xref: /linux/drivers/nvme/host/rdma.c (revision 140eb5227767c6754742020a16d2691222b9c19b)
1 /*
2  * NVMe over Fabrics RDMA host code.
3  * Copyright (c) 2015-2016 HGST, a Western Digital Company.
4  *
5  * This program is free software; you can redistribute it and/or modify it
6  * under the terms and conditions of the GNU General Public License,
7  * version 2, as published by the Free Software Foundation.
8  *
9  * This program is distributed in the hope it will be useful, but WITHOUT
10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
12  * more details.
13  */
14 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 #include <linux/module.h>
16 #include <linux/init.h>
17 #include <linux/slab.h>
18 #include <rdma/mr_pool.h>
19 #include <linux/err.h>
20 #include <linux/string.h>
21 #include <linux/atomic.h>
22 #include <linux/blk-mq.h>
23 #include <linux/blk-mq-rdma.h>
24 #include <linux/types.h>
25 #include <linux/list.h>
26 #include <linux/mutex.h>
27 #include <linux/scatterlist.h>
28 #include <linux/nvme.h>
29 #include <asm/unaligned.h>
30 
31 #include <rdma/ib_verbs.h>
32 #include <rdma/rdma_cm.h>
33 #include <linux/nvme-rdma.h>
34 
35 #include "nvme.h"
36 #include "fabrics.h"
37 
38 
39 #define NVME_RDMA_CONNECT_TIMEOUT_MS	3000		/* 3 second */
40 
41 #define NVME_RDMA_MAX_SEGMENTS		256
42 
43 #define NVME_RDMA_MAX_INLINE_SEGMENTS	1
44 
45 struct nvme_rdma_device {
46 	struct ib_device	*dev;
47 	struct ib_pd		*pd;
48 	struct kref		ref;
49 	struct list_head	entry;
50 };
51 
52 struct nvme_rdma_qe {
53 	struct ib_cqe		cqe;
54 	void			*data;
55 	u64			dma;
56 };
57 
58 struct nvme_rdma_queue;
59 struct nvme_rdma_request {
60 	struct nvme_request	req;
61 	struct ib_mr		*mr;
62 	struct nvme_rdma_qe	sqe;
63 	union nvme_result	result;
64 	__le16			status;
65 	refcount_t		ref;
66 	struct ib_sge		sge[1 + NVME_RDMA_MAX_INLINE_SEGMENTS];
67 	u32			num_sge;
68 	int			nents;
69 	bool			inline_data;
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 };
122 
123 static inline struct nvme_rdma_ctrl *to_rdma_ctrl(struct nvme_ctrl *ctrl)
124 {
125 	return container_of(ctrl, struct nvme_rdma_ctrl, ctrl);
126 }
127 
128 static LIST_HEAD(device_list);
129 static DEFINE_MUTEX(device_list_mutex);
130 
131 static LIST_HEAD(nvme_rdma_ctrl_list);
132 static DEFINE_MUTEX(nvme_rdma_ctrl_mutex);
133 
134 /*
135  * Disabling this option makes small I/O goes faster, but is fundamentally
136  * unsafe.  With it turned off we will have to register a global rkey that
137  * allows read and write access to all physical memory.
138  */
139 static bool register_always = true;
140 module_param(register_always, bool, 0444);
141 MODULE_PARM_DESC(register_always,
142 	 "Use memory registration even for contiguous memory regions");
143 
144 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
145 		struct rdma_cm_event *event);
146 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc);
147 
148 static const struct blk_mq_ops nvme_rdma_mq_ops;
149 static const struct blk_mq_ops nvme_rdma_admin_mq_ops;
150 
151 /* XXX: really should move to a generic header sooner or later.. */
152 static inline void put_unaligned_le24(u32 val, u8 *p)
153 {
154 	*p++ = val;
155 	*p++ = val >> 8;
156 	*p++ = val >> 16;
157 }
158 
159 static inline int nvme_rdma_queue_idx(struct nvme_rdma_queue *queue)
160 {
161 	return queue - queue->ctrl->queues;
162 }
163 
164 static inline size_t nvme_rdma_inline_data_size(struct nvme_rdma_queue *queue)
165 {
166 	return queue->cmnd_capsule_len - sizeof(struct nvme_command);
167 }
168 
169 static void nvme_rdma_free_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
170 		size_t capsule_size, enum dma_data_direction dir)
171 {
172 	ib_dma_unmap_single(ibdev, qe->dma, capsule_size, dir);
173 	kfree(qe->data);
174 }
175 
176 static int nvme_rdma_alloc_qe(struct ib_device *ibdev, struct nvme_rdma_qe *qe,
177 		size_t capsule_size, enum dma_data_direction dir)
178 {
179 	qe->data = kzalloc(capsule_size, GFP_KERNEL);
180 	if (!qe->data)
181 		return -ENOMEM;
182 
183 	qe->dma = ib_dma_map_single(ibdev, qe->data, capsule_size, dir);
184 	if (ib_dma_mapping_error(ibdev, qe->dma)) {
185 		kfree(qe->data);
186 		return -ENOMEM;
187 	}
188 
189 	return 0;
190 }
191 
192 static void nvme_rdma_free_ring(struct ib_device *ibdev,
193 		struct nvme_rdma_qe *ring, size_t ib_queue_size,
194 		size_t capsule_size, enum dma_data_direction dir)
195 {
196 	int i;
197 
198 	for (i = 0; i < ib_queue_size; i++)
199 		nvme_rdma_free_qe(ibdev, &ring[i], capsule_size, dir);
200 	kfree(ring);
201 }
202 
203 static struct nvme_rdma_qe *nvme_rdma_alloc_ring(struct ib_device *ibdev,
204 		size_t ib_queue_size, size_t capsule_size,
205 		enum dma_data_direction dir)
206 {
207 	struct nvme_rdma_qe *ring;
208 	int i;
209 
210 	ring = kcalloc(ib_queue_size, sizeof(struct nvme_rdma_qe), GFP_KERNEL);
211 	if (!ring)
212 		return NULL;
213 
214 	for (i = 0; i < ib_queue_size; i++) {
215 		if (nvme_rdma_alloc_qe(ibdev, &ring[i], capsule_size, dir))
216 			goto out_free_ring;
217 	}
218 
219 	return ring;
220 
221 out_free_ring:
222 	nvme_rdma_free_ring(ibdev, ring, i, capsule_size, dir);
223 	return NULL;
224 }
225 
226 static void nvme_rdma_qp_event(struct ib_event *event, void *context)
227 {
228 	pr_debug("QP event %s (%d)\n",
229 		 ib_event_msg(event->event), event->event);
230 
231 }
232 
233 static int nvme_rdma_wait_for_cm(struct nvme_rdma_queue *queue)
234 {
235 	wait_for_completion_interruptible_timeout(&queue->cm_done,
236 			msecs_to_jiffies(NVME_RDMA_CONNECT_TIMEOUT_MS) + 1);
237 	return queue->cm_error;
238 }
239 
240 static int nvme_rdma_create_qp(struct nvme_rdma_queue *queue, const int factor)
241 {
242 	struct nvme_rdma_device *dev = queue->device;
243 	struct ib_qp_init_attr init_attr;
244 	int ret;
245 
246 	memset(&init_attr, 0, sizeof(init_attr));
247 	init_attr.event_handler = nvme_rdma_qp_event;
248 	/* +1 for drain */
249 	init_attr.cap.max_send_wr = factor * queue->queue_size + 1;
250 	/* +1 for drain */
251 	init_attr.cap.max_recv_wr = queue->queue_size + 1;
252 	init_attr.cap.max_recv_sge = 1;
253 	init_attr.cap.max_send_sge = 1 + NVME_RDMA_MAX_INLINE_SEGMENTS;
254 	init_attr.sq_sig_type = IB_SIGNAL_REQ_WR;
255 	init_attr.qp_type = IB_QPT_RC;
256 	init_attr.send_cq = queue->ib_cq;
257 	init_attr.recv_cq = queue->ib_cq;
258 
259 	ret = rdma_create_qp(queue->cm_id, dev->pd, &init_attr);
260 
261 	queue->qp = queue->cm_id->qp;
262 	return ret;
263 }
264 
265 static void nvme_rdma_exit_request(struct blk_mq_tag_set *set,
266 		struct request *rq, unsigned int hctx_idx)
267 {
268 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
269 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
270 	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
271 	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
272 	struct nvme_rdma_device *dev = queue->device;
273 
274 	nvme_rdma_free_qe(dev->dev, &req->sqe, sizeof(struct nvme_command),
275 			DMA_TO_DEVICE);
276 }
277 
278 static int nvme_rdma_init_request(struct blk_mq_tag_set *set,
279 		struct request *rq, unsigned int hctx_idx,
280 		unsigned int numa_node)
281 {
282 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
283 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
284 	int queue_idx = (set == &ctrl->tag_set) ? hctx_idx + 1 : 0;
285 	struct nvme_rdma_queue *queue = &ctrl->queues[queue_idx];
286 	struct nvme_rdma_device *dev = queue->device;
287 	struct ib_device *ibdev = dev->dev;
288 	int ret;
289 
290 	ret = nvme_rdma_alloc_qe(ibdev, &req->sqe, sizeof(struct nvme_command),
291 			DMA_TO_DEVICE);
292 	if (ret)
293 		return ret;
294 
295 	req->queue = queue;
296 
297 	return 0;
298 }
299 
300 static int nvme_rdma_init_hctx(struct blk_mq_hw_ctx *hctx, void *data,
301 		unsigned int hctx_idx)
302 {
303 	struct nvme_rdma_ctrl *ctrl = data;
304 	struct nvme_rdma_queue *queue = &ctrl->queues[hctx_idx + 1];
305 
306 	BUG_ON(hctx_idx >= ctrl->ctrl.queue_count);
307 
308 	hctx->driver_data = queue;
309 	return 0;
310 }
311 
312 static int nvme_rdma_init_admin_hctx(struct blk_mq_hw_ctx *hctx, void *data,
313 		unsigned int hctx_idx)
314 {
315 	struct nvme_rdma_ctrl *ctrl = data;
316 	struct nvme_rdma_queue *queue = &ctrl->queues[0];
317 
318 	BUG_ON(hctx_idx != 0);
319 
320 	hctx->driver_data = queue;
321 	return 0;
322 }
323 
324 static void nvme_rdma_free_dev(struct kref *ref)
325 {
326 	struct nvme_rdma_device *ndev =
327 		container_of(ref, struct nvme_rdma_device, ref);
328 
329 	mutex_lock(&device_list_mutex);
330 	list_del(&ndev->entry);
331 	mutex_unlock(&device_list_mutex);
332 
333 	ib_dealloc_pd(ndev->pd);
334 	kfree(ndev);
335 }
336 
337 static void nvme_rdma_dev_put(struct nvme_rdma_device *dev)
338 {
339 	kref_put(&dev->ref, nvme_rdma_free_dev);
340 }
341 
342 static int nvme_rdma_dev_get(struct nvme_rdma_device *dev)
343 {
344 	return kref_get_unless_zero(&dev->ref);
345 }
346 
347 static struct nvme_rdma_device *
348 nvme_rdma_find_get_device(struct rdma_cm_id *cm_id)
349 {
350 	struct nvme_rdma_device *ndev;
351 
352 	mutex_lock(&device_list_mutex);
353 	list_for_each_entry(ndev, &device_list, entry) {
354 		if (ndev->dev->node_guid == cm_id->device->node_guid &&
355 		    nvme_rdma_dev_get(ndev))
356 			goto out_unlock;
357 	}
358 
359 	ndev = kzalloc(sizeof(*ndev), GFP_KERNEL);
360 	if (!ndev)
361 		goto out_err;
362 
363 	ndev->dev = cm_id->device;
364 	kref_init(&ndev->ref);
365 
366 	ndev->pd = ib_alloc_pd(ndev->dev,
367 		register_always ? 0 : IB_PD_UNSAFE_GLOBAL_RKEY);
368 	if (IS_ERR(ndev->pd))
369 		goto out_free_dev;
370 
371 	if (!(ndev->dev->attrs.device_cap_flags &
372 	      IB_DEVICE_MEM_MGT_EXTENSIONS)) {
373 		dev_err(&ndev->dev->dev,
374 			"Memory registrations not supported.\n");
375 		goto out_free_pd;
376 	}
377 
378 	list_add(&ndev->entry, &device_list);
379 out_unlock:
380 	mutex_unlock(&device_list_mutex);
381 	return ndev;
382 
383 out_free_pd:
384 	ib_dealloc_pd(ndev->pd);
385 out_free_dev:
386 	kfree(ndev);
387 out_err:
388 	mutex_unlock(&device_list_mutex);
389 	return NULL;
390 }
391 
392 static void nvme_rdma_destroy_queue_ib(struct nvme_rdma_queue *queue)
393 {
394 	struct nvme_rdma_device *dev;
395 	struct ib_device *ibdev;
396 
397 	if (!test_and_clear_bit(NVME_RDMA_Q_TR_READY, &queue->flags))
398 		return;
399 
400 	dev = queue->device;
401 	ibdev = dev->dev;
402 
403 	ib_mr_pool_destroy(queue->qp, &queue->qp->rdma_mrs);
404 
405 	/*
406 	 * The cm_id object might have been destroyed during RDMA connection
407 	 * establishment error flow to avoid getting other cma events, thus
408 	 * the destruction of the QP shouldn't use rdma_cm API.
