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/blk-mq-rdma.h>
23 #include <linux/types.h>
24 #include <linux/list.h>
25 #include <linux/mutex.h>
26 #include <linux/scatterlist.h>
27 #include <linux/nvme.h>
28 #include <asm/unaligned.h>
30 #include <rdma/ib_verbs.h>
31 #include <rdma/rdma_cm.h>
32 #include <linux/nvme-rdma.h>
38 #define NVME_RDMA_CONNECT_TIMEOUT_MS 3000 /* 3 second */
40 #define NVME_RDMA_MAX_SEGMENTS 256
42 #define NVME_RDMA_MAX_INLINE_SEGMENTS 1
45 * We handle AEN commands ourselves and don't even let the
46 * block layer know about them.
48 #define NVME_RDMA_NR_AEN_COMMANDS 1
49 #define NVME_RDMA_AQ_BLKMQ_DEPTH \
50 (NVME_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS)
52 struct nvme_rdma_device
{
53 struct ib_device
*dev
;
56 struct list_head entry
;
65 struct nvme_rdma_queue
;
66 struct nvme_rdma_request
{
67 struct nvme_request req
;
69 struct nvme_rdma_qe sqe
;
70 struct ib_sge sge
[1 + NVME_RDMA_MAX_INLINE_SEGMENTS
];
74 struct ib_reg_wr reg_wr
;
75 struct ib_cqe reg_cqe
;
76 struct nvme_rdma_queue
*queue
;
77 struct sg_table sg_table
;
78 struct scatterlist first_sgl
[];
81 enum nvme_rdma_queue_flags
{
83 NVME_RDMA_Q_DELETING
= 1,
86 struct nvme_rdma_queue
{
87 struct nvme_rdma_qe
*rsp_ring
;
90 size_t cmnd_capsule_len
;
91 struct nvme_rdma_ctrl
*ctrl
;
92 struct nvme_rdma_device
*device
;
97 struct rdma_cm_id
*cm_id
;
99 struct completion cm_done
;
102 struct nvme_rdma_ctrl
{
103 /* read only in the hot path */
104 struct nvme_rdma_queue
*queues
;
106 /* other member variables */
107 struct blk_mq_tag_set tag_set
;
108 struct work_struct delete_work
;
109 struct work_struct err_work
;
111 struct nvme_rdma_qe async_event_sqe
;
113 struct delayed_work reconnect_work
;
115 struct list_head list
;
117 struct blk_mq_tag_set admin_tag_set
;
118 struct nvme_rdma_device
*device
;
122 struct sockaddr_storage addr
;
123 struct sockaddr_storage src_addr
;
125 struct nvme_ctrl ctrl
;
128 static inline struct nvme_rdma_ctrl
*to_rdma_ctrl(struct nvme_ctrl
*ctrl
)
130 return container_of(ctrl
, struct nvme_rdma_ctrl
, ctrl
);
133 static LIST_HEAD(device_list
);
134 static DEFINE_MUTEX(device_list_mutex
);
136 static LIST_HEAD(nvme_rdma_ctrl_list
);
137 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex
);
140 * Disabling this option makes small I/O goes faster, but is fundamentally
141 * unsafe. With it turned off we will have to register a global rkey that
142 * allows read and write access to all physical memory.
144 static bool register_always
= true;
145 module_param(register_always
, bool, 0444);
146 MODULE_PARM_DESC(register_always
,
147 "Use memory registration even for contiguous memory regions");
149 static int nvme_rdma_cm_handler(struct rdma_cm_id
*cm_id
,
150 struct rdma_cm_event
*event
);
151 static void nvme_rdma_recv_done(struct ib_cq
*cq
, struct ib_wc
*wc
);
153 static const struct blk_mq_ops nvme_rdma_mq_ops
;
154 static const struct blk_mq_ops nvme_rdma_admin_mq_ops
;
156 /* XXX: really should move to a generic header sooner or later.. */
157 static inline void put_unaligned_le24(u32 val
, u8
*p
)
164 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue
*queue
)
166 return queue
- queue
->ctrl
->queues
;
169 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue
*queue
)
171 return queue
->cmnd_capsule_len
- sizeof(struct nvme_command
);
174 static void nvme_rdma_free_qe(struct ib_device
*ibdev
, struct nvme_rdma_qe
*qe
,
175 size_t capsule_size
, enum dma_data_direction dir
)
177 ib_dma_unmap_single(ibdev
, qe
->dma
, capsule_size
, dir
);
181 static int nvme_rdma_alloc_qe(struct ib_device
*ibdev
, struct nvme_rdma_qe
*qe
,
182 size_t capsule_size
, enum dma_data_direction dir
)
184 qe
->data
= kzalloc(capsule_size
, GFP_KERNEL
);
188 qe
->dma
= ib_dma_map_single(ibdev
, qe
->data
, capsule_size
, dir
);
189 if (ib_dma_mapping_error(ibdev
, qe
->dma
)) {
197 static void nvme_rdma_free_ring(struct ib_device
*ibdev
,
198 struct nvme_rdma_qe
*ring
, size_t ib_queue_size
,
199 size_t capsule_size
, enum dma_data_direction dir
)
203 for (i
= 0; i
< ib_queue_size
; i
++)
204 nvme_rdma_free_qe(ibdev
, &ring
[i
], capsule_size
, dir
);
208 static struct nvme_rdma_qe
*nvme_rdma_alloc_ring(struct ib_device
*ibdev
,
209 size_t ib_queue_size
, size_t capsule_size
,
210 enum dma_data_direction dir
)
212 struct nvme_rdma_qe
*ring
;
215 ring
= kcalloc(ib_queue_size
, sizeof(struct nvme_rdma_qe
), GFP_KERNEL
);
219 for (i
= 0; i
< ib_queue_size
; i
++) {
220 if (nvme_rdma_alloc_qe(ibdev
, &ring
[i
], capsule_size
, dir
))
227 nvme_rdma_free_ring(ibdev
, ring
, i
, capsule_size
, dir
);
231 static void nvme_rdma_qp_event(struct ib_event
*event
, void *context
)
233 pr_debug("QP event %s (%d)\n",
234 ib_event_msg(event
->event
), event
->event
);
238 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue
*queue
)
240 wait_for_completion_interruptible_timeout(&queue
->cm_done
,
241 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS
) + 1);
242 return queue
->cm_error
;
245 static int nvme_rdma_create_qp(struct nvme_rdma_queue
*queue
, const int factor
)
247 struct nvme_rdma_device
*dev
= queue
->device
;
248 struct ib_qp_init_attr init_attr
;
251 memset(&init_attr
, 0, sizeof(init_attr
));
252 init_attr
.event_handler
= nvme_rdma_qp_event
;
254 init_attr
.cap
.max_send_wr
= factor
* queue
->queue_size
+ 1;
256 init_attr
.cap
.max_recv_wr
= queue
->queue_size
+ 1;
257 init_attr
.cap
.max_recv_sge
= 1;
258 init_attr
.cap
.max_send_sge
= 1 + NVME_RDMA_MAX_INLINE_SEGMENTS
;
259 init_attr
.sq_sig_type
= IB_SIGNAL_REQ_WR
;
260 init_attr
.qp_type
= IB_QPT_RC
;
261 init_attr
.send_cq
= queue
->ib_cq
;
262 init_attr
.recv_cq
= queue
->ib_cq
;
264 ret
= rdma_create_qp(queue
->cm_id
, dev
->pd
, &init_attr
);
266 queue
->qp
= queue
->cm_id
->qp
;
270 static int nvme_rdma_reinit_request(void *data
, struct request
*rq
)
272 struct nvme_rdma_ctrl
*ctrl
= data
;
273 struct nvme_rdma_device
*dev
= ctrl
->device
;
274 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
277 ib_dereg_mr(req
->mr
);
279 req
->mr
= ib_alloc_mr(dev
->pd
, IB_MR_TYPE_MEM_REG
,
281 if (IS_ERR(req
->mr
)) {
282 ret
= PTR_ERR(req
->mr
);
287 req
->mr
->need_inval
= false;
293 static void nvme_rdma_exit_request(struct blk_mq_tag_set
*set
,
294 struct request
*rq
, unsigned int hctx_idx
)
296 struct nvme_rdma_ctrl
*ctrl
= set
->driver_data
;
297 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
298 int queue_idx
= (set
== &ctrl
->tag_set
) ? hctx_idx
+ 1 : 0;
299 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[queue_idx
];
300 struct nvme_rdma_device
*dev
= queue
->device
;
303 ib_dereg_mr(req
->mr
);
305 nvme_rdma_free_qe(dev
->dev
, &req
->sqe
, sizeof(struct nvme_command
),
309 static int nvme_rdma_init_request(struct blk_mq_tag_set
*set
,
310 struct request
*rq
, unsigned int hctx_idx
,
311 unsigned int numa_node
)
313 struct nvme_rdma_ctrl
*ctrl
= set
->driver_data
;
314 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
315 int queue_idx
= (set
== &ctrl
->tag_set
) ? hctx_idx
+ 1 : 0;
316 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[queue_idx
