2 * NVMe over Fabrics RDMA host code.
3 * Copyright (c) 2015-2016 HGST, a Western Digital Company.
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 #include <linux/module.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <rdma/mr_pool.h>
19 #include <linux/err.h>
20 #include <linux/string.h>
21 #include <linux/atomic.h>
22 #include <linux/blk-mq.h>
23 #include <linux/blk-mq-rdma.h>
24 #include <linux/types.h>
25 #include <linux/list.h>
26 #include <linux/mutex.h>
27 #include <linux/scatterlist.h>
28 #include <linux/nvme.h>
29 #include <asm/unaligned.h>
31 #include <rdma/ib_verbs.h>
32 #include <rdma/rdma_cm.h>
33 #include <linux/nvme-rdma.h>
39 #define NVME_RDMA_CONNECT_TIMEOUT_MS 3000 /* 3 second */
41 #define NVME_RDMA_MAX_SEGMENTS 256
43 #define NVME_RDMA_MAX_INLINE_SEGMENTS 4
45 struct nvme_rdma_device {
46 struct ib_device *dev;
49 struct list_head entry;
50 unsigned int num_inline_segments;
59 struct nvme_rdma_queue;
60 struct nvme_rdma_request {
61 struct nvme_request req;
63 struct nvme_rdma_qe sqe;
64 union nvme_result result;
67 struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
70 struct ib_reg_wr reg_wr;
71 struct ib_cqe reg_cqe;
72 struct nvme_rdma_queue *queue;
73 struct sg_table sg_table;
74 struct scatterlist first_sgl[];
77 enum nvme_rdma_queue_flags {
78 NVME_RDMA_Q_ALLOCATED = 0,
80 NVME_RDMA_Q_TR_READY = 2,
83 struct nvme_rdma_queue {
84 struct nvme_rdma_qe *rsp_ring;
86 size_t cmnd_capsule_len;
87 struct nvme_rdma_ctrl *ctrl;
88 struct nvme_rdma_device *device;
93 struct rdma_cm_id *cm_id;
95 struct completion cm_done;
98 struct nvme_rdma_ctrl {
99 /* read only in the hot path */
100 struct nvme_rdma_queue *queues;
102 /* other member variables */
103 struct blk_mq_tag_set tag_set;
104 struct work_struct err_work;
106 struct nvme_rdma_qe async_event_sqe;
108 struct delayed_work reconnect_work;
110 struct list_head list;
112 struct blk_mq_tag_set admin_tag_set;
113 struct nvme_rdma_device *device;
117 struct sockaddr_storage addr;
118 struct sockaddr_storage src_addr;
120 struct nvme_ctrl ctrl;
121 bool use_inline_data;
124 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
126 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
129 static LIST_HEAD(device_list);
130 static DEFINE_MUTEX(device_list_mutex);
132 static LIST_HEAD(nvme_rdma_ctrl_list);
133 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
136 * Disabling this option makes small I/O goes faster, but is fundamentally
137 * unsafe. With it turned off we will have to register a global rkey that
138 * allows read and write access to all physical memory.
140 static bool register_always = true;
141 module_param(register_always, bool, 0444);
142 MODULE_PARM_DESC(register_always,
143 "Use memory registration even for contiguous memory regions");
145 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
146 struct rdma_cm_event *event);
147 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
149 static const struct blk_mq_ops nvme_rdma_mq_ops;
150 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
152 /* XXX: really should move to a generic header sooner or later.. */
153 static inline void put_unaligned_le24(u32 val, u8 *p)
160 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
162 return queue - queue->ctrl->queues;
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);
445 queue->device = nvme_rdma_find_get_device(queue->cm_id);
446 if (!queue->device) {
447 dev_err(queue->cm_id->device->dev.parent,
448 "no client data found!\n");
449 return -ECONNREFUSED;
451 ibdev = queue->device->dev;
454 * Spread I/O queues completion vectors according their queue index.
455 * Admin queues can always go on completion vector 0.
