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