?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Merge tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost
[cris-mirror.git] / drivers / nvme / host / rdma.c
blob4d84a73ee12d06907ea94ef77e3e654d6e2b2033
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
15 #include <linux/module.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <rdma/mr_pool.h>
19 #include <linux/err.h>
20 #include <linux/string.h>
21 #include <linux/atomic.h>
22 #include <linux/blk-mq.h>
23 #include <linux/blk-mq-rdma.h>
24 #include <linux/types.h>
25 #include <linux/list.h>
26 #include <linux/mutex.h>
27 #include <linux/scatterlist.h>
28 #include <linux/nvme.h>
29 #include <asm/unaligned.h>
30
31 #include <rdma/ib_verbs.h>
32 #include <rdma/rdma_cm.h>
33 #include <linux/nvme-rdma.h>
34
35 #include "nvme.h"
36 #include "fabrics.h"
37
38
39 #define NVME_RDMA_CONNECT_TIMEOUT_MS 3000 /* 3 second */
40
41 #define NVME_RDMA_MAX_SEGMENTS 256
42
43 #define NVME_RDMA_MAX_INLINE_SEGMENTS 1
44
45 struct nvme_rdma_device {
46 struct ib_device *dev;
47 struct ib_pd *pd;
48 struct kref ref;
49 struct list_head entry;
50 };
51
52 struct nvme_rdma_qe {
53 struct ib_cqe cqe;
54 void *data;
55 u64 dma;
56 };
57
58 struct nvme_rdma_queue;
59 struct nvme_rdma_request {
60 struct nvme_request req;
61 struct ib_mr *mr;
62 struct nvme_rdma_qe sqe;
63 union nvme_result result;
64 __le16 status;
65 refcount_t ref;
66 struct ib_sge sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
67 u32 num_sge;
68 int nents;
69 struct ib_reg_wr reg_wr;
70 struct ib_cqe reg_cqe;
71 struct nvme_rdma_queue *queue;
72 struct sg_table sg_table;
73 struct scatterlist first_sgl[];
74 };
75
76 enum nvme_rdma_queue_flags {
77 NVME_RDMA_Q_ALLOCATED = 0,
78 NVME_RDMA_Q_LIVE = 1,
79 NVME_RDMA_Q_TR_READY = 2,
80 };
81
82 struct nvme_rdma_queue {
83 struct nvme_rdma_qe *rsp_ring;
84 int queue_size;
85 size_t cmnd_capsule_len;
86 struct nvme_rdma_ctrl *ctrl;
87 struct nvme_rdma_device *device;
88 struct ib_cq *ib_cq;
89 struct ib_qp *qp;
90
91 unsigned long flags;
92 struct rdma_cm_id *cm_id;
93 int cm_error;
94 struct completion cm_done;
95 };
96
97 struct nvme_rdma_ctrl {
98 /* read only in the hot path */
99 struct nvme_rdma_queue *queues;
100
101 /* other member variables */
102 struct blk_mq_tag_set tag_set;
103 struct work_struct err_work;
104
105 struct nvme_rdma_qe async_event_sqe;
106
107 struct delayed_work reconnect_work;
108
109 struct list_head list;
110
111 struct blk_mq_tag_set admin_tag_set;
112 struct nvme_rdma_device *device;
113
114 u32 max_fr_pages;
115
116 struct sockaddr_storage addr;
117 struct sockaddr_storage src_addr;
118
119 struct nvme_ctrl ctrl;
120 };
121
122 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
123 {
124 return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
125 }
126
127 static LIST_HEAD(device_list);
128 static DEFINE_MUTEX(device_list_mutex);
129
130 static LIST_HEAD(nvme_rdma_ctrl_list);
131 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
132
133 /*
134 * Disabling this option makes small I/O goes faster, but is fundamentally
135 * unsafe. With it turned off we will have to register a global rkey that
136 * allows read and write access to all physical memory.
137 */
138 static bool register_always = true;
139 module_param(register_always, bool, 0444);
140 MODULE_PARM_DESC(register_always,
141 "Use memory registration even for contiguous memory regions");
142
143 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
144 struct rdma_cm_event *event);
145 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
146
147 static const struct blk_mq_ops nvme_rdma_mq_ops;
148 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
149
150 /* XXX: really should move to a generic header sooner or later.. */
151 static inline void put_unaligned_le24(u32 val, u8 *p)
152 {
153 *p++ = val;
154 *p++ = val >> 8;
155 *p++ = val >> 16;
156 }
157
158 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
159 {
160 return queue - queue->ctrl->queues;
161 }
162
163 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
164 {
165 return queue->cmnd_capsule_len - sizeof(struct nvme_command);
166 }
167
168 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
169 size_t capsule_size, enum dma_data_direction dir)
170 {
171 ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
172 kfree(qe->data);
173 }
174
175 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
176 size_t capsule_size, enum dma_data_direction dir)
177 {
178 qe->data = kzalloc(capsule_size, GFP_KERNEL);
179 if (!qe->data)
180 return -ENOMEM;
181
182 qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
183 if (ib_dma_mapping_error(ibdev, qe->dma)) {
184 kfree(qe->data);
185 return -ENOMEM;
186 }
187
188 return 0;
189 }
190
191 static void nvme_rdma_free_ring(struct ib_device *ibdev,
192 struct nvme_rdma_qe *ring, size_t ib_queue_size,
193 size_t capsule_size, enum dma_data_direction dir)
194 {
195 int i;
196
197 for (i = 0; i < ib_queue_size; i++)
198 nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
199 kfree(ring);
200 }
201
202 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
203 size_t ib_queue_size, size_t capsule_size,
204 enum dma_data_direction dir)
205 {
206 struct nvme_rdma_qe *ring;
207 int i;
208
209 ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
210 if (!ring)
211 return NULL;
212
213 for (i = 0; i < ib_queue_size; i++) {
214 if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
215 goto out_free_ring;
216 }
217
218 return ring;
219
220 out_free_ring:
221 nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
222 return NULL;
223 }
224
225 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
226 {
227 pr_debug("QP event %s (%d)\n",
228 ib_event_msg(event->event), event->event);
229
230 }
231
232 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
233 {
234 wait_for_completion_interruptible_timeout(&queue->cm_done,
235 msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
236 return queue->cm_error;
237 }
238
239 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
240 {
241 struct nvme_rdma_device *dev = queue->device;
242 struct ib_qp_init_attr init_attr;
243 int ret;
244
245 memset(&init_attr, 0, sizeof(init_attr));
246 init_attr.event_handler = nvme_rdma_qp_event;
247 /* +1 for drain */
248 init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
249 /* +1 for drain */
250 init_attr.cap.max_recv_wr = queue->queue_size + 1;
251 init_attr.cap.max_recv_sge = 1;
252 init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
253 init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
254 init_attr.qp_type = IB_QPT_RC;
255 init_attr.send_cq = queue->ib_cq;
256 init_attr.recv_cq = queue->ib_cq;
257
258 ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
259
260 queue->qp = queue->cm_id->qp;
261 return ret;
262 }
263
264 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
265 struct request *rq, unsigned int hctx_idx)
266 {
267 struct nvme_rdma_ctrl *ctrl = set->driver_data;
268 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
269 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
270 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
271 struct nvme_rdma_device *dev = queue->device;
272
273 nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
274 DMA_TO_DEVICE);
275 }
276
277 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
278 struct request *rq, unsigned int hctx_idx,
279 unsigned int numa_node)
280 {
281 struct nvme_rdma_ctrl *ctrl = set->driver_data;
282 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
283 int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
284 struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
285 struct nvme_rdma_device *dev = queue->device;
286 struct ib_device *ibdev = dev->dev;
287 int ret;
288
289 ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
290 DMA_TO_DEVICE);
291 if (ret)
292 return ret;
293
294 req->queue = queue;
295
296 return 0;
297 }
298
299 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
300 unsigned int hctx_idx)
301 {
302 struct nvme_rdma_ctrl *ctrl = data;
303 struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
304
305 BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
306
307 hctx->driver_data = queue;
308 return 0;
309 }
310
311 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
312 unsigned int hctx_idx)
313 {
314 struct nvme_rdma_ctrl *ctrl = data;
315 struct nvme_rdma_queue *queue = &ctrl->queues[0];
