xref: /openbmc/linux/drivers/nvme/host/rdma.c (revision 2eb3ed33e55d003d721d4d1a5e72fe323c12b4c0)
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 <linux/err.h>
19 #include <linux/string.h>
20 #include <linux/atomic.h>
21 #include <linux/blk-mq.h>
22 #include <linux/blk-mq-rdma.h>
23 #include <linux/types.h>
24 #include <linux/list.h>
25 #include <linux/mutex.h>
26 #include <linux/scatterlist.h>
27 #include <linux/nvme.h>
28 #include <asm/unaligned.h>
29 
30 #include <rdma/ib_verbs.h>
31 #include <rdma/rdma_cm.h>
32 #include <linux/nvme-rdma.h>
33 
34 #include "nvme.h"
35 #include "fabrics.h"
36 
37 
38 #define NVME_RDMA_CONNECT_TIMEOUT_MS	3000		/* 3 second */
39 
40 #define NVME_RDMA_MAX_SEGMENTS		256
41 
42 #define NVME_RDMA_MAX_INLINE_SEGMENTS	1
43 
44 /*
45  * We handle AEN commands ourselves and don't even let the
46  * block layer know about them.
47  */
48 #define NVME_RDMA_NR_AEN_COMMANDS      1
49 #define NVME_RDMA_AQ_BLKMQ_DEPTH       \
50 	(NVME_AQ_DEPTH - NVME_RDMA_NR_AEN_COMMANDS)
51 
52 struct nvme_rdma_device {
53 	struct ib_device       *dev;
54 	struct ib_pd	       *pd;
55 	struct kref		ref;
56 	struct list_head	entry;
57 };
58 
59 struct nvme_rdma_qe {
60 	struct ib_cqe		cqe;
61 	void			*data;
62 	u64			dma;
63 };
64 
65 struct nvme_rdma_queue;
66 struct nvme_rdma_request {
67 	struct nvme_request	req;
68 	struct ib_mr		*mr;
69 	struct nvme_rdma_qe	sqe;
70 	struct ib_sge		sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
71 	u32			num_sge;
72 	int			nents;
73 	bool			inline_data;
74 	struct ib_reg_wr	reg_wr;
75 	struct ib_cqe		reg_cqe;
76 	struct nvme_rdma_queue  *queue;
77 	struct sg_table		sg_table;
78 	struct scatterlist	first_sgl[];
79 };
80 
81 enum nvme_rdma_queue_flags {
82 	NVME_RDMA_Q_LIVE		= 0,
83 	NVME_RDMA_Q_DELETING		= 1,
84 };
85 
86 struct nvme_rdma_queue {
87 	struct nvme_rdma_qe	*rsp_ring;
88 	atomic_t		sig_count;
89 	int			queue_size;
90 	size_t			cmnd_capsule_len;
91 	struct nvme_rdma_ctrl	*ctrl;
92 	struct nvme_rdma_device	*device;
93 	struct ib_cq		*ib_cq;
94 	struct ib_qp		*qp;
95 
96 	unsigned long		flags;
97 	struct rdma_cm_id	*cm_id;
98 	int			cm_error;
99 	struct completion	cm_done;
100 };
101 
102 struct nvme_rdma_ctrl {
103 	/* read only in the hot path */
104 	struct nvme_rdma_queue	*queues;
105 
106 	/* other member variables */
107 	struct blk_mq_tag_set	tag_set;
108 	struct work_struct	delete_work;
109 	struct work_struct	err_work;
110 
111 	struct nvme_rdma_qe	async_event_sqe;
112 
113 	struct delayed_work	reconnect_work;
114 
115 	struct list_head	list;
116 
117 	struct blk_mq_tag_set	admin_tag_set;
118 	struct nvme_rdma_device	*device;
119 
120 	u32			max_fr_pages;
121 
122 	struct sockaddr_storage addr;
123 	struct sockaddr_storage src_addr;
124 
125 	struct nvme_ctrl	ctrl;
126 };
127 
128 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
129 {
130 	return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
131 }
132 
133 static LIST_HEAD(device_list);
134 static DEFINE_MUTEX(device_list_mutex);
135 
136 static LIST_HEAD(nvme_rdma_ctrl_list);
137 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
138 
139 /*
140  * Disabling this option makes small I/O goes faster, but is fundamentally
141  * unsafe.  With it turned off we will have to register a global rkey that
142  * allows read and write access to all physical memory.
143  */
144 static bool register_always = true;
145 module_param(register_always, bool, 0444);
146 MODULE_PARM_DESC(register_always,
147 	 "Use memory registration even for contiguous memory regions");
148 
149 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
150 		struct rdma_cm_event *event);
151 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
152 
153 static const struct blk_mq_ops nvme_rdma_mq_ops;
154 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
155 
156 /* XXX: really should move to a generic header sooner or later.. */
157 static inline void put_unaligned_le24(u32 val, u8 *p)
158 {
159 	*p++ = val;
160 	*p++ = val >> 8;
161 	*p++ = val >> 16;
162 }
163 
164 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
165 {
166 	return queue - queue->ctrl->queues;
167 }
168 
169 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
170 {
171 	return queue->cmnd_capsule_len - sizeof(struct nvme_command);
172 }
173 
174 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
175 		size_t capsule_size, enum dma_data_direction dir)
176 {
177 	ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
178 	kfree(qe->data);
179 }
180 
181 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
182 		size_t capsule_size, enum dma_data_direction dir)
183 {
184 	qe->data = kzalloc(capsule_size, GFP_KERNEL);
185 	if (!qe->data)
186 		return -ENOMEM;
187 
188 	qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
189 	if (ib_dma_mapping_error(ibdev, qe->dma)) {
190 		kfree(qe->data);
191 		return -ENOMEM;
192 	}
193 
194 	return 0;
195 }
196 
197 static void nvme_rdma_free_ring(struct ib_device *ibdev,
198 		struct nvme_rdma_qe *ring, size_t ib_queue_size,
199 		size_t capsule_size, enum dma_data_direction dir)
200 {
201 	int i;
202 
203 	for (i = 0; i < ib_queue_size; i++)
204 		nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
205 	kfree(ring);
206 }
207 
208 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
209 		size_t ib_queue_size, size_t capsule_size,
210 		enum dma_data_direction dir)
211 {
212 	struct nvme_rdma_qe *ring;
213 	int i;
214 
215 	ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
216 	if (!ring)
217 		return NULL;
218 
219 	for (i = 0; i < ib_queue_size; i++) {
220 		if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
221 			goto out_free_ring;
222 	}
223 
224 	return ring;
225 
226 out_free_ring:
227 	nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
228 	return NULL;
229 }
230 
231 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
232 {
233 	pr_debug("QP event %s (%d)\n",
234 		 ib_event_msg(event->event), event->event);
235 
236 }
237 
238 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
239 {
240 	wait_for_completion_interruptible_timeout(&queue->cm_done,
241 			msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
242 	return queue->cm_error;
243 }
244 
245 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
246 {
247 	struct nvme_rdma_device *dev = queue->device;
248 	struct ib_qp_init_attr init_attr;
249 	int ret;
250 
251 	memset(&init_attr, 0, sizeof(init_attr));
252 	init_attr.event_handler = nvme_rdma_qp_event;
253 	/* +1 for drain */
254 	init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
255 	/* +1 for drain */
256 	init_attr.cap.max_recv_wr = queue->queue_size + 1;
257 	init_attr.cap.max_recv_sge = 1;
258 	init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
259 	init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
260 	init_attr.qp_type = IB_QPT_RC;
261 	init_attr.send_cq = queue->ib_cq;
262 	init_attr.recv_cq = queue->ib_cq;
263 
264 	ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
265 
266 	queue->qp = queue->cm_id->qp;
267 	return ret;
268 }
269 
270 static int nvme_rdma_reinit_request(void *data, struct request *rq)
271 {
272 	struct nvme_rdma_ctrl *ctrl = data;
273 	struct nvme_rdma_device *dev = ctrl->device;
274 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
275 	int ret = 0;
276 
277 	ib_dereg_mr(req->mr);
278 
279 	req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
280 			ctrl->max_fr_pages);
281 	if (IS_ERR(req->mr)) {
282 		ret = PTR_ERR(req->mr);
283 		req->mr = NULL;
284 		goto out;
285 	}
286 
287 	req->mr->need_inval = false;
288 
289 out:
290 	return ret;
291 }
292 
293 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
294 		struct request *rq, unsigned int hctx_idx)
295 {
296 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
297 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
298 	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
299 	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
300 	struct nvme_rdma_device *dev = queue->device;
301 
302 	if (req->mr)
303 		ib_dereg_mr(req->mr);
304 
305 	nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
306 			DMA_TO_DEVICE);
307 }
308 
309 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
310 		struct request *rq, unsigned int hctx_idx,
