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_CONNECT_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_timeout(&queue->cm_done,
252 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
257 WARN_ON_ONCE(queue->cm_error > 0);
258 return queue->cm_error;
261 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
263 struct nvme_rdma_device *dev = queue->device;
264 struct ib_qp_init_attr init_attr;
267 memset(&init_attr, 0, sizeof(init_attr));
268 init_attr.event_handler = nvme_rdma_qp_event;
270 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
272 init_attr.cap.max_recv_wr = queue->queue_size + 1;
273 init_attr.cap.max_recv_sge = 1;
274 init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
275 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
276 init_attr.qp_type = IB_QPT_RC;
277 init_attr.send_cq = queue->ib_cq;
278 init_attr.recv_cq = queue->ib_cq;
279 if (queue->pi_support)
280 init_attr.create_flags |= IB_QP_CREATE_INTEGRITY_EN;
281 init_attr.qp_context = queue;
283 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
285 queue->qp = queue->cm_id->qp;
289 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
290 struct request *rq, unsigned int hctx_idx)
292 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
294 kfree(req->sqe.data);
297 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
298 struct request *rq, unsigned int hctx_idx,
299 unsigned int numa_node)
301 struct nvme_rdma_ctrl *ctrl = set->driver_data;
302 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
303 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
304 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
306 nvme_req(rq)->ctrl = &ctrl->ctrl;
307 req->sqe.data = kzalloc(sizeof(struct nvme_command), GFP_KERNEL);
311 /* metadata nvme_rdma_sgl struct is located after command's data SGL */
312 if (queue->pi_support)
313 req->metadata_sgl = (void *)nvme_req(rq) +
314 sizeof(struct nvme_rdma_request) +
315 NVME_RDMA_DATA_SGL_SIZE;
318 nvme_req(rq)->cmd = req->sqe.data;
323 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
324 unsigned int hctx_idx)
326 struct nvme_rdma_ctrl *ctrl = data;
327 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
329 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
331 hctx->driver_data = queue;
335 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
336 unsigned int hctx_idx)
338 struct nvme_rdma_ctrl *ctrl = data;
339 struct nvme_rdma_queue *queue = &ctrl->queues[0];
341 BUG_ON(hctx_idx != 0);
343 hctx->driver_data = queue;
347 static void nvme_rdma_free_dev(struct kref *ref)
349 struct nvme_rdma_device *ndev =
350 container_of(ref, struct nvme_rdma_device, ref);
352 mutex_lock(&device_list_mutex);
353 list_del(&ndev->entry);
354 mutex_unlock(&device_list_mutex);
356 ib_dealloc_pd(ndev->pd);
360 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
362 kref_put(&dev->ref, nvme_rdma_free_dev);
365 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
367 return kref_get_unless_zero(&dev->ref);
370 static struct nvme_rdma_device *
371 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
373 struct nvme_rdma_device *ndev;
375 mutex_lock(&device_list_mutex);
376 list_for_each_entry(ndev, &device_list, entry) {
377 if (ndev->dev->node_guid == cm_id->device->node_guid &&
378 nvme_rdma_dev_get(ndev))
382 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
386 ndev->dev = cm_id->device;
387 kref_init(&ndev->ref);
389 ndev->pd = ib_alloc_pd(ndev->dev,
390 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
391 if (IS_ERR(ndev->pd))
394 if (!(ndev->dev->attrs.device_cap_flags &
395 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
396 dev_err(&ndev->dev->dev,
397 "Memory registrations not supported.\n");
401 ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
402 ndev->dev->attrs.max_send_sge - 1);
403 list_add(&ndev->entry, &device_list);
405 mutex_unlock(&device_list_mutex);
409 ib_dealloc_pd(ndev->pd);
413 mutex_unlock(&device_list_mutex);
417 static void nvme_rdma_free_cq(struct nvme_rdma_queue *queue)
419 if (nvme_rdma_poll_queue(queue))
420 ib_free_cq(queue->ib_cq);
422 ib_cq_pool_put(queue->ib_cq, queue->cq_size);
425 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
427 struct nvme_rdma_device *dev;
428 struct ib_device *ibdev;
430 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
436 if (queue->pi_support)
437 ib_mr_pool_destroy(queue->qp, &queue->qp->sig_mrs);
438 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
441 * The cm_id object might have been destroyed during RDMA connection
442 * establishment error flow to avoid getting other cma events, thus
443 * the destruction of the QP shouldn't use rdma_cm API.
445 ib_destroy_qp(queue->qp);
446 nvme_rdma_free_cq(queue);
448 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
449 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
451 nvme_rdma_dev_put(dev);
454 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev, bool pi_support)
456 u32 max_page_list_len;
459 max_page_list_len = ibdev->attrs.max_pi_fast_reg_page_list_len;
461 max_page_list_len = ibdev->attrs.max_fast_reg_page_list_len;
463 return min_t(u32, NVME_RDMA_MAX_SEGMENTS, max_page_list_len - 1);
466 static int nvme_rdma_create_cq(struct ib_device *ibdev,
467 struct nvme_rdma_queue *queue)
469 int ret, comp_vector, idx = nvme_rdma_queue_idx(queue);
470 enum ib_poll_context poll_ctx;
473 * Spread I/O queues completion vectors according their queue index.
474 * Admin queues can always go on completion vector 0.
476 comp_vector = (idx == 0 ? idx : idx - 1) % ibdev->num_comp_vectors;
478 /* Polling queues need direct cq polling context */
479 if (nvme_rdma_poll_queue(queue)) {
480 poll_ctx = IB_POLL_DIRECT;
481 queue->ib_cq = ib_alloc_cq(ibdev, queue, queue->cq_size,
482 comp_vector, poll_ctx);
484 poll_ctx = IB_POLL_SOFTIRQ;
485 queue->ib_cq = ib_cq_pool_get(ibdev, queue->cq_size,
486 comp_vector, poll_ctx);
489 if (IS_ERR(queue->ib_cq)) {
490 ret = PTR_ERR(queue->ib_cq);
497 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
499 struct ib_device *ibdev;
500 const int send_wr_factor = 3; /* MR, SEND, INV */
501 const int cq_factor = send_wr_factor + 1; /* + RECV */
502 int ret, pages_per_mr;
504 queue->device = nvme_rdma_find_get_device(queue->cm_id);
505 if (!queue->device) {
506 dev_err(queue->cm_id->device->dev.parent,
507 "no client data found!\n");
508 return -ECONNREFUSED;
510 ibdev = queue->device->dev;
512 /* +1 for ib_stop_cq */
513 queue->cq_size = cq_factor * queue->queue_size + 1;
515 ret = nvme_rdma_create_cq(ibdev, queue);
519 ret = nvme_rdma_create_qp(queue, send_wr_factor);
521 goto out_destroy_ib_cq;
523 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
524 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
525 if (!queue->rsp_ring) {
531 * Currently we don't use SG_GAPS MR's so if the first entry is
532 * misaligned we'll end up using two entries for a single data page,
533 * so one additional entry is required.