];
317 struct nvme_rdma_device
*dev
= queue
->device
;
318 struct ib_device
*ibdev
= dev
->dev
;
321 ret
= nvme_rdma_alloc_qe(ibdev
, &req
->sqe
, sizeof(struct nvme_command
),
326 req
->mr
= ib_alloc_mr(dev
->pd
, IB_MR_TYPE_MEM_REG
,
328 if (IS_ERR(req
->mr
)) {
329 ret
= PTR_ERR(req
->mr
);
338 nvme_rdma_free_qe(dev
->dev
, &req
->sqe
, sizeof(struct nvme_command
),
343 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx
*hctx
, void *data
,
344 unsigned int hctx_idx
)
346 struct nvme_rdma_ctrl
*ctrl
= data
;
347 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[hctx_idx
+ 1];
349 BUG_ON(hctx_idx
>= ctrl
->ctrl
.queue_count
);
351 hctx
->driver_data
= queue
;
355 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx
*hctx
, void *data
,
356 unsigned int hctx_idx
)
358 struct nvme_rdma_ctrl
*ctrl
= data
;
359 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[0];
361 BUG_ON(hctx_idx
!= 0);
363 hctx
->driver_data
= queue
;
367 static void nvme_rdma_free_dev(struct kref
*ref
)
369 struct nvme_rdma_device
*ndev
=
370 container_of(ref
, struct nvme_rdma_device
, ref
);
372 mutex_lock(&device_list_mutex
);
373 list_del(&ndev
->entry
);
374 mutex_unlock(&device_list_mutex
);
376 ib_dealloc_pd(ndev
->pd
);
380 static void nvme_rdma_dev_put(struct nvme_rdma_device
*dev
)
382 kref_put(&dev
->ref
, nvme_rdma_free_dev
);
385 static int nvme_rdma_dev_get(struct nvme_rdma_device
*dev
)
387 return kref_get_unless_zero(&dev
->ref
);
390 static struct nvme_rdma_device
*
391 nvme_rdma_find_get_device(struct rdma_cm_id
*cm_id
)
393 struct nvme_rdma_device
*ndev
;
395 mutex_lock(&device_list_mutex
);
396 list_for_each_entry(ndev
, &device_list
, entry
) {
397 if (ndev
->dev
->node_guid
== cm_id
->device
->node_guid
&&
398 nvme_rdma_dev_get(ndev
))
402 ndev
= kzalloc(sizeof(*ndev
), GFP_KERNEL
);
406 ndev
->dev
= cm_id
->device
;
407 kref_init(&ndev
->ref
);
409 ndev
->pd
= ib_alloc_pd(ndev
->dev
,
410 register_always
? 0 : IB_PD_UNSAFE_GLOBAL_RKEY
);
411 if (IS_ERR(ndev
->pd
))
414 if (!(ndev
->dev
->attrs
.device_cap_flags
&
415 IB_DEVICE_MEM_MGT_EXTENSIONS
)) {
416 dev_err(&ndev
->dev
->dev
,
417 "Memory registrations not supported.\n");
421 list_add(&ndev
->entry
, &device_list
);
423 mutex_unlock(&device_list_mutex
);
427 ib_dealloc_pd(ndev
->pd
);
431 mutex_unlock(&device_list_mutex
);
435 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue
*queue
)
437 struct nvme_rdma_device
*dev
;
438 struct ib_device
*ibdev
;
442 rdma_destroy_qp(queue
->cm_id
);
443 ib_free_cq(queue
->ib_cq
);
445 nvme_rdma_free_ring(ibdev
, queue
->rsp_ring
, queue
->queue_size
,
446 sizeof(struct nvme_completion
), DMA_FROM_DEVICE
);
448 nvme_rdma_dev_put(dev
);
451 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue
*queue
)
453 struct ib_device
*ibdev
;
454 const int send_wr_factor
= 3; /* MR, SEND, INV */
455 const int cq_factor
= send_wr_factor
+ 1; /* + RECV */
456 int comp_vector
, idx
= nvme_rdma_queue_idx(queue
);
459 queue
->device
= nvme_rdma_find_get_device(queue
->cm_id
);
460 if (!queue
->device
) {
461 dev_err(queue
->cm_id
->device
->dev
.parent
,
462 "no client data found!\n");
463 return -ECONNREFUSED
;
465 ibdev
= queue
->device
->dev
;
468 * Spread I/O queues completion vectors according their queue index.
469 * Admin queues can always go on completion vector 0.
471 comp_vector
= idx
== 0 ? idx
: idx
- 1;
473 /* +1 for ib_stop_cq */
474 queue
->ib_cq
= ib_alloc_cq(ibdev
, queue
,
475 cq_factor
* queue
->queue_size
+ 1,
476 comp_vector
, IB_POLL_SOFTIRQ
);
477 if (IS_ERR(queue
->ib_cq
)) {
478 ret
= PTR_ERR(queue
->ib_cq
);
482 ret
= nvme_rdma_create_qp(queue
, send_wr_factor
);
484 goto out_destroy_ib_cq
;
486 queue
->rsp_ring
= nvme_rdma_alloc_ring(ibdev
, queue
->queue_size
,
487 sizeof(struct nvme_completion
), DMA_FROM_DEVICE
);
488 if (!queue
->rsp_ring
) {
496 ib_destroy_qp(queue
->qp
);
498 ib_free_cq(queue
->ib_cq
);
500 nvme_rdma_dev_put(queue
->device
);
504 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl
*ctrl
,
505 int idx
, size_t queue_size
)
507 struct nvme_rdma_queue
*queue
;
508 struct sockaddr
*src_addr
= NULL
;
511 queue
= &ctrl
->queues
[idx
];
513 init_completion(&queue
->cm_done
);
516 queue
->cmnd_capsule_len
= ctrl
->ctrl
.ioccsz
* 16;
518 queue
->cmnd_capsule_len
= sizeof(struct nvme_command
);
520 queue
->queue_size
= queue_size
;
521 atomic_set(&queue
->sig_count
, 0);
523 queue
->cm_id
= rdma_create_id(&init_net
, nvme_rdma_cm_handler
, queue
,
524 RDMA_PS_TCP
, IB_QPT_RC
);
525 if (IS_ERR(queue
->cm_id
)) {
526 dev_info(ctrl
->ctrl
.device
,
527 "failed to create CM ID: %ld\n", PTR_ERR(queue
->cm_id
));
528 return PTR_ERR(queue
->cm_id
);
531 if (ctrl
->ctrl
.opts
->mask
& NVMF_OPT_HOST_TRADDR
)
532 src_addr
= (struct sockaddr
*)&ctrl
->src_addr
;
534 queue
->cm_error
= -ETIMEDOUT
;
535 ret
= rdma_resolve_addr(queue
->cm_id
, src_addr
,
536 (struct sockaddr
*)&ctrl
->addr
,
537 NVME_RDMA_CONNECT_TIMEOUT_MS
);
539 dev_info(ctrl
->ctrl
.device
,
540 "rdma_resolve_addr failed (%d).\n", ret
);
541 goto out_destroy_cm_id
;
544 ret
= nvme_rdma_wait_for_cm(queue
);
546 dev_info(ctrl
->ctrl
.device
,
547 "rdma_resolve_addr wait failed (%d).\n", ret
);
548 goto out_destroy_cm_id
;
551 clear_bit(NVME_RDMA_Q_DELETING
, &queue
->flags
);
556 rdma_destroy_id(queue
->cm_id
);
560 static void nvme_rdma_stop_queue(struct nvme_rdma_queue
*queue
)
562 if (!test_and_clear_bit(NVME_RDMA_Q_LIVE
, &queue
->flags
))
565 rdma_disconnect(queue
->cm_id
);
566 ib_drain_qp(queue
->qp
);
569 static void nvme_rdma_free_queue(struct nvme_rdma_queue
*queue
)
571 if (test_and_set_bit(NVME_RDMA_Q_DELETING
, &queue
->flags
))
574 nvme_rdma_destroy_queue_ib(queue
);
575 rdma_destroy_id(queue
->cm_id
);
578 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl
*ctrl
)
582 for (i
= 1; i
< ctrl
->ctrl
.queue_count
; i
++)
583 nvme_rdma_free_queue(&ctrl
->queues
[i
]);
586 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl
*ctrl
)
590 for (i
= 1; i
< ctrl
->ctrl
.queue_count
; i
++)
591 nvme_rdma_stop_queue(&ctrl
->queues
[i
]);
594 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl
*ctrl
, int idx
)
599 ret
= nvmf_connect_io_queue(&ctrl
->ctrl
, idx
);
601 ret
= nvmf_connect_admin_queue(&ctrl
->ctrl
);
604 set_bit(NVME_RDMA_Q_LIVE
, &ctrl
->queues
[idx
].flags
);
606 dev_info(ctrl
->ctrl
.device
,
607 "failed to connect queue: %d ret=%d\n", idx
, ret
);
611 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl
*ctrl
)
615 for (i
= 1; i
< ctrl
->ctrl
.queue_count
; i
++) {
616 ret
= nvme_rdma_start_queue(ctrl
, i
);
618 goto out_stop_queues
;
624 for (i
--; i
>= 1; i
--)
625 nvme_rdma_stop_queue(&ctrl
->queues
[i
]);
629 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl
*ctrl
)
631 struct nvmf_ctrl_options
*opts
= ctrl
->ctrl
.opts
;
632 struct ib_device
*ibdev
= ctrl
->device
->dev
;
633 unsigned int nr_io_queues
;
636 nr_io_queues
= min(opts
->nr_io_queues
, num_online_cpus());
639 * we map queues according to the device irq vectors for
640 * optimal locality so we don't need more queues than
641 * completion vectors.
