xref: /openbmc/linux/drivers/nvme/host/rdma.c (revision 0c874100)
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	4
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 	unsigned int		num_inline_segments;
51 };
52 
53 struct nvme_rdma_qe {
54 	struct ib_cqe		cqe;
55 	void			*data;
56 	u64			dma;
57 };
58 
59 struct nvme_rdma_queue;
60 struct nvme_rdma_request {
61 	struct nvme_request	req;
62 	struct ib_mr		*mr;
63 	struct nvme_rdma_qe	sqe;
64 	union nvme_result	result;
65 	__le16			status;
66 	refcount_t		ref;
67 	struct ib_sge		sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
68 	u32			num_sge;
69 	int			nents;
70 	struct ib_reg_wr	reg_wr;
71 	struct ib_cqe		reg_cqe;
72 	struct nvme_rdma_queue  *queue;
73 	struct sg_table		sg_table;
74 	struct scatterlist	first_sgl[];
75 };
76 
77 enum nvme_rdma_queue_flags {
78 	NVME_RDMA_Q_ALLOCATED		= 0,
79 	NVME_RDMA_Q_LIVE		= 1,
80 	NVME_RDMA_Q_TR_READY		= 2,
81 };
82 
83 struct nvme_rdma_queue {
84 	struct nvme_rdma_qe	*rsp_ring;
85 	int			queue_size;
86 	size_t			cmnd_capsule_len;
87 	struct nvme_rdma_ctrl	*ctrl;
88 	struct nvme_rdma_device	*device;
89 	struct ib_cq		*ib_cq;
90 	struct ib_qp		*qp;
91 
92 	unsigned long		flags;
93 	struct rdma_cm_id	*cm_id;
94 	int			cm_error;
95 	struct completion	cm_done;
96 };
97 
98 struct nvme_rdma_ctrl {
99 	/* read only in the hot path */
100 	struct nvme_rdma_queue	*queues;
101 
102 	/* other member variables */
103 	struct blk_mq_tag_set	tag_set;
104 	struct work_struct	err_work;
105 
106 	struct nvme_rdma_qe	async_event_sqe;
107 
108 	struct delayed_work	reconnect_work;
109 
110 	struct list_head	list;
111 
112 	struct blk_mq_tag_set	admin_tag_set;
113 	struct nvme_rdma_device	*device;
114 
115 	u32			max_fr_pages;
116 
117 	struct sockaddr_storage addr;
118 	struct sockaddr_storage src_addr;
119 
120 	struct nvme_ctrl	ctrl;
121 	bool			use_inline_data;
122 };
123 
124 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
125 {
126 	return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
127 }
128 
129 static LIST_HEAD(device_list);
130 static DEFINE_MUTEX(device_list_mutex);
131 
132 static LIST_HEAD(nvme_rdma_ctrl_list);
133 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
134 
135 /*
136  * Disabling this option makes small I/O goes faster, but is fundamentally
137  * unsafe.  With it turned off we will have to register a global rkey that
138  * allows read and write access to all physical memory.
139  */
140 static bool register_always = true;
141 module_param(register_always, bool, 0444);
142 MODULE_PARM_DESC(register_always,
143 	 "Use memory registration even for contiguous memory regions");
144 
145 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
146 		struct rdma_cm_event *event);
147 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
148 
149 static const struct blk_mq_ops nvme_rdma_mq_ops;
150 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
151 
152 /* XXX: really should move to a generic header sooner or later.. */
153 static inline void put_unaligned_le24(u32 val, u8 *p)
154 {
155 	*p++ = val;
156 	*p++ = val >> 8;
157 	*p++ = val >> 16;
158 }
159 
160 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
161 {
162 	return queue - queue->ctrl->queues;
163 }
164 
165 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
166 {
167 	return queue->cmnd_capsule_len - sizeof(struct nvme_command);
168 }
169 
170 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
171 		size_t capsule_size, enum dma_data_direction dir)
172 {
173 	ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
174 	kfree(qe->data);
175 }
176 
177 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
178 		size_t capsule_size, enum dma_data_direction dir)
179 {
180 	qe->data = kzalloc(capsule_size, GFP_KERNEL);
181 	if (!qe->data)
182 		return -ENOMEM;
183 
184 	qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
185 	if (ib_dma_mapping_error(ibdev, qe->dma)) {
186 		kfree(qe->data);
187 		return -ENOMEM;
188 	}
189 
190 	return 0;
191 }
192 
193 static void nvme_rdma_free_ring(struct ib_device *ibdev,
194 		struct nvme_rdma_qe *ring, size_t ib_queue_size,
195 		size_t capsule_size, enum dma_data_direction dir)
196 {
197 	int i;
198 
199 	for (i = 0; i < ib_queue_size; i++)
200 		nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
201 	kfree(ring);
202 }
203 
204 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
205 		size_t ib_queue_size, size_t capsule_size,
206 		enum dma_data_direction dir)
207 {
208 	struct nvme_rdma_qe *ring;
209 	int i;
210 
211 	ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
212 	if (!ring)
213 		return NULL;
214 
215 	for (i = 0; i < ib_queue_size; i++) {
216 		if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
217 			goto out_free_ring;
218 	}
219 
220 	return ring;
221 
222 out_free_ring:
223 	nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
224 	return NULL;
225 }
226 
227 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
228 {
229 	pr_debug("QP event %s (%d)\n",
230 		 ib_event_msg(event->event), event->event);
231 
232 }
233 
234 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
235 {
236 	int ret;
237 
238 	ret = wait_for_completion_interruptible_timeout(&queue->cm_done,
239 			msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
240 	if (ret < 0)
241 		return ret;
242 	if (ret == 0)
243 		return -ETIMEDOUT;
244 	WARN_ON_ONCE(queue->cm_error > 0);
245 	return queue->cm_error;
246 }
247 
248 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
249 {
250 	struct nvme_rdma_device *dev = queue->device;
251 	struct ib_qp_init_attr init_attr;
252 	int ret;
253 
254 	memset(&init_attr, 0, sizeof(init_attr));
255 	init_attr.event_handler = nvme_rdma_qp_event;
256 	/* +1 for drain */
257 	init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
258 	/* +1 for drain */
259 	init_attr.cap.max_recv_wr = queue->queue_size + 1;
260 	init_attr.cap.max_recv_sge = 1;
261 	init_attr.cap.max_send_sge = 1 + dev->num_inline_segments;
262 	init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
263 	init_attr.qp_type = IB_QPT_RC;
264 	init_attr.send_cq = queue->ib_cq;
265 	init_attr.recv_cq = queue->ib_cq;
266 
267 	ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
268 
269 	queue->qp = queue->cm_id->qp;
270 	return ret;
271 }
272 
273 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
274 		struct request *rq, unsigned int hctx_idx)
275 {
276 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
277 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
278 	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
279 	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
280 	struct nvme_rdma_device *dev = queue->device;
281 
282 	nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
283 			DMA_TO_DEVICE);
284 }
285 
286 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
287 		struct request *rq, unsigned int hctx_idx,
288 		unsigned int numa_node)
289 {
290 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
291 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
292 	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
293 	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
294 	struct nvme_rdma_device *dev = queue->device;
295 	struct ib_device *ibdev = dev->dev;
296 	int ret;
297 
298 	nvme_req(rq)->ctrl = &ctrl->ctrl;
299 	ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
300 			DMA_TO_DEVICE);
301 	if (ret)
302 		return ret;
303 
304 	req->queue = queue;
305 
306 	return 0;
307 }
308 
309 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
310 		unsigned int hctx_idx)
311 {
312 	struct nvme_rdma_ctrl *ctrl = data;
313 	struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
314 
315 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
316 
317 	hctx->driver_data = queue;
318 	return 0;
319 }
320 
321 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
322 		unsigned int hctx_idx)
323 {
324 	struct nvme_rdma_ctrl *ctrl = data;
325 	struct nvme_rdma_queue *queue = &ctrl->queues[0];
326 
327 	BUG_ON(hctx_idx != 0);
328 
329 	hctx->driver_data = queue;
330 	return 0;
331 }
332 
333 static void nvme_rdma_free_dev(struct kref *ref)
334 {
335 	struct nvme_rdma_device *ndev =
336 		container_of(ref, struct nvme_rdma_device, ref);
337 
338 	mutex_lock(&device_list_mutex);
339 	list_del(&ndev->entry);
340 	mutex_unlock(&device_list_mutex);