535 pages_per_mr = nvme_rdma_get_max_fr_pages(ibdev, queue->pi_support) + 1;
536 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
541 dev_err(queue->ctrl->ctrl.device,
542 "failed to initialize MR pool sized %d for QID %d\n",
543 queue->queue_size, nvme_rdma_queue_idx(queue));
544 goto out_destroy_ring;
547 if (queue->pi_support) {
548 ret = ib_mr_pool_init(queue->qp, &queue->qp->sig_mrs,
549 queue->queue_size, IB_MR_TYPE_INTEGRITY,
550 pages_per_mr, pages_per_mr);
552 dev_err(queue->ctrl->ctrl.device,
553 "failed to initialize PI MR pool sized %d for QID %d\n",
554 queue->queue_size, nvme_rdma_queue_idx(queue));
555 goto out_destroy_mr_pool;
559 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
564 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
566 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
567 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
569 rdma_destroy_qp(queue->cm_id);
571 nvme_rdma_free_cq(queue);
573 nvme_rdma_dev_put(queue->device);
577 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
578 int idx, size_t queue_size)
580 struct nvme_rdma_queue *queue;
581 struct sockaddr *src_addr = NULL;
584 queue = &ctrl->queues[idx];
585 mutex_init(&queue->queue_lock);
587 if (idx && ctrl->ctrl.max_integrity_segments)
588 queue->pi_support = true;
590 queue->pi_support = false;
591 init_completion(&queue->cm_done);
594 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
596 queue->cmnd_capsule_len = sizeof(struct nvme_command);
598 queue->queue_size = queue_size;
600 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
601 RDMA_PS_TCP, IB_QPT_RC);
602 if (IS_ERR(queue->cm_id)) {
603 dev_info(ctrl->ctrl.device,
604 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
605 ret = PTR_ERR(queue->cm_id);
606 goto out_destroy_mutex;
609 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
610 src_addr = (struct sockaddr *)&ctrl->src_addr;
612 queue->cm_error = -ETIMEDOUT;
613 ret = rdma_resolve_addr(queue->cm_id, src_addr,
614 (struct sockaddr *)&ctrl->addr,
615 NVME_RDMA_CONNECT_TIMEOUT_MS);
617 dev_info(ctrl->ctrl.device,
618 "rdma_resolve_addr failed (%d).\n", ret);
619 goto out_destroy_cm_id;
622 ret = nvme_rdma_wait_for_cm(queue);
624 dev_info(ctrl->ctrl.device,
625 "rdma connection establishment failed (%d)\n", ret);
626 goto out_destroy_cm_id;
629 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
634 rdma_destroy_id(queue->cm_id);
635 nvme_rdma_destroy_queue_ib(queue);
637 mutex_destroy(&queue->queue_lock);
641 static void __nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
643 rdma_disconnect(queue->cm_id);
644 ib_drain_qp(queue->qp);
647 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
649 mutex_lock(&queue->queue_lock);
650 if (test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
651 __nvme_rdma_stop_queue(queue);
652 mutex_unlock(&queue->queue_lock);
655 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
657 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
660 rdma_destroy_id(queue->cm_id);
661 nvme_rdma_destroy_queue_ib(queue);
662 mutex_destroy(&queue->queue_lock);
665 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
669 for (i = 1; i < ctrl->ctrl.queue_count; i++)
670 nvme_rdma_free_queue(&ctrl->queues[i]);
673 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
677 for (i = 1; i < ctrl->ctrl.queue_count; i++)
678 nvme_rdma_stop_queue(&ctrl->queues[i]);
681 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
683 struct nvme_rdma_queue *queue = &ctrl->queues[idx];
687 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
689 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
692 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
694 if (test_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
695 __nvme_rdma_stop_queue(queue);
696 dev_info(ctrl->ctrl.device,
697 "failed to connect queue: %d ret=%d\n", idx, ret);
702 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
706 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
707 ret = nvme_rdma_start_queue(ctrl, i);
709 goto out_stop_queues;
715 for (i--; i >= 1; i--)
716 nvme_rdma_stop_queue(&ctrl->queues[i]);
720 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
722 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
723 struct ib_device *ibdev = ctrl->device->dev;
724 unsigned int nr_io_queues, nr_default_queues;
725 unsigned int nr_read_queues, nr_poll_queues;
728 nr_read_queues = min_t(unsigned int, ibdev->num_comp_vectors,
729 min(opts->nr_io_queues, num_online_cpus()));
730 nr_default_queues = min_t(unsigned int, ibdev->num_comp_vectors,
731 min(opts->nr_write_queues, num_online_cpus()));
732 nr_poll_queues = min(opts->nr_poll_queues, num_online_cpus());
733 nr_io_queues = nr_read_queues + nr_default_queues + nr_poll_queues;
735 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
739 if (nr_io_queues == 0) {
740 dev_err(ctrl->ctrl.device,
741 "unable to set any I/O queues\n");
745 ctrl->ctrl.queue_count = nr_io_queues + 1;
746 dev_info(ctrl->ctrl.device,
747 "creating %d I/O queues.\n", nr_io_queues);
749 if (opts->nr_write_queues && nr_read_queues < nr_io_queues) {
751 * separate read/write queues
752 * hand out dedicated default queues only after we have
753 * sufficient read queues.
755 ctrl->io_queues[HCTX_TYPE_READ] = nr_read_queues;
756 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_READ];
757 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
758 min(nr_default_queues, nr_io_queues);
759 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
762 * shared read/write queues
763 * either no write queues were requested, or we don't have
764 * sufficient queue count to have dedicated default queues.
766 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
767 min(nr_read_queues, nr_io_queues);
768 nr_io_queues -= ctrl->io_queues[HCTX_TYPE_DEFAULT];
771 if (opts->nr_poll_queues && nr_io_queues) {
772 /* map dedicated poll queues only if we have queues left */
773 ctrl->io_queues[HCTX_TYPE_POLL] =
774 min(nr_poll_queues, nr_io_queues);
777 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
778 ret = nvme_rdma_alloc_queue(ctrl, i,
779 ctrl->ctrl.sqsize + 1);
781 goto out_free_queues;
787 for (i--; i >= 1; i--)
788 nvme_rdma_free_queue(&ctrl->queues[i]);
793 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
796 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
797 struct blk_mq_tag_set *set;
801 set = &ctrl->admin_tag_set;
802 memset(set, 0, sizeof(*set));
803 set->ops = &nvme_rdma_admin_mq_ops;
804 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
805 set->reserved_tags = NVMF_RESERVED_TAGS;
806 set->numa_node = nctrl->numa_node;
807 set->cmd_size = sizeof(struct nvme_rdma_request) +
808 NVME_RDMA_DATA_SGL_SIZE;
809 set->driver_data = ctrl;
810 set->nr_hw_queues = 1;
811 set->timeout = NVME_ADMIN_TIMEOUT;
812 set->flags = BLK_MQ_F_NO_SCHED;
814 set = &ctrl->tag_set;
815 memset(set, 0, sizeof(*set));
816 set->ops = &nvme_rdma_mq_ops;
817 set->queue_depth = nctrl->sqsize + 1;
818 set->reserved_tags = NVMF_RESERVED_TAGS;
819 set->numa_node = nctrl->numa_node;
820 set->flags = BLK_MQ_F_SHOULD_MERGE;
821 set->cmd_size = sizeof(struct nvme_rdma_request) +
822 NVME_RDMA_DATA_SGL_SIZE;
823 if (nctrl->max_integrity_segments)
824 set->cmd_size += sizeof(struct nvme_rdma_sgl) +
825 NVME_RDMA_METADATA_SGL_SIZE;
826 set->driver_data = ctrl;
827 set->nr_hw_queues = nctrl->queue_count - 1;
828 set->timeout = NVME_IO_TIMEOUT;
829 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
832 ret = blk_mq_alloc_tag_set(set);
839 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
843 blk_cleanup_queue(ctrl->ctrl.admin_q);
844 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
845 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
847 if (ctrl->async_event_sqe.data) {
848 cancel_work_sync(&ctrl->ctrl.async_event_work);
849 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
850 sizeof(struct nvme_command), DMA_TO_DEVICE);
851 ctrl->async_event_sqe.data = NULL;
853 nvme_rdma_free_queue(&ctrl->queues[0]);
856 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
859 bool pi_capable = false;
862 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
866 ctrl->device = ctrl->queues[0].device;
867 ctrl->ctrl.numa_node = ibdev_to_node(ctrl->device->dev);
870 if (ctrl->device->dev->attrs.kernel_cap_flags &
871 IBK_INTEGRITY_HANDOVER)
874 ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev,
878 * Bind the async event SQE DMA mapping to the admin queue lifetime.