311 		unsigned int numa_node)
312 {
313 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
314 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
315 	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
316 	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
317 	struct nvme_rdma_device *dev = queue->device;
318 	struct ib_device *ibdev = dev->dev;
319 	int ret;
320 
321 	ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
322 			DMA_TO_DEVICE);
323 	if (ret)
324 		return ret;
325 
326 	req->mr = ib_alloc_mr(dev->pd, IB_MR_TYPE_MEM_REG,
327 			ctrl->max_fr_pages);
328 	if (IS_ERR(req->mr)) {
329 		ret = PTR_ERR(req->mr);
330 		goto out_free_qe;
331 	}
332 
333 	req->queue = queue;
334 
335 	return 0;
336 
337 out_free_qe:
338 	nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
339 			DMA_TO_DEVICE);
340 	return -ENOMEM;
341 }
342 
343 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
344 		unsigned int hctx_idx)
345 {
346 	struct nvme_rdma_ctrl *ctrl = data;
347 	struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
348 
349 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
350 
351 	hctx->driver_data = queue;
352 	return 0;
353 }
354 
355 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
356 		unsigned int hctx_idx)
357 {
358 	struct nvme_rdma_ctrl *ctrl = data;
359 	struct nvme_rdma_queue *queue = &ctrl->queues[0];
360 
361 	BUG_ON(hctx_idx != 0);
362 
363 	hctx->driver_data = queue;
364 	return 0;
365 }
366 
367 static void nvme_rdma_free_dev(struct kref *ref)
368 {
369 	struct nvme_rdma_device *ndev =
370 		container_of(ref, struct nvme_rdma_device, ref);
371 
372 	mutex_lock(&device_list_mutex);
373 	list_del(&ndev->entry);
374 	mutex_unlock(&device_list_mutex);
375 
376 	ib_dealloc_pd(ndev->pd);
377 	kfree(ndev);
378 }
379 
380 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
381 {
382 	kref_put(&dev->ref, nvme_rdma_free_dev);
383 }
384 
385 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
386 {
387 	return kref_get_unless_zero(&dev->ref);
388 }
389 
390 static struct nvme_rdma_device *
391 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
392 {
393 	struct nvme_rdma_device *ndev;
394 
395 	mutex_lock(&device_list_mutex);
396 	list_for_each_entry(ndev, &device_list, entry) {
397 		if (ndev->dev->node_guid == cm_id->device->node_guid &&
398 		    nvme_rdma_dev_get(ndev))
399 			goto out_unlock;
400 	}
401 
402 	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
403 	if (!ndev)
404 		goto out_err;
405 
406 	ndev->dev = cm_id->device;
407 	kref_init(&ndev->ref);
408 
409 	ndev->pd = ib_alloc_pd(ndev->dev,
410 		register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
411 	if (IS_ERR(ndev->pd))
412 		goto out_free_dev;
413 
414 	if (!(ndev->dev->attrs.device_cap_flags &
415 	      IB_DEVICE_MEM_MGT_EXTENSIONS)) {
416 		dev_err(&ndev->dev->dev,
417 			"Memory registrations not supported.\n");
418 		goto out_free_pd;
419 	}
420 
421 	list_add(&ndev->entry, &device_list);
422 out_unlock:
423 	mutex_unlock(&device_list_mutex);
424 	return ndev;
425 
426 out_free_pd:
427 	ib_dealloc_pd(ndev->pd);
428 out_free_dev:
429 	kfree(ndev);
430 out_err:
431 	mutex_unlock(&device_list_mutex);
432 	return NULL;
433 }
434 
435 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
436 {
437 	struct nvme_rdma_device *dev;
438 	struct ib_device *ibdev;
439 
440 	dev = queue->device;
441 	ibdev = dev->dev;
442 	rdma_destroy_qp(queue->cm_id);
443 	ib_free_cq(queue->ib_cq);
444 
445 	nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
446 			sizeof(struct nvme_completion), DMA_FROM_DEVICE);
447 
448 	nvme_rdma_dev_put(dev);
449 }
450 
451 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
452 {
453 	struct ib_device *ibdev;
454 	const int send_wr_factor = 3;			/* MR, SEND, INV */
455 	const int cq_factor = send_wr_factor + 1;	/* + RECV */
456 	int comp_vector, idx = nvme_rdma_queue_idx(queue);
457 	int ret;
458 
459 	queue->device = nvme_rdma_find_get_device(queue->cm_id);
460 	if (!queue->device) {
461 		dev_err(queue->cm_id->device->dev.parent,
462 			"no client data found!\n");
463 		return -ECONNREFUSED;
464 	}
465 	ibdev = queue->device->dev;
466 
467 	/*
468 	 * Spread I/O queues completion vectors according their queue index.
469 	 * Admin queues can always go on completion vector 0.
470 	 */
471 	comp_vector = idx == 0 ? idx : idx - 1;
472 
473 	/* +1 for ib_stop_cq */
474 	queue->ib_cq = ib_alloc_cq(ibdev, queue,
475 				cq_factor * queue->queue_size + 1,
476 				comp_vector, IB_POLL_SOFTIRQ);
477 	if (IS_ERR(queue->ib_cq)) {
478 		ret = PTR_ERR(queue->ib_cq);
479 		goto out_put_dev;
480 	}
481 
482 	ret = nvme_rdma_create_qp(queue, send_wr_factor);
483 	if (ret)
484 		goto out_destroy_ib_cq;
485 
486 	queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
487 			sizeof(struct nvme_completion), DMA_FROM_DEVICE);
488 	if (!queue->rsp_ring) {
489 		ret = -ENOMEM;
490 		goto out_destroy_qp;
491 	}
492 
493 	return 0;
494 
495 out_destroy_qp:
496 	ib_destroy_qp(queue->qp);
497 out_destroy_ib_cq:
498 	ib_free_cq(queue->ib_cq);
499 out_put_dev:
500 	nvme_rdma_dev_put(queue->device);
501 	return ret;
502 }
503 
504 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
505 		int idx, size_t queue_size)
506 {
507 	struct nvme_rdma_queue *queue;
508 	struct sockaddr *src_addr = NULL;
509 	int ret;
510 
511 	queue = &ctrl->queues[idx];
512 	queue->ctrl = ctrl;
513 	init_completion(&queue->cm_done);
514 
515 	if (idx > 0)
516 		queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
517 	else
518 		queue->cmnd_capsule_len = sizeof(struct nvme_command);
519 
520 	queue->queue_size = queue_size;
521 	atomic_set(&queue->sig_count, 0);
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_resolve_addr wait failed (%d).\n", ret);
548 		goto out_destroy_cm_id;
549 	}
550 
551 	clear_bit(NVME_RDMA_Q_DELETING, &queue->flags);
552 
553 	return 0;
554 
555 out_destroy_cm_id:
556 	rdma_destroy_id(queue->cm_id);
557 	return ret;
558 }
559 
560 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
561 {
562 	if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
563 		return;
564 
565 	rdma_disconnect(queue->cm_id);
566 	ib_drain_qp(queue->qp);
567 }
568 
569 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
570 {
571 	if (test_and_set_bit(NVME_RDMA_Q_DELETING, &queue->flags))
572 		return;
573 
574 	if (nvme_rdma_queue_idx(queue) == 0) {
575 		nvme_rdma_free_qe(queue->device->dev,
576 			&queue->ctrl->async_event_sqe,
577 			sizeof(struct nvme_command), DMA_TO_DEVICE);
578 	}
579 
580 	nvme_rdma_destroy_queue_ib(queue);
581 	rdma_destroy_id(queue->cm_id);
582 }
583 
584 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
585 {
586 	int i;
587 
588 	for (i = 1; i < ctrl->ctrl.queue_count; i++)
589 		nvme_rdma_free_queue(&ctrl->queues[i]);
590 }
591 
592 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
593 {
594 	int i;
595 
596 	for (i = 1; i < ctrl->ctrl.queue_count; i++)
597 		nvme_rdma_stop_queue(&ctrl->queues[i]);
598 }
599 
600 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
601 {
602 	int ret;
603 
604 	if (idx)
605 		ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
606 	else
607 		ret = nvmf_connect_admin_queue(&ctrl->ctrl);
608 
609 	if (!ret)
610 		set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags);
611 	else
612 		dev_info(ctrl->ctrl.device,
613 			"failed to connect queue: %d ret=%d\n", idx, ret);
614 	return ret;
615 }
616 
617 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
618 {
619 	int i, ret = 0;
620 
621 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
622 		ret = nvme_rdma_start_queue(ctrl, i);
623 		if (ret)
624 			goto out_stop_queues;
625 	}
626 
627 	return 0;
628 
629 out_stop_queues:
630 	for (i--; i >= 1; i--)
631 		nvme_rdma_stop_queue(&ctrl->queues[i]);
632 	return ret;
633 }
634 
635 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
636 {
637 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
638 	struct ib_device *ibdev = ctrl->device->dev;
639 	unsigned int nr_io_queues;
640 	int i, ret;
641 
642 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
643 
644 	/*
645 	 * we map queues according to the device irq vectors for
646 	 * optimal locality so we don't need more queues than
647 	 * completion vectors.
