]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blame - drivers/nvme/host/rdma.c
projects / mirror_ubuntu-artful-kernel.git / blame
? search:
summary | shortlog | log | commit | commitdiff | tree
blob | blame (incremental) | history | HEAD
blk-wbt: don't throttle discard or write zeroes
[mirror_ubuntu-artful-kernel.git] / drivers / nvme / host / rdma.c
CommitLineData
71102307
CH		 1/*
2 * NVMe over Fabrics RDMA host code.
3 * Copyright (c) 2015-2016 HGST, a Western Digital Company.
4 *
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.
8 *
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
12 * more details.
13 */
14#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
71102307
CH		 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>
71102307
CH		 20#include <linux/atomic.h>
21#include <linux/blk-mq.h>
22#include <linux/types.h>
23#include <linux/list.h>
24#include <linux/mutex.h>
25#include <linux/scatterlist.h>
26#include <linux/nvme.h>
71102307
CH		 27#include <asm/unaligned.h>
28
29#include <rdma/ib_verbs.h>
30#include <rdma/rdma_cm.h>
71102307
CH		 31#include <linux/nvme-rdma.h>
32
33#include "nvme.h"
34#include "fabrics.h"
35
36
37#define NVME_RDMA_CONNECT_TIMEOUT_MS 1000 /* 1 second */
38
39#define NVME_RDMA_MAX_SEGMENT_SIZE 0xffffff /* 24-bit SGL field */
40
41#define NVME_RDMA_MAX_SEGMENTS 256
42
43#define NVME_RDMA_MAX_INLINE_SEGMENTS 1
44
71102307
CH		 45/*
46 * We handle AEN commands ourselves and don't even let the
47 * block layer know about them.
48 */
49#define NVME_RDMA_NR_AEN_COMMANDS 1
50#define NVME_RDMA_AQ_BLKMQ_DEPTH \
51 (NVMF_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS)
52
53struct nvme_rdma_device {
54 struct ib_device *dev;
55 struct ib_pd *pd;
71102307
CH		 56 struct kref ref;
57 struct list_head entry;
58};
59
60struct nvme_rdma_qe {
61 struct ib_cqe cqe;
62 void *data;
63 u64 dma;
64};
65
66struct nvme_rdma_queue;
67struct nvme_rdma_request {
d49187e9 68 struct nvme_request req;
71102307
CH		 69 struct ib_mr *mr;
70 struct nvme_rdma_qe sqe;
71 struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
72 u32 num_sge;
73 int nents;
74 bool inline_data;
71102307
CH		 75 struct ib_reg_wr reg_wr;
76 struct ib_cqe reg_cqe;
77 struct nvme_rdma_queue *queue;
78 struct sg_table sg_table;
79 struct scatterlist first_sgl[];
80};
81
82enum nvme_rdma_queue_flags {
83 NVME_RDMA_Q_CONNECTED = (1 << 0),
f361e5a0 84 NVME_RDMA_IB_QUEUE_ALLOCATED = (1 << 1),
e89ca58f 85 NVME_RDMA_Q_DELETING = (1 << 2),
71102307
CH		 86};
87
88struct nvme_rdma_queue {
89 struct nvme_rdma_qe *rsp_ring;
90 u8 sig_count;
91 int queue_size;
92 size_t cmnd_capsule_len;
93 struct nvme_rdma_ctrl *ctrl;
94 struct nvme_rdma_device *device;
95 struct ib_cq *ib_cq;
96 struct ib_qp *qp;
97
98 unsigned long flags;
99 struct rdma_cm_id *cm_id;
100 int cm_error;
101 struct completion cm_done;
102};
103
104struct nvme_rdma_ctrl {
105 /* read and written in the hot path */
106 spinlock_t lock;
107
108 /* read only in the hot path */
109 struct nvme_rdma_queue *queues;
110 u32 queue_count;
111
112 /* other member variables */
71102307
CH		 113 struct blk_mq_tag_set tag_set;
114 struct work_struct delete_work;
115 struct work_struct reset_work;
116 struct work_struct err_work;
117
118 struct nvme_rdma_qe async_event_sqe;
119
120 int reconnect_delay;
121 struct delayed_work reconnect_work;
122
123 struct list_head list;
124
125 struct blk_mq_tag_set admin_tag_set;
126 struct nvme_rdma_device *device;
127
128 u64 cap;
129 u32 max_fr_pages;
130
131 union {
132 struct sockaddr addr;
133 struct sockaddr_in addr_in;
134 };
135
136 struct nvme_ctrl ctrl;
137};
138
139static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
140{
141 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
142}
143
144static LIST_HEAD(device_list);
145static DEFINE_MUTEX(device_list_mutex);
146
147static LIST_HEAD(nvme_rdma_ctrl_list);
148static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
149
150static struct workqueue_struct *nvme_rdma_wq;
151
152/*
153 * Disabling this option makes small I/O goes faster, but is fundamentally
154 * unsafe. With it turned off we will have to register a global rkey that
155 * allows read and write access to all physical memory.
156 */
157static bool register_always = true;
158module_param(register_always, bool, 0444);
159MODULE_PARM_DESC(register_always,
160 "Use memory registration even for contiguous memory regions");
161
162static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
163 struct rdma_cm_event *event);
164static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
71102307
CH		 165
166/* XXX: really should move to a generic header sooner or later.. */
167static inline void put_unaligned_le24(u32 val, u8 *p)
168{
169 *p++ = val;
170 *p++ = val >> 8;
171 *p++ = val >> 16;
172}
173
174static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
175{
176 return queue - queue->ctrl->queues;
177}
178
179static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
180{
181 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
182}
183
184static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
185 size_t capsule_size, enum dma_data_direction dir)
186{
187 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
188 kfree(qe->data);
189}
190
191static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
192 size_t capsule_size, enum dma_data_direction dir)
193{
194 qe->data = kzalloc(capsule_size, GFP_KERNEL);
195 if (!qe->data)
196 return -ENOMEM;
197
198 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
199 if (ib_dma_mapping_error(ibdev, qe->dma)) {
200 kfree(qe->data);
201 return -ENOMEM;
202 }
203
204 return 0;
205}
206
207static void nvme_rdma_free_ring(struct ib_device *ibdev,
208 struct nvme_rdma_qe *ring, size_t ib_queue_size,
209 size_t capsule_size, enum dma_data_direction dir)
210{
211 int i;
212
213 for (i = 0; i < ib_queue_size; i++)
214 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
215 kfree(ring);
216}
217
218static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
219 size_t ib_queue_size, size_t capsule_size,
220 enum dma_data_direction dir)
221{
222 struct nvme_rdma_qe *ring;
223 int i;
224
225 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
226 if (!ring)
227 return NULL;
228
229 for (i = 0; i < ib_queue_size; i++) {
230 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
231 goto out_free_ring;
232 }
233
234 return ring;
235
236out_free_ring:
237 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
238 return NULL;
239}
240
241static void nvme_rdma_qp_event(struct ib_event *event, void *context)
242{
27a4beef
MG		 243 pr_debug("QP event %s (%d)\n",
244 ib_event_msg(event->event), event->event);
245
71102307
CH		 246}
247
248static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
249{
250 wait_for_completion_interruptible_timeout(&queue->cm_done,
251 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
252 return queue->cm_error;
253}
254
