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
= blk_mq_reinit_tagset(&ctrl
->admin_tag_set
,
778 nvme_rdma_reinit_request
);
783 error
= nvme_rdma_start_queue(ctrl
, 0);
785 goto out_cleanup_queue
;
787 error
= ctrl
->ctrl
.ops
->reg_read64(&ctrl
->ctrl
, NVME_REG_CAP
,
790 dev_err(ctrl
->ctrl
.device
,
791 "prop_get NVME_REG_CAP failed\n");
792 goto out_cleanup_queue
;
796 min_t(int, NVME_CAP_MQES(ctrl
->ctrl
.cap
), ctrl
->ctrl
.sqsize
);
798 error
= nvme_enable_ctrl(&ctrl
->ctrl
, ctrl
->ctrl
.cap
);
800 goto out_cleanup_queue
;
802 ctrl
->ctrl
.max_hw_sectors
=
803 (ctrl
->max_fr_pages
- 1) << (ilog2(SZ_4K
) - 9);
805 error
= nvme_init_identify(&ctrl
->ctrl
);
807 goto out_cleanup_queue
;
809 error
= nvme_rdma_alloc_qe(ctrl
->queues
[0].device
->dev
,
810 &ctrl
->async_event_sqe
, sizeof(struct nvme_command
),
813 goto out_cleanup_queue
;
819 blk_cleanup_queue(ctrl
->ctrl
.admin_q
);
822 nvme_rdma_free_tagset(&ctrl
->ctrl
, true);
824 nvme_rdma_free_queue(&ctrl
->queues
[0]);
828 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl
*ctrl
,
831 nvme_rdma_stop_io_queues(ctrl
);
833 blk_cleanup_queue(ctrl
->ctrl
.connect_q
);
834 nvme_rdma_free_tagset(&ctrl
->ctrl
, false);
836 nvme_rdma_free_io_queues(ctrl
);
839 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl
*ctrl
, bool new)
843 ret
= nvme_rdma_alloc_io_queues(ctrl
);
848 ctrl
->ctrl
.tagset
= nvme_rdma_alloc_tagset(&ctrl
->ctrl
, false);
849 if (IS_ERR(ctrl
->ctrl
.tagset
))
850 goto out_free_io_queues
;
852 ctrl
->ctrl
.connect_q
= blk_mq_init_queue(&ctrl
->tag_set
);
853 if (IS_ERR(ctrl
->ctrl
.connect_q
)) {
854 ret
= PTR_ERR(ctrl
->ctrl
.connect_q
);
855 goto out_free_tag_set
;
858 ret
= blk_mq_reinit_tagset(&ctrl
->tag_set
,
859 nvme_rdma_reinit_request
);
861 goto out_free_io_queues
;
863 blk_mq_update_nr_hw_queues(&ctrl
->tag_set
,
864 ctrl
->ctrl
.queue_count
- 1);
867 ret
= nvme_rdma_start_io_queues(ctrl
);
869 goto out_cleanup_connect_q
;
873 out_cleanup_connect_q
:
875 blk_cleanup_queue(ctrl
->ctrl
.connect_q
);
878 nvme_rdma_free_tagset(&ctrl
->ctrl
, false);
880 nvme_rdma_free_io_queues(ctrl
);
884 static void nvme_rdma_free_ctrl(struct nvme_ctrl
*nctrl
)
886 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(nctrl
);
888 if (list_empty(&ctrl
->list
))
891 mutex_lock(&nvme_rdma_ctrl_mutex
);
892 list_del(&ctrl
->list
);
893 mutex_unlock(&nvme_rdma_ctrl_mutex
);
896 nvmf_free_options(nctrl
->opts
);
901 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl
*ctrl
)
903 /* If we are resetting/deleting then do nothing */
904 if (ctrl
->ctrl
.state
!= NVME_CTRL_RECONNECTING
) {
905 WARN_ON_ONCE(ctrl
->ctrl
.state
== NVME_CTRL_NEW
||
906 ctrl
->ctrl
.state
== NVME_CTRL_LIVE
);
910 if (nvmf_should_reconnect(&ctrl
->ctrl
)) {
911 dev_info(ctrl
->ctrl
.device
, "Reconnecting in %d seconds...\n",
912 ctrl
->ctrl
.opts
->reconnect_delay
);
913 queue_delayed_work(nvme_wq
, &ctrl
->reconnect_work
,
914 ctrl
->ctrl
.opts
->reconnect_delay
* HZ
);
916 dev_info(ctrl
->ctrl
.device
, "Removing controller...\n");
917 queue_work(nvme_wq
, &ctrl
->delete_work
);
921 static void nvme_rdma_reconnect_ctrl_work(struct work_struct
*work
)
923 struct nvme_rdma_ctrl
*ctrl
= container_of(to_delayed_work(work
),
924 struct nvme_rdma_ctrl
, reconnect_work
);
928 ++ctrl
->ctrl
.nr_reconnects
;
930 if (ctrl
->ctrl
.queue_count
> 1)
931 nvme_rdma_destroy_io_queues(ctrl
, false);
933 nvme_rdma_destroy_admin_queue(ctrl
, false);
934 ret
= nvme_rdma_configure_admin_queue(ctrl
, false);
938 if (ctrl
->ctrl
.queue_count
> 1) {
939 ret
= nvme_rdma_configure_io_queues(ctrl
, false);
944 changed
= nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_LIVE
);
946 /* state change failure is ok if we're in DELETING state */
947 WARN_ON_ONCE(ctrl
->ctrl
.state
!= NVME_CTRL_DELETING
);
951 ctrl
->ctrl
.nr_reconnects
= 0;
953 nvme_start_ctrl(&ctrl
->ctrl
);
955 dev_info(ctrl
->ctrl
.device
, "Successfully reconnected\n");
960 dev_info(ctrl
->ctrl
.device
, "Failed reconnect attempt %d\n",
961 ctrl
->ctrl
.nr_reconnects
);
962 nvme_rdma_reconnect_or_remove(ctrl
);
965 static void nvme_rdma_error_recovery_work(struct work_struct
*work
)
967 struct nvme_rdma_ctrl
*ctrl
= container_of(work
,
968 struct nvme_rdma_ctrl
, err_work
);
970 nvme_stop_keep_alive(&ctrl
->ctrl
);
972 if (ctrl
->ctrl
.queue_count
> 1) {
973 nvme_stop_queues(&ctrl
->ctrl
);
974 nvme_rdma_stop_io_queues(ctrl