879 * It's safe, since any chage in the underlying RDMA device will issue
880 * error recovery and queue re-creation.
882 error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
883 sizeof(struct nvme_command), DMA_TO_DEVICE);
888 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
889 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
890 error = PTR_ERR(ctrl->ctrl.admin_tagset);
891 goto out_free_async_qe;
894 ctrl->ctrl.fabrics_q = blk_mq_init_queue(&ctrl->admin_tag_set);
895 if (IS_ERR(ctrl->ctrl.fabrics_q)) {
896 error = PTR_ERR(ctrl->ctrl.fabrics_q);
897 goto out_free_tagset;
900 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
901 if (IS_ERR(ctrl->ctrl.admin_q)) {
902 error = PTR_ERR(ctrl->ctrl.admin_q);
903 goto out_cleanup_fabrics_q;
907 error = nvme_rdma_start_queue(ctrl, 0);
909 goto out_cleanup_queue;
911 error = nvme_enable_ctrl(&ctrl->ctrl);
915 ctrl->ctrl.max_segments = ctrl->max_fr_pages;
916 ctrl->ctrl.max_hw_sectors = ctrl->max_fr_pages << (ilog2(SZ_4K) - 9);
918 ctrl->ctrl.max_integrity_segments = ctrl->max_fr_pages;
920 ctrl->ctrl.max_integrity_segments = 0;
922 nvme_start_admin_queue(&ctrl->ctrl);
924 error = nvme_init_ctrl_finish(&ctrl->ctrl);
926 goto out_quiesce_queue;
931 nvme_stop_admin_queue(&ctrl->ctrl);
932 blk_sync_queue(ctrl->ctrl.admin_q);
934 nvme_rdma_stop_queue(&ctrl->queues[0]);
935 nvme_cancel_admin_tagset(&ctrl->ctrl);
938 blk_cleanup_queue(ctrl->ctrl.admin_q);
939 out_cleanup_fabrics_q:
941 blk_cleanup_queue(ctrl->ctrl.fabrics_q);
944 blk_mq_free_tag_set(ctrl->ctrl.admin_tagset);
946 if (ctrl->async_event_sqe.data) {
947 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
948 sizeof(struct nvme_command), DMA_TO_DEVICE);
949 ctrl->async_event_sqe.data = NULL;
952 nvme_rdma_free_queue(&ctrl->queues[0]);
956 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
960 blk_cleanup_queue(ctrl->ctrl.connect_q);
961 blk_mq_free_tag_set(ctrl->ctrl.tagset);
963 nvme_rdma_free_io_queues(ctrl);
966 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
970 ret = nvme_rdma_alloc_io_queues(ctrl);
975 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
976 if (IS_ERR(ctrl->ctrl.tagset)) {
977 ret = PTR_ERR(ctrl->ctrl.tagset);
978 goto out_free_io_queues;
981 ret = nvme_ctrl_init_connect_q(&(ctrl->ctrl));
983 goto out_free_tag_set;
986 ret = nvme_rdma_start_io_queues(ctrl);
988 goto out_cleanup_connect_q;
991 nvme_start_queues(&ctrl->ctrl);
992 if (!nvme_wait_freeze_timeout(&ctrl->ctrl, NVME_IO_TIMEOUT)) {
994 * If we timed out waiting for freeze we are likely to
995 * be stuck. Fail the controller initialization just
999 goto out_wait_freeze_timed_out;
1001 blk_mq_update_nr_hw_queues(ctrl->ctrl.tagset,
1002 ctrl->ctrl.queue_count - 1);
1003 nvme_unfreeze(&ctrl->ctrl);
1008 out_wait_freeze_timed_out:
1009 nvme_stop_queues(&ctrl->ctrl);
1010 nvme_sync_io_queues(&ctrl->ctrl);
1011 nvme_rdma_stop_io_queues(ctrl);
1012 out_cleanup_connect_q:
1013 nvme_cancel_tagset(&ctrl->ctrl);
1015 blk_cleanup_queue(ctrl->ctrl.connect_q);
1018 blk_mq_free_tag_set(ctrl->ctrl.tagset);
1020 nvme_rdma_free_io_queues(ctrl);
1024 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
1027 nvme_stop_admin_queue(&ctrl->ctrl);
1028 blk_sync_queue(ctrl->ctrl.admin_q);
1029 nvme_rdma_stop_queue(&ctrl->queues[0]);
1030 nvme_cancel_admin_tagset(&ctrl->ctrl);
1032 nvme_start_admin_queue(&ctrl->ctrl);
1033 nvme_rdma_destroy_admin_queue(ctrl, remove);
1036 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
1039 if (ctrl->ctrl.queue_count > 1) {
1040 nvme_start_freeze(&ctrl->ctrl);
1041 nvme_stop_queues(&ctrl->ctrl);
1042 nvme_sync_io_queues(&ctrl->ctrl);
1043 nvme_rdma_stop_io_queues(ctrl);
1044 nvme_cancel_tagset(&ctrl->ctrl);
1046 nvme_start_queues(&ctrl->ctrl);
1047 nvme_rdma_destroy_io_queues(ctrl, remove);
1051 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
1053 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1055 if (list_empty(&ctrl->list))
1058 mutex_lock(&nvme_rdma_ctrl_mutex);
1059 list_del(&ctrl->list);
1060 mutex_unlock(&nvme_rdma_ctrl_mutex);
1062 nvmf_free_options(nctrl->opts);
1064 kfree(ctrl->queues);
1068 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
1070 /* If we are resetting/deleting then do nothing */
1071 if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
1072 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
1073 ctrl->ctrl.state == NVME_CTRL_LIVE);
1077 if (nvmf_should_reconnect(&ctrl->ctrl)) {
1078 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
1079 ctrl->ctrl.opts->reconnect_delay);
1080 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
1081 ctrl->ctrl.opts->reconnect_delay * HZ);
1083 nvme_delete_ctrl(&ctrl->ctrl);
1087 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
1092 ret = nvme_rdma_configure_admin_queue(ctrl, new);
1096 if (ctrl->ctrl.icdoff) {
1098 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1102 if (!(ctrl->ctrl.sgls & (1 << 2))) {
1104 dev_err(ctrl->ctrl.device,
1105 "Mandatory keyed sgls are not supported!\n");
1109 if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
1110 dev_warn(ctrl->ctrl.device,
1111 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1112 ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
1115 if (ctrl->ctrl.sqsize + 1 > NVME_RDMA_MAX_QUEUE_SIZE) {
1116 dev_warn(ctrl->ctrl.device,
1117 "ctrl sqsize %u > max queue size %u, clamping down\n",
1118 ctrl->ctrl.sqsize + 1, NVME_RDMA_MAX_QUEUE_SIZE);
1119 ctrl->ctrl.sqsize = NVME_RDMA_MAX_QUEUE_SIZE - 1;
1122 if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1123 dev_warn(ctrl->ctrl.device,
1124 "sqsize %u > ctrl maxcmd %u, clamping down\n",
1125 ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1126 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1129 if (ctrl->ctrl.sgls & (1 << 20))
1130 ctrl->use_inline_data = true;
1132 if (ctrl->ctrl.queue_count > 1) {
1133 ret = nvme_rdma_configure_io_queues(ctrl, new);
1138 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1141 * state change failure is ok if we started ctrl delete,
1142 * unless we're during creation of a new controller to
1143 * avoid races with teardown flow.