255static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
256{
257 struct nvme_rdma_device *dev = queue->device;
258 struct ib_qp_init_attr init_attr;
259 int ret;
260
261 memset(&init_attr, 0, sizeof(init_attr));
262 init_attr.event_handler = nvme_rdma_qp_event;
263 /* +1 for drain */
264 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
265 /* +1 for drain */
266 init_attr.cap.max_recv_wr = queue->queue_size + 1;
267 init_attr.cap.max_recv_sge = 1;
268 init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
269 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
270 init_attr.qp_type = IB_QPT_RC;
271 init_attr.send_cq = queue->ib_cq;
272 init_attr.recv_cq = queue->ib_cq;
273
274 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
275
276 queue->qp = queue->cm_id->qp;
277 return ret;
278}
279
280static int nvme_rdma_reinit_request(void *data, struct request *rq)
281{
282 struct nvme_rdma_ctrl *ctrl = data;
283 struct nvme_rdma_device *dev = ctrl->device;
284 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
285 int ret = 0;
286
f5b7b559 287 if (!req->mr->need_inval)
71102307
CH		 288 goto out;
289
290 ib_dereg_mr(req->mr);
291
292 req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
293 ctrl->max_fr_pages);
294 if (IS_ERR(req->mr)) {
71102307 295 ret = PTR_ERR(req->mr);
458a9632 296 req->mr = NULL;
1bda18de 297 goto out;
71102307
CH		 298 }
299
f5b7b559 300 req->mr->need_inval = false;
71102307
CH		 301
302out:
303 return ret;
304}
305
306static void __nvme_rdma_exit_request(struct nvme_rdma_ctrl *ctrl,
307 struct request *rq, unsigned int queue_idx)
308{
309 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
310 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
311 struct nvme_rdma_device *dev = queue->device;
312
313 if (req->mr)
314 ib_dereg_mr(req->mr);
315
316 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
317 DMA_TO_DEVICE);
318}
319
320static void nvme_rdma_exit_request(void *data, struct request *rq,
321 unsigned int hctx_idx, unsigned int rq_idx)
322{
323 return __nvme_rdma_exit_request(data, rq, hctx_idx + 1);
324}
325
326static void nvme_rdma_exit_admin_request(void *data, struct request *rq,
327 unsigned int hctx_idx, unsigned int rq_idx)
328{
329 return __nvme_rdma_exit_request(data, rq, 0);
330}
331
332static int __nvme_rdma_init_request(struct nvme_rdma_ctrl *ctrl,
333 struct request *rq, unsigned int queue_idx)
334{
335 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
336 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
337 struct nvme_rdma_device *dev = queue->device;
338 struct ib_device *ibdev = dev->dev;
339 int ret;
340
341 BUG_ON(queue_idx >= ctrl->queue_count);
342
343 ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
344 DMA_TO_DEVICE);
345 if (ret)
346 return ret;
347
348 req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
349 ctrl->max_fr_pages);
350 if (IS_ERR(req->mr)) {
351 ret = PTR_ERR(req->mr);
352 goto out_free_qe;
353 }
354
355 req->queue = queue;
356
357 return 0;
358
359out_free_qe:
360 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
361 DMA_TO_DEVICE);
362 return -ENOMEM;
363}
364
365static int nvme_rdma_init_request(void *data, struct request *rq,
366 unsigned int hctx_idx, unsigned int rq_idx,
367 unsigned int numa_node)
368{
369 return __nvme_rdma_init_request(data, rq, hctx_idx + 1);
370}
371
372static int nvme_rdma_init_admin_request(void *data, struct request *rq,
373 unsigned int hctx_idx, unsigned int rq_idx,
374 unsigned int numa_node)
375{
376 return __nvme_rdma_init_request(data, rq, 0);
377}
378
379static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
380 unsigned int hctx_idx)
381{
382 struct nvme_rdma_ctrl *ctrl = data;
383 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
384
385 BUG_ON(hctx_idx >= ctrl->queue_count);
386
387 hctx->driver_data = queue;
388 return 0;
389}
390
391static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
392 unsigned int hctx_idx)
393{
394 struct nvme_rdma_ctrl *ctrl = data;
395 struct nvme_rdma_queue *queue = &ctrl->queues[0];
396
397 BUG_ON(hctx_idx != 0);
398
399 hctx->driver_data = queue;
400 return 0;
401}
402
403static void nvme_rdma_free_dev(struct kref *ref)
404{
405 struct nvme_rdma_device *ndev =
406 container_of(ref, struct nvme_rdma_device, ref);
407
408 mutex_lock(&device_list_mutex);
409 list_del(&ndev->entry);
410 mutex_unlock(&device_list_mutex);
411
71102307 412 ib_dealloc_pd(ndev->pd);
71102307
CH		 413 kfree(ndev);
414}
415
416static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
417{
418 kref_put(&dev->ref, nvme_rdma_free_dev);
419}
420
421static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
422{
423 return kref_get_unless_zero(&dev->ref);
424}
425
426static struct nvme_rdma_device *
427nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
428{
429 struct nvme_rdma_device *ndev;
430
431 mutex_lock(&device_list_mutex);
432 list_for_each_entry(ndev, &device_list, entry) {
433 if (ndev->dev->node_guid == cm_id->device->node_guid &&
434 nvme_rdma_dev_get(ndev))
435 goto out_unlock;
436 }
437
438 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
439 if (!ndev)
440 goto out_err;
441
442 ndev->dev = cm_id->device;
443 kref_init(&ndev->ref);
444
11975e01
CH		 445 ndev->pd = ib_alloc_pd(ndev->dev,
446 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
71102307
CH		 447 if (IS_ERR(ndev->pd))
448 goto out_free_dev;
449
71102307
CH		 450 if (!(ndev->dev->attrs.device_cap_flags &
451 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
452 dev_err(&ndev->dev->dev,
453 "Memory registrations not supported.\n");
11975e01 454 goto out_free_pd;
71102307
CH		 455 }
456
457 list_add(&ndev->entry, &device_list);
458out_unlock:
459 mutex_unlock(&device_list_mutex);
460 return ndev;
461
71102307
CH		 462out_free_pd:
463 ib_dealloc_pd(ndev->pd);
464out_free_dev:
465 kfree(ndev);
466out_err:
467 mutex_unlock(&device_list_mutex);
468 return NULL;
469}
470
471static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
472{
f361e5a0
SW		 473 struct nvme_rdma_device *dev;
474 struct ib_device *ibdev;
71102307 475
f361e5a0
SW		 476 if (!test_and_clear_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags))
477 return;
478
479 dev = queue->device;
480 ibdev = dev->dev;
71102307
CH		 481 rdma_destroy_qp(queue->cm_id);
482 ib_free_cq(queue->ib_cq);
483
484 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
485 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
486
487 nvme_rdma_dev_put(dev);
488}
489
490static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue,
491 struct nvme_rdma_device *dev)
492{
493 struct ib_device *ibdev = dev->dev;
494 const int send_wr_factor = 3; /* MR, SEND, INV */
495 const int cq_factor = send_wr_factor + 1; /* + RECV */
496 int comp_vector, idx = nvme_rdma_queue_idx(queue);
497
498 int ret;
499
500 queue->device = dev;
501
502 /*
503 * The admin queue is barely used once the controller is live, so don't
504 * bother to spread it out.