409 	 */
410 	ib_destroy_qp(queue->qp);
411 	ib_free_cq(queue->ib_cq);
412 
413 	nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
414 			sizeof(struct nvme_completion), DMA_FROM_DEVICE);
415 
416 	nvme_rdma_dev_put(dev);
417 }
418 
419 static int nvme_rdma_get_max_fr_pages(struct ib_device *ibdev)
420 {
421 	return min_t(u32, NVME_RDMA_MAX_SEGMENTS,
422 		     ibdev->attrs.max_fast_reg_page_list_len);
423 }
424 
425 static int nvme_rdma_create_queue_ib(struct nvme_rdma_queue *queue)
426 {
427 	struct ib_device *ibdev;
428 	const int send_wr_factor = 3;			/* MR, SEND, INV */
429 	const int cq_factor = send_wr_factor + 1;	/* + RECV */
430 	int comp_vector, idx = nvme_rdma_queue_idx(queue);
431 	int ret;
432 
433 	queue->device = nvme_rdma_find_get_device(queue->cm_id);
434 	if (!queue->device) {
435 		dev_err(queue->cm_id->device->dev.parent,
436 			"no client data found!\n");
437 		return -ECONNREFUSED;
438 	}
439 	ibdev = queue->device->dev;
440 
441 	/*
442 	 * Spread I/O queues completion vectors according their queue index.
443 	 * Admin queues can always go on completion vector 0.
444 	 */
445 	comp_vector = idx == 0 ? idx : idx - 1;
446 
447 	/* +1 for ib_stop_cq */
448 	queue->ib_cq = ib_alloc_cq(ibdev, queue,
449 				cq_factor * queue->queue_size + 1,
450 				comp_vector, IB_POLL_SOFTIRQ);
451 	if (IS_ERR(queue->ib_cq)) {
452 		ret = PTR_ERR(queue->ib_cq);
453 		goto out_put_dev;
454 	}
455 
456 	ret = nvme_rdma_create_qp(queue, send_wr_factor);
457 	if (ret)
458 		goto out_destroy_ib_cq;
459 
460 	queue->rsp_ring = nvme_rdma_alloc_ring(ibdev, queue->queue_size,
461 			sizeof(struct nvme_completion), DMA_FROM_DEVICE);
462 	if (!queue->rsp_ring) {
463 		ret = -ENOMEM;
464 		goto out_destroy_qp;
465 	}
466 
467 	ret = ib_mr_pool_init(queue->qp, &queue->qp->rdma_mrs,
468 			      queue->queue_size,
469 			      IB_MR_TYPE_MEM_REG,
470 			      nvme_rdma_get_max_fr_pages(ibdev));
471 	if (ret) {
472 		dev_err(queue->ctrl->ctrl.device,
473 			"failed to initialize MR pool sized %d for QID %d\n",
474 			queue->queue_size, idx);
475 		goto out_destroy_ring;
476 	}
477 
478 	set_bit(NVME_RDMA_Q_TR_READY, &queue->flags);
479 
480 	return 0;
481 
482 out_destroy_ring:
483 	nvme_rdma_free_ring(ibdev, queue->rsp_ring, queue->queue_size,
484 			    sizeof(struct nvme_completion), DMA_FROM_DEVICE);
485 out_destroy_qp:
486 	rdma_destroy_qp(queue->cm_id);
487 out_destroy_ib_cq:
488 	ib_free_cq(queue->ib_cq);
489 out_put_dev:
490 	nvme_rdma_dev_put(queue->device);
491 	return ret;
492 }
493 
494 static int nvme_rdma_alloc_queue(struct nvme_rdma_ctrl *ctrl,
495 		int idx, size_t queue_size)
496 {
497 	struct nvme_rdma_queue *queue;
498 	struct sockaddr *src_addr = NULL;
499 	int ret;
500 
501 	queue = &ctrl->queues[idx];
502 	queue->ctrl = ctrl;
503 	init_completion(&queue->cm_done);
504 
505 	if (idx > 0)
506 		queue->cmnd_capsule_len = ctrl->ctrl.ioccsz * 16;
507 	else
508 		queue->cmnd_capsule_len = sizeof(struct nvme_command);
509 
510 	queue->queue_size = queue_size;
511 
512 	queue->cm_id = rdma_create_id(&init_net, nvme_rdma_cm_handler, queue,
513 			RDMA_PS_TCP, IB_QPT_RC);
514 	if (IS_ERR(queue->cm_id)) {
515 		dev_info(ctrl->ctrl.device,
516 			"failed to create CM ID: %ld\n", PTR_ERR(queue->cm_id));
517 		return PTR_ERR(queue->cm_id);
518 	}
519 
520 	if (ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
521 		src_addr = (struct sockaddr *)&ctrl->src_addr;
522 
523 	queue->cm_error = -ETIMEDOUT;
524 	ret = rdma_resolve_addr(queue->cm_id, src_addr,
525 			(struct sockaddr *)&ctrl->addr,
526 			NVME_RDMA_CONNECT_TIMEOUT_MS);
527 	if (ret) {
528 		dev_info(ctrl->ctrl.device,
529 			"rdma_resolve_addr failed (%d).\n", ret);
530 		goto out_destroy_cm_id;
531 	}
532 
533 	ret = nvme_rdma_wait_for_cm(queue);
534 	if (ret) {
535 		dev_info(ctrl->ctrl.device,
536 			"rdma connection establishment failed (%d)\n", ret);
537 		goto out_destroy_cm_id;
538 	}
539 
540 	set_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags);
541 
542 	return 0;
543 
544 out_destroy_cm_id:
545 	rdma_destroy_id(queue->cm_id);
546 	nvme_rdma_destroy_queue_ib(queue);
547 	return ret;
548 }
549 
550 static void nvme_rdma_stop_queue(struct nvme_rdma_queue *queue)
551 {
552 	if (!test_and_clear_bit(NVME_RDMA_Q_LIVE, &queue->flags))
553 		return;
554 
555 	rdma_disconnect(queue->cm_id);
556 	ib_drain_qp(queue->qp);
557 }
558 
559 static void nvme_rdma_free_queue(struct nvme_rdma_queue *queue)
560 {
561 	if (!test_and_clear_bit(NVME_RDMA_Q_ALLOCATED, &queue->flags))
562 		return;
563 
564 	if (nvme_rdma_queue_idx(queue) == 0) {
565 		nvme_rdma_free_qe(queue->device->dev,
566 			&queue->ctrl->async_event_sqe,
567 			sizeof(struct nvme_command), DMA_TO_DEVICE);
568 	}
569 
570 	nvme_rdma_destroy_queue_ib(queue);
571 	rdma_destroy_id(queue->cm_id);
572 }
573 
574 static void nvme_rdma_free_io_queues(struct nvme_rdma_ctrl *ctrl)
575 {
576 	int i;
577 
578 	for (i = 1; i < ctrl->ctrl.queue_count; i++)
579 		nvme_rdma_free_queue(&ctrl->queues[i]);
580 }
581 
582 static void nvme_rdma_stop_io_queues(struct nvme_rdma_ctrl *ctrl)
583 {
584 	int i;
585 
586 	for (i = 1; i < ctrl->ctrl.queue_count; i++)
587 		nvme_rdma_stop_queue(&ctrl->queues[i]);
588 }
589 
590 static int nvme_rdma_start_queue(struct nvme_rdma_ctrl *ctrl, int idx)
591 {
592 	int ret;
593 
594 	if (idx)
595 		ret = nvmf_connect_io_queue(&ctrl->ctrl, idx);
596 	else
597 		ret = nvmf_connect_admin_queue(&ctrl->ctrl);
598 
599 	if (!ret)
600 		set_bit(NVME_RDMA_Q_LIVE, &ctrl->queues[idx].flags);
601 	else
602 		dev_info(ctrl->ctrl.device,
603 			"failed to connect queue: %d ret=%d\n", idx, ret);
604 	return ret;
605 }
606 
607 static int nvme_rdma_start_io_queues(struct nvme_rdma_ctrl *ctrl)
608 {
609 	int i, ret = 0;
610 
611 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
612 		ret = nvme_rdma_start_queue(ctrl, i);
613 		if (ret)
614 			goto out_stop_queues;
615 	}
616 
617 	return 0;
618 
619 out_stop_queues:
620 	for (i--; i >= 1; i--)
621 		nvme_rdma_stop_queue(&ctrl->queues[i]);
622 	return ret;
623 }
624 
625 static int nvme_rdma_alloc_io_queues(struct nvme_rdma_ctrl *ctrl)
626 {
627 	struct nvmf_ctrl_options *opts = ctrl->ctrl.opts;
628 	struct ib_device *ibdev = ctrl->device->dev;
629 	unsigned int nr_io_queues;
630 	int i, ret;
631 
632 	nr_io_queues = min(opts->nr_io_queues, num_online_cpus());
633 
634 	/*
635 	 * we map queues according to the device irq vectors for
636 	 * optimal locality so we don't need more queues than
637 	 * completion vectors.
