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 <linux/err.h>
19 #include <linux/string.h>
20 #include <linux/atomic.h>
21 #include <linux/blk-mq.h>
22 #include <linux/types.h>
23 #include <linux/list.h>
24 #include <linux/mutex.h>
25 #include <linux/scatterlist.h>
26 #include <linux/nvme.h>
27 #include <asm/unaligned.h>
29 #include <rdma/ib_verbs.h>
30 #include <rdma/rdma_cm.h>
31 #include <linux/nvme-rdma.h>
37 #define NVME_RDMA_CONNECT_TIMEOUT_MS 1000 /* 1 second */
39 #define NVME_RDMA_MAX_SEGMENT_SIZE 0xffffff /* 24-bit SGL field */
41 #define NVME_RDMA_MAX_SEGMENTS 256
43 #define NVME_RDMA_MAX_INLINE_SEGMENTS 1
45 static const char *const nvme_rdma_cm_status_strs
[] = {
46 [NVME_RDMA_CM_INVALID_LEN
] = "invalid length",
47 [NVME_RDMA_CM_INVALID_RECFMT
] = "invalid record format",
48 [NVME_RDMA_CM_INVALID_QID
] = "invalid queue ID",
49 [NVME_RDMA_CM_INVALID_HSQSIZE
] = "invalid host SQ size",
50 [NVME_RDMA_CM_INVALID_HRQSIZE
] = "invalid host RQ size",
51 [NVME_RDMA_CM_NO_RSC
] = "resource not found",
52 [NVME_RDMA_CM_INVALID_IRD
] = "invalid IRD",
53 [NVME_RDMA_CM_INVALID_ORD
] = "Invalid ORD",
56 static const char *nvme_rdma_cm_msg(enum nvme_rdma_cm_status status
)
58 size_t index
= status
;
60 if (index
< ARRAY_SIZE(nvme_rdma_cm_status_strs
) &&
61 nvme_rdma_cm_status_strs
[index
])
62 return nvme_rdma_cm_status_strs
[index
];
64 return "unrecognized reason";
68 * We handle AEN commands ourselves and don't even let the
69 * block layer know about them.
71 #define NVME_RDMA_NR_AEN_COMMANDS 1
72 #define NVME_RDMA_AQ_BLKMQ_DEPTH \
73 (NVMF_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS)
75 struct nvme_rdma_device
{
76 struct ib_device
*dev
;
79 struct list_head entry
;
88 struct nvme_rdma_queue
;
89 struct nvme_rdma_request
{
90 struct nvme_request req
;
92 struct nvme_rdma_qe sqe
;
93 struct ib_sge sge
[1 + NVME_RDMA_MAX_INLINE_SEGMENTS
];
97 struct ib_reg_wr reg_wr
;
98 struct ib_cqe reg_cqe
;
99 struct nvme_rdma_queue
*queue
;
100 struct sg_table sg_table
;
101 struct scatterlist first_sgl
[];
104 enum nvme_rdma_queue_flags
{
105 NVME_RDMA_Q_CONNECTED
= (1 << 0),
106 NVME_RDMA_IB_QUEUE_ALLOCATED
= (1 << 1),
107 NVME_RDMA_Q_DELETING
= (1 << 2),
108 NVME_RDMA_Q_LIVE
= (1 << 3),
111 struct nvme_rdma_queue
{
112 struct nvme_rdma_qe
*rsp_ring
;
115 size_t cmnd_capsule_len
;
116 struct nvme_rdma_ctrl
*ctrl
;
117 struct nvme_rdma_device
*device
;
122 struct rdma_cm_id
*cm_id
;
124 struct completion cm_done
;
127 struct nvme_rdma_ctrl
{
128 /* read and written in the hot path */
131 /* read only in the hot path */
132 struct nvme_rdma_queue
*queues
;
135 /* other member variables */
136 struct blk_mq_tag_set tag_set
;
137 struct work_struct delete_work
;
138 struct work_struct reset_work
;
139 struct work_struct err_work
;
141 struct nvme_rdma_qe async_event_sqe
;
144 struct delayed_work reconnect_work
;
146 struct list_head list
;
148 struct blk_mq_tag_set admin_tag_set
;
149 struct nvme_rdma_device
*device
;
155 struct sockaddr addr
;
156 struct sockaddr_in addr_in
;
159 struct nvme_ctrl ctrl
;
162 static inline struct nvme_rdma_ctrl
*to_rdma_ctrl(struct nvme_ctrl
*ctrl
)
164 return container_of(ctrl
, struct nvme_rdma_ctrl
, ctrl
);
167 static LIST_HEAD(device_list
);
168 static DEFINE_MUTEX(device_list_mutex
);
170 static LIST_HEAD(nvme_rdma_ctrl_list
);
171 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex
);
173 static struct workqueue_struct
*nvme_rdma_wq
;
176 * Disabling this option makes small I/O goes faster, but is fundamentally
177 * unsafe. With it turned off we will have to register a global rkey that
178 * allows read and write access to all physical memory.
180 static bool register_always
= true;
181 module_param(register_always
, bool, 0444);
182 MODULE_PARM_DESC(register_always
,
183 "Use memory registration even for contiguous memory regions");
185 static int nvme_rdma_cm_handler(struct rdma_cm_id
*cm_id
,
186 struct rdma_cm_event
*event
);
187 static void nvme_rdma_recv_done(struct ib_cq
*cq
, struct ib_wc
*wc
);
189 /* XXX: really should move to a generic header sooner or later.. */
190 static inline void put_unaligned_le24(u32 val
, u8
*p
)
197 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue
*queue
)
199 return queue
- queue
->ctrl
->queues
;
202 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue
*queue
)
204 return queue
->cmnd_capsule_len
- sizeof(struct nvme_command
);
207 static void nvme_rdma_free_qe(struct ib_device
*ibdev
, struct nvme_rdma_qe
*qe
,
208 size_t capsule_size
, enum dma_data_direction dir
)
210 ib_dma_unmap_single(ibdev
, qe
->dma
, capsule_size
, dir
);
214 static int nvme_rdma_alloc_qe(struct ib_device
*ibdev
, struct nvme_rdma_qe
*qe
,
215 size_t capsule_size
, enum dma_data_direction dir
)
217 qe
->data
= kzalloc(capsule_size
, GFP_KERNEL
);
221 qe
->dma
= ib_dma_map_single(ibdev
, qe
->data
, capsule_size
, dir
);
222 if (ib_dma_mapping_error(ibdev
, qe
->dma
)) {
230 static void nvme_rdma_free_ring(struct ib_device
*ibdev
,
231 struct nvme_rdma_qe
*ring
, size_t ib_queue_size
,
232 size_t capsule_size
, enum dma_data_direction dir
)
236 for (i
= 0; i
< ib_queue_size
; i
++)
237 nvme_rdma_free_qe(ibdev
, &ring
[i
], capsule_size
, dir
);
241 static struct nvme_rdma_qe
*nvme_rdma_alloc_ring(struct ib_device
*ibdev
,
242 size_t ib_queue_size
, size_t capsule_size
,
243 enum dma_data_direction dir
)
245 struct nvme_rdma_qe
*ring
;
248 ring
= kcalloc(ib_queue_size
, sizeof(struct nvme_rdma_qe
), GFP_KERNEL
);
252 for (i
= 0; i
< ib_queue_size
; i
++) {
253 if (nvme_rdma_alloc_qe(ibdev
, &ring
[i
], capsule_size
, dir
))
260 nvme_rdma_free_ring(ibdev
, ring
, i
, capsule_size
, dir
);
264 static void nvme_rdma_qp_event(struct ib_event
*event
, void *context
)
266 pr_debug("QP event %s (%d)\n",
267 ib_event_msg(event
->event
), event
->event
);
271 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue
*queue
)
273 wait_for_completion_interruptible_timeout(&queue
->cm_done
,
