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 struct nvme_rdma_device {
38 struct ib_device *dev;
41 struct list_head entry;
42 unsigned int num_inline_segments;
51 struct nvme_rdma_queue;
52 struct nvme_rdma_request {
53 struct nvme_request req;
55 struct nvme_rdma_qe sqe;
56 union nvme_result result;
59 struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
62 struct ib_reg_wr reg_wr;
63 struct ib_cqe reg_cqe;
64 struct nvme_rdma_queue *queue;
65 struct sg_table sg_table;
66 struct scatterlist first_sgl[];
69 enum nvme_rdma_queue_flags {
70 NVME_RDMA_Q_ALLOCATED = 0,
72 NVME_RDMA_Q_TR_READY = 2,
75 struct nvme_rdma_queue {
76 struct nvme_rdma_qe *rsp_ring;
78 size_t cmnd_capsule_len;
79 struct nvme_rdma_ctrl *ctrl;
80 struct nvme_rdma_device *device;
85 struct rdma_cm_id *cm_id;
87 struct completion cm_done;
90 struct nvme_rdma_ctrl {
91 /* read only in the hot path */
92 struct nvme_rdma_queue *queues;
94 /* other member variables */
95 struct blk_mq_tag_set tag_set;
96 struct work_struct err_work;
98 struct nvme_rdma_qe async_event_sqe;
100 struct delayed_work reconnect_work;
102 struct list_head list;
104 struct blk_mq_tag_set admin_tag_set;
105 struct nvme_rdma_device *device;
109 struct sockaddr_storage addr;
110 struct sockaddr_storage src_addr;
112 struct nvme_ctrl ctrl;
113 bool use_inline_data;
114 u32 io_queues[HCTX_MAX_TYPES];
117 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
119 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
122 static LIST_HEAD(device_list);
123 static DEFINE_MUTEX(device_list_mutex);
125 static LIST_HEAD(nvme_rdma_ctrl_list);
126 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
129 * Disabling this option makes small I/O goes faster, but is fundamentally
130 * unsafe. With it turned off we will have to register a global rkey that
131 * allows read and write access to all physical memory.
133 static bool register_always = true;
134 module_param(register_always, bool, 0444);
135 MODULE_PARM_DESC(register_always,
136 "Use memory registration even for contiguous memory regions");
138 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
139 struct rdma_cm_event *event);
140 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
142 static const struct blk_mq_ops nvme_rdma_mq_ops;
143 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
145 /* XXX: really should move to a generic header sooner or later.. */
146 static inline void put_unaligned_le24(u32 val, u8 *p)
153 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
155 return queue - queue->ctrl->queues;
158 static bool nvme_rdma_poll_queue(struct nvme_rdma_queue *queue)
160 return nvme_rdma_queue_idx(queue) >
161 queue->ctrl->io_queues[HCTX_TYPE_DEFAULT] +
162 queue->ctrl->io_queues[HCTX_TYPE_READ];
165 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
167 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
170 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
171 size_t capsule_size, enum dma_data_direction dir)
173 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
177 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
178 size_t capsule_size, enum dma_data_direction dir)
180 qe->data = kzalloc(capsule_size, GFP_KERNEL);
184 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
185 if (ib_dma_mapping_error(ibdev, qe->dma)) {
194 static void nvme_rdma_free_ring(struct ib_device *ibdev,
195 struct nvme_rdma_qe *ring, size_t ib_queue_size,
196 size_t capsule_size, enum dma_data_direction dir)
200 for (i = 0; i < ib_queue_size; i++)
201 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
205 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
206 size_t ib_queue_size, size_t capsule_size,
207 enum dma_data_direction dir)
209 struct nvme_rdma_qe *ring;
212 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
216 for (i = 0; i < ib_queue_size; i++) {
217 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
224 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
228 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
230 pr_debug("QP event %s (%d)\n",
231 ib_event_msg(event->event), event->event);
235 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
239 ret = wait_for_completion_interruptible_timeout(&queue->cm_done,
240 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
245 WARN_ON_ONCE(queue->cm_error > 0);
246 return queue->cm_error;
249 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
251 struct nvme_rdma_device *dev = queue->device;
252 struct ib_qp_init_attr init_attr;
255 memset(&init_attr, 0, sizeof(init_attr));
256 init_attr.event_handler = nvme_rdma_qp_event;
258 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
260 init_attr.cap.max_recv_wr = queue->queue_size + 1;
261 init_attr.cap.max_recv_sge = 1;
262 init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
263 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
264 init_attr.qp_type = IB_QPT_RC;
265 init_attr.send_cq = queue->ib_cq;
266 init_attr.recv_cq = queue->ib_cq;
268 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
270 queue->qp = queue->cm_id->qp;
274 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
275 struct request *rq, unsigned int hctx_idx)
277 struct nvme_rdma_ctrl *ctrl = set->driver_data;
278 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
279 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
280 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
281 struct nvme_rdma_device *dev = queue->device;
283 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
287 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
288 struct request *rq, unsigned int hctx_idx,
289 unsigned int numa_node)
291 struct nvme_rdma_ctrl *ctrl = set->driver_data;
292 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
293 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
294 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
295 struct nvme_rdma_device *dev = queue->device;
296 struct ib_device *ibdev = dev->dev;
299 nvme_req(rq)->ctrl = &ctrl->ctrl;
300 ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
310 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
311 unsigned int hctx_idx)
313 struct nvme_rdma_ctrl *ctrl = data;
314 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
316 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
318 hctx->driver_data = queue;
322 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
323 unsigned int hctx_idx)
325 struct nvme_rdma_ctrl *ctrl = data;
326 struct nvme_rdma_queue *queue = &ctrl->queues[0];
328 BUG_ON(hctx_idx != 0);
330 hctx->driver_data = queue;
334 static void nvme_rdma_free_dev(struct kref *ref)
336 struct nvme_rdma_device *ndev =
337 container_of(ref, struct nvme_rdma_device, ref);
339 mutex_lock(&device_list_mutex);
340 list_del(&ndev->entry);
341 mutex_unlock(&device_list_mutex);
343 ib_dealloc_pd(ndev->pd);
347 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
349 kref_put(&dev->ref, nvme_rdma_free_dev);
352 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
354 return kref_get_unless_zero(&dev->ref);
357 static struct nvme_rdma_device *
358 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
360 struct nvme_rdma_device *ndev;
362 mutex_lock(&device_list_mutex);
363 list_for_each_entry(ndev, &device_list, entry) {
364 if (ndev->dev->node_guid == cm_id->device->node_guid &&
365 nvme_rdma_dev_get(ndev))
369 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
373 ndev->dev = cm_id->device;
374 kref_init(&ndev->ref);
376 ndev->pd = ib_alloc_pd(ndev->dev,
377 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
378 if (IS_ERR(ndev->pd))
381 if (!(ndev->dev->attrs.device_cap_flags &
382 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
383 dev_err(&ndev->dev->dev,
384 "Memory registrations not supported.\n");
388 ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
389 ndev->dev->attrs.max_send_sge - 1);
390 list_add(&ndev->entry, &device_list);
392 mutex_unlock(&device_list_mutex);
396 ib_dealloc_pd(ndev->pd);
400 mutex_unlock(&device_list_mutex);
404 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
406 struct nvme_rdma_device *dev;
407 struct ib_device *ibdev;
409 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
415 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
418 * The cm_id object might have been destroyed during RDMA connection
419 * establishment error flow to avoid getting other cma events, thus
420 * the destruction of the QP shouldn't use rdma_cm API.