648 	 */
649 	nr_io_queues = min_t(unsigned int, nr_io_queues,
650 				ibdev->num_comp_vectors);
651 
652 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
653 	if (ret)
654 		return ret;
655 
656 	ctrl->ctrl.queue_count = nr_io_queues + 1;
657 	if (ctrl->ctrl.queue_count < 2)
658 		return 0;
659 
660 	dev_info(ctrl->ctrl.device,
661 		"creating %d I/O queues.\n", nr_io_queues);
662 
663 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
664 		ret = nvme_rdma_alloc_queue(ctrl, i,
665 				ctrl->ctrl.sqsize + 1);
666 		if (ret)
667 			goto out_free_queues;
668 	}
669 
670 	return 0;
671 
672 out_free_queues:
673 	for (i--; i >= 1; i--)
674 		nvme_rdma_free_queue(&ctrl->queues[i]);
675 
676 	return ret;
677 }
678 
679 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl, bool admin)
680 {
681 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
682 	struct blk_mq_tag_set *set = admin ?
683 			&ctrl->admin_tag_set : &ctrl->tag_set;
684 
685 	blk_mq_free_tag_set(set);
686 	nvme_rdma_dev_put(ctrl->device);
687 }
688 
689 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
690 		bool admin)
691 {
692 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
693 	struct blk_mq_tag_set *set;
694 	int ret;
695 
696 	if (admin) {
697 		set = &ctrl->admin_tag_set;
698 		memset(set, 0, sizeof(*set));
699 		set->ops = &nvme_rdma_admin_mq_ops;
700 		set->queue_depth = NVME_RDMA_AQ_BLKMQ_DEPTH;
701 		set->reserved_tags = 2; /* connect + keep-alive */
702 		set->numa_node = NUMA_NO_NODE;
703 		set->cmd_size = sizeof(struct nvme_rdma_request) +
704 			SG_CHUNK_SIZE * sizeof(struct scatterlist);
705 		set->driver_data = ctrl;
706 		set->nr_hw_queues = 1;
707 		set->timeout = ADMIN_TIMEOUT;
708 	} else {
709 		set = &ctrl->tag_set;
710 		memset(set, 0, sizeof(*set));
711 		set->ops = &nvme_rdma_mq_ops;
712 		set->queue_depth = nctrl->opts->queue_size;
713 		set->reserved_tags = 1; /* fabric connect */
714 		set->numa_node = NUMA_NO_NODE;
715 		set->flags = BLK_MQ_F_SHOULD_MERGE;
716 		set->cmd_size = sizeof(struct nvme_rdma_request) +
717 			SG_CHUNK_SIZE * sizeof(struct scatterlist);
718 		set->driver_data = ctrl;
719 		set->nr_hw_queues = nctrl->queue_count - 1;
720 		set->timeout = NVME_IO_TIMEOUT;
721 	}
722 
723 	ret = blk_mq_alloc_tag_set(set);
724 	if (ret)
725 		goto out;
726 
727 	/*
728 	 * We need a reference on the device as long as the tag_set is alive,
729 	 * as the MRs in the request structures need a valid ib_device.
730 	 */
731 	ret = nvme_rdma_dev_get(ctrl->device);
732 	if (!ret) {
733 		ret = -EINVAL;
734 		goto out_free_tagset;
735 	}
736 
737 	return set;
738 
739 out_free_tagset:
740 	blk_mq_free_tag_set(set);
741 out:
742 	return ERR_PTR(ret);
743 }
744 
745 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
746 		bool remove)
747 {
748 	nvme_rdma_stop_queue(&ctrl->queues[0]);
749 	if (remove) {
750 		blk_cleanup_queue(ctrl->ctrl.admin_q);
751 		nvme_rdma_free_tagset(&ctrl->ctrl, true);
752 	}
753 	nvme_rdma_free_queue(&ctrl->queues[0]);
754 }
755 
756 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
757 		bool new)
758 {
759 	int error;
760 
761 	error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
762 	if (error)
763 		return error;
764 
765 	ctrl->device = ctrl->queues[0].device;
766 
767 	ctrl->max_fr_pages = min_t(u32, NVME_RDMA_MAX_SEGMENTS,
768 		ctrl->device->dev->attrs.max_fast_reg_page_list_len);
769 
770 	if (new) {
771 		ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
772 		if (IS_ERR(ctrl->ctrl.admin_tagset)) {
773 			error = PTR_ERR(ctrl->ctrl.admin_tagset);
774 			goto out_free_queue;
775 		}
776 
777 		ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
778 		if (IS_ERR(ctrl->ctrl.admin_q)) {
779 			error = PTR_ERR(ctrl->ctrl.admin_q);
780 			goto out_free_tagset;
781 		}
782 	} else {
783 		error = blk_mq_reinit_tagset(&ctrl->admin_tag_set,
784 					     nvme_rdma_reinit_request);
785 		if (error)
786 			goto out_free_queue;
787 	}
788 
789 	error = nvme_rdma_start_queue(ctrl, 0);
790 	if (error)
791 		goto out_cleanup_queue;
792 
793 	error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
794 			&ctrl->ctrl.cap);
795 	if (error) {
796 		dev_err(ctrl->ctrl.device,
797 			"prop_get NVME_REG_CAP failed\n");
798 		goto out_cleanup_queue;
799 	}
800 
801 	ctrl->ctrl.sqsize =
802 		min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
803 
804 	error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
805 	if (error)
806 		goto out_cleanup_queue;
807 
808 	ctrl->ctrl.max_hw_sectors =
809 		(ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
810 
811 	error = nvme_init_identify(&ctrl->ctrl);
812 	if (error)
813 		goto out_cleanup_queue;
814 
815 	error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
816 			&ctrl->async_event_sqe, sizeof(struct nvme_command),
817 			DMA_TO_DEVICE);
818 	if (error)
819 		goto out_cleanup_queue;
820 
821 	return 0;
822 
823 out_cleanup_queue:
824 	if (new)
825 		blk_cleanup_queue(ctrl->ctrl.admin_q);
826 out_free_tagset:
827 	if (new)
828 		nvme_rdma_free_tagset(&ctrl->ctrl, true);
829 out_free_queue:
830 	nvme_rdma_free_queue(&ctrl->queues[0]);
831 	return error;
832 }
833 
834 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
835 		bool remove)
836 {
837 	nvme_rdma_stop_io_queues(ctrl);
838 	if (remove) {
839 		blk_cleanup_queue(ctrl->ctrl.connect_q);
840 		nvme_rdma_free_tagset(&ctrl->ctrl, false);
841 	}
842 	nvme_rdma_free_io_queues(ctrl);
843 }
844 
845 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
846 {
847 	int ret;
848 
849 	ret = nvme_rdma_alloc_io_queues(ctrl);
850 	if (ret)
851 		return ret;
852 
853 	if (new) {
854 		ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
855 		if (IS_ERR(ctrl->ctrl.tagset)) {
856 			ret = PTR_ERR(ctrl->ctrl.tagset);
857 			goto out_free_io_queues;
858 		}
859 
860 		ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
861 		if (IS_ERR(ctrl->ctrl.connect_q)) {
862 			ret = PTR_ERR(ctrl->ctrl.connect_q);
863 			goto out_free_tag_set;
864 		}
865 	} else {
866 		ret = blk_mq_reinit_tagset(&ctrl->tag_set,
867 					   nvme_rdma_reinit_request);
868 		if (ret)
869 			goto out_free_io_queues;
870 
871 		blk_mq_update_nr_hw_queues(&ctrl->tag_set,
872 			ctrl->ctrl.queue_count - 1);
873 	}
874 
875 	ret = nvme_rdma_start_io_queues(ctrl);
876 	if (ret)
877 		goto out_cleanup_connect_q;
878 
879 	return 0;
880 
881 out_cleanup_connect_q:
882 	if (new)
883 		blk_cleanup_queue(ctrl->ctrl.connect_q);
884 out_free_tag_set:
885 	if (new)
886 		nvme_rdma_free_tagset(&ctrl->ctrl, false);
887 out_free_io_queues:
888 	nvme_rdma_free_io_queues(ctrl);
889 	return ret;
890 }
891 
892 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
893 {
894 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
895 
896 	if (list_empty(&ctrl->list))
897 		goto free_ctrl;
898 
899 	mutex_lock(&nvme_rdma_ctrl_mutex);
900 	list_del(&ctrl->list);
901 	mutex_unlock(&nvme_rdma_ctrl_mutex);
902 
903 	kfree(ctrl->queues);
904 	nvmf_free_options(nctrl->opts);
905 free_ctrl:
906 	kfree(ctrl);
907 }
908 
909 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
910 {
