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 1
45 struct nvme_rdma_device
{
46 struct ib_device
*dev
;
49 struct list_head entry
;
58 struct nvme_rdma_queue
;
59 struct nvme_rdma_request
{
60 struct nvme_request req
;
62 struct nvme_rdma_qe sqe
;
63 union nvme_result result
;
66 struct ib_sge sge
[1 + NVME_RDMA_MAX_INLINE_SEGMENTS
];
69 struct ib_reg_wr reg_wr
;
70 struct ib_cqe reg_cqe
;
71 struct nvme_rdma_queue
*queue
;
72 struct sg_table sg_table
;
73 struct scatterlist first_sgl
[];
76 enum nvme_rdma_queue_flags
{
77 NVME_RDMA_Q_ALLOCATED
= 0,
79 NVME_RDMA_Q_TR_READY
= 2,
82 struct nvme_rdma_queue
{
83 struct nvme_rdma_qe
*rsp_ring
;
85 size_t cmnd_capsule_len
;
86 struct nvme_rdma_ctrl
*ctrl
;
87 struct nvme_rdma_device
*device
;
92 struct rdma_cm_id
*cm_id
;
94 struct completion cm_done
;
97 struct nvme_rdma_ctrl
{
98 /* read only in the hot path */
99 struct nvme_rdma_queue
*queues
;
101 /* other member variables */
102 struct blk_mq_tag_set tag_set
;
103 struct work_struct err_work
;
105 struct nvme_rdma_qe async_event_sqe
;
107 struct delayed_work reconnect_work
;
109 struct list_head list
;
111 struct blk_mq_tag_set admin_tag_set
;
112 struct nvme_rdma_device
*device
;
116 struct sockaddr_storage addr
;
117 struct sockaddr_storage src_addr
;
119 struct nvme_ctrl ctrl
;
122 static inline struct nvme_rdma_ctrl
*to_rdma_ctrl(struct nvme_ctrl
*ctrl
)
124 return container_of(ctrl
, struct nvme_rdma_ctrl
, ctrl
);
127 static LIST_HEAD(device_list
);
128 static DEFINE_MUTEX(device_list_mutex
);
130 static LIST_HEAD(nvme_rdma_ctrl_list
);
131 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex
);
134 * Disabling this option makes small I/O goes faster, but is fundamentally
135 * unsafe. With it turned off we will have to register a global rkey that
136 * allows read and write access to all physical memory.
138 static bool register_always
= true;
139 module_param(register_always
, bool, 0444);
140 MODULE_PARM_DESC(register_always
,
141 "Use memory registration even for contiguous memory regions");
143 static int nvme_rdma_cm_handler(struct rdma_cm_id
*cm_id
,
144 struct rdma_cm_event
*event
);
145 static void nvme_rdma_recv_done(struct ib_cq
*cq
, struct ib_wc
*wc
);
147 static const struct blk_mq_ops nvme_rdma_mq_ops
;
148 static const struct blk_mq_ops nvme_rdma_admin_mq_ops
;
150 /* XXX: really should move to a generic header sooner or later.. */
151 static inline void put_unaligned_le24(u32 val
, u8
*p
)
158 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue
*queue
)
160 return queue
- queue
->ctrl
->queues
;
163 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue
*queue
)
165 return queue
->cmnd_capsule_len
- sizeof(struct nvme_command
);
168 static void nvme_rdma_free_qe(struct ib_device
*ibdev
, struct nvme_rdma_qe
*qe
,
169 size_t capsule_size
, enum dma_data_direction dir
)
171 ib_dma_unmap_single(ibdev
, qe
->dma
, capsule_size
, dir
);
175 static int nvme_rdma_alloc_qe(struct ib_device
*ibdev
, struct nvme_rdma_qe
*qe
,
176 size_t capsule_size
, enum dma_data_direction dir
)
178 qe
->data
= kzalloc(capsule_size
, GFP_KERNEL
);
182 qe
->dma
= ib_dma_map_single(ibdev
, qe
->data
, capsule_size
, dir
);
183 if (ib_dma_mapping_error(ibdev
, qe
->dma
)) {
191 static void nvme_rdma_free_ring(struct ib_device
*ibdev
,
192 struct nvme_rdma_qe
*ring
, size_t ib_queue_size
,
193 size_t capsule_size
, enum dma_data_direction dir
)
197 for (i
= 0; i
< ib_queue_size
; i
++)
198 nvme_rdma_free_qe(ibdev
, &ring
[i
], capsule_size
, dir
);
202 static struct nvme_rdma_qe
*nvme_rdma_alloc_ring(struct ib_device
*ibdev
,
203 size_t ib_queue_size
, size_t capsule_size
,
204 enum dma_data_direction dir
)
206 struct nvme_rdma_qe
*ring
;
209 ring
= kcalloc(ib_queue_size
, sizeof(struct nvme_rdma_qe
), GFP_KERNEL
);
213 for (i
= 0; i
< ib_queue_size
; i
++) {
214 if (nvme_rdma_alloc_qe(ibdev
, &ring
[i
], capsule_size
, dir
))
221 nvme_rdma_free_ring(ibdev
, ring
, i
, capsule_size
, dir
);
225 static void nvme_rdma_qp_event(struct ib_event
*event
, void *context
)
227 pr_debug("QP event %s (%d)\n",
228 ib_event_msg(event
->event
), event
->event
);
232 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue
*queue
)
234 wait_for_completion_interruptible_timeout(&queue
->cm_done
,
235 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS
) + 1);
236 return queue
->cm_error
;
239 static int nvme_rdma_create_qp(struct nvme_rdma_queue
*queue
, const int factor
)
241 struct nvme_rdma_device
*dev
= queue
->device
;
242 struct ib_qp_init_attr init_attr
;
245 memset(&init_attr
, 0, sizeof(init_attr
));
246 init_attr
.event_handler
= nvme_rdma_qp_event
;
248 init_attr
.cap
.max_send_wr
= factor
* queue
->queue_size
+ 1;
250 init_attr
.cap
.max_recv_wr
= queue
->queue_size
+ 1;
251 init_attr
.cap
.max_recv_sge
= 1;
252 init_attr
.cap
.max_send_sge
= 1 + NVME_RDMA_MAX_INLINE_SEGMENTS
;
253 init_attr
.sq_sig_type
= IB_SIGNAL_REQ_WR
;
254 init_attr
.qp_type
= IB_QPT_RC
;
255 init_attr
.send_cq
= queue
->ib_cq
;
256 init_attr
.recv_cq
= queue
->ib_cq
;
258 ret
= rdma_create_qp(queue
->cm_id
, dev
->pd
, &init_attr
);
260 queue
->qp
= queue
->cm_id
->qp
;
264 static void nvme_rdma_exit_request(struct blk_mq_tag_set
*set
,
265 struct request
*rq
, unsigned int hctx_idx
)
267 struct nvme_rdma_ctrl
*ctrl
= set
->driver_data
;
268 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
269 int queue_idx
= (set
== &ctrl
->tag_set
) ? hctx_idx
+ 1 : 0;
270 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[queue_idx
];
271 struct nvme_rdma_device
*dev
= queue
->device
;
273 nvme_rdma_free_qe(dev
->dev
, &req
->sqe
, sizeof(struct nvme_command
),
277 static int nvme_rdma_init_request(struct blk_mq_tag_set
*set
,
278 struct request
*rq
, unsigned int hctx_idx
,
279 unsigned int numa_node
)
281 struct nvme_rdma_ctrl
*ctrl
= set
->driver_data
;
282 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
283 int queue_idx
= (set
== &ctrl
->tag_set
) ? hctx_idx
+ 1 : 0;
284 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[queue_idx
];
285 struct nvme_rdma_device
*dev
= queue
->device
;
286 struct ib_device
*ibdev
= dev
->dev
;
289 ret
= nvme_rdma_alloc_qe(ibdev
, &req
->sqe
, sizeof(struct nvme_command
),
299 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx
*hctx
, void *data
,
300 unsigned int hctx_idx
)
302 struct nvme_rdma_ctrl
*ctrl
= data
;
303 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[hctx_idx
+ 1];
305 BUG_ON(hctx_idx
>= ctrl
->ctrl
.queue_count
);
307 hctx
->driver_data
= queue
;
311 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx
*hctx
, void *data
,
312 unsigned int hctx_idx
)
314 struct nvme_rdma_ctrl
*ctrl
= data
;
315 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[0];
317 BUG_ON(hctx_idx
!= 0);
319 hctx
->driver_data
= queue
;
323 static void nvme_rdma_free_dev(struct kref
*ref
)
325 struct nvme_rdma_device
*ndev
=
326 container_of(ref
, struct nvme_rdma_device
, ref
);
328 mutex_lock(&device_list_mutex
);
329 list_del(&ndev
->entry
);
330 mutex_unlock(&device_list_mutex
);
332 ib_dealloc_pd(ndev
->pd
);
336 static void nvme_rdma_dev_put(struct nvme_rdma_device
*dev
)
338 kref_put(&dev
->ref
, nvme_rdma_free_dev
);
341 static int nvme_rdma_dev_get(struct nvme_rdma_device
*dev
)
343 return kref_get_unless_zero(&dev
->ref
);
346 static struct nvme_rdma_device
*
347 nvme_rdma_find_get_device(struct rdma_cm_id
*cm_id
)
349 struct nvme_rdma_device
*ndev
;
351 mutex_lock(&device_list_mutex
);
352 list_for_each_entry(ndev
, &device_list
, entry
) {
353 if (ndev
->dev
->node_guid
== cm_id
->device
->node_guid
&&
354 nvme_rdma_dev_get(ndev
))
358 ndev
= kzalloc(sizeof(*ndev
), GFP_KERNEL
);
362 ndev
->dev
= cm_id
->device
;
363 kref_init(&ndev
->ref
);
365 ndev
->pd
= ib_alloc_pd(ndev
->dev
,
366 register_always
? 0 : IB_PD_UNSAFE_GLOBAL_RKEY
);
367 if (IS_ERR(ndev
->pd
))
370 if (!(ndev
->dev
->attrs
.device_cap_flags
&
371 IB_DEVICE_MEM_MGT_EXTENSIONS
)) {
372 dev_err(&ndev
->dev
->dev
,
373 "Memory registrations not supported.\n");
377 list_add(&ndev
->entry
, &device_list
);
379 mutex_unlock(&device_list_mutex
);
383 ib_dealloc_pd(ndev
->pd
);
387 mutex_unlock(&device_list_mutex
);
391 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue
*queue
)
393 struct nvme_rdma_device
*dev
;
394 struct ib_device
*ibdev
;
396 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY
, &queue
->flags
))
402 ib_mr_pool_destroy(queue
->qp
, &queue
->qp
->rdma_mrs
);
405 * The cm_id object might have been destroyed during RDMA connection
406 * establishment error flow to avoid getting other cma events, thus
407 * the destruction of the QP shouldn't use rdma_cm API.