274 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS
) + 1);
275 return queue
->cm_error
;
278 static int nvme_rdma_create_qp(struct nvme_rdma_queue
*queue
, const int factor
)
280 struct nvme_rdma_device
*dev
= queue
->device
;
281 struct ib_qp_init_attr init_attr
;
284 memset(&init_attr
, 0, sizeof(init_attr
));
285 init_attr
.event_handler
= nvme_rdma_qp_event
;
287 init_attr
.cap
.max_send_wr
= factor
* queue
->queue_size
+ 1;
289 init_attr
.cap
.max_recv_wr
= queue
->queue_size
+ 1;
290 init_attr
.cap
.max_recv_sge
= 1;
291 init_attr
.cap
.max_send_sge
= 1 + NVME_RDMA_MAX_INLINE_SEGMENTS
;
292 init_attr
.sq_sig_type
= IB_SIGNAL_REQ_WR
;
293 init_attr
.qp_type
= IB_QPT_RC
;
294 init_attr
.send_cq
= queue
->ib_cq
;
295 init_attr
.recv_cq
= queue
->ib_cq
;
297 ret
= rdma_create_qp(queue
->cm_id
, dev
->pd
, &init_attr
);
299 queue
->qp
= queue
->cm_id
->qp
;
303 static int nvme_rdma_reinit_request(void *data
, struct request
*rq
)
305 struct nvme_rdma_ctrl
*ctrl
= data
;
306 struct nvme_rdma_device
*dev
= ctrl
->device
;
307 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
310 if (!req
->mr
->need_inval
)
313 ib_dereg_mr(req
->mr
);
315 req
->mr
= ib_alloc_mr(dev
->pd
, IB_MR_TYPE_MEM_REG
,
317 if (IS_ERR(req
->mr
)) {
318 ret
= PTR_ERR(req
->mr
);
323 req
->mr
->need_inval
= false;
329 static void __nvme_rdma_exit_request(struct nvme_rdma_ctrl
*ctrl
,
330 struct request
*rq
, unsigned int queue_idx
)
332 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
333 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[queue_idx
];
334 struct nvme_rdma_device
*dev
= queue
->device
;
337 ib_dereg_mr(req
->mr
);
339 nvme_rdma_free_qe(dev
->dev
, &req
->sqe
, sizeof(struct nvme_command
),
343 static void nvme_rdma_exit_request(void *data
, struct request
*rq
,
344 unsigned int hctx_idx
, unsigned int rq_idx
)
346 return __nvme_rdma_exit_request(data
, rq
, hctx_idx
+ 1);
349 static void nvme_rdma_exit_admin_request(void *data
, struct request
*rq
,
350 unsigned int hctx_idx
, unsigned int rq_idx
)
352 return __nvme_rdma_exit_request(data
, rq
, 0);
355 static int __nvme_rdma_init_request(struct nvme_rdma_ctrl
*ctrl
,
356 struct request
*rq
, unsigned int queue_idx
)
358 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
359 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[queue_idx
];
360 struct nvme_rdma_device
*dev
= queue
->device
;
361 struct ib_device
*ibdev
= dev
->dev
;
364 BUG_ON(queue_idx
>= ctrl
->queue_count
);
366 ret
= nvme_rdma_alloc_qe(ibdev
, &req
->sqe
, sizeof(struct nvme_command
),
371 req
->mr
= ib_alloc_mr(dev
->pd
, IB_MR_TYPE_MEM_REG
,
373 if (IS_ERR(req
->mr
)) {
374 ret
= PTR_ERR(req
->mr
);
383 nvme_rdma_free_qe(dev
->dev
, &req
->sqe
, sizeof(struct nvme_command
),
388 static int nvme_rdma_init_request(void *data
, struct request
*rq
,
389 unsigned int hctx_idx
, unsigned int rq_idx
,
390 unsigned int numa_node
)
392 return __nvme_rdma_init_request(data
, rq
, hctx_idx
+ 1);
395 static int nvme_rdma_init_admin_request(void *data
, struct request
*rq
,
396 unsigned int hctx_idx
, unsigned int rq_idx
,
397 unsigned int numa_node
)
399 return __nvme_rdma_init_request(data
, rq
, 0);
402 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx
*hctx
, void *data
,
403 unsigned int hctx_idx
)
405 struct nvme_rdma_ctrl
*ctrl
= data
;
406 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[hctx_idx
+ 1];
408 BUG_ON(hctx_idx
>= ctrl
->queue_count
);
410 hctx
->driver_data
= queue
;
414 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx
*hctx
, void *data
,
415 unsigned int hctx_idx
)
417 struct nvme_rdma_ctrl
*ctrl
= data
;
418 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[0];
420 BUG_ON(hctx_idx
!= 0);
422 hctx
->driver_data
= queue
;
426 static void nvme_rdma_free_dev(struct kref
*ref
)
428 struct nvme_rdma_device
*ndev
=
429 container_of(ref
, struct nvme_rdma_device
, ref
);
431 mutex_lock(&device_list_mutex
);
432 list_del(&ndev
->entry
);
433 mutex_unlock(&device_list_mutex
);
435 ib_dealloc_pd(ndev
->pd
);
439 static void nvme_rdma_dev_put(struct nvme_rdma_device
*dev
)
441 kref_put(&dev
->ref
, nvme_rdma_free_dev
);
444 static int nvme_rdma_dev_get(struct nvme_rdma_device
*dev
)
446 return kref_get_unless_zero(&dev
->ref
);
449 static struct nvme_rdma_device
*
450 nvme_rdma_find_get_device(struct rdma_cm_id
*cm_id
)
452 struct nvme_rdma_device
*ndev
;
454 mutex_lock(&device_list_mutex
);
455 list_for_each_entry(ndev
, &device_list
, entry
) {
456 if (ndev
->dev
->node_guid
== cm_id
->device
->node_guid
&&
457 nvme_rdma_dev_get(ndev
))
461 ndev
= kzalloc(sizeof(*ndev
), GFP_KERNEL
);
465 ndev
->dev
= cm_id
->device
;
466 kref_init(&ndev
->ref
);
468 ndev
->pd
= ib_alloc_pd(ndev
->dev
,
469 register_always
? 0 : IB_PD_UNSAFE_GLOBAL_RKEY
);
470 if (IS_ERR(ndev
->pd
))
473 if (!(ndev
->dev
->attrs
.device_cap_flags
&
474 IB_DEVICE_MEM_MGT_EXTENSIONS
)) {
475 dev_err(&ndev
->dev
->dev
,
476 "Memory registrations not supported.\n");
480 list_add(&ndev
->entry
, &device_list
);
482 mutex_unlock(&device_list_mutex
);
486 ib_dealloc_pd(ndev
->pd
);
490 mutex_unlock(&device_list_mutex
);
494 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue
*queue
)
496 struct nvme_rdma_device
*dev
;
497 struct ib_device
*ibdev
;
499 if (!test_and_clear_bit(NVME_RDMA_IB_QUEUE_ALLOCATED
, &queue
->flags
))
504 rdma_destroy_qp(queue
->cm_id
);
505 ib_free_cq(queue
->ib_cq
);
507 nvme_rdma_free_ring(ibdev
, queue
->rsp_ring
, queue
->queue_size
,
508 sizeof(struct nvme_completion
), DMA_FROM_DEVICE
);
510 nvme_rdma_dev_put(dev
);
513 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue
*queue
,
514 struct nvme_rdma_device
*dev
)
516 struct ib_device
*ibdev
= dev
->dev
;
517 const int send_wr_factor
= 3; /* MR, SEND, INV */
518 const int cq_factor
= send_wr_factor
+ 1; /* + RECV */
519 int comp_vector
, idx
= nvme_rdma_queue_idx(queue
);
526 * The admin queue is barely used once the controller is live, so don't
527 * bother to spread it out.