457 comp_vector = idx == 0 ? idx : idx - 1;
459 /* +1 for ib_stop_cq */
460 queue->ib_cq = ib_alloc_cq(ibdev, queue,
461 cq_factor * queue->queue_size + 1,
462 comp_vector, IB_POLL_SOFTIRQ);
463 if (IS_ERR(queue->ib_cq)) {
464 ret = PTR_ERR(queue->ib_cq);
468 ret = nvme_rdma_create_qp(queue, send_wr_factor);
470 goto out_destroy_ib_cq;
472 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
473 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
474 if (!queue->rsp_ring) {
479 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
482 nvme_rdma_get_max_fr_pages(ibdev));
484 dev_err(queue->ctrl->ctrl.device,
485 "failed to initialize MR pool sized %d for QID %d\n",
486 queue->queue_size, idx);
487 goto out_destroy_ring;
490 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
495 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
496 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
498 rdma_destroy_qp(queue->cm_id);
500 ib_free_cq(queue->ib_cq);
502 nvme_rdma_dev_put(queue->device);
506 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
507 int idx, size_t queue_size)
509 struct nvme_rdma_queue *queue;
510 struct sockaddr *src_addr = NULL;
513 queue = &ctrl->queues[idx];
515 init_completion(&queue->cm_done);
518 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
520 queue->cmnd_capsule_len = sizeof(struct nvme_command);
522 queue->queue_size = queue_size;
524 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
525 RDMA_PS_TCP, IB_QPT_RC);
526 if (IS_ERR(queue->cm_id)) {
527 dev_info(ctrl->ctrl.device,
528 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
529 return PTR_ERR(queue->cm_id);
532 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
533 src_addr = (struct sockaddr *)&ctrl->src_addr;
535 queue->cm_error = -ETIMEDOUT;
536 ret = rdma_resolve_addr(queue->cm_id, src_addr,
537 (struct sockaddr *)&ctrl->addr,
538 NVME_RDMA_CONNECT_TIMEOUT_MS);
540 dev_info(ctrl->ctrl.device,
541 "rdma_resolve_addr failed (%d).\n", ret);
542 goto out_destroy_cm_id;
545 ret = nvme_rdma_wait_for_cm(queue);
547 dev_info(ctrl->ctrl.device,
548 "rdma connection establishment failed (%d)\n", ret);
549 goto out_destroy_cm_id;
552 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
557 rdma_destroy_id(queue->cm_id);
558 nvme_rdma_destroy_queue_ib(queue);
562 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
564 if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
567 rdma_disconnect(queue->cm_id);
568 ib_drain_qp(queue->qp);
571 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
573 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
576 nvme_rdma_destroy_queue_ib(queue);
577 rdma_destroy_id(queue->cm_id);
580 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
584 for (i = 1; i < ctrl->ctrl.queue_count; i++)
585 nvme_rdma_free_queue(&ctrl->queues[i]);
588 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
592 for (i = 1; i < ctrl->ctrl.queue_count; i++)
593 nvme_rdma_stop_queue(&ctrl->queues[i]);
596 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
601 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx, false);
603 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
606 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags);
608 dev_info(ctrl->ctrl.device,
609 "failed to connect queue: %d ret=%d\n", idx, ret);
613 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
617 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
618 ret = nvme_rdma_start_queue(ctrl, i);
620 goto out_stop_queues;
626 for (i--; i >= 1; i--)
627 nvme_rdma_stop_queue(&ctrl->queues[i]);
631 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
633 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
634 struct ib_device *ibdev = ctrl->device->dev;
635 unsigned int nr_io_queues;
638 nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
641 * we map queues according to the device irq vectors for
642 * optimal locality so we don't need more queues than
643 * completion vectors.
645 nr_io_queues = min_t(unsigned int, nr_io_queues,
646 ibdev->num_comp_vectors);
648 nr_io_queues += min(opts->nr_write_queues, num_online_cpus());
650 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
654 ctrl->ctrl.queue_count = nr_io_queues + 1;
655 if (ctrl->ctrl.queue_count < 2)
658 dev_info(ctrl->ctrl.device,
659 "creating %d I/O queues.\n", nr_io_queues);
661 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
662 ret = nvme_rdma_alloc_queue(ctrl, i,
663 ctrl->ctrl.sqsize + 1);
665 goto out_free_queues;
671 for (i--; i >= 1; i--)
672 nvme_rdma_free_queue(&ctrl->queues[i]);
677 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl,
678 struct blk_mq_tag_set *set)
680 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
682 blk_mq_free_tag_set(set);
683 nvme_rdma_dev_put(ctrl->device);
686 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
689 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
690 struct blk_mq_tag_set *set;
694 set = &ctrl->admin_tag_set;
695 memset(set, 0, sizeof(*set));
696 set->ops = &nvme_rdma_admin_mq_ops;
697 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
698 set->reserved_tags = 2; /* connect + keep-alive */
699 set->numa_node = nctrl->numa_node;
700 set->cmd_size = sizeof(struct nvme_rdma_request) +
701 SG_CHUNK_SIZE * sizeof(struct scatterlist);
702 set->driver_data = ctrl;
703 set->nr_hw_queues = 1;
704 set->timeout = ADMIN_TIMEOUT;
705 set->flags = BLK_MQ_F_NO_SCHED;
707 set = &ctrl->tag_set;
708 memset(set, 0, sizeof(*set));
709 set->ops = &nvme_rdma_mq_ops;
710 set->queue_depth = nctrl->sqsize + 1;
711 set->reserved_tags = 1; /* fabric connect */
712 set->numa_node = nctrl->numa_node;
713 set->flags = BLK_MQ_F_SHOULD_MERGE;
714 set->cmd_size = sizeof(struct nvme_rdma_request) +
715 SG_CHUNK_SIZE * sizeof(struct scatterlist);
716 set->driver_data = ctrl;
717 set->nr_hw_queues = nctrl->queue_count - 1;
718 set->timeout = NVME_IO_TIMEOUT;
719 set->nr_maps = 2 /* default + read */;
722 ret = blk_mq_alloc_tag_set(set);
727 * We need a reference on the device as long as the tag_set is alive,
728 * as the MRs in the request structures need a valid ib_device.