409 ib_destroy_qp(queue
->qp
);
410 ib_free_cq(queue
->ib_cq
);
412 nvme_rdma_free_ring(ibdev
, queue
->rsp_ring
, queue
->queue_size
,
413 sizeof(struct nvme_completion
), DMA_FROM_DEVICE
);
415 nvme_rdma_dev_put(dev
);
418 static int nvme_rdma_get_max_fr_pages(struct ib_device
*ibdev
)
420 return min_t(u32
, NVME_RDMA_MAX_SEGMENTS
,
421 ibdev
->attrs
.max_fast_reg_page_list_len
);
424 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue
*queue
)
426 struct ib_device
*ibdev
;
427 const int send_wr_factor
= 3; /* MR, SEND, INV */
428 const int cq_factor
= send_wr_factor
+ 1; /* + RECV */
429 int comp_vector
, idx
= nvme_rdma_queue_idx(queue
);
432 queue
->device
= nvme_rdma_find_get_device(queue
->cm_id
);
433 if (!queue
->device
) {
434 dev_err(queue
->cm_id
->device
->dev
.parent
,
435 "no client data found!\n");
436 return -ECONNREFUSED
;
438 ibdev
= queue
->device
->dev
;
441 * Spread I/O queues completion vectors according their queue index.
442 * Admin queues can always go on completion vector 0.
444 comp_vector
= idx
== 0 ? idx
: idx
- 1;
446 /* +1 for ib_stop_cq */
447 queue
->ib_cq
= ib_alloc_cq(ibdev
, queue
,
448 cq_factor
* queue
->queue_size
+ 1,
449 comp_vector
, IB_POLL_SOFTIRQ
);
450 if (IS_ERR(queue
->ib_cq
)) {
451 ret
= PTR_ERR(queue
->ib_cq
);
455 ret
= nvme_rdma_create_qp(queue
, send_wr_factor
);
457 goto out_destroy_ib_cq
;
459 queue
->rsp_ring
= nvme_rdma_alloc_ring(ibdev
, queue
->queue_size
,
460 sizeof(struct nvme_completion
), DMA_FROM_DEVICE
);
461 if (!queue
->rsp_ring
) {
466 ret
= ib_mr_pool_init(queue
->qp
, &queue
->qp
->rdma_mrs
,
469 nvme_rdma_get_max_fr_pages(ibdev
));
471 dev_err(queue
->ctrl
->ctrl
.device
,
472 "failed to initialize MR pool sized %d for QID %d\n",
473 queue
->queue_size
, idx
);
474 goto out_destroy_ring
;
477 set_bit(NVME_RDMA_Q_TR_READY
, &queue
->flags
);
482 nvme_rdma_free_ring(ibdev
, queue
->rsp_ring
, queue
->queue_size
,
483 sizeof(struct nvme_completion
), DMA_FROM_DEVICE
);
485 rdma_destroy_qp(queue
->cm_id
);
487 ib_free_cq(queue
->ib_cq
);
489 nvme_rdma_dev_put(queue
->device
);
493 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl
*ctrl
,
494 int idx
, size_t queue_size
)
496 struct nvme_rdma_queue
*queue
;
497 struct sockaddr
*src_addr
= NULL
;
500 queue
= &ctrl
->queues
[idx
];
502 init_completion(&queue
->cm_done
);
505 queue
->cmnd_capsule_len
= ctrl
->ctrl
.ioccsz
* 16;
507 queue
->cmnd_capsule_len
= sizeof(struct nvme_command
);
509 queue
->queue_size
= queue_size
;
511 queue
->cm_id
= rdma_create_id(&init_net
, nvme_rdma_cm_handler
, queue
,
512 RDMA_PS_TCP
, IB_QPT_RC
);
513 if (IS_ERR(queue
->cm_id
)) {
514 dev_info(ctrl
->ctrl
.device
,
515 "failed to create CM ID: %ld\n", PTR_ERR(queue
->cm_id
));
516 return PTR_ERR(queue
->cm_id
);
519 if (ctrl
->ctrl
.opts
->mask
& NVMF_OPT_HOST_TRADDR
)
520 src_addr
= (struct sockaddr
*)&ctrl
->src_addr
;
522 queue
->cm_error
= -ETIMEDOUT
;
523 ret
= rdma_resolve_addr(queue
->cm_id
, src_addr
,
524 (struct sockaddr
*)&ctrl
->addr
,
525 NVME_RDMA_CONNECT_TIMEOUT_MS
);
527 dev_info(ctrl
->ctrl
.device
,
528 "rdma_resolve_addr failed (%d).\n", ret
);
529 goto out_destroy_cm_id
;
532 ret
= nvme_rdma_wait_for_cm(queue
);
534 dev_info(ctrl
->ctrl
.device
,
535 "rdma connection establishment failed (%d)\n", ret
);
536 goto out_destroy_cm_id
;
539 set_bit(NVME_RDMA_Q_ALLOCATED
, &queue
->flags
);
544 rdma_destroy_id(queue
->cm_id
);
545 nvme_rdma_destroy_queue_ib(queue
);
549 static void nvme_rdma_stop_queue(struct nvme_rdma_queue
*queue
)
551 if (!test_and_clear_bit(NVME_RDMA_Q_LIVE
, &queue
->flags
))
554 rdma_disconnect(queue
->cm_id
);
555 ib_drain_qp(queue
->qp
);
558 static void nvme_rdma_free_queue(struct nvme_rdma_queue
*queue
)
560 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED
, &queue
->flags
))
563 if (nvme_rdma_queue_idx(queue
) == 0) {
564 nvme_rdma_free_qe(queue
->device
->dev
,
565 &queue
->ctrl
->async_event_sqe
,
566 sizeof(struct nvme_command
), DMA_TO_DEVICE
);
569 nvme_rdma_destroy_queue_ib(queue
);
570 rdma_destroy_id(queue
->cm_id
);
573 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl
*ctrl
)
577 for (i
= 1; i
< ctrl
->ctrl
.queue_count
; i
++)
578 nvme_rdma_free_queue(&ctrl
->queues
[i
]);
581 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl
*ctrl
)
585 for (i
= 1; i
< ctrl
->ctrl
.queue_count
; i
++)
586 nvme_rdma_stop_queue(&ctrl
->queues
[i
]);
589 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl
*ctrl
, int idx
)
594 ret
= nvmf_connect_io_queue(&ctrl
->ctrl
, idx
);
596 ret
= nvmf_connect_admin_queue(&ctrl
->ctrl
);
599 set_bit(NVME_RDMA_Q_LIVE
, &ctrl
->queues
[idx
].flags
);
601 dev_info(ctrl
->ctrl
.device
,
602 "failed to connect queue: %d ret=%d\n", idx
, ret
);
606 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl
*ctrl
)
610 for (i
= 1; i
< ctrl
->ctrl
.queue_count
; i
++) {
611 ret
= nvme_rdma_start_queue(ctrl
, i
);
613 goto out_stop_queues
;
619 for (i
--; i
>= 1; i
--)
620 nvme_rdma_stop_queue(&ctrl
->queues
[i
]);
624 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl
*ctrl
)
626 struct nvmf_ctrl_options
*opts
= ctrl
->ctrl
.opts
;
627 struct ib_device
*ibdev
= ctrl
->device
->dev
;
628 unsigned int nr_io_queues
;
631 nr_io_queues
= min(opts
->nr_io_queues
, num_online_cpus());
634 * we map queues according to the device irq vectors for
635 * optimal locality so we don't need more queues than
636 * completion vectors.