341 
342 	ib_dealloc_pd(ndev->pd);
343 	kfree(ndev);
344 }
345 
346 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
347 {
348 	kref_put(&dev->ref, nvme_rdma_free_dev);
349 }
350 
351 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
352 {
353 	return kref_get_unless_zero(&dev->ref);
354 }
355 
356 static struct nvme_rdma_device *
357 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
358 {
359 	struct nvme_rdma_device *ndev;
360 
361 	mutex_lock(&device_list_mutex);
362 	list_for_each_entry(ndev, &device_list, entry) {
363 		if (ndev->dev->node_guid == cm_id->device->node_guid &&
364 		    nvme_rdma_dev_get(ndev))
365 			goto out_unlock;
366 	}
367 
368 	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
369 	if (!ndev)
370 		goto out_err;
371 
372 	ndev->dev = cm_id->device;
373 	kref_init(&ndev->ref);
374 
375 	ndev->pd = ib_alloc_pd(ndev->dev,
376 		register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
377 	if (IS_ERR(ndev->pd))
378 		goto out_free_dev;
379 
380 	if (!(ndev->dev->attrs.device_cap_flags &
381 	      IB_DEVICE_MEM_MGT_EXTENSIONS)) {
382 		dev_err(&ndev->dev->dev,
383 			"Memory registrations not supported.\n");
384 		goto out_free_pd;
385 	}
386 
387 	ndev->num_inline_segments = min(NVME_RDMA_MAX_INLINE_SEGMENTS,
388 					ndev->dev->attrs.max_send_sge - 1);
389 	list_add(&ndev->entry, &device_list);
390 out_unlock:
391 	mutex_unlock(&device_list_mutex);
392 	return ndev;
393 
394 out_free_pd:
395 	ib_dealloc_pd(ndev->pd);
396 out_free_dev:
397 	kfree(ndev);
398 out_err:
399 	mutex_unlock(&device_list_mutex);
400 	return NULL;
401 }
402 
403 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
404 {
405 	struct nvme_rdma_device *dev;
406 	struct ib_device *ibdev;
407 
408 	if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
409 		return;
410 
411 	dev = queue->device;
412 	ibdev = dev->dev;
413 
414 	ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
415 
416 	/*
417 	 * The cm_id object might have been destroyed during RDMA connection
418 	 * establishment error flow to avoid getting other cma events, thus
419 	 * the destruction of the QP shouldn't use rdma_cm API.
420 	 */
421 	ib_destroy_qp(queue->qp);
422 	ib_free_cq(queue->ib_cq);
423 
424 	nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
425 			sizeof(struct nvme_completion), DMA_FROM_DEVICE);
426 
427 	nvme_rdma_dev_put(dev);
428 }
429 
430 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev)
431 {
432 	return min_t(u32, NVME_RDMA_MAX_SEGMENTS,
433 		     ibdev->attrs.max_fast_reg_page_list_len);
434 }
435 
436 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
437 {
438 	struct ib_device *ibdev;
439 	const int send_wr_factor = 3;			/* MR, SEND, INV */
440 	const int cq_factor = send_wr_factor + 1;	/* + RECV */
441 	int comp_vector, idx = nvme_rdma_queue_idx(queue);
442 	int ret;
443 
444 	queue->device = nvme_rdma_find_get_device(queue->cm_id);
445 	if (!queue->device) {
446 		dev_err(queue->cm_id->device->dev.parent,
447 			"no client data found!\n");
448 		return -ECONNREFUSED;
449 	}
450 	ibdev = queue->device->dev;
451 
452 	/*
453 	 * Spread I/O queues completion vectors according their queue index.
454 	 * Admin queues can always go on completion vector 0.
455 	 */
456 	comp_vector = idx == 0 ? idx : idx - 1;
457 
458 	/* +1 for ib_stop_cq */
459 	queue->ib_cq = ib_alloc_cq(ibdev, queue,
460 				cq_factor * queue->queue_size + 1,
461 				comp_vector, IB_POLL_SOFTIRQ);
462 	if (IS_ERR(queue->ib_cq)) {
463 		ret = PTR_ERR(queue->ib_cq);
464 		goto out_put_dev;
465 	}
466 
467 	ret = nvme_rdma_create_qp(queue, send_wr_factor);
468 	if (ret)
469 		goto out_destroy_ib_cq;
470 
471 	queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
472 			sizeof(struct nvme_completion), DMA_FROM_DEVICE);
473 	if (!queue->rsp_ring) {
474 		ret = -ENOMEM;
475 		goto out_destroy_qp;
476 	}
477 
478 	ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
479 			      queue->queue_size,
480 			      IB_MR_TYPE_MEM_REG,
481 			      nvme_rdma_get_max_fr_pages(ibdev));
482 	if (ret) {
483 		dev_err(queue->ctrl->ctrl.device,
484 			"failed to initialize MR pool sized %d for QID %d\n",
485 			queue->queue_size, idx);
486 		goto out_destroy_ring;
487 	}
488 
489 	set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
490 
491 	return 0;
492 
493 out_destroy_ring:
494 	nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
495 			    sizeof(struct nvme_completion), DMA_FROM_DEVICE);
496 out_destroy_qp:
497 	rdma_destroy_qp(queue->cm_id);
498 out_destroy_ib_cq:
499 	ib_free_cq(queue->ib_cq);
500 out_put_dev:
501 	nvme_rdma_dev_put(queue->device);
502 	return ret;
503 }
504 
505 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
506 		int idx, size_t queue_size)
507 {
508 	struct nvme_rdma_queue *queue;
509 	struct sockaddr *src_addr = NULL;
510 	int ret;
511 
512 	queue = &ctrl->queues[idx];
513 	queue->ctrl = ctrl;
514 	init_completion(&queue->cm_done);
515 
516 	if (idx > 0)
517 		queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
518 	else
519 		queue->cmnd_capsule_len = sizeof(struct nvme_command);
520 
521 	queue->queue_size = queue_size;
522 
523 	queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
524 			RDMA_PS_TCP, IB_QPT_RC);
525 	if (IS_ERR(queue->cm_id)) {
526 		dev_info(ctrl->ctrl.device,
527 			"failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
528 		return PTR_ERR(queue->cm_id);
529 	}
530 
531 	if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
532 		src_addr = (struct sockaddr *)&ctrl->src_addr;
533 
534 	queue->cm_error = -ETIMEDOUT;
535 	ret = rdma_resolve_addr(queue->cm_id, src_addr,
536 			(struct sockaddr *)&ctrl->addr,
537 			NVME_RDMA_CONNECT_TIMEOUT_MS);
538 	if (ret) {
539 		dev_info(ctrl->ctrl.device,
540 			"rdma_resolve_addr failed (%d).\n", ret);
541 		goto out_destroy_cm_id;
542 	}
543 
544 	ret = nvme_rdma_wait_for_cm(queue);
545 	if (ret) {
546 		dev_info(ctrl->ctrl.device,
547 			"rdma connection establishment failed (%d)\n", ret);
548 		goto out_destroy_cm_id;
549 	}
550 
551 	set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
552 
553 	return 0;
554 
555 out_destroy_cm_id:
556 	rdma_destroy_id(queue->cm_id);
557 	nvme_rdma_destroy_queue_ib(queue);
558 	return ret;
559 }
560 
561 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
562 {
563 	if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
564 		return;
565 
566 	rdma_disconnect(queue->cm_id);
567 	ib_drain_qp(queue->qp);
568 }
569 
570 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
571 {
572 	if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
573 		return;
574 
575 	nvme_rdma_destroy_queue_ib(queue);
576 	rdma_destroy_id(queue->cm_id);
577 }
578 
579 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
580 {
581 	int i;
582 
583 	for (i = 1; i < ctrl->ctrl.queue_count; i++)
584 		nvme_rdma_free_queue(&ctrl->queues[i]);
585 }
586 
587 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
588 {
589 	int i;
590 
591 	for (i = 1; i < ctrl->ctrl.queue_count; i++)
592 		nvme_rdma_stop_queue(&ctrl->queues[i]);
593 }
594 
595 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
596 {
597 	int ret;
598 
599 	if (idx)
600 		ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
601 	else
602 		ret = nvmf_connect_admin_queue(&ctrl->ctrl);
603 
604 	if (!ret)
605 		set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags);
606 	else
607 		dev_info(ctrl->ctrl.device,
608 			"failed to connect queue: %d ret=%d\n", idx, ret);
609 	return ret;
610 }
611 
612 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
613 {
614 	int i, ret = 0;
615 
616 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
617 		ret = nvme_rdma_start_queue(ctrl, i);
618 		if (ret)
619 			goto out_stop_queues;
620 	}
621 
622 	return 0;
623 
624 out_stop_queues:
625 	for (i--; i >= 1; i--)
626 		nvme_rdma_stop_queue(&ctrl->queues[i]);
627 	return ret;
628 }
629 
630 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
631 {
632 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
633 	struct ib_device *ibdev = ctrl->device->dev;
634 	unsigned int nr_io_queues;
635 	int i, ret;
636 
637 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
638 
639 	/*
640 	 * we map queues according to the device irq vectors for
641 	 * optimal locality so we don't need more queues than
642 	 * completion vectors.