643 nr_io_queues
= min_t(unsigned int, nr_io_queues
,
644 ibdev
->num_comp_vectors
);
646 ret
= nvme_set_queue_count(&ctrl
->ctrl
, &nr_io_queues
);
650 ctrl
->ctrl
.queue_count
= nr_io_queues
+ 1;
651 if (ctrl
->ctrl
.queue_count
< 2)
654 dev_info(ctrl
->ctrl
.device
,
655 "creating %d I/O queues.\n", nr_io_queues
);
657 for (i
= 1; i
< ctrl
->ctrl
.queue_count
; i
++) {
658 ret
= nvme_rdma_alloc_queue(ctrl
, i
,
659 ctrl
->ctrl
.sqsize
+ 1);
661 goto out_free_queues
;
667 for (i
--; i
>= 1; i
--)
668 nvme_rdma_free_queue(&ctrl
->queues
[i
]);
673 static void nvme_rdma_free_tagset(struct nvme_ctrl
*nctrl
, bool admin
)
675 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(nctrl
);
676 struct blk_mq_tag_set
*set
= admin
?
677 &ctrl
->admin_tag_set
: &ctrl
->tag_set
;
679 blk_mq_free_tag_set(set
);
680 nvme_rdma_dev_put(ctrl
->device
);
683 static struct blk_mq_tag_set
*nvme_rdma_alloc_tagset(struct nvme_ctrl
*nctrl
,
686 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(nctrl
);
687 struct blk_mq_tag_set
*set
;
691 set
= &ctrl
->admin_tag_set
;
692 memset(set
, 0, sizeof(*set
));
693 set
->ops
= &nvme_rdma_admin_mq_ops
;
694 set
->queue_depth
= NVME_RDMA_AQ_BLKMQ_DEPTH
;
695 set
->reserved_tags
= 2; /* connect + keep-alive */
696 set
->numa_node
= NUMA_NO_NODE
;
697 set
->cmd_size
= sizeof(struct nvme_rdma_request
) +
698 SG_CHUNK_SIZE
* sizeof(struct scatterlist
);
699 set
->driver_data
= ctrl
;
700 set
->nr_hw_queues
= 1;
701 set
->timeout
= ADMIN_TIMEOUT
;
703 set
= &ctrl
->tag_set
;
704 memset(set
, 0, sizeof(*set
));
705 set
->ops
= &nvme_rdma_mq_ops
;
706 set
->queue_depth
= nctrl
->opts
->queue_size
;
707 set
->reserved_tags
= 1; /* fabric connect */
708 set
->numa_node
= NUMA_NO_NODE
;
709 set
->flags
= BLK_MQ_F_SHOULD_MERGE
;
710 set
->cmd_size
= sizeof(struct nvme_rdma_request
) +
711 SG_CHUNK_SIZE
* sizeof(struct scatterlist
);
712 set
->driver_data
= ctrl
;
713 set
->nr_hw_queues
= nctrl
->queue_count
- 1;
714 set
->timeout
= NVME_IO_TIMEOUT
;
717 ret
= blk_mq_alloc_tag_set(set
);
722 * We need a reference on the device as long as the tag_set is alive,
723 * as the MRs in the request structures need a valid ib_device.
725 ret
= nvme_rdma_dev_get(ctrl
->device
);
728 goto out_free_tagset
;
734 blk_mq_free_tag_set(set
);
739 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl
*ctrl
,
742 nvme_rdma_free_qe(ctrl
->queues
[0].device
->dev
, &ctrl
->async_event_sqe
,
743 sizeof(struct nvme_command
), DMA_TO_DEVICE
);
744 nvme_rdma_stop_queue(&ctrl
->queues
[0]);
746 blk_cleanup_queue(ctrl
->ctrl
.admin_q
);
747 nvme_rdma_free_tagset(&ctrl
->ctrl
, true);
749 nvme_rdma_free_queue(&ctrl
->queues
[0]);
752 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl
*ctrl
,
757 error
= nvme_rdma_alloc_queue(ctrl
, 0, NVME_AQ_DEPTH
);
761 ctrl
->device
= ctrl
->queues
[0].device
;
763 ctrl
->max_fr_pages
= min_t(u32
, NVME_RDMA_MAX_SEGMENTS
,
764 ctrl
->device
->dev
->attrs
.max_fast_reg_page_list_len
);
767 ctrl
->ctrl
.admin_tagset
= nvme_rdma_alloc_tagset(&ctrl
->ctrl
, true);
768 if (IS_ERR(ctrl
->ctrl
.admin_tagset
))
771 ctrl
->ctrl
.admin_q
= blk_mq_init_queue(&ctrl
->admin_tag_set
);
772 if (IS_ERR(ctrl
->ctrl
.admin_q
)) {
773 error
= PTR_ERR(ctrl
->ctrl
.admin_q
);
774 goto out_free_tagset
;
777 error
= nvme_reinit_tagset(&ctrl
->ctrl
, ctrl
->ctrl
.admin_tagset
);
782 error
= nvme_rdma_start_queue(ctrl
, 0);
784 goto out_cleanup_queue
;
786 error
= ctrl
->ctrl
.ops
->reg_read64(&ctrl
->ctrl
, NVME_REG_CAP
,
789 dev_err(ctrl
->ctrl
.device
,
790 "prop_get NVME_REG_CAP failed\n");
791 goto out_cleanup_queue
;
795 min_t(int, NVME_CAP_MQES(ctrl
->ctrl
.cap
), ctrl
->ctrl
.sqsize
);
797 error
= nvme_enable_ctrl(&ctrl
->ctrl
, ctrl
->ctrl
.cap
);
799 goto out_cleanup_queue
;
801 ctrl
->ctrl
.max_hw_sectors
=
802 (ctrl
->max_fr_pages
- 1) << (ilog2(SZ_4K
) - 9);
804 error
= nvme_init_identify(&ctrl
->ctrl
);
806 goto out_cleanup_queue
;
808 error
= nvme_rdma_alloc_qe(ctrl
->queues
[0].device
->dev
,
809 &ctrl
->async_event_sqe
, sizeof(struct nvme_command
),
812 goto out_cleanup_queue
;
818 blk_cleanup_queue(ctrl
->ctrl
.admin_q
);
821 nvme_rdma_free_tagset(&ctrl
->ctrl
, true);
823 nvme_rdma_free_queue(&ctrl
->queues
[0]);
827 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl
*ctrl
,
830 nvme_rdma_stop_io_queues(ctrl
);
832 blk_cleanup_queue(ctrl
->ctrl
.connect_q
);
833 nvme_rdma_free_tagset(&ctrl
->ctrl
, false);
835 nvme_rdma_free_io_queues(ctrl
);
838 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl
*ctrl
, bool new)
842 ret
= nvme_rdma_alloc_io_queues(ctrl
);
847 ctrl
->ctrl
.tagset
= nvme_rdma_alloc_tagset(&ctrl
->ctrl
, false);
848 if (IS_ERR(ctrl
->ctrl
.tagset
))
849 goto out_free_io_queues
;
851 ctrl
->ctrl
.connect_q
= blk_mq_init_queue(&ctrl
->tag_set
);
852 if (IS_ERR(ctrl
->ctrl
.connect_q
)) {
853 ret
= PTR_ERR(ctrl
->ctrl
.connect_q
);
854 goto out_free_tag_set
;
857 ret
= nvme_reinit_tagset(&ctrl
->ctrl
, ctrl
->ctrl
.tagset
);
859 goto out_free_io_queues
;
861 blk_mq_update_nr_hw_queues(&ctrl
->tag_set
,
862 ctrl
->ctrl
.queue_count
- 1);
865 ret
= nvme_rdma_start_io_queues(ctrl
);
867 goto out_cleanup_connect_q
;
871 out_cleanup_connect_q
:
873 blk_cleanup_queue(ctrl
->ctrl
.connect_q
);
876 nvme_rdma_free_tagset(&ctrl
->ctrl
, false);
878 nvme_rdma_free_io_queues(ctrl
);
882 static void nvme_rdma_free_ctrl(struct nvme_ctrl
*nctrl
)
884 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(nctrl
);
886 if (list_empty(&ctrl
->list
))
889 mutex_lock(&nvme_rdma_ctrl_mutex
);
890 list_del(&ctrl
->list
);
891 mutex_unlock(&nvme_rdma_ctrl_mutex
);
894 nvmf_free_options(nctrl
->opts
);
899 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl
*ctrl
)
901 /* If we are resetting/deleting then do nothing */
902 if (ctrl
->ctrl
.state
!= NVME_CTRL_RECONNECTING
) {
903 WARN_ON_ONCE(ctrl
->ctrl
.state
== NVME_CTRL_NEW
||
904 ctrl
->ctrl
.state
== NVME_CTRL_LIVE
);
908 if (nvmf_should_reconnect(&ctrl
->ctrl
)) {
909 dev_info(ctrl
->ctrl
.device
, "Reconnecting in %d seconds...\n",
910 ctrl
->ctrl
.opts
->reconnect_delay
);
911 queue_delayed_work(nvme_wq
, &ctrl
->reconnect_work
,
912 ctrl
->ctrl
.opts
->reconnect_delay
* HZ
);
914 dev_info(ctrl
->ctrl
.device
, "Removing controller...\n");
915 queue_work(nvme_wq
, &ctrl
->delete_work
);
919 static void nvme_rdma_reconnect_ctrl_work(struct work_struct
*work
)
921 struct nvme_rdma_ctrl
*ctrl
= container_of(to_delayed_work(work
),
922 struct nvme_rdma_ctrl
, reconnect_work
);
926 ++ctrl
->ctrl
.nr_reconnects
;
928 if (ctrl
->ctrl
.queue_count
> 1)
929 nvme_rdma_destroy_io_queues(ctrl
, false);
931 nvme_rdma_destroy_admin_queue(ctrl
, false);
932 ret
= nvme_rdma_configure_admin_queue(ctrl
, false);
936 if (ctrl
->ctrl
.queue_count
> 1) {
937 ret
= nvme_rdma_configure_io_queues(ctrl
, false);
942 changed
= nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_LIVE
);
944 /* state change failure is ok if we're in DELETING state */
945 WARN_ON_ONCE(ctrl
->ctrl
.state
!= NVME_CTRL_DELETING
);
949 ctrl
->ctrl
.nr_reconnects
= 0;
951 nvme_start_ctrl(&ctrl
->ctrl
);
953 dev_info(ctrl
->ctrl
.device
, "Successfully reconnected\n");
958 dev_info(ctrl
->ctrl
.device
, "Failed reconnect attempt %d\n",
959 ctrl
->ctrl
.nr_reconnects
);
960 nvme_rdma_reconnect_or_remove(ctrl
);
963 static void nvme_rdma_error_recovery_work(struct work_struct
*work
)
965 struct nvme_rdma_ctrl
*ctrl
= container_of(work
,
966 struct nvme_rdma_ctrl
, err_work
);
968 nvme_stop_keep_alive(&ctrl
->ctrl
);
970 if (ctrl
->ctrl
.queue_count
> 1) {
971 nvme_stop_queues(&ctrl
->ctrl
);
972 nvme_rdma_stop_io_queues(ctrl
);
974 blk_mq_quiesce_queue(ctrl
->ctrl
.admin_q
);
975 nvme_rdma_stop_queue(&ctrl
->queues
[0]);
977 /* We must take care of fastfail/requeue all our inflight requests */
978 if (ctrl
->ctrl
.queue_count
> 1)
979 blk_mq_tagset_busy_iter(&ctrl
->tag_set
,
980 nvme_cancel_request
, &ctrl
->ctrl
);
981 blk_mq_tagset_busy_iter(&ctrl
->admin_tag_set
,
982 nvme_cancel_request
, &ctrl
->ctrl
);
985 * queues are not a live anymore, so restart the queues to fail fast
988 blk_mq_unquiesce_queue(ctrl
->ctrl
.admin_q
);
989 nvme_start_queues(&ctrl
->ctrl
);
991 nvme_rdma_reconnect_or_remove(ctrl
);
994 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl
*ctrl
)
996 if (!nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_RECONNECTING
))
999 queue_work(nvme_wq
, &ctrl
->err_work
);
1002 static void nvme_rdma_wr_error(struct ib_cq
*cq
, struct ib_wc
*wc
,
1005 struct nvme_rdma_queue
*queue
= cq
->cq_context
;
1006 struct nvme_rdma_ctrl
*ctrl
= queue
->ctrl
;
1008 if (ctrl
->ctrl
.state
== NVME_CTRL_LIVE
)
1009 dev_info(ctrl
->ctrl
.device
,
1010 "%s for CQE 0x%p failed with status %s (%d)\n",
1012 ib_wc_status_msg(wc
->status
), wc
->status
);
1013 nvme_rdma_error_recovery(ctrl
);
1016 static void nvme_rdma_memreg_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1018 if (unlikely(wc
->status
!= IB_WC_SUCCESS
))
1019 nvme_rdma_wr_error(cq
, wc
, "MEMREG");
1022 static void nvme_rdma_inv_rkey_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1024 if (unlikely(wc
->status
!= IB_WC_SUCCESS
))
1025 nvme_rdma_wr_error(cq
, wc
, "LOCAL_INV");
1028 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue
*queue
,
1029 struct nvme_rdma_request
*req
)
1031 struct ib_send_wr
*bad_wr
;
1032 struct ib_send_wr wr
= {
1033 .opcode
= IB_WR_LOCAL_INV
,
1037 .ex
.invalidate_rkey
= req
->mr
->rkey
,
1040 req
->reg_cqe
.done
= nvme_rdma_inv_rkey_done
;
1041 wr
.wr_cqe
= &req
->reg_cqe
;
1043 return ib_post_send(queue
->qp
, &wr
, &bad_wr
);
1046 static void nvme_rdma_unmap_data(struct nvme_rdma_queue
*queue
,
1049 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1050 struct nvme_rdma_ctrl
*ctrl
= queue
->ctrl
;
1051 struct nvme_rdma_device
*dev
= queue
->device
;
1052 struct ib_device
*ibdev
= dev
->dev
;
1055 if (!blk_rq_bytes(rq
))
1058 if (req
->mr
->need_inval
) {
1059 res
= nvme_rdma_inv_rkey(queue
, req
);
1060 if (unlikely(res
< 0)) {
1061 dev_err(ctrl
->ctrl
.device
,
1062 "Queueing INV WR for rkey %#x failed (%d)\n",
1063 req
->mr
->rkey
, res
);
1064 nvme_rdma_error_recovery(queue
->ctrl
);
1068 ib_dma_unmap_sg(ibdev
, req
->sg_table
.sgl
,
1069 req
->nents
, rq_data_dir(rq
) ==
1070 WRITE
? DMA_TO_DEVICE
: DMA_FROM_DEVICE
);
1072 nvme_cleanup_cmd(rq
);
1073 sg_free_table_chained(&req
->sg_table
, true);
1076 static int nvme_rdma_set_sg_null(struct nvme_command
*c
)
1078 struct nvme_keyed_sgl_desc
*sg
= &c
->common
.dptr
.ksgl
;
1081 put_unaligned_le24(0, sg
->length
);
1082 put_unaligned_le32(0, sg
->key
);
1083 sg
->type
= NVME_KEY_SGL_FMT_DATA_DESC
<< 4;
1087 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue
*queue
,
1088 struct nvme_rdma_request
*req
, struct nvme_command
*c
)
1090 struct nvme_sgl_desc
*sg
= &c
->common
.dptr
.sgl
;
1092 req
->sge
[1].addr
= sg_dma_address(req
->sg_table
.sgl
);
1093 req
->sge
[1].length
= sg_dma_len(req
->sg_table
.sgl
);
1094 req
->sge
[1].lkey
= queue
->device
->pd
->local_dma_lkey
;
1096 sg
->addr
= cpu_to_le64(queue
->ctrl
->ctrl
.icdoff
);
1097 sg
->length
= cpu_to_le32(sg_dma_len(req
->sg_table
.sgl
));
1098 sg
->type
= (NVME_SGL_FMT_DATA_DESC
<< 4) | NVME_SGL_FMT_OFFSET
;
1100 req
->inline_data
= true;
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
;
1125 * Align the MR to a 4K page size to match the ctrl page size and
1126 * the block virtual boundary.