532 comp_vector
= idx
% ibdev
->num_comp_vectors
;
535 /* +1 for ib_stop_cq */
536 queue
->ib_cq
= ib_alloc_cq(dev
->dev
, queue
,
537 cq_factor
* queue
->queue_size
+ 1, comp_vector
,
539 if (IS_ERR(queue
->ib_cq
)) {
540 ret
= PTR_ERR(queue
->ib_cq
);
544 ret
= nvme_rdma_create_qp(queue
, send_wr_factor
);
546 goto out_destroy_ib_cq
;
548 queue
->rsp_ring
= nvme_rdma_alloc_ring(ibdev
, queue
->queue_size
,
549 sizeof(struct nvme_completion
), DMA_FROM_DEVICE
);
550 if (!queue
->rsp_ring
) {
554 set_bit(NVME_RDMA_IB_QUEUE_ALLOCATED
, &queue
->flags
);
559 ib_destroy_qp(queue
->qp
);
561 ib_free_cq(queue
->ib_cq
);
566 static int nvme_rdma_init_queue(struct nvme_rdma_ctrl
*ctrl
,
567 int idx
, size_t queue_size
)
569 struct nvme_rdma_queue
*queue
;
572 queue
= &ctrl
->queues
[idx
];
574 init_completion(&queue
->cm_done
);
577 queue
->cmnd_capsule_len
= ctrl
->ctrl
.ioccsz
* 16;
579 queue
->cmnd_capsule_len
= sizeof(struct nvme_command
);
581 queue
->queue_size
= queue_size
;
583 queue
->cm_id
= rdma_create_id(&init_net
, nvme_rdma_cm_handler
, queue
,
584 RDMA_PS_TCP
, IB_QPT_RC
);
585 if (IS_ERR(queue
->cm_id
)) {
586 dev_info(ctrl
->ctrl
.device
,
587 "failed to create CM ID: %ld\n", PTR_ERR(queue
->cm_id
));
588 return PTR_ERR(queue
->cm_id
);
591 queue
->cm_error
= -ETIMEDOUT
;
592 ret
= rdma_resolve_addr(queue
->cm_id
, NULL
, &ctrl
->addr
,
593 NVME_RDMA_CONNECT_TIMEOUT_MS
);
595 dev_info(ctrl
->ctrl
.device
,
596 "rdma_resolve_addr failed (%d).\n", ret
);
597 goto out_destroy_cm_id
;
600 ret
= nvme_rdma_wait_for_cm(queue
);
602 dev_info(ctrl
->ctrl
.device
,
603 "rdma_resolve_addr wait failed (%d).\n", ret
);
604 goto out_destroy_cm_id
;
607 clear_bit(NVME_RDMA_Q_DELETING
, &queue
->flags
);
608 set_bit(NVME_RDMA_Q_CONNECTED
, &queue
->flags
);
613 nvme_rdma_destroy_queue_ib(queue
);
614 rdma_destroy_id(queue
->cm_id
);
618 static void nvme_rdma_stop_queue(struct nvme_rdma_queue
*queue
)
620 rdma_disconnect(queue
->cm_id
);
621 ib_drain_qp(queue
->qp
);
624 static void nvme_rdma_free_queue(struct nvme_rdma_queue
*queue
)
626 nvme_rdma_destroy_queue_ib(queue
);
627 rdma_destroy_id(queue
->cm_id
);
630 static void nvme_rdma_stop_and_free_queue(struct nvme_rdma_queue
*queue
)
632 if (test_and_set_bit(NVME_RDMA_Q_DELETING
, &queue
->flags
))
634 nvme_rdma_stop_queue(queue
);
635 nvme_rdma_free_queue(queue
);
638 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl
*ctrl
)
642 for (i
= 1; i
< ctrl
->queue_count
; i
++)
643 nvme_rdma_stop_and_free_queue(&ctrl
->queues
[i
]);
646 static int nvme_rdma_connect_io_queues(struct nvme_rdma_ctrl
*ctrl
)
650 for (i
= 1; i
< ctrl
->queue_count
; i
++) {
651 ret
= nvmf_connect_io_queue(&ctrl
->ctrl
, i
);
653 dev_info(ctrl
->ctrl
.device
,
654 "failed to connect i/o queue: %d\n", ret
);
655 goto out_free_queues
;
657 set_bit(NVME_RDMA_Q_LIVE
, &ctrl
->queues
[i
].flags
);
663 nvme_rdma_free_io_queues(ctrl
);
667 static int nvme_rdma_init_io_queues(struct nvme_rdma_ctrl
*ctrl
)
671 for (i
= 1; i
< ctrl
->queue_count
; i
++) {
672 ret
= nvme_rdma_init_queue(ctrl
, i
,
673 ctrl
->ctrl
.opts
->queue_size
);
675 dev_info(ctrl
->ctrl
.device
,
676 "failed to initialize i/o queue: %d\n", ret
);
677 goto out_free_queues
;
684 for (i
--; i
>= 1; i
--)
685 nvme_rdma_stop_and_free_queue(&ctrl
->queues
[i
]);
690 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl
*ctrl
)
692 nvme_rdma_free_qe(ctrl
->queues
[0].device
->dev
, &ctrl
->async_event_sqe
,
693 sizeof(struct nvme_command
), DMA_TO_DEVICE
);
694 nvme_rdma_stop_and_free_queue(&ctrl
->queues
[0]);
695 blk_cleanup_queue(ctrl
->ctrl
.admin_q
);
696 blk_mq_free_tag_set(&ctrl
->admin_tag_set
);
697 nvme_rdma_dev_put(ctrl
->device
);
700 static void nvme_rdma_free_ctrl(struct nvme_ctrl
*nctrl
)
702 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(nctrl
);
704 if (list_empty(&ctrl
->list
))
707 mutex_lock(&nvme_rdma_ctrl_mutex
);
708 list_del(&ctrl
->list
);
709 mutex_unlock(&nvme_rdma_ctrl_mutex
);
712 nvmf_free_options(nctrl
->opts
);
717 static void nvme_rdma_reconnect_ctrl_work(struct work_struct
*work
)
719 struct nvme_rdma_ctrl
*ctrl
= container_of(to_delayed_work(work
),
720 struct nvme_rdma_ctrl
, reconnect_work
);
724 if (ctrl
->queue_count
> 1) {
725 nvme_rdma_free_io_queues(ctrl
);
727 ret
= blk_mq_reinit_tagset(&ctrl
->tag_set
);
732 nvme_rdma_stop_and_free_queue(&ctrl
->queues
[0]);
734 ret
= blk_mq_reinit_tagset(&ctrl
->admin_tag_set
);
738 ret
= nvme_rdma_init_queue(ctrl
, 0, NVMF_AQ_DEPTH
);
742 blk_mq_start_stopped_hw_queues(ctrl
->ctrl
.admin_q
, true);
744 ret
= nvmf_connect_admin_queue(&ctrl
->ctrl
);
748 set_bit(NVME_RDMA_Q_LIVE
, &ctrl
->queues
[0].flags
);
750 ret
= nvme_enable_ctrl(&ctrl
->ctrl
, ctrl
->cap
);
754 nvme_start_keep_alive(&ctrl
->ctrl
);
756 if (ctrl
->queue_count
> 1) {
757 ret
= nvme_rdma_init_io_queues(ctrl
);
761 ret
= nvme_rdma_connect_io_queues(ctrl
);
766 changed
= nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_LIVE
);
767 WARN_ON_ONCE(!changed
);
769 if (ctrl
->queue_count
> 1) {
770 nvme_start_queues(&ctrl
->ctrl
);
771 nvme_queue_scan(&ctrl
->ctrl
);
772 nvme_queue_async_events(&ctrl
->ctrl
);
775 dev_info(ctrl
->ctrl
.device
, "Successfully reconnected\n");
780 blk_mq_stop_hw_queues(ctrl
->ctrl
.admin_q
);
782 /* Make sure we are not resetting/deleting */
783 if (ctrl
->ctrl
.state
== NVME_CTRL_RECONNECTING
) {
784 dev_info(ctrl
->ctrl
.device
,
785 "Failed reconnect attempt, requeueing...\n");
786 queue_delayed_work(nvme_rdma_wq
, &ctrl
->reconnect_work
,
787 ctrl
->reconnect_delay
* HZ
);
791 static void nvme_rdma_error_recovery_work(struct work_struct
*work
)
793 struct nvme_rdma_ctrl
*ctrl
= container_of(work
,
794 struct nvme_rdma_ctrl
, err_work
);
797 nvme_stop_keep_alive(&ctrl
->ctrl
);
799 for (i
= 0; i
< ctrl
->queue_count
; i
++) {
800 clear_bit(NVME_RDMA_Q_CONNECTED
, &ctrl
->queues
[i
].flags
);
801 clear_bit(NVME_RDMA_Q_LIVE
, &ctrl
->queues
[i
].flags
);
804 if (ctrl
->queue_count
> 1)
805 nvme_stop_queues(&ctrl
->ctrl
);
806 blk_mq_stop_hw_queues(ctrl
->ctrl
.admin_q
);
808 /* We must take care of fastfail/requeue all our inflight requests */
809 if (ctrl
->queue_count
> 1)
810 blk_mq_tagset_busy_iter(&ctrl
->tag_set
,
811 nvme_cancel_request
, &ctrl
->ctrl
);
812 blk_mq_tagset_busy_iter(&ctrl
->admin_tag_set
,
813 nvme_cancel_request
, &ctrl
->ctrl
);
815 dev_info(ctrl
->ctrl
.device
, "reconnecting in %d seconds\n",
816 ctrl
->reconnect_delay
);
818 queue_delayed_work(nvme_rdma_wq
, &ctrl
->reconnect_work
,
819 ctrl
->reconnect_delay
* HZ
);
822 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl
*ctrl
)
824 if (!nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_RECONNECTING
))
827 queue_work(nvme_rdma_wq
, &ctrl
->err_work
);
830 static void nvme_rdma_wr_error(struct ib_cq
*cq
, struct ib_wc
*wc
,
833 struct nvme_rdma_queue
*queue
= cq