643 	 */
644 	nr_io_queues = min_t(unsigned int, nr_io_queues,
645 				ibdev->num_comp_vectors);
646 
647 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
648 	if (ret)
649 		return ret;
650 
651 	ctrl->ctrl.queue_count = nr_io_queues + 1;
652 	if (ctrl->ctrl.queue_count < 2)
653 		return 0;
654 
655 	dev_info(ctrl->ctrl.device,
656 		"creating %d I/O queues.\n", nr_io_queues);
657 
658 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
659 		ret = nvme_rdma_alloc_queue(ctrl, i,
660 				ctrl->ctrl.sqsize + 1);
661 		if (ret)
662 			goto out_free_queues;
663 	}
664 
665 	return 0;
666 
667 out_free_queues:
668 	for (i--; i >= 1; i--)
669 		nvme_rdma_free_queue(&ctrl->queues[i]);
670 
671 	return ret;
672 }
673 
674 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl,
675 		struct blk_mq_tag_set *set)
676 {
677 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
678 
679 	blk_mq_free_tag_set(set);
680 	nvme_rdma_dev_put(ctrl->device);
681 }
682 
683 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
684 		bool admin)
685 {
686 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
687 	struct blk_mq_tag_set *set;
688 	int ret;
689 
690 	if (admin) {
691 		set = &ctrl->admin_tag_set;
692 		memset(set, 0, sizeof(*set));
693 		set->ops = &nvme_rdma_admin_mq_ops;
694 		set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
695 		set->reserved_tags = 2; /* connect + keep-alive */
696 		set->numa_node = NUMA_NO_NODE;
697 		set->cmd_size = sizeof(struct nvme_rdma_request) +
698 			SG_CHUNK_SIZE * sizeof(struct scatterlist);
699 		set->driver_data = ctrl;
700 		set->nr_hw_queues = 1;
701 		set->timeout = ADMIN_TIMEOUT;
702 		set->flags = BLK_MQ_F_NO_SCHED;
703 	} else {
704 		set = &ctrl->tag_set;
705 		memset(set, 0, sizeof(*set));
706 		set->ops = &nvme_rdma_mq_ops;
707 		set->queue_depth = nctrl->sqsize + 1;
708 		set->reserved_tags = 1; /* fabric connect */
709 		set->numa_node = NUMA_NO_NODE;
710 		set->flags = BLK_MQ_F_SHOULD_MERGE;
711 		set->cmd_size = sizeof(struct nvme_rdma_request) +
712 			SG_CHUNK_SIZE * sizeof(struct scatterlist);
713 		set->driver_data = ctrl;
714 		set->nr_hw_queues = nctrl->queue_count - 1;
715 		set->timeout = NVME_IO_TIMEOUT;
716 	}
717 
718 	ret = blk_mq_alloc_tag_set(set);
719 	if (ret)
720 		goto out;
721 
722 	/*
723 	 * We need a reference on the device as long as the tag_set is alive,
724 	 * as the MRs in the request structures need a valid ib_device.
725 	 */
726 	ret = nvme_rdma_dev_get(ctrl->device);
727 	if (!ret) {
728 		ret = -EINVAL;
729 		goto out_free_tagset;
730 	}
731 
732 	return set;
733 
734 out_free_tagset:
735 	blk_mq_free_tag_set(set);
736 out:
737 	return ERR_PTR(ret);
738 }
739 
740 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
741 		bool remove)
742 {
743 	if (remove) {
744 		blk_cleanup_queue(ctrl->ctrl.admin_q);
745 		nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
746 	}
747 	if (ctrl->async_event_sqe.data) {
748 		nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
749 				sizeof(struct nvme_command), DMA_TO_DEVICE);
750 		ctrl->async_event_sqe.data = NULL;
751 	}
752 	nvme_rdma_free_queue(&ctrl->queues[0]);
753 }
754 
755 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
756 		bool new)
757 {
758 	int error;
759 
760 	error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
761 	if (error)
762 		return error;
763 
764 	ctrl->device = ctrl->queues[0].device;
765 
766 	ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
767 
768 	error = nvme_rdma_alloc_qe(ctrl->device->dev, &ctrl->async_event_sqe,
769 			sizeof(struct nvme_command), DMA_TO_DEVICE);
770 	if (error)
771 		goto out_free_queue;
772 
773 	if (new) {
774 		ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
775 		if (IS_ERR(ctrl->ctrl.admin_tagset)) {
776 			error = PTR_ERR(ctrl->ctrl.admin_tagset);
777 			goto out_free_async_qe;
778 		}
779 
780 		ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
781 		if (IS_ERR(ctrl->ctrl.admin_q)) {
782 			error = PTR_ERR(ctrl->ctrl.admin_q);
783 			goto out_free_tagset;
784 		}
785 	}
786 
787 	error = nvme_rdma_start_queue(ctrl, 0);
788 	if (error)
789 		goto out_cleanup_queue;
790 
791 	error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
792 			&ctrl->ctrl.cap);
793 	if (error) {
794 		dev_err(ctrl->ctrl.device,
795 			"prop_get NVME_REG_CAP failed\n");
796 		goto out_stop_queue;
797 	}
798 
799 	ctrl->ctrl.sqsize =
800 		min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
801 
802 	error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
803 	if (error)
804 		goto out_stop_queue;
805 
806 	ctrl->ctrl.max_hw_sectors =
807 		(ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
808 
809 	error = nvme_init_identify(&ctrl->ctrl);
810 	if (error)
811 		goto out_stop_queue;
812 
813 	return 0;
814 
815 out_stop_queue:
816 	nvme_rdma_stop_queue(&ctrl->queues[0]);
817 out_cleanup_queue:
818 	if (new)
819 		blk_cleanup_queue(ctrl->ctrl.admin_q);
820 out_free_tagset:
821 	if (new)
822 		nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
823 out_free_async_qe:
824 	nvme_rdma_free_qe(ctrl->device->dev, &ctrl->async_event_sqe,
825 		sizeof(struct nvme_command), DMA_TO_DEVICE);
826 out_free_queue:
827 	nvme_rdma_free_queue(&ctrl->queues[0]);
828 	return error;
829 }
830 
831 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
832 		bool remove)
833 {
834 	if (remove) {
835 		blk_cleanup_queue(ctrl->ctrl.connect_q);
836 		nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
837 	}
838 	nvme_rdma_free_io_queues(ctrl);
839 }
840 
841 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
842 {
843 	int ret;
844 
845 	ret = nvme_rdma_alloc_io_queues(ctrl);
846 	if (ret)
847 		return ret;
848 
849 	if (new) {
850 		ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
851 		if (IS_ERR(ctrl->ctrl.tagset)) {
852 			ret = PTR_ERR(ctrl->ctrl.tagset);
853 			goto out_free_io_queues;
854 		}
855 
856 		ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
857 		if (IS_ERR(ctrl->ctrl.connect_q)) {
858 			ret = PTR_ERR(ctrl->ctrl.connect_q);
859 			goto out_free_tag_set;
860 		}
861 	} else {
862 		blk_mq_update_nr_hw_queues(&ctrl->tag_set,
863 			ctrl->ctrl.queue_count - 1);
864 	}
865 
866 	ret = nvme_rdma_start_io_queues(ctrl);
867 	if (ret)
868 		goto out_cleanup_connect_q;
869 
870 	return 0;
871 
872 out_cleanup_connect_q:
873 	if (new)
874 		blk_cleanup_queue(ctrl->ctrl.connect_q);
875 out_free_tag_set:
876 	if (new)
877 		nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
878 out_free_io_queues:
879 	nvme_rdma_free_io_queues(ctrl);
880 	return ret;
881 }
882 
883 static void nvme_rdma_teardown_admin_queue(struct nvme_rdma_ctrl *ctrl,
884 		bool remove)
885 {
886 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
887 	nvme_rdma_stop_queue(&ctrl->queues[0]);
888 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set, nvme_cancel_request,
889 			&ctrl->ctrl);
890 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
891 	nvme_rdma_destroy_admin_queue(ctrl, remove);
892 }
893 
894 static void nvme_rdma_teardown_io_queues(struct nvme_rdma_ctrl *ctrl,
895 		bool remove)
896 {
897 	if (ctrl->ctrl.queue_count > 1) {