911 	/* If we are resetting/deleting then do nothing */
912 	if (ctrl->ctrl.state != NVME_CTRL_RECONNECTING) {
913 		WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
914 			ctrl->ctrl.state == NVME_CTRL_LIVE);
915 		return;
916 	}
917 
918 	if (nvmf_should_reconnect(&ctrl->ctrl)) {
919 		dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
920 			ctrl->ctrl.opts->reconnect_delay);
921 		queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
922 				ctrl->ctrl.opts->reconnect_delay * HZ);
923 	} else {
924 		dev_info(ctrl->ctrl.device, "Removing controller...\n");
925 		queue_work(nvme_wq, &ctrl->delete_work);
926 	}
927 }
928 
929 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
930 {
931 	struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
932 			struct nvme_rdma_ctrl, reconnect_work);
933 	bool changed;
934 	int ret;
935 
936 	++ctrl->ctrl.nr_reconnects;
937 
938 	if (ctrl->ctrl.queue_count > 1)
939 		nvme_rdma_destroy_io_queues(ctrl, false);
940 
941 	nvme_rdma_destroy_admin_queue(ctrl, false);
942 	ret = nvme_rdma_configure_admin_queue(ctrl, false);
943 	if (ret)
944 		goto requeue;
945 
946 	if (ctrl->ctrl.queue_count > 1) {
947 		ret = nvme_rdma_configure_io_queues(ctrl, false);
948 		if (ret)
949 			goto requeue;
950 	}
951 
952 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
953 	if (!changed) {
954 		/* state change failure is ok if we're in DELETING state */
955 		WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
956 		return;
957 	}
958 
959 	ctrl->ctrl.nr_reconnects = 0;
960 
961 	nvme_start_ctrl(&ctrl->ctrl);
962 
963 	dev_info(ctrl->ctrl.device, "Successfully reconnected\n");
964 
965 	return;
966 
967 requeue:
968 	dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
969 			ctrl->ctrl.nr_reconnects);
970 	nvme_rdma_reconnect_or_remove(ctrl);
971 }
972 
973 static void nvme_rdma_error_recovery_work(struct work_struct *work)
974 {
975 	struct nvme_rdma_ctrl *ctrl = container_of(work,
976 			struct nvme_rdma_ctrl, err_work);
977 
978 	nvme_stop_keep_alive(&ctrl->ctrl);
979 
980 	if (ctrl->ctrl.queue_count > 1) {
981 		nvme_stop_queues(&ctrl->ctrl);
982 		nvme_rdma_stop_io_queues(ctrl);
983 	}
984 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
985 	nvme_rdma_stop_queue(&ctrl->queues[0]);
986 
987 	/* We must take care of fastfail/requeue all our inflight requests */
988 	if (ctrl->ctrl.queue_count > 1)
989 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
990 					nvme_cancel_request, &ctrl->ctrl);
991 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
992 				nvme_cancel_request, &ctrl->ctrl);
993 
994 	/*
995 	 * queues are not a live anymore, so restart the queues to fail fast
996 	 * new IO
997 	 */
998 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
999 	nvme_start_queues(&ctrl->ctrl);
1000 
1001 	nvme_rdma_reconnect_or_remove(ctrl);
1002 }
1003 
1004 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
1005 {
1006 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING))
1007 		return;
1008 
1009 	queue_work(nvme_wq, &ctrl->err_work);
1010 }
1011 
1012 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
1013 		const char *op)
1014 {
1015 	struct nvme_rdma_queue *queue = cq->cq_context;
1016 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1017 
1018 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1019 		dev_info(ctrl->ctrl.device,
1020 			     "%s for CQE 0x%p failed with status %s (%d)\n",
1021 			     op, wc->wr_cqe,
1022 			     ib_wc_status_msg(wc->status), wc->status);
1023 	nvme_rdma_error_recovery(ctrl);
1024 }
1025 
1026 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1027 {
1028 	if (unlikely(wc->status != IB_WC_SUCCESS))
1029 		nvme_rdma_wr_error(cq, wc, "MEMREG");
1030 }
1031 
1032 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1033 {
1034 	if (unlikely(wc->status != IB_WC_SUCCESS))
1035 		nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1036 }
1037 
1038 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1039 		struct nvme_rdma_request *req)
1040 {
1041 	struct ib_send_wr *bad_wr;
1042 	struct ib_send_wr wr = {
1043 		.opcode		    = IB_WR_LOCAL_INV,
1044 		.next		    = NULL,
1045 		.num_sge	    = 0,
1046 		.send_flags	    = 0,
1047 		.ex.invalidate_rkey = req->mr->rkey,
1048 	};
1049 
1050 	req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1051 	wr.wr_cqe = &req->reg_cqe;
1052 
1053 	return ib_post_send(queue->qp, &wr, &bad_wr);
1054 }
1055 
1056 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1057 		struct request *rq)
1058 {
1059 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1060 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1061 	struct nvme_rdma_device *dev = queue->device;
1062 	struct ib_device *ibdev = dev->dev;
1063 	int res;
1064 
1065 	if (!blk_rq_bytes(rq))
1066 		return;
1067 
1068 	if (req->mr->need_inval) {
1069 		res = nvme_rdma_inv_rkey(queue, req);
1070 		if (unlikely(res < 0)) {
1071 			dev_err(ctrl->ctrl.device,
1072 				"Queueing INV WR for rkey %#x failed (%d)\n",
1073 				req->mr->rkey, res);
1074 			nvme_rdma_error_recovery(queue->ctrl);
1075 		}
1076 	}
1077 
1078 	ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1079 			req->nents, rq_data_dir(rq) ==
1080 				    WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1081 
1082 	nvme_cleanup_cmd(rq);
1083 	sg_free_table_chained(&req->sg_table, true);
1084 }
1085 
1086 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1087 {
1088 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1089 
1090 	sg->addr = 0;
1091 	put_unaligned_le24(0, sg->length);
1092 	put_unaligned_le32(0, sg->key);
1093 	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1094 	return 0;
1095 }
1096 
1097 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1098 		struct nvme_rdma_request *req, struct nvme_command *c)
1099 {
1100 	struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1101 
1102 	req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
1103 	req->sge[1].length = sg_dma_len(req->sg_table.sgl);
1104 	req->sge[1].lkey = queue->device->pd->local_dma_lkey;
1105 
1106 	sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1107 	sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
1108 	sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1109 
1110 	req->inline_data = true;
1111 	req->num_sge++;
1112 	return 0;
1113 }
1114 
1115 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1116 		struct nvme_rdma_request *req, struct nvme_command *c)
1117 {
1118 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1119 
1120 	sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1121 	put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1122 	put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1123 	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1124 	return 0;
1125 }
1126 
1127 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1128 		struct nvme_rdma_request *req, struct nvme_command *c,
1129 		int count)
1130 {
1131 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1132 	int nr;
1133 
1134 	/*
1135 	 * Align the MR to a 4K page size to match the ctrl page size and
1136 	 * the block virtual boundary.