505 */
506 if (idx == 0)
507 comp_vector = 0;
508 else
509 comp_vector = idx % ibdev->num_comp_vectors;
510
511
512 /* +1 for ib_stop_cq */
513 queue->ib_cq = ib_alloc_cq(dev->dev, queue,
514 cq_factor * queue->queue_size + 1, comp_vector,
515 IB_POLL_SOFTIRQ);
516 if (IS_ERR(queue->ib_cq)) {
517 ret = PTR_ERR(queue->ib_cq);
518 goto out;
519 }
520
521 ret = nvme_rdma_create_qp(queue, send_wr_factor);
522 if (ret)
523 goto out_destroy_ib_cq;
524
525 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
526 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
527 if (!queue->rsp_ring) {
528 ret = -ENOMEM;
529 goto out_destroy_qp;
530 }
f361e5a0 531 set_bit(NVME_RDMA_IB_QUEUE_ALLOCATED, &queue->flags);
71102307
CH		 532
533 return 0;
534
535out_destroy_qp:
536 ib_destroy_qp(queue->qp);
537out_destroy_ib_cq:
538 ib_free_cq(queue->ib_cq);
539out:
540 return ret;
541}
542
543static int nvme_rdma_init_queue(struct nvme_rdma_ctrl *ctrl,
544 int idx, size_t queue_size)
545{
546 struct nvme_rdma_queue *queue;
547 int ret;
548
549 queue = &ctrl->queues[idx];
550 queue->ctrl = ctrl;
551 init_completion(&queue->cm_done);
552
553 if (idx > 0)
554 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
555 else
556 queue->cmnd_capsule_len = sizeof(struct nvme_command);
557
558 queue->queue_size = queue_size;
559
560 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
561 RDMA_PS_TCP, IB_QPT_RC);
562 if (IS_ERR(queue->cm_id)) {
563 dev_info(ctrl->ctrl.device,
564 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
565 return PTR_ERR(queue->cm_id);
566 }
567
568 queue->cm_error = -ETIMEDOUT;
569 ret = rdma_resolve_addr(queue->cm_id, NULL, &ctrl->addr,
570 NVME_RDMA_CONNECT_TIMEOUT_MS);
571 if (ret) {
572 dev_info(ctrl->ctrl.device,
573 "rdma_resolve_addr failed (%d).\n", ret);
574 goto out_destroy_cm_id;
575 }
576
577 ret = nvme_rdma_wait_for_cm(queue);
578 if (ret) {
579 dev_info(ctrl->ctrl.device,
580 "rdma_resolve_addr wait failed (%d).\n", ret);
581 goto out_destroy_cm_id;
582 }
583
3b4ac786 584 clear_bit(NVME_RDMA_Q_DELETING, &queue->flags);
71102307
CH		 585 set_bit(NVME_RDMA_Q_CONNECTED, &queue->flags);
586
587 return 0;
588
589out_destroy_cm_id:
f361e5a0 590 nvme_rdma_destroy_queue_ib(queue);
71102307
CH		 591 rdma_destroy_id(queue->cm_id);
592 return ret;
593}
594
595static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
596{
597 rdma_disconnect(queue->cm_id);
598 ib_drain_qp(queue->qp);
599}
600
601static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
602{
603 nvme_rdma_destroy_queue_ib(queue);
604 rdma_destroy_id(queue->cm_id);
605}
606
607static void nvme_rdma_stop_and_free_queue(struct nvme_rdma_queue *queue)
608{
e89ca58f 609 if (test_and_set_bit(NVME_RDMA_Q_DELETING, &queue->flags))
71102307
CH		 610 return;
611 nvme_rdma_stop_queue(queue);
612 nvme_rdma_free_queue(queue);
613}
614
615static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
616{
617 int i;
618
619 for (i = 1; i < ctrl->queue_count; i++)
620 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
621}
622
623static int nvme_rdma_connect_io_queues(struct nvme_rdma_ctrl *ctrl)
624{
625 int i, ret = 0;
626
627 for (i = 1; i < ctrl->queue_count; i++) {
628 ret = nvmf_connect_io_queue(&ctrl->ctrl, i);
629 if (ret)
630 break;
631 }
632
633 return ret;
634}
635
636static int nvme_rdma_init_io_queues(struct nvme_rdma_ctrl *ctrl)
637{
638 int i, ret;
639
640 for (i = 1; i < ctrl->queue_count; i++) {
c5af8654
JF		 641 ret = nvme_rdma_init_queue(ctrl, i,
642 ctrl->ctrl.opts->queue_size);
71102307
CH		 643 if (ret) {
644 dev_info(ctrl->ctrl.device,
645 "failed to initialize i/o queue: %d\n", ret);
646 goto out_free_queues;
647 }
648 }
649
650 return 0;
651
652out_free_queues:
f361e5a0 653 for (i--; i >= 1; i--)
71102307
CH		 654 nvme_rdma_stop_and_free_queue(&ctrl->queues[i]);
655
656 return ret;
657}
658
659static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl)
660{
661 nvme_rdma_free_qe(ctrl->queues[0].device->dev, &ctrl->async_event_sqe,
662 sizeof(struct nvme_command), DMA_TO_DEVICE);
663 nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
664 blk_cleanup_queue(ctrl->ctrl.admin_q);
665 blk_mq_free_tag_set(&ctrl->admin_tag_set);
666 nvme_rdma_dev_put(ctrl->device);
667}
668
669static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
670{
671 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
672
673 if (list_empty(&ctrl->list))
674 goto free_ctrl;
675
676 mutex_lock(&nvme_rdma_ctrl_mutex);
677 list_del(&ctrl->list);
678 mutex_unlock(&nvme_rdma_ctrl_mutex);
679
71102307
CH		 680 kfree(ctrl->queues);
681 nvmf_free_options(nctrl->opts);
682free_ctrl:
683 kfree(ctrl);
684}
685
686static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
687{
688 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
689 struct nvme_rdma_ctrl, reconnect_work);
690 bool changed;
691 int ret;
692
693 if (ctrl->queue_count > 1) {
694 nvme_rdma_free_io_queues(ctrl);
695
696 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
697 if (ret)
698 goto requeue;
699 }
700
701 nvme_rdma_stop_and_free_queue(&ctrl->queues[0]);
702
703 ret = blk_mq_reinit_tagset(&ctrl->admin_tag_set);
704 if (ret)
705 goto requeue;
706
707 ret = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
708 if (ret)
709 goto requeue;
710
711 blk_mq_start_stopped_hw_queues(ctrl->ctrl.admin_q, true);
712
713 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
714 if (ret)
715 goto stop_admin_q;
716
717 ret = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
718 if (ret)
719 goto stop_admin_q;
720
721 nvme_start_keep_alive(&ctrl->ctrl);
722
723 if (ctrl->queue_count > 1) {
724 ret = nvme_rdma_init_io_queues(ctrl);
725 if (ret)
726 goto stop_admin_q;
727
728 ret = nvme_rdma_connect_io_queues(ctrl);
729 if (ret)
730 goto stop_admin_q;
731 }
732
733 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
734 WARN_ON_ONCE(!changed);
735
5f372eb3 736 if (ctrl->queue_count > 1) {
71102307 737 nvme_start_queues(&ctrl->ctrl);
5f372eb3 738 nvme_queue_scan(&ctrl->ctrl);
3ef1b4b2 739 nvme_queue_async_events(&ctrl->ctrl);
5f372eb3 740 }
71102307
CH		 741
742 dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
743
744 return;
745
746stop_admin_q:
747 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
748requeue:
749 /* Make sure we are not resetting/deleting */
750 if (ctrl->ctrl.state == NVME_CTRL_RECONNECTING) {
751 dev_info(ctrl->ctrl.device,
752 "Failed reconnect attempt, requeueing...\n");
753 queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
754 ctrl->reconnect_delay * HZ);
755 }
756}
757
758static void nvme_rdma_error_recovery_work(struct work_struct *work)
759{
760 struct nvme_rdma_ctrl *ctrl = container_of(work,
761 struct nvme_rdma_ctrl, err_work);
e89ca58f 762 int i;
71102307
CH		 763
764 nvme_stop_keep_alive(&ctrl->ctrl);
e89ca58f
SG		 765
766 for (i = 0; i < ctrl->queue_count; i++)
767 clear_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[i].flags);
768
71102307
CH		 769 if (ctrl->queue_count > 1)
770 nvme_stop_queues(&ctrl->ctrl);
771 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
772
773 /* We must take care of fastfail/requeue all our inflight requests */
774 if (ctrl->queue_count > 1)
775 blk_mq_tagset_busy_iter(&ctrl->tag_set,
776 nvme_cancel_request, &ctrl->ctrl);
777 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
778 nvme_cancel_request, &ctrl->ctrl);
779
780 dev_info(ctrl->ctrl.device, "reconnecting in %d seconds\n",
781 ctrl->reconnect_delay);
782
783 queue_delayed_work(nvme_rdma_wq, &ctrl->reconnect_work,
784 ctrl->reconnect_delay * HZ);
785}
786
787static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
788{
789 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
790 return;
791
792 queue_work(nvme_rdma_wq, &ctrl->err_work);
793}
794
795static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
796 const char *op)
797{
798 struct nvme_rdma_queue *queue = cq->cq_context;
799 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
800
801 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
802 dev_info(ctrl->ctrl.device,
803 "%s for CQE 0x%p failed with status %s (%d)\n",
804 op, wc->wr_cqe,
805 ib_wc_status_msg(wc->status), wc->status);
806 nvme_rdma_error_recovery(ctrl);
807}
808
809static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
810{
811 if (unlikely(wc->status != IB_WC_SUCCESS))
812 nvme_rdma_wr_error(cq, wc, "MEMREG");
813}
814
815static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
816{
817 if (unlikely(wc->status != IB_WC_SUCCESS))
818 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
819}
820
821static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
822 struct nvme_rdma_request *req)
823{
824 struct ib_send_wr *bad_wr;
825 struct ib_send_wr wr = {
826 .opcode = IB_WR_LOCAL_INV,
827 .next = NULL,
828 .num_sge = 0,
829 .send_flags = 0,
830 .ex.invalidate_rkey = req->mr->rkey,
831 };
832
833 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
834 wr.wr_cqe = &req->reg_cqe;
835
836 return ib_post_send(queue->qp, &wr, &bad_wr);
837}
838
839static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
840 struct request *rq)
841{
842 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
843 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
844 struct nvme_rdma_device *dev = queue->device;
845 struct ib_device *ibdev = dev->dev;
846 int res;
847
848 if (!blk_rq_bytes(rq))
849 return;
850
f5b7b559 851 if (req->mr->need_inval) {
71102307
CH		 852 res = nvme_rdma_inv_rkey(queue, req);
853 if (res < 0) {
854 dev_err(ctrl->ctrl.device,
855 "Queueing INV WR for rkey %#x failed (%d)\n",
856 req->mr->rkey, res);
857 nvme_rdma_error_recovery(queue->ctrl);
858 }
859 }
860
861 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
862 req->nents, rq_data_dir(rq) ==