638 nr_io_queues
= min_t(unsigned int, nr_io_queues
,
639 ibdev
->num_comp_vectors
);
641 ret
= nvme_set_queue_count(&ctrl
->ctrl
, &nr_io_queues
);
645 ctrl
->ctrl
.queue_count
= nr_io_queues
+ 1;
646 if (ctrl
->ctrl
.queue_count
< 2)
649 dev_info(ctrl
->ctrl
.device
,
650 "creating %d I/O queues.\n", nr_io_queues
);
652 for (i
= 1; i
< ctrl
->ctrl
.queue_count
; i
++) {
653 ret
= nvme_rdma_alloc_queue(ctrl
, i
,
654 ctrl
->ctrl
.sqsize
+ 1);
656 goto out_free_queues
;
662 for (i
--; i
>= 1; i
--)
663 nvme_rdma_free_queue(&ctrl
->queues
[i
]);
668 static void nvme_rdma_free_tagset(struct nvme_ctrl
*nctrl
,
669 struct blk_mq_tag_set
*set
)
671 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(nctrl
);
673 blk_mq_free_tag_set(set
);
674 nvme_rdma_dev_put(ctrl
->device
);
677 static struct blk_mq_tag_set
*nvme_rdma_alloc_tagset(struct nvme_ctrl
*nctrl
,
680 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(nctrl
);
681 struct blk_mq_tag_set
*set
;
685 set
= &ctrl
->admin_tag_set
;
686 memset(set
, 0, sizeof(*set
));
687 set
->ops
= &nvme_rdma_admin_mq_ops
;
688 set
->queue_depth
= NVME_AQ_MQ_TAG_DEPTH
;
689 set
->reserved_tags
= 2; /* connect + keep-alive */
690 set
->numa_node
= NUMA_NO_NODE
;
691 set
->cmd_size
= sizeof(struct nvme_rdma_request
) +
692 SG_CHUNK_SIZE
* sizeof(struct scatterlist
);
693 set
->driver_data
= ctrl
;
694 set
->nr_hw_queues
= 1;
695 set
->timeout
= ADMIN_TIMEOUT
;
696 set
->flags
= BLK_MQ_F_NO_SCHED
;
698 set
= &ctrl
->tag_set
;
699 memset(set
, 0, sizeof(*set
));
700 set
->ops
= &nvme_rdma_mq_ops
;
701 set
->queue_depth
= nctrl
->opts
->queue_size
;
702 set
->reserved_tags
= 1; /* fabric connect */
703 set
->numa_node
= NUMA_NO_NODE
;
704 set
->flags
= BLK_MQ_F_SHOULD_MERGE
;
705 set
->cmd_size
= sizeof(struct nvme_rdma_request
) +
706 SG_CHUNK_SIZE
* sizeof(struct scatterlist
);
707 set
->driver_data
= ctrl
;
708 set
->nr_hw_queues
= nctrl
->queue_count
- 1;
709 set
->timeout
= NVME_IO_TIMEOUT
;
712 ret
= blk_mq_alloc_tag_set(set
);
717 * We need a reference on the device as long as the tag_set is alive,
718 * as the MRs in the request structures need a valid ib_device.
720 ret
= nvme_rdma_dev_get(ctrl
->device
);
723 goto out_free_tagset
;
729 blk_mq_free_tag_set(set
);
734 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl
*ctrl
,
737 nvme_rdma_stop_queue(&ctrl
->queues
[0]);
739 blk_cleanup_queue(ctrl
->ctrl
.admin_q
);
740 nvme_rdma_free_tagset(&ctrl
->ctrl
, ctrl
->ctrl
.admin_tagset
);
742 nvme_rdma_free_queue(&ctrl
->queues
[0]);
745 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl
*ctrl
,
750 error
= nvme_rdma_alloc_queue(ctrl
, 0, NVME_AQ_DEPTH
);
754 ctrl
->device
= ctrl
->queues
[0].device
;
756 ctrl
->max_fr_pages
= nvme_rdma_get_max_fr_pages(ctrl
->device
->dev
);
759 ctrl
->ctrl
.admin_tagset
= nvme_rdma_alloc_tagset(&ctrl
->ctrl
, true);
760 if (IS_ERR(ctrl
->ctrl
.admin_tagset
)) {
761 error
= PTR_ERR(ctrl
->ctrl
.admin_tagset
);
765 ctrl
->ctrl
.admin_q
= blk_mq_init_queue(&ctrl
->admin_tag_set
);
766 if (IS_ERR(ctrl
->ctrl
.admin_q
)) {
767 error
= PTR_ERR(ctrl
->ctrl
.admin_q
);
768 goto out_free_tagset
;
772 error
= nvme_rdma_start_queue(ctrl
, 0);
774 goto out_cleanup_queue
;
776 error
= ctrl
->ctrl
.ops
->reg_read64(&ctrl
->ctrl
, NVME_REG_CAP
,
779 dev_err(ctrl
->ctrl
.device
,
780 "prop_get NVME_REG_CAP failed\n");
781 goto out_cleanup_queue
;
785 min_t(int, NVME_CAP_MQES(ctrl
->ctrl
.cap
), ctrl
->ctrl
.sqsize
);
787 error
= nvme_enable_ctrl(&ctrl
->ctrl
, ctrl
->ctrl
.cap
);
789 goto out_cleanup_queue
;
791 ctrl
->ctrl
.max_hw_sectors
=
792 (ctrl
->max_fr_pages
- 1) << (ilog2(SZ_4K
) - 9);
794 error
= nvme_init_identify(&ctrl
->ctrl
);
796 goto out_cleanup_queue
;
798 error
= nvme_rdma_alloc_qe(ctrl
->queues
[0].device
->dev
,
799 &ctrl
->async_event_sqe
, sizeof(struct nvme_command
),
802 goto out_cleanup_queue
;
808 blk_cleanup_queue(ctrl
->ctrl
.admin_q
);
811 nvme_rdma_free_tagset(&ctrl
->ctrl
, ctrl
->ctrl
.admin_tagset
);
813 nvme_rdma_free_queue(&ctrl
->queues
[0]);
817 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl
*ctrl
,
820 nvme_rdma_stop_io_queues(ctrl
);
822 blk_cleanup_queue(ctrl
->ctrl
.connect_q
);
823 nvme_rdma_free_tagset(&ctrl
->ctrl
, ctrl
->ctrl
.tagset
);
825 nvme_rdma_free_io_queues(ctrl
);
828 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl
*ctrl
, bool new)
832 ret
= nvme_rdma_alloc_io_queues(ctrl
);
837 ctrl
->ctrl
.tagset
= nvme_rdma_alloc_tagset(&ctrl
->ctrl
, false);
838 if (IS_ERR(ctrl
->ctrl
.tagset
)) {
839 ret
= PTR_ERR(ctrl
->ctrl
.tagset
);
840 goto out_free_io_queues
;
843 ctrl
->ctrl
.connect_q
= blk_mq_init_queue(&ctrl
->tag_set
);
844 if (IS_ERR(ctrl
->ctrl
.connect_q
)) {
845 ret
= PTR_ERR(ctrl
->ctrl
.connect_q
);
846 goto out_free_tag_set
;
849 blk_mq_update_nr_hw_queues(&ctrl
->tag_set
,
850 ctrl
->ctrl
.queue_count
- 1);
853 ret
= nvme_rdma_start_io_queues(ctrl
);
855 goto out_cleanup_connect_q
;
859 out_cleanup_connect_q
:
861 blk_cleanup_queue(ctrl
->ctrl
.connect_q
);
864 nvme_rdma_free_tagset(&ctrl
->ctrl
, ctrl
->ctrl
.tagset
);
866 nvme_rdma_free_io_queues(ctrl
);
870 static void nvme_rdma_stop_ctrl(struct nvme_ctrl
*nctrl
)
872 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(nctrl
);
874 cancel_work_sync(&ctrl
->err_work
);
875 cancel_delayed_work_sync(&ctrl
->reconnect_work
);
878 static void nvme_rdma_free_ctrl(struct nvme_ctrl
*nctrl
)
880 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(nctrl
);
882 if (list_empty(&ctrl
->list
))
885 mutex_lock(&nvme_rdma_ctrl_mutex
);
886 list_del(&ctrl
->list
);
887 mutex_unlock(&nvme_rdma_ctrl_mutex
);
890 nvmf_free_options(nctrl
->opts
);
895 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl
*ctrl
)
897 /* If we are resetting/deleting then do nothing */
898 if (ctrl
->ctrl
.state
!= NVME_CTRL_CONNECTING
) {
899 WARN_ON_ONCE(ctrl
->ctrl
.state
== NVME_CTRL_NEW
||
900 ctrl
->ctrl
.state
== NVME_CTRL_LIVE
);
904 if (nvmf_should_reconnect(&ctrl
->ctrl
)) {
905 dev_info(ctrl
->ctrl
.device
, "Reconnecting in %d seconds...\n",
906 ctrl
->ctrl
.opts