1145 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1146 ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1152 nvme_start_ctrl(&ctrl->ctrl);
1156 if (ctrl->ctrl.queue_count > 1) {
1157 nvme_stop_queues(&ctrl->ctrl);
1158 nvme_sync_io_queues(&ctrl->ctrl);
1159 nvme_rdma_stop_io_queues(ctrl);
1160 nvme_cancel_tagset(&ctrl->ctrl);
1161 nvme_rdma_destroy_io_queues(ctrl, new);
1164 nvme_stop_admin_queue(&ctrl->ctrl);
1165 blk_sync_queue(ctrl->ctrl.admin_q);
1166 nvme_rdma_stop_queue(&ctrl->queues[0]);
1167 nvme_cancel_admin_tagset(&ctrl->ctrl);
1168 nvme_rdma_destroy_admin_queue(ctrl, new);
1172 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1174 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1175 struct nvme_rdma_ctrl, reconnect_work);
1177 ++ctrl->ctrl.nr_reconnects;
1179 if (nvme_rdma_setup_ctrl(ctrl, false))
1182 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1183 ctrl->ctrl.nr_reconnects);
1185 ctrl->ctrl.nr_reconnects = 0;
1190 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1191 ctrl->ctrl.nr_reconnects);
1192 nvme_rdma_reconnect_or_remove(ctrl);
1195 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1197 struct nvme_rdma_ctrl *ctrl = container_of(work,
1198 struct nvme_rdma_ctrl, err_work);
1200 nvme_stop_keep_alive(&ctrl->ctrl);
1201 flush_work(&ctrl->ctrl.async_event_work);
1202 nvme_rdma_teardown_io_queues(ctrl, false);
1203 nvme_start_queues(&ctrl->ctrl);
1204 nvme_rdma_teardown_admin_queue(ctrl, false);
1205 nvme_start_admin_queue(&ctrl->ctrl);
1207 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1208 /* state change failure is ok if we started ctrl delete */
1209 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING &&
1210 ctrl->ctrl.state != NVME_CTRL_DELETING_NOIO);
1214 nvme_rdma_reconnect_or_remove(ctrl);
1217 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1219 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1222 dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1223 queue_work(nvme_reset_wq, &ctrl->err_work);
1226 static void nvme_rdma_end_request(struct nvme_rdma_request *req)
1228 struct request *rq = blk_mq_rq_from_pdu(req);
1230 if (!refcount_dec_and_test(&req->ref))
1232 if (!nvme_try_complete_req(rq, req->status, req->result))
1233 nvme_rdma_complete_rq(rq);
1236 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1239 struct nvme_rdma_queue *queue = wc->qp->qp_context;
1240 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1242 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1243 dev_info(ctrl->ctrl.device,
1244 "%s for CQE 0x%p failed with status %s (%d)\n",
1246 ib_wc_status_msg(wc->status), wc->status);
1247 nvme_rdma_error_recovery(ctrl);
1250 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1252 if (unlikely(wc->status != IB_WC_SUCCESS))
1253 nvme_rdma_wr_error(cq, wc, "MEMREG");
1256 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1258 struct nvme_rdma_request *req =
1259 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1261 if (unlikely(wc->status != IB_WC_SUCCESS))
1262 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1264 nvme_rdma_end_request(req);
1267 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1268 struct nvme_rdma_request *req)
1270 struct ib_send_wr wr = {
1271 .opcode = IB_WR_LOCAL_INV,
1274 .send_flags = IB_SEND_SIGNALED,
1275 .ex.invalidate_rkey = req->mr->rkey,
1278 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1279 wr.wr_cqe = &req->reg_cqe;
1281 return ib_post_send(queue->qp, &wr, NULL);
1284 static void nvme_rdma_dma_unmap_req(struct ib_device *ibdev, struct request *rq)
1286 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1288 if (blk_integrity_rq(rq)) {
1289 ib_dma_unmap_sg(ibdev, req->metadata_sgl->sg_table.sgl,
1290 req->metadata_sgl->nents, rq_dma_dir(rq));
1291 sg_free_table_chained(&req->metadata_sgl->sg_table,
1292 NVME_INLINE_METADATA_SG_CNT);
1295 ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1297 sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1300 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1303 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1304 struct nvme_rdma_device *dev = queue->device;
1305 struct ib_device *ibdev = dev->dev;
1306 struct list_head *pool = &queue->qp->rdma_mrs;
1308 if (!blk_rq_nr_phys_segments(rq))
1311 if (req->use_sig_mr)
1312 pool = &queue->qp->sig_mrs;
1315 ib_mr_pool_put(queue->qp, pool, req->mr);
1319 nvme_rdma_dma_unmap_req(ibdev, rq);
1322 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1324 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1327 put_unaligned_le24(0, sg->length);
1328 put_unaligned_le32(0, sg->key);
1329 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1333 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1334 struct nvme_rdma_request *req, struct nvme_command *c,
1337 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1338 struct ib_sge *sge = &req->sge[1];
1339 struct scatterlist *sgl;
1343 for_each_sg(req->data_sgl.sg_table.sgl, sgl, count, i) {
1344 sge->addr = sg_dma_address(sgl);
1345 sge->length = sg_dma_len(sgl);
1346 sge->lkey = queue->device->pd->local_dma_lkey;
1351 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1352 sg->length = cpu_to_le32(len);
1353 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1355 req->num_sge += count;
1359 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1360 struct nvme_rdma_request *req, struct nvme_command *c)
1362 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1364 sg->addr = cpu_to_le64(sg_dma_address(req->data_sgl.sg_table.sgl));
1365 put_unaligned_le24(sg_dma_len(req->data_sgl.sg_table.sgl), sg->length);
1366 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1367 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1371 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1372 struct nvme_rdma_request *req, struct nvme_command *c,
1375 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1378 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1379 if (WARN_ON_ONCE(!req->mr))
1383 * Align the MR to a 4K page size to match the ctrl page size and
1384 * the block virtual boundary.