898 		nvme_stop_queues(&ctrl->ctrl);
899 		nvme_rdma_stop_io_queues(ctrl);
900 		blk_mq_tagset_busy_iter(&ctrl->tag_set, nvme_cancel_request,
901 				&ctrl->ctrl);
902 		if (remove)
903 			nvme_start_queues(&ctrl->ctrl);
904 		nvme_rdma_destroy_io_queues(ctrl, remove);
905 	}
906 }
907 
908 static void nvme_rdma_stop_ctrl(struct nvme_ctrl *nctrl)
909 {
910 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
911 
912 	cancel_work_sync(&ctrl->err_work);
913 	cancel_delayed_work_sync(&ctrl->reconnect_work);
914 }
915 
916 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
917 {
918 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
919 
920 	if (list_empty(&ctrl->list))
921 		goto free_ctrl;
922 
923 	mutex_lock(&nvme_rdma_ctrl_mutex);
924 	list_del(&ctrl->list);
925 	mutex_unlock(&nvme_rdma_ctrl_mutex);
926 
927 	nvmf_free_options(nctrl->opts);
928 free_ctrl:
929 	kfree(ctrl->queues);
930 	kfree(ctrl);
931 }
932 
933 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
934 {
935 	/* If we are resetting/deleting then do nothing */
936 	if (ctrl->ctrl.state != NVME_CTRL_CONNECTING) {
937 		WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
938 			ctrl->ctrl.state == NVME_CTRL_LIVE);
939 		return;
940 	}
941 
942 	if (nvmf_should_reconnect(&ctrl->ctrl)) {
943 		dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
944 			ctrl->ctrl.opts->reconnect_delay);
945 		queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
946 				ctrl->ctrl.opts->reconnect_delay * HZ);
947 	} else {
948 		nvme_delete_ctrl(&ctrl->ctrl);
949 	}
950 }
951 
952 static int nvme_rdma_setup_ctrl(struct nvme_rdma_ctrl *ctrl, bool new)
953 {
954 	int ret = -EINVAL;
955 	bool changed;
956 
957 	ret = nvme_rdma_configure_admin_queue(ctrl, new);
958 	if (ret)
959 		return ret;
960 
961 	if (ctrl->ctrl.icdoff) {
962 		dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
963 		goto destroy_admin;
964 	}
965 
966 	if (!(ctrl->ctrl.sgls & (1 << 2))) {
967 		dev_err(ctrl->ctrl.device,
968 			"Mandatory keyed sgls are not supported!\n");
969 		goto destroy_admin;
970 	}
971 
972 	if (ctrl->ctrl.opts->queue_size > ctrl->ctrl.sqsize + 1) {
973 		dev_warn(ctrl->ctrl.device,
974 			"queue_size %zu > ctrl sqsize %u, clamping down\n",
975 			ctrl->ctrl.opts->queue_size, ctrl->ctrl.sqsize + 1);
976 	}
977 
978 	if (ctrl->ctrl.sqsize + 1 > ctrl->ctrl.maxcmd) {
979 		dev_warn(ctrl->ctrl.device,
980 			"sqsize %u > ctrl maxcmd %u, clamping down\n",
981 			ctrl->ctrl.sqsize + 1, ctrl->ctrl.maxcmd);
982 		ctrl->ctrl.sqsize = ctrl->ctrl.maxcmd - 1;
983 	}
984 
985 	if (ctrl->ctrl.sgls & (1 << 20))
986 		ctrl->use_inline_data = true;
987 
988 	if (ctrl->ctrl.queue_count > 1) {
989 		ret = nvme_rdma_configure_io_queues(ctrl, new);
990 		if (ret)
991 			goto destroy_admin;
992 	}
993 
994 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
995 	if (!changed) {
996 		/* state change failure is ok if we're in DELETING state */
997 		WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
998 		ret = -EINVAL;
999 		goto destroy_io;
1000 	}
1001 
1002 	nvme_start_ctrl(&ctrl->ctrl);
1003 	return 0;
1004 
1005 destroy_io:
1006 	if (ctrl->ctrl.queue_count > 1)
1007 		nvme_rdma_destroy_io_queues(ctrl, new);
1008 destroy_admin:
1009 	nvme_rdma_stop_queue(&ctrl->queues[0]);
1010 	nvme_rdma_destroy_admin_queue(ctrl, new);
1011 	return ret;
1012 }
1013 
1014 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
1015 {
1016 	struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
1017 			struct nvme_rdma_ctrl, reconnect_work);
1018 
1019 	++ctrl->ctrl.nr_reconnects;
1020 
1021 	if (nvme_rdma_setup_ctrl(ctrl, false))
1022 		goto requeue;
1023 
1024 	dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
1025 			ctrl->ctrl.nr_reconnects);
1026 
1027 	ctrl->ctrl.nr_reconnects = 0;
1028 
1029 	return;
1030 
1031 requeue:
1032 	dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
1033 			ctrl->ctrl.nr_reconnects);
1034 	nvme_rdma_reconnect_or_remove(ctrl);
1035 }
1036 
1037 static void nvme_rdma_error_recovery_work(struct work_struct *work)
1038 {
1039 	struct nvme_rdma_ctrl *ctrl = container_of(work,
1040 			struct nvme_rdma_ctrl, err_work);
1041 
1042 	nvme_stop_keep_alive(&ctrl->ctrl);
1043 	nvme_rdma_teardown_io_queues(ctrl, false);
1044 	nvme_start_queues(&ctrl->ctrl);
1045 	nvme_rdma_teardown_admin_queue(ctrl, false);
1046 
1047 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1048 		/* state change failure is ok if we're in DELETING state */
1049 		WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1050 		return;
1051 	}
1052 
1053 	nvme_rdma_reconnect_or_remove(ctrl);
1054 }
1055 
1056 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1057 {
1058 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
1059 		return;
1060 
1061 	queue_work(nvme_wq, &ctrl->err_work);
1062 }
1063 
1064 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1065 		const char *op)
1066 {
1067 	struct nvme_rdma_queue *queue = cq->cq_context;
1068 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1069 
1070 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1071 		dev_info(ctrl->ctrl.device,
1072 			     "%s for CQE 0x%p failed with status %s (%d)\n",
1073 			     op, wc->wr_cqe,
1074 			     ib_wc_status_msg(wc->status), wc->status);
1075 	nvme_rdma_error_recovery(ctrl);
1076 }
1077 
1078 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1079 {
1080 	if (unlikely(wc->status != IB_WC_SUCCESS))
1081 		nvme_rdma_wr_error(cq, wc, "MEMREG");
1082 }
1083 
1084 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1085 {
1086 	struct nvme_rdma_request *req =
1087 		container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1088 	struct request *rq = blk_mq_rq_from_pdu(req);
1089 
1090 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1091 		nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1092 		return;
1093 	}
1094 
1095 	if (refcount_dec_and_test(&req->ref))
1096 		nvme_end_request(rq, req->status, req->result);
1097 
1098 }
1099 
1100 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1101 		struct nvme_rdma_request *req)
1102 {
1103 	struct ib_send_wr wr = {
1104 		.opcode		    = IB_WR_LOCAL_INV,
1105 		.next		    = NULL,
1106 		.num_sge	    = 0,
1107 		.send_flags	    = IB_SEND_SIGNALED,
1108 		.ex.invalidate_rkey = req->mr->rkey,
1109 	};
1110 
1111 	req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1112 	wr.wr_cqe = &req->reg_cqe;
1113 
1114 	return ib_post_send(queue->qp, &wr, NULL);
1115 }
1116 
1117 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1118 		struct request *rq)
1119 {
1120 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1121 	struct nvme_rdma_device *dev = queue->device;
1122 	struct ib_device *ibdev = dev->dev;
1123 
1124 	if (!blk_rq_payload_bytes(rq))
1125 		return;
1126 
1127 	if (req->mr) {
1128 		ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1129 		req->mr = NULL;
1130 	}
1131 
1132 	ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1133 			req->nents, rq_data_dir(rq) ==
1134 				    WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1135 
1136 	nvme_cleanup_cmd(rq);
1137 	sg_free_table_chained(&req->sg_table, true);