->reconnect_delay
);
907 queue_delayed_work(nvme_wq
, &ctrl
->reconnect_work
,
908 ctrl
->ctrl
.opts
->reconnect_delay
* HZ
);
910 nvme_delete_ctrl(&ctrl
->ctrl
);
914 static void nvme_rdma_reconnect_ctrl_work(struct work_struct
*work
)
916 struct nvme_rdma_ctrl
*ctrl
= container_of(to_delayed_work(work
),
917 struct nvme_rdma_ctrl
, reconnect_work
);
921 ++ctrl
->ctrl
.nr_reconnects
;
923 ret
= nvme_rdma_configure_admin_queue(ctrl
, false);
927 if (ctrl
->ctrl
.queue_count
> 1) {
928 ret
= nvme_rdma_configure_io_queues(ctrl
, false);
933 changed
= nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_LIVE
);
935 /* state change failure is ok if we're in DELETING state */
936 WARN_ON_ONCE(ctrl
->ctrl
.state
!= NVME_CTRL_DELETING
);
940 nvme_start_ctrl(&ctrl
->ctrl
);
942 dev_info(ctrl
->ctrl
.device
, "Successfully reconnected (%d attempts)\n",
943 ctrl
->ctrl
.nr_reconnects
);
945 ctrl
->ctrl
.nr_reconnects
= 0;
950 nvme_rdma_destroy_admin_queue(ctrl
, false);
952 dev_info(ctrl
->ctrl
.device
, "Failed reconnect attempt %d\n",
953 ctrl
->ctrl
.nr_reconnects
);
954 nvme_rdma_reconnect_or_remove(ctrl
);
957 static void nvme_rdma_error_recovery_work(struct work_struct
*work
)
959 struct nvme_rdma_ctrl
*ctrl
= container_of(work
,
960 struct nvme_rdma_ctrl
, err_work
);
962 nvme_stop_keep_alive(&ctrl
->ctrl
);
964 if (ctrl
->ctrl
.queue_count
> 1) {
965 nvme_stop_queues(&ctrl
->ctrl
);
966 blk_mq_tagset_busy_iter(&ctrl
->tag_set
,
967 nvme_cancel_request
, &ctrl
->ctrl
);
968 nvme_rdma_destroy_io_queues(ctrl
, false);
971 blk_mq_quiesce_queue(ctrl
->ctrl
.admin_q
);
972 blk_mq_tagset_busy_iter(&ctrl
->admin_tag_set
,
973 nvme_cancel_request
, &ctrl
->ctrl
);
974 nvme_rdma_destroy_admin_queue(ctrl
, false);
977 * queues are not a live anymore, so restart the queues to fail fast
980 blk_mq_unquiesce_queue(ctrl
->ctrl
.admin_q
);
981 nvme_start_queues(&ctrl
->ctrl
);
983 if (!nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_CONNECTING
)) {
984 /* state change failure is ok if we're in DELETING state */
985 WARN_ON_ONCE(ctrl
->ctrl
.state
!= NVME_CTRL_DELETING
);
989 nvme_rdma_reconnect_or_remove(ctrl
);
992 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl
*ctrl
)
994 if (!nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_RESETTING
))
997 queue_work(nvme_wq
, &ctrl
->err_work
);
1000 static void nvme_rdma_wr_error(struct ib_cq
*cq
, struct ib_wc
*wc
,
1003 struct nvme_rdma_queue
*queue
= cq
->cq_context
;
1004 struct nvme_rdma_ctrl
*ctrl
= queue
->ctrl
;
1006 if (ctrl
->ctrl
.state
== NVME_CTRL_LIVE
)
1007 dev_info(ctrl
->ctrl
.device
,
1008 "%s for CQE 0x%p failed with status %s (%d)\n",
1010 ib_wc_status_msg(wc
->status
), wc
->status
);
1011 nvme_rdma_error_recovery(ctrl
);
1014 static void nvme_rdma_memreg_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1016 if (unlikely(wc
->status
!= IB_WC_SUCCESS
))
1017 nvme_rdma_wr_error(cq
, wc
, "MEMREG");
1020 static void nvme_rdma_inv_rkey_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1022 struct nvme_rdma_request
*req
=
1023 container_of(wc
->wr_cqe
, struct nvme_rdma_request
, reg_cqe
);
1024 struct request
*rq
= blk_mq_rq_from_pdu(req
);
1026 if (unlikely(wc
->status
!= IB_WC_SUCCESS
)) {
1027 nvme_rdma_wr_error(cq
, wc
, "LOCAL_INV");
1031 if (refcount_dec_and_test(&req
->ref
))
1032 nvme_end_request(rq
, req
->status
, req
->result
);
1036 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue
*queue
,
1037 struct nvme_rdma_request
*req
)
1039 struct ib_send_wr
*bad_wr
;
1040 struct ib_send_wr wr
= {
1041 .opcode
= IB_WR_LOCAL_INV
,
1044 .send_flags
= IB_SEND_SIGNALED
,
1045 .ex
.invalidate_rkey
= req
->mr
->rkey
,
1048 req
->reg_cqe
.done
= nvme_rdma_inv_rkey_done
;
1049 wr
.wr_cqe
= &req
->reg_cqe
;
1051 return ib_post_send(queue
->qp
, &wr
, &bad_wr
);
1054 static void nvme_rdma_unmap_data(struct nvme_rdma_queue
*queue
,
1057 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1058 struct nvme_rdma_device
*dev
= queue
->device
;
1059 struct ib_device
*ibdev
= dev
->dev
;
1061 if (!blk_rq_payload_bytes(rq
))
1065 ib_mr_pool_put(queue
->qp
, &queue
->qp
->rdma_mrs
, req
->mr
);
1069 ib_dma_unmap_sg(ibdev
, req
->sg_table
.sgl
,
1070 req
->nents
, rq_data_dir(rq
) ==
1071 WRITE
? DMA_TO_DEVICE
: DMA_FROM_DEVICE
);
1073 nvme_cleanup_cmd(rq
);
1074 sg_free_table_chained(&req
->sg_table
, true);
1077 static int nvme_rdma_set_sg_null(struct nvme_command
*c
)
1079 struct nvme_keyed_sgl_desc
*sg
= &c
->common
.dptr
.ksgl
;
1082 put_unaligned_le24(0, sg
->length
);
1083 put_unaligned_le32(0, sg
->key
);
1084 sg
->type
= NVME_KEY_SGL_FMT_DATA_DESC
<< 4;
1088 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue
*queue
,
1089 struct nvme_rdma_request
*req
, struct nvme_command
*c
)
1091 struct nvme_sgl_desc
*sg
= &c
->common
.dptr
.sgl
;
1093 req
->sge
[1].addr
= sg_dma_address(req
->sg_table
.sgl
);
1094 req
->sge
[1].length
= sg_dma_len(req
->sg_table
.sgl
);
1095 req
->sge
[1].lkey
= queue
->device
->pd
->local_dma_lkey
;
1097 sg
->addr
= cpu_to_le64(queue
->ctrl
->ctrl
.icdoff
);
1098 sg
->length
= cpu_to_le32(sg_dma_len(req
->sg_table
.sgl
));
1099 sg
->type
= (NVME_SGL_FMT_DATA_DESC
<< 4) | NVME_SGL_FMT_OFFSET
;
1105 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue
*queue
,
1106 struct nvme_rdma_request
*req
, struct nvme_command
*c
)
1108 struct nvme_keyed_sgl_desc
*sg
= &c
->common
.dptr
.ksgl
;
1110 sg
->addr
= cpu_to_le64(sg_dma_address(req
->sg_table
.sgl
));
1111 put_unaligned_le24(sg_dma_len(req
->sg_table
.sgl
), sg
->length
);
1112 put_unaligned_le32(queue
->device
->pd
->unsafe_global_rkey
, sg
->key
);
1113 sg
->type
= NVME_KEY_SGL_FMT_DATA_DESC
<< 4;
1117 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue
*queue
,
1118 struct nvme_rdma_request
*req
, struct nvme_command
*c
,
1121 struct nvme_keyed_sgl_desc
*sg
= &c
->common
.dptr
.ksgl
;
1124 req
->mr
= ib_mr_pool_get(queue
->qp
, &queue
->qp
->rdma_mrs
);
1125 if (WARN_ON_ONCE(!req
->mr
))
1129 * Align the MR to a 4K page size to match the ctrl page size and
1130 * the block virtual boundary.