422 ib_destroy_qp(queue->qp);
423 ib_free_cq(queue->ib_cq);
425 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
426 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
428 nvme_rdma_dev_put(dev);
431 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev)
433 return min_t(u32, NVME_RDMA_MAX_SEGMENTS,
434 ibdev->attrs.max_fast_reg_page_list_len);
437 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
439 struct ib_device *ibdev;
440 const int send_wr_factor = 3; /* MR, SEND, INV */
441 const int cq_factor = send_wr_factor + 1; /* + RECV */
442 int comp_vector, idx = nvme_rdma_queue_idx(queue);
443 enum ib_poll_context poll_ctx;
446 queue->device = nvme_rdma_find_get_device(queue->cm_id);
447 if (!queue->device) {
448 dev_err(queue->cm_id->device->dev.parent,
449 "no client data found!\n");
450 return -ECONNREFUSED;
452 ibdev = queue->device->dev;
455 * Spread I/O queues completion vectors according their queue index.
456 * Admin queues can always go on completion vector 0.
458 comp_vector = idx == 0 ? idx : idx - 1;
460 /* Polling queues need direct cq polling context */
461 if (nvme_rdma_poll_queue(queue))
462 poll_ctx = IB_POLL_DIRECT;
464 poll_ctx = IB_POLL_SOFTIRQ;
466 /* +1 for ib_stop_cq */
467 queue->ib_cq = ib_alloc_cq(ibdev, queue,
468 cq_factor * queue->queue_size + 1,
469 comp_vector, poll_ctx);
470 if (IS_ERR(queue->ib_cq)) {
471 ret = PTR_ERR(queue->ib_cq);
475 ret = nvme_rdma_create_qp(queue, send_wr_factor);
477 goto out_destroy_ib_cq;
479 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
480 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
481 if (!queue->rsp_ring) {
486 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
489 nvme_rdma_get_max_fr_pages(ibdev));
491 dev_err(queue->ctrl->ctrl.device,
492 "failed to initialize MR pool sized %d for QID %d\n",
493 queue->queue_size, idx);
494 goto out_destroy_ring;
497 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
502 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
503 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
505 rdma_destroy_qp(queue->cm_id);
507 ib_free_cq(queue->ib_cq);
509 nvme_rdma_dev_put(queue->device);
513 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
514 int idx, size_t queue_size)
516 struct nvme_rdma_queue *queue;
517 struct sockaddr *src_addr = NULL;
520 queue = &ctrl->queues[idx];
522 init_completion(&queue->cm_done);
525 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
527 queue->cmnd_capsule_len = sizeof(struct nvme_command);
529 queue->queue_size = queue_size;
531 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
532 RDMA_PS_TCP, IB_QPT_RC);
533 if (IS_ERR(queue->cm_id)) {
534 dev_info(ctrl->ctrl.device,
535 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
536 return PTR_ERR(queue->cm_id);
539 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
540 src_addr = (struct sockaddr *)&ctrl->src_addr;
542 queue->cm_error = -ETIMEDOUT;
543 ret = rdma_resolve_addr(queue->cm_id, src_addr,
544 (struct sockaddr *)&ctrl->addr,
545 NVME_RDMA_CONNECT_TIMEOUT_MS);
547 dev_info(ctrl->ctrl.device,
548 "rdma_resolve_addr failed (%d).\n", ret);
549 goto out_destroy_cm_id;
552 ret = nvme_rdma_wait_for_cm(queue);
554 dev_info(ctrl->ctrl.device,
555 "rdma connection establishment failed (%d)\n", ret);
556 goto out_destroy_cm_id;
559 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
564 rdma_destroy_id(queue->cm_id);
565 nvme_rdma_destroy_queue_ib(queue);
569 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
571 if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
574 rdma_disconnect(queue->cm_id);
575 ib_drain_qp(queue->qp);
578 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
580 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
583 nvme_rdma_destroy_queue_ib(queue);
584 rdma_destroy_id(queue->cm_id);
587 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
591 for (i = 1; i < ctrl->ctrl.queue_count; i++)
592 nvme_rdma_free_queue(&ctrl->queues[i]);
595 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
599 for (i = 1; i < ctrl->ctrl.queue_count; i++)
600 nvme_rdma_stop_queue(&ctrl->queues[i]);
603 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
605 struct nvme_rdma_queue *queue = &ctrl->queues[idx];
606 bool poll = nvme_rdma_poll_queue(queue);
610 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx, poll);
612 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
615 set_bit(NVME_RDMA_Q_LIVE, &queue->flags);
617 dev_info(ctrl->ctrl.device,
618 "failed to connect queue: %d ret=%d\n", idx, ret);
622 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
626 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
627 ret = nvme_rdma_start_queue(ctrl, i);
629 goto out_stop_queues;
635 for (i--; i >= 1; i--)
636 nvme_rdma_stop_queue(&ctrl->queues[i]);
640 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
642 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
643 struct ib_device *ibdev = ctrl->device->dev;
644 unsigned int nr_io_queues;
647 nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
650 * we map queues according to the device irq vectors for
651 * optimal locality so we don't need more queues than
652 * completion vectors.