1386 nr = ib_map_mr_sg(req->mr, req->data_sgl.sg_table.sgl, count, NULL,
1388 if (unlikely(nr < count)) {
1389 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1396 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1398 req->reg_cqe.done = nvme_rdma_memreg_done;
1399 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1400 req->reg_wr.wr.opcode = IB_WR_REG_MR;
1401 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1402 req->reg_wr.wr.num_sge = 0;
1403 req->reg_wr.mr = req->mr;
1404 req->reg_wr.key = req->mr->rkey;
1405 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1406 IB_ACCESS_REMOTE_READ |
1407 IB_ACCESS_REMOTE_WRITE;
1409 sg->addr = cpu_to_le64(req->mr->iova);
1410 put_unaligned_le24(req->mr->length, sg->length);
1411 put_unaligned_le32(req->mr->rkey, sg->key);
1412 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1413 NVME_SGL_FMT_INVALIDATE;
1418 static void nvme_rdma_set_sig_domain(struct blk_integrity *bi,
1419 struct nvme_command *cmd, struct ib_sig_domain *domain,
1420 u16 control, u8 pi_type)
1422 domain->sig_type = IB_SIG_TYPE_T10_DIF;
1423 domain->sig.dif.bg_type = IB_T10DIF_CRC;
1424 domain->sig.dif.pi_interval = 1 << bi->interval_exp;
1425 domain->sig.dif.ref_tag = le32_to_cpu(cmd->rw.reftag);
1426 if (control & NVME_RW_PRINFO_PRCHK_REF)
1427 domain->sig.dif.ref_remap = true;
1429 domain->sig.dif.app_tag = le16_to_cpu(cmd->rw.apptag);
1430 domain->sig.dif.apptag_check_mask = le16_to_cpu(cmd->rw.appmask);
1431 domain->sig.dif.app_escape = true;
1432 if (pi_type == NVME_NS_DPS_PI_TYPE3)
1433 domain->sig.dif.ref_escape = true;
1436 static void nvme_rdma_set_sig_attrs(struct blk_integrity *bi,
1437 struct nvme_command *cmd, struct ib_sig_attrs *sig_attrs,
1440 u16 control = le16_to_cpu(cmd->rw.control);
1442 memset(sig_attrs, 0, sizeof(*sig_attrs));
1443 if (control & NVME_RW_PRINFO_PRACT) {
1444 /* for WRITE_INSERT/READ_STRIP no memory domain */
1445 sig_attrs->mem.sig_type = IB_SIG_TYPE_NONE;
1446 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1448 /* Clear the PRACT bit since HCA will generate/verify the PI */
1449 control &= ~NVME_RW_PRINFO_PRACT;
1450 cmd->rw.control = cpu_to_le16(control);
1452 /* for WRITE_PASS/READ_PASS both wire/memory domains exist */
1453 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->wire, control,
1455 nvme_rdma_set_sig_domain(bi, cmd, &sig_attrs->mem, control,
1460 static void nvme_rdma_set_prot_checks(struct nvme_command *cmd, u8 *mask)
1463 if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_REF)
1464 *mask |= IB_SIG_CHECK_REFTAG;
1465 if (le16_to_cpu(cmd->rw.control) & NVME_RW_PRINFO_PRCHK_GUARD)
1466 *mask |= IB_SIG_CHECK_GUARD;
1469 static void nvme_rdma_sig_done(struct ib_cq *cq, struct ib_wc *wc)
1471 if (unlikely(wc->status != IB_WC_SUCCESS))
1472 nvme_rdma_wr_error(cq, wc, "SIG");
1475 static int nvme_rdma_map_sg_pi(struct nvme_rdma_queue *queue,
1476 struct nvme_rdma_request *req, struct nvme_command *c,
1477 int count, int pi_count)
1479 struct nvme_rdma_sgl *sgl = &req->data_sgl;
1480 struct ib_reg_wr *wr = &req->reg_wr;
1481 struct request *rq = blk_mq_rq_from_pdu(req);
1482 struct nvme_ns *ns = rq->q->queuedata;
1483 struct bio *bio = rq->bio;
1484 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1487 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->sig_mrs);
1488 if (WARN_ON_ONCE(!req->mr))
1491 nr = ib_map_mr_sg_pi(req->mr, sgl->sg_table.sgl, count, NULL,
1492 req->metadata_sgl->sg_table.sgl, pi_count, NULL,
1497 nvme_rdma_set_sig_attrs(blk_get_integrity(bio->bi_bdev->bd_disk), c,
1498 req->mr->sig_attrs, ns->pi_type);
1499 nvme_rdma_set_prot_checks(c, &req->mr->sig_attrs->check_mask);
1501 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1503 req->reg_cqe.done = nvme_rdma_sig_done;
1504 memset(wr, 0, sizeof(*wr));
1505 wr->wr.opcode = IB_WR_REG_MR_INTEGRITY;
1506 wr->wr.wr_cqe = &req->reg_cqe;
1508 wr->wr.send_flags = 0;
1510 wr->key = req->mr->rkey;
1511 wr->access = IB_ACCESS_LOCAL_WRITE |
1512 IB_ACCESS_REMOTE_READ |
1513 IB_ACCESS_REMOTE_WRITE;
1515 sg->addr = cpu_to_le64(req->mr->iova);
1516 put_unaligned_le24(req->mr->length, sg->length);
1517 put_unaligned_le32(req->mr->rkey, sg->key);
1518 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1523 ib_mr_pool_put(queue->qp, &queue->qp->sig_mrs, req->mr);
1530 static int nvme_rdma_dma_map_req(struct ib_device *ibdev, struct request *rq,
1531 int *count, int *pi_count)
1533 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1536 req->data_sgl.sg_table.sgl = (struct scatterlist *)(req + 1);
1537 ret = sg_alloc_table_chained(&req->data_sgl.sg_table,
1538 blk_rq_nr_phys_segments(rq), req->data_sgl.sg_table.sgl,
1539 NVME_INLINE_SG_CNT);
1543 req->data_sgl.nents = blk_rq_map_sg(rq->q, rq,
1544 req->data_sgl.sg_table.sgl);
1546 *count = ib_dma_map_sg(ibdev, req->data_sgl.sg_table.sgl,
1547 req->data_sgl.nents, rq_dma_dir(rq));
1548 if (unlikely(*count <= 0)) {
1550 goto out_free_table;
1553 if (blk_integrity_rq(rq)) {
1554 req->metadata_sgl->sg_table.sgl =
1555 (struct scatterlist *)(req->metadata_sgl + 1);
1556 ret = sg_alloc_table_chained(&req->metadata_sgl->sg_table,
1557 blk_rq_count_integrity_sg(rq->q, rq->bio),
1558 req->metadata_sgl->sg_table.sgl,
1559 NVME_INLINE_METADATA_SG_CNT);
1560 if (unlikely(ret)) {
1565 req->metadata_sgl->nents = blk_rq_map_integrity_sg(rq->q,
1566 rq->bio, req->metadata_sgl->sg_table.sgl);
1567 *pi_count = ib_dma_map_sg(ibdev,
1568 req->metadata_sgl->sg_table.sgl,
1569 req->metadata_sgl->nents,
1571 if (unlikely(*pi_count <= 0)) {
1573 goto out_free_pi_table;
1580 sg_free_table_chained(&req->metadata_sgl->sg_table,
1581 NVME_INLINE_METADATA_SG_CNT);
1583 ib_dma_unmap_sg(ibdev, req->data_sgl.sg_table.sgl, req->data_sgl.nents,
1586 sg_free_table_chained(&req->data_sgl.sg_table, NVME_INLINE_SG_CNT);
1590 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1591 struct request *rq, struct nvme_command *c)
1593 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1594 struct nvme_rdma_device *dev = queue->device;
1595 struct ib_device *ibdev = dev->dev;
1600 refcount_set(&req->ref, 2); /* send and recv completions */
1602 c->common.flags |= NVME_CMD_SGL_METABUF;
1604 if (!blk_rq_nr_phys_segments(rq))
1605 return nvme_rdma_set_sg_null(c);
1607 ret = nvme_rdma_dma_map_req(ibdev, rq, &count, &pi_count);
1611 if (req->use_sig_mr) {
1612 ret = nvme_rdma_map_sg_pi(queue, req, c, count, pi_count);
1616 if (count <= dev->num_inline_segments) {
1617 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1618 queue->ctrl->use_inline_data &&
1619 blk_rq_payload_bytes(rq) <=
1620 nvme_rdma_inline_data_size(queue)) {
1621 ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1625 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1626 ret = nvme_rdma_map_sg_single(queue, req, c);
1631 ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1634 goto out_dma_unmap_req;
1639 nvme_rdma_dma_unmap_req(ibdev, rq);
1643 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1645 struct nvme_rdma_qe *qe =
1646 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1647 struct nvme_rdma_request *req =