);
976 blk_mq_quiesce_queue(ctrl
->ctrl
.admin_q
);
977 nvme_rdma_stop_queue(&ctrl
->queues
[0]);
979 /* We must take care of fastfail/requeue all our inflight requests */
980 if (ctrl
->ctrl
.queue_count
> 1)
981 blk_mq_tagset_busy_iter(&ctrl
->tag_set
,
982 nvme_cancel_request
, &ctrl
->ctrl
);
983 blk_mq_tagset_busy_iter(&ctrl
->admin_tag_set
,
984 nvme_cancel_request
, &ctrl
->ctrl
);
987 * queues are not a live anymore, so restart the queues to fail fast
990 blk_mq_unquiesce_queue(ctrl
->ctrl
.admin_q
);
991 nvme_start_queues(&ctrl
->ctrl
);
993 nvme_rdma_reconnect_or_remove(ctrl
);
996 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl
*ctrl
)
998 if (!nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_RECONNECTING
))
1001 queue_work(nvme_wq
, &ctrl
->err_work
);
1004 static void nvme_rdma_wr_error(struct ib_cq
*cq
, struct ib_wc
*wc
,
1007 struct nvme_rdma_queue
*queue
= cq
->cq_context
;
1008 struct nvme_rdma_ctrl
*ctrl
= queue
->ctrl
;
1010 if (ctrl
->ctrl
.state
== NVME_CTRL_LIVE
)
1011 dev_info(ctrl
->ctrl
.device
,
1012 "%s for CQE 0x%p failed with status %s (%d)\n",
1014 ib_wc_status_msg(wc
->status
), wc
->status
);
1015 nvme_rdma_error_recovery(ctrl
);
1018 static void nvme_rdma_memreg_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1020 if (unlikely(wc
->status
!= IB_WC_SUCCESS
))
1021 nvme_rdma_wr_error(cq
, wc
, "MEMREG");
1024 static void nvme_rdma_inv_rkey_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1026 if (unlikely(wc
->status
!= IB_WC_SUCCESS
))
1027 nvme_rdma_wr_error(cq
, wc
, "LOCAL_INV");
1030 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue
*queue
,
1031 struct nvme_rdma_request
*req
)
1033 struct ib_send_wr
*bad_wr
;
1034 struct ib_send_wr wr
= {
1035 .opcode
= IB_WR_LOCAL_INV
,
1039 .ex
.invalidate_rkey
= req
->mr
->rkey
,
1042 req
->reg_cqe
.done
= nvme_rdma_inv_rkey_done
;
1043 wr
.wr_cqe
= &req
->reg_cqe
;
1045 return ib_post_send(queue
->qp
, &wr
, &bad_wr
);
1048 static void nvme_rdma_unmap_data(struct nvme_rdma_queue
*queue
,
1051 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1052 struct nvme_rdma_ctrl
*ctrl
= queue
->ctrl
;
1053 struct nvme_rdma_device
*dev
= queue
->device
;
1054 struct ib_device
*ibdev
= dev
->dev
;
1057 if (!blk_rq_bytes(rq
))
1060 if (req
->mr
->need_inval
) {
1061 res
= nvme_rdma_inv_rkey(queue
, req
);
1062 if (unlikely(res
< 0)) {
1063 dev_err(ctrl
->ctrl
.device
,
1064 "Queueing INV WR for rkey %#x failed (%d)\n",
1065 req
->mr
->rkey
, res
);
1066 nvme_rdma_error_recovery(queue
->ctrl
);
1070 ib_dma_unmap_sg(ibdev
, req
->sg_table
.sgl
,
1071 req
->nents
, rq_data_dir(rq
) ==
1072 WRITE
? DMA_TO_DEVICE
: DMA_FROM_DEVICE
);
1074 nvme_cleanup_cmd(rq
);
1075 sg_free_table_chained(&req
->sg_table
, true);
1078 static int nvme_rdma_set_sg_null(struct nvme_command
*c
)
1080 struct nvme_keyed_sgl_desc
*sg
= &c
->common
.dptr
.ksgl
;
1083 put_unaligned_le24(0, sg
->length
);
1084 put_unaligned_le32(0, sg
->key
);
1085 sg
->type
= NVME_KEY_SGL_FMT_DATA_DESC
<< 4;
1089 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue
*queue
,
1090 struct nvme_rdma_request
*req
, struct nvme_command
*c
)
1092 struct nvme_sgl_desc
*sg
= &c
->common
.dptr
.sgl
;
1094 req
->sge
[1].addr
= sg_dma_address(req
->sg_table
.sgl
);
1095 req
->sge
[1].length
= sg_dma_len(req
->sg_table
.sgl
);
1096 req
->sge
[1].lkey
= queue
->device
->pd
->local_dma_lkey
;
1098 sg
->addr
= cpu_to_le64(queue
->ctrl
->ctrl
.icdoff
);
1099 sg
->length
= cpu_to_le32(sg_dma_len(req
->sg_table
.sgl
));
1100 sg
->type
= (NVME_SGL_FMT_DATA_DESC
<< 4) | NVME_SGL_FMT_OFFSET
;
1102 req
->inline_data
= true;
1107 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue
*queue
,
1108 struct nvme_rdma_request
*req
, struct nvme_command
*c
)
1110 struct nvme_keyed_sgl_desc
*sg
= &c
->common
.dptr
.ksgl
;
1112 sg
->addr
= cpu_to_le64(sg_dma_address(req
->sg_table
.sgl
));
1113 put_unaligned_le24(sg_dma_len(req
->sg_table
.sgl
), sg
->length
);
1114 put_unaligned_le32(queue
->device
->pd
->unsafe_global_rkey
, sg
->key
);
1115 sg
->type
= NVME_KEY_SGL_FMT_DATA_DESC
<< 4;
1119 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue
*queue
,
1120 struct nvme_rdma_request
*req
, struct nvme_command
*c
,
1123 struct nvme_keyed_sgl_desc
*sg
= &c
->common
.dptr
.ksgl
;
1127 * Align the MR to a 4K page size to match the ctrl page size and
1128 * the block virtual boundary.