->cq_context
;
834 struct nvme_rdma_ctrl
*ctrl
= queue
->ctrl
;
836 if (ctrl
->ctrl
.state
== NVME_CTRL_LIVE
)
837 dev_info(ctrl
->ctrl
.device
,
838 "%s for CQE 0x%p failed with status %s (%d)\n",
840 ib_wc_status_msg(wc
->status
), wc
->status
);
841 nvme_rdma_error_recovery(ctrl
);
844 static void nvme_rdma_memreg_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
846 if (unlikely(wc
->status
!= IB_WC_SUCCESS
))
847 nvme_rdma_wr_error(cq
, wc
, "MEMREG");
850 static void nvme_rdma_inv_rkey_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
852 if (unlikely(wc
->status
!= IB_WC_SUCCESS
))
853 nvme_rdma_wr_error(cq
, wc
, "LOCAL_INV");
856 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue
*queue
,
857 struct nvme_rdma_request
*req
)
859 struct ib_send_wr
*bad_wr
;
860 struct ib_send_wr wr
= {
861 .opcode
= IB_WR_LOCAL_INV
,
865 .ex
.invalidate_rkey
= req
->mr
->rkey
,
868 req
->reg_cqe
.done
= nvme_rdma_inv_rkey_done
;
869 wr
.wr_cqe
= &req
->reg_cqe
;
871 return ib_post_send(queue
->qp
, &wr
, &bad_wr
);
874 static void nvme_rdma_unmap_data(struct nvme_rdma_queue
*queue
,
877 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
878 struct nvme_rdma_ctrl
*ctrl
= queue
->ctrl
;
879 struct nvme_rdma_device
*dev
= queue
->device
;
880 struct ib_device
*ibdev
= dev
->dev
;
883 if (!blk_rq_bytes(rq
))
886 if (req
->mr
->need_inval
) {
887 res
= nvme_rdma_inv_rkey(queue
, req
);
889 dev_err(ctrl
->ctrl
.device
,
890 "Queueing INV WR for rkey %#x failed (%d)\n",
892 nvme_rdma_error_recovery(queue
->ctrl
);
896 ib_dma_unmap_sg(ibdev
, req
->sg_table
.sgl
,
897 req
->nents
, rq_data_dir(rq
) ==
898 WRITE
? DMA_TO_DEVICE
: DMA_FROM_DEVICE
);
900 nvme_cleanup_cmd(rq
);
901 sg_free_table_chained(&req
->sg_table
, true);
904 static int nvme_rdma_set_sg_null(struct nvme_command
*c
)
906 struct nvme_keyed_sgl_desc
*sg
= &c
->common
.dptr
.ksgl
;
909 put_unaligned_le24(0, sg
->length
);
910 put_unaligned_le32(0, sg
->key
);
911 sg
->type
= NVME_KEY_SGL_FMT_DATA_DESC
<< 4;
915 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue
*queue
,
916 struct nvme_rdma_request
*req
, struct nvme_command
*c
)
918 struct nvme_sgl_desc
*sg
= &c
->common
.dptr
.sgl
;
920 req
->sge
[1].addr
= sg_dma_address(req
->sg_table
.sgl
);
921 req
->sge
[1].length
= sg_dma_len(req
->sg_table
.sgl
);
922 req
->sge
[1].lkey
= queue
->device
->pd
->local_dma_lkey
;
924 sg
->addr
= cpu_to_le64(queue
->ctrl
->ctrl
.icdoff
);
925 sg
->length
= cpu_to_le32(sg_dma_len(req
->sg_table
.sgl
));
926 sg
->type
= (NVME_SGL_FMT_DATA_DESC
<< 4) | NVME_SGL_FMT_OFFSET
;
928 req
->inline_data
= true;
933 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue
*queue
,
934 struct nvme_rdma_request
*req
, struct nvme_command
*c
)
936 struct nvme_keyed_sgl_desc
*sg
= &c
->common
.dptr
.ksgl
;
938 sg
->addr
= cpu_to_le64(sg_dma_address(req
->sg_table
.sgl
));
939 put_unaligned_le24(sg_dma_len(req
->sg_table
.sgl
), sg
->length
);
940 put_unaligned_le32(queue
->device
->pd
->unsafe_global_rkey
, sg
->key
);
941 sg
->type
= NVME_KEY_SGL_FMT_DATA_DESC
<< 4;
945 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue
*queue
,
946 struct nvme_rdma_request
*req
, struct nvme_command
*c
,
949 struct nvme_keyed_sgl_desc
*sg
= &c
->common
.dptr
.ksgl
;
952 nr
= ib_map_mr_sg(req
->mr
, req
->sg_table
.sgl
, count
, NULL
, PAGE_SIZE
);
959 ib_update_fast_reg_key(req
->mr
, ib_inc_rkey(req
->mr
->rkey
));
961 req
->reg_cqe
.done
= nvme_rdma_memreg_done
;
962 memset(&req
->reg_wr
, 0, sizeof(req
->reg_wr
));
963 req
->reg_wr
.wr
.opcode
= IB_WR_REG_MR
;
964 req
->reg_wr
.wr
.wr_cqe
= &req
->reg_cqe
;
965 req
->reg_wr
.wr
.num_sge
= 0;
966 req
->reg_wr
.mr
= req
->mr
;
967 req
->reg_wr
.key
= req
->mr
->rkey
;
968 req
->reg_wr
.access
= IB_ACCESS_LOCAL_WRITE
|
969 IB_ACCESS_REMOTE_READ
|
970 IB_ACCESS_REMOTE_WRITE
;
972 req
->mr
->need_inval
= true;
974 sg
->addr
= cpu_to_le64(req
->mr
->iova
);
975 put_unaligned_le24(req
->mr
->length
, sg
->length
);
976 put_unaligned_le32(req
->mr
->rkey
, sg
->key
);
977 sg
->type
= (NVME_KEY_SGL_FMT_DATA_DESC
<< 4) |
978 NVME_SGL_FMT_INVALIDATE
;
983 static int nvme_rdma_map_data(struct nvme_rdma_queue
*queue
,
984 struct request
*rq
, struct nvme_command
*c
)
986 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
987 struct nvme_rdma_device
*dev
= queue
->device
;
988 struct ib_device
*ibdev
= dev
->dev
;
992 req
->inline_data
= false;
993 req
->mr
->need_inval
= false;
995 c
->common
.flags
|= NVME_CMD_SGL_METABUF
;
997 if (!blk_rq_bytes(rq
))
998 return nvme_rdma_set_sg_null(c
);
1000 req
->sg_table
.sgl
= req
->first_sgl
;
1001 ret
= sg_alloc_table_chained(&req
->sg_table
,
1002 blk_rq_nr_phys_segments(rq
), req
->sg_table
.sgl
);
1006 req
->nents
= blk_rq_map_sg(rq
->q
, rq
, req
->sg_table
.sgl
);
1008 count
= ib_dma_map_sg(ibdev
, req
->sg_table
.sgl
, req
->nents
,
1009 rq_data_dir(rq
) == WRITE
? DMA_TO_DEVICE
: DMA_FROM_DEVICE
);
1010 if (unlikely(count
<= 0)) {
1011 sg_free_table_chained(&req
->sg_table
, true);
1016 if (rq_data_dir(rq
) == WRITE
&& nvme_rdma_queue_idx(queue
) &&
1017 blk_rq_payload_bytes(rq
) <=
1018 nvme_rdma_inline_data_size(queue
))
1019 return nvme_rdma_map_sg_inline(queue
, req
, c
);
1021 if (dev
->pd
->flags
& IB_PD_UNSAFE_GLOBAL_RKEY
)
1022 return nvme_rdma_map_sg_single(queue
, req
, c
);
1025 return nvme_rdma_map_sg_fr(queue
, req
, c
, count
);
1028 static void nvme_rdma_send_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1030 if (unlikely(wc
->status
!= IB_WC_SUCCESS
))
1031 nvme_rdma_wr_error(cq
, wc
, "SEND");
1034 static int nvme_rdma_post_send(struct nvme_rdma_queue
*queue
,
1035 struct nvme_rdma_qe
*qe
, struct ib_sge
*sge
, u32 num_sge
,
1036 struct ib_send_wr
*first
, bool flush
)
1038 struct ib_send_wr wr
, *bad_wr
;
1041 sge
->addr
= qe
->dma
;
1042 sge
->length
= sizeof(struct nvme_command
),
1043 sge
->lkey
= queue
->device
->pd
->local_dma_lkey
;
1045 qe
->cqe
.done
= nvme_rdma_send_done
;
1048 wr
.wr_cqe
= &qe
->cqe
;
1050 wr
.num_sge
= num_sge
;
1051 wr
.opcode
= IB_WR_SEND
;
1055 * Unsignalled send completions are another giant desaster in the
1056 * IB Verbs spec: If we don't regularly post signalled sends
1057 * the send queue will fill up and only a QP reset will rescue us.
1058 * Would have been way to obvious to handle this in hardware or
1059 * at least the RDMA stack..
1061 * This messy and racy code sniplet is copy and pasted from the iSER
1062 * initiator, and the magic '32' comes from there as well.
1064 * Always signal the flushes. The magic request used for the flush
1065 * sequencer is not allocated in our driver's tagset and it's
1066 * triggered to be freed by blk_cleanup_queue(). So we need to
1067 * always mark it as signaled to ensure that the "wr_cqe", which is
1068 * embeded in request's payload, is not freed when __ib_process_cq()
1069 * calls wr_cqe->done().