730 ret = nvme_rdma_dev_get(ctrl->device);
733 goto out_free_tagset;
739 blk_mq_free_tag_set(set);
744 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
748 blk_cleanup_queue(ctrl->ctrl.admin_q);
749 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
751 if (ctrl->async_event_sqe.data) {
752 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
753 sizeof(struct nvme_command), DMA_TO_DEVICE);
754 ctrl->async_event_sqe.data = NULL;
756 nvme_rdma_free_queue(&ctrl->queues[0]);
759 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
764 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
768 ctrl->device = ctrl->queues[0].device;
769 ctrl->ctrl.numa_node = dev_to_node(ctrl->device->dev->dma_device);
771 ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
773 error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
774 sizeof(struct nvme_command), DMA_TO_DEVICE);
779 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
780 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
781 error = PTR_ERR(ctrl->ctrl.admin_tagset);
782 goto out_free_async_qe;
785 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
786 if (IS_ERR(ctrl->ctrl.admin_q)) {
787 error = PTR_ERR(ctrl->ctrl.admin_q);
788 goto out_free_tagset;
792 error = nvme_rdma_start_queue(ctrl, 0);
794 goto out_cleanup_queue;
796 error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
799 dev_err(ctrl->ctrl.device,
800 "prop_get NVME_REG_CAP failed\n");
805 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
807 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
811 ctrl->ctrl.max_hw_sectors =
812 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
814 error = nvme_init_identify(&ctrl->ctrl);
821 nvme_rdma_stop_queue(&ctrl->queues[0]);
824 blk_cleanup_queue(ctrl->ctrl.admin_q);
827 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
829 nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
830 sizeof(struct nvme_command), DMA_TO_DEVICE);
831 ctrl->async_event_sqe.data = NULL;
833 nvme_rdma_free_queue(&ctrl->queues[0]);
837 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
841 blk_cleanup_queue(ctrl->ctrl.connect_q);
842 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
844 nvme_rdma_free_io_queues(ctrl);
847 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
851 ret = nvme_rdma_alloc_io_queues(ctrl);
856 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
857 if (IS_ERR(ctrl->ctrl.tagset)) {
858 ret = PTR_ERR(ctrl->ctrl.tagset);
859 goto out_free_io_queues;
862 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
863 if (IS_ERR(ctrl->ctrl.connect_q)) {
864 ret = PTR_ERR(ctrl->ctrl.connect_q);
865 goto out_free_tag_set;
868 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
869 ctrl->ctrl.queue_count - 1);
872 ret = nvme_rdma_start_io_queues(ctrl);
874 goto out_cleanup_connect_q;
878 out_cleanup_connect_q:
880 blk_cleanup_queue(ctrl->ctrl.connect_q);
883 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
885 nvme_rdma_free_io_queues(ctrl);
889 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
892 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
893 nvme_rdma_stop_queue(&ctrl->queues[0]);
894 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set, nvme_cancel_request,
896 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
897 nvme_rdma_destroy_admin_queue(ctrl, remove);
900 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
903 if (ctrl->ctrl.queue_count > 1) {
904 nvme_stop_queues(&ctrl->ctrl);
905 nvme_rdma_stop_io_queues(ctrl);
906 blk_mq_tagset_busy_iter(&ctrl->tag_set, nvme_cancel_request,
909 nvme_start_queues(&ctrl->ctrl);
910 nvme_rdma_destroy_io_queues(ctrl, remove);
914 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
916 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
918 cancel_work_sync(&ctrl->err_work);
919 cancel_delayed_work_sync(&ctrl->reconnect_work);
922 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
924 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
926 if (list_empty(&ctrl->list))
929 mutex_lock(&nvme_rdma_ctrl_mutex);
930 list_del(&ctrl->list);
931 mutex_unlock(&nvme_rdma_ctrl_mutex);
933 nvmf_free_options(nctrl->opts);
939 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
941 /* If we are resetting/deleting then do nothing */
942 if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
943 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
944 ctrl->ctrl.state == NVME_CTRL_LIVE);
948 if (nvmf_should_reconnect(&ctrl->ctrl)) {
949 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
950 ctrl->ctrl.opts->reconnect_delay);
951 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
952 ctrl->ctrl.opts->reconnect_delay * HZ);
954 nvme_delete_ctrl(&ctrl->ctrl);
958 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
963 ret = nvme_rdma_configure_admin_queue(ctrl, new);
967 if (ctrl->ctrl.icdoff) {
968 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
972 if (!(ctrl->ctrl.sgls & (1 << 2))) {
973 dev_err(ctrl->ctrl.device,
974 "Mandatory keyed sgls are not supported!\n");
978 if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
979 dev_warn(ctrl->ctrl.device,
980 "queue_size %zu > ctrl sqsize %u, clamping down\n",
981 ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
984 if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
985 dev_warn(ctrl->ctrl.device,
986 "sqsize %u > ctrl maxcmd %u, clamping down\n",