654 nr_io_queues = min_t(unsigned int, nr_io_queues,
655 ibdev->num_comp_vectors);
657 if (opts->nr_write_queues) {
658 ctrl->io_queues[HCTX_TYPE_DEFAULT] =
659 min(opts->nr_write_queues, nr_io_queues);
660 nr_io_queues += ctrl->io_queues[HCTX_TYPE_DEFAULT];
662 ctrl->io_queues[HCTX_TYPE_DEFAULT] = nr_io_queues;
665 ctrl->io_queues[HCTX_TYPE_READ] = nr_io_queues;
667 if (opts->nr_poll_queues) {
668 ctrl->io_queues[HCTX_TYPE_POLL] =
669 min(opts->nr_poll_queues, num_online_cpus());
670 nr_io_queues += ctrl->io_queues[HCTX_TYPE_POLL];
673 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
677 ctrl->ctrl.queue_count = nr_io_queues + 1;
678 if (ctrl->ctrl.queue_count < 2)
681 dev_info(ctrl->ctrl.device,
682 "creating %d I/O queues.\n", nr_io_queues);
684 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
685 ret = nvme_rdma_alloc_queue(ctrl, i,
686 ctrl->ctrl.sqsize + 1);
688 goto out_free_queues;
694 for (i--; i >= 1; i--)
695 nvme_rdma_free_queue(&ctrl->queues[i]);
700 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl,
701 struct blk_mq_tag_set *set)
703 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
705 blk_mq_free_tag_set(set);
706 nvme_rdma_dev_put(ctrl->device);
709 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
712 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
713 struct blk_mq_tag_set *set;
717 set = &ctrl->admin_tag_set;
718 memset(set, 0, sizeof(*set));
719 set->ops = &nvme_rdma_admin_mq_ops;
720 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
721 set->reserved_tags = 2; /* connect + keep-alive */
722 set->numa_node = nctrl->numa_node;
723 set->cmd_size = sizeof(struct nvme_rdma_request) +
724 SG_CHUNK_SIZE * sizeof(struct scatterlist);
725 set->driver_data = ctrl;
726 set->nr_hw_queues = 1;
727 set->timeout = ADMIN_TIMEOUT;
728 set->flags = BLK_MQ_F_NO_SCHED;
730 set = &ctrl->tag_set;
731 memset(set, 0, sizeof(*set));
732 set->ops = &nvme_rdma_mq_ops;
733 set->queue_depth = nctrl->sqsize + 1;
734 set->reserved_tags = 1; /* fabric connect */
735 set->numa_node = nctrl->numa_node;
736 set->flags = BLK_MQ_F_SHOULD_MERGE;
737 set->cmd_size = sizeof(struct nvme_rdma_request) +
738 SG_CHUNK_SIZE * sizeof(struct scatterlist);
739 set->driver_data = ctrl;
740 set->nr_hw_queues = nctrl->queue_count - 1;
741 set->timeout = NVME_IO_TIMEOUT;
742 set->nr_maps = nctrl->opts->nr_poll_queues ? HCTX_MAX_TYPES : 2;
745 ret = blk_mq_alloc_tag_set(set);
750 * We need a reference on the device as long as the tag_set is alive,
751 * as the MRs in the request structures need a valid ib_device.
753 ret = nvme_rdma_dev_get(ctrl->device);
756 goto out_free_tagset;
762 blk_mq_free_tag_set(set);
767 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
771 blk_cleanup_queue(ctrl->ctrl.admin_q);
772 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
774 if (ctrl->async_event_sqe.data) {
775 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
776 sizeof(struct nvme_command), DMA_TO_DEVICE);
777 ctrl->async_event_sqe.data = NULL;
779 nvme_rdma_free_queue(&ctrl->queues[0]);
782 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
787 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
791 ctrl->device = ctrl->queues[0].device;
792 ctrl->ctrl.numa_node = dev_to_node(ctrl->device->dev->dma_device);
794 ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
796 error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
797 sizeof(struct nvme_command), DMA_TO_DEVICE);
802 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
803 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
804 error = PTR_ERR(ctrl->ctrl.admin_tagset);
805 goto out_free_async_qe;
808 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
809 if (IS_ERR(ctrl->ctrl.admin_q)) {
810 error = PTR_ERR(ctrl->ctrl.admin_q);
811 goto out_free_tagset;
815 error = nvme_rdma_start_queue(ctrl, 0);
817 goto out_cleanup_queue;
819 error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
822 dev_err(ctrl->ctrl.device,
823 "prop_get NVME_REG_CAP failed\n");
828 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
830 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
834 ctrl->ctrl.max_hw_sectors =
835 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
837 error = nvme_init_identify(&ctrl->ctrl);
844 nvme_rdma_stop_queue(&ctrl->queues[0]);
847 blk_cleanup_queue(ctrl->ctrl.admin_q);
850 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
852 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
853 sizeof(struct nvme_command), DMA_TO_DEVICE);
854 ctrl->async_event_sqe.data = NULL;
856 nvme_rdma_free_queue(&ctrl->queues[0]);
860 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
864 blk_cleanup_queue(ctrl->ctrl.connect_q);
865 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
867 nvme_rdma_free_io_queues(ctrl);
870 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
874 ret = nvme_rdma_alloc_io_queues(ctrl);
879 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
880 if (IS_ERR(ctrl->ctrl.tagset)) {
881 ret = PTR_ERR(ctrl->ctrl.tagset);
882 goto out_free_io_queues;
885 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
886 if (IS_ERR(ctrl->ctrl.connect_q)) {
887 ret = PTR_ERR(ctrl->ctrl.connect_q);
888 goto out_free_tag_set;
891 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
892 ctrl->ctrl.queue_count - 1);
895 ret = nvme_rdma_start_io_queues(ctrl);
897 goto out_cleanup_connect_q;
901 out_cleanup_connect_q:
903 blk_cleanup_queue(ctrl->ctrl.connect_q);
906 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
908 nvme_rdma_free_io_queues(ctrl);
912 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
915 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
916 nvme_rdma_stop_queue(&ctrl->queues[0]);
917 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set, nvme_cancel_request,
919 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
920 nvme_rdma_destroy_admin_queue(ctrl, remove);
923 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
926 if (ctrl->ctrl.queue_count > 1) {
927 nvme_stop_queues(&ctrl->ctrl);
928 nvme_rdma_stop_io_queues(ctrl);
929 blk_mq_tagset_busy_iter(&ctrl->tag_set, nvme_cancel_request,
932 nvme_start_queues(&ctrl->ctrl);
933 nvme_rdma_destroy_io_queues(ctrl, remove);
937 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
939 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
941 if (list_empty(&ctrl->list))
944 mutex_lock(&nvme_rdma_ctrl_mutex);
945 list_del(&ctrl->list);
946 mutex_unlock(&nvme_rdma_ctrl_mutex);
948 nvmf_free_options(nctrl->opts);
954 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