1132 nr
= ib_map_mr_sg(req
->mr
, req
->sg_table
.sgl
, count
, NULL
, SZ_4K
);
1133 if (unlikely(nr
< count
)) {
1134 ib_mr_pool_put(queue
->qp
, &queue
->qp
->rdma_mrs
, req
->mr
);
1141 ib_update_fast_reg_key(req
->mr
, ib_inc_rkey(req
->mr
->rkey
));
1143 req
->reg_cqe
.done
= nvme_rdma_memreg_done
;
1144 memset(&req
->reg_wr
, 0, sizeof(req
->reg_wr
));
1145 req
->reg_wr
.wr
.opcode
= IB_WR_REG_MR
;
1146 req
->reg_wr
.wr
.wr_cqe
= &req
->reg_cqe
;
1147 req
->reg_wr
.wr
.num_sge
= 0;
1148 req
->reg_wr
.mr
= req
->mr
;
1149 req
->reg_wr
.key
= req
->mr
->rkey
;
1150 req
->reg_wr
.access
= IB_ACCESS_LOCAL_WRITE
|
1151 IB_ACCESS_REMOTE_READ
|
1152 IB_ACCESS_REMOTE_WRITE
;
1154 sg
->addr
= cpu_to_le64(req
->mr
->iova
);
1155 put_unaligned_le24(req
->mr
->length
, sg
->length
);
1156 put_unaligned_le32(req
->mr
->rkey
, sg
->key
);
1157 sg
->type
= (NVME_KEY_SGL_FMT_DATA_DESC
<< 4) |
1158 NVME_SGL_FMT_INVALIDATE
;
1163 static int nvme_rdma_map_data(struct nvme_rdma_queue
*queue
,
1164 struct request
*rq
, struct nvme_command
*c
)
1166 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1167 struct nvme_rdma_device
*dev
= queue
->device
;
1168 struct ib_device
*ibdev
= dev
->dev
;
1172 refcount_set(&req
->ref
, 2); /* send and recv completions */
1174 c
->common
.flags
|= NVME_CMD_SGL_METABUF
;
1176 if (!blk_rq_payload_bytes(rq
))
1177 return nvme_rdma_set_sg_null(c
);
1179 req
->sg_table
.sgl
= req
->first_sgl
;
1180 ret
= sg_alloc_table_chained(&req
->sg_table
,
1181 blk_rq_nr_phys_segments(rq
), req
->sg_table
.sgl
);
1185 req
->nents
= blk_rq_map_sg(rq
->q
, rq
, req
->sg_table
.sgl
);
1187 count
= ib_dma_map_sg(ibdev
, req
->sg_table
.sgl
, req
->nents
,
1188 rq_data_dir(rq
) == WRITE
? DMA_TO_DEVICE
: DMA_FROM_DEVICE
);
1189 if (unlikely(count
<= 0)) {
1190 sg_free_table_chained(&req
->sg_table
, true);
1195 if (rq_data_dir(rq
) == WRITE
&& nvme_rdma_queue_idx(queue
) &&
1196 blk_rq_payload_bytes(rq
) <=
1197 nvme_rdma_inline_data_size(queue
))
1198 return nvme_rdma_map_sg_inline(queue
, req
, c
);
1200 if (dev
->pd
->flags
& IB_PD_UNSAFE_GLOBAL_RKEY
)
1201 return nvme_rdma_map_sg_single(queue
, req
, c
);
1204 return nvme_rdma_map_sg_fr(queue
, req
, c
, count
);
1207 static void nvme_rdma_send_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1209 struct nvme_rdma_qe
*qe
=
1210 container_of(wc
->wr_cqe
, struct nvme_rdma_qe
, cqe
);
1211 struct nvme_rdma_request
*req
=
1212 container_of(qe
, struct nvme_rdma_request
, sqe
);
1213 struct request
*rq
= blk_mq_rq_from_pdu(req
);
1215 if (unlikely(wc
->status
!= IB_WC_SUCCESS
)) {
1216 nvme_rdma_wr_error(cq
, wc
, "SEND");
1220 if (refcount_dec_and_test(&req
->ref
))
1221 nvme_end_request(rq
, req
->status
, req
->result
);
1224 static int nvme_rdma_post_send(struct nvme_rdma_queue
*queue
,
1225 struct nvme_rdma_qe
*qe
, struct ib_sge
*sge
, u32 num_sge
,
1226 struct ib_send_wr
*first
)
1228 struct ib_send_wr wr
, *bad_wr
;
1231 sge
->addr
= qe
->dma
;
1232 sge
->length
= sizeof(struct nvme_command
),
1233 sge
->lkey
= queue
->device
->pd
->local_dma_lkey
;
1236 wr
.wr_cqe
= &qe
->cqe
;
1238 wr
.num_sge
= num_sge
;
1239 wr
.opcode
= IB_WR_SEND
;
1240 wr
.send_flags
= IB_SEND_SIGNALED
;
1247 ret
= ib_post_send(queue
->qp
, first
, &bad_wr
);
1248 if (unlikely(ret
)) {
1249 dev_err(queue
->ctrl
->ctrl
.device
,
1250 "%s failed with error code %d\n", __func__
, ret
);
1255 static int nvme_rdma_post_recv(struct nvme_rdma_queue
*queue
,
1256 struct nvme_rdma_qe
*qe
)
1258 struct ib_recv_wr wr
, *bad_wr
;
1262 list
.addr
= qe
->dma
;
1263 list
.length
= sizeof(struct nvme_completion
);
1264 list
.lkey
= queue
->device
->pd
->local_dma_lkey
;
1266 qe
->cqe
.done
= nvme_rdma_recv_done
;
1269 wr
.wr_cqe
= &qe
->cqe
;
1273 ret
= ib_post_recv(queue
->qp
, &wr
, &bad_wr
);
1274 if (unlikely(ret
)) {
1275 dev_err(queue
->ctrl
->ctrl
.device
,
1276 "%s failed with error code %d\n", __func__
, ret
);
1281 static struct blk_mq_tags
*nvme_rdma_tagset(struct nvme_rdma_queue
*queue
)
1283 u32 queue_idx
= nvme_rdma_queue_idx(queue
);
1286 return queue
->ctrl
->admin_tag_set
.tags
[queue_idx
];
1287 return queue
->ctrl
->tag_set
.tags
[queue_idx
- 1];
1290 static void nvme_rdma_async_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1292 if (unlikely(wc
->status
!= IB_WC_SUCCESS
))
1293 nvme_rdma_wr_error(cq
, wc
, "ASYNC");
1296 static void nvme_rdma_submit_async_event(struct nvme_ctrl
*arg
)
1298 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(arg
);
1299 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[0];
1300 struct ib_device
*dev
= queue
->device
->dev
;
1301 struct nvme_rdma_qe
*sqe
= &ctrl
->async_event_sqe
;
1302 struct nvme_command
*cmd
= sqe
->data
;
1306 ib_dma_sync_single_for_cpu(dev
, sqe
->dma
, sizeof(*cmd
), DMA_TO_DEVICE
);
1308 memset(cmd
, 0, sizeof(*cmd
));
1309 cmd
->common
.opcode
= nvme_admin_async_event
;
1310 cmd
->common
.command_id
= NVME_AQ_BLK_MQ_DEPTH
;
1311 cmd
->common
.flags
|= NVME_CMD_SGL_METABUF
;
1312 nvme_rdma_set_sg_null(cmd
);
1314 sqe
->cqe
.done
= nvme_rdma_async_done
;
1316 ib_dma_sync_single_for_device(dev
, sqe
->dma
, sizeof(*cmd
),
1319 ret
= nvme_rdma_post_send(queue
, sqe
, &sge
, 1, NULL
);
1323 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue
*queue
,
1324 struct nvme_completion
*cqe
, struct ib_wc
*wc
, int tag
)
1327 struct nvme_rdma_request
*req
;
1330 rq
= blk_mq_tag_to_rq(nvme_rdma_tagset(queue
), cqe
->command_id
);
1332 dev_err(queue
->ctrl
->ctrl
.device
,
1333 "tag 0x%x on QP %#x not found\n",
1334 cqe
->command_id
, queue
->qp
->qp_num
);
1335 nvme_rdma_error_recovery(queue
->ctrl
);
1338 req
= blk_mq_rq_to_pdu(rq
);
1340 req
->status
= cqe
->status
;
1341 req
->result
= cqe
->result
;
1343 if (wc
->wc_flags
& IB_WC_WITH_INVALIDATE
) {
1344 if (unlikely(wc
->ex
.invalidate_rkey
!= req
->mr
->rkey
)) {
1345 dev_err(queue
->ctrl
->ctrl
.device
,
1346 "Bogus remote invalidation for rkey %#x\n",
1348 nvme_rdma_error_recovery(queue
->ctrl
);
1350 } else if (req
->mr
) {
1351 ret
= nvme_rdma_inv_rkey(queue
, req
);
1352 if (unlikely(ret
< 0)) {
1353 dev_err(queue
->ctrl
->ctrl
.device
,
1354 "Queueing INV WR for rkey %#x failed (%d)\n",
1355 req
->mr
->rkey
, ret
);
1356 nvme_rdma_error_recovery(queue
->ctrl
);
1358 /* the local invalidation completion will end the request */
1362 if (refcount_dec_and_test(&req
->ref
)) {
1365 nvme_end_request(rq
, req
->status
, req
->result
);
1371 static int __nvme_rdma_recv_done(struct ib_cq
*cq
, struct ib_wc
*wc
, int tag
)
1373 struct nvme_rdma_qe
*qe
=
1374 container_of(wc
->wr_cqe
, struct nvme_rdma_qe
, cqe
);
1375 struct nvme_rdma_queue
*queue
= cq
->cq_context
;
1376 struct ib_device
*ibdev
= queue
->device
->dev
;
1377 struct nvme_completion
*cqe
= qe
->data
;
1378 const size_t len
= sizeof(struct nvme_completion
);
1381 if (unlikely(wc
->status
!= IB_WC_SUCCESS
)) {
1382 nvme_rdma_wr_error(cq
, wc
, "RECV");
1386 ib_dma_sync_single_for_cpu(ibdev
, qe
->dma
, len
, DMA_FROM_DEVICE
);
1388 * AEN requests are special as they don't time out and can
1389 * survive any kind of queue freeze and often don't respond to
1390 * aborts. We don't even bother to allocate a struct request
1391 * for them but rather special case them here.