987 ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
988 ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
991 if (ctrl->ctrl.sgls & (1 << 20))
992 ctrl->use_inline_data = true;
994 if (ctrl->ctrl.queue_count > 1) {
995 ret = nvme_rdma_configure_io_queues(ctrl, new);
1000 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1002 /* state change failure is ok if we're in DELETING state */
1003 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1008 nvme_start_ctrl(&ctrl->ctrl);
1012 if (ctrl->ctrl.queue_count > 1)
1013 nvme_rdma_destroy_io_queues(ctrl, new);
1015 nvme_rdma_stop_queue(&ctrl->queues[0]);
1016 nvme_rdma_destroy_admin_queue(ctrl, new);
1020 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1022 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1023 struct nvme_rdma_ctrl, reconnect_work);
1025 ++ctrl->ctrl.nr_reconnects;
1027 if (nvme_rdma_setup_ctrl(ctrl, false))
1030 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1031 ctrl->ctrl.nr_reconnects);
1033 ctrl->ctrl.nr_reconnects = 0;
1038 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1039 ctrl->ctrl.nr_reconnects);
1040 nvme_rdma_reconnect_or_remove(ctrl);
1043 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1045 struct nvme_rdma_ctrl *ctrl = container_of(work,
1046 struct nvme_rdma_ctrl, err_work);
1048 nvme_stop_keep_alive(&ctrl->ctrl);
1049 nvme_rdma_teardown_io_queues(ctrl, false);
1050 nvme_start_queues(&ctrl->ctrl);
1051 nvme_rdma_teardown_admin_queue(ctrl, false);
1053 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1054 /* state change failure is ok if we're in DELETING state */
1055 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1059 nvme_rdma_reconnect_or_remove(ctrl);
1062 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1064 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1067 queue_work(nvme_wq, &ctrl->err_work);
1070 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1073 struct nvme_rdma_queue *queue = cq->cq_context;
1074 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1076 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1077 dev_info(ctrl->ctrl.device,
1078 "%s for CQE 0x%p failed with status %s (%d)\n",
1080 ib_wc_status_msg(wc->status), wc->status);
1081 nvme_rdma_error_recovery(ctrl);
1084 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1086 if (unlikely(wc->status != IB_WC_SUCCESS))
1087 nvme_rdma_wr_error(cq, wc, "MEMREG");
1090 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1092 struct nvme_rdma_request *req =
1093 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1094 struct request *rq = blk_mq_rq_from_pdu(req);
1096 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1097 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1101 if (refcount_dec_and_test(&req->ref))
1102 nvme_end_request(rq, req->status, req->result);
1106 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1107 struct nvme_rdma_request *req)
1109 struct ib_send_wr wr = {
1110 .opcode = IB_WR_LOCAL_INV,
1113 .send_flags = IB_SEND_SIGNALED,
1114 .ex.invalidate_rkey = req->mr->rkey,
1117 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1118 wr.wr_cqe = &req->reg_cqe;
1120 return ib_post_send(queue->qp, &wr, NULL);
1123 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1126 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1127 struct nvme_rdma_device *dev = queue->device;
1128 struct ib_device *ibdev = dev->dev;
1130 if (!blk_rq_payload_bytes(rq))
1134 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1138 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1139 req->nents, rq_data_dir(rq) ==
1140 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1142 nvme_cleanup_cmd(rq);
1143 sg_free_table_chained(&req->sg_table, true);
1146 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1148 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1151 put_unaligned_le24(0, sg->length);
1152 put_unaligned_le32(0, sg->key);
1153 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1157 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1158 struct nvme_rdma_request *req, struct nvme_command *c,
1161 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1162 struct scatterlist *sgl = req->sg_table.sgl;
1163 struct ib_sge *sge = &req->sge[1];
1167 for (i = 0; i < count; i++, sgl++, sge++) {
1168 sge->addr = sg_dma_address(sgl);
1169 sge->length = sg_dma_len(sgl);
1170 sge->lkey = queue->device->pd->local_dma_lkey;
1174 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1175 sg->length = cpu_to_le32(len);
1176 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1178 req->num_sge += count;
1182 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1183 struct nvme_rdma_request *req, struct nvme_command *c)
1185 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1187 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1188 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1189 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1190 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1194 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1195 struct nvme_rdma_request *req, struct nvme_command *c,
1198 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1201 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1202 if (WARN_ON_ONCE(!req->mr))
1206 * Align the MR to a 4K page size to match the ctrl page size and
1207 * the block virtual boundary.