1138 }
1139 
1140 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1141 {
1142 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1143 
1144 	sg->addr = 0;
1145 	put_unaligned_le24(0, sg->length);
1146 	put_unaligned_le32(0, sg->key);
1147 	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1148 	return 0;
1149 }
1150 
1151 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1152 		struct nvme_rdma_request *req, struct nvme_command *c,
1153 		int count)
1154 {
1155 	struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1156 	struct scatterlist *sgl = req->sg_table.sgl;
1157 	struct ib_sge *sge = &req->sge[1];
1158 	u32 len = 0;
1159 	int i;
1160 
1161 	for (i = 0; i < count; i++, sgl++, sge++) {
1162 		sge->addr = sg_dma_address(sgl);
1163 		sge->length = sg_dma_len(sgl);
1164 		sge->lkey = queue->device->pd->local_dma_lkey;
1165 		len += sge->length;
1166 	}
1167 
1168 	sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1169 	sg->length = cpu_to_le32(len);
1170 	sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1171 
1172 	req->num_sge += count;
1173 	return 0;
1174 }
1175 
1176 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1177 		struct nvme_rdma_request *req, struct nvme_command *c)
1178 {
1179 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1180 
1181 	sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1182 	put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1183 	put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1184 	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1185 	return 0;
1186 }
1187 
1188 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1189 		struct nvme_rdma_request *req, struct nvme_command *c,
1190 		int count)
1191 {
1192 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1193 	int nr;
1194 
1195 	req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1196 	if (WARN_ON_ONCE(!req->mr))
1197 		return -EAGAIN;
1198 
1199 	/*
1200 	 * Align the MR to a 4K page size to match the ctrl page size and
1201 	 * the block virtual boundary.
1202 	 */
1203 	nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1204 	if (unlikely(nr < count)) {
1205 		ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1206 		req->mr = NULL;
1207 		if (nr < 0)
1208 			return nr;
1209 		return -EINVAL;
1210 	}
1211 
1212 	ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1213 
1214 	req->reg_cqe.done = nvme_rdma_memreg_done;
1215 	memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1216 	req->reg_wr.wr.opcode = IB_WR_REG_MR;
1217 	req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1218 	req->reg_wr.wr.num_sge = 0;
1219 	req->reg_wr.mr = req->mr;
1220 	req->reg_wr.key = req->mr->rkey;
1221 	req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1222 			     IB_ACCESS_REMOTE_READ |
1223 			     IB_ACCESS_REMOTE_WRITE;
1224 
1225 	sg->addr = cpu_to_le64(req->mr->iova);
1226 	put_unaligned_le24(req->mr->length, sg->length);
1227 	put_unaligned_le32(req->mr->rkey, sg->key);
1228 	sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1229 			NVME_SGL_FMT_INVALIDATE;
1230 
1231 	return 0;
1232 }
1233 
1234 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1235 		struct request *rq, struct nvme_command *c)
1236 {
1237 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1238 	struct nvme_rdma_device *dev = queue->device;
1239 	struct ib_device *ibdev = dev->dev;
1240 	int count, ret;
1241 
1242 	req->num_sge = 1;
1243 	refcount_set(&req->ref, 2); /* send and recv completions */
1244 
1245 	c->common.flags |= NVME_CMD_SGL_METABUF;
1246 
1247 	if (!blk_rq_payload_bytes(rq))
1248 		return nvme_rdma_set_sg_null(c);
1249 
1250 	req->sg_table.sgl = req->first_sgl;
1251 	ret = sg_alloc_table_chained(&req->sg_table,
1252 			blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1253 	if (ret)
1254 		return -ENOMEM;
1255 
1256 	req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1257 
1258 	count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1259 		    rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1260 	if (unlikely(count <= 0)) {
1261 		ret = -EIO;
1262 		goto out_free_table;
1263 	}
1264 
1265 	if (count <= dev->num_inline_segments) {
1266 		if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1267 		    queue->ctrl->use_inline_data &&
1268 		    blk_rq_payload_bytes(rq) <=
1269 				nvme_rdma_inline_data_size(queue)) {
1270 			ret = nvme_rdma_map_sg_inline(queue, req, c, count);
1271 			goto out;
1272 		}
1273 
1274 		if (count == 1 && dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY) {
1275 			ret = nvme_rdma_map_sg_single(queue, req, c);
1276 			goto out;
1277 		}
1278 	}
1279 
1280 	ret = nvme_rdma_map_sg_fr(queue, req, c, count);
1281 out:
1282 	if (unlikely(ret))
1283 		goto out_unmap_sg;
1284 
1285 	return 0;
1286 
1287 out_unmap_sg:
1288 	ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1289 			req->nents, rq_data_dir(rq) ==
1290 			WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1291 out_free_table:
1292 	sg_free_table_chained(&req->sg_table, true);
1293 	return ret;
1294 }
1295 
1296 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1297 {
1298 	struct nvme_rdma_qe *qe =
1299 		container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1300 	struct nvme_rdma_request *req =
1301 		container_of(qe, struct nvme_rdma_request, sqe);
1302 	struct request *rq = blk_mq_rq_from_pdu(req);
1303 
1304 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1305 		nvme_rdma_wr_error(cq, wc, "SEND");
1306 		return;
1307 	}
1308 
1309 	if (refcount_dec_and_test(&req->ref))
1310 		nvme_end_request(rq, req->status, req->result);
1311 }
1312 
1313 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1314 		struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1315 		struct ib_send_wr *first)
1316 {
1317 	struct ib_send_wr wr;
1318 	int ret;
1319 
1320 	sge->addr   = qe->dma;
1321 	sge->length = sizeof(struct nvme_command),
1322 	sge->lkey   = queue->device->pd->local_dma_lkey;
1323 
1324 	wr.next       = NULL;
1325 	wr.wr_cqe     = &qe->cqe;
1326 	wr.sg_list    = sge;
1327 	wr.num_sge    = num_sge;
1328 	wr.opcode     = IB_WR_SEND;
1329 	wr.send_flags = IB_SEND_SIGNALED;
1330 
1331 	if (first)
1332 		first->next = &wr;
1333 	else
1334 		first = &wr;
1335 
1336 	ret = ib_post_send(queue->qp, first, NULL);
1337 	if (unlikely(ret)) {
1338 		dev_err(queue->ctrl->ctrl.device,
1339 			     "%s failed with error code %d\n", __func__, ret);
1340 	}
1341 	return ret;
1342 }
1343 
1344 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1345 		struct nvme_rdma_qe *qe)
1346 {
1347 	struct ib_recv_wr wr;
1348 	struct ib_sge list;
1349 	int ret;
1350 
1351 	list.addr   = qe->dma;
1352 	list.length = sizeof(struct nvme_completion);
1353 	list.lkey   = queue->device->pd->local_dma_lkey;
1354 
1355 	qe->cqe.done = nvme_rdma_recv_done;
1356 
1357 	wr.next     = NULL;
1358 	wr.wr_cqe   = &qe->cqe;
1359 	wr.sg_list  = &list;
1360 	wr.num_sge  = 1;
1361 
1362 	ret = ib_post_recv(queue->qp, &wr, NULL);
1363 	if (unlikely(ret)) {
1364 		dev_err(queue->ctrl->ctrl.device,
1365 			"%s failed with error code %d\n", __func__, ret);
1366 	}