1393 if (unlikely(nvme_rdma_queue_idx(queue
) == 0 &&
1394 cqe
->command_id
>= NVME_AQ_BLK_MQ_DEPTH
))
1395 nvme_complete_async_event(&queue
->ctrl
->ctrl
, cqe
->status
,
1398 ret
= nvme_rdma_process_nvme_rsp(queue
, cqe
, wc
, tag
);
1399 ib_dma_sync_single_for_device(ibdev
, qe
->dma
, len
, DMA_FROM_DEVICE
);
1401 nvme_rdma_post_recv(queue
, qe
);
1405 static void nvme_rdma_recv_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1407 __nvme_rdma_recv_done(cq
, wc
, -1);
1410 static int nvme_rdma_conn_established(struct nvme_rdma_queue
*queue
)
1414 for (i
= 0; i
< queue
->queue_size
; i
++) {
1415 ret
= nvme_rdma_post_recv(queue
, &queue
->rsp_ring
[i
]);
1417 goto out_destroy_queue_ib
;
1422 out_destroy_queue_ib
:
1423 nvme_rdma_destroy_queue_ib(queue
);
1427 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue
*queue
,
1428 struct rdma_cm_event
*ev
)
1430 struct rdma_cm_id
*cm_id
= queue
->cm_id
;
1431 int status
= ev
->status
;
1432 const char *rej_msg
;
1433 const struct nvme_rdma_cm_rej
*rej_data
;
1436 rej_msg
= rdma_reject_msg(cm_id
, status
);
1437 rej_data
= rdma_consumer_reject_data(cm_id
, ev
, &rej_data_len
);
1439 if (rej_data
&& rej_data_len
>= sizeof(u16
)) {
1440 u16 sts
= le16_to_cpu(rej_data
->sts
);
1442 dev_err(queue
->ctrl
->ctrl
.device
,
1443 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1444 status
, rej_msg
, sts
, nvme_rdma_cm_msg(sts
));
1446 dev_err(queue
->ctrl
->ctrl
.device
,
1447 "Connect rejected: status %d (%s).\n", status
, rej_msg
);
1453 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue
*queue
)
1457 ret
= nvme_rdma_create_queue_ib(queue
);
1461 ret
= rdma_resolve_route(queue
->cm_id
, NVME_RDMA_CONNECT_TIMEOUT_MS
);
1463 dev_err(queue
->ctrl
->ctrl
.device
,
1464 "rdma_resolve_route failed (%d).\n",
1466 goto out_destroy_queue
;
1472 nvme_rdma_destroy_queue_ib(queue
);
1476 static int nvme_rdma_route_resolved(struct nvme_rdma_queue
*queue
)
1478 struct nvme_rdma_ctrl
*ctrl
= queue
->ctrl
;
1479 struct rdma_conn_param param
= { };
1480 struct nvme_rdma_cm_req priv
= { };
1483 param
.qp_num
= queue
->qp
->qp_num
;
1484 param
.flow_control
= 1;
1486 param
.responder_resources
= queue
->device
->dev
->attrs
.max_qp_rd_atom
;
1487 /* maximum retry count */
1488 param
.retry_count
= 7;
1489 param
.rnr_retry_count
= 7;
1490 param
.private_data
= &priv
;
1491 param
.private_data_len
= sizeof(priv
);
1493 priv
.recfmt
= cpu_to_le16(NVME_RDMA_CM_FMT_1_0
);
1494 priv
.qid
= cpu_to_le16(nvme_rdma_queue_idx(queue
));
1496 * set the admin queue depth to the minimum size
1497 * specified by the Fabrics standard.
1499 if (priv
.qid
== 0) {
1500 priv
.hrqsize
= cpu_to_le16(NVME_AQ_DEPTH
);
1501 priv
.hsqsize
= cpu_to_le16(NVME_AQ_DEPTH
- 1);
1504 * current interpretation of the fabrics spec
1505 * is at minimum you make hrqsize sqsize+1, or a
1506 * 1's based representation of sqsize.
1508 priv
.hrqsize
= cpu_to_le16(queue
->queue_size
);
1509 priv
.hsqsize
= cpu_to_le16(queue
->ctrl
->ctrl
.sqsize
);
1512 ret
= rdma_connect(queue
->cm_id
, ¶m
);
1514 dev_err(ctrl
->ctrl
.device
,
1515 "rdma_connect failed (%d).\n", ret
);
1516 goto out_destroy_queue_ib
;
1521 out_destroy_queue_ib
:
1522 nvme_rdma_destroy_queue_ib(queue
);
1526 static int nvme_rdma_cm_handler(struct rdma_cm_id
*cm_id
,
1527 struct rdma_cm_event
*ev
)
1529 struct nvme_rdma_queue
*queue
= cm_id
->context
;
1532 dev_dbg(queue
->ctrl
->ctrl
.device
, "%s (%d): status %d id %p\n",
1533 rdma_event_msg(ev
->event
), ev
->event
,
1536 switch (ev
->event
) {
1537 case RDMA_CM_EVENT_ADDR_RESOLVED
:
1538 cm_error
= nvme_rdma_addr_resolved(queue
);
1540 case RDMA_CM_EVENT_ROUTE_RESOLVED
:
1541 cm_error
= nvme_rdma_route_resolved(queue
);
1543 case RDMA_CM_EVENT_ESTABLISHED
:
1544 queue
->cm_error
= nvme_rdma_conn_established(queue
);
1545 /* complete cm_done regardless of success/failure */
1546 complete(&queue
->cm_done
);
1548 case RDMA_CM_EVENT_REJECTED
:
1549 nvme_rdma_destroy_queue_ib(queue
);
1550 cm_error
= nvme_rdma_conn_rejected(queue
, ev
);
1552 case RDMA_CM_EVENT_ROUTE_ERROR
:
1553 case RDMA_CM_EVENT_CONNECT_ERROR
:
1554 case RDMA_CM_EVENT_UNREACHABLE
:
1555 nvme_rdma_destroy_queue_ib(queue
);
1556 case RDMA_CM_EVENT_ADDR_ERROR
:
1557 dev_dbg(queue
->ctrl
->ctrl
.device
,
1558 "CM error event %d\n", ev
->event
);
1559 cm_error
= -ECONNRESET
;
1561 case RDMA_CM_EVENT_DISCONNECTED
:
1562 case RDMA_CM_EVENT_ADDR_CHANGE
:
1563 case RDMA_CM_EVENT_TIMEWAIT_EXIT
:
1564 dev_dbg(queue
->ctrl
->ctrl
.device
,
1565 "disconnect received - connection closed\n");
1566 nvme_rdma_error_recovery(queue
->ctrl
);
1568 case RDMA_CM_EVENT_DEVICE_REMOVAL
:
1569 /* device removal is handled via the ib_client API */
1572 dev_err(queue
->ctrl
->ctrl
.device
,
1573 "Unexpected RDMA CM event (%d)\n", ev
->event
);
1574 nvme_rdma_error_recovery(queue
->ctrl
);
1579 queue
->cm_error
= cm_error
;
1580 complete(&queue
->cm_done
);
1586 static enum blk_eh_timer_return
1587 nvme_rdma_timeout(struct request
*rq
, bool reserved
)
1589 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1591 dev_warn(req
->queue
->ctrl
->ctrl
.device
,
1592 "I/O %d QID %d timeout, reset controller\n",
1593 rq
->tag
, nvme_rdma_queue_idx(req
->queue
));
1595 /* queue error recovery */
1596 nvme_rdma_error_recovery(req
->queue
->ctrl
);
1598 /* fail with DNR on cmd timeout */
1599 nvme_req(rq
)->status
= NVME_SC_ABORT_REQ
| NVME_SC_DNR
;
1604 static blk_status_t
nvme_rdma_queue_rq(struct blk_mq_hw_ctx
*hctx
,
1605 const struct blk_mq_queue_data
*bd
)
1607 struct nvme_ns
*ns
= hctx
->queue
->queuedata
;
1608 struct nvme_rdma_queue
*queue
= hctx