1648 container_of(qe, struct nvme_rdma_request, sqe);
1650 if (unlikely(wc->status != IB_WC_SUCCESS))
1651 nvme_rdma_wr_error(cq, wc, "SEND");
1653 nvme_rdma_end_request(req);
1656 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1657 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1658 struct ib_send_wr *first)
1660 struct ib_send_wr wr;
1663 sge->addr = qe->dma;
1664 sge->length = sizeof(struct nvme_command);
1665 sge->lkey = queue->device->pd->local_dma_lkey;
1668 wr.wr_cqe = &qe->cqe;
1670 wr.num_sge = num_sge;
1671 wr.opcode = IB_WR_SEND;
1672 wr.send_flags = IB_SEND_SIGNALED;
1679 ret = ib_post_send(queue->qp, first, NULL);
1680 if (unlikely(ret)) {
1681 dev_err(queue->ctrl->ctrl.device,
1682 "%s failed with error code %d\n", __func__, ret);
1687 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1688 struct nvme_rdma_qe *qe)
1690 struct ib_recv_wr wr;
1694 list.addr = qe->dma;
1695 list.length = sizeof(struct nvme_completion);
1696 list.lkey = queue->device->pd->local_dma_lkey;
1698 qe->cqe.done = nvme_rdma_recv_done;
1701 wr.wr_cqe = &qe->cqe;
1705 ret = ib_post_recv(queue->qp, &wr, NULL);
1706 if (unlikely(ret)) {
1707 dev_err(queue->ctrl->ctrl.device,
1708 "%s failed with error code %d\n", __func__, ret);
1713 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1715 u32 queue_idx = nvme_rdma_queue_idx(queue);
1718 return queue->ctrl->admin_tag_set.tags[queue_idx];
1719 return queue->ctrl->tag_set.tags[queue_idx - 1];
1722 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1724 if (unlikely(wc->status != IB_WC_SUCCESS))
1725 nvme_rdma_wr_error(cq, wc, "ASYNC");
1728 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1730 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1731 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1732 struct ib_device *dev = queue->device->dev;
1733 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1734 struct nvme_command *cmd = sqe->data;
1738 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1740 memset(cmd, 0, sizeof(*cmd));
1741 cmd->common.opcode = nvme_admin_async_event;
1742 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1743 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1744 nvme_rdma_set_sg_null(cmd);
1746 sqe->cqe.done = nvme_rdma_async_done;
1748 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1751 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1755 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1756 struct nvme_completion *cqe, struct ib_wc *wc)
1759 struct nvme_rdma_request *req;
1761 rq = nvme_find_rq(nvme_rdma_tagset(queue), cqe->command_id);
1763 dev_err(queue->ctrl->ctrl.device,
1764 "got bad command_id %#x on QP %#x\n",
1765 cqe->command_id, queue->qp->qp_num);
1766 nvme_rdma_error_recovery(queue->ctrl);
1769 req = blk_mq_rq_to_pdu(rq);
1771 req->status = cqe->status;
1772 req->result = cqe->result;
1774 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1775 if (unlikely(!req->mr ||
1776 wc->ex.invalidate_rkey != req->mr->rkey)) {
1777 dev_err(queue->ctrl->ctrl.device,
1778 "Bogus remote invalidation for rkey %#x\n",
1779 req->mr ? req->mr->rkey : 0);
1780 nvme_rdma_error_recovery(queue->ctrl);
1782 } else if (req->mr) {
1785 ret = nvme_rdma_inv_rkey(queue, req);
1786 if (unlikely(ret < 0)) {
1787 dev_err(queue->ctrl->ctrl.device,
1788 "Queueing INV WR for rkey %#x failed (%d)\n",
1789 req->mr->rkey, ret);
1790 nvme_rdma_error_recovery(queue->ctrl);
1792 /* the local invalidation completion will end the request */
1796 nvme_rdma_end_request(req);
1799 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1801 struct nvme_rdma_qe *qe =
1802 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1803 struct nvme_rdma_queue *queue = wc->qp->qp_context;
1804 struct ib_device *ibdev = queue->device->dev;
1805 struct nvme_completion *cqe = qe->data;
1806 const size_t len = sizeof(struct nvme_completion);
1808 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1809 nvme_rdma_wr_error(cq, wc, "RECV");
1813 /* sanity checking for received data length */
1814 if (unlikely(wc->byte_len < len)) {
1815 dev_err(queue->ctrl->ctrl.device,
1816 "Unexpected nvme completion length(%d)\n", wc->byte_len);
1817 nvme_rdma_error_recovery(queue->ctrl);
1821 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1823 * AEN requests are special as they don't time out and can
1824 * survive any kind of queue freeze and often don't respond to
1825 * aborts. We don't even bother to allocate a struct request
1826 * for them but rather special case them here.
1828 if (unlikely(nvme_is_aen_req(nvme_rdma_queue_idx(queue),
1830 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1833 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1834 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1836 nvme_rdma_post_recv(queue, qe);
1839 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1843 for (i = 0; i < queue->queue_size; i++) {
1844 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1852 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1853 struct rdma_cm_event *ev)
1855 struct rdma_cm_id *cm_id = queue->cm_id;
1856 int status = ev->status;
1857 const char *rej_msg;
1858 const struct nvme_rdma_cm_rej *rej_data;
1861 rej_msg = rdma_reject_msg(cm_id, status);
1862 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1864 if (rej_data && rej_data_len >= sizeof(u16)) {
1865 u16 sts = le16_to_cpu(rej_data->sts);
1867 dev_err(queue->ctrl->ctrl.device,
1868 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1869 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1871 dev_err(queue->ctrl->ctrl.device,
1872 "Connect rejected: status %d (%s).\n", status, rej_msg);
1878 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1880 struct nvme_ctrl *ctrl = &queue->ctrl->ctrl;
1883 ret = nvme_rdma_create_queue_ib(queue);
1887 if (ctrl->opts->tos >= 0)
1888 rdma_set_service_type(queue->cm_id, ctrl->opts->tos);
1889 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1891 dev_err(ctrl->device, "rdma_resolve_route failed (%d).\n",
1893 goto out_destroy_queue;
1899 nvme_rdma_destroy_queue_ib(queue);
1903 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1905 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1906 struct rdma_conn_param param = { };
1907 struct nvme_rdma_cm_req priv = { };
1910 param.qp_num = queue->qp->qp_num;
1911 param.flow_control = 1;
1913 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1914 /* maximum retry count */
1915 param.retry_count = 7;
1916 param.rnr_retry_count = 7;
1917 param.private_data = &priv;
1918 param.private_data_len = sizeof(priv);
1920 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1921 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1923 * set the admin queue depth to the minimum size
1924 * specified by the Fabrics standard.
1926 if (priv.qid == 0) {
1927 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1928 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1931 * current interpretation of the fabrics spec
1932 * is at minimum you make hrqsize sqsize+1, or a
1933 * 1's based representation of sqsize.