638 	 */
639 	nr_io_queues = min_t(unsigned int, nr_io_queues,
640 				ibdev->num_comp_vectors);
641 
642 	ret = nvme_set_queue_count(&ctrl->ctrl, &nr_io_queues);
643 	if (ret)
644 		return ret;
645 
646 	ctrl->ctrl.queue_count = nr_io_queues + 1;
647 	if (ctrl->ctrl.queue_count < 2)
648 		return 0;
649 
650 	dev_info(ctrl->ctrl.device,
651 		"creating %d I/O queues.\n", nr_io_queues);
652 
653 	for (i = 1; i < ctrl->ctrl.queue_count; i++) {
654 		ret = nvme_rdma_alloc_queue(ctrl, i,
655 				ctrl->ctrl.sqsize + 1);
656 		if (ret)
657 			goto out_free_queues;
658 	}
659 
660 	return 0;
661 
662 out_free_queues:
663 	for (i--; i >= 1; i--)
664 		nvme_rdma_free_queue(&ctrl->queues[i]);
665 
666 	return ret;
667 }
668 
669 static void nvme_rdma_free_tagset(struct nvme_ctrl *nctrl,
670 		struct blk_mq_tag_set *set)
671 {
672 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
673 
674 	blk_mq_free_tag_set(set);
675 	nvme_rdma_dev_put(ctrl->device);
676 }
677 
678 static struct blk_mq_tag_set *nvme_rdma_alloc_tagset(struct nvme_ctrl *nctrl,
679 		bool admin)
680 {
681 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
682 	struct blk_mq_tag_set *set;
683 	int ret;
684 
685 	if (admin) {
686 		set = &ctrl->admin_tag_set;
687 		memset(set, 0, sizeof(*set));
688 		set->ops = &nvme_rdma_admin_mq_ops;
689 		set->queue_depth = NVME_AQ_MQ_TAG_DEPTH;
690 		set->reserved_tags = 2; /* connect + keep-alive */
691 		set->numa_node = NUMA_NO_NODE;
692 		set->cmd_size = sizeof(struct nvme_rdma_request) +
693 			SG_CHUNK_SIZE * sizeof(struct scatterlist);
694 		set->driver_data = ctrl;
695 		set->nr_hw_queues = 1;
696 		set->timeout = ADMIN_TIMEOUT;
697 		set->flags = BLK_MQ_F_NO_SCHED;
698 	} else {
699 		set = &ctrl->tag_set;
700 		memset(set, 0, sizeof(*set));
701 		set->ops = &nvme_rdma_mq_ops;
702 		set->queue_depth = nctrl->opts->queue_size;
703 		set->reserved_tags = 1; /* fabric connect */
704 		set->numa_node = NUMA_NO_NODE;
705 		set->flags = BLK_MQ_F_SHOULD_MERGE;
706 		set->cmd_size = sizeof(struct nvme_rdma_request) +
707 			SG_CHUNK_SIZE * sizeof(struct scatterlist);
708 		set->driver_data = ctrl;
709 		set->nr_hw_queues = nctrl->queue_count - 1;
710 		set->timeout = NVME_IO_TIMEOUT;
711 	}
712 
713 	ret = blk_mq_alloc_tag_set(set);
714 	if (ret)
715 		goto out;
716 
717 	/*
718 	 * We need a reference on the device as long as the tag_set is alive,
719 	 * as the MRs in the request structures need a valid ib_device.
720 	 */
721 	ret = nvme_rdma_dev_get(ctrl->device);
722 	if (!ret) {
723 		ret = -EINVAL;
724 		goto out_free_tagset;
725 	}
726 
727 	return set;
728 
729 out_free_tagset:
730 	blk_mq_free_tag_set(set);
731 out:
732 	return ERR_PTR(ret);
733 }
734 
735 static void nvme_rdma_destroy_admin_queue(struct nvme_rdma_ctrl *ctrl,
736 		bool remove)
737 {
738 	nvme_rdma_stop_queue(&ctrl->queues[0]);
739 	if (remove) {
740 		blk_cleanup_queue(ctrl->ctrl.admin_q);
741 		nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
742 	}
743 	nvme_rdma_free_queue(&ctrl->queues[0]);
744 }
745 
746 static int nvme_rdma_configure_admin_queue(struct nvme_rdma_ctrl *ctrl,
747 		bool new)
748 {
749 	int error;
750 
751 	error = nvme_rdma_alloc_queue(ctrl, 0, NVME_AQ_DEPTH);
752 	if (error)
753 		return error;
754 
755 	ctrl->device = ctrl->queues[0].device;
756 
757 	ctrl->max_fr_pages = nvme_rdma_get_max_fr_pages(ctrl->device->dev);
758 
759 	if (new) {
760 		ctrl->ctrl.admin_tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, true);
761 		if (IS_ERR(ctrl->ctrl.admin_tagset)) {
762 			error = PTR_ERR(ctrl->ctrl.admin_tagset);
763 			goto out_free_queue;
764 		}
765 
766 		ctrl->ctrl.admin_q = blk_mq_init_queue(&ctrl->admin_tag_set);
767 		if (IS_ERR(ctrl->ctrl.admin_q)) {
768 			error = PTR_ERR(ctrl->ctrl.admin_q);
769 			goto out_free_tagset;
770 		}
771 	}
772 
773 	error = nvme_rdma_start_queue(ctrl, 0);
774 	if (error)
775 		goto out_cleanup_queue;
776 
777 	error = ctrl->ctrl.ops->reg_read64(&ctrl->ctrl, NVME_REG_CAP,
778 			&ctrl->ctrl.cap);
779 	if (error) {
780 		dev_err(ctrl->ctrl.device,
781 			"prop_get NVME_REG_CAP failed\n");
782 		goto out_cleanup_queue;
783 	}
784 
785 	ctrl->ctrl.sqsize =
786 		min_t(int, NVME_CAP_MQES(ctrl->ctrl.cap), ctrl->ctrl.sqsize);
787 
788 	error = nvme_enable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
789 	if (error)
790 		goto out_cleanup_queue;
791 
792 	ctrl->ctrl.max_hw_sectors =
793 		(ctrl->max_fr_pages - 1) << (ilog2(SZ_4K) - 9);
794 
795 	error = nvme_init_identify(&ctrl->ctrl);
796 	if (error)
797 		goto out_cleanup_queue;
798 
799 	error = nvme_rdma_alloc_qe(ctrl->queues[0].device->dev,
800 			&ctrl->async_event_sqe, sizeof(struct nvme_command),
801 			DMA_TO_DEVICE);
802 	if (error)
803 		goto out_cleanup_queue;
804 
805 	return 0;
806 
807 out_cleanup_queue:
808 	if (new)
809 		blk_cleanup_queue(ctrl->ctrl.admin_q);
810 out_free_tagset:
811 	if (new)
812 		nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.admin_tagset);
813 out_free_queue:
814 	nvme_rdma_free_queue(&ctrl->queues[0]);
815 	return error;
816 }
817 
818 static void nvme_rdma_destroy_io_queues(struct nvme_rdma_ctrl *ctrl,
819 		bool remove)
820 {
821 	nvme_rdma_stop_io_queues(ctrl);
822 	if (remove) {
823 		blk_cleanup_queue(ctrl->ctrl.connect_q);
824 		nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
825 	}
826 	nvme_rdma_free_io_queues(ctrl);
827 }
828 
829 static int nvme_rdma_configure_io_queues(struct nvme_rdma_ctrl *ctrl, bool new)
830 {
831 	int ret;
832 
833 	ret = nvme_rdma_alloc_io_queues(ctrl);
834 	if (ret)
835 		return ret;
836 
837 	if (new) {
838 		ctrl->ctrl.tagset = nvme_rdma_alloc_tagset(&ctrl->ctrl, false);
839 		if (IS_ERR(ctrl->ctrl.tagset)) {
840 			ret = PTR_ERR(ctrl->ctrl.tagset);
841 			goto out_free_io_queues;
842 		}
843 
844 		ctrl->ctrl.connect_q = blk_mq_init_queue(&ctrl->tag_set);
845 		if (IS_ERR(ctrl->ctrl.connect_q)) {
846 			ret = PTR_ERR(ctrl->ctrl.connect_q);
847 			goto out_free_tag_set;
848 		}
849 	} else {
850 		blk_mq_update_nr_hw_queues(&ctrl->tag_set,
851 			ctrl->ctrl.queue_count - 1);
852 	}
853 
854 	ret = nvme_rdma_start_io_queues(ctrl);
855 	if (ret)
856 		goto out_cleanup_connect_q;
857 
858 	return 0;
859 
860 out_cleanup_connect_q:
861 	if (new)
862 		blk_cleanup_queue(ctrl->ctrl.connect_q);
863 out_free_tag_set:
864 	if (new)
865 		nvme_rdma_free_tagset(&ctrl->ctrl, ctrl->ctrl.tagset);
866 out_free_io_queues:
867 	nvme_rdma_free_io_queues(ctrl);
868 	return ret;
869 }
870 
871 static void nvme_rdma_free_ctrl(struct nvme_ctrl *nctrl)
872 {
873 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(nctrl);
874 
875 	if (list_empty(&ctrl->list))
876 		goto free_ctrl;
877 
878 	mutex_lock(&nvme_rdma_ctrl_mutex);
879 	list_del(&ctrl->list);
880 	mutex_unlock(&nvme_rdma_ctrl_mutex);
881 
882 	kfree(ctrl->queues);
883 	nvmf_free_options(nctrl->opts);
884 free_ctrl:
885 	kfree(ctrl);
886 }
887 
888 static void nvme_rdma_reconnect_or_remove(struct nvme_rdma_ctrl *ctrl)
889 {
890 	/* If we are resetting/deleting then do nothing */
891 	if (ctrl->ctrl.state != NVME_CTRL_RECONNECTING) {
892 		WARN_ON_ONCE(ctrl->ctrl.state == NVME_CTRL_NEW ||
893 			ctrl->ctrl.state == NVME_CTRL_LIVE);
894 		return;
895 	}
896 