1130 nr
= ib_map_mr_sg(req
->mr
, req
->sg_table
.sgl
, count
, NULL
, SZ_4K
);
1131 if (unlikely(nr
< count
)) {
1137 ib_update_fast_reg_key(req
->mr
, ib_inc_rkey(req
->mr
->rkey
));
1139 req
->reg_cqe
.done
= nvme_rdma_memreg_done
;
1140 memset(&req
->reg_wr
, 0, sizeof(req
->reg_wr
));
1141 req
->reg_wr
.wr
.opcode
= IB_WR_REG_MR
;
1142 req
->reg_wr
.wr
.wr_cqe
= &req
->reg_cqe
;
1143 req
->reg_wr
.wr
.num_sge
= 0;
1144 req
->reg_wr
.mr
= req
->mr
;
1145 req
->reg_wr
.key
= req
->mr
->rkey
;
1146 req
->reg_wr
.access
= IB_ACCESS_LOCAL_WRITE
|
1147 IB_ACCESS_REMOTE_READ
|
1148 IB_ACCESS_REMOTE_WRITE
;
1150 req
->mr
->need_inval
= true;
1152 sg
->addr
= cpu_to_le64(req
->mr
->iova
);
1153 put_unaligned_le24(req
->mr
->length
, sg
->length
);
1154 put_unaligned_le32(req
->mr
->rkey
, sg
->key
);
1155 sg
->type
= (NVME_KEY_SGL_FMT_DATA_DESC
<< 4) |
1156 NVME_SGL_FMT_INVALIDATE
;
1161 static int nvme_rdma_map_data(struct nvme_rdma_queue
*queue
,
1162 struct request
*rq
, struct nvme_command
*c
)
1164 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1165 struct nvme_rdma_device
*dev
= queue
->device
;
1166 struct ib_device
*ibdev
= dev
->dev
;
1170 req
->inline_data
= false;
1171 req
->mr
->need_inval
= false;
1173 c
->common
.flags
|= NVME_CMD_SGL_METABUF
;
1175 if (!blk_rq_bytes(rq
))
1176 return nvme_rdma_set_sg_null(c
);
1178 req
->sg_table
.sgl
= req
->first_sgl
;
1179 ret
= sg_alloc_table_chained(&req
->sg_table
,
1180 blk_rq_nr_phys_segments(rq
), req
->sg_table
.sgl
);
1184 req
->nents
= blk_rq_map_sg(rq
->q
, rq
, req
->sg_table
.sgl
);
1186 count
= ib_dma_map_sg(ibdev
, req
->sg_table
.sgl
, req
->nents
,
1187 rq_data_dir(rq
) == WRITE
? DMA_TO_DEVICE
: DMA_FROM_DEVICE
);
1188 if (unlikely(count
<= 0)) {
1189 sg_free_table_chained(&req
->sg_table
, true);
1194 if (rq_data_dir(rq
) == WRITE
&& nvme_rdma_queue_idx(queue
) &&
1195 blk_rq_payload_bytes(rq
) <=
1196 nvme_rdma_inline_data_size(queue
))
1197 return nvme_rdma_map_sg_inline(queue
, req
, c
);
1199 if (dev
->pd
->flags
& IB_PD_UNSAFE_GLOBAL_RKEY
)
1200 return nvme_rdma_map_sg_single(queue
, req
, c
);
1203 return nvme_rdma_map_sg_fr(queue
, req
, c
, count
);
1206 static void nvme_rdma_send_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1208 if (unlikely(wc
->status
!= IB_WC_SUCCESS
))
1209 nvme_rdma_wr_error(cq
, wc
, "SEND");
1213 * We want to signal completion at least every queue depth/2. This returns the
1214 * largest power of two that is not above half of (queue size + 1) to optimize
1215 * (avoid divisions).
1217 static inline bool nvme_rdma_queue_sig_limit(struct nvme_rdma_queue
*queue
)
1219 int limit
= 1 << ilog2((queue
->queue_size
+ 1) / 2);
1221 return (atomic_inc_return(&queue
->sig_count
) & (limit
- 1)) == 0;
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
, bool flush
)
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
;
1235 qe
->cqe
.done
= nvme_rdma_send_done
;
1238 wr
.wr_cqe
= &qe
->cqe
;
1240 wr
.num_sge
= num_sge
;
1241 wr
.opcode
= IB_WR_SEND
;
1245 * Unsignalled send completions are another giant desaster in the
1246 * IB Verbs spec: If we don't regularly post signalled sends
1247 * the send queue will fill up and only a QP reset will rescue us.
1248 * Would have been way to obvious to handle this in hardware or
1249 * at least the RDMA stack..
1251 * Always signal the flushes. The magic request used for the flush
1252 * sequencer is not allocated in our driver's tagset and it's
1253 * triggered to be freed by blk_cleanup_queue(). So we need to
1254 * always mark it as signaled to ensure that the "wr_cqe", which is
1255 * embedded in request's payload, is not freed when __ib_process_cq()
1256 * calls wr_cqe->done().