1137 	 */
1138 	nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1139 	if (unlikely(nr < count)) {
1140 		if (nr < 0)
1141 			return nr;
1142 		return -EINVAL;
1143 	}
1144 
1145 	ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1146 
1147 	req->reg_cqe.done = nvme_rdma_memreg_done;
1148 	memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1149 	req->reg_wr.wr.opcode = IB_WR_REG_MR;
1150 	req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1151 	req->reg_wr.wr.num_sge = 0;
1152 	req->reg_wr.mr = req->mr;
1153 	req->reg_wr.key = req->mr->rkey;
1154 	req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1155 			     IB_ACCESS_REMOTE_READ |
1156 			     IB_ACCESS_REMOTE_WRITE;
1157 
1158 	req->mr->need_inval = true;
1159 
1160 	sg->addr = cpu_to_le64(req->mr->iova);
1161 	put_unaligned_le24(req->mr->length, sg->length);
1162 	put_unaligned_le32(req->mr->rkey, sg->key);
1163 	sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1164 			NVME_SGL_FMT_INVALIDATE;
1165 
1166 	return 0;
1167 }
1168 
1169 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1170 		struct request *rq, struct nvme_command *c)
1171 {
1172 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1173 	struct nvme_rdma_device *dev = queue->device;
1174 	struct ib_device *ibdev = dev->dev;
1175 	int count, ret;
1176 
1177 	req->num_sge = 1;
1178 	req->inline_data = false;
1179 	req->mr->need_inval = false;
1180 
1181 	c->common.flags |= NVME_CMD_SGL_METABUF;
1182 
1183 	if (!blk_rq_bytes(rq))
1184 		return nvme_rdma_set_sg_null(c);
1185 
1186 	req->sg_table.sgl = req->first_sgl;
1187 	ret = sg_alloc_table_chained(&req->sg_table,
1188 			blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1189 	if (ret)
1190 		return -ENOMEM;
1191 
1192 	req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1193 
1194 	count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1195 		    rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1196 	if (unlikely(count <= 0)) {
1197 		sg_free_table_chained(&req->sg_table, true);
1198 		return -EIO;
1199 	}
1200 
1201 	if (count == 1) {
1202 		if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1203 		    blk_rq_payload_bytes(rq) <=
1204 				nvme_rdma_inline_data_size(queue))
1205 			return nvme_rdma_map_sg_inline(queue, req, c);
1206 
1207 		if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
1208 			return nvme_rdma_map_sg_single(queue, req, c);
1209 	}
1210 
1211 	return nvme_rdma_map_sg_fr(queue, req, c, count);
1212 }
1213 
1214 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1215 {
1216 	if (unlikely(wc->status != IB_WC_SUCCESS))
1217 		nvme_rdma_wr_error(cq, wc, "SEND");
1218 }
1219 
1220 /*
1221  * We want to signal completion at least every queue depth/2.  This returns the
1222  * largest power of two that is not above half of (queue size + 1) to optimize
1223  * (avoid divisions).
1224  */
1225 static inline bool nvme_rdma_queue_sig_limit(struct nvme_rdma_queue *queue)
1226 {
1227 	int limit = 1 << ilog2((queue->queue_size + 1) / 2);
1228 
1229 	return (atomic_inc_return(&queue->sig_count) & (limit - 1)) == 0;
1230 }
1231 
1232 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1233 		struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1234 		struct ib_send_wr *first, bool flush)
1235 {
1236 	struct ib_send_wr wr, *bad_wr;
1237 	int ret;
1238 
1239 	sge->addr   = qe->dma;
1240 	sge->length = sizeof(struct nvme_command),
1241 	sge->lkey   = queue->device->pd->local_dma_lkey;
1242 
1243 	qe->cqe.done = nvme_rdma_send_done;
1244 
1245 	wr.next       = NULL;
1246 	wr.wr_cqe     = &qe->cqe;
1247 	wr.sg_list    = sge;
1248 	wr.num_sge    = num_sge;
1249 	wr.opcode     = IB_WR_SEND;
1250 	wr.send_flags = 0;
1251 
1252 	/*
1253 	 * Unsignalled send completions are another giant desaster in the
1254 	 * IB Verbs spec:  If we don't regularly post signalled sends
1255 	 * the send queue will fill up and only a QP reset will rescue us.
1256 	 * Would have been way to obvious to handle this in hardware or
1257 	 * at least the RDMA stack..
1258 	 *
1259 	 * Always signal the flushes. The magic request used for the flush
1260 	 * sequencer is not allocated in our driver's tagset and it's
1261 	 * triggered to be freed by blk_cleanup_queue(). So we need to
1262 	 * always mark it as signaled to ensure that the "wr_cqe", which is
1263 	 * embedded in request's payload, is not freed when __ib_process_cq()
1264 	 * calls wr_cqe->done().
1265 	 */
1266 	if (nvme_rdma_queue_sig_limit(queue) || flush)
1267 		wr.send_flags |= IB_SEND_SIGNALED;
1268 
1269 	if (first)
1270 		first->next = &wr;
1271 	else
1272 		first = &wr;
1273 
1274 	ret = ib_post_send(queue->qp, first, &bad_wr);
1275 	if (unlikely(ret)) {
1276 		dev_err(queue->ctrl->ctrl.device,
1277 			     "%s failed with error code %d\n", __func__, ret);
1278 	}
1279 	return ret;
1280 }
1281 
1282 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1283 		struct nvme_rdma_qe *qe)
1284 {
1285 	struct ib_recv_wr wr, *bad_wr;
1286 	struct ib_sge list;
1287 	int ret;
1288 
1289 	list.addr   = qe->dma;
1290 	list.length = sizeof(struct nvme_completion);
1291 	list.lkey   = queue->device->pd->local_dma_lkey;
1292 
1293 	qe->cqe.done = nvme_rdma_recv_done;
1294 
1295 	wr.next     = NULL;
1296 	wr.wr_cqe   = &qe->cqe;
1297 	wr.sg_list  = &list;
1298 	wr.num_sge  = 1;
1299 
1300 	ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1301 	if (unlikely(ret)) {
1302 		dev_err(queue->ctrl->ctrl.device,
1303 			"%s failed with error code %d\n", __func__, ret);
1304 	}
1305 	return ret;
1306 }
1307 
1308 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1309 {
1310 	u32 queue_idx = nvme_rdma_queue_idx(queue);
1311 
1312 	if (queue_idx == 0)
1313 		return queue->ctrl->admin_tag_set.tags[queue_idx];
1314 	return queue->ctrl->tag_set.tags[queue_idx - 1];
1315 }
1316 
1317 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg, int aer_idx)
1318 {
1319 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1320 	struct nvme_rdma_queue *queue = &ctrl->queues[0];
1321 	struct ib_device *dev = queue->device->dev;
1322 	struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1323 	struct nvme_command *cmd = sqe->data;
1324 	struct ib_sge sge;
1325 	int ret;
1326 
1327 	if (WARN_ON_ONCE(aer_idx != 0))
1328 		return;
1329 
1330 	ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1331 
1332 	memset(cmd, 0, sizeof(*cmd));
1333 	cmd->common.opcode = nvme_admin_async_event;
1334 	cmd->common.command_id = NVME_RDMA_AQ_BLKMQ_DEPTH;
1335 	cmd->common.flags |= NVME_CMD_SGL_METABUF;
1336 	nvme_rdma_set_sg_null(cmd);
1337 
1338 	ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1339 			DMA_TO_DEVICE);
1340 
1341 	ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL, false);
1342 	WARN_ON_ONCE(ret);
1343 }
1344 
1345 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1346 		struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1347 {
1348 	struct request *rq;
1349 	struct nvme_rdma_request *req;
1350 	int ret = 0;
1351 
1352 	rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1353 	if (!rq) {
1354 		dev_err(queue->ctrl->ctrl.device,
1355 			"tag 0x%x on QP %#x not found\n",
1356 			cqe->command_id, queue->qp->qp_num);
1357 		nvme_rdma_error_recovery(queue->ctrl);
1358 		return ret;
1359 	}
1360 	req = blk_mq_rq_to_pdu(rq);
1361 
1362 	if (rq->tag == tag)
1363 		ret = 1;
1364 
1365 	if ((wc->wc_flags & IB_WC_WITH_INVALIDATE) &&
1366 	    wc->ex.invalidate_rkey == req->mr->rkey)
1367 		req->mr->need_inval = false;
1368 
1369 	nvme_end_request(rq, cqe->status, cqe->result);
1370 	return ret;
1371 }
1372 
1373 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1374 {
1375 	struct nvme_rdma_qe *qe =
1376 		container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1377 	struct nvme_rdma_queue *queue = cq->cq_context;
1378 	struct ib_device *ibdev = queue->device->dev;
1379 	struct nvme_completion *cqe = qe->data;
1380 	const size_t len = sizeof(struct nvme_completion);
1381 	int ret = 0;
1382 
1383 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1384 		nvme_rdma_wr_error(cq, wc, "RECV");
1385 		return 0;
1386 	}
1387 
1388 	ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1389 	/*
1390 	 * AEN requests are special as they don't time out and can
1391 	 * survive any kind of queue freeze and often don't respond to
1392 	 * aborts.  We don't even bother to allocate a struct request
1393 	 * for them but rather special case them here.