863 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
864
865 nvme_cleanup_cmd(rq);
866 sg_free_table_chained(&req->sg_table, true);
867}
868
869static int nvme_rdma_set_sg_null(struct nvme_command *c)
870{
871 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
872
873 sg->addr = 0;
874 put_unaligned_le24(0, sg->length);
875 put_unaligned_le32(0, sg->key);
876 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
877 return 0;
878}
879
880static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
881 struct nvme_rdma_request *req, struct nvme_command *c)
882{
883 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
884
885 req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
886 req->sge[1].length = sg_dma_len(req->sg_table.sgl);
887 req->sge[1].lkey = queue->device->pd->local_dma_lkey;
888
889 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
890 sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
891 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
892
893 req->inline_data = true;
894 req->num_sge++;
895 return 0;
896}
897
898static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
899 struct nvme_rdma_request *req, struct nvme_command *c)
900{
901 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
902
903 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
904 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
11975e01 905 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
71102307
CH		 906 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
907 return 0;
908}
909
910static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
911 struct nvme_rdma_request *req, struct nvme_command *c,
912 int count)
913{
914 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
915 int nr;
916
917 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, PAGE_SIZE);
918 if (nr < count) {
919 if (nr < 0)
920 return nr;
921 return -EINVAL;
922 }
923
924 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
925
926 req->reg_cqe.done = nvme_rdma_memreg_done;
927 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
928 req->reg_wr.wr.opcode = IB_WR_REG_MR;
929 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
930 req->reg_wr.wr.num_sge = 0;
931 req->reg_wr.mr = req->mr;
932 req->reg_wr.key = req->mr->rkey;
933 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
934 IB_ACCESS_REMOTE_READ |
935 IB_ACCESS_REMOTE_WRITE;
936
f5b7b559 937 req->mr->need_inval = true;
71102307
CH		 938
939 sg->addr = cpu_to_le64(req->mr->iova);
940 put_unaligned_le24(req->mr->length, sg->length);
941 put_unaligned_le32(req->mr->rkey, sg->key);
942 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
943 NVME_SGL_FMT_INVALIDATE;
944
945 return 0;
946}
947
948static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
949 struct request *rq, unsigned int map_len,
950 struct nvme_command *c)
951{
952 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
953 struct nvme_rdma_device *dev = queue->device;
954 struct ib_device *ibdev = dev->dev;
955 int nents, count;
956 int ret;
957
958 req->num_sge = 1;
959 req->inline_data = false;
f5b7b559 960 req->mr->need_inval = false;
71102307
CH		 961
962 c->common.flags |= NVME_CMD_SGL_METABUF;
963
964 if (!blk_rq_bytes(rq))
965 return nvme_rdma_set_sg_null(c);
966
967 req->sg_table.sgl = req->first_sgl;
968 ret = sg_alloc_table_chained(&req->sg_table, rq->nr_phys_segments,
969 req->sg_table.sgl);
970 if (ret)
971 return -ENOMEM;
972
973 nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
974 BUG_ON(nents > rq->nr_phys_segments);
975 req->nents = nents;
976
977 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, nents,
978 rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
979 if (unlikely(count <= 0)) {
980 sg_free_table_chained(&req->sg_table, true);
981 return -EIO;
982 }
983
984 if (count == 1) {
985 if (rq_data_dir(rq) == WRITE &&
986 map_len <= nvme_rdma_inline_data_size(queue) &&
987 nvme_rdma_queue_idx(queue))
988 return nvme_rdma_map_sg_inline(queue, req, c);
989
11975e01 990 if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
71102307
CH		 991 return nvme_rdma_map_sg_single(queue, req, c);
992 }
993
994 return nvme_rdma_map_sg_fr(queue, req, c, count);
995}
996
997static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
998{
999 if (unlikely(wc->status != IB_WC_SUCCESS))
1000 nvme_rdma_wr_error(cq, wc, "SEND");
1001}
1002
1003static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1004 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1005 struct ib_send_wr *first, bool flush)
1006{
1007 struct ib_send_wr wr, *bad_wr;
1008 int ret;
1009
1010 sge->addr = qe->dma;
1011 sge->length = sizeof(struct nvme_command),
1012 sge->lkey = queue->device->pd->local_dma_lkey;
1013
1014 qe->cqe.done = nvme_rdma_send_done;
1015
1016 wr.next = NULL;
1017 wr.wr_cqe = &qe->cqe;
1018 wr.sg_list = sge;
1019 wr.num_sge = num_sge;
1020 wr.opcode = IB_WR_SEND;
1021 wr.send_flags = 0;
1022
1023 /*
1024 * Unsignalled send completions are another giant desaster in the
1025 * IB Verbs spec: If we don't regularly post signalled sends
1026 * the send queue will fill up and only a QP reset will rescue us.
1027 * Would have been way to obvious to handle this in hardware or
1028 * at least the RDMA stack..
1029 *
1030 * This messy and racy code sniplet is copy and pasted from the iSER
1031 * initiator, and the magic '32' comes from there as well.
1032 *
1033 * Always signal the flushes. The magic request used for the flush
1034 * sequencer is not allocated in our driver's tagset and it's
1035 * triggered to be freed by blk_cleanup_queue(). So we need to
1036 * always mark it as signaled to ensure that the "wr_cqe", which is
1037 * embeded in request's payload, is not freed when __ib_process_cq()
1038 * calls wr_cqe->done().
1039 */
1040 if ((++queue->sig_count % 32) == 0 || flush)
1041 wr.send_flags |= IB_SEND_SIGNALED;
1042
1043 if (first)
1044 first->next = &wr;
1045 else
1046 first = &wr;
1047
1048 ret = ib_post_send(queue->qp, first, &bad_wr);
1049 if (ret) {
1050 dev_err(queue->ctrl->ctrl.device,
1051 "%s failed with error code %d\n", __func__, ret);
1052 }
1053 return ret;
1054}
1055
1056static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1057 struct nvme_rdma_qe *qe)
1058{
1059 struct ib_recv_wr wr, *bad_wr;
1060 struct ib_sge list;
1061 int ret;
1062
1063 list.addr = qe->dma;
1064 list.length = sizeof(struct nvme_completion);
1065 list.lkey = queue->device->pd->local_dma_lkey;
1066
1067 qe->cqe.done = nvme_rdma_recv_done;
1068
1069 wr.next = NULL;
1070 wr.wr_cqe = &qe->cqe;
1071 wr.sg_list = &list;
1072 wr.num_sge = 1;
1073
1074 ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1075 if (ret) {
1076 dev_err(queue->ctrl->ctrl.device,
1077 "%s failed with error code %d\n", __func__, ret);
1078 }
1079 return ret;
1080}
1081
1082static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1083{
1084 u32 queue_idx = nvme_rdma_queue_idx(queue);
1085
1086 if (queue_idx == 0)
1087 return queue->ctrl->admin_tag_set.tags[queue_idx];
1088 return queue->ctrl->tag_set.tags[queue_idx - 1];
1089}
1090
1091static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1092{
1093 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1094 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1095 struct ib_device *dev = queue->device->dev;
1096 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1097 struct nvme_command *cmd = sqe->data;
1098 struct ib_sge sge;
1099 int ret;
1100
1101 if (WARN_ON_ONCE(aer_idx != 0))
1102 return;
1103
1104 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1105
1106 memset(cmd, 0, sizeof(*cmd));
1107 cmd->common.opcode = nvme_admin_async_event;
1108 cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1109 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1110 nvme_rdma_set_sg_null(cmd);
1111
1112 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1113 DMA_TO_DEVICE);
1114
1115 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1116 WARN_ON_ONCE(ret);
1117}
1118
1119static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1120 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1121{
71102307
CH		 1122 struct request *rq;
1123 struct nvme_rdma_request *req;
1124 int ret = 0;
1125
71102307
CH		 1126 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1127 if (!rq) {
1128 dev_err(queue->ctrl->ctrl.device,
1129 "tag 0x%x on QP %#x not found\n",
1130 cqe->command_id, queue->qp->qp_num);
1131 nvme_rdma_error_recovery(queue->ctrl);
1132 return ret;
1133 }
1134 req = blk_mq_rq_to_pdu(rq);
1135
71102307
CH		 1136 if (rq->tag == tag)
1137 ret = 1;
1138
1139 if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1140 wc->ex.invalidate_rkey == req->mr->rkey)
f5b7b559 1141 req->mr->need_inval = false;
71102307 1142
d49187e9
CH		 1143 req->req.result = cqe->result;
1144 blk_mq_complete_request(rq, le16_to_cpu(cqe->status) >> 1);
71102307
CH		 1145 return ret;
1146}
1147
1148static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1149{
1150 struct nvme_rdma_qe *qe =
1151 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1152 struct nvme_rdma_queue *queue = cq->cq_context;
1153 struct ib_device *ibdev = queue->device->dev;
1154 struct nvme_completion *cqe = qe->data;
1155 const size_t len = sizeof(struct nvme_completion);
1156 int ret = 0;
1157
1158 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1159 nvme_rdma_wr_error(cq, wc, "RECV");
1160 return 0;
1161 }
1162
1163 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1164 /*
1165 * AEN requests are special as they don't time out and can
1166 * survive any kind of queue freeze and often don't respond to
1167 * aborts. We don't even bother to allocate a struct request
1168 * for them but rather special case them here.