316
317 BUG_ON(hctx_idx != 0);
318
319 hctx->driver_data = queue;
320 return 0;
321 }
322
323 static void nvme_rdma_free_dev(struct kref *ref)
324 {
325 struct nvme_rdma_device *ndev =
326 container_of(ref, struct nvme_rdma_device, ref);
327
328 mutex_lock(&device_list_mutex);
329 list_del(&ndev->entry);
330 mutex_unlock(&device_list_mutex);
331
332 ib_dealloc_pd(ndev->pd);
333 kfree(ndev);
334 }
335
336 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
337 {
338 kref_put(&dev->ref, nvme_rdma_free_dev);
339 }
340
341 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
342 {
343 return kref_get_unless_zero(&dev->ref);
344 }
345
346 static struct nvme_rdma_device *
347 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
348 {
349 struct nvme_rdma_device *ndev;
350
351 mutex_lock(&device_list_mutex);
352 list_for_each_entry(ndev, &device_list, entry) {
353 if (ndev->dev->node_guid == cm_id->device->node_guid &&
354 nvme_rdma_dev_get(ndev))
355 goto out_unlock;
356 }
357
358 ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
359 if (!ndev)
360 goto out_err;
361
362 ndev->dev = cm_id->device;
363 kref_init(&ndev->ref);
364
365 ndev->pd = ib_alloc_pd(ndev->dev,
366 register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
367 if (IS_ERR(ndev->pd))
368 goto out_free_dev;
369
370 if (!(ndev->dev->attrs.device_cap_flags &
371 IB_DEVICE_MEM_MGT_EXTENSIONS)) {
372 dev_err(&ndev->dev->dev,
373 "Memory registrations not supported.\n");
374 goto out_free_pd;
375 }
376
377 list_add(&ndev->entry, &device_list);
378 out_unlock:
379 mutex_unlock(&device_list_mutex);
380 return ndev;
381
382 out_free_pd:
383 ib_dealloc_pd(ndev->pd);
384 out_free_dev:
385 kfree(ndev);
386 out_err:
387 mutex_unlock(&device_list_mutex);
388 return NULL;
389 }
390
391 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
392 {
393 struct nvme_rdma_device *dev;
394 struct ib_device *ibdev;
395
396 if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
397 return;
398
399 dev = queue->device;
400 ibdev = dev->dev;
401
402 ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
403
404 /*
405 * The cm_id object might have been destroyed during RDMA connection
406 * establishment error flow to avoid getting other cma events, thus
407 * the destruction of the QP shouldn't use rdma_cm API.
408 */
409 ib_destroy_qp(queue->qp);
410 ib_free_cq(queue->ib_cq);
411
412 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
413 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
414
415 nvme_rdma_dev_put(dev);
416 }
417
418 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev)
419 {
420 return min_t(u32, NVME_RDMA_MAX_SEGMENTS,
421 ibdev->attrs.max_fast_reg_page_list_len);
422 }
423
424 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
425 {
426 struct ib_device *ibdev;
427 const int send_wr_factor = 3; /* MR, SEND, INV */
428 const int cq_factor = send_wr_factor + 1; /* + RECV */
429 int comp_vector, idx = nvme_rdma_queue_idx(queue);
430 int ret;
431
432 queue->device = nvme_rdma_find_get_device(queue->cm_id);
433 if (!queue->device) {
434 dev_err(queue->cm_id->device->dev.parent,
435 "no client data found!\n");
436 return -ECONNREFUSED;
437 }
438 ibdev = queue->device->dev;
439
440 /*
441 * Spread I/O queues completion vectors according their queue index.
442 * Admin queues can always go on completion vector 0.
443 */
444 comp_vector = idx == 0 ? idx : idx - 1;
445
446 /* +1 for ib_stop_cq */
447 queue->ib_cq = ib_alloc_cq(ibdev, queue,
448 cq_factor * queue->queue_size + 1,
449 comp_vector, IB_POLL_SOFTIRQ);
450 if (IS_ERR(queue->ib_cq)) {
451 ret = PTR_ERR(queue->ib_cq);
452 goto out_put_dev;
453 }
454
455 ret = nvme_rdma_create_qp(queue, send_wr_factor);
456 if (ret)
457 goto out_destroy_ib_cq;
458
459 queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
460 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
461 if (!queue->rsp_ring) {
462 ret = -ENOMEM;
463 goto out_destroy_qp;
464 }
465
466 ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
467 queue->queue_size,
468 IB_MR_TYPE_MEM_REG,
469 nvme_rdma_get_max_fr_pages(ibdev));
470 if (ret) {
471 dev_err(queue->ctrl->ctrl.device,
472 "failed to initialize MR pool sized %d for QID %d\n",
473 queue->queue_size, idx);
474 goto out_destroy_ring;
475 }
476
477 set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
478
479 return 0;
480
481 out_destroy_ring:
482 nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
483 sizeof(struct nvme_completion), DMA_FROM_DEVICE);
484 out_destroy_qp:
485 rdma_destroy_qp(queue->cm_id);
486 out_destroy_ib_cq:
487 ib_free_cq(queue->ib_cq);
488 out_put_dev:
489 nvme_rdma_dev_put(queue->device);
490 return ret;
491 }
492
493 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
494 int idx, size_t queue_size)
495 {
496 struct nvme_rdma_queue *queue;
497 struct sockaddr *src_addr = NULL;
498 int ret;
499
500 queue = &ctrl->queues[idx];
501 queue->ctrl = ctrl;
502 init_completion(&queue->cm_done);
503
504 if (idx > 0)
505 queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
506 else
507 queue->cmnd_capsule_len = sizeof(struct nvme_command);
508
509 queue->queue_size = queue_size;
510
511 queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
512 RDMA_PS_TCP, IB_QPT_RC);
513 if (IS_ERR(queue->cm_id)) {
514 dev_info(ctrl->ctrl.device,
515 "failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
516 return PTR_ERR(queue->cm_id);
517 }
518
519 if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
520 src_addr = (struct sockaddr *)&ctrl->src_addr;
521
522 queue->cm_error = -ETIMEDOUT;
523 ret = rdma_resolve_addr(queue->cm_id, src_addr,
524 (struct sockaddr *)&ctrl->addr,
525 NVME_RDMA_CONNECT_TIMEOUT_MS);
526 if (ret) {
527 dev_info(ctrl->ctrl.device,
528 "rdma_resolve_addr failed (%d).\n", ret);
529 goto out_destroy_cm_id;
530 }
531
532 ret = nvme_rdma_wait_for_cm(queue);
533 if (ret) {
534 dev_info(ctrl->ctrl.device,
535 "rdma connection establishment failed (%d)\n", ret);
536 goto out_destroy_cm_id;
537 }
538
539 set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
540
541 return 0;
542
543 out_destroy_cm_id:
544 rdma_destroy_id(queue->cm_id);
545 nvme_rdma_destroy_queue_ib(queue);
546 return ret;
547 }
548
549 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
550 {
551 if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
552 return;
553
554 rdma_disconnect(queue->cm_id);
555 ib_drain_qp(queue->qp);
556 }
557
558 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
559 {
560 if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
561 return;
562
563 if (nvme_rdma_queue_idx(queue) == 0) {
564 nvme_rdma_free_qe(queue->device->dev,
565 &queue->ctrl->async_event_sqe,
566 sizeof(struct nvme_command), DMA_TO_DEVICE);
567 }
568
569 nvme_rdma_destroy_queue_ib(queue);
570 rdma_destroy_id(queue->cm_id);
571 }
572
573 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
574 {
575 int i;
576
577 for (i = 1; i < ctrl->ctrl.queue_count; i++)
578 nvme_rdma_free_queue(&ctrl->queues[i]);
579 }
580
581 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
582 {
583 int i;
584
585 for (i = 1; i < ctrl->ctrl.queue_count; i++)
586 nvme_rdma_stop_queue(&ctrl->queues[i]);
587 }
588
589 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
590 {
591 int ret;
592
593 if (idx)
594 ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
595 else
596 ret = nvmf_connect_admin_queue(&ctrl->ctrl);
597
598 if (!ret)
599 set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags);
600 else
601 dev_info(ctrl->ctrl.device,
602 "failed to connect queue: %d ret=%d\n", idx, ret);
603 return ret;
604 }
605
606 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
607 {
608 int i, ret = 0;
609
610 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
611 ret = nvme_rdma_start_queue(ctrl, i);
612 if (ret)
613 goto out_stop_queues;
614 }
615
616 return 0;
617
618 out_stop_queues:
619 for (i--; i >= 1; i--)
620 nvme_rdma_stop_queue(&ctrl->queues[i]);
621 return ret;
622 }
623
624 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
625 {
626 struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
627 struct ib_device *ibdev = ctrl->device->dev;
628 unsigned int nr_io_queues;
629 int i, ret;
630
631 nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
632
633 /*
634 * we map queues according to the device irq vectors for
635 * optimal locality so we don't need more queues than
636 * completion vectors.