1367 	return ret;
1368 }
1369 
1370 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1371 {
1372 	u32 queue_idx = nvme_rdma_queue_idx(queue);
1373 
1374 	if (queue_idx == 0)
1375 		return queue->ctrl->admin_tag_set.tags[queue_idx];
1376 	return queue->ctrl->tag_set.tags[queue_idx - 1];
1377 }
1378 
1379 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1380 {
1381 	if (unlikely(wc->status != IB_WC_SUCCESS))
1382 		nvme_rdma_wr_error(cq, wc, "ASYNC");
1383 }
1384 
1385 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1386 {
1387 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1388 	struct nvme_rdma_queue *queue = &ctrl->queues[0];
1389 	struct ib_device *dev = queue->device->dev;
1390 	struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1391 	struct nvme_command *cmd = sqe->data;
1392 	struct ib_sge sge;
1393 	int ret;
1394 
1395 	ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1396 
1397 	memset(cmd, 0, sizeof(*cmd));
1398 	cmd->common.opcode = nvme_admin_async_event;
1399 	cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1400 	cmd->common.flags |= NVME_CMD_SGL_METABUF;
1401 	nvme_rdma_set_sg_null(cmd);
1402 
1403 	sqe->cqe.done = nvme_rdma_async_done;
1404 
1405 	ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1406 			DMA_TO_DEVICE);
1407 
1408 	ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1409 	WARN_ON_ONCE(ret);
1410 }
1411 
1412 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1413 		struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1414 {
1415 	struct request *rq;
1416 	struct nvme_rdma_request *req;
1417 	int ret = 0;
1418 
1419 	rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1420 	if (!rq) {
1421 		dev_err(queue->ctrl->ctrl.device,
1422 			"tag 0x%x on QP %#x not found\n",
1423 			cqe->command_id, queue->qp->qp_num);
1424 		nvme_rdma_error_recovery(queue->ctrl);
1425 		return ret;
1426 	}
1427 	req = blk_mq_rq_to_pdu(rq);
1428 
1429 	req->status = cqe->status;
1430 	req->result = cqe->result;
1431 
1432 	if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1433 		if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1434 			dev_err(queue->ctrl->ctrl.device,
1435 				"Bogus remote invalidation for rkey %#x\n",
1436 				req->mr->rkey);
1437 			nvme_rdma_error_recovery(queue->ctrl);
1438 		}
1439 	} else if (req->mr) {
1440 		ret = nvme_rdma_inv_rkey(queue, req);
1441 		if (unlikely(ret < 0)) {
1442 			dev_err(queue->ctrl->ctrl.device,
1443 				"Queueing INV WR for rkey %#x failed (%d)\n",
1444 				req->mr->rkey, ret);
1445 			nvme_rdma_error_recovery(queue->ctrl);
1446 		}
1447 		/* the local invalidation completion will end the request */
1448 		return 0;
1449 	}
1450 
1451 	if (refcount_dec_and_test(&req->ref)) {
1452 		if (rq->tag == tag)
1453 			ret = 1;
1454 		nvme_end_request(rq, req->status, req->result);
1455 	}
1456 
1457 	return ret;
1458 }
1459 
1460 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1461 {
1462 	struct nvme_rdma_qe *qe =
1463 		container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1464 	struct nvme_rdma_queue *queue = cq->cq_context;
1465 	struct ib_device *ibdev = queue->device->dev;
1466 	struct nvme_completion *cqe = qe->data;
1467 	const size_t len = sizeof(struct nvme_completion);
1468 	int ret = 0;
1469 
1470 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1471 		nvme_rdma_wr_error(cq, wc, "RECV");
1472 		return 0;
1473 	}
1474 
1475 	ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1476 	/*
1477 	 * AEN requests are special as they don't time out and can
1478 	 * survive any kind of queue freeze and often don't respond to
1479 	 * aborts.  We don't even bother to allocate a struct request
1480 	 * for them but rather special case them here.
1481 	 */
1482 	if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1483 			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1484 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1485 				&cqe->result);
1486 	else
1487 		ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1488 	ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1489 
1490 	nvme_rdma_post_recv(queue, qe);
1491 	return ret;
1492 }
1493 
1494 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1495 {
1496 	__nvme_rdma_recv_done(cq, wc, -1);
1497 }
1498 
1499 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1500 {
1501 	int ret, i;
1502 
1503 	for (i = 0; i < queue->queue_size; i++) {
1504 		ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1505 		if (ret)
1506 			goto out_destroy_queue_ib;
1507 	}
1508 
1509 	return 0;
1510 
1511 out_destroy_queue_ib:
1512 	nvme_rdma_destroy_queue_ib(queue);
1513 	return ret;
1514 }
1515 
1516 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1517 		struct rdma_cm_event *ev)
1518 {
1519 	struct rdma_cm_id *cm_id = queue->cm_id;
1520 	int status = ev->status;
1521 	const char *rej_msg;
1522 	const struct nvme_rdma_cm_rej *rej_data;
1523 	u8 rej_data_len;
1524 
1525 	rej_msg = rdma_reject_msg(cm_id, status);
1526 	rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1527 
1528 	if (rej_data && rej_data_len >= sizeof(u16)) {
1529 		u16 sts = le16_to_cpu(rej_data->sts);
1530 
1531 		dev_err(queue->ctrl->ctrl.device,
1532 		      "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1533 		      status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1534 	} else {
1535 		dev_err(queue->ctrl->ctrl.device,
1536 			"Connect rejected: status %d (%s).\n", status, rej_msg);
1537 	}
1538 
1539 	return -ECONNRESET;
1540 }
1541 
1542 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1543 {
1544 	int ret;
1545 
1546 	ret = nvme_rdma_create_queue_ib(queue);
1547 	if (ret)
1548 		return ret;
1549 
1550 	ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1551 	if (ret) {
1552 		dev_err(queue->ctrl->ctrl.device,
1553 			"rdma_resolve_route failed (%d).\n",
1554 			queue->cm_error);
1555 		goto out_destroy_queue;
1556 	}
1557 
1558 	return 0;
1559 
1560 out_destroy_queue:
1561 	nvme_rdma_destroy_queue_ib(queue);
1562 	return ret;
1563 }
1564 
1565 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1566 {
1567 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1568 	struct rdma_conn_param param = { };
1569 	struct nvme_rdma_cm_req priv = { };
1570 	int ret;
1571 
1572 	param.qp_num = queue->qp->qp_num;
1573 	param.flow_control = 1;
1574 
1575 	param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1576 	/* maximum retry count */
1577 	param.retry_count = 7;
1578 	param.rnr_retry_count = 7;
1579 	param.private_data = &priv;
1580 	param.private_data_len = sizeof(priv);
1581 
1582 	priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1583 	priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1584 	/*
1585 	 * set the admin queue depth to the minimum size
1586 	 * specified by the Fabrics standard.
1587 	 */
1588 	if (priv.qid == 0) {
1589 		priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1590 		priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1591 	} else {
1592 		/*
1593 		 * current interpretation of the fabrics spec
1594 		 * is at minimum you make hrqsize sqsize+1, or a
1595 		 * 1's based representation of sqsize.