1169 */
1170 if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1171 cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
7bf58533
CH		 1172 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1173 &cqe->result);
71102307
CH		 1174 else
1175 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1176 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1177
1178 nvme_rdma_post_recv(queue, qe);
1179 return ret;
1180}
1181
1182static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1183{
1184 __nvme_rdma_recv_done(cq, wc, -1);
1185}
1186
1187static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1188{
1189 int ret, i;
1190
1191 for (i = 0; i < queue->queue_size; i++) {
1192 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1193 if (ret)
1194 goto out_destroy_queue_ib;
1195 }
1196
1197 return 0;
1198
1199out_destroy_queue_ib:
1200 nvme_rdma_destroy_queue_ib(queue);
1201 return ret;
1202}
1203
1204static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1205 struct rdma_cm_event *ev)
1206{
1207 if (ev->param.conn.private_data_len) {
1208 struct nvme_rdma_cm_rej *rej =
1209 (struct nvme_rdma_cm_rej *)ev->param.conn.private_data;
1210
1211 dev_err(queue->ctrl->ctrl.device,
1212 "Connect rejected, status %d.", le16_to_cpu(rej->sts));
1213 /* XXX: Think of something clever to do here... */
1214 } else {
1215 dev_err(queue->ctrl->ctrl.device,
1216 "Connect rejected, no private data.\n");
1217 }
1218
1219 return -ECONNRESET;
1220}
1221
1222static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1223{
1224 struct nvme_rdma_device *dev;
1225 int ret;
1226
1227 dev = nvme_rdma_find_get_device(queue->cm_id);
1228 if (!dev) {
1229 dev_err(queue->cm_id->device->dma_device,
1230 "no client data found!\n");
1231 return -ECONNREFUSED;
1232 }
1233
1234 ret = nvme_rdma_create_queue_ib(queue, dev);
1235 if (ret) {
1236 nvme_rdma_dev_put(dev);
1237 goto out;
1238 }
1239
1240 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1241 if (ret) {
1242 dev_err(queue->ctrl->ctrl.device,
1243 "rdma_resolve_route failed (%d).\n",
1244 queue->cm_error);
1245 goto out_destroy_queue;
1246 }
1247
1248 return 0;
1249
1250out_destroy_queue:
1251 nvme_rdma_destroy_queue_ib(queue);
1252out:
1253 return ret;
1254}
1255
1256static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1257{
1258 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1259 struct rdma_conn_param param = { };
0b857b44 1260 struct nvme_rdma_cm_req priv = { };
71102307
CH		 1261 int ret;
1262
1263 param.qp_num = queue->qp->qp_num;
1264 param.flow_control = 1;
1265
1266 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
2ac17c28
SG		 1267 /* maximum retry count */
1268 param.retry_count = 7;
71102307
CH		 1269 param.rnr_retry_count = 7;
1270 param.private_data = &priv;
1271 param.private_data_len = sizeof(priv);
1272
1273 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1274 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
f994d9dc
JF		 1275 /*
1276 * set the admin queue depth to the minimum size
1277 * specified by the Fabrics standard.
1278 */
1279 if (priv.qid == 0) {
1280 priv.hrqsize = cpu_to_le16(NVMF_AQ_DEPTH);
1281 priv.hsqsize = cpu_to_le16(NVMF_AQ_DEPTH - 1);
1282 } else {
c5af8654
JF		 1283 /*
1284 * current interpretation of the fabrics spec
1285 * is at minimum you make hrqsize sqsize+1, or a
1286 * 1's based representation of sqsize.
1287 */
f994d9dc 1288 priv.hrqsize = cpu_to_le16(queue->queue_size);
c5af8654 1289 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
f994d9dc 1290 }
71102307
CH		 1291
1292 ret = rdma_connect(queue->cm_id, &param);
1293 if (ret) {
1294 dev_err(ctrl->ctrl.device,
1295 "rdma_connect failed (%d).\n", ret);
1296 goto out_destroy_queue_ib;
1297 }
1298
1299 return 0;
1300
1301out_destroy_queue_ib:
1302 nvme_rdma_destroy_queue_ib(queue);
1303 return ret;
1304}
1305
71102307
CH		 1306static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1307 struct rdma_cm_event *ev)
1308{
1309 struct nvme_rdma_queue *queue = cm_id->context;
1310 int cm_error = 0;
1311
1312 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1313 rdma_event_msg(ev->event), ev->event,
1314 ev->status, cm_id);
1315
1316 switch (ev->event) {
1317 case RDMA_CM_EVENT_ADDR_RESOLVED:
1318 cm_error = nvme_rdma_addr_resolved(queue);
1319 break;
1320 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1321 cm_error = nvme_rdma_route_resolved(queue);
1322 break;
1323 case RDMA_CM_EVENT_ESTABLISHED:
1324 queue->cm_error = nvme_rdma_conn_established(queue);
1325 /* complete cm_done regardless of success/failure */
1326 complete(&queue->cm_done);
1327 return 0;
1328 case RDMA_CM_EVENT_REJECTED:
1329 cm_error = nvme_rdma_conn_rejected(queue, ev);
1330 break;
1331 case RDMA_CM_EVENT_ADDR_ERROR:
1332 case RDMA_CM_EVENT_ROUTE_ERROR:
1333 case RDMA_CM_EVENT_CONNECT_ERROR:
1334 case RDMA_CM_EVENT_UNREACHABLE:
1335 dev_dbg(queue->ctrl->ctrl.device,
1336 "CM error event %d\n", ev->event);
1337 cm_error = -ECONNRESET;
1338 break;
1339 case RDMA_CM_EVENT_DISCONNECTED:
1340 case RDMA_CM_EVENT_ADDR_CHANGE:
1341 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1342 dev_dbg(queue->ctrl->ctrl.device,
1343 "disconnect received - connection closed\n");
1344 nvme_rdma_error_recovery(queue->ctrl);
1345 break;
1346 case RDMA_CM_EVENT_DEVICE_REMOVAL:
e87a911f
SW		 1347 /* device removal is handled via the ib_client API */
1348 break;
71102307
CH		 1349 default:
1350 dev_err(queue->ctrl->ctrl.device,
1351 "Unexpected RDMA CM event (%d)\n", ev->event);
1352 nvme_rdma_error_recovery(queue->ctrl);
1353 break;
1354 }
1355
1356 if (cm_error) {
1357 queue->cm_error = cm_error;
1358 complete(&queue->cm_done);
1359 }
1360
1361 return 0;
1362}
1363
1364static enum blk_eh_timer_return
1365nvme_rdma_timeout(struct request *rq, bool reserved)
1366{
1367 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1368
1369 /* queue error recovery */
1370 nvme_rdma_error_recovery(req->queue->ctrl);
1371
1372 /* fail with DNR on cmd timeout */
1373 rq->errors = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1374
1375 return BLK_EH_HANDLED;
1376}
1377
1378static int nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1379 const struct blk_mq_queue_data *bd)
1380{
1381 struct nvme_ns *ns = hctx->queue->queuedata;
1382 struct nvme_rdma_queue *queue = hctx->driver_data;
1383 struct request *rq = bd->rq;
1384 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1385 struct nvme_rdma_qe *sqe = &req->sqe;
1386 struct nvme_command *c = sqe->data;
1387 bool flush = false;
1388 struct ib_device *dev;
1389 unsigned int map_len;
1390 int ret;
1391
1392 WARN_ON_ONCE(rq->tag < 0);
1393
1394 dev = queue->device->dev;
1395 ib_dma_sync_single_for_cpu(dev, sqe->dma,
1396 sizeof(struct nvme_command), DMA_TO_DEVICE);
1397
1398 ret = nvme_setup_cmd(ns, rq, c);
bac0000a 1399 if (ret != BLK_MQ_RQ_QUEUE_OK)
71102307
CH		 1400 return ret;
1401
71102307
CH		 1402 blk_mq_start_request(rq);
1403
1404 map_len = nvme_map_len(rq);
1405 ret = nvme_rdma_map_data(queue, rq, map_len, c);
1406 if (ret < 0) {
1407 dev_err(queue->ctrl->ctrl.device,
1408 "Failed to map data (%d)\n", ret);
1409 nvme_cleanup_cmd(rq);
1410 goto err;
1411 }
1412
1413 ib_dma_sync_single_for_device(dev, sqe->dma,
1414 sizeof(struct nvme_command), DMA_TO_DEVICE);
1415
1416 if (rq->cmd_type == REQ_TYPE_FS && req_op(rq) == REQ_OP_FLUSH)
1417 flush = true;
1418 ret = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
f5b7b559 1419 req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
71102307
CH		 1420 if (ret) {
1421 nvme_rdma_unmap_data(queue, rq);
1422 goto err;
1423 }
1424
1425 return BLK_MQ_RQ_QUEUE_OK;
1426err:
1427 return (ret == -ENOMEM || ret == -EAGAIN) ?
