s failed with error code %d\n", __func__, ret);
1678 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1679 struct nvme_rdma_qe *qe)
1681 struct ib_recv_wr wr;
1685 list.addr = qe->dma;
1686 list.length = sizeof(struct nvme_completion);
1687 list.lkey = queue->device->pd->local_dma_lkey;
1689 qe->cqe.done = nvme_rdma_recv_done;
1692 wr.wr_cqe = &qe->cqe;
1696 ret = ib_post_recv(queue->qp, &wr, NULL);
1697 if (unlikely(ret)) {
1698 dev_err(queue->ctrl->ctrl.device,
1699 "%s failed with error code %d\n", __func__, ret);
1704 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1706 u32 queue_idx = nvme_rdma_queue_idx(queue);
1709 return queue->ctrl->admin_tag_set.tags[queue_idx];
1710 return queue->ctrl->tag_set.tags[queue_idx - 1];
1713 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1715 if (unlikely(wc->status != IB_WC_SUCCESS))
1716 nvme_rdma_wr_error(cq, wc, "ASYNC");
1719 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1721 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1722 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1723 struct ib_device *dev = queue->device->dev;
1724 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1725 struct nvme_command *cmd = sqe->data;
1729 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1731 memset(cmd, 0, sizeof(*cmd));
1732 cmd->common.opcode = nvme_admin_async_event;
1733 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1734 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1735 nvme_rdma_set_sg_null(cmd);
1737 sqe->cqe.done = nvme_rdma_async_done;
1739 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1742 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1746 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1747 struct nvme_completion *cqe, struct ib_wc *wc)
1750 struct nvme_rdma_request *req;
1752 rq = nvme_find_rq(nvme_rdma_tagset(queue), cqe->command_id);
1754 dev_err(queue->ctrl->ctrl.device,
1755 "got bad command_id %#x on QP %#x\n",
1756 cqe->command_id, queue->qp->qp_num);
1757 nvme_rdma_error_recovery(queue->ctrl);
1760 req = blk_mq_rq_to_pdu(rq);
1762 req->status = cqe->status;
1763 req->result = cqe->result;
1765 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1766 if (unlikely(!req->mr ||
1767 wc->ex.invalidate_rkey != req->mr->rkey)) {
1768 dev_err(queue->ctrl->ctrl.device,
1769 "Bogus remote invalidation for rkey %#x\n",
1770 req->mr ? req->mr->rkey : 0);
1771 nvme_rdma_error_recovery(queue->ctrl);
1773 } else if (req->mr) {
1776 ret = nvme_rdma_inv_rkey(queue, req);
1777 if (unlikely(ret < 0)) {
1778 dev_err(queue->ctrl->ctrl.device,
1779 "Queueing INV WR for rkey %#x failed (%d)\n",
1780 req->mr->rkey, ret);
1781 nvme_rdma_error_recovery(queue->ctrl);
1783 /* the local invalidation completion will end the request */
1787 nvme_rdma_end_request(req);
1790 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1792 struct nvme_rdma_qe *qe =
1793 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1794 struct nvme_rdma_queue *queue = wc->qp->qp_context;
1795 struct ib_device *ibdev = queue->device->dev;
1796 struct nvme_completion *cqe = qe->data;
1797 const size_t len = sizeof(struct nvme_completion);
1799 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1800 nvme_rdma_wr_error(cq, wc, "RECV");
1804 /* sanity checking for received data length */
1805 if (unlikely(wc->byte_len < len)) {
1806 dev_err(queue->ctrl->ctrl.device,
1807 "Unexpected nvme completion length(%d)\n", wc->byte_len);
1808 nvme_rdma_error_recovery(queue->ctrl);
1812 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1814 * AEN requests are special as they don't time out and can
1815 * survive any kind of queue freeze and often don't respond to
1816 * aborts. We don't even bother to allocate a struct request
1817 * for them but rather special case them here.