897 	if (nvmf_should_reconnect(&ctrl->ctrl)) {
898 		dev_info(ctrl->ctrl.device, "Reconnecting in %d seconds...\n",
899 			ctrl->ctrl.opts->reconnect_delay);
900 		queue_delayed_work(nvme_wq, &ctrl->reconnect_work,
901 				ctrl->ctrl.opts->reconnect_delay * HZ);
902 	} else {
903 		dev_info(ctrl->ctrl.device, "Removing controller...\n");
904 		nvme_delete_ctrl(&ctrl->ctrl);
905 	}
906 }
907 
908 static void nvme_rdma_reconnect_ctrl_work(struct work_struct *work)
909 {
910 	struct nvme_rdma_ctrl *ctrl = container_of(to_delayed_work(work),
911 			struct nvme_rdma_ctrl, reconnect_work);
912 	bool changed;
913 	int ret;
914 
915 	++ctrl->ctrl.nr_reconnects;
916 
917 	ret = nvme_rdma_configure_admin_queue(ctrl, false);
918 	if (ret)
919 		goto requeue;
920 
921 	if (ctrl->ctrl.queue_count > 1) {
922 		ret = nvme_rdma_configure_io_queues(ctrl, false);
923 		if (ret)
924 			goto destroy_admin;
925 	}
926 
927 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
928 	if (!changed) {
929 		/* state change failure is ok if we're in DELETING state */
930 		WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
931 		return;
932 	}
933 
934 	nvme_start_ctrl(&ctrl->ctrl);
935 
936 	dev_info(ctrl->ctrl.device, "Successfully reconnected (%d attempts)\n",
937 			ctrl->ctrl.nr_reconnects);
938 
939 	ctrl->ctrl.nr_reconnects = 0;
940 
941 	return;
942 
943 destroy_admin:
944 	nvme_rdma_destroy_admin_queue(ctrl, false);
945 requeue:
946 	dev_info(ctrl->ctrl.device, "Failed reconnect attempt %d\n",
947 			ctrl->ctrl.nr_reconnects);
948 	nvme_rdma_reconnect_or_remove(ctrl);
949 }
950 
951 static void nvme_rdma_error_recovery_work(struct work_struct *work)
952 {
953 	struct nvme_rdma_ctrl *ctrl = container_of(work,
954 			struct nvme_rdma_ctrl, err_work);
955 
956 	nvme_stop_keep_alive(&ctrl->ctrl);
957 
958 	if (ctrl->ctrl.queue_count > 1) {
959 		nvme_stop_queues(&ctrl->ctrl);
960 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
961 					nvme_cancel_request, &ctrl->ctrl);
962 		nvme_rdma_destroy_io_queues(ctrl, false);
963 	}
964 
965 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
966 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
967 				nvme_cancel_request, &ctrl->ctrl);
968 	nvme_rdma_destroy_admin_queue(ctrl, false);
969 
970 	/*
971 	 * queues are not a live anymore, so restart the queues to fail fast
972 	 * new IO
973 	 */
974 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
975 	nvme_start_queues(&ctrl->ctrl);
976 
977 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING)) {
978 		/* state change failure should never happen */
979 		WARN_ON_ONCE(1);
980 		return;
981 	}
982 
983 	nvme_rdma_reconnect_or_remove(ctrl);
984 }
985 
986 static void nvme_rdma_error_recovery(struct nvme_rdma_ctrl *ctrl)
987 {
988 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RESETTING))
989 		return;
990 
991 	queue_work(nvme_wq, &ctrl->err_work);
992 }
993 
994 static void nvme_rdma_wr_error(struct ib_cq *cq, struct ib_wc *wc,
995 		const char *op)
996 {
997 	struct nvme_rdma_queue *queue = cq->cq_context;
998 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
999 
1000 	if (ctrl->ctrl.state == NVME_CTRL_LIVE)
1001 		dev_info(ctrl->ctrl.device,
1002 			     "%s for CQE 0x%p failed with status %s (%d)\n",
1003 			     op, wc->wr_cqe,
1004 			     ib_wc_status_msg(wc->status), wc->status);
1005 	nvme_rdma_error_recovery(ctrl);
1006 }
1007 
1008 static void nvme_rdma_memreg_done(struct ib_cq *cq, struct ib_wc *wc)
1009 {
1010 	if (unlikely(wc->status != IB_WC_SUCCESS))
1011 		nvme_rdma_wr_error(cq, wc, "MEMREG");
1012 }
1013 
1014 static void nvme_rdma_inv_rkey_done(struct ib_cq *cq, struct ib_wc *wc)
1015 {
1016 	struct nvme_rdma_request *req =
1017 		container_of(wc->wr_cqe, struct nvme_rdma_request, reg_cqe);
1018 	struct request *rq = blk_mq_rq_from_pdu(req);
1019 
1020 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1021 		nvme_rdma_wr_error(cq, wc, "LOCAL_INV");
1022 		return;
1023 	}
1024 
1025 	if (refcount_dec_and_test(&req->ref))
1026 		nvme_end_request(rq, req->status, req->result);
1027 
1028 }
1029 
1030 static int nvme_rdma_inv_rkey(struct nvme_rdma_queue *queue,
1031 		struct nvme_rdma_request *req)
1032 {
1033 	struct ib_send_wr *bad_wr;
1034 	struct ib_send_wr wr = {
1035 		.opcode		    = IB_WR_LOCAL_INV,
1036 		.next		    = NULL,
1037 		.num_sge	    = 0,
1038 		.send_flags	    = IB_SEND_SIGNALED,
1039 		.ex.invalidate_rkey = req->mr->rkey,
1040 	};
1041 
1042 	req->reg_cqe.done = nvme_rdma_inv_rkey_done;
1043 	wr.wr_cqe = &req->reg_cqe;
1044 
1045 	return ib_post_send(queue->qp, &wr, &bad_wr);
1046 }
1047 
1048 static void nvme_rdma_unmap_data(struct nvme_rdma_queue *queue,
1049 		struct request *rq)
1050 {
1051 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1052 	struct nvme_rdma_device *dev = queue->device;
1053 	struct ib_device *ibdev = dev->dev;
1054 
1055 	if (!blk_rq_bytes(rq))
1056 		return;
1057 
1058 	if (req->mr) {
1059 		ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1060 		req->mr = NULL;
1061 	}
1062 
1063 	ib_dma_unmap_sg(ibdev, req->sg_table.sgl,
1064 			req->nents, rq_data_dir(rq) ==
1065 				    WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1066 
1067 	nvme_cleanup_cmd(rq);
1068 	sg_free_table_chained(&req->sg_table, true);
1069 }
1070 
1071 static int nvme_rdma_set_sg_null(struct nvme_command *c)
1072 {
1073 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1074 
1075 	sg->addr = 0;
1076 	put_unaligned_le24(0, sg->length);
1077 	put_unaligned_le32(0, sg->key);
1078 	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1079 	return 0;
1080 }
1081 
1082 static int nvme_rdma_map_sg_inline(struct nvme_rdma_queue *queue,
1083 		struct nvme_rdma_request *req, struct nvme_command *c)
1084 {
1085 	struct nvme_sgl_desc *sg = &c->common.dptr.sgl;
1086 
1087 	req->sge[1].addr = sg_dma_address(req->sg_table.sgl);
1088 	req->sge[1].length = sg_dma_len(req->sg_table.sgl);
1089 	req->sge[1].lkey = queue->device->pd->local_dma_lkey;
1090 
1091 	sg->addr = cpu_to_le64(queue->ctrl->ctrl.icdoff);
1092 	sg->length = cpu_to_le32(sg_dma_len(req->sg_table.sgl));
1093 	sg->type = (NVME_SGL_FMT_DATA_DESC << 4) | NVME_SGL_FMT_OFFSET;
1094 
1095 	req->inline_data = true;
1096 	req->num_sge++;
1097 	return 0;
1098 }
1099 
1100 static int nvme_rdma_map_sg_single(struct nvme_rdma_queue *queue,
1101 		struct nvme_rdma_request *req, struct nvme_command *c)
1102 {
1103 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1104 
1105 	sg->addr = cpu_to_le64(sg_dma_address(req->sg_table.sgl));
1106 	put_unaligned_le24(sg_dma_len(req->sg_table.sgl), sg->length);
1107 	put_unaligned_le32(queue->device->pd->unsafe_global_rkey, sg->key);
1108 	sg->type = NVME_KEY_SGL_FMT_DATA_DESC << 4;
1109 	return 0;
1110 }
1111 
1112 static int nvme_rdma_map_sg_fr(struct nvme_rdma_queue *queue,
1113 		struct nvme_rdma_request *req, struct nvme_command *c,
1114 		int count)
1115 {
1116 	struct nvme_keyed_sgl_desc *sg = &c->common.dptr.ksgl;
1117 	int nr;
1118 
1119 	req->mr = ib_mr_pool_get(queue->qp, &queue->qp->rdma_mrs);
1120 	if (WARN_ON_ONCE(!req->mr))
1121 		return -EAGAIN;
1122 
1123 	/*
1124 	 * Align the MR to a 4K page size to match the ctrl page size and
1125 	 * the block virtual boundary.