1128 nr
= ib_map_mr_sg(req
->mr
, req
->sg_table
.sgl
, count
, NULL
, SZ_4K
);
1129 if (unlikely(nr
< count
)) {
1135 ib_update_fast_reg_key(req
->mr
, ib_inc_rkey(req
->mr
->rkey
));
1137 req
->reg_cqe
.done
= nvme_rdma_memreg_done
;
1138 memset(&req
->reg_wr
, 0, sizeof(req
->reg_wr
));
1139 req
->reg_wr
.wr
.opcode
= IB_WR_REG_MR
;
1140 req
->reg_wr
.wr
.wr_cqe
= &req
->reg_cqe
;
1141 req
->reg_wr
.wr
.num_sge
= 0;
1142 req
->reg_wr
.mr
= req
->mr
;
1143 req
->reg_wr
.key
= req
->mr
->rkey
;
1144 req
->reg_wr
.access
= IB_ACCESS_LOCAL_WRITE
|
1145 IB_ACCESS_REMOTE_READ
|
1146 IB_ACCESS_REMOTE_WRITE
;
1148 req
->mr
->need_inval
= true;
1150 sg
->addr
= cpu_to_le64(req
->mr
->iova
);
1151 put_unaligned_le24(req
->mr
->length
, sg
->length
);
1152 put_unaligned_le32(req
->mr
->rkey
, sg
->key
);
1153 sg
->type
= (NVME_KEY_SGL_FMT_DATA_DESC
<< 4) |
1154 NVME_SGL_FMT_INVALIDATE
;
1159 static int nvme_rdma_map_data(struct nvme_rdma_queue
*queue
,
1160 struct request
*rq
, struct nvme_command
*c
)
1162 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1163 struct nvme_rdma_device
*dev
= queue
->device
;
1164 struct ib_device
*ibdev
= dev
->dev
;
1168 req
->inline_data
= false;
1169 req
->mr
->need_inval
= false;
1171 c
->common
.flags
|= NVME_CMD_SGL_METABUF
;
1173 if (!blk_rq_bytes(rq
))
1174 return nvme_rdma_set_sg_null(c
);
1176 req
->sg_table
.sgl
= req
->first_sgl
;
1177 ret
= sg_alloc_table_chained(&req
->sg_table
,
1178 blk_rq_nr_phys_segments(rq
), req
->sg_table
.sgl
);
1182 req
->nents
= blk_rq_map_sg(rq
->q
, rq
, req
->sg_table
.sgl
);
1184 count
= ib_dma_map_sg(ibdev
, req
->sg_table
.sgl
, req
->nents
,
1185 rq_data_dir(rq
) == WRITE
? DMA_TO_DEVICE
: DMA_FROM_DEVICE
);
1186 if (unlikely(count
<= 0)) {
1187 sg_free_table_chained(&req
->sg_table
, true);
1192 if (rq_data_dir(rq
) == WRITE
&& nvme_rdma_queue_idx(queue
) &&
1193 blk_rq_payload_bytes(rq
) <=
1194 nvme_rdma_inline_data_size(queue
))
1195 return nvme_rdma_map_sg_inline(queue
, req
, c
);
1197 if (dev
->pd
->flags
& IB_PD_UNSAFE_GLOBAL_RKEY
)
1198 return nvme_rdma_map_sg_single(queue
, req
, c
);
1201 return nvme_rdma_map_sg_fr(queue
, req
, c
, count
);
1204 static void nvme_rdma_send_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1206 if (unlikely(wc
->status
!= IB_WC_SUCCESS
))
1207 nvme_rdma_wr_error(cq
, wc
, "SEND");
1211 * We want to signal completion at least every queue depth/2. This returns the
1212 * largest power of two that is not above half of (queue size + 1) to optimize
1213 * (avoid divisions).
1215 static inline bool nvme_rdma_queue_sig_limit(struct nvme_rdma_queue
*queue
)
1217 int limit
= 1 << ilog2((queue
->queue_size
+ 1) / 2);
1219 return (atomic_inc_return(&queue
->sig_count
) & (limit
- 1)) == 0;
1222 static int nvme_rdma_post_send(struct nvme_rdma_queue
*queue
,
1223 struct nvme_rdma_qe
*qe
, struct ib_sge
*sge
, u32 num_sge
,
1224 struct ib_send_wr
*first
, bool flush
)
1226 struct ib_send_wr wr
, *bad_wr
;
1229 sge
->addr
= qe
->dma
;
1230 sge
->length
= sizeof(struct nvme_command
),
1231 sge
->lkey
= queue
->device
->pd
->local_dma_lkey
;
1233 qe
->cqe
.done
= nvme_rdma_send_done
;
1236 wr
.wr_cqe
= &qe
->cqe
;
1238 wr
.num_sge
= num_sge
;
1239 wr
.opcode
= IB_WR_SEND
;
1243 * Unsignalled send completions are another giant desaster in the
1244 * IB Verbs spec: If we don't regularly post signalled sends
1245 * the send queue will fill up and only a QP reset will rescue us.
1246 * Would have been way to obvious to handle this in hardware or
1247 * at least the RDMA stack..
1249 * Always signal the flushes. The magic request used for the flush
1250 * sequencer is not allocated in our driver's tagset and it's
1251 * triggered to be freed by blk_cleanup_queue(). So we need to
1252 * always mark it as signaled to ensure that the "wr_cqe", which is
1253 * embedded in request's payload, is not freed when __ib_process_cq()
1254 * calls wr_cqe->done().
1256 if (nvme_rdma_queue_sig_limit(queue
) || flush
)
1257 wr
.send_flags
|= IB_SEND_SIGNALED
;
1264 ret
= ib_post_send(queue
->qp
, first
, &bad_wr
);
1265 if (unlikely(ret
)) {
1266 dev_err(queue
->ctrl
->ctrl
.device
,
1267 "%s failed with error code %d\n", __func__
, ret
);
1272 static int nvme_rdma_post_recv(struct nvme_rdma_queue
*queue
,
1273 struct nvme_rdma_qe
*qe
)
1275 struct ib_recv_wr wr
, *bad_wr
;
1279 list
.addr
= qe
->dma
;
1280 list
.length
= sizeof(struct nvme_completion
);
1281 list
.lkey
= queue
->device
->pd
->local_dma_lkey
;
1283 qe
->cqe
.done
= nvme_rdma_recv_done
;
1286 wr
.wr_cqe
= &qe
->cqe
;
1290 ret
= ib_post_recv(queue
->qp
, &wr
, &bad_wr
);
1291 if (unlikely(ret
)) {
1292 dev_err(queue
->ctrl
->ctrl
.device
,
1293 "%s failed with error code %d\n", __func__
, ret
);
1298 static struct blk_mq_tags
*nvme_rdma_tagset(struct nvme_rdma_queue
*queue
)
1300 u32 queue_idx
= nvme_rdma_queue_idx(queue
);
1303 return queue
->ctrl
->admin_tag_set
.tags
[queue_idx
];
1304 return queue
->ctrl
->tag_set
.tags
[queue_idx
- 1];
1307 static void nvme_rdma_submit_async_event(struct nvme_ctrl
*arg
, int aer_idx
)
1309 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(arg
);
1310 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[0];
1311 struct ib_device
*dev
= queue
->device
->dev
;
1312 struct nvme_rdma_qe
*sqe
= &ctrl
->async_event_sqe
;
1313 struct nvme_command
*cmd
= sqe
->data
;
1317 if (WARN_ON_ONCE(aer_idx
!= 0))
1320 ib_dma_sync_single_for_cpu(dev
, sqe
->dma
, sizeof(*cmd
), DMA_TO_DEVICE
);
1322 memset(cmd
, 0, sizeof(*cmd
));
1323 cmd
->common
.opcode
= nvme_admin_async_event
;
1324 cmd
->common
.command_id
= NVME_RDMA_AQ_BLKMQ_DEPTH
;
1325 cmd
->common
.flags
|= NVME_CMD_SGL_METABUF
;
1326 nvme_rdma_set_sg_null(cmd
);
1328 ib_dma_sync_single_for_device(dev
, sqe
->dma
, sizeof(*cmd
),
1331 ret
= nvme_rdma_post_send(queue
, sqe
, &sge
, 1, NULL
, false);
1335 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue
*queue
,
1336 struct nvme_completion
*cqe
, struct ib_wc
*wc
, int tag
)
1339 struct nvme_rdma_request
*req
;
1342 rq
= blk_mq_tag_to_rq(nvme_rdma_tagset(queue
), cqe
->command_id
);
1344 dev_err(queue
->ctrl
->ctrl
.device
,
1345 "tag 0x%x on QP %#x not found\n",
1346 cqe
->command_id
, queue
->qp
->qp_num
);
1347 nvme_rdma_error_recovery(queue
->ctrl
);
1350 req
= blk_mq_rq_to_pdu(rq
);
1355 if ((wc
->wc_flags
& IB_WC_WITH_INVALIDATE
) &&
1356 wc
->ex
.invalidate_rkey
== req
->mr
->rkey
)
1357 req
->mr
->need_inval
= false;
1359 nvme_end_request(rq
, cqe
->status
, cqe
->result
);
1363 static int __nvme_rdma_recv_done(struct ib_cq
*cq
, struct ib_wc
*wc
, int tag
)
1365 struct nvme_rdma_qe
*qe
=
1366 container_of(wc
->wr_cqe
, struct nvme_rdma_qe
, cqe
);
1367 struct nvme_rdma_queue
*queue
= cq
->cq_context
;
1368 struct ib_device
*ibdev
= queue
->device
->dev
;
1369 struct nvme_completion
*cqe
= qe
->data
;
1370 const size_t len
= sizeof(struct nvme_completion
);
1373 if (unlikely(wc
->status
!= IB_WC_SUCCESS
)) {
1374 nvme_rdma_wr_error(cq
, wc
, "RECV");
1378 ib_dma_sync_single_for_cpu(ibdev
, qe
->dma
, len
, DMA_FROM_DEVICE
);
1380 * AEN requests are special as they don't time out and can
1381 * survive any kind of queue freeze and often don't respond to
1382 * aborts. We don't even bother to allocate a struct request
1383 * for them but rather special case them here.