1071 if ((++queue
->sig_count
% 32) == 0 || flush
)
1072 wr
.send_flags
|= IB_SEND_SIGNALED
;
1079 ret
= ib_post_send(queue
->qp
, first
, &bad_wr
);
1081 dev_err(queue
->ctrl
->ctrl
.device
,
1082 "%s failed with error code %d\n", __func__
, ret
);
1087 static int nvme_rdma_post_recv(struct nvme_rdma_queue
*queue
,
1088 struct nvme_rdma_qe
*qe
)
1090 struct ib_recv_wr wr
, *bad_wr
;
1094 list
.addr
= qe
->dma
;
1095 list
.length
= sizeof(struct nvme_completion
);
1096 list
.lkey
= queue
->device
->pd
->local_dma_lkey
;
1098 qe
->cqe
.done
= nvme_rdma_recv_done
;
1101 wr
.wr_cqe
= &qe
->cqe
;
1105 ret
= ib_post_recv(queue
->qp
, &wr
, &bad_wr
);
1107 dev_err(queue
->ctrl
->ctrl
.device
,
1108 "%s failed with error code %d\n", __func__
, ret
);
1113 static struct blk_mq_tags
*nvme_rdma_tagset(struct nvme_rdma_queue
*queue
)
1115 u32 queue_idx
= nvme_rdma_queue_idx(queue
);
1118 return queue
->ctrl
->admin_tag_set
.tags
[queue_idx
];
1119 return queue
->ctrl
->tag_set
.tags
[queue_idx
- 1];
1122 static void nvme_rdma_submit_async_event(struct nvme_ctrl
*arg
, int aer_idx
)
1124 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(arg
);
1125 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[0];
1126 struct ib_device
*dev
= queue
->device
->dev
;
1127 struct nvme_rdma_qe
*sqe
= &ctrl
->async_event_sqe
;
1128 struct nvme_command
*cmd
= sqe
->data
;
1132 if (WARN_ON_ONCE(aer_idx
!= 0))
1135 ib_dma_sync_single_for_cpu(dev
, sqe
->dma
, sizeof(*cmd
), DMA_TO_DEVICE
);
1137 memset(cmd
, 0, sizeof(*cmd
));
1138 cmd
->common
.opcode
= nvme_admin_async_event
;
1139 cmd
->common
.command_id
= NVME_RDMA_AQ_BLKMQ_DEPTH
;
1140 cmd
->common
.flags
|= NVME_CMD_SGL_METABUF
;
1141 nvme_rdma_set_sg_null(cmd
);
1143 ib_dma_sync_single_for_device(dev
, sqe
->dma
, sizeof(*cmd
),
1146 ret
= nvme_rdma_post_send(queue
, sqe
, &sge
, 1, NULL
, false);
1150 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue
*queue
,
1151 struct nvme_completion
*cqe
, struct ib_wc
*wc
, int tag
)
1154 struct nvme_rdma_request
*req
;
1157 rq
= blk_mq_tag_to_rq(nvme_rdma_tagset(queue
), cqe
->command_id
);
1159 dev_err(queue
->ctrl
->ctrl
.device
,
1160 "tag 0x%x on QP %#x not found\n",
1161 cqe
->command_id
, queue
->qp
->qp_num
);
1162 nvme_rdma_error_recovery(queue
->ctrl
);
1165 req
= blk_mq_rq_to_pdu(rq
);
1170 if ((wc
->wc_flags
& IB_WC_WITH_INVALIDATE
) &&
1171 wc
->ex
.invalidate_rkey
== req
->mr
->rkey
)
1172 req
->mr
->need_inval
= false;
1174 req
->req
.result
= cqe
->result
;
1175 blk_mq_complete_request(rq
, le16_to_cpu(cqe
->status
) >> 1);
1179 static int __nvme_rdma_recv_done(struct ib_cq
*cq
, struct ib_wc
*wc
, int tag
)
1181 struct nvme_rdma_qe
*qe
=
1182 container_of(wc
->wr_cqe
, struct nvme_rdma_qe
, cqe
);
1183 struct nvme_rdma_queue
*queue
= cq
->cq_context
;
1184 struct ib_device
*ibdev
= queue
->device
->dev
;
1185 struct nvme_completion
*cqe
= qe
->data
;
1186 const size_t len
= sizeof(struct nvme_completion
);
1189 if (unlikely(wc
->status
!= IB_WC_SUCCESS
)) {
1190 nvme_rdma_wr_error(cq
, wc
, "RECV");
1194 ib_dma_sync_single_for_cpu(ibdev
, qe
->dma
, len
, DMA_FROM_DEVICE
);
1196 * AEN requests are special as they don't time out and can
1197 * survive any kind of queue freeze and often don't respond to
1198 * aborts. We don't even bother to allocate a struct request
1199 * for them but rather special case them here.
1201 if (unlikely(nvme_rdma_queue_idx(queue
) == 0 &&
1202 cqe
->command_id
>= NVME_RDMA_AQ_BLKMQ_DEPTH
))
1203 nvme_complete_async_event(&queue
->ctrl
->ctrl
, cqe
->status
,
1206 ret
= nvme_rdma_process_nvme_rsp(queue
, cqe
, wc
, tag
);
1207 ib_dma_sync_single_for_device(ibdev
, qe
->dma
, len
, DMA_FROM_DEVICE
);
1209 nvme_rdma_post_recv(queue
, qe
);
1213 static void nvme_rdma_recv_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1215 __nvme_rdma_recv_done(cq
, wc
, -1);
1218 static int nvme_rdma_conn_established(struct nvme_rdma_queue
*queue
)
1222 for (i
= 0; i
< queue
->queue_size
; i
++) {
1223 ret
= nvme_rdma_post_recv(queue
, &queue
->rsp_ring
[i
]);
1225 goto out_destroy_queue_ib
;
1230 out_destroy_queue_ib
:
1231 nvme_rdma_destroy_queue_ib(queue
);
1235 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue
*queue
,
1236 struct rdma_cm_event
*ev
)
1238 struct rdma_cm_id
*cm_id
= queue
->cm_id
;
1239 int status
= ev
->status
;
1240 const char *rej_msg
;
1241 const struct nvme_rdma_cm_rej
*rej_data
;
1244 rej_msg
= rdma_reject_msg(cm_id
, status
);
1245 rej_data
= rdma_consumer_reject_data(cm_id
, ev
, &rej_data_len
);
1247 if (rej_data
&& rej_data_len
>= sizeof(u16
)) {
1248 u16 sts
= le16_to_cpu(rej_data
->sts
);
1250 dev_err(queue
->ctrl
->ctrl
.device
,
1251 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1252 status
, rej_msg
, sts
, nvme_rdma_cm_msg(sts
));
1254 dev_err(queue
->ctrl
->ctrl
.device
,
1255 "Connect rejected: status %d (%s).\n", status
, rej_msg
);
1261 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue
*queue
)
1263 struct nvme_rdma_device
*dev
;
1266 dev
= nvme_rdma_find_get_device(queue
->cm_id
);
1268 dev_err(queue
->cm_id
->device
->dma_device
,
1269 "no client data found!\n");
1270 return -ECONNREFUSED
;
1273 ret
= nvme_rdma_create_queue_ib(queue
, dev
);
1275 nvme_rdma_dev_put(dev
);
1279 ret
= rdma_resolve_route(queue
->cm_id
, NVME_RDMA_CONNECT_TIMEOUT_MS
);
1281 dev_err(queue
->ctrl
->ctrl
.device
,
1282 "rdma_resolve_route failed (%d).\n",
1284 goto out_destroy_queue
;
1290 nvme_rdma_destroy_queue_ib(queue
);
1295 static int nvme_rdma_route_resolved(struct nvme_rdma_queue
*queue
)
1297 struct nvme_rdma_ctrl
*ctrl
= queue
->ctrl
;
1298 struct rdma_conn_param param
= { };
1299 struct nvme_rdma_cm_req priv
= { };
1302 param
.qp_num
= queue
->qp
->qp_num
;
1303 param
.flow_control
= 1;
1305 param
.responder_resources
= queue
->device
->dev
->attrs
.max_qp_rd_atom
;
1306 /* maximum retry count */
1307 param
.retry_count
= 7;
1308 param
.rnr_retry_count
= 7;
1309 param
.private_data
= &priv
;
1310 param
.private_data_len
= sizeof(priv
);
1312 priv
.recfmt
= cpu_to_le16(NVME_RDMA_CM_FMT_1_0
);
1313 priv
.qid
= cpu_to_le16(nvme_rdma_queue_idx(queue
));
1315 * set the admin queue depth to the minimum size
1316 * specified by the Fabrics standard.
1318 if (priv
.qid
== 0) {
1319 priv
.hrqsize
= cpu_to_le16(NVMF_AQ_DEPTH
);
1320 priv
.hsqsize
= cpu_to_le16(NVMF_AQ_DEPTH
- 1);
1323 * current interpretation of the fabrics spec
1324 * is at minimum you make hrqsize sqsize+1, or a
1325 * 1's based representation of sqsize.