1258 if (nvme_rdma_queue_sig_limit(queue
) || flush
)
1259 wr
.send_flags
|= IB_SEND_SIGNALED
;
1266 ret
= ib_post_send(queue
->qp
, first
, &bad_wr
);
1267 if (unlikely(ret
)) {
1268 dev_err(queue
->ctrl
->ctrl
.device
,
1269 "%s failed with error code %d\n", __func__
, ret
);
1274 static int nvme_rdma_post_recv(struct nvme_rdma_queue
*queue
,
1275 struct nvme_rdma_qe
*qe
)
1277 struct ib_recv_wr wr
, *bad_wr
;
1281 list
.addr
= qe
->dma
;
1282 list
.length
= sizeof(struct nvme_completion
);
1283 list
.lkey
= queue
->device
->pd
->local_dma_lkey
;
1285 qe
->cqe
.done
= nvme_rdma_recv_done
;
1288 wr
.wr_cqe
= &qe
->cqe
;
1292 ret
= ib_post_recv(queue
->qp
, &wr
, &bad_wr
);
1293 if (unlikely(ret
)) {
1294 dev_err(queue
->ctrl
->ctrl
.device
,
1295 "%s failed with error code %d\n", __func__
, ret
);
1300 static struct blk_mq_tags
*nvme_rdma_tagset(struct nvme_rdma_queue
*queue
)
1302 u32 queue_idx
= nvme_rdma_queue_idx(queue
);
1305 return queue
->ctrl
->admin_tag_set
.tags
[queue_idx
];
1306 return queue
->ctrl
->tag_set
.tags
[queue_idx
- 1];
1309 static void nvme_rdma_submit_async_event(struct nvme_ctrl
*arg
, int aer_idx
)
1311 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(arg
);
1312 struct nvme_rdma_queue
*queue
= &ctrl
->queues
[0];
1313 struct ib_device
*dev
= queue
->device
->dev
;
1314 struct nvme_rdma_qe
*sqe
= &ctrl
->async_event_sqe
;
1315 struct nvme_command
*cmd
= sqe
->data
;
1319 if (WARN_ON_ONCE(aer_idx
!= 0))
1322 ib_dma_sync_single_for_cpu(dev
, sqe
->dma
, sizeof(*cmd
), DMA_TO_DEVICE
);
1324 memset(cmd
, 0, sizeof(*cmd
));
1325 cmd
->common
.opcode
= nvme_admin_async_event
;
1326 cmd
->common
.command_id
= NVME_RDMA_AQ_BLKMQ_DEPTH
;
1327 cmd
->common
.flags
|= NVME_CMD_SGL_METABUF
;
1328 nvme_rdma_set_sg_null(cmd
);
1330 ib_dma_sync_single_for_device(dev
, sqe
->dma
, sizeof(*cmd
),
1333 ret
= nvme_rdma_post_send(queue
, sqe
, &sge
, 1, NULL
, false);
1337 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue
*queue
,
1338 struct nvme_completion
*cqe
, struct ib_wc
*wc
, int tag
)
1341 struct nvme_rdma_request
*req
;
1344 rq
= blk_mq_tag_to_rq(nvme_rdma_tagset(queue
), cqe
->command_id
);
1346 dev_err(queue
->ctrl
->ctrl
.device
,
1347 "tag 0x%x on QP %#x not found\n",
1348 cqe
->command_id
, queue
->qp
->qp_num
);
1349 nvme_rdma_error_recovery(queue
->ctrl
);
1352 req
= blk_mq_rq_to_pdu(rq
);
1357 if ((wc
->wc_flags
& IB_WC_WITH_INVALIDATE
) &&
1358 wc
->ex
.invalidate_rkey
== req
->mr
->rkey
)
1359 req
->mr
->need_inval
= false;
1361 nvme_end_request(rq
, cqe
->status
, cqe
->result
);
1365 static int __nvme_rdma_recv_done(struct ib_cq
*cq
, struct ib_wc
*wc
, int tag
)
1367 struct nvme_rdma_qe
*qe
=
1368 container_of(wc
->wr_cqe
, struct nvme_rdma_qe
, cqe
);
1369 struct nvme_rdma_queue
*queue
= cq
->cq_context
;
1370 struct ib_device
*ibdev
= queue
->device
->dev
;
1371 struct nvme_completion
*cqe
= qe
->data
;
1372 const size_t len
= sizeof(struct nvme_completion
);
1375 if (unlikely(wc
->status
!= IB_WC_SUCCESS
)) {
1376 nvme_rdma_wr_error(cq
, wc
, "RECV");
1380 ib_dma_sync_single_for_cpu(ibdev
, qe
->dma
, len
, DMA_FROM_DEVICE
);
1382 * AEN requests are special as they don't time out and can
1383 * survive any kind of queue freeze and often don't respond to
1384 * aborts. We don't even bother to allocate a struct request
1385 * for them but rather special case them here.