637 */
638 nr_io_queues = min_t(unsigned int, nr_io_queues,
639 ibdev->num_comp_vectors);
640
641 ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
642 if (ret)
643 return ret;
644
645 ctrl->ctrl.queue_count = nr_io_queues + 1;
646 if (ctrl->ctrl.queue_count < 2)
647 return 0;
648
649 dev_info(ctrl->ctrl.device,
650 "creating %d I/O queues.\n", nr_io_queues);
651
652 for (i = 1; i < ctrl->ctrl.queue_count; i++) {
653 ret = nvme_rdma_alloc_queue(ctrl, i,
654 ctrl->ctrl.sqsize + 1);
655 if (ret)
656 goto out_free_queues;
657 }
658
659 return 0;
660
661 out_free_queues:
662 for (i--; i >= 1; i--)
663 nvme_rdma_free_queue(&ctrl->queues[i]);
664
665 return ret;
666 }
667
668 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl,
669 struct blk_mq_tag_set *set)
670 {
671 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
672
673 blk_mq_free_tag_set(set);
674 nvme_rdma_dev_put(ctrl->device);
675 }
676
677 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
678 bool admin)
679 {
680 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
681 struct blk_mq_tag_set *set;
682 int ret;
683
684 if (admin) {
685 set = &ctrl->admin_tag_set;
686 memset(set, 0, sizeof(*set));
687 set->ops = &nvme_rdma_admin_mq_ops;
688 set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
689 set->reserved_tags = 2; /* connect + keep-alive */
690 set->numa_node = NUMA_NO_NODE;
691 set->cmd_size = sizeof(struct nvme_rdma_request) +
692 SG_CHUNK_SIZE * sizeof(struct scatterlist);
693 set->driver_data = ctrl;
694 set->nr_hw_queues = 1;
695 set->timeout = ADMIN_TIMEOUT;
696 set->flags = BLK_MQ_F_NO_SCHED;
697 } else {
698 set = &ctrl->tag_set;
699 memset(set, 0, sizeof(*set));
700 set->ops = &nvme_rdma_mq_ops;
701 set->queue_depth = nctrl->opts->queue_size;
702 set->reserved_tags = 1; /* fabric connect */
703 set->numa_node = NUMA_NO_NODE;
704 set->flags = BLK_MQ_F_SHOULD_MERGE;
705 set->cmd_size = sizeof(struct nvme_rdma_request) +
706 SG_CHUNK_SIZE * sizeof(struct scatterlist);
707 set->driver_data = ctrl;
708 set->nr_hw_queues = nctrl->queue_count - 1;
709 set->timeout = NVME_IO_TIMEOUT;
710 }
711
712 ret = blk_mq_alloc_tag_set(set);
713 if (ret)
714 goto out;
715
716 /*
717 * We need a reference on the device as long as the tag_set is alive,
718 * as the MRs in the request structures need a valid ib_device.
719 */
720 ret = nvme_rdma_dev_get(ctrl->device);
721 if (!ret) {
722 ret = -EINVAL;
723 goto out_free_tagset;
724 }
725
726 return set;
727
728 out_free_tagset:
729 blk_mq_free_tag_set(set);
730 out:
731 return ERR_PTR(ret);
732 }
733
734 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
735 bool remove)
736 {
737 nvme_rdma_stop_queue(&ctrl->queues[0]);
738 if (remove) {
739 blk_cleanup_queue(ctrl->ctrl.admin_q);
740 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
741 }
742 nvme_rdma_free_queue(&ctrl->queues[0]);
743 }
744
745 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
746 bool new)
747 {
748 int error;
749
750 error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
751 if (error)
752 return error;
753
754 ctrl->device = ctrl->queues[0].device;
755
756 ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
757
758 if (new) {
759 ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
760 if (IS_ERR(ctrl->ctrl.admin_tagset)) {
761 error = PTR_ERR(ctrl->ctrl.admin_tagset);
762 goto out_free_queue;
763 }
764
765 ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
766 if (IS_ERR(ctrl->ctrl.admin_q)) {
767 error = PTR_ERR(ctrl->ctrl.admin_q);
768 goto out_free_tagset;
769 }
770 }
771
772 error = nvme_rdma_start_queue(ctrl, 0);
773 if (error)
774 goto out_cleanup_queue;
775
776 error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
777 &ctrl->ctrl.cap);
778 if (error) {
779 dev_err(ctrl->ctrl.device,
780 "prop_get NVME_REG_CAP failed\n");
781 goto out_cleanup_queue;
782 }
783
784 ctrl->ctrl.sqsize =
785 min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
786
787 error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
788 if (error)
789 goto out_cleanup_queue;
790
791 ctrl->ctrl.max_hw_sectors =
792 (ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
793
794 error = nvme_init_identify(&ctrl->ctrl);
795 if (error)
796 goto out_cleanup_queue;
797
798 error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
799 &ctrl->async_event_sqe, sizeof(struct nvme_command),
800 DMA_TO_DEVICE);
801 if (error)
802 goto out_cleanup_queue;
803
804 return 0;
805
806 out_cleanup_queue:
807 if (new)
808 blk_cleanup_queue(ctrl->ctrl.admin_q);
809 out_free_tagset:
810 if (new)
811 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
812 out_free_queue:
813 nvme_rdma_free_queue(&ctrl->queues[0]);
814 return error;
815 }
816
817 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
818 bool remove)
819 {
820 nvme_rdma_stop_io_queues(ctrl);
821 if (remove) {
822 blk_cleanup_queue(ctrl->ctrl.connect_q);
823 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
824 }
825 nvme_rdma_free_io_queues(ctrl);
826 }
827
828 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
829 {
830 int ret;
831
832 ret = nvme_rdma_alloc_io_queues(ctrl);
833 if (ret)
834 return ret;
835
836 if (new) {
837 ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
838 if (IS_ERR(ctrl->ctrl.tagset)) {
839 ret = PTR_ERR(ctrl->ctrl.tagset);
840 goto out_free_io_queues;
841 }
842
843 ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
844 if (IS_ERR(ctrl->ctrl.connect_q)) {
845 ret = PTR_ERR(ctrl->ctrl.connect_q);
846 goto out_free_tag_set;
847 }
848 } else {
849 blk_mq_update_nr_hw_queues(&ctrl->tag_set,
850 ctrl->ctrl.queue_count - 1);
851 }
852
853 ret = nvme_rdma_start_io_queues(ctrl);
854 if (ret)
855 goto out_cleanup_connect_q;
856
857 return 0;
858
859 out_cleanup_connect_q:
860 if (new)
861 blk_cleanup_queue(ctrl->ctrl.connect_q);
862 out_free_tag_set:
863 if (new)
864 nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
865 out_free_io_queues:
866 nvme_rdma_free_io_queues(ctrl);
867 return ret;
868 }
869
870 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
871 {
872 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
873
874 if (list_empty(&ctrl->list))
875 goto free_ctrl;
876
877 mutex_lock(&nvme_rdma_ctrl_mutex);
878 list_del(&ctrl->list);
879 mutex_unlock(&nvme_rdma_ctrl_mutex);
880
881 kfree(ctrl->queues);
882 nvmf_free_options(nctrl->opts);
883 free_ctrl:
884 kfree(ctrl);
885 }
886
887 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
888 {
889 /* If we are resetting/deleting then do nothing */
890 if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
891 WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
892 ctrl->ctrl.state == NVME_CTRL_LIVE);
893 return;
894 }
895
896 if (nvmf_should_reconnect(&ctrl->ctrl)) {