1327 priv
.hrqsize
= cpu_to_le16(queue
->queue_size
);
1328 priv
.hsqsize
= cpu_to_le16(queue
->ctrl
->ctrl
.sqsize
);
1331 ret
= rdma_connect(queue
->cm_id
, ¶m
);
1333 dev_err(ctrl
->ctrl
.device
,
1334 "rdma_connect failed (%d).\n", ret
);
1335 goto out_destroy_queue_ib
;
1340 out_destroy_queue_ib
:
1341 nvme_rdma_destroy_queue_ib(queue
);
1345 static int nvme_rdma_cm_handler(struct rdma_cm_id
*cm_id
,
1346 struct rdma_cm_event
*ev
)
1348 struct nvme_rdma_queue
*queue
= cm_id
->context
;
1351 dev_dbg(queue
->ctrl
->ctrl
.device
, "%s (%d): status %d id %p\n",
1352 rdma_event_msg(ev
->event
), ev
->event
,
1355 switch (ev
->event
) {
1356 case RDMA_CM_EVENT_ADDR_RESOLVED
:
1357 cm_error
= nvme_rdma_addr_resolved(queue
);
1359 case RDMA_CM_EVENT_ROUTE_RESOLVED
:
1360 cm_error
= nvme_rdma_route_resolved(queue
);
1362 case RDMA_CM_EVENT_ESTABLISHED
:
1363 queue
->cm_error
= nvme_rdma_conn_established(queue
);
1364 /* complete cm_done regardless of success/failure */
1365 complete(&queue
->cm_done
);
1367 case RDMA_CM_EVENT_REJECTED
:
1368 cm_error
= nvme_rdma_conn_rejected(queue
, ev
);
1370 case RDMA_CM_EVENT_ADDR_ERROR
:
1371 case RDMA_CM_EVENT_ROUTE_ERROR
:
1372 case RDMA_CM_EVENT_CONNECT_ERROR
:
1373 case RDMA_CM_EVENT_UNREACHABLE
:
1374 dev_dbg(queue
->ctrl
->ctrl
.device
,
1375 "CM error event %d\n", ev
->event
);
1376 cm_error
= -ECONNRESET
;
1378 case RDMA_CM_EVENT_DISCONNECTED
:
1379 case RDMA_CM_EVENT_ADDR_CHANGE
:
1380 case RDMA_CM_EVENT_TIMEWAIT_EXIT
:
1381 dev_dbg(queue
->ctrl
->ctrl
.device
,
1382 "disconnect received - connection closed\n");
1383 nvme_rdma_error_recovery(queue
->ctrl
);
1385 case RDMA_CM_EVENT_DEVICE_REMOVAL
:
1386 /* device removal is handled via the ib_client API */
1389 dev_err(queue
->ctrl
->ctrl
.device
,
1390 "Unexpected RDMA CM event (%d)\n", ev
->event
);
1391 nvme_rdma_error_recovery(queue
->ctrl
);
1396 queue
->cm_error
= cm_error
;
1397 complete(&queue
->cm_done
);
1403 static enum blk_eh_timer_return
1404 nvme_rdma_timeout(struct request
*rq
, bool reserved
)
1406 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1408 /* queue error recovery */
1409 nvme_rdma_error_recovery(req
->queue
->ctrl
);
1411 /* fail with DNR on cmd timeout */
1412 rq
->errors
= NVME_SC_ABORT_REQ
| NVME_SC_DNR
;
1414 return BLK_EH_HANDLED
;
1418 * We cannot accept any other command until the Connect command has completed.
1420 static inline bool nvme_rdma_queue_is_ready(struct nvme_rdma_queue
*queue
,
1423 if (unlikely(!test_bit(NVME_RDMA_Q_LIVE
, &queue
->flags
))) {
1424 struct nvme_command
*cmd
= nvme_req(rq
)->cmd
;
1426 if (!blk_rq_is_passthrough(rq
) ||
1427 cmd
->common
.opcode
!= nvme_fabrics_command
||
1428 cmd
->fabrics
.fctype
!= nvme_fabrics_type_connect
)
1435 static int nvme_rdma_queue_rq(struct blk_mq_hw_ctx
*hctx
,
1436 const struct blk_mq_queue_data
*bd
)
1438 struct nvme_ns
*ns
= hctx
->queue
->queuedata
;
1439 struct nvme_rdma_queue
*queue
= hctx
->driver_data
;
1440 struct request
*rq
= bd
->rq
;
1441 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1442 struct nvme_rdma_qe
*sqe
= &req
->sqe
;
1443 struct nvme_command
*c
= sqe
->data
;
1445 struct ib_device
*dev
;
1448 WARN_ON_ONCE(rq
->tag
< 0);
1450 if (!nvme_rdma_queue_is_ready(queue
, rq
))
1451 return BLK_MQ_RQ_QUEUE_BUSY
;
1453 dev
= queue
->device
->dev
;
1454 ib_dma_sync_single_for_cpu(dev
, sqe
->dma
,
1455 sizeof(struct nvme_command
), DMA_TO_DEVICE
);
1457 ret
= nvme_setup_cmd(ns
, rq
, c
);
1458 if (ret
!= BLK_MQ_RQ_QUEUE_OK
)
1461 blk_mq_start_request(rq
);
1463 ret
= nvme_rdma_map_data(queue
, rq
, c
);
1465 dev_err(queue
->ctrl
->ctrl
.device
,
1466 "Failed to map data (%d)\n", ret
);
1467 nvme_cleanup_cmd(rq
);
1471 ib_dma_sync_single_for_device(dev
, sqe
->dma
,
1472 sizeof(struct nvme_command
), DMA_TO_DEVICE
);
1474 if (req_op(rq
) == REQ_OP_FLUSH
)
1476 ret
= nvme_rdma_post_send(queue
, sqe
, req
->sge
, req
->num_sge
,
1477 req
->mr
->need_inval
? &req
->reg_wr
.wr
: NULL
, flush
);
1479 nvme_rdma_unmap_data(queue
, rq
);
1483 return BLK_MQ_RQ_QUEUE_OK
;
1485 return (ret
== -ENOMEM
|| ret
== -EAGAIN
) ?
1486 BLK_MQ_RQ_QUEUE_BUSY
: BLK_MQ_RQ_QUEUE_ERROR
;
1489 static int nvme_rdma_poll(struct blk_mq_hw_ctx
*hctx
, unsigned int tag
)
1491 struct nvme_rdma_queue
*queue
= hctx
->driver_data
;
1492 struct ib_cq
*cq
= queue
->ib_cq
;
1496 ib_req_notify_cq(cq
, IB_CQ_NEXT_COMP
);
1497 while (ib_poll_cq(cq
, 1, &wc
) > 0) {
1498 struct ib_cqe
*cqe
= wc
.wr_cqe
;
1501 if (cqe
->done
== nvme_rdma_recv_done
)
1502 found
|= __nvme_rdma_recv_done(cq
, &wc
, tag
);
1511 static void nvme_rdma_complete_rq(struct request
*rq
)
1513 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1514 struct nvme_rdma_queue
*queue
= req
->queue
;
1517 nvme_rdma_unmap_data(queue
, rq
);
1519 if (unlikely(rq
->errors
)) {
1520 if (nvme_req_needs_retry(rq
, rq
->errors
)) {
1521 nvme_requeue_req(rq
);
1525 if (blk_rq_is_passthrough(rq
))
1528 error
= nvme_error_status(rq
->errors
);
1531 blk_mq_end_request(rq
, error
);
1534 static struct blk_mq_ops nvme_rdma_mq_ops
= {
1535 .queue_rq
= nvme_rdma_queue_rq
,
1536 .complete
= nvme_rdma_complete_rq
,
1537 .init_request
= nvme_rdma_init_request
,
1538 .exit_request
= nvme_rdma_exit_request
,
1539 .reinit_request
= nvme_rdma_reinit_request
,
1540 .init_hctx
= nvme_rdma_init_hctx
,
1541 .poll
= nvme_rdma_poll
,
1542 .timeout
= nvme_rdma_timeout
,
1545 static struct blk_mq_ops nvme_rdma_admin_mq_ops
= {
1546 .queue_rq
= nvme_rdma_queue_rq
,
1547 .complete
= nvme_rdma_complete_rq
,
1548 .init_request
= nvme_rdma_init_admin_request
,
1549 .exit_request
= nvme_rdma_exit_admin_request
,
1550 .reinit_request
= nvme_rdma_reinit_request
,
1551 .init_hctx
= nvme_rdma_init_admin_hctx
,
1552 .timeout
= nvme_rdma_timeout
,
1555 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl
*ctrl
)
1559 error
= nvme_rdma_init_queue(ctrl
, 0, NVMF_AQ_DEPTH
);
1563 ctrl
->device
= ctrl
->queues
[0].device
;
1566 * We need a reference on the device as long as the tag_set is alive,
1567 * as the MRs in the request structures need a valid ib_device.