1394 	 */
1395 	if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1396 			cqe->command_id >= NVME_RDMA_AQ_BLKMQ_DEPTH))
1397 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1398 				&cqe->result);
1399 	else
1400 		ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1401 	ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1402 
1403 	nvme_rdma_post_recv(queue, qe);
1404 	return ret;
1405 }
1406 
1407 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1408 {
1409 	__nvme_rdma_recv_done(cq, wc, -1);
1410 }
1411 
1412 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1413 {
1414 	int ret, i;
1415 
1416 	for (i = 0; i < queue->queue_size; i++) {
1417 		ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1418 		if (ret)
1419 			goto out_destroy_queue_ib;
1420 	}
1421 
1422 	return 0;
1423 
1424 out_destroy_queue_ib:
1425 	nvme_rdma_destroy_queue_ib(queue);
1426 	return ret;
1427 }
1428 
1429 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1430 		struct rdma_cm_event *ev)
1431 {
1432 	struct rdma_cm_id *cm_id = queue->cm_id;
1433 	int status = ev->status;
1434 	const char *rej_msg;
1435 	const struct nvme_rdma_cm_rej *rej_data;
1436 	u8 rej_data_len;
1437 
1438 	rej_msg = rdma_reject_msg(cm_id, status);
1439 	rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1440 
1441 	if (rej_data && rej_data_len >= sizeof(u16)) {
1442 		u16 sts = le16_to_cpu(rej_data->sts);
1443 
1444 		dev_err(queue->ctrl->ctrl.device,
1445 		      "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1446 		      status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1447 	} else {
1448 		dev_err(queue->ctrl->ctrl.device,
1449 			"Connect rejected: status %d (%s).\n", status, rej_msg);
1450 	}
1451 
1452 	return -ECONNRESET;
1453 }
1454 
1455 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1456 {
1457 	int ret;
1458 
1459 	ret = nvme_rdma_create_queue_ib(queue);
1460 	if (ret)
1461 		return ret;
1462 
1463 	ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1464 	if (ret) {
1465 		dev_err(queue->ctrl->ctrl.device,
1466 			"rdma_resolve_route failed (%d).\n",
1467 			queue->cm_error);
1468 		goto out_destroy_queue;
1469 	}
1470 
1471 	return 0;
1472 
1473 out_destroy_queue:
1474 	nvme_rdma_destroy_queue_ib(queue);
1475 	return ret;
1476 }
1477 
1478 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1479 {
1480 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1481 	struct rdma_conn_param param = { };
1482 	struct nvme_rdma_cm_req priv = { };
1483 	int ret;
1484 
1485 	param.qp_num = queue->qp->qp_num;
1486 	param.flow_control = 1;
1487 
1488 	param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1489 	/* maximum retry count */
1490 	param.retry_count = 7;
1491 	param.rnr_retry_count = 7;
1492 	param.private_data = &priv;
1493 	param.private_data_len = sizeof(priv);
1494 
1495 	priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1496 	priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1497 	/*
1498 	 * set the admin queue depth to the minimum size
1499 	 * specified by the Fabrics standard.
1500 	 */
1501 	if (priv.qid == 0) {
1502 		priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1503 		priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1504 	} else {
1505 		/*
1506 		 * current interpretation of the fabrics spec
1507 		 * is at minimum you make hrqsize sqsize+1, or a
1508 		 * 1's based representation of sqsize.
1509 		 */
1510 		priv.hrqsize = cpu_to_le16(queue->queue_size);
1511 		priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1512 	}
1513 
1514 	ret = rdma_connect(queue->cm_id, &param);
1515 	if (ret) {
1516 		dev_err(ctrl->ctrl.device,
1517 			"rdma_connect failed (%d).\n", ret);
1518 		goto out_destroy_queue_ib;
1519 	}
1520 
1521 	return 0;
1522 
1523 out_destroy_queue_ib:
1524 	nvme_rdma_destroy_queue_ib(queue);
1525 	return ret;
1526 }
1527 
1528 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1529 		struct rdma_cm_event *ev)
1530 {
1531 	struct nvme_rdma_queue *queue = cm_id->context;
1532 	int cm_error = 0;
1533 
1534 	dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1535 		rdma_event_msg(ev->event), ev->event,
1536 		ev->status, cm_id);
1537 
1538 	switch (ev->event) {
1539 	case RDMA_CM_EVENT_ADDR_RESOLVED:
1540 		cm_error = nvme_rdma_addr_resolved(queue);
1541 		break;
1542 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
1543 		cm_error = nvme_rdma_route_resolved(queue);
1544 		break;
1545 	case RDMA_CM_EVENT_ESTABLISHED:
1546 		queue->cm_error = nvme_rdma_conn_established(queue);
1547 		/* complete cm_done regardless of success/failure */
1548 		complete(&queue->cm_done);
1549 		return 0;
1550 	case RDMA_CM_EVENT_REJECTED:
1551 		nvme_rdma_destroy_queue_ib(queue);
1552 		cm_error = nvme_rdma_conn_rejected(queue, ev);
1553 		break;
1554 	case RDMA_CM_EVENT_ROUTE_ERROR:
1555 	case RDMA_CM_EVENT_CONNECT_ERROR:
1556 	case RDMA_CM_EVENT_UNREACHABLE:
1557 		nvme_rdma_destroy_queue_ib(queue);
1558 	case RDMA_CM_EVENT_ADDR_ERROR:
1559 		dev_dbg(queue->ctrl->ctrl.device,
1560 			"CM error event %d\n", ev->event);
1561 		cm_error = -ECONNRESET;
1562 		break;
1563 	case RDMA_CM_EVENT_DISCONNECTED:
1564 	case RDMA_CM_EVENT_ADDR_CHANGE:
1565 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1566 		dev_dbg(queue->ctrl->ctrl.device,
1567 			"disconnect received - connection closed\n");
1568 		nvme_rdma_error_recovery(queue->ctrl);
1569 		break;
1570 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1571 		/* device removal is handled via the ib_client API */
1572 		break;
1573 	default:
1574 		dev_err(queue->ctrl->ctrl.device,
1575 			"Unexpected RDMA CM event (%d)\n", ev->event);
1576 		nvme_rdma_error_recovery(queue->ctrl);
1577 		break;
1578 	}
1579 
1580 	if (cm_error) {
1581 		queue->cm_error = cm_error;
1582 		complete(&queue->cm_done);
1583 	}
1584 
1585 	return 0;
1586 }
1587 
1588 static enum blk_eh_timer_return
1589 nvme_rdma_timeout(struct request *rq, bool reserved)
1590 {
1591 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1592 
1593 	/* queue error recovery */
1594 	nvme_rdma_error_recovery(req->queue->ctrl);
1595 
1596 	/* fail with DNR on cmd timeout */
1597 	nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1598 
1599 	return BLK_EH_HANDLED;
1600 }
1601 
1602 /*
1603  * We cannot accept any other command until the Connect command has completed.