956 /* If we are resetting/deleting then do nothing */
957 if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
958 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
959 ctrl->ctrl.state == NVME_CTRL_LIVE);
963 if (nvmf_should_reconnect(&ctrl->ctrl)) {
964 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
965 ctrl->ctrl.opts->reconnect_delay);
966 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
967 ctrl->ctrl.opts->reconnect_delay * HZ);
969 nvme_delete_ctrl(&ctrl->ctrl);
973 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
978 ret = nvme_rdma_configure_admin_queue(ctrl, new);
982 if (ctrl->ctrl.icdoff) {
983 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
987 if (!(ctrl->ctrl.sgls & (1 << 2))) {
988 dev_err(ctrl->ctrl.device,
989 "Mandatory keyed sgls are not supported!\n");
993 if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
994 dev_warn(ctrl->ctrl.device,
995 "queue_size %zu > ctrl sqsize %u, clamping down\n",
996 ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
999 if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
1000 dev_warn(ctrl->ctrl.device,
1001 "sqsize %u > ctrl maxcmd %u, clamping down\n",
1002 ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
1003 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
1006 if (ctrl->ctrl.sgls & (1 << 20))
1007 ctrl->use_inline_data = true;
1009 if (ctrl->ctrl.queue_count > 1) {
1010 ret = nvme_rdma_configure_io_queues(ctrl, new);
1015 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1017 /* state change failure is ok if we're in DELETING state */
1018 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1023 nvme_start_ctrl(&ctrl->ctrl);
1027 if (ctrl->ctrl.queue_count > 1)
1028 nvme_rdma_destroy_io_queues(ctrl, new);
1030 nvme_rdma_stop_queue(&ctrl->queues[0]);
1031 nvme_rdma_destroy_admin_queue(ctrl, new);
1035 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1037 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1038 struct nvme_rdma_ctrl, reconnect_work);
1040 ++ctrl->ctrl.nr_reconnects;
1042 if (nvme_rdma_setup_ctrl(ctrl, false))
1045 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1046 ctrl->ctrl.nr_reconnects);
1048 ctrl->ctrl.nr_reconnects = 0;
1053 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1054 ctrl->ctrl.nr_reconnects);
1055 nvme_rdma_reconnect_or_remove(ctrl);
1058 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1060 struct nvme_rdma_ctrl *ctrl = container_of(work,
1061 struct nvme_rdma_ctrl, err_work);
1063 nvme_stop_keep_alive(&ctrl->ctrl);
1064 nvme_rdma_teardown_io_queues(ctrl, false);
1065 nvme_start_queues(&ctrl->ctrl);
1066 nvme_rdma_teardown_admin_queue(ctrl, false);
1068 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1069 /* state change failure is ok if we're in DELETING state */
1070 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1074 nvme_rdma_reconnect_or_remove(ctrl);
1077 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1079 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1082 queue_work(nvme_wq, &ctrl->err_work);
1085 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1088 struct nvme_rdma_queue *queue = cq->cq_context;
1089 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1091 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1092 dev_info(ctrl->ctrl.device,
1093 "%s for CQE 0x%p failed with status %s (%d)\n",
1095 ib_wc_status_msg(wc->status), wc->status);
1096 nvme_rdma_error_recovery(ctrl);
1099 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1101 if (unlikely(wc->status != IB_WC_SUCCESS))
1102 nvme_rdma_wr_error(cq, wc, "MEMREG");
1105 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1107 struct nvme_rdma_request *req =
1108 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1109 struct request *rq = blk_mq_rq_from_pdu(req);
1111 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1112 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1116 if (refcount_dec_and_test(&req->ref))
1117 nvme_end_request(rq, req->status, req->result);
1121 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1122 struct nvme_rdma_request *req)
1124 struct ib_send_wr wr = {
1125 .opcode = IB_WR_LOCAL_INV,
1128 .send_flags = IB_SEND_SIGNALED,
1129 .ex.invalidate_rkey = req->mr->rkey,
1132 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1133 wr.wr_cqe = &req->reg_cqe;
1135 return ib_post_send(queue->qp, &wr, NULL);
1138 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1141 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1142 struct nvme_rdma_device *dev = queue->device;
1143 struct ib_device *ibdev = dev->dev;
1145 if (!blk_rq_nr_phys_segments(rq))
1149 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1153 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1154 req->nents, rq_data_dir(rq) ==
1155 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1157 nvme_cleanup_cmd(rq);
1158 sg_free_table_chained(&req->sg_table, true);
1161 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1163 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1166 put_unaligned_le24(0, sg->length);
1167 put_unaligned_le32(0, sg->key);
1168 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1172 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1173 struct nvme_rdma_request *req, struct nvme_command *c,
1176 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1177 struct scatterlist *sgl = req->sg_table.sgl;
1178 struct ib_sge *sge = &req->sge[1];
1182 for (i = 0; i < count; i++, sgl++, sge++) {
1183 sge->addr = sg_dma_address(sgl);
1184 sge->length = sg_dma_len(sgl);
1185 sge->lkey = queue->device->pd->local_dma_lkey;
1189 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1190 sg->length = cpu_to_le32(len);
1191 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1193 req->num_sge += count;
1197 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1198 struct nvme_rdma_request *req, struct nvme_command *c)
1200 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1202 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1203 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1204 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1205 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1209 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1210 struct nvme_rdma_request *req, struct nvme_command *c,
1213 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1216 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1217 if (WARN_ON_ONCE(!req->mr))
1221 * Align the MR to a 4K page size to match the ctrl page size and
1222 * the block virtual boundary.