1126 	 */
1127 	nr = ib_map_mr_sg(req->mr, req->sg_table.sgl, count, NULL, SZ_4K);
1128 	if (unlikely(nr < count)) {
1129 		ib_mr_pool_put(queue->qp, &queue->qp->rdma_mrs, req->mr);
1130 		req->mr = NULL;
1131 		if (nr < 0)
1132 			return nr;
1133 		return -EINVAL;
1134 	}
1135 
1136 	ib_update_fast_reg_key(req->mr, ib_inc_rkey(req->mr->rkey));
1137 
1138 	req->reg_cqe.done = nvme_rdma_memreg_done;
1139 	memset(&req->reg_wr, 0, sizeof(req->reg_wr));
1140 	req->reg_wr.wr.opcode = IB_WR_REG_MR;
1141 	req->reg_wr.wr.wr_cqe = &req->reg_cqe;
1142 	req->reg_wr.wr.num_sge = 0;
1143 	req->reg_wr.mr = req->mr;
1144 	req->reg_wr.key = req->mr->rkey;
1145 	req->reg_wr.access = IB_ACCESS_LOCAL_WRITE |
1146 			     IB_ACCESS_REMOTE_READ |
1147 			     IB_ACCESS_REMOTE_WRITE;
1148 
1149 	sg->addr = cpu_to_le64(req->mr->iova);
1150 	put_unaligned_le24(req->mr->length, sg->length);
1151 	put_unaligned_le32(req->mr->rkey, sg->key);
1152 	sg->type = (NVME_KEY_SGL_FMT_DATA_DESC << 4) |
1153 			NVME_SGL_FMT_INVALIDATE;
1154 
1155 	return 0;
1156 }
1157 
1158 static int nvme_rdma_map_data(struct nvme_rdma_queue *queue,
1159 		struct request *rq, struct nvme_command *c)
1160 {
1161 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1162 	struct nvme_rdma_device *dev = queue->device;
1163 	struct ib_device *ibdev = dev->dev;
1164 	int count, ret;
1165 
1166 	req->num_sge = 1;
1167 	req->inline_data = false;
1168 	refcount_set(&req->ref, 2); /* send and recv completions */
1169 
1170 	c->common.flags |= NVME_CMD_SGL_METABUF;
1171 
1172 	if (!blk_rq_bytes(rq))
1173 		return nvme_rdma_set_sg_null(c);
1174 
1175 	req->sg_table.sgl = req->first_sgl;
1176 	ret = sg_alloc_table_chained(&req->sg_table,
1177 			blk_rq_nr_phys_segments(rq), req->sg_table.sgl);
1178 	if (ret)
1179 		return -ENOMEM;
1180 
1181 	req->nents = blk_rq_map_sg(rq->q, rq, req->sg_table.sgl);
1182 
1183 	count = ib_dma_map_sg(ibdev, req->sg_table.sgl, req->nents,
1184 		    rq_data_dir(rq) == WRITE ? DMA_TO_DEVICE : DMA_FROM_DEVICE);
1185 	if (unlikely(count <= 0)) {
1186 		sg_free_table_chained(&req->sg_table, true);
1187 		return -EIO;
1188 	}
1189 
1190 	if (count == 1) {
1191 		if (rq_data_dir(rq) == WRITE && nvme_rdma_queue_idx(queue) &&
1192 		    blk_rq_payload_bytes(rq) <=
1193 				nvme_rdma_inline_data_size(queue))
1194 			return nvme_rdma_map_sg_inline(queue, req, c);
1195 
1196 		if (dev->pd->flags & IB_PD_UNSAFE_GLOBAL_RKEY)
1197 			return nvme_rdma_map_sg_single(queue, req, c);
1198 	}
1199 
1200 	return nvme_rdma_map_sg_fr(queue, req, c, count);
1201 }
1202 
1203 static void nvme_rdma_send_done(struct ib_cq *cq, struct ib_wc *wc)
1204 {
1205 	struct nvme_rdma_qe *qe =
1206 		container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1207 	struct nvme_rdma_request *req =
1208 		container_of(qe, struct nvme_rdma_request, sqe);
1209 	struct request *rq = blk_mq_rq_from_pdu(req);
1210 
1211 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1212 		nvme_rdma_wr_error(cq, wc, "SEND");
1213 		return;
1214 	}
1215 
1216 	if (refcount_dec_and_test(&req->ref))
1217 		nvme_end_request(rq, req->status, req->result);
1218 }
1219 
1220 static int nvme_rdma_post_send(struct nvme_rdma_queue *queue,
1221 		struct nvme_rdma_qe *qe, struct ib_sge *sge, u32 num_sge,
1222 		struct ib_send_wr *first)
1223 {
1224 	struct ib_send_wr wr, *bad_wr;
1225 	int ret;
1226 
1227 	sge->addr   = qe->dma;
1228 	sge->length = sizeof(struct nvme_command),
1229 	sge->lkey   = queue->device->pd->local_dma_lkey;
1230 
1231 	wr.next       = NULL;
1232 	wr.wr_cqe     = &qe->cqe;
1233 	wr.sg_list    = sge;
1234 	wr.num_sge    = num_sge;
1235 	wr.opcode     = IB_WR_SEND;
1236 	wr.send_flags = IB_SEND_SIGNALED;
1237 
1238 	if (first)
1239 		first->next = &wr;
1240 	else
1241 		first = &wr;
1242 
1243 	ret = ib_post_send(queue->qp, first, &bad_wr);
1244 	if (unlikely(ret)) {
1245 		dev_err(queue->ctrl->ctrl.device,
1246 			     "%s failed with error code %d\n", __func__, ret);
1247 	}
1248 	return ret;
1249 }
1250 
1251 static int nvme_rdma_post_recv(struct nvme_rdma_queue *queue,
1252 		struct nvme_rdma_qe *qe)
1253 {
1254 	struct ib_recv_wr wr, *bad_wr;
1255 	struct ib_sge list;
1256 	int ret;
1257 
1258 	list.addr   = qe->dma;
1259 	list.length = sizeof(struct nvme_completion);
1260 	list.lkey   = queue->device->pd->local_dma_lkey;
1261 
1262 	qe->cqe.done = nvme_rdma_recv_done;
1263 
1264 	wr.next     = NULL;
1265 	wr.wr_cqe   = &qe->cqe;
1266 	wr.sg_list  = &list;
1267 	wr.num_sge  = 1;
1268 
1269 	ret = ib_post_recv(queue->qp, &wr, &bad_wr);
1270 	if (unlikely(ret)) {
1271 		dev_err(queue->ctrl->ctrl.device,
1272 			"%s failed with error code %d\n", __func__, ret);
1273 	}
1274 	return ret;
1275 }
1276 
1277 static struct blk_mq_tags *nvme_rdma_tagset(struct nvme_rdma_queue *queue)
1278 {
1279 	u32 queue_idx = nvme_rdma_queue_idx(queue);
1280 
1281 	if (queue_idx == 0)
1282 		return queue->ctrl->admin_tag_set.tags[queue_idx];
1283 	return queue->ctrl->tag_set.tags[queue_idx - 1];
1284 }
1285 
1286 static void nvme_rdma_async_done(struct ib_cq *cq, struct ib_wc *wc)
1287 {
1288 	if (unlikely(wc->status != IB_WC_SUCCESS))
1289 		nvme_rdma_wr_error(cq, wc, "ASYNC");
1290 }
1291 
1292 static void nvme_rdma_submit_async_event(struct nvme_ctrl *arg)
1293 {
1294 	struct nvme_rdma_ctrl *ctrl = to_rdma_ctrl(arg);
1295 	struct nvme_rdma_queue *queue = &ctrl->queues[0];
1296 	struct ib_device *dev = queue->device->dev;
1297 	struct nvme_rdma_qe *sqe = &ctrl->async_event_sqe;
1298 	struct nvme_command *cmd = sqe->data;
1299 	struct ib_sge sge;
1300 	int ret;
1301 
1302 	ib_dma_sync_single_for_cpu(dev, sqe->dma, sizeof(*cmd), DMA_TO_DEVICE);
1303 
1304 	memset(cmd, 0, sizeof(*cmd));
1305 	cmd->common.opcode = nvme_admin_async_event;
1306 	cmd->common.command_id = NVME_AQ_BLK_MQ_DEPTH;
1307 	cmd->common.flags |= NVME_CMD_SGL_METABUF;
1308 	nvme_rdma_set_sg_null(cmd);
1309 
1310 	sqe->cqe.done = nvme_rdma_async_done;
1311 
1312 	ib_dma_sync_single_for_device(dev, sqe->dma, sizeof(*cmd),
1313 			DMA_TO_DEVICE);
1314 
1315 	ret = nvme_rdma_post_send(queue, sqe, &sge, 1, NULL);
1316 	WARN_ON_ONCE(ret);
1317 }
1318 
1319 static int nvme_rdma_process_nvme_rsp(struct nvme_rdma_queue *queue,
1320 		struct nvme_completion *cqe, struct ib_wc *wc, int tag)
1321 {
1322 	struct request *rq;
1323 	struct nvme_rdma_request *req;
1324 	int ret = 0;
1325 
1326 	rq = blk_mq_tag_to_rq(nvme_rdma_tagset(queue), cqe->command_id);
1327 	if (!rq) {
1328 		dev_err(queue->ctrl->ctrl.device,
1329 			"tag 0x%x on QP %#x not found\n",
1330 			cqe->command_id, queue->qp->qp_num);
1331 		nvme_rdma_error_recovery(queue->ctrl);
1332 		return ret;
1333 	}
1334 	req = blk_mq_rq_to_pdu(rq);
1335 
1336 	req->status = cqe->status;
1337 	req->result = cqe->result;
1338 
1339 	if (wc->wc_flags & IB_WC_WITH_INVALIDATE) {
1340 		if (unlikely(wc->ex.invalidate_rkey != req->mr->rkey)) {
1341 			dev_err(queue->ctrl->ctrl.device,
1342 				"Bogus remote invalidation for rkey %#x\n",
1343 				req->mr->rkey);
1344 			nvme_rdma_error_recovery(queue->ctrl);
1345 		}
1346 	} else if (req->mr) {
1347 		ret = nvme_rdma_inv_rkey(queue, req);
1348 		if (unlikely(ret < 0)) {
1349 			dev_err(queue->ctrl->ctrl.device,
1350 				"Queueing INV WR for rkey %#x failed (%d)\n",
1351 				req->mr->rkey, ret);
1352 			nvme_rdma_error_recovery(queue->ctrl);
1353 		}
1354 		/* the local invalidation completion will end the request */
1355 		return 0;
1356 	}
1357 
1358 	if (refcount_dec_and_test(&req->ref)) {
1359 		if (rq->tag == tag)
1360 			ret = 1;
1361 		nvme_end_request(rq, req->status, req->result);
1362 	}
1363 
1364 	return ret;
1365 }
1366 
1367 static int __nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc, int tag)
1368 {
1369 	struct nvme_rdma_qe *qe =
1370 		container_of(wc->wr_cqe, struct nvme_rdma_qe, cqe);
1371 	struct nvme_rdma_queue *queue = cq->cq_context;
1372 	struct ib_device *ibdev = queue->device->dev;
1373 	struct nvme_completion *cqe = qe->data;
1374 	const size_t len = sizeof(struct nvme_completion);
1375 	int ret = 0;
1376 
1377 	if (unlikely(wc->status != IB_WC_SUCCESS)) {
1378 		nvme_rdma_wr_error(cq, wc, "RECV");
1379 		return 0;
1380 	}
1381 
1382 	ib_dma_sync_single_for_cpu(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1383 	/*
1384 	 * AEN requests are special as they don't time out and can
1385 	 * survive any kind of queue freeze and often don't respond to
1386 	 * aborts.  We don't even bother to allocate a struct request
1387 	 * for them but rather special case them here.