->driver_data
;
1609 struct request
*rq
= bd
->rq
;
1610 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1611 struct nvme_rdma_qe
*sqe
= &req
->sqe
;
1612 struct nvme_command
*c
= sqe
->data
;
1613 struct ib_device
*dev
;
1617 WARN_ON_ONCE(rq
->tag
< 0);
1619 ret
= nvmf_check_if_ready(&queue
->ctrl
->ctrl
, rq
,
1620 test_bit(NVME_RDMA_Q_LIVE
, &queue
->flags
), true);
1624 dev
= queue
->device
->dev
;
1625 ib_dma_sync_single_for_cpu(dev
, sqe
->dma
,
1626 sizeof(struct nvme_command
), DMA_TO_DEVICE
);
1628 ret
= nvme_setup_cmd(ns
, rq
, c
);
1632 blk_mq_start_request(rq
);
1634 err
= nvme_rdma_map_data(queue
, rq
, c
);
1635 if (unlikely(err
< 0)) {
1636 dev_err(queue
->ctrl
->ctrl
.device
,
1637 "Failed to map data (%d)\n", err
);
1638 nvme_cleanup_cmd(rq
);
1642 sqe
->cqe
.done
= nvme_rdma_send_done
;
1644 ib_dma_sync_single_for_device(dev
, sqe
->dma
,
1645 sizeof(struct nvme_command
), DMA_TO_DEVICE
);
1647 err
= nvme_rdma_post_send(queue
, sqe
, req
->sge
, req
->num_sge
,
1648 req
->mr
? &req
->reg_wr
.wr
: NULL
);
1649 if (unlikely(err
)) {
1650 nvme_rdma_unmap_data(queue
, rq
);
1656 if (err
== -ENOMEM
|| err
== -EAGAIN
)
1657 return BLK_STS_RESOURCE
;
1658 return BLK_STS_IOERR
;
1661 static int nvme_rdma_poll(struct blk_mq_hw_ctx
*hctx
, unsigned int tag
)
1663 struct nvme_rdma_queue
*queue
= hctx
->driver_data
;
1664 struct ib_cq
*cq
= queue
->ib_cq
;
1668 while (ib_poll_cq(cq
, 1, &wc
) > 0) {
1669 struct ib_cqe
*cqe
= wc
.wr_cqe
;
1672 if (cqe
->done
== nvme_rdma_recv_done
)
1673 found
|= __nvme_rdma_recv_done(cq
, &wc
, tag
);
1682 static void nvme_rdma_complete_rq(struct request
*rq
)
1684 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1686 nvme_rdma_unmap_data(req
->queue
, rq
);
1687 nvme_complete_rq(rq
);
1690 static int nvme_rdma_map_queues(struct blk_mq_tag_set
*set
)
1692 struct nvme_rdma_ctrl
*ctrl
= set
->driver_data
;
1694 return blk_mq_rdma_map_queues(set
, ctrl
->device
->dev
, 0);
1697 static const struct blk_mq_ops nvme_rdma_mq_ops
= {
1698 .queue_rq
= nvme_rdma_queue_rq
,
1699 .complete
= nvme_rdma_complete_rq
,
1700 .init_request
= nvme_rdma_init_request
,
1701 .exit_request
= nvme_rdma_exit_request
,
1702 .init_hctx
= nvme_rdma_init_hctx
,
1703 .poll
= nvme_rdma_poll
,
1704 .timeout
= nvme_rdma_timeout
,
1705 .map_queues
= nvme_rdma_map_queues
,
1708 static const struct blk_mq_ops nvme_rdma_admin_mq_ops
= {
1709 .queue_rq
= nvme_rdma_queue_rq
,
1710 .complete
= nvme_rdma_complete_rq
,
1711 .init_request
= nvme_rdma_init_request
,
1712 .exit_request
= nvme_rdma_exit_request
,
1713 .init_hctx
= nvme_rdma_init_admin_hctx
,
1714 .timeout
= nvme_rdma_timeout
,
1717 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl
*ctrl
, bool shutdown
)
1719 if (ctrl
->ctrl
.queue_count
> 1) {
1720 nvme_stop_queues(&ctrl
->ctrl
);
1721 blk_mq_tagset_busy_iter(&ctrl
->tag_set
,
1722 nvme_cancel_request
, &ctrl
->ctrl
);
1723 nvme_rdma_destroy_io_queues(ctrl
, shutdown
);
1727 nvme_shutdown_ctrl(&ctrl
->ctrl
);
1729 nvme_disable_ctrl(&ctrl
->ctrl
, ctrl
->ctrl
.cap
);
1731 blk_mq_quiesce_queue(ctrl
->ctrl
.admin_q
);
1732 blk_mq_tagset_busy_iter(&ctrl
->admin_tag_set
,
1733 nvme_cancel_request
, &ctrl
->ctrl
);
1734 blk_mq_unquiesce_queue(ctrl
->ctrl
.admin_q
);
1735 nvme_rdma_destroy_admin_queue(ctrl
, shutdown
);
1738 static void nvme_rdma_delete_ctrl(struct nvme_ctrl
*ctrl
)
1740 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl
), true);
1743 static void nvme_rdma_reset_ctrl_work(struct work_struct
*work
)
1745 struct nvme_rdma_ctrl
*ctrl
=
1746 container_of(work
, struct nvme_rdma_ctrl
, ctrl
.reset_work
);
1750 nvme_stop_ctrl(&ctrl
->ctrl
);
1751 nvme_rdma_shutdown_ctrl(ctrl
, false);
1753 if (!nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_CONNECTING
)) {
1754 /* state change failure should never happen */
1759 ret
= nvme_rdma_configure_admin_queue(ctrl
, false);
1763 if (ctrl
->ctrl
.queue_count
> 1) {
1764 ret
= nvme_rdma_configure_io_queues(ctrl
, false);
1769 changed
= nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_LIVE
);
1771 /* state change failure is ok if we're in DELETING state */
1772 WARN_ON_ONCE(ctrl
->ctrl
.state
!= NVME_CTRL_DELETING
);
1776 nvme_start_ctrl(&ctrl
->ctrl
);
1781 ++ctrl
->ctrl
.nr_reconnects
;
1782 nvme_rdma_reconnect_or_remove(ctrl
);
1785 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops
= {
1787 .module
= THIS_MODULE
,
1788 .flags
= NVME_F_FABRICS
,
1789 .reg_read32
= nvmf_reg_read32
,
1790 .reg_read64
= nvmf_reg_read64
,
1791 .reg_write32
= nvmf_reg_write32
,
1792 .free_ctrl
= nvme_rdma_free_ctrl
,
1793 .submit_async_event
= nvme_rdma_submit_async_event
,
1794 .delete_ctrl
= nvme_rdma_delete_ctrl
,
1795 .get_address
= nvmf_get_address
,
1796 .stop_ctrl
= nvme_rdma_stop_ctrl
,
1800 __nvme_rdma_options_match(struct nvme_rdma_ctrl
*ctrl
,
1801 struct nvmf_ctrl_options
*opts
)
1803 char *stdport
= __stringify(NVME_RDMA_IP_PORT
);
1806 if (!nvmf_ctlr_matches_baseopts(&ctrl
->ctrl
, opts
) ||
1807 strcmp(opts
->traddr
, ctrl
->ctrl
.opts
->traddr
))
1810 if (opts
->mask
& NVMF_OPT_TRSVCID
&&
1811 ctrl
->ctrl
.opts
->mask
& NVMF_OPT_TRSVCID
) {
1812 if (strcmp(opts
->trsvcid
, ctrl
->ctrl
.opts
->trsvcid
))
1814 } else if (opts
->mask
& NVMF_OPT_TRSVCID
) {
1815 if (strcmp(opts
->trsvcid
, stdport
))
1817 } else if (ctrl
->ctrl
.opts
->mask
& NVMF_OPT_TRSVCID
) {
1818 if (strcmp(stdport
, ctrl
->ctrl
.opts
->trsvcid
))
1821 /* else, it's a match as both have stdport. Fall to next checks */
1824 * checking the local address is rough. In most cases, one
1825 * is not specified and the host port is selected by the stack.