1385 if (unlikely(nvme_rdma_queue_idx(queue
) == 0 &&
1386 cqe
->command_id
>= NVME_RDMA_AQ_BLKMQ_DEPTH
))
1387 nvme_complete_async_event(&queue
->ctrl
->ctrl
, cqe
->status
,
1390 ret
= nvme_rdma_process_nvme_rsp(queue
, cqe
, wc
, tag
);
1391 ib_dma_sync_single_for_device(ibdev
, qe
->dma
, len
, DMA_FROM_DEVICE
);
1393 nvme_rdma_post_recv(queue
, qe
);
1397 static void nvme_rdma_recv_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1399 __nvme_rdma_recv_done(cq
, wc
, -1);
1402 static int nvme_rdma_conn_established(struct nvme_rdma_queue
*queue
)
1406 for (i
= 0; i
< queue
->queue_size
; i
++) {
1407 ret
= nvme_rdma_post_recv(queue
, &queue
->rsp_ring
[i
]);
1409 goto out_destroy_queue_ib
;
1414 out_destroy_queue_ib
:
1415 nvme_rdma_destroy_queue_ib(queue
);
1419 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue
*queue
,
1420 struct rdma_cm_event
*ev
)
1422 struct rdma_cm_id
*cm_id
= queue
->cm_id
;
1423 int status
= ev
->status
;
1424 const char *rej_msg
;
1425 const struct nvme_rdma_cm_rej
*rej_data
;
1428 rej_msg
= rdma_reject_msg(cm_id
, status
);
1429 rej_data
= rdma_consumer_reject_data(cm_id
, ev
, &rej_data_len
);
1431 if (rej_data
&& rej_data_len
>= sizeof(u16
)) {
1432 u16 sts
= le16_to_cpu(rej_data
->sts
);
1434 dev_err(queue
->ctrl
->ctrl
.device
,
1435 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1436 status
, rej_msg
, sts
, nvme_rdma_cm_msg(sts
));
1438 dev_err(queue
->ctrl
->ctrl
.device
,
1439 "Connect rejected: status %d (%s).\n", status
, rej_msg
);
1445 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue
*queue
)
1449 ret
= nvme_rdma_create_queue_ib(queue
);
1453 ret
= rdma_resolve_route(queue
->cm_id
, NVME_RDMA_CONNECT_TIMEOUT_MS
);
1455 dev_err(queue
->ctrl
->ctrl
.device
,
1456 "rdma_resolve_route failed (%d).\n",
1458 goto out_destroy_queue
;
1464 nvme_rdma_destroy_queue_ib(queue
);
1468 static int nvme_rdma_route_resolved(struct nvme_rdma_queue
*queue
)
1470 struct nvme_rdma_ctrl
*ctrl
= queue
->ctrl
;
1471 struct rdma_conn_param param
= { };
1472 struct nvme_rdma_cm_req priv
= { };
1475 param
.qp_num
= queue
->qp
->qp_num
;
1476 param
.flow_control
= 1;
1478 param
.responder_resources
= queue
->device
->dev
->attrs
.max_qp_rd_atom
;
1479 /* maximum retry count */
1480 param
.retry_count
= 7;
1481 param
.rnr_retry_count
= 7;
1482 param
.private_data
= &priv
;
1483 param
.private_data_len
= sizeof(priv
);
1485 priv
.recfmt
= cpu_to_le16(NVME_RDMA_CM_FMT_1_0
);
1486 priv
.qid
= cpu_to_le16(nvme_rdma_queue_idx(queue
));
1488 * set the admin queue depth to the minimum size
1489 * specified by the Fabrics standard.
1491 if (priv
.qid
== 0) {
1492 priv
.hrqsize
= cpu_to_le16(NVME_AQ_DEPTH
);
1493 priv
.hsqsize
= cpu_to_le16(NVME_AQ_DEPTH
- 1);
1496 * current interpretation of the fabrics spec
1497 * is at minimum you make hrqsize sqsize+1, or a
1498 * 1's based representation of sqsize.
1500 priv
.hrqsize
= cpu_to_le16(queue
->queue_size
);
1501 priv
.hsqsize
= cpu_to_le16(queue
->ctrl
->ctrl
.sqsize
);
1504 ret
= rdma_connect(queue
->cm_id
, ¶m
);
1506 dev_err(ctrl
->ctrl
.device
,
1507 "rdma_connect failed (%d).\n", ret
);
1508 goto out_destroy_queue_ib
;
1513 out_destroy_queue_ib
:
1514 nvme_rdma_destroy_queue_ib(queue
);
1518 static int nvme_rdma_cm_handler(struct rdma_cm_id
*cm_id
,
1519 struct rdma_cm_event
*ev
)
1521 struct nvme_rdma_queue
*queue
= cm_id
->context
;
1524 dev_dbg(queue
->ctrl
->ctrl
.device
, "%s (%d): status %d id %p\n",
1525 rdma_event_msg(ev
->event
), ev
->event
,
1528 switch (ev
->event
) {
1529 case RDMA_CM_EVENT_ADDR_RESOLVED
:
1530 cm_error
= nvme_rdma_addr_resolved(queue
);
1532 case RDMA_CM_EVENT_ROUTE_RESOLVED
:
1533 cm_error
= nvme_rdma_route_resolved(queue
);
1535 case RDMA_CM_EVENT_ESTABLISHED
:
1536 queue
->cm_error
= nvme_rdma_conn_established(queue
);
1537 /* complete cm_done regardless of success/failure */
1538 complete(&queue
->cm_done
);
1540 case RDMA_CM_EVENT_REJECTED
:
1541 nvme_rdma_destroy_queue_ib(queue
);
1542 cm_error
= nvme_rdma_conn_rejected(queue
, ev
);
1544 case RDMA_CM_EVENT_ROUTE_ERROR
:
1545 case RDMA_CM_EVENT_CONNECT_ERROR
:
1546 case RDMA_CM_EVENT_UNREACHABLE
:
1547 nvme_rdma_destroy_queue_ib(queue
);
1548 case RDMA_CM_EVENT_ADDR_ERROR
:
1549 dev_dbg(queue
->ctrl
->ctrl
.device
,
1550 "CM error event %d\n", ev
->event
);
1551 cm_error
= -ECONNRESET
;
1553 case RDMA_CM_EVENT_DISCONNECTED
:
1554 case RDMA_CM_EVENT_ADDR_CHANGE
:
1555 case RDMA_CM_EVENT_TIMEWAIT_EXIT
:
1556 dev_dbg(queue
->ctrl
->ctrl
.device
,
1557 "disconnect received - connection closed\n");
1558 nvme_rdma_error_recovery(queue
->ctrl
);
1560 case RDMA_CM_EVENT_DEVICE_REMOVAL
:
1561 /* device removal is handled via the ib_client API */
1564 dev_err(queue
->ctrl
->ctrl
.device
,
1565 "Unexpected RDMA CM event (%d)\n", ev
->event
);
1566 nvme_rdma_error_recovery(queue
->ctrl
);
1571 queue
->cm_error
= cm_error
;
1572 complete(&queue
->cm_done
);
1578 static enum blk_eh_timer_return
1579 nvme_rdma_timeout(struct request
*rq
, bool reserved
)
1581 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1583 /* queue error recovery */
1584 nvme_rdma_error_recovery(req
->queue
->ctrl
);
1586 /* fail with DNR on cmd timeout */
1587 nvme_req(rq
)->status
= NVME_SC_ABORT_REQ
| NVME_SC_DNR
;
1589 return BLK_EH_HANDLED
;
1593 * We cannot accept any other command until the Connect command has completed.