1935 priv.hrqsize = cpu_to_le16(queue->queue_size);
1936 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1939 ret = rdma_connect_locked(queue->cm_id, ¶m);
1941 dev_err(ctrl->ctrl.device,
1942 "rdma_connect_locked failed (%d).\n", ret);
1949 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1950 struct rdma_cm_event *ev)
1952 struct nvme_rdma_queue *queue = cm_id->context;
1955 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1956 rdma_event_msg(ev->event), ev->event,
1959 switch (ev->event) {
1960 case RDMA_CM_EVENT_ADDR_RESOLVED:
1961 cm_error = nvme_rdma_addr_resolved(queue);
1963 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1964 cm_error = nvme_rdma_route_resolved(queue);
1966 case RDMA_CM_EVENT_ESTABLISHED:
1967 queue->cm_error = nvme_rdma_conn_established(queue);
1968 /* complete cm_done regardless of success/failure */
1969 complete(&queue->cm_done);
1971 case RDMA_CM_EVENT_REJECTED:
1972 cm_error = nvme_rdma_conn_rejected(queue, ev);
1974 case RDMA_CM_EVENT_ROUTE_ERROR:
1975 case RDMA_CM_EVENT_CONNECT_ERROR:
1976 case RDMA_CM_EVENT_UNREACHABLE:
1977 case RDMA_CM_EVENT_ADDR_ERROR:
1978 dev_dbg(queue->ctrl->ctrl.device,
1979 "CM error event %d\n", ev->event);
1980 cm_error = -ECONNRESET;
1982 case RDMA_CM_EVENT_DISCONNECTED:
1983 case RDMA_CM_EVENT_ADDR_CHANGE:
1984 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1985 dev_dbg(queue->ctrl->ctrl.device,
1986 "disconnect received - connection closed\n");
1987 nvme_rdma_error_recovery(queue->ctrl);
1989 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1990 /* device removal is handled via the ib_client API */
1993 dev_err(queue->ctrl->ctrl.device,
1994 "Unexpected RDMA CM event (%d)\n", ev->event);
1995 nvme_rdma_error_recovery(queue->ctrl);
2000 queue->cm_error = cm_error;
2001 complete(&queue->cm_done);
2007 static void nvme_rdma_complete_timed_out(struct request *rq)
2009 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2010 struct nvme_rdma_queue *queue = req->queue;
2012 nvme_rdma_stop_queue(queue);
2013 nvmf_complete_timed_out_request(rq);
2016 static enum blk_eh_timer_return
2017 nvme_rdma_timeout(struct request *rq, bool reserved)
2019 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2020 struct nvme_rdma_queue *queue = req->queue;
2021 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
2023 dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
2024 rq->tag, nvme_rdma_queue_idx(queue));
2026 if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
2028 * If we are resetting, connecting or deleting we should
2029 * complete immediately because we may block controller
2030 * teardown or setup sequence
2031 * - ctrl disable/shutdown fabrics requests
2032 * - connect requests
2033 * - initialization admin requests
2034 * - I/O requests that entered after unquiescing and
2035 * the controller stopped responding
2037 * All other requests should be cancelled by the error
2038 * recovery work, so it's fine that we fail it here.
2040 nvme_rdma_complete_timed_out(rq);
2045 * LIVE state should trigger the normal error recovery which will
2046 * handle completing this request.
2048 nvme_rdma_error_recovery(ctrl);
2049 return BLK_EH_RESET_TIMER;
2052 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
2053 const struct blk_mq_queue_data *bd)
2055 struct nvme_ns *ns = hctx->queue->queuedata;
2056 struct nvme_rdma_queue *queue = hctx->driver_data;
2057 struct request *rq = bd->rq;
2058 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2059 struct nvme_rdma_qe *sqe = &req->sqe;
2060 struct nvme_command *c = nvme_req(rq)->cmd;
2061 struct ib_device *dev;
2062 bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
2066 WARN_ON_ONCE(rq->tag < 0);
2068 if (!nvme_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
2069 return nvme_fail_nonready_command(&queue->ctrl->ctrl, rq);
2071 dev = queue->device->dev;
2073 req->sqe.dma = ib_dma_map_single(dev, req->sqe.data,
2074 sizeof(struct nvme_command),
2076 err = ib_dma_mapping_error(dev, req->sqe.dma);
2078 return BLK_STS_RESOURCE;
2080 ib_dma_sync_single_for_cpu(dev, sqe->dma,
2081 sizeof(struct nvme_command), DMA_TO_DEVICE);
2083 ret = nvme_setup_cmd(ns, rq);
2087 blk_mq_start_request(rq);
2089 if (IS_ENABLED(CONFIG_BLK_DEV_INTEGRITY) &&
2090 queue->pi_support &&
2091 (c->common.opcode == nvme_cmd_write ||
2092 c->common.opcode == nvme_cmd_read) &&
2094 req->use_sig_mr = true;
2096 req->use_sig_mr = false;
2098 err = nvme_rdma_map_data(queue, rq, c);
2099 if (unlikely(err < 0)) {
2100 dev_err(queue->ctrl->ctrl.device,
2101 "Failed to map data (%d)\n", err);
2105 sqe->cqe.done = nvme_rdma_send_done;
2107 ib_dma_sync_single_for_device(dev, sqe->dma,
2108 sizeof(struct nvme_command), DMA_TO_DEVICE);
2110 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
2111 req->mr ? &req->reg_wr.wr : NULL);
2118 nvme_rdma_unmap_data(queue, rq);
2121 ret = nvme_host_path_error(rq);
2122 else if (err == -ENOMEM || err == -EAGAIN)
2123 ret = BLK_STS_RESOURCE;
2125 ret = BLK_STS_IOERR;
2126 nvme_cleanup_cmd(rq);
2128 ib_dma_unmap_single(dev, req->sqe.dma, sizeof(struct nvme_command),
2133 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, struct io_comp_batch *iob)
2135 struct nvme_rdma_queue *queue = hctx->driver_data;
2137 return ib_process_cq_direct(queue->ib_cq, -1);
2140 static void nvme_rdma_check_pi_status(struct nvme_rdma_request *req)
2142 struct request *rq = blk_mq_rq_from_pdu(req);
2143 struct ib_mr_status mr_status;
2146 ret = ib_check_mr_status(req->mr, IB_MR_CHECK_SIG_STATUS, &mr_status);
2148 pr_err("ib_check_mr_status failed, ret %d\n", ret);
2149 nvme_req(rq)->status = NVME_SC_INVALID_PI;
2153 if (mr_status.fail_status & IB_MR_CHECK_SIG_STATUS) {
2154 switch (mr_status.sig_err.err_type) {
2155 case IB_SIG_BAD_GUARD:
2156 nvme_req(rq)->status = NVME_SC_GUARD_CHECK;
2158 case IB_SIG_BAD_REFTAG:
2159 nvme_req(rq)->status = NVME_SC_REFTAG_CHECK;
2161 case IB_SIG_BAD_APPTAG:
2162 nvme_req(rq)->status = NVME_SC_APPTAG_CHECK;
2165 pr_err("PI error found type %d expected 0x%x vs actual 0x%x\n",
2166 mr_status.sig_err.err_type, mr_status.sig_err.expected,
2167 mr_status.sig_err.actual);
2171 static void nvme_rdma_complete_rq(struct request *rq)
2173 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
2174 struct nvme_rdma_queue *queue = req->queue;
2175 struct ib_device *ibdev = queue->device->dev;
2177 if (req->use_sig_mr)
2178 nvme_rdma_check_pi_status(req);
2180 nvme_rdma_unmap_data(queue, rq);
2181 ib_dma_unmap_single(ibdev, req->sqe.dma, sizeof(struct nvme_command),
2183 nvme_complete_rq(rq);
2186 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
2188 struct nvme_rdma_ctrl *ctrl = set->driver_data;
2189 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
2191 if (opts->nr_write_queues && ctrl->io_queues[HCTX_TYPE_READ]) {
2192 /* separate read/write queues */
2193 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2194 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2195 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2196 set->map[HCTX_TYPE_READ].nr_queues =
2197 ctrl->io_queues[HCTX_TYPE_READ];
2198 set->map[HCTX_TYPE_READ].queue_offset =
2199 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2201 /* shared read/write queues */
2202 set->map[HCTX_TYPE_DEFAULT].nr_queues =
2203 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2204 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
2205 set->map[HCTX_TYPE_READ].nr_queues =
2206 ctrl->io_queues[HCTX_TYPE_DEFAULT];
2207 set->map[HCTX_TYPE_READ].queue_offset = 0;