897 dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
898 ctrl->ctrl.opts->reconnect_delay);
899 queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
900 ctrl->ctrl.opts->reconnect_delay * HZ);
901 } else {
902 dev_info(ctrl->ctrl.device, "Removing controller...\n");
903 nvme_delete_ctrl(&ctrl->ctrl);
904 }
905 }
906
907 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
908 {
909 struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
910 struct nvme_rdma_ctrl, reconnect_work);
911 bool changed;
912 int ret;
913
914 ++ctrl->ctrl.nr_reconnects;
915
916 ret = nvme_rdma_configure_admin_queue(ctrl, false);
917 if (ret)
918 goto requeue;
919
920 if (ctrl->ctrl.queue_count > 1) {
921 ret = nvme_rdma_configure_io_queues(ctrl, false);
922 if (ret)
923 goto destroy_admin;
924 }
925
926 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
927 if (!changed) {
928 /* state change failure is ok if we're in DELETING state */
929 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
930 return;
931 }
932
933 nvme_start_ctrl(&ctrl->ctrl);
934
935 dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
936 ctrl->ctrl.nr_reconnects);
937
938 ctrl->ctrl.nr_reconnects = 0;
939
940 return;
941
942 destroy_admin:
943 nvme_rdma_destroy_admin_queue(ctrl, false);
944 requeue:
945 dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
946 ctrl->ctrl.nr_reconnects);
947 nvme_rdma_reconnect_or_remove(ctrl);
948 }
949
950 static void nvme_rdma_error_recovery_work(struct work_struct *work)
951 {
952 struct nvme_rdma_ctrl *ctrl = container_of(work,
953 struct nvme_rdma_ctrl, err_work);
954
955 nvme_stop_keep_alive(&ctrl->ctrl);
956
957 if (ctrl->ctrl.queue_count > 1) {
958 nvme_stop_queues(&ctrl->ctrl);
959 blk_mq_tagset_busy_iter(&ctrl->tag_set,
960 nvme_cancel_request, &ctrl->ctrl);
961 nvme_rdma_destroy_io_queues(ctrl, false);
962 }
963
964 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
965 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
966 nvme_cancel_request, &ctrl->ctrl);
967 nvme_rdma_destroy_admin_queue(ctrl, false);
968
969 /*
970 * queues are not a live anymore, so restart the queues to fail fast
971 * new IO
972 */
973 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
974 nvme_start_queues(&ctrl->ctrl);
975
976 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
977 /* state change failure should never happen */
978 WARN_ON_ONCE(1);
979 return;
980 }
981
982 nvme_rdma_reconnect_or_remove(ctrl);
983 }
984
985 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
986 {
987 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
988 return;
989
990 queue_work(nvme_wq, &ctrl->err_work);
991 }
992
993 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
994 const char *op)
995 {
996 struct nvme_rdma_queue *queue = cq->cq_context;
997 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
998
999 if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1000 dev_info(ctrl->ctrl.device,
1001 "%s for CQE 0x%p failed with status %s (%d)\n",
1002 op, wc->wr_cqe,
1003 ib_wc_status_msg(wc->status), wc->status);
1004 nvme_rdma_error_recovery(ctrl);
1005 }
1006
1007 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1008 {
1009 if (unlikely(wc->status != IB_WC_SUCCESS))
1010 nvme_rdma_wr_error(cq, wc, "MEMREG");
1011 }
1012
1013 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1014 {
1015 struct nvme_rdma_request *req =
1016 container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1017 struct request *rq = blk_mq_rq_from_pdu(req);
1018
1019 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1020 nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1021 return;
1022 }
1023
1024 if (refcount_dec_and_test(&req->ref))
1025 nvme_end_request(rq, req->status, req->result);
1026
1027 }
1028
1029 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1030 struct nvme_rdma_request *req)
1031 {
1032 struct ib_send_wr *bad_wr;
1033 struct ib_send_wr wr = {
1034 .opcode = IB_WR_LOCAL_INV,
1035 .next = NULL,
1036 .num_sge = 0,
1037 .send_flags = IB_SEND_SIGNALED,
1038 .ex.invalidate_rkey = req->mr->rkey,
1039 };
1040
1041 req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1042 wr.wr_cqe = &req->reg_cqe;
1043
1044 return ib_post_send(queue->qp, &wr, &bad_wr);
1045 }
1046
1047 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1048 struct request *rq)
1049 {
1050 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1051 struct nvme_rdma_device *dev = queue->device;
1052 struct ib_device *ibdev = dev->dev;
1053
1054 if (!blk_rq_payload_bytes(rq))
1055 return;
1056
1057 if (req->mr) {
1058 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1059 req->mr = NULL;
1060 }
1061
1062 ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1063 req->nents, rq_data_dir(rq) ==
1064 WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1065
1066 nvme_cleanup_cmd(rq);
1067 sg_free_table_chained(&req->sg_table, true);
1068 }
1069
1070 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1071 {
1072 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1073
1074 sg->addr = 0;
1075 put_unaligned_le24(0, sg->length);
1076 put_unaligned_le32(0, sg->key);
1077 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1078 return 0;
1079 }
1080
1081 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1082 struct nvme_rdma_request *req, struct nvme_command *c)
1083 {
1084 struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1085
1086 req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
1087 req->sge[1].length = sg_dma_len(req->sg_table.sgl);
1088 req->sge[1].lkey = queue->device->pd->local_dma_lkey;
1089
1090 sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1091 sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
1092 sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1093
1094 req->num_sge++;
1095 return 0;
1096 }
1097
1098 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1099 struct nvme_rdma_request *req, struct nvme_command *c)
1100 {
1101 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1102
1103 sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1104 put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1105 put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1106 sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1107 return 0;
1108 }
1109
1110 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1111 struct nvme_rdma_request *req, struct nvme_command *c,
1112 int count)
1113 {
1114 struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1115 int nr;
1116
1117 req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1118 if (WARN_ON_ONCE(!req->mr))
1119 return -EAGAIN;
1120
1121 /*
1122 * Align the MR to a 4K page size to match the ctrl page size and
1123 * the block virtual boundary.