1595 static inline blk_status_t
1596 nvme_rdma_queue_is_ready(struct nvme_rdma_queue
*queue
, struct request
*rq
)
1598 if (unlikely(!test_bit(NVME_RDMA_Q_LIVE
, &queue
->flags
))) {
1599 struct nvme_command
*cmd
= nvme_req(rq
)->cmd
;
1601 if (!blk_rq_is_passthrough(rq
) ||
1602 cmd
->common
.opcode
!= nvme_fabrics_command
||
1603 cmd
->fabrics
.fctype
!= nvme_fabrics_type_connect
) {
1605 * reconnecting state means transport disruption, which
1606 * can take a long time and even might fail permanently,
1607 * so we can't let incoming I/O be requeued forever.
1608 * fail it fast to allow upper layers a chance to
1611 if (queue
->ctrl
->ctrl
.state
== NVME_CTRL_RECONNECTING
)
1612 return BLK_STS_IOERR
;
1613 return BLK_STS_RESOURCE
; /* try again later */
1620 static blk_status_t
nvme_rdma_queue_rq(struct blk_mq_hw_ctx
*hctx
,
1621 const struct blk_mq_queue_data
*bd
)
1623 struct nvme_ns
*ns
= hctx
->queue
->queuedata
;
1624 struct nvme_rdma_queue
*queue
= hctx
->driver_data
;
1625 struct request
*rq
= bd
->rq
;
1626 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1627 struct nvme_rdma_qe
*sqe
= &req
->sqe
;
1628 struct nvme_command
*c
= sqe
->data
;
1630 struct ib_device
*dev
;
1634 WARN_ON_ONCE(rq
->tag
< 0);
1636 ret
= nvme_rdma_queue_is_ready(queue
, rq
);
1640 dev
= queue
->device
->dev
;
1641 ib_dma_sync_single_for_cpu(dev
, sqe
->dma
,
1642 sizeof(struct nvme_command
), DMA_TO_DEVICE
);
1644 ret
= nvme_setup_cmd(ns
, rq
, c
);
1648 blk_mq_start_request(rq
);
1650 err
= nvme_rdma_map_data(queue
, rq
, c
);
1651 if (unlikely(err
< 0)) {
1652 dev_err(queue
->ctrl
->ctrl
.device
,
1653 "Failed to map data (%d)\n", err
);
1654 nvme_cleanup_cmd(rq
);
1658 ib_dma_sync_single_for_device(dev
, sqe
->dma
,
1659 sizeof(struct nvme_command
), DMA_TO_DEVICE
);
1661 if (req_op(rq
) == REQ_OP_FLUSH
)
1663 err
= nvme_rdma_post_send(queue
, sqe
, req
->sge
, req
->num_sge
,
1664 req
->mr
->need_inval
? &req
->reg_wr
.wr
: NULL
, flush
);
1665 if (unlikely(err
)) {
1666 nvme_rdma_unmap_data(queue
, rq
);
1672 if (err
== -ENOMEM
|| err
== -EAGAIN
)
1673 return BLK_STS_RESOURCE
;
1674 return BLK_STS_IOERR
;
1677 static int nvme_rdma_poll(struct blk_mq_hw_ctx
*hctx
, unsigned int tag
)
1679 struct nvme_rdma_queue
*queue
= hctx
->driver_data
;
1680 struct ib_cq
*cq
= queue
->ib_cq
;
1684 while (ib_poll_cq(cq
, 1, &wc
) > 0) {
1685 struct ib_cqe
*cqe
= wc
.wr_cqe
;
1688 if (cqe
->done
== nvme_rdma_recv_done
)
1689 found
|= __nvme_rdma_recv_done(cq
, &wc
, tag
);
1698 static void nvme_rdma_complete_rq(struct request
*rq
)
1700 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1702 nvme_rdma_unmap_data(req
->queue
, rq
);
1703 nvme_complete_rq(rq
);
1706 static int nvme_rdma_map_queues(struct blk_mq_tag_set
*set
)
1708 struct nvme_rdma_ctrl
*ctrl
= set
->driver_data
;
1710 return blk_mq_rdma_map_queues(set
, ctrl
->device
->dev
, 0);
1713 static const struct blk_mq_ops nvme_rdma_mq_ops
= {
1714 .queue_rq
= nvme_rdma_queue_rq
,
1715 .complete
= nvme_rdma_complete_rq
,
1716 .init_request
= nvme_rdma_init_request
,
1717 .exit_request
= nvme_rdma_exit_request
,
1718 .init_hctx
= nvme_rdma_init_hctx
,
1719 .poll
= nvme_rdma_poll
,
1720 .timeout
= nvme_rdma_timeout
,
1721 .map_queues
= nvme_rdma_map_queues
,
1724 static const struct blk_mq_ops nvme_rdma_admin_mq_ops
= {
1725 .queue_rq
= nvme_rdma_queue_rq
,
1726 .complete
= nvme_rdma_complete_rq
,
1727 .init_request
= nvme_rdma_init_request
,
1728 .exit_request
= nvme_rdma_exit_request
,
1729 .init_hctx
= nvme_rdma_init_admin_hctx
,
1730 .timeout
= nvme_rdma_timeout
,
1733 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl
*ctrl
, bool shutdown
)
1735 cancel_work_sync(&ctrl
->err_work
);
1736 cancel_delayed_work_sync(&ctrl
->reconnect_work
);
1738 if (ctrl
->ctrl
.queue_count
> 1) {
1739 nvme_stop_queues(&ctrl
->ctrl
);
1740 blk_mq_tagset_busy_iter(&ctrl
->tag_set
,
1741 nvme_cancel_request
, &ctrl
->ctrl
);
1742 nvme_rdma_destroy_io_queues(ctrl
, shutdown
);
1746 nvme_shutdown_ctrl(&ctrl
->ctrl
);
1748 nvme_disable_ctrl(&ctrl
->ctrl
, ctrl
->ctrl
.cap
);
1750 blk_mq_quiesce_queue(ctrl
->ctrl
.admin_q
);
1751 blk_mq_tagset_busy_iter(&ctrl
->admin_tag_set
,
1752 nvme_cancel_request
, &ctrl
->ctrl
);
1753 blk_mq_unquiesce_queue(ctrl
->ctrl
.admin_q
);
1754 nvme_rdma_destroy_admin_queue(ctrl
, shutdown
);
1757 static void nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl
*ctrl
)
1759 nvme_remove_namespaces(&ctrl
->ctrl
);
1760 nvme_rdma_shutdown_ctrl(ctrl
, true);
1761 nvme_uninit_ctrl(&ctrl
->ctrl
);
1762 nvme_put_ctrl(&ctrl
->ctrl
);
1765 static void nvme_rdma_del_ctrl_work(struct work_struct
*work
)
1767 struct nvme_rdma_ctrl
*ctrl
= container_of(work
,
1768 struct nvme_rdma_ctrl
, delete_work
);
1770 nvme_stop_ctrl(&ctrl
->ctrl
);
1771 nvme_rdma_remove_ctrl(ctrl
);
1774 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl
*ctrl
)
1776 if (!nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_DELETING
))
1779 if (!queue_work(nvme_wq
, &ctrl
->delete_work
))
1785 static int nvme_rdma_del_ctrl(struct nvme_ctrl
*nctrl
)
1787 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(nctrl
);
1791 * Keep a reference until all work is flushed since
1792 * __nvme_rdma_del_ctrl can free the ctrl mem
1794 if (!kref_get_unless_zero(&ctrl
->ctrl
.kref
))
1796 ret
= __nvme_rdma_del_ctrl(ctrl
);
1798 flush_work(&ctrl
->delete_work
);
1799 nvme_put_ctrl(&ctrl
->ctrl
);
1803 static void nvme_rdma_reset_ctrl_work(struct work_struct
*work
)
1805 struct nvme_rdma_ctrl
*ctrl
=
1806 container_of(work
, struct nvme_rdma_ctrl
, ctrl
.reset_work
);
1810 nvme_stop_ctrl(&ctrl
->ctrl
);
1811 nvme_rdma_shutdown_ctrl(ctrl
, false);
1813 ret