1387 if (unlikely(nvme_rdma_queue_idx(queue
) == 0 &&
1388 cqe
->command_id
>= NVME_RDMA_AQ_BLKMQ_DEPTH
))
1389 nvme_complete_async_event(&queue
->ctrl
->ctrl
, cqe
->status
,
1392 ret
= nvme_rdma_process_nvme_rsp(queue
, cqe
, wc
, tag
);
1393 ib_dma_sync_single_for_device(ibdev
, qe
->dma
, len
, DMA_FROM_DEVICE
);
1395 nvme_rdma_post_recv(queue
, qe
);
1399 static void nvme_rdma_recv_done(struct ib_cq
*cq
, struct ib_wc
*wc
)
1401 __nvme_rdma_recv_done(cq
, wc
, -1);
1404 static int nvme_rdma_conn_established(struct nvme_rdma_queue
*queue
)
1408 for (i
= 0; i
< queue
->queue_size
; i
++) {
1409 ret
= nvme_rdma_post_recv(queue
, &queue
->rsp_ring
[i
]);
1411 goto out_destroy_queue_ib
;
1416 out_destroy_queue_ib
:
1417 nvme_rdma_destroy_queue_ib(queue
);
1421 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue
*queue
,
1422 struct rdma_cm_event
*ev
)
1424 struct rdma_cm_id
*cm_id
= queue
->cm_id
;
1425 int status
= ev
->status
;
1426 const char *rej_msg
;
1427 const struct nvme_rdma_cm_rej
*rej_data
;
1430 rej_msg
= rdma_reject_msg(cm_id
, status
);
1431 rej_data
= rdma_consumer_reject_data(cm_id
, ev
, &rej_data_len
);
1433 if (rej_data
&& rej_data_len
>= sizeof(u16
)) {
1434 u16 sts
= le16_to_cpu(rej_data
->sts
);
1436 dev_err(queue
->ctrl
->ctrl
.device
,
1437 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1438 status
, rej_msg
, sts
, nvme_rdma_cm_msg(sts
));
1440 dev_err(queue
->ctrl
->ctrl
.device
,
1441 "Connect rejected: status %d (%s).\n", status
, rej_msg
);
1447 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue
*queue
)
1451 ret
= nvme_rdma_create_queue_ib(queue
);
1455 ret
= rdma_resolve_route(queue
->cm_id
, NVME_RDMA_CONNECT_TIMEOUT_MS
);
1457 dev_err(queue
->ctrl
->ctrl
.device
,
1458 "rdma_resolve_route failed (%d).\n",
1460 goto out_destroy_queue
;
1466 nvme_rdma_destroy_queue_ib(queue
);
1470 static int nvme_rdma_route_resolved(struct nvme_rdma_queue
*queue
)
1472 struct nvme_rdma_ctrl
*ctrl
= queue
->ctrl
;
1473 struct rdma_conn_param param
= { };
1474 struct nvme_rdma_cm_req priv
= { };
1477 param
.qp_num
= queue
->qp
->qp_num
;
1478 param
.flow_control
= 1;
1480 param
.responder_resources
= queue
->device
->dev
->attrs
.max_qp_rd_atom
;
1481 /* maximum retry count */
1482 param
.retry_count
= 7;
1483 param
.rnr_retry_count
= 7;
1484 param
.private_data
= &priv
;
1485 param
.private_data_len
= sizeof(priv
);
1487 priv
.recfmt
= cpu_to_le16(NVME_RDMA_CM_FMT_1_0
);
1488 priv
.qid
= cpu_to_le16(nvme_rdma_queue_idx(queue
));
1490 * set the admin queue depth to the minimum size
1491 * specified by the Fabrics standard.
1493 if (priv
.qid
== 0) {
1494 priv
.hrqsize
= cpu_to_le16(NVME_AQ_DEPTH
);
1495 priv
.hsqsize
= cpu_to_le16(NVME_AQ_DEPTH
- 1);
1498 * current interpretation of the fabrics spec
1499 * is at minimum you make hrqsize sqsize+1, or a
1500 * 1's based representation of sqsize.
1502 priv
.hrqsize
= cpu_to_le16(queue
->queue_size
);
1503 priv
.hsqsize
= cpu_to_le16(queue
->ctrl
->ctrl
.sqsize
);
1506 ret
= rdma_connect(queue
->cm_id
, ¶m
);
1508 dev_err(ctrl
->ctrl
.device
,
1509 "rdma_connect failed (%d).\n", ret
);
1510 goto out_destroy_queue_ib
;
1515 out_destroy_queue_ib
:
1516 nvme_rdma_destroy_queue_ib(queue
);
1520 static int nvme_rdma_cm_handler(struct rdma_cm_id
*cm_id
,
1521 struct rdma_cm_event
*ev
)
1523 struct nvme_rdma_queue
*queue
= cm_id
->context
;
1526 dev_dbg(queue
->ctrl
->ctrl
.device
, "%s (%d): status %d id %p\n",
1527 rdma_event_msg(ev
->event
), ev
->event
,
1530 switch (ev
->event
) {
1531 case RDMA_CM_EVENT_ADDR_RESOLVED
:
1532 cm_error
= nvme_rdma_addr_resolved(queue
);
1534 case RDMA_CM_EVENT_ROUTE_RESOLVED
:
1535 cm_error
= nvme_rdma_route_resolved(queue
);
1537 case RDMA_CM_EVENT_ESTABLISHED
:
1538 queue
->cm_error
= nvme_rdma_conn_established(queue
);
1539 /* complete cm_done regardless of success/failure */
1540 complete(&queue
->cm_done
);
1542 case RDMA_CM_EVENT_REJECTED
:
1543 nvme_rdma_destroy_queue_ib(queue
);
1544 cm_error
= nvme_rdma_conn_rejected(queue
, ev
);
1546 case RDMA_CM_EVENT_ROUTE_ERROR
:
1547 case RDMA_CM_EVENT_CONNECT_ERROR
:
1548 case RDMA_CM_EVENT_UNREACHABLE
:
1549 nvme_rdma_destroy_queue_ib(queue
);
1550 case RDMA_CM_EVENT_ADDR_ERROR
:
1551 dev_dbg(queue
->ctrl
->ctrl
.device
,
1552 "CM error event %d\n", ev
->event
);
1553 cm_error
= -ECONNRESET
;
1555 case RDMA_CM_EVENT_DISCONNECTED
:
1556 case RDMA_CM_EVENT_ADDR_CHANGE
:
1557 case RDMA_CM_EVENT_TIMEWAIT_EXIT
:
1558 dev_dbg(queue
->ctrl
->ctrl
.device
,
1559 "disconnect received - connection closed\n");
1560 nvme_rdma_error_recovery(queue
->ctrl
);
1562 case RDMA_CM_EVENT_DEVICE_REMOVAL
:
1563 /* device removal is handled via the ib_client API */
1566 dev_err(queue
->ctrl
->ctrl
.device
,
1567 "Unexpected RDMA CM event (%d)\n", ev
->event
);
1568 nvme_rdma_error_recovery(queue
->ctrl
);
1573 queue
->cm_error
= cm_error
;
1574 complete(&queue
->cm_done
);
1580 static enum blk_eh_timer_return
1581 nvme_rdma_timeout(struct request
*rq
, bool reserved
)
1583 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1585 /* queue error recovery */
1586 nvme_rdma_error_recovery(req
->queue
->ctrl
);
1588 /* fail with DNR on cmd timeout */
1589 nvme_req(rq
)->status
= NVME_SC_ABORT_REQ
| NVME_SC_DNR
;
1591 return BLK_EH_HANDLED
;
1595 * We cannot accept any other command until the Connect command has completed.