1827 * Assume no match if:
1828 * local address is specified and address is not the same
1829 * local address is not specified but remote is, or vice versa
1830 * (admin using specific host_traddr when it matters).
1832 if (opts
->mask
& NVMF_OPT_HOST_TRADDR
&&
1833 ctrl
->ctrl
.opts
->mask
& NVMF_OPT_HOST_TRADDR
) {
1834 if (strcmp(opts
->host_traddr
, ctrl
->ctrl
.opts
->host_traddr
))
1836 } else if (opts
->mask
& NVMF_OPT_HOST_TRADDR
||
1837 ctrl
->ctrl
.opts
->mask
& NVMF_OPT_HOST_TRADDR
)
1840 * if neither controller had an host port specified, assume it's
1841 * a match as everything else matched.
1848 * Fails a connection request if it matches an existing controller
1849 * (association) with the same tuple:
1850 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1852 * if local address is not specified in the request, it will match an
1853 * existing controller with all the other parameters the same and no
1854 * local port address specified as well.
1856 * The ports don't need to be compared as they are intrinsically
1857 * already matched by the port pointers supplied.
1860 nvme_rdma_existing_controller(struct nvmf_ctrl_options
*opts
)
1862 struct nvme_rdma_ctrl
*ctrl
;
1865 mutex_lock(&nvme_rdma_ctrl_mutex
);
1866 list_for_each_entry(ctrl
, &nvme_rdma_ctrl_list
, list
) {
1867 found
= __nvme_rdma_options_match(ctrl
, opts
);
1871 mutex_unlock(&nvme_rdma_ctrl_mutex
);
1876 static struct nvme_ctrl
*nvme_rdma_create_ctrl(struct device
*dev
,
1877 struct nvmf_ctrl_options
*opts
)
1879 struct nvme_rdma_ctrl
*ctrl
;
1884 ctrl
= kzalloc(sizeof(*ctrl
), GFP_KERNEL
);
1886 return ERR_PTR(-ENOMEM
);
1887 ctrl
->ctrl
.opts
= opts
;
1888 INIT_LIST_HEAD(&ctrl
->list
);
1890 if (opts
->mask
& NVMF_OPT_TRSVCID
)
1891 port
= opts
->trsvcid
;
1893 port
= __stringify(NVME_RDMA_IP_PORT
);
1895 ret
= inet_pton_with_scope(&init_net
, AF_UNSPEC
,
1896 opts
->traddr
, port
, &ctrl
->addr
);
1898 pr_err("malformed address passed: %s:%s\n", opts
->traddr
, port
);
1902 if (opts
->mask
& NVMF_OPT_HOST_TRADDR
) {
1903 ret
= inet_pton_with_scope(&init_net
, AF_UNSPEC
,
1904 opts
->host_traddr
, NULL
, &ctrl
->src_addr
);
1906 pr_err("malformed src address passed: %s\n",
1912 if (!opts
->duplicate_connect
&& nvme_rdma_existing_controller(opts
)) {
1917 ret
= nvme_init_ctrl(&ctrl
->ctrl
, dev
, &nvme_rdma_ctrl_ops
,
1918 0 /* no quirks, we're perfect! */);
1922 INIT_DELAYED_WORK(&ctrl
->reconnect_work
,
1923 nvme_rdma_reconnect_ctrl_work
);
1924 INIT_WORK(&ctrl
->err_work
, nvme_rdma_error_recovery_work
);
1925 INIT_WORK(&ctrl
->ctrl
.reset_work
, nvme_rdma_reset_ctrl_work
);
1927 ctrl
->ctrl
.queue_count
= opts
->nr_io_queues
+ 1; /* +1 for admin queue */
1928 ctrl
->ctrl
.sqsize
= opts
->queue_size
- 1;
1929 ctrl
->ctrl
.kato
= opts
->kato
;
1932 ctrl
->queues
= kcalloc(ctrl
->ctrl
.queue_count
, sizeof(*ctrl
->queues
),
1935 goto out_uninit_ctrl
;
1937 changed
= nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_CONNECTING
);
1938 WARN_ON_ONCE(!changed
);
1940 ret
= nvme_rdma_configure_admin_queue(ctrl
, true);
1942 goto out_kfree_queues
;
1944 /* sanity check icdoff */
1945 if (ctrl
->ctrl
.icdoff
) {
1946 dev_err(ctrl
->ctrl
.device
, "icdoff is not supported!\n");
1948 goto out_remove_admin_queue
;
1951 /* sanity check keyed sgls */
1952 if (!(ctrl
->ctrl
.sgls
& (1 << 20))) {
1953 dev_err(ctrl
->ctrl
.device
, "Mandatory keyed sgls are not support\n");
1955 goto out_remove_admin_queue
;
1958 if (opts
->queue_size
> ctrl
->ctrl
.maxcmd
) {
1959 /* warn if maxcmd is lower than queue_size */
1960 dev_warn(ctrl
->ctrl
.device
,
1961 "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1962 opts
->queue_size
, ctrl
->ctrl
.maxcmd
);
1963 opts
->queue_size
= ctrl
->ctrl
.maxcmd
;
1966 if (opts
->queue_size
> ctrl
->ctrl
.sqsize
+ 1) {
1967 /* warn if sqsize is lower than queue_size */
1968 dev_warn(ctrl
->ctrl
.device
,
1969 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1970 opts
->queue_size
, ctrl
->ctrl
.sqsize
+ 1);
1971 opts
->queue_size
= ctrl
->ctrl
.sqsize
+ 1;
1974 if (opts
->nr_io_queues
) {
1975 ret
= nvme_rdma_configure_io_queues(ctrl
, true);
1977 goto out_remove_admin_queue
;
1980 changed
= nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_LIVE
);
1981 WARN_ON_ONCE(!changed
);
1983 dev_info(ctrl
->ctrl
.device
, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1984 ctrl
->ctrl
.opts
->subsysnqn
, &ctrl
->addr
);
1986 nvme_get_ctrl(&ctrl
->ctrl
);
1988 mutex_lock(&nvme_rdma_ctrl_mutex
);
1989 list_add_tail(&ctrl
->list
, &nvme_rdma_ctrl_list
);
1990 mutex_unlock(&nvme_rdma_ctrl_mutex
);
1992 nvme_start_ctrl(&ctrl
->ctrl
);
1996 out_remove_admin_queue
:
1997 nvme_rdma_destroy_admin_queue(ctrl
, true);
1999 kfree(ctrl
->queues
);
2001 nvme_uninit_ctrl(&ctrl
->ctrl
);
2002 nvme_put_ctrl(&ctrl
->ctrl
);
2005 return ERR_PTR(ret
);
2008 return ERR_PTR(ret
);
2011 static struct nvmf_transport_ops nvme_rdma_transport
= {
2013 .module
= THIS_MODULE
,
2014 .required_opts
= NVMF_OPT_TRADDR
,
2015 .allowed_opts
= NVMF_OPT_TRSVCID
| NVMF_OPT_RECONNECT_DELAY
|
2016 NVMF_OPT_HOST_TRADDR
| NVMF_OPT_CTRL_LOSS_TMO
,
2017 .create_ctrl
= nvme_rdma_create_ctrl
,
2020 static void nvme_rdma_remove_one(struct ib_device
*ib_device
, void *client_data
)
2022 struct nvme_rdma_ctrl
*ctrl
;
2023 struct nvme_rdma_device
*ndev
;
2026 mutex_lock(&device_list_mutex
);
2027 list_for_each_entry(ndev
, &device_list
, entry
) {
2028 if (ndev
->dev
== ib_device
) {
2033 mutex_unlock(&device_list_mutex
);
2038 /* Delete all controllers using this device */
2039 mutex_lock(&nvme_rdma_ctrl_mutex
);
2040 list_for_each_entry(ctrl
, &nvme_rdma_ctrl_list
, list
) {
2041 if (ctrl
->device
->dev
!= ib_device
)
2043 nvme_delete_ctrl(&ctrl
->ctrl
);
2045 mutex_unlock(&nvme_rdma_ctrl_mutex
);
2047 flush_workqueue(nvme_delete_wq
);
2050 static struct ib_client nvme_rdma_ib_client
= {
2051 .name
= "nvme_rdma",
2052 .remove
= nvme_rdma_remove_one
2055 static int __init
nvme_rdma_init_module(void)
2059 ret
= ib_register_client(&nvme_rdma_ib_client
);
2063 ret
= nvmf_register_transport(&nvme_rdma_transport
);
2065 goto err_unreg_client
;
2070 ib_unregister_client(&nvme_rdma_ib_client
);
2074 static void __exit
nvme_rdma_cleanup_module(void)
2076 nvmf_unregister_transport(&nvme_rdma_transport
);
2077 ib_unregister_client(&nvme_rdma_ib_client
);
2080 module_init(nvme_rdma_init_module
);
2081 module_exit(nvme_rdma_cleanup_module
);
2083 MODULE_LICENSE("GPL v2");