1124 */
1125 nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1126 if (unlikely(nr < count)) {
1127 ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1128 req->mr = NULL;
1129 if (nr < 0)
1130 return nr;
1131 return -EINVAL;
1132 }
1133
1134 ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1135
1136 req->reg_cqe.done = nvme_rdma_memreg_done;
1137 memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1138 req->reg_wr.wr.opcode = IB_WR_REG_MR;
1139 req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1140 req->reg_wr.wr.num_sge = 0;
1141 req->reg_wr.mr = req->mr;
1142 req->reg_wr.key = req->mr->rkey;
1143 req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1144 IB_ACCESS_REMOTE_READ |
1145 IB_ACCESS_REMOTE_WRITE;
1146
1147 sg->addr = cpu_to_le64(req->mr->iova);
1148 put_unaligned_le24(req->mr->length, sg->length);
1149 put_unaligned_le32(req->mr->rkey, sg->key);
1150 sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1151 NVME_SGL_FMT_INVALIDATE;
1152
1153 return 0;
1154 }
1155
1156 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1157 struct request *rq, struct nvme_command *c)
1158 {
1159 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1160 struct nvme_rdma_device *dev = queue->device;
1161 struct ib_device *ibdev = dev->dev;
1162 int count, ret;
1163
1164 req->num_sge = 1;
1165 refcount_set(&req->ref, 2); /* send and recv completions */
1166
1167 c->common.flags |= NVME_CMD_SGL_METABUF;
1168
1169 if (!blk_rq_payload_bytes(rq))
1170 return nvme_rdma_set_sg_null(c);
1171
1172 req->sg_table.sgl = req->first_sgl;
1173 ret = sg_alloc_table_chained(&req->sg_table,
1174 blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1175 if (ret)
1176 return -ENOMEM;
1177
1178 req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1179
1180 count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1181 rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1182 if (unlikely(count <= 0)) {
1183 sg_free_table_chained(&req->sg_table, true);
1184 return -EIO;
1185 }
1186
1187 if (count == 1) {
1188 if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1189 blk_rq_payload_bytes(rq) <=
1190 nvme_rdma_inline_data_size(queue))
1191 return nvme_rdma_map_sg_inline(queue, req, c);
1192
1193 if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
1194 return nvme_rdma_map_sg_single(queue, req, c);
1195 }
1196
1197 return nvme_rdma_map_sg_fr(queue, req, c, count);
1198 }
1199
1200 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1201 {
1202 struct nvme_rdma_qe *qe =
1203 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1204 struct nvme_rdma_request *req =
1205 container_of(qe, struct nvme_rdma_request, sqe);
1206 struct request *rq = blk_mq_rq_from_pdu(req);
1207
1208 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1209 nvme_rdma_wr_error(cq, wc, "SEND");
1210 return;
1211 }
1212
1213 if (refcount_dec_and_test(&req->ref))
1214 nvme_end_request(rq, req->status, req->result);
1215 }
1216
1217 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1218 struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1219 struct ib_send_wr *first)
1220 {
1221 struct ib_send_wr wr, *bad_wr;
1222 int ret;
1223
1224 sge->addr = qe->dma;
1225 sge->length = sizeof(struct nvme_command),
1226 sge->lkey = queue->device->pd->local_dma_lkey;
1227
1228 wr.next = NULL;
1229 wr.wr_cqe = &qe->cqe;
1230 wr.sg_list = sge;
1231 wr.num_sge = num_sge;
1232 wr.opcode = IB_WR_SEND;
1233 wr.send_flags = IB_SEND_SIGNALED;
1234
1235 if (first)
1236 first->next = &wr;
1237 else
1238 first = &wr;
1239
1240 ret = ib_post_send(queue->qp, first, &bad_wr);
1241 if (unlikely(ret)) {
1242 dev_err(queue->ctrl->ctrl.device,
1243 "%s failed with error code %d\n", __func__, ret);
1244 }
1245 return ret;
1246 }
1247
1248 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1249 struct nvme_rdma_qe *qe)
1250 {
1251 struct ib_recv_wr wr, *bad_wr;
1252 struct ib_sge list;
1253 int ret;
1254
1255 list.addr = qe->dma;
1256 list.length = sizeof(struct nvme_completion);
1257 list.lkey = queue->device->pd->local_dma_lkey;
1258
1259 qe->cqe.done = nvme_rdma_recv_done;
1260
1261 wr.next = NULL;
1262 wr.wr_cqe = &qe->cqe;
1263 wr.sg_list = &list;
1264 wr.num_sge = 1;
1265
1266 ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1267 if (unlikely(ret)) {
1268 dev_err(queue->ctrl->ctrl.device,
1269 "%s failed with error code %d\n", __func__, ret);
1270 }
1271 return ret;
1272 }
1273
1274 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1275 {
1276 u32 queue_idx = nvme_rdma_queue_idx(queue);
1277
1278 if (queue_idx == 0)
1279 return queue->ctrl->admin_tag_set.tags[queue_idx];
1280 return queue->ctrl->tag_set.tags[queue_idx - 1];
1281 }
1282
1283 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1284 {
1285 if (unlikely(wc->status != IB_WC_SUCCESS))
1286 nvme_rdma_wr_error(cq, wc, "ASYNC");
1287 }
1288
1289 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1290 {
1291 struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1292 struct nvme_rdma_queue *queue = &ctrl->queues[0];
1293 struct ib_device *dev = queue->device->dev;
1294 struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1295 struct nvme_command *cmd = sqe->data;
1296 struct ib_sge sge;
1297 int ret;
1298
1299 ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1300
1301 memset(cmd, 0, sizeof(*cmd));
1302 cmd->common.opcode = nvme_admin_async_event;
1303 cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1304 cmd->common.flags |= NVME_CMD_SGL_METABUF;
1305 nvme_rdma_set_sg_null(cmd);
1306
1307 sqe->cqe.done = nvme_rdma_async_done;
1308
1309 ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1310 DMA_TO_DEVICE);
1311
1312 ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1313 WARN_ON_ONCE(ret);
1314 }
1315
1316 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1317 struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1318 {
1319 struct request *rq;
1320 struct nvme_rdma_request *req;
1321 int ret = 0;
1322
1323 rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1324 if (!rq) {
1325 dev_err(queue->ctrl->ctrl.device,
1326 "tag 0x%x on QP %#x not found\n",
1327 cqe->command_id, queue->qp->qp_num);
1328 nvme_rdma_error_recovery(queue->ctrl);
1329 return ret;
1330 }
1331 req = blk_mq_rq_to_pdu(rq);
1332
1333 req->status = cqe->status;
1334 req->result = cqe->result;
1335
1336 if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1337 if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1338 dev_err(queue->ctrl->ctrl.device,
1339 "Bogus remote invalidation for rkey %#x\n",
1340 req->mr->rkey);
1341 nvme_rdma_error_recovery(queue->ctrl);
1342 }
1343 } else if (req->mr) {
1344 ret = nvme_rdma_inv_rkey(queue, req);
1345 if (unlikely(ret < 0)) {
1346 dev_err(queue->ctrl->ctrl.device,
1347 "Queueing INV WR for rkey %#x failed (%d)\n",
1348 req->mr->rkey, ret);
1349 nvme_rdma_error_recovery(queue->ctrl);
1350 }
1351 /* the local invalidation completion will end the request */
1352 return 0;
1353 }
1354
1355 if (refcount_dec_and_test(&req->ref)) {
1356 if (rq->tag == tag)
1357 ret = 1;
1358 nvme_end_request(rq, req->status, req->result);
1359 }
1360
1361 return ret;
1362 }
1363
1364 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1365 {
1366 struct nvme_rdma_qe *qe =
1367 container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1368 struct nvme_rdma_queue *queue = cq->cq_context;
1369 struct ib_device *ibdev = queue->device->dev;
1370 struct nvme_completion *cqe = qe->data;
1371 const size_t len = sizeof(struct nvme_completion);
1372 int ret = 0;
1373
1374 if (unlikely(wc->status != IB_WC_SUCCESS)) {
1375 nvme_rdma_wr_error(cq, wc, "RECV");
1376 return 0;
1377 }
1378
1379 ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1380 /*
1381 * AEN requests are special as they don't time out and can
1382 * survive any kind of queue freeze and often don't respond to
1383 * aborts. We don't even bother to allocate a struct request
1384 * for them but rather special case them here.