1209 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1210 if (unlikely(nr < count)) {
1211 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1218 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1220 req->reg_cqe.done = nvme_rdma_memreg_done;
1221 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1222 req->reg_wr.wr.opcode = IB_WR_REG_MR;
1223 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1224 req->reg_wr.wr.num_sge = 0;
1225 req->reg_wr.mr = req->mr;
1226 req->reg_wr.key = req->mr->rkey;
1227 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1228 IB_ACCESS_REMOTE_READ |
1229 IB_ACCESS_REMOTE_WRITE;
1231 sg->addr = cpu_to_le64(req->mr->iova);
1232 put_unaligned_le24(req->mr->length, sg->length);
1233 put_unaligned_le32(req->mr->rkey, sg->key);
1234 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1235 NVME_SGL_FMT_INVALIDATE;
1240 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1241 struct request *rq, struct nvme_command *c)
1243 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1244 struct nvme_rdma_device *dev = queue->device;
1245 struct ib_device *ibdev = dev->dev;
1249 refcount_set(&req->ref, 2); /* send and recv completions */
1251 c->common.flags |= NVME_CMD_SGL_METABUF;
1253 if (!blk_rq_payload_bytes(rq))
1254 return nvme_rdma_set_sg_null(c);
1256 req->sg_table.sgl = req->first_sgl;
1257 ret = sg_alloc_table_chained(&req->sg_table,
1258 blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1262 req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1264 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1265 rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1266 if (unlikely(count <= 0)) {
1268 goto out_free_table;
1271 if (count <= dev->num_inline_segments) {
1272 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1273 queue->ctrl->use_inline_data &&
1274 blk_rq_payload_bytes(rq) <=
1275 nvme_rdma_inline_data_size(queue)) {
1276 ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1280 if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1281 ret = nvme_rdma_map_sg_single(queue, req, c);
1286 ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1294 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1295 req->nents, rq_data_dir(rq) ==
1296 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1298 sg_free_table_chained(&req->sg_table, true);
1302 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1304 struct nvme_rdma_qe *qe =
1305 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1306 struct nvme_rdma_request *req =
1307 container_of(qe, struct nvme_rdma_request, sqe);
1308 struct request *rq = blk_mq_rq_from_pdu(req);
1310 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1311 nvme_rdma_wr_error(cq, wc, "SEND");
1315 if (refcount_dec_and_test(&req->ref))
1316 nvme_end_request(rq, req->status, req->result);
1319 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1320 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1321 struct ib_send_wr *first)
1323 struct ib_send_wr wr;
1326 sge->addr = qe->dma;
1327 sge->length = sizeof(struct nvme_command),
1328 sge->lkey = queue->device->pd->local_dma_lkey;
1331 wr.wr_cqe = &qe->cqe;
1333 wr.num_sge = num_sge;
1334 wr.opcode = IB_WR_SEND;
1335 wr.send_flags = IB_SEND_SIGNALED;
1342 ret = ib_post_send(queue->qp, first, NULL);
1343 if (unlikely(ret)) {
1344 dev_err(queue->ctrl->ctrl.device,
1345 "%s failed with error code %d\n", __func__, ret);
1350 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1351 struct nvme_rdma_qe *qe)
1353 struct ib_recv_wr wr;
1357 list.addr = qe->dma;
1358 list.length = sizeof(struct nvme_completion);
1359 list.lkey = queue->device->pd->local_dma_lkey;
1361 qe->cqe.done = nvme_rdma_recv_done;
1364 wr.wr_cqe = &qe->cqe;
1368 ret = ib_post_recv(queue->qp, &wr, NULL);
1369 if (unlikely(ret)) {
1370 dev_err(queue->ctrl->ctrl.device,
1371 "%s failed with error code %d\n", __func__, ret);
1376 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1378 u32 queue_idx = nvme_rdma_queue_idx(queue);
1381 return queue->ctrl->admin_tag_set.tags[queue_idx];
1382 return queue->ctrl->tag_set.tags[queue_idx - 1];
1385 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1387 if (unlikely(wc->status != IB_WC_SUCCESS))
1388 nvme_rdma_wr_error(cq, wc, "ASYNC");
1391 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1393 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1394 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1395 struct ib_device *dev = queue->device->dev;
1396 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1397 struct nvme_command *cmd = sqe->data;
1401 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1403 memset(cmd, 0, sizeof(*cmd));
1404 cmd->common.opcode = nvme_admin_async_event;
1405 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1406 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1407 nvme_rdma_set_sg_null(cmd);
1409 sqe->cqe.done = nvme_rdma_async_done;
1411 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1414 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1418 static void nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1419 struct nvme_completion *cqe, struct ib_wc *wc)
1422 struct nvme_rdma_request *req;
1424 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1426 dev_err(queue->ctrl->ctrl.device,
1427 "tag 0x%x on QP %#x not found\n",
1428 cqe->command_id, queue->qp->qp_num);
1429 nvme_rdma_error_recovery(queue->ctrl);
1432 req = blk_mq_rq_to_pdu(rq);
1434 req->status = cqe->status;
1435 req->result = cqe->result;
1437 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1438 if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1439 dev_err(queue->ctrl->ctrl.device,
1440 "Bogus remote invalidation for rkey %#x\n",
1442 nvme_rdma_error_recovery(queue->ctrl);
1444 } else if (req->mr) {
1447 ret = nvme_rdma_inv_rkey(queue, req);
1448 if (unlikely(ret < 0)) {
1449 dev_err(queue->ctrl->ctrl.device,
1450 "Queueing INV WR for rkey %#x failed (%d)\n",
1451 req->mr->rkey, ret);
1452 nvme_rdma_error_recovery(queue->ctrl);
1454 /* the local invalidation completion will end the request */
1458 if (refcount_dec_and_test(&req->ref))
1459 nvme_end_request(rq, req->status, req->result);
1462 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1464 struct nvme_rdma_qe *qe =
1465 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1466 struct nvme_rdma_queue *queue = cq->cq_context;
1467 struct ib_device *ibdev = queue->device->dev;
1468 struct nvme_completion *cqe = qe->data;
1469 const size_t len = sizeof(struct nvme_completion);
1471 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1472 nvme_rdma_wr_error(cq, wc, "RECV");
1476 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1478 * AEN requests are special as they don't time out and can
1479 * survive any kind of queue freeze and often don't respond to
1480 * aborts. We don't even bother to allocate a struct request
1481 * for them but rather special case them here.