1597 static inline blk_status_t
1598 nvme_rdma_queue_is_ready(struct nvme_rdma_queue
*queue
, struct request
*rq
)
1600 if (unlikely(!test_bit(NVME_RDMA_Q_LIVE
, &queue
->flags
))) {
1601 struct nvme_command
*cmd
= nvme_req(rq
)->cmd
;
1603 if (!blk_rq_is_passthrough(rq
) ||
1604 cmd
->common
.opcode
!= nvme_fabrics_command
||
1605 cmd
->fabrics
.fctype
!= nvme_fabrics_type_connect
) {
1607 * reconnecting state means transport disruption, which
1608 * can take a long time and even might fail permanently,
1609 * so we can't let incoming I/O be requeued forever.
1610 * fail it fast to allow upper layers a chance to
1613 if (queue
->ctrl
->ctrl
.state
== NVME_CTRL_RECONNECTING
)
1614 return BLK_STS_IOERR
;
1615 return BLK_STS_RESOURCE
; /* try again later */
1622 static blk_status_t
nvme_rdma_queue_rq(struct blk_mq_hw_ctx
*hctx
,
1623 const struct blk_mq_queue_data
*bd
)
1625 struct nvme_ns
*ns
= hctx
->queue
->queuedata
;
1626 struct nvme_rdma_queue
*queue
= hctx
->driver_data
;
1627 struct request
*rq
= bd
->rq
;
1628 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1629 struct nvme_rdma_qe
*sqe
= &req
->sqe
;
1630 struct nvme_command
*c
= sqe
->data
;
1632 struct ib_device
*dev
;
1636 WARN_ON_ONCE(rq
->tag
< 0);
1638 ret
= nvme_rdma_queue_is_ready(queue
, rq
);
1642 dev
= queue
->device
->dev
;
1643 ib_dma_sync_single_for_cpu(dev
, sqe
->dma
,
1644 sizeof(struct nvme_command
), DMA_TO_DEVICE
);
1646 ret
= nvme_setup_cmd(ns
, rq
, c
);
1650 blk_mq_start_request(rq
);
1652 err
= nvme_rdma_map_data(queue
, rq
, c
);
1653 if (unlikely(err
< 0)) {
1654 dev_err(queue
->ctrl
->ctrl
.device
,
1655 "Failed to map data (%d)\n", err
);
1656 nvme_cleanup_cmd(rq
);
1660 ib_dma_sync_single_for_device(dev
, sqe
->dma
,
1661 sizeof(struct nvme_command
), DMA_TO_DEVICE
);
1663 if (req_op(rq
) == REQ_OP_FLUSH
)
1665 err
= nvme_rdma_post_send(queue
, sqe
, req
->sge
, req
->num_sge
,
1666 req
->mr
->need_inval
? &req
->reg_wr
.wr
: NULL
, flush
);
1667 if (unlikely(err
)) {
1668 nvme_rdma_unmap_data(queue
, rq
);
1674 if (err
== -ENOMEM
|| err
== -EAGAIN
)
1675 return BLK_STS_RESOURCE
;
1676 return BLK_STS_IOERR
;
1679 static int nvme_rdma_poll(struct blk_mq_hw_ctx
*hctx
, unsigned int tag
)
1681 struct nvme_rdma_queue
*queue
= hctx
->driver_data
;
1682 struct ib_cq
*cq
= queue
->ib_cq
;
1686 while (ib_poll_cq(cq
, 1, &wc
) > 0) {
1687 struct ib_cqe
*cqe
= wc
.wr_cqe
;
1690 if (cqe
->done
== nvme_rdma_recv_done
)
1691 found
|= __nvme_rdma_recv_done(cq
, &wc
, tag
);
1700 static void nvme_rdma_complete_rq(struct request
*rq
)
1702 struct nvme_rdma_request
*req
= blk_mq_rq_to_pdu(rq
);
1704 nvme_rdma_unmap_data(req
->queue
, rq
);
1705 nvme_complete_rq(rq
);
1708 static int nvme_rdma_map_queues(struct blk_mq_tag_set
*set
)
1710 struct nvme_rdma_ctrl
*ctrl
= set
->driver_data
;
1712 return blk_mq_rdma_map_queues(set
, ctrl
->device
->dev
, 0);
1715 static const struct blk_mq_ops nvme_rdma_mq_ops
= {
1716 .queue_rq
= nvme_rdma_queue_rq
,
1717 .complete
= nvme_rdma_complete_rq
,
1718 .init_request
= nvme_rdma_init_request
,
1719 .exit_request
= nvme_rdma_exit_request
,
1720 .init_hctx
= nvme_rdma_init_hctx
,
1721 .poll
= nvme_rdma_poll
,
1722 .timeout
= nvme_rdma_timeout
,
1723 .map_queues
= nvme_rdma_map_queues
,
1726 static const struct blk_mq_ops nvme_rdma_admin_mq_ops
= {
1727 .queue_rq
= nvme_rdma_queue_rq
,
1728 .complete
= nvme_rdma_complete_rq
,
1729 .init_request
= nvme_rdma_init_request
,
1730 .exit_request
= nvme_rdma_exit_request
,
1731 .init_hctx
= nvme_rdma_init_admin_hctx
,
1732 .timeout
= nvme_rdma_timeout
,
1735 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl
*ctrl
, bool shutdown
)
1737 cancel_work_sync(&ctrl
->err_work
);
1738 cancel_delayed_work_sync(&ctrl
->reconnect_work
);