1385 */
1386 if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1387 cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1388 nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1389 &cqe->result);
1390 else
1391 ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1392 ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1393
1394 nvme_rdma_post_recv(queue, qe);
1395 return ret;
1396 }
1397
1398 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1399 {
1400 __nvme_rdma_recv_done(cq, wc, -1);
1401 }
1402
1403 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1404 {
1405 int ret, i;
1406
1407 for (i = 0; i < queue->queue_size; i++) {
1408 ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1409 if (ret)
1410 goto out_destroy_queue_ib;
1411 }
1412
1413 return 0;
1414
1415 out_destroy_queue_ib:
1416 nvme_rdma_destroy_queue_ib(queue);
1417 return ret;
1418 }
1419
1420 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1421 struct rdma_cm_event *ev)
1422 {
1423 struct rdma_cm_id *cm_id = queue->cm_id;
1424 int status = ev->status;
1425 const char *rej_msg;
1426 const struct nvme_rdma_cm_rej *rej_data;
1427 u8 rej_data_len;
1428
1429 rej_msg = rdma_reject_msg(cm_id, status);
1430 rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1431
1432 if (rej_data && rej_data_len >= sizeof(u16)) {
1433 u16 sts = le16_to_cpu(rej_data->sts);
1434
1435 dev_err(queue->ctrl->ctrl.device,
1436 "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1437 status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1438 } else {
1439 dev_err(queue->ctrl->ctrl.device,
1440 "Connect rejected: status %d (%s).\n", status, rej_msg);
1441 }
1442
1443 return -ECONNRESET;
1444 }
1445
1446 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1447 {
1448 int ret;
1449
1450 ret = nvme_rdma_create_queue_ib(queue);
1451 if (ret)
1452 return ret;
1453
1454 ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1455 if (ret) {
1456 dev_err(queue->ctrl->ctrl.device,
1457 "rdma_resolve_route failed (%d).\n",
1458 queue->cm_error);
1459 goto out_destroy_queue;
1460 }
1461
1462 return 0;
1463
1464 out_destroy_queue:
1465 nvme_rdma_destroy_queue_ib(queue);
1466 return ret;
1467 }
1468
1469 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1470 {
1471 struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1472 struct rdma_conn_param param = { };
1473 struct nvme_rdma_cm_req priv = { };
1474 int ret;
1475
1476 param.qp_num = queue->qp->qp_num;
1477 param.flow_control = 1;
1478
1479 param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1480 /* maximum retry count */
1481 param.retry_count = 7;
1482 param.rnr_retry_count = 7;
1483 param.private_data = &priv;
1484 param.private_data_len = sizeof(priv);
1485
1486 priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1487 priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1488 /*
1489 * set the admin queue depth to the minimum size
1490 * specified by the Fabrics standard.
1491 */
1492 if (priv.qid == 0) {
1493 priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1494 priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1495 } else {
1496 /*
1497 * current interpretation of the fabrics spec
1498 * is at minimum you make hrqsize sqsize+1, or a
1499 * 1's based representation of sqsize.
1500 */
1501 priv.hrqsize = cpu_to_le16(queue->queue_size);
1502 priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1503 }
1504
1505 ret = rdma_connect(queue->cm_id, &param);
1506 if (ret) {
1507 dev_err(ctrl->ctrl.device,
1508 "rdma_connect failed (%d).\n", ret);
1509 goto out_destroy_queue_ib;
1510 }
1511
1512 return 0;
1513
1514 out_destroy_queue_ib:
1515 nvme_rdma_destroy_queue_ib(queue);
1516 return ret;
1517 }
1518
1519 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1520 struct rdma_cm_event *ev)
1521 {
1522 struct nvme_rdma_queue *queue = cm_id->context;
1523 int cm_error = 0;
1524
1525 dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1526 rdma_event_msg(ev->event), ev->event,
1527 ev->status, cm_id);
1528
1529 switch (ev->event) {
1530 case RDMA_CM_EVENT_ADDR_RESOLVED:
1531 cm_error = nvme_rdma_addr_resolved(queue);
1532 break;
1533 case RDMA_CM_EVENT_ROUTE_RESOLVED:
1534 cm_error = nvme_rdma_route_resolved(queue);
1535 break;
1536 case RDMA_CM_EVENT_ESTABLISHED:
1537 queue->cm_error = nvme_rdma_conn_established(queue);
1538 /* complete cm_done regardless of success/failure */
1539 complete(&queue->cm_done);
1540 return 0;
1541 case RDMA_CM_EVENT_REJECTED:
1542 nvme_rdma_destroy_queue_ib(queue);
1543 cm_error = nvme_rdma_conn_rejected(queue, ev);
1544 break;
1545 case RDMA_CM_EVENT_ROUTE_ERROR:
1546 case RDMA_CM_EVENT_CONNECT_ERROR:
1547 case RDMA_CM_EVENT_UNREACHABLE:
1548 nvme_rdma_destroy_queue_ib(queue);
1549 case RDMA_CM_EVENT_ADDR_ERROR:
1550 dev_dbg(queue->ctrl->ctrl.device,
1551 "CM error event %d\n", ev->event);
1552 cm_error = -ECONNRESET;
1553 break;
1554 case RDMA_CM_EVENT_DISCONNECTED:
1555 case RDMA_CM_EVENT_ADDR_CHANGE:
1556 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1557 dev_dbg(queue->ctrl->ctrl.device,
1558 "disconnect received - connection closed\n");
1559 nvme_rdma_error_recovery(queue->ctrl);
1560 break;
1561 case RDMA_CM_EVENT_DEVICE_REMOVAL:
1562 /* device removal is handled via the ib_client API */
1563 break;
1564 default:
1565 dev_err(queue->ctrl->ctrl.device,
1566 "Unexpected RDMA CM event (%d)\n", ev->event);
1567 nvme_rdma_error_recovery(queue->ctrl);
1568 break;
1569 }
1570
1571 if (cm_error) {
1572 queue->cm_error = cm_error;
1573 complete(&queue->cm_done);
1574 }
1575
1576 return 0;
1577 }
1578
1579 static enum blk_eh_timer_return
1580 nvme_rdma_timeout(struct request *rq, bool reserved)
1581 {
1582 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1583
1584 dev_warn(req->queue->ctrl->ctrl.device,
1585 "I/O %d QID %d timeout, reset controller\n",
1586 rq->tag, nvme_rdma_queue_idx(req->queue));
1587
1588 /* queue error recovery */
1589 nvme_rdma_error_recovery(req->queue->ctrl);
1590
1591 /* fail with DNR on cmd timeout */
1592 nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1593
1594 return BLK_EH_HANDLED;
1595 }
1596
1597 /*
1598 * We cannot accept any other command until the Connect command has completed.