1428 BLK_MQ_RQ_QUEUE_BUSY : BLK_MQ_RQ_QUEUE_ERROR;
1429}
1430
1431static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1432{
1433 struct nvme_rdma_queue *queue = hctx->driver_data;
1434 struct ib_cq *cq = queue->ib_cq;
1435 struct ib_wc wc;
1436 int found = 0;
1437
1438 ib_req_notify_cq(cq, IB_CQ_NEXT_COMP);
1439 while (ib_poll_cq(cq, 1, &wc) > 0) {
1440 struct ib_cqe *cqe = wc.wr_cqe;
1441
1442 if (cqe) {
1443 if (cqe->done == nvme_rdma_recv_done)
1444 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1445 else
1446 cqe->done(cq, &wc);
1447 }
1448 }
1449
1450 return found;
1451}
1452
1453static void nvme_rdma_complete_rq(struct request *rq)
1454{
1455 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1456 struct nvme_rdma_queue *queue = req->queue;
1457 int error = 0;
1458
1459 nvme_rdma_unmap_data(queue, rq);
1460
1461 if (unlikely(rq->errors)) {
1462 if (nvme_req_needs_retry(rq, rq->errors)) {
1463 nvme_requeue_req(rq);
1464 return;
1465 }
1466
1467 if (rq->cmd_type == REQ_TYPE_DRV_PRIV)
1468 error = rq->errors;
1469 else
1470 error = nvme_error_status(rq->errors);
1471 }
1472
1473 blk_mq_end_request(rq, error);
1474}
1475
1476static struct blk_mq_ops nvme_rdma_mq_ops = {
1477 .queue_rq = nvme_rdma_queue_rq,
1478 .complete = nvme_rdma_complete_rq,
71102307
CH		 1479 .init_request = nvme_rdma_init_request,
1480 .exit_request = nvme_rdma_exit_request,
1481 .reinit_request = nvme_rdma_reinit_request,
1482 .init_hctx = nvme_rdma_init_hctx,
1483 .poll = nvme_rdma_poll,
1484 .timeout = nvme_rdma_timeout,
1485};
1486
1487static struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1488 .queue_rq = nvme_rdma_queue_rq,
1489 .complete = nvme_rdma_complete_rq,
71102307
CH		 1490 .init_request = nvme_rdma_init_admin_request,
1491 .exit_request = nvme_rdma_exit_admin_request,
1492 .reinit_request = nvme_rdma_reinit_request,
1493 .init_hctx = nvme_rdma_init_admin_hctx,
1494 .timeout = nvme_rdma_timeout,
1495};
1496
1497static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl)
1498{
1499 int error;
1500
1501 error = nvme_rdma_init_queue(ctrl, 0, NVMF_AQ_DEPTH);
1502 if (error)
1503 return error;
1504
1505 ctrl->device = ctrl->queues[0].device;
1506
1507 /*
1508 * We need a reference on the device as long as the tag_set is alive,
1509 * as the MRs in the request structures need a valid ib_device.
1510 */
1511 error = -EINVAL;
1512 if (!nvme_rdma_dev_get(ctrl->device))
1513 goto out_free_queue;
1514
1515 ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
1516 ctrl->device->dev->attrs.max_fast_reg_page_list_len);
1517
1518 memset(&ctrl->admin_tag_set, 0, sizeof(ctrl->admin_tag_set));
1519 ctrl->admin_tag_set.ops = &nvme_rdma_admin_mq_ops;
1520 ctrl->admin_tag_set.queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
1521 ctrl->admin_tag_set.reserved_tags = 2; /* connect + keep-alive */
1522 ctrl->admin_tag_set.numa_node = NUMA_NO_NODE;
1523 ctrl->admin_tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1524 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1525 ctrl->admin_tag_set.driver_data = ctrl;
1526 ctrl->admin_tag_set.nr_hw_queues = 1;
1527 ctrl->admin_tag_set.timeout = ADMIN_TIMEOUT;
1528
1529 error = blk_mq_alloc_tag_set(&ctrl->admin_tag_set);
1530 if (error)
1531 goto out_put_dev;
1532
1533 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
1534 if (IS_ERR(ctrl->ctrl.admin_q)) {
1535 error = PTR_ERR(ctrl->ctrl.admin_q);
1536 goto out_free_tagset;
1537 }
1538
1539 error = nvmf_connect_admin_queue(&ctrl->ctrl);
1540 if (error)
1541 goto out_cleanup_queue;
1542
1543 error = nvmf_reg_read64(&ctrl->ctrl, NVME_REG_CAP, &ctrl->cap);
1544 if (error) {
1545 dev_err(ctrl->ctrl.device,
1546 "prop_get NVME_REG_CAP failed\n");
1547 goto out_cleanup_queue;
1548 }
1549
1550 ctrl->ctrl.sqsize =
1551 min_t(int, NVME_CAP_MQES(ctrl->cap) + 1, ctrl->ctrl.sqsize);
1552
1553 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->cap);
1554 if (error)
1555 goto out_cleanup_queue;
1556
1557 ctrl->ctrl.max_hw_sectors =
1558 (ctrl->max_fr_pages - 1) << (PAGE_SHIFT - 9);
1559
1560 error = nvme_init_identify(&ctrl->ctrl);
1561 if (error)
1562 goto out_cleanup_queue;
1563
1564 error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
1565 &ctrl->async_event_sqe, sizeof(struct nvme_command),
1566 DMA_TO_DEVICE);
1567 if (error)
1568 goto out_cleanup_queue;
1569
1570 nvme_start_keep_alive(&ctrl->ctrl);
1571
1572 return 0;
1573
1574out_cleanup_queue:
1575 blk_cleanup_queue(ctrl->ctrl.admin_q);
1576out_free_tagset:
1577 /* disconnect and drain the queue before freeing the tagset */
1578 nvme_rdma_stop_queue(&ctrl->queues[0]);
1579 blk_mq_free_tag_set(&ctrl->admin_tag_set);
1580out_put_dev:
1581 nvme_rdma_dev_put(ctrl->device);
1582out_free_queue:
1583 nvme_rdma_free_queue(&ctrl->queues[0]);
1584 return error;
1585}
1586
1587static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl)
1588{
1589 nvme_stop_keep_alive(&ctrl->ctrl);
1590 cancel_work_sync(&ctrl->err_work);
1591 cancel_delayed_work_sync(&ctrl->reconnect_work);
1592
1593 if (ctrl->queue_count > 1) {
1594 nvme_stop_queues(&ctrl->ctrl);