1596 		 */
1597 		priv.hrqsize = cpu_to_le16(queue->queue_size);
1598 		priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1599 	}
1600 
1601 	ret = rdma_connect(queue->cm_id, &param);
1602 	if (ret) {
1603 		dev_err(ctrl->ctrl.device,
1604 			"rdma_connect failed (%d).\n", ret);
1605 		goto out_destroy_queue_ib;
1606 	}
1607 
1608 	return 0;
1609 
1610 out_destroy_queue_ib:
1611 	nvme_rdma_destroy_queue_ib(queue);
1612 	return ret;
1613 }
1614 
1615 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1616 		struct rdma_cm_event *ev)
1617 {
1618 	struct nvme_rdma_queue *queue = cm_id->context;
1619 	int cm_error = 0;
1620 
1621 	dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1622 		rdma_event_msg(ev->event), ev->event,
1623 		ev->status, cm_id);
1624 
1625 	switch (ev->event) {
1626 	case RDMA_CM_EVENT_ADDR_RESOLVED:
1627 		cm_error = nvme_rdma_addr_resolved(queue);
1628 		break;
1629 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
1630 		cm_error = nvme_rdma_route_resolved(queue);
1631 		break;
1632 	case RDMA_CM_EVENT_ESTABLISHED:
1633 		queue->cm_error = nvme_rdma_conn_established(queue);
1634 		/* complete cm_done regardless of success/failure */
1635 		complete(&queue->cm_done);
1636 		return 0;
1637 	case RDMA_CM_EVENT_REJECTED:
1638 		nvme_rdma_destroy_queue_ib(queue);
1639 		cm_error = nvme_rdma_conn_rejected(queue, ev);
1640 		break;
1641 	case RDMA_CM_EVENT_ROUTE_ERROR:
1642 	case RDMA_CM_EVENT_CONNECT_ERROR:
1643 	case RDMA_CM_EVENT_UNREACHABLE:
1644 		nvme_rdma_destroy_queue_ib(queue);
1645 		/* fall through */
1646 	case RDMA_CM_EVENT_ADDR_ERROR:
1647 		dev_dbg(queue->ctrl->ctrl.device,
1648 			"CM error event %d\n", ev->event);
1649 		cm_error = -ECONNRESET;
1650 		break;
1651 	case RDMA_CM_EVENT_DISCONNECTED:
1652 	case RDMA_CM_EVENT_ADDR_CHANGE:
1653 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1654 		dev_dbg(queue->ctrl->ctrl.device,
1655 			"disconnect received - connection closed\n");
1656 		nvme_rdma_error_recovery(queue->ctrl);
1657 		break;
1658 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1659 		/* device removal is handled via the ib_client API */
1660 		break;
1661 	default:
1662 		dev_err(queue->ctrl->ctrl.device,
1663 			"Unexpected RDMA CM event (%d)\n", ev->event);
1664 		nvme_rdma_error_recovery(queue->ctrl);
1665 		break;
1666 	}
1667 
1668 	if (cm_error) {
1669 		queue->cm_error = cm_error;
1670 		complete(&queue->cm_done);
1671 	}
1672 
1673 	return 0;
1674 }
1675 
1676 static enum blk_eh_timer_return
1677 nvme_rdma_timeout(struct request *rq, bool reserved)
1678 {
1679 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1680 
1681 	dev_warn(req->queue->ctrl->ctrl.device,
1682 		 "I/O %d QID %d timeout, reset controller\n",
1683 		 rq->tag, nvme_rdma_queue_idx(req->queue));
1684 
1685 	/* queue error recovery */
1686 	nvme_rdma_error_recovery(req->queue->ctrl);
1687 
1688 	/* fail with DNR on cmd timeout */
1689 	nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1690 
1691 	return BLK_EH_DONE;
1692 }
1693 
1694 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1695 		const struct blk_mq_queue_data *bd)
1696 {
1697 	struct nvme_ns *ns = hctx->queue->queuedata;
1698 	struct nvme_rdma_queue *queue = hctx->driver_data;
1699 	struct request *rq = bd->rq;
1700 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1701 	struct nvme_rdma_qe *sqe = &req->sqe;
1702 	struct nvme_command *c = sqe->data;
1703 	struct ib_device *dev;
1704 	bool queue_ready = test_bit(NVME_RDMA_Q_LIVE, &queue->flags);
1705 	blk_status_t ret;
1706 	int err;
1707 
1708 	WARN_ON_ONCE(rq->tag < 0);
1709 
1710 	if (!nvmf_check_ready(&queue->ctrl->ctrl, rq, queue_ready))
1711 		return nvmf_fail_nonready_command(&queue->ctrl->ctrl, rq);
1712 
1713 	dev = queue->device->dev;
1714 	ib_dma_sync_single_for_cpu(dev, sqe->dma,
1715 			sizeof(struct nvme_command), DMA_TO_DEVICE);
1716 
1717 	ret = nvme_setup_cmd(ns, rq, c);
1718 	if (ret)
1719 		return ret;
1720 
1721 	blk_mq_start_request(rq);
1722 
1723 	err = nvme_rdma_map_data(queue, rq, c);
1724 	if (unlikely(err < 0)) {
1725 		dev_err(queue->ctrl->ctrl.device,
1726 			     "Failed to map data (%d)\n", err);
1727 		nvme_cleanup_cmd(rq);
1728 		goto err;
1729 	}
1730 
1731 	sqe->cqe.done = nvme_rdma_send_done;
1732 
1733 	ib_dma_sync_single_for_device(dev, sqe->dma,
1734 			sizeof(struct nvme_command), DMA_TO_DEVICE);
1735 
1736 	err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1737 			req->mr ? &req->reg_wr.wr : NULL);
1738 	if (unlikely(err)) {
1739 		nvme_rdma_unmap_data(queue, rq);
1740 		goto err;
1741 	}
1742 
1743 	return BLK_STS_OK;
1744 err:
1745 	if (err == -ENOMEM || err == -EAGAIN)
1746 		return BLK_STS_RESOURCE;
1747 	return BLK_STS_IOERR;
1748 }
1749 
1750 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1751 {
1752 	struct nvme_rdma_queue *queue = hctx->driver_data;
1753 	struct ib_cq *cq = queue->ib_cq;
1754 	struct ib_wc wc;
1755 	int found = 0;
1756 
1757 	while (ib_poll_cq(cq, 1, &wc) > 0) {
1758 		struct ib_cqe *cqe = wc.wr_cqe;
1759 
1760 		if (cqe) {
1761 			if (cqe->done == nvme_rdma_recv_done)
1762 				found |= __nvme_rdma_recv_done(cq, &wc, tag);
1763 			else
1764 				cqe->done(cq, &wc);
1765 		}
1766 	}
1767 
1768 	return found;
1769 }
1770 
1771 static void nvme_rdma_complete_rq(struct request *rq)
1772 {
1773 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1774 
1775 	nvme_rdma_unmap_data(req->queue, rq);
1776 	nvme_complete_rq(rq);
1777 }
1778 
1779 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1780 {
1781 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
1782 
1783 	return blk_mq_rdma_map_queues(set, ctrl->device->dev, 0);
1784 }
1785 
1786 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1787 	.queue_rq	= nvme_rdma_queue_rq,
1788 	.complete	= nvme_rdma_complete_rq,
1789 	.init_request	= nvme_rdma_init_request,
1790 	.exit_request	= nvme_rdma_exit_request,
1791 	.init_hctx	= nvme_rdma_init_hctx,
1792 	.poll		= nvme_rdma_poll,
1793 	.timeout	= nvme_rdma_timeout,
1794 	.map_queues	= nvme_rdma_map_queues,
1795 };
1796 
1797 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1798 	.queue_rq	= nvme_rdma_queue_rq,
1799 	.complete	= nvme_rdma_complete_rq,
1800 	.init_request	= nvme_rdma_init_request,
1801 	.exit_request	= nvme_rdma_exit_request,
1802 	.init_hctx	= nvme_rdma_init_admin_hctx,
1803 	.timeout	= nvme_rdma_timeout,
1804 };
1805 
1806 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1807 {
1808 	nvme_rdma_teardown_io_queues(ctrl, shutdown);
1809 	if (shutdown)
1810 		nvme_shutdown_ctrl(&ctrl->ctrl);
1811 	else
1812 		nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1813 	nvme_rdma_teardown_admin_queue(ctrl, shutdown);
1814 }
1815 
1816 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1817 {
1818 	nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1819 }
1820 
1821 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1822 {
1823 	struct nvme_rdma_ctrl *ctrl =
1824 		container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1825 
1826 	nvme_stop_ctrl(&ctrl->ctrl);
1827 	nvme_rdma_shutdown_ctrl(ctrl, false);
1828 
1829 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING)) {
1830 		/* state change failure should never happen */
1831 		WARN_ON_ONCE(1);
1832 		return;
1833 	}
1834 
1835 	if (nvme_rdma_setup_ctrl(ctrl, false))
1836 		goto out_fail;
1837 
1838 	return;
1839 
1840 out_fail:
1841 	++ctrl->ctrl.nr_reconnects;
1842 	nvme_rdma_reconnect_or_remove(ctrl);
1843 }
1844 
1845 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1846 	.name			= "rdma",
1847 	.module			= THIS_MODULE,
1848 	.flags			= NVME_F_FABRICS,
1849 	.reg_read32		= nvmf_reg_read32,
1850 	.reg_read64		= nvmf_reg_read64,
1851 	.reg_write32		= nvmf_reg_write32,
1852 	.free_ctrl		= nvme_rdma_free_ctrl,
1853 	.submit_async_event	= nvme_rdma_submit_async_event,
1854 	.delete_ctrl		= nvme_rdma_delete_ctrl,
1855 	.get_address		= nvmf_get_address,
1856 	.stop_ctrl		= nvme_rdma_stop_ctrl,
1857 };
1858 
1859 /*
1860  * Fails a connection request if it matches an existing controller
1861  * (association) with the same tuple:
1862  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1863  *
1864  * if local address is not specified in the request, it will match an
1865  * existing controller with all the other parameters the same and no
1866  * local port address specified as well.