1599 */
1600 static inline blk_status_t
1601 nvme_rdma_is_ready(struct nvme_rdma_queue *queue, struct request *rq)
1602 {
1603 if (unlikely(!test_bit(NVME_RDMA_Q_LIVE, &queue->flags)))
1604 return nvmf_check_init_req(&queue->ctrl->ctrl, rq);
1605 return BLK_STS_OK;
1606 }
1607
1608 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1609 const struct blk_mq_queue_data *bd)
1610 {
1611 struct nvme_ns *ns = hctx->queue->queuedata;
1612 struct nvme_rdma_queue *queue = hctx->driver_data;
1613 struct request *rq = bd->rq;
1614 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1615 struct nvme_rdma_qe *sqe = &req->sqe;
1616 struct nvme_command *c = sqe->data;
1617 struct ib_device *dev;
1618 blk_status_t ret;
1619 int err;
1620
1621 WARN_ON_ONCE(rq->tag < 0);
1622
1623 ret = nvme_rdma_is_ready(queue, rq);
1624 if (unlikely(ret))
1625 return ret;
1626
1627 dev = queue->device->dev;
1628 ib_dma_sync_single_for_cpu(dev, sqe->dma,
1629 sizeof(struct nvme_command), DMA_TO_DEVICE);
1630
1631 ret = nvme_setup_cmd(ns, rq, c);
1632 if (ret)
1633 return ret;
1634
1635 blk_mq_start_request(rq);
1636
1637 err = nvme_rdma_map_data(queue, rq, c);
1638 if (unlikely(err < 0)) {
1639 dev_err(queue->ctrl->ctrl.device,
1640 "Failed to map data (%d)\n", err);
1641 nvme_cleanup_cmd(rq);
1642 goto err;
1643 }
1644
1645 sqe->cqe.done = nvme_rdma_send_done;
1646
1647 ib_dma_sync_single_for_device(dev, sqe->dma,
1648 sizeof(struct nvme_command), DMA_TO_DEVICE);
1649
1650 err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1651 req->mr ? &req->reg_wr.wr : NULL);
1652 if (unlikely(err)) {
1653 nvme_rdma_unmap_data(queue, rq);
1654 goto err;
1655 }
1656
1657 return BLK_STS_OK;
1658 err:
1659 if (err == -ENOMEM || err == -EAGAIN)
1660 return BLK_STS_RESOURCE;
1661 return BLK_STS_IOERR;
1662 }
1663
1664 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1665 {
1666 struct nvme_rdma_queue *queue = hctx->driver_data;
1667 struct ib_cq *cq = queue->ib_cq;
1668 struct ib_wc wc;
1669 int found = 0;
1670
1671 while (ib_poll_cq(cq, 1, &wc) > 0) {
1672 struct ib_cqe *cqe = wc.wr_cqe;
1673
1674 if (cqe) {
1675 if (cqe->done == nvme_rdma_recv_done)
1676 found |= __nvme_rdma_recv_done(cq, &wc, tag);
1677 else
1678 cqe->done(cq, &wc);
1679 }
1680 }
1681
1682 return found;
1683 }
1684
1685 static void nvme_rdma_complete_rq(struct request *rq)
1686 {
1687 struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1688
1689 nvme_rdma_unmap_data(req->queue, rq);
1690 nvme_complete_rq(rq);
1691 }
1692
1693 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1694 {
1695 struct nvme_rdma_ctrl *ctrl = set->driver_data;
1696
1697 return blk_mq_rdma_map_queues(set, ctrl->device->dev, 0);
1698 }
1699
1700 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1701 .queue_rq = nvme_rdma_queue_rq,
1702 .complete = nvme_rdma_complete_rq,
1703 .init_request = nvme_rdma_init_request,
1704 .exit_request = nvme_rdma_exit_request,
1705 .init_hctx = nvme_rdma_init_hctx,
1706 .poll = nvme_rdma_poll,
1707 .timeout = nvme_rdma_timeout,
1708 .map_queues = nvme_rdma_map_queues,
1709 };
1710
1711 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1712 .queue_rq = nvme_rdma_queue_rq,
1713 .complete = nvme_rdma_complete_rq,
1714 .init_request = nvme_rdma_init_request,
1715 .exit_request = nvme_rdma_exit_request,
1716 .init_hctx = nvme_rdma_init_admin_hctx,
1717 .timeout = nvme_rdma_timeout,
1718 };
1719
1720 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1721 {
1722 cancel_work_sync(&ctrl->err_work);
1723 cancel_delayed_work_sync(&ctrl->reconnect_work);
1724
1725 if (ctrl->ctrl.queue_count > 1) {
1726 nvme_stop_queues(&ctrl->ctrl);
1727 blk_mq_tagset_busy_iter(&ctrl->tag_set,
1728 nvme_cancel_request, &ctrl->ctrl);
1729 nvme_rdma_destroy_io_queues(ctrl, shutdown);
1730 }
1731
1732 if (shutdown)
1733 nvme_shutdown_ctrl(&ctrl->ctrl);
1734 else
1735 nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1736
1737 blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1738 blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1739 nvme_cancel_request, &ctrl->ctrl);
1740 blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1741 nvme_rdma_destroy_admin_queue(ctrl, shutdown);
1742 }
1743
1744 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1745 {
1746 nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1747 }
1748
1749 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1750 {
1751 struct nvme_rdma_ctrl *ctrl =
1752 container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1753 int ret;
1754 bool changed;
1755
1756 nvme_stop_ctrl(&ctrl->ctrl);
1757 nvme_rdma_shutdown_ctrl(ctrl, false);
1758
1759 if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1760 /* state change failure should never happen */
1761 WARN_ON_ONCE(1);
1762 return;
1763 }
1764
1765 ret = nvme_rdma_configure_admin_queue(ctrl, false);
1766 if (ret)
1767 goto out_fail;
1768
1769 if (ctrl->ctrl.queue_count > 1) {
1770 ret = nvme_rdma_configure_io_queues(ctrl, false);
1771 if (ret)
1772 goto out_fail;
1773 }
1774
1775 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1776 if (!changed) {
1777 /* state change failure is ok if we're in DELETING state */
1778 WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1779 return;
1780 }
1781
1782 nvme_start_ctrl(&ctrl->ctrl);
1783
1784 return;
1785
1786 out_fail:
1787 ++ctrl->ctrl.nr_reconnects;
1788 nvme_rdma_reconnect_or_remove(ctrl);
1789 }
1790
1791 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1792 .name = "rdma",
1793 .module = THIS_MODULE,
1794 .flags = NVME_F_FABRICS,
1795 .reg_read32 = nvmf_reg_read32,
1796 .reg_read64 = nvmf_reg_read64,
1797 .reg_write32 = nvmf_reg_write32,
1798 .free_ctrl = nvme_rdma_free_ctrl,
1799 .submit_async_event = nvme_rdma_submit_async_event,
1800 .delete_ctrl = nvme_rdma_delete_ctrl,
1801 .get_address = nvmf_get_address,
1802 };
1803
1804 static inline bool
1805 __nvme_rdma_options_match(struct nvme_rdma_ctrl *ctrl,
1806 struct nvmf_ctrl_options *opts)
1807 {
1808 char *stdport = __stringify(NVME_RDMA_IP_PORT);
1809
1810
1811 if (!nvmf_ctlr_matches_baseopts(&ctrl->ctrl, opts) ||
1812 strcmp(opts->traddr, ctrl->ctrl.opts->traddr))
1813 return false;
1814
1815 if (opts->mask & NVMF_OPT_TRSVCID &&
1816 ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1817 if (strcmp(opts->trsvcid, ctrl->ctrl.opts->trsvcid))
1818 return false;
1819 } else if (opts->mask & NVMF_OPT_TRSVCID) {
1820 if (strcmp(opts->trsvcid, stdport))
1821 return false;
1822 } else if (ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1823 if (strcmp(stdport, ctrl->ctrl.opts->trsvcid))
1824 return false;
1825 }
1826 /* else, it's a match as both have stdport. Fall to next checks */
1827
1828 /*
1829 * checking the local address is rough. In most cases, one
1830 * is not specified and the host port is selected by the stack.
1831 *
1832 * Assume no match if:
1833 * local address is specified and address is not the same
1834 * local address is not specified but remote is, or vice versa
1835 * (admin using specific host_traddr when it matters).