1483 if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1484 cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1485 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1488 nvme_rdma_process_nvme_rsp(queue, cqe, wc);
1489 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1491 nvme_rdma_post_recv(queue, qe);
1494 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1498 for (i = 0; i < queue->queue_size; i++) {
1499 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1501 goto out_destroy_queue_ib;
1506 out_destroy_queue_ib:
1507 nvme_rdma_destroy_queue_ib(queue);
1511 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1512 struct rdma_cm_event *ev)
1514 struct rdma_cm_id *cm_id = queue->cm_id;
1515 int status = ev->status;
1516 const char *rej_msg;
1517 const struct nvme_rdma_cm_rej *rej_data;
1520 rej_msg = rdma_reject_msg(cm_id, status);
1521 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1523 if (rej_data && rej_data_len >= sizeof(u16)) {
1524 u16 sts = le16_to_cpu(rej_data->sts);
1526 dev_err(queue->ctrl->ctrl.device,
1527 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1528 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1530 dev_err(queue->ctrl->ctrl.device,
1531 "Connect rejected: status %d (%s).\n", status, rej_msg);
1537 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1541 ret = nvme_rdma_create_queue_ib(queue);
1545 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1547 dev_err(queue->ctrl->ctrl.device,
1548 "rdma_resolve_route failed (%d).\n",
1550 goto out_destroy_queue;
1556 nvme_rdma_destroy_queue_ib(queue);
1560 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1562 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1563 struct rdma_conn_param param = { };
1564 struct nvme_rdma_cm_req priv = { };
1567 param.qp_num = queue->qp->qp_num;
1568 param.flow_control = 1;
1570 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1571 /* maximum retry count */
1572 param.retry_count = 7;
1573 param.rnr_retry_count = 7;
1574 param.private_data = &priv;
1575 param.private_data_len = sizeof(priv);
1577 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1578 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1580 * set the admin queue depth to the minimum size
1581 * specified by the Fabrics standard.
1583 if (priv.qid == 0) {
1584 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1585 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1588 * current interpretation of the fabrics spec
1589 * is at minimum you make hrqsize sqsize+1, or a
1590 * 1's based representation of sqsize.
1592 priv.hrqsize = cpu_to_le16(queue->queue_size);
1593 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1596 ret = rdma_connect(queue->cm_id, ¶m);
1598 dev_err(ctrl->ctrl.device,
1599 "rdma_connect failed (%d).\n", ret);
1600 goto out_destroy_queue_ib;
1605 out_destroy_queue_ib:
1606 nvme_rdma_destroy_queue_ib(queue);
1610 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1611 struct rdma_cm_event *ev)
1613 struct nvme_rdma_queue *queue = cm_id->context;
1616 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1617 rdma_event_msg(ev->event), ev->event,
1620 switch (ev->event) {
1621 case RDMA_CM_EVENT_ADDR_RESOLVED:
1622 cm_error = nvme_rdma_addr_resolved(queue);
1624 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1625 cm_error = nvme_rdma_route_resolved(queue);
1627 case RDMA_CM_EVENT_ESTABLISHED:
1628 queue->cm_error = nvme_rdma_conn_established(queue);
1629 /* complete cm_done regardless of success/failure */
1630 complete(&queue->cm_done);
1632 case RDMA_CM_EVENT_REJECTED:
1633 nvme_rdma_destroy_queue_ib(queue);
1634 cm_error = nvme_rdma_conn_rejected(queue, ev);
1636 case RDMA_CM_EVENT_ROUTE_ERROR:
1637 case RDMA_CM_EVENT_CONNECT_ERROR:
1638 case RDMA_CM_EVENT_UNREACHABLE:
1639 nvme_rdma_destroy_queue_ib(queue);
1641 case RDMA_CM_EVENT_ADDR_ERROR:
1642 dev_dbg(queue->ctrl->ctrl.device,
1643 "CM error event %d\n", ev->event);
1644 cm_error = -ECONNRESET;
1646 case RDMA_CM_EVENT_DISCONNECTED:
1647 case RDMA_CM_EVENT_ADDR_CHANGE:
1648 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1649 dev_dbg(queue->ctrl->ctrl.device,
1650 "disconnect received - connection closed\n");
1651 nvme_rdma_error_recovery(queue->ctrl);
1653 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1654 /* device removal is handled via the ib_client API */
1657 dev_err(queue->ctrl->ctrl.device,
1658 "Unexpected RDMA CM event (%d)\n", ev->event);
1659 nvme_rdma_error_recovery(queue->ctrl);
1664 queue->cm_error = cm_error;
1665 complete(&queue->cm_done);
1671 static enum blk_eh_timer_return
1672 nvme_rdma_timeout(struct request *rq, bool reserved)
1674 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1676 dev_warn(req->queue->ctrl->ctrl.device,
1677 "I/O %d QID %d timeout, reset controller\n",
1678 rq->tag, nvme_rdma_queue_idx(req->queue));