1867  *
1868  * The ports don't need to be compared as they are intrinsically
1869  * already matched by the port pointers supplied.
1870  */
1871 static bool
1872 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1873 {
1874 	struct nvme_rdma_ctrl *ctrl;
1875 	bool found = false;
1876 
1877 	mutex_lock(&nvme_rdma_ctrl_mutex);
1878 	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1879 		found = nvmf_ip_options_match(&ctrl->ctrl, opts);
1880 		if (found)
1881 			break;
1882 	}
1883 	mutex_unlock(&nvme_rdma_ctrl_mutex);
1884 
1885 	return found;
1886 }
1887 
1888 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1889 		struct nvmf_ctrl_options *opts)
1890 {
1891 	struct nvme_rdma_ctrl *ctrl;
1892 	int ret;
1893 	bool changed;
1894 
1895 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1896 	if (!ctrl)
1897 		return ERR_PTR(-ENOMEM);
1898 	ctrl->ctrl.opts = opts;
1899 	INIT_LIST_HEAD(&ctrl->list);
1900 
1901 	if (!(opts->mask & NVMF_OPT_TRSVCID)) {
1902 		opts->trsvcid =
1903 			kstrdup(__stringify(NVME_RDMA_IP_PORT), GFP_KERNEL);
1904 		if (!opts->trsvcid) {
1905 			ret = -ENOMEM;
1906 			goto out_free_ctrl;
1907 		}
1908 		opts->mask |= NVMF_OPT_TRSVCID;
1909 	}
1910 
1911 	ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1912 			opts->traddr, opts->trsvcid, &ctrl->addr);
1913 	if (ret) {
1914 		pr_err("malformed address passed: %s:%s\n",
1915 			opts->traddr, opts->trsvcid);
1916 		goto out_free_ctrl;
1917 	}
1918 
1919 	if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1920 		ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1921 			opts->host_traddr, NULL, &ctrl->src_addr);
1922 		if (ret) {
1923 			pr_err("malformed src address passed: %s\n",
1924 			       opts->host_traddr);
1925 			goto out_free_ctrl;
1926 		}
1927 	}
1928 
1929 	if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1930 		ret = -EALREADY;
1931 		goto out_free_ctrl;
1932 	}
1933 
1934 	INIT_DELAYED_WORK(&ctrl->reconnect_work,
1935 			nvme_rdma_reconnect_ctrl_work);
1936 	INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1937 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1938 
1939 	ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1940 	ctrl->ctrl.sqsize = opts->queue_size - 1;
1941 	ctrl->ctrl.kato = opts->kato;
1942 
1943 	ret = -ENOMEM;
1944 	ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1945 				GFP_KERNEL);
1946 	if (!ctrl->queues)
1947 		goto out_free_ctrl;
1948 
1949 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1950 				0 /* no quirks, we're perfect! */);
1951 	if (ret)
1952 		goto out_kfree_queues;
1953 
1954 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_CONNECTING);
1955 	WARN_ON_ONCE(!changed);
1956 
1957 	ret = nvme_rdma_setup_ctrl(ctrl, true);
1958 	if (ret)
1959 		goto out_uninit_ctrl;
1960 
1961 	dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1962 		ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1963 
1964 	nvme_get_ctrl(&ctrl->ctrl);
1965 
1966 	mutex_lock(&nvme_rdma_ctrl_mutex);
1967 	list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1968 	mutex_unlock(&nvme_rdma_ctrl_mutex);
1969 
1970 	return &ctrl->ctrl;
1971 
1972 out_uninit_ctrl:
1973 	nvme_uninit_ctrl(&ctrl->ctrl);
1974 	nvme_put_ctrl(&ctrl->ctrl);
1975 	if (ret > 0)
1976 		ret = -EIO;
1977 	return ERR_PTR(ret);
1978 out_kfree_queues:
1979 	kfree(ctrl->queues);
1980 out_free_ctrl:
1981 	kfree(ctrl);
1982 	return ERR_PTR(ret);
1983 }
1984 
1985 static struct nvmf_transport_ops nvme_rdma_transport = {
1986 	.name		= "rdma",
1987 	.module		= THIS_MODULE,
1988 	.required_opts	= NVMF_OPT_TRADDR,
1989 	.allowed_opts	= NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
1990 			  NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
1991 	.create_ctrl	= nvme_rdma_create_ctrl,
1992 };
1993 
1994 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1995 {
1996 	struct nvme_rdma_ctrl *ctrl;
1997 	struct nvme_rdma_device *ndev;
1998 	bool found = false;
1999 
2000 	mutex_lock(&device_list_mutex);
2001 	list_for_each_entry(ndev, &device_list, entry) {
2002 		if (ndev->dev == ib_device) {
2003 			found = true;
2004 			break;
2005 		}
2006 	}
2007 	mutex_unlock(&device_list_mutex);
2008 
2009 	if (!found)
2010 		return;
2011 
2012 	/* Delete all controllers using this device */
2013 	mutex_lock(&nvme_rdma_ctrl_mutex);
2014 	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2015 		if (ctrl->device->dev != ib_device)
2016 			continue;
2017 		nvme_delete_ctrl(&ctrl->ctrl);
2018 	}
2019 	mutex_unlock(&nvme_rdma_ctrl_mutex);
2020 
2021 	flush_workqueue(nvme_delete_wq);
2022 }
2023 
2024 static struct ib_client nvme_rdma_ib_client = {
2025 	.name   = "nvme_rdma",
2026 	.remove = nvme_rdma_remove_one
2027 };
2028 
2029 static int __init nvme_rdma_init_module(void)
2030 {
2031 	int ret;
2032 
2033 	ret = ib_register_client(&nvme_rdma_ib_client);
2034 	if (ret)
2035 		return ret;
2036 
2037 	ret = nvmf_register_transport(&nvme_rdma_transport);
2038 	if (ret)
2039 		goto err_unreg_client;
2040 
2041 	return 0;
2042 
2043 err_unreg_client:
2044 	ib_unregister_client(&nvme_rdma_ib_client);
2045 	return ret;
2046 }
2047 
2048 static void __exit nvme_rdma_cleanup_module(void)
2049 {
2050 	nvmf_unregister_transport(&nvme_rdma_transport);
2051 	ib_unregister_client(&nvme_rdma_ib_client);
2052 }
2053 
2054 module_init(nvme_rdma_init_module);
2055 module_exit(nvme_rdma_cleanup_module);
2056 
2057 MODULE_LICENSE("GPL v2");
2058