1388 	 */
1389 	if (unlikely(nvme_rdma_queue_idx(queue) == 0 &&
1390 			cqe->command_id >= NVME_AQ_BLK_MQ_DEPTH))
1391 		nvme_complete_async_event(&queue->ctrl->ctrl, cqe->status,
1392 				&cqe->result);
1393 	else
1394 		ret = nvme_rdma_process_nvme_rsp(queue, cqe, wc, tag);
1395 	ib_dma_sync_single_for_device(ibdev, qe->dma, len, DMA_FROM_DEVICE);
1396 
1397 	nvme_rdma_post_recv(queue, qe);
1398 	return ret;
1399 }
1400 
1401 static void nvme_rdma_recv_done(struct ib_cq *cq, struct ib_wc *wc)
1402 {
1403 	__nvme_rdma_recv_done(cq, wc, -1);
1404 }
1405 
1406 static int nvme_rdma_conn_established(struct nvme_rdma_queue *queue)
1407 {
1408 	int ret, i;
1409 
1410 	for (i = 0; i < queue->queue_size; i++) {
1411 		ret = nvme_rdma_post_recv(queue, &queue->rsp_ring[i]);
1412 		if (ret)
1413 			goto out_destroy_queue_ib;
1414 	}
1415 
1416 	return 0;
1417 
1418 out_destroy_queue_ib:
1419 	nvme_rdma_destroy_queue_ib(queue);
1420 	return ret;
1421 }
1422 
1423 static int nvme_rdma_conn_rejected(struct nvme_rdma_queue *queue,
1424 		struct rdma_cm_event *ev)
1425 {
1426 	struct rdma_cm_id *cm_id = queue->cm_id;
1427 	int status = ev->status;
1428 	const char *rej_msg;
1429 	const struct nvme_rdma_cm_rej *rej_data;
1430 	u8 rej_data_len;
1431 
1432 	rej_msg = rdma_reject_msg(cm_id, status);
1433 	rej_data = rdma_consumer_reject_data(cm_id, ev, &rej_data_len);
1434 
1435 	if (rej_data && rej_data_len >= sizeof(u16)) {
1436 		u16 sts = le16_to_cpu(rej_data->sts);
1437 
1438 		dev_err(queue->ctrl->ctrl.device,
1439 		      "Connect rejected: status %d (%s) nvme status %d (%s).\n",
1440 		      status, rej_msg, sts, nvme_rdma_cm_msg(sts));
1441 	} else {
1442 		dev_err(queue->ctrl->ctrl.device,
1443 			"Connect rejected: status %d (%s).\n", status, rej_msg);
1444 	}
1445 
1446 	return -ECONNRESET;
1447 }
1448 
1449 static int nvme_rdma_addr_resolved(struct nvme_rdma_queue *queue)
1450 {
1451 	int ret;
1452 
1453 	ret = nvme_rdma_create_queue_ib(queue);
1454 	if (ret)
1455 		return ret;
1456 
1457 	ret = rdma_resolve_route(queue->cm_id, NVME_RDMA_CONNECT_TIMEOUT_MS);
1458 	if (ret) {
1459 		dev_err(queue->ctrl->ctrl.device,
1460 			"rdma_resolve_route failed (%d).\n",
1461 			queue->cm_error);
1462 		goto out_destroy_queue;
1463 	}
1464 
1465 	return 0;
1466 
1467 out_destroy_queue:
1468 	nvme_rdma_destroy_queue_ib(queue);
1469 	return ret;
1470 }
1471 
1472 static int nvme_rdma_route_resolved(struct nvme_rdma_queue *queue)
1473 {
1474 	struct nvme_rdma_ctrl *ctrl = queue->ctrl;
1475 	struct rdma_conn_param param = { };
1476 	struct nvme_rdma_cm_req priv = { };
1477 	int ret;
1478 
1479 	param.qp_num = queue->qp->qp_num;
1480 	param.flow_control = 1;
1481 
1482 	param.responder_resources = queue->device->dev->attrs.max_qp_rd_atom;
1483 	/* maximum retry count */
1484 	param.retry_count = 7;
1485 	param.rnr_retry_count = 7;
1486 	param.private_data = &priv;
1487 	param.private_data_len = sizeof(priv);
1488 
1489 	priv.recfmt = cpu_to_le16(NVME_RDMA_CM_FMT_1_0);
1490 	priv.qid = cpu_to_le16(nvme_rdma_queue_idx(queue));
1491 	/*
1492 	 * set the admin queue depth to the minimum size
1493 	 * specified by the Fabrics standard.
1494 	 */
1495 	if (priv.qid == 0) {
1496 		priv.hrqsize = cpu_to_le16(NVME_AQ_DEPTH);
1497 		priv.hsqsize = cpu_to_le16(NVME_AQ_DEPTH - 1);
1498 	} else {
1499 		/*
1500 		 * current interpretation of the fabrics spec
1501 		 * is at minimum you make hrqsize sqsize+1, or a
1502 		 * 1's based representation of sqsize.
1503 		 */
1504 		priv.hrqsize = cpu_to_le16(queue->queue_size);
1505 		priv.hsqsize = cpu_to_le16(queue->ctrl->ctrl.sqsize);
1506 	}
1507 
1508 	ret = rdma_connect(queue->cm_id, &param);
1509 	if (ret) {
1510 		dev_err(ctrl->ctrl.device,
1511 			"rdma_connect failed (%d).\n", ret);
1512 		goto out_destroy_queue_ib;
1513 	}
1514 
1515 	return 0;
1516 
1517 out_destroy_queue_ib:
1518 	nvme_rdma_destroy_queue_ib(queue);
1519 	return ret;
1520 }
1521 
1522 static int nvme_rdma_cm_handler(struct rdma_cm_id *cm_id,
1523 		struct rdma_cm_event *ev)
1524 {
1525 	struct nvme_rdma_queue *queue = cm_id->context;
1526 	int cm_error = 0;
1527 
1528 	dev_dbg(queue->ctrl->ctrl.device, "%s (%d): status %d id %p\n",
1529 		rdma_event_msg(ev->event), ev->event,
1530 		ev->status, cm_id);
1531 
1532 	switch (ev->event) {
1533 	case RDMA_CM_EVENT_ADDR_RESOLVED:
1534 		cm_error = nvme_rdma_addr_resolved(queue);
1535 		break;
1536 	case RDMA_CM_EVENT_ROUTE_RESOLVED:
1537 		cm_error = nvme_rdma_route_resolved(queue);
1538 		break;
1539 	case RDMA_CM_EVENT_ESTABLISHED:
1540 		queue->cm_error = nvme_rdma_conn_established(queue);
1541 		/* complete cm_done regardless of success/failure */
1542 		complete(&queue->cm_done);
1543 		return 0;
1544 	case RDMA_CM_EVENT_REJECTED:
1545 		nvme_rdma_destroy_queue_ib(queue);
1546 		cm_error = nvme_rdma_conn_rejected(queue, ev);
1547 		break;
1548 	case RDMA_CM_EVENT_ROUTE_ERROR:
1549 	case RDMA_CM_EVENT_CONNECT_ERROR:
1550 	case RDMA_CM_EVENT_UNREACHABLE:
1551 		nvme_rdma_destroy_queue_ib(queue);
1552 	case RDMA_CM_EVENT_ADDR_ERROR:
1553 		dev_dbg(queue->ctrl->ctrl.device,
1554 			"CM error event %d\n", ev->event);
1555 		cm_error = -ECONNRESET;
1556 		break;
1557 	case RDMA_CM_EVENT_DISCONNECTED:
1558 	case RDMA_CM_EVENT_ADDR_CHANGE:
1559 	case RDMA_CM_EVENT_TIMEWAIT_EXIT:
1560 		dev_dbg(queue->ctrl->ctrl.device,
1561 			"disconnect received - connection closed\n");
1562 		nvme_rdma_error_recovery(queue->ctrl);
1563 		break;
1564 	case RDMA_CM_EVENT_DEVICE_REMOVAL:
1565 		/* device removal is handled via the ib_client API */
1566 		break;
1567 	default:
1568 		dev_err(queue->ctrl->ctrl.device,
1569 			"Unexpected RDMA CM event (%d)\n", ev->event);
1570 		nvme_rdma_error_recovery(queue->ctrl);
1571 		break;
1572 	}
1573 
1574 	if (cm_error) {
1575 		queue->cm_error = cm_error;
1576 		complete(&queue->cm_done);
1577 	}
1578 
1579 	return 0;
1580 }
1581 
1582 static enum blk_eh_timer_return
1583 nvme_rdma_timeout(struct request *rq, bool reserved)
1584 {
1585 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1586 
1587 	dev_warn(req->queue->ctrl->ctrl.device,
1588 		 "I/O %d QID %d timeout, reset controller\n",
1589 		 rq->tag, nvme_rdma_queue_idx(req->queue));
1590 
1591 	/* queue error recovery */
1592 	nvme_rdma_error_recovery(req->queue->ctrl);
1593 
1594 	/* fail with DNR on cmd timeout */
1595 	nvme_req(rq)->status = NVME_SC_ABORT_REQ | NVME_SC_DNR;
1596 
1597 	return BLK_EH_HANDLED;
1598 }
1599 
1600 /*
1601  * We cannot accept any other command until the Connect command has completed.