= nvme_rdma_configure_admin_queue(ctrl
, false);
1817 if (ctrl
->ctrl
.queue_count
> 1) {
1818 ret
= nvme_rdma_configure_io_queues(ctrl
, false);
1823 changed
= nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_LIVE
);
1824 WARN_ON_ONCE(!changed
);
1826 nvme_start_ctrl(&ctrl
->ctrl
);
1831 dev_warn(ctrl
->ctrl
.device
, "Removing after reset failure\n");
1832 nvme_rdma_remove_ctrl(ctrl
);
1835 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops
= {
1837 .module
= THIS_MODULE
,
1838 .flags
= NVME_F_FABRICS
,
1839 .reg_read32
= nvmf_reg_read32
,
1840 .reg_read64
= nvmf_reg_read64
,
1841 .reg_write32
= nvmf_reg_write32
,
1842 .free_ctrl
= nvme_rdma_free_ctrl
,
1843 .submit_async_event
= nvme_rdma_submit_async_event
,
1844 .delete_ctrl
= nvme_rdma_del_ctrl
,
1845 .get_address
= nvmf_get_address
,
1846 .reinit_request
= nvme_rdma_reinit_request
,
1849 static struct nvme_ctrl
*nvme_rdma_create_ctrl(struct device
*dev
,
1850 struct nvmf_ctrl_options
*opts
)
1852 struct nvme_rdma_ctrl
*ctrl
;
1857 ctrl
= kzalloc(sizeof(*ctrl
), GFP_KERNEL
);
1859 return ERR_PTR(-ENOMEM
);
1860 ctrl
->ctrl
.opts
= opts
;
1861 INIT_LIST_HEAD(&ctrl
->list
);
1863 if (opts
->mask
& NVMF_OPT_TRSVCID
)
1864 port
= opts
->trsvcid
;
1866 port
= __stringify(NVME_RDMA_IP_PORT
);
1868 ret
= inet_pton_with_scope(&init_net
, AF_UNSPEC
,
1869 opts
->traddr
, port
, &ctrl
->addr
);
1871 pr_err("malformed address passed: %s:%s\n", opts
->traddr
, port
);
1875 if (opts
->mask
& NVMF_OPT_HOST_TRADDR
) {
1876 ret
= inet_pton_with_scope(&init_net
, AF_UNSPEC
,
1877 opts
->host_traddr
, NULL
, &ctrl
->src_addr
);
1879 pr_err("malformed src address passed: %s\n",
1885 ret
= nvme_init_ctrl(&ctrl
->ctrl
, dev
, &nvme_rdma_ctrl_ops
,
1886 0 /* no quirks, we're perfect! */);
1890 INIT_DELAYED_WORK(&ctrl
->reconnect_work
,
1891 nvme_rdma_reconnect_ctrl_work
);
1892 INIT_WORK(&ctrl
->err_work
, nvme_rdma_error_recovery_work
);
1893 INIT_WORK(&ctrl
->delete_work
, nvme_rdma_del_ctrl_work
);
1894 INIT_WORK(&ctrl
->ctrl
.reset_work
, nvme_rdma_reset_ctrl_work
);
1896 ctrl
->ctrl
.queue_count
= opts
->nr_io_queues
+ 1; /* +1 for admin queue */
1897 ctrl
->ctrl
.sqsize
= opts
->queue_size
- 1;
1898 ctrl
->ctrl
.kato
= opts
->kato
;
1901 ctrl
->queues
= kcalloc(ctrl
->ctrl
.queue_count
, sizeof(*ctrl
->queues
),
1904 goto out_uninit_ctrl
;
1906 ret
= nvme_rdma_configure_admin_queue(ctrl
, true);
1908 goto out_kfree_queues
;
1910 /* sanity check icdoff */
1911 if (ctrl
->ctrl
.icdoff
) {
1912 dev_err(ctrl
->ctrl
.device
, "icdoff is not supported!\n");
1914 goto out_remove_admin_queue
;
1917 /* sanity check keyed sgls */
1918 if (!(ctrl
->ctrl
.sgls
& (1 << 20))) {
1919 dev_err(ctrl
->ctrl
.device
, "Mandatory keyed sgls are not support\n");
1921 goto out_remove_admin_queue
;
1924 if (opts
->queue_size
> ctrl
->ctrl
.maxcmd
) {
1925 /* warn if maxcmd is lower than queue_size */
1926 dev_warn(ctrl
->ctrl
.device
,
1927 "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1928 opts
->queue_size
, ctrl
->ctrl
.maxcmd
);
1929 opts
->queue_size
= ctrl
->ctrl
.maxcmd
;
1932 if (opts
->queue_size
> ctrl
->ctrl
.sqsize
+ 1) {
1933 /* warn if sqsize is lower than queue_size */
1934 dev_warn(ctrl
->ctrl
.device
,
1935 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1936 opts
->queue_size
, ctrl
->ctrl
.sqsize
+ 1);
1937 opts
->queue_size
= ctrl
->ctrl
.sqsize
+ 1;
1940 if (opts
->nr_io_queues
) {
1941 ret
= nvme_rdma_configure_io_queues(ctrl
, true);
1943 goto out_remove_admin_queue
;
1946 changed
= nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_LIVE
);
1947 WARN_ON_ONCE(!changed
);
1949 dev_info(ctrl
->ctrl
.device
, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1950 ctrl
->ctrl
.opts
->subsysnqn
, &ctrl
->addr
);
1952 kref_get(&ctrl
->ctrl
.kref
);
1954 mutex_lock(&nvme_rdma_ctrl_mutex
);
1955 list_add_tail(&ctrl
->list
, &nvme_rdma_ctrl_list
);
1956 mutex_unlock(&nvme_rdma_ctrl_mutex
);
1958 nvme_start_ctrl(&ctrl
->ctrl
);
1962 out_remove_admin_queue
:
1963 nvme_rdma_destroy_admin_queue(ctrl
, true);
1965 kfree(ctrl
->queues
);
1967 nvme_uninit_ctrl(&ctrl
->ctrl
);
1968 nvme_put_ctrl(&ctrl
->ctrl
);
1971 return ERR_PTR(ret
);
1974 return ERR_PTR(ret
);
1977 static struct nvmf_transport_ops nvme_rdma_transport
= {
1979 .required_opts
= NVMF_OPT_TRADDR
,
1980 .allowed_opts
= NVMF_OPT_TRSVCID
| NVMF_OPT_RECONNECT_DELAY
|
1981 NVMF_OPT_HOST_TRADDR
| NVMF_OPT_CTRL_LOSS_TMO
,
1982 .create_ctrl
= nvme_rdma_create_ctrl
,
1985 static void nvme_rdma_remove_one(struct ib_device
*ib_device
, void *client_data
)
1987 struct nvme_rdma_ctrl
*ctrl
;
1989 /* Delete all controllers using this device */
1990 mutex_lock(&nvme_rdma_ctrl_mutex
);
1991 list_for_each_entry(ctrl
, &nvme_rdma_ctrl_list
, list
) {
1992 if (ctrl
->device
->dev
!= ib_device
)
1994 dev_info(ctrl
->ctrl
.device
,
1995 "Removing ctrl: NQN \"%s\", addr %pISp\n",
1996 ctrl
->ctrl
.opts
->subsysnqn
, &ctrl
->addr
);
1997 __nvme_rdma_del_ctrl(ctrl
);
1999 mutex_unlock(&nvme_rdma_ctrl_mutex
);
2001 flush_workqueue(nvme_wq
);
2004 static struct ib_client nvme_rdma_ib_client
= {
2005 .name
= "nvme_rdma",
2006 .remove
= nvme_rdma_remove_one
2009 static int __init
nvme_rdma_init_module(void)
2013 ret
= ib_register_client(&nvme_rdma_ib_client
);
2017 ret
= nvmf_register_transport(&nvme_rdma_transport
);
2019 goto err_unreg_client
;
2024 ib_unregister_client(&nvme_rdma_ib_client
);
2028 static void __exit
nvme_rdma_cleanup_module(void)
2030 nvmf_unregister_transport(&nvme_rdma_transport
);
2031 ib_unregister_client(&nvme_rdma_ib_client
);
2034 module_init(nvme_rdma_init_module
);
2035 module_exit(nvme_rdma_cleanup_module
);
2037 MODULE_LICENSE("GPL v2");