1595 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1596 nvme_cancel_request, &ctrl->ctrl);
1597 nvme_rdma_free_io_queues(ctrl);
1598 }
1599
45862ebc 1600 if (test_bit(NVME_RDMA_Q_CONNECTED, &ctrl->queues[0].flags))
71102307
CH		 1601 nvme_shutdown_ctrl(&ctrl->ctrl);
1602
1603 blk_mq_stop_hw_queues(ctrl->ctrl.admin_q);
1604 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1605 nvme_cancel_request, &ctrl->ctrl);
1606 nvme_rdma_destroy_admin_queue(ctrl);
1607}
1608
2461a8dd
SG		 1609static void __nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1610{
1611 nvme_uninit_ctrl(&ctrl->ctrl);
1612 if (shutdown)
1613 nvme_rdma_shutdown_ctrl(ctrl);
a34ca17a
SG		 1614
1615 if (ctrl->ctrl.tagset) {
1616 blk_cleanup_queue(ctrl->ctrl.connect_q);
1617 blk_mq_free_tag_set(&ctrl->tag_set);
1618 nvme_rdma_dev_put(ctrl->device);
1619 }
1620
2461a8dd
SG		 1621 nvme_put_ctrl(&ctrl->ctrl);
1622}
1623
71102307
CH		 1624static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1625{
1626 struct nvme_rdma_ctrl *ctrl = container_of(work,
1627 struct nvme_rdma_ctrl, delete_work);
1628
2461a8dd 1629 __nvme_rdma_remove_ctrl(ctrl, true);
71102307
CH		 1630}
1631
1632static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1633{
1634 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1635 return -EBUSY;
1636
1637 if (!queue_work(nvme_rdma_wq, &ctrl->delete_work))
1638 return -EBUSY;
1639
1640 return 0;
1641}
1642
1643static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1644{
1645 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
cdbecc8d 1646 int ret = 0;
71102307 1647
cdbecc8d
SW		 1648 /*
1649 * Keep a reference until all work is flushed since
1650 * __nvme_rdma_del_ctrl can free the ctrl mem
1651 */
1652 if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1653 return -EBUSY;
71102307 1654 ret = __nvme_rdma_del_ctrl(ctrl);
cdbecc8d
SW		 1655 if (!ret)
1656 flush_work(&ctrl->delete_work);
1657 nvme_put_ctrl(&ctrl->ctrl);
1658 return ret;
71102307
CH		 1659}
1660
1661static void nvme_rdma_remove_ctrl_work(struct work_struct *work)
1662{
1663 struct nvme_rdma_ctrl *ctrl = container_of(work,
1664 struct nvme_rdma_ctrl, delete_work);
1665
2461a8dd 1666 __nvme_rdma_remove_ctrl(ctrl, false);
71102307
CH		 1667}
1668
1669static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1670{
1671 struct nvme_rdma_ctrl *ctrl = container_of(work,
1672 struct nvme_rdma_ctrl, reset_work);
1673 int ret;
1674 bool changed;
1675
1676 nvme_rdma_shutdown_ctrl(ctrl);
1677
1678 ret = nvme_rdma_configure_admin_queue(ctrl);
1679 if (ret) {
1680 /* ctrl is already shutdown, just remove the ctrl */
1681 INIT_WORK(&ctrl->delete_work, nvme_rdma_remove_ctrl_work);
1682 goto del_dead_ctrl;
1683 }
1684
1685 if (ctrl->queue_count > 1) {
1686 ret = blk_mq_reinit_tagset(&ctrl->tag_set);
1687 if (ret)
1688 goto del_dead_ctrl;
1689
1690 ret = nvme_rdma_init_io_queues(ctrl);
1691 if (ret)
1692 goto del_dead_ctrl;
1693
1694 ret = nvme_rdma_connect_io_queues(ctrl);
1695 if (ret)
1696 goto del_dead_ctrl;
1697 }
1698
1699 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1700 WARN_ON_ONCE(!changed);
1701
1702 if (ctrl->queue_count > 1) {
1703 nvme_start_queues(&ctrl->ctrl);
1704 nvme_queue_scan(&ctrl->ctrl);
3ef1b4b2 1705 nvme_queue_async_events(&ctrl->ctrl);
71102307
CH		 1706 }
1707
1708 return;
1709
1710del_dead_ctrl:
1711 /* Deleting this dead controller... */
1712 dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1713 WARN_ON(!queue_work(nvme_rdma_wq, &ctrl->delete_work));
1714}
1715
1716static int nvme_rdma_reset_ctrl(struct nvme_ctrl *nctrl)
1717{
1718 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1719
1720 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1721 return -EBUSY;
1722
1723 if (!queue_work(nvme_rdma_wq, &ctrl->reset_work))
1724 return -EBUSY;
1725
1726 flush_work(&ctrl->reset_work);
1727
1728 return 0;
1729}
1730
1731static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1732 .name = "rdma",
1733 .module = THIS_MODULE,
1734 .is_fabrics = true,
1735 .reg_read32 = nvmf_reg_read32,
1736 .reg_read64 = nvmf_reg_read64,
1737 .reg_write32 = nvmf_reg_write32,
1738 .reset_ctrl = nvme_rdma_reset_ctrl,
1739 .free_ctrl = nvme_rdma_free_ctrl,
1740 .submit_async_event = nvme_rdma_submit_async_event,
1741 .delete_ctrl = nvme_rdma_del_ctrl,
1742 .get_subsysnqn = nvmf_get_subsysnqn,
1743 .get_address = nvmf_get_address,
1744};
1745
1746static int nvme_rdma_create_io_queues(struct nvme_rdma_ctrl *ctrl)
1747{
1748 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
1749 int ret;
1750
1751 ret = nvme_set_queue_count(&ctrl->ctrl, &opts->nr_io_queues);
1752 if (ret)
1753 return ret;
1754
1755 ctrl->queue_count = opts->nr_io_queues + 1;
1756 if (ctrl->queue_count < 2)
1757 return 0;
1758
1759 dev_info(ctrl->ctrl.device,
1760 "creating %d I/O queues.\n", opts->nr_io_queues);
1761
1762 ret = nvme_rdma_init_io_queues(ctrl);
1763 if (ret)
1764 return ret;
1765
1766 /*
1767 * We need a reference on the device as long as the tag_set is alive,
1768 * as the MRs in the request structures need a valid ib_device.