1602  */
1603 static inline blk_status_t
1604 nvme_rdma_is_ready(struct nvme_rdma_queue *queue, struct request *rq)
1605 {
1606 	if (unlikely(!test_bit(NVME_RDMA_Q_LIVE, &queue->flags)))
1607 		return nvmf_check_init_req(&queue->ctrl->ctrl, rq);
1608 	return BLK_STS_OK;
1609 }
1610 
1611 static blk_status_t nvme_rdma_queue_rq(struct blk_mq_hw_ctx *hctx,
1612 		const struct blk_mq_queue_data *bd)
1613 {
1614 	struct nvme_ns *ns = hctx->queue->queuedata;
1615 	struct nvme_rdma_queue *queue = hctx->driver_data;
1616 	struct request *rq = bd->rq;
1617 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1618 	struct nvme_rdma_qe *sqe = &req->sqe;
1619 	struct nvme_command *c = sqe->data;
1620 	struct ib_device *dev;
1621 	blk_status_t ret;
1622 	int err;
1623 
1624 	WARN_ON_ONCE(rq->tag < 0);
1625 
1626 	ret = nvme_rdma_is_ready(queue, rq);
1627 	if (unlikely(ret))
1628 		return ret;
1629 
1630 	dev = queue->device->dev;
1631 	ib_dma_sync_single_for_cpu(dev, sqe->dma,
1632 			sizeof(struct nvme_command), DMA_TO_DEVICE);
1633 
1634 	ret = nvme_setup_cmd(ns, rq, c);
1635 	if (ret)
1636 		return ret;
1637 
1638 	blk_mq_start_request(rq);
1639 
1640 	err = nvme_rdma_map_data(queue, rq, c);
1641 	if (unlikely(err < 0)) {
1642 		dev_err(queue->ctrl->ctrl.device,
1643 			     "Failed to map data (%d)\n", err);
1644 		nvme_cleanup_cmd(rq);
1645 		goto err;
1646 	}
1647 
1648 	sqe->cqe.done = nvme_rdma_send_done;
1649 
1650 	ib_dma_sync_single_for_device(dev, sqe->dma,
1651 			sizeof(struct nvme_command), DMA_TO_DEVICE);
1652 
1653 	err = nvme_rdma_post_send(queue, sqe, req->sge, req->num_sge,
1654 			req->mr ? &req->reg_wr.wr : NULL);
1655 	if (unlikely(err)) {
1656 		nvme_rdma_unmap_data(queue, rq);
1657 		goto err;
1658 	}
1659 
1660 	return BLK_STS_OK;
1661 err:
1662 	if (err == -ENOMEM || err == -EAGAIN)
1663 		return BLK_STS_RESOURCE;
1664 	return BLK_STS_IOERR;
1665 }
1666 
1667 static int nvme_rdma_poll(struct blk_mq_hw_ctx *hctx, unsigned int tag)
1668 {
1669 	struct nvme_rdma_queue *queue = hctx->driver_data;
1670 	struct ib_cq *cq = queue->ib_cq;
1671 	struct ib_wc wc;
1672 	int found = 0;
1673 
1674 	while (ib_poll_cq(cq, 1, &wc) > 0) {
1675 		struct ib_cqe *cqe = wc.wr_cqe;
1676 
1677 		if (cqe) {
1678 			if (cqe->done == nvme_rdma_recv_done)
1679 				found |= __nvme_rdma_recv_done(cq, &wc, tag);
1680 			else
1681 				cqe->done(cq, &wc);
1682 		}
1683 	}
1684 
1685 	return found;
1686 }
1687 
1688 static void nvme_rdma_complete_rq(struct request *rq)
1689 {
1690 	struct nvme_rdma_request *req = blk_mq_rq_to_pdu(rq);
1691 
1692 	nvme_rdma_unmap_data(req->queue, rq);
1693 	nvme_complete_rq(rq);
1694 }
1695 
1696 static int nvme_rdma_map_queues(struct blk_mq_tag_set *set)
1697 {
1698 	struct nvme_rdma_ctrl *ctrl = set->driver_data;
1699 
1700 	return blk_mq_rdma_map_queues(set, ctrl->device->dev, 0);
1701 }
1702 
1703 static const struct blk_mq_ops nvme_rdma_mq_ops = {
1704 	.queue_rq	= nvme_rdma_queue_rq,
1705 	.complete	= nvme_rdma_complete_rq,
1706 	.init_request	= nvme_rdma_init_request,
1707 	.exit_request	= nvme_rdma_exit_request,
1708 	.init_hctx	= nvme_rdma_init_hctx,
1709 	.poll		= nvme_rdma_poll,
1710 	.timeout	= nvme_rdma_timeout,
1711 	.map_queues	= nvme_rdma_map_queues,
1712 };
1713 
1714 static const struct blk_mq_ops nvme_rdma_admin_mq_ops = {
1715 	.queue_rq	= nvme_rdma_queue_rq,
1716 	.complete	= nvme_rdma_complete_rq,
1717 	.init_request	= nvme_rdma_init_request,
1718 	.exit_request	= nvme_rdma_exit_request,
1719 	.init_hctx	= nvme_rdma_init_admin_hctx,
1720 	.timeout	= nvme_rdma_timeout,
1721 };
1722 
1723 static void nvme_rdma_shutdown_ctrl(struct nvme_rdma_ctrl *ctrl, bool shutdown)
1724 {
1725 	cancel_work_sync(&ctrl->err_work);
1726 	cancel_delayed_work_sync(&ctrl->reconnect_work);
1727 
1728 	if (ctrl->ctrl.queue_count > 1) {
1729 		nvme_stop_queues(&ctrl->ctrl);
1730 		blk_mq_tagset_busy_iter(&ctrl->tag_set,
1731 					nvme_cancel_request, &ctrl->ctrl);
1732 		nvme_rdma_destroy_io_queues(ctrl, shutdown);
1733 	}
1734 
1735 	if (shutdown)
1736 		nvme_shutdown_ctrl(&ctrl->ctrl);
1737 	else
1738 		nvme_disable_ctrl(&ctrl->ctrl, ctrl->ctrl.cap);
1739 
1740 	blk_mq_quiesce_queue(ctrl->ctrl.admin_q);
1741 	blk_mq_tagset_busy_iter(&ctrl->admin_tag_set,
1742 				nvme_cancel_request, &ctrl->ctrl);
1743 	blk_mq_unquiesce_queue(ctrl->ctrl.admin_q);
1744 	nvme_rdma_destroy_admin_queue(ctrl, shutdown);
1745 }
1746 
1747 static void nvme_rdma_delete_ctrl(struct nvme_ctrl *ctrl)
1748 {
1749 	nvme_rdma_shutdown_ctrl(to_rdma_ctrl(ctrl), true);
1750 }
1751 
1752 static void nvme_rdma_reset_ctrl_work(struct work_struct *work)
1753 {
1754 	struct nvme_rdma_ctrl *ctrl =
1755 		container_of(work, struct nvme_rdma_ctrl, ctrl.reset_work);
1756 	int ret;
1757 	bool changed;
1758 
1759 	nvme_stop_ctrl(&ctrl->ctrl);
1760 	nvme_rdma_shutdown_ctrl(ctrl, false);
1761 
1762 	if (!nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_RECONNECTING)) {
1763 		/* state change failure should never happen */
1764 		WARN_ON_ONCE(1);
1765 		return;
1766 	}
1767 
1768 	ret = nvme_rdma_configure_admin_queue(ctrl, false);
1769 	if (ret)
1770 		goto out_fail;
1771 
1772 	if (ctrl->ctrl.queue_count > 1) {
1773 		ret = nvme_rdma_configure_io_queues(ctrl, false);
1774 		if (ret)
1775 			goto out_fail;
1776 	}
1777 
1778 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1779 	if (!changed) {
1780 		/* state change failure is ok if we're in DELETING state */
1781 		WARN_ON_ONCE(ctrl->ctrl.state != NVME_CTRL_DELETING);
1782 		return;
1783 	}
1784 
1785 	nvme_start_ctrl(&ctrl->ctrl);
1786 
1787 	return;
1788 
1789 out_fail:
1790 	dev_warn(ctrl->ctrl.device, "Removing after reset failure\n");
1791 	nvme_remove_namespaces(&ctrl->ctrl);
1792 	nvme_rdma_shutdown_ctrl(ctrl, true);
1793 	nvme_uninit_ctrl(&ctrl->ctrl);
1794 	nvme_put_ctrl(&ctrl->ctrl);
1795 }
1796 
1797 static const struct nvme_ctrl_ops nvme_rdma_ctrl_ops = {
1798 	.name			= "rdma",
1799 	.module			= THIS_MODULE,
1800 	.flags			= NVME_F_FABRICS,
1801 	.reg_read32		= nvmf_reg_read32,
1802 	.reg_read64		= nvmf_reg_read64,
1803 	.reg_write32		= nvmf_reg_write32,
1804 	.free_ctrl		= nvme_rdma_free_ctrl,
1805 	.submit_async_event	= nvme_rdma_submit_async_event,
1806 	.delete_ctrl		= nvme_rdma_delete_ctrl,
1807 	.get_address		= nvmf_get_address,
1808 };
1809 
1810 static inline bool
1811 __nvme_rdma_options_match(struct nvme_rdma_ctrl *ctrl,
1812 	struct nvmf_ctrl_options *opts)
1813 {
1814 	char *stdport = __stringify(NVME_RDMA_IP_PORT);
1815 
1816 
1817 	if (!nvmf_ctlr_matches_baseopts(&ctrl->ctrl, opts) ||
1818 	    strcmp(opts->traddr, ctrl->ctrl.opts->traddr))
1819 		return false;
1820 
1821 	if (opts->mask & NVMF_OPT_TRSVCID &&
1822 	    ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1823 		if (strcmp(opts->trsvcid, ctrl->ctrl.opts->trsvcid))
1824 			return false;
1825 	} else if (opts->mask & NVMF_OPT_TRSVCID) {
1826 		if (strcmp(opts->trsvcid, stdport))
1827 			return false;
1828 	} else if (ctrl->ctrl.opts->mask & NVMF_OPT_TRSVCID) {
1829 		if (strcmp(stdport, ctrl->ctrl.opts->trsvcid))
1830 			return false;
1831 	}
1832 	/* else, it's a match as both have stdport. Fall to next checks */
1833 
1834 	/*
1835 	 * checking the local address is rough. In most cases, one
1836 	 * is not specified and the host port is selected by the stack.