1819 if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1821 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1824 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1825 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1827 nvme_rdma_post_recv(queue, qe);
1830 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1834 for (i = 0; i < queue->queue_size; i++) {
1835 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1843 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1844 struct rdma_cm_event *ev)
1846 struct rdma_cm_id *cm_id = queue->cm_id;
1847 int status = ev->status;
1848 const char *rej_msg;
1849 const struct nvme_rdma_cm_rej *rej_data;
1852 rej_msg = rdma_reject_msg(cm_id, status);
1853 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1855 if (rej_data && rej_data_len >= sizeof(u16)) {
1856 u16 sts = le16_to_cpu(rej_data->sts);
1858 dev_err(queue->ctrl->ctrl.device,
1859 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1860 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1862 dev_err(queue->ctrl->ctrl.device,
1863 "Connect rejected: status %d (%s).\n", status, rej_msg);
1869 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1871 struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1874 ret = nvme_rdma_create_queue_ib(queue);
1878 if (ctrl->opts->tos >= 0)
1879 rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1880 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1882 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1884 goto out_destroy_queue;
1890 nvme_rdma_destroy_queue_ib(queue);
1894 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1896 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1897 struct rdma_conn_param param = { };
1898 struct nvme_rdma_cm_req priv = { };
1901 param.qp_num = queue->qp->qp_num;
1902 param.flow_control = 1;
1904 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1905 /* maximum retry count */
1906 param.retry_count = 7;
1907 param.rnr_retry_count = 7;
1908 param.private_data = &priv;
1909 param.private_data_len = sizeof(priv);
1911 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1912 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1914 * set the admin queue depth to the minimum size
1915 * specified by the Fabrics standard.
1917 if (priv.qid == 0) {
1918 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1919 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1922 * current interpretation of the fabrics spec
1923 * is at minimum you make hrqsize sqsize+1, or a
1924 * 1's based representation of sqsize.
1926 priv.hrqsize = cpu_to_le16(queue->queue_size);
1927 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1930 ret = rdma_connect_locked(queue->cm_id, ¶m);
1932 dev_err(ctrl->ctrl.device,
1933 "rdma_connect_locked failed (%d).\n", ret);
1940 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1941 struct rdma_cm_event *ev)
1943 struct nvme_rdma_queue *queue = cm_id->context;
1946 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1947 rdma_event_msg(ev->event), ev->event,
1950 switch (ev->event) {
1951 case RDMA_CM_EVENT_ADDR_RESOLVED:
1952 cm_error = nvme_rdma_addr_resolved(queue);
1954 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1955 cm_error = nvme_rdma_route_resolved(queue);
1957 case RDMA_CM_EVENT_ESTABLISHED:
1958 queue->cm_error = nvme_rdma_conn_established(queue);
1959 /* complete cm_done regardless of success/failure */
1960 complete(&queue->cm_done);
1962 case RDMA_CM_EVENT_REJECTED:
1963 cm_error = nvme_rdma_conn_rejected(queue, ev);
1965 case RDMA_CM_EVENT_ROUTE_ERROR:
1966 case RDMA_CM_EVENT_CONNECT_ERROR:
1967 case RDMA_CM_EVENT_UNREACHABLE:
1968 case RDMA_CM_EVENT_ADDR_ERROR:
1969 dev_dbg(queue->ctrl->ctrl.device,
1970 "CM error event %d\n", ev->event);
1971 cm_error = -ECONNRESET;
1973 case RDMA_CM_EVENT_DISCONNECTED:
1974 case RDMA_CM_EVENT_ADDR_CHANGE:
1975 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1976 dev_dbg(queue->ctrl->ctrl.device,
1977 "disconnect received - connection closed\n");
1978 nvme_rdma_error_recovery(queue->ctrl);
1980 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1981 /* device removal is handled via the ib_client API */
1984 dev_err(queue->ctrl->ctrl.device,
1985 "Unexpected RDMA CM event (%d)\n", ev->event);
1986 nvme_rdma_error_recovery(queue->ctrl);
1991 queue->cm_error = cm_error;
1992 complete(&queue->cm_done);
1998 static void nvme_rdma_complete_timed_out(struct request *rq)
2000 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2001 struct nvme_rdma_queue *queue = req->queue;
2003 nvme_rdma_stop_queue(queue);
2004 if (blk_mq_request_started(rq) && !blk_mq_request_completed(rq)) {
2005 nvme_req(rq)->status = NVME_SC_HOST_ABORTED_CMD;
2006 blk_mq_complete_request(rq);
2010 static enum blk_eh_timer_return
2011 nvme_rdma_timeout(struct request *rq, bool reserved)
2013 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2014 struct nvme_rdma_queue *queue = req->queue;
2015 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
2017 dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
2018 rq->tag, nvme_rdma_queue_idx(queue));
2020 if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
2022 * If we are resetting, connecting or deleting we should
2023 * complete immediately because we may block controller
2024 * teardown or setup sequence
2025 * - ctrl disable/shutdown fabrics requests
2026 * - connect requests
2027 * - initialization admin requests
2028 * - I/O requests that entered after unquiescing and
2029 * the controller stopped responding
2031 * All other requests should be cancelled by the error
2032 * recovery work, so it's fine that we fail it here.