1836 */
1837 if (opts->mask & NVMF_OPT_HOST_TRADDR &&
1838 ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR) {
1839 if (strcmp(opts->host_traddr, ctrl->ctrl.opts->host_traddr))
1840 return false;
1841 } else if (opts->mask & NVMF_OPT_HOST_TRADDR ||
1842 ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
1843 return false;
1844 /*
1845 * if neither controller had an host port specified, assume it's
1846 * a match as everything else matched.
1847 */
1848
1849 return true;
1850 }
1851
1852 /*
1853 * Fails a connection request if it matches an existing controller
1854 * (association) with the same tuple:
1855 * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1856 *
1857 * if local address is not specified in the request, it will match an
1858 * existing controller with all the other parameters the same and no
1859 * local port address specified as well.
1860 *
1861 * The ports don't need to be compared as they are intrinsically
1862 * already matched by the port pointers supplied.
1863 */
1864 static bool
1865 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1866 {
1867 struct nvme_rdma_ctrl *ctrl;
1868 bool found = false;
1869
1870 mutex_lock(&nvme_rdma_ctrl_mutex);
1871 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1872 found = __nvme_rdma_options_match(ctrl, opts);
1873 if (found)
1874 break;
1875 }
1876 mutex_unlock(&nvme_rdma_ctrl_mutex);
1877
1878 return found;
1879 }
1880
1881 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1882 struct nvmf_ctrl_options *opts)
1883 {
1884 struct nvme_rdma_ctrl *ctrl;
1885 int ret;
1886 bool changed;
1887 char *port;
1888
1889 ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1890 if (!ctrl)
1891 return ERR_PTR(-ENOMEM);
1892 ctrl->ctrl.opts = opts;
1893 INIT_LIST_HEAD(&ctrl->list);
1894
1895 if (opts->mask & NVMF_OPT_TRSVCID)
1896 port = opts->trsvcid;
1897 else
1898 port = __stringify(NVME_RDMA_IP_PORT);
1899
1900 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1901 opts->traddr, port, &ctrl->addr);
1902 if (ret) {
1903 pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1904 goto out_free_ctrl;
1905 }
1906
1907 if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1908 ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1909 opts->host_traddr, NULL, &ctrl->src_addr);
1910 if (ret) {
1911 pr_err("malformed src address passed: %s\n",
1912 opts->host_traddr);
1913 goto out_free_ctrl;
1914 }
1915 }
1916
1917 if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1918 ret = -EALREADY;
1919 goto out_free_ctrl;
1920 }
1921
1922 ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1923 0 /* no quirks, we're perfect! */);
1924 if (ret)
1925 goto out_free_ctrl;
1926
1927 INIT_DELAYED_WORK(&ctrl->reconnect_work,
1928 nvme_rdma_reconnect_ctrl_work);
1929 INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1930 INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1931
1932 ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1933 ctrl->ctrl.sqsize = opts->queue_size - 1;
1934 ctrl->ctrl.kato = opts->kato;
1935
1936 ret = -ENOMEM;
1937 ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1938 GFP_KERNEL);
1939 if (!ctrl->queues)
1940 goto out_uninit_ctrl;
1941
1942 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
1943 WARN_ON_ONCE(!changed);
1944
1945 ret = nvme_rdma_configure_admin_queue(ctrl, true);
1946 if (ret)
1947 goto out_kfree_queues;
1948
1949 /* sanity check icdoff */
1950 if (ctrl->ctrl.icdoff) {
1951 dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1952 ret = -EINVAL;
1953 goto out_remove_admin_queue;
1954 }
1955
1956 /* sanity check keyed sgls */
1957 if (!(ctrl->ctrl.sgls & (1 << 20))) {
1958 dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1959 ret = -EINVAL;
1960 goto out_remove_admin_queue;
1961 }
1962
1963 if (opts->queue_size > ctrl->ctrl.maxcmd) {
1964 /* warn if maxcmd is lower than queue_size */
1965 dev_warn(ctrl->ctrl.device,
1966 "queue_size %zu > ctrl maxcmd %u, clamping down\n",
1967 opts->queue_size, ctrl->ctrl.maxcmd);
1968 opts->queue_size = ctrl->ctrl.maxcmd;
1969 }
1970
1971 if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
1972 /* warn if sqsize is lower than queue_size */
1973 dev_warn(ctrl->ctrl.device,
1974 "queue_size %zu > ctrl sqsize %u, clamping down\n",
1975 opts->queue_size, ctrl->ctrl.sqsize + 1);
1976 opts->queue_size = ctrl->ctrl.sqsize + 1;
1977 }
1978
1979 if (opts->nr_io_queues) {
1980 ret = nvme_rdma_configure_io_queues(ctrl, true);
1981 if (ret)
1982 goto out_remove_admin_queue;
1983 }
1984
1985 changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1986 WARN_ON_ONCE(!changed);
1987
1988 dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1989 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1990
1991 nvme_get_ctrl(&ctrl->ctrl);
1992
1993 mutex_lock(&nvme_rdma_ctrl_mutex);
1994 list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1995 mutex_unlock(&nvme_rdma_ctrl_mutex);
1996
1997 nvme_start_ctrl(&ctrl->ctrl);
1998
1999 return &ctrl->ctrl;
2000
2001 out_remove_admin_queue:
2002 nvme_rdma_destroy_admin_queue(ctrl, true);
2003 out_kfree_queues:
2004 kfree(ctrl->queues);
2005 out_uninit_ctrl:
2006 nvme_uninit_ctrl(&ctrl->ctrl);
2007 nvme_put_ctrl(&ctrl->ctrl);
2008 if (ret > 0)
2009 ret = -EIO;
2010 return ERR_PTR(ret);
2011 out_free_ctrl:
2012 kfree(ctrl);
2013 return ERR_PTR(ret);
2014 }
2015
2016 static struct nvmf_transport_ops nvme_rdma_transport = {
2017 .name = "rdma",
2018 .module = THIS_MODULE,
2019 .required_opts = NVMF_OPT_TRADDR,
2020 .allowed_opts = NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2021 NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
2022 .create_ctrl = nvme_rdma_create_ctrl,
2023 };
2024
2025 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2026 {
2027 struct nvme_rdma_ctrl *ctrl;
2028
2029 /* Delete all controllers using this device */
2030 mutex_lock(&nvme_rdma_ctrl_mutex);
2031 list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2032 if (ctrl->device->dev != ib_device)
2033 continue;
2034 dev_info(ctrl->ctrl.device,
2035 "Removing ctrl: NQN \"%s\", addr %pISp\n",
2036 ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2037 nvme_delete_ctrl(&ctrl->ctrl);
2038 }
2039 mutex_unlock(&nvme_rdma_ctrl_mutex);
2040
2041 flush_workqueue(nvme_delete_wq);
2042 }
2043
2044 static struct ib_client nvme_rdma_ib_client = {
2045 .name = "nvme_rdma",
2046 .remove = nvme_rdma_remove_one
2047 };
2048
2049 static int __init nvme_rdma_init_module(void)
2050 {
2051 int ret;
2052
2053 ret = ib_register_client(&nvme_rdma_ib_client);
2054 if (ret)
2055 return ret;
2056
2057 ret = nvmf_register_transport(&nvme_rdma_transport);
2058 if (ret)
2059 goto err_unreg_client;
2060
2061 return 0;
2062
2063 err_unreg_client:
2064 ib_unregister_client(&nvme_rdma_ib_client);
2065 return ret;
2066 }
2067
2068 static void __exit nvme_rdma_cleanup_module(void)
2069 {
2070 nvmf_unregister_transport(&nvme_rdma_transport);
2071 ib_unregister_client(&nvme_rdma_ib_client);
2072 }
2073
2074 module_init(nvme_rdma_init_module);
2075 module_exit(nvme_rdma_cleanup_module);
2076
2077 MODULE_LICENSE("GPL v2");
CRIS linux kernel tree