1837 	 *
1838 	 * Assume no match if:
1839 	 *  local address is specified and address is not the same
1840 	 *  local address is not specified but remote is, or vice versa
1841 	 *    (admin using specific host_traddr when it matters).
1842 	 */
1843 	if (opts->mask & NVMF_OPT_HOST_TRADDR &&
1844 	    ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR) {
1845 		if (strcmp(opts->host_traddr, ctrl->ctrl.opts->host_traddr))
1846 			return false;
1847 	} else if (opts->mask & NVMF_OPT_HOST_TRADDR ||
1848 		   ctrl->ctrl.opts->mask & NVMF_OPT_HOST_TRADDR)
1849 		return false;
1850 	/*
1851 	 * if neither controller had an host port specified, assume it's
1852 	 * a match as everything else matched.
1853 	 */
1854 
1855 	return true;
1856 }
1857 
1858 /*
1859  * Fails a connection request if it matches an existing controller
1860  * (association) with the same tuple:
1861  * <Host NQN, Host ID, local address, remote address, remote port, SUBSYS NQN>
1862  *
1863  * if local address is not specified in the request, it will match an
1864  * existing controller with all the other parameters the same and no
1865  * local port address specified as well.
1866  *
1867  * The ports don't need to be compared as they are intrinsically
1868  * already matched by the port pointers supplied.
1869  */
1870 static bool
1871 nvme_rdma_existing_controller(struct nvmf_ctrl_options *opts)
1872 {
1873 	struct nvme_rdma_ctrl *ctrl;
1874 	bool found = false;
1875 
1876 	mutex_lock(&nvme_rdma_ctrl_mutex);
1877 	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
1878 		found = __nvme_rdma_options_match(ctrl, opts);
1879 		if (found)
1880 			break;
1881 	}
1882 	mutex_unlock(&nvme_rdma_ctrl_mutex);
1883 
1884 	return found;
1885 }
1886 
1887 static struct nvme_ctrl *nvme_rdma_create_ctrl(struct device *dev,
1888 		struct nvmf_ctrl_options *opts)
1889 {
1890 	struct nvme_rdma_ctrl *ctrl;
1891 	int ret;
1892 	bool changed;
1893 	char *port;
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 		port = opts->trsvcid;
1903 	else
1904 		port = __stringify(NVME_RDMA_IP_PORT);
1905 
1906 	ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1907 			opts->traddr, port, &ctrl->addr);
1908 	if (ret) {
1909 		pr_err("malformed address passed: %s:%s\n", opts->traddr, port);
1910 		goto out_free_ctrl;
1911 	}
1912 
1913 	if (opts->mask & NVMF_OPT_HOST_TRADDR) {
1914 		ret = inet_pton_with_scope(&init_net, AF_UNSPEC,
1915 			opts->host_traddr, NULL, &ctrl->src_addr);
1916 		if (ret) {
1917 			pr_err("malformed src address passed: %s\n",
1918 			       opts->host_traddr);
1919 			goto out_free_ctrl;
1920 		}
1921 	}
1922 
1923 	if (!opts->duplicate_connect && nvme_rdma_existing_controller(opts)) {
1924 		ret = -EALREADY;
1925 		goto out_free_ctrl;
1926 	}
1927 
1928 	ret = nvme_init_ctrl(&ctrl->ctrl, dev, &nvme_rdma_ctrl_ops,
1929 				0 /* no quirks, we're perfect! */);
1930 	if (ret)
1931 		goto out_free_ctrl;
1932 
1933 	INIT_DELAYED_WORK(&ctrl->reconnect_work,
1934 			nvme_rdma_reconnect_ctrl_work);
1935 	INIT_WORK(&ctrl->err_work, nvme_rdma_error_recovery_work);
1936 	INIT_WORK(&ctrl->ctrl.reset_work, nvme_rdma_reset_ctrl_work);
1937 
1938 	ctrl->ctrl.queue_count = opts->nr_io_queues + 1; /* +1 for admin queue */
1939 	ctrl->ctrl.sqsize = opts->queue_size - 1;
1940 	ctrl->ctrl.kato = opts->kato;
1941 
1942 	ret = -ENOMEM;
1943 	ctrl->queues = kcalloc(ctrl->ctrl.queue_count, sizeof(*ctrl->queues),
1944 				GFP_KERNEL);
1945 	if (!ctrl->queues)
1946 		goto out_uninit_ctrl;
1947 
1948 	ret = nvme_rdma_configure_admin_queue(ctrl, true);
1949 	if (ret)
1950 		goto out_kfree_queues;
1951 
1952 	/* sanity check icdoff */
1953 	if (ctrl->ctrl.icdoff) {
1954 		dev_err(ctrl->ctrl.device, "icdoff is not supported!\n");
1955 		ret = -EINVAL;
1956 		goto out_remove_admin_queue;
1957 	}
1958 
1959 	/* sanity check keyed sgls */
1960 	if (!(ctrl->ctrl.sgls & (1 << 20))) {
1961 		dev_err(ctrl->ctrl.device, "Mandatory keyed sgls are not support\n");
1962 		ret = -EINVAL;
1963 		goto out_remove_admin_queue;
1964 	}
1965 
1966 	if (opts->queue_size > ctrl->ctrl.maxcmd) {
1967 		/* warn if maxcmd is lower than queue_size */
1968 		dev_warn(ctrl->ctrl.device,
1969 			"queue_size %zu > ctrl maxcmd %u, clamping down\n",
1970 			opts->queue_size, ctrl->ctrl.maxcmd);
1971 		opts->queue_size = ctrl->ctrl.maxcmd;
1972 	}
1973 
1974 	if (opts->queue_size > ctrl->ctrl.sqsize + 1) {
1975 		/* warn if sqsize is lower than queue_size */
1976 		dev_warn(ctrl->ctrl.device,
1977 			"queue_size %zu > ctrl sqsize %u, clamping down\n",
1978 			opts->queue_size, ctrl->ctrl.sqsize + 1);
1979 		opts->queue_size = ctrl->ctrl.sqsize + 1;
1980 	}
1981 
1982 	if (opts->nr_io_queues) {
1983 		ret = nvme_rdma_configure_io_queues(ctrl, true);
1984 		if (ret)
1985 			goto out_remove_admin_queue;
1986 	}
1987 
1988 	changed = nvme_change_ctrl_state(&ctrl->ctrl, NVME_CTRL_LIVE);
1989 	WARN_ON_ONCE(!changed);
1990 
1991 	dev_info(ctrl->ctrl.device, "new ctrl: NQN \"%s\", addr %pISpcs\n",
1992 		ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
1993 
1994 	nvme_get_ctrl(&ctrl->ctrl);
1995 
1996 	mutex_lock(&nvme_rdma_ctrl_mutex);
1997 	list_add_tail(&ctrl->list, &nvme_rdma_ctrl_list);
1998 	mutex_unlock(&nvme_rdma_ctrl_mutex);
1999 
2000 	nvme_start_ctrl(&ctrl->ctrl);
2001 
2002 	return &ctrl->ctrl;
2003 
2004 out_remove_admin_queue:
2005 	nvme_rdma_destroy_admin_queue(ctrl, true);
2006 out_kfree_queues:
2007 	kfree(ctrl->queues);
2008 out_uninit_ctrl:
2009 	nvme_uninit_ctrl(&ctrl->ctrl);
2010 	nvme_put_ctrl(&ctrl->ctrl);
2011 	if (ret > 0)
2012 		ret = -EIO;
2013 	return ERR_PTR(ret);
2014 out_free_ctrl:
2015 	kfree(ctrl);
2016 	return ERR_PTR(ret);
2017 }
2018 
2019 static struct nvmf_transport_ops nvme_rdma_transport = {
2020 	.name		= "rdma",
2021 	.required_opts	= NVMF_OPT_TRADDR,
2022 	.allowed_opts	= NVMF_OPT_TRSVCID | NVMF_OPT_RECONNECT_DELAY |
2023 			  NVMF_OPT_HOST_TRADDR | NVMF_OPT_CTRL_LOSS_TMO,
2024 	.create_ctrl	= nvme_rdma_create_ctrl,
2025 };
2026 
2027 static void nvme_rdma_remove_one(struct ib_device *ib_device, void *client_data)
2028 {
2029 	struct nvme_rdma_ctrl *ctrl;
2030 
2031 	/* Delete all controllers using this device */
2032 	mutex_lock(&nvme_rdma_ctrl_mutex);
2033 	list_for_each_entry(ctrl, &nvme_rdma_ctrl_list, list) {
2034 		if (ctrl->device->dev != ib_device)
2035 			continue;
2036 		dev_info(ctrl->ctrl.device,
2037 			"Removing ctrl: NQN \"%s\", addr %pISp\n",
2038 			ctrl->ctrl.opts->subsysnqn, &ctrl->addr);
2039 		nvme_delete_ctrl(&ctrl->ctrl);
2040 	}
2041 	mutex_unlock(&nvme_rdma_ctrl_mutex);
2042 
2043 	flush_workqueue(nvme_wq);
2044 }
2045 
2046 static struct ib_client nvme_rdma_ib_client = {
2047 	.name   = "nvme_rdma",
2048 	.remove = nvme_rdma_remove_one
2049 };
2050 
2051 static int __init nvme_rdma_init_module(void)
2052 {
2053 	int ret;
2054 
2055 	ret = ib_register_client(&nvme_rdma_ib_client);
2056 	if (ret)
2057 		return ret;
2058 
2059 	ret = nvmf_register_transport(&nvme_rdma_transport);
2060 	if (ret)
2061 		goto err_unreg_client;
2062 
2063 	return 0;
2064 
2065 err_unreg_client:
2066 	ib_unregister_client(&nvme_rdma_ib_client);
2067 	return ret;
2068 }
2069 
2070 static void __exit nvme_rdma_cleanup_module(void)
2071 {
2072 	nvmf_unregister_transport(&nvme_rdma_transport);
2073 	ib_unregister_client(&nvme_rdma_ib_client);
2074 }
2075 
2076 module_init(nvme_rdma_init_module);
2077 module_exit(nvme_rdma_cleanup_module);
2078 
2079 MODULE_LICENSE("GPL v2");
2080