2034 nvme_rdma_complete_timed_out(rq);
2039 * LIVE state should trigger the normal error recovery which will
2040 * handle completing this request.
2042 nvme_rdma_error_recovery(ctrl);
2043 return BLK_EH_RESET_TIMER;
2046 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
2047 const struct blk_mq_queue_data *bd)
2049 struct nvme_ns *ns = hctx->queue->queuedata;
2050 struct nvme_rdma_queue *queue = hctx->driver_data;
2051 struct request *rq = bd->rq;
2052 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2053 struct nvme_rdma_qe *sqe = &req->sqe;
2054 struct nvme_command *c = sqe->data;
2055 struct ib_device *dev;
2056 bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
2060 WARN_ON_ONCE(rq->tag < 0);
2062 if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2063 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
2065 dev = queue->device->dev;
2067 req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
2068 sizeof(struct nvme_command),
2070 err = ib_dma_mapping_error(dev, req->sqe.dma);
2072 return BLK_STS_RESOURCE;
2074 ib_dma_sync_single_for_cpu(dev, sqe->dma,
2075 sizeof(struct nvme_command), DMA_TO_DEVICE);
2077 ret = nvme_setup_cmd(ns, rq, c);
2081 blk_mq_start_request(rq);
2083 if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
2084 queue->pi_support &&
2085 (c->common.opcode == nvme_cmd_write ||
2086 c->common.opcode == nvme_cmd_read) &&
2088 req->use_sig_mr = true;
2090 req->use_sig_mr = false;
2092 err = nvme_rdma_map_data(queue, rq, c);
2093 if (unlikely(err < 0)) {
2094 dev_err(queue->ctrl->ctrl.device,
2095 "Failed to map data (%d)\n", err);
2099 sqe->cqe.done = nvme_rdma_send_done;
2101 ib_dma_sync_single_for_device(dev, sqe->dma,
2102 sizeof(struct nvme_command), DMA_TO_DEVICE);
2104 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
2105 req->mr ? &req->reg_wr.wr : NULL);
2112 nvme_rdma_unmap_data(queue, rq);
2114 if (err == -ENOMEM || err == -EAGAIN)
2115 ret = BLK_STS_RESOURCE;
2117 ret = BLK_STS_IOERR;
2118 nvme_cleanup_cmd(rq);
2120 ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
2125 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
2127 struct nvme_rdma_queue *queue = hctx->driver_data;
2129 return ib_process_cq_direct(queue->ib_cq, -1);
2132 static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req)
2134 struct request *rq = blk_mq_rq_from_pdu(req);
2135 struct ib_mr_status mr_status;
2138 ret = ib_check_mr_status(req->mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
2140 pr_err("ib_check_mr_status failed, ret %d\n", ret);
2141 nvme_req(rq)->status = NVME_SC_INVALID_PI;
2145 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
2146 switch (mr_status.sig_err.err_type) {
2147 case IB_SIG_BAD_GUARD:
2148 nvme_req(rq)->status = NVME_SC_GUARD_CHECK;
2150 case IB_SIG_BAD_REFTAG:
2151 nvme_req(rq)->status = NVME_SC_REFTAG_CHECK;
2153 case IB_SIG_BAD_APPTAG:
2154 nvme_req(rq)->status = NVME_SC_APPTAG_CHECK;
2157 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
2158 mr_status.sig_err.err_type, mr_status.sig_err.expected,
2159 mr_status.sig_err.actual);
2163 static void nvme_rdma_complete_rq(struct request *rq)
2165 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2166 struct nvme_rdma_queue *queue = req->queue;
2167 struct ib_device *ibdev = queue->device->dev;
2169 if (req->use_sig_mr)
2170 nvme_rdma_check_pi_status(req);
2172 nvme_rdma_unmap_data(queue, rq);
2173 ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
2175 nvme_complete_rq(rq);
2178 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
2180 struct nvme_rdma_ctrl *ctrl = set->driver_data;
2181 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2183 if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2184 /* separate read/write queues */
2185 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2186 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2187 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2188 set->map[HCTX_TYPE_READ].nr_queues =