1224 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1225 if (unlikely(nr < count)) {
1226 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1233 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1235 req->reg_cqe.done = nvme_rdma_memreg_done;
1236 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1237 req->reg_wr.wr.opcode = IB_WR_REG_MR;
1238 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1239 req->reg_wr.wr.num_sge = 0;
1240 req->reg_wr.mr = req->mr;
1241 req->reg_wr.key = req->mr->rkey;
1242 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1243 IB_ACCESS_REMOTE_READ |
1244 IB_ACCESS_REMOTE_WRITE;
1246 sg->addr = cpu_to_le64(req->mr->iova);
1247 put_unaligned_le24(req->mr->length, sg->length);
1248 put_unaligned_le32(req->mr->rkey, sg->key);
1249 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1250 NVME_SGL_FMT_INVALIDATE;
1255 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1256 struct request *rq, struct nvme_command *c)
1258 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1259 struct nvme_rdma_device *dev = queue->device;
1260 struct ib_device *ibdev = dev->dev;
1264 refcount_set(&req->ref, 2); /* send and recv completions */
1266 c->common.flags |= NVME_CMD_SGL_METABUF;
1268 if (!blk_rq_nr_phys_segments(rq))
1269 return nvme_rdma_set_sg_null(c);
1271 req->sg_table.sgl = req->first_sgl;
1272 ret = sg_alloc_table_chained(&req->sg_table,
1273 blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1277 req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1279 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1280 rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1281 if (unlikely(count <= 0)) {
1283 goto out_free_table;
1286 if (count <= dev->num_inline_segments) {
1287 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1288 queue->ctrl->use_inline_data &&
1289 blk_rq_payload_bytes(rq) <=
1290 nvme_rdma_inline_data_size(queue)) {
1291 ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1295 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1296 ret = nvme_rdma_map_sg_single(queue, req, c);
1301 ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1309 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1310 req->nents, rq_data_dir(rq) ==
1311 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1313 sg_free_table_chained(&req->sg_table, true);
1317 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1319 struct nvme_rdma_qe *qe =
1320 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1321 struct nvme_rdma_request *req =
1322 container_of(qe, struct nvme_rdma_request, sqe);
1323 struct request *rq = blk_mq_rq_from_pdu(req);
1325 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1326 nvme_rdma_wr_error(cq, wc, "SEND");
1330 if (refcount_dec_and_test(&req->ref))
1331 nvme_end_request(rq, req->status, req->result);
1334 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1335 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1336 struct ib_send_wr *first)
1338 struct ib_send_wr wr;
1341 sge->addr = qe->dma;
1342 sge->length = sizeof(struct nvme_command),
1343 sge->lkey = queue->device->pd->local_dma_lkey;
1346 wr.wr_cqe = &qe->cqe;
1348 wr.num_sge = num_sge;
1349 wr.opcode = IB_WR_SEND;
1350 wr.send_flags = IB_SEND_SIGNALED;
1357 ret = ib_post_send(queue->qp, first, NULL);
1358 if (unlikely(ret)) {
1359 dev_err(queue->ctrl->ctrl.device,
1360 "%s failed with error code %d\n", __func__, ret);
1365 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1366 struct nvme_rdma_qe *qe)
1368 struct ib_recv_wr wr;
1372 list.addr = qe->dma;
1373 list.length = sizeof(struct nvme_completion);
1374 list.lkey = queue->device->pd->local_dma_lkey;
1376 qe->cqe.done = nvme_rdma_recv_done;
1379 wr.wr_cqe = &qe->cqe;
1383 ret = ib_post_recv(queue->qp, &wr, NULL);
1384 if (unlikely(ret)) {
1385 dev_err(queue->ctrl->ctrl.device,
1386 "%s failed with error code %d\n", __func__, ret);
1391 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1393 u32 queue_idx = nvme_rdma_queue_idx(queue);
1396 return queue->ctrl->admin_tag_set.tags[queue_idx];
1397 return queue->ctrl->tag_set.tags[queue_idx - 1];
1400 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1402 if (unlikely(wc->status != IB_WC_SUCCESS))
1403 nvme_rdma_wr_error(cq, wc, "ASYNC");
1406 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1408 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1409 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1410 struct ib_device *dev = queue->device->dev;
1411 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1412 struct nvme_command *cmd = sqe->data;
1416 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1418 memset(cmd, 0, sizeof(*cmd));
1419 cmd->common.opcode = nvme_admin_async_event;
1420 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1421 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1422 nvme_rdma_set_sg_null(cmd);
1424 sqe->cqe.done = nvme_rdma_async_done;
1426 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1429 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1433 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1434 struct nvme_completion *cqe, struct ib_wc *wc)
1437 struct nvme_rdma_request *req;
1439 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1441 dev_err(queue->ctrl->ctrl.device,
1442 "tag 0x%x on QP %#x not found\n",
1443 cqe->command_id, queue->qp->qp_num);
1444 nvme_rdma_error_recovery(queue->ctrl);
1447 req = blk_mq_rq_to_pdu(rq);
1449 req->status = cqe->status;
1450 req->result = cqe->result;
1452 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1453 if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1454 dev_err(queue->ctrl->ctrl.device,
1455 "Bogus remote invalidation for rkey %#x\n",
1457 nvme_rdma_error_recovery(queue->ctrl);
1459 } else if (req->mr) {
1462 ret = nvme_rdma_inv_rkey(queue, req);
1463 if (unlikely(ret < 0)) {
1464 dev_err(queue->ctrl->ctrl.device,
1465 "Queueing INV WR for rkey %#x failed (%d)\n",
1466 req->mr->rkey, ret);
1467 nvme_rdma_error_recovery(queue->ctrl);
1469 /* the local invalidation completion will end the request */