1570 if (!nvme_rdma_dev_get(ctrl
->device
))
1571 goto out_free_queue
;
1573 ctrl
->max_fr_pages
= min_t(u32
, NVME_RDMA_MAX_SEGMENTS
,
1574 ctrl
->device
->dev
->attrs
.max_fast_reg_page_list_len
);
1576 memset(&ctrl
->admin_tag_set
, 0, sizeof(ctrl
->admin_tag_set
));
1577 ctrl
->admin_tag_set
.ops
= &nvme_rdma_admin_mq_ops
;
1578 ctrl
->admin_tag_set
.queue_depth
= NVME_RDMA_AQ_BLKMQ_DEPTH
;
1579 ctrl
->admin_tag_set
.reserved_tags
= 2; /* connect + keep-alive */
1580 ctrl
->admin_tag_set
.numa_node
= NUMA_NO_NODE
;
1581 ctrl
->admin_tag_set
.cmd_size
= sizeof(struct nvme_rdma_request
) +
1582 SG_CHUNK_SIZE
* sizeof(struct scatterlist
);
1583 ctrl
->admin_tag_set
.driver_data
= ctrl
;
1584 ctrl
->admin_tag_set
.nr_hw_queues
= 1;
1585 ctrl
->admin_tag_set
.timeout
= ADMIN_TIMEOUT
;
1587 error
= blk_mq_alloc_tag_set(&ctrl
->admin_tag_set
);
1591 ctrl
->ctrl
.admin_q
= blk_mq_init_queue(&ctrl
->admin_tag_set
);
1592 if (IS_ERR(ctrl
->ctrl
.admin_q
)) {
1593 error
= PTR_ERR(ctrl
->ctrl
.admin_q
);
1594 goto out_free_tagset
;
1597 error
= nvmf_connect_admin_queue(&ctrl
->ctrl
);
1599 goto out_cleanup_queue
;
1601 set_bit(NVME_RDMA_Q_LIVE
, &ctrl
->queues
[0].flags
);
1603 error
= nvmf_reg_read64(&ctrl
->ctrl
, NVME_REG_CAP
, &ctrl
->cap
);
1605 dev_err(ctrl
->ctrl
.device
,
1606 "prop_get NVME_REG_CAP failed\n");
1607 goto out_cleanup_queue
;
1611 min_t(int, NVME_CAP_MQES(ctrl
->cap
) + 1, ctrl
->ctrl
.sqsize
);
1613 error
= nvme_enable_ctrl(&ctrl
->ctrl
, ctrl
->cap
);
1615 goto out_cleanup_queue
;
1617 ctrl
->ctrl
.max_hw_sectors
=
1618 (ctrl
->max_fr_pages
- 1) << (PAGE_SHIFT
- 9);
1620 error
= nvme_init_identify(&ctrl
->ctrl
);
1622 goto out_cleanup_queue
;
1624 error
= nvme_rdma_alloc_qe(ctrl
->queues
[0].device
->dev
,
1625 &ctrl
->async_event_sqe
, sizeof(struct nvme_command
),
1628 goto out_cleanup_queue
;
1630 nvme_start_keep_alive(&ctrl
->ctrl
);
1635 blk_cleanup_queue(ctrl
->ctrl
.admin_q
);
1637 /* disconnect and drain the queue before freeing the tagset */
1638 nvme_rdma_stop_queue(&ctrl
->queues
[0]);
1639 blk_mq_free_tag_set(&ctrl
->admin_tag_set
);
1641 nvme_rdma_dev_put(ctrl
->device
);
1643 nvme_rdma_free_queue(&ctrl
->queues
[0]);
1647 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl
*ctrl
)
1649 nvme_stop_keep_alive(&ctrl
->ctrl
);
1650 cancel_work_sync(&ctrl
->err_work
);
1651 cancel_delayed_work_sync(&ctrl
->reconnect_work
);
1653 if (ctrl
->queue_count
> 1) {
1654 nvme_stop_queues(&ctrl
->ctrl
);
1655 blk_mq_tagset_busy_iter(&ctrl
->tag_set
,
1656 nvme_cancel_request
, &ctrl
->ctrl
);
1657 nvme_rdma_free_io_queues(ctrl
);
1660 if (test_bit(NVME_RDMA_Q_CONNECTED
, &ctrl
->queues
[0].flags
))
1661 nvme_shutdown_ctrl(&ctrl
->ctrl
);
1663 blk_mq_stop_hw_queues(ctrl
->ctrl
.admin_q
);
1664 blk_mq_tagset_busy_iter(&ctrl
->admin_tag_set
,
1665 nvme_cancel_request
, &ctrl
->ctrl
);
1666 nvme_rdma_destroy_admin_queue(ctrl
);
1669 static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl
*ctrl
, bool shutdown
)
1671 nvme_uninit_ctrl(&ctrl
->ctrl
);
1673 nvme_rdma_shutdown_ctrl(ctrl
);
1675 if (ctrl
->ctrl
.tagset
) {
1676 blk_cleanup_queue(ctrl
->ctrl
.connect_q
);
1677 blk_mq_free_tag_set(&ctrl
->tag_set
);
1678 nvme_rdma_dev_put(ctrl
->device
);
1681 nvme_put_ctrl(&ctrl
->ctrl
);
1684 static void nvme_rdma_del_ctrl_work(struct work_struct
*work
)
1686 struct nvme_rdma_ctrl
*ctrl
= container_of(work
,
1687 struct nvme_rdma_ctrl
, delete_work
);
1689 __nvme_rdma_remove_ctrl(ctrl
, true);
1692 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl
*ctrl
)
1694 if (!nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_DELETING
))
1697 if (!queue_work(nvme_rdma_wq
, &ctrl
->delete_work
))
1703 static int nvme_rdma_del_ctrl(struct nvme_ctrl
*nctrl
)
1705 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(nctrl
);
1709 * Keep a reference until all work is flushed since
1710 * __nvme_rdma_del_ctrl can free the ctrl mem
1712 if (!kref_get_unless_zero(&ctrl
->ctrl
.kref
))
1714 ret
= __nvme_rdma_del_ctrl(ctrl
);
1716 flush_work(&ctrl
->delete_work
);
1717 nvme_put_ctrl(&ctrl
->ctrl
);
1721 static void nvme_rdma_remove_ctrl_work(struct work_struct
*work
)
1723 struct nvme_rdma_ctrl
*ctrl
= container_of(work
,
1724 struct nvme_rdma_ctrl
, delete_work
);
1726 __nvme_rdma_remove_ctrl(ctrl
, false);
1729 static void nvme_rdma_reset_ctrl_work(struct work_struct
*work
)
1731 struct nvme_rdma_ctrl
*ctrl
= container_of(work
,
1732 struct nvme_rdma_ctrl
, reset_work
);
1736 nvme_rdma_shutdown_ctrl(ctrl
);
1738 ret
= nvme_rdma_configure_admin_queue(ctrl
);
1740 /* ctrl is already shutdown, just remove the ctrl */
1741 INIT_WORK(&ctrl
->delete_work
, nvme_rdma_remove_ctrl_work
);
1745 if (ctrl
->queue_count
> 1) {
1746 ret
= blk_mq_reinit_tagset(&ctrl
->tag_set
);
1750 ret
= nvme_rdma_init_io_queues(ctrl
);
1754 ret
= nvme_rdma_connect_io_queues(ctrl
);
1759 changed
= nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_LIVE
);
1760 WARN_ON_ONCE(!changed
);
1762 if (ctrl
->queue_count
> 1) {
1763 nvme_start_queues(&ctrl
->ctrl
);
1764 nvme_queue_scan(&ctrl
->ctrl
);
1765 nvme_queue_async_events(&ctrl
->ctrl
);
1771 /* Deleting this dead controller... */
1772 dev_warn(ctrl
->ctrl
.device
, "Removing after reset failure\n");
1773 WARN_ON(!queue_work(nvme_rdma_wq
, &ctrl
->delete_work
));
1776 static int nvme_rdma_reset_ctrl(struct nvme_ctrl
*nctrl
)
1778 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(nctrl
);
1780 if (!nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_RESETTING
))
1783 if (!queue_work(nvme_rdma_wq
, &ctrl
->reset_work
))
1786 flush_work(&ctrl
->reset_work
);
1791 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops
= {
1793 .module
= THIS_MODULE
,
1795 .reg_read32
= nvmf_reg_read32
,
1796 .reg_read64
= nvmf_reg_read64
,
1797 .reg_write32
= nvmf_reg_write32
,
1798 .reset_ctrl
= nvme_rdma_reset_ctrl
,
1799 .free_ctrl
= nvme_rdma_free_ctrl
,
1800 .submit_async_event
= nvme_rdma_submit_async_event
,
1801 .delete_ctrl
= nvme_rdma_del_ctrl
,
1802 .get_subsysnqn
= nvmf_get_subsysnqn
,
1803 .get_address
= nvmf_get_address
,
1806 static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl
*ctrl
)
1808 struct nvmf_ctrl_options
*opts
= ctrl
->ctrl
.opts
;
1811 ret
= nvme_set_queue_count(&ctrl
->ctrl
, &opts
->nr_io_queues
);
1815 ctrl
->queue_count
= opts
->nr_io_queues
+ 1;
1816 if (ctrl
->queue_count
< 2)
1819 dev_info(ctrl
->ctrl
.device
,
1820 "creating %d I/O queues.\n", opts
->nr_io_queues
);
1822 ret
= nvme_rdma_init_io_queues(ctrl
);
1827 * We need a reference on the device as long as the tag_set is alive,
1828 * as the MRs in the request structures need a valid ib_device.