2189 ctrl->io_queues[HCTX_TYPE_READ];
2190 set->map[HCTX_TYPE_READ].queue_offset =
2191 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2193 /* shared read/write queues */
2194 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2195 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2196 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2197 set->map[HCTX_TYPE_READ].nr_queues =
2198 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2199 set->map[HCTX_TYPE_READ].queue_offset = 0;
2201 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
2202 ctrl->device->dev, 0);
2203 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
2204 ctrl->device->dev, 0);
2206 if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2207 /* map dedicated poll queues only if we have queues left */
2208 set->map[HCTX_TYPE_POLL].nr_queues =
2209 ctrl->io_queues[HCTX_TYPE_POLL];
2210 set->map[HCTX_TYPE_POLL].queue_offset =
2211 ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2212 ctrl->io_queues[HCTX_TYPE_READ];
2213 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2216 dev_info(ctrl->ctrl.device,
2217 "mapped %d/%d/%d default/read/poll queues.\n",
2218 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2219 ctrl->io_queues[HCTX_TYPE_READ],
2220 ctrl->io_queues[HCTX_TYPE_POLL]);
2225 static const struct blk_mq_ops nvme_rdma_mq_ops = {
2226 .queue_rq = nvme_rdma_queue_rq,
2227 .complete = nvme_rdma_complete_rq,
2228 .init_request = nvme_rdma_init_request,
2229 .exit_request = nvme_rdma_exit_request,
2230 .init_hctx = nvme_rdma_init_hctx,
2231 .timeout = nvme_rdma_timeout,
2232 .map_queues = nvme_rdma_map_queues,
2233 .poll = nvme_rdma_poll,
2236 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
2237 .queue_rq = nvme_rdma_queue_rq,
2238 .complete = nvme_rdma_complete_rq,
2239 .init_request = nvme_rdma_init_request,
2240 .exit_request = nvme_rdma_exit_request,
2241 .init_hctx = nvme_rdma_init_admin_hctx,
2242 .timeout = nvme_rdma_timeout,
2245 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
2247 nvme_rdma_teardown_io_queues(ctrl, shutdown);
2248 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
2250 nvme_shutdown_ctrl(&ctrl->ctrl);
2252 nvme_disable_ctrl(&ctrl->ctrl);
2253 nvme_rdma_teardown_admin_queue(ctrl, shutdown);
2256 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
2258 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
2261 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
2263 struct nvme_rdma_ctrl *ctrl =
2264 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
2266 nvme_stop_ctrl(&ctrl->ctrl);
2267 nvme_rdma_shutdown_ctrl(ctrl, false);
2269 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2270 /* state change failure should never happen */
2275 if (nvme_rdma_setup_ctrl(ctrl, false))
2281 ++ctrl->ctrl.nr_reconnects;
2282 nvme_rdma_reconnect_or_remove(ctrl);
2285 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
2287 .module = THIS_MODULE,
2288 .flags = NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED,
2289 .reg_read32 = nvmf_reg_read32,
2290 .reg_read64 = nvmf_reg_read64,
2291 .reg_write32 = nvmf_reg_write32,
2292 .free_ctrl = nvme_rdma_free_ctrl,
2293 .submit_async_event = nvme_rdma_submit_async_event,
2294 .delete_ctrl = nvme_rdma_delete_ctrl,
2295 .get_address = nvmf_get_address,
2296 .stop_ctrl = nvme_rdma_stop_ctrl,
2300 * Fails a connection request if it matches an existing controller
2301 * (association) with the same tuple:
2302 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
2304 * if local address is not specified in the request, it will match an
2305 * existing controller with all the other parameters the same and no
2306 * local port address specified as well.
2308 * The ports don't need to be compared as they are intrinsically
2309 * already matched by the port pointers supplied.