1473 if (refcount_dec_and_test(&req->ref))
1474 nvme_end_request(rq, req->status, req->result);
1477 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1479 struct nvme_rdma_qe *qe =
1480 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1481 struct nvme_rdma_queue *queue = cq->cq_context;
1482 struct ib_device *ibdev = queue->device->dev;
1483 struct nvme_completion *cqe = qe->data;
1484 const size_t len = sizeof(struct nvme_completion);
1486 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1487 nvme_rdma_wr_error(cq, wc, "RECV");
1491 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1493 * AEN requests are special as they don't time out and can
1494 * survive any kind of queue freeze and often don't respond to
1495 * aborts. We don't even bother to allocate a struct request
1496 * for them but rather special case them here.
1498 if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1499 cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1500 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1503 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1504 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1506 nvme_rdma_post_recv(queue, qe);
1509 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1513 for (i = 0; i < queue->queue_size; i++) {
1514 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1516 goto out_destroy_queue_ib;
1521 out_destroy_queue_ib:
1522 nvme_rdma_destroy_queue_ib(queue);
1526 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1527 struct rdma_cm_event *ev)
1529 struct rdma_cm_id *cm_id = queue->cm_id;
1530 int status = ev->status;
1531 const char *rej_msg;
1532 const struct nvme_rdma_cm_rej *rej_data;
1535 rej_msg = rdma_reject_msg(cm_id, status);
1536 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1538 if (rej_data && rej_data_len >= sizeof(u16)) {
1539 u16 sts = le16_to_cpu(rej_data->sts);
1541 dev_err(queue->ctrl->ctrl.device,
1542 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1543 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1545 dev_err(queue->ctrl->ctrl.device,
1546 "Connect rejected: status %d (%s).\n", status, rej_msg);
1552 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1556 ret = nvme_rdma_create_queue_ib(queue);
1560 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1562 dev_err(queue->ctrl->ctrl.device,
1563 "rdma_resolve_route failed (%d).\n",
1565 goto out_destroy_queue;
1571 nvme_rdma_destroy_queue_ib(queue);
1575 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1577 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1578 struct rdma_conn_param param = { };
1579 struct nvme_rdma_cm_req priv = { };
1582 param.qp_num = queue->qp->qp_num;
1583 param.flow_control = 1;
1585 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1586 /* maximum retry count */
1587 param.retry_count = 7;
1588 param.rnr_retry_count = 7;
1589 param.private_data = &priv;
1590 param.private_data_len = sizeof(priv);
1592 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1593 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1595 * set the admin queue depth to the minimum size
1596 * specified by the Fabrics standard.
1598 if (priv.qid == 0) {
1599 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1600 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1603 * current interpretation of the fabrics spec
1604 * is at minimum you make hrqsize sqsize+1, or a
1605 * 1's based representation of sqsize.
1607 priv.hrqsize = cpu_to_le16(queue->queue_size);
1608 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1611 ret = rdma_connect(queue->cm_id, ¶m);
1613 dev_err(ctrl->ctrl.device,
1614 "rdma_connect failed (%d).\n", ret);
1615 goto out_destroy_queue_ib;
1620 out_destroy_queue_ib:
1621 nvme_rdma_destroy_queue_ib(queue);
1625 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1626 struct rdma_cm_event *ev)
1628 struct nvme_rdma_queue *queue = cm_id->context;
1631 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1632 rdma_event_msg(ev->event), ev->event,
1635 switch (ev->event) {
1636 case RDMA_CM_EVENT_ADDR_RESOLVED:
1637 cm_error = nvme_rdma_addr_resolved(queue);
1639 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1640 cm_error = nvme_rdma_route_resolved(queue);
1642 case RDMA_CM_EVENT_ESTABLISHED:
1643 queue->cm_error = nvme_rdma_conn_established(queue);
1644 /* complete cm_done regardless of success/failure */
1645 complete(&queue->cm_done);
1647 case RDMA_CM_EVENT_REJECTED:
1648 nvme_rdma_destroy_queue_ib(queue);
1649 cm_error = nvme_rdma_conn_rejected(queue, ev);
1651 case RDMA_CM_EVENT_ROUTE_ERROR:
1652 case RDMA_CM_EVENT_CONNECT_ERROR:
1653 case RDMA_CM_EVENT_UNREACHABLE:
1654 nvme_rdma_destroy_queue_ib(queue);
1656 case RDMA_CM_EVENT_ADDR_ERROR:
1657 dev_dbg(queue->ctrl->ctrl.device,
1658 "CM error event %d\n", ev->event);
1659 cm_error = -ECONNRESET;
1661 case RDMA_CM_EVENT_DISCONNECTED:
1662 case RDMA_CM_EVENT_ADDR_CHANGE:
1663 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1664 dev_dbg(queue->ctrl->ctrl.device,
1665 "disconnect received - connection closed\n");
1666 nvme_rdma_error_recovery(queue->ctrl);
1668 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1669 /* device removal is handled via the ib_client API */
1672 dev_err(queue->ctrl->ctrl.device,
1673 "Unexpected RDMA CM event (%d)\n", ev->event);
1674 nvme_rdma_error_recovery(queue->ctrl);
1679 queue->cm_error = cm_error;
1680 complete(&queue->cm_done);
1686 static enum blk_eh_timer_return
1687 nvme_rdma_timeout(struct request *rq, bool reserved)
1689 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1690 struct nvme_rdma_queue *queue = req->queue;
1691 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1693 dev_warn(ctrl->ctrl.device, "I/O %d QID %d timeout\n",
1694 rq->tag, nvme_rdma_queue_idx(queue));
1696 if (ctrl->ctrl.state != NVME_CTRL_LIVE) {
1698 * Teardown immediately if controller times out while starting
1699 * or we are already started error recovery. all outstanding
1700 * requests are completed on shutdown, so we return BLK_EH_DONE.