1680 /* queue error recovery */
1681 nvme_rdma_error_recovery(req->queue->ctrl);
1683 /* fail with DNR on cmd timeout */
1684 nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1689 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1690 const struct blk_mq_queue_data *bd)
1692 struct nvme_ns *ns = hctx->queue->queuedata;
1693 struct nvme_rdma_queue *queue = hctx->driver_data;
1694 struct request *rq = bd->rq;
1695 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1696 struct nvme_rdma_qe *sqe = &req->sqe;
1697 struct nvme_command *c = sqe->data;
1698 struct ib_device *dev;
1699 bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1703 WARN_ON_ONCE(rq->tag < 0);
1705 if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1706 return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
1708 dev = queue->device->dev;
1709 ib_dma_sync_single_for_cpu(dev, sqe->dma,
1710 sizeof(struct nvme_command), DMA_TO_DEVICE);
1712 ret = nvme_setup_cmd(ns, rq, c);
1716 blk_mq_start_request(rq);
1718 err = nvme_rdma_map_data(queue, rq, c);
1719 if (unlikely(err < 0)) {
1720 dev_err(queue->ctrl->ctrl.device,
1721 "Failed to map data (%d)\n", err);
1722 nvme_cleanup_cmd(rq);
1726 sqe->cqe.done = nvme_rdma_send_done;
1728 ib_dma_sync_single_for_device(dev, sqe->dma,
1729 sizeof(struct nvme_command), DMA_TO_DEVICE);
1731 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1732 req->mr ? &req->reg_wr.wr : NULL);
1733 if (unlikely(err)) {
1734 nvme_rdma_unmap_data(queue, rq);
1740 if (err == -ENOMEM || err == -EAGAIN)
1741 return BLK_STS_RESOURCE;
1742 return BLK_STS_IOERR;
1745 static void nvme_rdma_complete_rq(struct request *rq)
1747 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1749 nvme_rdma_unmap_data(req->queue, rq);
1750 nvme_complete_rq(rq);
1753 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1755 struct nvme_rdma_ctrl *ctrl = set->driver_data;
1757 set->map[HCTX_TYPE_DEFAULT].queue_offset = 0;
1758 set->map[HCTX_TYPE_READ].nr_queues = ctrl->ctrl.opts->nr_io_queues;
1759 if (ctrl->ctrl.opts->nr_write_queues) {
1760 /* separate read/write queues */
1761 set->map[HCTX_TYPE_DEFAULT].nr_queues =
1762 ctrl->ctrl.opts->nr_write_queues;
1763 set->map[HCTX_TYPE_READ].queue_offset =
1764 ctrl->ctrl.opts->nr_write_queues;
1766 /* mixed read/write queues */
1767 set->map[HCTX_TYPE_DEFAULT].nr_queues =
1768 ctrl->ctrl.opts->nr_io_queues;
1769 set->map[HCTX_TYPE_READ].queue_offset = 0;
1771 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_DEFAULT],
1772 ctrl->device->dev, 0);
1773 blk_mq_rdma_map_queues(&set->map[HCTX_TYPE_READ],
1774 ctrl->device->dev, 0);
1778 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1779 .queue_rq = nvme_rdma_queue_rq,
1780 .complete = nvme_rdma_complete_rq,
1781 .init_request = nvme_rdma_init_request,
1782 .exit_request = nvme_rdma_exit_request,
1783 .init_hctx = nvme_rdma_init_hctx,
1784 .timeout = nvme_rdma_timeout,
1785 .map_queues = nvme_rdma_map_queues,
1788 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1789 .queue_rq = nvme_rdma_queue_rq,
1790 .complete = nvme_rdma_complete_rq,
1791 .init_request = nvme_rdma_init_request,
1792 .exit_request = nvme_rdma_exit_request,
1793 .init_hctx = nvme_rdma_init_admin_hctx,
1794 .timeout = nvme_rdma_timeout,
1797 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1799 nvme_rdma_teardown_io_queues(ctrl, shutdown);
1801 nvme_shutdown_ctrl(&ctrl->ctrl);
1803 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1804 nvme_rdma_teardown_admin_queue(ctrl, shutdown);
1807 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1809 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1812 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1814 struct nvme_rdma_ctrl *ctrl =
1815 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1817 nvme_stop_ctrl(&ctrl->ctrl);
1818 nvme_rdma_shutdown_ctrl(ctrl, false);
1820 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1821 /* state change failure should never happen */
1826 if (nvme_rdma_setup_ctrl(ctrl, false))
1832 ++ctrl->ctrl.nr_reconnects;
1833 nvme_rdma_reconnect_or_remove(ctrl);
1836 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1838 .module = THIS_MODULE,
1839 .flags = NVME_F_FABRICS,
1840 .reg_read32 = nvmf_reg_read32,
1841 .reg_read64 = nvmf_reg_read64,
1842 .reg_write32 = nvmf_reg_write32,
1843 .free_ctrl = nvme_rdma_free_ctrl,
1844 .submit_async_event = nvme_rdma_submit_async_event,
1845 .delete_ctrl = nvme_rdma_delete_ctrl,
1846 .get_address = nvmf_get_address,
1847 .stop_ctrl = nvme_rdma_stop_ctrl,
1851 * Fails a connection request if it matches an existing controller
1852 * (association) with the same tuple:
1853 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1855 * if local address is not specified in the request, it will match an
1856 * existing controller with all the other parameters the same and no
1857 * local port address specified as well.