1740 if (ctrl
->ctrl
.queue_count
> 1) {
1741 nvme_stop_queues(&ctrl
->ctrl
);
1742 blk_mq_tagset_busy_iter(&ctrl
->tag_set
,
1743 nvme_cancel_request
, &ctrl
->ctrl
);
1744 nvme_rdma_destroy_io_queues(ctrl
, shutdown
);
1748 nvme_shutdown_ctrl(&ctrl
->ctrl
);
1750 nvme_disable_ctrl(&ctrl
->ctrl
, ctrl
->ctrl
.cap
);
1752 blk_mq_quiesce_queue(ctrl
->ctrl
.admin_q
);
1753 blk_mq_tagset_busy_iter(&ctrl
->admin_tag_set
,
1754 nvme_cancel_request
, &ctrl
->ctrl
);
1755 blk_mq_unquiesce_queue(ctrl
->ctrl
.admin_q
);
1756 nvme_rdma_destroy_admin_queue(ctrl
, shutdown
);
1759 static void nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl
*ctrl
)
1761 nvme_remove_namespaces(&ctrl
->ctrl
);
1762 nvme_rdma_shutdown_ctrl(ctrl
, true);
1763 nvme_uninit_ctrl(&ctrl
->ctrl
);
1764 nvme_put_ctrl(&ctrl
->ctrl
);
1767 static void nvme_rdma_del_ctrl_work(struct work_struct
*work
)
1769 struct nvme_rdma_ctrl
*ctrl
= container_of(work
,
1770 struct nvme_rdma_ctrl
, delete_work
);
1772 nvme_stop_ctrl(&ctrl
->ctrl
);
1773 nvme_rdma_remove_ctrl(ctrl
);
1776 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl
*ctrl
)
1778 if (!nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_DELETING
))
1781 if (!queue_work(nvme_wq
, &ctrl
->delete_work
))
1787 static int nvme_rdma_del_ctrl(struct nvme_ctrl
*nctrl
)
1789 struct nvme_rdma_ctrl
*ctrl
= to_rdma_ctrl(nctrl
);
1793 * Keep a reference until all work is flushed since
1794 * __nvme_rdma_del_ctrl can free the ctrl mem
1796 if (!kref_get_unless_zero(&ctrl
->ctrl
.kref
))
1798 ret
= __nvme_rdma_del_ctrl(ctrl
);
1800 flush_work(&ctrl
->delete_work
);
1801 nvme_put_ctrl(&ctrl
->ctrl
);
1805 static void nvme_rdma_reset_ctrl_work(struct work_struct
*work
)
1807 struct nvme_rdma_ctrl
*ctrl
=
1808 container_of(work
, struct nvme_rdma_ctrl
, ctrl
.reset_work
);
1812 nvme_stop_ctrl(&ctrl
->ctrl
);
1813 nvme_rdma_shutdown_ctrl(ctrl
, false);
1815 ret
= nvme_rdma_configure_admin_queue(ctrl
, false);
1819 if (ctrl
->ctrl
.queue_count
> 1) {
1820 ret
= nvme_rdma_configure_io_queues(ctrl
, false);
1825 changed
= nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_LIVE
);
1826 WARN_ON_ONCE(!changed
);
1828 nvme_start_ctrl(&ctrl
->ctrl
);
1833 dev_warn(ctrl
->ctrl
.device
, "Removing after reset failure\n");
1834 nvme_rdma_remove_ctrl(ctrl
);
1837 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops
= {
1839 .module
= THIS_MODULE
,
1840 .flags
= NVME_F_FABRICS
,
1841 .reg_read32
= nvmf_reg_read32
,
1842 .reg_read64
= nvmf_reg_read64
,
1843 .reg_write32
= nvmf_reg_write32
,
1844 .free_ctrl
= nvme_rdma_free_ctrl
,
1845 .submit_async_event
= nvme_rdma_submit_async_event
,
1846 .delete_ctrl
= nvme_rdma_del_ctrl
,
1847 .get_address
= nvmf_get_address
,
1850 static struct nvme_ctrl
*nvme_rdma_create_ctrl(struct device
*dev
,
1851 struct nvmf_ctrl_options
*opts
)
1853 struct nvme_rdma_ctrl
*ctrl
;
1858 ctrl
= kzalloc(sizeof(*ctrl
), GFP_KERNEL
);
1860 return ERR_PTR(-ENOMEM
);
1861 ctrl
->ctrl
.opts
= opts
;
1862 INIT_LIST_HEAD(&ctrl
->list
);
1864 if (opts
->mask
& NVMF_OPT_TRSVCID
)
1865 port
= opts
->trsvcid
;
1867 port
= __stringify(NVME_RDMA_IP_PORT
);
1869 ret
= inet_pton_with_scope(&init_net
, AF_UNSPEC
,
1870 opts
->traddr
, port
, &ctrl
->addr
);
1872 pr_err("malformed address passed: %s:%s\n", opts
->traddr
, port
);
1876 if (opts
->mask
& NVMF_OPT_HOST_TRADDR
) {
1877 ret
= inet_pton_with_scope(&init_net
, AF_UNSPEC
,
1878 opts
->host_traddr
, NULL
, &ctrl
->src_addr
);
1880 pr_err("malformed src address passed: %s\n",
1886 ret
= nvme_init_ctrl(&ctrl
->ctrl
, dev
, &nvme_rdma_ctrl_ops
,
1887 0 /* no quirks, we're perfect! */);
1891 INIT_DELAYED_WORK(&ctrl
->reconnect_work
,
1892 nvme_rdma_reconnect_ctrl_work
);
1893 INIT_WORK(&ctrl
->err_work
, nvme_rdma_error_recovery_work