1604  */
1605 static inline blk_status_t
1606 nvme_rdma_queue_is_ready(struct nvme_rdma_queue *queue, struct request *rq)
1607 {
1608 	if (unlikely(!test_bit(NVME_RDMA_Q_LIVE, &queue->flags))) {
1609 		struct nvme_command *cmd = nvme_req(rq)->cmd;
1610 
1611 		if (!blk_rq_is_passthrough(rq) ||
1612 		    cmd->common.opcode != nvme_fabrics_command ||
1613 		    cmd->fabrics.fctype != nvme_fabrics_type_connect) {
1614 			/*
1615 			 * reconnecting state means transport disruption, which
1616 			 * can take a long time and even might fail permanently,
1617 			 * fail fast to give upper layers a chance to failover.
1618 			 * deleting state means that the ctrl will never accept
1619 			 * commands again, fail it permanently.
1620 			 */
1621 			if (queue->ctrl->ctrl.state == NVME_CTRL_RECONNECTING ||
1622 			    queue->ctrl->ctrl.state == NVME_CTRL_DELETING) {
1623 				nvme_req(rq)->status = NVME_SC_ABORT_REQ;
1624 				return BLK_STS_IOERR;
1625 			}
1626 			return BLK_STS_RESOURCE; /* try again later */
1627 		}
1628 	}
1629 
1630 	return 0;
1631 }
1632 
1633 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1634 		const struct blk_mq_queue_data *bd)
1635 {
1636 	struct nvme_ns *ns = hctx->queue->queuedata;
1637 	struct nvme_rdma_queue *queue = hctx->driver_data;
1638 	struct request *rq = bd->rq;
1639 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1640 	struct nvme_rdma_qe *sqe = &req->sqe;
1641 	struct nvme_command *c = sqe->data;
1642 	bool flush = false;
1643 	struct ib_device *dev;
1644 	blk_status_t ret;
1645 	int err;
1646 
1647 	WARN_ON_ONCE(rq->tag < 0);
1648 
1649 	ret = nvme_rdma_queue_is_ready(queue, rq);
1650 	if (unlikely(ret))
1651 		return ret;
1652 
1653 	dev = queue->device->dev;
1654 	ib_dma_sync_single_for_cpu(dev, sqe->dma,
1655 			sizeof(struct nvme_command), DMA_TO_DEVICE);
1656 
1657 	ret = nvme_setup_cmd(ns, rq, c);
1658 	if (ret)
1659 		return ret;
1660 
1661 	blk_mq_start_request(rq);
1662 
1663 	err = nvme_rdma_map_data(queue, rq, c);
1664 	if (unlikely(err < 0)) {
1665 		dev_err(queue->ctrl->ctrl.device,
1666 			     "Failed to map data (%d)\n", err);
1667 		nvme_cleanup_cmd(rq);
1668 		goto err;
1669 	}
1670 
1671 	ib_dma_sync_single_for_device(dev, sqe->dma,
1672 			sizeof(struct nvme_command), DMA_TO_DEVICE);
1673 
1674 	if (req_op(rq) == REQ_OP_FLUSH)
1675 		flush = true;
1676 	err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1677 			req->mr->need_inval ? &req->reg_wr.wr : NULL, flush);
1678 	if (unlikely(err)) {
1679 		nvme_rdma_unmap_data(queue, rq);
1680 		goto err;
1681 	}
1682 
1683 	return BLK_STS_OK;
1684 err:
1685 	if (err == -ENOMEM || err == -EAGAIN)
1686 		return BLK_STS_RESOURCE;
1687 	return BLK_STS_IOERR;
1688 }
1689 
1690 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1691 {
1692 	struct nvme_rdma_queue *queue = hctx->driver_data;
1693 	struct ib_cq *cq = queue->ib_cq;
1694 	struct ib_wc wc;
1695 	int found = 0;
1696 
1697 	while (ib_poll_cq(cq, 1, &wc) > 0) {
1698 		struct ib_cqe *cqe = wc.wr_cqe;
1699 
1700 		if (cqe) {
1701 			if (cqe->done == nvme_rdma_recv_done)
1702 				found |= __nvme_rdma_recv_done(cq, &wc, tag);
1703 			else
1704 				cqe->done(cq, &wc);
1705 		}
1706 	}
1707 
1708 	return found;
1709 }
1710 
1711 static void nvme_rdma_complete_rq(struct request *rq)
1712 {
1713 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1714 
1715 	nvme_rdma_unmap_data(req->queue, rq);
1716 	nvme_complete_rq(rq);
1717 }
1718 
1719 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1720 {
1721 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
1722 
1723 	return blk_mq_rdma_map_queues(set, ctrl->device->dev, 0);
1724 }
1725 
1726 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1727 	.queue_rq	= nvme_rdma_queue_rq,
1728 	.complete	= nvme_rdma_complete_rq,
1729 	.init_request	= nvme_rdma_init_request,
1730 	.exit_request	= nvme_rdma_exit_request,
1731 	.init_hctx	= nvme_rdma_init_hctx,
1732 	.poll		= nvme_rdma_poll,
1733 	.timeout	= nvme_rdma_timeout,
1734 	.map_queues	= nvme_rdma_map_queues,
1735 };
1736 
1737 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1738 	.queue_rq	= nvme_rdma_queue_rq,
1739 	.complete	= nvme_rdma_complete_rq,
1740 	.init_request	= nvme_rdma_init_request,
1741 	.exit_request	= nvme_rdma_exit_request,
1742 	.init_hctx	= nvme_rdma_init_admin_hctx,
1743 	.timeout	= nvme_rdma_timeout,
1744 };
1745 
1746 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1747 {
1748 	cancel_work_sync(&ctrl->err_work);
1749 	cancel_delayed_work_sync(&ctrl->reconnect_work);
1750 
1751 	if (ctrl->ctrl.queue_count > 1) {
1752 		nvme_stop_queues(&ctrl->ctrl);
1753 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
1754 					nvme_cancel_request, &ctrl->ctrl);
1755 		nvme_rdma_destroy_io_queues(ctrl, shutdown);
1756 	}
1757 
1758 	if (shutdown)
1759 		nvme_shutdown_ctrl(&ctrl->ctrl);
1760 	else
1761 		nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1762 
1763 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1764 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1765 				nvme_cancel_request, &ctrl->ctrl);
1766 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1767 	nvme_rdma_destroy_admin_queue(ctrl, shutdown);
1768 }
1769 
1770 static void nvme_rdma_remove_ctrl(struct nvme_rdma_ctrl *ctrl)
1771 {
1772 	nvme_remove_namespaces(&ctrl->ctrl);
1773 	nvme_rdma_shutdown_ctrl(ctrl, true);
1774 	nvme_uninit_ctrl(&ctrl->ctrl);
1775 	nvme_put_ctrl(&ctrl->ctrl);
1776 }
1777 
1778 static void nvme_rdma_del_ctrl_work(struct work_struct *work)
1779 {
1780 	struct nvme_rdma_ctrl *ctrl = container_of(work,
1781 				struct nvme_rdma_ctrl, delete_work);
1782 
1783 	nvme_stop_ctrl(&ctrl->ctrl);
1784 	nvme_rdma_remove_ctrl(ctrl);
1785 }
1786 
1787 static int __nvme_rdma_del_ctrl(struct nvme_rdma_ctrl *ctrl)
1788 {
1789 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_DELETING))
1790 		return -EBUSY;
1791 
1792 	if (!queue_work(nvme_wq, &ctrl->delete_work))
1793 		return -EBUSY;
1794 
1795 	return 0;
1796 }
1797 
1798 static int nvme_rdma_del_ctrl(struct nvme_ctrl *nctrl)
1799 {
1800 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
1801 	int ret = 0;
1802 
1803 	/*
1804 	 * Keep a reference until all work is flushed since
1805 	 * __nvme_rdma_del_ctrl can free the ctrl mem
1806 	 */
1807 	if (!kref_get_unless_zero(&ctrl->ctrl.kref))
1808 		return -EBUSY;
1809 	ret = __nvme_rdma_del_ctrl(ctrl);
1810 	if (!ret)
1811 		flush_work(&ctrl->delete_work);
1812 	nvme_put_ctrl(&ctrl->ctrl);
1813 	return ret;
1814 }
1815 
1816 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1817 {
1818 	struct nvme_rdma_ctrl *ctrl =
1819 		container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1820 	int ret;