1831 if (!nvme_rdma_dev_get(ctrl
->device
))
1832 goto out_free_io_queues
;
1834 memset(&ctrl
->tag_set
, 0, sizeof(ctrl
->tag_set
));
1835 ctrl
->tag_set
.ops
= &nvme_rdma_mq_ops
;
1836 ctrl
->tag_set
.queue_depth
= ctrl
->ctrl
.opts
->queue_size
;
1837 ctrl
->tag_set
.reserved_tags
= 1; /* fabric connect */
1838 ctrl
->tag_set
.numa_node
= NUMA_NO_NODE
;
1839 ctrl
->tag_set
.flags
= BLK_MQ_F_SHOULD_MERGE
;
1840 ctrl
->tag_set
.cmd_size
= sizeof(struct nvme_rdma_request
) +
1841 SG_CHUNK_SIZE
* sizeof(struct scatterlist
);
1842 ctrl
->tag_set
.driver_data
= ctrl
;
1843 ctrl
->tag_set
.nr_hw_queues
= ctrl
->queue_count
- 1;
1844 ctrl
->tag_set
.timeout
= NVME_IO_TIMEOUT
;
1846 ret
= blk_mq_alloc_tag_set(&ctrl
->tag_set
);
1849 ctrl
->ctrl
.tagset
= &ctrl
->tag_set
;
1851 ctrl
->ctrl
.connect_q
= blk_mq_init_queue(&ctrl
->tag_set
);
1852 if (IS_ERR(ctrl
->ctrl
.connect_q
)) {
1853 ret
= PTR_ERR(ctrl
->ctrl
.connect_q
);
1854 goto out_free_tag_set
;
1857 ret
= nvme_rdma_connect_io_queues(ctrl
);
1859 goto out_cleanup_connect_q
;
1863 out_cleanup_connect_q
:
1864 blk_cleanup_queue(ctrl
->ctrl
.connect_q
);
1866 blk_mq_free_tag_set(&ctrl
->tag_set
);
1868 nvme_rdma_dev_put(ctrl
->device
);
1870 nvme_rdma_free_io_queues(ctrl
);
1874 static int nvme_rdma_parse_ipaddr(struct sockaddr_in
*in_addr
, char *p
)
1876 u8
*addr
= (u8
*)&in_addr
->sin_addr
.s_addr
;
1877 size_t buflen
= strlen(p
);
1879 /* XXX: handle IPv6 addresses */
1881 if (buflen
> INET_ADDRSTRLEN
)
1883 if (in4_pton(p
, buflen
, addr
, '\0', NULL
) == 0)
1885 in_addr
->sin_family
= AF_INET
;
1889 static struct nvme_ctrl
*nvme_rdma_create_ctrl(struct device
*dev
,
1890 struct nvmf_ctrl_options
*opts
)
1892 struct nvme_rdma_ctrl
*ctrl
;
1896 ctrl
= kzalloc(sizeof(*ctrl
), GFP_KERNEL
);
1898 return ERR_PTR(-ENOMEM
);
1899 ctrl
->ctrl
.opts
= opts
;
1900 INIT_LIST_HEAD(&ctrl
->list
);
1902 ret
= nvme_rdma_parse_ipaddr(&ctrl
->addr_in
, opts
->traddr
);
1904 pr_err("malformed IP address passed: %s\n", opts
->traddr
);
1908 if (opts
->mask
& NVMF_OPT_TRSVCID
) {
1911 ret
= kstrtou16(opts
->trsvcid
, 0, &port
);
1915 ctrl
->addr_in
.sin_port
= cpu_to_be16(port
);
1917 ctrl
->addr_in
.sin_port
= cpu_to_be16(NVME_RDMA_IP_PORT
);
1920 ret
= nvme_init_ctrl(&ctrl
->ctrl
, dev
, &nvme_rdma_ctrl_ops
,
1921 0 /* no quirks, we're perfect! */);
1925 ctrl
->reconnect_delay
= opts
->reconnect_delay
;
1926 INIT_DELAYED_WORK(&ctrl
->reconnect_work
,
1927 nvme_rdma_reconnect_ctrl_work
);
1928 INIT_WORK(&ctrl
->err_work
, nvme_rdma_error_recovery_work
);
1929 INIT_WORK(&ctrl
->delete_work
, nvme_rdma_del_ctrl_work
);
1930 INIT_WORK(&ctrl
->reset_work
, nvme_rdma_reset_ctrl_work
);
1931 spin_lock_init(&ctrl
->lock
);
1933 ctrl
->queue_count
= opts
->nr_io_queues
+ 1; /* +1 for admin queue */
1934 ctrl
->ctrl
.sqsize
= opts
->queue_size
- 1;
1935 ctrl
->ctrl
.kato
= opts
->kato
;
1938 ctrl
->queues
= kcalloc(ctrl
->queue_count
, sizeof(*ctrl
->queues
),
1941 goto out_uninit_ctrl
;
1943 ret
= nvme_rdma_configure_admin_queue(ctrl
);
1945 goto out_kfree_queues
;
1947 /* sanity check icdoff */
1948 if (ctrl
->ctrl
.icdoff
) {
1949 dev_err(ctrl
->ctrl
.device
, "icdoff is not supported!\n");
1950 goto out_remove_admin_queue
;
1953 /* sanity check keyed sgls */
1954 if (!(ctrl
->ctrl
.sgls
& (1 << 20))) {
1955 dev_err(ctrl
->ctrl
.device
, "Mandatory keyed sgls are not support\n");
1956 goto out_remove_admin_queue
;
1959 if (opts
->queue_size
> ctrl
->ctrl
.maxcmd
) {
1960 /* warn if maxcmd is lower than queue_size */
1961 dev_warn(ctrl
->ctrl
.device
,
1962 "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1963 opts
->queue_size
, ctrl
->ctrl
.maxcmd
);
1964 opts
->queue_size
= ctrl
->ctrl
.maxcmd
;
1967 if (opts
->queue_size
> ctrl
->ctrl
.sqsize
+ 1) {
1968 /* warn if sqsize is lower than queue_size */
1969 dev_warn(ctrl
->ctrl
.device
,
1970 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1971 opts
->queue_size
, ctrl
->ctrl
.sqsize
+ 1);
1972 opts
->queue_size
= ctrl
->ctrl
.sqsize
+ 1;
1975 if (opts
->nr_io_queues
) {
1976 ret
= nvme_rdma_create_io_queues(ctrl
);
1978 goto out_remove_admin_queue
;
1981 changed
= nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_LIVE
);
1982 WARN_ON_ONCE(!changed
);
1984 dev_info(ctrl
->ctrl
.device
, "new ctrl: NQN \"%s\", addr %pISp\n",
1985 ctrl
->ctrl
.opts
->subsysnqn
, &ctrl
->addr
);
1987 kref_get(&ctrl
->ctrl
.kref
);
1989 mutex_lock(&nvme_rdma_ctrl_mutex
);
1990 list_add_tail(&ctrl
->list
, &nvme_rdma_ctrl_list
);
1991 mutex_unlock(&nvme_rdma_ctrl_mutex
);
1993 if (opts
->nr_io_queues
) {
1994 nvme_queue_scan(&ctrl
->ctrl
);
1995 nvme_queue_async_events(&ctrl
->ctrl
);
2000 out_remove_admin_queue
:
2001 nvme_stop_keep_alive(&ctrl
->ctrl
);
2002 nvme_rdma_destroy_admin_queue(ctrl
);
2004 kfree(ctrl
->queues
);
2006 nvme_uninit_ctrl(&ctrl
->ctrl
);
2007 nvme_put_ctrl(&ctrl
->ctrl
);
2010 return ERR_PTR(ret
);
2013 return ERR_PTR(ret
);
2016 static struct nvmf_transport_ops nvme_rdma_transport
= {
2018 .required_opts
= NVMF_OPT_TRADDR
,
2019 .allowed_opts
= NVMF_OPT_TRSVCID
| NVMF_OPT_RECONNECT_DELAY
,
2020 .create_ctrl
= nvme_rdma_create_ctrl
,
2023 static void nvme_rdma_add_one(struct ib_device
*ib_device
)
2027 static void nvme_rdma_remove_one(struct ib_device
*ib_device
, void *client_data
)
2029 struct nvme_rdma_ctrl
*ctrl
;
2031 /* Delete all controllers using this device */
2032 mutex_lock(&nvme_rdma_ctrl_mutex
);
2033 list_for_each_entry(ctrl
, &nvme_rdma_ctrl_list
, list
) {
2034 if (ctrl
->device
->dev
!= ib_device
)
2036 dev_info(ctrl
->ctrl
.device
,
2037 "Removing ctrl: NQN \"%s\", addr %pISp\n",
2038 ctrl
->ctrl
.opts
->subsysnqn
, &ctrl
->addr
);
2039 __nvme_rdma_del_ctrl(ctrl
);
2041 mutex_unlock(&nvme_rdma_ctrl_mutex
);
2043 flush_workqueue(nvme_rdma_wq
);
2046 static struct ib_client nvme_rdma_ib_client
= {
2047 .name
= "nvme_rdma",
2048 .add
= nvme_rdma_add_one
,
2049 .remove
= nvme_rdma_remove_one
2052 static int __init
nvme_rdma_init_module(void)
2056 nvme_rdma_wq
= create_workqueue("nvme_rdma_wq");
2060 ret
= ib_register_client(&nvme_rdma_ib_client
);
2062 destroy_workqueue(nvme_rdma_wq
);
2066 nvmf_register_transport(&nvme_rdma_transport
);
2070 static void __exit
nvme_rdma_cleanup_module(void)
2072 nvmf_unregister_transport(&nvme_rdma_transport
);
2073 ib_unregister_client(&nvme_rdma_ib_client
);
2074 destroy_workqueue(nvme_rdma_wq
);
2077 module_init(nvme_rdma_init_module
);
2078 module_exit(nvme_rdma_cleanup_module
);
2080 MODULE_LICENSE("GPL v2");