1859 * The ports don't need to be compared as they are intrinsically
1860 * already matched by the port pointers supplied.
1863 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1865 struct nvme_rdma_ctrl *ctrl;
1868 mutex_lock(&nvme_rdma_ctrl_mutex);
1869 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1870 found = nvmf_ip_options_match(&ctrl->ctrl, opts);
1874 mutex_unlock(&nvme_rdma_ctrl_mutex);
1879 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1880 struct nvmf_ctrl_options *opts)
1882 struct nvme_rdma_ctrl *ctrl;
1886 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1888 return ERR_PTR(-ENOMEM);
1889 ctrl->ctrl.opts = opts;
1890 INIT_LIST_HEAD(&ctrl->list);
1892 if (!(opts->mask & NVMF_OPT_TRSVCID)) {
1894 kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
1895 if (!opts->trsvcid) {
1899 opts->mask |= NVMF_OPT_TRSVCID;
1902 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1903 opts->traddr, opts->trsvcid, &ctrl->addr);
1905 pr_err("malformed address passed: %s:%s\n",
1906 opts->traddr, opts->trsvcid);
1910 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1911 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1912 opts->host_traddr, NULL, &ctrl->src_addr);
1914 pr_err("malformed src address passed: %s\n",
1920 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1925 INIT_DELAYED_WORK(&ctrl->reconnect_work,
1926 nvme_rdma_reconnect_ctrl_work);
1927 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1928 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1930 ctrl->ctrl.queue_count = opts->nr_io_queues + opts->nr_write_queues + 1;
1931 ctrl->ctrl.sqsize = opts->queue_size - 1;
1932 ctrl->ctrl.kato = opts->kato;
1935 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1940 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1941 0 /* no quirks, we're perfect! */);
1943 goto out_kfree_queues;
1945 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
1946 WARN_ON_ONCE(!changed);
1948 ret = nvme_rdma_setup_ctrl(ctrl, true);
1950 goto out_uninit_ctrl;
1952 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1953 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1955 nvme_get_ctrl(&ctrl->ctrl);
1957 mutex_lock(&nvme_rdma_ctrl_mutex);
1958 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1959 mutex_unlock(&nvme_rdma_ctrl_mutex);
1964 nvme_uninit_ctrl(&ctrl->ctrl);
1965 nvme_put_ctrl(&ctrl->ctrl);
1968 return ERR_PTR(ret);
1970 kfree(ctrl->queues);
1973 return ERR_PTR(ret);
1976 static struct nvmf_transport_ops nvme_rdma_transport = {
1978 .module = THIS_MODULE,
1979 .required_opts = NVMF_OPT_TRADDR,
1980 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
1981 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO |
1982 NVMF_OPT_NR_WRITE_QUEUES,
1983 .create_ctrl = nvme_rdma_create_ctrl,
1986 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1988 struct nvme_rdma_ctrl *ctrl;
1989 struct nvme_rdma_device *ndev;
1992 mutex_lock(&device_list_mutex);
1993 list_for_each_entry(ndev, &device_list, entry) {
1994 if (ndev->dev == ib_device) {
1999 mutex_unlock(&device_list_mutex);
2004 /* Delete all controllers using this device */
2005 mutex_lock(&nvme_rdma_ctrl_mutex);
2006 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2007 if (ctrl->device->dev != ib_device)
2009 nvme_delete_ctrl(&ctrl->ctrl);
2011 mutex_unlock(&nvme_rdma_ctrl_mutex);
2013 flush_workqueue(nvme_delete_wq);
2016 static struct ib_client nvme_rdma_ib_client = {
2017 .name = "nvme_rdma",
2018 .remove = nvme_rdma_remove_one
2021 static int __init nvme_rdma_init_module(void)
2025 ret = ib_register_client(&nvme_rdma_ib_client);
2029 ret = nvmf_register_transport(&nvme_rdma_transport);
2031 goto err_unreg_client;
2036 ib_unregister_client(&nvme_rdma_ib_client);
2040 static void __exit nvme_rdma_cleanup_module(void)
2042 nvmf_unregister_transport(&nvme_rdma_transport);
2043 ib_unregister_client(&nvme_rdma_ib_client);
2046 module_init(nvme_rdma_init_module);
2047 module_exit(nvme_rdma_cleanup_module);
2049 MODULE_LICENSE("GPL v2");