1769 */
1770 ret = -EINVAL;
1771 if (!nvme_rdma_dev_get(ctrl->device))
1772 goto out_free_io_queues;
1773
1774 memset(&ctrl->tag_set, 0, sizeof(ctrl->tag_set));
1775 ctrl->tag_set.ops = &nvme_rdma_mq_ops;
c5af8654 1776 ctrl->tag_set.queue_depth = ctrl->ctrl.opts->queue_size;
71102307
CH		 1777 ctrl->tag_set.reserved_tags = 1; /* fabric connect */
1778 ctrl->tag_set.numa_node = NUMA_NO_NODE;
1779 ctrl->tag_set.flags = BLK_MQ_F_SHOULD_MERGE;
1780 ctrl->tag_set.cmd_size = sizeof(struct nvme_rdma_request) +
1781 SG_CHUNK_SIZE * sizeof(struct scatterlist);
1782 ctrl->tag_set.driver_data = ctrl;
1783 ctrl->tag_set.nr_hw_queues = ctrl->queue_count - 1;
1784 ctrl->tag_set.timeout = NVME_IO_TIMEOUT;
1785
1786 ret = blk_mq_alloc_tag_set(&ctrl->tag_set);
1787 if (ret)
1788 goto out_put_dev;
1789 ctrl->ctrl.tagset = &ctrl->tag_set;
1790
1791 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
1792 if (IS_ERR(ctrl->ctrl.connect_q)) {
1793 ret = PTR_ERR(ctrl->ctrl.connect_q);
1794 goto out_free_tag_set;
1795 }
1796
1797 ret = nvme_rdma_connect_io_queues(ctrl);
1798 if (ret)
1799 goto out_cleanup_connect_q;
1800
1801 return 0;
1802
1803out_cleanup_connect_q:
1804 blk_cleanup_queue(ctrl->ctrl.connect_q);
1805out_free_tag_set:
1806 blk_mq_free_tag_set(&ctrl->tag_set);
1807out_put_dev:
1808 nvme_rdma_dev_put(ctrl->device);
1809out_free_io_queues:
1810 nvme_rdma_free_io_queues(ctrl);
1811 return ret;
1812}
1813
1814static int nvme_rdma_parse_ipaddr(struct sockaddr_in *in_addr, char *p)
1815{
1816 u8 *addr = (u8 *)&in_addr->sin_addr.s_addr;
1817 size_t buflen = strlen(p);
1818
1819 /* XXX: handle IPv6 addresses */
1820
1821 if (buflen > INET_ADDRSTRLEN)
1822 return -EINVAL;
1823 if (in4_pton(p, buflen, addr, '\0', NULL) == 0)
1824 return -EINVAL;
1825 in_addr->sin_family = AF_INET;
1826 return 0;
1827}
1828
1829static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1830 struct nvmf_ctrl_options *opts)
1831{
1832 struct nvme_rdma_ctrl *ctrl;
1833 int ret;
1834 bool changed;
1835
1836 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1837 if (!ctrl)
1838 return ERR_PTR(-ENOMEM);
1839 ctrl->ctrl.opts = opts;
1840 INIT_LIST_HEAD(&ctrl->list);
1841
1842 ret = nvme_rdma_parse_ipaddr(&ctrl->addr_in, opts->traddr);
1843 if (ret) {
1844 pr_err("malformed IP address passed: %s\n", opts->traddr);
1845 goto out_free_ctrl;
1846 }
1847
1848 if (opts->mask & NVMF_OPT_TRSVCID) {
1849 u16 port;
1850
1851 ret = kstrtou16(opts->trsvcid, 0, &port);
1852 if (ret)
1853 goto out_free_ctrl;
1854
1855 ctrl->addr_in.sin_port = cpu_to_be16(port);
1856 } else {
1857 ctrl->addr_in.sin_port = cpu_to_be16(NVME_RDMA_IP_PORT);
1858 }
1859
1860 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1861 0 /* no quirks, we're perfect! */);
1862 if (ret)
1863 goto out_free_ctrl;
1864
1865 ctrl->reconnect_delay = opts->reconnect_delay;
1866 INIT_DELAYED_WORK(&ctrl->reconnect_work,
1867 nvme_rdma_reconnect_ctrl_work);
1868 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1869 INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1870 INIT_WORK(&ctrl->reset_work, nvme_rdma_reset_ctrl_work);
1871 spin_lock_init(&ctrl->lock);
1872
1873 ctrl->queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
c5af8654 1874 ctrl->ctrl.sqsize = opts->queue_size - 1;
71102307
CH		 1875 ctrl->ctrl.kato = opts->kato;
1876
1877 ret = -ENOMEM;
1878 ctrl->queues = kcalloc(ctrl->queue_count, sizeof(*ctrl->queues),
1879 GFP_KERNEL);
1880 if (!ctrl->queues)
1881 goto out_uninit_ctrl;
1882
1883 ret = nvme_rdma_configure_admin_queue(ctrl);
1884 if (ret)
1885 goto out_kfree_queues;
1886
1887 /* sanity check icdoff */
1888 if (ctrl->ctrl.icdoff) {
1889 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1890 goto out_remove_admin_queue;
1891 }
1892
1893 /* sanity check keyed sgls */
1894 if (!(ctrl->ctrl.sgls & (1 << 20))) {
1895 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1896 goto out_remove_admin_queue;
1897 }
1898
1899 if (opts->queue_size > ctrl->ctrl.maxcmd) {
1900 /* warn if maxcmd is lower than queue_size */
1901 dev_warn(ctrl->ctrl.device,
1902 "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1903 opts->queue_size, ctrl->ctrl.maxcmd);
1904 opts->queue_size = ctrl->ctrl.maxcmd;
1905 }
1906
76c08bf4
SJ		 1907 if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
1908 /* warn if sqsize is lower than queue_size */
1909 dev_warn(ctrl->ctrl.device,
1910 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1911 opts->queue_size, ctrl->ctrl.sqsize + 1);
1912 opts->queue_size = ctrl->ctrl.sqsize + 1;
1913 }
1914
71102307
CH		 1915 if (opts->nr_io_queues) {
1916 ret = nvme_rdma_create_io_queues(ctrl);
1917 if (ret)
1918 goto out_remove_admin_queue;
1919 }
1920
1921 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1922 WARN_ON_ONCE(!changed);
1923
1924 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISp\n",
1925 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1926
1927 kref_get(&ctrl->ctrl.kref);
1928
1929 mutex_lock(&nvme_rdma_ctrl_mutex);
1930 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1931 mutex_unlock(&nvme_rdma_ctrl_mutex);
1932
1933 if (opts->nr_io_queues) {
1934 nvme_queue_scan(&ctrl->ctrl);
1935 nvme_queue_async_events(&ctrl->ctrl);
1936 }
1937
1938 return &ctrl->ctrl;
1939
1940out_remove_admin_queue:
1941 nvme_stop_keep_alive(&ctrl->ctrl);
1942 nvme_rdma_destroy_admin_queue(ctrl);
1943out_kfree_queues:
1944 kfree(ctrl->queues);
1945out_uninit_ctrl:
1946 nvme_uninit_ctrl(&ctrl->ctrl);
1947 nvme_put_ctrl(&ctrl->ctrl);
1948 if (ret > 0)
1949 ret = -EIO;
1950 return ERR_PTR(ret);
1951out_free_ctrl:
1952 kfree(ctrl);
1953 return ERR_PTR(ret);
1954}
1955
1956static struct nvmf_transport_ops nvme_rdma_transport = {
1957 .name = "rdma",
1958 .required_opts = NVMF_OPT_TRADDR,
2ac17c28 1959 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY,
71102307
CH		 1960 .create_ctrl = nvme_rdma_create_ctrl,
1961};
1962
e87a911f
SW		 1963static void nvme_rdma_add_one(struct ib_device *ib_device)
1964{
1965}
1966
1967static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1968{
1969 struct nvme_rdma_ctrl *ctrl;
1970
1971 /* Delete all controllers using this device */
1972 mutex_lock(&nvme_rdma_ctrl_mutex);
1973 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1974 if (ctrl->device->dev != ib_device)
1975 continue;
1976 dev_info(ctrl->ctrl.device,
1977 "Removing ctrl: NQN \"%s\", addr %pISp\n",
1978 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1979 __nvme_rdma_del_ctrl(ctrl);
1980 }
1981 mutex_unlock(&nvme_rdma_ctrl_mutex);
1982
1983 flush_workqueue(nvme_rdma_wq);
1984}
1985
1986static struct ib_client nvme_rdma_ib_client = {
1987 .name = "nvme_rdma",
1988 .add = nvme_rdma_add_one,
1989 .remove = nvme_rdma_remove_one
1990};
1991
71102307
CH		 1992static int __init nvme_rdma_init_module(void)
1993{
e87a911f
SW		 1994 int ret;
1995
71102307
CH		 1996 nvme_rdma_wq = create_workqueue("nvme_rdma_wq");
1997 if (!nvme_rdma_wq)
1998 return -ENOMEM;
1999
e87a911f
SW		 2000 ret = ib_register_client(&nvme_rdma_ib_client);
2001 if (ret) {
2002 destroy_workqueue(nvme_rdma_wq);
2003 return ret;
2004 }
2005
71102307
CH		 2006 nvmf_register_transport(&nvme_rdma_transport);
2007 return 0;
2008}
2009
2010static void __exit nvme_rdma_cleanup_module(void)
2011{
71102307 2012 nvmf_unregister_transport(&nvme_rdma_transport);
e87a911f 2013 ib_unregister_client(&nvme_rdma_ib_client);
71102307
CH		 2014 destroy_workqueue(nvme_rdma_wq);
2015}
2016
2017module_init(nvme_rdma_init_module);
2018module_exit(nvme_rdma_cleanup_module);
2019
2020MODULE_LICENSE("GPL v2");
Ubuntu Artful kernel mirror