2312 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
2314 struct nvme_rdma_ctrl *ctrl;
2317 mutex_lock(&nvme_rdma_ctrl_mutex);
2318 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2319 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2323 mutex_unlock(&nvme_rdma_ctrl_mutex);
2328 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
2329 struct nvmf_ctrl_options *opts)
2331 struct nvme_rdma_ctrl *ctrl;
2335 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2337 return ERR_PTR(-ENOMEM);
2338 ctrl->ctrl.opts = opts;
2339 INIT_LIST_HEAD(&ctrl->list);
2341 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2343 kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
2344 if (!opts->trsvcid) {
2348 opts->mask |= NVMF_OPT_TRSVCID;
2351 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2352 opts->traddr, opts->trsvcid, &ctrl->addr);
2354 pr_err("malformed address passed: %s:%s\n",
2355 opts->traddr, opts->trsvcid);
2359 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2360 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2361 opts->host_traddr, NULL, &ctrl->src_addr);
2363 pr_err("malformed src address passed: %s\n",
2369 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2374 INIT_DELAYED_WORK(&ctrl->reconnect_work,
2375 nvme_rdma_reconnect_ctrl_work);
2376 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2377 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2379 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2380 opts->nr_poll_queues + 1;
2381 ctrl->ctrl.sqsize = opts->queue_size - 1;
2382 ctrl->ctrl.kato = opts->kato;
2385 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2390 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2391 0 /* no quirks, we're perfect! */);
2393 goto out_kfree_queues;
2395 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2396 WARN_ON_ONCE(!changed);
2398 ret = nvme_rdma_setup_ctrl(ctrl, true);
2400 goto out_uninit_ctrl;
2402 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2403 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2405 mutex_lock(&nvme_rdma_ctrl_mutex);
2406 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2407 mutex_unlock(&nvme_rdma_ctrl_mutex);
2412 nvme_uninit_ctrl(&ctrl->ctrl);
2413 nvme_put_ctrl(&ctrl->ctrl);
2416 return ERR_PTR(ret);
2418 kfree(ctrl->queues);
2421 return ERR_PTR(ret);
2424 static struct nvmf_transport_ops nvme_rdma_transport = {
2426 .module = THIS_MODULE,
2427 .required_opts = NVMF_OPT_TRADDR,
2428 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2429 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2430 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2432 .create_ctrl = nvme_rdma_create_ctrl,
2435 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2437 struct nvme_rdma_ctrl *ctrl;
2438 struct nvme_rdma_device *ndev;
2441 mutex_lock(&device_list_mutex);
2442 list_for_each_entry(ndev, &device_list, entry) {
2443 if (ndev->dev == ib_device) {
2448 mutex_unlock(&device_list_mutex);
2453 /* Delete all controllers using this device */
2454 mutex_lock(&nvme_rdma_ctrl_mutex);
2455 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2456 if (ctrl->device->dev != ib_device)
2458 nvme_delete_ctrl(&ctrl->ctrl);
2460 mutex_unlock(&nvme_rdma_ctrl_mutex);
2462 flush_workqueue(nvme_delete_wq);
2465 static struct ib_client nvme_rdma_ib_client = {
2466 .name = "nvme_rdma",
2467 .remove = nvme_rdma_remove_one
2470 static int __init nvme_rdma_init_module(void)
2474 ret = ib_register_client(&nvme_rdma_ib_client);
2478 ret = nvmf_register_transport(&nvme_rdma_transport);
2480 goto err_unreg_client;
2485 ib_unregister_client(&nvme_rdma_ib_client);
2489 static void __exit nvme_rdma_cleanup_module(void)
2491 struct nvme_rdma_ctrl *ctrl;
2493 nvmf_unregister_transport(&nvme_rdma_transport);
2494 ib_unregister_client(&nvme_rdma_ib_client);
2496 mutex_lock(&nvme_rdma_ctrl_mutex);
2497 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2498 nvme_delete_ctrl(&ctrl->ctrl);
2499 mutex_unlock(&nvme_rdma_ctrl_mutex);
2500 flush_workqueue(nvme_delete_wq);
2503 module_init(nvme_rdma_init_module);
2504 module_exit(nvme_rdma_cleanup_module);
2506 MODULE_LICENSE("GPL v2");