2209 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
2210 ctrl->device->dev, 0);
2211 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
2212 ctrl->device->dev, 0);
2214 if (opts->nr_poll_queues && ctrl->io_queues[HCTX_TYPE_POLL]) {
2215 /* map dedicated poll queues only if we have queues left */
2216 set->map[HCTX_TYPE_POLL].nr_queues =
2217 ctrl->io_queues[HCTX_TYPE_POLL];
2218 set->map[HCTX_TYPE_POLL].queue_offset =
2219 ctrl->io_queues[HCTX_TYPE_DEFAULT] +
2220 ctrl->io_queues[HCTX_TYPE_READ];
2221 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
2224 dev_info(ctrl->ctrl.device,
2225 "mapped %d/%d/%d default/read/poll queues.\n",
2226 ctrl->io_queues[HCTX_TYPE_DEFAULT],
2227 ctrl->io_queues[HCTX_TYPE_READ],
2228 ctrl->io_queues[HCTX_TYPE_POLL]);
2233 static const struct blk_mq_ops nvme_rdma_mq_ops = {
2234 .queue_rq = nvme_rdma_queue_rq,
2235 .complete = nvme_rdma_complete_rq,
2236 .init_request = nvme_rdma_init_request,
2237 .exit_request = nvme_rdma_exit_request,
2238 .init_hctx = nvme_rdma_init_hctx,
2239 .timeout = nvme_rdma_timeout,
2240 .map_queues = nvme_rdma_map_queues,
2241 .poll = nvme_rdma_poll,
2244 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
2245 .queue_rq = nvme_rdma_queue_rq,
2246 .complete = nvme_rdma_complete_rq,
2247 .init_request = nvme_rdma_init_request,
2248 .exit_request = nvme_rdma_exit_request,
2249 .init_hctx = nvme_rdma_init_admin_hctx,
2250 .timeout = nvme_rdma_timeout,
2253 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
2255 cancel_work_sync(&ctrl->err_work);
2256 cancel_delayed_work_sync(&ctrl->reconnect_work);
2258 nvme_rdma_teardown_io_queues(ctrl, shutdown);
2259 nvme_stop_admin_queue(&ctrl->ctrl);
2261 nvme_shutdown_ctrl(&ctrl->ctrl);
2263 nvme_disable_ctrl(&ctrl->ctrl);
2264 nvme_rdma_teardown_admin_queue(ctrl, shutdown);
2267 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
2269 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
2272 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
2274 struct nvme_rdma_ctrl *ctrl =
2275 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
2277 nvme_stop_ctrl(&ctrl->ctrl);
2278 nvme_rdma_shutdown_ctrl(ctrl, false);
2280 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
2281 /* state change failure should never happen */
2286 if (nvme_rdma_setup_ctrl(ctrl, false))
2292 ++ctrl->ctrl.nr_reconnects;
2293 nvme_rdma_reconnect_or_remove(ctrl);
2296 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
2298 .module = THIS_MODULE,
2299 .flags = NVME_F_FABRICS | NVME_F_METADATA_SUPPORTED,
2300 .reg_read32 = nvmf_reg_read32,
2301 .reg_read64 = nvmf_reg_read64,
2302 .reg_write32 = nvmf_reg_write32,
2303 .free_ctrl = nvme_rdma_free_ctrl,
2304 .submit_async_event = nvme_rdma_submit_async_event,
2305 .delete_ctrl = nvme_rdma_delete_ctrl,
2306 .get_address = nvmf_get_address,
2310 * Fails a connection request if it matches an existing controller
2311 * (association) with the same tuple:
2312 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
2314 * if local address is not specified in the request, it will match an
2315 * existing controller with all the other parameters the same and no
2316 * local port address specified as well.
2318 * The ports don't need to be compared as they are intrinsically
2319 * already matched by the port pointers supplied.
2322 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
2324 struct nvme_rdma_ctrl *ctrl;
2327 mutex_lock(&nvme_rdma_ctrl_mutex);
2328 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2329 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
2333 mutex_unlock(&nvme_rdma_ctrl_mutex);
2338 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
2339 struct nvmf_ctrl_options *opts)
2341 struct nvme_rdma_ctrl *ctrl;
2345 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
2347 return ERR_PTR(-ENOMEM);
2348 ctrl->ctrl.opts = opts;
2349 INIT_LIST_HEAD(&ctrl->list);
2351 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
2353 kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
2354 if (!opts->trsvcid) {
2358 opts->mask |= NVMF_OPT_TRSVCID;
2361 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2362 opts->traddr, opts->trsvcid, &ctrl->addr);
2364 pr_err("malformed address passed: %s:%s\n",
2365 opts->traddr, opts->trsvcid);
2369 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
2370 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
2371 opts->host_traddr, NULL, &ctrl->src_addr);
2373 pr_err("malformed src address passed: %s\n",
2379 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
2384 INIT_DELAYED_WORK(&ctrl->reconnect_work,
2385 nvme_rdma_reconnect_ctrl_work);
2386 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
2387 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
2389 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
2390 opts->nr_poll_queues + 1;
2391 ctrl->ctrl.sqsize = opts->queue_size - 1;
2392 ctrl->ctrl.kato = opts->kato;
2395 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
2400 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
2401 0 /* no quirks, we're perfect! */);
2403 goto out_kfree_queues;
2405 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
2406 WARN_ON_ONCE(!changed);
2408 ret = nvme_rdma_setup_ctrl(ctrl, true);
2410 goto out_uninit_ctrl;
2412 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
2413 nvmf_ctrl_subsysnqn(&ctrl->ctrl), &ctrl->addr);
2415 mutex_lock(&nvme_rdma_ctrl_mutex);
2416 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2417 mutex_unlock(&nvme_rdma_ctrl_mutex);
2422 nvme_uninit_ctrl(&ctrl->ctrl);
2423 nvme_put_ctrl(&ctrl->ctrl);
2426 return ERR_PTR(ret);
2428 kfree(ctrl->queues);
2431 return ERR_PTR(ret);
2434 static struct nvmf_transport_ops nvme_rdma_transport = {
2436 .module = THIS_MODULE,
2437 .required_opts = NVMF_OPT_TRADDR,
2438 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2439 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2440 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES |
2442 .create_ctrl = nvme_rdma_create_ctrl,
2445 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2447 struct nvme_rdma_ctrl *ctrl;
2448 struct nvme_rdma_device *ndev;
2451 mutex_lock(&device_list_mutex);
2452 list_for_each_entry(ndev, &device_list, entry) {
2453 if (ndev->dev == ib_device) {
2458 mutex_unlock(&device_list_mutex);
2463 /* Delete all controllers using this device */
2464 mutex_lock(&nvme_rdma_ctrl_mutex);
2465 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2466 if (ctrl->device->dev != ib_device)
2468 nvme_delete_ctrl(&ctrl->ctrl);
2470 mutex_unlock(&nvme_rdma_ctrl_mutex);
2472 flush_workqueue(nvme_delete_wq);
2475 static struct ib_client nvme_rdma_ib_client = {
2476 .name = "nvme_rdma",
2477 .remove = nvme_rdma_remove_one
2480 static int __init nvme_rdma_init_module(void)
2484 ret = ib_register_client(&nvme_rdma_ib_client);
2488 ret = nvmf_register_transport(&nvme_rdma_transport);
2490 goto err_unreg_client;
2495 ib_unregister_client(&nvme_rdma_ib_client);
2499 static void __exit nvme_rdma_cleanup_module(void)
2501 struct nvme_rdma_ctrl *ctrl;
2503 nvmf_unregister_transport(&nvme_rdma_transport);
2504 ib_unregister_client(&nvme_rdma_ib_client);
2506 mutex_lock(&nvme_rdma_ctrl_mutex);
2507 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list)
2508 nvme_delete_ctrl(&ctrl->ctrl);
2509 mutex_unlock(&nvme_rdma_ctrl_mutex);
2510 flush_workqueue(nvme_delete_wq);
2513 module_init(nvme_rdma_init_module);
2514 module_exit(nvme_rdma_cleanup_module);
2516 MODULE_LICENSE("GPL v2");