1702 flush_work(&ctrl->err_work);
1703 nvme_rdma_teardown_io_queues(ctrl, false);
1704 nvme_rdma_teardown_admin_queue(ctrl, false);
1708 dev_warn(ctrl->ctrl.device, "starting error recovery\n");
1709 nvme_rdma_error_recovery(ctrl);
1711 return BLK_EH_RESET_TIMER;
1714 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1715 const struct blk_mq_queue_data *bd)
1717 struct nvme_ns *ns = hctx->queue->queuedata;
1718 struct nvme_rdma_queue *queue = hctx->driver_data;
1719 struct request *rq = bd->rq;
1720 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1721 struct nvme_rdma_qe *sqe = &req->sqe;
1722 struct nvme_command *c = sqe->data;
1723 struct ib_device *dev;
1724 bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1728 WARN_ON_ONCE(rq->tag < 0);
1730 if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1731 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
1733 dev = queue->device->dev;
1734 ib_dma_sync_single_for_cpu(dev, sqe->dma,
1735 sizeof(struct nvme_command), DMA_TO_DEVICE);
1737 ret = nvme_setup_cmd(ns, rq, c);
1741 blk_mq_start_request(rq);
1743 err = nvme_rdma_map_data(queue, rq, c);
1744 if (unlikely(err < 0)) {
1745 dev_err(queue->ctrl->ctrl.device,
1746 "Failed to map data (%d)\n", err);
1747 nvme_cleanup_cmd(rq);
1751 sqe->cqe.done = nvme_rdma_send_done;
1753 ib_dma_sync_single_for_device(dev, sqe->dma,
1754 sizeof(struct nvme_command), DMA_TO_DEVICE);
1756 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1757 req->mr ? &req->reg_wr.wr : NULL);
1758 if (unlikely(err)) {
1759 nvme_rdma_unmap_data(queue, rq);
1765 if (err == -ENOMEM || err == -EAGAIN)
1766 return BLK_STS_RESOURCE;
1767 return BLK_STS_IOERR;
1770 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx)
1772 struct nvme_rdma_queue *queue = hctx->driver_data;
1774 return ib_process_cq_direct(queue->ib_cq, -1);
1777 static void nvme_rdma_complete_rq(struct request *rq)
1779 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1781 nvme_rdma_unmap_data(req->queue, rq);
1782 nvme_complete_rq(rq);
1785 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1787 struct nvme_rdma_ctrl *ctrl = set->driver_data;
1789 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
1790 set->map[HCTX_TYPE_DEFAULT].nr_queues =
1791 ctrl->io_queues[HCTX_TYPE_DEFAULT];
1792 set->map[HCTX_TYPE_READ].nr_queues = ctrl->io_queues[HCTX_TYPE_READ];
1793 if (ctrl->ctrl.opts->nr_write_queues) {
1794 /* separate read/write queues */
1795 set->map[HCTX_TYPE_READ].queue_offset =
1796 ctrl->io_queues[HCTX_TYPE_DEFAULT];
1798 /* mixed read/write queues */
1799 set->map[HCTX_TYPE_READ].queue_offset = 0;
1801 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
1802 ctrl->device->dev, 0);
1803 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
1804 ctrl->device->dev, 0);
1806 if (ctrl->ctrl.opts->nr_poll_queues) {
1807 set->map[HCTX_TYPE_POLL].nr_queues =
1808 ctrl->io_queues[HCTX_TYPE_POLL];
1809 set->map[HCTX_TYPE_POLL].queue_offset =
1810 ctrl->io_queues[HCTX_TYPE_DEFAULT];
1811 if (ctrl->ctrl.opts->nr_write_queues)
1812 set->map[HCTX_TYPE_POLL].queue_offset +=
1813 ctrl->io_queues[HCTX_TYPE_READ];
1814 blk_mq_map_queues(&set->map[HCTX_TYPE_POLL]);
1819 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1820 .queue_rq = nvme_rdma_queue_rq,
1821 .complete = nvme_rdma_complete_rq,
1822 .init_request = nvme_rdma_init_request,
1823 .exit_request = nvme_rdma_exit_request,
1824 .init_hctx = nvme_rdma_init_hctx,
1825 .timeout = nvme_rdma_timeout,
1826 .map_queues = nvme_rdma_map_queues,
1827 .poll = nvme_rdma_poll,
1830 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1831 .queue_rq = nvme_rdma_queue_rq,
1832 .complete = nvme_rdma_complete_rq,
1833 .init_request = nvme_rdma_init_request,
1834 .exit_request = nvme_rdma_exit_request,
1835 .init_hctx = nvme_rdma_init_admin_hctx,
1836 .timeout = nvme_rdma_timeout,
1839 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1841 cancel_work_sync(&ctrl->err_work);
1842 cancel_delayed_work_sync(&ctrl->reconnect_work);
1844 nvme_rdma_teardown_io_queues(ctrl, shutdown);
1846 nvme_shutdown_ctrl(&ctrl->ctrl);
1848 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1849 nvme_rdma_teardown_admin_queue(ctrl, shutdown);
1852 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1854 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1857 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1859 struct nvme_rdma_ctrl *ctrl =
1860 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1862 nvme_stop_ctrl(&ctrl->ctrl);
1863 nvme_rdma_shutdown_ctrl(ctrl, false);
1865 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1866 /* state change failure should never happen */
1871 if (nvme_rdma_setup_ctrl(ctrl, false))
1877 ++ctrl->ctrl.nr_reconnects;
1878 nvme_rdma_reconnect_or_remove(ctrl);
1881 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1883 .module = THIS_MODULE,
1884 .flags = NVME_F_FABRICS,
1885 .reg_read32 = nvmf_reg_read32,
1886 .reg_read64 = nvmf_reg_read64,
1887 .reg_write32 = nvmf_reg_write32,
1888 .free_ctrl = nvme_rdma_free_ctrl,
1889 .submit_async_event = nvme_rdma_submit_async_event,
1890 .delete_ctrl = nvme_rdma_delete_ctrl,
1891 .get_address = nvmf_get_address,
1895 * Fails a connection request if it matches an existing controller
1896 * (association) with the same tuple:
1897 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1899 * if local address is not specified in the request, it will match an
1900 * existing controller with all the other parameters the same and no
1901 * local port address specified as well.