);
1894 INIT_WORK(&ctrl
->delete_work
, nvme_rdma_del_ctrl_work
);
1895 INIT_WORK(&ctrl
->ctrl
.reset_work
, nvme_rdma_reset_ctrl_work
);
1897 ctrl
->ctrl
.queue_count
= opts
->nr_io_queues
+ 1; /* +1 for admin queue */
1898 ctrl
->ctrl
.sqsize
= opts
->queue_size
- 1;
1899 ctrl
->ctrl
.kato
= opts
->kato
;
1902 ctrl
->queues
= kcalloc(ctrl
->ctrl
.queue_count
, sizeof(*ctrl
->queues
),
1905 goto out_uninit_ctrl
;
1907 ret
= nvme_rdma_configure_admin_queue(ctrl
, true);
1909 goto out_kfree_queues
;
1911 /* sanity check icdoff */
1912 if (ctrl
->ctrl
.icdoff
) {
1913 dev_err(ctrl
->ctrl
.device
, "icdoff is not supported!\n");
1915 goto out_remove_admin_queue
;
1918 /* sanity check keyed sgls */
1919 if (!(ctrl
->ctrl
.sgls
& (1 << 20))) {
1920 dev_err(ctrl
->ctrl
.device
, "Mandatory keyed sgls are not support\n");
1922 goto out_remove_admin_queue
;
1925 if (opts
->queue_size
> ctrl
->ctrl
.maxcmd
) {
1926 /* warn if maxcmd is lower than queue_size */
1927 dev_warn(ctrl
->ctrl
.device
,
1928 "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1929 opts
->queue_size
, ctrl
->ctrl
.maxcmd
);
1930 opts
->queue_size
= ctrl
->ctrl
.maxcmd
;
1933 if (opts
->queue_size
> ctrl
->ctrl
.sqsize
+ 1) {
1934 /* warn if sqsize is lower than queue_size */
1935 dev_warn(ctrl
->ctrl
.device
,
1936 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1937 opts
->queue_size
, ctrl
->ctrl
.sqsize
+ 1);
1938 opts
->queue_size
= ctrl
->ctrl
.sqsize
+ 1;
1941 if (opts
->nr_io_queues
) {
1942 ret
= nvme_rdma_configure_io_queues(ctrl
, true);
1944 goto out_remove_admin_queue
;
1947 changed
= nvme_change_ctrl_state(&ctrl
->ctrl
, NVME_CTRL_LIVE
);
1948 WARN_ON_ONCE(!changed
);
1950 dev_info(ctrl
->ctrl
.device
, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1951 ctrl
->ctrl
.opts
->subsysnqn
, &ctrl
->addr
);
1953 kref_get(&ctrl
->ctrl
.kref
);
1955 mutex_lock(&nvme_rdma_ctrl_mutex
);
1956 list_add_tail(&ctrl
->list
, &nvme_rdma_ctrl_list
);
1957 mutex_unlock(&nvme_rdma_ctrl_mutex
);
1959 nvme_start_ctrl(&ctrl
->ctrl
);
1963 out_remove_admin_queue
:
1964 nvme_rdma_destroy_admin_queue(ctrl
, true);
1966 kfree(ctrl
->queues
);
1968 nvme_uninit_ctrl(&ctrl
->ctrl
);
1969 nvme_put_ctrl(&ctrl
->ctrl
);
1972 return ERR_PTR(ret
);
1975 return ERR_PTR(ret
);
1978 static struct nvmf_transport_ops nvme_rdma_transport
= {
1980 .required_opts
= NVMF_OPT_TRADDR
,
1981 .allowed_opts
= NVMF_OPT_TRSVCID
| NVMF_OPT_RECONNECT_DELAY
|
1982 NVMF_OPT_HOST_TRADDR
| NVMF_OPT_CTRL_LOSS_TMO
,
1983 .create_ctrl
= nvme_rdma_create_ctrl
,
1986 static void nvme_rdma_remove_one(struct ib_device
*ib_device
, void *client_data
)
1988 struct nvme_rdma_ctrl
*ctrl
;
1990 /* Delete all controllers using this device */
1991 mutex_lock(&nvme_rdma_ctrl_mutex
);
1992 list_for_each_entry(ctrl
, &nvme_rdma_ctrl_list
, list
) {
1993 if (ctrl
->device
->dev
!= ib_device
)
1995 dev_info(ctrl
->ctrl
.device
,
1996 "Removing ctrl: NQN \"%s\", addr %pISp\n",
1997 ctrl
->ctrl
.opts
->subsysnqn
, &ctrl
->addr
);
1998 __nvme_rdma_del_ctrl(ctrl
);
2000 mutex_unlock(&nvme_rdma_ctrl_mutex
);
2002 flush_workqueue(nvme_wq
);
2005 static struct ib_client nvme_rdma_ib_client
= {
2006 .name
= "nvme_rdma",
2007 .remove
= nvme_rdma_remove_one
2010 static int __init
nvme_rdma_init_module(void)
2014 ret
= ib_register_client(&nvme_rdma_ib_client
);
2018 ret
= nvmf_register_transport(&nvme_rdma_transport
);
2020 goto err_unreg_client
;
2025 ib_unregister_client(&nvme_rdma_ib_client
);
2029 static void __exit
nvme_rdma_cleanup_module(void)
2031 nvmf_unregister_transport(&nvme_rdma_transport
);
2032 ib_unregister_client(&nvme_rdma_ib_client
);
2035 module_init(nvme_rdma_init_module
);
2036 module_exit(nvme_rdma_cleanup_module
);
2038 MODULE_LICENSE("GPL v2");