1821 	bool changed;
1822 
1823 	nvme_stop_ctrl(&ctrl->ctrl);
1824 	nvme_rdma_shutdown_ctrl(ctrl, false);
1825 
1826 	ret = nvme_rdma_configure_admin_queue(ctrl, false);
1827 	if (ret)
1828 		goto out_fail;
1829 
1830 	if (ctrl->ctrl.queue_count > 1) {
1831 		ret = nvme_rdma_configure_io_queues(ctrl, false);
1832 		if (ret)
1833 			goto out_fail;
1834 	}
1835 
1836 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1837 	WARN_ON_ONCE(!changed);
1838 
1839 	nvme_start_ctrl(&ctrl->ctrl);
1840 
1841 	return;
1842 
1843 out_fail:
1844 	dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1845 	nvme_rdma_remove_ctrl(ctrl);
1846 }
1847 
1848 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1849 	.name			= "rdma",
1850 	.module			= THIS_MODULE,
1851 	.flags			= NVME_F_FABRICS,
1852 	.reg_read32		= nvmf_reg_read32,
1853 	.reg_read64		= nvmf_reg_read64,
1854 	.reg_write32		= nvmf_reg_write32,
1855 	.free_ctrl		= nvme_rdma_free_ctrl,
1856 	.submit_async_event	= nvme_rdma_submit_async_event,
1857 	.delete_ctrl		= nvme_rdma_del_ctrl,
1858 	.get_address		= nvmf_get_address,
1859 };
1860 
1861 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1862 		struct nvmf_ctrl_options *opts)
1863 {
1864 	struct nvme_rdma_ctrl *ctrl;
1865 	int ret;
1866 	bool changed;
1867 	char *port;
1868 
1869 	ctrl = kzalloc(sizeof(*ctrl), GFP_KERNEL);
1870 	if (!ctrl)
1871 		return ERR_PTR(-ENOMEM);
1872 	ctrl->ctrl.opts = opts;
1873 	INIT_LIST_HEAD(&ctrl->list);
1874 
1875 	if (opts->mask & NVMF_OPT_TRSVCID)
1876 		port = opts->trsvcid;
1877 	else
1878 		port = __stringify(NVME_RDMA_IP_PORT);
1879 
1880 	ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1881 			opts->traddr, port, &ctrl->addr);
1882 	if (ret) {
1883 		pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1884 		goto out_free_ctrl;
1885 	}
1886 
1887 	if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1888 		ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1889 			opts->host_traddr, NULL, &ctrl->src_addr);
1890 		if (ret) {
1891 			pr_err("malformed src address passed: %s\n",
1892 			       opts->host_traddr);
1893 			goto out_free_ctrl;
1894 		}
1895 	}
1896 
1897 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1898 				0 /* no quirks, we're perfect! */);
1899 	if (ret)
1900 		goto out_free_ctrl;
1901 
1902 	INIT_DELAYED_WORK(&ctrl->reconnect_work,
1903 			nvme_rdma_reconnect_ctrl_work);
1904 	INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1905 	INIT_WORK(&ctrl->delete_work, nvme_rdma_del_ctrl_work);
1906 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1907 
1908 	ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1909 	ctrl->ctrl.sqsize = opts->queue_size - 1;
1910 	ctrl->ctrl.kato = opts->kato;
1911 
1912 	ret = -ENOMEM;
1913 	ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1914 				GFP_KERNEL);
1915 	if (!ctrl->queues)
1916 		goto out_uninit_ctrl;
1917 
1918 	ret = nvme_rdma_configure_admin_queue(ctrl, true);
1919 	if (ret)
1920 		goto out_kfree_queues;
1921 
1922 	/* sanity check icdoff */
1923 	if (ctrl->ctrl.icdoff) {
1924 		dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1925 		ret = -EINVAL;
1926 		goto out_remove_admin_queue;
1927 	}
1928 
1929 	/* sanity check keyed sgls */
1930 	if (!(ctrl->ctrl.sgls & (1 << 20))) {
1931 		dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1932 		ret = -EINVAL;
1933 		goto out_remove_admin_queue;
1934 	}
1935 
1936 	if (opts->queue_size > ctrl->ctrl.maxcmd) {
1937 		/* warn if maxcmd is lower than queue_size */
1938 		dev_warn(ctrl->ctrl.device,
1939 			"queue_size %zu > ctrl maxcmd %u, clamping down\n",
1940 			opts->queue_size, ctrl->ctrl.maxcmd);
1941 		opts->queue_size = ctrl->ctrl.maxcmd;
1942 	}
1943 
1944 	if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
1945 		/* warn if sqsize is lower than queue_size */
1946 		dev_warn(ctrl->ctrl.device,
1947 			"queue_size %zu > ctrl sqsize %u, clamping down\n",
1948 			opts->queue_size, ctrl->ctrl.sqsize + 1);
1949 		opts->queue_size = ctrl->ctrl.sqsize + 1;
1950 	}
1951 
1952 	if (opts->nr_io_queues) {
1953 		ret = nvme_rdma_configure_io_queues(ctrl, true);
1954 		if (ret)
1955 			goto out_remove_admin_queue;
1956 	}
1957 
1958 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1959 	WARN_ON_ONCE(!changed);
1960 
1961 	dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1962 		ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1963 
1964 	kref_get(&ctrl->ctrl.kref);
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 	nvme_start_ctrl(&ctrl->ctrl);
1971 
1972 	return &ctrl->ctrl;
1973 
1974 out_remove_admin_queue:
1975 	nvme_rdma_destroy_admin_queue(ctrl, true);
1976 out_kfree_queues:
1977 	kfree(ctrl->queues);
1978 out_uninit_ctrl:
1979 	nvme_uninit_ctrl(&ctrl->ctrl);
1980 	nvme_put_ctrl(&ctrl->ctrl);
1981 	if (ret > 0)
1982 		ret = -EIO;
1983 	return ERR_PTR(ret);
1984 out_free_ctrl:
1985 	kfree(ctrl);
1986 	return ERR_PTR(ret);
1987 }
1988 
1989 static struct nvmf_transport_ops nvme_rdma_transport = {
1990 	.name		= "rdma",
1991 	.required_opts	= NVMF_OPT_TRADDR,
1992 	.allowed_opts	= NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
1993 			  NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
1994 	.create_ctrl	= nvme_rdma_create_ctrl,
1995 };
1996 
1997 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
1998 {
1999 	struct nvme_rdma_ctrl *ctrl;
2000 
2001 	/* Delete all controllers using this device */
2002 	mutex_lock(&nvme_rdma_ctrl_mutex);
2003 	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2004 		if (ctrl->device->dev != ib_device)
2005 			continue;
2006 		dev_info(ctrl->ctrl.device,
2007 			"Removing ctrl: NQN \"%s\", addr %pISp\n",
2008 			ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2009 		__nvme_rdma_del_ctrl(ctrl);
2010 	}
2011 	mutex_unlock(&nvme_rdma_ctrl_mutex);
2012 
2013 	flush_workqueue(nvme_wq);
2014 }
2015 
2016 static struct ib_client nvme_rdma_ib_client = {
2017 	.name   = "nvme_rdma",
2018 	.remove = nvme_rdma_remove_one
2019 };
2020 
2021 static int __init nvme_rdma_init_module(void)
2022 {
2023 	int ret;
2024 
2025 	ret = ib_register_client(&nvme_rdma_ib_client);
2026 	if (ret)
2027 		return ret;
2028 
2029 	ret = nvmf_register_transport(&nvme_rdma_transport);
2030 	if (ret)
2031 		goto err_unreg_client;
2032 
2033 	return 0;
2034 
2035 err_unreg_client:
2036 	ib_unregister_client(&nvme_rdma_ib_client);
2037 	return ret;
2038 }
2039 
2040 static void __exit nvme_rdma_cleanup_module(void)
2041 {
2042 	nvmf_unregister_transport(&nvme_rdma_transport);
2043 	ib_unregister_client(&nvme_rdma_ib_client);
2044 }
2045 
2046 module_init(nvme_rdma_init_module);
2047 module_exit(nvme_rdma_cleanup_module);
2048 
2049 MODULE_LICENSE("GPL v2");
2050