1903 * The ports don't need to be compared as they are intrinsically
1904 * already matched by the port pointers supplied.
1907 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1909 struct nvme_rdma_ctrl *ctrl;
1912 mutex_lock(&nvme_rdma_ctrl_mutex);
1913 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1914 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
1918 mutex_unlock(&nvme_rdma_ctrl_mutex);
1923 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1924 struct nvmf_ctrl_options *opts)
1926 struct nvme_rdma_ctrl *ctrl;
1930 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1932 return ERR_PTR(-ENOMEM);
1933 ctrl->ctrl.opts = opts;
1934 INIT_LIST_HEAD(&ctrl->list);
1936 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
1938 kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
1939 if (!opts->trsvcid) {
1943 opts->mask |= NVMF_OPT_TRSVCID;
1946 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1947 opts->traddr, opts->trsvcid, &ctrl->addr);
1949 pr_err("malformed address passed: %s:%s\n",
1950 opts->traddr, opts->trsvcid);
1954 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1955 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1956 opts->host_traddr, NULL, &ctrl->src_addr);
1958 pr_err("malformed src address passed: %s\n",
1964 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1969 INIT_DELAYED_WORK(&ctrl->reconnect_work,
1970 nvme_rdma_reconnect_ctrl_work);
1971 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1972 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1974 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues +
1975 opts->nr_poll_queues + 1;
1976 ctrl->ctrl.sqsize = opts->queue_size - 1;
1977 ctrl->ctrl.kato = opts->kato;
1980 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1985 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1986 0 /* no quirks, we're perfect! */);
1988 goto out_kfree_queues;
1990 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
1991 WARN_ON_ONCE(!changed);
1993 ret = nvme_rdma_setup_ctrl(ctrl, true);
1995 goto out_uninit_ctrl;
1997 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1998 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2000 nvme_get_ctrl(&ctrl->ctrl);
2002 mutex_lock(&nvme_rdma_ctrl_mutex);
2003 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
2004 mutex_unlock(&nvme_rdma_ctrl_mutex);
2009 nvme_uninit_ctrl(&ctrl->ctrl);
2010 nvme_put_ctrl(&ctrl->ctrl);
2013 return ERR_PTR(ret);
2015 kfree(ctrl->queues);
2018 return ERR_PTR(ret);
2021 static struct nvmf_transport_ops nvme_rdma_transport = {
2023 .module = THIS_MODULE,
2024 .required_opts = NVMF_OPT_TRADDR,
2025 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2026 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
2027 NVMF_OPT_NR_WRITE_QUEUES | NVMF_OPT_NR_POLL_QUEUES,
2028 .create_ctrl = nvme_rdma_create_ctrl,
2031 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2033 struct nvme_rdma_ctrl *ctrl;
2034 struct nvme_rdma_device *ndev;
2037 mutex_lock(&device_list_mutex);
2038 list_for_each_entry(ndev, &device_list, entry) {
2039 if (ndev->dev == ib_device) {
2044 mutex_unlock(&device_list_mutex);
2049 /* Delete all controllers using this device */
2050 mutex_lock(&nvme_rdma_ctrl_mutex);
2051 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2052 if (ctrl->device->dev != ib_device)
2054 nvme_delete_ctrl(&ctrl->ctrl);
2056 mutex_unlock(&nvme_rdma_ctrl_mutex);
2058 flush_workqueue(nvme_delete_wq);
2061 static struct ib_client nvme_rdma_ib_client = {
2062 .name = "nvme_rdma",
2063 .remove = nvme_rdma_remove_one
2066 static int __init nvme_rdma_init_module(void)
2070 ret = ib_register_client(&nvme_rdma_ib_client);
2074 ret = nvmf_register_transport(&nvme_rdma_transport);
2076 goto err_unreg_client;
2081 ib_unregister_client(&nvme_rdma_ib_client);
2085 static void __exit nvme_rdma_cleanup_module(void)
2087 nvmf_unregister_transport(&nvme_rdma_transport);
2088 ib_unregister_client(&nvme_rdma_ib_client);
2091 module_init(nvme_rdma_init_module);
2092 module_exit(nvme_rdma_cleanup_module);
2094 MODULE_LICENSE("GPL v2");